Datasets:

ArmelR

/

the-pile-splitted

like 15

--- abstract: 'Polarized magnetic dipole (M1) emission lines provide important diagnostics for the magnetic field dominating the evolution of the solar corona. This paper advances a multi-line technique using specific combinations of M1 lines to infer the full vector magnetic field for regions of optically thin emission that can be localized along a given line of sight. Our analytical formalism is a generalization of the “single-point inversion" approach introduced by @Plowman14. We show that combinations of M1 transitions for which each is either a $J=1\rightarrow0$ transition or has equal Landé g-factors for the upper and lower levels contain degenerate spectropolarimetric information that prohibits the application of the single-point inversion technique. This may include the pair of lines discussed by @Plowman14. We identify the 10747 Å and 14301 Å lines as one alternative combination for implementing this technique. Our sensitivity analysis, based on coronal loop properties, suggests that for photon noise levels around $10^{-4}$ of the line intensity, which will be achievable with the National Science Foundation’s *Daniel K. Inouye Solar Telescope*, magnetic fields with sufficient strength (${\sim}10$ G) and not severely inclined to the line-of-sight ($\lesssim 35^{\circ}$) can be recovered with this method. Degenerate solutions exist; though, we discuss how added constraints may help resolve them or reduce their number.' author: - 'Gabriel I. Dima' - 'Thomas A. Schad' bibliography: - 'gdrefs.bib' --- Introduction {#sec:intro} ============ In the low density environment of the solar corona, the magnetic field governs both the atmospheric structure and the energetics of active solar phenomena like flares, coronal mass ejections and the solar wind. Two promising techniques available for remotely measuring the coronal magnetic field are spectropolarimetric measurements of emission lines in the optical part of the spectrum and radio observations from coronal regions with different plasma frequencies [@Casini17]; however, each face observational and interpretation challenges. The brightest observable visible/infrared (Vis/IR) coronal emission lines are magnetic dipole (M1) lines [@Judge98; @DelZanna18a]. Photoexcitation of these lines by anisotropic disc radiation leads to scattering induced linear polarization with amplitudes of ${\sim}10^{-2}$ to $10^{-1}$ of the total line intensity [e.g., @Arnaud87; @Tomczyk08]). The presence of the magnetic field modifies the linear polarization through the saturated Hanle effect (*i.e.*, its amplitude is not sensitive to the magnetic field intensity) and induces a small circularly polarized signal through the longitudinal Zeeman effect with relative amplitudes of ${\sim}10^{-4}$ to $10^{-3}$ for a 10 G magnetic field [@Harvey69; @Lin00; @Lin04]. Achieving high sensitivity measurements of these signals is challenged by the large ${\sim}10^{5}$ brightness contrast between the solar disc and coronal emission in the Vis/IR [@Arnaud82; @Arnaud87; @Kuhn96; @Madsen19] and requires mitigation of scattered light from the sky and telescope optics. Observations at longer wavelengths are typically more sensitive because the scattered light performance of telescope optics improves and the Zeeman splitting increases faster than the thermal line width [@Penn14]. The upcoming 4m *Daniel K. Inouye Solar Telescope* (DKIST) will have two first light spectropolarimetric instruments targeting multiple IR emission lines built to address these challenges: DL-NIRSP and Cryo-NIRSP [@Rimmele15]. Key theoretical advances in understanding the formation of polarized M1 lines in the corona have enabled new diagnostic capabilities based on synthesizing polarimetric signals through the optically thin corona [@Charvin65; @SahalBrechot77; @House77; @Casini99; @Casini00]. Due in part to the current sparsity of circular polarization detections, most progress has been made by comparing observed linear polarized signals of a single M1 line to those synthesized through forward models [@Judge06; @Dove11; @Rachmeler13; @Dalmasse19] or by using time-based tomographic methods to invert the global coronal magnetic structure [@Kramar06; @Kramar13], again primarily using a single M1 line. These techniques necessarily consider the emission contributed by all the plasma along each line of sight. Work by [@Liu08] succeeded in comparing measured circular polarization signals with those synthesized through an active region. In contrast to the above approaches, here we focus on using full Stokes observations of multiple M1 lines to invert the full vector magnetic field for an isolated region with uniform magnetohydrodynamic properties: density, temperature and magnetic field. This region either dominates the total emission along the line of sight or can potentially be separated from the foreground/background emission with subtraction methods. Good candidate targets are coronal loops that appear in contrast against the background corona in ultraviolet [@Brooks12; @Aschwanden13; @Xie17] and IR [@Tomczyk08]. @Plowman14 demonstrated that, under this “single point approximation”, using a combination of full Stokes observations from the 10747 Å and 10798 Å lines, all three components of the magnetic field could be analytically calculated without explicit knowledge of the thermodynamic parameters that influence the amplitude of the linear polarization. In this work we examine the single point inversion method further by generalizing the approach of @Plowman14 to consider, in principle, any combination of M1 lines. The analytical formalism we derive in Section \[analytic\_solution\] for two line observations reveals that only specific line combinations may be used for the inversion due to fundamental quantum mechanical degeneracies in the polarized emission coefficients. As a result, we learn that the line pair used by @Plowman14 may not be adequate for single point inversions if LS coupling is valid for these lines. In Section \[sec:noise\] we propose combining the 10747 Å and 14301 Å coronal lines to implement the single point inversion. We also evaluate the impact photon noise has on this line combination when attempting to infer the magnetic field vector. In Section \[discussion\] we discuss ambiguities inherent in the present inversion approach and as well as other related diagnostics. Analytic inverse solution {#analytic_solution} ========================= Starting from equations 35(a-c) in @Casini99 [@Casini00], and making use of the geometric tensors for magnetic dipole radiation given by their equations 39(a-d), the polarized emission coefficients ($\epsilon_{i}$) for the Stokes parameters ($i=I,Q,U,V$) of an M1 line can be written as a function of wavelength (as opposed to angular frequency) as: $$\begin{aligned} \epsilon_I(\lambda) &= C\left[1 + \frac{3}{2\sqrt{2}}D\sigma^2_0\left(\cos^2{\Theta_B} - \frac{1}{3}\right)\right]\phi(\lambda-\lambda_0)\label{eqI0}\\ \epsilon_Q(\lambda) &= \frac{3}{2\sqrt{2}}CD\sigma^2_0\sin^2{\Theta_B}\cos{(2\gamma_B)}\phi(\lambda-\lambda_0) \label{eqQ0}\\ \epsilon_U(\lambda) &= -\frac{3}{2\sqrt{2}}CD\sigma^2_0\sin^2{\Theta_B}\sin{(2\gamma_B)}\phi(\lambda-\lambda_0) \label{eqU0}\\ \epsilon_V(\lambda) &= -C\frac{\lambda^2}{c}\nu_L\cos{\Theta_B}\left[\bar{g}+E\sigma^2_0\right]\phi'(\lambda-\lambda_0) \label{eqV0}\end{aligned}$$ where $C$ (*i.e.*, $C_{JJ_{0}}$ in the original notation) is the emission coefficient (without stimulated emission) in the unpolarized case and depends on the total upper level population and the Einstein A coefficient for the transition. $D=D_{JJ_{0}}$, $E=E_{JJ_{0}}$ and $\bar{g}=\bar{g}_{JJ_{0}}$ in the original notation are atomic parameters that depend on the quantum numbers of the upper and lower levels of the transition, $J$ and $J_0$ respectively. $\Theta_B$ is the inclination angle of the magnetic field with respect to the line of sight while $\gamma_B$ is the azimuth angle of the reference direction of linear polarization (r.d.l.p) measured from the projected magnetic field direction (see Figure \[fig:coordinate\_system\] as well as Figure 5 in [@Casini99]). The Larmor frequency $\nu_{L}$ is proportional to the magnetic field strength B. $\phi(\lambda-\lambda_0)$ refers to the Doppler dominated normalized ($\int \phi{\ }d\lambda = 1$) spectral profile centered at $\lambda_0$ while $\phi'(\lambda-\lambda_0)$ is its derivative with respect to wavelength. The scalar quantity $\sigma^2_0$ refers to the fractional atomic alignment of the upper level of the transition which describes the imbalance of the magnetic sublevel populations for different absolute values of the magnetic quantum number. It is equal to $\rho^2_0/\rho^0_0$ where the tensor $\rho^{K}_{Q}$ is the irreducible spherical statistical tensor representation of the atomic density matrix. $\sigma^2_0$ can be positive or negative and is generally a function of the exciting photospheric radiation anisotropy (which varies with height), the plasma density and temperature, and the local inclination angle of the magnetic field in the local radial frame of reference $\theta_B$. For the M1 lines of interest, $\sigma^2_0$ can often be well approximated by a positive definite factor, $k_{J}(T_{e},n_{e},h),$ dependent on the electron temperature, electron density, and height multiplied by a functional dependence on $\theta_B$ given by $3 \cos^2{\theta_B} - 1$ [@Judge07], which gives rise to the Van Vleck ambiguity discussed in Section \[ambiguities\]. Note that $\sigma^2_0$ does not depend on the magnetic field intensity. In principle, given proper constraints on the local plasma properties and the radiation field, as well as numerical calculations of $k_{J}$, full Stokes observations of a single M1 line can be used to derive the full magnetic field vector, within certain limitations and within the single point approximation as discussed in [@Judge07]. In absence of such extra constraints, a full set of Stokes observations of a single M1 line provide four measurements for five unknowns (*i.e.*, $C,\sigma^2_0,\gamma_B, \Theta_{B},B$). $\gamma_B$ can be derived using the ratio of U and Q; but the remaining variables remain coupled. With some limitations, to be discussed in Section \[blos\], the projected magnetic field intensity along the line of sight ($B_{LOS}\equiv B\cos\Theta_B$) can be derived from polarimetric observations of a single line. Likewise, the value of $C$ can be determined from I by neglecting the atomic alignment contribution; but $\sigma^2_0$ and $\Theta_B$ cannot be directly determined. ![Angular coordinates between three vectors passing through a point of emission in the corona: the outward radial direction (**R**), magnetic field (**B**) and line of sight (**k**). The projected magnetic field in a plane perpendicular to the line of sight is characterized by the azimuth angle $\gamma_B$ measured clockwise towards the reference direction for linear polarization (r.d.l.p.) with a 180[$^{\circ}$]{}ambiguity since the linear polarization direction is not oriented. Figure adapted from Figure 5 in @Casini99.[]{data-label="fig:coordinate_system"}](coordinate_system_6.pdf){width="3.3in"} @Plowman14 showed that with full Stokes measurements of two M1 lines, in particular the 10747, 10798 Å line pair, it is possible to obtain explicit analytic formulas for all the unknown parameters in equations \[eqI0\]-\[eqV0\] under the single point assumption. An additional spectral line provides four more measured quantities but only two additional unknowns, as the magnetic field properties are assumed to be the same. @Plowman14 focused only on the lines, using analytic formulas for the D and E parameters from Equation 38(a-b) in @Casini99 before solving for the unknown parameters. To extend this approach to other combinations of M1 emission lines, here we retain the symbolic form for the parameters D, E. Under the single point approximation, the line is strongly dominated by emission from a single region in space so the measured polarized Stokes profiles $S_i(\lambda)$ can be approximated as: $$\begin{aligned} \begin{array}{ll} S_i(\lambda) = \int_0^{\infty}\epsilon_i(\lambda)ds \approx \epsilon_i(\lambda)\Delta s & (i=I,Q,U,V) \end{array}\end{aligned}$$ where the integral is over emission along the line of sight and $\Delta s$ is the length of the homogeneous single point region. From the measured dispersed Stokes profiles the wavelength-independent Stokes amplitudes for the two lines can then be derived as: $$\begin{aligned} I &= \frac{S_I}{\phi(\lambda-\lambda_{0})} \approx \frac{\epsilon_I\Delta s}{\phi(\lambda-\lambda_{0})} \nonumber \\ &\approx I_0\left[1 + \frac{3}{2\sqrt{2}}D\sigma^2_0\left(\cos^2{\Theta_B} - \frac{1}{3}\right)\right] \label{eqI}\\ L &= \frac{\sqrt{S_Q^2 + S_U^2}}{\phi(\lambda-\lambda_{0})} \approx \frac{\sqrt{\epsilon_Q^2 + \epsilon_U^2}\Delta s}{\phi(\lambda-\lambda_{0})} \nonumber \\ &\approx \pm \frac{3}{2\sqrt{2}} D I_0 \sigma^2_0 \sin^2{\Theta_B} \label{eqL}\\ V &= \frac{cS_V}{\lambda^2\phi'(\lambda-\lambda_{0})} \approx \frac{c\epsilon_V\Delta s}{\lambda^2\phi'(\lambda-\lambda_{0})} \nonumber \\ &\approx -\frac{1}{\sqrt{2}}I_0\nu_L\cos{\Theta_B}\left[\left(\sqrt{2}+D\sigma^2_0\right)\bar{g}+FD\sigma^2_0\right] \label{eqV}\end{aligned}$$ where $I_0 \equiv C\Delta s$ is a constant with units of intensity. We combine the Q and U Stokes parameters into a single parameter L, which is the total linear polarization. We keep both signs of the quadratic sum because, as already stated, the sign of the atomic alignment $\sigma^2_0$ can be either positive or negative while the measured quantity L is always positive. We also introduce the parameter F defined as: $$\begin{aligned} F &\equiv \sqrt{2}\frac{E}{D} - \bar{g} \nonumber\\ &= \frac{3}{4}[J(J+1) - J_0(J_0+1) -2](g_1 - g_0) \label{eqF}\end{aligned}$$ where $g_1$ and $g_0$ are the Landé factors for the upper and lower levels in the transition. The analytic expression for F can be derived from Equations 36b-c in @Casini99 for $D$ and $E$ by expanding the Wigner 6-j and 9-j coefficients and is valid for all M1 transitions with $\Delta J=0,\pm1$. This parameter is introduced in hindsight since it will be shown to play an important role in discriminating between lines that contain non-degenerate information. The key to the two-line inversion involves solving for the angle $\Theta_B$ that gives the same value for B assuming lines are emitted from the same spatial point with the same magnetic field properties. We first write the magnetic field strength B as a function of $\Theta_B$ by substituting equations \[eqI\] and \[eqL\] into equation \[eqV\] and using the expression for the Larmor frequency (*i.e.*, $\nu_L=(\mu_B/h)B$): $$\begin{aligned} B = \frac{h}{\mu_B}\frac{-V}{\left[(I \pm L)\bar{g} \pm \frac{2}{3}F\frac{L}{\sin^2{\Theta_B}}\right]\cos{\Theta_B}} \label{eqB_num1}\end{aligned}$$ where $h$ and the $\mu_B$ are Planck’s constant and the Bohr magneton respectively. Using this equation and subscripts 1 and 2 for the two lines, we assume $B_{1} \equiv B_{2}$ and $\Theta_{B1} \equiv \Theta_{B2}$ and then solve for $\sin^{2}\Theta_{B}$. The other unknown parameters then follow from Equations \[eqI\], \[eqL\],and \[eqB\_num1\]. To summarize, the analytical inversion equations are as follows: $$\begin{aligned} \sin^{2}\Theta_{B} &= \pm \frac{2}{3}\frac{F_1L_1V_2 - F_2L_2V_1}{\bar{g}_2V_1(I_2 \pm L_2) - \bar{g}_1V_2(I_1 \pm L_1)}, \label{eqsin}\\[5pt] I_{0i} &= I_i \pm L_i\left(1 - \frac{2}{3\sin^2{\Theta_B}}\right), \\[5pt] \sigma^{2}_{0i} &= \frac{\pm L_i}{\sqrt{\frac{9}{8}} I_{0i} D_i \sin^{2}\Theta_{B}}, \\[5pt] B &= \frac{h}{\mu_B}\frac{-V_{i}}{\left[(I_i \pm L_i)\bar{g}_i \pm \frac{2}{3}F_i\frac{L_i}{\sin^{2}{\Theta_{B}}}\right]\cos{\Theta_{B}}},\label{eqB}\\ \gamma_B &= \left\{ \begin{array}{ll} -\frac{1}{2}\arctan\left(\frac{U_i}{Q_i}\right) + \{0,\pi\}, & \sigma^2_{0i} > 0 \\[5pt] -\frac{1}{2}\arctan\left(\frac{U_i}{Q_i}\right) + \{\frac{\pi}{2}, \frac{3\pi}{2}\}, & \sigma^2_{0i} < 0 \\ \end{array} \right. \label{eqgamma}\end{aligned}$$ where the arctangent function must account for the signs of the arguments to determine the quadrant for the result (usually implemented numerically as the arctan2 function). Since M1 lines are in the saturated Hanle regime, the magnetic field projection ($\gamma_B$) is parallel or perpendicular to the direction of linear polarization depending on sign of the alignment and the field inclination relative to the outward radial direction $\theta_B$ (Van Vleck effect, see discussion in Section \[ambiguities\]). In addition, the reference direction of linear polarization has a $180^{\circ}$ ambiguity, which introduces the additional solutions indicated in the curly brackets of Equation \[eqgamma\]. Consequently, given two measured Stokes vectors this inversion method produces up to four degenerate solutions for the magnetic field (discussed further in Sections \[sec:noise\] and  \[ambiguities\]). The sign for $\cos\Theta_B$ is chosen such that B is positive in equation \[eqB\] since physically the magnetic field strength is a positive quantity. An important result can be seen from equation \[eqsin\]: emission lines with $F=0$ contain degenerate information about the magnetic field such that no unique solution for $\Theta_B$ can be deduced from the two Stokes vectors. To apply this type of inversion emission lines must be selected such that at least one line has $F\ne0$. From Equation \[eqF\], F is zero when the upper and lower levels have equal Landé g-factors ($g_1=g_0$) or for $J=1\rightarrow0$ transitions. Table \[tab:lines\] reports the values of the D and F atomic parameters for a list of candidate visible and infrared M1 lines identified by @Judge01 as having the best coronagraphic potential. The values are calculated assuming LS coupling such that the Landé factors for each level are given by $$\begin{aligned} g = \frac{3}{2} + \frac{S(S+1) - L(L+1)}{2J(J+1)},\label{eq:LScoupling}\end{aligned}$$ where S and L are the total spin and orbital angular momenta. We find that only three of the candidate lines have $F\ne0$— 5303 Å, 14301 Å, and 30285 Å. The 6375 Å line can only be circularly polarized because the upper level has $J=1/2$ and it therefore cannot be atomically aligned. Transitions with an upper level that is not polarizable have $D=0$, $F=0$ and no alignment correction is needed to measure an unbiased value for the line of sight magnetic field (see discussion in Section \[blos\]). Based on these results, the 10747 Å and 10798 Å lines considered by @Plowman14 each have $F=0$ and thus contain degenerate information. Consequently, this line pair does not allow one to invert for the magnetic field vector. The discrepancy between his results and ours arises due to the use of Equation 38b from @Casini99 which contains a small algebraic error leading to a spurious value for the coefficient $E$ for the 10798 Å line.[^1] That said, the information in these line pairs is only degenerate if LS coupling holds for the energy levels. This may not be case for the 10747, 10798 Å pair because the separation in the energy levels for the two transitions does not obey Landé’s interval rule, which states that the splitting between two fine structure levels of the same term is proportional to the larger J value. The ratio of energies between the $^3P_2 \rightarrow\ ^3P_1$ and $^3P_1 \rightarrow\ ^3P_0$ transitions in the ground term of is 9258.6/9303.1 $\sim1$ instead of 2 as would be expected from Landé’s interval rule. It may be that the line pair does contain non-degenerate information but confirmation of the Landé factors for the levels is needed before the line can be used as a diagnostic for the magnetic field. Recent laboratory experiments by @Hensel16 and @Bekker18, using electron beam ion traps, succeeded in measuring the Landé g-factors for visible transitions from highly ionized iron, specifically 5303 Å, 6475 Å and 7892 Å. The g-factors measured for 7892 Å indicate significant departures from LS coupling; however, the g-factors for the two levels of the transition are equal within the experimental uncertainties which still means $F=0$ for this line. Extending such laboratory measurements to the infrared transitions of iron, as well as the other species in Table \[tab:lines\], is necessary to ensure the accuracy of coronal magnetic field measurements. [lr|ccccccccccccc]{} & 5303 & $^2P_{3/2} \rightarrow\ ^2P_{1/2}$ & 0.71 & 0.5\ & 6375 & $^2P_{1/2} \rightarrow\ ^2P_{3/2}$ & 0.0 & -\ & 7892 & $^3P_1 \rightarrow\ ^3P_2$ & 0.11 & 0.0\ & 10746 & $^3P_1 \rightarrow\ ^3P_0$ & 1.0 & 0.0\ & 10798 & $^3P_2 \rightarrow\ ^3P_1$ & 0.6 & 0.0\ & 12524 & $^3P_1 \rightarrow\ ^3P_2$ & 0.1 & 0.0\ & 14301 & $^2P_{3/2} \rightarrow\ ^2P_{1/2}$ & 0.71 & 0.5\ & 19201 & $^3P_1 \rightarrow\ ^3P_0$ & 1.0 & 0.0\ & 24826 & $^3P_2 \rightarrow\ ^3P_1$ & 0.1 & 0.0\ & 30285 & $^2P_{3/2} \rightarrow\ ^2P_{1/2}$ & 0.71 & 0.5\ & 39343 & $^3P_1 \rightarrow\ ^3P_0$ & 1.0 & 0.0\ & 55033 & $^3P_2 \rightarrow\ ^3P_1$ & 0.6 & 0.0\ & 90090 & $^3P_1 \rightarrow\ ^3P_0$ & 1.0 & 0.0 The influence of observational noise\ on inversions of Fe XIII and Si X {#sec:noise} ===================================== Coronal measurements are background-limited [@Penn04], and current observatories like COMP [@Tomczyk08] can only routinely achieve noise levels at the $10^{-3}$ noise level, while upcoming facilities like DKIST [@Rimmele15] and COSMO [@Tomczyk16] will more routinely achieve observations at the $10^{-4}$ noise level and higher. Following @Plowman14, here we consider the influence of observational noise on the inverted quantities using a line combination that meets the criteria set forth above. We propose the 10747 Å ($F=0$) and 14301 Å ($F=0.5$) line combination as a promising alternative to the line pair. Of the lines with $F\ne0$ identified, the 14301 Å has perhaps the best potential for full Stokes measurements on account of its brightness and heightened Zeeman sensitivity due to its longer wavelength [@Judge01]. Meanwhile 10747 Å has already been demonstrated to be a line with good potential for coronal magnetometry [@Lin00; @Lin04]. Both lines are photoexcited from the lower level of the ground state and have large values for D, which in turn leads to larger atomic alignments and larger linear polarization. While the temperatures of peak ionization fraction for and are different (${\sim}$1.7 MK and 1.4 MK, respectively), the contribution functions do overlap significantly such that structures like coronal loops between 1.5 and 1.6 MK will have measurable emission in both lines. Furthermore, both lines have now been measured simultaneously during eclipses and using ground-based coronagraphic linear polarimetry [@Dima18; @Dima19a]. One drawback to using line pairs from different ionized species is the increased likelihood that emission in each line may not originate in the same coronal region. However, as discussed in Section \[sec:implementation\] analyzing the polarization angle for each line offers an important check on whether the emission is co-spatial. Here we investigate the effect of noise on inversions of this line combination by performing a Monte Carlo experiment based on inverting synthetic observations of the two lines. Using a modified version of the polarized coronal line emission code [CLE, @Judge01b], which numerically calculates the atomic alignment and level populations for statistical equilibrium of a multi-level atom including photoexcitation and electron collisions, we generated dispersed full Stokes line profiles for the 10747 Å and 14301 Å lines sampled at 0.035 Å. Example line profiles of the 10747 Å line are shown in Figure \[fig:line\_fit\_noise\] for normally distributed noise in each pixel with standard deviations of $10^{-3}I_{peak}$ and $10^{-4}I_{peak}$. The profiles are comparable and not shown. The spectra shown are for emission located in the plane of the sky passing through the center of the Sun and at a height of $0.1$ [$R_{\Sun}$]{} above the photosphere. The reference direction for linear polarization is perpendicular to the plane containing the line-of-sight and the radial direction. The input magnetic field has local radial coordinates (B, $\theta_B$, $\phi_B$) = (25 G, 90[$^{\circ}$]{}, 40[$^{\circ}$]{}) and line of sight coordinates (B, $\Theta_B$, $\gamma_B$) = (25 G, 40[$^{\circ}$]{}, 0[$^{\circ}$]{}). The electron density is set to $n_e = 10^{8.5}$ cm$^{-3}$ and the temperature is $\sim1.55$ MK. As we are using relative noise values, we do not need to set a value for the total emission measure (or length of plasma along line of sight); however, the general scenario here is that of an over-dense closed coronal loop observed near its apex above the solar limb [see, *e.g.* @Aschwanden13]. ![Synthetic noisy line profiles (black) and fits (red) for the 10747 Å line. The spectra sampling is 0.035 Å and normally distributed random photon noise is added to each pixel with the standard deviation indicated above each column. The emission region is located at a height of 0.1 [$R_{\Sun}$]{} and the scattering angle is $\theta_B = 90^{\circ}$. The input magnetic field has line of sight coordinates (B, $\Theta_B$, $\gamma_B$) = (25 G, 40[$^{\circ}$]{}, 0[$^{\circ}$]{}).[]{data-label="fig:line_fit_noise"}](line_fit_noise_90.pdf){width="3.3in"} Like [@Plowman14] we test the single point inversion for different inclination angles relative to the line of sight ($\Theta_B$) for constant values of the atomic alignments ($\sigma^2_0$). To achieve this we fix the strength of the field B and $\theta_{B}$ while varying the azimuth $\phi_B$ in the plane perpendicular to the local radial direction (see Figure \[fig:coordinate\_system\]), which in turn varies $\Theta_B$. Physically this is akin to changing the orientation of a coronal loop from an edge-on ($\Theta_{B} = 0^{\circ}$) to a face-on ($\Theta_{B} = 90^{\circ}$) observational geometry. This approach, in addition to fixing $B \equiv 25$ G, is comparable to that of [@Plowman14]. Note, however, that the errors derived will depend upon the magnetic field intensity and the atomic alignments; here only one representative example is considered. For values of $\Theta^{in}_B$ ranging from $0^{\circ}$ to $90^{\circ}$, where the added superscript indicates “input” values, we produce a sample of 500 profiles with different realization of random noise. For each profile, we determine the Stokes amplitude values of I, L, and V in Equations \[eqI\]-\[eqV\] by fitting in a least-squares sense for the normalized profile $\phi(\lambda-\lambda_{0})$ and its derivative $\phi'(\lambda-\lambda_{0})$. The inverted parameters at each noise instance are then obtained from Equations \[eqsin\]-\[eqgamma\] and the solutions for each parameter are binned into histograms at each input $\Theta^{in}_B$. $I_{01}$,$\sigma^2_{01}$ and $I_{02}$,$\sigma^2_{02}$ refer to the and lines, respectively. The input values for alignments are $\sigma^2_{01}=-0.021$ and $\sigma^2_{02}=-0.012$. The normalized histograms for the negative and positive solution branches of Equations \[eqsin\]-\[eqgamma\] are shown in Figures \[fig:negative\_sol\] and \[fig:positive\_sol\]. Note that $\gamma_B$ and $\gamma_B + \pi$ are also solutions in either case. In this particular example the input atomic alignments are negative and correct solutions for the parameters are recovered by the negative solution branch (Figure \[fig:negative\_sol\]). The positive solution branch shown in Figure \[fig:positive\_sol\] recovers formally consistent solutions (i.e. solutions with $\mid\cos\Theta_B\mid \leq 1$) for $\Theta^{in}_B > 17^{\circ}$ while for $\Theta^{in}_B<17^{\circ}$ the inversions produce inconsistent solutions with $\cos \Theta_B > 1$. In general the range of $\Theta_B$ for which the number of degenerate solutions decreases is a function of both the lines analyzed and the atomic alignments as discussed in Section \[ambiguities\]. {width="6.5in"} {width="6.5in"} We find similar noise effects on the inverted parameters using the , line combination to results discussed by @Plowman14. Noise levels of $\sim10^{-4}I_{peak}$ ($\sim10^{-5}I_{peak}$) are needed to recover input magnetic field inclinations $\Theta^{in}_B < 36^{\circ}$ ( $\Theta^{in}_B < 66^{\circ}$). Parameters like the atomic alignments and the line intensities are recovered well for noise levels at the $10^{-4}I_{peak}$ levels. The line of sight magnetic field, $B_{LOS}$, and projected magnetic field angle, $\gamma_B$, are particular robust to noise. In general the full magnetic field vector can be recovered for field inclinations within some limiting angle to the line of sight, $\Theta_B < \Theta^{lim}_B$. The value $\Theta^{lim}_B$ depends on the quantum mechanical properties of the lines used in the inversion, the atomic alignments and the observation noise levels. In the example considered here with a moderately strong field $B=25$ G and small values for the atomic alignments, $\Theta^{lim}_B \approx 36^{\circ}$ for $\sim10^{-4}I_{peak}$ noise levels. Larger noise levels, smaller values for B and smaller atomic alignments all lead to a decrease in $\Theta^{lim}_B$ and a shrinking of the range of invertible solutions. In agreement with @Plowman14, we note that the inverted $\Theta_B$ value is biased towards $0^{\circ}$ for larger relative noise and for more transverse input fields directions. @Plowman14 alleviated this bias by applying a Bayesian MCMC method and appropriate priors to an observation and mapping the posterior probability distribution at each input value of $\Theta_B$. Such a method could be applied to any line pair assuming, as demonstrated, at least one of the lines has $F\ne0$. Discussion ========== Ambiguous Solutions {#ambiguities} ------------------- In the most general case equations \[eqsin\]-\[eqgamma\] return four consistent solutions ($\mid\cos\Theta_B\mid \leq 1$) for the inverted set of parameters. Two of the solutions have projected directions $\gamma_B$ at 90[$^{\circ}$]{} to each other and map to the classical Van Vleck ambiguity. However, it is notable from looking at Figures \[fig:negative\_sol\] and \[fig:positive\_sol\] that the ambiguity extends to all the inverted parameters. Thus solving this ambiguity may be accomplished by imposing physical limits on any of the solution parameters. The other two ambiguous solutions have parameters identical to the first two except the projected angles $\gamma_B$ are different by 180[$^{\circ}$]{} (see example in Figure \[fig:all\_solutions\]). The number of solutions depends on both the intrinsic quantum mechanical properties of the transitions through the $D$, $F$ and $\bar{g}$ parameters and on the atomic alignments $\sigma^2_{0i}$ which are functions of the extrinsic thermodynamic properties of the plasma and radiation anisotropy. An example of this is discussed in Section \[sec:noise\], where the number of solutions decreases to only two (180[$^{\circ}$]{} ambiguity) when $\Theta_B\lesssim 17^{\circ}$. When the emission lines used in the inversion have equal $D$ and $F$ values (e.g. 5303 Å and 14301 Å or 30285 Å) the number of degenerate solutions, in the noise free case, reduces to two for all values of $\Theta_B$ and atomic alignments. While the multi-line method recovers (at most) four possible degenerate solutions for the vector magnetic field and atomic alignment, it does not provide direct knowledge of the scattering angle $\theta_{obs}$, the location of the emission along the line of sight, nor the magnetic field direction in a coordinate frame local to the Sun. The field direction in the solar frame can be calculated from the inferred (degenerate) LOS-frame field vector(s) for a given scattering angle using spherical trigonometry (see equations 42,44 in @Casini99). With knowledge of the projected height above the solar surface and the inferred values of atomic alignment, numerical calculations (as discussed above) can determine a range of temperature and densities required to be consistent with the observations at a given scattering angle. Consequently, independent observations of the temperature and density can constrain the scattering geometry itself and the field in the local frame as discussed in the next section. See also further discussion by @Judge07 and @Schad18. Even without knowledge of the plasma density and temperature, however, some of the degenerate solutions may be ruled out for particular scattering angles based on the relationship between the sign of $\sigma^2_0$ and $\theta_B$, given by the ($3 \cos^2{\theta_B} - 1$) term discussed above [@Judge07]. At $\theta_{B} \approx 54.74^{\circ}$ (and ${\approx}125.26^{\circ}$), referred to as the Van Vleck angle, the atomic alignment changes sign. At each scattering angle along the line of sight, the solution pairs ($\sigma^2_0$, $\theta_B$) may be valid or determined to be invalid based on the sign relationship, or as discussed in @Schad18, ruled out if the scattering angle is unphysically large given the observations. ![Plot of four degenerate solutions shown in the LOS frame of reference. Scattering angle is $\theta_{obs} = 90^{\circ}$ so the emission is located in the plane perpendicular to the LOS and bisecting the Sun. The input field is (B, $\Theta_B$, $\gamma_B$) = (25 G, 35[$^{\circ}$]{}, -59[$^{\circ}$]{}) and corresponds one of the arrows shown in blue. The other blue arrow corresponds to the 180[$^{\circ}$]{}ambiguous solution due to the lack of orientation in the reference direction for the linear polarization (r.d.l.p). The red pair of arrows represents the degenerate solutions due to the Van Vleck ambiguity. The dashed lines represent the projections of the solutions on the plane perpendicular to the line of sight.[]{data-label="fig:all_solutions"}](all_solutions_3.pdf){width="3.3in"} Combining multi-line M1 inversions with other plasma diagnostics ---------------------------------------------------------------- Combining multi-line M1 inversions with spectroscopic plasma diagnostics for density and temperature provides an opportunity to further constrain the ambiguity of the M1 inversion, as discussed by [@Judge13], as well as a means to cross-validate numerical calculations of the atomic alignment with model independent empirical determinations. In addition, it is anticipated a simultaneous model fit of two or more M1 lines (as long as at least one of the lines has $F\ne0$) will further improve resilience of the solution to observational noise. Once again, the atomic alignment depends on electron density ($n_e$) and temperature ($T_e$), height above the photosphere and the magnetic field orientation relative to the local radial direction $\theta_B$ (see Figure \[fig:sigma\_density\]). Emission line ratios of both ultraviolet and infrared transitions may be used as plasma diagnostics to independently measure $n_e$ and $T_e$ [e.g. see reviews by @DelZanna18a; @delZanna18b]. This information can be used, in addition to knowledge of the projected height above the limb, to determine the scattering angle consistent with the multi-line M1 inverted parameters. Separately, the atomic alignment can be numerically calculated given the inferred plasma parameters, geometry, and field direction and therefore compared with the inferred value of $\sigma^2_0$ as a way to benchmark the theoretical calculations. However, we note that very high signal-to-noise is required to precisely determine the alignment (see Figure \[fig:negative\_sol\]). ![Variation of the atomic alignment for the 10747 Å line with density, temperature and magnetic field inclination $\theta_B$ with respect to the incident radiation axis of symmetry. The atomic alignment is calculated at a height of 0.1 [$R_{\Sun}$]{}(thin lines) and 0.3 [$R_{\Sun}$]{}(thick lines).[]{data-label="fig:sigma_density"}](sigma_density2.pdf){width="3.3in"} A single line measurement of B LOS at a single point {#blos} ---------------------------------------------------- While so far we have focused on multi-line diagnostics, the formalism we presented in Section \[analytic\_solution\] also helps identify diagnostic techniques available for single M1 line spectropolarimetric observations. In particular, it has been known since [@Casini99] that the atomic alignment plays a role in the circular polarized amplitude and that the typical weak field magnetograph formula cannot be directly applied to measure the longitudinal component of magnetic field (*i.e.* along the line-of-sight (LOS)) without an associated scale error. Interestingly, [@Plowman14] demonstrated via his Equation 14 that the scale error could by mitigated by using the observed linear polarized amplitude as a correction. In our generalization of this approach, the updated form of this relation is given via Equation \[eqB\] as $$\begin{aligned} B^{obs}_{LOS} &= \frac{h}{\mu_B}\frac{-V}{\left[(I \pm L)\bar{g} \pm \frac{2}{3}F\frac{L}{\sin^{2}{\Theta_{B}}}\right]}. \label{eqBLOS}\end{aligned}$$ where $B^{obs}_{LOS}$ is the observationally determined longitudinal magnetic field strength. Two solutions for $B^{obs}_{LOS}$ are obtained which correspond to the positive and negative branches of equation \[eqBLOS\]. For lines with $F=0$ (left panels in Figure \[fig:blos\_measurement\]), the solution branch that matches the sign of the atomic alignment accurately measures the input magnetic field ($B^{in}_{LOS}$) independent of $\Theta^{in}_B$ and $\sigma^{2in}_0$. The magnetic field measured for the degenerate solution over-predicts the input magnetic field by an amount that increases with both $\Theta^{in}_B$ and $\sigma^{2in}_0$ so fields more tangential to the LOS have a larger bias than fields parallel to the LOS. An example of this effect can be observed for the example discussed in Section \[sec:noise\]. Note how for the negative solution branch (Figure \[fig:negative\_sol\]) the inverted value of $B_{LOS}$ correlates much more tightly with the input value than the positive solution branch (Figure \[fig:positive\_sol\]). As the F value for 10747 Å is zero, V and L measurements provide tight constraints on $B_{LOS}$ via Equation \[eqBLOS\] that are independent of $\Theta_{B}$ and track the true value of $B_{LOS}$ when the chosen solution matches the sign of the input alignment, in this case the negative branch. For this example the bias is small since the atomic alignment ($\sigma^2_{01}=-0.021$) is small compared to the alignment values used to produce Figure \[fig:blos\_measurement\]. Meanwhile, for transitions with $F\ne0$ (right panels in Figure \[fig:blos\_measurement\]), both solutions for $B^{obs}_{LOS}$ are functions of $\Theta^{in}_B$ and $\sigma^{2in}_0$. Therefore, in principle, without a priori knowledge of these parameters both the measured values for $B^{obs}_{LOS}$ will have a bias relative to the true value $B^{in}_{LOS}$. It is possible to show that the bias is minimized at all values of $\Theta^{in}_B$ if a guess value of $\Theta^g_B=90^{\circ}$ is used in equation \[eqBLOS\]. Other guess values for $\Theta^{g}_B$ will improve agreement locally but induce larger errors at all other angles (e.g. see dashed lines with $\Theta^g_B=45^{\circ}$ in the right panels of Figure \[fig:blos\_measurement\]). As discussed by @Casini99 the weak field approximation or magnetograph formula (MF) relating $B_{LOS}$ to the Stokes measurements: $$\begin{aligned} B^{obs}_{LOS,MF} &= \frac{h}{\mu_B\bar{g}}\left(\frac{-V}{I}\right) \label{eqMF}\end{aligned}$$ biases the measurements by an amount that depends on the atomic alignment since it effectively assumes $L=0$. Plots for the value of $B^{obs}_{LOS}$ using equation \[eqMF\] are shown in blue in Figure \[fig:blos\_measurement\] and the values fall between the two solutions. ![Angle dependent ratios of the measured (or “observed”) line of sight magnetic field $B^{obs}_{LOS}$ to the true input magnetic field $B^{in}_{LOS}$. The panels on the left correspond to measurements for a line with $F=0$ while the panels on the right correspond to a line with $F=0.5$, with the top (bottom) panels having positive (negative) input atomic alignments $\sigma^{2in}_0$. The black (green) curves represent $B^{obs}_{LOS}$ values calculated with the positive (negative) sign in equation \[eqBLOS\] and having set $\Theta^g_B=90^{\circ}$ which minimizes the bias at all values for $\Theta_B$. For comparison the dashed lines assume $\Theta^g_B=45^{\circ}$. The blue curves represent $B^{obs}_{LOS}$ calculated using the magnetograph formula (MF, equation \[eqMF\]).[]{data-label="fig:blos_measurement"}](blos_measurement_6.pdf){width="3.3in"} Approaches for implementation and limits to the application of the single point approximation {#sec:implementation} --------------------------------------------------------------------------------------------- A reliable implementation of the single point inversion, using for example observations to be made available by the upcoming DKIST instruments, relies not only on using appropriate line combinations and sufficient single-to-noise but also on techniques to target and isolate regions of emission along the line of sight, which may be aided by observational checks on the validity of the single point approximation to a given observation. Coronal loop structures provide one viable target for applying the single point approximation as has been done by many authors who study the plasma properties along loops using spectroscopic and spectro-imaging techniques [e.g. @Aschwanden99; @Brooks12; @Xie17]. In these studies the importance of background/foreground emission subtraction is well known [see @Terzo10 and references within]. The intensity contrast of loops observed by TRACE (using the most sensitive filter for each loop) was generally found to be $\sim15-20\%$ [@Aschwanden05]. Warm ($\sim$1-2 MK) active region loops tend to have higher densities compared to diffuse emission surrounding the loops [@Brooks12; @Brooks19]. As seen in Figure \[fig:sigma\_density\], the atomic alignment decreases sharply with increases in the density (the temperature dependence is much smaller). Therefore, although loops have increased density and correspondingly larger brightness, the linearly polarized signals may be lower than for the background plasma. As such, the single-to-noise requirements become more demanding in order to implement a reliable background subtraction technique, such as fitting the cross section of the polarized emission profile across the loop. To understand how best to achieve this subtraction, synthesizing polarized emission through high resolution magnetohydrodynamic (MHD) simulations of the active corona [e.g. @Rempel17] and then testing methods may be helpful. Given a particular observation, one test for the validity of the single point approximation is provided by the two independent measurements of projected linear polarization angle. Since M1 lines are in the saturated Hanle regime, the angles $\gamma_B$ measured with different lines emitted from the same region should be equal. The condition is necessary but not sufficient because the possibility exists that background linear polarization may satisfy this condition while obscuring the signal. Resulting signals that have background subtraction applied should also satisfy this condition. Summary and outlook =================== In this article, we have presented a generalization of the dual-line inversion algorithm for the vector magnetic field, proposed by @Plowman14 for the 10747, 10798 Å line pair, to other combinations of magnetic dipole (M1) lines. A number of results that emerged from the analysis: 1. Assuming emission is concentrated at a single point in the corona it is possible to combine polarized observations from two or more M1 lines to invert for the vector magnetic field. However, only some combinations of M1 lines contain non-degenerate information such that a finite set of degenerate solutions can be calculated. Transitions between levels with equal upper/lower level Landé factors or $J=1 \rightarrow 0$ transitions cannot be used without adding another line where these conditions do not hold. 2. One pair of lines that satisfies this requirement is 10747 Å and 14301 Å. Using atomic modelling of the polarized emission coefficients under realistic physical conditions for a coronal loop, we showed that the inclination of the magnetic field with respect to the line of sight, $\Theta_B$, and the magnetic field strength B can be recovered for photon noise levels $\sim10^{-4}I_{peak}$ if the magnetic field is oriented close to the line of sight. 3. In general the method produces four degenerate magnetic field solutions in the line of sight frame of reference consistent with the Van Vleck and 180[$^{\circ}$]{} ambiguities. However, the number of allowed solutions may be smaller depending on the field geometry and the emission lines selected. Independent measurements of the height of emission (through stereoscopy) can be used together with an atomic model and the measured atomic alignment to determine the density and temperature of the emission region. Conversely, measurements of density and temperature (through coordinated observations with diagnostic UV and IR lines) can be used to constrain the height of emission. 4. The line of sight magnetic field $B_{LOS}$, obtained from polarized measurements of a single M1 line, is robust to noise but can be measured without bias only for M1 lines that have equal Landé factors for the upper and lower levels, or $J=1 \rightarrow 0$ transitions. Other M1 lines that do not satisfy these conditions will give a biased measurement value for $B_{LOS}$ that varies based on the inclination of the magnetic field with respect to the line of sight $\Theta_B$. However, this bias can be minimized for all values of $\Theta_B$ to better than 10% even for large values of the atomic alignment if the inclination is assumed to be $\Theta_B=90^{\circ}$. The coronagraphic polarimetric capabilities of the upcoming National Science Foundation’s *Daniel K. Inouye Solar Telescope* provides an unprecedented opportunity to observe M1 emission from the solar corona at high resolution and high sensitivity such that multi-line techniques for coronal magnetic field diagnostics, including the techniques developed here, can be put into practice. Acknowledgments =============== We extend our thanks to Roberto Casini, Phil Judge, and Joe Plowman for providing feedback and comments on this work. The National Solar Observatory (NSO) is operated by the Association of Universities for Research in Astronomy, Inc. (AURA), under cooperative agreement with the National Science Foundation. [^1]: We have confirmed this error with the authors of [@Casini99] and [@Plowman14].

ArXiv

--- author: - | Johanna Erdmenger\ Max-Planck-Institut für Physik (Werner-Heisenberg-Institut)\ Föhringer Ring 6, 80805 München, Germany - | Veselin Filev\ Max-Planck-Institut für Physik (Werner-Heisenberg-Institut)\ Föhringer Ring 6, 80805 München, Germany\ &\ School of Theoretical Physics\ Dublin Institute for Advanced Studies\ 10 Burlington Road, Dublin 4, Ireland - | Dimitrios Zoakos\ Centro de Física do Porto & Departamento de Física e Astronomia\ Faculdade de Ciências da Universidade do Porto\ Rua do Campo Alegre 687, 4169–007 Porto, Portugal title: Magnetic catalysis with massive dynamical flavours --- Introduction ============ In recent years, research on the AdS/CFT correspondence [@Maldacena:1997re] and its applications has expanded significantly, leading to substantial new results. At present, holographic descriptions of non-perturbative phenomena range from applications to condensed matter systems (e.g superconductivity, superfluidity, quantum Hall effect) to applications in high energy physics and to the quark-gluon plasma (e.g. confinement/deconfinement phase transitions, chiral symmetry breaking, elliptic flow, hadronization). Despite the remarkable insights gained by studying holographic gauge theories, the application of the correspondence to phenomenologically relevant gauge theories continues to remain a challenge. A way to further improve the phenomenological relevance of the AdS/CFT correspondence is to explore further phenomena of universal nature. The results of such studies are expected to capture aspects of the qualitative behaviour of realistic gauge theories (such as QCD), which do not possess holographically dual supergravity backgrounds. #### An important example in this class of phenomena is the effect of mass generation and spontaneous chiral symmetry breaking in the presence of an external magnetic field. This effect has been shown to be model-independent and therefore insensitive to the microscopic physics underlying the low energy effective theory. Its essence is the dimensional reduction $D \to D-2$, (3+1 $\to$ 1+1) in the dynamics of fermion pairing in a magnetic field. The enhanced infrared divergences in lower dimensions suggests that the dynamics of fermion pairing is governed by the lowest Landau level, which hints at the universal nature of the phenomenon. Furthermore, one can show that the $(D-2)$-dimensional dynamics of the Landau level favours the condensation of the fermion pairs. This effect is known as Magnetic Catalysis. Magnetic Catalysis has been demonstrated in various (1+2)- and (1+3)-dimensional field theories [@Gusynin:1994re]-[@Klimenko:1992ch] using conventional field theoretical methods. The holographic approach to this effect has been initiated in [@Filev:2007gb], where the (1+3)-dimensional holographic gauge theory dual to the D3/D7–brane intersection has been considered[^1]. Additional holographic studies of magnetic catalysis at finite temperature or chemical potential for both (1+3)- and (1+2)-dimensional systems have been performed in [@Filev:2007qu]-[@Bolognesi:2011un]. #### In the holographic description of magnetic catalysis, the flavour degrees of freedom are introduced by an additional stack of flavour D–branes. The most understood and widely applied regime of this approach is when the flavour branes are in the probe limit and their backreaction to the ambient supersymmetric background is neglected [@Karch:2002sh]. On the field theory side, this corresponds to the “quenched" approximation in which the number of flavour degrees of freedom, $N_f$, is much smaller than the number of colour degrees of freedom, $N_c$. In terms of Feynman diagrams this implies ignoring the contribution from internal quark loops (windows) in the planar diagrams of the corresponding large $N_c$ expansion. At present all holographic studies of magnetic catalysis are in the quenched approximation. An obvious question is to ask how corrections due to internal quark loops would influence the effect of magnetic catalysis. Our goal is to provide such an estimate in a perturbative expansion in the ratio between the number of flavour and colour degrees of freedom, $N_f/N_c$. #### For obtaining an unquenched holographic description of magnetic catalysis, we have to take into account the backreaction of the flavour branes to the supergravity background sourced by the colour branes. Ideally such a background would describe localized branes, however for technical reasons this is a difficult task even in the supersymmetric case. To circumvent these difficulties, the flavour D–branes may be distributed along the compact directions of the supergravity background. This procedure is called smearing[^2]. The smearing restores a significant part of the global symmetry of the geometry and hence simplifies the corresponding Einstein equations. #### Supersymmetric backgrounds of smeared massless flavour D–branes have been constructed in [@hep-th/0612118], and for flavours with finite bare mass in [@arXiv:0807.0298]. In the case of massive flavours, these backgrounds display a hollow cavity in the bulk of the geometry, where the supergravity solution is sourced solely by the colour branes. The radius of this cavity is related to the bare mass of the fundamental flavors. In the limit of vanishing bare mass the cavity shrinks to the radius of the compact part of the geometry and the supergravity background has an essential singularity at the origin of the non-compact part of the geometry. In both cases the dilaton field diverges at large radial distances. This corresponds to the Landau pole that the dual field theory develops in the UV, due to its positive beta function $\beta\propto N_f/N_c$. #### We can imagine that a non-supersymmetric background interpolating between two supersymmetric backgrounds, corresponding to massless flavors in the UV and massive flavors in the IR, would describe dynamical mass generation. The radius of the hollow cavity then corresponds to the dynamically generated constituent mass of the fundamental flavors. A promising framework for the construction of such a geometry was developed in [@arXiv:0909.2865], where the ten-dimensional black-hole solution dual to the non-conformal plasma of flavoured ${\cal N} = 4$ supersymmetric Yang-Mills theory is presented[^3]. The authors outline the smearing procedure, derive the corresponding equations of motion and present a perturbative solution for general massless non-supersymmetric flavour D7–brane embeddings. #### The first steps towards unquenching the holographic description of magnetic catalysis have been undertaken in [@arXiv:1106.1330], where the authors, following the approach of [@arXiv:0909.2865], construct a perturbative non-supersymmetric background with a non-vanishing $B$-field, which corresponds to an external magnetic field coupled to the fundamental degrees of freedom of the dual gauge theory. In the case of massless fundamental fields and sufficiently strong magnetic field, the supergravity background is unstable, suggesting that the theory undergoes a phase transition to a stable phase with dynamically generated mass for the matter fields. #### In section 2 of this paper we complete the studies initiated in [@arXiv:1106.1330] constructing a perturbative non-supersymmetric background with a non-trivial $B$-field, for massive flavour fields. Our solution has a hollow cavity in the bulk of the geometry where it is very similar to the supergravity dual of a non-commutative supersymmetric Yang-Mills theory, [@Maldacena:1999mh] & [@Hashimoto:1999ut]. The difference is in the presence of a squashed $S^5$, instead of a non-squashed one, which breaks the supersymmetry. As suggested above, the radius of this cavity, $r_q$, is related to the dynamically generated mass of the fundamental fields. #### For radial distances greater than $r_q$ the solution is characterized by a non-vanishing density for the smeared D7–brane charge. At sufficiently large radial distance our solution approaches the supesymmetric one, constructed in [@arXiv:0807.0298]. Following the prescription of [@arXiv:0909.2865], we introduce an additional large radial parameter $r_*\gg r_q$ (corresponding to a finite UV cutoff ), at which we match our solution to the supersymmetric one. Furthermore, we identify the value of the $B$-field at $r_*$ as the magnetic field of the dual gauge theory, $H_*\equiv B(r_*)$. #### For radial distances greater than $r_*$ the supergravity background is well approximated by the non-perturbative supersymmetric background [@arXiv:0807.0298]. This enables us to relate non-perturbatively the UV parameters of the theory, namely the finite cutoff $\Lambda_{UV}\propto r_*$ and the energy scale corresponding to the landau pole of the theory $\Lambda_{\rm LP}\propto r_{\rm LP}$, where $r_{\rm LP}$ is the radial distance at which the dilaton field diverges. #### Our supergravity construction has the following renormalization group flow interpretation: #### At the energy scale set by the finite cutoff ($\Lambda_{UV}\propto r_*$) the dual gauge theory is a commutative ${\cal N}=4$ Supersymmetric Yang-Mills theory coupled to $N_f$ flavours of ${\cal N}=2$ hypermultiplet fundamental fields. The fundamental hypermultiplets are coupled to a constant external magnetic field $H_*$, which breaks the supersymmetry. Decreasing the energy scale the Yang-Mills theory becomes non-commutative and the parameter of non-commutativity (roughly the non-trivial part of the $B$-fleld) is proportional to the ratio $N_f/N_c$. #### At energy scales of the order of the physical mass of the fundamental fields, $M_q$ (roughly $M_q\sim r_q$) the flavour fields decouple (the D7–brane charge density vanishes). At lower energy scales (inside the cavity, $r<r_q$) the dual gauge theory is a pure (only adjoint degrees of freedom) non-commutative Yang-Mills theory. To leading order the parameter of non-commutativity in the plane perpendicular to the magnetic field scales as $\Theta^{23}\propto \frac{N_f}{N_c}\frac{1}{H_*}$. #### Note that in the bare Lagrangian of the dual gauge theory the external magnetic field $H_*$ couples explicitly only to the fundamental degrees of freedom. Therefore the non-commutativity of the adjoint degrees of freedom cannot be captured by the quenched approximation and is one of the novel results of our analysis. #### Finally, in section 3 of this work we apply our construction to study the effect of magnetic catalysis. We develop an appropriate renormalization scheme and compute the free energy and the fundamental condensate of the holographically dual gauge theory as a function of the bare mass of the fundamental degrees of freedom. Our studies show that to leading order in a perturbative expansion in the ratio $N_f/N_c$, the free energy and fundamental condensate of the theory agree with the results obtained in the quenched approximation. Furthermore, at next order in $N_f/N_c$ we show that the effect of magnetic catalysis is enhanced and the contribution to the condensate of the theory from internal fundamental loops runs logarithmically with the finite cutoff $\Lambda_{UV}$. Constructing the Background =========================== In the present section we will construct the supergravity background necessary for the holographic study of the phenomenon of magnetic catalysis. The field theory duals are realized on the intersection between a set of $N_c$ [*colour*]{} D3-branes and a set of $N_f$, homogeneously smeared, [*flavour*]{} D7–branes, with an additional coupling between the fundamental fields and an external magnetic field. The [*colour*]{} D3-branes are placed at the tip of a Calabi-Yau (CY) cone over a Sasaki-Einstein manifold $X_5$, where the latter can be expressed as a $U(1)$ fiber bundle over a four-dimensional Kähler-Einstein base (KE). The [*flavour*]{} D7–branes extend along the radial direction, wrap a submanifold $X_3$ of $X_5$ and smear homogeneously over the transverse space [@hep-th/0602027; @hep-th/0505140]. Ansatz & smearing of the flavours --------------------------------- To take into account the contribution from internal fundamental loops in the Veneziano limit of the dual gauge theory, we need to consider the backreaction of the flavour branes. Ideally the corresponding supergravity solution would describe localized D7–branes which break the global symmetry of the internal subspace of the supergravity background from $SO(6)$ down to $SO(4)\times SO(2)$ (when $X_5$ is $S^5$). However, even in the supersymmetric case (no external magnetic field), this is a very difficult task. One way to circumnavigate the technical difficulties is to construct a solution with smeared D7–branes. In general the smearing procedure involves distributing the branes at different locations in the transverse space subspace and consider a course grained approximation in which the sum over all individual embeddings becomes an integral over a distribution of branes. It is somewhat analogous to the smearing of point-like charges in electrostatics in 1+3 dimensions to obtain: one, two or three dimensional charge densities. In our case the smearing is performed in such a way that the isometries of the fibered Kähler-Einstein space are kept unbroken, allowing for an ansatz where all the unknown functions just depend on a single radial coordinate. Based on this assumption we adopt the following ansatz for the metric $$ds_{10}^2 = h^{-\frac{1}{2}}\left[-dt^2 + dx_1^2+b(dx_2^2+dx_3^2)\right] + h^\frac{1}{2} \left[ b^2 S^8F^2 d\sigma^2 + S^2 ds_{CP^2}^2 + F^2 (d\tau + A_{CP^2})^2 \right] \, , \label{10dmetric}$$ where the $CP^2$ metric is given by $$\begin{aligned} ds_{CP^2}^2&=&\frac{1}{4} d\chi^2+ \frac{1}{4} \cos^2 \frac{\chi}{2} (d\theta^2 + \sin^2 \theta d\varphi^2) + \frac{1}{4} \cos^2 \frac{\chi}{2} \sin^2 \frac{\chi}{2}(d\psi + \cos \theta d\varphi)^2 \quad \& \nonumber \\ A_{CP^2}&=& \frac12\cos^2 \frac{\chi}{2}(d\psi + \cos \theta d\varphi)\,\,. \label{cp2metric}\end{aligned}$$ The range of the angles is $0\leq (\chi, \theta) \leq \pi$, $0\leq \varphi, \tau < 2\pi$, $0\leq \psi< 4 \pi$. The ansatz for the NSNS and the RR field strengths is given by $$\begin{aligned} \label{NS+RR} & B_{2} = H dx^2\wedge dx^3 \, , \quad C_{2} = J \,dt \wedge dx^1 \,, & \nonumber \\ \nonumber \\ & F_{5} = Q_c\,(1\,+\,*)\varepsilon(S^5)\, , \quad F_{1} = Q_f \, p(\sigma) \,(d\tau + A_{CP^2})\, , \quad F_{3} = d C_2\,+\, B_{2} \wedge F_1 \, , &\end{aligned}$$ where $\varepsilon(S_5)$ is the volume element of the internal space[^4] and $Q_{c}, Q_{f}$ are related to the number of different colours and flavours in the following way $$N_c = \frac{Q_c\, Vol(X_5)}{(2\pi)^4g_s \,\alpha'^2} \quad \& \quad N_f = \frac{4\,Q_f\,Vol(X_5)}{Vol(X_3) g_s} \, .\label{QcQf}$$ In our case $X_5=S^5$ and the volume of the three sphere is $2\pi^2$. All the functions that appear in the ansatz, $h,b,S,F, \Phi, J \, \& \, H$, depend on the radial variable $\sigma$ only. In the convention we follow, $S \, \& \, F$ have dimensions of length, $p,b,h,J \, \& \, H$ are dimensionless and $\sigma$ has a dimension of length${}^{-4}$. Furthermore, $\sigma=\infty$ at the origin and decreases to $\sigma_*$ at the boundary. The function $b$ in the ansatz for the metric reflects the breaking of the $SO(1,3)$ Lorentz symmetry down to $SO(1,1)\times SO(2)$. The function $p(\sigma)$ in $F_{(1)}$ and $F_{(3)}$, determines the distribution of the brane embeddings and has a characteristic asymptotic behavior. In fact $p(\sigma)$ encodes the bare mass, $m_q$ and the fundamental condensate of the dual gauge theory. It vanishes at energy scales smaller than the quarks’ mass while it asymptotes to $1$ in the UV. This leads to the formation of a spherical cavity inside the bulk of geometry. The radius of this cavity sets the energy scale related to the physical mass of the quark, $M_q$[^5] . We can understand better the structure of the distribution function $p(\sigma)$ if we consider a representative D7–brane embedding, the so called “fiducial embedding". The fiducial embedding is an auxiliary D7–brane embedding, which probes the backreacted geometry, its shape determines the distribution function $p(\sigma)$. To leading order in a multipole expansion we can describe the profile of the fiducial embedding by functions, which depend only on the radial (holographic) coordinate $\sigma$. The fiducial embedding relevant for our study wraps an internal three-cycle parameterized by $\theta,\varphi,\psi$, extends along $\sigma$ and sits at a fixed value of $\tau$. To obtain the distribution function $p(\sigma)$ we smear the fiducial embedding by acting with the symmetries of the internal space (see appendix of [@arXiv:0909.2865]). In this way we obtain $$p(\sigma)=\cos^4\frac{\chi_{q}}{2}\, . \label{p-sigma}$$ To clarify the relationship between the distribution function and the profile of the fiducial embedding (\[p-sigma\]), in figure \[fig:fig(-1)\] we have presented plots of a family of D7–brane embeddings obtained by acting on the fiducial one with a discrete subgroup of the symmetry group of the internal space. Figures \[fig:fig(-1)a\] and \[fig:fig(-1)b\] represent probe D7–brane embeddings terminating at the same minimal radial distance $r_q$. The first one is a supersymmetric embedding (vanishing magnetic field), while the second one is a non-supersymmetric one (finite magnetic field) and exhibits mass generation. Figures \[fig:fig(-1)c\] and \[fig:fig(-1)d\] represent the smearing of the corresponding probe D7–brane embeddings. One can clearly see the formation of a spherical cavity in the bulk of the geometry. On can also compare the radial distributions corresponding to supersymmetric and non-supersymmetric fiducial embeddings[^6]. \ In the next subsection following the strategy initiated in [@arXiv:0909.2865] we will describe the whole system in terms of a one-dimensional effective action, from which all the equations of motion can be produced. In principle the same system of equations can be explicitly derived also from a ten-dimensional point of view after inserting the above ansatz in the ten-dimensional equations of motion plus the Bianchi identities. The equations of motion and the corresponding Bianchi identities for the NSNS and RR fields are relatively easy to obtain. However, writing down the ten-dimensional Einstein equations is a difficult task. The crucial step is to obtain an effective ten-dimensional expression for the smeared the DBI action, in the case of massive non–supersymmetric probes. While in the massive supersymmetric case such a construction is possible using calibrated geometry (see e.g. [@arXiv:1002.1088; @arXiv:0811.3646]) in the non-supersymmetric case this is a non-trivial task. To circumvent this difficulty we derive the equations of motion from a one dimensional effective action. The key point is that even if we were able to obtain a simple ten dimensional term for the smeared DBI action we would still adopt the ansatz (\[10dmetric\])-(\[NS+RR\]) for the supergravity fields, which is equivalent to reducing the action to one dimension. Therefore we could first reduce the eight dimensional DBI action of every individual flavour brane to one dimension and then sum them. In this context the statement that the flavour branes are smeared is equivalent to the statement that their reduced one dimensional actions are equivalent. In other words, the smearing of the DBI action is equivalent to reducing the eight dimensional DBI action of a fiducial flavour brane embedding to one dimension and multiplying the result by the number of flavour branes $N_f$. Effective action & the equations of motion ------------------------------------------ The action for the Type IIB supergravity plus the contribution from the $N_f$ D7–branes in the Einstein frame is $$S=S_{IIB} + S_{fl} \, , \label{genact}$$ where the relevant terms of the $S_{IIB}$ action are $$\begin{aligned} \label{TypeIIB action} S_{IIB}&=&\frac{1}{2\kappa_{10}^2}\int d^{10} x \sqrt{-g} \Bigg[ R - {1 \over 2} \partial_M\Phi \partial^M\Phi - {1 \over 2} e^{2\Phi}F_{(1)}^2 - {1 \over 2} \frac{1}{3!} e^{\Phi} F_{(3)}^2 - {1 \over 2} \frac{1}{5!} F_{(5)}^2 \nonumber\\ \nonumber\\ &&\qquad \qquad \qquad \qquad \qquad \quad \qquad \qquad \, - {1 \over 2} \frac{1}{3!} e^{-\Phi} H_{(3)}^2\Bigg] - \frac{1}{2\kappa_{10}^2}\, \int C_4\wedge H_3\wedge F_{3} \, ,\end{aligned}$$ and the action for the flavour D7–branes takes the usual DBI+WZ form $$S_{fl} = -T_7 \sum_{N_f} \Bigg[ \int d^8x\, e^\Phi \sqrt{-\det (\hat{G}+e^{-\Phi/2} {\cal F}}) \, - \, \int \left(\hat{C}_8 + \hat{C}_6 \wedge B_2 \right) \Bigg]\,, \label{actionflav}$$ with ${\cal F} \equiv B + 2 \pi \alpha' F$. In those expressions $B$ denotes a non-constant magnetic field, $F$ the worldvolume gauge field and the hat refers to the pullback of the quantities, along the worldvolume directions of the D7–brane. The gravitational constant and D7–brane tension, in terms of string parameters, are $$\frac{1}{2\kappa_{10}^2} = \frac{T_7}{g_s} = \frac{1}{(2\pi)^7g_s^2 \alpha'^4} \, .$$ The first step in our analysis is deriving the equation of motion for the fiducial embedding, which follows from the action $S_{fl}=S_{DBI}+S_{WZ}$. The DBI action for the D7–brane is given by $$S_{DBI} = -\frac{T_7}{8} \int d^8x\, e^{\Phi} \,b^2 \, \sin\theta \, S^6 \, F^2 \, \cos^3 \frac{\chi_{q}}{2} \, \Xi_1 \, \, \Xi_2 \, \, \Xi_3 \, , \label{DBI}$$ where we have introduced the following auxiliary dimensionless quantities $$\Xi_1 \equiv \sqrt{\cos^2 \frac{\chi_{q}}{2}+ \frac{S^2}{F^2}\sin^2 \frac{\chi_{q}}{2}}\,\,, \quad \Xi_2\equiv \sqrt{1+\frac{(\partial_\sigma\chi_{q})^2}{4b^2 \,S^6\,F^2 }} \quad \& \quad \Xi_3\equiv \sqrt{1+\frac{e^{-\Phi}\,H^2\,h}{b^2}}\ ,$$ with $\chi_{q}$ being a function of $\sigma$ determining the brane embedding. The WZ piece of the action is $$S_{WZ} = - \frac{T_7}{32} \, Q_f \int d^8x \sin\theta \, p(\sigma) \, b^2 \,e^{2\Phi} \, S^8 \, \Xi_3^2 \, \cos^4 \frac{\chi_{q}}{2} \,. \label{WZ}$$ The corresponding equation of motion for $\chi_{q}$ is given by $$\begin{aligned} \label{EOM-chi} 0&=&\frac{1}{2} \partial_\sigma\left[e^\Phi \cos^3 \frac{\chi_{q}}{2}\, \frac{\Xi_1 \Xi_3}{\Xi_2}\,(\partial_\sigma \chi_{q}) \right]+\\ &+&e^\Phi b^2 \, S^6 F^2 \, \Xi_3 \, \cos^2 \frac{\chi_{q}}{2}\sin \frac{\chi_{q}}{2} \Bigg[3 \, \Xi_1 \, \Xi_2 + \cos^2 \frac{\chi_{q}}{2} \left(1-\frac{S^2}{F^2}\right) \frac{\Xi_2}{\Xi_1} +Q_f \, e^\Phi \frac{S^2}{F^2}\cos \frac{\chi_{q}}{2} \, p(\sigma) \, \Xi_3 \Bigg] \, . \nonumber\end{aligned}$$ As we commented in the previous subsection all functions depend only on $\sigma$, hence it is possible to describe the system in terms of a one-dimensional effective action. Inserting all the ingredients in we obtain: $$S_{eff}=\frac{\pi ^3 V_{1,3}}{2\kappa_{10}^2}\int {\cal L}_{1d} \, d\sigma \label{S-eff-1}\ ,$$ where $V_{1,3}$ is the volume of the Minkowski space and ${\cal L}_{1d} $ is given by the following expression $$\begin{aligned} \label{L-effective1} {\cal L}_{1d} &=& -\frac{1}{2} \left(\frac{h'}{h}\right)^2 + 12 \left(\frac{S'}{S}\right)^2 + 8 \, \frac{F' S'}{F S} + 24\,b^2\, F^2\,S^6 - 4\, b^2\, F^4\,S^4 \nonumber \\ \nonumber \\ &+&\frac{b'}{b}\,\left( \frac{h'}{h}+ 8 \,\frac{S'}{S}+ 2\, \frac{F'}{F} \right)+ \frac{1}{2}\, \left(\frac{b'}{b}\right)^2 - \frac{b^2 Q_c^2}{2 h^2} -\frac{1}{2}\,Q_f^2\, p^2\, b^2e^{2\Phi} S^8 \,\Xi_3^2 \\ \nonumber \\ &-&4\,Q_f \, e^{\Phi}\, b^2\,F^2\,S^6 \, \Xi_1 \, \Xi_2 \, \Xi_3 \, \cos^3 \frac{\chi_{q}}{2} - \frac{1}{2}\,\Phi'^2 - \frac{1}{2}\,\frac{e^{-\Phi}\,H'^2\,h}{b^2} \left(1 - \frac{e^{2\Phi}\,J'^2\,b^2}{H'^{2}}\right) -Q_c H J' \, . \nonumber \end{aligned}$$ Since the potential $J$ enters the effective action only through its derivative, it corresponds to a “constant of motion". This new parameter is related to the value of the magnetic field close to the boundary through the equations of motion for $F_3$, coming from the 10d supergravity [@arXiv:1106.1330]. We will fix this constant of motion in the following way $$\label{defJ} \frac{\partial {\cal L}_{1d}}{\partial J'} \, \equiv - \, Q_c H_{\star} \quad \Rightarrow \quad J' \,=\, \frac{e^{-\Phi}\,Q_c}{h} \left(H-\, H_{\star}\right)\, .$$ The next step is to use equation to eliminate $J'$, in favor of $H_*$, in equation after performing the following Legendre transformation $$\label{L-effective2} \tilde {\cal L}_{1d} = L_{1d}-\frac{\delta L_{1d}}{\delta J'} \, J'\Bigg|_{J' \equiv J'(H,H_*)} \, ,$$ and then calculate the Euler-Lagrange equations from the transformed action . Defining the following auxiliary (dimensionless) expressions $$\begin{aligned} && \Xi_4 \equiv 1- \cot^2 \frac{\chi_q}{2}\,\, \frac{F^2}{S^2} \, , \quad \xi \equiv \cos^3 \frac{\chi_q}{2}\, \frac{\Xi_1 \,\Xi_2}{\Xi_3}\, , \\ && \beta_2 \equiv 1 + \frac{e^{2\Phi}\,J'^{2}\,b^2}{H'^2} \quad \& \quad \beta_3 \equiv 1 + \frac{e^{-2\Phi}\,H'^{2}\,\beta_2}{Q_f^2\,p^2\,H^2\, b^2\,S^8} \nonumber\ ,\end{aligned}$$ we can write the equations of motion in the following compact way $$\begin{aligned} \partial_\sigma^2(\log b)&=&-\, 4 Q_f\,H^2\, h \,S^6 \, F^2\, \xi \,-\,e^{\Phi}\,H^2\,Q_f^2\, p^2\, h \,S^8\,\beta_3 \label{diff-b}\ , \\ \nonumber \\ \partial_\sigma^2(\log h)&=&- \, Q_c^2 \frac{b^2}{h^2}\, -\,\frac{1}{2}\,e^{\Phi}\,H^2\,Q_f^2\,p^2\,h \,S^8\,\beta_3\, +\,\left(1-\beta_2\right)\,\frac{e^{-\Phi}\,h\,H'^2}{b^2} \label{diff-h}\ , \\ \nonumber \\ &-& 2 Q_f\,H^2\, h \,S^6 F^2 \xi\,\ , \nonumber \\ \nonumber \\ \partial_\sigma^2(\log S)&=& -2\, b^2 \, F^4 \, S^4 + 6\, b^2 \, F^2 \, S^6 + \frac{1}{4} \, Q_f^2 \, p^2 \, e^{\Phi} \, H^2\, h \,S^8\, \beta_3 \label{diff-S}\ , \\ \nonumber \\ &-&\frac{1}{2}\,Q_f \, e^\Phi \, b^2 \, F^2 \, S^6\, \cos^3 \frac{\chi_q}{2} \,\, \Xi_3 \Bigg[\cos^2 \frac{\chi_q}{2} \, \frac{\Xi_2}{\Xi_1} \, + \, \frac{\Xi_1}{\Xi_2}\left(1\,-\,\Xi^2_2\,+\, 2\frac{\Xi_2^2}{\Xi_3^2} \right)\Bigg]\ , \nonumber \\ \nonumber \\ \partial_\sigma^2(\log F)&=& 4\, b^2 \,F^4 \,S^4 - \frac{1+\Xi_3^2}{4} \, Q_f^2 \, p^2 \, e^{2\Phi} \, b^2\, S^8\, +\,\frac{1}{4}\,\frac{e^{-\Phi}\,h\,H'^2\,\beta_2}{b^2} \label{diff-F}\ , \\ \nonumber \\ &-& 2 \, Q_f \, e^{\Phi} \, b^2\, S^8 \, \cos^3 \frac{\chi_q}{2} \, \sin^2 \frac{\chi_q}{2} \, \frac{\Xi_2}{\Xi_1 \Xi_3} \, \Bigg[ 1\, - \, \frac{1}{2}\, \, \Xi_4 \, \left(1 - \Xi_3^2 \right) \Bigg]\ , \nonumber \\ \nonumber \\ \partial_\sigma^2\Phi&=& \frac{1+\Xi_3^2}{2} \Bigg[Q_f^2\,p^2\,e^{2\Phi}\,b^2\,S^8 + 4 Q_f\,b^2\, e^\Phi\,F^2\, S^6 \, \xi \Bigg] -\frac{1}{2}\,\frac{e^{-\Phi}\,h\,H'^2\,\beta_2}{b^2} \label{diff-Phi}\ , \\ \nonumber \\ \partial_{\sigma} \left[\frac{e^{-\Phi}\,h\,H'}{b^2}\right]&=& e^{\Phi}\,Q_f^2 \, p^2\,H\,h\,S^8\,+\,Q_c\,J' +\, 4 Q_f \,H \, h \, S^6 \, F^2 \, \xi \, . \label{diff-H}\end{aligned}$$ Together with the above system of EOM we get the following “zero-energy” constraint $$\begin{aligned} \label{constraint} 0&=& -\frac{1}{2} \left(\frac{h'}{h}\right)^2 + 12 \left(\frac{S'}{S}\right)^2 + 8 \, \frac{F' S'}{F S} - 24\,b^2\, F^2\,S^6 + 4\, b^2\, F^4\,S^4 \, - \, \frac{1}{2}\,\Phi'^2 \nonumber \\ \nonumber \\ &+&\frac{b'}{b}\,\left( \frac{h'}{h}+ 8 \,\frac{S'}{S}+ 2\, \frac{F'}{F} \right)+ \frac{1}{2}\, \left(\frac{b'}{b}\right)^2 + \frac{b^2 Q_c^2}{2 h^2} \, - \frac{1}{2}\,\frac{e^{-\Phi}\,H'^2\,h}{b^2} \left(1 - \frac{e^{2\Phi}\,J'^2\,b^2}{ H'^{2}}\right) \\ \nonumber \\ &+&\frac{1}{2}\,Q_f^2\, p^2 \, b^2 \, e^{2\Phi} S^8 \, \Xi_3 \, + \, 4\,Q_f\, b^2\,e^{\Phi}\,F^2\,S^6 \,\frac{ \Xi_1 \, \Xi_3}{\Xi_2 } \, \cos^3 \frac{\chi_{q}}{2} \, . \nonumber \end{aligned}$$ Equation (\[constraint\]) is obtained by requiring invariance of the one dimensional effective action (\[S-eff-1\]) under an infinitesimal reparameterization $\sigma\to(1+\delta\lambda)\sigma$. This is equivalent to requiring that the one-dimensional Hamiltonian ${\cal H}_{1d}$ corresponding to ${\cal L}_{1d}$ vanishes. An important observation is that in generating the Hamiltonian ${\cal H}_{1d}$ one should Legendre transform with respect to all fields in ${\cal L}_{1d}$, including the field $\chi_q(\sigma)$ specifying the profile of the fiducial embedding. This suggests that we can treat $\chi_q(\sigma)$ as a dynamical variable, therefore we can also obtain the equation of motion (\[EOM-chi\]) from varying the effective lagrangian (\[L-effective1\]). Remarkably if we use equation (\[p-sigma\]) to substitute for $p(\sigma)$ in the effective Lagrangian ${\cal L}_{1d}$ and derive the corresponding Euler-Lagrange equation for $\chi_q(\sigma)$, we reproduce exactly the equation of motion (\[EOM-chi\]). This provides a non-trivial self-consistency check of the construction outlined above. The system & – allows for a systematic expansion of all the functions in power series of $Q_f$, as defined in equation (\[QcQf\]). In fact physically it is more relevant to expand in the parameter, $\epsilon_*$ $${\epsilon_*}\equiv Q_f\, e^{\Phi_*}\ ,\label{epsilonstar}$$ which takes into account the running of the effective ’t Hooft coupling (through the dilaton factor $e^{\Phi_*}$ in ). We consider the following first order expansion in $\epsilon_*$: $$\begin{aligned} & b \, = \, 1 \, + \, \epsilon_* b_1 \, , \quad h \, = \, \frac{R^4}{r^4} \, \left(1+\epsilon_*h_1\right) \, , \quad S \, = \, r \, \left(1+\epsilon_*S_1\right)\ , & \label{expansion1} \\ \nonumber \\ & F\, = \, r \, \left(1+\epsilon_*F_1\right) \, , \quad \Phi \, = \, \Phi_*+\epsilon_*\Phi_1 \, , \quad H \, = \, H_* \left( 1 \, + \, \epsilon_* H_1 \right) \quad \& \quad \chi_q \, = \, \chi_0 + \epsilon_* \chi_1 \, ,& \nonumber\end{aligned}$$ together with the reparametrization: $$r_m^4=e^{-\Phi_*}H_*^2R^4 \quad \& \quad \tilde r=\frac{r}{r_m} \, ,\label{dimless-r}$$ where $R^4\equiv Q_c/4$. The result is a coupled system of second order differential equations which can be decoupled by the transformations: $$\Delta_1 \, \equiv \, F_1 \, - \, S_1 \, , \quad \quad \Lambda_1 \, \equiv \, F_1 \, + \, 4 \, S_1 \, + \, \frac{5}{4} \, \, b_1 \quad \quad \& \quad \quad \lambda_1 \, \equiv \, h_1 \, - b_1 \, \ .$$ For the decoupled system we obtain: $$\begin{aligned} \label{EOMs} && \chi_0'' \, + \, \frac{5}{\tilde{r}} \, \frac{{\tilde r}^4 \,+\,\frac{3}{5}}{{\tilde r}^4\,+\,1} \,\chi_0' +\, \frac{{\tilde r} \left({\tilde r}^4 \,+\,\frac{1}{2}\right)}{{\tilde r}^4\,+\,1} \,\chi_0'^{\,3} \, = \, \, - \, \frac{6}{{\tilde r}^2} \, \tan \frac{\chi_{0}(\tilde{r})}{2} \, \sqrt{1 \, + \, \frac{\tilde{r}^2}{4} \, \chi_{0}'(\tilde{r})^2} \, , \nonumber \\ \nonumber \\ && \lambda_1'' \, + \, \frac{5}{\tilde{r}} \, \lambda_1' \, -\, \frac{32}{\tilde{r}^2} \, \lambda_1 \, = \, \frac{1}{2} \, J_{\lambda_1} \, , \qquad \qquad \,\, H_1'' \, + \, \frac{1}{\tilde{r}} \, H_1' \, -\, \frac{16}{\tilde{r}^2} \, H_1 \, = \, \frac{1}{2}\, J_{H_1} \, , \nonumber \\ \nonumber \\ && \Phi_1'' \, + \, \frac{5}{\tilde{r}} \, \Phi_1' \, = \, \frac{1}{2} \, J_{\Phi_1} \, , \qquad \qquad \qquad \qquad \qquad \qquad \quad b_1'' \, + \, \frac{5}{\tilde{r}} \, b_1' \, = \, \frac{1}{2} \, J_{b_1} \, , \nonumber \\ \nonumber \\ && \Delta_1'' \, + \, \frac{5}{\tilde{r}} \, \Delta_1' \, -\, \frac{12}{\tilde{r}^2} \, \Delta_1 \, = \, \frac{1}{2}\, J_{\Delta_1} \, , \qquad \qquad \Lambda_1'' \, + \, \frac{5}{\tilde{r}} \, \Lambda_1' \, -\, \frac{32}{\tilde{r}^2} \, \Lambda_1 \, = \, \frac{1}{2}\, J_{\Lambda_1} \, , $$ where the analytic expressions for the sources $J_{\lambda_1}, J_{H_1}, J_{\Phi_1}, J_{b_1}, J_{\Delta_1}, J_{\Lambda_1}$ appear in Appendix \[nonsusy-ingredients\]. The equation coming from the constraint (\[constraint\]) is: $$\label{EOM-eta} \eta_1' \, - \, \frac{4}{{\tilde r}} \, \eta_1 \, = \, - \, J_{\eta_1} ({\tilde r})$$ with $\eta_1 \, \equiv 2 \Lambda_1\, + \, \lambda_1$ and $J_{\eta_1}$ given in Appendix \[nonsusy-ingredients\]. As one can see all of the equations of motion, except the one for the fiducial embedding, in are linear, therefore it is possible to obtain their solution in an integral form in terms of appropriate Greens functions. On the other hand the non-linear equation of motion for $\chi_0$ is the same as the one for a probe D7–brane studied in ref. \[ours\]. In fact it is the only non-linear equation that we have in our construction and we solve it numerically. The general solution for the classical D7–brane has the following expansion at large $\tilde r$ $$\sin\frac{\chi_0(\tilde r)}{2}=\frac{\tilde m_0}{\tilde r}+\frac{\tilde c}{\tilde r^3}+O\left(\frac{1}{\tilde r^5}\right)\ ,$$ where $\tilde m_0\equiv m_0/r_m$ & $\tilde c\equiv c/r_m^3$, while $m_0$ & $c$ are proportional to the bare mass and fundamental condensate of the dual field theory. Inside the bulk of the geometry $\chi_0$ terminates at a given radial distance $\tilde r_{\rm min}=\tilde r_q$ at which the $S^3$ cycle wrapped by the D7–brane vanishes. As described at the beginning of this section the smearing procedure produces a spherical cavity of radius $\tilde r_q$. Inside this cavity the solution is sourced solely by the colour D3–branes through the self dual RR $F_{5}$ form. Before presenting the solution for the functions of the first order perturbative expansion , we will elaborate more on the equations of motion inside the cavity. Vacuum solution inside the cavity --------------------------------- The smearing of massive flavour D7–branes produces a hollow spherical cavity at the origin of the subspace transverse to the colour D3–branes. Inside this cavity the gravitational background is a solution of the vacuum equations of motion (setting $Q_f = 0$ in equations & –). It turns out that it is possible to obtain non-perturbative solutions for all background functions, except for the functions $S$ and $F$ describing the radii of the $CP_2$ internal subspace and corresponding $U(1)$ fiber bundle. ### The equations of motion in the vacuum The equations of motion inside the cavity are given by $$\begin{aligned} \partial_\sigma^2(\log b)&=&-\frac{e^{-\Phi}\,h\,H'^2}{b^2} -e^{\Phi}\,h\,J'^2 \, , \label{diff-bv} \\ \nonumber \\ \partial_\sigma^2(\log h)&=&-Q_c^2\frac{b^2}{h^2}-\frac{1}{2}\frac{e^{-\Phi}\,h\,H'^2}{b^2} -\frac{3}{2}e^{\Phi}\,h\,J'^2 \, , \label{diff-hv} \\ \nonumber \\ \partial_\sigma^2(\log S)&=& -2\, b^2 F^4 S^4 + 6\, b^2 F^2 S^6+\frac{1}{4}\frac{e^{-\Phi}\,h\,H'^2}{b^2} +\frac{1}{4}e^{\Phi}\,h\,J'^2 \, , \label{diff-Sv} \\ \nonumber \\ \partial_\sigma^2(\log F)&=& 4\,b^2 F^4 S^4+\frac{1}{4}\frac{e^{-\Phi}\,h\,H'^2}{b^2} +\frac{1}{4}e^{\Phi}\,h\,J'^2 \, , \label{diff-Fv} \\ \nonumber \\ \partial_\sigma^2\Phi&=& -\frac{1}{2}\frac{e^{-\Phi}\,h\,H'^2}{b^2} -\frac{1}{2}e^{\Phi}\,h\,J'^2 \, , \label{diff-Phiv} \\ \nonumber \\ \partial_{\sigma} \left[\frac{e^{-\Phi}\,h\,H'}{b^2}\right]&=& \,Q_c\,J'\, , \label{diff-Hv} \\ \nonumber \\ \partial_{\sigma} \left[e^{\Phi}\,h\,J'\right]&=& \,Q_c\,H'\, . \label{diff-Jv}\end{aligned}$$ Adding and subtracting in various ways , , & we easily obtain the following system of equations without sources $$\begin{aligned} \partial_\sigma^2\log \frac{F^2}{S^2} \, &=& \, - \, 12\, \left(b \, S^4\right)^2 \, \frac{F^2}{S^2} \, \left(1 \, - \, \frac{F^2}{S^2}\right) \, , \label{diff-F/S} \\ \nonumber \\ \partial_\sigma^2\log(bS^4) \, &=& \, 8\, \left(b \, S^4\right)^2 \, \frac{F^2}{S^2} \, \left(3 \, - \, \frac{F^2}{S^2}\right) \, , \label{diff-bS^4} \\ \nonumber \\ \partial_\sigma^2\log e^{-2\Phi}b \, &=& \, 0 \label{diff-be^-2Phi}\ .\end{aligned}$$ While strongly suggests $e^{2\Phi}\propto b$ for & we need to define a new set of variables $$U \, \equiv \, b \, S^4 \qquad \& \qquad V \, \equiv \,\frac{F^2}{S^2} \, .$$ In these variables the equations of motion for $V$ & $U$ decouple in the following way $$\begin{aligned} \partial_\sigma^2\log U \, & = & \, 8U^2V \, \left(3 \, - \, V \right) \, , \label{diff-U} \\ \nonumber \\ \partial_\sigma^2\log V \, & = & \, - \, 12 U^2V \, \left( 1 \, - \, V \right)\ . \label{diff-V} \end{aligned}$$ Notice that $V$ is the ratio between the radii of the $CP_2$ and the fiber in the $S^5$, therefore it is a measure for the relative squashing. The following solution to & $$V(\sigma) \, = \, 1 \quad \& \quad U(\sigma) \, = \, \frac{1}{4\sigma+{\rm const}}\, \ , \label{SUSY-S5}$$ corresponds to a non-squashed $S^5$. Since in our case the vacuum solution at the boundary of the cavity should match the flavour background, we are interested in deformations of corresponding to a squashed $S^5$. We have no reason to expect enhancement of the global symmetry of the theory in the deep IR. From a field theory point of view the squashing of the $S^5$ corresponds to a breaking of the global $SU(4)$ symmetry down to $SU(3)\times U(1)$. Inside the cavity the gravitational background corresponds to the effective field theory obtained after integrating out the flavours. Now let us briefly discuss the rest of the vacuum equations of motion. A natural candidate for a solution consistent with our ansatz is the supergravity background dual to a non-commutative Yang-Mills, which can be obtained as a near horizon limit of the supergravity solution corresponding to the D3-D1 bound state. Indeed one can check that upon the following substitution $$\begin{aligned} b(\hat r) \, &=& \, \frac{1 + c_b\Theta^2}{1+\Theta^4\hat r^4} \, , \label{sol-b} \\ \nonumber \\ e^{2\Phi(\hat r)} \, &=& \, e^{2\Phi_*}\frac{1+c_{\Phi}\Theta^2}{1+\Theta^4\hat r^4}\, , \\ \nonumber \\ H(\hat r) \, &=& \, H_{*}-\Theta^2e^{\frac{\Phi_*}{2}}\frac{\hat r^4}{R^2} \frac{(1+c_b\Theta^2)^{\frac{1}{2}}}{1+\Theta^4\hat r^4}(1+c_{\Phi}\Theta^2)^{\frac{1}{4}} \, , \label{sol-H}\\ \nonumber \\ J(\hat r) \, &=& \, \Theta^2e^{-\frac{\Phi_*}{2}}\frac{\hat r^4}{R^2}(1+c_b\Theta^2)^{-\frac{1}{2}}(1+c_{\Phi}\Theta^2)^{-\frac{1}{4}} \, , \\ \nonumber \\ h(\hat r) \, &=& \, \frac{R^4}{\hat r^4}\frac{1+c_b\Theta^2}{\sqrt{1+\Theta^4\hat r^4}} \, , \label{sol-h}\\ \nonumber \\ \sigma(\hat r)&=&\frac{1}{4\hat r^4}+c_\sigma\Theta^2 \, ;~~~~\Theta^2\equiv e^{\frac{\Phi_*}{2}}\Theta^{23}/R^2\ , \label{sol-sigm}\end{aligned}$$ equations , , , & are satisfied where $\Theta^{23}$ is the parameter of non-commutativity of the dual field theory in the ($x^2,x^3$)-plane. The constants of integration $c_b\, ,c_{\Phi}$ and $c_\sigma$ will be fixed by matching to the flavour part of the background at the boundary of the cavity. Notice that the system of equations & is completely decoupled and describes the different deformations of the compact part of the geometry. Expressions - together with constitute a full solution corresponding to the gravity dual of a non-commutative SYM with $SU(4)$ global symmetry. It is possible to find a more general class of solutions corresponding to a non-commutative SYM with $SU(3)\times U(1)$ global symmetry and this will be the topic of the next subsection. ### Deforming the sphere We now construct a non-commutative non-supersymmetric background by considering perturbative solutions of the system & around starting from the following ansatz: $$U(\sigma)=\frac{1}{4\sigma} \, \left(1+\Sigma\,u_1 \right) \quad \& \quad V(\sigma) \, = \, 1+\Sigma\,v_1 \, , \label{pert-UV}$$ where $u_1$ & $v_1$ are functions of $\sigma$ to be determined after substituting in & and expanding in $\Sigma$. At first order we have $$u_1''-\frac{2}{\sigma^2} \, u_1-\frac{1}{2\sigma^2} \, v_1 \, = \, 0 \quad \& \quad v_1'' - \frac{3}{4\sigma^2} \, v_1 \, = \, 0 \, , \label{FOSigma}$$ and after an appropriate redefinition $$\beta \, \equiv \, \frac{5}{2} \, u_1 +v_1 \quad \Rightarrow \quad \beta''-\frac{2}{\sigma^2} \, \beta \, = \, 0 \quad \& \quad v_1''-\frac{3}{4\sigma^2} \, v_1 \, = \, 0 \, .$$ The general solution is given by $$\beta(\sigma) \, = \, \tilde c_{1}\sigma^2 + \tilde c_{2}\sigma^{-1} \quad \& \quad v_1 \, = \, \tilde c_3\sigma^{3/2} + \tilde c_4\sigma^{-1/2} \, ,$$ and in terms of the radial variable $\hat r$ defined in the functions $u_1$ & $v_1$ are given by $$\begin{aligned} u_1&=&c_1\frac{1}{\hat r^8}+c_2{\hat r^4}-\frac{2}{5}c_3\hat r^{-6}-\frac{2}{5}c_4\hat r^2\ , \\ v_1&=&c_3\hat r^{-6}+c_4\hat r^2\ , \label{pert-uv}\end{aligned}$$ where the constants $c_i$ are related to the constants $\tilde c_i$ by a non-singular diagonal linear transformation. Notice that in the formalism of the parameter of non-commutativity $\Theta^2$ is proportional to $\Sigma$ and hence $\sigma=1/(4\hat r^4)+O(\Sigma)$. Analyzing , the general solution regular at $\hat r=0$ is spanned by $c_2\neq0, c_4\neq0\ , c_1=c_3=0$. The constants $c_2$ and $c_4$ will be determined by matching to the flavour solution at the boundary of the cavity (after setting $\Sigma \sim \epsilon_*$). Constructing the full perturbative solution ------------------------------------------- Our construction is a perturbative expansion near a supersymmetric AdS$_5\times S^5$ background. Requiring that the leading order corrections inside and outside the cavity agree implies that $\Theta^2\sim\epsilon_*$. Furthermore, we introduce the radial variable $r$ related to $\sigma$ via $r^4\equiv\frac{1}{4\sigma}$ and its dimensionless analog $\tilde r$ defined in equation (\[dimless-r\]). The resulting expansions are $$\begin{aligned} b&=&1+c_b\Theta^2 \, , \label{per-b}\\ \Phi&=&\Phi_*+\frac{1}{2} \, c_{\phi} \, \Theta^2 \, , \label{per-Phi}\\ H(r)&=&H_*-\Theta^2e^{\frac{\Phi_*}{2}}\,\frac{r_m^4}{R^2}\,\tilde r^4 \, , \label{per-H}\\ J(r)&=&\Theta^2e^{-\frac{\Phi_*}{2}}\,\frac{r_m^4}{R^2} \, {\tilde r^4} \, , \label{per-J}\\ h(r)&=&\frac{R^4}{r^4}\Bigg[1+\Theta^2\left(c_b-4c_{\sigma}r_m^4\tilde r^4\right)\Bigg]\ . \label{per-h}\end{aligned}$$ Since does not depend on any constants of integration we can use the perturbative expansion of the flavour solution to determine the exact relation between $\Theta^2$ and $\epsilon_*$. Indeed comparing and we conclude that for $\tilde r\leq \tilde r_q$ we have $$\epsilon_*H_1(\tilde r)=-\Theta^2e^{\frac{\Phi_*}{2}}\,\frac{r_m^4}{R^2}\,\tilde r^4\, , \label{H1-1}$$ where $\tilde r_q$ is the radius of the cavity. In the following subsections we will present semi-analytic solutions for all the functions of the first order perturbative expansion, requiring specific behaviors inside the cavity emerging from the expansions -. ### Solving for the $B$-field In order to solve the equation of motion for $H_1$ in we need the Greens function $G_{H_1}(\tilde r, \tilde r_1)$ satisfying the following equation $$\begin{aligned} \frac{d^2}{d{\tilde r}^2} \, G(\tilde r,\tilde r_1) \, + \, \frac{1}{\tilde r} \, \frac{d}{d\tilde r} \, G(\tilde r,\tilde r_1) \, - \, \frac{16}{\tilde r^2} \, G(\tilde r,\tilde r_1)\, & = & \, \delta(\tilde r-\tilde r_1) \nonumber \\ \nonumber \\ \quad \quad \quad \Rightarrow \quad \quad \quad G_{H_1}(\tilde r,\tilde r_1) \, & = & \, \frac{1}{8} \, \frac{{\tilde r}^4}{{\tilde r_1^3}} \, \left(1\, - \, \frac{\tilde r_1^8}{\tilde r^8} \right) \, \Theta(\tilde r-\tilde r_1) \, . \label{Green-H1}\end{aligned}$$ The solution for $H_1$ that inside the cavity is $\propto\tilde r^4$ and vanishes at $\tilde r=\tilde r_*$ can be written in the following integral form $$\label{semi-H1} H_1(\tilde r) \, = \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{H_1}({\tilde r}_1) \Bigg[G_{H_1}(\tilde r,\tilde r_1) \, - \, \frac{\tilde r^4}{\tilde r_*^4} \, G_{H_1}(\tilde r_*,\tilde r_1)\Bigg] \, .$$ Elaborating on inside the cavity ($\tilde r\leq\tilde r_q$) we have $$H_1(\tilde r)=-H_*\#\left(\tilde r_q,\tilde r_*\right) \,\tilde r^4\propto\tilde r^4 \, , \label{H1-2}$$ where the expression for the constant $\#\left(\tilde r_q,\tilde r_*\right)$ is $$\label{constant} \#\left(\tilde r_q,\tilde r_*\right)=\int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r\frac{\left(1-\frac{\tilde r^8}{\tilde r_*^8}\right) }{2\tilde r^3\sqrt{{\tilde r}^4 \, + 1}} \tilde\xi\left(\tilde r,\tilde r_q\right) \, .$$ Comparing and we obtain the promised relation connecting the non-commutativity parameter to the number of flavours $$\Theta^2=\#\left(\tilde r_q,\tilde r_*\right)\frac{1}{r_m^2}\, \epsilon_*\propto \frac{N_f}{N_c}\frac{1}{H_*}\, . \label{non-com}$$ ### Solving for the dilaton & $b_1$ In order to solve the equations of motion for both the dilaton and $b_1$ in we need the Greens function $G_{\{\Phi_1, b_1\}}(\tilde r, \tilde r_1)$ satisfying the following equation $$\begin{aligned} \frac{d^2}{d{\tilde r}^2} \, G(\tilde r,\tilde r_1) \, + \, \frac{5}{\tilde r} \, \frac{d}{d\tilde r} \, G(\tilde r,\tilde r_1) \, & = & \, \delta(\tilde r-\tilde r_1) \nonumber \\ \nonumber \\ \quad \quad \quad \Rightarrow \quad \quad \quad G_{\{\Phi_1, b_1\}}(\tilde r,\tilde r_1) \, & = & \,\frac{\tilde r_1}{4} \, \left(1\, - \, \frac{\tilde r_1^4}{\tilde r^4} \right) \, \Theta(\tilde r-\tilde r_1) \, . \label{Green-H1}\end{aligned}$$ Looking at and we see that both the dilaton and $b$ are constant inside the cavity. The solution for $\Phi_1$ & $b_1$ which is constant inside the cavity and vanishing at $\tilde r=\tilde r_*$ can be written in the following integral form $$\begin{aligned} \Phi_1(\tilde r) \, &=& \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{\Phi_1}(\tilde r_1) \, \Bigg[G_{\{\Phi_1, b_1\}}(\tilde r,\tilde r_1) \, - \, G_{\{\Phi_1, b_1\}}(\tilde r_*,\tilde r_1) \Bigg] \, , \label{semi-Phi1} \\ b_1(\tilde r) \, &=& \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{b_1}(\tilde r_1) \, \Bigg[G_{\{\Phi_1, b_1\}}(\tilde r,\tilde r_1) \, - \, G_{\{\Phi_1, b_1\}}(\tilde r_*,\tilde r_1)\Bigg] \, . \label{semi-b1}\end{aligned}$$ Inside the cavity both $\Phi_1$ and $b_1$ are constant and are given by the following expressions $$\begin{aligned} \Phi_1(\tilde r)&=& \, - \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{\Phi_1}(\tilde r_1)\, G_{\{\Phi_1, b_1\}}(\tilde r_*,\tilde r_1) \, , \label{phi1-const}\\ b_1(\tilde r)&=& \, - \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{b_1}(\tilde r_1)\, G_{\{\Phi_1, b_1\}}(\tilde r_*,\tilde r_1)\, \ .\label{b1-const}\end{aligned}$$ Equations , and can be used to fix the constants $c_b$ and $c_{\Phi}$. ### Solving for $\lambda_1$ Proceeding in the same way we obtain the Green’s function for the equation of motion for $\lambda_1$, defined in : $$\begin{aligned} \frac{d^2}{d{\tilde r}^2} \, G(\tilde r,\tilde r_1) \, + \, \frac{5}{\tilde r}\frac{d}{d\tilde r} \, G(\tilde r,\tilde r_1) \, - \, \frac{32}{\tilde{r}^2} \, G(\tilde r,\tilde r_1) \, & = & \, \delta(\tilde r-\tilde r_1) \nonumber \\ \nonumber \\ \quad \quad \quad \Rightarrow \quad \quad \quad G_{\lambda_1}(\tilde r,\tilde r_1) \, & = & \, \frac{\tilde r^4}{12 \, \tilde r_1^3} \, \left(1\, - \, \frac{\tilde r_1^{12}}{\tilde r^{12}} \right) \, \Theta(\tilde r-\tilde r_1) \, , \label{Green-lambda1}\end{aligned}$$ and look for a solution that inside the cavity is $\propto \tilde r^4$ and vanishes at $\tilde r=\tilde r_*$ $$\lambda_1 \, = \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{\lambda_1}({\tilde r}_1) \, \Bigg[G_{\lambda_1}(\tilde r,\tilde r_1) \, - \, \frac{\tilde r^4}{\tilde r_*^4} \, G_{\lambda_1}(\tilde r_*,\tilde r_1)\Bigg]\ . \label{semi-lambda1}$$ ### Solving for $\Delta_1$ Proceeding in the same way for $\Delta_1$ we have $$\begin{aligned} \frac{d^2}{d{\tilde r}^2} \, G(\tilde r,\tilde r_1) \, + \, \frac{5}{\tilde r}\frac{d}{d\tilde r} \, G(\tilde r,\tilde r_1) \, - \, \frac{12}{\tilde{r}^2} \, G(\tilde r,\tilde r_1) \, & = & \, \delta(\tilde r-\tilde r_1) \nonumber \\ \nonumber \\ \quad \quad \quad \Rightarrow \quad \quad \quad G_{\Delta_1}(\tilde r,\tilde r_1) \, & = & \, \frac{\tilde r^2}{8 \, \tilde r_1} \, \left(1\, - \, \frac{\tilde r_1^{8}}{\tilde r^{8}} \right) \, \Theta(\tilde r-\tilde r_1) \, .\end{aligned}$$ Looking for a solution which behaves as $\Delta_1(\tilde r)\propto \tilde r^2$ inside the cavity and asymptotes to the supersymmetric solution at $\tilde r_*$ $$\eqref{susy-Delta1} \qquad \Rightarrow \qquad \Delta_1({\tilde r}_*) \, = \, - \, \frac{1}{12} \, \left(1 \, - \frac{{\tilde m}_0^2}{{\tilde r}_*^2} \right)^3\ ,$$ we have $$\Delta_1(\tilde r) \, = \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{\Delta_1}(\tilde r_1) \, \Bigg[G_{\Delta_1}(\tilde r,\tilde r_1)-\frac{\tilde r^2}{\tilde r_*^2}G_{\Delta_1}(\tilde r_*,\tilde r_1)\Bigg] \, - \, \frac{1}{12} \, \frac{\tilde r^2}{\tilde r_*^2} \, \left(1 \, - \frac{{\tilde m}_0^2}{{\tilde r}_*^2} \right)^3\, . \label{semi-Delta1}$$ ### Solving for $\Lambda_1$ The Greens function for $\Lambda_1$ is given by so we proceed immediately in looking for a solution which behaves as $\Lambda_1(\tilde r)\propto \tilde r^4$ inside the cavity and asymptotes to the supersymmetric solution at $\tilde r_*$ $$\eqref{susy-Lambda1} \qquad \Rightarrow \qquad \Lambda_1({\tilde r}_*) \, = \, \ , \frac{1}{18} \, \left(1 \, - \frac{{\tilde m}_0^2}{{\tilde r}_*^2} \right)^3$$ we have $$\Lambda_1(\tilde r) \, = \, \frac{1}{2} \, \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r_1 \, J_{\Lambda_1}(\tilde r_1) \, \Bigg[G_{\Lambda_1}(\tilde r,\tilde r_1)-\frac{\tilde r^4}{\tilde r_*^4}G_{\Lambda_1}(\tilde r_*,\tilde r_1)\Bigg] \, + \, \frac{1}{18} \, \frac{\tilde r^4}{\tilde r_*^4} \, \left(1 \, - \frac{{\tilde m}_0^2}{{\tilde r}_*^2} \right)^3\, . \label{semi-Lambda1}$$ ### Solving the constraint for $\eta_1$ The first order differential equation for $\eta$ can be integrated to give $$\eta_1({\tilde r}) \, = \, \tilde r^4 \Bigg[ \frac{\eta_1({\tilde r}_*)}{{\tilde r}_*^4} \, + \, \int \limits_{{\tilde r}}^{\tilde r_*}d\hat r \, \frac{J_{\eta_1}({\hat r})}{\hat r^4}\,\Theta(\tilde r-\tilde r_q) \Bigg] \quad \text{with} \quad \eta_1(\tilde r_*) \, \Rightarrow \, \frac{1}{9} \, \left(1 \, - \frac{{\tilde m}_0^2}{{\tilde r}_*^2} \right)^3 \, .$$ The value of $\eta_1(\tilde r_*)$ is fixed by the requirement of obtaining the supersymmetric solution at $\tilde r=\tilde r_*$. ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](H1new.pdf "fig:"){width="3.4in"} ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](BigLambda1.pdf "fig:"){width="3.4in"} ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](Smalllambda1.pdf "fig:"){width="3.4in"} ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](Phi1.pdf "fig:"){width="3.4in"} ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](b1.pdf "fig:"){width="3.4in"} ![Family of solutions of the background functions for different bare mass parameter $\tilde m$.[]{data-label="fig:fig0"}](Delta1.pdf "fig:"){width="3.4in"} Numerical solution ------------------ As we learned from the previous subsection it is possible to obtain a solution for our supergravity background in a semi-analytic form in which all background functions are written in an integral form. The solution is completely determined by the D7–brane charge distribution function $p(\tilde r)$ or equivalently by the profile of the fiducial embedding $\chi_q(\tilde r)$. Furthermore, in the perturbative expansion the equations of motion for the first order corrections to the background fields are sourced by the zeroth order expansion of $\chi_q(\tilde r)$ (namely $\chi_0(\tilde r)$), which is known from the probe approximation. The equation of motion for $\chi_0(\tilde r)$ is non-linear and we solve it numerically. An obvious approach to obtain a numerical solution for the background functions would be to solve numerically the integral expressions presented in section 2. Actually it is more convenient to solve the equations of motion completely numerically using Mathematica’s built in function NDSolve. To generate the solution we employ standard shooting techniques and sew the numerical solution to the analytic solution inside the cavity at $\tilde r=\tilde r_q\,$. The constants of integration (which specify the shooting parameters) are obtained by matching the solution to the supersymmetric one at $\tilde r =\tilde r_*\,$. This approach has the disadvantage that the accumulated numerical error grows strongly with the parameter $\tilde r_*$, which we need to keep numerically high. It can be improved if we approximate the function $\chi_0$ with a taylor series expansion in $1/\tilde r$ for $\tilde r > \tilde r_{sew}$ ($\tilde r_{sew}$ is a number of order ten) and solve analytically for the background fields in the region $\tilde r_{sew}\leq \tilde r\leq \tilde r_*$ by matching the solution to the numerical one at $\tilde r_{sew}$ and to the supersymmetric one at $\tilde r_*$. Repeating the procedure outlined above for different values of the radius of the cavity $\tilde r_q$ generates solutions corresponding to different quark masses. In figure \[fig:fig0\] we have presented plots of the background functions for the following range of the bare mass parameter $\tilde m_0\in [0,2.3]$, which is the one needed for the analysis in section 5. As one can see the solutions corresponding to different values of $\tilde m$ differ significantly in the infrared $\tilde r \ll \tilde r_*$ and are very close to each other for $\tilde r\lesssim\tilde r_*$. In fact one can check that, except for the function $H_1$, for $\tilde r\lesssim \tilde r_*$ the background functions are very well approximated by their corresponding functions from the supersymmetric limit. The function $H_1$ follows the pattern $H_1(\tilde r)\approx\frac{1}{4}(\frac{\tilde r^4}{\tilde r_*}-1)$ represented by the black dashed curve in figure \[fig:fig0\]. Apparently at $\tilde r_*$, $H_1(\tilde r_*)$ vanishes, but has a non-vanishing first derivative, therefore the matching to the supersymmetric solution is not smooth at $\tilde r_*$. However, one can check that $H_1'(\tilde r_*)=1/\tilde r_*$ which is sub-leading for $\tilde r_*\gg 1$. Overall we conclude that our solution is very well approximated by the supersymmetric background near $\tilde r_*$ and hence one can approximate the background with the supersymmetric one for $\tilde r >\tilde r_*$. This is particularly useful to investigate the UV behaviour of the background, because it relates non-perturbatively the arbitrary UV scale $r_*$ and the parameter of perturbative expansion $\epsilon_*$ to the position of the Landau pole $r_{\rm LP}$. In the next subsection we provide a detailed description of the hierarchy of scales and regime of validity of our perturbative solution. Hierarchy of scales and regime of validity ------------------------------------------- In this subsection we analyze the regime of validity of our perturbative solution and the hierarchy of energy scales (in terms of radial scales) of the theory. Our analysis follows closely section 2.4 of ref. [@arXiv:0909.2865], where the finite temperature system has been unquenched. In terms of radial coordinates the hierarchy of scales can be written in the following way: $$0<r_m\sim r_q \ll r_* \ll r_a < r_{\rm LP}\ ,$$ where $r_m$ and $r_q$ represent IR energy scales related to the energy scale set by the magnetic field and the one corresponding to the physical mass of the fundamental matter. The radial scales $r_a$ and $r_{\rm LP}$ represent UV energy scales corresponding to the scale at which the supergravity solution develops pathologies and the Landau pole of the theory at which the effective ’t Hooft coupling blows up. For energy scales close to $r_*$ our solution, as can be seen from the plots in figure [\[fig:fig0\]]{}, is well approximated by the supersymmetric solution corresponding to vanishing magnetic field [@arXiv:0807.0298], which (given our choice $r_q\ll r_*$) is well approximated by the supersymmetric solution corresponding to massless fundamental fields [@hep-th/0612118]. This is why the analysis of the UV energy scales is exactly the same as in ref. [@arXiv:0909.2865] and in particular the relation between the “position" of the Landau pole $r_{\rm LP}$, the finite cutoff $r_*$ and the perturbative parameter $\epsilon_*$ is: $$\frac{r_*}{r_{\rm LP}}\approx e^{-1/\epsilon_*}\ll 1\ .$$ Furthermore in order for our perturbative solution to be valid in the region $r_q\leq r\leq r_*$ we need $e^{\Phi(r_q)}/e^{\Phi_*}\sim 1$, which requires $\epsilon_*\,|\Phi_1(\tilde r_q)|\ll 1$ (because $|\Phi_1(\tilde r_q)|$ is the biggest off all first order corrections). However, as one can check from the plot in Figure \[fig:fig0\], ${\rm max}{}_{\tilde r_q}|\Phi_1(\tilde r_q)|\sim \log\frac{r_*}{r_q}$. Therefore we need $\frac{r_q}{r_*}\gg e^{-1/\epsilon_*}$. In addition we would like to be able to neglect terms sub-leading in the $r_q/r_*$ expansion. Therefore we have to make sure that the corrections $\epsilon_*$ that we are considering are much larger than the one that we ignore. This requires $\epsilon \gg \frac{r_q}{ r_*}$. In summary we have: $$e^{-1/\epsilon_*}\ll\frac{M_q}{\Lambda_{\rm UV}}\sim \frac{r_q}{r_*}\ll\epsilon_*\sim \frac{\lambda_*\, N_f}{8\pi^2\, N_c}\ll 1\ .$$ Finally, validity of the supergravity approximation requires that we ignore closed string loops ($N_c\gg 1$) and $\alpha'$ corrections ($\lambda_q \gg 1$), where $\lambda_q$ is the effective ’t Hooft coupling at the energy scale set by $r_q\sim M_q$. It is related to $\lambda_*$ via $\lambda_q=\frac{e^{\Phi(r_q)}}{e^{\Phi_*}}\lambda_*$. In addition, validity of the smearing approximation suggests large $N_f$. In summary we have: $$N_c \gg 1,\quad \lambda_{q}\gg 1\ ,\quad \epsilon_{q}\equiv \frac{\lambda_q\,N_f}{8\pi^2\,N_c}\ll 1\ ,$$ where we have defined the IR perturbative parameter $\epsilon_q$. Clearly it is related to $\epsilon_*$ (defined in equation ) via: $$\epsilon_{q}=\epsilon_*\frac{e^{\Phi(\tilde r_q)}}{e^{\Phi_*}}=\epsilon_*(1+\epsilon_*\Phi_1(\tilde r_q))+O(\epsilon_*^3)\ ,$$ which also implies: $$\epsilon_{*}=\epsilon_q(1-\epsilon_q\Phi_1(\tilde r_q))+O(\epsilon_q^3)\ .\label{epsstar}$$ Finally requiring that $\alpha'$ corrections, which scale as $\lambda_q^{-3/2}$, (for more details look at ref. [@arXiv:0909.2865]) are sub-leading relative to flavour corrections controlled by $\epsilon_q$ requires: $$\lambda_q^{-3/2}\ll \epsilon_q \ .$$ We close this section by commenting that the numerical values used in our analysis are in the regime of validity specified above. Free Energy and Condensate ========================== In this section we calculate the free energy density and fundamental condensate of the dual field theory directly from the supergravity background. Helmholtz versus Gibbs free energy ---------------------------------- Following the general prescription of [@Hawking:1995fd] we identify the on-shell Euclidean action with the Helmholtz free energy. The Euclidean action has contributions from two terms ${\cal I}_{\rm bulk}$ and ${\cal I}_{\rm surf}$ given by: $$\begin{aligned} {\cal I}_{\rm bulk}&=&-\frac{V_{4}\pi^3}{2\kappa_{10}^2}\int {\cal L}_{\rm{II B}} \, d\sigma\ , \\ {\cal I}_{\rm surf}&=&-\frac{V_{4}\pi^3}{\kappa_{10}^2}\sqrt{\gamma}K \, \ ,\label{surf}\end{aligned}$$ where ${\cal I}_{{\rm IIB}}$ is the wick rotated action (\[genact\]) and in (\[surf\]) is the standard Gibbons-Hawking term. In fact one can check that the one dimensional effective lagrangian ${\cal L}_{1d}$ defined in (\[L-effective1\]) already includes this boundary term. Therefore we can write: $${\cal I}={\cal I}_{\rm bulk}+{\cal I}_{\rm surf}=-\frac{V_{4}\pi^3}{2\kappa_{10}^2}\int {\cal L}_{1d}(B,B',J',\psi_a,\psi_a') \, d\sigma\ ,$$ where $\psi_a$ is a collective notation for the background functions $(b,h,F,S,\Phi)$. Note that: $$\frac{\delta {\cal I}}{\delta J}\propto\frac{\partial {\cal L}_{1d}}{\partial J'}\propto H_*\ ,{\label{MagJ}}$$ where we have used (\[defJ\]). Given that we make the identification ${\cal I}/V_4={\cal F}/V_3$, where $\cal F$ refers to the Helmholtz free energy and we have that $\frac{\delta{\cal F}}{\delta M}=H_*$, it is natural to relate the field $J$ to the magnetization $M$ of the system. However we are interested in a thermodynamic ensemble in which we keep the external magnetic field $H_*$ fixed. Therefore the proper thermodynamic potential is given by the Gibbs free energy ${\cal G}\equiv{\cal F}-M H_*$. Equation (\[MagJ\]) implies: $${\cal G} \, = \, \frac{V_3}{V_4} \, \tilde{\cal I} \, = \, - \, \frac{V_{3}\pi^3}{2\kappa_{10}^2}\int \tilde{\cal L}_{1d}(B,B',H_*,\psi_a,\psi_a') \, d\sigma\ ,\label{Gibbsfree}$$ were $\tilde{\cal L}_{1d}$ refers to the Legendre transform of ${\cal L}_{1d}$ defined in (\[L-effective2\]). Next we expand in $\epsilon_*$ using (\[expansion1\]) and define dimensionless variables along . To first order one has: $$\tilde{\cal I}={\cal I}_0+\epsilon_*{\cal I}_{\rm DBI}+\epsilon_*{\cal I}_{\rm bound}+O(\epsilon_*^2)\ ,$$ and more explicitly: $$\begin{aligned} -\frac{2\kappa_{10}^2}{V_{3}\pi^3}{\cal G}&=&6r_*^4-4\epsilon_*r_m^4\int\limits_{0}^{\tilde r_*}d\tilde r\left[-\frac{d}{d\tilde r}\left(\tilde r^4\, \eta(\tilde r)\right)+\tilde r\sqrt{1+\tilde r^4}\tilde\xi(\tilde r)\right]={\label{Gibbs1}}\\ &=&\nonumber6r_*^4+4\epsilon_*r_*^4\eta(r_*)-4\epsilon_*\,r_m^4\,\int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r\tilde r\,\sqrt{1+\tilde r^4}\tilde\xi(\tilde r)+O(\epsilon_*^2)\ .\nonumber\end{aligned}$$ The first two terms in (\[Gibbs1\]) can be cancelled by appropriate counter terms (we refer the reader to the next section for more details), while the last term is precisely the DBI-term describing a probe D7–brane. Therefore to first order in $\epsilon_*\sim N_f/N_c$ the Gibbs free energy of the unquenched system coincides with the free energy calculated in the quenched approximation. This feature has been observed also in the finite temperature case studied in [@arXiv:0909.2865]. This suggests that to first order in $\epsilon_*$ one cannot study the effect of the dynamically generated energy scale set by the Landau pole of the dual gauge theory. Therefore we need to compute the second order contribution to the Gibbs free energy[^7]. Furthermore, despite the fact that we are dealing with a finite cut off set by the parameter $r_*$ we can still perform a holographic renormalization of the free energy resulting in an expression finite in the large $r_*$ limit. Holographic renormalization --------------------------- In this subsection we will regularize the Gibbs free energy defined in (\[Gibbsfree\]). One can show, that modulo logarithmic divergences, the counter terms needed to regulate (\[Gibbsfree\]) are the same as ones needed to regulate the supersymmetric background corresponding to vanishing magnetic field (the $r_m\to 0$ limit). Therefore, it is natural to regulate the on-shell action in (\[Gibbs1\]) by subtracting the supersymmetric on-shell action. Note that this would suggest regulating the DBI contribution to the free energy by subtracting the DBI term for the supersymmetric case. Such an approach would make comparison to the results obtained in the quenched approximation difficult, where the on-shell action is regulated by the addition of the appropriate boundary terms. Furthermore, logarithmic divergences depending on the magnetic field do not have analogues in the supersymmetric background. To rectify this we consider a mixed regularization scheme: we regulate logarithmic and divergences due to the DBI term by adding appropriate covariant boundary counter terms, while the rest we directly cancel by subtracting the relevant part of the supersymmetric action. ### The supersymmetric case In the supersymmetric limit of vanishing magnetic field, equations (\[EOM-chi\]) and (\[diff-b\])-(\[diff-H\]) can be integrated to give a BPS system of first order differential equations[^8]. The full non-perturbative solution for the corresponding supersymmetric background was obtained in [@arXiv:0807.0298] and for more details we refer the reader to the appendix \[app-susy\]. Here we present the expansion to second order in $\epsilon_*$ of the supersymmetric Euclidean “on-shell action" and analyze its divergences. We obtain: $$\begin{aligned} -\frac{2\kappa_{10}^2}{V_{4}\pi^3}{\cal I}_{\rm susy}&=&6r_*^4+4\epsilon_*\left[r_*^4\eta(r_*)-\int\limits_{ m_0}^{r_*}dr\,r^3\, \cos^3\frac{\chi_0}{2}\sqrt{1+\frac{r^2}{4}\chi_0'^2}\right]+\label{SUSYaction}\\ &+& \epsilon_*^2\left[ 4r_*^4\eta_2(r_*)-\frac{r^5\cos^3\frac{\chi_0}{2}\chi_0'}{\sqrt{1+\frac{r^2}{4}\chi_0'^2}}\chi_1\Big|_{r=r_*}-\frac{4}{5}r_*^4\left(4{\Delta_1^s(r_*)}^2+\Lambda_1^s(r_*)^2\right)\right .+\nonumber\\ &+&\left.\frac{4}{5}r_*^5\left(\Lambda_1^s\,{\Lambda_1^s}'-\Delta_1^s\,{\Delta_1^s}'\right)\Big|_{r=r_*}-\frac{1}{2}\int\limits_{m_0}^{r_*}drr^3\,\cos^8\frac{\chi_0}{2}-\frac{1}{2}\int\limits_{m_0}^{r_*}drr^5J_{\Phi}^s\Phi^s+\right. \nonumber\\ &+&\left. \frac{4}{5}\int\limits_{m_0}^{r_*}drr^5\left(J_{\Lambda_1}^s\Lambda_1^s-J_{\Delta_1^s}\Delta_1^s\right) \right]+O(\epsilon_*^3)\ .\nonumber\end{aligned}$$ As we pointed out above the divergences of the action (\[SUSYaction\]) completely cancel (modulo logarithmic terms) those of the non-supersymmetric action. However we would like to regulate together the DBI term in the first order expansion and the term depending on $\chi_1$ in the second order expansion (coming from expanding $\chi=\chi_0+\epsilon_*\chi_1$ in the DBI term) by adding covariant counter-terms. We make this choice in order for our regularization scheme to be compatible with the one employed in [@Albash:2007bk], for the probe approximation. Furthermore, by construction, boundary terms which do not contain derivatives of the fields are exactly the same with those of the non-supersymmetric action and can be directly subtracted. We define a subtracting action: $${\cal I}_{\rm subt}\equiv {\cal I}_{\rm subt}^{\rm bdr}+{\cal I}_{\rm subt}^{\rm bulk}\ ,\label{Isubt}$$ where: $$\begin{aligned} \frac{2\kappa_{10}^2}{V_{4}\pi^3}{\cal I}_{\rm subt}^{\rm bdr}&\equiv&6r_*^4+4\epsilon_* r_*^4\eta(r_*)+\epsilon_*^2\left[ 4r_*^4\eta_2(r_*)-\frac{4}{5}r_*^4\left(4{\Delta_1^s(r_*)}^2+\Lambda_1^s(r_*)^2\right)\right ]\ ,\\ \frac{2\kappa_{10}^2}{V_{4}\pi^3}{\cal I}_{\rm subt}^{\rm bulk}&\equiv&\epsilon_*^2\left[ \frac{4}{5}r_*^5\left(\Lambda_1^s\,{\Lambda_1^s}'-\Delta_1^s\,{\Delta_1^s}'\right)\Big|_{r=r_*}-\frac{1}{2}\int\limits_{m_0}^{r_*}drr^3\,\cos^8\frac{\chi_0}{2}-\frac{1}{2}\int\limits_{m_0}^{r_*}drr^5J_{\Phi}^s\Phi^s+\right. \nonumber\\ &+&\left. \frac{4}{5}\int\limits_{m_0}^{r_*}drr^5\left(J_{\Lambda_1}^s\Lambda_1^s-J_{\Delta_1^s}\Delta_1^s\right) \right]=-\epsilon_*^2\frac{(r_*^2-m^2)^5(11r_*^2+7m^2)}{162r_*^8}\ ,\end{aligned}$$ and we have explicitly evaluated ${\cal I}_{\rm subt}^{\rm bulk}$. Next we proceed with the regularization of: $$\begin{aligned} \frac{2\kappa_{10}^2}{V_{4}\pi^3}({\cal I}_{\rm susy}+{\cal I}_{\rm subt})&=&4\epsilon_*\int\limits_{ m_0}^{r_*}dr\,r^3\, \cos^3\frac{\chi_0}{2}\sqrt{1+\frac{r^2}{4}\chi_0'^2}+\epsilon_*^2\frac{r^5\cos^3\frac{\chi_0}{2}\chi_0'}{\sqrt{1+\frac{r^2}{4}\chi_0'^2}}\chi_1\Big|_{r=r_*}+{\cal O} (\epsilon_*^3) \nonumber \\ &=&4\epsilon_*\int\limits_{ m_0}^{r_*}dr\,r^3\, \cos^3\frac{\chi}{2}\sqrt{1+\frac{r^2}{4}\chi'^2} + {\cal O} (\epsilon_*^3) \, . \label{susysubtr}\end{aligned}$$ Note that in terms of the full function $\chi(r)$ the subtracted supersymmetric action looks just like the DBI term of the probe approximation. However, the function $\chi(r)$ is not the same (namely $2\arcsin(m/r) $). #### The $\kappa$-symmetry condition for the embedding can be solved in a new radial coordinate $\hat r(r)$: $$\hat r(r)=r\left(1+\epsilon_*\rho_1(r)+O(\epsilon_*^2)\right)\ .$$ where $\rho_1(r)$ is given in (\[rho1\]), with $\rho_1(r_*)=0$ and hence $\hat r(r_*)=r_*$. It is in this new radial coordinate that the fiducial embedding satisfies: $$\chi(\hat r)=2\arcsin\frac{\hat m}{\hat r}\ ,$$ where $\hat m=\hat r(m_0)$. In fact we identify $\hat m$ as the full bare mass parameter $\hat m\equiv\hat r(m_0)=m_0+\epsilon_*m_1+O(\epsilon_*^2)$. Clearly this defines $m_1$ in terms of $m_0$. Furthermore, since $\hat r(r_*)=r_*$ one can write: $$\chi(r_*)=\frac{2m}{r_*}+ {\cal O}\left(\frac{1}{r_*^3}\right)=\chi_0(r_*)+\epsilon_*\chi_1(r_*)+ {\cal O}\left(\epsilon_*^2\right)=\frac{2m_0}{r_*}+\epsilon_*\chi_{1}(r_*)+ {\cal O}\left(\frac{1}{r_*^3}\right)+ {\cal O}(\epsilon_*^2)\ ,$$ suggesting the natural relation between $m_1$ and the leading term in the expansion of $\chi_1(r_*)$: $$\chi_1(r_*)=\frac{2m_1}{r_*}+ {\cal O}\left(\frac{1}{r_*^3}\right)\ . \label{bare mass1}$$ Equation (\[bare mass1\]) is an alternative definition of the first correction $m_1$ to the bare mass, which does not require the existence of a special radial variable $\hat r$. We will use this definition when we study the theory at finite magnetic field. The fact that at $r_*$ the fiducial embedding still has the same expansion as in the quenched approximation suggests that one can use the same counter terms. Indeed one can check that the following counter terms: $$\frac{-2\kappa_{10}^2}{R^4V_{4}\pi^3}\frac{{\cal I}_{\rm count}}{\sqrt{\gamma}}=\epsilon_*\left(1-\frac{\chi^2}{2}+\frac{5\chi^4}{48}\right)_{r_*}=\epsilon_*\left(1-\frac{\chi_0^2}{2}+\frac{5\chi_0^4}{48}\right)_{r_*}+\epsilon_*^2\left(-{\chi_0}\chi_1+\frac{5\chi_0^3\chi_1}{12}\right)_{r_*}+O(\epsilon_*^3)\ ,\label{Icount}$$ completely cancel the first order term in (\[susysubtr\]) and leaves only sub-leading terms at second order in $\epsilon_*$. More precisely: $${\cal I}_{\rm susy}+{\cal I}_{\rm subt}+{\cal I}_{\rm count}=0+\epsilon_*^2\left[0+ {\cal O} \left(\#(m)\frac{\log r_*}{r_*^2}\right)\right]+{\cal O}\left(\epsilon_*^3\right)\ .\label{susyregreg}$$ If the theory were UV complete the addition of the regulating terms ${\cal I}_{\rm subt}$ and ${\cal I}_{\rm count}$ to the “on-shell" action would provide a vanishing mass dependent expression for the free energy of the supersymmetric background. However, due to the existence of a Landau pole our UV cutoff $r_*$ is finite. In general this can lead to spurious mass dependence of the free energy (and hence spurious condensate) due to the $\#(m)$ term in (\[susyregreg\]). Fortunately a similar contribution from sub-leading terms is present in the non-supersymmetric (finite magnetic field) case too. One can verify that the two contributions are approximately equal for large bare masses, which is also the regime when these contributions are significant. Therefore, ignoring sub-leading terms is a very good approximation, because they can always be canceled by a redefinition of ${\cal I}_{\rm subt}$. In the next subsection we will regularize the free energy at finite magnetic field by adding ${\cal I}_{\rm subt}+{\cal I}_{\rm count}$ to the action $\tilde {\cal I}$. ### Regularization at finite magnetic field In this subsection we regularize the “on-shell" action $\tilde {\cal I}$ by adding the term ${\cal I}_{\rm subt}$ defined in (\[Isubt\]) and the counter terms ${\cal I}_{\rm count}$ defined in (\[Icount\]). It is convenient to split the contributions to $\tilde{\cal I}$ into two parts $\tilde{\cal I}=\tilde {\cal I}_{\chi}+\tilde {\cal I}_{\rm bulk}$, regulated by ${\cal I}_{\rm count}$ and ${\cal I}_{\rm subt}$ respectively. We find: $$\begin{aligned} &&\frac{2\kappa_{10}^2}{V_{4}\pi^3}(\tilde{\cal I}_{\rm bulk}+{\cal I}_{\rm subt})=\epsilon_*^2r_m^4\left[\int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r\tilde r^5 \left[\frac{1}{4}J_{b_1}b_1+\frac{1}{2}J_{\lambda_1}\lambda_1-\frac{4}{5}\left(J_{\Lambda_1}\Lambda_1-J_{\Delta_1}\Delta_1\right)+\frac{1}{2}J_{\Phi}\Phi\right] +\right.\label{actionbulk} \\ &&\left. +\frac{1}{2}\int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r\tilde r\left[J_{H_1}H_1+\left(\tilde r^2+\frac{1}{\tilde r^2}\right)\cos^8\frac{\chi_0}{2}\right]- \frac{4}{5}\tilde r_*^5\left(\Lambda_1\,{\Lambda_1}'-\Delta_1\,{\Delta_1}'\right)\Big|_{\tilde r=\tilde r_*}\right]-\epsilon_*^2\frac{(r_*^2-m_0^2)^5(11r_*^2+7m_0^2)}{162r_*^8}\ .\nonumber\end{aligned}$$ The expression in equation (\[actionbulk\]) can be evaluated numerically and analytically in the limit $r_m\ll m_0$ (weak magnetic field). In the limit of weak magnetic field one finds that the subtracting term regulates all divergences. We will study this in more details in the next subsection. Let us now focus on the remaining part of the action: $$\begin{aligned} \frac{2\kappa_{10}^2}{V_{4}\pi^3}(\tilde{\cal I}_{\rm \chi}+{\cal I}_{\rm count})&=&\epsilon_*\left[4\,r_m^4\int\limits_{ \tilde r_q}^{\tilde r_*}d\tilde r\,\tilde r\,\sqrt{1+\tilde r^4}\, \cos^3\frac{\chi_0}{2}\sqrt{1+\frac{\tilde r^2}{4}\chi_0'^2}-\tilde r_*^4\left(1-\frac{\chi_0^2}{2}+\frac{5\chi_0^4}{48}\right)_{\tilde r_*}\right]+\nonumber\\ &+&\epsilon_*^2r_m^4\left[\frac{\tilde r^3\sqrt{1+\tilde r^4}\cos^3\frac{\chi_0}{2}\chi_0'}{\sqrt{1+\frac{\tilde r^2}{4}\chi_0'^2}}\Big|_{\tilde r=\tilde r_*}+\tilde r_*^4\left(-{\chi_0}+\frac{5\chi_0^3}{12}\right)_{r_*}\right]\chi_1(\tilde r_*)\ .\end{aligned}$$ The first term in the first order contribution to the action ${\cal I}_{\rm \chi}+{\cal I}_{\rm count}$ is the same as in the quenched approximation and hence one needs to complete the counter term action to subtract the $\propto r_m^4\log r_*$ divergence. The complete counter term action is: $$\frac{2\kappa_{10}^2}{V_{4}\pi^3}\tilde {\cal I}_{\rm count}=\frac{2\kappa_{10}^2}{V_{4}\pi^3}{\cal I}_{\rm count}-\epsilon_*\,B_{\alpha\beta}B^{\alpha\beta}\Big |_{r_*}\log\frac{r_*}{R}\ .$$ Now using the fact that $\chi_0(\tilde r_*)=2\tilde m_0/\tilde r_*+2\tilde c_0/\tilde r_*^3+O(1/\tilde r_*^5)$ one obtains: $$\frac{2\kappa_{10}^2}{V_{4}\pi^3}(\tilde{\cal I}_{\rm \chi}+\tilde{\cal I}_{\rm count})=4\epsilon_*\,r_m^4\,\left[\tilde {\cal I}_{\rm D7}(\tilde m_0)+\epsilon_*\left(-2\tilde c_0\,\tilde m_1+O\left(\frac{\log \tilde r_*}{\tilde r_*^2}\right)\right)\right]+O(\epsilon_*^3)\ ,\label{Ichireg}$$ where we have defined $\tilde {\cal I}_{\rm D7}(\tilde m)$ and $\tilde m_1$ via the following relations: $$\begin{aligned} \tilde{\cal I}_{\rm D7}(\tilde m_0)&=&\int\limits_{\tilde r_q}^{\tilde r_*}d\,\tilde r\, \tilde r \sqrt{1+\tilde r^4}\tilde\xi(\tilde r)-\frac{\tilde r_*^4}{4}\left(1-\frac{\chi_0^2}{2}+\frac{5\chi_0^4}{48}\right)_{\tilde r_*}-\frac{1}{2}\log\frac{r_*}{R}\ ,\label{ID7} \\ \chi_1(\tilde r_*)&=&\frac{2\tilde m_1}{\tilde r_*}+O\left(\frac{\log\tilde r_*}{\tilde r_*^3}\right)\ .\end{aligned}$$ The function $\tilde {\cal I}_{\rm D7}(\tilde m_0)$ is proportional to the expression for the free energy of the fundamental matter in the quenched approximation. It is known that (see for example [@Mateos:2007vn]) ${d\tilde{\cal I}_{\rm D7}(\tilde m_0)}/{d\tilde m_0}=-2\tilde c_0$. Using this relation (\[Ichireg\]) can be written as: $$\begin{aligned} \frac{2\kappa_{10}^2}{V_{4}\pi^3}(\tilde{\cal I}_{\rm \chi}+\tilde{\cal I}_{\rm count})&=&4\epsilon_*\,r_m^4\,\left[\tilde {\cal I}_{\rm D7}(\tilde m_0)+\tilde{\cal I}_{\rm D7}'(\tilde m_0)(\epsilon_*\, \tilde m_1)+\epsilon_*\left(O\left(\frac{\log \tilde r_*}{\tilde r_*^2}\right)\right)\right]+O(\epsilon_*^3)= \nonumber\\ &=&4\epsilon_*\,r_m^4\,\tilde {\cal I}_{\rm D7}(\tilde m)+\epsilon_*^2\left(O\left(\frac{\log \tilde r_*}{\tilde r_*^2}\right)\right)+O(\epsilon_*^3)\ ,\end{aligned}$$ where we have defined the complete bare mass parameter as $m=m_0+\epsilon_*\,m_1+O(\epsilon_*^2)$. Note that modulo sub-leading terms the only contribution from the $\epsilon_*^2$ term in (\[Ichireg\]) is the one needed to complete the argument of $\tilde {\cal I}_{\rm D7}$. Now we can write our final expression for the regularized action $\tilde {\cal I}_{\rm reg}\equiv\tilde{\cal I}+\tilde{\cal I}_{\rm subt}+\tilde{\cal I}_{\rm count}$: $$\frac{2\kappa_{10}^2}{V_{4}\pi^3}\tilde{\cal I}_{\rm reg}=4\epsilon_*\,r_m^4\,\tilde {\cal I}_{\rm D7}(\tilde m)+\epsilon_*^2\,r_m^4\tilde {\cal I}_{\rm D7}^{(2)}(\tilde m)+O(\tilde\epsilon_*^3)\ ,\label{regAction}$$ where the quantity $\tilde {\cal I}_{\rm D7}^{(2)}(\tilde m)$ is given by the right-hand side of (\[actionbulk\]) divided by $\epsilon_*^2\, r_m^4$ and we have completed its argument to a full bare mass $m$. We are now ready to proceed with the evaluation of the free energy and the fundamental condensate as a function of the bare mass. We will study first the limit of weak magnetic field $(\tilde m\gg 1)$, which can be analyzed analytically. Free energy and condensate at weak magnetic field ------------------------------------------------- In this section we calculate analytically the Gibbs free energy ${\cal G}$ and the fundamental condensate of the theory in the limit of weak magnetic field. Given that $H_*\propto r_m^2$ and $\tilde m=m/r_m$ this corresponds to the $1/\tilde m$ expansion in equation (\[regAction\]). It has been shown in [@Albash:2007bk] that: $$\tilde{\cal I}_{\rm D7}(\tilde m)=-\frac{1}{2}\log\tilde m+{\rm const_1}+O(1/\tilde m^4)\ .$$ Furthermore, using the approximate analytic solution for the background (see Appendix \[WMF\]) one can show that: $$\tilde{\cal I}_{\rm D7}^{(2)}(\tilde m)={\rm const_2 }+O\left(\frac{\log (\tilde r_*/\tilde m)}{\tilde m^4}\right)\ .\label{Actsecweak}$$ Note that, since we are dealing with a finite cutoff, numerically, the logarithmic term in (\[Actsecweak\]) is of order one and hence the sub-leading term is of order $1/\tilde m^4$. All together we obtain: $$\frac{2\kappa_{10}^2}{V_{4}\pi^3}\tilde{\cal I}_{\rm reg}(\tilde m)=-2\epsilon_*(\,r_m^4\log\tilde m+{\rm const})+O(\epsilon_*^3)\propto -\epsilon_*[2H_*^2\log m +\#( H_*)]+O(\epsilon_*^3)\ ,$$ where we have ignored terms of order $(1/\tilde m^4)$. For the leading order contribution to the free energy we recover the result from the quenched approximation: $$\frac{1}{V_3}{{\cal G}(H_*,m)}=-\epsilon_*\frac{\pi^3}{\kappa_{10}^2}r_m^4\log\,m+O(\epsilon_*^3)\ ,$$ where we have skipped the mass independent term $\#(H_*)$. Now we can define the fundamental condensate via: $${\langle\bar\psi\psi\rangle}\equiv\frac{1}{V_3}\frac{\partial{\cal G}(H_*,m)}{\partial m}=-\epsilon_*\,\frac{\pi^3}{\kappa_{10}^2}\frac{r_m^4}{m}+O(H_*^4)+O(\epsilon_*^3)\ .\label{condWeak}$$ Equation (\[condWeak\]) suggests that to order $\epsilon_*^3$ the fundamental condensate is the same as in the quenched approximation. However, as commented in section 4.6, the relevant perturbative parameter which takes into account the running of the effective coupling is $\epsilon_q$. Using (\[epsstar\]) we arrive at the following result for the condensate: $$\frac{\kappa_{10}^2}{\pi^3}{\langle\bar\psi\psi\rangle}=-\epsilon_q\,\frac{r_m^4}{m}\left(1+\epsilon_q\left(\frac{3}{4}+\log\frac{r_*}{m} \right)\right)+O(\epsilon_q^3)\ ,$$ where we have used that for $\tilde r_q\gg1$ to leading order one has: $$\Phi_1(\tilde r_q)\approx-\frac{3}{4}-\log\frac{r_*}{r_q}\approx-\frac{3}{4}-\log\frac{r_*}{m}\ .$$ Therefore our conclusion is that at first order in $\epsilon_q$ the result is the same as in the quenched approximation. However, at second order in $\epsilon_q$ we observe a logarithmic running of the condensate as a function of the bare mass and the energy scale set by the finite UV cutoff $\Lambda_{\rm UV}\sim r_*$, which reflects the positive beta function of the theory. We proceed with calculating the Gibbs free energy and the fundamental condensate at strong magnetic field. Free energy and condensate at strong magnetic field --------------------------------------------------- In this subsection we calculate the Gibbs free energy and the condensate of the theory at strong magnetic field evaluating numerically the functions $\tilde{\cal I}_{\rm D7}(\tilde m)$ and $\tilde{\cal I}_{\rm D7}^{(2)}(\tilde m)$. As commented above the first function describes the Gibbs free energy to first order in $\epsilon_*$ and has been studied in [@Erdmenger:2007bn; @Albash:2007bk]. Apparently it also describes the first order contribution to the fundamental condensate given by: $$\frac{\kappa_{10}^2}{2\pi^3r_m^3}\langle\bar\psi\psi\rangle= \epsilon_*\tilde{\cal I}_{\rm D7}'(\tilde m)+O(\epsilon_*^2)=-2\epsilon_*\,\tilde c_0(\tilde m)+O(\epsilon_*^2)\ .$$ In figure \[fig:fig1\] we present plots of $\tilde{\cal I}_{\rm D7}(\tilde m)$ and $-\tilde c_0(\tilde m)$. ![Plots of $\tilde{\cal I}_{\rm D7}$ and $\tilde c_0$ versus the bare mass parameter $\tilde m$.[]{data-label="fig:fig1"}](ID7.pdf "fig:"){width="3.4in"} ![Plots of $\tilde{\cal I}_{\rm D7}$ and $\tilde c_0$ versus the bare mass parameter $\tilde m$.[]{data-label="fig:fig1"}](cond0.pdf "fig:"){width="3.4in"} As one can see from the plot of $\tilde{\cal I}_{\rm D7}(\tilde m)$ for sufficiently small $\tilde m$ the theory has multiple phases. However only the lowest branch is thermodynamically stable and studies of the meson spectrum in the quenched approximation [@Filev:2007qu] verified this stability. In the second plot of Figure [\[fig:fig1\]]{} we see the condensate, but only for the stable phase. At vanishing bare mass the theory has a negative condensate, while for large bare masses the condensate has the $-1/(4\tilde m)\propto H_*^2/m$ dependence governing the weak magnetic field limit. #### Our next task is to obtain similar plots for the second order contribution in $\epsilon_*$ to the Gibbs free energy and the fundamental condensate. We will restrict ourselves in studying only the stable (to first order in $\epsilon_*$) phase of the theory. Using the numerical solution for the background we study $\tilde{\cal I}_{\rm D7}^{(2)}(\tilde m)$ and its derivative. Note that (\[regAction\]) and the definition of the condensate: $$\langle\bar\psi\psi\rangle\equiv\frac{1}{V_3}\frac{\partial{\cal G}(H_*,m)}{\partial m}=\frac{1}{V_4}\frac{\partial{\tilde{\cal I}}_{\rm reg}(H_*,m)}{\partial m}\ ,$$ imply the relation: $$\frac{\kappa_{10}^2}{2\pi^3r_m^3}\langle\bar\psi\psi\rangle= \epsilon_*\tilde{\cal I}_{\rm D7}'(\tilde m)+\frac{1}{4}\epsilon_*^2{\tilde{\cal I}_{\rm D7}^{(2)}}{}'(\tilde m)+O(\epsilon_*^3)=-2\epsilon_*\,\tilde c_0(\tilde m)-2\epsilon_*^2\tilde c_2(\tilde m)+O(\epsilon_*^3)\ ,\label{condfull}$$ where $\tilde c_2$ is defined by ${\tilde{\cal I}_{\rm D7}^{(2)}}{}'(\tilde m)=-8\tilde c_2(\tilde m) $. ![Plots of $\tilde{\cal I}_{\rm D7}^{(2)}$ and $-\tilde c_2$ versus the bare mass parameter $\tilde m$.[]{data-label="fig:fig2"}](ID7_2.pdf "fig:"){width="3.4in"} ![Plots of $\tilde{\cal I}_{\rm D7}^{(2)}$ and $-\tilde c_2$ versus the bare mass parameter $\tilde m$.[]{data-label="fig:fig2"}](cond2.pdf "fig:"){width="3.4in"} The corresponding plots are presented in figure \[fig:fig2\] and from these one can see that $\tilde c_2$ approaches zero much faster than $\tilde c_0$ in Figure \[fig:fig1\]. Surprisingly the second order contribution to the fundamental condensate is positive. However the relevant perturbative parameter is $\epsilon_q$ (defined in section 2.6) therefore in order to obtain the second order correction to the condensate we need to trade $\epsilon_*$ for $\epsilon_q$ in (\[condfull\]). Using (\[epsstar\]) and the fact that $\Phi_1(\tilde r_q)<0$ (see Figure \[fig:fig0\]) we obtain: $$\begin{aligned} \frac{\kappa_{10}^2}{2\pi^3r_m^3}\langle\bar\psi\psi\rangle&=& -2\epsilon_q\,\tilde c_0(\tilde m)-2\epsilon_q^2(\tilde c_2(\tilde m)+|\Phi(\tilde r_q)|\tilde c_0(\tilde m ))+O(\epsilon_q^3)\label{condfullQ}\\ &=& -2\epsilon_q\,\tilde c_0(\tilde m)-2\epsilon_q^2(\hat c_2(\tilde m))+O(\epsilon_q^3)\ , \nonumber\end{aligned}$$ where we have defined: $$\hat c_2(\tilde m)\equiv\tilde c_2(\tilde m)+|\Phi(\tilde r_q)|\tilde c_0(\tilde m )\ .$$ As one can see from figure \[fig:fig0\]$, |\Phi(\tilde r_q)|>1$. Furthermore, comparing Figure \[fig:fig1\] and Figure \[fig:fig2\] one can see that $|\tilde c_0(\tilde m)|>|\tilde c_2(\tilde m)|$. Therefore we conclude that $-\hat c_2(\tilde m)<0$ and the second order correction to the condensate is negative. This implies that taking into account internal fundamental loops in the calculation of the condensate does not change qualitatively the results obtained in the quenched approximation. ![A plot of $-\hat c_2$ versus $\tilde m$.[]{data-label="fig:fig3"}](hatcond.pdf){width="7in"} In fact it is not difficult to obtain a plot of $\hat c_2(\tilde m)$. Our results are presented in Figure \[fig:fig3\]. One can see that indeed the second order contribution to the fundamental condensate is negative. The black dashed curve represents the function $\frac{1}{4\tilde m}(\frac{3}{4}+\log\frac{\tilde r_*}{\tilde m})$, which governs the asymptotic behavior of the condensate at large $\tilde m$ (weak magnetic field). Overall our conclusion is that to leading order in a perturbative expansion in the ratio $N_f/N_c$, the free energy and fundamental condensate of the theory agree with the results obtained in the quenched approximation. Furthermore, at second order in $N_f/N_c$ we observe that the effect of magnetic catalysis is enhanced and the contribution to the condensate of the theory from internal fundamental loops runs logarithmically with the finite cutoff $\Lambda_{UV}$. Summary & Conclusions ===================== Let us briefly summarize our conclusions and outline possible directions for future studies: #### In section 2 of this paper we have constructed a backreacted supergravity background holographically dual to an SU($N_c$) ${\cal N}=4$ SYM theory coupled to $N_f$ flavours of ${\cal N}=2$ hypermultiplets in the presence of external magnetic field $H_*$. Our solution is perturbative in a parameter which counts the number of internal fundamental loops. At small distances the geometry has a hollow cavity, where it is similar to the supergravity dual of non-commutative SYM theory. The radius of this cavity, $r_q$, sets the energy scale corresponding to the physical mass of the fundamental fields. From holographic point of view, the supergravity solution inside the cavity corresponds to the low energy effective field theory obtained after integrating out the massive flavour fields. The non-commutative nature of the theory reflects that fact that the fundamental fields are coupled to an external magnetic field, $H_*$. We have shown that the parameter of non-commutative along the plane perpendicular to the magnetic field, $\Theta^{23}$, scales as $\Theta^{23}\sim \frac{N_f}{N_c}\frac{1}{H_*}$. #### At large radial distances, our solution is well-approximated by the non-perturbative supersymmetric solution corresponding to a vanishing magnetic field [@arXiv:0807.0298]. We introduce a sufficiently large radial parameter ($r_*\gg r_q$) at which we match our perturbative solution to the supersymmetric one. In the holographically dual field theory, the radial parameter $r_*$ corresponds to a finite cutoff $\Lambda_{UV}$. At a given radial distance $r_{LP}$ ($r_{LP}\gg r_*$ ), the dilaton field diverges, reflecting the positive $\beta$-function of the dual gauge theory and the existence of a Landau pole at an energy scale, $\Lambda_{LP}$ ($\Lambda_{LP}\sim r_{LP}$). We have the following relation [@arXiv:0807.0298], $\epsilon_*\propto \log\frac{\Lambda_{LP}}{\Lambda_{UV}}$ implying that to keep the finite cutoff sufficiently far away from the Landau pole we need to keep the parameter $\epsilon_*\sim \frac{N_f}{N_c}$ sufficiently low. Clearly in the limit $\epsilon_*\to 0$ we can make the cutoff $\Lambda_{UV}$ arbitrarily large, on field theory side this corresponds to the quenched approximation. In fact our studies of the free energy and the fundamental condensate of the dual gauge theory confirm that at leading order in $\epsilon_*$, we reproduce the results of the quenched approximation. #### In section 3 we have applied our construction to the phenomenon of magnetic catalysis, i.e. of mass generation in an external magnetic field, beyond the quenched approximation. Note that our studies are at fixed value of the external magnetic field, $H_*$. This implies that the relevant thermodynamic potential is the Gibbs free energy of the dual gauge theory, as opposed to the Helmholtz free energy. In our holographic setup, the Gibbs free energy is given by the Legendre transform of the “on-shell" action with respect to the magnitude of the Ramond-Ramond two-form at $r_*$, $J(r_*)$. To calculate the free energy, we have developed an appropriate holographic renormalisation scheme. This scheme is designed to facilitate the comparison of our results to the results of ref. [@Filev:2007gb; @Erdmenger:2007bn; @Albash:2007bk] obtained in the quenched approximation. To this end we regularize the contribution to the free energy which due to the DBI term of the smeared flavour branes by the addition of appropriate covariant counterterms. The remaining contributions to the “on-shell” action are regularized by subtraction, using the supersymmetric solution corresponding to the limit of a vanishing external magnetic field [@arXiv:0807.0298] as a reference background. #### The application of our regularization scheme to the supersymmetric background [@arXiv:0807.0298] implies the identification of the bare mass parameter of the fundamental fields $m$ with the leading order coefficient in the large $r_*$ expansion of the profile of the fiducial embedding $\chi(r)$, namely $\chi(r_*)=2m/r_*+\dots$. In the limit $\epsilon_*\to 0$ ($r_*\to\infty$), this definition of the bare mass parameter $m$ agrees with the definition in the quenched approximation given in ref. [@Karch:2002sh] . Moreover we define the fundamental condensate of the theory as the derivative of the Gibbs free energy with respect to the bare mass. At first order in $\epsilon_*$, our results agree with the results obtained in the quenched approximation. At next order in $N_f/N_c$, the relevant perturbative parameter is the IR parameter $\epsilon_q$ defined via $\epsilon_q\equiv e^{\Phi(r_q)}/e^{\Phi_*}\epsilon_*\,$. Note that $\epsilon_q$ takes into account the running of the effective ’t Hooft coupling, encoded in the running of the dilaton field[^9]. Our results for the second order in $\epsilon_q$ contribution to the fundamental condensate show that the effect of magnetic catalysis is enhanced. Furthermore we observe a logarithmic running of the condensate as a function of the UV cutoff of the theory, reflecting the positive $\beta$-function of the theory. This is the main result of our study of the effect of magnetic catalysis beyond the quenched approximation. #### An obvious direction for future studies is to investigate the effect of finite temperature. From supergravity point of view, this corresponds to generalizing our perturbative construction by substituting the AdS$_5\times S^5$ zeroth order supersymmetric background with the AdS$_5\times S^5$ black hole. Such a study would enable us to study the phase diagram of the theory, which in the quenched approximation was analyzed in refs. [@Erdmenger:2007bn; @Albash:2007bk], beyond the quenched approximation. It would also be interesting to generalize our holographic setup by turning on a finite chemical potential and to investigate the effect of internal fundamental loops to the quantum critical points reported in refs. [@Evans:2010iy; @Jensen:2010vd]. #### Finally it would be interesting to generalize our holographic setup to other Dp/Dq–brane intersections and to apply our framework to related phenomena also taking place at finite magnetic field. For example, it would be interesting to construct the analogue of the effect of Inverse Magnetic Catalysis [@Preis:2010cq], observed in the Sakai-Sugimoto Model, for holographic gauge theories dual to the Dp/Dq–intersections and to take into account the backreaction of the flavour branes. Other effects of potential interest include the Chiral Magnetic Effect [@Kharzeev:2007jp], investigated holographically in refs. [@Rebhan:2009vc; @Gorsky:2010xu], and the recently proposed effect of superconductivity from $\rho$ meson condensation in the QCD vacuum in a strong magnetic field [@Chernodub:2010qx], investigated holographically in refs. [@Callebaut:2011ab; @Ammon:2011je]. Acknowledgments =============== We would like to thank M. Ammon, F. Bigazzi, A. L. Cotrone, K. Y. Kim, D. O’Connor, A. Paredes, R. Rashkov, J. Shock and J. Tarrio for useful comments and suggestions. Moreover we would like to thank J. Shock for critically reading the manuscript. D. Z. is funded by the FCT fellowship SFRH/BPD/62888/2009. Centro de Física do Porto is partially funded by FCT through the projects PTDC/FIS/099293/2008 & CERN/FP/109306/2009. The work of V. F. is funded by an INSPIRE IRCSET-Marie Curie International Mobility Fellowship. V. F. would like to thank the organizers of the GGI Workshop “Large-N Gauge Theories" in Florence and the Institute for Theoretical Physics, at Vienna University of Technology for hospitality during the early stages of this project. D. Z. thanks the gauge/gravity duality group at the Max-Planck-Institut for Physics in Munich for hospitality during the final stages of this work. The work of J. E. is partially funded by the ‘Excellence Cluster for Fundamental Physics: Origin and Structure of the Universe’. Analytic expressions for the calculation of the EOM and the Free Energy {#nonsusy-ingredients} ======================================================================= The analytic expressions for the sources appearing in and are $$\begin{aligned} J_{\lambda_1}(\tilde{r}) \, & = & \, \frac{4 \, {\tilde \xi}(\tilde{r}) }{{\tilde r}^4 \sqrt{{\tilde r}^4 \, + \, 1}} \, , \quad \quad J_{b_1}(\tilde{r}) \, = \, -\frac{8 \, {\tilde\xi}(\tilde{r})}{{\tilde r}^4 \sqrt{{\tilde r}^4 \, + \, 1}} \, , \nonumber \\ \nonumber \\ J_{\Phi_1}(\tilde{r}) \, & = & \, 8 \, {\tilde \xi}(\tilde{r}) \, \, \frac{{\tilde r}^4 \,+\,\frac{1}{2}}{{\tilde r}^4 \sqrt{{\tilde r}^4 \, + \, 1}} \, , \quad \quad J_{\Lambda_1}(\tilde{r}) \, = \, - \, 8 \, {\tilde \xi}(\tilde{r}) \, \frac{\sqrt{1+\tilde{r}^4}}{\tilde{r}^4} \, \frac{1 \, + \, \frac{\tilde{r}^2}{8} \, \chi_{0}'(\tilde{r})^2}{1 \, + \, \frac{\tilde{r}^2}{4} \, \chi_{0}'(\tilde{r})^2} \, , \nonumber \\ \nonumber \\ J_{\eta_1}(\tilde{r}) \, & = & \, \frac{\sqrt{{\tilde r}^4 \, + \, 1}}{{\tilde r}^3} \, \frac{ {\tilde \xi}(\tilde{r})}{ 1 \, + \, \frac{\tilde{r}^2}{4} \, \chi_{0}'(\tilde{r})^2 } \, , \quad \quad J_{H_1}(\tilde{r}) \, = \, \frac{8 \, \tilde\xi(\tilde{r})}{\sqrt{{\tilde r}^4 \, + \, 1}} \, , \nonumber \\ \nonumber \\ J_{\Delta_1}(\tilde{r}) \, & = & \, {\tilde \xi}(\tilde{r}) \, \frac{\sqrt{1+\tilde{r}^4}}{\tilde{r}^4} \, \Bigg[ \, \frac{1}{{1 \, + \, \frac{\tilde{r}^2}{4} \, \chi_{0}'(\tilde{r})^2} } \, - \, \frac{3}{2} \, \left( 1 \, - \, \frac{5}{3}\, \cos \chi_{0}(\tilde{r}) \right) \Bigg] \, , \label{nonsusyJ}\end{aligned}$$ with $${\tilde\xi}(\tilde{r}) \equiv \cos^3 \frac{\chi_{0}(\tilde{r})}{2} \, \sqrt{1 \, + \, \frac{\tilde{r}^2}{4} \, \chi_{0}'(\tilde{r})^2} \, . \label{xi}$$ Useful ingredients from the massive supersymmetric solution {#app-susy} =========================================================== The massive supersymmetric solution was found in [@arXiv:0807.0298; @arXiv:0909.2865] and it is easy to check that the following set of first order differential equations $$\begin{aligned} && \partial_{\sigma} h \, = \, - \, Q_c \, , \quad \qquad \partial_{\sigma} F \, = \, S^4 \, F \Bigg[ 3-2 \, \frac{F^2}{S^2} - \frac{Q_f}{2} \, e^\Phi \, \cos^4\frac{\chi_{q}}{2} \, \Bigg] \, , \label{massiveBPS} \\ \nonumber \\ && \partial_{\sigma} S \, = \, S^3 \, F^2 \, , \quad\qquad \partial_{\sigma} \chi_{q} = -2 \, S^4 \, \tan \frac{\chi_{q}}{2} \, , \quad\quad \partial_{\sigma} \Phi \, = \, Q_f \, S^4 \, e^\Phi \cos^4\frac{\chi_{q}}{2} \, . \nonumber \end{aligned}$$ together with $b\, =\, 1$ & $H\, = \, 0$ solve the full set of equations of motion, & - . Changing coordinates from $\sigma$ to $\rho$, according to $d \rho = S^4 \, d \sigma$, we can integrate immediately the equation of motion for the embedding $$\sin\frac{\chi_{wv}}{2} \, = \, {e^{\rho_q} \over e^{\rho}} \label{susyembedding-rho}$$ where $\rho_q$ is an integration constant related to the bare quark mass. Substituting in we obtain the following set of explicit expressions for $S$, $F$, $\Phi$ & $h$ for $\rho > \rho_q$ $$\begin{aligned} && S \, = \, e^\rho \, \left[1 \, + \, \epsilon_* \, A_{S} \right]^\frac{1}{6} \, , \qquad \qquad \quad F \, = \, e^\rho\, \ , \frac{\left[1 \, + \, \epsilon_* \, A_{\Phi}\right]^{\frac{1}{2}}}{\left[1 \, + \, \epsilon_* \, A_{S} \right]^\frac{1}{3}} \, \label{susysolmassive} \\ \nonumber \\ && \Phi \, = \, \Phi_* \, - \, \log \left[ 1\, + \, \epsilon_* \, A_{\Phi} \right] \, , \qquad \partial_{\rho} h \, = \, - \, Q_c \, e^{-4\rho} \left[1 \, + \, \epsilon_* \, A_{S} \right]^\frac{2}{3}\ , \nonumber\end{aligned}$$ with $$\begin{aligned} A_{\Phi} \, & \equiv & \, \rho_*-\rho-e^{2\rho_q-2\rho} \, + \, \frac{1}{4} \, e^{4\rho_q-4\rho} \, + \, e^{2\rho_q-2\rho_*} \, - \, \frac{1}{4} \, e^{4\rho_q-4\rho_*}\ , \\ \nonumber \\ A_{S} \, & \equiv & \, \frac{1}{6} \, + \, \rho_* \, - \, \rho \, - \, \frac{1}{6} \, e^{6\rho_q-6\rho} \, - \, \frac{3}{2} \, e^{2\rho_q-2\rho} \, + \, \frac{3}{4} \, e^{4\rho_q-4\rho} \, - \, \frac{1}{4} \, e^{4\rho_q-4\rho_*} \, + \, e^{2\rho_q-2\rho_*} \, \ , \nonumber\end{aligned}$$ where $\rho_*$ is the UV scale where the dilaton takes the value $\Phi_*$ and $\epsilon_*=Q_f\,e^{\Phi_*}$. Following the analysis of the main part of the paper, we redefine the radial variable in such a way that the warp factor keeps the standard $AdS$ form $$h \, = \, \frac{R^4}{r^4} \, \qquad \text{with} \qquad R^4\equiv \frac{1}{4} Q_c\ ,$$ and expand $\partial_{\rho} h$ from in powers of $\epsilon_*$. Integrating, we obtain an expression for $r$ as a function of $\rho$, while we fix the integration constant by requiring $r(\rho_*) \, \equiv \, r_*\, = \, e^{\rho_*}$. Inverting that relation we have $$\rho \, = \, \log r \, + \, \epsilon_* \, \rho_1 \, , \label{rho0}$$ with $$\begin{aligned} && \rho_1 \, = \, \frac{1}{720} \Bigg[ 120 \log \frac{r}{r_*} \, + \, 10 \left(1 - \frac{r^4}{r_*^4}\right) \left(1 - 3 \, \frac{m_0^4}{r^4}\right)\, +\, 8 \, \frac{m_0^6}{r^6} \, \left(1 - \frac{r^{10}}{r_*^{10}}\right) \label{rho1} \\ \nonumber \\ && \qquad \qquad \qquad \qquad \qquad \qquad \,\,\,\, + \, 120 \, \, \frac{m_0^2}{r^2} \, \left(1 - \frac{r^2}{r_*^2}\right) \, - \, 15 \, \, \frac{m_0^4}{r^4} \left(1 \, - \, \frac{r^8}{r_*^8} \right) \Bigg] \, . \nonumber\end{aligned}$$ Substituting and in we have $$\begin{aligned} && S \, = \, r \left(1 \, + \, \epsilon_* \, S_1 \right) \, , \quad F\, = \, r \left( 1 \, + \, \epsilon_* \, F_1 \right) \, , \nonumber \\ \nonumber \\ && \Phi \, = \, \Phi_* \, + \, \epsilon_* \, \Phi_1 \, , \quad \& \quad \chi_q \, = \, 2\, \arcsin \frac{m_0}{r} \, + \epsilon_* \chi_1 \, ,\end{aligned}$$ with $$\begin{aligned} F_1 &=& - \, \frac{1}{24} \left(1 + \frac{1}{3} \, \frac{r^4}{r_*^4} \right) \, + \, \frac{1}{6} \, \frac{m_0^2}{r^2} \, - \, \frac{3}{16} \, \frac{m_0^4}{r^4} \left(1 - \frac{1}{9} \, \frac{r^8}{r_*^8} \right) \, + \frac{1}{15} \, \frac{m_0^6}{r^6} \left(1 - \frac{1}{6} \, \frac{r^{10}}{r_*^{10}} \right) \,, \\ \nonumber \\ S_1 &=& \frac{1}{24} \left(1 - \frac{1}{3} \, \frac{r^4}{r_*^4} \right) \, - \, \frac{1}{12} \, \frac{m_0^2}{r^2} \, + \, \frac{1}{16} \, \frac{m_0^4}{r^4} \left(1 + \frac{1}{3} \, \frac{r^8}{r_*^8} \right) \, - \frac{1}{60} \, \frac{m_0^6}{r^6} \left(1 + \frac{2}{3} \, \frac{r^{10}}{r_*^{10}} \right) \,, \\ \nonumber \\ \Phi_1 &=& \log\frac{r}{r_*} \, + \, \frac{m_0^2}{r^2} \left(1\, - \, \frac{r^2}{r_*^2}\right) \, - \, \frac{m_0^4}{4 \, r^4} \left(1\, - \, \frac{r^4}{r_*^4}\right) \, , \\ \nonumber \\ \chi_1 &=& \frac{1}{360} \, \frac{m_0}{\sqrt{r^2 -m_0^2}} \Bigg[ -120 \, \ln \frac{r}{m_0} \, + \, 8 \left(1 \, - \, \frac{m_0^6}{r^6} \right) \, - \, 45 \left(1 \, - \, \frac{m_0^4}{r^4} \right) \nonumber \\ \nonumber \\ && \qquad \quad \quad \, \, \, \, \, \, + \, 120 \left(1 \, - \, \frac{m_0^2}{r^2} \right) \, + \, \frac{r^4}{r_0^4} \left(1\, - \, \frac{m_0^4}{r^4}\right) \left(10 \, - \, 15 \, \frac{m_0^4}{r_0^4} \, + \, 8 \, \frac{m_0^6}{r_0^6} \right) \Bigg] \, . \label{susysolmassive-r}\end{aligned}$$ Since they are needed for the analysis of the free energy we explicitly construct $\Lambda_1$ & $\Delta_1$ $$\begin{aligned} \Delta_1 &=& \, \frac{1}{12} \, \frac{m_0^6}{r^6} \, \left(1 \, - \, \frac{r^2}{m_0^2} \right)^3 \,, \label{susy-Delta1} \\ \nonumber \\ \Lambda_1 &=& \frac{5}{72} \left(1 - \, \frac{r^4}{r_0^4} \right) \, - \, \frac{5}{48} \, \frac{m_0^4}{r^4} \, \left(1 - \, \frac{r^8}{r_0^8} \right) \, - \, \frac{1}{6} \, \frac{m_0^2}{r^2} \left(1 - \, \frac{m_0^2}{r^2} \right) \, + \, \frac{1}{18} \, \left(1 - \frac{m_0^6}{r^6} \, \frac{r^{10}}{r_0^{10}} \right) \, . \label{susy-Lambda1}\end{aligned}$$ Weak magnetic field {#WMF} =================== It is possible to perform analytic computations at the limit of weak magnetic field. The definition of the weak magnetic field that we will adopt is that the energy scale associated to the magnetic field $r_m$ is much smaller than the bare mass parameter $m$, namely $r_m\ll m_0$. In this limit to zeroth order in $\epsilon_*$ one has the following expansion for the profile of the fiducial embedding $\chi_0(r)$ and the bare mass parameter $m_0$ $$\begin{aligned} \chi_0(r) \, &=& \, 2 \arcsin\left(\frac{r_q}{r}\right) \, - \, \frac{\sqrt{r^2-r_q^2}}{2r_q^3r^2}r_m^4 \, + \, {\cal O}\left(r_m^6\right) \, , \\ m_0 \, &=& \, r_q-\frac{r_m^4}{4r_q^3} \, + \, {\cal O} \left(r_m^6\right )\nonumber \, .\end{aligned}$$ Expanding in $r_m$ we get $$\label{WMF-emb} \chi_0(r) \, = \, \frac{2 m_0}{r} \, + \, \frac{\frac{m_0^3}{3}\, +\, \frac{r_m^4}{2 m_0} }{r^3} \, + \, {\cal O}\left(r_m^6\right) \,$$ which is the expression that we will use to calculate . Using we can evaluate the expressions for all the functions of the background through , , , , & and then perform the different integrals in . After a long but straightforward calculation we obtain . Here we give some details on obtaining the contribution from $\Lambda_1$, since all the rest can be calculated in a similar manner. Using dimensionless variables we have $$\begin{aligned} J_{\Lambda_1}(\tilde{r}) \, & \sim & \, - \, \frac{8}{\tilde{r}^2} \, + \, \frac{12 \, \tilde{m}_0^2}{\tilde{r}^4} \,+\, \frac{2 - 4\tilde{m}_0^4 }{\tilde{r}^6} \, + \, {\cal O}(\tilde{r}^{-6}) \quad \& \\ \nonumber \int\limits_{\tilde r_q}^{\tilde r_*}d\tilde r\tilde r^5 J_{\Lambda_1}\Lambda_1 \, & \sim & \, - \, \frac{13 {\tilde r}_*^4 }{144} \, + \, \frac{23 {\tilde m}_0^2 {\tilde r}_*^2 }{36} \, + \, \frac{1}{4}\, \left( 1 \, - \, 6 {\tilde m}_0^4\right) \, \log \frac{{\tilde r}_*}{{\tilde m}_0} \, + \, {\cal O}(\tilde{r}_*^{-1}) \, .\end{aligned}$$ [99]{} J. M. Maldacena, Adv. Theor. Math. Phys.  [**2**]{}, 231 (1998) \[Int. J. Theor. Phys.  [**38**]{}, 1113 (1999)\] \[arXiv:hep-th/9711200\]. V. P. Gusynin, V. A. Miransky and I. A. Shovkovy, Phys. Rev. Lett.  [**73**]{}, 3499 (1994) \[Erratum-ibid.  [**76**]{}, 1005 (1996)\] \[arXiv:hep-ph/9405262\]. V. P. Gusynin, V. A. Miransky and I. A. Shovkovy, Phys. Lett.  B [**349**]{}, 477 (1995) \[arXiv:hep-ph/9412257\]. D. K. Hong, Y. Kim and S. J. Sin, Phys. Rev.  D [**54**]{}, 7879 (1996) \[arXiv:hep-th/9603157\]. K. G. Klimenko, Theor. Math. Phys.  [**89**]{}, 1161 (1992) \[Teor. Mat. Fiz.  [**89**]{}, 211 (1991)\]. K. G. Klimenko, Z. Phys.  C [**54**]{}, 323 (1992). K. G. Klimenko, Theor. Math. Phys.  [**90**]{}, 1 (1992) \[Teor. Mat. Fiz.  [**90**]{}, 3 (1992)\]. V. G. Filev, C. V. Johnson, R. C. Rashkov and K. S. Viswanathan, JHEP [**0710**]{}, 019 (2007) \[arXiv:hep-th/0701001\]. V. G. Filev, JHEP [**0804**]{}, 088 (2008) \[arXiv:0706.3811 \[hep-th\]\]. V. G. Filev and R. C. Raskov, Adv. High Energy Phys.  [**2010**]{}, 473206 (2010) \[arXiv:1010.0444 \[hep-th\]\]. J. Erdmenger, R. Meyer and J. P. Shock, JHEP [**0712**]{}, 091 (2007) \[arXiv:0709.1551 \[hep-th\]\]. J. Erdmenger, N. Evans, I. Kirsch and E. Threlfall, Eur. Phys. J.  A [**35**]{}, 81 (2008) \[arXiv:0711.4467 \[hep-th\]\]. T. Albash, V. G. Filev, C. V. Johnson and A. Kundu, JHEP [**0807**]{}, 080 (2008) \[arXiv:0709.1547 \[hep-th\]\]. C. V. Johnson and A. Kundu, JHEP [**0812**]{}, 053 (2008) \[arXiv:0803.0038 \[hep-th\]\]. A. V. Zayakin, JHEP [**0807**]{} (2008) 116 \[arXiv:0807.2917 \[hep-th\]\]. V. G. Filev, C. V. Johnson and J. P. Shock, JHEP [**0908**]{}, 013 (2009) \[arXiv:0903.5345 \[hep-th\]\]. V. G. Filev, JHEP [**0911**]{}, 123 (2009) \[arXiv:0910.0554 \[hep-th\]\]. N. Evans, A. Gebauer, K. Y. Kim and M. Magou, JHEP [**1003**]{}, 132 (2010) \[arXiv:1002.1885 \[hep-th\]\]. K. Jensen, A. Karch and E. G. Thompson, JHEP [**1005**]{}, 015 (2010) \[arXiv:1002.2447 \[hep-th\]\]. K. Jensen, A. Karch, D. T. Son and E. G. Thompson, Phys. Rev. Lett.  [**105**]{}, 041601 (2010) \[arXiv:1002.3159 \[hep-th\]\]. N. Evans, A. Gebauer, K. Y. Kim and M. Magou, Phys. Lett.  B [**698**]{}, 91 (2011) \[arXiv:1003.2694 \[hep-th\]\]. N. Evans, A. Gebauer and K. Y. Kim, JHEP [**1105**]{}, 067 (2011) \[arXiv:1103.5627 \[hep-th\]\]. S. Bolognesi and D. Tong, arXiv:1110.5902 \[hep-th\]. A. Karch and E. Katz, JHEP [**0206**]{}, 043 (2002) \[arXiv:hep-th/0205236\]. C. Nunez, A. Paredes and A. V. Ramallo, Adv. High Energy Phys.  [**2010**]{}, 196714 (2010) \[arXiv:1002.1088 \[hep-th\]\]. A. Paredes, JHEP [**0612**]{}, 032 (2006) \[arXiv:hep-th/0610270\].\ F. Benini, F. Canoura, S. Cremonesi, C. Nuñez and A. V. Ramallo, JHEP [**0709**]{}, 109 (2007) \[arXiv:0706.1238 \[hep-th\]\].\ R. Casero, C. Nuñez and A. Paredes, Phys. Rev.  D [**77**]{}, 046003 (2008) \[arXiv:0709.3421 \[hep-th\]\].\ E. Caceres, R. Flauger, M. Ihl and T. Wrase, JHEP [**0803**]{}, 020 (2008) \[arXiv:0711.4878 \[hep-th\]\].\ F. Canoura, P. Merlatti and A. V. Ramallo, JHEP [**0805**]{}, 011 (2008) \[arXiv:0803.1475 \[hep-th\]\].\ F. Bigazzi, A. L. Cotrone, C. Nuñez and A. Paredes, Phys. Rev.  D [**78**]{}, 114012 (2008) \[arXiv:0806.1741 \[hep-th\]\].\ C. Hoyos-Badajoz, C. Nuñez and I. Papadimitriou, Phys. Rev.  D [**78**]{}, 086005 (2008) \[arXiv:0807.3039 \[hep-th\]\].\ D. Arean, P. Merlatti, C. Nuñez and A. V. Ramallo, JHEP [**0812**]{}, 054 (2008) \[arXiv:0810.1053 \[hep-th\]\].\ J. Gaillard and J. Schmude, JHEP [**0901**]{}, 079 (2009) \[arXiv:0811.3646 \[hep-th\]\].\ A. V. Ramallo, J. P. Shock and D. Zoakos, JHEP [**0902**]{}, 001 (2009) \[arXiv:0812.1975 \[hep-th\]\].\ F. Bigazzi, A. L. Cotrone, A. Paredes and A. V. Ramallo, JHEP [**0903**]{}, 153 (2009) \[arXiv:0812.3399 \[hep-th\]\].\ F. Bigazzi, A. L. Cotrone, A. Paredes and A. V. Ramallo, JHEP [**0905**]{}, 034 (2009) \[arXiv:0903.4747 \[hep-th\]\].\ C. Nunez, M. Piai and A. Rago, Phys. Rev.  D [**81**]{}, 086001 (2010) \[arXiv:0909.0748 \[hep-th\]\].\ D. Arean, E. Conde and A. V. Ramallo, JHEP [**0912**]{}, 006 (2009) \[arXiv:0909.3106 \[hep-th\]\].\ D. Elander, JHEP [**1003**]{}, 114 (2010). \[arXiv:0912.1600 \[hep-th\]\].\ D. Arean, E. Conde, A. V. Ramallo and D. Zoakos, JHEP [**1006**]{}, 095 (2010). \[arXiv:1004.4212 \[hep-th\]\].\ J. Gaillard, D. Martelli, C. Nunez and I. Papadimitriou, Nucl. Phys.  B [**843**]{}, 1 (2011) \[arXiv:1004.4638 \[hep-th\]\].\ J. Schmude, \[arXiv:1007.1201 \[hep-th\]\].\ E. Conde, J. Gaillard, Nucl. Phys.  [**B848**]{}, 431-473 (2011). \[arXiv:1011.1451 \[hep-th\]\].\ E. Caceres, C. Nunez, L. A. Pando-Zayas, JHEP [**1103**]{}, 054 (2011). \[arXiv:1101.4123 \[hep-th\]\].\ D. Elander, J. Gaillard, C. Nunez, M. Piai, JHEP [**1107**]{}, 056 (2011) \[arXiv:1104.3963 \[hep-th\]\].\ E. Conde and A. V. Ramallo, JHEP [**1107**]{}, 099 (2011) \[arXiv:1105.6045 \[hep-th\]\].\ E. Conde, J. Gaillard and A. V. Ramallo, JHEP [**1110**]{}, 023 (2011) \[arXiv:1107.3803 \[hep-th\]\].\ A. Barranco, E. Pallante and J. G. Russo, JHEP [**1109**]{}, 086 (2011) \[arXiv:1107.4002 \[hep-th\]\]. F. Benini, F. Canoura, S. Cremonesi, C. Nunez and A. V. Ramallo, JHEP [**0702**]{}, 090 (2007) \[hep-th/0612118\]. F. Bigazzi, A. L. Cotrone and A. Paredes, JHEP [**0809**]{}, 048 (2008) \[arXiv:0807.0298 \[hep-th\]\]. F. Bigazzi, A. L. Cotrone, J. Mas, A. Paredes, A. V. Ramallo and J. Tarrio, JHEP [**0911**]{}, 117 (2009) \[arXiv:0909.2865 \[hep-th\]\]. F. Bigazzi, A. L. Cotrone, J. Tarrio, JHEP [**1002**]{}, 083 (2010) \[arXiv:0912.3256 \[hep-th\]\].\ F. Bigazzi, A. L. Cotrone, JHEP [**1008**]{}, 128 (2010) \[arXiv:1006.4634 \[hep-ph\]\]. F. Bigazzi, A. L. Cotrone, J. Mas, D. Mayerson and J. Tarrio, JHEP [**1104**]{}, 060 (2011) \[arXiv:1101.3560 \[hep-th\]\]. F. Bigazzi, A. L. Cotrone, J. Mas, D. Mayerson and J. Tarrio, arXiv:1110.1744 \[hep-th\]. V. G. Filev and D. Zoakos, JHEP [**1108**]{}, 022 (2011) \[arXiv:1106.1330 \[hep-th\]\]. J. M. Maldacena and J. G. Russo, JHEP [**9909**]{}, 025 (1999) \[arXiv:hep-th/9908134\]. A. Hashimoto and N. Itzhaki, Phys. Lett. B [**465**]{} (1999) 142 \[hep-th/9907166\]. R. Casero, C. Nunez and A. Paredes, Phys. Rev. D [**73**]{}, 086005 (2006) \[hep-th/0602027\]. F. Bigazzi, R. Casero, A. L. Cotrone, E. Kiritsis and A. Paredes, JHEP [**0510**]{}, 012 (2005) \[hep-th/0505140\]. J. Gaillard and J. Schmude, JHEP [**0901**]{}, 079 (2009) \[arXiv:0811.3646 \[hep-th\]\]. S. W. Hawking and G. T. Horowitz, Class. Quant. Grav.  [**13**]{}, 1487 (1996) \[arXiv:gr-qc/9501014\]. D. Mateos, R. C. Myers and R. M. Thomson, JHEP [**0705**]{}, 067 (2007) \[hep-th/0701132\]. F. Preis, A. Rebhan and A. Schmitt, JHEP [**1103**]{}, 033 (2011) \[arXiv:1012.4785 \[hep-th\]\]. D. E. Kharzeev, L. D. McLerran and H. J. Warringa, Nucl. Phys.  A [**803**]{}, 227 (2008) \[arXiv:0711.0950 \[hep-ph\]\].\ K. Fukushima, D. E. Kharzeev and H. J. Warringa, Phys. Rev.  D [**78**]{}, 074033 (2008) \[arXiv:0808.3382 \[hep-ph\]\].\ S. i. Nam, Phys. Rev.  D [**80**]{}, 114025 (2009) \[arXiv:0911.0509 \[hep-ph\]\]. A. Rebhan, A. Schmitt and S. A. Stricker, JHEP [**1001**]{}, 026 (2010) \[arXiv:0909.4782 \[hep-th\]\]. A. Gorsky, P. N. Kopnin and A. V. Zayakin, Phys. Rev. D [**83**]{} (2011) 014023 \[arXiv:1003.2293 \[hep-ph\]\]. M. N. Chernodub, Phys. Rev.  D [**82**]{}, 085011 (2010) \[arXiv:1008.1055 \[hep-ph\]\].\ M. N. Chernodub, Phys. Rev. Lett.  [**106**]{}, 142003 (2011) \[arXiv:1101.0117 \[hep-ph\]\].\ V. V. Braguta, P. V. Buividovich, M. N. Chernodub and M. I. Polikarpov, arXiv:1104.3767 \[hep-lat\]. N. Callebaut, D. Dudal and H. Verschelde, arXiv:1105.2217 \[hep-th\]. M. Ammon, J. Erdmenger, P. Kerner and M. Strydom, Phys. Lett. B [**706**]{}, 94 (2011) \[arXiv:1106.4551 \[hep-th\]\]. [^1]: For a comprehensive review we refer the reader to [@Filev:2010pm]. [^2]: For a detailed review on the smearing see [@arXiv:1002.1088] while for other solutions employing this technique see [@allsusyunquenched]. [^3]: All the hydrodynamic transport coefficients of the model were analyzed in [@hydro], while the addition of a finite baryon density was presented in [@Bigazzi:2011it]. For a review on unquenching the Quark Gluon Plasma see [@Bigazzi:2011db]. [^4]: with $\int \, \varepsilon(S_5) \, = \, \text{Vol}(S^5) \, = \, \pi^3$ [^5]: We remind the reader that beyond the quenched approximation the dual field theory has a positive $\beta$-function. As a result the mass of the fundamental fields runs with the energy scale and even in the supersymmetric case the physical mass $M_q$ differs from the bare mass $m_q$. [^6]: Note that we have taken advantage of the fact that to leading order in the ratio $N_f/N_c$ the profile of the fiducial embeddings is well approximated by the profile of the probe embeddings. [^7]: Note that from a holographic point of view even at first order in $\epsilon_*$ our ansatz provides a novel feature, namely the relation between the field $J(r)$ and the magnetization of the dual gauge theory. [^8]: Note that in this limit $b\equiv 1$ and $H\equiv J\equiv 0. $ [^9]: Note also that to first order in $\epsilon_*$, we have $\epsilon_q=\epsilon_*+O\left(\epsilon_*^2\right).$

ArXiv

--- abstract: 'Unstable dynamics characterizes the evolution of most solid tumors. Because of an increased failure of maintaining genome integrity, a cumulative increase in the levels of gene mutation and loss is observed. Previous work suggests that instability thresholds to cancer progression exist, defining phase transition phenomena separating tumor-winning scenarios from tumor extinction or coexistence phases. Here we present an integral equation approach to the quasispecies dynamics of unstable cancer. The model exhibits two main phases, characterized by either the success or failure of cancer tissue. Moreover, the model predicts that tumor failure can be due to either a reduced selective advantage over healthy cells or excessive instability. We also derive an approximate, analytical solution that predicts the front speed of aggressive tumor populations on the instability space.' author: - 'Daniel R. Amor$^{1,2}$, Ricard V. Solé$^{1,2,3}$' title: | Catastrophic shifts and lethal thresholds in a propagating front\ model of unstable tumor progression --- Introduction ============ Cancer is a disease that can be initiated by the failure of a single cell. Whenever such failure leads to some proliferation advantage over the neighboring somatic cells, this single cell is prone to originate its own cell lineage within its host tissue. After many rounds of replication, additional failures may occur, eventually generating a large population of abnormal, proliferating cells. This would be a rough description of the disease, but it would be more appropriate to say that cancer is an evolutionary dynamic process \[1,2\]. Changes occur in time and accumulate over generations and the final success of the tumor requires an appropriate accumulation of changes affecting different types of genes. We can classify cancer genes into three basic categories \[3\]: (a) oncogenes, (b) tumor suppressor genes and (c) stability-related genes. These groups corresponds to genes that (a) increase replication due to mutation, (b) increase cell growth when the gene is silenced or lost and (c) modify genome stability due to failures in cell division, repair and maintenance mechanisms \[4-9\]. All these changes occur through the process of cell replication, when cancer genes are likely to experience mutations or losses \[10\] leading to the emergence of fitter mutant clones. In order to understand the evolution of cancer, a huge amount of mathematical models have analyzed the impact of selection \[11\] on the evolution of clones. The stochasticity of mutations has also been shown to play a major role in triggering the clonal competition among different mutants, and could be a principal reason for the high heterogeneities (and the long waiting time to malignancy) observed in cancer development \[12\]. Another very relevant mechanism is spatial structure, which is also very significant in many ecological and evolutionary processes \[13\]. For example, in the context of asexual evolving populations, the spatial competition between different clones slows down the establishment of driver mutations (i. e. mutations causing a selective advantage) \[14\] if the population exceeds a critical size. Analogous results have been found in the context of cancer evolution, where space can increase the waiting time to tumor malignancy \[15\]. In other cases, the spatial invasion of tumors has been modeled as a propagating front \[16\]. This has permitted to, e.g., compare the role of advection (chemotaxis) and cell diffusion on the invasion speed of glioblastomas \[17,18\], or analyze the invasiveness enhancement by acidic pH gradients at tumor-host interfaces \[19\]. Although most classic models of cancer evolution deal with those factors associated with growth and competition among clones, a specially important characteristic of most tumors is precisely the increased levels of instability associated to progression. Instability can be understood in terms of mutations but also of losses and gains of genetic components that modify genome stability, making cells more prone to errors while replicating \[20\]. Mutations have been an intrinsic part of all evolutionary models of population dynamics (including cancer) but it is typically assumed that mutation rate remains constant over time. In genomically unstable tumors, the failure of the repair mechanisms, along with the generation of aneuploidy, makes possible to damage key components associated to the maintenance of genome integrity \[4-9,20\]. With their loss or failure, further increases of instability are expected to occur, since other genes linked to stability and repair are more likely to be damaged. As a consequence, instability itself can evolve over time. Such evolvable trait raises the question of how much instability can accumulate through carcinogenesis. It has been suggested that optimal instability rates \[4\] as well as thresholds to instability exist. The latter define the transition boundaries between viable and non-viable cancer populations \[21-24\]. There are actually examples of phase transitions defining the boundaries of viability in RNA viruses\[25-30\]. RNA virus populations are quasispecies \[22,30\] i.e. highly heterogeneous, related genotypes. Critical thresholds of mutation have been predicted and later experimentally tested \[31-33\] using in-vitro scenarios. The presence of such critical transitions have also received a great interest from the field of statistical physics. The nature of the resulting phase transitions have been analyzed for several fitness landscapes, both for finite-size competing molecules \[34\] and in the limit of infinitely large chains \[35\]. Error thresholds have also been reported in asexual evolutionary scenarios beyond the RNA viruses. Remarkably, it has been reported that natural selection, favoring immediate fitness benefits, may permit the hitchhiking of deleterious mutations that will finally lead to the population’s extinction in the long term\[36\]. The similarities between unstable cancer and RNA viruses suggests a therapeutically very interesting possibility: the use of additional instability as anticancer therapy \[23,37\]. That means that, instead of trying to decrease the tumor cells’ mutagenesis, an attempt to increase it towards non-viable levels could be a suitable way to fight the disease. Due to the qualitatively sharp change associated to the presence of instability thresholds, a physics approach to phase transitions in cancer quasispecies can be successfully used \[23,24,38-40\]. In this paper we explore the dynamics and phases of unstable cancer by constructing an analytical model of tumor progression to be defined as a front propagation problem \[41\] in the space of instability. By using this approximation we provide a better and easily extendable formal description of tumors that allows to characterize both the presence of transitions and the population structure that emerges in each phase. It also provides a well defined, formal approach to predict the speed of cancer propagation. The paper is organized as follows. In section II we present the rationale for the presence of a phase transition phenomenon separating a phase where the tumor will fail to succeed due to a high instability from another phase where it is expected to win. In section III we revisit the previous linear, discrete model of cancer cells dynamics and we explain some of its limitations. Section IV is devoted to present the integral model of unstable cancer, that improves the previous mathematical description of the disease. Section V presents several scenarios for tumor evolution predicted by the integral model (an analysis of the resulting phase space is included). In section VI we derive an approximate, analytical expression for the tumor front speed on the instability space, and we compare it with some numerical solutions for the model equations. The last section is devoted to discuss the potential implications of our results. Transitions in tumor instability ================================ In order to provide a rationale for the existence (and potential implications) of instability thresholds, let us first consider a mean field, two-compartment model of unstable cancer dynamics. In this model the population will be composed of two cell species, namely, host cells $H$ and cancer cells $C$. If we indicate as $r_{n}$ and $r_{c}$ the rates of growth of normal (host) and cancer cells, respectively, we can write the following evolution equations: $$\begin{aligned} {\frac{dH}{dt}} &=&r_{n}H-H\phi (H,C) \\ {\frac{dC}{dt}} &=&r_{c}C-C\phi (H,C),\end{aligned}$$ where $\phi (H,C)$ is an outflow term that represents the competition between both species. If we consider that the overall cell population $H+C$ is constant (because cells fill a given fixed space) the function $\phi$ reads $\phi =r_{n}H+r_{c}C$ which is actually the average rate of growth. The following step consists in defining the growth rates for each species. Regarding normal cells, it is sensible to assume a constant growth rate $% r_{n}$ (normal cells are renewed in a stable way to ensure that body functions are properly carried on). The situation is different for cancer cells, for which their average growth rate $r_{c}$ will depend on how much mutations have been accumulated in the cells’ genome. Concretely, the growth rate $r_{c}$ can be increased by the effects of driver mutations (i.e., mutations promoting cell replication) or decreased by deleterious mutations. Let us denote $\mu $ as the probability that a mutation takes place when replicating a given gene. Thus, $\mu $ is a measure of the genetic instability of the population. Consider that there exist a number $N_{r}$ of growth-related genes. If a growth-related gene is damaged (mutated) during the cell replication process, an average increase $\delta _{r}$ in the growth rate is expected. Thus, the average increase in $r_{c}$ due to driver mutations affecting our cancer population is simply: $$f_{1}(\mu )=N_{r}\mu \delta_{r}.$$ Similarly, we should expect a decrease in the growth rate due to the potential damage produced if a house keeping gene is damaged or lost. If $% N_{h}$ indicates the number of such genes, the probability that no one is damaged will read $$f_{2}(\mu )=(1-\mu )^{N_{h}}.$$ Available estimates indicate that $N_h \sim 500-600$ essential genes exist \[42\] whereas $N_r$ can be smaller or higher depending on the type of cancer considered. Around one percent of genes in the human genome appear related to the emergence of cancer \[43\]. In the above eqs. (3) and (4) we have assumed that the mutation and replication rates are the same for all genes, as well as a constant fitness benefit $\delta _{r}$ from driver mutations. Obviously, this is only simple approach to the more complex reality in which each gene mutates with a different probability and provides a different fitness benefit (or loss). Thus, the final rate of replication will be the product: $$r_{c}(\mu )=f_{1}(\mu )f_{2}(\mu )=\left( r_{n}+\mu N_{r}\delta r\right) (1-\mu )^{N_{h}},$$ where we have assumed that, in the absence of genetic instability ($\mu =0$) the normal cells replication rate $r_{n}$ is recovered. The above function (5) has a maximum at a given optimal instability rate. This is shown in Fig. 1, where we plot $r_{c}(\mu )$ for a given combination of parameters. The maximum is achieved at an optimal instability level $\mu _{o}\approx 1/N_{h}$. Considering Eq. (5) and the constant population constraint, it is possible to reduce the system of Eqs. (1)-(2) into a single equation describing the dynamics of the system in terms of the genetic instability $\mu $. Thus, the cancer cell population is now captured by a logistic-like nonlinear equation: $${\frac{dC}{dt}}=r_{n}(\Gamma (\mu )-1)C(1-C)$$ Two fixed points are present: the zero-population one $C^{\ast }=0$ and the maximum population state, here $C^{\ast }=1$. It is easy to see that the first is stable if $\Gamma (\mu )<1$ and unstable otherwise. By properly defining the function $\Gamma (\mu )$ we might be able to define the conditions under which genetic instability allows cancer growth to occur and overcome the host tissue. The critical mutation rate separating the two scenarios is sharp and defines a phase transition. The presence of a phase transition in this toy mean field model involving competition between two homogeneous populations offers an interesting prediction: further increases of instability can force cancer cells to enter the lethal phase. However, understanding how such shifts can occur requires a better understanding of the ways cancer cell populations evolve. Cancer cell populations are highly heterogeneous \[44,45\] and that means that we need to depart from the previous model approach. Linear model of unstable cancer =============================== In an early paper \[46\] a discrete, sequential model of unstable cancer was introduced. The model considered a population of cancer cells having different levels of instability and competing among them and with the normal tissue (figure 2). This led to a description of $M$ levels of instability describing an heterogeneous cancer cells population, which was governed by the following set of $M$ differential equations:$$\frac{dC_{i}}{dt}=f_{i-1}\mu _{i-1}C_{i-1}+f_{i}(1-\mu _{i})C_{i}-C_{i}\Phi (H,\mathbf{C}), \label{2}$$ where $i=1,3,...,M$, $\mathbf{C}=(C_{1},...,C_{M})$ and we consider the terms $\mu_{0}=\mu_{M}=0$ so that they properly define the first and last of the equations in (7). In the set (7), $H$ indicates the host (healthy) population, whose dynamics would be described by an additional equation $dH/dt=f_{H}(H,C)$ which takes the general form $$\frac{dH}{dt}=G(H)-H\Phi (H,\mathbf{C}) \label{10}$$Here $G(H)$ introduces the explicit form of growth characterizing the normal tissue. A constant population constraint (CPC) was also introduced, namely a total constant population size $H+\sum_{i}C_{i}=1$. This leads to an explicit form of the competition $\phi $ function, namely $$\Phi (H,\mathbf{C})=G(H)+\sum_{k=1}^{M}f_{k}C_{k}$$which is nothing but the average replication rate. A numerical analysis of this system was performed for some parameter values, showing that the population dynamics of the cancer population spread over mutation space as a wave until a stable distribution (showing a single peak) around high instability levels was observed. However, no systematic analysis was performed in order to characterize potential phases and their implications. In particular, it was not studied the behavior exhibited by the heterogeneous population close to the optimal/lethal thresholds. Moreover, the linear model above is an oversimplification and a better description is needed in order to make reliable predictions. Integral equation expansion =========================== The linear instability model reveals an important dynamical feature of unstable dynamics: a propagating front is formed and moves through instability space. Fronts (and their propagation dynamics) are a well known characteristic of many relevant biological processes \[41,44,45\] and can be analyzed in a systematic way through well known methods. Our first step here will be to convert the discrete model presented above into a more general, analytically tractable integral equation form. Such model will allow us exploring the phase space of our system and to make some analytic estimates of propagation speed. An integral equation model can be derived starting from the previous linear model. Let us first notice that the equations (7) for $C_{i}$ can be re-written as $$\frac{dC_{i}}{dt}=\sum_{1}^{M}f_{j}C_{j}w_{ji}-C_{i}\Phi (H,\mathbf{C}). \label{4}$$This is done by introducing the following notation: $$w_{ji}=\delta _{j,i-1}\mu _{j}+(1-\mu _{j})\delta _{ij}, \label{3}$$and, as explained in the previous section, we consider the condition $\mu_{0}=\mu_{M}=0$ to properly describe the evolution of $C_{0}$ and $C_{M}$. An integral equation can be now constructed, using the continuous variable $% C_{i}(t)=\Delta \mu \cdot c(\mu ,t)$. Moreover, we need to generalize the functional connection between different instability levels, which was assumed to be a simple function in (7) but could adopt different forms. A general integral equation can be constructed, namely: $$c(\mu,t+T)=c(\mu,t)+T\int\limits_{-\mu}^{0}f(\mu+\Delta _{\mu})c(\mu+\Delta_{\mu},t)\omega (\Delta _{\mu})d\Delta _{\mu}-c(\mu,t)T\phi (H,\mathbf{c}), \label{6}$$ where we have used a continuous dispersal kernel $\omega (\Delta _{\mu })$ \[16,41,47\] which provides the probability density that cancer cells in $c(\mu -\left\vert \Delta _{\mu }\right\vert ,t)$ produce offspring, after a given time $T$, within the $\mu $-coordinate i.e. further cells within the $c(\mu ,t+T)$. Moreover, in Eq. (12) we have changed the notation of the average fitness from $\Phi (H,\mathbf{C})$ \[as it appears in Eqs. (7) and (8)\] in order to remark that we now use a continuous description for cancer cells. While in the linear model \[Eqs. (7) and (8)\] the average fitness depends on $H$ and $\mathbf{C}$, in our integral model \[Eq. (12)\] the average fitness must depend on both $H$ and the continuous distribution of cancer cells at time t (that we have writen as $\mathbf{c}$ to explicitly indicate its correspondence to $\mathbf{C}$ in the linear model). Following analogous steps to those for cancer cells, we can also develop the differential Eq. (8) for healthy cells so that we obtain an explicit expression for the population $H$ at time $(t+T)$. This yields: $$H(t+T)=H(t) + T[G(H)-H\phi (H,\mathbf{c})] \label{10bis}$$ The constant population requirement (defined above as $C+N=1$ for the mean field model) can be expressed here as $$\begin{aligned} H(t)+\int_{0}^{M}c(\mu ,t)d\mu &=&1 \\ H(t+T)+\int_{0}^{M}c(\mu ,t+T)d\mu &=&1\end{aligned}$$and we assume that $M$ is large enough so that we can ensure that $c(M,t)=0$. In this paper we will use this integral equation approach to describe our cancer quasispecies model. This model allows us to properly study the way the instability wave can (or cannot) propagate and some other phenomena including the catastrophic collapse of the cancer population once the unstable wave crosses some given thresholds. Using the previous condition and definitions, it is possible to develop our model equation. Let us indicate as $\phi =\phi (H,\mathbf{c})$, $f_{H}=G(H)-H\phi $ \[which, considering Eq. (13), can be understood as the change in $H$ cells per unit time\], and compute the total cancer cells population as: $$\Lambda (t)=\int_{0}^{M}c(\mu ,t)d\mu.$$ Note that $\Lambda (t)$ strictly depends on $(M,t)$, but the dependence on $M$ has been omited because (as mentioned above) we consider $M$ is high enough to satisfy the condition $c(M,t)=0$. Thus, it is possible to see that our system is described by the following mathematical expressions: $$\begin{aligned} T f_{H} +H+ \Lambda (t) +T \int\limits_{0}^{M}\int\limits_{-\mu}^{0}c(\mu+\Delta _{\mu},t)f(\mu+\Delta _{\mu})\omega (\Delta _{\mu})d\Delta _{\mu}d\mu-T\phi \Lambda (t)= 1 \\ \Rightarrow 1+TH\phi+T\phi \Lambda (t) =TG(H)+H+\Lambda (t)+\int\limits_{0}^{M}T\int% \limits_{-\mu}^{0}c(\mu+\Delta _{\mu},t)f(\mu+\Delta _{\mu})\omega (\Delta_{\mu})d\Delta _{\mu}d\mu, \label{7}\end{aligned}$$ From the above equation (17) it is easy to derive the following expression for the average fitness of the population (that includes normal tissue and tumor cells): $$\phi (H,\mathbf{c})=G(H)+\int\limits_{0}^{M}\int\limits_{-\mu }^{0}c(\mu +\Delta _{\mu },t)f(\mu +\Delta _{\mu })\omega (\Delta _{\mu })d\Delta _{\mu }d\mu . \label{fitness}$$ It is worth to note that the integro-difference equation (\[6\]) permits to analyze several dynamical properties of the system which cannot be attained by means of the previous linear model (\[4\]). In the linear model, the offspring of tumor cells in a given stage $i$ may grow either in the same stage $i$ or in the subsequent $i+1$. A desirable feature of the continuous description from (\[6\]) is that the dispersal kernel can easily model different forms of instability-driven spread in the genetic landscape. In the following section, we analyze a simple case in which migration probability decays exponentially with the jumping distance $\Delta _{\mu }$. The linear model can also be recovered from Eq. (\[6\]) by introducing a dispersal kernel that restricts mutations to discrete points in the $\mu $-space. In order to derive some analytical solutions of the system, such simplified dispersal kernels will be shown to be specially useful. Wave fronts in instability space ================================ In this section, we present several scenarios in which a tumor can either collapse or succeed over a healthy tissue. According to the integral model \[Eqs. (\[6\]) and (\[fitness\])\], tumor evolution is mainly governed by competition. As explained above, this competition involves not only the fight between cancer cells and healthy cells, but also the struggle within cancer cell clones. In the previous section we have presented a model that is mainly based on two dynamical features of tumors: replication (introduced by the growth function $f(\mu )$) and mutation (given by the dispersal kernel $\omega (\Delta _{\mu })$). Concerning the replication process, below we consider some specific growth functions involving a constant reproduction rate for healthy cells, so that $G(H)=r_{n}H$. For tumor cells, the growth function depends on instability as $f(\mu )=r_{n}\left( 1+\alpha \mu \right) exp(-\mu /\mu _c)$. This was derived in \[23\] from the probabilistic condition defined by equation (5). The rate $\alpha $ introduces a selective advantage for cancer cells over healthy cells. The constant $\mu _{c}$ refers to a characteristic instability rate. In order to model the mutant trend of cancer cells, let us consider the following continuous function for the dispersal kernel: $$\omega (\Delta _{\mu })=\frac{1}{\mu _{disp}} \exp\left( \frac{-\left\vert\Delta _{\mu }\right\vert}{\mu _{disp}}\right) . \label{kernel}$$ According to Eq. (\[kernel\]), a parent cell generates offspring at similar instability domains (i.e. situated at $\Delta _{\mu }\rightarrow 0$) with higher probability than new cells presenting much higher instability (i.e., living at $\Delta _{\mu }>>0$). The parameter $\mu _{disp}$ represents a characteristic (within generation) instability increment. Since we have $$\int_{-\infty}^{0}\omega (\Delta _{\mu })d\Delta _{\mu }=1$$ the dispersal kernel distributes the cells of the new generation in the instability space, but it does not modify the total number of cancer cells in the system. Tumor wins phase ---------------- Figure 3a shows the evolution of a population of cancer cells which initially composes the $0.001\%$ of the cells in the system. Cancer cells at $t=0$ have been equally distributed within a range of low instability (namely, $\mu \in (0,2\cdot 10^{-4}]$). We observe an early stage ($t\in \lbrack 0,150]$) in which tumor cells remain at low values of the population density $c(\mu ,t)$. Within this initial period, cancer cells do not overcome healthy cells because their selective advantage is not significant (i.e., $f(\mu )\simeq G(H)$ because $\mu \simeq 0$). The dispersal kernel $\omega (\Delta _{\mu })$ pushes forward the tumor population towards higher instability domains. In other words, at each time step a fraction of the cancer cells offspring becomes sensibly more unstable than their parent cells. A rapid increase in cancer cells population density is observed about $t=200$ generations. The rapid growth affects cells whose genetic instability is above a certain threshold (see the region above $\mu =1.5\cdot 10^{-2}$). This indicates that such degree of instability provides for significant selective advantage over other cells in the system. During the fast growth phase, the population not only attains a large fraction of the total population, but it also continues migrating (see the left to right dispersion of the population wave). At the end of the time series in Fig. 3a, the concentration of cancer cells in the system is about $50\%$ (we consider this condition is enough to cause the death of the host). This is an example of the dynamics at the cancer expansion phase. Tumor failure phase ------------------- It seems reasonable to think that increasing the characteristic migration distance $\mu _{disp}$ should accelerate tumor proliferation, because cancer cells will reach optimal instability domains faster. However, increasing $% \mu _{disp}$ does not necessarily lead to the tumor-win phase. It can actually jeopardize cancer propagation even when an already established population is formed. If a tumor cell produces highly mutant descendants (i.e., new cells accumulating many new mutations) with high probability, it follows that the probability of generating descendants without additional mutations cannot be very large. Figure 3b depicts an example of the tumor-failure phase. In this case the selective advantage presents a higher value (namely, $\alpha =50$) than that for the tumor in the previous scenario. Here we observe a tumor population wave diffusing in the instability space, always coexisting with normal cells (the total number of cancer cells $\Lambda(t)$ do not exceed the $12\%$ at any generation). Despite the relatively high selective advantage $\alpha$, the high value of $\mu _{disp}$ prevents the tumor population to remain at the optimal instability domain, and hence cancer cells cannot grow fast. The tumor moves towards excessive instability, and cancer replication becomes smaller than that of the host tissue. These conditions define the tumor extinction phase. Catastrophic tumor decay ------------------------ A qualitatively different and somewhat unexpected outcome is displayed in Fig. 3c, where we have set a lower value of $\mu _{disp}$. As a result, a fast extinction of healthy cells occurs and cancer cells invade all the available space before $t=200$. Here we let the system evolve beyond the absence of healthy cells. Even if this situation typically involves the elimination of host cells, it could be observed in cell culture conditions. Moreover, we need to consider a potentially relevant situation, namely when a given tumor has expanded within large parts of the organ, as it occurs with many malignant cancers. After the rapid increase in cancer cells population density ($t\simeq 200$), the tumor continues its migration towards higher instability. Since the value of $\mu _{disp}$ is relatively high, the tumor population is unable to stay within the optimal region. At every new generation, a large fraction of the progeny accumulates new mutations. The final outcome is very interesting: a collapse finally occurs. This is illustrated in figure 3d, where we plot the total cancer population and the average instability (inset) for the example of figure 3c. Around $1700$ generations, cancer cells have accumulated so many mutations that they are almost unable to produce viable descendants. After $t=2000$ there is no significant cancer cells population. Despite the slow growth of $\langle \mu \rangle $, a catastrophic shift occurs, with a rapid decay of the tumor. Catastrophic shifts have been previously described within ecological and social systems \[48\] and are characterized by sudden system responses triggered by slow, continuous changes of given external control parameters. The novelty of our observation is that the changing parameter is affected by (and affects) population dynamics and thus is not externally tuned but internally increased. Phase space ----------- A systematic exploration of the parameter space provides a picture of the two main phases, as shown in Fig. \[7\]. The two axes involve a wide range of values for both $\alpha $ and $\mu _{disp}$. In the first phase (gray squares), the tumor is driven to extinction. Extinction arises as a combination of two components: *i)* an insufficient fitness advantage of the early cancer cells (the cancer population progressively decays without reaching enough instability to develop), or *ii)* the tumor inability to keep the optimal instability (when this happens, a moderate population growth precedes the tumor failure). The second region (white area) stands for tumors that grow enough to overcome the healthy tissue. The transition between the two regions is also marked by a rapid increase in the transient time. In Fig. \[7\]b we have depicted the transient time steps (i.e., generations of cancer cells) to reach either the tumor extinction or its stable expansion to equilibrium values. As expected, longer times are needed near the phase transition. Tumor front speed ================= In the previous section, we have seen how some tumor population waves diffuse in the instability space. A relevant feature of propagating fronts, with direct importance for tumor growth, is the propagation speed of the front. Such speed has been actually calculated for spatially growing tumors \[17,49,50\] and the front is thus a spatially defined one. Although we are here considering front propagation through instability space, the same reasoning applies. Here we derive an analytical, approximate solution for the front speed of the tumor. This will provide a quantitative measure of how fast cancer instability propagates. Since deriving an exact analytical expression for the front speed can be extremely cumbersome, some approximations are required. First, let us consider early stages in tumor development (such as the first $% 150$ in Fig. \[Twins\]). Here the system is mostly composed of healthy cells, and few of cancer cells. This permits to approximate the complex expression for the average fitness \[see Eq. (\[fitness\])\] as the reproduction rate of healthy cells, i.e.,$$\phi (H,\mathbf{c})\simeq G(H)\simeq r_{n}. \label{early}$$The second approximation we will consider refers to the dispersal kernel. According to Eq. (\[kernel\]) in the previous section, the dispersal kernel is a continuous function defined in the interval $[-\infty,0]$. In this section we will consider the following simpler, discrete dispersal kernel: $$\omega (\Delta _{\mu })=2p_{e}\delta (\Delta _{\mu })+(1-p_{e})\delta (\Delta_{\mu }+\mu _{disp})\gamma(\mu_{disp}), \label{dirackernel}$$ where $\delta (\Delta _{\mu })$ corresponds to the Dirac delta function operating on the variable $\Delta _{\mu }$, and $\gamma(\mu_{disp})=1$ if $\mu _{disp}\cancel{=} \mu$ and $2$ otherwise \[51\]. The above discrete kernel (\[dirackernel\]) considers that every new cell can either stay at the same instability $\mu $ of the parent cell (with probability $p_{e}$, which is called persistence) or jump into a higher instability $\mu +\mu _{disp}$ \[with probability $% (1-p_{e})$\]. Although the discrete kernel (\[dirackernel\]) is much simpler than the continuous kernel (\[kernel\]), it also models a major feature in cancer cells replication (see the previous section), that is: the stronger the mutant trend of cancer cells, the weaker the ability of the population to keep an optimal instability. Thus, according to Eqs. (\[early\]) and (\[dirackernel\]) above, our approximation to Eq. (\[6\]) reads: $$c(\mu,t+1)=c(\mu,t)+\int\limits_{-\mu}^{0}f(\mu+\Delta _{\mu})c(\mu+\Delta_{\mu},t)(2p_{e}\delta(\Delta_{\mu})\\ +(1-p_{e})\delta(\Delta_{\mu}+\mu_{disp})\gamma(\mu_{disp}))d\Delta _{\mu}-c(\mu,t)r_{n}. \label{dirac1}$$ Taking into account the integrative properties of the Dirac delta function $% \delta (\Delta _{\mu })$, the above equation \[dirac1\] can be rewritten in terms of a much simpler functional form: $$\begin{array}{c} c(\mu ,t+1)=c(\mu ,t)+p_{e}c(\mu ,t)f(\mu ) +(1-p_{e})c(\mu -\mu _{disp},t)f(\mu -\mu _{disp})-c(\mu ,t)r_{n},% \end{array} \label{diracmodel}$$ where we have assumed that the condition $\mu\geq\mu_{disp}$ holds, since we are interested in the propagation of the tumor front. Indeed, for $\mu<\mu_{disp}$ cancer cells evolve as in Eq. (\[diracmodel\]) but neglecting the third term on the RHS. The front speed from reaction-dispersal integro-difference equations such as (\[6\]) can be obtained under some general assumptions \[16,17\] associated with the shape to be expected for the propagating front. Here we are interested in the simplified version (\[diracmodel\]) of the model. Thus we only need to assume that there exist constant shape solutions of the form $$c(\mu ,t)=c_{0}\exp \left [ -\lambda z \right]$$ for large values of the coordinate $z \equiv (\mu -vt)$. This yields the following approximate, analytic relation between the tumor front speed and the wave front shape parameter $\lambda$: $$v(\lambda)= \frac{1}{\lambda }\ln \left [ p_{e}f(\mu ) +(1-p_{e})f(\mu -\mu _{disp}) e^{\lambda \mu _{disp}} -r_{n}+1 \right ] . \label{speed}$$ Finally, by means of the standard, marginal stability condition \[16\] the minimal speed $v_{*}=\min_{\lambda >0}v(\lambda)$ is the one selected by the front. In the analysis below, the following method has been used to find the value of the approximate speed $v_{*}$. First, we plotted the numerical solution to Eq. (25) on the $\lambda >0$ axis. As usual, we obtain a function $v(\lambda)$ that is convex from below \[16\]. Thus, we look for the minimum value of this function, which reveals the approximate value for the front speed of the tumor as it propagates through instability space (i.e., $v_{*}$). The approximate front speed $v_{*}$ should not be taken as a general trend in tumor evolution, since it is subject to the approximations explained above. Indeed, for cases in which healthy cells overcome the tumor it eventually predicts negative values of the front speed. However, predicting a negative front speed can also be seen as the retreat (i.e., the death) of the cancer population (which at early times is only composed by a few cancer cells with $\mu \rightarrow 0$). Nevertheless, the approximate speed $v_{*}$ provides remarkably good results for the front speeds of lethal tumors (i.e., for tumors within the parameter region in which the tumor succeeds), as we show in Fig. \[8\]. Fig. \[8\] shows a comparison between the numerical and the approximate analytical speed $v_{*}$ for the tumor front speed as a function of the characteristic dispersal distance $\mu _{disp}$. Numerical solutions for the front speed have been computed by numerically solving \[52\] the model Eqs. (\[6\]) and (\[fitness\]) using the discrete version of the dispersal kernel (\[dirackernel\]). For both the numerical and the approximate analytical solutions, the front speed monotonically increases with the characteristic distance $\mu _{disp}$. As far as the order of magnitude is concerned, the approximate analytical speed $v_{*}$ is able to predict the more exact numerical results for the tumor front speed. Furthermore, relative differences (which are typically above $15\%$) between the analytical results and the numerical solutions are approximately independent of $\mu _{disp}$. The solutions for the front speed on the instability space in Fig. \[8\] exhibit an approximately linear dependence on the characteristic dispersal distance $\mu _{disp}$. Since the instability $\mu $ of tumor cells is directly related with the selection of cells during tumor evolution, $\mu _{disp}$ can also be interpreted as a parameter that determines the selective intensity on cancer cells. Previous models on adaptive front waves have also yielded adaptation speeds that depend on selective intensity parameters. In this context, in \[14\] the speed of adaptation for asexual clones that compete for space scales as $\mu ^{1/2}$ in one-dimensional habitats and $\mu ^{1/3} $ in two-dimensional ones (in their model, $\mu $ stands for the rate at which new mutations appear at each lattice site). Also, in Ref. \[53\] the authors found that, when a continuous two-dimensional space is considered, genetic wave speeds are proportional to $s^{1/2}$, where $s$ represents the small fitness effects $s$ of beneficial mutations. In our model we have studied the front speed of the tumor on the instability space, without taking into account neither spatial structure nor the physical environment. However, both $\mu $ and $\mu _{disp}$ have similar interpretations to the selective parameters in Refs. \[14,53\]. Future work could be directed to introduce spatial effects into our model, and it would be interesting to explore if similar scaling exponents arise in the dependence of the tumor invasion speed and genetic instability rates. Discussion ========== In this paper we have presented an integral model for the evolution of unstable tumors. Our model improves a previous compartment description of the cancer cells population, because we consider the genetic instability as a continuous variable that characterizes the state of the cell. The model considers a population of tumor cells that replicate and migrate (mutate) in the instability space, while competing for available resources (a limited population constraint has been applied). This model is based on several simplifying assumptions, from the linear nature of interactions between instability levels to the dispersal kernels used. We have presented an extended analysis of unstable cancer evolution over the two most relevant parameters of the model: the selective advantage of cancer cells over the healthy cells population, and the characteristic migration distance within instability space (which determines the mutant tendency of cancer cells). Several outcomes of the process have been found. Two of them are expected: either the growth or the failure of cancer to succeed are predicted by the simplest mean field model that can be defined, as discussed in section II. The integral equation approach confirms such prediction, although it allows to substantiate it in more accurate ways, providing a formal framework to calculate useful quantities, particularly the front speed of our population through the $\mu $-space. Moreover, this formal approach provides a natural way to properly introduce population heterogeneity. An additional scenario has also been found, namely the catastrophic shift phase, where the tumor grows, eventually expanding over a significant part of the total available space, with a steady growth of instability. However, at some point the excessive instability level leads to a population collapse, with no cancer cells in the end. Our model does not consider immune components or other biologically relevant factors \[54\]. Instead, the key factor responsible for the tumor collapse is high genetic instability. This result provides further support to the original proposal that lethal thresholds of instability exist in cancer \[23,24\] which could be exploited for therapeutic purposes, even when major success of the cancer population is observable. Future work should further explore this observation, adding also other known threats to cancer progression, such as starvation or hypoxia, which could further enhance the frequency and sharpness of these thresholds. We thank the members of the CSL for useful discussions. This work was supported by grants from the Fundacion Botin, a MINECO grant and the Santa Fe Institute. —————————————————————————– References ========== 1. M. Greaves and C. C. Maley, Nature **481**, 306 (2012). 2. L. M. F. Merlo, J. W. Pepper, B. J. Reid and C. C. Maley, Nature Rev Cancer **6**, 924 (2006). 3. B. Vogelstein and K. W. Kinzler, Nature Medicine **10**, 789 (2004). 4. D. P. Cahill, K. W. Kinzler, B. Vogelstein and C. Lengauer, Trends Genet. **15**, M57 (1999). 5. M. Chow and H. Rubin, Cancer Res. **60**, 6510 (2000). 6. J. Jackson L. A. Loeb Genetics **148**, 1483 (1998). 7. C. Lengauer, K. W. Kinzler and B. Vogelstei, Nature **396**, 643 (1998). 8. L. A. Loeb, Cancer Res. **51**, 3075 (1991). 9. L. A. Loeb, Cancer Res. **54**, 5059 (1994). 10. B. Vogelstein, N. Papadopoulos, V. E. Velculescu, S. Zhou, L. A. Diaz Jr. and K. W. Kinzler, Science **339**, 1546 (2013). 11. F. Michor, Y. Iwasa and M. A. Nowak, Nat. Rev. Cancer **4**, 197 (2004). 12. N. Beerenwinkel, T. Antal, D. Dingli, A. Traulsen, K. Kinzler, V. E. Velvulescu, B. Vogelstein and M. A. Nowak, PLOS Comp. Biol. **3**, e225 (2007). 13. R. V. Solé and J. Bascompte, *Self-Organization in Complex Ecosystems* (Princeton U. Press, Princeton, New Jersey, 2007). 14. E.A. Martens and O. Hallatschek, Genetics **189**, 1045 (2011). 15. E.A. Martens, R. Kostadinov, C. C. Maley and O. Hallatschek, New J. Phys. **13**, 115014 (2011). 16. J. Fort and T. Pujol, Rep. Prog. Phys. **71**, 086001 (2008). 17. J. Fort and R. V. Solé, New J. Phys. **15**, 055001 (2013). 18. E. Khain, L. M. Sander, and A. M. Stein, Complexity **11**, 53 (2005). 19. R. A. Gatenby and E. T. Gawlinski, Cancer Res. **56**, 5745 (1996). 20. S. Negrini, V. G. Gorgoulis and T. D. Halazonetis, Nat. Rev. Mol. Cell Biol. **11**, 220 (2010). 21. R. A. Gatenby and B. R. Frieden, Cancer Res. **62**, 3675 (2002). 22. R. A. Gatenby and B. R. Frieden, Mutat. Res. **568,** 259 (2004). 23. R. V. Solé, Europ. J. Phys. B **35**, 117 (2003). 24. R. V. Solé and T. Deisboeck, J. Theor. Biol. **178**, 47 (2004). 25. J. M. Coffin, Science **267**, 483 (1995). 26. E. Domingo, J. J. Holland, C. Biebricher and M. Eigen, in: *Molecular Evolution of the Viruses*, edited by A. Gibbs, C. Calisher and F. Garcia-Arenal (Cambridge U. Press, Cambridge, 1995). 27. Domingo, E. and Holland, J. J. in: *The evolutionary biology of RNA viruses*, pp. 161-183, edited by S. Morse (Raven Press, New York, 1994). 28. M. Eigen, Naturwiss. **58**, 465 (1971). 29. M. Eigen, J. McCaskill and P. Schuster, Adv. Chem. Phys. **75**, 149 (1987). 30. P. Schuster, in: *Complexity: metaphors, models and reality*, pp. 383-418, edited by G. A. Cowan, D. Pines and D. Meltzer (Addison-Wesley, Reading, MA 1994). 31. S. Cottry, C. E. Cameron, and R. Andino, Proc. Natl. Acad. Sci. USA **98**, 6895 (2001). 32. L. A. Loeb, J. M. Essigmann, F. Kazazi, J. Zhang, K. D. Rose and J. I. Mullins, Proc. Natl. Acad. Sci. USA **96**, 1492 (1999). 33. J. J. Holland, E. Domingo, J. C. de la Torre and D. A. Steinhauer, J. Virol. **64**, 3960 (1999). 34. P. R. A. Campos and J. F. Fontanari, PRE **58**, 2664 (1998). 35. P. Tarazona, Phys. Rev. A **45**, 6038 (1992). 36. P. J. Gerrish, A. Colato, A. S. Perelson , and P. D. Sniegowski, PNAS **104**, 6266 (2007). 37. L. A. Loeb, Nat. Rev. Cancer **11**, 450 (2009). 38. Q. Zhang and R. H. Austin, Annu. Rev. Cond. Mat. Phys **3**, 363 (2012). 39. R. Pastor-Satorras and R. V. Solé, Phys. Rev. E **64**, 051909 (2001). 40. D. B. Saakian, E. Munoz, C. K. Hu, and M. W. Deem. Phys. Rev. E **73**, 041913 (2006). 41. V. Ortega-Cejas, J. Fort and V. Mendez, Ecology **85**, 258 (2004). 42. E. Eisenberg and E. Y. Levanon, Trends Genet. **19**, 362 (2003). 43. P. A. Futreal, L. Coin, M. Marshall, T. Down, T. Hubbard, Richard Wooster, N. Rahman and M. R. Stratton, Nature Rev. Cancer **4**, 177 (2004). 44. I. Gonzalez-Garcia, R. V. Solé and J. Costa, Proc. Natl. Acad. Sci. USA **99**, 13085 (2001). 45. A. Marusyk, V. Almendro, K. Polyak, Nat. Rev. Cancer **12**, 323 (2012). 46. R. V. Solé, C. Rodriguez-Caso, T. Deisboeck, and J. Saldanya, J. Theor. Biol. **253**, 629 (2008). 47. N. Isern, J. Fort and J. Pérez-Losada, J. Stat Mech: Theor Exp **10**, P10012 (2008). 48. M. Scheffer, *Critical transitions in Nature and Society* (Princeton U Press, New York 2009); R. Solé, *Phase Transitions* (Princeton U Press, New York, 2011). 49. K. R. Swanson, C. Bridge, J. D. Murray, and J. E. C. Alvord, J. Neurol. Sci. **216**, 1 (2003). 50. K. R. Swanson, J. E. C. Alvord, and J. D. Murray, Math. Comp. Modell. **37**, 1177 (2003). 51. Note that the $2$ factors in Eq. (22) provide the necessary corrections in case the arguments of the Dirac delta functions in Eq. (22) coincide with the integration limits in Eqs. (12) or (24). On the one hand, the factor $2$ appearing explicitly in Eq. (22) is necessary for the dispersal kernel to satisfy the condition $\int_{-\infty}^{0}\omega (\Delta _{\mu })d\Delta _{\mu }=1$ (see Section V). For an analogous reason we consider the factor $2$ in $\gamma(\mu_{disp})$, so that cells at $c(\mu_{disp},t+1)$ receive the proper fraction \[i.e., $(1-p_{e})$ instead of $(1-p_{e})/2$\] from the parent cells at $c(0,t)$. 52. When numerically solving the integral model, the position of the edge of the front can be obtained for each generation (time step). The numerical solutions for the front speed (Fig. 5) have been obtained by linear regression of the time-dependent position of the front. For proper comparison, all the analytical solutions and the numerical results in Fig. 5 have been computed at transient times such that the edge of the front is close to $\mu =0.01$. 53. E. S. Claudino, M. L. Lyra and I. Gleria, Phys. Rev E **87**, 032711 (2013). 54. R. P. Garay and R. Lefever, J. Theor. Biol. **73**, 417 (1978).

ArXiv

--- author: - 'M. Mladenović' - 'E. Roueff' date: 'Received February 28th, 2014; accepted April 14th, 2014' title: 'Ion-molecule reactions involving HCO$^+$ and N$_2$H$^+$: Isotopologue equilibria from new theoretical calculations and consequences for interstellar isotope fractionation ' --- [[ We revisit with new augmented accuracy the theoretical dynamics of basic isotope exchange reactions involved in the $^{12}$C/$^{13}$C, $^{16}$O/$^{18}$O, and $^{14}$N/$^{15}$N balance because these reactions have already been studied experimentally in great detail. ]{} ]{} [Electronic structure methods were employed to explore potential energy surfaces, full-dimensional rovibrational calculations to compute rovibrational energy levels that are numerically exact, and chemical network models to estimate the abundance ratios under interstellar conditions.]{} [New exothermicities, derived for HCO$^+$ reacting with CO, provide rate coefficients markedly different from previous theoretical values in particular at low temperatures, resulting in new abundance ratios relevant for carbon chemistry networks. [ In concrete terms, we obtain a reduction in the abundance of H$^{12}$C$^{18}$O$^+$ and an increase in the abundance of H$^{13}$C$^{16}$O$^+$ and D$^{13}$C$^{16}$O$^+$. ]{} In all studied cases, the reaction of the ion with a neutral polarizable molecule proceeds through the intermediate proton-bound complex found to be very stable. [ For the complexes [OCH$^+\cdots$CO]{}, [OCH$^+\cdots$OC]{}, [COHOC$^+$]{}, [N$_2 \cdots$HCO$^+$]{}, [N$_2$H$^+\cdots$OC]{}, and [N$_2$HN$_2^+$]{}, we also calculated vibrational frequencies and dissociation energies. ]{} ]{} [The linear proton-bound complexes possess sizeable dipole moments, which may facilitate their detection.]{} Introduction {#sec:intro} ============ Isotopic fractionation reactions have already been invoked by [@watson76a] and [@dalgarno76] to explain the enrichment of heavy isotopes of molecules in dark cold interstellar cloud environments. The exothermicity involved in the isotopic exchange reaction directly depends on the difference of the zero-point energies (ZPE) between the two isotopes, if one assumes that the reaction proceeds in the ground-rovibrational states of both the reactant and product molecule. This assumption has been questioned for the reaction H$_3^+$+HD$\rightleftharpoons$H$_2$D$^+$+H$_2$, where some rotational excitation in H$_2$ may reduce the efficiency of the reverse reaction [@pagani92; @hugo09]. In this paper we revisit some fractionation reactions involved in the $^{12}$C/$^{13}$C, $^{16}$O/$^{18}$O, and $^{14}$N/$^{15}$N balance by reinvestigating the potential energy surfaces involved in the isotopic exchange reactions. Within the Born-Oppenheimer approximation, a single nuclear-mass-independent potential energy surface (PES) is considered for all isotopic variants of molecules under consideration. The nuclear motions are introduced subsequently and isotopologues, molecules of different isotopic compositions and thus different masses, possess different rotational constants, different vibrational frequencies, and different ground-state (zero-point) vibrational energies, in other words, different thermodynamic properties [@urey47]. Differences in zero-point energies can become important under cool interstellar cloud conditions where molecules rather undergo isotopic exchange (fractionation) than react chemically. This thermodynamic effect may result in isotopologue abundance ratios (significantly) deviating from the elemental isotopic ratios. Knowledge of the abundance ratios may in return provide valuable information on molecular processes at low collision energies. As far as astrophysical models are concerned, $^{13}$C and $^{18}$O isotopic fractionation studies involving CO and HCO$^+$ [@lebourlot93; @liszt07; @roellig13; @maret13] are based on the pionneering paper by [@langer84], who referred to the experimental studies by [@smith80a] and used theoretical spectroscopic parameters for the isotopic variants of HCO$^+$ reported by [@henning77]. [@lohr98] derived the harmonic frequencies and equilibrium rotational constants for CO, HCO$^+$, and HOC$^+$ at the configuration interaction (including single and double excitations) level of theory (CISD/6-31G\*\*) and tabulated reduced partition function ratios and isotope exchange equilibrium constants for various isotope exchange reactions between CO and HCO$^+$. Surprisingly, this paper has not received much attention in the astrophysical literature, and its conclusions have never been applied. The studies of [@langer84] and [@lohr98] led to qualitatively different conclusions regarding the following fractionation reaction: $$\begin{aligned} {^{13}\mathrm{C}}{^{16}\mathrm{O}} + \mathrm{H}^{12}\mathrm{C}{^{18}\mathrm{O}}^+ \rightarrow \mathrm{H}^{13}\mathrm{C}{^{16}\mathrm{O}}^+ + {^{12}\mathrm{C}}{^{18}\mathrm{O}} + \Delta E , ~~ \label{reaction_diff} \end{aligned}$$ which was found to be endothermic with $\Delta E/k_{\mathrm{B}}=-5$ K by [@langer84] and exothermic with $\Delta E/k_{\mathrm{B}} = 12.5$ K by [@lohr98], where $k_{\mathrm{B}}$ is the Boltzmann constant. To clear up this discrepancy, we carried out numerically exact calculations for the vibrational ground state of HCO$^+$ using a potential energy surface previously developed by [@mladenovic98b]. Our calculations gave $\Delta E/k_{\mathrm{B}} = 11.3$ K for reaction (\[reaction\_diff\]), in good agreement with the harmonic value of [@lohr98]. In addition, we noticed that the $\Delta E/k_{\mathrm{B}}$ values of @henning77 for the reactions $$\begin{aligned} {^{13}\mathrm{C}}{^{16}\mathrm{O}} + \mathrm{H}^{12}\mathrm{C}{^{16}\mathrm{O}}^+ & \rightleftharpoons & \mathrm{H}^{13}\mathrm{C}{^{16}\mathrm{O}}^+ + {^{12}\mathrm{C}}{^{16}\mathrm{O}} ~~~~ \label{reaction2} \end{aligned}$$ and $$\begin{aligned} {^{12}\mathrm{C}}{^{18}\mathrm{O}} + \mathrm{H}{^{12}\mathrm{C}}{^{16}\mathrm{O}}^+ & \rightleftharpoons & \mathrm{H}{^{12}\mathrm{C}}{^{18}\mathrm{O}}^+ + {^{12}\mathrm{C}}{^{16}\mathrm{O}} ~~~~ \label{reaction3} \end{aligned}$$ were quoted as 17$\pm$1 K and 7$\pm$1 K by [@smith80a] and as 9 and 14 K by [@langer84]. Reconsidering the original values of [@henning77], we found that [@langer84] permuted the zero-point energies for H$^{13}$C$^{16}$O$^+$ and H$^{12}$C$^{18}$O$^+$ in Table 2 of their paper. From the original spectroscopic parameters of [@henning77], we derive $\Delta E/k_{\mathrm{B}} = 10.2$ K for reaction (\[reaction\_diff\]), in good agreement with our result and the result of [@lohr98]. The permutation of the zero-point vibrational energies of H$^{13}$C$^{16}$O$^+$ and H$^{12}$C$^{18}$O$^+$ affects the exothermicities and rate coefficients summarized in Table 1 and Table 3 of the paper by [@langer84]. These data are actually incorrect for all isotope fractionation reactions CO+HCO$^+$, except for $$\begin{aligned} {^{13}\mathrm{C}}{^{18}\mathrm{O}} + \mathrm{H}^{12}\mathrm{C}{^{16}\mathrm{O}}^+ & \rightleftharpoons & \mathrm{H}^{13}\mathrm{C}{^{18}\mathrm{O}}^+ + {^{12}\mathrm{C}}{^{16}\mathrm{O}} . ~~~~ \label{reaction_isto} \end{aligned}$$ The rate coefficients reported by [@langer84] are still widely used when including isotopes such as $^{13}$C and $^{18}$O into chemical (molecular) networks [@maret13; @roellig13]. With these points in mind, our goal is to provide reliable theoretical estimates for the zero-point vibrational energies first of H/DCO$^+$ and to derive proper rate coefficients for the related fractionation reactions. Our improved results for the exothermicities and rate coefficients are summarized in Tables \[table\_kelvin\] and \[table\_rate\]. [@henning77] also reported spectroscopic parameters for various isotopic variants of N$_2$H$^+$. This was our initial motivation to expand the present study to ion-molecule reactions between N$_2$H$^+$ and N$_2$. $^{15}$N fractionation in dense interstellar clouds has been first considered by [@terzieva00], who referred to the experimental information of the selected ion flow-tube (SIFT) studies at low temperatures of [@adams81]. [The reactions discussed in this paper, CO+HCO$^+$ and N$_2$+HN$_2^+$, are the most obvious candidates for isotopic fractionation. In addition, they have been studied in the laborataory, which allows a detailed discussion. A similar reaction has been invoked for CN [@milam09], but no experimental and/or theoretical information is available there. ]{} In the Langevin model, the long-range contribution to the intermolecular potential is described by the isotropic interaction between the charge of the ion and the induced dipole of the neutral. Theoretical approaches based on this standard assumption may qualitatively explain the behaviour of the association rates. However, they generally provide rate coefficients that are higher than experimental results [@langer84]. The rate coefficients for ion-molecule reactions are quite constant at higher temperatures but increase rapidly at lower temperatures. The latter feature is an indication of barrierless potential energy surfaces. The electrostatic forces are always attractive and can be experienced over large distances even at extremely low temperatures relevant for dark cloud enviroments. Short-range forces appear in closer encounters of interacting particles and may (prominently) influence the overall reaction rate. To explore the short-range effects we also undertake a study of linear proton-bound ionic complexes arising in the reactions involving HCO$^+$, HOC$^+$, and N$_2$H$^+$ with CO and N$_2$, which are common interstellar species. Our theoretical approach is described in Sect. \[sec:calculations\]. The specific aspects of the fractionation reactions of HCO$^+$ and HOC$^+$ with CO are reanalysed in Sect. \[sec:hco+\] and the fractionation reactions N$_2$H$^+$+N$_2$ in Sect. \[sec:nnh+\]. We discuss the equilibrium constants and rate coefficients of CO+HCO$^+$/HOC$^+$ in Sect. \[sec\_discussion:hco+\], providing the astrochemical implications of the new exothermicities in Sect. \[discussion:astro\]. The isotope fractionation reactions N$_2$H$^+$+N$_2$ are considered including the nuclear spin angular momentum selection rules in Sect. \[sec\_discussion:nnh+\]. The linear proton-bound cluster ions are analysed in Sect. \[discussion:pes\]. Our concluding remarks are given in Sect. \[sec:conclusion\]. Calculations {#sec:calculations} ============ The global three-dimensional potential energy surfaces (PES) developed by [@mladenovic98b] for the isomerizing system HCO$^+$/HOC$^+$ and by [@schmatz97] for the isoelectronic species N$_2$H$^+$ were used in the rovibrational calculations. These two PESs still provide the most comprehensive theoretical descriptions of the spectroscopic properties for HCO$^+$, HOC$^+$, and N$_2$H$^+$ and are valid up to the first dissociation limit. Potential energy representations recently developed by [@spirko08] and by [@huang10] reproduce the experimental fundamental transitions within 11\[6\] and 4\[3\] [[cm$^{-1}$]{}]{} for N$_2$H$^+$\[N$_2$D$^+$\], respectively, whereas the PES of [@schmatz97] predicts the fundamental transitions for both N$_2$H$^+$ and N$_2$D$^+$ within 2 [[cm$^{-1}$]{}]{}. The rovibrational energy levels of HCO$^+$/HOC$^+$ and N$_2$H$^+$ are calculated by a numerically exact quantum mechanical method, involving no dynamical approximation and applicable to any potential energy representation. The computational strategy is based on the discrete variable representation of the angular coordinate in combination with a sequential diagonalization/truncation procedure [@mladenovic90; @mladenovic98b]. For both molecular systems, the rovibrational states are calculated for the total angular momentum $J=0-15$. These rovibrational energies are used to evaluate theoretical partition functions and to model rate coefficients for proton transfer reactions involving HCO$^+$ and N$_2$H$^+$. To gain a first insight into dynamical features of ion-molecule reactions, additional electronic structure calculations were carried out for linear proton-bound cluster ions of HCO$^+$, HOC$^+$, and N$_2$H$^+$ with CO and N$_2$. The PESs were scanned by means of the coupled cluster method with single and double excitations including perturbative corrections for triple excitations \[CCSD(T)\] in combination with the augmented correlation consistent triple $\zeta$ basis set (aug-cc-pVTZ). Only valence electrons were correlated. The ab initio calculations were carried out with the MOLPRO [@MOLPRO_brief] and CFOUR [@CFOUR_brief] quantum chemistry program packages. Results {#sec:results} ======= The PES of [@mladenovic98b] provides a common potential energy representation for the formyl cation, HCO$^+$, and the isoformyl cation, HOC$^+$, where the local HOC$^+$ minimum is 13878 cm$^{-1}$ (166 kJ mol$^{-1}$) above the global HCO$^+$ minimum. Inclusion of the zero-point energy reduces this separation by 640–650 [[cm$^{-1}$]{}]{} for the hydrogen-containing isotopologues and by 570–580 [[cm$^{-1}$]{}]{} for the deuterium variants. The angular motion is described by a double-minimum anharmonic potential with a non-linear saddle point at 26838 [[cm$^{-1}$]{}]{} (321 kJ mol$^{-1}$) above the HCO$^+$ minimum, such that low-lying states of HCO$^+$ and HOC$^+$ are well separated. The potential energy surface of [@schmatz97] for N$_2$H$^+$ (dyazenilium) has two equivalent colinear minima as a consequence of the $S_2$ permutation symmetry, separated by an isomerization barrier 17137 [[cm$^{-1}$]{}]{} (205 kJ mol$^{-1}$) above the energy of the linear geometries. Low-lying states of N$_2$H$^+$ are, thus, localized in one of the two wells. The double-well symmetry and nuclear spin symmetries are lifted for mixed nitrogen isotope forms. Reaction of CO with HCO$^+$ and HOC$^+$ {#sec:hco+} --------------------------------------- The ground-state vibrational energies calculated in this work for isotopic variants of HCO$^+$ and HOC$^+$ are collected in Table \[table\_zpe\]. There we additionally show the harmonic zero-point energy estimates of [@lohr98] [and the anharmonic values of [@martin93a] available only for three isotopologues, as well as]{} the values obtained by [@langer84] and in the present work from the spectroscopic \[CI(corr)\] parameters of [@henning77]. Our values for CO are computed at the theoretical level used to construct the potential energy surface for HCO$^+$/HOC$^+$ \[CCSD(T)/cc-pVQZ\]. The isotopologues in Table \[table\_zpe\] are arranged in order of increasing total molecular mass. For CO and H/DOC$^+$, the zero-point energies decrease as the total molecular mass increases, which is not the case for H/DCO$^+$. Inspection of the table shows that the substitution of the central atom by its heavier isotope ($^{12}$C$\rightarrow$$^{13}$C in H/DCO$^+$ and $^{16}$O$\rightarrow$$^{18}$O in H/DOC$^+$) results in a more pronounced decrease of the zero-point energy than the isotopic substitution of the terminal atom ($^{16}$O$\rightarrow$$^{18}$O in H/DCO$^+$ and $^{12}$C$\rightarrow$$^{13}$C in H/DOC$^+$). This feature shared by H/DCO$^+$ and H/DOC$^+$ in Table \[table\_zpe\] is easy to rationalize since a central atom substitution affects all three vibrational frequencies. The zero-point energy differences for the proton transfer reactions CO+HCO$^+$/HOC$^+$ are listed in Table \[table\_kelvin\]. The reactions involving the formyl cation are labelled with F and the reactions involving the isoformyl cation with I. The deuterium variant of reaction F1 is denoted by F1(D) and similar for all other reactions. The reactions F1, F2, F3, F4, F5, and F6 are numbered as 1004, 3408, 3407, 3457, 3406, and 3458 by [@langer84]. In Table \[table\_kelvin\], our results, the values rederived from the spectroscopic parameters of Henning et al. (column HKD77), and the harmonic values of Lohr (column L98) all agree within less than 5 K. These three data sets predict the same direction for all listed reactions, whereas Langer et al. (column LGFA84) reported reaction F6 as endothermic. The replacement of our theoretical values for CO by the experimental values taken from [@huber79] affects the zero-point energy differences by at most 0.4K. The general trend seen in Table \[table\_kelvin\] is that $^{13}$C is preferentially placed in H/DCO$^+$ and $^{18}$O in H/DOC$^+$. This is in accordance with Table \[table\_zpe\], showing a stronger decrease of the zero-point energy upon isotopic substitution of the central atom. The substitution of the two $^{16}$O by $^{18}$O or the two $^{12}$C by $^{13}$C has nearly no influence on the exothermicities, as seen by comparing $\Delta E$ for reactions F1, F1(D), I1, I1(D) with $\Delta E$ for reactions F2, F2(D), I2, I2(D) and silimar for reactions F3, F3(D), I3, I3(D) versus F4, F4(D), I4, I4(D). Slightly higher exothermicities appear for reactions involving deuterium. The exothermicities for the reactions with the isoformyl isomers are lower than for the reactions with the formyl forms. From the measured forward reaction $k_f$ and backward reaction $k_r$ rate coefficients, [@smith80a] calculated the experimental zero-point energy differences using $$\begin{aligned} \frac{k_f}{k_r} = K_e = e^{{\Delta E}/{k_{\mathrm{B}} T}} , \label{ratio_sa}\end{aligned}$$ where $K_e$ is the equilibrium constant. The total estimated error on $k_f$ and $k_r$ is reported to be $\pm$25% at 80K. Table \[table\_kelvin\] indicates that the new/improved theoretical values, and the experimental finding for reaction F1 agree within the experimental uncertainty. For reactions F3 and F6, we see that the theoretical results consistently predict a higher $\Delta E$ value for H$^{12}$C$^{18}$O$^+$ reacting with $^{13}$C$^{16}$O (reaction F6) than for H$^{12}$C$^{16}$O$^+$ reacting with $^{12}$C$^{18}$O (reaction F3), whereas the opposite was derived experimentally. Note that [@smith80a] reported for ${^{13}\mathrm{C}}{^{16}\mathrm{O}}$ reacting with $\mathrm{H}{^{12}\mathrm{C}}{^{18}\mathrm{O}}^+$ in addition to reaction F6 also a yield of 10% for the rearrangement channel $$\begin{aligned} {^{13}\mathrm{C}}{^{16}\mathrm{O}} + \mathrm{H}^{12}\mathrm{C}{^{18}\mathrm{O}}^+ & \rightarrow& \mathrm{H}^{13}\mathrm{C}{^{18}\mathrm{O}}^+ + {^{12}\mathrm{C}}{^{16}\mathrm{O}} . \label{reaction_diff_jos} \end{aligned}$$ The latter transformation is not of a simple proton-transfer type (but bond-rearrangement type) and must involve a more complicated chemical mechanism probably including an activation energy barrier. Reaction of N$_2$ with N$_2$H$^+$ {#sec:nnh+} --------------------------------- The zero-point vibrational energies calculated for N$_2$H$^+$ are summarized in Table \[table\_zpe\_n2h\]. In addition to the results obtained for the potential energy surface of [@schmatz97], Table \[table\_zpe\_n2h\] also provides the values we derived from the spectroscopic parameters of [@huang10] (column HVL10) and of [@henning77] (column HKD77). The values for N$_2$ are taken from [@huber79]. The zero-point energy differences are given in Table \[table\_kelvin\_n2h\]. As seen there, our results agree with the values obtained from the spectroscopic parameters of [@huang10] within 0.4 K. The reactions involving diazenylium (or dinitrogen monohydride cation) are labelled with D in Table \[table\_kelvin\_n2h\]. The $^{14}$N/$^{15}$N substitution at the central-atom position lowers the zero-point energy more than the terminal-atom substitution (Table \[table\_zpe\_n2h\]), such that $^{15}$N preferentially assumes the central position in N-N-H$^+$ in all reactions of N$_2$H$^+$ with N$_2$ in Table \[table\_kelvin\_n2h\]. The exothermicities are found to be slightly higher for the reactions involving deuterium. In Table \[table\_kelvin\_n2h\], the experimental (SIFT) results of [@adams81] are listed as given in their paper. Note, however, that the elementary isotope fractionation reactions D2 and D3 $$\begin{aligned} {^{14}\mathrm{N}}_2{\mathrm{H}}^+ + {^{15}\mathrm{N}}{^{14}\mathrm{N}} \mathrel{\mathop{\rightleftharpoons}^{k_2}_{k_{-2}}} {^{14}\mathrm{N}}{^{15}\mathrm{N}}{\mathrm{H}}^+ + {^{14}\mathrm{N}}_2, \label{reaction_n2h_2} \\ {^{14}\mathrm{N}}_2{\mathrm{H}}^+ + {^{15}\mathrm{N}}{^{14}\mathrm{N}} \mathrel{\mathop{\rightleftharpoons}^{k_3}_{k_{-3}}} {^{15}\mathrm{N}}{^{14}\mathrm{N}}{\mathrm{H}}^+ + {^{14}\mathrm{N}}_2, \label{reaction_n2h_3}\end{aligned}$$ involve common reactants, whereas reactions D4 and D5 $$\begin{aligned} {^{14}\mathrm{N}}{^{15}\mathrm{N}}{\mathrm{H}}^+ + {^{15}\mathrm{N}}_2 \mathrel{\mathop{\rightleftharpoons}^{k_4}_{k_{-4}}} {^{15}\mathrm{N}}_2{\mathrm{H}}^+ + {^{15}\mathrm{N}}{^{14}\mathrm{N}}, \label{reaction_n2h_4} \\ {^{15}\mathrm{N}}{^{14}\mathrm{N}}{\mathrm{H}}^+ + {^{15}\mathrm{N}}_2 \mathrel{\mathop{\rightleftharpoons}^{k_5}_{k_{-5}}} {^{15}\mathrm{N}}_2{\mathrm{H}}^+ + {^{15}\mathrm{N}}{^{14}\mathrm{N}}, \label{reaction_n2h_5}\end{aligned}$$ have common products. The two reaction pairs are related by the $^{14}$N$\rightarrow {^{15}}$N substitution. Using thermodynamic reasoning, it is easy to verify that the following relationship $$\begin{aligned} \frac{K_e^{(2)}}{K_e^{(3)}} = \frac{K_e^{(5)}}{K_e^{(4)}} = {K_e^{(6)}} \label{keq_relations}\end{aligned}$$ is strictly fulfilled for $K_e^{(i)}=k_i/k_{-i}$, where ${K_e^{(6)}}$ corresponds to reaction D6, $$\begin{aligned} {^{15}\mathrm{N}}{^{14}\mathrm{N}}{\mathrm{H}}^+ + {^{15}\mathrm{N}}{^{14}\mathrm{N}} \mathrel{\mathop{\rightleftharpoons}^{k_6}_{k_{-6}}} {^{14}\mathrm{N}}{^{15}\mathrm{N}}{\mathrm{H}}^+ + {^{14}\mathrm{N}}{^{15}\mathrm{N}} . \label{reaction_n2h_6} \end{aligned}$$ Note that the factor $1/K_e^{(6)}$ also provides the thermal population of $^{15}$N$^{14}$NH$^+$ relative to $^{14}$N$^{15}$NH$^+$. Discussion {#sec:discussion} ========== The equilibrium constant $K_e$ for the proton transfer reaction $$\begin{aligned} {\mathrm A} + {\mathrm{HB}} \mathrel{\mathop{\rightleftharpoons}^{k_f}_{k_r}} \mathrm{HA} + {\mathrm B} \label{reaction_proton}\end{aligned}$$ under thermal equilibrium conditions is given by $$\begin{aligned} K_e = \frac{k_f}{k_r} = \frac{Q({\mathrm{HA}})}{Q({\mathrm{HB}})} \frac{Q({\mathrm{B}})}{Q({\mathrm{A}})} , \label{equilibrium_constant_0}\end{aligned}$$ where $Q(\mathrm{X})$ is the full partition function for the species X. Making the translation contribution explicit, we obtain $$\begin{aligned} K_e = f_m^{3/2} \, \frac{Q_{\mathrm{int}}({\mathrm{HA}})}{Q_{\mathrm{int}}({\mathrm{HB}})} \frac{Q_{\mathrm{int}}({\mathrm{B}})}{Q_{\mathrm{int}}({\mathrm{A}})} e^{\Delta E/k_{\mathrm{B}} T}, \label{equilibrium_constant}\end{aligned}$$ where the mass factor $f_m$ is given by $$\begin{aligned} f_m= \frac{m({\mathrm{HA}})\, m({\mathrm{B}})}{m({\mathrm{HB}})\, m({\mathrm{A}})} \label{mass_factor}\end{aligned}$$ for $m$(X) denoting the mass of the species X, whereas $\Delta E$ stands for the zero-point energy difference between the reactants and the products, $$\begin{aligned} \Delta E = E_0^{\mathrm{HA}} + E_0^{\mathrm{B}} - E_0^{\mathrm{HB}} - E_0^{\mathrm{A}} . \label{delta_e}\end{aligned}$$ The zero-point energies $E_0$ are measured on an absolute energy scale. For isotope fractionation reactions, the internal partition function, $Q_{\mathrm{int}}$, includes only the rovibrational degrees of freedom (no electronic contribution) and is given by the standard expression $$\begin{aligned} Q_{\mathrm{int}} = g \sum_J \sum_i \left(2J+1\right) e^{-\varepsilon_i^J/k_{\mathrm{B}} T}, \label{q_int} \end{aligned}$$ where $\varepsilon_i^J = E_i^J-E_0^0$ for a total angular momentum $J$ is the rovibrational energy measured relative to the corresponding zero-point energy ($J$=0). The factor $(2J+1)$ accounts for the degeneracy relative to the space-fixed reference frame and $g$ for the nuclear spin (hyperfine) degeneracy, $$\begin{aligned} g=\Pi_{\alpha} (2I_{N,\alpha}+1), \label{nuc_spin_g}\end{aligned}$$ in which $\alpha$ labels the constituent nuclei having the nuclear spin $I_{N,\alpha}$. For the nuclei considered in the present work, we have $I_N($H$)=1/2$, $I_N($D$)=1$, $I_N(^{12}$C$)=0$, $I_N(^{13}$C$)=1/2$, $I_N(^{16}$O$)=0$, $I_N(^{18}$O$)=0$, $I_N(^{14}$N$)=1$, and $I_N(^{15}$N$)=1/2$. Introducing the ratio $$\begin{aligned} R^{\mathrm{X}}_{\mathrm{Y}} & = & \frac{Q_{\mathrm{int}}({\mathrm{X}})}{Q_{\mathrm{int}}({\mathrm{Y}})}, \label{ratio_q}\end{aligned}$$ the equilibrium constant is compactly written as $$\begin{aligned} K_e = \frac{k_f}{k_r} = F_q \, e^{\Delta E/k_{\mathrm{B}} T}, \label{equilibrium_constant_compact}\end{aligned}$$ where the partition function factor $F_q$ is $$\begin{aligned} F_q & = & f_m^{3/2} \, R^{\mathrm{HA}}_{\mathrm{HB}} \, R^{\mathrm{B}}_{\mathrm{A}} . \label{equilibrium_constant_factor}\end{aligned}$$ For reactions proceeding in the ground-rovibrational states of the reactants and the products, the partition function ratios $R_{\mathrm{HB}}^{\mathrm{HA}}$ and $R_{\mathrm{A}}^{\mathrm{B}}$ are both equal to 1. Even then the corresponding partition function factor $F_q$ of Eq. (\[equilibrium\_constant\_factor\]) is, strictly speaking, different from 1 because of the mass term $f_m$ defined by Eq. (\[mass\_factor\]). For the reactions F1–F6 in Table \[table\_kelvin\], for instance, the $f_m^{3/2}$ values are 0.998, 0.998, 0.997, 0.997, 0.995, and 1.002, respectively, which are different from 1 at most by 0.5%. [The $f_m^{3/2}$ values are computed from Eq. (\[mass\_factor\]) using the following atomic masses $m(\mathrm{H})=1.007825035$, $m(\mathrm{D})=2.014101779$, $m(^{12}\mathrm{C})=12$, $m(^{13}\mathrm{C})=13.003354826$, $m(^{16}\mathrm{O})=15.99491463$, and $m(^{18}\mathrm{O})=17.9991603$ u, as given by [@mills93]. ]{} The terms of $Q_{\mathrm{int}}$ in Eq. (\[q\_int\]) decrease rapidly with energy and J. In the low-temperature limit relevant for dark cloud conditions, the discrete rotational structure of the ground-vibrational state provides the main contribution to $Q_{\mathrm{int}}$. That said, the rotational energy cannot be treated as continuous and one must explicitly sum the terms to obtain $Q_{\mathrm{int}}$. With increasing temperature, the rotational population in the ground-vibrational state increases and other vibrational states may also become accessible, leading to partition function factors $F_q$, which may show (weak) temperature dependences. [For a given potential energy surface, numerically exact full-dimenional strategies insure the determination of accurate level energies and therefrom accurate partition functions and equilibrium constants. To predict/estimate rate coefficients, we may use kinetic models, such as e.g. the Langevin collision rate model for ion-molecule reactions. Uncertainties in the rate coefficients are thus defined by uncertainties in the model parameters. In the case of the system CO+HCO$^+$, we employ the total rate coefficients from Table 3 of [@langer84] and the uncertainties of these quantities also provide the uncertainties of the rate coefficients derived in the present work. ]{} Reaction of CO with HCO$^+$ {#sec_discussion:hco+} --------------------------- The equilibrium constants for HCO$^+$ reacting with CO are given in Table \[table\_rate\]. Our $K_e$ values are obtained in accordance with Eq. (\[equilibrium\_constant\_compact\]) by direct evaluation of the internal partition functions $Q_{\mathrm{int}}$ from the computed rovibrational energies. The forward reaction $k_f$ and backward reaction $k_r$ rate coefficients are calculated using our $\Delta E$ values and the total temperature-dependent rate coefficients $k_{\mathrm{T}}$ given by [@langer84], where $$\begin{aligned} k_{\mathrm{T}} = k_f+k_r , \label{rate_langer}\end{aligned}$$ such that $$\begin{aligned} k_f & = & k_{\mathrm{T}} K_e/(K_e+1), \label{kf_langer}\\ k_r & = & k_{\mathrm{T}} /(K_e+1) . \label{kr_langer}\end{aligned}$$ The results for the deuterium variants are also listed in Table \[table\_rate\]. Their rate coefficients $k_f$ and $k_r$ are calculated assuming the same total rate coefficients $k_{\mathrm{T}}$ as for the H-containing forms (due to nearly equal reduced masses). For the purpose of comparison, note that the Langevin rate for CO+HCO$^+$ is $k_{\mathrm{L}} = 8.67 \times 10^{-10}$ cm$^3$s$^{-1}$. The partition function factors $F_q$ deviate from 1 by approximately 2% in Table \[table\_rate\]. They also exhibit marginal temperature dependences. This reflects the influence of rotational and vibrational excitations in the reactants and the products. Only the rotationally excited ground-vibrational states contribute to $Q_{\mathrm{int}}$ at temperatures $T < 200$ K. The contribution of the bending $\nu_2$ level is 0.5% at 200 K and 3.6–3.8% at 300 K, whereas the contributions from 2$\nu_2$ are 0.1% at 300 K. To appreciate the effect of $F_q$, we employed the rate coefficients measured at 80 K by [@smith80a] to determine the $\Delta E$ value for reactions F1, F3, and F6 by means of Eq. (\[equilibrium\_constant\_compact\]). Using the $F_q$ values from Table \[table\_rate\], we obtain $\Delta E/k_{\mathrm{B}}$ of 13.8, 15.1, and 4.8 K, respectively. For $F_q=1$, we find 12.3, 14.6, and 3.8 K, which are lower by 1.5 K (12%), 0.5 K (3%), and 1 K (26%) than the former $F_q \ne 1$ results. The equilibrium constants $K_e$ reported by [@langer84] deviate from the present results and those of [@lohr98] very prominently at low temperatures in Table \[table\_rate\]. At 10 K, we see deviations of 43% and 217% with respect to our values for reactions F1 and F3, respectively, and the related rate coefficients $k_f, k_r$ are accordingly different. An even larger discrepancy is seen for reaction F6 of Eq. (\[reaction\_diff\]), which was previously predicted to be endothermic. In accordance with this, the values of $k_f$ and $k_r$ derived by [@langer84] are given in the reverse positions as $(k_r,k_f)$ for reaction F6 in Table \[table\_rate\]. The deuterium variants in Table \[table\_rate\] are associated with slightly lower $F_q$ values and somewhat higher low-temperature $K_e$, resulting in somewhat faster foward reactions and slower backward reactions. Astrochemical implications {#discussion:astro} -------------------------- We investigated the role of these new derived exothermicities under different density conditions relevant to cold dark interstellar clouds. We display in Table \[tab:result\] steady-state results for isotopic ratios of CO, HCO$^+$ and DCO$^+$ for two chemical models performed at a temperature of 10 K with a cosmic ionization rate $\zeta$ of 1.3 $\times$ 10$^{-17}$ s$^{-1}$ per H$_2$ molecule with the old $\Delta E$ values by Langer et al. (Model A: LGFA84) and the present $\Delta E$ values listed in Table \[table\_kelvin\] (Model B). The ratios of the principal isotope to the minor isotope obtained for Model A, $R_{\mathrm{A}}$, and for Model B, $R_{\mathrm{B}}$, are compared using the relative difference $\delta$, $$\begin{aligned} \delta = 1-{R_{\mathrm{B}}}/{R_{\mathrm{A}}}.\end{aligned}$$ The chemical network contains 288 chemical species including $^{13}$C and $^{18}$O containing molecules as well as deuterated species and more than 5000 reactions. We assumed that the elemental $^{12}$C/$^{13}$C and $^{16}$O/$^{18}$O isotopic ratios are 60 and 500, so that any deviation relative to these values measures the amount of enrichment/depletion with respect to the elemental ratios. For the $^{13}$C$^{18}$O-containing molecules the value of 30000 is the reference. [The zero-point energies of other isotopic substitutes do not pose any problem because the reactions involved in the interstellar chemical networks are significantly exothermic and the solutions of the chemical equations are independent of these quantities. ]{} The isotopic fractionation reactions are introduced explicitly in the chemical network, whereas the other reactions involving isotopologues are built automatically from the reactions involving the main isotope in the chemical code. The adopted method has first been presented in [@lebourlot93], where statistical arguments were used to derive the various branching ratios in the chemical reactions. The procedure is limited to three carbon-containing molecules (oxygen-containing molecules have a maximum of two oxygen atoms in our chemical network) and does not disitinguish between C$^{13}$CC- or $^{13}$CCC-containing species. A similar approach has recently been applied by [@roellig13] for photon-dominated region models. However, [@roellig13] used the old (LGFA84) exothermicity values. We also explicitly introduce the relation given by [@langer84] that $k_f + k_r = k_{\mathrm{T}}$. The forward reaction $k_f$ and reverse reaction $k_r$ rate coefficients involved in the isotopic exchange reaction are then evaluated from the total rate coefficient $k_{\mathrm{T}}$ as follows $$\begin{aligned} k_f = k_{\mathrm{T}} \frac{1}{1+\exp(-\Delta E/k_{\mathrm{B}} T)} \label{evelyne1}\end{aligned}$$ and $$\begin{aligned} k_r = k_{\mathrm{T}} \frac{\exp(-\Delta E/k_{\mathrm{B}} T)}{1+\exp(-\Delta E/k_{\mathrm{B}} T)} . \label{evelyne2}\end{aligned}$$ These expressions have also been included in the study of fractionation in diffuse clouds presented by [@liszt07]. The results summarized in Table \[tab:result\] show that CO/$^{13}$CO has the elemental value, whereas rarer isotopologues are very slightly depleted. The results for Models A and B are also very similar because no differences were used for the reaction rate coefficients between $^{13}$C$^+$ and CO. However, more significant are the differences for the results for the isotopic ratio of HCO$^+$, which directly arise from the variations of the exothermicities found in the present work. We also introduced a fractionation reaction for the deuterated isotope, whose rotational frequencies have been measured in the laboratory [@caselli05] and are detected in the interstellar medium [@guelin82; @caselli02]. As the exothermicity of the deuterated isotopologues is somewhat higher, the isotopic $^{13}$C ratio is somewhat lower than in the hydrogenic counterpart. The general trend seen in Table \[tab:result\] is that the new Model B predicts lower fractional abundances $x$ for H$^{12}$C$^{16}$O$^+$ (up to 2%) and D$^{12}$C$^{16}$O$^+$ (up to 5%), lower relative abundances $R_{\mathrm{B}}$ of H$^{12}$C$^{18}$O$^+$ (7–21%), and higher relative abundances of the $^{13}$C-containing isotopologues (up to 40% for the hydrogenic forms and up to 75% for the deuterated forms) than Model A. Reaction of N$_2$ with N$_2$H$^+$ {#sec_discussion:nnh+} --------------------------------- Molecular nitrogen is a homonuclear diatomic molecule with a $X {^1}\Sigma_g^+$ ground-electronic state with the three naturally occurring isotopologues: $^{14}$N$_2$, $^{14}$N$^{15}$N, and $^{15}$N$_2$. Whereas $^{14}$N is a spin-1 boson, $^{15}$N is a spin-1/2 fermion, such that the two symmetric forms $^{14}$N$_2$ and $^{15}$N$_2$ follow different nuclear spin statistics. In the states with a higher nuclear spin degeneracy (ortho states), we have $g=(I_N+1)(2 I_N + 1)$, whereas $g=I_N(2 I_N + 1)$ holds for the states with lower nuclear spin degeneracy (para states). To properly account for this effect, we evaluated the internal partition functions separately for even and odd $J$ values, $$\begin{aligned} Q_{\mathrm{\mathrm{evenJ}}} & = & {\sum_{J=0}}' \sum_i (2J+1) \, e^{-\varepsilon_i^J/k_{\mathrm{B}} T} , \\ Q_{\mathrm{\mathrm{oddJ}}} & = & {\sum_{J=1}}' \sum_i (2J+1) \, e^{-\varepsilon_i^J/k_{\mathrm{B}} T} , ~~~\end{aligned}$$ where $\Sigma'$ denotes summation in steps of 2. Multiplying each term by the appropriate nuclear spin (hyperfine) degeneracy factor, we obtain the partition function for N$_2$ as $$\begin{aligned} Q_{\mathrm{int}}(^{14}\mathrm{N}_2) & = & 6 Q_{\mathrm{evenJ}} + 3 Q_{\mathrm{oddJ}} , \label{n14} \\ Q_{\mathrm{int}}(^{15}\mathrm{N}_2) & = & 3 Q_{\mathrm{oddJ}} + Q_{\mathrm{evenJ}} . \label{n15}\end{aligned}$$ $^{14}$N$^{15}$N is not a homonuclear diatomic molecule, such that $$\begin{aligned} Q_{\mathrm{int}}(^{14}\mathrm{N}^{15}\mathrm{N}) & = & 6 \left( Q_{\mathrm{evenJ}} + Q_{\mathrm{oddJ}} \right) .\end{aligned}$$ The equilibrium constants $K_e$ and rate coefficients for the isotopic variants of N$_2$H$^+$ reacting with N$_2$ are shown in Table \[table\_rate\_n2h\]. There we assumed the total rate coefficient $k_{\mathrm{T}}$ given by the Langevin collision rate (in SI units) $$\begin{aligned} k_{\mathrm{L}} & = & e \sqrt{{\pi \, \alpha(\mathrm{N}_2)}/{\mu_R \varepsilon_0}} , \label{langevin}\end{aligned}$$ where $e$ is the elementary charge, $\mu_R$ the reduced mass for the collision, and $\alpha(\mathrm{N}_2)$ the polarizability of N$_2$ \[$\alpha(\mathrm{N}_2)=1.710$ [Å]{}$^3$ [@olney97]\], giving thus $k_{\mathrm{T}}=k_{\mathrm{L}} = 8.11 \times 10^{-10}$ cm$^3$s$^{-1}$. The rate coefficients $k_f$ and $k_r$ are determined from $k_{\mathrm{T}}$ and $K_e$ with the help of Eqs. (\[kf\_langer\]) and (\[kr\_langer\]), respectively. Spectroscopic parameters of [@trickl95] and [@bendtsen01] were used for the $X {^1\Sigma}^+_g$ states of $^{14}$N$_2$, $^{14}$N$^{15}$N, and $^{15}$N$_2$. The nuclear spin degeneracy affects the equilibrium constants for the reactions involving either $^{14}$N$_2$ or $^{15}$N$_2$. At higher temperatures, $K_e$ in Table \[table\_rate\_n2h\] approaches 1/2 for reactions D2 and D3 having ${^{14}\mathrm{N}}_2$ as a product, and 2 for reactions D4 and D5 having ${^{15}\mathrm{N}}_2$ as a reactant. For reaction D1, $$\begin{aligned} {^{14}\mathrm{N}}_2{\mathrm{H}}^+ + {^{15}\mathrm{N}}_2 \mathrel{\mathop{\rightleftharpoons}^{k_1}_{k_{-1}}} {^{15}\mathrm{N}}_2{\mathrm{H}}^+ + {^{14}\mathrm{N}}_2, \label{reaction_n2h_1}\end{aligned}$$ the effects from nuclear spin statistics cancel out and $K_e^{(1)} \rightarrow 1$ as the temperature increases. From Eqs. (\[n14\]) and (\[n15\]), the ortho-to-para ratio is given by $R_{14}=6 Q_{\mathrm{evenJ}}/3 Q_{\mathrm{oddJ}}$ for $^{14}$N$_2$ and by $R_{15}=3 Q_{\mathrm{oddJ}}/ Q_{\mathrm{evenJ}}$ for $^{15}$N$_2$. We may note that $R_{14}$ assumes a value of 2.41 (2.01) and $R_{15}$ a value of 2.60 (2.99) at 5 K (10 K). At high temperature equilibrium, we have $R_{14}=2$ and $R_{15}=3$. [@adams81] employed normal nitrogen (ratio 2:1 of ortho vs para $^{14}$N$_2$) in the SIFT experimental study of N$_2$H$^+$+N$_2$. To measure the forward reaction and backward reaction rate coefficients at a given temperature, they interchanged the ion-source gas and reactant gas. Using mass-selected samples, these authors, however, were unable to distinguish between the isotopomers $^{14}$N$^{15}$NH$^+$ and $^{15}$N$^{14}$NH$^+$, such that their results provide the overall yield of these cations (no information on the relative yields). This applies to the competing reactions D2 and D3 on one side and the competing reactions D4 and D5 on the other side. $^{14}$N$^{15}$NH$^+$ and $^{15}$N$^{14}$NH$^+$ are expected to be differently fractionated (see Table \[table\_kelvin\_n2h\]). To simulate the experimental conditions of [@adams81], we introduced the overall forward $k_{23}$ and overall reverse $k_{-23}$ rate coefficients for reactions D2 and D3, $$\begin{aligned} k_{23} & = & k_2 + k_3, \label{rate_23_f}\\ k_{-23} & = & \left[ k_{-2} {K_e^{(6)}} + k_{-3} \right] \frac{1}{1+K_e^{(6)}} , \label{rate_23_r}\end{aligned}$$ and the overall forward $k_{45}$ and overall reverse $k_{-45}$ rate coefficients for reactions D4 and D5, $$\begin{aligned} k_{-45} & = & k_{-4} + k_{-5}, \label{rate_45_r} \\ k_{45} & = & \left[ k_4 K_e^{(6)} + k_5 \right] \frac{1}{1 + K_e^{(6)}} . \label{rate_45_f} \end{aligned}$$ Here we explicitly assumed an equilibrium distribution between $^{14}$N$^{15}$NH$^+$ and $^{15}$N$^{14}$NH$^+$. The term $1+K_e^{(6)}$ is the state-distribution normalization factor. 0.25cm 0.25cm 0.25cm The variation of the rate coefficients with the temperature is displayed in Fig. \[fig.rate\]. The common feature seen there is that the forward reaction becomes faster and the backward reaction slower with decreaseing temperature. We also see that $k_f$ and $k_r$ exhibit a very weak temperature dependence for $T>50$ K. For reactions D1 and D6, $k_f$ and $k_r$ approach the same value ($k_{\mathrm{L}}/2$ in our model) at higher temperatures in Fig. \[fig.rate\](a) and Table \[table\_rate\_n2h\]. The high temperature limits of $k_i$ and $k_{-i}$ for $i=2-5$ are, however, different because of the nuclear spin restrictions, as clearly seen in Figs. \[fig.rate\](b) and \[fig.rate\](c). The reverse rate coefficients $k_{-2}$ and $k_{-3}$ in Fig. \[fig.rate\](b) become even higher than $k_{2}$ and $k_{3}$ for $T>14.7$ K and $T>3.4$ K, respectively, inverting thus the reaction direction. [@herbst03] also found $k_{-3} > k_3$ at $T=10$ K for reaction D3 assuming a different rate-coefficient model. For the overall state-averaged rate coefficients in Fig. \[fig.rate\] and Table \[table\_rate\_n2h\], we have $k_{23} > k_{-23}$ and $k_{45}>k_{-45}$ for all temperatures shown. This is in accordance with the SIFT experiment of [@adams81]. The rate coefficients $k_{\pm23}$ and $k_{\pm 45}$ appear 30% higher than the experimental finding, reported with an error of $\pm$25% at 80K. Note, however, that the ratios $k_{23}/k_{-23}$ and $k_{45}/k_{-45}$ agree within 6% with the corresponding experimental values. Due to the nuclear spin angular momentum selection rules, the high-temperature limits (for $K_e^{(6)} \rightarrow 1$) of $k_{\pm 23}$ and $k_{\pm 45}$ are different from the high-temperature limits of $k_{\pm 1}$ and $k_{\pm 6}$. From Eqs. (\[rate\_23\_f\])-(\[rate\_45\_f\]) and the relationship of Eq. (\[keq\_relations\]), we easily obtain $$\begin{aligned} \frac{k_{23}}{k_{-23}} & = & {K_e^{(2)}}+{K_e^{(3)}}\end{aligned}$$ and $$\begin{aligned} \frac{k_{45}}{k_{-45}} & = & \frac{{K_e^{(4)}} {K_e^{(5)}}}{{K_e^{(4)}}+{K_e^{(5)}}} .\end{aligned}$$ Following the procedure of [@adams81], we may model the temperature dependence of the latter ratios as $e^{\Delta E_{ij}/k_{\mathrm{B}}T}$ \[compare with Eq. (\[ratio\_sa\])\]. Using our results from Table \[table\_rate\_n2h\] for the overall forward and overall reverse rate coefficients calculated at the temperatures of the SIFT experimental study, $T=80$ K and $T=292$ K, we derive $\Delta E_{23}, \Delta E_{45}=6.5$ K for both reaction pairs. [@adams81] estimated the zero-point energy difference of $9 \pm 3$ K for reactions D2 and D4 (see Table \[table\_kelvin\_n2h\]). In accordance with the analysis presented here, we see, however, that the results of [@adams81] should be attributed to the reaction pairs $\{$D2,D3$\}$ and $\{$D4,D5$\}$. This also explains a large discrepancy seen in Table \[table\_kelvin\_n2h\] between the theoretical estimates and experimental finding for reaction D4. In recent studies of [@bizzocchi10; @bizzocchi13], $^{14}$N$^{15}$NH$^+$ and $^{15}$N$^{14}$NH$^+$ were both detected in a prototypical starless core L1544 of low central temperature and an abundance ratio $R^{14,15}_{15,14}=[{^{14}\mathrm{N}}{^{15}\mathrm{N}}\mathrm{H}^+]/[{^{15}\mathrm{N}}{^{14}\mathrm{N}}\mathrm{H}^+]$ of $1.1 \pm 0.3$ was derived. [Note that the ratio $R^{14,15}_{15,14}$ correlates with $K_e^{(6)}$ describing reaction D6 of Eq. (\[reaction\_n2h\_6\]). As seen in Table \[table\_rate\_n2h\], we obtain $K_e^{(6)}$ of 1.22-1.02 for $T=40-292$ K; the additional calculation at T=30 K gave $R^{14,15}_{15,14}=1.31$. Also note that the earlier model of [@rodgers04] has led to the ratio $R^{14,15}_{15,14}$ of 1.8–2.3, which correlates with our $K_e^{(6)}$ value of 2.27 (1.72) at $T=10$K (15K). ]{} Ionic complexes {#discussion:pes} --------------- Ion-molecule reactions were additionally examined using electronic structure calculations, carried out for the linear approach of the neutral CO and N$_2$ to the linear cations HCO$^+$, HOC$^+$, and N$_2$H$^+$. The corresponding minimum-energy paths (MEPs) are displayed in Fig. \[fig.mep\]. The MEPs are obtained optimizing three intramolecular distances for various monomer separations. Our calculations were performed at the CCSD(T)/aug-cc-pVTZ level of theory [employing the standard MOLPRO and CFOUR optimization/threshold parameters. ]{} The lower MEP in Fig. \[fig.mep\](c) is related to the reaction $$\begin{aligned} {\mathrm{N}}_2{\mathrm{H}}^+ + \mathrm{CO} \rightarrow \mathrm{HCO}^+ + \mathrm{N}_2, && \label{reaction_n2h_destruction}\end{aligned}$$ which is considered to be the main destruction path for N$_2$H$^+$ when CO is present in the gas phase at standard abundances \[CO\]/\[H$_2$\] $\sim 10^{-4}$ [@snyder77; @jorgensen04]. For this reaction, [@herbst75] reported a rate coefficient of 8.79$\times$10$^{-10}$ cm$^3$ s$^{-1}$ at 297$\pm$2 K. No reverse reaction was detected [@anicich93]. For reactions involving HOC$^+$, [@freeman87] measured a rate coefficient $k$ of 6.70$\times$10$^{-10}$ cm$^3$ s$^{-1}$ for the following reaction $$\begin{aligned} \mathrm{HOC}^+ + \mathrm{CO} \rightarrow {\mathrm{HCO}}^+ + \mathrm{CO} , \label{reaction_hoc.2}\end{aligned}$$ whereas [@wagner-redeker85] reported $k$ as 6.70$\times$10$^{-10}$ cm$^3$ s$^{-1}$ for $$\begin{aligned} \mathrm{HOC}^+ + \mathrm{N}_2 \rightarrow {\mathrm{N}}_2{\mathrm{H}}^+ + \mathrm{CO} . \label{reaction_hoc}\end{aligned}$$ The Langevin collision rate is $k_{\mathrm{L}} = 8.67 \times 10^{-10}$ cm$^3$s$^{-1}$ for reactions (\[reaction\_n2h\_destruction\]) and (\[reaction\_hoc.2\]) involving CO and $k_{\mathrm{L}} = 8.11 \times 10^{-10}$ cm$^3$s$^{-1}$ for reaction (\[reaction\_hoc\]) involving N$_2$. The common feature in Fig. \[fig.mep\] is the formation of a linear proton-bound ionic complex, which is 2000-7000 [[cm$^{-1}$]{}]{} more stable than the separated monomers. The properties of the complexes are summarized in Table \[table\_freq\], where we give the geometric parameters $r_i$, the equilibrium rotational constants $B_e$, the harmonic wavenumbers $\omega_i$ for the main and deuterated isotopologues, and the harmonic zero-point energies $E_0$. The corresponding results for the constituent monomers are listed in Table \[table\_freq-monomers\]. Note that the monomer values $E_0$ in Table \[table\_freq-monomers\] are harmonic and therefore different from the anharmonic results of Table \[table\_zpe\]. The coordinates $r_i (i=1-4)$ for A–B–H–C–D denote $r_1=r(\mathrm{A-B}), r_2=r(\mathrm{B-H}), r_3=r(\mathrm{H-C})$ and $r_4=r(\mathrm{C-D})$ in Table \[table\_freq\] and similar in Table \[table\_freq-monomers\]. The dipole moments $\mu_z$ and the quadrupole moments $\Theta_{zz}$ in Table \[table\_freq\] and \[table\_freq-monomers\] are given with respect to the inertial reference frame with the origin in the complex centre of mass, where the position of the first atom A of A–B–H–C–D or A–B–C along the $z$ axis is chosen to be the most positive. The ionic complexes N$_2$HN$_2^+$ and COHOC$^+$ have linear centrosymmetric equilibrium structures. The complex OCH$^+$$\cdots$CO is asymmetric with a barrier height to the centrosymmetric saddle point OCHCO$^+$(TS), seen at 358 [[cm$^{-1}$]{}]{} in Fig. \[fig.mep\](a). In the mixed-cluster ions, the proton is bound either to CO, when N$_2\cdots$HCO$^+$ is formed, or to N$_2$, when N$_2$H$^+$$\cdots$OC is formed. Comparison of Tables \[table\_freq\] and \[table\_freq-monomers\] shows that the geometric parameters experience prominent changes (up to 0.01–0.02 Å) upon complexation. In this fashion, the ionic (molecular) complexes differ from van der Waals complexes, in which the monomers preserve their geometric parameters to a great extent. The transformations in Fig. \[fig.mep\] are all of the proton transfer type. The neutral CO may approach H$^+$ of the triatomic cation either with C or O since both C and O possess lone electron pairs. The proton attachment from the C side leads to a more stable complex. As seen in Fig. \[fig.mep\](a), the complex OCH$^+$$\cdots$OC is 1785 [[cm$^{-1}$]{}]{} above OCH$^+$$\cdots$CO and 9615 [[cm$^{-1}$]{}]{} below COHOC$^+$. We also see that N$_2$$\cdots$HCO$^+$ is 6996 [[cm$^{-1}$]{}]{} more stable than N$_2$H$^+$$\cdots$OC. In all cases, the energy separation between the HCO$^+$- and HOC$^+$-containing complexes is smaller than the separation between free HCO$^+$ and HOC$^+$, seen to be 13820 [[cm$^{-1}$]{}]{} in Fig. \[fig.mep\]. The results of Fig. \[fig.mep\] are consistent with the fact that the proton tends to localize on the species with higher proton affinity. The experimental proton affinity is 594 kJ mol$^{-1}$ (49654 cm$^{-1}$) for CO on the C end and 427 kJ mol$^{-1}$ (35694 cm$^{-1}$) for CO on the O end [@freeman87]. The experimental proton affinity of 498 kJ mol$^{-1}$ (41629 cm$^{-1}$) was determined for N$_2$ [@ruscic91]. The harmonic wavenumbers for the ionic complexes occurring in the course of reactions F1–F6 are provided in Table \[table\_freq-complex\]. In addition to the spectroscopic properties, we also give the harmonic zero-point energies of the complexes $E_0$, the reactants $E_0^r$, and the products $E_0^p$, as well as the dissociation energies including the harmonic zero-point energy correction in the direction of the reactants, $D_0^r=D_e+E_0^r-E_0$, and in the direction of the products, $D_0^p=D_e+E_0^p-E_0$, where $D_e$ is the classical dissociation energy. In Table \[table\_freq-complex\], the vibrational mode $\omega_2$, which is predominantly the diatom CO stretching vibration, is the most sensitive to isotopic substitutions. Compared with $\omega$ of free CO, $\omega_2$ exhibits a blue-shift of 93 [[cm$^{-1}$]{}]{} for the main isotopologue (Table \[table\_freq-monomers\] vs. Table \[table\_freq\]). The modes $\omega_1$ and $\omega_3$, highly sensitive to the H$\rightarrow$D substitution (Table \[table\_freq\]), can be considered as the H-C-O stretching modes. The intermolecular stretching mode is $\omega_8$. The zero-point-corrected dissociation energies in Table \[table\_freq-complex\] are approximately 240 [[cm$^{-1}$]{}]{} lower than the electronic dissociation energy of 4876 cm$^{-1}$ (Fig. \[fig.mep\]). The harmonic $\Delta E^h/k_{\mathrm{B}}$ values in Table \[table\_freq-complex\] and anharmonic $\Delta E/k_{\mathrm{B}}$ values in Table \[table\_kelvin\] agree within 0.5 K. The proton-bound complexes OCH$^+\cdots$CO and N$_2 \cdots$HCO$^+$ have large dipole moments $\mu_e$ of 2.94 D and 3.53 D (Table \[table\_freq\]). For OCH$^+\cdots$CO, the most intense infrared transitions are expected for $\omega_3$ (with harmonic intensity $I^h_3$ of 2440 km mol$^{-1}$) and $\omega_1$ ($I^h_1=536$ km mol$^{-1}$), whereas the intermolecular stretch $\omega_8$ has $I^h_8=232$ km mol$^{-1}$. The fundamental (anharmonic) transitions $(\nu_1,\nu_2,\nu_3,\nu_{4,5},\nu_{6,7},\nu_8,\nu_{9,10})$ are calculated to be (2267, 2236, 1026, 1136, 346, 186, 208) for the main isotopologue (in [[cm$^{-1}$]{}]{}). The most intense infrared active transitions for N$_2 \cdots$HCO$^+$ are $\omega_1$ ($I^h_1=1\,034$ km mol$^{-1}$), $\omega_3$ ($I^h_3=814$ km mol$^{-1}$), and $\omega_8$ ($I^h_8=154$ km mol$^{-1}$). For this complex, the fundamental vibrational $(\nu_1,\nu_2,\nu_3,\nu_{4,5},\nu_{6,7},\nu_8,\nu_{9,10})$ transitions are determined to be (2357, 2321, 1876, 1045, 127, 186, 113) (in [[cm$^{-1}$]{}]{}). The anharmonic transitions are calculated from the cubic and semi-diagonal quartic force field in a normal coordinate representation by means of vibrational second-order perturbation theory, as implemented in CFOUR [@CFOUR_brief]. [ Regarding the CCSD(T)/aug-cc-pVTZ method used here, we may note that our value of 358 [[cm$^{-1}$]{}]{} in Fig. \[fig.mep\](a) for the barrier height of OC+HCO$^+$ agrees reasonably well with previous theoretical results of 382 [[cm$^{-1}$]{}]{} \[the CCSD(T)/cc-pVQZ approach of [@botschwina01]\] and 398 [[cm$^{-1}$]{}]{} \[the CCSD(T)/aug-cc-pVXZ approach of [@terrill10] at the complete basis-set limit\]. A classical dissociation energy was previously determined to be 4634 [[cm$^{-1}$]{}]{} for OCH$^+\cdots$CO and 5828 [[cm$^{-1}$]{}]{} for N$_2$HN$_2^+$ at the complete basis-set limit [@terrill10].]{} The use of larger basis sets would ultimately be needed for converging theoretical results to stable values. [Our primary goal here is the acquisition of first information relevant for the physical behaviour of the ionic complexes involving HCO$^+$ and N$_2$H$^+$. For these initial explorations of the potential energy surfaces, the CCSD(T)/aug-cc-pVTZ approach is of satisfactory quality. A more detailed analysis of various basis-set effects, including the basis-set superposition error in systems with significantly deformed monomers, is being prepared and will be presented elsewhere. ]{} Conclusion {#sec:conclusion} ========== Ion-molecule reactions are common in interstellar space, and investigating them helps to quantitatively understand the molecular universe [@watson76a]. We studied the isotope fractionation reactions of HCO$^+$/HOC$^+$ with CO and N$_2$H$^+$ with N$_2$, as well as the linear proton-bound complexes formed in the course of these reactions. For OCH$^+$+CO, we pointed out inaccuracies of previous exothermicity values that are commonly employed in chemical networks. The new exothermicities affect particularly prominently the rate coefficients derived at temperatures of dark interstellar cloud environments, which markedly changes the abundance ratios of the $^{13}$C- and $^{18}$O-containing formyl isotopologues. The linear proton-bound cluster ions are found to be strongly bound (2000–7000 [cm$^{-1}$]{}). The ionic complexes OCH$^+\cdots$CO and OCH$^+\cdots$N$_2$ have sizeable dipole moments (2.9-3.5 D) and rotational constants of approximately 2000 MHz. If stabilized by means of collision and/or radiative processes, their high rotational population may facilitate the detection of these ions at low temperatures. MM is grateful to Geerd H. F. Diercksen for sending her a copy of the MPI/PAE Astro 135 report. Marius Lewerenz is acknowledged for helpful discussions. Mila Lewerenz is thanked for helping with the literature search. [43]{} natexlab\#1[\#1]{} Adams, N. G. & Smith, D. 1981, ApJ, 247, L123 Anicich, V. G. 1993, J. Phys. Chem. Ref. Data, 22, 1469 Bendtsen, J. 2001, J. Ram. Spectrosc., 32, 989 , L., [Caselli]{}, P., & [Dore]{}, L. 2010, A&A, 510, L5 , L., [Caselli]{}, P., [Leonardo]{}, E., & [Dore]{}, L. 2013, A&A, 555, A109 , P., [Dutoi]{}, T., [Mladenovi[ć]{}]{}, M., [et al.]{} 2001, Faraday Discussions, 118, 433 , P. & [Dore]{}, L. 2005, A&A, 433, 1145 , P., [Walmsley]{}, C. M., [Zucconi]{}, A., [et al.]{} 2002, ApJ, 565, 331 , A. & [Black]{}, J. H. 1976, Rep. Prog. Phys., 39, 573 Freeman, C. G., Knight, J. S., Love, J. G., & McEwan, M. J. 1987, Int. J. Mass. Spec. Ion. Proc., 80, 255 , M., [Langer]{}, W. D., & [Wilson]{}, R. W. 1982, A&A, 107, 107 Henning, P., Kraemer, W. P., & Diercksen, G. H. F. 1977, [Internal Report, MPI/PAE Astro 135]{}, Max-Planck Institut, M[ü]{}nchen , E. 2003, Space Sci. Rev., 106, 293 Herbst, E., Payzant, J. D., Schiff, H. I., & Bohme, D. K. 1975, ApJ, 201, 603 Huang, X., Valeev, E. F., & Lee, T. J. 2010, J. Chem. Phys., 133, 244108 Huber, K. P. & Herzberg, G. 1979, Molecular Spectra & Molecular Structure Vol. IV. Constants of Diatomic Molecules (Englewood Cliffs, NJ: Prentice-Hall) , E., [Asvany]{}, O., & [Schlemmer]{}, S. 2009, J. Chem. Phys., 130, 164302 , J. K., [Sch[ö]{}ier]{}, F. L., & [van Dishoeck]{}, E. F. 2004, A&A, 416, 603 Langer, W. D., Graedel, T. E., Frerking, M. A., & Armentrout, P. B. 1984, ApJ, 277, 581 , J., [Pineau Des Forets]{}, G., [Roueff]{}, E., & [Flower]{}, D. R. 1993, A&A, 267, 233 , H. S. 2007, , 476, 291 Lohr, L. L. 1998, J. Chem. Phys., 108, 8012 , S., [Bergin]{}, E. A., & [Tafalla]{}, M. 2013, A&A, 559, A53 Martin, J. M. L., Taylor, P. R., & Lee, T. J. 1993, J. Chem. Phys., 99, 286 , S. N., [Woolf]{}, N. J., & [Ziurys]{}, L. M. 2009, Astrophys. J., 690, 837 Mills, I., Cvitaš, T., Homann, K., Kallay, N., & Kuchitsu, K. 1993, Quantities, Units and Symbols in Physical Chemistry, Second Edition (Oxford: Blackwell Scientific Publications) Mladenović, M. & Bačić, Z. 1990, J. Chem. Phys., 93, 3039 Mladenović, M. & Schmatz, S. 1998, J. Chem. Phys., 109, 4456 Olney, T. N., Cann, N. M., Cooper, G., & Brion, C. E. 1997, Chem. Phys., 223, 59 , L., [Salez]{}, M., & [Wannier]{}, P. G. 1992, A&A, 258, 479 , S. D. & [Charnley]{}, S. B. 2004, MNRAS, 352, 600 , M. & [Ossenkopf]{}, V. 2013, A&A, 550, A56 , B. & [Berkowitz]{}, J. 1991, J. Chem. Phys., 95, 4378 Schmatz, S. & Mladenović, M. 1997, Ber. Bunsenges. Phys. Chemie, 101, 372 Smith, D. & Adams, N. G. 1980, ApJ, 242, 424 , L. E., [Hollis]{}, J. M., & [Watson]{}, W. D. 1977, ApJ, 212, 79 pirko, V., Bludsk[' y]{}, O., & Kraemer, W. P. 2008, Collect. Czech. Chem. Commun., 73, 873 Stanton, J. F., Gauss, J., Harding, M. E., Szalay, P. G., [et al.]{}, [CFOUR]{}, a quantum chemical program package, for the current version, see http://www.cfour.de , K. & [Nesbitt]{}, D. J. 2010, Phys. Chem. Chem. Phys., 12, 8311 Terzieva, R. & Herbst, E. 2000, MNRAS, 317, 563 Trickl, T., Proch, D., & Kompa, K. L. 1995, J. Mol. Spectrosc., 171, 374 Urey, H. C. 1947, J. Chem. Soc., 562 Wagner-Redeker, W., Kemper, P. R., Jarrold, M. F., & Bowers, M. T. 1985, J. Chem. Phys., 83, 1121 Watson, W. D. 1976, Rev. Mod. Phys., 48, 513 Werner, H.-J., Knowles, P. J., Knizia, G., [et al.]{} 2012, [MOLPRO]{}, a package of ab initio programs, for the current version, see http://www.molpro.net

ArXiv

--- abstract: 'We present a theoretical study of the magnetic band structure of conduction and valence states in Quantum Well Wires in high magnetic fields. We show that hole mixing results in a very complex behavior of valence edge states with respect to conduction states, a fact which is likely to be important in magneto-transport in the Quantum Hall regime. We show how the transition from one-dimensional subbands to edge states and to Landau levels can be followed by optical experiments by choosing the appropriate, linear or circular, polarization of the light, yielding information on the one-dimensional confinement.' address: - | Departement Natuurkunde, Universiteit Antwerpen (UIA), Universiteitsplein 1, B-2610 Antwerpen, Belgium\ and Istituto Nazionale Fisica della Materia, Dipartimento di Fisica, Università di Modena,\ Via Campi 213/A, I-41100 Modena, Italy - | Istituto Nazionale Fisica della Materia, Dipartimento di Fisica, Università di Modena,\ Via Campi 213/A, I-41100 Modena, Italy author: - 'G. Goldoni' - 'A. Fasolino' title: ' Valence band structure, edge states and interband absorption in Quantum Well Wires in high magnetic fields ' --- =10000 Quasi-one-dimensional (Q1D) semiconductor structures are promising systems for the investigation of novel optical and transport properties, as well as for potential technological applications. [@cingo] The nature of [*magnetic*]{} states in Q1D structures is of primary importance in the understanding of transport properties, e.g., in the Quantum Hall regime. Besides, characterization of the currently challenging growth of these structures and informations on the lateral confinement can be gained by use of high magnetic fields; in fact, in Q1D structures, the magnetic field provides an additional tunable confinement length which can range from much larger to much smaller than the lateral dimension of the wires and can therefore be used to characterize the lateral confinement length in these structures. Optical characterization of Q1D structures provides very detailed information on electronic states; band structure effects, such as the coupling between heavy-hole (HH) and light-hole (LH) valence states, have been shown to affect the interband optical absorption of Q1D and of laterally modulated structures at zero and low magnetic field, [@Bockelmann92] giving rise to anisotropic absorption for linearly polarized light. In this paper we present a theoretical study of the electronic and optical properties up to high magnetic fields of Quantum Well Wires (QWW) obtained by lateral etching of a Quantum Well (QW). [@marzin] The alternative method of growth of Q1D structures by deposition on non-planar substrates is also used to achieve smaller confinement lengths. [@kapon; @rinaldi] In QWWs, electronic states have either a quasi-2D (Q2D) or a Q1D character, depending on the size of the additional lateral confinement with respect to the underlying QW. As mentioned above, theoretical [@Bockelmann92; @citrin] and experimental [@rinaldi; @kohl; @tanaka; @notomi] investigations have shown that, at zero and low magnetic field, the in-plane anisotropy induced by 1D quantization can be probed by linearly polarized interband spectroscopy. Here we show that, in high magnetic fields, instead, the anisotropic absorption for circularly polarized light of different helicity is more appropriate to study the features resulting from the field-induced hole-mixing; A comparison of linearly and circularly polarized light absorption as a function of the field directly probes the lateral potential by monitoring the transition from the low field regime, dominated by 1D confinement, to the high field regime, dominated by Landau quantization. In Sec. \[sec:method\] we outline the method of calculation of conduction and valence states in Q1D structures from zero to high magnetic fields. In Sec. \[sec:edgestates\] we focus on the different nature of valence and conduction edge states in QWWs. In Sec. \[sec:optics\] we present the calculated optical absorption spectra in linearly and circularly polarized light. Finally, summary and conclusions are given in Sec. \[sec:conclusions\]. method {#sec:method} ====== We consider a QWW such as shown in Fig. 1, which we model by a (001)-grown QW, of length $d_z$, and a lateral infinite square well along (100), of length $d_x$. In the following we call $V^e(z), V^h(z) $ the QW confining potential for electrons and holes, respectively, and $V( x)$ the infinite well potential acting laterally on both electrons and holes. We will also consider this system in presence of a magnetic field B directed along the (001) direction. Previous calculations of hole states in $B\ne0$ have considered weak sinusoidal potentials, which allow a simpler analytical approach. [@Bockelmann92] For both the B=0 and B$\neq$0 cases we will also calculate the absorption strength for dipolar interband transitions for linearly and circularly polarized light perpendicular to the layers, as illustrated in Fig. 1. As we will show below, the assumption of an infinite well in the $x$-direction allows to reduce the 2D problem to two 1D problems. In other words, it allows to factorize the electron envelope function $\Psi^e({\bf r})$ into the product of the envelope functions of the QW bound states $\psi^e(z)$ times the envelope functions $\phi(x)$ of the lateral potential $V(x)$, also in presence of a magnetic field. Furthermore, the same factorization applied to the uncoupled HH and LH states serves as a suitable representation to include the hole mixing. Conversely, the choice of a QW profile in the $z$ direction can be easily extended to more complex potential profiles, such as coupled QWs. We obtain the electron and hole energy levels and envelope functions by solving separately the effective-mass electron and hole Hamiltonians. Case $B=0$ ---------- In zero field the effective-mass electron Hamiltonian reads $$H^e = -\frac{\hbar^2}{2} \frac{\partial}{\partial z}\frac{1}{m^*(z)}\frac{\partial}{\partial z} +\frac{\hbar^2}{2m^*(z)} \left[ -\frac{\partial^2}{\partial x^2} + k_y^2 \right] + V(x) + V^e(z) \,, \label{He}$$ where $m^*(z)$ is the bulk conduction electron effective mass which depend on the material and, therefore, on the $z$ coordinate; the usual current-conserving kinetic operator [@BenDaniel] has been used accordingly. We neglect non-parabolicity effects. The hole Hamiltonian is $${\bf H}^h= \bordermatrix{&+3/2&-1/2&+1/2&-3/2 \cr & H_h & c_+ & b_+ & 0 \cr & c_- & H_l & 0 & -b_+ \cr & b_- & 0 & H_l & c_+ \cr & 0 & -b_- & c_- & H_h \cr } + \left[V(x) + V^h(z)\right] {\bf I}_4 \,, \label{Hh}$$ with $$\begin{aligned} H_h & = & \frac{\hbar^2}{2m_0} \left[\left(\gamma_1+\gamma_2\right)\left(\frac{\partial^2} {\partial x^2}-k_y^2\right) + \frac{\partial}{\partial z} \left(\gamma_1-2\gamma_2\right) \frac{\partial}{\partial z} \right] \,, \\ H_l & = & \frac{\hbar^2}{2m_0} \left[\left(\gamma_1-\gamma_2\right)\left(\frac{\partial^2}{\partial x^2}-k_y^2\right) + \frac{\partial}{\partial z} \left(\gamma_1+2\gamma_2\right) \frac{\partial}{\partial z} \right] \,, \\ c_\pm & = & -\sqrt{3}\frac{\hbar^2}{2m_0} \left[ \gamma_2 \left(\frac{\partial^2}{\partial x^2}+k_y^2\right) \pm 2 \gamma_3 k_y \frac{\partial}{\partial x}\right] \,,\\ b_\pm & = & -\sqrt{3} \frac{\hbar^2}{2m_0} \left( \frac{\partial}{\partial x}\pm k_y\right) \left(\gamma_3\frac{\partial}{\partial z}+ \frac{\partial}{\partial z}\gamma_3\right) \,,\end{aligned}$$ where $m_0$ is the free-electron mass, $\gamma_1,\gamma_2$ and $\gamma_3$ are the Luttinger parameters and ${\bf I}_4$ is the identity $4\times 4$ matrix. Equation (\[Hh\]) is obtained from the Luttinger Hamiltonian [@Luttinger56] written in the basis of the eigenstates of the $J=3/2$ total angular momentum, $$|\frac{3}{2},+\frac{3}{2}\rangle\,, |\frac{3}{2},-\frac{1}{2}\rangle\,, |\frac{3}{2},+\frac{1}{2}\rangle\,, |\frac{3}{2},-\frac{3}{2}\rangle\,, \label{Jbasis}$$ after the substitutions $k_x\rightarrow -i\partial/\partial x, k_z\rightarrow -i\partial/\partial z $, and symmetrization of the non-commuting products [@bastard-wm] between $\partial/\partial z$ and the $z$-dependent $\gamma_i$’s. The eigenstates of $H^e$ can be factorized as $\Psi^e_{n,m} ({\bf r})=e^{ik_yy} \phi_m(x) \psi_n^e(z)$, where $n$ labels the bound states of the QW potential $V^e(z)$, and $m$ the analytical solutions of the infinite well $V(x)$. Consequently, the eigenvalues of (\[He\]) depend quadratically on the in-wire momentum $k_y$. The solution of the valence states is more complex because the off-diagonal terms in ${\bf H}^h$ couple the solutions of the diagonal, electron-like terms $H_h$ and $H_l$. Following Bockelmann and Bastard, [@Bockelmann92] we first solve the electron-like HH and LH Hamiltonians, $H_h+V^h(z)+V(x)$ and $H_l+V^h(z)+V(x)$, giving the factorized solutions $\Psi^{HH}_{n_h,m}=e^{ik_yy}\psi_{n_h}^{HH}(z)\phi_m(x) $ and $\Psi^{LH}_{n_l,m}=e^{ik_yy}\psi_{n_l}^{LH}(z)\phi_m(x) $, where $n_h, n_l$ label the HH and LH states. Then we expand the solution of the hole Hamiltonian ${\bf H}^h$ onto the following basis set $$\begin{aligned} |HH^+;n_h,m\rangle & = & (\Psi^{HH}_{n_h,m},0,0,0), \\ |LH^-;n_l,m\rangle & = & (0,\Psi^{LH}_{n_l,m},0,0), \\ |LH^+;n_l,m\rangle & = & (0,0,\Psi^{LH}_{n_l,m},0), \\ |HH^-;n_h,m\rangle & = & (0,0,0,\Psi^{HH}_{n_h,m}). \end{aligned}$$ In the expansion, we let $n_h,n_l$ run over all $N_h$ bound HH and $N_l$ bound LH states of $V^h(z)$, while $m$ runs over $M$ eigenstates of $V(x)$ (typically we take $M=15$). Finally we diagonalize the resulting $2 \times (N_h+N_l) \times M $ Hamiltonian. Due to the off-diagonal couplings, therefore, the valence band states cannot be assigned to a single set of indices $(n,m)$, in contrast to conduction electrons; this mixing results in a strongly non-parabolic in-wire energy dispersion. When we diagonalize the hole Hamiltonian ${\bf H}^h$, we do not make the axial approximation, as assumed in Ref. , which is more severe in QWWs than in QWs as shown later in Section III. Case $B\neq0$ ------------- When we include a uniform magnetic field $\mbox{\boldmath $B$}=(0,0,B)$ along the QW growth direction, described by the vector potential $\mbox{\boldmath $A$}=(0,Bx,0)$, the electron Hamiltonian $H^e$, neglecting the electron spin splitting, reads $$-\frac{\hbar^2}{2} \frac{\partial}{\partial z} \frac{1}{m^*(z)} \frac{\partial}{\partial z}+ \frac{1}{m^*(z)/m_0}\left[-\frac{\hbar^2}{2m_0} \frac{\partial^2}{\partial x^2}+ \frac{m_0\omega^2}{2}\left(x-x_0\right)^2\right] + V^e(z) + V(x)$$ and the hole Hamiltonian ${\bf H}^h$ becomes $${\bf H}^h -2\kappa\mu_BB{\bf J}_z\,,$$ where ${\bf H}^h$ is given by (\[Hh\]) with $$\begin{aligned} H_h & = & \left(\gamma_1+\gamma_2\right) \left[\frac{\hbar^2}{2m_0} \frac{\partial^2}{\partial x^2}- \frac{m_0\omega^2}{2}\left(x-x_0\right)^2\right] + \frac{\hbar^2}{2m_0} \frac{\partial}{\partial z} \left(\gamma_1-2\gamma_2\right) \frac{\partial}{\partial z} + V^h(z)+V(x)\,, \\ H_l & = & \left(\gamma_1-\gamma_2\right) \left[\frac{\hbar^2}{2m_0} \frac{\partial^2}{\partial x^2}- \frac{m_0\omega^2}{2}\left(x-x_0\right)^2\right] + \frac{\hbar^2}{2m_0} \frac{\partial}{\partial z} \left(\gamma_1+2\gamma_2\right) \frac{\partial}{\partial z} + V^h(z)+V(x)\,, \\ c_\pm & = & -\sqrt{3}\left[\gamma_2\left(\frac{\hbar^2}{2m_0}\frac{\partial^2}{\partial x^2}+\frac{m_0\omega^2} {2}\left(x-x_0\right)^2\right)\pm\gamma_3\frac{\hbar\omega}{2} \left(1+2\left(x-x_0\right)\frac{\partial}{\partial x}\right)\right] \,,\\ b_\pm & = & -\sqrt{3} \left[\frac{\hbar^2}{2m_0} \frac{\partial}{\partial x}\pm\frac{\hbar\omega}{2}\left(x-x_0\right)\right] \left(\gamma_3\frac{\partial}{\partial z}+ \frac{\partial}{\partial z}\gamma_3\right) \,.\end{aligned}$$ ${\bf J}_z$ is the diagonal matrix representing the $z$-component of the angular momentum operator in the basis (\[Jbasis\]); $\kappa $ is an additional Luttinger parameter, $\omega=e B/m_0$ is the cyclotron frequency and $x_0=-l_m^2k_y$ is the semiclassical ‘orbit center’, where $l_m=(\hbar/eB)^{1/2}$ is the magnetic length. The hole eigenstate dependence on the in-wire wavevector $k_y$ of the $B=0$ case, is now replaced by the dependence on $x_0$. Both for $H^e$ and for the diagonal terms of ${\bf H}^h$ the lateral potential in the $x$ direction is, at $B\neq0$, the sum of the infinite well plus a parabolic effective potential. The latter term is equal for the two Hamiltonians so that, at each $B$ and each $x_0$, the electron and hole $\phi_m(x)$’s are the same, as in the $B=0$ case. In high magnetic fields, the $\phi_m(x)$’s reduce to harmonic oscillator eigenstates when $x_0$ is far from the barrier, i.e., when the confinement along $x$ due the parabolic magnetic potential is not affected by the infinite barriers; instead, when $x_0$ is near the barriers the $\phi_m(x)$’s are the so-called edge states. [@MacDonald84] In view of the possible application to other confinement potentials, we have found convenient to compute both the $\psi_n(z)$’s and the $\phi_m(x)$’s numerically by direct integration of the Schrödinger equation in real space. Energy levels and edge states of QWWs {#sec:edgestates} ===================================== In Fig. 2(a) and Fig. 2(b) we show the calculated hole subbands for QWWs formed by a GaAs/Al$_{0.35}$Ga$_{0.65}$As QW of width $d_z=10\,\mbox{nm}$ and by lateral infinite wells of width $d_x=30\,\mbox{nm}$ and $d_x=100\,\mbox{nm}$ respectively. Each HH and LH bound state of the QW is split by the lateral potential into a series of 1D subbands of mixed character. In Table I we give the orbital composition of selected states as compared to the purely HH or LH states of the QW at zero in-plane wavevector. As expected, the additional confinement energy and hole mixing is much less for $d_x=100\,\mbox{nm}$ than for $d_x=30\,\mbox{nm}$. Furthermore, in Fig. 2(a) and Table I we compare the calculated 1D hole subbands with those obtained within the axial approximation for the same structure. The anisotropy-induced couplings have a rather strong effect on the curvature of the subbands; in particular, the electron-like curvature of the light hole derived 1D subbands is strongly reduced in the axial approximation. In fact, as sketched in the inset of Fig 2(b), due to the quantization of the lateral wavevector $k_x$, each in-wire momentum $k_y$ corresponds to an effective in-plane wavevector ${\bf k}_\parallel=(k^{\mbox{\small eff}}_x,k_y)$, away from the (010) direction, of the underlying, anisotropic QW band structure. [@Altarelli85] Here $k^{\mbox{\small eff}}_x=m\pi/d_x $, where $m$ is the QWW subband index. Therefore, as it is apparent in Fig. 2(a), in QWWs not only the in-wire dispersion but also the hole confinement energies are affected by the axial approximation. In Fig. 3 we show the electron and hole magnetic levels at $B=10\,\mbox{T}$ for the same QWW of Fig. 2(b). We plot the magnetic levels as a function of the semiclassical ‘orbit center’ $x_0$. The most striking effect is that, while the electrons (Fig. 3(a)) display a ladder of edge states with well defined increasing oscillator number, [@buttiker] the valence edge states (Fig. 3(b)) have complicated shapes which are likely to be of relevance for magneto-transport, particularly in the integer Quantum Hall regime. As for the 1D subbands, also here the axial approximation is very severe. The comparison of Figs. 3(a) and 3(b) clearly shows that, for holes, the cross over from bulk Landau levels to edge states close to the barrier is less straightforward than for electrons. In fact, each hole state is a mixture of oscillator states of different quantum number $m$. As a consequence, each component starts feeling the effect of the barrier at a different value of the magnetic field. As shown in Table II, the second hole level is the only one almost purely $m=0$ and, indeed, it starts deviating from a flat dispersion at the same $x_0$ as the $m=0$ electron state, while all others have a mixed composition resulting in more complicated dispersions. This point is further clarified by aid of Fig. 4 and Table II. In Fig. 4 we compare the calculated hole magnetic levels for the two QWWs of Fig. 2 ($d_z=10\,\mbox{nm}$ and either $d_x=30\,\mbox{nm}$ or $d_x=100\,\mbox{nm}$) in a magnetic field of $10\,\mbox{T}$ and $30\,\mbox{T}$. We recall that at $B=10\,\mbox{T}$,  $l_m \simeq 8\,\mbox{nm}$, i.e., comparable to half the thinner $d_x$, while at $B=30\,\mbox{T}$, $l_m \simeq 4.7\,\mbox{nm}$, i.e., smaller than half $d_x$ for both structures. Therefore, at $B=10\,\mbox{T}$, for $d_x=30\,\mbox{nm}$ (panel 4(a)) all Landau levels are perturbed by the lateral barriers also at the center of the well, while for $d_x=100\,\mbox{nm}$ (panel 4(c)) the first edge states evolve to flat Landau levels away from the barrier. Conversely, at $B=30\,\mbox{T}$ (panels 4(b), 4(d)) we find that the energy structure of all edge states down to -40 meV is the same for the two values of $d_x$ and is dominated by Landau quantization, due to the field, rather than by the one-dimensional confinement. Table II gives a rationale for this behavior. It can be seen that very few states are basically composed of only one harmonic oscillator number and that the mixing increases with increasing B. Furthermore, at high fields, the eigenvalues at the center of the well of the low oscillator index levels do not depend any more on the lateral potential. This transition takes place at different values of the magnetic field for each level as illustrated in Fig. 5. Here we show the evolution of electron and hole states as a function of the field for two values of $d_x$. The striking difference between electron and hole levels is related to the mixed composition of the hole states, given in Table II; The non-trivial behavior of hole levels in Fig. 5 suggests that the transition from the low field regime, dominated by 1D quantization, to the high field regime, dominated by Landau quantization, is associated with large changes in the wavefunction composition, with consequent influence on the matrix elements for optical transitions, which we calculate in the next section. Our aim is to show that the transition from the low to the high field regime can be monitored by optical experiments with linearly and circularly polarized light, giving unambiguous information on the lateral confinement. Optical absorption in QWW {#sec:optics} ========================= At B=0, the two in-plane linear polarizations of the light induce different inter-band absorption probability, as a consequence of the one-dimensionality of the hole states, [@Bockelmann92] while the two circular polarizations give the same absorption intensity. At the opposite limit of very large fields, the situation is reversed: we expect the two in-plane linear polarizations to give degenerate spectra, due to the recovered two-dimensional character of hole states, while the two circular polarizations of the light, being coupled to different components of the spin-orbit coupled states, give different spectra. Therefore, in the intermediate regime, we expect large changes in the absorption spectra for both circularly and linearly polarized light, as shown next. Strictly speaking, a calculation of magneto-optical properties in Q1D structures should include excitonic effects. The exact inclusion of the Coulomb interaction is, however, a complicated task in itself and theoretical calculations exist only for model, purely 1D structures in zero field. [@ogawa] However, several studies [@ancil; @goldoni] have shown that, in Q2D structures, a perturbative excitonic correction [@hasegawa] added to the one-particle absorption spectra yield a good description of magneto-optical experiments in the high field regime, provided the hole mixing is included. Therefore, we expect our results, obtained in a one-particle approximation, to be qualitatively correct, particularly in the high field regime close to the transition to a 2D behaviour. In Fig. 6, we compare the calculated optical absorption spectra for linear and circular light polarization at $B=10\,\mbox{T}$ and $B=30\,\mbox{T}$ and for $d_x=30\,\mbox{nm}$; we have labelled the main features according to the orbital character of the initial (hole) and final (electron) states involved in the transitions. The mixing of hole states makes that very many hole levels have not vanishing oscillator strength for the transition to the same electron state, making the spectra rich and informative. For instance, at $10\,\mbox{T}$ the transition $\mbox{LH}^-_{10}\rightarrow e^\downarrow_{10}$, which is induced by $\sigma^+$ polarized light, gives rise to a single peak in the spectra, while the transition $\mbox{LH}^+_{10}\rightarrow e^\uparrow_{10}$ is split into three peaks in $\sigma^-$ polarization. In fact, as shown in Table II, there is only one level (level 6) with strong LH$^-$ character and a strong $m=0$ component, while there are three levels with LH$^+$ and $m=0$ character (levels 5, 10 and a deeper one which is not reported in the table). Figure 6 shows that, while the spectra for light linearly polarized along the two in-plane direction of the QWW become very similar at high fields, as we expect from the isotropic character of the orbital motion, large differences are present in the spectra calculated for the two circular polarization of the light, as observed in Q2D structures. [@ancil] Hence, these differences can be regarded as a fingerprint of the transition from 1D dynamics to Landau quantization; by studying the evolution of the linear and circular absorption, the transition from the low field to the high field regime can be followed, yielding quantitative information on the effective length of the 1D confinement. Figure 7 shows how the anisotropy between spectra obtained from linearly polarized light decreases as a function of the field. Calculations are performed for the same sample of Fig. 6. The anisotropy is expressed as $100\times (I_x-I_y)/(I_x+I_y)$, where $I_x$ and $I_y$ are the heights of the lowest energy peak of the absorption spectra calculated for the $x$ and $y$ polarizations, respectively. As expected from the above discussion, the anisotropy is strongly quenched by the field and decreases monotonously from $10\,\%$ at $B=0$ to $3\,\%$ at $B=20\,\mbox{T}$ for this sample. At fields larger than $20\,\mbox{T}$, this peak splits into a double peak, as shown in Fig. 6 at $30\,\mbox{T}$, and the above definition does not apply. Finally, in Fig. 8 we show the evolution of the calculated spectra for the two circular polarizations of the light as a function of magnetic field for the $d_z=10\,\mbox{nm}$, $d_x=30\,\mbox{nm}$ sample. Hole mixing is responsible of complex features in the low-energy range, which could hopefully be experimentally detected. In Fig. 8 the calculated spectra for the $\sigma^-$ polarization show a clear anticrossing behavior around $1590\,\mbox{meV}$. This feature, which is due to HH-LH mixing and would be absent in a simple, uncoupled-hole model, can be seen, again, as a fingerprint of the transition between 1D subbands to Landau levels of hole states. Note that this anticrossing is totally absent in the corresponding peaks for the $\sigma^+$ polarization as a consequence of the smaller mixing of HH$^-$ and LH$^-$ states as compared to HH$^+$ and LH$^+$ states (see Table II). conclusions {#sec:conclusions} =========== We have presented a study of the electronic structure of rectangular QWWs following the transition from the 1D quantization at zero magnetic fields of the subband structure to Landau quantization at high magnetic fields. We have focused on the behavior of the spin-orbit coupled valence band, and, to this purpose, we have developed a method of solution of the Luttinger Hamiltonian for Q1D structures both at B=0 and B$\neq$0. We have shown that the hole mixing gives rise to a rather complex edge state structure compared to electron states. Furthermore, we have shown that a study of the optical interband transition for linearly and circularly polarized light could be used to get informations on the lateral confinement. As we discussed in Sec. III, excitonic effects, which are neglected in the present calculations, should not change qualitatively our results, particularly in the regime of high fields. Furthermore, the present one-particle calculation, which takes into account the coupled nature of the valence subbands, is a necessary ingredient for a successive calculation of the magneto-exciton. This study is also preliminary to a study of more complex Q1D structures [@kapon; @rinaldi] showing very promising optical properties. In these structures, the two confinement directions are of comparable width and have potential barriers of the same high and, therefore, should be treated on the same foot. Theoretical investigations of valence states in these structures are currently in progress. We hope that the present work will stimulate further experimental and theoretical work both on the transport and optical properties of these interesting systems. We thank F. Rossi for fruitful discussions. We acknowledge the project ESPRIT 7260 (SOLDES) for partial financial support. Part of this work has been carried out within the European Human Capital & Mobility Network magNET, n. ERBCHRXCT 920062 for for a recent review see, e.g., R. Cingolani and R. Rinaldi, Rivista del Nuovo Cimento [**16**]{}, 1 (1993). U. Bockelmann and G. Bastard, Phys. Rev. [**B 45**]{}, 1688 (1992); U. Bockelmann and G. Bastard, Phys. Rev. [**B 45**]{}, 1700 (1992). See, e.g., J. Y. Marzin, A. Izrael, and L. Birotheau, Solid State Electr. [**37**]{}, 1091 (1994) and Ref. . M. Walther, E. Kapon, C. Caneau, D. M. Hwang, and L. M. Schiavone, Appl. Phys. Lett. [**62**]{}, 2170 (1993). R. Rinaldi, R. Cingolani, M. Lepore, M. Ferrara, I. M. Catalano, F. Rossi, L. Rota, E. Molinari, P. Lugli, U. Marti, D. Martin, F. Morier-Gemoud, P. Ruterana, and F. K. Reinhart, Phys. Rev. Lett. [ **73**]{}, 2899 (1994); R. Rinaldi, M. Ferrara, R. Cingolani, U. Marti, D. Martin, F. Morier-Gemoud, P. Ruterana, and F. K. Reinhart, Phys. Rev. B [**50**]{}, 11795 (1994). D.S. Citrin and Y.-C. Chang, Phys. Rev. [**B 43**]{}, 11703 (1992). M. Kohl, D. Heitman, P. Granbow, and K. Ploog, Phys. Rev. Lett. [**63**]{}, 2124 (1989). M. Tanaka, J. Motohisa, and H. Sakaki, Surf. Sci. [**228**]{}, 408 (1990). M. Notomi, M. Okamoto, H. Iwamura, and T. Tamamura, Appl. Phys. Lett. [**62**]{}, 1094 (1993). D. J. Ben-Daniel and C. B. Duke, Phys. Rev. [**152**]{}, 683 (1966). J. M. Luttinger, Phys. Rev. [**102**]{}, 1030 (1956). G. Bastard, [*Wave mechanics applied to semiconductor heterostructures*]{} (Les éditions des physique, Les Ulis, Paris, France 1988). A. H. MacDonald and P. Streda, Phys. Rev. B [**29**]{}, 1616 (1984). M. Altarelli, U. Ekenberg, and A. Fasolino, Phys. Rev. B [**32**]{}, 5138 (1985). M. Buttiker, Phys. Rev. [**B 38**]{}, 9375 (1988). T. Ogawa and T. Takagehara, Phys. Rev. B [**43**]{}, 14325 (1991); T. Ogawa and T. Takagehara, Phys. Rev. B [**44**]{}, 8138 (1992). F. Ancilotto, A. Fasolino, and J. C. Maan Phys. Rev. [**B 38**]{}, 1788 (1988). G. Goldoni, T. Ruf, V. F. Sapega, A. Fainstein, and M. Cardona, Phys. Rev. B, in press. O. Akimoto and H. Hasegawa, J. Phys. Soc. Japan [**22**]{}, 181 (1967). -------- -------------- -------- ------- -------- -------- ------- -------- -------- ------- -------- QW $d_x=\infty$ E E HH/LH $m=0 $ E HH/LH $m=0 $ E HH/LH $m=0 $ HH$_1$ -7.39 -7.62 100 100 -9.94 98 98 -10.14 98 99 LH$_1$ -22.55 -22.52 85 59 -23.06 84 81 -23.41 89 88 HH$_2$ -29.39 -30.19 85 85 -35.89 67 62 -35.64 63 57 -------- -------------- -------- ------- -------- -------- ------- -------- -------- ------- -------- : Comparison of orbital character of the QW bound states and of the QWW levels at $k_y=0$ having mainly $m=0$ character at $B=0$ for the two differently confined QWWs of Fig. 2. For each QWW, we give the energy (in meV) of the lowest level deriving from each pure QW state, the percentage of HH or LH character, and the percentage of $m=0$ character. The pure QW HH$_1$, LH$_1$ and HH$_2$ levels are further confined by the QWW potential $V(z)$ and acquire a mixed HH-LH character. [clllllllll]{}\ band & E(meV) & $HH_1^+$ & m & $HH_1^-$ & m & $LH_1^-$ & m & $LH_1^+$ & m\ 1 & -10.24 & 0.89 & 0 & - & - & 0.07 & 2 & - & -\ 2 & -10.69 & - & - & 0.98 & 0 & - & - & 0.01 & 2\ 3 & -16.29 & 0.72 & 1 & - & - & 0.20 & 3 & - & -\ 4 & -17.50 & - & - & 0.70 & 1 & - & - & 0.20 & 1,3\ 5 & -19.35 & - & - & 0.36 & 2,0 & - & - & 0.54 & 0,2\ 6 & -19.47 & 0.03 & 2 & - & - & 0.82 & 0 & - & -\ 7 & -21.09 & - & - & 0.28 & 3 & - & - & 0.52 & 1\ 8 & -23.12 & 0.53 & 2 & - & - & 0.36 & 2,4 & - & -\ 9 & -24.01 & 0.09 & 3 & - & - & 0.76 & 1 & - & -\ 10 & -25.91 & - & - & 0.09 & 4 & - & - & 0.78 & 0,2,4\ \ band & E(meV) & $HH_1^+$ & m & $HH_1^-$ & m & $LH_1^-$ & m & $LH_1^+$ & m\ 1 & -11.52 & 0.77 & 0 & - & - & 0.17 & 2 & - & -\ 2 & -15.96 & - & - & 0.95 & 0 & - & - & 0.05 & 2\ 3 & -19.89 & 0.02 & 2 & - & - & 0.85 & 0 & - & -\ 4 & -20.27 & 0.61 & 1 & - & - & 0.29 & 3 & - & -\ 5 & -20.27 & - & - & 0.15 & 3 & - & - & 0.64 & 1\ 6 & -20.78 & - & - & 0.24 & 2 & - & - & 0.70 & 0\ 7 & -27.38 & - & - & 0.23 & 4 & - & - & 0.64 & 2\ 8 & -28.42 & 0.51 & 2 & - & - & 0.36 & 4 & - & -\ 9 & -29.47 & - & - & 0.58 & 1 & - & - & 0.25 & 3\ 10 & -30.50 & 0.07 & 3 & - & - & 0.78 & 1 & - & -\ \ band & E(meV) & $HH_1^+$ & m & $HH_1^-$ & m & $LH_1^-$ & m & $LH_1^+$ & m\ 1 & -11.56 & 0.76 & 0 & - & - & 0.17 & 2 & - & -\ 2 & -15.93 & - & - & 0.95 & 0 & - & - & 0.05 & 2\ 3 & -19.87 & 0.03 & 2 & - & - & 0.84 & 0 & - & -\ 4 & -20.37 & - & - & 0.15 & 3 & - & - & 0.64 & 1\ 5 & -20.41 & 0.60 & 1 & - & - & 0.29 & 3 & - & -\ 6 & -20.79 & - & - & 0.24 & 2 & - & - & 0.70 & 0\ 7 & -27.42 & - & - & 0.26 & 4 & - & - & 0.61 & 2\ 8 & -28.54 & 0.49 & 2 & - & - & 0.38 & 4 & - & -\ 9 & -29.37 & - & - & 0.54 & 1 & - & - & 0.29 & 3\ 10 & -30.48 & 0.09 & 3 & - & - & 0.76 & 1 & - & -\

ArXiv

--- abstract: 'In this paper, we study the efficiency of a [**R**]{}estarted [**S**]{}ub[**G**]{}radient (RSG) method that periodically restarts the standard subgradient method (SG). We show that, when applied to a broad class of convex optimization problems, RSG method can find an $\epsilon$-optimal solution with a low complexity than SG method. In particular, we first show that RSG can reduce the dependence of SG’s iteration complexity on the distance between the initial solution and the optimal set to that between the $\epsilon$-level set and the optimal set. In addition, we show the advantages of RSG over SG in solving three different families of convex optimization problems. (a) For the problems whose epigraph is a polyhedron, RSG is shown to converge linearly. (b) For the problems with local quadratic growth property, RSG has an $O(\frac{1}{\epsilon}\log(\frac{1}{\epsilon}))$ iteration complexity. (c) For the problems that admit a local Kurdyka-Łojasiewicz property with a power constant of $\beta\in[0,1)$, RSG has an $O(\frac{1}{\epsilon^{2\beta}}\log(\frac{1}{\epsilon}))$ iteration complexity. On the contrary, with only the standard analysis, the iteration complexity of SG is known to be $O(\frac{1}{\epsilon^2})$ for these three classes of problems. The novelty of our analysis lies at exploiting the lower bound of the first-order optimality residual at the $\epsilon$-level set. It is this novelty that allows us to explore the local properties of functions (e.g., local quadratic growth property, local Kurdyka-Łojasiewicz property, more generally local error bounds) to develop the improved convergence of RSG. We demonstrate the effectiveness of the proposed algorithms on several machine learning tasks including regression and classification.' bibliography: - 'all.bib' - 'lewa.bib' - 'icml11.bib' title: 'RSG: Beating Subgradient Method without Smoothness and Strong Convexity' --- subgradient method, improved convergence, local error bound, machine learning Introduction ============ We consider the following generic optimization problem $$\label{eqn:prob0} f_*:=\min_{\w\in\Omega} f(\w)$$ where $f:\mathbb{R}^d\rightarrow(-\infty,+\infty]$ is an extended-valued, lower semicontinuous and convex function, and $\Omega\subseteq\R^d$ is a closed convex set such that $\Omega\subseteq\text{dom}(f)$. Here, we do not assume the smoothness of $f$ on $\text{dom}(f)$. During the past several decades, many fast convergent (especially linearly convergent) optimization algorithms have been developed for when $f$ is smooth and/or strongly convex. On the contrary, there are relatively fewer techniques for solving generic non-smooth and non-strongly convex optimization problems which still have many applications in machine learning, statistics, computer vision, and etc. To solve with $f$ being potentially non-smooth and non-strongly convex, one of the simplest algorithms to use is the subgradient (SG) method. When $f$ is Lipschitz-continuous, it is known that SG method requires $O(1/\epsilon^2)$ iterations for obtaining an $\epsilon$-optimal solution [@rockafellar1970convex; @opac-b1104789]. It has been shown that this iteration complexity is unimprovable for general non-smooth and non-strongly convex problems in a black-box first-order oracle model of computation [@opac-b1091338]. However, better iteration complexity can be achieved by other first-order algorithms for certain class of $f$ where additional structural information is available [@Nesterov:2005:SMN; @DBLP:journals/mp/GilpinPS12; @2015arXivRobert; @2014arXivJames; @2015arXivJames; @2016arXivJames]. In this paper, we present a generic restarted subgradient (RSG) method for solving which runs in multiple stages with each stage warm-started by the solution from the previous stage. Within each stage, the standard projected subgradient descent is performed for a fixed number of iterations with a constant step size. This step size is reduced geometrically from stage to stage. With these schemes, we show that RSG can achieve a lower iteration complexity than the classical SG method when $f$ belongs to some classes of functions. In particular, we summarize the main results and properties of RSG below: - For the general problem , under mild assumptions (see Assumption \[ass:rsg\]), RSG has an iteration complexity of $O(\frac{1}{\epsilon^2}\log(\frac{1}{\epsilon}))$ which has an addition $\log(\frac{1}{\epsilon})$ term but has significantly smaller constant in $O(\cdot)$ compared to SG. In particular, SG’s complexity depends on the distance from the initial solution to the optimal set while RSG’s complexity only depends on the distance from the $\epsilon$-level set (defined in ) to the optimal set, which is much smaller than the former distance. - When $f$ is locally quadratically growing (see Definition \[def:ssc\]), which is a weaker condition than strong convexity, RSG can achieve an $O(\frac{1}{\epsilon}\log(\frac{1}{\epsilon}))$ iteration complexity. - When $f$ admits a local Kurdyka-Łojasiewicz property (see Definition \[def:KL\]) with a power desingularizing function of degree $1-\beta$ where $\beta\in[0,1)$, RSG can achieve an $O(\frac{1}{\epsilon^{2\beta}}\log(\frac{1}{\epsilon}))$ complexity. - When the epigraph of $f$ over $\Omega$ is a polyhedron, RSG can achieve linear convergence, i.e., an $O(\log(\frac{1}{\epsilon}))$ iteration complexity. These results, except for the first one, are derived from a generic complexity of RSG for the problem satisfying a *local error condition* , which has a close connection to the existing error bound conditions and growth conditions [@Pang:1997; @Pang:1987:PEB:35577.35584; @Luo:1993; @DBLP:journals/corr/nesterov16linearnon; @Bolte:2006:LIN:1328019.1328299] in the literature. In spite of its simplicity, the analysis of RSG provides additional insight on improving first-order methods’ iteration complexity via restarting. It is known that restarting can improve the theoretical complexity of (stochastic) SG method for non-smooth problem when strongly convexity is assumed [@DBLP:journals/siamjo/GhadimiL13; @NIPS2012_4543; @hazan-20110-beyond] but we show that restarting can be still helpful for SG methods under other (weaker) assumptions. Although we focus on SG methods, the similar idea can be incorporated into various existing algorithms, leading to different variants of restarted first-order methods. [In particular, built on the groundwork (in particular Lemma \[lem:lowf\] and Lemma \[lem:generalp\]) laid in this paper [^1], several pieces of studies have improved the convergence of non-smooth optimization with a special structure [@DBLP:journals/corr/abs-1607-03815], stochastic subgradient methods [@DBLP:journals/corr/abs-1607-01027] and projection-reduced first-order methods [@DBLP:journals/corr/abs-non]. ]{} [We organize the reminder of the paper as follows. Section \[sec:rw\] reviews some related work. Section \[sec:main\] presents some preliminaries and notations. Section \[sec:RSG\] presents the algorithm of RSG and the general theory of convergence in the Euclidean norm. Section \[sec:spec\] considers several classes of non-smooth and non-strongly convex problems and shows the improved iteration complexities of RSG. Section \[sec:variant\] generalizes the algorithm and theory into the $p$-norm space using dual averaging method in each stage. Section \[sec:prac\] presents a parameter-free variant of RSG. Section \[sec:exp\] presents some experimental results. Finally, we conclude in Section \[sec:conc\]. ]{} Related Work {#sec:rw} ============ Smoothness and strong convexity are two key properties of a convex optimization problem that determine the iteration complexity of finding an $\epsilon$-optimal solution by first-order methods. In general, a lower iteration complexity is expected when the problem is either smooth or strongly convex. We refer the reader to [@opac-b1104789; @opac-b1091338] for the optimal iteration complexity of first-order methods when applied to the problems with different properties of smoothness and convexity. Recently there emerges a surge of interests in further accelerating first-order methods for non-strongly convex or non-smooth problems that satisfy some particular conditions [@DBLP:conf/nips/BachM13; @DBLP:journals/jmlr/WangL14; @DBLP:journals/corr/So13; @DBLP:conf/nips/HouZSL13; @DBLP:conf/icml/ZhouZS15; @DBLP:journals/corr/GongY14; @DBLP:journals/mp/GilpinPS12; @2015arXivRobert]. The key condition for us to develop the improved complexity is a local error bound condition  which is closely related to the error bound conditions in [@Pang:1987:PEB:35577.35584; @Pang:1997; @Luo:1993; @DBLP:journals/corr/nesterov16linearnon; @Bolte:2006:LIN:1328019.1328299; @HuiZhang16a]. Various error bounds have been exploited in many studies to analyze the convergence of optimization algorithms. For example, [@Luo:1992a; @Luo:1992b; @Luo:1993] established the asymptotic linear convergence of a class of feasible descent algorithms for smooth optimization, including coordinate descent method and projected gradient method, based on a local error bound condition. Their results on coordinate descent method were further extended for a more general class of objective functions and constraints in [@TsengYun:2009; @TsengYun:2010]. [@DBLP:journals/jmlr/WangL14] showed that a global error bound holds for a family of non-strongly convex and smooth objective functions for which feasible descent methods can achieve a global linear convergence rate. Recently, these error bounds have been generalized and leveraged to show faster convergence for structured convex optimization that consists of a smooth function and a simple non-smooth function [@DBLP:conf/nips/HouZSL13; @ZhouSo15; @DBLP:conf/icml/ZhouZS15]. We would like to emphasize that the aforementioned error bounds are different from the local error bound explored in this paper. They bound the distance of a point to the optimal set by the norm of the projected gradient or proximal gradient at the point, thus requiring the smoothness of the objective function. In contrast, we bound the distance of a point to the optimal set by its objective residual with respect to the optimal value, covering a much broader family of functions. More recently, there appear many studies that consider smooth optimization or composite smooth optimization problems whose objective functions satisfy different error bound conditions, growth conditions or other non-degeneracy conditions and established the linear convergence rates of several first-order methods including proximal-gradient method, accelerated gradient method, prox-linear method and so on [@DBLP:journals/corr/GongY14; @DBLP:journals/corr/nesterov16linearnon; @HuiZhang:2015; @HuiZhang16a; @Karimi15; @Drusvyatskiy16a; @Drusvyatskiy16b; @DBLP:conf/nips/HouZSL13; @DBLP:conf/icml/ZhouZS15]. The relative strength and relationships between some of those conditions are studied by [@DBLP:journals/corr/nesterov16linearnon] and [@HuiZhang16a]. For example, the authors in [@DBLP:journals/corr/nesterov16linearnon] showed that under the smoothness assumption the second order growth condition is equivalent to the error bound condition in [@DBLP:journals/jmlr/WangL14]. It was brought to our attention that the local error bound condition in the present paper is closely related to metric subregularity of subdifferentials [@artacho:2008; @kruger2015; @drusmordnhgia; @RePEc:spr:jglopt:v:63:y:2015:i:4:p:777-795]. Nevertheless, to our best knowledge, this is the first work that leverages the considered local error bound to improve the convergence of subgradient method for non-smooth and non-strongly convex optimization. The aforementioned works assume the objective function is smooth or is a summation of a smooth function and a simple non-smooth function, which is not only critical for defining some error bounds they assumed but also necessary for obtaining the linear rates for the algorithms they studied. On the contrary, we focus on minimizing non-smooth objective functions with less structure using the SG method. Exploiting the error bound , we develop restarting scheme in the RSG method and obtain different convergence rates from the works mentioned above. Moreover, even without condition  or any other error bounds, we still obtain a new complexity that can be better than the standard complexity of SG method in some scenarios (see Corollary \[lem:3\] and the remark thereafter.) [@DBLP:journals/mp/GilpinPS12] established a polyhedral error bound condition (presented in Lemma \[lem:polyeb\] below). Using this polyhedral error bound condition, they study a two-person zero-sum game and proposed an restarted first-order method based on Nesterov’s smoothing technique [@Nesterov:2005:SMN] that can find the Nash equilibrium in a linear convergence rate. This error bound condition was shown in [@DBLP:journals/mp/GilpinPS12] to hold for the objective function whose epigraph is polyhedral and the domain is a bounded polytope [@DBLP:journals/mp/GilpinPS12]. Here, we slightly generalize their result to allow the domain to be an unbounded polyhedron which is the case for many important applications. In addition, we consider a general condition that contains this polyhedral error bound condition as a special case and we try to solve the general problem rather than the bilinear saddle-point problem in [@DBLP:journals/mp/GilpinPS12]. In his recent work [@2014arXivJames; @2015arXivJames; @2016arXivJames], Renegar presented a framework of applying first-order methods to general conic optimization problems by transforming the original problem into an equivalent convex optimization problem with only linear equality constraints and a Lipschitz-continuous objective function. This framework greatly extends the applicability of first-order methods to the problems with general linear inequality constraints and leads to new algorithms and new iteration complexity. One of his results related to this work is Corollary 3.4 of [@2015arXivJames], which implies, if the objective function has a polyhedral epigraph and the optimal objective value is known beforehand, a subgradient method can have a linear convergence rate. Compared to his work, our method does not need to the know the optimal value but instead requires an upper bound of the initial optimality gap and knowing a growth constant in the error bound condition of the polyhedral objective function. In addition, our results include improved iteration complexities for a broader family of objective functions than the ones with a polyhedral epigraph. More recently, [@2015arXivRobert] proposed a new SG method by assuming that a strict lower bound of $f_*$, denoted by $f_{slb}$, is known and $f$ satisfies a growth condition, $\|\w-\w^*\|_2\leq \mathcal{G}\cdot(f(\w)-f_{slb})$, where $\w^*$ is the optimal solution closest to $\w$ and $\mathcal{G}$ is a growth rate constant depending on $f_{slb}$. Using a novel step size that incorporates $f_{slb}$, for non-smooth optimization, their SG method achieves an iteration complexity of $O(\mathcal{G}^2(\frac{\log H}{\epsilon'} + \frac{1}{\epsilon'^2}))$ for finding a solution $\hat{\w}$ such that $f(\hat{\w}) - f_*\leq \epsilon'(f_* - f_{slb})$, where $H = \frac{f(\w_0)-f_{slb}}{f_*-f_{slb}}$ and $\w_0$ is the initial solution. We note that there are several key differences in the theoretical properties and implementations between our work and [@2015arXivRobert]: (i) Their growth condition has a similar form to the inequality  we prove for a general function but there are still noticeable differences in the both sides and the growth constants. (ii) The convergence results in [@2015arXivRobert] are established based on finding an solution $\hat{\w}$ with a relative error of $\epsilon'$ while we consider absolute error. (iii) By rewriting the convergence results in [@2015arXivRobert] in terms of absolute accuracy $\epsilon$ with $\epsilon=\epsilon'(f_* - f_{slb})$, the complexity in [@2015arXivRobert] will strongly depend on $f_*-f_{slb}$ and can be worse than ours if $f_*-f_{slb}$ is large. (iv) Their SG descent method keeps track of the best solution up to the current iteration while our method maintains the average of all solutions within each epoch. We will compared our RSG method with the method in [@2015arXivRobert] with more details in Section \[sec:RSG\]. Restarting and multi-stage strategies have been utilized to achieve the (uniformly) optimal theoretical complexity of (stochastic) SG methods when $f$ is strongly convex [@DBLP:journals/siamjo/GhadimiL13; @NIPS2012_4543; @hazan-20110-beyond] [or uniformly convex [@Nesterov:2014:uniform_convex]]{}. Here, we show that restarting can be still helpful even without [uniform or]{} strong convexity. Furthermore, in all the algorithms proposed in [@DBLP:journals/siamjo/GhadimiL13; @NIPS2012_4543; @hazan-20110-beyond; @Nesterov:2014:uniform_convex], the number of iterations per stage increases between stages while our algorithm uses the same number of iterations in all stages. This provides a different possibility of designing restarted algorithms for a better complexity. Preliminaries {#sec:main} ============= In this section, we define some notations used in this paper and present the main assumptions needed to establish our results. We use $\partial f(\w)$ to denote the set of subgradients (the subdifferential) of $f$ at $\w$. Let $\mathcal G(\cdot)$ denote a first-order oracle that returns a subgradient of $f(\cdot)$, namely, $\G(\w) \in \partial f(\w)$ for any $\w\in\Omega$. Since the objective function is not necessarily strongly convex, the optimal solution is not necessarily unique. We denote by $\Omega_*$ the optimal solution set and by $f_*$ the unique optimal objective value. We denote by $\|\cdot\|_2$ the Euclidean norm in $\mathbb{R}^d$. Throughout the paper, we make the following assumptions unless stated otherwise. \[ass:rsg\] For the convex minimization problem (\[eqn:prob0\]), we assume - For any $\w_0\in\Omega$, we know a constant $\epsilon_0\geq 0$ such that $f(\w_0) - f_*\leq \epsilon_0$. - There exists a constant $G$ such that $\|\G(\w)\|_2\leq G$ for any $\w\in\Omega$. - $\Omega_*$ is a non-empty convex compact set. We make several remarks about the above assumptions: (i) Assumption \[ass:rsg\].a is equivalent to assuming we know a lower bound of $f_*$ which is one of the assumptions made in [@2015arXivRobert]. In machine learning applications, $f_*$ is usually bounded below by zero, i.e., $f_*\geq 0$, so that $\epsilon_0 = f(\w_0)$ for any $\w_0\in\R^d$ will satisfy the condition; (ii) Assumption \[ass:rsg\].b is a standard assumption also made in many previous subgradient-based methods. (iii) Assumption \[ass:rsg\].c simply assumes the optimal set is closed and bounded. Let $\w^*$ denote the closest optimal solution in $\Omega_*$ to $\w$ measured in terms of norm $\|\cdot\|_2$, i.e., $$\w^* := \arg\min_{\u\in\Omega_*}\|\u - \w\|^2_2.$$ Note that $\w^*$ is uniquely defined for any $\w$ due to the convexity of $\Omega_*$ and that $\|\cdot\|_2^2$ is strongly convex. We denote by $\V_\epsilon$ the $\epsilon$-level set of $f(\w)$ and by $\S_\epsilon$ the $\epsilon$-sublevel set of $f(\w)$, respectively, i.e., $$\begin{aligned} \label{def:LeSe} \V_\epsilon := \{\w\in\Omega: f(\w) = f_* + \epsilon\}\quad\text{and}\quad\S_\epsilon := \{\w\in\Omega: f(\w) \leq f_* + \epsilon\}.\end{aligned}$$ Since $\Omega_*$ is assumed to be bounded and $f$ is convex, it follows from [@rockafellar1970convex] (Corollary 8.7.1) that the sublevel set $\S_\epsilon$ must be bounded for any $\epsilon\geq 0$ and so is the level set $\V_\epsilon$. Let $B_\epsilon$ be the maximum distance between the points in the $\epsilon$-level set $\V_\epsilon$ and the optimal set $\Omega_*$, i.e., $$\label{eqn:keyB} B_{\epsilon}:= \max_{\w\in\V_\epsilon}\min_{\u\in\Omega_*}\|\w - \u\|_2 = \max_{\w\in\V_\epsilon}\|\w - \w^*\|_2.$$ We follow the convention that $B_{\epsilon}=0$ if $\V_\epsilon=\emptyset$. Because $\V_\epsilon$ and $\Omega_*$ are bounded, $B_\epsilon$ is finite. Let $\w_{\epsilon}^\dagger$ denote the closest point in the $\epsilon$-sublevel set to $\w$, i.e., $$\label{eq:wproj} \begin{aligned} \w_\epsilon^{\dagger}:=\arg\min_{\u\in\S_\epsilon}\|\u - \w\|^2_2\end{aligned}$$ Denote by $\Omega\backslash\S=\{\w\in\Omega: \w\not\in\S\}$. It is easy to show that $\w_\epsilon^{\dagger}\in\V_\epsilon$ when $\w\in\Omega\backslash\S_\epsilon$ (using the optimality condition of ). Given $\w\in\Omega$, we denote the normal cone of $\Omega$ at $\w$ by $\mathcal{N}_{\Omega}(\w)$. Formally, $\mathcal N_\Omega(\w)= \{\v\in\R^d: \v^{\top}(\u - \w)\leq 0, \forall \u\in\Omega\}$. Define $\|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_2$ as $$\label{eq:bdsubgrad} \|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_2:=\min_{\mathbf g\in\partial f(\w),\mathbf v\in\mathcal{N}_{\Omega}(\w)} \|\mathbf g+\mathbf v\|_2.$$ Note that $\w\in\Omega_*$ if and only if $\|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_2=0$. Therefore, we call $\|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|$ the *first-order optimality residual* of at $\w\in\Omega$. Given any $\epsilon>0$ such that $\V_\epsilon\neq\emptyset$, we define a constant $\rho_\epsilon$ as $$\label{eq:defrho} \rho_\epsilon:=\min_{\w\in\V_\epsilon} \|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_2.$$ Given the notations above, we provide the following lemma which is the key to our analysis. \[lem:1\] For any $\epsilon>0$ such that $\V_\epsilon\neq\emptyset$ and any $\w\in\Omega$, we have $$\label{eqn:keyii} \begin{aligned} \|\w - \w_\epsilon^\dagger\|_2\leq \frac{1}{\rho_\epsilon}(f(\w) - f(\w_\epsilon^\dagger)). \end{aligned}$$ Since the conclusion holds trivially if $\w\in\S_\epsilon$ (so that $\w^\dagger_\epsilon=\w$), we assume $\w\in\Omega\backslash\S_\epsilon$. According to the first-order optimality conditions of (\[eq:wproj\]), there exist a scalar $\zeta\geq 0$ (the Lagrangian multiplier of the constraint $f(\u)\leq f_* + \epsilon$ in (\[eq:wproj\])), a subgradient $\mathbf{g}\in \partial f(\w^\dagger_\epsilon)$ and a vector $\mathbf v\in\mathcal{N}_{\Omega}(\w^\dagger_\epsilon)$ such that $$\label{eqn:o2} \begin{aligned} &\w^\dagger_\epsilon - \w + \zeta \mathbf{g}+\mathbf v=0. \end{aligned}$$ The definition of normal cone leads to $ (\w_\epsilon^\dagger - \w)^{\top}\mathbf v\geq 0$. This inequality and the convexity of $f(\cdot)$ imply $$\zeta\left(f(\w) - f(\w^\dagger_\epsilon)\right)\geq \zeta(\w - \w^\dagger_\epsilon)^{\top}\mathbf g \geq (\w - \w^\dagger_\epsilon)^{\top}\left(\zeta\mathbf g+\mathbf v\right) = \|\w- \w^\dagger_\epsilon\|_2^2$$ where the equality is due to . Since $\w\in\Omega\backslash\S_\epsilon$, we must have $\|\w- \w^\dagger_\epsilon\|_2>0$ so that $\zeta>0$. Therefore, $\w^\dagger_\epsilon\in\V_\epsilon$ by complementary slackness. Dividing the inequality above by $\zeta$ gives $$\begin{aligned} \label{eq:lemma2} f(\w) - f(\w^\dagger_\epsilon)\geq \frac{\|\w- \w^\dagger_\epsilon\|_2^2}{\zeta}=\|\w- \w^\dagger_\epsilon\|_2\|\mathbf g+\mathbf v/\zeta \|_2\geq\rho_\epsilon\|\w- \w^\dagger_\epsilon\|_2,\end{aligned}$$ where the equality is due to and the last inequality is due to the definition of $\rho_\epsilon$ in . The lemma is then proved. The inequality takes a similar form to the growth condition, $$\begin{aligned} \label{eqn:FreundCond} \|\w-\w^*\|_2\leq \mathcal{G}\cdot(f(\w)-f_{slb}),\end{aligned}$$ where $f_{slb}$ is a strict lower bound of $f^*$, by Freund and Lu [@2015arXivRobert] but with some striking differences: the left-hand side is the distance of $\w$ to the optimal set in while it is the distance of $\w$ to the $\epsilon$-sublevel set in ; the right-hand side is the objective gap with respect to $f_{slb}$ in and it is the objective gap with respect to $f^*$ in ; the growth constant $\mathcal{G}$ in varies with $f_{slb}$ and $\rho_\epsilon$ in may depend on $\epsilon$ in general. The inequality in (\[eqn:keyii\]) is the key to achieve improved convergence by RSG, which hinges on the condition that the first-order optimality residual on the $\epsilon$-level set is lower bounded. It is important to note that (i) the above result depends on $f$ rather than the optimization algorithm applied; and (ii) the above result can be generalized to use other norm such as the $p$-norm $\|\w\|_p$ ($p\in(1,2]$) to measure the distance between $\w$ and $\w^\dagger_\epsilon$ (Section \[sec:variant\]) and use the corresponding dual norm to define the lower bound of the residual in and . This generalization allows us to design mirror decent [@Nemirovski:2009:RSA:1654243.1654247] variant of RSG. We defer the details into Section \[sec:variant\]. To our best knowledge, this is the first work that leverages the lower bound of the optimal residual to improve the convergence for non-smooth convex optimization. In next several sections, we will exhibit the value of $\rho_\epsilon$ for different classes of problems and discuss its impact on the convergence. Restarted SubGradient (RSG) Method and Its Complexity for General Problem {#sec:RSG} ========================================================================= In this section, we present the proposed restarted subgradient (RSG) method and prove its general convergence result using Lemma \[lem:1\]. In next sections, we will present improved convergence of RSG for problems of different classes. The steps of RSG are presented in Algorithm \[alg:1\] where SG is a subroutine of projected subgradient descent given in Algorithm \[alg:0\] and $\Pi_{\Omega}[\w]$ is defined as $$\Pi_{\Omega}[\w] = \arg\min_{\u\in\Omega}\|\u - \w\|_2^2.$$ [The values of $K$ and $t$ in RSG will be revealed later for proving the convergence of RSG to an $2\epsilon$-optimal solution. ]{} The RSG algorithm runs in stages and calls SG once in each stage. The subroutine SG performs projected subgradient descent with a fixed step size $\eta$ for a [**fixed number of iterations $t$**]{}, using the solution returned by SG from the previous stage of RSG as the starting point. The RSG algorithm geometrically decreases the step size $\eta_k$ between stages. [The output solution of RSG is the solution returned by SG in the $K$-th stage, i.e., $\w_K$.]{} The number of iterations $t$ is the only varying parameter in RSG that depends on the classes of problems. The parameter $\alpha$ could be any value larger than $1$ (e.g., 2) and it only has a small influence on the iteration complexity. We emphasize that (i) RSG is a generic algorithm that is applicable to a broad family of non-smooth and/or non-strongly convex problems without changing updating schemes except for one tuning parameter, the number of iterations per stage, whose best value varies with problems; (ii) RSG has different variants with different subroutine in stages. In fact, we can use other optimization algorithms than subgradient descent as the subroutine in Algorithm \[alg:1\], as long as a similar convergence result to Lemma \[lem:GD\] is guaranteed. Examples include SG as shown in Algorithm \[alg:0\], dual averaging [@Nesterov:2009:PSM:1530733.1530741], and the regularized dual averaging [@NIPS2012_4543] [^2]. In the following discussions, we will focus on using SG as the subroutine unless stated otherwise. **Input**: a step size $\eta$, the number of iterations $T$, and the initial solution $\w_1\in\Omega$ Query the subgradient oracle to obtain $\G(\w_t)$ Update $\w_{t+1} =\Pi_{\Omega}[ \w_t - \eta \G(\w_t)]$ **Output**: $\wh_T = \sum_{t=1}^T\frac{\w_t}{T}$ **Input**: the number of stages $K$ and the number of iterations $t$ per-stage, $\w_0 \in\Omega$, and $\alpha>1$. Set $\eta_1=\epsilon_0/(\alpha G^2)$, where $\epsilon_0$ is from Assumption \[ass:rsg\].a Call subroutine SG to obtain $\w_k = \text{SG}(\w_{k-1}, \eta_k, t)$ Set $\displaystyle \eta_{k+1} = \eta_{k}/\alpha$ **Output**: $\w_K$ It was also brought to our attention that the current common practice for training deep neural network (yet with no formal justification) is very similar to the proposed RSG. In particular, one usually employs stochastic subgradient with momentum (with similar convergence guarantee to SG for non-smooth optimization [@yang2016unified]) with a fixed step size by a certain number of iterations and then decreases the step size by a certain number of times [@DBLP:conf/nips/KrizhevskySH12]. Next, we establish the convergence of RSG. It relies on the convergence result of the SG subroutine which is given in the lemma below. [@zinkevich-2003-online; @opac-b1104789] \[lem:GD\] If Algorithm \[alg:0\] runs for $T$ iterations, we have, for any $\w\in\Omega$, $$f(\wh_T) - f(\w) \leq \frac{G^2\eta}{2} + \frac{\|\w_{1} - \w\|_2^2}{2\eta T}.$$ We omit the proof because it follows a standard analysis and can be found in cited papers. With the above lemma, we can prove the following convergence of RSG. We would like to emphasize here the difference between our analysis of convergence for RSG and previous analysis of SG. In the previous analysis, $\w$ is set to an optimal solution $\w_*$ in Lemma \[lem:GD\]. Assuming $\|\w_1 - \w_*\|_2\leq D$, one can optimize $\eta$ to obtain an $O(GD/\sqrt{T})$ convergence rate for SG. In our analysis, we leverage Lemma \[lem:1\] to bound $\|\w_{k-1} - \w\|_2$ for the $k$-th stage and prove the convergence of RSG by induction. \[thm:GDr\] Suppose Assumption \[ass:rsg\] holds. If $t\geq \frac{\alpha^2 G^2}{\rho_\epsilon^2}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$ in Algorithm \[alg:1\], with at most $K$ stages, Algorithm \[alg:1\] returns a solution $\w_K$ such that $f(\w_K)-f_*\leq 2\epsilon$. In other word, the total number of iterations for Algorithm \[alg:1\] to find an $2\epsilon$-optimal solution is at most $T= O\left(t\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$ where $t\geq \frac{\alpha^2 G^2}{\rho_\epsilon^2}$. In particular, if $\frac{\alpha^2 G^2}{\rho_\epsilon^2}\leq t=O\left( \frac{\alpha^2 G^2}{\rho_\epsilon^2}\right)$, the total number of iterations for Algorithm \[alg:1\] to find an $2\epsilon$-optimal solution is at most $T= O\left(\frac{\alpha^2G^2}{\rho_\epsilon^2}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$. Let $\w_{k,\epsilon}^\dagger$ denote the closest point to $\w_k$ in the $\epsilon$-sublevel set. Let $\epsilon_k:=\frac{\epsilon_0}{\alpha^k}$ so that $\eta_k = \epsilon_k/G^2$ because $\eta_1=\epsilon_0/(\alpha G^2)$ and $\eta_{k+1} = \eta_{k}/\alpha$. We will show by induction that $$\label{eqn:mainineq1} f(\w_k) - f_*\leq \epsilon_k +\epsilon$$ for $k=0,1,\dots,K$ which leads to our conclusion if we let $k=K$. Note that holds obviously for $k=0$. Suppose it holds for $k-1$, namely, $f(\w_{k-1}) - f_*\leq \epsilon_{k-1} + \epsilon$. We want to prove for $k$. We apply Lemma \[lem:GD\] to the $k$-th stage of Algorithm \[alg:1\] and get $$\label{eqn:mainineq2} f(\w_k) - f(\w_{k-1, \epsilon}^\dagger)\leq \frac{G^2\eta_k}{2} + \frac{\|\w_{k-1} - \w_{k-1,\epsilon}^\dagger\|_2^2}{2\eta_k t}.$$ We now consider two cases for $\w_{k-1}$. First, assume $f(\w_{k-1}) - f_* \leq \epsilon$, i.e., $\w_{k-1}\in\mathcal S_\epsilon$. Then $\w^\dagger_{k-1,\epsilon} = \w_{k-1}$ and $f(\w_k) - f(\w^\dagger_{k-1,\epsilon}) \leq \frac{G^2\eta_k}{2} = \frac{\epsilon_k}{2}$. As a result, $$f(\w_k) - f_* \leq f(\w_{k-1,\epsilon}^\dagger) - f_* + \frac{\epsilon_k}{2} \leq \epsilon + \epsilon_k$$ Next, we consider the case that $f(\w_{k-1}) - f_*>\epsilon$, i.e., $\w_{k-1}\not\in\S_\epsilon$. Then we have $f(\w_{k-1,\epsilon}^\dagger) = f_* + \epsilon$. By Lemma \[lem:1\], we have $$\label{eqn:mainineq3} \begin{aligned} \|\w_{k-1} - \w^\dagger_{k-1,\epsilon}\|_2&\leq \frac{1}{\rho_\epsilon}(f(\w_{k-1}) - f(\w_{k-1,\epsilon}^\dagger))=\frac{ f(\w_{k-1}) - f_* + (f_* - f(\w^\dagger_{k-1,\epsilon}))}{\rho_\epsilon}\\ &\leq \frac{\epsilon_{k-1}+\epsilon - \epsilon}{\rho_\epsilon}. \end{aligned}$$ Combining and and using the facts that $\displaystyle \eta_k =\frac{\epsilon_k}{G^2}$ and $\displaystyle t \geq \frac{\alpha^2 G^2}{\rho_\epsilon^2}$, we have $$f(\w_k) - f(\w^\dagger_{k-1,\epsilon})\leq \frac{\epsilon_k}{2} + \frac{\epsilon_{k-1}^2}{2\epsilon_k\alpha^2} = \epsilon_k $$ which, together with the fact that $ f(\w^\dagger_{k-1,\epsilon})=f_*+\epsilon$, implies for $k$. Therefore, by induction, we have holds for $k=1,2,\dots,K$ so that $$f(\w_K) - f_*\leq \epsilon_K + \epsilon = \frac{\epsilon_0}{\alpha^K} +\epsilon\leq 2\epsilon,$$ where the last inequality is due to the definition of $K$. In Theorem \[thm:GDr\], the iteration complexity of RSG for the general problem is given in terms of $\rho_\epsilon$. Next, we show that $\rho_\epsilon\geq\frac{\epsilon}{B_\epsilon}$ which allows us to choose the number of iterations $t$ in each stage according to $B_\epsilon$ instead of $\rho_\epsilon$ so that corresponding complexity can be given in terms of $B_\epsilon$. \[lem:lowf\] For any $\epsilon>0$ such that $\V_\epsilon\neq\emptyset$, we have $ \rho_\epsilon\geq \frac{\epsilon}{B_\epsilon}$, where $B_\epsilon$ is defined in (\[eqn:keyB\]), and for any $\w\in\Omega$ $$\label{eqn:keyto} \|\w - \w_\epsilon^\dagger\|_2 \leq \frac{\|\w_\epsilon^\dagger - \w^*_\epsilon\|_2}{\epsilon}(f(\w) - f(\w_\epsilon^\dagger))\leq \frac{B_\epsilon}{\epsilon}(f(\w) - f(\w_\epsilon^\dagger)),$$ where $\w_\epsilon^*$ is the closest point in $\Omega_*$ to $\w^\dagger_\epsilon$. ![A geometric illustration of the inequality (\[eqn:keyto\]), where $dist(w^\dagger_\epsilon,\Omega_*) = |w^\dagger_\epsilon - w^*_\epsilon|$. []{data-label="fig:2"}](geometric.eps) Given any $\u\in\V_\epsilon$, let $\mathbf g_\u$ be any subgradient in $\partial f(\u)$ and $\mathbf v_\u$ be any vector in $\mathcal{N}_{\Omega}(\u)$. By the convexity of $f(\cdot)$ and the definition of normal cone, we have $$f(\u^*) - f(\u)\geq (\u^* - \u)^{\top}\mathbf g_\u \geq(\u^* - \u)^{\top}\left(\mathbf g_\u+\mathbf v_\u\right),$$ where $\u^*$ is the closest point in $\Omega_*$ to $\u$. This inequality further implies $$\label{eq:lowboundeq1} \begin{aligned} \|\u^* - \u\|_2&\|\mathbf g_\u+\mathbf v_\u\|_2\geq f(\u) - f(\u^*) = \epsilon, \quad\forall \g_\u\in\partial f(\u) \text{ and } \v_\u\in\mathcal N_{\Omega}(\u) \end{aligned}$$ where the equality is because $\u\in\V_\epsilon$. By and the definition of $B_\epsilon$, we obtain $$B_\epsilon\|\mathbf g_\u+\mathbf v_\u\|_2\geq \epsilon\Longrightarrow \|\mathbf g_\u+\mathbf v_\u\|_2\geq \epsilon/B_\epsilon.$$ Since $\mathbf g_\u+\mathbf v_\u$ can be any element in $\partial f(\u)+\mathcal{N}_{\Omega}(\u)$, we have $\rho_\epsilon\geq \frac{\epsilon}{B_\epsilon}$ by the definition . To prove (\[eqn:keyto\]), we assume $\w\in\Omega\backslash\S_\epsilon$ and thus $\w^\dagger_\epsilon\in\V_\epsilon$; otherwise it is trivial. In the proof of Lemma \[lem:1\], we have shown that (see ) there exists $\g\in\partial f(\w_\epsilon^\dagger)$ and $\v\in\mathcal N_\Omega(\w^\dagger_\epsilon)$ such that $ f(\w) - f(\w^\dagger_\epsilon)\geq\|\w- \w^\dagger_\epsilon\|_2\|\mathbf g+\mathbf v/\zeta \|_2 $, which, according to (\[eq:lowboundeq1\]) with $\u=\w^\dagger_\epsilon$, $\mathbf g_\u=\mathbf g$ and $\mathbf v_\u=\mathbf v/\zeta$, leads to (\[eqn:keyto\]). A geometric explanation of the inequality (\[eqn:keyto\]) in one dimension is shown in Figure \[fig:2\]. With Lemma \[lem:lowf\], the iteration complexity of RSG can be stated in terms of $B_\epsilon$ in the following corollary of Theorem \[thm:GDr\]. \[lem:3\] Suppose Assumption \[ass:rsg\] holds. The iteration complexity of RSG for obtaining an $2\epsilon$-optimal solution is $O(\frac{\alpha^2G^2B_\epsilon^2}{\epsilon^2}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil)$ provided $\frac{\alpha^2G^2B_\epsilon^2}{\epsilon^2}\leq t =O\left( \frac{\alpha^2G^2B_\epsilon^2}{\epsilon^2}\right) $ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$. [**Remark: Compared to the standard SG**]{} whose iteration complexity is known as $O(\frac{G^2\|\w_0 - \w^*_0\|_2^2}{\epsilon^2})$ for achieving an $2\epsilon$-optimal solution, RSG’s iteration complexity in Corollary \[lem:3\] depends on $B_\epsilon^2$ instead of $\|\w_0 - \w_0^*\|_2^2$ and only has a logarithmic dependence on $\epsilon_0$, the upper bound of $f(\w_0) - f_*$. When the initial solution is far from the optimal set so that $B_\epsilon^2\ll\|\w_0 - \w_0^*\|_2^2$, the proposed RSG can be much faster. In some special cases, e.g., when $f$ satisfies the local error bound condition (\[eqn:ler\]) so that $B_\epsilon=\Theta(\epsilon^\theta)$ with $\theta\in(0,1]$, RSG only needs $O\left(\frac{1}{\epsilon^{2(1-\theta)}}\log\left(\frac{1}{\epsilon}\right)\right)$ iterations (see Section \[sec:ler\]), which as a better dependency on $\epsilon$ than the complexity of standard SG method. [**Remark: Compared to the SG method in [@2015arXivRobert]**]{} whose iteration complexity is $O(G^2\mathcal{G}^2(\frac{\log H}{\epsilon'} + \frac{1}{\epsilon'^2}))$ for finding a solution $\hat{\w}$ such that $f(\hat{\w}) - f_*\leq \epsilon'(f_* - f_{slb})$, where $f_{slb}$ and $\mathcal{G}$ are defined in and $H = \frac{f(\w_0)-f_{slb}}{f_*-f_{slb}}$, our RSG can be better if $f_*-f_{slb}$ is large. To see this, we represent the complexity in [@2015arXivRobert] in terms of the absolute error $\epsilon$ with $\epsilon=\epsilon'(f_* - f_{slb})$ and obtain $O(G^2\mathcal{G}^2(\frac{(f_*-f_{slb})\log H}{\epsilon} + \frac{(f_*-f_{slb})^2}{\epsilon^2}))$. If the gap $f_* - f_{slb}$ is large, e.g., $O(f(\w_0)-f_{slb})$, the second term is dominating, which is at least $O(\frac{G^2\|\w_0 - \w^*_0\|_2^2}{\epsilon^2})$ due to the definition of $\mathcal{G}$ in $\eqref{eqn:FreundCond}$. This complexity has the same order of magnitude as the standard SG method so that RSG can be better for the reason in the last remark. When the gap $f_* - f_{slb}$ is small, e.g. $O(\epsilon)$, the first term is dominating, which is $O(G^2\mathcal{G}^2\log (\frac{\epsilon_0}{\epsilon} ))$. In this case, RSG has a lower complexity when $\mathcal{G}^2\geq\frac{B_\epsilon^2}{\epsilon^2}$. Note that, in general, $\mathcal{G}$ is non-increasing in $f_* - f_{slb}$ but the exact dependency is not clear. However, compare to the standard SG and the method in [@2015arXivRobert], RSG does require knowing additional information about $f$. In particular, the issue with RSG is that its improved complexity is obtained by choosing $t =O\left(\frac{\alpha^2G^2B_\epsilon^2}{\epsilon^2}\right)$ which requires knowing the order of magnitude of $B_\epsilon$, if not its exact value. To address the issue of unknown $B_\epsilon$ for general problems, in the next section, we consider the family of problems that admit a local error bound and show that the requirement of knowing $B_\epsilon$ is relaxed to knowing some particular parameters related to the local error bound. Complexity for Some Classes of Non-smooth Non-strongly Convex Optimization {#sec:spec} ========================================================================== In this section, we consider a particular family of problems that admit local error bounds and show the improved iteration complexities of RSG compared to standard SG method. Complexity for the Problems with Local Error Bounds {#sec:ler} --------------------------------------------------- We first define local error bound of the objective function. \[def:ler\] We say $f(\cdot)$ admits a **local error bound** on the $\epsilon$-sublevel set $\S_\epsilon$ if $$\label{eqn:ler} \|\w - \w^*\|_2 \leq c(f(\w) - f_*)^{\theta}, \quad \forall \w\in\S_\epsilon,$$ where $\w^*$ is the closet point in $\Omega_*$ to $\w$, $\theta\in(0,1]$ and $c>0$ are constants. Because $\S_{\epsilon_2}\subset\S_{\epsilon_1}$ for $\epsilon_2\leq\epsilon_1$, if holds for some $\epsilon$, it will always hold when $\epsilon$ decreases to zero with the same $\theta$ and $c$, which is the case we are interested in. If the problem admits a local error bound like (\[eqn:ler\]), RSG can achieve a better iteration complexity than $\widetilde O(1/\epsilon^2)$. In particular, the property (\[eqn:ler\]) implies $$\label{eqn:Bebound} B_\epsilon\leq c\epsilon^{\theta}.$$ Replacing $B_\epsilon$ in Corollary \[lem:3\] by this upper bound and choosing $t=\frac{\alpha^2G^2c^2}{\epsilon^{2(1-\theta)}}$ in RSG if $c$ and $\theta$ are known, we obtain the following complexity of RSG. \[lem:leb\_corr\] Suppose Assumption \[ass:rsg\] holds and $f(\cdot)$ admits a local error bound on $\S_\epsilon$. The iteration complexity of RSG for obtaining an $2\epsilon$-optimal solution is $O\left(\frac{\alpha^2G^2c^2}{\epsilon^{2(1-\theta)}}\log_\alpha\left(\frac{\epsilon_0}{\epsilon}\right)\right)$ provided $t =\frac{\alpha^2G^2c^2}{\epsilon^{2(1-\theta)}}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$. [[@Nesterov:2014:uniform_convex] considered subgradient methods for with $f$ being *uniformly convex*, namely, $$f(\alpha\w+ (1-\alpha)\v)\leq \alpha f(\w)+(1-\alpha)f(\v)-\frac{1}{2}\mu\alpha(1-\alpha)[\alpha^{\rho-1}+(1-\alpha)^{\rho-1}]\|\w-\v\|_2^\rho$$ for any $\w$ and $\v$ in $\Omega$ and any $\alpha\in[0,1]$[^3], where $\rho\in[2,+\infty]$ and $\mu\geq0$. In this case, the method by [@Nesterov:2014:uniform_convex] has an iteration complexity of $O\left(\frac{G^2}{\mu^{2/\rho}\epsilon^{2(\rho-1)/\rho}}\right)$. The uniform convexity of $f$ further implies $f(\w)-f_*\geq\frac{1}{2}\mu\|\w-\w^*\|_2^\rho$ for any $\w\in\Omega$ so that $f(\cdot)$ admits a local error bound on the $\epsilon$-sublevel set $\S_\epsilon$ with $c=\left(\frac{2}{\mu}\right)^{\frac{1}{\rho}}$ and $\theta=\frac{1}{\rho}$. Therefore, our RSG has a complexity of $O\left(\frac{G^2}{\mu^{2/\rho}\epsilon^{2(\rho-1)/\rho}}\log(\frac{\epsilon_0}{\epsilon})\right)$ according to Corollary \[lem:leb\_corr\]. Compared to [@Nesterov:2014:uniform_convex], our complexity is higher by a logarithmic factor. However, we only require the local error bound property of $f$ that is weaker than uniform convexity.]{} Next, we will consider different convex optimization problems that admit a local error bound on $\S_\epsilon$ with different $c$ and $\theta$ and show the faster convergence of RSG when applied to these problems. Linear Convergence for Polyhedral Convex Optimization {#sec:poly} ----------------------------------------------------- In this subsection, we consider a special family of non-smooth and non-strongly convex problems where the epigraph of $f(\cdot)$ over $\Omega$ is a polyhedron. In this case, we call a **polyhedral convex minimization** problem. We show that, in polyhedral convex minimization problem, $f(\cdot)$ has a linear growth property and admits a local error bound with $\theta=1$ so that $B_\epsilon\leq c\epsilon$ for a constant $c$. \[lem:polyeb\] Suppose $\Omega$ is a polyhedron and the epigraph of $f(\cdot)$ is also polyhedron. There exists a constant $\kappa>0$ such that $$\|\w - \w^*\|_2\leq \frac{f(\w)-f_*}{\kappa}, \quad\forall \w\in\Omega.$$ Thus, $f(\cdot)$ admits a local error bound on $\S_\epsilon$ with $\theta=1$ and $c=\frac{1}{\kappa}$ [^4] (so $B_\epsilon \leq \frac{\epsilon}{\kappa}$) for any $\epsilon>0$. [**Remark:**]{} We remark that the above result can be extended to any valid norm to measure the distance between $\w$ and $\w_*$. The proof is included in the appendix. Lemma \[lem:polyeb\] above generalizes Lemma 4 by [@DBLP:journals/mp/GilpinPS12], which requires $\Omega$ to be a bounded polyhedron, to a similar result where $\Omega$ can be an unbounded polyhedron. This generalization is simple but useful because it helps the development of efficient algorithms based on this error bound for unconstrained problems without artificially including a box constraint. Different from [@DBLP:journals/mp/GilpinPS12] that used their Lemma 4 and Nesterov’s smoothing technique [@Nesterov:2005:SMN] to develop a linearly convergent algorithm for solving the Nash equilibrium of a two-person zero-sum games, we show below that Lemma \[lem:polyeb\] provides the basis for RSG to achieve a linear convergence for the polyhedral convex minimization problems. In fact, the following linear convergence of RSG can be obtained if we plugin the values of $\theta=1$ and $c=\frac{1}{\kappa}$ into Corollary \[lem:leb\_corr\]. \[cor:5\] Suppose Assumption 1 holds and is a polyhedral convex minimization problem. The iteration complexity of RSG for obtaining an $\epsilon$-optimal solution is $O(\frac{\alpha^2G^2}{\kappa^2}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil)$ provided $t = \frac{\alpha^2 G^2}{\kappa^2}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$. We want to point out that Corollary \[cor:5\] can be proved directly by replacing $\w^\dagger_{k-1,\epsilon}$ by $\w^*_{k-1}$ and replacing $\rho_\epsilon$ by $\kappa$ in the proof of Theorem \[thm:GDr\]. Here, we derive it as a corollary of a more general result. We also want to mention that, as shown by [@2015arXivJames], the linear convergence rate in Corollary \[cor:5\] can be also obtained by the SG method for the historically best solution, provided either $\kappa$ or $f^*$ is known. #### **Examples** Many non-smooth and non-strongly convex machine learning problems satisfy the assumptions of Corollary \[cor:5\], for example, [**$\ell_1$ or $\ell_\infty$ constrained or regularized piecewise linear loss minimization**]{}. In many machine learning tasks (e.g., classification and regression), there exists a set of data $\{(\x_i,y_i)\}_{i=1,2,\dots,n}$ and one often needs to solve the following empirical risk minimization problem $$\begin{aligned} \min_{\w\in\R^d}f(\w) \triangleq \frac{1}{n}\sum_{i=1}^n\ell(\w^{\top}\x_i, y_i) + R(\w),\end{aligned}$$ where $R(\w)$ is a regularization term and $\ell(z,y)$ denotes a loss function. We consider a special case where (a) $R(\w)$ is a $\ell_1$ regularizer, $\ell_\infty$ regularizer or an indicator function of a $\ell_1$/$\ell_\infty$ ball centered at zero; and (b) $\ell(z,y)$ is any piecewise linear loss function, including hinge loss $\ell(z,y)= \max(0, 1-yz)$, absolute loss $\ell(z,y) = |z- y|$, $\epsilon$-insensitive loss $\ell(z,y) = \max(|z-y| - \epsilon, 0)$, and etc [@DBLP:journals/ml/YangMJZ14Non]. It is easy to show that the epigraph of $f(\w)$ is a polyhedron if $f(\w)$ is defined as a sum of any of these regularization terms and any of these loss functions. In fact, a piecewise linear loss functions can be generally written as $$\label{eq:polyell} \ell(\w^{\top}\x, y) = \max_{1\leq j\leq m}a_{j}\w^{\top}\x+b_j,$$ where $(a_j,b_j)$ for $j=1,2,\dots,m$ are finitely many pairs of scalars. The formulation indicates that $\ell(\w^{\top}\x, y)$ is a piecewise affine function so that its epigraph is a polyhedron. In addition, the $\ell_1$ or $\ell_\infty$ norm is also a polyhedral function because we can represent them as $$\|\w\|_1 = \sum_{i=1}^d\max(w_i, -w_i), \quad \|\w\|_\infty=\max_{1\leq i\leq d}|w_i| = \max_{1\leq i\leq d}\max(w_i, -w_i).$$ Since the sum of finitely many polyhedral functions is also a polyhedral function, the epigraph of $f(\w)$ is a polyhedron. Another important family of problems whose objective function has a polyhedral epigraph is [**submodular function minimization**]{}. Let $V = \{1,\ldots, d\}$ be a set and $2^V$ denote its power set. A submodular function $F(A): 2^{V}\rightarrow\R$ is a set function such that $F(A) + F(B)\geq F(A\cup B)+ F(A\cap B)$ for all subsets $A, B\subseteq V$ and $F(\emptyset)=0$. A submodular function minimization can be cast into a non-smooth convex optimization using the Lovász extension [@DBLP:journals/ftml/Bach13]. In particular, let the base polyhedron $B(F)$ be defined as $$B(F) = \{\s\in\R^d, \s(V) = F(V), \forall A\subseteq V, \s(A)\leq F(A)\},$$ where $\s(A) = \sum_{i\in A}s_i$. Then the Lovász extension of $F(A)$ is $f(\w) = \max_{\s\in B(F)}\w^{\top}\s$, and $\min_{A\subseteq V}F(A) = \min_{\w\in[0,1]^d}f(\w)$. As a result, a submodular function minimization is essentially a non-smooth and non-strongly convex optimization with a polyhedral epigraph. Improved Convergence for Locally Semi-Strongly Convex Problems -------------------------------------------------------------- First, we give a definition of local semi-strong convexity. \[def:ssc\] A function $f(\w)$ is semi-strongly convex on the $\epsilon$-sublevel set $\S_\epsilon$ if there exists $\lambda>0$ such that $$\label{eqn:strong} \frac{\lambda}{2}\|\w - \w^* \|_2^2\leq f(\w) - f(\w^*),\quad\forall \w\in\S_\epsilon$$ where $\w^*$ is the closest point to $\w$ in the optimal set. We refer to the property (\[eqn:strong\]) as *local* semi-strong convexity when $\S_\epsilon\neq\Omega$. The two papers [@DBLP:journals/corr/GongY14; @DBLP:journals/corr/nesterov16linearnon] have explored the semi-strong convexity on the whole domain $\Omega$ to prove linear convergence of smooth optimization problems. In [@DBLP:journals/corr/nesterov16linearnon], the inequality (\[eqn:strong\]) is also called **second-order growth property**. They have also shown that a class of problems satisfy (\[eqn:strong\]) (see examples given below). The inequality indicates that $f(\cdot)$ admits a local error bound on $\S_\epsilon$ with $\theta=\frac{1}{2}$ and $c=\sqrt{\frac{2}{\lambda}}$, which leads to the following the corollary about the iteration complexity of RSG for locally semi-strongly convex problems. Suppose Assumption \[ass:rsg\] holds and $f(\w)$ is semi-strongly convex on $\S_\epsilon$. Then $B_\epsilon \leq \sqrt{\frac{2\epsilon}{\lambda}}$ [^5]and the iteration complexity of RSG for obtaining an $2\epsilon$-optimal solution is $O(\frac{2\alpha^2G^2}{\lambda \epsilon}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil)$ provided $t = \frac{2\alpha^2 G^2}{\lambda\epsilon}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$. [**Remark:**]{} To the best of our knowledge, the previous subgradient methods can achieve the $O(1/\epsilon)$ iteration complexity only by assuming strong convexity [@DBLP:journals/siamjo/GhadimiL12; @DBLP:journals/siamjo/GhadimiL13; @NIPS2012_4543; @hazan-20110-beyond]. Here, we obtain an $\widetilde O(1/\epsilon)$ iteration complexity ($\widetilde O(\cdot)$ suppresses constants and logarithmic terms) only with local semi-strong convexity. It is obvious that strong convexity implies local semi-strong convexity [@hazan-20110-beyond] but not vice versa. #### **Examples** Consider a family of functions in the form of $f(\w) = h(X\w) + r(\w)$, where $X\in\mathbb{R}^{n\times d}$, $h(\cdot)$ is smooth and *strongly convex* on any *compact set* and $r(\cdot)$ has a polyhedral epigraph. According to [@DBLP:journals/corr/GongY14; @DBLP:journals/corr/nesterov16linearnon], such a function $f(\w)$ satisfies (\[eqn:strong\]) for any $\epsilon\leq \epsilon_0$ with a constant value for $\lambda$. Although smoothness is assumed for $h(\cdot)$ in [@DBLP:journals/corr/GongY14; @DBLP:journals/corr/nesterov16linearnon], we find that it is not necessary for proving (\[eqn:strong\]). We state this result as the lemma below. Suppose Assumption \[ass:rsg\] holds, $\Omega=\{\w\in \mathbb{R}^{d}|C\w\leq \mathbf{b}\}$ with $C\in\mathbb{R}^{k\times d}$ and $\mathbf{b}\in\mathbb{R}^k$, and $f(\w) = h(X\w) + r(\w)$ where $h:\mathbb{R}^n\rightarrow\mathbb{R}$ satisfies $\text{dom}(h)=\mathbb{R}^k$ and is a strongly convex function on any compact set in $\mathbb{R}^n$, and $r(\w)$ has a polyhedral epigraph. Then, $f(w)$ satisfies (\[eqn:strong\]) for any $\epsilon\leq \epsilon_0$. The proof of this lemma is almost identical to the proof of Lemma 1 in [@DBLP:journals/corr/GongY14] which assumes $h(\cdot)$ is smooth. Here, we show that a similar result holds without the smoothness of $h(\cdot)$. Since $r(\cdot)$ has a polyhedral epigraph, we can assume $r(\w)=\q^T\w$ for some $q\in\mathbb{R}^d$ without loss of generality. In fact, when $r(\cdot)$ has a polyhedral epigraph and $\Omega=\mathbb{R}^d$, we can introduce a new variable $w\in\mathbb{R}$ so that can be equivalently represented as $$\begin{aligned} \label{eqn:prob0_poly_old} \min_{(\w,w)\in\mathbb{R}^{d+1}}& h(X\w)+w\\ \text{s.t. }&\w\in\Omega,r(\w)\leq w,\end{aligned}$$ where feasible set is a polyhedron in $\mathbb{R}^{d+1}$ and the corresponding $r(\cdot)$ becomes linear. Since $h(\cdot)$ is a strongly convex function on any compact set, following a standard argument (see [@DBLP:journals/corr/nesterov16linearnon] for example), we can show that there exist $\r^*\in\mathbb{R}^n$ and $s^*\in\mathbb{R}$ such that $$\Omega_*=\{\w\in\mathbb{R}^d | X\w=\r^*, \q^T\w= s^*,C\w\leq \mathbf{b} \},$$ which is a polyhedron. By Hoffman’s bound, there exists a constant $\zeta>0$ such that, for any $\w\in\Omega$, we have $$\begin{aligned} \label{eqn:hoffman} \|\w-\w^*\|^2\leq \zeta^2\left(\|X\w-\r^*\|^2+(\q^T\w-s^*)^2\right).\end{aligned}$$ where $\w^*$ is the closest point to $\w$ in $\Omega_*$. By the compactness of $\S_\epsilon$ and the strong convexity of $h(\cdot)$ on $\S_\epsilon$, there exists a constant $\mu>0$ such that $$\begin{aligned} \label{eqn:strongh} f(\w)-f(\w^*)&\geq&\mathbf{\xi}_{\w_*}^T(X\w-X\w^*)+\q^T(\w-\w^*)+\frac{\mu}{2}\|X\w-X\w^*\|^2\\\nonumber &\geq&\frac{\mu}{2}\|X\w-\r^*\|^2\end{aligned}$$ for some $\mathbf{\xi}_{\w^*}\in\partial h(\r^*)$ for any $\w\in\S_\epsilon$, where the second inequality is due to the optimality condition of $\w^*$. Note that $\mathbf{\xi}_{\w^*}$ in may change with $\w^*$ (and thus with $\w$). With the same $\mathbf{\xi}_{\w^*}$ and $\w$ as above, we can also show that $$\begin{aligned} \label{eqn:stronghr} (\q^T\w-s^*)^2&=&((X^T\mathbf{\xi}_{\w^*}+\q-X^T\mathbf{\xi}_{\w^*})^T(\w-\w^*))^2\\\nonumber &\leq&2((X^T\mathbf{\xi}_{\w^*}+\q)^T(\w-\w^*))^2+2(\mathbf{\xi}_{\w^*}^T(X\w-X\w^*))^2\\\nonumber &\leq&2(f(\w)-f^*)^2+2\|\partial h(\r^*)\|^2\|X\w-\r^*\|^2.\end{aligned}$$ Here, $\|\partial h(\r^*)\|^2<+\infty$ because $\text{dom}(h)=\mathbb{R}^n$. Applying and to , we obtain for any $\w\in\S_\epsilon$ that $$\begin{aligned} \|\w-\w^*\|^2\leq 2\zeta^2\left(\frac{1+2\|\partial h(\r^*)\|^2}{\mu}(f(\w)-f^*)+(f(\w)-f^*)^2\right),\end{aligned}$$ which further implies $$\begin{aligned} \frac{1}{2}\|\w-\w^*\|^2\leq \zeta^2\left(\frac{1+2\|\partial h(\r^*)\|^2}{\mu}+\epsilon\right)(f(\w)-f^*).\end{aligned}$$ using the fact that $f(\w)-f^*\leq\epsilon$ for $\w\in\S_\epsilon$. The function of this type covers some commonly used loss functions and regularization terms in machine learning and statistics. See the following examples. [**Robust regression**]{} $$\label{eqn:rr} \min_{\w\in\R^d}\frac{1}{n}\sum\limits_{i=1}^n|\x_i^{\top}\w - y_i|^p,$$ where $p\in(1,2)$, $\x_i\in\R^d$ denotes the feature vector and $y_i$ is the target output. The objective function is in the form of $h(X\w)$ where $X$ is a $n\times d$ matrix with $\x_1,\x_2,\dots,\x_n$ being its rows and $h(\u):=\sum_{i=1}^n|u_i-y_i|^p$. According to [@Goebel_localstrong], $h(\u)$ is a strongly convex function on any compact set so that the objective function above is semi-strongly convex on $\S_\epsilon$ for any $\epsilon\leq \epsilon_0$. Improved Convergence for Convex Problems with KL property --------------------------------------------------------- Lastly, we consider a family of non-smooth functions with a local Kurdyka-Łojasiewicz (KL) property. The definition of KL property is given below. \[def:KL\] The function $f(\w)$ has the Kurdyka - Łojasiewicz (KL) property at $\bar\w$ if there exist $\eta\in(0,\infty]$, a neighborhood $U_{\bar\w}$ of $\bar\w$ and a continuous concave function $\varphi:[0, \eta) \rightarrow \R_+$ such that (i) $\varphi(0) = 0$; (ii) $\varphi$ is continuous on $(0, \eta)$; (iii)for all $s\in(0, \eta)$, $\varphi'(s)>0$; (iv) and for all $\w\in U_{\bar\w}\cap \{\w: f(\bar\w)< f(\w)< f(\bar\w)+\eta\}$, the Kurdyka - Łojasiewicz (KL) inequality holds $$\begin{aligned} \label{eq:KLineq} \varphi'(f(\w) - f(\bar\w))\|\partial f(\w)\|_2\geq 1, \end{aligned}$$ where $\|\partial f(\w)\|_2:=\min_{\mathbf g\in\partial f(\w)} \|\mathbf g\|_2$. The function $\varphi$ is called the **desingularizing function** of $f$ at $\bar\w$, which sharpens the function $f(\w)$ by reparameterization. An important desingularizing function is in the form of $\varphi(s) = cs^{1-\beta}$ for some $c>0$ and $\beta\in[0,1)$, by which, gives the KL inequality $$\|\partial f(\w)\|_2\geq \frac{1}{c(1-\beta)}(f(\w) - f(\bar\w))^{\beta}.$$ Note that all semi-algebraic functions satisfy the KL property at any point [@Bolte:2014:PAL:2650160.2650169]. Indeed, all the concrete examples given before satisfy the Kurdyka - Łojasiewicz property. For more discussions about the KL property, we refer readers to [@Bolte:2014:PAL:2650160.2650169; @journals/siamjo/BolteDLS07; @DBLP:journals/corr/SchneiderU14; @journals/mp/AttouchBS13; @Bolte:2006:LIN:1328019.1328299]. The following corollary states the iteration complexity of RSG for unconstrained problems that have the KL property at each $\bar\w\in\Omega_*$ . \[thm:KL\] Suppose Assumption \[ass:rsg\] holds, $f(\w)$ satisfies a (uniform) Kurdyka - Łojasiewicz property at any $\bar\w\in\Omega_*$ with the same desingularizing function $\varphi$ and constant $\eta$, and $$\label{eqn:smallS} \S_\epsilon\subset\cup_{\bar\w\in\Omega_*}\left[U_{\bar\w}\cap \{\w: f(\bar\w)< f(\w)< f(\bar\w)+\eta\}\right].$$ The iteration complexity of RSG for obtaining an $2\epsilon$-optimal solution is given by $O\left(\alpha^2G^2(\varphi(\epsilon)/\epsilon)^2\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$ provided $\alpha^2G^2(\varphi(\epsilon)/\epsilon)^2\leq t=O\left(\alpha^2G^2(\varphi(\epsilon)/\epsilon)^2\right)$. In addition, if $\varphi(s) = cs^{1-\beta}$ for some $c>0$ and $\beta\in[0,1)$, the iteration complexity of RSG is $O( \frac{\alpha^2 G^2c^2(1-\beta)^2}{\epsilon^{2\beta}}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil)$ provided $t = \frac{\alpha^2G^2c^2}{\epsilon^{2\beta}}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$. We can prove the above corollary following a result in [@arxiv:1510.08234] as presented in Proposition \[prop:KL\] in the appendix. According to Proposition \[prop:KL\], if $f(\cdot)$ satisfies the KL property at $\bar\w$, then for all $\w\in U_{\bar \w}\cap \{\w: f(\bar\w)< f(\w)< f(\bar \w)+ \eta\}$ it holds that $ \|\w - \w^*\|_2\leq \varphi(f(\w) - f(\bar \w))$. It then, under the uniform condition in (\[eqn:smallS\]), implies that, for any $\w\in\S_\epsilon$ $$\|\w - \w^*\|_2 \leq \varphi(f(\w) - f_*)\leq \varphi(\epsilon),$$ where we use the monotonic property of $\varphi$. Then the first conclusion follows similarly as Corollary \[lem:3\] by noting $B_\epsilon \leq \varphi(\epsilon)$. The second conclusion immediately follows by setting $\varphi(s) = cs^{1-\beta}$ in the first conclusion. While the conclusion in Corollary \[thm:KL\] hinges on a condition in (\[eqn:smallS\]), in practice many convex functions satisfy the KL property with $U =\R^d$ and $\eta=\infty$ [@Attouch:2010:PAM:1836121.1836131]. It is worth mentioning that to our best knowledge, the present work is the first to leverage the KL property for developing improved subgradient methods, though it has been explored in non-convex and convex optimization for deterministic descent methods for smooth optimization [@arxiv:1510.08234; @Bolte:2014:PAL:2650160.2650169; @Attouch:2010:PAM:1836121.1836131; @Karimi15]. A variant of RSG with $p$-norm and Dual Averaging Method {#sec:variant} ======================================================== To demonstrate the flexibility of our restarting strategy, in this section, we present a variant of RSG that utilizes the $p$-norm with $p\in(1,2]$ and a Nesterov’s dual averaging (DA) algorithm as its subroutine. Let $\|\w\|_p=(\sum_{i=1}^d|w_i|^p)^{1/p}$ denote the $p$-norm. It is known that $\frac{1}{2}\|\w\|_p^2$ is $(p-1)$-strongly convex. Let $q$ and $p$ be conjugate constants, namely, $1/p+1/q=1$. Using the $p$-norm, we redefine some of the previous notations: $$\begin{aligned} \w^\dagger_\epsilon &:= \arg\min_{\u\in\S_\epsilon}\|\u- \w\|_p^2=\arg\min_{\u\in\Omega}\|\u- \w\|_p^2, \quad \text{s.t.}\quad f(\u)\leq f_*+\epsilon,\\ \w^* &:= \arg\min_{\u\in\Omega_*}\|\u - \w\|_p^2,\\ B_\epsilon&:=\max_{\w\in\V_\epsilon}\min_{\u\in\Omega_*}\|\w - \u\|_p = \max_{\w\in\V_\epsilon}\|\w - \w^*\|_p,\end{aligned}$$ $$\begin{aligned} \|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_q&:=\min_{\mathbf g\in\partial f(\w),\mathbf v\in\mathcal{N}_{\Omega}(\w)} \|\mathbf g+\mathbf v\|_q\\ \rho_\epsilon&:=\min_{\w\in\V_\epsilon} \|\partial f(\w)+\mathcal{N}_{\Omega}(\w)\|_q.\end{aligned}$$ We first generalize Lemma \[lem:1\] with the $p$-norm. \[lem:p-norm\] For any $\epsilon>0$ such that $\V_\epsilon\neq\emptyset$ and any $\w\in\Omega$, we have $$\|\w - \w_\epsilon^\dagger\|_p\leq \frac{1}{\rho_\epsilon}(f(\w) - f(\w_\epsilon^\dagger)).$$ Let $[\w]_i$ represent the $i$th coordinate of $\w$. Note that $$\frac{\partial \|\w\|_p}{\partial w_i} = \frac{|w_i|^{p-1}\text{sign}(w_i)}{\|\w\|_p^{p-1}}, \quad \left[\nabla\|\w\|_p^2\right]_i = 2\|\w\|_p^{2-p}|w_i|^{p-1}\text{sign}(w_i).$$ Since the conclusion holds trivially if $\w\in\S_\epsilon$ (so that $\w^\dagger_\epsilon=\w$), we assume $\w\in\Omega\backslash\S_\epsilon$. According to the definition of $\w^\dagger_\epsilon$ and the associated first-order optimality conditions, there exist a scalar $\zeta\geq 0$, a subgradient $\mathbf{g}\in \partial f(\w^\dagger_\epsilon)$ and a vector $\mathbf v\in\mathcal{N}_{\Omega}(\w^\dagger_\epsilon)$ such that $$\label{eqn:o2_pnorm} \|\w^\dagger_\epsilon - \w\|^{2-p}_p|[\w^\dagger_{\epsilon}-\w]_i|^{p-1}\text{sign}([\w^\dagger_{\epsilon}-\w]_i)+ \zeta [\mathbf g]_i+[\mathbf v]_i= 0, \quad i=1,2,\dots,d.$$ By the convexity of $f(\cdot)$ and the definition of normal cone, we have $$\label{eq:gqnorm1} \begin{split} \zeta(f(\w) - f(\w^\dagger_\epsilon))&\geq (\w - \w^\dagger_\epsilon)^{\top}(\zeta\mathbf g +\mathbf v)\\ &= \|\w^\dagger_\epsilon - \w\|^{2-p}_p\sum_{i=1}^d|[\w^\dagger_{\epsilon}-\w]_i|^{p-1}\text{sign}([\w^\dagger_{\epsilon}-\w]_i)[\w^\dagger_{\epsilon}-\w]_i\\ &=\|\w^\dagger_\epsilon - \w\|^{2-p}_p\|\w^\dagger_\epsilon- \w\|_p^p = \|\w^\dagger_\epsilon - \w\|^{2}_p. \end{split}$$ It is clear that $\zeta>0$, since, otherwise, $\w=\w^\dagger_\epsilon$ which contradicts with $\w\not\in\S_\epsilon$. It is from that $$\label{eq:gqnorm2} \begin{split} &\zeta^q\sum_{i}|[\mathbf g+\mathbf v/\zeta]_i|^q = \|\w^\dagger_\epsilon - \w\|_p^{q(2-p)}\sum_{i=1}^d|[\w^\dagger_{\epsilon}-\w]_i|^{q(p-1)} \\ &= \|\w^\dagger_\epsilon - \w\|_p^{q(2-p)}\sum_{i=1}^d|[\w^\dagger_{\epsilon}-\w]_i|^{p} = \|\w^\dagger_\epsilon - \w\|_p^{q(2-p)+p}= \|\w^\dagger_\epsilon - \w\|_p^{p/(p-1)}. \end{split}$$ Organizing terms in the equality above gives $$\frac{1}{\zeta}= \frac{\|\mathbf g+\mathbf v/\zeta\|_q}{\|\w^\dagger_\epsilon - \w\|_p^{p/(q(p-1))}} \geq \frac{\rho_\epsilon}{\|\w^\dagger_\epsilon - \w\|_p},$$ which further implies$$f(\w) - f(\w^\dagger_\epsilon)\geq \rho_\epsilon\|\w^\dagger_\epsilon - \w\|_p.$$ Similar to Lemma \[lem:lowf\], we can also lower bound $\rho_\epsilon$ by $\epsilon/B_\epsilon$. In particular, we have the following lemma. \[lem:generalp\] For any $\epsilon>0$ such that $\V_\epsilon\neq\emptyset$ and any $\w\in\Omega$, we have $$\|\w - \w_\epsilon^\dagger\|_p\leq \frac{\|\w^\dagger_\epsilon-\w^*_{\epsilon}\|_p}{\epsilon}(f(\w) - f(\w_\epsilon^\dagger))\leq\frac{B_\epsilon}{\epsilon}(f(\w) - f(\w_\epsilon^\dagger)) .$$ where $\w^*_\epsilon$ is the closest point in $\Omega_*$ to $\w^\dagger_\epsilon$ measured in the $p$-norm. Since the conclusion holds trivially if $\w\in\S_\epsilon$ (so that $\w^\dagger_\epsilon=\w$), we assume $\w\in\Omega\backslash\S_\epsilon$ so that $\w_\epsilon^\dagger\in\mathcal L_\epsilon$. From in the proof of Lemma \[lem:p-norm\], we have $\|\mathbf g+\mathbf v/\zeta\|_q=\|\w^\dagger_\epsilon - \w\|_p/\zeta$, where $\zeta>0$, $\g\in\partial f(\w_\epsilon^\dagger)$ and $\v\in\mathcal N_\Omega(\w_\epsilon^\dagger)$. Applying this relationship to in the proof of Lemma \[lem:p-norm\], we have $$f(\w) - f(\w_\epsilon^\dagger)\geq \|\mathbf g+\mathbf v/\zeta\|_q\|\w -\w^\dagger_\epsilon \|_p.$$ By the convexity of $f(\w)$ and the definition of normal cone, we have $$f(\w_\epsilon^*) - f(\w^\dagger_\epsilon)\geq (\w_\epsilon^* - \w^\dagger_\epsilon)^{\top}\mathbf g \geq(\w_\epsilon^* - \w^\dagger_\epsilon)^{\top}\left(\mathbf g+\mathbf v/\zeta\right),$$ Then, $$\|\g + \v/\zeta\|_q \|\w^*_\epsilon - \w^\dagger_\epsilon\|_p\geq f(\w_\epsilon^\dagger) - f(\w_\epsilon^*) = \epsilon$$ As a result, $$f(\w) - f(\w_\epsilon^\dagger)\geq\frac{\epsilon}{ \|\w^*_\epsilon - \w^\dagger_\epsilon\|_p}\|\w - \w_\epsilon^\dagger\|_p$$ which completes the proof. **Input**: a step size $\eta$, the number of iterations $T$, and the initial solution $\w_1$, $\gh_1=0$, $\Lambda_0 = 0$ Query the gradient oracle to obtain $\G(\w_t)$ Set $\lambda_t=1$ or $\lambda_t = 1/\|\G(\w_t)\|_q$, and $\Lambda_t=\Lambda_{t-1}+\lambda_t$ $\gh_{t+1}=\gh_t + \lambda_t \G(\w_t)$ update $\w_{t+1} =\arg\min_{\w}\eta\gh_{t+1}^{\top}\w + \frac{1}{2}\|\w - \w_1\|_p^2$ **Output**: $\wh_T =\frac{ \sum_{t=1}^T\lambda_t\w_t}{\Lambda_T}$ **Input**: the number of iterations $t$ per-epoch, $\w_0 \in \R^d$, $\alpha>1$. Set $\eta_1=\epsilon_0(p-1)/(\alpha G)$ or $\eta_1 = \epsilon_0(p-1)/(\alpha G^2)$ corresponding to the two choices of $\lambda_t$ in DA$_p$, where $\epsilon_0$ is from Assumption \[ass:sdap\].a Call subroutine to obtain $\w_k = \text{DA}_p(\w_{k-1}, \eta_k, t)$ Set $\eta_{k+1}=\eta_k/\alpha$ **Output**: $\w_K$ In Algorithm \[alg:11\], we present a modified version of RSG that uses Algorithm \[alg:sdap\] as a subroutine. Here, Algorithm \[alg:sdap\] is a variant of Nesterov’s dual averaging (DA) algorithm. We denote Algorithm \[alg:11\] by RSG-DA$_p$ to distinguish it from the RSG method in Algorithm \[alg:1\]. Before presenting the convergence property of RSG-DA$_p$, we first need to address two issues: (i) how to compute $\w_{t+1}$ in Line 7 of DA$_p$ and (ii) what the convergence rate of DA$_p$ is. The proposition below shows that $\w_{t+1}$ can be computed in a closed form [when $\Omega = \R^d$]{}. The solution to $\w^+ = \arg\min_{\u}\g^{\top}\u + \frac{1}{2}\|\u-\w\|_p^2$ is given by $$[\w^+]_i= [\w]_i - \|\g\|_q^{(p-q)/p}\text{sign}([\g]_i)|[\g]_i|^{q-1},\quad i=1,2,\dots,d.$$ By the optimality condition, $\w^+$ satisfies $$\label{eqn:point} \|\w^+ - \w\|^{2-p}_p|[\w^+-\w]_i|^{p-1}\text{sign}([\w^+-\w]_i)+ [\g]_i= 0, \quad i=1,2,\dots,d.$$ It is a simple exercise to verify that $\|\g\|_q = \|\w^+-\w\|_p$. From (\[eqn:point\]), we have $$(\w - \w^+)^{\top}\g = \|\w^+ - \w\|_p^2 = \|\w^+ - \w\|_p\|\g\|_q.$$ This means the Holder’s inequality, $(\w - \w^+)^{\top}\g \leq \|\w^+ - \w\|_p\|\g\|_q$, holds as an equality and this can happens only if $|[\w^+ - \w]_i| = c|[\g]_i|^{q-1}$ for a constant $c$. From (\[eqn:point\]), we can also see that $[\w^+ - \w]_i$ has the same sign as $-[\g]_i$ so that $$[\w^+]_i - [\w]_i = - c\text{sign}([\g]_i)|[\g]_i|^{q-1}.$$ It remains to derive the value of $c$. To do so, we observe that $$\sum_{i=1}|[\w^+-\w]_i|^p = c^p\sum_{i=1}^d|[\g]_i|^{p(q-1)} = c^p\sum_{i=1}^d|[\g]_i|^{q},$$ which indicates $\|\w^+ - \w\|_p^p = c^p\|\g\|_q^q$. Since $\|\w^+ - \w\|_p =\|\g\|_q$, we have $$c =\left( \frac{\|\g\|_q^p}{\|\g\|_q^q}\right)^{1/p} = \|\g\|_q^{(p-q)/p}.$$ The following Lemma by [@Nesterov:2009:PSM:1530733.1530741] characterizes the convergence property of DA$_p$. \[lem:sgdp\][@Nesterov:2009:PSM:1530733.1530741] Suppose $\|\G(\w_t)\|_q\leq G$. Let Algorithm \[alg:sdap\] run for $T$ iterations. Then, for any $\w\in\Omega$, we have $$f(\wh_T) - f(\w) \leq \frac{\|\w - \w_1\|_p^2}{2\eta\Lambda_T} + \frac{\eta\sum_{t=1}^T\lambda^2_t\|\G(\w_t)\|_q^2}{2(p-1)\Lambda_T}.$$ If $\lambda_t=1$, we have $$f(\wh_T) - f(\w) \leq \frac{\|\w - \w_1\|_p^2}{2\eta T} + \frac{\eta G^2}{2(p-1)}.$$ If $\lambda_t = 1/\|\G(\w_t)\|_q$, we have $$f(\wh_T) - f(\w)\leq \frac{G\|\w - \w_1\|_p^2}{2\eta T} + \frac{\eta G}{2(p-1)}.$$ We omit the proof because it can be found in [@Nesterov:2009:PSM:1530733.1530741]. To state the convergence of RSG-DA$_p$. We need the following assumptions which is the same as Assumption \[ass:rsg\] except that the Euclidean norm is replaced by the $q$-norm in item **b**: \[ass:sdap\] For a convex minimization problem (1), we assume - There exist $\w_0\in\Omega$ and $\epsilon_0\geq 0$ such that $f(\w_0) - \min_{\w\in\Omega}f(\w)\leq \epsilon_0$. - There exists a constant $G$ such that $\|\G(\w)\|_q\leq G$ for any $\w\in\Omega$. - $\Omega_*$ is a non-empty convex compact set. The following theorem characterizes the iteration complexity of RSG-DA$_p$ whose proof is similar to that of Theorem \[thm:GDr\] and is thus omitted. \[thm:SDApr\] Suppose [Assumption \[ass:sdap\]]{} hods. If $t\geq \frac{\alpha^2 G^2B_\epsilon^2}{\epsilon^2}$ and $K = \lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil$ in Algorithm \[alg:11\], with at most $K$ stages, Algorithm \[alg:11\] returns a solution $\w_K$ such that $f(\w_K)-f_*\leq 2\epsilon$. In other word, the total number of iterations for Algorithm \[alg:11\] to find an $2\epsilon$-optimal solution is at most $T= O\left(t\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$ where $t\geq \frac{\alpha^2 G^2B_\epsilon^2}{\epsilon^2}$. In particular, if $\frac{\alpha^2 G^2B_\epsilon^2}{\epsilon^2}\leq t=O\left( \frac{\alpha^2 G^2B_\epsilon^2}{\epsilon^2}\right)$, the total number of iterations for Algorithm \[alg:11\] to find an $2\epsilon$-optimal solution is at most $T= O\left(\frac{\alpha^2G^2B_\epsilon^2}{\epsilon^2}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$. We want to emphasis that, although the conclusion of Theorem \[thm:SDApr\] is similar to Corollary \[lem:3\], $B_\epsilon$ here is defined using $p$-norm and $G$ here is defined using the $q$-norm. Depending on the problem, Algorithm \[alg:11\] may have a lower complexity than Algorithm \[alg:1\]. [For an example, we can consider empirical loss minimization in machine learning for finding a sparse model with high-dimensional data. Suppose the loss function $\ell(z,y)$ is $1$-Lipschitz continuous, if $f(\w)$ is the average of $\ell(\w^{\top}\x_i, y_i)$ over training examples $(\x_i, y_i), i=1,\ldots, n$, where $\x_i\in\R^d$ ($d\gg 1$), then $G = \max_{i}\|\x_i\|_q$. Let $p=\frac{2\ln d}{2\ln d -1}$ and $q = 2\ln d$ [^6]. Then, using the $p$-norm to define $B_\epsilon$ and $q$-norm to define $G$, we have $B_\epsilon G =\max_{\w\in\mathcal L_\epsilon}\|\w - \w^*\|_p\max_{i}\|\x_i\|_q \approx \max_{\w\in\mathcal L_\epsilon}\|\w - \w^*\|_1\max_{\x_i}\|\x\|_\infty $. In contrast, if we use the Euclidean norm definitions, we have $B_\epsilon G = \max_{\w\in\mathcal L_\epsilon}\|\w - \w^*\|_2\max_{i}\|\x\|_2 $. If we assume $\|\x_i\|_\infty\leq 1$ and $\w\in\S_\epsilon$ is approximately $s$-sparse with $s\ll d$ such that $\frac{\|\w - \w^*\|_1}{\|\w - \w^*\|_2}\approx \sqrt{s}$ [@DBLP:conf/icml/Yang0JZ15a; @DBLP:journals/corr/abs-1109-4299]. In light of $\frac{\|\x_i\|_\infty}{\|\x_i\|_2}\approx \frac{1}{\sqrt{d}}$, the magnitude of $B_\epsilon G$ using the special $p$-norm and $q$-norm definitions is much smaller than $B_\epsilon G$ using the Euclidean norm definitions, making RSG-DA$_p$ converge faster than RSG. ]{} Variants of RSG without knowing the constant $c$ and the exponent $\theta$ in the local error bound {#sec:prac} =================================================================================================== In Section \[sec:spec\], we have discussed the local error bound and presented several classes of problems to reveal the magnitude of $B_\epsilon$, i.e., $B_\epsilon = c\epsilon^\theta$. For some problems, the value of $\theta$ is exhibited. However, the value of the constant $c$ could be still difficult to estimate, which renders it challenging to set the appropriate value $t= \frac{\alpha^2c^2 G^2}{\epsilon^{2(1-\theta)}}$ for inner iterations of RSG. In practice, one might use a sufficiently large $c$ to set up the value of $t$. However, such an approach is vulnerable to both over-estimation and under-estimation of $t$. Over-estimating the value of $t$ leads to a waste of iterations while under-estimation leads to an less accurate solution that might not reach to the target accuracy level. In addition, for other problems the value of $\theta$ is still an open problem. One interesting family of objective functions in machine learning is the sum of piecewise linear loss over training data and overlapped or non-overlapped group lasso. In this section, we present variants of RSG that can be implemented without knowing the value of $c$ in the local error bound condition and even the value of exponent $\theta$, and prove their improved convergence. RSG without knowing $c$ ----------------------- The key idea is to use an increasing sequence of $t$ and another level of restarting for RSG. The detailed steps are presented in Algorithm \[alg:2\], to which we refer as R$^2$SG. With mild conditions on $t_1$ in R$^2$SG, the complexity of R$^2$SG for finding an $\epsilon$ solution is given by the theorem below. \[thm:R2SG\] Suppose $\epsilon\leq \epsilon_0/4$ and $K =\lceil \log_\alpha(\epsilon_0/\epsilon)\rceil$. Let $t_1$ in Algorithm \[alg:2\] be large enough so that there exists $\hat\epsilon_1$ such that $\epsilon\leq \hat\epsilon_1\leq \epsilon_0/2$ and $t_1 = \frac{\alpha^2c^2G^2}{\hat\epsilon_1^{2(1-\theta)}}$. In addition, suppose $f(\cdot)$ admits an error bound condition with $\theta\in(0,1)$ and $c$ on $\S_{\hat\epsilon_1}$. Then, with at most $S=\lceil\log_2(\hat\epsilon_1/\epsilon)\rceil +1$ calls of RSG in Algorithm \[alg:2\], we find a solution $\w^S$ such that $f(\w^S) - f_*\leq 2\epsilon$. The total number of iterations of R$^2$SG for obtaining $2\epsilon$-optimal solution is upper bounded by $T_S = O\left(\frac{c^2G^2}{\epsilon^{2(1-\theta)}}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$. Since $K=\lceil\log_\alpha(\epsilon_0/\epsilon)\rceil\geq \lceil\log_\alpha(\epsilon_0/\hat\epsilon_1)\rceil$ and $t_1 = \frac{\alpha^2c^2G^2}{\hat\epsilon_1^{2(1-\theta)}}$, we can apply Corollary \[lem:leb\_corr\] with $\epsilon=\hat\epsilon_1$ to the first call of RSG in Algorithm \[alg:2\] so that the output $\w^1$ satisfies $$\label{eqn:RSG-s1} f(\w^1) - f_* \leq 2\hat\epsilon_1.$$ Then, we consider the second call of RSG with the initial solution $\w^1$ satisfying (\[eqn:RSG-s1\]). By the setup $K = \lceil\log_\alpha(\epsilon_0/\epsilon)\rceil\geq \lceil\log_\alpha(2\hat\epsilon_1/(\hat\epsilon_1/2))\rceil$ and $t_2 =t_1 2^{2(1-\theta)}= \frac{c^2G^2}{(\hat\epsilon_1/2)^{2(1-\theta)}}$, we can apply Corollary \[lem:leb\_corr\] with $\epsilon=\hat\epsilon_1/2$ and $\epsilon_0=2\hat\epsilon_1$ so that the output $\w^2$ of the second call satisfies $ f(\w^2) - f_* \leq \hat\epsilon_1$. By repeating this argument for all the subsequent calls of RSG, with at most $S = \lceil \log_2(\hat\epsilon_1/\epsilon)\rceil + 1$ calls, Algorithm \[alg:2\] ensures that $$f(\w^S) - f_* \leq 2\hat\epsilon_1/2^{S-1} \leq 2\epsilon$$ The total number of iterations during the $S$ calls of RSG is bounded by $$\begin{aligned} T_S &= K\sum_{s=1}^St_s =K \sum_{s=1}^St_12^{2(s-1)(1-\theta)} = Kt_1 2^{2(S-1)(1-\theta)}\sum_{s=1}^S \left(\frac{1}{2^{2(1-\theta)}}\right)^{S-s}\\ &\leq \frac{Kt_12^{2(S-1)(1-\theta)}}{1 - 1/2^{2(1-\theta)}} \leq O\left(Kt_1 \left(\frac{\hat\epsilon_1}{\epsilon}\right)^{2(1-\theta)}\right) = O\left(\frac{c^2G^2}{\epsilon^{2(1-\theta)}}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right). \end{aligned}$$ **Input**: the number of iterations $t_1$ in each stage of the first call of RSG and the number of stages $K$ in each call of RSG **Initialization:** $\w^0 \in \Omega$; Let $\w^s= \text{RSG}(\w^{s-1}, K, t_s, \alpha)$ Let $t_{s+1} = t_s 2^{2(1-\theta)}$ **Remark:** We make several remarks about Algorithm \[alg:2\] and Theorem \[thm:R2SG\]: (i) Theorem \[thm:R2SG\] applies only when $\theta\in(0,1)$. If $\theta=1$, in order to have an increasing sequence of $t_s$, we can set $\theta$ in Algorithm \[alg:2\] to a little smaller value than $1$ in practical implementation. (ii) the $\epsilon_0$ in the implementation of RSG (Algorthm \[alg:1\]) can be re-calibrated for $s\geq 2$ to improve the performance (e.g., one can use the relationship $f(\w_{s-1}) - f_* = f(\w_{s-2}) - f_* + f(\w_{s-1}) - f(\w_{s-2})$ to do re-calibration); (iii) as a tradeoff, the exiting criterion of R$^2$SG is not as automatic as RSG. In fact, the total number of calls $S$ of RSG for obtaining an $2\epsilon$-optimal solution depends on an unknown parameter (namely $\hat\epsilon_1$). In practice, one could use other stopping criteria to terminate the algorithm. For example, in machine learning applications one can monitor the performance on the validation data set to terminate the algorithm. (vi) The quantities $\hat\epsilon_1$, $S$ in the proof above are implicitly determined by $t_1$ and one do not need to compute $\hat\epsilon_1$ and $S$ in order to apply Algorithm \[alg:2\]. RSG for unknown $\theta$ and $c$ -------------------------------- Without knowing $\theta\in(0,1]$ and $c$ to get a sharper local error bound, we can simply let $\theta=0$ and $c = B_{\epsilon'}$ with $\epsilon'\geq \epsilon$, which still render the local error bound condition hold (c.f. Definition \[def:ler\]). Then we can employ the doubling trick to increase the values of $t$. In particular, we start with a sufficiently large value of $t$ and run RSG with $K=\lceil \log_\alpha(\epsilon_0/\epsilon)\rceil$ stages, and then double the value of $t$ and repeat the process. \[thm:2RSG\] Suppose $\epsilon\leq \epsilon_0/4$ and $K =\lceil \log_\alpha(\epsilon_0/\epsilon)\rceil$. Let $\theta=0$ and $t_1$ in Algorithm \[alg:2\] be large enough so that there exists $\hat\epsilon_1$ such that $\epsilon\leq \hat\epsilon_1\leq \epsilon_0/2$ and $t_1 = \frac{\alpha^2B_{\hat\epsilon_1}^2G^2}{\hat\epsilon_1^{2}}$. Then, with at most $S=\lceil\log_2(\hat\epsilon_1/\epsilon)\rceil +1$ calls of RSG in Algorithm \[alg:2\], we find a solution $\w^S$ such that $f(\w^S) - f_*\leq 2\epsilon$. The total number of iterations of R$^2$SG for obtaining $2\epsilon$-optimal solution is upper bounded by $T_S = O\left(\frac{B_{\hat\epsilon_1}^2G^2}{\epsilon^{2}}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right)$. **Remark:** Compared to the vanilla SG, the above iteration complexity is still an improved one with a smaller factor $B_{\hat\epsilon_1}^2$ compared to $\|\w_0 - \w_0^*\|_2^2$ in the iteration complexity of SG. The proof is similar to that of Theorem \[thm:R2SG\] except that we let $c = B_{\hat\epsilon_1}$ and $\theta=0$. Since $K=\lceil\log_\alpha(\epsilon_0/\epsilon)\rceil\geq \lceil\log_\alpha(\epsilon_0/\hat\epsilon_1)\rceil$ and $t_1 = \frac{\alpha^2B_{\hat\epsilon_1}^2G^2}{\hat\epsilon_1^{2}}$, we can apply Corollary \[lem:leb\_corr\] with $\epsilon=\hat\epsilon_1$ to the first call of RSG in Algorithm \[alg:2\] so that the output $\w^1$ satisfies $$\label{eqn:RSG-s2} f(\w^1) - f_* \leq 2\hat\epsilon_1.$$ Then, we consider the second call of RSG with the initial solution $\w^1$ satisfying (\[eqn:RSG-s2\]). By the setup $K = \lceil\log_\alpha(\epsilon_0/\epsilon)\rceil\geq \lceil\log_\alpha(2\hat\epsilon_1/(\hat\epsilon_1/2))\rceil$ and $t_2 =t_1 2^{2}= \frac{B_{\hat\epsilon_1}^2G^2}{(\hat\epsilon_1/2)^{2}}$, we can apply Corollary \[lem:leb\_corr\] with $\epsilon=\hat\epsilon_1/2$ and $\epsilon_0=2\hat\epsilon_1$ so that the output $\w^2$ of the second call satisfies $ f(\w^2) - f_* \leq \hat\epsilon_1$. By repeating this argument for all the subsequent calls of RSG, with at most $S = \lceil \log_2(\hat\epsilon_1/\epsilon)\rceil + 1$ calls, Algorithm \[alg:2\] ensures that $$f(\w^S) - f_* \leq 2\hat\epsilon_1/2^{S-1} \leq 2\epsilon.$$ The total number of iterations during the $S$ calls of RSG is bounded by $$\begin{aligned} T_S &= K\sum_{s=1}^St_s =K \sum_{s=1}^St_12^{2(s-1)} = Kt_1 2^{2(S-1)}\sum_{s=1}^S \left(\frac{1}{2^{2}}\right)^{S-s}\\ &\leq \frac{Kt_12^{2(S-1)}}{1 - 1/2^{2}} \leq O\left(Kt_1 \left(\frac{\hat\epsilon_1}{\epsilon}\right)^{2}\right) = O\left(\frac{B_{\hat\epsilon_1}^2G^2}{\epsilon^{2}}\lceil \log_\alpha(\frac{\epsilon_0}{\epsilon})\rceil\right). \end{aligned}$$ Experiments {#sec:exp} =========== In this section, we present some experiments to demonstrate the effectiveness of RSG. We focus on applications in machine learning, in particular regression and classification. #### Robust Regression The regression problem is to predict an output $y$ based on a feature vector $\x\in\R^d$. Given a set of training examples $(\x_i,y_i), i=1,\ldots, n$, a linear regression model can be found by solving the optimization problem in (\[eqn:rr\]). We solve two instances of the problem with $p=1$ and $p=1.5$. We conduct experiments on two data sets from libsvm website [^7], namely housing ($n=506$ and $d=13$) and space-ga ($n=3107$ and $d=6$). We first examine the convergence behavior of RSG with different values for the number of iterations per-stage $t=10^2$, $10^3$, and $10^4$. The value of $\alpha$ is set to $2$ in all experiments. The initial step size of RSG is set to be proportional to $\epsilon_0/2$ with the same scaling parameter for different variants. We plot the results on housing data in Figure \[fig:4\] (a,b) and on space-ga data in Figure \[fig:3\] (a,b). In each figure, we plot the objective value vs number of stages and the log difference between the objective value and the converged value (to which we refer as level gap). We can clearly see that with different $t$ RSG converges to an $\epsilon$-level set and the convergence rate is linear in terms of the number of stages, which is consistent with our theory. Secondly, we compare with SG to verify the effectiveness of RSG. The baseline SG is implemented with a decreasing step size proportional to $1/\sqrt{\tau}$, where $\tau$ is the iteration index. The initial step size of SG is tuned in a wide range to give the fastest convergence. The initial step size of RSG is also tuned around the best initial step size of SG. The results are shown in Figure \[fig:4\](c,d) and Figure \[fig:3\](c,d), where we show RSG with two different values of $t$ and also R$^2$SG with an increasing sequence of $t$. In implementing R$^2$SG, we restart RSG for every $5$ stages, and increase the number of iterations by a certain factor. In particular,we increase $t$ by a factor of $1.15$ and $1.5$ respectively for $p=1$ and $p=1.5$. From the results, we can see that (i) RSG with a smaller value of $t=10^3$ can quickly converge to an $\epsilon$-level, which is less accurate than SG after running a sufficiently large number of iterations; (ii) RSG with a relatively large value $t=10^4$ can converge to a much more accurate solution; (iv) R$^2$SG converges much faster than SG and can bridge the gap between RSG-$t=10^3$ and RSG-$t=10^4$. #### SVM Classification with a graph-guided fused lasso The classification problem is to predict a binary class label $y\in\{1,-1\}$ based on a feature vector $\x\in\R^d$. Given a set of training examples $(\x_i,y_i), i=1,\ldots, n$, the problem of training a linear classification model can be cast into $$\min_{\w\in\R^d}F(\w):=\frac{1}{n}\sum_{i=1}^n\ell(\w^{\top}\x_i, y_i) + R(\w).$$ Here we consider the hinge loss $\ell(z, y) = \max(0, 1- yz)$ as in support vector machine (SVM) and a graph-guided fused lasso (GFlasso) regularizer $R(\w) = \lambda \|F\w\|_1$ [@DBLP:journals/bioinformatics/KimSX09], where $F=[F_{ij}]_{m\times d}\in\R^{m\times d}$ encodes the edge information between variables. Suppose there is a graph $\mathcal G=\{\mathcal V, \mathcal E\}$ where nodes $\mathcal V$ are the attributes and each edge is assigned a weight $s_{ij}$ that represents some kind of similarity between attribute $i$ and attribute $j$. Let $\mathcal E=\{e_1,\ldots, e_m\}$ denote a set of $m$ edges, where an edge $e_\tau=(i_\tau, j_\tau)$ consists of a tuple of two attributes. Then the $\tau$-th row of $F$ matrix can be formed by setting $F_{\tau, i_\tau}=s_{i_\tau, j_\tau}$ and $F_{\tau, j_\tau} = - s_{i_\tau, j_\tau}$ for $(i_\tau, j_\tau)\in\mathcal E$, and zeros for other entries. Then the GFlasso becomes $R(\w) = \lambda \sum_{(i,j)\in\mathcal E}s_{ij}|w_i - w_j|$. Previous studies have found that a carefully designed GFlasso regularization helps in reducing the risk of over-fitting. In this experiment, we follow [@DBLP:conf/icml/OuyangHTG13] to generate a dependency graph by sparse inverse covariance selection [@citeulike:2134265]. To this end, we first generate a sparse inverse covariance matrix using the method in [@citeulike:2134265] and then assign an equal weight $s_{ij}=1$ to all edges that have non-zero entries in the resulting inverse covariance matrix. We conduct the experiment on the dna data ($n=2000$ and $d=180$) from the libsvm website, which has three class labels. We solve the above problem to classify class 3 versus the rest. Besides SG, we also compare with another baseline, namely alternating direction method of multiplers (ADMM). A stochastic variant of ADMM has been employed to solve the above problem in [@DBLP:conf/icml/OuyangHTG13] by writing the problem as $$\min_{\w\in\R^d} \frac{1}{n}\sum_{i=1}^n\ell(\w^{\top}\x_i, y_i) + \lambda \|\u\|_1, \quad s.t. \quad \u = F\w$$ For fairness, we compare with deterministic ADMM. Since the intermediate problems associated with $\w$ are also difficult to be solved, we therefore follow the approach in [@conf/icml/Suzuki13] [^8] to linearize the hinge loss part at every iteration. The difference between the approach in [@conf/icml/Suzuki13] and [@DBLP:conf/icml/OuyangHTG13] is that the former uses a special proximal term $\frac{1}{2\eta_{\tau}}(\w - \w_{\tau})^{\top}G_\tau(\w - \w_{\tau})$ to compute $\w_{\tau+1}$ at each iteration while the latter simply uses $\frac{1}{2\eta_{\tau}}\|\w - \w_{\tau}\|_2^2$, where $\eta_\tau$ is the step size and $G_{\tau}$ is a PSD matrix. We compared these two approaches and found that the variant in [@conf/icml/Suzuki13] works better for this problem and hence we only report its performance. The comparison between different algorithms starting from an initial solution with all zero entries for solving the above problem with $\lambda=0.1$ is presented in Figure \[fig:5\](a). For R$^2$SG, we start from $t_1 = 10^3$ and restart it every $10$ stages with $t$ increased by a factor of $1.15$. The initial step sizes for all algorithms are tuned, and so is the penalty parameter in ADMM. Finally, we compare the dependence of R$^2$SG’s convergence on the initial solution with that of SG. We use two different initial solutions (the first initial solution $\w_0=0$ and the second initial solution $\w_0$ is generated once from a normal Gaussian distribution). The convergence curves of the two algorithms from the two different initial solutions are plotted in Figure \[fig:5\](b). Note that the initial step sizes of SG and R$^2$SG are separately tuned for each initial solution. We can see that R$^2$SG is much less sensitive to a bad initial solution than SG consistent with our theory. Conclusion {#sec:conc} ========== In this work, we have proposed a novel restarted subgradient method for non-smooth and/or non-strongly convex optimization for obtaining an $\epsilon$-optimal solution. By leveraging the lower bound of the first-order optimality residual, we establish a generic complexity of RSG that improves over standard subgradient method. We have also considered several classes of non-smooth and non-strongly convex problems that admit a local error bound condition and derived the improved order of iteration complexities for RSG. Several extensions have been made to generalize the theory of RSG to a $p$-norm space and to design a parameter-free variant of RSG without requiring the knowledge of the multiple constant in the local error bound condition. Experimental results on several machine learning tasks have demonstrated the effectiveness of the proposed algorithms in comparison to the subgradient method. Acknolwedgements {#acknolwedgements .unnumbered} ================ We thank James Renegar for pointing out the connection to his work and for his valuable comments on the difference between the two work. We also thank to Nghia T.A. Tran for pointing out the connection between the local error bound and metric subregularity of subdifferentials. Thanks to Mingrui Liu for spotting an error in the formulation of the $F$ matrix for GFlasso in earlier versions. T. Yang is supported by NSF (1463988, 1545995). Proofs ====== Proof of Lemma \[lem:GD\] ------------------------- For any $\u\in\R^d$, we have $$\begin{aligned} & -(\u - \w_t )^{\top}\G(\w_t)\leq \frac{1}{\eta}(\u - \w_t)^{\top}(\w_{t+1} - \w_t)\leq \frac{1}{2\eta}(\|\u-\w_t\|^2_2 + \|\w_t-\w_{t+1}\|_2^2 - \|\u-\w_{t+1}\|_2^2)\\ &\leq \frac{\|\u - \w_t\|_2^2 - \|\u - \w_{t+1}\|_2^2}{2\eta} + \frac{\eta}{2}\|\G(\w_t)\|_2^2.\end{aligned}$$ Therefore, by summing the inequality above for $t=1,2,\dots,T$, we have $$\sum_{t=1}^T(f(\w_t) - f(\u))\leq \sum_{t=1}^T(\w_t - \u)^{\top}\G(\w_t)\leq \frac{\|\w_1 - \u\|_2^2}{2\eta} + \frac{\eta G^2T}{2}.$$ If $\wh_T = \frac{1}{T}\sum_{t=1}\w_t$, then $Tf(\wh_T)\leq \sum_{t=1}^Tf(\w_t)$ so that $$f(\wh_T) - f(\u)\leq \frac{\|\w_1 - \u\|_2^2}{2\eta T} + \frac{\eta G^2}{2}.$$ Proof of Lemma \[lem:polyeb\] ----------------------------- The proof below is motivated by the proof of Lemma 4 in [@DBLP:journals/mp/GilpinPS12]. However, since our Lemma \[lem:polyeb\] generalizes Lemma 4 in [@DBLP:journals/mp/GilpinPS12] by allowing the feasible set $\Omega$ to be unbounded, additional technical challenges are introduced in its proof, which lead to a parameter $\kappa$ with a definition different from the parameter $\delta$ in [@DBLP:journals/mp/GilpinPS12]. In the proof, we let $\|\cdot\|$ denote any valid norm. Since the epigraph is polyhedron, by Minkowski-Weyl theorem [@12891400], there exist finitely many feasible solutions $\{\w_i, i=1,\ldots, M\}$ with the associated objective values $\{f_i=f_i(\w_i), i=1,\ldots, M\}$ and finitely many directions $V=\left\{\left(\begin{array}{c}\u_1\\ s_1\end{array}\right),\ldots, \left(\begin{array}{c}\u_E\\ s_E\end{array}\right)\right\}$ in $\mathbb{R}^{p+1}$ such that $$\begin{aligned} &\text{epi}(f) = \text{conv}\{(\w_i, f_i), i=1,\ldots, M\} + \text{cone}(V)\\ &=\left\{(\w, t): \left(\begin{array}{c}\w\\ \\ t\end{array}\right) = \left(\begin{array}{c}\sum_{i=1}^M\lambda_i\w_i + \sum_{j=1}^E\gamma_j\u_j\\ \\ \sum_{i=1}^M\lambda_if_i + \sum_{j=1}^E\gamma_js_j\end{array}\right), \lambda\in\Delta, \gamma\in\R_+\right\}.\end{aligned}$$ Thus, we can express $f(\w)$ as $$\begin{aligned} \label{eqn:fw} f(\w) = \min_{\lambda, \gamma}\left\{\sum_{i=1}^M\lambda_if_i + \sum_{j=1}^E\gamma_j s_j: \w = \sum_{i=1}^M\lambda_i\w_i + \sum_{j=1}^E\gamma_j\u_j, \lambda\in\Delta, \gamma\in\R_+\right\}. \end{aligned}$$ Since $\min_{\w\in\Omega}f(\w) = f_*$, we have $\min_{1\leq i\leq M}f_i = f_*$ and $s_j\geq 0, \forall j$. We temporarily assume that $$f_1\geq f_2\geq \ldots\geq f_N> f_*= f_{N+1}=\ldots = f_M$$ for $N\geq1$. We denote by $\S\subset[E]$ the indices such that $s_j\neq 0$ and by $\S^c$ the complement. For any $\gamma\in\R^E_+$, we let $\gamma_{\S}$ denote a vector that contains elements $\gamma_i$ such that $i\in\S$. From (\[eqn:fw\]), we can see that there exist $\lambda\in\Delta$ and $\gamma\in\R_+$, such that $$\w = \sum_{i=1}^M\lambda_i \w_i + \sum_{j\in\S}\gamma_j\u_j + \sum_{j\in\S^c}\gamma_j\u_j, \quad\text{ and }\quad f(\w) = \sum_{i=1}^N\lambda_i f_i + f_*\sum_{i=N+1}^M\lambda_i +\sum_{j\in\S}\gamma_js_j.$$ Define $\w_s= \sum_{i=1}^M\lambda_i\w_i + \sum_{j\in\S}\gamma_j\u_j $. Then $$\begin{aligned} \label{eqn:key0} \|\w - \w^+\| & \leq \min_{\lambda'\in\Delta, \gamma'\in\R_+}\left\|\w - \left(\sum_{i=N+1}^M\lambda'_i\w_i + \sum_{j\in\S^c}\gamma'_j\u_j\right)\right\| \notag\\ & =\min_{\lambda'\in\Delta, \gamma'\in\R_+}\left\|\w_s + \sum_{j\in\S^c}\gamma_j\u_j - \left(\sum_{i=N+1}^M\lambda'_i\w_i + \sum_{j\in\S^c}\gamma'_j\u_j\right)\right\| \notag \\ &\leq \min_{\lambda'\in\Delta}\left\|\w_s - \sum_{i=N+1}^M\lambda_i'\w_i\right\| + \min_{\gamma'\in\R_+}\left\|\sum_{j\in\S^c}\gamma_j\u_j - \sum_{j\in\S^c}\gamma'_j\u_j\right\|\notag \\ &= \min_{\lambda'\in\Delta}\left\|\sum_{i=1}^M\lambda_i\w_i + \sum_{j\in\S}\gamma_j\u_j - \sum_{i=N+1}^M\lambda_i'\w_i\right\|\notag\\ &\leq \min_{\lambda'\in\Delta}\left\|\sum_{i=1}^M\lambda_i\w_i - \sum_{i=N+1}^M\lambda_i'\w_i\right\| +\left\|\sum_{j\in\S}\gamma_j\u_j\right\|, $$ where the first inequality is due to that $\sum_{i=N+1}^M\lambda'_i\w_i + \sum_{j\in\S^c}\gamma_j'\u_j\in\Omega_*$. Next, we will bound the two terms in the R.H.S of the above inequality. To proceed, we construct $\delta$ and $\sigma$ as follows: $$\begin{aligned} \delta& = \frac{f_N - f_*}{\max_{i, \u}\{\|\w_i -\u\|: i=1,\ldots, N, \u\in\Omega_*\}}>0,\\ \sigma& = \min_{j\in\S: \|\u_j\|\neq 0}\frac{s_j}{\|\u_j\|}>0. $$ Let $\w' = \sum_{i=1}^M\lambda_i\w_i$ and $\mu = \sum_{i=1}^N\lambda_i$. Suppose $\mu >0$. Let $\hat\lambda_i, i=1,\ldots, N$ be defined as $\hat\lambda_i = \frac{\lambda_i}{\mu}, i=1,\ldots, N$. Further define $$\wh = \sum_{i=1}^N\hat\lambda_i\w_i,\quad\text{ and }\quad \wt = \left\{\begin{array}{ll}\sum_{i=N+1}^M\w_i\frac{\lambda_i}{1-\mu}&\text{if } \mu<1\\ \\ \w_M&\text{ if }\mu = 1\end{array}\right..$$ As a result, $\w' = \mu\wh + (1-\mu) \wt$. For the first term in the R.H.S of (\[eqn:key0\]), we have $$\min_{\lambda'\in\Delta}\left\|\sum_{i=1}^M\lambda_i\w_i - \sum_{i=N+1}^M\lambda_i'\w_i\right\| = \min_{\lambda'\in\Delta}\left\|\w' - \sum_{i=N+1}^M\lambda_i'\w_i\right\|\leq \|\w' - \wt\|.$$ To continue, $$\begin{aligned} \|\w' - \wt\| &= \mu\|\wh - \wt\| \leq \mu\left\|\sum_{i=1}^N\hat\lambda_i(\w_i - \wt)\right\|\leq \mu\sum_{i=1}^N\hat\lambda_i\|\w_i - \wt\|\\ & \leq \mu \max_i\{\|\w_i - \wt\|, i=1,\ldots, N, \wt\in\Omega_*\} \leq \frac{\mu(f_N - f_*)}{\delta}.\end{aligned}$$ where the last inequality is due to the definition of $\delta$. On the other hand, we can represent $$f(\w)= \sum_{i=1}^M\lambda_if_i + \sum_{j\in\S}\gamma_js_j= \mu\sum_{i=1}^N\hat\lambda_if_i + f_*(1-\mu) + \sum_{j\in\S}\gamma_js_j.$$ Since $s_j\geq 0$ for all $j$, we have $$f(\w) - f_* \geq \mu\sum_{i=1}^N\hat\lambda_i (f_i - f_*) + \sum_{j\in\S}\gamma_js_j\geq \max\left\{\mu (f_N - f_*),\sum_{j\in\S}\gamma_js_j\right\}.$$ Thus, $$\min_{\lambda'\in\Delta}\left\|\sum_{i=1}^M\lambda_i\w_i - \sum_{i=N+1}^M\lambda_i'\w_i\right\| \leq \|\w' - \wt\|\leq \frac{f(\w) -f_*}{\delta}.$$ Suppose $\mu =0$ (and thus $\lambda_i=0$ for $i=1,2,\dots,N$). We can still have $$\min_{\lambda'\in\Delta}\left\|\sum_{i=1}^M\lambda_i\w_i - \sum_{i=N+1}^M\lambda_i'\w_i\right\| \leq \min_{\lambda'\in\Delta}\left\|\sum_{i=N+1}^M\lambda_i\w_i - \sum_{i=N+1}^M\lambda_i'\w_i\right\| =0\leq \frac{f(\w) -f_*}{\delta},$$ and can still represent $$f(\w)= \sum_{i=1}^M\lambda_if_i + \sum_{j\in\S}\gamma_js_j= f_* + \sum_{j\in\S}\gamma_js_j$$ so that $$f(\w)-f_*\geq \sum_{j\in\S}\gamma_js_j.$$ As a result, no matter $\mu=0$ or $\mu>0$, we can bound the first term in the R.H.S of (\[eqn:key0\]) by $\frac{f(\w) -f_*}{\delta}$. For the second term in the R.H.S of (\[eqn:key0\]), by the definition of $\sigma$ , we have $$\left\|\sum_{j\in\S}\gamma_j\u_j\right\|\leq \sum_{j\in\S}\gamma_j\|\u_j\| \leq \frac{1}{\sigma}\sum_{j\in\S}\gamma_j s_j\leq \frac{f(\w) - f_*}{\sigma},$$ Combining the results above with $\displaystyle\frac{1}{\kappa} =\left(\frac{1}{\delta} +\frac{1}{\sigma}\right)$, we have $$\|\w-\w^+\|\leq \frac{1}{\kappa}(f(\w) - f_*).$$ Finally, we note that when $f_1=\ldots=f_M=f_*$, the Lemma is trivially proved following the same analysis except that $\delta>0$ can be any positive value. A proposition needed to prove Corollary \[thm:KL\] -------------------------------------------------- The proof of Corollary \[thm:KL\] leverages the following result from [@arxiv:1510.08234]. \[prop:KL\][@arxiv:1510.08234 Theorem 5] Let $f(x)$ be an extended-valued, proper, convex and lower semicontinuous function that satisfies the KL inequality at $x_*\in\arg\min f(\cdot)$ for all $x\in U\cap \{x: f(x_*)< f(x)< f(x_*)+\eta\}$, where $U$ is a neighborhood of $x_*$, then $dist(x, \arg\min f(\cdot))\leq \varphi(f(x) - f(x_*))$ for all $x\in U\cap \{x: f(x_*)< f(x)< f(x_*)+\eta\}$. Proof of Lemma \[lem:sgdp\] --------------------------- According to [@Nesterov:2009:PSM:1530733.1530741], we have $$\sum_{t=1}^T\lambda_t\G(\w_t)^{\top}(\w_t - \w)\leq \frac{1}{2\eta}\|\w - \w_1\|_p^2 + \frac{\eta}{2(p-1)}\sum_{t=1}^T\lambda^2_t\|\G(\w_t)\|_q^2.$$ Taking expectation on both sides and by the convexity of $f(\w)$, we have $$\sum_{t=1}^T\lambda_t(f(\w_t) - f(\w))\leq \frac{1}{2\eta}\|\w - \w_1\|_p^2 +\frac{\eta}{2(p-1)}\sum_{t=1}^T\lambda^2_t\|\G(\w_t)\|_q^2.$$ By the definition of $\wh_T$ and the convexity of $f(\w)$, we have $$f(\wh_T) - f(\w)\leq \frac{\|\w - \w_1\|_p^2}{2\eta \Lambda_T}+ \frac{\eta\sum_{t=1}^T\lambda^2_t\|\G(\w_t)\|_q^2 }{2(p-1)\Lambda_T}.$$ The lemma can be proved by noting that $\|\G(\w_t)\|_q\leq G$. [^1]: The first version of this paper appeared on arXiv on December 9, 2015 and a complete version of this paper appeared on arXiv on April 5, 2016. [^2]: Without assuming smoothness and strong convexity. [^3]: The Euclidean norm in the definition here can be replaced by a general norm as in [@Nesterov:2014:uniform_convex]. [^4]: In fact, this property of $f(\cdot)$ is a global error bond on $\Omega$. [^5]: Recall . [^6]: Other choices are possible [@DBLP:conf/colt/DuchiSST10]. [^7]: <https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/> [^8]: The OPG variant.

ArXiv

--- title: | -10pt TYPED FEATURE STRUCTURES AS DESCRIPTIONS Paul John King[^1] Seminar für Sprachwissenschaft, Eberhard-Karls-Universität[^2] -30pt --- = 16mm = -17mm = -2mm = 244mm =eufm10 [ ]{} \#1 [ ]{} \#1 [\#1]{} [ ]{} \#2 [ ]{} [ ]{} [ \#1 ]{} [ ]{} \#1 \#1[\#1]{} § ¶ Å ABSTRACT {#abstract .unnumbered} -------- A description is an entity that can be interpreted as true or false of an object, and using feature structures as descriptions accrues several computational benefits. In this paper, I create an explicit interpretation of a typed feature structure used as a description, define the notion of a satisfiable feature structure, and create a simple and effective algorithm to decide if a feature structure is satisfiable. 1. INTRODUCTION {#introduction .unnumbered} --------------- Describing objects is one of several purposes for which linguists use feature structures. A description is an entity that can be interpreted as true or false of an object. For example, the conventional interpretation of the description ‘it is black’ is true of a soot particle, but false of a snowflake. Therefore, any use of a feature structure to describe an object demands that the feature structure can be interpreted as true or false of the object. In this paper, I tailor the semantics of [[@King; @1989]]{} to suit the typed feature structures of [[@Carpenter; @1992]]{}, and so create an explicit interpretation of a typed feature structure used as a description. I then use this interpretation to define the notion of a satisfiable feature structure. Though no feature structure algebra provides descriptions as expressive as those provided by a feature logic, using feature structures to describe objects profits from a large stock of available computational techniques to represent, test and process feature structures. In this paper, I demonstrate the computational benefits of marrying a tractable syntax and an explicit semantics by creating a simple and effective algorithm to decide the satisfiability of a feature structure. Gerdemann and Götz’s Troll type resolution system implements both the semantics and an efficient refinement of the satisfiability algorithm I present here (see , [[@Gerdemann; @and; @King; @1994]]{} and [[@Gerdemann; @(fc)]]{}). 2. A FEATURE STRUCTURE SEMANTICS {#a-feature-structure-semantics .unnumbered} -------------------------------- A signature provides the symbols from which to construct typed feature structures, and an interpretation gives those symbols meaning. Henceforth, I tacitly work with a signature $\sequence{\Q,\T,\preceq,\S,\A,\F}$. I call members of $\Q$ states, members of $\T$ types, $\preceq$ subsumption, members of $\S$ species, members of $\A$ attributes, and $\F$ appropriateness. Suppose that $I$ is an interpretation $\sequence{U,S,A}$. I call each member of $U$ an object in $I$. Each type denotes a set of objects in $I$. The denotations of the species partition $U$, and $S$ assigns each object in $I$ the unique species whose denotation contains the object: object $u$ is in the denotation of species $\sigma$ iff $\sigma=S(u)$. Subsumption encodes a relationship between the denotations of species and types: object $u$ is in the denotation of type $\tau$ iff $\tau\preceq S(u)$. So, if $\tau_1\preceq\tau_2$ then the denotation of type $\tau_1$ contains the denotation of type $\tau_2$. Each attribute denotes a partial function from the objects in $I$ to the objects in $I$, and $A$ assigns each attribute the partial function it denotes. Appropriateness encodes a relationship between the denotations of species and attributes: if $\F\sequence{\sigma,\alpha}$ is defined then the denotation of attribute $\alpha$ acts upon each object in the denotation of species $\sigma$ to yield an object in the denotation of type $\F\sequence{\sigma,\alpha}$, but if $\F\sequence{\sigma,\alpha}$ is undefined then the denotation of attribute $\alpha$ acts upon no object in the denotation of species $\sigma$. So, if $\F\sequence{\tau,\alpha}$ is defined then the denotation of attribute $\alpha$ acts upon each object in the denotation of type $\tau$ to yield an object in the denotation of type $\F\sequence{\tau,\alpha}$. I call a finite sequence of attributes a path, and write $\P$ for the set of paths. I write $P_I$ for the path interpretation function under $I$. Each feature structure is a connected Moore machine (see [[@Moore; @1956]]{}) with finitely many states, input alphabet $\A$, and output alphabet $\T$. 3. MORPHS {#morphs .unnumbered} --------- The abundance of interpretations seems to preclude an effective algorithm to decide if a feature structure is satisfiable. However, I insert [*morphs*]{} between feature structures and objects to yield an interpretation free characterisation of a satisfiable feature structure. Each morph is the Moshier abstraction (see [[@Moshier; @1988]]{}) of a connected and totally well-typed (see [[@Carpenter; @1992]]{}) Moore machine with possibly infinitely many states, input alphabet $\A$, and output alphabet $\S$. I thus write of [*the*]{} abstraction of $u$ under $I$. I write $\UU$ for the set of standard objects, write $\SS$ for the total function from $\UU$ to $\S$, where and write $\AA$ for the total function from $\A$ to the set of partial functions from $\UU$ to $\UU$, where I write $\II$ for $\sequence{\UU,\SS,\AA}$. A feature structure approximates a morph iff the Moshier abstraction of the feature structure abstractly subsumes (see [[@Carpenter; @1992]]{}) the morph. 4. RESOLVED FEATURE STRUCTURES {#resolved-feature-structures .unnumbered} ------------------------------ Though theorem \[F\_is\_satisfiable\_iff\_F\_approximates\_some\_morph\] gives an interpretation free characterisation of a satisfiable feature structure, the characterisation still seems to admit of no effective algorithm to decide if a feature structure is satisfiable. However, I use theorem \[F\_is\_satisfiable\_iff\_F\_approximates\_some\_morph\] and [*resolved feature structures*]{} to yield a less general interpretation free characterisation of a satisfiable feature structure that admits of such an algorithm. Each resolved feature structure is a well-typed (see [[@Carpenter; @1992]]{}) feature structure with output alphabet $\S$. 5. A SATISFIABILITY ALGORITHM {#a-satisfiability-algorithm .unnumbered} ----------------------------- In this section, I use theorem \[F\_is\_satisfiable\_iff\_F\_has\_a\_resolvant\] to show how – given a rational signature that meets reasonable computational conditions – to construct an effective algorithm to decide if a feature structure is satisfiable. Gerdemann and Götz’s Troll system (see , [[@Gerdemann; @and; @King; @1994]]{} and [[@Gerdemann; @(fc)]]{}) employs an efficient refinement of $\Res$ to test the satisfiability of feature structures. In fact, Troll represents each feature structure as a disjunction of the resolvants of the feature structure. Loosely speaking, the resolvants of a feature structure have the same underlying finite state automaton as the feature structure, and differ only in their output function. Troll exploits this property to represent each feature structure as a finite state automaton and a set of output functions. The Troll unifier is closed on these representations. Thus, though $\Res$ is computationally expensive, Troll uses $\Res$ only during compilation, never during run time. References {#references .unnumbered} ---------- [^1]: The research presented in this paper was sponsored by Teilprojekt B4 “Constraints on Grammar for Efficient Generation” of the Sonderforschungsbereich 340 of the Deutsche Forschungsgemeinschaft. I also wish to thank Bob Carpenter, Dale Gerdemann, Thilo Götz and Jennifer King for their invaluable help with this paper. [^2]: Wilhelmstr. 113, 72074 Tübingen, Germany. Email: king@sfs.nphil.uni-tuebingen.de.

ArXiv

--- author: - | G. Ambrosi$^{\;(1)}$, M. Ambrosio$^{\;(2)}$, C. Aramo$^{\;(2)}$, E. Bissaldi$^{\; \dagger \,(3,4)}$, M. Capasso$^{\;(5)}$, D. Corti$^{\;(6)}$, A. de Angelis$^{\;(6)}$, F. de Palma$^{\;(7,8)}$, F. Ferrarotto$^{\;(9)}$, A. Ferri$^{\;(10)}$, S. Garrappa$^{\;(5)}$, N. Giglietto$^{\;(5,8)}$, $^{\;\; \dagger \;(5,8)}$, A. Gola$^{\;(10)}$, M. Ionica$^{\;(1)}$, M. Iori$^{\;(9,11)}$, F. Licciulli$^{\;(12)}$, M. Mariotti$^{\;(6)}$, C. Marzocca$^{\;(8,12)}$, R. Paoletti$^{\;(13)}$, C. Piemonte$^{\;(10)}$, V. Postolache$^{\;(1)}$, R. Rando$^{\;(6)}$, C. Stella$^{\;(4,14)}$, P. Vallania$^{\;(15,16)}$, C. Vigorito$^{\;(15,17)}$\ $^{(1)}$INFN– Sezione di Perugia; $^{(2)}$INFN– Sezione di Napoli; $^{(3)}$Dipartimento di Fisica, Università di Trieste; $^{(4)}$INFN– Sezione di Trieste–Udine; $^{(5)}$Dipartimento Interateneo di Fisica, Università e Politecnico di Bari; $^{(6)}$INFN– Sezione di Padova; $^{(7)}$Università Telematica Pegaso; $^{(8)}$INFN– Sezione di Bari; $^{(9)}$INFN– Sezione di Roma I; $^{(10)}$FBK– Trento; $^{(11)}$Sapienza– Università di Roma; $^{(12)}$Politecnico di Bari; $^{(13)}$INFN– Sezione di Pisa; $^{(14)}$Dipartimento di Chimica, Fisica e Ambiente, Università di Udine; $^{(15)}$INAF– Osservatorio Astrofisico di Torino; $^{(16)}$INFN– Sezione di Torino; $^{(17)}$Dipartimento di Fisica, Università di Torino.\ $\dagger$ title: Measurements and tests on FBK silicon sensors with an optimized electronic design for a CTA camera --- The prototype ============= The Sensors ----------- The devices employed for the realization of the matrix prototype make use of the FBK [@FBK14] Near Ultra-Violet (NUV) SiPM technology. The basic structure of these SiPMs consists of a p$^+-$n junction, whose design is optimized for the detection in the blue-NUV region of the electro-magnetic spectrum [@PRO13]. They show low breakdown voltages (around 25.6 V) with a slight temperature dependence of about 24 mV/$^{\rm o}$C. Moreover, these devices also show a Photon Detection Efficiency (PDE) of about 24% at 380 nm for 2 V of over-voltage (OV). In particular, the matrix we tested consists of 16 SiPMs with 50$\mu$m cells, for a total equipped area of 12$\times$12 mm$^2$ each (see Figure \[Fig\_1\], [*left panel*]{}). -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![[*Left panel*]{}: On the right, we can see the assembled matrix with the protective layer of epoxy resin. On the left, the printed circuit board (PCB) with the 16 AD8000 OPAs. On the top, Dante’s face for size comparison. [*Right panel:*]{} Another PCB showing 16 LEMO connectors for the readout.[]{data-label="Fig_1"}](SiPM_1.eps "fig:"){height="55.00000%"} ![[*Left panel*]{}: On the right, we can see the assembled matrix with the protective layer of epoxy resin. On the left, the printed circuit board (PCB) with the 16 AD8000 OPAs. On the top, Dante’s face for size comparison. [*Right panel:*]{} Another PCB showing 16 LEMO connectors for the readout.[]{data-label="Fig_1"}](SiPM_2.eps "fig:"){height="55.00000%"} -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Front End Electronics --------------------- We designed a preamplifier based on an AD8000 operational amplifier (OPA) in transimpe-dance configuration. The design consists of 20 $\Omega$ decoupling resistances with a 1 k$\Omega$ feedback resistance. The current signal from the SiPM is fed to the preamplifier in DC mode, as illustrated in Figure \[Fig\_2\]. ![Schematic view of the SiPM and the AD8000 OPA. The Pole-Zero cancellation circuit is also drawn.[]{data-label="Fig_2"}](ad8000.eps){height="35.00000%"} The [*left panel*]{} of Figure \[Fig\_3\] shows the current signal from the preamplifier corresponding to five photo-electrons (p.e.). The very fast rise time is due to the development of the current signal in the SiPM junction and it is about 100 ps long. Moreover, a fast decay followed by a very long tail is clearly visible. This long tail is due to the recovery time obtained by the product of the quenching resistor times the single cell capacitance value, and it is about 100 ns long. Such a long recovery time can be problematic in high-background environments, like the one expected for the CTA experiment. For this purpose, a Pole-Zero cancellation circuit [@GOL13] was added to the preamplifier design (see Figure \[Fig\_2\]). A 470 pF capacitance was placed in parallel to a 1 k$\Omega$ resistance and closed over a 50 $\Omega$ load. The signal taken from this Pole-Zero cancellation network is shown in the [*right panel*]{} of Figure \[Fig\_3\]. The absence of the long tail is clearly visible, and the same amplitude is almost preserved. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Current signals from the preamplifier obtained without ([*left panel*]{}) and with the Pole-Zero cancellation circuit ([*right panel*]{}).[]{data-label="Fig_3"}](PCB_1.eps "fig:"){height="0.24\textheight"} ![Current signals from the preamplifier obtained without ([*left panel*]{}) and with the Pole-Zero cancellation circuit ([*right panel*]{}).[]{data-label="Fig_3"}](PCB_2.eps "fig:"){height="0.24\textheight"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Prototype performance ===================== We tested the coupling of the SiPM matrix with the electronics by means of a pulsed-mode laser light. Four out of the 16 signals were sent to a TDS5104B oscilloscope, with a signal sampling of 5 GHz. Two screen-shots are shown in Figure \[Fig\_4\]. The [*left panel*]{} refers to the dark count signals. Here, the average amplitude per single p.e. is about 5 mV and the average distance time is about 200 ns, corresponding to a total dark rate of about 500 kHz/mm$^2$. It is worth noting that the SiPMs adopted for this analysis represent a very preliminary production by FBK. More recently developed sensors show an improvement in the dark count rate by an order of magnitude. The [*right panel*]{} of Figure \[Fig\_4\] shows the signal sampling in the case of a laser event, characterized by the four SiPM signals being coincident in time. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Screenshots from the oscilloscope’s monitor. [*Left panel*]{}: Signals sampled in the dark; [*Right panel*]{}: Signals sampled in coincidence with the laser pulse.[]{data-label="Fig_4"}](CRD_1.eps "fig:"){height="35.00000%"} ![Screenshots from the oscilloscope’s monitor. [*Left panel*]{}: Signals sampled in the dark; [*Right panel*]{}: Signals sampled in coincidence with the laser pulse.[]{data-label="Fig_4"}](CRD_2.eps "fig:"){height="35.00000%"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ We studied the amplitude of the signal from a single SiPM by covering all other SiPMs. An example amplitude spectrum is shown in the [*top left panel*]{} of Figure \[Fig\_5\]. The amplitude distribution of all p.e. events clearly exhibits a Poissonian shape, as can be seen [*top right panel*]{} of Figure \[Fig\_5\]. We also studied the linearity of the SiPM electronics chain. The [*bottom left panel*]{} of Figure \[Fig\_5\] displays a linear fit, which indicates a good linear correlation of the signal amplitude vs. the number of converted photo-electrons. Finally we evaluated the signal-to-noise ratio (SNR) by calculating the ratio between the gain (i.e. the position of the single p.e. amplitude) to the width of the electronic noise (see the [*bottom right panel*]{} of Figure \[Fig\_5\]). ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![ [*Top left panel:*]{} Amplitude spectrum from a single SiPM. [*Top right panel:*]{} Amplitude distribution. [*Bottom left panel:*]{} Gain. [*Bottom right panel:*]{} SNR. []{data-label="Fig_5"}](Ampiezze.eps "fig:"){height="30.00000%"} ![ [*Top left panel:*]{} Amplitude spectrum from a single SiPM. [*Top right panel:*]{} Amplitude distribution. [*Bottom left panel:*]{} Gain. [*Bottom right panel:*]{} SNR. []{data-label="Fig_5"}](AREA_VS_PE.eps "fig:"){height="30.00000%"} ![ [*Top left panel:*]{} Amplitude spectrum from a single SiPM. [*Top right panel:*]{} Amplitude distribution. [*Bottom left panel:*]{} Gain. [*Bottom right panel:*]{} SNR. []{data-label="Fig_5"}](CENTER_VS_PE_Matrix.eps "fig:"){height="30.00000%"} ![ [*Top left panel:*]{} Amplitude spectrum from a single SiPM. [*Top right panel:*]{} Amplitude distribution. [*Bottom left panel:*]{} Gain. [*Bottom right panel:*]{} SNR. []{data-label="Fig_5"}](SNR_VS_PE_Matrix.eps "fig:"){height="30.00000%"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- In summary, we conclude that our device shows a gain of 5.6 mV/p.e. with a SNR of about 8.3 for an OV of roughly 2 V. Future Work =========== INFN is currently producing a cluster of 7 or more 3 $\times$ 3 (or 6 $\times$ 6) SiPMs of 1 inch diameter, which can be suitable for telescopes with big cameras, like CTA’s large size telescopes (LSTs). The [*left panel*]{} of Figure \[Fig\_Last\] shows a preliminary design of a PCB for housing a matrix of 16 SiPMs and an electronics optimized to collect the signals of all these sensors and to fed out the signal of only one pixel. The coupling of this high dimensions pixel with a light guide (shown in the [*right panel*]{}) is also being exploited. The first prototypes have been produced with a cooperation between INFN Padova and INFN Perugia and will be tested on the MAGIC telescope. Moreover, FBK is continuously producing and upgrading devices with a reduced dark count rate, an increased PDE and an optimized geometry, in order to reduce the dead area zone. Further studies are in progress in order to optimize the current signal tail in order to limit the impact of the diffuse light background to the trigger efficiency and signal reconstruction. For this purpose, more tests are currently under investigation. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![High dimensions pixel: [*Left panel:*]{} array of SiPMs and electronics; [*Right panel:*]{} the SiPM array seen through the light guide.[]{data-label="Fig_Last"}](Fig_6a.eps "fig:"){height="29.00000%"} ![High dimensions pixel: [*Left panel:*]{} array of SiPMs and electronics; [*Right panel:*]{} the SiPM array seen through the light guide.[]{data-label="Fig_Last"}](Fig_6b.eps "fig:"){height="29.00000%"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [99]{} http://www.fbk.eu/ Pro, T., et al., 2013, “New Developments of Near-UV SiPMs at FBK”, [*IEEE TRANSACTIONS ON NUCLEAR SCIENCE*]{}, Volume: 60, Issue:3, Part:3 Gola, A., Piemonte, C., Tarolli A., 2013, “Analog Circuit for Timing Measurements With Large Area SiPMs Coupled to LYSO Crystals”, [*IEEE TRANSACTIONS ON NUCLEAR SCIENCE*]{}, Volume: 60, Issue:2, Part:2

ArXiv

--- author: - | Lillian Lee, Cornell University\ http://www.cs.cornell.edu/home/llee title: | A Matter of Opinion: Sentiment Analysis and Business Intelligence\ (position paper) --- [A Matter of Opinion: Sentiment Analysis and Business Intelligence (position paper)]{}\ [Lillian Lee, Cornell University]{}\ Presented at the IBM Faculty Summit on the Architecture of On-Demand Business, May 2004 #### Motivation In the novel [*Hard Times*]{}, Charles Dickens described the fictional “Coketown” as follows: > Fact, fact, fact, everywhere in the material aspect of the town; fact, fact, fact, everywhere in the immaterial. The M’Choakumchild school was all fact, and the school of design was all fact, and the relations between master and man were all fact, and [everything was fact between the lying-in hospital and the cemetery, and what you couldn’t state in figures, or show to be purchasable in the cheapest market and salable in the dearest, was not]{}, and never should be, world without end, Amen. In real-life business intelligence, facts are of course very important, but [*opinion*]{} also plays a crucial role. Consider, for instance, the following scenario. A major computer manufacturer, disappointed with unexpectedly low sales, finds itself confronted with this question: Why aren’t consumers buying our laptop? While concrete data such as the laptop’s weight or the price of a competitor’s model are obviously relevant, answering this question requires focusing more on people’s personal [*views*]{} of such objective characteristics. Moreover, subjective judgments regarding intangible qualities — e.g., “the design is tacky” or “customer service was condescending” — or even misperceptions — “updated device drivers aren’t available” — must be taken into account as well. [*Sentiment-analysis technologies*]{} for extracting opinions from unstructured human-authored documents would be excellent tools for handling many business-intelligence tasks related to the one just described. Continuing with our example scenario: it would be difficult to try to directly survey laptop purchasers who [*haven’t*]{} bought the company’s product. Rather, we could employ a system that (a) finds reviews or other expressions of opinion on the Web — newsgroups, individual blogs, and aggregation sites such as epinions.com are likely to be productive sources — and then (b) creates condensed versions of the reviews or a digest of the overall consensus. This would save the analyst from having to read potentially dozens or even hundreds of versions of the same complaints. Note that Internet sources can vary wildly in form, tenor, and even grammaticality; this fact underscores the need for robust techniques even when only one language (e.g., English) is considered. #### Challenges in sentiment classification Given the multitude of potential applications, researchers have been devoting more and more attention to sentiment analysis. Much of the current work is devoted to [*classification*]{} problems: determining whether a particular document or portion thereof is subjective or not, and/or determining whether the opinion it expresses is positive or negative. At first blush, this might not appear so hard: one might expect that we need simply look for obvious sentiment-indicating words, such as “great”. The difficulty lies in the richness of human language use. First, there can be an amazingly large number of ways to say the same thing (especially, it seems, when that thing is a negative perception); this complicates the task of finding a high-coverage set of indicators. Furthermore, the same indicator may admit several different interpretations. Consider, for example, the following sentences: - This laptop is . - of media attention surrounded the release of the new laptop model. - If you think this laptop is , I’ve got a nice bridge you might be interested in. Each of these sentences contains the three words “a great deal”, but the opinions expressed are, respectively, positive, neutral, and negative. The first two sentences use the same phrase to mean different things. The last sentence involves sarcasm, which, along with related rhetorical devices, is an intrinsic feature of texts from unrestricted domains such as blogs and newsgroup postings. In general, researchers have adopted one of two approaches to meeting the challenges that sentiment analysis presents. Many groups are working to directly improve the selection and interpretation of indicators through the incorporation of linguistic knowledge; given the subtleties of natural language, such efforts will be critical to building operational systems. Others have been pursuing a different tack: employing [*learning algorithms*]{} that can automatically infer from text samples what indicators are useful. Besides being potentially more cost-effective, more easily ported to other domains and languages, and more robust to grammatical mistakes, learning-based systems can also discover indicators that humans might neglect. For example, in our own work, we found that the phrase “still,” (comma included) is a better indicator of positive sentiment than “good” — a typical instance of use would be a sentence like “Still, despite these flaws, I’d go with this laptop”. Nevertheless, it bears repeating that incorporating deep knowledge about language will be absolutely crucial to developing systems capable of high-quality (as opposed to merely high-throughput) sentiment analysis. Both the linguistic and the learning approach have considerable merits; it seems very safe to say that the community will need to turn towards finding ways to combine their advantages. #### Related problems, new directions The classification problems discussed above only involve the determination of sentiment. However, there is growing interest in capturing interactions between [*subjectivity*]{} and [*subject*]{} — we not only need to know what an author’s opinion is, but what that opinion is about. For example, while in a broad sense a review of a particular laptop is only about one topic (the laptop itself), it almost surely discusses various specific aspects of the machine. We would ideally like a sentiment-analysis system to reveal whether there are particular features that the review’s author disapproves of even if his or her overall impression was positive. Another interesting research direction of potentially great importance is to integrate into sentiment analysis the notion of the [*status*]{} of an opinion holder, perhaps via adaptation of the hubs-and-authorities techniques used in Web search or link-analysis methods in reputation systems. For example, we might want to identify [*bellwethers*]{} — thought leaders with enough influence that others explicitly adopt their opinions — or [*barometers*]{} — those whose opinions are generally held by the majority of the population of interest. Tracking the views of these two types of people could both streamline and enhance the process of gathering business intelligence to a large degree. Surely that sounds like a great deal!

ArXiv

--- abstract: 'Nucleosynthesis in the $s$ process takes place in the He burning layers of low mass AGB stars and during the He and C burning phases of massive stars. The $s$ process contributes about half of the element abundances between Cu and Bi in solar system material. Depending on stellar mass and metallicity the resulting $s$-abundance patterns exhibit characteristic features, which provide comprehensive information for our understanding of the stellar life cycle and for the chemical evolution of galaxies. The rapidly growing body of detailed abundance observations, in particular for AGB and post-AGB stars, for objects in binary systems, and for the very faint metal-poor population represents exciting challenges and constraints for stellar model calculations. Based on updated and improved nuclear physics data for the $s$-process reaction network, current models are aiming at ab initio solution for the stellar physics related to convection and mixing processes. Progress in the intimately related areas of observations, nuclear and atomic physics, and stellar modeling is reviewed and the corresponding interplay is illustrated by the general abundance patterns of the elements beyond iron and by the effect of sensitive branching points along the $s$-process path. The strong variations of the $s$-process efficiency with metallicity bear also interesting consequences for Galactic chemical evolution.' author: - 'F. Käppeler' - 'R. Gallino' - 'S. Bisterzo' - Wako Aoki title: 'The $s$ Process: Nuclear Physics, Stellar Models, Observations' --- Introduction \[sec1\] ===================== Neutron capture nucleosynthesis during stellar He burning contributes about half of the elemental abundances between Fe and Bi. Substantial progress in the quantitative description of this slow neutron capture ($s$) process has been achieved by an interdisciplinary approach involving improved nuclear physics input, advanced stellar model codes, and a wealth of data from astronomical observations and from the analysis of circumstellar dust grains. These three topics and their mutual connections are briefly described, before each topic is addressed in detail in Secs. II to IV. In the final section the synergies between the main topics are outlined with particular emphasis on the open quests and future prospects of this field. The phenomenological picture of the classical $s$ process was formulated about 50 years ago in the seminal papers of @BBF57 (hereafter referred to as B$^2$FH) and of @Cam57, where the entire $s$-process panorama was already sketched in its essential parts. These ideas were worked out in the following decades by @CFH61, @SFC65, @ClR67, and @ClW74. The distinction of a slow and a rapid ($s$ and $r$) neutron capture process follows from the isotopic pattern in the chart of nuclides (Fig. \[fig:1\]), which shows that the $s$ process follows the stability valley because the neutron capture time scale is slower than that for $\beta$ decay. The neutron-rich isotopes outside the $s$ path are ascribed to the $r$ process, which occurs under explosive conditions, presumably in supernovae. The decay of the reaction products from the $r$-process path on the far neutron-rich side of the stability valley forms the $r$-only isotopes. It also contributes to most of the other isotopes, except for those, which are shielded by stable isobars. The corresponding ensembles of $s$- and $r$-only isotopes are important for the separation of the respective abundance distributions. The subset of stable isotopes on the proton-rich side are ascribed to the $p$ process, which is likely to occur also in supernova explosions [@ArG03]. With a few exceptions, the $p$ abundances are much smaller than the $s$ and $r$ components. ![\[fig:1\] An illustration of the neutron capture processes responsible for the formation of the nuclei between iron and the actinides. The observed abundance distribution in the inset shows characteristic twin peaks. These result from the nuclear properties where the $s$- and $r$-reaction paths encounter magic neutron numbers. Note that a $p$ process has to be invoked for producing the proton rich nuclei that are not reached by neutron capture reactions. (For details see discussion in text.)](rmp_fig1.eps){width="12cm"} The decisive role of nuclear physics for a quantitative model of the $s$ process was clearly expressed already by B$^2$FH. In spite of the fact that neither the neutron source reactions nor the neutron capture cross sections in the astrophysically relevant energy range were known apart from some scatterted and uncertain information, all essential features had been inferred from this meager information: The product of the stellar ($n, \gamma$) cross sections and of the resulting $s$ abundances, $\langle \sigma \rangle N_s$, which represents the reaction flow, was found to be a smooth function of mass number $A$. From the composite slope of this function, the two different $s$ processes had already been postulated. The steep decline between $A \approx 63$ and 100 was interpreted as the result of an $s$-process site with not enough neutrons available per $^{56}$Fe seed to build the nuclei to their saturation abundances. In the mass region beyond $A \approx 100$, the much smaller slope was suggesting that steady flow was achieved and that all of these nuclei reached their saturation abundances. It was concluded that “two different processes might have occurred in two different red-giant stars (B$^2$FH)”. In the 1990s, however, improvements in the accuracy of the nuclear input data revealed that the classical $s$ process suffered from inconsistencies in the description of the abundance signatures in $s$-process branchings. Because such patterns are particularly sensitive to neutron density and temperature, this implied that these parameters were not constant in time as assumed in the formulation of the classical model [@Kae99]. A few years before, stellar models of the He burning stages of stellar evolution started to provide an increasingly realistic picture of $s$-process nucleosynthesis. The prospects of this approach were clearly superior to that of the classical model because it could be directly linked to astronomical observations. A first generation of models [@HoI88; @GBP88] was soon replaced by scenarios related to core He [@Heg06; @Lim06] and shell C burning [@RBG91b; @RGB93; @LSC00] in massive stars for the weak $s$ process on the one hand, and to thermally pulsing low mass asymptotic giant branch (AGB) stars [@SGB95; @GAB98; @AKW99b] for the main $s$ process on the other hand. The current status of AGB evolution includes phenomena such as hot bottom burning, the ab initio treatment of third dredge up and related mixing processes as well as the effect of metallicity and initial stellar mass [@Her05]. The latter aspects are particularly important for the discussion of the $s$-process component in galactic chemical evolution [@TGA04]. The success of the stellar models could be impressively verified by comparison with the solar $s$ component and with a large body of data obtained from analyses of presolar material in form of refractive dust grains of circumstellar origin [@Zin98]. With respect to the origin of the heavy elements, observations of $s$-process abundances in AGB stars began 1952 with the discovery of Tc lines in red giant stars of spectral type S by @Mer52b. Ever since, spectral observations of peculiar red giants turned out to be a prolific source of $s$-process information for the He burning stage of stellar evolution [@Gus89]. Spectroscopy of astronomical objects has made spectacular progress over the last decades by the deployment of new telescopes on the ground and in space and by the impressive increase in computing power, that led to enormous improvement in the modeling of stellar atmospheres and in synthetic spectrum calculations [@Asp05]. Our understanding of stellar and galactic evolution has been promoted accordingly, e.g. via refined studies along the AGB [@Lam91; @LSB95; @Her05]. Separation of the $s$ and $r$ components in solar material through careful evaluation of the $s$ abundances [@AKW99b] provided the key for the abundance distributions in the oldest, very metal-poor stars in the Universe [@CoS06; @CHB01; @SCB98], which were found to scale with the solar $r$-process distribution [@BeC05]. The composition of planetary nebulae [@PeB94] and circumstellar envelopes [@Hab96] could be investigated by IR observations, while the composition of interstellar matter [@SaS96] is inferred from UV absorption line diagnostics. At higher energies, X-ray [@FRA97] and $\gamma$-ray astronomy [@Die98] have produced exciting new vistas of explosive nucleosynthesis [@CLT92; @DBB97]. The three aspects of $s$-process research are addressed in the following sections with an attempt to illustrate their mutual connections. Nuclear physics \[sec2\] ======================== The discussion of the nuclear part concentrates on the neutron capture reactions and $\beta$-decay rates needed to calculate the $s$ abundances between Cu and Bi. For a summary of the charged particle reactions, which are of key importance for energy and neutron production at the various $s$-process sites and for further stellar evolution see the compiled data of the NACRE collaboration [@AAR99] and of @IDS01 as well as recent work on $^{12}$C [@AFL06; @PHK05] and on the two major neutron source reactions $^{13}$C($\alpha, n$)$^{16}$O [@HKT08] and $^{22}$Ne($\alpha, n$)$^{25}$Mg [@JKM01]. In their discussion of the $s$-abundance characteristics B$^2$FH noted that more detailed conclusions were impeded by the lack of reliable neutron capture cross sections and emphasized that “unambiguous results would be obtained by measuring the total absorption cross sections. It is our view that such measurements would serve as a crucial test of the validity of the $s$ process.” The importance of a complete set of experimental data for the reliable description of the $\langle \sigma \rangle N_s$-curve is illustrated in Fig. \[fig:2\], which corresponds to the situation obtained by the time of the cross section compilation of @BBK00, when experimentally determined cross sections were available for the majority of the involved isotopes. ![\[fig:2\] The characteristic product of cross section times $s$-process abundance, $\langle \sigma \rangle N_s$, plotted as a function of mass number. The thick solid line represents the main component obtained by means of the classical model, and the thin line corresponds to the weak component in massive stars (see text). Symbols denote the empirical products for the $s$-only nuclei. Some important branchings of the neutron capture chain are indicated as well.](rmp_fig2.eps){width="12cm"} Apart from the clear separation of the two $s$-process components, Fig. \[fig:2\] also shows the pronounced effect of $s$-process branchings, which could not be addressed by B$^2$FH simply because the data at the time were far too uncertain to reveal their signatures in the $\langle \sigma \rangle N_s$-curve. These branchings are the result of the competition between neutron capture and $\beta$-decay at unstable isotopes in the half-life range from a few weeks to a few years. For the same reason, also the effect of stellar temperature on the $\beta$-decay half-lives had not been anticipated. $^{79}$Se represents such an example, where the drastically reduced stellar half-life gives rise to a pronounced branching that can be characterized by the strongly different $\langle \sigma \rangle N_s$-values of the $s$-only isotopes $^{80}$Kr and $^{82}$Kr [@KlK88]. This section on the nuclear physics of the $s$ process starts with a summary of current techniques for the experimental determination of stellar neutron capture rates (Sec. \[sec2A\]), followed by the theoretical aspects, which have to be considered in the step from laboratory measurements to stellar applications (Sec. \[sec2B\]). The problems related to the often dramatic enhancement of $\beta$-decay rates under stellar conditions are discussed in Sec. \[sec2C\]. The status of stellar ($n, \gamma$) rates and further improvements by new experimental facilities and advanced techniques are addressed in Sec. \[sec2D\]. \[sec2A\]Measurement of neutron capture rates --------------------------------------------- ### Pulsed neutron sources \[sec2A1\] The laboratory neutron sources used in nuclear astrophysics measurements cover a variety of facilities, which differ in many aspects. The discussion presented here is focused on the main concepts and does, therefore, not include rarely used options such as filtered reactor beams [@BPS79] and radioactive sources [@Kno79]. At small accelerators, neutrons are produced by nuclear reactions, such as the $^7$Li($p, n$)$^7$Be reaction, with the possibility of tailoring the neutron spectrum exactly to the stellar energy range between 0.3 and $\approx$500 keV. In many cases, the limited source strength can be compensated by low backgrounds and the use of comparably short neutron flight paths [@WiK81; @NIM91; @JaK96a]. Operated in DC mode this type of accelerator can also be used for the simulation of stellar neutron spectra, which are important in applications of the activation technique (see below). Much higher intensities can be achieved via ($\gamma, n$) reactions at electron linear accelerators, such as GELINA at Geel, Belgium, and ORELA at Oak Ridge, USA, by bombarding heavy metal targets with electron beams of typically 50 to 100 MeV. These so-called white neutron sources provide continuous neutron spectra over an energy range from thermal to some tens of MeV. Measurements at these facilities need to be carried out at larger neutron flight paths because of the strong $\gamma$ flash from the impact of the lectron beam. In turn, the longer flight paths provide the possibility to study the resolved resonance region with high resolution (see for example @KSW96). Spallation reactions induced by energetic particle beams constitute the most prolific pulsed sources of fast neutrons suited for time-of-flight (TOF) measurements. Presently, two such spallation sources are in operation, LANSCE at Los Alamos [@LBR90] and the n$\_$TOF facility at CERN [@AAA03]. The main advantage of these facilities is the superb efficiency for neutron production due to the high primary proton beam energies of 800 MeV and 20 GeV at LANSCE and n$\_$TOF, respectively. At n$\_$TOF, for example, 300 neutrons are produced per incident proton, which makes this facility the most luminous white neutron source presently available. Due to their excellent efficiency, spallation sources can be operated at rather low repetition rates while still maintaining high average intensities. The situation at LANSCE is characterized by a comparably short flight path of 20 m, a time resolution of 250 ns, and a repetition rate of 50 Hz, similar to the performance of the SNS at Oak Ridge [@SNS04] and at J-PARC in Japan [@JPC04]. The n$\_$TOF facility at CERN represents a complementary approach aiming at higher resolution (185 m flight path, 7 ns pulse width) and even lower repetition rates of typically 0.4 Hz [@AAA03]. The astrophysics options at various white neutron sources have been compared by @Koe01b with respect to measurements on radioactive samples. As expected, spallation sources are unique for their superior peak neutron fluxes in the astrophysically relevant keV region. However, only the n$\_$TOF facility exhibits a neutron energy resolution comparable to that of electron linear accelerators. ### Time-of-flight methods \[sec2A2\] The aim of the energy-differential TOF methods is to measure the neutron capture cross sections over a sufficiently large neutron energy range that Maxwellian averaged cross sections (MACS) can be determined from these data for any stellar temperature of interest. Recent developments and improvements in pulsed neutron sources and detection techniques have led to ($n, \gamma$) cross section measurements with improved accuracy, in many cases with uncertainties of a few percent. This progress is essential for obtaining the $s$ abundances accurately enough to infer the physical conditions at the stellar site by analysis of the abundance patterns of $s$-process branchings either in solar material or in presolar grains. (i)Total absorption calorimeters The energy sum of the $\gamma$-ray cascade emitted in the decay of the compound nucleus corresponds to the binding energy of the captured neutron. Therefore, this neutron separation energy represents the best signature of a capture event. Hence, $4\pi$ detectors with an efficiency close to 100% are the most direct way to unambiguously identify ($n, \gamma$) reactions and to determine capture cross sections. This calorimetric approach started with the use of large liquid scintillator tanks, which are meanwhile replaced by arrays of BaF$_2$ crystals because of their superior resolution in $\gamma$-ray energy and their correspondingly lower backgrounds. A detector of this type consisting of 42 modules was developed at Karlsruhe [@WGK90b] and is also in use at the n\_TOF facility at CERN [@HRF01]. In this design the BaF$_2$ crystals are shaped as truncated pyramids, forming a fullerene-type geometry where each module covers the same solid angle with respect to the sample. A somewhat simpler approach was chosen at ORNL [@GSK97b] and at FZ Rossendorf [@KAB07], where cylindrical BaF$_2$ arrays have been constructed with hexagonal crystals. Recent examples of accurate cross section measurements made with the Karlsruhe 4$\pi$ detector comprise the unstable branch point isotope $^{151}$Sm [@WVK06c] and the Lu- and Hf-isotopes [@WVK06a; @WVK06b]. These results are essential for constraining the temperature at the $s$-process site via the branchings at $A=151$, 175 and 179. A higher segmentation of such a 4$\pi$ detector is of advantage for separating true capture events from backgrounds and for handling the data rates in measurements on radioactive samples. The state-of-the-art in this respect is the DANCE array with 162 BaF$_2$ modules that is operated at the LANSCE facility in Los Alamos [@RBA04]. The high efficiency of $4\pi$ arrays in combination with intense pulsed neutron sources provides the possibility for measurements on very small samples. In general, this is important for any samples, where only small quantities are available, and in particular for radioactive isotopes, where the background from the activity of the sample needs to be kept at minimum. An illustrative example for the first aspect is the keV ($n, \gamma$) cross section of $^{180}$Ta where the sample consisted of 6.7 mg of $^{180}$Ta immersed in 145 mg of $^{181}$Ta [@WVA04]. Even smaller samples of about 200–400 $\mu$g have recently been used in the DANCE array to determine the keV ($n, \gamma$) cross sections of actinide samples [@JBB08; @ERB08]. A detailed survey for future measurements of neutron capture cross sections on radioactive isotopes with special emphasis on branching points along the $s$-process path is given by @CoR07. The main problem in using 4$\pi$ arrays arises from their response to neutrons scattered in the sample. Although the scintillator is selected to consist of nuclei with small ($n, \gamma$) cross sections, about 10% of the scattered neutrons are captured in the scintillator. The resulting background can be attenuated by an absorber shell around the sample, preferentially consisting of $^6$LiH [@RBA04] or a $^6$Li containing compound [@HRF01]. Such an absorber is not required in the setup at Karlsruhe if the neutron spectrum is limited to energies below 225 keV, which allows one to separate the background from sample scattered neutrons via TOF. This type of background becomes crucial in measurements on neutron magic nuclei and on light isotopes, where neutron scattering dominates the rather weak capture by orders of magnitude. Therefore, large detector arrays are less suited for these isotopes, which are of fundamental importance because they act as bottlenecks in the $s$-process path or as potential neutron poisons. The potential of a 4$\pi$ BaF$_2$ array was extensively used at Karlsruhe for determination of accurate MACS values for an almost complete set of lanthanide isotopes between $^{141}$Pr [@VWA99] and $^{176}$Lu [@WVK06a] including all $s$-only nuclei (for a complete list of references see http://www.kadonis.org). The group of the lanthanide isotopes are particularly suited for a precise test of $s$-process nucleosynthesis concepts, because the relative abundances of the lanthanides are very well known [@AGS09] so that the $s$process reaction chain and the associated branchings can be consistently followed. In fact, the failure of the classical $s$ process and the success of the stellar $s$ process in thermally pulsing low mass AGB stars [@AKW99b] was possible after accurate cross sections for the Nd isotopes and in particular for the $s$-only nucleus $^{142}$Nd became available [@WVK98a] (Sec. \[sec3A\]). Comprehensive measurements were also performed at Karlsruhe for the long isotope chains of Cd [@WVK02], Sn [@WVT96a], Te [@WVK92], and Ba [@VWG94b] to provide detailed information for studying the full mass range of the main $s$-process component with well-defined MACS data. Reliable cross sections are also instrumental for defining the strength of the branchings in the reaction path, where the specific abundance patterns yield constraints for important parameters of the stellar plasma, i.e. neutron density, temperature, pressure, and mixing phenomena. Such examples are the branchings at A=122/123, 128, and 147/148, which represent sensitive tests for the quasi-equilibrium of the $s$-process reaction flow [@WVK92] and for details of the stellar $s$-process conditions, i.e. for the convective velocities [@RKV04] and the neutron density [@WGV93] during He shell flashes in thermally pulsing low-mass AGB stars. Complementary information on the $s$-process temperature can be obtained from the branchings at $^{151}$Sm [@AAA04c; @WVK06c; @MAA06] and at $^{175}$Lu [@WVK06a; @WVK06b]. These important aspects of the $s$-abundance distribution are the subject of Secs. \[sec3C\] and \[sec4C\]. \(ii) Detectors with low neutron sensitivity Originally, the neutron sensitivity problem led to the development of Moxon-Rae type detectors [@MoR63]. The idea was to design a $\gamma$-ray detector with an efficiency proportional to the energy deposited. With this feature, the probability for detecting a capture event becomes $$\label{eq:yield} \varepsilon_{casc} = \sum_{i=1}^{m} \varepsilon_{i} (E_{\gamma}^i) = \sum_{i=1}^{m} k \times E_{\gamma}^i = k\times E_{\gamma}^{tot}$$ independent of the cascade multiplicity $m$ and of the $\gamma$ energies. To avoid systematic uncertainties, the efficiency of Moxon-Rae detectors had to be small enough that no more than one $\gamma$ ray was detected per cascade. In order to improve the overall efficiency, the principle of Moxon-Rae detectors was generalized by introducing the pulse height weighting technique (PHWT) [@MaG67b; @Rau63], where the proportionality between deposited energy and $\gamma$-ray efficiency is achieved a posteriori by an off-line weighting function applied to the detector signals. The PHWT technique was first used in experiments with C$_6$F$_6$ liquid scintillators. Although smaller than for scintillators containing hydrogen, the neutron sensitivity of C$_6$F$_6$ detectors gave still rise to large systematic uncertainties as illustrated by @KWG00 and @GLS05 [@GLS05b]. This problem was reduced in a second generation of detectors, which are based on deuterated benzene (C$_6$D$_6$) because of the smaller capture cross section of deuterium. Further improvement was achieved by minimizing the construction materials and by replacing aluminum and steel by graphite or carbon fiber, resulting in a solution, where the background due to scattered neutrons is practically negligible [@PHK03]. The accuracy of the PHWT has been an issue for a long time. When the technique was proposed, the uncertainties introduced by the weighting function (WF) were about 20% in some particular cases [@Mac87]. Dedicated measurements and Monte Carlo (MC) calculations [@CPL88; @PJG88] have led to gradually improved WFs. With present advanced MC codes, realistic detector response functions and WFs could be determined by means of precise and detailed computer models of the experimental setup [@KSW96; @TGC02; @AAA04a; @BAG05b]. A dedicated set of measurements at the n\_TOF facility confirmed that WFs obtained by such refined simulations allows one to determine neutron capture cross sections with a systematic accuracy of better than 2% [@AAA04a]. Recent applications of the improved PHWT technique with optimized C$_6$D$_6$ detectors are the measurements of ($n, \gamma$) cross sections on isotopes at or near magic neutron numbers, which are characterized by small capture/scattering ratios. Examples for such measurements are the studies of the n\_TOF collaboration on $^{209}$Bi [@DAA06a] and on a sequence of stable Pb isotopes [@DAA06b; @DAA07a; @DAA07b]. Several of the involved resonances show the effect of neutron sensitivity as illustrated in Fig. \[fig:3\] for two cases in the $^{209}$Bi cross section, where the resonance yields obtained from previous data are clearly overestimated. ![\[fig:3\] (Color online) R-matrix analysis of the second resonance in bismuth. The dashed line corresponds to the yield calculated with the resonance parameters from the ENDF/B-VI.8 evaluation which exhibits the effect of neutron sensitivity in previous data.](rmp_fig3.eps){width="8.5cm"} Other neutron magic isotopes, which were recently studied with improved accuracy, are $^{139}$La [@TAA07] and $^{90}$Zr [@TFM08]. The importance of $^{139}$La results from the fact that it is abundantly produced by the $s$ process (e.g. 70% of solar La). Because La is easily detectable by stellar spectroscopy, it can be used as an indicator for the onset of the $s$ process in low-metallicity stars, which were formed early in galactic history. It is interesting to note that the $^{139}$La cross section was accurately confirmed by independent activation measurements [@BDH03; @WDH06]. The position of $^{90}$Zr in the $s$-process reaction chain is exactly located at the matching point between the mass regions dominated by the weak and main $s$ process. Therefore, this isotope assumes a key role among the isotopes with N = 50. As in the Bi and in the Pb isotopes, the new data are significantly smaller than measured previously due to neutron sensitivity problems in the past. The potential of TOF measurements is illustrated in Table \[tab1\] at the example of the important pairs of $s$-only isotopes, which can be used to define branching ratios $$B=\frac{\langle \sigma \rangle_p\,N_p}{\langle \sigma \rangle_f\,N_f}$$ where $\langle \sigma \rangle_p$ denotes the MACS of the partially bypassed isotope and $N_p$ its isotopic abundance. The corresponding values for the heavier isotope, which experiences the full reaction flow, are indicated by index $f$. One such pair are the isotopes $^{80}$Kr and $^{82}$Kr, which characterize the branching at $^{79}$Se as indicated in Fig. \[fig:1\]. Pairs of $s$-only isotopes are particularly valuable because these branchings can be evaluated without interference by the uncertainties related to elemental abundance values. For this reason, the MACSs of these isotopes have been determined with the highest possible accuracy. The uncertainties of the respective branching ratios in the last column of Table \[tab1\] are refer only the MACS uncertainties although isotopic abundance ratios exhibit also non-negligible uncertainties [@RoT98]. For example, a 6% uncertainty has been attributed to the abundance ratio of $^{122}$Te and $^{124}$Te. The quoted values, which include small corrections due to minor deviations from reaction flow equilibrium [@AKW99b], were estimated via the classical approach. Note that weak branchings can only be analyzed with confidence if the MACS are accurately known as in case of the $^{148,150}$Sm pair. ### Data acquisition and analysis techniques \[sec2A3\] Modern electronic techniques have led to substantial improvements in data acquisition and analysis. At the front end, flash analog-to-digital converters (FADCs) provide the fastest and safest way to record the complete information contained in the analogue detector output by digitizing the entire wave-form of the signals. This has the obvious advantage that the raw data are preserved for repeated and refined off-line analysis, which allows for the efficient identification and correction of baseline shifts, pile-up, and noise, resulting in a rigorous assessment of systematic uncertainties. Specifically designed pulse shape analysis (PSA) algorithms can be used for evaluating the relevant signal parameters, i.e. time information, amplitude, area, and shape, for each type of detector. In case that these algorithms are improved at some point, one can always return to the original information and repeat the analysis of a particular experiment. Another advantage is that accidental errors due to the failure of an electronic module are minimized, simply because there is much less electronics needed to run an experiment. It is sufficient, for example, to connect the anode signal of a photomultiplier tube with an FADC-channel, which is then read by a computer. These features were recently used in ($n, \gamma$) studies with 4$\pi$ BaF$_2$ arrays, where flexible algorithms have been employed to reduce background events in the scintillator via $n/\gamma$-identification [@MBB06] and to suppress the intrinsic $\alpha$ background in BaF$_2$ crystals [@RBA04]. A good example is the data acquisition system at n\_TOF [@AAA05b], which is based on 8-bit FADC modules, with sampling rates up to 2 GHz and 8 or 16 Mbyte memory. The low repetition rate of $\sim$0.4 Hz leaves enough time to digitize and store all the raw FADC information accumulated during each neutron bunch and for all detectors employed. Although peak rates of 8 Mbyte are reached per burst and per detector due to the very high instantaneous neutron flux, the data acquisition system works practically dead-time free, except for a narrow interval of 15-20 ns that is needed to separate two consecutive signals unambiguously. The use of FADCs in TOF measurements at facilities with repetition rates above about 1 kHz is hampered by rapidly increasing dead times caused by the transfer of the large amount of data to the storage medium, a problem that is presently studied by the EFNUDAT collaboration [@Pla09]. The main drawback of a FADC-based acquisition system is the huge amount of accumulated data, which demands large storage capabilities and high data transfer rates. While this difficulty can be mitigated by applying a zero suppression algorithm on the fly [@AAA05b], the enormous improvement in computing power and in the capacity of storage media was essential for the handling and analysis of Terabytes of data taken with FADC systems. Another general aspect of data analysis techniques is the increasing importance of Monte Carlo simulations, which are becoming standard tools for planning of measurements and for analyzing experimental data. The efficient application of the GEANT [@Gea03]and MCNP [@MCN07] software packages has also been favored by the recent advances in computing power. ### Activations \[sec2A4\] Activation in a quasi-stellar neutron spectrum provides a completely different approach for the determination of stellar ($n, \gamma$) rates. Apart from the fact that the method is restricted to cases, where neutron capture produces an unstable nucleus, it has a number of appealing features. - It was found that stellar neutron spectra can be very well approximated in the laboratory so that MACS measurements can be directly performed by irradiation and subsequent determination of the induced activity. - Technically, the method is comparably simple and can be performed at small electrostatic accelerators with standard equipment for $\gamma$ spectroscopy. - The sensitivity is orders of magnitude better than for TOF experiments because the accelerator can be operated in DC mode and because the sample can be placed directly at the neutron production target in the highest possible neutron flux. This feature opens the possibility for measurements on sub-$\mu$g samples and on rare isotopes, an important advantage if one deals with radioactive materials. - In most cases the induced activity can be measured via the $\gamma$ decay of the product nucleus. This implies favorable signal/background ratios and unambiguous identification of the reaction products. The excellent selectivity achieved in this way can often be used to study more than one reaction in a single irradiation, either by using elemental samples of natural composition or suited chemical compounds. In an astrophysical environment with temperature $T$, interacting particles are quickly thermalized by collisions in the stellar plasma. The neutron energy distribution corresponds to a Maxwell-Boltzmann spectrum, $$\Phi = dN/dE_n \sim \sqrt{E_n}{\rm exp}(-E_n/kT).$$ So far experimental neutron spectra, which simulate the energy dependence of the flux $v \Phi \sim E_n {\rm exp}(-E_n/kT)$, have been produced by three reactions. The $^7$Li($p, n$)$^7$Be reaction allows one to simulate the spectrum for a thermal energy of $kT$ = 25 keV [@BeK80; @RaK88] very close to the 23 keV effective thermal energy in He shell flashes of low mass AGB stars, where neutrons are produced via the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction. More recently, the $^{18}$O($p, n$)$^{18}$F reaction has been shown to provide a spectrum for $kT=5$ keV [@HDJ05], which is well suited for $s$-process studies of the main neutron source in these stars, the $^{13}$C($\alpha, n$)$^{16}$O reaction that operates at 8 keV thermal energy. While these two quasi-stellar spectra allow one to determine the MACSs necessary for studies of the main $s$ component, the weak component associated with massive stars is characterized by higher temperatures, i.e. 26 keV thermal energy during core He burning and about 90 keV in the shell C burning phase. The situation during core He burning is again well described by the $^7$Li($p, n$)$^7$Be reaction, but the high temperatures during shell C burning are only roughly represented by means of the $^3$H($p, n$)$^3$He reaction, which provides a spectrum for $kT$ = 52 keV [@KNA87]. In this case, the measured MACSs have to be extrapolated by statistical model calculations. Because the proton energies for producing these quasi-stellar spectra are only slightly higher than the reaction thresholds, all neutrons are emitted in forward direction. In this way, the samples are exposed to the full spectrum and backgrounds from scattered neutrons are negligible. With a proton beam current of 100 $\mu$A on target intensities of the order of 10$^9$, 10$^8$, and 10$^5$ s$^{-1}$ can be achieved for the ($p, n$) reactions on $^7$Li, $^3$H, and $^{18}$O with present electrostatic accelerators. Future developments, however, will provide much higher beam currents and correspondingly higher neutron fluxes (Sec. \[sec2A4\]). Already at present, the neutron intensities for activation measurements exceed the fluxes obtainable in TOF measurements by orders of magnitude. For example, the highest neutron flux reached at an experimental TOF setup is 5$\times$10$^5$ s$^{-1}$ at the DANCE array in Los Alamos. Accordingly, activation represents the most sensitive method for ($n, \gamma$) measurements in the astrophysically relevant energy range. This feature is unique for the possibility to measure the MACSs of neutron poisons, abundant light isotopes with very small cross sections, as well as for the use of extremely small sample masses. The latter aspect is most important for the determination of MACSs of unstable isotopes, which are needed for investigating unstable nuclei of relevance for $s$-process branchings. In most cases TOF measurements on unstable branch point isotopes are challenged by the background due to the sample activity (Sec. \[sec2A4\]). Illustrative examples in this respect are the successful measurements of the MACS of $^{60}$Fe [@URS09] and $^{147}$Pm [@RAH03]. In the first case, the sample [@SND10] consisted of $1.4\times 10^{16}$ atoms or 1.4 $\mu$g and the activation was complicated by the 6 min half life of the $^{61}$Fe, which required 47 repeated irradiations, and by the small capture cross section of 5.7 mb. Note that the number of atoms and the cross section result reduced by a factor 1.75 compared to the original paper [@URS09] because of a new precise half-life determination for $^{60}$Fe [@RFK09]. The second experiment was performed with an even smaller sample of only 28 ng or $1.1\times10^{14}$ atoms in order to keep the $^{147}$Pm activity ($t_{1/2}=2.6$ yr) at a reasonable value. In this case, the small sample mass could be used because of the half live of $^{148}$Pm and the cross section were conveniently large. Another advantage of the activation method is that it is insensitive to the reaction mechanism. In particular, it includes the contributions from Direct Capture (DC), where the neutron is captured directly into a bound state. The DC component, which contributes substantially to the ($n, \gamma$) cross sections of light nuclei, is extremely difficult to determine in TOF measurements (for an exception see the specialized setup used by @INM95). Apart from measurements on unstable isotopes as discussed below, the excellent sensitivity of the activation technique has been extensively used for the determination of cross section with a non-negligible DC component, for the determination of partial cross sections for the population of isomeric states, for the measurement of small cross sections in general. Prominent examples of the latter type are the series of measurements between Fe and Sr, which are related to the reaction flow of the weak component [@RDF07; @HKU08a; @HKU08b; @MDD09a]. These data were consistently smaller than previous TOF results, which evidently suffered from an underestimated neutron sensitivity. Fig. \[fig:4\] shows that these changes gave rise to strong propagation effects in the abundance distribution. These effects are to a large part originating from the high-temperature phase during shell C burning that operates at $kT=90$ keV. In order to reduce the uncertainties in the extrapolation from the measured MACS at 25 keV complementary TOF measurements on the stable Fe and Ni isotopes are under way at CERN. ![\[fig:4\] Nucleosynthesis yields between Fe and Nb illustrating the final $s$-process yields after shell C burning for a 25 $M_\odot$ star with solar metallicity. To illustrate the combined effect of new cross sections for $^{58}$Fe, $^{59}$Co, $^{64}$Ni, $^{63,65}$Cu the yield is plotted relative to the case obtained with the previous cross sections of @BBK00. Even and odd Z elements are distinguished by black and open symbols, respectively. The thin lines demonstrate the uncertainties stemming from the extrapolation of the measured cross sections to higher and lower energies. ](rmp_fig4.eps){width="8.5cm"} For some branchings the population of long-lived isomers plays an important role [@War77]. In most cases the respective partial cross sections feeding the isomers have to be determined in activation measurements because this information is very difficult to obtain via the TOF technique. Important isomers are those in $^{176}$Lu and in $^{180}$Ta for example. The isomer in $^{176}$Lu at 123 keV is the key for the interpretation of the mother/daughter ratio of the $s$-only isotopes $^{176}$Lu and $^{176}$Hf as an $s$-process thermometer for the He shell flashes in low mass AGB stars [@KKB91b; @DBJ99]. At temperatures above about 150 MK the initial population probabilities of isomer ($t_{1/2}=3.68$ h) and ground state ($t_{1/2}=37.5$ Gyr) [@WVK06a; @HWD08] are altered by a delicate interplay between neutron density and temperature, depending on subtle details in the nuclear structure of $^{176}$Lu [@MBG09]. The case of $^{180}$Ta is of interest because this is the only isotope in nature, which is (almost) stable in its isomeric state. Again, the survival of $^{180}$Ta$^{\rm m}$ depends on the ($n, \gamma$) cross section [@WVA01; @WVA04; @KAH04] and on the effect of the high temperatures at the $s$-process site [@BAB99; @BAB02] in depopulating the isomer to the ground state. A few years ago, the potential of the activation technique was considerably extended where activation products are counted directly by Accelerator Mass Spectrometry (AMS) instead of the induced activity (Sec. \[sec2D2\]). The complementarity between TOF and activation measurements is illustrated in Table \[tab2\] for examples between Fe and Au. Overall, there is good agreement between the results obtained with both methods. As far as the related uncertainties are concerned one could be tempted to assume that the small, resonance-dominated cross sections of the Fe group and at magic neutron numbers are more accurately determined by activation, whereas the smooth cross sections of the heavier nuclei are generally better determined in TOF experiments. However, this is no general rule as indicated by the cases of $^{88}$Sr and $^{197}$Au, but the quality of the data depends on individual details of the respective experiments. ### Studies on radioactive isotopes \[sec2A5\] Although many cross sections of the stable isotopes are still rather uncertain and need to be improved, a major challenge of future experiments is to extend such measurements to the largely unexplored subset of unstable branch point nuclei. The concept of $s$-process branchings was formulated by @WNC76 to obtain information on [*average*]{} $s$-process neutron densities and temperatures. With the advent of quantitative stellar models the abundance patterns of the branchings were understood to represent sensitive tests for the time-dependence of these parameters during the various $s$-process episodes in thermally pulsing low mass AGB stars [@GAB98; @AKW99b] and in massive stars [@RBG91b]. Branchings in the $s$ path occur whenever $\lambda_n \approx \lambda_\beta$. The $\beta$ decay rate is $\lambda_\beta=ln2/t_{1/2}$, whereas the neutron capture rate $\lambda_n = n_n\,\langle\sigma\,v\rangle\,v_T$ is the product of neutron density, MACS, and mean thermal velocity. Depending on neutron density and MACS of the unstable branch point isotope, the stellar half lives of branch point nuclei may range between several days and several years. In addition, stellar half lives can be strongly enhanced compared to the corresponding terrestrial values as discussed in Sec. \[sec2C\]. The branch point isotope $^{79}$Se in Fig. \[fig:1\] represents such a case, where the decay is accelerated by thermal population of a short-lived excited state. In contrast, the other two branch points indicated in Fig. \[fig:1\], $^{63}$Ni and $^{85}$Kr, are not or only weakly altered by temperature effects. All three branch points fall in the mass range of the weak component associated with massive stars. Measurements of the MACS for the branch points of the weak component are hampered by the high specific activity of $^{85}$Kr and by the lack of sample material in case of $^{79}$Se. Only for $^{63}$Ni a TOF measurement has been recently performed with the DANCE detector at Los Alamos using a mildly enriched nickel sample containing 11% $^{63}$Ni [@Cou09]. For the main component experimental information could be obtained for some important branch point isotopes. With one exception these measurements were performed via the activation technique, where the activity problem was relaxed because the high sensitivity of the method allows one to use $\mu$g or even sub-$\mu$g samples. Results were reported for $^{135}$Cs [@PAK04], $^{147}$Pm [@RAH03], $^{163}$Ho [@JaK96a], $^{155}$Eu [@JaK95b], and $^{182}$Hf [@VDH07]. As noted before the data obtained via activation usually represent the respective MACS values at $kT=25$ keV and have to be extrapolated to higher and lower temperatures by means of theoretical data (Sec. \[sec2B\]). TOF measurements on branch point isotopes of the main component have been reported for the quasi-stable nuclei $^{99}$Tc [@Mac82b; @WiM87], $^{107}$Pd [@Mac85a], and $^{129}$I [@Mac83]. So far $^{151}$Sm is the only branch point with a half-life shorter than 100 yr where the ($n, \gamma$) cross section has been studied by means of the TOF technique over a wide energy range. Combination of the accurate and comprehensive data measured at Karlsruhe [@WVK06c] and CERN [@AAA04c] provided a full set of MACS values for $^{151}$Sm. The measurement of the $^{14}$C($n, \gamma$)$^{15}$C cross section [@RHF08] was of interest because this reaction determines whether the neutron balance of the $s$ process could be affected by neutron induced CNO cycles [@WGS99], it contributes to the reaction flow in neutrino driven wind scenarios of the $r$ process [@TSK01], and is important for validating the ($n, \gamma$) cross sections calculated via detailed balance from the inverse Coulomb dissociation reaction [@WGT90; @TBD06]. The possibility to complement ($n, \gamma$) experiments by studies of the inverse ($\gamma, n$) reactions has been invoked by @SMV03 and @MSU04 for the branch point isotope $^{185}$W. Other examples of ($\gamma, n$) measurements refer to applications in the $p$ rather than in the $s$ process [@VMB01; @SVG04; @SVG05]. Another indirect approach for obtaining information on ($n, \gamma$) cross sections of unstable nuclei is the surrogate method [@DiE07; @EAB05], which uses the assumption that the reaction of interest proceeds via the formation of a compound nucleus and that formation and decay of the compound state can be separated, provided that both steps are independent of each other. In many cases, the formation cross section can be calculated reasonably well by using optical potentials, but theoretical decay probabilities are often quite uncertain. In the surrogate approach the compound nucleus is produced via an alternative direct reaction and its decay probability is then measured. There are several challenges of the surrogate method, in particular the “J population mismatch”, which means that in the desired reaction different compound states might be populated, the difficulty to convert the experimental observables into decay probabilities, the role of pre-equilibrium reactions, where the intermediate configuration decays before a compound nucleus is formed, and the role of projectile break-up, which may disturb the proper identification of the surrogate reaction [@FAB05]. A recent example for the application of this method is the work of @BDW06. Although the indirect approaches rely in essential parts on theory and are therefore limited in accuracy, they often provide valuable information, which is impossible to obtain otherwise. \[sec2B\]Cross section calculations ----------------------------------- As already pointed out in the previous sections, measurements cannot be performed at all energies and for all relevant isotopes. In addition, the reaction rates in a stellar environment require estimation of reaction processes for nuclei in their excited states, which are impossible to measure under laboratory conditions. Therefore, a close collaboration between experiment and theory remains crucial for establishing the complete nuclear physics input for $s$-process studies. On the other hand, experimental information is also mandatory for guiding and testing developments in theory in the region of unstable nuclei, a necessary step towards quantitative models of explosive nucleosynthesis. ### Statistical model \[sec2B1\] The key approach for the calculation of stellar $s$-process reaction rates is based on the Hauser-Feshbach statistical model (HFSM), which has been formulated over 30 years ago [@Mol75]. The model relies essentially on two basic assumptions, the validity of the compound nucleus reaction mechanism and a statistical distribution of nuclear excited states. With these assumptions, the reaction cross section (e.g. for neutron capture) can be written in terms of model parameters such as the energy-dependent neutron transmission functions $T_{n,ls}$ and the $\gamma$-ray transmission functions $T_{\gamma,J}$. The general expression reads $$\sigma_{n,\gamma}(E_{n}) = \frac{\pi}{k_{n}^2} \sum_{J,\pi} g_{J} \frac{\sum_{ls}T_{n,ls} T_{\gamma,J}}{\sum_{ls} T_{n,ls} + \sum_{ls} T_{n',ls} + T_{\gamma,J}} W_{\gamma,J} \label{HFSM}$$ where $E_{n}$ is the incident neutron energy, $k_{n}$ the wave number, $s=1/2$ the intrinsic spin of the incident particle, and $l$ the orbital angular momentum of neutron and nucleus. The $g_{J} = (2J+1)(2s+1)^{-1}(2I+1)^{-1}$ is a statistical weight factor for target nuclei of spin $I$ and compound states of total angular momentum $J$ compatible with spin and parity conservation laws. The width fluctuation factor $W_{\gamma}$ takes the different statistical properties of the $\gamma$-decay channel and of the competing neutron elastic ($n, n$) and inelastic ($n, n'$) channels into account. The various HFSM approaches differ by the particular nuclear structure and de-excitation models adopted for calculating the nuclear quantities in Eq. \[HFSM\]. Examples of widely used HFSM approaches for applications in nuclear astrophysics are those of @HWF76, @Har81, and the latest “NON-SMOKER” [@Rau01], “MOST” [@Gor05], and “TALYS” [@KHD05] versions. Most of the quoted references include also HFSM computer codes for calculation of reaction cross sections. A repository of parameters and systematics of nuclear structure quantities can be found in the “RIPL” initiative [@Bel05]. Additional model codes have been used for individual reaction rate calculations. ### Maxwellian averaged cross sections \[sec2B2\] The neutron spectrum typical of the various $s$-process sites is described by a Maxwell-Boltzmann distribution, because neutrons are quickly thermalized in the dense stellar plasma. The effective stellar reaction cross sections are therefore obtained by averaging the experimental data over that spectrum. The resulting Maxwellian averaged cross sections (MACS) $$\label{eq:macs} \langle \sigma \rangle_{kT}=\frac{2}{\sqrt{\pi}}\frac{\int_0^\infty \sigma(E_n)~E_n~{\rm e}^{-E_n/kT}~dE_n}{\int_0^\infty E_n~{\rm e}^{-E_n/kT}~dE_n}.$$ are commonly compared for a thermal energy of $kT=30$ keV, but for realistic $s$-process scenarios a range of thermal energies has to be considered, from about 8 keV in the $^{13}$C pocket of thermally pulsing low mass AGB stars to about 90 keV during carbon shell burning in massive stars. To cover this full range, energy-differential cross sections $\sigma(E_n)$ are needed in the energy region $0.1 \leq E_n \leq 500$ keV. Whenever experimental data are available only for part of this range, cross section calculations are required for filling these gaps. Correspondingly, calculated cross sections are needed to extrapolate the results from activation measurements to the entire temperature range of the various $s$-process sites (see Ref. [@HKU08a]). As a service to the community a complete set of MACS data is available from the continuously updated compilation described in Sec. \[sec2D1\]. ### Stellar enhancement factors \[sec2B3\] Apart from the need of cross section calculations for filling the gaps in experimental data, theory is indispensable for adapting the experimental data to the stellar environment. With respect to MACSs, this refers to the fact that excited nuclear states are populated under stellar conditions due to interactions with the hot thermal photon bath. Because of the high photon intensity, all states with excitation energies $E_i$ are in thermal equilibrium with population probabilities $$p_i = \frac{(2J_i + 1)\,{\rm exp}(-E_i/kT)}{\sum_m (2J_m + 1)\,{\rm exp}(-E_m/kT)}$$ where $J$ denotes the level spin and the denominator represents the nuclear partition function. The capture cross section of excited states can be modeled as for ground states. However, in reactions on excited states an additional possibility for inelastic scattering must be considered, the so-called “super-elastic” channel, in which the incident neutron gains in energy, leaving the target nucleus in a lower state. This particularly relevant process has to be taken into account in the HFSM equation (Eq. \[HFSM\]) by adding the transmission coefficients $T_{n',ls}$ for the open super-elastic channels. Possibilities for testing the calculations of the stellar MACS are the comparison of measured and calculated inelastic scattering cross sections, which provides a good benchmark for the neutron-nucleus interaction needed to obtain the transmission functions in such calculations and the comparison with the experimental capture cross section for the ground state. Unfortunately, experimental data for the inelastic channel in the astrophysically relevant energy range are rather scarce. In practice, these effects are taken into account by the so-called stellar enhancement factor $$SEF = \frac{\langle\sigma\rangle^{*}}{\langle\sigma\rangle^{lab}}$$ where the MACS labeled by $*$ and $lab$ indicate the stellar average over the thermally populated states and for the laboratory (ground state) cross section, respectively. On average, these factors are below 10% at the $s$-process temperatures in low mass AGB stars, but can reach values of more than 40% during shell C burning in massive stars, especially for the heavy odd isotopes with low lying excited states. An important example for the role of SEFs is the Re/Os nuclear cosmochronology [@Cla64], where the MACSs of the $s$-only isotopes $^{186}$Os and $^{187}$Os are of key importance. The aspects related to the neutron physics of this clock has been studied by the n\_TOF collaboration in a series of three papers dealing with cross section measurements in the astrophysically relevant energy range [@MFM10; @MHK10] and the SEF calculations based on the HFSM approach with model parameters tuned to reproduce the experimental ($n, \gamma$) and ($n, n'$) cross sections [@FMM10]. The SEF corrections are particularly relevant for $^{187}$Os, where the ground state is populated by only about 30% at $kT=30$ keV, while 70% of the nuclei exist in excited states, 47% alone in the first excited state at 9.75 keV, the state which strongly dominates the competition by inelastic and superelastic scattering. The comparison of the SEF values for $^{186}$Os and $^{187}$Os in Fig. \[fig:5\] underlines the importance of this correction for $^{187}$Os in the relevant range of thermal energies around $kT=25$ keV. A relatively small uncertainty of $\pm$4% could be estimated for the SEF of $^{187}$Os from the difference between the results obtained with a spherical and a deformed optical model potential for the neutron-nucleus interaction. ![\[fig:5\] (Color online) Stellar enhancement factor (SEF) for the $^{187}$Os($n, \gamma$)$^{188}$Os reaction. The first ten excited states in $^{187}$Os, up to an excitation energy of 260 keV, have been included in the calculation. The difference between the results obtained with a spherical and a deformed optical model potential for the neutron-nucleus interaction could be used for estimating the uncertainties of the SEF results.](rmp_fig5.eps){width="8.5cm"} The impact of the estimated SEF uncertainties and of other relevant nuclear physics input on the galactic age (or on the duration of nucleosynthesis) have been estimated by means of the simple model proposed by @FoH60 to be less than 1 Gyr [@FMM10]. This means that uncertainties of the experimental data are no longer limiting a revision of the Re/Os chronometer. Further work can now concentrate of the related astrophysical issues, which are mostly due to the time-dependence of the production rate of $^{187}$Re and the related problem of astration, i.e. that $^{187}$Re is partially destroyed in later stellar generations. The example of $^{187}$Re shows that uncertainties of $\approx$4 - 5% can be obtained in SEF calculations, provided that the measured capture cross section for the ground-state is known with sufficient accuracy and that the other parameters in HFSM calculations can be derived from measured quantities. \[sec2C\]Beta decay under stellar conditions -------------------------------------------- A detailed evaluation of beta decay rates for $s$-process analyses has been presented by @TaY87 on the basis of a thorough classification of possible contributions from thermally excited states and by considering the relevant effects related to the high degree of ionization in the stellar plasma. The most spectacular consequence of ionization is the enormous enhancement of decays with small Q$_{\beta}$ values, where the decay electrons can be emitted into unoccupied atomic orbits. This bound beta decay was eventually confirmed in storage ring experiments with fully stripped $^{163}$Ho and $^{187}$Re atoms at GSI Darmstadt [@JBB92; @BFF96]. The quantitative assessment of the temperature dependent decay rates of the key branch point isotopes requires more experimental information on log$_{10}$ ft values for the decay of excited states as well as more storage ring experiments to expand our knowledge of bound beta decay rates. Experimental possibilities have been discussed [@Kae99], but must be extended by the successful recent application of ($d, ^2$He) reactions [@Fre04]. \[sec2D\] Status and Prospects ------------------------------ ### Compilations of stellar ($n, \gamma$) cross sections and further requirements \[sec2D1\] Stellar neutron capture cross sections have first been compiled in 1971 by @AGM71, who presented a set of recommended ($n, \gamma$) cross sections averaged over a Maxwell-Boltzmann distribution for a thermal energy of $kT$ = 30 keV. This first collection of Maxwellian averaged cross sections (MACS) comprised already 130 experimental cross sections with typical uncertainties between 10 and 25%. These data were complemented by 109 semi-empirical values estimated from the cross section trends with neutron number of neighboring nuclei to provide a full set of nuclear data for quantitative studies of the $s$-process. The next compilation of experimental and theoretical stellar neutron cross sections for $s$-process studies, which was published 16 years later by @BaK87, included cross sections for ($n, \gamma$) reactions between $^{12}$C and $^{209}$Bi, some ($n, p$) and ($n, \alpha$) reactions (from $^{33}$Se to $^{59}$Ni), and also ($n, \gamma$) and ($n, f$) reactions for long-lived actinides. Also in this version MACSs were given at a single thermal energy of $kT = 30$ keV, sufficient for studies with the canonical $s$-process formulated by @SFC65 for a constant temperature and neutron density scenario. A major achievement, however, was the significant improvement of the accuracy, which was reaching the 1 - 2% level for a number of important $s$-process isotopes. Meanwhile, the canonical or “classical” approach had been challenged by refined stellar models, which indicated different sites for the $s$-process, from He shell burning in thermally pulsing low mass AGB stars [@GBP88; @HoI88] to shell C burning in massive stars [@RBG91a; @RBG91b], where ($\alpha, n$) reactions on $^{13}$C and $^{22}$Ne were identified as the dominant neutron sources, respectively. The fact that the temperatures at the various sites require MACS data for thermal energies between 8 and 90 keV was taken into account in the compilation of @BVW92, which listed values in the range $5 \leq kT \leq 100$ keV. The following compilation of @BBK00 was extended to cover a network of 364 ($n, \gamma$) reactions, including also relevant partial cross sections. This work presents detailed information on previous MACS results, which were eventually condensed into recommended values. Again, data are given for thermal energies from 5 to 100 keV. For isotopes without experimental cross section information, recommended values were derived from calculations with the Hauser-Feshbach statistical model code [NON-SMOKER]{} [@RaT00], which were empirically corrected for known systematic deficiencies in the nuclear input of the calculation. For the first time, stellar enhancement factors (SEF), which take the effect of thermally excited nuclear states into account, were included as well. For easy access, the compilation of @BBK00 was published in electronic form via the <span style="font-variant:small-caps;">KADoNiS</span> project (http://www.kadonis.org) [@DHK05]. The current version <span style="font-variant:small-caps;">KADoNiS</span> v0.3 [@DPK09] is already the third update and includes – compared to the Bao [*et al.*]{} compilation [@BBK00] – recommended values for 38 improved and 14 new cross sections. In total, data sets are available for 356 isotopes, including 77 radioactive nuclei on or close to the $s$-process path. For 13 of these radioactive nuclei, experimental data is available, i.e. for $^{14}$C, $^{60}$Fe, $^{93}$Zr, $^{99}$Tc, $^{107}$Pd, $^{129}$I, $^{135}$Cs, $^{147}$Pm, $^{151}$Sm, $^{155}$Eu, $^{163}$Ho, $^{182}$Hf, and $^{185}$W. The remaining 64 radioactive nuclei are not (yet) measured in the stellar energy range and are represented only by empirically corrected Hauser-Feshbach rates with typical uncertainties of 25 to 30%. Almost all of the ($n, \gamma$) cross sections of the 277 stable isotopes have been measured. The few exceptions are $^{17}$O, $^{36,38}$Ar, $^{40}$K, $^{50}$V, $^{70}$Zn, $^{72,73}$Ge, $^{77,82}$Se, $^{98,99}$Ru, $^{131}$Xe, $^{138}$La, $^{158}$Dy, and $^{195}$Pt, which lie mostly outside the $s$-process path in the proton-rich $p$-process domain. These cross sections are difficult to determine because they are often not accessible by activation measurements or not available in sufficient amounts and/or enrichment for time-of-flight measurements. The actual status of the ($n, \gamma$) cross sections for $s$-process nucleosynthesis calculations is summarized in Fig. \[fig:6\], which shows the respective uncertainties as a function of mass number. Though the necessary accuracy of 1 to 5% has been locally achieved, further improvements are clearly required, predominantly in the mass region below $A$ = 120 and above $A$ = 180. Further efforts in this field are the more important as Fig. \[fig:6\] reflects only the situation for a thermal energy of 30 keV. In most cases, however, extrapolation to lower and higher temperatures implies still larger uncertainties. The lack of accurate data is particularly crucial for the weak $s$-process in massive stars, which is responsible for most of the $s$ abundances between Cu and Sr. Since the neutron exposure of the the weak $s$ process is not sufficient for achieving flow equilibrium, cross section uncertainties may affect the abundances of a sequence of heavier isotopes (see Sec. \[sec3B\]). ![\[fig:6\] Uncertainties of the stellar ($n, \gamma$) cross sections for $s$-process nucleosynthesis. These values refer to a thermal energy of $kT$ = 30 keV, but may be considerably larger at lower and higher temperatures.](rmp_fig6.eps){width="12cm"} The present version of <span style="font-variant:small-caps;">KADoNiS</span> consists of two parts: the $s$-process library and a collection of available experimental $p$-process reactions. The $s$-process library will be complemented in the near future by some ($n, p$) and ($n, \alpha$) cross sections measured at $kT$ = 30 keV, as it was already included in [@BaK87]. The $p$-process database will be a collection of all available charged-particle reactions measured within or close to the Gamow window of the $p$ process ($T_9$= 2-3 GK). In a further extension of <span style="font-variant:small-caps;">KADoNiS</span> it is planned to include more radioactive isotopes, which are relevant for $s$-process nucleosynthesis at higher neutron densities (up to 10$^{11}$ cm$^{-3})$ [@CGS06]. Since these isotopes are more than one atomic mass unit away from the “regular” $s$-process path on the neutron-rich side of stability, their stellar ($n, \gamma$) values have to be extrapolated from known cross sections by means of the statistical Hauser-Feshbach model. The present list covers 73 new isotopes and is available on the <span style="font-variant:small-caps;">KADoNiS</span> homepage. ### Measurements on rare and unstable samples \[sec2D2\] The continuous development and optimization of techniques and facilities remains a most vital aspect of the field, especially with respect to unstable isotopes. This concerns the production of higher fluxes, i.e. by means of shorter flight paths at existing spallation sources, e.g. a 20 m station at n\_TOF, or beam lines for keV neutrons at the new generation of high intensity accelerators such as the japanese proton accelerator complex J-PARC [@JPC04]. A completely new approach is presently developed at the University of Frankfurt [@RCM07]. The Frankfurt Neutron source at the Stern-Gerlach-Zentrum (FRANZ) will provide short neutron pulses by bombardment of a $^7$Li target with an intense proton beam. The proton energy range is limited to 2.0$\pm$0.2 MeV and the pulse rate will be typically 250 kHz. The scheme of the accelerator is sketched in Fig. \[fig:7\], starting with a volume-type proton source on a 150 kV high voltage platform, followed by a 100 ns chopper in the low energy beam transport line to the acceleration stage consisting of an RFQ with an exit energy of 700 keV and IH-DTL structure with an effective energy gain of 1.4 MeV. The macro pulses at the exit contain up to 10 rf bunches, which are compressed into 1 ns pulses for neutron production in the Li target. This is achieved by a second chopper, which deflects the bunches to traces of different path length in a Mobley-type buncher system [@CMR06; @MCM06; @RCM07]. ![\[fig:7\] (Color online) Schematic layout and main parameters of the Frankfurt intense neutron source (see text).](rmp_fig7.eps){width="12cm"} In this way, intense pulses of up to $5 \times 10^{10}$ protons can be focused onto the Li target within 1 ns. The average beam current is expected to reach 2 mA, corresponding to a neutron flux of about 10$^5$ n/cm$^{2}$/s/keV at a distance of 80 cm from the target, more than an order of magnitude higher than is achieved at present spallation sources. Since the amount of sample material can be reduced by the gain in flux, TOF measurements at FRANZ appear to be feasible with samples of 10$^{14}$ to 10$^{16}$ atoms. The use of such minute amounts represents a break-through with respect to the production of unstable samples, because beam intensities of the order of 10$^{10}$ to 10$^{12}$ s$^{-1}$ are expected at future Rare Isotope Facilities such as RIKEN [@Tan98], @Fai07, or @Fri10. The gain in beam intensity and the related reduction in sample mass implies that TOF measurements can be carried out on unstable nuclei with correspondingly higher specific activities, including a number of branch point isotopes, which are not accessible by activation techniques. The concise list of important branch point isotopes in Table \[tab3\] shows that most of these samples are accessible to TOF measurements at FRANZ thanks to the excellent sensitivity of that facility. It should be noted, however, that Table \[tab3\] includes a number of cases, which could, in principle, be studied at existing facilities. Such measurements are impeded, however, because isotopically enriched samples are not available in sufficient amounts. Among the 20 listed unstable isotopes in Table \[tab3\] there are only three cases, where the specific $\gamma$ activities seem to be too high for a promising experiment in the foreseeable future. In addition to the gain in TOF sensitivity, FRANZ is also perfectly suited for the simulation of stellar neutron spectra via $^7$Li($p, n$)$^7$Be reaction (Sec. \[sec2A4\]). In fact the larger dispersion in proton energy will result in a closer approximation of the stellar spectrum as was pointed out by @MMP09. Similar to the situation sketched for TOF measurements, the higher flux will enable one to apply the activation method to a largely extended set of shorter-lived unstable nuclei, an important aspect for investigating the $s$-process during shell C-burning in massive stars, where the reaction path is shifted by a few mass units from the valley of stability due to neutron densities in excess of $10^{12}$ cm$^{-3}$ [e.g. @PGH10]. In combination with AMS it will even be possible to use the double neutron capture method to obtain the MACS for such crucial cases as $^{59}$Fe, $^{125}$Sn, and $^{181}$Hf. Apart from their importance for specific problems and their direct use in $s$-process networks, MACSs of unstable isotopes represent also most valuable information in a wider sense, particularly for testing and improving statistical model calculations in areas, which are not accessible to experiments in the near future. $s$-Process Models \[sec3\] =========================== Stellar models for the He burning stage of stellar evolution have been worked out in great detail over the past decade, both for low mass stars in the AGB phase while suffering recurrent thermal pulses in the He shell, and for massive stars. There, besides central He burning in the convective core, neutrons are released in the subsequent convective shell C burning phase, which involves a large fraction of the final ejected mass in the supernova event. Accounting for a continuous updated network of neutron captures and charged particle reaction rates, the $s$ process taking place in both low-mass stars and massive stars will be discussed. Classical approach \[sec3A\] ---------------------------- Shortly after stellar spectroscopy of the unstable element Tc provided evidence for active neutron capture nucleosynthesis in red giant stars [@Mer52b], the canonical or classical model of the $s$ process was suggested by B$^2$FH [@BBF57]. Although it was argued that the He-burning zones of red giants were the most promising site of the $s$ process, the lack of detailed stellar models led to a phenomenological solution. Within this approach it is empirically assumed that a certain fraction $G$ of the observed $^{56}$Fe abundance was irradiated by an exponential distribution of neutron exposures [@SFC65]. In this case, an analytical solution can be obtained if a possible time dependence of the neutron capture rates, $\lambda_n = n_n \langle\sigma\rangle v_T$ is neglected. In other words, it is assumed that temperature and neutron density, $n_n$, are constant. Then, the product of stellar cross section and resulting $s$ abundance, which characterizes the reaction flow, can be given by $$\langle \sigma \rangle_{(A)} N_{s(A)} = \frac{G \cdot N^{\odot}_{56}}{\tau_0} \prod_{i=56}^{A} (1 + \frac{1}{\tau_0 \langle \sigma \rangle_i})^{-1}.$$ Apart from the two parameters $G$ and $\tau_0$ (which are adjusted by fitting the abundances of the $s$-only nuclei), the only remaining input for this expression are the stellar (n,$\gamma$) cross sections $\langle\sigma\rangle$ [@KBW89; @WIP97]. Given the very schematic nature of this classical approach, it surprisingly provided an excellent description of the $s$-process abundances in the solar system (Fig. \[fig:2\]). One finds that equilibrium in the neutron capture flow was obtained between magic neutron numbers, where the $\langle \sigma \rangle N_{s}$-curve is almost constant. The small cross sections of the neutron magic nuclei around A$\sim$88, 140, and 208 act as bottlenecks for the capture flow, resulting in distinct steps in the $\langle \sigma \rangle N_{s}$-curve. The global parameters, $G$ and $\tau_0$, which determine the overall shape of this curve, represent a first constraint for the stellar $s$-process site with respect to the required seed abundance and total neutron exposure. It is found that 0.04% of the $^{56}$Fe abundance observed in solar system material is a sufficient seed, and that on average about 15 neutrons are captured by each seed nucleus [@KGB90]. These numbers refer to the [ *main*]{} $s$-process component, which dominates the $s$ abundances for A$>$90. The rather steep increase of the $\langle \sigma \rangle N_{s}$-curve below A$=$90 requires an additional component, the [*weak*]{} component. The weak component is not firmly described by the classical analysis, because there are only six $s$-only isotopes below $A=90$ ($^{70}$Ge, $^{76}$Se, $^{80,82}$Kr, $^{86,87}$Sr), which are also partly produced by the main component and possibly even by the $p$ process. Moreover, $^{80}$Kr and $^{86,87}$Sr are affected by branchings in the $s$-process path. The temperature and neutron density determining these branchings are to be treated as free parameters. Accordingly, it is difficult within the classical approach to distinguish between a single component or an exponential distribution of neutron exposures [@Bee86; @BeM89; @KBW89]. For about 40 years the classical model was quite successful in describing the solar $s$-process abundances [@KBW82; @KGB90]. In fact, the empirical $\langle \sigma \rangle N_{s}$-values of the $s$-only isotopes that are not affected by branchings, are reproduced with a mean square deviation of only 3% [@KGB90] as illustrated in Fig. \[fig:2\]. About 20 years ago the development of new experimental techniques led to a set of accurate neutron capture cross sections, which ultimately revealed that the classical $s$ process suffered from inherent inconsistencies. This was convincingly demonstrated at the example of the neutron-magic $s$-only isotope $^{142}$Nd [@AKW99b]. At this point, the classical approach was replaced by a first generation of stellar $s$-process models, where such problems could be successfully avoided [@GAB98; @AKW99b]. Nevertheless, the classical model provides still a useful tool for estimating the $s$ abundances, in particular in mass regions between magic numbers, where the cross sections are large enough that reaction flow equilibrium was actually reached. This is illustrated by the example of the solar $r$-process distribution obtained with the $r$-residual method, $$N_r = N_{\odot} - N_s,$$ which is obtained by the difference between the solar abundances and the corresponding $s$-process yields. The $r$ distributions obtained via the classical approach or via stellar models are quite similar and match very well with the abundances of the $r$-only isotopes as shown in Fig. \[fig:8\]. ![ (Color online) The $r$-process abundances (open squares) obtained via the $r$-process residual method, N$_r$ = N$_{\odot}$ - N$_s$, using the classical and stellar $s$-process model [@AKW99b]. The $r$-only nuclei are represented by full squares. (N$_i$ relative to Si $\equiv$ 10$^6$). \[fig:8\]](rmp_fig8.eps){width="14cm"} The separation of the solar abundance distribution into the $s$ and $r$ components became also important for the interpretation of abundance patterns of ultra-metal-poor stars [@WSG00; @HCB02; @CSB02; @SCL03], which turned out to agree very well with the scaled solar $r$ component for the elements heavier than Ba. Massive stars {#sec3B} ------------- As discussed at the end of the previous Sec. \[sec3A\], the classical analysis of the $s$ process fails to reproduce the cosmic abundances of the $s$ isopotes below A = 90. The same problem is found for stellar models of low-mass AGB stars, as described in detail in Sec. \[sec3C\]. Actually, a complementary weak $s$ process occurs in massive stars ($M > 8 M_\odot$), which explode as supernovae of type II. Slow neutron captures are driven by the reaction $^{22}$Ne($\alpha, n$)$^{25}$Mg during convective core He burning at temperatures around 3 $\times$ 10$^8$ K as well as in the subsequent convective shell C burning at 1 $\times$ 10$^9$ K [@CSA74; @LHT77; @KBW89; @PHN90; @BWK92b; @RGB93]. The available $^{22}$Ne is produced via the reaction sequence $^{14}$N($\alpha, \gamma$)$^{18}$F($\beta^+ \nu$)$^{18}$O($\alpha, \gamma$)$^{22}$Ne, where $^{14}$N derives from the CNO cycle in the previous H-burning phase. Consequently, this weak $s$-process component produced in massive stars is secondary like, decreasing with metallicity. At He exhaustion, not all the $^{22}$Ne is consumed [e.g. @PHN90], and neutron production via $^{22}$Ne($\alpha, n$)$^{25}$Mg continues during shell C-burning by means of the $\alpha$-particles provided by the reaction channel $^{12}$C($^{12}$C, $\alpha$)$^{20}$Ne [@ArT69]. In the ejecta of SN II, the chemical composition of the core up to a mass of   3.5 $M_\odot$ (for a star of 25 $M_\odot$) is modified by explosive nucleosynthesis, which destroys any previous $s$-process signature. However, the ejecta still contain also an important mass fraction of   2.5 $M_\odot$, which preserves the original $s$-process abundances produced by the hydrostatic nucleosynthesis phases of the presupernova evolution. This scenario was confirmed by post-process models and full stellar models describing the evolution of massive stars up to the final burning phases and the SN explosion [@RGB93; @WoW95; @LSC00; @WHW02; @TEM07; @ETM09; @PGH10]. In contrast to the main $s$-process component, the neutron fluence in the weak $s$ process is too low for achieving reaction flow equilibrium. This has the important consequence that a particular MACS does not only determine the abundance of the respective isotope, but affects also the abundances of all heavier isotopes as well [@PGH10]. This propagation effect is particularly critical for the abundant isotopes near the iron seed. A prominent example is the case of the $^{62}$Ni($n, \gamma$)$^{63}$Ni cross section, where the effect was discussed first [@RHH02; @RaG02; @RaG05]. This problem has triggered a series of experimental studies on that isotope [@NPA05; @TTS05; @ABE08; @LBC10] and on other key reactions of the weak $s$ process between Fe and Sr, which could be considerably improved [e.g. @HKU08a; @HKU08b]. For a full account see the recent update of the KADoNiS library [@DPK09]. The cumulated uncertainties of the propagation effect are significant even for the heavier isotopes of the weak $s$ process, up to Kr and Sr, with a possible, minor contribution to the Y and Zr abundances [@PGH10]. The corresponding uncertainties will be partly solved once the neutron capture cross sections of the isotopes between Fe up to Sr will be measured with an accuracy of 5% (Fig. \[fig:6\]). Apart from these problems with the neutron capture cross sections, it has been pointed out that the weak $s$ process is also still affected by large uncertainties of several charged-particle reactions during He and C burning [e.g. @TEM07; @BHP10]. \[sec3C\]AGB stars ------------------ ![(Color online) Structural characteristics versus age of an AGB model with initial mass $M^{\rm AGB}_{ini}$ = 2 $M_\odot$ and solar metallicity. Upper panel, top to bottom: Temporal evolution of the mass coordinates of the inner border of the convective envelope, the mass location of maximum energy production in the H-burning shell, and the maximum energy production within the H-depleted core. During each interpulse period, the flat segment of the lowest (blue) line corresponds to the location of the radiative burning of the $^{13}$C($\alpha, n$)$^{16}$O reaction in the pocket. Lower panel: Temporal evolution of the H-burning and He-burning contributions to luminosity (from @CPS09a).[]{data-label="fig:9"}](rmp_fig9.eps){width="10cm"} ![The formation of the $^{13}$C pocket according to @SCG09. The sequence of the panels shows the evolution of the chemical composition in the transition zone between the H-rich envelope and the H-exhausted core. The various lines represent the abundances of H (crosses), $^{12}$C (dotted), $^{13}$C (solid), and $^{14}$N (dashed). Panel (a): The TDU just occurred and the convective envelope is receding. Panel (b): Production of $^{13}$C by proton capture on the abundant $^{12}$C starts in the hotter region. Panel (c): Some $^{14}$N is also produced, and panel (d): the $^{13}$C (and $^{14}$N) pocket is fully developed.[]{data-label="fig:10"}](rmp_fig10_rev.eps){width="13cm"} During the AGB phase, the H- and He-burning shells are activated alternately on top of the degenerate C-O core. These two shells are separated by a thin zone in radiative equilibrium, the so-called He-intershell, enriched in He and C. The H-burning shell erodes the bottom layers of the envelope and produces He. The He-intershell grows in mass and is progressively compressed and heated until He burning is triggered in a quasi explosive way (thermal pulse, TP; @ScH65 [@Wei66]). For a general discussion see @IbR83 [@BGW99; @Her04; @SGC06; @SCG08]. The sudden release of energy due to a TP drives convection in the whole intershell for a short period of time. During a TP, partial He burning occurs producing a large amount of $^{12}$C. The envelope expands and the H shell is temporarily extinguished. He shell burning continues radiatively for another few thousand years, and then H shell burning starts again. After a limited number of TPs, when the mass of the H-exhausted core reaches $\sim$ 0.6 $M_\odot$ and the H-shell is inactive, the convective envelope penetrates into the top region of the He-intershell and mixes newly synthesised material to the surface (third dredge-up, TDU). The star undergoes recurrent TDU episodes, whose occurrence and efficiency depend on the physical and numerical treatment of the convective borders. The TDU is influenced by the parameters affecting the H-burning rate, such as the metallicity, the mass of the H-exhausted core, and the mass of the envelope, which in turn depends on the choice of the mass loss rate by stellar winds [see discussion in @SGC06]. The upper panel of Fig. \[fig:9\] illustrates the changes of the structural characteristics of a given AGB model with time [@CSG09; @CPS09a], the position in mass coordinates of the inner border of the convective envelope, the maximum energy production of the H-burning shell and the maximum energy production within the H-depleted core. During each interpulse period, the flat segment of the lowest (blue) line corresponds to the radiative burning of the $^{13}$C($\alpha, n$)$^{16}$O reaction. The lower panel shows the H-burning and He-burning contributions to the luminosity. During the TP-AGB phase, the envelope becomes progressively enriched in primary $^{12}$C and in $s$-process elements. TDU drives a chemical discontinuity between the H-rich envelope and the He-intershell, where a few protons likely penetrate into the top layers of the He intershell. At hydrogen reignition, these protons are captured by the abundant $^{12}$C forming $^{13}$C via $^{12}$C($p, \gamma$)$^{13}$N($\beta^+ \nu$)$^{13}$C in a thin region of the He-intershell ($^{13}$C pocket). Neutrons are released in the pocket under radiative conditions by the $^{13}$C($\alpha, n$)$^{16}$O reaction at $T \sim 0.9 \times 10^8$ K. This neutron exposure lasts for about 10,000 years with a relatively low neutron density of 10$^{6}$ to 10$^{8}$ cm$^{-3}$. The pocket, strongly enriched in $s$-process elements, is then engulfed by the subsequent convective TP. Models including rotation [@LHW99] or gravity waves [@DeT03] have obtained a partial mixing zone at the base of the convective envelope during the TDU episodes, which leads to the formation of a $^{13}$C-rich layer of limited mass extension. @HBS97 and @Her00 [@Her04], guided by dynamical simulations of @FLS96, introduced an exponential diffusive overshoot at the borders of all convective zones. A formally similar algorithm based on a non-diffusive mixing scheme has been proposed by @SGC06. Applying the Schwarzschild criterion to determine the border of the H-rich convective envelope and the inner He-rich and C-rich intershell, a thermodynamical instability would ensue. In order to handle this instability, @SGC06 assumed an exponentially decaying profile of the convective velocity $$v = v_{bce} {\rm exp}(-d/\beta H_P)$$ where $d$ is distance from the convective boundary, $v_{\rm bce}$ the average element velocity at the convective boundary (as derived by means of the mixing length theory), $H_P$ the pressure scale height at the convective boundary, and $\beta$ a free parameter (for a proper choice see @CSG09). Fig. \[fig:10\] illustrates the formation of the $^{13}$C pocket according to the full evolutionary model described by @SCG09. The hydrogen profile adopted in the pocket and the consequent amount of $^{13}$C (and of $^{14}$N) determines the final $s$-abundance distribution. At the maximum extension of the convective TP, when the temperature at the base of the convective zone exceeds $2.5 \times 10^{8}$ K, a second neutron burst is powered for a few years by the marginal activation of the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction. This neutron burst is characterized by a low neutron exposure and a high neutron density up to 10$^{10}$ cm$^{-3}$, depending on the maximum temperature reached at the bottom of the TP. The dynamical conditions in which the two neutron sources operate are defining the final abundances of nuclei involved in branchings along the $s$-process path. A comparison of low mass AGB models computed with different evolutionary codes has been discussed by @LHL03. In AGB stars of intermediate mass ($4 < M/M_{\odot} < 8$) the maximum temperature during a TP reaches about $3.5 \times 10^8$ K, leading to a substantial neutron production via the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction. However, both the mass of the He-intershell and the TDU efficiency are much smaller in these stars than in low-mass AGBs. Consequently, the predicted $s$-process abundances in the envelope are fairly low. The production of the $s$ elements at very low metallicity (\[Fe/H\] $<$ $-$2.5) may be affected by a new phenomenon, which is limited to AGB stars of the lowest mass leading to TDU and takes place only during the first fully developed TP. There, the reduction of CNO catalysts is compensated by an increase of the temperature in the H-shell and, consequently, the entropy in the H-shell decreases. Under these conditions, the first convective He instability may expand over the H-shell, thus engulfing protons from the envelope, which are instantly captured by the abundant $^{12}$C in the convective region. The $^{13}$C($\alpha, n$)$^{16}$O reaction is now occurring during the TP, in competition with the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction at the bottom of the TP. This complex feature has been found by many authors despite the different physics adopted in the various works [@HIF90; @FII00; @IKM04; @SCG04; @CSL07; @SKY08; @CaL08; @WHP08; @LST09]. First calculations of the consequences of this mechanism for the $s$ process have been made by @CPS09a for an AGB model of an initial mass $M = 1.5 M_\odot$, a metallicity of \[Fe/H\] = $-$2.6, and no enhancement of the $\alpha$ elements. The convective shell was found to split into two sub-shells: the lower one boosted by the $^{13}$C($\alpha, n$)$^{16}$O reactions and the upper one by the CNO cycle. Once the splitting has occurred, the nucleosynthesis in the two shells exhibits a completely different behavior. In the upper shell, the very large $^{13}$C abundance is marginally consumed via $^{13}$C($\alpha$, n)$^{16}$O, leading to the production of a corresponding amount of ls elements (For the definition of ls and hs elements see Sec. \[sec3D\] below). This peculiar phase is followed by a deep third dredge-up episode, which mixes freshly synthesized $^{13}$C, $^{14}$N, and ls elements into the envelope. The second TDU carries a large amount of hs elements and Pb to the surface, which was previously synthesized by $^{13}$C($\alpha$, n)$^{16}$O reactions in the lower splitted shell. After this initial event, the subsequent series of TPs and TDUs follow the standard pattern. The whole problem and its consequences on the surface abundances are currently a matter of intense study. This phenomenon may be of interest for the analysis of some Carbon-Enhanced Metal-Poor (CEMP) stars showing $s$-process enhancement (CEMP-$s$ stars) (Secs. \[sec3D\] and \[sec4B2\]). Theoretical AGB results \[sec3D\] --------------------------------- The $s$ process in AGB stars is not a unique process, but depends on the initial mass, metallicity, the strength of the $^{13}$C pocket, and the choice of the mass loss rate. Important observational constraints derive from the abundances of elements belonging to the three $s$-process peaks located at the magic neutron numbers N = 50, 82, and 126. The peaks occur because the low neutron capture cross sections of Sr, Y, Zr (light $s$-process elements, ls), Ba, La, Ce, Nd, Sm (heavy $s$-process elements, hs), and Pb act as bottlenecks for the $s$-process reaction path. For a given metallicity, a spread in the three $s$-process peaks is observed in stars of spectral type MS, S, C(N) and Ba stars of the Galactic disk (see @BLB95 [@BGL01; @ABG01; @ADG02; @GWB05]). At low metallicities, high-resolution spectroscopic measurements of CEMP-$s$ stars showed an even larger spread [@ISG05; @ABG06; @TIB08; @RFS08; @SCG08]. The observed spread can be interpreted by assuming different $s$-process efficiencies of these stars, corresponding to a change in the amount of $^{13}$C in the pocket. We discuss here theoretical results from models based on the FRANEC code (Frascati Raphson-Newton Evolutionary Code, @ChS89), coupled with a post-process code that includes a full $s$-process network up to Bi [@BGS10]. In the post-process code, the prescriptions for the amount of the dredged-up mass, the number of TDUs, the choice of the mass loss rate as well as for the temporal history of the temperature and density during the TPs were adopted from @SDC03 [@SGC06]. The $^{13}$C pocket is artificially introduced starting from the ST case adopted by @GAB98 and @AKW99b, which was shown to reproduce the solar main $s$-process component as the average of the 1.5 and 3.0 $M_{\odot}$ models at half solar metallicity (Sect.\[sec3E\]). The $^{13}$C (and $^{14}$N) abundances in the pocket were then multiplied by different factors. A minimum $^{13}$C pocket may be defined as the one that affects the final $s$-process distribution. Higher $^{13}$C-pocket efficiencies than case ST $\times$ 2 would not produce a correspondingly higher abundance of $^{13}$C, because of the increasing competition by $^{13}$C($p, \gamma$)$^{14}$N reactions. For any given model, the efficiency of the $^{13}$C pocket is assumed to be constant for all TPs. The theoretical predictions in the envelope for elements from C to Bi (\[El/Fe\]) versus atomic number $Z$ are shown in Fig. \[fig:11\] for an AGB star with an initial mass $M$ = 1.5 $M_{\odot}$ and for a range of $^{13}$C pockets (from case ST down to ST/12) at solar metallicity (top panel) and at \[Fe/H\] = $-$0.5 (bottom panel) (see @BGS10; @HGB09). With decreasing metallicity, the $s$-process distribution is shifted toward heavier elements. This is the consequence of the primary nature of the $^{13}$C neutron source: while $^{56}$Fe (the seed of the $s$-process) decreases with metallicity, the number of neutrons available per iron seed increases [@Cla88]. Hence, at halo metallicities, the neutron fluence overcomes the first two peaks, directly feeding $^{208}$Pb (@GAB98 [@GoM00; @TGB01b]; see Sect.\[sec3E\]). In Fig. \[fig:12\], the relative behaviour of the three $s$-process peaks ls, hs, and Pb are analyzed using the definitions \[[ls/Fe]{}\] = 1/2(\[[Y/Fe]{}\] + \[[Zr/Fe]{}\]) and \[[hs/Fe]{}\] = 1/3(\[[La/Fe]{}\] + \[[Nd/Fe]{}\] + \[[Sm/Fe]{}\]). In general, the choice of the specific elements considered in the average ls and hs abundances varies from author to author and depends on the quality of the spectra available. Our choice is made because, at disk metallicity, Sr and Ba have few and saturated lines (see @BLB95 [@BGL01]) and may be affected by non-local thermodynamical equilibrium (NLTE) corrections, in particular at low metallicities (@ASK09 and references therein). The \[hs/ls\] and \[Pb/hs\] ratios are indexes of the whole $s$-process distribution. They remain unchanged both in the envelope of the AGB companion (now a white dwarf) and in the envelope of the observed star after accretion of AGB winds. The top panel of Fig. \[fig:12\] shows \[hs/ls\] predictions for AGB stars of an initial mass of 1.5 $M_\odot$ as a function of \[Fe/H\] for different $^{13}$C-pocket efficiencies. The figure also contains representative abundance data discussed in Sec. \[sec4\]. Considering the ST case as a function of metallicity in the top panel of Fig. \[fig:12\], one finds that the \[hs/ls\] ratio first increases with decreasing \[Fe/H\], reaching a maximum at \[Fe/H\] = $-$1.3 and then decreases again. This behaviour is due to the progressive build-up of hs elements and, subsequently, of the third $s$-peak at $^{208}$Pb. Note that $^{208}$Pb becomes dominant over the ls and hs components already at a metallicity of \[Fe/H\] = $-$1. In the bottom panel of Fig. \[fig:12\] theoretical predictions for \[Pb/hs\] versus \[Fe/H\] are compared with spectroscopic observations of Ba stars, CEMP-$s$ and CEMP-$s/r$ stars. For halo metallicities, the spread in \[Pb/hs\] is found to be about 2 dex. A sample of a hundred CEMP-$s$ stars have been observed (see @SCG08 and references therein) including main-sequence stars, subgiants, or giants, far from the AGB phase where the $s$-process is taking place. Moreover, to be observed today, these stars have long lifetimes and correspondingly low initial masses ($M$ $\leq$ 0.9 $M_\odot$). Therefore, the hypothesis of mass accretion of $s$-rich material from a more massive AGB companion becomes essential to explain the overabundances detected in their spectra. The spectroscopic $s$-process abundances in CEMP-$s$ stars depend on the fraction of the AGB mass transferred by stellar winds, whereas the $s$-process indexes \[hs/ls\] and \[Pb/hs\] remain unchanged. The mass transfer can be simulated by introducing a dilution factor between the accreted AGB mass and the original envelope of the observed star. The dilution factor $dil$ can be defined as the logarithmic ratio between the mass of the convective envelope of the observed star before the mixing, and the total transferred AGB mass. About half of the known CEMP-$s$ stars are also $r$-process-rich with \[Eu/Fe\] comparable to \[La/Fe\]. Among the sample of stars reported in the literature with Eu measurements, six stars show an $r$-process enrichment of $\sim$2 dex, among which are HE 2148-1247, the first CEMP-$s/r$ star discovered [@CCQ03], CS 29497-030 [@ISG05], HE 0338-3945 [@JBG06], and CS 22898-027 [@ABC07]. The $s$ and $r$ processes occur in completely different astrophysical scenarios, the $s$ process in low mass AGB stars and the $r$ process during explosive nucleosynthesis in massive stars. While Eu is a typical $r$-process element (about 94% of solar Eu is of $r$ origin), the elements of the hs peak, as Ba and La, are mainly synthesized by the $s$ process. In particular, 70% of solar La is produced by the $s$ process. AGB $s$-process predictions give \[La/Eu\]$_{\rm s}$ $\sim$ 1. Consequently, spectroscopic observations of enhanced \[La/Fe\] and comparable \[Eu/Fe\] cannot be explained by $s$-process AGB models alone. @CVH97 and @VaC98 showed through numerical simulations how the supernova ejecta at high velocities interact with a nearby molecular cloud inducing Rayleigh-Taylor instabilities in the cloud, polluting it with freshly synthesized material, and at the same time triggering the condensation of a low mass binary system. This scenario may explain the high fraction of CEMP-$s/r$ stars. Other explanations for these peculiar CEMP-$s/r$ stars have also been advanced, e.g. by @CCQ03 and @WNI06. On the other hand, 13 CEMP-$s$ stars do not show any $r$-process-enhancement, i.e. CS 22880-074 [@ABC07], HE 0024-2523 [@LGC03], HE 2158-0348 [@CMS06], and HE 0202-2203 [@BCB05]. In Fig. \[fig:13\] we show theoretical predictions of \[hs/ls\] vs \[Fe/H\] for AGB models with initial mass $M = 1.3 M_{\odot}$, different $^{13}$C-pocket efficiencies, and an initial $r$-process enrichment \[r/Fe\]$^{ini}$ = 2.0. The label “n5” indicates the number of TDU episodes suffered by the 1.3 M$_\odot$ model. For a given element, the initial $r$ enrichment is determined via the solar $r$-process contribution per isotope using the $r$-residual method [@AKW99b]. Fig. \[fig:13\] shows also abundance data of all main-sequence/turnoff CEMP-$s/r$ stars (big full red diamonds). The two CEMP-$s/r$ stars CS 29497-030 and CS 31062-050, which will be discussed in detail in Sec. \[sec4B\], are indicated by a big filled diamond and a full rotated triangle, respectively. Due to the small number of TDUs in the 1.3 $M_\odot$ AGB model, the final surface distribution is affected by the initial $r$ enrichment. In particular, at low metallicities, the predicted \[hs/ls\] ratios reach values as high as 1.3 dex. In the top panel of Fig. \[fig:12\], however, the predictions reach a maximum value \[hs/ls\]=1, independent of whether an $r$ enhancement of 2 dex is included or not. Therefore, the most important message is that theoretical predictions given in Fig. \[fig:13\] match the observations of main sequence/turn-off CEMP-$s/r$ stars quite well. As an example, we show in Fig. \[fig:14\] the two CEMP-$s/r$ stars CS 29497–-030 [@ISG05] and CS 31062–-050 (@JoB04, @ANR02b [@ARN02; @ABG06; @ABC07]), which are spectroscopically well studied. A more detailed discussion is presented in Sec. \[sec4B\]. While the AGB model calculations are admittedly quite complex and subject to a number of free and model-dependent parameters, they have been successful in reproducing the abundance patterns of the CEMP-$s$ stars. The main $s$ component and the role of GCE\[sec3E\] --------------------------------------------------- The $s$-process abundance distribution of the heavy isotopes beyond A $\sim$ 90 must be considered as the result of all previous generations of AGB stars that were polluting the interstellar medium before the formation of the Solar System. Therefore, the cosmic $s$-process abundances have to be explained by means of a general Galactic Chemical Evolution (GCE) model. It was shown by @GAB98 and @AKW99b that the solar main component can be reproduced by assuming a standard $^{13}$C pocket, a metallicity of \[Fe/H\] = $-$0.3, and by averaging between stellar models of $M$ = 1.5 and 3 $M_\odot$. The distribution obtained with this prescription and updated nuclear input by @BGS10 is plotted in the top panel of Fig. \[fig:15\] normalized to the $s$-only isotope $^{150}$Sm. The set of $s$-only isotopes indicated by full circles is well reproduced. Open symbols have been used for $^{128}$Xe, $^{152}$Gd, and $^{164}$Er, which have a non-negligible $p$ contribution (10% for Xe), for $^{176}$Lu, a long-lived isotope (3.8 $\times$ 10$^{10}$ yr), which decays into $^{176}$Hf, for $^{187}$Os, which is affected by the long-lived decay of $^{187}$Re (4.1 $\times$ 10$^{10}$ yr), and for $^{180}$Ta, which receives also contributions from the $p$-process and from $\nu$-nucleus interactions in massive stars. The black full square corresponds to $^{208}$Pb, the only heavy isotope, which is clearly underproduced by the main $s$ component [@ClR67]. Employing a GCE code, in which the Galaxy is subdivided into three zones (halo, thick, and thin disks), adopting the $s$-process yields from AGB stars of different mass and metallicity, and accounting for their respective lifetimes, @TGA04 determined the temporal variation of the $s$-process abundances in the interstellar medium. Their results were recently updated by @SGT09. The resulting $s$-process distribution at the epoch of the Solar System formation is plotted in the bottom panel of Fig. \[fig:15\]. One finds that GCE calculation [@TGA04; @SGT09] yield good agreement with the solar abundance values of the $s$-only isotopes between $^{134,136}$Ba and $^{204}$Pb. Moreover, also the solar abundance of $^{208}$Pb is well reproduced by the contributions from low metallicity AGB stars, thus solving the long-standing problem of the origin of the strong $s$-component in a natural way. Below magic neutron number N = 82, however, there is a significant discrepancy between the abundance distribution obtained by the GCE approach and the Solar System values. It turned out that GCE models underproduce the $s$-process component of the solar-system abundances of Sr-Y-Zr by about 20-30% [@TGA04]. A similar deficit holds for the $s$-only isotopes from $^{96}$Mo up to $^{130}$Xe. This finding prompted @TGA04 to postulate another source of neutron-capture nucleosynthesis in the Galaxy defined as Light Element Primary Process (LEPP). The LEPP process is different from the $s$ process in AGB stars and also different from the weak $s$-process component occurring in massive stars. @TGA04 suggested that 8% of solar Sr and 18% of solar Y and Zr must come from the LEPP. There is general consensus for the need of an additional LEPP source of yet unknown origin for the light isotopes, including Sr, Y, and Zr, an issue that is still highly debated [@FKP10; @PGH10; @QiW07]. @MBC07 surveyed the possible ranges of parameters (e.g. for the neutron density) that reproduce the abundance patterns of HD 122563, a very metal-poor star representing the yields of the LEPP [@HAI06]. Comparisons of the GCE models with observations of Galactic stars are discussed in Sec. \[sec4E\]. Observational constraints \[sec4\] ================================== Models of nuclear reactions and enrichment of elements in the Universe are examined by astronomical observations as well as by the analysis of solar-system material. The mechanisms and astronomical sites of the $s$ process are constrained by the comparison of chemical abundance patterns with model predictions. Some key observational studies obtained until the middle of 1990s have been outlined in the review paper by @WIP97. In the following sections a summary of the observational studies on the $s$ process obtained in the past 15 years is presented. Detailed chemical abundances have been determined for solar-system material. As described before, the classical model of the $s$ process has been constructed by fitting the ${\sigma}N$ curve to those nuclei, which are only produced by the $s$ process [@KBW89]. A more physical approach is to fit the abundances predicted by AGB models to the solar $s$-process component [@AKW99b]. The chemical abundances of the Solar System reflect the composition of a particular site of the Galaxy at 4.6 billion years ago, to which numerous nuclear processes in a variety of stars have contributed. Therefore, the abundance pattern of the $s$-process nuclei represents an average over the products from a variety of objects. In addition, information on individual nucleosynthesis events are provided by the rapid progress in the measurements on pre-solar grains, which provide accurate isotope ratios of important elements [@Nit09; @ClN04; @Zin98]. Stellar abundances \[sec4A\] ---------------------------- Useful information on chemical $s$-process yields by individual objects is obtained by observations of stars and planetary nebulae. Stellar abundances are derived from high resolution spectra of the ultraviolet, optical and infrared ranges. In general, only elemental abundances can be determined in analyses of stellar spectral lines and the measurable elements are limited compared with the analyses of solar-system material. However, observations for appropriately selected targets (e.g. AGB stars) enable one to investigate directly the products of individual processes, and to identify the respective astrophysical sites. Moreover, measurements of isotope abundance ratios have been made for a few exceptional elements from detailed analyses of spectral line profiles (see Sec. \[sec4C\]). Measurable elements are dependent on objects. Stellar abundances are usually derived from the absorption features formed at the surface of stars in the so-called photosphere. In cool stars, neutral species of alkali metals (e.g. Na, K, Rb) and singly ionized species of alkaline earth metals (e.g. Mg, Ca) show strong resonance doublet lines. Singly ionized Sr and Ba atoms, which belong to elements of the $s$-process abundance peaks at neutron magic numbers 50 and 82, exhibit strong absorption features in the optical range, which makes it possible to determine their abundances even in very metal-poor stars. Molecular absorption dominates in very cool stars. In such cases, molecular bands, e.g for ZrO, can be used to estimate the elemental and isotopic abundances of heavy neutron-capture elements. The abundances of noble gases (e.g. Ne, Ar) are very difficult to determine in cool stars as well as in the Sun because of the lack of useful spectral lines. Such elements are observable from emission lines in planetary nebulae, where these elements are ionized and excited by the ultraviolet photons from the hot central star. Hence, observations of planetary nebulae provide complementary information to the chemical yields obtained from stellar observations. Considerable progress has been achieved in the past few decades in stellar spectroscopy related to the $s$ process. They are promoted by high resolution spectrographs mounted on large telescopes, application of spectrum synthesis technique to high resolution spectra of cool stars, and surveys of metal-poor stars that provide useful samples of binary stars (see below). The main targets of $s$-process observations are (i) AGB stars, post-AGB stars and planetary nebulae evolving to white dwarfs, and (ii) binary companions accreting the products of primary AGB stars. AGB stars are objects where the $s$ process is currently taking place (or recently took place), giving direct information on the $s$-process nucleosynthesis in such stars. However, the surface of AGB stars is very cool and the spectra are crowded by numerous molecular lines (Fig. \[fig:17\]). Accordingly, abundance analyses from such spectra are very difficult and the number of measurable elements is limited. ![Comparison of spectra around the Ba [II]{} 4934 [Å]{} line for AGB, post-AGB and CEMP stars. Spectral data were obtained with the High Dispersion Spectrograph (HDS) of the Subaru Telescope with a resolving power of 60,000 or higher. The spectrum of the metal-poor star is normalized to the continuum level (bottom). The atomic absorption lines are clearly seen in the spectrum and are useful for abundance analyses. In contrast, molecular absorption (mostly of C$_2$) is dominant in the AGB star (top), and the continuum level is very uncertain. Molecular absorption of the post AGB star is not as severe as in the AGB star, depending on stellar temperature. \[fig:17\]](rmp_fig16_rev.eps){width="10.5cm"} On the other hand, in binary systems with appropriate separation, material from the surface of the primary star starts to accrete onto the secondary companion when the primary evolves to an AGB star. Such mass transfer across a binary system is expected by stellar winds from the AGB primary or by forming a common envelope when the separation is sufficiently small. The secondary is usually an unevolved main-sequence star with a much longer timescale of stellar evolution compared to the AGB primary. After the primary evolves to a faint white dwarf, only the companion, which preserves the material provided by the AGB primary, is observable. In such cases, the $s$-process yields are easily measurable at the surface of the secondary, because the object is still relatively warm and molecular absorption is not severe (Fig. \[fig:17\]). ### AGB, post-AGB, and planetary nebulae \[sec4A1\] As described in Sec. \[sec3\] heavy $s$-process elements are enriched on the surface of AGB stars as the result of repeated TPs and TDU episodes. In such objects, the initial CNO abundances at the surface are significantly changed. In particular, carbon is enhanced and might become comparable or more abundant than oxygen. These stars are classified as spectroscopic types S or N, respectively. In massive AGB stars, which are affected by the hot bottom burning (HBB) process (@Her05 and references therein), the surface composition changes toward the equilibrium values of the CNO cycle with a reduced C/O ratio. Some $s$-process elements might also be enriched in such stars. The surface of AGB stars is cool ($\sim 3000$ K) and the optical and near infrared spectra are dominated by molecular absorption lines, e.g., from CO, TiO, and C$_{2}$, depending on chemical composition and temperature. While the abundances of C, N, O, and some other elements can be determined via molecular lines [e.g. @LGE86], analyses of atomic spectra are very difficult. However, recent analyses based on the spectrum synthesis technique provided important results on the surface abundances of heavy elements in cool AGB stars (see Sec. \[sec4D\]). Molecular absorption features like ZrO are also useful to determine abundances of neutron-capture elements and their isotope ratios [@LSB95; @Zoo85; @PeB70]. Excesses of heavy $s$-process elements are also found in post-AGB stars, which are quickly losing their residual envelopes while evolving from AGB stars to planetary nebulae. The surface of such objects becomes warmer and molecular absorption becomes weaker as this transition proceeds, but the atmospheric structure is unstable and quite complicated as that of supergiants, making spectral analyses difficult (Fig. \[fig:17\]). The duration of this evolutionary stage is very short, indicating that one must observe apparently faint objects to increase the sample, even though the luminosity of such objects is high. Useful results have been obtained by recent observations with large telescopes and detailed analyses of high resolution spectra. Planetary nebulae are formed by the material ejected from evolved low- and intermediate-mass stars and ionized by the central star that is evolving to a hot white dwarf. Hence, the spectra of nebulae provide direct information on the yields that these stars contribute to the chemical enrichment of the Galaxy. The light elements (e.g., C, O) in planetary nebulae are accessible by X-ray observations [e.g. @MKM06], while neutron-capture elements can be studied by optical spectroscopy (Sec. \[sec4D\]). ### Binary systems affected by mass transfer \[sec4A2\] Detailed information on the abundance patterns of heavy elements produced by AGB stars can be obtained by analyses of a binary companion that is affected by mass accretion from the primary AGB star. In these cases, the target objects are main-sequence or red giant stars, but they are distinguished from normal stars by excesses of carbon and heavy $s$-process elements like Ba. Such stars with high metallicity similar to the Sun are known as Ba stars [@BiK51]. Since the CNO abundances and, in particular, the oxygen abundance in stars with solar metallicity are already high, mass accretion from carbon-enriched AGB stars does not significantly change the molecular features of their spectra. Instead, strong absorption features of Ba can be a signature for the stellar classification. Periodic variations of the radial velocity have been found for many Ba stars, supporting the scenario that these stars experienced accretion of $s$-process enhanced material from a primary AGB star in a binary system [@Mcc84b; @McW90]. Mass accretion from AGB stars has stronger impacts on spectra of metal-poor stars. Such objects are classified into CH stars [@Kee42] or subgiant CH stars [@Bon74]. Excesses of heavy elements have also been detected in such objects, and the binarity has been confirmed by radial velocity monitoring [@Mcc84b; @McW90; @PrS01]. Recent surveys of metal-poor stars based on the weakness of calcium absorption lines have detected a number of carbon-enhanced objects, e.g. the HK-survey [@BPS92; @BSM07], the Hamburg/ESO Survey [@Chr03], the Sloan Digital Sky Survey (SDSS, @YAA00) including the subprogram SEGUE (Sloan Extension for Galactic Understanding and Exploration) and the SEGUE Stellar Parameter Pipeline (SSPP; @LBS08a [@LBS08b]). Follow-up high-resolution spectroscopy has been made by the ESO Large Programme First Stars with VLT/UVES [e.g. @SBB06], the Chemical Abundances of Stars in the Halo (CASH) Project [@RFS08], and others [e.g. @ABC07]. Enhancements of heavy elements are found for most of these stars, indicating that their peculiar abundances originate from AGB nucleosynthesis and mass transfer in binary systems, as in the case of CH stars. The sample of such objects contains stars with a variety of metallicities, and enables one to investigate the metallicity dependence of the $s$ process. Abundance patterns of heavy elements \[sec4B\] ---------------------------------------------- ### Overall abundance patterns covering three magic neutron numbers \[sec4B1\] Measurements of abundance patterns of heavy elements in AGB stars or objects affected by AGB nucleosynthesis and comparisons with model predictions are important for understanding the $s$-process mechanisms. The $s$-process abundance pattern exhibits three peaks at $^{88}$Sr, $^{138}$Ba, and $^{208}$Pb, corresponding to the magic neutron numbers 50, 82, and 126, respectively. The singly ionized Sr and Ba have strong resonance lines in the optical range, making such elements detectable even in spectra of very metal-poor stars. There are also many useful lines in the near UV and blue spectral range for the determination of e.g., Y, Zr, La, Nd, and Eu abundances. For this reason, measurements of light (ls, comprising Sr, Y, and Zr) and heavy (hs, with Ba, La, etc.) neutron-capture elements have been performed for many CH and Ba stars in order to estimate the neutron exposure from the ls/hs ratios. Abundance studies of cool AGB stars are very difficult due to the dominant molecular absorption features. Important progress is, however, obtained by recent work. @ADG02 have determined chemical abundances of N type AGB stars. By means of the spectrum synthesis technique abundances of neutron-capture elements were determined for a large sample of AGB stars with excesses of such elements. The abundance ratios could be explained by $s$-process nucleosynthesis in AGB stars with $M\leq 3M_\odot$. However, the large scatter in the abundance ratios (e.g. \[hs/ls\]) suggests that the efficiency of the $s$ process is affected by yet uncertain model parameters (Figs. \[fig:12\] and \[fig:13\]). The $s$-process products of AGB stars are also recorded in Ba stars. @AlB06 studied neutron-capture elements for 26 Ba stars, separating the contribution of the main $s$ process by the progenitor (i.e. the former primary star in the binary system that provided the $s$-enhanced material) from other original components of the observed Ba stars and compared the observational results with the model predictions by @Mal87a and @Mal87b for single and exponential neutron-exposures of the $s$ process. Also @SPS07 determined abundances of neutron-capture elements in Ba stars. Excesses of heavy neutron-capture elements and the hs/ls ratios are discussed in their work, while no significant $s$ process effects were found for Cu, Mn, V and Sc. A recent important progress is that measurements have been extended to Pb, the stable element at the third abundance peak of the $s$ process. The lack of strong spectral features in the optical range makes the abundance measurements of Pb from stellar spectra very difficult. The detection of Pb was reported for the post-AGB star FG Sge by @GLW98. Measurements of Pb abundances have been made by @ANR00 [@ARN01] for CEMP stars, by @VGJ01 for CH stars, and by @AlB06 for Ba stars using the absorption line of neutral Pb at 4058 [Å]{}. These observations made it possible to investigate the overall $s$-process abundance pattern more consistently. By now abundance studies for Pb have been made for more than 20 carbon-enhanced objects [@ARN02; @VGJ03; @LGC03; @CCQ03; @SBM04; @SBB06; @ABS08]. Moreover, @ISG05 reported the detection of Bi as well as Pb for the carbon-enhanced metal-poor star CS 29497-030. Fig. \[fig:14\] shows the abundance pattern of this and another well studied carbon-enhanced star (see below). The Pb abundances determined for carbon-enhanced metal-poor stars revealed that the abundance ratios of elements between the second and third abundance peaks (e.g. \[Pb/hs\]) show large star-to-star scatter (Fig. \[fig:12\]) as discussed in Sec. \[sec3\]. A similar scatter is also found in the abundance ratios of elements between the first and second abundance peaks. This scatter can be interpreted as due to a dispersion of the hydrogen mixing into the $^{13}$C pocket that determines the efficiency of the $s$ process in AGB stars [e.g. @BGW99]. The efficiency is expected to depend on the metallicity, which determines the ratio of neutrons to seed nuclei at the $s$-process site. However, Fig. \[fig:12\] shows no clear correlation between the metallicity or iron abundance and the \[Pb/hs\] or \[hs/ls\] abundance ratios. Thus, the reason for the scatter of the abundance ratios is still unclear. Using the technique described in Sec. \[sec3D\] and AGB stellar models of disk metallicity, @HGB09 were able to fit the Ba stars in a range of $^{13}$C-pocket efficiencies ST$\times$2 down to ST/3 and initial AGB masses between 1.5 and 3 $M_\odot$. The results are plotted in Fig. \[fig:12\] for \[hs/ls\] and for the somewhat more uncertain \[Pb/hs\]. ### Detailed abundance patterns: CEMP-$s$ and -$s/r$ stars \[sec4B2\] In addition to the overall abundance patterns, which are represented by the abundance ratios between the three $s$-process peaks, detailed abundance measurements for CEMP stars with \[C/Fe\] $> 1$ provide useful constraints on the origin of the heavy elements and the role of the $s$ process. An illustrative case are the high Eu abundances, relative to Ba and other $s$-process elements, in some CEMP stars, which can not be explained by standard $s$-process models [@HBS00; @ARN02; @CCQ03]. Although the $r$ process is the dominant source of Eu in solar-system material (Eu is sometimes called an $r$-process element), this element is also produced by the $s$ process. Therefore, large enhancements of heavy elements in carbon-enhanced objects can also provide some excess Eu. However, the measured abundance ratios of Eu with respect to Ba, La and other elements around Eu are significantly higher than in model predictions as well as in the $s$-process component of solar-system material. For this reason, the Eu excess suggests a large $r$-process contribution in addition to the $s$-process component, and such stars are sometimes called CEMP-$s/r$ stars [@BeC05]. As discussed in Sec. \[sec3D\], a preliminary comparison of CEMP-$s$ with the theoretical models by @BGS06 has provided reasonable interpretations for all CEMP-$s$ and CEMP-$s/r$ stars published so far as illustrated for the $s$-process indexes \[hs/ls\] and \[Pb/hs\] in Fig. \[fig:12\]. A full analysis with updated AGB models is underway [@BGS10]. Examples of the detailed elemental abundance patterns of two CEMP-$s/r$ stars have already been presented in Fig. \[fig:14\], where the top panel shows the situation for the blue metal-poor (BMP) star CS 29497–030 with \[Fe/H\] = $-$2.57 [@ISG05]. A large number of elements have been detected in that star besides the usual ls and hs elements and Pb, including Nb, Bi, some of the heaviest rare-earth elements, and a few useful upper limits. The abundance ratios with respect to Fe are normalized to the solar values, and compared to an AGB model of the same metallicity, for an inferred initial mass of 1.3 $M_\odot$, and a rather small $^{13}$C pocket (case ST/9). Note that no dilution is required to fit this CEMP-$s$ star, suggesting that the material from the AGB companion dominates at the surface of this object. The very low ratio \[hs/Eu\] $\sim$ 0 indicates that the original composition of the binary system had a huge $r$-process enhancement of \[Eu/Fe\] $\sim$2 dex, even higher than the known $r$-II stars[^1]. CS 31062–-050 is a subgiant CH star with a large enhancement in $s$- and $r$-process elements. Lines of interesting elements such as Os and Ir [@ABG06] and Na [@ABC07] have been detected due to its relatively cool temperature and low gravity ($T_{\rm eff}$ = 5600 $\pm$ 150 K and log$_{10}$ $g$ = 3.0 $\pm$ 0.3). This star has probably undergone the first dredge-up episode, where the convective envelope extends over about 80% of the stellar mass. This implies that the $s$-process rich material accreted from the AGB star by stellar winds needs to be diluted by a large extent with the original material of the observed star. In the bottom panel of Fig. \[fig:14\] the observed abundances are compared with AGB yields from models of 1.5 $M_\odot$ (case ST/2.7) and $dil$ = 1.0 dex, which corresponds to ten parts of original material mixed with one part of accreted material from the AGB star. This star appears to be particularly highly enhanced in both $s$- and $r$-process elements. \[Eu/Fe\] observed is about 1 dex higher than expected for a pure $s$-process AGB prediction (blue dashed line), indicating an important initial $r$-process contribution corresponding to \[r/Fe\]$^{ini}$ = 1.6. Ba is possibly overestimated with respect to the other heavy-$s$ elements [@ABG06]. Detailed abundance ratios provide other constraints on AGB models. For instance, the observation of Zr and Nb permits the immediate confirmation of the extrinsic AGB nature of CS 29497-030. In AGB stars Nb is produced by the radiogenic decay of the long-lived $^{93}$Zr ($\tau_{\rm 1/2}~=~1.5 \times 10^6$ yr), which will mostly decay later in the interstellar medium to $^{93}$Nb. In the envelope of an intrinsic AGB one would expect \[Zr/Nb\] $=$ 1, but in an extrinsic (mass receiving) star like CS 29497–030 all $^{93}$Zr should have already decayed to $^{93}$Nb, and thus one expects \[Zr/Nb\] = 0. This ratio were potentially a powerful tool for determining the state of $s$-process enhanced stars, but, unfortunately, Nb has only a single easily observable transition. The Nb II 3215.6 Å line falls in the crowded near-UV spectral region, and no systematic survey of Nb abundances in CEMP-$s$ stars has been undertaken so far. Na in CS 31062-050, for which a single line has been detected [@ABC07], is lower by approximately 1 dex than the model predictions for the $M = 1.5 M_\odot$ case (Fig. \[fig:14\]). A lower Na yield is obtained with models of lower mass AGB stars and would agree with an AGB model of $M$ = 1.3 $M_\odot$ (case ST/10 and \[r/Fe\]$^{ini}$ = 1.6, and a negligible dilution of 0.2 dex, Fig. \[fig:18\]). This solution would only be compatible with a subgiant star before the first dredge-up. Note that the exact evaluation of the effective temperature at which the first dredge-up occurs may be affected by the treatment of gravitational settling in binary systems with mass transfer as it is the case in CEMP stars. Another possibility is to consider low mass AGB models at [*very*]{} low metallicities. As discussed before, low metallicity AGBs with initial mass $M < 1.5 M_\odot$ may undergo a deep first TDU episode (Sec. \[sec3C\]). Branchings \[sec4C\] -------------------- Detailed abundance measurements, in particular determinations of isotope ratios, are useful for the analysis of the branching points of the $s$ process to derive observational constraints on the temperature and neutron density of the $s$-process sites as illustrated at the following examples. ### $^{85}$Kr branch \[sec4C1\] The ground state of the unstable isotope $^{85}$Kr has a half-life of 10.7 years. At low neutron density, e.g. $n_n<10^{7}$ cm$^{-3}$, the $\beta$-decay of $^{85}$Kr dominates over the neutron capture rate and the $s$-process path runs from $^{84}$Kr to $^{85}$Rb to $^{86}$Sr. At high neutron density the neutron capture chain runs from $^{84}$Kr to $^{86}$Kr and on to $^{87}$Rb (see @KBW89). These main paths are somewhat affected by the minor branching at $^{86}$Rb as well. As a result, a high Rb abundance ratio with respect to Sr, Y, or Zr signifies an $s$ process at high neutron density ($>10^{8}$ cm$^{-3}$), because of the small neutron capture cross section of the neutron magic isotope $^{87}$Rb. In thermally pulsing AGB stars low neutron densities are expected during the inter-pulse phase, where the neutron source is the $^{13}$C($\alpha, n$)$^{16}$O reaction, while high neutron densities are obtained during the He shell flashes, when the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction is activated at the higher temperatures at the bottom of the reaction zone. In very cool stars Rb abundances are measurable via resonance lines of neutral Rb at 7800 [Å]{} and 7947 [Å]{}. This is not the case for warmer stars because the ionization potential of this element is very low, and ionized species have no measurable lines in the optical and the near infrared ranges. Rb abundance ratios were obtained for AGB stars of spectral type M, MS and S by @LSB95, who derived values of Rb/Sr $\sim $ 0.05 for $s$-processed material from stellar surface abundances and taking the dilution by envelope material into consideration. The low Rb/Sr ratio is consistent with the predicted low neutron density during the inter-pulse phase in low-mass AGB stars. These authors also measured the Zr isotope ratios from ZrO molecular bands, and investigated the branching at $^{95}$Zr, an isotope with a $\beta$-decay half life of 65 days. No evidence of existence of $^{96}$Zr was found, supporting the low neutron density estimated from the Rb abundances. A similar conclusion was obtained for carbon stars by @ABG01, who measured the abundances of Rb and other elements for 21 N-type carbon stars and found that the \[Rb/Sr, Y, Zr\] abundances are better explained by AGB models for low-mass ($M\lesssim 3M_{\odot}$) than for intermediate-mass stars, although the determination of Rb is uncertain for carbon stars and some calibration of abundance ratios to a non $s$-process-enhanced AGB star (the J-type carbon star WZ Cas) was required. On the other hand, @GGP07 investigated OH/IR stars, oxygen-rich AGB stars that are believed to have high masses (4–8 $M_{\odot}$). They found large enhancements of Rb in these objects over a wide range ($-1.0<$\[Rb/Fe\]$<2.6$). Given the only mild excesses of $s$ material (\[Zr/Fe\]$<0.5$), the Rb/Zr ratios are significantly higher than in low-mass AGB stars (S-type and carbon stars) investigated in the above studies. Accordingly, the high Rb/Zr ratios were interpreted as evidence for the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction during thermal pulses in massive AGB stars. Measurements of Rb abundances were extended to metal-deficient ($-2.0<$\[Fe/H\]$<0.0$) stars in the Galactic disk and halo by @ToL99. Excluding CH stars, the Y, Zr, and Ba in their sample with \[Fe/H\]$<-0.5$ are underabundant, while Rb is overabundant (\[Rb/Fe\]$=0.21$) on average. The high Rb abundance ratio with respect to Y and Zr are at least partially attributable to the larger component of the $r$ process in metal-deficient stars than in solar-system material. However, an alternative possibility is that the $s$-process neutron density is higher in metal-deficient AGB stars than that in stars with solar-like composition. Rb abundances in globular cluster stars with $-1.7<$\[Fe/H\]$<-1.2$ have been studied by @YAL06 [@YLP08]. While Rb, as well as other $s$-process elements, is overabundant in M4 (see Sec. \[sec4E\]), the \[Rb/Fe\] ratios in NGC 6752 and M13 are similar to the solar value. The \[Rb/Y\] and \[Rb/Zr\] ratios are constant within the uncertainty of the measurement, suggesting that the nature of the $s$ process that has contributed to these clusters is similar (although the $s$ contribution is significantly larger in M4 than in the other two cases). The discussion might be, however, more complicated if the contribution of the $r$ process is fully taken into consideration. Such measurements are also important for the understanding of the abundance variation of light elements (e.g., O, Na, Mg) found in some globular clusters [@YAL06]. ### $^{151}$Sm branch \[sec4C2\] The branching at $^{151}$Sm is of particular interest because the 93 yr half-life of $^{151}$Sm decreases by about a factor of 30 at $s$-process temperatures [@TaY87]. This branch has been studied using the solar-system abundance ratios of Gd isotopes (e.g. @WVK95b). However, the physical conditions of the $s$ process in an individual site (object) are not obtained from the analysis of solar-system abundances. Moreover, the Gd isotopes are also affected by contamination of the $p$ process. The effect of the $^{151}$Sm branching also appears in the abundance ratio of the two stable europium isotopes $^{151}$Eu and $^{153}$Eu. Although measurements of isotope ratios from stellar spectra are difficult in general, Eu lines show relatively large hyperfine splitting, which is characteristic of the two isotopes (Fig. \[fig:16\]). Thus, detailed analyses of Eu absorption line profiles enable one to estimate the Eu isotope ratios [@LWD01]. @SCL02 and @AHB03 determined the Eu isotope ratios for $r$-process-enhanced, metal-poor stars. The results agree well with the Eu isotope ratio of the $r$-process component of solar-system material (with a 48% fraction of $^{151}$Eu), confirming the usefulness of the line profile analysis. ![Measurements of Eu isotope ratios from the Eu II absorption line profile for an $r$-process-enhanced star (top) and CEMP-$s$ and CEMP-$s/r$ stars. Adapted from Aoki [*et al.*]{} (2003). \[fig:16\]](rmp_fig18_rev.eps){width="9.5cm"} @ARI03 applied such an analysis to two CEMP stars that show large excesses of $s$-process elements. The fraction of $^{151}$Eu derived are 55 and 60%, slightly higher than was found for $r$-process-enhanced stars. These results are consistent with the predictions of $s$-process models for wide ranges of neutron density and temperature, given the uncertainties of the measurements (3%). However, the high neutron density case might be preferable for explaining the results if recent measurements of $^{151}$Sm neutron capture cross section are adopted [@MAA06; @WVK06c]. Further detailed analyses for a larger sample will give stronger constraints on neutron density and temperature at $s$-process sites of low metallicity. Stellar evolution \[sec4D\] --------------------------- Observations of characteristic $s$-process elements are important for probing the evolution of AGB stars (Sec. \[sec3C\]). Tc has no stable isotope, and the first isotope was artificially produced in 1937. While $^{98}$Tc has the longest half-life ($t_{1/2} = 4 \times 10^{6}$ yr), $^{99}$Tc ($t_{1/2} = 2 \times 10^5$ yr) is expected to be the most abundant isotope produced by the $s$-process in AGB stars. The discovery of Tc in the spectrum of an AGB star of spectral type S by @Mer52b provided firm evidence for the synthesis of heavy elements in such evolved stars, and for stellar nucleosynthesis in general. Abundance studies for Tc in AGB stars have been made by @LLB87, @SmL88 for example. Recently, @VSL07 compiled the observational results for S stars and investigated the correlation of the spectral features between Tc and Li, because a strong Li line indicates the contribution of HBB in massive AGB stars. Tc line absorption is detected in 28 stars of their sample, and nine of these stars show also a strong Li line. Tc is expected to be observed in low-mass AGB stars, which are actively producing the $s$ abundances, the simultaneous detection of Tc and Li suggests the existence of a concomitant production mechanism for Li, although HBB is not expected in AGB stars. @GGP07 studied Zr as well as Li in Galactic OH/IR stars, which are oxygen-rich AGB stars with relatively high masses (3-4 $M_\odot$). These stars are expected to show HBB, the CNO cycle at the bottom of hydrogen-rich envelope, that makes the C/O ratio lower than unity and produces high Li abundances. Although the effect of HBB is confirmed by the Li over-abundances in a half of their sample, no excess Zr was found in these cases. This led to the conclusion that the high mass AGB stars in our Galaxy do not show any significant $s$-process enrichments, in contrast to the results derived for AGB stars in the Magellanic clouds, which are lower in metallicity by a factor of two or three than Galactic objects. @PSL93 and @SPL95 found effects of HBB for luminous (massive) AGB stars in the Magellanic clouds that show $s$-process enhancement as well. These observations confirm that the $s$ process in massive AGB stars depends strongly on metallicity. The heavy neutron-capture elements in planetary nebulae have been recently studied by optical and near-infrared spectroscopy. @SZW07 measured weak emission lines in the optical range for five planetary nebulae. These observations included neutron-capture elements, for which the required atomic data for a reliable abundance determination became available recently. The discovered excesses of Kr and Xe in three objects can be assigned to the first and second $s$-process abundance peaks, although a large correction for the $r$-process contribution is required for Xe. These two elements are enhanced by a similar factor in three planetary nebulae (\[Xe/Kr\]$\sim 0$), suggesting the effect of an $s$ process with a significant neutron-exposure. @StD08 measured abundances of the light neutron-capture elements Kr and Se for a large number of planetary nebulae (81 objects for Kr and 120 for Se). The abundances were determined from emission features in near infrared (2.2 $\mu$m) and showed that 44% of the sample, which corresponds to 20% of all planetary nebulae in the Galaxy, are $s$-process-enriched (\[Kr, Se/Ar\] $>0.3$). Contribution to the Galactic chemical evolution {#sec4E} ----------------------------------------------- The stars in the substructures of the Milky-Way, the thin and thick disks, the bulge, and the halo, differ with respect to metallicity and kinematical properties, i.e. their orbital motion around the Galactic center. The formation timescale of the Galactic structure is usually estimated by the abundance ratios between the $\alpha$ elements, e.g. Mg, and iron [@Mcw97]. The thin disk is the relatively new component of the Galaxy, and the Sun is involved in this structure. The metallicity is similar to the solar one, while the ages of stars range between zero to ten billion years. The thick disk consists of old stars of lower metallicity. The formation of this component is still debated, a possible scenario being the burst-like star formation when a small galaxy merged with the Milky Way at some early epoch of the Galactic history [e.g. @FrB02]. The bulge consists mostly of old stars with a broad metallicity distribution, including stars of solar metallicity. It has been suggested that the bulge was formed at very early times of the Universe and its evolution during Galactic history has been discussed. So far, abundance measurements of neutron-capture elements for the bulge are still quite limited because of its long distance and severe interstellar extinction. Therefore, future studies with larger telescopes are required. The halo structure consists of old, metal-deficient stars and extends to distances of 100 kparsec around the disks. The formation timescale is estimated to be one or at most a few billion years after the Big Bang. About 150 globular clusters have been found in the Galaxy, which also belong to the old population of the Galaxy and are believed to be related to the formation of the halo and bulge structures. There are tens of satellite dwarf galaxies around the Milky Way, which probably formed and evolved by interactions with the Galaxy. Spectroscopy for individual stars in such dwarf galaxies has revealed significant $s$-process contributions in some cases, providing observational constraints on the $s$ process at different metallicities and their roles in chemical enrichment [@THT09]. ### $s$-Process contributions to Galactic field stars \[sec4E1\] Because the lifetimes of low- to intermediate-mass stars, the progenitors of AGB stars that are responsible for the main component of the $s$ process, are longer than for massive stars ($M\gtrsim 8-10 M_{\odot}$), it is not expected that the main $s$ process deriving from AGB stars of all masses and metallicities contributed to the chemical evolution of the early Galaxy for \[Fe/H\] $< -1.5$ [@TGA04]. As for the weak $s$ process, which is to be ascribed to the presupernova evolution of massive stars, we point out that, besides the present MACS uncertainties in the region below $A = 90$ (see Sec. \[sec3B\]), neutron capture in massive stars is driven by the $^{22}$Ne($\alpha, n$)$^{25}$Mg reaction, where $^{22}$Ne acts as a secondary-like source. In fact $^{22}$Ne comes from the original CNO abundances, which are transmuted into $^{14}$N during H burning and then converted to $^{22}$Ne by the chain $^{14}$N($\alpha, \gamma$)$^{18}$F($\beta^+\nu$)$^{18}$O($\alpha, \gamma$)$^{22}$Ne in the early phase of core He burning (@RGB93 and references therein). Consequently, also the weak $s$ process is not expected to play any role in Galactic Halo stars. The weak $s$ process is also believed to occur in massive stars (Sec. \[sec3C\]). However, the process requires high metallicities and is, therefore, effective only in young, metal-rich stars. Although the abundance patterns produced by the weak $s$ process are not clearly found in stellar atmospheres, some constraints have been obtained from Cu, Zn, Ga, and Ge abundances in Galactic stars [@PGH10]. With respect to Zn, the most abundant isotope $^{64}$Zn (48.6% of solar Zn) derives from SNe of Type II in the $\alpha$-rich freezout of neutrino winds [@WoH92] or in hypernovae [@UmN02] while the other 50% of solar Zn are almost fully ascribable to the weak $s$ process (Bisterzo et al. 2004). Assuming that SNII produce about 1/3 of solar Fe, this means that the ratio \[Zn/Fe\] in the Halo should be a bit positive, about 0.2 dex on average, consistent with spectroscopic observations [e.g. @CDS04]. Concerning Cu, the weak $s$ process accounts for 90% and the main $s$ component for 5% of the solar abundance, whereas SN Ia are not predicted to contribute any Cu [@TNY86]. This implies that \[Cu/Fe\] in the halo should be constant and strongly negative, around $-$0.8 dex. The origin of this small primary Cu contribution may be ascribable to the explosive nucleosynthesis in massive stars. These expectations are confirmed by theoretical expectations of @WoW95 for a range of massive stars with metallicities from 0 to solar. A large number of spectroscopic observations of \[Cu/Fe\] versus \[Fe/H\] exist in the literature, again confirming the above expectations [e.g. @BGP04; @RoM07; @PGH10]. The nucleosynthetic origin of primary Zn and Cu is still a subject of debate. This may imply different processes, e.g. the so-called $\nu p$ process [@FML06]. Other light elements beyond the Fe group, like Ga and Ge, should behave like Cu, that is with a major weak and secondary-like $s$-process contribution [@PGH10]. For the few spectroscopic observations available in Halo stars, we refer to @CSB05. Although the astrophysical sites of the $r$ process are not well identified, massive stars that terminate their lives by core-collapse supernovae would be promising candidates for the progenitor. Due to the short lifetimes of massive stars it appears plausible that the $r$ process contributed significantly to the enrichment of the early Galaxy. Apart from the $r$ process and weak $s$ process, recent studies on light neutron-capture elements suggest a LEPP as an additional source of these elements in the very early Galaxy [@TCP02; @TGA04; @AHB05] as discussed in Sec. \[sec3\]. The abundances of Ba (or La) and Eu are used as indicators of the $s$- and $r$-process contributions to the origin of the heavy elements. Analyses of the abundances in the Solar System showed that about 80% of Ba [@TGG99; @SGT09] and 70% of La [@WDH06] originate from the $s$ process, while about 95% of Eu comes from the $r$ process [@KBW89; @AKW99b]. Heavy neutron-capture elements of metal-poor stars in the halo have been recently studied with high resolution spectroscopy [@MPS95; @RNB96; @BPA00; @SSC04; @HAK04; @AHB05; @FDH07; @SCG08]. On the other hand, the chemical composition of a large sample of disk stars has been studied in the past two decades [e.g. @EAG93]. Recently, @RTL03 and @RLA06 determined abundances of many elements including Y, Ba and Eu for 181 thin disk stars and 95 thick disk stars, respectively, based on high quality spectra. The elemental abundance ratios for Ba and Eu, as a function of metallicity, are shown in Fig. \[fig:19\], which includes also examples for theoretical GCE expectations. The distributions were obtained by considering the $s$ and $r$ components separately. The Galactic abundances of these elements were computed from the sum of both processes by comparing model results [@SGT09; @TGA04] with the available spectroscopic observations of field stars at different metallicities. For the $s$ component of each isotope at the epoch of the formation of the Solar System only the contributions of AGB stars are considered. Subsequently, the $r$-process residual method ($r = 1- s$) was used to determine the respective solar system $r$-process fractions, assuming that the production of $r$ nuclei is a primary process occurring in Type II supernovae, independent of metallicity. Given the present theoretical uncertainties of the $r$-process modeling, the $r$-process residual method is to be considered as an approximation of the fractional $r$ abundances in Solar System material. It does not exclude, however, that some spread in the $r$-process ratios, e.g. \[Ba/La\]$_r$, \[Ba/Eu\]$_r$, \[Eu/Pb\]$_r$, \[Eu/Th\]$_r$, may exist in low metallicity stars, indicating a multiplicity of $r$-process components. Significant examples concerning \[Eu/Th\]$_r$ are already reported from spectroscopic observations [@PHC04; @RKF09]. Concerning the ratio \[Ba/Eu\]$_r$, a certain spread is apparent in Fig. \[fig:19\], but a more detailed analysis of the spectroscopic data and the related uncertainties reported by different authors might be be necessary. From the theoretical point of view, a range of $r$-process predictions has been advanced in recent works by @KFP07a [@FKP10]. ![(Color online) Top panel: The evolution of the $s$-process fractions of \[Ba/Fe\] versus \[Fe/H\] in the Galactic halo as well as in the thick and thin disk (dashed lines) and theoretical predictions of the total $s/r$ abundances (solid lines) from @TGA04 [@SGT09]. The spectroscopic data from observations of Galactic disk and halo stars are collected from the literature [@EAG93; @GrS94; @MPS95; @Mcw98; @JMN99; @ToL99; @BPA00; @Ful00; @NRB01; @MaG01; @MiK01; @MaZ06; @ISS06; @ABS08; @AoH08; @LBJ08; @CMC07; @NCK07; @FCN07; @MZG08; @RLS08; @AHB05; @FDH07; @CCM08; @AFC06; @YGG05; @VGJ03; @CSB02]. Error bars are plotted only when reported for individual objects by the authors. The dotted line connects a star observed by different authors. Analogous plots are shown for \[Eu/Fe\] (middle panel) and \[Ba/Eu\] (bottom panel).\[fig:19\]](rmp_fig19.eps){width="13cm"} The top and middle panels refer to the typical $s$- and $r$-process elements \[Ba/Fe\] and \[Eu/Fe\], whereas the bottom panel shows their ratio \[Ba/Eu\]. Theoretical GCE expectations using only the $s$-process products from AGB stars in the Galactic halo as well as in the thick and thin disk are separately indicated by dashed lines. Theoretical predictions of the total ($s + r$) yields are shown as solid lines. We recall that the elemental composition of the $r$ process has been obtained via the $r$-residual method described before. Below \[Fe/H\] $< -1.5$ the $r$ process dominates the theoretical expectations. According to our prescriptions the $r$ process is to be considered of primary origin, i.e. to originate from reactions starting from H and He. However, as discussed in @TGA04, calculations of the GCE trend versus metallicity have been made assuming that only a small range of massive stars, with initial masses of $8-10$ $M_\odot$ are involved in the $r$-process production. This implies that \[Ba/Fe\] as well \[Eu/Fe\] decrease below \[Fe/H\] $< -2.3$, but also other choices may be invoked to explain the general decrease of spectroscopic observations. As to the large observed spread of \[Ba,Eu/Fe\], an easy explanation could be that at those metallicities the interstellar medium in the Halo was not fully homogenized (Travaglio, Galli and Burkert 2001, Raiteri et al. 1999, Ishimaru and Wanajo 1999, Ishimaru et al. 2004). Notice that the observed ratio \[Ba/Eu\] stays almost flat (bottom panel of Fig. \[fig:19\]). In conclusion, the \[Eu/Fe\] shown in Fig. \[fig:19\] is basically explained by the $r$ process. In the metal-rich range, a decreasing trend of \[Eu/Fe\] with increasing metallicity is found as in the case of \[$\alpha$/Fe\] [e.g. @Mcw97]. Actually, the decreasing trend of \[Eu/Fe\] in the disk versus higher metallicities is not due to a decreasing efficiency of the $r$ process, but to an increasing apport of Fe in the interstellar medium from the long-lived SNe of type Ia. By contrast, the model curves for \[Ba/Fe\] (top panel of Fig. \[fig:19\]) indicate that the main $s$ process is the dominant contributor to Ba in the metal-rich range (\[Fe/H\]$>-1$). Indeed, there is no decreasing trend of \[Ba/Fe\] with increasing metallicity. The models also indicate that the Ba in metal-poor stars is provided by the $r$ process as in the case of Eu. \[La/Eu\] is another indicator for the relative contributions of the two processes. @SSC04 suggested a correlation with kinematical properties, because it was found that stars with high \[La/Eu\] show low velocities with respect to the Galactic plane. Therefore, these stars possibly indicate the existence of a substructure in the halo to which the $s$ process has contributed. In view of the long time scale of the $s$-process effects, such a substructure has an impact on the formation history of the halo. On average, stars in the thick disk have lower metallicity than thin disk stars. Low metallicity thick disk stars have 0.3-0.5 dex higher \[Eu/Fe\] than the solar value, while \[Ba/Fe\] is slightly lower. The overabundance of Eu implies that the formation time scale of the thick disk is shorter than that of type Ia supernovae (approximately a few billion years), which is also supported by the abundance ratios between $\alpha$ elements and iron. The \[Ba/Eu\] ratios of $-$0.5 to $-$0.7 agree with the $r$-process composition, indicating that the dominant source of heavy neutron-capture elements in thick disk stars is indeed the $r$ process. The absence or minor contribution of the $s$ process supports the rapid formation of the thick disk. However, there are several thick disk stars with higher metallicity than thin disk stars, which have similar \[Ba/Fe\] to solar. Their existence indicates that star formation continued longer in the thick disk, although the fraction of such stars is quite small. ![(Color online) Same as Fig. \[fig:19\], but for \[Y/Fe\] (top) and \[Pb/Fe\] (bottom). The observational data of Y are from the references given in Fig. \[fig:19\]. The Pb data are from @SCB98 [@TGB01b; @RLS08; @RKF09; @ISS06; @AoH08].[]{data-label="fig:20"}](rmp_fig20_new.eps){width="11cm"} Fig. \[fig:20\] shows the light and heavy $s$-process elements \[Y/Fe\] and \[Pb/Fe\] as a function of metallicity. Pb, which is formally produced by the $s$ process in low metallicity AGB stars (Sec. \[sec3E\]), represents a good $s$-process indicator. @TGB01b provided model predictions of Pb enrichments in the Galaxy compared to observations in several metal-poor stars, but the observational constraint was rather weak due to the small sample size. Recently, Pb abundances of 12 giants in the halo were determined by @ABS08. The \[Pb/Fe\] and \[Pb/Eu\] ratios of these stars are constant within the observational uncertainties, and the \[Pb/Eu\] values agree with the inferred $r$-process component in the Solar System, clearly indicating that there is no significant $s$-process contribution in these stars (see also @RKF09 for the latest compilation). However, the sample for \[Fe/H\]$>-1.5$, where the $s$ process is expected to become significant, is still scarce. Further measurements of Pb abundances for halo stars, particularly in those stars with high \[La/Eu\], will be useful for understanding of the formation time scale of the halo. The situation is more complicated in the enrichment of the light neutron-capture elements. Fig. \[fig:20\] (top panel) shows \[Y/Fe\] as an example. As discussed in Sec. \[sec3E\], the Galactic chemical enrichment of Y by the apport of all previous generations of AGB stars at the epoch of the solar system formation explains only 67% of the solar Y abundance [@TGA04; @SGT09]. According to that discussion, the solid line in the plot includes contributions of 67%, $5-10$%, about 8%, and $15-20$ % from the main and weak $s$ process, from the $r$ process, and from the LEPP, respectively. The flat theoretical GCE prediction of \[Y/Fe\] = $-$0.2 dex in the Halo is based upon the assumption that the primary LEPP component is obtained in all massive stars, which are exploding as SNII. We recall instead that the primary $r$ process was assumed to derive from a small range of massive stars, between 8 and 10 $M_\odot$ (see discussion for the \[Eu/Fe\] trend and Fig. \[fig:19\]. The \[Y/Fe\] in extremely metal-poor stars (\[Fe/H\]$<-3$) shows very large scatter, indicating a diversity of the relative contributions by the LEPP and the $r$ process, as well as incomplete mixing in the gas cloud from which these stars have formed. The Y abundances of thick disk stars are similar to the stars in the thin disk. However, the origin of Y would be significantly different, given the different \[Y/Eu\] ratios of the two structures. The dominant source of this element in thin disk stars, as in the Solar System, is the (main) $s$ process [@TGA04]. On the other hand, the $r$ process and the LEPP should be significant sources of Y in thick disk stars. The thick disk is currently assumed to have suffered a rapid evolution and a high star formation rate, such that contributions from SNIa or from the main $s$ process by AGB stars should be absent. Due to the high star formation rate, however, higher \[Fe/H\] values are observed compared to the Halo. Indeed, \[Eu/Fe\] is almost flat with an average of 0.5 dex, the same as observed in the Halo. A similar behavior is found for the so-called $\alpha$-enhancement. Therefore, we may expect \[Y/Fe\] = $-$0.2 dex, as discussed before in the analysis of Fig. \[fig:20\] (top panel). Consequently, the small variation of \[Y/Fe\] in thick disk stars with respect to the thin disk is not surprising. ### Globular clusters and galaxies in the local group \[sec4E2\] The Ba and Eu abundances measured in globular clusters have been summarized by @GSC04. Fig. 6 of their paper indicates that there are no significant variations in the abundances of Ba and Eu compared to the observational uncertainties. One remarkable exception is the metal-poor cluster M15, which shows significant star-to-star scatter of Ba and Eu abundances [@SCK97], presumably as the result of “local” massive star nucleosynthesis. In general, however, globular clusters are modestly $r$-process-rich compared with Solar System material. There are several exceptional clusters that exhibit significant $s$-process contributions. A few clusters show variations in metallicity and abundance ratios. Among those, $\omega$ Cen represents a remarkable case. This cluster has a metallicity distribution between \[Fe/H\]$=-1.8$ and $-0.8$, implying chemical evolution inside the cluster. Studies of neutron-capture elements for a significant sample of cluster stars by @NoD95 and @SSC00 have shown that the \[Ba/Fe\] and \[La/Fe\] ratios increase with increasing metallicity. This behavior can be explained by the contributions of the $s$ process in 1–3 M$_{\odot}$ AGB stars. Accordingly, star formation in this cluster continues over a period, which is longer than the lifetime of these stars, i.e. longer than one billion years. Other exceptions are clusters without significant variation of chemical abundances but with an enhancement in $s$-process elements. A well known example is M4 (\[Fe/H\]$=-1.2$), where \[Ba/Fe\] is 0.6 dex higher than in other clusters with similar metallicity [@ISK99]. The fact that no variations of metallicity and chemical abundance ratios are observed suggests that the origin of these heavy elements is primordial. In other words, the cluster forming cloud was already polluted by the ejecta of previous generations of stars. This is a strong constraint for the formation timescale of the cluster. A few other clusters are also suggested to have excesses of $s$-process elements, though they are not as clear as in M4. A good probe of the $s$-process contribution is Pb. Abundances of this element have been measured in four clusters including M4 by @YAL06 and @YLP08. In M4, Pb was found to be enhanced as expected from other $s$-process elements, but no such excess was found for three other clusters, which showed Pb abundances compatible with that in halo stars [@ABS08]. Heavy element abundances have been measured for bright stars (supergiants and stars at the red giant branch tip) in the Magellanic clouds, the irregular satellite galaxies of the Milky Way. Recently, @PHS08 reported chemical abundances of a large sample of disk stars in the Large Magellanic cloud. Y, Zr, Ba, and La abundances were determined for 30–50 red giant stars covering a metallicity range of $-1.3<$\[Fe/H\]$<-0.3$. The heavy neutron-capture elements Ba and La turned out to be overabundant, while the ls elements Y and Zr are underabundant. Although the measurement of Eu was not available for the sample, Ba and La are likely of $s$-process origin. The high abundance ratios between the hs and ls elements are compatible with the model prediction for the $s$ process in metal-deficient AGB stars. Measurements of heavy elements in stars of dwarf spheroidal galaxies around the Milky Way have been made in the past decade [e.g. @SCS01; @SVT03; @SAI04]. Although the sample size is still small, in particular for the measurement of Eu abundances, the neutron-capture elements in metal-deficient (\[Fe/H\]$<-1.0$) stars in dwarf galaxies can be explained by the $r$ process, while $s$-process contributions are seen in metal-rich stars of some galaxies, e.g. in Fornax and Carina [@SVT03; @THT09]. The $s$-process contributions from low- and intermediate-mass AGB stars are an important probe for the star formation history of galaxies. The $\alpha$/Fe abundance ratios of stars in dwarf galaxies are lower than those of halo stars in general, indicating the possible role of type Ia supernovae. This is unlikely the case in galaxies without significant $s$-process contribution, because the time scales for type Ia supernovae are longer than the evolution of intermediate-mass stars. Summary ======= The $s$-process, which is ascribed to low mass stars during the TP-AGB phase (main and the strong components), and to massive stars (the weak component), is discussed with respect to the underlying nuclear physics, current stellar models, and the rapidly growing observational evidence. From the nuclear physics side, there is an increasingly complete set of neutron capture measurements that provide the necessary Maxwellian averaged cross sections for detailed network calculation of the $s$-abundance patterns of the various scenarios. Over the last two decades considerable experimental progress was achieved on the basis of new and improved neutron facilities and detector developments. Although the productive facilities at Oak Ridge and Karlsruhe have been closed recently, replacement became available through intense pulsed neutron sources using spallation reactions (n\_TOF at CERN, J-PARC in Japan, and LANSCE at Los Alamos) or the $^{7}$Li($p, n$) reaction (FRANZ at Frankfurt/Germany and SARAF at the Weizmann Institute in Jerusalem, both under construction). Apart from the high fluxes, which these facilities have in common, they exhibit widely complementary characteristics, thus providing promising solutions for a variety of improved TOF measurements. Such measurements benefit also from developments in detector technology, aiming at higher efficiency (total absorption calorimeters) or minimized neutron sensitivity. Combined with new data acquisition systems and rapidly growing computing power, a new generation of experiments has already provided a number of very accurate cross sections at astrophysically relevant energies. In parallel, the activation method proved to play an important role because of the superior sensitivity, which enabled first measurements on unstable branch-point nuclei along the $s$-process path. Future efforts in $s$-process experiments would be most useful in the following areas: Improvements in the accuracy of ($n, \gamma$) cross sections are needed in mass regions where present uncertainties are still exceeding the $3-5$% level, i.e. around magic neutron numbers, in the Fe-Sr region, and for the lighter elements. The persisting problem of the cross sections for the neutron source reactions $^{13}$C($\alpha, n$)$^{16}$O and $^{22}$Ne($\alpha, n$)$^{25}$Mg in the respective Gamow windows is still unsolved. Together with the abundant light elements, which act as neutron poisons, the source reactions determine the $s$-process neutron balance and represent, therefore, important constraints for stellar models, i.e. for the role of the $^{13}$C pocket in thermally pulsing low-mass AGB stars. The scarce information for ($n, \gamma$) cross sections of unstable isotopes, which are crucial for the analysis of $s$-process branchings, need to be completed. The feasibility of such measurements will benefit from progress in neutron facilities and advanced experimental techniques. Ultimately, they will also be needed to treat the extended reaction paths into the neutron rich region, which follow from the high neutron densities during C shell burning in massive stars and during the first, strong pulse in low metallicity AGB stars. Last but not least, theory remains indispensable for complementing the experimental ($n, \gamma$) information, either by closing gaps in the data, where measured cross sections are not (yet) available or by providing stellar enhancement factors to correct the laboratory results for the effect of thermally populated excited nuclear states in the hot stellar plasma. Correspondingly, the even more pronounced enhancement of the weak interaction rates as a function of neutron and electron density at the $s$-process sites remains an important domain of theoretical studies, especially because experimental work in this field had been long neglected. The weak $s$ process, which produces a large fraction of the $s$ isotopes between Fe and Sr during convective core He burning and subsequent convective shell C burning, is of secondary nature. The neutron source is driven by ($\alpha, n$) reactions on $^{22}$Ne deriving from the conversion of initial CNO nuclei to $^{14}$N during core H burning via the sequence $^{14}$N($\alpha, \gamma$)$^{18}$F($\beta^+ \nu$)$^{18}$O, and subsequently by $^{18}$O($\alpha, \gamma$)$^{22}$Ne reactions at the beginning of convective core He burning, when the central temperature raises above $2.5 \times 10^8$ K. The weak $s$ contribution to the solar abundances is not easy to estimate in a quantitative way due to the present uncertainties of the stellar cross sections in the range from Fe to Se and to the physical uncertainties in the treatment of the pre-explosive and explosive nucleosynthesis in supernovae. However, half of solar Zn and about $70-80$% of solar Cu, Ga, Ge, and As are to be ascribed to the weak $s$ process. The interplay between theory and spectroscopic observations is briefly discussed. All $s$ isotopes beyond $A = 90$ and of about half of solar Pb are contributed by the main $s$ process. The second half of solar Pb is produced by low mass AGB stars at low metallicities (strong component). Below $A=90$ the contribution of the main component to the $s$-process abundances decreases rapidly. The main $s$ process is not a unique process, but depends on the initial mass, metallicity, the strength of the $^{13}$C-pocket, the efficiency of the TDU, and the choice of the mass loss rate. Stellar models could be verified by comparison with a large body of data obtained from analyses of presolar material in form of circumstellar dust grains and by the conspicuous number of observations of MS, S, C(N) and Ba stars of the Galactic disk as well as of CH stars in the halo. The $s$-process contribution to the cosmic abundances in the interstellar medium in the mass range $A > 90$ is the result of all previous generations of AGB stars that polluted the interstellar medium before the formation of the Solar System. The impact of the chemical evolution of the Galaxy is analyzed for the representative elements Y, Ba, and Pb corresponding to the $s$-abundance peaks at magic neutron numbers 50, 82, and 126. The origin of the heavy neutron-capture elements is partly due to the main $s$ process and partly to the primary $r$ process. Usually the $r$-process contribution to each isotope in the solar system is estimated using the so-called $r$-residual method by subtracting the well defined $s$-process components from the respective solar abundances. This approach provides a fair basis for comparison with the still uncertain predictions from current $r$-process models. Progress in stellar modeling depends on a continuous interplay between theory and observation. This is particularly true for recent observations of the rare class of carbon-enhanced metal-poor stars with $s$-process enhancements, the CEMP-$s$ stars, which are main sequence, turnoff, or giant stars of low mass ($M \sim 0.8 M_\odot$) in close binary systems. The primary more massive companion (now a white dwarf) evolved along the AGB and polluted the envelope of the observed star with C and $s$-process elements when it lost its entire envelope at the end of the AGB phase. A strongly debated issue is the subclass of CEMP-$s/r$ stars, showing $s$- and $r$-process contributions at the same time, although both processes are of completely different astrophysical origin. Some of them are highly enhanced in Ba, Ce, and La, which belong to the second $s$-process peak at $N=82$, as well as in Eu, which is a typical $r$-process element. In fact, the $s$ and $r$ elements beyond Ba are enhanced at the same level in these very metal-poor stars. A plausible scenario considers the formation of binary systems in giant clouds that were locally polluted by the ejecta of type II supernovae. Apart from the contributions from the weak and main $s$ process and from the $r$ process, the light $s$-process elements Sr, Y, and Zr in the solar system contain an additional component contributed by a primary source of still unknown origin, the Light Element Primary Process (LEPP). The different proposed hypotheses, all related to the most advanced phases of pre-explosive and explosive nucleosynthesis in massive stars, represent a most relevant issue of present nucleosynthesis research. These intriguing problems have been recognized by the rapid increase of observational data in the past decade, using high-resolution and high signal-to-noise spectra. Further chemical abundance studies, in particular isotope abundance measurements for key elements, will provide useful hints and constraints for understanding the physical processes behind these unsolved problems. Further improvement of the models related to the weak, main, and strong $s$-process components coming from massive and intermediate-to-low mass stars, will have a strong impact on studies of the Milky Way and surrounding smaller galaxies. Great efforts have been made to understand the chemical evolution and formation history of these galaxies, which are traced by chemical abundance ratios as well as kinematic properties of individual stars. Abundance ratios of $s$-process elements provide useful constraints for the chemical evolution models, i.e. on the time scale of the star formation history and on the initial mass function. Acknowledgements ================ We are deeply indebted to O. Straniero and S. Cristallo for continuous clarifying discussions concerning the modeling of AGB stars. We would like to thank also the referees for their suggestions, which led to a substantial improvement of the manuscript. [391]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , , , , , , , *et al.*, , **, , , . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , ****, . , , and , , ****, . , , , , , , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , and , , ****, . , and , , ****, . , , , , , , , , , , , and , , ****, . , , , , , , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , ****, . , , , , , , , , and , , ****, . , , , , , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , and , , ****, . , , , , , , , , , and , , ****, . , , , and , , ****, . , , , , and , , ****, . , , , , and , , ****, . , , , , , , , , and , , ****, . , , , , , , , , and , , ****, . , , , , , and , , ****, . , , , , , , and , , ****, . , and , , ****, . , and , , ****, . , , ****, . , , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , ****, . , and , , ****, . , , , , , , , , , and , , ****, . , , in **, edited by , , , , and (, ), p. . , and , , ****, . , , and , , ****, . , and , , ****, . , , and , , ****, . , , and , , ****, . , and , , ****, . , , and , , ****, . , , , , , and , , ****, . , , **, , , . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , and , , ****, . , and , , ****, . , , , , , and , , ****, . , , , , , and , , , ****, . , , , , , and , , ****, . , , ****, . , , , , , and , , in **, edited by , , , and (, ), pp. , . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , and , , in **, edited by , , and (, ), p. . , , , , and , , **, , , . , , , and , , ****, . , , , , , and , , ****, . , , , , and , , ****, . , , and , , ****, . , , , , , and , , ****, . , , **, , . , , and , , in ** (, ), AIP Conf. Proceedings, Volume 402, pp. . , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , in **, p. , . , and , , ****, . , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , ****, . , , ****, . , , , and , , ****, . , , , , and , , ****, . , and , , ****, . , and , , ****, . , and , , ****, . , , , , , , , and , , ****, . , , , and , , ****, . , , , , , , , and , , ****, . , , , , , , , , and , , ****, . , , , and , , ****, . , , and , , ****, . , , , . , and , , ****, . , and , , ****, . , , , , , , , , and , , ****, . , , , , , , , , , , and , , ****, . , , , , and , , ****, . , , , , , and , , ****, . , , , , , and , , ****, . , , , and , , ****, . , and , , ****, . , , in **, edited by , , , and (, ), p. . , and , , ****, . , , , . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , in **, edited by and (, ), p. , , . , , , and , , in ** (, ), . , , , and , , ****, . , and , , ****, . , and , , ****, . , and , , ****, . , and , , ****, . , , , , , , , and , , ****, . , , , , , and , , ****, . , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , and , , ****, . , , **, , . , , , , , and , , ****, . , , , , , , and , , ****, . , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , and , , ****, . , and , , ****, . , , ****, . , , and , , ****, . . , , and , , ****, . , and , , ****, . , and , , ****, . , , , and , , ****, . , , , and , , ****, . , , , , , , , , , , , , *et al.*, , ****, . , , , , , , , , and , , ****, . , , , , and , , ****, . , , ****, . , , ****, . $s$-only isotopes Abundance ratio [@RoT98] Branching ratio ------------------- --------------- -------------------------- ----------------- --------------- $^{80,82}$Kr 267$\pm$14 90$\pm$6 2.28/11.58 0.61$\pm$0.05 $^{122,124}$Te 295$\pm$3 155$\pm$2 2.55/4.74 1.06$\pm$0.02 $^{128,130}$Xe 262.5$\pm$3.7 132.0$\pm$2.1 1.92/4.08 0.96$\pm$0.02 $^{134,136}$Ba 176.0$\pm$5.6 61.2$\pm$2.0 2.417/7.854 0.94$\pm$0.04 $^{148,150}$Sm 241$\pm$2 422$\pm$4 11.24/7.38 0.88$\pm$0.01 : MACS results at $kT$=30 keV for elements with a pair of $s$-only isotopes. (Data are from KADoNiS [@DHK05], see Sec. \[sec2D\]). \[tab1\] ---------------- --------------- ------------------- ----------------- ------------------- Target isotope $^{58}$Fe 12.1$\pm$1.1 [@KWH83] 13.5$\pm$0.7 [@HKU08a] $^{59}$Co 38$\pm$3 [@SpM76] 39.6$\pm$2.7 [@HKU08a] $^{62}$Ni 25.8$\pm$3.7 [@ABE08] 20.2$\pm$2.1 [@NPA05; @Dil09a] 37.0$\pm$3.2 [@TTS05] 23.4$\pm$4.6 [@DFK10] $^{87}$Rb 15.5$\pm$1.5 [@JaK96b] 15.8$\pm$0.9 [@HKU08b] $^{88}$Sr 6.01$\pm$0.17 [@KWG00; @KWG01a] 6.13$\pm$0.18 [@KZB90] $^{89}$Y 21$\pm$3 [@MAM78c] 19.0$\pm$0.6 [@KZB90] $^{139}$La 32.4$\pm$3.1 [@TAA07] 31.6$\pm$0.8 [@BDH03] $^{146}$Nd 91.2$\pm$1.0 [@WVK98a] 87.1$\pm$4.0 [@TDK95] $^{148}$Nd 146.6$\pm$1.9 [@WVK98a] 152$\pm$9 [@TDK95] $^{176}$Lu 1639$\pm$14 [@WVK06a] 1599$\pm$85$^*$ [@BeK80] $^{180}$Hf 156.5$\pm$1.9 [@WVK06b] 168$\pm$9 [@BKW82] $^{197}$Au 588$\pm$20 [@MHW75; @Mac82a] 592$\pm$9 [@RaK88] ---------------- --------------- ------------------- ----------------- ------------------- : Recent MACS results at $kT$=30 keV obtained via TOF techniques and with the activation method. (Data are from KADoNiS [@DHK05], see Sec. \[sec2D\]). \[tab2\] $^*$Original value renormalized by @BBK00. ------------ ------------- ------------------ ------------------------------------------------------------------------------------ Sample Half life $Q$ value Comment (yr) (MeV) $^{63}$Ni 100.1 $\beta^-$, 0.066 TOF work in progress [@Cou09], sample with low enrichment $^{79}$Se 2.95 $10^5$ $\beta^-$, 0.159 important branching, constrains $s$-process temperature in massive stars $^{81}$Kr 2.29 $10^5$ EC, 0.322 part of $^{79}$Se branching $^{85}$Kr 10.73 $\beta^-$, 0.687 important branching, constrains neutron density in massive stars $^{95}$Zr 64.02 d $\beta^-$, 1.125 not feasible in near future, but important for neutron density low mass AGB stars $^{134}$Cs 2.0652 $\beta^-$, 2.059 important branching at $A=134,135$, sensitive to $s$-process temperature in low-mass AGB stars, measurement not feasible in near future $^{135}$Cs 2.3 $10^6$ $\beta^-$, 0.269 so far only activation measurement at $kT=25$ keV by @PAK04 $^{147}$Nd 10.981 d $\beta^-$, 0.896 important branching at $A=147/148$, constrains neutron density in low-mass AGB stars $^{147}$Pm 2.6234 $\beta^-$, 0.225 part of branching at $A=147/148$ $^{148}$Pm 5.368 d $\beta^-$, 2.464 not feasible in near future $^{151}$Sm 90 $\beta^-$, 0.076 existing TOF measurements, full set of MACS data available [@WVK06c; @AAA04c] $^{154}$Eu 8.593 $\beta^-$, 1.978 complex branching at $A=154,155$, sensitive to temperature and neutron density $^{155}$Eu 4.753 $\beta^-$, 0.246 so far only activation measurement at $kT=25$ keV by @JaK95b $^{153}$Gd 0.658 EC, 0.244 part of branching at $A=154,155$ $^{160}$Tb 0.198 $\beta^-$, 1.833 weak temperature-sensitive branching, very challenging experiment $^{163}$Ho 4570 EC, 0.0026 branching at $A=163$ sensitive to mass density during $s$ process, so far only activation measurement at $kT=25$ keV by @JaK96a $^{170}$Tm 0.352 $\beta^-$, 0.968 important branching, constrains neutron density in low-mass AGB stars $^{171}$Tm 1.921 $\beta^-$, 0.098 part of branching at $A=170,171$ $^{179}$Ta 1.82 EC, 0.115 crucial for $s$-process contribution to $^{180}$Ta, nature’s rarest stable isotope $^{185}$W 0.206 $\beta^-$, 0.432 important branching, sensitive to neutron density and $s$-process temperature in low-mass AGB stars $^{204}$Tl 3.78 $\beta^-$, 0.763 determines $^{205}$Pb/$^{205}$Tl clock for dating of early solar system ------------ ------------- ------------------ ------------------------------------------------------------------------------------ : Feasibility of future TOF measurements on unstable branch point isotopes at the FRANZ facility. \[tab3\] [^1]: Such stars are labeled $r$-II by @CBB04, defined as those very metal-poor (\[Fe/H\] $<$ $-$2.5) stars with \[Eu/Fe\] $>$ 1.0 and \[Eu/Ba\] $<$ 0.0 (e.g., CS 31082-001 by @HCB02, CS 22183-031 by @HAK04, HE 1523-0901 by @FCN07, CS 29497-004 by @CBB04). A few other stars, such as BD+17 3248 [@CSB02] and CS 30306-132 [@HAK04], would formally be classified in this system as $r$-II stars (+0.3 $\leq$ \[Eu/Fe\] $\leq$ +1.0 and \[Eu/Ba\] $<$ 0.0).

ArXiv

--- abstract: 'A subgroup $H$ of a compact abelian group $X$ is said to be characterized by a sequence $(v_n)$ of characters of $X$, if $H = \left\{x\in X:v_n(x)\to0 \text{ in }{\mathbb{T}}\right\}$. We discuss various known facts and open problems about the Borel complexity of characterized subgroups of compact abelian groups.' author: - Dikran Dikranjan and Daniele Impieri title: Questions on the Borel Complexity of Characterized Subgroups --- Introduction ============ In the sequel ${\mathbb{T}}={\mathbb{R}}/{\mathbb{Z}}$ will be denote the circle group written additively, identifying each $x\in {\mathbb{T}}$ with an obviously determined element of the interval $[0,1)$. In these terms we consider the norm in ${\mathbb{T}}$ defined by $\|x\| = \min\{x, 1-x\}$ for $x\in {\mathbb{T}}$. It defines an invariant metric $d$ on ${\mathbb{T}}$ by letting $d(x,y) = \|x-y\|$, for $x,y\in {\mathbb{T}}$. For a topological abelian group $X$, a continuous homomorphism $\chi:X\to{\mathbb{T}}$ is called a *character* of $X$. Denote by $\widehat{X}$ the group of all characters of $X$, that is the Pontryagin dual of $X$. For a compact abelian group $X$ and a sequence ${\mathbf{v}}=(v_n)_{n\in{\mathbb{N}}}$ of characters of $X$ let $$\operatorname{s_{{\mathbf{v}}}}(X):=\left\{x\in X:v_n(x)\to0 \text{ in }{\mathbb{T}}\right\}.$$ One can easily check that $ \operatorname{s_{{\mathbf{v}}}}(X)$ is a subgroup of $X$. \[DefChar\] [[@DMT]]{} A subgroup $H$ of a compact abelian group $X$ is called *characterized* if there exists a sequence of characters ${\mathbf{v}}$ such that $H=\operatorname{s_{{\mathbf{v}}}}(X)$. We also say that ${\mathbf{v}}$ *characterizes* $H$ and denote by $\operatorname{\mathfrak{Char}}(X)$ the family of all characterized subgroups of $X$. The following general problem has been studied in [@DG]. \[probBor\] Describe the Borel complexity of $\operatorname{s_{{\mathbf{v}}}}(X)$ for a compact abelian group $X$. The Borel complexity of a Borel set is the lowestclass in the Borel Hierarchy where such a set belongs to. Let us recall the first six classes of the Borel Hierarchy. \[nsub\] $$\begin{matrix} & \{\text{open sets}\} \!&\!\subseteq\! &\! \{G_\delta\text{-sets}\} \!&\!\subseteq \!&\!\{\text{$G_{\delta\sigma}$-sets}\}&\\ & \{\text{closed sets}\} \!&\! \subseteq \!& \!\{F_\sigma\text{-sets}\}&\subseteq \!&\!\{\text{$F_{\sigma\delta}$-sets}\}& \end{matrix},$$ where $G_\delta$-sets are countable intersection of open sets, $F_\sigma$-sets are countable union of closed sets, $G_{\delta\sigma}$ are countable union of $G_\delta$-sets and $F_{\sigma\delta}$-sets are countable intersection of $F_\sigma$-sets. In this paper we are interested in Borel subgroups, hence we shall denote by $\operatorname{\mathfrak{G}_\delta}(X)$, $\operatorname{\mathfrak{F}_\sigma}(X)$, $\operatorname{\mathfrak{G}_{\delta\sigma}}(X)$ and $\operatorname{\mathfrak{F}_{\sigma\delta}}(X)$ the class, respectively, of $G_{\delta}$-subgroups, $F_{\sigma}$-subgroups, $G_{\delta\sigma}$-subgroups and $F_{\delta\sigma}$-subgroups of a given topological group $X$ (e.g., accordingly, an $F_{\sigma\delta}$-subgroup means a subgroup that is an $F_{\sigma\delta}$-set as a subset, etc.). \[rem1\] Every characterized subgroup $H$ of a compact abelian group $X$ is an $F_{\sigma\delta}$-subgroup of $X$, and hence $H$ is a Borel subset of $X$. Indeed, if $H$ is characterized by a sequence ${\mathbf{v}}$, this fact directly follows from the equality $$\operatorname{s_{{\mathbf{v}}}}(X) = \bigcap_{M=1}^\infty \bigcup_{m=0}^\infty \left( \bigcap_{n\geq m}^\infty \left\{ x\in X : {\left\|}v_n(x){\right\|}\leq \frac{1}{M} \right\} \right).$$ Since the characterized subgroups are Borel, if $H\in\operatorname{\mathfrak{Char}}(X)$ then either $|H|\le\aleph_0$ or $|H|={\mathfrak{c}}$ by Alexandroff-Hausdorff’s theorem [@KK §37, Theorem 3]. One can define characterized subgroups of an arbitrary abelian topological group. Similarly as in Remark \[rem1\], one can see that they are always Borel subgroups (actually, $F_{\delta\sigma}$-subgroups). On the other hand, Borel subgroups of the uncountable Polish abelian groups need not be characterized in general. Indeed, the above remark establishes an upper bound for the Borel complexity of characterized subgroups. On the other hand, every uncountable Polish group contains arbitrarily complicated Borel subgroups [@FS Theorem 2.1]. This implies that every uncountable Polish group contains a Borel subgroup that is not characterized. The main issue discussed in this paper is when characterized subgroups are $F_\sigma$, as well as the opposite question: when an $F_\sigma$-subgroup of a compact abelian group is a characterized subgroup. The paper is organized as follows. Section 2 is dedicated to the general case of arbitrary compact abelian groups. In §2.1 we recall some known facts about characterized subgroups and their Borel complexity, starting from the case of countable subgroups and $G_\delta$-subgroups. Then comes the reduction to the case of metrizable compact abelian groups and some facts related to $F_\sigma$-subgroups. This introduces in a natural way the next topic, in §2.2, of the next sharper necessary condition for a subgroup to be characterized, namely, to be Polishable (beyond being a Borel subgroup). This gives rise to a finer Polish topology $\tau(H)$ (witnessing Polishability) for a characterized subgroup $H$ of a compact metrizable abelian group. Further properties of this topology are discussed in §2.3, local compactness being the prominent among all. These stronger properties of $\tau(H)$ turn out to be sufficient for a group $H$ to be characterized. In §2.4 comes a general criterion for a characterized subgroup $H$ to be an $F_\sigma$-subgroup in terms of a new topology, obtained by a suitable modification of $\tau(H)$. Section 3 is dedicated to the case of subgroups of ${\mathbb{T}}$. Since $\widehat {\mathbb{T}}= {\mathbb{Z}}$, now the sequences of characters are simply sequences of integers. It turns out that the proper characterized subgroups of ${\mathbb{T}}$ whose associated Polish topology is locally compact are precisely the countable ones. Moreover, among all compact abelian groups only the group ${\mathbb{T}}$ has this property. A special attention is dedicated to sequences ${\mathbf{u}}$ such that $u_n | u_{n+1}$. In this case we obtain a complete description of the characterized subgroups of ${\mathbb{T}}$ that are $F_\sigma$. Surprisingly, these are precisely the countable subgroups of ${\mathbb{T}}$. Characterized Subgroups of Compact Abelian Groups ================================================= Some known results on Borel Complexity -------------------------------------- It was proved by B' ir' o, Deshouillers and Sós [@BDS] that all countable subgroups of ${\mathbb{T}}$ are characterized. These authors conjectured that this fact could be extended to all compact metrizable abelian groups, although they gave no rigorous definition of a characterized subgroup of a compact abelian group. The relevant definition \[DefChar\] was provided later only in [@DMT], where cyclic subgroups of some compact metrizable abelian groups were proved to be characterized. The fact conjectured by B' ir' o, Deshouillers and Sós was proved to be true: \[ThmDK\] The countable subgroups of a compact metrizable abelian group are characterized. \[l21\] It was pointed out in [@VN] that metrizability is a necessary condition in the above theorem. Indeed, for a sequence ${\mathbf{v}}$ of characters of a compact abelian groups $X$, the subgroup $K_{\mathbf{v}}=\bigcap_{n\in\omega} \ker v_n$ is a closed $G_\delta$-subgroup of $X$ contained in $\operatorname{s_{{\mathbf{v}}}}(X)$. Hence, if the (countable) subgroup $\{0\}$ of $X$ is characterized, then $\{0\}$ is a $G_\delta$-subgroup of $X$, so $X$ is metrizable. According to the next theorem, the $G_\delta$-subgroups of a compact abelian group $X$ are precisely the closed characterized subgroups of $X$. \[TGchar\] Let $X$ be a compact abelian group. - $\operatorname{\mathfrak{G}_\delta}(X)\subseteq\{\text{closed subgroups of }X\}$. - $\operatorname{\mathfrak{G}_\delta}(X)\subseteq\operatorname{\mathfrak{Char}}(X)$; more precisely, a closed subgroup $H$ of $X$ is characterized if and only if $H$ is a $G_\delta$-subgroup. Note that if $X$ is metric then the inclusion of item (a) of the theorem becomes an equality. In case $X = {\mathbb{T}}$, item (b) of the above theorem is an easy consequence of item (a), as the closed subgroups of ${\mathbb{T}}$ are precisely the finite ones and they have the form ${\mathbb{T}}[m]=\{x\in{\mathbb{T}}:mx=0\}$ for some $m\in{\mathbb{N}}$ (${\mathbb{T}}[m]$ is trivially characterized by the constant sequence ${\mathbf{u}}=(m)_{n\in{\mathbb{N}}}$). For a sequence ${\mathbf{v}}$ of characters of a compact abelian groups $X$ the subgroup $\operatorname{s_{{\mathbf{v}}}}(X)/K_{\mathbf{v}}$ of $X/K_{\mathbf{v}}$ is characterized. More generally one has: \[ThB\] For a subgroup $H$ of a compact abelian group $X$ the following are equivalent: - $H$ is characterized; - $H$ contains a closed (necessarily characterized) $G_\delta$-subgroup $K$ of $X$ such that $H/K$ is a characterized subgroup of the compact metrizable group $X/K$. The previous theorem, reduces the study of characterizable subgroups of compact abelian groups, to the study of these subgroups in compact metrizable abelian groups. The general case of Theorem \[ThB\], namely the case of an arbitrary abelian group $X$, is considered in [@Ga5]. Following [@DG], call a subgroup $H$ of a compact abelian group $X$ [*countable modulo compact*]{} (briefly, [*CMC*]{}) if $H$ has a compact subgroup $K$ such that $H/K$ is countable and $K$ is a $G_\delta$-set of $X$. Clearly, CMC subgroups are $F_\sigma$, but they are also characterized subgroups: \[CDKB\] Let $ X$ be a compact abelian group and let $H$ be a CMC subgroup of $X$. Then $H$ is characterized. This follows easily from Theorems \[ThmDK\] and \[ThB\] (see Proposition \[new:prop\] for a stronger result). The above corollary was inspired by the following theorem. The subgroups characterized in the following theorem are particular CMC subgroups (in fact they are *countable torsion modulo a compact subgroup*. Let $X$ be a compact abelian group and $\{F_n\}_{n\in{\mathbb{N}}}$ a family of closed subgroups of $X$ such that, $F_n\lneq F_{n+1}$ for all $n\in{\mathbb{N}}$. Then $H=\bigcup F_n\in\operatorname{\mathfrak{Char}}(X)$ if and only if there exists $m\in{\mathbb{N}}$ such that $X/F_m$ is metrizable and $|F_{n+1}:F_n|<\infty$ for all $n\ge m$. One can prove that every characterized subgroup $H$ of $X$ is CMC if and only if $X$ has finite exponent [@DG]. A natural question arising from Corollary \[CDKB\] is the following: When an $F_\sigma$-subgroup $H$ of a compact metrizable abelian group $X$ is characterized? The next theorem sharpening Remark \[rem1\] was proved in [@DG]: \[teoDG1\] [[@DG]]{} For every infinite compact abelian group $X$, the following inclusions hold: $$\label{*} \operatorname{\mathfrak{G}_\delta}(X) \subsetneqq \operatorname{\mathfrak{Char}}(X) \subsetneqq \operatorname{\mathfrak{F}_{\sigma\delta}}(X)\ \ \mbox{ and } \ \ \operatorname{\mathfrak{F}_\sigma}(X) \not \subseteq \operatorname{\mathfrak{Char}}(X).$$ If in addition $X$ has finite exponent, then $$\label{dag} \operatorname{\mathfrak{Char}}(X) \subseteq \operatorname{\mathfrak{F}_\sigma}(X).$$ Clearly, one has $\operatorname{\mathfrak{Char}}(X) \subsetneqq \operatorname{\mathfrak{F}_\sigma}(X)$ in (\[dag\]), due to the second part of (\[\*\]). Gabriyelyan [@Ga2] proved that the implication in the final part of the above Theorem can be inverted: \[ThGa\] $\operatorname{\mathfrak{Char}}(X) \subseteq \operatorname{\mathfrak{F}_\sigma}(X)$ for a compact abelian group $X$ if and only if $X$ has finite exponent. In other words, for every compact abelian group of infinite exponent, he produced a characterized subgroup that is not $F_\sigma$. The first example of a characterized subgroup of ${\mathbb{T}}$ that is not an $F_\sigma$-subgroup of ${\mathbb{T}}$ was given by Bukovský, Kholshevikova, Repický [@BKR] (see Example \[Exa\_NonFS\] in §\[secaseq\]). The above discussion motivates the following general question When uncountable characterized subgroups of the compact metrizable groups are $F_\sigma$? This issue will be discussed in \[FS\], while the case of characterized subgroups of ${\mathbb{T}}$ is discussed in more detail in §\[Toro\]. The proof of the above theorem uses a relevant property of characterized subgroups, noticed first by Biró [@B] and independently, by Gabriyelyan [@Ga1], [@Ga2]. Namely they are Polishable, as we see in the next section. Polishability {#SPol} ------------- Let us recall the notion of Polishable subgroup that plays a key role in the study of characterized subgroups. It was introduced in [@KL]. \[DPol\] A [*Polishable subgroup*]{} $H$ of a Polish group $G$ is a subgroup that satisfies one of the following equivalent conditions: - there exists a Polish group topology $\tau$ on $H$ having the same Borel sets as $H$ when considered as a topological subgroup of $G$; - there exists a continuous isomorphism from a Polish group $P$ to $H$; - there exists a continuous surjective homomorphism from a Polish group $P$ onto $H$. The topology witnessing the polishability of a subgroup is unique [@S]. Answering negatively the first named author’s question on whether $\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})$ is contained in $\operatorname{\mathfrak{Char}}({\mathbb{T}})$, Biró proved in 2005 the more precise Theorem \[BiroTh\]. Before stating the theorem, let us recall the definition of Kronecker set. A non empty compact subset $K$ of an infinite compact metrizable abelian group $X$ is called a *Kronecker set*, if for every continuous function $f:K\to{\mathbb{T}}$ and $\varepsilon>0$ there exists a character $v\in\widehat{X}$ such that $$\max\left\{{\left\|}f(x)-v(x){\right\|}:x\in K\right\}<{\varepsilon}.$$ If $X$ is a group and $A\subseteq X$, denote by $\langle A\rangle$ the subgroup of $X$ generated by $A$. \[BiroTh\] If $K$ is an uncountable Kronecker set of ${\mathbb{T}}$, then $\langle K\rangle\in\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})\setminus\operatorname{\mathfrak{Char}}({\mathbb{T}})$. Inspired by ideas of Aaronson and Nadkarni [@AN], he proved that every characterized subgroup of ${\mathbb{T}}$ is Polishable, while the subgroup $\langle K\rangle$ generated by any uncountable Kronecker set $K$, is not polishable and hence not characterized. Unaware of this result, Gabriyelyan [@Ga1] generalized significantly this theorem in several directions. \[Tpolgen\] Let $X$ be a compact metrizable abelian group. Then - [[@Ga1 Theorem 1]]{} $\operatorname{s_{{\mathbf{v}}}}(X)$ is Polishable for every sequence ${\mathbf{v}}$ of characters of $X$; - [[@Ga1 Theorem 2]]{} if $K$ is an uncountable Kronecker set in $X$, then $\langle K\rangle$ is not Polishable; in particular, $\langle K\rangle$ is not characterized. The topology witnessing the polishability of $\operatorname{s_{{\mathbf{v}}}}(X)$ is induced by the following metric. Let $X$ be a compact metrizable abelian group, $\delta$ be a compatible invariant metric on $X$ and ${\mathbf{v}}=(v_n)$ be a sequence of characters of $X$. Let $x,y\in X$ and $$\varrho_{\mathbf{v}}(x,y)=\sup_{n\in{\mathbb{N}}}\{\delta(x,y),d(v_n(x),v_n(y))\}.$$ Denote by $\tau_{\varrho_{\mathbf{v}}}$ the topology of $X$ generated by the metric $\varrho_{\mathbf{v}}$. By the uniqueness of the Polish topology, the restriction $\tau_{\varrho_{\mathbf{u}}}\restriction_{\operatorname{s_{{\mathbf{u}}}}(X)}$ does not depend on ${\mathbf{u}}$, in the sense that $\tau_{\varrho_{\mathbf{u}}}\restriction_{\operatorname{s_{{\mathbf{u}}}}(X)}= \tau_{\varrho_{\mathbf{v}}}\restriction_{\operatorname{s_{{\mathbf{v}}}}(X)}$ whenever $\operatorname{s_{{\mathbf{u}}}}(X)= \operatorname{s_{{\mathbf{v}}}}(X)$. This is why we denote by $\tau(H)$ the unique Polish topology of a characterized subgroup $H$ of $X$. Let us note here that [*only this restriction*]{} $\tau(H)$ was considered in [@B; @Ga1]. The topology $\tau_{\varrho_{\mathbf{v}}}$ on [*the whole*]{} group $X$ appeared for the first time only in [@DG]. The assignments ${\mathbf{u}}\mapsto \operatorname{s_{{\mathbf{u}}}}(X)$ and ${\mathbf{u}}\mapsto \tau_{\varrho_{\mathbf{u}}}$ give rise to two natural equivalence relations between sequences of characters of $X$. For two sequences of characters ${\mathbf{u}}$ and ${\mathbf{v}}$ of a compact abelian group $X$ we write - ${\mathbf{u}}\sim {\mathbf{v}}$, if $\operatorname{s_{{\mathbf{u}}}}(X)= \operatorname{s_{{\mathbf{v}}}}(X)$ - ${\mathbf{u}}\approx {\mathbf{v}}$, if $\tau_{\varrho_{\mathbf{u}}}= \tau_{\varrho_{\mathbf{v}}}$. As mentioned above, ${\mathbf{u}}\sim {\mathbf{v}}$ always implies also $\tau_{\varrho_{\mathbf{u}}}\restriction_{\operatorname{s_{{\mathbf{u}}}}(X)}= \tau_{\varrho_{\mathbf{v}}}\restriction_{\operatorname{s_{{\mathbf{v}}}}(X)}$, but we do not know if ${\mathbf{u}}\sim {\mathbf{v}}$ implies ${\mathbf{u}}\approx {\mathbf{v}}$ in general (see Question \[Qtopocoinc\]). We shall see below that the answer is positive if the group $\operatorname{s_{{\mathbf{u}}}}(X)= \operatorname{s_{{\mathbf{v}}}}(X)$ is $F_\sigma$ (Theorem \[CorC1\]). Let $\tau$ be the compact metrizable topology on $X$, it is clear that $\tau_{\varrho_{\mathbf{v}}}\supseteq\tau$ and hence $\tau(H) \supseteq \tau\restriction_H$ for $H\in\operatorname{\mathfrak{Char}}(X)$. \[11Ott\] It is important to mention that $\tau(H) = \tau\restriction_H$, if and only if $H$ is closed. Indeed, $\tau\restriction_H$ is Polish if and only if $H\in\operatorname{\mathfrak{G}_\delta}(X)$. By item (a) of Theorem \[TGchar\] the class $\operatorname{\mathfrak{G}_\delta}(X)$ coincides with class of closed subgroups of $X$. Using the uniqueness of $\tau(H)$, we can conclude that $\tau(H) = \tau\restriction_H$ precisely when $H$ is closed. In the light of the above mentioned results of Biró and Gabriyelyan, it makes sense to introduce also the notation $\operatorname{\mathfrak{Pol}}(X)$ for the collection of all Polishable subgroups of $X$. Clearly, Theorem \[Tpolgen\] (a) can be written briefly as $$\label{Eq1} \operatorname{\mathfrak{Char}}(X)\subseteq\operatorname{\mathfrak{Pol}}(X).$$ One can expect to have equality in (\[Eq1\]) at least for some infinite compact metrizable abelian groups $X$. That this cannot occur follows from a general result of Hjorth [@Hj]. Answering a question of Farah and Solecki [@FS Question 6.1], he established that an uncountable abelian Polish group contains Polishable subgroups of unbounded Borel complexity. According to Remark \[rem1\], none of these Polishable subgroups can be characterized. Therefore, equality in (\[Eq1\]) fails for every uncountable abelian Polish group $X$. In [@Ga4; @Ga1] Gabriyelyan produced compact groups $X$ witnessing $\operatorname{\mathfrak{Pol}}(X)\nsubseteq\operatorname{\mathfrak{Char}}(X)$ using a different idea, since the Polishable subgroup he produced have some additional properties, being among others $F_\sigma$-subgroups. In this way he established for these particular groups the following stronger non-inclusion $$\label{Eq93} \operatorname{\mathfrak{Pol}}(X)\cap\operatorname{\mathfrak{F}_\sigma}(X)\not \subseteq\operatorname{\mathfrak{Char}}(X).$$ \[QG61\] Does (\[Eq93\]) holds for all compact metrizable abelian groups? Recently Gabriyelyan proved in [@Ga6] that *the non-inclusion (\[Eq93\]) holds for every compact metrizable non-totally disconnected abelian group*. Further properties of the polish topology of a characterized subgroup {#FurtherProp} --------------------------------------------------------------------- A subset $A$ of a topological abelian group $G$ is called [*quasi-convex*]{} if for every $g\in G\setminus A$ there exists $\chi\in \widehat{G}$ such that $$\chi(A)\subset {\mathbb{T}}_+,\,\,\,\text{but}\,\,\chi(g)\not\in {\mathbb{T}}_+ \, ,$$ where ${\mathbb{T}}_+ $ is the image of the segment $[-\frac{1}{4},\frac{1}{4} ]$ with respect to the natural quotient map ${\mathbb{R}}\rightarrow {\mathbb{T}}$. A topological group $(G,\tau)$ (as well as its topology $\tau$) is called [*locally quasi-convex*]{} if $\mathcal{N}(G,\tau)$ $G$ has a basis of neighborhoods of $0$ consisting of quasi-convex subsets of $(G,\tau)$. In connection to the inclusion (\[Eq1\]), Gabriyelyan [@Ga1] added in item (a) of Theorem \[Tpolgen\] one more property of the finer Polish topology of a characterized subgroup $H$ of a compact metrizable abelian group $X$; namely, it is also locally quasi-convex. Therefore, if $\operatorname{\mathfrak{Pol}}_{qc}(X)$ denotes the family of all subgroups of $X$ that admits a finer polish locally quasi-convex topology, he refined (\[Eq1\]) with the following inclusion: $$\label{Eq1qc} \operatorname{\mathfrak{Char}}(X)\subseteq\operatorname{\mathfrak{Pol}}_{qc}(X).$$ On the other hand, he produced in [@Ga1] a compact metrizable abelian group $X$ with a locally quasi-convex Polishable $F_\sigma$-subgroup that is not characterized. Hence the following non-inclusion that improves (\[Eq93\]) holds: $$\label{Eq93qc} \operatorname{\mathfrak{Pol}}_{qc}(X)\cap\operatorname{\mathfrak{F}_\sigma}(X)\not \subseteq\operatorname{\mathfrak{Char}}(X).$$ The group produced in [@Ga1] that witnesses (\[Eq93qc\]) is ${\mathbb{T}}^{\mathbb{N}}$. It makes sense to pose the following question related to Question \[QG61\]. \[QG61qc\] Does (\[Eq93qc\]) holds for all compact metrizable abelian groups? In the light of Biró’s result it would be natural to ask whether (\[Eq93\]) holds for ${\mathbb{T}}$, namely whether $\operatorname{\mathfrak{Pol}}({\mathbb{T}})\cap\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})\subseteq\operatorname{\mathfrak{Char}}({\mathbb{T}})$. A negative answer to this question can be found in [@Ga4]. Indeed, the group $G_2$ defined there is Polishable and $F_\sigma$ (by [@Ga4 Proposition 1]). On the other hand, $G_2$ is not locally quasi-convex (by [@Ga4 Theorem 2]), so $G_2$ is not characterized. So it makes sense to ask Question \[QG61qc\] for $X={\mathbb{T}}$. \[QG61qcT\] Does (\[Eq93qc\]) holds for ${\mathbb{T}}$? A natural class of groups having both properties (namely, being simultaneously Polish and locally quasi convex) is the class of second countable locally compact abelian groups. Therefore, it makes sense to consider also the subfamily $\operatorname{\mathfrak{LC}}(X)$ of $\operatorname{\mathfrak{Pol}}_{qc}(X)$, consisting of those subgroups $H\in \operatorname{\mathfrak{Pol}}(X)$ whose Polish topology is [*locally compact*]{}. This is justified by the following result from [@N], answering a question from [@Ga3]: \[negro\] If $G$ is a second countable locally compact abelian group and $p: G \to X$ is a continuous injective homomorphism into a compact metrizable abelian group $X$, then $p(X)$ is a characterized subgroup of $X$. This theorem provides a useful and handy sufficient condition for a subgroup of a compact metrizable abelian group to be characterized: \[corollary:negro\] $\operatorname{\mathfrak{LC}}(X)\subseteq\operatorname{\mathfrak{Char}}(X)$ for every compact metrizable abelian group $X$. Indeed, if $H \in \operatorname{\mathfrak{LC}}(X)$ and $G$ denotes the group $H$ equipped with the finer locally compact group topology, then the inclusion map $p: G \hookrightarrow X$ is a continuous injective homomorphism, so $H = p(G)$ is a characterized subgroup of $X$ by Theorem \[negro\]. The next chain of inclusions summarizes (\[Eq1\]), (\[Eq1qc\]) and the above corollary $$\operatorname{\mathfrak{LC}}(X)\subseteq\operatorname{\mathfrak{Char}}(X)\subseteq\operatorname{\mathfrak{Pol}}_{qc}(X).$$ Recall that a topological space is *$\sigma$-compact* if it is a countable union of compact subspace. Note that if $H\in\operatorname{\mathfrak{LC}}(X)$, then $H$ is $\sigma$-compact (see [@K]) and hence obviously an $F_\sigma$-subgroup. Therefore, the following more precise inclusion holds $$\label{eqdag} \operatorname{\mathfrak{LC}}(X)\subseteq\operatorname{\mathfrak{Char}}(X)\cap\operatorname{\mathfrak{F}_\sigma}(X).$$ Obviously, $\operatorname{\mathfrak{LC}}(X)$ contains all countable subgroups as well as all clopen subgroups of a compact metrizable group $X$. Sometimes, $\operatorname{\mathfrak{LC}}(X)$ may be limited to exactly those subgroups of $X$, as in the case of the group ${\mathbb{J}}_p$ of $p$-adic integers, i.e., $$\operatorname{\mathfrak{LC}}({\mathbb{J}}_p)=\{\text{countable subgroups}\}\cup\{\text{clopen subgroups}\}.$$ It is easy to check, that every CMC subgroup of $X$ is in $ \operatorname{\mathfrak{LC}}(X)$: \[new:prop\] The CMC subgroups of a compact metrizable abelian group $X$ belong to $\operatorname{\mathfrak{LC}}(X)$, so they are characterized $F_\sigma$-subgroups of $X$. Note that this proposition provides a new proof of Corollary \[CDKB\]. The next theorem describes when the implication of Proposition \[new:prop\] can be inverted. \[new:prop1\][@HH] For a compact metrizable abelian group $X$ the following are equivalent: - there exist no continuous injective homomorphisms ${\mathbb{R}}\to X$; - all subgroups in $\operatorname{\mathfrak{LC}}(X)$ are CMC. Obviously, $X$ satisfies (a) whenever $X$ is totally disconnected. More precisely, $X$ satisfies (a) if and only if its arc-component is either trivial or isomorphic to ${\mathbb{T}}$ (for more detail see [@HH]). When characterized subgroups of the compact metrizable groups are $F_\sigma$? {#FS} ------------------------------------------------------------------------------ According to Theorem \[ThGa\], the answer to the above question is always, when $X$ is of finite exponent. For groups $X$ of infinite exponent, the same theorem provides always some characterized subgroup of $X$ that is not $F_\sigma$. Our aim is to precisely describe when a characterized subgroup of $X$ is $F_\sigma$. To this end one can introduce the following topology, obtained by a modification of $\tau_{\varrho_{\mathbf{v}}}$. \[DefTestTopo\] [[@DI3]]{} Let ${\mathbf{v}}$ be a sequence of characters on a compact metrizable abelian group $(X,\tau)$ and $\tau_{\varrho_{\mathbf{v}}}^*$ be the topology on $X$ having as a filter of neighborhoods of $0$ in $X$ the family $$\left\{W_n=\overline{B^{\varrho_{\mathbf{v}}}_{1/n}(0)}:n\in{\mathbb{N}}\right\},$$ where $B^{\varrho_{\mathbf{v}}}_{1/n}(0)$ is the ball around 0 of radius ${1/n}$ with respect to the metric $\varrho_{\mathbf{v}}$ and $\overline{B^{\varrho_{\mathbf{v}}}_{1/n}(0)}$ denotes its closure in $(X,\tau)$. We can refer to $\tau_{\varrho_{\mathbf{v}}}^*$ as the *$\operatorname{\mathfrak{F}_\sigma}$-test topology* with respect to the sequence ${\mathbf{v}}$. Note that this topology is metrizable since it has a countable local base. Moreover if $\tau$ is the original compact topology on $X$, the following inclusions hold. $$\tau\subseteq\tau_{\varrho_{\mathbf{v}}}^*\subseteq\tau_{\varrho_{\mathbf{v}}}.$$ Clearly, $\tau_{\varrho_{\mathbf{v}}}$ is discrete if and only if $\tau_{\varrho_{\mathbf{v}}}^*$ is discrete. The following theorem, providing a simple criterion for the solution of Problem \[probBor\], justifies the term test-topology for $\tau_{\varrho_{\mathbf{v}}}^*$. \[ThC\] Let $X$ be a metrizable compact abelian group and ${\mathbf{v}}\in \widehat{X}^{\mathbb{N}}$. Then $$\operatorname{s_{{\mathbf{v}}}}(X)\in\operatorname{\mathfrak{F}_\sigma}(X)\Longleftrightarrow\operatorname{s_{{\mathbf{v}}}}(X)\in\tau_{\varrho_{\mathbf{v}}}^*.$$ An important consequence of the above theorem is the independence of the topology $\tau_{\varrho_{\mathbf{v}}}$ on the choice of the characterizing sequence whenever $\operatorname{s_{{\mathbf{v}}}}(X)$ is an $F_\sigma$-subgroup. In such a case $(X,\tau_{\varrho_{\mathbf{v}}})$ is Polish precisely when $\operatorname{s_{{\mathbf{v}}}}(X)$ is an open subgroup of $X$: \[CorC1\] Let $X$ be a metrizable compact abelian group and ${\mathbf{v}}$ be a sequence of characters such that $\operatorname{s_{{\mathbf{v}}}}(X)\in\operatorname{\mathfrak{F}_\sigma}(X)$. - If ${\mathbf{u}}$ is a sequence of characters such that ${\mathbf{v}}\sim{\mathbf{u}}$, then ${\mathbf{v}}\approx{\mathbf{u}}$. - The following are equivalent: - $(X,\tau_{\varrho_{\mathbf{v}}})$ is Polish; - $(X,\tau_{\varrho_{\mathbf{v}}})$ is separable; - $\operatorname{s_{{\mathbf{v}}}}(X)$ is a closed finite-index (so, open) subgroup of $X$. It is not known if item (a) of the above theorem remains valid without the assumption that $H$ is an $F_\sigma$-subgroup (see Question \[Qtopocoinc\]). \[new:cor\] Let $X$ be a compact metrizable abelian group. If $H \in \operatorname{\mathfrak{LC}}(X)$, then for every pair of characterizing sequences ${\mathbf{u}}$ and ${\mathbf{v}}$ for $H$ the topologies $\tau_{\varrho_{\mathbf{u}}}$ and $\tau_{\varrho_{\mathbf{v}}}$ coincide. Follows from Theorem \[CorC1\] and (\[eqdag\]). From Theorem \[ThC\] we deduce now that $\tau_{\varrho_{\mathbf{v}}}$ is discrete on the whole $X$ whenever $\operatorname{s_{{\mathbf{v}}}}(X)$ is countable. \[CorC2\] Let $X$ be a metrizable compact abelian group and ${\mathbf{v}}\in \widehat{X}^{\mathbb{N}}$. Then the following are equivalent: - $\operatorname{s_{{\mathbf{v}}}}(X)$ is countable; - $|\operatorname{s_{{\mathbf{v}}}}(X)|<{\mathfrak{c}}$; - $\tau_{\varrho_{\mathbf{v}}}$ is discrete; - $\tau_{\varrho_{\mathbf{v}}}^*$ is discrete; - $\tau_{\varrho_{\mathbf{v}}}^*\restriction_{\operatorname{s_{{\mathbf{v}}}}(X)}$ is discrete; - $\tau({\operatorname{s_{{\mathbf{v}}}}(X)})$ is discrete. (a)$\Leftrightarrow$(b). It is a consequence of Remark \[rem1\] (a)$\Rightarrow(c)$. Indeed if $\operatorname{s_{{\mathbf{v}}}}(X)$ is countable, then it is obviously in $\operatorname{\mathfrak{F}_\sigma}(X)$. Hence, $\operatorname{s_{{\mathbf{v}}}}(X)$ is $\tau_{\varrho_{\mathbf{v}}}$-open, by Theorem \[ThC\]. Moreover $\tau_{\varrho_{\mathbf{v}}}$ is discrete when restricted to $\operatorname{s_{{\mathbf{v}}}}(X)$ since the topology that witnesses the polishability of $\operatorname{s_{{\mathbf{v}}}}(X)$ is unique. Thus, $\operatorname{s_{{\mathbf{v}}}}(X)$ is both $\tau_{\varrho_{\mathbf{v}}}$-open and $\tau_{\varrho_{\mathbf{v}}}$-discrete. Consequently, $\tau_{\varrho_{\mathbf{v}}}$ is discrete. (c)$\Rightarrow$(d) follows from the definition of $\tau_{\varrho_{\mathbf{v}}}^*$, while (d)$\Rightarrow$(e) is obvious. (e)$\Rightarrow$(f) follows from the fact that $\tau({\operatorname{s_{{\mathbf{v}}}}(X)})=\tau_{\varrho_{\mathbf{v}}}\restriction_{\operatorname{s_{{\mathbf{v}}}}(X)}$ is finer then $\tau_{\varrho_{\mathbf{v}}}^*\restriction_{\operatorname{s_{{\mathbf{v}}}}(X)}$. (f)$\Rightarrow$(a) follows from the fact that $(\operatorname{s_{{\mathbf{v}}}}(X),\tau({\operatorname{s_{{\mathbf{v}}}}(X)}))$ is separable. This theorem shows, that ${\mathbf{u}}\approx {\mathbf{v}}$ does not imply ${\mathbf{u}}\sim {\mathbf{v}}$. Indeed, if $H$ and $H'$ two distinct countable subgroups, then they are characterized by Theorem \[ThmDK\], say $H= \tau_{\varrho_{\mathbf{u}}}$ and $H'=\tau_{\varrho_{\mathbf{v}}}$. According to Theorem \[CorC2\], $\tau_{\varrho_{\mathbf{u}}}=\tau_{\varrho_{\mathbf{v}}}$ are discrete. So ${\mathbf{u}}\approx {\mathbf{v}}$, while $H' \ne H$, i.e., ${\mathbf{u}}\not\sim {\mathbf{v}}$. Characterized subgroups of the Circle Group {#Toro} =========================================== Preliminaries ------------- For a topological abelian group $X$ and a sequence ${\mathbf{u}}=(u_n)_{n\in{\mathbb{N}}}$ of integers, let $$\operatorname{t_{{\mathbf{u}}}}(X)=\left\{x\in X:u_nx\to0 \text{ in } X \right\}.$$ This subset is actually a subgroup of $X$ and has been studied in [@Arm; @D; @BDMW2]. Here we consider only the case when $X = {\mathbb{T}}$. Since $\widehat{{\mathbb{T}}}={\mathbb{Z}}$, the notion defined above coincides (in the case of $X={\mathbb{T}}$) with the notion of characterized subgroup of ${\mathbb{T}}$. In fact, ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}=\operatorname{s_{{\mathbf{v}}}}({\mathbb{T}})$ where for all $n\in{\mathbb{N}}$ the character $v_n$ of ${\mathbb{T}}$ is defined by $v_n:x\mapsto u_nx$. In other words, $H\le{\mathbb{T}}$ is characterized if and only if there exists ${\mathbf{u}}\in{\mathbb{Z}}^{{\mathbb{N}}}$ such that $H={\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$. The above described circumstance provides additional tools to use in the study of characterized subgroups of ${\mathbb{T}}$, namely the specific properties of ${\mathbb{Z}}$. As a first easy reduction, one can note that if ${\mathbf{u}}$ does not have any constant subsequences one can find a strictly increasing sequence ${\mathbf{u}}^*$ of non-negative integers such that ${\mathbf{u}}^*\sim {\mathbf{u}}$ (see [@BDMW1 Proposition 2.5]). Another powerful tool to use in the study of characterized subgroups of ${\mathbb{T}}$ is the sequence of ratios $(q_n):=(\frac{u_n}{u_{n-1}})$ for a sequence ${\mathbf{u}}$ of non-zero integers. Furthermore, let $$q^{\mathbf{u}}: = \limsup_n\frac{u_{n+1}}{u_n}\mbox{ and }q_{\mathbf{u}}: = \liminf_n\frac{u_{n+1}}{u_n}.$$ Note that the sequence of ratios $(q_n)$ is bounded precisely when $q^{\mathbf{u}}$ is finite (one can also say that ${\mathbf{u}}$ is *$q$-bounded*), while $(q_n)$ converges to infinity precisely when $q_{\mathbf{u}}= \infty$ (i.e. ${\mathbf{u}}$ is *$q$-divergent*). In these terms, Eggleston [@Egg] proved the following remarkable theorem (see also [@BDMW1]). \[TEgg\] Let ${\mathbf{u}}=(u_n)$ a sequence of integers. - If $q^{\mathbf{u}}< \infty$ (i.e., the sequence of ratios $(q_n)$ is bounded), then ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is countable. - If $q_{\mathbf{u}}= \infty$ (i.e., the sequence of ratios $(q_n)$ converges to infinity), then ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is uncountable. One can show that the implications of Eggleston Theorem \[TEgg\] cannot be inverted: \[B-M-W\] If $H$ is an infinite characterized subgroup of ${\mathbb{T}}$, with characterizing sequence ${\mathbf{u}}\in{\mathbb{Z}}^{\mathbb{N}}$, then for every $m\in{\mathbb{N}}$ one can find a strictly increasing sequence ${\mathbf{u}}^*\sim {\mathbf{u}}$ such that $q^{{\mathbf{u}}^*}=\infty$ (i.e., $\limsup_n\frac{u_{n+1}^*}{u_n^*}=\infty$) and $q_{{\mathbf{u}}^*}=m$ (i.e., $\liminf_n\frac{u_{n+1}^*}{u_n^*}=m$). This remark shows that the properties of having bounded ratios, or having ratios converging to $\infty$ are not $\sim$-invariant. Actually, if one takes a sequence ${\mathbf{u}}$ with bounded ratios, and then a sequence ${\mathbf{u}}^*\sim {\mathbf{u}}$ as in the remark, then one will have also ${\mathbf{u}}^*\approx {\mathbf{u}}$, according to Theorem \[CorC1\], while ${\mathbf{u}}^*$ will fail to have the property of having bounded ratios. On the other hand the following remark, answering a question posed in [@BDMW1], is a sort of inversion of the first Eggleston implication. Every countable subgroup of the circle group, admits a $q$-bounded characterizing sequence. If $(q_n)$ is bounded, then ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is countable and hence obviously $F_\sigma$. One can study the relations between these three properties, that are not equivalent in general (see Theorem \[ThE\] and remarks \[Ranotd\] and \[Rbnotd\]). To this end we distinguish two cases, the first one is the case of a general sequence of integers. The second one is the case of a particular kind of sequences of integers, namely the arithmetic sequences (see Definition \[Def\_a\_seq\]). In this case the three properties mentioned above are equivalent. General Sequences ----------------- According to Theorem \[CorC2\], when $\operatorname{s_{{\mathbf{v}}}}(X)$ is countable, then the topology $\tau_{\varrho_{\mathbf{v}}}$ is discrete on $X$, where $X$ is a compact metrizable abelian group. Hence, for $X={\mathbb{T}}$ and a sequence of integers ${\mathbf{u}}$, having a bounded sequence of ratios $(q_n)$, the topology $\tau_{\varrho_{\mathbf{v}}}$ is discrete on ${\mathbb{T}}$ in view of Theorem \[TEgg\] (a). However, one can say something more precise than Theorem \[CorC2\]: \[ThA\] Let ${\mathbf{u}}$ be a strictly increasing sequence of positive integers such that its sequence of ratios $(q_n)$ is bounded. In that case $\tau_{\varrho_{\mathbf{u}}}$ is discrete. In particular if $0<C\in{\mathbb{R}}$ and $q_n\le C$ for all $n\in{\mathbb{N}}$, then $B_{\frac{1}{2C}}^{\varrho_{\mathbf{u}}}(0) = \{0\}$ in ${\mathbb{T}}$. The proof of the next theorem uses a property of ${\mathbb{T}}$ first noticed by Hewitt [@H], and then extended to all subgroups of ${\mathbb{T}}$ (and elsewhere) in [@HH]: [*if $H$ is a subgroup of ${\mathbb{T}}$ and $\tau$ is a strictly finer locally compact group topology on $H$, then $\tau$ is discrete.*]{} Hence one can add to Theorem \[CorC2\] one more equivalent condition in the case $X={\mathbb{T}}$, given by the next theorem. \[ThD\] Let $H\lneq{\mathbb{T}}$. Then $H\in\operatorname{\mathfrak{LC}}({\mathbb{T}})$, if and only if $H$ is countable. Let us recall that $\operatorname{\mathfrak{LC}}({\mathbb{T}})\subset\operatorname{\mathfrak{Char}}({\mathbb{T}})$ (by Corollary \[corollary:negro\]) and $H={\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ coincides with ${\mathbb{T}}$ if and only if ${\mathbf{u}}$ is definitely $0$ (see [@BDMW1 Example 2.8]). It turns out that this theorem cannot be proved in more general situation, i.e., it is specific for the group ${\mathbb{T}}$. \[ThDD\] Let $X$ be an infinite compact abelian group. Then the following are equivalent: - $X \cong {\mathbb{T}}$; - whenever $H\in\operatorname{\mathfrak{LC}}(X)$ and $H \ne X$, then $H$ is countable. Arithmetic Sequences {#secaseq} -------------------- In the case of ${\mathbb{T}}$ one has the following two examples of characterized non-$F_\sigma$-subgroups: \[Exa\_NonFS\] - For the sequence ${\mathbf{u}}= (2^{2^n})$ Bukovský, Kholshevikova, Repický [@BKR] proved that the characterized subgroup ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ of ${\mathbb{T}}$ is not an $F_\sigma$-set. - For the sequence ${\mathbf{u}}= (n!)$ Gabriyelyan [@Ga2] proved that the subgroup ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ of ${\mathbb{T}}$ is not an $F_\sigma$-subgroup of ${\mathbb{T}}$. The obvious common features between (a) and (b) are $u_n | u_{n+1}$ and $q_n= \frac{u_{n+1}}{u_n} \to \infty$. We are going to use the first one to define the following new notion. \[Def\_a\_seq\] A strictly increasing sequence of positive integers ${\mathbf{u}}= (u_n)$ is called [*arithmetic*]{} (or briefly, an [*a-sequence*]{}) if $u_n|u_{n+1}$ for all $n\in{\mathbb{N}}$. Let ${\mathbb{P}}$ be the set of all prime numbers. If $p\in{\mathbb{P}}$, then ${\mathbb{Z}}(p^\infty)$ is the Prüfer $p$-group and ${\mathbb{Z}}(p^n)$ the cyclic group of order $p^n$ where $n\in{\mathbb{N}}$. Note that in this case $(q_n)$ is a sequence of integers. The problem of the description of the structure of ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ in the case of arithmetic sequences has been raised in [@DPS Chap.4], where some partial results can be found. Much earlier, Armacost [@Arm] considered two special kinds of a-sequences, $u_n = n!$ and $u_n = p^n$ for some $p \in {\mathbb{P}}$. The subgroup characterized by the former type is ${\mathbb{Z}}(p^\infty)$, as shown in [@Arm]. The subgroup characterized by $u_n = n!$ was denoted by ${\mathbb{T}}!$, its description was left as an open problem in [@Arm], resolved independently by Borel [@Bo2] and [@DPS]. Further results on subgroups of ${\mathbb{T}}$ characterized by a-sequences can be found in [@DdS; @DI]. A complete description of this class of subgroups of ${\mathbb{T}}$ is given in [@DI]. \[RemBDMW\]([@BDMW]) If ${\mathbf{u}}$ is an a-sequence and $(q_n)$ is bounded, then $${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\cong\left(\bigoplus_{i=1}^s{\mathbb{Z}}(p_i^{k_i})\right) \oplus\left(\bigoplus_{j=1}^r{\mathbb{Z}}(t_j^\infty)\right);$$ where $p_i\in{\mathbb{P}}$, $k_i=|\{n:p_i|q_n\}|<\infty$ and $t_j$ are primes that divide infinitely many $q_n$. It is easy to see that the subgroups of this form of ${\mathbb{T}}$ are countably many. In the case of arithmetic sequences one can add some further equivalent conditions to the conditions of Theorem \[CorC2\] as follows. \[ThE\] The following are equivalent for an a-sequence ${\mathbf{u}}$ in ${\mathbb{Z}}$: - ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\le{\mathbb{Q}}/{\mathbb{Z}}$; - $(q_n)$ is bounded; - ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is countable; - ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in \operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})$; - ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in\tau_{\varrho_{\mathbf{u}}}$. The implications (b)$\Rightarrow$(c)$\Rightarrow$(d) are immediate (the first one due to Theorem \[TEgg\] (a)), while the equivalence between (a) and (b) is proved in [@DdS; @DI]. The implication (d)$\Rightarrow$(e) is a consequence of the Theorem \[ThC\] and the fact that $\tau_\varrho^*\subseteq\tau_\varrho$. The last implication (e)$\Rightarrow$(b) is proved in [@DI3]. From the equivalence of (a) and (b) in the above theorem and Example \[RemBDMW\], we deduce that [ *only countably many subgroups of ${\mathbb{Q}}/{\mathbb{Z}}$ can be characterized by means of an a-sequence ${\mathbf{u}}$ in ${\mathbb{Z}}$*]{}. Let us see that some of the implications of Theorem \[ThE\] do not hold for a general ${\mathbf{u}}\in{\mathbb{Z}}^{\mathbb{N}}$. \[Rbnotd\] If ${\mathbf{u}}$ is the Fibonacci’s sequence, then $(q_n)$ is bounded, so ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is countable (by virtue of Theorem \[TEgg\]), i.e. (b) and (c) of Theorem \[ThE\] holds. Indeed $u_n=u_{n-1}+u_{n-2}$ for all $n>1$ and $u_0=u_1=1$. Hence, $q_n=\frac{u_n}{u_{n-1}} = 1 + \frac{u_{n-2}}{u_{n-1}} \le2$ for all $n\in{\mathbb{N}}$. On the other hand, ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is the infinite cyclic group generated by the fractional part of the golden ratio [@La; @BDMW; @BDS]. In particular, ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ is not torsion. The following remark shows that (a)$\nRightarrow$(b)$\nLeftarrow$(c) in Theorem \[ThE\] in case ${\mathbf{u}}$ is not an a-sequence. \[Ranotd\] Take any infinite subgroup $H$ of ${\mathbb{Q}}/{\mathbb{Z}}$. By Remark \[B-M-W\], $H$ has a characterizing sequence such that its sequence of ratios is unbounded. This proves the non-implication (a)$\nRightarrow$(b) in Theorem \[ThE\]. As (a) implies (c), this witnesses also the non-implication (c)$\nRightarrow$(b) We do not know if the remaining implication (d)$\Rightarrow$(c) is true in the general case (see Question \[Q1\]). The following diagram summarizes Theorem \[ThE\] and remarks \[Rbnotd\] and \[Ranotd\], where the question mark denotes the unknown implication and the slashes denote the failing implications. $${\xymatrix@!0@C4.2cm@R=1.6cm{ (a) \ar@{=>}@/_/[rd] \ar|{/}@/_/[dd] & & \\ &(c)\ar|{/}@/_/[ul] \ar|{/}@/_/[dl]\ar@{=>}@/_/[r] &(d) \ar@/_/_{?}[l]\\ (b) \ar|{/}@/_/[uu] \ar@{=>}@/_/[ur]& & } }$$ Questions and Remarks ===================== We start with some questions and remarks concerning characterizing sequences. If $\operatorname{\mathfrak{a-Char}}({\mathbb{T}})$ denotes the set of all subgroups of ${\mathbb{T}}$ characterized by an a-sequence, then $$\{\text{closed subgroups}\}=\operatorname{\mathfrak{G}_\delta}({\mathbb{T}})\subsetneq\operatorname{\mathfrak{a-Char}}({\mathbb{T}})\cap\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}}).$$ An example witnessing the above proper inclusion is ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})\setminus\operatorname{\mathfrak{G}_\delta}({\mathbb{T}})$ where ${\mathbf{u}}=(p^n)$. More generally all countably infinite ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in \operatorname{\mathfrak{a-Char}}({\mathbb{T}})$ witness that proper inclusion, where ${\mathbf{u}}$ is an a-sequence. To give an exhaustive description of $\operatorname{\mathfrak{Char}}({\mathbb{T}})$ and $\operatorname{\mathfrak{a-Char}}({\mathbb{T}})$, in terms of Problem \[probBor\], it is needed to establish if there exists a sequence ${\mathbf{u}}$ such that ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\notin\operatorname{\mathfrak{G}_{\delta\sigma}}({\mathbb{T}})$. This kind of sets is called $F_{\sigma\delta}$-complete (or $\Pi_3^0$-complete using the Descriptive Set-Theoretic terminology, see [@K]), that is, in plain words, a set among the most complex in $\operatorname{\mathfrak{F}_{\sigma\delta}}({\mathbb{T}})$. Hence question of whether $\operatorname{\mathfrak{Char}}({\mathbb{T}})\subseteq \operatorname{\mathfrak{G}_{\delta\sigma}}({\mathbb{T}})$ remains open: - ([@Ga2]) Does there exists a sequence of integers ${\mathbf{u}}$ such that ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\notin\operatorname{\mathfrak{G}_{\delta\sigma}}({\mathbb{T}})$? - What about a-sequences ${\mathbf{u}}$? Does there exists an explicit method to decide whether ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in\operatorname{\mathfrak{G}_{\delta\sigma}}({\mathbb{T}})$ for a sequence of integers ${\mathbf{u}}$? What about a-sequences ${\mathbf{u}}$? Since $t({\mathbb{T}}) = {\mathbb{Q}}/{\mathbb{Z}}$ is countable, all ($\mathfrak c$ many) torsion subgroups of ${\mathbb{T}}$ can be characterized, by Theorem \[ThmDK\]. On the other hand, not all the torsion subgroups of ${\mathbb{T}}$ can be characterized by an a-sequence. Indeed, by Example \[RemBDMW\] and Theorem \[ThE\], all torsion subgroups of ${\mathbb{T}}$ of the form $\bigoplus_{i\in I}{\mathbb{Z}}(p_i^{k_i})$ with $k_i\in{\mathbb{N}}\cup\{\infty\}$ for all $i\in I, p_i\in{\mathbb{P}}$, and $p_i\neq p_j$ for $i\neq j$ and $\left|I\right|=\aleph_0$ cannot be characterized by an a-sequence. Hence, as far as only torsion subgroups of ${\mathbb{T}}$ are concerned, one can say that $\operatorname{\mathfrak{a-Char}}({\mathbb{T}})$ is much smaller with respect to $\operatorname{\mathfrak{Char}}({\mathbb{T}})$ (as $\operatorname{\mathfrak{a-Char}}({\mathbb{T}})$ contains only countably many torsion subgroups, while $\operatorname{\mathfrak{Char}}({\mathbb{T}})$ contains $\mathfrak c$ many torsion subgroups of ${\mathbb{T}}$). According to Theorem \[ThDD\], every compact metrizable abelian group $X$ non-isomorphic to the circle group, has a proper uncountable $F_\sigma$-subgroup $H$ that is characterized. Indeed, such an $H$ can be chosen to admit a finer locally compact Polish topology, hence $H$ is characterized by Corollary \[corollary:negro\] and $F_\sigma$ by (\[eqdag\]). On the other hand, for $X={\mathbb{T}}$ this matter remains unclear. This leaves open the following question, related to the implication (d)$\Rightarrow$(c) in Theorem \[ThE\]. \[Q1\] If ${\mathbf{u}}\in{\mathbb{Z}}^{\mathbb{N}}$ is not definitely $0$ and ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}\in\operatorname{\mathfrak{F}_\sigma}({\mathbb{T}})$, must ${\operatorname{t_{{\mathbf{u}}}}({\mathbb{T}})}$ be necessarily countable? The final questions concern Polishable subgroups. According to Theorem \[CorC1\], the topology $\tau_{\varrho_{\mathbf{v}}}$ does not depend on the choice of ${\mathbf{v}}$ in case $\operatorname{s_{{\mathbf{v}}}}(X)$ is an $F_\sigma$-subgroup of $X$, however the questions remain open in the general case: \[Qtopocoinc\] If $\operatorname{s_{{\mathbf{v}}}}(X)=\operatorname{s_{{\mathbf{u}}}}(X)\notin\operatorname{\mathfrak{F}_\sigma}(X)$, is $\tau_{\varrho_{\mathbf{v}}}=\tau_{\varrho_{\mathbf{u}}}$ in the whole $X$? In Theorem \[CorC1\] we described when $(X,\tau_{\varrho_{\mathbf{v}}})$ is Polish, provided $\operatorname{s_{{\mathbf{v}}}}(X)$ is an $F_\sigma$-subgroup of $X$. \[QtopoPol\] Describe when $(X,\tau_{\varrho_{\mathbf{v}}})$ is Polish in case $\operatorname{s_{{\mathbf{v}}}}(X)\notin\operatorname{\mathfrak{F}_\sigma}(X)$. Acknowledgements {#acknowledgements .unnumbered} ================ It is a pleasure to thank S. S. Gabriyelyan for his helpful remarks and suggestions that helped us to improve our paper. [1000MM.]{} J. Aaronson and M. Nadkarni, $L_\infty$ eigenvalues and $L_2$ spectra of no-singular transformations, Proc. London Math. Soc. [**55**]{} (3) (1987) pp. 538–570. D. Armacost, [*The structure of locally compact abelian groups*]{}, Monographs and Textbooks in Pure and Applied Mathematics [**68**]{}, Marcel Dekker Inc., New York, (1981). A. Bir' o, *Characterizations of groups generated by Kronecker sets*, Journal de th' eorie des nombres de Bordeaux [**19**]{} (3) (2007) pp. 567–582. G. Barbieri, D. Dikranjan, C. Milan, H. Weber, [*Answer to Raczkowki’s questions on convergent sequences of integers*]{}, Topology Appl. [**132**]{} (2003) pp. 89-101. G. Barbieri, D. Dikranjan, C. Milan, H. Weber, [*$\mathfrak t$-dense subgroups of topological abelian groups*]{}, Questions & Answers Gen. Topology [**24**]{} (2) (2006) pp. 99–118. G. Barbieri, D. Dikranjan, C. Milan, H. Weber, [*Topological torsion related to some recursive sequences of integer*]{}, Math. Nachrichten [**281**]{} (7) (2008) pp. 930-950. A. B' ir' o, J. -M. Deshouillers and V.T. Sós, [*Good approximation and characterization of subgroups of ${\mathbb{R}}/{\mathbb{Z}}$*]{}, Studia Sci. Math. Hungar. [**38**]{} (2001) pp. 97–113. M. Beiglb" ock, C. Steineder, and R. Winkler, *Sequences and filters of characters characterizing subgroups of compact abelian groups*, Topology Appl. [**153**]{} (11) (2006) pp. 1682–1695. J. P. Borel, [*Sous-groupes de ${\mathbb{R}}$ liés \` a répartition modulo $1$ de suites*]{} Ann. Fac. Sci. Toulouse Math. [**5**]{} (3-4) (1983) pp. 217–235. J.-P. Borel, [*Sur certains sous-groupes de ${\mathbb{R}}$ li' es ' a la suite des factorielles*]{}, Colloq. Math. [**62**]{} (1) (1991) pp. 21–30. L. Bukovský, N. N. Kholshevikova, M. Repický, *Thin sets in harmonic analysis and infinite combinatorics*, Real Anal. Exchange [**20**]{} (1994/1995) pp. 454–509. D. Dikranjan, [*Topologically torsion elements of topological groups*]{}, Topology Proc. [**26**]{} (2001-2002) pp. 505–532. D. Dikranjan, [*Closure operators in topological groups related to von Neumann’s kernel*]{}, Topology Appl. [**153**]{} (2006) pp. 1930–1955. D. Dikranjan and R. Di Santo, [*Answer to Armacost’s quest on topologically torsion elements of the circle group*]{}, Comm. Algebra [**32**]{} (2004) pp. 133–146. D. Dikranjan and D. Impieri, [*Topologically torsion elements of the circle group*]{}, Comm. Algebra [**42**]{} (2014) pp. 600–614. D. Dikranjan and D. Impieri, [*On the Borel complexity of characterized subgroups*]{}, preprint. D. Dikranjan and D. Impieri, [*On Hewitt groups and finer locally compact group topologies*]{}, preprint. D. Dikranjan and [S. S. Gabriyelyan]{}, [*On characterized subgroups of compact abelian groups*]{}, Topology Appl., [**160**]{} (18) (2013) pp. 2427–2442. D. Dikranjan, [S. S. Gabriyelyan]{} and V. Tarieladze, [*Characterizing sequences for precompact group topologies*]{}, J. Math. Analysis and Applications, [**412**]{} (1) (2014) pp. 505–519. D. Dikranjan and K. Kunen, *Characterizing countable subgroups of compact abelian groups*, J. Pure Appl. Algebra [**208**]{} (2007) pp. 285–291. D. Dikranjan, C. Milan and A. Tonolo, [*A characterization of the maximally almost periodic abelian groups*]{}, J. Pure Appl. Algebra [**197**]{} (1–3) (2005) pp. 23–41. D. Dikranjan, Iv. Prodanov and L. Stoyanov, [*Topological Groups: Characters, Dualities and Minimal Group Topologies*]{}, Pure and Applied Mathematics, Vol. [**130**]{}, Marcel Dekker Inc., New York-Basel, (1989). H. G. Eggleston [*Sets of fractional dimensions which occur in some problems of number theory*]{}, Proc. London Math. Soc. [**54**]{} (2) (1952) pp. 42–93. I. Farah, S. Solecki, *Borel subgroups of Polish groups*, Adv. Math. [**199**]{} (2) (2006) pp. 499–541. , *Group of quasi-invariance and the Pontryagin duality*, Topology Appl. [**157**]{} (2010) pp. 2786–2802. , *On $T$-sequence and characterized subgroups*, Topology Appl. [**157**]{} (2010) pp. 2834–2843. , *Characterizable groups: some results and open questions*, Topology Appl. [**159**]{} (2012) pp. 2378–2391. , *On $T$-characterized subgroups of compact Abelian groups*, submitted. , *On characterized subgroups of Abelian topological groups $X$ and the group of all $X$-valued null sequences*, Comment. Math. Univ. Caroli. to appear. *On Borel complexity of characterized subgroups*, Work in progress. E. Hewitt, [*A remark on characters of locally compact Abelian groups*]{}, Fund. Math. [**53**]{} (1963) pp. 55–64. G. Hjorth, [*Subgroups of abelian Polish groups*]{}, Set theory, Trends Math., BirkhŠuser, Basel, (2006) pp. 297–308. A.S. Kechris, [*Classical descriptive set theory*]{}, Graduate Texts in Mathematics [**156**]{}, Springer, (1995). A.S. Kechris and Louveau , [*The classification of the structure of hypersmooth Borel equivalence relations*]{}, J. Amer. Math. Soc. [**10**]{} (1997) pp. 215-242. K. Kuratowski, *Topology, Vol. 1*, Academic Press/PWN , New York-London-Warszawa, (1966). G. Larcher, [*A convergence problem connected with continued fractions*]{}, Proc. Amer. Math. Soc [**103**]{} (3) (1988) pp. 718–722. G. Negro, *Polish LCA groups are strongly characterizable*, Topology and its Applications [**162**]{} (2014) pp. 66–75. S. Solecki, *Polish group topologies*, in Sets and Proofs, London Mathematical Society Lecture Notes Series [**258**]{}, Cambridge University Press (1999), pp. 339–364

ArXiv

--- bibliography: - 'primer.bib' --- <span style="font-variant:small-caps;">May 2019\ Pierre Martin-Dussaud[^1]</span>\ ------------------------------------------------------------------------ \ [ **A primer of group theory\ for Loop Quantum Gravity\ and Spin-foams\ **]{} ------------------------------------------------------------------------ \ [**M. F. Atiyah**, [@Atiyah1974].]{} The aim of a theory really is, to a great extent, that of systematically organizing past experience in\ such a way that the next generation, our students and their students and so on, will be able to\ absorb the essential aspects in as painless a way as possible, and this is the only way in which you\ can go on cumulatively building up any kind of scientific activity without eventually coming to a dead end. Foreword {#foreword .unnumbered} ======== Calculations in Loop Quantum Gravity (LQG) and spin-foams theory rely heavily on group theory of $SU(2)$ and $SL_2(\mathbb{C})$. Even though many monographs exist devoted to this theory, the different tools needed (e.g. representation theory, harmonic analysis, recoupling theory...) are often dispersed in different books, with different conventions and notations. This was the initial motivation for the compilation of the present document. Generally, these notes can serve three main purposes: 1. A concise introduction for students to the essential mathematical tools of LQG. It bridges a gap between the level of students at the end of a Master programme, and the minimum level required to start doing research in LQG. In case the pace is too fast, paragraphs have been inserted for a quick refresher. They are written in small font size, and introduced by “[$\clubsuit$ <span style="font-variant:small-caps;">Reminder</span>]{}”. Instead of introducing new formulae out of nowhere, we insist on motivations for doing it. The proofs are done, if helpful for understanding, but are sometimes only sketched . They are written in small font size introduced by “[$\blacktriangleright$ <span style="font-variant:small-caps;">Proof</span>]{}”, so that they can also be skipped easily. 2. A convenient compendium for researchers. Instead of having each formula each in a different heavy, old book, the most useful ones are now gathered in a short toolbox. 3. A translational hub between the conventions of the main references. For many notions, no standard notations are universally used. Each author tends to use their own notations, which makes it difficult to switch easily from one reference to another. We have made some choices ourselves, but we show explicitly how they relate to various major references: such discussions are done in small font size introduced by “$\star$ [<span style="font-variant:small-caps;">Nota Bene</span>]{}”. We see this attempt as a step towards a more widely accepted use of common notations. In particular, we give the conventions of the Wolfram Language, which are helpful for implementing numerical computations. Although most of the content is not new, we also offer some new derivations of results, simpler than what can be found elsewhere, or sometimes not written anywhere else. A commented bibliography is provided at the end to give a panorama of the existing literature and to help readers looking for more details. These notes are aimed both at physicists, caring about their tools being mathematically well grounded, and at mathematicians, curious about how some of their familiar abstract structures can reveal the beauty of quantum gravity. For the latter, we have introduced specific short paragraphs in small font size, introduced by “[$\varheart$ <span style="font-variant:small-caps;">Physics</span>]{}”, that provide general ideas and references on how the mathematics has been exploited by theoretical physicists, especially in quantum gravity. The first chapter introduces $SU(2)$ and $SL_2(\mathbb{C})$ the two Lie groups of main interest for quantum gravity. It is aimed at gathering the main ingredients which will be extensively used later, fixing notations and refreshing memories of the reader. Chapter 2 deals with representations of $SU(2)$, and present various possible realisations that are used in the literature. Chapter 3 condenses the main results of recoupling theory of $SU(2)$. Chapter 4 deals with the representations of $SL_2(\mathbb{C})$. Finally chapter 5 wraps everything up into a mathematical presentation of quantum gravity (with loops and spin-foams). In writing these notes, I received much help from Simone Speziale, Giorgio Sarno, Fabio D’Ambrosio, Alexander Thomas, Daniel Martinez and Carlo Rovelli, who I warmly thank for this. Warmup ====== The central role of group theory in physics has been largely revealed in the modern theories of the [XX]{}$^{th}$ century. Quantum gravity makes no exception. The two main groups of interest for quantum gravity are $SL_2(\mathbb{C})$ and its subgroup $SU(2)$. This may seem natural since $SL_2(\mathbb{C})$ is, in some sense, the “quantum version” (more precisely the universal cover) of the restricted Lorentz group $SO^+(3,1)$, which is the symmetry group of Minkowski spacetime, and $SU(2)$ is the subgroup obtained when a preferred time-slice is chosen; but the reason why these groups come out in quantum gravity is actually more subtle (see section \[presentation of LQG\]). This first chapter is a quick and dense summary of the very basic tools and facts about $SL_2(\mathbb{C})$ and $SU(2)$, which will be later used extensively. It also fixes many of the notations. If you already feel warmed up, you would do well skipping this chapter. If you have never seen these notions in your life, you would do better to first learn them with an introductory book (see the references in the commented bibliography, appendix \[commented\]). Basics of SL(2,C) ----------------- The algebra of $2 \times 2$ complex matrices is denoted $\mathcal{M}_2(\mathbb{C})$. The linear group $$GL_2(\mathbb{C}) \overset{\rm def}= \left \{ M \in \mathcal{M}_2(\mathbb{C}) \mid \det M \neq 0 \right \}$$ is a Lie group. Its associated Lie algebra $\mathfrak{gl}_2(\mathbb{C})$ is actually isomorphic to $\mathcal{M}_2(\mathbb{C})$ endowed with the Lie product $\left[M,N\right] = MN-NM$. The set $\mathcal{M}_2(\mathbb{C})$ is an *algebra* because it is a vector space (with addition of matrices) endowed with a bilinear product (the usual matrix product). $GL_2(\mathbb{C})$ is not a linear subspace, but it is a complex differential manifold of dimension $4$. What’s more, the inversion and the multiplication are both analytical maps, which makes $GL_2(\mathbb{C})$ a complex *Lie group*. A complex Lie group of dimension $n$ is also a real Lie group of dimension $2n$. Therefore $GL_2(\mathbb{C})$ is a real Lie group of dimension $8$. The *associated Lie algebra* is the tangent space over the neutral element of the Lie group, usually denoted with Gothic letters. It is indeed a *Lie algebra*, that is to say an algebra where the bilinear product, called the *Lie bracket*, is antisymmetric and satisfies the Jacobi identity. The subset of invertible uni-modular matrices, $$SL_2(\mathbb{C}) \overset{\rm def}= \left \{ M \in GL_2(\mathbb{C}) \mid \det M =1 \right \},$$ is a complex Lie sub-group of dimension $3$, called the special linear group $SL_2(\mathbb{C})$. Therefore, it is also a real Lie group of dimension $6$. Topologically, $SL_2(\mathbb{C})$ is not compact, but it is simply connected. The Lie algebra of $SL_2(\mathbb{C})$ is $$\mathfrak{sl}_2(\mathbb{C}) =\left \{ M \in \mathcal{M}_2(\mathbb{C}) \mid {\rm Tr \ }M = 0 \right \}.$$ It is a $3$-dimensional complex Lie sub-algebra of $\mathfrak{gl}_2(\mathbb{C})$. A basis is given by the three *Pauli matrices*: $$\begin{aligned} \sigma_1 \overset{\rm def}=\begin{pmatrix} 0& 1 \\ 1 & 0 \end{pmatrix}, && \sigma_2 \overset{\rm def}=\begin{pmatrix} 0&-i\\i&0 \end{pmatrix}, && \sigma_3 \overset{\rm def}= \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix}.\end{aligned}$$ Interestingly, they satisfy[^2] $$[\sigma_i,\sigma_j] = 2i \, \epsilon_{ijk} \, \sigma_k.$$ With the identity matrix $\mathbb{1}$, the Pauli matrices also provide a basis for $\mathcal{M}_2(\mathbb{C})$: any $a \in \mathcal{M}_2(\mathbb{C})$ can be written uniquely $$\label{decomposition Pauli} a = a_0 \mathbb{1} + \sum_{k=1}^3 a_k \sigma_k \quad \text{with} \quad a_0,a_1,a_2,a_3 \in \mathbb{C}.$$ Note that in this basis, the determinant reads $\det a = a_0^2 - a_1^2 - a_2^2 - a_3^2$. Moreover, the identity matrix and the Pauli matrices also provide a basis to the *real* vector space of $2 \times 2$ hermitian matrices, defined by $$H_2(\mathbb{C}) \overset{\rm def}= \{ M \in \mathcal{M}_2(\mathbb{C}) \mid M^\dagger = M \}.$$ Any $h \in H_2(\mathbb{C})$ can be written uniquely as $$\label{hermitian decomposition Pauli} h = h_0 \mathbb{1} + \sum_{k=1}^3 h_k \sigma_k \quad \text{with} \quad h_0,h_1,h_2,h_3 \in \mathbb{R}.$$ The restricted Lorentz group SO+(3,1) ------------------------------------- The spacetime of special relativity is *Minkowski spacetime* $\mathbb{M}$. Mathematically, it is the vector space $\mathbb{R}^4$, endowed with a *Lorentzian inner product*, whose signature is either $(-,+,+,+)$ (general relativists convention) or $(+,-,-,-)$ (particle physicists convention). The group of all isometries (distance-preserving transformations) of Minkoswki spacetime is called the *Poincaré group* (sometimes the *inhomogeneous Lorentz group*). The isometries that leave the origin fixed form a linear subgroup, called the *Lorentz group* (sometimes the *homogeneous Lorentz group*), and denoted $O(3,1)$ (or $O(1,3)$). It is composed of four connected components related to each other by the operators of parity (space inversion) and time-reversal. The identity component forms a subgroup of $O(3,1)$, made of transformations that preserves orientation and the direction of time. It is called the *proper orthochronous Lorentz group*, or the *restricted Lorentz group*, denoted $SO^+(3,1)$. As a real vector space, Minkowski spacetime $\mathbb{M}$ is isomorphic to $H_2(\mathbb{C})$, with the map $$\label{hermitian Minkowski} X = (t,x,y,z) \mapsto h = t \mathbb{1} + x \sigma_1 + y \sigma_2 + z \sigma_3 = \begin{pmatrix} t + z & x -iy \\ x+iy & t-z \end{pmatrix}.$$ The inverse map is given by $$h \mapsto X = \frac 12 (\Tr h, \Tr h\sigma_1,\Tr h \sigma_2,\Tr h \sigma_3),$$ and the pseudo-scalar product (with convention $(+,-,-,-)$) $$X \cdot X' = tt' - xx' - yy'-zz' = - \frac 14 \Tr \left( hh' - h\sigma_1h'\sigma_1- h\sigma_2h'\sigma_2 - h \sigma_3 h' \sigma_3 \right)$$ Note that the pseudo-norm of $\mathbb{M}$ is mapped to the determinant over $H_2(\mathbb{C})$: $$X \cdot X = \det h.$$ From the latter property, we see that the action of $a \in SL_2(\mathbb{C})$ upon $h \in H_2(\mathbb{C})$, given by $$\label{SL2C action hermitian} h \mapsto a h a^\dagger,$$ defines a linear isometry on $\mathbb{M}$. Thus, it defines an homomorphism between $SO^+(3,1)$ and $SL_2(\mathbb{C})$, and it is easy to show the following isomorphism of groups $$\label{SO SL} SL_2(\mathbb{C}) / \{\mathbb{1} , -\mathbb{1} \} \cong SO^+(3,1).$$ $SL_2(\mathbb{C})$ is said to be the double cover, or the universal cover, of $SO^+(3,1)$, or sometimes also the *Lorentz spin group*. This gives a first glimpse on the role of $SL_2(\mathbb{C})$ in fundamental physics. The sub-groups of SL(2,C) {#subgroups} ------------------------- There are many sub-groups of $SL_2(\mathbb{C})$. We describe below the main ones. The figure \[graph\] shows the relations of inclusion between them. ![*This graph represents the relations of inclusions between the subgroups of $SL_2(\mathbb{C})$.*[]{data-label="graph"}](graph.png) $\blacksquare$ **$\pmb{SU(2)}$, the unitary special group,** is defined by: $$\label{def SU(2)} SU(2) \overset{\rm def}= \left \{ u \in SL_2(\mathbb{C}) \mid u^\dagger u = \mathbb{1} \right \}.$$ Any $u \in SU(2)$ can be uniquely written as $$u = u_0 e + i \sum_{k=1}^3 u_k \sigma_k \quad \text{with} \quad u_0,u_1,u_2,u_3 \in \mathbb{R} \quad {\rm and} \quad \sum_{k=0}^3 u_k^2 = 1.$$ Through the isomorphism \[hermitian Minkowski\] and the action \[SL2C action hermitian\], the definition \[def SU(2)\] enables to see $SU(2)$ as the stabiliser of the unit time vector $(1,0,0,0)$. Sometimes, the stabiliser is also called the *little group* or the *isotropy group*. Physically, it means that $SU(2)$ only acts over the space, and not in the time direction. Choosing another time direction, related to $(1,0,0,0)$ by a boost $\Lambda$, would have defined another stabilizer, isomorphic to $SU(2)$, which makes physicists sometimes talk of *a* $SU(2)$, as if there were several. $\blacksquare$ **$\pmb{SU(1,1)}$** is defined by: $$SU(1,1) \overset{\rm def}= \left \{ v \in SL_2(\mathbb{C}) \mid v^\dagger \sigma_3 v= \sigma_3 \right \}.$$ Any $v \in SU(1,1)$ can be uniquely written as: $$v = v_0 e + v_1 \sigma_1 + v_2 \sigma_2 + iv_3 \sigma_3 \quad \text{with} \quad v_0,v_1,v_2,v_3 \in \mathbb{R} \quad {\rm and} \quad v_0^2 - v_1^2 - v_2^2 + v_3^2 = 1.$$ Similarly to the $SU(2)$ case, $SU(1,1)$ can be understood by its action in Minkoswki spacetime as the stabiliser of $(0,0,0,1)$. $\blacksquare$ **$\pmb{SL_2(\mathbb{R})}$, the real linear special group,** is defined by $$SL_2(\mathbb{R}) \overset{\rm def}= \left \{ a \in \mathcal{M}_2(\mathbb{R}) \mid \det a = 1 \right \},$$ and interestingly it is also $$SL_2(\mathbb{R}) = \left \{a \in SL_2(\mathbb{C}) \mid a^\dagger \sigma_2 a = \sigma_2 \right \}.$$ Any $a \in SU(1,1)$ can be uniquely written as: $$a = a_0 e + a_1 \sigma_1 + ia_2 \sigma_2 + a_3 \sigma_3 \quad \text{with} \quad a_0,a_1,a_2,a_3 \in \mathbb{R} \quad {\rm and} \quad a_0^2 - a_1^2 + a_2^2 - a_3^2 = 1.$$ Similarly to the $SU(2)$ case, $SL_2(\mathbb{R})$ can be understood by its action in Minkoswki spacetime as the stabiliser of $(0,0,1,0)$. Following the previous sequence, it would be fair to expect next sub-group to be the one defined by$$\left \{b \in SL(2,\mathbb{C}) \mid b^\dagger \sigma_1 = \sigma_1 b^{-1} \right \}.$$ To the knowledge of the author, this group has no name and is not much studied in the literature. $\blacksquare$ **The upper $\pmb{K_+}$ and lower $\pmb{K_-}$ triangular group** are defined by: $$\begin{aligned} K_+ \overset{\rm def}= \left \{ \begin{pmatrix} \lambda^{-1} & \mu \\ 0 & \lambda \end{pmatrix} \mid \lambda \in \mathbb{C}^* \quad {\rm and} \quad \mu \in \mathbb{C} \right \} && K_- \overset{\rm def}= \left \{ \begin{pmatrix} \lambda^{-1} & 0 \\ \mu & \lambda \end{pmatrix} \mid \lambda \in \mathbb{C}^* \quad {\rm and} \quad \mu \in \mathbb{C} \right \} .\end{aligned}$$ They are also called the Borel sub-groups or the parabolic sub-groups. $\blacksquare$ **The subgroups $\pmb{Z_+}$ and $\pmb{Z_-}$** are defined by: $$\begin{aligned} Z_+ \overset{\rm def}= \left \{ \begin{pmatrix} 1 & z \\ 0 & 1 \end{pmatrix} \mid z \in \mathbb{C} \right \} && Z_- \overset{\rm def}= \left \{ \begin{pmatrix} 1 & 0 \\ z & 1 \end{pmatrix} \mid z \in \mathbb{C} \right \} .\end{aligned}$$ $\blacksquare$ **The diagonal group $\pmb{D}$** is defined by: $$D \overset{\rm def}= \left \{ \begin{pmatrix} \delta & 0 \\ 0 & \delta^{-1} \end{pmatrix} \mid \delta \in \mathbb{C}^* \right \}.$$ $\blacksquare$ **$\pmb{U(1)}$, the uni-dimensional unitary group,** seen as a subgroup of $SL_2(\mathbb{C})$ through the group isomorphism $$U(1) \overset{\rm def}= \{e^{i\theta } \mid \theta \in \mathbb{R} \} \cong \left \{ \begin{pmatrix} e^{i\theta} & 0 \\ 0 & e^{-i \theta} \end{pmatrix} \mid \theta \in \mathbb{R} \right \}.$$ $U(1)$ is also the maximal torus (biggest compact, connected, abelian Lie subgroup) of $SU(2)$. $\blacksquare$ **$\pmb{Z_2}$, the center,** defined as the subset of matrices which commute with all $SL_2(\mathbb{C})$, is shown to be $$Z_2 =\left \{ \mathbb{1},-\mathbb{1} \right \}.$$ Since it is a normal subgroup (as any center of any group), the quotient $SL_2(\mathbb{C})/Z_2$ is also a group, which can be shown to be isomorphic to the restricted Lorentz group $SO^+(3,1)$, as was already said in \[SO SL\]. Decomposing SL(2,C) {#decomposing SL} ------------------- The structural properties of a matrix group can be grasped through the study of its decompositions. We are going to present four different decompositions of $SL_2(\mathbb{C})$. First of all, let us introduce the notation $Sp(M)$ for the set of the eigenvalues of $M$, and $$H_2^{++}(\mathbb{C}) \overset{\rm def}= \left \{ M \in H_2(\mathbb{C}) \mid \forall \lambda \in Sp(H), \quad \lambda > 0 \right \},$$ for the set of $2 \times 2$ hermitian positive-definite matrices. Recall that any hermitian matrix is diagonalisable. It is the content of the spectral theorem that says that $H$ is an hermitian matrix if and only if there exists a unitary matrix $U$ and a diagonal matrix $D$ with real coefficients such that$$\begin{aligned} H = U^\dagger D U.\end{aligned}$$ The coefficients of $D$ are the eigenvalues of $H$. $\blacksquare$ **Polar decomposition.** *For all $M \in GL_2(\mathbb{C})$, there exists a unique unitary matrix $U \in U(2)$ and a unique positive-definite hermitian matrix $H \in H_2^{++}(\mathbb{C})$ such that: $$M = HU.$$* The polar decomposition actually works for the set $H_2^{++}(\mathbb{C})$ of $n \times n$ positive-definite hermitian matrices. Recall that if $H \in H_n^{++} (\mathbb{C})$, then there exists a unique matrix $S \in H_n^{++}(\mathbb{C})$, called the square root of $H$, such that $H = S^2$. Existence. Let $M \in GL_n(\mathbb{C})$ be the matrix to be decomposed, then $M^\dagger M$ is hermitian and positive. Thus we have a unique matrix $S \in H_n^{++}(\mathbb{C})$ such that $M^\dagger M = S^2$. We check finally that $MS^{-1}$ is unitary. Uniqueness. If $M=QS$ with $Q$ unitary and $S$ definite-positive hermitian, then $M^\dagger=SQ^{-1}$, so $M^\dagger M=S^2$. But $M^\dagger M$ is positive hermitian and has therefore a unique positive hermitian square root, so that $S$ is this square root and $Q$ is equal to $MS^{-1}$. For $n=1$, it is the decomposition $z=re^{i\theta}$ of a non-zero complex number. It is the reason why we call it polar decomposition (kind of generalisation of polar coordinates). Remarks: 1. The order does not matter, and the theorem would also be true with $M = UH$. 2. If $M = HU \in SL_2(\mathbb{C})$, then $U \in SU(2)$ and $\det H = 1$. The polar decomposition has been used notably by Thiemann and Winkler in their analysis of the coherent states of quantum gravity (see [@Thiemann:2000bw; @Thiemann:2000bx; @Thiemann:2000by; @Thiemann:2000aa]). $\blacksquare$ **Cartan decomposition.** *For all $g \in SL_2(\mathbb{C})$, there exists $u,v \in SU(2)$ and $r \in \mathbb{R}_+$ such that: $$g = u e^{r \sigma_3 /2} v^{-1}.$$* The proof is essentially the same as previously. Existence. Let $g \in SL_2(\mathbb{C})$ be decomposed. $g^\dagger g$ is positive-definite hermitian. With the spectral theorem, we have $v \in U(2)$ and $d$ a real diagonal matrix with strictly positive coefficients such that $g^\dagger g = v^\dagger d v$. If $\det v= e^{i \theta}$, then $u = e^{-i\theta} v \in SU(2)$ and $g^\dagger g = u^\dagger d u$. Since $\det d = 1$, one can write $d = e^{r \sigma_3}$. Note that $$e^{r\sigma_3/2} = \begin{pmatrix} e^{r/2}&0\\0&e^{-r/2} \end{pmatrix},$$ Then we show $g u e^{r \sigma_3/2} \in SU(2)$. Remark: 1. The number $r$ is called *the rapidity of the boost along the axis $z$*. 2. This theorem can be generalized to the case $SL_n(\mathbb{C})$. 3. Given the decomposition $g = u e^{r \sigma_3 /2} v^{-1}$, $r$ is uniquely determined, but $u$ and $v$ are not. The other possible choices are obtained by $(u,v) \mapsto (u e^{i \theta \sigma_3} , v e^{i \theta \sigma_3} )$, with $\theta \in \mathbb{R}$. 4. The polar decomposition of $SL_2(\mathbb{C})$ is a particular case of the Cartan decomposition where $v^{-1} = \mathbb{1}$. This decomposition has been used notably for the asymptotics of spin-foams amplitude[@Speziale2017], and also for twisted geometries in [@Langvik2016]. $\blacksquare$ **Gauss decomposition.** *Let $g \in SL_2(\mathbb{C})$ such that $g_{22} \neq 0$. There exists a unique triplet $(z_+,d,z_-) \in Z_+ \times D \times Z_-$ such that $$\label{Gauss decomposition} g=z_+dz_-$$* Explicit computation. If $g = \begin{pmatrix} \alpha & \beta \\ \gamma & \delta \end{pmatrix} \in SL(2,\mathbb{C})$, then one can write: $$g = \begin{pmatrix} 1 & \beta \delta^{-1} \\ 0 & 1 \end{pmatrix} \begin{pmatrix} \delta^{-1} & 0 \\ 0 & \delta \end{pmatrix} \begin{pmatrix} 1 & 0 \\ \gamma \delta^{-1} & 1 \end{pmatrix}$$ $\blacksquare$ **Iwasawa decomposition.** *According to the Japanese mathematician (1917 - 1998), for any matrix $M \in SL_2(\mathbb{C})$, there exists a unique triplet $(Z,D,U) \in Z_+ \times D_{\mathbb{R}_+} \times SU(2)$ that decomposes $M$: $$M = ZDU = \begin{pmatrix} 1 & z \\ 0 & 1 \end{pmatrix} \begin{pmatrix} \lambda^{-1} & 0 \\ 0 & \lambda \end{pmatrix} \begin{pmatrix} \alpha & - \beta^* \\ \beta & \alpha^* \end{pmatrix}$$ with $(z,\lambda,\alpha, \beta) \in \mathbb{C} \times \mathbb{R}^*_+ \times \mathbb{C}^2$.* This decomposition has been used notably for the study of covariant twisted geometries [@Livine2011a]. Basics of SU(2) --------------- Let us now focus on the special unitary subgroup $$SU(2) \overset{\rm def}= \left \{ u \in SL_2(\mathbb{C}) \mid u^\dagger u = \mathbb{1} \right \}.$$ It is a $3$-dimensional real Lie subgroup of the $6$-dimensional real Lie group $SL_2(\mathbb{C})$. Any $u \in SU(2)$ can be uniquely written as $$u = \begin{pmatrix} \alpha & - \beta^* \\ \beta & \alpha^* \end{pmatrix} \quad {\rm with} \quad (\alpha, \beta) \in \mathbb{C}^2, |\alpha|^2 + |\beta|^2 = 1,$$ or equivalently $$u = \begin{pmatrix} a+ib & -c+id \\ c+id & a-ib \end{pmatrix} \quad {\rm with} \quad (a,b,c,d) \in \mathbb{R}^4, a^2 + b^2 + c^2 + d^2 = 1.$$ The latter expression shows that $SU(2)$ is diffeomorphic to $S^3$ (the unit sphere of $\mathbb{R}^4$). Therefore it is connected, simply connected and compact. The center of $SU(2)$ is $Z_2 = \{ \mathbb{1},-\mathbb{1} \}$, and the quotient $SU(2)/Z_2$ is also a group, which happens to be isomorphic to $SO(3)$ (see section \[rotation\]). The group $SU(2)$ is central in quantum physics. First, it appears for the theory of the angular momentum (spin). Historically, it was also used as an approximate symmetry group for the isospin that relates protons and neutrons. Then it reappeared to describe the electro-weak interaction. In LQG, $SU(2)$ comes with the holonomies, which are obtained by exponentiation of the Ashtekar variables, used for the quantization (see section \[presentation of LQG\]). The real Lie algebra of $SU(2)$ is $$\mathfrak{su}(2) = \left \{ M \in \mathcal{M}_2(\mathbb{C}) \mid M^\dagger = - M \ {\rm and } \ {\rm Tr \ } M = 0 \right \}.$$ It is a real vector space, of which a basis is given by $( i \sigma_1, i \sigma_2,i\sigma_3 )$. Since $SU(2)$ is a compact Lie group, any element of $SU(2)$ can be written (non uniquely) as the exponential of an element of the associated Lie algebra $\mathfrak{su}(2)$ (it is a general theorem for compact Lie groups). $\blacksquare$ **Exponential decomposition.** If $u \in {\rm SU}(2)$, there exists a (non-unique) $\vec{\alpha} \in \mathbb{R}^3$ such that $$\label{exponential decomposition} u = e^{i \vec{\alpha} \cdot \vec{\sigma}} = \cos \| \vec \alpha \| I + i \sin \| \vec \alpha \| \frac{\vec \alpha}{\| \vec \alpha \|} \cdot \vec \sigma.$$ Let $M = \begin{pmatrix} a & -b^* \\ b & a^* \end{pmatrix} \in SU(2)$. The equality is easy to check for $r = \arccos(\Re a)$ and $\alpha_1 = \frac{r}{\sin r} \Im b$, $\alpha_2 = - \frac{r}{\sin r} \Re b$, $\alpha_3 = \frac{r}{\sin r} \Im a$. $\blacksquare$ **Euler angles decomposition.** For all $u \in SU(2)$, there exists $\alpha, \beta, \gamma \in \mathbb{R}$ (called Euler angles) such that: $$\label{Euler decomposition} u = e^{-\frac{i\alpha}{2} \sigma_3} e^{-\frac{i \beta}{2} \sigma_2} e^{-\frac{i \gamma}{2} \sigma_3}$$ The choice can be made unique by restricting the domain of definition of the angles, with for instance $\alpha \in ]-2\pi,2\pi [, \beta \in [0,\pi] \ \text{and} \ \gamma \in [|\alpha| , 4\pi- |\alpha|[$. Explicit computation. The right hand side (RHS) gives $$\begin{pmatrix} e^{-\frac{i(\alpha + \gamma)}{2}} \cos \beta/2 & - e^{\frac{i ( \gamma - \alpha)}{2}} \sin \beta /2\\ e^{- \frac{i(\gamma - \alpha)}{2}} \sin \beta/2 & e^{\frac{i(\alpha + \gamma)}{2}} \cos \beta /2 \end{pmatrix}$$ and for any $u \in SU(2)$, it is clearly possible to find $\alpha, \beta, \gamma \in \mathbb{R}$ to write $u$ in this form. Note that there are other conventions for the definition of Euler angles. The definition we have chosen is the one of Varshalovich ([@Varshalovich1987] p. 27) and Sakurai ([@Sakurai2011] p. 177). Instead, Rühl ([@ruhl] p. 43) and the Wolfram Language have chosen the convention $u = e^{\frac{i\alpha}{2} \sigma_3} e^{\frac{i \beta}{2} \sigma_2} e^{\frac{i \gamma}{2} \sigma_3}$. The rotations SO(3) {#rotation} ------------------- As we have said in section \[subgroups\], the action of $SU(2)$ over Minkowski space, given by \[SL2C action hermitian\], preserves the time direction. Then, $SU(2)$ acts on the spatial dimensions as the group of rotations over the Euclidean space $\mathbb{R}^3$. The group of rotations of $\mathbb{R}^3$ is $$SO(3) \overset{\rm def}= \left \{ M \in \mathcal{M}_3(\mathbb{R}) \mid M^T M = \mathbb{1} \ {\rm and} \ \det M = 1 \right \}.$$ It is a Lie group whose Lie algebra is $$\mathfrak{so}(3) = \left \{ M \in \mathcal{M}_3(\mathbb{R}) \mid M^T + M = 0 \ {\rm and} \ \Tr M = 0 \right \}.$$ We can show the following isomorphism of Lie algebra $$\mathfrak{so}(3) \cong \mathfrak{su}(2).$$ Besides we have the following group isomorphism: $$SO(3) \cong SU(2) / Z_2.$$ This can be seen with the map $$\label{homomorphism} g : \begin{pmatrix} \alpha & - \beta^* \\ \beta & \alpha^* \end{pmatrix} \mapsto \begin{pmatrix} \frac{1}{2} ( \alpha^2 + \alpha^{*2} - \beta^2 - \beta^{*2} ) & \frac{i}{2} ( \alpha^2 - \alpha^{*2} - \beta^2 + \beta^{*2} ) & \alpha \beta^* + \alpha^* \beta \\ \frac{i}{2} ( - \alpha^2 + \alpha^{*2} - \beta^2 + \beta^{*2} ) & \frac{1}{2} ( \alpha^2 + \alpha^{*2} + \beta^2 + \beta^{*2} ) & i (-\alpha \beta^* + \alpha^* \beta) \\ - \alpha \beta - \alpha^* \beta^* & i(- \alpha \beta + \alpha^* \beta^*) & \alpha \alpha^* - \beta \beta^* \end{pmatrix}$$ which is a $2$-to-$1$ onto homomorphism from $SU(2)$ to $SO(3)$. It satisfies notably $g(u) = g(-u)$. Topologically, $SO(3)$ is homeomorphic to the sphere $S^3$ where the antipodal points have been identified. It is notably connected, but not simply connected. The action of the homomorphism \[homomorphism\] over the Euler decomposition \[Euler decomposition\], shows that any rotation $r \in SO(3)$, can be decomposed as $$\begin{gathered} r = r_z(\alpha) r_y(\beta) r_z(\gamma) \\ \text{with} \ r_z(\phi) = \begin{pmatrix} \cos \phi & \sin \phi & 0 \\ - \sin \phi & \cos \phi & 0 \\ 0 & 0 & 1 \end{pmatrix} \quad \text{and} \quad r_y(\beta) = \begin{pmatrix} \cos \beta & 0 & \sin \beta \\ 0 & 1 & 0 \\ - \sin \beta & 0 & \cos \beta \end{pmatrix}.\end{gathered}$$ where $(\alpha, \beta, \gamma)$ are (any choice of) Euler angles of (any choice of) one of the two antecedents of $r$ by $g$. The unicity of the decomposition can be obtained for instance with the restriction $\alpha \in ]-\pi,\pi [, \beta \in [0,\pi] \ \text{and} \ \gamma \in [|\alpha| , 2\pi- |\alpha|[$. Integrable functions over SU(2) ------------------------------- ##### Haar measure. It can be shown that there exists a unique quasi-regular Borel measure $\mu$ over $SU(2)$ which satisfies 1. Invariant: $\mu(u) = \mu(gu) = \mu(ug)$; 2. Normalised: $\mu(SU(2)) = 1$. It is called the (two-sided normalised) Haar measure of $SU(2)$. A Borel set in $SU(2)$ is any subset of $SU(2)$ obtained from open sets through countable union, countable intersection, or taking the complement. All Borel sets form an algebra called the Borel algebra $\mathcal{B}(SU(2))$. A Borel $SU(2)$-measure $\mu$ is a non-negative function over $\mathcal{B}(SU(2))$ for which $\mu(\emptyset)=0$, and which is countable additive (the measure of a disjoint union is the sum of the disjoint sets). A Borel measure is said to be quasi-regular if it is outer regular ($ \mu(S) = \inf \{\mu(U) \mid S \subseteq U, U \ \text{open}\}$) and inner regular: $\mu(S) = \sup \{\mu(K) \mid K \subseteq S, K \ \text{compact}$). The Haar measure enables definition of integrals of functions $f$ over $SU(2)$: $$\int_{SU(2)} f(u) \, \dd \mu(u) \quad \text{also denoted} \quad \int_{SU(2)} f(u) \, \dd u$$ ##### The Hilbert space $L^2(SU(2))$. The space of complex functions over $SU(2)$ satisfying $$\int_{SU(2)} |f(u)|^2 du < \infty,$$ is denoted $L^2(SU(2))$. It is an infinite-dimensional Hilbert space with the scalar product $$(f_1,f_2) \overset{\rm def}= \int_{SU(2)} f_1^*(u) f_2(u) du.$$ Integrable functions over SL(2,C) --------------------------------- In Vilenkin ([@Gelfand:1966uq] pp. 214-215), an invariant measure $da$ over $SL_2(\mathbb{C})$ is defined, so that for any $g \in SL_2(\mathbb{C})$, we have $$da = d(ga) = d(ag) = d(a^{-1}).$$ It is given explicitly by $$da = \left( \frac{i}{2} \right)^3 |a_{12}|^{-2} \, da_{11} d\overline{a_{11}} \, da_{12} d\overline{a_{12}} \, da_{22} d\overline{a_{22}} .$$ In Rühl [@ruhl], the invariant measure is given in terms of the coefficients in the decomposition \[decomposition Pauli\], by $$da = \pi^{-4} \delta \left( a_0 - \sum_{k=1}^3 a_k^2 - 1 \right) da_0d\overline{a_0} \, da_1 d\overline{a_1} \,da_2d\overline{a_2} \,da_3d\overline{a_3}.$$ It is normalised so that the induced measure over $SU(2)$ is the same Haar measure defined in the previous section. In Rühl ([@ruhl] p. 285), it is shown that using the Cartan decomposition $a = u e^{r \sigma_3/2} v^{-1}$, we have: $$d\mu(a) = \frac{1}{4 \pi} \sinh^2r dr dudv.$$ Representation theory of SU(2) {#rep theory SU(2)} ============================== Motivated by their omnipresence in quantum physics, we are going to study the representations of $SU(2)$ over finite-dimensional Hilbert spaces[^3]. Now, due to Peter-Weyl’s theorem (see appendix \[representations and intertwiners\]), a complex finite representation of a compact group can be decomposed into a direct sum of irreps. We thus focus on irreps of $SU(2)$. Irreps of SU(2) {#abstract representation} --------------- To start with, we make use of the following correspondence, which is a particular case of the so-called *Weyl’s unitary trick*. For finite dimensional representations, we can show that the following sets of representations are in one-to-one correspondence: 1. Holomorphic representations[^4] of $SL_2(\mathbb{C})$, 2. Representations of $SU(2)$, 3. Representations of $\mathfrak{su}(2)$, 4. $\mathbb{C}$-linear representations[^5] of $\mathfrak{sl}_2(\mathbb{C})$. Moreover, this correspondence preserves invariant subspaces and equivalences of representations. We can show the following bijections [@Knapp1986]: - Restriction of the action of $SL_2(\mathbb{C})$ to that of its subgroup $SU(2)$. - Differentiation as shown in appendix \[representations and intertwiners\]. - Use the fact that $\mathfrak{sl}_2(\mathbb{C}) \cong \mathfrak{su}(2) \oplus i \mathfrak{su}(2)$. - If $G$ and $H$ are two analytical groups with $G$ simply connected, and if $\phi : \mathfrak{g} \rightarrow \mathfrak{h}$ is an homomorphism between there Lie algebras, then there exists a smooth homomorphism $\Phi : G \rightarrow H$ whose differential at the identity is $\phi$. In our case, $SL_2(\mathbb{C})$ is simply connected, and since $\phi$ is assumed $\mathbb{C}$-linear, $\Phi$ is holomorphic. We can now describe all the $\mathbb{C}$-linear irreps of $\mathfrak{sl}_2(\mathbb{C})$. ##### Theorem. *For all $n \in \mathbb{N}$, there exists an $n$-dimensional $\mathbb{C}$-linear irreducible representation of $\mathfrak{sl}_2(\mathbb{C})$, unique up to equivalence.* The proof is constructive. Analysis: assume the existence of an $(n+1)$-dimensional representation over a vector space $\mathcal{H}$. One shows then that there exists a basis $(v_i)$ with $i=0..n$ such that the action of the elements $h,e,f \in \mathfrak{sl}_2(\mathbb{C})$, defined by $$\begin{aligned} h \overset{\rm def}=\sigma_3 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix} && e \overset{\rm def}=(\sigma_1+i\sigma_2)/2 = \begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix} && f \overset{\rm def}=(\sigma_1-i\sigma_2)/2 = \begin{pmatrix} 0 & 0 \\ 1 & 0 \end{pmatrix},\end{aligned}$$ and which satisfy the commuting relations $$\begin{aligned} [h,e]=2e && [h,f]=-2f && [e,f]=h,\end{aligned}$$ is given by: $$\begin{aligned} h \cdot v_k = 2(j-k)v_k, && e \cdot v_k = k(n-k+1)v_{k-1}, && f \cdot v_k = v_{k+1}.\end{aligned}$$ Synthesis: one choose any basis $(v_i)$ and defines the action of $h$, $e$, and $f$ by the previous formulas. One checks easily that the obtained representation is irreducible. The $3$-dimensional complex vector space $\mathfrak{sl}_2(\mathbb{C})$ can also be seen as a $6$-dimensional real vector space, which has $\mathfrak{su}(2)$ as its subspace. Thus, by restriction of the action of $\mathfrak{sl}_2(\mathbb{C})$ to $\mathfrak{su}(2)$, the previously found $\mathbb{C}$-linear irreps of $\mathfrak{sl}_2(\mathbb{C})$, define also irreps of $\mathfrak{su}(2)$. Finally, by exponentiating with \[exponential decomposition\], we find all irreps of $SU(2)$ over complex vector spaces. Angular momentum realisation {#angular momentum realisation} ---------------------------- In physics textbooks, the representations of $\mathfrak{sl}_2(\mathbb{C})$ are indexed by a half-integer, called a *spin*. To each spin $j\in \mathbb{N}/2$ is associated a Hilbert space $\mathcal{Q}_j$ of dimension $2j+1$. The canonical basis, also called *magnetic basis*, is composed of the vectors (or “kets” in the Dirac language) denoted $\ket{j,m}$ with $m \in \{-j,...,0,...,j \}$. It is made orthonormal by choosing the scalar product that satisfies the property (in Dirac notations): $$\braket{j,m}{ j,n} = \delta_{mn}.$$ We now define the *angular momentum observables* $J_i \overset{\rm def}= \frac{1}{2} \sigma_i$, sometimes called simply*generators of $SU(2)$* or *generators of rotations*. They satisfy: $$\left[ J_i,J_j \right] = i \epsilon_{ijk} J_k.$$ We then define their linear action over $\mathcal{Q}_j$ by $$\begin{aligned} &J_1 \ket{j,m} = \frac 1 2 \sqrt{(j - m)(j+m+1)} \ket{j,m+1} + \frac 1 2 \sqrt{(j+m)(j-m+1)} \ket{j,m-1}, \\ &J_2 \ket{j,m} = \frac{1}{2i} \sqrt{(j - m)(j+m+1)} \ket{j,m+1} - \frac{1}{2i} \sqrt{(j+m)(j-m+1)} \ket{j,m-1}, \\ &J_3 \ket{j,m} = m \ket{j,m}. \end{aligned}$$ In some textbooks, the generators are defined as $J_i \overset{\rm def}= \frac{\hbar}{2} \sigma_i$ where $\hbar$ is the reduced Planck constant, which has the dimension of an action. Indeed, this realisation originally comes from atomic physics, where the $J_i$ represent “observables” of angular momentum. For simplicity, we are working in the units where $\hbar = 1$, keeping in mind the possibility to restore $\hbar$ explicitly at any moment by dimensional analysis. By the way, notice also that since “observables” are required to be hermitian operators, the $J_i$ are elements of $i \mathfrak{su}(2)$ and not of $\mathfrak{su}(2)$. The action of the generators over $\mathcal{Q}_j$ defines a $(2j+1)$-dimensional irrep of $SU(2)$, called the *spin-$j$ representation*. This is shown by exhibiting the following equivalence[^6] with the irreps defined in the previous section: $\ket{jm} \cong v_{j-m}$. The generators are related by $J_3 \cong h/2$, $ J_+ \cong e$ and $J_- \cong f$, where we define the *ladder operators* $J_+ \overset{\rm def}= J_1+iJ_2 $ (up) and $J_- \overset{\rm def}= J_1 - iJ_2 $ (down). Their action is $$\begin{aligned} &J_+ \ket{j,m} = \sqrt{(j - m)(j+m+1)} \ket{j,m+1}, \\ &J_- \ket{j,m} = \sqrt{(j+m)(j-m+1)} \ket{j,m-1}. \end{aligned}$$ Sometimes the action of $J_+$ over $\ket{jm}$ is written with a constant phase $e^{i\delta}$. This choice defines an equivalent representation, but the choice of a null phase (called the *Condon-Shortley convention* from [@Condon1959]) is the most widespread. Finally notice that $\ket{jm}$ is also an eigenvector of the *total angular momentum* $\vec{J}^2 \overset{\rm def}= J_1^2 + J_2^2 + J_3^2$: $$\vec{J}^2 \ket{jm} = j (j+1) \ket{jm}.$$ In fact, the $\ket{jm}$ form the unique orthonormal basis that diagonalises simultaneously the commuting operators $J_3$ and $\vec{J}^2$. We say that $J_3$ and $\vec{J}^2$ form a *complete set of commuting operators (CSCO)*. From a mathematical perspective, notice also that $\vec{J}^2$ is not a element of the algebra $i \mathfrak{su}(2)$, but an element of the *universal enveloping algebra* $\mathcal{U}(i \mathfrak{su}(2))$ whose action can be easily computed by successive action of $\mathfrak{su}(2)$. Since $\vec{J}^2$ has the property to be a quadratic element that commutes with all of $\mathcal{U}(i \mathfrak{su}(2))$, it is called the *Casimir operator* of $\mathcal{U}(i \mathfrak{su}(2))$. ##### Wigner matrix. The action of the generators of $SU(2)$ over $\mathcal{Q}_j$ defines a linear action of the group $SU(2)$ by exponentiation (see appendix \[representations and intertwiners\]). The *Wigner matrix* $D^j(g)$, is the matrix that represents the action of $g \in SU(2)$ in the $\ket{j,m}$ basis. It is thus a square matrix of size $2j+1$. By definition, the coefficients of the Wigner matrix are the functions $$D^j_{mn} (g) \overset{\rm def}= \matrixel{j,m}{ g }{ j,n}.$$ One should be aware of a small ambiguity in the notation “$\matrixel{j,m}{ g }{ j,n}$” that arises when $g$ is a matrix that belongs simultaneously to $SU(2)$ and to $\mathfrak{su}(2)$. Then it should be said explicitly if one considers the group action or the algebra action when computing $\matrixel{j,m}{ g }{ j,n}$, because it gives a different result. This ambiguity comes from the fact that physicists do not usually write explicitly if they consider the group representation $\rho$, or its differential $D\rho$. Mathematicians would write $\matrixel{j,m}{\rho(g)}{j,n}$, or $ \matrixel{j,m}{D\rho(g)}{ j,n}$. From equation \[exponential decomposition\], if $ g = e^a \in SU(2) \cap \mathfrak{su}(2)$, with $a \in \mathfrak{su}(2)$, then we know from \[representations and intertwiners\], that $\rho(g) = e^{D\rho(a)}$, but $\rho(g) \neq {D\rho(e^a)} = D\rho(g)$. In the definition of the Wigner matrix above, it is the group action which is considered. The functions $D^j_{mn}$ form an orthogonal family of $L^2(SU(2))$: $$\label{ortho D} \int_{SU(2)} dg \ \overline{ D^{j'}_{m'n'}(g)} D^j_{mn}(g) = \frac{1}{2j+1} \delta_{jj'} \delta_{mm'} \delta_{nn'}.$$ The left hand side (LHS) is the coefficient $\matrixel{j'n'}{A}{jn}$ of the operator $A : \mathcal{Q}_j \to \mathcal{Q}_{j'}$ defined by $$A \overset{\rm def}= \int du \ u^\dagger \dyad{j'm'}{jm} u.$$ We first show that $A$ is an intertwiner. If $j \neq j'$, then, by Schur’s lemmas (cf. section \[representations and intertwiners\]), $A = 0$. Otherwise, $j'=j$, and $A$ is bijective, and there exists $\lambda \in \mathbb{C}$ so that $A = \lambda \mathbb{1}$. Taking the trace on both sides, we see that $\lambda = \frac{\delta_{mm'}}{2j+1}$. In fact, the Peter-Weyl theorem (see \[representations and intertwiners\]) even asserts that the functions $D^j_{mn}$ form a basis of $L^2(SU(2))$: any function $f \in L^2(SU(2))$ can be written $$f(g) = \sum_{j \in \mathbb{N}/2} \sum_{m = -j}^j \sum_{n = -j}^j f^j_{mn} D^j_{mn}(g),$$ with coefficients $f^j_{mn} \in \mathbb{C}$. It implies notably an equivalence between the following Hilbert spaces $$\label{Peter L(SU(2))} L^2(SU(2)) \cong \bigoplus_{j \in \mathbb{N}/2} (\mathcal{Q}_j \otimes \mathcal{Q}_j^*).$$ We are going to derive explicit expressions for computing $D^j_{mn}(g)$, but we first need to introduce another realisation of the spin-$j$ irreps. Homogeneous realisation {#section homogeneous} ----------------------- Let $\mathbb{C}_{2j}[z_0,z_1]$ be the vector space of polynomials of two complex variables, homogeneous of degree $2j \in \mathbb{N}$. If $P(z_0,z_1) \in \mathbb{C}_{2j}[z_0,z_1]$, it can be written as $$P(z_0,z_1) = \sum_{k=0}^{2j} a_k z_0^k z_1^{2j-k},$$ with coefficients $a_0,...,a_{2j} \in \mathbb{C}$. The action of $SU(2)$ given by $$\label{action homogeneous} g \cdot P(\mathbf{z}) = P(g^T \mathbf{z})$$ defines a $(2j+1)$-dimensional group representation. The action satisfies $$\begin{aligned} (g_1g_2) \cdot P = g_1 \cdot (g_2 \cdot P) && {\rm and} && e \cdot P = P,\end{aligned}$$ which defines a group action over the vector space $\mathbb{C}_{2j}[z_0,z_1]$. We could also have defined the action by $P(a^{-1} \pmb z)$, $P(a^\dagger \pmb z)$ or $P(\pmb z a)$. In fact $P(\pmb z a) = P(a^T \pmb z)$, defines the same action. The convention that we have chosen here is the one of Rovelli ([@CLQG] p. 173). Moreover our convention is consistent with the choice we have made later for the representations of the principal series (cf. \[principal homogeneous\]). In Bernard ([@Bernard2012] p. 128) the convention $P(a^{-1} \pmb z)$ is used. This group action induces the following action of the generators[^7] $$\label{generators homogeneous} J_+ \cong z_0 \frac{\partial}{\partial z_1} \qquad J_- \cong z_1 \frac{\partial}{\partial z_0} \qquad J_3 \cong \frac{1}{2} \left( z_0 \frac{\partial}{\partial z_0} - z_1 \frac{\partial}{\partial z_1} \right).$$ Let $g(t) = e^{t M}$. The differential of the representation is given by (see \[representations and intertwiners\]): $$\begin{aligned} M \cdot P(z_0,z_1) &= \frac{{\rm d}}{{\rm d} t} \left( g(t) \cdot P(z_0,z_1) \right) \big|_{t=0} \\ &= \frac{{\rm d}}{{\rm d} t} \left( P(g_{11}(t) z_0 + g_{21}(t)z_1,g_{12}(t) z_0 + g_{22}(t)z_1) \right) \big|_{t=0} \\ &= (M_{11}z_0 + M_{21} z_1) \frac{\partial P}{\partial z_0} (z_0,z_1) + (M_{12}z_0 + M_{22} z_1) \frac{\partial P}{\partial z_1} (z_0 , z_1).\end{aligned}$$ Then it suffices to apply to $J_3, J_+, J_-$. This representation is equivalent to the spin-$j$ irrep through the correspondence: $$\ket{j,m} \cong \left( \frac{(2j)!}{(j+m)!(j-m)!} \right)^{1/2} z_0^{j+m} z_1^{j-m},$$ The RHS is sometimes denoted with Dirac notations $\braket{z_0z_1}{jm}$. One looks for an intertwiner $\Phi : \mathcal{Q}_j \to \mathbb{C}_{2j}[z_0,z_1]$, such that for all $i$, $J_i \cdot \Phi(\ket{j,m}) = \Phi (J_i \ket{j,m})$. We show easily the necessary condition: $$\Phi(\ket{j,m}) = c \left( \frac{(2j)!}{(j+m)!(j-m)!} \right)^{1/2} z_0^{j+m} z_1^{j-m},$$ We choose $c=1$. The homogeneous realisation is very convenient to derive an explicit expression for the Wigner matrix coefficients, as we show now. ##### Wigner matrix formula. $$\label{Wigner matrix} D^j_{mn}(g) = \left( \frac{(j+m)!(j-m)!}{(j+n)!(j-n)!} \right)^{1/2} \sum_k \binom{j+n}{k} \binom{j-n}{j+m-k} g_{11}^k g_{21}^{j+n-k} g_{12}^{j+m-k} g_{22}^{k-m-n}.$$ The sum is done over the integers $k \in \{\max (0,m+n),...,\min(j+m,j+n)\}$. First of all remark that the action \[action homogeneous\] is well defined for any $g = \begin{pmatrix} g_{11} & g_{12} \\ g_{21} & g_{22} \end{pmatrix} \in GL_2(\mathbb{C})$. Explicitly, it acts over the canonical basis of $\mathbb{C}_{2j}[z_0,z_1]$ like $$\begin{aligned} g \cdot z_0^k z_1^{2j - k} &= P\left( \begin{pmatrix} g_{11} & g_{21} \\ g_{12} & g_{22} \end{pmatrix} \binom{z_0}{z_1} \right) = P \left( \binom{g_{11} z_0 + g_{21} z_1}{g_{12} z_0 + g_{22} z_1} \right) \\ &= (g_{11} z_0 + g_{21} z_1)^k (g_{12} z_0 + g_{22} z_1)^{2j-k} \\ &= \left( \sum_{i=0}^k \binom{k}{i} (g_{11} z_0)^i (g_{21} z_1)^{k-i} \right) \left( \sum_{l=0}^{2j-k} \binom{2j-k}{l} (g_{12} z_0)^l ( g_{22} z_1)^{2j-k-l} \right) && \text{[binomial theorem]} \\ &= \sum_{i=0}^k \sum_{l=0}^{2j-k} \binom{k}{i} \binom{2j-k}{l} g_{11}^i g_{21}^{k-i} g_{12}^l g_{22}^{2j-k-l} z_0^{i+l} z_1^{2j-i-l} \\ &= \sum_{i=0}^{k} \sum_{n=i}^{2j-k+i} \left( \binom{k}{i} \binom{2j-k}{n-i} g_{11}^i g_{21}^{k-i} g_{12}^{n-i} g_{22}^{2j-k-n+i}\right) z_0^{n} z_1^{2j-n} && [n:=i+l] \\ &= \sum_{n=0}^{2j} \left( \sum_{i= \max (0,n+k-2j)}^{\min (k,n)} \binom{k}{i} \binom{2j-k}{n-i} g_{11}^i g_{21}^{k-i} g_{12}^{n-i} g_{22}^{2j-k-n+i} \right) z_0^{n} z_1^{2j-n}\end{aligned}$$ Up to there, nothing but simple computation. Now, a subtle change of variables: $$m := n-j \qquad q:=k-j.$$ (The subtlety is that $j$ is a spin, and we generalise the notation $\sum$ for half-integers bounds with still a step of $1$. A general polynomial is written $$P(\mathbf{z}) = \sum_{m=-j}^{j} a_{j+m} z_0^{j+m} z_1^{j-m}.$$ The action of $GL_2(\mathbb{C})$ is: $$\begin{aligned} g \cdot z_0^{j+q} z_1^{j-q} &= \sum_{m=-j}^{j} \left( \sum_{i= \max (0,m+q)}^{ \min(j+m,j+q)} \binom{q+j}{i} \binom{j-q}{m+j-i} g_{11}^i g_{21}^{q+j-i} g_{12}^{m+j-i} g_{22}^{-q-m+i} \right) z_0^{j+m} z_1^{j-m}.\end{aligned}$$ This explicit formula is useful to show some symmetry properties like $$\label{symmetry D conjugate} \overline{D^j_{mn}(u)} = (-1)^{m-n} D^j_{-m,-n}(u).$$ This formula is much used for numerical computations with spin-networks and spin-foams, like in [@Barrett2010]. ##### Euler angles expression. Wigner proposed also another explicit expression for its matrix, in terms of the so-called Euler angles. If $u \in SU(2)$, and $\alpha, \beta, \gamma \in \mathbb{R}^3$ are the Euler angles of $u$, such that $u = e^{-\frac{i\alpha}{2} \sigma_3} e^{-\frac{i \beta}{2} \sigma_2} e^{-\frac{i \gamma}{2} \sigma_3}$, then $$\begin{aligned} D^j_{m'm}(u) = e^{-i (\alpha m' + \gamma m )} d^j_{m'm}(\beta)\end{aligned}$$ with the *reduced Wigner matrix* $$\begin{gathered} \label{reduced Wigner} d^j_{m'm}(\beta) = \left( \frac{(j+m')!(j-m')!}{(j+m)!(j-m)!} \right)^{\frac 12} \sum_{k= \max(0,m'+m)}^{\min(j+m',j+m)} (-1)^{m'+j-k} \binom{j+m}{k} \binom{j-m}{j-k+m'} \\ \times \left(\cos \frac \beta 2 \right)^{2k-m-m'} \left(\sin \frac \beta 2 \right)^{m+m'+2j-2k}.\end{gathered}$$ The proof is very easy once given formula \[Wigner matrix\]. It only consists in applying it to the matrix $$\begin{aligned} e^{-\frac{i\alpha}{2} \sigma_3} e^{-\frac{i \beta}{2} \sigma_2} e^{-\frac{i \gamma}{2} \sigma_3} = \begin{pmatrix} e^{-\frac{i(\alpha + \gamma)}{2}} \cos \beta/2 & - e^{\frac{i ( \gamma - \alpha)}{2}} \sin \beta /2\\ e^{- \frac{i(\gamma - \alpha)}{2}} \sin \beta/2 & e^{\frac{i(\alpha + \gamma)}{2}} \cos \beta /2 \end{pmatrix}.\end{aligned}$$ This is what Rühl does ([@ruhl] p. 43), except that in his convention, the Euler angles are defined with no minus sign in front, so $\alpha \mapsto - \alpha$, $\beta \mapsto - \beta$, $\gamma \mapsto - \gamma$, and so $d^j_{m'm}(\beta)|_{\text{Rühl}} = d^j_{m'm} (-\beta)$, which finally also equals the RHO of \[reduced Wigner\] provided the coefficient $(-1)^{m'+j-k}$ is changed into $(-1)^{m+j-k}$. It is also the convention chosen by the Wolfram Language to define its function `WignerD`, hence: $$\texttt{WignerD}[\{j,m,n\},\alpha,\beta,\gamma]= e^{i (\alpha m + \gamma n )} d^j_{mn}(-\beta).$$ A proof that does not presuppose formula \[Wigner matrix\] can be found in Sakurai ([@Sakurai2011] p. 236-238), which has the same convention as ours for Euler angles. It uses the Schwinger’s oscillator model for angular momentum. It obtains the same formula as ours, but written in a slightly different way, changing the index of summation $k \rightarrow j+m-k$. Finally, Varshalovich ([@Varshalovich1987] p. 76), which also has the same convention, gives a series of variations in the way of writing the above formula. Projective realisation {#projective} ---------------------- The spin-$j$ irrep can be realised over $\mathbb{C}_{2j}[z]$, the vector space of complex polynomials of one variable $z$ of degree at most $2j$. This realisation is obtained from the $\mathbb{C}_{2j}[z_0,z_1]$ realisation by the map: $$\left\{ \begin{array}{ll} \mathbb{C}_{2j}[z_0,z_1] &\to \mathbb{C}_{2j}[z] \\ P(z_0,z_1) &\mapsto P(z,1) \end{array} \right.$$ This map is constructed from a projection from $\mathbb{C}^2$ to $\mathbb{C}$, hence the name “projective” that we give to this realisation. Sometimes it is also named the “holomorphic” realisation. From this we deduce the action of $SU(2)$ $$a \cdot f(z) = (a_{12} z + a_{22})^{2j} f \left(\frac{a_{11} z + a_{21}}{a_{12} z +a_{22}}\right),$$ and of the algebra $\mathfrak{su}(2)$ $$J_+ \cong - z^2 \frac{d}{dz} + 2jz \quad J_3 \cong z\frac{d}{dz} - j \quad J_- \cong \frac{d}{dz}.$$ The canonical basis becomes $$\ket{j,m} \cong \sqrt{\frac{(2j)!}{(j+m)!(j-m)!}} z^{j+m}.$$ We can give the following explicit expression for the scalar product that makes the canonical basis orthonormal: $$\braket{f}{g} \overset{\rm def}= \frac{i}{2} \frac{2j+1}{\pi} \int_\mathbb{C} \overline{f(z)} g(z) \frac{dzd\overline{z}}{(1+|z|^2)^{2j+2}}.$$ Spinorial realisation --------------------- The following realisation of the spin-$j$ irreps relies on notations which have been much developed by Penrose [@Penrose1984]. It was found useful for *twistor theory* [@Penrose1986], and later in quantum gravity for the so-called *twisted geometries* [@Freidel2010a; @Langvik2016]. ##### Abstract indices. We are going to use the clever *conventions of abstract indices* of Penrose ([@Penrose1984] pp. 68-115). To start with we need a set of “abstract indices” $\mathcal{L}$, that is to say a countable set of symbols. We use for instance capital letters: $$\mathcal{L} \overset{\rm def}= \{ A, B, ..., Z, A_0, ..., Z_0, A_1,... \}.$$ Then we denote $\mathfrak{S}^\bullet \overset{\rm def}= \mathbb{C}^2$, and for any abstract index $A \in \mathcal{L}$, $\mathfrak{S}^A \overset{\rm def}= \mathfrak{S}^\bullet \times \{ A \}$. Obviously $\mathfrak{S}^A$ is isomorphic to $\mathbb{C}^2$ as a complex vector space. An element of $\mathfrak{S}^A$ will be typically denoted $z^A = (z,A) \in \mathfrak{S}^A$. The abstract index $A$ serves as a marker to “type” the vector $z \in \mathbb{C}^2$ (thus $z^A \neq z^B$). This notation is very efficient to deal with several copies of the same space (here $\mathbb{C}^2$), like in tensor theory. The vector space of linear forms from $\mathbb{C}^2$ to $\mathbb{C}$, is called the dual space, and denoted $\mathfrak{S}_\bullet$. Similarly, we denote $\mathfrak{S}_A \overset{\rm def}= \mathfrak{S}_\bullet \times \{A\}$, which is trivially isomorphic to the dual space of $\mathfrak{S}^A$. Its elements, called covectors, are denoted with an abstract lower capital index, $z_A$. Then the evaluation of a covector $y_A = (y,A)$ on a vector $z^A = (z,A)$ (called a “contraction” or a “scalar product”) is denoted $y_A z^A = y(z) \in \mathbb{C}$ (the order does not matter $y_A z^A = z^A y_A$). ##### Spinors. Consider the space of formal (commutative and associative) finite sums of formal (commutative and associative) products of elements, one from each $\mathfrak{S}^{A_1} , ... , \mathfrak{S}^{A_p} , \mathfrak{S}_{B_1} , ... , \mathfrak{S}_{B_q}$. A typical element can be written: $$t^{A_1 ... A_p}_{\hspace{30 pt} B_1 ... B_q} = \sum_{i=1}^m z_{1,i}^{\hspace{10 pt} A_1} \, ... \, z_{p,i}^{\hspace{10 pt} A_p} \ y^{1,i}_{\hspace{10 pt} B_1} \, ... \, y^{q,i}_{\hspace{10 pt} B_q}.$$ Then impose the rules 1. (Homogeneity) $\forall \alpha \in \mathbb{C}, \quad (\alpha z_1^{\ A_1}) \, z_2^{\ A_2} \, ... \, z_p^{\ A_p} = z_1^{\ A_1} \, (\alpha z_2^{\ A_2}) \, ... \, z_p^{\ A_p}$ 2. (Distributivity) $(z_1^{\ A_1} + z_2^{\ A_2}) \, z_3^{A_3} \, ... \, z_p^{\ A_p} = z_1^{\ A_1} \, z_3^{\ A_3} \, ... \, z_p^{\ A_p} + z_2^{\ A_2} \, z_3^{\ A_3} \, ... \, z_p^{\ A_p}$. The resulting space is a vector space denoted $\mathfrak{S}^{A_1 ... A_p}_{B_1 ... B_q}$. Its elements are called *spinors* of type $(p,q)$, and its dimension is $2^{p+q}$. The spinor space $\mathfrak{S}^{A_1 ... A_p}_{B_1 ... B_q}$ is isomorphic, but not equal, to $\mathfrak{S}^{A_1} \otimes ... \otimes \mathfrak{S}^{A_p} \otimes \mathfrak{S}_{B_1} \otimes ... \otimes \mathfrak{S}_{B_q}$. The difference is the commutativity of the product. For instance the formal product of $\mathfrak{S}^{AB}$ is commutative (by assumption) in the sense that, for $z^A \in \mathfrak{S}^A$ and $y^B \in \mathfrak{S}^B$, $z^A y^B = y^B z^A$, whereas the tensor product is not, $z^A \otimes y^B \neq y^B \otimes z^A$, simply because $z^A \otimes y^B \in \mathfrak{S}^A \otimes \mathfrak{S}^B$ and $y^B \otimes z^A \in \mathfrak{S}^B \otimes \mathfrak{S}^A$ do not belong to the same set. Heuristically it can be said that the abstract indices keep track of the position in the tensor product. The spinor space is endowed with a bunch a basic operations defined by a set of rules. It would be utterly unpedagogical to state these rules in the most general case. On the contrary they are very intuitive for simple examples, and generalise without ambiguities for higher order spinors. 1. (Index substitution) If $z^A = (z,A) \in \mathfrak{S}^A$, we denote $z^B = (z,B) \in \mathfrak{S}^B$. Thus $z^A \neq z^B$. 2. (Index permutation) If $t^{AB} = \sum_i z_i^A y_i^B \in \mathfrak{S}^{AB}$, we denote $t^{BA} = \sum_i z_i^B y_i^A \in \mathfrak{S}^{AB}$. 3. (Symmetrisation) $t^{(AB)} \overset{\rm def}= \frac 12 (t^{AB} + t^{BA} )$ or generally $z^{(A_1...A_n)} \overset{\rm def}= \frac{1}{n!} \sum_{\sigma \in S_n}z^{A_{\sigma(1)}...A_{\sigma(n)}}$. 4. (Anti-symmetrisation) $t^{[AB]} \overset{\rm def}= \frac 12 (t^{AB} - t^{BA} )$ or generally $z^{[A_1...A_n]} \overset{\rm def}= \frac{1}{n!} \sum_{\sigma \in S_n} \epsilon_\sigma z^{A_{\sigma(1)}...A_{\sigma(n)}}$, with $\epsilon_\sigma$ the signature of the permutation $\sigma$. 5. (Contraction) If $\sum_i z_i^A y^i_B = t^A_{\ B}$, then $t^A_{\ A} = \sum_i z_i^A y^i_A \in \mathbb{C}$. ##### Index dualisation We denote the canonical basis of $\mathbb{C}^2$: $$e_0 \overset{\rm def}= \binom{1}{0} \hspace{3 cm} e_1 \overset{\rm def}= \binom{0}{1}.$$ It is easy to show that there exists a unique normalised skew-symmetric spinor of type $(0,2)$. It is denoted $\epsilon_{AB}$, and satisfies by definition: $$\epsilon_{AB} = - \epsilon_{BA}, \hspace{3 cm} \epsilon_{AB} e_0^A e_1^B = 1.$$ $\epsilon_{AB}$ corresponds over $\mathbb{C}^2$ to the unique $2$-form $\epsilon$ normalised by the condition $\epsilon(e_0,e_1)=1$, which is nothing but the determinant over $\mathbb{C}^2$. For two vectors $z = (z_0,z_1)$ and $y=(y_0,y_1)$, we show easily that: $$\label{determinant spinor} \epsilon_{AB} z^A y^B = z_0 y_1 - z_1 y_0.$$ Interestingly, $\epsilon_{AB}$ defines a canonical mapping between $\mathfrak{S}^A$ and $\mathfrak{S}_A$, given by $$z^A \mapsto z_A = z^B \epsilon_{BA}.$$ It is called *index dualisation*. Actually the mapping could also have been defined by $z^A \mapsto z_A = z^B \epsilon_{AB}$. It gives the same mapping “up to a sign”, since $\epsilon_{AB}$ is skew-symmetric. This other convention is chosen by Rovelli ([@CLQG] p. 23). Our convention is the one of Penrose ([@Penrose1984] p. 104). The covectors of the dual space $\mathfrak{S}_\bullet$ can also be described by a pair of components in the dual basis. The index dualisation can then be expressed in components as $((z_0,z_1),A) \mapsto ((-z_1,z_0),A)$. Similarly to the usual Dirac notation $\ket{z} = (z_0,z_1)$, a notation is sometimes introduced for the dual $[z| = (-z_1,z_0)$. With this choice, the RHS of \[determinant spinor\] reads $[z|y\rangle$. ##### Conjugation. The conjugation of $z \in \mathbb{C}$ is denoted $\bar z$ or $z^*$. We define the conjugation over $\mathfrak{S}^A$ by $\overline{z^A} = \overline{(z,A)} \overset{\rm def}=(\overline{z},\dot A) = \bar z^{\dot A} \in \mathfrak{S}^{\dot A}$. Thus we have introduced a new set of abstract indices, the dotted indices: $$\dot{\mathcal{L}} \overset{\rm def}= \{ \dot A, \dot B, ..., \dot Z, \dot A_0, ..., \dot Z_0, \dot A_1,... \}.$$ We impose moreover that $\overline{\bar z^{\dot A}} = z^A$, i.e. $\ddot A = A$, so that the conjugation is an involution. Importantly, we regard the set $\mathcal{L}$ and $\dot{\mathcal{L}}$ as incompatible classes of abstract indices, meaning that we forbid index substitution between them two. In other words, dotted and undotted indices commute: for any $t^{A \dot B} \in \mathfrak{S}^{A \dot B}$, we have $t^{A \dot B} = t^{\dot B A}$. One way to formalise this “incompatibility” between dotted and undotted indices would be to define rather $z^A = (z,A,0)$ and $z^{\dot A} = (z,A,1)$. Thus the index substitution $z^A = (z,A,0) \mapsto z^A = (z,B,0) = z^B $ clearly does not enable to translate from a dotted to an undotted index. Only the complex conjugation can through $z^A = (z,A,0) \mapsto (\overline{z},A,1) = z^{\dot A}$. ##### Inner product. We define the map $J$ by: $$J \binom{z_0}{z_1} = \binom{-\overline{z_1}}{\overline{z_0}}.$$ Using the previously introduced generalised Dirac notation, we read $J \ket{z} = |z]$. Since $J^2 = - 1$, the map $J$ behaves over $\mathbb{C}^2$ very much as the imaginary number $i$ behaves over $\mathbb{C}$. For this reason the map $J$ is said to define a *complex structure* over $\mathbb{C}^2$. A combination of $\epsilon_{AB}$ and $J$ defines an inner product over $\mathfrak{S}^A$: $$- \epsilon_{AB} (Jz)^A y^B = \overline{z_0}y_0 + \overline{z_1} y_1.$$ In generalised Dirac notations, we read $- [Jz | y\rangle = \braket{z}{y}$, which is consistent with the usual Dirac notation for the scalar product. With the matrix action over spinors, defined just below, equation \[action over spinors\], we can see that the inner product is invariant under the action of $SU(2)$: $\braket{u \cdot z }{u \cdot y} = \braket{z }{y}$. There is no surprise since it is actually one way of defining $SU(2)$. However the inner product is not invariant under $SL_2(\mathbb{C})$ (contrary to the determinant). So, to “another choice of $SU(2)$”, in the sense of a stabilizer of a time direction (see \[subgroups\]), would correspond another invariant inner product, and thus another complex structure $J$. For instance, in [@Livine2011a], they choose rather $(-J)$ for the complex structure. The choice we have made here is the one of Rovelli ([@CLQG] p. 24). ##### Representation. The vector space $\mathcal{M}_2(\mathbb{C})$ is isomorphic to $\mathfrak{S}^A_B$, through the isomorphism that associates to any $t \in \mathcal{M}_2(\mathbb{C})$ the unique spinor $t^A_{\ \ B}$ such that: $$\label{action over spinors} \forall z \in \mathbb{C}^2, \quad (tz)^A = t^A_{\ \ B} \, z^B.$$ Then the groups $SL_2(\mathbb{C})$ and $SU(2)$ can be represented over $\mathfrak{S}^{A_1...A_p}$ such as: $$u \cdot z^{A_1...A_p} = u^{A_1}_{\ \ B_1} ... u^{A_p}_{\ \ B_p} \, z^{B_1...B_p}.$$ Yet this representation is not irreducible, since it is stable over the subspace of completely symmetric spinors $\mathfrak{S}^{(A_1...A_p)}$. Let’s see that the action is stable over this subspace. Suppose $z^{A_1...A_p} $ is completely symmetric, i.e. $z^{(A_1...A_p)} = z^{A_1...A_p}$. Then (sketching the proof, the details are left to the reader): $$\begin{aligned} u \cdot z^{(A_1...A_p)} &= u^{(A_1}_{\ \ B_1} ... u^{A_p)}_{\ \ B_p} \, z^{B_1...B_p} \\ &= u^{A_1}_{\ \ (B_1} ... u^{A_p}_{\ \ B_p)} \, z^{B_1...B_p} \\ &= u^{A_1}_{\ \ B_1} ... u^{A_p}_{\ \ B_p} \, z^{(B_1...B_p)} \\ &= u^{A_1}_{\ \ B_1} ... u^{A_p}_{\ \ B_p} \, z^{B_1...B_p} \\ &= u \cdot z^{A_1...A_p} \end{aligned}$$ thus $u \cdot z^{A_1...A_p} $ is also completely symmetric. Thus $SL_2(\mathbb{C})$ and $SU(2)$ can be represented of the vector space $\mathfrak{S}^{(A_1...A_p)}$ of dimension $p+1$. A basis is given by $$\{ e^{(A_1}_{i_1}... e^{A_p)}_{i_p} \mid i_1,...,i_p \in \{0,1\} \} = \{ e^{(A_1}_0...e^{A_m}_0 e^{A_{m+1}}_1 ... e^{A_p)}_1 \mid m \in \{0,...,p\} \}.$$ This representation is irreducible and equivalent to the spin $p/2$ representation through the intertwiner:$$e^{(A_1}_0...e^{A_m}_0 e^{A_{m+1}}_1 ... e^{A_p)}_1 \cong z_0^m z_1^{p-m}.$$ To see this, it is simpler to write the spinors in the canonical basis of $\mathfrak{S}^{A_1...A_p}$:$$\xi^{A_1...A_p} = \sum_{i_1,i_2,...,i_p=0}^1 c^{i_1,...,i_p} \, e^{A_1}_{i_1}... e^{A_p}_{i_p}$$ The total symmetry of $\xi^{A_1...A_p}$ imposes a total symmetry of the coefficients $c^{i_1,...,i_p}$. Then we define the following bijection between $\mathfrak{S}^{(A_1...A_p)}$ and $\mathbb{C}_p[z_0,z_1]$: $$\xi^{A_1...A_p} \cong \sum_{p_1,...p_n=0}^1 c^{p_1...p_n} z_{p_1} ... z_{p_2}$$ This defines an intertwiner as can be checked by looking at the action of a group element. Recoupling theory of SU(2) ========================== The $SU(2)$ irreps provide the fundamental building blocks of quantum space-time. From a mathematical perspective, irreps are the fundamental bricks from which other representations are built. Indeed any finite representation of $SU(2)$ is completely reducible (i.e. it can be written as a direct sum of irreps). In particular, a tensor product of irreps can be decomposed into a direct sum of irreps: i.e. there exists a bijective intertwiner that maps the tensor product to a direct sum of irreps. Such an intertwiner is sometimes called a “*coupling tensor*” (see Moussouris [@Moussouris:1983uq] pp. 10-11). This naming comes from quantum physics: when two systems *couple* (i.e. interact), the total system is described by states of the tensor product of the Hilbert spaces of the subsystems. Notice that there may exist several coupling tensors between a tensor product and its corresponding sum of irreps. It is precisely the goal of “*recoupling theory*” to describe these coupling tensors and to understand how one can translate from one decomposition to another. Clebsch-Gordan coefficients --------------------------- Given $\mathcal{Q}_{j_1}$ and $\mathcal{Q}_{j_2}$, two irreps of $SU(2)$, the tensor representation is defined over $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2}$. The canonical basis of $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2}$ is given by the elements $$\ket{j_1m_1;j_2m_2} \overset{\rm def}= \ket{j_1,m_1} \otimes \ket{j_2 ,m _2}$$ where $m_1$ and $m_2$ belong to the usual range of magnetic indices. Interestingly, this basis is the unique orthonormal basis that diagonalises simultaneously the commuting operators: $$J_3 \otimes \mathbb{1}, \quad \mathbb{1} \otimes J_3, \quad \vec{J}^2 \otimes \mathbb{1}, \quad \mathbb{1} \otimes \vec{J}^2.$$ Another complete set of commuting operators is given by $$J_3 \otimes \mathbb{1} + \mathbb{1} \otimes J_3, \quad \left( \vec{J} \otimes \mathbb{1} + \mathbb{1} \otimes \vec{J} \right)^2, \quad \vec{J}^2 \otimes \mathbb{1}, \quad \mathbb{1} \otimes \vec{J}^2.$$ Therefore, there exists an orthonormal basis that diagonalises them simultaneously. It is given by $$\ket{j_1j_2;k;n} \quad {\rm with} \ k \in \{|j_1-j_2|,...,j_1+j_2 \} \quad {\rm and} \ n\in \{-k,...,k\},$$ and characterised by the action of the operators: $$\begin{aligned} &(J_3 \otimes \mathbb{1} + \mathbb{1} \otimes J_3 ) \ket{j_1j_2;k;n} = n \ket{j_1j_2;k;n}\\ &\left( \vec{J} \otimes \mathbb{1} + \mathbb{1} \otimes \vec{J} \right)^2 \ket{j_1j_2;k;n} = k(k+1)\ket{j_1j_2;k;n}\\ &\vec{J}^2 \otimes \mathbb{1}\ket{j_1j_2;k;n} = j_1(j_1+1)\ket{j_1j_2;k;n} \\ &\mathbb{1} \otimes \vec{J}^2\ket{j_1j_2;k;n} = j_2 (j_2+1) \ket{j_1j_2;k;n}.\end{aligned}$$ A proof can be found in Sakurai ([@Sakurai2011] pp. 217-231). This result proves that $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2}$ can be decomposed into a direct sum of irreps, namely we have the following equivalence of representations $$\label{decomposition of 2} \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \cong \bigoplus_{k=|j_1-j_2|}^{j_1+j_2} \mathcal{Q}_k.$$ The equivalence is given by the bijective intertwiner $\iota : \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \to \bigoplus_{k=|j_1-j_2|}^{j_1+j_2} \mathcal{Q}_k$, which satisfies $$\label{intertwiner iota} \iota\ket{j_1,j_2;km} = \ket{km} .$$ We define the *Clebsch-Gordan coefficients* by the scalar product $$C^{jm}_{j_1m_1j_2m_2} \overset{\rm def}= \braket{j_1m_1j_2m_2}{j_1j_2;jm}.$$ Said differently, we have $$\label{CG translation} \ket{j_1j_2;j,m} = \sum_{m_1 = -j_1}^{j_1} \sum_{m_2 = -j_2}^{j_2} C^{jm}_{j_1m_1j_2m_2} \ket{j_1m_1j_2m_2}.$$ The Clebsh-Gordan coefficients can be seen as the matrix coefficients of the intertwiner in the canonical bases. The Clebsh-Gordan coefficients appear largely in the quantum theory of angular momentum. Two spin-systems are described as one single spin-system with a larger total angular momentum. ##### Remarks 1. Due to the Condon-Shortley convention for the $SU(2)$-action, we have $C^{jm}_{j_1m_1j_2m_2} \in \mathbb{R}$. 2. The coefficients $C^{jm}_{j_1m_1j_2m_2}$ are well-defined and non-zero, only if the following *Clebsch-Gordan inequality* (aka *triangle inequality*) is satisfied $$|j_1-j_2|\leq j \leq j_1+j_2.$$ Otherwise, we choose by convention, that $C^{jm}_{j_1m_1j_2m_2} = 0$. 3. If $m \neq m_1+m_1$, then $C^{jm}_{j_1m_1j_2m_2} =0$. 4. Since the $\ket{jm}$ form an orthonormal basis, we have the following “orthogonality relations” $$\label{CG orthogonalite} \sum_{m_1 = -j_1}^{j_1} \sum_{m_2 = -j_2}^{j_2} C^{jm}_{j_1m_1j_2m_2} C^{j'm'}_{j_1m_1j_2m_2} = \delta_{j,j'} \delta_{m,m'}.$$ The Clebsch-Gordan coefficients are numbers, but their definition is quite implicit. Hopefully, there are also explicit formulas to compute them! ##### Explicit formula $$\begin{gathered} \label{CG explicit} C^{jm}_{j_1m_1j_2m_2} = \delta_{m,m_1+m_2} \sqrt{2j+1} \sqrt{\frac{(j+m)!(j-m)!(-j+j_1+j_2)!(j-j_1+j_2)!(j+j_1-j_2)!}{(j+j_1+j_2+1)!(j_1+m_1)!(j_1-m_1)!(j_2+m_2)!(j_2-m_2)!}} \\ \times \sum_k \frac{(-1)^{k+j_2+m_2} (j+j_2+m_1-k)!(j_1-m_1+k)!}{(j-j_1+j_2-k)!(j+m-k)!k!(k+j_1-j_2-m)!}\end{gathered}$$ Other similar expressions can be found in Varshalovich ([@Varshalovich1987] p. 238), with references to various proofs. It can be notably convenient to recognise in the sum above the so-called *hyper-geometrical function ${}_3F_2$*:$$\begin{gathered} C^{jm}_{j_1m_1j_2m_2} = \delta_{m,m_1+m_2} \sqrt{2j+1}\frac{\sqrt{(j+j_1-j_2)!(j-j_1+j_2)!}}{\sqrt{(-j+j_1+j_2)!(j+j_1+j_2+1)!}} \frac{\sqrt{(j+m)!(j-m)!(j_1+m_1)!(j_2-m_2)!}}{\sqrt{(j_1-m_1)!(j_2+m_2)!}} \\ \times {}_3F_2 (j-j_1-j_2,m_1-j_1,-j_2-m_2 ;j-j_2+m_1+1,j+j_1-m_2+1;1).\end{gathered}$$ In the Wolfram Language, they are implemented as $$C^{j_3m_3}_{j_1m_1j_2m_2} = \texttt{ClebschGordan}[\{j_1, m_1\}, \{j_2, m_2\}, \{j_3, m_3\}] .$$. ##### Exercise. Show that $$\label{exercise 1} D^{j_1}_{m_1n_1}(g) D^{j_2}_{m_2n_2} (g) = \sum_{j \in \mathbb{N}/2} \sum_{m=-j}^j \sum_{m'=-j}^{j} C^{jm}_{j_1m_1j_2m_2} C^{jm'}_{j_1n_1j_2n_2} D^j_{mm'}(g).$$ $$\begin{aligned} D^{j_1}_{m_1n_1}(g) D^{j_2}_{m_2n_2} (g) &= \bra{j_1m_1} g \ket{j_1n_1} \bra{j_2m_2} g \ket{j_2n_2} \\ &= \bra{j_1m_1 ; j_2m_2} g \ket{j_1n_1 ; j_2n_2} \\ &= \sum_{j \in \mathbb{N}/2} \sum_{m=-j}^j \sum_{j' \in \mathbb{N}/2} \sum_{m'=-j'}^{j'} \braket{j_1m_1j_2m_2}{jm} \bra{jm} g \ket{j'm'} \braket{j'm'}{j_1n_1j_2n_2} \\ &= \sum_{j \in \mathbb{N}/2} \sum_{m=-j}^j \sum_{m'=-j}^{j} C^{jm}_{j_1m_1j_2m_2} C^{jm'}_{j_1n_1j_2n_2} D^j_{mm'}(g)\end{aligned}$$ Invariant subspace ------------------ A general tensor product of $n$ irreps can be decomposed into a direct sum $$\bigotimes_{i=1}^n \mathcal{Q}_{j_i} \cong \bigoplus_{k=0}^{J} \left( \underbrace{\mathcal{Q}_{k} \oplus ... \oplus \mathcal{Q}_k}_{d_k \ {\rm times}} \right),$$ where $J=\sum_i j_i$ and $d_k$ is the degeneracy of the irrep $\mathcal{Q}_k$. Here, “decomposing” means “finding a bijective intertwiner between the two spaces”. Concretely, such a decomposition is obtained by applying successively the decomposition of only two, given by \[decomposition of 2\]. The operator $\mathbb{1} \otimes ... \otimes J_i \otimes ... \otimes \mathbb{1}$ corresponding to $J_i$ acting on the $k^{th}$ Hilbert space of the product is denoted $(J_i)_k$. The three components $(J_1)_k, (J_2)_k, (J_3)_k$ form the vectorial operator $\vec{J}_k$. In quantum gravity, such tensor spaces appear in the *kinematical Hilbert space* $\mathcal{H}$. The description of the dynamics requires to impose constraints that select subspaces of $\mathcal{H}$. One important constraint is the *Gauss constraint* which reduces $\mathcal{H}$ to its $SU(2)$-invariant subspace ${\rm Inv}_{SU(2)} \mathcal{H}$ that we define below. We define the $SU(2)$-invariant subspace as $${\rm Inv}_{SU(2)} \left( \bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right) \overset{\rm def}= \left\{ \psi \in \bigotimes_{i=1}^n \mathcal{Q}_{j_i} \mid \forall g \in SU(2), \ g \cdot \psi = \psi \right\}.$$ It can also be characterized rather by the action of the algebra: $$\label{invariant J=0} {\rm Inv}_{SU(2)} \left( \bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right) = \left\{ \psi \in \bigotimes_{i=1}^n \mathcal{Q}_{j_i} \mid \forall s \in \mathfrak{su}(2), \ s \cdot \psi = 0 \right\}.$$ From this, it is easy to see that $${\rm Inv}_{SU(2)} \left( \bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right) \cong \underbrace{\mathcal{Q}_{0} \oplus ... \oplus \mathcal{Q}_0}_{d_0 \ {\rm times}},$$ where $\mathcal{Q}_0 \cong \mathbb{C}$ is the trivial representation. Interestingly, we also have the following isomorphism: $$\label{iso intertwiners} {\rm Inv}_{SU(2)} \left(\bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right) \cong {\rm Hom}_{SU(2)} \left(\bigotimes_{i=1}^n \mathcal{Q}_{j_i} , \mathcal{Q}_{0} \right),$$ where the RHS is the vector space of $SU(2)$-intertwiners between $\bigotimes_{i=1}^n \mathcal{Q}_{j_i}$ and $\mathcal{Q}_{0}$. Let $T:\bigotimes_{k=1}^n \mathcal{Q}_{j_k} \rightarrow \mathcal{Q}_0$ be an intertwiner. Since $T$ is a linear form, there exists $\psi_T \in \bigotimes_{k=1}^n \mathcal{Q}_{j_k}$ such that $T (\phi) = \braket{\phi}{\psi_T}$. Since $T$ is also an intertwiner, we have for all $\psi \in \bigotimes_{k=1}^n \mathcal{Q}_{j_k}$ and $u \in SU(2)$, $\braket{\phi}{u \cdot \psi_T} = \braket{u^\dagger \cdot \phi}{ \psi_T} = T(u^\dagger \cdot \phi) = u^\dagger \cdot T(\phi) = T(\phi) = \braket{\phi}{\psi_T}$. So $u \cdot \psi_T = \psi_T$. We can check that the map $T \mapsto \psi_T$ is linear and bijective. QED. ##### Orthogonal projector. By definition, the orthogonal projector $P$ over ${\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right)$ satisfies $$P^2 = P \quad \text{and} \quad P^\dagger = P.$$ It is easy to show that $$\label{projector integral} P = \int_{SU(2)} dg \bigotimes_{k=1}^n D^{j_k}(g).$$ Remember also that if $\ket{j}$ is an orthonormal basis of ${\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^n \mathcal{Q}_{j_i} \right)$, then $P$ can also be written as $$\label{projector dyad} P = \sum_{j} \dyad{j}{j}.$$ Wigner’s 3jm-symbol ------------------- We can decompose $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3}$ into a direct sum by applying \[decomposition of 2\] first on the left tensor product: $$\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \rightarrow \left( \bigoplus_{j_{12}} \mathcal{Q}_{j_{12}} \right) \otimes \mathcal{Q}_{j_3} \rightarrow \bigoplus_{j_{12}=|j_1-j_2|}^{j_1+j_2} \bigoplus_{k=|j_{12}-j_3|}^{j_{12}+j_3} \mathcal{Q}_k$$ Thus we construct an orthonormal basis of $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3}$ given by the states $$\begin{gathered} \label{3CG translation} \ket{(j_1j_2)j_3;j_{12}kn} = \sum_{m_1,m_2,m_3,m_{12}} C^{j_{12}m_{12}}_{j_1m_1j_2m_2} C^{kn}_{j_{12}m_{12}j_3m_3} \bigotimes_{i=1}^3 \ket{j_i,m_i}, \\ {\rm with} \ j_{12} \in \{|j_1-j_2|,...,j_1+j_2 \} \quad {\rm and} \ k \in \{|j_{12}-j_3|,...,j_{12} + j_3 \} \quad {\rm and} \ n\in \{-k,...,k\}.\end{gathered}$$ This orthonormal basis is obtained by twice applying the relation \[3CG translation\], first to the left tensor product. Notice that applying it first on the right would instead build the states: $$\ket{j_1(j_2j_3);j_{23}kn} = \sum_{m_1,m_2,m_3,m_{23}} C^{j_{23}m_{23}}_{j_2m_2j_3m_3} C^{kn}_{j_1m_1j_{23}m_{23}} \bigotimes_{i=1}^3 \ket{j_i,m_i}.$$ Then we show $${\rm Inv}_{SU(2)} \left( \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \right) = {\rm Span} \left\{ \ket{(j_1j_2)j_3;j_300} \right\}.$$ First we show the equivalence: $$\begin{aligned} \vec{J}^2 \ket{(j_1j_2)j_3;j_{12}kn} = 0 \quad \Leftrightarrow \quad k=0.\end{aligned}$$ We conclude using the characterisation \[invariant J=0\]. Thus if the Clebsch-Gordan condition is satisfied ($|j_1-j_2| \leq j_3 \leq j_1+j_2$), ${\rm Inv}_{SU(2)} ( \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3})$ is one dimensional. Otherwise ${\rm Inv}_{SU(2)} ( \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3}) = \{ 0 \}$. Now supposing that the condition is satisfied, there exists a unique unit vector in ${\rm Inv}_{SU(2)} ( \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3}) $, $$\ket{0} = \sum_{m_1,m_2,m_3} \begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix} \bigotimes_{k=1}^3 \ket{j_k,m_k},$$ such that the coefficients $\begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix}$ in the canonical basis are real and satisfy the symmetry properties: $$\label{sym 3jm def} \begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix} =\begin{pmatrix} j_3 & j_1 & j_2 \\ m_3 & m_1 & m_2 \end{pmatrix} =\begin{pmatrix} j_2 & j_3 & j_1 \\ m_2 & m_3 & m_1 \end{pmatrix}.$$ The coefficients $\begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix}$ are called the *Wigner’s $3jm$-symbol* and are related to the Clebsch-Gordan coefficients by $$\label{definition 3jm} \begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix} = \frac{(-1)^{j_1 - j_2 - m_3}}{\sqrt{2 j_3 + 1}} C^{j_3,-m_3}_{j_1m_1j_2m_2}.$$ In Mathematica, they are given by `ThreeJSymbol`$[\{j_1, m_1 \}, \{j_2, m_2 \}, \{j_3, m_3 \}]$. In ${\rm Inv}_{SU(2)} \left( \mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \right)$, all vectors are proportional to $$\begin{aligned} \ket{(j_1j_2)j_3;j_300} &= \sum_{m_1,m_2,m_3,m} C^{00}_{jmj_3m_3} C^{jm}_{j_1m_1;j_2m_2} \ket{j_1m_1;j_2m_2;j_3m_3} & \text { [applying \ref{3CG translation}]} \\ &= \sum_{m_1,m_2,m_3,m} \delta_{m,-m_3} \delta_{j,j_3} \frac{(-1)^{j_3 + m_3}}{\sqrt{2j_3 +1}} C^{jm}_{j_1m_1;j_2m_2} \ket{j_1m_1;j_2m_2;j_3m_3} & \text { [computing \ref{CG explicit}]} \\ &= \sum_{m_1,m_2,m_3} \frac{(-1)^{j_3 + m_3}}{\sqrt{2j_3 +1}} C^{j_3, -m_3}_{j_1m_1;j_2m_2} \ket{j_1m_1;j_2m_2;j_3m_3} & \text { [simplifying]}.\end{aligned}$$ The proportionality factor is chosen to be $(-1)^{j_1 - j_2 + j_3}$ to match the reality and the symmetry requirements. ##### Remarks. 1. These symbols satisfy nice symmetry properties such as: $$\label{sym 3jm} \begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix} = (-1)^{j_1+j_2+j_3} \begin{pmatrix} j_2 & j_1 & j_3 \\ m_2 & m_1 & m_3 \end{pmatrix} = (-1)^{j_1+j_2+j_3} \begin{pmatrix} j_1 & j_2 & j_3 \\ -m_1 & -m_2 & -m_3 \end{pmatrix}.$$ 2. The orthogonality relations \[CG orthogonalite\] become $$\sum_{jm} (2j+1) \begin{pmatrix} j_1 & j_2 & j \\ m_1 & m_2 & m \end{pmatrix} \begin{pmatrix} j_1 & j_2 & j \\ m_1' & m_2' & m \end{pmatrix} = \delta_{m_1 m_1'} \delta_{m_2 m_2'}$$ $$\label{ortho 3jm sum mm} \sum_{m_1m_2} (2j+1) \begin{pmatrix} j_1 & j_2 & j \\ m_1 & m_2 & m \end{pmatrix} \begin{pmatrix} j_1 & j_2 & j' \\ m_1 & m_2 & m' \end{pmatrix} = \delta_{jj'} \delta_{m m'}$$ ##### Exercise. Show that $$\label{projector 3D} \int_{SU(2)} D^{j_1}_{m_1n_1} (u) D^{j_2}_{m_2n_2}(u) D^{j_3}_{m_3n_3}(u) du = \begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix} \begin{pmatrix} j_1 & j_2 & j_3 \\ n_1 & n_2 & n_3 \end{pmatrix}$$ The result can be directly obtained by equating \[projector integral\] and \[projector dyad\], and expressing the equality in the magnetic basis. Wigner’s 4jm-symbol ------------------- Similarly to the previous section we can decompose $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \otimes \mathcal{Q}_{j_4}$ into a direct sum by applying \[decomposition of 2\] successively: $$\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \otimes \mathcal{Q}_{j_4} \cong \bigoplus_{j_{12}=|j_1-j_2|}^{j_1+j_2} \bigoplus_{k=|j_{12}-j_3|}^{j_{12}+j_3} \bigoplus_{l=|k-j_4|}^{k+j_4} \mathcal{Q}_l$$ In particular, we can see that $$\begin{gathered} {\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^4 \mathcal{Q}_{j_i} \right) \cong \underbrace{\mathcal{Q}_0 \oplus ... \oplus \mathcal{Q}_0}_{d_0 \ \text{times}} \\ \text{with} \ d_0 =\max (|j_1-j_2|,|j_3-j_4|) - \min (j_1+j_2,j_3+j_4).\end{gathered}$$ An orthonormal basis of ${\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^4 \mathcal{Q}_{j_i} \right) $ is given by $$\begin{gathered} \label{intertwiner 4jm} \ket{j}_{12} = \sum_{m_1,m_2,m_3,m_4} \sqrt{2j+1} \begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j)} \bigotimes_{k=1}^4 \ket{j_k,m_k}, \\ {\rm with} \ j \in \{ \max (|j_1-j_2|,|j_3-j_4|) ,...,\min (j_1+j_2,j_3+j_4) \}\end{gathered}$$ and $$\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j)} \overset{\rm def}= \sum_{m} (-1)^{j - m} \begin{pmatrix} j_1 & j_2 & j \\ m_1 & m_2 & m \end{pmatrix} \begin{pmatrix} j & j_3 & j_4 \\ -m & m_3 & m_4 \end{pmatrix}$$ First, we construct an orthonormal basis of $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2} \otimes \mathcal{Q}_{j_3} \otimes \mathcal{Q}_{j_4}$ given by the states $$\begin{gathered} \ket{((j_1j_2)j_3)j_4;jklm} = \sum_{m_1,m_2,m_3,m,n,m_4} C^{jm}_{j_1m_1j_2m_2} C^{kn}_{jmj_3m_3} C^{lm}_{knj_4m_4} \bigotimes_{i=1}^4 \ket{j_i,m_i}, \\ \hfill {\rm with} \ j \in \{|j_1-j_2|,...,j_1+j_2 \} \quad {\rm and} \ k \in \{|j-j_3|,...,j + j_3 \} \\ {\rm and} \ l \in \{|k-j_4|,...,k + j_4 \} \quad {\rm and} \ n\in \{-l,...,l\}.\end{gathered}$$ ${\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^4 \mathcal{Q}_{j_i} \right)$ is spanned by the vectors with $l=0$. Similarly to the case $n=3$, we compute $$\begin{aligned} &\ket{((j_1j_2)j_3)j_4;jj_400} &= (-1)^{j_4 + j_1 - j_2 - j_3} \sum_{m_1,m_2,m_3,m_4} \sqrt{2j + 1} \begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j)} \bigotimes_{i=1}^4 \ket{j_i,m_i}.\end{aligned}$$ This basis has the interesting property that it diagonalises $(\vec{J}_1 + \vec{J}_2)^2$: $$(\vec{J}_1 + \vec{J}_2)^2 \ket{j}_{12} = j(j+1) \ket{j}_{12}.$$ The $4jm$-symbol also satisfy orthogonality relations: $$\sum_{m_1,m_2,m_3} {\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j_{12})}} {\begin{pmatrix} j_1 & j_2 & j_3 & l_4 \\ m_1 & m_2 & m_3 & n_4 \end{pmatrix}^{(l_{12})}} = \frac{\delta_{j_{12}l_{12}}}{d_{j_{12}}} \frac{\delta_{j_{4}l_{4}} \delta_{m_4n_4}}{d_{j_{4}}}.\label{orto4j}$$ Finally we can show, similarly to \[projector 3D\], that $$\begin{gathered} \label{4 projected D} \int_{SU(2)} D^{j_1}_{m_1n_1} (u) D^{j_2}_{m_2n_2}(u) D^{j_3}_{m_3n_3}(u) D^{j_4}_{m_4n_4} (u) du \\ = \sum_{j} (2j+1) \begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j)} \begin{pmatrix} j_1 & j_2 & j_3 & j_4\\ n_1 & n_2 & n_3 & n_4 \end{pmatrix}^{(j)}.\end{gathered}$$ Wigner’s 6j-symbol ------------------ In the previous section, we have exhibited an orthonormal basis for ${\rm Inv}_{SU(2)}\left(\bigotimes_{i=1}^4 \mathcal{Q}_{j_i} \right)$. It is built from one possible decomposition of $\bigotimes_{i=1}^4 \mathcal{Q}_{j_i}$ into irreps. Another possible decomposition leads to another basis $$\ket{j}_{23} = \sum_{m_1,m_2,m_3,m_4} \sqrt{2j +1} \begin{pmatrix} j_4 & j_1 & j_2 & j_3 \\ m_4 & m_1 & m_2 & m_3 \end{pmatrix}^{(j)} \bigotimes_{i=1}^4 \ket{j_i,m_i}$$ From $$\label{3CG translation2} \ket{(j_1(j_2j_3))j_4;jklm} = \sum_{m_1,m_2,m_3,m,n,m_4} C^{jm}_{j_2m_2j_3m_3} C^{kn}_{j_1m_1jm} C^{lm}_{knj_4m_4} \bigotimes_{i=1}^4 \ket{j_i,m_i},$$ we show that $$\begin{aligned} &\ket{(j_1(j_2j_3))j_4;jj_400} &= (-1)^{j_1+j_2-j_3+j_4} \sum_{m_1,m_2,m_3,m_4} \sqrt{2j_{23} +1} \begin{pmatrix} j_4 & j_1 & j_2 & j_3 \\ m_4 & m_1 & m_2 & m_3 \end{pmatrix}^{(j_{23})} \bigotimes_{i=1}^4 \ket{j_i,m_i} .\end{aligned}$$ The change of basis is given by $${}_{12}\braket{j}{k}_{23} = \sqrt{2j+1} \sqrt{2k +1} (-1)^{j_1 + j_2 + j_3 - j_4 - 2j - 2k} \begin{Bmatrix} j_1 & j_2 & j \\ j_3 & j_4 & k \end{Bmatrix}$$ where we have defined a new symbol: $$\begin{gathered} \begin{Bmatrix} j_1 & j_2 & j_3\\ j_4 & j_5 & j_6 \end{Bmatrix} \overset{\rm def}= \sum_{m_1, \dots, m_6} (-1)^{\sum_{i=1}^6 (j_i -m_i)} \begin{pmatrix} j_1 & j_2 & j_3\\ -m_1 & -m_2 & -m_3 \end{pmatrix} \begin{pmatrix} j_1 & j_5 & j_6\\ m_1 & -m_5 & m_6 \end{pmatrix} \\ \times \begin{pmatrix} j_4 & j_2 & j_6\\ m_4 & m_2 & -m_6 \end{pmatrix} \begin{pmatrix} j_3 & j_4 & j_5\\ m_3 & -m_4 & m_5 \end{pmatrix}\end{gathered}$$ In the Wolfram Language, it is returned by the function `SixJSymbol`$[\{j_1, j_2, j_3\}, \{j_4, j_5, j_6\}]$. These symbols satisfy the symmetries $$\begin{Bmatrix} j_1 & j_2 & j_3\\ j_4 & j_5 & j_6 \end{Bmatrix} = \begin{Bmatrix} j_2 & j_1 & j_3\\ j_5 & j_4 & j_6 \end{Bmatrix} = \begin{Bmatrix} j_3 & j_2 & j_1\\ j_6 & j_5 & j_4 \end{Bmatrix} = \begin{Bmatrix} j_4 & j_2 & j_3\\ j_1 & j_5 & j_6 \end{Bmatrix}.$$ Similarly one can define the symbols $9j$ and $15j$. The $\{6j\}$-symbol appeared in quantum gravity when Ponzano and Regge realised that the $\{6j\}$-symbol approximate the action of general relativity in the semi-classical limit [@Ponzano:1968uq]. More precisely they have shown that $$\label{PonzanoRegge} \begin{Bmatrix} j_1 & j_2 & j_3\\ j_4 & j_5 & j_6 \end{Bmatrix} \underset{j_i \to \infty}\sim \frac{1}{\sqrt{12 \pi V}} \cos ( S(j_i) + \pi/4) \,,$$ where $V$ is the volume of a tetrahedron whose edges have a length of $j_i + 1/2$, and $S(j_i)$ is the so-called *Regge action*, which is a discrete $3$-dimensional version of the Einstein-Hilbert action. This result was a important source of inspiration for later development of spin-foams. Graphical calculus {#graphical calculus} ------------------ The recoupling theory of $SU(2)$ can be nicely implemented graphically. The underlying philosophy of it is to take advantage of the $2$-dimensional surfaces offered by our sheets of paper and our black-boards to literally *draw our calculations*, rather than restricting oneself to the usual one dimensional lines of calculations. If done properly, the method can help to understand the structure of analytical expressions, and make computations faster. Of course, the first principle of graphical calculus is that there should be a one-to-one correspondence between analytical expressions and diagrams. There exists many conventions for this correspondence in the literature, so we have chosen one that seems to be quite popular [@Sarno2018], and which is described in details by Varshalovich ([@Varshalovich1987], Chap. 11). ##### Definitions. The basic object of this graphical calculus is the $3$-valent node, that represent the Wigner’s $3jm$ symbol: $$\label{graphical 3jm} {\begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix}} = \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \put (45,5) {$-$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_3$} \put (90,55) {$j_2$} \put (45,5) {$+$} \end{overpic} \end{array}.$$ Remark that 1. The signs $+/-$ on the nodes indicate the sense of rotation (anticlockwise/clockwise) in which the spins must be read. To alleviate notations we can decide not to write them by choosing conventionally that the default sign of the nodes is minus, if not otherwise specified. The arrows on the wires will be used below to define the operation of summation. 2. Everywhere we implicitly assume that the Clebsch-Gordan inequalities are satisfied. 3. The magnetic indices are implicit in the diagram, which creates no ambiguity, as long as we associate the $m_i$ to the spin $j_i$. 4. The symmetry properties \[sym 3jm def\] are naturally implemented on the diagram, which also guarantees the one-to-one correspondence between the analytical expression and the diagram. 5. Only the topology of the diagram matters, which means that all topological deformations are allowed. $$\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-star.png} \put (5,25) {$j_1$} \put (50,70) {$j_2$} \put (80,40) {$j_3$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-octopus.png} \put (5,60) {$j_1$} \put (50,75) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array}$$ This principle of topological equivalence is a strong principle of graphical calculus, that will hold for any other diagram constructed later. Then we can define graphically the two basic operations of algebra: multiplication and summation. Multiplication is implemented simply by juxtaposition of two diagrams: $$\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array} \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_4$} \put (50,65) {$j_5$} \put (90,55) {$j_6$} \end{overpic} \end{array} = {\begin{pmatrix} j_1 & j_2 & j_3 \\ m_1 & m_2 & m_3 \end{pmatrix}} {\begin{pmatrix} j_4 & j_5 & j_6 \\ m_4 & m_5 & m_6 \end{pmatrix}}$$ To define the summation, we shall first tell more about the orientation of external wires. As you may have noticed, the arrows on the wires are all outgoing. Now we define also the ingoing orientation with the general rule that inverting the orientation[^8] of an external line $(jm)$ amounts analytically to transforming $m$ to $-m$ and multiplying the overall expression by a factor $(-1)^{j-m}$. For instance $$\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-outin.png} \put (5,45) {$j_1$} \put (55,65) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array}=(-1)^{j_1-m_1}{\begin{pmatrix} j_1 & j_2 & j_3 \\ -m_1 & m_2 & m_3 \end{pmatrix}}\,.$$ The summation over a magnetic index $m$ (from $-j$ to $j$) is now represented by gluing two external wires with the same spin $j$ and magnetic index $m$, but of opposite directions, like: $$\label{def summation} \begin{array}{c} \begin{overpic}[scale = 0.6]{4CG.png} \put (-5,55) {$j_1$} \put (25,55) {$j_2$} \put (70,55) {$j_3$} \put (95,55) {$j_4$} \put (60,5) {$j$} \end{overpic} \end{array} = \sum_{m=-j}^j \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (55,65) {$j_2$} \put (90,55) {$j$} \end{overpic} \end{array} \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-outin.png} \put (5,45) {$j$} \put (55,65) {$j_3$} \put (90,55) {$j_4$} \end{overpic} \end{array}$$ On the RHS, we recognise the definition of the $4jm$-symbol, so that $$\begin{array}{c} \begin{overpic}[scale = 0.6]{4CG.png} \put (-5,55) {$j_1$} \put (25,55) {$j_2$} \put (70,55) {$j_3$} \put (95,55) {$j_4$} \put (60,5) {$j$} \end{overpic} \end{array} = {\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(j)}} \label{CG4}$$ The line between two nodes, whose magnetic index is summed over, is called an internal line, in opposition to external lines, which have a free hand. Contrary to the previous rule of inversion for external lines, it is easy to show that changing the orientation of an internal line gives a phase: $$\begin{array}{c} \begin{overpic}[scale = 0.6]{4CG.png} \put (-5,55) {$j_1$} \put (25,55) {$j_2$} \put (70,55) {$j_3$} \put (95,55) {$j_4$} \put (60,5) {$j$} \end{overpic} \end{array} =(-1)^{2j} \begin{array}{c} \begin{overpic}[scale = 0.6]{4CG-right.png} \put (0,55) {$j_1$} \put (25,55) {$j_2$} \put (60,55) {$j_3$} \put (90,55) {$j_4$} \put (60,5) {$j$} \end{overpic} \end{array}. \nonumber$$ A powerful aspect of graphical calculus comes from the representation of the Kronecker delta with a single line $$\begin{array}{c} \begin{overpic}[scale = 0.6]{up.png} \put (-30,85) {$(j_1,m_1)$} \put (-25,15) {$(j_2,m_2)$} \end{overpic} \end{array}= \delta_{j_1j_2} \delta_{m_1m_2} \qquad \text{or} \qquad \begin{array}{c} \begin{overpic}[scale = 0.6]{up.png} \put (10,60) {$j$} \put (15,95) {$m$} \put (20,5) {$n$} \end{overpic} \end{array}= \delta_{mn} \,.$$ The rule of summation applied to it enables to compute its trace: $$\begin{array}{c} \begin{overpic}[scale = 0.6]{circle.png} \put (5,60) {$j$} \end{overpic} \end{array} = 2j+1$$ For instance, the orthogonality relation \[ortho 3jm sum mm\] now reads $$\begin{array}{c} \begin{overpic}[scale = 0.6]{loop.png} \put (30,90) {$(j_1,m_1)$} \put (30,5) {$(j_2,m_2)$} \put (5,60) {$j_3$} \put (40,60) {$j_4$} \end{overpic} \end{array} \begin{array}{c} \begin{overpic}[scale = 0.6]{circle.png} \put (5,70) {$j_1$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.6]{down.png} \put (25,90) {$(j_1,m_1)$} \put (25,5) {$(j_2,m_2)$} \end{overpic} \end{array} \label{Rule}$$ ##### Lemmas. From all the rules described above, the following lemma can already be checked as an exercise. 1. Reversing all external lines has no effect: $$\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array} =\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-in.png} \put (5,45) {$j_1$} \put (55,65) {$j_2$} \put (90,55) {$j_3$} \end{overpic} \end{array} \,.\label{keep}$$ 2. Changing the sign of the node gives a phase $$\begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \put (45,5) {$-$} \end{overpic} \end{array} = (-1)^{j_1+j_2+j_3} \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (5,45) {$j_1$} \put (50,65) {$j_2$} \put (90,55) {$j_3$} \put (45,5) {$+$} \end{overpic} \end{array} \,.$$ 3. The evaluation of the so-called $\Theta$-graph: $$\begin{array}{c} \begin{overpic}[scale = 0.6]{theta.png} \put (20,70) {$j_1$} \put (55,50) {$j_2$} \put (75,10) {$j_3$} \end{overpic} \end{array}=1.$$ 4. Similarly, \[orto4j\] implies $$\begin{array}{c} \begin{overpic}[scale = 0.6]{doubletheta.png} \put (60,90) {$j_1$} \put (30,55) {$j_2$} \put (55,40) {$j_3$} \put (70,10) {$j_4$} \put (5,40) {$i$} \put (85,40) {$k$} \end{overpic} \end{array} =\frac{\delta_{i,k}}{d_i}. \label{thetaeq}$$ ##### Invariant functions. One nice thing about this graphical calculus is that it makes easy to represent and to remember the Wigner $6j$-symbols: $$\begin{Bmatrix} j_1 & j_2 & j_3\\ j_4 & j_5 & j_6 \end{Bmatrix} = \begin{array}{c} \begin{overpic}[scale = 0.6]{tetrahedron.png} \put (25,63) {$j_1$} \put (47,60) {$j_2$} \put (75,60) {$j_3$} \put (40,35) {$j_4$} \put (25,20) {$j_5$} \put (75,20) {$j_6$} \end{overpic} \end{array}$$ As we can see, the $6j$-symbol looks like a tetrahedron. All magnetic indices are summed over, so that it is only a function of the spins $j_i$, what we can call an *invariant function*. It gives us the idea to define other invariant functions by finding other diagrams with no external links. For instance we define the $9j$-symbol: $$\label{9j} \begin{Bmatrix} j_1 & j_2 & j_3\\ j_4 & j_5 & j_6\\ j_7 & j_8 & j_9 \end{Bmatrix} = \begin{array}{c} \begin{overpic}[scale = 0.6]{9CG.png} \put (25,80) {$j_1$} \put (47,65) {$j_2$} \put (73,73) {$j_3$} \put (60,40) {$j_4$} \put (80,40) {$j_5$} \put (65,15) {$j_6$} \put (25,15) {$j_7$} \put (25,35) {$j_8$} \put (5,45) {$j_9$} \end{overpic} \end{array}$$ Notice that we could have also defined the $9j$-symbol to be rather $$\begin{array}{c} \begin{overpic}[scale = 0.6]{9j-alter.png} \put (25,90) {$j_1$} \put (45,75) {$j_2$} \put (63,83) {$j_3$} \put (50,55) {$j_4$} \put (75,40) {$j_5$} \put (55,5) {$j_6$} \put (10,15) {$j_7$} \put (25,35) {$j_8$} \put (0,45) {$j_9$} \end{overpic} \end{array}$$ but this one can be actually rewritten as the product of two $6j$-symbol. Such a decomposition cannot be done with the $9j$-symbol \[9j\], so that it is said “irreducible”[^9]. We also have the $15j$-symbol: $$\begin{Bmatrix} j_1 & j_2 & j_{11} \\ j_4 & j_5 & j_{15} \\ j_7 & j_3 & j_{14} \\ j_9 & j_6 & j_{13} \\ j_8 & j_{10} & j_{12} \end{Bmatrix} = \begin{array}{c} \begin{overpic}[scale = 0.6]{15j.png} \put (25,75) {$j_1$} \put (33,47) {$j_2$} \put (60,50) {$j_3$} \put (80,70) {$j_4$} \put (50,55) {$j_5$} \put (55,35) {$j_6$} \put (95,35) {$j_7$} \put (37,37) {$j_8$} \put (55,5) {$j_9$} \put (0,35) {$j_{10}$} \put (45,81) {$j_{11}$} \put (0,60) {$j_{12}$} \put (15,5) {$j_{13}$} \put (80,10) {$j_{14}$} \put (93,60) {$j_{15}$} \end{overpic} \end{array}$$ which is the definition used by [@Sarno2018]. It is different from the convention chosen in [@Ooguri:1992eb], which is $$\begin{Bmatrix} l_1 & l_2 & l_3 & l_4 & l_5 \\ j_1 & j_2 & j_3 & j_4 & j_5 \\ l_{10} & l_9 & l_8 & l_7 & l_6 \end{Bmatrix} = \begin{array}{c} \begin{overpic}[scale = 0.6]{15j-ooguri.png} \put (25,75) {$l_2$} \put (35,60) {$j_4$} \put (50,45) {$j_3$} \put (80,70) {$l_{10}$} \put (60,57) {$j_1$} \put (60,30) {$j_5$} \put (95,35) {$l_8$} \put (37,30) {$j_2$} \put (55,5) {$l_6$} \put (0,35) {$l_4$} \put (45,81) {$l_1$} \put (0,60) {$l_3$} \put (15,5) {$l_5$} \put (80,10) {$l_7$} \put (93,60) {$l_9$} \end{overpic} \end{array}$$ Contrary to the $6j$-symbol, there is no consensus about what is called the $15 j$-symbol, but in all cases it corresponds to an invariant function associated to a $3$-valent graph with 15 links. Actually, we can build $5$ topologically different $15j$-symbols[^10]. Here we see the power of graphical calculus: imagine if we had given the analytical formula for it... that is doable, but unreadable. In the spirit of the result of Ponzano and Regge \[PonzanoRegge\], Ooguri used the $15j$-symbol to provide a model of quantum gravity [@Ooguri:1992eb]. It still plays a major role in the EPRL model [@Speziale2017]. Representation theory of SL(2,C) ================================ To put it in a nutshell, the kinematic aspects of loop quantum gravity deal with representations of $SU(2)$, while the dynamics, in its spin-foam formulation, lie in the representation theory of $SL_2(\mathbb{C})$. The current models of spin-foams, like the EPRL one, use extensively the principal series of $SL_2(\mathbb{C})$. Fortunately all the irreps of $SL_2(\mathbb{C})$ are known. Whether or not all representations, including non-reducible ones, have been classified is unknown to the author. In section \[finite irreps\], we present the finite-dimensional ones. In section \[infinite irreps\], we summarize the infinite-dimensional ones. Finally, in section \[principal series\], we focus on the principal series, which is the main interest for quantum gravity. Finite irreps {#finite irreps} ------------- The finite irreps of $SL_2(\mathbb{C})$ are well-known. They can be obtained from the irreps of $\mathfrak{sl}_2(\mathbb{C})$. In section \[abstract representation\], we have already seen the finite irreps of its $3$-dimensional (complex) Lie algebra $\mathfrak{sl}_2(\mathbb{C})$: they are indexed by a spin $j \in \mathbb{N}/2$. It is also possible to see $\mathfrak{sl}_2(\mathbb{C})$ as a real Lie algebra of dimension $6$, in which case, we will rather denote it $\mathfrak{sl}_2(\mathbb{C})_\mathbb{R}$. In this section, we will describe the (real) linear representations of $\mathfrak{sl}_2(\mathbb{C})_\mathbb{R}$. We have the following isomorphism between real vector spaces: $$\label{iso sl su} \mathfrak{sl}_2(\mathbb{C})_{\mathbb{R}} \cong \mathfrak{su}(2) \oplus \mathfrak{su}(2) .$$ The algebra $\mathfrak{su}(2) \oplus \mathfrak{su}(2)$ is the Lie algebra of $SU(2) \times SU(2)$. A consequence of Peter-Weyl’s theorem is that the irreps of a cartesian product are tensor products of the irreps of the factors ([@Knapp1986] p. 32). Thus the irreps of $\mathfrak{sl}_2(\mathbb{C})_{\mathbb{R}} $ are given by the usual tensor representation over $\mathcal{Q}_{j_1} \otimes \mathcal{Q}_{j_2}$, abbreviated by $(j_1,j_2)$. The action is given by: $$a \cdot (\ket{j_1,m_2} \otimes \ket{j_2,m_2}) \overset{\rm def}= (a \ket{j_1,m_1}) \otimes \ket{j_2,m_2} + \ket{j_1,m_1} \otimes ( a \ket{j_2,m_2}).$$ The isomorphism \[iso sl su\] provides naturally a basis of $\mathfrak{sl}_2(\mathbb{C})_\mathbb{R}$, given by the three Pauli matrices $\sigma_i \in i \mathfrak{su}(2)$ and the three matrices $i \sigma_i \in \mathfrak{su}(2)$. To match the earlier notations introduced in section \[angular momentum realisation\], we often denote the *rotation generators* $J_i \overset{\rm def} = \frac{1}{2} \sigma_i$ and the *boost generators* $K_i \overset{\rm def}= \frac{i}{2} \sigma_i$. These generators satisfy the commutation relations: $$\begin{aligned} \left[ J_i,J_j \right]= i \varepsilon_{ijk} J_k && \left[ J_i,K_j \right]= i \varepsilon_{ijk} K_k && \left[ K_i,K_j \right]= - i \varepsilon_{ijk} J_k.\end{aligned}$$ Another basis is given by the *complexified generators*. Posing $A_i = \frac 12 (J_i + i K_i)$ and $B_i = \frac 12 (J_i-i K_i)$, the commutation relations become: $$\begin{aligned} \left[A_i ,A_j\right] = i\varepsilon_{ijk}A_k && \left[B_i ,B_j\right] = i\varepsilon_{ijk}B_k && \left[A_i ,B_j\right] = 0.\end{aligned}$$ We can also define the scale operators. Posing $K_\pm \overset{\rm def}=K_1 \pm i K_2 $ and $J_\pm \overset{\rm def}= J_1 \pm i J_2$, the scale operators satisfy: $$\begin{aligned} &\left[J_3,J_\pm\right] = \pm J_\pm && \left[J_+,J_-\right] = 2 J_3 \\ &\left[K_3,K_\pm\right]=\mp J_\pm && \left[K_+,K_-\right]=-2J_3 \\ &\left[J_+,K_+\right]=\left[J_-,K_-\right]=\left[J_3,K_3\right]=0 \\ &\left[K_3,J_\pm\right]=\pm K_\pm && \left[J_\pm,K_\mp\right]= \pm 2 K_3 && \left[J_3,K_\pm\right]=\pm K_\pm\end{aligned}$$ Then the various realisations, which where described in chapter \[rep theory SU(2)\] for the action of $\mathfrak{su}(2)$, can be adapted to $\mathfrak{sl}_2(\mathbb{C})_\mathbb{R}$. ##### Homogeneous realisation. Let $m,n \geq 2$, and $\mathbb{C}_{(m,n)}[z_0,z_1;\overline{z_0},\overline{z_1}]$ the vector space of homogeneous polynomials of degree $m$ in $(z_0,z_1)$ and homogeneous of degree $n$ in $(\overline{z_0},\overline{z_1})$. The action of $SL_2(\mathbb{C})$ is given by $$\begin{aligned} g \cdot P(z_0,z_1) = P \left(g^T \binom{z_0}{z_1} \right).\end{aligned}$$ The associated action of the algebra $\mathfrak{sl}_2(\mathbb{C})_\mathbb{R}$ is given by $$\begin{aligned} J_+ \cong z_0 \frac{\partial}{\partial z_1} + \overline{z_0} \frac{\partial}{\partial \overline{z_1}} && J_- \cong z_1 \frac{\partial}{\partial z_0} + \overline{z_1} \frac{\partial}{\partial \overline{z_0}} && J_3 \cong \frac{1}{2} \left( z_0 \frac{\partial}{\partial z_0} - z_1 \frac{\partial}{\partial z_1} + \overline{z_0} \frac{\partial}{\partial \overline{z_0}} - \overline{z_1} \frac{\partial}{\partial \overline{z_1}} \right).\end{aligned}$$ ##### Projective realisation. As seen in section \[projective\], the restriction of $P(z_0,z_1)$ to $P(z,1)$ induces another realisation over the space of polynomials of degree at most $m$ in $z$ and at most $n$ in $\overline{z}$. The action is given by $$\begin{aligned} g \cdot \phi(\xi) = (g_{12} \xi + g_{22})^m \overline{(g_{12} \xi + g_{22})}^n \phi \left( \frac{g_{11} \xi + g_{21}}{g_{12}\xi+g_{22}} \right).\end{aligned}$$ ##### Spinorial realisation. Finally the $(m,n)$ representation of $SL_2(\mathbb{C})$ can be realised over the space of totally symmetric spinors $\mathfrak{S}^{(A_1...A_m)(\dot A_1... \dot A_n)}$ such as (see Penrose [@Penrose1984] p. 142) $$u \cdot z^{A_1...A_m \dot A_1... \dot A_n} = u^{A_1}_{\ \ B_1} ... u^{A_m}_{\ \ B_m} \, \overline{u}^{\dot A_1}_{\ \ \dot B_1} ... \overline{u}^{\dot A_n}_{\ \ \dot B_n} \, z^{B_1...B_m \dot B_1... \dot B_n}.$$ Finite representations of $SL_2(\mathbb{C})$ cannot be unitary (except the trivial one), because it is a simply connected non-compact Lie group. If we want unitary representations, we shall turn to infinite ones. Infinite irreps {#infinite irreps} --------------- In this section, we describe all the infinite-dimensional irreps of $SL_2(\mathbb{C})$. All the unitary irreps of the Lorentz group have been found simultaneously in 1946 by Gel’fand and Naimark [@Gelfand1946], by Harish-Chandra [@Harish-Chandra1947] and by Bargmann [@Bargmann2010]. It seems nevertheless that Gel’fand and Naimark were the first to publish (unfortunately their article is only in Russian). The question remained to find all the irreps, unitary or not, and this was solved also by Naimark in [@Naimark1954]. In 1963, Gel’fand, Minlos and Shapiro published the first book (with english translation) that reviews all these results [@Gelfand:1963fk]. In 1964, Naimark wrote a more detailed and well-written book that wraps up the subject for mathematically-orientated physicists [@naimark]. The infinite irreps of $SL_2(\mathbb{C})$ are parametrised by $(m,\rho) \in \mathbb{Z} \times \mathbb{C}$ with $\Im \rho \geq 0$ and $\rho^2 \neq -(|m| +2n)^2 \in \mathbb{N}^*$. A realisation is given over the Hilbert space $L^2(\mathbb{C})$ endowed with the scalar product $$\braket{\varphi }{ \phi} = \frac{i}{2} \int_{\mathbb{C}} \overline{\varphi(\xi)} \phi(\xi) (1+|\xi|^2)^{- \Im \rho} d\xi d\overline{\xi},$$ and the action $$\label{action Naimark} a \cdot f(z) = (a_{12}z+a_{22})^{\frac {m}{2} + \frac{i\rho}{2} -1} \overline{(a_{12} z + a_{22})}^{-\frac {m}{2} + \frac{i \rho}{2}-1} f\left(\frac{a_{11}z+a_{21}}{a_{12}z+a_{22}}\right).$$ **Remarks:** 1. If $\rho^2 = -(|m| +2n)^2$ with $n \in \mathbb{N}^*$, the Hilbert space and the action still defines a representation, but a reducible one. Then, if one restricts the action to the subspace of polynomials of degree at most $p= \frac m2 + i \frac \rho 2-1$ in $z$ and $q= -\frac m2 + i \frac \rho 2 -1$ in $\overline{z}$, the representation is irreducible and equivalent to the finite-dimensional representation $(p,q)$. 2. Not all the representations $(m,\rho)$ are unitary. They are unitary in only two cases: when $\rho \in \mathbb{R}$ (principal series); when $m = 0$ and $i \rho \in ]-2,0[$ (complementary series), provided another scalar product chosen in the latter case (see below). 3. Among these infinite irreps, only the principal representations $(\rho,k)$ and $(-\rho,-k)$ are equivalent. 4. A proof of the result above can be found in Naimark ([@naimark] pp. 294-295). A sketch of it in the case of the principal series can be found in appendix \[induced\]. **Principal series.** When $\rho \in \mathbb{R}$, the scalar product over $L^2(\mathbb{C})$ becomes the more usual $$(f_1,f_2) \overset{\rm def}= \frac i 2\int_\mathbb{C} \overline{f_1(z)} f_2(z) dz d\overline{z},$$ and the representation $(\rho,m) \in \mathbb{R} \times \mathbb{Z}$ is unitary. It is called the principal series. Moreover the representations $(\rho,m)$ and $(-\rho,- m)$ are unitarily equivalent. We have used below the convention of Naimark for indexing the representations of the principal unitary series, $(\rho,m) \in \mathbb{R} \times \mathbb{Z}$ (see [@naimark] p. 150). Vilenkin uses the same convention ([@Gelfand:1966uq] p. 191). Rühl uses rather the convention $(\rho_R,m_R) = (\rho, - m)$ ([@ruhl] p. 54). Gel’fand uses the convention $(\rho_G, m_G) = (\rho/2,m)$ ([@Gelfand:1963fk] p. 247). In LQG, it is common to use the conventions $(p,k)=(\rho/2, m/2)$, as Rovelli does ([@CLQG] p. 182) or Barrett [@Barrett2010]. Usually in LQG, the parameter $k$ is even restricted to be non-negative, which actually does not suppress any representation since $(p,k)$ and $(-p,-k)$ are equivalent. **Complementary series**. When $m=0$ and $i \rho \in ]-2,0[$, the action becomes $$a \cdot \phi(\xi) = |a_{12} \xi + a_{22}|^{i\rho-2} \phi\left( \frac{a_{11} \xi + a_{21}}{a_{12} \xi + a_{22}} \right).$$ It also defines a unitary representation for the scalar product $$\braket{\varphi }{\phi} = \left( \frac{i}{2} \right)^2 \int_{\mathbb{C}^2} \frac{\overline{\varphi(\xi)} \phi(\eta)}{|\xi - \eta|^{2+ i \rho}} d\xi d\overline{\xi} d\eta d\overline{\eta}.$$ Principal unitary series {#principal series} ------------------------ In this section, we will present three realisations of the unitary principal series. The construction of the principal series by Gel’fand and Naimark is based on the induced representations method (see appendix \[induced\] for a sketch of the proof). Though rigorous from the mathematical point of view, it is not very intuitive, especially for physicists. In 1962, Gel’fand, Graev and Vilenkin published a book (referred to as “the Vilenkin” hereafter) where they build the principal series from a space of homogeneous functions, which may seem more natural ([@Gelfand:1966uq] pp. 139-201). A beautiful and concise exposure can be found in the article of Dao and Nguyen ([@Dao1967] pp. 18-21). We present the method in the following subsection. ### Homogeneous realisation {#principal homogeneous} Consider $\mathcal{F}(\mathbb{C}^2)$, the vector space of the complex functions over $\mathbb{C}^2$. A function $F \in \mathcal{F}(\mathbb{C}^2)$ is said to be *homogeneous of degree $(\lambda,\mu) \in \mathbb{C}^2$* if it satisfies for all $\alpha \in \mathbb{C}$ : $$F(\alpha \pmb z) = \alpha^\lambda \overline{\alpha}^\mu F(\pmb z).$$ To be consistently defined (when $\alpha = e^{2i\pi n}$), the degree should satisfy the condition : $$\mu - \lambda \in \mathbb{Z}.$$ Instead of $(\lambda, \mu)$, we will instead use in the following, the parameters $(p = \frac{\mu + \lambda + 2}{2i}, k =\frac{\lambda - \mu }{2})$ (same choice of parameters as Rovelli [@CLQG] p. 182). Define $\mathcal{D}^{(p,k)}[z_0,z_1]$ as the subspace of homogeneous functions of degree $(\lambda,\mu)$ infinitely differentiable over $\mathbb{C}^2 \setminus \{ 0 \}$ in the variables $z_0,z_1,\bar{z}_0$ et $\bar{z}_1$ with a certain topology[^11]. We define a continuous representation $SL_2(\mathbb{C})$ over $\mathcal{D}^{(p,k)}[z_0,z_1]$ by $$a \cdot F( \pmb z) \overset{\rm def}= F(a^T \pmb z).$$ We could also have defined the action by $F(a^{-1} \pmb z)$, $F(a^\dagger \pmb z)$ or $F(\pmb z a)$. In fact $F(\pmb z a) = F(a^T \pmb z)$, defines the same action. The action with $a^\dagger$ is obtained by the transformation $a_{ij} \mapsto \overline{a_{ij}}$. The action with $a^{-1}$ is obtained by the transformation $a_{ij} \mapsto (2 \delta_{ij} -1 ) \sum_{kl} (1 - \delta_{ik})(1- \delta_{jl}) a_{kl}$. The convention that we have chosen here is the one of Vilenkin ([@Gelfand:1966uq] p. 145), Dao and Nguyen ([@Dao1967] p. 18), Rühl ([@ruhl] p. 53), Rovelli ([@CLQG] p. 182) and Barrett [@Barrett2010]. Knapp ([@Knapp1986] p. 28) is using the convention $F(a^{-1} \pmb z)$. Now define the following $2$-form over $\mathbb{C}^2$: $$\begin{aligned} \Omega (z_0,z_1) = \frac i2 (z_0dz_1 - z_1 dz_0) \wedge (\overline{z_0}d\overline{z_1} - \overline{z_1} d\overline{z_0}).\end{aligned}$$ Interestingly, it is invariant for the action of $SL_2(\mathbb{C})$: $\Omega(a \pmb z)=\Omega (\pmb z)$. Let $\Gamma$ be a path in $\mathbb{C}^2$ that intersects each projective line exactly once. Then define the scalar product over $\mathcal{D}^{(p,k)}[z_0,z_1]$: $$\begin{aligned} (F,G) = \int_\Gamma \overline{F(\pmb z)} G(\pmb z) \Omega(\pmb z).\end{aligned}$$ Thus $\mathcal{D}^{(p,k)}[z_0,z_1]$ is an Hilbert space. Interestingly, the result does not depend on the path $\Gamma$ provided $p \in \mathbb{R}$, which we consider to be the case in the following. This scalar product is invariant for $SL_2(\mathbb{C})$: $(a \cdot F, a \cdot G) = (F,G)$. Thus the representation is unitary. It could be also shown to be irreducible. In the next subsection, we will see that the representation $\mathcal{D}^{(p,k)}[z_0,z_1]$ is equivalent to the representation $(\rho = 2 p,m = 2k)$ of the principal series described in section \[infinite irreps\]. ### Projective realisation {#projective-realisation} Consider the map $$\begin{aligned} \iota : \left\{ \begin{array}{lll} \mathbb{C} & \rightarrow &\mathbb{C}^2 \\ \zeta & \mapsto & (\zeta,1) \end{array} \right.\end{aligned}$$ $\iota$ is a diffeomorphism from $\mathbb{C}$ to its range. It parametrises a horizontal straight line of the complex plane. The projective construction consists in restricting the domain of definition of the homogeneous function to this line. If $F \in \mathcal{F}(\mathbb{C}^2)$, define $\iota^*F \in \mathcal{F}(\mathbb{C})$ as $$\iota^* F (z) \overset{\rm def}= F \circ \iota (z) = F(z,1).$$ The $2$-form $\Omega$ becomes similarly $$\begin{aligned} \iota^* \Omega (z) = \frac i 2 dz \wedge d\overline{z},\end{aligned}$$ which is nothing but the usual Lebesgue measure over $\mathbb{C}$. Thus we define the Hilbert space $L^2(\mathbb{C})$ with the the scalar product $$\begin{aligned} (f,g) = \frac i 2 \int_\mathbb{C} \overline{f(z)}g(z) dz \wedge d\overline{z}.\end{aligned}$$ Thus we have a map $\iota^* : \mathcal{D}^{(p,k)}[z_0,z_1] \rightarrow L^2(\mathbb{C})$. In fact $\iota^*$ is bijective: for all $f \in L^2(\mathbb{C})$, there exists a unique $F \in \mathcal{D}^{(p,k)}[z_0,z_1]$ such that $f = \iota^* F$. $F$ is given explicitly by $$F (z_0,z_1) = z_1^{-1+ip+k} \bar{z}_1^{-1+ip-k} f \left( \frac{z_0}{z_1} \right).$$ Importantly, $\iota^*$ induces naturally an action of $SL_2(\mathbb{C})$ over $\mathcal{F}(\mathbb{C})$, such that $\iota^*$ becomes an intertwiner between two equivalent representations. After computation, we obtain: $$a \cdot f(z) \overset{\rm def}= (a_{12}z+a_{22})^{-1+ip +k} \overline{(a_{12}z + a_{22})}^{-1+ip -k} f\left( \frac{a_{11}z+a_{21}}{a_{12}z+a_{22}} \right).$$ This formula is exactly the same formula as \[action Naimark\], with the indices $(p,k) = (\rho/2,m/2)$. Thus we have constructed explicitly the representations of the principal series, and we have shown the equivalence of the realisations $\mathcal{D}^{(p,k)}[z_0,z_1]$ and $L^2(\mathbb{C})$. ### SU(2)-realisation We are going to build another realisation of the unitary principal representations. It is based on a space of “$U(1)$-covariant” functions over $SU(2)$. The general idea lies over the observation that $SU(2)/U(1) \cong \mathbb{C}P^1$. To be precise, consider the map $$\begin{aligned} \kappa : \left\{ \begin{array}{lll} SU(2) & \rightarrow & \mathbb{C}^2 \\ u & \mapsto & (u_{21},u_{22}) \end{array} \right.\end{aligned}$$ $\kappa$ is a diffeomorphism to its range. In some sense $SU(2)$ can be seen as the “unit circle” of $\mathbb{C}^2$, so that $\kappa$ can be seen as the injection of the circle in the plane $\mathbb{C}^2$. Then, define $\kappa^* : \mathcal{F}(\mathbb{C}^2) \rightarrow \mathcal{F}(SU(2))$ such that $$\label{kappa^*} \kappa^* F (u) \overset{\rm def}= F \circ \kappa (u) = F \left( u_{21}, u_{22} \right).$$ If $F \in \mathcal{D}^{(p,k)}[z_0,z_1]$, then we show easily that $\kappa^*F$ satisfies the covariance property $$\label{covariance property} \kappa^* F (e^{i\theta \sigma_3} u) = e^{-2 i \theta k} \kappa^* F (u).$$ We denote $\mathcal{D}^{(p,k)}[u] \overset{\rm def}= \kappa^* \mathcal{D}^{(p,k)}[z_0,z_1]$. Thus $\kappa^*: \mathcal{D}^k[z_0,z_1] \to \mathcal{D}^{(p,k)}[u]$ is a bijection. Its inverse is given explicitly by $$\label{kappa*-1} F (z_0 , z_1) = (|z_0|^2 + |z_1|^2)^{-1+ip} \phi \left( \frac{1}{\sqrt{|z_0|^2+|z_1|^2}} \begin{pmatrix} z_1^* & -z_0^* \\ z_0 & z_1 \end{pmatrix} \right).$$ We have to check that \[kappa\*-1\] is the inverse expression of \[kappa\^\*\]. Then we check the homogeneity property $(\kappa^*)^{-1}\phi(\alpha (z_0,z_1)) = \alpha^\lambda \overline{\alpha}^\mu (\kappa^*)^{-1}\phi(z_0,z_1) $. We could also translate the measure $\kappa^* \Omega$, and thus endow $\mathcal{D}^{(p,k)}[u]$ with the structure of an Hilbert space. Interestingly, it is a subspace of $L^2(SU(2))$. As previously, one can translate the action of $SL_2(\mathbb{C})$ over $\mathcal{D}^{(p,k)}[u]$ such that $\kappa^*$ become a bijective intertwiner, and we obtain $$a \cdot \phi(u) = (|\beta_{a,u}|^2 + |\alpha_{a,u}|^2)^{-1+ip} \phi \left( \frac{1}{\sqrt{|\beta_{a,u}|^2+|\alpha_{a,u}|^2}} \begin{pmatrix} \alpha_{a,u} & -\beta_{a,u}^* \\ \beta_{a,u} & \alpha_{a,u}^* \end{pmatrix} \right),$$ with $\alpha_{a,u} \overset{\rm def}= (u_{21} a_{12} + u_{22} a_{22})^*$ and $\beta_{a,u} \overset{\rm def}= u_{21} a_{11} + u_{22} a_{21}$. Thus $\mathcal{D}^{(p,k)}[u]$ is a third equivalent realisation of the unitary principal series. The equivalence with $L^2(\mathbb{C})$ is made through $(\kappa \circ \iota^{-1} )^*$ which gives explicitly $$\phi (u)= u_{22}^{-1+ip+k} \overline{u_{22}}^{-1+ip-k} f\left(\frac{u_{21}}{u_{22}} \right) ,$$ and conversely $$f (z) =(1+|z|^2)^{-1+ip} \phi \left( \frac{1}{\sqrt{1+|z|^2}} \begin{pmatrix} 1 & -z^* \\ z & 1 \end{pmatrix} \right).$$ ### Canonical basis The advantage of the $SU(2)$-realisation is that we already know interesting function over $SU(2)$, namely the coefficients of the Wigner matrix $D^j_{pq}$. Are they elements of $\mathcal{D}^{(p,k)}[u]$? They are indeed, provided they satisfy the covariance property \[covariance property\]. We compute easily $$\begin{aligned} D^j_{pq}\left( e^{i\theta \sigma_3} u \right) = \sum_{k=-j}^j \bra{j,p} e^{i\theta \sigma_3} \ket{j,k} D^j_{kq}\left( u \right) = \sum_{k=-j}^j \delta_{pk} e^{2ik\theta} D^j_{kq}\left( u \right) = e^{2ip\theta} D^j_{pq}\left( u \right).\end{aligned}$$ Thus, the covariance property it is satisfied if $p= -k$, and so $$\label{Dbasis} \forall j \in \{ |k|,|k|+1,... \}, \quad \forall q \in \{-j,..., j \} , \quad D^j_{-k,q} \in \mathcal{D}^{(p,k)}[u].$$ Since the $D^j_{mn}(u)$ form a basis of $L^2(SU(2))$, we show easily that the subset exhibited in \[Dbasis\] form a basis of $\mathcal{D}^{(p,k)}[u]$. Another consequence is the following decomposition of $\mathcal{D}^{(p,k)}[u]$ into irreps of $SU(2)$: $$\mathcal{D}^{(p,k)}[u] \cong \bigoplus_{j = |k|}^\infty \mathcal{Q}_j.$$ We then call *canonical basis* of $\mathcal{D}^{(p,k)}[u]$ the set of functions: $$\label{canonical SU(2)} \phi^{(p,k)}_{jm}(u) \overset{\rm def}= \sqrt{\frac{2j+1}{\pi}} D^j_{-k,m}(u), \quad \text{with} \ j = |k|, |k|+1,... \ \text{and} \ -j \leq m \leq j.$$ From \[ortho D\], we see that they satisfy the orthogonality relations $$\int_{SU(2)} \dd u \ \overline{\phi^{(p,k)}_{jm}(u)} \phi^{(p,k)}_{ln} (u) = \frac 1 \pi \delta_{jl} \delta_{mn}.$$ In Rühl ([@ruhl] p. 59), the factor $\frac{1}{\sqrt{\pi}}$ is absent from the definition of the $\phi^{(p,k)}_{jm}$. Thus, the orthogonality relations do not show a factor $\frac{1}{\pi}$ on the RHS. We have chosen this factor so that the canonical basis $f^{(p,k)}_{jm}$ of $L^2(\mathbb{C})$ (see below \[canonical complex\]) is orthonormal for the usual scalar product with the Lebesgue measure ${\rm d}z$ (for Rühl the measure is ${\rm d}z / \pi$). Moreover $\phi^{(p,k)}_{jm}$ could have been defined with a phase factor $e^{i\psi(p,j)}$. This is is set to zero in some literature including [@ruhl; @Barrett2010], and we follow that convention here. An alternative phase convention leading to real $SL_2(\mathbb{C})$-Clebsch-Gordan coefficients is obtained for the choice [@Kerimov1978; @Speziale2017] $e^{i\psi(p,j)} = (-1)^{-\frac{j}2} \frac{\Gamma(j+i\rho+1)}{|\Gamma(j+i\rho+1)|}$. This phase convention is the one used in the approach of [@Speziale2017] to compute the coefficients and vertex amplitudes and the numerical analysis [@Dona2018b; @Dona2018c; @Fanizza] that implements this approach. From the perspective of the large spin asymptotics this additional phase plays no role and can be taken into account independently, if so desired. An intermediate choice of phase is the one of [@Dao1967; @Rashid2003], which has the advantage of simplifying the recursion relations satisfied by the Clebsch-Gordan coefficients [@Anderson1970; @Anderson1970a]. The latter are now either real or purely imaginary. The intertwiner $\kappa^*$ enables to translate this basis in $\mathcal{D}^{(p,k)}[z_0,z_1]$, and we obtain the canonical basis: $$\begin{aligned} F^{(p,k)}_{jm} (z_0,z_1) &= \sqrt{\frac{2j+1}{\pi}} (|z_0|^2 + |z_1|^2)^{ip - 1 } D^j_{-k,m} \left( \frac{1}{\sqrt{|z_0|^2+|z_1|^2}} \begin{pmatrix} z_1^* & -z_0^* \\ z_0 & z_1 \end{pmatrix} \right),\end{aligned}$$ where an explicit expression for $D^j_{-k,m}$ is given by equation \[Wigner matrix\]. The same is done with the intertwiner $\iota^*$ to $L^2(\mathbb{C})$, and we obtain the canonical basis: $$\label{canonical complex} f^{(p,k)}_{jm} (z) = \sqrt{\frac{2j+1}{\pi}} (1+|z|^2)^{ip -1 -j } D^j_{-k,m} \begin{pmatrix} 1 & -z^* \\ z & 1 \end{pmatrix}$$ From \[Wigner matrix\] we have $$\label{Wig mat 11} D^j_{-k,m}\begin{pmatrix} 1 & -z^* \\ z & 1 \end{pmatrix} = \left( \frac{(j-k)!(j+k)!}{(j+m)!(j-m)!} \right)^{1/2} \sum_{i= \max (0,m-k)}^{ \min(j-k,j+m)} \binom{j+m}{i} \binom{j-m}{j-k-i} z^{j+m-i} (-z^*)^{j-k-i}.$$ The constant factors of \[canonical SU(2)\] have been chosen so that $$\begin{aligned} \frac{i}{2} \int_\mathbb{C} \overline{f^{(p,k)}_{jm}(z)} f^{(p,k)}_{ln} (z) dz d\overline{z} = \delta_{jl} \delta_{mn}.\end{aligned}$$ Finally, in ket notations, the canonical basis is denoted $\ket{p,k,jm}$. ### Action of the generators Similarly to \[generators homogeneous\], the action of the $SL_2(\mathbb{C})$-generators can be computed from the action of the group (see Naimark [@naimark] pp. 104-117). The generators of the rotations “stay inside” the same $SU(2)$-irreps: $$\left\{ \begin{array}{l} J_3 \ket{p,k,j,m} = m \ket{p,k,j,m} \\ J_+ \ket{p,k,j,m} = \sqrt{(j+m+1)(j-m)} \ket{p,k,j,m+1} \\ J_- \ket{p,k,j,m} = \sqrt{(j+m)(j-m+1)} \ket{p,k,j,m-1} \end{array} \right.$$ The generators of the boost spread over the neighbouring subspaces: $$\begin{gathered} K_3 \ket{p,k,j,m} = \alpha_j \sqrt{j^2 - m^2} \ket{p,k,j-1,m} + \gamma_j m \ket{p,k,j,m} \\ - \alpha_{j+1} \sqrt{(j+1)^2 - m^2} \ket{p,k,j+1,m},\end{gathered}$$ $$\begin{gathered} K_+ \ket{p,k,j,m} = \alpha_j \sqrt{(j-m)(j-m-1)} \ket{p,k,j-1,m+1} \\ + \gamma_j \sqrt{(j-m)(j+m+1)} \ket{p,k,j,m+1} \\ + \alpha_{j+1} \sqrt{(j+m+1)(j+m+2)} \ket{p,k,j+1,m+1}\end{gathered}$$ $$\begin{gathered} K_- \ket{p,k,j,m} = - \alpha_j \sqrt{(j+m)(j+m-1)} \ket{p,k,j-1,m-1} \\ + \gamma_j \sqrt{(j+m)(j-m+1)} \ket{p,k,j,m-1} \\ - \alpha_{j+1} \sqrt{(j-m+1)(j-m+2)} \ket{p,k,j+1,m-1}\end{gathered}$$ with $\gamma_j \overset{\rm def}= \frac{k p}{j(j+1)}$ and $\alpha_j \overset{\rm def}= i \sqrt{\frac{(j^2-k^2)(j^2+p^2)}{j^2(4j^2-1)}}$. From these expressions, it is possible to compute the action of the two Casimir operators: $$\begin{aligned} &(\vec{K}^2 - \vec{L}^2) \ket{p,k,j,m} = (p^2- k^2 + 1) \ket{p,k,j,m} \ , \\ &\vec K \cdot \vec L \ket{p,k,j,m} = pk \ket{k,p,j,m} \ . \end{aligned}$$ ### SL(2,C) Wigner’s matrix We define the $SL_2(\mathbb{C})$ Wigner’s matrix by its coefficients $$D^{(p,k)}_{j_1q_1j_2q_2}(a) \overset{\rm def}= \bra{p,k;j_1q_1} a \ket{p,k;j_2q_2}.$$ These coefficients satisfy the orthogonality relations: $$\int_{SL_2(\mathbb{C})} dh D^{(p_1,k_1)}_{j_1m_1l_1n_1}(h) D^{(p_2,k_2)}_{j_2m_2l_2n_2}(h) = \frac{1}{4(p_1^2 + k_1^2)} \delta(p_1-p_2) \delta_{k_1k_2} \delta_{j_1j_2} \delta_{l_1l_2} \delta_{m_1m_2} \delta_{n_1n_2}.$$ ##### Explicit expressions. To compute explicitly, we use a concrete realisation, for instance $$D^{(p,k)}_{jplq} (g) = \frac{i}{2} \int_{\mathbb{C}} \dd \omega \dd \overline{\omega} \, \overline{f^{(p,k)}_{jp} (\omega)} (g_{12} \omega + g_{22})^{-1+ip +k} \overline{(g_{12} \omega + g_{22})}^{-1+ip-k} f^{(p,k)}_{lq} \left( \frac{g_{11} \omega + g_{21}}{g_{12} \omega + g_{22}}\right).$$ Cartan decomposition states that any $g \in SL_2(\mathbb{C})$ can be written (non uniquely) as $g = u e^{r \sigma_3 /2} v^{-1}$, with $u,v \in SU(2)$ and $r \in \mathbb{R}_+$ (see \[decomposing SL\]). Then, we show that $$D^{(p,k)}_{jmln}(g) = \sum_{q=-\min(j,l)}^{\min(j,l)} D^j_{mq}(u) d^{(p,k)}_{jlq}(r) D^l_{qn}(v^{-1}).$$ with the *reduced $SL_2(\mathbb{C})$ Wigner’s matrix* defined as $$d^{(p,k)}_{jlm}(r) \overset{\rm def}= D^{(p,k)}_{jmlm} (e^{r\sigma_3/2}).$$ $$\begin{aligned} D^{(p,k)}_{jmln}(g) &= \sum_{p \geq |k|} \sum_{q=-p}^p \sum_{p'\geq |k|} \sum_{q'=-p'}^{p'} D^{(p,k)}_{jmpq}(u) D^{(p,k)}_{pqp'q'}(e^{r \sigma_3/2}) D^{(p,q)}_{p'q'ln}(v^{-1})\\ &= \sum_{p \geq |k|} \sum_{q=-p}^p \sum_{p'\geq |k|} \sum_{q'=-p'}^{p'} \delta_{jp} D^j_{mq}(u) D^{(p,k)}_{pqp'q'}(e^{r \sigma_3/2}) \delta_{p'l} D^l_{q'n}(v^{-1})\\ &= \sum_{q=-j}^j \sum_{q'=-l}^{l} D^j_{mq}(u) D^{(p,k)}_{jqlq'}(e^{r \sigma_3/2}) D^l_{q'n}(v^{-1}) \\ &= \sum_{q=-j}^j \sum_{q'=-l}^{l} D^j_{mq}(u) \matrixel{p,k,jq }{ e^{r \sigma_3/2}}{ p,k,lq'} D^l_{q'n}(v^{-1}) \\ &= \sum_{q=-\min(j,l)}^{\min(j,l)} D^j_{mq}(u) D^{(p,k)}_{jqlq} (e^{r\sigma_3/2}) D^l_{qn}(v^{-1})\end{aligned}$$ We have the following symmetry properties: $$d^{(p,k)}_{jlm}(r) = d^{(-p,k)}_{ljm}(-r) = d^{(p,-k)}_{jl,-m}(r) = (-1)^{j-l} d^{(-p,-k)}_{ljm}(r)=\overline{d^{(p,k)}_{ljm}(-r) }.$$ $\blacksquare$ **Integral formula 1.** $$\begin{gathered} d^{(p,k)}_{jlm}(r) = \sqrt{(2j+1)(2l+1)} \left( \frac{(j-k)!(j+k)!}{(j+p)!(j-p)!} \right)^{1/2} \left( \frac{(l-k)!(l+k)!}{(l+p)!(l-p)!} \right)^{1/2} \\ \times \sum_{i= \max (0,p-k)}^{ \min(j-k,j+p)} \sum_{i'= \max (0,p-k)}^{ \min(l-k,l+p)} \Bigg[ (-1)^{j+l-2k-i-i'} \binom{j+p}{i} \binom{j-p}{j-k-i} \binom{l+p}{i'} \binom{l-p}{l-k-i'} \\ \times e^{r(ip-1+p-k-2i')} \int^\infty_0 2 |\omega| (1+|\omega|^2)^{-ip -1 -j } (e^{-2r}+|\omega|^2)^{ip -1 -l } |\omega|^{2(j+l-i'-i+p-k)} d|\omega| \Bigg]\end{gathered}$$ $$\begin{aligned} D^{(p,k)}_{jplp} (e^{r\sigma_3/2}) &= \int_{\mathbb{C}} d\omega \overline{f^{(p,k)}_{jp} (\omega)} e^{r(1-ip)} f^{(p ,k)}_{lp} (e^r \omega) \\ &= \int_{\mathbb{C}} d\omega \sqrt{\frac{2j+1}{\pi}} (1+|\omega|^2)^{-ip -1 -j } \overline{D}^j_{-k,p} \begin{pmatrix} 1 & -\omega^* \\ \omega & 1 \end{pmatrix} e^{r(1-ip)} \\ & \quad \times \sqrt{\frac{2l+1}{\pi}} (1+e^{2r}|\omega|^2)^{ip -1 -l } D^l_{-k,p} \begin{pmatrix} 1 & -e^r \omega^* \\ e^{r} \omega & 1 \end{pmatrix}.\end{aligned}$$ We conclude using the explicit expression of equation \[Wig mat 11\]. $\blacksquare$ **Integral formula 2.** $$\begin{gathered} d^{(p,k)}_{jlm}(r) = \sqrt{(2j+1)(2l+1)} \sqrt{\frac{(j-k)!(j+k)!(l-k)!(l+k)!}{(j+m)!(j-m)!(l+m)!(l-m)!}} \\ \times \sum_{n_1,n_2} \Bigg[ (-1)^{j+l+2m-n_1-n_2} \binom{j+m}{n_1} \binom{j-m}{j-k-n_1} \binom{l+m}{n_2} \binom{l-m}{l-k-n_2} \\ \times e^{r(ip -1-2n_2-k +m)} \int_0^1 dt \left[1-(1-e^{-2r})t \right]^{ip -1-l} t^{n_1+n_2+k-m} (1-t)^{j+l-n_1-n_2-k+m} \Bigg].\end{gathered}$$ Change of variables: $|\omega|^2 = \frac{1-t}{t} \Leftrightarrow t = \frac{1}{1+|\omega|^2}$ et $d\omega^2 = -\frac{1}{t^2} dt$. $\blacksquare$ **Hyper-geometrical formula.** $$\begin{gathered} d^{(p,k)}_{jlm}(r) = \frac{\sqrt{(2j+1)(2l+1)} }{(j+l+1)!} \sqrt{\frac{(j-k)!(j+k)!(l-k)!(l+k)!}{(j-m)!(j+m)!(l-m)!(l+m)!}} \\ \times \sum_{n_1,n_2} (-1)^{n_1+n_2} \binom{j+m}{n_1} \binom{j-m}{n_1-m-k} \binom{l+m}{n_2} \binom{l-m}{n_2-m-k} \\ \times (j+l-n_1-n_2+m+k)!(n_1+n_2-m-k)!e^{r(m+k-ip-1 - 2n_1)} \\ \times {}_2F_1(n_1+n_2-m-k+1, j+ip+1, j+l+2;1-e^{-2r}).\end{gathered}$$ We have used the integral expression of the hyper-geometrical function ${}_2F_1$, $${}_2F_1 (a,b,c;z) = \frac{\Gamma (c)}{\Gamma(b) \Gamma(c-b)} \int_0^1 dt \ t^{b-1} (1-t)^{c-b-1} (1-zt)^{-a}.$$ $\blacksquare$ **Rühl’s formula.** ([@ruhl] p. 64) $$\begin{gathered} d^{(p,k)}_{jlq} (r) = (2j+1)^{1/2} (2l+1)^{1/2} \int_0^1 dt ((1-t) e^r + t e^{-r})^{-1 + i p} d^j_{-k,q} \left( \arccos(2t-1) \right) \\ \times d^l_{-k,q} \left( \arccos \left( \frac{t e^{-r} - (1-t)e^r }{t e^{-r} + (1-t)e^r } \right) \right) \end{gathered}$$ Recoupling of SL(2,C) {#recoupling SL2C} --------------------- ##### $SL_2(\mathbb{C})$-Clebsch-Gordan coefficients. Similarly to the $SU(2)$ case, the tensor product of two irreps of $SL_2(\mathbb{C})$ can be decomposed into a direct sum of irreps: $$\mathcal{D}^{(p_1,k_1)} \otimes \mathcal{D}^{(p_2,k_2)} \cong \int_{\mathbb{R}} dp \bigoplus_{\substack{k \in \mathbb{Z}/2 \\ k_1 + k_2 + k \in \mathbb{N}}} \mathcal{D}^{(p,k)}.$$ Kerimov and Verdiev first got interested in the generalisation of the Clebsch-Gordan coefficients to the irreps of $SL_2(\mathbb{C})$ [@Kerimov1978]. The $SL_2(\mathbb{C})$-Clebsch-Gordan coefficients are defined by the relation $$\ket{p,k ; j,m} = \int dp_1 dp_2 \sum_{k_1j_1m_1} \sum_{k_2j_2m_2} C^{p kjm}_{p_1 k_1j_1m_1, p_2 k_2j_2m_2} \ket{p_1,k_1;j_1m_1} \otimes \ket{p_2,k_2;j_2,m_2}.$$ The coefficients are non-zero only when $k_1 + k_2 + k_3 \in \mathbb{N}$, in addition to the usual triangle inequality $|j_1-j_2| \leq j_3 \leq j_1 + j_2$. We have explicit expression for the $SL_2(\mathbb{C})$-Clebsch-Gordan coefficients but they are a bit tough. First of all, remark that the magnetic part factorises as $$C^{p_3 k_3j_3m_3}_{p_1 k_1j_1m_1, p_2 k_2j_2m_2} = \chi (p_1, p_2, p_3, k_1, k_2,k_3;j_1,j_2,j_3) C^{j_3m_3}_{j_1m_1j_2m_2}.$$ $\chi$ is a function of 9 variables which can be computed by the following expression (found initially in [@Kerimov1978] but corrected slightly in [@Speziale2017]): $$\begin{gathered} \chi (p_1, p_2, p_3, k_1, k_2,k_3;j_1,j_2,j_3)= \kappa (-1)^{(j_1+j_2+j_3+k_1+k_2+k_3)/2} (-1)^{ - k_2 - k_1} N_{p_1}^{j_1} N_{p_2}^{j_2} \overline{N_{p_3}^{j_3}} \\ \times \frac{ 1}{4\sqrt{2\pi}} \sqrt{(2j_1+1)(2j_2+1)(2j_3+1)} \left( \frac{(j_1-k_1)!(j_2+k_2)!}{(j_1+k_1)!(j_2-k_2)!}\right)^{1/2} \\ \times \Gamma(1 - \nu_3+\mu_3 ) \Gamma(1-\nu_3-\mu_3) \sum_{n=-j_1}^{j_1} \left( \frac{(j_1-n)!(j_2+k_3-n)!}{(j_1+n)!(j_2-k_3+n)!} \right)^{1/2} C^{j_3 k_3}_{j_1n;j_2,k_3-n}\\ \times \sum_{l_1 = \max(k_1,n)}^{\min(j_1,k_3+j_2)} \sum_{l_2 = \max(-k_2,n-k_3)}^{j_2} \frac{(j_1+l_1)!(j_2+l_2)!(-1)^{l_1 - k_1 + l_2 + k_2}}{(j_1-l_1)!(l_1-k_1)!(l_1-n)!(j_2-l_2)!(l_2+k_2)!(l_2-n+k_3)!}\\ \times \\ \times \frac{\Gamma(2 - \nu_1 -\nu_2 - \nu_3 + \mu_1 + l_1 + l_2 -n) \Gamma (1- \nu_1 + \mu_3 + l_1) \Gamma (1- \nu_2 - \mu_3 + l_2)}{\Gamma(2-\nu_1 - \nu_2 + l_1 + l_2) \Gamma (1- \nu_3 + \mu_1 - n) \Gamma(2-\nu_1-\nu_3+l_1) \Gamma(2-\nu_3 - \nu_2 +l_2)}\end{gathered}$$ with $$\begin{array}{l} \nu_1 = \frac{1}{2} (1 + ip_1 - i p_2 - ip_3) \\ \nu_2 = \frac{1}{2} (1 - ip_1 + i p_2 - ip_3) \\ \nu_3 = \frac{1}{2} (1 + ip_1 + i p_2 + ip_3) \\ \mu_1 = \frac{1}{2} (- k_1 + k_2 + k_3) \\ \mu_2 = \frac{1}{2} (k_1 - k_2 + k_3) \\ \mu_3 = \frac{1}{2} (- k_1 - k_2 - k_3) \\ \end{array}$$ and a phase $$\kappa = \frac{\Gamma (\nu_1 + \mu_1) \Gamma(\nu_2 + \mu_2) \Gamma(\nu_3 + \mu_3) \Gamma(-1+\nu_1+\nu_2+\nu_3+\mu_1+\mu_2+\mu_3)}{|\Gamma(\nu_1+\mu_1)\Gamma(\nu_2+\mu_2) \Gamma(\nu_3+\mu_3)\Gamma(-1+\nu_1+\nu_2+\nu_3+\mu_1+\mu_2+\mu_3)|},$$ and $$N_p^j = \frac{\Gamma(1+j+ip)}{|\Gamma(1+j+ip)|},$$ and the usual gamma function defined over $\mathbb{C}$ by analytic continuation of $$\Gamma ( z) = \int_0^{+\infty} t^{z-1}\,e^{-t}\,\dd t, \quad {\rm with} \quad \Re z >0.$$ The phase $\kappa$ satisfying $|\kappa|=1$ was chosen to make the $SL_2(\mathbb{C})$-Clebsch-Gordan coefficients real (equivalent to the Condon-Shortley convention in the $SU(2)$ case). Contrary to the usual $SU(2)$-Clebsh-Gordan coefficients, there is no consensual convention for this phase. The choice of Kerimov differs from that of Anderson [@Anderson1970] or Speziale [@Speziale2017]. These seemingly intricate expressions have nevertheless been used with much efficiency in [@Speziale2017] to numerically compute spin-foam amplitudes. The formula is indeed interesting because it is expressed with only finite sums. ##### Graphical calculus. When willing to define a graphical calculus for $SL_2(\mathbb{C})$ one encounters the difficulty of finding a good $SL_2(\mathbb{C})$-equivalent to the $3jm$-symbol of $SU(2)$ recoupling theory, so that it would satisfy the good symmetry relations to be well-represented by a $3$-valent vertex. This issue is investigated in [@Anderson1970], but the symmetry relations are intricate and depend on the convention chosen for the phase $\kappa$. As a result there is no consensus about the definition of the rules of graphical calculus for $SL_2(\mathbb{C})$. Following the phase convention of [@Speziale2017], we define then $${\begin{pmatrix} (p_1,k_1) & (p_2,k_2) & (p_3,k_3) \\ (j_1,m_1) & (j_2,m_2) & (j_3,m_3) \end{pmatrix}} \overset{\rm def}= (-1)^{2j_1-j_2+j_3-m_3} C^{p_3 k_3j_3,-m_3}_{p_1 k_1j_1m_1, p_2 k_2j_2m_2}.$$ Graphically it corresponds to the $3$-valent vertex $${\begin{pmatrix} (p_1,k_1) & (p_2,k_2) & (p_3,k_3) \\ (j_1,m_1) & (j_2,m_2) & (j_3,m_3) \end{pmatrix}} = \begin{array}{c} \begin{overpic}[scale = 0.6]{3CG-out.png} \put (-15,70) {$(p_1,k_1)$} \put (37,70) {$(p_2,k_2)$} \put (90,65) {$(p_3,k_3)$} \end{overpic} \end{array}$$ With the same rules of orientation and summation as that of section \[graphical calculus\], we can then fully develop the graphical calculus of $SL_2(\mathbb{C})$. For instance, we can define $SL_2(\mathbb{C})$-invariant functions, like the $(6p,6k)$-symbol. The $SL_2(\mathbb{C})$-$15j$-symbol can be used to define the spin-foam amplitude (see section \[section spin-foam\]). Loops and Foams in a nutshell {#presentation of LQG} ============================= Loop Quantum Gravity (LQG) is a good candidate theory for quantum gravity. It is obtained by the canonical quantisation of general relativity and describes the quantum states of space with the so-called spin-networks. Spin-foam theory is a later spinoff of both LQG and the sum-over-histories approach to quantum gravity. It describes quantum space-time, seen as the time evolution of spin-networks. Most of the main textbooks provide a derivation of the theory, following more or less its historical developments through the process of quantisation [@CLQG; @Baez99; @introductory]. Here we will only introduce the general mathematical framework of the theory, trying to be as concise as possible, since we believe that a full-fledged fundamental theory should come to a point where it stands on its own, with its mathematical framework and physical principles, without any reference to older approximate theories like general relativity or non-relativistic quantum mechanics. Spin-network ------------ As any good quantum theory, LQG comes with an Hilbert space. It is the mathematical space of the various possible states of physical space. A very convenient basis is parametrised by the so-called spin-networks that we first define. ##### Spin-network. An *abstract[^12] directed graph $\Gamma$* is an ordered pair $\Gamma = (\mathcal{N}, \mathcal{L})$, where $\mathcal{N} = \{ n_1,...,n_N\}$ is a finite set of $N$ nodes, and $\mathcal{L} = \{ l_1,...,l_L\}$ a finite set of $L$ links[^13], endowed with a target map $t:\mathcal{L} \to \mathcal{N}$ and a source map $s:\mathcal{L} \to \mathcal{N}$, assigning each link to its endpoints (respectively the head or the tail, defined by the orientation). We denote $\mathcal{S}(n)$ (resp. $\mathcal{T}(n)$) the set of links for which the node $n$ is the source (resp. the target). The valency of a node $n$ is the number of links which have $n$ as a endpoint. A graph is said to be $p$-valent if the valency of each node is $p$. Given a directed graph $\Gamma$, we denote $\Lambda_\Gamma$ the set of labellings $j$ that assign to any link $l \in \mathcal{L}$, an $SU(2)$-irrep $j_l \in \mathbb{N}/2$. Given a labelling $j \in \Lambda_\Gamma$, we denote $$\label{inter hilb node} {\rm Inv} (n, j) \overset{\rm def}= {\rm Inv}_{SU(2)} \left(\bigotimes_{l \in \mathcal{S}(n)} \mathcal{Q}_{j_l} \otimes \bigotimes_{l \in \mathcal{T}(n)} \mathcal{Q}_{j_l}^* \right).$$ The tensor product above assumes that an ordering of the links around a node, i.e. a sense of rotation and a starting link, has been prescribed. A *spin-network* is a triple $\Sigma = (\Gamma, j, \iota)$, where $\Gamma$ is a directed graph, $j \in \Lambda_\Gamma$, and $\iota$ is a map that assigns to any $n \in \mathcal{N}$ an intertwiner $\ket{\iota_n} \in {\rm Inv}(n,j)$. Figure \[spinnetwork\] shows a pictorial representation of a $4$-valent spin-network. [4spinnetwork.png]{} (45,70) [$j_1$]{} (85,65) [$j_2$]{} (33,53) [$j_3$]{} (12,22) [$j_4$]{} (35,25) [$j_5$]{} (55,20) [$j_6$]{} (65,28) [$j_7$]{} (80,20) [$j_8$]{} (10,65) [$\ket{\iota_1}$]{} (90,45) [$\ket{\iota_2}$]{} (55,35) [$\ket{\iota_3}$]{} (35,5) [$\ket{\iota_4}$]{} ##### Hilbert space. The Hilbert space of LQG is given by $$\mathcal{H} = \bigoplus_\Gamma \mathcal{H}_\Gamma$$ where the direct sum is made over all possible directed $4$-valent graphs $\Gamma$, and $\mathcal{H}_\Gamma$ is $$\mathcal{H}_\Gamma = \bigoplus_{j \in \Lambda_\Gamma} \bigotimes_{n \in \mathcal{N}} {\rm Inv} (n,j)$$ It is spanned by the set of *spin-networks states* $$\ket{\Gamma, j, \iota} = \bigotimes_{n \in \mathcal{N}} \ket{\iota_n}$$ where $\Gamma$ range over all possible $4$-valent graphs, $j$ over $\Lambda_\Gamma$, and $\ket{\iota_n}$ over an orthonormal basis of ${\rm Inv}(n,j)$. By definition of the invariant space ${\rm Inv} (n, j)$, it is straightforward to see that “the action of any $g_n \in SU(2)$ over a node $n$”, i.e. over ${\rm Inv} (n, j)$, leaves the spin-network states invariant: $$\label{Gauss constraint} g_n \cdot \ket{\Gamma, j, \iota} = \ket{\Gamma, j, \iota}.$$ With this property, the spin-network states are said to satisfy the *Gauss constraint*[^14] at each node. Since we only consider $4$-valent graphs, an orthonormal basis of ${\rm Inv}(n,j)$ is given by the states of equation \[intertwiner 4jm\]. Thus, instead of writing the abstract states $\ket{\iota}$, it is equivalent to split each $4$-valent node (according to the prescribed ordering of the links around the nodes), like $$\label{split} \begin{array}{c} \begin{overpic}[scale = 0.8]{4valent.png} \put (20,80) {$j_1$} \put (60,80) {$j_2$} \put (60,15) {$j_3$} \put (25,15) {$j_4$} \put (60,45) {$\ket{\iota}_{12}$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.8]{split.png} \put (5,75) {$j_1$} \put (55,75) {$j_2$} \put (50,5) {$j_3$} \put (20,5) {$j_4$} \put (45,45) {$\iota$} \end{overpic} \end{array}$$ and then associate to the virtual link the spin $\iota \in \{ \max (|j_1-j_2|,|j_3-j_4|) ,...,\min (j_1+j_2,j_3+j_4) \} $, which parametrises the basis $\ket{\iota}_{12}$ of equation \[intertwiner 4jm\]. By metonymy the spin $\iota$ is also called an intertwiner. Thus the spin-network of Figure \[spinnetwork\], becomes $$\begin{overpic}[scale = 0.8]{4spinnetwork-splitted.png} \put (45,65) {$j_1$} \put (85,65) {$j_2$} \put (45,35) {$j_3$} \put (12,22) {$j_4$} \put (42,15) {$j_5$} \put (62,15) {$j_6$} \put (65,37) {$j_7$} \put (80,15) {$j_8$} \put (8,60) {$\iota_1$} \put (90,45) {$\iota_2$} \put (62,25) {$\iota_3$} \put (50,0) {$\iota_4$} \end{overpic}$$ ##### Spin-network wave function. The isomorphism \[Peter L(SU(2))\], deduced from Peter-Weyl’s theorem, offers another possible realisation of $\mathcal{H}_\Gamma$, as a subspace of $L^2(SU(2)^N)$, denoted $L^2(SU(2)^N)_\Gamma$. A spin-network state $\ket{\Gamma, j, \iota}$ becomes a function $\Psi_{(\Gamma,j,\iota)} \in L^2(SU(2)^N)_\Gamma$, obtained with the following procedure: 1. Associate to each link $l$ $$\label{line} \begin{array}{c} \begin{overpic}[scale = 0.8]{line.png} \put (40,20) {$j_l$} \put (0,25) {$n_l$} \put (95,25) {$m_l$} \end{overpic} \end{array} \cong D^{j_l}_{m_ln_l}(g_l)$$ with the magnetic indices $m_l$ or $n_l$, depending of the orientation, and the variable $g_l \in SU(2)$. 2. Associate to each (splitted) node a $4jm$ symbol, like $$\begin{array}{c} \begin{overpic}[scale = 0.8]{split.png} \put (5,75) {$j_1$} \put (55,75) {$j_2$} \put (50,5) {$j_3$} \put (20,5) {$j_4$} \put (45,45) {$\iota$} \end{overpic} \end{array} \cong (-1)^{j_4-n_4} {\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & -n_4 \end{pmatrix}^{(\iota)}}$$ with an index $-n$ and a phase $(-1)^{j-n}$ for outgoing links. 3. Finally multiply all together, and sum over all the magnetic indices. Thus, we obtain a set of *spin-network wave functions* $\Psi_{(\Gamma,j,\iota)}(g_{l_1},...,g_{l_L})$ that span a peculiar subspace of $L^2(SU(2)^N)$, denoted[^15] $L^2(SU(2)^N)_\Gamma$. For instance, the spin-network $$\begin{array}{c} \begin{overpic}[scale = 0.8]{doublethetaspin.png} \put (15,75) {$j_1$} \put (55,70) {$j_2$} \put (55,40) {$j_3$} \put (60,5) {$j_4$} \put (5,50) {$\iota$} \put (85,50) {$\kappa$} \end{overpic} \end{array}$$ encodes the function $$\begin{gathered} \Psi(u_1,u_2,u_3,u_4) \\ = \sum_{m_i,n_i} (-1)^{\sum_i (j_i - n_i)} {\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ -n_1 & -n_2 & -n_3 & -n_4 \end{pmatrix}^{(\iota)}} {\begin{pmatrix} j_1 & j_2 & j_3 & j_4 \\ m_1 & m_2 & m_3 & m_4 \end{pmatrix}^{(\kappa)}} \prod_{i=1}^4 D^{j_i}_{m_in_i}(u_i).\end{gathered}$$ From the isomorphism \[Peter L(SU(2))\], we can express the Gauss constraint \[Gauss constraint\] as an invariance of the functions $\Psi_{(\Gamma,j,\iota)}(g_{l_1},...,g_{l_N})$: for all sets $(u_n) \in SU(2)^L$, parametrised by the nodes $n \in \mathcal{N}$, we have $$\label{Gauss constraint functions} \Psi_{(\Gamma,j,\iota)}(g_{l_1},...,g_{l_N}) = \Psi_{(\Gamma,j,\iota)}(u_{s(l_1)} g_{l_1} u^{-1}_{t(l_1)} ,..., u_{s(l_N)} g_{l_N}u^{-1}_{t(l_1)})$$ with $s$ and $t$ the source and target map of the gr aph. In fact, $L^2(SU(2)^N)_\Gamma$ can be characterized as the subspace of functions of $L^2(SU(2)^N)$ that satisfy this property. Notice finally that evaluating the function at the identity on all links result in the graphical calculus previously defined in section \[graphical calculus\]. ##### Algebra of observables. In fact, there is not much information in the Hilbert space itself. What really matters physically is the algebra of observables $\mathcal{A}$ acting upon it. The observables of LQG are obtained by the principle of correspondence. Thus, they come with a geometrical interpretation: they correspond notably to measurements of area or measurement of volume. The Hilbert space $\mathcal{H}$ is built from the building block spaces $\mathcal{Q}_{j_l}$, where $j_l$ labels a link $l$. Similarly, the algebra of observables is built from the action of $\mathfrak{su}(2)$ (the flux) and $SU(2)$ (the holonomy) over $\mathcal{Q}_{j_l}$. Notice that an observable should not “go out” of $\mathcal{H}$: in other words, an observable should commute with the Gauss constraint. Given a graph $\Gamma$, the observable of area associated to a link $l$ is $$\hat A_l \overset{\rm def}= 8 \pi \frac{\hbar G}{c^3} \gamma \sqrt{\vec{J}^2_l},$$ where $\gamma$ is a real parameter called the *Immirzi parameter*, and $\vec{J}_l$ are the generators of $SU(2)$ acting over $\mathcal{Q}_{j_l}$. The spin-network basis diagonalises $\hat A_l$: $$\begin{aligned} \hat A_l \ket{\Gamma, j, \iota} \overset{\rm def}= j_l (j_l+1) \ket{\Gamma, j, \iota}.\end{aligned}$$ It also diagonalises the observable $(\vec{J}_1 + \vec{J}_2)^2$, acting over a node $n$, $$\label{node} \begin{array}{c} \begin{overpic}[scale = 0.8]{split.png} \put (5,75) {$j_1$} \put (55,75) {$j_2$} \put (50,5) {$j_3$} \put (20,5) {$j_4$} \put (45,45) {$\iota_{12}$} \end{overpic} \end{array},$$ so that $$\begin{aligned} (\vec{J}_1 + \vec{J}_2)^2 \ket{\Gamma, j, \iota} \overset{\rm def}= \iota_{12} (\iota_{12}+1) \ket{\Gamma, j, \iota}.\end{aligned}$$ The latter observable encodes a notion of “angle” between the links $j_1$ and $j_2$. Given a graph $\Gamma$, the set of area observables associated to each link and the set of “angle operators” like $(\vec{J}_1 + \vec{J}_2)^2$ (one per each node), define a complete set of commuting observables over $\mathcal{H}_\Gamma$, diagonalised by the spin-network basis. On each node like \[node\], we can also define the volume operator $$\hat V_n = \frac{\sqrt{2}}{3} (8 \pi G \hbar \gamma)^{3/2} \sqrt{|\vec{J}_1 \cdot (\vec{J}_2 \times \vec{J}_3)|}.$$ It is not diagonalised by the spin-network basis, but its eigenvalues can be computed numerically. It does not commute with $(\vec{J}_1 + \vec{J}_2)^2$ but it does with the areas, so that the areas $\hat A_l$ and the volumes $\hat V_n$ form another complete set of commuting observables (diagonalised by another basis than that of spin-networks). These geometric operators of area, volume or angle, built from the principle of correspondance, suggest a vision of the “quantum geometry”. It is obtained as the dual picture of a graph $\Gamma$: a tetrahedron is associated to each node, and they glue together along faces (whose area is given by the eigenvalue of $\hat A_l$) dual to links. Spin-foam {#section spin-foam} --------- ##### Dynamics. The latter mathematical framework of loop quantum gravity is obtained through the canonical quantisation of general relativity: the spin-network states represent quantum states of space. The time evolution of these states should be found by finding the subspace formed by the solution to the hamiltonian constraint $\hat H \ket{\Psi} = 0$, where $\ket{\Psi}$ is a superposition of spin-network states, $\hat{H}$ the quantized hamiltonian. This hard path of finding the dynamics was followed notably by Thiemann [@ThiemannBook]. Below we present a way to short-circuit the issue, called spin-foams, which takes inspiration from former sum-over-histories approaches to quantum gravity. Spin-foams can be seen as the time evolution of spin-networks, or also as quantum states of space-time. ##### Spin-foams. Spin-foams can be seen both as a higher dimensional version of Feynman diagrams propagating the gravitational field, and as the time evolution of spin-networks. Spin-foams are built out of combinatorial objects, which generalise graphs to higher dimensions, called *piecewise linear cell complexes*, often abbreviated as *complexes*. An *oriented $2$-complex* is an ordered triple $\kappa = (\mathcal{E}, \mathcal{V}, \mathcal{F})$, with a finite set $\mathcal{E} = \{ e_1,...e_E\}$ of edges, a finite set $\mathcal{V} = \{ v_1,...v_V\}$ of vertices, and a finite set $\mathcal{F} = \{ f_1,...,f_F\}$ of faces, such that they all “glue consistently”[^16]. The orientation is given on the edges by a target map $t:\mathcal{E} \to \mathcal{V}$ and a source map $s:\mathcal{E} \to \mathcal{V}$, and and the orientation of each faces gives a cyclic ordering of its bounding vertices. Given an oriented $2$-complex $\kappa$, we denote $\Lambda_\kappa$ the set of labellings $j$ that assign an $SU(2)$-irrep $j_f \in \mathbb{N}/2$ to any face $f \in \mathcal{F}$. Similarly we denote $I_\kappa$ the set of labellings $\iota$ that assign to each edge $e$ an intertwiner $\ket{\iota_e}$, $$\label{inter hilb vertex} \ket{\iota_e} \in {\rm Inv} (e, j) \overset{\rm def}= {\rm Inv}_{SU(2)} \left(\bigotimes_{f \in \mathcal{F}(e)} \mathcal{Q}_{j_f} \otimes \bigotimes_{f \in \mathcal{F}^*(e)} \mathcal{Q}_{j_f}^* \right),$$ where $\mathcal{F}(e)$ and $\mathcal{F}^*(e)$ are the set of faces adjacent to the edge $e$, whose orientation respectivelly matches and does not match that of $e$. A *spin-foam* is a triple $F = (\kappa, j, \iota)$. where $\kappa$ is an oriented $2$-complex, $j \in \Lambda_\kappa$, and $\iota \in I_\kappa$. We can stick to a purely “abstract” combinatorial definition of $2$-complexes, but we can also adopt a geometrical “hypostasis” that represents “faces” as polygons. For instance, Figure \[spin-foam\] shows a spin-foam embedded into $3$-dimensional euclidean space. [spinfoam.png]{} Notice that such a graphical representation is not always possible in $3$ dimensions, and sometimes a fourth dimension can be required. Interestingly, the boundary of a $2$-complex[^17] is a graph, as can be seen on Figure \[spin-foam\]. Thus, the boundary of a spin-foam is spin-network. The vertices and the edges of the boundary are called respectively nodes and links. Each link bounds an inside face, so that the spin of the link is also the spin of the face. Similarly, each node is an endpoint of an inside edge, so that the associated intertwiners match. ##### Spin-foam amplitude. To each spin-foam we associate an amplitude, which is like the propagator associated to a Feynman diagram. Its interpretation is made precise below. Given a spin-foam $(\kappa, j , \iota)$, we define the spin-foam amplitude as $$\mathcal{A}(\kappa, j , \iota) = \left(\prod_{f \in \mathcal{F}} (2j_f + 1) \right) \left(\prod_{e \in \mathcal{E}} (2\iota_e + 1) \right) \left(\prod_{v \in \mathcal{V}} A_v (j, \iota)\right).$$ $A_v$ is called the vertex amplitude. In the short history of spin-foam amplitudes there has already been many various formulae proposed for the vertex amplitude. First, let us say that for quantum gravity, it is sufficient to consider spin-foams whose vertices are $5$-valent ($5$ edges attached to it) and whose edges are $4$-valent ($4$ faces attached to it). This restriction comes from the fact that the $2$-complexes of quantum gravity are built by dualising the triangulation of a $4$-dimensional manifold. Unfortunately there is no possible nice picture as Figure \[spin-foam\] to visualise such a $2$-complex since it cannot be embedded into $3$-dimensional euclidean space. However it is sufficient to get an idea of the combinatorial structure of each vertex by representing the adjacent edges with dots and the faces with lines, so that we draw the *vertex graph* $$\label{vertex graph} \begin{array}{c} \begin{overpic}[scale = 0.6]{pentagram.png} \put (25,70) {$j_1$} \put (38,47) {$j_2$} \put (60,45) {$j_3$} \put (75,60) {$j_4$} \put (45,60) {$j_5$} \put (55,25) {$j_6$} \put (75,15) {$j_7$} \put (30,30) {$j_8$} \put (35,5) {$j_9$} \put (0,35) {$j_{10}$} \put (50,80) {$\ket{\iota_1}$} \put (0,60) {$\ket{\iota_2}$} \put (10,10) {$\ket{\iota_3}$} \put (70,0) {$\ket{\iota_4}$} \put (95,50) {$\ket{\iota_5}$} \end{overpic} \end{array} = \begin{array}{c} \begin{overpic}[scale = 0.6]{splitted-pentagram.png} \put (25,75) {$j_1$} \put (33,47) {$j_2$} \put (60,50) {$j_3$} \put (80,70) {$j_4$} \put (50,55) {$j_5$} \put (55,35) {$j_6$} \put (95,35) {$j_7$} \put (37,37) {$j_8$} \put (55,5) {$j_9$} \put (0,35) {$j_{10}$} \put (45,81) {$\iota_1$} \put (0,60) {$\iota_2$} \put (15,5) {$\iota_3$} \put (80,10) {$\iota_4$} \put (93,60) {$\iota_5$} \end{overpic} \end{array}$$ The orientation and the spin of the links, and the intertwiners of the nodes are naturally inherited from the underlying spin-foam, so that the vertex graph is a spin-network. To avoid confusion, let us recap. Each spin-foam comes with a *boundary spin-network*, and also with a *vertex graph* for each of its vertex. If the spin-foam is made of only one vertex, then the boundary spin-network and the vertex graph coincides. Contrary to the the boundary spin-network, there is in general no interpretation of the vertex graphs in terms of quantum states of space. The combinatorial shape of each vertex suggests to define the amplitude $A_v$ as the value obtained with the rules of graphical calculus of $SU(2)$ recoupling theory, defined in section \[graphical calculus\]. This is precisely what Ooguri did in [@Ooguri:1992eb] by defining the vertex amplitude as the $\{ 15 j \}$-symbol, but it later appeared not to be a good candidate for quantum gravity. Since then many other models were suggested [@Perez2012]. They all consist in finding other rules than that of $SU(2)$ recoupling theory to assign a value to the vertex graph \[vertex graph\]. The EPRL model, introduced in [@Engle2008], is a model that is still considered a good candidate for quantum gravity. The vertex amplitude is computed from the vertex graph \[vertex graph\] with the following rules: 1. Compute the spin-network wave function as shown in the previous section. We obtain a function of $L^2(SU(2)^{10})$ which satisfies the Gauss constraint \[Gauss constraint functions\]: $$\Psi_{(\Gamma,j,\iota)}(g_{l_1},...,g_{l_{10}})$$ 2. Apply the so-called *$Y_\gamma$-map*, which is the linear map $Y_\gamma : L^2(SU(2)^{10}) \to \mathcal{F}(SL_2(\mathbb{C})^{10})$ defined over the canonical basis of Wigner matrix coefficients by $$Y_\gamma \left( \prod_i D^{j_i}_{m_in_i} \right) = \prod_i D^{(\gamma j_i,j_i)}_{j_im_ij_in_i} ,$$ where $\gamma$ is the Immirzi parameter. We thus obtain a function of $\mathcal{F}(SL_2(\mathbb{C})^{10})$ $$Y_\gamma\Psi_{(\Gamma,j,\iota)}(h_{l_1},...,h_{l_{10}}) .$$ It still satisfies the invariance of the Gauss constraint \[Gauss constraint functions\] for $SU(2)$ action, be not for $SL_2(\mathbb{C})$. 3. Project down to the $SL_2(\mathbb{C})$-invariant subspace on each node with the projector $P_{SL_2(\mathbb{C})}$ acting as $$P_{SL_2(\mathbb{C})} Y_\gamma \Psi_{(\Gamma,j,\iota)}(h_{l_1},...,h_{l_{10}}) = \int_{SL_2(\mathbb{C})} \left(\delta(a_{n_5}) \prod_{n \in \mathcal{N}} \dd a_n \right) \Psi_{(\Gamma,j,\iota)}(a_{s(l_1)} h_{l_1} a^{-1}_{t(l_1)} ,..., a_{s(l_{10})} h_{l_{10}} a^{-1}_{t(l_1)}).$$ with $n_5$ any of the 5 nodes. The delta function $\delta(a)$ (only non-vaninshing when $a = \mathbb{1}$) is required to avoid the divergence of the integration, but the final result does not depend on the choice of node $n_5$. To put it differently the integration is only effective over (any) four nodes, while the fifth $a_{n_5}$ is fixed to the identity $\mathbb{1}$. 4. Evaluate all the variables $h_l$ to $\mathbb{1}$. So if $(\Gamma,j,\iota)$ is the vertex graph of a vertex $v$ in a spin-foam $(\kappa,j,\iota)$, we can finally write in a nutshell $$A_v (j, \iota) = \left( P_{SL_2(\mathbb{C})} Y_\gamma \Psi_{(\Gamma,j,\iota)} \right) (\mathbb{1}).$$ Thus we have fully defined the spin-foam amplitude $\mathcal{A}(\kappa,j,\iota)$ of the EPRL model. The specificity of this model is the $Y_\gamma$-map which selects only the irreps $(p=\gamma j, k=j)$ among the principal series of $SL_2(\mathbb{C})$. It implements the so-called *simplicity constraints*, which enable formulation of general relativity as a BF theory [@Baez99]. Besides, the apparently sophisticated procedure should not hide the fact that the value of $A_v(j,\iota)$ is the same as that obtained from the $SL_2(\mathbb{C})$ graphical calculus, defined in section \[recoupling SL2C\], when the simplicity constraint is applied. For those only interested in the actual computation of the amplitude of a given vertex graph, we can summarise the previous procedure with the following algorithm: 1. Associate a variable $h_p \in SL_2(\mathbb{C})$ to each intertwiner $\iota_p$. 2. Associate to each link $$\begin{array}{c} \begin{overpic}[scale = 0.8]{line.png} \put (40,20) {$j$} \put (95,20) {$\iota_{p}$} \put (0,20) {$\iota_{q}$} \end{overpic} \end{array} \cong D^{(\gamma j,j)}_{jmjn}(h^{-1}_p h_q).$$ 3. Associate a $3jm$-symbol to each node as in usual graphical calculus (equation \[graphical 3jm\]). 4. Multiply everything together and sum over all the magnetic indices $m$ and $n$. 5. Integrate over (any) four of the five $SL_2(\mathbb{C})$ variables $h_p$, and fix the fourth to the identity $\mathbb{1}$. ##### Interpretation. The interpretation of spin-foams relies on the general boundary formulation of quantum mechanics which was introduced by Oeckl [@Oeckl:2003vu; @Oeckl:2005bv]. Consider a finite region of spacetime. Its boundary $\Sigma$ is a $3$-dimensional hypersurface which constitutes the quantum system under consideration. Its space of states is the Hilbert space of LQG, $\mathcal{H}$, spanned by the spin-network states. An observer $\mathcal{O}$ may know some partial information about the state $\psi$ of $\Sigma$, which can be expressed by the fact that $\psi \in \mathcal{S}$, where $\mathcal{S}$ is a linear subspace of $\mathcal{H}$. Then $\mathcal{O}$ can carry out measurements with the operators of the algebra, to determine information about $\psi$. If $\mathcal{A}$ is a linear subspace of $\mathcal{S}$, then the probability to find $\psi \in \mathcal{A}$ is $$P(\mathcal{A} | \mathcal{S}) = \frac{\sum_{i \in I} |\rho(\xi_i)|^2}{\sum_{j \in J} |\rho(\zeta_j)|^2},$$ where $\xi_i$ (resp. $\zeta_j$) is an orthonormal basis of $\mathcal{A}$ (resp. $\mathcal{S}$). $\rho : \mathcal{H} \to \mathbb{C}$ is a linear map, called the *transition amplitude* defined for a spin-network state $\Psi$ by $$\rho(\Psi) \overset{\rm def}= \sum_\sigma W_\sigma (\Psi)$$ where the sum is done over all possible spin-foams $\sigma$ which have $\Psi$ as a boundary, and $W_\sigma(\Psi)$ is the $2$-complex amplitude defined as $$W_\sigma(\Psi) = \sum_j \sum_\iota \mathcal{A}(\sigma, j , \iota)$$ where the sum is done over all the possible spin labelling $j \in \Lambda_\kappa$, and intertwiner labellings $\iota \in I_\kappa$, that are compatible with the spin-network $\Psi$ at the boundary. This completes the mathematical formulation of the theory and its probabilistic interpretation. Of course, much remains to be discovered. In particular, the theory has yet to meet the benchmark of experimental evidence, but this would be another story to tell! Representations and intertwiners {#representations and intertwiners} ================================ ##### Representation of groups. A good way to understand the structural properties of a group is to look how it can act on vector spaces. By “action”, I mean a linear action that preserves the group product: it is called a “representation”. In physics, notably in quantum mechanics, we often focus on representations over Hilbert spaces. Let $G$ be a locally compact group, and $GL(\mathcal{H})$ the group of bounded linear operators over a Hilbert space $\mathcal{H}$ that admit a bounded inverse. A (bounded continuous) *representation* $\rho$ of $G$ over $\mathcal{H}$ is a homomorphism $\rho : G \to GL(\mathcal{H})$, such that the resulting map $G \times \mathcal{H} \to \mathcal{H}$ is continuous. In the case of a finite dimensional Hilbert space $\mathcal{H}$, $GL(\mathcal{H})$ is just the space of invertible linear maps, and a representation is any linear action of $G$ over $\mathcal{H}$. It is called *unitary* if it preserves the scalar product. ##### Representation of Lie algebras. There are also representations of Lie algebras, which are linear actions preserving the Lie bracket. Any representation of a Lie group defines, by differentiation, a representation of its Lie algebra. Precisely, if $\rho: G \rightarrow GL(\mathcal{H})$ is a representation of a Lie group $G$, the *differential of the representation $\rho$*, is the linear map $D\rho: \mathfrak{g} \rightarrow \mathfrak{gl}(\mathcal{H})$ defined for all $X \in \mathfrak{g}$ by: $$(D\rho)(X) = \frac{d}{dt}\rho(e^{tX})|_{t=0}.$$ It is shown to be a Lie algebra representation. Moreover, for all $X \in \mathfrak{g}$, $$\rho(e^X) = e^{D\rho(X)}.$$ One shows 1. If $F \subset \mathcal{H}$ is stable for $\rho$, then $F$ is also stable for $D \rho$. 2. If $D\rho$ is irreducible, then $\rho$ is also irreducible. 3. If $G$ is connected, the converses of (1) and (2) are also true. Conversely, given a Lie algebra $\mathfrak{g}$, there is no unique Lie group associated to it, but there is a unique simply connected one $G$, which is obtained by exponentiation of $\mathfrak{g}$. Then given any morphism of Lie algebra $\phi$, there exists a morphism of a Lie group $\rho$ such that $\phi = D\rho$. Thus a representation of $\mathfrak{g}$ will infer a representation on each of its associated Lie groups. ##### Irrep. A representation is *irreducible* if it admits no closed stable subspace other than $\left\{ 0 \right\}$ and $\mathcal{H}$. For brevity, we commonly say “irrep” instead of “irreducible representation”. They can be seen as the building blocks of the other representations. From two representations, one can build others using the direct sum and the tensor product notably. If $V$ and $W$ are two vector spaces of representation for a group $G$ and its algebra $\mathfrak{g}$, we define a representation over the direct sum $V \oplus W$ by $$\begin{aligned} \forall g \in G, \quad g \cdot (v + w) = g \cdot v + g \cdot w \\ \forall X \in \mathfrak{g}, \quad X \cdot (v + w) = X \cdot v + X \cdot w .\end{aligned}$$ We also define a representation over the tensor product $V \otimes W$ $$\begin{aligned} \forall g \in G, \quad g \cdot (v \otimes w) = (g \cdot v) \otimes (g\cdot w) \\ \forall X \in \mathfrak{g}, \quad X \cdot (v \otimes w) = (X \cdot v) \otimes w + v \otimes (X \cdot w). \end{aligned}$$ ##### Intertwiners. If $V$ and $W$ are two vector spaces of representation of a group $G$ and its algebra $\mathfrak{g}$, an *intertwiner* (or equivariant map or intertwining operator) is a linear map $T : V \rightarrow W$ satisfying: $$T( g \cdot v) = g \cdot T(v).$$ The space of intertwiners, denoted ${\rm Hom}_G(V,W)$, is a subspace of the vector space of linear maps ${\rm Hom}(V,W)$. A useful result, (generalising equation \[iso intertwiners\]) is the following isomorphism $${\rm Hom}_G(V,W) \cong {\rm Inv}_G(V \otimes W^*),$$ where ${\rm Inv}_G (E) \overset{\rm def}= \left\{ \psi \in E \mid \forall g \in G, \ g \cdot \psi = \psi \right\}$. Two representations are *equivalent* if there is an invertible intertwiner between them. An invertible intertwiner is a way to *identify* two representations, as if there were only a change of notation between them. In the language of category theory, an intertwiner is nothing but a *natural transformation* between two *functors*, each functor being a representation of the group. In the main text, we have chosen to alleviate the notations by making the intertwiner implicit, so that we write for instance (see section \[section homogeneous\]) $$J_+ \cong z_0 \frac{\partial}{\partial z_1} \quad \text{and} \quad \ket{j,m} \cong \left( \frac{(2j)!}{(j+m)!(j-m)!} \right)^{1/2} z_0^{j+m} z_1^{j-m},$$ where the symbol of congruence “$\cong$” should be understood as “equal from the perspective of the group representation”. Thus, two equivalent representations will often be presented as two *realisations* of the same representation. But of course $\cong$ is not a strict equality “$=$” in the mathematical sense since for instance $\mathbb{C}_{2j}[z_0,z_1]$ carries other mathematical structures to which the intertwiner is blind. ##### Schur’s Lemma. If $T : V \to W$ is an intertwiner between two finite irreps of $G$, then either $T=0$, or $T$ is bijective. Moreover, if the irreps are unitary and $T$ is bijective, then for any other bijective intertwiner $T'$ there exists $\lambda \in \mathbb{C}$ such that $T' = \lambda T$. ##### Peter-Weyl’s theorem. An important case is when the group $G$ is compact (e.g. $SU(2)$, but not $SL_2(\mathbb{C)}$). In this case we have the following properties: 1. Any (complex) finite representation of $G$ can be endowed with an hermitian product which makes the representation unitary. 2. Any unitary irrep of $G$ is finite-dimensional. 3. Any unitary representation can be decomposed into a direct sum of irreps. Theses results justify notably that focusing on unitary irreps of $SU(2)$, as we do in chapter \[rep theory SU(2)\], is sufficient to describe all possible finite or unitary representations of $SU(2)$. Finally, the compactness of $G$ enables to define the space of square-integrable functions $L^2(G)$ with the Haar measure, and we have 4. the linear span of all matrix coefficients of all finite unitary irreps of $G$ is dense in $L^2(G)$. A proof can be found in Knapp ([@Knapp1986] pp. 17-20). Induced representation {#induced} ====================== There is a well-known method to build a representation of group, *induced* from a representation of one of its subgroups. We present below two possible formal definitions of the method (see the book [@Maurin1997] for more details). Then we apply it to the case of $SL_2(\mathbb{C})$. Let $K$ be a subgroup of $G$, and $\rho$ a representation of $K$ over a vector space $V$. ##### Definition 1. To build a representation of $G$ starting from $\rho$, we first build a vector space $\mathcal{H}^\rho$, then a group homomorphism $U^\rho: G \rightarrow GL(\mathcal{H}^\rho)$. Let $\mathcal{H}^\rho$ be the vector space of functions $f:G \rightarrow V$ such that $$\forall g \in G, \quad \forall k \in K, \quad f(gk) = \rho(k)f(g).$$ For all $g \in G$, we define the linear map $U^\rho(g) : \mathcal{H}^\rho \rightarrow \mathcal{H}^\rho$ by $$\forall f \in \mathcal{H}^\rho, \quad \forall x \in G, \quad U^\rho(g)f(x) = f(g^{-1}x).$$ Thus $(U^\rho, \mathcal{H}^\rho)$ is the representation of $G$ *induced* from the representation $(\rho,V)$ of the subgroup $K$. ##### Definition 2. Denote the quotient $M\overset{\rm def}=G/K$. Let $P(M,K)$ be a $K$-principal bundle. Denote $P \times_\rho V \rightarrow M$ the associated vector bundle. It is a bundle of base $M$ and fibre $V$. Let $$\mathcal{H}^\rho = \left\{ \text{sections $f$ of the bundle $P \times_\rho V \rightarrow M$} \right\}.$$ For all $g \in G$, we define the linear map $U^\rho(g)$ by $$\forall f \in \mathcal{H}^\rho, \quad \forall x \in G/K, \quad (U^\rho(g) f )(x) = f(g^{-1}x).$$ Thus $(U^\rho, \mathcal{H}^\rho)$ is the representation of $G$ *induced* from the representation $(\rho,V)$ of the subgroup $K$. It is equivalent to the first definition. ##### Example. Consider the trivial subgroup $\{e\}$ of a Lie group $G$, and its trivial representation over $\mathbb{C}$. The induced representation is then given by the Hilbert space $L^2(G)$, endowed with a left-invariant (resp. right-invariant) measure, and the linear action $g \cdot f(h) = f(g^{-1} h)$ (resp. $g \cdot f(h) = f(hg)$). It is also called the left (resp. right) regular representation. ##### Application to $SL_2(\mathbb{C})$. Naimark has built the unitary representations of $SL_2(\mathbb{C})$ induced by the uni-dimensional representations of the upper-triangular subgroup $K_+$ [@naimark]. First he shows the following diffeomorphism between differentiable manifolds: $$SL_2(\mathbb{C}) / K_+ \cong \bar{\mathbb{C}}.$$ Let $H$ be a subgroup of $G$. The quotient $G/H$ is the set of all sets $Hg$ with $g \in G$. From Gauss decomposition, equation \[Gauss decomposition\], if $g \in SL_2(\mathbb{C})$ such that $g_{22} \neq 0$, there exists a unique $(k,z) \in K_+ \times Z_-$ such that: $$g = kz.$$ The group $Z_-$ is isomorphic to $\mathbb{C}$. If $a \in SL_2(\mathbb{C})$ satisfies $a_{22} = 0$ then it belongs to the class of $\begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ since $$\begin{pmatrix} a_{11} & a_{12} \\ a_{21} & 0 \end{pmatrix} = \begin{pmatrix} a_{21}^{-1} & a_{11} \\ 0 & a_{21} \end{pmatrix} \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}.$$ Thus, identifying the class of $\begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ with $\{ \infty \}$, we see that $SL_2(\mathbb{C})/K_+ \cong \bar{\mathbb{C}}$. Then we compute the expression of the induced linear action of $SL_2(\mathbb{C})$ over $\bar{\mathbb{C}}$: $$a \cdot z = \frac{a_{11} z + a_{21}}{a_{12} z + a_{22}}.$$ It is nothing but the so-called Möbius transformation. An action of $SL_2(\mathbb{C})$ over $SL_2(\mathbb{C}) / K_+$, is naturally given by: $$a \cdot Kg = Kga.$$ It leads to the given expression over $\bar{\mathbb{C}}$. Consider the Hilbert space of square integrable complex functions $L^2(\mathbb{C})$ with the scalar product: $$(f_1,f_2) \overset{\rm def}= \frac i2 \int_\mathbb{C} \overline{f_1}(z) f_2(z) \, dz \wedge d\overline{z}.$$ Here we use the usual Lebesgue measure over $\mathbb{C}$. We look for a unitary representation over $L^2(\mathbb{C})$ of the form: $$a \cdot f(z) = \alpha(z,a) f(a \cdot z).$$ Then it can be shown (after lines of computation) that for all $(\rho,m) \in \mathbb{R} \times \mathbb{Z}$, there exists a unitary representation of $SL_2(\mathbb{C})$ over $L_2(\mathbb{C})$ given by $$V_af(z) = (a_{12}z+a_{22})^{\frac {m}{2} + \frac{i\rho}{2} -1} \overline{(a_{12} z + a_{22})}^{-\frac {m}{2} + \frac{i \rho}{2}-1} f(\frac{a_{11}z+a_{21}}{a_{12}z+a_{22}}).$$ The set of representations is called the principal unitary series of $SL_2(\mathbb{C})$. It can be shown these representations are irreducible! ##### Variation. Rühl constructs an induced representation in [@ruhl] (p. 57). Formally, he first observes the following diffeomorphism between manifolds: $$SL_2(\mathbb{C})/K_+ \cong SU(2) / U(1).$$ A convenient way to see it is to decompose $a \in SL_2(\mathbb{C})$ with $k \in K_+$ and $u \in SU(2)$ such that: $$a=ku.$$ Such a decomposition exists since for all $\theta \in \mathbb{R}$: $$\begin{pmatrix} a_{11} & a_{12} \\ a_{21} & a_{22} \end{pmatrix} = \begin{pmatrix} \frac{e^{-i\theta}}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} & \frac{a_{12} a_{22}^* + a_{11}a_{21}^*}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} e^{i\theta}\\ 0 & \sqrt{|a_{21}|^2 + |a_{22}|^2}e^{i\theta} \end{pmatrix} \begin{pmatrix} \frac{a_{22}^* e^{i\theta}}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} & - \frac{a_{21}^*e^{i\theta}}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} \\ \frac{a_{21}e^{-i\theta}}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} & \frac{a_{22} e^{-i\theta}}{\sqrt{|a_{21}|^2 + |a_{22}|^2}} \end{pmatrix}.$$ But there is no uniqueness of the decomposition. If $u \in SU(2)$ decomposes $a$, then also does for all $\theta \in \mathbb{R}$ $$\begin{pmatrix} e^{i \theta} & 0 \\ 0 & e^{-i \theta} \end{pmatrix} u .$$ Conversely, it is easy to show that these are the only possible matrices of $SU(2)$ decomposing $a$. Instead of constructing a space of functions over $SL_2(\mathbb{C})/K_+$, it is then equivalent to consider functions over $SU(2)$ satisfying a covariance condition for the group $U(1)$ (in the spirit of definition 1). Thus, we consider functions over $SU(2)$ satisfying the condition: $$\phi \left(\begin{pmatrix} e^{i \theta} & 0 \\ 0 & e^{-i \theta} \end{pmatrix} u \right) = e^{i n \theta} \phi(u)$$ with $n \in \mathbb{Z}$. The choice of the factor $e^{i n \theta}$ corresponds to uni-dimensional representations of $U(1)$. This way another representation of $SL_2(\mathbb{C})$ can be built. It is equivalent to the previous one of Naimark provided that $n=-2k$. Commented bibliography {#commented} ====================== This commented bibliography gathers the main textbooks that will provide more details than this primer. Warmup {#warmup-1 .unnumbered} ------ [@Knapp1986] A. W. Knapp, *Representation Theory of Semisimple Groups*, Princeton University Press, 1986. > This book is a clear and exhaustive introduction to the representation theory of semi-simple groups, with specific focus on $SL_2(\mathbb{C})$ and its subgroup. It nevertheless requires to have followed a first semester course on Lie groups and algebras. [@Hall2003] Brian C. Hall, *Lie Groups, Lie Algebras, and Representations*, Springer, 2003. > This book provides a beautifully written introduction for physicists. [@Bernard2012] D. Bernard, Y. Laszlo and D. Renard, *[É]{}l[é]{}ments de th[é]{}orie des groupes et sym[é]{}tries quantiques*, cours de l’[É]{}cole polytechnique, 2012. > This very pedagogical introduction to groups gets inspiration from physics. It takes on a wide series of subjects in a concise manner. Unfortunately, there is only a French version. Representation and recoupling of $SU(2)$ {#representation-and-recoupling-of-su2 .unnumbered} ---------------------------------------- [@Sakurai2011] J.J Sakurai and Jim Napolitano, *Modern Quantum Mechanics*, Addison-Wesley, 2011. > This classic book is an introduction to quantum mechanics. The chapter 3 deals with the theory of angular momentum. A number of basic formulae can be found there. Here, we have used it for the Euler angles decomposition. Many other classical textbooks cover the angular momentum with small (but interesting) variations like [@Edmonds1957], [@Condon1959] and [@Yutsis1962]. [@Varshalovich1987] D. A. Varshalovich, A. N. Moskalev and V. K. Khersonskii, *Quantum theory of angular momentum*, World Scientific, 1987. > As the title suggests, this book could be looked at as the bible for the quantum aspects of angular momentum. It is supposed to be exhaustive in terms of formulae, so it is not really the kind of book you read, but rather something like a directory when you need something specific and not very memorable. [@Moussouris:1983uq] John P. Moussouris, *Quantum Models of Space-Time based on Recoupling Theory*, PhD. thesis (Oxford), 1983. > This is a beautifully written PhD thesis by Moussouris, under the supervision of Roger Penrose. It deals notably with the recoupling theory of $SU(2)$, and its link to space-time. Unfortunately, the document is not easily accessible. Representation of $SL_2(\mathbb{C})$ {#representation-of-sl_2mathbbc .unnumbered} ------------------------------------ [@ruhl] W. R[ü]{}hl, *The Lorentz Group and Harmonic Analysis*, W. A. Benjamin, Inc, 1970. > This old book was written by a physicist and is maybe too sloppy in its mathematical exposure. It is nevertheless a classic textbook with a lot of useful formulae. It focuses on the study of $SL_2(\mathbb{C})$ and $SL_2(\mathbb{R})$. [@Gelfand:1963fk] I. M. Gel’fand, R. A. Minlos and Z. Ya. Shapiro, *Representations of the rotation and Lorentz groups and their applications*, Pergamon Press, 1963. > This book proposes a self-contained presentation of the representations of the rotation and Lorentz groups. However, its rudimentary page layout makes it a bit hard to read. From that respect, the book of Naimark, one year later, is a better introduction (and is also probably more detailed in its content). [@naimark] M. A. Naimark, *Linear Representations of the Lorentz Group*, Pergamon Press, 1964. > This book introduces the subject to physicists. It is well-written, very introductory in the beginning, complete on the subject and quite rigorous (through not reaching the usual purely mathematical standards). Unfortunately the formalism and the notation start getting old and sometimes look a bit clumsy, which make the reading a bit bumpy. [@Gelfand:1966uq] I. M. Gel’fand, M. I. Graev and N. Ya. Vilenkin, *Generalized Functions: Volume 5, Integral Geometry and Representation Theory*, Academic Press, 1966. > This book is the English translation of the Russian version, published in 1962. The chapters of interest for us are chapter III devoted to the representations of $SL_2(\mathbb{C})$ and chapter IV for its harmonic analysis. Loop Quantum Gravity and Spin-Foams {#loop-quantum-gravity-and-spin-foams .unnumbered} ----------------------------------- [@quantum_gravity] C. Rovelli, *Quantum Gravity*, Cambridge University Press, 2004. > This major textbook is recommended for its insistence on underlying physical ideas. The mathematical formulae are also present but some of the tools of representation and recoupling theories are assumed to be already known. [@CLQG] C. Rovelli and F. Vidotto, *Covariant Loop Quantum Gravity*, Cambridge University Press, 2014. > This book is a concise exposition of the covariant formulation of LQG, also known as the spin-foam formalism. It gathers all the main achievements of the theory. It can alternatively be used as a technical toolbox ready for use or as a general introduction that sketches the programme and the physical ideas upon which it relies. Nevertheless, the mathematics are not explained in detail (through lots of formulae are found) and it is sometimes a bit sloppy with the mathematical accuracy. [^1]: Aix Marseille Univ, Université de Toulon, CNRS, CPT, Marseille, France\ pmd@cpt.univ-mrs.fr [^2]: Here and everywhere else, Einstein notation is understood over repeated indices. [^3]: Hilbert spaces are of main interest for quantum physics. But many physicists may not be aware of the real representations of $SU(2)$, which are much less studied (see [@Itzkowitz1991] for more about them). [^4]: “Holomorphic representation” means that the map defined by the representation over the vector space is holomorphic. [^5]: As we will see later, $\mathfrak{sl}_2(\mathbb{C})$ can be seen both as a complex or as a real vector space. “$\mathbb{C}$-linear representations” means that we care about the complex structure of $\mathfrak{sl}_2(\mathbb{C})$. We will care about the $\mathbb{R}$-linear representations of $\mathfrak{sl}_2(\mathbb{C})$ in section \[finite irreps\]. [^6]: Two representations are equivalent if there exists a bijective intertwiner, i.e. a bijective map $T$ between the two vector spaces, so that it commutes with the action of the group (see appendix \[representations and intertwiners\]). It would be too heavy to write explicitly $\ket{jm} = T ( v_{j-m})$, so that we choose to write rather $\ket{jm} = T( v_{j-m})$. Similarly, for the operators we write $ J_+ \cong e$, rather than $ J_+ = T \circ e \circ T^{-1}$. \[note cong\] [^7]: See the footnote of page for the notation $\cong$. [^8]: It can be seen as the contraction with the “metric tensor” $\epsilon_{mm'} = (-1)^{j-m} \delta_{m,-m'}$, as it is often said. [^9]: For more details on the $9j$-symbol, see [@Edmonds1957] pp. 100-114. [^10]: For more details on the $15j$-symbol, see [@Yutsis1962] pp. 65-70. [^11]: The topology is defined by the following property of convergence: a sequence $F_n(z_0,z_1)$ is said to converge to $0$ if it converges to zero uniformly together with all its derivatives on any compact set in the $(z_0,z_1)$-p1ane which does not contain $(0,0)$ (see Vilenkin [@Gelfand:1966uq] p. 142). [^12]: Strictly speaking “LQG” refers to the canonical approach for which the spin-networks are embedded. Here, we adopt a more abstract point of view, sometimes called “covariant LQG”, which is motivated by spin-foams. This alternative construction raises difficulties for defining the hamiltonian, but they are circumvented by the spin-foam formalism. [^13]: Mathematicians usually say *edge* or *arrow*, but not “link”, which has another meaning in knot theory. The terminology of LQG keeps “edge” for spin-foams (see below), and uses “link” for spin-networks. [^14]: This designation comes from an analogy with Maxwell theory of electromagnetism. [^15]: This subspace is sometimes denoted $L^2 \left(SU(2)^L / SU(2)^N \right)$, but this is not mathematically rigorous. [^16]: There is a way to give a precise meaning to this gluing, but it will be sufficient to keep it intuitive below, and to avoid these technicalities. [^17]: The notion of boundary of an abstract $2$-complex requires a formal definition, but we keep it intuitive below for simplicity. We can admit that any $2$-complex comes with a boundary.

ArXiv

--- abstract: 'In the present work, electroencephalographic recordings of healthy human participants were performed to study the entrainment of brainwaves using a variety of stimulus. First, periodic entrainment of the brainwaves was studied using two different stimuli in the form of periodic auditory and visual signals. The entrainment with the periodic visual stimulation was consistently observed, whereas the auditory entrainment was inconclusive. Hence, a photic (Visual) stimulus, where two frequencies were presented to the subject simultaneously was used to further explore the bifrequency entrainment of human brainwaves. Subsequently, the evolution of brainwaves as a result of an aperiodic stimulation was explored, wherein an entrainment to the predetermined aperiodic pattern was observed. These results suggest that aperiodic entrainment could be used as a tool for guided modification of brainwaves. This could find possible applications in processes such as epilepsy suppression and biofeedback.' author: - Richa Phogat - 'P. Parmananda' title: Provoking Predetermined Aperiodic Patterns in Human Brainwaves --- [**Variations in the environment of a system can alter its dynamics. In this work, the phenomenon of the entrainment of human brainwaves to a predetermined aperiodic photic stimulus is presented. As a precursor, entrainment to a single frequency and a bi-frequency signal is studied. Subsequently, the presence of aperiodic entrainment was quantified using short-time Fourier transform (STFT) and by calculating the cross correlation between the STFT of the aperiodic signal and that of the observed EEG dynamics. This guided modification of brainwaves may find possible applications in suppressing some types of epilepsy and in biofeedback.**]{} Entrainment is the process of adjusting the dynamics of a system to that of an external rhythm. This is observed in a wide variety of natural as well as laboratory systems [@Parmananda; @Zlotnik; @Klein; @Freedman; @Stokkan; @Takahashi]. In mammals, entrainment of the circadian rhythms as a function of various factors such as illumination, body temperature, social cues and food availability is well documented in literature [@Klein; @Freedman; @Stokkan; @Takahashi]. Another interesting observation in this field is the phenomenon of brainwave entrainment [@Notbohm]. This phenomenon leads back to the initial experiments done to study the brain dynamics [@Adrian; @Walter], wherein flickering lights at different frequencies were used to study the modification of brainwaves in human as well as animal subjects. A recent interest has emerged in the entrainment of brainwaves using a variety of stimulation such as, audio-visual stimulation (AVS) or transcranial alternating current stimulation and its possible applications [@Aftanas; @Helfrich; @Teplan2; @Teplan3]. Research has also been carried out to study the individual effects of auditory [@Karino; @Neher; @Will] and photic [@Mori] entrainment of the brainwaves. Noise along with a subthreshold photic stimulus has previously been shown to enhance the periodicity in brainwaves via stochastic resonance [@Mori; @Benzi]. However, there have been contradicting reports regarding the effects of auditory stimulation on the brainwave entrainment [@Karino; @Neher; @Lopez]. The auditory stimulation is conventionally given in the form of binaural beats [@Will; @Teplan2; @Teplan3; @Karino; @Lopez] or repeating drum sounds [@Neher]. Similarly, the photic entrainment is studied using both colored [@Teplan2; @Teplan3] and white light [@Mori] LEDs flickering at a desired frequency. In the present work, entrainment is studied using white light LEDs to avoid the psychological effects of the coloured light [@Elliot], if any. Also, the auditory counterpart of white light i.e. white sound was switched on and off periodically to explore the auditory entrainment of brainwaves.\ The next step after studying single frequency entrainment would be bifrequency entrainment, wherein two rhythmic photic stimuli are presented to the subject simultaneously. Recently, in synthetic gene oscillators, entrainment was studied using aperiodic signals  [@Butzin]. Another interesting extension to be explored in this direction would be an aperiodic entrainment of human brainwaves. We have studied the effects of an aperiodic photic stimulation on the electrical activity of the brain. The organization of this paper is as follows. In part I, the protocol employed for the experiments is described. In part II, results for the periodic photic stimulation are presented. Bifrequency and aperiodic entrainment are reported in part III and IV respectively. A discussion on the results follows in part V. Experimental Protocol ===================== The experiments were performed on five healthy adults (26.12$\pm$1.86) who volunteered for the experiments. All participants were informed about the experimental protocol beforehand and the experiments were performed only after the participants signed the Informed Consent Form(ICF). The ICF was approved by the institute ethics committee of IIT Bombay. A set of 20 cerebral electrodes with the help of a adhesive electrode paste were used to record the EEG data at a sampling frequency of 256 Hz. The 10-20 electrode placement system was used for the positioning of the electrodes on the scalp [@Homan]. Four additional electrodes were used to capture the eye movement and heart beat. Fpz was used as the reference and nasion was grounded. The data recorded was first cleaned visually for the artefacts in the EEGlab [@Delorme] toolbox of MATLAB^^ and then analysed using in house MATLAB^^ codes.\ The photic stimulus used for the experiments is a square wave of required frequency distribution, created using MATLAB^^. This digital square signal was supplied using a digital to analogue converter (MCC USB-1616HS-4) to a set of 8 LEDs mounted on a board. When maintained in a state of constant illumination, without any flicker, these LEDs have an intensity of 1125$\pm$98 Lux at a distance of $\approx$ 78 cm from the LED board. This was the distance maintained between the LEDs and the eyes of the subject. Therefore, when the stimulation is provided, the LEDs flicker with the given frequency distribution between the maximum illumination of 1125$\pm$98 Lux and a minimum illumination of 0 Lux. The subjects were sitting comfortably in a chair with their eyes closed to ensure alpha range (8-13 Hz) as the baseline [@Adrian; @Dolce]. The experiments were performed in a dark space so that the photic stimulus provided was the only source of light. The protocol for all the experiments was as follows: 0-8 minutes : Relaxed state (Part I)\ 8-18 minutes : Stimulus applied (Part II)\ 18-26 minutes : Relaxed state (Part III)\ 26-36 minutes : Stimulus applied (Part IV)\ 36-44 minutes : Relaxed state (Part V) Periodic Photic Stimulation =========================== The light signal was provided using a set of 8 white light LEDs mounted on a board. The first set of experiments were performed with all these LEDs synchronously flickering at 10 Hz frequency. This value of frequency was chosen to check for the periodic entrainment of the brainwaves with a stimulation in the baseline frequency range (alpha range: 8-13 Hz). Scalp maps were used to ensure maximum entrainment in the occipital head region. The analysis of the results was then performed using the short-time Fourier transform (STFT) to see the evolution of brainwaves in both time as well as the frequency domain. The spectrogram function of MATLAB^^ was used for this purpose. The STFT was calculated using a Gaussian window of 6 s with an overlap of 5.96 s between consecutive windows. The EEG recordings for the Oz electrode were used for the analysis. This was done considering the symmetric location of this electrode between the left and right hemispheres, thus minimizing the effects of right or left handedness of the subject, if any. To quantify the entrainment observed, $\zeta$ is calculated and compared across various experimental conditions for Oz electrode. It gives a measure of the increment in the power of the Oz electrode when the photic stimulus is provided as compared to when no stimulus is given. $\zeta = \int_{f_1}^{f_2} \int_{t_1}^{t_2} P(f,t) df dt$ $P$ represents the power spectral density (PSD) in frequency $f$ at time $t$. Unless otherwise specified, the values of $f_1$ and $f_2$ were kept constant at 5 Hz and 40 Hz respectively. This was done to study only the fundamental and super-harmonic entrainment of brainwaves. This frequency range includes all the discernible changes at fundamental and harmonic frequency observed with the entrainment. First two minutes of the cleaned EEG data were used for the analysis across the subjects and for all the experiments. Hence, $t_1$ = 0 s and $t_2$ = 2 min. To study the effects of the stimulation on the amplitude of the brainwaves, the variance in the amplitude with and without stimulus was compared. The amplitude variance was calculated as follows: $Amplitude Variance=\frac{1}{N-1} \sum_{i=1}^{N} |A_i - \mu|^2$, Where N is the total number of data points in the time-series A and $\mu = \frac{1}{N} \sum_{i=1}^{N} A_i$ is the mean of A.\ The effect of the 10 Hz photic stimulation on the brainwaves calculated using the techniques mentioned above are presented in Figure 1 and Figure 2. Figure 1 shows the STFT for one of the subjects without (upper panel) and with (lower panel) the stimulus. An increment in the power across the fundamental frequency and the subsequent harmonics can be observed when the stimulus was provided as compared to when no stimulus was given. A discernible increment in $\zeta$ (left panel) and the amplitude variance (right panel) of the brainwaves when the stimulus is provided (Stimulus Status: $\bf{On}$) can be seen from Figure 2. Sub-harmonic entrainment of the brainwaves at this frequency was also studied by filtering the data in the frequency range of 4.9-5.1 Hz (sub-harmonic for 10 Hz is at 5 Hz) and studying both $\zeta$ and amplitude variance as a function of stimulus status. A persistent increment in both the quantities when the stimulus is provided was observed across all the subjects (results in supplementary material).\ ![STFT of the brainwaves (Oz electrode) for the first 2 minutes of without and with 10 Hz photic stimulation. An increment in power at 10 Hz and the subsequent harmonics (20 and 30 Hz) can be observed in the sub-plot where the stimulus is provided (lower panel).[]{data-label="1"}](fig1){width="9cm"} ![Evolution of the quantities $\zeta$ and Amplitude Variance as a function of stimulus status for all the subjects. A universal increment for both these quantities when the stimulus is provided (Stimulus Status: $\bf{On}$) is evident from the box plots in both the sub-figures.[]{data-label="2"}](fig2){width="9cm"} Upon observing entrainment using a 10 Hz photic stimulation, the effects of a 6 Hz photic stimulation on the brainwaves were explored. This was done to study brainwave entrainment when the entrainment frequency is relatively farther from the baseline frequency range. As shown in Figure 3, entrainment at fundamental frequency and its subsequent harmonics can be observed when the stimulus is provided (lower panel). However, the entrainment observed at 6 Hz is weaker as compared to the entrainment observed at 10 Hz. This is evident by lesser percentage increment in $\zeta$ and amplitude variance in Figure 4 as compared to Figure 2. Also, the sub-harmonic entrainment for 6 Hz calculated by filtering the brainwaves in the frequency range of 2.9 - 3.1 Hz (sub-harmonic for 6 Hz at 3 Hz), was not observed across all the subjects (results in supplementary material).\ ![STFT of the brainwaves (Oz electrode) for the first 2 minutes of without and with 6 Hz photic stimulation. An increment in power at 6 Hz and the subsequent harmonics (12, 18, 24 , 30 and 36 Hz) is detected in the sub-plot with the stimulus.[]{data-label="3"}](fig3){width="9cm"} ![Evolution of the quantities $\zeta$ and Amplitude Variance as a function of stimulus status for 6 Hz photic stimulus across all the subjects. A consistent increment in both these quantities when the stimulus is provided (Stimulus Status: $\bf{On}$) is indicated by the box plots in both the figures.[]{data-label="4"}](fig4){width="9cm"} After studying photic entrainment, the next step was to study the effects of auditory stimulation on the brainwaves. For this purpose, an auditory analogue of white light i.e. white sound was used. This sound consists of the frequencies in the auditory range i.e. 22-22000 Hz distributed uniformly. The required auditory stimulation signal was then consturcted by oscillating between no sound and white sound at the desired frequencies (6 and 10 Hz). However, a consistent increment across all the subjects was not observed using this form of auditory stimulation. Amongst the subjects that did show entrainment, a significantly smaller increment in power was observed as compared to the photic stimulation (a detailed comparison for the same is presented in supplementary material). The results for auditory stimulation in this case are in agreement with those reported in [@Lopez] for binaural beat entrainment. Bifrequency Entrainment ======================= Since a single frequency stimulus was able to entrain the brainwaves, two frequencies were provided simultaneously and the evolution of brainwaves was analysed. For this purpose, half the LEDs were flickering at 6 Hz and the other half at 10 Hz. An entrainment to both the frequencies as well as their subsequent harmonics as shown in Figure 5 was observed. Also, an entrainment to the sum of the two different frequencies (6+10 = 16 Hz) was observed. Figure 6 shows the consistency of this pattern across all the subjects for both $\zeta$ and amplitude variance. Bifrequency entrainment was further explored by supplying one frequency to each eye. A persistence of entrainment at the harmonics and summation of the two frequencies was observed. This indicates to the integration of the information received by each eye in the visual pathways of the brain. Since sub-harmonic entrainment was observed using 10 Hz photic stimulation and not with 6 Hz photic stimulation, it was not pursued for the analysis in the subsequent set of experiments. As seen in Figure 5, the bifrequency photic stimulation was able to simultaneously provoke a range of frequencies in the brainwaves. Subsequently, the time required to entrain the brainwaves was investigated. In [@Notbohm], it was observed that only a periodic stimulus provoked a higher strength of entrainment demonstrated by a better phase locking between the forcing and the brainwaves. This motivated us to find the minimum duration of periodic signal required for entrainment. This was checked by alternating the forcing frequency between 6 and 10 Hz. The time interval of exposure to one frequency (6 Hz) before shifting to the next frequency (10 Hz) was monotonically decreased. It was observed that the brainwave entrainment successfully shifts from one frequency value (6 Hz) to the next frequency value (10 Hz) and vice versa with the duration of exposure to each frequency being as low as 1 s. A further decrement in frequency exposure time below 1 s was not possible because of the resolution limitations posed by the sampling frequency of the instrument. ![STFT of the brainwaves (Oz electrode) for the first 2 minutes of without and with bifrequency (simultaneous 6 and 10 Hz) photic stimulation. A discernible increment in power at the fundamental frequencies (6 and 10 Hz) and the subsequent harmonics (12, 18, 20, 24, 30 and 36 Hz) can be observed from the sub-plot with the stimulus (lower panel). Also entrainment at the sum of the fundamental frequencies (6 + 10 = 16 Hz) is observed.[]{data-label="5"}](fig5){width="9cm"} ![This figure shows the variations in $\zeta$ and Amplitude Variance as the bifrequency (simultaneous 6 and 10 Hz) photic stimulation was used for entrainment. A consistent increment for both these quantities when the stimulus is provided (Stimulus Status: $\bf{On}$) can be seen from the box plots in both the figures.[]{data-label="6"}](fig6){width="9cm"} Aperiodic Entrainment ===================== As mentioned in the previous section, entrainment to a single frequency was possible by an exposure to that frequency for one second. Also, the entrainment shifts to the next frequency as the stimulus updates to the subsequent frequency. Hence, an aperiodic signal was made with a uniform random distribution of frequencies in the range of 5-15 Hz. The frequency of the signal kept changing randomly from one value to the next every second. This signal was fed to the LEDs and its effects on the brainwaves evaluated. As shown in Figure 7, unlike previous forms of entrainment, no consistent entrainment at a single frequency is observed. This is because the entrainment state is shifting every second due to the change in frequency values every second. In Figure 8, the increment in $\zeta$ and amplitude variance as the stimulus is provided (Stimulus Status: $\bf{On}$) persists as a consequence of entrainment. The response provoked in the brainwaves by the sequence of frequencies provided by the light signal was visually inspected. A local increment at the instantaneous stimulation frequency was observed in the brainwaves. As the stimulus changes to the next frequency, the brainwave entrainment was also found to shift towards this next frequency in the sequence. To further check if the brainwaves were following the predetermined aperiodic signal supplied through the LEDs, information transfer between the light signal and the brainwaves was quantified. In the data cleaned for this purpose, the previously rejected epochs of noisy data were replaced with zeros. This was done to maintain a uniform length of the time series. $P_k(f_i,t_i)$ represents the PSD in the frequency range of $f_i$ to $f_i + 1$ Hz and a time window of $t_i$ to $t_i + 1$ s. The time windows in this case are non-overlapping. The subscript $k$ indicates to the five parts of the experiment as mentioned in the protocol. To reiterate, during part 2 and 4, the photic stimulation was provided and part 1, 3 and 5 are without stimulation. $P_s(f_i,t_i)$ denotes the PSD for the light signal. PSD from part 1, 3 and 5 was used for defining a threshold $(th)$ which is employed as an indicator of power in a frequency band without the stimulus. The PSD was then modified as follows:\ $$P_k'(f_i,t_i) = \begin{cases} P_k(f_i,t_i) \hspace{1cm} \forall \hspace{0.5cm} P_k(f_i,t_i) > th \\ 0 \hspace{2.2cm} \forall \hspace{0.5cm} P_k(f_i,t_i) < th \end{cases}$$ $C_k$ is a measure of correlation between the STFT of the brainwaves and the light signal. It is defined as follows:\ $$C_k = \frac{\sum_{f_i = 5}^{15}\sum_{t_i = 0}^{400}P_k'(f_i,t_i)P_s(f_i,t_i)}{N_k}$$ $P_s(f_i,t_i)$ and $P_k'(f_i,t_i)$ have been defined previously. First 400 s of the data was used for calculating the information transfer in the fundamental frequency range (5-15 Hz). For this analysis, $f_i$ and $t_i$ increase in steps of 1 Hz and 1 s respectively. The normalizing constant $N_k$ denotes the number of non zero data points in the brain data for the corresponding part of the experiment. In Figure 9, an increment/decrement in $C_k$ as a function of stimulus status ($\bf{On}$/$\bf{Off}$) can be observed. To confirm $C_k$ as a reliable measure of correlation between the given aperiodic signal and the brainwaves, a comparison of the value of $C_k$ obtained using the original aperiodic signal to that of the surrogate signals was performed in Figure 10. The blue plot shows the correlation measure $C_k$ between the original aperiodic signal and the brainwaves. The other five plots measure $C_k$ between the surrogate aperiodic signals and the brainwaves. The value of $C_k$ was found to be considerably higher for the original signal when the stimulus is provided (Stimulus Status: $\bf{On}$).\ ![STFT of the brainwaves (Oz electrode) for the first 2 minutes of without and with predetermined aperiodic photic stimulation. An increment in power at the fundamental frequency band (5 - 15 Hz) and the harmonic band (10 - 40 Hz) can be observed from the sub-plot with the stimulus (lower panel).[]{data-label="7"}](fig7){width="9cm"} ![Evolution of the quantities $\zeta$ and Amplitude Variance as a function of stimulus status for a predetermined aperiodic photic stimulus across all the subjects. A consistent increment for both these quantities when the stimulus is provided (Stimulus Status: $\bf{On}$) is evident from the box plots in both the figures.[]{data-label="8"}](fig8){width="9cm"} ![Correlation measure ($C_k$) as a function of stimulus status. An increment in $C_k$ when the stimulus is provided (Stimulus Status: $\bf{On}$) can be seen from the figure. $C_k$ was found to be robust against the surrogate signals.[]{data-label="9"}](fig9){width="9cm"} ![Comparison of correlation measure ($C_k$) obtained with the surrogate signals to that obtained with the original signal. The blue plot shows $C_k$ values for the original signal. The other five plots are the $C_k$ values for the surrogate data. It can be seen from the graph that $C_k$ values for the original signal when the stimulus is provided (Stimulus Status: $\bf{On}$) are considerably higher than those for the surrogate data.[]{data-label="10"}](fig10){width="9cm"} Discussion ========== In the results presented above, entrainment is found to persist with different types of stimulus including periodic, bifrequency and aperiodic photic stimulation. The STFT plots across various forms of stimulus show entrainment for one subject while the box plots represent the robustness of the phenomenon against subject variability. The decrement in $\zeta$ and amplitude variance after the stimulus is removed is a manifestation of the underlying information processing area. The effect of stimulation dies down after the system stops receiving the external information. The increment in amplitude variance of the EEG signal when the stimulus is provided is an indicator of increased neuronal firing in the corresponding region of the brain. This can be loosely compared to a stimulus provoked increased blood flow in a specific brain region observed using various neuroimaging techniques. In periodic entrainment, as shown using the power spectral density (PSD) in the supplementary material (Supplementary Figure 5 and 6), the increment in power is relatively higher with the forcing frequency of 10 Hz as compared to the 6 Hz stimulation. This is in agreement with [@Notbohm], where maximum strength of entrainment for a specified light intensity is observed when the frequency mismatch between the dominant alpha frequency and the forcing frequency is minimum. Therefore, the absence of sub-harmonic entrainment in case of 6 Hz stimulation can be explained by the relative weakness of entrainment observed with 6 Hz forcing frequency. In bifrequency entrainment, a simultaneous increment in power at the fundamental, harmonic and summation of frequencies was observed. Aperiodic entrainment, in our opinion, is of special interest as it might have wide applicability. One possible application could be in the field of biofeedback. The brainwaves of a healthy individual during various stages could be recorded and fed back to them as a visual stimulus. In certain pathological conditions such as some forms of epilepsy (Petit Mal Seizures), low frequency rhythmic activity in the brainwaves is observed. An aperiodic pattern lying outside this rhythmic activity and centred near the baseline state of the subject might lead to the modification of the brainwaves to the desired state. An aperiodic entrainment of circadian rhythms may also be looked into. Supplementary Material ====================== In the supplementary material, the results of the sub-harmonic entrainment for the two periodic photic stimulations are reported. A detailed comparison between the periodic auditory and periodic visual entrainment using white sound and white light stimulus respectively is also presented in the supplementary material. Acknowledgements ================ The authors would like to thank DST (India), (Project ref no. EMR/2016/000275) for financial assistance. Richa Phogat would also like to acknowledge CSIR (India) for financial assistance. We would also like to acknowledge Guru Vamsi Policharla, Kanishk Chauhan, Keshav Srinivasan and all the past and present members of the NLD lab group of IIT-Bombay for help during various stages of the study. [99]{} P. Parmananda, M. Rivera, & R. Madrigal, Electrochimica acta $\bf{44}$, 4677-4683 (1999). A. Zlotnik, R. Nagao, I. Z. Kiss, & J. S. Li, Nature communications $\bf{7}$, 10788 (2016). D. C.Klein, R. Y. Moore, & S. M. Reppert, (Eds.), Oxford University Press, USA (1991). M. S. Freedman, R. J. Lucas, B. Soni, M. von Schantz, M. Muñoz, Z. David-Gray, & R. Foster, Science $\bf{284}$, 502-504 (1999). K. A. Stokkan, S. Yamazaki, H. Tei, Y. Sakaki, & M. Menaker, Science $\bf{291}$, 490-493 (2001). K. Takahashi, & T. Deguchi, Physiology & behavior $\bf{31}$, 373-378 (1983). E. D. Adrian, & B. H. Matthews, Brain $\bf{57}$, 355-385 (1934). W. G. Walter, V. J. Dovey, & H. Shipton, Nature $\bf{158}$, 540-541 (1946). L. I. Aftanas, P. V. Miroshnikova, N. B. Morozova, S. V. Yarosh, O. M. Gilinskaya, & G. R. Khazankin, Frontiers in Human Neuroscience Conference Abstract: SAN2016 Meeting (2016). R. F. Helfrich, T. R. Schneider, S. Rach, S. A. Trautmann-Lengsfeld, A. K. Engel, & C. S. Herrmann, Current Biology $\bf{24(3)}$, 333-339 (2014). M. Teplan, A. Krakovská, & S. Štolc, Computer methods and programs in biomedicine $\bf{102}$, 17-24 (2011). M. Teplan, A. Krakovská, & S. Štolc, International journal of psychophysiology $\bf{59}$ 81-90 (2006). U. Will, & E. Berg, Neuroscience Letters $\bf{424}$, 55-60 (2007). S. Karino, M. Yumoto, K. Itoh, A. Uno, K. Yamakawa, S. Sekimoto, & K. Kaga, Journal of neurophysiology $\bf{96}$, 1927-1938 (2006). A. Neher, Electroencephalography and clinical neurophysiology $\bf{13}$, 449-451 (1961). T. Mori, & S. Kai, Phys. Rev. Lett. $\bf{88}$, 218101 (2002). R. Benzi, A. Sutera, & A. Vulpiani, Journal of Physics A: mathematical and general [**14**]{} L453 (1981). F. López-Caballero, & C. Escera, Frontiers in human neuroscience $\bf{11}$, 557 (2017). A. J. Elliot, & M. A. Maier, Annual review of psychology $\bf{65}$, 95-120 (2014). N. C. Butzin, P. Hochendoner, C. T. Ogle, P. Hill, & W. H. Mather, ACS synthetic biology $\bf{5(2)}$, 146-153 (2015). R. W. Homan, J. Herman, & P. Purdy, Electroencephalography and clinical neurophysiology $\bf{66}$, 376-382 (1987). A. Delorme, & S. Makeig, Journal of Neuroscience Methods $\bf{139}$, 9-21 (2004). G. Dolce, & H. Waldeier, Electroencephalography and Clinical Neurophysiology, $\bf{36}$, 577-584 (1974). A. Notbohm, J. Kurths, & C. S. Herrmann, Frontiers in human neuroscience $\bf{10}$, 10 (2016).

ArXiv

--- abstract: | Although it has recently been proved that the packing chromatic number is unbounded on the class of subcubic graphs, there exists subclasses in which the packing chromatic number is finite (and small). These subclasses include subcubic trees, base-3 Sierpiński graphs and hexagonal lattices. In this paper we are interested in the packing chromatic number of subcubic outerplanar graphs. We provide asymptotic bounds depending on structural properties of the outerplanar graphs and determine sharper bounds for some classes of subcubic outerplanar graphs. [**Keywords:**]{} packing colouring, packing chromatic number, outerplanar graphs, subcubic graphs. [**AMS Subject Classification: 05C12, 05C15, 05C70**]{} author: - 'Nicolas Gastineau$^{1}$, Přemysl Holub$^{2}$ and Olivier Togni$^{3}$' title: On the packing chromatic number of subcubic outerplanar graphs --- Introduction ============ Throughout this paper, we consider undirected simple graphs only, and for definitions and notations not defined here we refer to [@bondy_murty]. Let $G$ be a graph and $c$ a vertex $k$-colouring of $G$, i.e., a mapping $c: V(G) \rightarrow \{1,2,\dots, k\}$. We say that $c$ is a [*packing $k$-colouring*]{} of $G$ if vertices coloured with the same colour $i$ have pairwise distance greater than $i$. The [*packing chromatic number*]{} of $G$, denoted by $\chi_{\rho}(G)$ is the smallest integer $k$ such that $G$ has a packing $k$-colouring; if there is no such integer $k$ then we set $ \chi_{\rho}(G)= \infty$. For a class of graphs $\mathcal{C}$, we say that the packing chromatic number of $\mathcal{C}$ is finite if there exists a positive integer $k$ such that $\chi_{\rho}(G)\leq k$ for every graph $G\in \mathcal{C}$. The concept of a packing colouring of a graph, introduced by Goddard et al. in [@God] under the name broadcast colouring, is inspired by frequency planning in wireless systems, in which it emphasizes the fact that signals can have different powers, providing a model for the frequency assignment problem. The packing chromatic number of lattices has been studied by several authors: for the infinite square lattice $\mathbb{Z}^{2}$, Soukal and Holub in [@SO2010] proved that $\chi_{\rho}(\mathbb{Z}^{2})\le 17$, while Ekstein et al. in [@EK2010] showed that $12\le\chi_{\rho}(\mathbb{Z}^{2})$. Recently, Martin et al. in [@Martsquare] improve the bounds by showing that $13\le \chi_{\rho}(\mathbb{Z}^{2}) \le 15$. For the infinite hexagonal grid $\mathscr{H}$, Fiala et al. in [@FI2009] showed that $\chi_{\rho}(\mathscr{H})\le7$, Kor[ž]{}e and Vesel in [@AV2007] proved that $\chi_{\rho}(\mathscr{H})\ge7$. Finbow and Rall in [@Fin] proved that the infinite triangular grid $\mathscr{T}$ is not packing colourable, i.e., $\chi_{\rho}(\mathscr{T})=\infty$. Moreover, Kor[ž]{}e and Vesel in [@AV2018] proved that the infinite octagonal lattice has packing chromatic number $7$. The packing chromatic number of the Cartesian product of some graphs was investigated in [@BrePa; @FI2009; @jonck]. Also, the packing chromatic number has been studied for further graph classes in [@BrePa; @BresKlav; @God; @HolubD; @TO2010]. The computational complexity has been also studied: determining whether a graph has packing chromatic number at most $4$ is an NP-complete problem [@God] and determining whether a tree has packing chromatic number at most $k$ (with a tree and $k$ on input) is also an NP-complete problem [@FiCo]. Sloper in [@Slo] showed that the infinite complete ternary tree $T$ has $\chi_{\rho}(T)=\infty$ while any subcubic tree $T$ is packing $7$-colourable, hence it is natural to ask if all graphs with maximum degree $3$ (often so-called [*subcubic graphs*]{}) have finite packing chromatic number. This question was raised by Goddard et al. [@God]. Recently, a second open question has been proposed about the packing chromatic number of $S(G)$, when $G$ is subcubic [@BresKlav3; @Gasto] ($S(G)$ being the graph obtained from $G$ by subdividing each edge once). Recently, Balogh, Kostochka and Liu [@Bal17] proved that, for any integer $k$, almost all cubic graphs of order $n$ and of girth at least $2k+2$ have packing chromatic number greater than $k$, hence answering negatively the question of Goddard et al. Moreover, an explicit construction of an infinite family of subcubic graphs with unbounded packing chromatic number have been found very recently [@BreFe]. Some subclasses of subcubic graphs were also under consideration, see e.g. [@BrePa; @BresKlav]. Outerplanar graphs form a class of structured graphs (containing the class of trees), which are generally easy to colour. Our aim is to find some classes of subcubic outerplanar graphs, which have finite packing chromatic number. We define these classes by giving restrictions on their structure (number of faces of different types), or, equivalently, on their weak dual. Note that, when a graph is not connected, we can colour each component separately satisfying the distance constraints of a packing colouring and the resulting colouring is packing as well. Thus, throughout the rest of this paper, we will consider connected outerplanar graphs only. The paper is organized as follows. Section 2 presents an upper bound for $2$-connected subcubic outerplanar graphs without internal face, i.e., for which the weak dual is a path. Then, in Section 3, we use results from Section 2 in order to determine asymptotic bounds for some larger classes of subcubic outerplanar graphs restricted by the number of (internal) faces. In Section 4, we improve bounds from Section 3 for some specific classes of subcubic outerplanar graphs with a specific structure. Finally, in the last section, we present lower bounds for the packing chromatic number of subcubic outerplanar graphs and give concluding remarks. Table \[table1\] summarizes the main results of this paper. Condition on the subcubic outerplanar graph $G$ $\ell$ Section ------------------------------------------------------------------------ ---------------------- --------- $G$ is $2$-connected with no internal face $15$ 2 $G$ is $2$-connected with at most $k$ internal faces $17 \times 6^{3k}-2$ 3 $G$ is connected with at most $k'$ faces $ 9 \times 6^{k'}-2$ 3 $G$ is $2$-connected with one internal face 51 4 $G$ is connected with no internal face and with the block graph a path 305 4 : Classes of subcubic outerplanar graphs and values of $\ell$ for which every relevant graph $G$ satisfies $\chi_{\rho}(G)\le \ell$.[]{data-label="table1"} Preliminaries ------------- Let $G$ be a graph and $A\subset V(G)$. We denote $G-A$ the subgraph of $G$ after deletion of all vertices of $A$ from $G$ and all edges incident to some vertex of $A$ in $G$. We further denote $G[A]$ the subgraph of $G$ induced by $A$, or equivalently, $G[A]=G-(V(G)\setminus A)$. Specifically, for $x\in V(G)$, $G-x$ denotes the subgraph of $G$ after deletion of $x$ and all edges incident to $x$ from $G$. An [*outerplanar graph*]{} $G$ is a planar graph such that there exists a planar drawing of $G$ for which all vertices belong to the outer face. When it is $2$-connected, it can be represented by a [*boundary cycle*]{} $C$ containing all vertices of $G$, with non-crossing chords dividing the interior of $C$ into [*faces*]{}. A face $F$ of $G$ is called an [*internal face*]{} if $F$ contains more than two chords of $G$, and an [*end face*]{} of $G$ if $F$ contains only one chord of $G$; note that all remaining edges of an end face belong to $C$. The [*weak dual*]{} of $G$, denoted by $\mathcal{T}_G$, is the graph with the set of all faces of $G$, except the outer face, as vertex set, and the edge set $E(G)=\left\{ FF' \vert \, F \mbox{ and }F' \mbox{ have an edge in common}\right\}$. We denote by $u_F$ the vertex of $\mathcal{T}_G$ corresponding to the face $F$ of $G$ and sometimes we identify a face $F$ and the corresponding vertex $u_F$ of $\mathcal{T}_G$. It is well known that the weak dual of a connected outerplanar graph is a forest and of a $2$-connected outerplanar graph is a tree. Note that an end face of an outerplanar graph $G$ corresponds to a leaf of $\mathcal{T}_G$ and that an internal face of $G$ corresponds to a vertex of degree at least $3$ in $\mathcal{T}_G$. Obviously, every end face of a $2$-connected outerplanar graph contains at least one vertex of degree 2. For a graph $G$, the [*block graph*]{} of $G$, denoted by $\mathcal{B}_G$, is the graph where vertices of $\mathcal{B}_G$ represent all maximal 2-connected subgraphs of $G$ (usually called [*blocks*]{}) and two vertices of $\mathcal{B}_G$ are adjacent whenever the corresponding blocks share a cut vertex. For any $G_1\subset G$, let $N(G_1)=\{u\in V(G)| \ uv \in E(G) \mbox{ for some } v\in V(G_1)\}$ be the [*neighbourhood*]{} of $G_1$ in $G$. Specifically, if $G_1=\{v\}$, let $N(v)$ denote the neighbourhood of $v$ in $G$. For $X,Y\subseteq V(G)$, a [*shortest ($X$,$Y$)-path*]{} is a shortest path in $G$ between some vertex of $X$ and some vertex of $Y$. If $X$ contains a vertex $u$ only, then we write $u$ instead of $\{u\}$. Let $d_G(u,v)$ denote the [*distance between two vertices*]{} $u$ and $v$ in $G$, i.e., the length of a shortest $(u,v)$-path. Analogously, $d_G(X,Y)$ denote the [*distance between $X$ and $Y$*]{}, i.e., the length of a shortest $(X,Y)$-path in $G$. Let $P_{\infty}$ denote the two-way infinite path, i.e., $V(P_{\infty})=\mathbb{Z}$ and $E(P_{\infty})=\{i\ i+1|\ i\in\mathbb{Z}\}$ and let $P^{+}_{\infty}$ denote the one-way infinite path, i.e., $V(P^{+}_{\infty})=\mathbb{N}$ and $E(P^{+}_{\infty})=\{i\ i+1|\ i\in\mathbb{N}\}$. In our proofs we will use the following statement presented by Goddard et al. in [@God]. [[@God]]{} \[preliminaries\] Let $k$ be a positive integer. Then 1. Every cycle has packing chromatic number at most $4$; 2. There is a packing colouring of $P_{\infty}$ with colours $\{k,k+1,\dots, \, 3k+2\}$; 3. If $k\ge 34$, then there is a packing colouring of $P_{\infty}$ with colours $\{k,k+1,\dots, \, 3k-1\}$. $2$-connected subcubic outerplanar graphs with the weak dual a path =================================================================== In the different proofs of this paper, we say that we denote the vertices of a path of order $n$ by $x^1,x^2,\ldots,x^n$ in an ordering starting by $x$ and finishing by $y$, in the case $x^i$ and $x^{i+1}$ are adjacent, for $1\le i\le n-1$, $x^i$ and $x$ denote the same vertex and $x^n$ and $y$ denote the same vertex. The following observation will be used in order to construct some useful shortest path in $2$-connected subcubic outerplanar graphs. Moreover, it gives a description of $2$-connected outerplanar graphs with the weak dual of these graphs. Let $G$ be a $2$-connected outerplanar graph that is not a cycle. Then $G$ contains at least two end faces. Moreover, if $G$ contains no internal face, then $G$ has exactly two end faces. Considering the weak dual, $\mathcal{T}_G$ is a tree by connectedness of $G$. Since every nontrivial tree has at least two leaves and each leaf of $\mathcal{T}_G$ corresponds to some end face of $G$, each $2$-connected outerplanar graph that is not a cycle contains at least two end faces. In particular, if $G$ has no internal face, then $\mathcal{T}_G$ is a path (the converse also holds), implying that $G$ has exactly two end faces. We begin this section with the following lemma. This lemma will be used in Sections 3 and 4. \[lemma\] Let $G$ be a $2$-connected subcubic outerplanar graph with no internal face. Let $x,y$ be a pair of vertices of degree 2 in $G$ such that $x$ belongs to one of the end faces of $G$ and $y$ to the other one, and let $P$ be a shortest $x,y$-path in $G$. Then there exists a packing colouring of $G$ such that the vertices of $V(G)\setminus V(P)$ are coloured with colours from $\{1,2,3,4\}$. Let $C$ denote the boundary cycle of $G$. If there is no chord in $G$, then $G$ is a cycle and, by Proposition \[preliminaries\].i), $\chi_{\rho}(G)\leq 4$. Thus we may assume that $C$ contains some chords in $G$. Note that $C-P$ is not necessarily connected, but, since there is no internal face, each component of $C-P$ is an induced path of $G$. Let $D_i$, $i=1,\dots, \,k$, denote the components of $C-P$ in an ordering from $x$ to $y$ (i.e., for $i<j$, $D_i$ has a neighbour in $P$ that is closest to $x$ than any neighbour of $D_j$ in $P$), and let $l_i$ denote the length of $D_i$. We further denote the vertices of each $D_i$ by $x_i^1, x_i^2, \dots, x_i^{l_i}$ in an ordering starting from a vertex of $D_i$ which is closest to $x$ in $D_i$. The described structure is shown in Fig. \[figL5\], where the thick $x,y$-path depicts $P$. units &lt;1mm,0.7mm&gt; x from 0 to 155, y from 0 to 55 axes ratio 1:1 180 degrees from 25 45 center at 25 25 axes ratio 1:1 -90 degrees from 25 4.8 center at 25 25 axes ratio 1:1 -90 degrees from 25 5.2 center at 25 25 axes ratio 1:1 180 degrees from 130 5 center at 130 25 axes ratio 1:1 90 degrees from 130 5.2 center at 130 25 axes ratio 1:1 90 degrees from 130 4.8 center at 130 25 25 45 130 45 / 25 5 130 5 / 25 4.8 35.1 4.8 50.1 44.8 60 44.8 59.85 45 59.85 5 60 4.8 80.1 4.8 95.1 44.8 104.9 44.8 119.9 4.8 130 4.8 / 25 5.2 34.9 5.2 49.9 45.2 60 45.2 60.15 45 60.15 5 60 5.2 79.9 5.2 94.9 45.2 105.1 45.2 120.1 5.2 130 5.2 / 25 5 25 45 / 35 5 50 45 / 60 5 60 45 / 70 5 70 45 / 80 5 95 45 / 120 5 105 45 / 125 5 120 45 / at 8 25 at 148 25 at 11 25 at 144 25 at 25 45 at 30 45 at 35 45 at 40 45 at 45 45 at 50 45 at 55 45 at 60 45 at 65 45 at 70 45 at 75 45 at 80 45 at 85 45 at 90 45 at 95 45 at 100 45 at 105 45 at 110 45 at 115 45 at 120 45 at 125 45 at 130 45 at 25 5 at 30 5 at 35 5 at 40 5 at 45 5 at 50 5 at 55 5 at 60 5 at 65 5 at 70 5 at 75 5 at 80 5 at 85 5 at 90 5 at 95 5 at 100 5 at 105 5 at 110 5 at 115 5 at 120 5 at 125 5 at 130 5 at 25 50 at 30 50 at 35 50 at 40 50 at 45 50 at 40 0 at 45 0 at 50 0 at 55 0 at 65 50 at 70 50 at 75 50 at 90 50 at 85 0 at 90 0 at 115 0 at 110 50 at 115 50 at 130 50 We colour each component $D_i$ of $C-P$ with a pattern $1,2,1,3$ starting from $x_i^1$ ($i=1,2,\dots, k$), i.e., for each odd $j$, $x_i^j$ is coloured with colour $1$, for each $j$ divisible by $4$, $x_i^j$ obtains colour $3$, and, for each even $j$ not divisible by $4$, we colour vertex $x_i^j$ with colour $2$, $j=1,2,\dots, l_i$. We denote by $\chi$ the defined colouring. Since $P$ is a shortest path and $G$ is subcubic, we are going to show that there is no collision between any pair of vertices coloured with colour $1$ or $2$, respectively. Suppose to the contrary that there is a pair of clashing vertices $a$ and $b$ coloured with colour $1$. Clearly $a$ and $b$ belong to the same component $D_i$ of $C-P$ and $ab\in E(G)\setminus E(C)$ by the definition of $\chi$. But then we get a contradiction with the fact that $D_i$ is an induced path of $G$. Now suppose that there is a pair of clashing vertices $a$ and $b$ coloured with colour $2$, i.e., $d_G(a,b)\leq 2$. Again, $a$ and $b$ must belong to the same component $D_i$ of $C-P$, otherwise $d_G(a,b)>2$ and we obtain a contradiction. Analogously as for colour $1$ we can show that $ab\not\in E(G)$. Thus $a$ and $b$ must have a common neighbour $c$ in $G$. From the definition of $\chi$ and also because $D_i$ is an induced path in $G$, $c\not\in V(D_i)$, and $c\not\in P$ since $G$ is subcubic, a contradiction with the existence of $c$. Therefore the only possible collision in the defined colouring $\chi$ could be between vertices coloured with colour $3$. Analogously as for colours $1$ and $2$, any pair of clashing vertices $a$ and $b$ cannot be at distance one or two apart. Therefore any such collision happens for $a$ and $b$ with $d_G(a,b)=3$. We will check and modify collisions in the defined colouring $\chi$ of the components $D_1,D_2,\dots, D_k$ of $C-P$ one-by-one starting from $D_1$ and from the vertex $x_i^1$ in each $D_i$. Note that, in each step of the modification process, we check the modified colouring, not the original one. The following possible collisions can occur: $a$ and $b$ belong to different components $D_i$ and $D_j$ of $C-P$, $i,j\in\{ 1,2,\dots, k \}$. Since $d_G(a,b)=3$, (up to a symmetry) $a=x_i^{l_i}$, $b$ belongs to the chord of $D_j\cup P$ which is closest to $a$, and $j=i+1$, otherwise we get $d_G(a,b)>3$. Then we can modify the colouring $\chi$ of $D_j$ by recolouring vertices $x_j^s$, $s=3,\dots, l_j-1$, with $\chi(x_j^s):= \chi(x_j^{s+1})$ and we set $\chi(x_j^{l_j})\in \{1,2,3\}$ depending on the continuation of the pattern $1,2,1,3$ in $D_j$. $a$ and $b$ belong to the same component $D_i$. Since $d_G(a,b)=3$, $a$ and $b$ must belong to consecutive chords of $P\cup D_i$ and there is no vertex between these chords on $P$. We call such a pair of vertices coloured with colour $3$ a [*critical pair*]{}. Consider a critical pair $a$ and $b$ such that $a=x_i^m$, $b=x_i^n$, $m<n$, and that there is no critical pair $a'$ and $a$ with $a'=x_i^o$, $o<m$. Then we modify the colouring of the vertices of $D_i$ starting at vertex $a$ by $\underline{3},1,4,1,\underline{2},1,3,1,2,\dots$ instead of $\underline{3},1,2,1,\underline{3},1,2,1,3 \dots$, i.e., we recolour the vertex $x_i^{m+2}$ with colour $4$ and we switch colours $2$ and $3$ of the vertices $x_i^j$ for even $j>m+2$. Note that the underlined colours represent the critical pair $a$ and $b$. It is easy to verify that vertices coloured with colour $4$ are mutually at distance more than $4$ apart, implying that there is no collision between any pair of vertices coloured with colour $4$. After these modifications we obtain a colouring of all the vertices of $G-P$ with colours $\{1,2,3,4\}$ satisfying the distance constraints of a packing colouring. \[out1\] If $G$ is a $2$-connected subcubic outerplanar graph with no internal face, then $\chi_{\rho}(G)\leq 15$. Let $x,y$ be any pair of vertices of degree 2 in $G$ belonging to distinct end faces of $G$. Let $P$ be a shortest $x,y$-path in $G$. By Lemma \[lemma\], the vertices of $V(G)\setminus V(P)$ can be coloured with colours from $\{1,2,3,4\}$. Then the colouring can be completed in a packing 15-colouring of $G$ by colouring the vertices along the path $P$ starting at $x$ and using a packing colouring of the infinite path (since $P$ is a shortest path in $G$, then the distance between any pair of vertices of $P$ is the same on $P$ and on $G$). For this, we repeat the following pattern with colours from $\{5,\ldots,15\}$ of length 36 along the vertices of $P$ starting at $x$: $$\label{pat5-15} 5,6,7,9,13,12,5,8,6,10,7,11,5,9,14,6,8,15,5 ,7,13,10,6,11,5,8,9,7,12,6,5,14,10,15,8,11$$ It is easy to check that any two colours $i$ in this repeating sequence are separated by at least $i$ integers. Note that the previous pattern was found by a computer search. Asymptotic results for subcubic outerplanar graphs ================================================== The main goal of this paper is to study the finiteness of the packing chromatic number of subcubic outerplanar graphs, i.e, we ask whether the packing chromatic number of an outerplanar graph with maximum degree at most $3$ depends on the order of the graph or not. In this section we prove that, for any $2$-connected outerplanar graph with a fixed number of internal faces and for any connected outerplanar graph with a fixed number of faces, the packing chromatic number does not depend on the order of the graph. We begin this section by proving the following useful lemma that will also be used in Section 4. We recall that the weak dual of a $2$-connected outerplanar graph is a tree and that $u_F$ is the vertex of the weak dual corresponding to the face $F$. \[out2\] There exists a packing colouring of $P^{+}_{\infty}$ with colours $\{5,\ldots,15\}$ such that the first vertex along the path is at distance at least $\lceil (i-5)/2 \rceil$ of any vertex of colour $i$. By considering Pattern  from the proof of Theorem \[out1\] starting at the first vertex of the path (the vertex of degree $1$), we can easily check that the first six vertices of $P^{+}_{\infty}$ satisfy the property. Since the colours used in Pattern  are bounded by $15$, the other vertices (other than the first six vertices) satisfy the property as well. For positive integers $i$, $j$ and $k$, let $r^k_{i,j}\in \mathbb{Z}$ such that $ r^k_{i,j}\equiv i-j \pmod{k}$ with minimum absolute value. The value $|r^k_{i,j}|$ corresponds to the distance between two vertices $i$ and $j$ in a cycle $C_k$ with vertex set $\{0,\ldots, k-1\}$ (the vertices are enumerated along the cycle). A subset of vertices $A$ of a graph $G$ is a *cycle-distance-preserved* set if there exists an ordering $v_{A}^{0},\ldots,v_{A}^{|A|-1}$ of the vertices of $A$ satisfying $d_G(v^j_{A},v^{j'}_{A})\ge |r^{|A|}_{j,j'}|$, for integers $0\le j<j'\le |A|-1$. \[acycle\] For any positive integers $k$ and $n>2$, there exists a packing colouring of the cycle $C_n$ with colours from $\{k,\ldots,6k+4 \}$. Let $C_n$ be a cycle of length $n$. First, if $n\le 5k+5$, then we can colour each vertex of $C_n$ with a different colour from $\{k,\ldots,6k+4 \}$. Second, if $5k+5 <n\le 6k+5$, then we colour $3k$ consecutive vertices of $C$ with colours $k,\ldots, 3k-1,k,\dots, 2k-1$, and colour the remaining $n-3k\le 3k+5$ vertices of $C_n$ with mutually distinct colours from $\{ 3k, \ldots, 6k+4\}$. Third, suppose $n> 6k+5$. By Proposition \[preliminaries\].ii), we can colour $P_{\infty}$ with colours from $\{k,\ldots, 3k+2\}$. Let $P'$ be any subpath of $C_n$ on $3k+2$ vertices. Since the distance between the two ends of $C_n-P'$ is at least $3k+3$ in $C_n-P'$ and exactly $3k+3$ in $C_n$, we can colour the vertices of $C_n-P'$ with the colours $\{k,\ldots, 3k+2\}$ (using Proposition \[preliminaries\].ii) ) and the vertices of $P'$ with mutually distinct colours from $\{ 3k +3,\ldots, 6k+ 4 \} $. A subset of vertices $A$ of a graph $G$ is *decomposable* into $r$ cycle-distance-preserved sets if there exist $r$ sets of vertices $A_1$, $A_2$, $\ldots$, $A_r$, such that $A_1\cup \ldots \cup A_r=A$ and for each integer $i$, $A_i$ is a cycle-distance-preserved set. The following lemma will be useful in order to prove Theorems \[k-inner face structure\] and \[k-face structure\]. \[disjoint cycles\] Let $G$ be a graph and let $A\subseteq V(G)$ be a subset decomposable into $r$ cycle-distance-preserved sets. The vertices of $A$ can be packing-coloured with colours $\{k,\ldots,6^{r} (k+1)-2 \}$, for any positive integer $k$. We proceed by induction on $r$. For $r=1$, since $A$ is a cycle-distance-preserved set, by Lemma \[acycle\], we can colour the vertices of $A$ with colours $\{k,\ldots,6k+4 \}$. Now suppose that a subset $A\subset V(G)$ is decomposable into $r+1$ cycle-distance-preserved sets. Using induction hypothesis we can colour the vertices of $A_1$, $\ldots$, $A_r$ with colours $\{k,\ldots,6^{r} (k+1)-2 \}$. For the vertices of $A_{r+1}$, by Lemma \[acycle\], we can use colours $\{6^{r} (k+1)-1, \ldots, k '\}$, where $k'= 6( 6^{r} (k+1) -1)+4= 6^{r+1} (k+1)-2$. Note that we do not need to change colours of the vertices from $\cup_{i=1}^r (A_i \cap A_{r+1})$ (in the case it is not empty). The following theorem is one of our main results. It shows that the packing chromatic number of a 2-connected subcubic outerplanar graph with bounded number of internal faces is bounded by a constant. \[k-inner face structure\] If $G$ is a $2$-connected subcubic outerplanar graph with $r$ internal faces, then $\chi_{\rho}(G)\le 17\times 6^{3r}-2$. Let $F_1, \ldots, F_r$ denote the $r$ distinct internal faces of $G$, and $u_{F_1}, \dots u_{F_r}$ the corresponding vertices of $\mathcal{T}_G$ (note that each $u_{F_i}$ has degree at least $3$ in $\mathcal{T}_G$). Figure \[figadd1\] illustrates a subcubic outerplanar graph and its dual. Then, removing the vertices $u_{F_1}, \dots, u_{F_r}$ from $\mathcal{T}_G$, we obtain a union of disjoint paths. The connected components with one end vertex adjacent in $\mathcal{T}_G$ to $u_{F_i}$ and the other end vertex adjacent in $\mathcal{T}_G$ to $u_{F_j}$ are denoted by $U_{i,j}$. In the case $u_{F_i}$ and $u_{F_j}$ are adjacent, $U_{i,j}$ is not defined (all the vertices in the face “between” $F_i$ and $F_j$ are colored in Step 3). For any $u_{F_i}$, $i=1,\dots, r$, the paths with one end vertex of degree $1$ and the other one adjacent to $u_{F_i}$, are denoted by $U_{i}^{1}, \ldots, U_{i}^{\ell_i}$, where $\ell_i$ is the number of such paths for $u_{F_i}$. Note that some of the paths $U_{i,j}$,$U_{i}^{q}$ may be trivial or empty. (-4,0) – (4,0); (-4,1.5) – (4,1.5); (-1,0) – (0,1.5); (1,0) – (1,1.5); (-2,0) – (-3,1.5); (2,1.5) – (3,0); (4,0) – (4,1.5); (-4,0) – (-4,1.5); (4,0) – (4.8,-0.7); (6.2,-0.7) – (4.8,-0.7); (6.2,-0.7) – (7,0); (7,0) – (7,1.5); (7,1.5) – (6.2,2.2); (6.2,2.2) – (4.8,2.2); (4.8,2.2) – (4,1.5); (-4,0) – (-4.8,-0.7); (-6.2,-0.7) – (-4.8,-0.7); (-6.2,-0.7) – (-7,0); (-7,0) – (-7,1.5); (-7,1.5) – (-6.2,2.2); (-6.2,2.2) – (-4.8,2.2); (-4.8,2.2) – (-4,1.5); (-4.8,2.2) – (-4.8,3); (-6.2,2.2) – (-6.2,3); (-4.8,3) – (-5.5,3.7); (-6.2,3) – (-5.5,3.7); (-4.8,3) – (-6.2,3); (4.8,2.2) – (5.5,3); (6.2,2.2) – (5.5,3); (7,1.5) – (11,1.5); (7,0) – (11,0); (8,0) – (9,1.5); (11,0) – (10,1.5); (11,0) – (11,1.5); (-7,0) – (-7.3,-1); (-6.2,-0.7) – (-7.3,-1); (4.8,-0.7) – (4.8,-1.6); (6.2,-0.7) – (6.2,-1.6); (4.8,-1.6) – (6.2,-1.6); (0,1.5) – (2,1.5); (-1,0) – (0,1.5); (-1,0) – (-2,0); (-2,0) – (-3,1.5); (2,1.5) – (3,0); (-5.5,3.7) – (-6.2,3); (8,0) – (9,1.5); (9,1.5) – (11,1.5); at (-4,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (0,0) \[circle,draw=black,fill=black,scale=0.5\]; at (1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (4,0) \[circle,draw=black,fill=black,scale=0.5\]; at (3,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (0,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (1,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (2,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (4,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (3,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (4.8,-0.7) \[circle,draw=black,fill=black,scale=0.5\]; at (6.2,-0.7) \[circle,draw=black,fill=black,scale=0.5\]; at (7,0) \[circle,draw=black,fill=black,scale=0.5\]; at (7,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (6.2,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (4.8,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.8,-0.7) \[circle,draw=black,fill=black,scale=0.5\]; at (-6.2,-0.7) \[circle,draw=black,fill=black,scale=0.5\]; at (-7,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-7,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-6.2,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.8,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.8,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-6.2,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,3.7) \[circle,draw=black,fill=black,scale=0.5\]; at (8,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (9,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (10,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (11,1.5) \[circle,draw=black,fill=black,scale=0.5\]; at (8,0) \[circle,draw=black,fill=black,scale=0.5\]; at (9,0) \[circle,draw=black,fill=black,scale=0.5\]; at (10,0) \[circle,draw=black,fill=black,scale=0.5\]; at (11,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-7.3,-1) \[circle,draw=black,fill=black,scale=0.5\]; at (4.8,-1.6) \[circle,draw=black,fill=black,scale=0.5\]; at (6.2,-1.6) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,0.75) ; at (5.5,0.75) ; at (-5.5,3.25) ; at (-6.8,-0.55) ; at (5.5,2.55) ; at (10.7,1) ; at (5.5,-1.2) ; at (-5.2,3.8) ; at (-6.9,-1.1) ; at (5.8,3.1) ; at (11.2,1.8) ; at (6.5,-1.8) ; at (8.2,-0.3) ; at (9.2,1.8) ; at (10.2,1.8) ; at (-6.7,3) ; at (-2.8,1.8) ; at (-1.8,-0.3) ; at (-0.8,-0.3) ; at (0.2,1.8) ; at (1.2,1.8) ; at (2.2,1.8) ; at (3.2,-0.3) ; (5.5,0.75) circle (2.75cm); (2,1.5) circle (0.7cm); (9,1.5) circle (0.7cm); (-5.5,0.75) circle (2.75cm); (-2,0) circle (0.7cm); (-5.5,3.7) circle (0.7cm); (-3,0) – (3,0); (-3,0) – (-3.25,0.25); (-3.25,0.25) – (-3.5,0.5); (-3.5,0.5) – (-3.75,0.75); (-3.75,0.75) – (-4,1); (-3,0) – (-3.25,-0.25); (-3.25,-0.25) – (-3.5,-0.5); (-3,0) – (5,0); (-3,0) – (-2.5,0); (-2,0) – (-1.5,0); (-1,0) – (-0.5,0); (0,0) – (0.5,0); (1,0) – (1.5,0); (2,0) – (2.5,0); (3,0) – (3.35,0); (3.7,0) – (4.05,0); (4.4,0) – (4.75,0); (3,0) – (3.25,0.25); (3.25,0.25) – (3.5,0.5); (3,0) – (3.25,-0.25); (3.25,-0.25) – (3.5,-0.5); at (-3,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (3,0) \[circle,draw=black,fill=black,scale=0.5\]; at (2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (0,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,-0.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,0.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,1) \[circle,draw=black,fill=black,scale=0.5\]; at (3.5,-0.5) \[circle,draw=black,fill=black,scale=0.5\]; at (3.5,0.5) \[circle,draw=black,fill=black,scale=0.5\]; at (3.7,0) \[circle,draw=black,fill=black,scale=0.5\]; at (4.4,0) \[circle,draw=black,fill=black,scale=0.5\]; at (5.1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.7,-0.3) ; at (2.9,-0.3) ; at (0.65,-0.5) ; For any $i$, $1\le i\le r$, let $B_i=N(F_i)\setminus \bigcup_{i=1}^r V(F_i)$, and let $B=\bigcup_{i=1}^r (F_i\cup B_i)$. Let $i$ and $q$ be integers such that $1\le i\le r$, $1\le q\le \ell_i$. Consider an end face $\hat F_{i}^{q}$ in $G$ corresponding to an end vertex of $V(U_{i}^{q})$ of degree 1 in $\mathcal{T}_G$. We denote by $y_i^{q}$ a vertex of $\hat F_{i}^{q}$ of degree $2$ (note that such a vertex always exists) and by $P_i^{q}$ a shortest $(B_i,y_i^{q})$-path in $G$. Let $(p_{i}^{q})_1,(p_{i}^{q})_2,\ldots$ denote vertices of $P_i^{q}$ in an ordering starting from the vertex of $B_i$. Choose any vertex $z$ of $\mathcal{T}_G$ such that $z$ has degree $3$ in $\mathcal{T}_G$ and let $\overrightarrow{\mathcal{T}_G}$ be the digraph obtained from $\mathcal{T}_G$ by replacing each edge $uv\in E(\mathcal{T}_G)$ satisfying $d_{\mathcal{T}_G}(u,z)<d_{\mathcal{T}_G}(v,z)$ with an arc from $u$ to $v$. Now we consider the paths $U_{i,j}$ in $\mathcal{T}_G$. Let $i,j$ be positive integers such that $1\le i<j\le r$ and $U_{i,j}$ is defined and has length at least one. Let $P_{i,j}$ be a shortest $(B_i, B_j)$-path and $p_{i,j}^1,p_{i,j}^2 \ldots,p_{i,j}^{l_{i,j}}$ its vertices in an ordering starting from the vertex of $B_i$, if $d_{\mathcal{T}_G}(u_{F_i},z)<d_{\mathcal{T}_G}(u_{F_j},z)$, or from the vertex of $B_j$ otherwise (by $l_{i,j}$ we mean the order of $P_{i,j}$). The upper part of Figure \[figadd1\] illustrates the notations used in this proof. Let $P =\left(\bigcup\limits_{1\le i\le r} \bigcup\limits_{1\le q\le \ell_i } V(P_{i}^{q})\right)\,\, \bigcup \,\, \left(\bigcup\limits_{1\le i<j\le k} V(P_{i,j})\right)$. Colouring the vertices of $V(G)\setminus (B\cup P)$ with colours $\{1,2,3,4\}$. We colour the vertices of $V(G)\setminus (B\cup P)$ by colouring each connected component (one by one) of $G- (B\cup P)$ in the same way as in the proof of Lemma \[lemma\], i.e., we use the pattern $1,2,1,3$. Note that the distance between any two vertices from $V(G)\setminus (B\cup P)$ in two different connected components of $G-B$ is at least $5$. Moreover, we proceed as in the proof of Lemma \[lemma\] to avoid clashing vertices of colour $3$, i.e., we use colour $4$. Colouring vertices of $P$. Let $i$, $j$, $i'$ and $q$ be integers such that $U_{i,j}$ and $U_{i'}^{q}$ are defined. For the vertices of $P_{i'}^{q}$, we use Pattern  and Lemma \[out2\] starting at the vertex $(p_{i'}^{q})_3$. For the vertices of $P_{i,j}$, we use Pattern  and Lemma \[out2\], starting at the vertex $p_{i,j}^3$ and finishing at the vertex $p_{i,j}^{k_{i,j}-3}$. Note that every vertex of $V(\overrightarrow{\mathcal{T}_G})$ has in-degree at most one. This property, along with Lemma \[out2\], ensure us that a vertex coloured with colour $a$ in $P_{i,j}$, $a\in\{5,\ldots,15\}$, is at distance at least $a+1$ from any other vertex coloured by $a$ in $P_{\bar i,\bar j}$, for $1\le \bar i< \bar j\le r$. Colouring the remaining vertices of $G$. Let $w_{i,j}$ be a vertex among $\{p_{i,j}^{2} ,p_{i,j}^{l_{i,j}-2}\}$ at distance $2$ from a vertex of $V(F_i)$ (when $U_{i,j}$ is defined). Let $D_i$ be the set $\{(p_{i}^{A})_2|\ 1\le A\le \ell_i\}\cup \{w_{i,j}|\ U_{i,j} \text{ is defined,}\ 1\le j\le k\}$. Since the sets $V(F_i)$, $B_i$ and $D_i$ , $1\le i\le r$, are cycle-distance-preserved sets, the set $\bigcup_{i=1}^r V(F_i)\bigcup_{i=1}^r B_i \bigcup_{i=1}^r D_i$ is decomposable into $3r$ cycle-distance-preserved sets. Hence, using Lemma \[disjoint cycles\], the remaining uncoloured vertices can be coloured with colours $\{16,\ldots,17\times 6^{3r}-2 \}$. Sloper in [@Slo] defined an [*expandable broadcast-colouring*]{} of a complete binary tree $T$ as a colouring $c$ of $V(T)$ with colours $1,2,\dots, 7$ such that: $\forall u,v\in V(T)$ $c(u)=c(v) \Rightarrow d_T(u,v)>c(u)$, the root $x$ of $T$ has colour $1$, all vertices at even distance from $x$ have colour $1$, every vertex of colour $1$ has at least one child of colour $2$ or $3$, $c(u)=6, c(v)=7 \Rightarrow d_T(u,v)\ge 5$, $c(u)\in \{4,5,6,7\} \Rightarrow$ $u$’s children each have children coloured with $2$ and $3$. Notice that an expandable broadcast-colouring of a tree is a packing 7-colouring. Sloper has shown that given an expandable colouring of a (complete) binary tree of height $n$, it is possible to create an expandable colouring of a (complete) binary tree of height $(n + 1)$ by using the colouring for the tree of height $n$ as a basis [@Slo]. Note that the colouring of a complete binary tree of height $3$ that consists in giving the colours $2$ and $3$ to the two neighbours of $x$, giving the colour $1$ to the vertex at distance $2$ from $x$, and giving colours from $\{2,3,4,5\}$ to the remaining vertices is an expendable broadcast-colouring. This colouring is described in Figure \[figadd2\]. (0,3) – (-3,2); (0,3) – (3,2); (-3,2) – (-4.5,1); (-3,2) – (-1.5,1); (3,2) – (4.5,1); (3,2) – (1.5,1); (-4.5,1) – (-5.25,0); (-4.5,1) – (-3.75,0); (-1.5,1) – (-2.25,0); (-1.5,1) – (-0.75,0); (4.5,1) – (5.25,0); (4.5,1) – (3.75,0); (1.5,1) – (2.25,0); (1.5,1) – (0.75,0); at (0,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,2) \[circle,draw=black,fill=black,scale=0.5\]; at (3,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (4.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.25,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.75,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.25,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-0.75,0) \[circle,draw=black,fill=black,scale=0.5\]; at (5.25,0) \[circle,draw=black,fill=black,scale=0.5\]; at (3.75,0) \[circle,draw=black,fill=black,scale=0.5\]; at (2.25,0) \[circle,draw=black,fill=black,scale=0.5\]; at (0.75,0) \[circle,draw=black,fill=black,scale=0.5\]; at (0,3.35) [$x$]{}; at (0,2.65) [$1$]{}; at (-3,1.65) [$2$]{}; at (3,1.65) [$3$]{}; at (-4.5,0.65) [$1$]{}; at (-1.5,0.65) [$1$]{}; at (1.5,0.65) [$1$]{}; at (4.5,0.65) [$1$]{}; at (-5.25,-0.35) [$4$]{}; at (-3.75,-0.35) [$3$]{}; at (-2.25,-0.35) [$5$]{}; at (-0.75,-0.35) [$3$]{}; at (5.25,-0.35) [$4$]{}; at (3.75,-0.35) [$2$]{}; at (2.25,-0.35) [$1$]{}; at (0.75,-0.35) [$2$]{}; The following statement is true for a more general class of graphs than in Theorem \[k-inner face structure\] since it gives an upper bound for all connected outerplanar graphs (not necessarily $2$-connected). However, since the parameter is the number of faces, the bound is weaker than the bound in Theorem \[k-inner face structure\]. \[k-face structure\] If $G$ is a connected subcubic outerplanar graph with $r$ (non external) faces, then $\chi_{\rho}(G)\le 9\times 6^{r}-2$. Let $F_1, \ldots, F_r$ denote the $r$ (non external) faces of $G$. The graph $O=G-\bigcup_{i=1}^r V(F_i)$ consists of components $O_1,O_2, \dots, O_s$ such that each $O_j$ ($j=1,\dots, s$) is a tree. And, since $G$ is subcubic, each $O_j$ is subcubic as well. In the weak dual $\mathcal{T}_G$ of $G$, choose arbitrary vertex $z$ and let $F_z$ denote a face corresponding to $z$ in $G$. We colour the vertices of $G$ in two steps. Colouring the vertices of $O$ with colours $\{1, \ldots,7\}$. Consider each component $O_j$ of $O$ separately ($i=1,\dots, s$) and let $z_j$ denote the vertex of $O_j$ closest to $F_z$. Then we use an expandable broadcast-colouring to colour vertices of $O_j$ with colours $1,2,\dots, 7$ such that the vertices at distance at most $3$ from $z_j$ are coloured as in Figure \[figadd2\] (by considering $z_j$ as $x$ in this figure). Using the result of Sloper [@Slo], it is possible to extend this packing colouring to the graph $O_j$. Note that $z_j$ has colour $1$, the neighbour(s) of $z_j$ in $O_j$ has (have) colour $2$ (and $3$), vertices of $O_j$ at distance $2$ from $z_j$ are coloured with colour $1$ and vertices of $O_j$ at distance $3$ have colours $2$, $3$, $4$ and $5$. Obviously, since $G$ is subcubic, $z_j$ is at distance at least $3$ from any vertex of any $O_\ell\not=O_i$. Note that, except one vertex, every vertex of $O$ having a neighbor in $F_i$ is at shortest distance of $F_z$ (compared to the other vertices in the same component of $O$). Consequently, by definition of $z_j$, in every face $F_i$ there is at most one vertex which has a neighbour in $O$ which is not $z_j$ for some $j\in\{1,\ldots,s\}$. Let $w_i$ be this possible neighbour in $O$. Note that $w_i$ can have any colour among $\{1,\ldots,7\}$. If $w_i$ has colour in $\{2,3\}$, then the other vertices of colour $2$ or $3$ at close distance from vertices of $F_i$ are the neighbours of the vertices $z_j$, $j\in\{1,\ldots,s\}$, which are at distance $4$ from $w_i$. If $w_i$ has colour in $\{4,5\}$, then, also, the vertices of colour $4$ or $5$ at close distance from vertices of $F_i$ are the vertices at distance $3$ (in $O_j$) of the vertices $z_j$, $j\in\{1,\ldots,s\}$, and these vertices are at distance $6$ from $w_i$. Finally, since the vertices at even distance of $z_j$ in $O_j$ are coloured with colour $1$, the other vertices of colour $6$ or $7$ are at distance at least $8$ from $w_i$. Hence, the above defined colouring satisfies the distance constraints of a packing colouring. Colouring the remaining vertices of $G$. The sets $V(F_1),\ldots, V(F_r)$ are cycle-distance-preserved sets. Hence, by Lemma \[disjoint cycles\], the remaining uncoloured vertices can be coloured with colours $\{8,\ldots, 9\times 6^{r}-2 \}$. Some $2$-connected outerplanar graphs with finite packing chromatic number ========================================================================== In this section we consider some special classes of subcubic outerplanar graphs for which we can decrease the upper bound on the packing chromatic number given in Theorem \[k-inner face structure\]. \[thm star structure\] If $G$ is a $2$-connected subcubic outerplanar graph with exactly one internal face, then $\chi_{\rho}(G)\leq 51$. Suppose $G$ is a $2$-connected subcubic outerplanar graph with exactly one internal face. Let $C$ denote the boundary cycle of $G$ and $F$ the internal face of $G$. Let $C'=\{v_0,\ldots, v_{N-1}\}$ denote the set of vertices which belong to $F$, with $v_i$ adjacent to $v_{i+1}$, for $0\le i<N$. When $N$ is odd, we suppose that $v_{N-1}$ is a vertex with $d_G(v_{N-1})=2$. Such a vertex exists since the number of vertices of degree 3 in $C'$ is even. By removing the edges of $C\cap F$ from $G$, and by removing the isolated vertices from the resulting graph, we obtain a graph $G'$ which is a disjoint union of $2$-connected outerplanar graphs having no internal face. (0,0) circle (3cm); (2.5,1.7) .. controls (6,2.8) and (6,1) .. (2.97,0.3); (2.5,-1.7) .. controls (6,-2.8) and (6,-1) .. (2.97,-0.3); (1.7,-2.5) .. controls (3.3,-5.5) and (1.5,-5.5) .. (0.3,-2.97); (-2.5,1.7) .. controls (-6,2.8) and (-6,1) .. (-2.97,0.3); (2.97,-0.3) .. controls (3,0) .. (2.97,0.3); (2.5,-1.7) .. controls (2.2,-2.1) .. (1.7,-2.5); (0.3,-2.97) .. controls (-1.7,-2.97) and (-3 ,-1.95) .. (-2.97,0.3); (2.5,1.7) .. controls (1.4,3.4) and (-1.4,3.4) .. (-2.5,1.7); at (0,0) [$F$]{}; at (4,1.3) [$F_{s-1} $]{}; at (4,-1.3) [$F_0 $]{}; at (1.5,-3.7) [$F_1 $]{}; at (-4,1.3) [$F_i $]{}; at (2.1,-1.6) [$u_0^b$]{}; at (2.5,-1.7) \[circle,draw=black,fill=black,scale=0.5\]; at (2.5,0.3) [$u_{s-1}^b$]{}; at (2.97,0.3) \[circle,draw=black,fill=black,scale=0.5\]; at (2.6,-0.3) [$u_0^a$]{}; at (2.97,-0.3) \[circle,draw=black,fill=black,scale=0.5\]; at (2.1,1.6) [$u_{s-1}^a$]{}; at (2.5,1.7) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,-2.1) [$u_1^b$]{}; at (1.7,-2.5) \[circle,draw=black,fill=black,scale=0.5\]; at (0.3,-2.6) [$u_1^a$]{}; at (0.3,-2.97) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.1,1.6) [$u_i^b$]{}; at (-2.5,1.7) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.6,0.3) [$u_i^a$]{}; at (-2.97,0.3) \[circle,draw=black,fill=black,scale=0.5\]; Let $F_0$, $\ldots$, $F_{s-1}$ denote the $2$-connected components of $G'$, enumerated in the clockwise order along the cycle $C$ in $G$ (for details, see Fig. \[out11\]). Note that, since any $F_i$ contains no internal face, each $F_i$ has exactly two end faces or $F_i$ is a cycle. Let $i$ be an integer with $0\le i< s$, and let $u_i^a$ and $u_i^b$ denote the two adjacent vertices of degree $3$ in $G$ which belong to $V(F_i)\cap C'$, as it is depicted in Fig. \[out11\]. Let $y_i$ be a vertex of degree $2$ in the end face of $F_i$ which does not contain $u_i^a$ (for $F_i$ a cycle we denote by $y_i$ a vertex of $F_i$ at maximum distance from $F$ in $G$). Let $x_i\in \{u_i^a, u_i^b\}$ denote a vertex at minimal distance from $y_i$. Finally, let $P_i$ be a shortest $(x_i,y_i)$-path in $G$. We further denote the vertices of each $P_i$ by $x_i, p_i^1,p_i^2, \ldots, y_i$ in an ordering starting from $x_i$. Let $D_i^1$, $\ldots$, $D_i^{k_i}$ denote the connected components of $F_i-P_i$ with $D_i^1$ containing a vertex among $u_i^a$ and $u_i^b$ and with $D_i^{k}$ being at larger distance than $D_{i}^{k-1}$ from $x_i$, $2\le k\le k_i$. The proof will be organized as follows. First, we will colour the vertices of $C'$. Second, we will colour the vertices of $\cup_{0\le i < s} F_i-P_i$ with colour $1$, $2$ and $3$. Note that the obtained colouring does not necessarily satisfy the distance constraints of a packing colouring of $G$. Third, we will modify colouring of some vertices of $F_i-P_i$ ($i=0,\dots, s-1$) to save colour $1$ for some vertices of the paths $P_i$ and to prevent collisions in colour $2$. Fourth, we will recolour some vertices of $F_0,\dots, F_{s-1}$ with colour $4$ in order to satisfy the distance constraints of a packing colouring. Finally, we will colour vertices of the paths $\cup_{0\le i < s} P_i\setminus \{x_i\}$. (-2,0) .. controls (0,0.3) .. (2,0); (-1.3,0.1) .. controls (-1.5,4.8) and (1.5,4.8) .. (1.3,0.1); (-1.3,0.8) – (1.2,2); (-1.3,0.1) .. controls (-1.3,1.5) .. (-1.1,2.4); at (-1.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (1.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (1.2,2) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.2,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3,-0.3) [$x_i $]{}; at (1.3,-0.3) [$x'_i $]{}; at (-1.6,1.6) [$p_i^2$]{}; at (-1.7,0.9) [$p_i^1$]{}; at (1.6,2) [$w_i $]{}; at (1.6,2.5) [$\ldots$]{}; at (1.8,0.8) [$\ldots$]{}; at (0,-0.8) [(a)]{}; (10,0) .. controls (12,0.3) .. (14,0); (10.7,0.1) .. controls (10.5,4.8) and (13.5,4.8) .. (13.3,0.1); (10.7,0.8) – (13.2,2); (10.8,1.6) – (12.9,2.9); (10.7,0.1) .. controls (10.7,0.8) .. (10.75,1.6); (10.8,1.6) – (12.9,2.9); at (10.7,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (13.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (13.2,2) \[circle,draw=black,fill=black,scale=0.7\]; at (10.8,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (10.95,2.4) \[circle,draw=black,fill=black,scale=0.7\]; at (12.9,2.9) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7,-0.3) [$x_i $]{}; at (13.3,-0.3) [$x'_i $]{}; at (10.5,2.4) [$w'_i $]{}; at (10.4,1.6) [$p_i^2$]{}; at (10.3,0.9) [$p_i^1$]{}; at (13.6,2) [$w_i $]{}; at (13.4,3) [$p_i^3$]{}; at (13.6,2.5) [$\ldots$]{}; at (13.8,0.8) [$\ldots$]{}; at (12,-0.8) [(c)]{}; (4,0) .. controls (6,0.3) .. (8,0); (4.7,0.1) .. controls (4.5,4.8) and (7.5,4.8) .. (7.3,0.1); (4.7,0.8) – (7.2,2); (4.8,1.6) – (6.9,2.9); (4.7,0.1) .. controls (4.7,1.5) .. (4.9,2.4); at (4.7,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (7.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (4.7,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (4.8,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (7.2,2) \[circle,draw=black,fill=black,scale=0.7\]; at (6.9,2.9) \[circle,draw=black,fill=black,scale=0.7\]; at (4.7,-0.3) [$x_i $]{}; at (7.3,-0.3) [$x'_i $]{}; at (4.3,0.9) [$p_i^1$]{}; at (4.4,1.6) [$p_i^2$]{}; at (7.6,2.5) [$\ldots$]{}; at (7.6,2) [$w_i $]{}; at (7.4,2.9) [$w'_i $]{}; at (7.8,0.8) [$\ldots$]{}; at (6,-0.8) [(b)]{}; (-2,0) .. controls (0,0.3) .. (2,0); (-1.3,0.1) .. controls (-1.5,4.8) and (1.5,4.8) .. (1.3,0.1); (-1.2,1.6) – (1.3,0.8); (-1.3,0.1) .. controls (-1.3,1.5) .. (-1.1,2.4); at (-1.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (1.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.2,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (1.3,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3,-0.3) [$x_i $]{}; at (1.3,-0.3) [$x'_i $]{}; at (-1.6,1.6) [$p_i^2$]{}; at (-1.7,0.9) [$p_i^1$]{}; at (1.7,0.8) [$w'_i $]{}; at (1.7,1.6) [$\ldots$]{}; at (0,-0.8) [(d)]{}; (10-6,0) .. controls (12-6,0.3) .. (14-6,0); (10.7-6,0.1) .. controls (10.5-6,4.8) and (13.5-6,4.8) .. (13.3-6,0.1); (10.7-6,0.8) – (13.2-6,1.7); (10.7-6,0.1) – (10.7-6,0.8); (10.7-6,0.8) – (13.2-6,1.7); (13-6,2.6) .. controls (13.1-6,2.3) .. (13.2-6,1.7); at (10.7-6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (13.3-6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7-6,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (13.2-6,1.7) \[circle,draw=black,fill=black,scale=0.7\]; at (13.3-6,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7-6,-0.3) [$x_i $]{}; at (13.3-6,-0.3) [$x'_i $]{}; at (13.7-6,1.7) [$p_i^2 $]{}; at (10.3-6,0.9) [$p_i^1$]{}; at (13.6-6,2.5) [$\ldots$]{}; at (13.8-6,0.8) [$w'_i$]{}; at (12-6,-0.8) [(e)]{}; (4+6,0) .. controls (6+6,0.3) .. (8+6,0); (4.7+6,0.1) .. controls (4.5+6,4.8) and (7.5+6,4.8) .. (7.3+6,0.1); (4.8+6,1.6) – (6.9+6,2.9); (4.7+6,0.1) .. controls (4.7+6,0.8) .. (4.75+6,1.6); (4.8+6,1.6) – (6.9+6,2.9); at (4.7+6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (7.3+6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (4.7+6,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (4.8+6,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (6.9+6,2.9) \[circle,draw=black,fill=black,scale=0.7\]; at (4.95+6,2.4) \[circle,draw=black,fill=black,scale=0.7\]; at (4.5+6,2.4) [$w'_i $]{}; at (4.7+6,-0.3) [$x_i $]{}; at (7.3+6,-0.3) [$x'_i $]{}; at (4.4+6,1.6) [$p_i^2$]{}; at (4.3+6,0.9) [$p_i^1$]{}; at (7.4+6,3) [$p_i^3$]{}; at (7.7+6,1.4) [$\ldots$]{}; at (6+6,-0.8) [(f)]{}; (-2+6,0) .. controls (0+6,0.3) .. (2+6,0); (-1.3+6,0.1) .. controls (-1.5+6,4.8) and (1.5+6,4.8) .. (1.3+6,0.1); (-1.2+6,1.6) – (1.25+6,1.6); (-1.3+6,0.1) .. controls (-1.3+6,1.5) .. (-1.1+6,2.4); at (-1.3+6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (1.3+6,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3+6,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (1.25+6,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.2+6,1.6) \[circle,draw=black,fill=black,scale=0.7\]; at (-1.3+6,-0.3) [$x_i $]{}; at (1.3+6,-0.3) [$x'_i $]{}; at (-1.6+6,1.6) [$p_i^2$]{}; at (-1.7+6,0.9) [$p_i^1$]{}; at (1.8+6,1.6) [$w'_i $]{}; at (1.6+6,2.5) [$\ldots$]{}; at (1.8+6,0.8) [$\ldots$]{}; at (0+6,-0.8) [(g)]{}; (10,0) .. controls (12,0.3) .. (14,0); (10.7,0.1) .. controls (10.5,4.8) and (13.5,4.8) .. (13.3,0.1); (10.7,0.8) – (13.2,1.7); (10.7,0.1) – (10.7,0.8); (10.7,0.8) – (13.2,1.7); (13,2.6) .. controls (13.1,2.3) .. (13.2,1.7); at (10.7,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (13.3,0.1) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7,0.8) \[circle,draw=black,fill=black,scale=0.7\]; at (13.2,1.7) \[circle,draw=black,fill=black,scale=0.7\]; at (13.3,1.1) \[circle,draw=black,fill=black,scale=0.7\]; at (10.7,-0.3) [$x_i $]{}; at (13.3,-0.3) [$x'_i $]{}; at (13.7,1.7) [$p_i^2 $]{}; at (10.3,0.9) [$p_i^1$]{}; at (13.6,2.5) [$\ldots$]{}; at (13.8,1.1) [$w'_i$]{}; at (13.8,0.55) [$\ldots$]{}; at (12,-0.8) [(h)]{}; Colouring the vertices of $C'$ with colours $1$, $2$, $29, 30,\dots, 45$. For integers $j$, $j'$, let $r_{j,j'}$ be an integer such that $r_{j,j'}\equiv j-j' \pmod{N}$ and $- \lfloor N/2 \rfloor\le r_{j,j'}\le \lfloor N/2 \rfloor$. Note that $d_G(v_j,v_{j'}) =|r_{j,j'}|$. We begin with a partitioning of $C'$ into five subsets: $C'_1=\{v_j|\ j\equiv 0\pmod{2},\ 0\le j< N\}$, $C'_2=\{v_j|\ j\equiv 1\pmod{4},\ 0\le j< N\}$, $C'_3=\{v_j|\ j\equiv 3\pmod{12},\ 0\le j< N\}$, $C'_4=\{v_j|\ j\equiv 7\pmod{12},\ 0\le j< N\}$ and $C'_5=\{v_j|\ j\equiv 11\pmod{12},\ 0\le j< N\}$. Let $m_k$ denote the vertex with the largest index in $C'_k$, for $k\in\{1,2,3,4,5\}$. We use the following patterns to colour the vertices of $C'$. 1. if $|C'_1|\equiv 0\pmod{2}$ (or $|C'_1|\equiv 1\pmod{2}$), then we colour all vertices of $C'_1$ (or $C'_1\setminus \{m_1\}$, respectively) with colour $1$; 2. we colour all vertices of $C'_2\setminus \{m_2\}$ with colour $2$; 3. if $|C'_3|\equiv 0\pmod{4}$ (or $|C'_3|\equiv 1\pmod{4}$, respectively), then we use the pattern $29,30,35,36,$ $29,30,35,36,\ldots, 29,30,35,36$ to colour the vertices of $C'_3$ (or $C'_3\setminus \{m_3\}$, respectively); if $|C'_3|\equiv 2\pmod{4}$ (or $|C'_3|\equiv 3\pmod{4}$), then we use the pattern $29,30,35,$$36,$ $29,30,$$35,36,\ldots,29,30,35,$ $29,30,36$ to colour the vertices of $C'_3$ (or $C'_3\setminus \{ m_3\}$, respectively); 4. if $|C'_4|\equiv 0\pmod{4}$ (or $|C'_4|\equiv 1\pmod{4}$), then we use the pattern $31,32,37,$ $38,$ $31,32,$ $37,38,\ldots,31,32,37,38$ to colour the vertices of $C'_4$ (or $C'_4\setminus \{ m_4\}$, respectively); if $|C'_4|\equiv 2\pmod{4}$ (or $|C'_4|\equiv 3\pmod{4}$), then we use the pattern $31,32,37,$ $38,31,$ $32,$ $37,38,\ldots,31,$ $32,37,31,32,38$ to colour the vertices of $C'_4$ (or $C'_4\setminus \{m_4\}$, respectively); 5. if $|C'_5|\equiv 0\pmod{4}$ (or $|C'_5|\equiv 1\pmod{4}$), then we use the pattern $33,34,39,$ $40,$ $33,34,$ $39,40,\ldots,33,$ $34,39,40$ to colour the vertices of $C'_5$ (or $C'_5\setminus \{m_5\}$, respectively); if $|C'_5|\equiv 2\pmod{4}$ (or $|C'_5|\equiv 3\pmod{4}$), then we use the pattern $33,34,39,$ $40,$ $33,34,$ $39,40,\ldots,33,$ $34,39,$ $33,34,40$ to colour the vertices of $C'_5$ (or $C'_5\setminus \{ m_5\}$, respectively); 6. when it is necessary, we use the colours $41$, $42$, $43$, $44$, $45$ to colour the vertices of $\{m_k |\ 1\le k\le 5\}$. One can check that, for any pair of vertices $u,v$ of $C'$ with the same colour $k$, $d_G(u,v)>k$. For example, in the pattern $29,30,35,36$ of length four, two vertices with the same colour are at distance at least $48$, since the pattern has length four and we colour vertices with the same remainder modulo $12$. The same goes for the pattern $29,30,35,29,30,36$ of length six. Note that, for every pair of vertices $(u_i^a,u_i^b)$, at least one of them is coloured with $1$. Colouring the vertices of $F_i- P_i$ with colours $1$, $2$ and $3$, for every $i=0,\ldots, s-1$. Let $x'_i$ be the vertex among $u_i^a$ and $u_i^b$ different from $x_i$. Let $l_i$ be the order of $D_i^1$ and let $x'_i, x_i^1, \ldots, x_{i}^{ l_{i}-1}$ be the vertices of $D_i^1$ in an ordering starting from $x'_i$. If $x'_i$ is coloured with colour $1$, then we use the pattern $3,1,2,1$ to colour the vertices $x_i^1, \ldots, x_{i}^{l_{i}-1}$. If $x'_i$ is not coloured with colour $1$, then we use the pattern $1,3,1,2$ to colour the vertices $x_i^1, \ldots, x_{i}^{l_i-1}$. Analogously as in the proof of Lemma \[lemma\], we colour vertices of $D_i^j$ ($j=2,3,\dots, k_i$) using the pattern $1,2,1,3$ starting from the vertex of $D_i^j$ at shortest distance from $C'$. At this step we do not change the colouring in order to avoid clashing vertices of colour $3$. Recolouring some vertices in $F_i$, for every $i=0,\ldots, s-1$. In this step we deal with possible collisions in colour $2$ between vertices of $F$ and vertices of $D_i^1$ at distance $2$ from $F$. We also change colours of neighbours of $p_i^2$ coloured with $1$ since, in Step $5$, we will colour $p_i^2$ with $1$ for reducing the number of colours used for the whole graph $G$. Since we used the patterns $1,3,1,2$ and $3,1,2,1$ to colour the vertices of $V(D_i^1)\setminus \{x_i\}$, no vertex at distance $2$ from $x'_i$ has colour $2$. For any $i=0,\dots, s-1$, let $w_i$ denote the vertex of $F_i-P_i$ at distance $2$ from $x_i$ and let $w'_i$ be the possible neighbour of $p_i^2$ in $F_i-P_i$. 1. [*$w_i$ has colour $2$.*]{} First suppose that $p_i^2$ has no neighbour of colour $1$ (see Fig. \[out12\](a)) or $p_i^2$ has a neighbour $w'_i$ with colour $1$ in $D_i^1$ (see Fig. \[out12\](b)). In both possibilities we recolour the vertices $x_i^j=w_i, x_i^{j+1}, x_i^{l_i-1}$ of $D_i^1$ with $\underline{4},2,1,3,1,2,1\ldots$ instead of $\underline{2},1,3,1,2,\ldots$. Note that the underlined colours belong to the vertex $w_i$. Now we assume that $p_i^2$ has a neighbour $w'_i$ of colour $1$ which does not belong to $D_i^1$. Thus $p_i^2 p_i^3$ is a chord (see Fig. \[out12\](c)). We recolour the vertices $x_i^j=w_i, x_i^{j+1}, \dots, x_i^{l_i-1}$ of $D_i^1$ with $\underline{4},1,2,1,3,\ldots$ instead of $\underline{2},1,3,1,2,\ldots$ (the underlined colours belong to $w_i$), and the vertices of $D_i^2$ with pattern $\underline{2},1,3,1,\ldots$ instead of $\underline{1},2,1,3,\ldots$ (the underlined colours belong to $w'_i$). Note that if $w'_i$ had colour $1$, the colour $1$ was changed. 2. [*$w_i$ does not have colour $2$*]{}. If $p_i^2$ has no neighbour $w'_i$ of colour $1$, then we do not modify the colouring of $D_i^1$ in this step. Suppose that $p_i^2$ has a neighbour $w'_i$ of colour $1$. Suppose that $w'_i$ is a neighbour of $x'_i$ (see Fig. \[out12\](d,e)). Clearly $x'_i$ does not have colour $1$ and $x_i$ has colour $1$ (by Step 1). Then we modify the path $P_i$ by replacing vertex $x_i$ with $x'_i$ and $p_i^1$ with $w'_i$ and recolour the modified path $D^i_1$ with pattern $3,1,2,1,3,\dots $. Now suppose that $w'_i$ is at distance at least $2$ from $x'_i$ and that $p_i^2 w'_i$ and $p_i^1 p_i^2$ are not chords (see Fig. \[out12\](f)). We recolour vertices of $D_2^i$ with pattern $\underline{2},1,3,1,\ldots$ instead of $\underline{1},2,1,3,\ldots$. If $w'_i$ is at distance at least $2$ from $x'_i$ and $p_i^2 w'_i$ is a chord (see Fig. \[out12\](g)), we recolour vertices $x_i^j=w'_i,x_i^{j+1}, \dots x_i^{l_i-1}$ of $D_1^i$ with $\underline{4},1,2,1,3,\ldots$ instead of $\underline{1},2,1,3,\ldots$ or $\underline{1},3,1,2,\ldots$. Finally, if $w'_i$ is at distance at least $2$ from $x'_i$ and $p_i^1 p_i^2$ is a chord (see Fig. \[out12\](h)), then we change the colour of $w'_i$ to $4$. Note again that, in each possibility, the underlined colours belong to $w'_i$. Avoid collisions in colouring of vertices of $F_i-P_i$ for $i=0,\dots,s-1 $. Now we check and modify (analogously as in the proof of Lemma \[lemma\]) the defined colouring of $F_i-P_i$ to avoid collisions between pairs of vertice with the same colour. Obviously, there is no collision between vertices coloured with colour $1$ or $2$. Hence the only possible collision is in colour $3$. Let $a$ and $b$ be a pair of clashing vertices in colour $3$. If $a$ and $b$ belong to the same component $D_i^k$ of $F_i - P_i$, $k,\in\{1,2,\ldots,k_i\}$, then we proceed as in Case 2 of the proof of Lemma \[lemma\]. Thus we may assume that $a$ and $b$ belong to different components $D_i^k$ and $D_i^{k'}$ of $F_i - P_i$, $k,k'\in\{1,2,\ldots,k_i\}$. If we changed the colouring of $D_i^2$ in Step 3, then we recolour the vertices of $D_i^2$ starting from $w'_i$ (also defined in Step 3) with pattern $2,3,1,2,1,3,1,\dots $ instead of $2,1,3,1,2,1,\dots$. Then we proceed as in Case 1 of the proof of Lemma \[lemma\]. Now we have to check that the vertices coloured with colour $4$ in Step 3 are pairwise at distance at least $5$, and that the vertices coloured with colour $4$ in Step 3 are pairwise at distance at least $5$ from the added vertices of colour $4$ in Step $4$ of applying Lemma \[lemma\]. By definition, $w_i$ is at distance at most $2$ from both $x_i$ and $x'_i$. Moreover, since $w'_i$ is the neighbour of $p_i^2$ and since we did not recolour $w'_i$ with colour $4$ in the configurations described in Fig. \[out12\](d,e) of Step 3, $w'_i$ is either at distance at least $2$ from both $x_i$ and $x'_i$ or does not have colour $4$. Thus, the vertices recoloured with colour $4$ in Step 3 are at mutual distance at least $5$. Let $a$ be a vertex of colour $4$ from Step 3 (one of $w_i$, $w'_i$ denoted in Step 3). Suppose that $b$ is a vertex of colour $4$ not in $D_i^1$. The minimal distance between any vertex of $D_i^1$ and any vertex of $D_i^2$ is at least $3$. Moreover, because we have proceeded as in the proof of Lemma \[lemma\], $b$ is at distance at least $2$ from a vertex at minimal distance from $a$. Hence, $d(a,b)\ge 5$. Now suppose $b$ is a vertex of colour $4$ in $D_i^1$. Since, in every case, $a$ is at distance at least $3$ from another vertex of colour $3$ in $D_i^1$, we obtain that $d(a,b)\ge 5$. Colouring the vertices of $P_i\setminus\{x_i\}$ with colours 5 to 28 and 46 to 51, for every $i=0,\ldots, s-1$. We start with colouring of the vertices $p^2_i$ by $1$ for each $i=0,\dots s-1$. Since we have changed the colours of the eventual neighbours of $p^2_i$ of colour $1$ in Step 3, there are no possible collisions. For the vertices of $P_i$, we use Pattern  beginning at the fourth vertex of $P_i$, i.e., the vertex $p^3_i$. By the proof of Lemma \[out2\], we know that such a colouring satisfies the distance constraints of a packing colouring. Let $B=\{p^1_{i}|\ 0\le i< s \} $. We colour the vertices of $B$ with colours $16$ to $28$ and (if necessary) $46$ to $51$. For integers $j$, $j'$, let $r_{j,j'}$ be an integer such that $r_{j,j'}\equiv j-j' \pmod{s}$ and $- \lfloor s/2 \rfloor\le r_{j,j'}\le \lfloor s /2 \rfloor$. Note that the vertices $p^1_j$ and $p^1_{j'}$ are at distance $2 \vert r_{j,j'}\vert +1$. We begin by a partitioning of $B$ into three subsets $B_1$, $B_2$ and $B_3$, with $B_k=\{p_i^1|\ i\equiv k-1\pmod{3},\ 0\le i< s \}$, $k=1,2,3$. Let $m_k$ ($m'_k$) denote the vertex with the largest (second largest, respectively) index in $B_k$, for $k\in\{1,2,3\}$. We use the following patterns to colour the vertices of $B$. 1. For vertices of $B_1$, we use the pattern $16, 17, 18, 16, 17, 18, \ldots, 16, 17, 18$. If $|B_1|\equiv 1\pmod{3}$ (or $|B_1|\equiv 2\pmod{3}$), then we erase colours of $m_1$ (or of $m_1$ and $m'_1$, respectively). 2. For vertices of $B_2$, we use the pattern $$19, 20, 21, 25, 26, 19, 20, 21, 25, 26, \ldots, 19, 20, 21, 25, 26$$ when $\vert B_2\vert \equiv 0,1,2 \pmod{5}$, or the pattern $$19, 20, 21, 25, 26, 19, 20, 21, 25, 26, \ldots, 19, 20, 21, 25, 26, 19, 20, 21, 25, 19, 20, 21, 26$$ when $\vert B_2\vert \equiv 3,4 \pmod{5}$. Then, for $\vert B_2\vert \equiv 1,4 \pmod{5}$ we erase colour of $m_2$, and for $\vert B_2\vert \equiv 2 \pmod{5}$ we erase colours of $m_2$ and $m'_2$. 3. For vertices of $B_3$, we use the pattern $$22,23,24,27,28,22,23,24,27,28,\ldots,22,23,24,27,28$$ when $\vert B_3\vert \equiv 0,1,2 \pmod{5}$, or the pattern $$27,28,22,23,24,27,28,\ldots,22,23,24,27,28,22,23,24,27,22,23,24,28$$ when $\vert B_3\vert \equiv 3,4 \pmod{5}$. Then, for $\vert B_3\vert \equiv 1,4 \pmod{5}$ we erase colour of $m_3$, and for $\vert B_3\vert \equiv 2 \pmod{5}$ we erase colours of $m_3$ and $m'_3$. 4. When it is necessary, we use the colours $46$, $47$, $48$, $49$, $50$, $51$ to colour the vertices of $\{m_k,m'_k |\ 1\le k\le 3\}$. For checking that the defined colouring satisfies the distance constraints of a packing colouring, we recall that any two consecutive vertices in each $B_k$ ($k=1,2,3$) are pairwise at distance at least $7$, implying that vertices having the same colour are pairwise at distance at least $19$ in $B_1$ and at distance at least $31$ in $B_2$ and in $B_3$ (except some vertices of colours from $19$ to $24$ that can be at distance $25$ apart). Note that in some cases (depending on the size of $B$ and $C'$) we can decrease the upper bound $51$ of Theorem \[thm star structure\]. For example, if the internal face $C'$ has length $4k$ ($k\in\mathbb{N}$) and if the number of $2$-connected components of $G-C'$ is $15r$, $r\in \mathbb{N}$, then we can use only $40$ colours instead of $51$. The following statement will be used in the proof of Theorem \[thm path structure\]. \[pathgap\] There is a packing colouring of even vertices of $P_{\infty}$ with colours $\{k,k+1, \dots, 2k-1\}$. For the colouring of the vertices of the path $P_{\infty}$ we use pattern $1,k,1,k+1,1,k+2, \dots, 1,2k-1$ and after deleting colour $1$ we get the required colouring. Note that the distance between any pair of vertices coloured with the same colour in two consecutive copies of this pattern is $2k$. The last class of outerplanar graphs we consider in this paper is a class of not necessarily $2$-connected graphs. \[thm path structure\] Let $G$ be a connected subcubic outerplanar graph with no internal face such that the block graph of $G$ is a path. Then $\chi_{\rho}(G)\leq 305$. Let $G$ be a sucubic outerplanar graph, $B_G$ the block graph of $G$ and let $B_1, \dots, \, B_k$ denote the blocks of $G$ such that $B_i, B_{i+1}$ are consecutive in $B_G$, $i=1,\dots, k-1$ (i.e., since $\Delta(G)\le 3$, they are connected by a path which intersects no other block of $G$). Let $C_i$ denote the boundary cycle of $B_i$, $i=1,\dots, k$. Since $G$ contains no internal face, each $B_i$ contains no internal face as well, implying that every $B_i$ which is not a cycle contains exactly two end faces. Let $x_1$ denote any vertex of degree two (in $B_1$) in one end face of $B_1$, $x_k$ any vertex of degree two (in $B_k$) in one end face of $B_k$ and let $P$ denote a shortest $x_1,x_k$-path in $G$. Among all possible choices of the vertices $x_1,x_k$ we choose $x_1'$ and $x_k'$ such that the path $P$ is shortest possible. Let $P_i=P\cap B_i$, $i=1,\dots, k$. Obviously $P$ goes through all the blocks of $G$, hence $P_i$ is nonempty for each $i=1,\dots, k$, every $P_i$ is a path since the block graph of $G$ is a path, and each $P_i$ is shortest in $G$ since $P$ is shortest in $G$. In an orientation of $P$ from $x_1$ to $x_k$, we denote by $z_i$ the first vertex of $P$ in $B_i$ and by $z_i'$ the neighbour of $z_i$ in $B_i$ which does not belong to $P$. Note that there must be exactly one such vertex $z_i'$ in each $B_i$ since $P$ is shortest possible and, clearly, each $z_i$ must have degree three. For every block which is not a cycle, we further denote by $x_i$ any vertex of degree two (in $B_i$) in one end face of $B_i$, $i=2,\dots, k-1$ and by $y_i$ any vertex of degree two (in $B_i$) in the end face of $B_i$ which does not contain vertex $x_i$, $i=1,\dots, k$. units &lt;0.8mm,1mm&gt; x from -10 to 180, y from 20 to 90 axes ratio 2:7 360 degrees from 20 50 center at 10 50 at 10 82 1.5 64 18 67 / 0.3 55 20 51 / 0.2 45 18.9 38 / 3 30 16.6 29 / at 21 71 at -1 27 at 17.4 69 at 3.8 28 axes ratio 2:7 360 degrees from 50 50 center at 40 50 at 40 82 32 67 48.5 64 / 30.5 58 49.5 56 / 30 50 50 50 / 30.7 40 48.9 37 / 32 33 47 30 / at 29 72 at 50 26.5 at 32.7 69 at 46.4 28 axes ratio 2:7 360 degrees from 80 50 center at 70 50 at 70 82 63 70 78 67 / 60.5 58 79.8 55 / 60.7 40 79.6 44 / 62.1 33 78.5 36 / at 81 71.5 at 82 33 at 77.3 69.5 at 78.4 34 axes ratio 5:2 360 degrees from 90 50 center at 110 50 at 110 60 94 46.1 96 54.6 / 102 44.1 104 56 / 114 43.8 115 56.1 / 120 44.5 123.5 54.7 / at 88.5 45 at 128.5 57 at 92 46.9 at 125.5 54 at 145 50 axes ratio 2:7 360 degrees from 180 50 center at 170 50 at 170 82 161.5 64 178 67 / 160.3 55 180 51 / 160.2 45 178.9 38 / 163 30 176.6 29 / at 158 68 at 180 26 at 161.9 66 at 176 27 18 67.2 1.5 64.2 1.5 63.8 18.2 66.85 18.2 67 17.6 69 17.2 69 17.8 67 / axes ratio 2:7 40 degrees from 1.17 64 center at 10 50 axes ratio 2:7 41 degrees from 1.6 64 center at 10 50 0.2 45.2 18.9 38.2 18.9 37.8 0.2 44.8 / 18.8 38 19.25 41 19.65 41 19.2 37.8 / 19.35 40.75 30.3 55.75 30.3 56.25 19.25 41.25 19.25 41 30.3 56 / 30.05 56 30.25 58.2 49.5 56.2 49.5 55.8 30.65 57.8 30.45 56 / axes ratio 2:7 12.5 degrees from 49.8 50 center at 40 50 axes ratio 2:7 12.8 degrees from 50.2 49.8 center at 40 50 30 50.2 50 50.2 50 49.8 30 49.8 / axes ratio 2:7 21.7 degrees from 29.8 50.2 center at 40 50 axes ratio 2:7 21.8 degrees from 30.2 50 center at 40 50 30.5 39.8 48.9 36.8 48.9 37.2 30.7 40.2 30.7 40 / 48.7 37 49.1 40 49.5 40 49.1 36.8 / 49.3 39.8 60.25 41.8 60.5 39.75 79.6 43.8 79.6 44.2 60.9 40.2 60.65 42.2 49.3 40.2 49.3 40 60.45 42 / 79.6 44 79.4 41.8 79.8 42 80 44.2 / 79.6 41.8 96 41 106 43.7 / 79.6 41.96 96 41.17 105.7 43.75 / 79.4 41.64 96 40.83 106.3 43.65 / 106 43.52 110 43.42 110 43.78 105.6 43.88 105.8 43.7 110 43.6 / 110 43.6 120 40 136 50 / 109.75 43.6 120 39.82 136.25 50 / 110.25 43.6 120 40.18 135.75 50 / 155 49.75 160.2 55.75 160.2 56.25 155 50.25 155 50 160.2 56 / axes ratio 2:7 -24 degrees from 160.05 56 center at 170 50 axes ratio 2:7 -24 degrees from 160.4 55.8 center at 170 50 Among all possible choices of the vertices $x_i,y_i$ ($i=1,\dots, k$) we choose $x_i', y_i'$ in such a way that $x_i',P$-path and $y_i',P$-path, respectively, is shortest possible. For each $i=1,\dots, \, k$, let $Q_{2i-1}$ be a shortest $y'_i,P$-path in $G$, and let $Q_{2i}$ be a shortest $x'_i,P$-path in $G$. Note that $Q_2$ and $Q_{2k}$ are empty and that some other paths $Q_{i}$ may be trivial or empty (e.g., in the case when $B_i$ is a cycle). The structure of the graph $G$ is depicted in Fig. \[fig path structure\] (the thick path represents the path $P$) and the paths $Q_{2i-1}, Q_{2i}$ are illustrated in Fig. \[figadd44\]. Consider each block $B_i$ separately. Note that, for each $i=1,\dots, \,k$, the graph $G_i=G[B_i-(V(P_i)\cup V(Q_{2i-1})\cup V(Q_{2i}))]$ consists of path components. Thus, analogously as in the proof of Lemma \[lemma\], we can colour vertices of $G_i$ with colours $1,2,3,4$ using the periodic pattern $1,2,1,3$ and modifications introduced in the proof of Lemma \[lemma\], starting at the end face of $B_i$ containing vertex $x_i$. Note that, if $B_i$ is a cycle, $\chi_{\rho}(B_i)\leq 4$. Moreover, we can colour the vertices of $G_i$ in such a way that the vertex $z_i'$ gets colour $1$ (if not so, then we can interchange roles of $x_i$ and $y_i$ for colouring of $V(G_i)$, i.e., we start such a colouring from $y_i$ instead of $x_i$). Now we check that the defined colouring meets the conditions of a packing colouring. First, there is no collision in colours $1$ and $2$, since $\mbox{dist}_G(a,b)\geq 3$ for any $a\in V(G_i)$ and $b\in V(G_j), \, i<j$. From the modifications described in the proof of Lemma \[lemma\], it follows that no end vertex of any path component of any $G_i$ is coloured with colour $4$ since colour $4$ was used for a vertex between vertices of a critical pair belonging to one path component. This implies that the distance between two vertices coloured with $4$ which belong to different block $B_i,B_j$ is at least $5$, hence there is no collision in colour $4$. For colour $3$, since no $z_i'$ is coloured with colour $3$, there is no collision in colour $3$ as well. And since there is no edge in $G$ connecting the path components of the blocks $B_i$, the defined colouring meets the distance constraints of a packing colouring of $\bigcup\limits_{i=1}^k \left[ B_i-(V(P_i)\cup V(Q_{2i-1})\cup V(Q_{2i}))\right]$. (0,0) ellipse (6cm and 1cm); (0,3) ellipse (6cm and 1cm); (-1,1)– (-1,2); (-1,2)– (0.5,2); (1,4)– (0.5,2); (1,4)– (1.5,3.95); (2,3.9)– (1.5,3.95); (5,-0.55)– (5.5,-1.1); (2,3.9)– (2.5,4.4); (-1,1)– (-0.5,1); (-0.5,1)– (0,-1); (1.5,-1)– (0,-1); (1.5,-1)– (2,-0.95); (2.5,-0.9)– (2,-0.95); (2.5,-0.9)– (3,-0.85); (3.5,-0.8)– (3,-0.85); (3.5,-0.8)– (4,-0.75); (4.5,-0.65)– (4,-0.75); (4.5,-0.65)– (5,-0.55); (5.5,-0.4)– (5,-0.55); (5.5,-0.4)– (6,0); (5.5,3.4)– (5,3.55); (5,3.55)– (4.5,3.65); (4.5,3.65)– (4,3.75); (3.5,3.8)– (4,3.75); (3.5,3.8)– (3,3.85); (2.5,3.9)– (3,3.85); (2.5,3.9)– (2,3.95); (-1.5,2)– (-1,2); (-1.5,2)– (-2,2.05); (-2.5,2.1)– (-2,2.05); (-2.5,2.1)– (-3,3.85); (-3.5,3.8)– (-3,3.85); (-3.5,3.8)– (-4,3.75); (-4.5,3.65)– (-4,3.75); (-4.5,3.65)– (-5,3.55); (-1.5,1)– (-1,1); (-1.5,1)– (-2,0.95); (-2.5,0.9)– (-2,0.95); (-2.5,0.9)– (-3,0.85); (-3.5,0.8)– (-3,0.85); (-3.5,0.8)– (-4,0.75); (-4.5,0.65)– (-4,0.75); (-4.5,0.65)– (-5,0.55); (-5.5,0.4)– (-5,0.55); (-0.5,2)– (-1,4); (0.5,2)– (1,4); (-2,2.05)– (-1.5,4); (-2.5,2.1)– (-3,3.85); (-3.5,2.2)– (-4,3.75); (-5.5,2.6)– (-4.5,3.65); (2,2.1)– (1.5,4); (2.5,2.1)– (2.5,3.9); (3,2.15)– (3.5,3.8); (4,2.25)– (4,3.75); (5,2.45)– (4.5,3.65); (5.5,2.6)– (5,3.55); (0.5,2-3)– (1,4-3); (0,2-3)– (-0.5,4-3); (2,2.05-3)– (1.5,4-3); (2.5,2.1-3)– (3,3.85-3); (3.5,2.2-3)– (4,3.75-3); (5.5,2.6-3)– (4.5,3.65-3); (-2,2.1-3)– (-1.5,4-3); (-2.5,2.1-3)– (-2.5,3.9-3); (-3,2.15-3)– (-3.5,3.8-3); (-4,2.25-3)– (-4,3.75-3); (-5,2.45-3)– (-4.5,3.65-3); (-5.5,2.6-3)– (-5,3.55-3); at (0.5,2) \[circle,draw=black,fill=black,scale=0.5\]; at (0,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-0.5,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,2.05) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.5,2.1) \[regular polygon,regular polygon sides=3,draw=black,fill=green,scale=0.4\]; at (-3,2.15) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,2.25) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.5,2.35) \[circle,draw=black,fill=black,scale=0.5\]; at (-5,2.45) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,2.6) \[circle,draw=black,fill=black,scale=0.5\]; at (-6,3) \[circle,draw=black,fill=black,scale=0.5\]; at (1,2) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,2) \[circle,draw=black,fill=black,scale=0.5\]; at (2,2.05) \[circle,draw=black,fill=black,scale=0.5\]; at (2.5,2.1) \[circle,draw=black,fill=black,scale=0.5\]; at (3,2.15) \[circle,draw=black,fill=black,scale=0.5\]; at (3.5,2.2) \[circle,draw=black,fill=black,scale=0.5\]; at (4,2.25) \[circle,draw=black,fill=black,scale=0.5\]; at (4.5,2.35) \[circle,draw=black,fill=black,scale=0.5\]; at (5,2.45) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,2.6) \[circle,draw=black,fill=black,scale=0.5\]; at (6,3) \[circle,draw=black,fill=black,scale=0.5\]; at (0.5,4) \[circle,draw=black,fill=black,scale=0.5\]; at (0,4) \[circle,draw=black,fill=black,scale=0.5\]; at (-0.5,4) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,4) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,4) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,3.95) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.5,3.9) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,3.85) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,3.8) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,3.75) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.5,3.65) \[circle,draw=black,fill=black,scale=0.5\]; at (-5,3.55) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,3.4) \[circle,draw=black,fill=black,scale=0.5\]; at (1,4) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,4) \[circle,draw=black,fill=black,scale=0.5\]; at (2,3.95) \[circle,draw=black,fill=black,scale=0.5\]; at (2.5,3.9) \[circle,draw=black,fill=black,scale=0.5\]; at (3,3.85) \[rectangle,draw=black,fill=red,scale=0.7\]; at (3.5,3.8) \[circle,draw=black,fill=black,scale=0.5\]; at (4,3.75) \[circle,draw=black,fill=black,scale=0.5\]; at (4.5,3.65) \[circle,draw=black,fill=black,scale=0.5\]; at (5,3.55) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,3.4) \[circle,draw=black,fill=black,scale=0.5\]; at (0.5,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (0,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-0.5,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,2.05-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.5,2.1-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,2.15-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,2.2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,2.25-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.5,2.35-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-5,2.45-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,2.6-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-6,3-3) \[circle,draw=black,fill=black,scale=0.5\]; at (1,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (2,2.05-3) \[circle,draw=black,fill=black,scale=0.5\]; at (2.5,2.1-3) \[circle,draw=black,fill=black,scale=0.5\]; at (3,2.15-3) \[circle,draw=black,fill=black,scale=0.5\]; at (3.5,2.2-3) \[circle,draw=black,fill=black,scale=0.5\]; at (4,2.25-3) \[circle,draw=black,fill=black,scale=0.5\]; at (4.5,2.35-3) \[circle,draw=black,fill=black,scale=0.5\]; at (5,2.45-3) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,2.6-3) \[circle,draw=black,fill=black,scale=0.5\]; at (6,3-3) \[rectangle,draw=black,fill=red,scale=0.7\]; at (0.5,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (0,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-0.5,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,3.95-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-2.5,3.9-3) \[regular polygon,regular polygon sides=3,draw=black,fill=green,scale=0.4\]; at (-3,3.85-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-3.5,3.8-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,3.75-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4.5,3.65-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-5,3.55-3) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.5,3.4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (1,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (2,3.95-3) \[circle,draw=black,fill=black,scale=0.5\]; at (2.5,3.9-3) \[circle,draw=black,fill=black,scale=0.5\]; at (3,3.85-3) \[circle,draw=black,fill=black,scale=0.5\]; at (3.5,3.8-3) \[circle,draw=black,fill=black,scale=0.5\]; at (4,3.75-3) \[circle,draw=black,fill=black,scale=0.5\]; at (4.5,3.65-3) \[circle,draw=black,fill=black,scale=0.5\]; at (5,3.55-3) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,3.4-3) \[circle,draw=black,fill=black,scale=0.5\]; at (5.5,-1.1)\[circle,draw=black,fill=black,scale=0.5\]; at (2.5,4.4) \[circle,draw=black,fill=black,scale=0.5\]; at (-5.1,3.9) [$x'_i$]{}; at (0,3) [$B_i$]{}; at (5.5,3.7) [$y'_i$]{}; at (-5.5,0.7) [$x'_{i+1}$]{}; at (0.3,0) [$B_{i+1}$]{}; at (6.5,0) [$y'_{i+1}$]{}; at (5.5,-1.4) ; at (2.5,4.6) ; Now we colour the paths $Q_j$, $j=1,2,\dots, \, 2k$. The distance from $P$ to any vertex of $Q_j$ in $G$ is the same as on $Q_j$, hence each $Q_j$ is a shortest path in $G$ between $P$ and the relevant vertex $x'_i$ or $y'_i$, respectively. Thus, by Lemma \[out2\], we can colour the vertices of each path $Q_{2i-1}$ ($i=1,2,\dots, k$) with Pattern , starting at the vertex of $Q_{2i-1}$ at distance two from $P$ (square vertices in Fig. \[figadd44\]). Analogously we can colour the vertices of the paths $Q_{2i}$ ($i=1,2,\dots, k$) using Pattern , starting at the vertex of $Q_{2i}$ at distance three from $P$ (triangle vertices in Fig. \[figadd44\]). Then the distance between vertices on distinct paths $Q_m$, $Q_n$ coloured with colour $5$ is at least $3+1+2$, the distance between vertices on distinct paths $Q_m$, $Q_n$ coloured with colour $6$ is at least $4+1+3$, etc. Therefore the defined colouring of the paths $Q_j$, $j=1,\dots, 2k$ satisfies the distance constraints of a packing colouring. Now we colour the remaining vertices of $G$. We start with colouring of the path $P$ with a pattern using colours $16,17,\dots, 50$ by Proposition \[preliminaries\].ii). Then we colour all uncoloured vertices of $Q_j$ ($j=1,\dots, 2k$) at distance one from $P$ with colours $51, 52, \dots, 152$ by Proposition \[preliminaries\].iii). For the remaining vertices of $Q_{2i}$ at distance two from $P$, the distance between any such vertices on $Q_{2m}$ and $Q_{2n}$ ($m,n\in \{ 1,2,\dots, k-1 \}$, $m\neq n$) is at least $2\vert m-n\vert +4$. Hence we can colour these vertices with colours $153, \dots, 305$ by Proposition \[pathgap\]. Concluding remarks ================== In the previous sections, we have determined some classes of outerplanar graphs with finite packing chromatic number. As for lower bounds, we are (only) able to state the following, where symbols $\square$ and $\boxtimes$ denote the Cartesian and strong product of graph, respectively (see [@kla]). There exists an infinite family of $2$-connected subcubic outerplanar graphs without internal faces and with packing chromatic number $5$. It has been proven in [@God] that $\chi_{\rho}(G)=5$ for $G=P_n\square P_2$ and $n\ge6$. The graph $G$ illustrated in Figure \[lowerbound\] is a $2$-connected subcubic outerplanar graph with packing chromatic number $7$. We verified by computer that every proper colouring of $G$ with $6$ colours is not a packing colouring and we found a packing colouring of $G$ with $7$ colours. (-2,0) – (2,0); (-2,0) – (-1,2); (2,0) – (1,2); (-1,2) – (1,2); (-1,0) – (-1.5,1); (1,0) – (1.5,1); (-1,2) – (-2,3); (-2,5) – (-2,3); (-2,5) – (-1,6); (1,6) – (-1,6); (1,6) – (2,5); (2,3) – (2,5); (2,3) – (1,2); (-2,3) – (-4,2); (-2,5) – (-4,6); (-4,2) – (-4,6); (-4,3) – (-3,2.5); (-4,5) – (-3,5.5); (2,3) – (4,2); (2,5) – (4,6); (4,2) – (4,6); (4,3) – (3,2.5); (4,5) – (3,5.5); (1,6) – (2,8); (-1,6) – (-2,8); (-2,8) – (2,8); (-1,8) – (-1.5,7); (1,8) – (1.5,7); at (2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (1,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (1,0) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,1) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,2) \[circle,draw=black,fill=black,scale=0.5\]; at (2,8) \[circle,draw=black,fill=black,scale=0.5\]; at (1.5,7) \[circle,draw=black,fill=black,scale=0.5\]; at (1,6) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,8) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,8) \[circle,draw=black,fill=black,scale=0.5\]; at (1,8) \[circle,draw=black,fill=black,scale=0.5\]; at (-1.5,7) \[circle,draw=black,fill=black,scale=0.5\]; at (-1,6) \[circle,draw=black,fill=black,scale=0.5\]; at (2,3) \[circle,draw=black,fill=black,scale=0.5\]; at (2,5) \[circle,draw=black,fill=black,scale=0.5\]; at (4,2) \[circle,draw=black,fill=black,scale=0.5\]; at (4,6) \[circle,draw=black,fill=black,scale=0.5\]; at (4,3) \[circle,draw=black,fill=black,scale=0.5\]; at (4,5) \[circle,draw=black,fill=black,scale=0.5\]; at (3,5.5) \[circle,draw=black,fill=black,scale=0.5\]; at (3,2.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-2,5) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,2) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,6) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,3) \[circle,draw=black,fill=black,scale=0.5\]; at (-4,5) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,5.5) \[circle,draw=black,fill=black,scale=0.5\]; at (-3,2.5) \[circle,draw=black,fill=black,scale=0.5\]; Bre[š]{}ar et al. [@BrePa] have proven that for any finite graph $G$, the graph $G\boxtimes P_{\infty}$ has finite packing chromatic number. The degree of $G\boxtimes P_{\infty}$ can be arbitrary large. This property illustrates the fact that the degree of a graph is not the only parameter to consider in order to have finite packing chromatic number. Maybe the fact that the weak dual is a path (and is not any tree) helps to bound the packing chromatic number. It remains an open question to determine if the packing chromatic number of subcubic outerplanar graphs is finite or not. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank the referees for their judicious comments and Mahmoud Omidvar for his precious help, in particular for providing Pattern  of proof of Theorem \[out1\]. First author was partly supported by the Burgundy Council under grant \#CRB2011-9201AAO048S05587. Second author was partly supported by project P202/12/G061 of the Czech Science Foundation. [10]{} J. Balogh, A. Kostochka, and X. Liu: Packing chromatic number of cubic graphs. [*Discrete Appl. Math.*]{} **341**(2):474–483, 2018. J. A. Bondy and U.S.R. Murty: [*Graph Theory*]{}. Springer, 2008. B. Bre[š]{}ar, J. Ferme: An infinite family of subcubic graphs with unbounded packing chromatic number. [*Discrete Math.*]{} **341**:2337–2342, 2018. B.  Bre[š]{}ar, S. Klav[ž]{}ar, and D. F. Rall: On the packing chromatic number of [C]{}artesian products, hexagonal lattice, and trees. [*Discrete Appl. Math.*]{} **155**(17):2303–2311, 2007. B.  Bre[š]{}ar, S. Klav[ž]{}ar, and D. F. Rall: Packing [C]{}hromatic [N]{}umber of [B]{}ase-3 [S]{}ierpiński [G]{}raphs. [*Graphs and Combin.*]{} **32**(4):1315–1327, 2016. B.  Bre[š]{}ar, S. Klav[ž]{}ar, D. F. Rall, and K. Wash: Packing chromatic number under local changes in a graph. [*Discrete Math.*]{} **340**:1110–1115, 2017. B.  Bre[š]{}ar, S. Klav[ž]{}ar, D. F. Rall, and K. Wash: Packing chromatic number, $(1,1,2,2)$-colorings, and characterizing the Petersen graph. [*Aequat. Math.*]{} **91**:169-184, 2017. J. Ekstein, J. Fiala, P. Holub, and B. Lidický: The packing chromatic number of the square lattice is at least 12. *arXiv*:1003.2291v1, 2010. J. Fiala and P. A. Golovach: Complexity of the packing coloring problem for trees. *Discrete Appl. Math.* **158**:771–778, 2010. J. Fiala and P. A. Golovach: Complexity of the packing coloring problem for trees. [*Discrete Appl. Math.*]{}, **158**(7):771–778, 2010. J. Fiala, S. Klav[ž]{}ar, and B. Lidický: The packing chromatic number of infinite product graphs. [*European J. Combin.*]{} **30**(5):1101–1113, 2009. A.S. Finbow and D.F. Rall: On the packing chromatic number of some lattices. [*Discrete Appl. Math.*]{} **158**:1224-1228, 2010. N. Gastineau and O. Togni: $S$-Packing Colorings of Cubic Graphs. [*Discrete Math.*]{} **339**:2461–2470, 2016. W. Goddard, S. M. Hedetniemi, S. T. Hedetniemi, J. M. Harris, and D. F. Rall: Broadcast chromatic numbers of graphs. [*Ars Comb.*]{} [**86**]{}:33-49, 2008. R. Hammack, W. Himrich and S. Klavžar, Handbook of Graph Product, 2nd edition, CRC Press, 2011. J. Ekstein, P. Holub, and O. Togni: The packing colouring of distance graph $D(k,t)$. [*D*iscrete Appl. Math.]{} [**1**67]{}:100–106, 2014. Y. Jacobs, E. Jonck, and E. Joubert: A lower bound for the packing chromatic number of the Cartesin product of cycles. [*C*entr. Eur. J. Math.]{} [**1**1(7)]{}:1344–1357, 2013. D. Kor[ž]{}e and A. Vesel: On the packing chromatic number of square and hexagonal lattice. *Ars Mathematica Contemporanea* **7**:13–22, 2014. D. Kor[ž]{}e and A. Vesel: $(d,n)$-packing colorings of infinite lattices. [*D*iscrete Appl. Math.]{} **237**:97–108, 2018. B. Martin, F. Raimondi, T. Chen, and J. Martin: The packing chromatic number of the infinite square lattice is between 13 and 15. [*Discrete Appl. Math.*]{} **225**: 136–142, 2017. C. Sloper: Broadcast-coloring of trees. [*Reports in Informatics*]{} **233**:1-11, 2002. R. Soukal and P. Holub: A Note on Packing Chromatic Number of the Square Lattice. *Electron. J. of Combinatorics* **17**:447–468, 2010. O. Togni: On packing Colorings of Distance Graphs. *Discrete Appl. Math.* **167**:280–289, 2014.

ArXiv

--- author: - 'M.A. Thompson' - 'G.H. Macdonald' date: 'Received 8 April 2003 / Accepted 21 May 2003' title: A molecular line survey of the candidate massive Class 0 protostar IRAS 23385+6053 --- Introduction ============ The evolutionary process in which a high-mass star develops within a molecular cloud core is still subject to some uncertainty. Low-mass star formation, on the other hand, is underpinned by an extremely successful classification system (e.g. André et al. [@awb00]), stretching from the earliest protostellar phase (Class 0) to obscured young stellar objects and T-Tauri stars (Class II/III). No analogous scheme exists for high-mass protostars and Young Stellar Objects (YSOs), although a tentative scheme linking the development of hot molecular cores to that of ultracompact HII regions has recently been proposed (Kurtz et al. [@kcchw00]). In part this is due to the observational challenges posed by objects that are less numerous, are at much greater distances, preferentially form in compact clusters and evolve more rapidly than low-mass star-forming regions. Until recently, the earliest known phase in the development of a massive YSO was the ultracompact HII region (Churchwell [@churchwell02]), which traces the development of a high-mass star some 10$^{5}$ years after its formation (De Pree et al. [@dprg95]). Much effort in recent years has been placed into identifying the earlier phases of high mass star formation, i.e. the massive protostellar precursors of ultracompact HII regions (e.g. Molinari et al. [@mol96], [@mol98a], [@mol00]; Sridharan et al. [@sbsmw02], Beuther et al. [@bsmmsw02], Lumsden et al. [@lhor02]). A large number of candidate massive protostars and YSOs have been identified in these studies and the physical conditions of the pre-ultracompact HII region phase are beginning to emerge. One of the prototypes of these candidate massive protostars is , identified in the search of Molinari et al. ([@mol96], [@mol98a]). It is comprised of a compact dense molecular core, massing some 370 [$M_{\odot}$]{}, associated with a powerful molecular outflow seen in HCO$^{+}$ and SiO emission (Molinari et al. [@mol98b]). Although there is nearby extended radio and mid-infrared emission (Molinari et al. [@mol98a], [@mol98b]) associated with two YSO clusters, the dense molecular core of IRAS 23385+6053 is coincident with neither (see Fig. \[fig:msxfig\]) and is detected only by its far-infrared to sub-millimetre continuum emission. Molinari et al. ([@mol98b]) determined the sub-millimetre to bolometric luminosity ($L_{\rm submm}/L_{\rm bol}$) and envelope mass to core mass ($M_{\rm env}/M_{*}$) ratios for , which are standard diagnostic indicators for the evolutionary class of low-mass protostars (André et al. [@awb93]). Based on these indicators and the large bolometric luminosity ($\sim1.6\times10^{4}$ [$L_{\odot}$]{} for an assumed kinematic distance of 4.9 kpc) Molinari et al. ([@mol98b]) concluded that was the first bona fide example of a massive class 0 protostar. Although a wealth of information on the physical properties of massive protostars and YSOs is becoming available there is as yet little information on their chemistry or molecular inventory. The gas-phase molecular chemistry of star-forming regions has long been regarded as a potential diagnostic of their age (e.g. Brown et al. [@bcm88]). Many chemical studies have been made of hot molecular cores and ultracompact HII region (Gibb et al. [@gniwb00]; Thompson et al. [@tm99]; Hatchell et al. [@htmm98a], [@htmm98b], Schilke et al. [@sgbp97]), but none to date of candidate massive protostellar objects. Molecular line surveys allow the abundances of a large number of molecular species to be determined, characterising the chemistry of the molecular gas and providing a list of molecular tracers that may be used to probe the physical conditions of massive protostars (e.g. to determine the gas kinetic temperature or density, or for use in infall searches). For these reasons we have carried out a molecular line survey of the candidate massive protostar IRAS 23385+6053 with the James Clerk Maxwell Telescope. The observations and data reduction procedure are detailed in the next section. The identified molecular lines detected in the survey and their derived column densities and abundances are presented in Sect. \[sect:analysis\]. The results of the survey and their implications for the protostellar nature of and its relation to the more evolved hot molecular cores and ultracompact HII regions are discussed in Sect. \[sect:discuss\]. Observations and data reduction =============================== Observations ------------ The observations were carried out at the James Clerk Maxwell Telescope (JCMT[^1]) during the August 1999 to January 2000 semester. The observations were scheduled according to the flexible-scheduling backup mode in operation at the JCMT, whereby scientific programmes are accorded a priority within a specified weather band defined by the atmospheric opacity at 225 GHz and then carried out by visiting observers according to suitable weather conditions. The data were thus obtained piecemeal over a period of several months, depending upon the local weather conditions and the visibility of the target source . The 330–360 GHz region was selected for the molecular line survey to match as closely as possible the FWHM beam-width of the JCMT at this frequency (14 at 345 GHz) to the observed angular diameter of the molecular core (FWHM $\sim 5$) as measured in the HCO$^{+}$ J=1–0 transition, Molinari et al. [@mol98b]). This approach was crucial to minimise the beam dilution caused by observing such a compact source and also to avoid possible contamination from gas associated with the nearby mid-infrared sources (see Fig. \[fig:msxfig\]). A secondary concern for choosing this frequency range was to facilitate the comparison of the survey results with the many surveys of hot cores and ultracompact HII regions carried out in this range (Thompson et al. [@tm99]; Schilke et al. [@sgbp97]; Macdonald et al. [@mghm96], Jewell et al. [@jhls89]). ![2MASS $K_{s}$ band image of the region surrounding . The contours represent the 8 $\mu$m emission measured by MSX, with a starting contour level of $10^{-6}$ Wm$^{-2}$sr$^{-1}$ and contour spacing of $10^{-6}$ Wm$^{-2}$sr$^{-1}$. The position of the millimetre peak identified by Molinari et al.([@mol98b]) is marked with a vertical cross (+) and the two adjacent radio sources and are shown by diagonal crosses ($\times$). The molecular line survey position and beam FWHM is indicated by a white circle.[]{data-label="fig:msxfig"}](h4440f1.ps) All observations were made at the coordinates of the millimetre peak of given by Molinari et al. ([@mol98b]), i.e. $\alpha$(2000) = 23$^{\rm h}$40$^{\rm m}$545, $\delta$(2000) = +611028. The pointing accuracy of the telescope was checked hourly against standard pointing calibrators and was found to be good to within 3–4. The spectra were taken in beam-switching mode, in which the secondary mirror is chopped from on-source to off-source at a frequency of 1 Hz. Beam-switching is much superior to position-switching (i.e. moving the primary mirror to the off-source position) for obtaining the extremely flat baselines required to search for faint line emission. A chop throw of 3 in RA was used to keep a constant reference position from spectrum to spectrum. This throw was more than sufficient to reach a clean reference position for all species except CO, which in massive star-forming regions is typically widespread over a much larger area. The spectra were observed using the facility heterodyne receiver B3 (RxB3), which is a dual-SIS junction receiver operating in the 345 GHz band, coupled to the facility Digital Autocorrelation Spectrometer (DAS). The receiver was used in dual sideband mode in which the resulting spectra are comprised of two frequency bands (the upper and lower sidebands) folded over one another. The “main band” of the spectrum may be set to either the upper or lower sideband and the remaining sideband is often referred to as the “image band”. The main and image bands are separated from one another by roughly twice the intermediate frequency or IF of the receiver, which for RxB3 is 4 GHz, leading to a band separation of $\sim$ 8 GHz. The receiver may also be used in single-sideband mode where the image band of the receiver is attenuated by roughly a factor of $\sim$ 20 using a dual-beam Mach-Zender interferometer. Strong lines in the image band may thus leak through to the main band, mimicking faint main band lines and the sideband rejection factor is not accurately calibrated over the entire passband of the receiver. We used the dual-sideband mode of receiver B3 to avoid problems in relative image band rejection across the passband and to instantaneously obtain twice the frequency coverage of single-sideband mode. The sideband of each detected line was determined by taking an additional spectrum with the local oscillator frequency shifted by +10 MHz. In the shifted spectrum lines in the image band appear to shift frequency by $\pm20$ MHz relative to the lines in the main band. The DAS was set to a bandwidth of 920 MHz to exploit the maximum dual-sideband bandwidth of RxB3. Our observing strategy was to take spectra with the main band set to the lower sideband, increment the central frequency of the main band by 800 MHz and repeat the spectra until the frequency range covered in the lower sideband reached that of the upper sideband from the first spectrum. A block of 10 spectra observed in this manner sample contiguous 8 GHz frequency ranges in both the lower and upper sidebands. Two of these blocks of dual-sideband spectra are sufficient to cover the entire 330–360 GHz frequency range, with a 1 GHz extension to either side. The 800 MHz frequency increment allows for a 120 MHz overlap between spectra, allowing the the raised noise at the edge of the passband to be avoided. Unfortunately, during the observations the second block of spectra was mistakenly observed in the the upper sideband rather than the lower sideband, leading to a duplication of the 338–342 GHz frequency range and a gap in the coverage between 352–359 GHz. An attempt was made to take extra spectra to cover the gaps in frequency coverage, although the lack of suitable weather over the semester meant that the gaps could not be completely filled. The resulting frequency coverage of the survey is 27.2 GHz with gaps at 334.6–337, 344.2–345 and 352-358.6 GHz. Fig. \[fig:coverage\] shows the total frequency coverage of the survey. Despite the gaps in frequency coverage the molecular line survey is 85% complete over the 329–361 GHz range originally intended and covers the frequencies of many important astrophysical transitions. The implications of the gaps in frequency coverage and their impact on the survey are discussed further in Sect. \[sect:discuss\]. With the DAS set to a 920 MHz bandwidth the width of each channel was 625 kHz, with a spectral resolution of 756 kHz. Later in the data reduction procedure all spectra were binned to a channel width of 1.25 MHz to improve the signal to noise ratio. The standard three-load chopper-wheel calibration method of Kutner & Ulich ([@ku81]) was used to obtain line temperatures on the [$T_{A}^{*}$]{} scale, i.e. corrected for the atmosphere, resistive telecope losses and rearward spillover and scattering. All line temperatures quoted in this paper are on the [$T_{A}^{*}$]{} scale unless explicitly stated otherwise. Values of [$T_{A}^{*}$]{} may be converted to the corrected receiver temperature [$T_{R}^{*}$]{} by dividing by the forward spillover and scattering coefficient [$\eta_{\rm fss}$]{}, which for the JCMT and RxB3 is 0.7 at 345 GHz. Absolute calibration of the [$T_{A}^{*}$]{} scale was determined by regular observations of the standard spectral line calibrator NGC 7538 IRS1 and was found to be accurate to within 10%. Data reduction and line identification -------------------------------------- The data were reduced using the Starlink millimetre-wave spectroscopy package SPECX (Prestage et al. [@specx]). Linear baselines were subtracted from each spectrum and the line parameters of peak temperature ([$T_{A}^{*}$]{}), central observed frequency ($\nu$(obs)) and line width at half maximum ($\Delta\nu_{1/2}$) were measured. A value of $-$51.0 kms$^{-1}$ for the LSR velocity of was assumed, following Molinari et al. ([@mol98b]). Sample spectra are shown in Figure \[fig:spectra\]. The noise in each spectrum was estimated using the line-free channels and was found to be typically 50 mK per 1.25 MHz wide channel. Line features below a minimum detection limit of 3$\sigma$ (0.15 K) were ignored to avoid inaccurate line identifications. Each line was identified at least twice in separate spectra, including the 10MHz shifted spectra and the 120 MHz overlaps between adjacent spectra. The values of [$T_{A}^{*}$]{}, $\nu$(obs) and $\Delta\nu_{1/2}$ were averaged together for multiple detections of the same line and these average quantities are listed in Table \[tbl:linetable\]. ------------ ------------- ------------------- ------------------ --------------------- ------------- --------------- ------------------------------------- $\nu$(obs) $T_{A}^{*}$ $\Delta\nu_{1/2}$ Species Transition $\nu$(rest) $E_{\rm u}/k$ Notes (GHz) (K) (MHz) (GHz) (K) 330.587 6.89 4.4 $^{13}$CO 3–2 330.588 31.7 337.060 0.90 2.7 C$^{17}$O 3–2 337.061 32.4 338.124 0.33 3.3 CH$_{3}$OH 7(0)–6(0) E 338.125 76.9 338.344 0.73 4.6 CH$_{3}$OH 7($-$1)–6($-$1) E 338.345 69.4 338.408 0.89 4.6 CH$_{3}$OH 7(0)–6(0) A+ 338.409 65.0 338.515 0.21 3.9 CH$_{3}$OH 7(2)–6(2) A$-$ 338.513 103 338.615 0.22 3.9 CH$_{3}$OH 7(1)–6(1) E 338.615 84.9 338.722 0.27 3.6 CH$_{3}$OH 7(2)–6(2) E 338.722 86.1 340.032 0.25 2.8 CN 3–2 2.5 3.5–1.5 2.5 340.032 16.3 sl-blend 340.035 0.24 2.8 CN 3–2 2.5 1.5–1.5 0.5 340.035 16.3 h/fine blend with 2.5 2.5-1.5 1.5 340.247 0.59 3.5 CN 3–2 3.5 4.5–2.5 3.5 340.248 16.3 340.714 0.26 3.8 SO 7(8)–6(7) 340.714 81.2 341.350 0.15 2.4 HCS$^{+}$ 8–7 341.350 73.7 341.415 0.38 4.2 CH$_{3}$OH 7(1)–6(1) A$-$ 341.416 80.1 342.883 1.25 3.8 CS 7–6 342.883 65.8 345.339 0.15 3.7 H$^{13}$CN 4–3 345.340 41.3 \* SO$_{2}$ 13(2,12)–12(2,11) 345.339 93.0 345.795 12.19 – CO 3–2 345.796 33.2 strongly self-absorbed 346.527 0.35 4.7 SO 9(8)–8(7) 346.529 78.8 346.998 0.19 4.4 H$^{13}$CO$^{+}$ 4–3 346.999 41.6 347.387 0.17 5.0 U 348.534 0.15 5.0 H$_{2}$CS 10(1,9)–9(1,8) 348.532 105 349.338 0.92 4.9 C$_{2}$H 4.5–3.5 349.338 41.9 349.400 0.61 3.7 C$_{2}$H 3.5–2.5 349.401 41.9 350.688 0.47 4.4 CH$_{3}$OH 4(0)–3($-$1) E 350.688 35.1 poss. blend with NO 3.5 0.5–2.5 0.5 350.905 0.38 4.9 CH$_{3}$OH 1(1)–0(0) A+ 350.905 16.8 351.768 1.60 4.9 H$_{2}$CO 5(1,5)–4(1,4) 351.769 62.5 358.606 0.23 3.9 CH$_{3}$OH 4(1)–3(0) E 358.606 43.1 ------------ ------------- ------------------- ------------------ --------------------- ------------- --------------- ------------------------------------- The lines were identified with molecular transitions by comparing their central observed frequencies to those listed in the JPL Molecular Spectroscopy Database (available from `http://spec.jpl.nasa.gov`). Other line lists used include the methanol lists of Anderson et al. ([@anderson]) and the observational lists of lines detected by Jewell et al. ([@jhls89]) toward Orion-A, Macdonald et al. ([@mghm96]) toward G34.3+0.15, Schilke et al. ([@sgbp97]) toward Orion-KL and Thompson & Macdonald ([@tm99]) toward G5.89$-$0.39. A list of species identified in the molecular gas of may be found in Table \[tbl:speciestable\]. The identification of asymmetric rotors (such as H$_{2}$CS) from single line detections must be viewed with caution as these species possess many other transitions within the observed frequency range. ------------------ ---------------- Species Number of detected lines CO 1 $^{13}$CO 1 C$^{17}$O 1 CH$_{3}$OH 10 CN 3 SO 2 HCS$^{+}$ 1 CS 1 H$^{13}$CN 1 H$^{13}$CO$^{+}$ 1 H$_{2}$CS 1 C$_{2}$H 2 H$_{2}$CO 1 ------------------ ---------------- : Molecular species (including isotopomers) identified in the survey. We note that the identification of certain species (e.g. H$_{2}$CS) from single line detections must be regarded with caution.[]{data-label="tbl:speciestable"} Results and analysis {#sect:analysis} ==================== Identified molecular lines -------------------------- A total of 27 lines were identified in the survey, originating from 11 species, and are shown in Table \[tbl:linetable\]. One further line at 347.387 GHz could not be identified with any transition listed in the JPL Molecular Spectroscopy Database and is also not listed in any of the other observational lists that were checked. A list of the identified species (including isotopomers) and the number of detected lines from each is contained in Table \[tbl:speciestable\]. Most of the lines identified in the survey originate from the J=7–6 transitions of methanol (CH$_{3}$OH). With the exception of the $^{12}$CO J=3–2 line the lines are narrow, with typical FWHM linewidths of $\sim$ 4 MHz, which corresponds to a width of 3.5 kms$^{-1}$ at 345 GHz. Emission from the commonly used tracers of temperature and column density CH$_{3}$CN (methyl cyanide) and CH$_{3}$CCH (propyne or methyl acetylene) was not detected down to a level of $\sim 0.15$ K. If the most likely identification of the line at 345.399 GHz is the J=4–3 transition of H$^{13}$CN then the SO$_{2}$ molecule was also not detected toward . No transitions from the organic asymmetric rotors (e.g. CH$_{3}$OCH$_{3}$, HCOOCH$_{3}$, NH$_{2}$CHO, H$_{2}$CCHCN or C$_{2}$H$_{5}$OH) observed to be widespread in hot molecular cores (e.g. Hatchell et al.[@htmm98a]) were detected. The non-detections of these various transitions most likely arise from a mix of excitation and chemical effects. The narrow linewidths and faint molecular lines observed in the survey indicate that the molecular gas of is cold, meaning that there is an insufficient kinetic temperature to adequately excite the higher energy transitions. High excitation lines require hot gas in order to be excited. The highest excitation transition observed in the survey was the 7(2)–6(2) A$-$ CH$_{3}$OH line at 338.515 GHz, which possesses an upper energy level of $E_{\rm u}/k = 103$ K. It is likely that the non-detection of certain species whose lines in this frequency range are of high excitation (e.g. HC$_{3}$N with an $E_{\rm u}$/k $\sim$ 300 K) is reflected by the inability of the cold gas to excite their transitions adequately rather than a genuinely low gas-phase abundance. However, asymmetric rotors such as SO$_{2}$, HCOOCH$_{3}$ and C$_{2}$H$_{5}$OH possess many lines across the frequency range of the survey with a wide range of excitation. The non-detection of the low-excitation lines of these species must arise from their low gas-phase abundance rather than excitation effects. We will further explore these issues and their implications for the chemistry of in Sect. \[sect:discuss\]. Rotation diagram analysis ------------------------- The temperature and column density of the gas may be determined via rotation diagrams (Turner [@turner91]), also known as Boltzmann plots (e.g. Brand et al. [@bcpm01]). The main assumptions in the rotation diagram approach are that the gas is optically thin, in Local Thermodynamic Equilibrium (LTE) and can be described by a single rotational temperature $T_{\rm rot}$. For these assumptions the column density equation may be written as (e.g. Thompson, Macdonald & Millar [@tmm99]) $$\label{eqn:coldensity} N = \frac{3k}{8\pi^{3}}\, \frac{\int T_{R} \,\,{\rm d}v}{\nu S \mu^{2} g_{I}g_{K}} \,Q(T_{\rm rot})\,{\rm exp}\left(\frac{E_{\rm u}}{k T_{\rm rot}}\right),$$ where $\int T_{R} \,\,{\rm d}v$ is the integrated intensity of the line, $\nu$ is the line frequency, $S$ is the line strength, $\mu$ is the permanent dipole moment, $g_{I}$ and $g_{K}$ are, respectively, the reduced nuclear spin and K-level degeneracy of the molecule. The energy of the upper level of the line is represented by $E_{\rm u}$. The partition function as a function of $T_{\rm rot}$ is written as $Q(T_{\rm rot})$ and represents the partitioning of the total level energies into each rotational level. The column density $N$ derived from Eq. \[eqn:coldensity\] is beam-averaged rather than source-averaged, as the beam-averaged receiver temperature [$T_{R}^{*}$]{} is used instead of the source brightness temperature $T_{\rm b}$. To calculate a source-averaged column density the beam-filling factor of the emission (and hence the emitting area of the line) must be known. Each line will trace slightly different environments depending upon their critical densities and excitation temperature, thus there is no guarantee that the beam filling factors are constant for all species and lines. As the beam filling factors are unknown for the transitions in question we derive beam-averaged column densities throughout this paper. Equation \[eqn:coldensity\] is often known as the column density equation and, by taking logarithms of each side, it is possible to rearrange this equation to that of a straight line, i.e. $$\label{eqn:rotdiag} \log \left(\frac{3k}{8\pi^{3}} \frac{\int T_{R} \,\,{\rm d}v}{\nu S \mu^{2} g_{I}g_{K}} \right) = \log\left(\frac{N}{Q(T_{\rm rot})}\right) - \frac{E_{\rm u}}{k}\,\frac{\log e}{T_{\rm rot}}$$ This equation is more commonly known as the rotation diagram equation (e.g. Macdonald et al. [@mghm96]). The left hand quantity in Equation \[eqn:rotdiag\] is more commonly known as $L$ and a plot of $L$ against $E_{\rm u}/k$ results in a straight line with a gradient of $-log\,\,e/T_{\rm rot}$ and an intercept of $\log\left(N/Q(T_{\rm rot})\right)$. Methanol is the only species in our survey with a sufficient number of identified lines to plot rotation diagrams. Methanol is an asymmetric top molecule with its rotation-torsion levels split into two types of sublevel: A (symmetric) and E (degenerate). Both types must be analysed separately to take into account differences in sublevel populations. We plotted separate rotation diagrams for the A and E-type methanol transitions identified in the survey, which may be seen in Fig. \[fig:meth\_rotdiag\]. Straight lines were fitted to the data using a least-squares fitting routine. The results of the straight line fits for the A-type transitions are $T_{\rm rot} = 52 \pm 20$ K and $N = 2.4 \pm 1.5\,\,10^{14}$ cm$^{-2}$. The fits to the E-type transitions yield a rotation temperature and column density of $T_{\rm rot} = 16 \pm 6$ K and $N = 9.4 \pm 8.9\,\,10^{14}$ cm$^{-2}$. The partition function $Q(T_{\rm rot})$ was evaluated via interpolation from the values quoted in the JPL Molecular Spectroscopy Database to the rotation temperature $T_{\rm rot}$ derived from the rotation diagram. The column densities were then determined using Eq. \[eqn:rotdiag\] and the y-intercept value from the least-squares fit. ![Rotation diagrams of the A and E forms of methanol. Solid lines represent linear least-squares fits to all the data points, whereas dashed lines are least-squares fits to lines with $E_{\rm u}/k < 50$ K (as mentioned in the text).[]{data-label="fig:meth_rotdiag"}](h4440f4.ps) What can be immediately seen from the two rotation diagrams in Figure \[fig:meth\_rotdiag\] is that the lower excitation transitions are inconsistent with the straight line fits. When transitions with values of $E_{\rm u}/k$ less than 50 K are excluded from the straight line fits then the straight line correlation coefficient improves markedly for both rotation diagrams. We carefully inspected the data for any irregularities (e.g. pointing offsets, line misidentifications or blending, calibration errors) which may have caused the integrated intensity or upper level energy of any of the methanol lines to be over- or under-estimated. No such cause was found and we conclude that it is likely to be a genuine physical effect rather than an artefact of the observations. Discounting the low excitation lines with $E_{\rm u}/k \leq 50$ K from the straight line fit brings the rotation temperatures and column densities derived from each subtype into closer agreement, with $T_{\rm rot} \simeq 20$ K and $N \simeq 10^{15}$ cm$^{-2}$. These straight line fits are indicated in Fig. \[fig:meth\_rotdiag\] by dashed lines. The most likely cause of the difference in the low-excitation transitions is a breakdown in the assumptions behind the rotation diagram approach, i.e. the gas may not be described by a single temperature or it is optically thick. If the former were true then the low excitation lines would predominantly originate from colder gas than the high excitation lines and one would expect the low-excitation lines to follow a steeper (i.e. colder) gradient. If the latter were true then the integrated intensity of the more optically thick lower excitation lines would be underestimated compared with the less optically thick higher excitation lines. It is unlikely that the methanol lines are appreciably optically thick, as they are relatively faint and show no evidence of the line core saturation or asymmetric line profiles expected for optically thick lines. We do not detect the low excitation lines of the $^{13}$CH$_{3}$OH isotopomer, although the upper limit on their detection ([$T_{A}^{*}$]{}$\leq 0.15$ K) does not allow stringent limits to be placed on the optical depth of the low excitation lines ($\tau_{\rm max} \leq 20$). There is some evidence for a two-temperature distribution in the E-type rotation diagram, as the two lowest excitation lines appear to follow a steeper gradient. Further observations of lower excitation lines, for example in the 210–280 GHz window, are required to confirm this hypothesis. Lower limits to column density ------------------------------ Eq. \[eqn:coldensity\] may be used to yield lower limits to the beam-averaged column density for species with only one or two detected lines. To derive a lower limit to the beam-averaged column density we set the derivative of the temperature-dependent part of Eq. \[eqn:coldensity\] to zero, i.e. $$\frac{\rm d}{{\rm d}T_{\rm ex}}\left[Q(T_{\rm ex})\,\exp(E_{\rm u}/kT_{\rm ex})\right] = 0$$ where $T_{\rm ex}$ represents the excitation temperature of the transition. Values of $T_{\rm ex}$ satisfying this condition are turning points in the column density equation. Using the Turner ([@turner91]) high temperature approximations for the partition function $Q(T_{\rm ex})$, it is easy to show that these turning points are minima (as the second derivative is positive) and also that they occur at values of $T_{\rm ex} = E_{\rm u}/k$ for linear molecules and $T_{\rm ex} = 2\,E_{\rm u}/3\,k$ for symmetric and asymmetric top molecules. The resulting equations for the minimum beam-averaged column density may thus be written as: $$\begin{aligned} \label{eqn:nmin_lin} N_{\rm min} & = & \frac{3k}{8\pi^{3}}\,\frac{\int T_{R} {\rm d}v}{\nu S\mu^{2}g_{I}g_{K}}\,Q\left(E_{\rm u}/k\right)\, e \\ \label{eqn:nmin_asymm} N_{\rm min} & = & \frac{3k}{8\pi^{3}}\,\frac{\int T_{R} {\rm d}v}{\nu S\mu^{2}g_{I}g_{K}}\,Q\left(2E_{\rm u}/3k\right)\, e^{3/2} \end{aligned}$$ where Eq. \[eqn:nmin\_lin\] is used for linear molecules such as CO and HC$_{3}$N and Eq. \[eqn:nmin\_asymm\] is appropriate for symmetric and asymmetric top molecules such as CH$_{3}$CN and H$_{2}$CS. We used Equations \[eqn:nmin\_lin\] and \[eqn:nmin\_asymm\] to determine the beam-averaged column density lower limits ($N_{\rm min}$) for the lines detected in the survey. Values of $N_{\rm min}$ are shown in Table \[tbl:nmin\]. Self-absorbed and blended lines were excluded from the analysis. Species N$_{\rm min}$ (cm$^{-2}$) ------------------ --------------------------- $^{13}$CO 1.910$^{16}$ C$^{17}$O 1.5$10^{15}$ SO 2.710$^{13}$ HCS$^{+}$ 1.610$^{12}$ CS 1.710$^{13}$ H$^{13}$CN 1.410$^{12}$ H$^{13}$CO$^{+}$ 5.110$^{11}$ H$_{2}$CS 1.810$^{13}$ C$_{2}$H 1.610$^{14}$ H$_{2}$CO 4.110$^{13}$ : Lower limits to column density for the molecular species detected in the survey.[]{data-label="tbl:nmin"} Upper limits to column density for undetected species {#sect:uplim} ----------------------------------------------------- The column density equation (Eq. \[eqn:coldensity\]) may be used to determine an upper limit to the beam-averaged column density of molecular species that were not detected in the survey (e.g. Hatchell et al. [@htmm98a]). In this case assumed values for the maximum integrated intensity of an undetected line and the excitation temperature of the gas must be used in place of $\int T_{R}\,\,{\rm d}v$ and $T_{\rm rot}$. The maximum integrated intensity for an undetected line is estimated from a 3$\sigma$ upper limit to the receiver temperature and an assumed value for the linewidth. For all the upper limits derived in this paper we take the assumed linewidth to be the average linewidth of the detected lines, i.e. 4 MHz or 3.5 kms$^{-1}$ at 345 GHz. We calculate upper limits to column density for two assumed temperatures (20 and 40 K), to bracket the rotation temperature of $\sim$ 20 K as derived from the methanol rotation diagrams and also the values derived by Molinari et al. ([@mol96], [@mol98b]) from ammonia and millimetre-wave dust continuum observations (27 and 40 K respectively). It should be noted that the latter temperature estimate from Molinari et al. ([@mol98b]) is of the temperature at the outer radius of the millimetre-wave core. Upper column density limits were calculated for various astrochemically important species that have been observed towards hot molecular cores and ultracompact HII regions, so that we may investigate the evolution of the molecular gas towards these later stages. We also selected species that are contained in chemical models of these regions (e.g. Viti & Williams [@vw99]). The upper column density limits are given in Table \[tbl:nmax\], along with the line used for the calculation and its rest frequency. -------------- --------------------- ------------- --------------------- --------------------- Species Line $\nu$(rest) (GHz) $T_{\rm ex} = 20$ K $T_{\rm ex} = 40$ K CH$_{3}$CN 18(0)–17(0) 331.072 2.910$^{15}$ 1.910$^{14}$ CH$_{3}$CCH 20(0)–19(0) 341.741 1.010$^{16}$ 3.710$^{14}$ HCOOCH$_{3}$ 17(5,12)–16(4,13) A 343.150 7.910$^{14}$ 2.310$^{14}$ HC$_{3}$N 38–37 345.609 1.910$^{18}$ 1.210$^{15}$ SO$_{2}$ 4(3,1)–3(2,2) 332.505 3.610$^{13}$ 4.610$^{13}$ OCS 28–27 340.449 1.510$^{18}$ 7.810$^{15}$ -------------- --------------------- ------------- --------------------- --------------------- It can be seen from Table \[tbl:nmax\] that upper limits for species with only high-excitation lines ($E_{\rm u}/k \ge 150$ K) are not well constrained for the $T_{\rm ex} = 20$ K case. This is due to the exponential term in Eq. \[eqn:coldensity\], which significantly raises the column density upper limit for lines possessing large values of $E_{\rm u}/k$ in relation to the excitation temperature of the gas. As the assumed excitation temperature rises the derived column density upper limit decreases, until the point where the partition function term $Q(T_{\rm ex})$ begins to dominate over the exponential ($T_{\rm ex} \simeq Eu/k$). The low temperatures used in the calculation may not be appropriate for species that are thought to predominantly originate from the evaporation of dust grain ice mantles (CH$_{3}$CN and HCOOCH$_{3}$). These species, if they are present in the gas phase, are expected to have evaporated from the ice mantles at gas kinetic temperatures above $\sim$ 90 K (e.g. Hatchell et al. [@htmm98a]). Purely gas-phase chemical models predict extremely low abundances of these molecules (Millar et al. [@mmg97]). In order to account for the possibility that contains a small embedded hot molecular core heated by the central protostar we have evaluated column density upper limits for CH$_{3}$CN and HCOOCH$_{3}$ assuming an excitation temperature of 150 K. These limits are: $N_{\rm max} = 8.8\,\,10^{13}$ for CH$_{3}$CN and $N_{\rm max} = 3.1\,\,10^{14}$ for HCOOCH$_{3}$. We will investigate the likelihood that contains an embedded hot molecular core and set limits on its possible size in Sect. \[sect:hotcore\]. Discussion {#sect:discuss} ========== The molecular inventory of --------------------------- The results of our survey show that has a simple molecular inventory. The gas is comprised of simple molecules with emission detected from CO, CN, SO, HCS$^{+}$, CS, HCN, HCO$^{+}$, C$_{2}$H, H$_{2}$CO, CH$_{3}$OH and H$_{2}$CS. The latter molecule, H$_{2}$CS, is an uncertain identification as several other low-excitation lines of this asymmetric top molecule lying in the frequency range of the survey were not detected. Only low-excitation lines of these species were detected with values of $E_{\rm u}/k \le 100$ K. To this inventory we may add the following species observed toward by other authors: NH$_{3}$ (Molinari et al. [@mol96]), SiO (Molinari et al. [@mol98b]) and CH$_{3}$CCH (Brand et al. [@bcpm01]). We did not detect these last two species (SiO and CH$_{3}$CCH) in our survey due to a combination of excitation and the sensitivity of our observations. The lines of these species in the 330–360 GHz frequency range are of moderate to high excitation; $E_{\rm u}/k = 75$ K in the case of the SiO 8–7 line and $E_{\rm u}/k = 179$ K for the CH$_{3}$CCH 20(0)–19(0) line. The upper limit that we derived for CH$_{3}$CCH in Sect. \[sect:uplim\] for an excitation temperature of 40 K is larger than the column density measured by Brand et al. ([@bcpm01]) for the CH$_{3}$CCH J=6–5 K-ladder lines by roughly an order of magnitude and thus our observations are consistent with those of Brand et al. ([@bcpm01]). Molinari et al.([@mol98b]) detected the SiO 2–1 line with a source-integrated column density of 1.410$^{14}$ cm$^{-2}$ for an assumed excitation temperature of 30 K. Using Eq. \[eqn:coldensity\] we calculate a beam-averaged upper column density limit (based upon our non-detection of the SiO J=8–7 line at 347.331 GHz) of 1.010$^{12}$ cm$^{-2}$. This is inconsistent with the Molinari et al. ([@mol98b]) detection and indicates that either the assumed excitation temperature is too large or that the emitting area of the SiO J=8–7 line is small enough for the beam-dilution in our 14 beam to reduce the emission below our sensitivity limit ([$T_{A}^{*}$]{}(rms) = 43 mK). We did not detect emission from any of the complex organic molecules whose high abundances are a classic signature of hot molecular cores (e.g. CH$_{3}$CN, HCOOCH$_{3}$ or C$_{2}$H$_{5}$OH). These molecules are predominantly thought to originate from the evaporation of dust grain ice mantles and their non-detection towards indicates that the presence of a hot molecular core within is unlikely. A more sensitive search for the hot core tracer methyl cyanide (CH$_{3}$CN) was conducted recently by Pankonin et al. ([@pcwb01]) towards a number of candidate massive protostars. Their non-detection of emission from the CH$_{3}$CN J=12–11 K-ladder to a 1$\sigma$ r.m.s. level of [$T_{A}^{*}$]{} = 13mK allows us to set an upper limit to the CH$_{3}$CN column density of 5.010$^{12}$ cm$^{-2}$ for an assumed excitation temperature of 40 K (a factor of $\sim 40$ lower than our limit based on the non-detection of the J=18–17 K-ladder). Important molecular lines that fall in the frequency bands missing from the survey are the HCO$^{+}$ J=3–2 and HCN J=4–3 lines, plus several lines of SO$_{2}$ and CH$_{3}$OH. For the former two species we have detected their less common $^{13}$C isotopomers and are thus certain that the more common isotopomers are present in the molecular gas. The remaining two species are asymmetric rotors and there are many transitions lying in the frequency bands of the survey. Thus we are reasonably confident that the missing frequency bands do not affect the final conclusions of the survey. The molecular inventory of is that of a cold dense molecular core with depleted abundances of complex molecules and exhibiting emission only from low-excitation lines of chemically simple species. In many respects the molecular inventory of shares many characteristics with the line-poor ultracompact HII regions observed by Hatchell et al. ([@htmm98a]). Both types of object exhibit emission from low excitation lines with $E_{\rm u}/k \le 100$ K and show no evidence for hot, dense molecular gas. If IRAS 23385 can be thought of as a representative massive protostar this may indicate that the molecular inventory of massive star-forming regions evolves from line-poor protostars to line-rich hot cores as the molecular ices are evaporated from dust grains, then back to line-poor ultracompact HII regions as the chemically rich dense gas of the hot core is dispersed by the radiation from the newly-born massive stars. In summary, the molecular inventory of suggests that the molecular gas is cold and dense, with depleted abundances of the more complex saturated species that trace hot molecular cores. The composition of the gas appears to be similar to the line-poor ultracompact HII regions observed by Hatchell et al. ([@htmm98a]), whose limited molecular inventories are ascribed to the lack of a hot dense molecular core. We investigate this possibility in the following section. Could contain a hot core? {#sect:hotcore} ------------------------- does not exhibit any of the classic signs of a hot molecular core, i.e. high excitation lines or high abundances of saturated molecules such as CH$_{3}$CN or HCOOCH$_{3}$. The highest excitation lines detected in our survey have values of $E_{\rm u}/k \le 100$ K and in the previous section, using the observations of Pankonin et al. ([@pcwb01]), we set an upper limit for the beam-averaged CH$_{3}$CN column density of 5.010$^{12}$ cm$^{-2}$, which assumes an excitation temperature of 40 K. Nevertheless, given the small filling factor of the millimetre core (Molinari et al. [@mol98b]) and the beams of our and Pankonin et al’s study it is possible that a small, optically thick hot molecular core may lie below our sensitivity limits. To explore this possibility and set an upper limit on the size of any hot molecular core we modelled the expected CH$_{3}$CN emission from a compact molecular core using the LTE technique described by Hatchell et al. ([@htmm98a]). This technique assumes that the emission is in thermal equilibrium, but does not assume optically thin emission. The density of the millimetre core is $\sim$ 10$^{7}$ cm$^{-3}$ (Molinari et al. [@mol98b]) and this is more than sufficient to thermalise the CH$_{3}$CN transitions. The LTE analysis technique predicts the line receiver temperature [$T_{R}^{*}$]{}as a function of a column density $N$, the gas kinetic temperature $T_{\rm kin}$ and the angular source FWHM $\Theta_{\rm s}$. Here, as we can only assume an upper limit for the observed line receiver temperature, we do not have sufficient information to extract either the kinetic temperature or column density and so we model the emission based on assumed values of the kinetic temperature and the CH$_{3}$CN source-averaged column density. We assume that as the gas is thermalised $T_{\rm kin} \simeq T_{\rm ex}$. We use an upper limit for the line receiver temperature [$T_{R}^{*}$]{} of 3$\sigma$, which is 51 mK in the Pankonin et al. ([@pcwb01]) study for the J=12–11 K-ladder and 116 mK for the CH$_{3}$CN J=19–18 ladder that was undetected in our survey. Within both ladders we selected the K=0 component for our LTE calculations. To calculate the CH$_{3}$CN source-averaged column density we assume a standard hot core abundance for CH$_{3}$CN of 10$^{-8}$ relative to H$_{2}$ and use density and radius values for the millimetre core of $\sim 10^{7}$ cm$^{-3}$ and 0.048 pc as derived from the envelope model of Molinari et al. ([@mol98b]). The LTE models show that as the H$_{2}$ density of the millimetre core is high, the CH$_{3}$CN emission is extremely optically thick for both the J=12–11 and J=19–18 K-ladders and remains so even for high gas temperatures, small core radii and low CH$_{3}$CN abundances. In the optically thick case, the predicted receiver temperatures are practically independent of the column density or abundance of CH$_{3}$CN and depend solely on the kinetic temperature and beam filling factor of the core (see Equation 1 of Hatchell et al. [@htmm98b]). For a kinetic temperature of 150 K we determine a maximum hot core angular diameter of 07, using Pankonin et al’s detection limit and their FWHM beamwidth of 36. The angular size of an optically thick core is inversely proportional to the square root of the kinetic temperature, so there is a slight increase in the maximum core size for colder cores (08 for a 100 K core) and a corresponding decrease for hotter cores (05 for a 300 K core). Even though our survey is a factor of two less sensitive than the work of Pankonin et al. ([@pcwb01]), we obtain tighter constraints upon the maximum hot core angular diameter because our FWHM beamwidth (14) is less than half that of Pankonin et al. For a [$T_{R}^{*}$]{} 3$\sigma$ upper limit of 112 mK we determine maximum hot core angular diameters of 05, 04 and 03 for kinetic temperatures of 100, 150 and 300 K respectively. We have taken a kinetic temperature of 100 K as a lower limit to the temperature of a hot molecular core. The high abundances of saturated molecules that are present in hot cores are predominantly thought to arise from the evaporation of dust grain ice mantles (e.g. Brown et al. [@bcm88]), and these mantles would remain in the solid phase at temperatures much below 100 K. Thus, the maximum angular size of any hot molecular core that may be associated with is 05, which corresponds to a spatial radius of 0.006 pc at the assumed distance to of 4.9 kpc. This is roughly an eighth of the radius that Molinari et al. ([@mol98b]) derive for the millimetre core. The radii of hot molecular cores known to be associated with ultracompact HII regions has been measured as typically 0.03–0.06 pc via similar LTE modelling (Hatchell et al. [@htmm98a]) or interferometric CH$_{3}$CN observations Olmi et al. [@ocnw96]; Hofner et al. [@hkcwc96]). Using a simple argument based upon the Stefan-Boltzmann law we can estimate the minimum angular size of a hot core that is consistent with the bolometric luminosity of . The equilibrium temperature $T_{\rm eq}$ of the core is inversely proportional to the angular radius $\theta$ of the core (strictly, $T_{\rm eq}^{\,\,\,4} \propto \tan^{-2}\,\theta$) and so increasing the equilibrium temperature of the core decreases its angular radius. Assuming that the majority of the bolometric luminosity originates from the hot core and that the bolometric luminosity and distance of the core are 1.610$^{4}$ [$L_{\odot}$]{} and 4.9 kpc respectively (Molinari et al. [@mol98b]), we derive a minimum angular diameter for a 100 K equilibrium temperature core of 28. The validity of this approach may be checked by comparing the predicted equilibrium temperature of the IRAS 23385+6053 millimetre-wave core to that determined by Molinari et al. ([@mol98b]) from an envelope radiative transfer model. Using a core angular radius of 2 the predicted equilibrium temperature is 60 K, fully consistent with the temperature of 40 K determined by Molinari et al. (note that their temperature is that at the external radius of the core and increases towards the centre). The maximum possible hot core diameter for T $=$ 100 K that we have determined from the CH$_{3}$CN upper limit (05) is considerably smaller than the minimum possible diameter derived from the simple Stefan-Boltzmann luminosity argument (28). Only by increasing the temperature of the core to greater than 300 K can we reconcile the CH$_{3}$CN non-detection and the Stefan-Boltzmann predicted core size. However, such a high-temperature core is inconsistent with the non-detection of IRAS 23385+6053 at 15 $\mu$m (Molinari et al. [@mol98b]) for all but extremely low-mass cores. We thus conclude that IRAS 23385+6053 is not associated with a hot molecular core. The only supporting evidence for hot molecular gas towards that would strengthen the case for a hot core is the association with an H$_{2}$O maser (Molinari et al. [@mol96]). The collisional pumping of H$_{2}$O masers requires the presence of gas densities $\sim10^{7}$ cm$^{-3}$ and kinetic temperatures of several hundred K (Elitzur et al. [@ehm89]). However, the positional accuracy of the H$_{2}$O maser emission is only good to within $\sim 1$ (it has only been observed to date as part of the Medicina H$_{2}$O maser survey, Valdettaro et al.[@h2omedicina]). Thus the maser is as likely to originate from shocks within the outflow associated with than within a possible hot molecular core and by itself does not provide conclusive evidence for the presence of a hot molecular core. IRAS 23385+6053 is not associated with either a hot molecular core or an ultracompact HII region. There is thus a strong possibility that it is in an evolutionary phase prior to these two phenomena. We will explore the nature of IRAS 23385+6053 in the next section, paying particular attention to the constraints that chemistry may place upon its evolutionary phase. The physical and chemical nature of ------------------------------------ Our molecular line survey reveals that is comprised of cold gas with a molecular inventory limited to simple species such as CO, CS, SO, HCN and CH$_{3}$OH. We do not detect emission from highly excited molecular lines (all lines detected in the survey have values of $E_{\rm u}/k \le 100$ K), nor lines from saturated molecules believed to arise in the gas-phase from evaporation of dust grain molecular ice mantles. As a result it is unlikely that is associated with a hot molecular core, although our data do not preclude the existence of a small hot core with a radius less than 0.006 pc. A natural supposition given the non-detection of evaporated molecular species, the gas temperature of $\sim 20$K determined from the methanol rotation diagrams and the candidate protostellar nature of is that the molecular gas is in a cold, depleted “pre-switch on” state. In this scenario the massive protostar at the heart of the millimetre-wave core has either not yet begun to warm the surrounding gas or has not yet warmed a sufficient volume of gas to be detectable. Thus many molecular species could be currently frozen out in dust grain ice mantles. The temperature of the gas as probed by methanol in this study, CH$_{3}$CCH in the study of Brand et al. ([@bcpm01]) and the NH$_{3}$ study of Molinari et al. ([@mol96]) support this hypothesis, with gas temperatures ranging from 20–40 K. The small linewidths seen in our survey show that the gas is relatively quiescent compared to ultracompact HII regions and hot molecular cores which typically exhibit much broader linewidths of around 8–10 kms$^{-1}$(Hatchell et al. [@htmm98a]). is revealed as a cold, dense, relatively quiescent molecular core, displaying protostellar hallmarks such as a strong outflow and high sub-millimetre to bolometric luminosity ratio (Molinari et al. [@mol98b]). Does the chemistry of also support this hypothesis? Recent chemical models (Rodgers & Charnley [@rc03]; Viti & Williams [@vw99]) investigate the time-dependent evaporation of molecular ices within massive star-forming cores and their subsequent chemical evolution. The models predict that certain molecular species will be evaporated from the grain mantles at different epochs, depending upon their binding energy to the grain surface and the rise in luminosity of the central protostar or YSO. For example, in the collapse model of Rodgers & Charnley ([@rc03]) the abundance of SO$_{2}$ is predicted to peak later than SO. Before contrasting the column density ratios of various species to those predicted by the models it is useful to stress that due to the paucity of molecular lines detected in the survey we are able to provide stringent column density limits for only one species (methanol). The column density upper and lower limits that we have derived are all beam-averaged across a 14 FWHM beam and it is not at all certain that the molecular species trace the same emitting area of gas. The high critical density lines of CH$_{3}$CN with its large dipole moment are unlikely to trace the same gas as, for example, the low critical density lines of CH$_{3}$OH with its relatively low dipole moment. With these caveats in mind, it is nevertheless instructive to compare the observed chemistry of to that predicted by molecular evaporation chemical models in order to determine whether the models can constrain the evolutionary state of . We have chosen to calculate the column density ratios as a fraction of the methanol column density so that the resulting ratio is either an upper or lower limit depending upon whether the other species is a detection or a non-detection. Column density ratios were calculated for those species modelled by Rodgers & Charnley ([@rc03]) and Viti & Williams ([@vw99]) and are found in Table \[tbl:Nratio\]. In order to calculate the column densities of more common isotopomers from their less common variants (e.g. HCO$^{+}$ from H$^{13}$CO$^{+}$) we assumed that the more common isotopomer was optically thin and used standard interstellar isotope ratios. For the undetected species with upper limits to column density we used values assuming an excitation temperature of 20 K to maintain consistency with the rotation temperature drived from the methanol rotation diagrams. This may be inappropriate for the hot core tracer molecules HCOOCH$_{3}$ and CH$_{3}$CN, but as their column density upper limits slowly decrease with temperature until $T_{\rm ex} \simeq E_{\rm u/k}$ the minimum column density ratios quoted in Table \[tbl:Nratio\] are still valid for excitation temperatures less than this value. Species $N/N_{CH_{3}OH}$ -------------- ------------------ SO $> 0.03$ CS $> 1.5$ HCO$^{+}$ $> 0.03$ H$_{2}$CS $> 0.02$ H$_{2}$CO $>0.04$ CH$_{3}$CN $ < 0.01$ HCOOCH$_{3}$ $< 0.8$ SO$_{2}$ $< 0.04$ OCS $<1500$ : Column density ratios between species modelled in the evaporation models of Rodgers & Charnley ([@rc03]) and Viti & Williams ([@vw99]).Upper and lower limits to the ratios are indicated.[]{data-label="tbl:Nratio"} The chemical models of Rodgers & Charnley ([@rc03]) and Viti & Williams ([@vw99]) were not explicitly calculated for the molecular core and so we cannot constrain the chemical timescale in an absolute sense. However, Viti & Williams ([@vw99]) showed that whilst the evaporation timescale depends upon the rapidity of the molecular ice evaporation, the subsequent gas-phase evolution is similar for all evaporation models. We can thus use the column density ratios in Table \[tbl:Nratio\] to determine if the chemistry of IRAS 23385+6053 is consistent with the early, middle or late phases predicted by the chemical models. These three phases represent the early stages of the evaporation in which the chemistry is dominated by small simple molecules with low surface binding energies (e.g. CO and CH$_{4}$), the middle phase when most species have just evaporated from the grains and the late phase where the chemistry is dominated by the gas-phase reactions between evaporated species. Broadly speaking the chemistry of is consistent with the middle evaporation phases predicted by both models. In the model of Viti & Williams ([@vw99]) methanol comes off the grains roughly in the middle of the evaporation process, preceded by simple molecules with low surface binding energies and followed by the more complex species such as HCOOCH$_{3}$, CH$_{3}$CO and C$_{2}$H$_{5}$OH. The abundance of methanol after the initial evaporation rises rapidly, with a plateau at an abundance of $\sim 10^{-11}$ which lasts for a few 10$^{3}$ years followed by a second rapid rise to a final plateau at an abundance of $\sim 10^{-8}$. The observed column density ratios in Table \[tbl:Nratio\] show that methanol is more abundant than almost every other species with the exception of OCS and CS. The ratios for SO, CS, HCO$^{+}$, H$_{2}$CS and H$_{2}$CO are all lower limits and it is possible that these species are more abundant than methanol. Based upon the assumed excitation temperatures used in the calculation of the column density lower limit and the estimated kinetic temperature of the gas from the methanol rotation diagrams we expect the true column density to be underestimated by at most a factor of 10 (e.g. Thompson et al. [@tmm99]). For the ratios that are upper limits (CH$_{3}$CN, HCOOCH$_{3}$, SO$_{2}$ and OCS) we can be reasonably certain that the only species that may be more abundant than methanol is OCS. Methanol is predicted in the model of Viti & Williams ([@vw99]) to be more abundant than almost every other molecule in Table \[tbl:Nratio\] during the early evaporation phase. The exceptions to this rule are OCS which is more abundant than methanol at all times, CS which has roughly the same abundance during its early evaporation, H$_{2}$CO which is always more abundant than methanol and CH$_{3}$CN which evaporates from grains marginally before methanol and is more abundant until methanol evaporates. In the later phases many species are predicted to have larger abundances than methanol, in particular CS, H$_{2}$CS, SO, and the aforementioned OCS and H$_{2}$CO. The models of Rodgers & Charnley ([@rc03]) broadly agree with these predictions for the late phase; they do not calculate abundance ratios for the early ice evaporation phase (note that their Figure 11 begins at a time 100 years *after* the evaporation phase). On balance the chemistry of resembles that predicted by Viti & Williams ([@vw99]) for the middle ice evaporation phase immediately prior to the development of a hot molecular core. The abundance of SO rises faster than that of methanol after the middle phase. HCO$^{+}$ is more abundant than methanol during the early phase of evaporation and less abundant in the middle and later phases. The SO$_{2}$ abundance is lower than that of methanol in the early and middle stages, rising toward that of methanol in the late stages. These results constrain the chemical timescale of IRAS 23385+6053 to that of the middle phase, i.e. where the majority of species are beginning to be evaporated from the dust grain ice mantles. The major inconsistencies that arise between the observed ratios and the model predictions are for CS and OCS, where the observed ratios are larger than the model predictions. It is possible that the CS abundance is severely underestimated due to high optical depth. We can discount this possibility due to the non-detection of the C$^{34}$S line at 337.397 GHz, which allows us to set a maximum optical depth of 2.6 for the main isotopomer. The column density upper limit for OCS cannot be more accurately constrained as the only OCS lines lying in our survey range are of high excitation, possessing values of $E_{\rm u}/k \simeq 250$ K. Both the physical and chemical properties of point towards its protostellar nature and identify the millimetre core as being on the verge of developing into a hot molecular core. However, in order to confirm this hypothesis the chemical abundances of the molecular gas must be constrained by further observations of lower excitation lines and more detailed modelling. In particular as the molecular core is unresolved by our survey observations the relative emitting sizes and beam filling factors of the various species are extremely uncertain. Interferometric observations are a priority to try and resolve the emission and derive firm abundance limits for the chemistry of . Summary and conclusions ======================= We have carried out a molecular line survey with a total frequency range of 27.2 GHz of the candidate massive Class 0 protostar . We detected emission from 27 lines originating from 11 molecular species. One line feature could not be identified with any known lines in the JPL Molecular Spectroscopy Database or with other observational line lists (Jewell et al. ([@jhls89]); Macdonald et al. ([@mghm96]); Schilke et al. ([@sgbp97]); Thompson & Macdonald ([@tm99])). No emission was detected from high excitation lines ($E_{\rm u}/k \ge 100$ K) or from the complex saturated molecules (e.g. CH$_{3}$CN, HCOOCH$_{3}$ or C$_{2}$H$_{5}$OH) observed toward hot molecular cores. Over a third of the identified lines originate from methanol (CH$_{3}$OH). We derived the rotation temperature and column density for methanol using the rotation diagram approach and estimate lower beam-averaged column density limits for molecular species with one or two line detections (e.g. Thompson & Macdonald [@tm99]). Upper limits for undetected species were determined from the r.m.s. noise level of appropriate spectra. We draw the following conclusions from our survey: 1. The molecular inventory of resembles that of the line-poor ultracompact HII regions observed by Hatchell et al. ([@htmm98a]). They both exhibit emission from low excitation transitions of simple molecules and show no signs of hot core emission. Given the supposed massive protostellar nature of this may indicate that the molecular inventory of massive star-forming regions evolves from line-poor protostellar cores through a line-rich hot molecular core phase and back to a line-poor ultracompact HII region phase when the hot dense molecular core has been dispersed. 2. We see no evidence of a hot molecular core associated with . No emission was detected from high-excitation lines that are tracers of hot molecular gas and the rotation temperature derived from the methanol emission is $\sim 20$ K. We rule out the presence of a hot core by a combination of the LTE modelling of the CH$_{3}$CN emission that is consistent with our non-detection of the CH$_{3}$CN J=19–18 K-ladder and a simple argument based upon the Stefan-Boltzmann law and the bolometric luminosity of any possible hot molecular core. The H$_{2}$O maser associated with (Molinari et al. [@mol96]) must be associated with outflow shocks rather than the dense molecular core. 3. The chemical composition of the molecular core is consistent with the predictions of Viti & Williams ([@vw99]) for a molecular core in the middle evaporation phase, i.e. when the majority of species are beginning to be evaporated from dust grain ice mantles. This confirms the hypothesis of Molinari et al. ([@mol98b]) that is an extremely young massive protostellar object, possibly on the verge of developing into a hot molecular core. The authors would like to thank Neil Alvey and Samantha Large for their assistance with the data reduction and line identifications. We would also like to thank the referee, Todd Hunter, for providing several useful suggestions which considerably improved this paper, particularly the argument for the lack of a hot molecular core. MAT is supported by a PPARC postdoctoral grant. This research has made use of the SIMBAD astronomical database service operated at CDS, Strasbourg, France and the NASA Astrophysics Data System Bibliographic Services. Quicklook 2MASS images were obtained as part of the Two Micron All Sky Survey (2MASS), a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. MSX 8 $\mu$m data were obtained from the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Anderson, T., Herbst, E., Delucia, F.C., 1993, J. Mol. Spectrosc., 159, 410 André, P., Ward-Thompson, D., Barsony, M., 1993, ApJ, 406, 122 André, P., Ward-Thompson, D., Barsony, M., 2000, in Protostars & Planets IV, ed. V. Mannings, A. Boss & S. Russel (Tucson:Univ. Arizona Press), 59 Beuther, H., Schilke, P., Menten, K. M., Motte, F., Sridharan, T. K., & Wyrowski, F. 2002, [ApJ]{}, 566, 945 Brand, J., Cesaroni, R., Palla, F., & Molinari, S. 2001, [A&A]{}, 370, 230 Brown, P. D., Charnley, S. B., & Millar, T. J. 1988, [MNRAS]{}, 231, 409 Churchwell, E. 2002, [ARA&A]{}, 40, 27 de Pree, C. G., Rodriguez, L. F., & Goss, W. M. 1995, Revista Mexicana de Astronomia y Astrofisica, 31, 39 Elitzur, M., Hollenbach, D. J., & McKee, C. F. 1989, [ApJ]{}, 346, 983 Gibb, E., Nummelin, A., Irvine, W. M., Whittet, D. C. B., & Bergman, P. 2000, [ApJ]{}, 545, 309 Hatchell, J., Thompson, M. A., Millar, T. J., & Macdonald, G. H. 1998a, [A&AS]{}, 133, 29 Hatchell, J., Thompson, M. A., Millar, T. J., & Macdonald, G. H. 1998b, [A&A]{}, 338, 713 Hofner, P., Kurtz, S., Churchwell, E., Walmsley, C. M., & Cesaroni, R. 1996, [ApJ]{}, 460, 359 Jewell, P. R., Hollis, J. M., Lovas, F. J., & Snyder, L. E. 1989, [ApJS]{}, 70, 833 Kurtz, S., Cesaroni, R., Churchwell, E., Hofner, P., Walmsley, C.M., 2000, Protostars and Planets IV (Tuscon, Arizona Press), eds. v. Mannings, A.P. Boss, S.Russell, p299 Kutner, M.L., Ulich, B.L, 1981, ApJ, 250, 341 Lumsden, S. L., Hoare, M. G., Oudmaijer, R. D., & Richards, D. 2002, [MNRAS]{}, 336, 621 Macdonald, G. H., Gibb, A. G., Habing, R. J., & Millar, T. J. 1996, [A&AS]{}, 119, 333 Millar, T. J., Macdonald, G. H., & Gibb, A. G. 1997, [A&A]{}, 325, 1163 Molinari, S., Brand, J., Cesaroni, R., & Palla, F. 1996, [A&A]{}, 308, 573 Molinari, S., Brand, J., Cesaroni, R., Palla, F., & Palumbo, G. G. C. 1998a, [A&A]{}, 336, 339 Molinari, S., Testi, L., Brand, J., Cesaroni, R., & Pallo, F. 1998b, [ApJ]{}, 505, L39 Molinari, S., Brand, J., Cesaroni, R., & Palla, F. 2000, [A&A]{}, 355, 617 Olmi, L., Cesaroni, R., Neri, R., & Walmsley, C. M. 1996, [A&A]{}, 315, 565 Pankonin, V., Churchwell, E., Watson, C., & Bieging, J. H. 2001, [ApJ]{}, 558, 194 Prestage, R.M., Meyerdierks, H., Lightfoot, J.F., Jenness, T., Tilanus, R.P.J., Padman, R., Chipperfield, A.J., 2000, Starlink User Note 17, Starlink Project, CCLRC Rodgers, S. D. & Charnley, S. B. 2003, [ApJ]{}, 585, 355 Schilke, P., Groesbeck, T. D., Blake, G. A., & Phillips, T. G. 1997, [ApJS]{}, 108, 301 Sridharan, T. K., Beuther, H., Schilke, P., Menten, K. M., & Wyrowski, F. 2002, [ApJ]{}, 566, 931 Thompson, M. A. & Macdonald, G. H. 1999, [A&AS]{}, 135, 531 Thompson, M. A., Macdonald, G. H., & Millar, T. J. 1999, [A&A]{}, 342, 809 Turner, B. E. 1991, [ApJS]{}, 76, 617 Valdettaro, R. et al. 2001, [A&A]{}, 368, 845 Viti, S. & Williams, D. A. 1999, [MNRAS]{}, 305, 755 [^1]: The JCMT is operated by the Joint Astronomy Centre on behalf of PPARC for the United Kingdom, the Netherlands Organisation of Scientific Research, and the National Research Council of Canada.

ArXiv

--- abstract: 'Probing the real time dynamics of a reacting molecule remains one of the central challenges in chemistry. In this letter we show how the time-dependent wavefunction of an excited-state reacting molecule can be completely reconstructed from resonant coherent anti-Stokes Raman spectroscopy. The method assumes knowledge of the ground-state potential but not of any excited-state potential, although we show that the latter can be computed once the time-dependent excited-state wavefunction is known. The formulation applies to polyatomics as well as diatomics and to bound as well as dissociative excited potentials. We demonstrate the method on the Li$_2$ molecule with its bound first excited-state, and on a model Li$_2$-like system with a dissociative excited state potential.' author: - David Avisar - 'David J. Tannor' title: | Complete Reconstruction of the Wavefunction of a Reacting Molecule\ by Four-Wave Mixing Spectroscopy --- 0.25cm For several decades now, femtosecond pump-probe spectroscopies have been employed to study transition states of molecules reacting on excited potential surfaces [@zewail; @polanyi; @mathies; @Takeuchi; @ruhman]. Although these studies have shed a tremendous amount of light on excited-state dynamics, none of the methods in use provides complete information on the excited-state wavefunction. The need for an experimental method that will provide this information is compounded by the fact that theoretical ab initio calculations for excited states are difficult and of limited accuracy. There have been several theoretical proposals for complete reconstruction of an excited-state molecular wavefunction from spectroscopic signals [@shapiro_imaging; @cina3]. These studies, however, generally assume that one or more excited-state potentials (or the corresponding vibrational eigenstates) is known. A notable exception is a recently developed iterative method for excited-state potential reconstruction from electronic transition dipole matrix elements [@shapiro2] but this method does not appear to be applicable to dissociative potentials. Experimental work has focused on wavepacket interferometry of vibrational wavepackets [@scherer; @ohmori] as well as electronic Rydberg wavepackets [@weinacht; @girard]. The approach we present here assumes knowledge of the ground-state potential but not of any excited potential. In principle, the approach is completely general for polyatomics. Our strategy is to express the reacting-molecule wavefunction, $|\Psi(t)\rangle$, as a superposition of the vibrational eigenstates of the ground-state Hamiltonian, ${\{ |\psi_{g}\rangle \}}$: $$\begin{aligned} |\Psi(t)\rangle = \sum_{g} |\psi_{g}\rangle \langle \psi_{g} |\Psi(t)\rangle \equiv \sum_{g} C_{g}(t) |\psi_{g}\rangle . \label{Psi_superpos}\end{aligned}$$ Since the vibrational eigenstates $\{|\psi_{g}\rangle\}$ are assumed known, the challenge is to find the time-dependent superposition coefficients $C_{g}(t)$. Consider a two-state molecular system within the Born-Oppenheimer approximation. The nuclear Hamiltonians $H_g$ and $H_e$ correspond, respectively, to the (known) ground and (unknown) excited potentials, which can be of any dimension. For simplicity, we consider a $\delta$-pulse excitation as well as a coordinate-independent electronic transition dipole, $\mu$ (Condon approximation). Applying first-order time-dependent perturbation theory, the wavepacket that we want to reconstruct is [@tannor_book] $$\begin{aligned} | \Psi(t) \rangle= -i e^{-iH_{e}t} \left\lbrace -{\mu} \varepsilon_{1} \right\rbrace | \psi_{0} \rangle, \label{1st_order_wf_deltpuls}\end{aligned}$$ where the initial state, $ {| \psi_{0} \rangle}$, is the vibrational ground-state of $H_{g}$ with the eigenfrequency $\omega_{0}$, $\varepsilon_{1}$ is the amplitude of the pulse and $t$ is the propagation time on the excited state measured from the time of pulse excitation. (Here and henceforth we take $\hbar=1$.) Substituting Eq. (\[1st\_order\_wf\_deltpuls\]) into the definition of $C_{g}(t)$, we find that the superposition coefficients are given by $$\begin{aligned} C_{g}(t) = i{ \mu} \varepsilon_{1} \langle \psi_{g} | e^{-iH_{e}t} | \psi_{0} \rangle \equiv i{ \mu} \varepsilon_{1} c_{g}(t). \label{C_coeff_delt_condon}\end{aligned}$$ Hence, the central quantities required for reconstructing $|\Psi(t)\rangle$ are the cross-correlation functions $c_{g}(t)$. It has long been recognized that these correlation functions appear (up to ${\mu}$) in the time-dependent formulation of resonance Raman scattering (RRS) [@heller_tdraman]; however, the experimental RRS signal involves the absolute-value-squared of the half-Fourier transform of the correlation function, hence the latter cannot be recovered from that signal. Fully resonant coherent anti-Stokes Raman scattering (CARS) has been shown to be a powerful probe of ground and excited electronic states properties [@decola; @mathew]. In this letter we show that the correlation functions $\{c_{g}(t)\}$ may be completely recovered from femtosecond resonant CARS spectroscopy, allowing complete reconstruction of the excited-state wavepacket. The formula for the CARS signal produced by a three-pulse pump-dump-pump sequence is ${P^{(3)}(\tau) = \langle \psi^{(0)}(\tau) | \hat{\mu} | \psi^{(3)}(\tau)\rangle + {\rm c.c.}}$ [@faeder], where $\psi^{(3)}(\tau)$ is the third-order wavefunction and ${\psi^{(0)}(\tau)= e^{-iH_{g}\tau }\psi_{0}}$. Within the $\delta$-pulse and Condon approximations, $P^{(3)}$ takes the form $$\begin{aligned} P^{(3)}(\bm{\tau}) = \widetilde{\varepsilon} \langle \psi_{0} | e^{-iH_{e} \tau_{43}} e^{-i \widetilde{H}_{g}\tau_{32}} e^{-iH_{e}\tau_{21}} |\psi_{0}\rangle, \label{3rd_polariz_tdelays_condon}\end{aligned}$$ where ${\tau_{ij}=\tau_{i}-\tau_{j}}$ is the (positive) time-delay between the centers of the $i$th and $j$th pulses and ${\tau_{43}=\tau-\tau_{3}}$ with $\tau$ being the time of signal measurement. We have denoted ${\widetilde{H}_{g}=H_{g}-\omega_{0}}$, ${\widetilde{\varepsilon} = i^{3} {\mu}^{4}\varepsilon_{1} \varepsilon_{2} \varepsilon_{3} e^{i\omega_{0}(\tau_{21}+\tau_{43})}}$ with $\varepsilon_{1,2,3}$ as the first, second and third pulse amplitudes, respectively, and $\bm{\tau}\equiv[\tau_{21}, \tau_{32}, \tau_{43}]$. In writing $P^{(3)}(\bm{\tau})$ as a complex quantity we have assumed the signal is measured in a heterodyne fashion. As illustrated in Fig. \[CARS\_scheme2\], Eq. (\[3rd\_polariz\_tdelays\_condon\]) has the following physical interpretation: A first laser pulse, the pump pulse, transfers amplitude to the excited potential surface creating a wavepacket whose time-dependence we are interested in reconstructing. After evolving on the excited state for some time, a second laser pulse, the dump pulse, transfers part of this amplitude back to the ground state where it evolves for a second interval of time. Finally, a third laser pulse excites part of the second-order amplitude to the excited state, generating the third-order polarization that produces the CARS signal, measured at later times. The desired wavefunction $|\Psi(t)\rangle$ (Eq. (\[1st\_order\_wf\_deltpuls\])) may already be recognized in Eq. (\[3rd\_polariz\_tdelays\_condon\]). ![ []{data-label="CARS_scheme2"}](CARS_scheme2B.ps){width="4cm"} The reconstruction of $|\Psi(t)\rangle$ from $P^{(3)}$ proceeds in five steps: 1\. *Insert a complete set of ground vibrational states*. Introducing ${ \sum_{g}| \psi_{g} \rangle \langle \psi_{g}| }=\hat{\textbf{1}}$ into Eq. (\[3rd\_polariz\_tdelays\_condon\]) we obtain $$\begin{aligned} P^{(3)}(\bm{\tau}) = \widetilde{\varepsilon} \sum_{g=0}^{N} e^{-i\widetilde{\omega}_{g} \tau_{32}} P_{g}^{(3)}(\tau_{43},\tau_{21}), \label{3rd_polariz_closr}\end{aligned}$$ where ${P_{g}^{(3)} (\tau_{43},\tau_{21}) = \langle \psi_{0} | e^{-iH_{e} \tau_{43}} |\psi_{g} \rangle \langle \psi_{g}| e^{-iH_{e}\tau_{21}} |\psi_{0} \rangle}$, and ${\widetilde{\omega}_{g} = \omega_{g} - \omega_{0}}$. $N$ is determined by the number of ground vibrational states required to expand $|\Psi(t)\rangle$. Note that the desired correlation functions $c_{g}(t)$ may already be recognized in $P_{g}^{(3)} $. 2\. *Fourier-transform $P^{(3)}$ with respect to $\tau_{32}$*. The transformation resolves $P^{(3)}$ into individual ground-state components, $P_{g}^{(3)}$. Since $\tau_{32}$ is defined to be positive, we multiply Eq. (\[3rd\_polariz\_closr\]), prior to the transformation, by the rectangular function that takes the value 1 for the $\tau_{32}$ domain and 0 elsewhere. Using the Fourier convolution theorem we obtain a [sinc]{}-type of spectrum with peaks at the frequencies $\omega=\widetilde{\omega}_{g}$: $$\begin{aligned} \widetilde{P}^{(3)}(\tau_{43},\omega,\tau_{21}) = \sum_{g=0}^{N} S(\omega,g) P_{g}^{(3)} (\tau_{43},\tau_{21}), \label{3rd_polariz_ft}\end{aligned}$$ where ${S(\omega,g) = 2 T\widetilde{\varepsilon} e^{i(\omega - \widetilde{\omega}_{g})(\check{\tau}_{32}+T)} {\rm sinc} [(\omega -\widetilde{\omega}_{g})T]}$, ${2T=\hat{\tau}_{32}-\check{\tau}_{32}}$, and $\check{\tau}_{32}$ ($\hat{\tau}_{32}$) is the minimal (maximal) value of $\tau_{32}$. Fixing $(\tau_{43},\tau_{21})$, Eq. (\[3rd\_polariz\_ft\]) can be written as a matrix equation: $$\begin{aligned} \mathbf{\widetilde{P}_{\bm{\omega}}^{(3)}} = \mathbf{S_{\bm{\omega} g}} \mathbf{P_{g}^{(3)}}. \label{matrx_eq}\end{aligned}$$ 3\. *Invert the matrix equation (\[matrx\_eq\])*. The equation $\mathbf{P_{g}^{(3)}} = \mathbf{S_{g \bm{\omega}}^{-1}} \mathbf{\widetilde{P}_{\bm{\omega}}^{(3)}}$ isolates the two-dimensional functions $P_{g}^{(3)} (\tau_{43},\tau_{21})$. In inverting $\mathbf{S}$ we choose the number of frequency elements ($\omega$) equal to the number of the $g$ elements so that the matrix is square. For numerical accuracy, the inversion is implemented separately around each of the peaks at $\widetilde{\omega}_{g}$. 4\. *Take the square-root of $P_{g}^{(3)}$*. Assuming the functions $\{\psi_{g}(x)\} $ are real, we can rewrite $P_{g}^{(3)}$ as $$\begin{aligned} P_{g}^{(3)} (\tau_{43},\tau_{21}) = \langle \psi_{g} | e^{-iH_{e}\tau_{43}} | \psi_{0} \rangle \langle \psi_{g} | e^{-iH_{e}\tau_{21}} | \psi_{0} \rangle. \label{two_d_polariz_2ndequ}\end{aligned}$$ Taking the square-root of the diagonal of $P_{g}^{(3)} (\tau_{43},\tau_{21})$ (i.e. ${\tau_{43}=\tau_{21}= t}$), we recover the $c_{g}(t)$ up to a sign: $$\begin{aligned} \sqrt {P_{g}^{(3)} (t)}=a_{g} \langle \psi_{g} | e^{-iH_{e}t} | \psi_{0} \rangle \equiv \langle \widetilde{\psi}_{g} | e^{-iH_{e}t} |\psi_{0} \rangle, \label{cross_cor_sqrt}\end{aligned}$$ where $a_{g}=\pm1$ and the sign of $\widetilde{\psi}_{g}(x)$ is as yet undetermined. By demanding continuity of the cross-correlation functions (and their derivatives), the coefficients $a_{g}$ can be regarded as time-independent. Substituting Eq. (\[cross\_cor\_sqrt\]) instead of $c_{g}(t)$ into Eq. (\[C\_coeff\_delt\_condon\]) and using the resulting $C_{g}(t)$ in Eq. (\[Psi\_superpos\]) yields $$\begin{aligned} |\widetilde{\Psi}(t)\rangle = i \mu \varepsilon_{1} \sum_{g=0}^{N} |\psi_{g}\rangle \langle \widetilde{\psi}_{g} | e^{-iH_{e}t} | \psi_{0} \rangle. \label{psi_1t}\end{aligned}$$ The different sign combinations of $\widetilde{\psi}_{g}(x)$ generate $2^{N+1}$ possible superpositions. (In fact, only $2^{N}$ are physically meaningful since we are free to set the sign of one of the $g$-components.) Only one out of the $2^{N}$ $|\widetilde{\Psi}(t)\rangle$ coincides with $|{\Psi}(t)\rangle$: the $|\widetilde{\Psi}(t)\rangle$ for which the sign combination satisfies ${\sum_{g} |\psi_{g}\rangle \langle \widetilde{\psi}_{g} | = \mathbb{1}}$. 5\. *Discriminating ${|\Psi}(t)\rangle$ from the set $\{| \widetilde{\Psi}(t)\rangle \}$*. The set of wavefunctions $\{|\widetilde{\Psi}(t)\rangle\}$ are all consistent with the CARS signal at a specific value of ${\tau_{43}=\tau_{21} }$ [@footnt]. However, only one $|\widetilde{\Psi}(t)\rangle$ is consistent with the signal *derivatives*. To see this, consider the *n*th derivative of the experimental signal, Eq. (\[3rd\_polariz\_tdelays\_condon\]), with respect to $\tau_{21}$: $$\begin{aligned} &~&\frac{\partial^{n} P^{(3)}(\bm{\tau})}{\partial \tau^{n}_{21}}=\varepsilon^{\dag} \langle \Psi^{*}(\tau_{43})|e^{-iH_{g}\tau_{32}} \widetilde{H}_{e}^{n} |\Psi(\tau_{21}) \rangle~~~~ \nonumber \\ \nonumber \\ &~&=\varepsilon^{\dag} \sum_{g,g'} e^{-i\omega_{g}\tau_{32}} C_{g}(\tau_{43})C_{g'}(\tau_{21}) \widetilde{H}_{e,gg'}^{n}, \label{Polder_dep_on_Psi}\end{aligned}$$ where $\varepsilon^{\dag}=(-i)^{n-1}\mu^{2} \varepsilon_{1}^{-1} \varepsilon_{2} \varepsilon_{3} e^{i\omega_{0}\tau_{41}}$, $\tau_{41}=\tau-\tau_{1}$, ${\widetilde{H}_{e}^{n}=(H_{e}-\omega_{0}})^{n}$, and ${\widetilde{H}_{e,gg'}^{n}=\langle \psi_{g}|\widetilde{H}_{e}^{ n}| \psi_{g'} \rangle}$. Substituting $|\widetilde{\Psi}(t)\rangle$ instead of $|\Psi(t)\rangle$, into Eq. (\[Polder\_dep\_on\_Psi\]) gives $$\begin{aligned} \frac{ \partial^{n} \widetilde{P^{(3)}}(\bm{\tau} )}{\partial \tau_{21}^{n}} &=&\varepsilon^{\dag} \sum_{g,g'} e^{-i\omega_{g}\tau_{32}} a_{g}a_{g'}C_{g}(\tau_{43})C_{g'}(\tau_{21})\widetilde{H}_{e,gg'}^{n}. \nonumber \\ \vspace{-0.4cm} \label{plug_psi_toPder}\end{aligned}$$ Accordingly, the $|\widetilde{\Psi}(t)\rangle$ for which ${\frac { \partial^{n} \widetilde{P^{(3)}}(\bm{\tau} )}{\partial \tau_{21}^{n}}= \frac{\partial^{n} P^{(3)}(\bm{\tau})}{\partial \tau_{21}^{n}}}$ for all $n$, is the wavefunction that coincides with $|\Psi(t)\rangle$ of Eq. (\[1st\_order\_wf\_deltpuls\]), and hence, is the reconstruction solution. In practice, we proceed as follows. We invert the time-dependent Schrödinger equation to calculate a set of potentials from each $|\widetilde{\Psi}(t)\rangle$: $$\begin{aligned} V(x) &=& \frac{1}{\widetilde{\Psi}(x,t)} \left[ i\frac{\partial }{\partial t} + \frac{1}{2m} \frac{\partial^{2}}{\partial x^{2}} \right]\widetilde{\Psi}(x,t),~~~ \label{td_pot_recons}\end{aligned}$$ where $m$ is the system’s reduced mass. One can show that the potentials calculated by the $|\widetilde{\Psi}(t)\rangle$ that do not coincide with $|\Psi(t)\rangle$, are time-dependent [@avisar]. Only the potential calculated with $|\widetilde{\Psi}(t)\rangle= |\Psi(t)\rangle$ is time-independent and hence corresponds to the excited-state Hamiltonian $H_{e}$ of the measured system. Thus, in order to find the correct wavefunction we use the set of calculated potentials, as if they were time-independent, to propagate the corresponding $\{|\widetilde{\Psi}(t)\rangle\}$ back to time zero. Of all the potentials, only the truly time-independent one will propagate the corresponding $|\widetilde{\Psi}(t)\rangle$ correctly back to $|\psi_{0}\rangle$, and therefore this $|\widetilde{\Psi}(t)\rangle$ is the correct wavefunction. Note that the above procedure requires knowing the signal as a function only of $\tau_{32}$ and $\tau_{21}=\tau_{43}$. [c @ | @ c c c]{}   &$X$ & $A$ & $\widetilde{A}$\ \ $D$ & 0.0378492 & 0.0426108& 9.11267$\times10^{-5}$\ $b$ & 0.4730844 & 0.3175063& 1.5875317\ $x_{0}$ & 5.0493478 & 5.8713786& 7.3699313\ $T$ & 0 & 0.0640074 & 0.0640074\ To test the above reconstruction methodology, we simulated the CARS signal by calculating $ {\langle \psi^{(0)}(\tau) | \hat{\mu}| \psi^{(3)}(\tau)\rangle}$ as a function of the three time-delays, for two one-dimensional systems. The first is the Li$_{2}$ molecule, with its ground ($X$) and first-excited ($A$) electronic states as Morse-type potentials, ${V(x) = D(1-e^{-b(x-x_{0})})^{2} + T}$. The second system, henceforth denoted d-Li$_{2}$, has the Li$_{2}$ ground state ($X$) but a dissociative excited potential of the form ${V(x) = De^{-b(x-x_{0})} + T}$ (denoted $\widetilde{A}$). Table \[table\_pot\] gives the potential parameters in atomic units used for the simulations. The parameters for the Morse-type potentials are based on data published in [@herzberg]. The wavepacket propagations employed in simulating $P^{(3)}$ were performed using the split-operator method [@feit] on a spatial grid of 256 points in the range of $2$–$12$a.u. with time spacing of $\Delta t$=0.1fs. A constant transition-dipole of $2$a.u. was used, and the pulse amplitudes $\varepsilon_{1,2,3}$ were taken to be $10^{-4}$a.u. The ranges of time-delay for the Li$_2$ (d-Li$_{2}$) system were $\tau_{21,43}=0 - 200$fs ($0 - 80$fs) with spacing of $0.2$fs. For both systems, we took $\tau_{32}=3 - 6000$fs with $1$fs spacing. For Li$_2$, we inverted Eq. (\[matrx\_eq\]) for each of the first 25 peaks of $\widetilde{P}^{3}(\omega)$ using the matrices $\mathbf{S}$ with 25 frequency grid points centered around the peaks at $\widetilde{\omega}_{g}$. This produced 25 two-dimensional functions $P_{g}^{(3)}$, $g=0,\ldots,24$. For d-Li$_2$ the procedure was performed for the first 40 peaks, producing 40 two-dimensional functions $P_{g}^{(3)}$, $g=0,\ldots,39$. ![ []{data-label="reconwf_sqrt_XA_Li2_0_200_all"}](recon_wf_sqrt_bndX_bndA_Li2_0_200_all.ps){width="8.5cm"} In Figs. \[reconwf\_sqrt\_XA\_Li2\_0\_200\_all\] and \[reconwf\_sqrt\_X\_Disso\_Li2\_0\_80\_part\] we present snapshots of the real part of the reconstructed first-order wavefunction for the Li$_2$ and the d-Li$_{2}$ molecules, respectively. For Li$_2$ (d-Li$_{2}$) we superpose the first 25 (40) eigenfunctions $\psi_{g}(x)$ using the cross-correlation functions obtained by the CARS analysis and maintaining $\sum_{g}|\psi_{g}\rangle \langle \widetilde{\psi}_{g}|=\mathbb{1}$. ![ []{data-label="reconwf_sqrt_X_Disso_Li2_0_80_part"}](recon_wf_sqrt_bndX_Disso_Li2_0_80_part.ps){width="8.5cm"} The reconstructed wavefunctions are seen to be in excellent agreement with the exact ones, obtained by direct calculation of the first-order population, for all propagation times. For the Li$_2$ system, a high quality reconstruction is already obtained by superposing just 20 basis functions. ![ []{data-label="recon_potential_sqrt_bndX_Disso_Li2_0_80_02dt_all"}](recon_pot_bound_Li2_0_200_all_new.ps){width="8.7cm"} ![ []{data-label="recon_potential_sqrt_bndX_Disso_Li2_0_80_02dt_all"}](recon_potential_sqrt_bndX_Disso_Li2_0_80_02dt_all.ps){width="8.7cm"} Having determined the wavefunctions we calculate the corresponding excited potential surfaces from Eq. (\[td\_pot\_recons\]) using eight-point (three-point) central finite-differencing for the time (spatial) derivatives. The time-step used was 0.2fs but very good results were also obtained using 0.5fs. Figures \[reconst\_pot\_bound\_Li2\_02dt\_all\] and \[recon\_potential\_sqrt\_bndX\_Disso\_Li2\_0\_80\_02dt\_all\] compare the reconstructed vs. the exact potentials. The wavefunction (absolute value) used in calculating the potential is shown by a black solid line. Note from Figs. \[reconst\_pot\_bound\_Li2\_02dt\_all\] and \[recon\_potential\_sqrt\_bndX\_Disso\_Li2\_0\_80\_02dt\_all\] that combining the reconstructed potential from two points in time (e.g. $5$ and $70$fs for Li$_2$ and $5$ and $79$fs for d-Li$_2$) is sufficient to reconstruct the potential over the full range of interest (2–5Å). Once the potential is known one can calculate the excited-state wavefunction as a function of time for any excitation pulse sequence without the need for any additional laboratory experiments. To conclude, we have presented a methodology for the complete reconstruction of the excited-state wavefunction of a reacting molecule by analyzing a multi-dimensional resonant CARS signal. The methodology applies to polyatomics as well as diatomics. We have assumed that only the ground-state potential is known. The approach is very compelling since the desired excited-state wavefunction is explicitly contained in the formula for the CARS signal. Highly accurate reconstruction is obtained even far from the Franck-Condon region. In fact, in practice the method may be more accurate far from the Franck-Condon region, since the frequency shift between the pump and dump pulses will be more effective in discriminating unwanted processes that may contribute to the measured signal at $\mathbf{k=k_{1}-k_{2}+k_{3}}$. We simplified matters by considering $\delta$-function pulse excitations, a coordinate-independent transition dipole moment and only one excited-state potential. In future work we will test the removal of all these assumptions. We have shown that once the time-dependent wavefunction is found, the excited potential can be reconstructed with quite high accuracy. It will be of great interest to test the method on polyatomics, where obtaining multidimensional potential surfaces from spectroscopic data has been one of the longstanding challenges of molecular spectroscopy. An important application of excited-state potential reconstruction will be the ab initio simulations of laser control of chemical bond breaking. Experimental laser control has been greatly hindered by the lack of detailed theoretical guidance, which in turn is due to the lack of accurate excited-state potentials. The present methodology could have a significant impact in this field by providing the necessary information about excited-state potentials. This research was supported by the Minerva Foundation and made possible, in part, by the historic generosity of the Harold Perlman family. [99]{} A.H. Zewail, Science [**242**]{}, 1645 (1988). J.C. Polanyi and A.H. Zewail, Acc. Chem. Res. [**28**]{}, 119 (1995). P. Kukura, D.W. McCamant, S. Yoon, D.B. Wandschneider and R.A. Mathies, Science [**310**]{}, 1006 (2005). S. Takeuchi, S. Ruhman, T. Tsuneda, M. Chiba, T. Taketsugu and T. Tahara, Science [**322**]{}, 1073 (2008). U. Banin and S. Ruhman, J. Chem. Phys. [**99**]{}, 9318 (1993). M. Shapiro, J. Chem. Phys. [**103**]{}, 1748 (1995); C. Leichtle, W.P. Schleich, I.Sh. Averbukh and M. Shapiro, Phys. Rev. Lett. [**80**]{}, 1418 (1998). T.S. Humble and J.A. Cina, Phys. Rev. Lett. [**93**]{}, 060402-1 (2004); J.A. Cina, Annu. Rev. Phys. Chem. [**59**]{}, 319 (2008). X. Li, C. Menzel-Jones, D. Avisar and M. Shapiro, Phys. Chem. Chem. Phys. [**12**]{}, 15760 (2010). N.F. Scherer et al., J. Chem. Phys. [**95**]{}, 1487 (1991). K. Ohmori et al., Phys. Rev. Lett. [**96**]{}, 093002 (2006); K. Ohmori, Annu. Rev. Phys. Chem. [**60**]{}, 487 (2009). T.C. Weinacht, J. Ahn and P.H. Bucksbaum, Phys. Rev. Lett. [**80**]{}, 5508 (1998). A. Monmayrant, B. Chatel and B. Girard, Phys. Rev. Lett. [**96**]{}, 103002 (2006). D.J. Tannor, *Introduction to Quantum Mechanics: A Time-Dependent Perspective* (University Science Books Sausalito, 2007), Eq. (13.8). Soo-Y. Lee and E.J. Heller, J. Chem. Phys. [**71**]{}, 4777 (1979); E.J. Heller, R.L. Sundberg and D. Tannor, J. Phys. Chem. [**86**]{}, 1822 (1982); A.B. Myers, R.A. Mathies, D.J. Tannor and E.J. Heller, J. Chem. Phys. [**77**]{}, 3857 (1982); D. Imre, J.L. Kinsey, A. Sinha and J. Krenos, J. Phys. Chem. [**88**]{}, 3956 (1983). P.L. Decola, J.R. Andrews and R.M. Hochstrasser, J. Chem. Phys. [**73**]{}, 4695 (1980). N.A. Mathew et al., J. Phys. Chem. A [**114**]{}, 817 (2010). J. Faeder, I. Pinkas, G. Knopp, Y. Prior and D.J. Tannor, J. Chem. Phys. [**115**]{}, 8440 (2001). In fact, the set of wavefunctions given by (\[psi\_1t\]) are consistent with the CARS signal for any pair $(\tau_{21},\tau_{43})$. See Supplementary Online Material. G. Herzberg, in [*Molecular Spectra and Molecular Structure; I. Spectra of Diatomic Molecules,*]{} Krieger Publishing Company, Malabar, Florida, (1950). J.M.D. Feit, J.A. Fleck and A. Steinger, J. Comput. Phys. [**47**]{}, 412 (1982).   **Supplementary Online Material – Determining the Correct Wavefunction out of the Set $\{\widetilde{\Psi}_{i}(t)\}$** In this supplement we explain how we determine the correct wavefunction out of the set of wavefunctions $\{\widetilde{\Psi}_{i}(t)\}$, $i = 1, 2, ..., 2^{N}$, where $N$ is the number of basis functions $\{\psi_{g}\}$ needed to span $\Psi(t)$ (ref \[22\] in the paper). We have defined a set of wavefunctions that can be constructed using the information obtained from the CARS signal: $$\begin{aligned} |\widetilde{\Psi}(t)\rangle &\equiv& \sum_{g} |\psi_{g}\rangle \langle \widetilde{\psi}_{g} | e^{-iH_{e}t} | \psi_{0} \rangle \nonumber \\ &=&\sum_{g} |\psi_{g}\rangle \langle \widetilde{\psi}_{g} | \Psi(t) \rangle \equiv \widetilde{\mathbb{1}}|\Psi(t) \rangle. \label{psi_1t}\end{aligned}$$ (In writing Eq. (\[psi\_1t\]) we omitted the proportionality coefficient $i \mu \varepsilon_{1}$ relative to Eq. (10) in the paper.) Recall that $\widetilde{\mathbb{1}} \equiv \sum_{g} |\psi_{g}\rangle \langle \widetilde{\psi}_{g} | \equiv \sum_{g} |\psi_{g}\rangle a_{g} \langle \psi_{g} |$ where $a_{g}$ may take one out of two possible values: $\pm 1$. A useful property of the operator $\widetilde{\mathbb{1}}$ is that its square equals the identity operator $\mathbb{1}$: $$\begin{aligned} \widetilde{\mathbb{1}}\widetilde{\mathbb{1}} &=& \sum_{gg'} a_{g}a_{g'}|\psi_{g}\rangle \langle \psi_{g} |\psi_{g'}\rangle \langle \psi_{g'} | \nonumber \\ &=& \sum_{g} a^{2}_{g}|\psi_{g}\rangle \langle \psi_{g} | = \sum_{g} |\psi_{g}\rangle \langle \psi_{g} | = \mathbb{1}.\end{aligned}$$ We can derive an equation of motion for $\widetilde{\Psi}(t)$: $$\begin{aligned} \frac{\partial}{\partial t}|\widetilde{\Psi}(t)\rangle &= & \frac{\partial}{\partial t} \widetilde{\mathbb{1}} | \Psi(t) \rangle = \widetilde{\mathbb{1}} \frac{\partial}{\partial t} | \Psi(t) \rangle = -i\widetilde{\mathbb{1}}H_{e}|\Psi(t)\rangle \nonumber \\ &=&-i\widetilde{\mathbb{1}}H_{e}\widetilde{\mathbb{1}}\widetilde{ \mathbb {1}}|\Psi(t)\rangle = -i\widetilde{\mathbb{1}}H_{e}\widetilde{\mathbb {1}}|\widetilde{\Psi}(t)\rangle \nonumber \\ &\equiv& -i \widetilde{H}_{e}|\widetilde{\Psi}(t)\rangle, \label{tder_psi_1t}\end{aligned}$$ where, we have used the fact that $\widetilde{\mathbb{1}}$ is time-independent and therefore commutes with $\frac{\partial}{\partial t}$. Equation (\[tder\_psi\_1t\]) shows that $\widetilde{\Psi}(t)$ obeys a time-dependent Schrödinger equation with the effective Hamiltonian $\widetilde{H}_{e} = \widetilde{\mathbb{1}}H_{e}\widetilde{\mathbb {1}}$. The Hamiltonian $H_{e}$ has the conventional form of $H_{e} = V_{e} +T$, where $T$ is the kinetic-energy operator. The Hamiltonian $\widetilde{H}_{e}$ therefore takes the form: $$\begin{aligned} \widetilde{H}_{e} \equiv \widetilde{\mathbb{1}}H_{e}\widetilde{\mathbb {1}} = \widetilde{\mathbb{1}} V_{e} \widetilde{\mathbb{1}}+\widetilde{\mathbb{1}} T\widetilde{\mathbb {1}} \equiv \widetilde{V}_{e} + \widetilde{T}, \label{He_tilde}\end{aligned}$$ Note that the operator $\widetilde{\mathbb{1}}$ does not commute with $V_{e}$, $T$ or $H_{e}$ since it does not share a common basis of eigenvectors with the last three operators. Note also that the operators $\widetilde{V}_{e}$, $\widetilde{T}$ and $\widetilde{H}_{e}$ are all time-independent. Rearranging Eq. (\[tder\_psi\_1t\]), we obtain: $$\begin{aligned} \widetilde{V}_{e} &=& \frac{1}{\widetilde{\Psi}(t)} \left[ i\frac{\partial }{\partial t} - \widetilde{T} \right]\widetilde{\Psi}(t). \label{psitilde_SE}\end{aligned}$$ where we emphasize that $\widetilde{V}_{e}$ is time-independent. Let us now define the related quantity $$\begin{aligned} V &=& \frac{1}{\widetilde{\Psi}(t)} \left[ i\frac{\partial }{\partial t} - T \right]\widetilde{\Psi}(t), \label{fic_SE}\end{aligned}$$ where $T$ is the usual kinetic energy operator. Obviously, for $\widetilde{\Psi}(t) \equiv \Psi(t)$ Eq. (\[fic\_SE\]) is equivalent to the usual time-dependent Schrödinger equation for $\Psi(t)$ and therefore $V \equiv V_{e}$ is time-independent. We claim that for any other, incorrect, wavefunction $\widetilde{\Psi}(t)$, Eq. (\[fic\_SE\]) results in a time-*dependent* potential $V$. In order to show this we substitute $\widetilde{T} = T + \Delta T$ in Eq. (\[psitilde\_SE\]), where $\Delta T= \widetilde{T} - T$, and obtain: $$\begin{aligned} \widetilde{V}_{e} &=& \frac{1}{\widetilde{\Psi}(t)} \left[ i\frac{\partial }{\partial t} - (T + \Delta T) \right]\widetilde{\Psi}(t) \nonumber \\ &=& V - \frac{1}{\widetilde{\Psi}(t)} \left[ \Delta T \right]\widetilde{\Psi}(t). \label{psitilde_SE2}\end{aligned}$$ The term $\frac{1}{\widetilde{\Psi}(t)} [ \Delta T ] \widetilde{\Psi}(t)$ is time-*dependent* (unless $\widetilde{\Psi}(t)$ is an eigenfunction of $\Delta T$, which has no general reason to hold. In addition, in the Appendix we show that $\Delta T$ is generally different from zero). Therefore, in order to preserve the time-independence of $\widetilde{V}_{e} \equiv \widetilde{\mathbb{1}} V_{e} \widetilde{\mathbb{1}}$, $V$ must also be time-*dependent*. To summarize: in order to determine the correct wavefunction out of the set of wavefunctions $\{\widetilde{\Psi}_{i}(t) \}, i=1, 2, ..., 2^{N}$, we use the fictitious Schrödinger equation, (\[fic\_SE\]), to calculate a potential, $V$, from each wavefunction $\widetilde{\Psi}(t)$ of the set. At different times, $t$, the wavefunctions $\widetilde{\Psi}(t)$ will give different potentials $V$ except for the one correct wavefunction, $\Psi(t)$, that corresponds to the correct Schrödinger equation and therefore will give the same potential, $V \equiv V_{e}$, at all times. Thus, the correct wavefunction $\Psi(t)$ can be selected from the set $\{\widetilde{\Psi}_{i}(t)\}$ as the one that provides a time-*independent* potential via Eq. (\[fic\_SE\]). Alternatively, as described in the paper, the wavefunction $\widetilde{\Psi}(t)$ that propagates back to the known $\Psi(0) \equiv \psi_{0}$, using the corresponding potential calculated by Eq. (\[fic\_SE\]), is guaranteed to be the correct reconstructed wavefunction, $\Psi(t)$. Appendix -------- We show that $\Delta T \neq 0$: $$\begin{aligned} \Delta T &=& ( \widetilde{T}- T ) = \widetilde{\mathbb{1}}\widetilde{\mathbb{1}}(\widetilde{\mathbb{1}}T \widetilde{\mathbb{1}}-T ) \nonumber \\ &=& \widetilde{\mathbb{1}}(T\widetilde{\mathbb{1}} - \widetilde{\mathbb{1}}T) = \widetilde{\mathbb{1}}[T,\widetilde{\mathbb{1}}]. \label{T_Ttilde}\end{aligned}$$ The commutator $[T,\widetilde{\mathbb{1}}]$ is not identically zero. Therefore, $\widetilde{\mathbb{1}}[T,\widetilde{\mathbb{1}}] \equiv \Delta T$ is not identically zero as well.

ArXiv

--- author: - Greg Martin title: 'The Smallest Solution of $\phi(30n+1)<\phi(30n)$ is …' --- \#1[[ (mod $#1$)]{}]{} In a previous issue of the [*American Mathematical Monthly*]{}, D. J. Newman [@newman] showed that for any positive integers $a$, $b$, $c$, and $d$ with $ad\ne bc$, there exist infinitely many positive integers $n$ for which $\phi(an+b)<\phi(cn+d)$, where $\phi(m)$ is the familiar Euler totient function, the number of positive integers less than and relatively prime to $m$. In particular, it must be the case that $\phi(30n+1)<\phi(30n)$ infinitely often; however, Newman mentions that there are no solutions of this inequality with $n\le{}$20,000,000, and he states that a solution “is not explicitly available and it may be beyond the reach of any possible computers”. The purpose of this note is to describe a method for computing solutions to inequalities of this type that avoids the need to factor large numbers. In particular, we explicitly compute the smallest number $n$ satisfying $\phi(30n+1)<\phi(30n)$. It is quite easy to compute values of $n$ for which $\phi(30n+1)$ is relatively small by imposing many congruence conditions on $n$ modulo primes, so that $30n+1$ is highly composite. However, the numbers $n$ that arise in this way are quite large, having hundreds of digits. Computing $\phi(30n)$ exactly relies on the factorization of $30n$, which for integers of this size is not possible to find in a reasonable amount of time with today’s computers and factoring algorithms. The idea underlying our method is to use partial knowledge of the factorization of a large number $m$ to get an estimate for $\phi(m)$. We rely on the following claim: [*Proof.*]{} Let $t$ be the number of distinct prime factors of $m$, and let the prime factors be $p_{i_1}$, …, $p_{i_t}$ with $i_1<\dots<i_t$. Since none of the primes $p_1$, …, $p_r$ divide $m$, it must be the case that $i_1\ge r+1$, $i_2\ge r+2$, and so on. So if we define $k=\prod_{j=r+1}^{r+t}p_j$, we see that $k\le\prod_{j=1}^tp_{i_j}\le m$. But $m\le q$ by assumption, and so $k\le q$, which can clearly only be the case if $t\le s$. This proves part (a) of the claim. As for part (b), we use the fact that the function $\phi(m)/m$ can be written as a product over primes dividing $m$: $${\phi(m)\over m} = \prod_{p\mid m} \big( 1-\frac1p \big).$$ With $k$ defined as above, notice that $${\phi(m)\over m} = \prod_{j=1}^t \big( 1-\frac1{p_{i_j}} \big) \ge \prod_{j=1}^t \big( 1-\frac1{p_{r+j}} \big) = {\phi(k)\over k},$$ since $1-1/p$ is an increasing function of $p$. On the other hand, since $t\le s$ by assumption, we have $${\phi(k)\over k} = \prod_{j=r+1}^{r+t} \big( 1-\frac1{p_j} \big) \ge \prod_{j=r+1}^{r+s} \big( 1-\frac1{p_j} \big) = {\phi(q)\over q},$$ since each $1-1/p$ is less than 1. This proves part (b) of the claim. We now proceed to find the smallest solution of $\phi(30n+1)<\phi(30n)$, though it must be pointed out that the method applies to any inequality of the form $\phi(an+b)<\phi(cn+d)$. Clearly $30n+1\equiv1\mod{30}$ no matter what $n$ is. Also, if $n$ is a solution of $\phi(30n+1)<\phi(30n)$, then we must have $${\phi(30n+1)\over30n+1} < {\phi(30n)\over30n+1} < {\phi(30)n\over30n} = \frac4{15} = 0.26666\dots,$$ since the inequality $\phi(ab)\le\phi(a)b$ holds for any numbers $a$ and $b$. Thus it makes sense to look for numbers that satisfy both these conditions. [*Proof.*]{} A computation shows that $z$ is indeed congruent to $1\mod{30}$ and that $$\frac{\phi(z)}z = \bigg( \prod_{i=4}^{383} \big( 1-\frac1{p_i} \big) \bigg) \big( 1-\frac1{p_{385}} \big) \big( 1-\frac1{p_{388}} \big) = 0.2666117\ldots<\frac4{15}.$$ Suppose $m$ is an integer satisfying $m\equiv1\mod{30}$ and $\phi(m)/m<4/15$. Because of the congruence condition, $m$ cannot be divisible by 2, 3, or 5. If we define $q_1=\prod_{i=4}^{384}p_i$, then we can compute that $\phi(q_1)/q_1 = 0.26671\dots$, and so $\phi(q_1)/q_1>\phi(m)/m$. Thus if we apply part (b) of Claim 1 with $r=3$ and $s=381$, we conclude that $m$ must have more than 381 distinct prime factors. \#1\#2\#3[p\_[38\#1]{}p\_[38\#2]{}p\_[38\#3]{}]{} Another computation reveals that the only numbers with at least 382 distinct prime factors that are less than $z$ are the numbers $p_4p_5\dotsm p_{382}m'$, where $m'\in\{\ppp345$, $\ppp346$, $\ppp356$, $\ppp347$, $\ppp357$, $\ppp456$, $\ppp348$, $\ppp367\}$; and none of these numbers are congruent to $1\mod{30}$. Let us define $n=(z-1)/30$, which by Claim 2 is both an integer and the smallest possible solution of $\phi(30n+1)<\phi(30n)$. (Small wonder that we haven’t stumbled across any solutions of this inequality—$n$ has 1,116 digits!) It would be quite gracious of $n$ to be an actual solution, and indeed it is. First we show that $\phi(30n+1)/(30n+1)<\phi(30n)/30n$. We have already computed that $${\phi(30n+1)\over30n+1} = \frac{\phi(z)}z = 0.2666117\dots. \label{small}$$ It turns out that $n$ is divisible by both 60 and $p_{4,874}={}$47,279, so let us define $n'=n/(60p_{4,874})$. We can compute that $n'$ is not divisible by any of the first 80,000 primes. This computation can be done most quickly by multiplying the primes together in blocks of 1,000, say, and computing the greatest common divisor of $n'$ and the product. Since computing greatest common divisors is a very fast operation, checking that $n'$ is not divisible by any of the first 80,000 primes takes only a few minutes on a workstation—much more reasonable than trying to factor a number with over a thousand digits. Now define $q_2=\prod_{i=80,001}^{80,186}p_i$. We compute that $q_2$ has 1,118 digits and so $q_2>n>n'$. By using parts (a) and (b) of Claim 1 with $r={}$80,000 and $s=186$, we see that $\phi(n')/n'\ge\phi(q_2)/q_2$. Therefore, since $\phi(ab)=\phi(a)\phi(b)$ when $a$ and $b$ are relatively prime, we compute that $${\phi(30n)\over30n} = {\phi(30\cdot60p_{4,874})\over 30\cdot60p_{4,874}} {\phi(n')\over n'} \ge \frac4{15} \big( 1-\frac1{47,279} \big) \frac{\phi(q_2)}{q_2} = 0.2666124\dots. \label{big}$$ This shows that $\phi(30n+1)/(30n+1)<\phi(30n)/30n$, which doesn’t quite imply that $\phi(30n+1)<\phi(30n)$, but only $\phi(30n+1)<\phi(30n)(1+1/(30n))$. However, the numbers computed in (\[small\]) and (\[big\]) differ in the sixth decimal place, while multiplying by $1+1/(30n)$ leaves a number unchanged until past the 1100th decimal place. Therefore we have proved: [**Theorem. *The smallest solution of $\phi(30n+1)<\phi(30n)$ is***]{} $$\openup-1\jot %%% to get the following lines of digits closer %%% together, since we're using the \scriptstyle-sized %%% numbers \begin{split} \scriptstyle n=&\scriptstyle232,909,810,175,496,793,814,049,684,205,233,780,004,859,885,966,051,235,363,345,311,075,888,344,528,723,154,527,984,\\ &\scriptstyle260,176,895,854,182,634,802,907,109,271,610,432,287,652,976,907,467,574,362,400,134,090,318,355,962,121,476,785,712,\\ &\scriptstyle891,544,538,210,966,704,036,990,885,292,446,155,135,679,717,565,808,063,766,383,846,220,120,606,143,826,509,433,540,\\ &\scriptstyle250,085,111,624,970,464,541,380,934,486,375,688,208,918,750,640,674,629,942,465,499,369,036,578,640,331,759,035,979,\\ &\scriptstyle369,302,685,371,156,272,245,466,396,227,865,621,951,101,808,240,692,259,960,203,091,330,589,296,656,888,011,791,011,\\ &\scriptstyle416,062,631,565,320,593,772,287,118,913,728,608,997,901,791,216,356,108,665,476,306,080,740,121,528,236,888,680,120,\\ &\scriptstyle152,479,138,327,451,088,404,280,929,048,314,912,122,784,879,758,304,016,832,436,751,532,255,185,640,249,324,065,492,\\ &\scriptstyle491,511,072,521,585,980,547,438,748,689,307,159,363,481,233,965,802,331,725,033,663,862,618,957,168,974,043,547,448,\\ &\scriptstyle879,663,217,971,081,445,619,618,789,985,472,074,303,100,303,636,078,827,273,695,551,162,089,725,435,110,246,701,964,\\ &\scriptstyle021,045,849,081,811,604,427,331,227,553,783,590,821,510,091,607,567,178,842,569,576,699,548,038,217,673,171,895,383,\\ &\scriptstyle249,326,800,667,432,993,531,186,437,659,910,632,865,419,892,370,957,722,154,266,351,039,808,548,150,828,868,968,820,\\ &\scriptstyle675,198,820,381,135,523,646,361,202,383,915,218,571,017,801,463,011,491,108,784,343,253,284,393,511,650,254,506,597,\\ &\scriptstyle923,969,653,616,813,897,710,621,756,693,827,471,154,701,151,222,320,443,347,408,180,047,964,860. \end{split}$$ I would like to thank Mike Bennett for verifying the above computations and to acknowledge the support of National Science Foundation grant DMS 9304580. [1]{} D. J. Newman, Euler’s $\phi$ function on arithmetic progressions, [*Amer. Math. Monthly*]{} **104** (1997), 256–257. =cmti10 School of Mathematics, Institute for Advanced Study, Olden Lane, Princeton, NJ 08540, U.S.A. gerg@@math.ias.edu

ArXiv

--- author: - bibliography: - '0-string.bib' - 'bib.bib' title: '[**[SIBRA]{}: [Scalable Internet Bandwidth Reservation Architecture]{}**]{}' --- IEEEpubidpullup[9]{} Introduction {#sec:introduction} ============ \[sec:intro\] Goals, Assumptions, and the Adversary {#sec:probdef} ===================================== [SIBRA]{}Design {#sec:path} =============== Implementation {#sec:imple} ============== \[sec:implementation\] Evaluation {#sec:eval} ========== \[sec:evaluation\] Incremental Deployment ====================== \[sec:deployment\] Use Cases ========= Discussion {#sec:discussion} ========== \[sec:discuss\] Related Work {#sec:related} ============ Conclusions {#sec:conclusion} =========== Acknowledgments =============== We would like to thank Virgil Gligor, Chris Pappas, Christian Rossow, Stephen Shirley, and Laurent Vanbever for insightful discussions and their valuable comments throughout the evolution of this project. We also thank Xiaoyou Wang, Dominik Roos, and Takayuki Sasaki for their help with the implementation and evaluation of [SIBRA]{}. The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement 617605. We also gratefully acknowledge support by ETH Zurich, and NSF under award number CNS-1040801. The research was also supported by a gift from KDDI.

ArXiv

--- author: - 'Anqi Cheng,$^1$' - 'Anna Hasenfratz,$^1$' - 'Yuzhi Liu,$^1$' - Gregory Petropoulos$^1$ - 'and David Schaich$^{1, 2}$' bibliography: - 'gradflow.bib' title: | Improving the continuum limit\ of gradient flow step scaling --- Introduction ============ Asymptotically-free SU($N$) gauge theories coupled to $N_f$ massless fundamental fermions are conformal in the infrared if $N_f$ is sufficiently large, $N_f \geq N_f^{(c)}$. Their renormalization group (RG) [$\beta$ ]{}functions possess a non-trivial infrared fixed point (IRFP) where the gauge coupling is an irrelevant operator. Although this IRFP can be studied perturbatively for large $N_f$ near the value at which asymptotic freedom is lost [@Caswell:1974gg; @Banks:1981nn], as $N_f$ decreases the fixed point becomes strongly coupled. Systems around $N_f \approx N_f^{(c)}$ are particularly interesting strongly-coupled quantum field theories, with non-perturbative conformal or near-conformal dynamics. Their most exciting phenomenological application is the possibility of a light composite Higgs boson from dynamical electroweak symmetry breaking [@Fodor:2012ty; @Matsuzaki:2012xx; @Appelquist:2013sia; @Fodor:2014pqa; @Aoki:2014oha]. Due to the strongly-coupled nature of these systems, lattice gauge theory calculations are a crucial non-perturbative tool with which to investigate them from first principles. Many lattice studies of potentially IR-conformal theories have been carried out in recent years (cf. the recent reviews [@Neil:2012cb; @Giedt:2012it] and references therein). While direct analysis of the RG [$\beta$ ]{}function may appear an obvious way to determine whether or not a given system flows to a conformal fixed point in the infrared, in practice this is a difficult question to address with lattice techniques. In particular, extrapolation to the infinite-volume continuum limit is an essential part of such calculations. In the case of SU(3) gauge theory with $N_f = 12$ fundamental fermions, several lattice groups have investigated the step scaling function, the discretized form of the [$\beta$ ]{}function. To date, these studies either did not reach a definite conclusion [@Hasenfratz:2010fi; @Lin:2012iw] or may be criticized for not properly taking the infinite-volume continuum limit [@Appelquist:2007hu; @Appelquist:2009ty; @Hasenfratz:2010fi; @Hasenfratz:2011xn; @Itou:2012qn; @Petropoulos:2013gaa]. At the same time, complementary numerical investigations have been carried out, considering for example the spectrum, or bulk and finite-temperature phase transitions [@Deuzeman:2009mh; @Fodor:2011tu; @Appelquist:2011dp; @DeGrand:2011cu; @Cheng:2011ic; @Cheng:2013eu; @Fodor:2012uw; @Fodor:2012et; @Aoki:2012eq; @Aoki:2013zsa; @Jin:2012dw]. The different groups performing these studies have not yet reached consensus regarding the infrared behavior of the 12-flavor system. Our own $N_f = 12$ results favor the existence of a conformal IRFP, which we observe in Monte Carlo RG studies [@Hasenfratz:2011xn; @Petropoulos:2013gaa]. Our zero- and finite-temperature studies of the lattice phase diagram show a bulk transition consistent with conformal dynamics [@Schaich:2012fr; @Hasenfratz:2013uha]. From the Dirac eigenvalue spectrum [@Cheng:2013eu], and from finite-size scaling of mesonic observables [@Cheng:2013xha], we obtain consistent predictions for a relatively small fermion mass anomalous dimension: ${\ensuremath{\gamma} }_m^{\star} = 0.32(3)$ and 0.235(15), respectively. While this conclusion, if correct, would render the 12-flavor system unsuitable for composite Higgs phenomenology, we consider $N_f = 12$ to remain an important case to study. Considerable time and effort has already been invested to obtain high-quality lattice data for the 12-flavor system. Until different methods of analyzing and interpreting these data can be reconciled – or the causes of any remaining disagreements can be clarified – it will not be clear which approaches are most reliable and most efficient to use in other contexts. The recent development of new running coupling schemes based on the gradient flow [@Luscher:2009eq; @Luscher:2010iy; @Fodor:2012td; @Fodor:2012qh; @Fritzsch:2013je] provides a promising opportunity to make progress. In this work we investigate step scaling using the gradient flow running coupling.[^1] We begin by introducing a non-perturbative improvement to this technique, which increases our control over the continuum extrapolation by reducing the leading-order cut-off effects. While this improvement is phenomenological in the sense that we have not derived it systematically through a full improvement program, it is generally applicable to any lattice gauge theory of interest and can remove all ${\ensuremath{\mathcal O} }(a^2)$ cut-off effects. We illustrate it first for 4-flavor SU(3) gauge theory, a system where the running coupling has previously been studied with both Wilson [@Tekin:2010mm] and staggered [@PerezRubio:2010ke; @Fodor:2012td; @Fodor:2012qh] fermions. We then turn to $N_f = 12$, where we show that the infinite-volume continuum limit is well defined and predicts an IRFP. In both the 4- and 12-flavor systems, our improvement can remove all observable ${\ensuremath{\mathcal O} }(a^2)$ effects, despite the dramatically different IR dynamics. We conclude with some comments on other systems where improved gradient flow step scaling may profitably be applied. Improving gradient flow step scaling ==================================== The gradient flow is a continuous invertible smearing transformation that systematically removes short-distance lattice cut-off effects [@Luscher:2009eq; @Luscher:2010iy]. At flow time $t = a^2 t_{\rm{lat}}$ it can be used to define a renormalized coupling at scale $\mu = 1 / \sqrt{8t}$ $$\label{eq:def_g2} {\ensuremath{g_{\rm GF}^2} }(\mu = 1 / \sqrt{8t}) = \frac{1}{{\ensuremath{\mathcal N} }} {\ensuremath{\left\langle t^2 E(t) \right\rangle} },$$ where “$a$” is the lattice spacing, $t_{\rm{lat}}$ is dimensionless, and the energy density $E(t) = -\frac{1}{2}\mbox{ReTr}\left[G_{\mu\nu}(t) G^{\mu\nu}(t)\right]$ is calculated at flow time $t$ with an appropriate lattice operator. We evolve the gradient flow with the Wilson plaquette term and use the usual “clover” or “symmetric” definition of $G_{\mu\nu}(t)$. The normalization [$\mathcal N$ ]{}is set such that ${\ensuremath{g_{\rm GF}^2} }(\mu)$ agrees with the continuum [$\overline{\mbox{MS}}$ ]{}coupling at tree level. If the flow time is fixed relative to the lattice size, $\sqrt{8t} = cL$ with $c$ constant, the scale of the corresponding coupling ${\ensuremath{g_c^2} }(L)$ is set by the lattice size. Like the well-known Schrödinger functional (SF) coupling, ${\ensuremath{g_c^2} }(L)$ can be used to compute a step scaling function [@Fodor:2012td; @Fodor:2012qh; @Fritzsch:2013je]. The greater flexibility of the gradient flow running coupling is a significant advantage over the more traditional SF coupling. A single measurement of the gradient flow will provide [$g_c^2$ ]{}for a range of $c$. In our study we obtain [$g_c^2$ ]{}for all $0 \leq c \leq 0.5$ separated by ${\ensuremath{\delta} }t_{\rm{lat}} = 0.01$. Each choice of $c$ corresponds to a different renormalization scheme, which can be explored simultaneously on the same set of configurations [@Fritzsch:2013je]. The normalization factor [$\mathcal N$ ]{}in finite volume has been calculated for anti-periodic boundary conditions (BCs) in refs. [@Fodor:2012td; @Fodor:2012qh], and for SF BCs in [ref. [@Fritzsch:2013je]]{}. In this work we use anti-periodic BCs, for which $$\begin{aligned} \frac{1}{{\ensuremath{\mathcal N} }} & = \frac{128\pi^2}{3(N^2 - 1)(1 + {\ensuremath{\delta} }(c))} & {\ensuremath{\delta} }(c) & = \vartheta^4\left(e^{-1 / c^2}\right) - 1 - \frac{c^4 \pi^2}{3},\end{aligned}$$ where $\vartheta(x) = \sum_{n = -\infty}^{\infty} x^{n^2}$ is the Jacobi elliptic function. For $0 \leq c \leq 0.3$ the finite-volume correction ${\ensuremath{\delta} }(c)$ computed in [ref. [@Fodor:2012td]]{} is small, $|{\ensuremath{\delta} }(c)| \leq 0.03$. As explained in refs. [@Fodor:2012td; @Fodor:2012qh], the RG [$\beta$ ]{}function of [$g_{\rm GF}^2$ ]{}is two-loop universal with SF BCs, but only one-loop universal with anti-periodic BCs. At non-zero lattice spacing [$g_{\rm GF}^2$ ]{}has cut-off corrections. These corrections could be ${\ensuremath{\mathcal O} }(a)$ for unimproved actions, and even ${\ensuremath{\mathcal O} }(a)$-improved actions could have large ${\ensuremath{\mathcal O} }(a^2 [\log a]^n)$-type corrections [@Balog:2009yj; @Balog:2009np]. In existing numerical studies of staggered or ${\ensuremath{\mathcal O} }(a)$-improved Wilson fermions the leading lattice corrections appear to be ${\ensuremath{\mathcal O} }(a^2)$ [@Fritzsch:2013je; @Sommer:2014mea], $$\label{eq:lat_g2} {\ensuremath{g_{\rm GF}^2} }(\mu; a) = {\ensuremath{g_{\rm GF}^2} }(\mu; a = 0) + a^2 {\ensuremath{\mathcal C} }+ {\ensuremath{\mathcal O} }(a^4 [\log a]^n, a^4).$$ It is possible to remove, or at least greatly reduce, the ${\ensuremath{\mathcal O} }(a^2)$ corrections in [eq. \[eq:lat\_g2\]]{} by defining $$\label{eq:t-shift} {\ensuremath{\widetilde g_{\rm GF}^2} }(\mu; a) = \frac{1}{{\ensuremath{\mathcal N} }} {\ensuremath{\left\langle t^2 E(t + \tau_0 a^2) \right\rangle} }\, ,$$ where $\tau_0 \ll t / a^2$ is a small shift in the flow time. In the continuum limit $\tau_0 a^2 \to 0$ and ${\ensuremath{\widetilde g_{\rm GF}^2} }(\mu) = {\ensuremath{g_{\rm GF}^2} }(\mu)$. There are several possible interpretations of the $t$-shift in [eq. \[eq:t-shift\]]{}. The gradient flow is an invertible smearing transformation, so one can consider $\tau_0$ as an initial flow that does not change the IR properties of the system but leads to a new action. The gradient flow coupling [$\widetilde g_{\rm GF}^2$ ]{}in [eq. \[eq:t-shift\]]{} is calculated for this new action. Alternatively one can consider the replacement of ${\ensuremath{\left\langle t^2 E(t) \right\rangle} }$ with ${\ensuremath{\left\langle t^2 E(t + \tau_0 a^2) \right\rangle} }$ as an improved operator for the energy density. In either case the $t$-shift changes the ${\ensuremath{\mathcal O} }(a^2)$ term of ${\ensuremath{g_{\rm GF}^2} }(\mu; a)$. If we expand ${\ensuremath{\widetilde g_{\rm GF}^2} }(\mu)$ in $\tau_0 a^2$, $$\label{eq:expand} {\ensuremath{\widetilde g_{\rm GF}^2} }(\mu; a) = \frac{1}{{\ensuremath{\mathcal N} }} {\ensuremath{\left\langle t^2 E(t) \right\rangle} } + \frac{a^2 \tau_0}{{\ensuremath{\mathcal N} }} {\ensuremath{\left\langle t^2 \frac{\partial E(t)}{\partial t} \right\rangle} },$$ and choose $\tau_0$ such that the second term in [eq. \[eq:expand\]]{} cancels the $a^2 {\ensuremath{\mathcal C} }$ term in [eq. \[eq:lat\_g2\]]{}, we remove the leading lattice artifacts $${\ensuremath{\widetilde g_{\rm opt}^2} }(\mu; a) = {\ensuremath{g_{\rm GF}^2} }(\mu; a = 0) + {\ensuremath{\mathcal O} }(a^4 [\log a]^n, a^4).$$ Full ${\ensuremath{\mathcal O} }(a^2)$ improvement through a systematic improvement program would require adding terms to improve the flow equation, the action, the boundary conditions, and the energy density operator ${\ensuremath{\left\langle t^2 E(t) \right\rangle} }$ [@Sommer:2014mea]. Since our proposed improvement involves only a single parameter $\tau_0$, this $\tau_0$ itself must depend on other parameters, most importantly on ${\ensuremath{\widetilde g_{\rm GF}^2} }(\mu)$ and on the bare coupling through the lattice spacing dependence of the term ${\ensuremath{\left\langle t^2 \frac{\partial E(t)}{\partial t} \right\rangle} }$ in [eq. \[eq:expand\]]{}. Optimizing $\tau_0$ both in the renormalized and bare couplings could remove the predictive power of the method. Fortunately, as we will see in the next section, our numerical tests indicate that it is sufficient to choose $\tau_0$ to be a constant or only weakly ${\ensuremath{\widetilde g_{\rm GF}^2} }(\mu)$ dependent to remove most ${\ensuremath{\mathcal O} }(a^2)$ lattice artifacts. Since the gradient flow is evaluated through numerical integration, the replacement ${\ensuremath{g_{\rm GF}^2} }\to {\ensuremath{\widetilde g_{\rm GF}^2} }$ can be done by a simple shift of $t$ without incurring any additional computational cost. The optimal $t$-shift [$\tau_{\rm opt}$ ]{}can be identified by a simple procedure when the gradient flow is used for scale setting, which we will consider in a future publication. In this paper we concentrate on the step scaling function and find the [$\tau_{\rm opt}$ ]{}that removes the ${\ensuremath{\mathcal O} }(a^2)$ terms of the discrete [$\beta$ ]{}function corresponding to scale change $s$, $$\label{eq:beta_lat} {\ensuremath{\beta} }_{\rm lat}({\ensuremath{g_c^2} }; s; a) = \frac{{\ensuremath{\widetilde g_c^2} }(L; a) - {\ensuremath{\widetilde g_c^2} }(sL; a)}{\log(s^2)}.$$ Testing improvement with 4-flavor SU(3) gauge theory ==================================================== ![\[fig:Nf4\] Continuum extrapolations of the discrete ${\ensuremath{\beta} }_{\rm lat}$ function of the $N_f = 4$ system at ${\ensuremath{\widetilde g_c^2} }(L) = 2.2$ with several different values of the $t$-shift coefficient $\tau_0$. The dotted lines are independent linear fits at each $\tau_0$, which predict a consistent continuum value.](dg_nf4_c025_22){width="0.75\linewidth"} ![\[fig:Nf4\_many\] Continuum extrapolations of the discrete ${\ensuremath{\beta} }_{\rm lat}$ function of the $N_f = 4$ system for several different ${\ensuremath{\widetilde g_c^2} }(L)$ values. For ${\ensuremath{\widetilde g_c^2} }(L) = 1.8$, 2.2 and 2.6 $\tau_0 = -0.02$ is near-optimal, while the larger couplings ${\ensuremath{\widetilde g_c^2} }(L) = 3.0$ and 3.4 require $\tau_0 = -0.01$ to remove most ${\ensuremath{\mathcal O} }(a^2)$ effects. The colored points at $(a / L)^2 = 0$ are the continuum extrapolated results, while the black crosses at $(a / L)^2 = 0$ show the corresponding two-loop perturbative predictions.](dg_nf4_c025_many){width="0.75\linewidth"} We illustrate the $t$-shift improvement with the $N_f = 4$ SU(3) system. This theory was recently studied by refs. [@Fodor:2012td; @Fodor:2012qh] using gradient flow step scaling with staggered fermions. The 4-flavor SF running coupling was previously considered in [ref. [@Tekin:2010mm]]{} using ${\ensuremath{\mathcal O} }(a)$-improved Wilson fermions, and in [ref. [@PerezRubio:2010ke]]{} using staggered fermions. In our calculations we use nHYP-smeared [@Hasenfratz:2001hp; @Hasenfratz:2007rf] staggered fermions and a gauge action that includes an adjoint plaquette term in order to move farther away from a well-known spurious fixed point in the adjoint–fundamental plaquette plane [@Cheng:2011ic]. As in [ref. [@Fodor:2012td]]{} we impose anti-periodic BCs in all four directions, which allows us to carry out computations with exactly vanishing fermion mass, $m = 0$. For the discrete [$\beta$ ]{}function we consider the scale change $s = 3 / 2$ and compare lattice volumes $12^4 \to 18^4$, $16^4 \to 24^4$ and $20^4 \to 30^4$. We accumulated 500–600 measurements of the gradient flow coupling, with each measurement separated by 10 molecular dynamics time units (MDTU), at 7–8 values of the bare gauge coupling on each volume. We consider the $c = 0.25$ scheme, as opposed to $c = 0.3$ used in [ref. [@Fodor:2012td]]{}, because smaller $c$ gives better statistics at the expense of larger lattice artifacts. As discussed above, we aim to reduce these lattice artifacts through the non-perturbative improvement we have introduced. We follow the fitting procedure described in [ref. [@Tekin:2010mm]]{}. Full details of this study will be presented in [ref. [@Cheng:2014]]{}. Here we provide a representative illustration of the $t$-shift optimization. Figure \[fig:Nf4\] shows the dependence of the discrete [$\beta$ ]{}function on $(a / L)^2$ when ${\ensuremath{\widetilde g_c^2} }(L) = 2.2$ with several values of the $t$-shift parameter $\tau_0$. The red triangles correspond to no improvement, $\tau_0 = 0$. The data are consistent with linear dependence on $a^2$ and extrapolate to 0.262(17), about $2\sigma$ below the two-loop perturbative value of 0.301. The slope of the extrapolation is already rather small, $b = 11(3)$. By adding a small shift this slope can be increased or decreased. With $\tau_0 = -0.02$ no ${\ensuremath{\mathcal O} }(a^2)$ effects can be observed – the corresponding slope is $b = 1.5(3.1)$ – and we identify this value as near the optimal ${\ensuremath{\tau_{\rm opt}} }$. The data at different $\tau_0$ extrapolate to the same continuum value, even when the slope $b$ is larger than that for $\tau_0 = 0$. This is consistent with the expectation that the $t$-shift changes the ${\ensuremath{\mathcal O} }(a^2)$ behavior of the system but does not affect the continuum limit. Since our action produces relatively small ${\ensuremath{\mathcal O} }(a^2)$ corrections even without improvement, the $t$-shift optimization has little effect on the continuum extrapolation, though the consistency between different values of $\tau_0$ is reassuring. It is interesting that the cut-off effects in our unimproved results, characterized by the slope $b$ of the red triangles in [figure \[fig:Nf4\]]{}, are more than three times smaller than those shown in fig. 4 of [ref. [@Fodor:2012td]]{}. This difference grows to about a factor of four when we consider the larger $c = 0.3$ used in that study, suggesting that the $t$-shift optimization could have a more pronounced effect with the action used in [ref. [@Fodor:2012td]]{}. The cause of the reduced lattice artifacts with our action is not obvious. Both our action and that used by [ref. [@Fodor:2012td]]{} are based on smeared staggered fermions, though we use different smearing schemes. The different smearing might have an effect, as might the inclusion of the adjoint plaquette term in our gauge action. This question is worth investigating in the future. In principle [$\tau_{\rm opt}$ ]{}could be different at different [$g_c^2$ ]{}couplings but in practice we found little variation. Figure \[fig:Nf4\_many\] shows near-optimal continuum extrapolations of the discrete [$\beta$ ]{}function at several values of ${\ensuremath{\widetilde g_c^2} }(L)$. At each ${\ensuremath{\widetilde g_c^2} }(L)$ the continuum extrapolated result is consistent within $\sim$$2\sigma$ with the two-loop perturbative prediction, denoted by a black cross in [figure \[fig:Nf4\_many\]]{}. Comparable consistency with perturbation theory was found in previous studies [@Tekin:2010mm; @PerezRubio:2010ke; @Fodor:2012td; @Fodor:2012qh]. Infrared fixed point in 12-flavor SU(3) gauge theory ==================================================== ![\[fig:gradient\_flow\]The $N_f = 12$ running coupling ${\ensuremath{g_c^2} }(L)$ versus the bare coupling ${\ensuremath{\beta} }_F$ on several volumes, for $c = 0.2$. Crossings between results from different volumes predict the finite volume IRFP coupling ${\ensuremath{g_{\star}^2} }(L)$ in this scheme.](wflow){width="0.75\linewidth"} We use the same lattice action with $N_f = 12$ as with $N_f = 4$ and consider six different volumes: $12^4$, $16^4$, $18^4$, $24^4$, $32^4$ and $36^4$. This range of volumes allows us to carry out step scaling analyses with scale changes $s = 4 / 3$, $3 / 2$ and 2. As for $N_f = 4$ we performed simulations in the $m = 0$ chiral limit with anti-periodic BCs in all four directions. Depending on the volume and bare coupling ${\ensuremath{\beta} }_F$ we accumulated 300–1000 measurements of the gradient flow coupling [$g_c^2$ ]{}for $0 \leq c \leq 0.5$, with 10 MDTU separating subsequent measurements. Here we will consider only $c = 0.2$. Full details of our ensembles and measurements, studies of their auto-correlations, and additional analyses for $c = 0.25$ and 0.3 will appear in [ref. [@Cheng:2014]]{}. The choice of $c = 0.2$ minimizes the statistical errors, and we find the IRFP in this scheme to be at a weaker coupling than for larger $c$, which is numerically easier to reach. The typical trade-off for these smaller statistical errors would be larger cut-off effects, but as discussed in previous sections these cut-off effects can be reduced by our non-perturbative improvement. Figure \[fig:gradient\_flow\] shows the running coupling ${\ensuremath{g_c^2} }(L)$ as the function of the bare gauge coupling ${\ensuremath{\beta} }_F$ for different volumes. The interpolating curves are from fits similar to those in [ref. [@Tekin:2010mm]]{}. The curves from different volumes cross in the range $6.0 \leq {\ensuremath{\beta} }_F \leq 6.5$. The crossing from lattices with linear size $L$ and $sL$ defines the finite-volume IRFP coupling ${\ensuremath{g_{\star}^2} }(L; s)$: $${\ensuremath{g_c^2} }(L) = {\ensuremath{g_c^2} }(sL) \implies {\ensuremath{g_{\star}^2} }(L; s) = {\ensuremath{g_c^2} }(L).$$ If the IRFP exists in the continuum limit then the extrapolation $$\lim_{(a / L)^2 \to 0} {\ensuremath{g_{\star}^2} }(L; s) \equiv {\ensuremath{g_{\star}^2} }$$ has to be finite and independent of $s$.[^2] Figure \[fig:Nf12\] illustrates the continuum extrapolation of ${\ensuremath{g_{\star}^2} }(L)$ with scale change $s = 2$ for various choices of the $t$-shift parameter $\tau_0$. The red triangles correspond to no shift, $\tau_0 = 0$. Their $(a / L)^2 \to 0$ continuum extrapolation has a negative slope, and the leading lattice cut-off effects are removed with a positive $t$-shift, ${\ensuremath{\tau_{\rm opt}} }\approx 0.04$. A joint linear extrapolation of the $\tau_0 = 0$, 0.02, 0.04 and 0.06 results, constrained to have the same continuum limit at $(a / L)^2 = 0$, predicts ${\ensuremath{g_{\star}^2} }= 6.21(25)$. However, these results all come from the same measurements, and are therefore quite correlated. While it is an important consistency check that the continuum limit does not change with ${\ensuremath{\tau_{\rm opt}} }$, just as for $N_f = 4$, the uncertainty in the continuum-extrapolated [$g_{\star}^2$ ]{}from this joint fit is not reliable. Instead, we should consider only the results with the near-optimal ${\ensuremath{\tau_{\rm opt}} }\approx 0.04$. As we show in [figure \[fig:Nf12\_all\]]{}, ${\ensuremath{\tau_{\rm opt}} }\approx 0.04$ is also near-optimal for scale changes $s = 3 / 2$ and $4 / 3$. None of these results have any observable ${\ensuremath{\mathcal O} }(a^2)$ effect, making the extrapolation to the continuum very stable. Each scale change predicts a continuum IRFP for $N_f = 12$. The three sets of results in [figure \[fig:Nf12\_all\]]{} come from matching different volumes, making a joint fit legitimate. This continuum extrapolation predicts that the IR fixed point is located at renormalized coupling ${\ensuremath{g_{\star}^2} }= 6.18(20)$ in the $c = 0.2$ scheme. ![\[fig:Nf12\]Continuum extrapolations of the 12-flavor finite volume IRFP ${\ensuremath{g_{\star}^2} }(L)$, with several different $t$-shift coefficients $\tau_0$ for fixed scale change $s = 2$. The dotted lines are a joint linear fit constrained to have the same $(a / L)^2 = 0$ intercept, which gives ${\ensuremath{g_{\star}^2} }= 6.21(25)$.](fit_s2_c02){width="0.75\linewidth"} ![\[fig:Nf12\_all\]Continuum extrapolations of the 12-flavor finite volume IRFP ${\ensuremath{g_{\star}^2} }(L)$, with several different scale changes for the near-optimal improvement coefficient ${\ensuremath{\tau_{\rm opt}} }\approx 0.04$. The $s = 4 / 3$ and $3 / 2$ data points are horizontally displaced for greater clarity. The dashed lines are a joint linear fit constrained to have the same $(a / L)^2 = 0$ intercept, which gives ${\ensuremath{g_{\star}^2} }= 6.18(20)$.](fit_c02_tau004){width="0.75\linewidth"} Conclusion and summary ====================== We have considered step scaling based on the gradient flow renormalized coupling, introducing a non-perturbative ${\ensuremath{\mathcal O} }(a^2)$ improvement that removes, or at least greatly reduces, leading-order cut-off effects. This phenomenological improvement increases our control over the extrapolation to the infinite-volume continuum limit, as we demonstrated first for the case of SU(3) gauge theory with $N_f = 4$ massless staggered fermions. Turning to $N_f = 12$, we found that the continuum limit was well defined and predicted an infrared fixed point even without improvement. Applying our proposed improvement reinforced this conclusion by removing all observable ${\ensuremath{\mathcal O} }(a^2)$ effects. For the finite-volume gradient flow renormalization scheme defined by $c = 0.2$, we find the continuum conformal fixed point to be located at ${\ensuremath{g_{\star}^2} }= 6.18(20)$. The 12-flavor system has been under investigation for some time, and other groups have studied its step scaling function [@Lin:2012iw; @Appelquist:2007hu; @Appelquist:2009ty; @Itou:2012qn]. However, this work is the first to observe an IRFP in the infinite-volume continuum limit. There are likely several factors contributing to this progress. While we did not invest more computer time than other groups, we have employed a well-designed lattice action. The adjoint plaquette term in our gauge action moves us farther away from a well-known spurious fixed point, while nHYP smearing allows us to simulate at relatively strong couplings. The gradient flow coupling itself appears to be a significant improvement over other schemes,[^3] and our non-perturbative improvement also contributes to obtaining more reliable continuum extrapolations. Our non-perturbative improvement is general and easy to use in other systems. It does not rely on the lattice action or fermion discretization, though we suspect that the improvement may not be effective if there are ${\ensuremath{\mathcal O} }(a)$ artifacts, e.g. for unimproved Wilson fermions. Since ${\ensuremath{\mathcal O} }(a)$-improved lattice actions are standard, this does not appear to be a practical limitation. We look forward to seeing our proposal applied both to QCD and to other conformal or near-conformal systems. Acknowledgments {#acknowledgments .unnumbered} =============== We thank J. Kuti for suggesting a perturbative interpretation of our proposed improvement, and D. Nógrádi for useful discussions of the continuum limit. Both J. Kuti and C.-J. D. Lin have kindly told us about their groups’ ongoing investigations of the $N_f = 12$ gradient flow step scaling function. A. H. is grateful for the hospitality of the Brookhaven National Laboratory HET group and of the Kobayashi–Maskawa Institute at Nagoya University during her extended visits, as well as the Japan Society for the Promotion of Science Fellowship that made the latter visit possible. This research was partially supported by the U.S. Department of Energy (DOE) through Grant Nos. DE-SC0010005 (A.C., A.H., Y.L. and D.S.), DE-SC0008669 and DE-SC0009998 (D.S.), and by the DOE Office of Science Graduate Fellowship Program under Contract No. DE-AC05-06OR23100 (G.P.). Our code is based in part on the MILC Collaboration’s public lattice gauge theory software.[^4] Numerical calculations were carried out on the University of Colorado HEP-TH cluster and on the Janus cluster partially funded by U.S. National Science Foundation (NSF) Grant No. CNS-0821794; at Fermilab under the auspices of USQCD supported by the DOE; and at the San Diego Computing Center through XSEDE supported by NSF Grant No. OCI-1053575. [^1]: We are aware of two other ongoing investigations of the $N_f = 12$ gradient flow step scaling function, by the authors of [ref. [@Fodor:2012td]]{} and [ref. [@Lin:2012iw]]{}. [^2]: We thank D. Nógrádi for useful discussions of the continuum limit. [^3]: C.-J. D. Lin has told us about dramatic improvements in auto-correlations when using the gradient flow coupling compared to the twisted Polyakov loop coupling of [ref. [@Lin:2012iw]]{}. [^4]: `http://www.physics.utah.edu/\simdetar/milc/`

ArXiv

--- abstract: 'We identify various structures on the configuration space $C$ of a flying saucer, moving in a three-dimensional smooth manifold $M$. Always $C$ is a five-dimensional contact manifold. If $M$ has a projective structure, then $C$ is its twistor space and is equipped with an almost contact Legendrean structure. Instead, if $M$ has a conformal structure, then the saucer moves according to a CR structure on $C$. With yet another structure on $M$, the contact distribution in $C$ is equipped with a cone over a twisted cubic. This defines a certain type of Cartan geometry on $C$ (more specifically, a type of ‘parabolic geometry’) and we provide examples when this geometry is ‘flat,’ meaning that its symmetries comprise the split form of the exceptional Lie algebra ${\mathfrak{g}}_2$.' address: - | -School of Mathematical Sciences\ University of Adelaide\ SA 5005\ Australia - '-Centrum Fizyki Teoretycznej, Polska Akademia Nauk, Al. Lotników 32/46, 02-668 Warszawa, Poland' author: - Michael Eastwood - Paweł Nurowski title: Aerodynamics of flying saucers --- [^1] Introduction ============ Throughout this article $M$ will be a $3$-dimensional smooth oriented manifold. For $x\in M$, a non-zero element $\omega\in T_x^*M$ defines an oriented $2$-plane $\{X\in T_xM\mid X{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\omega=0\}$ at $x$. (40,25)(0,33) (37.32,55) (34.82,59.33) (17.50,49.33) (17.50,49.33) (0.18,39.33) (2.68,35) (2.68,35) (5.18,30.67) (22.5,40.67) (22.5,40.67) (39.82,50.67) (37.32,55) (20,45)[(0,0)[$\bullet$]{}]{} (20,45)[(-2,3)[6]{}]{} (20.15,45.1)[(-2,3)[5]{}]{} (19.85,44.9)[(-2,3)[5]{}]{} Thus, we may realise the bundle of oriented two-planes in $TM$ as $$\label{C} \frac{\{\omega\in T^*M\setminus\{\mbox{the zero section}\}\}} {\omega\sim\lambda\omega\enskip\mbox{for}\enskip\lambda>0} ={\mathrm{Gr}}_2^+(TM)\xrightarrow{\,\pi\,}M.$$ We shall write $C\xrightarrow{\,\pi\,}M$ for this [*configuration space*]{} of oriented saucers in $M$ (flying saucers are oriented as they traditionally have a cockpit). The basic intrinsic structure on $C$ is a contact distribution $H\subset TC$. We shall see that a saucer moves along a path in $C$ that is everywhere tangent to $H$ if and only if its motion in $M$ is in directions taken from its own disc. With this specification, arbitrary ‘rolls’ are allowed but, with more structure on $M$, these rolls are constrained and this gives rise to differential geometries on $C$ expressed in terms of various algebraic structures on $H$. This article is organised as follows. Section \[contact\] discusses the contact geometry on $C$. In Section \[attack\] it is supposed that $M$ has a projective structure and consequently we shall find an almost contact Legendrean structure on $C$ (originally due to Takeuchi [@T] using methods due to Tanaka). Then $M$ is projectively flat if and only if $C$ has maximal symmetry ${\mathfrak{sl}}(4,{\mathbb{R}})$. Usual CR structures emerge in Section \[landing\] (following LeBrun [@LeB1]). In each case there are links with twistor theory. Section \[G2\] explains how to endow $C$ with a geometric structure modelled on the contact homogeneous space for the split form of the exceptional Lie group $G_2$ and we present some examples for which this structure turns out to be ‘flat,’ meaning that it is locally isomorphic to the flat model. Section \[geometry\_on\_M\] further investigates the geometry on $M$ that is needed to construct this ‘$G_2$ contact structure’ on $C$. This article is concerned only with the geometry of $C$ and especially its construction from, and relationship to, various geometrical features on $M$. In a companion article [@EN], we simply start with Euclidean space $M={\mathbb{R}}^3$ and explain how the various aerodynamic options considered here are reflected in the aerobatic man[œ]{}uvres available to a pilot flying according to these options. The authors would like to thank Katja Sagerschnig and Travis Willse for many helpful conversations. The contact structure on $C$ {#contact} ============================ In fact, the intrinsic contact structure on $C$ is, in addition, filtered. Specifically, we shall find canonically defined subbundles $$TC\supset H\supset V,$$ where $H$ is the contact distribution and $V$ is the vertical subbundle of $\pi:C\to M$. To define $H$, we note the canonical identification $$TC/V=\pi^*TM$$ and observe that a point in $C$ is precisely a point $x\in M$ together with an oriented $2$-plane in $T_xM$. In other words, we have a tautologically defined rank $2$ subbundle $P\subset\pi^*TM$ recording this subspace and we may define $H$ as the inverse image of $P$ under $TC\to\pi^*TM$. In summary, we have a canonical filtration $$\label{canonical_filtration} TC=\overbrace{\vphantom{(}L+{}\hspace{10pt}}^{\makebox[0pt]{$=\pi^*TM$}} \hspace{-11pt}\underbrace{P+V\!}_{\textstyle= H}\,,$$ where $L$ is, by definition, the line bundle $TC/H$ and we are recording here the composition factors, with the rightmost bundle $V$ being the natural subbundle. It remains to see that $H$ is, indeed, a contact distribution. This is a calculation in local coördinates. Specifically, we recall that the cotangent bundle $T^*M$ of any smooth manifold is equipped with the well-known tautological $1$-form $\theta$. In ‘canonical coördinates’ $(x^a,p_a)$ on $T^*M$, we have $\theta=p_a\,dx^a$ (for details, see [@C]). On $C$, we may use an affine chart $(x,y,z,a,b)\mapsto(x,y,z,a,b,1)$ to embed $C$ in $T^*M$ and pull-back $\theta$ to the $1$-form $a\,dx+b\,dy+dz$ whose kernel is $H$. Then $$d(a\,dx+b\,dy+dz)=da\wedge dx+db\wedge dy$$ is the Levi form on $H$, which is manifestly non-degenerate. Alternatively, the Levi form on $H$ may be seen as arising from the canonical symplectic form on $T^*M$ as follows. In canonical coördinates $(x^a,p_a)$ on $T^*M$, the symplectic form is $d\theta=dp_a\wedge dx^a$. It means that if we use canonical coördinates on the total space of $T^*M\xrightarrow{\,\nu\,}M$ to split its tangent bundle as $$\label{splitting}T(T^*M)= \begin{array}c\nu^*TM\\[-4pt] \oplus\\[-2pt] \nu^*T^*M\end{array}\ni \left[\!\!\begin{array}{c}X^a\\ \omega_a\end{array}\!\!\right] \leftarrowtail \left[\!\!\begin{array}{c}0\\ p_a\end{array}\!\!\right]=\theta$$ then the symplectic form is $$\label{canonical_symplectic_form} \left[\!\!\begin{array}{c}X^a\\ \omega_a\end{array}\!\!\right]\otimes \left[\!\!\begin{array}{c}\tilde X^b\\ \tilde\omega_b\end{array}\!\!\right] \longmapsto X^b\tilde\omega_b-\tilde X^b\omega_b,$$ (and is independent of choice of coördinates (as we shall see in the next section by a different argument)). Viewing $C$ as in (\[C\]), we see its tangent bundle as $$\left\{\left[\!\!\begin{array}{c}X^a\\ \omega_a\end{array}\!\!\right]\right\} \raisebox{-3pt}{\LARGE$/$}\begin{array}{l} \omega_a\sim\omega_a+tp_a\\ \mbox{for}\enskip t\in{\mathbb{R}}\end{array}$$ and $H$ as the subbundle for which $X^ap_a=0$. Evidently, the form (\[canonical\_symplectic\_form\]) descends to $H$ (and is easily verified to be the Levi form). Flying tangent to $H$ in $C$ is saying exactly that the velocity of the saucer in space is constrained to lie in its own disc. Otherwise, the pilot is free to make arbitrary ‘rolls’ and the Chow–Rashevskii Theorem [@M] in this context implies that a pilot flying with these man[œ]{}uvres may park her craft in an arbitrary location and orientation. Finally in this section, we consider the abstract structure on $C$ arising from its being a configuration space. Recall from (\[canonical\_filtration\]) that $C$ is equipped with a filtration on its tangent bundle $$\label{LPV}TC=L+P+V$$ in which $H=P+V$ is contact and the two-dimensional subbundle $V$ is integrable. In fact, there are no local invariants of this arrangement. Suppose $C$ is a five-dimensional contact manifold with contact distribution $H$. Suppose $V$ is a rank two integrable subbundle of $H$. Then we may find local coördinates $(x,y,z,a,b)$ on $C$ so that - $H$ is defined by the $1$-form$\lambda\equiv dz+a\,dx+b\,dy$, - $V$ is defined by $\lambda$ and the two $1$-forms $dx$ and $dy$. Thus, it is as if $C$ were defined by $M$ with local coördinates $(x,y,z)$. The following argument pertains locally. Choose $1$-forms $\lambda,\mu,\nu$ so that $$H=\lambda^\perp\quad\mbox{and}\quad V=(\lambda,\mu,\nu)^\perp.$$ Integrability of $V$ ensures, by Frobenius, that we can find coördinates $(x,y,z,u,v)$ so that $$\lambda,\mu,\nu\in{\mathrm{span}}\{dx,dy,dz\}$$ and, since $\lambda\not=0$, we may rescale it and subtract appropriate multiples thereof from $\mu$ and $\nu$ to suppose, without loss of generality, that $$\label{lambda_mu_nu} \lambda=dz+a\,dx+b\,dy\qquad \mu=p\,dx+q\,dy\qquad\nu=r\,dx+s\,dy,$$ for suitable smooth functions $(a,b,p,q,r,s)$. Now, since $H$ is contact, $$0\not=\lambda\wedge d\lambda\wedge d\lambda =2\,dx\wedge dy\wedge dz\wedge da\wedge db$$ so $(x,y,z,a,b)$ may be used as local coördinates instead. Finally, $$0\not=\lambda\wedge\mu\wedge\nu=(ps-qr)\,dx\wedge dy\wedge dz$$ so we may replace $\{\mu,\nu\}$ by $\{dx,dy\}$ without changing their span. The almost contact Legendrean geometry on $C$ {#attack} ============================================= Firstly, we revisit the splitting (\[splitting\]), now using torsion-free connections instead of choosing coördinates. As is well-known [@L], a connection on $T^*M$ may be viewed as a splitting (\[splitting\]) of $T(T^*M)$, into [*horizontal*]{} and vertical subbundles. If we change connections, say $$\label{connection_change} \widehat\nabla_aX^c=\nabla_aX^c+\Gamma_{ab}{}^cX^b,$$ (using Penrose’s abstract index notation [@OT]) then the splitting changes according to $$\label{change_of_splitting} \widehat{\left[\!\!\begin{array}{c}X^b\\ \omega_b\end{array}\!\!\right]} =\left[\!\!\begin{array}{c}X^b\\ \omega_b+X^a\Gamma_{ab}{}^cp_c\end{array}\!\!\right].$$ Now, if we insist on using [*torsion-free*]{} connections, as we may, then the [*skew*]{} form (\[canonical\_symplectic\_form\]) is manifestly invariant because $\Gamma_{ab}{}^c$ is [*symmetric*]{}. In order to navigate in $M$, we now suppose that this manifold is endowed with a [*projective differential geometric*]{} structure. A detailed discussion, specifically in $3$ dimensions, may be found in [@DE]. We shall therefore be brief in recalling the salient features. Although a projective structure may be viewed as a type of path geometry (eminently suitable for flying in $M$) an operational viewpoint on projective structures is as an equivalence class of torsion-free connections, where the notion of equivalence is that $$\label{projective_change} \widehat\nabla_a\phi_b=\nabla_a\phi_b-\Upsilon_a\phi_b-\Upsilon_b\phi_a$$ for an arbitrary $1$-form $\Upsilon_a$. In (\[connection\_change\]) it means that $$\Gamma_{ab}{}^c=\Upsilon_a\delta_b{}^c+\Upsilon_b\delta_a{}^c$$ where $\delta_b{}^c$ is canonical pairing between vectors and covectors. Hence, with (\[projective\_change\]) in place, the formula (\[change\_of\_splitting\]) for the change in splitting becomes $$\label{projective_change_of_splitting} \widehat{\left[\!\!\begin{array}{c}X^b\\ \omega_b\end{array}\!\!\right]} =\left[\!\!\begin{array}{c}X^b\\ \omega_b+X^a\Upsilon_ap_b+X^cp_c\Upsilon_b\end{array}\!\!\right].$$ But with a chosen connection and hence a chosen splitting in place, $$P\enskip\mbox{is the subspace of}\enskip\pi^*TM\enskip \mbox{given by}\enskip\{X^b\mid X^bp_b=0\}$$ and, in any case $V$ is the quotient of $\pi^*T^*M$ given by $$\{\omega_b\}/\sim\enskip\mbox{where}\enskip \omega_a\sim\omega_b+tp_a\enskip\forall\,t\in{\mathbb{R}}.$$ [From]{} (\[projective\_change\_of\_splitting\]) it follows at once that the splitting $H=P\oplus V$ is projectively invariant. In summary, we have proved the following. \[projective\_theorem\] A projective structure on $M$ gives rise to extra structure on its configuration space $C$. Specifically, the contact distribution $H$ canonically splits as $$\label{H_splits}H=P\oplus V.$$ Both $P$ and $V$ are null with respect to the Levi form ${\mathcal{L}}:\Wedge^2H\to L$, which otherwise restricts to a non-degenerate pairing $P\otimes V\to L$. In general, if $C$ is a manifold with contact distribution $H\subset TC$, then a splitting $H=P\oplus V$ into null subspaces for the Levi form is a type of parabolic geometry [@CS] called [*almost contact Legendrean*]{}. Projective differential geometry is another type of parabolic geometry and the construction of this section may be viewed in Dynkin diagram notation as ${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle $}} \put(20,8){\makebox(0,0){$\scriptstyle $}} \put(36,8){\makebox(0,0){$\scriptstyle $}} \end{picture}}\xrightarrow{\,\pi\,}{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle $}} \put(20,8){\makebox(0,0){$\scriptstyle $}} \put(36,8){\makebox(0,0){$\scriptstyle $}} \end{picture}}$. Furthermore, $$\label{dynkin}TC=L+ \begin{array}{c}P\\[-4pt] \oplus\\[-1pt] V\end{array}= \;{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}+ \begin{array}{c}{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}\\[-2pt] \oplus\\[-1pt] {\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}\end{array}$$ and the harmonic curvature splits into $3$ pieces $$H\^2(\_[-1]{},(4,)) = [l]{} (40,11) (4,1.5)[(0,0)[$\times$]{}]{} (20,1.5)[(0,0)[$\bullet$]{}]{} (36,1.5)[(0,0)[$\times$]{}]{} (4,1.5)[(1,0)[32]{}]{} (4,8)[(0,0)[$\scriptstyle -4$]{}]{} (20,8)[(0,0)[$\scriptstyle 1$]{}]{} (36,8)[(0,0)[$\scriptstyle 2$]{}]{} { ---------------------- obstruction to integrability of $P$ ---------------------- \ \ (40,11) (4,1.5)[(0,0)[$\times$]{}]{} (20,1.5)[(0,0)[$\bullet$]{}]{} (36,1.5)[(0,0)[$\times$]{}]{} (4,1.5)[(1,0)[32]{}]{} (4,8)[(0,0)[$\scriptstyle -3$]{}]{} (20,8)[(0,0)[$\scriptstyle 4$]{}]{} (36,8)[(0,0)[$\scriptstyle -3$]{}]{} \ \ (40,11) (4,1.5)[(0,0)[$\times$]{}]{} (20,1.5)[(0,0)[$\bullet$]{}]{} (36,1.5)[(0,0)[$\times$]{}]{} (4,1.5)[(1,0)[32]{}]{} (4,8)[(0,0)[$\scriptstyle 2$]{}]{} (20,8)[(0,0)[$\scriptstyle 1$]{}]{} (36,8)[(0,0)[$\scriptstyle -4$]{}]{} { ----------------------- obstruction to integrability of $V$. ----------------------- $$ Meanwhile, as detailed in [@DE], the harmonic curvature of $3$-dimensional projective geometry (usually known as the projective [*Weyl curvature*]{}) lies in ${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -4$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 2$}} \end{picture}}$. One can easily check that, in case $C$ is constructed from such a $3$-dimensional projective $M$, as above, then the harmonic curvature of $C$ lies only in ${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -4$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 2$}} \end{picture}}$ and that it is the pull-back of the Weyl curvature. In particular, the contact Legendrean structure on $C$ is flat if and only if the projective structure on $M$ is flat. Aerobatics may now be restricted by requiring, not only that the trajectory in $C$ be everywhere tangent to $H$, but also that the tangent vector be null with respect to the neutral signature conformal metric on $H$ given by the non-degenerate pairing $P\otimes V\to L$. Some special man[œ]{}uvres are permitted. Firstly, there is the option of remaining stationary in $M$ whilst changing the saucer orientation arbitrarily. In other words, since the fibres of $\pi:C\to M$ are null, it is permitted to move along them as one wishes. The second option is to move along a projective geodesic in $M$, with any initial orientation, lifting this curve to $C$ in accordance with the projectively invariant splitting (\[H\_splits\]). It is a common experience in usual aerobatics, that one carries along one’s own frame of reference! Indeed, any curve starting at $x\in M$ with an initial choice of orientation in $\pi^{-1}(x)$ can be uniquely lifted into $C$ in accordance with (\[H\_splits\]). This may be viewed as the difference between ‘gliding’ and ‘powered flight.’ In any case, null man[œ]{}uvring now has the geometric interpretation that, when moving in $M$, ‘rolls’ are restricted to be about one’s axis of flight (the ‘slow roll’ in usual aerobatics). Using only the two special man[œ]{}uvres of stationary rolling and gliding, as above, it is already clear that a pilot may park her craft in an arbitrary location and orientation. A CR structure on $C$ {#landing} ===================== In the previous section we saw that a projective structure on $M$ is exactly what is needed to define what might be called ‘attack mode,’ in which a saucer is permitted only to make rolls about its axis of flight. In coming in to land, however, this type of man[œ]{}uvre is unsuitable, even dangerous! More suitable for landing is the motion often observed in falling leaves, whereby rolls are constrained to be about axes orthogonal to the direction of flight. To make sense to this ‘landing mode,’ one clearly needs a notion of orthogonality in the disc of the saucer. It is natural to suppose that this notion is induced from $M$ itself. In other words, we shall suppose that $M$ is endowed with a [*conformal metric*]{}. If two Riemannian metrics $g_{ab}$ and $\widehat{g}_{ab}$ are conformally related, it is convenient to write $\widehat{g}_{ab}=\Omega^2g_{ab}$ for a smooth positive function $\Omega$. We shall suppose that $M$ is oriented and write $\epsilon_{abc}$ for the volume form associated to the metric $g_{ab}$. A conformal change of metric $\widehat{g}_{ab}=\Omega^2g_{ab}$ induces a change of volume form $\widehat{\epsilon}_{abc}=\Omega^3\epsilon_{abc}$ (we say that $\epsilon_{abc}$ has [*conformal weight*]{} $3$) and the corresponding Levi-Civita connections are related according to $$\label{conformal_change_of_connection} \widehat\nabla_a\phi_b=\nabla_a\phi_b-\Upsilon_a\phi_b-\Upsilon_b\phi_a +\Upsilon^c\phi_cg_{ab},$$ where $\Upsilon_a=\nabla_a\log\Omega$. We may choose a metric in the conformal class and use its Levi-Civita connection to write $$T(T^*M)= \begin{array}c\nu^*TM\\[-4pt] \oplus\\[-2pt] \nu^*T^*M\end{array}\ni \left[\!\!\begin{array}{c}X^a\\ \omega_a\end{array}\!\!\right].$$ According to (\[change\_of\_splitting\]) and  (\[conformal\_change\_of\_connection\]), if $\widehat{g}_{ab}=\Omega^2g_{ab}$, then $$\widehat{\left[\!\!\begin{array}{c}X^b\\ \omega_b\end{array}\!\!\right]} =\left[\!\!\begin{array}{c}X^b\\ \omega_b+X^a\Upsilon_ap_b+X^cp_c\Upsilon_b-X_b\Upsilon^cp_c \end{array}\!\!\right].$$ On $H$, since $\omega_b$ is only defined modulo $p_b$, and since $X^cp_c=0$, we can drop two of these terms to obtain $$\label{conformal_change_of_splitting} \widehat{\left[\!\!\begin{array}{c}X^b\\ \omega_b\end{array}\!\!\right]} =\left[\!\!\begin{array}{c}X^b\\ \omega_b-X_b\Upsilon^cp_c \end{array}\!\!\right]\enskip\mbox{on}\enskip H=\begin{array}{c}P\\[-4pt] \oplus\\[-2pt] V\end{array}\!\!\! \twoheadleftarrow\mbox{in the presence of~$g_{ab}$}.$$ Instead of $\omega_b$ up to multiples of $p_b$, we may use the conformal metric to suppose that $\omega^ap_a=0$ (i.e., normalise by $\omega_b\mapsto\omega_b-(\omega^cp_c/p^ap_a)p_b$). The change in splitting respects this normalisation (since $X^bp_b=0$). So now we have, for a chosen metric in the conformal class, $$H=\left\{\left[\!\!\begin{array}{c}X^b\\ \omega^b\end{array}\!\!\right] \mbox{s.t.\ }X^bp_b=0\enskip\mbox{and}\enskip \omega^bp_b=0\right\},$$ where $\omega^b$ has conformal weight $-2$ and, if $\widehat{g}_{ab}=\Omega^2g_{ab}$, then (\[conformal\_change\_of\_splitting\]) applies. We define $J:H\to H$ by $$\label{J} \left[\!\!\begin{array}{c}X^b\\ \omega^b\end{array}\!\!\right]\longmapsto \frac{1}{\sqrt{p^dp_d}}\left[\!\!\begin{array}{c}\epsilon^{abc}X_ap_c\\ \epsilon^{abc}\omega_ap_c\end{array}\!\!\right].$$ It respects the change (\[conformal\_change\_of\_splitting\]) and is hence well-defined. Since $$\epsilon^{abc}\epsilon_{ade} =\delta_d{}^b\delta_e{}^c-\delta_e{}^b\delta_d{}^c$$ it follows that $J^2=-{\mathrm{Id}}$ and we have defined an [*almost CR structure*]{}. In fact, we may check that this almost CR structure is integrable as follows. Since $P$ and $V$ are both two-dimensional, we need only check that, for the Nijenhuis tensor $N(\underbar\enskip,\underbar\enskip)$, $$N\left(\left[\!\!\begin{array}{c}X^a\\ 0\end{array}\!\!\right], \left[\!\!\begin{array}{c}0\\ \omega_b\end{array}\!\!\right]\right)=0.$$ This requirement expands to the vanishing of $$\label{what_we_would_like_to_vanish}\left[\!\!\begin{array}{c} -\omega_b\partial^bX^a\!-\!J(\omega_b\partial^b(JX^a)) \!-\!J((J\omega_b)\partial^bX^a)\!+\!(J\omega_b)\partial^b(JX^a)\\ X^a\nabla_a\omega_b\!+\!J(X^a\nabla_a(J\omega_b)) \!+\!J((JX^a)\nabla_a\omega_b)\!-\!(JX^a)\nabla_a(J\omega_b) \end{array}\!\!\right]$$ where $\partial^a=\partial/\partial p_a$ and may be verified as follows. Firstly, $$\partial^b(JX^a) =\partial^b\Big(\frac1{\sqrt{p^ep_e}}\epsilon^{cad}X_cp_d\Big),$$ which may be expanded by the Leibniz rule as $$\frac1{\sqrt{p^ep_e}}\epsilon^{cad}(\partial^bX_c)p_d -\frac{p^b}{(p^ep_e)^{3/2}}\epsilon^{cad}X_cp_d +\frac1{\sqrt{p^ep_e}}\epsilon^{cab}X_c.$$ Therefore $$\omega_b\partial^b(JX^a) =J(\omega_b\partial^bX^a) +\frac1{\sqrt{p^ep_e}}\epsilon^{abc}\omega_bX_c$$ so $$J(\omega_b\partial^b(JX^a))=-\omega_b\partial^bX^a +\frac1{p^ep_e}(g^{ab}g^{cd}-g^{ac}g^{bd})\omega_bX_cp_d,$$ which, bearing in mind that $X^dp_d=0$ and $\omega^dp_d=0$, reduces to $$J(\omega_b\partial^b(JX^a))=-\omega_b\partial^bX^a.$$ Similarly, $$(J\omega_b)\partial^b(JX^a) =J((J\omega_b)\partial^bX^a)$$ and all terms in the first line of (\[what\_we\_would\_like\_to\_vanish\]) cancel. For the second line, it is evident that $$X^a\nabla_a(J\omega_b)=J(X^a\nabla_a\omega_b) \quad\mbox{and}\quad (JX^a)\nabla_a(J\omega_b)=J((JX^a)\nabla_a\omega_b)$$ and, again, all terms cancel. The ‘landing mode,’ informally described at the beginning of this section is now formally defined by the restriction $X^bJ\omega_b=0$, noting from (\[conformal\_change\_of\_splitting\]) and (\[J\]) that this constraint is invariantly defined. Such a man[œ]{}uvre is very much at odds with the ‘attack mode’ of the previous section. A conformal structure on $M$ does not allow slow rolls to be defined: one sees from (\[conformal\_change\_of\_splitting\]) that the restriction $X^b\omega_b=0$ is always ill-defined unless one restricts to constant rescalings of the metric. In both modes, however, flying is restricted by requiring that the allowed curves in $C$ are not only tangent to $H$, but also that they be [*null*]{} for an appropriately defined neutral signature metric on $H$ (with values in the line bundle $L$). These metrics are $$\label{modes} \left\|\left[\!\!\begin{array}{c}X\\ \omega\end{array}\!\!\right] \right\|^2\!=\!X{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\omega\;\mbox{(attacking)}\qquad \left\|\left[\!\!\begin{array}{c}X\\ \omega\end{array}\!\!\right] \right\|^2\!=\!X{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}J\omega\;\mbox{(landing)}.$$ As tensors on $H$, we have the usual compatibility $J_\alpha{}^\beta=\Omega_{\alpha\gamma}g^{\beta\gamma}$ in which any two of the Levi form $\Omega_{\alpha\beta}$, the inverse metric $g^{\alpha\beta}$, and the endomorphism $J_\alpha{}^\beta$ determine the third. In landing mode, we have $J^2=-{\mathrm{Id}}$. In attacking mode, the endomorphism $J:H\to H$ instead satisfies $J^2={\mathrm{Id}}$, being the identity on $V$ and minus the identity on $P$ (for any given metric in the conformal class on $M$). For either of these geometries, stationary rolling is allowed since $V$ is null in either case. A metric on $M$ induces a splitting $$0\to V\longrightarrow H\longrightarrow \raisebox{-10pt}{\begin{picture}(0,0) \put(-23,2){\LARGE$\smile$} \put(-23,11){\vector(-1,1){0}} \end{picture}}P\to 0$$ and, as in Section \[attack\], the ‘gliding’ man[œ]{}uvre is now available. In other words, we may use a horizontal lift to arrive at $\omega=0$ in either of the neutral signature metrics (\[modes\]). Therefore, parking in an arbitrary location and orientation is easily achievable in the CR case. The only difference is that the splitting, and hence the particular man[œ]{}uvring to be used, depends on choosing a metric in the conformal class. The Dynkin diagram notation for this construction is $${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle $}} \put(20,8){\makebox(0,0){$\scriptstyle $}} \put(36,8){\makebox(0,0){$\scriptstyle $}} \end{picture}}\xrightarrow{\,\pi\,}{\begin{picture}(28,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(24,1.5){\makebox(0,0){$\bullet$}} \put(5,2.5){\line(1,0){20}} \put(5,0.5){\line(1,0){20}} \put(14,1.5){\makebox(0,0){$\rangle$}} \put(4,8){\makebox(0,0){$\scriptstyle $}} \put(24,8){\makebox(0,0){$\scriptstyle $}} \end{picture}}.$$ The harmonic curvature in three-dimensional conformal geometry is the Cotton tensor in ${\begin{picture}(28,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(24,1.5){\makebox(0,0){$\bullet$}} \put(5,2.5){\line(1,0){20}} \put(5,0.5){\line(1,0){20}} \put(14,1.5){\makebox(0,0){$\rangle$}} \put(4,8){\makebox(0,0){$\scriptstyle -5$}} \put(24,8){\makebox(0,0){$\scriptstyle 4$}} \end{picture}}$, which pulls back to ${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -3$}} \put(20,8){\makebox(0,0){$\scriptstyle 4$}} \put(36,8){\makebox(0,0){$\scriptstyle -3$}} \end{picture}}$. For more details concerning this construction, and especially a characterisation of the $5$-dimensional CR manifolds that arise in this way, see [@LeB1] (and also [@LeB2] for some very interesting consequences in the real-analytic setting). The flat model of this construction is when $M=S^3\hookrightarrow S^4$ is the standard inclusion of round spheres and $$\begin{array}{ccccl} C&\hookrightarrow&{\mathbb{CP}}_3\\ \pi\!\!\downarrow&&\tau\!\!\downarrow\\ S^3&\hookrightarrow&S^4\end{array}$$ where $\tau:{\mathbb{CP}}_3\to S^4$ is the twistor fibration and $C\hookrightarrow{\mathbb{CP}}_3$ is the Levi indefinite hyperquadric. Finally, we remark that if we would like to have at our disposal both the ‘attack mode’ of Section \[attack\] and the ‘landing mode’ of the current section, then we require compatible projective and conformal structures on $M$. If $\nabla_a$ represents a projective structure and $g_{ab}$ is a metric, then one can check that the $1$-form $$\omega_a\equiv 4g^{bc}\nabla_bg_{ac}-\epsilon^{bcd}\nabla_a\epsilon_{bcd},$$ where $\epsilon_{abc}$ is the volume form of $g_{ab}$, is projectively invariant. To find a metric in the conformal class of $g_{ab}$ whose Levi-Civita connection is in the given projective class, it is firstly necessary that $\omega_a$ be exact. In this case, a further necessary and sufficient condition is that $$40\nabla_ag_{bc}+2\omega_ag_{bc}-3\omega_bg_{ac} -3\omega_cg_{ab}=0,$$ where $\nabla_a$ has been chosen from the projective class so that $\nabla_a\epsilon_{bcd}=0$ (cf. [@MT]). In this case, we are obliged to have a Riemannian metric on $M$ defined up to homothety, i.e. only constant rescalings are allowed. More severely, as discussed in [@N], the generic projective structure does not arise from a metric at all. In summary, for a Riemannian metric on $M$ there are two possible flying modes, namely the ‘attacking mode’ of §\[attack\], which sees only the induced projective structure on $M$, and the ‘landing mode’ of §\[landing\], which sees only the induced conformal structure on $M$. In this case, on the configuration space $C\to M$, there are [*two different*]{} neutral signature conformal metrics on the contact distribution $H\subset TC$. The vertical bundle $V\hookrightarrow H$ is null for either of these conformal metrics on $H$, as is the horizontal bundle $P\hookrightarrow H$ (defined only by the projective structure on $M$). So, when $M$ is Riemannian, a flying saucer may be fitted with a switch that enables its pilot to toggle between attacking and landing. A $G_2$ contact structure on $C$ {#G2} ================================ Recall that the contact distribution $H\subset TC$ on a five-dimensional contact manifold $C$ is endowed with its Levi form $\Wedge^2H\to L\equiv TC/H$, a non-degenerate symplectic form defined up to scale, thereby reducing the structure group of frames for $H$ from ${\mathrm{GL}}(4,{\mathbb{R}})$ to the conformal symplectic group ${\mathrm{CSp}}(4,{\mathbb{R}})$. A [*$G_2$ contact structure*]{} on $C$ is a further reduction of structure group to ${\mathrm{GL}}(2,{\mathbb{R}})\subset{\mathrm{CSp}}(4,{\mathbb{R}})$, realised by the representation of ${\mathrm{GL}}(2,{\mathbb{R}})$ on the third symmetric power $\bigodot^3\!{\mathbb{R}}^2$ of the standard representation of ${\mathrm{GL}}(2,{\mathbb{R}})$ on ${\mathbb{R}}^2$. As $\bigodot^3\!{\mathbb{R}}^2$ is $4$-dimensional, we have ${\mathrm{GL}}(2,{\mathbb{R}})\hookrightarrow{\mathrm{GL}}(4,{\mathbb{R}})$ and, since $$\label{Wedge2H}\textstyle\Wedge^2\bigodot^3\!{\mathbb{R}}^2 =(\bigodot^4\!{\mathbb{R}}^2\otimes\Wedge^2{\mathbb{R}}^2)\oplus (\Wedge^2{\mathbb{R}}^2\otimes\Wedge^2{\mathbb{R}}^2 \otimes\Wedge^2{\mathbb{R}}^2),$$ the second summand of which is $1$-dimensional, this homomorphism maps to ${\mathrm{CSp}}(4,{\mathbb{R}})\subset{\mathrm{GL}}(4,{\mathbb{R}})$. (See [@B] for a discussion of similar reductions for the frame bundle of a four-dimensional manifold.) Equivalently, a $G_2$ structure on a five-dimensional contact manifold $C$ is a rank two vector bundle $S\to C$ together with an identification of vector bundles, compatible with the Levi form, $$\label{H_is_sym_3}\textstyle\bigodot^3\!S=H,$$ where $H$ is the contact distribution and $\bigodot^3\!S$ is the third symmetric power of $S$. It is named for a ‘flat model’ in parabolic geometry, namely the homogeneous space for the split real form of the Lie group $G_2$. The tangent bundle of this homogeneous space is an extension $$\label{extension}T(\raisebox{1.5pt}{${\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle $}} \put(24,8){\makebox(0,0){$\scriptstyle $}} \end{picture}}$})= {\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(24,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}\enskip+\enskip{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}},$$ and $H$ has the required form for $S={\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(24,8){\makebox(0,0){$\scriptstyle \enskip-1/3$}} \end{picture}}$.For this flat model, the local infinitesimal symmetries are isomorphic to the split real form of the exceptional Lie algebra $G_2$. Yet a third interpretation of a $G_2$ contact structure is as a field of twisted cubic cones inside $H$ (akin to the interpretation of a Lorentzian conformal structure as a field of quadratic cones, i.e. the null vectors). Specifically, writing $H$ in the form (\[H\_is\_sym\_3\]) defines a cone $$\textstyle s\odot s\odot s\in\bigodot^3\!S=H$$ of simple vectors in each fibre of $H$. It is easy to check that this cone determines the structure. Geometrically, it can be regarded as a smoothly varying family of twisted cubics in the bundle ${\mathbb{P}}(H)$ of three-dimensional projective spaces (again, see [@B] for the corresponding geometry in four dimensions). In any case, our aim in this section is to equip the configuration space of a flying saucer with a $G_2$ contact structure. In Sections \[attack\] and \[landing\], the structures on the configuration space $C$ were nicely determined by suitable differential geometric structures on $M$, specifically [*projective*]{} in Section \[attack\] and [*conformal*]{} in Section \[landing\]. Now we shall firstly suppose that $M$ has a projective structure and also, for the moment, a fixed volume form $\epsilon_{bcd}$. In this case, we can specify a unique connection from the projective class by insisting that $\nabla_a\epsilon_{bcd}=0$. In any case, as in Section \[attack\], we have a well-defined splitting $H=P\oplus V$. To complete the geometric structure on $M$, we suppose that we are given two linearly independent $1$-forms, say $\phi$ and $\psi$. These forms are sufficient to define a $G_2$ contact structure on (an open subset of) $C$ as follows. For $P\hookrightarrow C$, define a frame $e^1,e^2\in\Gamma(P)$ by requiring that $$\begin{array}{cc} e^1{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\pi^*\phi=1&\enskip e^1{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\pi^*\psi=0\\[4pt] e^2{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\pi^*\phi=0&\enskip e^2{\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}\pi^*\psi=1. \end{array}$$ This is legitimate wherever $\pi^*\phi$ and $\pi^*\psi$ are linearly independent when restricted to $H\subset TC$ (and defines an open subset of $C$). Next, recall the canonical short exact sequence $$0\to P\to\pi^*TM\to L\to 0$$ on $C$ and hence a canonical identification $$L^*=\pi^*\Wedge_M^3\otimes\Wedge^2P.$$ Thus, we may use the nowhere vanishing sections $\pi^*\epsilon_{bcd}\in\Gamma(\pi^*\Wedge_M^3)$ and $e^1\wedge e^2\in\Gamma(\Wedge^2P)$ to trivialise $L^*$. With this trivialisation, the Levi form $P\otimes V\to L$ identifies $V=P^*$ and we take $e_1,e_2\in\Gamma(V)$ to be the dual frame to $e^1,e^2\in\Gamma(P)$. Finally, we define a twisted cubic ${\mathbb{RP}}_1\hookrightarrow{\mathbb{P}}(P\oplus V)={\mathbb{P}}(H)$ by $$\label{twisted_cubic}{\mathbb{RP}}_1\ni[s,t]\longmapsto [s^3\,e^1+s^2t\,e^2+t^3\,e_1-3st^2\,e_2] \in{\mathbb{P}}(H).$$ The seemingly peculiar choice of constants here is to ensure that this cubic induce the existing Levi form on $H$, specifically that $$\begin{array}{l}(s\tilde{t}-t\tilde{s})^3\\ =(s^3\,e^1+s^2t\,e^2){\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}(\tilde{t^3}\,e_1-3\tilde{s}\tilde{t}^2\,e_2) -(\tilde{s}^3\,e^1+\tilde{s}^2\tilde{t}\,e^2){\,\rule{5pt}{.3pt}\rule{.3pt}{7pt}\;}(t^3\,e_1-3st^2\,e_2) \end{array}$$ in accordance with (\[canonical\_symplectic\_form\]). There is some ‘gauge freedom’ associated with this arrangement, i.e. changes in the data $(\epsilon,\phi,\psi)$ on $M$ that do not affect the associated $G_2$ projective structure on $C$. Specifically, if we replace $(\epsilon_{bcd},\phi_b,\psi_b)$ by $$\label{freedom}\hat\epsilon_{bcd}=\Omega^4\epsilon_{bcd}\qquad \hat\phi_b=h^3\phi_b\qquad\hat\psi_b=\Omega\psi_b,$$ for arbitrary smooth nowhere vanishing functions $\Omega$ and $h$ on $M$, then $$\hat e^1=h^{-3}e^1\qquad\hat e^2=\Omega^{-1}e^2\qquad \hat e_1=h^6\Omega^{-3}e_1\qquad\hat e_2=h^3\Omega^{-2}e_2,$$ giving rise to the twisted cubic $$[\hat s,\hat t]\longmapsto [h^{-3}\hat s^3\,e^1+\Omega^{-1}\hat s^2\hat t\,e^2 +h^6\Omega^{-3}\hat t^3\,e_1-3h^3\Omega^{-2}\hat s\hat t^2\,e_2],$$ which is simply a reparameterisation, namely $[\hat s,\hat t]=[hs,h^{-2}\Omega t]$, of the original cubic (\[twisted\_cubic\]). Using the language of [*projective weights*]{}, e.g. as in [@DE], we may regard $\epsilon_{bcd}$ as the tautologically defined section of $\Wedge^3(4)$ and the true data as a $1$-form $\phi\in\Gamma(\Wedge^1)$, defined only up to scale, and $\psi\in\Gamma(\Wedge^1(1))$, a $1$-form of projective weight $1$. The construction above is almost captured geometrically as follows. Firstly, the projective structure on $M$ splits the contact distribution as $H=P\oplus V$, which, in the bundle ${\mathbb{P}}(H)\to M$ of $3$-dimensional projective spaces, may be viewed as a family of skew lines $$\begin{picture}(100,100) \put(110,90){\makebox(0,0){${\mathbb{P}}(H_x)$}} \put(-29,90){\makebox(0,0){${\mathbb{P}}(V_x)$}} \put(-29,10){\makebox(0,0){${\mathbb{P}}(P_x)$}} \put(-3,98.5){\line(-2,-1){10}} \put(-3,1.5){\line(-2,1){10}} \thicklines \put(0,100){\line(1,-1){100}} \put(0,0){\line(4,1){72}} \put(88,22){\line(4,1){75}} \put(0,100){\makebox(0,0){\Large$\circ$}} \put(0,0){\makebox(0,0){\Large$\circ$}} \end{picture}$$ and the family of twisted cubics (\[twisted\_cubic\]) looks like this: $$\begin{picture}(100,110)(0,-10)\thicklines \put(0,100){\line(1,-1){100}} \put(0,0){\line(4,1){72}} \put(88,22){\line(4,1){75}} \put(20,80){\makebox(0,0){$\bullet$}} \put(24,6){\makebox(0,0){$\bullet$}} \qbezier(20,80)(80,20)(24,6) \qbezier(20,80)(0,100)(-20,50) \qbezier(-20,50)(-40,0)(0,20) \qbezier(0,20)(40,40)(0,40) \qbezier(0,40)(-20,40)(-20,40) \qbezier(-28,40)(-50,40)(-55,20) \qbezier(-55,20)(-60,0)(-20,0) \qbezier(-20,0)(0,0)(24,6) \put(32,87){\makebox(0,0){$[e_1]$}} \put(35,-3){\makebox(0,0){$[e^1]$}} \end{picture}$$ intersecting ${\mathbb{P}}(P)$ and ${\mathbb{P}}(V)$ tangentially at $[e^1]$ and $[e_1]$, respectively, as can be seen from (\[twisted\_cubic\]). Requiring that the twisted cubic be compatible with the Levi form, as we do, is insufficient to fix it. The gauge freedom explained above says that there is just one more scalar-valued piece of information required at each point $x\in M$ and this may be interpreted as our requiring $\psi\in\Gamma(\Wedge^1(1))$. As is detailed in [@EN], there are two examples of this construction for which $C$ turns out to be (locally) the flat model for the split real form of the Lie group $G_2$. For both, we start with the standard flat projective structure on ${\mathbb{R}}^3$ with usual coördinates $(x,y,z)$ and take $$\epsilon_{abc}=dx\wedge dy\wedge dz,$$ the standard volume form. We take $$\phi=dx\quad\mbox{and}\quad\psi=dy.$$ With slightly different coördinates, this example is due to Engel [@E]. In [@EN] we write down all its $14$-dimensional symmetries. We take $$\phi=x\,dy-y\,dx\quad\mbox{and}\quad\psi=y^{-1}dy.$$ The geometry on $M$ appears to be different from the first example but we shall see in Theorem \[G2\_flat\_thm\] below that the induced $G_2$ contact structure on $C$ is $G_2$-flat and therefore isomorphic. The geometry on $M$ {#geometry_on_M} =================== In this section we speculate on the geometry on $M$ that is needed to generate the $G_2$ contact structure on $C\to M$, as in §\[G2\]. We have already seen in (\[freedom\]) that $\psi$ should have projective weight $1$ whilst $\phi$ may be arbitrarily rescaled. In fact, we shall suppose that $\phi$ has projective weight $2$. There are several reasons for this, the most naïve of which is as follows. For a $1$-form of projective weight $w$, the formula (\[projective\_change\]) for projective change becomes $$\label{projective_change_with_weight}\widehat\nabla_a\phi_b =\nabla_a\phi_b+(w-1)\Upsilon_a\phi_b-\Upsilon_b\phi_a$$ and when $w=2$, we see that $\nabla_{(a}\phi_{b)}$ is invariant, where the round brackets mean to take the symmetric part. We may, therefore, assume that $\nabla_{(a}\phi_{b)}=0$ as a sort of [*compatibility*]{} between $\phi_b$ and the projective structure defined by $\nabla_a$. We shall come back to this shortly but an immediate and congenial consequence of imposing $\nabla_{(a}\phi_{b)}=0$ is that $\phi_b$ is then determined up to an overall constant: $$\nabla_{(a}(f\phi_{b)})=f\nabla_{(a}\phi_{b)}+(\nabla_{(a}f)\phi_{b)} \enskip\Rightarrow\enskip\nabla_af=0.$$ In summary, the data we are supposing on $M$ is as follows. $$\label{data}\begin{array}{ll} \bullet&\mbox{A projective structure, determined by~$\nabla_a$},\\ \bullet&\mbox{$\phi_b\in\Gamma(M,\Wedge^1(2))$, such that $\nabla_{(a}\phi_{b)}=0$},\\ \bullet&\mbox{$\psi_b\in\Gamma(M,\Wedge^1(1))$}, \end{array}$$ and we note that the two examples from §\[G2\] above satisfy $\nabla_{(a}\phi_{b)}=0$, as requested. There are various projective invariants that we may generate from this data. The [*concircularity*]{} operator $$\textstyle\theta^b\mapsto(\nabla_a\theta^b)_\circ \equiv\nabla_a\theta^b-\frac13\delta_a{}^b\nabla_c\theta^c$$ (where $\circ$ means to take the trace-free part) is projectively invariant if $\theta^a$ has projective weight $-1$. Meanwhile, the tautological form $\epsilon^{bcd}$ has projective weight $-4$ so $\epsilon^{bcd}\phi_c\psi_d$ has projective weight $-1$ and hence $(\nabla_a(\epsilon^{bcd}\phi_c\psi_d))_\circ$ is invariant. It may be rewritten as $$\nabla_a(\phi_{[b}\psi_{c]})-\nabla_{[a}(\phi_b\psi_{c]}).$$ It is easily verified that this expression vanishes for the two examples given at the end of the previous section. When $w=1$ the projective change (\[projective\_change\_with\_weight\]) reads $$\widehat\nabla_b\psi_c=\nabla_b\psi_c-\Upsilon_c\psi_b\qquad\mbox{whence} \qquad\psi_{[a}\nabla_{b]}\psi_c\enskip\mbox{is invariant}$$ and also vanishes for our two examples. \[G2\_flat\_thm\] For the data to define a flat $G_2$ contact structure on $C$, it is necessary and sufficient that $$\label{G2_flat_conditions} \nabla_a\mbox{ be projectively flat},\enskip \nabla_a(\phi_{[b}\psi_{c]})=\nabla_{[a}(\phi_b\psi_{c]}),\enskip \psi_{[a}\nabla_{b]}\psi_c=0.$$ The proof will be given shortly but, firstly, some discussion. This theorem leads us to the following. We take the standard flat projective structure on ${\mathbb{R}}^3$ with usual coördinates $(x,y,z)$ and $$\phi=x\,dy-y\,dx\quad\mbox{and}\quad\psi=zy^{-1}dy-dz.$$ It is easily verified that all conditions (\[G2\_flat\_conditions\]) hold and the corresponding $G_2$ contact structure on the configuration space $C$ is, therefore, flat. Although these examples seem naïvely to be distinct (and from the Riemannian viewpoint, this is true), they are, in fact, projectively equivalent. Specifically, if we set $$\hat x=-x/y\qquad\hat y=-1/y\qquad\hat z=-z/y$$ then $$d\hat x=y^{-2}(x\,dy-y\,dx)\quad\mbox{and}\quad d\hat y=y^{-2}dy$$ and, taking into account that $\phi$ should have projective weight $2$ and $\psi$ should have projective weight $1$, it follows that the first example, $(\phi=d\hat x, \psi=d\hat y)$ is converted into $(\phi=x\,dy-y\,dx,\psi=y^{-1}\,dy)$, which is the second example. Similarly, the projective change $$\hat x=-x/y\qquad\hat y=-z/y\qquad\hat z=1/y$$ converts $(\phi=d\hat x, \psi=d\hat y)$ into $(\phi=x\,dy-y\,dx,\psi=zy^{-1}\,dy-dz)$, which is the third example. To some extent, this projective equivalence of our three examples justifies our request that the $1$-form $\phi$ should have projective weight $2$ but there is another good reason for this, as follows. Recall, in (\[freedom\]), that $\phi$ may be arbitrarily rescaled without effecting the resulting $G_2$ contact geometry on $C$. Its kernel $D\subset TM$ is therefore canonically defined and we may write $$\label{SES} 0\to D\to TM\xrightarrow{\,\phi\,}\xi\to 0,$$ a short exact sequence tautologically defining a line bundle $\xi$ on $M$. Let us temporarily suppose that $D$ is a contact distribution (even though this is false in our three examples). In three dimensions, the Levi form then provides a canonical isomorphism $\Wedge^2D=\xi$ and, feeding this back into (\[SES\]), we may identify $\Wedge^3TM=\xi^2$. That $M$ is oriented allows us to identify $\xi$ as the bundle of densities of projective weight $2$ (whether or not $M$ has a projective structure). In summary, when $D$ is a contact distribution we are forced to regard $\phi\in\Gamma(M,\Wedge^0(2))$ and, even when $D$ is integrable, we may [*choose*]{} to do this. As already observed, in the presence of a projective structure $[\nabla_a]$, we may also insist that $\nabla_{(a}\phi_{b)}=0$. In case $D$ is a contact distribution, this is exactly the compatibility required between $D$ and a projective structure in order that the pair $([\nabla_a],D)$ define a [*contact projective structure*]{} in the sense of Harrison [@H] and/or Fox [@F]. It is a type of parabolic geometry, the flat model of which is ${\mathbb{RP}}_3$ under the action of ${\mathrm{Sp}}(4,{\mathbb{R}})$. It would be nice to construct a flat $G_2$ contact structure starting with this flat contact projective structure. Unfortunately, this seems to be impossible. Specifically, in standard coördinates $(x,y,z)$ on ${\mathbb{R}}^3$, we may arrange that $$\phi=x\,dy-y\,dx+dz.$$ On the other hand, the general solution of the concircularity equation $(\nabla_a\theta^b)_\circ=0$ is $$\theta =a\frac\partial{\partial x}+b\frac\partial{\partial y} +c\frac\partial{\partial z} +e\Big(x\frac\partial{\partial x}+y\frac\partial{\partial y} +z\frac\partial{\partial z}\Big),$$ in which case $$\theta^b\phi_b=c+bx-ay+ez.$$ Therefore, we cannot find $\psi_d\not=0$ such that $(\nabla_a(\epsilon^{bcd}\phi_c\psi_d))_\circ=0$, which is the second condition from (\[G2\_flat\_conditions\]). As we shall see in the proof of Theorem \[construct\] below, the quantity $\epsilon^{bcd}\phi_c\psi_d$ naturally arises in constructing the $G_2$ contact geometry on $C$ and, even if Theorem \[G2\_flat\_thm\] is too restrictive, one would expect the projective invariant $(\nabla_a(\epsilon^{bcd}\phi_c\psi_d))_\circ$ to be part of the harmonic curvature of the $G_2$ contact geometry on $C$ (which, in general, is a binary septic: in fact, a section of ${\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 7$}} \put(24,8){\makebox(0,0){$\scriptstyle -4$}} \end{picture}}$). Taking $\phi\in\Gamma(M,\Wedge^1(2))$ and $\psi\in\Gamma(M,\Wedge^1(1))$, allows us to write the twisted cubic cone (\[twisted\_cubic\]) inside the contact distribution $H\subset TC$ more invariantly than was done in §\[G2\]. We obtain the following. \[construct\] Suppose $M$ is a three-dimensional smooth manifold and write $C\xrightarrow{\,\pi\,}M$ for the configuration space of flying saucers in $M$. Given data in the form  on $M$, we may canonically construct a $G_2$ contact structure on a suitable open subset of $C$ so that the contact distribution $H\subset TM$ is written as $$\textstyle H=\bigodot^3\!S,\quad\mbox{where } S=\pi^*\Wedge^0(2/3)\oplus\pi^*\Wedge^0(-1/3).$$ Recall from (\[dynkin\]) that $$P={\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}\quad\mbox{and}\quad V={\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}.$$ Also, from (\[dynkin\]) we have $$0\to{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}\to\pi^*TM\to{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}\to 0,$$ a canonical short exact sequence on $C$ and, in particular, canonical surjections $$\pi^*\Wedge_M^1\to\big(\,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}\;\big)^*={\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -2$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}} \quad\mbox{and}\quad\pi^*\Wedge_M^1(w)\to\enskip{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle w-2\;{}$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}$$ for any projective weight $w$. Writing $\pi^!$ for the pullback $\pi^*$ followed by this surjection, firstly gives $$\Theta\equiv\pi^!\phi\wedge\pi^!\psi \in\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}\wedge{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}})=\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}})$$ and then, on the open set where $\Theta$ is non-vanishing, $$E_1\equiv\Theta\,\pi^!\phi\in\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}})\qquad E_2\equiv\Theta\,\pi^!\psi\in\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}})$$ and $$E^1\equiv-\Theta^{-1}\pi^!\psi\in\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}})\qquad E^2\equiv\Theta^{-1}\pi^!\phi\in\Gamma(C,{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}).$$ Bearing in mind that $$\begin{array}{ll} {\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}=V\otimes\Wedge_C^0(1)&\qquad {\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}=V\\[4pt] {\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -1$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}=P\otimes\Wedge_C^0(-2)&\qquad {\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}=P\otimes\Wedge_C^0(-1)\,, \end{array}$$ where $\Wedge_C^0(w)={\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\times$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle w$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}=\pi^*\Wedge_M^0(w)$, we conclude that $$S\equiv\pi^*\Wedge^0(2/3)\oplus\pi^*\Wedge^0(-1/3) \ni\hspace{-77pt}\begin{array}[t]{cl}(\sigma,\tau)\\ \begin{picture}(10,15) \put(5,5){\makebox(0,0){$\downarrow$}} \put(5,10.2){\makebox(0,0){${}-{}$}}\end{picture}\\ \sigma^3\,E^1+\sigma^2\tau\,E^2+\tau^3\,E_1-3\sigma\tau^2\,E_2& {}\hspace{-8pt}\in P\oplus V=H \end{array}$$ is well defined. It is an invariant formulation of (\[twisted\_cubic\]) whose range defines a twisted cubic cone in ${\mathbb{P}}(H)$ compatible with the Levi form and hence a $G_2$ contact structure on $C$. We remark that, although the projective weights $2/3$ and $-1/3$ in the identification of $S$ may look contrived, they yield $$\Wedge^2S\otimes\Wedge^2S\otimes\Wedge^2S =\pi^*\Wedge^0(1)$$ and therefore, in accordance with the vector bundle version of (\[Wedge2H\]), that $\pi^*\Wedge^0(w)={\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(24,8){\makebox(0,0){$\scriptstyle w$}} \end{picture}}$, as one might expect. [*Proof of Theorem \[G2\_flat\_thm\].*]{}Since $\nabla_a$ is projectively flat we may suppose, without loss of generality, that our manifold is ${\mathbb{R}}^3\hookrightarrow{\mathbb{RP}}_3$ with $\nabla_a$ the standard flat connection. The operator $\phi_b\mapsto\nabla_{(a}\phi_{b)}$ is the first BGG operator $${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}\xrightarrow{\,\nabla\,}{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -2$}} \put(20,8){\makebox(0,0){$\scriptstyle 2$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}\quad\mbox{on ${\mathbb{RP}}_3$}$$ with kernel the irreducible representation ${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\bullet$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 0$}} \end{picture}}=\Wedge^2{\mathbb{R}}^4$ of ${\mathrm{SL}}(4,{\mathbb{R}})$ (acting by projective transformations on ${\mathbb{RP}}_3$). There are two non-zero orbits for the action of ${\mathrm{SL}}(4,{\mathbb{R}})$ on $\Wedge^4{\mathbb{R}}^4$ depending on rank and the non-degenerate case is represented by $\phi=x\,dy-y\,dx+dz$, which we have already seen to be incompatible with the second condition of (\[G2\_flat\_conditions\]). It follows that, without loss of generality, we may suppose $\phi=dx$. The operator $\theta^b\mapsto(\nabla_a\theta^b)_\circ$ is also a first BGG operator $${\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}\xrightarrow{\,\nabla\,}{\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\times$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle -2$}} \put(20,8){\makebox(0,0){$\scriptstyle 1$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}$$ whose solution space is ${\mathbb{R}}^4={\begin{picture}(40,11) \put(4,1.5){\makebox(0,0){$\bullet$}} \put(20,1.5){\makebox(0,0){$\bullet$}} \put(36,1.5){\makebox(0,0){$\bullet$}} \put(4,1.5){\line(1,0){32}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(20,8){\makebox(0,0){$\scriptstyle 0$}} \put(36,8){\makebox(0,0){$\scriptstyle 1$}} \end{picture}}$ as an ${\mathrm{SL}}(4,{\mathbb{R}})$-module. The degenerate $2$-form corresponding to $\phi$ specifies a $2$-plane in ${\mathbb{R}}^4$ and so there are just two cases for $\theta^b$ depending on whether the corresponding vector in ${\mathbb{R}}^4$ lies in this plane or not. This is exactly whether $\theta^a\phi_a$ vanishes or not and, with $\theta^a$ being of the form $\epsilon^{abc}\phi_b\psi_c$, it must vanish. Therefore, without loss of generality $\theta^a=\partial/\partial z$. We have reached the following normal forms for $\phi$ and $\psi$: $$\phi=dx\quad\mbox{and}\quad\psi=\xi(x,y,z)\,dx+dy$$ where $\xi(x,y,z)$ is an arbitrary smooth function. It remains to consider the remaining condition from (\[G2\_flat\_conditions\]), namely that $\psi_{[a}\nabla_{b]}\psi_c=0$. It means that $\xi=\xi(x,y)$, a function of $(x,y)$ alone, and that $\xi\xi_y=\xi_x$. In the computation that follows, we shall find $\xi\xi_y-\xi_x$ as the only non-trivial component of the harmonic curvature for the associated $G_2$ contact structure and our proof will be complete. The harmonic curvature is computed in Theorem \[harmonic\_curvature\] below. To use it we must specify the $G_2$ contact structure on $C$ in terms of an adapted co-frame (\[fully\_adapted\]). Starting with $$\omega^1=\phi=dx\quad\mbox{and}\quad\omega^2=\psi=\xi(x,y)\,dx+dy$$ on $M={\mathbb{R}}^3$ we may take $$\textstyle\omega^0=dz-a\,dx-b\,dy\qquad\omega^3=-\frac13\,db\qquad \omega^4=da-\xi(x,y)\,db$$ in local coördinates $(x,y,z,a,b)$ on $C$, as in §\[contact\]. These satisfy (\[fully\_adapted\]) with $\chi\equiv 1$. We compute $$\begin{array}{ll}d\omega^1=0\quad& d\omega^2=d\xi\wedge dx=-\xi_y\,\omega^1\wedge\omega^2\\[4pt] d\omega^3=0&d\omega^4=-d\xi\wedge db =3(\xi_x-\xi\xi_y)\,\omega^1\wedge\omega^3+3\xi_y\,\omega^2\wedge\omega^3 \end{array}$$ and we see that the only non-zero coefficients in (\[we\_write\]) are $$a^2{}_{12}=-\xi_y\qquad a^4{}_{13}=3(\xi_x-\xi\xi_y)\qquad a^4{}_{23}=3\xi_y.$$ Substituting into (\[psi\]) gives $\psi_0=\psi_1=\psi_2=\psi_3=\psi_4 =\psi_5=\psi_7=0$ and $\psi_6=6(\xi\xi_y-\xi_x)$, as claimed. $\square$ The partial differential equation $\xi\xi_y=\xi_x$ has plenty of local solutions. Indeed, if $F(t)$ is an arbitrary smooth function and we define $\xi(x,y)$ implicitly by the equation $$F(\xi)=x\xi+y,$$ then $\xi\xi_y=\xi_x$. In particular, the foliation of ${\mathbb{R}}^3$ defined by $\psi=\xi\,dx+dy$ need not be the planar foliation exhibited in the three examples above. Therefore, we have found many projectively inequivalent examples of structures given by data of the form (\[data\]) that all give rise to the same (flat) $G_2$ contact structure on the associated configuration space. Appendix: harmonic curvature of a $G_2$ contact structure {#appendix-harmonic-curvature-of-a-g_2-contact-structure .unnumbered} ========================================================= As already mentioned (see  [@CS] for the general theory), the harmonic curvature of a $G_2$ contact structure is a section of the bundle ${\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 7$}} \put(24,8){\makebox(0,0){$\scriptstyle -4$}} \end{picture}}$. This binary septic may be obtained by the Cartan equivalence method. Specifically, a $G_2$ contact structure on a five-dimensional manifold $C$ may be specified in terms of an adapted co-frame as follows. Firstly, we choose a $1$-form $\omega^0$ whose kernel is the contact distribution $H\subset TC$. Non-degeneracy of the contact distribution says that $\omega^0\wedge d\omega^0\wedge d\omega^0\not=0$. The $G_2$ structure is then determined by completing $\omega^0$ to a co-frame $\omega^0,\omega^1,\omega^2,\omega^3,\omega^4$ so that $$\label{adapted} d\omega^0=\chi(\omega^1\wedge\omega^4-3\omega^2\wedge\omega^3) \enskip\bmod\omega^0$$ for some smooth function $\chi$. Specifically, if $X_0,X_1,X_2,X_3,X_4$ is the dual frame, then $H={\mathrm{span}}\{X_1,X_2,X_3,X_4\}$ and the twisted cubic (\[twisted\_cubic\]) may be given as $$\label{twisted_cubic_again} (s,t)\mapsto s^3X_1+s^2tX_2+st^2X_3+t^3X_4,$$ compatibility with the Levi form being a consequence of (\[adapted\]). Imposing the structure equations (\[adapted\]) leaves precisely the following freedom in choice of co-frame: $$\label{thisisG} \left[\!\!\begin{array}c\tilde\omega^0\\ \tilde\omega^1\\ \tilde\omega^2\\ \tilde\omega^3\\ \tilde\omega^4\end{array}\!\!\right] \!\!=\!\!\left[\!\!\addtolength{\arraycolsep}{-1pt}\begin{array}{ccccc} t_9&0&0&0&0\\ t_{10}&t_5{}^3&3t_5{}^2t_6&3t_5t_6{}^2&t_6{}^3\\ t_{11}&t_5{}^2t_7&t_5(t_5t_8+2t_6t_7)&t_6(2t_5t_8+t_6t_7)&t_6{}^2t_8\\ t_{12}&t_5t_7{}^2&t_7(2t_5t_8+t_6t_7)&t_8(t_5t_8+2t_6t_7)&t_6t_8{}^2\\ t_{13}&t_7{}^3&3t_7{}^2t_8&3t_7t_8{}^2&t_8{}^3 \end{array}\!\!\right]\!\! \left[\!\!\begin{array}c\omega^0\\ \omega^1\\ \omega^2\\ \omega^3\\ \omega^4\end{array}\!\!\right]$$ for arbitrary functions $t_5,t_6,t_7,t_8,t_9,t_{10},t_{11},t_{12},t_{13}$ on $C$ subject only to $t_9(t_5t_8-t_6t_7)\not=0$. (The functions $t_5,t_6,t_7,t_8$ correspond to $$\raisebox{4pt}{$(s,t)\mapsto(s,t)$} \left(\!\begin{array}{cc}t_5&t_7\\ t_6&t_8\end{array}\right)$$ as a change of parameterisation in (\[twisted\_cubic\_again\]) whilst $t_9,t_{10},t_{11},t_{12},t_{13}$ modify the co-frame with multiples of $\omega^0$.) To proceed with Cartan’s method of equivalence, we now pass to the bundle $\widetilde{C}\to C$ whose sections are frames adapted according to (\[adapted\]). It is a $G$-principal bundle where $G$ is the $9$-dimensional Lie subgroup of ${\mathrm{GL}}(5,{\mathbb{R}})$ given in (\[thisisG\]) and comes tautologically equipped with $1$-forms $\theta^0,\theta^1,\theta^2,\theta^3,\theta^4$ whose pull-backs along a section are $\omega^0,\omega^1,\omega^2,\omega^3,\omega^4$, the co-frame on $C$ corresponding to that section. Cartan’s aim is to extend this to an invariant co-frame on $\widetilde{C}$ by making various normalisations. For our purposes we need not take these normalisations too far. For calculation, we choose a co-frame on $C$ adapted according to (\[adapted\]) so that $\widetilde{C}$ is identified as $G\times C$ and the forms $\theta^0,\theta^1\,\theta^2,\theta^3,\theta^4$ are given as $$\left[\!\!\begin{array}c\theta^0\\ \theta^1\\ \theta^2\\ \theta^3\\ \theta^4\end{array}\!\!\right] \!\!=\!\!\left[\!\!\addtolength{\arraycolsep}{-1pt}\begin{array}{ccccc} t_9&0&0&0&0\\ t_{10}&t_5{}^3&3t_5{}^2t_6&3t_5t_6{}^2&t_6{}^3\\ t_{11}&t_5{}^2t_7&t_5(t_5t_8+2t_6t_7)&t_6(2t_5t_8+t_6t_7)&t_6{}^2t_8\\ t_{12}&t_5t_7{}^2&t_7(2t_5t_8+t_6t_7)&t_8(t_5t_8+2t_6t_7)&t_6t_8{}^2\\ t_{13}&t_7{}^3&3t_7{}^2t_8&3t_7t_8{}^2&t_8{}^3 \end{array}\!\!\right]\!\! \left[\!\!\begin{array}c\omega^0\\ \omega^1\\ \omega^2\\ \omega^3\\ \omega^4\end{array}\!\!\right].$$ Normalise the co-frame on $C$ so that $$\label{fully_adapted} d\omega^0=\chi(\omega^1\wedge\omega^4-3\omega^2\wedge\omega^3).$$ This is easily achieved by the freedom $$\tilde\omega^1=\omega^1+t_{10}\omega^0,\enskip \tilde\omega^2=\omega^1+t_{11}\omega^0,\enskip \tilde\omega^3=\omega^1+t_{12}\omega^0,\enskip \tilde\omega^4=\omega^1+t_{13}\omega^0$$ and determines the functions $t_{10},t_{11},t_{12},t_{13}$. Find $\theta^5$ such that $$d\theta^0 =-6\theta^0\wedge\theta^5+\theta^1\wedge\theta^4-3\theta^2\wedge\theta^3.$$ This may be achieved by setting $-t_9=\Delta^3/\chi$, where $\Delta=t_6t_7-t_5t_8$ and then $$\theta^5=-\frac16\frac{d\chi}{\chi}+\frac12\frac{d\Delta}{\Delta} +\frac{\chi}{\Delta^3}\Big(\frac{t_{13}\theta^1}6-\frac{t_{12}\theta^2}2 +\frac{t_{11}\theta^3}2-\frac{t_{10}\theta^4}6\Big)+s\theta^0$$ for some function $s$. Find $\theta^7,\theta^8,\theta^9$ such that $$E^1\equiv d\theta^1-(6\theta^0\wedge\theta^9-3\theta^1\wedge\theta^5 -3\theta^1\wedge\theta^8+3\theta^2\wedge\theta^7)$$ is of the form $c^1{}_{\mu\nu}\theta^\mu\wedge\theta^v$ for $\mu,\nu=0,1,2,3,4$. It follows that $$E^1\wedge\theta^0\wedge\theta^1\wedge\theta^2 =c^1{}_{34}\theta^0\wedge\theta^1\wedge\theta^2\wedge\theta^4\wedge\theta^5.$$ It turns out that $$c^1{}_{34}=\frac{t_5{}^7}{\Delta^5} \Big(\psi_0+\psi_1s+\psi_2s^2+\psi_3s^3+\psi_4s^4+\psi_5s^5+\psi_6s^6 +\psi_7s^7\Big),$$ where $\psi_0,\psi_1,\psi_2,\psi_3,\psi_4,\psi_5,\psi_6,\psi_7$ are functions on $C$. These are the coefficients of the invariantly defined harmonic curvature. In fact, if $$d\omega^0=\chi(\omega^1\wedge\omega^4-3\omega^2\wedge\omega^3)$$ on $C$ and we write $$\label{we_write} d\omega^i=\sum_{1\leq j<k\leq4}a^i{}_{jk}\omega^j\wedge\omega^k \enskip\bmod\omega^0,\enskip\mbox{for }i=1,2,3,4$$ then $$\label{psi}\begin{array}{rcl}\psi_0&=&a^1{}_{34}\\ \psi_1&=&-2a^1{}_{24}+3a^2{}_{34}\\ \psi_2&=&3a^1{}_{14}+a^1{}_{23}-6a^2{}_{24}+3a{}^3{}_{34}\\ \psi_3&=&-2a^1{}_{13}+9a^2{}_{14}+3a^2{}_{23}-6a^3{}_{24}+a^4{}_{34}\\ \psi_4&=&a^1{}_{12}-6a^2{}_{13}+9a^3{}_{14}+3a^3{}_{23}-2a^4{}_{24}\\ \psi_5&=&3a^2{}_{12}-6a^3{}_{13}+3a^4{}_{14}+a^4{}_{23}\\ \psi_6&=&3a^3{}_{12}-2a^4{}_{13}\\ \psi_7&=&a^4{}_{12}. \end{array}$$ [From]{} the general theory of parabolic geometry [@CS Theorem 3.1.12] we find the following characterisation of flat $G_2$ contact structures. \[harmonic\_curvature\] The local symmetry algebra of the $G_2$ contact structure specified by a co-frame adapted according to is the split exceptional Lie algebra $G_2$ if and only if $\psi_0,\psi_1,\psi_2,\psi_3,\psi_4,\psi_5,\psi_6,\psi_7$ given by all vanish. Moreover, in this case, the manifold $C$ is locally isomorphic to the homogeneous model. The formul[æ]{} (\[we\_write\]) and (\[psi\]) may alternatively be derived as follows. According to (\[extension\]), the exterior derivative $d:\Wedge^1\to\Wedge^2$ gives rise, via the diagram $$\begin{array}{ccccccccc} 0&\to&{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(24,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}&\to&\Wedge^1&\to&{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -2$}} \end{picture}}&\to&0\\ &&&&d\downarrow\phantom{d}\\ 0&\to&{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -3$}} \end{picture}}&\to&\Wedge^2&\to& {\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 0$}} \put(24,8){\makebox(0,0){$\scriptstyle -1$}} \end{picture}}\oplus{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 4$}} \put(24,8){\makebox(0,0){$\scriptstyle -3$}} \end{picture}}&\to&0,\\ \end{array}$$ to an invariantly defined first order differential operator $$\label{rumin}\nabla:{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -2$}} \end{picture}}\to{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 4$}} \put(24,8){\makebox(0,0){$\scriptstyle -3$}} \end{picture}}.$$ In fact, this is nothing more than the second operator in the Rumin complex [@R], which depends only on the contact structure on $C$. The exterior derivative, on the other hand, may be written as $\omega_b\mapsto\nabla_{[a}\omega_{b]}$ for any [*torsion-free*]{} connection $\nabla_a$ on $TC$. Thus, the Rumin operator (\[rumin\]) may be written with [*spinor indices*]{} [@OT], adapted to our cause, as $$\label{rumin_with_spinors} \omega_{ABC}\longmapsto\nabla_{(AB}{}^H\omega_{CD)H}.$$ One may readily check that the formul[æ]{} (\[we\_write\]) and (\[psi\]) amount to the stipulation that $$\label{stip} \psi_{ABCDEFG}\pi^A\pi^B\pi^C\pi^D\pi^E\pi^F\pi^G =\pi^A\pi^B\pi^C\pi^D\nabla_{AB}{}^H(\pi_C\pi_D\pi_H)$$ for all sections $\pi_A$ of $S^*={\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(24,8){\makebox(0,0){$\scriptstyle \enskip-2/3$}} \end{picture}}$. Note, by the Leibniz rule $$\begin{array}{rcl}\pi^A\pi^B\pi^C\pi^D\nabla_{AB}{}^H(f\pi_C\pi_D\pi_H) &=&f\pi^A\pi^B\pi^C\pi^D\nabla_{AB}{}^H(\pi_C\pi_D\pi_H)\\ &&\enskip{}+\pi^A\pi^B\underbrace{\pi^C\pi^D\pi_C}_{=0}\pi_D\pi_H \nabla_{AB}{}^Hf,\end{array}$$ that the right hand side of (\[stip\]) is homogeneous of degree $7$ over the functions and, therefore, automatically of the form given on the left. It follows that $\psi_{ABCDEFG}$ is the obstruction to writing (\[rumin\]) as $$\omega_{ABC}\longmapsto{\mathcal{D}}_{(AB}{}^H\omega_{CD)H},$$ where ${\mathcal{D}}_{ABC}$ is induced by a connection on $S^*={\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 1$}} \put(24,8){\makebox(0,0){$\scriptstyle \enskip-2/3$}} \end{picture}}$, because the Leibniz rule for such a connection would imply that $${\mathcal{D}}_{AB}{}^H(\pi_C\pi_D\pi_H) =\pi_C\pi_D{\mathcal{D}}_{AB}{}^H\pi_H +\pi_C\pi_H{\mathcal{D}}_{AB}{}^H\pi_D +\pi_D\pi_H{\mathcal{D}}_{AB}{}^H\pi_C$$ and the right hand side of (\[stip\]) would therefore vanish. In other words, the formula (\[rumin\_with\_spinors\]) depends on $\nabla_a$ being torsion-free and $\psi_{ABCDEFG}$ may be seen as some invariant part of the [*partial torsion*]{} of a freely chosen spinor connection ${\mathcal{D}}_a:S\to\Wedge^1\otimes S$. More specifically, suppose ${\mathcal{D}}_a:S\to\Wedge^1\otimes S$ is any connection and define its [*partial torsion*]{} $T_{ABCD}{}^{EFG}=T_{(ABCD)}{}^{(EFG)}$ according to $${\mathcal{D}}_{(AB}{}^E{\mathcal{D}}_{CD)E}f =T_{ABCD}{}^{EFG}{\mathcal{D}}_{EFG}f,\enskip \forall\mbox{ smooth functions }f.$$ Changing the connection ${\mathcal{D}}_a$, leads to a change of [*partial connection*]{} ${\mathcal{D}}_{ABC}:S^*\to{\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -2$}} \end{picture}}\otimes S^*$ according to $$\widehat{\mathcal{D}}_{ABC}\pi_D ={\mathcal{D}}_{ABC}\pi_D-\Gamma_{ABCD}{}^E\pi_E, \enskip\mbox{where }\Gamma_{ABCD}{}^E=\Gamma_{(ABC)D}{}^E$$ and, therefore, an induced change on ${\begin{picture}(28,10) \put(4,1.1){\makebox(0,0){$\bullet$}} \put(24,1.5){\makebox(0,0){$\times$}} \put(4,3.5){\line(1,0){18}} \put(4,1.5){\line(1,0){20}} \put(4,-.5){\line(1,0){18}} \put(14,1.5){\makebox(0,0){$\langle$}} \put(4,8){\makebox(0,0){$\scriptstyle 3$}} \put(24,8){\makebox(0,0){$\scriptstyle -2$}} \end{picture}}=\bigodot^3\!S^*$, namely $$\widehat{\mathcal{D}}_{ABC}\omega_{DEF}= {\mathcal{D}}_{ABC}\omega_{DEF}-3\Gamma_{ABC(D}{}^G\omega_{EF)G}.$$ It follows that $$\widehat{\mathcal{D}}_{AB}{}^E\widehat{\mathcal{D}}_{CDE}f ={\mathcal{D}}_{AB}{}^E{\mathcal{D}}_{CDE}f -3\Gamma_{AB}{}^E{}_{(C}{}^G{\mathcal{D}}_{DE)G}f$$ and, therefore, that $$\widehat{\mathcal{D}}_{(AB}{}^E\widehat{\mathcal{D}}_{CD)E}f \!=\!{\mathcal{D}}_{(AB}{}^E{\mathcal{D}}_{CD)E}f -2\Gamma_{(AB}{}^E{}_C{}^F{\mathcal{D}}_{D)EF}f +\Gamma_{H(AB}{}^{HE}{\mathcal{D}}_{CD)E}f$$ whence the partial torsion of ${\mathcal{D}}_a$ changes according to $$\widehat{T}_{ABCD}{}^{EFG}=T_{ABCD}{}^{EFG} -2\Gamma_{(AB}{}^{(E}{}_C{}^F\delta_{D)}{}^{G)} +\Gamma_{H(AB}{}^{H(E}\delta_C{}^F\delta_{D)}{}^{G)}.$$ In particular, the trace-free part of $T_{ABCD}{}^{EFG}$, equivalently $$\psi_{ABCDEFG}\equiv T_{(ABCDEFG)},$$ is an invariant of the $G_2$ contact structure. [11]{} R.L. Bryant, [*Two exotic holonomies in dimension four, path geometries, and twistor theory*]{}, Complex Geometry and Lie Theory, Proc. Sympos. Pure Math. vol. 53, Amer. Math. Soc. 1991, pp. 33–88. A. Čap and J. Slovák, [*Parabolic Geometries I: Background and General Theory*]{}, American Mathematical Society 2009. A. Cannas da Silva, [*Lectures on Symplectic Geometry*]{}, Springer 2001. M. Dunajski and M.G. Eastwood, [*Metrisability of three-dimensional path geometries*]{}, Eur. Jour. Math. [**2**]{} (2016) 809–834. M.G. Eastwood and P. Nurowski, [*Aerobatics of flying saucers*]{}, (2018). M.F. Engel, [*Sur un groupe simple á quatorze paramètres*]{}, C.R. Acad. Sci. Paris [**116**]{} (1893) 786–788. D.J.F. Fox, [*Contact projective structures*]{}, Indiana Univ. Math. Jour. [**54**]{} (2005) 1547–1598. J.R. Harrison, [*Some Problems in the Invariant Theory of Parabolic Geometries*]{}, PhD thesis, University of Edinburgh 1995. C.R. LeBrun, [*Twistor CR manifolds and three-dimensional conformal geometry*]{}, Trans. Amer. Math. Soc. [**284**]{} (1984) 601–616. C.R. LeBrun, [*${\mathcal{H}}$-space with a cosmological constant*]{}, Proc. Roy. Soc. London [**A 380**]{} (1982) 171–185. J.M. Lee, [*Manifolds and Differential Geometry*]{}, American Mathematical Society 2009. V.S. Matveev and A. Trautman, [*A criterion for compatibility of conformal and projective structures*]{}, Comm. Math. Phys. [**329**]{} (2014) 821–825. R. Montgomery, [*A Tour of Subriemannian Geometries, their Geodesics and Applications*]{}, American Mathematical Society 2002. P. Nurowski, [*Projective versus metric structures*]{}, Jour. Geom. Phys. [**62**]{} (2012) 657–674. R. Penrose and W. Rindler, [*Spinors and Space-time, vol. 1*]{}, Cambridge University Press 1984. M. Rumin, Un complexe de formes différentielles sur les variétés de contact, C.R. Acad. Sci. Paris Math. [**310**]{} (1990) 401–404. M. Takeuchi, [*Lagrangean contact structures on projective cotangent bundles*]{}, Osaka Jour. Math. [**31**]{} (1994) 837–860. [^1]: This work was supported by the Simons Foundation grant 346300 and the Government MNiSW 2015–2019 matching fund. It was written whilst the first was visiting the Banach Centre at IMPAN in Warsaw for the Simons ‘Symmetry and Geometric Structures’ and during another visit to supported by the Polish National Science Centre (NCN) via the POLONEZ grant 2016/23/P/ST1/04148, which received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sk[ł]{}odowska-Curie grant agreement No. 665778.

ArXiv

--- abstract: 'A rigorous *ab initio* derivation of the (square of) Dirac’s equation for a single particle with spin is presented. The general Hamilton-Jacobi equation for the particle expressed in terms of a background Weyl’s conformal geometry is found to be linearized, exactly and in closed form, by an *ansatz* solution that can be straightforwardly interpreted as the “quantum wave function” $\psi_4$ of the 4-spinor Dirac’s equation. In particular, all quantum features of the model arise from a subtle interplay between the conformal curvature of the configuration space acting as a potential and Weyl’s “pre-potential”, closely related to $\psi_4$, which acts on the particle trajectory. The theory, carried out here by assuming a Minkowsky metric, can be easily extended to arbitrary space-time Riemann metric, e.g. the one adopted in the context of General Relativity. This novel theoretical scenario, referred to as “Affine Quantum Mechanics”, appears to be of general application and is expected to open a promising perspective in the modern endeavor aimed at the unification of the natural forces with gravitation.' author: - Enrico Santamato - Francesco De Martini title: '*Ab initio* derivation of the quantum Dirac’s equation by conformal differential geometry: the “Affine Quantum Mechanics”' --- Introduction ============ The particle spin theory is one of the cornerstones of quantum mechanics. Consequently, being the spin a peculiar feature of the quantum world, any attempt to find a classical system behaving as a spinning quantum particle is generally considered hopeless. In this work, we show that a wave equation for quantum spin (and, in particular, the square of Dirac’s spin 1/2 equation) may be derived from the mechanics of the relativistic top in a curved configuration space.\ Our approach is based on the theory proposed some time ago by one of us to derive the Schrödinger and the Klein-Gordon equations from mechanics in curved spaces [@santamato84; @*santamato85]. The main physical assumption of this theory, we shall refer to henceforth as “Affine Quantum Mechanics” (AQM), is that the origin of quantum effects is a feedback between geometry and dynamics [^1]. More precisely, in this work we ascribe the quantum effects to the presence of a not trivial parallel transport law in the particle configuration space bearing a nonzero curvature. The metric has no role here and can be arbitrarily prescribed. In fact, at the very fundamental level, the space curvature is originated by the affine connections and not by the metric tensor. The AQM assumes that the actual affine connections (and hence the scalar curvature) of the particle configuration space are affected by the dynamics of the particle itself and that, in turn, the space scalar curvature acts on the particle as a potential. Thus, the particle motion and the space affine connections must be determined consistently. The overall physical picture is analogous to the situation prevailing in general relativity: geometry is not prescribed; rather it is determined by the physical reality. In turn, geometry acts as a “guidance field” for matter. The idea of a “guidance field” to explain quantum phenomena dates back to De Broglie. The AQM identifies the origin of the already mysterious De Broglie field with the curvature of space-time and casts its effects on a firm and plausible theoretical frame. However, unlike in general relativity, the space curvature is ascribed by AQM to the affine connections rather than to the metric of the geometry. In this way, gravitational and quantum phenomena share a common geometrical origin, but are based on *independent* geometrical objects: the metric tensor for the former and the affine connections for the latter. As in general relativity, the geometric approach forces to describe matter as a fluid or as a bundle of elementary trajectories rather than as a single point particle moving along one trajectory. In this respect, AQM is somewhat related to the hydrodynamic approach to quantum mechanics first proposed by Madelung [@madelung26] and then developed by Bohm [@bohm83a; @*bohm83b]. In the Madelung-Bohm approach the particle trajectories are deterministically governed by classical mechanics and quantum effects are due to a “quantum potential” of quite mysterious origin, whose gradient acts as a newtonian force on the particle. According to the AQM the active potential originates from geometry, as does gravitation, and arises form the space curvature due to the presence of the non trivial affine connections of the Weyl conformal geometry [@weyl18; @*weyl]. The relativistic top ==================== We start considering the simplest model for the relativistic spinning particle, namely the top described by six Euler angles, as made, for example, by Frenkel [@frenkel26], Thomas [@thomas26] and in the classic work by Hanson and Regge [@hanson74]. Subsequent important works on relativistic spinning particles can be found in many textbook [@souriau; @corben; @sudarshan]. Here, we simply imagine that the particle follows a path $x^\mu=x^\mu(\sigma)$ in space-time, where $\sigma$ is an arbitrary parameter along the path, and that it carries along with itself a moving fourleg $e^\mu_a=e^\mu_a(\sigma)$ ($\mu,a=0,\dots,3$). The fourleg vectors $e^\mu_a$ are normalized according to $g_{\mu\nu}e^\mu_ae^\nu_b=g_{ab}$ where $g_{\mu\nu}=g_{ab}={\mbox{diag}}(-1,1,1,1)$ is the Minkowski metric tensor. Setting $\Lambda(\sigma)=\{e^\mu_a(\sigma)\}$ and $G=\{g_{\mu\nu}\}$, the normalization relations can be cast in the matrix form as $\Lambda^T G\Lambda=G$, showing that the $4{\times}4$ matrix $\Lambda(\sigma)\in SO(3,1)$ is a proper Lorentz matrix. The derivative of $e^\mu_a(\sigma)$ with respect to $\sigma$ can be written as $de^\mu_a/d\sigma=\omega^\mu_\nu e^\nu_a$. The contravariant tensor $\omega^{\mu\nu}=\omega^\mu_\sigma g^{\sigma\nu}$ is skewsymmetric and can be considered as the “angular velocity” of the top in space-time. The free Lagrangian of this minimal relativistic top is $$\begin{aligned} \label{eq:L0} L_0 &=& mc\sqrt{-g_{\mu\nu}\frac{dx^\mu}{d\sigma} \frac{dx^\nu}{d\sigma} - a^2 g_{\mu\nu}g^{ab}\frac{de^\mu_a}{d\sigma}\frac{de^\nu_b}{d\sigma}}= \nonumber \\ &=& mc\sqrt{-g_{\mu\nu}\frac{dx^\mu}{d\sigma} \frac{dx^\nu}{d\sigma} - a^2 \omega_{\mu\nu} \omega^{\mu\nu}},\end{aligned}$$ where $m$ is the particle mass, $c$ is the speed of light, and $a$ is a constant having the dimension of a length. For a quantum particle of mass $m$ we expect $a$ to be of the order of the particle Compton wavelength. The square root in Eq. (\[eq:L0\]) ensures that $L_0$ is parameter invariant. In the presence of an external electromagnetic field, the total Lagrangian becomes $L=L_0+L_{em}$, where the electromagnetic interaction Lagrangian is taken as $$\label{eq:Lem} L_{em} = \frac{e}{c}A_\mu \frac{dx^\mu}{d\sigma} + \frac{\kappa e}{4c}\;a^2 F_{\mu\nu}\omega^{\mu\nu}$$ where $e$ is the particle charge and $F_{\mu\nu}$ is given by $F_{\mu\nu}=\partial A_\nu/\partial x^\mu-\partial A_\mu/\partial x^\nu$ with four-potential $A_\mu$ given by $A_\mu=(-\phi,\bm A)$, $\phi$, $\bm A$ being the scalar and vector electromagnetic potentials, respectively. Finally, $\kappa$ is a numeric constant that will be identified as the particle gyromagnetic ratio. The fourleg components $e^\mu_a$ (and the $SO(3,1)$ group) are parametrized by six “Euler angles” $\theta^\alpha$ ($\alpha=1,\dots,6$), so that the configuration space spanned by the space-time coordinates and the Euler angles is ten dimensional. When $\omega^{\mu\nu}$ is written in terms of the angles $\theta^\alpha$ and their derivatives, the free-particle Lagrangian $L_0$ assumes the standard form $$\label{eq:L0geo} L_0=mc\frac{ds}{d\sigma}= mc\sqrt{-g_{ij}\frac{dq^i}{d\sigma}\frac{dq^j}{d\sigma}},$$ where $q^i=\{x^\mu,\theta^\alpha\}$ $(i=0,\dots,9)$ are the ten coordinates spanning the dynamical configuration space of the top [^2]. Similarly, the electromagnetic interaction Lagrangian $L_{em}$ assumes the standard form $L_{em}=(e/c)A_idq^i/d\sigma$, where $A_i=(A_\mu,A_\alpha)$ is a ten dimensional covariant vector. The last six entries $A_\alpha$ of $A_i$ are linear combinations of the components of the magnetic and electric fields $\bm H(x)$ and $\bm E(x)$, respectively [^3]. The quantities $g_{ij}$ in Eq. (\[eq:L0geo\]) define the distance $ds=\sqrt{-g_{ij}dq^idq^j}$ in the top configuration space, which is so converted into the 10-D Riemann metric space $V_{10}={\cal M}_4{\times}SO(3,1)$. The extremal curves of $L_0$ are the geodetics of this space. The metric tensor $g_{ij}$ in $V_{10}$ has the diagonal block form $g_{ij}=\pmatrix{ g_{\mu\nu} & 0 \cr 0 & g_{\alpha\beta} }$, where $g_{\mu\nu}$ is the Minkowski metric and $g_{\alpha\beta}$ is the metric of the parameter space of Lorentz group with signature $(+,+,+,-,-,-)$ and we assumed the Euler angles $\theta^\alpha$ ordered so that the first three angles $\theta^\alpha$ for $\alpha=1,2,3$ are associated with space rotations, and the last three angles $\theta^\alpha$ for $\alpha=4,5,6$ to Lorentz boosts. The classical mechanics induced by the extremals of the Lagrangian $L$ on the space $V_{10}$ is well known [@souriau; @corben; @sudarshan]. Here it is enough noticing that neither the time-like vector $e^\mu_0$ of the moving fourleg is identified by the particle four-velocity $u^\mu=dx^\mu/d\tau$, $d\tau=\sqrt{-g_{\mu\nu}dx^\mu dx^\nu}$ being the proper time, nor Weysenhoff’s kinematical constraints $\omega_{\mu\nu}u^\nu=0$ are imposed, in general. As a consequence, the so called “center-of-mass” space-time trajectory $x^\mu(\tau)$ and the so called “center-of-energy” space-time trajectory $y^\mu(\tau)$ of our top (obtained from $dy^\mu/d\tau=e^\mu_0$) are different [see Ref.  @sudarshan chap. 20]. The main advantage of using a top described by six Euler angles is that the usual methods of analytical mechanics can be applied without worrying about kinematical constraints; but the usual picture of spin as the coadjoint action of the little Poincaré group on the particle momentum space [see Ref.  @souriau Chap. 3, sec. 13] is generally lost. The conformal relativistic top ============================== The main result of the present work is to show that the square of Dirac’s equation for the quantum spin 1/2 particle can be obtained by a simple change of Lagrangian so to provide Weyl’s conformal invariance to the particle dynamics without introducing any concept extraneous to the classical world. In other words, to describe the quantum spinning particle, we propose to use in place of the Lagrangian $L=L_0+L_{em}$ a new Lagrangian $\bar L$ which is invariant under the conformal change $g_{ij}\rightarrow\rho(q)g_{ij}$ of the configuration space metric. As suggested elsewhere for spinless particles [@santamato84; @*santamato85], we introduce conformal invariance by assuming that the configuration space of the top is a Weyl space with metric $g_{ij}$ and integrable Weyl’s connections $\Gamma^i_{jk}$ given by $$\label{eq:GammaWeyl} \Gamma^i_{jk}=-{\ensuremath{\left\{^{\,\,i}_{jk}\right\}}}+ \delta^i_j\phi_k+\delta^i_k\phi_j+g_{jk}\phi^i,$$ where ${\ensuremath{\left\{^{\,\,i}_{jk}\right\}}}$ are the Cristoffel symbols out of the metric $g_{ij}$, $\phi^i=g^{il}\phi_l$, and $\phi_i$ is the Weyl potential that we assume to be integrable, viz. $\phi_i=\chi^{-1}\partial\chi/\partial q^i$. As Lagrangian we take $$\label{eq:Lbar} \bar L = \xi\hbar\sqrt{-R_W g_{ij}\frac{dq^i}{d\sigma}\frac{dq^j}{d\sigma}} + L_{em},$$ where $\xi$ is a numeric constant, $L_{em}$ is given by Eq. (\[eq:Lem\]) and $R_W$ is the Weyl scalar curvature calculated from the connections (\[eq:GammaWeyl\]), viz. $$\begin{aligned} \label{eq:RW} R_W &=& R + 2(n-1)\nabla_k\phi^k-(n-1)\phi_k\phi^k = \nonumber\\ &=& R + 2(n-1)\frac{\nabla_k\nabla^k\chi}{\chi} - n(n-1)\frac{\nabla_k\chi\nabla^k\chi}{\chi^2},\end{aligned}$$ where $\nabla_i$ denote the covariant derivatives built out from the Cristoffel symbols [[$\left\{^{\,\,i}_{jk}\right\}$]{}]{}. The Lagrangian (\[eq:Lbar\]) is manifestly invariant under the conformal changes of the metric and Weyl potential $g_{ij}\rightarrow\rho g_{ij}$ and $\phi_i\rightarrow\phi_i-\rho^{-1}\partial\rho/\partial q^i$, respectively, provided the fields $A_i$ are Weyl invariant. The dynamic theory derived from the Lagrangian $\bar L$ applying the extremal action principle is conformally invariant too. Moreover, it is worth noting that the Lagrangian $\bar L$ is massless, because the particle mass $m$ was replaced by Weyl’s curvature according to $mc\rightarrow \xi\hbar \sqrt{R_W}$. We will call the top described by the Lagrangian $\bar L$ the conformal relativistic top. The Weyl curvature field $R_W(q)$ in Eq. (\[eq:RW\]) acts as a scalar potential on the conformal top and, because it depends on $\chi$ and its derivatives, the field $\chi$ acts on the conformal top as a sort of pre-potential. The paths followed by the conformal top in the configuration space $V_{10}={\cal M}_4{\times}SO(3,1)$ are the extremal curves of the action integral $\int \bar L\:d\sigma$. Of particular importance are the bundles of extremals belonging to a family of equidistant hypersurfaces $S=\mathrm{const.}$ in the configuration space. These bundles are obtained from the solutions of the Hamilton-Jacobi equation associated with $\bar L$ $$\begin{aligned} \label{eq:HJbar} g^{ij} \left(\frac{\partial S}{\partial q^i}-\frac{e}{c}A_i\right) \left(\frac{\partial S}{\partial q^j}-\frac{e}{c}A_j\right)= \nonumber\\ = g^{ij}\left(D_i S-\frac{e}{c}A_i\right) \left(D_j S -\frac{e}{c}A_j\right) = -\hbar^2\xi^2 R_W\end{aligned}$$ by integrating the differential equations $$\label{eq:dqds} \frac{dq^i}{ds}= \frac{g^{ij}(\frac{\partial S}{\partial q^j}-\frac{e}{c}A_j)} {[g^{mn}(\frac{\partial S}{\partial q^m}-\frac{e}{c}A_m) (\frac{\partial S}{\partial q^n}-\frac{e}{c}A_m)]^{1/2}}.$$ Moreover, we assume that the action function $S$ obeys the auxiliary divergence condition $$\label{eq:div} D_k\left(D^k S-\frac{e}{c}A_k\right) = 0.$$ We may think this condition as stating that the trajectories in the bundle do not intersect in the considered region of the configuration space. In Eqs. (\[eq:HJbar\]) and (\[eq:div\]) $D_i$ denote the Weyl co-covariant derivatives with respect of the coordinate $q^i$ [^4]. The use of the co-covariant derivatives makes explicit the coordinate and conformal gauge covariance of Eqs. (\[eq:HJbar\]) and (\[eq:div\]). When written out in full, Eq. (\[eq:div\]) states the conservation of the Weyl-invariant current density $$\label{eq:j} j^i=\chi^{-(n-2)}\sqrt{g}\:g^{ij}(\partial S/\partial q^j-(e/c)A_j).$$ Equations (\[eq:HJbar\]) and (\[eq:div\]) are a set of nonlinear partial differential equations for the unknown functions $S(q)$ and $\chi(q)$, once the metric tensor $g_{ij}(q)$ is given. The nonlinear problem implied by Eqs. (\[eq:HJbar\]) and (\[eq:div\]) looks very hard at first glance. However, bu introducing the complex scalar function $\psi$ of Weyl’e type $w(\psi)=-(n-2)/4$ given by $$\label{eq:psidef} \psi(q)=\chi(q)^{-\frac{n-2}{2}}e^{i\frac{S(q)}{\hbar}}$$and fixing $\xi$ according to $$\label{eq:gamma} \xi^2=\frac{n-2}{4(n-1)}=\frac{2}{9},$$ where $n=10$ is the dimensionality of the top configuration space, converts Eqs. (\[eq:HJbar\]) and (\[eq:div\]) into the *linear* differential equation $$\label{eq:KG} g^{ij}\left(\hat p_i-\frac{e}{c}A_i\right) \left(\hat p_j-\frac{e}{c}A_j\right)\psi+ \hbar^2 \xi^2 R\psi=0.$$ where $\hat p_=-i\hbar \nabla_i$. This is a striking result as it demonstrates that the Hamilton-Jacobi equation, applied to a general dynamical problem can be transformed into a linear eigenvalue equation, the foremost ingredient of the formal structure of quantum mechanics and of the Hilbert space theory. Note that the transition from the Hamilton-Jacobi Eq. (\[eq:HJbar\]) to the quantum mechanical Eq. (\[eq:KG\]) implies the adoption of a formally trivial albeit conceptually significant transformation: $p_i = \partial S/\partial x^i \rightarrow (-i\hbar\partial/\partial x^i) \times(i S(x)/\hbar)$, where the two factors are commonly interpreted as momentum operator $(\hat p)$ and (complex) phase, respectively. This transformation precisely represents the transition from Hamilton’s classical dynamics to quantum mechanics in our theory. In the absence of the electromagnetic field ($A_i=0$), Eq. (\[eq:KG\]) reduces to $\hat L\psi=(-\Delta+\xi^2 R)\psi$, where $\Delta$ is the Laplace-Beltrami operator and $\hat L$ is the conformal Laplacian, also known as the Laplace-de Rham operator associated with the metric $g_{ij}$. The value of $\xi$ given by Eq. (\[eq:gamma\]) ensures that Eq. (\[eq:KG\]) is conformally invariant. The Laplace-de Rham Eq. (\[eq:KG\]) resembles the covariant quantum Klein-Gordon wave-equation in the configuration space with the mass term $m^2c^2$ replaced by the curvature potential term $\hbar^2\xi^2 R(q)$. What it is more surprising is that any explicit reference to the Weyl pre-potential $\chi(q)$ and to the Weyl curvature $R_W$ has been cancelled out from Eq. (\[eq:KG\]). In fact, the curvature $R(q)$ and the covariant derivatives $\nabla_i$ in Eq. (\[eq:KG\]) are calculated using the Cristoffel symbols derived from the metric $g_{ij}$. Finally, the (Riemann) curvature $R$ of the top configuration space is constant in our case, and it is given by $R=6/a^2$. Moreover, the conserved current density $j^i$ in Eq. (\[eq:j\]) can be written in the alternative form $$\label{eq:jalt} j^i=|\psi|^2\sqrt{g}\:g^{ij}(\partial S/\partial q^j-(e/c)A_j).$$ without any explicit reference to the underlying Weyl’s geometry. The current (\[eq:jalt\]) together with Eq. (\[eq:dqds\]) shows that the scalar density $|\psi|^2$ is transported along the particle trajectory in the configuration space, allowing the optional statistical interpretation of the wavefunction $\psi$ according to Born’s quantum mechanical rule [@santamato84; @*santamato85]. The reduction of Eqs. (\[eq:HJbar\]) and (\[eq:div\]) to the wave-equation (\[eq:KG\]) is the central result of this work, because it builds a bridge between the quantum and the classical worlds. The quantum wave equation (\[eq:KG\]) with the $|\psi|^2$ Born prescription is mathematically equivalent to the classical Hamilton-Jacobi Eq. (\[eq:HJbar\]) associated with the conformally invariant Lagrangian $\bar L$; Born’s rule comes out in a very natural way from the conformally invariant zero divergence current requirement along any Hamiltonian bundle of trajectories in the configuration space. It is also worth noting that Born’s rule relays on the particular choice of the conformal gauge made to obtain the Laplace-de Rham Eq. (\[eq:KG\]). By changing the gauge, we can make $|\psi|^2\rightarrow |\bar\psi|^2=1$ in which case the Weyl curvature reduces to the Riemann curvature $\bar R$ of the not trivial metric $\bar g_{ij}=\chi^{-2}g_{ij}=|\psi|^{4/(n-2)}g_{ij}$. Our final step is now to show that the wave equation (\[eq:KG\]) is able to account for the quantum spin 1/2.\ Equivalence with Dirac’s equation ================================= We first note that Eq, (\[eq:KG\]) is invariant under parity $P$, so we may look for solutions $\psi(q)$ which also are invariant under $P$. These solutions can be cast in the mode expansion form $$\begin{aligned} \label{eq:psiuv} \psi_{uv}(q) = D^{(u,v)}(\Lambda^{-1})^\sigma_{\sigma'}\psi^{\sigma'}_{\sigma}(x)&+& D^{(v,u)}(\Lambda^{-1})^{\dot\sigma}_{\dot\sigma'} \psi^{\dot\sigma'}_{\dot\sigma}(x)\nonumber \\ &&(u\le v)\end{aligned}$$ where $D^{(u,v)}(\Lambda)^\sigma_{\sigma'}$ are the $(2u+1){\times}(2v+1)$ matrices representing the Lorentz transformation $\Lambda(\theta)=\{e^\mu_a(\theta)\}$ in the irreducible representation labeled by the two numbers $u,v$ given by $2u,2v=0,1,2,\dots$, and the $\psi^{\sigma'}_\sigma(x)$ and $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ are expansion coefficients depending on the space-time coordinates $x^\mu$ alone. The matrices $D^{(u,v)}(\Lambda)$ and $D^{(v,u)}(\Lambda)$ depend on the Euler angles $\theta^\alpha$ only, and provide conjugate representations of the Lorentz transformations [^5]. As the notation suggests, the invariance of $\psi_{uv}(q)$ under Lorentz transformations implies that $\psi^{\sigma'}_{\sigma}(x)$ and $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ change as undotted and dotted contravariant spinors, respectively [^6]. In Eq. (\[eq:psiuv\]) both dotted and undotted spinors appear, because we are interested in solutions $\psi_{uv}(q)$ of Eq. (\[eq:KG\]) which are parity invariant. Indeed, both terms on the right of Eq. (\[eq:psiuv\]) obey Eq. (\[eq:KG\]) separately, each one providing not parity invariant solutions to Eq. (\[eq:KG\]). In the case of spin 1/2, the spinors $\psi^{\sigma'}_{\sigma}(x)$ and $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ have two components. The use of two-components spinors in place of the four-component Dirac’s spinors have been extensively discussed in the literature [@brown58]. In this paper, however, we will limit to parity invariant solutions of Eq. (\[eq:KG\]) described by four-component Dirac’s spinors. Insertion of the expansion (\[eq:psiuv\]) into the wave-equation (\[eq:KG\]) yields to the following equation for the coefficients $\psi^{\sigma'}_{\sigma}(x)$ and $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ $$\begin{aligned} \label{eq:coeff} \left[g^{\mu\nu}\left(\hat p_\mu -\frac{e}{c}A_\mu\right) \left(\hat p_\nu -\frac{e}{c}A_\nu\right)\right. &+&\left. \hbar^2\xi^2 R\right]\psi(x) +\nonumber \\ \mbox{} + \Delta_J \psi(x) &=& 0\end{aligned}$$ where $\hat p_\mu=-i\hbar\partial_\mu$ and $\psi(x)$ denotes either $\psi^{\sigma'}_{\sigma}(x)$ or $\psi^{\dot\sigma'}_{\dot\sigma}(x)$. Finally, $\Delta_J$ is a $(2u+1){\times}(2v+1)$ matrix depending on the space-time coordinates $x^\mu$ only, given by $$\label{eq:DeltaJ} \Delta_J = \left[\frac{\hbar}{a}\bm J - \frac{\kappa e a}{2c}\bm H\right]^2- \left[\frac{\hbar}{a}\bm K - \frac{\kappa e a}{2c}\bm E\right]^2.$$ Here $\bm J$ and $\bm K$ are the generators of the Lorentz group in the dotted or undotted (conjugate) representation, according if $\psi^{\sigma'}_\sigma(x)$ or $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ are considered. The connection with the spin $1/2$ Dirac’s theory is made by taking $(u,v)=(0,\frac{1}{2})$ in Eq. (\[eq:psiuv\]) so that $D^{(0,1/2)}(\Lambda)\in SL(2,C)$. Then, introducing the Dirac four-spinor $\Psi_D=\pmatrix{\psi^{\sigma'}_{\sigma}\cr \psi^{\dot\sigma'}_{\dot\sigma}}$ with $\sigma=\dot\sigma$ fixed and setting $\kappa=2$ for the electron, Eq. (\[eq:coeff\]) yields $$\begin{aligned} \label{eq:Dirac} \left[g^{\mu\nu}\left(\hat p_\mu -\frac{e}{c}A_\mu\right) \left(\hat p_\nu -\frac{e}{c}A_\nu\right)- \frac{e\hbar}{c}(\bm\Sigma{\ensuremath\cdot}\bm H - i\bm\alpha{\ensuremath\cdot}\bm E)\right]\Psi_D + \nonumber\\ \mbox{+} \left[\frac{e^2a^2}{c^2}(H^2-E^2)+ \frac{3\hbar^2}{2a^2}(1+4\xi^2)\right]\Psi_D = 0,\end{aligned}$$ where: $\bm\Sigma=\pmatrix{\bm\sigma & 0\cr 0 & \bm\sigma}$, $\bm\alpha=\pmatrix{\bm\sigma & 0\cr 0 & -\bm\sigma}$, and $\bm\sigma=\{\sigma_x,\sigma_y\sigma_z\}$ are the usual Pauli matrices. Setting $a=(\hbar/mc)\sqrt{3(1+4\xi^2)/2}$, where $m$ is the particle mass, and neglecting the term $(ea/c)^2(H^2-E^2)=(ea/c)^2)(\frac{1}{2}F_{\mu\nu}F^{\mu\nu})$, Eq. (\[eq:Dirac\]) reduces to the square of Dirac’s equation in its spinor representation, viz. [see, for example, Ref.  @landau4 Eq. (32,7a)] $$\label{eq:Diracgamma} \left[\gamma^\mu\gamma^\nu\left(\hat p_\mu -\frac{e}{c}A_\mu\right) \left(\hat p_\nu -\frac{e}{c}A_\nu\right)-m^2c^2\right]\Psi_D = 0,$$ where $\gamma^\mu$ are Dirac’s $4\times 4$ matrices in the spinorial representation. Equation (\[eq:Dirac\]), comprehensive of the electromagnetic term proportional to $F_{\mu\nu}F^{\mu\nu}$, was derived by Schulman by applying usual quantization rules to the relativistic top described by three Euler angles [@schulman70]. In his work, Schulman proposed also generalized wave equations for fields of arbitrary spin, which are equivalent to our Eqs. (\[eq:coeff\]) and (\[eq:DeltaJ\]). We will refer to Schulman’s paper for a detailed discussion about the physical implications of Eq. (\[eq:Dirac\]). However, it is worth noting that the term proportional to $F_{\mu\nu}F^{\mu\nu}$ in Eq. (\[eq:Dirac\]) can be cancelled out just transforming the Weyl curvature $R_W$ in the Lagrangian (\[eq:L0geo\]) according to $R_W\rightarrow R_W-(ea/c\hbar\xi)^2(\frac{1}{2} F_{\mu\nu}F^{\mu\nu})$ so that Eq. (\[eq:Dirac\]) would reproduce the square of Dirac’s equation exactly. Before concluding, we notice that the present approach applies to any spin (see Eq. \[eq:psiuv\]), as expected from a theory based on rotating fourlegs. When spin other than 1/2 are considered, the usual functional relationship for the particle mass $m(s)=a s(s+1)+b$, ($a,b$ constant) is found, because we considered a six internal degrees of freedom Lagrangian with no constraint. It would be then interesting to investigate if appropriate constraints can be imposed either to select the spin 1/2 only, as proposed in Ref. [@balachandran80; @balachandran], or to extend the present fixed mass and spin approach to include a family of particles collected in a different Regge trajectory [@chew61; @atre86; @biedenharn87].\ Interpretation ============== We derived the square of Dirac’s spin 1/2 equation in the framework of the Affine Quantum Mechanics. The spinning particle was described as a conformal relativistic top, obtained from the minimal relativistic top introduced long time ago by Frenkel [@frenkel26] and Thomas [@thomas26] by formally replacing the mass with the Weyl curvature of the top configuration space. The dynamics of the conformal top is invariant under conformal changes of the metric. All trajectories of the conformal top are extremal curves of the corresponding Lagrangian and they can form bundles described by the Hamilton-Jacobi equation. In summary, according to the present AQM interpretation the quantum wave equation related to any quantum mechanical problem (here the Dirac’s spin dynamics) should not be taken as the starting point of the theory, as usually done. It is rather a useful mathematical tool adopted to reduce a generally awkward nonlinear geometro-dynamical problem to a more tractable linear one, in order to obtain simultaneously the dynamical properties of the particle motion and the geometrical properties (Weyl’s potential and curvature) that determine that motion. Then, according to AQM, the physical interpretation of the dynamical theory is different from the usual one since the particle trajectories here are the extremal curves of a Lagrangian, i.e. the geodesics of the Weyl’s field acting on the particle [@wheeler; @*misner]. We may regard the space curvature as a “guidance field” which is ultimately imposed by the conformal invariance of the problem. A stochastic interpretation of the present theory is possible, as pointed out by [@santamato84; @*santamato85]. But it is not necessary since the particle motion on his trajectory can be traced with a precision allowed by the Heisenberg’s uncertainty relations, which are of course valid in the AQM theory. We do believe that the present AQM geometro-dynamical interpretation of Quantum Mechanics leads to two consequences that may result of large relevance in modern Physics. First, several alleged “mysteries of the quantum world” may be partially unveiled or at least find an alternative plausible interpretation. Among them the quantum superposition and, hopefully, the entanglement and quantum nonlocality involved in the EPR paradigm. Moreover, the well known Penrose’s claim for the role of gravitation in the state reduction process can be analyzed in concrete terms by the AQM theory by replacing the metric with the affine properties of space [@penrose1; @*penrose2]. Second, the AQM geometro-dynamical interpretation relates in an obvious and direct way the quantum world with general relativity, in particular with the modern paradigmatic endeavor of quantum gravity. Indeed it could suggest new research paths in that domain [@penrose3] [@wilczek05; @*kiefer08; @*wang06] [@rovelli]. We thank dott. Paolo Aniello for useful suggestions. [30]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ****, (). , ** (, , ), ed. , ****, (). , ****, (). , ****, (). , ** (, , ). , ** (, , ). , ** (, , ). , ****, (). , ** (, , ). , ****, (). , , , , ****, (). , , , , **, vol. of ** (, , ). , ****, (). , ****, (). , , , , , , , ****, (). , ** (, , ). , , , ** (, , ). , ** (, , ). , ** (, ). , ** in **, edited by , , (, , ). , ****, (). (), . , ****, (). , ** (, ). [^1]: According to the Felix Klein Erlangen program, the affine geometry deals with intrinsic geometric properties that remain unchanged under affine transformations (affinities). These preserve collinearity transformations e.g. sending parallels into parallels, and ratios of distances along parallel lines. The Weyl geometry is considered a kind of affine geometry preserving angles. Cfr: H. Coxeter, Introduction to Geometry (Wiley, New York 1969). [^2]: The configuration space of the top described by the Lagrangian $L$ is the principal fiber bundle whose base is the Minkowski space-time ${\cal M}_4$ and whose fiber is $SO(3,1)$, conceived as a proper Lorentz frame manifold. The dynamical invariance group is the whole Poincaré group of the inhomogeneous Lorentz transformations. [^3]: More specifically, $A_\alpha=\xi^a_\alpha(\theta)A_a(x)$ $(\alpha,a=1,\dots,6)$, where $A_a(x)=-\frac{\kappa}{2}\{\bm H(x),\bm E(x)\}$ and $\xi^a_\alpha(\theta)$ are the Killing vectors of the Lorentz group $SO(3,1)$. [^4]: The action of $D_i$ over a tensor field $F$ of Weyl type $w(F)$ is given by $D_i F= \nabla^{(\Gamma)}_i F -2w(F)\phi_iF$, where $\phi_i$ is the Weyl potential and $\nabla^{(\Gamma)}_i$ is the covariant derivative built up from the Weyl connections $\Gamma^i_{jk}$ given by Eq. (\[eq:gamma\]). The Weyl type of $D_i F$ is the same as of $F$. [^5]: The two matrices are related by $[D^{(u,v)}(\Lambda)]^\dag=[D^{(v,u)}(\Lambda)]^{-1}$. [^6]: The spinors $\psi^{\sigma'}_{\sigma}(x)$ and $\psi^{\dot\sigma'}_{\dot\sigma}(x)$ are invariant with respect to their lower indices, which are related to the spin component along the top axis.

ArXiv

--- abstract: 'We present a new software package designed to reduce the computational burden of hadron collider measurements in Parton Distribution Function (PDF) fits. The [APFELgrid]{} package converts interpolated weight tables provided by [APPLgrid]{} files into a more efficient format for PDF fitting by the combination with PDF and $\alpha_s$ evolution factors provided by [APFEL]{}. This combination significantly reduces the number of operations required to perform the calculation of hadronic observables in PDF fits and simplifies the structure of the calculation into a readily optimised scalar product. We demonstrate that our technique can lead to a substantial speed improvement when compared to existing methods without any reduction in numerical accuracy.' address: - | Rudolf Peierls Centre for Theoretical Physics,\ 1 Keble Road, University of Oxford, OX1 3NP, Oxford, UK - 'Theoretical Physics Department, CERN, Geneva, Switzerland' author: - Valerio Bertone - Stefano Carrazza - 'Nathan P. Hartland' bibliography: - 'paper.bib' title: '[APFELgrid]{}: a high performance tool for parton density determinations' --- QCD; parton distribution functions; fast predictions. [**Program Summary**]{}\ [*Manuscript Title:*]{} [APFELgrid]{}: a high performance tool for parton density determinations\ [*Authors:*]{} V. Bertone, S. Carrazza, N.P. Hartland\ [*Program Title:*]{} [APFELgrid]{}\ [*Journal Reference:*]{}\ [*Catalogue identifier:*]{}\ [*Licensing provisions:*]{} MIT license\ [*Programming language:*]{} C++\ [*Computer:*]{} PC/Mac\ [*Operating system:*]{} MacOS/Linux\ [*RAM:*]{} varying\ [*Keywords:*]{} QCD, PDF\ [*Classification:*]{} 11.1 General, High Energy Physics and Computing\ [*External routines/libraries:*]{} [APPLgrid]{}, [APFEL]{}\ [*Nature of problem:*]{}\ Fast computation of hadronic observables under the variation of parton distribution functions.\ [*Solution method:*]{}\ Combination of interpolated weight grids from [APPLgrid]{} files and evolution factors from [APFEL]{} into efficient [FastKernel]{} tables.\ [*Running time:*]{} varying\ \ Introduction {#sec:intro} ============ Measurements at colliders such as the Tevatron and Large Hadron Collider (LHC) have a unique capacity to shed light upon the internal dynamics of the proton and provide constraints upon proton PDFs [@Rojo:2015acz]. However including large hadron collider datasets in PDF fits can provide a significant challenge due to the large computational footprint of performing accurate theoretical predictions over the many iterations required in a fitting procedure. In order to make the fullest use of current and future LHC results, efficient strategies for the computation of these observables must therefore be employed. The typical Monte Carlo software packages used to perform predictions for hadron collider observables cannot be easily deployed in a PDF fit due the processing time required to obtain accurate results (usually of the order of a few hours or more per data point). To overcome such limitations, the typical strategy adopted for fast cross section prediction relies on the precomputation of the partonic hard cross sections in such a way that the standard numerical convolution with any set of PDFs can be reliably approximated by means of interpolation techniques. Such interpolation strategies are implemented in the [ APPLgrid]{} [@Carli:2010rw] and [ FastNLO]{} [@Wobisch:2011ij] projects. For the computation of the hard cross sections, these packages rely on external codes to which they are interfaced by means of a suite of functions allowing for the filling of PDF- and $\alpha_s$-independent look-up tables of cross section weights. Monte Carlo programs such as [MCFM]{} [@Campbell:2010ff] and [NLOJet++]{} [@Nagy:2003tz] have been interfaced directly to [APPLgrid]{}/[FastNLO]{} and more recently dedicated interfaces to automated general-purpose event generators have been developed. The [ aMCfast]{} [@Bertone:2014zva] and [ MCgrid]{} [@DelDebbio:2013kxa] codes can generate interpolation grids in [APPLgrid]{}/[FastNLO]{} format by extracting the relevant information from the [ MadGraph5\_aMC@NLO]{} [@Alwall:2014hca] and [ SHERPA]{} [@Gleisberg:2008ta] event generators respectively. While these tools have proven to be invaluable in the extraction of parton densities, the volume of experimental data made available by LHC collaborations for use in PDF fits is already stretching the capabilities of the typical fitting technology. A standard global PDF fit may now include thousands of hadronic data points for which predictions have to be computed thousands of times during the minimisation process. As a consequence, performing these predictions using the standard interpolating tools, $i.e.$ [APPLgrid]{} and [FastNLO]{}, starts to become prohibitively time-consuming. For this reason a high-performance tool tailored specifically to the requirements of PDF analysis becomes increasingly important. The [FastKernel]{} method was developed to address this problem in the context of the NNPDF global analyses [@Ball:2014uwa]. This method differs from the standard procedure à la [APPLgrid]{} or [FastNLO]{} in that it maximises the amount of information that is precomputed prior to fitting so as to minimise the amount of operations required during the fit. More specifically, the [FastKernel]{} method relies on the combination of precomputed hard cross sections with DGLAP evolution kernels into a single look-up table, here called a [FastKernel]{} ([FK]{}) table. In this way the prediction for a given hadronic observable can be obtained by performing a simple matrix product between the respective [FK]{} table and PDFs evaluated directly at the fitting scale. In this paper we present the [APFELgrid]{} package, a public implementation of the [FastKernel]{} method in which the hard partonic cross sections provided in an [APPLgrid]{} look-up table are combined with the DGLAP evolution kernels provided by the [ APFEL]{} package [@Bertone:2013vaa]. This paper proceeds as follows. In Sect. \[sec:FastKernel\] we present the technical details of the implementation of the [FastKernel]{} method. This is followed in Sect. \[sec:benchmark\] by a performance benchmark of the [APFELgrid]{} library and resulting [FK]{} tables. Finally, in Sect. \[sec:conclusion\] we summarise the results discussed in this work. Interpolation tools for collider observables {#sec:FastKernel} ============================================ Hadron collider observables are typically computed in QCD by means of a double convolution of parton densities with a hard scattering cross section. Consider for example the calculation of a general cross section $pp\to X$ with a set of PDFs $\{f\}$: $$\label{eq:hadconv} \sigma_{pp\to X} = \sum_{s}\sum_{p} \int dx_1\,dx_2\, \hat{\sigma}^{(p)(s)}\,\alpha_s^{p+p_{\rm LO}}(Q^2) \,F^{(s)}(x_1,x_2, Q^2)\,,$$ where $Q^2$ is the typical hard scale of the process, the index $s$ sums over the active partonic subprocesses in the calculation, $p$ sums over the perturbative orders used in the expansion, $p_{\rm LO}$ is the leading-order power of $\alpha_s$ for the process and $\hat{\sigma}^{(p)(s)}$ is the N$^p$LO contribution to the cross section for the partonic subprocess scattering $(s)\to X$. $F^{(s)}$ represents the subprocess parton density: $$\label{eq:APPLsubproc} F^{(s)}(x_1,x_2, Q^2) =\sum_{i,j} C^{(s)}_{ij} f_i(x_{1},Q^2)f_j(x_{2},Q^2)\,,$$ where the $C^{(s)}_{ij}$ matrix enumerates the combinations of PDFs contributing to the $s$-th subprocess. The central observation of tools such as [APPLgrid]{} and [FastNLO]{} is that the PDF and $\alpha_S$ dependence may be factorised out of the convolution via expansion over a set of interpolating functions, spanning $Q^2$ and the two values of parton-$x$. For example one may represent the subprocess PDFs and $\alpha_S$ in terms of Lagrange basis polynomials $\mathcal{I}_\tau(Q^2)$, $\mathcal{I}_\alpha(x_1)$ and $\mathcal{I}_\beta(x_2)$ as: $$\label{eq:interpolation} \begin{array}{c} \displaystyle \alpha_s^{p+p_{\rm LO}}(Q^2) \,F^{(s)}(x_1,x_2, Q^2) =\\ \\ \displaystyle \sum_{\alpha,\beta,\tau} \alpha_s^{p+p_{\rm LO}}(Q_\tau^2) \,F^{(s)}_{\alpha\beta , \tau} \,\mathcal{I}_\tau(Q^2)\,\mathcal{I}_\alpha(x_1) \,\mathcal{I}_\beta(x_2), \end{array}$$ where we use the shorthand $F^{(s)}_{\alpha\beta ,\tau} = F^{(s)}(x_\alpha, x_\beta,Q_\tau^2)$. Using these expressions in the double convolution of Eq. (\[eq:hadconv\]) one can finally obtain an expression for the desired cross section which depends upon the subprocess PDFs only through a simple product: $$\label{eq:applconv} \sigma_{pp\to X} = \sum_p \sum_{s}\sum_{\alpha,\beta,\tau} \alpha_s^{p+p_{\rm LO}}(Q^2_\tau)W_{\alpha\beta,\tau}^{(p)(s)} \, F_{\alpha\beta,\tau}^{(s)}\,,$$ where $$\label{eq:applgrid} W_{\alpha\beta,\tau}^{(p)(s)} = \mathcal{I}_\tau(Q^2)\int dx_1\,dx_2\, \hat{\sigma}^{(p)(s)}\,\mathcal{I}_\alpha(x_1) \,\mathcal{I}_\beta(x_2)\,,$$ consists of the convolution of the hard cross section with the interpolating polynomials. This information may be stored in a precomputed look-up table. The final expression for the cross section in Eq. (\[eq:applconv\]) is therefore a considerably simpler task to perform inside a fit than the direct evaluation of the double convolution. The FastKernel method --------------------- A number of tools ($e.g.$ [ APFEL]{}, [HOPPET]{} [@Salam:2008qg] and [QCDNUM]{} [@Botje:2010ay]) are available which perform PDF evolution via an analogous interpolation procedure. In such a way PDFs at a general scale $Q_\tau$ may be expressed as a product of PDFs at some initial fitting scale $Q_0$ and an *evolution operator* obtained by the solution of the DGLAP equation. $$\label{eq:fastPDFfinal_recalled} f_i(x_{\alpha},Q^2_\tau) = \sum_{k} \sum_\beta A^\tau_{\alpha\beta, ik}\; f_k(x_\beta,Q^2_0)\,,$$ where latin indices run over PDF flavour, greek indices run over points in an initial-scale interpolating $x$-grid and the evolution operator $A$ may be accessed directly in the [APFEL]{} package. Given this operator, we may replace the (general-scale) PDFs used in the subprocess parton density Eq. (\[eq:APPLsubproc\]) with their equivalent expressions evaluated at the fitting scale as $$\label{eq:FK1} \begin{array}{rcl} F^{(s)}_{\alpha\beta,\tau} &=& \displaystyle \sum_{i,j} \sum_{k,l} \sum_{\delta,\gamma} C^{(s)}_{ij} \left[ A^\tau_{\alpha\delta ik}\; f_k(x_\delta,Q^2_0) A^\tau_{\beta\gamma jl}\; f_l(x_\gamma,Q^2_0) \right]\;\;\; \\ \\ &=& \displaystyle \sum_{k,l}\sum_{\delta,\gamma} \widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta} f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0)\,, \end{array}$$ with the object $$\widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta} = \sum_{i,j} C^{(s)}_{ij} A^\tau_{\alpha\delta ik} A^\tau_{\beta\gamma jl}\,,$$ combining the operations of subprocess density construction and PDF evolution. Going further and using the expression for subprocess parton densities in Eq. (\[eq:FK1\]) in the full cross section calculation we obtain $$\sigma_{pp\to X} = \sum_{k,l}\sum_{\delta,\gamma}\sum_p \sum_{s} \sum_{\alpha,\beta} \sum_{\tau} \alpha_s^{p+p_{\rm LO}}(Q^2_\tau)W_{\alpha\beta,\tau}^{(p)(s)} \widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta} f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0)\,.$$ Performing some further contractions it is possible to obtain an extremely compact expression for the calculation of the cross section in question, in terms of only the initial-scale PDFs and summing only over the initial scale interpolating $x$-grid and the incoming parton flavours: $$\label{eq:FK} \sigma _{pp\to X} = \sum_{k,l}\sum_{\delta,\gamma} \widetilde{W}_{kl,\delta\gamma} \,f_k(x_\delta,Q^2_0) f_l(x_\gamma,Q^2_0),$$ where the object: $$\label{eq:FKTable} \widetilde{W}_{kl,\delta\gamma} = \sum_p\sum_{s}\sum_{\alpha,\beta} \sum_{\tau} \alpha_s^{p+p_{\rm LO}}(Q^2_\tau) W_{\alpha\beta,\tau}^{(p)(s)} \widetilde{C}^{(s),\tau}_{kl,\alpha\beta\gamma\delta}$$ is referred to here as an [FK]{} table, and combines the information stored in [APPLgrid]{}-style interpolated weight grids with analogously interpolated DGLAP evolution operators. This combination enables for a maximally efficient expression for the calculation of observables at hadron colliders under PDF variation, without invoking any additional approximation. Features and limitations of [FK]{} tables ----------------------------------------- The [FK]{} product of Eq. (\[eq:FK\]) differs with respect to the product in Eq. (\[eq:applconv\]) in several notable ways. Firstly the typical [APPLgrid]{} or [FastNLO]{} products use as input PDFs at a general scale, requiring that PDF evolution [*e.g.*]{} Eq. (\[eq:fastPDFfinal\_recalled\]) be performed for every variation of the PDFs during the fit. In the [FK]{} product this evolution is pre-cached at the stage of [FK]{} table generation, requiring only initial-scale PDFs at the time of fitting. This pre-caching of the evolution also removes the need to sum over hard scale and perturbative order during the fit, further reducing the number of operations required. As the [FK]{} product acts directly at the fitting scale, it benefits from the typically reduced number of active partonic modes, with the sum over flavours in Eq. (\[eq:FK\]) being limited to those directly parametrised in the fit. Having reduced the calculation to such a simple form, it is also straightforward to apply standard computational tools such as multi-threading through e.g OpenMP or Single Instruction Multiple Data (SIMD) operations such as SSE or AVX to further reduce computational expense. While these features provide significant performance enhancements, the [FK]{} table format is not suitable as a complete replacement for tools such as [APPLgrid]{}. The precomputation of the PDF evolution necessarily means that all theory parameters such as perturbative order, strong coupling and factorization/renormalization scales are inextricably embedded in each [FK]{} table. In order to perform PDF fits including variations of these parameters, multiple [FK]{} tables must be computed, each with different theory settings. While performing such a re-calculation directly from Monte Carlo codes would be exceptionally time consuming, the data representation in [ APPLgrid]{} files allows for an efficient (re)-combination with varying theory parameters. Performance benchmarks {#sec:benchmark} ====================== We shall now examine the performance differences between the two expressions for fast interpolated cross section prediction Eq. (\[eq:applconv\]) ([APPLgrid]{}) and Eq. (\[eq:FK\]) ([ FK]{}). In order to provide a comprehensive benchmark, we consider here a wide range of processes including LHC and Tevatron electroweak vector boson production measurements [@Aaij:2012mda; @Aaij:2012vn; @Chatrchyan:2013mza; @Chatrchyan:2013uja; @Chatrchyan:2012xt; @Aad:2013iua; @Aad:2011fp; @Aad:2011dm; @Aaltonen:2010zza], $t\bar{t}$ total cross sections [@ATLAS:2012aa; @Chatrchyan:2013faa; @Chatrchyan:2012bra; @Chatrchyan:2012ria], double-differential Drell-Yan cross sections [@Chatrchyan:2013tia; @CMS:2014jea] and inclusive jet data [@Chatrchyan:2012bja; @Aad:2011fc; @Aad:2013lpa; @Abazov:2007jy]. Predictions are performed over a wide range of kinematics, for a total of 52 source [APPLgrid]{} files corresponding to the majority of available LHC and Tevatron datasets applicable to PDF determination. While the source [APPLgrid]{} files have varying grid densities in $x$ and $Q^2$, for the purposes of comparison the corresponding [ FK]{} tables are produced consistently with 30 points in $x$, and at an initial scale below the charm threshold, therefore with seven active partonic species. These settings are chosen so as to provide a realistic comparison, in a production environment the density and distribution of the $x$-grid may be adjusted to match interpolation accuracy requirements. For these comparisons the [FK]{} table is stored as double-precision in memory for table generation and in single-precision for the purposes of computing the [FK]{} product. In Fig. \[fig:timings\] we compare the average time taken per datapoint for the [FK]{} and [APPLgrid]{} calculations, for all of the 52 tables. We show timings for the [FK]{} calculation in four different configurations: AVX-OpenMP 2x (2 CPU cores), AVX, SSE3 and the standard double precision product. Due to the inherent structural differences between the [FK]{} and [APPLgrid]{} procedures, results from the [FK]{} calculation are systematically faster than those from [APPLgrid]{}. In particular, when comparing [ FK]{} AVX-OpenMP 2x to [APPLgrid]{} timings we obtain minimally a factor of ten improvement in speed for electroweak vector boson production and a maximum factor of 2000 improvement in predictions for inclusive jet data. Across all processes and kinematic regions we observe significant performance improvements from using the [FK]{} calculation even without the use of SIMD or multi-threading. While sheer computational speed is typically the primary consideration when computing observables in a PDF fit, other factors such as table size in the filesystem and memory, along with the computational cost of pre-computing [FK]{} tables must be considered. Indeed, the computation of the [FK]{} table in Eq. (\[eq:FKTable\]) requires a great deal of operations which can be time consuming, particularly in the case of source [APPLgrid]{} files with very high interpolation precision. In Fig. \[fig:performance\] we examine the [FK]{} table generation time with [APFELgrid]{}, [FK]{} table file size and memory usage of the [FK]{} tables arising from the same source [APPLgrid]{} files as discussed in Fig. \[fig:timings\]. When examining the table generation time per point, we observe timings from a few milliseconds to 3.5 minutes per point, with differences arising from the varying grid densities used in the input [APPLgrid]{} files. In terms of the grid size on disk, [FK]{} tables are typically larger than their corresponding [APPLgrid]{} files, primarily as the [ FK]{} file format is encoded in plain text for compatibility whereas [APPLgrid]{} files are expressed in binary as [ROOT]{} files. However when measuring the in-memory resident set size used by the two procedures, the amount of system memory used by [ FK]{} tables is systematically less than [APPLgrid]{} files for all processes considered here by at least 75%. Note that this effect is in part due to the differing precisions of the default representations. ![Performance comparisons between [FK]{} with AVX-OpenMP 2x, AVX, SSE3, double precision convolution and [ APPLgrid]{} convolution time per point and process.[]{data-label="fig:timings"}](plots/t0a) ![ generation time per point and comparison to [APPLgrid]{} for grid size on disk and resident set size (RSS).[]{data-label="fig:performance"}](plots/t0b) Conclusion {#sec:conclusion} ========== In this work we have demonstrated that by employing the so-called [ FastKernel]{} method, it is possible to convert an [APPLgrid]{} weight table into a derived format, referred to as an [FK]{} table, including the effects of PDF and $\alpha_s$ evolution. This procedure has been implemented in the [APFELgrid]{} package, supplied as a set of [ C++]{} routines designed to supplement the PDF evolution library [ APFEL]{} with [FK]{} table generation capabilities. The [APFELgrid]{} package allows one to obtain a computationally efficient expression for the calculation of hadronic cross sections, in terms of only the initial-scale PDFs and summing only over the initial scale interpolating $x$-grid and the incoming parton flavours. The simple structure of the resulting product makes [FK]{} tables particularly suitable for the efficient use of computational tools such as [SIMD]{} and [ OpenMP]{}. We have shown that in several practical examples the numerical evaluation of an [FK]{} product is considerably faster than the corresponding [APPLgrid]{} product, even in the case where neither SIMD or multi-threading are applied. [FK]{} tables are supplied in a simple plain-text format in order to simplify the construction of user interfaces, and therefore are typically larger than corresponding [APPLgrids]{}. However we have shown that the in-memory resident set sizes occupied by [FK]{} tables are typically smaller than those required by [APPLgrids]{}, in our examples by at least 75%. The substantial speed improvement of [FK]{} tables with respect to [APPLgrid]{} in association with the reduction in memory footprint makes the [APFELgrid]{} code a valuable tool for modern PDF fits including large collider datasets. The [APFELgrid]{} package and associated documentation are publicly available from the webpage: <https://github.com/nhartland/APFELgrid> Acknowledgements {#acknowledgements .unnumbered} ================ The authors would like to thank members of the NNPDF Collaboration for their support and motivation for this work; particularly Juan Rojo, Luigi del Debbio and Alberto Guffanti. We would also like to thank Mark Sutton for helpful comments on the paper. V. B. and N. H. are supported by an European Research Council Starting Grant “PDF4BSM”. S. C. is supported by the HICCUP ERC Consolidator grant (614577).

ArXiv

--- abstract: 'Given two schemes $S$ and $S''$, we prove that every equivalence between ${{\operatorname{\textbf{\textup{Sch}}}}}_S$ and ${{\operatorname{\textbf{\textup{Sch}}}}}_{S''}$ comes from a unique isomorphism between $S$ and $S''$. This eliminates all Noetherian and finite type hypotheses from a result of Mochizuki [@Moch] and fully answers a programme set out by Brandenburg in a series of questions on MathOverflow in 2011 [@Brand1; @Brand2; @Brand3; @Brand4].' bibliography: - 'Automorphisms.bib' --- =1 ------------------------------------------------------------------------ Introduction {#Sec intro .unnumbered} ============ Let $\mathscr C$ and $\mathscr D$ be categories, and write ${\operatorname{\textbf{\textup{Isom}}}}(\mathscr C,\mathscr D)$ for the category whose objects are equivalences $F \colon \mathscr C \to \mathscr D$ and whose morphisms $\eta \colon F \to G$ are natural isomorphisms. Let ${\operatorname{Isom}}(\mathscr C,\mathscr D) = \pi_0({\operatorname{\textbf{\textup{Isom}}}}(\mathscr C,\mathscr D))$ be its set of isomorphism classes, i.e. the set of equivalences $F \colon \mathscr C \to \mathscr D$ up to natural isomorphism. We will study ${\operatorname{\textbf{\textup{Isom}}}}(\mathscr C,\mathscr D)$ when $\mathscr C$ and $\mathscr D$ are categories of schemes. The main result is the following. \[Thm main\] Let $S$ and $S'$ be schemes. Then the natural functor $${\operatorname{Isom}}(S,S') \to {\operatorname{\textbf{\textup{Isom}}}}({{\operatorname{\textbf{\textup{Sch}}}}}_{S'},{{\operatorname{\textbf{\textup{Sch}}}}}_S)$$ is an equivalence (where ${\operatorname{Isom}}(S,S')$ is a discrete category). A similar result for the category of locally Noetherian schemes with finite type morphisms was proven by Mochizuki [@Moch Thm. 1.7]. Our result completely eliminates all Noetherian and finite type hypotheses. Because we do not have access to the same finite type techniques, the proof is almost entirely independent from Mochizuki’s. A special case of interest is the case $S = S' = {\operatorname{Spec}}{\mathbf Z}$. This gives a positive answer to a question by Brandenburg [@Brand1]: \[Cor MO\] Let $F \colon {{\operatorname{\textbf{\textup{Sch}}}}}\to {{\operatorname{\textbf{\textup{Sch}}}}}$ be an equivalence. Then $F$ is isomorphic to the identity functor. While this paper was in preparation, was obtained independently by Pohl [@Pohl]. We also get a version for commutative rings: \[Thm rings\] Let $R$ and $R'$ be commutative rings. Then the natural functor $${\operatorname{Isom}}(R,R') \to {\operatorname{\textbf{\textup{Isom}}}}({{\operatorname{\textbf{\textup{Alg}}}}}_R,{{\operatorname{\textbf{\textup{Alg}}}}}_{R'})$$ is an equivalence. A first statement of this type was proven by Clark and Bergman in the case where $R = R'$ is a (commutative, unital) integral domain [@CB Thm. 5.5]. Their result is only on the level of $\pi_0$ (equivalences up to natural isomorphism), and does not address whether the ${\operatorname{\textbf{\textup{Isom}}}}$ category is a setoid. It also deals with categories of algebras that are not necessarily commutative or unital, under the same assumptions on $R = R'$. In the case of non-commutative algebras, one also gets the $(-){^{\operatorname{op}}}$ autoequivalence \[*loc. cit.*\]. We imagine that may have been known to experts, although we do not know a reference. The analogue of for commutative (unital) rings follows from [@CB]; see also the answers to the post [@Bel] for a number of easy alternative proofs. Strategy of proof {#strategy-of-proof .unnumbered} ----------------- The strategy of the proof of is to characterise certain properties of schemes and morphisms of schemes by purely categorical means. The general framework is that of *categorical reconstruction*, in the sense of below. Although this technique has been used before in many different contexts, as far as we know there is no systematic treatment in the literature. In , we recall the definitions and prove some basic lemmas, in particular in the setting of slice categories. We then specialise to slice categories of schemes. The underlying set of a scheme is easily found as the isomorphism classes of simple subobjects (). To find the topology (), we first relate locally closed immersions to regular monomorphisms (). Then we characterise spectra of valuation rings (), which with some work recovers the topology (). On the other hand, a variant of a standard argument due to Beck [@BeckThesis Ex. 8] recovers the category ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_S)$ as cogroup objects in the category $S/{{\operatorname{\textbf{\textup{Sch}}}}}_S$ of $S$-schemes with a section (). The centre of ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_S)$ is $\Gamma(S,\mathcal O_S)$, and a sheafy version of this statement recovers the structure sheaf $\mathcal O_S$ (). This finishes the reconstruction of a scheme isomorphic to $S$ from the slice category ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. By , a sufficiently functorial version of this immediately implies the main theorem. The first proof we found of required an additional step in between and , namely the reconstruction of affine morphisms. Pohl’s argument for shows that this step is not needed, because cogroups in $S/{{\operatorname{\textbf{\textup{Sch}}}}}_S$ are nilpotent thickenings, hence automatically affine. The now omitted characterisation of affine morphisms gives some results of independent interest, which will appear separately. Acknowledgements {#acknowledgements .unnumbered} ---------------- [I am grateful to Gregor Pohl for the shortcut provided by his argument for , bypassing entirely the characterisation of affine morphisms. I thank Raymond Cheng for many helpful discussions, and Bhargav Bhatt for helpful suggestions and further questions. Finally, I want to thank Johan de Jong for his encouragement to eliminate all remaining Noetherian and finite type hypotheses, as well as for providing most of . ]{} Notation and conventions {#notation-and-conventions .unnumbered} ------------------------ Categories will be written bold, like the categories of sets (${{\operatorname{\textbf{\textup{Set}}}}}$), topological spaces (${{\operatorname{\textbf{\textup{Top}}}}}$), rings (${{\operatorname{\textbf{\textup{Ring}}}}}$), $R$-algebras (${{\operatorname{\textbf{\textup{Alg}}}}}_R$), schemes (${{\operatorname{\textbf{\textup{Sch}}}}}$), locally ringed spaces (${{\operatorname{\textbf{\textup{LRS}}}}}$), etcetera. All rings and algebras will be commutative and unital. Given a morphism of schemes $f \colon X \to Y$, write $f^\#$ for the morphism of sheaves $\mathcal O_Y \to f_*\mathcal O_X$ on $Y$. Given an object $X$ in a category $\mathscr C$, we write $\mathscr C/X$ for the slice category of objects over $X$, and $X/\mathscr C$ for the coslice category of objects under $X$. In the former case, we sometimes write $\mathscr C_{/X}$ or simply $\mathscr C_X$ if this causes no confusion. For example, ${{\operatorname{\textbf{\textup{Sch}}}}}_S$ is the category ${{\operatorname{\textbf{\textup{Sch}}}}}/S$ of schemes over a base scheme $S$. Note however that ${{\operatorname{\textbf{\textup{Alg}}}}}_R$ is by convention the *coslice* category $R/{{\operatorname{\textbf{\textup{Ring}}}}}$. To deal with set-theoretic issues arising in the formation of ${\operatorname{\textbf{\textup{Isom}}}}({{\operatorname{\textbf{\textup{Sch}}}}}_{S'}, {{\operatorname{\textbf{\textup{Sch}}}}}_S)$ and other category theoretic constructions, one should either work with universes, or use ‘sufficiently large’ small categories of schemes, cf. e.g. [@Stacks Tag [000J](https://stacks.math.columbia.edu/tag/000J)]. Functorial reconstruction {#Sec categorical} ========================= The strategy of the proof of is to show that certain properties of schemes are *categorical*, in the sense of below. The mathematical foundations needed to define categorical reconstruction lie in the intersection of model theory and category theory. The notion appears to originate in [@CB §2], but we don’t know a systematic study in the literature, and it is not always possible to give detailed references. We will illustrate the definitions with examples from algebra and topology. \[Def categorical\] Let $\mathscr C$ be a category. Then a property $\mathcal P$ of objects $X \in \mathscr C$ or of morphisms $f \colon X \to Y$ in $\mathcal C$ is *categorical* if it is definable in terms of morphisms, compositions, and equality of morphisms. This in particular implies that if $F \colon \mathscr C \to \mathscr D$ is an equivalence, then $\mathcal P(X) {\Leftrightarrow}\mathcal P(FX)$ for an object $X$ in $\mathscr C$ (resp. $\mathcal P(f) {\Leftrightarrow}\mathcal P(Ff)$ for a morphism $f$ in $\mathscr C$). However, our language does not have a predicate for *equality* (rather than isomorphism) of objects. For example, the property that an object is the only one in its isomorphism class should not be a categorical one, as it is not stable under equivalence of categories. \[Def reconstructed\] Let $F \colon \mathscr C \to \mathscr D$ be a functor to a concretely definable category $\mathscr D$. Then a *categorical reconstruction* of $F$ is a pair $(F', \eta)$ consisting of a functor $F' \colon \mathscr C \to \mathscr D$ that is definable in terms of morphisms of $\mathscr C$ satisfying categorical properties, together with a natural isomorphism $\eta \colon F \to F'$. Here, a *concretely definable* category $\mathscr D$ (used loosely) means a category of some collection of sets satisfying some relations[^1]. For example, ${{\operatorname{\textbf{\textup{Sch}}}}}$ is concretely definable, because a scheme is a topological space with a sheaf of rings on it. However, a category abstractly equivalent to ${{\operatorname{\textbf{\textup{Sch}}}}}$ need not be concretely definable. Note that the roles of $\mathscr C$ and $\mathscr D$ are rather asymmetric in the definition of categorical reconstruction. The classical case is $\mathscr D = {{\operatorname{\textbf{\textup{Set}}}}}$: The forgetful functor ${{\operatorname{\textbf{\textup{Top}}}}}\to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically. Indeed, if $X$ is a topological space, then points of $X$ are in bijection with ${\operatorname{Mor}}(*,X)$, where $*$ is any terminal object. \[Lem auto-equivalence\] Let $F \colon \mathscr C \to \mathscr D$ be a functor that can be reconstructed categorically. If $G \colon \mathscr C \to \mathscr C$ is an auto-equivalence, then $FG \cong F$. For any $X \in \mathscr C$, the auto-equivalence $G$ takes the data used to define $F(X)$ to the data used to define $F(G(X))$. -It turns out to be very powerful to reconstruct the identity functor $\mathscr C \to \mathscr C$ of a concretely definable category: If the identity functor ${\operatorname{id}}\colon \mathscr C \to \mathscr C$ of a concretely definable category can be reconstructed categorically, then ${\operatorname{Aut}}(\mathscr C) = 1$. For example, by the following lemma we get ${\operatorname{Aut}}({{\operatorname{\textbf{\textup{Top}}}}}) = 1$. The identity ${{\operatorname{\textbf{\textup{Top}}}}}\to {{\operatorname{\textbf{\textup{Top}}}}}$ can be reconstructed categorically. The property that $X \in {{\operatorname{\textbf{\textup{Top}}}}}$ is (isomorphic to) the Sierpiński space $S$ is categorical: it is the unique two-point space for which the swap is not continuous, i.e. $\#{\operatorname{Mor}}(*,S) = 2$ and ${\operatorname{Aut}}(S) = 1$. Then the functor ${{\operatorname{\textbf{\textup{Top}}}}}{^{\operatorname{op}}}\to {{\operatorname{\textbf{\textup{Set}}}}}$ given by $(X,\mathcal T) \mapsto \mathcal T$ can also be reconstructed categorically, as it is corepresented by $S$. The open point $\eta \in S$ is characterised as the unique point such that the inclusion $$\begin{aligned} \iota_X \colon {\operatorname{Mor}}(X,S) &\hookrightarrow \mathcal P({\operatorname{Mor}}(*,X))\\ f &\mapsto \{g \in {\operatorname{Mor}}(*,X)\ |\ fg = \eta\}\end{aligned}$$ endows ${\operatorname{Mor}}(*,X)$ with a topology for all $X$ (equivalently, the system of sets $\iota_X(f)$ for $f \in {\operatorname{Mor}}(X,S)$ is closed under arbitrary unions). The functorial association $$X \mapsto \Big({\operatorname{Mor}}(*,X),\iota_X({\operatorname{Mor}}(X,S))\Big)$$ recovers a topological space naturally homeomorphic to $X$. -The categorical reconstruction of the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Sch}}}}}$ follows a similar strategy. We start by reconstructing the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ (), then upgrade this to ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Top}}}}}$ (), and finally we reconstruct ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Sch}}}}}$ (). There are many examples of categorical reconstruction theorems in the literature; a sample of well-known results includes the Neukirch–Uchida theorem [@Neu1; @Neu2; @Uch1; @Uch2], the Gabriel–Rosenberg theorem [@Gabriel; @Ros; @BrandRosenberg], the Bondal–Orlov theorem [@BonOrl], and Mochizuki’s results in anabelian geometry [@MochAnab1], [@MochAnab2] and for Noetherian schemes and log schemes [@Moch]. The most general setup is as follows. Let $\mathscr C$ be a finitely complete category, and let $\mathscr S \to \mathscr C$ be a fibred category. If $\mathscr S_X$ and $\mathscr S_Y$ are equivalent for $X, Y \in \mathscr C$, is it true that $X \cong Y$? In general it is too much to expect that the natural map $${\operatorname{Isom}}(X,Y) \to {\operatorname{Isom}}(\mathscr S_Y,\mathscr S_X)$$ is an isomorphism. For example, although Gabriel [@Gabriel] (resp. Rosenberg [@Ros], [@BrandRosenberg]) reconstruct a Noetherian scheme (resp. quasi-separated scheme) $X$ from its category ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$, the latter can have extra endomorphisms not coming from $X$: if $\mathscr L$ is a nontrivial line bundle, then $-\otimes \mathscr L$ is an auto-equivalence of ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$ that does not come from an automorphism of $X$ (e.g. since it does not fix $\mathcal O_X$). Similarly, Bondal and Orlov [@BonOrl] reconstruct a smooth projective variety $X$ with ample canonical or anti-canonical bundle from its derived category $D(X) = D^b_{{{\operatorname{\textbf{\textup{Coh}}}}}}(X)$, but the derived category has a shift $[1] \colon D(X) \to D(X)$ that is not induced by an automorphism of $X$. In fact, they prove [@BonOrl Thm. 3.1] that ${\operatorname{Aut}}(D(X)) \cong {\operatorname{Aut}}(X) \ltimes ({\operatorname{Pic}}(X) \oplus {\mathbf Z})$, where ${\mathbf Z}$ acts by shifting. However, for slice categories $\mathscr C_{/X}$ we have the following result, which is often implicitly reproved in applications. For example, this argument applies to the Noetherian version [@Moch], as well as some anabelian situations (taking slice categories ${{\operatorname{\textbf{\textup{F\'Et}}}}}_{/X}$ in the category of finite étale morphisms of schemes). \[Lem categorical reconstruction equivalence on ISOM\] Let $\mathscr C$ be a finitely complete category. Assume that the forgetful functors $F_X \colon \mathscr C_{/X} \to \mathscr C$ can be reconstructed categorically from $\mathscr C_{/X}$, by a formula that does not depend on $X$. Then the natural functors $$\begin{aligned} \Psi \colon {\operatorname{Isom}}(X,Y) &\to {\operatorname{\textbf{\textup{Isom}}}}(\mathscr C_{/Y},\mathscr C_{/X})\\ f &\mapsto f^*\end{aligned}$$ for $X, Y \in \mathscr C$ are equivalences of categories (where the left hand side is viewed as a discrete category). That is, ${\operatorname{\textbf{\textup{Isom}}}}(\mathscr C_{/Y},\mathscr C_{/X})$ is a setoid whose isomorphism classes are ${\operatorname{Isom}}(X,Y)$. For simplicity, assume the categorical reconstructions of the forgetful functors are *equal* (rather than isomorphic) to $F_X$ and $F_Y$; this does not affect the argument. If $F \colon \mathscr C_{/Y} \to \mathscr C_{/X}$ is an equivalence, then there exists a natural isomorphism $\eta \colon F_XF \stackrel\sim\to F_Y$. For any morphism $B \to Y$, let its image under $F$ be $A \to X$. Then naturality of $\eta$ gives a commutative diagram $$\begin{tikzcd}\label{Dia essentially surjective} A \ar{r}{\sim}[swap]{\eta_B}\ar{d} & B \ar{d}\\ X \ar{r}{\sim}[swap]{\eta_Y} & Y{\makebox[0pt][l]{\,,}} \end{tikzcd}$$ inducing functorial isomorphisms $\beta_B \colon A \stackrel\sim\to B \times_Y X$. Setting $f = \eta_Y \colon X \stackrel\sim\to Y$, we conclude that $\beta$ gives a natural isomorphism $\beta \colon F \stackrel\sim\to f^*$, showing that $\Psi$ is essentially surjective. For fully faithfulness, it suffices to show that ${\operatorname{Aut}}({\operatorname{id}}_{\mathscr C_{/X}}) = 1$ for all $X \in \mathscr C$. Let $\alpha \colon {\operatorname{id}}_{\mathscr C_{/X}} \stackrel\sim\to {\operatorname{id}}_{\mathscr C_{/X}}$ be an automorphism. For every morphism $f \colon A \to B$ in $\mathscr C_{/X}$, naturality of $\alpha$ gives a commutative diagram $$\begin{tikzcd} A \ar{d}[swap]{f}\ar{r}{\sim}[swap]{\alpha_A} & A\ar{d}{f}\\ B \ar{d}\ar{r}{\sim}[swap]{\alpha_B} & B\ar{d}\\ X \ar{r}{\sim}[swap]{\alpha_X} & X{\makebox[0pt][l]{\,.}} \end{tikzcd}$$ Applying this to all morphisms used in the categorical reconstruction of $F_X$, we conclude that $F_X(\alpha) = {\operatorname{id}}_{F_X}$. Since $F_X$ is faithful, this implies that $\alpha = {\operatorname{id}}$. The forgetful functor ${{\operatorname{\textbf{\textup{Set}}}}}_{/X} \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically: for a set $A$ over $X$, the points of $A$ correspond to (isomorphism classes of) simple subobjects of $A$. The proof of then reads as follows. - If $F \colon {{\operatorname{\textbf{\textup{Set}}}}}_{/Y} \to {{\operatorname{\textbf{\textup{Set}}}}}_{/X}$ is an equivalence, then looking at the simple subobjects of the terminal objects gives an isomorphism $\eta_Y \colon X \to Y$, and diagram (\[Dia essentially surjective\]) shows that $F \cong \eta_Y^*$. - If $\alpha \colon {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Set}}}}}_{/X}} \to {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Set}}}}}_{/X}}$ is an automorphism, then for any $A \in {{\operatorname{\textbf{\textup{Set}}}}}_{/X}$ and any simple subobject $a \to A$, naturality of $\alpha$ gives the diagram $$\begin{tikzcd} a \ar{d}\ar{r}{\sim}[swap]{\alpha_a} & a\ar{d}\\ A \ar{r}{\sim}[swap]{\alpha_A} & A{\makebox[0pt][l]{\,,}} \end{tikzcd}$$ showing that $\alpha_A$ fixes all points of $A$, i.e. $F_X(\alpha_A)$ agrees with ${\operatorname{id}}_{F_X(A)}$. To prove , it therefore suffices to show that the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Sch}}}}}$ can be reconstructed categorically from ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. Underlying set {#Sec set} ============== We begin by recovering the underlying set of an $S$-scheme $X$ from categorical information in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$; see below. \[Lem field\] The simple objects in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$ are the spectra of fields. Here, a *simple* objects is a nonempty object $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ whose only subobjects are $\varnothing$ and $X$. \[Lem points\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$. If ${\operatorname{Spec}}k \to X$ is a monomorphism from the spectrum of a field, then there exists a unique point $x \in X$ and a unique isomorphism ${\operatorname{Spec}}k \cong {\operatorname{Spec}}\kappa(x)$ over $X$. \[Lem connected\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$. Then $X$ is connected if and only if $X$ cannot be written as a coproduct of two nonempty $S$-schemes. To summarise the results of this section: \[Cor basic properties\] The following properties on objects $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ are categorical: 1. $X$ is the spectrum of a field;\[Cat field\] 2. $X$ is connected.\[Cat connected\] Moreover, the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically. Statements \[Cat field\] and \[Cat connected\] follow from and . For the final statement, by we may take the functor that takes $X$ to the set of (isomorphism classes of) simple subobjects of $X$. Topology {#Sec topology} ======== Having reconstructed the underlying set $|X|$ of an $S$-scheme $X$, we turn next to the topology on $|X|$. We first find locally closed immersions (), then describe spectra of valuation rings (), and finally recover the topology (). \[Lem immersion\] Let $f \colon X \to Y$ be a morphism in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. Then $f$ is a (locally closed) immersion if and only if $f$ can be written as a composition of two regular monomorphisms. Thus, the property that $f$ is an immersion is categorical. If $f$ is the equaliser of $g_{1,2} \colon Y \rightrightarrows Z$, then the diagram $$\begin{tikzcd}[column sep=2.8em] X \ar{d}[swap]{f}\ar{r} & Z \ar{d}{\Delta_Z}\\ Y\ar{r}{(g_1,g_2)} & Z \times Z \end{tikzcd}$$ is a pullback. Hence, $f$ is an immersion since $\Delta_Z$ is [@Stacks Tag [01KJ](https://stacks.math.columbia.edu/tag/01KJ)]. Thus, any regular monomorphism is an immersion, so the same goes for a composition of two regular monomorphisms [@Stacks Tag [02V0](https://stacks.math.columbia.edu/tag/02V0)]. Conversely, every immersion $f \colon X \to Y$ factors as $X \to U \to Y$, with $X \to U$ closed and $U \to Y$ open. Thus, it suffices to show that if $f$ is either an open immersion or a closed immersion, then $f$ is a regular monomorphism. If $f$ is an open immersion, then let $Z = Y \amalg_X Y$ be two copies of $Y$ glued along $X$, and let $g_{1,2} \colon Y \rightrightarrows Z$ be the two inclusions. Then $X$ is the equaliser of $Y \rightrightarrows Z$. If $f$ is a closed immersion, then let $Z = Y \amalg_X Y$ (which exists and is for example described in [@Stacks Tag [0B7M](https://stacks.math.columbia.edu/tag/0B7M)]). Once again, $X$ is the equaliser of $Y \rightrightarrows Z$. This proves the first statement, and the second is immediate. The proof actually shows that we may replace ‘composition of two regular monomorphisms’ by ‘composition of two effective monomorphisms’. One can combine the arguments in the open and closed cases to show that every immersion that factors as an open immersion inside a closed immersion is also a regular monomorphism [@Brand4] (but this no longer gives an effective monomorphism). As far as we know, the classification of regular monomorphisms in ${{\operatorname{\textbf{\textup{Sch}}}}}$ is still open. For example, it does not appear to be known whether every regular monomorphism factors as open inside closed; see [@Brand4]. \[Lem reduced\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$. Then $X$ is reduced if and only if every immersion $Z \to X$ that induces a bijection on points $|Z| \to |X|$ is an isomorphism. Thus, the property that $X$ is reduced is categorical. An immersion whose underlying set map is closed is a closed immersion [@Stacks Tag [01IQ](https://stacks.math.columbia.edu/tag/01IQ)]. This proves the first statement, and the second is immediate because immersions are categorical () and the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically (). A different characterisation was given by Moret-Bailly’s partial answer [@MBAns] to [@Brand1]: an object $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ is reduced if and only if the natural map $\coprod_{x \in X} {\operatorname{Spec}}\kappa(x) \to X$ is an epimorphism. Although this criterion is arguably more elementary, we find more intuitive. \[Prop valuation\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$, and let $x \in X$ be a point. Then $(X,x)$ is isomorphic to $({\operatorname{Spec}}R,\mathfrak m)$ for a valuation ring $R$ with maximal ideal $\mathfrak m$ if and only if all of the following hold: 1. $X$ is reduced and connected;\[Item reduced and connected\] 2. the category of immersions $Z \to X$ containing $x$ is a linear order;\[Item totally ordered\] 3. there exists a set $V \subseteq |X|$ that is the support of infinitely many immersions $Z \to X$ containing $x$.\[Item support\] In particular, the property that $(X,x) \cong ({\operatorname{Spec}}R,\mathfrak m)$ for a valuation ring $R$ with maximal ideal $\mathfrak m$ is categorical. If $(X,x) \cong ({\operatorname{Spec}}R,\mathfrak m)$ with $R$ a valuation ring, then $X$ is reduced and connected since $R$ is a domain. Moreover, the only open subset $U \subseteq X$ containing $x$ is $X$, hence every immersion $Z \to X$ containing $x$ is closed. Then the immersions containing $x$ are linearly ordered because the ideals of $R$ are linearly ordered. Finally, if $r \in R$ is a non-unit, then the ideals $(r) \supsetneq (r^2) \supsetneq \ldots$ all have the same underlying closed set $V \subseteq |X|$. Thus, \[Item reduced and connected\], \[Item totally ordered\], and \[Item support\] hold if $(X,x) \cong ({\operatorname{Spec}}R,\mathfrak m)$ for a valuation ring $R$ with maximal ideal $\mathfrak m$. Conversely, suppose $(X,x)$ satisfies \[Item reduced and connected\], \[Item totally ordered\], and \[Item support\]. Firstly, we show that every point specialises to $x$. For a point $y \in X$, write $V(y)$ for the closure of $y$, and $D(y)$ for its complement. First assume $y \in X$ neither specialises nor generalises to $x$, so $x \in D(y)$ and $y \in D(x)$. Define the locally closed sets $U = D(y)$ and $V = V(x) \cup V(y)$, both of which contain $x$. We have $y \in V$ and $y \not\in U$, hence $V \nsubseteq U$, so $U \subseteq V$ by \[Item totally ordered\]. This means that $D(y) \subseteq V(x)$, hence $D(x) \cap D(y) = \varnothing$. On the other hand, if $U = D(x) \triangle D(y)$ (the symmetric difference) and $V = V(x)$, then $U$ and $V$ both contain $x$, but $y$ is in $U$ and not in $V$. Hence, $U \nsubseteq V$, so \[Item totally ordered\] implies $V \subseteq U$. This means that $V(x) \subseteq D(y)$, hence $D(x) \cup D(y) = X$. Thus, $X = D(x) \amalg D(y)$, contradicting connectedness of $X$. Thus, every $y \in X$ either specialises to $x$ or generalises to $x$. Suppose that there exist strict specialisations $y {\rightsquigarrow}x {\rightsquigarrow}z$. Then the locally closed sets $U = D(z)$ and $V = V(y)$ contain $x$. But we have $y \in U$ but $y \not\in V$, and $z \in V$ but $z \not\in U$. This contradicts \[Item totally ordered\], and we conclude that one of the following holds: - Every point $y \in X$ specialises to $x$; - Every point $y \in X$ generalises to $x$. In the first case, the only open containing $x$ is $X$, so every immersion $Z \to X$ containing $x$ is closed. Similarly, in the second case every immersion $Z \to X$ containing $x$ is open. But since $X$ is reduced, the latter implies there is a unique scheme structure on every immersion $Z \to X$ containing $x$, contradicting \[Item support\]. So we conclude that every point $y \in X$ specialises to $x$, and every immersion $Z \to X$ containing $x$ is closed. Condition \[Item totally ordered\] now also implies that for any $y,z \in X$, we have either $V(y) \subseteq V(z)$ or $V(z) \subseteq V(y)$, hence $z {\rightsquigarrow}y$ or $y {\rightsquigarrow}z$. In particular, $X$ has a unique generic point. Since $X$ is also assumed reduced, this implies that $X$ is integral. Since $X$ is the only open containing $x$, we conclude that $X$ is affine; say $X = {\operatorname{Spec}}R$ for some domain $R$. Finally, \[Item totally ordered\] implies that the ideals in $R$ are linearly ordered, so $R$ is a valuation ring or a field. The maximal ideal $\mathfrak m \subseteq R$ corresponds to $x$, and $R$ is not a field by \[Item support\]. This shows that \[Item reduced and connected\], \[Item totally ordered\], and \[Item support\] imply that $(X,x) \cong ({\operatorname{Spec}}R, \mathfrak m)$ for a valuation ring $R$ with maximal ideal $\mathfrak m$. The final statement follows as immersions are categorical (), reducedness and connectedness are categorical (, ), and the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically (). \[Lem specialisation\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$, and let $x,y \in X$. Then $x {\rightsquigarrow}y$ if and only if there exists a morphism $f \colon Z \to X$ where $Z$ is the spectrum of a valuation ring $R$ such that $f(\mathfrak m) = y$ and $x \in {\operatorname{im}}(f)$. In particular, the property that $x$ specialises to $y$ is categorical. Assume that such a morphism $f \colon Z \to X$ exists. Then $f^{-1}(V(x))$ is a nonempty closed subset of $Z$, hence contains $\mathfrak m$. This forces $y \in V(x)$, so $x {\rightsquigarrow}y$. Conversely, assume $x {\rightsquigarrow}y$. The case $x = y$ is trivial, so we may assume $x \neq y$. Let $U \subseteq X$ be an affine open neighbourhood of $y$. Since $U$ is stable under generalisation, it contains $x$. If $U = {\operatorname{Spec}}A$ and $x,y \in U$ correspond to primes $\mathfrak p, \mathfrak q \subseteq A$ respectively, then we get a map $A \to B = (A/\mathfrak p)_{\mathfrak q}$. Since $\mathfrak p \subsetneq \mathfrak q$, the local domain $B$ is not a field. Hence, there exists a valuation ring $R \subseteq {\operatorname{Frac}}B$ dominating $B$ [@Stacks Tag [00IA](https://stacks.math.columbia.edu/tag/00IA)]. Thus, the point $\mathfrak m_R \in {\operatorname{Spec}}R$ maps to $y \in X$, and $(0) \in {\operatorname{Spec}}R$ maps to $x \in X$. This proves the first statement, and the final statement follows from and . -It is tempting at this point to try to classify closed immersions as immersions () that are closed under specialisation (). However, this is not true in general; see . Instead, we use the following: \[Lem closed immersion\] Given a map $f \colon X \to Y$ in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$, the following are equivalent: 1. $f$ is a closed immersion;\[Item closed immersion\] 2. $f$ is an immersion, and for every $Y' \to Y$ and every closed point $y \in Y'$ not in the image of the base change $f' \colon X' \to Y'$, the map $X' \amalg {\operatorname{Spec}}\kappa(y) \to Y'$ is an immersion.\[Item disjoint union with closed point\] In particular, closed immersions are categorical. If $f$ is a closed immersion, then clearly $f$ is an immersion, and for every $Y' \to Y$, the base change $f' \colon X' \to Y'$ is a closed immersion. If $y \not\in {\operatorname{im}}(f')$ is a closed point, then $g \colon X' \amalg {\operatorname{Spec}}\kappa(y) \to Y'$ is in fact a *closed* immersion by the Chinese remainder theorem. This proves $\ref{Item closed immersion} {\Rightarrow}\ref{Item disjoint union with closed point}$. Conversely, assume \[Item disjoint union with closed point\] holds. Since $f$ is an immersion, it suffices to show that the image of $f$ is closed [@Stacks Tag [01IQ](https://stacks.math.columbia.edu/tag/01IQ)]. This can be done on a cover by affine opens, so let $Y' \subseteq Y$ be an affine open, with preimage $X' \subseteq X$. If $y \in Y'$ is a closed point not in $X'$, then the map $X' \amalg {\operatorname{Spec}}\kappa(y) \to Y'$ is an immersion by \[Item disjoint union with closed point\]. In particular, the subspace topology on $X' \cup \{y\}$ is the disjoint union of $X'$ and $\{y\}$, hence there exists an open $U \subseteq Y'$ such that $U \cap (X' \cup \{y\}) = \{y\}$. This means that $U$ is an open neighbourhood of $y$ with $U \cap X' = \varnothing$, so $y \not \in \overline{X'}$. Therefore the closed set $Z = \overline{X'}\setminus X' \subseteq Y'$ has no closed points, which implies $Z = \varnothing$ since $Y'$ is affine. Hence, $X'$ is closed in $Y'$, which proves $\ref{Item disjoint union with closed point} {\Rightarrow}\ref{Item closed immersion}$. The final statement follows since immersions are categorical (), the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically (), and closed points are categorical (). \[Prop topology\] The forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Top}}}}}$ can be reconstructed categorically. By , the forgetful functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Set}}}}}$ can be reconstructed categorically. Moreover, a subset $U \subseteq |X|$ is open if and only if $U^c$ is the support of a closed immersion. The result now follows from . \[Cor topological\] The following properties for a morphism $f \colon X \to Y$ in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$ are categorical: 1. $f$ is quasi-compact;\[Cat qc\] 2. $f$ is quasi-separated;\[Cat qs\] 3. $f$ is separated.\[Cat separated\] Statement \[Cat qc\] is immediate from . Statements \[Cat qs\] and \[Cat separated\] follow from \[Cat qc\] and respectively, since $\Delta_{X/Y} \colon X \to X\times_Y X$ is functorially associated to $f$. \[Lem open immersion\] Let $f \colon X \to Y$ be a morphism. Then $f$ is an open immersion if and only if there exists an open set $U \subseteq |Y|$ such that $f$ is terminal among morphisms $Z \to Y$ landing in $U$. In particular, open immersions are categorical. The first statement is clear; the second follows from . -If $X$ is Noetherian, then it suffices to replace the condition that $U \subseteq |Y|$ be open by the condition that $f$ is an immersion. This is not true for arbitrary schemes, as can be seen by the following example. \[Ex closed immersion\] Let $R$ be a ring admitting a pure ideal $I \subseteq R$ that is not finitely generated. For example, let $R = C^\infty({\mathbf R})$, and let $I$ be the functions vanishing in a neighbourhood of $0$ [@Stacks Tag [052H](https://stacks.math.columbia.edu/tag/052H)]. Then there exists a multiplicative set $S \subseteq R$ such that $R/I \cong R[S^{-1}]$ as $R$-algebras [@Stacks Tag [04PS](https://stacks.math.columbia.edu/tag/04PS)]. Then the closed subset $V(I) \subseteq {\operatorname{Spec}}R$ is closed and stable under generalisation, but not open [@Stacks Tags [04PU](https://stacks.math.columbia.edu/tag/04PU) and [05KK](https://stacks.math.columbia.edu/tag/05KK)]. Similarly, the complementary open $D(I)$ is open and closed under specialisation, but not closed. Thus, we cannot get the closed immersions directly from and as the immersions that are closed under specialisation. Finally, the map ${\operatorname{Spec}}R/I \to {\operatorname{Spec}}R$ is terminal for maps landing in $V(I)$, because of the isomorphism $R/I \cong R[S^{-1}]$ and the universal property of rings of fractions. This shows that in , we cannot replace the assumption that $U$ is open by the weaker assumption that $U$ is locally closed. Quasi-coherent sheaves {#Sec quasi-coherent} ====================== We use a variant of the classical cogroup argument due to Beck [@BeckThesis Ex. 8] to recover the category of quasi-coherent sheaves on $X$ from the category ${{\operatorname{\textbf{\textup{Sch}}}}}_X$. For $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$, write ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$ for the category of quasi-coherent $\mathcal O_X$-modules. Given an object $\mathscr F \in {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$, denote by ${\operatorname{Sym}}^{\leq 1} \mathscr F$ the quasi-coherent $\mathcal O_X$-algebra ${\operatorname{Sym}}^* \mathscr F/{\operatorname{Sym}}^2 \mathscr F \cong \mathcal O_X \oplus \mathscr F$, and by ${\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F)$ the $X$-scheme ${\operatorname{\textbf{\textup{Spec}}}}_X({\operatorname{Sym}}^{\leq 1} \mathscr F)$. It is the *nilpotent thickening of $X$ by $\mathscr F$*. The natural $\mathcal O_X$-algebra map ${\operatorname{Sym}}^{\leq 1} \mathscr F \twoheadrightarrow \mathcal O_X = {\operatorname{Sym}}^{\leq 0} \mathscr F$ is a retraction of $\mathcal O_X \to {\operatorname{Sym}}^{\leq 1} \mathscr F$, hence induces a section $\sigma \colon X \to {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F)$. This realises ${\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F)$ as an object of the coslice category $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$. \[Rmk pushout of schemes\] For $X \stackrel{\sigma_i}\to Y_i \to X$ in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$ (for $i \in \{1,2\}$), the pushout $Y_1 \amalg_X Y_2$ exists in ${{\operatorname{\textbf{\textup{Sch}}}}}_X$ when each $\sigma_i$ is a closed immersion [@Stacks Tag [0B7M](https://stacks.math.columbia.edu/tag/0B7M)]. Moreover, $|Y_1 \amalg_X Y_2|$ is the topological pushout $|Y_1| \amalg_{|X|} |Y_2|$, with the sheaf of rings given by $\mathcal O_{Y_1} \times_{\mathcal O_X} \mathcal O_{Y_2}$. In particular, if $\mathscr F, \mathscr G \in {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$, then ${\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F) \amalg_X {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr G)$ exists, and the explicit description gives $${\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F) \underset X\amalg {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr G) \cong {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F \oplus \mathscr G).$$ Therefore, the functor ${\operatorname{\textbf{\textup{Nil}}}}_X(-) \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X){^{\operatorname{op}}}\to X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$ preserves finite coproducts. In particular, addition $\mathscr F \oplus \mathscr F \to \mathscr F$ induces a comultiplication $$c \colon {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F) \to {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F) \underset X\amalg {\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F),$$ and similarly for inversion $\iota$, making $({\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F),c,\iota)$ into an abelian cogroup object in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$. Here a cogroup object in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$ means for the monoidal structure given by the pushout $- \amalg_X -$; in particular if $Y$ is a cogroup in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$ this means that $Y \amalg_X Y$ exists. Since existence and description of pushouts is subtle in general, we have to impose some mild additional conditions. \[Prop cogroup essentially surjective\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$, and let $(Y, c, \iota)$ be a cogroup in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$. Assume that the underlying set of $Y \amalg_X Y$ agrees with the pushout $|Y| \amalg_{|X|} |Y|$ of sets. Then $(Y,c,\iota) \cong ({\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F),c,\iota)$ for a unique $\mathscr F \in {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$. \[Rmk pushout separated\] The assumption on $Y \amalg_X Y$ is for example satisfied if $Y \to X$ is separated, or if formation of $Y \amalg_X Y$ commutes with base change of $X$. Indeed, in the former case, the section $\sigma \colon X \to Y$ is a closed immersion. Then the pushout $Y \amalg_X Y$ is described in above, and agrees with the pushout as locally ringed spaces (in particular as sets). In the latter case, it suffices to look at each fibre. But a section ${\operatorname{Spec}}k \to Y$ to a $k$-scheme is always closed, so we may proceed as in the first case. In fact, the first case follows from the second, because formation of $Y \amalg_X Y$ commutes with base change if $\sigma \colon X \to Y$ is closed, since the formation of $B \times_A B$ for a split surjection $B \twoheadrightarrow A$ commutes with tensor products. Because $Y \amalg_X Y$ is also the pushout as sets, we get a cogroup in $|X|/{{\operatorname{\textbf{\textup{Set}}}}}_{|X|}$. Thus for each $x \in X$ we get a cogroup in $*/{{\operatorname{\textbf{\textup{Set}}}}}$ (pointed sets). But these are always trivial: the existence of a two-sided counit $\varepsilon \colon Y \to *$ implies that the compositions $$\begin{tikzcd} Y \ar{r}{c} & Y \vee Y \ar[yshift=.25em]{r}{1 \vee \varepsilon}\ar[yshift=-.25em]{r}[swap]{\varepsilon \vee 1} & Y \end{tikzcd}$$ are the identity, where $\vee$ denotes the wedge sum. The equaliser of $1 \vee \varepsilon$ and $\varepsilon \vee 1$ is the point $*$, showing that $Y = *$. Applying this to all fibres $Y_x \to x$, we conclude that $Y \to X$ is a bijection, with section $\sigma$. Hence, $Y \to X$ is a homeomorphism, hence affine [@Stacks Tag [04DE](https://stacks.math.columbia.edu/tag/04DE)]; say $Y = {\operatorname{\textbf{\textup{Spec}}}}_X(\mathscr A)$ for some quasi-coherent sheaf of $\mathcal O_X$-algebras $\mathscr A$. Then $\mathscr A$ is an abelian group object in ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)/\mathcal O_X$, say with multiplication $m = c^\#$ and retraction $\pi = \sigma^\#$. Write $\mathscr I = \ker(\mathscr A \stackrel\pi\to \mathcal O_X)$, so $\mathscr A \cong \mathcal O_X \oplus \mathscr I$ as $\mathcal O_X$-modules. We have natural maps $$\mathscr A \to \mathscr A \underset{\mathcal O_X}\times \mathscr A \cong \mathcal O_X \oplus \mathscr I \oplus \mathscr I \vspace{-.25em}$$ given by $(\pi,{\operatorname{id}}_\mathscr A)$ and $({\operatorname{id}}_\mathscr A,\pi)$. The relations of abelian groups imply that $m \circ ({\operatorname{id}}_\mathscr A,\pi) = {\operatorname{id}}_\mathscr A = m \circ (\pi,{\operatorname{id}}_{\mathscr A})$, so $m$ is necessarily given by $$\begin{aligned} m \colon \mathcal O_X \oplus \mathscr I \oplus \mathscr I &\to \mathcal O_X \oplus \mathscr I\\ (a,b,c) &\mapsto (a,b+c).\end{aligned}$$ But $m$ is also an $\mathcal O_X$-algebra homomorphism. Thus, $$(0,f)\cdot(0,g) = m(0,f,0)\cdot m(0,0,g) = m((0,f,0)\cdot (0,0,g)) = 0,$$ which means that $\mathscr I^2 = 0$. Thus, $(\mathscr A,m)$ is isomorphic to $({\operatorname{Sym}}^{\leq 1} \mathscr I,m)$, where $m$ is the group structure induced by the addition $\mathscr I \oplus \mathscr I \to \mathscr I$. This shows that $(Y,c,\iota)$ is isomorphic to $({\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr I),c,\iota)$. Since $\mathscr I = \ker(\mathscr A \to \mathcal O_X)$ can be recovered from $Y$, the uniqueness statement follows. -Write ${{\operatorname{\textbf{\textup{coAb}}}}^{\operatorname{sep}}}(X/{{\operatorname{\textbf{\textup{Sch}}}}}_X)$ for the category of abelian cogroup objects $(Y,c,\iota)$ in $X/{{\operatorname{\textbf{\textup{Sch}}}}}_X$ such that the structure map $Y \to X$ is separated (see ). \[Thm Qcoh\] Let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$. Then the functor $$\begin{aligned} F \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X){^{\operatorname{op}}}&\to {{\operatorname{\textbf{\textup{coAb}}}}^{\operatorname{sep}}}(X/{{\operatorname{\textbf{\textup{Sch}}}}}_X)\\ \mathscr F &\mapsto ({\operatorname{\textbf{\textup{Nil}}}}_X(\mathscr F),c,\iota)\end{aligned}$$ is an equivalence of categories. Essential surjectivity is . If $\mathscr F, \mathscr G \in {{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)$, then an $\mathcal O_X$-algebra homomorphism $f \colon \mathcal O_X \oplus \mathscr F \to \mathcal O_X \oplus \mathscr G$ preserving the surjections to $\mathcal O_X$ necessarily maps $\mathscr F$ into $\mathscr G$, hence comes from a unique $\mathcal O_X$-module map $\mathscr F \to \mathscr G$. Moveover, any such $f$ automatically preserves the abelian cogroup structure, showing that $F$ is fully faithful. Structure sheaf {#Sec structure sheaf} =============== We modify the argument of Brandenburg’s second answer [@BraAns] to [@Bel] to categorically reconstruct the structure sheaf on $S$. The proof is based on the well-known formula $${\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X)}) \cong \Gamma(X,\mathcal O_X).$$ However, the left hand side does not naturally come with functorial restriction maps for morphisms $f \colon Y \to X$, for the same reason that sheaves $\mathscr F, \mathscr G$ on a topological space $X$ do not have restriction maps $${\operatorname{Hom}}(\mathscr F(U),\mathscr G(U)) \to {\operatorname{Hom}}(\mathscr F(V),\mathscr G(V))$$ for opens $V \subseteq U$. For sheaves, the solution is to work with ${\operatorname{Hom}}(\mathscr F|_U,\mathscr G|_U)$ instead of ${\operatorname{Hom}}(\mathscr F(U),\mathscr G(U))$, and our solution will be similar. \[Def QCOH\] If $S$ is a scheme, write ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/S}$ for the category whose objects are pairs $(X,\mathscr F)$, where $X$ is an $S$-scheme and $\mathscr F$ is a quasi-coherent sheaf on $X$. A morphism $(f,\phi) \colon (X,\mathscr F) \to (Y,\mathscr G)$ consists of a morphism $f \colon X \to Y$ of $S$-schemes and a morphism $\phi \colon f^*\mathscr G \to \mathscr F$ of $\mathcal O_X$-modules (equivalently, a morphism $\mathscr G \to f_*\mathscr F$ of $\mathcal O_Y$-modules). The forgetful functor $p \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/S} \to {{\operatorname{\textbf{\textup{Sch}}}}}_S$ makes it into a fibred category [@Stacks Tag [03YM](https://stacks.math.columbia.edu/tag/03YM)]. The fibre ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{X/S} = p^{-1}({\operatorname{id}}_X)$ is canonically equivalent to ${{\operatorname{\textbf{\textup{Qcoh}}}}}(\mathcal O_X){^{\operatorname{op}}}$. Similarly, define the category ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S}$ whose objects are pairs $(Y \to X, \mathscr F)$, where $Y \to X$ is a morphism of $S$-schemes and $\mathscr F$ is a quasi-coherent sheaf on $Y$. A morphism $(f,g,\phi) \colon (Y \to X,\mathscr F) \to (Y' \to X', \mathscr F')$ in ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S}$ consists of morphisms $f \colon Y \to Y'$ and $g \colon X \to X'$ of $S$-schemes such that the diagram $$\label{Dia morphism in QCOH} \begin{tikzcd} Y \ar{r}{f}\ar{d} & Y'\ar{d}\\ X \ar{r}{g} & X' \end{tikzcd}$$ commutes, along with a morphism $\phi \colon f^*\mathscr F' \to \mathscr F$ of $\mathcal O_Y$-modules. It has a forgetful functor $q \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S} \to {{\operatorname{\textbf{\textup{Sch}}}}}_S$ mapping $(Y \to X,\mathscr F)$ to $X$, making it into a fibred category whose fibre ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S} = q^{-1}({\operatorname{id}}_X)$ is canonically equivalent to ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X}$. For a morphism $f \colon Y \to X$ in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$, define the pushforward functor $$\begin{aligned} f_* \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S} &\to {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}\\ (Z \to Y, \mathscr F) &\mapsto (Z \to X, \mathscr F),\\ (g,{\operatorname{id}}_Y,\phi) &\mapsto (g,{\operatorname{id}}_X,\phi).\end{aligned}$$ The pushforward $f_* \colon {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S} \to {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}$ allows us to turn the association $X \mapsto {\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}})$ into a functor ${\operatorname{End}}({\operatorname{id}})$: for a morphism $f \colon Y \to X$ in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$, define the pullback $$f^* \colon {\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}}) \to {\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S}})$$ by $f^*(\alpha)_y = \alpha_{f_*y}$ for any $\alpha \in \mathscr {\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}})$ and any $y \in {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S}$. \[Def E\] Let $\mathscr E \subseteq {\operatorname{End}}({\operatorname{id}})$ be the subfunctor consisting of $\alpha \in {\operatorname{End}}({\operatorname{id}})$ such that $\alpha_{(Z \to X,\mathscr F)} \colon (Z \to X,\mathscr F) \to (Z \to X,\mathscr F)$ is of the form[^2] $({\operatorname{id}}_Z,{\operatorname{id}}_X,\phi)$, for all $(Z \to X,\mathscr F)$ in ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S}$. This property is clearly preserved by the pullback $f^*$, so it defines a subfunctor. Concretely, an element $\alpha \in \mathscr E(X)$ consists of the following data: for every object $(Z \to X,\mathscr F) \in {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}$, a map $$\alpha_{(Z \to X,\mathscr F)} = ({\operatorname{id}}_Z,{\operatorname{id}}_X,\phi_{(Z \to X,\mathscr F)}) \colon (Z \to X,\mathscr F) \to (Z \to X,\mathscr F)\label{Eq pullback alpha}$$ in ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}$ for some morphism of $\mathcal O_Z$-modules $\phi_{(Z \to X,\mathscr F)} \colon \mathscr F \to \mathscr F$, such that for every map $(f,{\operatorname{id}}_X,\psi) \colon (Z \to X,\mathscr F) \to (Z' \to X,\mathscr F')$ in ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}$, the diagram $$\label{Dia End} \begin{tikzcd}[column sep=5em] f^* \mathscr F' \ar{r}{\phi_{(Z' \to X,\mathscr F')}}\ar{d}[swap]{\psi} & f^* \mathscr F' \ar{d}{\psi}\\ \mathscr F \ar{r}{\phi_{(Z \to X,\mathscr F)}} & \mathscr F \end{tikzcd}$$ commutes. If $f \colon Y \to X$ is a morphism, then the pullback $f^* \colon \mathscr E(X) \to \mathscr E(Y)$ is defined by sending $\alpha \in \mathscr E(X)$ to $f^* \alpha \colon {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S}} \to {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S}}$ given by $$\left(f^*\alpha\right)_{(Z \to Y,\mathscr F)} = ({\operatorname{id}}_Z,{\operatorname{id}}_Y,\phi_{(Z \to X,\mathscr F)}) \colon (Z \to Y,\mathscr F) \to (Z \to Y,\mathscr F)$$ for all $(Z \to Y,\mathscr F) \in {{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/Y/S}$. Thus, in (\[Eq pullback alpha\]) we may unambiguously write $\phi_{\mathscr F}$ instead of $\phi_{(Z \to X,\mathscr F)}$, because it does not depend on which $X$ we consider (and we think of $Z$ as understood). We think of $\mathscr E(X)$ as the association of an endomorphism $\phi_{\mathscr F}$ for every quasi-coherent sheaf $\mathscr F$ on an $X$-scheme $Z$, with the compatibility condition (\[Dia End\]). \[Prop structure sheaf\] Let $S$ be a scheme. Then the presheaf $\mathscr E \colon {{\operatorname{\textbf{\textup{Sch}}}}}_S{^{\operatorname{op}}}\to {{\operatorname{\textbf{\textup{Ring}}}}}$ is isomorphic to $\mathcal O \colon X \mapsto \Gamma(X,\mathcal O_X)$. We will define maps $a \colon \mathcal O \to \mathscr E$ and $b \colon \mathscr E \to \mathcal O$ that are mutual inverses. To define $a$, let $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ be given. If $f \in \mathcal O(X)$, then define $a(f) \in \mathscr E(X)$ by letting $\phi_{\mathscr F}$ for a sheaf $\mathscr F$ on $g \colon Z \to X$ be multiplication by $g^\# f \in \mathcal O(Z)$. Clearly the diagrams (\[Dia End\]) commute for every morphism in ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/X/S}$, so $a(f)$ defines an element of $\mathcal E(X)$. Similarly, the definition of the pullback $f^* \colon \mathscr E(X) \to \mathscr E(Y)$ immediately implies that $a$ is a morphism of presheaves. Conversely, if $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ and $\alpha \in \mathscr E(X)$, then set $b(\alpha) = \phi_{\mathcal O_X}(1) \in \mathcal O_X(X)$. For any $f \colon Y \to X$, commutativity of (\[Dia End\]) for $\alpha \in \mathscr E(X)$ and for the morphism $(f,f^\#) \colon (Y,\mathcal O_Y) \to (X,\mathcal O_X)$ gives $$f^*\phi_{\mathcal O_Y}(1) = \phi_{\mathcal O_Y}(1) = f^\#(b(\alpha)),$$ showing that $b$ is a morphism of presheaves. It is clear that $ba(f) = f$ for any $f \in \mathcal O(X)$. Conversely, given $X \in {{\operatorname{\textbf{\textup{Sch}}}}}_S$ and $\alpha \in \mathscr E(X)$, we must show that $\phi_{\mathscr F}$ is multiplication by $f = b(\alpha)$ for any quasi-coherent sheaf $\mathscr F$ on an $X$-scheme $Z$. Sections $\mathscr F(U)$ for $U \subseteq Z$ open are given by maps $\psi \colon \mathcal O_U \to \mathscr F|_U$, so the diagram (\[Dia End\]) for the morphisms $(U \to X,{\operatorname{id}}_X,\psi) \colon (U \to X,\mathscr F) \to (X \to X,\mathcal O_X)$ shows that each element of $\phi_{\mathscr F}(U)$ gets multiplied by $f$. \[Thm forgetful functor\] Let $S$ be a scheme. Then the forgetful functor $U \colon {{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Sch}}}}}$ can be reconstructed categorically from ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. By , the topology can be reconstructed categorically. Moreover, the fibred category ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S} \to {{\operatorname{\textbf{\textup{Sch}}}}}_S$ can be reconstructed categorically: by it is equivalent to the category of pairs $(Y \to X, \mathscr F)$ where $Y \to X$ is a morphism in ${{\operatorname{\textbf{\textup{Sch}}}}}_S$, and $\mathscr F$ a separated abelian cogroup object of $Y/{{\operatorname{\textbf{\textup{Sch}}}}}_Y$. Then the presheaf $\mathscr E \colon {{\operatorname{\textbf{\textup{Sch}}}}}_S{^{\operatorname{op}}}\to {{\operatorname{\textbf{\textup{Ring}}}}}$ of can be reconstructed categorically, which by is isomorphic to the sheaf $\mathcal O$ on the big Zariski site ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. This shows that the functor ${{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{LRS}}}}}$ given by $X \mapsto (|X|,\mathcal O_X)$ can be reconstructed categorically. Its essential image lands in the category of schemes. Proofs of main theorems {#Sec proofs of main theorems} ======================= With the results from , we are ready to prove and . By , the forgetful functor $U \colon {{\operatorname{\textbf{\textup{Sch}}}}}_S \to {{\operatorname{\textbf{\textup{Sch}}}}}$ can be reconstructed categorically from ${{\operatorname{\textbf{\textup{Sch}}}}}_S$. Then implies that the functor $$\begin{aligned} {\operatorname{Isom}}(S,S') &\to {\operatorname{\textbf{\textup{Isom}}}}({{\operatorname{\textbf{\textup{Sch}}}}}_{S'},{{\operatorname{\textbf{\textup{Sch}}}}}_S)\\ f &\mapsto f^*\end{aligned}$$ is an equivalence. Analogously to the construction of ${{\operatorname{\textbf{\textup{Qcoh}}}}}_{-/-/S}$, we can form a category ${{\operatorname{\textbf{\textup{Mod}}}}}_{-/-/R}$ of modules over $R \to A \to B$, and use the pushforward $$f_* \colon {{\operatorname{\textbf{\textup{Mod}}}}}_{-/B/R} \to {{\operatorname{\textbf{\textup{Mod}}}}}_{-/A/R}$$ to make $A \mapsto {\operatorname{End}}({\operatorname{id}}_{{{\operatorname{\textbf{\textup{Mod}}}}}_{-/A/R}})$ functorial in $A$. Similarly, we construct a subfunctor $$\mathscr E \colon {{\operatorname{\textbf{\textup{Alg}}}}}_R \to {{\operatorname{\textbf{\textup{Ring}}}}}$$ of endomorphisms $\alpha \colon {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Mod}}}}}_{-/A/R}} \to {\operatorname{id}}_{{{\operatorname{\textbf{\textup{Mod}}}}}_{-/A/R}}$ that are the identity on the rings $B/A/R$ (so they only act on the module). Analogously to , one shows that $\mathscr E$ is isomorphic to the forgetful functor ${{\operatorname{\textbf{\textup{Alg}}}}}_R \to {{\operatorname{\textbf{\textup{Ring}}}}}$. Then gives the result. -We note that the proof of only relies on the constructions from , so it is much shorter than the proof of Clark–Bergman [@CB], who only treat the case where the base is an integral domain. Already in this case, the statement that ${\operatorname{\textbf{\textup{Isom}}}}({{\operatorname{\textbf{\textup{Alg}}}}}_R,{{\operatorname{\textbf{\textup{Alg}}}}}_{R'})$ is a setoid appears to be new. [^1]: More precisely, there should be a predicate that takes a set and says whether it’s in $\mathscr D$; see e.g. [@Stacks Tag [0009](https://stacks.math.columbia.edu/tag/0009)] for the case $\mathscr D = {{\operatorname{\textbf{\textup{Sch}}}}}$. [^2]: A general $\alpha_{(Z \to X, \mathscr F)}$ has the form $(f,{\operatorname{id}}_X,\phi)$, so the extra assumption is $f = {\operatorname{id}}_Z$.

ArXiv

--- abstract: 'Layered copper oxides have highest superconducting transition temperatures at ambient pressure. Its mechanism remains a grand challenge in condensed matter physics. The essential physics lying in 2-dimensional copper-oxygen layers is well described by a single band Hubbard model or its strong coupling limit t-J model in 2-dimensional square lattice. Recently discovered high temperature superconductor Ba$_2$CuO$_{3+\delta}$ with $\delta \sim 0.2$ has different crystal structure with large portion of in-plane oxygen vacancies. We observe that an oxygen vacancy breaks the bond of its two neighboring copper atoms, and propose ordered vacancies in Ba$_2$CuO$_{3+\delta}$ lead to extended t-J model on an effective brick-wall lattice. For the nearest neighbor hopping, the brick-wall model can be mapped onto t-J model on honeycomb lattice. Our theory explains the superconductivity of the new compound at high charge carrier density, and predict a time reversal symmetry broken pairing state.' author: - Zhan Wang - Sen Zhou - Weiqiang Chen - 'Fu-Chun Zhang' bibliography: - 'BaCuO.bib' title: 't-J model on the effective brick-wall lattice for newly discovered high temperature superconducting Ba$_2$CuO$_{3+\delta}$' --- *Introduction.—*Understanding of high Tc copper oxides[@Keimer2015; @Anderson2004; @Rice2011; @Bednorz1986; @RevModPhys.78.17] remains a grand challenge in condensed matter physics despite intensive studies in the past over thirty years. It is generally accepted that the basic physics of high $T_c$ cuprates is on $\text{CuO}_2$ planes, and the half-filled $d_{x^2-y^2}$ holes of formal $\text{Cu}^{2+}$ form Mott insulator or charge transfer insulator. Chemical doping to the parent compounds introduces additional holes and leads to the $d$-wave superconductivity with the highest transition temperatures at ambient pressure. Recently discovered superconductor of $\text{Ba}_2\text{CuO}_{3+\delta}$ with $\delta \approx 0.2$ is a new family of high $T_c$ cuprate[@Li12156; @Scalapino2019]. The Meissner effect measurement shows that the compound has transition temperature $T_c = 73K$, which is much higher than that of the similar structure high Tc cuprate $\text{Ba}_x\text{La}_{1-x}\text{CuO}_4$. It is then interesting and important to examine the electronic structure and superconductivity of the newly discovered $\text{Ba}_2\text{CuO}_{3+\delta}$, which may also shed new light to our understanding of high $T_c$ cuprates in general. $\text{Ba}_2\text{CuO}_{3+\delta}$ has compressed octahedron with the apical oxygen atoms closer to the central Cu-atoms, the Cu-O bond length for the apical O is about 1.9 Å, shorter than the Cu-O bond length on the Cu-O plane. This is different from all other known cuprates, where the Cu-O bond along the apical O is longer. There are a large number of O-vacancies in $\text{Ba}_2\text{CuO}_{3+\delta}$, which are located on the $\text{CuO}_2$ plane as neutron data showed[@Huang]. These vacancies are expected to play important role in the electron structure and possibly superconductivity. First principles calculations have been reported by Liu et al[@PhysRevMaterials.3.044802], who considered various possible crystal structures. For $\delta=0.25$, they have found crystal structures corresponding to lower total energies. There have been a number of proposed theories for its electronic structure and superconductivity. In the usual cuprates, the formal valence of Cu is less than $+2.2$, corresponding to the hole concentration on the $\text{CuO}_2$ plane less than $20\%$. In $\text{Ba}_2\text{CuO}_{3+\delta}$, the average valence of Cu is $+2(1+\delta)$, corresponding to average hole concentration of $2\delta \gg 0.2$ for the value of given $\delta\sim0.2$. This has led to the proposal of 2-band Hubbard model to describe the new compound[@PhysRevB.99.224515; @2019arXiv190912620L]. Large number of O vacancies has led to the proposal of weakly linked 2-chain ladder for the superconductivity[@2019arXiv190908304L], whose related physics has been previously examined extensively in study of Hubbard or t-J ladders[@Dagotto1996]. In this paper, we take the viewpoint that the observed superconductivity in $\text{Ba}_2\text{CuO}_{3+\delta}$ is resulted from an ordered crystal structure and the O-vacancy effectively transforms the original square lattice of $\text{CuO}_2$ to a new lattice. We propose a crystal structure for $\text{Ba}_2\text{CuO}_{3+\delta}$, where layers of 1-dimensional CuO chains and layers of 2-dimensional $\text{CuO}_{1.5}$ plane are alternating as illustrated in Fig. \[model\](a). We expect the CuO chains to serve as charge reservoir, in analogy to the CuO chains in YBCO, and the CuO$_{1.5}$ planes to contain the essential physics for the superconductivity in Ba$_2$CuO$_{3+\delta}$. We argue that the effective Hamiltonian of a CuO$_{1.5}$ plane is described by a single band $t$-$t'$-$J$ model on an underlying brick-wall lattice, which is shown in fig. \[model\](b), with $t$ and $t'$ the nearest neighbor (n.n.) and next n.n. hopping integrals and $J$ the n.n. spin-spin coupling of the spin-1/2 Cu-holes. We use renormalized mean field theory[@Zhang_1988] to study the $t$-$t'$-$J$ model on the brick-wall lattice. We find that the superconductivity extends to hole density of as high as $40\%$, and the maximum of the pairing order parameter appears at a larger hole concentration than the other cuprates due to the shift of the van Hove singularity for the density of states in the brick-wall lattice. The pairing symmetry depends on the value of $t'/t$ and the hole concentration, and may break time reversal symmetry, which can be tested in muon spin rotation ($\mu$SR) experiment. The effects of bond disorder is also studied, and the superconductivity is expected to survive a weak bond disorder that deviates the structure from perfect brick-wall lattice. Our theory appears to be consistent with the available experiments and explains the superconductivity of $\text{Ba}_2\text{CuO}_{3+\delta}$ with relatively large hole concentration. We attribute the change of the electronic structure in the planar layer of the compound to the O-vacancy, which in turn changes the effective lattice. This may provide a new route to study high $T_c$ in future. *Model and effective lattice.—*As mentioned above, we consider a crystal structure shown in Fig. \[model\] (a), with alternative layers of $\text{CuO}$ chain and $\text{CuO}_{1.5}$ plane (x-y plane). The average $\delta$ is $0.25$ in such a structure. The $\text{O}$-vacancies (missing O-atoms) in $\text{CuO}_{1.5}$ layer form a square lattice and there are three O-atoms around each $\text{Cu}$-atom on the plane. The energy of the crystal structure has not been calculated, but should be similar to the one of the lowest energy crystal structure calculated for $\text{Ba}_2\text{CuO}_{3.25}$ by Liu. et al[@PhysRevMaterials.3.044802], with the difference that in the latter case the $\text{O}$-vacancies form an alternative chains along the x-direction. We will focus on the electronic structure of $\text{CuO}_{1.5}$ plane and consider CuO chains to serve as charge reservoir. While the average of the formal valence of Cu in $\text{Ba}_2\text{CuO}_{3.25}$ is $\text{Cu}^{2.5+}$, the formal valence of Cu-atom on the $\text{CuO}_{1.5}$ layer can be significantly smaller because of the compensation from the CuO chains. ![(a) Proposed crystal structure for $\text{Ba}_2\text{CuO}_{3.25}$ with alternating Cu-O chain layer (middle layer) and Cu-$\text{O}_{1.5}$ plane (top and bottom layers). The apical O-atoms are all occupied. The Cu-$\text{O}_{1.5}$ plane forms a brick-wall structure due to the missing O-atoms. (b)The brick-wall lattice formed by Cu (light blue for A sublattice and dark blue for B sublattice) and O (red) atoms. $\Delta_\alpha,\alpha=1,2,3$ are the superconducting order parameters on the corresponding bonds. The unit cell vectors are depicted as green dashed lines with length $a$.[]{data-label="model"}](Fig1.pdf){width="1.\linewidth"} For each Cu-atom on the $\text{CuO}_{1.5}$ plane, there are five nearest neighbor $\text{O}$ atoms including two apical O-atoms, two on the x-axis and one on the y-axis. The possible anti-bonding Cu-$3d$ orbitals here are $3d_{3z^2-r^2}$ or $3d_{z^2-x^2}$, whose relative energies depend on the difference or ratio of the bond lengths between the short apical Cu-O bond and the in-plane Cu-O bond. Here we assume orbital $3d_{3z^2-r^2}$ to have higher energy, which is consistent with a density functional calculation for proper parameters[@Le]. Then the lowest-energy of the atomic $3d$ hole state is $d_{3z^2-r^2}$, which replaces $3d_{x^2-y^2}$ in other cuprates as the relevant orbital. Considering Cu-$3d^{10}$ and O-$2p^6$ as the vacuum configuration, the formal $\text{Cu}^{2+}$ thus has one hole primarily on Cu-$3d_{3z^2-r^2}$. Because of the large repulsive interaction $U$ for two holes on the same Cu-site, the ground state at the half filled, namely one hole per Cu-atom in average, will be a Mott insulator, similar to that of $\text{La}_2\text{CuO}_4$. We next consider additional holes to the half filled case. As soft-X ray absorption experiment[@Li12156] shows, the additional hole largely goes to the O-2$p$ orbitals, which implies that the formal $\text{Cu}^{3+}$ has one hole on Cu-$3d_{3z^2-r^2}$ and the other on O-$2p$. We expect them to form a spin singlet, similar to the Zhang-Rice spin singlet[@PhysRevB.37.3759] formed in other cuprates, and moves through the lattice as a charge carrier. Since the hopping of the hole between the two neighboring Cu-atoms is essentially mediated by oxygen $2p$ orbitals in between, it is strongly suppressed if the oxygen between the two Cu-atoms is missing. Therefore, the nearest neighbor Cu-Cu bond may be effectively removed from the lattice if the O-atom between them is missing. This leads to an effective brick-wall lattice as shown in fig.\[model\](b). As we discussed above, the physics in Ba$_2$CuO$_{3+\delta}$ may be similar to that in other cuprates and we may consider the material as a doped Mott insulator within a single orbital model, although the relevant Cu-$3d$ orbital and the underlying lattice due to the missing O ions will be different from the other cuprates. The low energy physics of $\text{Ba}_2\text{CuO}_{3.25}$ is thus described by a 2-dimensional t-J model on a brick-wall lattice, $$\label{eq:1} \mathcal{H} = - \sum_{ij} \left(t_{ij} P_G c_{i \sigma}^{\dag} c_{j \sigma} P_G +\text{h.c.}\right)+ J\sum_{\left\langle ij \right\rangle} \mathbf{S}_i \cdot \mathbf{S}_j,$$ where $c_{i \sigma}$ and $c_{i \sigma}^{\dag}$ are the annihilation and creation operator of electrons with spin $\sigma$ at site $i$ respectively, $t_{ij}$ is the hopping integral of electrons between site $i$ and $j$, $P_G$ is the Gutzwiller projection operator to project out doubly occupied electron states on the Cu-sites and $\left\langle ij \right\rangle$ means n.n. pairs of $i$ and $j$. Repeated spin indices are summed. As we analyzed above, the dominant orbital is $3d_{3z^2-r^2}$, thus the hopping integrals are isotropic along the x- and y- axes. And we consider only the case with n.n. and next n.n. hoppings, whose hopping integrals are $t_{nn} = t$ and $t_{nnn} = t'$ respectively. *Mean field results.—*As shown in Fig. \[model\](b), the superconducting order parameters of the brick-wall lattice are denoted by $\Delta_{\alpha}$ with $\alpha = 1,2,3$. Usually, the symmetry of a system will impose some constraints on the order parameters. The point group symmetry of the brick-wall lattice is $\mathcal{D}_2$ which has four irreducible representations as listed in Table \[char\]. Since we are only interested in spin singlet pairing, only the $A$ and $B_1$ representations are relevant. For $A$ representation, We have real pairings with $\Delta_1 = \Delta_2$. For $B_1$ representation (will be denoted as $B$ hereafter), we have real pairings with $\Delta_1 = - \Delta_2, \Delta_3 = 0 $. ![Spin singlet pairing order parameters $\Delta_{xA}$, $\Delta_{xB}$ and $\Delta_y$ of Hamiltonian obtained from the renormalized mean field theory for $t'=0$ (a), $t'/t=0.1$ (b) and $t'/t=-0.1$ (c) with $J/t=0.4$ for all cases. $\Delta_{xA}(\Delta_{xB})$ indicates the contribution from $A$($B$) representation, respectively. The pairing order parameters on three distinct bonds are indicated in (d). The insets in (a-c) are the DOS and the corresponding Fermi energy at the optimal dopings marked by the arrows. []{data-label="result"}](Fig2.pdf){width="1.\linewidth"} Reps e $C_2(z)$ $\sigma_x$ $\sigma_y$ ------- ---- ---------- ------------ ------------ $A$ +1 +1 +1 +1 $B_1$ +1 +1 -1 -1 $B_2$ +1 -1 +1 -1 $B_3$ +1 -1 -1 +1 : character table of group $\mathcal{D}_2$ \[char\] $t'=0$ is a special case, where the Hamiltonian has a higher symmetry than the lattice, the $D_6$ symmetry. This can be understood as the following. The brick-wall lattice can be treated as a squeezed honeycomb lattice. If one considers only the n.n. coupling, the model can be mapped onto a t-J model with the n.n. hopping on an honeycomb lattice, and has a $D_6$ symmetry. Previous studies[@PhysRevB.81.085431; @Nandkishore2012; @PhysRevLett.100.146404; @PhysRevB.87.094521; @PhysRevB.85.035414] on t-J model on honeycomb lattice have suggested a $d+id$-wave topological superconducting phase which corresponds to a two-dimensional representation of $D_6$ group. Although the $D_6$ symmetry is reduced to $D_2$ as one turns on the next n.n. hopping, it is possible that the pairing symmetry may still be $d+id$, or more precisely $A+iB$ in the $D_2$ case, where the 2D representation reduces to $A\oplus B$, for small $t'$. In the following calculations, we will take the $A+iB$ ansatz $\Delta_1 = \Delta_2^{*} = \Delta_{xA} + i \Delta_{xB}$ and $\Delta_3 = \Delta_y$ where $\Delta_{xA}$, $\Delta_{xB}$, and $\Delta_y$ are all real, as shown in Fig. \[result\](d). $\Delta_{xA}$ and $\Delta_{xB}$ tracks the contribution from $A$ and $B$ respectively, a case with finite $\Delta_{xA}$ and vanished $\Delta_{xB}$ corresponds to pure $A$ phase, while a case with finite $\Delta_{xB}$ and $\Delta_{xA}=\Delta_y = 0$ corresponds to pure $B$ phase. {width="1.\linewidth"} . \[phasediagram\] Then we solve the superconducting order parameters with the standard renormalized mean field theory approach (see the details in appendix \[sec:renorm-mean-field\]), and the results for $t'=0$, $t'=0.1$, and $t'=-0.1$ are depicted in Fig. \[result\] (a), (b) and (c) respectively. The result for $t'=0$ shows a time-reversal symmetry breaking $A+iB$-wave ($d+id$-wave) phase in large doping regime, which is consistent with the previous studies on honeycomb lattice. The optimal doping of the superconducting dome corresponds to the van Hove singularities as shown in the inset of Fig. \[result\](a). For small but finite $t' = \pm 0.1 t$, we still have the $A+iB$ phase at low doping ($p < 0.25$). But with increasing doping, the $A+iB$ phase is replaced by a pure $B$ phase (for $t' = 0.1t$) or a pure $A$ phase (for $t' = -0.1t$). In each phase, we find that the doping with maximum superconducting order parameter corresponds to the peak in density of states, i.e. the van Hove singularities. We also perform the calculations for various dopings and $t'$, the resultant phase diagram is shown in Fig. \[phasediagram\]. One can find a robust $A+iB$ phase in rather large parameter regime and a phase transition from $A+iB$ phase to $A$ phase in $t'<0$ region and a transition from $A+iB$ to $B$ phase in $t'>0$ region. The different behavior between positive and negative $t'$ can be understood from the Fermi surface geometry. In Fig. \[phasediagram\], we depict the typical Fermi surface for each phases. In large doping regime, the underlying Fermi surface for $t' > 0$ has a very good nesting along $k_x$ direction that favors one dimensional instability, while the Fermi surface for $t' < 0$ does not have such a nesting and favors a two dimensional instability. On the other hand, the superconducting pairing in $B$ phase is one dimensional (because $\Delta_{y}=0$), while the pairing in $A$ phase is more two dimensional. Thus one have the $B$ phase at $t'>0$ and $A$ phase at $t'<0$ in large doping regime. As we analyzed above, the $A+iB$ phase at $t' = 0$ corresponds to the $d+id$ superconductivity, thus the $A+iB$ phase should also be a topological superconducting phase. To confirm that, we calculated the Chern number $\nu_{\text{Ch}}$ associated with each phase. We find that the $A+iB$ phase has a Chern number $\nu_{\text{Ch}}=2$, while the pure $A$ and $B$ phase is topologically trivial and has Chern number $\nu_{\text{Ch}}=0$. Thus, the phase transition between them is indeed a topological phase transition associated with a gap closing behavior. *Effects of bond disorder.—* It is natural to expect the existence of bond disorder in the system that deviates the structure from a perfect brick-wall lattice, and its effects on the superconductivity need to be investigated. ![(a) Ensemble averaged renormalized pairing $\tilde{\Delta}$ as a function of bond disorder strength $\eta$ at zero temperature. (b) Temperature dependence of the ensemble averaged $\tilde{\Delta}$ at bond disorder $\eta =0$, $\eta \simeq 0.039$, and $\eta \simeq 0.078$. Crosses represent data of each disorder realizations. []{data-label="disorder"}](fig4b.eps){width="1.\linewidth"} We start with a $N=32\times 32$ brick-wall lattice of periodic boundary condition, in which there are $3N/2$ n.n. bonds with nonzero hopping integral $t$ and superexchange $J$, forming the structure depicted in Fig. \[model\](b). Note that, as compared with the square lattice, there are $N/2$ n.n. bonds missing in the brick-wall lattice. To introduce bond disorder, $N_\text{dis}$ n.n. bonds are redistributed randomly. Explicitly, we take out $N_\text{dis}$ bonds randomly from the $3N/2$ n.n. bonds of the brick-wall lattice and then distribute them randomly to the place of $N/2$ previously missing bonds. Clearly, $N_\text{dis}$ takes an integer value between 0 and $N/2$, and the strength of bond disorder is measured by the value of $\eta =2N_\text{dis}/N$. At a given $\eta$, we generate 20 disorder realizations and obtain the ground state of each realization self-consistently. We stay with $t'=0$ and doping concentration $\rho =0.25$ where the superconducting brick-wall lattice is in the $A+iB$ phase. The average renormalized pairing of each disorder realization, $\tilde{\Delta} = \frac{2}{3N} \sum_{\langle i,j \rangle} g^t_{ij} |\Delta_{ij}|$, at zero temperature is plotted in Fig. \[disorder\](a) as crosses at various disorder strength $\eta$, with the solid line represent their ensemble averages. The ensemble averaged $\tilde{\Delta}$ decreases linearly as the disorder strength $\eta$ increases, and saturates to a nonzero minimum value at $\eta \simeq 0.09$. We thus expect the superconductivity on the perfect brick-wall lattice to survive weak bond disorders. To estimate the effects of bond disorder on the superconducting transition temperature $T_c$, the temperature dependence of $\tilde{\Delta}$ is plotted in Fig. \[disorder\](b) for $\eta \simeq 0.038$ and $\eta \simeq 0.078$, with the result in the absence of disorder ($\eta =0$) is also shown for comparison. Clearly, the suppression of the mean-field critical temperature by bond disorder is much weaker than the suppression of the ensemble averaged $\tilde{\Delta}$ at zero-temperature. *Summary.—* In summary, we have proposed an effective brick-wall $t-t'-J$ model for the newly discovered high Tc superconductor. By using renormalized mean field theory, we have demonstrated that the superconductivity extends to very large hole concentration, and the pairing order parameter is peaked at larger hole concentration. The pairing symmetry can be complex and breaks time reversal invariance, similar to the superconductivity theoretically studied for the $t-J$ model on the honeycomb lattice. It will be interesting to test this in $\mu$SR experiment. We note that Hubbard model on square lattice has recently been studied using more sophisticated numerical methods. Our study based on renormalized mean field theory on the brick-wall lattice may be viewed as a starting point for these more advanced numerical methods. Our mean field result on the bond disorder effect suggests a relatively weaker reduction to the superconductivity transition temperature. We wish to thank C. Q. Jin and Q. Z. Huang for many useful discussions on the experimental results, Congcong Le for the density functional theory calculations to estimate the energy of the local electronic configurations. We would like also to thank J. P. Hu, Z. Q. Wang, K Jiang for many stimulating discussions. This work is in part supported by the National Natural Science Foundation of China (No. 11674151, 11847612, 11974362 and 11674278), National Key Research and Development Program of China (Grant No. 2016YFA0300300), the Strategic Priority Research Program of CAS (No. XDB28000000) and Beijing Municipal Science and Technology Commission (No. Z181100004218001). Renormalized Mean Field Theory {#sec:renorm-mean-field} ============================== The basic idea of renormalized mean field theory (RMFT) is to approximate the Gutzwiller projection $P_G$ in by Gutzwiller factors, and the Hamiltonian becomes $$\mathcal{H} = - \sum_{ ij } (g^t_{ij} t_{ij} c_{i \sigma}^{\dag} c_{j \sigma} + h.c.) + J\sum_{\left\langle ij \right\rangle} g^J_{ij} \mathbf{S}_i \cdot \mathbf{S}_j,$$ where $g^t_{ij} = \sqrt{4\rho_i \rho_j / (1+\rho_i) (1+\rho_j)}$ and $g^J_{ij} = {4/ (1+\rho_i) (1+\rho_j)}$, with $\rho_i$ the local doping concentration on site $i$. The exchange term is then decoupled in terms of the mean fields of hopping $\chi^0_{ij} = \langle c^\dagger_{i\sigma} c_{j\sigma} \rangle$ and pairing $\Delta^0_{ij} = \langle \epsilon_{\sigma\sigma'} c_{i\sigma} c_{j\sigma'} \rangle$. The mean-field Hamiltonian reads $$\begin{aligned} \mathcal{H} =& - \sum_{ij} (g^t_{ij} t_{ij} c_{i \sigma}^{\dag} c_{j \sigma} + h.c.) - \frac{3J}{8} \sum_{\left\langle ij \right\rangle} g^J_{ij} \label{Hij}\\ & \times \left[\chi^{0*}_{ij} c^\dagger_{i\sigma} c_{j\sigma} +\Delta^{0*}_{ij} \epsilon_{\sigma\sigma'} c_{i\sigma} c_{j\sigma'} +h.c. -|\chi^0_{ij}|^2 -|\Delta^0_{ij}|^2 \right]. \nonumber\end{aligned}$$ The ground state is obtained by solving the hopping and pairing parameters in Eq. (\[Hij\]) self-consistently. In uniform solutions, $\rho_i = \rho$, and the Gutzwiller factors $g^t_{ij} = g_t = 2\rho/(1+\rho)$ and $g^J_{ij} = g_J = 4/(1+\rho)$. The three independent n.n. bonds are denoted as $\alpha = 1, 2, 3$, as shown in Fig. \[model\] (b). In the momentum space, the mean-field Hamiltonian reads $$\mathcal{H}^{MF}=\sum_k C^{\dag}_{\mathbf{k}} \left(\begin{array}{cccc} \epsilon(\mathbf{k}) & f(\mathbf{k}) & 0 & g_2^*(\mathbf{k}) \\ f^*(\mathbf{k}) & \epsilon(\mathbf{k}) & g_1^*(\mathbf{k}) & 0 \\ 0 & g_1(\mathbf{k}) & -\epsilon(\mathbf{k}) & -f(\mathbf{k}) \\ g_2(\mathbf{k}) & 0 & -f^*(\mathbf{k}) & -\epsilon(\mathbf{k}) \\ \end{array}\right) C_{\mathbf{k}},$$ where $C_{\mathbf{k}} = (C_{\mathbf{k}A\uparrow}, C_{\mathbf{k}B\uparrow}, C_{-\mathbf{k}A\downarrow}^\dagger, C_{-\mathbf{k}B\downarrow}^\dagger)^T$, and the matrix element is given by $$\begin{aligned} \epsilon(\mathbf{k})&=-\sum_{\delta}g_{t}t' \exp[i\mathbf{k}\cdot \mathbf{r}_{\delta}]-\mu\notag\\ f(\mathbf{k})&=\sum_\alpha(-g_tt-\frac{3}{8}g_JJ\chi^0_\alpha)\exp[i\mathbf{k}\cdot \mathbf{r}_{\alpha}]\notag\\ g_1(\mathbf{k})&=\sum_\alpha(-\frac{3}{8}g_J J \Delta^0_\alpha) \exp[i\mathbf{k}\cdot \mathbf{r}_{\alpha}]\notag\\ g_2(\mathbf{k})&= \sum_\alpha(-\frac{3}{8}g_J J \Delta^0_\alpha) \exp[-i\mathbf{k} \cdot \mathbf{r}_{\alpha}]\notag,\end{aligned}$$ where $\alpha = 1,2,3$ is the index of the n.n. site while $\delta = 1,2,3,4$ is the index of the next n.n. site. $H^{MF}$ can be diagonalized via Bogolyubov transformation at each $\vec k$ point $$\left(\begin{array}{c} C_{\mathbf{k}A\uparrow}\\ C_{\mathbf{k}B\uparrow}\\ C_{-\mathbf{k}A\downarrow}^\dagger \\ C_{-\mathbf{k}B\downarrow}^\dagger \end{array}\right)=U^{\mathbf{k}}\left(\begin{array}{c} \alpha_{\mathbf{k}\uparrow} \\ \beta_{\mathbf{k}\uparrow} \\ \alpha_{-\mathbf{k}\downarrow}^\dagger \\ \beta_{-\mathbf{k}\downarrow}^\dagger \end{array}\right)$$ And the self-consistent equations are given by $$\begin{aligned} \Delta^0_\alpha&=\frac{1}{N_S}\sum_{\mathbf{k}}\exp[i\mathbf{k}\cdot \mathbf{r}_{\alpha}](u_{13}^{*\mathbf{k}}u_{43}^\mathbf{k}+u_{14}^{*\mathbf{k}}u_{44}^\mathbf{k})\\ \chi^0_\alpha&=\frac{1}{N_S}\sum_{\mathbf{k}}\exp[-i\mathbf{k}\cdot \mathbf{r}_\alpha](u_{13}^{*\mathbf{k}}u_{23}^\mathbf{k}+u_{14}^{*\mathbf{k}}u_{24}^\mathbf{k}),\end{aligned}$$ where $N_S$ is the number of unit cells. The value of the mean fields can be obtained by solving the self-consistent equation. The physical superconducting order parameters are given by $$\begin{aligned} \label{eq:2} \Delta_{\alpha} & = g_t \Delta^0_{\alpha}.\end{aligned}$$

ArXiv

--- abstract: | Let $G$ be a finite simple graph on $n$ vertices and $J_G$ denote the corresponding binomial edge ideal in the polynomial ring $S = K[x_1, \ldots, x_n, y_1, \ldots, y_n].$ In this article, we compute the Hilbert series of binomial edge ideal of decomposable graphs in terms of Hilbert series of its indecomposable subgraphs. Also, we compute the Hilbert series of binomial edge ideal of join of two graphs and as a consequence we obtain the Hilbert series of complete $k$-partite graph, fan graph, multi-fan graph and wheel graph. address: 'Department of Mathematics, Indian Institute of Technology Madras, Chennai, INDIA - 60036' author: - Arvind Kumar - Rajib Sarkar bibliography: - 'binomial.bib' title: HILBERT SERIES OF BINOMIAL EDGE IDEALS --- [^1] Introduction ============ Let $G$ be a finite simple graph on the vertex set $[n]$. Herzog et al. in [@HH1] and Ohtani, independently in [@oh], introduced the notion of binomial edge ideal corresponding to a finite simple graph. Let $S=K[x_1, \ldots, x_n,y_1, \ldots, y_n]$, where $K$ is a field. The binomial edge ideal of the graph $G$ is $J_G =(x_i y_j - x_j y_i : \{i,j\} \in E(G), \; i <j)$. Researchers have been trying to relate the algebraic properties of $J_G$ with the combinatorial properties of $G$, see for example [@AB; @her1; @HH1; @JA1; @JNR; @KM6; @FM]. While the Castelnuovo-Mumford regularity of binomial edge ideals of several classes of graphs are known, not much is known about other invariants such as Betti numbers, Hilbert series and multiplicity. Betti numbers and Hilbert series of the binomial edge ideals of cycles were computed by Zafar and Zahid in [@SZ]. Mohammadi and Sharifan studied the Hilbert series of quasi cycles in [@FM]. In [@MR3195706], Schenzel and Zafar computed dimension, depth, multiplicity and Betti numbers of complete bipartite graphs. The aim of this article is to compute the Hilbert series of binomial edge ideals of graphs in terms of the Hilbert series of certain subgraphs. A graph $G$ is said to be decomposable if there exist induced subgraphs $G_1$ and $G_2$ such that $G= G_1 \cup G_2$, $V(G_1)\cap V(G_2)=\{v\}$ and $v$ is a free vertex of $G_1$ and $G_2$. A graph $G$ is indecomposable, if it is not decomposable. Upto permutation, $G$ has a unique decomposition into indecomposable subgraphs, i.e. there exist indecomposable subgraphs $G_1,\ldots,G_r$ of $G$ with $G=G_1\cup \cdots \cup G_r$ such that for each $i \neq j$, either $V(G_i) \cap V(G_j) = \phi$ or $V(G_i) \cap V(G_j) =\{v_{i,j}\}$ and $v_{i,j}$ is a free vertex of $G_i$ and $G_j$. In Section $3$, we obtain the Hilbert series and multiplicity of a decomposable graph in terms of the Hilbert series of its indecomposable subgraphs (Theorem \[3.3\]). As consequences we obtain the Hilbert series and multiplicity of Cohen-Macaulay closed graphs and $k$-handle lollipop graphs. Let $H$ and $H'$ be two graphs with the vertex sets $[p]$ and $[q]$, respectively. The join of $H$ and $H'$, denoted by $H*H'$ is the graph with vertex set $[p] \sqcup [q]$ and the edge set $E(H*H')= E(H) \cup E(H') \cup \{\{i,j\}| i \in [p], j \in [q]\}$. Our aim is to compute the Hilbert series of $H*H'$ in terms of the Hilbert series of $H$ and $H'$. First, we treat the case when both the graphs are disconnected (Theorem \[4.3\]). In order to compute the Hilbert series of join of two arbitrary graphs, it is necessary to understand the structure of some other intermediate graphs. We define a product operation of a graph with the complete graph and study the minimal primes and Hilbert series of the resulting graph. We further compute the Hilbert series of several intermediate graphs and use those results finally to obtain: **Theorem \[4.15\]**.   [*Let $H$ and $H'$ be two graphs on vertex sets $[p]$ and $[q]$, respectively. Let $G=H*H'$ be the join of $H$ and $H'$. Let $S_{H}=K[x_{i},y_{i} : i\in V(H)]$, $S_{H^{'}}=K[w_{i},z_{i} : i\in V(H^{'})]$ and $S=K[x_{i},y_{i},w_{j},z_{j} : i\in V(H), j\in V(H^{'})]$. Then*]{} $${\operatorname{Hilb}}_{S/J_{G}}(t) ={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+{\operatorname{Hilb}}_{S_{H^{'}}/J_{H^{'}}}(t) + \frac{(p+q-1)t+1}{(1-t)^{p+q+1}}-\frac{(p-1)t+1}{(1-t)^{p+1}}- \frac{(q-1)t+1}{(1-t)^{q+1}}.$$ As consequences, we obtain the Hilbert series of binomial edge ideal of complete $k$-partite graph, wheel graph, fan graph and multi-fan graph. **Acknowledgements:** The authors are grateful to their advisor A. V. Jayanthan for constant support, valuable ideas and suggestions. The first author thanks the National Board for Higher Mathematics, India for the financial support. The second author thanks University Grant Commission, Government of India for the financial support. Preliminaries ============= In this section we recall some notation and fundamental results which are used throughout this article. Let $G$ be a finite simple graph with the vertex set $V(G)$ and edge set $E(G)$. For a subset $A \subseteq V(G)$, $G[A]$ denotes the induced subgraph of $G$ on the vertex set $A$, that is, for $i, j \in A$, $\{i,j\} \in E(G[A])$ if and only if $ \{i,j\} \in E(G)$. A subset $U \subseteq V(G)$ is said to be a *clique* if $G[U]$ is a complete graph. A clique $U$ is said to be a maximal clique if for every $v \in V(G) \setminus U$, $U \cup \{v\}$ is not a clique. A vertex $v \in V(G)$ is called a *free vertex* if $v$ belongs to exactly one maximal clique. For a vertex $v$, $G \setminus v$ denotes the induced subgraph of $G$ on the vertex set $V(G) \setminus \{v\}$. A vertex $v \in V(G)$ is said to be a *cut* vertex if $G \setminus v$ has strictly more connected components than $G$. For a vertex $v$, $G_v$ denotes the graph on the vertex set $V(G)$ and edge set $E(G_v) =E(G) \cup \{ \{u,w\}: u,w \in N_G(v)\}$, where for any $x\in V(G)$, $N_G(x) = \{u \in V(G) : \{u,x\} \in E(G) \}$. Complement of $G$, denoted by $G^c$ is the graph on the vertex set $V(G)$ and the edge set $E(G^c)= \{\{i,j\} : i \neq j, \{i,j\} \notin E(G) \}$. For a subset $T$ of $[n]$, let $\bar{T} = [n]\setminus T$ and $c_G(T)$ denote the number of connected components of $G[\bar{T}]$. Let $G_1,\cdots,G_{c_{G}(T)}$ be connected components of $G[\bar{T}]$. For each $i$, let $\tilde{G_i}$ denote the complete graph on $V(G_i)$ and $$P_T(G) = (\underset{i\in T} \cup \{x_i,y_i\}, J_{\tilde{G_1}},\cdots, J_{\tilde{G}_{c_G(T)}}).$$ It was shown by Herzog et al. that $J_G = \underset{T \subseteq [n]}\cap P_T(G)$, [@HH1]. For each $i \in T$, if $i$ is a cut vertex of the graph $G[\bar{T} \cup \{i\}]$, then we say that $T$ has the cut point property. Let $\mathscr{C}(G) =\{\phi \} \cup \{ T: T \; \text{has cut point property} \}$ and $\mathscr{M}(G) = \{ P_T(G) : T \in \mathscr{C}(G)\}$. In [@HH1], the authors proved that $P$ is the minimal prime associated to $J_G$ if and only if $P \in \mathscr{M}(G)$. Let $M =\underset{k \in \mathbb{N}} \bigoplus M_k$ be a graded $S$-module such that for each $k \in \mathbb{N}$, $l(M_k) < \infty$. The function $H_M : \mathbb{N} \rightarrow \mathbb{N}$ defined as $H_M(k) = l(M_k)$ is called the Hilbert function of the module $M$. The Hilbert series of $M$ is the generating function of the Hilbert function $H_M$ and is denoted by, ${\operatorname{Hilb}}_{M}(t)= \underset{k \in \mathbb{N}} \sum l(M_k) t^k$. Let $M$ be a finitely generated graded $S$-module of dimension $d$. The Hilbert polynomial of $M$, denoted by $P_M(X)$ is the unique polynomial with rational coefficients such that $H_M(k) = P_M(k)$ for $k\gg0$. It is known that one can express the polynomial $P_M(X)$ as $$P_M(X) = \sum_{i=0}^{d-1} (-1)^{d-1-i} e_{d-1-i}(M) \binom{X+i} {i},$$ where $e_j(M)$’s are integers in [@bh Lemma 4.1.4]. The coefficient $e(M)=e_0(M)$ is called the multiplicity of $M$. Hilbert series of binomial edge ideal of decomposable graphs ============================================================ In this section, we compute the Hilbert series of decomposable graphs in terms of the Hilbert series of its indecomposable components. Throughout this section, by writing $G = G_1 \cup \cdots \cup G_r$, we mean that $G_i$’s are induced subgraphs of G such that $V(G_i) \cap V(G_j) = \phi$ or $\{v_{i,j}\}$ where $v_{i,j}$ is a free vertex in $G_i$ and $G_j$. We begin by recalling the Betti polynomial of $M$. Let $M$ be a finite graded $S$-module. Then we denote by $$B_{M}(s,t)=\displaystyle \sum _{i,j} \beta_{ij}(M)s^{i}t^{j},$$ the Betti polynomial of $M$, where $\beta_{ij}(M)$ are the graded Betti numbers of $M$. Recently J. Herzog and G. Rinaldo [@her2 Proposition 3] proved that if $G=G_{1}\cup G_{2}$, then $$B_{S/J_{G}}(s,t)= B_{S/J_{G_{1}}}(s,t) B_{S/J_{G_{2}}}(s,t)= B_{S_1/J_{G_1}}(s,t)B_{S_2/J_{G_2}}(s,t),$$ where $S_i = K[x_j,y_j:j \in V(G_i)]$ for $i \in \{1,2\}$. By [@bh Lemma 4.1.13], the Hilbert series of a graded $S$-module $S/J_G$ is $${\operatorname{Hilb}}_{S/J_{G}}(t)=\frac{B_{S/J_{G}}(-1,t)}{(1-t)^{2n}}.$$ As an immediate consequence, we obtain \[3.3\] Let $G=G_{1}\cup G_{2}$ be a graph on the vertex set $[n]$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)=(1-t)^{2} {\operatorname{Hilb}}_{S_{1}/J_{G_{1}}}(t) {\operatorname{Hilb}}_{S_{2}/J_{G_{2}}}(t),$$ where $S_{i}=K[{x_{j},y_{j}:j \in V(G_{i})}]$, for $i=1,2$. In particular, $$\dim(S/J_G) = \dim(S_1/J_{G_1}) +\dim(S_2/J_{G_2})-2 \ \textit{and} \ e(S/J_{G})=e(S_{1}/J_{G_{1}})e(S_{2}/J_{G_{2}}).$$ For $i=1,2$, let $G_i$ be a graph on vertex set $[m_i]$. Note that $n=m_1+m_2-1$. Now, $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t)&=&\frac{B_{S/J_{G}}(-1,t)}{(1-t)^{2n}} \\&=& \frac{B_{S/J_{G}}(-1,t)}{(1-t)^{2(m_{1}+m_{2}-1)}} \\ &= &(1-t)^{2} \frac{B_{S_{1}/J_{G_{1}}}(-1,t)}{(1-t)^{2 m_{1}}} \frac{B_{S_{2}/J_{G_{2}}}(-1,t)}{(1-t)^{2 m_{2}}} \\ &= &(1-t)^{2} {\operatorname{Hilb}}_{S_{1}/J_{G_{1}}}(t) {\operatorname{Hilb}}_{S_{2}/J_{G_{2}}}(t).\end{aligned}$$ Now, let $d= \dim(S/J_G) $ and $d_{i}$ = $\dim(S_{i}/J_{G_{i}})$, for $i=1,2$. It follows from [@bh Corollary 4.1.8] that $ {\operatorname{Hilb}}_{S/J_{G}}(t) = \frac{Q_{S/J_{G}}(t)}{(1-t)^{d}}$, $ {\operatorname{Hilb}}_{S_1/J_{G_1}}(t) = \frac{Q_{S_1/J_{G_1}}(t)}{(1-t)^{d_1}}$ and $ {\operatorname{Hilb}}_{S_2/J_{G_2}}(t) = \frac{Q_{S_2/J_{G_2}}(t)}{(1-t)^{d_2}}$. Therefore $$\begin{aligned} \frac{Q_{S/J_{G}}(t)}{(1-t)^{d}} &=&(1-t)^{2} \frac{Q_{S_{1}/J_{G_{1}}}(t)}{(1-t)^{d_{1}}}\frac{Q_{S_{2}/J_{G_{2}}}(t)}{(1-t)^{d_{2}}}.\end{aligned}$$ Hence $d=d_1+d_2-2$ and $Q_{S/J_{G}}(t) = Q_{S_{1}/J_{G_{1}}}(t)Q_{S_{2}/J_{G_{2}}}(t)$. Also, $e(S/J_{G}) = Q_{S/J_{G}}(1) =e(S_{1}/J_{G_{1}})e(S_{2}/J_{G_{2}})$. The following is an immediate consequence of previous result. \[3.5\] Let $G= G_1 \cup \cdots \cup G_r$ be a connected graph. Let $S_i = K[x_j,y_j : j \in V(G_i)]$, for each $i \in [r]$. Then $${\operatorname{Hilb}}_{S/J_G}(t) = (1-t)^{2r-2} \underset{i\in[r]} \prod {\operatorname{Hilb}}_{S_i/J_{G_i}}(t).$$ In particular, $\dim(S/J_G)=\displaystyle \sum_{i\in [r]} \dim(S_i/J_{G_i})-2r+2$ and $e(S/J_G) = \underset{i \in[r]} \prod e(S_i/J_{G_i})$. Note that if $G$ is the complete graph on $n$ vertices, then $S/J_G$ is a determinantal ring. It follows from [@ch Corollary 1] that ${\operatorname{Hilb}}_{S/J_G}(t) = \frac{(n-1)t+1}{ (1-t)^{n+1}}$. Hence by [@bh Proposition 4.1.9], $e(S/J_G) = n$, $e_1(S/J_G) = n-1$ and for $2\leq i \leq n$, $e_i(S/J_G) =0$. \[3.6\] Let $G=P_n$ be the path graph on the vertex set $[n]$. Then $${\operatorname{Hilb}}_{S/{J_G}}(t)= \frac{(1+t)^{n-1}}{(1-t)^{n+1}},$$ $e_i(S/{J_G}) = \binom{n-1}{i} 2^{n-1-i} $ for $0 \leq i \leq n-1$ and $e_n(S/{J_G}) =0$. The result follows for $n=1,2$. For $n \geq 3$, $G$ is a decomposable graph with $n-1$ indecomposable subgraphs and each indecomposable subgraph of $G$ is $K_2$. Thus, by Corollary \[3.5\], ${\operatorname{Hilb}}_{S/J_G}(t) = (1-t)^{2n-4} \Big(\frac{t+1}{(1-t)^3}\Big)^{n-1}= \frac{(t+1)^{n-1}}{(1-t)^{n+1}}= \frac{Q(t)}{(1-t)^{n+1}}$. For $0\leq i \leq n-1$, $Q^{(i)}(t) = {i !}\binom{n-1}{i}(1+t)^{n-1-i}$, where $Q^{(i)}(t)$ denotes the $i$-th derivative of $Q(t)$ with respect to $t$. Hence it follows from [@bh Proposition 4.1.9] that for $0 \leq i \leq n-1$, $e_i(S/J_G)= \binom{n-1}{i}2^{n-1-i}$ and $e_n(S/J_G) =0$. It follows from the previous Corollary that the Hilbert Polynomial of $P_n$ is $P_{S/J_{P_n}}(X) = \underset{i \in [n]} \sum (-1)^{n-i} 2^{i-1} \binom{n-1}{i-1}\binom{X+i}{i}$. Now, we obtain the Hilbert series of $k$-handle lollipop graph. Let $m \geq 2$ and $K_m$ be the complete graph on $[m]$. Let $L_{m,r_1,\ldots ,r_k}$ denote the graph obtained by identifying one free vertex of each path $P_{r_1}, \ldots ,P_{r_k}$ with $k$ distinct vertices of $K_m$. Such a graph is called $k$-handle lollipop graph. For an induced subgraph $H$ of $G$, set $S_H = K[x_j,y_j :j \in V(H)]$. \[3.7\] Let $G=L_{m,r_{1},...,r_{k}}$ be a $k$-handle lollipop graph. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)=\frac{((m-1)t+1)(1+t)^{r-k}}{(1-t)^{m+r-k+1}}, \ where \ r=r_{1}+\cdots +r_{k}.$$ In particular, $\dim(S/J_G)= m+r-k+1$ and $e(S/J_{G})=2^{r-k}m$. Note that $ G=K_m\cup P_{r_{1}}\cup P_{r_{2}} \cdots \cup P_{r_{k}}$. Therefore by Corollary \[3.5\], $${\operatorname{Hilb}}_{S/J_{G}}(t)=(1-t)^{2k}{\operatorname{Hilb}}_{S_{K_m}/J_{K_m}}(t) \underset{i \in [k]}\prod{\operatorname{Hilb}}_{S_{P_{r_i}}/J_{P_{r_i}}}(t).$$ By Corollary \[3.6\], we have $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &= &(1-t)^{2k} \frac{((m-1)t+1)}{(1-t)^{m+1}}\underset {i \in [k]} \prod \frac{(1+t)^{r_i -1}}{(1-t)^{r_i+1}} \\&=& \frac{((m-1)t+1)(1+t)^{r-k}}{(1-t)^{m+r-k+1}}. \end{aligned}$$ It follows from Proposition \[3.7\] and [@bh Proposition 4.1.9] that if $G=L_{m,r_1,\ldots, r_k}$, then for $1\leq i \leq r-k$, $e_i(S/J_G) = (m-1)\binom{r-k}{ i-1}2^{r-k-i+1} + m\binom{r-k}{ i}2^{r-k-i}$, $e_{r-k+1}(S/J_G) = m-1$ and $e_j(S/J_G) =0$ for $j \geq r-k+2$. Let $G$ be a Cohen-Macaulay closed graph on vertex set $[n]$. It follows from [@her1 Theorem 3.1] that $G=K_{m_1} \cup \cdots \cup K_{m_r}$ for some $m_i \geq 2$. Therefore by using Corollary \[3.5\], we recover the result of Ene-Herzog-Hibi([@her1 page 66]), $${\operatorname{Hilb}}_{S/J_G}(t) = (1-t)^{2r-2}\underset{i \in [r]} \prod {\operatorname{Hilb}}_{S_{K_{m_i}}/J_{K_{m_i}}}(t) = \frac{\underset{i \in[r]} \prod ((m_i-1)t+1)}{(1-t)^{n+1}}$$ and $e(S/J_G) = \underset{i \in [r]} \prod m_i$. Hilbert series of Binomial edge ideal of Join of Graphs ======================================================== In this section we compute the Hilbert series of binomial edge ideal of join of two graphs. As consequences of this we obtain the Hilbert series of the binomial edge ideal of complete $k$-partite graphs, wheel graphs, fan graphs and multi-fan graphs. Let $H$ and $H'$ be two graphs with the vertex sets $[p]$ and $[q]$, respectively. The *join* of $H$ and $H'$, denoted by $H*H'$ is the graph with vertex set $[p] \sqcup [q]$ and the edge set $E(H*H')= E(H) \cup E(H') \cup \{\{i,j\}| i \in [p], j \in [q]\}$. The graph $G$, given on the right, is the join of the complete graph $K_4$ and the complement graph of the complete graph $K_3$. (4,2)– (2,2); (2,2)– (4,3); (4,3)– (2,4); (2,4)– (4,2); (4,4)– (2,2); (2,4)– (4,4); (2,4)– (3,4.5); (3,4.5)– (3,1.5); (3,1.5)– (2,2); (2,4)– (2,2); (2,2)– (3,4.5); (2,4)– (3,1.5); (4,4)– (3,4.5); (4,4)– (3,1.5); (4,3)– (3,1.5); (4,3)– (3,4.5); (4,2)– (3,1.5); (4,2)– (3,4.5); (2,4) circle (1.5pt); (3,4.5) circle (1.5pt); (3,1.5) circle (1.5pt); (2,2) circle (1.5pt); (4,4) circle (1.5pt); (4,3) circle (1.5pt); (4,2) circle (1.5pt); For rest of the section we use the following notation: Let $H$ and $H'$ be two graphs on vertex sets $[p]$ and $[q]$, respectively. Let $G=H*H^{'}$ be the join of $H$ and $H'$. Set $S_{H}=K[x_{i},y_{i}: i\in V(H)]$, $ S_{H^{'}}=K[z_{j},z'_{j}: j\in V(H^{'})]$ and $S=K[x_{i},y_{i},z_{j},z'_{j}: i\in V(H), j\in V(H^{'})]$. \[4.3\] Let $H$ and $H^{'}$ be two disconnected graphs on vertex sets $[p]$ and $[q]$, respectively. Let $G=H*H^{'}$ be the join of $H$ and $H^{'}$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t) ={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+{\operatorname{Hilb}}_{S_{H^{'}}/J_{H^{'}}}(t) + \frac{(p+q-1)t+1}{(1-t)^{p+q+1}}-\frac{(p-1)t+1}{(1-t)^{p+1}}-\frac{(q-1)t+1}{(1-t)^{q+1}}.$$ It follows from [@KM6 Proposition 4.14] that $$\mathscr{C}(G)=\{\phi\} \cup \{T\sqcup[q]:T \in \mathscr{C}(H)\} \cup\{[p] \sqcup T': T' \in \mathscr{C}(H')\}.$$ Let $$Q_{1}=\underset{{\substack{T\in \mathscr{C}(G)\\ [p] \subseteq T}}}{\cap} P_{T}(G)=(x_{i},y_{i} : i\in [p])+J_{H^{'}}$$ and $$Q_{2} = \underset{{\substack{T\in \mathscr{C}(G) \\ T\neq \phi \ [p] \nsubseteq T}}} {\cap}P_{T}(G) =(z_{i},w_{i} : i\in [q])+J_{H}.$$ Now, by [@HH1 Corollary 3.9], $$J_G= \underset{T \in \mathscr{C}(G)} \cap P_T(G) = P_{\phi}(G) \cap Q_1 \cap Q_2 = J_{K_{p+q}} \cap Q_1 \cap Q_2.$$ Set $Q_3 = J_{K_{p+q}} \cap Q_2$ and consider the short exact sequence, $$0\longrightarrow \frac{S}{Q_{3}} \longrightarrow \frac{S}{J_{K_{p+q}}} \oplus \frac{S}{Q_{2}} \longrightarrow \frac{S}{J_{K_{p+q}} +Q_2} \longrightarrow 0 .$$ Note that $J_{K_{p+q}}+Q_2= (z_{i},w_{i} : i\in [q])+J_{K_{p}}.$ Therefore $$\begin{aligned} {\operatorname{Hilb}}_{S/Q_{3}}(t) &={\operatorname{Hilb}}_{S/J_{K_{p+q}}}(t) + {\operatorname{Hilb}}_{S/Q_2}(t) - {\operatorname{Hilb}}_{S/({J_{K_p+q}+Q_2})}(t) \\ &=\frac{(p+q-1)t+1}{(1-t)^{p+q+1}} + {\operatorname{Hilb}}_{S_{H}/J_{H}}(t) - \frac{(p-1)t+1}{(1-t)^{p+1}}. \end{aligned}$$ Now, consider the short exact sequence $$0\longrightarrow \frac{S}{J_{G}}\longrightarrow \frac{S}{Q_{1}}\oplus \frac{S}{Q_{3}}\longrightarrow \frac{S}{Q_{1}+Q_{3}} \longrightarrow 0 .$$ Note that $Q_{1}+Q_{3} =(x_{i},y_{i} : i\in [p])+ J_{K_{q}}$. Therefore $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &= {\operatorname{Hilb}}_{S/Q_{1}}(t)+ {\operatorname{Hilb}}_{S/Q_{3}}(t)- {\operatorname{Hilb}}_{S/Q_{1}+Q_{3}}(t) \\ &= {\operatorname{Hilb}}_{S_{H^{'}}/J_{H^{'}}}(t)+{\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ \\ & \ \ \ \ \ \ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}} - \frac{(p-1)t+1}{(1-t)^{p+1}}-\frac{(q-1)t+1}{(1-t)^{q+1}}. \end{aligned}$$ Now we move on to study the join of a disconnected graph with a complete graph. We first identify the subsets with cutpoint property. \[4.5\] Let $H$ be a disconnected graph on the vertex set $[p]$. Let $G=H*K_q$ be the join of $H$ and $K_q$. Then $$\mathscr{C}(G)= \{\phi \} \cup \{T\sqcup [q]: T\in \mathscr{C}(H) \}.$$ The proof is immediate from [@KM6 Proposition 4.5]. We now compute the Hilbert series: \[4.6\] Let $H$ be a disconnected graph on the vertex set $[p]$. Let $G=H*K_{q}$ be the join of $H$ and the complete graph $K_q$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}} -\frac{(p-1)t+1}{(1-t)^{p+1}}.$$ By [@HH1 Corollary 3.9], $J_G = \underset{T \in \mathscr{C}(G)} \cap P_T(G)$. Let $Q = \underset{{\substack{T\in \mathscr{C}(G)\\ [q] \subseteq T}}} \cap P_{T}(G)=(z_{j},w_{j}: j\in [q])+J_{H}$. Now it follows from Lemma \[4.5\] that $J_G = P_{\phi}(G) \cap Q = J_{K_{p+q}} \cap Q$. Consider the short exact sequence $$0\longrightarrow \frac{S}{J_{G}}\longrightarrow \frac{S}{J_{K_{p+q}}}\oplus \frac{S}{Q}\longrightarrow \frac{S}{J_{K_{p+q}} +Q} \longrightarrow 0.$$ Note that $J_{K_{p+q}} +Q = (z_j,w_j : j \in [q] ) + J_{K_p}$. Therefore $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &= {\operatorname{Hilb}}_{S/Q}(t)+{\operatorname{Hilb}}_{S/ J_{K_{p+q}}}(t)- {\operatorname{Hilb}}_{S/({J_{K_{p+q}} +Q})}(t) \\ &= {\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}}- \frac{(p-1)t+1}{(1-t)^{p+1}}.\end{aligned}$$ If $H$ is a disconnected graph, then by taking $q=1$ in Theorem \[4.6\], we obtain the Hilbert series of the cone of $H$. Of particular interest is the case of multi-fan graph. Let $P_{p_1},\ldots, P_{p_r}$ denote paths on vertices $p_1, \ldots, p_r$ respectively. The the graph $\{v\} * (\sqcup_{i=1}^r P_{p_i})$ is called a multi-fan graph, denoted by $M_{p_1, \ldots, p_r}.$ As a consequence of Theorem \[4.6\], we obtain the Hilbert series of binomial edge ideal of the multi-fan graphs. Let $G=M_{p_1,\ldots,p_r}$ be the multi-fan graph with $r \geq 2$. Let $p=\underset{i\in [r]}\sum p_{i}$. Then $${\operatorname{Hilb}}_{S/J_G} = \frac{(1+t)^{p-r}}{(1-t)^{p+r}} + \frac{(p-1)t^2 +2t}{(1-t)^{p+2}}.$$ In particular, $ \dim(S/J_{G}) =p+r$ and $$e(S/J_{G}) = \begin{cases} {2}^{p-r} &\quad\text{if r}> 2, \\ {2}^{p-r}+{p+1} &\quad\text{if r} = 2. \end{cases}$$ Let $H = \underset{ i \in [r]} \sqcup P_{p_i}$. Since $r \geq 2$, by Theorem \[4.6\] and Corollary \[3.6\] $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{pt+1}{(1-t)^{p+2}} - \frac{(p-1)t+1}{(1-t)^{p+1}} \\ &= \prod_{i=1}^r {\operatorname{Hilb}}_{S_{P_{p_{i}}}/J_{P_{p_{i}}}}(t) +\frac{(p-1)t^2+2t}{(1-t)^{p+2}} \\ &= \prod_{i=1}^{r} \frac{(1+t)^{p_{i}-1}}{(1-t)^{p_{i}+1}} +\frac{(p-1)t^2+2t}{(1-t)^{p+2}} \\ &= \frac{(1+t)^{p-r}}{(1-t)^{p+r}}+\frac{(p-1)t^2+2t}{(1-t)^{p+2}}. \end{aligned}$$ We proceed to study the connected case now. To understand $H*K_q$, where $H$ is a connected graph, we need more tools. We first treat the case $q=1$. \[4.9\] Let $H$ be a connected graph on $[p]$ and $G=H*\{v\}$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t) ={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ \frac{(p-1)t^2+2t}{(1-t)^{p+2}} .$$ In particular,$$e(S/J_{G}) = \begin{cases} {e(S}_{H}/J_{H}) &\quad\text{if dim(S}_{H}/J_{H})\geq p+3, \\ {p+1}+{e(S}_{H}/J_{H}) &\quad\text{if dim(S}_{H}/J_{H})= p+2, \\ {p+1} &\quad\text{if dim(S}_\text{H}/J_\text{H})=p+1. \end{cases}$$ If $H=K_p$, then the result is immediate. Assume that $H \neq K_p$, so that $v$ is not a free vertex in $G$. By [@oh Lemma 4.8], $J_G = J_{G_v} \cap ((x_v,y_v) + J_{G \setminus v})$. Note that $G_v = K_{p+1}$ and ${G \setminus v}= H$. It follows from the short exact sequence $$0\longrightarrow \frac{S}{J_{G}}\longrightarrow \frac{S}{J_{K_{p+1}}}\oplus \frac{S}{(x_v,y_v)+J_H}\longrightarrow \frac{S}{(x_v,y_v)+J_{K_{p}}} \longrightarrow 0$$ that $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{pt+1}{(1-t)^{p+2}}-\frac{(p-1)t+1}{(1-t)^{p+1}} \\ &={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{(p-1)t^2+2t}{(1-t)^{p+2}}. \end{aligned}$$ The formula for the multiplicity follows directly from the Hilbert series expression. As an immediate consequence, we have: Let $G=W_{p+1}$ be the wheel graph on $[p+1]$ with $p \geq 3$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)=\frac{(1-t^2)(1+t)^{p-1}+(p-1)t^{p}+t^{p+1}}{(1-t)^{p+1}}+\frac{(p-1)t^2+2t}{(1-t)^{p+2}}.$$ In particular, $\dim(S/J_G)=p+2$ and $e(S/J_G)=p+1$. Let $H = C_p $ be the cycle graph on $[p]$. Note that $G=C_{p}*\{v\}$. Therefore by Theorem \[4.9\], $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{(p-1)t^2+2t}{(1-t)^{p+2}}.$$ Now the result follows from [@SZ Theorem 10]. We now define a new product which is required in the computation of the Hilbert series of join of two arbitrary graphs. Let $H$ be a graph on $[p]$ and $H'$ be a graph on $[q]$. Let $v_{1},...,v_{r}$ be vertices of $H'$ with $1 \leq r \leq q$. The graph with vertex set $[p]\sqcup [q]$ and edge set $$E(H)\sqcup \left\{ (u,v_{i}):u \in V(H),i=1,...,r \right\}\sqcup E(H')$$ is denoted by $H (*)^{r} H'$. Note that if $r= q$, then $H(*)^q H'$ is the join of $H$ and $H'$. For example, the graph $G$ given on the right side is the graph $P_3(*)^2K_4$. (1,4.5)– (1,3.5); (1,3.5)– (1,2.5); (1,2.5)– (2,3); (2,3)– (1,3.5); (1,3.5)– (2,4); (2,4)– (1,2.5); (1,4.5)– (2,3); (2,4)– (1,4.5); (2,4)– (3,4); (3,4)– (3,3); (3,3)– (2,3); (2,4)– (2,3); (2,3)– (3,4); (2,4)– (3,3); (2,4) circle (1.5pt); (2.2,4.36) node [$v_1$]{}; (3,4) circle (1.5pt); (3,3) circle (1.5pt); (2,3) circle (1.5pt); (2.18,2.76) node [$v_2$]{}; (1,4.5) circle (1.5pt); (1,3.5) circle (1.5pt); (1,2.5) circle (1.5pt); Our next aim is to compute the Hilbert series of $H(*)^r K_q.$ First we identify the sets with cut point property. \[4.11\] Let $H$ be a graph on $[p]$ and $G=H (*)^{r} K_{q},$ where $q\geq 2$ and $1 \leq r <q$. Then $$\mathscr{C}(G)=\{\phi \} \cup \{ T\sqcup \{v_{1},...,v_{r} \} : T\in \mathscr{C}(H) \}.$$ Let $T\in \mathscr{C}(G)$ and $T\neq \phi$. First note that $\{v_1,...,v_r\} \subseteq T$. For if $v_{i}\notin T$ for some $i$, then $G_{\bar{T}}$ is a connected graph, which contradicts the fact that $T\in \mathscr{C}(G)$. Let $T'=T \setminus \{v_{1},...,v_{r}\}$. We need to show that $T^{'}\in \mathscr{C}(H)$. If $T'=\phi$, then we are through. Assume that $T^{'}\neq \phi$. If $u\in T^{'}$, then $G[{\bar{T} \cup \{ u \}}]=H[{\bar{T'}\cup \{ u \}}] \sqcup K_{q-r}$. Since $u$ is a cut vertex of $G[{\bar{T} \cup \{ u \}}]$, it is a cut vertex of $H[{\bar{T'}\cup \{ u \}}]$. Thus $T' \in \mathscr{C}(H)$. Conversely, let $T=T' \sqcup \{v_{1},...,v_{r} \}$, where $T'\in \mathscr{C}(H)$. Note that for $ i \in [r]$, $G[{\bar{T}\cup \{ v_{i} \}}]=H[{\bar{T'}}] (*)^{1} K_{q-r+1}$. So $v_{i}$ is a cut vertex of $ G[{\bar{T}\cup \{ v_{i} \}}]$. If $T' = \phi$, then we are done. Assume that $T' \neq \phi$ and let $ u \in T'$. Since $G[{\bar{T}\cup \{ u \}}]=H[{\bar{T'}\cup \{ u \}}] \sqcup K_{q-r}$, $u$ is a cut vertex of $G[{\bar{T}\cup \{ u \}}]$ and $T\in \mathscr{C}(G)$. Hence the result follows. \[4.12\] Let $H$ be a graph on $[p]$ and $G=H (*)^{r} K_{q}$ be the graph on vertex set $[p]\sqcup [q]$ with $q\geq 2$ and $1 \leq r <q$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t) ={\operatorname{Hilb}}_{S_{H}/J_{H}}(t) \frac{(q-r-1)t+1}{(1-t)^{q-r+1}}+\frac{(p+q-1)t+1}{(1-t)^{p+q+1}}-\frac{(p+q-r-1)t+1}{(1-t)^{p+q-r+1}}.$$ By [@HH1 Corollary 3.9], $J_G = \underset{ T \in \mathscr{C}(G)} \cap P_T(G)$. Now, by Lemma \[4.11\], $$\begin{aligned} J_{G} &=P_{\phi}(G)\cap \underset{T \in \mathscr{C}(G) , T\neq \phi} {\cap} P_{T}(G) \\ &= J_{K_{p+q}} \cap ( (x_{v_{i}},y_{v_{i}}: i \in [r]) +J_{H}+J_{K_{q-r}}).\end{aligned}$$ Let $Q=(x_{v_{i}},y_{v_{i}}: i \in [r])+J_{H}+J_{K_{q-r}}$. Note that $J_{K_{p+q}}+Q=(x_{v_{i}},y_{v_{i}}: i \in [r])+J_{K_{p+q-r}}.$ Then it follows from the exact sequence below $$0\longrightarrow \frac{S}{J_{G}}\longrightarrow \frac{S}{J_{K_{p+q}}}\oplus \frac{S}{Q}\longrightarrow \frac{S}{J_{K_{p+q}}+Q}\longrightarrow 0$$ that $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) &= {\operatorname{Hilb}}_{S/Q}(t)+{\operatorname{Hilb}}_{S/J_{K_{p+q}}}(t) - {\operatorname{Hilb}}_{S/(J_{K_{p+q}}+Q)}(t) \\ &={\operatorname{Hilb}}_{S_{H}/J_{H}}(t) \frac{(q-r-1)t+1}{(1-t)^{q-r+1}}+\frac{(p+q-1)t+1}{(1-t)^{p+q+1}}-\frac{(p+q-r-1)t+1}{(1-t)^{p+q-r+1}}.\end{aligned}$$ Now we compute the Hilbert series of a q-cone over a connected graph $H$, i.e., join of $H$ and $q$ isolated vertices. We denote the graph of q isolated vertices by $K_q^c$. \[4.13\] Let $H$ be a connected graph on the vertex set $[p]$. Let $G=H*K_{q}^{c}$ be the join of $H$ and $K_{q}^{c}$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t) + {\operatorname{Hilb}}_{S/J_{K_{p}*{K_{q}^c}}}(t) - \frac{(p-1)t+1}{(1-t)^{p+1}}.$$ We proceed by induction on $q$. For $q=1$, $G=H*\{v\}$ and the result follows from Theorem \[4.9\]. Now assume that $q> 1$ and the result is true for $q-1$. Let $G=H*K_{q}^{c}$ and $V(K_{q}^{c})=\left\{v_{1},v_{2},...,v_{q} \right\}.$ Let $Q_{1}=(x_{v_{q}},y_{v_{q}})+J_{G\setminus v_{q}}$ and $Q_{2}=J_{G_{v_{q}}}$. Note that $Q_{1}+Q_{2}=(x_{v_{q}},y_{v_{q}})+J_{G_{v_{q}}\setminus v_{q}}$. Since $v_{q}$ is not a free vertex, it follows from [@oh Lemma 4.8] that $J_{G}=Q_{1}\cap Q_{2}$. Let $R=K[x_{i},y_{i} : i\in V(G) \ and \ i\neq v_{q}]$. Then by induction hypothesis $${\operatorname{Hilb}}_{S/{Q_{1}}}(t)={\operatorname{Hilb}}_{R/J_{G\setminus v_{q}}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ {\operatorname{Hilb}}_{R/J_{K_{p}*K_{q-1}^c}}(t) -\frac{(p-1)t+1}{(1-t)^{p+1}}.$$ Consider the short exact sequence $$0\longrightarrow \frac{S}{J_{G}}\longrightarrow \frac{S}{Q_{1}}\oplus \frac{S}{Q_{2}}\longrightarrow \frac{S}{Q_{1}+Q_{2}} \longrightarrow 0.$$ Note that $ G_{v_q} = K_p*K_{q}^c$, $G_{v_q} \setminus v_q = K_p*K_{q-1}^c$ and $ G \setminus v_q =H*K_{q-1}^c$. Therefore, it follows from the above short exact sequence that $${\operatorname{Hilb}}_{S/J_{G}}(t) ={\operatorname{Hilb}}_{S_{H}/J_{H}}(t) + {\operatorname{Hilb}}_{S/J_{K_{p}*K_{q}^c}}(t) - \frac{(p-1)t+1}{(1-t)^{p+1}}.$$ As a consequence of the Theorems \[4.6\], \[4.9\] and \[4.13\], we compute the Hilbert series of fan graphs. Let $G=P_{p}*K_{q}^{c}$ be the fan graph. Then $${\operatorname{Hilb}}_{S/J_{G}}(t) =\frac{(1+t)^{p-1}}{(1-t)^{p+1}}+ \frac{1}{(1-t)^{2q}}+ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}} - \frac{(p-1)t+1}{(1-t)^{p+1}}- \frac{(q-1)t+1}{(1-t)^{q+1}}.$$ In particular, $\dim(S/J_{G})=\max\{2q,p+q+1 \}$ and $$e(S/J_{G}) = \begin{cases} {1} & \quad\text{if p}+1 < q, \\ {p+q+1} &\quad\text{if p}+1 = q, \\ {p+q} &\quad\text{if p}+1 > q. \end{cases}$$ By Theorem \[4.13\], $${\operatorname{Hilb}}_{S/J_{G}}(t) = \frac{(1+t)^{p-1}}{(1-t)^{p+1}}+ {\operatorname{Hilb}}_{S/J_{K_{p}*K_{q}^c}}(t) - \frac{(p-1)t+1}{(1-t)^{p+1}}.$$ If $q=1$, then the assertion follows from the Theorem \[4.9\]. Assume that $q > 1$. Now by Theorem \[4.6\], the assertion follows. \[4.14\] Let $H$ be a connected graph on $[p]$ and $G=H*K_{q}$, where $q\geq 2$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{(p+q-1)t+1}{(1-t)^{p+q+1}} - \frac{(p-1)t+1}{(1-t)^{p+1}}.$$ Let $H^{'}=H\sqcup \{ w\}$, $G^{'}=H^{'}*K_{q}^{c}$, $ S_{H^{'}}=S_{H}[x_{w},y_{w}]$ and $S^{'}=S[x_{w},y_{w}]$. By Theorem \[4.3\], $${\operatorname{Hilb}}_{S^{'}/J_{G^{'}}}(t) =\frac{1}{(1-t)^2} {\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+\frac{1}{(1-t)^{2q}}+ \frac{(p+q)t+1}{(1-t)^{p+q+2}}-\frac{pt+1}{(1-t)^{p+2}}-\frac{(q-1)t+1}{(1-t)^{q+1}}.$$ Since $w$ is not a free vertex of $G^{'}$, it follows from [@oh Lemma 4.8] that $J_{G^{'}}= ((x_{w},y_{w})+J_{G^{'}\setminus w})\cap J_{G^{'}_{w}}$. Let $Q_{1}=(x_{w},y_{w})+J_{G^{'} \setminus w}$, $Q_{2}=J_{G^{'}_{w}}$. Note that $Q_{1}+Q_{2}=(x_{w},y_{w})+J_{G^{'}_{w}\setminus w}$, $G^{'}_{w}=H (*)^{q} K_{q+1}$, ${G^{'} \setminus w}=H*K_{q}^{c}$ and ${G^{'}_{w} \setminus w}=H*K_{q}=G$. Consider the short exact sequence $$0\longrightarrow \frac{S^{'}}{J_{G^{'}}}\longrightarrow \frac{S^{'}}{Q_{1}}\oplus \frac{S^{'}}{Q_{2}}\longrightarrow \frac{S^{'}}{Q_{1}+Q_{2}} \longrightarrow 0 .$$ Then by Theorem \[4.12\] and Theorem \[4.13\] $$\begin{aligned} {\operatorname{Hilb}}_{S/J_{G}}(t) & ={\operatorname{Hilb}}_{S/J_{H*K_{q}^{c}}}(t)+{\operatorname{Hilb}}_{S^{'}/J_{H (*)^{q} K_{q+1}}}(t)-{\operatorname{Hilb}}_{S^{'}/J_{G^{'}}}(t) \\ &= {\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}}- \frac{(p-1)t+1}{(1-t)^{p+1}}.\end{aligned}$$ Now we are ready to compute the Hilbert series of the join of two arbitrary graphs. \[4.15\] Let $H$ and $H'$ be graphs on vertex sets $[p]$ and $[q]$, respectively. Let $G=H*H^{'}$ be the join of $H$ and $H'$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+{\operatorname{Hilb}}_{S_{H^{'}}/J_{H^{'}}}(t)+ \frac{(p+q-1)t+1}{(1-t)^{p+q+1}}-\frac{(p-1)t+1}{(1-t)^{p+1}}-\frac{(q-1)t+1}{(1-t)^{q+1}} .$$ If both $H$ and $H'$ are disconnected graphs, then the assertion is proved in Theorem \[4.3\]. Assume that $H$ is connected and $H'$ is disconnected. Let $w$ be a new vertex. Set $H^{''}=H\sqcup \left\{ w \right\},$ $G^{'}=H^{''}*H^{'}$, $S_{H^{''}}=S_{H}[x_{w},y_{w}]$ and $S^{'}=S[x_{w},y_{w}]$. Let $Q_{1}=(x_{w},y_{w})+J_{G' \setminus w}$ and $Q_{2}=J_{G^{'}_{w}}$. Since $w$ is not a free vertex of $G'$, it follow from [@oh Lemma 4.8] that $J_{G^{'}}=Q_{1}\cap Q_{2}$. Note that $Q_{1}+Q_{2}=(x_{w},y_{w})+J_{G^{'}_{w}\setminus w}$, $G^{'}_{w}=H (*)^{q} K_{q+1},$ $G^{'}_{w}\setminus w=H*K_{q}$ and $G' \setminus w= G$. Therefore, it follows from the short exact sequence $$0\longrightarrow \frac{S^{'}}{J_{G^{'}}}\longrightarrow \frac{S^{'}}{Q_{1}}\oplus \frac{S^{'}}{Q_{2}}\longrightarrow \frac{S^{'}}{Q_{1}+Q_{2}} \longrightarrow 0$$ that $${\operatorname{Hilb}}_{S/J_{G}}(t)={\operatorname{Hilb}}_{S^{'}/J_{G^{'}}}(t)+{\operatorname{Hilb}}_{S/J_{H*K_q}}(t)-{\operatorname{Hilb}}_{S^{'}/J_{H (*)^q K_{q+1}}}(t) .$$ Note that $G'$ is join of two disconnected graphs. Now, the assertion follows from Theorems \[4.3\], \[4.12\] and \[4.14\].\ Assume now that both $H$ and $H'$ are connected. Let $u$ be a new vertex. Set $H^{'''}=H'\sqcup \left\{ u \right\},$ $G^{''}=H*H^{'''}$, $S_{H^{'''}}=S_{H'}[x_{u},y_{u}]$ and $S^{'}=S[x_{u},y_{u}]$. Let $Q_{1}=(x_{u},y_{u})+J_{G}$ and $Q_{2}=J_{G^{''}_{u}}$. It follow from the [@oh Lemma 4.8] that $J_{G^{''}}=Q_{1}\cap Q_{2}$. Consider the short exact sequence $$0\longrightarrow \frac{S^{'}}{J_{G^{''}}}\longrightarrow \frac{S^{'}}{Q_{1}}\oplus \frac{S^{'}}{Q_{2}}\longrightarrow \frac{S^{'}}{Q_{1}+Q_{2}} \longrightarrow 0 .$$ Note that $Q_{1}+Q_{2}=(x_{u},y_{u})+J_{G^{''}_{u}\setminus u}$, $G^{''}_{u}=H' (*)^{q} K_{q+1},$ $G^{''}_{u}\setminus u=H'*K_{q}$ and $G'' \setminus u= G$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)= {\operatorname{Hilb}}_{S^{'}/J_{G^{''}}}(t)+ {\operatorname{Hilb}}_{S/J_{H'*K_q}}(t)-{\operatorname{Hilb}}_{S^{'}/J_{H' (*)^q K_{q+1}}}(t) .$$ Since $G''$ is join of a connected graph $H$ and a disconnected graph $H'''$, by previous case $${\operatorname{Hilb}}_{S'/J_{G''}}(t)={\operatorname{Hilb}}_{S_{H}/J_{H}}(t)+{\operatorname{Hilb}}_{S_{H^{'''}}/J_{H^{'''}}}(t)+ \frac{(p+q)t+1}{(1-t)^{p+q+2}}-\frac{(p-1)t+1}{(1-t)^{p+1}}-\frac{qt+1}{(1-t)^{q+2}} .$$ Now, the assertion follows from Theorems \[4.12\] and \[4.14\]. As an immediate consequence, we are able to compute the Hilbert series, dimension and multiplicity of complete multi-partite graphs. Let $G=K_{p_1,\ldots,p_k}$ be the complete $k$-partite graph on the vertex set $[n] = [p_1] \sqcup \cdots \sqcup [p_k]$ with $p_1 \geq \cdots \geq p_k$. Then $${\operatorname{Hilb}}_{S/J_{G}}(t)= \underset{ i \in [k]} \sum \frac{1}{(1-t)^{2p_i}}- \underset{ i \in [k]} \sum \frac{(p_i-1)t+1}{(1-t)^{p_i +1}}+ \frac{(n-1)t+1}{(1-t)^{n+1}} .$$ In particular, $\dim(S/J_G)=\max\{2p_1, n+1 \}$ and $$e(S/J_{G}) = \begin{cases} {n} & \quad\text{if p}_1=1 \quad\text{or } 2p_1 <n+1, \\ {1} &\quad\text{if }2p_1 > n+1, \\ {n+1} &\quad\text{if }2p_1 = n+1. \end{cases}$$ Note that $G= (\cdots(K_{p_1}^c*K_{p_2}^c)*\cdots)* K_{p_r}^c$. Now the result follows by recursively applying Theorem \[4.15\]. [^1]: AMS Subject Classification (2010): 13A02, 05E40

ArXiv

--- abstract: 'From the Press-Schechter mass function and the empirical X-ray cluster luminosity-temperature ($L$-$T$) relation, we construct an X-ray cluster luminosity function that can be applied to the growing number of high-redshift, X-ray cluster luminosity catalogs to constrain cosmological parameters. In this paper, we apply this luminosity function to the [*Einstein*]{} Medium Sensitivity Survey (EMSS) and the [*ROSAT*]{} Brightest Cluster Sample (BCS) luminosity function to constrain the value of $\Omega_m$. In the case of the EMSS, we find a factor of 4 $-$ 5 fewer X-ray clusters at redshifts above $z = 0.4$ than below this redshift at luminosities above $L_X = 7 \times 10^{44}$ erg s$^{-1}$ (0.3 - 3.5 keV), which suggests that the X-ray cluster luminosity function has evolved above $L_\star$. At lower luminosities, this luminosity function evolves only minimally, if at all. Using Bayesian inference, we find that the degree of evolution at high luminosities suggests that $\Omega_m = 0.96^{+0.36}_{-0.32}$, given the best-fit $L$-$T$ relation of Reichart, Castander, & Nichol (1998). When we account for the uncertainty in how the empirical $L$-$T$ relation evolves with redshift, we find that $\Omega_m \approx 1.0 \pm 0.4$. However, it is unclear to what degree systematic effects may affect this and similarly obtained results.' author: - 'D. E. Reichart, R. C. Nichol, F. J. Castander, D. J. Burke, A. K. Romer, B. P. Holden, C. A. Collins, M. P. Ulmer' title: 'A Deficit of High-Redshift, High-Luminosity X-Ray Clusters:Evidence for a High Value of $\Omega_m$?' --- Introduction ============ The combination of the Press-Schechter mass function (e.g., Press & Schechter 1974; Oukbir & Blanchard 1992; Lacey & Cole 1993) and present and future X-ray cluster catalogs presents a unique opportunity to constrain the cosmological mass density parameter, $\Omega_m$. The Press-Schechter approach offers a number of advantages over various, more traditional methods of measuring this parameter. First of all, numerical simulations reproduce the Press-Schechter mass function to a high degree of accuracy (e.g., Eke 1996; Bryan & Norman 1998; Borgani 1998). Secondly, unlike methods that only probe $\Omega_m$ over small spatial scales $-$ methods that may be insensitive to an underlying, more uniformly distributed component of the dark matter $-$ the Press-Schechter approach probes $\Omega_m$ over the scales of X-ray cluster catalogs, which now have limiting redshifts of about unity. Thirdly, the Press-Schechter approach appears to be relatively insensitive to a cosmological constant (e.g., Henry 1997; Mathiesen & Evrard 1998; Eke 1998; Viana & Liddle 1998); consequently, a Press-Schechter-based determination of $\Omega_m$ might be compared to independent determinations of the deceleration parameter, for example, to constrain the cosmological constant. Finally and most importantly, a number of independent, high-redshift, X-ray cluster luminosity catalogs with well-understood selection functions are and will soon be available (see §4). We discuss potential problems with the Press-Schechter approach in §4. The archetypal X-ray cluster catalog is the X-ray cluster subsample of the [*Einstein*]{} Medium Sensitivity Survey, which we refer to here as the EMSS. A complete description of this sample and its selection criteria can be found in Henry (1992) (see also Gioia 1990b; Stocke 1991; Gioia & Luppino 1994). The original EMSS consists of 93 X-ray clusters, of which 67 have redshifts $z \ge 0.14$. Nichol (1997) updated this $z \ge 0.14$ subsample with 21 [*ROSAT*]{} luminosities and optical information from the literature. This revised EMSS consists of 64 X-ray clusters, of which 25 have been updated. The redshift and luminosity ranges of the revised EMSS are $0.140 \le z \le 0.823$ and $7.5 \times 10^{43} \le L_X \le 2.336 \times 10^{45}$ erg s$^{-1}$ (0.3 - 3.5 keV). The EMSS is of great importance because at present, it is the only X-ray cluster catalog that probes masses above $M_\star$ at high redshifts, where the Press-Schechter mass function is the most sensitive to $\Omega_m$ (see §3). Assuming that X-ray clusters correspond to virialized, dark matter halos, the Press-Schechter mass function describes how the X-ray-selected cluster mass function evolves with redshift. Unfortunately, X-ray-selected cluster mass catalogs that span sufficiently broad ranges in $M$ and $z$ to constrain $\Omega_m$ do not yet exist. Since the Press-Schechter mass function already assumes that X-ray clusters are virialized, one may convert this mass function to a temperature function with the virial theorem (see §2); however, X-ray cluster temperature catalogs that span sufficiently broad ranges in $T$ and $z$ to strongly constrain $\Omega_m$ also do not yet exist (Viana & Liddle 1998; Blanchard, Bartlett, & Sadat 1998; however, see Henry 1997; Eke 1998). However, several X-ray cluster luminosity catalogs span sufficiently broad ranges in $L$ and $z$ to strongly constrain $\Omega_m$, and the number of such catalogs is growing. However, to fit the Press-Schechter mass function to such catalogs, one must invoke a luminosity-temperature ($L$-$T$) relation in addition to the virial theorem. Theoretically, a wide variety of $L$-$T$ relations have been proposed (e.g., Kaiser 1986; Evrard & Henry 1991; Kaiser 1991); consequently, the $L$-$T$ relation should be determined empirically. Until recently, the $L$-$T$ relations of temperature catalogs have suffered from much scatter (e.g., Edge & Stewart 1991; David 1993; Mushotzky & Scharf 1997); however, recently Markevitch (1998), Allen & Fabian (1998), and Arnaud & Evrard (1998) have published temperature catalogs with temperatures and luminosities that have either been corrected for, or avoided the effects of cooling flows; the result is a significant reduction of this scatter. Using Bayesian inference, Reichart, Castander, & Nichol (1998) have constrained the slope and the evolution of the empirical $L$-$T$ relation for the luminosity range $10^{44.5}$ erg s$^{-1}$ $\la$ $L_{bol} \la 10^{46.5}$ erg s$^{-1}$ and the redshift range $z \la 0.5$ from the Markevitch (1998) and Allen & Fabian (1998) catalogs. This latter work may be the key to determining cosmological parameters with X-ray cluster catalogs: given a well-constrained $L$-$T$ relation, the Press-Schechter mass function may be fitted to the growing number of independent, high-redshift, high-luminosity X-ray cluster catalogs with well-understood selection functions. Using the cooling flow corrected $L$-$T$ relation of Reichart, Castander, & Nichol (1998), we do this to the EMSS and the [*ROSAT*]{} Brightest Cluster Sample (BCS) luminosity function of Ebeling (1997) in §3. In §2, we model the X-ray cluster luminosity function; in §4, we draw conclusions and discuss future applications of this luminosity function to the Southern SHARC and the Bright SHARC. The Model ========= The X-ray Cluster Luminosity Function ------------------------------------- Assuming that X-ray clusters correspond to virialized, dark matter halos, we model the comoving number density of X-ray clusters with the Press-Schechter mass function, which is given by (e.g., Lacey & Cole 1993) $$\frac{dn_c(M,z)}{dM} = -\sqrt{\frac{2}{\pi}} \frac{\bar{\rho}_0}{M^2} \frac{d\ln{\sigma_0}(M)}{d\ln{M}} \frac{\delta_{c0}(z)}{\sigma_0(M)} \exp{\left[\frac{-\delta_{c0}^2(z)}{2\sigma_0^2(M)}\right]}, \label{ps}$$ where $n_c(M,z)$ is the comoving number density of X-ray clusters of mass $M$ at redshift $z$, $\delta_{c0}(z)$ is the present linear theory overdensity of perturbations that collapsed and virialized at redshift $z$, $\sigma_0(M)$ is the present linear theory variance of the mass density fluctuation power spectrum filtered on mass scale $M$, and $\bar{\rho}_0$ is the present mean mass density of the universe. In the case of zero cosmological constant, which we assume throughout this paper, the present overdensity is given by (Lacey & Cole 1993; Peebles 1980) $$\delta_{c0}(z) = \cases{\frac{3}{2} D(0) \left(\left(\frac{2\pi}{\sinh{\eta}-\eta}\right)^{2/3}+1\right) & ($\Omega_m < 1$) \cr \frac{3}{20} (12\pi)^{2/3} (1+z) & ($\Omega_m = 1$) \cr \frac{3}{2} D(0) \left(\left(\frac{2\pi}{\eta-\sin{\eta}}\right)^{2/3}-1\right) & ($\Omega_m > 1$)}, \label{del}$$ where $$D(0) = \cases{1 + \frac{3}{x_0} + \frac{3\sqrt{1+x_0}}{x_0^{3/2}} \ln{(\sqrt{1+x_0}-\sqrt{x_0})} & ($\Omega_m < 1$) \cr -1 + \frac{3}{x_0} - \frac{3\sqrt{1+x_0}}{x_0^{3/2}} \tan^{-1}{\sqrt{\frac{x_0}{1-x_0}}} & ($\Omega_m > 1$)},$$ $$x_0 = \left|\Omega_m^{-1} - 1\right|,$$ $$\eta = \cases{\cosh^{-1}{\left(\frac{2}{\Omega(z)} - 1\right)} & ($\Omega_m < 1$) \cr \cos^{-1}{\left(1 - \frac{2}{\Omega(z)}\right)} & ($\Omega_m > 1$)},$$ and $$\Omega(z) = \frac{\Omega_m (1+z)}{1 + \Omega_m z}.$$ We assume a scale-free mass density fluctuation power spectrum of power law index $n$, so the present variance is given by (e.g., Lacey & Cole 1993) $$\sigma_0(M) = \sigma_8 \left(\frac{M}{M_8}\right)^{-\frac{3+n}{6}},$$ where $\sigma_8$ is the amplitude of the mass density fluctuation power spectrum over spheres of radius $8 h^{-1}$ Mpc, and $M_8$ is the mean mass within these spheres. We now convert equation (\[ps\]) from a mass function to an appropriately defined luminosity function. Following the notation of Mathiesen & Evrard (1998), we begin by assuming that X-ray clusters’ bolometric luminosities scale as power laws in mass and redshift: $$L_{bol} \propto M^p (1+z)^s. \label{power}$$ As did Henry (1992) in the case of the EMSS, we find that the fractions of this luminosity that fall into the EMSS band of 0.3 - 3.5 keV and the BCS band of 0.1 - 2.4 keV are well approximated by power laws in X-ray cluster temperature: $$L_X = f_X T^{-\beta} L_{bol}, \label{fjc}$$ where $f_X = 0.989 \pm 0.014$ and $\beta = 0.407 \pm 0.008$ for the representative temperature range of the EMSS ($3 \la T \la 10$ keV), and $f_X = 1.033 \pm 0.012$ and $\beta = 0.472 \pm 0.008$ for the representative temperature range of the BCS ($1.5 \la T \la 12$ keV), where temperature is measured in keV. At lower temperatures, however, these approximations quickly fail. Equation (\[fjc\]) is independent of redshift because X-ray cluster luminosities are measured in the source frame. The temperature dependence introduced by equation (\[fjc\]) is removed with the virial theorem: $$T \propto M^\frac{2}{3} (1+z). \label{virial}$$ Technically, this expression holds only when $\Omega_m = 1$; however, we show in §4 that generalizing this expression has little effect upon our results. Together, equations (\[power\]), (\[fjc\]), and (\[virial\]) yield the following expression that relates an X-ray cluster’s mass to its observed luminosity, $L_X$: $$L_X \propto f_XM^{p-\frac{2\beta}{3}}(1+z)^{s-\beta}. \label{factor}$$ Substitution of equation (\[factor\]) into equation (\[ps\]) yields the following luminosity function: $$\frac{dn_c(L_X,z)}{dL_X} = f(z)L_X^{-1-\frac{3-n}{2(3p-2\beta)}} \exp{\left[-g(z)L_X^{\frac{3+n}{3p-2\beta}}\right]}, \label{ps2}$$ where $$f(z) = af_X^{\frac{3-n}{2(3p-2\beta)}}(1+z)^{\frac{(s-\beta)(3-n)}{2(3p-2\beta)}}\delta_{c0}(z),$$ $$g(z) = cf_X^{-\frac{3+n}{3p-2\beta)}}(1+z)^{-\frac{(s-\beta)(3+n)}{3p-2\beta}} \delta_{c0}^2(z),$$ and $a$ and $c$ depend upon $\sigma_8$ and the factor of proportionality of equation (\[factor\]). Instead of trying to model this factor of proportionality and fitting to $\sigma_8$, we simply fit to such degenerate, or grouped, combinations of these parameters in this paper (see §2.2). Since luminosities are computed from measured fluxes and redshifts, $L_X$ is a function of $H_0$ and $\Omega_m$. With one exception, all dependences upon $H_0$ can be grouped into the parameters $a$ and $c$, and this exception is noted in §2.2. The EMSS and the BCS provide luminosities in their respective X-ray bands that have been computed for $H_0 = 50$ km s$^{-1}$ Mpc$^{-1}$ and $\Omega_m = 1$; we denote these luminosities by $L_1$. The relationship between $L_X$ and $L_1$ is given by $$L_X = x(z)L_1, \label{lxl}$$ where $$x(z) = \cases{\left(\frac{d_L(\Omega_m)}{d_L(\Omega_m = 1)}\right)^2 \frac{f_F(d_A(\Omega_m = 1))}{f_F(d_A(\Omega_m))} & ({\it Einstein}) \cr \left(\frac{d_L(\Omega_m)}{d_L(\Omega_m = 1))}\right)^2 & ({\it ROSAT})}, \label{x}$$ $d_L(\Omega_m)$ is luminosity distance, $d_A(\Omega_m)$ is angular diameter distance, and $f_F(d_A(\Omega_m))$ is the fraction of an X-ray cluster’s flux that is detected in the 2.4 $\times$ 2.4 detect cell of the original EMSS.[^1] A complete description of this quantity can be found in Henry (1992). Since [*ROSAT*]{} measures total fluxes, $f_F(d_A(\Omega_m)) = 1$ here. So in the case of the BCS, the latter expression applies. However, the revised EMSS subsample that we fit to in §3.3 is a combination of 43 [*Einstein*]{} luminosities and 18 [*ROSAT*]{} luminosities (see §3). Fortunately, 36 of the 43 [*Einstein*]{} clusters have redshifts of $z < 0.33$, and 6 of the remaining 7 clusters have redshifts of $z < 0.47$. At these redshifts, the ratio of fractional fluxes in the former expression for $x(z)$ is within a few percent of unity for a wide range of values of $\Omega_m$. Furthermore, [*ROSAT*]{} luminosities are available for 7 of the 8 clusters that carry the majority of the weight in the fits of §3.3, and the remaining [*Einstein*]{} cluster is at a redshift of $z = 0.259$. Consequently, we also use the latter expression for $x(z)$ in the case of the EMSS. Besides, we show in §3 that the sensitivity of $x(z)$ to $\Omega_m$ plays only a tertiary role in the determination of this parameter. Substitution of equation (\[lxl\]) into equation (\[ps2\]) allows the luminosity function to be fitted to $L_1$ data without loss of generality: $$\frac{dn_c(L_1,z)}{dL_1} = f(z)L_1^{-1-\frac{3-n}{2(3p-2\beta)}} \exp{\left[-g(z )L_1^{\frac{3+n}{3p-2\beta}}\right]}, \label{ps3}$$ where $$f(z) = af_X^{\frac{3-n}{2(3p-2\beta)}}(1+z)^{\frac{(s-\beta)(3-n)}{2(3p-2\beta)}} \delta_{c0}(z) x^{-\frac{3-n }{2(3p-2\beta)}}(z), \label{f}$$ and $$g(z) = cf_X^{-\frac{3+n}{3p-2\beta}}(1+z)^{-\frac{(s-\beta)(3+n)}{3p-2\beta}} \delta_{c0}^2(z) x^{\frac{3+n} {3p-2\beta}}(z). \label{g}$$ The Selection Function ---------------------- Let $A(L_1,z)$ be the area of the sky that an X-ray survey samples at redshift $z$ as a function of luminosity $L_1$. In the case of the EMSS, this quantity is given by (Avni & Bahcall 1980; Henry 1992; Nichol 1997) $$A(L_1,z) = A(F_{lim} = F(L_1,z)), \label{area}$$ where $A(F_{lim})$ is the area of the sky that the EMSS surveyed below sensitivity limit $F_{lim}$ (see Henry 1992), $$F(L_1,z) = \frac{f_F(d_A(z))}{k(z)} \frac{h_{50}^2L_1}{4\pi d_L^2(z)}, \label{flux}$$ and $k(z)$ is the k-correction from the observer frame to the source frame for a $T = 6$ keV X-ray cluster; the exact dependence of $k(z)$ upon X-ray cluster temperature can be ignored for the representative temperature range of the EMSS. For the EMSS, we have computed $A(L_1,z)$ for 41 values of $L_1$ between $10^{43.5}$ and $10^{45.5}$ erg s$^{-1}$ (0.3 - 3.5 keV) and for $\Omega_m = 0$, 0.5, 1, and 1.5. For intermediate values of $L_1$ and $\Omega_m$, we use linear interpolation between $43.5 < \log{L_1} < 45.5$ and $0 < \Omega_m < 1.5$. The cases of $\Omega_m = 0$ and 1 are plotted in Figure 1. The dependence of $A(L_1,z)$ upon $H_0$ cannot be grouped into the parameters $a$ and $c$, unlike all of the other $H_0$ dependences in this analysis (§2.1). Instead of making $H_0$ a free parameter, we fix $H_0 = 50$ km s$^{-1}$ Mpc$^{-1}$ in this paper. However, if others wish to be more general, they need only consider the $H_0$ dependence of the selection function. The case of the BCS is treated separately in §3.2. The total number of X-ray clusters observed between luminosity and redshift limits $L_l < L_1 < L_u$ and $z_l < z < z_u$, i.e., the cumulative luminosity function, is given by $$N(L_l,L_u;z_l,z_u) = \int_{L_l}^{L_u}\int_{z_l}^{z_u}A(L_1,z)\frac{dn_c(L_1,z)}{dL_1}dL_1dV(z), \label{ps4}$$ where $$dV(z) = \frac{4c^3dz}{H_0^3\Omega_m^4(1+z)^3}\frac{(\Omega_m z + (\Omega_m-2)((\Omega_m z+1)^\frac{1}{2}-1))^2}{(1+\Omega_m z)^\frac{1}{2}} \label{vol}$$ is the comoving volume element. Hence, our model (luminosity function $+$ selection function) consists of nine parameters: $H_0$, $f_X$, $\beta$, $p$, $s$, $a$, $c$, $n$, and $\Omega_m$. We have fixed the value of $H_0$, and we have tightly constrained the values of $f_X$ and $\beta$ for the EMSS and the BCS (§2.1). In §3.1, we adopt values of $p$ and $s$ from the Bayesian inference analysis of the X-ray cluster $L$-$T$ relation of Reichart, Castander, & Nichol (1998), and in §3.3, the normalization parameter, $a$, drops out of the Bayesian inference analysis of this paper. This leaves three parameters: $\Omega_m$, $n$, and the grouped parameter $c$, which depends upon $\sigma_8$, the proportionality constant of equation (\[factor\]), and $H_0$. We determine credible intervals for the values of these three parameters in §3.3. Bayesian Inference ================== If properly applied, equation (\[ps4\]) can be an effective probe of $\Omega_m$. In this cumulative luminosity function, $\delta_{c0}(z)$, $A(L_1,z)$, $x(z)$, and $dV(z)$ depend upon $\Omega_m$. We now consider how sensitive each of these quantities is to $\Omega_m$. The comoving volume element $dV(z)$, is approximately given by (equation (\[vol\])) $$\frac{dV(z)}{dz} \propto \cases{\frac{z^{2.26}}{(1+z)^3} & ($\Omega_m = 0$) \cr \frac{z^{1.99}}{(1+z)^3} & ($\Omega_m = 1$)},$$ where the indices apply in the redshift range $0.14 < z < 0.6$. The luminosity conversion expression, $x(z)$, is approximately given by (equation (\[x\])) $$x(z) \approx 1 - \frac{1-\Omega_m}{4}z,$$ and in equation (\[ps3\]), it is always raised to a power that is between $\approx$ $-1.5$ and $\approx$ 0.5. Consequently, $dV(z)$ and $x(z)$ contribute only weak dependences upon $\Omega_m$ to equation (\[ps4\]). The present overdensity is a stronger function of $\Omega_m$ (equation (\[del\])): $$\delta_{c0}(z) = \cases{1.5 & ($\Omega_m = 0$) \cr 1.69(1+z) & ($\Omega_m = 1$)}. \label{del2}$$ Since this expression appears to the second power in the exponential cutoff of equation (\[ps3\]), it contributes a significant dependence upon $\Omega_m$ to the cumulative luminosity function. For example, if the luminosity function is observed to cut off prematurely at higher redshifts, i.e., if there is a deficit of high-redshift, luminous X-ray clusters, then higher values of $\Omega_m$ are favored. However, if little or no evolution is manifest in the observed X-ray cluster luminosity function, particularly above $L_\star$[^2], then lower values of $\Omega_m$ are favored. The surveyed area, $A(L_1,z)$, contributes a different type of dependence upon $\Omega_m$ to the cumulative luminosity function. In the case of the EMSS (see Figure 1), this dependence is negligible at low luminosities and redshifts. However, at luminosities $\ga L_\star$, $A(L_1,z)$ is a non-negligible, increasing function of $\Omega_m$ at sufficiently high redshifts. For example, in the case of a $L_1 = 10^{45}$ erg s$^{-1}$, z = 0.8 EMSS cluster, $A(L_1,z)$ is roughly twice as large in an $\Omega_m = 1$ universe than it is in an $\Omega_m = 0$ universe. Although this effect is suppressed by the exponential cutoff of equation (\[ps3\]) above $L_\star$, this effect is amplified about $L_\star$ by the fact that the luminosity function itself is a non-negligible, increasing function of $\Omega_m$ at luminosities $\la L_\star$ at these high redshifts (see §3.3). Consequently, we find that an overabundance of high-redshift, $\sim L_\star$ EMSS clusters favors higher values of $\Omega_m$ and not lower values of this parameter as is generally thought. We return to this idea in §3.3. Consider first the case of X-ray cluster luminosity data that lies within a narrow redshift band of effective redshift $z_{eff}$. Then, up to a factor of $A(L_1,z_{eff})dV(z_{eff})/dz$, the integrand of equation (\[ps4\]) (equation (\[ps3\])) is simply a power law in luminosity with an $\Omega_m$-dependent exponential cutoff. This exponential cutoff is a function of the parameters $f_X$, $\beta$, $p$, $n$, and $g_{eff} = g(z_{eff})$, which itself is a function of $\Omega_m$ (see below). We have already constrained the values of $f_X$ and $\beta$ (§2.1), and we adopt the Reichart, Castander, & Nichol (1998) value of $p$, as well as that of $s$, in §3.1. However, there are too few high-luminosity X-ray clusters to simultaneously constrain $n$ and $g_{eff}$. Fortunately, $n$ is also constrained by the low-luminosity, power-law limit of equation (\[ps3\]) for which data is more plentiful. Consequently, by fitting this luminosity function to data of this type, $n$ and $g_{eff}$ can be jointly constrained. By equation (\[g\]), the parameter $g_{eff}$ is a function of $z_{eff}$ and the parameters $f_X$, $\beta$, $p$, $s$, $n$, $c$, and $\Omega_m$. The effective redshift is a given and the parameters $f_X$, $\beta$, $p$, $s$, and $n$ can be constrained as described above. However, since $z \approx z_{eff}$, a constant, the parameters $c$ and $\Omega_m$ are degenerate; consequently, $\Omega_m$ can only be constrained if the value of $c$ is otherwise known, i.e., if the values of $\sigma_8$, the factor of proportionality of equation (\[factor\]), and $H_0$ are otherwise known (§2.2). Even in the event that X-ray cluster mass data is used instead of luminosity data, any fitted value of $\Omega_m$ will still depend strongly upon the assumed value of $\sigma_8$, as well as upon the assumed value of $n$, since mass data is not yet plentiful enough for the Press-Schechter mass function to constrain these parameters. However, now consider X-ray cluster luminosity data that spans a breadth of redshifts. Instead of constraining the single parameter $g_{eff}$, one instead constrains a distribution of such parameters with redshift, i.e., $g(z)$. The normalization of this distribution is $c$ and its shape yields $\Omega_m$ since the parameters $f_X$, $\beta$, $p$, $s$, and $n$ are otherwise constrained. Consequently, by fitting equation (\[ps4\]) to the EMSS, which spans a breadth of luminosities and redshifts, the parameters $n$, $c$, and $\Omega_m$ can be jointly constrained regardless of the values of $\sigma_8$ and the factor of proportionality of equation (\[factor\]) (but not regardless of the value of $H_0$ since $A(L_1,z)$ is a function of this parameter (§2.2)). We do this for the EMSS in §3.3. In §3.2, we better constrain the parameters $n$ and $c$ (actually, $g_{eff}$) with the local ($z_{eff} \sim 0.1$) luminosity function of the BCS. First however, we discuss the cooling flow corrected $L$-$T$ relation of Reichart, Castander, & Nichol (1998) and its implied values of $p$ and $s$, in §3.1. The $L$-$T$ Relation -------------------- The combination of equations (\[power\]) and (\[virial\]) yields the $L$-$T$ relation: $$L_{bol} \propto T^\frac{3p}{2}(1+z)^{s-\frac{3p}{2}}. \label{lt}$$ Reichart, Castander, & Nichol (1998) have constrained the slope and the evolution of equation (\[lt\]) using the cooling flow corrected X-ray cluster temperature catalogs of Markevitch (1998) and Allen & Fabian (1998), and Bayesian inference. For the luminosity ranges of the EMSS and the BCS, and the redshift range $z \la 0.5$, they find that $p = 1.86^{+0.10}_{-0.10}$ and $s = (3.77 - 0.63\Omega_m)^{+0.48}_{-1.22}$. However, when using the $L$-$T$ relation to fit the Press-Schechter mass function to X-ray cluster luminosity catalogs (1) that are not cooling flow corrected, and (2) for which X-ray cluster photon count rates have been converted to fluxes and luminosities by assuming a $T = 6$ keV thermal bremsstrahlung spectrum, they find that one should use $p = 1.77^{+0.16}_{-0.13}$ and $s = (3.14 - 0.65\Omega_m)^{+0.88}_{-0.86}$. In the case of the EMSS, this latter case applies. In this paper, we adopt the best-fit values: $p = 1.77$ and $s = 3.14 - 0.65\Omega_m$. The parameter $p$ is well-constrained, and in §3.3, we show that the uncertainty in the value of $s$ does not significantly affect our results. In the case of the BCS, luminosities are not cooling flow corrected, but they are not determined by assuming $T = 6$ keV for each X-ray cluster; instead, luminosities are determined by additionally requiring that each X-ray cluster satisfy a $L$-$T$ relation (Ebeling 1997). From Figure 1 of Reichart, Castander, & Nichol (1998), it is apparent that the assumption of $T = 6$ keV more strongly affects the value of $p$ than does the use of cooling flow corrected luminosities. Consequently, in this paper, we adopt $p = 1.86$ in the case of the BCS. We note, however, that such minor variations in this parameter do not significantly affect our results. Since the BCS is a local ($z_{eff} \sim 0.1$) catalog, the value of $s$ is unimportant in this case (see §3.2). The [*ROSAT*]{} BCS ------------------- The [*ROSAT*]{} BCS is a flux-limited sample of 199 bright X-ray clusters. A complete description of this sample and its selection criteria can be found in Ebeling (1997). The redshift range of the BCS is $z < 0.3$; however, most of the BCS clusters have redshifts of $z < 0.2$, and the effective redshift (see §3) of the sample is $z_{eff} \sim 0.1$. Consequently, the BCS samples the X-ray cluster population of the local universe. Although such a sample may not have enough redshift leverage to adequately probe $\Omega_m$, its large size makes it an excellent sample to constrain the parameters $n$ and $c$ (actually, $g_{eff}$, §3). These constraints can then be combined with the EMSS results of §3.3 to better constrain $\Omega_m$. However, since the BCS is not yet publicly available, we settle here for a simplified analysis of the binned BCS luminosity function of Figure 1 of Ebeling (1997), which we replot in Figure 2. Since the BCS spans a relatively narrow band of redshifts, we let $f(z) = f(z_{eff}) = f_{eff}$ and $g(z) = g(z_{eff}) = g_{eff}$ in equation (\[ps3\]). This approximation is reasonable, unless the value of $\Omega_m$ is high, in which case one expects a lower comoving number density of X-ray clusters in the highest-luminosity bins. This is because the highest-luminosity bins more strongly sample the highest-redshift BCS clusters than do the lower-luminosity bins. Given this approximation $-$ that $z = z_{eff}$, a constant $-$ equation (\[ps3\]) can only reproduce such a high-luminosity roll-over of the luminosity function by favoring an artificially high value of $n$. To safeguard against this potential bias, we first fit equation (\[ps3\]) to all 12 luminosity bins, then to all but the highest-luminosity bin, then to all but the 2 highest-luminosity bins, etc., until the fitted values of $n$ and $c$ do not change appreciably from fit to fit. Also, since Ebeling (1997) have already corrected this luminosity function for sample completeness, we set $A(L_1,z_{eff}) = 1$. Here, we ignore the dependence that this quantity has upon $\Omega_m$, which is a reasonable approximation since $z_{eff} < 0.3$. We find that only the highest-luminosity bin noticeably changes our results. When we fit equation (\[ps3\]) to all 12 luminosity bins, we find that $n = -0.47^{+0.32}_{-0.31}$ and $g_{eff} = 0.90^{+0.28}_{-0.14}$, where we have assumed a flat prior probability distribution between $-3 < n < 0$ and $0 < g_{eff} < 10$, and the likelihood function is given by $e^{-\chi^2/2}$.[^3] When we fit equation (\[ps3\]) to all but the highest-luminosity bin, we find that $n = -1.83^{+0.85}_{-0.15}$ and $g_{eff} = 1.20^{+0.70}_{-0.60}$. Ignoring additional high-luminosity bins does not appreciably change this result; consequently, in this paper, we use all but the highest-luminosity bin to determine a constraint from the BCS. We will explore what this highest-luminosity bin implies for the value of $\Omega_m$ in a later paper. In Figure 3, we plot the 1, 2, and 3 $\sigma$ credible regions in the $n-g_{eff}$ plane for both of the above fits. We do not plot credible regions in the $n-c$ plane because, by equation (\[g\]), the parameter $c$ is a degenerate function of the parameters $g_{eff}$ and $\Omega_m$ (see §3); however, credible regions in the $n-c$ plane are easily recovered, given a value of $\Omega_m$ (see §3.3). In Figure 2, we also plot the best-fit luminosity functions of these fits. The EMSS -------- As described in §3, the breadth of the luminosity and redshift ranges of the EMSS makes this catalog an ideal sample with which to probe $\Omega_m$. In Figure 4, we plot the $L_1-z$ distribution of the revised EMSS of Nichol (1997), as well as the $z < 0.14$ portion of the original EMSS X-ray cluster subsample. The solid curves are contours of constant sampled differential volume, i.e., $A(L_1,z)dV(z)/dz =$ constant. From left to right, these contours are equally spaced from zero (zero contour not shown). If the X-ray cluster luminosity function has not evolved over the redshift range of the EMSS, then at each luminosity, most of the observed X-ray clusters would be where most of the sampled differential volume is. This appears to be the case below $L_1 \sim 7 \times 10^{44}$ erg s$^{-1}$ (0.3 - 3.5 keV), which demonstrates agreement with the results of Collins (1997), Nichol (1997), Burke (1997), Rosati (1998), Jones (1998), and Vikhlinin (1998a), i.e., that the X-ray cluster luminosity function evolves only minimally, if at all, below $L_\star$ (see §4). However, as was originally found by Gioia (1990a) (see also Henry 1992), there appears to be a deficit of high-redshift X-ray clusters above $L_\star$. For example, given that six X-ray clusters were detected between $0.14 < z < 0.4$ at $L_1 > 7 \times 10^{44}$ erg s$^{-1}$ (0.3 - 3.5 keV), one would expect $\sim 9$ X-ray clusters between $0.4 < z < 0.6$ and $\sim 21$ X-ray clusters between $0.4 < z < 0.9$ at these luminosities if the luminosity function were not evolving, yet only 2 and 4 clusters were detected in these redshift ranges, respectively. Hence, we find a factor of 4 $-$ 5 fewer X-ray clusters at redshifts above $z = 0.4$ than below this redshift at these luminosities. This is only possible at the $\approx 1 - 2$% level with the no evolution model. This suggests that high values of $\Omega_m$ may be favored (§3); however, we offer alternative interpretations of this deficit in §4. By Bayes’ theorem, the posterior probability distribution for $\Omega_m$, $n$, and $c$, $P(\Omega_m,n,c)$, is given by normalizing the product of the prior probability distribution and the likelihood function (e.g., Gregory & Loredo 1992). Here, we assume a flat prior probability distribution between $0 < \Omega_m < 1.5$, $-3 < n < 0$, and $0 < c < 3$. The likelihood function, ${\cal{L}}(\Omega_m,n,c)$, is given by (e.g., Cash 1979) $${\cal{L}}(\Omega_m,n,c) = \prod_{i=1}^{N_{tot}}P(L_{1,i},z_i|\Omega_m,n,c), \label{likely}$$ where $P(L_{1,i},z_i|\Omega_m,n,c)$ is the probability that the $i$th X-ray cluster fits our model, given values of $\Omega_m$, $n$, and $c$. For our model (equation (\[ps4\])), this probability is given by (e.g., Cash 1979) $$P(L_1,z|\Omega_m,n,c) = \frac{A(L_1,z)}{N(L_l,L_u;z_l,z_u)} \frac{dV(z)}{dz} \frac{dn_c(L_1,z)}{dL_1}.$$ Since both $dn_c(L_1,z)/dL_1$ and $N(L_l,L_u;z_l,z_u)$ are proportional to the normalization parameter $a$, our results are clearly independent of this parameter. In equation (\[likely\]), $N_{tot}$ is the total number of X-ray clusters in the same region of the $L_1-z$ plane as that over which $N(L_l,L_u;z_l,z_u)$ is defined. This region should be as broad as is reasonably possible and it need not be rectangular, as the simple integration limits of equation (\[ps4\]) suggest. In this paper, we set $L_l < L_1 < L_u = 10^{45.5}$ erg s$^{-1}$, where $L_l$ is set by the limiting flux of the EMSS: $F(L_l,z) = 1.33$ x 10$^{-13}$ erg cm$^{-2}$ s$^{-1}$ (Henry 1992); hence, $L_l$ is a function of redshift.[^4] Also in this paper, we set $0.14 = z_l < z < z_u = 0.6$. We exclude higher redshifts because (1) the $L$-$T$ relation of Reichart, Castander, and Nichol (1998) is derived from $z \la 0.5$ X-ray clusters, so its accuracy should not be trusted at redshifts much in excess of this value, and (2) optical identification become more difficult at these high redshifts (§1). This excludes three $\sim$ $L_\star$ X-ray clusters, which reduces the total number of X-ray clusters in our sample from 64 to 61. Below, however, we repeat our analysis with $z_u = 0.9$ to show that the exclusion of these three X-ray clusters does not unfairly bias our results. In Figure 4, we mark this region of the $L_1-z$ plane with a dotted line; solid points are interior to this region. The posterior probability distribution for any two parameters, e.g., $P(\Omega_m,n)$, or one parameter, e.g., $P(\Omega_m)$, is given by marginalizing the posterior probability distribution for all three parameters, $P(\Omega_m,n,c)$, over the other parameters (e.g., Gregory & Loredo 1992). 1, 2, and 3 $\sigma$ credible regions are determined by integrating the posterior probability distribution over the most probable region of its parameter space until 68.3%, 95.4% and 99.73% of this distribution has been integrated (e.g., Gregory & Loredo 1992). In Figure 5, we plot the 1, 2, and 3 $\sigma$ credible regions of the two-dimensional posterior probability distributions $P(\Omega_m,n)$ (top panel) and $P(n,c)$ (bottom panel). The 1 $\sigma$ credible intervals of the one-dimensional posterior probability distributions, i.e., $P(\Omega_m)$, $P(n)$, and $P(c)$, are $\Omega_m = 0.96^{+0.48}_{-0.38}$, $n = -2.28^{+0.36}_{-0.25}$, and $c = 0.66^{+0.48}_{-0.23}$. Hence, the EMSS favors high values of $\Omega_m$ and low values of $n$. Although these results by themselves are constraining $-$ $\Omega_m < 0.2$ is ruled out at the 2.3 $\sigma$ credible level $-$ by combining the likelihood function of the EMSS with the posterior probability distribution of the BCS (Figure 3), stronger constraints can be placed upon these parameters. We now determine the combined posterior probability distribution of the EMSS and the BCS. Let the likelihood function be that of the EMSS, as given by equation (\[likely\]). However, instead of assuming a flat prior probability distribution for all three parameters, as we did above, only assume a flat prior probability distribution between $0 < \Omega_m < 1.5$, and use the posterior probability distribution of the BCS $-$ $P_{BCS}(n,g_{eff})$ (Figure 3) $-$ equation (\[g\]), and the effective redshift of the BCS, $z_{eff} \sim 0.1$, to determine the full prior probability distribution: $P_{BCS}(\Omega_m,n,c)$. In Figure 6, we plot credible regions of the two-dimensional combined EMSS/BCS posterior probability distributions $P(\Omega_m,n)$ and $P(n,c)$. In the three left panels of Figure 8, we plot the one-dimensional combined EMSS/BCS posterior probability distributions. The dotted lines in this figure mark the 1, 2, and 3 $\sigma$ credible intervals. We find that $\Omega_m = 0.96^{+0.36}_{-0.32}$, $n = -1.86^{+0.42}_{-0.34}$, and $c = 0.54^{+0.24}_{-0.12}$. Values of $\Omega_m < 0.2$ are ruled out at the 3.0 $\sigma$ credible level. To establish that the exclusion of the three $z > 0.6$, $\sim$ $L_\star$ EMSS clusters does not unfairly bias our results, we repeat this analysis for $z_u = 0.9$. In Figure 7, we plot credible regions of the two-dimensional combined EMSS/BCS posterior probability distributions $P(\Omega_m,n)$ and $P(n,c)$. In the three right panels of Figure 8, we plot the one-dimensional combined EMSS/BCS posterior probability distributions. In this case, we find that $\Omega_m = 0.93^{+0.33}_{-0.26}$, $n = -1.50^{+0.37}_{-0.36}$, and $c = 0.48^{+0.12}_{-0.12}$. Values of $\Omega_m < 0.2$ are ruled out at the 3.5 $\sigma$ credible level. Lower values of $\Omega_m$ are not favored for three reasons. First of all, it is clear from Figure 3 that we have added a great deal of volume for which there are no EMSS clusters above $L_\star$. Secondly, for the region of the $L_1-z$ plane occupied by these three clusters, the EMSS surveyed roughly twice as much area if $\Omega_m = 1$ than it did if $\Omega_m = 0$ (§3). This is primarily because higher values of $\Omega_m$ imply lower luminosity distances, which imply higher fluxes for a given luminosity and, consequently, greater surveyed areas by equation (\[area\]). The third reason, also mentioned in §3, is that at luminosities $\la L_\star$, the luminosity function itself is a non-negligible, increasing function of $\Omega_m$. At these luminosities, the redshift dependence of the luminosity function is dominated by the function $f(z)$. This function is approximately given by (equations (\[f\]), (\[del\]), and (\[x\])) $$f(z) \propto \cases{(1+z)^{0.33(3-n)} & ($\Omega_m = 0$) \cr (1+z)^{1+0.23(3-n)} & ($\Omega_m = 1$)}$$ Consequently, there is an additional factor of $\approx (1+z)^{1/2}$ in the $\Omega_m = 1$ case. Although our $z_u = 0.9$ results are more constraining than are our $z_u = 0.6$ results, for the reasons stated above, we feel less confident about these results than we do about our $z_u = 0.6$ results. We now show that the uncertainty in the value of $s$ does not significantly affect our results. In the three left panels of Figure 9, we plot the one-dimensional combined EMSS/BCS posterior probability distributions, once again for $z_u = 0.6$, except that we have now used the $-1$ $\sigma$ value of $s$ from Reichart, Castander, & Nichol (1998). In the three right panels of Figure 9, we plot the same distributions, but for the $+1$ $\sigma$ value of $s$ from Reichart, Castander, & Nichol (1998). The effect of varying the value of $s$ by $\pm 1$ $\sigma$ is a variation in the values of the fitted parameters by less than the extent of their $\pm 1$ $\sigma$ uncertainties. If we add these uncertainties in quadrature, we find that $\Omega_m \approx 1.0 \pm 0.4$, $n \approx -1.9 \pm 0.4$, and $c \approx 0.55^{+0.25}_{-0.15}$. One-dimensional credible intervals for all of the fitted values in this section are compiled in Table 1. Discussion & Conclusions ======================== In this paper, we have constructed from the Press-Schechter mass function and the empirical X-ray cluster $L$-$T$ relation of Reichart, Castander, & Nichol (1998) an X-ray cluster luminosity function that can be applied to the growing number of independent, high-redshift, X-ray cluster luminosity catalogs to constrain cosmological parameters. In particular, we have incorporated the evolution of the $L$-$T$ relation and all significant dependences upon $\Omega_m$ of the luminosity and selection functions into our Bayesian inference analysis. For a fixed value of $H_0$, we have applied this luminosity function to broad subsets of the revised EMSS X-ray cluster subsample of Nichol (1997) and to the [*ROSAT*]{} BCS luminosity function of Ebeling (1997) to constrain $\Omega_m$. For the 61 revised EMSS clusters between $0.14 < z < 0.6$, we find that $\Omega_m = 0.96^{+0.36}_{-0.32}$ and $n = -1.86^{+0.42}_{-0.34}$; for all 64 revised EMSS clusters, between $0.14 < z < 0.9$, we find that $\Omega_m = 0.93^{+0.33}_{-0.26}$ and $n = -1.50^{+0.37}_{-0.36}$. These high values of $\Omega_m$ are the result of an apparent deficit of high-redshift, luminous X-ray clusters, which suggests that the X-ray cluster luminosity function has evolved above $L_\star$. Nichol (1997) suggested that the statistical evidence for the evolution of the EMSS luminosity function was only minimal. At first glance, this appears to be in contradiction to one of the conclusions of this paper. However, since Nichol (1997) used a power-law luminosity function - which they did for purposes of comparison with the original EMSS result of Henry (1992) - instead of a luminosity function that permits different degrees of evolution below and above $L_\star$, as we have done in this paper, their results are most directly applicable below $L_\star$: this is the luminosity range of the vast majority of the EMSS clusters, so it is by this luminosity range that their results have most strongly been weighted. The fact that the X-ray cluster luminosity function does not evolve below $L_\star$ has since been shown by Collins (1997), Burke (1997), Rosati (1998), Jones (1998), and Vikhlinin (1998a). That the luminosity function appears to evolve above $L_\star$ is in agreement with the original EMSS findings of Gioia (1990a), as well as the findings of Vikhlinin (1998a,b) with the 160 deg$^2$ survey. The value of $\Omega_m$ that we find that this high-luminosity evolution in the EMSS corresponds to is consistent with the values found by Sadat, Blanchard, & Oukbir (1998) ($\Omega_m = 0.85 \pm 0.2$) and Blanchard & Bartlett (1998) ($\Omega_m \approx 1$), based upon the work of Oukbir & Blanchard (1992,1997). Our value of $\Omega_m$ is somewhat consistent with the values found by Henry (1997) ($\Omega_m = 0.50 \pm 0.14$) and Eke (1998) ($\Omega_m = 0.45 \pm 0.2$); however, Viana & Liddle (1998) have performed a more extensive error analysis upon a conservative subset of the data of these authors and find that $\Omega_m \sim 0.75$ with $\Omega_m > 0.3$ at the 90% confidence level and $\Omega_m \sim 1$ still viable. Blanchard, Bartlett, & Sadat (1998) find almost identical results ($\Omega_m \sim 0.74$, with $0.3 < \Omega_m < 1.2$ at the 95% confidence level) from these data. Finally, our value of $\Omega_m$ is inconsistent with the values found by Bahcall, Fan, & Cen (1997) ($\Omega_m = 0.3 \pm 0.1$), Fan, Bahcall, & Cen (1997) ($\Omega_m \approx 0.3 \pm 0.1$), and Bahcall & Fan (1998) ($\Omega_m = 0.2^{+0.3}_{-0.1}$). Our value of $n$ is consistent with the values found by Henry & Arnaud (1991) ($n = -1.7^{+0.65}_{-0.35}$) and Henry (1992) ($n = -2.10^{+0.27}_{-0.15}$), where these authors set $\Omega_m = 1$. Our value of $n$ is also consistent with the value found by Eke (1998) ($n = -1.69^{+0.12}_{-0.07}$), where these authors included $\Omega_m$ as a free parameter. Our value of $n$ is somewhat consistent with the value that Bahcall, Fan, & Cen (1997), Fan, Bahcall, & Cen (1997), and Bahcall & Fan (1998) adopted ($n = -1.4$). Taken as an ensemble, these results are perhaps discouraging in that they span the entire range of acceptable solutions: $0.2 \la \Omega_m \la 1$. This suggests that as yet unknown systematic effects may be plaguing some, if not all, of these results. We briefly identify seven areas where systematic effects could enter ours and similar analyses. (1) The first is the Press-Schechter mass function itself; however, numerical simulations (e.g., Eke 1996; Bryan & Norman 1997; Borgani 1998) consistently show that the Press-Schechter mass function is an adequate approximation. (2) The spherical collapse model of cluster formation (equations (\[del\]) and (\[del2\])) may be inadequate. For example, numerical simulations by Governato (1998) suggest that in equation (\[del2\]), the expression $1.69(1+z)$ may really be as low as $\sim 1.6(1+z)^{0.9}$. This suggests that use of the spherical collapse model may lead to underestimated values of $\Omega_m$; however, this is only a $\la 10$% effect. (3) Technically, equation (\[virial\]) only holds when $\Omega_m = 1$. Recently, Voit & Donahue (1998) derived a virial theorem that holds for all values of $\Omega_m$, and that allows for the fact that clusters grow gradually; their mass-temperature ($M$-$T$) relation reduces to equation (\[virial\]) when $\Omega_m = 1$. We find that $M$-$T$ relations with functional forms that are similar to that of the $M$-$T$ relation of Voit & Donahue (1998) reduce our fitted value of $\Omega_m$ by $\la 10$%; however, further investigation and use of this $M$-$T$ relation is clearly needed. (4) Also on the subject of the $M$-$T$ relation, care must be taken when fitting to X-ray cluster temperature catalogs: cooling flows lower the measured temperature of most X-ray clusters, which should systematically affect values of $\Omega_m$ that are determined in this way. (5) Based upon the cooling flow corrected X-ray cluster temperature catalogs of Markevitch (1998) and Allen & Fabian (1998), Reichart, Castander, & Nichol (1998) determined an empirical $L$-$T$ relation between measured luminosities and cooling flow corrected temperatures that holds for $z \la 0.5$ and for luminosities that are typical of X-ray cluster catalogs (see §3.1); however, more cooling flow corrected X-ray cluster temperature measurements are needed to determine what, if any, exceptions exist to this $L$-$T$ relation, and to extend it to higher redshifts. (6) The art of determining an X-ray cluster catalog’s selection function is a constantly improving science; modern selection functions are determined via extensive numerical simulations. An alternative explanation to our high-$\Omega_m$ result is that the EMSS, for whatever reasons, missed many high-redshift, high-luminosity X-ray clusters beyond what is accounted for by their selection function (see §3.3). However, given that the EMSS detected many high-redshift, low-luminosity X-ray clusters, this seems to be an unlikely scenario. (7) Finally, our cosmological model may be inadequate. We did not investigate the effects of a cosmological constant in this paper; however, many authors have demonstrated that the inclusion of a cosmological constant has little effect upon the determined value of $\Omega_m$ (see §1). Also, the effects of quintessent and other exotic cosmologies have not yet been investigated in this context. Some of these potential sources of systematic error can be safeguarded against. For example, the spherical collapse model, the $M$-$T$ relation, and the $L$-$T$ relation all have proportionality factors that are potential sources of systematic error. However, as we have shown in §2.1, all of these factors, as well as the parameter $\sigma_8$, group together, giving us our parameter $c$. Since we fit for $c$, these factors cannot bias our result. However, inadequate functional forms for these relations, as well as for the other functions listed above, can bias ours and others’ results. In addition to further theoretical and numerical development of this formalism, only the continued construction of X-ray cluster catalogs will act to further resolve these issues. Fortunately, the number of X-ray cluster luminosity catalogs is growing rapidly. One such catalog is the Southern Serendipitous High-Redshift Archival [*ROSAT*]{} Catalog (Southern SHARC) (Collins 1997; Burke 1997). The redshift and luminosity ranges of the Southern SHARC are $z < 0.7$ and $L_1 < 3 \times 10^{44}$ erg s$^{-1}$ (0.5 - 2.0 keV). Although the Southern SHARC does not span the luminosity range of the EMSS, it will provide a good consistency check of our EMSS results. Our analysis of this catalog is underway. Two similar X-ray cluster catalogs that can serve a similar purpose are the [*ROSAT*]{} Deep Cluster Survey (RDCS) (Rosati 1998) and the Wide Angle [*ROSAT*]{} Pointed Survey (WARPS) (Jones 1998). The RDCS spans the redshift and luminosity ranges $z < 0.8$ and $L_1 < 3 \times 10^{44}$ erg s$^{-1}$ (0.5 - 2.0 keV). The WARPS spans the redshift and luminosity ranges $z < 0.7$ and $L_1 < 2 \times 10^{44}$ erg s$^{-1}$ (0.5 - 2.0 keV). An analysis of the RDCS is also underway (Borgani 1998). The 160 deg$^2$ survey (Vikhlinin 1998b) and the Bright SHARC (Romer 1998), a high-luminosity extension of the Southern SHARC that is currently under construction, span redshift and luminosity ranges that rival those of the EMSS. Consequently, these catalogs will provide strong, independent checks of the EMSS results. Finally, local ($z_{eff} \sim 0.1$) X-ray cluster catalogs, such as the [*ROSAT*]{} BCS, are of great importance. Although these samples do not have the redshift leverage to constrain cosmological parameters, their large sizes make them excellent samples to better constrain the parameters $n$ and $c$. Samples like the EMSS, the 160 deg$^2$ survey, and the Bright SHARC do not have sufficient luminosity leverage to strongly constrain these parameters, which leads to weaker constraints upon the cosmological parameters. However, a simultaneous analysis of a local X-ray cluster catalog $-$ as opposed to a local X-ray cluster luminosity function as we have used in this paper $-$ and any of these high-redshift, high-luminosity X-ray cluster catalogs could lead to significantly improved constraints upon all of these parameters. This research has been partially funded by NASA grants NAG5-6548 and NAG5-2432. We are very grateful to H. Ebeling for providing us with the data for Figure 2. Also, we are grateful to A. Blanchard, S. Borgani, C. Graziani, D. Q. Lamb, C. A. Metzler, C. Scharf, M. S. Turner, and J. M. Quashnock for valuable discussions. We are also very grateful to our anonymous referee, whose input has greatly improved this paper. D. E. R. is especially grateful to Dr. and Mrs. Bernard Keisler for their hospitality during the summers of 1997 and 1998. [ccccccc]{} BCS & $-$ & $-$ & $-$ & $-1.83^{+0.85}_{-0.15}$ & $-$ & $-$ EMSS & $0.6$ & $3.14-0.65\Omega_m$ & $0.96^{+0.48}_{-0.38}$ & $-2.28^{+0.36}_{-0.25}$ & $0.66^{+0.48}_{-0.23}$ & 2.3 $\sigma$ EMSS+BCS & $0.6$ & $3.14-0.65\Omega_m$ & $0.96^{+0.36}_{-0.32}$ & $-1.86^{+0.42}_{-0.34}$ & $0.54^{+0.24}_{-0.12}$ & 3.0 $\sigma$ EMSS+BCS & $0.9$ & $3.14-0.65\Omega_m$ & $0.93^{+0.33}_{-0.26}$ & $-1.50^{+0.37}_{-0.36}$ & $0.48^{+0.12}_{-0.12}$ & 3.5 $\sigma$ EMSS+BCS & $0.6$ & $2.26-0.61\Omega_m$ & $0.69^{+0.36}_{-0.27}$ & $-1.74^{+0.42}_{-0.36}$ & $0.48^{+0.28}_{-0.12}$ & 2.6 $\sigma$ EMSS+BCS & $0.6$ & $3.79-0.59\Omega_m$ & $1.17^{+0.33}_{-0.28}$ & $-1.98^{+0.34}_{-0.30}$ & $0.66^{+0.25}_{-0.18}$ & 3.2 $\sigma$ Allen, S. W., & Fabian, A. C. 1998, MNRAS, 297, L57 Arnaud, M., & Evrard, A. E. 1998, MNRAS, submitted (astro-ph 9806353) Avni, Y., & Bahcall, J. N. 1980, ApJ, 235, 694 Bahcall, N. A., & Fan, X. 1998, ApJ, 504, 1 Bahcall, N. A., Fan, X., & Cen, R. 1997, ApJ, 485, L53 Blanchard, A., & Bartlett, J. G. 1997, A&A, 332, L49 Blanchard, A., Bartlett, J. G., & Sadat, R. 1998, in Les Comptes Rendus de l’Academie des Sciences, in press Borgani, S., 1998, ApJ, submitted Bryan, G. L., & Norman, M. L. 1997, ApJ, 495, 80 Burke, D. J., Collins, C. A., Sharples, R. M., Romer, A. K., Holden, B. P., & Nichol, R. C. 1997, ApJ, 488, L83 Cash, A. 1979, ApJ, 228, 939 Collins, C. A., Burke, D. J., Romer, A. K., Sharples, R. M., & Nichol, R. C. 1997, ApJ, 479, L117 David, L. P., Slyz, A., Jones, C., Forman, W., Vrtilek, S. D., & Arnaud, K. A. 1993, ApJ, 412, 479 Ebeling, H., Edge, A. C., Fabian, A. C., Allen, S. W., Crawford, C. S., & Böehringer, H. 1997, ApJ, 479, L101 Edge, A. C., & Stewart, G. C. 1991, MNRAS, 252, 414 Eke, V. R., Cole, S., Frenk, C. S., & Henry, J. P. 1998, MNRAS, 298, 1145 Eke, V. R., Cole, S., Frenk, C. S., & Navarro, J. F. 1996, MNRAS, 281, 703 Evrard, A. E., & Henry, J. P. 1991, ApJ, 383, 95 Fan, X., Bahcall, N. A., & Cen, R. 1997, ApJ, 490, L123 Gioia, I. M., & Luppino, G. A. 1995, ApJS, 94, 583 Gioia, I. M., Henry, J. P., Maccacaro, T., Morris, S. L., Stocke, J. T., & Wolter, A. 1990a ApJ, 356, L35 Gioia, I. M., Maccacaro, T., Schild, R. E., Wolter, A., Stocke, J. T., Morris, S. L., & Henry, J. P. 1990b, ApJS, 72, 567 Governato, F., Babul, A., Quinn, T., Tozzi, P., Baugh, C. M., Katz, N., & Lake, G. 1998, MNRAS, submitted (astro-ph 9810487) Gregory, P. C., & Loredo, T. J.. 1992, ApJ, 398, 146 Henry, J. P. 1997, ApJ, 489, L1 Henry, J. P., Gioia, I. M., Maccacaro, T., Morris, S. L., Stocke, J. T., & Wolter, A. 1992, ApJ, 386, 408 Henry, J. P., & Arnaud, K. A. 1991, ApJ, 372, 410 Jones, L. R., Scharf, C., Ebeling, H., Perlman, E., Wegner, G., Malkan, M., & Horner, D. 1998, ApJ, 495, 100 Kaiser, N. 1986, MNRAS, 222, 323 Kaiser, N. 1991, ApJ, 383, 104 Lacey, C., & Cole, S. 1993, MNRAS, 262, 627 Markevitch, M. 1998, ApJ, 504, 27 Mathiesen, B., & Evrard, A. E. 1998, MNRAS, 295, 769 Mushotzky, R. F., & Scharf, C. A. 1997, ApJ, 482, L13 Nichol, R. C., Holden, B. P., Romer, A. K., Ulmer, M. P., Burke, D. J., & Collins, C. A. 1997, ApJ, 481, 644 Oukbir, J., & Blanchard, A. 1997, A&A, 262, L21 Oukbir, J., & Blanchard, A. 1997, A&A, 317, 10 Peebles, P. J. E. 1980, The Large Scale Structure of the Universe (Princeton: Princeton University Press) Press, W. H., & Schechter, P. 1974, ApJ, 187, 425 Reichart, D. E., Castander, F. J., & Nichol, R. C. 1998, ApJ, in press (astro-ph 9810487) Romer, A. K., Nichol, R. C., Collins, C. A., Burke, D. J., Holden, B. P., Ulmer, M. P., Pildis, R. A., & Metevier, A. 1998, AN, 319, 83 Rosati, P., Della Ceca, R., Norman, C., & Giacconi, R. 1998, ApJ, 492, L21 Sadat, R., Blanchard, A., & Oukbir, J. 1998, A&A, 329, 21 Stocke, J. T., Morris, S. L., Gioia, I. M., Maccacaro, T., Schild, R., Wolter, A., Fleming, T. A., & Henry, J. P. 1991, ApJS, 76, 813 Viana, P. T. P., & Liddle, A. R. 1998, MNRAS, in press (astro-ph 9803244) Vikhlinin, A., McNamara, B. R., Forman, W., Jones, C., Quintana, H., & Hornstrup, A. 1998, ApJ, 498, L21 Vikhlinin, A., McNamara, B. R., Forman, W., & Jones, C. 1998, ApJ, 502, 558 Voit, G. M., & Donahue, M. 1998, ApJ, 500, L111 [^1]: The quantity $f_F(d_A(\Omega_m))$ is also a function of X-ray cluster core radius, $a_0$, which Henry (1992) found to be $a_0 \sim 0.25$ Mpc, assuming that $\Omega_m = 1$. Repeating their analysis for $\Omega_m = 0$, we find that this result again holds; consequently, we adopt this value of $a_0$ throughout this paper. [^2]: In this paper, $L_\star$ refers very generally to those luminosities at which the luminosity function, modeled by equation (\[ps3\]), appears to roll over from a power law to an exponential cutoff. [^3]: See, e.g., Gregory & Loredo (1992) for an excellent discussion of Bayesian inference. [^4]: In computing $L_l(z)$, we set $\Omega_m = 0$. This is because if we were to use a higher value of $\Omega_m$ when defining this curve, the selection function, $A(L_1,z)$, would be undefined for luminosities and redshifts near this curve when the value of $\Omega_m$ is lower than the curve-defining value (see equation (\[flux\])).

ArXiv

--- author: - Jürgen Pannek and Matthias Gerdts title: Performance of Sensitivity based NMPC Updates in Automotive Applications --- Introduction {#p41:sec_introduction} ============ Within the last decades, model predictive control (MPC) has grown mature for both linear and nonlinear systems, see, e.g., [@CamachoBordons2004; @RawlingsMayne2009; @GruenePannek2011]. Although analytically and numerically challenging, the method itself is attractive due to its simplicity and approximates an infinite horizon optimal control as follows: In a first step, a measurement of the current system state is obtained which in the second step is used to compute an optimal control over a finite optimization horizon. In the third and last step, a portion of this control is applied to the process and the entire problem is shifted forward in time rendering the scheme to be iteratively applicable. Unfortunately, stability and optimality of the closed loop may be lost due to considering finite horizons only. To ensure stability of the resulting closed loop, one may impose terminal point constraints as shown in [@KeerthiGilbert1988; @Alamir2006] or Lyapunov type terminal costs and terminal regions, see [@ChenAllgoewer1998; @MayneRawlingsRaoScokaert2000]. A third approach uses a relaxed Lyapunov condition presented in [@GrueneRantzer2008] which can be shown to hold if the system is controllable in terms of the stage costs [@Gruene2009; @GruenePannekSeehaferWorthmann2010]. Additionally, this method allows for computing an estimate on the degree of suboptimality with respect to the infinite horizon controller, see also [@ShammaXiong1997; @NevisticPrimbs1997] for earlier works on this topic. Here, we use an extension of the third approach to the case of parametric control systems and subsequent disturbance rejection updates. In particular, we focus on updating the MPC control law via sensitivities introduced in [@Fiacco1983]. Such updates have been analysed extensively for the case of open loop optimal controls, see, e.g, [@GroetschelKrumkeRambau2001], but were also applied in the MPC closed loop context in [@ZavalaBiegler2009; @PannekGerdts2012]. In order to avoid the usage of stabilizing Lyapunov type terminal costs and terminal regions and obtain performance results with respect to the infinite horizon controller, we utilize results from [@PannekGerdts2012] in an advanced step setting, see, e.g., [@FindeisenAllgoewer2004]. In the following, we present the considered half car model from [@SpeckertDresslerRuf2009; @PoppSchiehlen2010] and the imposed MPC setup. The obtained numerical results show that this approach is both realtime applicable and provides a cheap and yet significant performance improvement with respect to the comfort and handling objectives requested by our industrial partners. Problem setting {#p41:sec_setting} =============== Throughout this work we consider the control systems dynamics of a half car which originate from [@SpeckertDresslerRuf2009; @PoppSchiehlen2010] and are slightly modified to incorporate active dampers, see Fig. \[fig:halfcar\] for a schematical sketch. ![Schematical sketch of a halfcar subject to road excitation ${w}$[]{data-label="fig:halfcar"}](p41_halfcar-eps-converted-to){width="62.00000%"} The resulting second order dynamics read $$\begin{aligned} m_1 \ddot{{x}}_1 & = m_1 g + f_3 - f_1 \qquad & m_3 \ddot{{x}}_3 & = m_3 g - f_3 - f_4 \nonumber \\ \label{dynamic} m_2 \ddot{{x}}_2 & = m_2 g + f_4 - f_2 \qquad & I \ddot{{x}}_4 & = \cos({x}_4) ( b f_3 - a f_4 )\end{aligned}$$ where the control enters the forces $$\begin{aligned} f_1 & = k_1 ( {x}_1 - {w}_1 ) + d_1 (\dot{{x}}_1 - \dot{{w}}_1) \\ f_2 & = k_2 ( {x}_2 - {w}_2 ) + d_2 (\dot{{x}}_2 - \dot{{w}}_2) \\ f_3 & = k_3 ( {x}_3 - {x}_1 - b \sin({x}_4) ) + {u}_1 ( \dot{{x}}_3 - \dot{{x}}_1 - b \dot{{x}}_4 \cos({x}_4) ) \\ f_4 & = k_4 ( {x}_3 - {x}_2 + a \sin({x}_4) ) + {u}_2 ( \dot{{x}}_3 - \dot{{x}}_2 + a \dot{{x}}_4 \cos({x}_4) )\end{aligned}$$ Here, ${x}_1$ and ${x}_2$ denote the centers of gravity of the wheels, ${x}_3$ the respective center of the chassis and ${x}_4$ the pitch angle of the car. The disturbances ${w}_1$, ${w}_2$ are connected via ${w}_1(t) = {w}(t)$, ${w}_2(t) = {w}(t - \Delta)$ and the control constraints ${\mathbb{{U}}}= [ 0.2 kNs/m, 5 kNs/m ]^2$ limit the range of the active dampers. The remaining constants of the halfcar are displayed in Table \[tab:halfcar\]. [cccc]{} name & symbol & quantity & unit\ distance to joint & $a, b$ & $1$ & $m$\ mass wheel & $m_1, m_2$ & $15$ & $kg$\ mass chassis & $m_3$ & $750$ & $kg$\ inertia & $I$ & $500$ & $kg\, m^2$\ spring constant wheels & $k_1, k_2$ & $2\cdot 10^5$ & $kN/m$\ damper constant wheels & $d_1, d_2$ & $2\cdot 10^2$ & $kNs/m$\ spring constant chassis & $k_3, k_4$ & $1 \cdot 10^5$ & $kN/m$\ gravitational constant & $g$ & $9.81$ & $m/s^2$\ MPC Algorithm {#p41:sec_mpc} ============= In order to design a feedback for the half car problem , we impose the cost functional $$\begin{aligned} \label{costfunctional} {{J}_N}({x}, {u}, {w}) := \sum_{{k}= 0}^{N-1} \mu_R F_{R}( {k}) + \mu_A F_{A}( {k})\end{aligned}$$ following ISO 2631 with horizon length $N = 5$. The handling objective is implemented via $$\begin{aligned} & F_{R}( {k}) := \sum_{i = 1}^2 \int\limits_{{k}T}^{({k}+1) T} \left( \frac{[ k_i ( {x}_i(t) - {w}_i(t) ) + d_i ( \dot{{x}}_i(t) - \dot{{w}}_i(t) ) ] - F_i}{F_i} \right)^2 \, dt\end{aligned}$$ with nominal forces $$\begin{aligned} F_1 & = ( a \cdot g \cdot ( m_1 + m_2 + m_3 ) ) / ( a + b ) \\ F_2 & = ( b \cdot g \cdot ( m_1 + m_2 + m_3 ) ) / ( a + b )\end{aligned}$$ whereas minimizing the chassis jerk $$\begin{aligned} F_{A}({k}) := \int\limits_{{k}T}^{({k}+1)T} \left( m_3 \dddot{{x}}_3(t) \right)^2 \, dt\end{aligned}$$ is used to treat the comfort objective. Both integrals are equally weighted via $\mu_R = \mu_A = 1$ and are evaluated using a constant sampling rate of $T = 0.1s$ during which the control are held constant, i.e. the control is implemented in a zero–order hold manner. The nominal disturbance ${w}(\cdot)$ and the corresponding derivates are computed from road profile measurements taken at a sampling rate of $0.002s$ via a fast Fourier transformation (FFT). For the resulting finite time optimal control problem, we denote a minimizer of satisfying all constraints by ${{u}^\star}(\cdot, {x}, {w})$. Since the control must be readily computed at the time instant it is supposed to be applied, ${{u}^\star}(\cdot, {x}, {w})$ is computed in an advanced step setting, cf. [@FindeisenAllgoewer2004]. To this end, the initial state ${x}$ of the optimal control problem is predicted for a future time instant using the last known measurement and the intermediate control which is readily available from previous MPC iteration steps. Since we want to apply sensitivity updates in case of measurement/prediction deviations and disturbances, we additionally precompute sensitivity information along the optimal open loop solution with respect to the predicted state $\partial {{u}^\star}/\partial {x}(\cdot, {x}, {w})$ and the nominal disturbances $\partial {{u}^\star}/\partial {w}(\cdot, {x}, {w})$. Then, once the nominal control ${{u}^\star}(\cdot, {x}, {w})$ is to be applied, we use newly obtained state and disturbance information ${\overline{{x}}}$, ${\overline{{w}}}$ to update the control via $$\begin{aligned} \label{update formula} {\overline{{u}}}(\cdot, {\overline{{x}}}, {\overline{{w}}}) := {{u}^\star}(\cdot, {x}, {w}) + \begin{pmatrix} \frac{\partial {{u}^\star}}{\partial {x}}(\cdot, {x}, {w}) \\ \frac{\partial {{u}^\star}}{\partial {w}}(\cdot, {x}, {w}) \end{pmatrix}^\top \begin{pmatrix} {\overline{{x}}}(\cdot) - {x}(\cdot) \\ {\overline{{w}}}(\cdot) - {w}(\cdot) \end{pmatrix},\end{aligned}$$ see also [@GroetschelKrumkeRambau2001; @Fiacco1983] for details on the computation and limitations of sensitivities. For simplicity of exposition, we predict the initial state ${x}$ using two sampling intervals $T$ of the closed loop control. Note that although larger predictions are possible, robustness problems are more likely to occur since predicted and real solutions usually diverge, see, e.g., [@LimonAlamoRaimondoBravoMunoyFerramoscaCamacho2009; @FindeisenGruenePannekVarutti2011]. Numerical Results {#p41:sec_numericalresults} ================= During our simulations, we modified both the states of the system and the road profile measurements using a disturbance which is uniformly distributed in the interval $[-0.025 m, 0.025 m]$. For this setting, precomputation of ${{u}^\star}(\cdot, {x}, {w})$, $\partial {{u}^\star}/\partial {x}(\cdot, {x}, {w})$ and $\partial {{u}^\star}/\partial {w}(\cdot, {x}, {w})$ required at maximum $0.168s < 2 T = 0.2s$ which renders the scheme realtime applicable. As expected, the updated control law shows a better performance than the nominal control. The improvement cannot only be observed from Fig. \[fig:results\], but also in terms of the closed loop costs: For the considered race track road data we obtain an improvement of approximately $8.2 \%$ using the sensitivity update . Although this seems to be a fairly small improvement, the best possible result obtained by a full reoptimization reveals a reduction of approximately $10.5 \%$ of the closed loop costs. ![Comparison plot for the chassis jerk using MPC with ($\times$) and without sensitivity update ($\circ$).[]{data-label="fig:results"}](p41_jerk-eps-converted-to){width="62.00000%"} Note that due to the presence of constraints it is a priori unknown whether the conditions of the Sensitivity Theorem of [@Fiacco1983] hold at each visited point along the closed loop. Such an occurrance can be detected online by checking for violations of constraints or changes in the control structure. Yet, due to the structure of the MPC algorithm, such an event has to be treated if one of the constraints is violated at open loop time instant ${k}= 1$ only which was not the case for our example. This work was partially funded by the German Federal Ministry of Education and Research (BMBF), grant no. 03MS633G. [10]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{} Alamir, M.: Stabilization of nonlinear systems using receding-horizon control schemes. Springer (2006) Camacho, E., Bordons, C.: [Model Predictive Control]{}. Springer (2004) Chen, H., Allg[ö]{}wer, F.: A quasi-infinite horizon nonlinear model predictive control scheme with guaranteed stability. Automatica **34**(10), 1205–1218 (1998) Fiacco, A.: Introduction to sensitivity and stability analysis in nonlinear programming. Academic Press Inc. (1983) Findeisen, R., Allgöwer, F.: [Computational Delay in Nonlinear Model Predictive Control]{}. In: Proceedings of the International Symposium on Advanced Control of Chemical Processes (2004) Findeisen, R., Grüne, L., Pannek, J., Varutti, P.: [Robustness of Prediction Based Delay Compensation for Nonlinear Systems]{}. In: Proceedings of the 18th IFAC World Congress, pp. 203–208. Milan, Italy (2011) Grötschel, M., Krumke, S., Rambau, J.: Online Optimization of Large Scale Systems. Springer (2001) Gr[ü]{}ne, L.: Analysis and design of unconstrained nonlinear [MPC]{} schemes for finite and infinite dimensional systems. SIAM Journal on Control and Optimization **48**, 1206–1228 (2009) Grüne, L., Pannek, J.: Nonlinear Model Predictive Control: Theory and Algorithms. Springer (2011) Grüne, L., Pannek, J., Seehafer, M., Worthmann, K.: Analysis of unconstrained nonlinear [M]{}[P]{}[C]{} schemes with varying control horizon. SIAM Journal on Control and Optimization **48**(8), 4938–4962 (2010) Grüne, L., Rantzer, A.: On the infinite horizon performance of receding horizon controllers. IEEE Transactions on Automatic Control **53**(9), 2100–2111 (2008) Keerthi, S., Gilbert, E.: Optimal infinite horizon feedback laws for a general class of constrained discrete-time systems: stability and moving horizon approximations. Journal of Optimization Theory and Applications **57**, 265–293 (1988) Limon, D., Alamo, T., Raimondo, D.L., Bravo, J.M., Munoy de la Pena, D., Ferramosca, A., Camacho, E.F.: Input-to-state stability: an unifying framework for robust model predictive control, nonlinear model predictive control. In: L. Magni, D. Raimondo, F. Allg[ö]{}wer (eds.) Nonlinear Model Predictive Control: Towards New Challenging Applications, [LNCIS]{} 384, pp. 1–26. Springer (2009) Mayne, D., Rawlings, J., Rao, C., Scokaert, P.: [Constrained model predictive control: Stability and optimality]{}. Automatica **36**(6), 789–814 (2000) Nevistić, V., Primbs, J.A.: Receding horizon quadratic optimal control: Performance bounds for a finite horizon strategy. In: Proceedings of the European Control Conference (1997) Pannek, J., Gerdts, M.: Robust stability and performance bounds for nmpc with abstract updates. In: Proceedings of the 4th IFAC Nonlinear Model Predictive Control Conference, pp. 311–316 (2012) Popp, K., Schiehlen, W.: Ground Vehicle Dynamics. Springer (2010) Rawlings, J.B., Mayne, D.Q.: Model Predictive Control: Theory and Design. Nob Hill Publishing (2009) Shamma, J., Xiong, D.: Linear nonquadratic optimal control. IEEE Transactions on Automatic Control **42**(6), 875–879 (1997) Speckert, M., Dreßler, K., Ruf, N.: Undesired drift of multibody models excited by measured accelerations or forces. Tech. rep., ITWM Kaiserslautern (2009) Zavala, V.M., Biegler, L.T.: [The advanced-step NMPC controller: Optimality, stability and robustness]{}. Automatica **45**(1), 86–93 (2009)

ArXiv

--- abstract: 'Convolutional neural network (CNN) and recurrent neural network (RNN) models have become the mainstream methods for relation classification. We propose a unified architecture, which exploits the advantages of CNN and RNN simultaneously, to identify medical relations in clinical records, with only word embedding features. Our model learns phrase-level features through a CNN layer, and these feature representations are directly fed into a bidirectional gated recurrent unit (GRU) layer to capture long-term feature dependencies. We evaluate our model on two clinical datasets, and experiments demonstrate that our model performs significantly better than previous single-model methods on both datasets.' address: - 'Research Center of Language Technology, Harbin Institute of Technology, Harbin, China' - 'Department of Mathematics, Harbin Institute of Technology, Harbin, China' author: - Bin He - Yi Guan - Rui Dai bibliography: - '2018he.bib' title: Convolutional Gated Recurrent Units for Medical Relation Classification --- Relation classification; Clinical record; Convolutional neural network; Gated recurrent unit. Introduction ============ Relation classification, a natural language processing (NLP) task which identifies the relation between two entities in a sentence, is an important technique in many subsequent NLP applications, such as question answering and knowledge base completion. In the clinical domain, Informatics for Integrating Biology and the Bedside (i2b2) released an annotated relation dataset on clinical records and attracted considerable attention [@Uzuner2011]. Identifying relations in clinical records is a challenging task because one sentence from clinical records may contain more than two medical concepts and a concept may contain several words. Figure \[fig:sample\] illustrates relation samples in this task. ![\[fig:sample\] An example of medical relations in a sample sentence. TrIP, treatment improves medical problem; TrAP, treatment is administered for medical problem.](sample1.pdf){width="\linewidth"} Due to the powerful feature learning ability, convolutional neural network (CNN) and recurrent neural network (RNN) are the mainstream architectures in the relation classification task [@Zhang2015a; @liu-EtAl:2015:ACL-IJCNLP; @xu-EtAl:2015:EMNLP1; @Xu2015; @xu-EtAl:2016:COLING1; @miwa-bansal:2016:P16-1; @Wang2016; @Zhou2016]. In order to utilize the advantages of these two neural networks simultaneously, combinations of CNN and RNN turn into a research trend. The most direct way is to use the voting scheme [@vu-EtAl:2016:N16-1]. The second combination way is to feed features extracted by a RNN architecture into CNN [@Cai2016; @Raj2017], which can be seen as generating new input representations by RNN. The third way is to stack RNN on CNN. Even though this architecture has not been applied to identify medical relations from clinical text, its variants have achieved remarkable results in many other classification tasks [@Tang2015; @Nguyen2015a; @zhou2015c; @choi2017convolutional]. Deep learning methods have presented satisfactory results [@Socher2012; @yu2014factor; @Zhang2015a; @liu-EtAl:2015:ACL-IJCNLP; @xu-EtAl:2015:EMNLP1; @Xu2015; @xu-EtAl:2016:COLING1; @miwa-bansal:2016:P16-1] and make the models less dependent on manual feature engineering. Moreover, some researchers proposed models only with word representations as input features [@DBLP:conf/acl/SantosXZ15; @Zhou2016], which achieved outstanding results. Similarly, our goal is to propose a model for relation classification on clinical records, without using any external feature set. In this work, we follow the third combination way and design a two-layer architecture: input representations (word-level) are fed into a CNN layer to learn n-gram features (phrase-level), and these feature representations are directly used as the input of a bidirectional gated recurrent unit (GRU) [@bahdanau2014neural] layer to achieve the final sample representation (sentence-level). Our main contributions are as follows: (1) we propose a unified architecture to identify medical relations in clinical records, which has the ability to capture both local features (extracted by a CNN layer) and sequential correlations among these features (extracted by a bidirectional GRU layer); (2) we also explore training our model with attention mechanism (C-BGRU-Att) and compare the performance with the model using the conventional max-pooling operation (C-BGRU-Max); (3) experiments show our model achieves better performance than previous single-model methods, with only word embedding features. Methodology =========== Figure \[fig:arch\] describes the architecture of our model for medical relation classification on clinical records. This model learns a distributed representation for each relation sample, and calculates final scores with relation type representations. More details will be discussed in the following sections. ![\[fig:arch\] Architecture of our model for medical relation classification. In the input of this architecture, concept contents in the relation sample “she was treated with $[\text{steroids}]_{treatment}$ for $[\text{this\ swelling}]_{problem}$ at the outside hospital , and these were continued ." are replaced by their concept types.](workflow.pdf){width="\linewidth"} Word representation layer {#sec:wordrepresentation} ------------------------- With reference to a previous study on relation classification [@Zeng2014], word position features capture information of the relative position between words and target concepts. Therefore, an word embedding matrix $W^{w}\in\mathbb{R}^{d^w\times |V^w|}$ and an word position embedding matrix $W^{wp}\in\mathbb{R}^{d^p\times |V^p|}$ are given in this work, where $V^w$ is the vocabulary, $V^p$ is the word position set, and $d^w$ and $d^p$ are pre-set embedding sizes. Every word in the relation sample is mapped to a column vector $\mathbf{x}^w_i$ to represent the word feature. In addition, relative distances between the current word and the target concepts are mapped to word position vectors $\mathbf{x}^{p_1}_i$ and $\mathbf{x}^{p_2}_i$. Based on the above features, each word can be represented by $\mathbf{x}_i'={[{(\mathbf{x}^w_i)}^\mathrm{T}, {(\mathbf{x}^{p_1}_i)}^\mathrm{T}, {(\mathbf{x}^{p_2}_i)}^\mathrm{T}]}^\mathrm{T}$, and $\mathbf{x}_i'\in\mathbb{R}^{d^x}$, where $d^x=d^w+2d^p$. Convolutional layer {#sec:conv} ------------------- The semantic representations of n-grams are valuable features to the relation classification task, and convolution operation can capture this information by combining word embedding features in a fixed window. Given the input representation $\mathbf{x}'=(\mathbf{x}_1',\mathbf{x}_2',\dots,\mathbf{x}_n')$ and a context window size $k$, concatenation of successive words in this window size can be defined as $X_j={[{\mathbf{x}_j'}^\mathrm{T},\dots,{\mathbf{x}_{j+k-1}'}^\mathrm{T}]}^\mathrm{T}$, and the representation of this relation sample can be reformatted as $X=(X_1,\dots,X_{n-k+1})$. Given a weight matrix of the convolutional filters $W^{conv}$ and a linear bias $\mathbf{b}$, the local feature representations are computed: $$C_j=tanh(W^{conv}\cdot X_j + \mathbf{b}),$$ where $W^{conv}\in\mathbb{R}^{d^c\times d^xk}$, $\mathbf{b}\in\mathbb{R}^{d^c}$, and $tanh$ denotes the hyperbolic tangent function. Generally, this convolutional result will be fed into a max-pooling operation to extract the most significant features. However, these extracted features are independent, and the correlation information among the local features are not captured. GRU has the ability to make up for this deficiency by using a gating mechanism to capture short-term and long-term dependencies. Therefore, in this study, a GRU layer is stacked on top of the convolutional layer to continue the feature extraction work. GRU layer --------- Similar to the long short-term memory (LSTM) unit with a memory cell and three gating units [@Hochreiter1997; @graves2012supervised], GRU is much simpler to compute because only two gating units are used to adaptively capture dependencies over different time scales: one is the reset gate $\mathbf{r}_j$, which controls how much information from the previous hidden state is kept in the candidate hidden state; another is the update gate $\mathbf{z}_j$, which decides how much previous information contributes and how much information from the candidate hidden state is added. The computational process are demonstrated by the following equations: $$\begin{aligned} \mathbf{r}_j &= \sigma(W_r\cdot C_j + U_r\cdot \mathbf{h}_{j-1} + \mathbf{b}_r), \\ \mathbf{z}_j &= \sigma(W_z\cdot C_j + U_z\cdot \mathbf{h}_{j-1} + \mathbf{b}_z), \\ \tilde{\mathbf{h}}_j &= tanh(W_h\cdot C_j + \mathbf{r}_j\odot (U_h\cdot \mathbf{h}_{j-1}) + \mathbf{b}_h), \\ \mathbf{h}_j &= (1-\mathbf{z}_j)\odot \mathbf{h}_{j-1} + \mathbf{z}_j\odot \tilde{\mathbf{h}}_j,\end{aligned}$$ where $\sigma$ is the logistic sigmoid function, $\odot$ stands for the element-wise multiplication, $C_j$ is the current n-gram feature representation (mentioned in Section \[sec:conv\]), $\mathbf{h}_{j-1}$ and $\tilde{\mathbf{h}}_j$ are the previous and the candidate hidden state, respectively, and $\mathbf{h}_j\in\mathbb{R}^{d^h}$ is the current hidden state. $W_r$, $U_r$, $\mathbf{b}_r$, $W_z$, $U_z$, $\mathbf{b}_z$, $W_h$, $U_h$ and $\mathbf{b}_h$ are weight matrices to be learned. We use a bidirectional GRU [@bahdanau2014neural] to encode the n-gram feature representations, which contains a forward GRU and a backword GRU. A sequence of forward hidden states $(\overrightarrow{\mathbf{h}}_1, \dots, \overrightarrow{\mathbf{h}}_{n-k+1})$ and a sequence of backward hidden states $(\overleftarrow{\mathbf{h}}_1, \dots, \overleftarrow{\mathbf{h}}_{n-k+1})$ are obtained. The final $j$-th hidden state can be achieved by concatenating the $j$-th forward and backward hidden state: $\mathbf{h}_j = {[{\overrightarrow{\mathbf{h}}_j}^\mathrm{T}, {\overleftarrow{\mathbf{h}}_j}^\mathrm{T}]}^\mathrm{T}$, which contains the dependencies of the preceding and the following n-gram features. Pooling layer ------------- Two different kinds of pooling schemes are adopted to generate the semantic representation of the relation sample $\mathbf{rs}$. [**Max pooling**]{} can be seen as a down-sampling operation that aims to extract the most significant features. After using this operation in our network, the $i$-th feature value $\mathbf{rs}_i$ is calculated by $$\mathbf{rs}_i = max([\mathbf{h}_1]_i, \dots, [\mathbf{h}_{n-k+1}]_i),$$ where $[\mathbf{h}_j]_i$ denotes the $i$-th element in vector $\mathbf{h}_j$. And all these features constitute the semantic representation of the relation sample $\mathbf{rs} = {(\mathbf{rs}_1, \dots, \mathbf{rs}_{d^h})}^\mathrm{T}$. [**Attentive pooling**]{} Given the output of the GRU layer $H=[\mathbf{h}_{1}, \dots, \mathbf{h}_{n-k+1}]$, we follow the attention mechanism used in [@Zhou2016], and the representation $\mathbf{rs}$ is formed: $$\begin{aligned} \boldsymbol{\alpha} &= softmax(\mathbf{v}^\mathrm{T}\cdot tanh(H)), \\ \mathbf{rs} &= tanh(H\cdot {\boldsymbol{\alpha}}^\mathrm{T}),\end{aligned}$$ where $\mathbf{v}$ is a model parameter vector and $\boldsymbol{\alpha}$ is a weight vector to measure which parts of the GRU output are relatively significant for the relation classification. Fully connected layer --------------------- We apply a softmax classifier to achieve the confidence scores with a class embedding matrix $W^{cs}$: $$\mathbf{s}_{\theta} = softmax(W^{cs}\cdot \mathbf{rs}),$$ where $\theta$ is the model parameter set. $\mathbf{s}_{\theta}^y$ is the confidence score of the true relation type $y$, and the loss function can be defined as $$\mathcal{L} = -\frac{1}{m}\sum_{i=1}^{m}{\log \mathbf{s}_{\theta}^y} + \beta||\theta||^2,$$ where $m$ is the sample size and $\beta$ is the $l_2$ regularization parameter. Experiments =========== Dataset and experimental settings --------------------------------- Experiments are conducted on the 2010 i2b2/VA relation dataset[^1] and the WI relation dataset[^2]. The former dataset comprises 426 English discharge summaries (170 for training and 256 for test), and the latter dataset contains 992 Chinese clinical records (521 for training and 471 for test). The relation types and their counts in these two datasets are listed in Table \[tab:relationcount\]. As stipulated in the official evaluation metric in the 2010 i2b2/VA challenge, the model performance is based on the micro-averaged F1 score over all positive relation types. [XXXXXX]{} &\ (lr)[1-3]{} (lr)[4-6]{} Relation & Train & Test & Relation & Train & Test\ TrIP & 51 & 152 & TrID & 103 & 92\ TrWP & 24 & 109 & TrWD & 38 & 27\ TrCP & 184 & 342 & TrAD & 221 & 166\ TrAP & 885 & 1732 & NTrD & 675 & 656\ TrNAP & 62 & 112 & TrIS & 337 & 215\ NTrP & 1702 & 2759 & TrWS & 297 & 242\ TeRP & 993 & 2060 & TrCS & 125 & 176\ TeCP & 166 & 338 & TrAS & 334 & 238\ NTeP & 993 & 1974 & NTrS & 1062 & 901\ PIP & 755 & 1448 & TeRD & 301 & 227\ NPP & 4418 & 8089 & NTeD & 331 & 248\ (lr)[1-3]{} (lr)[1-3]{} SID & 969 & 620 & TeRS & 527 & 542\ DCS & 228 & 181 & TeAS & 313 & 564\ NDS & 777 & 635 & NTeS & 8628 & 7060\ Positive relations were annotated in both relation datasets, and samples of negative relation types (starting with “N" in this table) were extracted to ensure each concept pair within a sentence could be assigned a certain relation type. For more details of these relation types, please refer to [@Uzuner2011; @He2016]. In our methods, the initial word representations and the other matrices are randomly initialized by normalized initialization [@DBLP:journals/jmlr/GlorotB10], and a 5-fold cross-validation is used on the training set to tune the model hyperparameters. The selected hyperparameter values are: word embedding size $d^w$, 100; word position embedding size $d^p$, 10; convolutional size $d^c$, 200; context window size $k$, 3; GRU dimension $d^h$, 100; learning rate, 0.01. Adam technique [@kingma2014adam] is utilized to optimize our loss function. We use both $l_2$ regularization and dropout technique [@Srivastava2014] to avoid overfitting, and the values are set to 0.0001 and 0.5, respectively. Baselines --------- ### 2010 i2b2/VA relation dataset When doing experiments on this dataset, the previous methods [@DBLP:conf/coling/DSouzaN14; @Sahu2016Relation; @Raj2017] followed inconsistent data split schemes. In order to compare these methods together, we choose the split scheme in [@DBLP:conf/coling/DSouzaN14], which is also the official data split. [**SVM**]{}: due to the dataset available to the research community is only a subset of the dataset used in the 2010 i2b2/VA challenge, so @DBLP:conf/coling/DSouzaN14 reimplemented the state-of-the-art model in the challenge [@Rink2011] and reevaluated this model on the relation dataset accessible. [**SVM+ILP**]{}: @DBLP:conf/coling/DSouzaN14 also proposed a better single-model method and an ensemble-based method within an integer linear programming (ILP) framework. In these feature-based state-of-the-art methods, a variety of external features sets are used, such as part-of-speech (POS) tagging and dependency parsing. In this work, three previous neural network methods are reimplemented and reevaluated. [**CNN**]{}: a multiple-filter CNN with max-pooling proposed by @Sahu2016Relation. To evaluate the model performance independent of the external features, POS and chunk features used in this method are removed. [**CRNN-Max and CRNN-Att**]{}: a two-layer model comprising recurrent and convolutional layers with max and attentive pooling [@Raj2017]. However, only word embeddings were used in their work. In order to maintain a fair comparison, word position embeddings are added in our model reimplementation. In these three baseline reimplementations, we follow the selected hyperparameters used in the corresponding work and the word embeddings are pre-trained on the deidentified notes from the MIMIC-III database [@Johnson2016]. ### WI relation dataset [**SVM**]{}: this model is implemented using scikit-learn[^3]. And it involves the following features: entity $e_1$, entity $e_2$, entity type $et_1$, entity type $et_2$, distance between $e_1$ and $e_2$, words in $e_1$ and $e_2$, words between $e_1$ and $e_2$, words behind $e_2$, POS of words in $e_1$ and $e_2$, POS of words between $e_1$ and $e_2$, and POS of words behind $e_2$. [**CNN**]{}: the model version of C-BGRU-Max after removing the GRU layer, which is a CNN-based model. Experimental results -------------------- ### System performance The performance results are displayed in Table \[tab:classifiers\] and \[tab:classifiers\_ct\], including 95% confidence intervals for the models we implemented, which are derived using bootstrapping [@DiCiccio1996]. We use the same bootstrapping method described in [@gao2017hierarchical]. We observe that our C-BGRU-Max model outperforms the previous single-model methods significantly in both datasets, without using any external features. After using attentive pooling, the model performance on the two datasets shows different changes: drops on the 2010 i2b2/VA relation dataset but increases on the WI relation dataset. The intuitive explanation is that descriptions in English discharge summaries tend to be more colloquial, making specific features more difficult to capture. More details of the category-wise and class-wise performance comparisons are listed in Table \[tab:category\], \[tab:subclasses\], \[tab:category\_ct\], and \[tab:subclasses\_ct\]. [llXXX]{} Classifier & & P & R & F1\ \ SVM$^*$ [@Rink2011] & Set1 & 58.1 & **66.7 & 62.1\ SVM+ILP [@DBLP:conf/coling/DSouzaN14] & Set2 & **75.0 & 58.9 & 66.0\ & & & &\ & & & &\ & & & &\ & & & &\ & & & &\ \ Ensemble+ILP$^{\circ}$ [@DBLP:conf/coling/DSouzaN14] & Set2 & 66.7 & 72.9 & 69.6\ **** The symbol $*$ indicates that this model is reimplemented by [@DBLP:conf/coling/DSouzaN14] on the relation dataset available to the research community, due to the accessible dataset is only a subset of that used in the 2010 i2b2/VA challenge. The symbol $\circ$ indicates that this classifier is the ensemble of 5 independent models. The bold item is the best result. Set1: POS, chunk, semantic role labeler, word lemma, dependency parse, assertion type, sentiment category, Wikipedia. Set2: POS, chunk, semantic role labeler, word lemma, dependency parse, assertion type, sentiment category, Wikipedia, manually labeled patterns. POS, part-of-speech; ILP, integer linear programming. [lXXX]{} Classifier & P & R & F1\ SVM & 72.9 & 63.9 & 68.1\ & & &\ & & &\ & & &\ The bold item is the best result. [lXXXXXXXXX]{} Classifier & & &\ (lr)[2-4]{}(lr)[5-7]{}(lr)[8-10]{} & P & R & F1 & P & R & F1 & P & R & F1\ & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$)\ CNN [@Sahu2016Relation] & 60.9 & 48.2 & 53.8 & 75.8 & 69.2 & 72.3 & 64.8 & 43.3 & 51.9\ & 1.0 & 0.9 & 0.9 & 0.9 & 0.8 & 0.7 & 1.3 & 1.1 & 1.1\ CRNN-Max [@Raj2017] & 58.4 & 53.8 & 56.0 & 73.3 & 73.1 & 73.2 & 61.6 & 54.4 & 57.8\ & 0.9 & 0.9 & 0.8 & 0.8 & 0.8 & 0.7 & 1.1 & 1.2 & 1.0\ CRNN-Att [@Raj2017] & 55.2 & 50.8 & 52.9 & 70.1 & 73.8 & 71.9 & 63.3 & 46.3 & 53.5\ & 0.9 & 0.9 & 0.8 & 0.8 & 0.8 & 0.7 & 1.3 & 1.2 & 1.1\ C-BGRU-Max & 62.7 & 59.7 & & 78.4 & 77.5 & & 64.8 & 58.9 &\ & 0.9 & 0.9 & 0.8 & 0.8 & 0.8 & 0.6 & 1.2 & 1.1 & 1.0\ C-BGRU-Att & 63.9 & 57.1 & & 79.4 & 72.5 & & 62.8 & 60.0 &\ & 0.9 & 0.9 & 0.8 & 0.7 & 0.8 & 0.7 & 1.1 & 1.2 & 1.0\ TrP, Treatment-Problem; TeP, Test-Problem; PP, Problem-Problem. CI($\pm$) is confidence interval for P, R, and F1. The bold item is the best result. Compared with previous models, the underlined item is statistically significant. ----------------------------------------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- Classifier (lr)[2-4]{}(lr)[5-7]{}(lr)[8-10]{}(lr)[11-13]{}(lr)[14-16]{}(lr)[17-19]{}(lr)[20-22]{}(lr)[23-25]{} P R F1 P R F1 P R F1 P R F1 P R F1 P R F1 P R F1 P R F1 CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CI($\pm$) CNN [@Sahu2016Relation] 18.7 5.9 9.0 0.0 0.0 0.0 56.6 30.1 39.3 63.0 61.1 62.0 35.9 9.1 **14.5 & 77.7 & 77.8 & 77.7 & 45.0 & 16.9 & 24.5 & 64.8 & 43.3 & 51.9\ & 5.0 & 1.7 & 2.5 & 0.0 & 0.0 & 0.0 & 3.4 & 2.2 & 2.4 & 1.0 & 1.0 & 0.9 & 7.5 & 2.3 & 3.5 & 0.8 & 0.8 & 0.6 & 3.9 & 1.7 & 2.3 & 1.3 & 1.1 & 1.1\ CRNN-Max [@Raj2017] & 33.0 & 4.6 & 8.1 & 0.0 & 0.0 & 0.0 & 41.9 & 28.1 & 33.6 & 61.3 & 69.8 & 65.3 & 11.7 & 3.2 & 5.0 & 76.9 & 80.1 & 78.5 & 41.8 & 30.4 & 35.2 & 61.6 & 54.4 & 57.8\ & 8.9 & 1.5 & 2.6 & 0.0 & 0.0 & 0.0 & 2.9 & 2.1 & 2.3 & 1.0 & 1.0 & 0.8 & 5.1 & 1.4 & 2.2 & 0.8 & 0.8 & 0.6 & 2.8 & 2.2 & 2.2 & 1.1 & 1.2 & 1.0\ CRNN-Att [@Raj2017] & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 34.1 & 10.8 & 16.4 & 56.4 & 69.6 & 62.3 & 0.0 & 0.0 & 0.0 & 71.5 & 83.0 & 76.8 & 45.7 & 17.8 & 25.6 & 63.3 & 46.3 & 53.5\ & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 0.0 & 3.7 & 1.4 & 2.0 & 1.0 & 1.0 & 0.8 & 0.0 & 0.0 & 0.0 & 0.8 & 0.7 & 0.6 & 3.9 & 1.9 & 2.4 & 1.3 & 1.2 & 1.1\ C-BGRU-Max & 51.4 & 4.7 & 8.7 & 36.4 & 0.7 & & 52.1 & 38.9 & & 64.4 & 75.8 & & 38.5 & 7.1 & 12.0 & 79.9 & 84.4 & & 61.1 & 35.3 & & 64.8 & 58.9 &\ & 12.0 & 1.5 & 2.7 & 29.2 & 0.7 & 1.3 & 2.7 & 2.3 & 2.1 & 1.0 & 0.9 & 0.8 & 9.7 & 2.1 & 3.4 & 0.8 & 0.7 & 0.6 & 3.1 & 2.3 & 2.4 & 1.2 & 1.1 & 1.0\ C-BGRU-Att & 43.8 & 11.1 & & 26.7 & 2.9 & & 48.5 & 41.1 & & 67.7 & 70.9 & & 30.1 & 8.8 & 13.6 & 81.2 & 79.3 & & 58.7 & 30.8 & & 62.8 & 60.0 &\ & 7.0 & 2.2 & 3.3 & 10.9 & 1.4 & 2.4 & 2.5 & 2.2 & 2.0 & 0.9 & 1.0 & 0.8 & 7.0 & 2.3 & 3.4 & 0.8 & 0.8 & 0.6 & 3.2 & 2.2 & 2.4 & 1.1 & 1.2 & 1.0\ ** ----------------------------------------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- ----------- CI($\pm$) is confidence interval for P, R, and F1. The bold item is the best result. Compared with previous models, the underlined item is statistically significant. [lXXXXXXXXXXXXXXX]{} Classifier & & & & &\ (lr)[2-4]{}(lr)[5-7]{}(lr)[8-10]{}(lr)[11-13]{}(lr)[14-16]{} & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1\ & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$)\ CNN & 59.5 & 56.3 & 57.8 & 58.1 & 50.5 & 54.1 & 90.1 & 90.7 & 90.4 & 86.9 & 59.0 & 70.3 & 72.5 & 82.7 & 77.3\ & 2.6 & 2.7 & 2.5 & 1.5 & 1.5 & 1.3 & 1.7 & 1.7 & 1.2 & 1.1 & 1.3 & 1.1 & 1.3 & 1.2 & 1.0\ C-BGRU-Max & 59.9 & 60.4 & 60.1 & 59.3 & 54.2 & 56.7 & 91.7 & 92.0 & 91.8 & 85.5 & 64.7 & & 73.6 & 84.7 & **78.8\ & 2.4 & 2.5 & 2.3 & 1.5 & 1.5 & 1.3 & 1.6 & 1.5 & 1.1 & 1.1 & 1.3 & 1.0 & 1.2 & 1.1 & 1.0\ C-BGRU-Att & 61.8 & 58.7 & **60.2 & 64.4 & 55.5 & & 91.6 & 93.6 & **92.5 & 82.6 & 68.0 & & 75.0 & 80.8 & 77.8\ & 2.6 & 2.6 & 2.5 & 1.5 & 1.5 & 1.4 & 1.6 & 1.4 & 1.1 & 1.1 & 1.2 & 0.9 & 1.3 & 1.2 & 1.0\ ****** TrD, Treatment-Disease; TrS, Treatment-Symptom; TeD, Test-Disease; TeS, Test-Symptom; DS, Disease-Symptom. CI($\pm$) is confidence interval for P, R, and F1. The bold item is the best result. Compared with CNN, the underlined item is statistically significant. [lXXXXXXXXXXXXXXXXXX]{} Classifier & & & & & &\ (lr)[2-4]{}(lr)[5-7]{}(lr)[8-10]{}(lr)[11-13]{}(lr)[14-16]{}(lr)[17-19]{} & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1\ & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$)\ CNN & 53.7 & 39.6 & 45.6 & 53.4 & 28.9 & 37.5 & 62.0 & 70.0 & 65.8 & 50.6 & 43.7 & 46.9 & 71.5 & 69.0 & 70.2 & 69.0 & 20.2 & 31.3\ & 5.4 & 4.7 & 4.5 & 11.7 & 7.6 & 8.4 & 3.1 & 3.2 & 2.6 & 3.0 & 3.0 & 2.6 & 2.5 & 2.7 & 2.0 & 5.5 & 2.7 & 3.5\ C-BGRU-Max & 53.7 & 39.1 & 45.3 & 56.6 & 44.4 & **49.8 & 62.3 & 74.7 & **67.9 & 54.3 & 42.0 & 47.3 & 75.3 & 70.5 & **72.8 & 57.5 & 27.7 & 37.4\ & 5.4 & 4.4 & 4.2 & 9.5 & 8.6 & 8.0 & 2.9 & 3.0 & 2.4 & 3.4 & 3.0 & 2.7 & 2.5 & 2.6 & 2.0 & 4.7 & 2.9 & 3.3\ C-BGRU-Att & 57.9 & 40.7 & **47.8 & 58.7 & 40.0 & 47.6 & 63.5 & 71.7 & 67.3 & 58.8 & 50.0 & & 75.8 & 69.5 & 72.5 & 66.4 & 26.2 & **37.6\ & 5.5 & 4.6 & 4.4 & 10.4 & 8.4 & 8.1 & 3.1 & 3.1 & 2.6 & 3.2 & 3.0 & 2.7 & 2.5 & 2.6 & 2.1 & 5.0 & 2.9 & 3.4\ & & & & & &\ (lr)[2-4]{}(lr)[5-7]{}(lr)[8-10]{}(lr)[11-13]{}(lr)[14-16]{}(lr)[17-19]{} & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1 & P & R & F1\ & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$) & CI($\pm$)\ CNN & 50.2 & 60.3 & 54.8 & 90.1 & 90.7 & 90.4 & 88.8 & 84.1 & **86.4 & 82.9 & 35.0 & 49.2 & 56.6 & 63.2 & 59.7 & 77.0 & 88.4 & 82.3\ & 2.6 & 2.8 & 2.3 & 1.7 & 1.7 & 1.2 & 1.2 & 1.3 & 0.9 & 2.1 & 1.8 & 1.9 & 3.1 & 3.2 & 2.6 & 1.4 & 1.2 & 1.0\ C-BGRU-Max & 51.1 & 68.3 & 58.5 & 91.7 & 92.0 & 91.8 & 85.4 & 84.7 & 85.1 & 85.6 & 45.4 & & 56.3 & 69.0 & **62.0 & 79.1 & 89.4 & **83.9\ & 2.4 & 2.7 & 2.1 & 1.6 & 1.5 & 1.1 & 1.4 & 1.4 & 1.0 & 1.7 & 1.9 & 1.7 & 2.8 & 3.0 & 2.4 & 1.3 & 1.1 & 0.9\ C-BGRU-Att & 58.6 & 67.7 & & 91.6 & 93.6 & **92.5 & 80.6 & 86.0 & 83.2 & 86.2 & 50.7 & & 58.7 & 61.7 & 60.2 & 79.6 & 86.4 & 82.9\ & 2.6 & 2.6 & 2.2 & 1.6 & 1.4 & 1.1 & 1.4 & 1.3 & 1.0 & 1.6 & 1.9 & 1.7 & 3.2 & 3.2 & 2.7 & 1.4 & 1.3 & 1.0\ ****************** CI($\pm$) is confidence interval for P, R, and F1. The bold item is the best result. Compared with CNN, the underlined item is statistically significant. ### Discussion of attentive pooling As show in Table \[tab:classifiers\], the F1 scores of CRNN-Att and C-BGRU-Att are lower than that of CRNN-Max and C-BGRU-Max, respectively. This indicates that the attention mechanism, which presents a positive effect in the general domain [@Zhou2016; @kim2017multiple], does not show any performance improvement on the 2010 i2b2/VA relation dataset. In this dataset, there exist $\sim$3.3 entities in each sentence on average. Therefore, input representations of relation samples generated from the same sentence are quite similar, and the only difference is that some of the word position representations between these relation samples are different, which may not be able to show sufficient sample differentiation. In addition, attentive pooling does not extract the most significant features like max-pooling, which may lead to relative deficiencies in distinguishing model similar samples. We will try to analysis and validate these speculations in our future work. ### F1 score vs. distance Figure \[fig:distance\]a and \[fig:distance\_ct\]a show the frequency distribution of different distances in the two datasets, and Figure \[fig:distance\]b and \[fig:distance\_ct\]b depict the trend of the F1 score as the distance increases. The F1 score is the average value of the relation samples belonging to the distance window $[d - 2, d + 2]$. In order to ensure the reliability of the evaluation, the maximum distance value with a statistic greater than 20 is selected as the truncation of the distance value. On the 2010 i2b2/VA relation dataset, C-BGRU-Max and C-BGRU-Att outperform the baselines over all distances. On the WI relation dataset, C-BGRU-Max and CNN do not show significant differences when the distance is less than 20, but as the distance increases, the performance gap gradually expands. These results verifies that our model has the ability to learn long-term dependencies and this information works in the relation classification task. {width="\textwidth"} {width="\textwidth"} Related Work ============ Before deep learning research became popular, statistical machine learning methods were the main approaches in the relation classification task. Most of the researchers in the general and clinical domain focused on feature-based and kernel-based methods [@Bunescu2005; @hendrickx2009semeval; @Rink2010; @Zhu2013; @Kim2015]. In recent years, researchers have gradually tried the effect of deep learning methods in the relation classification task and achieved satisfactory results. A variety of deep architectures have been proposed to classify the relations, such as recurrent neural network (MV-RNN) [@Socher2012], CNN with softmax classification [@Zeng2014], factor-based compositional embedding model (FCM) [@yu2014factor], and word embedding-based models [@hashimoto-EtAl:2015:CoNLL]. Next, there exist many RNN-based and CNN-based variants. Because the max-pooling operation in CNN models will lose significant linguistic features in a sentence, some researchers introduced dependency trees for this work, e.g., bidirectional long short-term memory networks (BLSTM) [@Zhang2015a], dependency-based neural networks (DepNN) [@liu-EtAl:2015:ACL-IJCNLP], shortest dependency path-based CNN [@xu-EtAl:2015:EMNLP1], long short term memory networks along shortest dependency paths (SDP-LSTM) [@Xu2015], deep recurrent neural networks (DRNN) [@xu-EtAl:2016:COLING1], and jointed sequential and tree-structured LSTM-RNN [@miwa-bansal:2016:P16-1]. Although the above studies achieved solid results, further research was devoted to eliminating the dependence on the NLP parser because of its limited performance. @DBLP:conf/acl/SantosXZ15 proposed a new pairwise ranking loss function, and only two class representations were updated in every training round. Similarly, @Wang2016 introduced a pairwise margin-based loss function and multi-level attention mechanism and achieved the new state-of-the-art results for relation classification. More recently, neural network methods have show promising performance for relation classification on clinical records. @Sahu2016Relation proposed a multiple-filter CNN with some linguistic features, and experiments on the 2010 i2b2/VA relation dataset verified the effectiveness of the neural network model for medical relation classification. @Raj2017 trained a two-layer model by feeding short phrase features extracted by a bidirectional LSTM layer into CNN, and the model performed better than CNN on relation samples where the distance between the medical concepts are large. Different from @Raj2017’s study, we think n-gram features and sequential correlations among them are the key to relation classification, so we explore another unified architecture that utilizes the strengths of CNN and RNN simultaneously. Conclusion ========== In this paper, we present a unified architecture based on the combination of CNN and RNN to classify medical relations in English and Chinese clinical records. Our model captures long-term dependencies of phrase-level features through a bidirectional GRU layer and this information improves model performance. To the best of our knowledge, this is the first time that neural network methods have been used to classify relations in Chinese clinical text. Experiments show that the proposed model achieves a significant improvement over comparable methods on the 2010 i2b2/VA relation dataset and the WI relation dataset. Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to thank all the anonymous reviewers for their insightful comments. We would also like to thank the data support from the 2010 i2b2/VA challenge. [^1]: The relation dataset is available at <https://www.i2b2.org/NLP/Relations/>. [^2]: <https://github.com/WILAB-HIT/Resources/tree/master/entity_assertion_relation> [^3]: <http://scikit-learn.org/stable/>.

ArXiv

.25in \#1\#2\#3\#4[ [\#1]{} [**\#2**]{} [\#3]{} [(\#4)]{} ]{} \#1 \#1[\^[\#1]{}]{} \#1[\^[\#1]{}]{}\#1[\^[\#1]{}]{} \#1 \#1[\_[FT]{}]{} \#1[\_[FT]{}\^T]{} 0[A\_0]{} SNU-TP 01-005 3.5cm [Meson Exchange Effect on Color Superconductivity ]{} 2.3cm [Youngman Kim, Young-Ho Song and Dong-Pil Min ]{} 0.5cm [*Department of Phyiscs, Seoul National University, Seoul 151-742, Korea*]{} () **Abstract** 0.5cm We investigate the effects of pion and gluon exchanges on the formation of two-flavor color superconductivity at moderate density, $\mu <1 GeV$. The chiral quark model proposed by Manohar and Georgi containing pions as well as gluons is employed to show that the pion exchange reduces substantially the value of the superconducting gap gotten with the gluon exchange only. It turns out that the pion exchanges produce a repulsion between quark-quark pair in a spin and isospin singlet state. We suggest that the phase consisiting of pions, gluons and quarks is one of the candidates of in-medium QCD phase at moderate density. Introduction ============ Due to the attractive interactions between two quarks from one-gluon exchange or ’t Hooft interaction induced by instanton, there is a tendency toward spontaneous breaking of (color) gauge symmetry forming color supercondutivity and especially color-flavor-locking(CFL) [@csc][@arw]. The magnitude of the superconducting gap $\Delta$ is estimated by using models whose parameters are chosen to reproduce reasonable zero density physics [@csc][@arw] or from the perturbative one-gluon exchange calculations which are valid at asymptotically high density[@tf][@son]. Recently, it is shown that at intermediate densities $\mu\sim 500 MeV$ , large enough for the system to be in the quark phase but small enough to support nonperturbative interactions, Cooper pairs with nonzero total momentum are favored leading to gaps which vary in space[@sbmit]. For recent reviews on the color superconductivity in quark matter, see Ref. [@rwh]. On the other hand, the dense nuclear matter has been successfully described in terms of the hadrons, which will be called [*hadronic*]{} phase. It was already conceived that the boson condensation, particularly kaon condensation, can be produced at the relatively low density, about 3-4 times of the normal nuclear matter density. Indeed, the hadrons as the quasi-particles are subdued to the pattern of BR (Brown-Rho) scaling[@br], whose relation to the results based on the conventional hadronic many-body calculation[@rapp] is investigated in Ref.[@krbr]. Between the hadronic and quark phases, we could expect a phase whose relevant degrees of freedom are mesons as Goldstone bosons of spontaneous breaking of chiral symmetry by non-zero value of chiral order parameter $\la\bar\psi\psi\ra$ and gluons due to lack of confinement as well as constituent (quasi) quarks[@mg], and we will call this phase as the CQ(constituent or chiral quark) phase. Phase diagram for QCD involving hadronic(nuclear) matter and quark matter(2SC, CFL) is conjectured and depicted in [@unlock]. We propose that the CQ phase is one of the candidates of in-medium QCD phase at moderate density. A recent study on baryons in matter-free space reveals that the ground states and excitation spectra of light and strange baryons is governed by the Goldstone boson exchange(GBE) as well as the harmonic confinement potential between constituent quarks (quasi-quarks) [@gloz]. Based on the decoupling of vector interactions at high density, it is argued in Ref. [@gmc] that quasi-particle picture involving (quasi) quarks becomes more appropriate at higher densities as chiral restoration is approached. The goal of this work is to study two-flavor color superconductivity (2SC) in the CQ phase and therefore to investigate a phase boundary of hadronic matter and quark matter (especially 2SC). We calculate the gap of two-flavor color superconductivity in the CQ phase using the chiral quark model proposed by Manohar and Georgi[@mg]. Note that the model [@mg] is valid around the energy scale $E_{CQ}$ given in matter-free space $\Lambda_{QCD}<E_{CQ}<\Lambda_{\chi SB}\sim 1 GeV $. We expect, however, that the model is relevent to describe QCD at moderate density since we can argue that $\mu \sim E_{CQ}$ based on the renormalization group analysis at finite density[@man]. In section 2, we introduce the chiral quark model and derive a gap equation for the Cooper pair from one gluon and one pion exchange. To see the physics concealed in the gap equation economically, we adopt contact four-Fermi interactions, which are assumed to reproduce the physics of one gluon and one pion exchange, and solve the gap equation in section 3. Discussions are presented in section 4. Schwinger-Dyson equation in the Chiral Quark Model ================================================== The chiral quark model is defined between $\Lambda_{QCD}\sim 200 MeV$ and $\Lambda_{\chi SB}\sim 1 GeV$ and therefore contains mesons as Goldstone bosons of $SU(3)_L\times SU(3)_R$ spontaneous chiral symmetry breaking. And gluons as gauge bosons of $SU(3)_{color}$ should be also present in this phase of deconfinement. The chiral quark lagrangian is given by &=&[|]{}(i [ D]{}+[ V]{})+ g\_[A]{}[|]{}[ A]{}\_5 - m [|]{}+ \^0 &&+ f\_\^2 tr(\^\^\_) - tr(F\^F\_) + …\[lag\] where D\_&=&\_+i g G\_,    G\_=G\_\^a T\^a, V\_&=&[i 2]{}(\^\_+\_\^),A\_&=&[i 2]{}(\^\_-\_\^),&=&e\^[(i /f\_)]{},    f\_93  MeV,  =and =[1 ]{}( [ccc]{} \^0+& \^+& K\^+\ \^-&-\^0+& K\^0\ K\^- & |[K]{}\^0& -[2 ]{} ). Note that in the chiral quark model, [*we can establish a power counting argument to deal with non-renormalizable terms and to neglect internal gluon lines*]{} [@mg]. Since the chiral quark model is valid below the chiral symmetry breaking scale, the value of the chiral order parameter $\la\bar\psi\psi\ra$ is not zero. We expect, however, the value of $\la\bar\psi\psi\ra$ in CQ phase to be quite small compared to that in the matter-free space. To see the effects of GBE on color superconductivity, we consider one pion exchange as well as one gluon exchange with $N_f=2$. As it is described in the conjectured phase diagram for QCD at zero temperature  [@abr], when chemical potential exceeds the constituent strange quark mass the more relevant phase will be $2SC+s$ in which we have $ss$ condensates as well as Cooper pairs composed of $u$ and $d$ quarks. We can, however, neglect the difference between the $2SC$ and $2SC+s$ since the $ss$ condensate is expected to be small [@abr]. In this work, we ignore $ss$ condensate and concentrate on $2SC$. Since the value of chiral order parameter is assumed to be small, we may take the light chiral quark masses to be negligible compared to chemical potential. There are several works on the possibility of mixed phase of both chiral condensate and color superconductivity[@ms]. In this work, we use the Nambu-Gorkov formalism with $$\Psi \equiv \left( \begin{array}{c} \psi \\ {\bar\psi}^T \end{array} \right) \equiv\left( \begin{array}{c} \psi \\ \psi_c \end{array} \right).$$ The inverse quark propagator [^1] is S\^[-1]{}(q)=( [cc]{}[ q]{}+\_0-m&|\ & ([ q]{}-\_0+m)\^T ) where $\bar\Delta =\gamma_0\Delta^{\dag}\gamma_0$. To leading order in the perturbative expansion, the quark-gluon vertex matrix is $-i g \Gamma^a_{\mu}$ and the quark-pion vertex matrix is given by $- (g_A / f_\pi)\Gamma^i(q)$. \^a\_&=& ( [cc]{}T\^a\^ &0\ 0 &-(T\^a\^)\^T ) \^i(q)&=& ( [cc]{}\_5 &0\ 0 &-(\_5)\^T ) The gluon propagator in the hard dense loops (HDL) approximation is given by D\_(q)=+- where P\_[ij]{}\^T&=&\_[ij]{}-[q]{}\_i[q]{}\_j ,  P\_[00]{}\^T=P\_[0i]{}\^T=0 P\_\^L&=&-g\_+-P\_\^T For $q_0\ll{\vec q}\rightarrow 0 $ and to leading order in perturbation theory we have, F=2 m\^2,  G=[2]{}m\^2[q\_0 |[q]{}|]{},with $m^2=N_f g^2\mu^2/(4\pi^2) $. The pion propagator is $D(k)=1/k^2$. The Schwinger-Dyson equation for gap matrix $\Delta$ becomes, (k)&=&i g\^2 (-T\^a\^)\^T S\_[21]{}(q) (T\^b\^) D\_(q-k) &&+ i()\^2 (-([ q]{}-[ k]{})\_5)\^T S\_[21]{}(q) (([ k]{}-[ q]{})\_5)\^[ij]{} D(q-k) Let us take the form of the gap matrix as [@tf] \^[ab]{}\_[ij]{}(q)=(\_2)\^[ab]{}(\_2)\_[ij]{}C\_5 \[\_1(q\_0) +\_2(q\_0)\] where ${\vec \alpha}=\gamma_0 {\vec \gamma} $. By inverting the inverse quark propagator matrix $S^{-1}(q) $, we obtain the 21-component of $S(q)$ S\_[21]{}(q)=-(\_2\_2C\_5) ( \_[-]{} +\_[+]{})where $\Lambda_{+}=(1+{\vec \alpha}\cdot{\hat q})/2 $ and $\Lambda_{-}=(1-{\vec \alpha}\cdot{\hat q})/2 $. For color and isospin, we can use the following relation, (\_a)\^T\_2\_a=-\_2 =-\_2 , (N\_c =3) where we have used $(\lambda^a)_{ij}(\lambda^a)_{kl}=-2/N\delta_{ij}\delta_{kl} +2\delta_{il}\delta_{jk}$. For isospin $N_f=2$, we use $\frac{1}{4}(\tau_i)^T\tau_2\tau_i=-\frac{3}{4}\tau_2$. Then the gap equation becomes (k)&=& \_2\_2 C\_5 (\_1(k\_0) \_[+]{}+\_2(k\_0)\_[-]{}) &=& i g\^2(+\_2\_2 C\_5) \_(\_[-]{} && +\_[+]{} ) \_D\^(q-k) &&+i ()\^2 (+\_2\_2C\_5) ([ q]{}-[ k]{})\_5 (\_[-]{} && +\_[+]{} )( q- k) \_5 D(q-k) Following the same procedure shown in Ref. [@tf], we get the gap equation, (k\_0) &=& [g\^2 18\^2]{}\^\_0 dq\_0 ln([b |k\_0-q\_0|]{}) &&-[3 \^2 16 \^2]{}[g\_A\^2 f\_\^2]{}\^\_0 dq\_0 ln (). \[gapeq\] From (\[gapeq\]) we can see that one pion exchange gives repulsive interaction in the quark-quark potential and therefore the magnitude of the gap will be reduced in the presence of the GBE. Note that the quark pairs are in isosinglets and spin singlet state. The reason that one pion exchange(OPE) gives repulsive potential in the quark-quark potential can be understood from the similarity with nucleon-nucleon potential from OPE in the same isosinglet and spin singlet channel, which is repulsive[@ew] V\_(S=I=0)=. We note here that it is not simple to solve the gap equation (\[gapeq\]) analytically even with the approximation given in [@son]. A Toy Solution ============== The property of the gap equation (\[gapeq\]) can be understood by a toy solution in which we set forth the following four Fermi interactions mimicking the quark-quark interaction in the chiral quark model. Note that with a NJL type interaction our gap will have different dependence on the coupling constant $\Delta\sim\exp(-1/g^2)$ from the one with one-gluon exchange $\Delta\sim\exp(-1/g)$[@son]. We can see, however, the effects of the GBE on the gap explicitly in a simple manner with the NJL type interaction. \_[int]{} = -G d\^4 x ([|]{}\^a \^ ) ([|]{}\^a \_ ) +G\_ d\^4 x([|]{}\_5 )([|]{}\_5 )\[ml\] where the first term represents gluon induced interaction, and the second one is pion induced interaction. Here $\lambda^a$ is a generator of $SU(3)$ color group and $\vec\tau$ is for $SU(2)$ isospin group. Note that we choose positive sign in front of pion induced term considering repulsive OPE potential (\[gapeq\]) and take $G$ and $G_\pi$ to be positive. Since pion is almost massless, it is not simple to put the one-pion exchange into the four-Fermi form, but here we expect that the four-Fermi interaction with $G_\pi$ in (\[ml\]) could at least mimic the [*repulsive*]{} nature of the one-pion exchange calculated in section 2. Now it is straightforward to obtain a gap equation with the interacting Lagranigan (\[ml\]) and we get &=&(-) d q\_0 \[sgap\] Note that this gap equation can be traced from the equation (\[gapeq\]) by removing propagator dependence, [*i.e.*]{} by neglecting logarithmic term in (\[gapeq\]). For the sake of simplicity, we introduce cut-off $\Lambda$ and take $\Lambda\sim\mu$. Since the gap $\Delta$ is constant with the interaction (\[ml\]), we get a solution easily &=& 2 with $N \equiv \frac{2 G \mu^2}{3 \pi^2}-\frac{3 G_\pi \mu^2}{16 \pi^2}$. Now we can see that due to the repulsive nature of pion exchange potential, the value of the gap will be reduced by a factor $\sim\exp [-1/(G_\pi\mu^2) ]$. To estimate the value of the gap, we take $G\approx 5.8GeV^{-2}$ and G\_=c where $c$ is a constant. The value of $c$ could be determined by fully integrating out pions to obtain effective action containing four-Fermi interactions, but we take it as a free parameter and we shall use $g_A= 0.75 $ and $f_{\pi}= 93 MeV$ given in free space for numerical estimates. Taking $G\approx 5.8GeV^{-2}$ to reproduce the results from NJL model calculations with only single gloun exchange effects and choosing $c=0.05$($G_\pi\approx 3.3 GeV^{-2}$), we get the results depicted in Fig. \[cscp\]. We note here that when the value of $G_\pi$ is big enough to make $N$ to be negative, we see from (\[sgap\]) that there is a possibility to have no color superconducting gap at all. Discussion ========== In this work, we employ the chiral quark model to study the effect of one pion exchange on the two flavor color superconductivity. We found that OPE reduces the value of the gap $\Delta$, though it is difficult to be quantitative as to how much the gap will be reduced since we have no definte information on the value of $G_\pi$. Our work will have several important consequences : $\bullet$ At moderate density where CQ phase is plausible, pion as a Goldstone boson of spontaneous breaking of chiral symmetry by quark condensate $\la \bar \psi \psi \ra$ could do important role in the physics of color superconductivity, implying that at moderate density there could be some other interactions as well as one gluon exchange which have to be taken into account. Besides the GBE mode considered in this work, light vector mesons could come in as indicated by BR scaling [@br] and the “vector manifestation”[@hy] in which the chiral symmetry is restored at the critical point by the massless degenerate pion and (longitudinal) $\rho$-meson as the chiral partner. In the case of two-nucleon states of total isospin $0$ and spin $0$, the $\rho$ exchange potential is given by [@ew] V\_(S=I=0)=3\[- +2 \] where $f_\rho=(g_\rho +g_t)(m_\rho/2m_N)$. When we take the value of the couplings determined in free space, we can easily see that the $\rho$ exchange potential is repulsive. Therefore we expect [*naively*]{} that light vector mesons also produce repulsion between quark-quark pair we are considering. We note, however, that the potential becomes attractive when neglecting tensor coupling ($g_t=0$) and that density dependence of $g_\rho$ and $g_t$ should be properly taken into account. In the vector manifestation, we should also keep it in mind that the transverse $\rho$ is decoupled from the vector current. $\bullet$ The effects of diquark condensates on the cooling of compact stars is investigated by several authors, see for example  [@bt]. However it is expected that the quark matter core in neutron star is [*inert*]{} as far as cooling is concerned since the quasi-quarks contribution to the specific heat and neutrino emissivity is suppressed by a factor $\sim \exp (-\Delta /T )$ [@alford]. However, when GBE reduces the gap substantially, we expect that quark matter core could play important role in cooling of neutron star and the formation of neutron star [@cr]. $\bullet$ Extending our work to three-flavor QCD will be quite interesting. For example, flavor non-singlet nature of the GBE may induce mixing between $LL$ and $RR$ condensates which is not possible with only gluon exchange interactions and reduction in the value of the gap will affect meson masses in CFL phase [@genm] and the physics of kaon condensation in quark matter  [@sch], which will be reported elsewhere. [**Acknowledgements**]{} We would like to thank Mannque Rho and Deog Ki Hong for discussion. This work is supported partly by the BK21 project of the Ministry of Education, by KOSEF Grant 1999-2-111-005-5 and KRF Grant 1999-015-DI0023. [99]{} D. Bailin and A. Love, Phys. Rept. [**107**]{}, 325 (1984); M. Alford, K. Rajagopal and F. Wilczek, Phys. Lett. [**B422**]{}, 247 (1998); R. Rapp, T. Sch[ä]{}fer, E. V. Shuryak and M. Velkovsky, Phys. Rev. Lett. [**81**]{}, 53 (1998) M. Alford, K. Rajagopal and F. Wilczek, Nucl. Phys. [**B537**]{}, 443 (1999) T. Schaefer and F. Wilczek, Phys. Rev. [**D 60**]{}, 114033 (1999) D. T. Son, Phys. Rev. [**D59**]{}, 094019 (1999), hep-ph/9812287 R. Rapp, E. Shuryak and I. Zahed, “A chiral crystal in cold QCD matter at intermediates?,” hep-ph/0008207; M. Alford, J. Bowers and K. Rajagopal, “Crystalline Color Superconductivity,” hep-ph/0008208 K. Rajagopal and F. Wilczek, “The Condensed Matter Physics of QCD,” hep-ph/0011333; D. K. Hong, “Aspects of Color Superconductivity,” hep-ph/0101025 G.E. Brown and M. Rho, Phys. Rev. Lett. [**66**]{}, 2720 (1991) R. Rapp, G. Chanfray and J. Wambach, Nucl. Phys. [ **A617**]{}, 472 (1997) G.E. Brown, G.Q. Li, R. Rapp, M. Rho and J. Wambach, Acta Phys. Pol. [**B29**]{}, 2309 (1998) ; Y. Kim, R. Rapp, G.E. Brown and M. Rho, Phys. Rev. [**C 62**]{}, 015202 (2000) A. Manohar and H. Georgi, Nucl. Phys. [**B324**]{}, 189 (1984) M. Alford, J. Berges and K. Rajagopal, Nucl. Phys. [**558**]{}, 219 (1999), hep-ph/9903502 L. Ya. Glozman, Nucl. Phys. [**A663**]{} 103 (2000), hep-ph/9908423; L. Ya. Glozman and D. O. Riska, Phys. Rept. [**268**]{}, 263 (1996) G.E. Brown, M. Rho and C. Song, “ Strangeness Equilibration at GSI Energies,” nucl-th/001008 J.C. Collins and M.J. Perry, Phys. Rev. Lett, [**34**]{}, 1353 (1975); B.A. Freedman and L.D. McLerran, Phys. Rev. [**D16**]{} 1169 (1977); P.D. Morley and M.B. Kislinger, Phys. Rep. [**51**]{}, 63 (1979); Y. Kim and H. K. Lee, Phys. Rev. [**C62**]{}, 037901 (2000) T.M. Schwartz, S.P. Klevansky and G. Papp, Phys. Rev. [**C 60**]{}, 055205 (1999) T. Ericson and W. Weise, [*Pions and Nuclei*]{}, (Clarendon Press 1988) M. Alford, J. Berges and K. Rajagopal, Nucl. Phys. [ **B558**]{}, 219 (1999), hep-ph/9903502 J. Berges and K. Rajagopal, Nucl. Phys. [**B538**]{}, 215 (1999), hep-ph/9804233; G.W. Carter and D. Diakonov, Phys. Rev. [**D60**]{} 016004 (1999), hep-ph/9812445; T. M. Schwarz et al, Phys. Rev. [**C60**]{}, 055205 (1999), nucl-th/9903048; H. Mishra and J. C. Parikh, “Chiral symmetry breaking, color superconductivity and equation of state at high density: a variational approach,” hep-ph/0003019; B. Vanderheyden and A.D.Jackson, Phys. Rev. [**D62**]{}, 094010 (2000), hep-ph/0003150 M. Harada and K. Yamawaki, “Vector Manifestation of the Chiral Symmetry,” hep-ph/0010207 D. Blaschke, T. Klaehn and D. N. Voskresensky, “Diquark condensates and compact star cooling,” astro-ph/9908334 M. Alford, J. A. Bowers, K. Rajagopal, “Color superconductivity in compact stars,” hep-ph/0009357 G. W. Carter and S. Reddy, Phys. Rev. [**D62**]{}, 103002 (2000) D. K. Hong, M. Rho and I. Zahed, Phys. Lett. [**B468**]{}, 261 (1999), hep-ph/9906551; D. T. Son and M.A. Stephanov, Phys. Rev. [**D 61**]{} 074012 (2000), hep-ph/9910491; R. Casalbuoni and R. Gatto, “Pseudoscalar Masses in the Effective Theory for Color-Flavor Locking in high density QCD,”hep-ph/9911223; D. K. Hong, T. Lee and D.-P. Min, Phys. Lett. [**B477**]{}, 137 (2000), hep-ph/9912531; C. Manuel and M. H. G. Tytgat, Phys. Lett [**B 479**]{}, 190 (2000); M. Rho, E. Shuryak, A. Wirzba and I. Zahed, Nucl. Phys. [**A676**]{}, 273 (2000), hep-ph/0001104 ; S. R. Beane, P. F. Bedaque and M. J. Savage, Phys. Lett. [**B483**]{}, 131 (2000), hep-ph/0002209 T. Schäfer, “Kaon Condensation in High Density Quark Matter,”nucl-th/0007021 [^1]: Note that in the 2SC phase Cooper pairs of quarks cannot be color singlets and pick a color direction, the 3 direction for example. Therefore the color 3 quarks remain ungapped.

ArXiv

--- abstract: 'We carry out a quantum analysis of a dc SQUID mechanical displacement detector, comprising a SQUID with mechanically compliant loop segment, which is embedded in a microwave transmission line resonator. The SQUID is approximated as a nonlinear, current dependent inductance, inducing an external flux tunable, nonlinear Duffing self-interaction term in the microwave resonator mode equation. Motion of the compliant SQUID loop segment is transduced inductively through changes in the external flux threading SQUID loop, giving a ponderomotive, radiation pressure type coupling between the microwave and mechanical resonator modes. Expressions are derived for the detector signal response and noise, and it is found that a soft-spring Duffing self-interaction enables a closer approach to the displacement detection standard quantum limit, as well as cooling closer to the ground state.' author: - 'P. D. Nation' - 'M. P. Blencowe' - 'E. Buks' bibliography: - 'squid\_paperv10.bib' title: 'Quantum analysis of a nonlinear microwave cavity-embedded dc SQUID displacement detector' --- \[sec:introduction\]Introduction ================================ Recently there has been interest in exploiting the nonlinear dynamics of nanoelectromechanical systems (NEMS) for amplification.[@Krommer:2000p2946; @Aldridge:2005p3614; @Almog:2007p2875] The use of nonlinear mechanical resonators to some extent parallels investigations with systems comprising purely electronic degrees of freedom, such as nonlinear superconducting devices incorporating Josephson Junctions (JJ).[@Siddiqi:2004p1662; @Lupascu:2006p3289; @Lee:2007p3227] For example, it was shown that the bistable response of an RF-driven JJ can be employed as a low noise, high-sensitivity amplifier for superconducting qubits.[@Siddiqi:2004p1662] A similar setup consisting of a JJ embedded in a microwave cavity was used to measure the states of a quantronium qubit,[@Metcalfe:2007p1060] where the relevant cavity mode was found to obey the Duffing oscillator equation.[@Boaknin:2007p3308] One area of nanomechanics that has yet to fully explore the possibility of exploiting nonlinearities for sensitive detection involves setups in which a nanomechanical resonator couples either capacitively[@Xue:2007p2711; @Regal:2008p2560; @Teufel:2008p2398] or inductively[@Blencowe:2007p285; @Buks:2007p279] to a superconducting microwave transmission line resonator, combining elements from both the above-described NEMS and superconducting systems. Such setups are in some sense the solid-state analogues of optomechanical systems, which ponderomotively couple a movable mirror to the optical field inside a cavity using radiation pressure.[@Mancini:1994p2815; @LAW:1995p3039; @Jacobs:1999p2816; @Metzger:2004p2871; @Gigan:2006p1942; @Arcizet:2006p2873; @Schliesser:2006p2558; @Mavalvala:2007p160801; @Thompson:2008p72] In both areas, the focus has primarily been on operating in the regime where the cavity and resonator behave to a good approximation as harmonic oscillators interacting via their mutual ponderomotive coupling. However, in the case of microwave cavities, introducing an embedded JJ,[@Boaknin:2007p3308] or simply driving the cavity close to the superconducting critical temperature,[@Tholen:2007p1916] results in the breakdown of the harmonic mode approximation; nonlinear dynamical behavior of the cavity must be taken into account. Furthermore, the ponderomotive coupling term between the microwave or optical cavity mode and mechanical mode is by itself capable of inducing strong, effective nonlinearities in the respective mode equations. In optical systems, such nonlinearities can manifest themselves in the appearance of a bistable (or even multistable) region for the movable mirror.[@DORSEL:1983p2818; @Marquardt:2006p103901] By restricting ourselves to linear microwave cavities, we are overlooking a range of nonlinear phenomena that might enable a closer approach to quantum-limited detection, as well as cooling of the mechanical oscillator closer to its ground state. As an illustration, consider the phase sensitive Josephson parametric amplifier,[@yurke:1989p2519; @Yurke:2006p1911; @bergeal:2008; @castellanos:2008] which exploits the nonlinear effective inductance of the JJ to perform (in principle) noiseless amplification and quantum squeezing of the respective complimentary quadrature amplitudes of the signal oscillator. In this paper, we will go beyond the usually considered ponderomotively-coupled two oscillator system to include a Duffing nonlinearity in the microwave cavity mode equations. The closed system model Hamiltonian describing the nonlinear microwave-coupled mechanical oscillators is given by Eq. (\[eq:hamiltonian-closed\]). The nonlinear microwave mode is externally driven with a pump frequency $\omega_p$ that can be detuned from the transmission line mode frequency $\omega_T$. Our investigation will focus on the nonlinear amplifier created by embedding a dc-SQUID displacement detector into a superconducting microwave transmission line.[@Blencowe:2007p285] This has the advantage of significantly larger coupling strengths[@Devoret:2007p1782] as compared with existing geometrical coupling schemes.[@Xue:2007p2711; @Regal:2008p2560; @Teufel:2008p2398] The displacement detector comprises a SQUID with one segment consisting of a doubly-clamped mechanical resonator as shown in Fig. \[fig:diagram\]. The net flux, and therefore circulating current, is modulated by the mechanical motion, providing displacement transduction. The capacitively-coupled pump/probe feedline both drives and provides readout of the relevant transmission line resonator mode amplitude (or phase). We will assume transmission line losses are predominantly due to coupling with the feedline, and that the pump drive is coherent. The main irreducible noise source is therefore microwave photon shot noise from the drive that acts back on the mechanical oscillator via the intermediate nonlinear microwave resonator and SQUID. Environmental influences on the mechanical oscillator other than that due to the SQUID detector are simply modelled as a free oscillator thermal bath. By operating the amplifier well below the superconducting critical temperature, and with transmission line currents less than the SQUID JJ’s critical current threshold, resistive tunneling of electrons and the associated noise is a negligible contribution. Similar setups involving JJ elements have been considered previously.[@Blencowe:2007p285; @Buks:2007p279; @Xue:2007p2454; @Wang:2007p2819] With JJ plasma frequencies assumed to be larger than both the mechanical and transmission line fundamental mode frequencies, the SQUID can be considered as a passive, effective inductance element that depends on both the externally applied flux and drive current. The effective inductance can therefore be freely tuned by varying these external parameters. Previously, we considered only the lowest, zeroth order expansion of the inductance with respect to the current entering (or exiting) the SQUID,[@Blencowe:2007p285] yielding the usual ponderomotively-coupled double harmonic oscillator system. In this companion paper, we include the next leading, quadratic order term, resulting in a nonlinear current dependent inductance. Provided that the current magnitude is small as compared with the JJ’s critical current, neglecting higher order terms should not introduce significant errors. The nonlinear inductance induces an effective Duffing (i.e., cubic) self-interaction term in the microwave mode equations of motion. The results presented here apply to a broad class of bosonic detector, which includes optomechanical amplifiers with nonlinear cavities[@Drummond:1980p3191] that are describable by Hamiltonian (\[eq:hamiltonian-closed\]). A related analysis of quantum noise in a Duffing oscillator amplifier is given in Ref. . The paper is organized as follows. In Sec. \[sec:motion\] we first derive the truncated Hamiltonian (\[eq:hamiltonian-closed\]) that describes the closed system dynamics of the coupled cavity and mechanical resonator fundamental modes. We then derive the quantum Langevin equations of motion that describe the open system dynamics in the presence of the pump/probe line and mechanical oscillator’s external environment. In Sec. \[sec:response\] we find expressions for the detector signal response and noise using a semiclassical treatment of the detector’s linear response to the external noise input signal driving the mechanical oscillator. Section \[sec:bistability\] determines the critical drive current for the onset of bistability (not to be confused with the JJ critical current). Sections \[sec:detection\] and \[sec:cooling\] discuss the effects of the microwave mode Duffing nonlinearity on mechanical mode displacement detection and cooling, respectively, giving illustrative examples assuming achievable device parameters. Section \[sec:conclusion\] briefly concludes, while the more technical aspects of the signal and noise term derivations are relegated to the appendix. \[sec:motion\]Equations of Motion ================================= \[subsec:closed\]Closed System Hamiltonian ------------------------------------------ The displacement detector scheme is shown in Fig. \[fig:diagram\]. The device consists of a stripline resonator (transmission line) of length $l$ bisected by a SQUID. The transmission line is characterized by an inductance and capacitance per unit length $L_{T}$ and $C_{T}$ respectively. The SQUID comprises two identical Josephson junctions with critical current $I_{C}$ and capacitance $C_{J}$. A segment of the SQUID loop is mechanically compliant, forming a doubly clamped resonator of length $l_{\mathrm{osc}}$. We only take into account mechanical fundamental mode displacements in the plane of the loop and assume that the resonator can be modeled effectively as a harmonic oscillator with the $y$ coordinate giving the centre of mass displacement. The magnetic flux threading the loop is given by $\Phi_{\mathrm{ext}}(y)=\Phi_{\mathrm{ext}}(0)+\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}y$, where $\Phi_{\mathrm{ext}}(0)\equiv\Phi_{\mathrm{ext}}$ is the flux with the mechanical oscillator fixed at $y=0$, $B_{\mathrm{ext}}$ is the externally applied field in the vicinity of the resonator, and $\lambda$ is a geometrical factor that corrects for the non-uniform displacement of the oscillator along its length. The coupling between mechanical oscillator and external heat bath is characterized by the oscillator amplitude damping rate $\gamma_{bm}$, while the pump-probe line-transmission line coupling is characterized by the transmission line amplitude damping rate $\gamma_{pT}$. In what follows, we will assume weak couplings (i.e., large quality factors for the transmission line and mechanical oscillator) and also that the dominant dissipation mechanism for the transmission line is due to its coupling to the pump-probe line, $\gamma_{pT}$. ![Layout of the dc SQUID displacement detector. The dimensions of the SQUID have been enlarged relative to the transmission line to show the key characteristics employed in the analysis.[]{data-label="fig:diagram"}](circuitdiagram){width="5.0in"} In analyzing the SQUID dynamics, an appropriate choice of variables is $\gamma_{\pm}=(\phi_{1}\pm\phi_{2})/2$, where $\phi_{1}$ and $\phi_{2}$ are the gauge invariant phases across the Josephson junctions,[@orlando] while for the transmission line we choose the phase field $\phi(x,t)$. The transmission line current and voltage are described in terms of $\phi(x,t)$ using the standard telegraphic relations: $$\begin{aligned} I(x,t)&=&-\frac{\Phi_{0}}{2\pi L_{T}}\frac{\partial \phi(x,t)}{\partial x}, \\ V(x,t)&=&\frac{\Phi_{0}}{2\pi}\frac{\partial \phi(x,t)}{\partial t}, \label{eq:currentvoltage}\end{aligned}$$ where $\Phi_{0}=h/(2e)$ is the flux quantum. Assuming that the SQUID can be lumped at the midpoint $x=0$ of the transmission line, the equations of motion for the closed system comprising the SQUID, transmission line and mechanical oscillator are given by[@Buks:2006p346] $$\label{eq:wave} \frac{\partial^{2}\phi}{\partial t^{2}}=\frac{1}{L_{T}C_{T}}\frac{\partial^{2}\phi}{\partial x^{2}},$$ $$\label{eq:current-circ} {\omega_{J}^{-2}}\ddot{\gamma}_{-}+\cos(\gamma_{+})\sin(\gamma_{-})+{2}{\beta^{-1}_{L}}\left[\gamma_{-}-\pi\left(n+\frac{\Phi_{\mathrm{ext}}+\lambda B_{\mathrm{ext}}l_{\mathrm{osc}} y}{\Phi_{0}}\right)\right]=0,$$ $$\label{eq:current-avg} {\omega_{J}^{-2}}\ddot{\gamma}_{+}+\sin(\gamma_{+})\cos(\gamma_{-})+\frac{\Phi_{0}}{4\pi L_{T}I_{c}}\frac{\partial \phi(0,t)}{\partial x}=0,$$ and $$\label{eq:force} m\ddot{y}+m\omega_{m}^{2}y-\frac{\Phi_{0}}{\pi L}\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}\gamma_{-}=0,$$ where $\omega_{J}=\sqrt{2\pi I_{c}/(C_{J}\Phi_{0})}$ is the plasma frequency of the Josephson junctions, $\beta_{L}\equiv 2\pi L I_{c}/\Phi_{0}$ is a dimensionless parameter with $L$ the SQUID loop self-inductance and $I_c$ the Josephson junction critical current, and where $n$ takes on integer values arising from the requirement that the phase around the loop be single-valued. Equation (\[eq:wave\]) is the wave equation for the transmission line, equation (\[eq:current-circ\]) describes the current circling the loop, which depends on the external flux and oscillator position, equation (\[eq:current-avg\]) describes the average current in the loop, and Eq. (\[eq:force\]) is Newton’s second law for the mechanical oscillator with Lorentz force acting on the oscillator. The current and voltage across the SQUID must also obey the boundary conditions $$\begin{aligned} \frac{\partial \phi(\pm l/2,t)}{\partial x}=0;\ \ \frac{\partial \phi(0^{-},t)}{\partial x}=\frac{\partial \phi(0^{+},t)}{\partial x},\label{eq:currentbc}\\ \dot{\gamma}_{+}-\frac{L}{4 L_{T}}\frac{\partial^2 \phi(0,t)}{\partial t\partial x}=\frac{\partial \phi(0^{-},t)}{\partial t}-\frac{\partial \phi(0^{+},t)}{\partial t}.\label{eq:voltagebc}\end{aligned}$$ Using Eqs. (\[eq:wave\])-(\[eq:voltagebc\]), we shall now derive approximate equations of motion describing a single mode of the transmission line interacting with the mechanical oscillator, where the form of the interaction between the two oscillators is governed by the SQUID parameters and boundary conditions. Assume that the following conditions are satisfied: (a) $\omega_{J}\gg\omega_{T}\gg\omega_{m}$. (b) $\beta_{L}\ll 1$. (c) $|I(0,t)/I_{c}|\ll1$. (d) $|\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}y/{\Phi_{0}}|\ll 1$. Condition (a) states that the SQUID plasma frequency is much larger than the transmission line mode frequency of interest, $\omega_T$, and consequently we shall ignore the SQUID inertia terms in (\[eq:current-circ\]) and (\[eq:current-avg\]). Condition (b) allows us to neglect the SQUID loop self inductance and, together with (a), eliminate $\gamma_{\pm}$ from the equations by expressing them in terms of the transmission line and oscillator coordinates as series expansions in $\beta_{L}$. Conditions (c) and (d) allow us to expand the above equations in the transmission line current $I(0,t)\equiv I(t)$ at $x=0$ and in the oscillator displacement $y$. Keeping terms to first order in $y$ and to leading, second order in $I$, Eq. (\[eq:force\]) for the mechanical oscillator becomes approximately $$m\ddot{y}+m\omega_{m}^{2}y-L_{01}I^2/2 =0. \label{eq:force-approx}$$ The voltage boundary condition (\[eq:voltagebc\]) can be expressed as $$\frac{\partial}{\partial t}[L(I,y)I]=\frac{\Phi_{0}}{2\pi}\left[\frac{\partial \phi(0^{-},t)}{\partial t}-\frac{\partial \phi(0^{+},t)}{\partial t}\right]\label{eq:voltage},$$ where $L(I,y)$ is the effective inductance, which expanded to second order in $I$ takes the form $$\label{eq:inductance} L(I,y)=L_{00}+L_{20}\left({I}/{I_{c}}\right)^{2}+L_{01}y,$$ where the $L_{ij}$ coefficients are defined as $$\begin{aligned} L_{00}&=&\frac{\Phi_{0}}{4\pi I_{c}}\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\label{eq:L00}\\ L_{20}&=&\frac{\Phi_{0}}{96\pi I_{c}}\sec^{3}\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\label{eq:L20}\\ L_{01}&=&\frac{\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}}{4I_{c}}\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\tan\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right).\label{eq:L01}\end{aligned}$$ Note that we have neglected the $I^2 y$ term in (\[eq:inductance\]), restricting ourselves to the leading order coupling only between the transmission line and mechanical oscillator, as already stated. The above equations differ from those of the prequel [@Blencowe:2007p285] through the inclusion of the nonlinear, leading order current-dependent contribution \[$L_{20}(I/I_c)^2$\] to the effective inductance $L(I,y)$. The nonlinear voltage boundary condition (\[eq:voltage\]) with inductance given by Eq. (\[eq:inductance\]) generates frequency tripling harmonics of the transmission line resonator mode. Omitting for the time being the mechanical oscillator degree of freedom, a trial perturbative mode solution to the wave equation (\[eq:wave\]) that includes the leading harmonic and solves the current boundary conditions (\[eq:currentbc\]) is the following: $$\label{eq:phase-field} \phi(x,t)=\left\{\begin{array}{ccc} +A\cos(\omega t +\varphi)\cos\left[k(x-l/2)\right]+a A^3 \cos(3\omega t +3\varphi)\cos\left[3k(x-l/2)\right]; x>0\\ -A\cos(\omega t +\varphi)\cos\left[k(x+l/2)\right]-a A^3 \cos(3\omega t +3\varphi)\cos\left[3k(x+l/2)\right]; x<0, \end{array}\right.$$ where $k=k^{(0)} + k^{(1)}$ and $\omega=|k|/\sqrt{L_T C_T}$. The coefficients $a$, $k^{(0)}$ and $k^{(1)}$ are determined by substituting Eq. (\[eq:phase-field\]) into the voltage boundary condition (\[eq:voltage\]) and solving perturbatively to order $A^3$, with $k^{(1)}$ scaling as $A^2$. We obtain: $a=-1/48$, $$({k^{(0)}l}/{2})\tan\left({k^{(0)}l}/{2}\right)=\zeta^{-1} \label{eq:k0}$$ and $$k^{(1)} l=-\frac{1}{8}\zeta^3 A^2 (k^{(0)}l)^3 \sin^2 \left({k^{(0)}l}/{2}\right), \label{eq:k1}$$ where $$\zeta=\frac{L_{00}}{L_Tl}=\frac{\Phi_0}{4\pi L_T l I_c} \sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_0}\right). \label{eq:zeta}$$ Considering the transmission line phase field at the location $x=-l/2$, where the field is pumped and probed (see Fig. \[fig:diagram\]), the perturbative solution (\[eq:phase-field\]) can be obtained from the following single mode equation for $\phi(-l/2,t)\equiv\phi(t)$: $$\begin{aligned} \frac{d^2\phi}{dt^2}+\omega_T^2 \phi &+&\frac{1}{12}\omega_T^2 \left(1-18\zeta^3\left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2\right)\phi^3\cr &&+\frac{1}{12}\left(1-2\zeta^3\left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2\right)\frac{d^2(\phi^3)}{dt^2}=0, \label{eq:nonlinearmode}\end{aligned}$$ where $\omega_T=|k^{(0)}|/\sqrt{L_TC_T}$. The awkward nonlinear term $\ddot{\phi^3}$ can be eliminated by redefining the phase mode coordinate as $\phi=\psi (1+\Gamma\psi^2)$, provided $|\Gamma|\phi^2\ll 1$, where $$\Gamma=-\frac{1}{12}\left(1-2\zeta^3\left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2\right). \label{eq:Gamma}$$ The mode equation (\[eq:nonlinearmode\]) in terms of the redefined phase coordinate $\psi$ then becomes $$\ddot{\psi}+\omega^2_T \psi -\frac{4}{3}\omega_T^2\zeta^3 \left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2 \psi^3 =0. \label{eq:psimode}$$ Thus, embedding a SQUID in a microwave transmission line induces a cubic nonlinearity in the effective single mode equations (under the conditions of small currents as compared with the Josephson junction critical current), resulting in the familiar (undamped) Duffing oscillator. We now restore the mechanical degree of freedom $y(t)$ by assuming that for small and slow displacements \[conditions (a) and (d) above\], the interaction with $\psi$ can be obtained by expanding $\omega_T$ \[through its dependence on $\Phi_{\mathrm{ext}}(y)$\] to first order in $y$ in Eq. (\[eq:psimode\]) to obtain $$\begin{aligned} \ddot{\psi}+\omega^2_T \psi &-&\frac{4}{3}\omega_T^2\zeta^3 \left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2 \psi^3\cr &&=\frac{\lambda B_{\mathrm{ext}} l_{\mathrm{osc}} y}{\left(\Phi_0/\pi\right)}\frac{\Phi_0}{4\pi L_Tl I_c}\tan\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_0}\right)\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_0}\right)\omega^2_T\psi. \label{eq:mass-psimode}\end{aligned}$$ Equation (\[eq:force-approx\]) for the mechanical oscillator, together with Eq. (\[eq:mass-psimode\]) for the phase coordinate, follow from the Lagrangian: $$\begin{aligned} &&\mathcal{L}(\psi,y,\dot{\psi},\dot{y})=\frac{1}{2}m\dot{y}^{2}-\frac{1}{2}m\omega_{m}^{2}y^{2}+\frac{1}{2}C_{T}l\left(\frac{\Phi_{0}}{2\pi}\right)^{2}\sin^{2}(k_{0}l/2)\cr &&\times\left\{\frac{1}{2}\dot{\psi}^{2}(t)-\frac{1}{2}\left[1-\frac{\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}y}{(\Phi_{0}/\pi)}\frac{\Phi_{0}}{4\pi L_{T}lI_{c}}\tan\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\right]\omega_{T}^{2}\psi^{2}(t)\right.\cr &&\left.+\frac{1}{3}\omega^2_T\zeta^{3}\left[\left({k^{(0)}l}/{2}\right)\sin\left({k^{(0)}l}/{2}\right)\right]^2\psi^{4}\right\}.\end{aligned}$$ Introduce the phase momentum coordinate $p_{\psi}=\partial{\mathcal{L}}/\partial{\dot{\psi}}=m_{\psi}\dot{\psi}$ and raising (lowering) operators $$\begin{aligned} \hat{a}_{T}^{\pm}&=&\frac{1}{\sqrt{2m_{\psi}\hbar\omega_{T}}}\left(m_{\psi}\omega_{T}\hat{\psi}\mp i\hat{p}_{\psi}\right)\label{eq:aT}\\ \hat{a}_{m}^{\pm}&=&\frac{1}{\sqrt{2m\hbar\omega_{m}}}\left(m\omega_{m}\hat{y}\mp i\hat{p}_{y}\right)\label{eq:am}\end{aligned}$$ satisfying the usual commutation relations, where the effective phase mass is $m_{\psi}=\frac{1}{2}C_{T}l\left(\Phi_{0}/2\pi\right)^{2}\sin^{2}(k_{0}l/2)$. In terms of the raising (lowering) operators, the Hamiltonian operator is $$\label{eq:hamiltonian-closed} H=\hbar\omega_{T}{a}_{T}^{+}{a}_{T}+\frac{1}{12}\hbar\omega_{T}K_{d}({a}_{T}^{+}+{a}_{T})^{4}+\hbar\omega_{m}{a}_{m}^{+}{a}_{m} +\frac{1}{2}\hbar\omega_{T}K_{Tm}({a}_{T}^{+}+{a}_{T})^2({a}_{m}^{+}+{a}_{m}),$$ where, for notational convenience, hats on the operators and the minus superscript on the lowering operators will be suppressed from now on. The parameter characterizing the strength of the interaction between the transmission line mode and mechanical oscillator mode is $$K_{Tm}=\frac{\lambda B_{\mathrm{ext}}l_{\mathrm{osc}}\Delta_{zp}}{\left(\Phi_{0}/\pi\right)}\frac{\Phi_{0}}{4\pi L_{T}lI_{c}}\tan\left(\pi\Phi_{\mathrm{ext}}/\Phi_{0}\right)\sec\left(\pi\Phi_{\mathrm{ext}}/\Phi_{0}\right), \label{eq:ktm}$$ where $\Delta_{zp}=\sqrt{\hbar/(2m\omega_m)}$ is the zero-point displacement uncertainty. The parameter $K_d$ characterizing the strength of the Duffing nonlinear term takes the form $$K_d = -\left(k^{(0)} l\right)^2 \left(\frac{L_{00}}{L_Tl}\right)^3 \left[\frac{(2e)^2/(2C_Tl)}{\hbar\omega_T}\right], \label{eq:Kd}$$ which has been written in such a way as to make clear its various dependencies. In particular, $K_d$ depends essentially on the cube of the ratio of the linear SQUID effective inductance $L_{00}$ to transmission line inductance $L_Tl$, as well as on the ratio of the single Cooper pair charging energy to the microwave mode photon energy of the transmission line. Since the strength and sign of the linear SQUID inductance depends on the external flux $\Phi_{\mathrm{ext}}$ \[see Eq. (\[eq:L00\])\], it is possible to vary the strength as well as the sign of the Duffing constant by tuning the external flux either side of $\Phi_0/2$. Thus, we can have either spring hardening or spring softening of the transmission line oscillator mode. Previously this flux tunability was observed in the readout of a persistent current qubit.[@Lee:2007p3227] Note, however, that the perturbative approximations that go into deriving the above Hamiltonian (\[eq:hamiltonian-closed\]) do not allow too close an approach to the singular half-integer flux quantum point. In particular, the validity of the expansions in $I_T$ and $\beta_L$ properly require the following conditions to hold: $$\begin{aligned} \left|\frac{I}{I_{c}}\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\right|&\ll&1\label{eq:Iccondition}\\ \left|\beta_{L}\sec\left({\pi\Phi_{\mathrm{ext}}}/{\Phi_{0}}\right)\right|&\ll&1\label{eq:betalcondition}.\end{aligned}$$ As already noted, Eq. (\[eq:hamiltonian-closed\]) without the Duffing nonlinearity coincides with the Hamiltonian commonly used to describe the single mode of an optical cavity interacting with a mechanical mirror via the radiation pressure. However, we have just seen that embedding a SQUID within a microwave transmission line cavity induces a tunable Duffing self-interaction term as well; it is not so easy to achieve a similar, tunable nonlinearity in the optical cavity counterpart. \[sec:open\]Open System Dynamics -------------------------------- Up until now we have considered the transmission line, SQUID and mechanical oscillator as an isolated system. It is straightforward to couple the transmission line to an external pump-probe feedline and mechanical oscillator to a thermal bath using the ‘in-out’ formalism of Gardiner and Collett.[@Gardiner:1985p1483] Assuming weak system-bath couplings justify making the rotating wave approximation (RWA), and furthermore making a Markov approximation for the bath dynamics, the following Langevin equations can be derived for the system mode operators in the Heisenberg picture: $$\begin{aligned} \label{eq:motionm} \frac{d{a}_{m}}{dt}&=&-i\omega_{m}{a}_{m}+\frac{i}{\hbar}\sqrt{\frac{\hbar}{2m\omega_{m}}}F_{\mathrm{ext}}(t)-i\omega_{T}K_{Tm}{a}^{+}_{T}{a}_{T}\cr &&-\gamma_{bm}{a}_{m}-i\sqrt{2\gamma_{bm}}e^{i\phi_{bm}}{a}_{b}^{\mathrm{in}}(t)\end{aligned}$$ and $$\begin{aligned} \label{eq:motiont} \frac{d{a}_{T}}{dt}&=&-i\omega_{T}{a}_{T}-i\omega_{T}K_{d}{a}_{T}^{+}{a}_{T}{a}_{T}-i\omega_{T}K_{Tm}{a}_{T}({a}^{+}_{m}+{a}_{m})\cr &&-\gamma_{pT}{a}_{T}-i\sqrt{2\gamma_{pT}}e^{i\phi_{pT}}{a}_{p}^{\mathrm{in}}(t),\end{aligned}$$ where $\gamma_{bm}$ is the mechanical oscillator amplitude damping rate due to coupling to the bath, $\gamma_{pT}$ is the transmission line mode damping rate due to coupling to the pump-probe line, and we have also assumed that the small Duffing coupling $K_d$ and transmission line-mechanical oscillator coupling $K_{Tm}$ justify applying the RWA to the transmission line mode operator terms. The ‘in’ bath and probe line operators are defined as $$\label{eq:inoperators} {a}_{i}^{\mathrm{in}}(t)=\frac{1}{\sqrt{2\pi}}\int d\omega e^{-i\omega(t-t_{0})}{a}_{i}(\omega,t_{0}),$$ where $t>t_{0}$, with the states of the pump-probe line and oscillator bath assigned at $t_0$, interpreted as the initial time in the past before the measurement commences. For completeness, we have also included a classical, external time-dependent force $F_{\mathrm{ext}}$(t) acting on the mechanical oscillator, although we shall not address the force detection sensitivity in the present work. It will be convenient to work with the Fourier transformed Langevin equations. With $O(\omega)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}d\omega e^{i\omega t} O(t)$, Eqs. (\[eq:motionm\]) and (\[eq:motiont\]) become $$\begin{aligned} \label{eq:amw} {a}_{m}(\omega)&=&\frac{1}{\omega-\omega_{m}+i\gamma_{bm}}\left\{\sqrt{2\gamma_{bm}}e^{i\phi_{bm}}{a}_{b}^{\mathrm{in}}(\omega)-\frac{1}{\sqrt{2m\hbar\omega_{m}}}F_{\mathrm{ext}}(\omega)\right.\cr &&\left.+\frac{\omega_{T}K_{Tm}}{2\sqrt{2\pi}}\int_{-\infty}^{\infty}d\omega' \left[{a}_{T}(\omega'){a}_{T}^{+}(\omega'-\omega)+{a}_{T}^{+}(\omega'){a}_{T}(\omega'+\omega)\right]\right\} \end{aligned}$$ and $$\begin{aligned} \label{eq:atw} {a}_{T}(\omega)&=&\frac{1}{\omega-\omega_{T}+i\gamma_{pT}}\left\{\frac{}{}\sqrt{2\gamma_{pT}}e^{i\phi_{pT}}{a}_{p}^{\mathrm{in}}(\omega)+\frac{\omega_{T}K_{d}}{2\pi}\right.\cr &&\left.\times \int_{-\infty}^{\infty}d\omega'\int_{-\infty}^{\infty}d\omega''{a}_{T}^{+}(\omega''){a}_{T}(\omega'){a}_{T}(\omega+\omega''-\omega')\right.\cr &&\left.+\frac{\omega_{T}K_{Tm}}{\sqrt{2\pi}}\int_{-\infty}^{\infty}d\omega'{a}_{T}(\omega') \left[{a}_{m}(\omega-\omega')+{a}_{m}^{+}(\omega'-\omega)\right]\right\}.\end{aligned}$$ \[sec:response\]Detector Response ================================= The probe line observables are expressed in terms of the ‘out’ mode operator: $$\label{eq:outoperator} {a}_{p}^{\mathrm{out}}(t)=\frac{1}{\sqrt{2\pi}}\int d\omega e^{-i\omega(t-t_1)}{a}_{p}(\omega,t_1),$$ where $t_1>t$. The ‘out’ and ‘in’ probe operators are related via the following useful identity:[@Gardiner:1985p1483] $$\label{eq:inoutidentity} {a}_{p}^{\mathrm{out}}(t)=-i\sqrt{2\gamma_{pT}}e^{-i\phi_{pT}}{a}_{T}(t)+{a}_{p}^{\mathrm{in}}(t),$$ which allows us to obtain the expectation value of a given observable once $a_T(t)$ is determined. As illustrative expectation value, we shall consider the variance in the probe line reflected current in a given bandwidth $\delta\omega$ centered about the signal frequency of interest $\omega_s$:[@Blencowe:2007p285] $$\begin{aligned} \overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s},\delta\omega)\right]^{2}\rangle}&=&\frac{1}{Z_{p}}\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega_{1}d\omega_{2}}{2\pi}\hbar\omega_{1}\left(\frac{2\sin\left[(\omega_{1}-\omega_{2})T_{M}/2\right]}{(\omega_{1}-\omega_{2})T_{M}}\right)\cr &&\times\frac{1}{2}\langle{a}_{p}^{\mathrm{out}}(\omega_{1}){a}_{p}^{{\mathrm{out}}+}(\omega_{2})+{a}_{p}^{{\mathrm{out}}+}(\omega_{2}){a}_{p}^{\mathrm{out}}(\omega_{1})\rangle, \label{eq:currentvariance}\end{aligned}$$ where, in addition to the ensemble average, there is also a time average denoted by the overbar, with the averaging time taken to be the duration of the measurement $T_M$, assumed much longer than all other timescales associated with the detector dynamics. In particular, time averaging is required when $F_{\mathrm{ext}}(t)$ has a deterministic time dependence.[@Blencowe:2007p285] Expectation values of other observables, such as the reflected voltage variance and reflected power are simply obtained from Eq. (\[eq:currentvariance\]) with appropriate inclusions of the probe line impedance $Z_p=\sqrt{L_p/C_p}$: $P^{\mathrm{out}}=\overline{\langle [\delta V^{\mathrm{out}}]^2\rangle}/Z_p=\overline{\langle [\delta I^{\mathrm{out}}]^2\rangle} Z_p$. From the form of the $K_{Tm}$ coupling term in Eq. (\[eq:motiont\]), we can see that the motion of the mechanical resonator modulates the transmission line frequency, and thus a complimentary way to transduce displacements besides measuring the current amplitude, is to measure the frequency-dependent, relative phase shift between the ‘in’ pump drive current and ‘out’ probe current using the homodyne detection procedure.[@gardiner] While we shall focus on amplitude detection, the homodyne method can be straighforwardly addressed and is expected to give similar results for the quantum limited detection sensitivity. Substituting Eq. (\[eq:amw\]) into (\[eq:atw\]), we obtained the following single equation for the transmission line mode operator $a_T$: $$\begin{aligned} \label{eq:atwonly} {a}_{T}(\omega)&=&\int_{-\infty}^{\infty}d\omega'{a}_{T}(\omega-\omega')A(\omega,\omega')+\int_{-\infty}^{\infty}d\omega'B(\omega,\omega'){a}_{T}(\omega-\omega')\cr &&\times\int_{-\infty}^{\infty}d\omega''\left[{a}_{T}(\omega''){a}_{T}^{+}(\omega''-\omega')+{a}_{T}^{+}(\omega''){a}_{T}(\omega''+\omega')\right]\cr &&+D(\omega)\int_{-\infty}^{\infty}d\omega''\int_{-\infty}^{\infty}d\omega'{a}^{+}_{T}(\omega''){a}_{T}(\omega'){a}_{T}(\omega+\omega''-\omega')+C(\omega),\end{aligned}$$ where $$\label{eq:A} A(\omega,\omega')=\frac{\omega_T K_{Tm}}{\sqrt{2\pi}}\frac{1}{\omega-\omega_{T}+i\gamma_{pT}} \left[\frac{S_{m}(\omega')}{\omega'-\omega_{m}+i\gamma_{bm}}+\frac{S_{m}^{+}(-\omega')}{-\omega'-\omega_{m}-i\gamma_{bm}}\right],$$ $$\label{eq:B} B(\omega,\omega')=\frac{(\omega_T K_{Tm})^{2}}{4\pi}\frac{1}{\omega-\omega_{T}+i\gamma_{pT}} \left[\frac{1}{\omega'-\omega_{m}+i\gamma_{bm}}+\frac{1}{-\omega'-\omega_{m}-i\gamma_{bm}}\right],$$ $$\label{eq:C} C(\omega)=\frac{S_{T}(\omega)}{\omega-\omega_{T}+i\gamma_{pT}}$$ and $$\label{eq:D} D(\omega)=\frac{\omega_T K_d}{2\pi}\frac{1}{\omega-\omega_{T}+i\gamma_{pT}},$$ with mechanical signal operator $$S_{m}(\omega)=\sqrt{2\gamma_{bm}}e^{i\phi_{bm}}{a}_{b}^{\mathrm{in}}(\omega)-\frac{1}{\sqrt{2m\hbar\omega_{m}}}F_{\mathrm{ext}}(\omega)\label{eq:sm}$$ and noise operator $$S_{T}(\omega)=\sqrt{2\gamma_{pT}}e^{i\phi_{pT}}{a}_{p}^{\mathrm{in}}(\omega).\label{eq:st}$$ For small signal strength, it is assumed that Eq. (\[eq:atwonly\]) can be solved as a series expansion up to first order in $A(\omega,\omega')$, giving the usual linear-response approximation. I.e., $a_T(\omega)\approx a_T^{(0)}(\omega)+a_T^{(1)}(\omega)$, where the noise component $a_T^{(0)}(\omega)$ is the solution to Eq. (\[eq:atwonly\]) with the mechanical signal source term $A(\omega,\omega')$ set to zero, while the signal component $a_T^{(1)}(\omega)$ is the part of the solution to Eq. (\[eq:atwonly\]) that depends linearly on $A(\omega,\omega')$. Thus, from Eq. (\[eq:inoutidentity\]) we can express the ‘out’ probe mode operator as follows: $$a_p^{\mathrm{out}}(\omega)=\left[-i\sqrt{2\gamma_{pT}}e^{-i\phi_{pT}} a^{(1)}_T (\omega)\right]+\left[-i\sqrt{2\gamma_{pT}}e^{-i\phi_{pT}} a^{(0)}_T (\omega)+a_p^{\mathrm{in}}(\omega)\right], \label{eq:asignalnoise}$$ where the first square bracketed term gives the signal contribution to the detector response and the second square bracketed term gives the noise contribution. As ‘in’ states, we consider the mechanical oscillator bath to be in a thermal state at temperature $T$ and the pump line to be in a coherent state centered about the pump frequency $\omega_p$:[@Johansson:2006p2429] $$\label{eq:coherent} \left|\{\alpha(\omega)\}\right.\rangle_{p}=\exp\left(\int d\omega\alpha(\omega)\left[{a}_{p}^{\mathrm{in}+}(\omega)-{a}_{p}^{\mathrm{in}}(\omega)\right]\right)\left|0\right.\rangle_{p},$$ where $\left|0\right.\rangle_{p}$ is the vacuum state and $$\label{eq:alpha} \alpha(\omega)=-I_{0}\sqrt{\frac{Z_{p}T^{2}_{M}}{2\hbar}}\frac{e^{-(\omega-\omega_{p})^{2}T^{2}_{M}/2}}{\sqrt{\omega}}.$$ The coherent state coordinate $\alpha(\omega)$ is parametrized such that the expectation value of the right-propagating ‘in’ current $I^{\mathrm{in}}(x,t)$ with respect to this coherent state has amplitude $I_0$, where $$I^{\mathrm{in}}(x,t)=i\sqrt{\frac{\hbar}{4\pi Z_p}}\int_0^{\infty} d\omega \sqrt{\omega} \left[e^{i\omega (x/v_p -t)} a_p^{\mathrm{in}}(\omega)-e^{-i\omega (x/v_p -t)} a_p^{\mathrm{in}+}(\omega)\right], \label{eq:incurrentoperator}$$ with $v_p=1/\sqrt{L_pC_p}$ the wave propagation velocity in the pump probe line. With the pump probe line in a coherent state, we assume that for large drive currents Eq. (\[eq:atwonly\]) can be approximately solved using a semiclassical, ‘mean field’ approximation, where the quantum fluctuation $\delta a_T^{(0)}(\omega)$ in $a_T^{(0)}(\omega)=\langle a_T^{(0)}(\omega)\rangle +\delta a_T^{(0)}(\omega)$ is kept to first order only. However, the nonlinear Duffing and transmission line-mechanical oscillator interaction terms can give rise to a bistability in the transmission line oscillator dynamics and one must be careful when interpreting the results from the mean field approximation when operating close to a bifurcation point; large fluctuations can occur in the oscillator amplitude as it jumps between the two metastable amplitudes, which are not accounted for in the mean field approximation. (See Refs. and  for respective analyses of the classical and quantum oscillator fluctuation dynamics near a bifurcation point). This issue will be further discussed in the following sections. The solutions to the signal $a_T^{(1)}(\omega)$ and noise $a_T^{(0)}$ terms parallel closely our previous calculations, which omitted the Duffing nonlinearity;[@Blencowe:2007p285] the Duffing ($D$) term in Eq. (\[eq:atwonly\]) has a very similar form to the transmission line oscillator coupling ($B$) term, both involving $a_T^2 a_T^+$ operator combinations. We therefore relegate the solution details to the appendix, presenting only the essential results in this section. The solution to $\langle a _T^{(0)}(\omega)\rangle$ is sharply peaked about the pump frequency $\omega_p$ for large $T_M$ and so can be approximately expressed as a delta function: $\langle a _T^{(0)}(\omega)\rangle=\chi \delta (\omega-\omega_p)$. Substituting this expression into Eq. (\[eq:zeroth-atw-coherent\]), we obtain for the amplitude $\chi$: $$\label{eq:mean-field0} \chi=c+\left[2B(\omega_{p},0)+D(\omega_{p})\right]\chi\left|\chi\right|^{2},$$ with $$c=\frac{i\sqrt{2\pi}e^{i\phi_{pT}}}{\gamma_{pT}-i\Delta\omega}\sqrt{\frac{I_{0}^{2}Z_{p}\gamma_{pT}}{\hbar\omega_{p}}}, \label{eq:c}$$ where $\Delta\omega=\omega_{p}-\omega_{T}$ is the detuning of the pump frequency $\omega_p$ from the transmission line resonance frequency $\omega_T$. Using the expressions for $B(\omega_p,0)$ and $D(\omega_p)$, Eq. (\[eq:mean-field0\]) can be written as $$(\omega_T-\omega_p -i \gamma_{pT})\chi+\frac{\omega_T}{2\pi} {\mathcal{K}}\chi |\chi|^2 = e^{i\phi_{pT}}\sqrt{{{2\pi}I_0^2 Z_p \gamma_{pT}}/{(\hbar\omega_p})},\label{eq:meanfield}$$ where the effective Duffing coupling is defined as $${\mathcal{K}}=K_d -\frac{2 \omega_T \omega_m}{\omega_m^2 +\gamma_{bm}^2} K^2_{Tm}. \label{eq:k}$$ Notice that the interaction between the transmission line and mechanical oscillator induces an additional Duffing nonlinearity (the second term involving $K_{Tm}$ in ${\mathcal{K}}$) in the transmission line mode amplitude effective equations of motion (\[eq:meanfield\]). However, in contrast with the $K_d$ nonlinearity, which can be tuned to have either sign, the former mechanically-induced nonlinearity is always negative and thus has a “spring-softening" affect on the transmission line mode. Interestingly, by choosing an appropriate compensating “spring hardening" $K_d>0$, the effective Duffing constant ${\mathcal{K}}$ can in principle be completely suppressed so that the next non-vanishing higher order nonlinearity would govern the mode amplitude dynamics. Once we have the solution for $\langle a _T^{(0)}(\omega)\rangle$, the solutions for the quantum signal $a_T^{(1)}(\omega)$ and quantum noise $\delta a_T^{(0)}(\omega)$ are obtained from Eqs. (\[eq:first-atw\]) and (\[eq:zeroth-atw-quantum\]), respectively. These solutions can be expressed as follows: $${a}_{T}^{(1)}(\omega)=\alpha_{1}(\omega)A(\omega,\omega-\omega_{p})+\alpha_{2}(\omega)A(\omega-2\Delta\omega,\omega-\omega_{p})\label{eq:a1solution}$$ and $$\label{eq:a0solution} \delta{a}_{T}^{(0)}(\omega)=\beta_{1}(\omega)\delta C(\omega)+\beta_{2}(\omega)\delta C^{+}(2\omega_{p}-\omega),$$ where the $\alpha_{i}(\omega)$ and $\beta_{i}(\omega)$ functions are defined in Eqs. (\[eq:alpha1\]), (\[eq:alpha2\]), (\[eq:beta1\]), and (\[eq:beta2\]). Substituting Eqs. (\[eq:a1solution\]) and (\[eq:a0solution\]) into the expression (\[eq:asignalnoise\]) for $a^{\mathrm{out}}(\omega)$ and then evaluating the signal component of the detector response (\[eq:currentvariance\]), we obtain[@Blencowe:2007p285] $$\begin{aligned} \label{eq:current-signal} &&\left.\overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s},\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{signal}}=\left(\frac{I_{0}K_{Tm}\omega_{T}}{\gamma_{pT}}\right)^{2}\frac{\gamma_{pT}^{2}}{\gamma_{pT}^{2}+\Delta\omega^{2}}\cr &&\times\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega}{2\pi}\left(\frac{\omega}{\omega_{p}}\frac{\gamma_{pT}^{2}}{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{pT}^{2}}\right)\left|\frac{\alpha_{1}(\omega)}{c}+\frac{\alpha_{2}(\omega)}{c}\left(\frac{\omega-\omega_{p}+\Delta\omega+i\gamma_{pT}}{\omega-\omega_{p}-\Delta\omega+i\gamma_{pT}}\right)\right|^{2}\cr &&\times\left(\frac{2\gamma_{bm}}{(\omega-\omega_{p}-\omega_{m})^{2}+\gamma_{bm}^{2}}[2n(\omega-\omega_{p})+1]+\frac{2\gamma_{bm}}{(\omega_{p}-\omega-\omega_{m})^{2}+\gamma_{bm}^{2}}[2n(\omega_{p}-\omega)+1]\right)\cr &&+\left(\frac{I_{0}K_{Tm}\omega_{T}}{\gamma_{pT}}\right)^{2}\frac{\gamma_{pT}^{2}}{\gamma_{pT}^{2}+\Delta\omega^{2}}\frac{1}{2m\hbar\omega_{m}\gamma_{bm}}\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega d\omega'}{2\pi}\left(\frac{\omega}{\omega_{p}}\frac{\gamma_{pT}^{2}}{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{pT}^{2}}\right)\cr &&\times\left|\frac{\alpha_{1}(\omega)}{c}+\frac{\alpha_{2}(\omega)}{c}\left(\frac{\omega-\omega_{p}+\Delta\omega+i\gamma_{pT}}{\omega-\omega_{p}-\Delta\omega+i\gamma_{pT}}\right)\right|^{2}\frac{\sin\left[(\omega-\omega')T_{m}/2\right]}{(\omega-\omega')T_{m}/2}\cr &&\times\left(\frac{2\gamma_{bm}}{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{bm}^{2}}F_{\mathrm{ext}}(\omega-\omega_{p})F^{*}_{\mathrm{ext}}(\omega'-\omega_{p})\right.\cr &&\left.+\frac{2\gamma_{bm}}{(\omega_{p}-\omega-\omega_{m})^{2}+\gamma_{bm}^{2}}F_{\mathrm{ext}}(\omega_{p}-\omega)F^{*}_{\mathrm{ext}}(\omega_{p}-\omega')\right), \end{aligned}$$ where $n(\omega)=(e^{\hbar\omega/k_BT} -1)^{-1}$ is the thermal average occupation number for bath mode $\omega$. In the limit of small drive current amplitude $I_0\rightarrow 0$, we have $\alpha_1 (\omega)/c \rightarrow 1$, $\alpha_2 (\omega)/c \rightarrow 0$, and we see that the signal spectrum comprises two Lorentzian peaks centered at $\omega_p\pm\omega_m$. The $\omega_p+\omega_m$ peak corresponds to phase preserving detection, in the sense that $a_p^{\mathrm{out}}$ gives the amplified $a_b^{\mathrm{in}}$ signal, while the $\omega_p-\omega_m$ peak corresponds to phase conjugating detection, with $a_p^{\mathrm{out}}$ amplifying the $a^{\mathrm{in}+}_b$ signal.[@Caves:1982p1311] Increasing the drive current amplitude causes the peaks to shift, and the peak widths relative to their height to change, signifying renormalization of the mechanical oscillator frequency and damping rate. The noise component of the detector response is $$\begin{aligned} \label{eq:current-noise} &&\left.\overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s},\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{noise}}=\frac{1}{Z_{p}}\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega}{2\pi}\hbar\omega\frac{2\gamma_{pT}^{2}}{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{pT}^{2}}\cr &&\times\left(\left|\beta_{1}(\omega)\right|^{2}+\frac{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{pT}^{2}}{(\omega-\omega_{p}-\Delta\omega)^{2}+\gamma_{pT}^{2}}\left|\beta_{2}(\omega)\right|^{2} -{\mathrm{Re}}\left[\beta_{1}(\omega)\right]+\frac{(\omega-\omega_{p}+\Delta\omega)}{\gamma_{pT}}{\mathrm{Im}}\left[\beta_{1}(\omega)\right]\right)\cr &&+Z_{p}^{-1}\frac{\hbar\omega_{s}}{2}\frac{\delta\omega}{2\pi},\end{aligned}$$ where the integral term involving the $\beta_i (\omega)$ functions includes the back reaction noise on the mechanical oscillator and the term involving $Z_p$ describes the probe line zero-point fluctuations added at the output. In Sec. \[sec:detection\] we will numerically evaluate Eqs. (\[eq:current-signal\]) and (\[eq:current-noise\]) and in particular compare the detector noise with the minimum noise bound discussed by Caves that follows from the Heisenberg uncertainty principle for the detector: $$\begin{aligned} \label{eq:current-min} &&\left.\overline{\langle\left[\delta I^{\mathrm{out}}(\omega_{s},\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{min-noise}}=\left|Z_p^{-1}\frac{\hbar\omega_{s}}{2}\frac{\delta\omega}{2\pi}-\left(\frac{I_{0}K_{Tm}\omega_{T}}{\gamma_{pT}}\right)^{2}\frac{\gamma^{2}_{pT}}{\gamma^{2}_{pT}+\Delta\omega^{2}}\right.\cr &&\left.\times\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega}{2\pi}\left(\frac{\omega}{\omega_{p}}\frac{\gamma_{pT}^{2}}{(\omega-\omega_{p}+\Delta\omega)^{2}+\gamma_{pT}^{2}}\right)\left|\frac{\alpha_{1}(\omega)}{c}+\frac{\alpha_{2}(\omega)}{c}\left(\frac{\omega-\omega_{p}+\Delta\omega+i\gamma_{pT}}{\omega-\omega_{p}-\Delta\omega+i\gamma_{pT}}\right)\right|^{2}\right.\cr &&\left.\times \left(\frac{2\gamma_{bm}}{(\omega-\omega_{p}-\omega_{m})^{2}+\gamma^{2}_{bm}}-\frac{2\gamma_{bm}}{(\omega_{p}-\omega-\omega_{m})^{2}+\gamma^{2}_{bm}}\right)\right|.\end{aligned}$$ \[sec:bistability\]Bistability Conditions ========================================= We have seen \[Hamiltonian (\[eq:hamiltonian-closed\])\] that the current-dependent SQUID effective inductance gives rise to a transmission line Duffing type nonlinearity with strength $K_d$. Furthermore, there is a nonlinear coupling with strength $K_{Tm}$ between the transmission line and mechanical oscillator. These two nonlinearities correspond respectively to the cubic terms proportional to $K_{d}$ and $K^2_{Tm}$ in the mean transmission line coordinate amplitude $\chi$ equation (\[eq:mean-field0\]). For sufficiently large drive current amplitude $I_0$ and/or coupling strengths $K_{Tm}$, $K_d$, the cubic term $\chi\left|\chi\right|^2$ term in Eq. (\[eq:meanfield\]) becomes appreciable, resulting in three real solutions over a certain pump frequency range $\omega_p$. This parameter regime defines the bistable region of the detector phase space (the intermediate amplitude solution is unstable and cannot be realized in practice). In the following, we determine the conditions on the parameters for the bistable region employing the analysis of Ref. . We first express the transmission line mode coordinate in terms of its phase and amplitude: $$\begin{aligned} \label{eq:chi_B} \langle a_T^{(0)}(t)\rangle&=&Me^{-i(\omega_{p}t+\phi_{M})},\cr \chi&=&\sqrt{2\pi}M e^{-i\phi_M}, \end{aligned}$$ where the amplitude $M$ is a positive real constant and recall $\chi$ is defined through the relation $ \langle a_T^{(0)}(\omega)\rangle=\chi \delta(\omega-\omega_p)$. Equation (\[eq:meanfield\]) then becomes $$\label{eq:kerr-eq} \left(\omega_{T}-\omega_{p}-i\gamma_{pT}\right)M+\mathcal{K}\omega_T M^{3}=\sqrt{2\gamma_{pT}}\langle b_{pT}^{\mathrm{in}}\rangle e^{i(\phi_{pT}+\phi_{M})},$$ where $$\label{eq:bin} \langle b_{pT}^{\mathrm{in}}\rangle=\sqrt{\frac{I_{0}^{2}Z_{p}}{2\hbar\omega_{p}}}.$$ Multiplying both sides of Eq. (\[eq:kerr-eq\]) by their complex conjugates and substituting $E=M^{2}$, we obtain the following third-order polynomial in $E$: $$\label{eq:third_poly} E^{3}+\frac{2(\omega_{T}-\omega_{p})}{\omega_T\mathcal{K}}E^{2}+\frac{(\omega_{T}-\omega_{p})^{2}+\gamma_{pT}^{2}}{\omega_T^2\mathcal{K}^{2}}E=\frac{2\gamma_{pT}\langle b_{pT}^{\mathrm{in}}\rangle^{2}}{\omega_T^2\mathcal{K}^{2}}.$$ The bifurcation line in current drive and detuning parameter space that delineates between the single solution and bistable solution regions occurs where the susceptibility $\partial E/\partial\omega_{p}$ diverges. If we further impose the condition that the transition between the two regions is continuous, i.e. $\partial^{2}\omega_{p}/\partial^{2}E=0$, we obtain the bistability onset critical point. From Eq. (\[eq:third\_poly\]), these two requirements can be written $$\begin{aligned} &3\mathcal{K}^{2}E^{2}+4(\omega_{T}-\omega_{p})\mathcal{K}E+(\omega_{T}-\omega_{p})^{2}+\gamma_{pT}^{2}=0,\label{eq:susceptibility} \\ &6\mathcal{K}^{2}E+4(\omega_{T}-\omega_{p})\mathcal{K}=0.\end{aligned}$$ Solving these equations simultaneously for $E$ and $\Delta\omega=\omega_p-\omega_T$ yields the following bistability onset critical values: $$\begin{aligned} E_{{bi}}&=&\frac{2\gamma_{pT}}{\sqrt{3}\omega_T|\mathcal{K}|}\label{eq:bi}, \cr \Delta\omega_{{bi}}&=&\sqrt{3}\gamma_{pT}\frac{\mathcal{K}}{|\mathcal{K}|}.\label{eq:detune}\end{aligned}$$ Substituting these critical values into Eq. (\[eq:third\_poly\]) gives $$\langle b_{pT}^{\mathrm{in}}\rangle_{{bi}}^{2}=\frac{4\gamma_{pT}^{2}}{3\sqrt{3}\omega_T|\mathcal{K}|}.$$ Finally, using Eq. (\[eq:bin\]) we obtain the driving current critical amplitude: $$\label{eq:Ibi} I_{{bi}}=2\gamma_{pT}\sqrt{\frac{2\hbar\omega_{p}}{3\sqrt{3}\omega_T|\mathcal{K}|Z_{p}}}.$$ Note, the requirement that we operate below the Josephson critical current, $I_{0}<I_{c}$, gives a lower limit on the value of $|\mathcal{K}|$ for which our system can approach the bistability onset. The boundary of the bistable region that is given by the vanishing susceptibility equation (\[eq:susceptibility\]) can be expressed in units of the bistability onset critical current $I_{bi}$ and detuning value $\Delta\omega_{bi}$ using Eqs. (\[eq:Ibi\]) and (\[eq:detune\]) to obtain[@Dykman:1980p480] $$\label{eq:boundary} \frac{I}{I_{bi}}=\frac{1}{2}\left(\frac{\Delta\omega}{\Delta\omega_{bi}}\right)^{3/2}\left\{1+3\left(\frac{\Delta\omega_{bi}}{\Delta\omega}\right)^{2}\pm\left[1-\left(\frac{\Delta\omega_{bi}}{\Delta\omega}\right)^{2}\right]^{3/2}\right\}^{1/2},$$ where the $\pm$ roots give the upper and lower boundaries of the bistable region, respectively (see Fig. \[fig:region\]). As mentioned in the preceding section, care must be taken when applying our semiclassical, mean field approximations to the detector signal and noise response when approaching closely the bifurcation boundary lines. Fluctuation-induced jumps between the small and large amplitude solutions of the transmission line mode can occur that are not accounted for in the mean field approximation. Nevertheless, in the next two sections we shall in some instances evaluate the detector response close the boundaries of the bistability region. For example, we shall see that significant improvements in cooling can be achieved provided a way is found to keep the transmission line mode on the low amplitude solution branch when operating in the bistable region. ![Bistable region (shaded) of the cavity-oscillator system for negative, spring softening Duffing nonlinearity. The drive current and detuning are expressed in units of the bistability onset critical values $I_{bi}$ and $|\Delta\omega_{bi}|$. The labelled straight line traces correspond to detection ($d$) and cooling ($c$) current drive-detuning parameter examples considered in Secs \[sec:detection\] and \[sec:cooling\]. The arrows give the direction in which the drive current is varied in order to enter the bistable region on the small amplitude branch.[]{data-label="fig:region"}](region){width="2.7in"} \[sec:detection\]Displacement Detection ======================================= Assuming that $\gamma_{bm}\ll\gamma_{pT}$, i.e., the unrenormalized mechanical oscillator amplitude damping rate is much smaller than the transmission line oscillator amplitude damping rate, then the detector spectral noise and response in the mechanical signal bandwidth is approximately white over a large range of drive current and detuning parameter space. The mechanical signal and noise response spectra are therefore approximately Lorentzian and Eqs. (\[eq:current-signal\]) and (\[eq:current-noise\]) can be parametrized as $$\begin{aligned} &&\left.\overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s}=\omega_p\pm R_{\omega}\omega_m,\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{signal}} Z_p\cr &&=G_{\pm} \frac{\hbar}{2m R_{\omega}\omega_m}\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega}{2\pi} \frac{2\gamma_{bm}[2n(R_{\omega}\omega_m)+1]}{(\omega-\omega_p\mp R_{\omega}\omega_m)^2 +(R_{\gamma} \gamma_{bm})^2} \label{eq:paramcurrentsignal}\end{aligned}$$ and $$\begin{aligned} &&\left.\overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s}=\omega_p\pm R_{\omega}\omega_m,\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{noise}} Z_p\cr &&=G_{\pm} \frac{\hbar}{2m R_{\omega}\omega_m}\int_{\omega_{s}-\delta\omega/2}^{\omega_{s}+\delta\omega/2}\frac{d\omega}{2\pi} \frac{2\gamma_{\mathrm{back}}[2n^{\pm}_{\mathrm{back}}+1]}{(\omega-\omega_p\mp R_{\omega}\omega_m)^2 +(R_{\gamma} \gamma_{bm})^2}\cr &&+\left.\overline{\langle\left[{\delta I}^{\mathrm{out}}(\omega_{s}=\omega_p\pm R_{\omega}\omega_m,\delta\omega)\right]^{2}\rangle}\right|_{\mathrm{added}~\mathrm{noise}} Z_p, \label{eq:paramcurrentnoise}\end{aligned}$$ where $G_{\pm}$ is the phase preserving (conjugating) gain (in W$\cdot$m$^{-2}$), $n(R_{\omega}\omega_m)$ is the mechanical oscillator’s external bath occupation number at its renormalized frequency $R_{\omega} \omega_m$, $R_{\gamma} \gamma_{bm}$ is the renormalized (i.e., net) mechanical oscillator damping rate, and the detector back reaction noise on the oscillator is effectively that of a thermal bath with damping rate $\gamma_{\mathrm{back}}=\gamma_{bm} (R_{\gamma}-1)$ and thermal average occupation number $n^{\pm}_{\mathrm{back}}$. Note, from here on we do not consider an external classical force driving the mechanical oscillator; the focus is on displacement detection rather than force detection. The added noise term in Eq. (\[eq:paramcurrentnoise\]) comprises output noise that is not due to the action of the detector on the mechanical oscillator; the added noise is present even when there is no coupling to the mechanical oscillator, i.e., when $K_{Tm}=0$. In the absence of the transmission line Duffing nonlinearity, the added noise simply consists of the probe line zero-point fluctuations $\hbar\omega_s\delta\omega/(4\pi Z_p)$. However, with the Duffing nonlinearity present, the added noise will be in excess of the probe line zero-point fluctuations. The convenient Lorentzian parametrization approximations of the mechanical signal (\[eq:paramcurrentsignal\]) and noise response spectra (\[eq:paramcurrentnoise\]) that provide the above-described effective thermal description of the back reaction noise will break down as one approaches arbitrarily closely the jump points at the ends of the small or large amplitude transmission line oscillator solution branches occuring at the boundaries of the bistable region indicated in Fig. \[fig:region\]. This is a consequence of the diverging damping (i.e., ring-down) time of transmission line mode.[@Yurke:2006p1911] Thus, when numerically solving  (\[eq:current-signal\]) and (\[eq:current-noise\]) to extract the effective thermal properties of the detector back reaction, it is important to always check the accuracy of the Lorentzian spectrum approximation. For sufficiently large gain (i.e., large current drive amplitude), we can neglect the added noise contribution and we have for the noise-to-signal response ratio when the mechanical oscillator external bath is at absolute zero \[i.e., $n(R_{\omega} \omega_m)=0$\]: $$\frac{\langle\left[{\delta I}^{\mathrm{out}}\right]^{2}\rangle_{\mathrm{noise}}}{\langle\left[{\delta I}^{\mathrm{out}}\right]^{2}\rangle_{\mathrm{signal}}}=(2n^{\pm}_{\mathrm{back}}+1)\frac{ \gamma_{\mathrm{back}}}{\gamma_{bm}}. \label{eq:noisetosignal}$$ On the other hand, in the large gain limit the Caves noise lower bound (\[eq:current-min\]) gives a noise-to-signal ratio of one. For large gain, we typically have $|2n^{\pm}_{\mathrm{back}}+1|\gg 1$ and thus to approach the Caves bound necessarily requires $|\gamma_{\mathrm{back}}|\ll \gamma_{bm}$.[@Clerk:2004p245306] As an example, we numerically solve for the signal and noise contributions of the detector response, Eqs. (\[eq:current-signal\]) and (\[eq:current-noise\]) respectively, as well as the Caves lower bound on the quantum noise (\[eq:current-min\]). We consider Duffing nonlinearities $K_d=-3.4\times 10^{-6}$ and $K_d=0$ (i.e., no nonlinearity). The integrated signal and noise bandwidth is taken to be $\delta\omega=2R_{\gamma}\gamma_{bm}$. The corresponding parameter values are: probe line impedance $Z_p=50~{\mathrm{Ohms}}$, transmission line mode angular frequency $\omega_T/(2\pi)=5\times 10^9~{\mathrm{s}}^{-1}$, transmission line mode quality factor $Q_T=\omega_T/(2\gamma_{pT})=300$, mechanical frequency $\omega_m/(2\pi)=4\times 10^6~{\mathrm{s}}^{-1}$, mechanical quality factor $Q_m=\omega_m /(2\gamma_{bm})=10^3$, oscillator mass $m=10^{-16}~{\mathrm{kg}}$, Josephson junction critical current $I_{c}=4.5\times10^{-6}~\mathrm{A}$, junction capacitance $C_{J}=10^{-14}~\mathrm{F}$, external flux bias $\Phi_{\mathrm{ext}}=0.442\ \Phi_0$, and external field in the vicinity of the mechanical resonator $B_{\mathrm{ext}}=0.05~\mathrm{T}$. These values give a zero-point uncertainty $\Delta_{zp}=1.45 \times 10^{-13}~{\mathrm{m}}$ and transmission line-oscillator coupling $K_{Tm}=1.1\times10^{-5}$. ![Detector noise versus signal response at $\Delta\omega=0$ for harmonic ($K_d=0$) transmission line, Duffing ($K_d<0$) transmission line ($d_1$), and Caves’ bound (black-dashed). Noise for the nonlinear transmission line is also evaluated for blue detunings: $\Delta\omega=+0.2|\Delta\omega_{bi}|$ ($d_2$), and $+0.4|\Delta\omega_{bi}|$ ($d_3$). The labeled curves correspond to the traces in Fig. \[fig:region\]. The dashed, colored lines give Caves’ bound for the corresponding detuning values.[]{data-label="fig:detection"}](detection){width="4.0in"} The advantage of using a spring softening nonlinearity, $K_d<0$, is clearly evident in Fig. \[fig:detection\], where we plot the noise versus response signal under increasing current drive for a transmission line both with and without Duffing term driven on resonance, $\Delta\omega=\omega_p-\omega_T=0$. We also plot the response of the nonlinear transmission line for several positively detuned values, $\Delta\omega=\omega_p-\omega_T>0$. Termination of the curves indicates the signal value at which the damping renormalization $R_{\gamma}= 0$, beyond which the derived solutions become unphysical due to the net mechanical damping rate becoming negative and hence the motion unstable about the original fixed point. Note that the same criterion, namely $R_{\gamma}>0$, is employed throughout the paper in order to ensure stability of the system. Again, the semiclassical, mean field approximation is expected to break down in the vicinity of termination points, where large fluctuations in the mechanical oscillator amplitude occur. In Fig. \[fig:detection\], we see that with positive detuning, we can further approach the Caves bound. However, this is at the expense of reduced gain; depending on one’s point of view, large renormalizations of the mechanical oscillator damping rate (and frequency) due to detector back action may or may not be allowed in detector displacement sensitivity figures of merit, affecting the maximum achievable gain as one approaches more closely the Caves bound. The trends displayed in Fig. \[fig:detection\] can be partly explained by invoking Fig. \[fig:curves\], which indicates qualitatively the force on the mechanical oscillator due to the microwave transmission line ‘ponderomotive radiation pressure’ force, both with vanishing and with nonzero Duffing nonlinearity and also for ‘red’ and ‘blue’ pump frequency detunings. The work done on the mechanical oscillator by the radiation pressure force during one period of motion, due to the delayed transmission line resonator response, is given by the area enclosed within the hysteresis loop[@Kippenberg:2007p1995; @Marquardt:2008p1309] and can be related to the steady-state back action damping rate through $$\gamma_{\mathrm{back}}=-\frac{W}{\bar{E}}\frac{1}{\tau}, \label{eq:hysteresis}$$ where $W$ is the work done on the mechanical oscillator, $\bar{E}$ is the average oscillator energy and $\tau$ is the period of motion. ![Cartoon indicating the ‘radiation pressure’ force exerted on the mechanical oscillator by the transmission line mode during one cycle of mechanical motion: (a) the harmonic transmission line mode approximation; (b) approaching the onset of bistability. The work done on the oscillator is proportional to the area swept out during each cycle, considerably exaggerated here for clarity. Positive mechanical damping (red detuning) results on the positive slope side of the curves. Negative mechanical damping (blue detuning) results on the negative slope side. A spring softening nonlinearity can result in improved cooling for red detuning and improved signal-to-noise amplification for blue detuning. []{data-label="fig:curves"}](curves){width="5.5in"} When frequency pulling is taken into account, the usual notions of red-detuned ($\Delta\omega<0$) or blue-detuned ($\Delta\omega>0$) hold only in the weak drive limit. We will assume red (blue)-detuned to correspond to drive and detuning values $\Delta\omega$ where the net work done on the oscillator is negative (positive) as seen in Fig. \[fig:curves\]. For a harmonic transmission line and for low drive powers, the frequency pulling effects can be ignored, since the effective Duffing coupling Eq. (\[eq:k\]) is proportional to the square of the transmission line-mechanical oscillator coupling $K_{Tm}$, which contributes only weakly for the considered parameter values. Conversely, the Duffing term causes frequency pulling even at low input power and can significantly alter the slope of the response curve. From Eq. (\[eq:hysteresis\]), the decreased slope on the blue detuned side leads to a decrease in the damping rate magnitude which, through Eq. (\[eq:noisetosignal\]), leads as demonstrated above to a closer approach to the Caves’ limit. As mentioned above, benefits in lower noise-to-signal resulting from further detuning deep into the blue region are offset by diminished achievable signal gain levels. Tuning the sign of the Duffing coupling $\mathcal{K}$ (\[eq:k\]) to be positive, so that we have a hardening spring, results in an increased back action damping rate for blue detuning, and hence a corresponding decrease in signal to noise relative to the harmonic transmission line resonator detector case. \[sec:cooling\] Cooling ======================= Referring to the parametrizations (\[eq:paramcurrentsignal\]) and (\[eq:paramcurrentnoise\]) of the signal and noise components of the detector response, we define the mechanical oscillator’s net occupation number through the following equation: $$\gamma_{\mathrm{net}}(2n^{\pm}_{\mathrm{net}}+1)=\gamma_{bm}\left[2n(R_{\omega}\omega_{m})+1\right]+\gamma_{\mathrm{back}}(2n^{\pm}_{\mathrm{back}}+1), \label{eq:gammanetnnet}$$ where the net damping rate is $\gamma_{\mathrm{net}}=\gamma_{bm}+\gamma_{\mathrm{back}}=R_{\gamma}\gamma_{bm}$. The oscillator’s net occupation number is then $$2n_{\mathrm{net}}+1=R_{\gamma}^{-1}\left[2n(R_{\omega}\omega_{m})+1\right] +(1-R_{\gamma}^{-1}) (2n^{\pm}_{\mathrm{back}}+1). \label{eq:nnet}$$ In order to cool a mechanical oscillator to its ground state using detector back action, we therefore require a large detector back action damping rate, equivalently large damping rate renormalization $R_{\gamma}\gg 1$, together with a small detector back action effective occupation number $n^{\pm}_{\mathrm{back}}\ll 1$. Referring to Fig. \[fig:curves\], operating closer to the bistability increases the negative work done per cycle on the oscillator by the cavity and hence increases the back action damping rate for given current drive. In Fig. \[fig:gamma\], we plot the mechanical oscillator damping rate renormalization factor $R_{\gamma}$, using the same parameter values as in Sec. \[sec:detection\] (e.g., Duffing coupling $K_{d}=-3.4\times10^{-6}$), but with a larger yet still feasible mechanical quality factor $Q_m=10^4$ (which we shall adopt throughout this section). We clearly see the enhanced damping as one approaches the onset of bistability given by $I_{{bi}}$ (\[eq:Ibi\]) and $\Delta\omega_{{bi}}<0$ (\[eq:detune\]). ![Mechanical oscillator damping renormalization factor $R_{\gamma}$ for detunings both above and below the bistable detuning $\Delta\omega_{bi}$. The amplification region corresponds to negative back action damping, i.e., $R_{\gamma}<1$.[]{data-label="fig:gamma"}](gamma){width="4.0in"} For the example parameter choices of Sec. \[sec:detection\], we have $\omega_{m}/\gamma_{pT}\approx0.5$ and thus we are operating in the so-called bad cavity limit, where cooling close to the ground state (i.e., $n_{\mathrm{net}}\ll 1$) is not possible.[@Blencowe:2007p285; @Marquardt:2007p978; @WilsonRae:2007p502] While it is not difficult to achieve the good cavity limit $\omega_m>\gamma_{pT}$ simply by realizing sufficiently large quality factor superconducting microwave resonators, together with high frequency mechanical resonators,[@Teufel:2008p2398] it is nevertheless worthwhile to address how nonlinearities can improve on the cooling limits in the bad-cavity case. With the fundamental motivation to demonstrate macroscopic quantum behavior, the anticipated trend is to work with increasingly massive and hence lower frequency oscillators, making it progressively more difficult to achieve the good cavity limit. In Fig. \[fig:optimal\], we plot the dependence of detector’s noise effective back action occupation number $n_{\mathrm{back}}$ on microwave drive current amplitude at the detuning bias $\Delta\omega=-\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}$, where $|\Delta\omega|<|\Delta\omega_{bi}|$. This is the optimum detuning in the harmonic, transmission line oscillator approximation, i.e., when nonlinear effects are ignored. The noise effective occupation number is indicated for both a nonzero ($K_d=-3.4\times10^{-6}$) as well as zero ($K_d=0$) Duffing nonlinearity transmission line. We also show for comparison the effective back action occupation number when the frequency pulling effects of both the ponderomotive coupling $K_{Tm}$ and Duffing coupling $K_{d}$ are neglected. The latter case is obtained by dropping the nonlinear microwave mode amplitude term in the mean field equation (\[eq:mean-field0\]). The sharp rise in occupation number and associated sharp drop in damping renormalization at larger current drives is a consequence of crossing over into the amplification region due to negative frequency pulling of the cavity response relative to the fixed detuning. The decrease in occupation number as $I \rightarrow 0$ is accompanied by weak back action damping, which prevents cooling the mechanical oscillator to such occupation numbers. Note that at smaller current drives the damping renormalization in the presence of a Duffing nonlinearity peaks above the corresponding damping renormalization without the Duffing nonlinearity. This damping enhancement can be qualitatively explained with the aid of Fig. \[fig:curves\]. In the presence of the nonlinearity then, improved cooling can be achieved for smaller current drives. ![(a) Detector noise effective back action occupation number versus current drive when red-detuned at $\Delta\omega=-\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}$, $|\Delta\omega|<|\Delta\omega_{bi}|$, with a Duffing nonlinearity (solid line), without a Duffing nonlinearity (dashed line), and without both Duffing and ponderomotive nonlinearities (dotted line). (b) Oscillator coupling renormalization factor $R_{\gamma}$ for the corresponding back-action occupation number curves. These plots are obtained for the straight line trace labeled $c_1$ in Fig. \[fig:region\].[]{data-label="fig:optimal"}](optimal300){width="6.0in"} According to the above discussion, any improvements in mechanical oscillator cooling are due solely to enhancements in the detector’s back action damping rate for given drive; as can be seen from Fig. \[fig:optimal\], the absolute minimum attainable detector effective occupation number is the same both in the presence and absence of the transmission line resonator Duffing nonlinearity. While the effects of enhanced back action damping may be beneficial in situations where one is facing constraints on the maximum achievable drive power,[@Teufel:2008p2398] it would nevertheless be more significant if reductions in detector effective occupation number could similarly be achieved through nonlinear effects. To see how this might be possible, we consider detunings corresponding to the pump frequency being to the left and away from the cavity resonance, i.e., $|\Delta\omega| > |\Delta\omega_{bi}|$, $\Delta\omega<0$. For such detunings, the mechanical oscillator ‘sees’ a transmission line resonator effective quality factor that is determined by the steeper slope on the left side of the response curve (see Fig. \[fig:curves\]). As we drive the transmission line resonator towards the lower bistable boundary (see Fig. \[fig:region\]), the slope of the response curve increases sharply and mimics a resonator with larger quality factor, effectively getting closer to the good cavity limit and hence resulting in a lower detector occupation number.[@Blencowe:2007p285; @Marquardt:2007p978; @WilsonRae:2007p502] Continuing to drive the transmission line resonator into the bistable region, and assuming that the resonator can be maintained on the low amplitude, red-detuned solution branch,[@Naaman:2008] the detector effective occupation number further decreases while the back action damping rate on the mechanical resonator increases (as explained by Fig. \[fig:curves\]). Eventually, the transmission line resonator becomes unstable at the upper bistable boundary indicated in Fig. \[fig:region\], and the oscillator jumps to the larger amplitude, blue-detuned solution (see Fig. \[fig:response\]). ![Transmission line resonator response curve for $Q_{T}=300$ restricted to the small amplitude solution branch. The example drive currents are $I/I_{bi}=0.8$ (green), 0.95 (yellow), 1.15 (red), and 1.3 (blue). The jump between small (red-detuned) and large (blue-detuned) amplitude solutions is indicated by the dotted lines.[]{data-label="fig:response"}](response){width="3.2in"} In Fig. \[fig:goodcavity\], we plot the dependence of the detector effective occupation number $n_{\mathrm{back}}$ on current drive for an example detuning value of $\Delta\omega=-2\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}=1.3\Delta\omega_{bi}$. ![(a) Detector noise effective occupation number versus current drive when red detuned at $\Delta\omega=1.3\ \Delta\omega_{bi}$, corresponding to straight line trace $c'_1$ in Fig. \[fig:region\]. The Duffing nonlinear transmission line resonator occupation number (solid line) rapidly decreases as the resonator is driven towards the upper bistable boundary, assuming the resonator can be maintained on the small amplitude metastable stable solution branch. In contrast, a harmonic transmission line resonator (dashed line) or a cavity with neither Duffing nor ponderomotive nonlinearities in its mean field microwave mode equations (dotted line) shows no such decrease in the occupation number. (b) Mechanical oscillator damping renormalization factor for the same fixed detuning and drive current range. The dashed vertical lines indicate the boundaries of the bistable region for the given transmission line resonator parameters.[]{data-label="fig:goodcavity"}](2xoptimal300){width="6.0in"} Driving the nonlinear transmission line resonator towards the upper boundary of the bistable region (see Fig. \[fig:curves\]) produces a sharp decrease in detector occupation number, and an occupation number value of $(2n_{\mathrm{back}}+1)\approx 0.55$ can be obtained, well below that achievable when ignoring frequency pulling effects. The harmonic cavity shows no such decrease in occupation number, indicating the qualitatively different quantum dynamical dependencies on $K_{d}$ and $K_{Tm}$ and the necessity of the former. We can quantify the effect of frequency pulling by comparing with a harmonic transmission line resonator with a quality factor $Q_{T}^{\mathrm{eff}}$ value chosen so as to give the same detector effective occupation number. For the occupation number value $(2n_{\mathrm{back}}+1)\approx 0.55$, we have $Q_{T}^{\mathrm{eff}}\approx 600$, corresponding to $\omega_{m}/\gamma^{\mathrm{eff}}_{pT}=0.95$, and therefore the mechanical oscillator behaves as if it is coupled to a cavity with double the quality factor. This translates into lower net mechanical temperatures as shown in Fig. \[fig:cooling\], where we give the net oscillator occupation number $n_{\mathrm{net}}$ (\[eq:nnet\]) for various external bath temperatures. ![(a) Net mechanical occupation number at $\Delta\omega=-\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}$ for a harmonic transmission line resonator. External bath temperatures: $T=1$ (solid line), 10 (dashed line), 50 (dotted line) and 100 (dot-dashed line) $\mathrm{mK}$. (b) Dependence of the net mechanical oscillator occupation number on current drive for a Duffing transmission line with detuning $\Delta\omega=1.3\ \Delta\omega_{bi}$. The bistable region boundaries are indicated by the dashed vertical lines.[]{data-label="fig:cooling"}](cooling300){width="6.0in"} The combination of nonlinearly-enhanced coupling $R_{\gamma}\gamma_{bm}$ and enhanced transmission line effective quality factor can be seen to significantly affect cooling of the mechanical motion, even for relatively large external temperatures. In the numerical solutions to Eqs. (\[eq:current-signal\]) and (\[eq:current-noise\]), the Lorentzian parametrizations (\[eq:paramcurrentsignal\]) and (\[eq:paramcurrentnoise\]) were found to give good approximations even when the upper bistable boundary is approached quite closely. This is a consequence of the wide separation in the relaxation rates that determine the line widths of the harmonic transmission line resonator and unrenormalized mechanical oscillator modes, i.e., $\gamma_{bm}\ll\gamma_{pT}$. The upper bistable boundary has to be approached pretty closely in order for the nonlinear transmission line resonator ring-down time to exceed the renormalized mechanical oscillator damping time, resulting in the breakdown of the effective thermal description of the detector back reaction. In all of the plots shown in this section, the Lorentzian approximation is a good one over the resolvable scale of the plots. The actual minimum temperature that can be achieved depends on the upper drive threshold where the Lorentzian approximation breaks down, as well as on the ability to keep the transmission line resonator on the small amplitude solution branch; the latter condition becomes progressively more difficult to satisfy as the upper boundary is approached, owing to the increasing probability of noise-induced jumps to the large amplitude branch. A Duffing transmission line resonator nonlinearity can also produce cooling gains in the good cavity limit. In Fig. \[fig:1000\], we consider a transmission line resonator with $Q_{T}=1000$, giving $\omega_{m}/\gamma_{pT}=1.6$, and compare the nonlinear transmission line resonator with the harmonic resonator approximation at optimal harmonic detuning. Again, by detuning to twice the optimal harmonic resonator value, $\Delta\omega=-2\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}\approx2.2 \Delta\omega_{bi}$, we see that the effective back action occupation number decreases, while the back action damping increases as the system is driven towards the upper boundary of the bistable region. ![(a) Detector noise effective occupation number versus current drive when red-detuned at $\Delta\omega=-\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}$, $|\Delta\omega|<|\Delta\omega_{bi}|$, for a Duffing nonlinear (solid line), harmonic (dashed line) transmission line, and with the effects of frequency pulling due to both ponderomotive and Duffing nonlinearities neglected (dotted line). The vertical dashed lines give the bistable region boundaries for the Duffing and harmonic transmission line resonators. This plot is obtained for the straight line trace labeled $c_2$ in Fig. \[fig:region\] (b) Oscillator coupling renormalization factor $R_{\gamma}$ for optimal harmonic detuning. (c) Detector occupation number detuned at twice the harmonic optimum, $\Delta\omega=2.2\Delta\omega_{bi}$ and locked to the lower stable amplitude solution. This plot is obtained for the straight line trace labeled $c'_2$ in Fig. \[fig:region\] (d) Corresponding back-action damping rate when driven to the upper bistable boundary.[]{data-label="fig:1000"}](1000){width="6.0in"} Driving a Duffing transmission line resonator at twice the optimal harmonic detuning can yield a detector occupation number $(2n_{\mathrm{back}}+1)\approx 0.06$ just below the upper boundary of the bistable region, which is equivalent to an effective harmonic resonator quality factor of $Q_{T}^{\mathrm{eff}}\approx 1400$ or $\omega_{m}/\gamma^{\mathrm{eff}}_{pT}=2.2$. In comparison, the minimum effective detector occupation number ignoring nonlinear effects is $2n_{\mathrm{back}}+1=0.13$. In Fig. \[fig:cooling1000\], we plot the net mechanical occupation number for the good cavity transmission line resonator both in the presence and absence of the Duffing nonlinearity. ![(a) Net mechanical occupation number at $\Delta\omega=-\sqrt{\omega_{m}^{2}+\gamma_{pT}^{2}}$ for a harmonic transmission line resonator. External bath temperatures: $T=1$ (solid line), 10 (dashed line), 50 (dotted line) and 100(dot-dashed line) $\mathrm{mK}$. (b) Dependence of the net mechanical oscillator occupation number on current drive for a Duffing transmission line with detuning $\Delta\omega=2.2\Delta\omega_{bi}$.[]{data-label="fig:cooling1000"}](cooling1000){width="6.0in"} Again, we see the strong cooling effects provided by frequency pulling of the cavity response. As discussed above, the minimum achievable net occupation number will depend on the threshold drive for which the Lorentzian approximation breaks down, as well as on the ability to lock the transmission line resonator onto the small amplitude solution branch in the bistable region. \[sec:conclusion\]Conclusions ============================= We have provided a quantum analysis of a nonlinear microwave amplifier for displacement detection and cooling of a mechanical oscillator. The amplifier comprises a microwave stripline resonator with embedded dc SQUID. The SQUID gives rise to an effective, Duffing-type nonlinearity in the fundamental microwave mode equations, as well as a ponderomotive-type coupling between the microwave and fundamental mechanical modes. It was found that a spring-softening Duffing nonlinearity enables a closer approach to the standard quantum limit for position detection as expressed by the Caves bound, as well as cooling closer to the mechanical mode ground state. These findings can be qualitatively explained by considering the effects of frequency pulling in the response curve of the transmission line resonator ‘ponderomotive force’ acting on the mechanical oscillator (see Fig. \[fig:curves\]). With blue detuning, the decrease in damping allows for a closer approach to the quantum limit with large amplifier gain. Conversely, red detuning towards the bistable point of the force response curve increases the back action damping, improving the thermal contact to the detector ‘cold load’. Furthermore, effectively increasing the cavity quality factor due to the nonlinearity mimics the so-called good cavity limit in the harmonic case, allowing cooling closer to the ground state. The present investigation has by no means exhaustively searched the large parameter space of the transmission line resonator-embedded SQUID-mechanical resonator system for establishing the optimal displacement detection sensitivity and cooling parameters. Rather, our intention has been to point out general trends, using specific parameter values as illustrative examples. It may be that other choices of parameters (e.g., using a mechanical resonator with a smaller quality factor) lead to a closer approach to the standard quantum limit, or to cooling closer to the ground state. The semiclassical, mean field methods employed in the present work do not take into account classical or quantum noise-induced jumps between the small and large amplitude metastable solutions that become more likely as the bistability region boundaries are approached. Unless ways can be found to keep the transmission line resonator locked onto the smaller amplitude solution branch,[@Naaman:2008] the predicted effects of nonlinearity-induced cooling will be less substantial, as it will be necessary to operate deeper in the bistability region to avoid jumps. The driven microwave mode amplitude dynamics in the vicinity of the bistable region boundaries is still a relatively unexplored area that requires more sophisticated theoretical techniques in order to elucidate the fluctuations between the small and large amplitude metastable solution branches.[@Dykman:1980p480; @Dykman:2004p061102; @Dykman:2005p021102; @Dykman:2007p1864; @Serban:2007p3199; @Lifshitz:2007p040404; @Kogan:2008p0972] This will be the subject of a future investigation. Acknowledgements {#acknowledgements .unnumbered} ================ We thank A. Armour, M. Dykman, and R. Lifshitz for helpful discussions. This work was partly supported by the US-Israel Binational Science Foundation (BSF), the National Science Foundation (NSF) under NIRT grant CMS-0404031 (M.P.B.), and by the Israeli Science foundation (ISF), the Deborah Foundation and the Ministry of Science (E.B.). \[sec:appendix\] ================ In this appendix, we give the derivation of the signal $a_T^{(1)}(\omega)$ and noise $a_T^{(0)}(\omega)$ terms. Suppressing the signal dependent term $A(\omega,\omega')$ in Eq. (\[eq:atwonly\]), we obtain the noise equation $$\begin{aligned} \label{eq:zeroth-atw} {a}_{T}^{(0)}(\omega)&=&\int_{-\infty}^{\infty}d\omega'B(\omega,\omega'){a}_{T}^{(0)}(\omega-\omega')\cr &\times&\int_{-\infty}^{\infty}d\omega''\left[{a}_{T}^{(0)}(\omega''){a}_{T}^{{(0)}+}(\omega''-\omega')+{a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(0)}(\omega''+\omega')\right]\cr &+&D(\omega)\int_{-\infty}^{\infty}d\omega''\int_{-\infty}^{\infty}d\omega' {a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(0)}(\omega'){a}_{T}^{(0)}(\omega+\omega''-\omega')+C(\omega).\end{aligned}$$ Keeping only terms to first order in $A(\omega,\omega')$ or $a_T^{(1)}$ in Eq. (\[eq:atwonly\]), we obtain the signal equation $$\begin{aligned} \label{eq:first-atw} {a}_{T}^{(1)}(\omega)&=&\int_{-\infty}^{\infty}d\omega'{a}_{T}^{(0)}(\omega-\omega')A(\omega,\omega')+\int_{-\infty}^{\infty}d\omega'B(\omega,\omega'){a}_{T}^{(1)}(\omega-\omega')\cr &\times&\int_{-\infty}^{\infty}d\omega''\left[{a}_{T}^{(0)}(\omega''){a}_{T}^{{(0)}+}(\omega''-\omega')+{a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(0)}(\omega''+\omega')\right]\cr &+&\int_{-\infty}^{\infty}d\omega'B(\omega,\omega'){a}_{T}^{(0)}(\omega-\omega')\int_{-\infty}^{\infty}d\omega''\left[{a}_{T}^{(1)}(\omega''){a}_{T}^{{(0)}+}(\omega''-\omega')\right.\cr &+&\left.{a}_{T}^{{(1)}+}(\omega''){a}_{T}^{(0)}(\omega''+\omega') +{a}_{T}^{(0)}(\omega''){a}_{T}^{{(1)}+}(\omega''-\omega')+{a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(1)}(\omega''+\omega')\right]\cr &+&D(\omega)\int_{-\infty}^{\infty}d\omega''\int_{-\infty}^{\infty}d\omega'\left[{a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(0)}(\omega'){a}_{T}^{(1)}(\omega+\omega''-\omega')\right.\cr &+&\left.{a}_{T}^{{(0)}+}(\omega''){a}_{T}^{(1)}(\omega'){a}_{T}^{(0)}(\omega+\omega''-\omega')+{a}_{T}^{{(1)}+}(\omega''){a}_{T}^{(0)}(\omega'){a}_{T}^{(0)}(\omega+\omega''-\omega')\right].\end{aligned}$$ Decomposing $a_T^{(0)}(\omega)=\langle a_T^{(0)}(\omega)\rangle +\delta a_T^{(0)}(\omega)$ and expanding Eq. (\[eq:zeroth-atw\]) to first order in the quantum noise fluctuation $\delta a_T^{(0)}(\omega)$, we obtain the following two equations: $$\begin{aligned} \label{eq:zeroth-atw-coherent} \langle{a}_{T}^{(0)}(\omega)\rangle&=&\int_{-\infty}^{\infty}d\omega'B(\omega,\omega')\langle{a}_{T}^{(0)}(\omega-\omega')\rangle\cr &\times&\int_{-\infty}^{\infty}d\omega''\left[\langle{a}_{T}^{(0)}(\omega'')\rangle\langle{a}_{T}^{{(0)}+}(\omega''-\omega')\rangle+\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\langle{a}_{T}^{(0)}(\omega''+\omega')\rangle\right]\cr &+&D(\omega)\int_{-\infty}^{\infty}d\omega''\int_{-\infty}^{\infty}d\omega'\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\langle{a}_{T}^{(0)}(\omega')\rangle\langle{a}_{T}^{(0)}(\omega+\omega''-\omega')\rangle\cr &+&\langle C (\omega)\rangle\end{aligned}$$ and $$\begin{aligned} \label{eq:zeroth-atw-quantum} &&\delta{a}_{T}^{(0)}(\omega)=\int_{-\infty}^{\infty}d\omega'B(\omega,\omega')\delta{a}_{T}^{(0)}(\omega-\omega')\cr &\times&\int_{-\infty}^{\infty}d\omega''\left[\langle{a}_{T}^{(0)}(\omega'')\rangle\langle{a}_{T}^{{(0)}+}(\omega''-\omega')\rangle+\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\langle{a}_{T}^{(0)}(\omega''+\omega')\rangle\right]\cr &+&\int_{-\infty}^{\infty}d\omega'B(\omega,\omega')\langle{a}_{T}^{(0)}(\omega-\omega')\rangle \int_{-\infty}^{\infty}d\omega''\left[\delta{a}_{T}^{(0)}(\omega'')\langle{a}_{T}^{{(0)}+}(\omega''-\omega')\rangle\right.\cr &+&\left.\delta{a}_{T}^{{(0)}+}(\omega''-\omega')\langle{a}_{T}^{(0)}(\omega'')\rangle+\delta{a}_{T}^{{(0)}+}(\omega'')\langle{a}_{T}^{(0)}(\omega''+\omega')\rangle\right.\cr &+&\left.\delta{a}_{T}^{(0)}(\omega''+\omega')\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\right]\cr &+&D(\omega)\int_{-\infty}^{\infty}d\omega''\int_{-\infty}^{\infty}d\omega'\left[\delta{a}_{T}^{{(0)}+}(\omega'')\langle{a}_{T}^{(0)}(\omega')\rangle\langle{a}_{T}^{(0)}(\omega+\omega''-\omega')\rangle\right.\cr &+&\left.\delta{a}_{T}^{(0)}(\omega')\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\langle{a}_{T}^{(0)}(\omega+\omega''-\omega')\rangle+\delta{a}_{T}^{(0)}(\omega+\omega''-\omega')\langle{a}_{T}^{{(0)}+}(\omega'')\rangle\langle{a}_{T}^{(0)}(\omega')\rangle\right]\cr &+&\delta C(\omega).\end{aligned}$$ Assuming $\langle a_T^{(0)}(\omega)\rangle$ can be expressed approximately as a delta function, i.e., $\langle a_T^{(0)}(\omega)\rangle=\chi\delta(\omega-\omega_p)$, Eq. (\[eq:zeroth-atw-coherent\]) reduces to Eq. (\[eq:mean-field0\]) for $\chi$. The semiclassical approximation to Eq. (\[eq:first-atw\]) for $a_T^{(1)}(\omega)$, with $a_T^{(0)}(\omega)$ replaced by $\langle a_T^{(0)}(\omega)\rangle=\chi\delta(\omega-\omega_p)$, then becomes $$\begin{aligned} \label{eq:atw_chi} &&\left\{1-2\left|\chi\right|^{2}\left[B(\omega,0)+B(\omega,\omega-\omega_{p})+D(\omega)\right]\right\}{a}_{T}^{(1)}(\omega)\cr &&-\chi^{2}\left[2B(\omega,\omega-\omega_{p})+D(\omega)\right]{a}_{T}^{{(1)}+}(2\omega_{p}-\omega)=\chi A(\omega,\omega-\omega_{p}).\end{aligned}$$ In order to invert and obtain ${a}_{T}^{(1)}(\omega)$, we require a second, linearly independent equation that also depends on ${a}_{T}^{{(1)}+}(2\omega_{p}-\omega)$. Such an equation is obtained by making the replacement $\omega\rightarrow 2\omega_{p}-\omega$ in Eq. (\[eq:atw\_chi\]) and then taking the adjoint: $$\begin{aligned} \label{eq:atw_one_chi} &&\left\{1+2\left|\chi\right|^{2}\left[B(\omega-2\Delta\omega,0)+B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right]\right\}{a}_{T}^{{(1)}+}(2\omega_p-\omega)\cr &&+\chi^{*2}\left[2B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right]{a}_{T}^{{(1)}}(\omega)=-\chi^* A(\omega-2\Delta\omega,\omega-\omega_{p}).\end{aligned}$$ Now inverting, we obtain $${a}_{T}^{(1)}(\omega)=\alpha_{1}(\omega)A(\omega,\omega-\omega_{p})+\alpha_{2}(\omega)A(\omega-2\Delta\omega,\omega-\omega_{p}),$$ where $$\alpha_{1}(\omega)=\mathcal{D}(\omega)^{-1}\left\{1+2\left|\chi\right|^{2}\left[B(\omega-2\Delta\omega,0)+B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right]\right\}\chi,\label{eq:alpha1}$$ $$\alpha_{2}(\omega)=-\mathcal{D}(\omega)^{-1}\left[2B(\omega,\omega-\omega_{p})+D(\omega)\right]\left|\chi\right|^{2}\chi\label{eq:alpha2}$$ and the determinant is given by, $$\begin{aligned} \label{eq:determinant} \mathcal{D}(\omega)&=&\left\{1-2\left|\chi\right|^{2}\left[B(\omega,0)+B(\omega,\omega-\omega_{p})+D(\omega)\right]\right\}\cr &\times&\left\{1+2\left|\chi\right|^{2}\left[B(\omega-2\Delta\omega,0)+B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right]\right\}\cr &+&\left|\chi\right|^{4}\left[2B(\omega,\omega-\omega_{p})+D(\omega)\right] \left[2B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right].\end{aligned}$$ A similar approach is used to obtain $\delta{a}_{T}^{(0)}(\omega)$ from Eq. (\[eq:zeroth-atw-quantum\]), giving $$\delta{a}_{T}^{(0)}(\omega)=\beta_{1}(\omega)\delta C(\omega)+\beta_{2}(\omega)\delta C^{+}(2\omega_{p}-\omega),$$ with $$\beta_{1}(\omega)=\mathcal{D}(\omega)^{-1}\left\{1+2\left[B(\omega-2\Delta\omega,0)+B(\omega-2\Delta\omega,\omega-\omega_{p})+D(\omega-2\Delta\omega)\right]\left|\chi\right|^{2}\right\}\label{eq:beta1}$$ and $$\beta_{2}(\omega)=\mathcal{D}(\omega)^{-1}\left[2B(\omega,\omega-\omega_{p})+D(\omega)\right]\chi^{2}.\label{eq:beta2}$$

ArXiv

--- abstract: 'We present a distributed control law to assemble a cluster of satellites into an equally-spaced, planar constellation in a desired circular orbit about a planet. We assume each satellite only uses local information, transmitted through communication links with neighboring satellites. The same control law is used to maintain relative angular positions in the presence of disturbance forces. The stability of the constellation in the desired orbit is proved using a compositional approach. We first show the existence and uniqueness of an equilibrium of the interconnected system. We then certify each satellite and communication link is equilibrium-independent passive with respective storage functions. By leveraging the skew symmetric coupling structure of the constellation and the equilibrium-independent passivity property of each subsystem, we show that the equilibrium of the interconnected system is stable with a Lyapunov function composed of the individual subsystem storage functions. We further prove that the angular velocity of each satellite converges to the desired value necessary to maintain circular, areostationary orbit. Finally, we present simulation results to demonstrate the efficacy of the proposed control law in acquisition and station-keeping of an equally-spaced satellite constellation in areostationary orbit despite the presence of unmodeled disturbance forces.' author: - 'Emmanuel Sin, He Yin and Murat Arcak[^1] [^2] [^3]' title: '**Passivity-based distributed acquisition and station-keeping control of a satellite constellation in areostationary orbit** ' --- INTRODUCTION ============ A satellite constellation is a group of satellites that are coordinated to achieve objectives that may not be possible with a single satellite. Constellations have been applied to serve as telecommunications or broadcasting networks, provide global imagery and weather services, and enable global positioning and navigation capabilities. The control of such constellations can be divided into two different problems: *acquisition* and *station-keeping*. Acquisition refers to the process of forming the constellation once the satellites have been deployed by the delivery vehicle. For example, we may spread out a cluster of satellites in a desired orbital plane to form an equally-spaced constellation. Once the desired constellation is acquired, station-keeping refers to the process of maintaining relative positions and velocities in the presence of disturbances. The acquisition of a small spacecraft constellation in low Earth orbit, using a centralized approach, is studied in [@LP]. A centralized approach may be used if, for example, a large number of ground stations are available to measure and control the satellites. ![Depiction of constellation. Each satellite may share state information with its neighbors via communication links[]{data-label="fig:constellation"}](constellation.png){width=".48\textwidth"} In this paper, we shift our focus to a distributed approach of acquiring and station-keeping a constellation. A distributed control strategy is appealing for satellite constellations in situations where centralized control is difficult or impossible. For example, as thousands of satellites are employed in constellations, the resulting uplink/downlink demands on a network of Earth-based ground stations may become unmanageable. A distributed strategy is also critical for a constellation orbiting a planet without ground stations. Passivity-based methods are well suited for distributed control of large-scale, interconnected systems [@coord]–[@coop]. We model our constellation as an interconnected system where we assume each satellite has a communication link with neighboring satellites, sharing relative angular position information. An internal feedback control law is designed for the satellites and we certify that each satellite and communication link is equilibrium independent passive with respect to proposed storage functions. A constellation coordination control law is introduced to interconnect the subsystems in a skew-symmetric coupling structure. The equilibrium-independent passivity property of each subsystem and the skew-symmetry of their interconnection enables us to prove the stability of the constellation at equilibrium. Preliminaries ------------- We use a compositional approach to certify the stability of a large system consisting of interconnected, dissipative subsystems. We briefly state results that extend the works in [@bible], [@EID] and [@Burger], which are used in a later section to prove stability of the constellation under a closed-loop acquisition and station-keeping control law. Consider the system $\Sigma$ described by $$\begin{aligned} \label{eqn:f(x,u)} \dot{x}(t) = f(t,x(t),u(t)) \ , \quad y(t) = h(t,x(t),u(t)) \ ,\end{aligned}$$ where $x(t) \in \mathbb{R}^{n_x}$ is the state, $u(t) \in \mathbb{R}^{n_u}$ is the input, and $y(t) \in \mathbb{R}^{n_y}$ is the output. Furthermore, suppose there exists a nonempty set $\mathcal{X} \subset \mathbb{R}^{n_x}$ where, for every $\bar{x} \in \mathcal{X}$, there exists a unique $\bar{u} \in \mathbb{R}^{n_u}$ satisfying $f(t,\bar{x},\bar{u})=0$ [*Definition*]{} 1. The system (\[eqn:f(x,u)\]) is [*equilibrium independent dissipative*]{} (EID) with supply rate $s(\cdot,\cdot)$ if there exist continuously differentiable functions $V: \mathbb{R} \times \mathbb{R}^{n_x} \times \mathcal{X} \mapsto \mathbb{R}$ and ${\underaccent{\bar}{V}}: \mathbb{R}^{n_x} \times \mathcal{X} \mapsto \mathbb{R}$ satisfying the conditions \[eqn:EID\] $$\begin{aligned} &V(t,x,\bar{x}) \geq {\underaccent{\bar}{V}}(x,\bar{x}) > 0, \ \forall (x,\bar x) \ \text{s.t.} \ x \neq \bar x, \\ &V(t, \bar{x},\bar{x}) = 0, \ \ {\underaccent{\bar}{V}}(\bar{x},\bar{x}) = 0, \label{eqn:EIDa} \\ &\dot{V}(t,x,\bar{x}) := \nabla_t V(t,x,\bar{x}) + \nabla_{\scriptstyle x} V(t,x,\bar{x})^\top f(t,x,u) \nonumber \\ &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\leq s(u-\bar{u},y-\bar{y}) \label{eqn:EIDb} \ ,\end{aligned}$$ $\forall (t,x,\bar{x}, u, \bar{u}) \in \mathbb{R} \times \mathbb{R}^{n_x} \times \mathcal{X} \times \mathbb{R}^{n_u} \times \mathbb{R}^{n_u}$, where $\bar y = h(t, \bar x, \bar u)$. A system is [*equilibrium-independent passive*]{} (EIP) if it is EID with respect to the supply rate $$\begin{aligned} \label{eqn:EIP} s(u-\bar{u},y-\bar{y}) = (u-\bar{u})^\top(y-\bar{y})\end{aligned}$$ and it is [*output strictly equilibrium-independent passive*]{} (OSEIP) if, for some $\epsilon > 0$, it is EID with respect to $$\begin{aligned} \label{eqn:OSEIP} s(u-\bar{u},y-\bar{y}) = (u-\bar{u})^\top(y-\bar{y})- \epsilon (y-\bar{y})^\top(y-\bar{y}) \ .\end{aligned}$$ SYSTEM DYNAMICS =============== Instead of creating a monolithic model of the constellation, we decompose it into subsystems and consider the interconnections between them. By characterizing the input-output properties of each individual subsystem and the interconnections that exist between them, we may certify stability and convergence properties of the constellation. Satellite Model --------------- In our constellation, we refer to the constituent satellites as subsystems. Each satellite is under the influence of the gravitational pull from the central body, the thrust applied by the satellite, and natural perturbing forces (e.g., atmospheric drag, gravity from moons, solar radiation pressure). To model the motion of a satellite orbiting a planet, we start with the central-force problem (or restricted two-body problem) where we assume that the barycenter of the system is co-located with the center of a spherically, symmetric central body (i.e., the mass of the satellite is negligible). The satellite’s motion can be described by the following second-order ordinary differential equation known as the fundamental orbital differential equation (FODE) with specific force perturbations [@BMW]: $$\begin{aligned} \label{eqn:EOM} \ddot{\vec{r}} = -\frac{\mu}{\lVert\vec{r}\rVert_2^3}\vec{r} + \frac{1}{m} \vec{\tau} + \vec{a}_{perturb} \ ,\end{aligned}$$ where $\vec{r} \in \mathbb{R}^3$ is the position vector pointing from the center of the planet to the satellite, $\mu$ is the gravitational parameter of the central body (i.e., gravitational constant multiplied by the mass of the planet), $m$ is the mass of the satellite, $\vec{\tau} \in \mathbb{R}^3$ is thrust, and $\vec{a}_{perturb} \in \mathbb{R}^3$ represents the specific forces due to perturbations. It is well known that two-body motion in an inertial frame is planar. Since atmospheric drag acts against the direction of motion, a satellite under atmospheric drag remains in planar motion. Furthermore, if a satellite and the moons of a planet lie in the same plane (e.g., equatorial plane), then the gravitational perturbations from the moons may be approximated as planar. Hence, for certain examples, we may use a polar coordinate system to represent the satellite orbital kinematics in the plane: \[eqn:orbitalkinematics\] $$\begin{aligned} \vec{r} &= r\underline{e}_r \\ \dot{\vec{r}} &= \dot{r}\underline{e}_r + r\dot{\theta}\underline{e}_{\theta} \\ \ddot{\vec{r}} &= \left( \ddot{r}-r\dot{\theta}^2 \right)\underline{e}_r + \left( 2\dot{r}\dot{\theta} + r\ddot{\theta} \right)\underline{e}_{\theta} \ .\end{aligned}$$ We denote the magnitude of the radial position with $r$ and the angular position with $\theta$. We use $\underline{e}_r$ and $\underline{e}_{\theta}$ as the unit vectors in the radial and tangential directions of the orbital plane, respectively. If we include the specific forces from the right-hand side of (\[eqn:EOM\]), we get the following model representing the $i^{th}$ satellite’s motion in the radial and tangential directions, respectively: \[eqn:planarEOM\] $$\begin{aligned} \ddot{r}_i &= r_i\dot{\theta}_i^2 - \frac{\mu}{r_i^2} + \frac{1}{m_i} \tau_{r,i} + (\vec{a}_{perturb, i})_r \\ \ddot{\theta}_i &= \frac{-2\dot{r}_i\dot{\theta}_i}{r_i} + \frac{1}{m_i r_i} \tau_{\theta,i} + \frac{1}{r_i} (\vec{a}_{perturb, i})_{\theta} \ .\end{aligned}$$ Finally, if we implement a change of variables so that $v := \dot{r}$ and $\omega := \dot{\theta}$, we get the following set of first-order differential equations to describe each satellite of the constellation \[eqn:planarEOM1storder\] $$\begin{aligned} \dot{r}_i &= v_i \label{eqn:radial} \\ \dot{v}_i &= r_i \omega_i^2 - \frac{\mu}{r_i^2} + \frac{1}{m_i} \tau_{r,i} \label{eqn:radvel} \\ \dot{\omega}_i &= \frac{-2 v_i \omega_i}{r_i} + \frac{1}{m_i r_i} \tau_{\theta, i} \label{eqn:angvel} \ .\end{aligned}$$ Note that we exclude $\dot{\theta}_i = \omega_i$ from the set of equations. The $\theta$ state does not appear in the equations of motion (\[eqn:planarEOM1storder\]), hence, it is not needed in our state feedback controller design. Furthermore, we omit the terms representing specific forces due to perturbations. Through an example simulation we will show that our state feedback controller based on the model described by (\[eqn:planarEOM1storder\]) is robust to unmodeled disturbances that are present in the simulation model, described by (\[eqn:planarEOM\]). Interconnections ---------------- We assume that only neighboring satellites may communicate with each other. The topology of this particular information exchange is illustrated by the undirected graph shown in Fig \[fig:constellation\]. If the $i^{th}$ and $j^{th}$ subsystems have access to relative state information, then the $i^{th}$ and $j^{th}$ nodes of the graph are connected by a link $l=1,\ldots, M$. Although the communication is assumed to be bidirectional, we assign an orientation to the graph by considering one of the nodes of a link to be the positive end. As a convention, we set the direction of a communication link to point in the direction of the orbital motion. Hence, the incidence matrix $D$ of the graph is defined as: $$\nonumber \begin{aligned} D_{il} = \left\{ \begin{matrix} +1 & \text{if $i^{th}$ node is positive end of $l^{th}$ link} \\ -1 & \text{if $i^{th}$ node is negative end of $l^{th}$ link} \\ \phantom{\text{--}}0 & \text{otherwise. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ } \end{matrix} \right. \end{aligned}$$ In this application, for a constellation with $N$ satellites that only communicate with neighbors, the incidence matrix D is $$\begin{aligned} \label{eqn:Dmatrix} D = {\begin{bmatrix}1 & 0 & 0 \\-1 & \ddots & 0 \\ 0 & \ddots& 1 \\ 0 & 0 & -1 \end{bmatrix}} \in \mathbb{R}^{N \times M} \ ,\end{aligned}$$ where $M:= N-1$. Note that we assume the $1^{st}$ and $N^{th}$ satellites do not communicate; hence, they do not share a communication link. All other satellites have two links each. CONTROL STRATEGY ================ We now describe an internal feedback control strategy for each satellite that renders a linear map between the input (to be designed with a simple state feedback law) and the output variable of interest. Subsequently, we add a *constellation coordination* term that regulates the relative angular spacing error between neighboring satellites. Internal Feedback Control ------------------------- For each subsystem, we propose the following thrust control laws in the radial and tangential directions: \[eqn:thrustlaws\] $$\begin{aligned} \tau_{r,i} &= m_i\left(-r_i \omega_i^2 + \frac{\mu}{r_i^2}\right) - k_v (v_i - v_d) - k_r(r_i - r_d) \label{eqn:radialthrustlaw} \\ \tau_{\scriptscriptstyle \theta, i} &= m_i \left( 2 v_i \omega_i - k_{\scriptscriptstyle \omega} (\omega_i - \omega_d) + \frac{r_i}{k_c} u_i \right) \label{eqn:tangentialthrustlaw} ,\end{aligned}$$ where $r_d$, $v_d$, and $\omega_d$ are the desired radius, radial velocity, and angular velocity for every satellite to maintain an areostationary orbit. The term $u_i$ is a constellation coordination control law to be designed. The controller gains $k_r$, $k_v$, $k_\omega$, $k_c$ $> 0$ are discussed and chosen in the subsequent stability analysis and simulation results. If we substitute the thrust control laws - into the equations of motion -, the dynamics of each satellite, $\Sigma_i$ for $i=1,\ldots,N$, take the form of \[eqn:dynamicsmod\] $$\begin{aligned} \dot{r}_i &= v_i \label{eqn:radialmod} \\ \dot{v}_i &= - k_v (v_i - v_d) - k_r(r_i - r_d) \label{eqn:radvelmod} \\ \dot{\omega}_i &= - \frac{k_{\scriptscriptstyle \omega}}{r_i} (\omega_i - \omega_d) + \frac{1}{k_c}u_i \ \label{eqn:angvelmod} \\ z_i &= \omega_i \ , \label{eqn:satellite_output}\end{aligned}$$ where the output variable $z_i$ of interest is the angular velocity of the satellite. Note that we have transformed the radial dynamics - to be independent of the $\omega$ state. Constellation Coordination Control ---------------------------------- The subsystems are dynamically decoupled, however, we may coordinate their relative motion through a constellation coordination control law where we use feedback of local information from spatially neighboring subsystems. We assume that this local information is shared via inter-satellite communication links [@comm1]-[@comm2]. The links can be expressed as subsystems $\Lambda_l$ for $l=1,\ldots,M$ : \[eqn:dyn\_relang\] $$\begin{aligned} \dot{\theta}^{rel}_l &= e_l \label{eqn:link_state} \\ y_l &= h_l(\theta^{rel}_l) \label{eqn:link_output} \ ,\end{aligned}$$ where $e_l$ is the input and $y_l$ is the output of each communication link. The subsystem $\Lambda_l$ keeps track of a state $\theta^{rel}_l \in \mathbb{R}$ and outputs a signal of interest that is measured through the function $h_l: \mathbb{R} \mapsto \mathbb{R}$, that we assume is strictly increasing and onto, and $\lim_{a \rightarrow \infty}h_l(a) = \infty$. Let us refer to satellite inputs and outputs in compact form as $u := {\begin{bmatrix}u_1, \ldots, u_{N}\end{bmatrix}}^\top$ and $z := {\begin{bmatrix}z_1, \ldots, z_{N}\end{bmatrix}}^\top$, respectively. Similarly, we refer to the communication link inputs and outputs collectively as $e := {\begin{bmatrix}e_1, \ldots, e_{M}\end{bmatrix}}^\top$ and $y := {\begin{bmatrix}y_1, \ldots, y_M\end{bmatrix}}^\top$, respectively. ![Interconnected system[]{data-label="fig:system1"}](system1.png){width=".48\textwidth"} We construct an interconnection between the satellites $\Sigma_1, \ldots, \Sigma_N$ and the communication links $\Lambda_1, \ldots, \Lambda_M$ as shown in Fig \[fig:system1\] and define the following input-output mappings: \[eqn:interconnectionmappings\] $$\begin{aligned} e &:= D^\top z = {\begin{bmatrix} \omega_1-\omega_2 \\ \omega_2-\omega_3 \\ \vdots \\ \omega_{N-1} - \omega_{N} \end{bmatrix}} \equiv {\begin{bmatrix} \dot{\theta}^{rel}_1 \\ \dot{\theta}^{rel}_2 \\ \vdots \\ \dot{\theta}^{rel}_{M} \end{bmatrix}} =: \dot{\theta}^{rel} \label{eqn:errorvec} \\ u &:= -Dy = -D {\begin{bmatrix}h_1(\theta^{rel}_1)\\ \vdots \\ h_{M}(\theta^{rel}_M)\end{bmatrix}} = -D h(\theta^{rel}) \label{eqn:controlvec} \ .\end{aligned}$$ Note that the input applied to the $i^{th}$ satellite, $$\begin{aligned} \label{eqn:individualcontrol} u_i = - \sum_{l=1}^{M} D_{il} h_l(\theta^{rel}_l) \ ,\end{aligned}$$ is based only on local information since $D_{il}=0$ when the $i^{th}$ subsystem does not have access to information on the $l^{th}$ communication link. Hence, we have a distributed control architecture where local controllers act on local information. STABILITY ANALYSIS ================== We first show the existence and uniqueness of an equilibrium point whose stability will be subsequently analyzed. At equilibrium, the right-hand sides of , , for all $i=1,\ldots,N$, and for all $l=1,\ldots,M$ must equal zero. The equilibrium states of the radial dynamics – may be found by inspection to be $(\bar{r}_i,\bar{v}_i) = (r_d,v_d) = (r_d,0)$. For the right-hand side of to vanish, $e_l$ must equal zero for $l=1,\ldots M$. In other words, $$\begin{aligned} \label{eq:sol_omegabar} \bar e = D^\top \bar \omega = \mathbf{0}. \end{aligned}$$ By definition of $D$ given in , we have $D^\top \mathbf{1}=0$. Since $nullity(D^\top) = 1$, the span of $\mathbf{1}$ constitutes the entire null space of $D^\top$. Therefore, $\bar \omega = \omega_0 \mathbf{1}$ is the unique solution to , where $\omega_0$ is the common angular velocity of all $N$ satellites. That is, all satellites must have the same angular velocity. Finally, the right-hand side of must vanish: $$\begin{aligned} \label{eq:sol_ubar} - \frac{k_{\scriptscriptstyle \omega}}{r_i} (\omega_0 - \omega_d) + \frac{1}{k_c}\bar u_i = 0, \ \text{for} \ i = 1,...,N.\end{aligned}$$ From and the fact that $\mathbf{1}^\top D = \mathbf{0}^\top$, we have $\sum_{i=1}^N u_i = \mathbf{1}^\top u = -\mathbf{1}^\top D h(\theta^{rel}) = 0$. Adding from $i=1$ to $i=N$ yields the following equation: $$\begin{aligned} -(\omega_0 - \omega_d) \sum_{i=1}^N \frac{k_{\scriptscriptstyle \omega}}{r_i}= 0, \nonumber\end{aligned}$$ which requires that $\omega_0 = \omega_d$, and therefore $\bar \omega = \omega_d \mathbf{1}$. Substituting this value for $\omega_0$ back into , we get $$\begin{aligned} \bar u_i = - \sum_{l=1}^{M} D_{il} h_l(\bar \theta^{rel}_l) = 0 \ \text{for} \ i = 1,...,N,\end{aligned}$$ which amounts to $$\begin{aligned} h_1(\bar \theta^{rel}_1) &= 0, \nonumber \\ -h_{l-1}(\bar \theta^{rel}_{l-1}) + h_{l}(\bar \theta^{rel}_{l}) &= 0, \ l = 2,...,M, \label{eq:thetabar_def}\\ -h_{M}(\bar \theta^{rel}_{M}) &= 0. \nonumber\end{aligned}$$ A solution $\bar \theta^{rel}_{l}$ for $l=1,\ldots,M$ exists and is unique since $h_l$ is onto and strictly increasing. In summary, there exists a unique equilibrium point for a desired constellation given by $(\bar{r}_i, \bar{v}_i, \bar{\omega}_i) = (r_d, 0, \omega_d), \ i = 1,\ldots, N$ and $\bar{\theta}^{rel}_l, \ l = 1,\ldots,M$ that satisfy . Furthermore, we note that $\omega_d = \sqrt{\sfrac{\mu}{r_d^3}}$ for a circular orbit at a given altitude. We use a compositional approach to analyze the stability properties of the closed-loop constellation under our proposed internal feedback and coordination control laws. First, we show the stability of an equilibrium point for the radial component of each individual $\Sigma_i$ subsystem –. Second, we propose storage functions for each of the interconnected subsystems, comprised of the tangential component of the $\Sigma_i$ subsystems –, $i=1,\ldots,N$ and the $\Lambda_l$ subsystems , $l=1,\ldots,M$, and certify that they are EID as defined in . We then use the storage functions to compose a Lyapunov function for the interconnected system. For the radial component of the $\Sigma_i$ subsystem - , we choose $k_r$, $k_v$ so that the closed-loop system is stable. We define $r^e_i = r_i - \bar r_i$, and $v^e_i = v_i - \bar v_i = v_i$, then and can be rewritten as $$\begin{aligned} {\begin{bmatrix}\dot{r}^e_i \\ \dot{v}^e_i\end{bmatrix}} = {\begin{bmatrix}0 & 1 \\ - k_r & - k_v\end{bmatrix}} {\begin{bmatrix}r^e_i \\ v^e_i\end{bmatrix}} \label{eqn:radial_linear}.\end{aligned}$$ It can be verified that the equilibrium point $(\bar r_i, \bar v_i)$ of – is exponentially stable if and only if $k_r >0$ and $k_v >0$. We now proceed to prove stability of the tangential component of the subsystems under the influence of both the internal feedback law and the constellation coordination law . In the internal feedback law , we utilize a positive parameter $k_c$ to scale down the magnitude of the constellation coordination control input $u_i$. More specifically, we assume that $k_c$ is a time-varying parameter: $$\begin{aligned} \label{eqn:kc} k_c(t) \ge {\underaccent{\bar}{k}}_c > 0, \ \dot{k}_c(t) \leq 0, \ \forall \ t \ge 0,\end{aligned}$$ that decreases and converges to a positive limit ${\underaccent{\bar}{k}}_c$. We propose the following storage function for the $i^{th}$ subsystem: $$\begin{aligned} \label{eqn:storagefcn_semi_closedloop_angular_EOM} S_i(t,\omega_i,\bar{\omega}_i) = \frac{k_c(t)}{2} (\omega_i - \bar{\omega}_i)^2 \ .\end{aligned}$$ We can verify that $S_i(t,\omega_i,\bar{\omega}_i) \ge \frac{1}{2} {\underaccent{\bar}{k}}_c (\omega_i - \bar{\omega}_i)^2 > 0$, for all ($\omega_i$, $\bar{\omega}_i$) such that $\omega_i \neq \bar{\omega}_i$, and that $S_i(t,\bar{\omega}_i,\bar{\omega}_i)=0$. If we take the derivative of the storage function we get $$\begin{aligned} \label{eqn:deriv_storagefcn_semi_closedloop_angular_EOM} &\dot{S}_i(t,\omega_i,\bar{\omega}_i) = k_c(t) (\omega_i - \bar{\omega}_i) \dot{\omega}_i + \frac{\dot{k}_c(t)}{2}(\omega_i - \bar \omega_i)^2 \nonumber \\ &= k_c(t) (\omega_i - \bar{\omega}_i) \left( -\frac{k_{\scriptscriptstyle \omega}}{r_i} (\omega_i - \omega_d) + \frac{1}{k_c(t)} u_i \right) + \nonumber\\ &~~~ \frac{\dot k_c(t)}{2}(\omega_i - \bar \omega_i)^2 \\ &= (u_i - \bar{u}_i)(\omega_i - \bar{\omega}_i) - \left( \frac{k_c(t) k_{\scriptscriptstyle \omega}}{r_i} - \frac{\dot k_c(t)}{2} \right) (\omega_i - \bar{\omega}_i)^2\end{aligned}$$ where we have used $\bar \omega_i = \omega_d$, $\bar{u}_i = - \sum_{l=1}^{M} D_{il} h_l(\bar \theta^{rel}_l) = 0$. We note that $r_i(t) > 0, \ \forall i = 1,...,N$ is always satisfied (i.e., the radius is always positive). Hence, the storage function $S_i$, described by , certifies that the tangential component of the $\Sigma_i$ subsystems –, is OSEIP, as defined in . For the links $\Lambda_l$, we propose$$\begin{aligned} \label{eqn:storagefcn_links} T_l(\theta_l^{rel},\bar{\theta}_l^{rel}) = \int_{\bar{\theta}_l^{rel}}^{\theta_l^{rel}} \left( h_l(z) - h_l(\bar{\theta}_l^{rel}) \right) dz.\end{aligned}$$ Since $h_l$ is strictly increasing, we can verify that $T_l(\theta_l^{rel}, \bar{\theta}_l^{rel}) > 0$ for all $\theta_l^{rel} \neq \bar{\theta}_l^{rel}$ and $T_l(\bar{\theta}_l^{rel}, \bar{\theta}_l^{rel}) = 0$. If we take the derivative of the storage function we get $$\begin{aligned} \label{eqn:deriv_storagefcn_links} \dot{T}_l(\theta_l^{rel},\bar{\theta}_l^{rel}) &= \dot{\theta}_l^{rel} \left( h_l(\theta_l^{rel}) - h_l(\bar{\theta}_l^{rel}) \right) \nonumber \\ &= (e_l - \bar{e}_l) \left( y_l - \bar{y}_l \right) \end{aligned}$$ where we have used $\bar{e}_l = \sum_{i=1}^{N} D_{il} \bar{z}_i = \sum_{i=1}^{N} D_{il} \bar \omega_i = 0$ and $\bar{y}_l = h_l(\bar{\theta}_l^{rel})$. Since $\bar \omega$ is in the null space of $D^\top$, then $\bar e = D^\top \bar \omega$. We note that the storage function $T_l$ certifies that each communication link $\Lambda_l$ is EIP as defined in . ![Interconnected system in canonical form[]{data-label="fig:system2"}](system2.png){width=".32\textwidth"} Now that we have shown that each of the subsystems is equilibrium-independent passive, we note that the interconnected system as shown in Fig \[fig:system1\] may be brought into the canonical form of Fig \[fig:system2\] where the upper block has the subsystems along its diagonal and the lower block contains a skew symmetric matrix. As shown in [@bible], since the equilibrium-independent passive subsystems are coupled through a skew symmetric interconnection matrix, an equilibrium point of the interconnected system, if it exists, is stable and the sum of the individual subsystems provides a Lyapunov function. Let us sum the storage functions for all the $\Sigma_i$ subsystems and $\Lambda_l$ subsystems: $$\begin{aligned} \label{eq:Lyapfcn} V(t,x,\bar x) = \sum_{i=1}^{N} S_i(t,\omega_i,\bar{\omega}_i) + \sum_{l=1}^{M} T_l(\theta_l^{rel},\bar{\theta}_l^{rel}) \end{aligned}$$ where we use $x:=(\omega, \theta^{rel})$, $\bar{x}:=(\bar{\omega}, \bar{\theta}^{rel})$. The time-varying Lyapunov function can be lower and upper bounded: $$\begin{aligned} \label{eqn:bound1} {\underaccent{\bar}{V}} \left(x, \bar{x} \right) \leq V \left(t,x,\bar{x}\right) \leq \bar{V} \left(x, \bar{x}\right) \ ,\end{aligned}$$ where $$\begin{aligned} {\underaccent{\bar}{V}}(x,\bar x) = \sum_{i=1}^{N} \frac{{\underaccent{\bar}{k}}_c}{2} (\omega_i - \bar{\omega}_i)^2 + \sum_{l=1}^{M} T_l(\theta_l^{rel},\bar{\theta}_l^{rel}) \\ \bar{V}(x,\bar x) = \sum_{i=1}^{N} \frac{\bar{k}_c}{2} (\omega_i - \bar{\omega}_i)^2 + \sum_{l=1}^{M} T_l(\theta_l^{rel},\bar{\theta}_l^{rel}) \end{aligned}$$ and $\bar{k}_c := k_c(0) \ge k_c(t) \ge {\underaccent{\bar}{k}}_c \ \forall t \ge 0$. We note that ${\underaccent{\bar}{V}}(x,\bar x)$ and $\bar{V}(x,\bar x)$ are positive definite and radially unbounded. If we take the time derivative of , we get: $$\begin{aligned} &\dot{V}(t,x,\bar x) = \sum_{i=1}^{N} \dot{S}_i + \sum_{l=1}^{M} \dot{T}_l \nonumber \\ &= \sum_{i=1}^{N} \left\{ (u_i - \bar u_i) (\omega_i - \bar{\omega}_i) -\left( \frac{k_c(t) k_{\scriptscriptstyle \omega}}{r_i} - \frac{\dot k_c(t)}{2} \right) (\omega_i - \bar \omega_i)^2 \right\} \nonumber \\ &\quad + \sum_{l=1}^{M} \left\{ (e_l - \bar{e}_l) \left( y_l - \bar{y}_l \right) \right\}. \nonumber \\ \intertext{If we define $R:=blkdiag(r_1,...,r_N)$, and use $\bar e = D^\top \bar \omega$ then} &= -k_c(t) k_{\omega}(\omega - \bar{\omega})^\top R^{-1} (\omega - \bar{\omega}) + \frac{\dot k_c(t)}{2}(\omega - \bar \omega)^\top (\omega - \bar \omega) \nonumber \\ &\quad + (\omega - \bar{\omega})^\top (u - \bar u) + (\omega - \bar{\omega})^\top D \left( y - \bar{y} \right) \nonumber \\ \intertext{Finally, use our constellation coordination control law \eqref{eqn:controlvec} and $\bar u = -D\bar{y}$, then} &= -k_c(t) k_{\omega}(\omega - \bar{\omega})^\top R^{-1} (\omega - \bar{\omega}) + \frac{\dot k_c(t)}{2}(\omega - \bar \omega)^\top (\omega - \bar \omega). \label{eq:Vdot}\end{aligned}$$ Note that the expression above is negative semi-definite. As a result, $(\bar{r}_i, \bar{v}_i, \bar \omega_i, \bar \theta_l^{rel}) = (r_d, 0, \omega_d, \bar \theta_l^{rel})$, for all $\Sigma_i, \ i=1,\ldots N$ and all $\Lambda_l, \ l=1,\ldots M$ is a stable equilibrium point of the interconnected system shown in Fig \[fig:system1\], where $\bar \theta_l^{rel}$ satisfies equations . Due to the time-varying parameters $r$ and $k_c$, the interconnected constellation is a non-autonomous system for which the Lasalle-Krasovskii Invariance Principle is not applicable. Although we may not conclude asymptotic stability of an equilibrium, we may prove the weaker result [@Khalil] that $\omega_i, \ i = 1, \ldots, N$ converges to the desired $\omega_d$ value. Physically, this signifies that the constellation will maintain a circular orbit. As shown in [@Khalil], $x(t)$ is bounded by using and the dynamics are locally Lipschitz in $x$ and bounded in $t$, implying that $\dot{x}(t)$ is also bounded for all $t \geq 0$. Hence, $x(t)$ is uniformly continuous for $t \geq 0$. Define a negative semi-definite function $$\begin{aligned} W(x) = -{\underaccent{\bar}{k}}_c k_{\omega}(\omega - \bar{\omega})^\top R^{-1} (\omega - \bar{\omega}).\end{aligned}$$ As a result, $W(\cdot)$ is uniformly continuous on the bounded domain of $x(t)$. From we can verify that $$\begin{aligned} \dot{V}(t,x(t),\bar x) &\leq W(x(t)) \nonumber \\ \intertext{Integrate it over $[0,T]$, then} V(T,x(T),\bar x) - V(0,x(0),\bar x) &\leq \int_0^T W(x(t)) dt \ , \nonumber \\ \intertext{which implies} -\int_0^\infty W(x(t))dt &\leq V(0,x(0),\bar x) < \infty. \nonumber\end{aligned}$$ Using Barbalat’s Lemma, since $W \left(\cdot\right)$ is uniformly continuous and $\int_{0}^\infty W(x(t)) dt$ exists,  $W \left(x(t)\right) \rightarrow 0$ as $t \rightarrow \infty$, which implies that $x(t)$ approaches $E = \{ x: W \left(x\right) = 0 \}$. In other words, $\omega_i(t) \rightarrow \bar{\omega}_i = \omega_d$. EXAMPLE ======= Consider a cluster of $N=10$ satellites that have been batch deployed into a nearly-circular, equatorial, prograde orbit around the planet Mars at a desired altitude of approximately above the Martian surface. Assuming the equatorial radius of Mars is , each satellite in this orbit has desired equilibrium states of $(r_d,v_d,\omega_d) = (r_d,0,\sqrt{\sfrac{\mu}{r_d^3}})$ where $r_d=\SI{20428.2}{\km}$. This specific orbit, from the class of areosynchronous (i.e., Martian synchronous) orbits, is known as an areostationary orbit. Similar to satellites in geostationary orbit about Earth, the position of an areostationary satellite appears fixed in the sky relative to an observer on the surface of Mars. By equally spacing the 10 satellites within this orbit, the resulting constellation may serve as a telecommunication network or navigation system for the exploration of Mars. After deployment we assume the following initial conditions for all $i=1,\ldots,N$ satellites: $r_i = \SI[separate-uncertainty = true]{20428.0 (1)}{\km}$, $v_i = \SI[separate-uncertainty = true]{0 (1)e-08}{\meter\per\second}$, $\omega_i = \SI[separate-uncertainty = true]{7.0879 (100)e-05}{\radian\per\second}$, $\theta_i = \SI[separate-uncertainty = true]{0 (5)e-03}{\radian}$. Note that the initial conditions prescribe nearly circular orbits. The angular position $\theta_i$ is measured with respect to a reference horizontal line in the orbital plane. We assume each $m=\SI{100}{\kg}$ satellite is equipped with a throtteable, continuous-thrust propulsion system with a maximum thrust of $\tau_{max}=\SI{100}{\mN}$ in each of the radial and tangential directions of motion. In this example, we do not consider motion normal to the orbital plane. Solar electric propulsion systems, which use electricity generated by solar panels to accelerate propellant at high exhaust speeds, are capable of throtteable, continuous-thrust. Although electric propulsion systems have high specific impulse (i.e., they are fuel efficient), they have much weaker thrust compared to traditional chemical rockets. The NASA Evolutionary Xenon Thruster [@ion] is an example of a solar electric propulsion system with a maximum thrust of . We expect that the state-of-the-art will continue to develop, allowing for even higher thrust magnitudes in the future, but we maintain a conservative thrust limit for this example. In addition to the gravitational pull of Mars, we introduce perturbations due to the gravity of Mars’ two moons. Since the inclinations of Phobos and Deimos with respect to Mars’ equator are $1.093^\circ$ and $0.930^\circ$, respectively, we approximate their orbits as equatorial in this example. Note that since Phobos and Deimos have orbital eccentricities of $0.0151$ and $0.0003$, respectively, their orbits are nearly circular. We use the values of and for the radial distance of each moon’s orbit at its respective periapsis. Finally, we use values of $\mu = \SI{4.282837e+13}{\meter\cubed\per\second\squared}$, $\mu_{Phobos} = \SI{7.161e+05}{\meter\cubed\per\second\squared}$, and $\mu_{Deimos} = \SI{1.041e+05}{\meter\cubed\per\second\squared}$ for the standard gravitational parameter of Mars, Phobos, and Deimos, respectively. We find the specific force perturbation acting on each satellite by each moon, $\vec{a}_{p,i}$ (where $p=\{Phobos, Deimos\}$), by computing $$\begin{aligned} \label{eqn:accelmoon} {\begin{bmatrix}(\vec{a}_{p,i})_{r} \\ (\vec{a}_{p,i})_{\theta}\end{bmatrix}} = -\frac{\mu_{p}}{\lVert\vec{r}_{p, i}\rVert_2^3} {\begin{bmatrix}\phantom{\text{-}}\cos{\theta_i} & \sin{\theta_i} \\ \text{-}\sin{\theta_i} & \cos{\theta_i}\end{bmatrix}} \vec{r}_{p, i} \ ,\end{aligned}$$ where $\vec{r}_{p, i}$, the expression for the relative position of the $i^{th}$ satellite with respect to the moon $p$ in the Mars-centered inertial coordinate system, is \[eqn:relposmoon\] $$\begin{aligned} \vec{r}_{p, i} = {\begin{bmatrix}r_{i}\cos{\theta_i} - r_{p}\cos{\theta_{p}} \\ r_{i}\sin{\theta_i} - r_{m}\sin{\theta_{p}} \end{bmatrix}} \ .\end{aligned}$$ The radial and tangential components of the acceleration are found by rotating $\vec{r}_{p,i}$ by the appropriate rotation matrix. The *mission objectives* are (1) spread out the initial cluster of satellites into an equally-spaced constellation, and (2) regulate the satellites’ deviations from the desired areostationary orbit as well as their relative angular positions with respect to the desired spacings, in the presence of unmodeled perturbations. We call these distinct phases of the mission as *acquisition* and *station-keeping*. In the acquisition phase, we consider a generous acquisition time of $t_f = 355$ Martian days (Sols), or approximately 1 Earth year. Although the constellation may be acquired in less time, it may not be necessary. In various design proposals for manned missions to explore Mars [@vonBraun], plans include an initial uncrewed cargo mission so that supplies and infrastructure are in place before the crewed missions arrive. We assume that a satellite constellation to serve as a telecommunications network would be launched in this initial mission. Given that subsequent crewed missions would require approximately two years to arrive, due to launch window constraints, 1 Earth year would provide sufficient time to deploy and test the satellite constellation before use by a crewed mission. RESULTS ======= We implement the thrust controls laws described by (\[eqn:thrustlaws\]) where the formation control law $u_i$ for all $i=1,\ldots,N$ satellites is given by (\[eqn:individualcontrol\]) and the interconnection between satellites is described by the incidence matrix $D$ in (\[eqn:Dmatrix\]). In this example, the measurement output from each of the communication links, $h_l(\theta_l^{rel}), \ l=1,\ldots,M$, in is of the form: $$\begin{aligned} \label{eqn:h_expression} h_l(\theta^{rel}_l) = \theta^{rel}_l - \theta^{rel}_d,\end{aligned}$$ where $\theta^{rel}_d = \frac{2\pi}{N}$ represents the desired, equal angular spacing between neighboring satellites. The model (\[eqn:planarEOM\]) is used for simulation where the specific force perturbations due to Phobos and Deimos are included using (\[eqn:accelmoon\]). To regulate the radial distance, radial velocity, and angular velocity of each satellite about the areostationary orbit, we use the gains $k_r=\SI{1e-5}{}$,  $k_v=\SI{1e-4}{}$, and $k_w =\SI{1e4}{}$. In the acquisition phase $(0\leq t \leq t_f)$, we use a time-varying constellation coordination gain $$\begin{aligned} \label{eqn:exponentialkc} k_c(t) = (\bar{k}_c - {\underaccent{\bar}{k}}_c)\exp(-\tfrac{c}{t_f} t) + {\underaccent{\bar}{k}}_c \ , \end{aligned}$$ where $\bar{k}_c > {\underaccent{\bar}{k}}_c > 0$ and $c > 0$ $k_c(0) = \bar{k}_c$, and $k_c(\frac{3t_f}{c}) \approx {\underaccent{\bar}{k}}_c$ . We can simply calculate the time derivative of $k_c$ as $$\begin{aligned} \dot{k}_c(t) = -\frac{c}{t_f} (\bar{k}_c - {\underaccent{\bar}{k}}_c)\exp(-c \tfrac{t}{t_f}) < 0, \ \forall \ t \geq 0 \ .\end{aligned}$$ Note that the constellation coordination gain function, , satisfies the condition in (\[eqn:kc\]) used for the stability analysis. For this example, we choose $\bar{k}_c=\SI{1e11}{}$,  ${\underaccent{\bar}{k}}_c=\SI{1e9}{}$,  $c=30$. Since the relative angle $\theta_l^{rel}$ is far from the desired relative angle $\theta_d^{rel}$ at the beginning of the acquisition phase, the magnitude of control input $u_i$ derived with is large. We initially need a large $k_c$ to scale it down. As $\theta_l^{rel}$ converges to $\theta_d^{rel}$, the magnitude of $u_i$ decreases and we require less scaling. Therefore, the constantly decreasing parameter $k_c$ allows the thrust commands $\tau_{r, i}$ and $\tau_{\theta,i}$ in to stay within a reasonable range during the acquisition phase. After acquisition, we enter the station-keeping phase where we use a constant value of ${\underaccent{\bar}{k}}_c$. ![Absolute radial positions, radial and angular velocities, and relative angular spacing between neighboring satellites[]{data-label="fig:states"}](states.png){width=".49\textwidth"} ![Radial and tangential thrust commands to each satellite during acquisition phase[]{data-label="fig:inputs"}](inputs.png){width=".49\textwidth"} ![Orbital position of satellites during different stages of the 303.06 Sols acquisition phase[]{data-label="fig:phases"}](phases.png){width=".49\textwidth"} The simulated states of each satellite are shown in the first three subplots of Fig \[fig:states\]. Despite the perturbed initial conditions and the specific force perturbations due to Phobos and Deimos, each satellite regulates to the desired equilibrium point for an areostationary orbit (illustrated by the dotted lines). The fourth subplot of Fig \[fig:states\] shows that the angular spacing between each pair of satellites reaches the desired value of $36^{\circ}$. All angular spacings reach within a $0.5^\circ$ tolerance of the desired value in $303.06$ Sols (or approximately 311 solar Earth days). In Fig \[fig:inputs\], we plot the radial and tangential thrust inputs commanded by our feedback laws (\[eqn:thrustlaws\]). We observe that the control histories remain within the maximum thrust value of throughout the acquisition phase. We also note that, although the constellation coordination term appears in the tangential thrust control law, most of the control action occurs in the radial direction. This behavior signifies that the $\omega_i^2$ term in the radial thrust law (\[eqn:radialthrustlaw\]) dominates the other terms. The controller exhibits the same strategy as traditional station-keeping methods where orbital phasing maneuvers (i.e., adjusting a satellite’s position within an orbit) can be conducted by decreasing (increasing) the altitude of a spacecraft, causing it to speed up (slown down) in the tangential direction to gain (reduce) angular position. Finally, we present Fig \[fig:phases\], where the angular positions of the satellites are depicted at different times during the acquisition phase. The central red body represents Mars whereas the two gray bodies are the moons, Phobos and Deimos. We note that the orbit of the outer moon, Deimos, is very close to that of the areostationary orbit at a distance of approximately . Despite the close proximity, the effect of the unmodeled gravitational perturbation is mitigated by the proposed control law. An animation of the acquisition phase is available at <https://youtu.be/-2y_IWRPuzU>. CONCLUSION ========== We have presented a control strategy to coordinate a large number of satellites to not only acquire but also to maintain an equally-spaced constellation in areostationary orbit. The proposed distributed control law is implemented on each satellite using only local information from neighboring satellites. We proved that the closed-loop system, comprised of the satellites and communication links, is stable at equilibrium due to the equilibrium-independent passive property of each subsystem and the skew-symmetric coupling structure of their interconnections. We further proved that the angular velocities of each satellite converge to the desired value necessary for a circular, areostationary orbit. We then demonstrated the efficacy of the acquisition and station-keeping control strategy on a simulation example. Regarding the practical implementation of our approach to constellation acquisition and station-keeping, we note that although the proposed control strategy is not optimal (with respect to a minimum-acquisition-time or minimum-fuel objective), it is a simple, distributed, and computationally inexpensive approach that may be tuned to achieve specific mission constraints on time or fuel. Given the time and maximum thrust constraints of our example mission, our simulation results showed that the commanded thrust profiles are achievable with the current state-of-the-art in electric propulsion. We also note that the proposed strategy exhibits robustness to perturbed initial conditions and unmodeled disturbances. Future work will investigate delay robustness although we do not deem the communication delay between satellites to be significant relative to the slow time scales in which the constellation evolves in our example. If we assume that communication delay is proportional to inter-satellite link distance, the worst delay is when the areostationary constellation is completely acquired and the 10 satellites are equally spaced with a line-of-sight distance of between each pair. Considering that the delay between a ground station and a geostationary satellite at an altitude of is approximately a quarter of a second, we can deduce that the communication delay between our satellites will be relatively small compared to the time it takes a circular, areostationary orbit to be influenced by low-thrust propulsion or the time we allow for the acquisition phase. [99]{} E. Sin, M. Arcak, A. Packard, “Small Satellite Constellation Separation using Linear Programming based Differential Drag Commands,” in *Proceedings of the 2018 American Control Conference*, Milwaukee, WI, USA, August 2018. M. Arcak, “Passivity as a Design Tool for Group Coordination,” in [*IEEE Transactions on Automatic Control*]{}, vol. 52, no. 8, pp. 1380-1390, 2007. H. Bai, M. Arcak, and J. Wen, “Cooperative Control Design: A Systematic, Passivity-Based Approach,” [*Springer Communications and Control Engineering Series*]{}, New York, NY, USA 2011. C. Meissen, K. Klausen, M. Arcak, T.I. Fossen, A. Packard, “Passivity-based Formation Control for UAVs with a Suspended Load,” [*IFAC-PapersOnLine*]{}, vol. 50, no. 1, pp. 13150–13155, 2017. M. Arcak, C. Meissen, A. Packard, “Networks of Dissipative System: Compositional Certification of Stability, Performance, and Safety,” [*Springer Briefs in Control, Automation and Robotics*]{}, Switzerland, 2016. G. Hines, M. Arcak, A. Packard, “Equilibrium-independent passivity: a new definition and numerical certification,” [*in Automatica*]{}, vol. 47, no. 9, pp. 1949-1956, 2011. M. Bürger and D. Zelazo and F. Allgöwer, “Duality and network theory in passivity-based cooperative control,” [*in Automatica*]{}, vol. 50, no.8, pp. 2051-2061, 2014. R.R. Bate, D.D. Mueller, J.E. White, “Fundamentals of Astrodynamics,” Dover, New York, NY, USA, 1971. R. Radhakrishnan, W. W. Edmonson, F. Afghah, R. M. Rodriguez-Osorio, F. Pinto and S. C. Burleigh, “Survey of Inter-Satellite Communication for Small Satellite Systems: Physical Layer to Network Layer View,” in [*IEEE Communications Surveys & Tutorials*]{}, vol. 18, no. 4, pp. 2442-2473, 2016. National Aeronautics and Space Administration, “Communications- State of the Art of Small Spacecraft Technology,” <https://sst-soa.arc.nasa.gov/09-communications> H.K. Khalil, “Nonlinear Systems,” Prentice Hall, Upper Saddle River, NJ, USA, 2002. National Aeronautics and Space Administration, “NASA”s Evolutionary Xenon Thruster (NEXT),” <https://www1.grc.nasa.gov/space/sep/gridded-ion-thrusters-next-c/> W. von Braun, “Manned Mars Landing Presentation to the Space Task Group,” <https://www.nasa.gov/sites/default/files/atoms/files/19690804_manned_mars_landing_presentation_to_the_space_task_group_by_dr._wernher_von_braun.pdf>, 1969. [^1]: The first two authors contributed equally to this work. [^2]: E. Sin and H. Yin are Graduate Students of the Department of Mechanical Engineering at the University of California, Berkeley {[emansin,he\_yin }@berkeley.edu]{}. [^3]: M. Arcak is a Professor of the Department of Electrical Engineering & Computer Sciences at the University of California, Berkeley, [arcak@berkeley.edu]{}.

ArXiv

--- abstract: | Considering gravitino dark matter scenarios with a long-lived charged slepton, we show that collider measurements of the slepton mass and its lifetime can probe not only the gravitino mass but also the post-inflationary reheating temperature ${T_{{\mathrm{R}}}}$. In a model independent way, we derive upper limits on ${T_{{\mathrm{R}}}}$ and discuss them in light of the constraints from the primordial catalysis of $^6$Li through bound-state effects. In the collider-friendly region of slepton masses below 1 TeV, the obtained conservative estimate of the maximum reheating temperature is about ${T_{{\mathrm{R}}}}=3\times 10^9\,{\mathrm{GeV}}$ for the limiting case of a small gluino–slepton mass splitting and about ${T_{{\mathrm{R}}}}=10^8\,{\mathrm{GeV}}$ for the case that is typical for universal soft supersymmetry breaking parameters at the scale of grand unification. We find that a determination of the gluino–slepton mass ratio at the Large Hadron Collider will test the possibility of ${T_{{\mathrm{R}}}}> 10^9\,{\mathrm{GeV}}$ and thereby the viability of thermal leptogenesis with hierarchical heavy right–handed Majorana neutrinos. author: - Frank Daniel Steffen title: Probing the Reheating Temperature at Colliders and with Primordial Nucleosynthesis --- Introduction ============ Big bang nucleosynthesis (BBN) is a powerful tool to test physics beyond the Standard Model; cf. [@Dimopoulos:1988ue; @Jedamzik:2004er; @Kawasaki:2004yh; @Kawasaki:2004qu; @Kohri:2005wn; @Jedamzik:2006xz; @Kawasaki:2008qe] and references therein. Indeed, in supersymmetric (SUSY) theories, severe constraints appear owing to the existence of the gravitino ${{\widetilde{G}}}$ which is the gauge field of local SUSY transformations and whose mass is governed by the SUSY breaking scale. As the spin-3/2 superpartner of the graviton, the gravitino is an extremely weakly interacting particle with couplings suppressed by inverse powers of the (reduced) Planck scale ${\mathrm{M}_{\mathrm{P}}}=2.4\times 10^{18}\,{\mathrm{GeV}}$ [@Wess:1992cp]. Accordingly, once produced in thermal scattering of particles in the hot primordial plasma [@Ellis:1984eq; @Moroi:1993mb; @Ellis:1995mr; @Bolz:1998ek; @Bolz:2000fu; @Pradler:2006qh; @Pradler:2006hh; @Rychkov:2007uq], unstable gravitinos with a mass ${m_{\widetilde{G}}}{\buildrel<\over{_\sim}}5~{\mathrm{TeV}}$ have long lifetimes, $\tau_{{{\widetilde{G}}}} {\buildrel>\over{_\sim}}100~{\mathrm{s}}$, and decay during or after BBN. Since the decay products affect the abundances of the primordial light elements, successful BBN predictions imply a bound on the reheating temperature after inflation $T_{{\mathrm{R}}}$ which governs the abundance of gravitinos before their decay [@Ellis:1984eq; @Moroi:1993mb; @Ellis:1995mr; @Bolz:1998ek; @Bolz:2000fu; @Pradler:2006qh; @Pradler:2006hh; @Pradler:2007ne; @Rychkov:2007uq]: $T_{{\mathrm{R}}}{\buildrel<\over{_\sim}}10^8\,{\mathrm{GeV}}$ for ${m_{\widetilde{G}}}{\buildrel<\over{_\sim}}5~{\mathrm{TeV}}$ [@Kohri:2005wn; @Kawasaki:2008qe]. We consider SUSY extensions of the Standard Model in which the gravitino ${{\widetilde{G}}}$ is the lightest supersymmetric particle (LSP) and a charged slepton ${{\widetilde{l}_1}}$—such as the lighter stau ${{\widetilde{\tau}_1}}$—the next-to-lightest supersymmetric particle (NLSP). Assuming R-parity conservation, the gravitino LSP is stable and a promising candidate for dark matter; cf. [@Moroi:1993mb; @Bolz:1998ek; @Asaka:2000zh; @Bolz:2000fu; @Ellis:2003dn; @Feng:2004mt; @Cerdeno:2005eu; @Steffen:2006hw; @Pradler:2006qh; @Pradler:2006hh; @Pradler:2007ne; @Rychkov:2007uq; @Steffen:2007sp] and references therein. Because of the extremely weak interactions of the gravitino, the NLSP typically has a long lifetime before it decays into the gravitino. If these decays occur during or after BBN, the Standard Model particles emitted in addition to the gravitino can affect the abundances of the primordial light elements. For the charged slepton NLSP case, the BBN constraints associated with hadronic/electromagnetic energy injection have been estimated [@Ellis:2003dn; @Feng:2004mt; @Cerdeno:2005eu; @Steffen:2006hw]. Taking into account additional constraints from large-scale-structure formation, apparently viable gravitino dark matter scenarios had been identified (see, e.g., benchmark scenarios $A_{1,2}$–$C_{1,2}$ in Ref. [@Steffen:2006hw]) which are attractive for two reasons: (i) $T_{{\mathrm{R}}} > 10^9\,{\mathrm{GeV}}$ is possible so that thermal leptogenesis with hierarchical right-handed neutrinos [@Fukugita:1986hr; @Davidson:2002qv; @Buchmuller:2004nz; @Blanchet:2006be; @Antusch:2006gy][^1] remains a viable explanation of the baryon asymmetry [@Bolz:1998ek; @Fujii:2003nr; @Cerdeno:2005eu; @Steffen:2006hw; @Pradler:2006qh] and (ii) the gravitino mass can be close to the NLSP mass ${m_{{{\widetilde{l}_1}}}}$, i.e., $0.1\,{m_{{{\widetilde{l}_1}}}}\lesssim{m_{\widetilde{G}}}<{m_{{{\widetilde{l}_1}}}}$, so that a kinematical ${m_{\widetilde{G}}}$ determination appears viable [@Buchmuller:2004rq; @Martyn:2006as; @Hamaguchi:2006vu]. With a kinematically determined ${m_{\widetilde{G}}}$, one would be able to measure the Planck scale ${\mathrm{M}_{\mathrm{P}}}$ at colliders [@Buchmuller:2004rq; @Martyn:2006as; @Hamaguchi:2006vu] and to test the viability of thermal leptogenesis in the laboratory [@Pradler:2006qh]. Indeed, an agreement of the ${\mathrm{M}_{\mathrm{P}}}$ value determined in collider experiments with the one inferred from Newton’s constant [@Yao:2006px] $G_{\rm N} = 6.709\times 10^{-39}\,{\mathrm{GeV}}^{-2}$ would provide evidence for the existence of supergravity (SUGRA) in nature [@Buchmuller:2004rq]. Already in the early paper [@Dimopoulos:1989hk] it had been realized that long-lived negatively charged massive particles ${X^{\! -}}$ (such as long-lived ${{\widetilde{l}_1}}^-$’s) can form primordial bound states, which can affect the abundances of the primordial light elements. It was however only recently [@Pospelov:2006sc] when it was realized that bound-state formation of ${X^{\! -}}$ with ${{}^4 \mathrm{He}}$ can lead to a substantial overproduction of primordial ${{}^6 \mathrm{Li}}$ via the catalyzed BBN (CBBN) reaction $$({{}^4 \mathrm{He}}{X^{\! -}})+{\mathrm{D}}\rightarrow {{}^6 \mathrm{Li}}+ {X^{\! -}}\, . \label{Eq:CBBNof6Li}$$ With an ${X^{\! -}}$ abundance that is typical for an electrically charged massive thermal relic [@Asaka:2000zh], this reaction becomes so efficient that an ${X^{\! -}}$ lifetime of $\tau_{{X^{\! -}}} \gtrsim 5\times 10^3~{\mathrm{s}}$ is excluded by observationally inferred values of the primordial ${{}^6 \mathrm{Li}}$ abundance [@Pospelov:2006sc]. In the considered gravitino LSP scenarios, this bound applies directly to the lifetime of the ${{\widetilde{l}_1}}$ NLSP [@Pospelov:2006sc; @Cyburt:2006uv; @Steffen:2006wx; @Pradler:2006hh; @Hamaguchi:2007mp; @Kawasaki:2007xb; @Pradler:2007is; @Pradler:2007ar; @Steffen:2007sp; @Kawasaki:2008qe]: $\tau_{{{\widetilde{l}_1}}}\lesssim 5 \times 10^3~{\mathrm{s}}$.[^2] This implies ${m_{\widetilde{G}}}\lesssim 0.1\,{m_{{{\widetilde{l}_1}}}}$ in the collider-friendly mass range of ${m_{{{\widetilde{l}_1}}}}\lesssim 1~{\mathrm{TeV}}$. Accordingly, the region $0.1\,{m_{{{\widetilde{l}_1}}}}\lesssim{m_{\widetilde{G}}}<{m_{{{\widetilde{l}_1}}}}$, in which the kinematical ${m_{\widetilde{G}}}$ determination appears feasible, seems to be excluded by BBN constraints [@Steffen:2006wx]. Moreover, within the framework of the constrained minimal supersymmetric Standard Model (CMSSM), we have found that the ${{}^6 \mathrm{Li}}$ constraint implies the upper limit ${T_{{\mathrm{R}}}}\lesssim 10^7~{\mathrm{GeV}}$ in gravitino dark matter scenarios with unbroken R-parity and typical thermal ${{\widetilde{\tau}_1}}$ NLSP relic abundances [@Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar]. For a standard cosmological history, this finding clearly disfavors successful thermal leptogenesis within the CMSSM in the case of hierarchical heavy right-handed Majorana neutrinos. For a gravitino LSP mass range that is natural for gravity-mediated SUSY breaking, the ${{}^6 \mathrm{Li}}$ constraint can even point to a CMSSM mass spectrum which will be difficult to probe at the Large Hadron Collider (LHC) [@Cyburt:2006uv; @Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar]. This letter provides a model independent study of gravitino LSP scenarios with a charged slepton NLSP that has a $^6$Li-friendly lifetime of $\tau_{{{\widetilde{l}_1}}}\lesssim 10^4~{\mathrm{s}}$ and a collider-friendly mass of ${m_{{{\widetilde{l}_1}}}}\lesssim 1~{\mathrm{TeV}}$. In contrast to [@Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar; @Kersten:2007ab], the investigation presented in this work is not restricted to a constrained framework such as the CMSSM or to the gravitino mass range suggested by gravity-mediated SUSY breaking. Thereby, generic results are obtained with a range of validity that includes models with gauge-mediated SUSY breaking and/or non-standard mass spectra. The remainder of this letter is organized as follows. In the next section we review that ${m_{\widetilde{G}}}$ could be determined by measuring ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$ at colliders. We give also an upper limit on ${m_{\widetilde{G}}}$ that depends on ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$. This limit allows us to derive a lower limit for the gravitino density ${\Omega_{\widetilde{G}}}$. Since ${\Omega_{\widetilde{G}}}$ cannot exceed the dark matter density ${\Omega_{\mathrm{DM}}}$, this leads to conservative ${T_{{\mathrm{R}}}}$ limits which can be probed in measurements of ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$ (and the gluino mass $m_{{{\widetilde g}}}$). We use the derived expressions to translate the $\tau_{{{\widetilde{l}_1}}}$ constraint from CBBN of $^6$Li into robust upper limits on ${T_{{\mathrm{R}}}}$ that depend on the mass ratio $m_{{{\widetilde g}}}/{m_{{{\widetilde{l}_1}}}}$. Finally, we show that the requirement ${T_{{\mathrm{R}}}}> 10^9\,{\mathrm{GeV}}$ needed for successful standard thermal leptogenesis provides an upper limit on this ratio and thereby a testable prediction for LHC phenomenology. Probing ${m_{\widetilde{G}}}$ at Colliders {#Sec:mGravitinoCollider} ========================================== In the considered SUSY scenarios with the gravitino LSP being stable due to R-parity conservation,[^3] the charged slepton NLSP has a lifetime $\tau_{{{\widetilde{l}_1}}}$ that is governed by the decay ${{\widetilde{l}_1}}\to{{\widetilde{G}}}\tau$ and thus given by the SUGRA prediction for the associated partial width $$\tau_{{{\widetilde{l}_1}}} \simeq \frac{48 \pi {m_{\widetilde{G}}}^2 {\mathrm{M}_{\mathrm{P}}}^2}{{m_{{{\widetilde{l}_1}}}}^5} \left(1-\frac{{m_{\widetilde{G}}}^2}{{m_{{{\widetilde{l}_1}}}}^2}\right)^{-4} \gtrsim \frac{48 \pi {m_{\widetilde{G}}}^2 {\mathrm{M}_{\mathrm{P}}}^2}{{m_{{{\widetilde{l}_1}}}}^5} \, , \label{Eq:SleptonLifetime}$$ where the rightmost term underestimates $\tau_{{{\widetilde{l}_1}}}$ by at most 5% (30%) for ${m_{\widetilde{G}}}\lesssim 0.1\,{m_{{{\widetilde{l}_1}}}}$ (${m_{\widetilde{G}}}\lesssim 0.25\,{m_{{{\widetilde{l}_1}}}}$). Accordingly, using ${\mathrm{M}_{\mathrm{P}}}=2.4\times 10^{18}\,{\mathrm{GeV}}$ as inferred from Newton’s constant, the gravitino mass ${m_{\widetilde{G}}}$ can be determined by measuring both ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$ at future colliders [@Ambrosanio:2000ik]. Moreover, from the rightmost expression in (\[Eq:SleptonLifetime\]), one can extract an upper limit for the gravitino mass $${m_{\widetilde{G}}}\lesssim 0.41~{\mathrm{GeV}}\left(\frac{\tau_{{{\widetilde{l}_1}}}}{10^4~{\mathrm{s}}}\right)^{\!\!\frac{1}{2}}\!\! \left(\frac{{m_{{{\widetilde{l}_1}}}}}{100~{\mathrm{GeV}}}\right)^{\!\!\frac{5}{2}}\!\! \equiv {m_{{{\widetilde{G}}}}^{\max}}\, , \label{Eq:mGmax}$$ which turns into an equality for ${m_{\widetilde{G}}}\ll{m_{{{\widetilde{l}_1}}}}$. In Fig. \[Fig:ProbingMgravitino\] ![Contours of ${m_{\widetilde{G}}}^{(\max)}$ (solid lines) in the plane spanned by $\tau_{{{\widetilde{l}_1}}}$ and ${m_{{{\widetilde{l}_1}}}}$. BBN constraints from D and $^6$Li are indicated by the short-dash-dotted (blue in the web version) and by the long-dash-dotted (red in the web version) lines, respectively. The $^6$Li constraints are obtained from the CBBN treatment of [@Pradler:2007is] for upper limits on the primordial $^6\mathrm{Li/H}$ abundance of $2\times 10^{-11}$ (left line) and $6\times 10^{-11}$ (right line).[]{data-label="Fig:ProbingMgravitino"}](ProbingMgravitino_ite.eps){width="3.25in"} contours of ${m_{\widetilde{G}}}^{(\max)}$ (solid lines) are shown in the plane spanned by $\tau_{{{\widetilde{l}_1}}}$ and ${m_{{{\widetilde{l}_1}}}}$. Only $m_{{{\widetilde{l}_1}}} {\gtrsim}100\,{\mathrm{GeV}}$ is considered since long-lived charged sleptons should have otherwise been observed already at the Large Electron–Positron Collider (LEP) [@Yao:2006px]. The potential for collider measurements of ${m_{{{\widetilde{l}_1}}}}$ is promising since each heavier superpartner produced will cascade down to the ${{\widetilde{l}_1}}$ NLSP which will appear as a (quasi-) stable muon-like particle in the detector [@Drees:1990yw; @Nisati:1997gb; @Feng:1997zr]. For slow ${{\widetilde{l}_1}}$’s, the associated highly ionizing tracks and time–of–flight measurements can allow for a distinction from muons [@Drees:1990yw; @Nisati:1997gb; @Feng:1997zr; @Ambrosanio:2000ik]. With measurements of the ${{\widetilde{l}_1}}$ velocity $\beta_{{{\widetilde{l}_1}}} \equiv v_{{{\widetilde{l}_1}}}/c$ and its momentum $p_{{{\widetilde{l}_1}}}\equiv |\vec{p}_{{{\widetilde{l}_1}}}|$, ${m_{{{\widetilde{l}_1}}}}$ can be determined: ${m_{{{\widetilde{l}_1}}}}=p_{{{\widetilde{l}_1}}}(1-\beta_{{{\widetilde{l}_1}}}^2)^{1/2}/\beta_{{{\widetilde{l}_1}}}$ [@Ambrosanio:2000ik]. For the upcoming LHC experiments, studies of hypothetical scenarios with long-lived charged particles are actively pursued [@Ellis:2006vu; @Bressler:2007gk; @Zalewski:2007up]. In Ref. [@Ellis:2006vu], for example, it is shown that one should be able to measure the mass ${m_{{{\widetilde{l}_1}}}}$ of a (quasi-) stable slepton quite accurately at the LHC. The experimental determination of $\tau_{{{\widetilde{l}_1}}}$ will be substantially more difficult than the ${m_{{{\widetilde{l}_1}}}}$ measurement. If some of the sleptons decay already in the collider detectors, the statistical method proposed in [@Ambrosanio:2000ik] could allow one to measure $\tau_{{{\widetilde{l}_1}}}$. Moreover, ways to stop and collect charged long-lived particles for an analysis of their decays have been proposed for the LHC and for the International Linear Collider (ILC) [@Goity:1993ih; @Buchmuller:2004rq; @Hamaguchi:2004df; @Feng:2004yi; @DeRoeck:2005bw; @Martyn:2006as; @Hamaguchi:2006vu; @Cakir:2007xa; @Martyn:2007mj]. These challenging proposals could lead to a precise measurement of $\tau_{{{\widetilde{l}_1}}}$. In addition, these proposals could help to distinguish the case of the gravitino LSP from the one of the axino LSP [@Brandenburg:2005he+X; @Hamaguchi:2006vu], which is—as the fermionic superpartner of the axion—another well-motivated dark matter candidate [@Covi:1999ty; @Steffen:2007sp]. Indeed, also the axino LSP can be produced thermally in the early Universe [@Covi:2001nw; @Brandenburg:2004du] and can be associated with a ${{\widetilde{l}_1}}$ NLSP with ${{\cal O}}(\mathrm{ms})\lesssim\tau_{{{\widetilde{l}_1}}}\lesssim{{\cal O}}(\mathrm{days})$ [@Covi:2004rb; @Brandenburg:2005he+X]. In the axino LSP case, however, measurements of $\tau_{{{\widetilde{l}_1}}}$ and ${m_{{{\widetilde{l}_1}}}}$ will probe the Peccei–Quinn scale $f_a$ [@Brandenburg:2005he+X] instead of ${m_{\widetilde{G}}}$. Keeping in mind that the axino LSP could mimic the gravitino LSP at colliders (at first sight), we assume in the remainder of this work that it is the gravitino LSP scenario that is realized in nature. Probing ${T_{{\mathrm{R}}}}$ at Colliders ========================================= Within a standard cosmological history, the gravitino LSP can be produced in decays of scalar fields such as the inflaton [@Asaka:2006bv; @Takahashi:2007tz; @Endo:2007sz], in thermal scattering of particles in the primordial plasma [@Moroi:1993mb; @Bolz:1998ek; @Bolz:2000fu; @Pradler:2006qh; @Pradler:2006hh; @Rychkov:2007uq], and in NLSP decays [@Borgani:1996ag; @Asaka:2000zh; @Ellis:2003dn; @Feng:2004mt]. Focussing on the case ${m_{\widetilde{G}}}\gg 100~{\mathrm{eV}}$ in which gravitinos are never in thermal equilibrium due to their extremely weak interactions,[^4] the thermally produced gravitino density ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ depends basically linearly on ${T_{{\mathrm{R}}}}$—cf. (\[Eq:OmegaGrTPSU3\]) below—and thus can serve as a thermometer of the earliest moments of the radiation-dominated epoch [@Moroi:1993mb; @Bolz:1998ek; @Bolz:2000fu; @Pradler:2006qh; @Pradler:2006hh; @Rychkov:2007uq]. In particular, this allows for the derivation of upper limits on ${T_{{\mathrm{R}}}}$ since the relic gravitino density is bounded from above by the observed dark matter density [@Yao:2006px]: $${\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 \lesssim {\Omega_{\widetilde{G}}}h^2\leq {\Omega_{\mathrm{DM}}}h^2\simeq 0.1 \, , \label{Eq:OmegaDM}$$ where $h\simeq 0.7$ denotes the Hubble constant in units of $100~{\mathrm{km}}\,{\mathrm{Mpc}}^{-1}{\mathrm{s}}^{-1}$. Aiming at a lower limit of ${\Omega_{\widetilde{G}}}$ to arrive at a truly conservative upper limit on ${T_{{\mathrm{R}}}}$, we do not take into account the model dependent contributions from decays of scalar fields such as the inflaton [@Asaka:2006bv; @Takahashi:2007tz; @Endo:2007sz] or the ones from NLSP decays [@Borgani:1996ag; @Asaka:2000zh; @Feng:2004mt],[^5] which can be negligible anyhow, in particular, for ${m_{\widetilde{G}}}\lesssim 1~{\mathrm{GeV}}$. Indeed, we focus on the SUSY QCD contribution to the thermally produced gravitino density, as derived in the gauge-invariant calculation of [@Bolz:2000fu; @Pradler:2006qh], $$\begin{aligned} {\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c} &\!\!\!\!=\!\!& 0.117\,g_s^2({T_{{\mathrm{R}}}})\left[1+\frac{m_{\tilde{g}}^2({T_{{\mathrm{R}}}})}{3{m_{\widetilde{G}}}^2}\right] \ln\left[\frac{1.271}{g_s({T_{{\mathrm{R}}}})}\right] \nonumber\\ && \times \left(\frac{{m_{\widetilde{G}}}}{100~{\mathrm{GeV}}}\right) \left(\frac{T_{{\mathrm{R}}}}{10^{10}\,{\mathrm{GeV}}}\right) \label{Eq:OmegaGrTPSU3}\end{aligned}$$ with the strong coupling $g_s$ and the gluino mass $m_{{{\widetilde g}}}$ to be evaluated at the scale given by ${T_{{\mathrm{R}}}}$, i.e., $g_s(T_R)=[g_s^{-2}(M_Z)\!+\!3\ln(T_R/M_Z)/(8\pi^2)]^{-1/2}$ and $m_{{{\widetilde g}}}(T_R)=[g_s(T_R)/g_s(M_Z)]^2 m_{{{\widetilde g}}}(M_Z)$, where $g_s^2(M_Z)/(4\pi)=0.1172$ at $M_Z=91.188~{\mathrm{GeV}}$. By discarding the electroweak contributions, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ is underestimated typically by 20%–50% depending on the size of the gaugino masses in the electroweak sector [@Pradler:2006qh; @Pradler:2006hh; @Pradler:2007ne]. Thus, these additional contributions would tighten the ${T_{{\mathrm{R}}}}$ limits derived below only by a factor that is not smaller than about 2/3. For high-temperatures, $10^6\,{\mathrm{GeV}}\lesssim{T_{{\mathrm{R}}}}\lesssim 10^{10}\,{\mathrm{GeV}}$, where (\[Eq:OmegaGrTPSU3\]) derived in the weak coupling limit $g_s \ll 1$ is most reliable [@Bolz:2000fu; @Pradler:2006qh], we do now derive a lower limit $${\Omega_{\widetilde{G}}}^{\min} h^2 \lesssim {\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c} \label{Eq:OmegaGconsVSOmegaGrTPSU3}$$ by manipulating expression (\[Eq:OmegaGrTPSU3\]) as follows: - We use the replacement $$\ln\left[\frac{1.271}{g_s({T_{{\mathrm{R}}}})}\right] \to \ln\left[\frac{1.271}{g_s(10^6\,{\mathrm{GeV}})}\right] = 0.256 \, .$$ Thereby, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c}$ is underestimated at most at ${T_{{\mathrm{R}}}}=10^{10}\,{\mathrm{GeV}}$ and there by a factor of about 5/8. - We use $g_s(10^{10}\,{\mathrm{GeV}})=0.85$ in the numerator of (\[Eq:OmegaGrTPSU3\]) and to evolve $m_{{{\widetilde g}}}({T_{{\mathrm{R}}}})$ to the weak scale. Thereby, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c}$ is underestimated at most at ${T_{{\mathrm{R}}}}=10^{6}\,{\mathrm{GeV}}$ and there by a factor of about 2/5. - We use the constant $c>1$ to parametrize $m_{{{\widetilde g}}}(M_Z)$ in terms of the mass of the slepton NLSP: $$m_{{{\widetilde g}}}(M_Z) = c\, m_{{{\widetilde{l}_1}}} . \label{Eq:mgluinomslepton}$$ For concreteness, we use $m_{{{\widetilde g}}}(M_Z)$ to represent the gluino mass at the weak scale. While the running mass $m_{{{\widetilde g}}}$ decreases when evolved to higher energy scales, we assume implicitly $m_{{{\widetilde g}}}>m_{{{\widetilde{l}_1}}}$ (even in the limiting case $c\simeq 1$) at least up to energy scales accessible at the LHC and thereby up to temperatures well above the freeze-out temperature of the ${{\widetilde{l}_1}}$ NLSP. For example, since $m_{{{\widetilde g}}}$ decreases by about 20% when evolved from $M_Z$ to $10~{\mathrm{TeV}}$, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c}$ can thereby be underestimated for $c\simeq 1$ by a factor of about $(4/5)^2$. - We drop the contributions from the spin 3/2 components of the gravitino which are given by the term in (\[Eq:OmegaGrTPSU3\]) that is independent of $m_{\tilde{g}}$. Focussing on ${m_{\widetilde{G}}}\lesssim 0.1\,{m_{{{\widetilde{l}_1}}}}$, the relative importance of the spin 1/2 components is minimal for ${m_{\widetilde{G}}}\simeq 0.1\,{m_{{{\widetilde{l}_1}}}}$, $c\simeq 1$, and ${T_{{\mathrm{R}}}}=10^{10}\,{\mathrm{GeV}}$, where the associated second term in the first bracket of (\[Eq:OmegaGrTPSU3\]) becomes about 8 so that ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}h^2 |_{\mathrm{SU(3)}_c}$ is underestimated without the spin 3/2 term by a factor that is not smaller than about 8/9. Accordingly, we find a guaranteed gravitino density of $${\Omega_{\widetilde{G}}}^{\min} h^2 = 0.174 \left(\frac{1~{\mathrm{GeV}}}{{m_{\widetilde{G}}}}\right)\!\! \left(\frac{c\,{m_{{{\widetilde{l}_1}}}}}{100~{\mathrm{GeV}}}\right)^{\!\!2}\!\! \left(\frac{{T_{{\mathrm{R}}}}}{10^{10}\,{\mathrm{GeV}}}\right) . \label{Eq:OmegaGmin}$$ For ${m_{\widetilde{G}}}\lesssim 0.1\,{m_{{{\widetilde{l}_1}}}}$, $10^8\,{\mathrm{GeV}}\,(10^9\,{\mathrm{GeV}})\,\leq{T_{{\mathrm{R}}}}\leq 10^{10}\,{\mathrm{GeV}}$ and $c\,{m_{{{\widetilde{l}_1}}}}>200~{\mathrm{GeV}}~(100~{\mathrm{GeV}})$, this expression is associated with $1.57 \lesssim {\Omega_{\widetilde{G}}^{\mathrm{TP}}}|_{\mathrm{SU(3)}_c}/{\Omega_{\widetilde{G}}}^{\min}\lesssim 2.1$ ($1.9$). Accordingly, the limits derived below may be considered to be ‘too relaxed’ (or ‘too conservative’) by at least a factor of about 1.5. Indeed, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}|_{\mathrm{SU(3)}_c}/{\Omega_{\widetilde{G}}}^{\min}$ increases by decreasing ${T_{{\mathrm{R}}}}$ and/or $c\,{m_{{{\widetilde{l}_1}}}}$ or by increasing ${m_{\widetilde{G}}}$ so that the corresponding limits presented below will become even more conservative. For example, for ${T_{{\mathrm{R}}}}=10^6\,{\mathrm{GeV}}$, $c\,{m_{{{\widetilde{l}_1}}}}\simeq 100\,{\mathrm{GeV}}$, and ${m_{\widetilde{G}}}\lesssim 0.05\,{m_{{{\widetilde{l}_1}}}}$, ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}|_{\mathrm{SU(3)}_c}/{\Omega_{\widetilde{G}}}^{\min}\simeq 2.4$ is encountered so that the associated limits derived below may be considered as ‘too relaxed’ by at least a factor of about 2.4. With ${m_{{{\widetilde{G}}}}^{\max}}$ from (\[Eq:mGmax\]), expression (\[Eq:OmegaGmin\]) leads to a lower limit on ${\Omega_{\widetilde{G}}}$ in terms of ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$: $${\Omega_{\widetilde{G}}}h^2 \gtrsim 0.422 \, c^2 \left(\frac{10^4\,{\mathrm{s}}}{\tau_{{{\widetilde{l}_1}}}}\right)^{\!\!\frac{1}{2}}\!\! \left(\frac{100~{\mathrm{GeV}}}{{m_{{{\widetilde{l}_1}}}}}\right)^{\!\!\frac{1}{2}}\!\! \left(\frac{{T_{{\mathrm{R}}}}}{10^{10}\,{\mathrm{GeV}}}\right) . \label{Eq:OmegaGlimit}$$ Comparing this lower limit with the dark matter constraint (\[Eq:OmegaDM\]), we arrive at a conservative ${T_{{\mathrm{R}}}}$ limit that can be determined at colliders by measuring the slepton mass ${m_{{{\widetilde{l}_1}}}}$, its lifetime $\tau_{{{\widetilde{l}_1}}}$, and $c$ (or the gluino mass $m_{{{\widetilde g}}}$): $$\begin{aligned} {T_{{\mathrm{R}}}}&\leq& \frac{2.37\times 10^9~{\mathrm{GeV}}}{c^{2}} \left(\frac{\Omega_{{\mathrm{DM}}}h^2}{0.1}\right) \nonumber\\ &&\times \left(\frac{\tau_{{{\widetilde{l}_1}}}}{10^4~{\mathrm{s}}}\right)^{\!\!\frac{1}{2}}\!\! \left(\frac{{m_{{{\widetilde{l}_1}}}}}{100~{\mathrm{GeV}}}\right)^{\!\!\frac{1}{2}}\!\! \equiv {T_{{\mathrm{R}}}}^{\max} \, . \label{Eq:TR_max}\end{aligned}$$ This is one of the main results of this letter. Note that this limit can be refined easily. Once an additional contribution $\Omega_{\mathrm{x}}$ to $\Omega_{{\mathrm{DM}}}$—such as an axion density or a non-thermally produced gravitino density ${\Omega_{\widetilde{G}}^{\mathrm{NTP}}}$—is taken for granted, the resulting tighter ${T_{{\mathrm{R}}}}$ limits can be obtained from (\[Eq:TR\_max\]) after the replacement: ${\Omega_{\mathrm{DM}}}\to{\Omega_{\mathrm{DM}}}-\Omega_{\mathrm{x}}$. In fact, based on the interplay between ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ and the contribution from NLSP decays, ${\Omega_{\widetilde{G}}^{\mathrm{NTP}}}$-dependent upper limits on ${T_{{\mathrm{R}}}}$ have been derived to be tested at colliders [@Choi:2007rh]. At this point, one has to clarify to which ${T_{{\mathrm{R}}}}$ definition the limit (\[Eq:TR\_max\]) applies. The analytic expression (\[Eq:OmegaGrTPSU3\]) is derived by assuming a radiation-dominated epoch with an initial temperature of ${T_{{\mathrm{R}}}}$ [@Bolz:2000fu; @Pradler:2006qh]. In a numerical treatment, the epoch in which the coherent oscillations of the inflaton field dominate the energy density of the Universe can also be taken into account, where one usually defines ${T_{{\mathrm{R}}}}$ in terms of the decay width $\Gamma_{\phi}$ of the inflaton field [@Kawasaki:2004qu; @Pradler:2006hh]. In fact, the numerical result for ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ has been found to agree with the corresponding analytic expression for [@Pradler:2006hh] $${T_{{\mathrm{R}}}}\simeq \left[\frac{90}{g_*({T_{{\mathrm{R}}}})\pi^2}\right]^{1/4} \sqrt{\frac{\Gamma_{\phi}{\mathrm{M}_{\mathrm{P}}}}{1.8}} \label{Eq:TR_definition}$$ which satisfies $\Gamma_{\phi}\simeq 1.8 H_{\mathrm{rad}}({T_{{\mathrm{R}}}})$ with the Hubble parameter $H_{\mathrm{rad}}(T)=\sqrt{g_*(T)\pi^2/90}\, T^2/{\mathrm{M}_{\mathrm{P}}}$ and an effective number of relativistic degrees of freedom of $g_*({T_{{\mathrm{R}}}})=228.75$. Thus, (\[Eq:TR\_definition\]) provides the ${T_{{\mathrm{R}}}}$ definition to which the upper limit (\[Eq:TR\_max\]) applies. For an alternative ${T_{{\mathrm{R}}}}$ definition given by $\Gamma_{\phi}=\xi H_{\mathrm{rad}}({T_{{\mathrm{R}}}}^{[\xi]})$, $${T_{{\mathrm{R}}}}^{[\xi]} \equiv \left[\frac{90}{g_*({T_{{\mathrm{R}}}})\pi^2}\right]^{1/4} \sqrt{\frac{\Gamma_{\phi}{\mathrm{M}_{\mathrm{P}}}}{\xi}} \, , \label{Eq:TR_xi_definition}$$ the upper limit (\[Eq:TR\_max\]) can be translated accordingly $${T_{{\mathrm{R}}}}^{\max\,[\xi]}=\sqrt{\frac{1.8}{\xi}}\,{T_{{\mathrm{R}}}}^{\max} \, . \label{Eq:TR_max_xi}$$ In particular, the associated numerically obtained ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ can be reproduced with the analytical expression after substituting ${T_{{\mathrm{R}}}}$ with $\sqrt{\xi/1.8}\,{T_{{\mathrm{R}}}}^{[\xi]}$ [@Pradler:2006hh]. ![Contours of ${T_{{\mathrm{R}}}^{\max}}$ imposed by ${\Omega_{\widetilde{G}}}h^2\leq{\Omega_{\mathrm{DM}}}h^2\leq 0.126$ for $c=1$ (solid lines) and $c=7$ (dashed lines) in the plane spanned by $\tau_{{{\widetilde{l}_1}}}$ and ${m_{{{\widetilde{l}_1}}}}$. The BBN constraints are identical to the ones shown in Fig. \[Fig:ProbingMgravitino\].[]{data-label="Fig:ProbingTRmax"}](ProbingTRmax_ite_f.eps){width="3.25in"} In Fig. \[Fig:ProbingTRmax\] contours of the upper limit ${T_{{\mathrm{R}}}^{\max}}$ given in (\[Eq:TR\_max\]) are shown for $c=1$ (solid lines) and $c=7$ (dashed lines) as obtained with ${\Omega_{\mathrm{DM}}}h^2\leq 0.126$.[^6] The most conservative upper limit is represented by the limiting case[^7] $c=1$ which holds even without insights into $m_{{{\widetilde g}}}$. Indeed, in this exceptional case, we assume implicitly a mass difference, $\Delta m=m_{{{\widetilde g}}}-m_{{{\widetilde{l}_1}}}>0$, such that ${{\widetilde{l}_1}}$–${{\widetilde g}}$ coannihilation effects are negligible and such that long-lived gluinos do not appear (avoiding thereby possibly severe hadronic BBN constraints from late decaying ${{\widetilde g}}$’s [@Fujii:2003nr]). With experimental insights into $m_{{{\widetilde{l}_1}}}$ and $m_{{{\widetilde g}}}$ (and thereby into $c$), the limit (\[Eq:TR\_max\]) can become considerably more severe than in the $c=1$ case. In Fig. \[Fig:ProbingTRmax\] this is illustrated for $c=7$ which is a value that appears typically in constrained scenarios with universal soft SUSY breaking parameters at the scale of grand unification ${M_{\mathrm{GUT}}}\simeq 2\times 10^{16}\,{\mathrm{GeV}}$. For example, in the CMSSM in which the gaugino masses, the scalar masses, and the trilinear scalar couplings are assumed to take on the respective universal values $m_{1/2}$, $m_0$, and $A_0$ at ${M_{\mathrm{GUT}}}$, the ${{\widetilde{\tau}_1}}$ NLSP region has been found to be associated with $m_{{{\widetilde{\tau}_1}}}^2\leq 0.21\, m_{1/2}^2$ and thus with $c>6$ over the entire natural parameter range [@Pradler:2007is]. The ${T_{{\mathrm{R}}}}$ limits derived above are in spirit similar to the ${m_{\widetilde{G}}}$-dependent ${T_{{\mathrm{R}}}}$ limits given in Refs. [@Moroi:1993mb; @Asaka:2000zh; @Cerdeno:2005eu; @Steffen:2006hw; @Pradler:2006hh; @Rychkov:2007uq; @Choi:2007rh]. Also the sensitivity of these ${T_{{\mathrm{R}}}}$ limits on the gaugino masses has already been discussed and used to provide ${m_{\widetilde{G}}}$-dependent upper limits on the gaugino masses for given values of ${T_{{\mathrm{R}}}}$ [@Bolz:1998ek; @Bolz:2000fu; @Fujii:2003nr; @Pradler:2006qh]. In fact, once the gaugino masses and the gravitino mass ${m_{\widetilde{G}}}$ (inferred from $m_{{{\widetilde{l}_1}}}$ and $\tau_{{{\widetilde{l}_1}}}$) are known, numerical results such as the ones shown in Fig. 2 of Ref. [@Pradler:2006hh] will provide a ${T_{{\mathrm{R}}}}$ limit that is more restrictive than our analytic estimate (\[Eq:TR\_max\]). However, as long as insights into a SUSY model possibly realized in nature are missing, we find it important to provide a conservative and robust estimate of ${T_{{\mathrm{R}}}^{\max}}$ that is insensitive to details in the SUSY spectrum (other than the assumption of the ${{\widetilde{G}}}$ LSP and a ${{\widetilde{l}_1}}$ NLSP). In particular, the comparison of (\[Eq:TR\_max\]) with more model dependent limits—obtained, e.g., within the CMSSM [@Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar]—demonstrates very clearly the impact of restrictive assumptions on the soft SUSY breaking sector. Moreover, the derivation of ${T_{{\mathrm{R}}}^{\max}}$ contours as a function of ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$ may turn out to become very useful since ${m_{{{\widetilde{l}_1}}}}$ and $\tau_{{{\widetilde{l}_1}}}$ are the quantities that might be directly accessible in collider experiments. Finally, as demonstrated below, the presentation of ${T_{{\mathrm{R}}}^{\max}}$ in the plane spanned by $\tau_{{{\widetilde{l}_1}}}$ and ${m_{{{\widetilde{l}_1}}}}$ allows also for a convenient analysis of ${T_{{\mathrm{R}}}^{\max}}$ in light of recent BBN constraints. Before proceeding, we would like to stress that the relation of ${T_{{\mathrm{R}}}^{\max}}$ to quantities ${m_{{{\widetilde{l}_1}}}}$, $\tau_{{{\widetilde{l}_1}}}$, and $c$ (or $m_{{{\widetilde g}}}$), which could be accessible at future collider experiments, relies crucially on assumptions on the cosmological history and the evolution of physical parameters. For example, for a non-standard thermal history with late-time entropy production, the thermally produced gravitino abundance can be diluted ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}\to{\Omega_{\widetilde{G}}^{\mathrm{TP}}}/\delta$ by a factor $\delta>1$ so that ${T_{{\mathrm{R}}}^{\max}}\to\delta\,{T_{{\mathrm{R}}}^{\max}}$ [@Pradler:2006hh]. Moreover, the limit (\[Eq:TR\_max\]) can be evaded if the strong coupling $g_s$ levels off in a non-standard way at high temperatures [@Buchmuller:2003is]. This emphasizes that (\[Eq:TR\_max\]) relies on the assumptions of a standard cosmological history and a strong gauge coupling that behaves at high temperatures as described by the renormalization group equation in the minimal supersymmetric Standard Model (MSSM), i.e., as described by $g_s({T_{{\mathrm{R}}}})$ given below (\[Eq:OmegaGrTPSU3\]). While tests of these assumptions seem inaccessible to terrestrial accelerator experiments, futuristic space-based gravitational-wave detectors such as the Big Bang Observer (BBO) or the Deci-hertz Interferometer Gravitational Wave Observatory (DECIGO) [@Seto:2001qf] could allow for tests of the thermal history after inflation and could even probe ${T_{{\mathrm{R}}}}$ [@Nakayama:2008ip; @Nakayama:2008wy] in a way that is complementary to the approach presented in this letter. Probing ${T_{{\mathrm{R}}}}$ with Primordial Nucleosynthesis {#Sec:TRmaxBBN} ============================================================ Gravitino LSP scenarios with a long-lived charged slepton NLSP can affect BBN as described in the Introduction. Particularly severe is the catalytic effect of ($^4$He${{\widetilde{l}_1}}^-$)-bound states on the primordial abundance of $^6$Li [@Pospelov:2006sc]. Indeed, the CBBN reaction (\[Eq:CBBNof6Li\]) can become very efficient at temperatures $T\simeq 10~{\mathrm{keV}}$ depending on the ${{\widetilde{l}_1}}$ abundance at that time. Observationally inferred upper limits on the primordial $^6$Li/H abundance ${^6\mathrm{Li/H}|_{\mathrm{p}}}$ (cf. [@Asplund:2005yt]) can thus be translated into $\tau_{{{\widetilde{l}_1}}}$-dependent upper limits on the thermal relic abundance of the negatively charged ${{\widetilde{l}_1}}^-$ NLSP [@Pospelov:2006sc; @Hamaguchi:2007mp; @Kawasaki:2007xb; @Takayama:2007du; @Pradler:2007is; @Kawasaki:2008qe]. For a given SUSY model, this abundance can be calculated, for example, with the computer program `micrOMEGAs 2.1` [@Belanger:2001fz+X]. By confronting the obtained abundance with the $\tau_{{{\widetilde{l}_1}}}$-dependent upper limits (see, e.g., Fig. 1 in Ref. [@Pradler:2007is]), one can extract an upper limit on $\tau_{{{\widetilde{l}_1}}}$ [@Pospelov:2006sc; @Cyburt:2006uv; @Steffen:2006wx; @Pradler:2006hh; @Hamaguchi:2007mp; @Kawasaki:2007xb; @Takayama:2007du; @Pradler:2007is; @Kawasaki:2008qe]. This $\tau_{{{\widetilde{l}_1}}}$ limit can then be used to determine ${T_{{\mathrm{R}}}^{\max}}$ directly from (\[Eq:TR\_max\]). Let us perform this procedure explicitly for the following thermal relic ${{\widetilde{l}_1}}$ NLSP abundance after decoupling and prior to decay [@Asaka:2000zh; @Fujii:2003nr]: $$Y_{{{\widetilde{l}_1}}} \equiv \frac{n_{{{\widetilde{l}_1}}}}{s} = 2\,Y_{{{\widetilde{l}_1}}^-} = 0.7 \times 10^{-13}\left(\frac{{m_{{{\widetilde{l}_1}}}}}{100~{\mathrm{GeV}}}\right) , \label{Eq:Yslepton}$$ where $s$ denotes the entropy density and $n_{{{\widetilde{l}_1}}}$ the total ${{\widetilde{l}_1}}$ number density assuming an equal number density of positively and negatively charged ${{\widetilde{l}_1}}$’s. We confront (\[Eq:Yslepton\]) with the $Y_{{{\widetilde{l}_1}}^-}$ limits that emerge from a calculation of the $^6$Li abundance from CBBN, $^6$Li/H$|_{{\mathrm{CBBN}}}$, which uses the state-of-the-art result of the catalyzed $^6$Li production cross section [@Hamaguchi:2007mp] and the Boltzmann equation (instead of the Saha type approximation) to describe the time evolution of the ($^4$He${{\widetilde{l}_1}}^-$)-bound-state abundance [@Takayama:2007du; @Pradler:2007is].[^8] Contour lines of $^6$Li/H$|_{{\mathrm{CBBN}}}$ obtained in this calculation are shown, e.g., in Fig. 1 of Ref. [@Pradler:2007is].[^9] Working with an upper limit on the primordial $^6$Li/H abundance of [@Cyburt:2002uv] $${^6\mathrm{Li/H}|_{\mathrm{p}}}\leq 2\times 10^{-11} , \label{Eq:LiLimit1}$$ the corresponding contour given in that figure is the one that provides the upper limit for $Y_{{{\widetilde{l}_1}}^-}$. As a second more conservative limit, we consider in this work also $${^6\mathrm{Li/H}|_{\mathrm{p}}}\leq 6\times 10^{-11} , \label{Eq:LiLimit2}$$ which provides a more relaxed $Y_{{{\widetilde{l}_1}}^-}$ limit. In Figs. \[Fig:ProbingMgravitino\] and \[Fig:ProbingTRmax\], the long-dash-dotted (red in the web version) lines show the constraints obtained by confronting (\[Eq:Yslepton\]) with the $Y_{{{\widetilde{l}_1}}^-}$ limits for ${^6\mathrm{Li/H}|_{\mathrm{p}}}\leq 2\times 10^{-11}$ (left line) and $6\times 10^{-11}$ (right line). It is the region to the right of the respective line that is disfavored by $^6$Li/H$|_{{\mathrm{CBBN}}}$ exceeding the limit inferred from observations. Only the constraint from the primordial D abundance on hadronic energy release [@Kawasaki:2004qu; @Jedamzik:2006xz] in ${{\widetilde{l}_1}}$ decays [@Feng:2004mt; @Cerdeno:2005eu; @Steffen:2006hw] can be more severe than the CBBN constraints [@Cyburt:2006uv; @Steffen:2006wx; @Pradler:2006hh; @Kawasaki:2007xb; @Steffen:2007sp; @Kawasaki:2008qe].[^10] In Figs. \[Fig:ProbingMgravitino\] and \[Fig:ProbingTRmax\], the associated disfavored region is the one above the short-dash-dotted (blue in the web version) line. For details on this constraint, see [@Steffen:2006hw; @Pradler:2006hh] and references therein. In Fig. \[Fig:ProbingMgravitino\] one sees explicitly that the BBN constraints disfavor the region $0.1\,{m_{{{\widetilde{l}_1}}}}\lesssim{m_{\widetilde{G}}}<{m_{{{\widetilde{l}_1}}}}$ as already emphasized in the Introduction. Moreover, the $^6$Li constraint imposes for ${m_{\widetilde{G}}}\gtrsim 10~{\mathrm{GeV}}$ the lower limit ${m_{{{\widetilde{l}_1}}}}>400~{\mathrm{GeV}}$ which implies $m_{{{\widetilde g}}}(M_Z)>2.4~{\mathrm{TeV}}$ for $c>6$. This is in agreement with what had previously been realized in the CMSSM (where $c>6$ in the ${{\widetilde{\tau}_1}}$ NLSP region [@Pradler:2007is]): For ${m_{\widetilde{G}}}$ in the range that is natural for gravity-mediated SUSY breaking, the $^6$Li constraint can point to a SUSY mass spectrum that will be difficult to probe at the LHC [@Cyburt:2006uv; @Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar]. On the other hand, Fig. \[Fig:ProbingMgravitino\] shows very clearly that the $^6$Li constraint becomes negligible for ${m_{\widetilde{G}}}< 1~{\mathrm{GeV}}$, i.e., for the ${m_{\widetilde{G}}}$ range that is natural for gauge-mediated SUSY breaking scenarios. Figure \[Fig:ProbingTRmax\] allows us to read off the ${T_{{\mathrm{R}}}^{\max}}$ value imposed by the $^6$Li constraint in the collider-friendly region of ${m_{{{\widetilde{l}_1}}}}\lesssim 1~{\mathrm{TeV}}$. Our most conservative and thereby most robust limit is the one obtained in the limiting case $c=1$. There one finds the ${T_{{\mathrm{R}}}^{\max}}=3\times 10^9\,{\mathrm{GeV}}$ contour in the region that is allowed by the BBN constraints. If the gluino turns out to be significantly heavier than the ${{\widetilde{l}_1}}$ NLSP, the ${T_{{\mathrm{R}}}^{\max}}$ value will become considerably more severe. For example, for the case of $c=7$, shown by the dashed lines in Fig. \[Fig:ProbingTRmax\], one finds that a reheating temperature above ${T_{{\mathrm{R}}}^{\max}}=10^8\,{\mathrm{GeV}}$ is disfavored by BBN constraints. Note that these limit are quite robust since they emerge from the conservative limit (\[Eq:TR\_max\]). In fact, one may consider these limits as overly conservative by at least a factor of about 1.5 as discussed below (\[Eq:OmegaGmin\]). Indeed, a more restrictive limit of ${T_{{\mathrm{R}}}}\lesssim 10^7\,{\mathrm{GeV}}$ is found within the CMSSM by using the full expressions for ${\Omega_{\widetilde{G}}^{\mathrm{TP}}}$ given in [@Pradler:2006qh] which include also the electroweak contributions to thermal gravitino production [@Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar]. Let us comment on the model dependence of the BBN constraints. As described above, the BBN constraints shown in Figs. \[Fig:ProbingMgravitino\] and \[Fig:ProbingTRmax\] are derived with the yield (\[Eq:Yslepton\]). This does introduce a model dependence into the ${T_{{\mathrm{R}}}^{\max}}$ values discussed in the preceding paragraph. However, one should stress two points: (i) The yield (\[Eq:Yslepton\]) is quite typical for an electrically charged massive thermal relic [@Asaka:2000zh]. (ii) The $\tau_{{{\widetilde{l}_1}}}$-dependent upper limits on $Y_{{{\widetilde{l}_1}}^-}$ are very steep in the relevant region as can be seen, e.g., in Fig. 1 of Ref. [@Pradler:2007is]. Indeed, this steepness is reflected by the relatively weak $m_{{{\widetilde{l}_1}}}$ dependence of the $^6$Li constraints that can be seen in Figs. \[Fig:ProbingMgravitino\] and \[Fig:ProbingTRmax\]. These two points support that the shown BBN constraints and the associated ${T_{{\mathrm{R}}}^{\max}}$ values are quite robust. Only for very generous upper limits on the primordial $^6$Li abundance and/or exceptionally small $Y_{{{\widetilde{l}_1}}^-}$ values can the CBBN bound be relaxed substantially. The latter could be achieved, for example, with non-standard entropy production after thermal freeze out of the ${{\widetilde{l}_1}}$ NLSP and before BBN [@Buchmuller:2006tt; @Pradler:2006hh; @Hamaguchi:2007mp] which we do not consider since we assume a standard cosmological history throughout this letter. Probing ${T_{{\mathrm{R}}}}$ by Measuring the Gluino–Slepton Mass Ratio ======================================================================= In this section we present a way that could allow us to probe with collider experiments the ${T_{{\mathrm{R}}}^{\max}}$ value imposed by the $^6$Li constraint—or any other constraint that can be translated into a similar upper limit on $\tau_{{{\widetilde{l}_1}}}$—and the dark matter constraint (\[Eq:OmegaDM\]). As discussed above, the $^6$Li constraint implies basically an upper limit on the ${{\widetilde{l}_1}}$ NLSP lifetime $$\tau_{{{\widetilde{l}_1}}} \leq {\tau^{\max}}, \label{Eq:taumax}$$ which is, e.g., ${\tau^{\max}}\simeq 5\times 10^3\,{\mathrm{s}}$ for the case of the yield (\[Eq:Yslepton\]) as can be seen explicitly in Fig. \[Fig:ProbingTRmax\]. With a given ${\tau^{\max}}$ and (\[Eq:TR\_max\]) derived from the dark matter constraint (\[Eq:OmegaDM\]), one arrives immediately at the following ${m_{\widetilde{G}}}$-independent upper limit on the gluino–slepton mass ratio at the weak scale: $$\begin{aligned} c &\leq& \left(\frac{2.37\times 10^9~{\mathrm{GeV}}}{{T_{{\mathrm{R}}}^{\max}}}\right)^{\!\!\frac{1}{2}}\!\! \left(\frac{\Omega_{{\mathrm{DM}}}h^2}{0.1}\right)^{\!\!\frac{1}{2}}\!\! \nonumber\\ &&\times \left(\frac{{\tau^{\max}}}{10^4~{\mathrm{s}}}\right)^{\!\!\frac{1}{4}}\!\! \left(\frac{{m_{{{\widetilde{l}_1}}}}}{100~{\mathrm{GeV}}}\right)^{\!\!\frac{1}{4}}\!\! \equiv {c^{\max}}, \label{Eq:cmax}\end{aligned}$$ i.e., a reheating temperature of at most ${T_{{\mathrm{R}}}^{\max}}$ can viable with a given ${m_{{{\widetilde{l}_1}}}}$ only for a mass ratio $c\leq{c^{\max}}$. In other words, for a lifetime constraint ${\tau^{\max}}$ inferred from cosmological considerations (such as CBBN) and/or collider experiments, a simultaneous measurement of ${m_{{{\widetilde{l}_1}}}}$ and the mass ratio $c$ will provide a model-independent conservative limit ${T_{{\mathrm{R}}}^{\max}}$ for ${{\widetilde{G}}}$ LSP scenarios with a ${{\widetilde{l}_1}}$ NLSP. In Fig. \[Fig:ProbingC\] ![Upper limits on the mass ratio $c=m_{{{\widetilde g}}}(M_Z)/{m_{{{\widetilde{l}_1}}}}$ imposed by ${\Omega_{\widetilde{G}}}h^2\leq{\Omega_{\mathrm{DM}}}h^2\leq 0.126$ and $\tau_{{\widetilde{l}_1}}\leq3\times 10^3\,{\mathrm{s}}$ ($10^4\,{\mathrm{s}}$) are shown as a function of ${m_{{{\widetilde{l}_1}}}}$ by the solid lines (dashed lines) for values of ${T_{{\mathrm{R}}}^{\max}}$ ranging from $10^8\,{\mathrm{GeV}}$ up to $3\times 10^9\,{\mathrm{GeV}}$.[]{data-label="Fig:ProbingC"}](ProbingC_ite_f.eps){width="3.25in"} the solid lines (dashed lines) show ${c^{\max}}$ imposed by the lifetime constraint $\tau_{{{\widetilde{l}_1}}}\leq 3\times 10^3\,{\mathrm{s}}$ ($10^4\,{\mathrm{s}}$) and and the dark matter constraint ${\Omega_{\widetilde{G}}}h^2\leq{\Omega_{\mathrm{DM}}}h^2\leq 0.126$ as a function of ${m_{{{\widetilde{l}_1}}}}$ for values of ${T_{{\mathrm{R}}}^{\max}}$ ranging from $10^8\,{\mathrm{GeV}}$ up to $3\times 10^9\,{\mathrm{GeV}}$. The lifetime constraint $\tau_{{{\widetilde{l}_1}}}\leq 5\times 10^3\,{\mathrm{s}}$ is slightly weaker (and thereby more conservative) than the $^6$Li CBBN constraint obtained with the yield (\[Eq:Yslepton\]) for ${m_{{{\widetilde{l}_1}}}}> 100~{\mathrm{GeV}}$ as can be seen in Fig. \[Fig:ProbingTRmax\]. Accordingly, the ${T_{{\mathrm{R}}}^{\max}}$ values for $c=1$ and $c=7$ discussed in the previous section are more restrictive than the corresponding ${T_{{\mathrm{R}}}^{\max}}$ values in Fig. \[Fig:ProbingC\]. The even more conservative ${c^{\max}}$ curves obtained for $\tau_{{{\widetilde{l}_1}}}\leq 10^4\,{\mathrm{s}}$ are presented to cover also the case of a more relaxed CBBN constraint which could result, for example, from ${^6\mathrm{Li/H}|_{\mathrm{p}}}\leq 2.7\times 10^{-10}$ [@Jedamzik:2007qk], i.e., an upper limit that is about an order of magnitude more generous than (\[Eq:LiLimit1\]). This emphasizes the conservative character of the corresponding ${c^{\max}}$ curves shown in Fig. \[Fig:ProbingC\]. Testing the Viability of Thermal Leptogenesis at Colliders ========================================================== Thermal leptogenesis provides an attractive explanation of the baryon asymmetry in the Universe [@Fukugita:1986hr]. Since successful thermal leptogenesis with hierarchical right-handed heavy Majorana neutrinos requires a reheating temperature of ${T_{{\mathrm{R}}}}>10^9\,{\mathrm{GeV}}$ [@Davidson:2002qv; @Buchmuller:2004nz; @Blanchet:2006be; @Antusch:2006gy], one will be able to test its viability at the LHC with the method described in the previous section. From earlier studies of gravitino LSP scenarios [@Bolz:1998ek; @Bolz:2000fu; @Fujii:2003nr; @Pradler:2006qh], it is known that thermal leptogenesis can be associated with testable (${m_{\widetilde{G}}}$-dependent) upper limits on the masses of the gluino and the other gauginos. Nevertheless, the $\tau_{{{\widetilde{l}_1}}}$ constraint from CBBN of $^6$Li is not taken into account in any of these studies. In fact, as mentioned in the Introduction, this $\tau_{{{\widetilde{l}_1}}}$ constraint disfavors thermal leptogenesis in constrained scenarios with universal soft SUSY breaking parameters at ${M_{\mathrm{GUT}}}$ [@Pradler:2006hh; @Pradler:2007is; @Pradler:2007ar; @Kersten:2007ab]. Interestingly, in this work, it is this upper limit on $\tau_{{{\widetilde{l}_1}}}$ that allows us to arrive in a model independent way at a testable prediction of thermal leptogenesis that does not depend on the gravitino mass ${m_{\widetilde{G}}}$. As can be seen in Fig. \[Fig:ProbingC\], already the conservative limits imposed by $\tau_{{{\widetilde{l}_1}}}\leq 10^4\,{\mathrm{s}}$ show that ${T_{{\mathrm{R}}}^{\max}}=10^9\,{\mathrm{GeV}}$ is associated with a gluino-slepton mass ratio of $c<3$ in the collider-friendly region of $m_{{{\widetilde{l}_1}}}< 1~{\mathrm{TeV}}$. Thus, this ratio favors a gluino mass that will be accessible at the LHC. In this way, $c<3$ can be understood as a testable prediction of thermal leptogenesis for phenomenology at the LHC. Indeed, the realistic, less conservative constraints $\tau_{{{\widetilde{l}_1}}}\leq 5\times 10^3\,{\mathrm{s}}$ and ${T_{{\mathrm{R}}}}\geq 3\times 10^9\,{\mathrm{GeV}}$ point even to the exceptional case with a gluino that is only slightly heavier than the slepton NLSP: $c\lesssim 1.5$. If realized in nature, this will guarantee unique signatures at the LHC. Conclusion ========== The observation of a (quasi-) stable heavy charged slepton ${{\widetilde{l}_1}}$ as the lightest Standard Model superpartner at future colliders could provide a first hint towards the gravitino LSP being the fundamental constituent of dark matter. Under the assumption that the gravitino is the LSP and a long-lived ${{\widetilde{l}_1}}$ the NLSP, we have shown that measurements of the slepton mass $m_{{{\widetilde{l}_1}}}$ and its lifetime $\tau_{{{\widetilde{l}_1}}}$ can provide the gravitino mass ${m_{\widetilde{G}}}$ and upper limits on the reheating temperature ${T_{{\mathrm{R}}}}$ after inflation, which can be tightened once the gluino mass $m_{{{\widetilde g}}}$ is measured. The conceivable insights into ${m_{\widetilde{G}}}$ are a direct consequence of the SUGRA Lagrangian and of R-parity conservation. Thus, they will be valid for any cosmological scenario provided the ${{\widetilde{G}}}$ is the LSP and a ${{\widetilde{l}_1}}$ the NLSP. The possibility to probe ${T_{{\mathrm{R}}}}$ at colliders results from the ${T_{{\mathrm{R}}}}$ dependence of the thermally produced gravitino density and the fact that this density cannot exceed the dark matter density. Thereby, the given upper limits on ${T_{{\mathrm{R}}}}$ rely not only on the SUGRA Lagrangian and R-parity conservation but also on the assumptions of a standard cosmological history and on (gauge) couplings that evolve at high temperatures as expected from their standard MSSM renormalization group running. Nevertheless, under these assumptions, the presented ${T_{{\mathrm{R}}}}$ limits are very robust since they have been derived in a model independent way from a conservative lower limit on the (thermally produced) gravitino density. Prior to collider measurements of $m_{{{\widetilde{l}_1}}}$, $\tau_{{{\widetilde{l}_1}}}$, and $m_{{{\widetilde g}}}$, primordial nucleosynthesis can be used to constrain ${m_{\widetilde{G}}}$ and ${T_{{\mathrm{R}}}}$. In particular, we have found that the constraint $\tau_{{{\widetilde{l}_1}}}\lesssim 5\times 10^3\,{\mathrm{s}}$ from CBBN of $^6$Li implies a truly conservative limit of ${T_{{\mathrm{R}}}}\lesssim 3\times 10^9\,{\mathrm{GeV}}$ in the collider-friendly region of $m_{{{\widetilde{l}_1}}}\lesssim 1\,{\mathrm{TeV}}$ and in the limiting case of a small gluino–slepton mass splitting at the weak scale: $c=m_{{{\widetilde g}}}/{m_{{{\widetilde{l}_1}}}}\simeq 1$. Indeed, with $c=7$—as obtained in scenarios with universal soft SUSY breaking parameters at the scale of grand unification—our robust conservative upper limit is ${T_{{\mathrm{R}}}}\lesssim 10^8\,{\mathrm{GeV}}$ and thus as severe as in scenarios with an unstable gravitino of mass ${m_{\widetilde{G}}}\lesssim 5~{\mathrm{TeV}}$. The $\tau_{{{\widetilde{l}_1}}}$ constraint from CBBN of $^6$Li has allowed us to derive upper limits on ${T_{{\mathrm{R}}}}$ that depend on ${m_{{{\widetilde{l}_1}}}}$ and $c$ only. In particular, we find that the condition for successful thermal leptogenesis with hierarchical right-handed heavy Majorana neutrinos, ${T_{{\mathrm{R}}}}> 10^9\,{\mathrm{GeV}}$ ($3\times 10^9\,{\mathrm{GeV}}$), implies an upper limit of $c<3$ ($1.5$) for ${m_{{{\widetilde{l}_1}}}}<1~{\mathrm{TeV}}$. This is a prediction of thermal leptogenesis that will be testable at the upcoming LHC experiments. I am grateful to Josef Pradler for valuable discussions and comments on the manuscript. This research was partially supported by the Cluster of Excellence ‘Origin and Structure of the Universe.’ [88]{} S. Dimopoulos, R. Esmailzadeh, L.J. Hall, G.D. Starkman, Nucl. Phys. **B311**, 699 (1989) K. Jedamzik, Phys. Rev. **D70**, 063524 (2004), `astro-ph/0402344` M. Kawasaki, K. Kohri, T. Moroi, Phys. Lett. **B625**, 7 (2005), `astro-ph/0402490` M. Kawasaki, K. Kohri, T. Moroi, Phys. Rev. **D71**, 083502 (2005), `astro-ph/0408426` K. Kohri, T. Moroi, A. Yotsuyanagi, Phys. Rev. **D73**, 123511 (2006), `hep-ph/0507245` K. Jedamzik, Phys. Rev. **D74**, 103509 (2006), `hep-ph/0604251` M. Kawasaki, K. Kohri, T. Moroi, A. Yotsuyanagi (2008), `0804.3745` J. Wess, J. Bagger, *Supersymmetry and supergravity*, [Princeton]{} [University]{} [Press]{}, Princeton, USA (1992) J.R. Ellis, J.E. Kim, D.V. Nanopoulos, Phys. Lett. **B145**, 181 (1984) T. Moroi, H. Murayama, M. Yamaguchi, Phys. Lett. **B303**, 289 (1993) J.R. Ellis, D.V. Nanopoulos, K.A. Olive, S.J. Rey, Astropart. Phys. **4**, 371 (1996), `hep-ph/9505438` M. Bolz, W. Buchm[ü]{}ller, M. Pl[ü]{}macher, Phys. Lett. **B443**, 209 (1998), `hep-ph/9809381` M. Bolz, A. Brandenburg, W. Buchm[ü]{}ller, Nucl. Phys. **B606**, 518 (2001), `hep-ph/0012052` J. Pradler, F.D. Steffen, Phys. Rev. **D75**, 023509 (2007), `hep-ph/0608344` J. Pradler, F.D. Steffen, Phys. Lett. **B648**, 224 (2007), `hep-ph/0612291` J. Pradler, *Electroweak Contributions to Thermal Gravitino Production* (2006), `0708.2786` V.S. Rychkov, A. Strumia, Phys. Rev. **D75**, 075011 (2007), `hep-ph/0701104` T. Asaka, K. Hamaguchi, K. Suzuki, Phys. Lett. **B490**, 136 (2000), `hep-ph/0005136` J.R. Ellis, K.A. Olive, Y. Santoso, V.C. Spanos, Phys. Lett. **B588**, 7 (2004), `hep-ph/0312262` J.L. Feng, S. Su, F. Takayama, Phys. Rev. **D70**, 075019 (2004), `hep-ph/0404231` D.G. Cerdeno, K.Y. Choi, K. Jedamzik, L. Roszkowski, R. Ruiz de Austri, JCAP **0606**, 005 (2006), `hep-ph/0509275` F.D. Steffen, JCAP **0609**, 001 (2006), `hep-ph/0605306` F.D. Steffen (2007), `0711.1240` M. Fukugita, T. Yanagida, Phys. Lett. **B174**, 45 (1986) S. Davidson, A. Ibarra, Phys. Lett.  B [**535**]{} 25 (2002), `hep-ph/0202239` W. Buchm[ü]{}ller, P. Di Bari, M. Pl[ü]{}macher, Ann. Phys. **315**, 305 (2005), `hep-ph/0401240` S. Blanchet, P. Di Bari, JCAP **0703**, 018 (2007), `hep-ph/0607330` S. Antusch, A.M. Teixeira, JCAP **0702**, 024 (2007), `hep-ph/0611232` E. Nardi, Y. Nir, E. Roulet, J. Racker, JHEP [**0601**]{} 164 (2006), `hep-ph/0601084` A. Abada, S. Davidson, F. X. Josse-Michaux, M. Losada, A. Riotto, JCAP [**0604**]{} 004 (2006), `hep-ph/0601083`; *ibid.* JHEP [**0609**]{} 010 (2006), `hep-ph/0605281` M. Flanz, E. A. Paschos, U. Sarkar, Phys. Lett.  B [**345**]{}, 248 (1995); *Erratum-ibid.* B [**382**]{}, 447 (1996), `hep-ph/9411366` L. Covi, E. Roulet, Phys. Lett.  B [**399**]{}, 113 (1997), `hep-ph/9611425` A. Pilaftsis, Phys. Rev.  D [**56**]{}, 5431 (1997), `hep-ph/9707235` A. Anisimov, A. Broncano, M. Pl[ü]{}macher, Nucl. Phys.  B [**737**]{} 176 (2006), `hep-ph/0511248` F. Hahn-Woernle, M. Pl[ü]{}macher, `0801.3972` M. Fujii, M. Ibe, T. Yanagida, Phys. Lett. **B579**, 6 (2004), `hep-ph/0310142` W. Buchm[ü]{}ller, K. Hamaguchi, M. Ratz, T. Yanagida, Phys. Lett. **B588**, 90 (2004), `hep-ph/0402179` H.U. Martyn, Eur. Phys. J. **C48**, 15 (2006), `hep-ph/0605257` K. Hamaguchi, M.M. Nojiri, A. de Roeck, JHEP **03**, 046 (2007), `hep-ph/0612060` W.M. Yao et al. (Particle Data Group), J. Phys. **G33**, 1 (2006) S. Dimopoulos, D. Eichler, R. Esmailzadeh, G.D. Starkman, Phys. Rev. **D41**, 2388 (1990) M. Pospelov, Phys. Rev. Lett. **98**, 231301 (2007), `hep-ph/0605215` R.H. Cyburt, J.R. Ellis, B.D. Fields, K.A. Olive, V.C. Spanos, JCAP **0611**, 014 (2006), `astro-ph/0608562` F.D. Steffen, AIP Conf. Proc. **903**, 595 (2007), `hep-ph/0611027` K. Hamaguchi, T. Hatsuda, M. Kamimura, Y. Kino, T.T. Yanagida, Phys. Lett. **B650**, 268 (2007), `hep-ph/0702274` M. Kawasaki, K. Kohri, T. Moroi, Phys. Lett. **B649**, 436 (2007), `hep-ph/0703122` J. Pradler, F.D. Steffen (2007), `0710.2213` J. Pradler, F.D. Steffen (2007), `0710.4548` K. Kohri, F. Takayama, Phys. Rev. **D76**, 063507 (2007), `hep-ph/0605243` M. Kaplinghat, A. Rajaraman, Phys. Rev. **D74**, 103004 (2006), `astro-ph/0606209` C. Bird, K. Koopmans, M. Pospelov (2007), `hep-ph/0703096` K. Jedamzik, Phys. Rev. **D77**, 063524 (2008), `0707.2070` K. Jedamzik, JCAP **0803**, 008 (2008), `0710.5153` M. Pospelov, J. Pradler, F.D. Steffen, in preparation J. Kersten, K. Schmidt-Hoberg, JCAP **0801**, 011 (2008), `0710.4528` W. Buchm[ü]{}ller, L. Covi, K. Hamaguchi, A. Ibarra, T. Yanagida, JHEP **03**, 037 (2007), `hep-ph/0702184` S. Ambrosanio, B. Mele, S. Petrarca, G. Polesello, A. Rimoldi, JHEP **01**, 014 (2001), `hep-ph/0010081` M. Drees, X. Tata, Phys. Lett. **B252**, 695 (1990) A. Nisati, S. Petrarca, G. Salvini, Mod. Phys. Lett. **A12**, 2213 (1997), `hep-ph/9707376` J.L. Feng, T. Moroi, Phys. Rev. **D58**, 035001 (1998), `hep-ph/9712499` J.R. Ellis, A.R. Raklev, O.K. Oye, JHEP **10**, 061 (2006), `hep-ph/0607261` S. Bressler (ATLAS) (2007), `0710.2111` P. Zalewski (2007), `0710.2647` J.L. Goity, W.J. Kossler, M. Sher, Phys. Rev. **D48**, 5437 (1993), `hep-ph/9305244` K. Hamaguchi, Y. Kuno, T. Nakaya, M.M. Nojiri, Phys. Rev. **D70**, 115007 (2004), `hep-ph/0409248` J.L. Feng, B.T. Smith, Phys. Rev. **D71**, 015004 (2005), `hep-ph/0409278` A. De Roeck et al., Eur. Phys. J. **C49**, 1041 (2007), `hep-ph/0508198` O. Cakir, I.T. Cakir, J.R. Ellis, Z. Kirca (2007), `hep-ph/0703121` H.U. Martyn (2007), `0709.1030` A. Brandenburg, L. Covi, K. Hamaguchi, L. Roszkowski, F.D. Steffen, Phys. Lett. **B617**, 99 (2005), `hep-ph/0501287`; F.D. Steffen (2005), `hep-ph/0507003` L. Covi, J.E. Kim, L. Roszkowski, Phys. Rev. Lett. **82**, 4180 (1999), `hep-ph/9905212` L. Covi, H.B. Kim, J.E. Kim, L. Roszkowski, JHEP **05**, 033 (2001), `hep-ph/0101009` A. Brandenburg, F.D. Steffen, JCAP **0408**, 008 (2004), `hep-ph/0405158`; `hep-ph/0406021`; `hep-ph/0407324` L. Covi, L. Roszkowski, R. Ruiz de Austri, M. Small, JHEP **06**, 003 (2004), `hep-ph/0402240` T. Asaka, S. Nakamura, M. Yamaguchi, Phys. Rev. **D74**, 023520 (2006), `hep-ph/0604132` F. Takahashi, Phys. Lett. **B660**, 100 (2008), `0705.0579` M. Endo, F. Takahashi, T.T. Yanagida, Phys. Rev. **D76**, 083509 (2007), `0706.0986` S. Borgani, A. Masiero, M. Yamaguchi, Phys. Lett. **B386**, 189 (1996), `hep-ph/9605222` E.A. Baltz, H. Murayama, JHEP **05**, 067 (2003), `astro-ph/0108172` M. Fujii, T. Yanagida, Phys. Lett. **B549**, 273 (2002), `hep-ph/0208191` M. Fujii, M. Ibe, T. Yanagida, Phys. Rev. **D69**, 015006 (2004), `hep-ph/0309064` M. Lemoine, G. Moultaka, K. Jedamzik, Phys. Lett. **B645**, 222 (2007), `hep-ph/0504021` K. Jedamzik, M. Lemoine, G. Moultaka, Phys. Rev. **D73**, 043514 (2006), `hep-ph/0506129` K. Y. Choi, L. Roszkowski, R. Ruiz de Austri, JHEP [**0804**]{}, 016 (2008), `0710.3349` D.N. Spergel et al. (WMAP), Astrophys. J. Suppl. **170**, 377 (2007), `astro-ph/0603449` W. Buchm[ü]{}ller, K. Hamaguchi, M. Ratz, Phys. Lett. **B574**, 156 (2003), `hep-ph/0307181` N. Seto, S. Kawamura, T. Nakamura, Phys. Rev. Lett. **87**, 221103 (2001), `astro-ph/0108011` K. Nakayama, S. Saito, Y. Suwa, J. Yokoyama, Phys. Rev. **D77**, 124001 (2008), `0802.2452` K. Nakayama, S. Saito, Y. Suwa, J. Yokoyama (2008), `0804.1827` M. Asplund, D.L. Lambert, P.E. Nissen, F. Primas, V.V. Smith, Astrophys. J. **644**, 229 (2006), `astro-ph/0510636` F. Takayama (2007), `0704.2785` G. Belanger, F. Boudjema, A. Pukhov, A. Semenov, Comput. Phys. Commun. **149**, 103 (2002), `hep-ph/0112278`; *ibid.* **174**, 577 (2006), `hep-ph/0405253`; *ibid.* **177**, 894 (2007); *ibid.* (2008), `0803.2360` R.H. Cyburt, J.R. Ellis, B.D. Fields, K.A. Olive, Phys. Rev. **D67**, 103521 (2003), `astro-ph/0211258` M. Pospelov (2007), `0712.0647` W. Buchm[ü]{}ller, K. Hamaguchi, M. Ibe, T.T. Yanagida, Phys. Lett. **B643**, 124 (2006), `hep-ph/0605164` [^1]: Note that flavor effects [@Nardi:2006fx; @Abada:2006fw; @Blanchet:2006be; @Antusch:2006gy] do not change the lower bound $T_{{\mathrm{R}}} > 10^9\,{\mathrm{GeV}}$ required by successful thermal leptogenesis with hierarchical right-handed neutrinos [@Blanchet:2006be; @Antusch:2006gy]. However, in the case of (nearly) mass-degenerate heavy right-handed Majorana neutrinos, resonant leptogenesis can explain the baryon asymmetry at smaller values of $T_{{\mathrm{R}}}$ [@Flanz:1994yx; @Covi:1996fm; @Pilaftsis:1997jf; @Anisimov:2005hr]. Another example for a framework in which the limit $T_{{\mathrm{R}}} > 10^9\,{\mathrm{GeV}}$ is relaxed is non-thermal leptogenesis; see, e.g., [@HahnWoernle:2008pq] and references therein. [^2]: While numerous other bound-state effects can affect the abundances of $^6$Li and the other primordial light elements [@Kohri:2006cn; @Kaplinghat:2006qr; @Cyburt:2006uv; @Bird:2007ge; @Jedamzik:2007cp; @Jedamzik:2007qk], the approximate $\tau_{{{\widetilde{l}_1}}}$ bound is found to be quite robust; cf. [@Pospelov:MPP-2008-63]. [^3]: For the case of broken $\mathrm{R}$-parity, see, e.g., [@Buchmuller:2007ui]. [^4]: In gauge-mediated SUSY breaking scenarios, light gravitinos can be viable thermal relics if their abundance is diluted by entropy production, which can result, for example, from decays of messenger fields [@Baltz:2001rq; @Fujii:2002fv; @Fujii:2003iw; @Lemoine:2005hu; @Jedamzik:2005ir]. [^5]: For ${T_{{\mathrm{R}}}}$ limits obtained by taking into account contributions to ${\Omega_{\widetilde{G}}}$ from NLSP decays, see e.g. [@Asaka:2000zh; @Fujii:2003nr; @Cerdeno:2005eu; @Steffen:2006hw; @Pradler:2006qh; @Pradler:2006hh; @Choi:2007rh; @Pradler:2007ar]. [^6]: This ${\Omega_{\mathrm{DM}}}h^2$ value is used to allow for a convenient comparison with our previous works [@Pradler:2006qh; @Pradler:2006hh]. There ${\Omega_{\mathrm{DM}}}h^2\leq 0.126$ is understood as a nominal $3\sigma$ limit derived with a restrictive six-parameter “vanilla” model from the three year data set of the Wilkinson Microwave Anisotropy Probe (WMAP) satellite [@Spergel:2006hy]. [^7]: This limiting case has already been considered in the benchmark scenarios $C_{1,2}$ in Ref. [@Steffen:2006hw] which are now understood to be disfavored by the $^6$Li constraint from CBBN [@Pospelov:2006sc]. [^8]: We thank Josef Pradler for providing us with the $^6$Li/H$|_{{\mathrm{CBBN}}}$ data from the CBBN treatment of Ref. [@Pradler:2007is]. [^9]: Comparisons of [@Pradler:2007is] with [@Cyburt:2006uv; @Kawasaki:2007xb; @Jedamzik:2007qk; @Kawasaki:2008qe], in which also the possible destruction of $^6$Li due to ${{\widetilde{l}_1}}$ decays is considered, show that those effects affect the $^6$Li constraint only marginally. [^10]: The additional primordial bound-state effects discussed in [@Kohri:2006cn; @Kaplinghat:2006qr; @Cyburt:2006uv; @Bird:2007ge; @Kawasaki:2007xb; @Jedamzik:2007cp; @Jedamzik:2007qk; @Pospelov:2007js] do not affect the conclusions of this letter; cf. [@Pospelov:MPP-2008-63].

ArXiv

--- abstract: 'It is well known that a sequence of two non-collinear pure Lorentz transformations (boosts) is not a boost again, but involves a spatial rotation, the Wigner or Thomas-Wigner rotation. The formation of this rotation is visually illustrated by moving a Born-rigid object on a closed trajectory in several sections. Within each section the boost’s proper time duration is assumed to be the same and the object’s centre accelerates uniformly. Born-rigidity implies that the stern of this object accelerates faster than its bow. It is shown that at least five boosts are required to return the object’s centre to its start position. With these assumptions, the Thomas-Wigner rotation angle depends on a single parameter only, the maximum speed reached within each boost section. The visualization highlights the close relationship between the Thomas-Wigner rotation and the relativity of simultaneity. Furthermore, it is illustrated that accelerated motion implies the formation of an event horizon. The event horizons associated with the five boosts constitute a boundary to the rotated Born-rigid object and ensure its finite size.' address: 'Caputh, Germany' author: - 'G. Beyerle' title: 'Visualization of Thomas-Wigner rotations' --- Introduction ============ In 1926 the British physicist Llewellyn Hilleth Thomas (1903–1992) resolved a discrepancy between observed line splittings of atomic spectra in an external magnetic field (Zeeman effect) and theoretical calculations at that time [see e.g. @tomonaga97]. Thomas’ analysis [@thomas26; @thomas27] explains the observed deviations in terms of a special relativistic effect [@einstein05]. He recognized that a sequence of two non-collinear pure Lorentz transformations (boosts) cannot be expressed as one single boost. Rather, two non-collinear boosts correspond to a pure Lorentz transformation combined with a spatial rotation. This spatial rotation is known as Wigner rotation or Thomas-Wigner rotation, the corresponding rotation angle is the Thomas-Wigner angle [see e.g. @wigner39; @benmenahem85; @costella01; @cushing67; @ferraro99; @fisher72; @gelman90; @kennedy02; @mocanu92; @rhodes04; @rowe84; @ungar89; @ungar97; @gourgoulhon13; @misner73; @rebilas15; @steane12 and references therein]. The prevalent approach to visualize Thomas-Wigner rotations employs passive Lorentz transformations. An object $\mathcal{G}$ is simultaneously observed from two inertial reference frames, $[1]$ and $[N]$, and $\mathcal{G}$ is assumed to be at rest in both of them. Frame $[N]$ is related to $[1]$ by $N-1$ pure Lorentz transformations $$[1] \rightarrow [2] \rightarrow [3] \rightarrow \ldots \rightarrow [N] \quad .$$ As already mentioned, non-zero Thomas-Wigner rotations require two non-collinear boosts. For given non-collinear boosts $[1] \rightarrow [2]$ and $[2] \rightarrow [3]$ there exists a unique third boost $[3] \rightarrow [4]$, such that $\mathcal{G}$ is at rest with respect to both, frame $[1]$ and frame $[4]$. Thus, non-zero Thomas-Wigner rotations imply $N\ge4$. In the following frame $[1]$ is taken to be the laboratory frame. Following @jonsson07 in the present paper an alternative route to visualize Thomas-Wigner rotations using active or “physical” boosts is attempted. That is, $\mathcal{G}$ is accelerated starting from and ending at zero velocity in the laboratory frame (frame $[1]$). During its journey $\mathcal{G}$ performs several acceleration and/or deceleration manoeuvres. I suppose, that the mathematical complications created by using active boosts are outweighed by the visual impressions of $\mathcal{G}$ moving through the series of acceleration phases and finally coming to rest in a rotated orientation (see fig. \[fg:trjboostobj\] below). The paper is sectioned as follows. First, the approach is described in general terms and the basic assumptions are introduced. Second, uniform accelerations of Born-rigid objects are recalled. In the following section sequences of uniform, non-collinear accelerations for a given reference point and the motions of its adjacent grid vertices are calculated. The last two sections present the visualization results and discuss their implications. Details of the computer algebraic calculations performed in this study are given in \[se:appendix\]. For simplicity length units of light-seconds, abbreviated “ls” (roughly 300,000 km) are being used; in these units the velocity of light is $1~\text{ls}/\text{s}$. Method {#se:method} ====== We consider the trajectory of a square-shaped grid $\mathcal{G}$. The grid consists of $M$ vertices and is assumed to be Born-rigid, i.e. the distances between all grid points, as observed in the instantaneous comoving inertial frame, also known as the momentarily comoving inertial frame, remain constant [@born09]. $\mathcal{G}$’s central point $R$ serves as the reference point. $R$ is uniformly accelerated for a predefined proper time period. To obtain a closed trajectory several of these sections with constant proper acceleration, but different boost direction are joined together. In $R$’s instantaneous comoving frame the directions and magnitudes of the vertices’ proper accelerations $\vec{\alpha}_i$ ($i=1,\ldots,M$) change simultaneously and discontinuously at the switchover from one boost to the next. The subscript $i$ indicates that these accelerations are in general not the same for all vertex points. In reference frames other than the instantaneous comoving frame, the accelerations $\vec{\alpha}_i$ change asynchronously and $\mathcal{G}$, despite its Born-rigidity, appears distorted and twisted (see figure \[fg:trjboostobj\] below). On the other hand, $\mathcal{G}$’s Born-rigidity implies that the motion of the reference point $R$ determines the trajectories of all remaining $M-1$ vertices; it is sufficient to calculate $R$’s trajectory [@herglotz09; @noether10]. We note that the separations between individual switchover events are spacelike. I.e. the switchover events are causally disconnected and each vertex has to be “programmed” in advance to perform the required acceleration changes. In the following, $\alpha_R$ and $\Delta\tau_R$ denote the magnitude of the proper acceleration of $\mathcal{G}$’s reference point $R$ and the boost duration in terms of $R$’s proper time, respectively. To simplify the calculations we impose the following four conditions on all $N$ boosts. 1. \[as:bornrigid\] The grid $\mathcal{G}$ is Born-rigid. 2. \[as:posvelstartend\] At the start and after completion of the $N$th boost $\mathcal{G}$ is at rest in frame $[1]$; $R$ returns to its start position in frame $[1]$. 3. \[as:constacctime\] $R$’s proper acceleration $\alpha_R$ and the boost’s proper duration $\Delta\tau_R$ are the same in all $N$ sections. 4. \[as:geoplanar\] All boost directions and therefore all trajectories lie within the $xy$-plane. Let the unit vector $\hat{e}_{1}$ denote the direction of the first boost in frame $[1]$. This first boost lasts for a proper time $\Delta\tau_R$, as measured by $R$’s clock, when $R$ attains the final velocity $v_R \equiv \beta$ with respect to frame $[1]$. Frame $[2]$ is now defined as $R$’s comoving inertial frame at this instant of time. The corresponding Lorentz matrix transforming a four-vector from frame $[2]$ to frame $[1]$ is $$\begin{aligned} \label{eq:lotra1to2} \Lambda(\gamma,\hat{e}_1) \equiv \renewcommand*{\arraystretch}{1.5} \begin{pmatrix} \gamma \; , & \gamma\,\beta\,\hat{e}^T_1 \\ \gamma\,\beta\,\hat{e}_1 \; , & \mathds{1}_{3{\times}3} + (\gamma-1)\,\hat{e}_1\cdot\hat{e}^T_1 \end{pmatrix} \; .\end{aligned}$$ Here, $\mathds{1}_{3{\times}3}$ is the $3{\times}3$ unit matrix, the superscript $T$ denotes transposition, the Lorentz factor is $$\begin{aligned} \label{eq:defgamma} \gamma \equiv \frac{1}{\sqrt{1-\beta^2}}\end{aligned}$$ and, in turn, $\beta = \sqrt{\gamma^2-1} / \gamma$. Similarly, frame $[3]$ is $R$’s comoving inertial frame at the end of the second boost, etc. In general, the Lorentz transformation from frame $[i]$ to frame $[i+1]$ is given by eqn. \[eq:lotra1to2\], with $\hat{e}_{1}$ replaced by $\hat{e}_{i}$, the direction of the $i$th boost in frame $[i]$. Assumption \[\[as:constacctime\]\] implies that the sole unknowns, which need to be determined, are the angles between consecutive boosts, $$\begin{aligned} \label{eq:defbeta} \zeta_{i,i+1} &\equiv& \arccos \left( \hat{e}_i^T \cdot \hat{e}_{i+1} \right) \quad .\end{aligned}$$ The proper acceleration $\alpha_R$ and boost duration $\Delta\tau_R$ are taken to be given parameters. In the following the boost angles $\zeta$ will be replaced by the “half-angle” parametrization $$\begin{aligned} \label{eq:halfangleparm} T \equiv \tan\left(\frac{\zeta}{2}\right) \quad .\end{aligned}$$ Eqn. \[eq:halfangleparm\] allows us to write expressions involving $\sin(\zeta)$ and $\cos(\zeta)$ as polynomials in $T$ since $$\begin{aligned} \label{eq:halfanglesincos} \sin(\zeta) &=& \frac{2\,T} {1 + T^2} \\ \cos(\zeta) &=& \frac{1 - T^2} {1 + T^2} \quad . \nonumber\end{aligned}$$ We will find that a) no solutions exist if the number of boosts $N$ is four or less, b) for $N = 5$ the solution is unique and c) the derived boost angles $\zeta$ depend solely on the selected value of $\gamma = 1/\sqrt{1-\beta^2}$. Changing $\alpha_R$ and/or $\Delta\tau_R$ only affects the spatial and temporal scale of $R$’s trajectory (see below). In section \[se:sequniformacc\] it is shown that at least $N=5$ boost are necessary, in order to satisfy assumptions \[\[as:bornrigid\]\], \[\[as:posvelstartend\]\], \[\[as:constacctime\]\] and \[\[as:geoplanar\]\]. The derivations of $\zeta_{i,i+1} = \zeta_{i,i+1}(\gamma)$ are simplified by noting that the constraints \[\[as:posvelstartend\]\], \[\[as:constacctime\]\] and \[\[as:geoplanar\]\] imply time reversal invariance. I.e. $R$’s trajectory from destination to start followed backward in time is a valid solution as well and therefore $\zeta_{i,i+1} = \zeta_{N-i,N-i+1}$ for $i=1,\ldots,N-1$. Thus, for $N=5$ the number of unknown reduces from four to two, $\zeta_{1,2} = \zeta_{4,5}$ and $\zeta_{2,3} = \zeta_{3,4}$. Uniform acceleration of a Born-rigid object {#se:uniformaccmot} =========================================== Consider the uniform acceleration of the reference point $R$. We assume, the acceleration phase lasts for the proper time period $\Delta\tau_R$. During $\Delta\tau_R$ the reference point moves from location $\vec{r}_R(0)$ to location $$\begin{aligned} \label{eq:posattauA} \vec{r}_R(\Delta\tau_R) &=& \vec{r}_R(0) + \frac{1}{\alpha_R} \, \left( \cosh(\alpha_R\,\Delta\tau_R) - 1 \right) \, \hat{e}_B\end{aligned}$$ in the laboratory frame with unit vector $\hat{e}_B$ denoting the boost direction [see e.g. @hamilton78; @hobson06; @misner73; @rindler06; @steane12; @styer07]. The coordinate time duration ${\Delta}t_R$ corresponding to the proper time duration $\Delta\tau_R$ is $$\begin{aligned} \label{eq:timattauA} {\Delta}t_R &=& \frac{1}{\alpha_R}\,\sinh(\alpha_R\,\Delta\tau_R)\end{aligned}$$ and $R$’s final speed in the laboratory frame is $$\begin{aligned} \label{eq:velattauA} v_R = \tanh(\alpha_R\,\Delta\tau_R) = \beta \quad .\end{aligned}$$ Let $G$ be another vertex point of the grid $\mathcal{G}$ at location $\vec{r}_G(0)$ and $$\begin{aligned} \label{eq:defineofsb} b \equiv \left(\vec{r}_G(0) - \vec{r}_R(0) \right) \cdot \hat{e}_B\end{aligned}$$ the projection of distance vector from $R$ to $G$ onto the boost direction $\hat{e}_B$. In the laboratory frame the vertices $G$ and $R$ start to accelerate simultaneously since $\mathcal{G}$ is Born-rigid and analogous to eqns. \[eq:posattauA\], \[eq:timattauA\] and \[eq:velattauA\] we obtain for $G$’s trajectory $$\begin{aligned} \label{eq:postimattauB} \vec{r}_G(\Delta\tau_G) &=& \vec{r}_G(0) + \left( \frac{1}{\alpha_G} \, \left( \cosh(\alpha_G\,\Delta\tau_G) - 1 \right) + b \right) \, \hat{e}_B \\ {\Delta}t_G &=& \frac{1}{\alpha_G}\, \sinh(\alpha_G\,\Delta\tau_G) \nonumber \\ v_G &=& \tanh(\alpha_G\,\Delta\tau_G) \quad . \nonumber\end{aligned}$$ We note, that the boost phase ends simultaneously for all grid points in $R$’s instantaneous comoving frame. Since $\mathcal{G}$ is Born-rigid (assumption \[\[as:bornrigid\]\]), their speeds at the end of the boost phase have to be identical; in particular, $v_G = v_R$ and thus $$\begin{aligned} \label{eq:alphatautimesalphatau} \alpha_G\,\Delta\tau_G &=& \alpha_R\,\Delta\tau_R \quad .\end{aligned}$$ Furthermore, $\mathcal{G}$’s Born-rigidity implies that the spatial distance between $G$ and $R$ in $R$’s comoving rest frame at the end of the boost phase is the same as the distance at the beginning. A brief calculation leads to $$\begin{aligned} \label{eq:timspatialdist0} -\frac{1}{\alpha_G} \, (\gamma-1) + b\,\gamma + \frac{1}{\alpha_R} \, (\gamma-1) = b\end{aligned}$$ which simplifies to $$\begin{aligned} \label{eq:alphatfromalphar} \alpha_G &=& \frac{1}{1 + b \, \alpha_R} \, \alpha_R\end{aligned}$$ and with eqn. \[eq:alphatautimesalphatau\] $$\begin{aligned} \label{eq:tautfromtaur} \Delta\tau_G &=& \left(1 + b \, \alpha_R\right) \, \Delta\tau_R \quad .\end{aligned}$$ Eqn. \[eq:alphatfromalphar\] expresses the well-known fact that the proper accelerations aboard a Born-rigid grid may differ from one vertex to the next. More specifically, at a location trailing the reference point $R$ the acceleration exceeds $\alpha_R$, vertex points leading $R$ accelerate less than $\alpha_R$. (In relativistic space travel the passengers in the bow of the spaceship suffer lower acceleration forces than those seated in the stern. This amenity of a more comfortable acceleration, however, is counterbalanced by faster ageing of the space passengers (eqn. \[eq:tautfromtaur\]). Here it is assumed, that relativistic spaceships are Born-rigidly constructed.) This position-dependent acceleration is well-known from the Dewan-Beran-Bell spaceship paradox [@dewan59; @dewan63] and [@bell04 chapter 9]. Two spaceships, connected by a Born-rigid wire, accelerate along the direction separating the two. According to eqn. \[eq:alphatfromalphar\] the trailing ship has to accelerate faster than the leading one. Conversely, if both accelerated at the same rate in the laboratory frame, Born-rigidity could not be maintained and the wire connecting the two ships would eventually break. This well-known, but admittedly counterintuitive fact is not a paradox in the true sense of the word as discussed extensively in the relevant literature . Eqns. \[eq:alphatfromalphar\] and \[eq:tautfromtaur\] also imply, that $\alpha_G \rightarrow \infty$ and $\Delta\tau_G \rightarrow 0$, as the distance between a (trailing) vertex $G$ and the reference point $R$ approaches the critical value $$\begin{aligned} \label{eq:critdist} b^* \equiv -1/\alpha_R \quad .\end{aligned}$$ Clearly, the object $\mathcal{G}$ cannot extend beyond this boundary, which in the following is referred to as “event horizon”. In section \[su:eventhorizon\] we will discuss its consequences. Finally, we note that eqn. \[eq:tautfromtaur\] implies that a set of initially synchronized clocks mounted on a Born-rigid grid will in general fall out of synchronization once the grid is accelerated. Thus, the switchover events, which occur simultaneous in $R$’s instantaneous comoving frame, are not simultaneous with respect to the time displayed by the vertex clocks. Since switchover events are causally not connected and lie outside of each others’ lightcone, the acceleration changes have to be “programmed” into each vertex in advance [@eriksen82]. Sequence of uniform accelerations {#se:sequniformacc} ================================= In the previous section $R$’s trajectory throughout a specific acceleration phase (eqn. \[eq:postimattauB\]) was discussed. Now several of these segments are linked together to form a closed trajectory for $R$. Let $\text{A}^{[k]}$ denote $R$’s start event as observed in frame $[k]$ and $\text{B}^{[k]}$, $\text{C}^{[k]}$, etc.correspondingly denote the “switchover” events between $1^\text{st}$ and $2^\text{st}$ boost, $2^\text{nd}$ and $3^\text{rd}$ boost, etc., respectively. The bracketed superscripts indicate the reference frame. Frame $[1]$, i.e. $k=1$, is the laboratory frame, frame $[2]$ is obtained from frame $[1]$ using Lorentz transformation $\Lambda(\gamma,-\hat{e}_1)$ (eqn. \[eq:lotra1to2\]). Generally, frame $[k+1]$ is calculated from frame $[k]$ using the transformation matrix $\Lambda(\gamma,-\hat{e}_k)$. First, we determine the smallest number of boosts that satisfies the four assumptions listed in section \[se:method\]. Denoting the number of boosts by $N$, it is self-evident that $N \ge 3$, since for $N = 1$ the requirement of vanishing final velocity cannot be met if $v_R \ne 0$. And for $N=2$ the requirement of vanishing final velocity implies collinear boost directions. With two collinear boosts, however, the reference point $R$ does not move along a closed trajectory. In addition, we note, that collinear boosts imply vanishing Thomas-Wigner rotation [see e.g. @steane12]. Three boosts {#su:threeboosts} ------------ Consider three boosts of the reference point $R$ starting from location $\text{A}$ and returning to location $\text{D}$ via locations $\text{B}$ and $\text{C}$. In the laboratory frame (frame $[1]$) the four-position at the destination $\text{D}$ is given by $$\begin{aligned} \label{eq:threeboostpos} \textbf{\textit{P}}_\text{D}^{[1]} &=& \textbf{\textit{P}}_\text{A}^{[1]} + \textbf{\textit{S}}_{\text{A}{\rightarrow}\text{B}}^{[1]} \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \textbf{\textit{S}}_{\text{B}{\rightarrow}\text{C}}^{[2]} \nonumber \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \textbf{\textit{S}}_{\text{C}{\rightarrow}\text{D}}^{[3]} \nonumber\end{aligned}$$ and $$\begin{aligned} \label{eq:threeboostvel} \textbf{\textit{V}}_\text{D}^{[1]} &=& \Lambda \left(\gamma,-\hat{e}_1 \right) \cdot \, \Lambda \left(\gamma,-\hat{e}_2 \right) \cdot \, \Lambda \left(\gamma,-\hat{e}_3 \right) \cdot \, \textbf{\textit{V}}_\text{D}^{[4]}\end{aligned}$$ is the corresponding four-velocity. Here, the four-vector $$\begin{aligned} \label{eq:threeboostdeltapos} \textbf{\textit{S}}_{\text{A}{\rightarrow}\text{B}}^{[1]} &\equiv& \frac{1}{\alpha_R} \, \begin{pmatrix} \sinh(\alpha_R\,\Delta\tau_R) \\ (\cosh(\alpha_R\,\Delta\tau_R)-1)\,\hat{e}_1 \end{pmatrix} \\ &=& \frac{1}{\alpha_R} \, \begin{pmatrix} \gamma\,\beta \\ (\gamma-1)\,\hat{e}_1 \end{pmatrix} \nonumber\end{aligned}$$ describes $R$’s trajectory in spacetime from $\text{A}$ to $\text{B}$ (eqn. \[eq:postimattauB\]). $\textbf{\textit{S}}_{\text{B}{\rightarrow}\text{C}}^{[2]}$ and $\textbf{\textit{S}}_{\text{C}{\rightarrow}\text{D}}^{[3]}$ are defined correspondingly. Assumption \[\[as:posvelstartend\]\] implies that $$\begin{aligned} \label{eq:posinit3boost} \vec{P}_\text{A}^{[1]} = \vec{P}_\text{D}^{[1]} = \vec{P}_\text{A}^{[4]} = \vec{P}_\text{D}^{[4]} = \vec{0}\end{aligned}$$ and $$\begin{aligned} \label{eq:velinit3boost} \textbf{\textit{V}}_\text{A}^{[1]} = \textbf{\textit{V}}_\text{D}^{[1]} = \textbf{\textit{V}}_\text{A}^{[4]} = \textbf{\textit{V}}_\text{D}^{[4]} = \begin{pmatrix} 1 \\ \vec{0} \end{pmatrix} \quad .\end{aligned}$$ Inserting eqn. \[eq:lotra1to2\] into eqn. \[eq:threeboostvel\] yields $$\begin{aligned} \label{eq:threebooststvalsol} T_{12} &=& T_{23} = \pm\sqrt{2\,\gamma+1}\end{aligned}$$ (see \[se:appendix\]) and, in turn, using eqn. \[eq:posinit3boost\] we obtain $$\begin{aligned} \label{eq:threeboostspossol} \vec{P}_\text{D}^{[1]} = \frac{1}{(\gamma + 1)^2} \begin{pmatrix} -(\gamma - 1)^2\,(2\,\gamma + 1)^2 \\ -(\gamma - 1)\,(2\gamma + 1)^\frac{3}{2}\,(3\gamma + 1) \\ 0 \end{pmatrix} \overset{!}{=} \vec{0} \quad .\end{aligned}$$ Its only solution for real-valued $\beta$ is the trivial solution $\gamma=1$, i.e. $\beta=0$. Thus, there are no non-trivial solutions for $N=3$ boosts, which are consistent with the assumptions \[\[as:bornrigid\]\] to \[\[as:geoplanar\]\]. Four boosts {#su:fourboosts} ----------- For a sequence of four boosts time reversal symmetry implies that $R$’s velocity in the laboratory frame vanishes at event $\text{C}$ after the second boost, i.e. $\vec{\text{V}}_\text{C}^{[1]} = \vec{0}$. However, stationarity in the laboratory frame can only be achieved if the first two boosts $\text{A} \rightarrow \text{B}$ and $\text{B} \rightarrow \text{C}$ are collinear. In order to fulfil assumption \[\[as:posvelstartend\]\] the third and fourth boosts have to be collinear with the first (and second) boost as well. As already noted, a sequence of collinear boosts, however, does not produce a non-zero Thomas-Wigner rotation. Five boosts {#su:fiveboosts} ----------- For a sequence of five boosts, i.e. $N=5$, the expressions for four-position and four-velocity are $$\begin{aligned} \label{eq:fiveboostpos} \textbf{\textit{P}}_\text{F}^{[1]} &=& \textbf{\textit{P}}_\text{A}^{[1]} + \textbf{\textit{S}}_{\text{A}{\rightarrow}\text{B}}^{[1]} \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \textbf{\textit{S}}_{\text{B}{\rightarrow}\text{C}}^{[2]} \nonumber \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \textbf{\textit{S}}_{\text{C}{\rightarrow}\text{D}}^{[3]} \nonumber \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_3\right) \cdot \, \textbf{\textit{S}}_{\text{D}{\rightarrow}\text{E}}^{[4]} \nonumber \\ && + \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_3\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_4\right) \cdot \, \textbf{\textit{S}}_{\text{E}{\rightarrow}\text{F}}^{[5]} \nonumber\end{aligned}$$ and $$\begin{aligned} \label{eq:fiveboostvel} \textbf{\textit{V}}_\text{F}^{[1]} &=& \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_3\right) \\ && \cdot \, \Lambda\left(\gamma,-\hat{e}_4\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_5\right) \cdot \, \textbf{\textit{V}}_\text{F}^{[6]} \quad , \nonumber\end{aligned}$$ respectively. Analogous to eqns. \[eq:posinit3boost\] and \[eq:velinit3boost\] assumption \[\[as:posvelstartend\]\] implies that $$\begin{aligned} \label{eq:posinit5boost} \vec{P}_\text{A}^{[1]} = \vec{P}_\text{F}^{[1]} = \vec{P}_\text{A}^{[6]} = \vec{P}_\text{F}^{[6]} = \vec{0}\end{aligned}$$ and $$\begin{aligned} \label{eq:velinit5boost} \textbf{\textit{V}}_\text{A}^{[1]} = \textbf{\textit{V}}_\text{F}^{[1]} = \textbf{\textit{V}}_\text{A}^{[6]} = \textbf{\textit{V}}_\text{F}^{[6]} = \begin{pmatrix} 1 \\ \vec{0} \end{pmatrix} \quad .\end{aligned}$$ To simplify the expressions in eqns. \[eq:fiveboostpos\] and \[eq:fiveboostvel\] time reversal symmetry is invoked again. It implies that the set of boost vectors $-\hat{e}_5, -\hat{e}_4, \ldots, -\hat{e}_1$ constitutes a valid solution, provided $\hat{e}_1, \hat{e}_2, \ldots, \hat{e}_5$ is one and satisfies assumptions \[\[as:bornrigid\]\] to \[\[as:geoplanar\]\]. Thereby the number of unknown is reduced from four to two, the angle between the boost vectors $\hat{e}_1$ and $\hat{e}_2$, and the angle between $\hat{e}_2$ and $\hat{e}_3$ $$\begin{aligned} \label{eq:defbeta123} \zeta_{1,2} &\equiv& \arccos\left( \hat{e}_1^T \cdot \hat{e}_2\right) = \arccos\left( \hat{e}_4^T \cdot \hat{e}_5\right) \\ \zeta_{2,3} &\equiv& \arccos\left( \hat{e}_2^T \cdot \hat{e}_3\right) = \arccos\left( \hat{e}_3^T \cdot \hat{e}_4\right) \quad . \nonumber\end{aligned}$$ Fig. \[fg:trjbstrefptfr1and6\] illustrates the sequence of the five boosts in the laboratory frame (frame $[1]$). $R$ moves along a closed trajectory starting at location $\text{A}$ and returning to $\text{F}$ via $\text{B}$, $\text{C}$, $\text{D}$ and $\text{E}$. Since the start and final velocities are zero, the motion between $\text{A}$ and $\text{B}$ and, likewise, between $\text{E}$ and $\text{F}$ is rectilineal. In contrast, the trajectory connecting $\text{B}$ and $\text{E}$ (via $\text{C}$ and $\text{D}$) appears curved in frame $[1]$. As discussed and illustrated below the curved paths appear as straight lines in the corresponding frame (see fig.\[fg:trjboostrefpt\]). From eqns. \[eq:fiveboostvel\] and \[eq:velinit5boost\] it follows that $$\label{eq:t12dott23} \left( (T_{12})^2 - 2 \, \gamma - 1\right) \, (T_{23})^2 - 4 \, (1+\gamma) \, T_{12} \, T_{23} + (T_{12})^2 + 4 \, \gamma^2 + 2 \, \gamma - 1 = 0$$ with the two unknowns $T_{12}$ and $T_{23}$ (for details see \[se:appendix\]). Eqn. \[eq:t12dott23\] has the two solutions $$\begin{aligned} \label{eq:t23fromt12} T_{23}^{(\pm)} &=& \frac{1}{- (T_{12})^2 + 2 \gamma + 1} \\ && \times \bigg(- 2\,T_{12}\,\left(\gamma + 1\right) \nonumber \\ && \quad \pm \sqrt{- (T_{12})^4 + 8\,(T_{12})^2 \gamma + 6\,(T_{12})^2 + 8\,\gamma^3 + 8\,\gamma^2 - 1} \bigg) \nonumber\end{aligned}$$ provided $$\begin{aligned} \label{eq:t12exprneqzero} (T_{12})^2 - 2 \gamma - 1 \ne 0 \quad .\end{aligned}$$ Assumption \[\[as:posvelstartend\]\] implies that $\vec{P}^{[1]}_\text{F} = 0$, i.e. the spatial component of the event $\textbf{\textit{P}}_\text{F}^{[1]}$ vanishes. Since all motions are restricted to the $xy$-plane, it suffices to consider the $x$- and $y$-components of eqn. \[eq:posinit5boost\]. The $y$-component leads to the equations (see \[se:appendix\]) $$\begin{aligned} \label{eq:t12t23first} && (T_{12})^{8} \\ && + (T_{12})^{6} \, 4 \, \left(\gamma + 2\right) \nonumber \\ && + (T_{12})^4 \, (-2) \, \left(2\,\gamma+1\right) \, \left(2\,\gamma^2 + 8\,\gamma + 9\right) \nonumber \\ && + (T_{12})^2 \, (-4) \, \left(8 \gamma^4 + 28 \gamma^3 + 26 \gamma^2 + 5 \gamma - 2\right) \nonumber \\ && - \left(2 \gamma + 1\right)^3 \, \left(4\,\gamma^2 + 2 \gamma - 1\right) \nonumber \\ &=& 0 \nonumber\end{aligned}$$ or $$\begin{aligned} \label{eq:t12t23second} && (T_{12})^{8} \, \left(\gamma^2 + 6\,\gamma + 9 \right) \\ && + (T_{12})^{6} \, (-4) \, \left(6 \, \gamma^3 - 15\,\gamma^2 - 12\,\gamma + 5 \right) \nonumber \\ && + (T_{12})^4 \, (-2) \, \left(24\,\gamma^4 - 44\,\gamma^3 - 55\,\gamma^2 + 26\,\gamma + 1 \right) \nonumber \\ && + (T_{12})^2 \, (-4) \, \left( 8\,\gamma^5 - 16\,\gamma^4 - 14\,\gamma^3 + 5\,\gamma^2 + 1 \right) \nonumber \\ && + \left( 4 \, \gamma^2 + \gamma - 1 \right)^2 \nonumber \\ &=& 0 \quad . \nonumber\end{aligned}$$ For $\gamma=1$ eqn. \[eq:t12t23second\] yields $$\begin{aligned} \label{eq:t12t23atunitgamma} (T_{12})^8 + 4\,(T_{12})^6 + 6\,(T_{12})^4 + 4\,(T_{12})^2 + 1 = 0\end{aligned}$$ which has no real-valued solution for $T_{12}$. It turns out (see \[se:appendix\]) that equating the $x$-component of $\vec{P}^{[1]}_\text{F}=0$ to zero, results in an expression containing two factors as well, one of which is identical to the LHS of eqn. \[eq:t12t23first\]. Thus, eqn. \[eq:t12t23first\] satisfies the condition $\vec{P}^{[1]}_\text{F}=0$. Eqn. \[eq:t12t23first\] is a polynomial equation of degree four in terms of $(T_{12})^2$. Its solutions are classified according to the value of the discriminant $\Delta$, which for eqn. \[eq:t12t23first\] evaluates to (see e.g. [en.wikipedia.org/wiki/Quartic\_function](en.wikipedia.org/wiki/Quartic_function)) $$\begin{aligned} \label{eq:pol4degdelta} \Delta &=& -524288 \, \gamma \,(\gamma - 1)^3 \,(\gamma + 1)^7 \\ && \times\,(4\,\gamma^4 + 28\,\gamma^3 + 193\,\gamma^2 + 234\,\gamma + 81) \quad . \nonumber\end{aligned}$$ For non-trivial boost $\gamma>1$, the discriminant is negative and therefore the roots of the quartic polynomial consist of two pairs of real and complex conjugate numbers. The real-valued solutions are $$\begin{aligned} \label{eq:solvepol4deg} (T_{12})^2 &=& - (\gamma+2) + \mathcal{S} + \frac{1}{2} \, \sqrt{-4\,\mathcal{S}^2 - 2\,p - \frac{q}{\mathcal{S}}}\end{aligned}$$ and $$\begin{aligned} \label{eq:solvepol4degD} \left(T_{12}^{(b)}\right)^2 &=& - (\gamma+2) + \mathcal{S} - \frac{1}{2} \, \sqrt{-4\,\mathcal{S}^2 - 2\,p - \frac{q}{\mathcal{S}}}\end{aligned}$$ with $$\begin{aligned} \label{eq:pol4pars} p &\equiv& -2 \, \left( \gamma + 1 \right) \, \left( 4\gamma^2 + 17 \gamma + 21 \right) \\ q &\equiv& -16 \, \left( \gamma + 1 \right) \, \left( \gamma^2 - 2 \gamma -9 \right) \nonumber \\ \mathcal{S} &\equiv& \frac{1}{2\,\sqrt{3}} \, \sqrt{ -2 \, p + \mathcal{Q} + \frac{\Delta_0}{\mathcal{Q}} } \nonumber \\ \Delta_0 &\equiv& 16 \, \left( \gamma + 1 \right)^2 \, \, \left( 4\,\gamma^4 + 28\,\gamma^3 + 157\,\gamma^2 + 126\,\gamma + 9 \right) \nonumber \\ \mathcal{Q} &\equiv& 4 \, \sqrt[3]{ \mathcal{Q}_0 + 12 \, \sqrt{6} \, \sqrt{\mathcal{Q}_0} } \nonumber \\ \mathcal{Q}_0 &\equiv& \gamma \, ( \gamma - 1 )^3 \, ( \gamma + 1 )^7 \, ( 4 \, \gamma^4 + 28 \, \gamma^3 + 193 \, \gamma^2 + 234 \, \gamma + 81 ) \quad . \nonumber\end{aligned}$$ The solution listed in eqn. \[eq:solvepol4degD\] turns out to be negative and thus does not produce a real-valued solution for $T_{12}$. The remaining two solutions of eqn. \[eq:t12t23first\] $$\begin{aligned} \label{eq:t12othersols} \left(T_{12}^{(c)}\right)^2 &=& - (\gamma+2) - \mathcal{S} + \frac{1}{2} \, \sqrt{-4\,\mathcal{S}^2 - 2\,p + \frac{q}{\mathcal{S}}} \\ \left(T_{12}^{(d)}\right)^2 &=& - (\gamma+2) - \mathcal{S} - \frac{1}{2} \, \sqrt{-4\,\mathcal{S}^2 - 2\,p + \frac{q}{\mathcal{S}}} \nonumber\end{aligned}$$ correspond to replacing $\mathcal{S} \rightarrow -\mathcal{S}$ in eqns. \[eq:solvepol4deg\] and \[eq:solvepol4degD\], are complex-valued and therefore disregarded as well. The other unknown, $T_{23}$, follows from eqn. \[eq:t23fromt12\] by choosing the positive square root $+\sqrt{(T_{12})^2}$ and using $$\begin{aligned} \label{eq:t23fromt12sel} T_{23} &\equiv& T_{23}^{(+)}\end{aligned}$$ (see eqn. \[eq:t12dott23\]). For a given Lorentz factor $\gamma$ the angles between the boost directions $\hat{e}_i$ and $\hat{e}_{i+1}$ follow from $$\begin{aligned} \label{eq:zeta1fromgamma} \zeta_{1,2}(\gamma) &=& \arccos\left( \hat{e}^T_1 \cdot \hat{e}_2 \right) = 2 \, \arctan\left( +\sqrt{(T_{12}(\gamma))^2} \right) \\ \zeta_{2,3}(\gamma) &=& \arccos\left( \hat{e}^T_2 \cdot \hat{e}_3 \right) = 2 \, \arctan\left( +\sqrt{(T_{23}(\gamma))^2} \right) \nonumber\end{aligned}$$ and, with $\zeta_{4,5} = \zeta_{1,2}$ and $\zeta_{3,4} = \zeta_{2,3}$, the orientation of the five boost directions $\hat{e}_i$ for $i=1,\ldots,5$ within the $xy$-plane are obtained. ![ \[fg:boostanglesvsgamma\] The angle between the boost direction vectors $\hat{e}_1$ and $\hat{e}_2$ in frame $[1]$ (blue line), and the angle between $\hat{e}_2$ and $\hat{e}_3$ in frame $[2]$ (red) as a function of $\gamma$. The dotted line marks $+180^\circ$, the limit of $\zeta_{1,2}$ and $\zeta_{2,3}$ for $\gamma\rightarrow\infty$. ](memo44244-1-100000-1000-6.eps) Fig. \[fg:boostanglesvsgamma\] shows numerical values of the boost angles $\zeta_{1,2}$ and $\zeta_{2,3}$ as a function of $\gamma$. The angles increase from $$\zeta_{1,2}(\gamma=1) = 2 \, \arctan\left(\sqrt{-5 + 4\,\sqrt{10}}\right) \approx 140.2^\circ$$ and $$\zeta_{2,3}(\gamma=1) = 2 \, \arctan\left( \frac{\sqrt{-5 + 4\,\sqrt{10}} - \sqrt{3}\,\sqrt{-5 + 2\,\sqrt{10}}} {-2 + \sqrt{10}} \right) \approx 67.2^\circ$$ at $\gamma=1$ to $\zeta_{1,2}(\gamma \rightarrow \infty) = +180^\circ = \zeta_{2,3}(\gamma \rightarrow \infty)$ as $\gamma \rightarrow \infty$. Fig. \[fg:normboostdirs\] depicts the orientation of the five boost directions for several values of $\gamma$. For illustrative purposes the first boost vector $\hat{e}_1$ is taken to point along the $x$-axis. We note that the panels in fig. \[fg:normboostdirs\] do not represent a specific reference frame; rather, each vector $\hat{e}_k$ is plotted with respect to frame $[k]$ ($k=1,\ldots,5$). The four panels show the changes in boost directions for increasing values of $\gamma$. Interestingly, the asymptotic limits $\zeta_{1,2}(\gamma\rightarrow\infty) = +180^\circ$ and $\zeta_{2,3}(\gamma\rightarrow\infty) = +180^\circ$ imply that in the relativistic limit $\gamma\rightarrow\infty$ the trajectory of $R$ essentially reduces to one-dimensional motions along the $x$-axis. At the same time the Thomas-Wigner rotation angle increases to $+360^\circ$ as $\gamma\rightarrow\infty$ (see the discussion in section \[se:discuss\] below). ![ \[fg:normboostdirs\] Boost directions for four different values of $\gamma = 1 / \sqrt{1-\beta^2}$. $\hat{e}_k$ is assumed to point along the $x$-axis. In the relativistic limit $\gamma \rightarrow \infty$ the angles between $\hat{e}_k$ and $\hat{e}_{k+1}$ approach $+180^\circ$ and the motion of the reference point $R$ tends to be more and more restricted along the $x$-axis (see panel (4)), i.e. in the relativistic limit the object’s trajectory transitions from a two- to a one-dimensional motion. ](memo44208-1.eps) Since the accelerated object is Born-rigid, the trajectories of all grid vertices $G$ are uniquely determined once the trajectory of the reference point $R$ is known [@herglotz09; @noether10; @eriksen82]. Following the discussion in section \[se:uniformaccmot\] the position and coordinate time of a random vertex $G$, in the frame comoving with $R$ at the beginning of the corresponding acceleration phase, follows from eqns. \[eq:postimattauB\]. The resulting trajectories are discussed in the next section. ![ \[fg:trjbstrefptfr1and6\] Trajectories of reference point $R$ for $\gamma = 2/\sqrt{3} \approx 1.15$ as seen from reference frames $[1]$ (laboratory frame) and $[6]$. The two frames are stationary with respect to each other, but rotated by a Thomas-Wigner angle of $\theta_{TW} = 14.4^\circ$. ](memo44069-50-3.eps) Visualization {#se:visual} ============= The trajectory of the reference point $R$ within the $xy$-plane for a boost velocity of $\beta = 1/2$, corresponding to $\gamma = 2/\sqrt{3} \approx 1.15$, is displayed in fig. \[fg:trjbstrefptfr1and6\]. It is plotted in the laboratory frame (frame $[1]$) as black solid line. The same trajectory as it appears to an observer in frame $[6]$ is marked in grey. The two frames are stationary with respect to other, but rotated by a Thomas-Wigner angle of about $14.4^\circ$. In addition, dots show the spatial component of the four switchover events $\text{B}$, $\text{C}$, $\text{D}$ and $\text{E}$ in the two frames. As required by assumption \[\[as:posvelstartend\]\] the start and final positions, corresponding to the events $\text{A}$ and $\text{F}$, coincide. ![ \[fg:trjboostrefpt\] Trajectories of the reference point $R$ as seen from the six reference frames $[1],[2],\ldots, [6]$. The switchover points are marked by $X_i^{[k]}$ with $X = \text{A},\ldots,\text{F}$. Corresponding switchover points are connected by dashed lines. The Lorentz factor is $\gamma = 2/\sqrt{3} \approx 1.15$. ](memo44069-50-1.eps) Fig. \[fg:trjboostrefpt\] shows the same trajectory as fig. \[fg:trjbstrefptfr1and6\]. In addition, $R$’s trajectories as recorded by observers in the frames $[2],\ldots,[5]$ are plotted as well (solid coloured lines). Corresponding switchover events are connected by dashed lines. At $\text{B}^{[2]}$, $\text{C}^{[3]}$, $\text{D}^{[4]}$ and $\text{E}^{[5]}$ (and of course at the start event $\text{A}^{[1,6]}$ and destination event $\text{F}^{[1,6]}$) the reference point $R$ slows down and/or accelerates from zero velocity producing a kink in the trajectory. In all other cases the tangent vectors of the trajectories, i.e. the velocity vectors are continuous at the switchover points. ![ \[fg:trjboostobj\] A series of grid positions as seen in the laboratory frame. The boost velocity is taken to be $\beta = 0.7$, the resulting Thomas-Wigner rotation angle amounts to about $33.7^\circ$. Coordinate time is displayed in the top right corner of each panel. The five boost phases are distinguished by colour. It is evident that switchovers between boosts do not occur simultaneously in the laboratory frame. The reference point (marked in red) moves along its trajectory counterclockwise, whereas the grid Thomas-Wigner rotates clockwise. For details see text. ](memo44243-70-6.eps) With eqns. \[eq:solvepol4deg\] and \[eq:t23fromt12sel\] all necessary ingredients to visualize the relativistic motion of a Born-rigid object are available. In fig. \[fg:trjboostobj\] the object is modelled as a square-shaped grid of $11\times11$ points, arranged around the reference point $R$. The object uniformly accelerates in the $xy$-plane changing the boost directions by the four angles $\zeta_{1,2}$ (as measured in frame $[2]$), $\zeta_{2,3}$ (measured in frame $[3]$), $\zeta_{2,3}$ (measured in frame $[4]$) and finally $\zeta_{1,2}$ (measured in frame $[5]$). The vertices’ colour code indicates the corresponding boost section. The 16 panels depict the grid positions in the laboratory frame (frame $[1]$) for specific values of coordinate time displayed in the top right. To improve the visual impression the magnitude of the Thomas-Wigner rotation in fig. \[fg:trjboostobj\] is enlarged by increasing the boost velocity from $\beta = 0.5$ in figs. \[fg:trjbstrefptfr1and6\] and \[fg:trjboostrefpt\] to $\beta = 0.7$ corresponding to $\gamma \approx 1.4$. Despite appearance the grid $\mathcal{G}$ is Born-rigid; in $R$’s comoving inertial frame the grid maintains its original square shape. In the laboratory frame, however, $\mathcal{G}$ appears compressed, when it starts to accelerate or decelerate and sheared, when one part of $\mathcal{G}$ has not yet finished boost $k$, but the remaining part of $\mathcal{G}$ already has transitioned to the next boost section $k+1$. This feature is clearly evident from panels (4), (7), (10) or (13) in fig. \[fg:trjboostobj\] with the occurrence of two colours indicating two boost sections taking effect at the same epoch of coordinate time. We note, however, that the switchover events occur simultaneously for all grid points in $R$’s comoving frame. The non-uniform colourings illustrate the asynchronism of the switchovers in the laboratory frame and thereby the relationship between Thomas-Wigner rotations and the non-existence of absolute simultaneity. Discussion {#se:discuss} ========== In this final section the Thomas-Wigner rotation angle is calculated from the known boost angles $\zeta_{1,2}(\gamma)$ and $\zeta_{2,3}(\gamma)$ (eqn. \[eq:zeta1fromgamma\]). Second, the question of size limits of Born-rigid objects, Thomas-Wigner-rotated by a series of boosts, is addressed. Derivation of Thomas-Wigner angle {#su:derivethomaswigner} --------------------------------- From the preceding sections follows a straightforward calculation of the Thomas-Wigner angle as a function of Lorentz factor $\gamma$. Assumption \[\[as:posvelstartend\]\] implies that the sequence of the five Lorentz transformations $[6] \rightarrow [5] \rightarrow \ldots \rightarrow [1]$ is constructed such that frame $[6]$ is stationary with respect to frame $[1]$ and their spatial origins coincide. I.e. the combined transformation reduces to an exclusively spatial rotation and the corresponding Lorentz matrix can be written as $$\begin{aligned} \label{eq:thomaswignerrotmatrix} && \Lambda\left(\gamma,-\hat{e}_1\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_2\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_3\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_4\right) \cdot \, \Lambda\left(\gamma,-\hat{e}_5\right) \\ &=& \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & R_{1,1} & R_{1,2} & R_{1,3} \\ 0 & R_{2,1} & R_{2,2} & R_{2,3} \\ 0 & R_{3,1} & R_{3,2} & R_{3,3} \end{pmatrix} \quad . \nonumber\end{aligned}$$ Since the rotation is confined to the $xy$-plane, $R_{3,i} = 0 = R_{i,3}$ with $i=1,2,3$. The remaining elements $$\begin{aligned} \label{eq:thomaswignerrotmatelem} R_{1,1}(\gamma) &=& R_{2,2}(\gamma) \equiv \cos(\theta_{TW}(\gamma)) \\ R_{1,2}(\gamma) &=& -R_{2,1}(\gamma) \equiv \sin(\theta_{TW}(\gamma)) \nonumber\end{aligned}$$ yield the Thomas-Wigner rotation angle $$\begin{aligned} \label{eq:thowigrotangle} \tilde{\theta}_{TW}(\gamma) \equiv \text{atan2}( R_{2,1}(\gamma), R_{1,1}(\gamma)) \\ \theta_{TW}(\gamma) \equiv \left\{ \begin{array} {c@{\quad:\quad}l} \tilde{\theta}_{TW}(\gamma) & \tilde{\theta}_{TW}(\gamma) \ge 0 \\ \tilde{\theta}_{TW}(\gamma) + 2\,\pi & \tilde{\theta}_{TW}(\gamma) < 0 \end{array} \right. \nonumber\end{aligned}$$ with $\text{atan2}(\cdot,\cdot)$ denoting the four-quadrant inverse tangent. The definition of $\theta_{TW}(\gamma)$ in eqn. \[eq:thowigrotangle\] ensures that angles exceeding $+180^\circ$ are unwrapped and mapped into the interval $[0^\circ,+360^\circ]$ (see fig. \[fg:thomaswigneranglevsgamma\]). Using eqn. \[eq:thomaswignerrotmatrix\] the rotation matrix elements $R_{1,1}$ and $R_{2,1}$, expressed in terms of $T_{12}$ and $T_{23}$ (eqn. \[eq:halfangleparm\]), are (see \[se:appendix\]) $$\begin{aligned} \label{eq:thwirot11} R_{1,1}(\gamma) &=& -1 + \frac{\left(\gamma + 1\right)} {\left((T_{12})^2 + 1\right)^4 \left((T_{23})^2 + 1\right)^4} \\ && \times \bigg((T_{12})^4 \, (T_{23})^4 + 2 \, (T_{12})^4 \, (T_{23})^2 + (T_{12})^4 \nonumber \\ && - 4 \, (T_{12})^3 \, (T_{23})^3 \, \gamma + 4 \, (T_{12})^3 \, (T_{23})^3 - 4 \, (T_{12})^3 \, T_{23} \, \gamma \nonumber \\ && + 4 \, (T_{12})^3 \, T_{23} - 2 \, (T_{12})^2 \, (T_{23})^4 \, \gamma + 4 \, (T_{12})^2 \, (T_{23})^4 \nonumber \\ && + 4 \, (T_{12})^2 \, (T_{23})^2 \, \gamma^2 + 8 \, (T_{12})^2 \, (T_{23})^2 \, \gamma - 8 \, (T_{12})^2 \, (T_{23})^2 \nonumber \\ && - 4 \, (T_{12})^2 \, \gamma^2 + 10 \, (T_{12})^2 \, \gamma - 4 \, (T_{12})^2 \nonumber \\ && + 12 \, T_{12} \, (T_{23})^3 \, \gamma - 12 \, T_{12} \, (T_{23})^3 - 16 \, T_{12} \, T_{23} \, \gamma^2 \nonumber \\ && + 12 \, T_{12} \, T_{23} \, \gamma + 4 \, T_{12} \, T_{23} + 2 \, (T_{23})^4 \, \gamma \nonumber \\ && - (T_{23})^4 - 4 \, (T_{23})^2 \, \gamma^2 + 6 \, (T_{23})^2 + 4 \, \gamma^2 - 2 \, \gamma - 1\bigg)^2 \nonumber\end{aligned}$$ and $$\begin{aligned} \label{eq:thwirot21} R_{2,1}(\gamma) &=& \frac{4 \left( \gamma - 1\right) \, \left(\gamma + 1\right)}{\left((T_{12})^2 + 1\right)^4 \left((T_{23})^2 + 1\right)^4} \\ && \times \bigg((T_{12})^3 \, (T_{23})^2 + (T_{12})^3 + 3 \, (T_{12})^2 \, (T_{23})^3 \nonumber \\ && - 2 \, (T_{12})^2 \, T_{23} \, \gamma + (T_{12})^2 \, T_{23} + T_{12} \, (T_{23})^4 \nonumber \\ && - 2 \, T_{12} \, (T_{23})^2 \, \gamma - 3 \, T_{12} \, (T_{23})^2 + 2 \, T_{12} \, \gamma \nonumber \\ && - (T_{23})^3 + 2 \, T_{23} \, \gamma + T_{23}\bigg) \nonumber \\ && \times \bigg((T_{12})^4 \, (T_{23})^4 + 2 \, (T_{12})^4 \, (T_{23})^2 + (T_{12})^4 \nonumber \\ && - 4 \, (T_{12})^3 \, (T_{23})^3 \, \gamma + 4 \, (T_{12})^3 \, (T_{23})^3 - 4 \, (T_{12})^3 \, T_{23} \gamma \nonumber \\ && + 4 \, (T_{12})^3 \, T_{23} - 2 \, (T_{12})^2 \, (T_{23})^4 \gamma + 4 \, (T_{12})^2 \, (T_{23})^4 \nonumber \\ && + 4 \, (T_{12})^2 \, (T_{23})^2 \gamma^2 + 8 \, (T_{12})^2 \, (T_{23})^2 \gamma - 8 \, (T_{12})^2 \, (T_{23})^2 \nonumber \\ && - 4 \, (T_{12})^2 \, \gamma^2 + 10 \, (T_{12})^2 \, \gamma - 4 \, (T_{12})^2 \nonumber \\ && + 12 \, T_{12} \, (T_{23})^3 \, \gamma - 12 \, T_{12} \, (T_{23})^3 - 16 \, T_{12} \, T_{23} \, \gamma^2 \nonumber \\ && + 12 \, T_{12} \, T_{23} \, \gamma + 4 \, T_{12} \, T_{23} + 2 \, (T_{23})^4 \, \gamma \nonumber \\ && - (T_{23})^4 - 4 \, (T_{23})^2 \, \gamma^2 + 6 (T_{23})^2 + 4 \gamma^2 - 2 \gamma - 1\bigg) \nonumber\end{aligned}$$ with $T_{12} = T_{12}(\gamma)$ and $T_{23} = T_{23}(\gamma)$ given by eqns. \[eq:solvepol4deg\] and \[eq:t23fromt12sel\], respectively. ![ \[fg:thomaswigneranglevsgamma\] Thomas-Wigner rotation angle as a function of $\gamma$. For clarity the angle has been unwrapped to the range $[0^\circ, +360^\circ]$. ](memo44244-1-100000-1000-5.eps) The resulting angle $\theta_{TW}(\gamma)$ as a function of $\gamma$ is plotted in fig. \[fg:thomaswigneranglevsgamma\]. The plot suggests that $\theta_{TW} \rightarrow +360^\circ$ as $\gamma \rightarrow \infty$. As already mentioned in subsection \[su:fiveboosts\] (see fig. \[fg:normboostdirs\]) the boost angles $\zeta_{1,2} \rightarrow +180^\circ$ and $\zeta_{2,3} \rightarrow +180^\circ$ in the relativistic limit $\gamma \rightarrow \infty$. Notwithstanding that the $R$’s trajectory reduces to an one-dimensional motion along the $x$-axis as $\gamma \rightarrow \infty$, the grid’s Thomas-Wigner rotation angle approaches a full revolution of $+360^\circ$ in the laboratory frame. Event horizons {#su:eventhorizon} -------------- As illustrated by fig. \[fg:trjboostobj\] the Born-rigid object $\mathcal{G}$ rotates in the $xy$-plane. Clearly, $\mathcal{G}$’s spatial extent in the $x$- and $y$-directions has to be bounded by a maximum distance from the reference point $R$ on the order of $\Delta t / \theta_{TW}$ [@born09]. This boundary, which prevents paradoxical faster-than-light translations of sufficiently distant vertices, is produced by event horizons associated with the accelerations in the five boost sections. ![ \[fg:accgrid1dimevhrz\] Spacetime diagram of a one-dimensional grid consisting of seven points. The grid accelerates towards the positive $x$-direction. The trajectories are marked in blue, the mid point is taken as the reference $R$ and its worldline is colored in red. Dots indicate the lapse of 0.1 s in proper time. After 0.6 s have passed on $R$’s proper time close the acceleration stops and the points move with constant velocity (green lines). Dashed lines connect spacetime points of $R$’s comoving frame. The grey area indicates the inaccessible spacetime beyond the event horizon. ](memo44176-1.eps) Fig. \[fg:accgrid1dimevhrz\] exemplifies the formation of an event horizon for an accelerated object in $1+1$ (one time and one space) dimensions [see e.g. @desloge87; @eriksen82; @hamilton78; @semay06]. Here, the Born-rigid object is assumed to be one-dimensional and to consist of seven equidistant grid points. Each point accelerates for a finite time period towards the positive $x$-direction (blue worldlines); the reference point $R$, marked in red, accelerates with $\alpha_R \equiv 1~\text{ls}/\text{s}^2$. Contrary to the simulations discussed in fig. \[fg:trjboostobj\] above, in the present case the acceleration phase is not followed immediately by another boost. Rather, the object continues to move with constant velocity after the accelerating force has been switched off (green worldlines in fig. \[fg:accgrid1dimevhrz\]). The completion of the acceleration phase is synchronous in $R$’s instantaneous comoving frame (dashed-dotted line) and asynchronous in the laboratory frame. Fig. \[fg:accgrid1dimevhrz\] also illustrates that for an uniform acceleration the event horizon (black dot) is stationary with respect to the laboratory frame. In this simulation each vertex is assumed to be equipped with an ideal clock ticking at a proper frequency of 10 Hz, the corresponding ticks are marked by dots; the boost phase lasts for 0.6 s on $R$’s clock. The clocks of the left-most (trailing) and right-most (leading) vertex measure (proper time) boost durations of 0.3 s and 0.9 s, respectively. Thus, with respect to the instantaneous comoving frames (dashed lines) the vertex clocks run at different rates (see eqn. \[eq:tautfromtaur\]). The trailing clocks tick slower, the leading clocks faster than the reference clock at $R$. From eqns. \[eq:alphatfromalphar\] and \[eq:tautfromtaur\] it follows that the proper time variations are compensated by corresponding changes in proper acceleration experienced by the seven vertices. For the parameters used in fig. \[fg:accgrid1dimevhrz\] the accelerations of the trailing and leading vertex are $2\,\alpha_R$ and $2\,\alpha_R/3$, respectively. The spatial components of the inertial reference frames, comoving with $R$, are plotted in fig. \[fg:accgrid1dimevhrz\] as well. During the acceleration-free period following the boost phase the grid moves with constant velocity and the equal-time slices of the corresponding comoving frames (dotted lines) are oriented parallel to other. During the boost phase, however, the lines intersect and eqn. \[eq:postimattauB\] entails that the equal-time slices of the comoving frames all intersect in one spacetime point, the event horizon $x_{H} \equiv -1~\text{ls}$ (black dot at $x = -1$ ls and $t = 0$ s in fig. \[fg:accgrid1dimevhrz\]). A clock placed at $x_{H}$ does not tick, time is frozen at this spacetime point and eqns. \[eq:alphatfromalphar\] and \[eq:tautfromtaur\] imply that the corresponding proper acceleration diverges. Clearly a physical object accelerating towards positive $x$ in fig. \[fg:accgrid1dimevhrz\] cannot extend beyond this boundary at $x_{H}$. If the grid in fig. \[fg:accgrid1dimevhrz\] is regarded as realization of an accelerating coordinate system, this frame is bounded in the spatial dimension and ends at the coordinate value $x_H$. However, as soon as the grid’s acceleration stops, the event horizon disappears and coordinates $x < x_H$ are permissible. We note, that the event horizon in fig. \[fg:accgrid1dimevhrz\] is a zero-dimensional object, a point in $1+1$-dimensional spacetime considered here. The horizon is frozen in time and exists only for the instant $t=0$. Generalizing this result we find that the five boosts described in subsection \[su:fiveboosts\] and depicted in fig. \[fg:trjbstrefptfr1and6\] induce five event horizons in various orientations. It turns out that the accelerated object $\mathcal{G}$ is bounded by these horizons in all directions within the $xy$-plane. They limit $\mathcal{G}$’s maximum size [@born09] and thereby assure that all of its vertices obey the special relativistic speed limit [@einstein05]. Conclusions {#se:conclusion} =========== It is well known that pure Lorentz transformations do not form a group in the mathematical sense, since the composition of two transformations in general is not a pure Lorentz transformation again, but involves the Thomas-Wigner spatial rotation. The rotation is illustrated by uniformly accelerating a Born-rigid object, which is assumed to consist of a finite number of vertices, repeatedly, such that the object’s reference point returns to its start location. It turns out that at leasts five boosts are necessary, provided, first, the (proper time) duration and the magnitude of the proper acceleration within each boost is the same and, second, the object’s motion is restricted to the $xy$-plane. Analytic expressions are derived for the angles between adjacent boost directions. The visualizations illustrate the relationship between Thomas-Wigner rotations and the relativity of simultaneity. The transition from one boost section to the next occurs synchronously in the instantaneous comoving frame. In the laboratory frame, however, the trailing vertices conclude the current boost phase first and switch to the next one, which in general involves a direction change. In the laboratory frame the accelerated object not only contracts and expands along its direction of propagation, but also exhibits a shearing motion during the switchover phases. The simulations illustrate clearly that the aggregation of these shearing contributions finally adds up to the Thomas-Wigner rotation. Accelerated motions induce event horizons, which no part of a physical, Born-rigid object may overstep. Thus, the object’s size is limited to a finite volume or area (if its motion is restricted to two spatial dimension) and Thomas-Wigner rotations by construction observe the special relativistic speed limit. Auxiliary material ================== An MPEG-4 movie showing the Thomas-Wigner rotation of a grid object is available at the URL [www.gbeyerle.de/twr](www.gbeyerle.de/twr). In addition, the <span style="font-variant:small-caps;">Matlab</span> source code used to create fig. \[fg:trjboostobj\] and the “SymPy” script file discussed in \[se:appendix\] can be downloaded from the same site. Some calculations described in this paper were performed with the help of the computer algebra system “SymPy” [@joyner12], which is available for download from [www.sympy.org](www.sympy.org). “SymPy” is licensed under the General Public License; for more information see [www.gnu.org/licenses/gpl.html](www.gnu.org/licenses/gpl.html). Trademarks are the property of their respective owners. {#se:appendix} A number of equations in this paper were derived using the computer algebra system “SymPy” [@joyner12]. The corresponding “SymPy” source code files `vtwr3bst.py` (three boost case, see subsection \[su:threeboosts\]) and `vtwr5bst.py` (five boost case, see subsection \[su:fiveboosts\]) are available for download at [www.gbeyerle.de/twr](www.gbeyerle.de/twr). These scripts process eqns. \[eq:threeboostpos\], \[eq:threeboostvel\], \[eq:fiveboostpos\] and \[eq:fiveboostvel\] and derive the results given in eqns. \[eq:threebooststvalsol\], \[eq:threeboostspossol\], \[eq:t23fromt12\], \[eq:t12t23first\], \[eq:t12t23second\], \[eq:thwirot11\] and \[eq:thwirot21\]. The following offers a few explanatory comments. First, we address the case of three boosts (“SymPy” script `vtwr3bst.py`) and the derivation of eqn. \[eq:threebooststvalsol\]. The corresponding boost vectors in the $xy$-plane $\hat{e}^{(3B)}_i$ with $i=1,2,3$ are taken to be $$\begin{aligned} \label{eq:enrm2bst3} \hat{e}^{(3B)}_1 \equiv \begin{pmatrix} C_a \\ S_a \end{pmatrix} = \frac{1}{1+(T_{12})^2} \begin{pmatrix} 1-(T_{12})^2 \\ 2\,T_{12} \end{pmatrix} \\ \hat{e}^{(3B)}_2 \equiv \begin{pmatrix} 1 \\ 0 \end{pmatrix} \nonumber \\ \hat{e}^{(3B)}_3 \equiv \begin{pmatrix} C_a \\ -S_a \end{pmatrix} = \frac{1}{1+(T_{12})^2} \begin{pmatrix} 1-(T_{12})^2 \\ -2\,T_{12} \end{pmatrix} \nonumber\end{aligned}$$ with $$\begin{aligned} \label{eq:defs1c1} S_a &\equiv& \sin(\zeta_{1,2}) \qquad C_a \equiv \cos(\zeta_{1,2})\end{aligned}$$ in terms of the direction angle $\zeta_{1,2}$ and the half-angle approximation (eqn. \[eq:halfanglesincos\]). Here, the $z$-coordinate has been omitted since the trajectory is restricted to the $xy$-plane and $\zeta_{2,3} = \zeta_{1,2}$ from time reversal symmetry is being used. Inserting the corresponding Lorentz transformation matrices (eqn. \[eq:lotra1to2\]) into eqn. \[eq:threeboostvel\] and selecting the time component yields $$\begin{aligned} \label{eq:eval3bvel} (\gamma - 1)\, \frac{((T_{12})^2 - 2\,\gamma - 1)^2} {((T_{12})^2 + 1)^2} = 0\end{aligned}$$ which reduces to eqn. \[eq:threebooststvalsol\] if the trivial solution $\gamma=1$ is ignored. The case of five boosts (“SymPy” script `vtwr5bst.py`) and the derivation of eqn. \[eq:t12t23first\] is addressed next. In analogy to eqn. \[eq:enrm2bst3\] we define $$\begin{aligned} \label{eq:enrm2bst5} \hat{e}^{(5B)}_1 \equiv \begin{pmatrix} C_x \\ S_x \end{pmatrix} = \begin{pmatrix} C_{12}\,C_{23} - S_{12}\,S_{23} \\ -S_{12}\,C_{23} - C_{12}\,S_{23} \end{pmatrix} \\ \hat{e}^{(5B)}_2 \equiv \begin{pmatrix} C_y \\ S_y \end{pmatrix} = \begin{pmatrix} C_{23} \\ -S_{23} \end{pmatrix} = \frac{1}{1+(T_{23})^2} \begin{pmatrix} 1-(T_{23})^2 \\ -2\,T_{23} \end{pmatrix} \nonumber \\ \hat{e}^{(5B)}_3 \equiv \begin{pmatrix} 1 \\ 0 \end{pmatrix} \nonumber \\ \hat{e}^{(5B)}_4 \equiv \begin{pmatrix} C_y \\ -S_y \end{pmatrix} = \begin{pmatrix} C_{23} \\ S_{23} \end{pmatrix} = \frac{1}{1+(T_{23})^2} \begin{pmatrix} 1-(T_{23})^2 \\ 2\,T_{23} \end{pmatrix} \nonumber \\ \hat{e}^{(5B)}_5 \equiv \begin{pmatrix} C_x \\ -S_x \end{pmatrix} = \begin{pmatrix} C_{12}\,C_{23} - S_{12}\,S_{23} \\ S_{12}\,C_{23} + C_{12}\,S_{23} \end{pmatrix} \nonumber\end{aligned}$$ with eqn. \[eq:defs1c1\] and $$\begin{aligned} \label{eq:defs2c2} S_{12} &\equiv& \sin(\zeta_{1,2}) \qquad C_{12} \equiv \cos(\zeta_{1,2}) \\ S_{23} &\equiv& \sin(\zeta_{2,3}) \qquad C_{23} \equiv \cos(\zeta_{2,3}) \quad . \nonumber\end{aligned}$$ The corresponding Lorentz transformation matrices are too unwieldy to reproduce them here. “SymPy” script `vtwr5bst.py` calculates them, their products in terms of $T_{12}$ and $T_{23}$ and inserts them into eqn. \[eq:fiveboostvel\]. Selecting the time component yields the equation $$\begin{aligned} \label{eq:eqt12t23for5boost} \frac{\gamma - 1} {((T_{12})^2 + 1)^2\,((T_{23})^2 + 1)^2} \\ \times \left((T_{12})^2\,(T_{23})^2 + (T_{12})^2 - 4\,T_{12}\,T_{23}\,\gamma \right. \nonumber \\ \left. \qquad - 4\,T_{12}\,T_{23} - 2\,(T_{23})^2\,\gamma - (T_{23})^2 + 4\,\gamma^2 + 2\,\gamma - 1\right)^2 = 0 \quad . \nonumber\end{aligned}$$ We exclude the trivial solution $\gamma=1$ and restrict ourselves to real values of $T_{12}$ and $T_{23}$; eqn. \[eq:eqt12t23for5boost\] then leads to eqn. \[eq:t12dott23\], a second order polynomial with respect to $T_{23}$. The two solutions are given in eqn. \[eq:t23fromt12\]. Next insert $T_{23} = T_{23}(T_{12},\gamma)$ in eqn. \[eq:fiveboostpos\]. Since its time component involves the travel time of $R$ along its closed trajectory as an additional unknown and the $z$-coordinate vanishes by construction, we focus in the following on the $x$- and $y$-components of eqn. \[eq:fiveboostpos\]. The script `vtwr5bst.py` shows that the result for the $y$-component of the four-vector equation $\vec{P}_F^{[6]} = 0$ can be expressed in the form $$\begin{aligned} \label{eq:} -8\,((T_{12})^2 + 1)\,(\gamma + 1)^2 \, \frac{X_1(T_{12},\gamma) + X_2(T_{12},\gamma) \, \sqrt{X_3(T_{12},\gamma)}} {((T_{12})^2 - 2\,\gamma - 1)^6} = 0\end{aligned}$$ Here, $X_i(T_{12},\gamma)$ with $i=1,2,3$ are polynomials in $T_{12}$. For real $T_{12}$ and $\gamma \ge 1$ the numerator has to equate to zero. Moving the term involving the square root to the RHS and squaring both sides yields $$\begin{aligned} \left(X_1(T_{12},\gamma)\right)^2 - \left(X_2(T_{12},\gamma)\right)^2 \, X_3(T_{12},\gamma) = 0 \quad .\end{aligned}$$ Evaluation of this expression (see script `vtwr5bst.py`) leads to the product of two polynomial expression, each of which is fourth order with respect to $(T_{12})^2$ (see eqn. \[eq:t12t23first\] and \[eq:t12t23second\]). Repeating these calculations for the $x$-component of the four-vector equation $\vec{P}_F^{[6]} = 0$ leads also to the product of two polynomial expression, one of which is identical to eqn. \[eq:t12t23first\]. Thus, we have found a solution to eqn. \[eq:posinit5boost\]. The Thomas-Wigner angle $\theta_{TW}$ (eqn. \[eq:thowigrotangle\]) follows from the Lorentz matrix relating frame $[1]$ to frame $[6]$ (eqn. \[eq:thomaswignerrotmatrix\]). Script `vtwr5bst.py` evaluates the matrix elements $R_{1,1}$ and $R_{2,1}$ in terms of the variable $T_{12}$ and $T_{23}$. Again the expressions are unwieldy, but the derivation is straightforward. [42]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} J. S. Bell. *Speakable and Unspeakable in Quantum Mechanics*. Cambridge University Press, Cambridge, second edition, 2004. ISBN 978-0-521-52338-7. A. Ben-Menahem. Wigners rotation revisited. *Am. J. Phys.*, 530 (1):0 62–66, 1985. [doi: ]{}[10.1119/1.13953]{}. M. Born. Die [T]{}heorie des starren [E]{}lektrons in der [K]{}inematik des [R]{}elativitätsprinzips. *Annalen der Physik*, 3350 (11):0 1–56, 1909. [doi: ]{}[10.1002/andp.19093351102]{}. J. P. Costella, B. H. J. McKellar, A. A. Rawlinson, and G. J. [Stephenson Jr.]{} The [T]{}homas rotation. *Am. J. Phys.*, 690 (8):0 837–847, 2001. [doi: ]{}[10.1119/1.1371010]{}. J. T. Cushing. Vector [L]{}orentz transformations. *Am. J. Phys.*, 350 (9):0 858–862, 1967. [doi: ]{}[10.1119/1.1974267]{}. E. A. Desloge and R. J. Philpott. Uniformly accelerated reference frames in special relativity. *Am. J. Phys.*, 550 (3):0 252–261, 1987. [doi: ]{}[10.1119/1.15197]{}. E. Dewan and M. Beran. Note on stress effects due to relativistic contraction. *Am. J. Phys.*, 270 (7):0 517–518, 1959. [doi: ]{}[10.1119/1.1996214]{}. E. M. Dewan. Stress effects due to [L]{}orentz contraction. *Am. J. Phys.*, 310 (5):0 383–386, 1963. [doi: ]{}[10.1119/1.1969514]{}. A. Einstein. Zur [E]{}lektrodynamik bewegter [K]{}[ö]{}rper. *Annalen der Physik*, 3220 (10):0 891–921, 1905. [doi: ]{}[10.1002/andp.19053221004]{}. E. Eriksen, M. Mehlen, and J. M. Leinaas. Relativistic rigid motion in one dimension. *Phys. Scr.*, 250 (6B):0 905–910, 1982. [doi: ]{}[10.1119/1.1996214]{}. A. A. Evett. A relativistic rocket discussion problem. *Am. J. Phys.*, 400 (8):0 1170–1171, 1972. [doi: ]{}[10.1119/1.1986781]{}. A. A. Evett and R. K. Wangsness. Note on the separation of relativistically moving rockets. *Am. J. Phys.*, 280 (6):0 566–566, 1960. [doi: ]{}[10.1119/1.1935893]{}. F. Fernflores. Bell’s spaceships problem and the foundations of special relativity. *Int. Stud. Philos. Sci.*, 250 (4):0 351–370, 2011. [doi: ]{}[10.1080/02698595.2011.623364]{}. R. Ferraro and M. Thibeault. Generic composition of boosts: an elementary derivation of the [W]{}igner rotation. *Eur. J. Phys.*, 200 (3):0 143–151, 1999. [doi: ]{}[10.1088/0143-0807/20/3/003]{}. G. P. Fisher. The [T]{}homas precession. *Am. J. Phys.*, 400 (12):0 1772–1781, 1972. [doi: ]{}[10.1119/1.1987061]{}. J. Franklin. Lorentz contraction, [B]{}ell’s spaceships and rigid body motion in special relativity. *Eur. J. Phys.*, 310 (2):0 291–298, 2010. [doi: ]{}[10.1088/0143-0807/31/2/006]{}. H. Gelman. Sequences of co-moving [L]{}orentz frames. *J. Math. Anal. Appl.*, 1450 (2):0 524–538, 1990. [doi: ]{}[10.1016/0022-247X(90)90418-F]{}. E. Gourgoulhon. *Special Relativity in General Frames*. Springer, Berlin, Heidelberg, 2013. J. D. Hamilton. The uniformly accelerated reference frame. *Am. J. Phys.*, 460 (1):0 83–89, 1978. [doi: ]{}[10.1119/1.11169]{}. G. Herglotz. ber den vom [S]{}tandpunkt des [R]{}elativitätsprinzips aus als starr zu bezeichnenden [K]{}örper. *Annalen der Physik*, 3360 (2):0 393–415, 1909. [doi: ]{}[10.1002/andp.19103360208]{}. M. P. Hobson, G. P. Efstathiou, and A. N. Lasenby. *General relativity: [A]{}n introduction for physicists*. Cambridge University Press, Cambridge, 2006. ISBN 0-521-82951-8. R. M. Jonsson. Gyroscope precession in special and general relativity from basic principles. *Am. J. Phys.*, 750 (5):0 463–471, 2007. [doi: ]{}[10.1119/1.2719202]{}. D. Joyner, O. Čertík, A. Meurer, and B. E. Granger. Open source computer algebra systems: [S]{}ym[P]{}y. *ACM Communications in Computer Algebra*, 450 (3/4):0 225–234, 2012. [doi: ]{}[10.1145/2110170.2110185]{}. W. L. Kennedy. Thomas rotation: [A]{} [L]{}orentz matrix approach. *Eur. J. Phys.*, 230 (3):0 235–247, 2002. [doi: ]{}[10.1088/0143-0807/23/3/301]{}. C. W. Misner, K. S. Thorne, and J. A. Wheeler. *Gravitation*. Palgrave Macmillan, 1973. ISBN 978-0716703440. C. I. Mocanu. On the relativistic velocity composition paradox and the [T]{}homas rotation. *Found. Phys. Lett.*, 50 (5):0 443–456, 1992. [doi: ]{}[10.1007/BF00690425]{}. F. Noether. Zur [K]{}inematik des starren [K]{}örpers in der [R]{}elativtheorie. *Annalen der Physik*, 3360 (5):0 919–944, 1910. [doi: ]{}[10.1002/andp.19103360504]{}. K. Rȩbilas. Subtleties of the [T]{}homas precession. *Eur. J. Phys.*, 360 (4):0 1–13, 2015. [doi: ]{}[10.1088/0143-0807/36/4/045007]{}. D. V. Redzić. Note on [D]{}ewan–[B]{}eran–[B]{}ell’s spaceship problem. *Eur. J. Phys.*, 290 (3):0 N11–N19, 2008. [doi: ]{}[10.1088/0143-0807/29/3/N02]{}. J. A. Rhodes and M. D. Semon. Relativistic velocity space, [W]{}igner rotation, and [T]{}homas precession. *Am. J. Phys.*, 720 (7):0 943–960, 2004. [doi: ]{}[10.1119/1.1652040]{}. W. Rindler. *Relativity: [S]{}pecial, [G]{}eneral, and [C]{}osmological*. Oxford University Press, USA, 2006. E. G. P. Rowe. The [T]{}homas precession. *Eur. J. Phys.*, 50 (1):0 40–45, 1984. [doi: ]{}[10.1088/0143-0807/5/1/009]{}. C. Semay. Observer with a constant proper acceleration. *Eur. J. Phys.*, 270 (5):0 1157–1167, 2006. [doi: ]{}[10.1088/0143-0807/27/5/015]{}. A. M. Steane. *Relativity Made Relatively Easy*. Oxford University Press, 2012. ISBN 019966286X. D. F. Styer. How do two moving clocks fall out of sync? [A]{} tale of trucks, threads, and twins. *Am. J. Phys.*, 750 (9):0 805–814, 2007. [doi: ]{}[10.1119/1.2733691]{}. A. Tartaglia and M. L. Ruggiero. Lorentz contraction and accelerated systems. *Eur. J. Phys.*, 240 (2):0 215–220, 2003. [doi: ]{}[10.1088/0143-0807/24/2/361]{}. L. H. Thomas. The motion of the spinning electron. *Nature*, 1170 (2945):0 514–514, 1926. [doi: ]{}[10.1038/117514a0]{}. L. H. Thomas. The kinematics of an electron with an axis. *Philos. Mag.*, 30 (13):0 1–22, 1927. [doi: ]{}[10.1080/14786440108564170]{}. S.-I. Tomonaga. *The story of spin*. The University of Chicago Press, Chicago, 1997. ISBN 0-226-80794-0. A. A. Ungar. The relativistic velocity composition paradox and the [T]{}homas rotation. *Found. Phys.*, 190 (11):0 1385–1396, 1989. [doi: ]{}[10.1007/BF00732759]{}. A. A. Ungar. Thomas precession: [I]{}ts underlying gyrogroup axioms and their use in hyperbolic geometry and relativistic physics. *Found. Phys.*, 270 (6):0 881–951, 1997. [doi: ]{}[10.1007/BF02550347]{}. E. P. Wigner. On unitary representations of the inhomogeneous [L]{}orentz group. *Ann. Math.*, 400 (1):0 149–204, 1939. [doi: ]{}[10.2307/1968551]{}.

ArXiv

--- abstract: 'The need for transparency of predictive systems based on Machine Learning algorithms arises as a consequence of their everincreasing proliferation in the industry. Whenever blackbox algorithmic predictions influence human affairs, the inner workings of these algorithms should be scrutinised and their decisions explained to the relevant stakeholders, including the system engineers, the system’s operators and the individuals whose case is being decided. While a variety of interpretability and explainability methods is available, none of them is a panacea that can satisfy all diverse expectations and competing objectives that might be required by the parties involved. We address this challenge in this paper by discussing the promises of *Interactive* Machine Learning for improved transparency of blackbox systems using the example of contrastive explanations – a stateoftheart approach to *Interpretable* Machine Learning. Specifically, we show how to personalise counterfactual explanations by interactively adjusting their conditional statements and extract additional explanations by asking followup “What if?” questions. Our experience in building, deploying and presenting this type of system allowed us to list desired properties as well as potential limitations, which can be used to guide the development of interactive explainers. While customising the medium of interaction, i.e., the user interface comprising of various communication channels, may give an impression of personalisation, we argue that adjusting the explanation itself and its content is more important. To this end, properties such as breadth, scope, context, purpose and target of the explanation have to be considered, in addition to explicitly informing the explainee about its limitations and caveats. Furthermore, we discuss the challenges of mirroring the explainee’s mental model, which is the main building block of intelligible humanmachine interactions. We also deliberate on the risks of allowing the explainee to freely manipulate the explanations and thereby extracting information about the underlying predictive model, which might be leveraged by malicious actors to steal or game the model. Finally, building an endtoend interactive explainability system is a challenging engineering task; unless the main goal is its deployment, we recommend “Wizard of Oz” studies as a proxy for testing and evaluating standalone interactive explainability algorithms.' author: - Kacper Sokol - Peter Flach bibliography: - 'template.bib' date: '(Published in the *Künstliche Intelligenz* journal, special issue on *Challenges in Interactive Machine Learning*.)' subtitle: The Promise of Interactive Explanations for Machine Learning Transparency title: One Explanation Does Not Fit All --- Introduction\[sec:chap07:intro\] ================================ Given the opaque, “blackbox” nature of complex Machine Learning (ML) systems, their deployment in missioncritical domains is limited by the extent to which they can be interpreted or validated. In particular, predictions, (trained) models and (training) data should be accounted for. One way to achieve this is by “transparency by design”, so that all components of a predictive system are “glass boxes”, i.e., antehoc transparency [@rudin2019stop]. Alternatively, transparency might be achieved with posthoc tools, which have the advantage of not limiting the choice of a predictive model in advance [@ribeiro2016why]. The latter approaches can either be modelspecific or modelagnostic [@robnik2018perturbation]. Despite this wide range of available tools and techniques, many of them are noninteractive, providing the explainee (a recipient of an explanation) with a single explanation that has been optimised according to some predefined metric. While some of these methods simply cannot be customised by the end user without an indepth understanding of their inner workings, others can take direct input from users with a varying level of domain expertise: from a lay audience – e.g., selecting regions of an image in order to query their influence on the classification outcome – to domain experts – e.g., tuning explanation parameters such as the importance of neighbouring data points. A particular risk of a lack of interaction and personalisation mechanisms is that the explanation may not always align with users’ expectations, reducing its overall value and usefulness. Allowing the user to guide and customise an explanation can benefit the transparency of a predictive system by making it more suitable and appealing to the explainee, for example, by adjusting its content and complexity. Therefore, personalisation can be understood as modifying an explanation or an explanatory process to answer userspecific questions. For counterfactual explanations of the form: “had feature $X$ been different, the prediction of the model would have been different too”, these can be userdefined constrains on the number and type of features ($X$) that can and cannot appear in the conditional statement. Delegating the task of customising and personalising explanations to the end user via interaction mitigates the need for the difficult process of modelling the user’s mental model beforehand, rendering the task feasible and making the whole process feel more natural, engaging and less frustrating. In human interactions, understanding is naturally achieved via an *explanatory dialogue* [@miller2018explanation], possibly supported with visual aids. Mirroring this explanatory process for ML transparency would make it attractive and accessible to a wide audience. Furthermore, allowing the user to customise explanations extends their utility beyond ML transparency. The explainee can steer the explanatory process to inspect fairness (e.g., identify biases towards protected groups[^1]) [@kusner2017counterfactual], assess accountability (e.g., identify model errors such as nonmonotonic predictions with respect to monotonic features) [@lipton2017doctor] or debug predictive models [@kulesza2015principles; @sokol2019counterfactual]. In contrast to ML tasks [@kapoor2010interactive] – where any interaction may be impeded by humanincomprehensible internal representations utilised by a predictive model – interacting with explainability systems is feasible as the representation has to be humanunderstandable in the first place, thereby enabling a bidirectional *communication*. Interaction with explanatory systems also allows incorporating new knowledge into the underlying ML algorithm and building a mental model of the explainee, which will help to customise the resulting explanations. Consider the example of explaining an image with a local surrogate method that relies upon superpixel segmentation (e.g., LIME algorithm introduced by @ribeiro2016why). While superpixel discovery may be good at separating colour patches based on their edges, these segments do not often correspond to meaningful *concepts* such as ears or a tail for a dog image – see Figure \[fig:chap07:segmentation\] for an example. The explanation can be personalised by allowing the explainee to *merge* and *split* segments before analysing their influence on the output of a blackbox model, thereby implicitly answering what prompted the explainee to alter the segmentation. User input is a welcome addition given the complexity of images; a similar approach is possible for tabular and text data, although user input is often unnecessary in these two cases. For tabular data the explainee may select certain feature values that are of interest or create meaningful binning for some of the continuous features; for text data (treated as a bag of words) the user may group some words into a phrase that conveys the correct meaning in that particular sentence. This exchange of knowledge between the explainee and the explainability system can considerably increase the quality of explanations, but also poses a significant safety, security and privacy risk. A malicious explainee may use such a system to reveal sensitive data used to train the underlying predictive (or explanatory) model, extract proprietary model components, or learn its behaviour in an attempt to game it (see Section \[sec:chap07:glass-box\_desiderata\]). [0.45]{} {width="\textwidth"} [0.45]{} {width="\textwidth"} After @miller2018explanation’s [@miller2018explanation] seminal work – inspired by explanation research in the social sciences – drew attention to the lack of humanaspect considerations in the eXplainable Artificial Intelligence (XAI) literature – with many such systems being designed by the technical community for the technical community [@miller2017explainable] – researchers started acknowledging the end user when designing XAI solutions. While this has advanced humancentred design and validation of explanations produced by XAI systems, another of @miller2018explanation’s insights received relatively little attention: the interactive, dialoguelike nature of explanations. Many of the stateoftheart explainability approaches are static, oneoff systems that do not take user input or preferences into consideration beyond the initial configuration and parametrisation [@friedman2001greedy; @goldstein2015peeking; @lundberg2017unified; @ribeiro2016why; @ribeiro2018anchors; @wachter2017counterfactual].[^2] While sometimes the underlying explanatory algorithms are simply incapable of a meaningful interaction, others do apply a technique or utilise an explanatory artefact that can support it in principle. Part of this trend can be attributed to the lack of a welldefined protocol for evaluating interactive explanations and the challenging process of assessing their quality and effectiveness, which – in contrast to a oneshot evaluation – is a software system engineering challenge[^3] and requires time and resourceconsuming user studies. @schneider2019personalized noted that bespoke explanations in AI – achieved through interaction or otherwise – are largely absent within the existing literature. Research in this space usually touches upon three aspects of “personalised” explanations. First, there are interactive machine learning systems where the user input is harnessed to improve performance of a predictive model or align the data processing with its operator’s prior beliefs. While the classic active learning paradigm dominates this space, @kulesza2015principles designed a system that presents its users with classification explanations to help them refine and personalise the predictive task, hence focusing the interaction on the underlying ML model and not the explanations. Similarly, @kim2015ibcm introduced an interactive ML system with an explainability component, allowing its users to alter the data clustering based on their preferences. Secondly, the work of @krause2016interacting and @weld2018challenge focused on interactive (multimodal) explainability systems. Here, the interaction allows the explainee to elicit more information about an ML system by receiving a range of diverse explanations derived from a collection of XAI algorithms such as Partial Dependence (PD) [@friedman2001greedy] and Individual Conditional Expectation (ICE) [@goldstein2015peeking] plots. While this body of research illustrates what such an interaction (with multiple explanatory modalities) might look like and persuasively argues its benefits [@weld2018challenge], the advocated interaction is mostly with respect to the presentation medium itself – e.g., an interactive PD plot – and cannot be used to customise and personalise the explanation *per se*. Thirdly, @madumal2019grounded and @schneider2019personalized developed theoretical frameworks for interactive, personalised explainability that prescribe the interaction protocol and design of such systems. However, these theoretical foundations have not yet been utilised to design and implement an interactive explainability system coherent with XAI desiderata outlined by @miller2018explanation, which could offer customisable explanations. A more detailed overview and discussion of the relevant literature is given in Section \[sec:chap07:related\_work\]. In this paper we propose an architecture of a truly interactive explainability system, demonstrate how to build such a system, analyse its desiderata, and examine how a diverse range of explanations can be personalised (Section \[sec:chap07:glass-box\]). Furthermore, we discuss lessons learnt from presenting it to both a technical and a lay audience, and provide a plan for future research in this direction (Section \[sec:chap07:discussion\]). As a first attempt to build an XAI system that allows the explainee to customise and personalise the explanations, we decided to use a *decision tree* as the underlying predictive model. This choice simplifies many steps of our initial study, allowing us to validate (and guarantee correctness of) the explanations and reduce the overall complexity of the explanation generation and tuning process by inspecting the structure of the underlying decision tree. Using *antehoc* explanations derived from a single predictive model also allows us to mitigate engineering challenges that come with combining multiple independent XAI algorithms as proposed by @weld2018challenge. Furthermore, a decision tree can provide a wide range of diverse explanation types, many of which can be customised and personalised. Specifically, for global model explanations we provide - *model visualisation*, and - *feature importance*; while as prediction explanations we provide - *decision rule* – extracted from a roottoleaf path, - *counterfactual* – achieved by comparing decision rules for different tree leaves, and - *exemplar* – a similar training data point extracted from the tree leaves. When presented to the user, all of these explanations span a wide range of explanatory artefacts in visualisation (images) and textualisation (natural language) domains, thereby allowing us to test the extent to which they can be interactively personalised. Contrastive explanations, in particular *classcontrastive* counterfactual statements, are the foundation of our system. These take the form of: “had *one of the attributes been different in a particular way*, the classification outcome would have changed as follows….” Arguably, they are the most suitable, natural and appealing explanations targeted at humans [@miller2018explanation; @wachter2017counterfactual]. In addition to all of their desired properties grounded in the social sciences [@miller2018explanation] and legal considerations [@wachter2017counterfactual], they can be easily adapted to an interactive dialogue aimed at personalisation, which is not widely utilised. In our system they are delivered in an interactive dialogue – a natural language conversation, which is the most intuitive explanatory mechanism [@miller2018explanation]. In summary, our approach aims to build a holistic and diverse interactive XAI system where the interaction is focused on **personalising** explanations (in accordance with @miller2018explanation’s [@miller2018explanation] notion of XAI interactivity) as opposed to simply building an XAI system that provides explanations interactively (to explain different aspects of a blackbox system using a range of XAI algorithms) – a subtle but significant difference. Background and Related Work\[sec:chap07:related\_work\] ======================================================= Throughout our research we have identified three distinct research strands in the literature that are relevant to interactive explanations: - interactive Artificial Intelligence and Machine Learning (mostly from the perspective of HumanComputer Interaction), - interactive explainability tools, which are interactive with respect to the user interface that delivers the explanations, and - theory of explanatory interactions, e.g., a natural language dialogue, between two intelligent agents (be them humans, machines or one of each). The HumanComputer Interaction community has identified the benefits of human input for tools powered by AI and ML algorithms that extend beyond the active learning paradigm where people act as data labelling oracles [@amershi2014power]. For example, consider a movie recommendation system where the user provides both explicit feedback, such as movie ratings, and implicit feedback, e.g., movies that the user did not finish watching. In order to utilise the full potential of any feedback and ensure user satisfaction, the users have to understand how their input affects the system (in particular, its underlying predictive model). Among others, the users should be informed whether their feedback is incorporated into recommendations immediately or with a delay and how does “liking” a movie influences future recommendations (e.g., similar genre and shared cast members). Here, this understanding is mostly achieved (in the case of user studies) by inviting the users to onboarding sessions or (progressively) disclosing relevant information via the user interface, hence the explanation is provided outside of the autonomous system. These actions help the users build a correct mental model of the “intelligent agent” allowing them to seamlessly interact with it. Ideally, the users would develop a *structural* mental model that gives them a deep and indetailed understanding of how the ML or AI operates, however a *functional* mental model (a shallow understanding) often suffices. While explanations are often provided outside of the interactive agents, several researchers showed how to integrate them into the user interface of autonomous systems [@kulesza2010explanatory; @kulesza2013too; @kulesza2015principles; @kim2015ibcm]. This is especially useful when the system is dynamic – e.g., its underlying predictive model evolves over time – in which case the explanations support and inform users’ interaction with the system and guide the users towards achieving the desired objective. There are two prominent examples of such systems in the literature. @kulesza2015principles developed an interactive topicbased Naïve Bayes classifier for electronic mail to help the users “debug” and “personalise” the categorisation of emails. The users are presented with explanations pertaining to every classified email – words in the email that contribute towards and against a given class – and are allowed to adjust the weights of these factors if they do not agree with their premise, and hence refine and personalise the model in a process which the authors call *explanatory debugging* [@kulesza2010explanatory; @kulesza2013too; @kulesza2015principles]. @kim2015ibcm designed a similar system where the users can interactively personalise clustering results – which are explained with cluster centroids and prominent exemplars – by promoting and demoting data points within each cluster. In this literature, explanations of predictive models are used to improve users’ understanding (mental model) of an autonomous system to empower them to better utilise its capabilities (e.g., via improved personalisation) by interactively providing beneficial input. Hence, AI and ML explainability is not the main research objective in this setting and the explanations are not interactive themselves. The second research strand that we identified in the literature covers interactive, multimodal explainability tools in AI and ML. These systems allow investigating a blackbox model and its predictions by providing the user with a variety of explanations produced with a range of diverse explainability techniques delivered via (an interactive) user interface. For example, @krause2016interacting built an interactive system that allows its users to inspect Partial Dependence [@friedman2001greedy] of selected features (model explanation) and investigate how changing feature values for an individual data point would affect its classification (prediction explanation) [@krause2016interacting; @krause2016using]. Whereas combining multiple explainability techniques within a single system with a unified user interface is feasible, ensuring coherence of the diverse explanations that they produce poses significant challenges as some of the explanations may be at odds with each other and provide contradictory evidence for the same outcome. @weld2018challenge showed an idealised example of such a system and persuasively argued its benefits, however they have not discussed how to mitigate the issue with contradictory and competing explanations. While both of these explainability tools are *interactive*, the interaction is limited to the presentation medium of the explanations and a choice of explainability technique, which, we argue, is insufficient – the system is interactive but the explanation is not. *Truly interactive* explanations allow the user to tweak, tune and personalise them (i.e., their content) via an interaction, hence the explainee is given an opportunity to guide them in a direction that helps to answer selected questions. The third research strand in the literature characterises explanatory communication as interaction between two intelligent agents [@arioua2015formalizing; @walton2016dialogue; @madumal2019grounded; @schneider2019personalized]. @arioua2015formalizing formalised explanatory dialogues in Dung’s argumentation framework [@dung2009assumption] and introduced “questioning” dialogues to evaluate success of explanations. @walton2016dialogue introduced a similar *shift model* composed of two distinct dialogue modes: an explanation dialogue and an examination dialogue, where the latter is used to evaluate the success of the former [@walton2007dialogical; @walton2011dialogue; @walton2016dialogue]. @madumal2019grounded refined these two approaches and proposed an interactive communication schema that supports explanatory and questioning dialogues, which also allow the explainee to formally challenge and argue against some of the decisions and their explanations. @madumal2019grounded have also empirically evaluated their explanatory dialogue protocol on text corpora to show its flexibility and applicability to a range of different scenarios. @schneider2019personalized approached this problem on a more conceptual level discussing interactions with various explainability tools and showing examples of how they could allow for personalised explanation. Most of the work presented in this body of literature is purely theoretical and has not yet been embraced by practical explainability tools. These diverse research strands come together to help eXplainable AI and Interpretable Machine Learning (IML) researchers and practitioners design appealing and useful explainability tools with many of their recommendations originating from explanatory interactions between humans. @miller2018explanation reviewed a diverse body of social sciences literature on human explanations and proposed an agenda for humancentred explanations in AI and ML. @miller2017explainable noticed that explainability systems built for autonomous agents and predictive systems rarely ever consider the end users and their expectations as they are mostly “built by engineers, for engineers.” Since then, XAI and IML research has taken a more humancentred direction, with many academics and engineers [@wachter2017counterfactual; @waa2018contrastive; @schneider2019personalized; @weld2018challenge; @henin2019towards] evaluating their approaches against @miller2018explanation’s guidelines to help mitigate such issues. Two of @miller2018explanation’s recommendations are of particular importance: interactive, dialoguelike nature of explanations and popularity of contrastive explanations among humans. While interactivity of explanations [@schneider2019personalized] has been investigated from various viewpoints in the literature (and discussed earlier in this section), explanations delivered in a bidirectional conversation, giving the explainee the opportunity to customise and personalise them, have not seen much uptake in practice. Oneoff explanations are still the most popular operationalisation of explainability algorithms [@schneider2019personalized], where the explainer outputs a onesizefitsall explanation in an attempt to make the behaviour of a predictive system transparent. A slight improvement over this scenario is to enable the explainer to account for user preferences when generating the explanations [@lakkaraju2019faithful; @poyiadzi2019face], but this modality is not common either. Interactively personalising an explanation allows the users to adjust its complexity to suit their background knowledge, experience and mental capacity; for example, explaining a disease to a medical student should take a very different form from explaining it to a patient. Therefore, an interactive system can satisfy a wide range of explainees’ expectations, including objectives other than improving transparency such as inspecting individual fairness of algorithmic predictions [@kusner2017counterfactual]. The prominence of contrastive statements in human explanations is another important insight from the social sciences, which also highlights their capacity to be interactively customised and personalised. In the recent years this type of explanations has proliferated into the XAI and IML literature in the form of *class contrastive counterfactual* explanations: “had you earned twice as much, your loan application would have been successful.” This uptake can also be attributed to their legal compliance [@wachter2017counterfactual] with the “right to explanation” introduced by the European Union’s General Data Protection Regulation (GDPR). However, their capacity to be customised and personalised is often overlooked in practice [@miller2018explanation; @wachter2017counterfactual; @waa2018contrastive; @poyiadzi2019face]. All in all, many of @miller2018explanation’s [@miller2018explanation] insights from the social sciences have found their way into research and practice. An example of the latter is Google’s *People + AI Guidebook*[^4] describing best practices for designing humancentred AI and ML products and acknowledging the importance of interaction and explainability in such systems. The lack of customisable explanations has also received attention in the literature [@schneider2019personalized]. @schneider2019personalized have reviewed an array of explainability approaches focusing on their personalisation capabilities. They have observed that personalised explanations in XAI and IML are generally absent from the existing literature. To help researchers design and implement such methods, @schneider2019personalized proposed a generic framework for personalised explanations that identifies their three adjustable properties: complexity, content (called “decision information”, i.e., what to explain) and presentation (how to explain, e.g., figures vs. text); @eiband2018bringing discussed the latter two properties from a user interface design perspective. Furthermore, @schneider2019personalized highlighted that interactive personalisation of explanations can either be an *iterative*, e.g., a conversation, or a *oneoff* process, e.g., specifying constrains before the explanation is generated. The latter approach does not, however, require the explainability system to be interactive as the same personalisation can be achieved offline by extracting the personalisation specification from the explainee and subsequently incorporating it into the data or algorithm initialisation. Interaction with explainability systems has also been acknowledged by @henin2019towards, who proposed a generic mathematical formulation of blackbox explainers consisting of three distinct steps: sampling, generation and **interaction**. While some explainability approaches introduced in the literature are simply incapable of interactive personalisation – a number of them may still be personalised offline – others are, however this property is neither utilised [@wachter2017counterfactual] nor acknowledged. This lack of recognition may be because the explainability system designers do not see the benefits of this step or due to the difficulties with building such systems (from the engineering perspective) as well as evaluating them. To facilitate interactive personalisation the user interface has to be capable of delivering explanations and collecting explainees’ feedback, which may require an interdisciplinary collaboration with User Experience and HumanComputer Interaction researchers. Systematic evaluation and validation of this type of explainers is also more elaborate, possibly requiring multiple rounds of timeconsuming user studies. Despite these hurdles, a number of explainability tools and techniques enable the user to personalise explanations to some extent. @akula2019natural presented a dialoguedriven explainability system that uses contrastive explanations based on predictions derived from AndOr graphs and handcrafted ontology, however generalising this technique may be challenging as it requires handcrafting separate ontology and AndOr graph for each application. @lakkaraju2019faithful introduced rulebased explanations that the user can personalise by choosing which features will appear in the explanation – an offline personalisation. Google published their *whatif* tool[^5], which provides the explainee with an interactive interface that allows generating contrastive explanations of selected data points by modifying their features, i.e., asking “What if?” questions. In our work we strive to bring together the most important concepts from this wide spectrum of research as a generic and powerful aid to people building explainers of predictive systems that allow explanation personalisation via online interaction. To this end, we provide an overview of a voicedriven contrastive explainer built for an ML loan application model, which allows the explainee to interrogate its predictions by asking counterfactual questions [@sokol2018glass]. We discuss our experience from building, deploying and presenting the system, which allowed us to critically evaluate its properties and formulate further desiderata and lessons learnt. We present these observations in the following sections as guidelines for developing similar projects. Interactively Customisable Explanations\[sec:chap07:glass-box\] =============================================================== As a first step towards personalised, interactive XAI systems we developed *GlassBox* [@sokol2018glass]: a classcontrastive counterfactual explainability system that can be queried with a natural language dialogue (described in Section \[sec:chap07:glass-box\_design\]). It supports a range of “Why?” questions that can be posed either through a voice or chatbased interface. Building this system and testing it in the wild provided us with invaluable experience and insights, which we now share with the community as they may be useful to anybody attempting to develop and deploy a similar system – Sections \[sec:chap07:glass-box\_desiderata\] and \[sec:chap07:glass-box\_properties\] discuss interactive explainers *desiderata* and *properties* respectively. The feedback that helped us to refine our idea of interactive XAI systems producing personalised explanations (presented in Section \[sec:chap07:glass-box\_feedback\]) was collected while demonstrating GlassBox to a diverse audience consisting of both domain experts, approached during the 27^th^ International Joint Conference on Artificial Intelligence (IJCAI 2018), and a lay audience, approached during a local “Research without Borders” festival[^6] that is open to the public and attended by pupils from local schools. While at the time of presentation our system was limited to classcontrastive counterfactual explanations personalised by (implicitly) choosing data features that the counterfactual statements were conditioned on and provided to the user in natural language, we believe that our observations remain valid beyond this particular XAI technique. We hope to test this assumption in our future work – see Section \[sec:chap07:discussion\] for more details – by employing the remaining four decision tree explainability modalities listed in the introduction, albeit in an XAI system refined based on our experience to date. GlassBox Design\[sec:chap07:glass-box\_design\] ----------------------------------------------- GlassBox has been designed as a piece of hardware built upon the Google AIY (Artificial Intelligence Yourself) Voice Kit[^7] – a customisable hardware and software platform for development of voice interfaceenabled interactive agents. The first prototype of GlassBox utilised the Amazon Alexa skill Application Programming Interface, however the limitations of this platform at the time (the processing of data had to be deployed to an online server and invoked via an API call) have hindered the progress and prompted us to switch to the aforementioned Google AIY Voice device. These recent technological advancements in automated speechtotext transcription and speech synthesis provided as a service allowed us to utilise an offtheshelf, voiceenabled, virtual, digital assistant to process explainees’ speech and automatically answer their questions – something that would not have been feasible had we decided to build this component ourselves. We extended the voicedriven user interface with a (textual) chatbased web interface that displays the transcription of the conversation and its history – to improve accessibility of the system, among other things – in addition to allowing the explainee to type in the queries instead of saying them out loud. To avoid a lengthy and possibly offputting process of submitting (mock) personal details – i.e., a data point – to be predicted by the underlying Machine Learning model and explained by GlassBox, we opted for a predefined set of ten data points. Any of them could be selected and input to GlassBox by scanning a QR code placed on a printed card that also listed details of this fictional individual. Once a data point is selected, the explainee can alter personal details of this fictional individual by interacting with GlassBox, e.g., “I am 27 years old, not 45.” Any input to the system is passed to a natural language processing and understanding module built using *rasa*[^8]. Our deployment of the GlassBox system was based on the *UCI German credit* data set[^9] (using a subset of its features) for which a decision tree classifier was trained using *scikitlearn*[^10] [@scikit-learn]. Since the German credit data set has a binary target variable (“good” or “bad” credit score), the class contrast in the counterfactual explanations is implicit. Nevertheless, this could be easily generalised to a multiclass setting by requiring the explainee to explicitly specify the contrast class, taking the secondmost likely one or providing one explanation per class. A conceptual design of GlassBox is presented in Figure \[fig:chap07:glass-box\_design\]. ![GlassBox design and information flow.[]{data-label="fig:chap07:glass-box_design"}](fig07/glass-box.pdf){width=".45\textwidth"} To facilitate some of the user interactions the data set had to be manually annotated. This process allowed the generation of engaging natural language responses and enabled answering questions related to individual fairness. The latter functionality was achieved by indicating which features (and combinations thereof) should be treated as protected attributes (features), hence had a counterfactual data point conditioned on one of these features been found, GlassBox would indicate unfair treatment of this individual. This functionality could be invoked by asking “Is the decision fair?” question and further interrogating the resulting counterfactual explanation if one was found. Depending on the explainability and interactiveness requirements expected of the system, other data set annotations may be required. Since annotation is mostly a manual process, creating them can be time and resourceconsuming. As noted before, the main objective of GlassBox is to provide the users with personalised explanations whenever they decide to challenge the decision of the underlying Machine Learning model. The explainee can request and interactively customise the resulting counterfactual explanations through a natural language interface with appropriate dialogue cues. This can be done in three different ways by asking the following questions: - “Why?” – a plain counterfactual explanation – the system returns the shortest possible classcontrastive counterfactual; - “Why despite?” – a counterfactual explanation not conditioned on the indicated feature(s) – the system returns a classcontrastive counterfactual that does not use a specified (set of) feature(s) as its condition; and - “Why given?” – a (partially)fixed counterfactual explanation – the system returns a counterfactual that is conditioned on the specified (set of) feature(s). By repetitively asking any of the above “Why?” questions the system will enumerate all the possible explanations with the condition set (the features that need to change) increasing in quantity until no more explanations can be found. It is also possible to mix the latter two questions into “Why given …and despite …?”, thereby introducing even stronger restrictions on the counterfactual explanations. In addition to “Why?” questions the explainee can also ask **“What if?”** In this case it is the user who provides the contrast and wants to learn the classification outcome of this hypothetical data point. This question can be either applied to the selected data point (which is currently being explained) or any of the counterfactual data points offered by the system as an explanation. All of these requirements imposed by the user are processed by a simple logical unit that translates the user requests into constraints applied to the set of features that the counterfactual is allowed and/or required to be conditioned upon. All of these happen through a natural language dialogue, an example of which is depicted in Figure \[fig:chap07:glass-box\_conversation\]. ![An example explanatory conversation between GlassBox and an explainee who personalises the explanations by asking counterfactual questions.[]{data-label="fig:chap07:glass-box_conversation"}](fig07/conversation.pdf){width=".45\textwidth"} The method used to generate counterfactual explanations from the underlying decision tree classifier relies upon a bespoke leaftoleaf distance metric. It allows to find leaves of different classes to the one assigned to the selected data point that require the fewest possible changes to this data point in its feature space. One obvious solution to this problem is any neighbouring leaf of a different class; this requires just one feature to be altered. However, there may also exist leaves that are relatively distant in the decision tree structure but also require just one feature value change, for example, when these two decision tree paths do not share many features. This distance metric is computed by representing the tree structure in a binary *metafeature* space that is created by extracting all the unique feature partitions from the splits of the decision tree. Finally, an $L1$like metric (when a particular feature is present on one branch and absent on the other, this distance component is assumed to be $0$) is calculated and minimised to derive a list of counterfactual explanations ordered by their length. Explanation Desiderata\[sec:chap07:glass-box\_desiderata\] ---------------------------------------------------------- During the development stage and early trials of GlassBox we identified a collection of *desiderata* and *properties* that should be considered when building such systems. Some of these attributes are inspired by relevant literature [@weld2018challenge; @miller2018explanation; @kulesza2015principles; @schneider2019personalized], while others come from our experience gained in the process of building the system, presenting it to various audiences, discussing its properties at different events and collecting feedback about interacting with it. While this and the following sections focus on desiderata for interactive and customisable explanations, we provide an indepth discussion on this topic for generic explainability systems in our work on “Explainability Fact Sheets” [@sokol2020explainability]. The relevant subset of these desiderata are summarised in Table \[tab:chap07:glass-box\_desiderata\] as well as collected and discussed below. Section \[sec:chap07:glass-box\_properties\], on the other hand, examines the properties of interactive explainability systems. **Functional** **Operational** **Usability** --------------------------------------- --------------------------------- ---------------------- F3: Explanation Target O7: Function of the Explanation U3: Contextfullness F4: Explanation Breadth/Scope O8: Causality vs. Actionability U6: Chronology F7: Relation to the Predictive System U7: Coherence U8: Novelty U9: Complexity U10: Personalisation U11: Parsimony Given the complex nature of such systems, it would be expected that some of these objectives might be at odds with each other, their definition may be “fuzzy”, they might be difficult to operationalise, their “correct” application might depend on the use case, etc. Furthermore, striking the right balance between these desiderata can be challenging. Nevertheless, we argue that considering them while designing interactive explainers will improve the overall quality of the system, help the designers and users understand their strengths and limitations, and make the interaction feel more natural to humans. Furthermore, some of these desired properties can be achieved (and “optimised” for the explainees) by simply allowing user interaction, thereby alleviating the need of explicitly building them into the system. For example, interactive *personalisation* of the explanations (online, with user input) can mean that it does not have to be solved fully algorithmically offline. The main advantage of GlassBox interactiveness is the explainee’s ability to transfer knowledge onto the system – in this particular case various preferences with respect to the desired explanation – thereby *personalising* the resulting explanation [@sokol2020explainability property U10, see Table \[tab:chap07:glass-box\_desiderata\]]. In our experience, personalisation can come in many different shapes and forms, some of which are discussed below. By interacting with the system the explainee should be able to adjust the *breadth and scope* of an explanation [@sokol2020explainability property F4]. Given the complexity of the underlying predictive model, the explainee may start by asking for an explanation of a *single data point* (blackbox prediction) and continue the interrogation by generalising it to an explanation of a *subspace of the data space* (a cohort explanation) with the final stage entailing the explanation of the entire blackbox model. Such a shift in explainee’s interest may require the explainability method to adapt and respond by changing the *target* of the explanation [@sokol2020explainability property F3]. The user may request an explanation of a single data point or a summary of the whole data set (training, test, validation, etc.), an explanation of a predictive model (or its subspace) or any number of its predictions. Furthermore, interactive personalisation of an explanation can increase the overall versatility of such systems as customised explanations may serve different purposes and have different *functions* [@sokol2020explainability property O7]. An appropriately phrased explanation may be used as an evidence that the system is *fair* – either with respect to a group or an individual depending on the scope and breadth of the explanation – or that it is *accountable*, which again can be investigated with a varied scope, for example, a “What if?” question uncovering that two seemingly indistinguishable data points yield significantly different class assignment, aka adversarial examples [@goodfellow2015explaining]. Importantly, if the explainer is flexible enough and the interaction allows such customisation, however the explanations were designed to serve only one purpose, e.g., transparency, the explainee should be explicitly warned of such limitations to avoid any unintended consequences. For example, the explanations may be counterfactually actionable but they are not causal as they were not derived from a causal model [@sokol2020explainability property O8]. Some of the aforementioned principles can be observed in how GlassBox operates. The contrastive statements about the underlying blackbox model can be used to assess its transparency (their main purpose), fairness (disparate treatment via contrastive statements conditioned on protected attributes) and accountability (e.g., answers to “What if?” questions that indicate an unexpected nonmonotonic behaviour). The contrastive statements are personalised via userspecified constrains of the conditional part (foil) of the counterfactual explanation and by default are with respect to a single prediction. Cohortbased insights can be retrieved by asking “What if?” questions with regard to counterfactual explanations generated by GlassBox – Section \[sec:chap07:discussion\] discusses how the scope and the target of our explanations can be broadened to global explanations of the blackbox model. Given the wide range of possible explanations and their uses some systems may produce contradictory or competing explanations. GlassBox is less prone to such issues as the employed explainer is antehoc [@sokol2020explainability property F7], i.e., predictions and explanations are derived from the same ML model, hence they are always truthful with respect to the predictive model. This means that contradictory explanations are indicative of flaws in the underlying ML model, hence can be very helpful in improving its accountability. In daytoday human interactions we are able to communicate effectively and efficiently because we share common background knowledge about the world that surrounds us – a mental model of how to interact with the world and other people [@kulesza2012tell]. Often, humanmachine interactions lack this implicit link making the whole process feel unnatural and frustrating. Therefore, the creators of interactive explainability techniques should strive to make their systems *coherent* with the explainee’s mental model to mitigate this phenomenon as much as possible [@sokol2020explainability property U7]. While this objective may not be achievable in general, modelling a part of the user’s mental model, however small, can make a significant difference. The two main approaches to extracting an explainee’s mental model are interactive querying of the explainee in an iterative dialogue (online), or embedding the user’s characteristics and preferences in the data or in the parameters of the explainer (offline), both of which are discussed in Section \[sec:chap07:related\_work\]. For explainability systems this task is possible to some extent as their operation and purpose are limited in scope in contrast to more difficult tasks like developing a generic virtual personal assistant. Designers of such systems should also be aware that many interactions are underlined by implicit assumptions that are embedded in the explainee’s mental model and perceived as mundane, hence not voiced, for example, the context of a *followup* question. However, for humanmachine interactions the *context* and its dynamic changes can be more subtle, which may cause the coherence of the internal state of an explainer and the explainee’s mental model to diverge [@sokol2020explainability property U3]. This issue can be partially mitigated by explicitly grounding explanations in a context at certain stages, for example, whenever the context shifts, which will help the users to adapt by updating their mental model and assumptions. Contextfullness will also help the explainee better understand the limitations of the system, e.g., whether an explanation produced for a single prediction can (or must not) be generalised to other (similar) instances: “this explanation can be generalised to other data points that have all of the feature values the same but feature $x_5$, which can span the $0.4 \leq x_5 < 1.7$ range.” Regardless of the system’s interactivity, the explanations should be *parsimonious* – as short as possible but not shorter than necessary – to convey the required information without overwhelming the explainee [@sokol2020explainability property U11]. Maintaining a mental model of the user can help to achieve this objective as the system can provide the explainee only with *novel* explanations – accounting for factors that the user is not familiar with – therefore reducing the amount of information carried by the explanation [@sokol2020explainability property U8]. Another two usercentred aspects of an explanation are its *complexity* and granularity [@sokol2020explainability property U9]. The complexity of explanations should be adjusted according to the depth of the technical knowledge expected of the intended audience, and the level of detail chosen appropriately for their intended use. This can either be achieved by design (i.e., incorporated into the explainability technique), be part of the system configuration and parametrisation steps (offline) or adjusted interactively by the user as part of the explanatory dialogue (online). Another aspect of an explanation, which is often expected by humans [@miller2018explanation], is the *chronology* of factors presented therein: the explainee expects to hear more recent events first [@sokol2020explainability property U6]. While this property is data setspecific, the explainee should be given the opportunity to trace the explanation back in time, which can easily be achieved via interaction. GlassBox attempts to approximate its users’ mental models by mapping their interests and interaction context (inferred from posed questions) to data features that are used to compose counterfactual explanations. Memorising previous interactions, their sequence and the frequency of features mentioned by the user help to achieve this goal and avoid repeating the same answers – once all of the explanations satisfying given constraints were presented, the system explicitly states this fact. Contextfullness of explanations is based on user interactions and is implicitly preserved for followup queries in case of actions that do not alter the context and are initiated by the user – e.g., interrogative dialogue. Whenever the context shifts – e.g., a new personalised explanation is requested by the user or an interaction is initiated by GlassBox – it is explicitly communicated to the user. Contrastive explanations are inherently succinct, but a lack of parsimony could be observed for some of GlassBox explanations, which resulted in a long “monologue” delivered by the system. In most of the cases this was caused by the system “deciding” to repeat the personalisation conditions provided by the user to ensure their coherence with the explainee’s mental model. GlassBox is capable of producing novel explanations by using features that have not been acknowledged by the user during the interaction. Interestingly, there is a tradeoff between novelty of explanations and their coherence with the user’s mental model, which we have not explored when presenting our system but which should be navigated carefully to avoid jeopardising explainee’s trust. GlassBox was built to explain predictions of the underlying ML model and did not account for possible generalisation of its explanations to other data points (the users were informed about it prior to interacting with the device). However, the explainees can ask “What if?” questions with respect to the counterfactual explanations, e.g., using slight variations of the explained data point, to explicitly check whether their intuition about the broader scope of an explanation holds up. Finally, chronology was not required of GlassBox explanations as the data set used to train the underlying predictive model does not have any timeannotated features. GlassBox Properties\[sec:chap07:glass-box\_properties\] ------------------------------------------------------- In addition to a set of interactive explainability system desiderata, we consider a number of their general properties and requirements that should be considered prior to their development. These are summarised in Table \[tab:chap07:glass-box\_properties\] and discussed below. **Operational** **Usability** **Safety** --------------------------------- --------------------- ---------------------------- O1: Explanation Family U4: Interactiveness S3: Explanation Invariance O2: Explanatory Medium O3: System Interaction O4: Explanation Domain O5: Data and Model Transparency O6: Explanation Audience O10: Provenance Assuming that the system is interactive, the *communication protocol* between the explainee and the explainer should be carefully chosen to support the expected input and deliver the explanations in the most natural way possible. For example, clearly indicating which parts of the explanation can be personalised and the limitations of this process should be disclosed to the user [@sokol2020explainability property O3, see Table \[tab:chap07:glass-box\_properties\]]. The choice of *explanatory medium* used to convey the explanation is also crucial. Plots, interactive or not, can be very informative, but may not convey the whole story due to the curse of dimensionality and the limitations of the human visual system [@sokol2020explainability property O2]. Supporting visualisations with textual description can greatly improve their intelligibility, and vice versa, nevertheless in some cases this approach may be suboptimal, for example, explaining images using only a natural language interface. The *intended audience* should be considered in conjunction with the communication protocol to choose a suitable explanation type [@sokol2020explainability property O6]. Domain experts may prefer explanations expressed in terms of the internal parameters of the underlying predictive model, but a lay audience may rather prefer exemplar explanations that use relevant data points – choosing the appropriate *explanation domain* [@sokol2020explainability property O4]. The audience also determines the purpose of the explanation. For example, inspecting a predictive model for debugging purposes will need a different system than guiding the explainee with an actionable advice towards a certain goal like getting a loan. Interactive explainers can support a wide spectrum of these properties by allowing the explainee to personalise the output of the explainer as discussed in Section \[sec:chap07:glass-box\_desiderata\]. Achieving some of these objectives may require the features of the underlying data set or the predictive model itself to be *transparent* [@sokol2020explainability property O5]. For example, consider explaining a model trained on a data set with features that are object measurements in meters in contrast to magnitudes of embedding vectors. When the raw features (original domain) are not humaninterpretable, the system designer may decide to use an interpretable representation (transformed domain) to aid the explainee. Providing the users with the *provenance* of an explanation may help them to better understanding its origin, e.g., an explanation purely based on data, model parameters or both [@sokol2020explainability property O10]. Choosing the right *explanation family* is also important, for example: relation between data features and the prediction, relevant examples such as similar data points or causal mechanisms [@sokol2020explainability property O1]. Again, interactive explainers have the advantage of giving the user the opportunity to switch between multiple different explanation types. Furthermore, the design of the user interface should be grounded in the Interactive Machine Learning, HumanComputer Interaction, User Experience and Explainable Artificial Intelligence research to seemingly deliver the explanations. For example, the explainee should be given the opportunity to reverse the effect of any actions that may influence the internal state of the explainer and the system should always respect user’s preferences and feedback [@sokol2020explainability property U4]. Finally, if an explanation of the same event can change over time or is influenced by a random factor, user’s trust is at stake. The explainee should always be informed about the degree of *explanation invariance* and its manifestation in the output of an explainer [@sokol2020explainability property S3]. This property is vital to GlassBox’s success, which we discuss in more detail in Section \[sec:chap07:discussion\]. GlassBox Reception and Feedback\[sec:chap07:glass-box\_feedback\] ----------------------------------------------------------------- We presented GlassBox to domain experts (general AI background knowledge) and a lay audience with the intention to gauge their reception of our prototype and collect feedback that would help us revise and improve our explainability system. To this end, we opted for informal and unstructured freeform feedback, which was mostly userdriven and guided by reference questions (based on our list of desiderata) whenever necessary. We decided to take this approach given the nature of the events at which we presented our prototype – a scientific conference and a research festival. GlassBox is composed of multiple independent components, all of which play a role in the user’s reception of the system: - natural language understanding and generation, - speech transcription and synthesis, - voice and text user interfaces, and - a data set that determines the problem domain. Therefore, collecting freeform feedback at this early stage helped us to pinpoint components of the system that required more attention and identify possible avenues for formal testing and design of user studies. While presenting the device we only approached members of the audience who expressed an interest in interacting with the device and who afterwards were willing to describe their experience. In total, we collected feedback from 6 domain experts and 11 participants of the research festival of varying demographics. When introducing the system and its modes of operation to the participants, we assessed their level of AI and ML expertise by asking background questions, which allowed us to appropriately structure the feedback session. While discussing the system with the participants, we were mainly interested in their perception of its individual components and suggestions about how these can be improved. Most of the participants enjoyed asking questions and interacting with the device via the voice interface, however some of them found the speech synthesis module that answered their questions “slow”, “unnatural” and “clunky”. These observations have prompted some of the participants to disable voicebased responses and use the textbased chat interface to read the answers instead of listening to them. When asked about the quality of explanations, their comprehensibility and content, many participants were satisfied with received answers. They claimed that personalised explanations provided them with information that they were seeking for as opposed to the default explanation given at first. However, some of them expressed concerns regarding the deployment of such systems in everyday life and taking the human out of the loop. The most common worry was the impossibility to “argue” and “convince” the explainer that the decision is incorrect and the explanation does not capture the complexity of one’s case. Some participants were also sceptical of the general idea of interacting with an AI agent and the failsafe mode of the device, which produced “I cannot help you with this query.” response whenever the explainer could not answer the user’s question. We plan to use all of this feedback and our experience in building interactive explainers to refine the system focusing on its explanation personalisation aspect, and test this particular component with formal user studies. Isolating this module of the explainer will alleviate the influence of the user interface on the perception of the explanations, allowing us to investigate the effectiveness and reception of personalised explanations in a formal setting. Discussion\[sec:chap07:discussion\] =================================== Developing GlassBox and demonstrating it to a diverse audience provided us with a unique experience of building, deploying and refining interactive explainers. To help researchers and engineers with a similar agenda we summarise the lessons learnt in Section \[sec:chap07:lessons\]. We also discuss our next steps in interactive and personalised explainability research in Section \[sec:chap07:future\] to draw attention to interesting open questions. Lessons Learnt\[sec:chap07:lessons\] ------------------------------------ The major challenge of building GlassBox was the development overhead associated with setting up the hardware and software needed to make it voiceenabled and capable of processing the natural language. While readymade components were adapted for these purposes, the effort required to build such a system is still significant which may not always be justified. We encourage researchers to build such a system if the research value lies in the system itself or it is used as a means to an end, for example, research on interactive explainability systems. In this case, one should be aware of generalisability issues as each new data set used within such a system must be adapted by preparing appropriate annotations and (possibly) training a new natural language processing model. In many cases, based on our observations, it seems that all this effort is only justified when the creator of the system is committed to deploying it in real life. For research purposes, however, the engineering overhead can be overwhelming, in which case we suggest using the *Wizard of Oz* studies [@dahlback1993wizard] as an accessible alternative. Once GlassBox was operational, its major usability barrier was the timeconsuming process of inputting personal data when roleplaying the loan application process. At first, we implemented this step as a voicedriven questionanswering task but even with just 13 attributes (most of which were categorical) this proved to be a challenge for the explainee. We overcame this issue by predefining 10 individuals whom the explainee could impersonate. We then allowed the explainee to further customise the attributes of the selected individual by asking GlassBox to edit them (with voice and textbased commands). In hindsight, we believe that this kind of task should be completed by using a dedicated input form (e.g., a questionnaire delivered as a web page), thereby giving the explainee the full control of the data input process and mitigating the lengthy “interrogation” process. The interactive aspect of GlassBox (discussed in length in Section \[sec:chap07:glass-box\_design\]) provides many advantages from the point of view of explainability. For example, it enables the explainee to assess individual fairness of the underlying predictive model and personalise the explanations (see Sections \[sec:chap07:glass-box\_desiderata\] and \[sec:chap07:glass-box\_properties\] for more details). However, not all types of explainability algorithms allow for the resulting explanation to be interactively customised and personalised, restricting the set of tools that can be deployed in such a setting. If incorporating the user feedback (delivered as part of the interaction, e.g., via argumentation [@madumal2019grounded]) into the underlying predictive model is desired, this model has to support refinements beyond the training phase, further reducing the number of applicable Machine Learning and explainability techniques. As noted in Section \[sec:chap07:glass-box\_desiderata\], some of the interactivity and personalisation desiderata cannot be achieved without “simulating” the explainee’s mental model. While we believe that cracking this problem will be a corner stone of delivering explanations that feel natural to humans, we do not expect it to be solved across the board in the near future. In case of GlassBox, where the explanations are presented to the user as counterfactual statements, we observed a tendency amongst the explainees to generalise an explanation of a single data point to other, relatively similar, instances. However tempting, GlassBox explanations cannot be generalised as they are derived from a predictive model (structure of a decision tree) that does not encode and account for the causal structure of the underlying phenomenon. This can sometimes lead to contradictory explanations, which can be detrimental to the explainee’s trust. Since GlassBox uses an antehoc explainability algorithm (i.e., explanations and predictions are derived from the same ML model), contradictory, incorrect or incoherent explanations are indicative of issues embedded in the underlying predictive model, which should be reported to and addressed by the model creators. However, if a posthoc explainability tool is employed (explanations are not derived directly from the predictive model, e.g., surrogate explainers), contradictory explanations manifest a problem with the system. This issue cannot be uniquely pinpointed and can either be attributed to a lowfidelity explainer or to an underperforming predictive model, putting explainees’ trust at risk. Communicating the limitations of the explanations clearly can help to partially mitigate this problem; grounding the explanations in a context (see Section \[sec:chap07:glass-box\_desiderata\]) is another approach. While truthful to the underlying black box, an antehoc explainability approach may not be available for a chosen predictive model. For example, deep neural networks are intrinsically complex, which encumbers explaining them without resorting to proxies. This observation highlights another tradeoff of AIbased systems: predictive power vs. transparency [@gunning2019darpa]. Simpler models such as decision trees are less expressive but more interpretable. On the other hand, complex models such as deep neural networks are more powerful at the expense of opacity. It is still possible to explain the latter model family with proxies and posthoc approaches, but issues with the fidelity and truthfulness of such explanations can be unacceptable, e.g., in highstakes situations such as criminal justice or financial matters [@rudin2019stop]. These conclusions have led some researchers (e.g., @rudin2019stop) to deem posthoc explainers as outright harmful. Instead, they argue, developers behind predictive systems for highstake applications should invest more time in feature engineering and restrict their toolkit to inherently transparent ML models. As might be expected, the power and flexibility of GlassBox explanations comes at a cost. The interactiveness of the process enables malicious users to ask for explanations of arbitrary data points, which in large quantities may expose internals of the underlying predictive model [@sokol2020explainability properties S1 and S2]. Adversaries can misuse the information leaked by the system in an attempt to reverseengineer the underlying predictive model (which may be proprietary) or use this knowledge to game it. This is particularly visible in GlassBox as the condition of the counterfactual explanations is based on one of the splits in the underlying decision tree, thereby revealing the exact threshold applied to one of the features, e.g., “had you been older than 25,…” implies `age > 25` internal splitting node. Since every explanation reveals a part of the tree structure (at least one split), with a certain budget of queries the adversary can reconstruct the tree. This issue is intrinsic to antehoc explainers but may also affect highfidelity posthoc approaches, albeit to a lesser extent since in the latter case the explanations are not generated directly from the blackbox model. This undesired sideeffect can be controlled to some extent by limiting the explanation query budget for untrustworthy users or obfuscating the precise (numerical) thresholds. The latter can be achieved either by injecting random noise (possibly at the expense of explainees’ trust) or replacing the numerical values with *quantitative adjectives*, e.g., “slightly older” (shown to enhance user satisfaction [@biran2014justification]). The tradeoff between transparency and security of interactive explainers should be explicitly considered during their design stage, with appropriate mitigation technique implemented and documented. Next Steps\[sec:chap07:future\] ------------------------------- One of the main contributions of GlassBox lies in the composition of its software stack and hardware architecture. While investigating the challenge of readying such a system for a deployment is one possible avenue for future research, we believe that a more interesting direction is to design explainability tools and techniques that facilitate (interactive) personalisation of their explanations. Since the latter research aspect is conditioned upon the availability of the former, we plan to use the Wizard of Oz approach [@dahlback1993wizard] to mitigate the need for building an interactive user interface that is responsible for processing the natural language. In this scenario, the input handling and the output generation are done by a human disguised as an intelligent interface, who can access all the components of the tested explainability approach and is only allowed to take predefined actions. Therefore, bypassing an algorithmic natural language interface by using the Wizard of Oz [@dahlback1993wizard] approach will allow us to focus our research agenda on designing and evaluating the properties of personalised explanations. It will also ensure that our findings are not adversely affected by poor performance of the natural language interface. To this end, we will use a bespoke surrogate explainer of blackbox predictions similar to the Local Interpretable ModelAgnostic Explanations (LIME) [@ribeiro2016why] algorithm. Our explainer uses decision trees as the local surrogate model [@sokol2019blimey], whereas LIME is based on a sparse linear regression. Among others, this modification allows us to improve the fidelity of the explainer by reducing the number of conflicting explanations. Furthermore, a treebased surrogate inherits the best of both worlds: the explainer is modelagnostic, hence it can be used with any blackbox model, and it can take advantage of the wide variety of explanations supported by decision trees (as discussed in Section \[sec:chap07:intro\]). We plan to apply this approach to three different data domains: tabular data, text, and images, which tests its capacity of interactively generating personalised explanations for a range of tasks. We expect object recognition for images and sentiment analysis for text to be the most effective evaluation tasks as they do not require any background knowledge. In our studies, the explainees will be asked to interactively personalise two aspects of the explanations: an interpretable representation of the data features and their content. The objective of the first task depends on the data domain. For text, it will allow the explainees to adjust and introduce the concepts that can only be expressed with multiple words since the default interpretable representation would be a bag of words. For images, the users will modify the default superpixel segmentation to separate semantically meaningful regions – see the example shown in Figure \[fig:chap07:segmentation\] and discussed in Section \[sec:chap07:intro\]. For tabular data, the interpretable representation is achieved by discretising continuous features. Since the local surrogate model is a decision tree, this representation is learnt automatically and cannot be explicitly modified by the explainee. We will give the users indirect control over the feature splits by allowing them to adjust the tree structure in terms of depth, the number of data points required for a split and the minimum number of data points per leaf. The second personalisation objective will allow the explainee to choose the explanation type and customise it accordingly. The visualisation of the surrogate tree structure can either depict the whole tree or zoom in on its selected part. The explainee will also be able to inspect treebased feature importance either by viewing all of them in a list or by querying the importance of the chosen ones. These two explanation types will allow the user to grasp the overall behaviour of the blackbox model in the vicinity of the explained data point. For text and images these will be the interactions between the words and superpixels in that region, i.e., within a sentence and an image respectively, and for tabular data the influence of raw features and ranges of their values. Furthermore, the explainee will be able to get personalised explanations of individual predictions. A counterfactual retrieved from the local tree, e.g., “had these two superpixels/words not been there, the image/sentence would be classified differently”, will be customised by specifying constraints appearing in its condition. Next, the explainee can request a logical rule, e.g., “these three superpixels/words must be present and these two must be removed”, for any leaf in the tree, which will be extracted from the corresponding roottoleaf path. Both of these explanations will allow the user to understand how parts of an image or a sentence (superpixels and words respectively) come together to predict a data point. Finally, the user can view exemplar explanations of any prediction, which come from the part of the surrogate training set – generated by perturbing the selected data point, possibly in the interpretable domain – assigned to the relevant tree leaf. This means that the exemplars will be images with occluded superpixels, sentences with missing words and, for tabular data, slight variations of the explained data point in its original feature space. We believe that this diverse set of personalised explanations will encourage the user to investigate different aspects of the blackbox model thus lead to a better understanding of its behaviour. Interactions form another aspect of a system that delivers a multitude of different explanation types. A user who has learnt which features are important may want to know whether one of the counterfactual explanations is conditioned upon them. In particular, we want to investigate whether the user would discount the counterfactual explanations conditioned upon unimportant features and focus on the ones that include important factors. Also, we are interested in how the user’s confidence is affected upon discovering that most of the (counterfactual) explanations are based on features indicated as unimportant by a different explanation type. As part of the study we aim to recognise current limitations of the interaction and personalisation aspects of the system by taking note of the requests that failed from the user’s perspective. The possibility of retrieving multiple counterfactuals of the same length (the same number of conditions) brings up the question of their ordering. One approach is to use a predefined, featurespecific “cost” of including a condition on that feature into the explanation. This heuristic can be based on the purity (accuracy) of the counterfactual leaf, the cumulative importance of features that appear on the corresponding roottoleaf path, the collective importance of features listed in its conditional statement or, simply, the number of training data points falling into that leaf [@waa2018contrastive]. However, a more usercentred approach is to allow the explainee to supply this information either implicitly or explicitly during the interaction. To improve the quality of the interactions, we will build a partial mental model of the explainee using a formal argumentative [@dung2009assumption] dialogue introduced by @madumal2019grounded. Many user arguments can be parsed into logical requirements, allowing for further personalisation and more convincing explanations. The roles in this dialogue can also be reversed to assess and validate the explainee’s understanding of the blackbox model – the machine questioning the human [@walton2011dialogue; @walton2016dialogue]. In this interrogative dialogue, if an insight about the black box voiced by the user is incorrect, the system can provide a personalised explanation in an attempt to correct explainee’s beliefs. Asking the user “What if?” questions can further assist in this task by directing the explainee’s attention towards evidence relevant to the preconceived misconceptions. When an interaction is finished, a succinct excerpt summarising the whole explanatory process (similar to a court transcript) can be provided to the user as a reference material. This document should only contain explanations that the user has challenged or investigated in detail, avoiding the ones that agree with the explainee’s mental model. The mental model can also be utilised to adjust the granularity and complexity of explanations. For example, a disease can be explained in medical terms – e.g., on a bacterial level – or with easily observable external symptoms – e.g., cough and abnormal body temperature – depending on the audience. While solving this task for a generic case is currently not feasible, we will investigate possible approaches for a data set that exhibits a hierarchy of lowlevel features, which can be handcrafted and incorporated into the explainer. Summary and Conclusions ======================= In this paper we discussed the benefits that Interactive Machine Learning can bring to eXplainable Artificial Intelligence and Interpretable Machine Learning. We showed how personalised explanations can improve the transparency of a Machine Learning model and how they can be generated via a humanmachine interaction. While other aspects of an explainability system can also be made interactive, we argued that one of the major benefits comes from personalisation. In particular, we showed the difference between interactiveness of an explainability system – e.g., interactive plots – and interactiveness of an explanation – e.g., personalisation. We supported our discussion and claims with experience gained from building and demonstrating GlassBox: a classcontrastive counterfactual explainability system that communicates with the user via a natural language dialogue. To the best of our knowledge, this is the first XAI system tested in the wild that supports explanation customisation and personalisation via interaction. Our experience building GlassBox and experimenting with it helped to identify a collection of desired functionality and a set of properties that such systems should have. We discussed which ones are applicable to GlassBox, and summarised a list of lessons that we have learnt. The most important one draws attention to the engineering overhead required to build such a system despite adapting many offtheshelf components. We concluded that one should avoid this effort in favour of Wizard of Oz studies when the main objective is to use such a platform as a test bed for various explainability techniques, unless the intention is to deploy it afterwards. Other key observations concerned both the importance and impossibility of simulating an explainee’s mental model. While doing so is highly beneficial, fully satisfying this requirement is out of reach at present. Nevertheless, we observed that by using a formal argumentation framework to model a part of the usermachine interactions, it may be possible to extract some relevant knowledge from the explainee, which can be utilised to this end. To ground our study we have reviewed relevant literature, where we identified three related research strands and showed how our work has the potential to bridge them together. Our investigation has shown that while some of the explainability algorithms and tools are capable of explanation personalisation via user interaction, many more are not. A number of other explainability approaches, such as contrastive explanations, can easily support such interactions, however their implementations lack this functionality. Finally, these observations combined with our findings helped us to devise next steps for our research, which pivot around investigating personalised explanations and their properties in a more principled way. [^1]: A *protected group* is a subpopulation in a data set created by fixing a value of a protected attribute such as *age*, *gender* or *ethnicity*, which discriminating upon is illegal. [^2]: To clarify, the notion of interaction is with respect to the explanation, e.g., the ability of the explainee to personalise it, and not the overall interactiveness of the explainability system. [^3]: Building such systems requires a range of diverse components: user interface, natural language processing unit, natural language generation module, conversation management system and a suitable and welldesigned XAI algorithm. Furthermore, most of these components are domainspecific and cannot be generalised beyond the selected data set and use case. [^4]: <https://pair.withgoogle.com> [^5]: <https://pair-code.github.io/what-if-tool/> [^6]: The festival spans a wide range of research projects both in social sciences and engineering. [^7]: <https://aiyprojects.withgoogle.com/voice> [^8]: <https://github.com/RasaHQ/rasa> [^9]: <https://archive.ics.uci.edu/ml/datasets/statlog+(german+credit+data)> [^10]: <https://scikit-learn.org/>

ArXiv

--- abstract: 'We study matrix sketching methods for regularized variants of linear regression, low rank approximation, and canonical correlation analysis. Our main focus is on sketching techniques which preserve the objective function value for regularized problems, which is an area that has remained largely unexplored. We study regularization both in a fairly broad setting, and in the specific context of the popular and widely used technique of ridge regularization; for the latter, as applied to each of these problems, we show algorithmic resource bounds in which the [*statistical dimension*]{} appears in places where in previous bounds the rank would appear. The statistical dimension is always smaller than the rank, and decreases as the amount of regularization increases. In particular, for the ridge low-rank approximation problem $\min_{Y,X} {\lVert YX - A \rVert}_F^2 + \lambda{\lVert Y \rVert}_F^2 + \lambda{\lVert X \rVert}_F^2$, where $Y\in{{\mathbb R}}^{n\times k}$ and $X\in{{\mathbb R}}^{k\times d}$, we give an approximation algorithm needing $O(\operatorname{\mathtt{nnz}}(A)) + {\tilde{O}}((n+d){\varepsilon}^{-1}k \min\{k, {\varepsilon}^{-1}\operatorname{\mathtt{sd}}_\lambda(Y^*)\})+ {{\mathrm{poly}}}(\operatorname{\mathtt{sd}}_\lambda(Y^*) \epsilon^{-1})$ time, where $s_{\lambda}(Y^*)\le k$ is the statistical dimension of $Y^*$, $Y^*$ is an optimal $Y$, ${\varepsilon}$ is an error parameter, and $\operatorname{\mathtt{nnz}}(A)$ is the number of nonzero entries of $A$. This is faster than prior work, even when $\lambda=0$. We also study regularization in a much more general setting. For example, we obtain sketching-based algorithms for the low-rank approximation problem $\min_{X,Y} {\lVert YX - A \rVert}_F^2 + f(Y,X)$ where $f(\cdot,\cdot)$ is a regularizing function satisfying some very general conditions (chiefly, invariance under orthogonal transformations).' author: - 'Haim Avron[^1]\' - 'Kenneth L. Clarkson[^2]\' - 'David P. Woodruff[^3]' bibliography: - 'p.bib' title: Sharper Bounds for Regularized Data Fitting --- Basic Definitions and Notation ------------------------------ We denote scalars using Greek letters. Vectors are denoted by $x,y,\dots$ and matrices by $A,B,\dots$. We use the convention that vectors are column-vectors. We use $\operatorname{\mathtt{nnz}}(\cdot)$ to denote the number of nonzeros in a vector or matrix. We denote by $[n]$ the set ${1,\dots,n}$. The notation $\alpha = (1 \pm \gamma)\beta$ means that $(1- \gamma)\beta \leq \alpha \leq (1 + \gamma)\beta$. Throughout the paper, $A$ denotes an $n \times d$ matrix, and $\sigma_1 \geq \sigma_2 \ge \cdots \ge \sigma_{\min(n,d)}$ its singular values. The *Schatten $p$-norm* of $A$ is ${\lVert A \rVert}_{(p)}\equiv \left[\sum_i \sigma_i^p\right]^{1/p}$. Note that the trace (nuclear) norm ${\lVert A \rVert}_*={\lVert A \rVert}_{(1)}$, the Frobenius norm ${\lVert A \rVert}_F = {\lVert A \rVert}_{(2)}$, and the spectral norm ${\lVert A \rVert}_2 = {\lVert A \rVert}_{(\infty)}$. The notation ${\lVert \cdot \rVert}$ without a subscript denotes the $\ell_2$ norm for vectors, and the spectral norm for matrices. We use a subscript for other norms. We use $\operatorname{\mathtt{range}}(A)$ to denote the subspace spanned by the columns of $A$, i.e. $\operatorname{\mathtt{range}}(A) \equiv \{Ax\mid x\in{{\mathbb R}}^d\}$. ${{I}}_d$ denotes the $d\times d$ identity matrix, $0_d$ denotes the column vector comprising $d$ entries of zero, and $0_{a\times b}\in{{\mathbb R}}^{a\times b}$ denotes a zero matrix. The rank $\operatorname{\mathtt{rank}}(A)$ of a matrix $A$ is the dimension of the subspace $\operatorname{\mathtt{range}}(A)$ spanned by its columns (equivalently, the number of its non-zero singular values). Bounds on sketch sizes are often written in terms of the rank of the matrices involved. The *stable rank* $\operatorname{\mathtt{sr}}(A) \equiv {\lVert A \rVert}_F^2/{\lVert A \rVert}_2^2$. The stable rank satisfies $\operatorname{\mathtt{sr}}(A)\le \operatorname{\mathtt{rank}}(A)$. Omitted Proofs ============== Proof of Lemma \[lem reg\] {#subsec lemregproof} -------------------------- Proof of Lemma \[lem U1 size\] {#subsec lemUonesizeproof} ------------------------------ Proof of Lemma \[lem lam large\] {#subsec lemlamlargeproof} -------------------------------- Proof of Corollary \[cor size of S\] {#subsec corsizeofSproof} ------------------------------------ Proof of Theorem \[thm reg stacked\] {#subsec thmregstacked} ------------------------------------ Proof of Theorem \[thm:twoProp\] {#subsec thmtwoPropproof} -------------------------------- Acknowledgments {#acknowledgments .unnumbered} =============== The authors acknowledge the support from the XDATA program of the Defense Advanced Research Projects Agency (DARPA), administered through Air Force Research Laboratory contract FA8750-12-C-0323. [^1]: Tel Aviv University [^2]: IBM Research - Almaden [^3]: IBM Research - Almaden

ArXiv

--- abstract: '[Salient object detection is evaluated using binary ground truth with the labels being salient object class and background. In this paper, we corroborate based on three subjective experiments on a novel image dataset that objects in natural images are inherently perceived to have varying levels of importance. Our dataset, named SalMoN (saliency in multi-object natural images), has 588 images containing multiple objects. The subjective experiments performed record spontaneous attention and perception through eye fixation duration, point clicking and rectangle drawing. As object saliency in a multi-object image is inherently multi-level, we propose that salient object detection must be evaluated for the capability to detect all multi-level salient objects apart from the salient object class detection capability. For this purpose, we generate multi-level maps as ground truth corresponding to all the dataset images using the results of the subjective experiments, with the labels being multi-level salient objects and background. We then propose the use of mean absolute error, Kendall’s rank correlation and average area under precision-recall curve to evaluate existing salient object detection methods on our multi-level saliency ground truth dataset. Approaches that represent saliency detection on images as local-global hierarchical processing of a graph perform well in our dataset.]{}' address: author: - ', , , and' bibliography: - 'literature.bib' title: 'Evaluating Salient Object Detection in Natural Images with Multiple Objects having Multi-level Saliency' --- Introduction {#sec:introduction} ============ Visual saliency signifying the importance of an entity in an image is postulated to guide the deployment of human visual attention [@Baluch11], where both top-down and bottom-up processing are involved. As only the attended entities drive image understanding [@Baluch11; @Neibur98], saliency computation acts as a strategy used by humans to judiciously use the limited resources available by considering only pertinent visual sensory data [@Neibur98]. Salient object detection uses saliency to perform object detection in images, where objects are detected as important entities as opposed to the background. In recent studies, salient object detection has been evaluated for performance using binary images as ground truth, testing their capability to detect a single salient object class against the background [@Han2018]. [As shown in the examples of Figures \[fig:datasetComparisons\] and , each image in the well-known salient object detection evaluation datasets MSRA-1000 [@Achanta_frequency_tuned], SED-100 [@Alpert_image_segmentation], SOD [@Movahedi_design_and], PASCAL-S [@Sal_secret], SOC [@fan2018SOC], and DUT-OMRON [@Yang_saliency_detection] has a corresponding binary ground-truth image /map to be used for evaluation.]{} For an image with multiple objects, a binary ground truth image for evaluation assumes that all the objects in the image are either of equal importance or some are not salient at all. The important objects fall in the salient object class with others falling in the background class. Note that the works proposing the SOD and PASCAL-S datasets do consider unequal importance of objects in their dataset creation, but do not use it for ground truth generation and evaluation. Assuming equal importance of salient objects, the use saliency for object detection has progressed to object-level abstraction [[@Zhou2019]]{} and video salient object detection [[@Chen2018]]{}. [Firstly, in this paper, with the help of subjective experiments we corroborate the findings of [@Movahedi_design_and; @Sal_secret] that objects in natural images are seen and perceived to have varying levels of importance.]{} The saliency of objects in images with multiple objects is found to be multi-level through a study of spontaneous human visual attention on and visual perception of objects. [*Spontaneous Human Attention through Eye Fixation Duration:* ]{} According to Henderson et al. [@Henderson_eye_movements] and Yarbus [@Yarbus_eye_movements], a person spontaneously attends to visually distinctive image regions by fixating on them while free-viewing, and this spontaneous attention is mainly achieved by bottom-up processing involving distinctness computation. The fixation duration (when perception happens), which may differ from person to person, is usually longer for regions that are more informative [@Henderson_eye_movements; @Yarbus_eye_movements]. Therefore, recording eye fixations at a constant rate yields more fixations at higher informative regions indicating higher importance/saliency. Different fixation durations represented by different fixation densities at different objects shows that object saliency is inherently multi-level. [*Human Perception through Point Clicking and Rectangle Drawing:* ]{} While spontaneous attention driven by visual distinction is required for perceiving (during fixation) the information presented by an image region, activities such as point clicking and rectangle drawing are usually performed after perception. Hence, when a person clicks on or draws a rectangle around an image region by attending to them upon asked to do so for important regions, we may assume that the person considers the region to be salient after perception. This consideration of a region to be salient, which guides the attention that leads to image understanding, involves both bottom-up and top-down processing [@Baluch11]. Variation in the number of persons who consider different objects as salient shows that object saliency is also inherently multi-level after perception. Note that perception recorded through point clicking and rectangle drawing differ as one is related to marking the location of a salient region while the other requires consideration of a salient region’s size and shape. We created a dataset of 588 images having multiple objects per image to perform the said study of spontaneous attention on and perception of objects, and the three subjective experiments were performed as follows: $\ \ $Subjects were asked to free-view the images on a monitor and their eye fixations were recorded using an eye tracker. Here we consider the well-founded assumption that free-viewing images for a short period captures eye movements due to spontaneous attentional shifts [@Peters05]. An example eye fixation density map collected from multiple subjects on a single image is shown in Figure \[fig:multimodalGroundTruth\]. [In total, 95 people participated in this experiment, where each image is viewed by 24 subjects on an average (min: 13, max: 34).]{} $\ \ $Subjects recruited through crowd sourcing were asked to view the images and click on the important objects that they notice at first glance. A set of clicked points collected from multiple subjects on a single image is shown in Figure \[fig:multimodalGroundTruth\]. The “first glance” phrase was used to avoid multiple perceptual processing (by a subject) of a single image entity. [In this experiment, each image on an average is viewed by 33 subjects (min: 24, max: 38).]{} $\ \ $Subjects recruited through crowd sourcing were asked to view the images and draw rectangles around the important objects that they notice at first glance. A set of rectangles collected from multiple subjects on a single image is illustrated in Figure \[fig:multimodalGroundTruth\]. [In this experiment, each image on an average is viewed by 32 subjects (min: 15, max: 50).]{} \ Secondly, in this paper, we propose that as saliencies of objects in images are inherently multi-level, a salient object detection method must be evaluated for its capability to detect all the multi-level salient objects considering their levels of saliency. For this purpose, we generate multi-level (gray-level) images as ground truth using the data collected from the three subjective experiments. The ground truth is generated for each of the 588 images of the new dataset, which we refer to as the SalMoN (saliency in multi-object natural images) dataset, indicating a dataset of natural images with multiple objects having multi-level saliency. As the study of spontaneous attention on and perception of objects is subjective in nature, it is imperative that the ground truth generation must consider the data obtained from many persons collectively to account for inter-person variations [@Wolfe_guided_search]. For every image, three multi-level ground truth images are generated corresponding to each subjective experiment with the [help of ]{}manually marked object boundaries (see Figure \[fig:multimodalGroundTruth0\]) in the images. The multi(gray)-level ground truth images for the fixation density in Figure \[fig:multimodalGroundTruth\], the point-clicking in Figure \[fig:multimodalGroundTruth\] and the rectangle drawing in Figure \[fig:multimodalGroundTruth\] are shown respectively in Figures \[fig:multimodalGroundTruth\], and . Note that the multi-level ground truths should not be confused with normalized fixation density maps used to evaluate saliency map generation algorithms for eye fixation prediction as in [@Judd_learning_to; @Li_visual_saliency]. Unlike fixation density maps, our ground truths are for evaluating salient object detection results and not saliency map based eye fixation prediction results. Our salient object detection evaluation involves assessment of an approach’s capability of giving appropriate object boundaries, unlike evaluation of saliency map based eye fixation prediction. Moreover, our ground truth is based on both human spontaneous attention and perception, unlike fixation density maps that considers only spontaneous attention. Thirdly and finally, we propose simple yet effective performance evaluation approaches to evaluate detection of multiple objects with multi-level saliency, and consider them for evaluating a few (seventeen) well-known and latest methods on the natural images of our dataset. In this evaluation, we perform comparisons between the methods using the three multi-level ground truths [individually as well as collectively]{}. We propose specific ways of using them together such that certain biasing can be avoided. One important observation in the evaluation of salient object detection using our dataset is that the approaches found superior in the standard evaluation using existing datasets with binary ground truth maps may not be superior when applied on images with multiple multi-level salient objects. The rest of the paper is organized as follows. In Section \[sec:relatedWork\], we discuss the data collection procedures and the limitations of a few well-known saliency datasets. In Section \[sec:dataCollection\], we explain our dataset, the subjective experiments, and object saliency measurement from the data collected through the experiments. In Section \[sec:visualSaliencyOfObjects\], we discuss the inherent multi-level nature of object saliency and analyze the different kinds of saliencies measured from the different subjective experiments. In Section \[sec:salientObjectDetectors\], we evaluate state-of-the-art salient-object detectors on our dataset, and in this process, propose suitable approaches to evaluate multi-level salient object detection. Section \[sec:conclusion\], concludes our paper discussing the summary and future scope. \ \ Related Work {#sec:relatedWork} ============ The existing literature related to our work is on dataset generation for evaluation of salient object detection. Although many datasets for salient object detection evaluation exist in literature, we consider the following widely used datasets, MSRA-1000 [@Achanta_frequency_tuned], SED-100 [@Alpert_image_segmentation], SOD [@Movahedi_design_and], PASCAL-S [@Sal_secret] and [DUT-OMRON [@Yang_saliency_detection; @DUTOMRON]]{}, which have been used by some of the state-of-the-art salient object detectors among [@Achanta_frequency_tuned; @Yildirim_fast_accurate; @Cheng_global_contrast; @Jiang_automatic_salient; @Cheng_efficient_salient; @Perazzi_saliency_filters; @Shen_a_unified; @Li_contextual_hypergraph; @Jiang_saliency_detection; @Yang_saliency_detection; @Yildirim_saliency_detection; @Zhang2015; @Li2015; @Peng2017; @Huang2017]. [We also consider the datasets SOC [@fan2018SOC] and AugPASCAL-S [@islamsal18] that have been recently reported.]{} Even though we consider a sample of seven datasets from many existing ones, we do so without loss of generality in relation to the nature of ground truth maps provided and their use in evaluation. Here, we briefly explain the data collection and ground truth generation procedures of these existing datasets and then discuss their representation of multi-level object saliency in natural images with multiple objects. Data Collection and Ground Truth Generation ------------------------------------------- The MSRA-1000 consists of 1000 natural images that are taken from the larger MSRA dataset [@Liu_learning_to]. The images in the original MSRA dataset [@Liu_learning_to] were specifically selected such that they have a single distinctive foreground object. Then, nine subjects were asked to draw a rectangle around the most salient object in each image. In the derived MSRA-1000 dataset, these rectangles are used only to identify a single salient object segmented manually by a single person to produce a binary ground truth image /map as shown in Figure \[fig:datasetComparisons\]. Just like MSRA, all 100 images in the SED-100 dataset were selected such that they have a single clear foreground object. However, unlike MSRA-1000, rectangle drawing was not considered and the salient object in each image was directly segmented by three subjects. The segmentations of the subjects were combined into a binary ground truth image by considering a pixel to be salient when it is marked so by more than one subject. An example image and its ground truth are the ones given in Figure \[fig:datasetComparisons\]. The SOD dataset was formed using the BSD dataset of [@Martin_a_database]. In BSD dataset, 300 images were oversegmented by three persons. That is, the objects and background are divided into multiple sub-parts. To create the SOD dataset, seven subjects were asked to identify salient object/s by combining BSD sub-segments. Unlike the other existing datasets, during the creation of SOD, subjects were required to rank the objects with respect to their saliency, if they detect more than one salient object. As evident, this salient object ranking is an attempt to represent multi-level object saliency in images with multiple objects. Although salient object ranking is available in SOD dataset, state-of-the-art salient object detection evaluation using it ignores the ranking and considers only a binary ground truth image with all salient objects in a single salient object class as shown in Figure \[fig:datasetComparisons\]. The PASCAL-S dataset was generated from the validation set of PASCAL 2010 dataset [[@VOC2010]]{} containing 850 images. First, all the objects in all the images were manually segmented to separate them from the background. Then 12 subjects were asked to click on the salient objects. They were free to take their time and click as many objects as salient as they wanted. After the experiment, each object in an image was assigned a saliency value given by the ratio of the number of clicks to the number of subjects. As evident, this assignment of different saliency values to different objects is nothing but representing multi-level object saliency in images with multiple objects. However, this multi-level object saliency was used for a purpose other than evaluating salient object detection performance. For salient object detection evaluation, the ground truth images proposed to be used were generated as binary maps obtained by thresholding the multi-level object saliency values. An example image and it’s ground truth are the ones given in Figure \[fig:datasetComparisons\]. The available multi-level object saliency has been used in [@islamsal18] to augment the ground truth of PASCAL-S dataset with information about the relative ranks of the salient objects in an image. We refer to this augmented dataset as AugPASCAL-S dataset, an example of whose ground truth is shown in Figure \[fig:datasetComparisons\]. Pointing out that most datasets assume that images contain at least one salient object and have images that usually contain objects in low clutter, the SOC dataset was generated addressing the said issues. This dataset contains 6000 images which are categorized into 80 classes, where real-world conditions such as motion blur, occlusion, cluttered background and image with no salient object are present. 5 viewers were asked to mark salient objects using bounding boxes and annotate them. Only objects annotated by a majority of the viewers were considered salient, which yielded 3000 images with salient objects. Volunteers were then involved to mark precise boundaries (through accurate silhouetting) of the salient objects and validate the annotations. As evident, this dataset contains images with multiple salient objects which are individually annotated, and as shown in Figure \[fig:datasetComparisons\] the ground truth images provided for salient object detection evaluation are binary maps segregating all the salient objects from those which are not and the image background. Note that for images with no salient object, the ground truth is all zeros. [The DUT-OMRON dataset consists of 5172 natural images that are taken from the SUN dataset [@SUND] through random selection followed by criteria-based pruning with one of the criteria being the presence of an apparent foreground. So the selected natural images contain at least one foreground object, with many of them having multiple. At least five operators are asked to mark bounding boxes on the salient objects in an image, and the final bounding box for a salient object is taken as the binary bounding mask obtained with a threshold of 0.5 on the average of the operators’ binary bounding masks. An eye tracker was also employed and many operators’ eye fixation data were collected on the images by showing them the images for 2 seconds each. Processing to remove outliers were carried out following which around 150 fixations were available per image on an average. A saliency map, which is a gray map taking values from $[0, 1]$, was generated from the fixations in an image applying Gaussian masks. Finally, a threshold of 0.1 was used on the saliency map to get a binary map. It was suggested that salient object detection may be evaluated on the two binary maps separately using precision-recall and receiver-operator characteristics curves.]{} Representation of Multi-level Object Saliency {#sec:limitations} --------------------------------------------- A discussion on the representation of multi-level object saliency in the seven existing datasets is given here. The binary ground truths generated in the MSRA-1000 and SED-100 datasets marked a single prominent object as salient in the images. When a single object is considered salient in an image, it helps us understand the visual cues of objects that affect their saliency. However, the notion of single object salient oversimplifies what salient object detection should be evaluated for, especially as natural images can be complex and can include multiple objects. Obviously, in natural images with multiple objects, the MSRA-1000 and SED-100 datasets do not attempt to represent multi-level object saliency. During the creation of SOD dataset, salient object ranking was performed in images with multiple objects in order to represent multi-level object saliency. For evaluation, ground truth that represents multi-level object saliency is required. However, it is not trivial to generate such a ground truth from the salient object ranking. State-of-the-art salient object detection evaluation using SOD data considers only binary ground truth without involving multi-level object saliency. Salient object ranking does not contain information which can be used to quantify the difference between saliencies of any two objects, which makes the ground truth generation non-trivial. Moreover, there is no single appropriate way of combining the rankings given by the seven subjects. [Although the SOC dataset is a very useful one with a substantially large number of images where real-world scenarios are carefully captured and different salient objects in an image are annotated, the ground truth available for salient object detection evaluation is binary. This binarization is done while considering objects marked by a majority as salient, with an object’s bounding box determined using a threshold on the IoU of the viewers’ bounding boxes. Even though the information about the different categories of objects in an image is available through the annotation, mapping that information to multi-level saliency is not straightforward.]{} The PASCAL-S dataset assigned a different saliency value to each object in an image in order to represent multi-level object saliency in images with multiple objects. With all objects in an image already manually segmented, this representation of multi-level object saliency can be used as ground truth for evaluating salient object detection, which unfortunately was not considered. Although the multi-level object saliency representation is readily available to be used as ground truth, certain aspects of it need to be noted. The only task performed by the volunteers to generate the representation was to click on salient objects. Such an approach to mark salient objects follows after the human perception of information at fixations. It is well known that saliency detection by humans involves both bottom-up and top-down processes [@Baluch11]. As human perception involves cognitive entities such as memory, knowledge, and preference, the top-down process would dominate in the clicking on salient objects. Here, the lack of the bottom-up process’s role diminishes the direct influence of the visual cues /features of objects in their saliency. The multi-level object saliency representation to be used as ground truth for salient object detection evaluation must relate to both bottom-up and top-down processes, as the interaction between the processes is substantially unknown [@McMains11]. The bottom-up process is more objective in nature while the top-down process is more subjective, and hence, they represent different aspects of saliency detection by humans. In addition, as most state-of-the-art salient object detectors are heavily based on processing object’s visual cues, consideration of bottom-up process will evaluate the approaches on what they aim to do and the inclusion of top-down process will help in evaluating what an approach must look to achieve further. Moreover, only 12 persons were used to create the PASCAL-S dataset and allied multi-level object saliency representation, which is substantially less number of samples to capture sufficient variations in subjective top-down influences. [The above discussion is also applicable for the AugPASCAL-S dataset, where saliency ranks are also part of the ground truth along with binary salient object maps, but not multi-level object saliency.]{} Referring back to the salient object ranking in SOD dataset, it must be noted that the ranking was also performed there in a manner where top-down process dominates and only 7 persons were involved. [The DUT-OMRON dataset with a large number of images, is the only existing dataset which considers bounding box markings and eye fixations to generate ground truth for evaluating salient object detection. Bounding box marking and eye fixation can respectively capture top-down and bottom-up processes involved in saliency detection by humans. Hence, DUT-OMRON dataset is probably better suited for evaluation of salient object detection at different perceptual levels of humans than PASCAL-S dataset. However, similar to the PASCAL-S dataset, although inherently multi-level object saliency data were collected, they were binarized to yield binary ground truth maps for evaluating salient object detection, which is provided in the dataset. Further, although there are two different ground truth binary maps available related to each image obtained from the two subjective human experiments, no guidelines have been provided on performing a single evaluation using them. In [@Yang_saliency_detection], only the binary ground truth corresponding to bounding box marking is considered for salient object detection evaluation.]{} In the following section, we introduce a new dataset and discuss three different subjective experiments performed using the dataset. We also justify that the data collected through the three experiments can be used together to appropriately represent multi-level object saliency ground truth. Subjective Experiments and Data Collection {#sec:dataCollection} ========================================== In order to form our dataset, we examine a subset of the ImageNet [@Deng_imagenet] images, in which, object bounding boxes are provided. Among those examined, we select 588 natural images that have multiple objects (total 2434 objects in 588 images). Our dataset is judiciously built ensuring sufficient variations among images in terms of the number of objects, the object sizes and positions, the object color content and the color contrast between the object and the rest in the image (see Appendix 1 (Section \[appen\]) for further explanations). We retrieved the original high resolution versions of the images for our dataset from Flickr (http://www.flickr.com) and the larger dimension of all of the images is [kept at 1024 pixels]{}. Our dataset and codes are publicly available at **https://github.com/gokyildirim/salmon\_dataset**. Compared to the datasets described in Section \[sec:relatedWork\], our dataset is unique in the following aspects and hence more suitable: Each image contains multiple salient objects and the image sizes are large enough to properly accommodate objects of substantially different sizes. The saliency of objects is recorded from a sufficiently large number of subjects using multiple complementary subjective experiments. These factors help us appropriately investigate the saliency levels of objects with different features and sizes in the same image. As we discussed in Section \[sec:introduction\], the saliency level of an object can be studied from visual attention (spontaneous) and perception. For this study, we conducted three subjective experiments: eye-tracking, point-clicking, and rectangle-drawing. The tasks in the subjective experiments [record]{} spontaneous attention and different kinds of human perception. The eye-tracking experiments are not associated with any goal and conducted in a free-viewing task, where only spontaneous attention is involved. Point-clicking involves a clicking task, which requires positional awareness of a salient object imparted through low-level perception [@Shenton; @Chalmers92]. The rectangle-drawing experiments involve concepts of salient object shape and size/scale, where it is necessary to tightly fit a rectangle, and require higher-level perception [@Shenton; @Chalmers92]. These subjective experiments measure object saliency at various perceptual levels of humans that can be arranged in a hierarchy as illustrated in Table \[tab:hierarchy\]. Tasks in the three subjective experiments achieved through spontaneous attention and levels of perception are shown in the table indicating the progression from bottom-up to top-down process domination. [|c|c|c|c|c|]{} &\ & --------- Free Viewing --------- : Hierarchy of our subjective experiments & ---------- Object Location ---------- : Hierarchy of our subjective experiments & -------- Object Scale -------- : Hierarchy of our subjective experiments \ Eye Tracking & & &\ Point Clicking & & &\ Rectangle Drawing & & &\ \[tab:hierarchy\] [In Figures \[fig:multimodalGroundTruth\], , and , we visualize the object-level saliency analogs of eye-fixation, point-clicking and rectangle-drawing data. These maps represent object saliencies in the image at different levels of perception. Although object saliencies at different perceptual levels can vary, in general, one of the maps can not be recognized as more appropriate than the others.]{} [Instead, they reflect the hierarchical differences in perception (see Table \[tab:hierarchy\]) and represent these differences at the object level. This allows us to evaluate and verify the features and assumptions of existing salient object detection methods, which in turn should provide guidelines on how to improve them and on what type of applications we can use them.]{} In order to precisely measure object saliency, we manually segment all objects in all the images, and then study object saliency using the experimental data collected. Object Segmentation {#obsegmask} ------------------- As we investigate object saliency, we need to know which pixels belong to which object or background. Objects in images are separated from the background by segmenting them with pixel-precise outlines. The segmentation mask $\mathbf{M}^o$ of an object $o$ is illustrated in Figure \[fig:averageColorCalculation\]. We use this mask to measure object saliency. \ Eye-Tracking Experiments {#sec:eyeTrackingExperiments} ------------------------ In order to measure the human visual saliency of objects due to their distinct features and information content that attract spontaneous attention, we perform eye-tracking experiments. We used RED250[^1] infrared eye-tracking device with a sampling frequency of 250 Hz and 9-point gaze calibration was considered. In total, we collected eye-fixation data from 95 people (48% women, 52% men) within ages 18-34. Each person in the experiment was asked to freely view randomly chosen 200 images (two experiments with 100 images shown in each) on an LCD monitor screen with a resolution of $1680\times 1050$ pixels. Each image was shown to the subjects for five seconds and an empty gray screen was shown for two seconds between two images to destroy persistence. Each image was viewed by 24 subjects on an average (min: 13, max: 34). For an image, a fixation by a person is recorded as a point at a single pixel location and a fixation map of the image size is formed with each pixel location containing a non-negative integer denoting the number of fixations by the person on it. We convert the fixation points of a person $\delta$ on an image $k$ into a fixation-density map $\mathbf{S}_{et}^{\delta,k}$ using a circular Gaussian filter on the fixation map. The $\sigma$ parameter of this filter is calculated using (\[eq:eyeTrackingSigma\]), which is equal to the radius of the circle of foveal vision as shown in Figure \[fig:eyeTrackingSetup\]. $$\label{eq:eyeTrackingSigma} \sigma = d_v \cdot \frac{r_v}{h_m} \cdot \big( \tan(\alpha + \eta + \theta) - \tan(\theta) \big)$$ ![The configuration of the subject and the monitor during the eye-tracking experiments.[]{data-label="fig:eyeTrackingSetup"}](figures/eyeTrackingSetup.pdf){width="9cm"} Here, $\alpha = 1\degree$ is the half size of the human fovea, $\eta = 0.4\degree$ is the accuracy of the eye tracker, $d_v = 75\text{ cm}$ is the viewing distance, $r_v = 1050 \text{ pixels}$ is the vertical resolution of the monitor, and $h_m = 29.5 \text{ cm}$ is the height of the monitor. In our experiments $\sigma \approx 66$ pixels. After the Gaussian filtering, the fixation-density maps $\mathbf{S}_{et}^{\delta,k}$ are normalized between 0 and 1. Fixation-density maps indicate the visual saliency level of the object in an image as measured by recording gaze. We use these maps to measure the eye-tracking saliency", namely $s_{et}^o$, of an object $o$ in an image. It is possible that different people fixated at different distinctive parts of the same object which they deemed salient. Considering the fact that objects with a single distinctive part are more likely to be fixated upon [@Yarbus_eye_movements], to measure $s_{et}^o$, we use the maximum fixation-density values within the boundary of object $o$ in image $k$ as follows: $$s_{et}^o = \frac{\sum_\delta \displaystyle \max_{i \in \mathbf{M}^o} \big(\mathbf{S}_{et}^{\delta,k}(i)\big)}{N_{et}^k}$$ Here $N_{et}^k$ is the number of people who viewed the image $k$ during the eye-tracking experiments and $i$ is a pixel inside the object boundary ($i \in \mathbf{M}^o$). An illustration of this operation is given in Figure \[fig:eyeTrackingExperiments\]. [We convert the eye-fixation density [of a person $\delta$ on an image $k$ ]{}into a saliency value for a whole object $o$ via $\max_{i \in \mathbf{M}^o} \big(\mathbf{S}_{et}^{\delta,k}(i)\big)$. We do so considering the fact that in an image where objects lay in front of the background, a person fixating on a part of an object indicates interest on the entire object as a single entity. This enables us to evaluate salient object detection methods against the saliency provided by eye-tracking ground truth, as these methods focus on pixel-wise segmentation of the salient objects, rather than eye-fixation prediction.]{} Point-Clicking Experiments {#sec:pointClickingExperiments} -------------------------- In order to measure human visual saliency of objects at various perceptual levels, we perform a couple of experiments involving subjective tasks. First, we consider the measurement of saliency at low-level perception. Correlating low-level perception to the so-called ‘dorsal’ or ‘where’ stream of the two-streams hypothesis related to neural processing of visual signals [@Kandel], we consider point-clicking experiments, which only require positional awareness of salient objects. In order to perform the point-clicking experiments, we used a crowd-sourcing web site[^2]. We asked people to click on the “important objects” that they “notice at first glance”. The “first glance” phrase was used to avoid multiple perceptual processing of a single image entity by a subject to the extent possible. Such avoidance of multiple time perceptual processing makes this experiment akin to the eye tracking one where multiple processing is avoided due to the phenomenon of inhibition of return [@Itti_a_model]. The task duration was limited to 30 minutes, and in each task, randomly selected 42 images were shown simultaneously. Therefore, approximately 45 seconds were available to the subjects per image, which is sufficient as far as seeing the image and clicking is concerned. In this experiment, each image is viewed by 33 people on an average (min: 24, max: 38). We represent the set of points (pixels $i$) where a person $\delta$ clicked on image $k$ as $\mathbf{S}_{pc}^{\delta,k}$. In order to measure the point-clicking saliency", namely $s_{pc}^o$, of an object $o$ in image $k$, we count the number of people who clicked an object and normalized it with the number of subjects who viewed the image. [This is very similar to the procedure used in PASCAL-S [@Sal_secret], where an object is considered as salient by a person when they click within the boundary of that object.]{} Formally, it can be calculated as follows: $$\begin{aligned} s_{pc}^o &= \frac{\sum_\delta f(\mathbf{M}^o, \mathbf{S}_{pc}^{\delta,k})}{N_{pc}^k}\\ f(\mathbf{M}^o, \mathbf{S}_{pc}^{\delta,k}) &= \begin{cases} \exists i \in \mathbf{S}_{pc}^{\delta,k} \colon i \in \mathbf{M}^o\text{,} & 1 \\ \text{else,} & 0\end{cases} \end{aligned}$$ Here $N_{pc}^k$ is the number of people who viewed the image $k$ during point-clicking experiments. An illustration of this operation is given in Figure \[fig:pointClickingExperiments\]. Rectangle-Drawing Experiments {#sec:rectangleDrawingExperiments} ----------------------------- We consider a second experiment to measure of saliency at higher-level perception. Correlating higher-level perception to the so-called ‘ventral’ or ‘what’ stream of the two-streams hypothesis, we consider rectangle-drawing experiments, which involve concepts of shape and scale. Similar to the point clicking experiments, rectangle-drawing experiments are performed using crowd sourcing. We asked people to draw a tight rectangle around the “important objects” that they “notice at first glance”. Again, the task duration was limited to 30 minutes and in each task 42 images were shown. In this experiment, each image is viewed by 32 people on an average (min: 15, max: 50). We represent the set of rectangles which a person $\delta$ drew on image $k$ as $\mathbf{S}_{rd}^{\delta,k}$. In order to measure the rectangle-drawing saliency", namely $s_{rd}^o$, of an object $o$ in image $k$, we count the number of people, who drew a rectangle on an object with an intersection-over-union [(IoU)]{} score greater than 0.3 and normalize it using the number of subjects who viewed the image. [In object detection literature, IoU is used to verify whether an estimated rectangle is a correct detection [@Everingham_pascal_visual]. Here, analogously, an object is considered as salient by a person, when they draw a rectangle with IoU $ \ge 0.3$ with respect to the tight rectangle around a segmented object (see Figure \[fig:rectangleDrawingExperiments\]).]{} We can calculate this value as follows: $$\begin{aligned} s_{rd}^o &= \frac{\sum_\delta f(\mathbf{M}^o, \mathbf{S}_{rd}^{\delta,k})}{N_{rd}^k}\\ f(\mathbf{M}^o, \mathbf{S}_{rd}^{\delta,k}) &= \begin{cases} \exists r \in \mathbf{S}_{rd}^{\delta,k} \colon g(r,r^o) \geq 0.3\text{,} & 1 \\ \text{else,} & 0\end{cases} \end{aligned}$$ Here $N_{rd}^k$ is the number of people who viewed the image $k$ during rectangle-drawing experiments, $r$ is a rectangle in $\mathbf{S}_{rd}^{\delta,k}$, $g(.,.)$ is the function that computes the intersection-over-union score, and $r^o$ is the reference rectangle, which tightly encloses the object we segmented before. An illustration of this operation is given in Figure \[fig:rectangleDrawingExperiments\]. It is evident from all the three measures of object saliency, that they can be any value between 0 and 1. Certain objects which are considered salient by some, may not be considered salient by others, and this subjectivity is addressed by considering the collective response of a large number of persons. Following this, it becomes clear from the multiple values of object saliency that objects in natural images are seen and perceived to have varying levels of importance. Hence, while evaluating a salient object detection performance quantitatively, it would be imperative for a human to consider the different levels of object importance. Therefore, for an image with multiple objects, a salient object detection algorithm must be evaluated considering the varying level of importance of the objects, making the automatic objective evaluation more meaningful (human-like). Multi-Level Object Saliency {#sec:visualSaliencyOfObjects} =========================== Let us consider the varying levels of importance of objects, which makes object saliency inherently multi-level, further here. Multi-level object saliency implies that saliency of objects can take any value between 0 and 1, rather than having two distinct (i.e. binary) values. In Figures \[fig:objectSaliency\](a), (c), and (e), we illustrate the measured saliency values of the 2434 objects in our dataset in the form of distributions. We can clearly see that the objects are not equally salient, i.e. object saliency is multi-level, as far as collective eye-fixation durations and human perception are concerned. The measured saliency values of an object are different for three subjective experiments. In Figures \[fig:objectSaliency\](b), (d), and (f), we show the relationship between eye-tracking, point-clicking, and rectangle-drawing saliency values of the objects in our dataset in pairs. For each pair, we fit a gamma non-linearity that maps the values from the x-axis to the y-axis. The correlation coefficients (squared) or coefficient of determination for these mappings are given below the plots in Figure \[fig:objectSaliency\]. We see that the hierarchy of the subjective experiments given in Table \[tab:hierarchy\] is reflected by the coefficients. The eye tracking and rectangle drawing saliencies are more dissimilar from each other than they are from the point clicking saliency. This observation is related to the levels of human perception involved in the experiments, with eye tracking and rectangle drawing being the two opposite ends. \ \ On a different note, depending on the desired kind of object saliency, eye-tracking, point-clicking, and rectangle-drawing saliency values can be employed in different applications. For example, we can use the eye-tracking data as an overall indicator for human spontaneous attention in psychophysical studies [@Engelke_comparative_study; @Toet_computational_versus]. In point-clicking experiments, a subject is required to voluntarily move the mouse and to perform the click. So the point-click data does not comprise of components present in eye-tracking data that is due to involuntary eye-movements, experiment fatigue, etc., which are considered unwanted in certain cases. In addition, clicked-points can be very useful for resolving the ambiguity in eye-fixation density maps when two or more objects are spatially very close to each other. Therefore, compared to eye-tracking, point-clicking experiments provide more certain object-level saliency and can be used in areas such as web site design [@Ivory_improving_web]. In rectangle-drawing experiments, subjects needed to fit a rectangle around the objects of interest, which further specifies saliency compared to point-clicking experiment. For example, when a subject clicks on a human face in an image, it is not obvious whether the face or the whole body of that human is considered as salient. In rectangle-drawing experiments, however, a drawn rectangle would remove such uncertainty, which is valuable in objectness measurement [@Alexe_measuring_the]. In Section \[sec:introduction\], we referred to literature which suggests that humans fixate on (spontaneously attend to) objects that are distinct and more informative for a longer period of time compared to other parts of an image. In [@Alers_how_the], Alers et al. compare the eye-fixation trends of subjects under two conditions: free-viewing and task-driven. They observe that subjects have a tendency to fixate for more time on the distinct informative objects, when they are not given a certain task (free-viewing condition). As in our free-viewing eye-tracking experiments we take eye-fixation duration as a measure of saliency (importance), we implicitly consider saliency to be correlated to distinctness and informativeness. This is unlike the other two experiments, where the subjects are given explicit instructions to notice important objects. However, as it is well known that human vision performs image understanding by considering only pertinent visual data, the process of spontaneous attention and subsequent fixation duration obviously works to choose the pertinent data, and hence, our implicit consideration is well-founded. As we discussed in Sections \[sec:introduction\] and \[sec:dataCollection\], object saliency is subjective. Some objects which are considered salient by a few, may not be considered salient by others. When we measure object saliency, we take this subjectivity of saliency into account by considering collective human opinion /response. For example, in our point-clicking experiments, we asked a group of people to click on the important objects. The subjects of this experiment followed various strategies, such as clicking on all the objects, only on the most prominent object, or on the objects that are closer to the camera. We measure the collective object saliency by considering the ratio of the number of people who clicked on an object in an image to the number of people who viewed the image. We use both point-clicking and rectangle-drawing saliency values as measures based on collective human perception. Note that, the standard deviation of the collective opinion is correlated to the saliency measured. When the majority of the subjects agrees on the significance (or insignificance) of an object in an image, standard deviation of collective opinion is low and the saliency value of that object approaches to 1 (or 0). Whereas, a saliency value near 0.5 implies strong disagreement on object saliency with the subjects being equally divided into two groups and the standard deviation of collective opinion is high. Even with such careful and conceptually clear measurement of object saliency, there are some, mostly trivial issues that we overlooked in preference to simplicity. However, one important issue that we encountered is the order in which a subject looked at, clicked on, or drew on an object, which we found could be a factor influencing object saliency. For example, during point-clicking experiments, most subjects usually clicked in a horizontal order from left to right. This order of attention introduced a small but undesired bias in saliency measurements making the objects on the left slightly more salient than the objects on the right when no other factor plays a role. [In the future, this bias can possibly be alleviated by applying a random horizontal flip to the displayed images, so that the left to right preference is averaged out.]{} Performance Evaluation of Existing Salient Object Detection Approaches {#sec:salientObjectDetectors} ====================================================================== As we have seen till now that image objects are inherently perceived to possess multi-level saliency, salient object detection techniques must be evaluated considering multi-level object saliency ground truth, and hence, we present here such an evaluation of the state-of-the-art. Salient object detectors generate object saliency maps, which are gray-level maps, from which binary object saliency maps are subsequently generated during their standard evaluation against binary ground truth using precision-recall curves. Obviously, if multi-level object saliency ground truth is available, the gray-level object saliency maps generated during salient object detection could be directly used against them for evaluation. We generate multi-level object saliency ground truth corresponding to all the images in our dataset discussed earlier in Section \[sec:dataCollection\] using the results of the subjective experiments, and use them here for evaluating a few existing salient object detection approaches. We get the gray-level object saliency maps generated by the following salient object detectors using their publicly available codes: IT [@Itti_a_model], FT [@Achanta_frequency_tuned], FASA [@Yildirim_fast_accurate], HC [@Cheng_global_contrast], RC [@Cheng_global_contrast], CB [@Jiang_automatic_salient], GC [@Cheng_efficient_salient], SF [@Perazzi_saliency_filters], LR [@Shen_a_unified], CH [@Li_contextual_hypergraph], AMC [@Jiang_saliency_detection], GMR [@Yang_saliency_detection], HR [@Yildirim_saliency_detection], MB [@Zhang2015], IILP [@Li2015], SMD [@Peng2017], and MIL [@Huang2017]. We use these gray-level saliency maps to compare the salient object detectors using our dataset and ground truth. Note that, among the above detectors, the ones involving machine learning have been trained by the authors using binary ground-truth maps. The widely-used performance metric, precision-recall curve is applicable with the use of binary ground-truth maps. Multi-level ground-truth maps, however, consist of multiple values. Therefore, we evaluate and compare the performances of the existing methods on the following three different aspects: We use a standard regression loss, mean absolute error, to evaluate the accuracy in estimation of multiple saliency values. We consider a consistent utilization of the area under precision-recall curve measure to evaluate classification performance at multiple saliency levels. We use a standard ranking correlation measure, Kendall rank correlation coefficient, to evaluate the accuracy in ranking objects based on the estimated multi-level saliency. Forming Multi-Level Ground-Truth Maps ------------------------------------- We generate multi-level ground-truth maps from the eye-tracking ($s_{et}^o$), point-clicking ($s_{pc}^o$), and rectangle-drawing saliency ($s_{rd}^o$) values as follows: $$\begin{aligned} \mathbf{F}_{\gamma}^k = \sum_{\forall o \text{ in image } k} s_\gamma^o \cdot \mathbf{m}^o \end{aligned}$$ where $$\begin{aligned} \mathbf{m}^o(i,j)&=& 1, \forall (i,j)\in \mathbf{M}^o\\ \nonumber &=& 0, otherwise\end{aligned}$$ with $(i,j)$ representing a pixel in image $k$. Here, $\mathbf{F}_{\gamma}^k$ is the multi-level ground-truth map of the image $k$ using $\gamma$ saliency values, where $\gamma \in \{ et, pc, rd\}$ is one of the subjective experiments. This gives us three different ground-truth maps for each image in the dataset, which we will use in our analysis. Note that all the three ground truth maps take values from the interval $[0,1]$. Mean Absolute Error {#MAEss} ------------------- Mean absolute error (MAE) measures the absolute difference between estimated and measured (ground truth) object-saliency values.[We employ mean absolute error as it is a well-known regression loss, and computation of saliency values can be directly interpreted as an estimation of regressands based on the input data acting as the regressors.]{} We calculate the MAE with respect to our dataset as follows: $$\begin{aligned} \label{myeq2} \widehat{\text{MAE}}_{\gamma} &=& \frac{1}{N} \sum_{\forall o} \text{MAE}_{\gamma}^o\\ \nonumber \text{MAE}_{\gamma}^o &=& \frac{1}{|\mathbf{M}^o|} \sum_{(i,j) \in \mathbf{M}^o} \big|\mathrm{S}^o - \mathbf{F}_{\gamma}^k(i,j)\big|,\ \mathbf{M}^o \text{ in image } k \\ \label{myeq1} &=& \big|\mathrm{S}^o - s_\gamma^o\big|\\ \label{myeq0} \mathrm{S}^o &=& \frac{1}{|\mathbf{M}^o|} \sum_{(i,j) \in \mathbf{M}^o} \mathbf{S}(i,j) $$ Here, $\widehat{\text{MAE}}_{\gamma}$ is the average error over all objects, $N = 2434$ is the total number of objects in our dataset and $\text{MAE}_{\gamma}^o$ is the error on an object $o$ for saliency type $\gamma \in \{et,pc,rd\}$. In (\[myeq0\]), $\mathbf{S}$ is the estimated gray-level object saliency map (with value $\in [0,1]$) by an algorithm, and in (\[myeq1\]), $\mathrm{S}^o$ is considered as the estimated saliency value of the object $o$. Note that, $(i,j)$ represents a pixel and $\forall (i,j) \in \mathbf{M}^o$ (object mask), $\mathbf{F}_{\gamma}^k(i,j) = s_\gamma^o$. The three average errors that we get in (\[myeq2\]) for the three saliency types can be used to analyze the performance of an algorithm in terms of how it corresponds to object saliencies at different levels of human perception, and hence, are extremely useful. However, we require a single average error using the three saliency types together, which can be used to directly evaluate an algorithm in comparison to another. We define the error $\text{MAE}^o$ on an object $o$ when all the three ground truth saliencies are considered together as follows: $$\label{myeq3} \text{MAE}^o = \min_{\gamma \in \{ et, pc, rd\}} \text{MAE}_{\gamma}^o$$ and this $\text{MAE}^o$ is then used in (\[myeq2\]) instead of $\text{MAE}_{\gamma}^o$ to get the corresponding average error $\widehat{\text{MAE}}$ over all objects. As can be seen in (\[myeq3\]), for an object, the minimum of the three errors is considered. This signifies that we do not penalize an approach for not sufficiently conforming with two of the three ground truth saliencies when it does with the third, considering each object independently. In effect, we suggest that a salient object detection approach performs well when it estimates accurately the object saliency at any single human perceptual level, which allows us to fairly compare different algorithms that may contain modules inspired by different phenomena (at different levels) in human vision. Instead of the average error over all objects in the dataset, we can also obtain the average error for an image by considering only the objects in it while using (\[myeq2\]) and (\[myeq1\]), with the number of objects in the image (say $N_I$) in place of $N$. [We perform object-wise MAE computation and not pixel-wise, as we intend to evaluate the single saliency level of an object predicted by an algorithm. In object saliency estimation, an object irrespective of its size is a single entity. So, pixel-wise MAE will be strongly biased towards large objects, and will particularly neglect grave errors in estimating the saliency levels of small but important objects.]{} In Figures \[fig:multiLevelComparison\](a), (b), and (c), we illustrate the $\widehat{\text{MAE}}_{\gamma}$ values of the mentioned existing salient-object detectors and in Figure \[fig:multiLevelComparison0\](b), we show their $\widehat{\text{MAE}}$ values. Average area under Precision-Recall Curve {#AUCss} ----------------------------------------- Evaluation of salient object detection using binary images as ground truth is predominantly done based on precision-recall curves [@Han2018], where the approaches are tested in classification performance. Here, we propose to use area under precision-recall curve (AuPRC) to evaluate classification performance at multiple saliency levels given by our multi-level ground truth images. [In order to compute AuPRC, we binarize a multi-level ground truth image (of type $\gamma$) into multiple binary images by using the saliency levels of the objects in that image as thresholds. An example of this binarization is illustrated in Figure \[fig:aucVisualization\]]{}. \ [Let $A_\gamma^{(k,l)}$ be the AuPRC that is calculated using the binary map of type $\gamma$ for image $k$ at saliency level $l$. Then we compute: $$\begin{aligned} \label{auc2} \widehat{A}_\gamma &=& \frac{1}{N} \sum_{o} A_\gamma^{o},\ \ o=(k,l)\end{aligned}$$]{} where, $\widehat{A}_\gamma$ gives the average AuPRC over all objects for ground truth saliency type $\gamma \in \{ et, pc, rd\}$. In a manner similar to the case of MAE computation, we compute a single average AuPRC value $\widehat{A}$ considering all the three ground truth saliency types together as follows: $$\label{auc3} \widehat{A} = \frac{1}{N} \sum_{\forall o}\max_{\gamma \in \{ et, pc, rd\}}A_\gamma^o$$ where, we do not penalize an approach for not sufficiently conforming with two of the three ground truths when it does with the third, whose appropriateness has been discussed in Section \[MAEss\]. In Figures \[fig:multiLevelComparison\](a), (b), and (c), the $\widehat{A}_\gamma$ values of the existing salient object detectors are given and in Figure \[fig:multiLevelComparison0\](a), the $\widehat{A}$ values are shown. Just like case of MAE computation, one can check performance on an image by computing average AuPRC for all the objects only in the image using (\[auc2\])-(\[auc3\]), accordingly. In the above explanation of AuPRC usage for multi-level ground truth, it should be noted that the consideration of thresholds as shown in [Figure \[fig:aucVisualization\]]{} after ordering is [analogous]{} with the thresholding performed in the standard precision-recall curve computation. Salient Object Ranking Correlation ---------------------------------- Salient objects detected by an approach are usually used in further analysis performing image understanding and recognition. In many such cases, the highest to lowest ranking of the salient objects are more important than the absolute saliency values, as preferences in further analysis are given in that order. Therefore, we consider evaluation based on relative importance of salient objects in our dataset by ranking them with respect to their saliencies from highest to lowest. For the evaluation, we consider the correlation between the salient object ranking based on the gray-level saliency values provided by an algorithm and that based on the ground truth gray-level saliencies of type $\gamma$ by computing Kendall rank-correlation coefficient (Kendall’s $\tau_b$) [@Kendall] between them. In Figures \[fig:multiLevelComparison\](a), (b), and (c), the Kendall’s tau [($\tau^\gamma_b$)]{} values for the existing methods considering the three different ground truths are shown. Now, similar to [Sections \[MAEss\] and \[AUCss\]]{}, we require a single correlation value using the three saliency ground truths together, which can be used to directly evaluate an algorithm in comparison to another. For this, we introduce a slight modification in Kendall’s tau computation to get a correlation measure which considers all the three ground truth saliencies together. [Similar to the MAE and AUC computation cases]{}, the modification of Kendall’s tau by considering all the three ground truths together must indicate that a salient object detection approach performs well when it estimates accurately the object saliency rank at any single human perceptual level. That is, an approach is not penalized for not sufficiently conforming with two of the three ground truth saliencies when it does with the third, whose appropriateness has already been pointed out in Section \[MAEss\]. Let the set of type-$\gamma$ ground truth gray-level saliencies of all objects in our dataset be represented by $\rho_{\gamma}$, that is, $s_\gamma^o\in\rho_{\gamma}, \forall o$. Further, let the set of gray-level saliencies of all the corresponding objects estimated by an algorithm be $R$. The gray-level object saliency estimated by an algorithm for an object $o$ is given in (\[myeq0\]), and therefore, $\mathrm{S}^o\in R, \forall o$. Kendall’s tau $\tau^\gamma_b$ for each type of ground truth is calculated between $R$ and $\rho_\gamma$ using the standard way of Kendall’s tau computation as given in [@Kendall]. On the other hand, Kendall’s tau ($\tau_b$) considering the three ground truth types together are obtained using a modified Kendall’s tau computation as follows: $$\begin{aligned} \label{kennew} \tau_b=(C-D)\Big{/}\sqrt{(C+D+T_R)(C+D+T_\rho)}\end{aligned}$$ where, with all estimated saliency pairs $(\mathrm{S}^{x},\mathrm{S}^{y})$ and all ground truth saliency pairs $(s_\gamma^x, s_\gamma^y)$ employed, the $C$ and $D$ are numbers of concordant and discordant pairs, respectively, and the $T_R$ and $T_\rho$ are numbers of tied pairs in estimated and ground truth saliencies, respectively. The values of the parameters in (\[kennew\]) are computed as given below: $$\begin{aligned} \label{kennew1} C &= \sum_{x,y} {G^s_{x>y}} \cdot {E_{x>y}} + {G^s_{x<y}} \cdot {E_{x<y}}\\ \label{kennew2} D &= \sum_{x,y} {G^s_{x<y}} \cdot {G^i_{x\ley}} \cdot {E_{x>y}}+ {G^s_{x>y}} \cdot {G^i_{x\gey}} \cdot {E_{x<y}}\\ \label{kennew3} T_\rho &= \sum_{x,y} {G^i_{x=y}} \cdot ({E_{x>y}} + {E_{x<y}})\\ \label{kennew4} T_R &= \sum_{x,y} {G^s_{x\ney}} \cdot {E_{x=y}}\end{aligned}$$ with $$\begin{aligned} \nonumber {G^s_{x\bullety}} &= \max_\gamma \Delta(s_\gamma^x \bullet s_\gamma^y)\\ \nonumber {G^i_{x\bullety}} &= \min_\gamma \Delta(s_\gamma^x \bullet s_\gamma^y)\\ \nonumber {E_{x\bullety}} &= \Delta(S^x \bullet S^y)\end{aligned}$$ where, $\bullet \in \{<, >, \le, \ge, =, \ne\}$ is a generic operator and $\Delta(.)$ is an indicator function, which takes the value $1$ when the condition is satisfied and $0$ otherwise. In our modification of Kendall’s tau computation, to ensure that an approach is not penalized when it conforms with at least one of the three ground truths, the number of concordant pairs $C$ is incremented in (\[kennew1\]) when the concordance happens for at least one ground truth type. Further, the number of discordant pairs $D$ is incremented in (\[kennew2\]) only when there is no agreement with respect to all the three ground truths. The increments in $T_\rho$ and $T_R$ in (\[kennew3\]) and (\[kennew4\]) are deduced logically from the increment strategies of $C$ and $D$. Figure \[fig:multiLevelComparison0\](c) shows the modified Kendall’s tau ($\tau_b$) measure values for the existing methods considering the three ground truths together. Instead of computing the correlations considering all objects in the dataset, one can obtain the correlations considering only the objects in an image to check the performance on it. \ \ \[1\][&gt;m[\#1]{}]{} ![The multi-level ground truth saliency maps and the saliency maps that are estimated by a few well-performing methods on a few images where their qualitative performance significantly vary. Here, ET, PC, and RD indicate the ground truth maps that are generated using the eye-tracking, point-clicking, and rectangle-drawing subjective data, respectively.[]{data-label="fig:imageComparison"}](figures/imageComparisonFinal.png){width="8.4cm"} Discussion ---------- In Figures \[fig:multiLevelComparison\](a), (b), (c) and Figure \[fig:multiLevelComparison0\](b), which shows the MAE comparisons, we find that some of the approaches such as RC, HR, AMC do better in most of the cases than the others. Although from the values, the differences seem little, they are significant as they are obtained considering 2434 objects in 588 images. As can be seen, there are a few cases where an approach that performs well with respect to a ground truth, does not do well for the others, and vice versa. For example, HR does not do well for the eye-tracking saliency ground truth unlike for the others, and GMR appears among the top 3 only for the eye-tracking saliency ground truth. This is expected, as intentionally or not, approaches may be designed such that they work similar to that at a narrow band of perception from the entire zero (low) to high spectrum. That is why, Figure \[fig:multiLevelComparison0\](a) showing the MAE results when the three ground truths are considered together is essential. Note that the MAE values of all the approaches in Figure \[fig:multiLevelComparison0\](b) are lower compared to their values in Figures \[fig:multiLevelComparison\](a), (b) and (c) as a consequence of being fair by removing the bias of the evaluation in terms of perceptual level. In Figures \[fig:multiLevelComparison\](a), (b), (c) and Figure \[fig:multiLevelComparison0\](c), which shows the Kendall’s tau comparisons, we again find that some of the approaches such as GMR, MB, SMD do better in most of the cases than the others. Consider the top 3 performing approaches for the rectangle-drawing saliency. We see that two of them, namely HR and AMC appear way below in performance with respect to the other ground truth saliencies. Considering the ground truths separately, we cannot infer whether their good performance for the rectangle-drawing saliency is significant or not. Now, in terms of performance evaluation based on all the ground truths together, HR and AMC does not appear in top 3. This makes us capable to infer that the concerned case of good performance in one of the ground truth saliencies is not significant. This precisely shows how removing the perceptual level bias in the evaluation is appropriate and useful. It can also be noted that the Kendall’s tau values of all the approaches in Figure \[fig:multiLevelComparison0\](c) are higher compared to their values in Figures \[fig:multiLevelComparison\](a), (b) and (c), an effect similar to what was seen with the MAE values. When we compare using MAE and Kendall’s tau in Figures \[fig:multiLevelComparison\] and \[fig:multiLevelComparison0\], we observe that the order of performance of the existing methods is significantly different. Mean absolute error and salient object ranking correlation measure salient object detection performance very differently. In Table \[tab:metricsDiffer\], two hypothetical cases for multi-level salient-object detection is presented. In each case, there are two salient objects with different ground truth (GT) and estimated saliency values. In the first case, the mean absolute error measures a small error, which indicates good performance. On the other hand, due to the change in the order of saliency values, the salient object ranking correlation evaluates an inverse correlation, that is, a bad performance. In the second case, observation of the evaluations is reversed. This example shows that we should use a measure depending on how we intend to apply the multi-level saliency values in a task. Given two salient objects, absolute saliency values not only measure which one is more salient, but also quantify the difference in saliency. This approach can be useful in applications such as content-aware image compression, because absolute saliency values are directly used as an indicator to compression ratio. Salient object ranking is useful when the ordering of salient objects is more important than their actual saliency values. For example, in image tagging, processing the most important object would yield more relevant image tags compared to a randomly selected object. In literature, a few recently proposed existing approaches which [we have considered]{} in our analysis have been shown to outperform other existing ones based on binary ground truth datasets, and hence, they represent the state-of-the-art. In fact, most approaches for salient object detection are targeted to perform well on such binary datasets when evaluated using precision and recall measures. Therefore, the recently developed approaches may not perform well on our multi-level ground truth dataset when evaluated using MAE and Kendall’ tau, and hence, it would be appropriate to also evaluate existing approaches on binary maps generated from our multi-level ground truth saliencies. The average AuPRC based comparisons on our multi-level ground truth dataset shown in Figures \[fig:multiLevelComparison\](a), (b), (c) and Figure \[fig:multiLevelComparison0\](a) consider performance on binary maps generated from the multi-level ground truth maps. We find that some of the existing approaches such as HR, MB and AMC do, in general, better than others. Similar to the cases of MAE and Kendall’s tau based comparisons, we see here as well that most top performing approaches tend to do well in a particular type of ground truth and not in all types. Hence, the average AuPRC values shown in Figure \[fig:multiLevelComparison0\](a) considering the three ground truths together while removing perceptual bias is useful for evaluation. Note that the average AuPRC values of all the approaches in Figure \[fig:multiLevelComparison0\](a) are higher compared to their values in Figures \[fig:multiLevelComparison\](a), (b), (c), a result of being fair by removing perceptual bias. Like the disagreement between MAE and Kendall’s tau demonstrated using Table \[tab:metricsDiffer\], average AuPRC can also differ from both the MAE and Kendall’s tau values, which would depend on the comparative estimated saliency values and sizes of the objects in the images. It is interesting to find that approaches such as HR [@Yildirim_saliency_detection], AMC [@Jiang_saliency_detection], MB [@Zhang2015] and GMR [@Yang_saliency_detection], which consider salient object detection in images as a local to global hierarchical processing of a graph, in general, perform well in our multi-level saliency dataset. However, a couple of the recently proposed existing approaches do not perform well on our ground truth dataset, [even when evaluated using AuPRC]{}, and hence, may not be suitable for images with objects having multi-level saliency. [In Figure \[fig:BGTComparison\], we provide the average AuPRC results that evaluate the approaches considered in Figure \[fig:multiLevelComparison0\](a) on the binary ground truth obtained for our dataset images by considering all objects in them to be equally salient. As can be seen, some of the approaches perform well on the binary ground truth in comparison to others, although they did not do so on our multi-level saliency ground truth, and vice-versa. Note that, as expected, the average AuPRC results are higher for the binary ground truth than the multi-level ground truth.]{} Now, as our three evaluation measures can be used to assess multi-level salient object detection performance, designing of approaches in future can focus on detecting multiple salient objects with varying saliencies, which can be followed with further processing to yield binary (salient/non-salient) salient object detection outputs, if required, and this would be akin to how humans perform “salient object detection”. Gray-level object saliency maps obtained using a few approaches, which have performed well in our dataset, on a few images are shown in Figure \[fig:imageComparison\] along with the corresponding multi-level ground truth saliency maps. Case \#1 Case \#2 ------------------------------ ---------- ---------- GT Saliency Value \#1 0.48 0.3 GT Saliency Value \#2 0.52 0.8 Estimated Saliency Value \#1 0.51 0.0 Estimated Saliency Value \#2 0.49 0.5 Mean Absolute Error 0.03 0.3 Salient-Object Ranking -1 1 : Two hypothetical cases, where two evaluation measures significantly differ[]{data-label="tab:metricsDiffer"} ![Average area under precision-recall curves of the existing methods obtained on binary ground truth of our dataset images containing all salient object in one salient class.[]{data-label="fig:BGTComparison"}](figures/BinGTRes){width="\imwidth"} Conclusion {#sec:conclusion} ========== In this paper, subjective experiments have been conducted on natural images to [confirm]{} that all objects are inherently not equally important and salient object detection should be evaluated considering this aspect. For this, we have introduced an image dataset with each image having multiple objects, and for each image there are three types of multi-level ground-truth maps. For generating the ground truths, the saliency level of an object has been measured by conducting three subjective experiments: eye tracking, point clicking, and rectangle drawing. While eye-tracking experiments capture the effect of spontaneous attention and eye-fixation duration on object saliency, point-clicking and rectangle-drawing experiments measure the role of human perception at different levels. From this, we have shown that object saliency is inherently multi-level. Further, performance evaluation measures are given to evaluate multi-level salient object detection, and existing saliency detection approaches, some of which represent the state-of-the-art, are compared. [Approaches that perform a local to global hierarchical processing of images considering them as graphs are found to perform well in our dataset.]{} The dataset and measures reported in this paper add a different dimension to the evaluation of salient object detection designed for application in object localization, generic target detection, visual media description, compression and segmentation. Appendices ========== Appendix 1. Variations in the Images and Objects of the New Dataset {#appen} ------------------------------------------------------------------- In Section III of the paper, we mentioned that our dataset is balanced ensuring sufficient variations among images in terms of the number of objects, the object sizes and positions, the object color content and the color-contrast between the object and surroundings (local and global). Here, in Figures \[ch3-numberOfObjects\], \[ch3-objectDistributions1\], \[ch3-objectDistributions2\] and \[ch3-averageColorCalculation\], we illustrate the said characteristics of our dataset in terms of several related parameters. Consider Figure \[ch3-numberOfObjects\]. Figure \[ch3-numberOfObjects\](a) gives the distribution of the images in our dataset in terms of the number of salient objects contained in them. It is evident that the images in our dataset have a substantially varied number of objects, which is central to the discussions in the paper. [Figure \[ch3-numberOfObjects\](b) gives the distribution of different salient object categories in our dataset, where the terms ‘outdoor’ and ‘indoor’ stands for other generic objects found outdoors and indoors, respectively.]{} Figures \[ch3-numberOfObjects\](c), (d) and (e) display some example images from our dataset with the salient objects marked. \ In Figure \[ch3-objectDistributions1\], we demonstrate the variedness of the objects in our dataset with respect to their color contents. Figure \[ch3-objectDistributions1\](a) gives the distribution of the objects in our dataset in terms of their color entropies. Obviously, color entropies quantify within-object color variety. To compute color entropy of an object $o$ the CIE L$^*$a$^*$b$^*$ values at pixels within its segmentation mask $\mathbf{M}^o$ are considered. Segmentation mask formation for an object is described in Section \[obsegmask\]. The distribution of the object color values in the 3 dimensional L$^*$a$^*$b$^*$ space is then used to compute its entropy. It is evident from Figure \[ch3-objectDistributions1\](a) that the objects in our dataset vary substantially in within-object color variety. The 2D plots in Figure \[ch3-objectDistributions1\](b-d) show the L$^*$, a$^*$ and b$^*$ values of the objects distributed with respect to each other. Here, an object is represented by a single color value, which is the average of all the color vectors at pixels within the object’s segmentation mask $\mathbf{M}^o$. So, an element in the 2D plots corresponds to an object and gives its average 2D color value. As can be seen from all the three figures, significant average color variation exists between the objects of the dataset. \ In Figure \[ch3-objectDistributions2\], we demonstrate the variedness of the objects in our dataset with respect to their positions in images, and their shapes and sizes in terms of various parameters. [The objects are also categorized into six classes, namely, animal, plant, human, vehicle, other outdoor and indoor objects, and category-wise variedness with respect to the parameters are also presented.]{} To understand the parameters, consider Figure \[param\]. Figure \[ch3-objectDistributions2\](a) gives the distribution of the objects in our dataset in terms of their Euclidean distances to the image centers normalized with respect to underlying images’ diagonal lengths. Figures \[ch3-objectDistributions2\](b) and (c) respectively show the distributions of the objects in our dataset in terms of their widths and heights, which are computed as shown in the example of Figure \[param\]. Note that, the widths and heights have been computed in terms of the number of pixels, and then normalized with respect to their maximum values in the dataset. Figure \[ch3-objectDistributions2\](d) shows the distribution of the objects in our dataset in terms of their aspect ratios, which are calculated as the ratio of width and height. Finally, Figure \[ch3-objectDistributions2\](e) shows the distribution of the objects in our dataset in terms their areas, which are the number of pixels in the irregular object regions, normalized by the maximum value in the dataset. It is evident from these figures that the objects in our dataset vary substantially in terms of size and image position. [It is also evident that the said observation is also valid if we consider the object category-wise distributions shown in Figure \[ch3-objectDistributions2\](a)-(e).]{} \ \ ![Depiction of computation of various parameters related to an object’s size and location in the image.[]{data-label="param"}](figures/Objparam){width="\imwidth"} In Figure \[ch3-averageColorCalculation\] the variedness in contrasts of the objects in our dataset from their local neighborhood and global background is demonstrated. For an object in an image within a region given by the object segmentation mask $\mathbf{M}^o$, regions of local neighborhood and global background in the image are defined. For example, the local neighborhood mask $\mathbf{M}_l^o$ and the global background mask $\mathbf{M}_g^o$ for the object mask $\mathbf{M}^o$ in Figure \[fig:averageColorCalculation\](b) corresponding to an object in the image of Figure \[fig:averageColorCalculation\](a) are shown in Figures \[ch3-averageColorCalculation\](a) and (b), respectively. The local and global contrasts are then calculated as the $\chi^2$-distances between the distributions of the L$^*$a$^*$b$^*$ color values at the pixels within the object segmentation mask, and the local neighborhood and the global background masks, respectively. The distribution of the objects in our dataset in terms of the said distance values are shown in Figures \[ch3-averageColorCalculation\] (c) and (d) with the former depicting the global contrast distribution and the latter local contrast distribution. As can be seen, significant variations in local and global contrasts exist between different objects in our dataset. \ [^1]: [http://www.smivision.com/en/gaze-and-eye-tracking-systems/products/red250-red-500.html]{} [^2]: http://www.shorttask.com

ArXiv

--- abstract: 'We prove that the Kobayashi distance near boundary of a pseudoconvex Reinhardt domain $D$ increases asymptotically at most like $-\log d_D+C$. Moreover, for boundary points from $\text{int}\overline D$ the growth does not exceed $\frac{1}{2}\log(-\log d_D)+C$. The lower estimate by $-\frac{1}{2}\log d_D+C$ is obtained under additional assumptions of $\mathcal C^1$-smoothness of a domain and a non-tangential convergence.' address: 'Instytut Matematyki, Wydzia³ Matematyki i Informatyki, Uniwersytet Jagielloñski, ul. Prof. St. £ojasiewicza 6, 30-348 Kraków, Poland' author: - Tomasz Warszawski title: Boundary behavior of the Kobayashi distance in pseudoconvex Reinhardt domains --- Introduction and results ======================== The problem of a boundary behavior of the Kobayashi (pseudo)distance in pseudoconvex Reinhardt domains is connected with studying their Kobayashi completeness. The qualitative condition for the $k$-completeness of a bounded domain $D$ is $$k_{D}(z_0,z)\to\infty\text{ as } z\to\partial D.$$ The main fact is that if a pseudoconvex Reinhardt domain $D$ is hyperbolic then it is $k$-complete. At first Pflug [@pfl] proved it for bounded complete domains. A second step was done by Fu for bounded domains in [@fu]. The general case was finally solved by Zwonek in [@zw]. Hence it is natural to ask about a quantitative behavior of the function $k_{D}(z_0,\cdotp)$. Forstnerič and Rosay estimated it from below on bounded strongly pseudoconvex domains. Namely, it was proved in [@fr] that $$k_{D}(z_{1},z_{2})\geq-\frac{1}{2}\log d_{D}(z_{1})-\frac{1}{2}\log d_{D}(z_{2})+C$$ for $z_{j}$ near two distinct points $\zeta_{j}\in\partial D$, $j=1,2$. In the same paper the authors showed the opposite estimate for $\mathcal{C}^{1+\varepsilon}$-smooth domains with $z_{1},z_{2}$ near $\zeta_{0}\in\partial D$. This estimate in the bounded case follows from the inequality for the Lempert function of bounded $\mathcal{C}^{1+\varepsilon}$-smooth domains obtained by Nikolov, Pflug and Thomas $$\widetilde{k}_D(z_1,z_2)\leq-\frac{1}{2}\log d_D(z_1)-\frac{1}{2}\log d_D(z_2)+C,\,z_1,z_2\in D$$ in [@npt]. It was also proved that the above estimate fails in the $\mathcal{C}^1$-smooth case. The other general version of an upper estimate, for $\mathcal{C}^{2}$-smooth domains, can be found in [@jp]. The case of bounded convex domains was investigated by Mercer in [@mer]. For such domains we have $$-\frac{1}{2}\log d_{D}(z)+C'\leq k_{D}(z_{0},z)\leq-\alpha\log d_{D}(z)+C$$ with $\alpha>\frac{1}{2}$ and $z$ close to $\zeta_{0}\in\partial D$ (the constant $\alpha$ can not be replaced with $\frac{1}{2}$). An example $$D_\beta:=\{(z,w)\in\mathbb{C}^{2}:|z|^{\beta}+|w|^{\beta}<1\},\,0<\beta<1$$ shows that the lower estimate by $-\alpha\log d_{D}(z)+C$, where $\alpha>0$ — a constant independent on a domain, is not true for complete pseudoconvex Reinhardt domains. Easy calculations lead to $$k_{D_\beta}((0,0),(z,0))\leq-\frac{\beta}{2}\log d_{D_\beta}(z,0)+C$$ if $0<z<1$ and $(z,0)$ tends to $(1,0)$. In the paper we prove the following theorems. \[1\] Let $D\subset\mathbb{C}^{n}$ be a pseudoconvex Reinhardt domain. Fix $z_{0}\in D$ and $\zeta_0\in\partial D$. Then for some constant $C$ the inequality $$k_{D}(z_{0},z)\leq-\log d_{D}(z)+C$$ holds if $z\in D$ tends to $\zeta_0$. Additionally, for $\zeta_0\in\mathbb{C}^{n}_{*}$ the estimate can be improved to $$k_{D}(z_{0},z)\leq-\frac{1}{2}\log d_{{D}}(z)+C'$$ where $C'$ is a constant. \[9\] Let ${D}\subset\mathbb{C}^{n}$ be a pseudoconvex Reinhardt domain. Fix $z_{0}\in D$ and $\zeta_{0}\in\partial D\cap\textnormal{int}\overline{D}$. Then for some constant $C$ the inequality $$k_{{D}}(z_{0},z)\leq\frac{1}{2}\log(-\log d_D(z))+C$$ holds if $z\in D$ tends to $\zeta_{0}$. \[3\] Let ${D}\subset\mathbb{C}^{n}$ be a $\mathcal{C}^{1}$-smooth pseudoconvex Reinhardt domain. Fix $z_{0}\in D$ and $\zeta_{0}\in\partial D$. Then for some constant $C$ the inequality $$k_{{D}}(z_{0},z)\geq-\frac{1}{2}\log d_{{D}}(z)+C$$ holds if $z\in D$ tends non-tangentially to $\zeta_{0}$. Notations and definitions ========================= By $D$ we denote a domain in $\mathbb{C}^{n}$. The *Kobayashi *(*pseudo*)*distance* is defined as\ $k_{D}(w,z):=\sup\{d_{D}(w,z):(d_{D})$ is a family of holomorphically invariant pseudodistances less than or equal to $\widetilde{k}_{D}\}$, where $$\widetilde{k}_{D}(w,z):=\inf\{p(\lambda,\mu):\lambda,\mu\in\mathbb{D}\ \textnormal{and }\exists f\in \mathcal{O}(\mathbb{D},D):\,f(\lambda)=w,\,f(\mu)=z\}$$ is the *Lempert function* of $D$, $\mathbb{D}\subset\mathbb{C}$ — the unit disc and $p$ — the Poincaré distance on $\mathbb{D}$. For general properties of functions $k_{D}$ one can see [@jp]. Denote $z_{j}$ as the $j$-th coordinate of point $z\in\mathbb{C}^{n}$. A domain $D$ is called a *Reinhardt domain* if $(\lambda_{1}z_{1},\ldots,\lambda_{n}z_{n})\in D$ for all numbers $\lambda_{1},\ldots,\lambda_{n}\in\partial\mathbb{D}$ and points $z\in D$. A Reinhardt domain $D$ is *complete in $j$-th direction* if $$(\{1\}^{j-1}\times\overline{\mathbb{D}}\times\{1\}^{n-j})\,\cdotp D\subset D,$$ where $A\,\cdotp B:=\{(a_{1}b_{1},\ldots,a_{n}b_{n}):a\in A,\,b\in B\}$. Define subspaces $V_{j}^{n}:=\{z\in\mathbb{C}^{n}:z_{j}=0\}$ for $j=1,\ldots,n$. If a Reinhardt domain $D$ is complete in the $j$-th direction for all $j$ such that $D\cap V_{j}^{n}\neq\emptyset$ then $D$ is called *relatively complete*. Let us denote $A_{*}:=A\setminus\{0\}$ for a set $A\subset\mathbb{C}$ and $\mathbb{C}^{n}_{*}:=(\mathbb{C}_{*})^{n}$. By $d_{D}(z)$ denote a distance of a point $z\in D$ to $\partial D$ (here, exceptionally, $D$ can be a domain in $\mathbb{R}^{n}$) and by $\zeta_{D}(z)$ — one of points admitting a distance of a point $z\in D$ to $\partial D$. We will use the following main branch of the power $z^{\alpha}:=e^{\alpha\log z}=e^{\alpha(\log|z|+i\textnormal{Arg}\,z)}$, where the main argument $\textnormal{Arg}\,z\in(-\pi,\pi]$. Define $z^{\alpha}:=z_{1}^{\alpha_{1}}\cdotp\ldots\cdotp z_{n}^{\alpha_{n}}$, $|z|^{\alpha}:=|z_{1}|^{\alpha_{1}}\cdotp\ldots\cdotp |z_{n}|^{\alpha_{n}}$ for $z\in\mathbb{C}^{n}_{*}$ and $\alpha\in\mathbb{R}^{n}$. Moreover, let $|z|:=(|z_{1}|,\ldots,|z_{n}|)$ for $z\in\mathbb{C}^{n}$, $\log|z|:=(\log|z_{1}|,\ldots,\log|z_{n}|)$ for $z\in\mathbb{C}^{n}_{*}$ and $\log D:=\{\log|z|:z\in D\cap\mathbb{C}^{n}_{*}\}$ — a *logarithmic image* of $D$. We use $C$ to denote constants not necessarily the same in different places. We also need notations $f\lesssim g$ if there exists $C>0$ such that $f\leq Cg$; $f\approx g$ if $f\lesssim g$ and $g\lesssim f$. We call $D$ a $\mathcal{C}^{k}$*-smooth* domain if for any point $\zeta_{0}\in\partial D$ there exist its open neighbourhood $U\subset\mathbb{C}^{n}$ and a $\mathcal{C}^{k}$-smooth function $\rho:U\longrightarrow\mathbb{R}$ such that 1. $U\cap D=\{z\in U:\rho(z)<0\}$; 2. $U\setminus\overline{D}=\{z\in U:\rho(z)>0\}\neq\emptyset$; 3. $\nabla\rho:=\left(\frac{\partial\rho}{\partial\overline{z}_{1}},\ldots,\frac{\partial\rho}{\partial \overline{z}_{n}}\right)\neq 0$ on $U$. The function $\rho$ is called a *local defining function* for $D$ at the point $\zeta_{0}$. For a $\mathcal{C}^{1}$-smooth domain $D$ we define a *normal vector* to $\partial D$ at a point $\zeta_{0}\in\partial D$ as $$\nu_{D}(\zeta_{0}):=\frac{\nabla\rho(\zeta_{0})}{\|\nabla\rho(\zeta_{0})\|},$$ where $\rho$ is a local defining function for $D$ at $\zeta_{0}$. Clearly $$z=\zeta_{D}(z)-d_{D}(z)\nu_{D}(\zeta_{D}(z))$$ for $z\in D$ and $$\lim_{D\ni z\to\zeta_{0}}\nu_{D}(\zeta_{D}(z))=\nu_{D}(\zeta_{0})$$ for every choice of $\zeta_{D}(z)$. For the transparent notation we shorten the symbol $\nu_{D}(\zeta_{D}(z))$ to $\nu_{D}(z)$. To define a non-tangential convergence we need a concept of a $\textit{cone}$ with a vertex $x_{0}\in\mathbb{R}^{n}$, a semi-axis $\nu\in(\mathbb{R}^{n})_{*}$ and an angle $\alpha\in (0,\frac{\pi}{2})$. It is a set of $x\in\mathbb{R}^{n}\setminus\{x_{0}\}$ such that an angle between vectors $\nu$ and $x-x_{0}$ does not exceed $\alpha$. Let $D$ be a $\mathcal{C}^{1}$-smooth domain and $\zeta_{0}\in\partial D$. We say that $z\in D$ tends *non-tangentially* to $\zeta_{0}$ if there exist a cone $\mathcal{A}\subset\mathbb{C}^{n}\cong\mathbb{R}^{2n}$ with a vertex $\zeta_{0}$, a semi-axis $-\nu_{D}(\zeta_{0})$ and an angle $\alpha\in (0,\frac{\pi}{2})$ and an open neighbourhood $U\subset\mathbb{C}^{n}$ of $\zeta_{0}$ such that $U\cap\mathcal{A}\subset D$ and $z$ tends to $\zeta_{0}$ in $U\cap\mathcal{A}$. We say that a Reinhardt domain $D$ satisfies the *Fu condition* if for any $j\in\{1,\ldots,n\}$ the following implication holds $$\partial D\cap V_{j}^{n}\neq\emptyset\Longrightarrow D\cap V_{j}^{n}\neq\emptyset.$$ The following well-known properties of pseudoconvex Reinhardt domains will be used in the paper (see e.g. [@jp1]). \[4\] A Reinhardt domain $D$ is pseudoconvex if and only if $\,\log D$ is convex and $D$ relatively complete. \[5\] A $\mathcal{C}^{1}$-smooth Reinhardt domain satisfies the Fu condition. Proofs ====== We proceed as follows. The first step is to simplify the general case to ‘real’ coordinates, further we consider some parallelepipeds contained in the given domain and use the decreasing property of the Kobayashi distance. Finally, we explicitly calculate and estimate it in other domains — cartesian products of a belt and annuli in $\mathbb{C}$. To improve the estimate for a boundary point with all non-zero coordinates we use similar methods, but with intervals instead of parallelepipeds. Using some biholomorphism of the form $$w\ni\mathbb{C}^{n}\longmapsto(a_{1}w_{1},\ldots,a_{n}w_{n})\in\mathbb{C}^{n},\,a\in\mathbb{C}_{*}^{n}$$ and the triangle inequality for $k_{D}$, we can assume that $z_{0}=(1,\ldots,1)$ and $|\zeta_{0j}|\neq 1$ for $j=1,\ldots,n$. Notice that the proof can be reduced to $z\in D\cap\mathbb{C}_{*}^{n}$ near $\zeta_0$ and next to the case $$z\in D\cap(0,\infty)^n \text{ near } \zeta_0\in\partial D\cap([0,\infty)\setminus\{1\})^n.$$ Indeed, the first reduction follows from the continuity of $k_{D}$ and the triangle inequality for $k_{D}$. Now, if $z\to\zeta_0$ then $|z|\to|\zeta_0|\in\partial D$ and $$k_D(z_0,z)=k_D(\widetilde{z_0},|z|),$$ where $$\widetilde{z_0}:=\left(\frac{|z_1|}{z_1}z_{01},\ldots,\frac{|z_n|}{z_n}z_{0n}\right)\in T:=\{(\lambda_{1}z_{01},\ldots,\lambda_{n}z_{0n}):\lambda_{1},\ldots,\lambda_{n}\in\partial\mathbb{D}\}.$$ The continuity of $k_{D}$ gives $$\max_{T\times T}k_{D}=:C<\infty$$ and therefore $$k_D(\widetilde{z_0},|z|)\leq k_D(\widetilde{z_0},z_0)+k_D(z_0,|z|)\leq k_D(z_0,|z|)+C.$$ The property $d_D(|z|)=d_D(z)$ finishes this reduction. In what follows, we assume that points $z\in D\cap(0,\infty)^n$ are sufficiently close to $\zeta_0\in\partial D\cap([0,\infty)\setminus\{1\})^n$. Observe that $$d_{\log D}(\log z)\geq\varepsilon d_{D}(z)$$ for some $\varepsilon>0$. Indeed, for $u\in\mathbb{R}^{n} $, $\|u\|<1$ and $0\leq t\leq\varepsilon d_{D}(z)$, where $$\varepsilon:=\frac{1}{3(\|\zeta_0\|+1)},$$ we have $$\log z+tu\in\log D$$ if and only if $$\left(z_{1}e^{tu_{1}},\ldots,z_{n}e^{tu_{n}}\right)\in D$$ but this property follows from $$\left\|(z_{j}e^{tu_{j}})_{j=1}^{n}-z\right\|\leq\sqrt{\sum_{j=1}^{n}z_{j}^{2}(2t)^{2}}\leq 2t(\|\zeta_0\|+1)<d_{D}(z).$$ Moreover, for $\zeta_0=0$ a similar consideration leads to $$d_{\log D}(\log z)\geq\varepsilon'\frac{d_{D}(z)}{\|z\|}$$ for sufficiently small $\varepsilon'>0$. Indeed, there exists $\varepsilon'\in\left(0,\frac{1}{2}\right)$ such that the inequalities $$\left|e^{tu_{j}}-1\right|\leq 2t,\,j=1,\ldots,n$$ hold for $0\leq t\leq\varepsilon'$. Hence for $0\leq t\leq\varepsilon'\frac{d_{D}(z)}{\|z\|}$ we have $$\left\|(z_{j}e^{tu_{j}})_{j=1}^{n}-z\right\|\leq\sqrt{\sum_{j=1}^{n}z_{j}^{2}(2t)^{2}}\leq 2\varepsilon'\frac{d_{D}(z)}{\|z\|}\|z\|<d_{D}(z).$$ Denote $$\widetilde{d}_{D}(z):=\begin{cases}\varepsilon d_{D}(z),\,\zeta_0\neq 0\\\varepsilon''\frac{d_{D}(z)}{\|z\|},\,\zeta_0=0,\end{cases}$$ where $\varepsilon'':=\varepsilon' d_{\log D}(0)$. Let us define $$m_{z}:=\min\{0,\log z_{1}\},\,M_{z}:=\max\{0,\log z_{1}\}$$ and consider the set $$D_{z}:=\{w\in\mathbb{C}^{n}:m_{z}-\widetilde{d}_{D}(z)<\log|w_{1}|<M_{z}+\widetilde{d}_{D}(z),$$$$\frac{\log z_{j}}{\log z_{1}}\log|w_{1}|-\widetilde{d}_{D}(z)<\log|w_{j}|<\frac{\log z_{j}}{\log z_{1}}\log|w_{1}|+\widetilde{d}_{D}(z),\,j=2,\ldots,n\}.$$ Then $\log D_{z}$ is a domain in $\mathbb{R}^{n}$ containing points $0$ and $\log z$ but contained in a convex domain $\log D$. Define also $$G_{z}:=\{v\in\mathbb{C}^{n}:m_{z}-\widetilde{d}_{D}(z)<\textnormal{Re\,}v_{1}<M_{z}+\widetilde{d}_{D}(z),$$$$-\widetilde{d}_{D}(z)<\log|v_{j}|<\widetilde{d}_{D}(z),\,j=2,\ldots,n\}.$$ Hence the holomorphic map $$f_{z}(v):=\left(e^{v_{1}},v_{2}{e^{v_{1}\frac{\log z_{2}}{\log z_{1}}}},\ldots, v_{n}{e^{v_{1}\frac{\log z_{n}}{\log z_{1}}}}\right),\,v\in G_{z}$$ has values in $D_{z}$. Moreover $$w=f_{z}\left(\log w_{1},\frac{w_{2}}{w_{1}^{\frac{\log z_{2}}{\log z_{1}}}},\ldots,\frac{w_{n}}{w_{1}^{\frac{\log z_{n}}{\log z_{1}}}}\right)\textnormal{ for \,} w\in D_{z}.$$ Therefore $$k_{D}(z_{0},z)\leq k_{D_{z}}(z_{0},z)=k_{D_{z}}\left(f_{z}(0,1,\ldots,1),f_{z}\left(\log z_{1},\frac{z_{2}}{z_{1}^{\frac{\log z_{2}}{\log z_{1}}}},\ldots,\frac{z_{n}}{z_{1}^{\frac{\log z_{n}}{\log z_{1}}}}\right)\right)=$$$$=k_{D_{z}}(f_{z}(0,1,\ldots,1),f_{z}(\log z_{1},1,\ldots,1))\leq k_{G_{z}}((0,1,\ldots,1),(\log z_{1},1,\ldots,1))=$$ $$=\max\{k_{S_{z}}(0,\log z_{1}),k_{A_{z}}(1,1),\ldots, k_{A_{z}}(1,1)\}=k_{S_{z}}(0,\log z_{1}),$$ where $$S_{z}:=\left\{\lambda\in\mathbb{C}:m_{z}-\widetilde{d}_{D}(z)<\textnormal{Re\,}\lambda<M_{z}+\widetilde{d}_{D}(z)\right\}$$ and $$A_{z}:=\left\{\lambda\in\mathbb{C}:-\widetilde{d}_{D}(z)<\log|\lambda|<\widetilde{d}_{D}(z)\right\}.$$ Using suitable biholomorphisms, we calculate $$k_{S_{z}}(0,\log z_{1})=p\left(\frac{i-\exp\pi iP(z)}{i+\exp\pi iP(z)},\frac{i-\exp\pi iQ(z)}{i+\exp\pi iQ(z)}\right),$$ where $$P(z):=\frac{\widetilde{d}_{D}(z)-m_{z}}{2\widetilde{d}_{D}(z)+M_{z}-m_{z}},\,Q(z):=\frac{\log z_{1}+\widetilde{d}_{D}(z)-m_{z}}{2\widetilde{d}_{D}(z)+M_{z}-m_{z}}.$$ Analogously, after changing the index 1 to any of $2,\ldots,n$, we get $$k_{D}(z_{0},z)\leq\min_{j=1,\ldots,n}k_{S_{z}^{(j)}}(0,\log z_{j}),$$ where $$k_{S_{z}^{(j)}}(0,\log z_{j})=p\left(\frac{i-\exp\pi iP^{(j)}(z)}{i+\exp\pi iP^{(j)}(z)},\frac{i-\exp\pi iQ^{(j)}(z)}{i+\exp\pi iQ^{(j)}(z)}\right)$$ and $$S_{z}^{(j)}:=\left\{\lambda\in\mathbb{C}:m_{z}^{(j)}-\widetilde{d}_{D}(z)<\textnormal{Re\,} \lambda<M_{z}^{(j)}+\widetilde{d}_{D}(z)\right\},$$ $$m_{z}^{(j)}:=\min\{0,\log z_{j}\},\,M_{z}^{(j)}:=\max\{0,\log z_{j}\},\,j=1,\ldots,n,$$ $$P^{(j)}(z):=\frac{\widetilde{d}_{D}(z)-m_{z}^{(j)}}{2\widetilde{d}_{D}(z)+M_{z}^{(j)}-m_{z}^{(j)}},\,Q^{(j)}(z):=\frac{\log z_{1}+\widetilde{d}_{D}(z)-m_{z}^{(j)}}{2\widetilde{d}_{D}(z)+M_{z}^{(j)}-m_{z}^{(j)}}.$$ Consider two cases: $\zeta_0\neq 0$ and $\zeta_0=0$. If $\zeta_0\neq 0$ then choose $j\in\{1,\ldots,n\}$ such that $\zeta_{0j}\neq 0$ (recall that $\zeta_{0j}=|\zeta_{0j}|\neq 1$). In the case of $\zeta_{0j}>1$ we obtain $$\label{6}k_{S_{z}^{(j)}}(0,\log z_{j})=p\left(\frac{i-\exp\pi iT^{(j)}(z)}{i+\exp\pi iT^{(j)}(z)},\frac{i-\exp\pi iU^{(j)}(z)}{i+\exp\pi iU^{(j)}(z)}\right)\leq$$ $$\leq p\left(0,\frac{i-\exp\pi iT^{(j)}(z)}{i+\exp\pi iT^{(j)}(z)}\right)+p\left(0,\frac{i-\exp\pi iU^{(j)}(z)}{i+\exp\pi iU^{(j)}(z)}\right),$$ where $$T^{(j)}(z):=\frac{\varepsilon d_{D}(z)}{2\varepsilon d_{D}(z)+\log z_{j}},\,U^{(j)}(z):=\frac{\log z_{j}+\varepsilon d_{D}(z)}{2\varepsilon d_{D}(z)+\log z_{j}}.$$ We have, by Taylor expansion $$\frac{i-\exp\pi iT^{(j)}(z)}{i+\exp\pi iT^{(j)}(z)}=i-\pi iT^{(j)}(z)+O\left(d_D(z)^2\right).$$ Hence $$p\left(0,\frac{i-\exp\pi iT^{(j)}(z)}{i+\exp\pi iT^{(j)}(z)}\right)=p\left(0,i-\pi iT^{(j)}(z)+O\left(d_D(z)^2\right)\right)\leq$$$$\leq\frac{\log 2}{2}-\frac{1}{2}\log\left(1-\left|i-\pi iT^{(j)}(z)+O\left(d_D(z)^2\right)\right|\right)\leq$$$$\leq\frac{\log 2}{2}-\frac{1}{2}\log\left(1-\left|i-\pi iT^{(j)}(z)\right|-\left|O\left(d_D(z)^2\right)\right|\right)=$$$$=\frac{\log 2}{2}-\frac{1}{2}\log\left(\pi\frac{\varepsilon d_{D}(z)}{2\varepsilon d_{D}(z)+\log z_{j}}-O\left(d_D(z)^2\right)\right)\leq-\frac{1}{2}\log d_{D}(z)+C.$$ Similarly $$\frac{i-\exp\pi iU^{(j)}(z)}{i+\exp\pi iU^{(j)}(z)}=-i+\pi iT^{(j)}(z)+O\left(d_D(z)^2\right),$$ which gives the same estimation for the second summand. Otherwise if $\zeta_{0j}<1$, we have $$\label{7}k_{S_{z}^{(j)}}(0,\log z_{j})=p\left(\frac{i-\exp\pi iV^{(j)}(z)}{i+\exp\pi iV^{(j)}(z)},\frac{i-\exp\pi iW^{(j)}(z)}{i+\exp\pi iW^{(j)}(z)}\right),$$ where $$V^{(j)}(z):=\frac{\varepsilon d_{D}(z)-\log z_{j}}{2\varepsilon d_{D}(z)-\log z_{j}},\,W^{(j)}(z):=\frac{\varepsilon d_{D}(z)}{2\varepsilon d_{D}(z)-\log z_{j}}.$$ We see that the expression in is the expression in after substitute $\log z_j\leadsto-\log z_j$ and the estimates stay true. Assume $\zeta_0=0$. We have for $j=1,\ldots,n$ $$k_{S_{z}^{(j)}}(0,\log z_{j})=p\left(\frac{i-\exp\pi iX^{(j)}(z)}{i+\exp\pi iX^{(j)}(z)},\frac{i-\exp\pi iY^{(j)}(z)}{i+\exp\pi iY^{(j)}(z)}\right)\leq$$$$\leq p\left(0,\frac{i-\exp\pi iX^{(j)}(z)}{i+\exp\pi iX^{(j)}(z)}\right)+p\left(0,\frac{i-\exp\pi iY^{(j)}(z)}{i+\exp\pi iY^{(j)}(z)}\right),$$ where $$X^{(j)}(z):=\frac{\varepsilon'' d_{D}(z)\|z\|^{-1}-\log z_{j}}{2\varepsilon'' d_{D}(z)\|z\|^{-1}-\log z_{j}},\,Y^{(j)}(z):=\frac{\varepsilon'' d_{D}(z)\|z\|^{-1}}{2\varepsilon'' d_{D}(z)\|z\|^{-1}-\log z_{j}}.$$ Putting $$\delta^{(j)}(z):=\frac{\varepsilon'' d_{D}(z)}{\|z\|\log z_{j}}$$ we have $$X^{(j)}(z)=\frac{\delta^{(j)}(z)-1}{2\delta^{(j)}(z)-1},\,Y^{(j)}(z)=\frac{\delta^{(j)}(z)}{2\delta^{(j)}(z)-1}$$ and $\delta^{(j)}(z)\to 0$ as $z\to 0$. The analogous calculations as in the first case give $$\frac{i-\exp\pi iX^{(j)}(z)}{i+\exp\pi iX^{(j)}(z)}=-i+\pi iY^{(j)}(z)+O\left(\delta^{(j)}(z)^2\right)$$ and $$\frac{i-\exp\pi iY^{(j)}(z)}{i+\exp\pi iY^{(j)}(z)}=i-\pi iY^{(j)}(z)+O\left(\delta^{(j)}(z)^2\right).$$ Therefore $$p\left(0,\frac{i-\exp\pi iX^{(j)}(z)}{i+\exp\pi iX^{(j)}(z)}\right)\leq\frac{\log 2}{2}-\frac{1}{2}\log\left(\pi\frac{\delta^{(j)}(z)}{2\delta^{(j)}(z)-1}-O\left(\delta^{(j)}(z)^2\right)\right)\leq$$ $$\leq-\frac{1}{2}\log(-\delta^{(j)}(z))+C$$ and similarly $$p\left(0,\frac{i-\exp\pi iY^{(j)}(z)}{i+\exp\pi iY^{(j)}(z)}\right)\leq-\frac{1}{2}\log(-\delta^{(j)}(z))+C.$$ Finally $$\min_{j=1,\ldots,n}k_{S_{z}^{(j)}}(0,\log z_{j})\leq\min_{j=1,\ldots,n}-\log(-\delta^{(j)}(z))+C=$$$$=-\log d_{D}(z)+\log\|z\|+\min_{j=1,\ldots,n}\log(-\log z_{j})+C=$$$$=-\log d_{D}(z)+\log\|z\|+\log\left(-\log\max_{j=1,\ldots,n}z_{j}\right)+C\leq$$$$\leq-\log d_{D}(z)+\log\|z\|+\log(-\log\|z\|)+C\leq-\log d_D(z)+C.$$ For improving the estimate in the case of $\zeta_0\in\partial D\cap\mathbb{C}^{n}_{*}$, we may assume that $z_0\in\mathbb{C}_{*}^{n}$ and $|z_{0j}|,|\zeta_{0j}|\neq 1$ for $j=1,\ldots,n$. Since $\log D$ is a convex domain, the interval $$I_{z}:=\{t\log|z|+(1-t)\log|z_0|:t\in(-\varepsilon(z),1+\delta(z))\}$$ is contained in $\log D$ for some positive numbers $\delta(z)$, $\varepsilon(z)$. The number $\varepsilon(z)$ can be chosen as a sufficiently small positive constant $\varepsilon$ independent of $z$. Indeed, $$t\log|z|+(1-t)\log|z_0|=\log|z_0|+t(\log|z|-\log|z_0|)$$ and $\|\log|z|-\log|z_0|\|$ is bounded, say by $M$. Hence $$\varepsilon:=\frac{d_{\log D}(\log|z_0|)}{2M}$$ is good. Analogously, $$\frac{d_{\log D}(\log|z|)}{2M}$$ is a candidate for $\delta(z)$. We have $$\frac{d_{\log D}(\log|z|)}{2M}\geq\delta d_{D}(z)$$ for some $\delta>0$ (in fact, “$\geq$” can be replaced with “$\approx$”). Thus we can choose $\delta(z):=\delta d_{D}(z)$. From the inclusion $I_{z}\subset\log D$ it follows that $$\exp I_{z}\subset D$$ i.e. $$\left(\left|\frac{z_{1}}{z_{01}}\right|^{t}|z_{01}|,\ldots,\left|\frac{z_{n}}{z_{0n}}\right|^{t}|z_{0n}|\right)\in D$$ for $t\in(-\varepsilon,1+\delta d_{D}(z))$. Hence the holomorphic map $$f_z(\lambda):=\left(e^{i\arg z_{1}}\left|\frac{z_{1}}{z_{01}}\right|^{\lambda}|z_{01}|,\ldots,e^{i\arg z_{n}}\left|\frac{z_{n}}{z_{0n}}\right|^{\lambda}|z_{0n}|\right)$$ leading from the strip $$S_{z}:=\{\lambda\in\mathbb{C}:-\varepsilon<\textnormal{Re}\,\lambda<1+\delta d_{D}(z)\}$$ has values in $D$. Moreover $f_z(1)=z$ and $f_z(0)$ lies on the torus $$T:=\{(\lambda_{1}z_{01},\ldots,\lambda_{n}z_{0n}):\lambda_{1},\ldots,\lambda_{n}\in\partial\mathbb{D}\}.$$ Therefore $$k_{D}(z_{0},z)\leq k_{D}(z_{0},f_z(0))+k_{D}(f_z(0),z)\leq$$$$\leq k_{D}(f_z(0),f_z(1))+\max_{T\times T}k_{D}\leq k_{S_{z}}(0,1)+\max_{T\times T}k_{D}.$$ Calculating $k_{S_{z}}(0,1)$ we get $$k_{S_{z}}(0,1)=p\left(\frac{i-\exp\pi iP^{(j)}(z)}{i+\exp\pi iP^{(j)}(z)},\frac{i-\exp\pi iQ^{(j)}(z)}{i+\exp\pi iQ^{(j)}(z)}\right),$$ where $$P^{(j)}(z):=\frac{\varepsilon}{1+\varepsilon+\delta d_{D}(z)},\,Q^{(j)}(z):=\frac{1+\varepsilon}{1+\varepsilon+\delta d_{D}(z)}.$$ Certainly, first of the above argument of the function $p$ tends to some point from the unit disc. For the second we have $$\frac{i-\exp\pi iQ^{(j)}(z)}{i+\exp\pi iQ^{(j)}(z)}=-i+\pi i\frac{\delta d_{D}(z)}{1+\varepsilon+\delta d_{D}(z)}+O\left(d_D(z)^2\right).$$ Consequently $$p\left(0,\frac{i-\exp\pi iQ^{(j)}(z)}{i+\exp\pi iQ^{(j)}(z)}\right)\leq\frac{\log 2}{2}-\frac{1}{2}\log\left(\pi\frac{\delta d_{D}(z)}{1+\varepsilon+\delta d_{D}(z)}-O\left(d_{D}(z)^2\right)\right)\leq$$$$\leq-\frac{1}{2}\log d_{D}(z)+C.$$ The triangle inequality for $p$ finishes the proof. The proof is based on decreasing and product properties of the Kobayashi distance and need to consider some cases which form, in fact, an induction. Note that if $E\subset\mathbb{R}^n$ is a convex domain then $E=\text{int}\overline{E}$. The condition $\zeta_{0}\in\partial D\cap\text{int}\overline{D}$ implies $\zeta_{0}\notin\mathbb{C}^{n}_{*}$. To see this, assume that $\zeta_{0}\in\mathbb{C}^{n}_{*}$. An easy topological argument shows that $$\log|\zeta_0|\in(\partial\log D)\cap\text{int\,}\overline{\log D}=(\partial\log D)\cap\log D=\emptyset.$$ Assume, without loss of generality, that $$\zeta_{0}=(\zeta_{01},\ldots,\zeta_{0k},0,\ldots,0),$$ where $0\leq k\leq n-1$ and $\zeta_{0j}\neq 0$, $j\leq k$. Let $r>0$ be such that an open polydisc $P(\zeta_0,r)$ is contained in $\overline{D}$. Then $\log P(\zeta_0,r)\subset\log\overline{D}$. Taking interiors of both sides we get $$\log P(\zeta_0,r)\subset\text{int}\log\overline{D}=\text{int\,}\overline{\log D}=\log D.$$ Therefore $$\label{10}P(\zeta_0,r)\cap\mathbb{C}^{n}_{*}\subset D.$$ Clearly (for fixed small $r$)$$P(\zeta_0,r)\cap\mathbb{C}^{n}_{*}=\mathbb{D}(\zeta_{01},r)\times\ldots\times\mathbb{D}(\zeta_{0k},r)\times (r\mathbb{D}_*)^{n-k},$$ where $\mathbb{D}(\zeta_{0j},r)$ is a disc in $\mathbb{C}$ centered at $\zeta_{0j}$ with radius $r$. Hence, choosing any $z_0\in P(\zeta_0,r)\cap\mathbb{C}^{n}_{*}$, we have $$k_D(z_0,z)\leq\max\left\{\max_{j=1,\ldots,k}k_{\mathbb{D}(\zeta_{0j},r)}(z_{0j},z_j),\max_{j=k+1,\ldots,n} k_{r\mathbb{D}_*}(z_{0j},z_{j})\right\}$$ for $z\in D\cap\mathbb{C}^{n}_{*}$ near $\zeta_0$. For $j=1,\ldots,k$ the numbers $z_j$ tend to $\zeta_{0j}$, so the first of the above maxima is bounded by a constant. The well-known estimate for the punctured disc gives us $$k_{r\mathbb{D}_*}(z_{0j},z_{j})\leq\frac{1}{2}\log(-\log d_{r\mathbb{D}_*}(z_j))+C= \frac{1}{2}\log(-\log|z_j|)+C$$ for $j=k+1,\ldots,n$. Therefore $$\label{11}k_D(z_0,z)\leq\frac{1}{2}\log\left(-\log\min_{j=k+1,\ldots,n}|z_j|\right)+C.$$ The above estimate is not sufficiently good yet. Denote $z':=(z_1,\ldots,z_k)$. Note that $$\label{13}(z',0,\ldots,0)\in\partial D.$$ Indeed, $(z',0,\ldots,0)\in\overline{D}$. If $(z',0,\ldots,0)\in D$ then $D$ is complete in the directions $k+1,\ldots,n$ (Fact \[4\]). Moreover, $(\zeta_{01},\ldots,\zeta_{0k},r/2,\ldots,r/2)\in D$, which implies $(\zeta_{01},\ldots,\zeta_{0k},0,\ldots,0)\in D$ — a contradiction. We claim that for all $k+1\leq p<q\leq n$ $$\label{12}(z',0,\ldots,0,\underline{z_p},0,\ldots,0)\in\partial D \text{ or }(z',0,\ldots,0,\underline{z_q},0,\ldots,0)\in\partial D,$$ where the symbol $\underline{z_j}$ means that $z_j$ is on the $j$-th place. If it is not true then both points belong to $D$ (recall that $P(\zeta_0,r)\subset\overline{D}$). Hence $D$ is complete in the directions $k+1,\ldots,n$ and $(z',0,\ldots,0)\in D$, which contradicts . Therefore all points $$(z',0,\ldots,0,\underline{z_p},0,\ldots,0),\,p=k+1,\ldots,n,$$ except possibly one, belong to $\partial D$. Consider the following cases. Case 1.1. One of above points, say $(z',0,\ldots,0,z_n)$, does not belong to $\partial D$. Then it belongs to $D$. Hence $D$ is complete in the directions $k+1,\ldots,n-1$. Now the inclusion can be improved to $$P(\zeta_0,r)\cap(\mathbb{C}^{n-1}\times\mathbb{C}_*)\subset D$$ and $$P(\zeta_0,r)\cap(\mathbb{C}^{n-1}\times\mathbb{C}_*)=\mathbb{D}(\zeta_{01},r)\times\ldots\times\mathbb{D}(\zeta_{0k},r)\times (r\mathbb{D})^{n-k-1}\times r\mathbb{D}_*.$$ The estimate for $k_D(z_0,z)$ is improved to $$\max\left\{\max_{j=1,\ldots,k}k_{\mathbb{D}(\zeta_{0j},r)}(z_{0j},z_j),\max_{j=k+1,\ldots,n-1} k_{r\mathbb{D}}(z_{0j},z_{j}),\,k_{r\mathbb{D}_*}(z_{0n},z_{n})\right\}=$$$$= k_{r\mathbb{D}_*}(z_{0n},z_{n})\leq\frac{1}{2}\log(-\log|z_n|)+C.$$ It remains to notice that $$(z',z_{k+1},\ldots,z_{n-1},0)\in\partial D$$ since in the opposite case the domain $D$ would be complete in $n$-th direction and the property $(z',0,\ldots,0,z_n)\in D$ would imply $(z',0,\ldots,0)\in D$ — a contradiction with . Thus $$d_D(z)\leq\|z-(z',z_{k+1},\ldots,z_{n-1},0)\|=|z_n|,$$ which let us estimate $$\frac{1}{2}\log(-\log|z_n|)+C\leq\frac{1}{2}\log(-\log d_D(z))+C.$$ Case 1.2. All the points $$(z',0,\ldots,0,\underline{z_p},0,\ldots,0),\,p=k+1,\ldots,n$$ belong to $\partial D$. We claim that for all $k+1\leq p<q\leq n$ and $k+1\leq p'<q'\leq n$ with $\{p,q\}\neq\{p',q'\}$ $$(z',0,\ldots,0,\underline{z_p},0,\ldots,0,\underline{z_q},0,\ldots,0)\in\partial D \text{ or }(z',0,\ldots,0,\underline{z_{p'}},0,\ldots,0,\underline{z_{q'}},0,\ldots,0)\in\partial D.$$ Analogously as before we use an argument of completeness in the suitable directions to get $$(z',0,\ldots,0,\underline{z_j},0,\ldots,0)\in D$$ for some $j\in\{p,q,p',q'\}$ — a contradiction with the assumption of the case 1.2. Therefore all points $$(z',0,\ldots,0,\underline{z_p},0,\ldots,0,\underline{z_q},0,\ldots,0),\,k+1\leq p<q\leq n,$$ except possibly one, belong to $\partial D$. Again we consider two cases. Case 2.1. One of above points, say $(z',0,\ldots,0,z_{n-1},z_n)$, does not belong to $\partial D$. Then it belongs to $D$. We see, analogously as in the case 1.1, that $$P(\zeta_0,r)\cap(\mathbb{C}^{n-2}\times\mathbb{C}^2_*)\subset D,$$ $$k_D(z_0,z)\leq\frac{1}{2}\log\left(-\log\min_{j=n-1,n}|z_j|\right)+C,$$ $$(z',z_{k+1},\ldots,z_{n-2},z_{n-1},0),\,(z',z_{k+1},\ldots,z_{n-2},0,z_{n})\in\partial D,$$ $$d_D(z)\leq\min_{j=n-1,n}|z_{j}|.$$ Case 2.2. All the points $$(z',0,\ldots,0,\underline{z_p},0,\ldots,0,\underline{z_q},0,\ldots,0),\,k+1\leq p<q\leq n$$ belong to $\partial D$. We see, by induction, that in the $s$-th step ($s=3,\ldots,n-k-1$) all points $$(z',0,\ldots,0,\underline{z_{p_1}},0,\ldots,0,\underline{z_{p_s}},0,\ldots,0),\,k+1\leq p_1<\ldots p_s\leq n,$$ except possibly one, belong to $\partial D$. If one of these points, say $(z',0,\ldots,0,z_{n-s+1},\ldots,z_n)$, does not belong to $\partial D$ then it belongs to $D$ and $$P(\zeta_0,r)\cap(\mathbb{C}^{n-s}\times\mathbb{C}^s_*)\subset D,$$ $$k_D(z_0,z)\leq\frac{1}{2}\log(-\log\min_{j=n-s+1,\ldots,n}|z_j|)+C,$$ $$(z',z_{k+1},\ldots,z_{n-s},z_{n-s+1},\ldots,z_{j-1},0,z_{j+1},\ldots,z_n)\in\partial D,\,j=n-s+1,\ldots,n,$$ $$d_D(z)\leq\min_{j=n-s+1,\ldots,n}|z_{j}|,$$ which finishes the proof in the case $s.1$. If all the points $$(z',0,\ldots,0,\underline{z_{p_1}},0,\ldots,0,\underline{z_{p_s}},0,\ldots,0),\,k+1\leq p_1<\ldots p_s\leq n$$ belong to $\partial D$ then we “jump” from the case $s$.2 to the case $(s+1).1$, getting finally in the case $(n-k-1).1$ $$(z',0,z_{k+2},\ldots,z_n)\in D,$$ $$P(\zeta_0,r)\cap(\mathbb{C}^{k+1}\times\mathbb{C}^{n-k-1}_*)\subset D,$$ $$k_D(z_0,z)\leq\frac{1}{2}\log(-\log\min_{j=k+2,\ldots,n}|z_j|)+C,$$ $$(z',z_{k+1},z_{k+2},\ldots,z_{j-1},0,z_{j+1},\ldots,z_n)\in\partial D,\,j=k+2,\ldots,n,$$ $$d_D(z)\leq\min_{j=k+2,\ldots,n}|z_{j}|$$ or in the case $(n-k-1).2$ $$(z',z_{k+1},\ldots,z_{j-1},0,z_{j+1},\ldots,z_{n})\in\partial D,\,j=k+1,\ldots,n.$$ This property let us estimate $d_D(z)$ from above by $\min_{j=k+1,\ldots,n}|z_j|$ and use to finish the proof. The proof has two main parts; in the first the claim is proved for $\zeta_0\in\partial D\cap\mathbb{C}^{n}_{*}$ thanks to the effective formulas for the Kobayashi distance in special domains and in the second part the remaining case is amounted to the lower-dimensional situation with a boundary point having all non-zero coordinates. Let $\zeta_0\in\partial D\cap\mathbb{C}^{n}_{*}$ and consider $z\in D\cap\mathbb{C}^{n}_{*}$ close to $\zeta_0$. From the convexity of the set $\log{D}$ there exist $\alpha\in\mathbb{R}^{n}$ and $c>0$ such that the hyperplane $$\{x\in\mathbb{R}^{n}:\langle\alpha,x\rangle_{\mathbb{R}^{n}}=\log c\}$$ contains point $\log|\zeta_0|$ and $\log{D}$ lies on the one side of this hyperplane. Assume, without loss of generality, that this side is $\{x\in\mathbb{R}^{n}:\langle\alpha,x\rangle_{\mathbb{R}^{n}}<\log c\}$ since in the case of $\log{D}\subset\{x\in\mathbb{R}^{n}:\langle\alpha ',x\rangle_{\mathbb{R}^{n}}>\log c'\}$ it suffices to define $$\alpha:=-\alpha '\text{ and }c:=1/c'.$$ Therefore $$\{( e^{x_{1}},\ldots,e^{x_{n}}):x\in \log D\}\subset\{w\in\mathbb{C}^{n}:|w|^{\alpha}<c\}=:D_{\alpha,c}\footnote{These sets are called \textit{elementary Reinhardt domains}.},$$ where by a point satisfying the condition $|w|^{\alpha}<c$ we mean such point $w$ whose coordinate $w_{j}$ is non-zero when $\alpha_j<0$ (and satisfies $|w|^{\alpha}<c$ in the usual sense). To affirm that $D\subset D_{\alpha,c}$, we have to check that the above restriction for points $w$ does not remove from $D$ points with some zero coordinates. Indeed, if there is no such inclusion, we can assume that the order of zero coordinates of point $w\in{D}$ and negative terms of the sequence $\alpha$ is as follows: $$w_{1},\ldots,w_{k}\neq 0,\, w_{k+1},\ldots,w_{n}=0$$ $$\alpha_{k+1},\ldots,\alpha_{l}\geq 0,\,\alpha_{l+1},\ldots,\alpha_{n}<0,$$ where $1\leq k\leq l<n$. In some neighbourhood of the point $w$ contained in ${D}$ there exist points $v\in\mathbb{C}_{*}^{n}$ with coordinates $v_{j}$ such that $$|v_{1}|,\ldots,|v_{l}|>\varepsilon>0$$ and $|v_{l+1}|,\ldots,|v_{n}|$ arbitrarily close to zero (i.e. moved from $w$ in a direction of subspace $\{0\}^{l}\times\mathbb{C}^{n-l}$ and next moved from it by a constant vector in the direction $\mathbb{C}^{l}\times\{0\}^{n-l}$). Then there exist points $u\in\log{D}$ whose coordinates $u_{j}$ satisfy $$u_{1},\ldots,u_{l}>\log\varepsilon>-\infty$$ however $u_{l+1},\ldots,u_{n}$ are arbitrarily close to $-\infty$. But it contradicts a fact that values of the expression $$\sum_{j=l+1}^{n}\alpha_{j}u_{j}$$ are for these points $u$ bounded from above by a constant $\log c-\sum_{j=1}^{l}\alpha_{j}\log\varepsilon$. We will use effective formulas for the Kobayashi distance in domains $D_{\alpha,c}$ [@zw1]. Define $$l:=\#\{j=1,\ldots,n:\alpha_{j}<0\}$$ and $$\widetilde{\alpha}:=\min\{\alpha_{j}:\alpha_{j}>0\}\,\textnormal{ if }\,l<n.$$ We first consider a situation $l<n$. The formula in this case gives $$k_{D}(z_{0},z)\geq k_{D_{\alpha,c}}(z_{0},z)\geq p\left(0,\frac{|z|^{\alpha/\widetilde{\alpha}}}{c^{1/\widetilde{\alpha}}}\right)+C.$$ But $$z=\zeta_{D_{\alpha,c}}(z)-d_{D_{\alpha,c}}(z)\nu_{D_{\alpha,c}}(z)$$ and hence $$|z|^{\alpha/\widetilde{\alpha}}=\prod\limits_{j=1}^{n}|\zeta_{D_{\alpha,c}}(z)_{j}-d_{D_{\alpha,c}}{(z)} \nu_{D_{\alpha,c}}(z)_{j}|^{\alpha_{j}/\widetilde{\alpha}}=c^{1/\widetilde{\alpha}}-\rho(z)d_{D_{\alpha,c}}(z)$$ for some bounded positive function $\rho$. Thus $$p\left(0,\frac{|z|^{\alpha/\widetilde{\alpha}}}{c^{1/\widetilde{\alpha}}}\right)= p\left(0,1-\frac{\rho(z)}{c^{1/\widetilde{\alpha}}}d_{D_{\alpha,c}}(z)\right)\geq -\frac{1}{2}\log\left(\frac{\rho(z)}{c^{1/\widetilde{\alpha}}}d_{D_{\alpha,c}}(z)\right)\geq$$$$\geq -\frac{1}{2}\log d_{D_{\alpha,c}}(z)+C.$$ We will show that $$d_{D_{\alpha,c}}(z)\approx d_{{D}}(z)\textnormal{ as }z\to\zeta_{0} \textnormal{ non-tangentially}.$$ By the definition there exists a cone $\mathcal{A}$ with a vertex $\zeta_{0}$ and a semi-axis $-\nu_{D_{\alpha,c}}(\zeta_{0})$ which contains considered points $z$. Thanks to the $\mathcal{C}^{1}$-smoothness of ${D}$ we have a cone $\mathcal{B}$ with the vertex $\zeta_{0}$ and the semi-axis $-\nu_{D_{\alpha,c}}(\zeta_{0})$, whose intersection with some neighbourhood of the point $\zeta_{0}$ is contained in ${D}$ and contains in its interior the cone $\mathcal{A}$. Therefore $$1\geq\frac{{d_{{D}}(z)}}{d_{D_{\alpha,c}}(z)}=\frac{\|z-\zeta_{{D}}(z)\|}{\|z-\zeta_{D_{\alpha,c}}(z)\|}\geq \frac{\|z- \zeta_{{D}}(z)\|}{\|z-\zeta_{0}\|}\geq\frac{\|z-\zeta_{\mathcal{B}}(z)\|}{\|z-\zeta_{0}\|}=$$ $$=\sin\angle(z,\zeta_{0},\zeta_{\mathcal{B}}(z))\geq\sin\theta,$$ where $\angle(X,Y,Z)$ is an angle with vertex $Y$, whose arms contain points $X$, $Z$ and $\theta$ is an angle between these generatrices of cones $\mathcal{A}$, $\mathcal{B}$ which lie in one plane with the axis of both cones.[^1] The second case $l=n$ gives $$k_{D}(z_{0},z)\geq k_{D_{\alpha,c}}(z_{0},z)\geq p\left(0,\frac{|z|^{\alpha}}{c}\right)+C.$$ Similarly as before $$|z|^{\alpha}=\prod\limits_{j=1}^{n}|\zeta_{D_{\alpha,c}}(z)_{j}-d_{D_{\alpha,c}}{(z)} \nu_{D_{\alpha,c}}(z)_{j}|^{\alpha_{j}}=c-\sigma(z)d_{D_{\alpha,c}}(z)$$ with a bounded positive function $\sigma$. Hence $$p\left(0,\frac{|z|^{\alpha}}{c}\right)\geq-\frac{1}{2}\log d_{D_{\alpha,c}}(z)+C\geq-\frac{1}{2}\log d_D(z)+C.$$ Now, take $\zeta_{0}\in\partial D\setminus\mathbb{C}_{*}^{n}$. We may assume that the first $k$ coordinates of $\zeta_0$ are non-zero and the last $n-k$ are zero, where $0\leq k\leq n-1$. Notice that $k\neq 0$. Indeed, the assumption $k=0$ is equivalent to $0\in\partial{D}$. Using Facts \[4\] and \[5\] we see that the $\mathcal{C}^{1}$-smoothness of $D$ implies, thanks to the Fu condition, ${D}\cap V_{j}^{n}\neq\emptyset$ for $j=1,\ldots,n$. Hence $D$ is complete i.e. $0\in D$ — a contradiction. Finally, point $\zeta_{0}$ has a form $$\zeta_{0}=(\zeta_{01},\ldots,\zeta_{0k},0,\ldots,0),\,\zeta_{0j}\neq 0,\,1\leq j\leq k\leq n-1.$$ Consider the projection $\pi_{k}:\mathbb{C}^{n}\longrightarrow\mathbb{C}^{k}$ i.e. $$\pi_{k}(z)=(z_{1},\ldots,z_{k}).$$ We will show that ${D}_{k}:=\pi_{k}({D})$ is a $\mathcal{C}^{1}$-smooth pseudoconvex Reinhardt domain. A Reinhardt property is clear for ${D}_{k}$. To affirm the pseudoconvexity of $D_k$ it suffices to show that $${D}_{k}\times\{0\}^{n-k}={D}\cap(\mathbb{C}^{k}\times\{0\}^{n-k}).$$ Inclusion $${D}_{k}\times\{0\}^{n-k}\supset{D}\cap(\mathbb{C}^{k}\times\{0\}^{n-k})$$ is obvious. To prove the opposite inclusion we will use Facts \[4\] and \[5\] again. We have ${D}\cap V_{j}^{n}\neq\emptyset$ for $j=k+1,\ldots,n$, so $D$ is complete in $j$-th direction for $j=k+1,\ldots,n$. Take some $z\in D_{k}\times\{0\}^{n-k}$. Then $z=(z_{1},\ldots,z_{k},0\ldots,0)$ and $(z_{1},\ldots,z_{k},\widetilde{z}_{k+1},\ldots,\widetilde{z}_{n})\in D$ for some $\widetilde{z}_{k+1},\ldots,\widetilde{z}_{n}\in\mathbb{C}$. Thus $(z_{1},\ldots,z_{k},0\ldots,0)\in D$ i.e. $z\in D\cap(\mathbb{C}^{k}\times\{0\}^{n-k})$. The local defining function for ${D}_{k}$ at point $\zeta\in\partial{D}_{k}$ is $$\widetilde{\rho}(z_{1},\ldots,z_{k}):=\rho(z_{1},\ldots,z_{k},0,\ldots,0),\,(z_{1},\ldots,z_{k})\in\pi_{k}(U)\cap{D}_{k},$$ where $\rho:U\longrightarrow\mathbb{R}$ is the local defining function for ${D}$ at point $(\zeta,0,\ldots,0)$. Indeed, $\nabla\widetilde{\rho}\neq 0$ since - $\nabla\rho\neq 0$; - $\frac{\partial\widetilde{\rho}}{\partial\overline{z}_{j}}= \frac{\partial\rho}{\partial\overline{z}_{j}}$ for $j=1,\ldots,k$; - $\frac{\partial\rho}{\partial\overline{z}_{j}}=0$ for $j=k+1,\ldots,n$, however the two remaining conditions for a defining function follow easy from the definition of $\widetilde{\rho}$. If $z$ tends to $\zeta_{0}$ non-tangentially in a cone $\mathcal{A}\subset\mathbb{C}^{n}$ then $\pi_{k}(z)$ tends to $\pi_{k}(\zeta_{0})\in\mathbb{C}_{*}^{k}$ non-tangentially in a cone $\pi_{k}(\mathcal{A})\subset\mathbb{C}^{k}$. From the previous case $$k_{{D}}(z_{0},z)\geq k_{{D}_{k}}(\pi_{k}(z_{0}),\pi_{k}(z))\geq-\frac{1}{2}\log d_{{D}_{k}}(\pi_{k}(z))+C.$$ Hence to finish the proof it suffices to show that $$d_{{D}_{k}}(\pi_{k}(z))\lesssim d_{{D}}(z).$$ Consider a cone $\mathcal{B}$ with vertex $\zeta_{0}$ and semi-axis $-\nu_{D_{\alpha}}(\zeta_{0})$ whose intersection with some neighbourhood of the point $\zeta_{0}$ is contained in ${D}$ and contains in its interior the cone $\mathcal{A}$. Then $$1\geq\frac{d_{\mathcal{B}}(z)}{d_{{D}}(z)}=\frac{\|z-\zeta_{\mathcal{B}}(z)\|}{\|z-\zeta_{{D}}(z)\|}\geq \frac{\|z-\zeta_{\mathcal{B}}(z)\|}{\|z-\zeta_{0}\|}=\sin\angle(z,\zeta_{0},\zeta_{\mathcal{B}}(z))\geq\sin\theta,$$ where $\theta$ is an angle between these generatrices of the cones $\mathcal{A}$, $\mathcal{B}$ which lie in one plane with the axis of both cones. Analogously $$1\geq\frac{d_{\pi_{k}(\mathcal{B})}(\pi_{k}(z))}{d_{{D}_{k}}(\pi_{k}(z))}\geq\sin\theta',$$ where $\theta'$ depends only on $\mathcal{B}$. Therefore $$\frac{d_{{D}_{k}}(\pi_{k}(z))}{d_{{D}}(z)}\approx\frac{d_{\pi_{k}(\mathcal{B})}(\pi_{k}(z))}{d_{\mathcal{B}}(z)},$$ however $$\frac{d_{\pi_{k}(\mathcal{B})}(\pi_{k}(z))}{d_{\mathcal{B}}(z)}= \frac{\|\pi_{k}(z)-\zeta_{\pi_{k}(\mathcal{B})}(\pi_{k}(z))\|}{\|z-\zeta_{\mathcal{B}}(z)\|}=$$$$=\frac{\|\pi_{k}(z)-\pi_{k}(\zeta_{0})\|\sin\angle(\pi_{k}(z),\pi_{k}(\zeta_{0}),\zeta_{\pi_{k}(\mathcal{B})}(\pi_{k}(z)))} {\|z-\zeta_{0}\|\sin\angle(z,\zeta_{0},\zeta_{\mathcal{B}}(z))}\leq$$$$\leq\frac{\|\pi_{k}(z)-\pi_{k}(\zeta_{0})\|}{\|z-\zeta_{0}\|\sin\theta}\leq \frac{1}{\sin\theta}.\eqno\qedhere$$ The estimate from below by $-\frac{1}{2}\log d_D+C$ for the Carathéodory (pseudo) distance $c_D$ is not true even for smooth bounded complete pseudoconvex Reinhardt domain $D$ and its boundary point $\zeta_0\in\mathbb{C}^{n}_{*}$. Consider a domain $$D:=\left\{(z_1,z_2)\in\mathbb{C}^2:|z_1|<R_1,\,|z_2|<R_2,\,|z_1||z_2|^\alpha<R_3\right\},$$ where $R_1,R_2,R_3>0$, $\alpha\in(\mathbb{R}\setminus\mathbb{Q})_+$ and $R_{1}R_{2}^{\alpha}>R_3$. Fix $\zeta_0\in\partial D$ such that $|\zeta_{01}|<R_1,|\zeta_{02}|<R_2$. This domain is not smooth. Since $$\log D=\left\{(x_1,x_2)\in\mathbb{R}^2:x_1<\log R_1,\,x_2<\log R_2,\,x_1+\alpha x_2<\log R_3\right\},$$ it is easy to construct smooth bounded convex domain $E\subset\mathbb{R}^2$ such that $\log D\subset E$ and $\partial E$ contains the skew segment $$(\partial\log D)\cap\{(x_1,x_2)\in\mathbb{R}^2:x_1+\alpha x_2=\log R_3\}.$$ Let $\widetilde{D}\subset\mathbb{C}^2$ be a complete Reinhardt domain such that $\log\widetilde{D}=E$. Then $\widetilde{D}$ is bounded, smooth and, thanks to Fact \[4\], pseudoconvex. Moreover, $D\subset\widetilde{D}$ and $D$, $\widetilde{D}$ are identic in the neighbourhood of their common boundary point $\zeta_0$. We have from the Proposition 4.3.2 in [@zw1] $$a_\lambda:=\frac{g_D(\lambda\zeta_0,0)}{\log|\lambda|}\to\infty\text{ as }\lambda\to\partial\mathbb{D},$$ where $g_D$ is the pluricomplex Green function (general properties of the Carathéodory (pseudo)distance and the pluricomplex Green function one can find e.g. in [@jp] and [@zw1]). Certainly $$d_{\widetilde{D}}(\lambda\zeta_0)=d_D(\lambda\zeta_0)\approx 1-|\lambda| \text{ as } |\lambda|\to 1$$ and $$c_D(\lambda\zeta_0,0)\leq\tanh^{-1}\exp g_D(\lambda\zeta_0,0).$$ Therefore, if there exists a constant $C>0$ such that $$c_{\widetilde{D}}(\lambda\zeta_0,0)\geq-\frac{1}{2}\log d_{\widetilde{D}}(\lambda\zeta_0)+C,\,|\lambda|\to 1$$ then for $|\lambda|\to 1$ $$c_{D}(\lambda\zeta_0,0)\geq-\frac{1}{2}\log d_{D}(\lambda\zeta_0)+C,$$ $$-\frac{1}{2}\log(1-|\lambda|)+C\leq\tanh^{-1}|\lambda|^{a_\lambda},$$ $$\frac{1}{1-|\lambda|}\leq\frac{C'}{1-|\lambda|^{a_\lambda}}$$ with a constant $C'>0$. For $|\lambda|$ sufficiently close to 1 we have $a_\lambda\geq C'+1$. Therefore $$\frac{1}{1-|\lambda|}\leq\frac{C'}{1-|\lambda|^{C'+1}}$$ or equivalently $$\frac{1-|\lambda|^{C'+1}}{1-|\lambda|}\leq C'.$$ The left-hand side tends to $C'+1$ as $|\lambda|\to 1$. Open problems ============= (1) Can we improve the estimate from Theorem \[1\] to $-\frac{1}{2}\log d_{{D}}(z)+C$? (2) Let ${D}\subset\mathbb{C}^{n}$ be a pseudoconvex Reinhardt domain and $\zeta_{0}\in\partial D\cap\mathbb{C}^{n}_{*}$. Does it implies that for some constant $C$ the inequality $$k_{{D}}(z_{0},z)\geq-\frac{1}{2}\log d_{{D}}(z)+C$$ holds if $z\in D$ tends to $\zeta_{0}$? (3) Is it true for pseudoconvex Reinhardt domains $D\subset\mathbb{C}^{n}$ that if $$\#\{j:\zeta_{0j}=0\text{ and } D\cap V_{j}^{n}=\emptyset\}=0$$ then $$k_{D}(z_{0},z)\geq-\frac{1}{2}\log d_{D}(z)+C$$ and in the opposite case $$k_{D}(z_{0},z)\geq\frac{1}{2}\log(-\log d_{D}(z))+C$$ for $z\in D$ near $\zeta_0\in\partial D$?\ <span style="font-variant:small-caps;">Acknowledgements.</span> I would like to thank Professor W³odzimierz Zwonek for introduction to the problem and helpful suggestions. [9]{} , *Localization of the Kobayashi metric and the boundary continuity of proper holomorphic mappings*, Math. Ann. [**279**]{} (1987), 239–252. , *On completeness of invariant metrics of Reinhardt domains*, Arch. Math. (Basel) [**63**]{} (1994), 166–172. , *Invariant Distances and Metrics in Complex Analysis*, Walter de Gruyter, 1993. , *First Steps in Several Complex Variables: Reinhardt Domains*, European Mathematical Society Publishing House, 2008. , *Complex geodesics and iterates of holomorphic maps on convex domains in $\mathbb{C}^{n}$*, Trans. Amer. Math. Soc. [**338**]{} (1993), no. 1, 201–211. , *Upper bound for the Lempert function of smooth domains*, Math. Z. [**266**]{} (2010), no. 2, 425–430. , *About the Carathéodory completeness of all Reinhardt domains*, Functional Analysis, Holomorphy and Approximation Theory II, G. Zapata ed., 331–337, Amsterdam 1984. , *On hyperbolicity of pseudoconvex Reinhardt domains*, Arch. Math. (Basel) [**72**]{} (1999), 304–314. , *Completeness, Reinhardt domains and the method of complex geodesics in the theory of invariant functions*, Diss. Math. [**388**]{} (2000). [^1]: In other words, $\theta$ is a difference of angles appearing in the definitions of the cones $\mathcal{B}$, $\mathcal{A}$.

ArXiv

--- abstract: 'We derive a novel integral representations of Jacobi polynomials in terms of the Gauss hypergeometric function. Such representation is then used to give the explicit integral representation for the Heat kernel on the quantized Riemann sphere.' address: - '(A.H.) Centre Régional des Métiers de l’Education et de la Formationde kenitra, Morocco' - '(G.A.) Analysis, P.D.E. $\&$ Spectral Geometry, Lab MIA-SI, CeReMAR Department of Mathematics, Faculty of Sciences, P.O. Box 1014 Mohammed V University in Rabat, Morocco' author: - Ali Hafoud - Allal Ghanmi title: Integral representation for Jacobi polynomials and application to heat kernel on quantized sphere --- Introduction ============ Integral representation of orthogonal polynomials have potential applications in several branches of mathematical, physical, statistical and engineering sciences, see e.g. [@Askey1975; @Erdelyi1953; @Rainville71; @Whittaker-Watson1952]. The following one [@DijksmaKoornwinder1971 Theorem 2.2], $$\begin{aligned} \label{IntRepJPol} P_n^{(\alpha, \beta)}(1-2t^2)= c_{\alpha,\beta}^n \int_0 ^1 C_{2n}^{(\alpha+\beta+1)} (tu)(1-u^2)^{\alpha - \frac 12 } du,\end{aligned}$$ is well-known ones for Jacobi polynomials. Above $$c_{\alpha,\beta}^n := \frac{ 2 (-1)^n \Gamma( \alpha+\beta+1) \Gamma(n+\alpha+1)}{\sqrt{\pi} \Gamma(n+\alpha+\beta+1) \Gamma \left( \alpha+ \frac 12 \right) }.$$ In the present paper we provide in Section 2 new integral representations for Jacobi polynomials such as the one involving the product of the Gauss hypergeometric function $_2F_1$ and the Gegenbauer polynomials. Namely we prove $$\begin{aligned} P_\ell^{(n, m)}(\cos(2 \theta)) &= \frac{ 2n!(\ell+m)!}{\pi (\ell+n+m)!} \frac{1}{\cos ^{m}(\theta)} \int_{\theta}^{{\pi}/2} \frac{\sin(u)} { \sqrt {\cos^2(\theta)-\cos^2(u)}} \\& \qquad\qquad \times {_2 F_1}\left( \begin{array}{c} -m , m \\ \frac 12 \end{array} \Big| \frac{\cos (\theta) -\cos (u)}{2 \cos (\theta) }\right) C_{2l+m}^{n+1}(\cos u) du. \nonumber \end{aligned}$$ As immediate application, we give in Section 3 an explicit integral representation of the Heat kernel for the invariant magnetic Laplacian $$\begin{aligned} \label{MagnLap} \Delta_{\nu} = -(1+|z|^2)^2 \frac{\partial^ 2}{\partial z \partial \overline{z}} -\nu (1+|z|^2)\left( z\frac{\partial}{\partial z }-\overline{z}\frac{\partial}{\overline{z}} \right) +\nu^2 |z|^2 \end{aligned}$$ acting on the sections of the $ U(1)$-bundle for the ( quantized) Riemann unit sphere $ S^2$ identified to the extended complex plane $\mathbb{C} \cup \infty $, and describing the Dirac monopole with charge $ q=2\nu$; $\nu>0$, under the action of a constant quantized magnetic field of strength $ \nu \in \mathbb{Z}^+$. For complement, we also provide in Section 4 a new direct proof of which tied up to Dirichlet–Mehler integral and the Christoffel–Darboux formula for Jacobi polynomials. New integral representations of Jacobi polynomials ================================================== We begin with the following result which readily follows by specifying $y=1$ in the Christoffel–Darboux formula for Jacobi polynomials [@Szego1945 Theorem 3.2.2, p. 43] and next making use of the three terms recurrence formula for Jacobi polynomials in [@Askey1975 Eq. (2.17), p. 9] (see also [@Szego1945 Chap. 4] or [@Erdelyi1953 Chap. 10]). \[lemFundamentalRessult1\] The following formula $$\begin{aligned} \sum_{k=0}^{\ell} (2k+\alpha+\beta+1)\frac{\Gamma(k+\alpha+\beta +1)}{\Gamma(k+\beta+1)} P_k^{(\alpha, \beta)}(x)= \frac{\Gamma(\ell+\alpha+\beta +2)}{\Gamma(\ell+\beta +1)}P_\ell^{(\alpha +1, \beta)}(x) \end{aligned}$$ holds true for every $ \alpha > -1/2$, $ \beta > -1/2$ and $-1\leq x<1$. Using Lemma \[lemFundamentalRessult1\], the Dirichlet–Mehler integral for Legendre polynomials [@Fejer1908] (see also [@Askey1975 Eq. (3.1), p. 19]) $$\begin{aligned} \label{IntRepDirichlet--Mehler} P_{\ell}(\cos(2\theta ))=\frac{2}{\pi}\int_{\theta}^{{\pi}/2} \frac{\sin((2\ell+1)u)}{\sqrt{\cos^2(\theta)-\cos^2(u)}} du\end{aligned}$$ as well as the observation $$\begin{aligned} \label{sinsin} \frac{\lambda \sin (\lambda u)}{\sin(u) } = \frac{-1}{\sin(u)}\frac{d }{du}\left(\cos(\lambda u) \right),\end{aligned}$$ we can prove the following \[propFundamentalRessult2\] For every nonnegative integers $n,\ell$ we have $$\begin{aligned} \label{i} P_\ell^{(n, 0)}(\cos(2 \theta) ) &= \frac{2 \ell!}{\pi 2^{n} (\ell+n)!} \int_{\theta}^{{\pi}/2} \frac{\sin(u)}{\sqrt{\cos^2(\theta)-\cos^2(u)}}\\ &\qquad \times \left( \frac{-d}{\sin(u) du}\right) ^n \left( \frac{\sin((2\ell+n+1)u)}{\sin(u)}\right) du.\nonumber \end{aligned}$$ The proof of follows by mathematical induction on $n$. The case of $n=0$ is exactly the Dirichlet–Mehler integral for Legendre polynomials. Next, assume that for $P_k^{(n, 0)}(\cos 2 \theta )$ holds true for given fixed positive integer $n$ and all nonnegative integer $k$. Therefore, making use of Lemma \[lemFundamentalRessult1\] we get $$\begin{aligned} \frac{(\ell+n+1)!}{\ell !} P_\ell^{(n +1, 0)}(\cos(2\theta)) =\sum_{k=0}^{\ell} (2k+n+1)\frac{(k+n )!}{k!} P_k^{(n, 0)}(\cos(2\theta)). \end{aligned}$$ Hence, by induction hypothesis combined with the observation $$\begin{aligned} \label{sinsin} \frac{\lambda \sin (\lambda u)}{\sin(u) } = \frac{-1}{\sin(u)}\frac{d }{du}\left(\cos(\lambda u) \right) , \end{aligned}$$ we obtain $$\begin{aligned} \label{Recn1} \frac{(\ell+n+1)!}{\ell !} &&P_\ell^{(n +1, 0)}(\cos(2\theta)) \\\qquad & =& \frac{1}{2^{n-1}\pi } \int_{\theta}^{{\pi}/2} \frac{\sin(u)}{\sqrt{\cos^2(\theta)-\cos^2(u)}} \left( \frac{-d}{\sin(u) du}\right)^{n+1} \left(S_{\ell,n}(u) \right) du ,\nonumber \end{aligned}$$ with $$\begin{aligned} S_{\ell,z}(u) &:= \sum_{k=0}^{\ell} \cos((2k+z)u) = \frac{1}{2} \left( \frac{\sin (z-1)u}{\sin(u)} + \frac{\sin ((2\ell+z+1)u)}{\sin(u)}\right) \end{aligned}$$ which readily follows by direct computation. Therefore, by taking $z=n+1$ and using the fact $$\begin{aligned} \label{idsin} \left( \frac{-d}{\sin(u) du}\right) ^{n}\left( \frac{\sin (nu)}{\sin(u)} \right) =0 , \end{aligned}$$ we obtain $$\begin{aligned} \label{actionDer} \left( \frac{-d}{\sin(u) du}\right) ^{n+1} \left( S_{\ell,n+1}(u) \right) = \frac{1}{2} \left( \frac{-d}{\sin(u) du}\right) ^{n+1} \left(\frac{\sin ((2\ell+n+2)u)}{\sin(u)}\right). \end{aligned}$$ Substitution of in shows that holds true for rank $n+1$ and for every nonnegative integer $\ell$. This finishes the proof of Lemma \[propFundamentalRessult2\]. The identity is immediate for ${\sin (nu)}/{\sin(u)}$ being a ultraspherical polynomial in $\cos(u)$ of degree $n-1$ (see ). The previous result can be rewritten in terms of ultraspherical polynomials using as well as the well-known fact [@MagnusOberhettingerSoni1966 p. 218] $$\begin{aligned} \label{sinGrn} \frac{\sin (nu)}{\sin(u)}= C_{n-1}^{(1)}(\cos u); \quad n=1,2, \cdots.\end{aligned}$$ \[lemIntJacGegn0\] For every nonnegative integers $n,\ell$ we have $$\begin{aligned} \label{RelJacGegen} P_{\ell}^{(n, 0)}(2t^2-1)= \frac{2\ell!n! }{ \pi (l+n)!} \int_0^{1} \frac{C_{2\ell}^{(n+1)}(tv)}{ \sqrt{1-v^2} } dv. \end{aligned}$$ Recall first that the ultraspherical polynomials satisfy $$\begin{aligned} \label{dnGegen} \frac{d^n}{dx^n}C_{\ell+n}^{(\lambda) } (x)= \frac{2^n\Gamma(\lambda+n)}{\Gamma(\lambda)} C_{\ell}^{(\lambda+n) } (x) . \end{aligned}$$ This can be handled by induction starting from $ \frac{d}{dx}C_{\ell+1}^{(\lambda) } = 2 \lambda C_{\ell}^{(\lambda+1) } $. Then when combined with and the identity , it infers $$\begin{aligned} \left( \frac{-d}{\sin(u) du}\right) ^{n} \left( \frac{\sin ((2\ell+n+1)u)}{\sin(u)} \right) & = \left( \frac{-d}{\sin(u) du}\right) ^{n} \left( C_{2\ell+n}^{(1)} (\cos (u)) \right) \\& = 2^{n} n!C_{2\ell}^{(n+1)} (\cos (u)). \end{aligned}$$ Therefore, from one obtains by means of the changes $ t=\cos (\theta) $ and $v=\cos(u)/t$. This completes the proof. The identity appears as particular case of DijksmaKoornwinder integral representation of Jacobi polynomials given through . However, th can be use to reprove making use of Dirichlet–Mehler integral for the Legendre polynomials. Namely, we claim we have $$\begin{aligned} \label{ii} P_\ell^{(n, m)}(2t^2-1 )= d_{n,m}(\ell) \int_0 ^1 \left( 1-v^2\right) ^{m-\frac 12} C_{2\ell}^{(n+m+1)}(vt) dv , \end{aligned}$$ For every nonnegative integers $m,n,\ell$ such that $ n \geq m$, where $$\begin{aligned} \label{CSTlEM} d_{n,m}(\ell) =: \frac{2^{2m+1} (\ell+m)!m!(n+m)!}{\pi (2m)!(\ell+n+m)!} . \end{aligned}$$ Now, using the hypergeometric representation of ultraspherical polynomials, $$\begin{aligned} \label{GegenGauss} C_{2l}^{(\lambda)}(t) = (-1)^\ell \frac{\Gamma(\lambda+\ell)}{\ell! \Gamma(\lambda) } {_2 F_1}\left( \begin{array}{c} -\ell , \ell+\lambda \\ \frac 12 \end{array} \Big| t^2\right) ,\end{aligned}$$ we can rewrite in terms of the Gauss hypergeometric function $$\begin{aligned} P_{\ell}^{(n,m)}(2t^2-1) =\frac{2 (-1)^\ell (m+\ell)!}{\sqrt{\pi} \ell! \Gamma\left(m+\frac 12 \right) } \int_0^{1} (1-v^2)^{m-1/2} {_2 F_1}\left( \begin{array}{c} -\ell , \ell+n+m+1 \\ \frac 12 \end{array} \Big| t^2v^2\right) dv.\end{aligned}$$ Moreover, we can prove the following \[corIntRepJPGauss\] We have $$\begin{aligned} P_\ell^{(n, m)}(\cos(2 \theta)) &= \frac{ 2n!(\ell+m)!}{\pi (\ell+n+m)!} \frac{1}{\cos ^{m}(\theta)} \int_{\theta}^{{\pi}/2} \frac{\sin(u)} { \sqrt {\cos^2(\theta)-\cos^2(u)}} \\& \qquad\qquad \times {_2 F_1}\left( \begin{array}{c} -m , m \\ \frac 12 \end{array} \Big| \frac{\cos(\theta) -\cos (u)}{2 \cos (\theta) }\right) C_{2l+m}^{n+1}(\cos u) du. \nonumber \end{aligned}$$ An integration by parts starting from , keeping in mind yields $$\begin{aligned} P_{\ell}^{(n,m)}(2t^2-1) &= \frac{ (-1)^m n! d_{n,m}(\ell)}{2^m(m+n)! t^m} \int_0^{1} \frac{d^m}{dv^m}\left( (1-v^2)^{m-1/2} \right) C_{2l+m}^{n+1} (tv) dv, \end{aligned}$$ where $d_{n,m}(\ell)$ stands for the constant in . Now, by Rodrigues formula for Jacobi polynomials, we have $$\begin{aligned} \frac{d^m}{dv^m}\left( (1-v^2)^{m-1/2} \right) & = (-1)^m 2^m m! (1-v^2)^{-1/2} P_m^{(-1/2,-1/2)} (v) \\&= \frac{(-1)^m (2m)!}{2^{m} } (1-v^2)^{-1/2} {_2 F_1}\left( \begin{array}{c} -m , m \\ \frac 12\end{array} \Big| \frac{1-v}{2 }\right), \end{aligned}$$ it follows $$\begin{aligned} P_{\ell}^{(n,m)}(2t^2-1) &= \frac{(2m)!n! d_{n,m}(\ell)}{2^{2m}(m+n)! t^m} \int_0^{1} (1-v^2)^{-1/2} {_2 F_1}\left( \begin{array}{c} -m , m \\ \frac 12\end{array} \Big| \frac{1-v}{2 }\right) C_{2l+m}^{n+1} (tv) dv. \end{aligned}$$ Finally, the change of variables $t=\cos(\theta)$ and $v=\cos(u)/\cos(\theta)$ completes the proof of Theorem \[corIntRepJPGauss\]. Application to Heat kernel on the quantized Riemann sphere $ S^2 $ ================================================================== In the present section, we provide a concrete application of . Indeed, we give the explicit integral representation for the heat kernel $E_{\nu}(t,z,w) $ solving the following Heat problem $$\Delta_{\nu} E_{\nu}(t,z,z_0) =\frac{\partial}{\partial t} E_{\nu}(t,z,z_0) ; \quad, t>0 , \,z,z_0 \in S^2$$ and $$\lim_{t\rightarrow 0} \int_{S^2 }E_{\nu}(t,z,w)f(w)d\mu_{\nu}(w)=f(z) \in {\mathbb{C}}^{\infty}( S^2)$$ for $ \Delta_{\nu}$ in . The concrete spectral analysis of the magnetic Laplacian $\Delta_{\nu}$ on $ S^2$ follows from the one elaborated by Peetre and Zhang in [@PeetreZhang1993] for $$\widetilde{\Delta_{\nu}} = -(1+|z|^2)^2 \frac{\partial^ 2}{\partial z \partial \overline{z}} +2\nu (1+|z|^2) \overline{z}\frac{\partial}{\partial\overline{z}} ,$$ by observing that $\Delta_{\nu}$ and $\widetilde{\Delta_{\nu}}$ are unitary equivalent. In fact, for every sufficiently differential function $$f \in L^2(S^2)=L^2\left( S^2, d\mu\right) ; \quad d\mu(z) := \frac{ dxdy}{\pi (1+|z|^2)^{2}} ,$$ we have $$\Delta_{\nu} f = (1+|z|^2)^{-\nu}\left( \widetilde{\Delta_{\nu}} +\nu \right) \left( (1+|z|^2)^\nu f\right) .$$ Thus, the spectrum of $\Delta_{\nu}$ acting in the Hilbert space $L^2(S^2)$ is purely discrete and consists of an infinite number of eigenvalues $$\lambda_{\nu, m} = =\nu + m(m+2\nu+1) ; \, m=0,1,2, \cdots.$$ Therefore, the spectral decomposition of the Hilbert space $ L^2(S^2)$ in terms of the eigenspaces $$\mathcal{A}^{2,\nu}_\ell=\mathcal{A}^{2,\nu}_\ell(S^2) =\{ \phi:S^2 \rightarrow \mathbb{C} \in L^2(S^2) ; \quad \Delta_{\nu} \phi= \lambda_{\nu, \ell}\phi \}$$ reads $$L^2(S^2) = \bigoplus_{\ell=0}^{+\infty} \mathcal{A}^{2,\nu}_\ell(S^2) .$$ Moreover, the $m$-th eigenspace $\mathcal{A}^{2,\nu}_\ell$ is a finite dimensional vector space with dimension $ 2\ell+2\nu+ 1 $. Moreover, the closed expression of the corresponding reproducing kernel is given in [@PeetreZhang1993 Theorem 1, p. 231]. It can be rewritten as $$\begin{aligned} K^\nu_{m}(z,w) &=\frac{(2\nu+2\ell+1)(1+z\overline{w})^{2\nu}}{(1+|z|^2)^{\nu} (1+|w|^2)^{\nu}} {_2F_1}\left( \begin{array}{c} -\ell, \ell+2\nu+1 \\ 1 \end{array} \Big | \sin^2(d(z,w)) \right), \nonumber \\&=\frac{(2\nu+2\ell+1)(1+z\overline{w})^{2\nu}}{(1+|z|^2)^{\nu} (1+|w|^2)^{\nu}} P^{(0,2\nu)}_\ell(\cos^2(2d(z,w))). \label{repKer2}\end{aligned}$$ where $$d(z,w)= \frac{ |1+z\overline{w}|}{(1+|z|)^2 (1+|w|)^2},$$ thanks to $${_2F_1}\left(\begin{array}{c}-m,1+\alpha+\beta+m\\\alpha+1 \end{array} \Big | \tfrac{1}{2}(1-z)\right)=\frac{m!}{(\alpha+1)_m} P_m^{(\alpha,\beta)}(z).$$ Accordingly, we can provide an expansion series of the heat kernel $E_{\nu}(t,z,z_0) $. \[propHeatexp\] The heat kernel $E_{\nu}(t,z,w)$ has the following asymptotic decomposition $$E_{\nu}(t,z,w)=\frac{(1+z\overline{w})^{2\nu} e^{\nu t}}{(1+|z|^2)^{\nu} (1+|w|^2)^{\nu}} \sum_{\ell=0}^{+\infty} (2l+2\nu+1)e^{-l(l+2\nu+1)t} P_\ell^{(0,2\nu )}(\cos(2d(z,w))) .$$ The proof follows making use of the fact that for given self-adjoint operator with eigenvalues $\lambda_j$ and the corresponding eigenfunctions $\{e_j\}$ is a complete orthonormal system, the heat kernel $E(t,z,z_0)$ of is given by $$E(t,z,z_0)= \sum_{k=0}^\infty e^{- \lambda_k t} e_k(z)\overline{e_k(z_0)} .$$ See [@Davies] for example. Therefore, the expansion in Proposition \[propHeatexp\] readily follows by means of the closed formula of $K^{\nu}_m$ given through since $$\begin{aligned} E_{\nu}(t,z,z_0) &= \sum_{m=0}^{+\infty} e^{- \lambda_{\nu,m} t} \left( \sum_{j=-m}^{m+2\nu} \frac{\phi^\nu_{m, j}(z) \overline{\phi^\nu_{m, j}(z_0)} }{{\left\Vert\phi^\nu_{m, j}\right\Vert}^2_{L^2(S^2)}}\right) \\& = \sum_{m=0}^{+\infty} e^{- \lambda_{\nu,m} t} K^{\nu}_m(z,z_0) . \end{aligned}$$ \[rem\] By taking $\nu=0$, we recover the heat kernel associated to the Laplace–Beltrami operator $\frac{\partial^2}{\partial z \partial {\overline{z}}}$ on the Riemann sphere [@FisherJungsterWilliams1985], $$E_{0}(t;d)= \sum_{l=0}^{+\infty} (2l+1)e^{-l(l+1)t} P_l(cos(2d)) .$$ By means of the Dirichlet–Mehler integral representation for the Legendre polynomials , we can rewrite $ E_{0}(t;d)$ in Remark \[rem\] in terms of the usual theta function $$\theta_{2}(u)= \sum_{l=0}^{+\infty} e^{-l(l+1)t}\cos (2l +1)u$$ as $$\begin{aligned} E_{0}(t,z,w) = \frac{2 }{\pi } \int_{d}^{{\pi}/2} \frac{ \frac{d}{du} \Big( \theta_{2,0}(t,u) \Big)} { \sqrt {\cos^2(d)-\cos^2(u)}} du. \end{aligned}$$ More generally, we prove the following. The explicit real integral representation of the Heat kernel $E_{\nu}(t,z,w)$ for the invariant Laplacian $\Delta_\nu$ on the quantized Riemann sphere $ S^2 $ is given by $$\begin{aligned} E_{\nu}(t,z,w)& = \frac{2(1+z\overline{w})^{2\nu} e^{_\nu t}}{\pi (1+|z|^2)^{\nu} (1+|w|^2)^{\nu} \cos ^{2\nu}(d)} \int_{d}^{{\pi}/2} \frac{ \frac{d}{du} \Big( \theta_{2,\nu}(t,u) \Big)} { \sqrt {\cos^2(d)-\cos^2(u)}} \\& \qquad \qquad \times {_2 F_1}\left( \begin{array}{c} -2\nu , 2\nu \\ \frac 12 \end{array} \Big| \frac{\cos(d) -\cos (u)}{2 \cos(d) }\right) du.\end{aligned}$$ where $ d=d(z,w)$ and $ \theta_{2,\nu}(u)$ is given by $$\begin{aligned} \label{theta2nu} \theta_{2,\nu}(u)= \sum_{l=0}^{+\infty} e^{-l(l+2\nu+1)t}\cos (2l+2\nu +1)u . \end{aligned}$$ The closed integral representation of $ E_{\nu}(t,z,w)$ follows making use of Proposition \[propHeatexp\] as well as the integral representation of Jacobi polynomials given in Theorem \[corIntRepJPGauss\]. Indeed, $$\begin{aligned} E_{\nu}(t,z,w) &= \frac{2(1+z\overline{w})^{2\nu} e^{\nu t}}{\pi (1+|z|^2)^{\nu} (1+|w|^2)^{\nu} \cos ^{2\nu}(d)} \int_{d}^{{\pi}/2} \frac{\sin(u)} { \sqrt {\cos^2(d)-\cos^2(u)}} \\&\qquad \times {_2 F_1}\left( \begin{array}{c} -2\nu , 2\nu \\ \frac 12 \end{array} \Big| \frac{\cos(d) -\cos (u)}{2 \cos(d) }\right) R^\nu_\ell (u) du, \end{aligned}$$ where we have set $$R^\nu_\ell (u) := \sum_{\ell=0}^{+\infty} (2l+2\nu+1)e^{-l(l+2\nu+1)t} C_{2l+2\nu}^{1}(\cos u).$$ Finally, using , we can rewrite $R^\nu_\ell (u)$ in terms of $\theta_{2,\nu}$ in as $$R^\nu_\ell (u) =\frac {1}{\sin(u)} \frac{d}{du} \Big( \theta_{2,\nu}(t,u) \Big) .$$ A new proof of Dijksama-Koornwinder integral representation =========================================================== The integral representation , for Jacobi polynomials in terms of ultraspherical polynomials, appears a specific case of $$\begin{aligned} \label{IntRepProdDK1971} && P_n^{(\alpha, \beta)}(1-2t^2) P_n^{(\alpha, \beta)}(1-2s^2) = \frac{ \Gamma( \alpha+\beta+1) \Gamma(n+\alpha+1)\Gamma(n+\beta+1)}{\pi n!\Gamma( n+\alpha+\beta+1)\Gamma\left( \alpha+\frac 12\right) \Gamma\left( \beta+\frac 12\right) } \\& &\times \int_{-1}^1 \int_{-1}^1 C_{2n}^{(\alpha+\beta+1)}\left( stu+v\sqrt{(1-t^2)(1-s^2)}\right) (1-u^2)^{\alpha -\frac 12}(1-v^2)^{\beta -\frac 12} dudv \nonumber\end{aligned}$$ valid for two fixed complex numbers $\alpha,\beta$ such that $2\Re e (\alpha) > - 1$ and $2\Re e (\beta) > - 1$. The proof of requires special geometrical characterization of $P_n^{(\alpha, \beta)}$ (as invariant spherical harmonics under some orthogonal transformations in high dimensions) and the Laplace’s integral representation obtained by Braaksma and Meulenbeld in [@BraaksmaMeulenbeld1968]. The proof we propose for makes use of Dirichlet–Mehler integral for the Legendre polynomials and is contained in the following fundamental and elementary lemmas. In fact, we need only to establish for nonnegative integers $\alpha=n$ and $\beta =m$. The result for arbitrary complex numbers $\alpha,\beta$ such that $2\Re(\alpha)>-1$ and $2\Re(\beta)>-1$ follows by analytic continuation. We begin by noting that for every real $a$ such that $a\ne1$, we have the identity $$\begin{aligned} \label{fdleid} (1-v^2)^a \frac{\partial}{4t\partial t} \left( C_{k}^{(\lambda)}(tv) \right) &= -\frac{\lambda}{4(a+1)t^2} \frac{\partial}{\partial v} \left( (1-v^2)^{a+1} C_{k-1}^{(\lambda+1)}(tv) \right)\\ & \qquad + \frac{\lambda(\lambda+1)}{2(a+1)} (1-v^2)^{a+1} C_{k-2}^{(\lambda+2)}(tv) .\nonumber \end{aligned}$$ This is easy to handle by observing that $$\frac{\partial}{\partial t} \left( C_{k}^{(\lambda)}(tv) \right) = \frac{v}{t} \frac{\partial}{\partial v} \left( C_{k}^{(\lambda)}(tv) \right)$$ and next using the well-established facts $f'g'= (fg')'-fg"$ and $ \frac{d}{dx}C_{\ell+1}^{(\lambda) } = 2 \lambda C_{\ell}^{(\lambda+1) }$. Therefore, we get $$\begin{aligned} \label{fdleid2} \int_0^{1} (1-v^2)^a \frac{\partial}{4t\partial t} \left( C_{2\ell}^{(\lambda)}(tv) \right) dv = \frac{\lambda(\lambda+1)}{2(a+1)} \int_0^{1} (1-v^2)^{a+1} C_{2\ell-2}^{(\lambda+2)}(tv) dv \end{aligned}$$ for $C_{2\ell-1}^{(\lambda)}(0)=0$. More generally, an inductive reasoning making use of gives rise to $$\int_0^{1} (1-v^2)^a \left( \frac{\partial}{4t\partial t} \right) ^m \left( C_{2\ell}^{(\lambda)}(tv) \right) dv = d_{a,\lambda}(n) \int_0^{1} (1-v^2)^{a+m} C_{2\ell-2m}^{(\lambda+2m)}(tv) dv,$$ for some constant $d_{a,\lambda}(n)$ depending only in $a$, $\lambda$ and $n$. Now, by taking $a=-1/2$ and $\lambda=n-m+1$ with $n\geq m$, and using the explicit expression of the $m$-th derivative formula for the Jacobi polynomials [@Rainville71 p. 260] $$\left( \frac{d}{dx} \right) ^m P_{\ell +m}^{(n,0)}(x)=\frac{(\ell+n+2m)!}{2^m (n+m+\ell)!} P_{\ell}^{(n+m,m)}(x) ,$$ as well as Lemma \[lemIntJacGegn0\], we get $$\begin{aligned} \label{mthder} P_{\ell}^{(n,m)}(2t^2-1) &= \frac{2^m (\ell+n)!}{(\ell+n+m)!} \left( \frac{d}{4tdt}\right)^m P_{\ell+m}^{(n-m,0)}(2t^2-1) \\&\stackrel{\eqref{RelJacGegen}}{=} \widetilde{s_{n,m}(\ell)} \int_0^{1} \left( \frac{\partial}{4t\partial t}\right)^m \left( (1-v^2)^{-1/2} C_{2(\ell+m)}^{(n-m+1)}(tv) \right) dv \\&= s_{n,m}(\ell) \int_0^{1} (1-v^2)^{m-1/2} C_{2\ell}^{(n+m+1)}(tv) dv \end{aligned}$$ for every nonnegative integers $n\geq m$. The involved constant $s_{n,m}(\ell) $ is given by $$s_{n,m}(\ell) := \frac{2(n+m)! (m+\ell)!}{\sqrt{\pi} (n+m+\ell)! \Gamma\left(m+\frac 12 \right) }$$ and can be verified by taking $t=0$, keeping in mind the specific values of $$C_{2\ell}^{(\lambda)}(0)=(-1)^{\ell} \frac{\Gamma(\lambda+\ell)}{\ell!\Gamma(\lambda)},$$ $$\int_0^1(1-v^2)^{\alpha-1/2} dv =\frac{\sqrt{\pi} \Gamma\left( \alpha +\frac 12\right) }{\Gamma(\alpha+1)},$$ and $$P_{\ell}^{(\alpha,\beta)}(-1)=(-1)^\ell \frac{\Gamma(\beta+\ell+1)}{\ell!\Gamma(\beta+1)}.$$ This proves . Using the symmetry relation [@Askey1975 Eq. (2.13), p. 8] $$P_\ell^{(\alpha,\beta)}(-x) =(-1)^nP_\ell^{(\beta,\alpha)}(x),$$ we recover . One recovers Mehler’s form of Dirichlet’s integral for Legendre polynomials by taking $ \alpha =\beta=0$ in and making specific change of variables. [99]{} Askey R. Orthogonal polynomials and special functions. Society for Industrial and Applied Mathematics, Philadelphia, Pa., 1975. Erdélyi A., Magnus W., Oberhettinger F., Tricomi F.G. Higher Transcendental Functions, vol. 2, McGraw-Hill, New York, 1953. Rainville E.D. Special functions. Chelsea Publishing Co., Bronx, N.Y.; 1960. Whittaker E. T., Watson G.N., A Course of Modern Analysis, 4th ed., Cambridge University Press, Cambridge, 1952. Dijksma A., Koornwinder T.H., Spherical harmonics and the product of two Jacobi polynomials. Nederl. Akad. Wetensch. Proc. Ser. A 74=Indag. Math. 33 (1971), 191–196. Hafoud A., Intissar A., Représentation intégrale du noyau de la chaleur sur l’espace projectif complexe. C.R.Acad. Sci.Paris. Ser 1 335 (2002) 871-876 Braaksma B.L.J., Meulenbeld B. Jacobi polynomials as spherical harmonics. Nederl. Akad. Wetensch. Proc. Ser. A 71=Indag. Math. 30 1968;38:384–389. Fejer L., Sur le développement d’une fonction arbitraire suivant les fonctions de Laplace, C. R. Acad. Sci. Paris, 146 (1908), pp. 224-225. Szegö G. Orthogonal Polynomials, Colloquium Publications, vol. 23, 3rd ed., American Mathematical Society, Providence, R.I., 1975. Magnus W., Oberhettinger F., Soni R.P., Formulas and Theorems in the Special Functions of Mathematical Physics. Springer -Verlag, Berlin, 1966 Peetre J., Zhang G. Harmonic analysis on the quantized Riemann sphere. Internat. J. Math. Math. Sci. 1993 2;16:225–243. Davies E.B., [*Heat kernels and spectral theory*]{}. Cambridge Tracts in Mathematics, 92. Cambridge University Press: Cambridge; 1989. Fisher H.R.., Jungster J.J, Williams F.J., The Heat kernel on the two-sphere. J. Math.Anal . Appl.112 (1985) 328-334

ArXiv

--- abstract: 'A quantum control landscape is defined as the physical objective as a function of the control variables. In this paper the control landscapes for two-level open quantum systems, whose evolution is described by general completely positive trace preserving maps (i.e., Kraus maps), are investigated in details. The objective function, which is the expectation value of a target system operator, is defined on the Stiefel manifold representing the space of Kraus maps. Three practically important properties of the objective function are found: (a) the absence of local maxima or minima (i.e., false traps); (b) the existence of multi-dimensional sub-manifolds of optimal solutions corresponding to the global maximum and minimum; and (c) the connectivity of each level set. All of the critical values and their associated critical sub-manifolds are explicitly found for any initial system state. Away from the absolute extrema there are no local maxima or minima, and only saddles may exist, whose number and the explicit structure of the corresponding critical sub-manifolds are determined by the initial system state. There are no saddles for pure initial states, one saddle for a completely mixed initial state, and two saddles for partially mixed initial states. In general, the landscape analysis of critical points and optimal manifolds is relevant to explain the relative ease of obtaining good optimal control outcomes in the laboratory, even in the presence of the environment.' author: - | Alexander Pechen$^1$[^1], Dmitrii Prokhorenko$^2$, Rebing Wu$^1$\ and Herschel Rabitz$^1$[^2] title: 'Control landscapes for two-level open quantum systems' --- $^1$ Department of Chemistry, Princeton University, Princeton, New Jersey 08544, USA $^2$ Institute of Spectroscopy, Troitsk, Moscow Region 142190, Russia Introduction ============ A common goal in quantum control is to maximize the expectation value of a given target operator by applying a suitable external action to the system. Such an external action often can be realized by a tailored coherent control field steering the system from the initial state to a target state, which maximizes the expectation value of the target operator [@R; @R0; @R1; @R2; @RZ; @R3; @R4; @R5; @alessandro07]. Tailored coherent fields allow for controlling Hamiltonian aspects (i.e., unitary dynamics) of the system evolution. Another form of action on the system could be realized by tailoring the environment (e.g., incoherent radiation, or a gas of electrons, atoms, or molecules) to induce control through non-unitary system dynamics [@ice]. In this approach the control is the suitably optimized, generally non-equilibrium and time dependent distribution function of the environment; the optimization of the environment would itself be attained by application of a proper external action. Combining such incoherent control by the environment (ICE) with a tailored coherent control field provides a general tool for manipulating both the Hamiltonian and dissipative aspects of the system dynamics. A similar approach to incoherent control was also suggested in [@Romano] where, in difference with [@ice], finite-level ancilla systems are used as the control environment. The initial state of the field and the interaction Hamiltonian as the parameters for controlling non-unitary dynamics was also suggested in [@acim]. Non-unitary controlled quantum dynamics can also be realized by using as an external action suitably optimized quantum measurements which drive the system towards the desired control goal [@qm1; @qm2; @roa1; @qm; @feng]. General mathematical definitions for the controlled Markov dynamics of quantum-mechanical systems are formulated in [@belavkin]. In this paper we consider the most general physically allowed transformations of states of quantum open systems, which are represented by completely positive trace preserving maps (i.e., Kraus maps) [@CPmap; @AL; @NiCh; @breuer]. A typical control problem in this framework is to find, for a given initial state of the system, a Kraus map which transforms the initial state into the state maximizing the expected value $\langle\Theta\rangle$ of a target operator $\Theta$ of the system. Practical means to find such optimal Kraus maps in the laboratory could employ various procedures such as adaptive learning algorithms [@R1; @GA], which are capable of finding an optimal solution without detailed knowledge of the dynamics of the system. Kraus maps can be represented by matrices satisfying an orthogonality constraint (see Sec. II), which can be naturally parameterized by points in a Stiefel manifold [@Stiefel], and then various algorithms may be applied to perform optimization over the Stiefel manifold (e.g., steepest descent, Newton methods, etc. adapted for optimization over Stiefel manifolds) [@EdArSm; @Manton02]. The quantum control landscape is defined as the objective expectation value $\langle\Theta\rangle$ as a function of the control variables. The efficiency of various search algorithms (i.e., employed either directly in the laboratory or in numerical simulations) for finding the minimum or maximum of a specific objective function can depend on the existence and nature of the landscape critical points. For example, the presence of many local minima or maxima (i.e., false traps) could result in either permanent trapping of the search or possibly dwelling for a long time in some of them (i.e., assuming that the algorithm has the capability of extricating the search from a trap) thus lowering the search efficiency. In such cases stopping of an algorithm at some solution does not guarantee that this solution is a global optimum, as the algorithm can end the search at a local maximum of the objective function. [*A priori*]{} information about absence of local maxima could be very helpful in such cases to guarantee that the search will be stopped only at a global optimum solution. This situation makes important the investigation of the critical points of the control landscapes. Also, in the laboratory, evidence shows that it is relatively easy to find optimal solutions, even in the presence of an environment. Explanation of this fact similarly can be related with the structure of the control landscapes for open quantum systems. The critical points of the landscapes for closed quantum systems controlled by unitary evolution were investigated in [@U1; @U2; @U3; @U4; @U5], where it was found that there are no sub-optimal local maxima or minima and only saddles may exist in addition to the global maxima and minima. In particular, it was found that for a two-level system prepared initially in a pure state the landscape of the unitary control does not have critical points except for global minima and maxima. The capabilities of unitary control to maximize or minimize the expectation value of the target operator in the case of mixed initial states are limited, since unitary transformations can only connect states (i.e., density matrices) with the same spectrum. In going beyond the latter limitations, the dynamics may be extended to encompass non-unitary evolution by directing the controls to include the set of Kraus maps (i.e., dual manipulation of the system and the environment). Quantum systems which admit arbitrary Kraus map dynamics are completely controllable, since for any pair of states there exists a Kraus map which transforms one into the another [@rong]. In this paper the analysis of the landscape critical points is performed for two-level quantum systems controlled by Kraus maps. It is found that the objective function does not have sub-optimal local maxima or minima and only saddles may exist. The number of different saddle values and the structure of the corresponding critical sub-manifolds depend on the system initial state. For pure initial states the landscape has no saddles; for a completely mixed initial state the landscape has one saddle value; for other (i.e., partially mixed) initial states the landscape has two saddle values. For each case we explicitly find all critical sub-manifolds and critical values of the objective as functions of the Stokes vector of the initial density matrix. An investigation of the landscapes for multi-level open quantum systems with a different method may also be performed [@rw]. The absence of local minima or maxima holds also in the general case although an explicit description of the critical manifolds is difficult to provide for multi-level systems. The absence of false traps practically implies the relative ease of obtaining good optimal solutions using various search algorithms in the laboratory, even in the presence of an environment. It should be noted that the property of there being no false traps relies on the assumption of the full controllability of the system, i.e., assuming that an arbitrary Kraus map can be realized. Restrictions on the set of available Kraus maps can result in the appearance of false traps thus creating difficulties in the search for optimal solutions. Thus, it is important to consider possible methods for engineering arbitrary Kraus type evolution of a controlled system. One method is to put the system in contact with an ancilla and implement, on the coupled system, specific unitary evolution whose form is determined by the structure of the desired Kraus map [@ref1] (see also Sec. II). Lloyd and Viola proposed another method of engineering arbitrary Kraus maps, based on the combination of coherent control and measurements [@lloyd]. They show that the ability to perform a simple single measurement on the system together with the ability to apply coherent control to feedback the measurement results allows for enacting arbitrary Kraus map evolution at a finite time. A level set of the objective function is defined as the set of controls which produce the same outcome value for $\langle\Theta\rangle$. We investigate connectivity of the level sets of the objective functions for open quantum systems and show that each level set is connected, including the one which corresponds to the global maximum/minimum of the objective function. Connectivity of a level set implies that any two solutions from the same level set can be continuously mapped one into another via a pathway entirely passing through this level set. The proof of the connectivity of the level sets is based on a generalization of Morse theory. Experimental observations of level sets for quantum control landscapes can be practically performed, as it was recently demonstrated for control of nonresonant two-photon excitations [@roslund06].  In summary, the main properties of control landscapes for open quantum systems are: (a) the absence of false traps; (b) the existence of multi-dimensional sub-manifolds of global optimum solutions, and (c) the connectivity of each level set. The proof of the properties (a)–(c) is provided in the next sections for the two-level case. Figure \[fig1\] illustrates the properties (a), (b), and connectivity of the manifold of global maximum solutions; the figure does not serve to illustrate other properties such as connectivity of each level set. It is evident that the function drawn on figure \[fig1\] does not have local minima or maxima and the set of solutions for the global maximum is a connected sub-manifold (a curve in this case). A simple illustration is chosen for the figure since an exact objective function for an $N$-level quantum system depends on $D=2N^4-2N^2$ real variables (such that $D=24$ for $N=2$) and therefore can not be drawn. The present analysis is performed in the kinematic picture which uses Kraus maps to represent evolution of quantum open systems. An important future task is to investigate the structure of the control landscape in the dynamical picture, which can be based on the use of various dynamical master equations to describe the dynamics of quantum open systems [@breuer; @spohn; @spohn2; @dumcke; @APV; @ALV]. Such analysis may reveal landscape properties for quantum open systems under (possibly, restricted) control through manipulation by a specific type of the environment (e.g., incoherent radiation). In addition to optimizing expected value of a target operator, a large class of quantum control problems includes generation of a predefined unitary (e.g., phase or Hadamard) [@NiCh] or a non-unitary [@Tarasov] quantum gate (i.e., a quantum operation). This class of control problems is important for quantum computation and in this regard a numerical analysis of the problem of optimal controlled generation of unitary quantum gates for two-level quantum systems interacting with an environment is available [@Grace1; @Grace2]. Although the assumption of complete positivity of the dynamics of open quantum systems used in the present analysis is a generally accepted requirement, some works consider dynamics of a more general form [@nonCP; @nonCP2]. Such more general evolutions may result in different controllability and landscape properties. For example, for a two-level open quantum system positive and completely positive dynamics may have different accessibility properties [@romano05]. In this regard it would be interesting to investigate if such different types of the dynamics have distinct essential landscape properties. In Sec. \[sec1\] the optimal control problem for a general $N$-level open quantum system is formulated. Section \[sec2\] reduces the consideration to the case of a two-level system. In Sec. \[sec3\] a complete description is given of all critical points of the control landscape. The connectivity of the level sets is investigated in Sec. \[sec4\]. Formulation for an $N$-level system {#sec1} =================================== Let ${\cal M}_N$ be the linear space of $N\times N$ complex matrices. The density matrix $\rho$ of an $N$-level quantum system is a positive component in ${\cal M}_N$, $\rho\ge 0$, with unit trace, ${{\rm Tr}}\rho=1$ (Hermicity of $\rho$ follows from its positivity). Physically allowed evolution transformations of density matrices are given by completely positive trace preserving maps (i.e., Kraus maps) in ${\cal M}_N$. A linear Kraus map $\Phi:{\cal M}_N\to{\cal M}_N$ satisfies the following conditions [@CPmap]: - [Complete positivity. Let $\mathbb I_n$ be the identity matrix in ${\cal M}_n$. Complete positivity means that for any integer $n\in\mathbb N$ the map $\Phi\otimes\mathbb I_n$ acting in the space ${\cal M}_N\otimes {\cal M}_n$ is positive.]{} - [Trace preserving: $\forall{\rho}\in{\cal M}_{N}$, ${{\rm Tr}}\Phi(\rho)={{\rm Tr}}\rho$.]{} Any Kraus map $\Phi$ can be decomposed (non-uniquely) in the Kraus form [@choi; @alessandro07]: $$\label{eq4} \Phi(\rho)=\sum\limits_{l=1}^M K_l\rho K^\dagger_l,$$ where the Kraus operators $K_l$ satisfy the relation $\sum_{l=1}^M K^\dagger_lK_l=\mathbb I_N$. For an $N$-level quantum system it is sufficient to consider at most $M=N^2$ Kraus operators [@choi]. Let ${\cal H}_1=\mathbb C^N$ be the Hilbert space of the system under control. An arbitrary Kraus map of the form (\[eq4\]) can be realized by coupling the system to an ancilla system characterized by the Hilbert space ${\cal H}_2=\mathbb C^M$, and generating a unitary evolution operator $U$ acting in the Hilbert space of the total system ${\cal H}={\cal H}_1\otimes{\cal H}_2$ as follows [@ref1]. Choose in ${\cal H}_2$ a unit vector $|0\rangle$ and an orthonormal basis $|e_i\rangle$, $i=1,\dots,M$. For any $|\psi\rangle\in {\cal H}_1$ let $U (|\psi\rangle\otimes|0\rangle)=\sum_{i=1}^M K_i|\psi\rangle\otimes |e_i\rangle$. Such an operator can be extended to a unitary operator in $\cal H$ and for any $\rho$ one has $\Phi(\rho)={\rm Tr}_{{\cal H}_2}\,\{ U(\rho\otimes |0\rangle\langle 0|)U^\dagger\}$. Therefore the ability to dynamically create, for example via coherent control, an arbitrary unitary evolution of the system and ancilla allows for generating arbitrary Kraus maps of the controlled system. Let $\rho_0$ be the initial system density matrix. A typical optimization goal in quantum control is to maximize the expectation value $J=\langle\Theta\rangle$ of a target Hermitian operator $\Theta$ over an admissible set of dynamical transformations of the system density matrices. For coherent unitary control this expectation value becomes $$J[U]={{\rm Tr}}[U\rho_0 U^\dagger\Theta]$$ where $U=U(t,t_0)$ is a unitary matrix, $UU^\dagger=U^\dagger U=\mathbb I_N$, which describes the evolution of the system during the control period from the initial time $t_0$ until some final time $t$ and implicitly incorporates the action of the coherent control field on the system. In the present paper we consider general non-unitary controlled dynamics such that the controls are Kraus maps, for which the parametrization by Kraus operators is used. The corresponding objective function specifying the control landscape has the form $$\label{eq8} J[K_1,\dots,K_M]={{\rm Tr}}\Bigl[\sum\limits_{l=1}^M K_l\rho_0 K^\dagger_l\Theta\Bigr]$$ where the Kraus operators $\{K_l\}=\{K_l(t,t_0)\}$ describe evolution of the open quantum system from an initial time $t_0$ until some final time $t$. The control goal is to maximize the objective function over the set of all Kraus operators $K_1,\dots,K_M$ satisfying $\sum_{l=1}^M K^\dagger_lK_l=\mathbb I_N$, thereby forming a constrained optimization problem. Let $\mathbb F$ be a field of real or complex numbers, i.e., $\mathbb F=\mathbb R$ or $\mathbb F=\mathbb C$. A Stiefel manifold over $\mathbb F$, denoted $V_k(\mathbb F^n)$, is the set of all orthonormal $k$-frames in $\mathbb F^n$ (i.e., the set of ordered $k$-tuples of orthonormal vectors in $\mathbb F^n$). The case $\mathbb F=\mathbb R$ (respectively, $\mathbb F=\mathbb C$) corresponds to a real (complex) Stiefel manifold. Let $K$ be the $N\times (NM)$ matrix defined as $K=(K_1^{\rm T}\dots K_M^{\rm T})$, where $K_l^{\rm T}$ is the transpose of matrix $K_l$ and $M$ is the number of Kraus operators. Consider $N$ vectors $X_1,\dots, X_N\in\mathbb C^{NM}$ with components $(X_i)_j=K_{ij}$, i.e., vector $X_i$ is the $i$-th row of the matrix $K$. The constraint $\sum_{l=1}^M K^\dagger_lK_l=\mathbb I_N$ in terms of the vectors $X_1,\dots, X_N$ takes the form $\langle X_i,X_j\rangle={\delta}_{ij}$, where ${\delta}_{ij}$ is the Kronecker delta symbol. This constraint defines the complex Stiefel manifold $V_N(\mathbb C^{NM})$. Therefore optimization of the objective function $J[K_1,\dots,K_M]$ defined by Eq. (\[eq8\]) can be formulated as optimization over the complex Stiefel manifold $V_N(\mathbb C^{NM})$. Two-level system {#sec2} ================ In the following we consider the case of a two-level system in detail. Any density matrix of a two-level system can be represented as $$\rho=\frac{1}{2}[1+\langle{{{\bf w}},\sigma}\rangle]$$ where ${\bf\sigma}=(\sigma_1,\sigma_2,\sigma_3)\equiv(\sigma_x,\sigma_y,\sigma_z)$ is the vector of Pauli matrices and ${\bf w}\in\mathbb R^3$ is the Stokes vector, $\|{\bf w}\|\le 1$. Thus, the set of density matrices can be identified with the unit ball in $\mathbb R^3$, which is known as the Bloch sphere. Any Kraus map $\Phi$ on ${\cal M}_2$ can be represented using at most four Kraus operators $$K_l=\left(\begin{array}{cc} x_{l1} & x_{l3}\\ x_{l2} & x_{l4} \end{array}\right),\qquad l=1,2,3,4$$ as $\Phi(\rho)=\sum_{l=1}^4K_l\rho K_l^\dagger$, where the Kraus operators satisfy the constraint $$\label{eq1} \sum\limits_{l=1}^4 K_l^\dagger K_l=\mathbb I_2$$ Let $\rho_0$ be the initial system density matrix with Stokes vector $\bf w=(\alpha,\beta,\gamma)$, where $\|{{\bf w}}\|^2=\alpha^2+\beta^2+\gamma^2\le 1$, and let $\Theta$ be a Hermitian target operator. The objective functional for optimizing the expectation value of $\Theta$ has the form $J[K_1,K_2,K_3,K_4;\rho_0,\Theta]=\sum_{l=1}^4{{\rm Tr}}[K_l\rho_0 K^\dagger_l\Theta]$. The control goal is to find all quadruples of Kraus operators $(K_1,K_2,K_3,K_4)$ which maximize (or minimize, depending on the control goal) the objective functional $J$. The goal of the landscape analysis is to characterize all critical points of $J[K_1,K_2,K_3,K_4]$, including local extrema, if they exist. The analysis for an arbitrary $2\times2$ Hermitian matrix $\Theta$ can be reduced to the case $$\Theta_0=\left(\begin{array}{cc} 1 & 0\\ 0 & 0 \end{array}\right)$$ which we will consider in the sequel. This point follows, as an arbitrary Hermitian operator $\Theta\in{\cal M}_2$ has two eigenvalues ${\lambda}_1$ and ${\lambda}_2$ and can be represented in the basis of its eigenvectors as $$\Theta=\left(\begin{array}{cc} {\lambda}_1 & 0\\ 0 & {\lambda}_2 \end{array}\right)$$ where ${\lambda}_1\ge{\lambda}_2$. One has $\Theta=({\lambda}_1-{\lambda}_2)\Theta_0+{\lambda}_2\mathbb I_2$ and $$\begin{aligned} J[K_1,K_2,K_3,K_4;\rho_0,\Theta]&=&\sum\limits_{l=1}^4{{\rm Tr}}[K_l\rho_0 K^\dagger_l\Theta]\\ &=&({\lambda}_1-{\lambda}_2)\sum\limits_{l=1}^4{{\rm Tr}}[K_l\rho_0 K^\dagger_l\Theta_0]+{\lambda}_2\sum\limits_{l=1}^4{{\rm Tr}}[K_l\rho_0 K^\dagger_l]\\ &=&({\lambda}_1-{\lambda}_2)J[K_1,K_2,K_3,K_4;\rho_0,\Theta_0]+{\lambda}_2\end{aligned}$$ Therefore, the objective function for a general observable operator $\Theta$ depends linearly on the objective function defined for $\Theta_0$. We denote $J[K_1,K_2,K_3,K_4;{\bf w}]:=J[K_1,K_2,K_3,K_4;\rho_0,\Theta_0]$. In the trivial case $\Theta=\mathbb I_2$ the landscape is completely flat and no further analysis is needed. The critical points of the objective function landscape {#sec3} ======================================================= The Kraus operators for a two-level system can be parameterized by a pair of vectors $X,Y\in\mathbb C^8=\mathbb C^4\oplus\mathbb C^4$ of the form $X=u_1\oplus v_1$ and $Y=u_2\oplus v_2$, where $u_1=(x_{11},x_{21},x_{31},x_{41})$, $v_1=(x_{12},x_{22},x_{32},x_{42})$, $u_2=(x_{13},x_{23},x_{33},x_{43})$, and $v_2=(x_{14},x_{24},x_{34},x_{44})$. The objective function in terms of these vectors has the form $$\label{eq7} J[u_1,u_2,v_1,v_2;{\bf w}]=\frac{1}{2}\Bigl[(1+{\gamma})\|u_1\|^2+(1-{\gamma})\|u_2\|^2+2{\rm Re}[z_0\langle u_1,u_2\rangle]\Bigr]$$ where $z_0=\alpha-{{\rm i}}\beta$, $\langle\cdot,\cdot\rangle$ and $\|\cdot\|$ denote the standard inner product and the norm in $\mathbb C^N$ (here the numbers $\alpha,\beta,\gamma$ are the components of the Stokes vector ${{\bf w}}=(\alpha,\beta,\gamma)$ of the initial density matrix $\rho_0$, see Sec. \[sec2\]). The constraint (\[eq1\]) in terms of the vectors $X$ and $Y$ has the form $\|X\|=\|Y\|=1$, $\langle X,Y\rangle=0$ and determines the Stiefel manifold ${\cal M}=V_2(\mathbb C^8)$. The matrix constraint (\[eq1\]) in terms of the vectors $u_i$ and $v_i$ has the form $$\begin{aligned} \Phi_1(u_1,u_2,v_1,v_2)&:=&\|u_1\|^2+\|v_1\|^2-1=0\label{c1}\\ \Phi_2(u_1,u_2,v_1,v_2)&:=&\|u_2\|^2+\|v_2\|^2-1=0\label{c2}\\ \Phi_3(u_1,u_2,v_1,v_2)&:=&\langle u_1,u_2\rangle+\langle v_1,v_2\rangle=0\label{c3}\end{aligned}$$ If $z_0\ne 0$, then the objective function is diagonalized by introducing new coordinates $(\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2)$ in $\mathbb C^{16}$ according to the formulas $$\begin{aligned} u_1&=&\mu \tilde u_1-\nu \tilde u_2,\qquad u_2=\frac{z_0^*}{|z_0|}\nu \tilde u_1+\frac{z_0^*}{|z_0|}\mu \tilde u_2\label{nc1}\\ v_1&=&\mu \tilde v_1-\nu \tilde v_2,\qquad v_2=\frac{z_0^*}{|z_0|}\nu \tilde v_1+\frac{z_0^*}{|z_0|}\mu \tilde v_2\label{nc2}\end{aligned}$$ where $\mu=|z_0|/\sqrt{2\|{{\bf w}}\|(\|{{\bf w}}\|-{\gamma})}$ and $\nu=|z_0|/\sqrt{2\|{{\bf w}}\|(\|{{\bf w}}\|+{\gamma})}$. The objective function in these coordinates has the form $$\label{eq9} J[x;{\bf w}]={\lambda}_+\|\tilde u_1\|^2+{\lambda}_-\|\tilde u_2\|^2$$ where $x=(\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2)\in{\cal M}$ and ${\lambda}_\pm=(1\pm\|{{\bf w}}\|)/2$. If $z_0=0$ and ${\gamma}\ge 0$ (resp., ${\gamma}<0$), then the objective function (\[eq7\]) has the form (\[eq9\]) with $\tilde u_i=u_i, \tilde v_i=v_i$ for $i=1,2$ (resp., $\tilde u_1=u_2,\tilde u_2=u_1,\tilde v_1=v_2,\tilde v_2=v_1$). The constraints (\[c1\])–(\[c3\]) in the new coordinates have the same form $\Phi_i(\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2)=0$ for $i=1,2,3$. Let ${\bf w}=(\alpha,\beta,\gamma)\in\mathbb R^3$ be a real vector such that $\|{\bf w}\|\le 1$ and let ${\lambda}_\pm=(1\pm\|{{\bf w}}\|)/2$. For any such ${{\bf w}}$, the global maximum and minimum values of the objective function $J[\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2;{\bf w}]={\lambda}_+\|\tilde u_1\|^2+{\lambda}_-\|\tilde u_2\|^2$ are $$\begin{aligned} \min\limits_{(\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2)\in{\cal M}}J[\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2;{\bf w}]&=&0\\ \max\limits_{(\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2)\in{\cal M}}J[\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2;{\bf w}]&=&1.\end{aligned}$$ The critical sub-manifolds and other critical values of $J$ in ${\cal M}$ are the following: [**Case 1.**]{} ${{\bf w}}=0$ (the completely mixed initial state). The global minimum sub-manifold is ${\cal M}^{(0,0,0)}_{\rm min}=\{x\in{\cal M}\,|\,\tilde u_1=\tilde u_2=0\}$. The global maximum sub-manifold is ${\cal M}^{(0,0,0)}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=\tilde v_2=0\}$. The objective function has one saddle value $J=1/2$ with the corresponding critical sub-manifold ${\cal M}^{(0,0,0)}_{\rm saddle}=\{x\in{\cal M}\,|\, \tilde u_2=z\tilde u_1,\, \tilde v_1=-z^*\tilde v_2,\, z\in\mathbb C\}\bigcup\{x\in{\cal M}\,|\,\tilde u_1=\tilde v_2=0\}$. The Hessian of $J$ at any point at ${\cal M}^{(0,0,0)}_{\rm saddle}$ has $\nu_+=6$ positive, $\nu_-=6$ negative, and $\nu_0=16$ zero eigenvalues. [**Case 2.**]{} $0<\|{{\bf w}}\|<1$ (a partially mixed initial state). The global minimum sub-manifold is ${\cal M}^{{\bf w}}_{\rm min}=\{x\in{\cal M}\,|\,\tilde u_1=\tilde u_2=0\}$. The global maximum sub-manifold is ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=\tilde v_2=0\}$. The objective function has two saddle values: $$\label{s1} J_\pm({{\bf w}})=\frac{1\pm\|{\bf w}\|}{2}={\lambda}_\pm.$$ The corresponding critical sub-manifolds are ${\cal M}^{\bf w}_-=\{x\in{\cal M}\,|\, \tilde u_1=\tilde v_2=0\}$ and ${\cal M}^{\bf w}_+=\{x\in{\cal M}\,|\, \tilde u_2=\tilde v_1=0\}$. The Hessian of $J$ at any point at ${\cal M}^{{\bf w}}_-$ (resp., ${\cal M}^{{\bf w}}_+$) has $\nu_+=8$ positive, $\nu_-=6$ negative (resp., $\nu_+=6$ positive, $\nu_-=8$ negative), and $\nu_0=14$ zero eigenvalues. [**Case 3.**]{} $\|{{\bf w}}\|=1$ (a pure initial state). The global minimum sub-manifold is ${\cal M}^{{\bf w}}_{\rm min}=\{x\in{\cal M}\,|\,\tilde u_1=0\}$. The global maximum sub-manifold is ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=0\}$. The objective function has no saddles. [**Proof.**]{} The objective function has the form $J=\rho_{11}$, where $\rho_{11}$ is the diagonal matrix element of the density matrix. Therefore $0\le J\le1$ and the value $J=0$ (resp., $J=1$) corresponds to the global minimum (resp., maximum). The constraints can be included in the objective function (\[eq9\]) by adding the term $\Phi[\tilde u,\tilde v,\eta]=\eta_1\Phi_1+\eta_2\Phi_2+2{\rm Re}\,[\eta_3^*\Phi_3]$, where the two real and one complex Lagrange multipliers $\eta_1,\eta_2$, and $\eta_3$ correspond to the two real and one complex valued constraints $\Phi_1,\Phi_2$, and $\Phi_3$, respectively. Critical points of the function $J$ on the manifold $\cal M$ are given by the solutions of the following Euler-Lagrange equations for the functional $\widetilde{J}[\tilde u,\tilde v,{\lambda}]=J[\tilde u,\tilde v]+\Phi[\tilde u,\tilde v,\eta]$: $$\begin{aligned} 0=\nabla_{\tilde u^*_1} \widetilde{J}\Rightarrow\qquad 0&=&({\lambda}_++\eta_1)\tilde u_1+\eta_3\tilde u_2\label{e1}\\ 0=\nabla_{\tilde u^*_2} \widetilde{J}\Rightarrow\qquad 0&=&\eta^*_3\tilde u_1+({\lambda}_-+\eta_2)\tilde u_2\label{e2}\\ 0=\nabla_{\tilde v^*_1} \widetilde{J}\Rightarrow\qquad 0&=&\eta_1\tilde v_1+\eta_3\tilde v_2\label{e3}\\ 0=\nabla_{\tilde v^*_2} \widetilde{J}\Rightarrow\qquad 0&=&\eta^*_3\tilde v_1+\eta_2\tilde v_2\label{e4}\end{aligned}$$ where $\tilde u_1,\tilde u_2,\tilde v_1,\tilde v_2$ satisfy the constraints (\[c1\])–(\[c3\]). The proof of the theorem is based on the straightforward solution of the system (\[e1\])–(\[e4\]). The case 2 will be considered first, followed by the cases 1 and 3. [**Case 2.**]{} $0<\|{{\bf w}}\|<1$. Consider in $\cal M$ the open subset $\mathcal{O}_1=\{x \in \mathcal{M}\;|\;\tilde v_1\ne 0,\;\tilde v_2\ne 0\}$. Let us prove that the set of all critical points of $J$ in $\mathcal{O}_1$ is the set of all points of $\mathcal{M}$ such that $\tilde u_1=\tilde u_2=0$. Suppose that there are critical points in $\mathcal{O}_1$ such that $\tilde{u}_1\neq 0$ or $\tilde{u}_2\neq 0$. For such points the following identity holds $$\label{3} |\eta_3|^2=(\lambda_++\eta_1)(\lambda_-+\eta_2).$$ In $\mathcal{O}_1$, $\tilde{v}_1\ne 0$ and therefore $|\eta_3|^2=\eta_1\eta_2$. This equality together with (\[3\]) gives $$\label{eq11} \eta_2=-{\lambda}_-\left(1+\frac{\eta_1}{{\lambda}_+}\right)$$ Suppose that $\eta_3\neq 0$. Then, using (\[e1\]) and (\[e3\]), the constraint ${\Phi}_3$ gives $$(\lambda_++\eta_1)\|\tilde{u}_1\|^2+\eta_1\|\tilde{v}_1\|^2=0.$$ Constraint ${\Phi}_1$ gives $\|\tilde{v}_1\|^2=1-\|\tilde{u}_1\|^2$, and therefore $\eta_1=-\lambda_+\|\tilde{u}_1\|^2$. Similarly we find $\eta_2=-\lambda_-\|\tilde{u}_2\|^2$. Substituting these expressions for $\eta_1$ and $\eta_2$ into the (\[e1\]) and (\[e2\]) we find $$\label{eq2} \left\{\begin{array}{c} \sqrt{\lambda_-\lambda_+}\|\tilde{u}_1\|\|\tilde{u}_2\|^2= \lambda_+(1-\|\tilde{u}_1\|^2)\|\tilde{u}_1\|\\ \sqrt{\lambda_-\lambda_+}\|\tilde{u}_1\|^2\|\tilde{u}_2\|= \lambda_-(1-\|\tilde{u}_2\|^2)\|\tilde{u}_2\| \end{array}\right. \Rightarrow \left\{\begin{array}{c} \sqrt{\lambda_+}=\sqrt{\lambda_+}\|\tilde{u}_1\|^2+\sqrt{\lambda_-}\|\tilde{u}_2\|^2\\ \sqrt{\lambda_-}=\sqrt{\lambda_+}\|\tilde{u}_1\|^2+\sqrt{\lambda_-}\|\tilde{u}_2\|^2 \end{array}\right.$$ This system of equations implies $\lambda_-=\lambda_+\Leftrightarrow{{\bf w}}=0$ which is in contradiction with the assumption $\|\mathbf{w}\|>0$ for the present case. If $\eta_3=0$, then it follows from (\[e3\]), (\[e4\]) that $\eta_1=\eta_2=0$. In this case equations (\[e1\]) and (\[e2\]) have only the solution $\tilde{u}_1=\tilde{u}_2=0$. Points in ${\cal O}_1$ with $\tilde u_1=\tilde u_2=0$ form the global minimum manifold $\mathcal{M}^{{\bf w}}_{\rm min}=V_2(\mathbb{C}^4)$, which is a Stiefel manifold and hence is connected. In some small neighborhood of zero we can choose $\tilde{u}_1$ and $\tilde{u}_2$ as normal coordinates. So $\mathcal{M}^{{\bf w}}_{\rm min}$ is non degenerate. Similar treatment of the region ${\cal O}_2=\{x\in{\cal M}\,|\,\tilde u_1\ne0,\,\tilde u_2\ne 0\}$ gives the global maximum manifold ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=\tilde v_2=0\}$. Now consider the region $\mathcal{O}_3=\{x \in \mathcal{M}\,|\,\tilde u_2\neq 0,\,\tilde v_1\neq 0 \}$. In this region the objective function $J$ has the form $$\begin{aligned} J[\tilde{u}_1,\tilde{u}_2,\tilde{v}_1,\tilde{v}_2]=\lambda_-+ \lambda_+\|\tilde{u}_1\|^2-\lambda_-\|\tilde{v}_2\|^2.\end{aligned}$$ Using the analysis for the region ${\cal O}_1$, we conclude that the objective function has no critical points such that $\tilde v_2\neq 0$ in ${\cal O}_3$. Therefore all critical points in $\mathcal{O}_3$ are in the sub-manifold $\mathcal{N}=\{x \in \mathcal{M}\,|\,\tilde{v}_2=0\}\subset\mathcal{M}$. The restriction of $J$ to $\mathcal{N}$ has the form $$\begin{aligned} J[\tilde{u}_1,\tilde u_2,\tilde v_1,\tilde{v}_2]|_{\mathcal{N}}=\lambda_-+\lambda_+\|\tilde{u}_1\|^2.\end{aligned}$$ Note that $\mathcal{N}$ is a subset of all sets of vectors $(\tilde{u}_1,\tilde{u}_2,\tilde{v}_1)$ satisfying the constraints $$\|\tilde{u}_2\|^2=1,\qquad \|\tilde{u}_1\|^2+\|\tilde{v}_1\|^2=1,\qquad \langle\tilde{u}_1,\tilde{u}_2\rangle=0.$$ It is clear from this representation of $\mathcal{N}$ that $\nabla J|_{\mathcal{N}}=0$ if and only if $\tilde{u}_1=0$. This gives the critical sub-manifold ${\cal M}^{{\bf w}}_-=\{x\in{\cal M}\,|\,\tilde u_1=\tilde v_2=0\}$. The objective function has the value $J\mid_{{\cal M}^{{\bf w}}_-}={\lambda}_-$ on this manifold. To show that this is a saddle manifold, and not a local maximum or minimum, we calculate the Morse indices of the objective function on ${\cal M}^{{\bf w}}_-$ and show that both positive and negative Morse indices are different from zero (the Morse indices are the numbers of positive, negative and zero eigenvalues of the Hessian of $J$ and positive and negative Morse indices determine the number of local coordinates along which the function increases or decreases, respectively). With regard to this goal, consider the manifold $\mathcal{K}:=\{x \in \mathbb{C}^{16}\,|\,\Phi_1(\widetilde{u},\widetilde{v})=0, \;\Phi_2(\widetilde{u},\widetilde{v})=0\}$. Let $x \in \mathcal{M}$. Below we introduce some coordinates in a neighborhood of $x$ on $\mathcal{K}$. For any $z \in \mathbb{C}^4$ such that $z\neq0$ we define the unit vector $g(z)=z/\|z\|\in\mathbb{C}^4$. Let $\varphi_i,\;i=1,\dots,7$ be some coordinate system on $S^7$ (embedded in $\mathbb{C}^8$ as a unit sphere with the origin at zero) in some neighborhood $V_u$ of $g(\tilde{u}_2(x))$ and $\psi_i,\;i=1,\dots,7$ be some coordinate system on $S^7$ in some neighborhood $V_v$ of $g(\tilde{v}_1(x))$. We will use the following functions defined in some neighborhood of $x$ on $\mathcal{K}$ ($z \in \mathcal{K}$): $$\begin{aligned} \tilde{\varphi}_i(z)&=&\varphi_i\circ g\circ \tilde{u}_2(z),\qquad i=1,\dots,7\\ \tilde{\psi}_i(z)&=&\psi_i\circ g\circ \tilde{v}_1(z),\qquad i=1,\dots,7.\end{aligned}$$ Let $T_z S^7$ be the maximal complex subspace of the tangent space of $S^7$. For each $z \in V_u$ let $x_1,\dots,x_6$ be coordinates on $T_z {S^7}$ and for each $z \in V_v$ $y_1,\dots,y_6$ be coordinates on $T_z{S^7}$. Let $\tilde{x}_1,\dots,\tilde{x}_6$ and $\tilde{y}_1,\dots,\tilde{y}_6$ be functions on $\mathcal{K}$ defined as follows. Let $z=(\tilde{u}_1,\tilde{u}_2,\tilde{v}_1,\tilde{v}_2) \in \mathcal{K}$ be in a small enough neighborhood of $x$. By definition $\rm Pr \mit_u$ is the projection from $\mathbb{C}^4$ to $T_{g(\tilde{u}_2)}S^7$ and $\rm Pr \mit_v$ is the projection from $\mathbb{C}^4$ to $T_{g(\tilde{v}_1)}S^7$. By definition $$\begin{aligned} \tilde{x}_i&=&x_i\circ{\rm Pr}_u \circ \widetilde{u}_1, \qquad i=1,\dots,6,\\ \tilde{y}_i&=&y_i\circ{\rm Pr}_v \circ \widetilde{v}_2,\qquad i=1,\dots,6.\end{aligned}$$ Now let $\rm Pr \mit '_{u}$ and $\rm Pr \mit '_{v}$ be the complex-valued functions defined on $\mathbb{C}^4$ by the formulas $$\begin{aligned} {\rm Pr'}_{u}(f)&=&\langle g(\widetilde{u}_2),f\rangle,\qquad f \in \mathbb{C}^4\\ {\rm Pr'}_{v}(f)&=&\langle g(\widetilde{v}_1),f\rangle,\qquad f \in \mathbb{C}^4.\end{aligned}$$ By definition $$\begin{aligned} p:={\rm Pr'}_{u}\circ\widetilde{u}_1,\qquad q:={\rm Pr'}_{v}\circ\widetilde{v}_2.\end{aligned}$$ Thus, the functions $\tilde{\varphi}_i,\tilde{\psi}_i,\tilde{x}_k,\tilde{y}_l,p,q$, where $i,j=1,\dots,7$ and $k,l=1,\dots,6$, are coordinates on $\mathcal{K}$ in some neighborhood of the point $x$. Locally the manifold $\mathcal{M}$ is a sub-manifold of $\mathcal{K}$ defined by the constraint $\Phi_3=0$. In our coordinates this constraint has a form $$p\left(1-\sum \limits_{i=1}^{6}y_i^2-|q|^2\right)^{\frac{1}{2}}+ q\left(1-\sum \limits_{i=1}^{6}x_i^2-|p|^2\right)^{\frac{1}{2}}=0.$$ Therefore $\tilde{\varphi}_i,\tilde{\psi}_i,\tilde{x}_k,\tilde{y}_l,p$, where $i,j=1,\dots,7$ and $k,l=1,\dots,6$ are the coordinates on $\mathcal{M}$ in some neighborhood of $x$. The second differential of $J$ at the point $x$ in this coordinates has the form $$\begin{aligned} {{\rm d}}^2J=\lambda_+\sum \limits_{i=1}^{6}{{\rm d}}x_i^2- \lambda_- \sum \limits_{i=1}^{6}{{\rm d}}y_i^2+(\lambda_+-\lambda_-)|{{\rm d}}p|^2.\end{aligned}$$ Since ${\lambda}_+-{\lambda}_-=\|{{\bf w}}\|>0$ for the present case, the Morse indices of this point are $\nu_+=8,\;\nu_-=6$ (note that $p$ is a complex coordinate). Similar treatment of the region ${\cal O}_4=\{x\in{\cal M}\,|\,\tilde u_1\ne 0,\,\tilde v_2\ne 0\}$ shows the existence of the critical sub-manifold ${\cal M}^{{\bf w}}_+=\{x\in{\cal M}\,|\,\tilde u_2=\tilde v_1=0\}$. This sub-manifold corresponds to the critical value $J\mid_{{\cal M}^{{\bf w}}_+}=\lambda_+$ and its Morse indices are $\nu_+=6,\;\nu_-=8$. Since $\bigcup\limits_{i=0}^4{\cal O}_i={\cal M}$, this concludes the proof for the case $0<\|{{\bf w}}\|<1$. [**Case 1.**]{} [[**w**]{}]{}=0. Consider in $\cal M$ the open subset $\mathcal{O}_1$. Let $\eta_3=0$. Then in the region ${\cal O}_1$ Eqs. (\[e3\]) and (\[e4\]) imply that $\eta_1\tilde v_1=\eta_2\tilde v_2=0\Rightarrow\eta_1=\eta_2=0$. Equations (\[e1\]) and (\[e2\]) for such $\eta_i$ have only the solution $\tilde u_1=\tilde u_2=0$ which defines the global minimum manifold ${\cal M}^{(0,0,0)}_{\rm min}=\{x\in{\cal M}\,|\,\tilde u_1=\tilde u_2=0\}$. Now let $\eta_3\ne 0$ and $\tilde u_1\ne 0$ or $\tilde u_2\ne 0$. In this case Eqs. (\[e1\])–(\[e4\]) give $|\eta_3|^2=(1+\eta_1)(1+\eta_2)$ and $|\eta_3|^2=\eta_1\eta_2$, which imply $\eta_2=-1-\eta_1$ and $|\eta_3|^2=-\eta_1(1+\eta_1)$. Then Eqs. (\[e1\]) and (\[e4\]) have the solution $$\label{eq3} \tilde u_2=-\frac{1+\eta_1}{\eta_3}\tilde u_1=z\tilde u_1,\qquad \tilde v_1=-\frac{\eta_2}{\eta^*_3}\tilde v_2=-z^*\tilde v_2$$ where we used the notation $z=-(1+\eta_1)/\eta_3\in\mathbb C/\{0\}$ and the relation $-\eta_2/\eta^*_3=-z^*$. Note that for a given pair $(\tilde u_1,\tilde v_2)\in\mathbb C^8$, $z$ can be any non-zero complex number such that $(\tilde u_1,z\tilde u_1,-z^*\tilde v_2,\tilde v_2)\in{\cal M}$. The solutions of the form (\[eq3\]) constitute the critical set ${\cal T}=\{x\in{\cal O}_1\,|\,\tilde u_2=z\tilde u_1, \tilde v_1=-z^*\tilde v_2, z\in\mathbb C\}\subset {\cal M}^{(0,0,0)}_{\rm saddle}$. A similar treatment of the region ${\cal O}_2$ shows that the objective function in this region has as critical points only the global maximum manifold ${\cal M}^{(0,0,0)}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=\tilde v_2=0\}$ and the set $\cal T$. Now consider the region $\mathcal{O}_3$. Let $\eta_3=0$. Then in the region ${\cal O}_3$ Eqs. (\[e2\]) and (\[e3\]) imply $(1+\eta_2)\tilde u_2=\eta_1\tilde v_1=0\Rightarrow \eta_1=0$, $\eta_2=-1$. The solution of Eqs. (\[e1\]) and (\[e4\]) for such values of $\eta_i$ gives the critical set $\{x\in{\cal M}\,|\,\tilde u_1=\tilde v_2=0\}\subset{\cal M}^{(0,0,0)}_{\rm saddle}$. Let $\eta_3\ne 0$. The treatment is similar to the treatment of the case $\eta_3\ne 0$ for the region ${\cal O}_1$ and gives the critical set $\cal T$. A similar treatment of the region ${\cal O}_4$ shows that the set of critical points of the objective function in this region is $\{x\in{\cal M}\,|\,\tilde u_2=\tilde v_1=0\}\bigcup{\cal T}$. Combining together the results for the regions ${\cal O}_1$, ${\cal O}_2$, ${\cal O}_3$, and ${\cal O}_4$, we find that the critical manifolds are the global minimum manifold ${\cal M}^{(0,0,0)}_{\rm min}$, the global maximum manifold ${\cal M}^{(0,0,0)}_{\rm max}$, and the set ${\cal T}\bigcup\{x\in{\cal M}\,|\,\tilde u_2=\tilde v_1=0\}\bigcup\{x\in{\cal M}\,|\,\tilde u_1=\tilde v_2=0\}\equiv{\cal M}^{(0,0,0)}_{\rm saddle}$. Since $\bigcup\limits_{i=1}^4{\cal O}_i={\cal M}$, these manifolds are all critical manifolds of the objective function $J$ for the case ${{\bf w}}=0$. A simple computation using the constraints (\[c1\])–(\[c3\]) shows that the value of the objective function at any point $x\in{\cal M}^{(0,0,0)}_{\rm saddle}$ equals to $1/2$, i.e., $J|_{{\cal M}^0}=1/2$. Now we will find Morse indices of the critical manifold ${\cal M}^{(0,0,0)}_{\rm saddle}$. An arbitrary point $x=(u_1,u_2,v_1,v_2) \in \mathcal{M}^{(0,0,0)}_{\rm saddle}$ can be moved into the point $\tilde{x}=(\tilde{u}_1,\tilde{u}_2,\tilde{v}_1,\tilde{v}_2) \in \mathcal{M}^{(0,0,0)}_{\rm saddle}$ with $\tilde{u}_1=0$, $\tilde{v}_2=0$ by the following transformation: $$\begin{aligned} \tilde{u}_1&=&\alpha u_1+\beta u_2,\qquad \tilde{u}_2=- \beta^*u_1+\alpha^* u_2,\\ \tilde{v}_1&=&\alpha v_1+\beta v_2,\qquad \tilde{v}_2=- \beta^*v_1+\alpha^* v_2,\end{aligned}$$ where $\alpha,\beta \in \mathbb{C}$, $|\alpha|^2+|\beta|^2=1$. For example, $\alpha=-\beta z$ for $x=(u_1,zu_1,-z^*v_2,v_2) \in \mathcal{T}$. As in the analysis of the Morse indices for the case 2, in some neighborhood of $\tilde{x}$ we can introduce the coordinates $\tilde{\varphi}_i,\tilde{\psi}_i, \tilde{x}_k,\tilde{y}_l,p, q$, where $i,j=1,\dots,7$ and $k,l=1,\dots,6$. These coordinates satisfy the constraint: $$p\left(1-\sum \limits_{i=1}^6\tilde{y}_i^2-|q|^2\right)^{\frac{1}{2}}+q\left(1-\sum \limits_{i=1}^6\tilde{x}_i^2-|p|^2\right)^{\frac{1}{2}}=0.$$ The second differential of $J$ in these coordinates has the form: $$\begin{aligned} {\rm d}^2J=\sum \limits_{i=1}^6 {{\rm d}}\tilde{x}_i^2-\sum \limits_{i=1}^6 {{\rm d}}\tilde{y}_i^2+0\cdot |{{\rm d}}p|^2.\end{aligned}$$ It is easy to see that the tangent space to $\mathcal{M}^{(0,0,0)}_{\rm saddle}$ at the point $\tilde{x}$ is spanned by the vectors $$\frac{\partial}{\partial \tilde{\varphi}_i},\; \frac{\partial}{\partial \tilde{\psi}_i},\; \frac{\partial}{\partial \rm Re \mit p},\;\frac{\partial}{\partial \rm Im \mit p}.$$ Therefore $\mathcal{M}^{(0,0,0)}_{\rm saddle}$ is nondegenerate, ${\rm dim}\,{\cal M}^{(0,0,0)}_{\rm saddle}=16$ and the Morse indices of $\mathcal{M}^{(0,0,0)}_{\rm saddle}$ are $\nu_+=\nu_-=6$. [**Case 3.**]{} $ \|{{\bf w}}\|=1$. In this case ${\lambda}_-=0$, ${\lambda}_+=1$, and $$\begin{aligned} J[\tilde{u}_1,\tilde u_2,\tilde v_1,\tilde{v}_2]=\|\tilde{u}_1\|^2.\end{aligned}$$ Let $\mathcal{U}_1=\{x \in \mathcal{M}\,|\,\tilde{v}_1\neq 0\}$. Clearly, points in ${\cal U}_1$ with $\tilde u_1=0$ form the global minimum of the objective. Assume that there are critical points in $\mathcal{U}_1$ such that $\tilde u_1\ne 0$. For such points Eqs. (\[e1\])–(\[e4\]) imply the system of equations $$\begin{aligned} |\eta_3|^2&=&\eta_1\eta_2\\ |\eta_3|^2&=&\eta_2(1+\eta_1)\end{aligned}$$ which has only the solutions with $\eta_2=\eta_3=0$. But in the region $\mathcal{U}_1$, $\tilde{v}_1\neq 0$ and therefore Eq. (\[e3\]) implies $\eta_1=0$. Then, Eq. (\[e1\]) for $\eta_1=\eta_2=\eta_3=0$ has the solution $\tilde{u}_1=0$ which contradicts the assumption $\tilde u_1\ne 0$. As a result, the only critical points in $\mathcal{U}_1$ are with $\tilde{u}_1=0$. These points form the global minimum manifold ${\cal M}^{{\bf w}}_{\rm min}=\{x\in{\cal M}\,|\,\tilde u_1=0\}$. This manifold is diffeomorphic to the space bundle with $S^7$ as a base and $S^{14}$ as a fibre. Thus, ${\cal M}^{{\bf w}}_{\rm min}$ is connected. We can use $\tilde{u}_1$ as normal coordinates in some neighborhood of ${\cal M}^{{\bf w}}_{\rm min}$. Thus ${\cal M}^{{\bf w}}_{\rm min}$ is nondegenerate. The treatment of the region $\mathcal{U}_2=\{x \in \mathcal{M}\,|\,\tilde{u}_1\neq 0\}$ is equivalent to the previous consideration. The critical points in this region form the global maximum manifold ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,\tilde v_1=0\}$. Note that $\mathcal{U}_1\cup\mathcal{U}_2=\mathcal{M}$. Therefore, all critical points of $J$ correspond to the global minimum $J=0$ and global maximum $J=1$. The critical manifolds corresponding to the minimum and the maximum are connected and nondegenerate. $\Box$ The critical manifolds in terms of the original parametrization of the Kraus operators by $(u_1,u_2,v_1,v_2)$ can be obtained by expressing $\tilde u_i$ and $\tilde v_i$ in terms of $u_i$ and $v_i$. If $z_0\ne 0$, then it follows from (\[nc1\]) and (\[nc2\]) that $$\begin{aligned} \tilde u_1&=&\mu u_1+\frac{z_0}{|z_0|}\nu u_2,\qquad \tilde u_2=-\nu u_1+\frac{z_0}{|z_0|}\mu u_2\\ \tilde v_1&=&\mu v_1+\frac{z_0}{|z_0|}\nu v_2,\qquad \tilde v_2=-\nu v_1+\frac{z_0}{|z_0|}\mu v_2\end{aligned}$$ Thus, for $z_0\ne 0$ and $0<\|{{\bf w}}\|<1$ the critical manifolds are the following: the global minimum ${\cal M}^{{\bf w}}_{\rm min}=\{x\in{\cal M}\,|\,u_1=u_2=0\}$, the global maximum ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,v_1=v_2=0\}$, and the saddles ${\cal M}^{{\bf w}}_\pm=\{x\in{\cal M}\,|\,u_2=z_\pm u_1, v_1=-z^*_\pm v_2\}$. Here $z_\pm=z_0^*/({\gamma}\pm\|{{\bf w}}\|)$. For $z_0\ne 0$ and $\|{{\bf w}}\|=1$ (hence ${\gamma}\ne 1$), the critical manifolds are ${\cal M}^{{\bf w}}_{\rm min}=\{x\in{\cal M}\,|\,u_2=z_0^*u_1/({\gamma}-1)\}$, ${\cal M}^{{\bf w}}_{\rm max}=\{x\in{\cal M}\,|\,v_2=z_0^*v_1/({\gamma}-1)\}$, and there are no saddles. If $z_0=0$ and ${\gamma}\ge 0$, then $\tilde u_1=u_1$, $\tilde u_2=u_2$, $\tilde v_1=v_1$, and $\tilde v_2=v_2$. Thus for ${\gamma}=0$, ${\cal M}^{(0,0,0)}_{\rm min}=\{x\in{\cal M}\,|\, u_1=u_2=0\}$, ${\cal M}^{(0,0,0)}_{\rm max}=\{x\in{\cal M}\,|\, v_1=v_2=0\}$, and ${\cal M}^{(0,0,0)}_{\rm saddle}=\{x\in{\cal M}\,|\, u_2=zu_1, v_1=-z^*v_2, z\in\mathbb C\}\bigcup\{x\in{\cal M}\,|\, u_1=v_2=0\}$. For $0<{\gamma}<1$ the critical manifolds are ${\cal M}^{(0,0,{\gamma})}_{\rm min}=\{x\in{\cal M}\,|\, u_1=u_2=0\}$, ${\cal M}^{(0,0,{\gamma})}_{\rm max}=\{x\in{\cal M}\,|\, v_1=v_2=0\}$, and the saddles ${\cal M}^{(0,0,{\gamma})}_-=\{x\in{\cal M}\,|\, u_1=v_2=0\}$ and ${\cal M}^{(0,0,{\gamma})}_+=\{x\in{\cal M}\,|\, u_2=v_1=0\}$. For ${\gamma}=1$, ${\cal M}^{(0,0,1)}_{\rm min}=\{x\in{\cal M}\,|\, u_1=0\}$ and ${\cal M}^{(0,0,1)}_{\rm max}=\{x\in{\cal M}\,|\, v_1=0\}$. If $z_0=0$ and ${\gamma}<0$, then $\tilde u_1=u_2$, $\tilde u_2=u_1$, $\tilde v_1=v_2$, and $\tilde v_2=v_1$. In this case for $-1<{\gamma}<0$ the critical manifolds are the following: ${\cal M}^{(0,0,{\gamma})}_{\rm min}=\{x\in{\cal M}\,|\, u_1=u_2=0\}$, ${\cal M}^{(0,0,{\gamma})}_{\rm max}=\{x\in{\cal M}\,|\, v_1=v_2=0\}$, ${\cal M}^{(0,0,{\gamma})}_-=\{x\in{\cal M}\,|\, u_2=v_1=0\}$ and ${\cal M}^{(0,0,{\gamma})}_+=\{x\in{\cal M}\,|\, u_1=v_2=0\}$. For ${\gamma}=-1$, ${\cal M}^{(0,0,-1)}_{\rm min}=\{x\in{\cal M}\,|\, u_2=0\}$ and ${\cal M}^{(0,0,-1)}_{\rm max}=\{x\in{\cal M}\,|\, v_2=0\}$. [The values of the objective function at the saddle points satisfy the equality $J_+({{\bf w}})+J_-({{\bf w}})=1$. This fact is a consequence of the more general symmetry of the objective function, defined by the duality map $T:{\cal M}\to{\cal M}$ such that $T(u_1,u_2,v_1,v_2)=(v_1,v_2,u_1,u_2)$ as $J[x;{{\bf w}}]+J[T(x);{{\bf w}}]=1$ for any $x\in{\cal M}$. Thus, if the level set $\Gamma_{{\bf w}}(\alpha):=\{x\in{\cal M}\,|\,J[x,{{\bf w}}]=\alpha\}$ for some value $\alpha\in[0,1]$ is known then one immediately gets the level set for the value $1-\alpha$ as $\Gamma_{{\bf w}}(1-\alpha)=T(\Gamma_{{\bf w}}(\alpha))$.]{} Connectivity of the level sets {#sec4} ============================== The level set $\Gamma_{{\bf w}}(\mu)$ for an admissible objective value $\mu\in [0,1]$ is defined as the set of all controls $x=(u_1,u_2,v_1,v_2)\in{\cal M}$ which produce the same outcome value $\mu$ for the objective function $J[u_1,u_2,v_1,v_2;{{\bf w}}]$, i.e., $\Gamma_{{\bf w}}(\mu)=\{x\in{\cal M}\,|\,J[x;{{\bf w}}]=\mu\}$ (we omit the subscript ${{\bf w}}$ in the sequel). In this section it is shown that each level set for the function $J[\cdot;{{\bf w}}]$ is connected. This means that any pair of solutions in a level set $\Gamma(\mu)$ is connected via a continuous pathway of solutions entirely passing through $\Gamma(\mu)$. Practically, connectivity of the level sets implies the possibility to experimentally locate more desirable solutions via continuous variations of the control parameters while maintaining the same value of the objective function. The proof of the connectivity of the level sets for the objective functions defined by (\[eq7\]) is based on generalized Morse theory, which is presented in the remainder of this section. Theorem \[t1\] below formulates the conditions for a generalized Morse function to have connected level sets. These conditions are satisfied for the objective function $J[\cdot,{{\bf w}}]$ defined by (\[eq7\]), as stated in the end of this section. Formulation of Theorem \[t1\] includes a very general class of functions and can be applied to the investigation of connectivity of the level sets for situations beyond the scope of this paper, including landscapes for multilevel closed and open systems. Connectivity of level sets of generalized Morse functions --------------------------------------------------------- Let $M$ be a smooth compact manifold of dimension $d$, and let $f$ be a smooth function $f:M\rightarrow\mathbb{R}$. We suppose that the critical set of $f$, $S:=\{x \in M| {{\rm d}}f(x)=0\}$ is a disjoint union of smooth connected sub-manifolds $C_i$ $(i=1,2,\dots,n)$ of dimension $d_i$. Let $\mu_i=f|_{C_i}$. For each point $x \in C_i$ there exists an open neighborhood $U$ of $x$ and a coordinate system $\{x_l\}$ in $U$ such that $$\begin{aligned} C_i\cap U=\{x \in U|x_{d_i+1}=\cdots=x_n=0\}.\end{aligned}$$ Consider the following matrix $$\begin{aligned} J_i(x):= \left \| \frac{\partial^2 f(x)}{\partial x_l\partial x_m} \right\|_{l,m=d_i+1,\dots,d}, \qquad x\in C_i.\end{aligned}$$ It is easy to see that if $\{y_l\}$ is another coordinate system in $U$ such that $$\begin{aligned} C_i\cap U=\{y \in U|y_{d_i+1}=\cdots=y_n=0\},\end{aligned}$$ and $$\begin{aligned} \widetilde{J}_i(x):= \left \| \frac{\partial^2 f(x)}{\partial y_l\partial y_m} \right\|_{l,m=d_i+1,\dots,d}, \qquad x\in C_i\end{aligned}$$ then $$\begin{aligned} \rm rank \mit J_i(x)=\rm rank \mit \widetilde{J}_i(x).\end{aligned}$$ Therefore we can give the following The point $x \in C_i$ is said to be nondegenerate if ${\rm det}\, J_i(x)\neq 0$. A critical sub-manifold $C_i$ is said to be nondegenerate if $\forall x \in C_i$, $x$ is a nondegenerate point. Let $x \in C_i$ and $\lambda_i^+(x)$, $\lambda_i^-(x)$ be the numbers of positive and negative eigenvalues of the matrix $J_i(x)$. It is clear that $\lambda_i^+(x),\lambda_i^-(x)$ do not depend on the choice of coordinate system $\{x_i\}$ in the neighborhood of $x$. One can prove that $\lambda_i^+(x)$ and $\lambda_i^-(x)$ do not depend on the point $x \in C_i$ ($\lambda_i^+(x)$ and $\lambda_i^-(x)$ are continuous and $C_i$ is connected.). Let $\lambda_i^+:=\lambda_i^+(x)$ and $\lambda_i^-:=\lambda_i^-(x)$. $\lambda_i^+$ and $\lambda_i^-$ are called the indices of $C_i$. Let $M$ be a smooth compact connected manifold and $f:M\rightarrow\mathbb{R}$. Suppose that the critical set of $f$ is a disjoint union of (compact) connected nondegenerate sub-manifolds $C_i$. In this case we say that $f$ is a generalized Morse function. Sub-manifolds $C_i$ are called the critical sub-manifolds of $f$. \[t1\] Let $M$ be a smooth compact connected manifold and $f$ be a generalized Morse function. Let $C_i$, $i=1,\dots,n$ be critical sub-manifolds of $f$ and $\mu_i=f|_{C_i}$. We can assume that $\mu_{\min}:=\mu_1\leq\mu_2\leq\dots\leq \mu_n=:\mu_{\max}$. Suppose that the sub-manifold $C_{\max}:=f^{-1}(\mu_{\max})$ is connected. Suppose also that $\forall i=1,\dots,n-1$ the indices $\lambda_i^+\geq2$, $\lambda_i^-\geq 2$. Then $\forall \mu: \mu_{\min}\leq \mu\leq \mu_{\max}$ the set $\Gamma(\mu):=f^{-1}(\mu)$ is connected. [**Proof.**]{} We decompose the proof of the theorem into a sequence of several Lemmas. \[m:1\] There exists an open neighborhood $U$ of $C_{\max}$ such that $U$ is diffeomorphic to some bundle $\mathcal{E}$ with the base $C_{\max}$ and the fibre $B_{d-d_n}$. Here $B_k$ is a $k$-dimensional ball. **Proof.** M is a compact. Therefore there exists a Riemann metric $g \in \rm sym (\mit T^\ast M\otimes T^\ast M) $. (Here $T^\ast M$ is a cotangent bundle of $M$.) By definition, $\mathcal{L}$ is a restriction of the tangent bundle $TM$ to $C_{\max}$. Let $\mathcal{N}$ be a sub-bundle of $\mathcal{L}$ such that $\forall x \in C_{\max}$ the fiber $\mathcal{N}_x$ of $\mathcal{N}$ over $x$ is a subspace of $T_xM$ consisting of all vectors orthogonal to $T_xC_{\max}$. Let $\mathcal{B}_l$ be a sub-bundle of $\mathcal{N}$ such that $\forall x \in C_{\max}$ the fiber $(\mathcal{B}_l)_x$ of $\mathcal{B}_l$ is a set of all vectors $v$ of $\mathcal{N}_x$ satisfying the following inequality: $\|v\|< l$ (with respect to the metric $g$). Let $\gamma_v(x)(t)$ ($x \in M,v \in T_x M, t \in \mathbb{R}$) be a geodesic line, i.e., the solution of the following ordinary differential equation $$\begin{aligned} \nabla_{\dot{\gamma}_v(x)(t)}\dot{\gamma}_v(x)(t)=0\end{aligned}$$ with the following initial conditions $$\begin{aligned} \gamma_v(x)(0)&=&x,\nonumber\\ \dot{\gamma}_v(x)(t)|_{t=0}&=&v.\end{aligned}$$ Here $\nabla_v$ is a Levi-Civita connection on $M$ with respect to the metric $g$. The solution of this differential equation is defined on the whole real line because $M$ is compact. Let $F_l$ for $l \in (0,+\infty)$ be a map $\mathcal{B}_l\rightarrow M$ which assigns to each point $(x,v) \in \mathcal{B}_l$ ($x \in C_{\max}, v \in (\mathcal{B}_l)_x$) the point $\gamma_v(x)(1)$. It follows from the inverse function theorem that there exits a number $l_0>0$ such that $ F_l$ is a diffeomorphism on its image for all $l:0<l\leq l_0$. $\Box$ If $\varepsilon$ is small enough then $\forall \mu: \mu_{\max}>\mu>\mu_{\max}-\varepsilon$ the set $\Gamma(\mu)=f^{-1}(\mu)$ is connected. **Proof.** Let $l_0$ be a number from the previous Lemma. It follows from the Morse Lemma that for every $x \in C_{\max}$ we can choose coordinates $z_1,\dots,z_{d-d_{n}}$ on $(\mathcal{B}_{l_0})_x$ in some neighborhood $U$ of zero such that $$\begin{aligned} f\circ F_{l_0}|_U=z_1^2+\dots+z_{d-d_{n}}^2.\end{aligned}$$ Moreover, from construction of these coordinates it follows that in some neighborhood of every point $x_0 \in C_{\max}$ they are differentiable functions of $x$. Therefore, there exists a finite covering $\{U_i\}_{i=1,\dots,q}$ of $C_{\max}$ by open connected sets and a family of diffeomorphisms $g_i:U_i\times B_{d-d_n}\rightarrow \pi^{-1}(U_i)$ $(i=1,\dots,q)$ on its image commuting with the projections such that $$\begin{aligned} f\circ F_{l_0}\circ g_i=z_1^2+\dots+z^2_{d-d_{n}},\qquad i=1,\dots,q.\end{aligned}$$ Here $z_i$, $i=1,\dots,d-d_{n}$ are some coordinates on the ball $B_{d-d_n}$ and $\pi$ is a canonical projection from $\mathcal{B}_{l_0}$ to $C_{\max}$. We now prove that for every $l_1:0<l_1<l_0$ there exists $\varepsilon_1>0$ such that $\forall \mu: \mu_{\max}- \varepsilon_1 <\mu\leq \mu_{\max}$, $\Gamma(\mu) \subset F_{l_1}\mathcal{B}_{l_1}$. Suppose that $\forall n=1,2,\dots$ there exists a point $x_n$ such that $f(x_n)>\mu_{\max}-1/n$ and $x_n \notin F_{l_1}\mathcal{B}_{l_1}$. Because $M\setminus F_{l_1}\mathcal{B}_{l_1}$ is compact, then there exists a point $x_0 \in M\setminus F_{l_1}\mathcal{B}_{l1}$ and sub-sequence $\{x_{n_k}\}$ of $\{x_n\}$ such that $x_{n_k}\rightarrow x_0$ as $k \rightarrow \infty$. We find that $f(x_0)=\mu_{\max}$ and $x_0 \in C_{\max}$. This contradiction proves our statement. If $l_1$ is small enough then $\mathcal{B}_{l_1}\cap U_i\subset g_i(U_i\times B_{d-d_n})$ for all $i=1,\dots,q$. Therefore if $\mu>\mu_{\max}-\varepsilon_1$ then $f^{-1}(\mu)\cap \pi^{-1}(U_i)\subset g_i(U_i\times B_{d-d_n})$ and connected. So we find that $f^{-1}(\mu)$ is connected if $\mu>\mu_{\max}-\varepsilon_1$. $\Box$ Suppose that for some $\mu: \mu_i<\mu<\mu_{i+1}$ $(i=1,\dots,n-1)$ the set $\Gamma(\mu)$ is connected. Then $\forall \mu$ such that $\mu_i<\mu<\mu_{i+1}$, the set $\Gamma(\mu)$ is connected. **Proof.** Let $\nu \in \mathbb{R}:\mu_i<\nu<\mu_{i+1}$. Let us prove that $\Gamma(\nu)$ is connected. We can assume that $\nu<\mu$, and let $\varepsilon$ be a positive number such that $\mu_i<\nu-\varepsilon<\mu+\varepsilon<\mu_{i+1}$. Consider the following sets $$\begin{aligned} U_\varepsilon&=&\{x|\nu-\varepsilon<f(x)<\mu+\varepsilon\}, \nonumber\\ \overline{U}_\varepsilon&=&\{x|\nu-\varepsilon\leq f(x)\leq\mu+\varepsilon\}\end{aligned}$$ Consider also the following differential equation on $M$ $$\begin{aligned} \dot{\gamma}(t)=\frac{\rm grad \mit f(\gamma(t))}{\|\rm grad \mit f(\gamma(t)) \|^2}.\label{1}\end{aligned}$$ (Recall that $M$ has a Riemann metric). The right hand side of this equation is well defined on $U_\varepsilon$. The solution of (\[1\])is $\gamma_x(t)$ with the initial condition $$\begin{aligned} \gamma_x(0)=x,\qquad x \in \Gamma(\mu).\end{aligned}$$ By the extension theorem [@pontriagin; @arnold] this solution must leave the compact set $\overline{U}_{\varepsilon/2}$. It is easy to prove that $f(\gamma_x(t))=t+\mu$. So the solution $\gamma_x(t)$ is defined and unique on the interval $(\nu-\mu-\varepsilon/3,\mu+\varepsilon/3)$. Therefore we have a smooth map $\Delta_{\mu,\nu}:\Gamma(\mu)\rightarrow\Gamma(\nu)$, $x\mapsto \gamma_x(\nu-\mu)$. By the same means we can construct the map $\Delta_{\nu,\mu}:\Gamma(\nu)\rightarrow\Gamma(\mu)$. $\Delta_{\mu,\nu}(x)=y$ if and only if $x$ and $y$ lie on the same integral curve of (\[1\]). We have $\Delta_{\mu,\nu}\circ \Delta_{\nu,\mu}={\rm id}$ and $\Delta_{\nu,\mu}\circ \Delta_{\mu,\nu}={\rm id}$. So $\Gamma(\mu)$ and $\Gamma(\nu)$ are diffeomorphic. $\Box$ Suppose that the assumptions of the theorem hold. Let $\mu \in \mathbb{R}$: $\mu_i<\mu<\mu_{i+1}$, $\mu_i=2,\dots,n-1$, and $\Gamma(\mu)$ is connected. Then $\forall \nu$ such that $\mu_{i-1}<\nu<\mu_{i}$, the set $\Gamma(\nu)$ is also connected. **Proof.** We prove this lemma only for the case of connected $C_i$. The general case is analogous to this case. As in Lemma \[m:1\], let $\mathcal{B}_l$ be a bundle with the base $C_i$ which consists of all vectors $v$ normal to $C_i$ and such that $\|v\|< l$. We have $\mathcal{B}_{l_1}\subset\mathcal{B}_{l_2}$ for $l_1<l_2$. Let $F_l$ be a map $\mathcal{B}_l\rightarrow M$ constructed as in Lemma 1. As in Lemma \[m:1\], we find that $F_l$ is a diffeomorphism if $0<l\leq l_0$ for some positive number $l_0$. As in Lemma 1 we find that for every $l_0'<l_0$ there exists a covering $\{U_j\}_{j=1,\dots,p}$ of $C_i$ by open connected sets and the family of diffeomorphisms $g_j:U_j\times B_{d-d_i}\rightarrow \pi^{-1}(U_j)$ on its image commuting with the projections such that $$\begin{aligned} f\circ F_l\circ g_j=z_1^2+\dots+z^2_{\lambda_i^+}-z^2_{\lambda_i^+ +1}-\dots-z^2_{d-d_i}+\mu_i\end{aligned}$$ Here $B_{d-d_i}$ is a $d-d_i$-dimensional ball and $\pi$ a canonical projection from $\mathcal{B}_{l_0'}$ to $C_i$. It is easy to see that for every $l_0'<l_0$ there exists a positive number $l_1<l_0'$ such that $\forall j=1,\dots,p$ $\mathcal{B}_{l_1}\cap \pi^{-1}(U_j) \subset g_j(U_j\times B_{d-d_i})$. For every $l_1<l_0'$ there exists a positive number $\varepsilon_2$ such that $\forall x \in C_i$, $(\mathcal{B}_{l_1/2})_x\cap F_{l_0}^{-1}(\Gamma(\mu_i+\kappa))\neq\emptyset$ $\forall\kappa :|\kappa|<\varepsilon_2$. We now prove that $\mathcal{B}_{l_1}\cup \pi^{-1}(U_j)\cap F_{l_0}^{-1}( \Gamma(\mu+\kappa))$ is connected $\forall j=1,\dots,p$ if $|\kappa|<\varepsilon_2$. Indeed, let $x_1$ and $x_2$ be two points which lie in the set $\mathcal{B}_{l_1}\cup \pi^{-1}(U_j)\cap F_{l_0}^{-1}(\Gamma(\mu+\kappa))$. We can consider only the case $\kappa>0$. The set $g_j(U_j\times B_{d-d_i})\cap F^{-1}_{l_0}(\Gamma(\mu_i+\kappa))$ is diffeomorphic to $\mathbb{R}^{\lambda_i^+}\times S^{\lambda_i^--1}\times U_i$ and connected. Let $\gamma(t)\;t \in [0,1]$ be a path in $g_j(U_j\times B_{d-d_i})\cap F^{-1}_{l_0}( \Gamma(\mu_i+\kappa))$ such that $\gamma(0)=x_1$, $\gamma(1)=x_2$. Let $d(x)$ be a function on $\mathcal{B}_{l_0}$ defined as follows: $d((z,v))=\|v\|^2$, where $z \in C_i$ and $v \in (\mathcal{B}_{l_0})_x$. Let $x \in F_{l_0}^{-1}(\Gamma(\mu+\kappa))\cap g_j(U_j\times B_{d-d_{i}})$ and $w(x)$ be a projection of $\nabla d(x)$ to the tangent space of $F_{l_0}^{-1}(\Gamma(\mu+\kappa))$ at $x$. It is obvious that $w(x)\neq 0$ $\forall x \in F_{l_0'}^{-1}(\Gamma(\mu+\kappa))\cap g_j(U_j\times B_{d-d_{i}})\cap(\mathcal{B}_{l_0'}\setminus \mathcal{B}_{l_1})$ if $l_0'$ is a sufficiently small number. So we can retract the path $\gamma(t)$ along the vector field $w$ to the part $\tilde{\gamma}(t)$ which lies in $\mathcal{B}_{l_1}$ and connects the points $x_1$ and $x_2$. So $\mathcal{B}_{l_1}\cup \pi^{-1}(U_j)\cap F_{l_0}^{-1}( \Gamma(\mu+\kappa))$ is connected. Now we can find that $\mathcal{B}_{l_1}\cup \cap F_{l_0}^{-1}( \Gamma(\mu+\kappa))$ is connected. Now let $x_1, x_2 \in \Gamma(\mu)$, $\mu<\mu_i$, $|\mu-\mu_i|<\varepsilon_3$. Let $U=F_{l_0}(\mathcal{B})_{l_1/2}$, $V=F_{l_0}(\mathcal{B})_{l_1/3}$, $W=F_{l_0}(\mathcal{B})_{l_1/4}$. At first suppose that $x_1 \notin U$ and $x_2 \notin U$. Let $\gamma_{x_1}(t)$, $\gamma_{x_2}(t)$ be solutions of the differential equation (\[1\]) with initial conditions $x_1$ and $x_2$ respectively. The paths $\gamma_{x_1}(t)$ and $\gamma_{x_2}(t)$ intersect the sub-manifold $\Gamma(\mu+\varepsilon_3)$ at the points $y_1$ and $y_2$ if $\varepsilon_3$ is enough small. Let $\widetilde{\delta}(t), t \in [0,1]$ be a path such that $\forall t \in [0,1]\; \widetilde{\delta}(t) \in \Gamma(\mu+\varepsilon)$ and $y_1=\widetilde{\delta}(0), y_2=\widetilde{\delta}(1)$. We must consider the following two cases. 1\) $\widetilde{\delta}\cap V=\emptyset$. If $\varepsilon_3$ is small enough then we can deform the part $\widetilde{\delta}$ along the vector field $\nabla {f}/\| \nabla f\|^2$ to the part $\delta$ which lies on $\Gamma(\mu)$ and connects the points $x_1$ and $x_2$. 2\) $\widetilde{\delta}\cap V \neq\emptyset$. If $\varepsilon_3$ is small enough then $y_1,y_2 \notin V$. We can decompose the part $\widetilde{\delta}$ as $\widetilde{\delta}=\widetilde{\alpha}_1\circ\widetilde{\beta}\circ\tilde{\alpha}_2$, where $$\begin{aligned} \widetilde{\alpha_2}(1) \in \partial V,\qquad\forall t \in [0,1]\; \widetilde{\alpha_2}(t) \notin V \nonumber\\ \widetilde{\alpha_1}(0) \in \partial V,\qquad\forall t \in [0,1]\; \widetilde{\alpha_1}(t) \notin V.\end{aligned}$$ If $\varepsilon_3$ is a sufficiently small positive number we can deform the paths $\widetilde{\alpha}_1$ and $\widetilde{\alpha}_2$ along the vector field $\nabla f/\|\nabla f\|^2$ into the the paths $\alpha_1,\alpha_2 \subset \Gamma(\mu)$ such that $\alpha_1, \alpha_2 \nsubseteq W$ and $\alpha_2(0)=x_1, \alpha_2(1)\in U$, $\alpha_1(1)=x_2$ and $\alpha_1(0)\in U$. But, it has been proved that $U\cap\Gamma(\mu)$ is connected. Therefore, there exists a path $\beta \subset \Gamma(\mu)$ such that $\beta(1)=\alpha_1(0)$ and $\alpha_2(1)=\beta(0)$. We see that the path $\alpha_1\circ\beta\circ\alpha_2$ connects the point $x_1$ and $x_2$. Consideration of the case with $x_1 \in U$ or $x_2 \in \mathcal{B}_U$ is analogous to consideration of the previous case. The statement of the theorem follows from these four Lemmas. $\Box$ Each level set of the objective function $J[\cdot,{{\bf w}}]$ defined by (\[eq7\]) is connected. [**Proof.**]{} The objective function $J[\cdot,{{\bf w}}]$ is a generalized Morse function. The sub-manifold of solutions corresponding to the global maximum in the coordinates $\tilde u_1,\tilde u_2\in\mathbb C^4$ is defined by $\|\tilde u_1\|=\|\tilde u_2\|=1$, $\langle \tilde u_1,\tilde u_2\rangle=0$. It is a Stiefel manifold, ${\cal M}^{{\bf w}}_{\rm max}=V_2(\mathbb C^4)$, and hence is connected. The Morse indices of the function $J[\cdot,{{\bf w}}]$ are $\nu_\pm>2$ at any saddle sub-manifold. Therefore this function satisfies the conditions of Theorem \[t1\] and its each level set is connected. $\Box$ Conclusions =========== In this paper the landscape of the objective functions for open quantum systems controlled by general Kraus maps is investigated in detail for the two-level case. It is shown that a typical objective function has: (a) no false traps, (b) multi-dimensional sub-manifolds of the optimal global solutions, and (c) each level set is connected. These results may be generalized to systems of arbitrary dimension $N$, although a full enumeration of the critical sub-manifold dimensions remains open for analysis. The landscape analysis and the conclusions rest on assuming that the controls can manage the system and the environment. Managing the environment, in practice, is likely not highly demanding, as control over only the immediate environment of the system is most likely needed. The critical point topology of general controlled open system dynamics could provide a basis to explain the relative ease of practical searches for optimal solutions in the laboratory, even in the presence of an environment. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by the Department of Energy. A. Pechen acknowledges partial support from the grant RFFI 05-01-00884-a. The authors thank Jonathan Roslund for help with drawing the Figure \[fig1\]. [99]{} Butkovskiy A G and Samoilenko Yu I 1984 [*Control of Quantum-Mechanical Processes and Systems*]{} (Moscow: Nauka) \[in Russian\]; (Kluwer Academic, Dordrecht, 1990) \[in English\] Tannor D and Rice S A 1985 [*J. Chem. Phys.*]{} [**83**]{} 5013 Judson R S and Rabitz H 1992 [*Phys. Rev. Lett.*]{} [**68**]{} 1500 Warren W S, Rabitz H and Dahleh M 1993 [*Science*]{} [**259**]{} 1581 Rice S A and Zhao M 2000 [*Optical Control of Molecular Dynamics*]{} (New York: Wiley) Rabitz H, de Vivie-Riedle R, Motzkus M and Kompa K 2000 [*Science*]{} [**288**]{} 824 Shapiro M and Brumer P 2003 [*Principles of the Quantum Control of Molecular Processes*]{} (Hoboken, NJ: Wiley-Interscience) Dantus M and Lozovoy V V 2004 [*Chem. Rev.*]{} [**104**]{} 1813 D’Alessandro D 2007 [*Introduction to Quantum Control and Dynamics*]{} (Boca Raton: Chapman and Hall) Pechen A and Rabitz H 2006 [*Phys. Rev.*]{} A [**73**]{} 062102;\ [*E-print*]{}: http://xxx.lanl.gov/abs/quant-ph/0609097 Romano R and D’Alessandro D 2006 [*Phys. Rev.*]{} A [**73**]{} 022323 Accardi L and Imafuku K 2006 [*QP–PQ: Quantum Probability and White Noise Analysis*]{} vol [**XIX**]{} ed L Accardi, M Ohya and N Watanabe (Singapore: World Sci. Pub. Co.) pp 28–45 Vilela Mendes R and Man’ko V I 2003 [*Phys. Rev.*]{} A [**67**]{} 053404 Mandilara A and Clark J W 2005 [*Phys. Rev.*]{} A [**71**]{} 013406 Roa L, Delgado A, Ladron de Guevara M L and Klimov A B 2006 [*Phys. Rev.*]{} A [**73**]{} 012322 Pechen A, Il’in N, Shuang F and Rabitz H 2006 [*Phys. Rev.*]{} A [**74**]{} 052102;\ [*E-print*]{}: http://xxx.lanl.gov/abs/quant-ph/0606187 Shuang F, Pechen A, Ho T-S and Rabitz H 2007 [*J. Chem. Phys.*]{} [**126**]{} 134303;\ [*E-print*]{}: http://xxx.lanl.gov/abs/quant-ph/0609084 Belavkin V P 1983 [*Automatia and Remote Control*]{} [**44**]{} 178;\ [*E-print*]{}: http://arxiv.org/abs/quant-ph/0408003 Kraus K 1983 [*States, Effects, and Operations*]{} (Berlin, New York: Springer-Verlag) Alicki R and Lendi K 1987 [*Quantum Dynamical Semigroups and Applications*]{} (Berlin: Springer-Verlag) Nielsen M A and Chuang I L 2000 [*Quantum Computation and Quantum Information*]{} (Cambridge, UK: Cambridge University Press) Breuer H-P and Petruccione F 2007 [*The Theory of Open Quantum Systems*]{} (Oxford: Clarendon Press) Goldberg D E 1989 [*Genetic Algorithms in Search, Optimization and Machine Learning*]{} (Reading, MA: Addison-Wesley) Stiefel E 1935–36 [*Commentarii Math. Helvetici*]{} [**8**]{} 305 Edelman A, Arias T A and Smith S T 1998 [*SIAM J. Matrix Anal. Appl.*]{} [**20**]{} 303 Manton J H 2002 [*IEEE Transactions on Signal Processing*]{} [**50**]{} 635 Rabitz H, Hsieh M and Rosenthal C 2004 [*Science*]{} [**303**]{} 1998 Ho T-S and Rabitz H 2006 [*J. Photochemistry and Photobiology*]{} A [**180**]{} 226 Rabitz H, Hsieh M and Rosenthal C 2005 [*Phys. Rev.*]{} A [**72**]{} 052337 Hsieh M, Wu R and Rabitz H (unpublished) Wu R, Hsieh M and Rabitz H 2008 [*J. Phys. A: Math. Theor.*]{} [**41**]{} 015006 Wu R, Pechen A, Brif C and Rabitz H 2007 [*J. Phys. A: Math. Theor.*]{} [**40**]{} 5681;\ [*E-print*]{}: http://xxx.lanl.gov/abs/quant-ph/0611215 Tarasov V E 2002 [*J. Phys. A: Math. Gen.*]{} [[**35**]{} 5207](http://dx.doi.org/10.1088/0305-4470/35/25/305). Grace M, Brif C, Rabitz H, Walmsley I A, Kosut R L and Lidar D A 2007 [*J. Phys.*]{} B [**40**]{} S103. Grace M, Brif C, Rabitz H, Lidar D A, Walmsley I A and Kosut R L 2007 [*J. Modern Optics*]{} [**54**]{} 2339. Wu R, Pechen A, Rabitz H, Hsieh M and Tsou B\ [*E-print*]{}: http://xxx.lanl.gov/abs/0708.2119 Preskill J 2004 [*Lecture Notes on Quantum Information and Computation*]{}\ [*E-print*]{}: http://www.theory.caltech.edu/people/preskill/ph229 Lloyd S and Viola L 2001 [*Phys. Rev.*]{} A [**65**]{} 010101 Roslund J, Roth M and Rabitz H 2006 [*Phys. Rev.*]{} A [**74**]{} 043414 Spohn H and Lebowitz J L 1978 [*Adv. Chem. Phys.*]{} [**38**]{} 109 Spohn H 1980 [*Rev. Mod. Phys.*]{} [**53**]{} 569 Dümcke R 1985 [*Comm. Math. Phys.*]{} [**97**]{} 331 Accardi L, Pechen A N and Volovich I V 2003 [*Infinite Dimens. Anal. Quantum Probab. Rel. Top.*]{} [**6**]{} 431; [*E-print*]{}: http://xxx.lanl.gov/abs/math-ph/0206032 Accardi L, Lu Y G and Volovich I V 2002 [*Quantum Theory and Its Stochastic Limit*]{} (Berlin: Springer) Pechukas P 1994 [*Phys. Rev. Lett.*]{} [**73**]{} 1060 Shaji A and Sudarshan E C G 2005 [*Phys. Lett.*]{} A [**341**]{} 48 Romano R 2005 [*J. Phys. A: Math. Gen.*]{} [**38**]{} 9105 Choi M-D 1975 [*Linear Algebra and its Applications*]{} [**10**]{} 285 Pontriagin L S 1962 [*Ordinary Differential Equations*]{} (Reading, MA: Addison-Wesley) Arnold V I 1992 [*Ordinary Differential Equations*]{} (Berlin, New York: Springer-Verlag) [^1]: E-mail: apechen@princeton.edu [^2]: E-mail: hrabitz@princeton.edu

ArXiv

--- abstract: 'We study a classical multitype Galton–Watson process with mutation and selection. The individuals are sequences of fixed length over a finite alphabet. On the sharp peak fitness landscape together with independent mutations per locus, we show that, as the length of the sequences goes to $\infty$ and the mutation probability goes to $0$, the asymptotic relative frequency of the sequences differing on $k$ digits from the master sequence approaches $$(\s\exa-1)\frac{a^k}{k!}\sum_{i\geq1}\frac{i^k}{\s^i}\,,$$ where $\s$ is the selective advantage of the master sequence and $a$ is the product of the length of the chains with the mutation probability. The probability distribution $\cQ(\s,a)$ on the non negative integers given by the above formula is the quasispecies distribution with parameters $\s$ and $a$.' bibliography: - 'galwat.bib' --- Distribution of the quasispecies\ for a Galton–Watson process\ on the sharp peak landscape Joseba Dalmau Université Paris Sud and ENS Paris Introduction ============ Understanding the origin of life on Earth has always been a major objective of science. One of the many important contributions to the subject is the 1971 article by Manfred Eigen [@Eig], which theoretically explores the early stages of life on our planet. As an explanation of how evolution of very simple macromolecules might have occurred, Eigen proposed a model known today as Eigen’s model or the quasispecies model. The model aims at describing the evolution of a population of macromolecules, driven both by selection and mutation forces. Due to the simplicity of the reproducing individuals, Eigen’s model can be synthesised as a system of differential equations, obtained from the chemico–kinetic reactions the different macromolecules are subject to: $$x_k'(t)\,=\, \sum_{i=1}^N f(i) Q(i,k) x_i(t) -x_k(t)\sum_{i=1}^N f(i) x_i(t)\,,\qquad 1\,\leq\,k\,\leq\,N\,.$$ Here, the different possible genotypes are numbered from $1$ to $N$; $x_k(t)$ represents the concentration of individuals with genotype $k$ at time $t$; $f(i)$ is the fitness (reproductive rate) of the $i$–th genotype, and $Q(i,k)$ is the probability that an individual with genotype $i$ is transformed into an individual with genotype $k$ by mutation (unfaithful replication during reproduction). Thus, the first term in the differential equation accounts for the production of genotype $k$ individuals, while the second term accounts for the loss of individuals with genotype $k$; the second term is proportional to the concentration of genotype $k$ individuals as well as to the population’s average fitness, and it helps to keep the total concentration of chains constant. One of the simplest scenarios we can consider is that of the sharp peak landscape together with independent mutations per locus. In the sharp peak fitness landscape all sequences but one, the master sequence, have the same fitness, while the master sequence has a higher fitness than the rest. Mutations happen during reproduction independently on each locus of the sequence, with equal probability. Eigen studied this simple scenario and found that two major phenomena take place. The first is an error threshold phenomenon: there is a critical mutation probability such that for above–critical mutation probabilities the equilibrium state of the population is a totally disordered one. The second phenomenon is found for below–critical mutation probabilities: in this case the equilibrium state of the population is no longer disordered; it contains a positive concentration of the master sequence, together with a cloud of mutants that closely resemble the master sequence. This kind of distribution has come to be known as a quasispecies distribution. The concept of error threshold, as well as that of quasispecies, are very appealing to the scientific community, mostly due to their potential for qualitatively explaining the behaviour of a wide range of biological populations. Since Eigen introduced them, it has long been sought to extend the concepts to many other situations, both experimentally and theoretically. From a theoretical point of view, there are two main objections to the applicability of Eigen’s model to more complex kinds of populations. The first objection comes from considering at the same time finite chain length and infinite population size: if the individuals we seek to model are fairly complex, the number of possible genotypes largely exceeds the size of any viable population, a feature that Eigen’s model fails to account for. The second objection is due to the deterministic nature of Eigen’s model: again, for fairly complex individuals, the description of the reproduction mechanism by chemico–kinetic reactions is completely out of reach, and it is typically replaced by some random mechanism. The program is thus settled: to retrieve the error threshold phenomenon and a quasispecies distribution for finite population stochastic models. For a discussion on the several contributions to this program we refer the reader to [@CerfM; @CD]. In the series of papers [@CerfM; @CerfWF; @CD; @Dalmau], the authors studied the classical Moran and Wright–Fisher models, recovering both the error threshold phenomenon and a quasispecies distribution for mutation rates below the error threshold. Furthermore, the quasispecies distribution happens to be the same for both models, and an explicit expression was found: the concentration of sequences differing in exactly $k$ digits from the master sequence is given by $$(\s\exa-1)\frac{a^k}{k!}\sum_{i=1}^\infty \frac{i^k}{\s^i}\,,$$ where $\s>1$ is the reproductive advantage of the master sequence and $a$ is the product of the mutation probability with the length of the sequences. We call this distribution the quasispecies distribution with parameters $\s$ and $a$, and we denote it by $\cQ(\s,a)$. Both the Moran and the Wright–Fisher models are constant population models, since their aim is to describe a sufficiently large population which has stabilised in its environment. However, we might be interested in studying the evolution of a population in its early stages. The size of such a population is very likely to undergo significant fluctuations, the classical stochastic model for this situation is the Galton–Watson branching process. The aim of our article is to study a Galton–Watson branching process, with selection and mutation, in order to recover the phase transition phenomenon and the quasispecies distribution. Demetrius, Schuster and Sigmund [@DSS] already pursued this task in a more general context: a general fitness landscape as well as a general mutation kernel. In [@ABCJ], Antoneli, Bosco, Castro and Janini generalised the work in [@DSS] by studying a multivariate branching process, which incorporates neutral, deleterious and beneficial mutations. Our setting is closer to that of [@DSS] than [@ABCJ]; our aim is to show that for the sharp peak landscape along with per–locus independent mutations, the quasispecies distribution is again the one obtained for the Moran model and for the Wright–Fisher model. In [@DSS], it was proved that the relative frequencies of the genotypes converge to those given by the stationary solution of Eigen’s system of differential equations. However, the quasispecies distribution is a distribution on the Hamming classes of the sequence space, which arises in a particular asymptotic regime. Thus, we cannot apply the results in [@DSS] directly. Along the lines of [@CerfM; @CerfWF; @CD; @Dalmau], we develop our argument from scratch. We start by defining the Galton–Watson process on the genotypes, with selection and mutation. We formally show how to pass from the process on the genotypes to a Galton–Watson process on the Hamming classes. The relative frequencies of the classes are shown to converge to the stationary solution of the corresponding Eigen’s system, as shown in [@DSS]. Finally, the stationary solution to this particular Eigen’s system is shown to converge to the quasispecies distribution. Our article is organised as follows: first we define a multitype Galton–Watson process to model the evolution of a finite population. We state next the main result of the article, and all the remaining sections are devoted to the proof of the main result. The Galton–Watson process ========================= In this section we define a multitype Galton–Watson process driving the dynamics of a finite population, which incorporates both selection and mutation effects. Let us begin by introducing the individuals that will form the population. *Individuals.* Let $\cA$ be a finite alphabet of cardinality $\k\geq1$, and consider sequences of fixed length $\ell\geq1$ over the alphabet $\cA$. A sequence in $\cA^\ell$ represents the genotype of a haploid individual. We study the evolution of a population of such individuals, with selection and mutation. *Sharp peak landscape.* The selection mechanism is given by a fitness function $A:\al\lra\R_+$. Many fitness landscapes might be considered, but we choose to work with the sharp peak landscape: there is a particular sequence ${w^*\in\al}$, called the master sequence, whose fitness is $\s\geq1$, while every other sequence in $\al$ has fitness $1$. So, the fitness function in this case is given by $$\forall u\in\al\qquad A(u)\,=\,\begin{cases} \quad\s&\quad\text{if }\ u=w^*\,,\\ \quad 1&\quad\text{if }\ u\neq w^*\,. \end{cases}$$ *Independent mutations per locus.* Mutations happen randomly due to unfaithful replication of the chains, independently on each locus of the chain, with equal probability $q\in\,]0,1[\,$ for all loci. When an allele mutates, it does so to a randomly chosen letter, uniformly from the $\k-1$ letters still available in the alphabet $\cA$. This mutation mechanism can be encoded into a mutation kernel in the following manner: $$\forall u,v\in\al\qquad M(u,v)\,=\,\prod_{i=1}^\ell \Big( (1-q)1_{u(i)=v(i)} +\frac{q}{\k-1}1_{u(i)\neq v(i)} \Big)\,.$$ *The multitype Galton–Watson process* is a Markov chain with values in $\smash{\N^{\k^\ell}}$, $$X_n\,=\, \big(X_n(u), u\in\al\big)\,,\qquad n\geq 0\,.$$ For each $u\in\al$ and $n\geq 0$, $X_n(u)$ represents the number of individuals with genotype $u$ present in the population at time $n$. At each generation, each individual in the population gives birth to a random number of children, independently of the other individuals and of the past of the process. The number of offspring of an individual $u\in\al$ is distributed as a Poisson random variable with mean $A(u)$. The newborn individuals then mutate according to the kernel $M$. The new generation is formed by all the offspring, after mutation. *Generating functions.* The classical tool for studying the Galton–Watson process we just described is generating functions, which are also useful for formally defining the transition mechanism of the process. Let $u\in\al$ and define the function $f^u:[-1,1]^{\al}\lra\R$ by: $$\forall s\in [-1,1]^{\al}\qquad f^u(s)\,=\, \sum_{r\in\N^{\al}}p^u(r)\prod_{v\in\al}s(v)^{r(v)}\,,$$ where $p^u(r)$ represents the probability that an individual with genotype $u$ has $r(v)$ children with genotype $v$. For each $v\in\al$: $$\forall r\in\N^{\al}\qquad p^u(r)\,=\, e^{-A(u)}A(u)^{|r|_1} \prod_{v\in\al}\frac{M(u,v)^{r(v)}}{r(v)!}\,.$$ Here $|r|_1$ represents the usual 1–norm of the vector $r$, that is, the sum of its components. For an initial population $X_0$ consisting of one genotype $u$ individual only, $X_1$ is a random vector having generating function $f^u$. In general, for $n\geq0$, if $X_n=r\in\nal$, then $X_{n+1}$ is the sum of $|r|_1$ random vectors, where, for each $u\in\al$, $r(u)$ of the random vectors have generating function $f^u$. Note that the null vector is an absorbing state. Main result =========== Since we work with the sharp peak landscape fitness function, we can classify the sequences in $\al$ according to the number of digits they differ from the master sequence. Precisely, the Hamming distance between two sequences $u,v\in\al$ is defined as the number of digits where the two sequences differ: $$d_H(u,v)\,=\,\text{card}\big\lbrace\, i\in\ul : u(i)\neq v(i) \,\big\rbrace\,.$$ For each $k\in\zl$, let $\cC_k$ be the set of the sequences in $\al$ at Hamming distance $k$ from the master sequence: $$\cC_k\,=\,\lbrace\, u\in\al : d_H(u,w^*)=k \,\rbrace\,.$$ We refer to the set $\cC_k$ as the $k$–th Hamming class. Our aim is to study the concentration of the individuals of $X_n$ which are in the class $k$ in the following asymptotic regime: $$\ell\to\infty\,,\qquad\quad q\to 0\,,\qquad\quad \ell q \to a\in [0,\infty]\,.$$ We have the following result. The process $(X_n)_{n\geq 0}$ has a positive probability of survival. Conditioned on the event of non–extinction, if $\s\exa\leq 1$ then $$\forall k\geq 0\qquad \lim_{\genfrac{}{}{0pt}{1}{\ell\to\infty,\,q\to0}{\ell q\to a}} \,\lim_{n\to\infty}\, \frac{1}{|X_n|_1}\sum_{u\in\cC_k}X_n(u)\,=\,0\,.$$ If $\s\exa>1$ then $$\forall k\geq 0\qquad \lim_{\genfrac{}{}{0pt}{1}{\ell\to\infty,\,q\to0}{\ell q\to a}} \,\lim_{n\to\infty}\, \frac{1}{|X_n|_1}\sum_{u\in\cC_k}X_n(u)\,=\, (\s\exa-1)\frac{a^k}{k!} \sum_{i\geq1}\frac{i^k}{\s^i}\,.$$ The right hand side in this equation is the concentration of the $k$–th Hamming class in the distribution of the quasispecies $\cQ(\s,a)$ with parameters $\s$ and $a$. We devote the rest of the paper to the proof of this result. The occupancy process ===================== In this section we build an occupancy process $$\On\,=\,(Z_n(0),\dots,Z_n(\ell))_{n\geq 0}\,,$$ to keep track of the number of sequences in each of the Hamming classes. Here $Z_n(l)$ represents the number of individuals in $X_n$ that are at distance $l$ from the master sequence. In order to build the occupancy process formally, we use the classical lumping technique [@KS]. *Fitness.* The fitness function $A$ can be factorised into Hamming classes: define the function $A_H:\zl\lra\R_+$ by $$\forall l\in\zl\qquad A_H(l)\,=\,\begin{cases} \quad\s&\quad\text{if }\ l=0\,,\\ \quad 1&\quad\text{if }\ 1\leq l\leq \ell\,. \end{cases}$$ Then, for each $u\in\al$ we have $A(u)= A_H(d_H(u,w^*))$. *Mutations.* The mutation matrix $M$ can also be factorised into the Hamming classes. Indeed, for each $u\in\al$ and $c\in\zl$, the value $$\sum_{v\in\cC_c}M(u,v)$$ depends on $u$ through its Hamming class only (lemma 6.1 in [@CerfM]). For $b,c\in\zl$, let us call $M_H(b,c)$ this common value for $u$ in $\cC_b$. The coefficient $M_H(b,c)$ can be analytically expressed as $$\sum_{ \genfrac{}{}{0pt}{1}{0\leq k\leq\ell-b}{ \genfrac{}{}{0pt}{1} {0\leq l\leq b}{k-l=c-b} } } { \binom{\ell-b}{k}} {\binom{b}{l}} q^k (1-q)^{\ell-b-k} \Big(\frac{q}{\kappa-1}\Big)^l \Big(1-\frac{q}{\kappa-1}\Big)^{b-l}\,.$$ *Lumping.* Let $\cZ:\N^{\al}\lra\N^{\ell+1}$ be the map that associates to each population $r\in\N^{\al}$ the corresponding occupancy distribution: $$\forall r\in\N^{\al}\quad \forall l\in\zl\qquad \cZ(r)(l)\,=\, \sum_{u\in\cC_l}r(u)\,.$$ The occupancy process $\On$ is defined by $$\forall n\geq0\qquad \on\,=\,\cZ(X_n)\,.$$ We check next that the occupancy process is again a Galton–Watson process. Let $k\in\zl$, $u\in\cC_k$, and $z\in\nl$. We have $$\begin{aligned} \sum_{\genfrac{}{}{0pt}{1} {r\in\N^{\al}}{\cZ(r)=z}}p^u(r)\,&=\, \sum_{\genfrac{}{}{0pt}{1} {r\in\N^{\al}}{\cZ(r)=z}}e^{-A(u)} A(u)^{|r|_1} \prod_{v\in\al}\frac{M(u,v)^{r(v)}}{r(v)!} \\&=\, e^{-A_H(k)}A_H(k)^{|z|_1} \sum_{\genfrac{}{}{0pt}{1} {r\in\N^{\al}}{\cZ(r)=z}}\prod_{v\in\al} \frac{M(u,v)^{r(v)}}{r(v)!}\,.\end{aligned}$$ Decomposing the last sum and product into Hamming classes, we obtain $$\begin{aligned} \sum_{\genfrac{}{}{0pt}{1} {r\in\N^{\al}}{\cZ(r)=z}}\prod_{v\in\al} \frac{M(u,v)^{r(v)}}{r(v)!} \,&=\, \prod_{l=0}^\ell\bigg( \sum_{\genfrac{}{}{0pt}{1} {t\in\N^{\cC_l}}{|t|_1=z(l)}} %\prod_{l=0}^\ell \prod_{v\in\cC_l}\frac{M(u,v)^{t(v)}}{t(v)!} \bigg) %\\=\,\prod_{l=0}^\ell %\bigg( %\sum_{v\in\cC_l} M(u,v) %\bigg)^{z(l)}\,=\, \\&=\, \prod_{l=0}^\ell \frac{M_H(k,l)^{z(l)}}{z(l)!}\,.\end{aligned}$$ Let $$p^k(z)\,=\,\sum_{\genfrac{}{}{0pt}{1} {r\in\N^{\al}}{\cZ(r)=z}}p^u(r)\,=\, e^{-A_H(k)}A_H(k)^{|z|_1} \prod_{l=0}^\ell \frac{M_H(k,l)^{z(l)}}{z(l)!}\,.$$ Since this expression depends on $u$ only through $k$, the sum $$\sum_{\genfrac{}{}{0pt}{1} {r'\in\nl}{\cZ(r')=z'}} P\big( X_{n+1}=z'\,|\,X_n=r \big)$$ depends on $r$ only through $z=\cZ(r)$. Thus, by the classical lumping theorem, the process $\On$ is a Markov chain (the classical lumping theorem is stated in [@KS] for finite state space Markov chains, but both the result and the proof carry over word by word to the case of denumerable Markov chains). Let us define, for $k\in\zl$ and $s\in [0,1]^{\ell+1}$, $$f^k(s)\,=\,\sum_{z\in\nl}p^k(z)\prod_{l=0}^\ell s(l)^{z(l)}\,.$$ The process $\On$ is in fact a Galton–Watson process with $\ell+1$ types, having the following transition mechanism: for all $n\geq 0$ and $z\in\nl$, if $Z_n=z$, then $Z_{n+1}$ is the sum of $|z|_1$ independent random vectors, where, for each $k\in\zl$, $z(k)$ of the vectors have generating function $f^k$. Proof of theorem 1 ================== We use now the classical theory of branching processes [@Harris] in order to study the process $\Zn$. The mean matrix $W$ of $\Zn$ is the matrix with coefficients $W(i,j),0\leq i,j\leq\ell$, given by the expected number of class $j$ individuals in the first generation, when the process starts with a population consisting of just one individual in the class $i$. The mean matrix is thus given by: $$\forall i,j\in\zl\qquad W(i,j)\,=\,A_H(i)M_H(i,j)\,.$$ The entries of the matrix $W$ are all positive. By the Perron–Frobenius theorem, there exist a unique largest eigenvalue $\l$ of $W$ and a unique positive and unitary eigenvector $\rho$ associated to $\l$. By the general theory of multitype Galton–Watson processes (theorems 7.1 and 9.2 in chapter II of [@Harris]), if $\l\leq 1$ then the population goes extinct with probability 1. If $\l>1$ there is a positive probability of survival, and conditioned on the event of non extinction, we have $$\lim_{n\to\infty}\frac{Z_n(k)}{Z_n(0)+\dots+Z_n(\ell)}\,=\, \rho(k)\,,\qquad 0\leq k\leq\ell\,.$$ From their definition, $\l$ and $\rho$ satisfy $$\l\rho(k)\,=\, \sum_{i=0}^\ell \rho(i)A_H(i)M_H(i,k)\,,\qquad 0\leq k\leq\ell\,.$$ Summing the above expression over $k$, since $\rho$ is a unitary vector, we deduce that $$\l\,=\, \sum_{i=0}^\ell\rho(i)A(i)\,=\, (\s-1)\rho(0)+1\,.$$ Thus, the eigenvalue $\l$ is equal to the average fitness of a population whose concentrations are given by the vector $\rho$. We remark that solving the above system of equations is equivalent to finding the stationary solutions of the corresponding Eigen’s system of differential equations. From the above equation, we see that, in particular, $\l\in\,]1,\s[\,$, and this implies the first statement of the theorem: the process $\Zn$ has a positive probability of survival. It remains to study the asymptotic behaviour of $\l$ and $\rho$ when $\ell$ goes to $\infty$, $q$ goes to $0$ and $\ell q$ goes to $a$. In this asymptotic regime the mutation kernel $M_H$ converges to the following limiting expression: $$\forall i,k\geq0\qquad \lim_{\genfrac{}{}{0pt}{1}{\ell\to\infty,\,q\to0}{\ell q\to a}}\, M_H(i,k)\,=\,\begin{cases} \quad \displaystyle\exa\frac{a^{k-i}}{(k-i)!} &\quad\text{if }\ k\geq i\,,\\ \quad 0 &\quad\text{if }\ k< i\,. \end{cases}$$ Up to extraction of a subsequence, we can suppose that the following limits exist $$\l^*\,=\, \lim_{\genfrac{}{}{0pt}{1}{\ell\to\infty,\,q\to0}{\ell q\to a}}\, \l\,,\qquad \rho^*(k)\,=\, \lim_{\genfrac{}{}{0pt}{1}{\ell\to\infty,\,q\to0}{\ell q\to a}}\, \rho(k)\,,\quad k\geq 0\,.$$ Writing down the first equation of the system $\l\rho=\rho^T W$, we see that $$\s\rho M_H(0,0)\,<\, \l\rho(0)\,<\, \s\rho M_H(0,0)+ \max_{1\leq i\leq\ell}M_H(i,0)\,.$$ Since we also know that $\l>1$, we conclude that $\l^*\geq\max\lbrace\,1,\s\exa\,\rbrace$. As we have already pointed out, $$\l\,=\,(\s-1)\rho(0)+1\,.$$ Thus, taking the limits in the above two equations we deduce that $$\l^*\,=\,(\s-1)\rho^*(0)+1\,,\qquad \l^*\rho^*(0)\,=\,\s\rho^*(0)\exa\,.$$ Since $\l^*\geq\max\lbrace\,1,\s\exa\,\rbrace$, we conclude that $\bullet$ if $\s\exa\leq1$, then $\l^*=1$ and $\rho^*(0)=0$. $\bullet$ if $\s\exa>1$, then $$\l^*\,=\,\s\exa\qquad \text{and}\qquad \rho^*(0)\,=\,\frac{\s\exa-1}{\s-1}\,.$$ Finally, writing down the $k$–th equation of the system $\l\rho=\rho^T W$, we see that $$\begin{gathered} \s\rho_0M_H(0,k)+\sum_{i=1}^k\rho(i)M_H(i,k) \,<\,\l\rho(k)\,<\,\\ \s\rho_0M_H(0,k)+\sum_{i=1}^k\rho(i)M_H(i,k) +\max_{k<i\leq\ell}M_H(i,k)\,.\end{gathered}$$ Thus, taking the limit we obtain the recurrence relation $$\s\exa\rho^*(k)\,=\, \s\rho^*(0)\exa\frac{a^k}{k!} +\sum_{i=1}^k \rho^*(i)\exa\frac{a^{k-i}}{(k-i)!}\,,\qquad k\geq1\,.$$ We conclude that if $\s\exa\leq1$, then $\rho^*(k)=0$ for all $k\geq0$, and if $\s\exa>1$, then $$\rho^*(k)\,=\,(\s\exa-1)\frac{a^k}{k!} \sum_{i\geq1}\frac{i^k}{\s^i}\,,\qquad k\geq0\,.$$ This can be seen by solving the recurrence relation by the method of generating functions, see, for example [@CD].

ArXiv

--- abstract: 'We use particle data from the [Illustris]{} simulation, combined with individual kinematic constraints on the mass of the Milky Way (MW) at specific distances from the Galactic Centre, to infer the radial distribution of the MW’s dark matter halo mass. Our method allows us to convert any constraint on the mass of the MW within a fixed distance to a full circular velocity profile to the MW’s virial radius. As primary examples, we take two recent (and discrepant) measurements of the total mass within 50 [[kpc]{}]{} of the Galaxy and find that they imply very different mass profiles and stellar masses for the Galaxy. The dark-matter-only version of the [Illustris]{} simulation enables us to compute the effects of galaxy formation on such constraints on a halo-by-halo basis; on small scales, galaxy formation enhances the density relative to dark-matter-only runs, while the total mass density is approximately 20% lower at large Galactocentric distances. We are also able to quantify how current and future constraints on the mass of the MW within specific radii will be reflected in uncertainties on its virial mass: even a measurement of ${M(<\!50\,{\rm kpc})}$ with essentially perfect precision still results in a 20% uncertainty on the virial mass of the Galaxy, while a future measurement of ${M(<\!100\,{\rm kpc})}$ with 10% errors would result in the same level of uncertainty. We expect that our technique will become even more useful as (1) better kinematic constraints become available at larger distances and (2) cosmological simulations provide even more faithful representations of the observable Universe.' author: - | Corbin Taylor$^1$[^1], Michael Boylan-Kolchin$^2$[^2], Paul Torrey$^{3, 4}$, Mark Vogelsberger$^{3}$, and Lars Hernquist$^{5}$\ $^1$Department of Astronomy, University of Maryland, 1113 Physical Sciences Complex (Building 415), College Park, MD 20742-2421, USA\ $^2$Department of Astronomy, The University of Texas at Austin, 2515 Speedway, Stop C1400, Austin, TX 78712-1205, USA\ $^{3}$Department of Physics, Kavli Institute for Astrophysics and Space Research, Massachusetts Institute of Technology, Cambridge, MA 02139, USA\ $^{4}$TAPIR, Mailcode 350-17, California Institute of Technology, Pasadena, CA 91125, USA\ $^{5}$Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA, 02138, USA bibliography: - 'ms\_clean.bib' title: The Mass Profile of the Milky Way to the Virial Radius from the Illustris Simulation --- \[firstpage\] Galaxy: fundamental parameters–Galaxy: halo–Galaxy: structure–dark matter. Introduction ============ While living within the Milky Way (MW) galaxy does have its virtues, easily and accurately determining the mass distribution of the Galaxy’s dark matter halo is not one of them. This is not for lack of trying, naturally; a variety of techniques have been crafted for just this purpose, and multiple classes of kinematic tracers are available. The difficulty in measuring the MW’s mass distribution is two-fold. First, only line-of-sight information is available for the vast majority of kinematic measurements. While great strides are being made in measuring the proper motions of both individual stars [@cunningham2015] and dwarf galaxies (e.g., @Piatek2007 [@Sohn2013; @vanDerMarel2014; @pryor2015]) at large Galactocentric distances in the MW’s halo, the number of tracers at $\sim 50-100$ kpc with full 6D phase space information will remain small even in the Gaia era [@deBruijne2014]. Perhaps more importantly, the level of precision desired for the MW’s mass is simply higher than is the case for other galaxies. Whereas a factor of $\pm2$ uncertainty in the mass of a typical galaxy’s halo would be considered an excellent measurement, it is often thought of more as an embarrassment in the case of the MW. For example, if we take a dark matter halo mass of $10^{12}\,{M_{\odot}}$ as a fiducial estimate for the MW, changes by a factor of 2 in either direction are the difference between: (1) an implied conversion efficiency of baryons into stars of $\approx 70\%$ (at $M=5\times 10^{11}\,{M_{\odot}}$) and $16\%$ (at $2 \times 10^{12}\,{M_{\odot}}$); (2) eliminating the too-big-to-fail problem [@Boylan-Kolchin2011a; @Boylan-Kolchin2012] and severely exacerbating it [@Wang2012; @Vera-Ciro2013; @Jiang2015]; and (3) placing the Large Magellanic Cloud and the Leo I dwarf spheroidal on unbound versus bound orbits [@kallivayalil2006; @kallivayalil2013; @besla2007; @Boylan-Kolchin2013]. Our understanding of the MW in cosmological context relies on our ability to know its mass to high precision. While uncertainties are most pronounced in the outer dark matter halo of the MW, where there are few tracers of the total mass, they also persist at small Galactocentric distances: there are disagreements about the mass within the solar circle at the 25% level (e.g., @Bovy2012 [@schonrich2012]). At 40–80 kpc, estimates differ at the 50% level (see, e.g., @Williams2015). In this paper, we take an alternate approach to constraining the mass distribution of the MW. Cosmological hydrodynamic simulations are now producing galaxies that match a variety of observations both for statistical samples of galaxies and for individual galaxies themselves. In particular, both the [Illustris]{} [@Vogelsberger2014a] and Eagle [@Schaye2015] simulations use $\sim 10^{10}$ particles within $\sim 100$ Mpc boxes, meaning they contain thousands of haloes with masses comparable to that of the MW, each with of the order 1 million particles within the virial radius. The successes of these models, and the underlying successes of the $\Lambda$ cold dark matter ([$\Lambda$CDM]{}) model, motivate using the results of cosmological simulations to constrain the mass distribution of the MW. There are a number of ways one could use cosmological simulations for this purpose. Indeed, several previous works on the mass of the MW have used cosmological simulations in some capacity. One possibility is to use dark matter haloes from large cosmological simulations as point particles and calibrate the timing argument [@Kahn1959] for measuring the total mass of the MW [@Li2008; @Gonzalez2014]. Alternately, properties of satellites from cosmological simulations can be compared to those of MW satellites such as the Magellanic Clouds, yielding estimates of the virial mass of the MW [@Boylan-Kolchin2011b; @Busha2011; @Gonzalez2013; @Fattahi2016]. Yet another possibility is to use individual, high-resolution simulations of Milky Way-sized haloes in conjunction with kinematic information about dwarf satellites of the MW [@Boylan-Kolchin2013; @Barber2014]. Cosmological hydrodynamic simulations of individual MW-mass haloes have also been used to calibrate kinematic analyses of tracer populations in order to measure the mass of the MW [@Xue2008; @Rashkov2013; @Piffl2014; @Wang2015]. Our approach is to use importance sampling in a homogeneously-resolved, large-volume cosmological simulation, weighing each simulated halo by its level of consistency with the MW; for a clear description of this technique applied to cosmological simulations, see [@Busha2011]. By taking any individual constraint and using it to perform importance sampling from simulations, we can find the mass distributions of haloes that are consistent with the imposed constraint. An advantage of this technique is that it allows us to easily map different constraints, with different errors, on to mass distributions for the MW and its dark matter halo. Variants of importance sampling have been used to measure the mass of the MW [@Li2008; @Boylan-Kolchin2011b; @Busha2011; @Gonzalez2014]. However, previous work has generally focused on using dark-matter-only (DMO) simulations to measure the total (virial) mass of the MW. With hydrodynamic simulations, we are able to make two improvements. First, we are able to measure the mass *distribution* of the MW in simulations that self-consistently model the effects of galaxy formation on the dark matter haloes of galaxies. Secondly, we are able to compare our constraints directly to those obtained from DMO simulations, as a DMO version of [Illustris]{} is also publicly available. By matching objects between the two simulations, we are able to investigate, in detail, the effects of baryonic physics on inferences of the mass distribution of the MW from cosmological simulations. We generally use the mass within 50 kpc as our primary constraint, as this is approximately the largest radius where stellar kinematic tracers are found in large enough numbers to facilitate a mass measurement. We also provide estimates for how a measurement of the mass within 100 kpc – which future surveys may provide – will improve our knowledge of the mass distribution at even larger radii. This paper is structured as follows. Section \[sec:methods\] describes our basic approach, provides information about the Illustris simulation, and describes our primary analysis of the simulation. Section \[sec:results\] contains our main results regarding the mass distribution of the MW as derived from haloes taken from the [Illustris]{} simulation. We also quantify how inferences on the enclosed mass at large scales (at 250 kpc and various spherical overdensity values) depend on the measured mass within 50 kpc and quantify the stellar masses of galaxies having haloes consistent with the adopted mass constraint. A discussion of our results and prospects for future improvements is given in Section \[sec:discussion\]; our primary conclusions are given in Section \[sec:conclusions\]. Throughout this paper, error bars give 68% confidence intervals unless otherwise noted. Methods {#sec:methods} ======= Simulations and Importance Sampling {#sec:simulations} ----------------------------------- Our analysis is based on the [Illustris]{} suite of cosmological simulations [@Vogelsberger2014a], which consists of paired hydrodynamic and DMO simulations at three different resolution levels. Each simulation uses a periodic box of length $75\,h^{-1}$ Mpc and an initial redshift of $z=127$. The highest resolution simulation, [Illustris]{}-1, uses $1820^3$ dark matter particles and an equal number of hydrodynamic cells initially, with a spatial resolution of $1\,h^{-1}$ kpc for the dark matter. The DMO version of this simulation, [Illustris]{}-Dark-1, uses identical initial conditions but treats the baryonic component as collisionless mater. Two lower resolution simulation of the same volume, [Illustris]{}-2 and [Illustris]{}-3, were also performed, with 8 and 64 times fewer particles, respectively. The background cosmology for all of the simulations was chosen to be consistent with Wilkinson Microwave Anisotropy Probe-9 results [@Hinshaw2013]: $\Omega_{\rm m,0} = 0.2726$, $\Omega_{\Lambda, 0} = 0.7274$, $\Omega_{\rm b,0} = 0.0456$, $\sigma_{8} = 0.809$, $n_{\rm s} = 0.963$, and $h = 0.704$. Haloes and subhaloes in the [Illustris]{} simulations were identified using a friends-of-friends algorithm followed by SUBFIND [@Springel2001]. For further information about the [Illustris]{} suite,[^3] including details about the implementation of galaxy formation physics, see @Vogelsberger2013 [@Vogelsberger2014b]. Using Illustris to inform our understanding of the mass distribution of the MW requires calculations of the mass profiles of an unbiased sample of dark matter haloes within the simulation. Although the halo catalogues provide the centres for each halo (we only consider central halos, not subhaloes, as possible centres), a brute-force calculation of the mass profile for each halo is prohibitively expensive, as it requires repeated searches through the $\sim 10^{10}$ particles of the simulation. We instead use a K-D tree algorithm, taking into account the periodic boundary conditions of the simulation volume. The algorithm was verified against brute-force calculations applied to Illustris-3 and Illustris-2. In addition to considering the mass within spherical apertures, we also compute spherical overdensity masses with respect to three common overdensity choices: $M_{\rm 200,c}$ (measured with respect to $200\,\rho_{\rm crit}$), $M_{\rm 200,m}$ (measured with respect to $200\,\rho_{\rm m}\approx 55\,\rho_{\rm crit}$ for the Illustris cosmology at $z=0$), and $M_{\rm vir}$ (measured with respect to $\Delta_{\rm vir}\,\rho_{\rm crit}$; for the Illustris cosmology at $z=0$, $\Delta_{\rm vir} \approx 97$; @Bryan1998). {width="0.48\linewidth"} {width="0.48\linewidth"} Statistical Analysis -------------------- Our basic framework is to consider the Illustris simulation a plausible model of galaxies in our Universe, then to assign each halo in the simulation a weight based on how closely its enclosed mass at some radius[^4] (we typically use 50 kpc in what follows) matches observational data. The resulting weights for the halo sample then provide a constraint on the enclosed mass of the MW at other radii. In more detail, we take an observational measurement of the total MW mass within a specific radius and assign a weight to each halo in the [Illustris]{}galaxy catalog: assuming the observed mass has a value of $\mu$ and an associated (Gaussian) error of $\sigma$, then the weight $W_i$ contributed by an individual halo $i$ with enclosed mass $M_i$ at the specified radius is $$\centering W_i = \frac{1}{\sqrt{2\rm{\pi}}\sigma} \exp\left(\frac{-(M_i-\mu)^{2}}{2\sigma^{2}}\right)\,. \label{eq:singleprob}$$ We can then construct the full mass or circular velocity profile and compute the total stellar or halo mass that is consistent with the observed constraint by using the distribution of weights assigned to the haloes. In this analysis, we assume that observed constraints all follow Gaussian distributions, consistent with the analyses we incorporate, but this technique can be easily extended to any other analytic or numerical probability distribution. In what follows, we quote median values and confidence intervals that are centred on the median and contain 68% of the probability distribution. The primary observational constraint we use is the total mass of the MW within 50 kpc, ${M(<\!50\,{\rm kpc})}$. There are many literature estimates of the MW’s mass at approximately this scale (e.g., @Wilkinson1999 [@battaglia2005; @Xue2008; @brown2010; @gnedin2010; @Kafle2014; @Eadie2015]), in large part because (1) this is approximately the distance to which large samples of blue horizontal branch (BHB) stars are currently available from surveys such as the Sloan Digital Sky Survey, and (2) the LMC lies at a Galactocentric distance of $\approx 50\,{{\rm kpc}}$, meaning estimates of the MW mass based on LMC’s dynamics directly constrain ${M(<\!50\,{\rm kpc})}$. We focus on two recent and disparate measurements of ${M(<\!50\,{\rm kpc})}$: @deason2012 [hereafter D12], who used BHB stars and found ${M(<\!50\,{\rm kpc})}=4.2 \pm 0.4 \times 10^{11}\,{M_{\odot}}$, and @gibbons2014 [hereafter G14], who used the Sagittarius stream to measure ${M(<\!50\,{\rm kpc})}=2.9 \pm 0.4 \times 10^{11}\,{M_{\odot}}$ [@Gomez2015]. These measurements are clearly incompatible at the $3\,\sigma$ level and therefore are useful for showing the effects of varying ${M(<\!50\,{\rm kpc})}$ on the inferred mass distribution at larger radii. In Sec. \[subsec:mass\_at\_50\], we explicitly show how estimates of $M_{\rm 200,c}$ vary as a function of ${M(<\!50\,{\rm kpc})}$. In principle, a complete analysis would include every dark matter halo in Illustris. In practice, however, only a relatively narrow range of masses contribute any weight to our inferences. We therefore restrict our analysis to all haloes with $M_{\rm 200,c} = (0.1-10)\times10^{12}{M_{\odot}}$, which includes $14\,192$ haloes for Illustris-1, $14\,316$ haloes for Illustris-2, and $12\,885$ haloes for Illustris-3. As we show below, this mass range is more than sufficient for including all relevant haloes in our analysis and does not bias our results in any way. The Mass Distribution of the MW {#sec:results} =============================== The MW’s radial mass profile {#subsec:minr} ---------------------------- {width="0.495\linewidth"} {width="0.495\linewidth"} Fig. \[mInR\] presents the mass distributions obtained using the constraints on ${M(<\!50\,{\rm kpc})}$ from D12 (left-hand panel) and G14 (right-hand panel) from the Illustris-1 sample. The best-fitting Navarro-Frenk-White ([-@Navarro1997 hereafter, NFW]) profiles for the *total* mass distribution are given in the figure as well. The fits were performed over the radial range of 40–300 kpc, as we find a lack of convergence among different resolution versions of [Illustris]{} on smaller scales (see below; convergence in density profiles should occur at smaller scales, as density is a differential quantity while mass and circular velocity are cumulative quantities). Unsurprisingly, given the significantly higher value of ${M(<\!50\,{\rm kpc})}$ found in D12 relative to G14, the best-fitting NFW value of $M_{\rm 200,c}$ for D12 is much larger than for G14, $1.1\times10^{12}\,{M_{\odot}}$ versus $0.61\times 10^{12}\,{M_{\odot}}$. The best fitting concentration parameters are similar: $c_{\rm 200,c}=12.2\pm2.12$ for D12 and $c_{\rm 200,c}=13.2\pm 3.57$ for G14. Both of these concentrations are larger than those derived from large DMO simulations, which typically find $c_{\rm 200,c}\approx 8.33$ for haloes of $M_{\rm200,c} \approx 10^{12}\,{M_{\odot}}$ (e.g., @dutton2014). The lower panels of the figures show the fractional differences of Illustris-2 and Illustris-3 with respect to their high-resolution counterpart, with error bars representing 68% confidence intervals. There are relatively large differences between the different levels of resolution at relatively small radii ($r<30$ kpc), while differences are much less substantial farther away from Galactic Centre. With a gravitational softening length $\sim$ 1 kpc and baryonic sub-grid routines tailored specifically to the highest resolution simulation. This lack of convergence on small scales is not surprising. For instance, @Schaller2016 show that the dark matter density profiles of Eagle galaxy haloes are only converged at $\approx 20$ kpc (their fig. 3). We therefore strongly caution against extrapolating the NFW fits presented in this paper to small radii ($r \la 30$ kpc). If future generations of simulations provide well-converged results at smaller radii, the dark matter fraction within $\sim 2$ disc scale lengths will likely provide important constraints on feedback models [@Courteau2015]. The circular velocity profiles, $V_{\rm circ}(r)$, corresponding to the cumulative mass profiles of Fig. \[mInR\] are shown in the left-hand panel of Fig. \[vcirc\]. This highlights the large difference in the two determinations of the MW potential, as well as how this difference persists in predicted profiles out to 300 kpc. It is only at distances $>250$ kpc that the 68% confidence intervals begin to overlap. The distribution of mass among dark matter, stars, and gas within any given radius is interesting to consider: observationally, we can measure the stellar mass with reasonable accuracy and infer the dark matter mass, but constraining the distribution of the Galaxy’s gaseous component at large distances is much more difficult (see, e.g., @Gupta2012 [@Fang2013]). In the right-hand panel of Fig. \[vcirc\], we plot the circular velocity profile decomposed into the contributions from each of these components. Dark matter dominates the potential at all radii we study, and while stars substantially outweigh the gas for $r \la 50$ kpc, the two contribute approximately the same mass by $r \approx 100$ kpc. Mass constraints within specific radii {#subsec:mass_at_50} -------------------------------------- In this subsection, we explore predictions for enclosed masses at specific radii in more detail. In particular, we are interested in understanding how observational constraints at ${M(<\!50\,{\rm kpc})}$ translate into inferences on masses at other radii. We consider both individual physical radii (in particular, 100 and 250 kpc) and various definitions of spherical overdensity masses ($M_{\rm 200,c}$, $M_{\rm vir}$, and $M_{\rm 200,m}$). ![The probability distribution of ${M(<\!100\,{\rm kpc})}$ derived from the G14 (squares, connected by dashed lines) and D12 (circles, connected by solid lines) constraints on ${M(<\!50\,{\rm kpc})}$. The colours represent the individual resolution levels: Illustris-1 (black), Illustris-2 (blue), and Illustris-3 (red). The excellent agreement across the three levels of resolution indicates that the total mass profiles are well-converged in Illustris.[]{data-label="m100-m50"}](figure_3.pdf){width="\linewidth"} Fig. \[m100-m50\] presents the probability distribution for ${M(<\!100\,{\rm kpc})}$, with black, blue, and red symbols representing Illustris-1, Illustris-2, and Illustris-3 respectively. The results using the D12 constraint on ${M(<\!50\,{\rm kpc})}$ are presented as circles connected with solid lines, while those using the G14 constraint are shown as squares with dashed connecting lines. As expected, and shown previously, the D12 constraint on ${M(<\!50\,{\rm kpc})}$ results in a significantly higher predicted total mass within 100 kpc (approximately 0.2 dex). The smaller (relative) error quoted in D12 also results in a narrower distribution for ${M(<\!100\,{\rm kpc})}$. Perhaps the most important aspect of Fig. \[m100-m50\] is the *excellent* convergence seen across the three Illustris simulations (a factor of 64 in mass resolution and 4 in force resolution). Not only is the peak or median value well converged, the entire distribution is essentially identical in each case. This indicates that, while masses on small scales (10–30 kpc) are affected by resolution and baryonic physics, enclosed masses at larger radii are not subject to such effects. The consistency of the mass distributions at large radii, subject to a constraint at 50 kpc, points to robustness of our technique for constraining the mass distribution of the MW. \[table:confidenceIntervals\] Inferred values of aperture masses within 100 and 250 kpc and three different spherical overdensity masses, along with 68% and 90% confidence intervals, are given in Table \[table:confidenceIntervals\]. The estimated virial mass, ${M_{\rm{vir}}}$, using D12 is $1.3 \times 10^{12}\,{M_{\odot}}$, with a 90% confidence interval of $0.86-2.3\times 10^{12}\,{M_{\odot}}$. This is similar to the result of @Boylan-Kolchin2013, who found a 90% confidence interval of $1.0-2.4\times10^{12}\,{M_{\odot}}$ for ${M_{\rm{vir}}}$ based on the dynamics of the Leo I satellite galaxy. Using the G14 estimate of ${M(<\!50\,{\rm kpc})}$, we find a median value of ${M_{\rm{vir}}}=0.71 \times 10^{12}\,{M_{\odot}}$ with a 90% confidence interval of $0.44-1.2 \times 10^{12}\,{M_{\odot}}$, both of which are substantially lower than our inference based on the results of D12. These results highlight the importance of accurate determinations of ${M(<\!50\,{\rm kpc})}$ for understanding the large-scale properties of the MW. We note that the 99.95% confidence interval for haloes consistent with the D12 constraint is $5.17\times 10^{11}< M_{\rm 200,c}<5.06\times10^{12}\,{M_{\odot}}$ (the range for the G14 constraint is $2.12\times 10^{11}< M_{\rm 200,c}<3.48\times10^{12}\,{M_{\odot}}$), confirming that our range of $10^{11} \le M_{\rm 200, c} \le 10^{13}\,{M_{\odot}}$ is more than sufficient for inferences about the mass of the MW. ![The dependence of the inferred value of $M_{\rm 200,c}$ on the input (measured) value of ${M(<\!50\,{\rm kpc})}$. The data points with error bars show values of $M_{\rm 200,c}$ based on our weighting procedure, assuming a 10% error in ${M(<\!50\,{\rm kpc})}$. The G14 and D12 determinations of ${M(<\!50\,{\rm kpc})}$ are highlighted with yellow and magenta vertical bands, respectively, with the widths of the bands showing the 68% confidence intervals. The best-fitting log-quadratic relation between ${M(<\!50\,{\rm kpc})}$ and $M_{\rm 200,c}$ (given in equation \[eq:mfifty\_fit\]) is plotted as a solid grey line, while the dashed grey line shows the fit to the unweighted data; see the text for details. This relation can be used to map any constraint on ${M(<\!50\,{\rm kpc})}$ to an inferred value of $M_{\rm 200,c}$.[]{data-label="mcrit-function"}](figure_4.pdf){width="\linewidth"} Given the uncertainties in ${M(<\!50\,{\rm kpc})}$, it is also important to understand how inferences of spherical overdensity masses depend on ${M(<\!50\,{\rm kpc})}$. To do this, we assume that ${M(<\!50\,{\rm kpc})}$ can be measured with an accuracy of 10% (i.e., $X \pm 0.1\,X$) and compute the resulting median value and 68% confidence intervals for $M_{\rm 200,c}$. The resulting dependence of $M_{\rm 200,c}$ on ${M(<\!50\,{\rm kpc})}$ is shown in Fig. \[mcrit-function\], where the error bars show 68% confidence intervals. It is clear that there is a strong correlation between ${M(<\!50\,{\rm kpc})}$ and $M_{\rm 200,c}$. We fit this with a quadratic function in log space: $$\begin{aligned} &&\log_{10}\left(\frac{M_{\rm 200,c}}{{M_{\odot}}}\right) = A +B\,\mu+C\,\mu^2\,,\label{eq:mfifty_fit} \\ &&\mu = \log_{10}\left(\frac{M(<50\,{\rm kpc})}{4\times10^{11}\,{M_{\odot}}}\right)\,. \nonumber \end{aligned}$$ Fitting to the weighted results plotted in Fig. \[mcrit-function\], we find $A=12.0, \,B=1.60, \,C=0.373$ with an rms scatter of 0.069, whereas fitting the unweighted data, we find $A=12.0, \,B=1.62, \,C=0.325$ with an rms scatter of 0.067. The latter is offset slightly higher at fixed ${M(<\!50\,{\rm kpc})}$, as the weighted results naturally involve averaging over the dark halo mass function within each bin, which is a steeply declining function of mass, whereas the unweighted results do not. Equation \[eq:mfifty\_fit\] can be used to convert any constraint on ${M(<\!50\,{\rm kpc})}$ into a constraint on $M_{\rm 200,c}$. It is also straightforward to convert this fit to a constraint on $M_{\rm vir}$ or $M_{\rm 200,m}$, as ${M_{\rm{vir}}}\approx 1.17\, M_{\rm 200,c}$ and $M_{\rm 200,m}\approx 1.32\,M_{\rm 200,c}$ for the typical mass profiles in [Illustris]{}. If Equation \[eq:mfifty\_fit\] or a similar relation holds broadly for other hydrodynamic simulations with different galaxy formation physics implementations, then it will be of tremendous value for MW mass inference studies. We plan to examine this issue in more detail in future work (and see further discussion below). The Impact of Baryonic Physics {#subsec:baryonic_physics} ------------------------------ Our primary analysis, presented over the previous subsections, makes use of the highest resolution [Illustris]{} simulation. This, and all other hydrodynamic simulations of the evolution of a representative galaxy population over cosmic time, require a number of assumptions in order to produce a realistic set of galaxies. One of the primary calibrations for [Illustris]{}, for example, was to match the $z=0$ galaxy stellar mass function. As shown in fig. 7 of [@Vogelsberger2014b], the galaxy formation prescriptions in [Illustris]{}result in notable changes in the total masses of dark matter haloes over a wide range in halo mass. Moreover, these changes depend on specific choices made in the galaxy formation modelling, as the galaxy formation modelling within the Eagle simulation results in substantially different effects on halo masses (see fig. 1 of @Schaller2015). It is not a priori obvious whether using the DMO run should result in similar or different predictions from the fully hydrodynamic simulation, and if the results are different, it is not clear whether they will be higher or lower. Certainly, we expect that the formation of a galaxy will lead to a more centrally concentrated mass distribution relative to the DMO run, to some extent. Adiabatic contraction of the dark matter in response to gas cooling will also tend to increase the amount of dark matter in the central regions of the halo. On the other hand, it is well established that galaxy formation must be inefficient in [$\Lambda$CDM]{} (e.g., @fukugita2004), meaning that only a relatively small fraction of the baryonic allotment of a dark matter halo ($\sim 20\%$ for MW-mass haloes) will be converted into stars by $z=0$. Strong feedback from galaxy formation can change the structure of dark matter haloes, reducing their mass within a given radius compared to what would be obtained in a DMO version (e.g., @Vogelsberger2014a [@Schaller2015]). It is therefore of great interest to study precisely how inferences about the MW’s mass profile change from using DMO simulations – which, for given cosmological parameters, are uniquely predicted – to using cosmological hydrodynamic simulations. ![The cumulative mass distribution from [Illustris]{}-Dark-1 (black points with error bars), along with the best-fitting NFW profile (grey line), derived assuming the D12 constraint on ${M(<\!50\,{\rm kpc})}$. This figure can be directly compared to the left-hand panel of Fig. \[mInR\], which shows the same quantities from the full hydrodynamic run. While both versions of Illustris are well-fitted by NFW profiles, the fit parameters differ substantially between the two: the DMO run is fitted by a higher-mass (37% higher), lower-concentration (40% lower) halo. If DMO runs are used for modelling the MW mass distribution based on ${M(<\!50\,{\rm kpc})}$, or a similar constraint, this effect must be taken into account.[]{data-label="mInR-nfw"}](figure_5.pdf){width="\linewidth"} \[table:confidenceIntervals\_dmo\] The first test we perform to gauge the effects of including galaxy formation physics on mass inferences is to rerun our analysis on the DMO versions of [Illustris]{}. Fig. \[mInR-nfw\] shows the results of applying the D12 constraint to [Illustris]{}-Dark-1. It can be directly compared to Fig. \[mInR\], in which the D12 constraint was applied to the hydrodynamic version of [Illustris]{}-1. Relative to the full [Illustris]{} simulation, inferences based on the DMO version result in a significantly higher estimate of $M_{\rm 200,c}$ ($1.5\times 10^{12}$ versus $1.1\times 10^{12}\,{M_{\odot}}$) and a significantly lower version of the NFW concentration ($c=7.4$ versus 12.3). Table \[table:confidenceIntervals\_dmo\] provides an alternate version of Table \[table:confidenceIntervals\] in which all constraints are obtained using the DMO version of [Illustris]{}-1. In all cases, the net effect of using the DMO run rather than the hydrodynamic version is to infer *higher* values for a given aperture mass. We can use the [Illustris]{} suite to perform an additional test of the effects of galaxy formation on the mass distribution within dark matter halos (and for accompanying inferences on the mass distribution of the MW): since [Illustris]{}and [Illustris]{}-Dark share the same initial conditions, individual dark matter halos can be matched between the two simulations (for details, see section 3.2 of @Vogelsberger2014b). In this way, we can study the effects of galaxy formation on a halo-by halo basis by identifying the DMO analogue of each halo in the full [Illustris]{} run and comparing the resulting mass distributions. {width="0.49\linewidth"} {width="0.49\linewidth"} Fig. \[dark-v-hydro\] shows the results of this comparison, for which we use haloes in [Illustris]{}-1 falling within the 68% confidence interval of $M_{\rm 200,c}$ computed using the D12 constraint (see Table \[table:confidenceIntervals\]) – assigning equal weight to all such haloes – and their counterparts in the DMO run. The left-hand panel shows how the density profiles are affected at each radius. On small scales ($r \la 30\,{{\rm kpc}}$), the hydrodynamic run has higher densities on a halo-by-halo basis. This is caused by the formation of the central galaxy, both through its mass and through any adiabatic contraction. On larger scales ($r \ga 40 \,{{\rm kpc}}$), a given halo in the hydrodynamic run is less dense than its equivalent in the DMO run by approximately 20%. This reduction in density is likely caused by outflows and the loss of gas mass (or the prevention of gas accretion). The effect on the cumulative mass distribution is shown in the right-hand panel of Fig. \[dark-v-hydro\]. On a halo-by-halo basis, the hydrodynamic run results in larger masses out to $\approx 100$ kpc; on larger scales, the masses in the DMO run are larger, with the difference reaching an asymptotic value of $\approx 10\%$ at 250-300 kpc. As discussed in Section \[sec:discussion\], the details of the reduction in mass may depend on the adopted models of galaxy formation modelling. The Stellar Mass of the Galaxy ------------------------------ D12 G14 ------------- ------------------------------------------ ------------------------------------------ Illustris-1 $5.04^{+1.47\,(2.72)}_{-1.32\,(2.03)}$ $2.41^{+0.98 \,(1.74)}_{-0.72\, (1.12)}$ Illustris-2 $4.09^{+1.19 \, (2.15)}_{-1.15\,(1.66)}$ $1.83^{+0.76\,(1.36)}_{-0.59\, (0.90)}$ Illustris-3 $2.57^{+0.79\,(1.52)}_{-0.71\,(1.07)}$ $1.03^{+0.48\,(0.92)}_{-0.35\,(0.55)}$ : Inferred values of ${M_{\star}}$, in units of $10^{10}\,{M_{\odot}}$, using the D12 (column 2) and G14 (column 3) constraints on ${M(<\!50\,{\rm kpc})}$. The quoted errors are the 68% and 90% confidence intervals. The [Illustris]{} feedback prescriptions were calibrated for the highest-resolution simulation ([Illustris]{}-1), so perfect convergence in ${M_{\star}}$ across the three simulations is not expected. \[table:stellarMass\_mfifty\] We can also use the technique explored in the previous sections to compute the galaxy stellar masses from [Illustris]{} that are consistent with the adopted mass constraints at 50 kpc. Table \[table:stellarMass\_mfifty\] gives the median values as well as 68% and 90% confidence intervals based on the D12 and G14 constraints in each of the three [Illustris]{} resolution levels. Unlike the total enclosed mass at large radii, which is well-converged across the three different [Illustris]{} resolutions, the stellar masses in these haloes increase by a factor of $\sim 2$ from [Illustris]{}-3 to [Illustris]{}-1. This difference is not large enough to be reflected in stellar mass functions (which are reasonably similar for the different resolution levels studied here; see, e.g., @Vogelsberger2013 and @Torrey2014). It is larger than the uncertainty on the measured ${M_{\star}}$ of the MW, however: most recent estimates for the Galaxy fall in the range ${M_{\star}}= 5-6.5\times 10^{10}\,{M_{\odot}}$ (e.g., @McMillan2011 [@Bovy2013; @Licquia2015]).\ Differences in the simulated stellar masses at the factor of $\sim 2$ level are unsurprising, as the galaxy formation models used in the [Illustris]{} suite were calibrated at the resolution of [Illustris]{}-1; we would not expect the same models to work identically at significantly lower resolution. Specifically, the minimum resolution required for the feedback implementation in [Illustris]{} to produce a realistic galaxy population is not achieved in [Illustris]{}-3 [@Vogelsberger2013]. We therefore consider the results from [Illustris]{}-1 to be the most reasonable comparison to make with observations. We adopt the measurement of @Licquia2015 [hereafter, LN15], in which the authors used results derived in [@Bovy2013] to obtain ${M_{\star}}=6.08 \pm 1.14\times 10^{10}\,{M_{\odot}}$, as a representative value of the stellar mass of the MW and use it as a reference point in what follows. Comparing this number to the results for [Illustris]{}-1 in Table \[table:stellarMass\_mfifty\], we see that D12 agrees well with the observed value, while G14 is substantially lower. This is not surprising, given the results of Table \[table:confidenceIntervals\]. The very low value of $M_{\rm 200,c}$ obtained based on G14 is much lower than the typical value found for haloes with the stellar mass of the MW via either abundance matching [@guo2010; @behroozi2013c; @moster2013], galaxy-galaxy lensing [@Mandelbaum2016], or satellite kinematics (e.g., @watkins2010 [@Boylan-Kolchin2013]). Even accounting for possible differences in halo masses of red and blue galaxies at fixed stellar mass [@Mandelbaum2016], the MW would be a strong outlier if its mass is as low as the median value indicated by our analysis using the G14 constraint on ${M(<\!50\,{\rm kpc})}$. As noted in Section \[sec:methods\], our methodology for constraining the mass profile of the MW is quite general. While we have focused on constraining the total mass at large radii based on measurements of the total mass within 50 kpc, we can instead use other quantities – for example, ${M_{\star}}$ – for our inference. Following the same procedure outlined in Section \[sec:methods\], we estimate ${M(<\!50\,{\rm kpc})}$, ${M(<\!100\,{\rm kpc})}$, and the three spherical overdensity masses used above based on LN15’s determination of ${M_{\star}}$; the results are presented in the second column of Table \[table:mstar\]. The results are very similar to those obtained using the D12 determination of ${M(<\!50\,{\rm kpc})}$, with LN15-based estimates being 5-7% higher (the results are approximately a factor of 1.6–2 larger than G14-based estimates). ![The correlation between ${M_{\star}}$ and ${M(<\!50\,{\rm kpc})}$ for all haloes in the Illustris-1 sample; the haloes are coloured by the value of $M_{\rm 200,c}$. Vertical shaded bands show G14 (yellow) and D12 (magenta) determinations of ${M(<\!50\,{\rm kpc})}$ while the horizontal band shows the LN15 determination of ${M_{\star}}$ for the MW. Very few haloes agree with both the G14 measurement of the total mass at 50 kpc and the MW’s stellar mass; many more of the simulated galaxies match the D12 value for ${M(<\!50\,{\rm kpc})}$ and the LN14 ${M_{\star}}$ value simultaneously.[]{data-label="mstar-m50"}](figure_7.pdf){width="1.05\linewidth"} Since ${M_{\star}}$ and ${M(<\!50\,{\rm kpc})}$ can be considered independent variables, we can also study the joint probability of obtaining various mass measures conditioned on ${M_{\star}}$ and ${M(<\!50\,{\rm kpc})}$. These joint constraints, using D12’s estimate of ${M(<\!50\,{\rm kpc})}$, are given in the third column of Table \[table:mstar\]. The joint constraints are similar to both the estimates using ${M_{\star}}$ alone and the estimate using ${M(<\!50\,{\rm kpc})}$ (from D12) alone, which is a result of the good agreement of each of these estimates individually. Had we used the G14 value of ${M(<\!50\,{\rm kpc})}$, the constraints would have shifted substantially. This is highlighted in Fig. \[mstar-m50\], which shows the [Illustris]{}-1 data in ${M_{\star}}-{M(<\!50\,{\rm kpc})}$ space; each halo assigned a colour according to its value of $M_{\rm 200,c}$. The intersection of the D12 and LN15 constraints falls along the main locus of the points while the G14 constraint intersects the LN 15 constraint in a part of parameter space with very few haloes. The D12 and LN15 measurements are therefore in good agreement based on the [Illustris]{} haloes, while the G14 and LN15 measurements are not. \[table:mstar\] Discussion and future prospects {#sec:discussion} =============================== ![The precision attained in measurements of $M_{\rm 200,c}$ as a function of the precision in the input constraint. We consider input constraints of the total mass within 50, 80, and 100 kpc (black circles, red squares, and blue triangles, respectively) to show how more precise determinations of masses within larger radii can affect the inferred value of $M_{\rm 200,c}$. The figure shows the trade-off between precision and distance: for example, an error of 15% in ${M(<\!100\,{\rm kpc})}$ results in the same precision in the estimate of $M_{\rm 200,c}$ as an error of 9% in ${M(<\!50\,{\rm kpc})}$. []{data-label="fig:errorsVsRad"}](figure_8.pdf){width="\linewidth"} As larger samples of halo stars at greater distances become available, it may become possible to constrain the mass of the MW enclosed within 80 or even 100 kpc (see, e.g., @gnedin2010 [@Cohen2015] for initial work in this direction). Such measurements would have the benefit of providing stronger constraints on the virial mass of the MW. Fig. \[fig:errorsVsRad\] shows the fractional uncertainty in $M_{\rm 200,c}$ as a function of the error in the mass contained within 50 (black circles), 80 (red squares), and 100 kpc (blue triangles). At a fixed uncertainty in $M(<r)$, the implied uncertainty in $M_{\rm 200,c}$ does indeed become smaller as one moves to greater Galactocentric distance. The figure quantifies how improving uncertainties at a given distance will be reflected in uncertainties on $M_{\rm 200,c}$: for example, reducing the error on ${M(<\!50\,{\rm kpc})}$ from 10 to 5% would reduce the error on $M_{\rm 200,c}$ from 28 to 23%. On the other hand, an measurement of the mass within 80 kpc that is accurate to 10% results in an error of 22% in $M_{\rm 200,c}$, while the same accuracy on a measurement of the mass within 100 kpc of the Galaxy would yield errors of 19% in $M_{\rm 200,c}$. The figure also shows the fundamental limitations in extrapolating to $M_{\rm 200,c}$ based on measured aperture masses within smaller radii. Some level of irreducible uncertainty is unavoidable in standard cosmological models, as extrapolation from mass at a given radius to the virial radius depends on the halo concentration (e.g. @Navarro1997 [@Bullock2001]). For example, consider the recent study of @Williams2015, who found that ${M(<\!50\,{\rm kpc})}=4.48^{+0.15}_{-0.14}\times 10^{11}\,{M_{\odot}}$, or an error of approximately 3% on ${M(<\!50\,{\rm kpc})}$. Using this constraint, we obtain $M_{\rm 200,c}=1.25^{+0.35}_{-0.18}\times10^{12}\,{M_{\odot}}$; the uncertainty on the derived value of $M_{\rm 200,c}$ remains large in spite of the high precision of the input measurement. Fig. \[fig:errorsVsRad\] makes it clear that measurements of the mass within 50 (80, 100) kpc will result in an uncertainty on $M_{\rm 200,c}$ of no better than 23% (17%, 14%). A central assumption of the techniques we employ here is that [Illustris]{}-1 provides a faithful representation of galaxies and the effects of galaxy formation on dark matter halo structure. Since cosmological hydrodynamic simulations are still at the point of relying on subgrid models of physics, and will be for the foreseeable future, a logical extension of our work would be to investigate predictions in future generations of simulations to test the robustness of our results. It would also be interesting to compare the results we have obtained with [Illustris]{} to the Eagle simulations, as the galaxy formation modelling employed there is somewhat different. Given the differences seen in the ratio of masses in hydrodynamic to DMO simulations in [Illustris]{}versus Eagle (compare fig. 7 of @Vogelsberger2014b and fig. 1 of @Schaller2015), such a comparison would be timely. One effect that appears to be particularly important for setting the amount of mass reduction for a given halo in the hydrodynamic run relative to its counterpart in the DMO version is the underlying model of AGN feedback. @Vogelsberger2014a adopted an AGN model that drives *very* strong outflows, perhaps unrealistically so [@Genel2014]. Forthcoming updates to the [Illustris]{} suite will use modified versions of AGN feedback that are less powerful and may result in different modifications of the large-scale halo properties of galaxies, which may in turn affect how ${M(<\!50\,{\rm kpc})}$ maps on to $M_{\rm 200,c}$. To explore the potential impact of this effect on our results, we use the current generation of [Illustris]{} and compare the effects of BH mass for galaxies of a fixed halo mass (we use the haloes that are closest to the median value of $M_{\rm 200,c}$ found in [Illustris]{}-Dark-1 using the D12 constraint). We rank this sample according to black hole mass and then compute the difference in mass in the hydrodynamic simulation relative to the DMO run. There is indeed a difference: the galaxies with the highest-mass black holes show a 20% reduction in their overall mass, on average, while the galaxies with the lowest mass black holes see a 10% reduction in mass compared to their DMO counterparts. The total halo mass therefore appears to depend somewhat on the choice of black hole feedback model, although this does not appear to be a large source of uncertainty in our predictions. Future generations of [Illustris]{}-like simulations with modified black hole feedback models will allow us to directly test the effects on inferences regarding the MW mass. It is not entirely obvious how the effects of vigorous feedback propagate through our analysis, as this will depend on the change in enclosed mass within 50 kpc relative to the change in enclosed mass within larger radii. However, given that the black hole feedback in the current version of [Illustris]{} may be too effective and that the larger-mass black holes correlate with larger reductions in halo mass as compared to lower-mass black holes, it is likely that any modified prescriptions will result in slightly higher inferences on the total halo mass compared to our current results, should there be a difference. Future work would also benefit significantly from cosmological simulations with larger volumes and higher mass resolution. Importance sampling relies on having a well-sampled parameter space, which can be an issue if not many haloes match the desired constraint(s) (see @Busha2011 and @Gonzalez2014 for more details). Our current analysis has many haloes contributing significant weights: 870 and 2196 haloes contribute weights that are at least 10% of the maximum possible weight ($W_{\rm max}=1/\sqrt{2\rm{\pi}\sigma^2}$ from equation. \[eq:singleprob\]) for the D12 and G14 constraints, respectively. However, if we wish to add additional restrictions – based on morphology, disc size, star formation history, or specific star formation rate, for example – the sample would likely become significantly smaller, which would be the limiting factor in the conclusions we could draw. With larger sample sizes, such concerns would be eliminated. From Fig. \[mstar-m50\], joint constraints on ${M(<\!50\,{\rm kpc})}$ and ${M_{\star}}$ are unlikely to be strongly affected by sample size unless a much larger volume produced many haloes with much larger stellar masses at fixed halo mass \[in which case, the G14 measurement of ${M(<\!50\,{\rm kpc})}$ would be more consistent with the simulation results than it is at present\]. Conclusions {#sec:conclusions} =========== In this paper, we have explored how the [Illustris]{} suite can be used to inform our understanding of the mass distribution around the MW. Our main conclusions are as follows. - The mass profiles of haloes consistent with a given constraint on ${M(<\!50\,{\rm kpc})}$ differ substantially between DMO and hydrodynamic versions of [Illustris]{}. Using DMO simulations to extrapolate from 50 kpc to larger radii results in an overestimate of the halo mass and an underestimate of the halo concentration. - The effects of baryonic physics on the mass distribution of MW-like systems in [Illustris]{} are substantial: by matching haloes between the DMO and hydrodynamic simulations, we find that the latter have more mass on small scales and less mass on large scales. The asymptotic difference in the total mass density at large radii is approximately 20%. - Since different feedback models result in very different effects on the mass distribution of dark matter even at large distances from halo centres (e.g., fig. 7 of @Vogelsberger2014b compared to fig. 1 of @Schaller2015), it is imperative to test how inferences on the mass of the MW depend on galaxy formation modelling. - The mass distribution in the inner $\sim 20$ kpc is not converged in the [Illustris]{} suite \[see @Schaller2016 for similar results in the Eagle simulations\]; this is a much larger distance than the formal convergence radius for the dark matter simulations. Results regarding the density distribution for $r \la 20$ kpc must therefore be interpreted with caution, and our best-fitting NFW profiles for the hydrodynamic simulations, which were obtained over the radial range of 40-300 kpc, should not be extrapolated to smaller radii. - The relationship between ${M(<\!50\,{\rm kpc})}$ and $M_{\rm 200,c}$ in [Illustris]{}-1 is well-described by a log-quadratic relationship (equation. \[eq:mfifty\_fit\]). This relationship enables the translation of any existing or future constraint on ${M(<\!50\,{\rm kpc})}$ into a measurement $M_{\rm 200, c}$. - The constraints on ${M(<\!50\,{\rm kpc})}$ derived by D12 ($4.2 \pm 0.4\times 10^{11}\,{M_{\odot}}$) and G14 ($2.9 \pm 0.4 \times 10^{11}\,{M_{\odot}}$) predict very different values for the virial mass of the Galaxy’s halo when using [Illustris]{}: for D12, we find $M_{\rm 200,c}=1.12^{+0.37}_{-0.24}\times 10^{12}\,{M_{\odot}}$ (68% confidence), while for G14, we find $M_{\rm 200,c}=0.612^{+0.196}_{-0.148}\times10^{12}\,{M_{\odot}}$ (68% confidence). The values for ${M_{\rm{vir}}}$ and $M_{\rm 200,m}$ are 17% and 32% larger, respectively. - [Illustris]{} haloes that have galaxies with stellar masses consistent with measurements of the MW’s ${M_{\star}}$ have significantly more mass within $50\,{{\rm kpc}}$ than the result of G14; the measurements of D12 and @Williams2015 are in much better agreement with [Illustris]{} haloes that match the observed value of ${M_{\star}}$. In particular, almost no haloes in [Illustris]{} jointly satisfy the G14 constraint and the LN15 measurement of ${M_{\star}}$ for the MW. - From our analysis of the [Illustris]{} simulation, even an infinitely precise measurement of ${M(<\!50\,{\rm kpc})}$ would result in an uncertainty of $>$20% in $M_{\rm 200,c}$. The same uncertainty can be achieved for 10% errors on $M(<80\,{{\rm kpc}})$ or 12% errors on ${M(<\!100\,{\rm kpc})}$. A measurement of ${M(<\!100\,{\rm kpc})}$ that is accurate to 5% will translate into 15% uncertainties on $M_{\rm 200,c}$. As ever larger and ever more realistic hydrodynamic simulations become available, so too will better statistical constraints on the mass profile of our Galaxy. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Joss Bland-Hawthorn, Nitya Kallivayalil, Julio Navarro, and Annalisa Pillepich for helpful conversations. The analysis of the [Illustris]{} data sets for this paper was done using the Odyssey cluster, which is supported by the FAS Division of Science, Research Computing Group at Harvard University. MB-K acknowledges support provided by NASA through a *Hubble Space Telescope* theory grant (programme AR-12836) from the Space Telescope Science Institute (STScI), which is operated by the Association of Universities for Research in Astronomy (AURA), Inc., under NASA contract NAS5-26555. PT acknowledges support from NASA ATP Grant NNX14AH35G. LH acknowledges support from NASA grant NNX12AC67G and NSF grant AST-1312095. \[lastpage\] [^1]: cjtaylor@astro.umd.edu [^2]: mbk@astro.as.utexas.edu [^3]: The [Illustris]{} data are all publicly available (http://www.illustris-project.org/); see @Nelson2015 for further information. [^4]: Here and throughout this work, we use ’mass’ to refer to the enclosed mass (as opposed to mass within a spherical shell)

ArXiv

--- address: | Massachusetts Institute of Technology\ Cambridge, Massachusetts 20139, USA author: - Frank Ma for the CMS Collaboration title: Studies of Jet Quenching in HI Collisions at CMS --- {width="99.00000%"} \[fig:Cover\] Introduction ============ Heavy ion collisions at the Large Hadron Collider (LHC) allow one to study the thermodynamic properties of the fundamental theory of the strong interaction — Quantum Chromodynamics (QCD). Studying the modification of jets that are created from within the medium has long been proposed as a particularly useful tool for probing the QCD medium properties [@Appel:1985dq; @Blaizot:1986ma]. In the presence of a QCD medium, the partons may lose energy to the medium via elastic processes (collisional parton energy loss) or inelastic processes (radiative parton energy loss). The study of medium-induced modifications of dijet properties can therefore shed light on the transport properties of collective QCD matter created by heavy ion collisions. Experimental Methods ==================== This analysis was performed using the data collected in 2010 from PbPb collisions at a nucleon-nucleon center-of-mass energy of $\sqrt{s_{_{NN}}}=2.76$ TeV at the Compact Muon Solenoid (CMS) detector [@bib_CMS]. Jets were reconstructed with background subtraction based on their energy deposits in the CMS calorimeters [@Kodolova:2007hd], and the events were selected from a jet-triggered dataset [@CmsJetQuenching:2011]. Because heavy ions are extended objects, the impact parameter is an important characterization of the events. The amount of overlap between the two colliding nuclei is what we mean by “centrality” of the collision. In this analysis, centrality was determined from minimum events based on the total energy from both forward hadronic calorimeters [@CmsJetQuenching:2011]. Simulations can be used to correlate centrality, as quantified using the fraction of the total interaction cross section, with physically meaningful quantities such as the total number of nucleons in the two lead ($^{208}$Pb) nuclei which experienced at least one inelastic collision ([$N_{\rm part}$]{}). Results ======= Dijet Properties in pp and PbPb data {#sec:dijet_results} ------------------------------------ To obtain a clean dijet selection, we select events with a leading jet having corrected $p_{\mathrm{T},1} > 120$ [$\text{GeV/}c$]{}, a subleading jet with $p_{\mathrm{T},2} > 50$ [$\text{GeV/}c$]{}, and a minimum azimuthal angle between them ($\Delta \phi_{12} > 2\pi/3$). Only jets within $|\eta| < 2$ were considered. Given this selection, we observe a sharp $\Delta \phi_{12}$ correlation between leading and subleading jets [@CmsJetQuenching:2011] , indicating true dijet pairs. In-medium induced parton energy loss can significantly alter the detector level jet energy (and hence dijet energy balance) by either transporting energy outside of the jet cone or shifting the energy towards low momentum particles that will not be detected in the calorimeter. To characterize the dijet momentum balance quantitatively, we use the asymmetry ratio, $$\label{eq:aj} A_J = \frac{p_{\mathrm{T},1}-p_{\mathrm{T},2}}{p_{\mathrm{T},1}+p_{\mathrm{T},2}}~,$$ where ${\ensuremath{p_T}}$ is the corrected ${\ensuremath{p_T}}$ of the reconstructed calorimeter jet. The subscript $1$ always refers to the leading jet, so that $A_J$ is positive by construction. In Fig. \[fig:JetAsymm\] (a), the $A_J$ dijet asymmetry observable calculated by is compared to   data at $\sqrt{s}$ = 7 TeV. We see that data and event generator are found to be in excellent agreement, demonstrating that (at $\sqrt{s}$ = 2.76 TeV) can serve as a good reference for the dijet imbalance analysis in  collisions. Figs. \[fig:JetAsymm\] (b)-(f) show the centrality dependence of $A_J$ for  collisions. To separate effects due to the medium itself from effects simply due to reconstructing jets in the complicated environment of the underlying PbPb event, the reference dijet events were embedded into a minimum bias selection of PbPb events at the raw data level [@CmsJetQuenching:2011]. In contrast to , we see that data shows a dramatic decrease of balanced dijets with increasing centrality. ![Left 6 panels show dijet asymmetry distribution, $A_{J}$, of selected dijets for 7 TeV pp collisions (a) and 2.76 TeV  collisions in several centrality bins: (b) 50–100%, (c) 30–50%, (d) 20–30%, (e) 10–20% and (f) 0–10%. Data are shown as black points, while the histograms show (a) events and (b)-(f) events embedded into  data. Right panel shows fraction of selected dijets with $A_J<0.15$ out of all events with a leading jet with $p_{\mathrm{T},1} > 120$ [$\text{GeV/}c$]{} as a function of [$N_{\rm part}$]{}. The result for reconstructed dijet events (blue filled star) is plotted at [$N_{\rm part}$]{} = 2. The other points (from left to right) correspond to centrality bins shown in (b)-(f) in the left 6 panels. The red squares are for reconstruction of events and the filled circles are for the  data. For the data points, vertical bars and brackets represent the statistical and systematic uncertainties, respectively.[]{data-label="fig:JetAsymm"}](dijet_imbalance_all_cent_20101126_v0.pdf "fig:"){width="60.00000%"} ![Left 6 panels show dijet asymmetry distribution, $A_{J}$, of selected dijets for 7 TeV pp collisions (a) and 2.76 TeV  collisions in several centrality bins: (b) 50–100%, (c) 30–50%, (d) 20–30%, (e) 10–20% and (f) 0–10%. Data are shown as black points, while the histograms show (a) events and (b)-(f) events embedded into  data. Right panel shows fraction of selected dijets with $A_J<0.15$ out of all events with a leading jet with $p_{\mathrm{T},1} > 120$ [$\text{GeV/}c$]{} as a function of [$N_{\rm part}$]{}. The result for reconstructed dijet events (blue filled star) is plotted at [$N_{\rm part}$]{} = 2. The other points (from left to right) correspond to centrality bins shown in (b)-(f) in the left 6 panels. The red squares are for reconstruction of events and the filled circles are for the  data. For the data points, vertical bars and brackets represent the statistical and systematic uncertainties, respectively.[]{data-label="fig:JetAsymm"}](RB_015_vs_Npart20101216v1 "fig:"){width="39.00000%"} The centrality evolution of the dijet momentum balance can be explored more quantitatively by studying the fraction of balanced jets in the  events. The balanced fraction, $R_B({\ensuremath{A_J}}< 0.15)$, is plotted as a function of collision centrality (in terms of [$N_{\rm part}$]{}) in the right panel of Fig. \[fig:JetAsymm\]. It is defined as the fraction of all events with a leading jet having $p_{\mathrm{T},1} > 120$ [$\text{GeV/}c$]{} for which a subleading partner with ${\ensuremath{A_J}}< 0.15$ and ${\ensuremath{\Delta\phi}}_{12} > 2\pi/3$ is found. The [$A_J$]{} threshold of 0.15 was chosen because it is the median of the [$A_J$]{} distribution for selected dijets in pure events. In contrast to dijets, the  data show a rapid decrease in the fraction of balanced jets with collision centrality. The effect is much larger than the combined systematic uncertainties. These results imply a degradation of the parton energy, or jet quenching, in the medium produced in central  collisions. The final systematic uncertainties, stemming mainly from uncertainties in the jet energy scale, are described in  [@CmsJetQuenching:2011]. Overall Momentum balance of Dijet Events {#sec:missingpt} ---------------------------------------- We next turn to the question of where and how the medium energy loss occurs by exploiting additional information from the entire CMS tracker. We measure overall transverse momentum balance in the dijet events using the projection of missing ${\ensuremath{p_T}}$ of reconstructed charged tracks onto the leading jet axis, defined as, $$\displaystyle{\not} p_{\mathrm{T}}^{\parallel} = -\sum_{\rm i}{p_{\mathrm{T}}^{\rm i}\cos{(\phi_{\rm i}-\phi_{\rm Leading\ Jet})}}, $$ where the sum is over all tracks with ${\ensuremath{p_T}}> 0.5$ ${\ensuremath{\text{GeV/}c}\xspace}$ and $|\eta| < 2.4$. For this study, the leading and subleading jets are required to have a slightly smaller $\eta$ range ($|\eta| < 1.6$) to allow the jets to remain fully inside the CMS tracker acceptance. No background subtraction in the track distribution is needed since the underlying  tracks cancel in the $\displaystyle{\not} p_{\mathrm{T}}^{\parallel}$ sum. In Fig. \[fig:MissingpT\], $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ is shown as a function of [$A_J$]{}  in the 0–30% centrality bin, where we expect the medium effects to be the strongest. Here [$A_J$]{} is the same calorimeter jet [$A_J$]{} as described in Sec. \[sec:dijet\_results\]. The left column shows $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ using all selected tracks. One sees that in both data and simulation, the overall momentum balance of the events (shown as solid circles) is recovered within uncertainties even for dijets with large energy asymmetry. This cross-checks the soundness of the detector, since regardless of medium effects, net transverse momentum is conserved. The figure also shows the contributions to $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ for five transverse momentum ranges from 0.5–1 [$\text{GeV/}c$]{} to ${\ensuremath{p_T}}> 8$ [$\text{GeV/}c$]{}, shown as stacked histograms. ![Average missing transverse momentum, $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$, for tracks with ${\ensuremath{p_T}}> 0.5$ [$\text{GeV/}c$]{}, projected onto the leading jet axis (solid circles). The $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ values are shown as a function of dijet asymmetry $A_J$ in 0–30% central events, for the full event (left), inside ($\Delta R < 0.8$) one of the leading or subleading jet cones (middle) and outside ($\Delta R > 0.8$) the leading and subleading jet cones (right). For the solid circles, vertical bars and brackets represent the statistical and systematic uncertainties, respectively. Colored bands show the contribution to $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ for five ranges of track [$p_T$]{}. The top and bottom rows show results for and  data, respectively. For the individual ${\ensuremath{p_T}}$ ranges, the statistical uncertainties are shown as vertical bars. Note that as the underlying  event in both data and MC is not $\phi$-symmetric on an event-by-event basis, the back-to-back requirement was tightened to ${\ensuremath{\Delta\phi}}_{12} > 5 \pi/6$. []{data-label="fig:MissingpT"}](missingPtParallel-Corrected-data-InConeOutConeDPhiCut_layer10){width="70.00000%"} Important insights into the dijet asymmetry emerge when we look at the $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ differential in radial distance from the jets. The middle and right columns of Fig. \[fig:MissingpT\] show $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ separately for tracks inside cones of size $\Delta R = 0.8$ around the leading and subleading jet axes, and for tracks outside of these cones. We see that for both data and MC an in-cone imbalance of $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle \approx -20$ [$\text{GeV/}c$]{} is found for the ${\ensuremath{A_J}}> 0.33$ selection. This shows that track momentum sums within the leading and subleading jet cones confirm the calorimeter dijet asymmetry results showed earlier in Sec. \[sec:dijet\_results\]. In addition, both data and simulation show similar large negative contribution to $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle$ (i.e., in the direction of the leading jet) in the ${\ensuremath{p_T}}> 8$ [$\text{GeV/}c$]{} range. This cross-checks that the dijet energy asymmetry in data is not caused by fake jets from background fluctuation, because only genuine high ${\ensuremath{p_T}}$ jets give rise to high ${\ensuremath{p_T}}$ tracks. Looking now at the right column, we see that in both data and MC the in-cone energy difference is balanced by a corresponding out-of-cone imbalance of $\langle \displaystyle{\not} p_{\mathrm{T}}^{\parallel} \rangle \approx 20$ [$\text{GeV/}c$]{}. However, in the  data the out-of-cone contribution is carried almost entirely by tracks with $0.5 < {\ensuremath{p_T}}< 4$ [$\text{GeV/}c$]{} whereas in MC more than 50% of the balance is carried by tracks with ${\ensuremath{p_T}}> 4$ [$\text{GeV/}c$]{}, with a negligible contribution from ${\ensuremath{p_T}}< 1$ [$\text{GeV/}c$]{}. The results are indicative of semi-hard initial or final-state radiation as the underlying cause for large [$A_J$]{} events in the MC study. This is in contrast to the results for large-[$A_J$]{}  data, which show that a large part of the momentum balance is carried by soft particles (${\ensuremath{p_T}}< 2$ [$\text{GeV/}c$]{}) and radiated at large angles to the jet axes ($\Delta R > 0.8$). Summary and Conclusion ====================== A strong increase in the fraction of highly unbalanced jets has been seen in central PbPb collisions compared with peripheral collisions and model calculations, consistent with a high degree of parton energy loss in the produced QCD medium. A large fraction of the momentum balance of these unbalanced jets is carried by low-[$p_T$]{}particles at large radial distance, in contrast to simulations embedded into heavy ion events. The results provide qualitative constraints on the nature of the jet modification in  collisions and quantitative input to models of the transport properties of the medium created in these collisions. References {#references .unnumbered} ========== [99]{} D. A. Appel, “[Jets as a probe of quark-gluon plasmas]{}”, *Phys. Rev.* **D33** (1986) 717. [` doi:10.1103/PhysRevD.33.717`](http://dx.doi.org/10.1103/PhysRevD.33.717). J. P. Blaizot and L. D. McLerran, “[Jets in Expanding Quark - Gluon Plasmas]{}”, *Phys. Rev.* **D34** (1986) 2739. [` doi:10.1103/PhysRevD.34.2739`](http://dx.doi.org/10.1103/PhysRevD.34.2739). Collaboration, “The CMS experiment at the CERN LHC”, *JINST* **3** (2008) S08004. [` doi:10.1088/1748-0221/3/08/S08004`](http://dx.doi.org/10.1088/1748-0221/3/08/S08004). O. Kodolova, I. Vardanian, A. Nikitenko[ et al.]{}, “[The performance of the jet identification and reconstruction in heavy ions collisions with CMS detector]{}”, *Eur. Phys. J.* **C50** (2007) 117. [` doi:10.1140/epjc/s10052-007-0223-9`](http://dx.doi.org/10.1140/epjc/s10052-007-0223-9). CMS Collaboration, “[Observation and studies of jet quenching in PbPb collisions at nucleon-nucleon center-of-mass energy = 2.76 TeV]{}”, [` arXiv:1102.1957`](http://arxiv.org/abs/1102.1957).

ArXiv

--- abstract: 'Results from a realistic simulation of the heavy MSSM Higgs-bosons $A$ and $H$ production, $ \gamma \gamma \rightarrow A,H \rightarrow b\overline{b} $, at the Photon Collider at TESLA are reported. In the scenario where a light SM-like Higgs boson $h$ exists, we study (following M. Mühlleitner et al. [@MMuhlleitner]) Higgs bosons $A$ and $H$ for masses $ M_{A}=200,250,300,350 $ GeV and $\tan\beta=7$. This scenario corresponds to parameters region not accessible at LHC or at the first stage of the $e^+ e^-$ collider. NLO estimation of background, analysis of overlaying events, realistic $b$-tagging and corrections for escaping neutrinos were performed. The statistical precision of the cross-section measurement is estimated to be 8-20%.' author: - | \ Piotr Nieżurawski, Aleksander Filip Żarnecki\ [*Institute of Experimental Physics, Warsaw University, ul. Hoża 69, 00-681 Warsaw, Poland*]{}\ Maria Krawczyk\ [*Institute of Theoretical Physics, Warsaw University, ul. Hoża 69, 00-681 Warsaw, Poland*]{}\ title: | ----------------------- [IFT - 20/2003]{} \[-3mm\] [hep-ph/ ]{} \[3mm\] [July 2003]{} ----------------------- Measurement of the MSSM Higgs-bosons production\ in $ \gamma \gamma \rightarrow A,H \rightarrow b\overline{b} $\ at the Photon Collider at TESLA --- Introduction ============ A photon collider option of the TESLA collider [@TDR] offers a unique possibility to produce the Higgs boson as an $ s $-channel resonance. The neutral Higgs boson couples to the photons through a loop with massive charged particles. This loop-induced higgs-$\gaga$ coupling is sensitive to contributions of new particles, which appear in various extensions of the Standard Model (SM). Besides precision measurements, a photon collider is a candidate for the discovery machine [@Asner]. In case of Minimal Supersymmetric extension of the SM (MSSM) the photon collider will be able to cover so called “LHC wedge” around intermediate values of $ \tan \beta$, $ \tan \beta \sim 7 $, and for heavy neutral Higgs-bosons masses above 200 GeV, which will be inaccessible at LHC [@ATLAS] and at the first stage of $\epem$ linear colliders. In our analysis, we consider a SM-like scenario where the lightest Higgs boson $h$ has properties of the SM-Higgs boson, while heavy neutral Higgs bosons are degenerated in mass and have negligible couplings to the gauge bosons $W/Z$. We take MSSM parameters as in [@MMuhlleitner], , and consider the process $ \gagaAHbb $ at the Photon Collider at TESLA for Higgs-boson masses 200, 250, 300 and 350 GeV. The MSSM Higgs bosons $\AOHO$ with masses in this range are expected to decay into $ \bbbar $ state with branching ratios from around 80% to 20%. The preliminary results for the MSSM Higgs-bosons were presented in [@NZKamsterdam]. In our previous works [@NZKhbbm120appb; @NZKSMeps2003] we have considered a light SM Higgs-boson production at the Photon Collider at TESLA. The background treatment introduced in that analysis we apply now for the MSSM Higgs-bosons. Both, the signal and background events are generated according to a realistic photon–photon luminosity spectrum [@V.Telnov], parametrized by a CompAZ model [@CompAZ]. This study takes into account overlaying events [$ \gaga \rightarrow $ [*hadrons*]{}]{} for which full simulated photon–photon luminosity spectra [@V.Telnov] are used. We simulate the detector response according to the program SIMDET [@SIMDET401] and perform a realistic [$ b $-tagging]{} [@Btagging]. Photon–photon luminosity spectra ================================ The Compton back-scattering of a laser light off the high-energy electron beams is considered as a source of the highly energetic, highly polarized photon beams [@Ilya]. In photon–photon beams simulations for a photon collider option at TESLA, according to the current design [@TDR], the energy of the laser photons is assumed to be fixed for all considered electron-beam energies; the laser photons are assumed to have circular polarization $P_{c} = $ 100%, while the electrons longitudinal polarization is $P_{e} = $ 85%. In considered case the luminosity spectrum is peaked at high energy and we assume that the energy of primary electrons is adjusted in order to enhance the signal at a particular mass. In a generation of the processes [$ \gaga \rightarrow $ [*hadrons*]{}]{}, one has to take into account also the low energy events, since they contribute to overlaying events. To simulate them the realistic $\gaga$-luminosity spectra for the photon collider at TESLA [@V.Telnov] are used, with the non-linear corrections and higher order QED processes. For generation of the processes $\gagaAHbb$ and $\gagabbgccg$ we use the CompAZ parametrization [@CompAZ] of the $\gaga$-luminosity spectrum [@V.Telnov]. The results presented in this paper were obtained for an integrated luminosity expected for one year of the photon collider running for TESLA collider [@V.Telnov]. For example with 419 GeV – the optimal choice for 300 GeV – the total photon–photon luminosity per year is $L_{\gamma\gamma}=808$ fb$^{-1}$. The total photon-photon luminosity increases with energy from about 570 fb$^{-1}$ for 305 GeV (200 GeV) to about 937 fb$^{-1}$ for 473 GeV (350 GeV). Details of a simulation\ and the first results for 300 GeV ================================= To calculate the total widths and branching ratios of the MSSM Higgs bosons $\HO$ and $\AO$ we use the program HDECAY [@HDECAY] with MSSM parameters as in [@MMuhlleitner], , and including decays and loops of supersymmetric particles. In a generation of the signal events both $\AO$ and $\HO$ are included, due to their degeneracy in mass. This was done with the PYTHIA program [@PYTHIA]. A parton shower algorithm, implemented in PYTHIA, was used to generate the final-state particles. The background events due to processes $\gagabbgccg$ were generated using the program written by G. Jikia [@JikiaAndSoldner], where a complete NLO QCD calculation for the production of massive quarks is performed within the massive-quark scheme. The program includes exact one-loop QCD corrections to the lowest order (LO) process $\gagabbcc$ [@JikiaAndTkabladze], and in addition the non-Sudakov form factor in the double-logarithmic approximation, calculated up to four loops [@MellesStirlingKhoze]. The fragmentation into hadrons for all processes was performed using the PYTHIA program. Because of a large cross section, about two [$ \gaga \rightarrow $ [*hadrons*]{}]{} events[^1] are expected on average per bunch crossing at the TESLA Photon-Collider (for $ \sqrt{s_{ee}} \approx $ 400 GeV, at nominal luminosity). We generate these events according to PYTHIA and convolute with results of a full simulation of the photon–photon luminosity spectra [@V.Telnov], rescaled for the chosen beam energy. For each considered $\emem$ energy, $\sqrtsee$, an average number of the [$ \gaga \rightarrow $ [*hadrons*]{}]{} events per a bunch crossing is calculated. Then, for every signal $\gagaAHbb$ or background $\gagabbgccg$ event, the [$ \gaga \rightarrow $ [*hadrons*]{}]{} events are overlaid (added to the event record) according to the Poisson distribution. Program SIMDET version 4.01 [@SIMDET401] was used to simulate a detector performance. Because [$ \gaga \rightarrow $ [*hadrons*]{}]{} process has a forward-peaked distribution (see for example [@NZKamsterdam]) we decrease the influence of overlaying events by ignoring tracks and clusters with $|\cos(\theta_{i})|>\cos(\theta_{min})=0.9$ ($\theta_{min}=450$ mrad; the polar angle is measured in the laboratory frame). This cut is used only when overlaying events are included in the analysis. The jets were reconstructed using the Durham algorithm, with $ y_{cut} = 0.02 $; the distance measure was defined as $ y_{ij}=2\min (E^{2}_{i},E^{2}_{j})(1-\cos \theta _{ij})/E^{2}_{vis} $, where $E_{vis}$ is defined as the total energy measured in the detector. The following selection cuts were used to select the signal events, $ \gagaAHbb $: - since the Higgs bosons are expected to be produced almost at rest, we require that the ratio of the total longitudinal momentum of all observed particles to the total visible energy is $ |P_{z}|/E_{vis}<0.15 $, - we select two- and three-jet events, $ N_{jets}=2,\, 3 $, so that events with one additional jet due to a hard-gluon emission are also accepted, - for each jet we require $ |\cos \theta _{i}|<0.75 $, $i=1, ..., N_{jets}$. We use “ZVTOP-B-Hadron-Tagger” package for the TESLA collider [@Btagging] for realistic [$ b $-tagging]{} simulation. The package is based on the neural network algorithm trained on the $Z$ decays. For each jet it returns a “$b$-tag” value – the number between 0 and 1 corresponding to “$b$-jet” likelihood. In order to optimize the signal cross-section measurement, we choose the two-dimensional cut on $b$-tag values for 2-jet events. For 3-jet events three possible pairs of jets were considered and the event was accepted if at least one pair passed the $b\bar{b}$ cut. It was found that the cut optimal for $\sgagaAHbb$ measurement, including effects of overlaying events, corresponds to the [$ \bbbar $-tagging]{} efficiency $\varepsilon_{bb}=79\%$ and [$ \ccbar $-mistagging]{} probability $\varepsilon_{cc}=3.6\%$. However, if overlaying events are not included then the best choice corresponds to the efficiences $\varepsilon_{bb}=72\%$ and $\varepsilon_{cc}=1.7\%$. [![\[fig:HHOEBtag\] Reconstructed invariant-mass, $ W_{rec}\protect $, distributions for selected $\gagaAbb$ events for 300 GeV, after [$ b $-tagging]{}, without and with overlaying events (OE). ](Plots/plot_ptet_wwoe_m300_modmssm_a0_var1.eps "fig:")]{} The invariant-mass distributions for the $\gagaAbb$ events after [$ b $-tagging]{}, before and after taking into account overlaying events (OE), are compared in Fig. \[fig:HHOEBtag\] for 300 GeV. The mass resolutions from the Gaussian fits, in the region $ \mu - \sigma $ to $ \mu + 2 \sigma $, to both distributions are 8 and 20 GeV, respectively. The same resolutions are obtained for the selected $\gagaHbb$ events (not shown). As the expected mass difference between $\HO$ and $\AO$ is of the order of 1 GeV, it will not be possible to separate signals of scalar and pseudoscalar Higgs boson using the the invariant-mass distribution. In the following we will consider a simultaneous measurement of the total cross section for both processes $\gagaAbb$ and $\gagaHbb$. Despite our choice of a $\bbbar$ selection cut which corresponds to the greater [$ \bbbar $-tagging]{} efficiency than for the simulation without overlaying events, we observe a 16% drop of number of events (from about 440 to 370 events) for simulation which includes overlaying events. This is because energy deposits from the [$ \gaga \rightarrow $ [*hadrons*]{}]{} processes, remaining after the $\theta_{min}$ cut, “shift” jets nearer to the beam axis and the event can be rejected by the jet-angle cut. Moreover, the additional deposits and $\theta_{min}$-cut deform jets slightly what reduces a selection efficiency. After applying selection cuts described above and including [$ b $-tagging]{}, we obtain the distributions of the reconstructed $ \gaga $ invariant mass, $ W_{rec} $, shown in Figs. \[fig:ResultWithNLOBackgd\]. The $\Hbb$ and $\Abb$ signal, and the NLO background contributions, $\bbbar(g)$ and $\ccbar(g)$, are shown separately. Result obtained before (upper plot) and after (lower plot) taking into account overlaying events are compared. We observe that the overlaying events significantly smear out the Higgs-boson signal. Assuming that the signal for the Higgs-bosons production will be extracted by counting the number of $ \bbbar $ events in the mass window around the peak, $N_{obs}$, and subtracting the expected background events, $N_{bkgd}$, we can calculate the expected relative statistical error for the cross section $ \sgagaAHbb $ in the following way: $$\frac{\Delta \sgagaAHbb}{\sgagaAHbb} = \frac{\sqrt{N_{obs}}}{N_{obs}-N_{bkgd}}.$$ For Higgs-boson mass of 300 GeV, after taking into account overlaying events, the accuracy of 9.3% is expected for the reconstructed invariant-mass window between 275 and 405 GeV (see Fig. \[fig:ResultWithNLOBackgd\]). [![\[fig:ResultWithNLOBackgd\] Distribution of the reconstructed invariant mass, $ W_{rec}\protect $, for the selected $\bbbar$ events (300 GeV) before (upper plot) and after (lower plot) taking into account overlaying events (OE). Contributions of the $H$ and $A$ signal and of the heavy-quark background are shown separately. Arrows indicate the mass windows optimized for the measurement of $\sgagaAHbb$. ](Plots/plot_m300_modmssm_var1_oe0.eps "fig:")]{} [![\[fig:ResultWithNLOBackgd\] Distribution of the reconstructed invariant mass, $ W_{rec}\protect $, for the selected $\bbbar$ events (300 GeV) before (upper plot) and after (lower plot) taking into account overlaying events (OE). Contributions of the $H$ and $A$ signal and of the heavy-quark background are shown separately. Arrows indicate the mass windows optimized for the measurement of $\sgagaAHbb$. ](Plots/plot_m300_modmssm_var1_oe1.eps "fig:")]{} Final results for masses 200–350 GeV ==================================== As in [@NZKhbbm120appb] to correct for escaping neutrinos we use the corrected invariant-mass variable, defined as: $$\begin{aligned} W_{corr} & \equiv & \sqrt{W^{2}_{rec}+2P_{T}(E_{vis}+P_{T})} \; .\end{aligned}$$ In Fig. \[fig:HHOEBtagWcorr\] the distributions of $ W_{corr} $ for the selected $\gagaAbb$ events, obtained without and with overlaying events, are presented. The tail of events with the invariant masses below $\sim 280$ GeV is much smaller than for $W_{rec}$ (compare with Fig. \[fig:HHOEBtag\]). The mass resolutions, derived from the Gaussian fits to the $W_{corr}$ distributions in the region from $ \mu - 2 \sigma $ to $ \mu + \sigma $, without and with overlaying events, are equal to 8 and 13 GeV, respectively. For $\gagaHbb$ events we obtain a very similar distribution, as seen in Fig. \[fig:AHBtagWcorr\]. The $ W_{corr} $ distributions, obtained for the signal and background events, are shown in Figs. \[fig:Wcorr\]. Results obtained before (upper plot) and after (lower plot) taking into account overlaying events (OE) are compared. If overlaying events are included, the most precise measurement of the Higgs-boson cross section is obtained for the $ W_{corr} $ mass window between 290 and 415 GeV, as indicated by arrows. In the selected $ W_{corr} $ region one expects, after one year of the Photon Collider running at the nominal luminosity, about 610 reconstructed signal events and 2000 background events ($ S/B \approx 0.3 $). This corresponds to the measurement with the expected relative statistical precision of 8.3%. We have performed also a full simulation of the signal and background events for 200, 250 and 350 GeV, choosing for each mass an optimal $\emem$ beam energy. In Fig. \[fig:PrecisionSummary\] a statistical precision of $\sgagaAHbb$ measurement for all considered masses are presented. For comparison also our earlier estimates [@NZKamsterdam], obtained without overlaying events, are shown. Conclusions =========== In this paper we present results of a full simulation of a signal due to the heavy MSSM Higgs bosons $\AO$ and $\HO$ decaying into $\bbbar$ and background events for the Photon Collider at TESLA. We study masses 200, 250, 300 and 350 GeV, and for each mass we choose an optimal $\emem$ beam energy. Following [@MMuhlleitner], we study parameters of the MSSM, which correspond to the “LHC wedge”, for which $\AO$ and $\HO$ are almost degenerate in mass. We performed a realistic simulation with the NLO background, corrections for escaping neutrinos, with realistic [$ b $-tagging]{} and taking into account overlaying events, as in our SM-Higgs analysis [@NZKSMeps2003]. Our analysis shows that, for the MSSM Higgs-bosons at $\MAO \sim$ 300 GeV, the cross section $\sgagaAHbb$ can be measured with a statistical precision around 8%, for other masses it is lower – from 10% till 20%. Although this result is less optimistic than the earlier estimate [@MMuhlleitner], it is still true that a photon–photon collider gives opportunity of a precision measurement of $\sgagaAHbb$, assuming that we know the mass of a Higgs boson(s). The issue arises how to distinguish (separate) $\AO$ from $\HO$. To get a reliable answer to this question an additional study is needed. We confirm results of [@JikiaAndSoldner] which indicated that the reconstructed mass resolutions (without including overlaying events) are greater than $\sim 2$ GeV. A discovery of MSSM Higgs-bosons requires energy scanning or a run with a broad luminosity spectrum, perhaps followed by the run with a peaked one [@Asner]. We conclude, that there is still a room for an optimization of $\theta_{min}$ cut to minimize an influence of overlaying events, which may increase a statistical precision. [![\[fig:HHOEBtagWcorr\] Corrected invariant mass, $ W_{corr}\protect $, distributions for the selected $\gagaAbb$ events for 300 GeV, after [$ b $-tagging]{}, obtained without and with overlaying events (OE). ](Plots/plot_ptet_wwoe_m300_modmssm_a0_var4.eps "fig:")]{} [![\[fig:AHBtagWcorr\] Corrected invariant mass, $ W_{corr}\protect $, distributions for the selected $\gagaHbb$ and $\gagaAbb$ events for 300 GeV, after [$ b $-tagging]{}, with overlaying events. ](Plots/plot_ptet_ahh_m300_modmssm_hh0_var4.eps "fig:")]{} [![\[fig:Wcorr\] As in Figs. \[fig:ResultWithNLOBackgd\], for the corrected invariant-mass, $ W_{corr}\protect $, distributions. The statistical precision is 6.3% without overlaying events (upper plot) and 8.3% with overlaying events (lower plot). ](Plots/plot_m300_modmssm_var4_oe0.eps "fig:")]{} [![\[fig:Wcorr\] As in Figs. \[fig:ResultWithNLOBackgd\], for the corrected invariant-mass, $ W_{corr}\protect $, distributions. The statistical precision is 6.3% without overlaying events (upper plot) and 8.3% with overlaying events (lower plot). ](Plots/plot_m300_modmssm_var4_oe1.eps "fig:")]{} [![\[fig:PrecisionSummary\] Precisions of $\sgagaAHbb$ measurement are shown for 200-350 GeV and ; with and without overlaying events (OE), as indicated in the plot. The points are connected with lines to guide the eye. ](Plots/plot_precision_summary_modmssm.eps "fig:")]{} Acknowledgments {#acknowledgments .unnumbered} --------------- We would like to thank M. M. Mühlleitner, M. Spira and P. Zerwas for valuable discussions. M.K. acknowledges partial support by the Polish Committee for Scientific Research, Grants 2 P03B 05119 (2003), 5 P03B 12120 (2003), and by the European Community’s Human Potential Programme under contract HPRN-CT-2000-00149 Physics at Colliders. [1]{} M. M. Mühlleitner, M. Krämer, M. Spira, P. M. Zerwas, Phys. Lett. B 508 311-316 (2001), hep-ph/0101083. B. Badelek et al., TESLA Technical Design Report, Part VI, Chapter 1: The Photon Collider at TESLA, DESY 2001-011, ECFA 2001-209, hep-ex/0108012. David M. Asner, Jeffrey B. Gronberg, John F. Gunion, Phys.Rev. D 67 035009 (2003); D. Asner et al., hep-ph/0208219. ATLAS Coll., Technical Design Report, CERN-LHCC 99-14 (1999); CMS Coll., Technical Proposal, CERN-LHCC 94-38 (1994). P. Nieżurawski, A.F. Żarnecki, M. Krawczyk, *New results for $ \gamma \gamma \rightarrow H \rightarrow b \bar{b} $ in SM and MSSM*, talk presented at the 4th ECFA/DESY Workshop, Amsterdam, Netherlands, April 2003. P. Nieżurawski, A.F. Żarnecki, M. Krawczyk, *The SM Higgs-boson production in $ \gamma \gamma \rightarrow h\rightarrow b\bar{b} $ at the Photon Collider at TESLA*, Acta Physica Polonica B 34 177-187 (2003), CERN-TH-2002-166, IFT-2002-29, . P. Nieżurawski, A.F. Żarnecki, M. Krawczyk, *Light Higgs boson production at the Photon Collider at TESLA with an improved background analysis*, International Europhysics Conference on High Energy Physics EPS 2003, Aachen, Germany, abstract 595. V. Telnov, Nucl. Instrum. Meth. A 355, 3 (1995); V. Telnov, *A code PHOCOL for the simulation of luminosities and backgrounds at photon colliders*, talk presented at the Second Workshop of ECFA-DESY Study, Saint Malo, France, April 2002, http://www.desy.de/\~telnov/stmalo/stmalo1.ps.gz; for details of the simulation see also: http://www.desy.de/\~telnov/ggtesla/spectra/. A.F. Żarnecki, *CompAZ: parametrization of the photon collider luminosity spectra*,\ Acta Phys. Polon. B34, 2741-2758 (2003), hep-ex/0207021;\ http://info.fuw.edu.pl/\~zarnecki/compaz/compaz.html. M. Pohl and H. J. Schreiber, DESY-02-061, hep-ex/0206009. T. Kuhl and K. Harder, *B-tagging in SIMDET: 1st application to Higgs*, talk presented at the Second Workshop of ECFA-DESY Study, Saint Malo, France, April 2002, http://www-dapnia.cea.fr/ecfadesy-stmalo/Sessions/Higgs/session1/kuhl.ps. I. F. Ginzburg, G. L. Kotkin, V. G. Serbo and V. I. Telnov, 34, 514 (1981), 34, 491 (1982); Preprint INP 81-50, Novosibirsk, 1981 and Nucl. Instrum. Meth. A 205, 47 (1983); Preprint INP 81-102, Novosibirsk, 1981;\ I. F. Ginzburg, G. L. Kotkin, S. L. Panfil, V. G. Serbo and V. I. Telnov, A219, 5 (1984);\ V. I. Telnov, A294, 72 (1990) and A355, 3 (1995). A. Djouadi, J. Kalinowski and M. Spira, Comput. Phys. Commun. 108, 56 (1998), . T. Sjöstrand, P. Eden, C. Friberg, L. Lonnblad, G. Miu, S. Mrenna and E. Norrbin, Comput. Phys. Commun. 135, 238 (2001), hep-ph/0108264. G. Jikia and S. Söldner-Rembold, Nucl. Instrum. Meth. A 472, 133 (2001), hep-ex/0101056. S. Söldner-Rembold, talk presented at the 10th International Conference on Supersymmetry and Unification of Fundamental Interactions (SUSY02), Hamburg, Germany, Jun 2002. G. Jikia and A. Tkabladze, Nucl. Instrum. Meth. A 355, 81 (1995) and Phys. Rev. D 54, 2030 (1996), hep-ph/9406428. M. Melles and W. J. Stirling, Phys. Rev. D 59, 94009 (1999) and Eur. Phys. J. C 9, 101 (1999), hep-ph/9807332; M. Melles, W. J. Stirling and V. A. Khoze, Phys. Rev. D 61, 54015 (2000), hep-ph/9907238; M. Melles, Nucl. Instrum. Meth.  A 472, 128 (2001), hep-ph/0008125. [^1]: We consider only photon–photon events with $\Wgaga >4 $ GeV.

ArXiv

Michail Tsagris$^1$, Simon Preston$^2$ and Andrew T.A. Wood$^2$ 0.5cm $^1$ Department of Computer Science, University of Crete, Herakleion, Greece, <mtsagris@yahoo.gr>\ $^2$ School of Mathematical Sciences, University of Nottingham, Nottingham Park, Nottingham, UK [**Abstract**]{} In the context of data that lie on the simplex, we investigate use of empirical and exponential empirical likelihood, and Hotelling and James statistics, to test the null hypothesis of equal population means based on two independent samples. We perform an extensive numerical study using data simulated from various distributions on the simplex. The results, taken together with practical considerations regarding implementation, support the use of bootstrap-calibrated James statistic.\ \ **Keywords**: Compositional data, hypothesis testing, Hotelling test, James test, non parametric, empirical likelihood, bootstrap Introduction ============ Data that lie on the the simplex $$\begin{aligned} \label{simplex} S^d=\left\lbrace(x_1,...,x_D)^T \bigg\vert x_i \geq 0,\sum_{i=1}^Dx_i=1\right\rbrace, \end{aligned}$$ where $d=D-1$ are sometimes called compositional data, and they arise in many disciplines, including geology [@ait1982], economics [@fry2000], archaeology [@baxter2005] and political sciences [@rodrigues2009]. There has been extensive and sometimes highly-charged debate over how best to analyse compositional data; see @scealy2014 for a recent review. Some authors strongly advocate the use of a log-ratio transformation [@ait2003], while others advocate different approaches; e.g. [@baxter2005] consider the direct analysis of compositional data with no transformation. However, we stress that the debate concerning how best to analyse compositional data is outside the scope of the present paper. Here, our goal is purely to explore the performance of various nonparametric methods of testing for equality of means of two populations on the simplex using linear statistics based on independent samples from each population. For some readers the main interest may be in seeing the potential of these methods for the analysis of compositional data, while for others the main interest may be in seeing the performance of these nonparametric methods in a highly non-Gaussian setting (due to the simplex being a compact space with boundaries). We consider two forms of nonparametric likelihood: Empirical likelihood (EL) is a nonparametric likelihood which shares many of the properties of parametric likelihoods (see [@owen1988; @owen1990; @owen2001]; [@qin1994]); and Exponential Empirical Likelihood (EEL), due to @efron1981, who obtained it by exponential tilting. EEL has similar first-order asymptotic properties to EL but different second-order properties, e.g. in contrast to EL, which is Bartlett correctable [@diciccio1990], EEL is not Bartlett correctable [@jing1995]. [@zhu2008] consider a correction to EEL. However, in this paper we focus on higher-order corrections based on bootstrap calibration rather than Bartlett correction of other types of analytic correction (see [@hall1990], for discussion of these different approaches, and [@li2011]). Some kind of correction is usually needed in practice unless the sample size is large because, as has been shown in many simulation studies in a variety of contexts, EL and EEL likelihood ratio tests without correction do not do a good job of controlling Type I error; usually the actual Type I error is larger than the nominal Type I level. Examples of such simulation studies include [@diciccio1989], [@fisher1996], [@emerson2009], [@amaral2010] and [@preston2010]. Our results show that, without correction, EL and EEL based tests tend to be less accurate, in terms of control of Type I error, than other nonparametric methods, such as nonparametric bootstrap versions of the Hotelling and James statistics. Moreover, as we shall see, when bootstrap calibration is applied to EL and EEL testing, it only brings the performance of EL and EEL in line with, but does not surpass, the performance of bootstrapped Hotelling and James statistics. In view of the challenging computational issue in higher dimensions of finding points in the intersection of the supports of the two sample EL or EEL likelihoods, we conclude that, from a practical point of view, bootstrapped Hotelling and James statistics are preferable to use in the setting of the paper, since they are much easier to implement than, yet achieve control of Type I error and power which is as good as, that of bootstrap-calibrated EL and EEL tests. The outline of the paper is as follows. In Section 2 we review the test statistics to be studied: parametric and nonparametric bootstrap versions of the Hotelling and James statistics, and EL and EEl statistics with and without bootstrap calibration. In Section 3 we presents the results of an extensive simulation study and we present our conclusions in Section 4. Quadratic tests for two population mean vectors =============================================== The two quadratic-form test statistics we will use are the Hotelling statistic and James statistic defined as follows. Two-sample equality of mean vector test when $\pmb{\Sigma}_1=\pmb{\Sigma}_2$ (Hotelling test) --------------------------------------------------------------------------------------------- If the covariance matrices can be assumed equal, the Hotelling $T^2$ test statistic for two $d$-dimensional samples is given by [@mardia1979] $$\begin{aligned} \label{hotel} T^2=\left(\bar{{\bf x}}_1-\bar{{\bf x}}_2\right)^T\left[{\bf S}_p\left(\frac{1}{n_1}+\frac{1}{n_2}\right) \right]^{-1}\left(\bar{{\bf x}}_1-\bar{{\bf x}}_2\right),\end{aligned}$$ where ${\bf S}_p=\frac{\left(n_1-1\right){\bf S}_1+\left(n_2-1\right){\bf S}_2}{n_1+n_2-2}$ is the pooled covariance matrix with ${\bf S}_1$ and ${\bf S}_2$ being the two unbiased sample covariance matrices $$\begin{aligned} {\bf S}_1 &=& 1/(n_1-1) \sum_{i=1}^{n_1}\left[ (\mathbf{x}_1)_i- \bar{\bf x}_1\right]\left[ (\mathbf{x}_1)_i- \bar{\bf x}_1\right]^T \ \ \text{and} \\ {\bf S}_2 &=& 1/(n_2-1) \sum_{i=1}^{n_1}\left[ (\mathbf{x}_2)_i- \bar{\bf x}_2\right]\left[ (\mathbf{x}_2)_i- \bar{\bf x}_2\right]^T, \end{aligned}$$ where $\bar{{\bf x}}_1$ and $\bar{{\bf x}}_2$ are the two sample means and $n_1$ and $n_2$ are the two sample sizes. Under $H_0$ and when the central limit theorem holds true for each population we have that $$\begin{aligned} T^2 \sim \frac{\left(n_1+n_2\right)d}{n_1+n_2-d+1}F_{d,n_1+n_2-d+1}. \end{aligned}$$ Two-sample equality of mean vector test when $\pmb{\Sigma}_1 \neq \pmb{\Sigma}_2$ (James test) {#james} ---------------------------------------------------------------------------------------------- @james1954 proposed a test for linear form of hypotheses of the population means when the variances are not known. The test statistic for two $d$-dimensional samples is $$\begin{aligned} \label{james2} T^2_u=\left(\bar{{\bf x}}_1-\bar{{\bf x}}_2\right)^T\tilde{{\bf S}}^{-1}\left(\bar{{\bf x}}_1-\bar{{\bf x}}_2\right), \end{aligned}$$ where $\tilde{{\bf S}}=\tilde{{\bf S}}_1+\tilde{{\bf S}}_2=\frac{{\bf S}_1}{n_1}+\frac{{\bf S}_2}{n_2}$. @james1954 suggested that the test statistic is to be compared with $2h\left(\alpha\right)$, a corrected $\chi^2$ quantile whose form is $$\begin{aligned} 2h\left(\alpha\right)=\chi^2_{\nu,1-a}\left(A+B\chi^2_{\nu,1-a}\right),\end{aligned}$$ where $\chi^2_{\nu}$ is a chi-squared random variable with $\nu$ degrees of freedom, such that $P\left[\chi_{\nu}^2 \leq \chi_{\nu,1-\alpha}^2 \right]=1-\alpha$ and $$\begin{aligned} A &=& 1+\frac{1}{2d}\sum_{i=1}^2\frac{\left[\text{tr}\left( \tilde{{\bf S}}^{-1}\tilde{{\bf S}_i}\right)\right]^2}{n_i-1} \ \ \text{and} \\ B &=& \frac{1}{p\left(p+2\right)}\left[\frac{1}{2}\sum_{i=1}^2\frac{\text{tr}\left[\left(\tilde{{\bf S}}^{-1}\tilde{{\bf S}}_i\right)^2\right]}{n_i-1} +\frac{1}{2}\sum_{i=1}^2\frac{\left[\text{tr}\left( \tilde{{\bf S}}^{-1}\tilde{{\bf S}}_i\right)\right]^2}{n_i-1} \right],\end{aligned}$$ @krishnamoorthy2004 showed that under the multivariate normality assumption for each sample $$\begin{aligned} T^2_u \sim \frac{\nu d}{\nu-d+1}F_{d,\nu-d+1} \ \text{approximately},\end{aligned}$$ where $$\begin{aligned} \label{df} \nu=\frac{d+d^2}{\frac{1}{n_1}\left[tr\left(\tilde{{\bf S}}_1\tilde{{\bf S}}^{-1}\right)^2+ \left(tr\left(\tilde{{\bf S}}_1\tilde{{\bf S}}^{-1}\right)\right)^2 \right] + \frac{1}{n_2}\left[ tr\left(\tilde{{\bf S}}_2\tilde{{\bf S}}^{-1}\right)^2+ \left(tr\left(\tilde{{\bf S}}_2\tilde{{\bf S}}^{-1}\right)\right)^2 \right]}. \end{aligned}$$ The advantage of the calibration proposed by @james1954 is that it can be applied to more than two samples, whereas @krishnamoorthy2004 calculated the degrees of freedom of the $F$ distribution for the two samples only. Empirical likelihood for the two sample case {#manova} -------------------------------------------- @jing1995 and @liu2008 described the two-sample hypothesis testing using empirical likelihood. The $2$ constraints imposed by empirical likelihood $$\begin{aligned} \label{elcons} \frac{1}{n_j}\sum_{i=1}^{n_j}\left\lbrace\left[1+\pmb{\lambda}_j^T\left({\bf x}_{ji}-\pmb{\mu} \right)\right]^{-1}\left({\bf x}_{ij}-\pmb{\mu}\right)\right\rbrace={\bf 0}, \ \mbox{$j=1,2$},\end{aligned}$$ where the $\pmb{\lambda}_js$ are Lagrnagian parameters introduced to maximize (\[ellambda\]). The probabilities of each of the $j$ samples have the following form $$\begin{aligned} \label{pis} p_{ji}=\frac{1}{n_j} \left[1+\pmb{\lambda}_j^T \left({\bf x}_{ji}-\pmb{\mu} \right)\right]^{-1},\end{aligned}$$ where $\pmb{\lambda}_1 + \pmb{\lambda}_2=0$ is a convenient constraint that can be used. The log-likelihood ratio test statistic can be written as $$\begin{aligned} \label{ellambda} \Lambda = \sum_{j=1}^2\sum_{i=1}^{n_j}\log{n_jp_{ij}} = \sum_{j=1}^2n_j\left({\bf \bar{x}}_j-\pmb{\mu}\right)^T{\bf S}_j\left(\pmb{\mu}\right)^{-1}\left({\bf \bar{x}}_j-\pmb{\mu}\right)+o_p\left(1\right),\end{aligned}$$ where ${\bf S}_j\left(\pmb{\mu}\right)=\frac{n_j-1}{n_j}{\bf S}_j+\left(\bar{{\bf x}}-\pmb{\mu}\right)\left(\bar{{\bf x}}-\pmb{\mu} \right)^T$ with ${\bf S}_j$ denoting the sample covariance matrix. The maximization of (\[ellambda\]) is with respect to the $\pmb{\lambda}_js$. Asymptotically, under $H_0$ $\Lambda \sim \chi^2_d$, since ${\bf S}\left(\pmb{\mu}\right)\overset{p}{\rightarrow} \pmb{\Sigma}$, where $\pmb{\Sigma}$ is the population covariance matrix. A proof of the asymptotic distribution of the test statistic when we have more than two samples, both in the univariate and multivariate case, can be found in @owen2001. However, asymptotically, this test statistic is the same as the test statistic suggested by @james1954. As mentioned in Section \[james\], @james1954 used a corrected $\chi^2$ distribution whose form in the two sample means cases is given in Section \[james\]. Therefore we have strong grounds to suggest James corrected $\chi^2$ distribution for calibration of the empirical likelihood test statistic instead of the classical $\chi^2$ distribution. Maximization of (\[ellambda\]) with respect to a scalar $\lambda$, in the univariate case, is easy since a simple search over an interval is enough. In the multivariate case though, the difficulty increases with the dimensionality. Another important issue we highlight is that empirical likelihood test statistic will not be computed if $\pmb{\mu}$ lies within the convex hull of the data [@emerson2009]. This issue becomes more crucial again as the dimensions increase. As for the distribution of the test statistic under $H_1$ let us assume that each mean $\pmb{\mu}_j$ deviates from the common mean $\pmb{\mu}$ by a quantity which is a function of the sample covariance matrix, the sample size plus a constant vector $\pmb{\tau_j}$ for each sample. We can then write the mean as a function of the covariance matrix and of the sample size $$\begin{aligned} \label{alter1} \pmb{\mu}_j=\pmb{\mu}+\frac{\pmb{\Sigma} _j^{1/2}}{\sqrt{n_j}}\pmb{\tau_j},\end{aligned}$$ where $\pmb{\Sigma}_j^{1/2}$ is the true covariance matrix of the $j$-th sample and $$\begin{aligned} {\bf z}_j=\pmb{\Sigma}_j^{-1/2}\sqrt{n_j}\left({\bf \bar{x}}_j-\pmb{\mu}\right).\end{aligned}$$ Since ${\bf S}\left(\pmb{\mu}\right)\overset{p}{\rightarrow} \pmb{\Sigma}$ we have that $$\begin{aligned} \label{L1} \Lambda = \sum_{j=1}^2g_j\left(\pmb{\tau}_j\right)=\sum_{j=1}^k\left\lbrace\left({\bf z}_j-\pmb{\tau}_j\right)^T\left({\bf z}_j-\pmb{\tau}_j\right) \left[1+\frac{\left({\bf z}_j-\pmb{\tau}_j\right)^T\left({\bf z}_j-\pmb{\tau}_j\right)-1}{n_j} \right]^{-1}\right\rbrace,\end{aligned}$$ Asymptotically, the scalar factor $\left[1+\frac{\left({\bf z}_j-\pmb{\tau}_j\right)^T\left({\bf z}_j-\pmb{\tau}_j\right)-1}{n_j} \right]^{-1}$ will disappear and $H_1$ (\[L1\]) will be equal to the sum of $k-1$ independent non-central $\chi^2$ variables, where each of them have a non-centrality parameter equal to $\left|\left|\pmb{\tau}_j\right|\right|^2$. Consequently, (\[L1\]) follows asymptotically a non-central $\chi^2$ with non-centrality parameter $\sum_{j=1}^k\left|\left|\tau_j^2\right|\right|$. Exponential empirical likelihood for the two sample case {#exponential} -------------------------------------------------------- Exponential empirical likelihood or exponential tilting was first introduced by @efron1981 as a way to perform a “tilted” version of the bootstrap for the one sample mean hypothesis testing. Similarly to the empirical likelihood, positive weights $p_i$, which sum to one, are allocated to the observations, such that the weighted sample mean $\bar{x}$ is equal to a population mean $\mu$ under the null hypothesis. Under the alternative hypothesis the weights are equal to $\frac{1}{n}$, where $n$ is the sample size. The choice of $p_is$ will minimize the Kullback-Leibler distance from $H_0$ to $H_1$ [@efron1981] $$\begin{aligned} \label{expKL} D\left(L_0,L_1\right)=\sum_{i=1}^np_i\log\left(np_i\right),\end{aligned}$$ subject to the constraint $$\begin{aligned} \label{expconstrain} \sum_{i=1}^np_i{\bf x}_i=\pmb{\mu}. \end{aligned}$$ The probabilities take the following form $$\begin{aligned} p_i=\frac{e^{\pmb{\lambda}^T{\bf x}_i}}{\sum_{j=1}^ne^{\pmb{\lambda}^T{\bf x}_j}}\end{aligned}$$ and the constraint in (\[expconstrain\]) becomes $$\begin{aligned} \frac{\sum_{i=1}^ne^{\pmb{\lambda}^T{\bf x}_i}\left({\bf x}_i-\mu\right)}{\sum_{j=1}^ne^{\pmb{\lambda}^T{\bf x}_j}}=0 \Rightarrow \frac{\sum_{i=1}^nx_ie^{\pmb{\lambda}^T{\bf x}_i}}{\sum_{j=1}^ne^{\pmb{\lambda}^T{\bf x}_j}}-\mu=0.\end{aligned}$$ Similarly to the univariate empirical likelihood a numerical search over $\lambda$ is required. We can derive the asymptotic form of the test statistic (\[expKL\]) in the two sample means case but in a simpler form, using a rather somewhat path to the one @jing1997 followed, but for the multivariate case. The three constraints are $$\begin{aligned} \label{constraint} \begin{array}{ccc} \left(\sum_{j=1}^{n_1}e^{\pmb {\lambda}_1^T{\bf x}_j}\right)^{-1}\left(\sum_{i=1}^{n_1}{\bf x}_ie^{\pmb{\lambda}_1^T {\bf x}_i}\right) -\pmb{\mu} & = & {\bf 0} \\ \left(\sum_{j=1}^{n_2}e^{\pmb {\lambda}_2^T{\bf y}_j}\right)^{-1}\left(\sum_{i=1}^{n_2}{\bf y}_ie^{\pmb{\lambda}_2^T {\bf y}_i}\right) -\pmb{\mu} & = & {\bf 0} \\ n_1\pmb{\lambda}_1+n_2\pmb{\lambda}_2 & = & {\bf 0}. \end{array}\end{aligned}$$ Similarly to the empirical likelihood the sum of a linear combination of the $\pmb{\lambda}s$ is set to zero. We can equate the first two constraints of (\[constraint\]) $$\begin{aligned} \label{eel2} \left(\sum_{j=1}^{n_1}e^{\pmb {\lambda}_1^T{\bf x}_j}\right)^{-1}\left(\sum_{i=1}^{n_1}{\bf x}_ie^{\pmb{\lambda}_1^T {\bf x}_i}\right)= \left(\sum_{j=1}^{n_2}e^{\pmb {\lambda}_2^T{\bf y}_j}\right)^{-1}\left(\sum_{i=1}^{n_2}{\bf y}_ie^{\pmb{\lambda}_2^T {\bf y}_i}\right).\end{aligned}$$ Also, we can write the third constraint of (\[constraint\]) as $\pmb{\lambda}_2=-\frac{n_1}{n_2}\pmb{\lambda}_1$ and thus rewrite (\[eel2\]) as $$\begin{aligned} \left(\sum_{j=1}^{n_1}e^{\pmb{\lambda}^T{\bf x}_j}\right)^{-1}\left(\sum_{i=1}^{n_1}{\bf x}_ie^{\pmb{\lambda}^T {\bf x}_i}\right) = \left(\sum_{j=1}^{n_2}e^{-\frac{n_1}{n_2}\pmb{\lambda}^T{\bf y}_j}\right)^{-1}\left(\sum_{i=1}^{n_2}{\bf y}_ie^{-\frac{n_1}{n_2}\pmb{\lambda}^T {\bf y}_i}\right).\end{aligned}$$ This trick allows us to avoid the estimation of the common mean. It is not possible though to do this in the empirical likelihood method. Instead of minimisation of the sum of the one-sample test statistics from the common mean, we can define the probabilities by searching for the $\pmb{\lambda}$ which makes the last equation hold true. The third constraint of (\[constraint\]) is a convenient constraint, but @jing1997 mentions that even though as a constraint is simple it does not lead to second-order accurate confidence intervals unless the two sample sizes are equal. The asymptotic form of the test statistic under $H_1$ is equal to $$\begin{aligned} \label{L2} \Lambda = \sum_{j=1}^2\left\lbrace\left[\pmb{\Sigma}_j^{-1/2}\sqrt{n_j}\left({\bf \bar{x}}_j-\pmb{\mu}\right)-\pmb{\tau_j}\right]^T \left[\pmb{\Sigma}_j^{-1/2}\sqrt{n_j}\left({\bf \bar{x}}_j-\pmb{\mu}\right)-\pmb{\tau_j}\right]\left(1-\frac{1}{n_j} \right)^{-1}\right\rbrace\end{aligned}$$ When the sample sizes are large, the scalar $\left(1-\frac{1}{n_j} \right)^{-1}$ will disappear, and thus the asymptotic distribution is the sum of $k-1$ independent non-central $\chi^2$ distributions, where each of them has a non central parameter equal to $\left|\left|\tau_j\right|\right|^2$. $\Lambda$ (\[L2\]) follows asymptotically a non-central $\chi^2$ distribution with non-centrality parameter $\sum_{j=1}^k\left|\left|\tau_j^2\right|\right|$. Thus, under $H_1$, to the leading term the distribution of the exponential empirical likelihood test statistic is the same as that of the empirical likelihood. Non parametric bootstrap hypothesis testing ------------------------------------------- The non-parametric bootstrap procedure that we use in §3 is as follows. 1. Define the test statistic T as one of (\[hotel\]), (\[james2\]), (\[L1\]) or (\[L2\]) and define $T_{obs}$ to be $T$ calculated for the available data $(\mathbf{x}_1)_1, ... , (\mathbf{x}_1)_{n_1}$ and $(\mathbf{x}_2)_1, ... , (\mathbf{x}_2)_{n_2}$ with means $\bar{\bf x}_1$ and $\bar{\bf x}_2$ and covariance matrices ${\bf S}_1$ and ${\bf S}_2$ 2. Transform the data so that the null hypothesis is true $$\begin{aligned} (\mathbf{y}_1)_i = (\mathbf{x}_1)_i - \bar{\bf x}_1 + \hat{\pmb{\mu}}_c \ \ \text{and} \ \ (\mathbf{y}_2)_i = (\mathbf{x}_2)_i - \bar{\bf x}_2 + \hat{\pmb{\mu}}_c,\end{aligned}$$ where $\hat{\pmb{\mu}}_c = \left[(n_1-1){\bf S}_1^{-1} + (n_2-1){\bf S}_2^{-1} \right]^{-1}\left([n_1-1){\bf S}_1^{-1}\bar{\bf x}_1 + (n_2-1){\bf S}_2^{-1}\bar{\bf x}_2 \right]^T$ is the estimated common mean under the null hypothesis. 3. Generate two bootstrap samples by sampling with replacement $(y_1)_1,\ldots,(y_1)_{n_1}$ and $(y_2)_1,\ldots ,(y_2)_{n_2}$. 4. Define $T_b$ as the test statistic $T$ calculated for the bootstrap sample in step 3. 5. Repeat steps 3 and 4 B times to generate bootstrap statistics $T_b^1, ..., T_b^B$ and calculate the bootstrap p-value as $$\begin{aligned} p-value = \frac{\sum_{i=1}^B{\bf 1}\left(T^i_b>T_{obs}\right)+1}{B+1}. \end{aligned}$$ Simulations studies for the performance of the testing procedures applied to two compositional sample means {#comparisons} =========================================================================================================== The goal of this manuscript is to draw conclusions, via extensive simulation studies, about the testing procedures when applied to compositional data. We will compare the testing procedures and see if one is to be preferred to the others. In both scenarios considered here, the two populations have the same mean vectors but different covariance matrix structures. At first, we will apply the following transformation to compositional data $$\begin{aligned} {\bf y}={\bf H}{\bf x},\end{aligned}$$ where $\bf H$ is the Helmert sub-matrix (i.e. the Helmert matrix [@helm1965] with the the first row omitted) and ${\bf x} \in \mathbb{S}^d$ (\[simplex\]). The multiplication by the Helmert sub-matrix is essentially a linear transformation of the data (and of the simplex). This means that even if we applied the testing procedures on the raw (un-transformed) data the results would be the same (empirical likelihood is invariant under invertible transformations of the data, [@owen2001]). We do it though a) for convenience purposes and b) to speed up the computational time required by empirical and exponential empirical likelihood. Note that the Helmert sub-matrix appears also in the isometric log-ratio transformation [@ilr2003] and in [@tsagris2011]. The comparison of all the testing procedures was in terms of the probability of type I error and of the power. Bootstrap calibration was necessary for all tests in the small samples case even though it is quite computationally intensive for the empirical and exponential empirical likelihoods. The number of bootstrap replications was equal to $299$ and $1000$ simulations were performed. In each case a 4-dimensional simplex was used. When the estimated probability of Type I error falls within $\left(0.0365,0.0635\right)$ (theoretical $95\%$ confidence interval based on $1000$ simulations) we have evidence that the test attains the correct probability of Type I error. For the implementation of the empirical likelihood the R package *emplik* [@emplik2013] was used. The procedure was to calculate the common mean which minimizes the sum of the two empirical likelihood tests [@amaral2010]. We used the $\chi^2$ corrected distribution suggested by @james1954 and the $F$, with degrees of freedom given in (\[df\]), suggested by @krishnamoorthy2004. The EL and EEL stand for empirical likelihood and exponential empirical likelihood respectively. The term inside the parentheses indicates the calibration, $(\chi^2)$, $(F)$ or $(bootstrap)$, corresponding to the $\chi^2$ or the $F$ distribution and bootstrap respectively. Scenario 1. Simulated data from Dirichlet populations {#example_1} ----------------------------------------------------- Data were generated from two Dirichlet populations such that the two arithmetic means in $\mathbb{S}^d$ are the same. The first population was $\text{Dir}\left(0.148, 0.222, 0.296, 0.333\right)$ and the second came from a mixture of two Dirichlets: $$\begin{aligned} 0.3 \times \text{Dir}\left(0.889, 1.333, 1.778, 2.000\right)+0.7 \times \text{Dir}\left(1.481, 2.222, 2.963, 3.333\right).\end{aligned}$$ As for the estimation of the power, we will keep the mean vector of the mixture of two Dirichlets constant and change the mean vector of the other Dirichlet population. We select the fourth component (it has the largest variance) and change it so that the whole mean vector is moving along a straight line. For every change in the mean of this component, there is the same (across all three components) change in the opposite direction for the other three components. The second compositional mean vector is written as $$\begin{aligned} \label{change1} \pmb{\mu}=\left(\mu_1-\frac{\delta}{3},\mu_2-\frac{\delta}{3},\mu_3-\frac{\delta}{3},\mu_4+\delta\right),\end{aligned}$$ where $\delta$ ranges from $-0.21$ up to $0.21$ each time at a step equal to $0.03$. Scenario 2. Simulated data from different distributions {#example_2} ------------------------------------------------------- We will now present an example where the two datasets come from populations with different distributions, a Dirichlet and a logistic normal. $20,000,000$ observations from a were generated from $N_3\left(\pmb{\mu},\pmb{\Sigma}\right)$, where $$\begin{aligned} \pmb{\mu}=\left(1.548,0.747,-0.052\right)^T \ \text{and} \ \pmb{\Sigma}= \left[ \begin{array}{ccc} 0.083 & 0.185 & -0.169 \\ 0.185 & 0.547 & -0.671 \\ -0.169 & -0.671 & 1.110 \end{array} \right] \end{aligned}$$ Then the inverse of the additive log-ratio transformation [@ait2003] $$\begin{aligned} \label{alrinv} x_i = \frac{e^{v_i}}{1+\sum_{j=1}^de^{v_j}}, \ \text{for} \ i=1,\ldots,d \ \ \text{and} \ \ x_D= \frac{1}{1+\sum_{j=1}^de^{v_j}}.\end{aligned}$$ was applied to map the observations onto the simplex. The empirical population mean vector was equal to $\left(0.483,0.249,0.163,0.105\right)^T$. We then generated observations from the same multivariate normal distribution on $R^3$ and observations from a mixture of two Dirichlet distributions $$\begin{aligned} 0.3\times \text{Dir}\left(0.483,0.249,0.163,0.105\right)+0.7\times \text{Dir}\left(3.381, 1.743, 1.141, 0.735\right).\end{aligned}$$ As for the estimation of the powers, the direction of the alternatives was the same as before, but the the first component (it had the second largest variance) was changing now. The mean of the logistic normal distribution was kept constant. We chose the second sample (the mixture of two Dirichlets) and changed its first component. Every change in the mean of the first component resulted in an equal change of the opposite direction for the other three components $$\begin{aligned} \label{change2} \pmb{\mu}=\left(\mu_1+\delta,\mu_2-\frac{\delta}{3},\mu_3-\frac{\delta}{3},\mu_4-\frac{\delta}{3}\right),\end{aligned}$$ where $\delta$ ranges from $-0.21$ up to $0.21$ each time at a step equal to $0.03$. Table \[probs\] shows the $95\%$ confidence intervals for different levels of probabilities calculated using the Monte Carlo simulations error based on $1000$ simulations. They will help us compare the powers of the different testing procedures as a guide of how large is the simulations error at different levels of power. ------------- -------------------- ------------------ ------------------ ------------------ ------------------ Probability $0.05$ $0.10$ $0.20$ $0.30$ $0.40$ Intervals $(0.0365, 0.0635)$ $(0.081, 0.119)$ $(0.175, 0.225)$ $(0.272, 0.328)$ $(0.370, 0.430)$ Probability $0.50$ $0.60$ $0.70$ $0.80$ $0.90$ Intervals $(0.469, 0.531)$ $(0.570, 0.630)$ $(0.672, 0.728)$ $(0.775, 0.825)$ $(0.881, 0.919)$ ------------- -------------------- ------------------ ------------------ ------------------ ------------------ : $95\%$ confidence intervals for different levels of probability.[]{data-label="probs"} Results ------- ### Type I error with equal sample sizes The sample sizes were set equal for both samples and equal to $15$, $30$, $50$ and $100$. When the sample sizes were equal to $15$ the true Type I error was not achieved by any procedure. Bootstrap calibration however corrected the size of all testing procedures and for the sample sizes. For the sample sizes $30$ and $50$ we can see that empirical and exponential empirical likelihood calibrated with the $F$ distribution performed better than the other testing procedures. What is more, is that James test when calibrated using an $F$ rather than a corrected $\chi^2$ distribution, shows no significant improvement. Finally, when the sample sizes are large all procedures attain the nominal Type I error of the test. ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- Testing procedure $n=15$ $n=30$ $n=50$ $n=100$ $n=15$ $n=30$ $n=50$ $n=100$ Hotelling 0.097 0.067 0.073 [**0.05**]{} 0.09 0.083 0.078 [**0.061**]{} James($\chi^2$) 0.092 0.065 0.069 [**0.049**]{} 0.087 0.08 0.072 [**0.059**]{} James($F$) 0.09 0.065 0.069 [**0.048**]{} 0.078 0.078 0.072 [**0.059**]{} EEL($\chi^2$) 0.139 0.075 0.075 [**0.055**]{} 0.184 0.112 0.089 [**0.065**]{} EL($\chi^2$) 0.126 0.066 0.071 [**0.051**]{} 0.154 0.097 0.08 [**0.062**]{} EEL($F$) 0.095 [**0.056**]{} [**0.06**]{} [**0.05**]{} 0.114 0.083 0.076 [**0.062**]{} EL($F$) 0.08 [**0.052**]{} [**0.054**]{} [**0.046**]{} 0.099 0.072 0.064 [**0.056**]{} Hotelling(bootstrap) [**0.046**]{} [**0.052**]{} [**0.061**]{} [**0.041**]{} [**0.047**]{} [**0.055**]{} [**0.056**]{} [**0.05**]{} James(bootstrap) [**0.044**]{} [**0.052**]{} [**0.061**]{} [**0.041**]{} [**0.046**]{} [**0.055**]{} [**0.056**]{} [**0.05**]{} EEL(bootstrap) [**0.051**]{} [**0.046**]{} [**0.058**]{} [**0.043**]{} [**0.05**]{} [**0.059**]{} [**0.056**]{} [**0.057**]{} EL(bootstrap) [**0.049**]{} [**0.046**]{} [**0.057**]{} [**0.043**]{} [**0.046**]{} [**0.054**]{} [**0.054**]{} [**0.057**]{} ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- : Estimated probability of Type I error using different tests and a variety of calibrations. The nominal level of the Type I error was equal to $0.05$. The numbers in bold indicate that the estimated probability was within the acceptable limits.[]{data-label="ex_1"} ### Type I error with unequal sample sizes The second sample, which came from the mixture of two Dirichlets, had observations which were less spread (its covariance determinant was smaller) and for this reason it will now have a larger size. We can see in Table \[ex\_1a\] that Hotelling’s test clearly fails as expected. However, when bootstrap calibrated, it works reasonably well for large sample sizes. The $F$ calibration of the empirical and exponential empirical likelihood works better than the $\chi^2$ calibration. However, bootstrap is again necessary for the medium sizes. The conclusion is again that the bootstrap computation of the p-values does a very good job. ---------------------- ---------- --------------- --------------- --------------- --------------- --------------- Testing $n_1=15$ $n_1=30$ $n_1=50$ $n_1=15$ $n_1=30$ $n_1=50$ procedure $n_2=30$ $n_2=50$ $n_2=100$ $n_2=30$ $n_2=50$ $n_2=100$ Hotelling 0.222 0.154 0.160 0.236 0.189 0.178 James($\chi^2$) 0.142 0.086 [**0.053**]{} 0.106 0.087 [**0.054**]{} James($F$) 0.134 0.08 [**0.053**]{} 0.113 0.089 [**0.054**]{} EEL($\chi^2$) 0.174 0.08 [**0.049**]{} 0.211 0.128 0.072 EL($\chi^2$) 0.165 0.072 [**0.043**]{} 0.183 0.108 0.065 EEL($F$) 0.115 [**0.053**]{} [**0.039**]{} 0.139 0.089 [**0.056**]{} EL($F$) 0.104 [**0.045**]{} [**0.038**]{} 0.114 0.076 [**0.047**]{} Hotelling(bootstrap) 0.078 [**0.055**]{} [**0.05**]{} 0.073 [**0.060**]{} [**0.046**]{} James(bootstrap) 0.075 [**0.052**]{} [**0.04**]{} [**0.062**]{} [**0.055**]{} [**0.035**]{} EEL(bootstrap) 0.074 [**0.041**]{} [**0.037**]{} [**0.057**]{} [**0.062**]{} [**0.043**]{} EL(bootstrap) 0.072 [**0.039**]{} [**0.037**]{} [**0.052**]{} [**0.061**]{} [**0.035**]{} ---------------------- ---------- --------------- --------------- --------------- --------------- --------------- : Estimated probability of Type I error using different tests and a variety of calibrations. The nominal level of the Type I error was equal to $0.05$. The numbers in bold indicate that the estimated probability was within the acceptable limits.[]{data-label="ex_1a"} ### Estimated power of the tests with equal sample sizes Table \[pow\_1a\] shows the power of these testing procedures under some alternatives for different sample sizes. We have included four different sample sizes in the simulations studies when the null hypothesis is true. But, when examining the power of the testing procedures we considered only the bootstrap calibrated procedures. The reason for this is that the testing procedures were size correct when bootstrap calibration was implemented. -------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=15 Hotelling(bootstrap) 0.414 0.292 0.194 0.114 0.086 0.050 0.061 0.062 0.082 0.116 0.149 0.206 0.292 0.366 James(bootstrap) 0.410 0.291 0.189 0.114 0.086 0.048 0.061 0.062 0.081 0.115 0.149 0.208 0.291 0.362 Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=30 Hotelling(bootstrap) 0.803 0.606 0.420 0.262 0.153 0.082 0.042 0.069 0.100 0.154 0.258 0.359 0.524 0.630 James(bootstrap) 0.800 0.606 0.420 0.260 0.151 0.082 0.042 0.069 0.101 0.153 0.258 0.356 0.522 0.626 EEL(bootstrap) 0.729 0.555 0.395 0.250 0.145 0.092 0.041 0.061 0.081 0.125 0.229 0.307 0.481 0.583 EL(bootstrap) 0.734 0.554 0.403 0.254 0.144 0.095 0.042 0.063 0.082 0.129 0.234 0.323 0.498 0.600 Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=50 Hotelling(bootstrap) 0.966 0.884 0.726 0.472 0.246 0.111 0.051 0.094 0.117 0.246 0.411 0.582 0.755 0.897 James(bootstrap) 0.966 0.884 0.726 0.472 0.245 0.111 0.051 0.094 0.117 0.245 0.411 0.579 0.755 0.896 Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=100 Hotelling(bootstrap) 1.000 0.995 0.970 0.807 0.527 0.210 0.075 0.099 0.231 0.463 0.755 0.927 0.984 0.999 James(bootstrap) 1.000 0.995 0.970 0.807 0.527 0.210 0.075 0.099 0.231 0.463 0.755 0.927 0.984 0.999 EEL(bootstrap) 1.000 0.995 0.970 0.805 0.527 0.227 0.084 0.093 0.221 0.471 0.768 0.936 0.988 0.999 EL(bootstrap) 1.000 0.995 0.970 0.805 0.530 0.228 0.082 0.094 0.223 0.474 0.771 0.939 0.989 0.999 -------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- In all cases we can see that there little difference between the two quadratic tests when calibrated with bootstrap. What is evident from all testing procedures is that as the sample size increases the powers increase as expected. When looking at the case when the sample size is equal to $30$ (Table \[pow\_1a\]) we see that the power of the quadratic tests is higher than the power of the empirical likelihoods. When the alternative (change in the fourth component) is with a negative sign the power is higher than when the alternative is with a positive sign. But as we move towards the null hypothesis the differences between the two types of testing procedures decrease. When the sample sizes are equal to $100$ there are almost no differences between the quadratic tests and the empirical likelihood methods (Table \[pow\_1a\]). As seen from Table \[ex\_1a\] when the sample sizes are small, no procedure managed to attain the correct size. The $F$ calibration of the empirical likelihoods and bootstrap calibration of all tests decreased the Type I error, yet not enough. When the sample sizes increase, all the empirical likelihood methods estimate the probability of Type I error correctly only when the $F$ or bootstrap calibration is applied. As for the quadratic tests, bootstrap calibration has proved very useful too. Finally, when the sample sizes are large we can see that Hotelling test is not size correct, as expected (Hotelling assumes equality of the covariance matrices), but all the other testing procedures estimate the probability of Type I error within the acceptable limits regardless of bootstrap calibration. Table \[pow\_3a\] presents the estimated powers of the bootstrap calibrated testing procedures. The empirical likelihood methods with bootstrap was computationally heavy and for this reason we estimated the powers of these two methods only for two sample sizes, $30$ and $100$. -------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=15 Hotelling(bootstrap) 0.546 0.409 0.237 0.152 0.060 0.049 0.037 0.094 0.148 0.244 0.306 0.468 0.593 0.718 James(bootstrap) 0.544 0.408 0.233 0.150 0.060 0.050 0.034 0.091 0.148 0.241 0.304 0.468 0.591 0.716 Sample Testing $\delta$) size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=30 Hotelling(bootstrap) 0.942 0.833 0.665 0.403 0.214 0.081 0.045 0.103 0.207 0.347 0.530 0.693 0.796 0.889 James(bootstrap) 0.941 0.832 0.663 0.404 0.212 0.081 0.045 0.104 0.207 0.347 0.530 0.694 0.795 0.888 EEL(bootstrap) 0.919 0.817 0.648 0.433 0.251 0.112 0.070 0.100 0.215 0.322 0.511 0.694 0.817 0.884 EL(bootstrap) 0.913 0.813 0.638 0.422 0.236 0.099 0.064 0.102 0.212 0.324 0.519 0.696 0.813 0.886 Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=50 Hotelling(bootstrap) 0.997 0.986 0.919 0.756 0.440 0.200 0.056 0.123 0.276 0.514 0.741 0.901 0.968 0.987 James(bootstrap) 0.997 0.987 0.920 0.755 0.440 0.200 0.057 0.123 0.276 0.515 0.741 0.901 0.968 0.987 Sample Testing $\delta$ size procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} n=100 Hotelling(bootstrap) 1.000 1.000 0.999 0.982 0.817 0.426 0.105 0.169 0.514 0.858 0.988 0.998 1.000 1.000 James(bootstrap) 1.000 1.000 0.999 0.982 0.817 0.426 0.106 0.169 0.514 0.858 0.988 0.998 1.000 1.000 EEL(bootstrap) 1.000 1.000 0.999 0.987 0.835 0.461 0.127 0.146 0.483 0.839 0.986 0.997 1.000 1.000 EL(bootstrap) 1.000 1.000 0.999 0.987 0.837 0.462 0.127 0.143 0.479 0.840 0.986 0.998 1.000 1.000 -------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- -------------- -------------- -------------- -------------- -------------- -------------- -------------- Similarly to the previous example when both samples have the same size and come from Dirichlet populations the power of the James and Hotelling tests are very similar when bootstrap is employed. When the sample sizes are equal to $30$ the quadratic tests exhibit higher powers than than the empirical likelihood methods in the case of a negative change in the first component (see Table \[pow\_3a\]). This is not true though when the change is positive. In addition, when the negative change gets closer to zero, the power of the empirical likelihood methods is better than the power of the quadratic tests and when the positive change gets closer to zero the opposite is true. When the sample sizes are large (equal to $100$) the quadratic tests and the empirical likelihood methods seem to perform equally well as seen in Table \[pow\_3a\]. But, as the change approaches zero from the negative side, we can see that the tests based on the empirical likelihoods reject the null hypothesis more times than the quadratic tests and the converse is true when the change approaches zero from the positive side. [cc]{}\ ![Estimated powers for a range of alternatives. The sample sizes are equal to (a) $30$ and (b) $100$ for each sample. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_1"}](powdiri_1.eps "fig:") & ![Estimated powers for a range of alternatives. The sample sizes are equal to (a) $30$ and (b) $100$ for each sample. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_1"}](powdiri_2.eps "fig:")\ $n_1 = n_2 = 30$ & $n_1 = n_2 = 30$\ \ ![Estimated powers for a range of alternatives. The sample sizes are equal to (a) $30$ and (b) $100$ for each sample. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_1"}](powlogi_1.eps "fig:") & ![Estimated powers for a range of alternatives. The sample sizes are equal to (a) $30$ and (b) $100$ for each sample. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_1"}](powlogi_2.eps "fig:")\ $n_1 = n_2 = 100$ & $n_1 = n_2 = 100$\ ### Estimated power of the tests with unequal sample sizes The sample sizes of the two groups are the same as before, the second sample, which came from one Dirichlet population and was less spread and thus had larger size. The direction of the alternatives was the same as in the case of equal sample sizes. It is worthy to mention that when we had relatively small samples ($n_1=15$ and $n_2=30$) none of the tests was size correct even after bootstrap calibration was applied (see Table \[ex\_1a\]). Thus, we estimated the powers for the other two combinations of the sample sizes. Hotelling test seemed to perform slightly better than James in the small samples but in overall there was almost no difference between them. This difference was more obvious in the unequal sample sizes case, where James test showed evidence that is slightly more powerful than Hotelling, especially in the case where the change in the fourth component is positive (see Table \[pow\_2\]). When both sample sizes are large though, the powers of the testing procedures are almost the same. ----------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- Sample Testing $\delta$ sizes procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} $n_1=30$ Hotelling(bootstrap) 0.958 0.835 0.642 0.400 0.232 0.110 0.052 $n_2=50$ James(bootstrap) 0.964 0.844 0.644 0.423 0.246 0.115 0.056 EEL(bootstrap) 0.950 0.829 0.627 0.413 0.246 0.122 0.060 EL(bootstrap) 0.951 0.829 0.631 0.418 0.250 0.120 0.063 $\delta$ [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} $n_1=30$ Hotelling(bootstrap) 0.058 0.115 0.211 0.336 0.512 0.709 0.831 $n_2=50$ James(bootstrap) 0.062 0.131 0.225 0.381 0.563 0.754 0.874 EEL(bootstrap) 0.051 0.119 0.216 0.358 0.542 0.731 0.866 EL(bootstrap) 0.050 0.122 0.218 0.367 0.551 0.745 0.873 Sample Testing $\delta$ sizes procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} $n_1=50$ Hotelling(bootstrap) 0.999 0.991 0.934 0.771 0.471 0.196 0.069 $n_2=100$ James(bootstrap) 0.999 0.990 0.932 0.766 0.473 0.199 0.074 EEL(bootstrap) 0.998 0.988 0.927 0.762 0.462 0.205 0.082 EL(bootstrap) 0.998 0.989 0.929 0.763 0.466 0.206 0.083 $\delta$ [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} $n_1=50$ Hotelling(bootstrap) 0.077 0.176 0.411 0.648 0.852 0.965 0.993 $n_2=100$ James(bootstrap) 0.082 0.186 0.440 0.681 0.871 0.974 0.993 EEL(bootstrap) 0.083 0.180 0.434 0.673 0.868 0.971 0.994 EL(bootstrap) 0.088 0.181 0.438 0.677 0.871 0.973 0.995 ----------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- : Scenario 1. Estimated powers of the tests tests with bootstrap calibration when the sample sizes are different. The alternatives denote the change ($\delta$) in the $4$th component (\[change1\]).[]{data-label="pow_2"} It is worthy to mention that in when we had relatively small samples ($n_1=15$ and $n_2=30$) only James test was size correct after bootstrap calibration (see Table \[ex\_1a\]). The alternatives in this case were chosen as in the case of equal sample sizes. We chose the second mean of the mixture of two Dirichlet populations and changed it. The second sample (from the mixture of Dirichlet distributions) had always smaller size since its covariance determinant was larger. ----------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- Sample Testing $\delta$ sizes procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} $n_1=50$ Hotelling(bootstrap) 0.957 0.868 0.697 0.466 0.245 0.086 0.039 $n_2=30$ James(bootstrap) 0.940 0.825 0.656 0.417 0.196 0.070 0.040 EEL(bootstrap) 0.924 0.792 0.657 0.437 0.242 0.100 0.060 EL(bootstrap) 0.920 0.790 0.639 0.423 0.225 0.092 0.055 $\delta$ [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} $n_1=50$ Hotelling(bootstrap) 0.106 0.245 0.417 0.572 0.752 0.853 0.903 $n_2=30$ James(bootstrap) 0.092 0.218 0.388 0.534 0.716 0.815 0.866 EEl(bootstrap) 0.095 0.217 0.369 0.514 0.715 0.829 0.883 EL(bootstrap) 0.095 0.220 0.366 0.510 0.717 0.831 0.879 Sample Testing $\delta$ sizes procedure [**-0.21**]{} [**-0.18**]{} [**-0.15**]{} [**-0.12**]{} [**-0.09**]{} [**-0.06**]{} [**-0.03**]{} $n_1=100$ Hotelling(bootstrap) 0.998 0.990 0.940 0.762 0.480 0.222 0.060 $n_2=50$ James(bootstrap) 0.996 0.982 0.919 0.721 0.441 0.184 0.049 EEL(bootstrap) 0.997 0.988 0.927 0.745 0.482 0.231 0.078 EL(bootstrap) 0.997 0.988 0.925 0.741 0.472 0.223 0.078 $\delta$ [**0.03**]{} [**0.06**]{} [**0.09**]{} [**0.12**]{} [**0.15**]{} [**0.18**]{} [**0.21**]{} $n_1=100$ Hotelling(bootstrap) 0.131 0.307 0.598 0.803 0.943 0.980 0.997 $n_2=50$ James(bootstrap) 0.112 0.282 0.552 0.758 0.909 0.967 0.986 EEL(bootstrap) 0.103 0.265 0.531 0.753 0.909 0.970 0.988 EL(bootstrap) 0.100 0.259 0.529 0.754 0.907 0.970 0.987 ----------- ---------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- : Scenario 2. Estimated powers of the tests with bootstrap calibration when the sample sizes are unequal. The alternatives denote the change ($\delta$) in the $1$st component (\[change2\]).[]{data-label="pow_2a"} [cc]{}\ ![Estimated powers for a range of alternatives. The sample sizes are (a) equal to $n_1=30$ and $n_2=50$ and (b) equal to $n_1=50$ and $n_2=100$. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_2"}](powdiri_3.eps "fig:") & ![Estimated powers for a range of alternatives. The sample sizes are (a) equal to $n_1=30$ and $n_2=50$ and (b) equal to $n_1=50$ and $n_2=100$. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_2"}](powdiri_4.eps "fig:")\ $n_1 = 30 \ \ \& \ \ n_2 = 50$ & $n_1 = 50 \ \ \& \ \ n_2 = 100$\ \ ![Estimated powers for a range of alternatives. The sample sizes are (a) equal to $n_1=30$ and $n_2=50$ and (b) equal to $n_1=50$ and $n_2=100$. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_2"}](powlogi_3.eps "fig:") & ![Estimated powers for a range of alternatives. The sample sizes are (a) equal to $n_1=30$ and $n_2=50$ and (b) equal to $n_1=50$ and $n_2=100$. The solid horizontal line indicates the nominal level ($5\%$) and the two dashed lines are the lower and upper limits of the simulations error. The black and red lines refer to the Hotelling and James test respectively, the green and blue lines refer to the EEL and EL test respectively.[]{data-label="powers_2"}](powlogi_4.eps "fig:")\ $n_1 = 30 \ \ \& \ \ n_2 = 50$ & $n_1 = 50 \ \ \& \ \ n_2 = 100$\ When the sample sizes are relatively small, the powers of the quadratic tests is better than the power of the empirical methods regardless of the sign in the change. In fact Hotelling test performs better than James test and it is better when the change in the first component is positive and small. As for the larger sample sizes both quadratic tests and empirical likelihood methods perform very well with. But even then, Hotelling test still performs better than James test when the change in the alternative hypothesis decreases. Discussion and conclusions ========================== In most cases, the nominal level ($5\%$) of the Type I error was not attained by the procedures unless the sample sizes were large. In the small sample or unequal sample cases bootstrap calibration played an important role in correcting the test size. The cost of of this re-sampling procedure was time but only for the non-parametric likelihood methods. The quadratic tests require no numerical optimisation, only matrix calculations and with $299$ bootstrap re-samples, the calculation of the p-value requires less than a second. The exponential empirical likelihood requires a few seconds when calibrated using $299$ bootstrap samples, whereas the empirical likelihood requires a few minutes. We proposed the use of the $F$ distribution, with the degrees of freedom of the $F$ distribution as suggested by [@krishnamoorthy2004], for calibration of the empirical and the exponential empirical likelihood test statistics. Our results showed that it works better than the $\chi^2$. Another alternative is to use the corrected $\chi^2$ distribution [@james1954]. However, these alternative calibrations do not work when the sample sizes are small or very different. Bootstrap calibration on the other hand, performed very well in almost all cases. As for the power comparisons, the differences between the quadratic and the empirical likelihood tests were less than in the null case. However, since bootstrap was used to calibrate the test statistic when the null hypothesis was true, the same calibration had to be employed in testing the power under the different alternative hypotheses, when the null hypothesis was not true. The computational cost was high, since the quadratic tests require significantly less time than the non-parametric likelihood ones. This could be due to the fact that empirical likelihood required two optimisations, one to find the common mean and one to obtain the ratio test statistic value. Exponential empirical likelihood on the other hand requires one root search only. The conclusion that can be drawn is that the Hotelling test statistic or the James test statistic with bootstrap calibration is to be preferred when it comes to algorithmic simplicity and computational cost. Non-parametric likelihood methods perform equally well when bootstrap calibration is present but they require significantly more time than the James or Hotelling test statistics. Furthermore, we can see that the modified (in terms of the calibration) James test performs the same as the classical James test (using a corrected $\chi^2$ distribution). Time required by these two likelihoods for the simulations was counted in many weeks in clusters of computers, not just personal computers. This could be an evidence against the use of these non-parametric likelihoods. Based on our simulations we saw that when bootstrap calibration is applied, both methods tend to work almost equally well. If we had high computational power or an algorithm that would perform the empirical and exponential empirical likelihood testing procedures as quick as the James (or the Hotelling) test then we would say that the only reason to choose James test would be because of the convex hull limitation. The picture we got from the unequal sample sizes is similar to the one in the equal sample size cases. The conclusion drawn from this example is again that empirical likelihood methods are computationally expensive. Bootstrap calibration of the James test requires less than a second when $299$ bootstrap re-samples are implemented. Empirical likelihood methods on the other hand require more time which in the case of bootstrap is substantial, especially for the empirical likelihood. Even if we increase the number of bootstrap re-samples, James test will still require maybe a couple of seconds, whereas the empirical likelihood methods will probably require $10$ or more minutes. However, the availability of parallel computing in a desktop computer and a faster implementation of the non parametric likelihood tests, can reduce the time required to bootstrap calibrate the empirical likelihood. Even then, if one takes into account the fact that bootstrap calibration allowed for $299$ re-samples it becomes clear that the empirical likelihood is much more computationally expensive. The cost will still be high if data with many observations and or many dimensions are being examined. Aitchison, J. (1982). . , 44(2):139–177. Aitchison, J. (2003). . Reprinted by The Blackburn Press. Amaral, G. J. and Wood, A. T. (2010). Empirical likelihood methods for two-dimensional shape analysis. , 97(3):757–764. Baxter, M., Beardah, C., Cool, H., and Jackson, C. (2005). . , 37(2):183–196. DiCiccio, T. and Romano, J. (1990). Nonparametric confidence limits by resampling methods and least favorable families. , 58(1):59–76. Diciccio, T. J. and Romano, J. P. (1989). On adjustments based on the signed root of the empirical likelihood ratio statistic. , 76(3):447–456. Efron, B. (1981). Nonparametric standard errors and confidence intervals. , 9(2):139–158. Egozcue, J., Pawlowsky-Glahn, V., Mateu-Figueras, G., and Barcel[ó]{}-Vidal, C. (2003). . , 35(3):279–300. Emerson, S. (2009). . PhD thesis, Stanford university. Fisher, N. I., Hall, P., Jing, B.-Y., and Wood, A. T. (1996). Improved pivotal methods for constructing confidence regions with directional data. , 91(435):1062–1070. Fry, J., Fry, T., and McLaren, K. (2000). . , 32(8):953–959. Hall, P. and La Scala, B. (1990). Methodology and algorithms of empirical likelihood. , 58(2):109–127. James, G. (1954). Tests of linear hypotheses in univariate and multivariate analysis when the ratios of the population variances are unknown. , 41(1/2):19–43. Jing, B. (1995). . , 24(4):315–319. Jing, B. and Robinson, J. (1997). Two-sample nonparametric tilting method. , 39(1):25–34. Krishnamoorthy, K. and Yu, J. (2004). Modified nel and van der merwe test for the multivariate behrens-fisher problem. , 66(2):161–169. Lancaster, H. (1965). . , 72(1):4–12. Li, X., Chen, J., Wu, Y., and Tu, D. (2011). Constructing nonparametric likelihood confidence regions with high order precisions. , 21(4):1767–1783. Liu, Y., Zou, C., and Zhang, R. (2008). . , 78(5):548–556. Mardia, K., Kent, J., and Bibby, J. (1979). . London: Academic Press. Owen, A. (1988). Empirical likelihood ratio confidence intervals for a single functional. , 75(2):237–249. Owen, A. (1990). Empirical likelihood ratio confidence regions. , 18(1):90–120. Owen, A. (2001). . Boca Raton: Chapman & Hall/CRC. Preston, S. P. and Wood, A. T. (2010). Two-sample bootstrap hypothesis tests for three-dimensional labelled landmark data. , 37(4):568–587. Qin, J. and Lawless, J. (1994). Empirical likelihood and general estimating equations. , 22(1):300–325. Rodrigues, P. C. and Lima, A. T. (2009). Analysis of an european union election using principal component analysis. , 50(4):895–904. Scealy, J. L. and Welsh, A. H. (2014). Colours and cocktails: Compositional data analysis 2013 Lancaster lecture. , 56(2):145–169. Tsagris, M. T., Preston, S., and Wood, A. T. A. (2011). A data-based power transformation for compositional data. In [*Proceedings of the 4rth Compositional Data Analysis Workshop, Girona, Spain*]{}. Zhou, M. (2013). . R package version 0.9-9. Zhu, H., Zhou, H., Chen, J., Li, Y., Lieberman, J., and Styner, M. (2008). Adjusted exponentially tilted likelihood with applications to brain morphology. , 65(3):919–927.

ArXiv

--- abstract: 'Using measured instrumental response functions for data deconvolution is a known source of uncertainty. This problem is revisited here with Bayesian data analysis an Monte Carlo simulations. Noise correlation induced by the convolution operator is identified as a major source of uncertainty which has been neglected in previous treatments of this problem. Application to a luminescence lifetime measurement setup shows that existing approximate treatments are markedly defficient and that the correlation length of the noise is directly related to the lifetime to be estimated. Simple counteractive treatments are proposed to increase the accuracy of this procedure.' author: - | Pascal [<span style="font-variant:small-caps;">Pernot</span>]{}\ Laboratoire de Chimie Physique, (UMR 8000, associated to CNRS)\ Bât. 349, Université Paris-Sud, 91405 Orsay Cedex, France\ email: pascal.pernot@lcp.u-psud.fr title: | Bayesian analysis of signal deconvolution\ using measured instrument response functions --- Introduction ============ The deconvolution problem is a classical inverse problem, and has received a lot of attention in many scientific and engineering fields. The instrument response function (IRF), also called blurring function, is generally assumed to be accurately determined. It is however not uncommon that the IRF is measured with the same accuracy as the signal to be treated, due to instrumental or experimental design constraints. The impact of an uncertain IRF on the accuracy of the deconvolved signal has to be considered with care. This uncertainty propagation issue has been addressed in the past by Dose *et al.* [@Dose98]. We show here analytically and numerically that their approximate solution does not encompass important effects of noise correlation due to convolution. In this paper, we use Bayesian data analysis to derive an exact expression of the likelihood function in the case of gaussian additive noise. This solution is applied to the classical problem of lifetime estimation from luminescence data. Theory ====== The observed signal vector $\mathbf{s}$ (length $n$) is generally expressed as a linear reconvolution model $$\mathbf{s}=\mathbf{Hm}+\mathbf{e}_{s},\label{eq1}$$ where $\mathbf{m}$ is a vector of values of the model function at the measurement points, $\mathbf{H}$ is a $n\times n$ zero-padded lower triangular Toeplitz matrix built from the IRF $\mathbf{h}$ of length $n_{h}$ $$\mathbf{H}=\left(\begin{array}{ccccc} h_{1} & 0 & \cdots & 0 & 0\\ h_{2} & h_{1} & \cdots & 0 & 0\\ \vdots & \vdots & \ddots & 0 & 0\\ h_{n_{h}} & h_{n_{h}-1} & & h_{1} & 0\\ 0 & h_{n_{h}} & \cdots & h_{2} & h_{1}\end{array}\right)\label{mat}$$ and $\mathbf{e}_{s}$ is an additive noise with multinormal statistics and covariance matrix $\mathbf{R}_{s}$: $$\mathbf{e}_{s}\sim\mathcal{N}_{n}(0,\mathbf{R}_{s}).\label{eq:noise}$$ Note that we use bold lowercase symbols for vectors ($\mathbf{s}$) and bold capitals for matrices ($\mathbf{H}$). Using the symmetry property of convolution, Eq. \[eq1\] can also be written$$\mathbf{s}=\mathbf{Mh}+\mathbf{e}_{s}\label{eq2}$$ where $\mathbf{M}$ is a $n\times n_{h}$ lower triangular Toeplitz matrix built from the model vector $\mathbf{m}$ as shown above (Eq. \[mat\]). As the exact IRF is generally not known, a measured IRF is used instead to solve Eq.\[eq1\] or Eq.\[eq2\]. A multinormal additive noise model is also used for the IRF$$\mathbf{h}=\hat{\mathbf{h}}+\mathbf{e}_{h},$$ where $\hat{\mathbf{h}}$ is the unknown exact IRF and $\mathbf{e}_{h}\sim\mathcal{N}_{n_{h}}(0,\mathbf{R}_{h}).$ The problem is to reconstruct the model vector $\mathbf{m}$, knowing $\mathbf{s}$, $\mathbf{R}_{s}$, $\mathbf{h}$ and $\mathbf{R}_{h}$, and to evaluate the impact of the measurement uncertainties of $\mathbf{h}$ on $\mathbf{m}$. This is an inverse problem doubled with an uncertainty propagation problem. Bayesian data analysis is very well suited to handle this kind of problem [@Gelman95; @Sivia96; @Press89; @Hanson99a]. Bayesian data analysis ---------------------- The posterior probability density function (pdf) for $\mathbf{m}$ is obtained by Bayes’s formula$$p(\mathbf{m}|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})=\frac{p(\mathbf{m})}{p(\mathbf{s})}p(\mathbf{s}|\mathbf{m},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h}),$$ where $p(\mathbf{m})$ and $p(\mathbf{s})$ are the prior pdf’s for $\mathbf{m}$ and $\mathbf{s}$, and where $p(\mathbf{s}|\mathbf{m},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})$ is the likelihood function. Given our model, we do not know explicitely this latter function. Instead, we know the explicit expression for the likelihood when the exact IRF $\hat{\mathbf{h}}$ is considered (cf. Eq. \[eq:noise\])$$p(\mathbf{s}|\mathbf{m},\mathbf{R}_{s},\hat{\mathbf{h}})\sim\mathcal{N}_{n}(\mathbf{M}\hat{\mathbf{h}},\mathbf{R}_{s}).$$ Applying the marginalization rule and knowing the expression of the pdf for $\hat{\mathbf{h}}$$$p(\hat{\mathbf{h}}|\mathbf{h},\mathbf{R}_{h})\sim\mathcal{N}_{n}(\mathbf{h},\mathbf{R}_{h}),$$ we can write$$p(\mathbf{m}|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})=\frac{p(\mathbf{m})}{p(\mathbf{s})}\int d\hat{\mathbf{h}}\: p(\mathbf{s}|\mathbf{m},\mathbf{R}_{s},\hat{\mathbf{h}})p(\hat{\mathbf{h}}|\mathbf{h},\mathbf{R}_{h}).$$ Considering that in our model $p(\mathbf{s})$ is a normalization constant, and expliciting the pdf’s, one gets$$p(\mathbf{m}|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})\propto p(\mathbf{m})\int d\hat{\mathbf{h}}\:\exp\left(-\frac{1}{2}J\right),$$ where$$\begin{aligned} J & = & (\mathbf{s}-\mathbf{M}\hat{\mathbf{h}})^{T}\mathbf{R}_{s}^{-1}(\mathbf{s}-\mathbf{M}\hat{\mathbf{h}})\nonumber \\ & + & (\mathbf{h}-\hat{\mathbf{h}})^{T}\mathbf{R}_{h}^{-1}(\mathbf{h}-\hat{\mathbf{h}}).\end{aligned}$$ This quantity is rearranged in order to enable analytical integration$$\begin{aligned} J & = & (\hat{\mathbf{h}}-\mathbf{h}_{0})^{T}\mathbf{P}^{-1}(\hat{\mathbf{h}}-\mathbf{h}_{0})-\mathbf{h}_{0}^{T}\mathbf{P}^{-1}\mathbf{h}_{0}\nonumber \\ & + & \mathbf{s}^{T}\mathbf{R}_{s}^{-1}\mathbf{s}+\mathbf{h}^{T}\mathbf{R}_{h}^{-1}\mathbf{h},\end{aligned}$$ where$$\left\{ \begin{array}{rcl} \mathbf{P} & = & \left(\mathbf{M}^{T}\mathbf{R}_{s}^{-1}\mathbf{M}+\mathbf{R}_{h}^{-1}\right)^{-1}\\ \mathbf{h}_{0} & = & \mathbf{P}(\mathbf{M}^{T}\mathbf{R}_{s}^{-1}\mathbf{s}+\mathbf{R}_{h}^{-1}\mathbf{h})\\ & = & \mathbf{h}+\mathbf{PM}^{T}\mathbf{R}_{s}^{-1}(\mathbf{s}-\mathbf{Mh})\end{array}\right.$$ Integration over $\hat{\mathbf{h}}$ finally leads to$$p(\mathbf{m}|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})\propto\frac{p(\mathbf{m})}{|\mathbf{P}|^{1/2}}\exp\left(-\frac{1}{2}(\mathbf{s}-\mathbf{Mh})^{T}\mathbf{K}(\mathbf{s}-\mathbf{Mh})\right),$$ where$$\mathbf{K}=\mathbf{R}_{s}^{-1}-\mathbf{R}_{s}^{-1}\mathbf{MPM}^{T}\mathbf{R}_{s}^{-1}.\label{result}$$ This expression for the posterior pdf calls for a few comments : - convolution of the model vector with a noisy IRF leads to a ”noisy model” $\tilde{\mathbf{m}}=\mathbf{Mh}$, affected by correlated noise with covariance matrix $\mathbf{K}^{-1}$, the structure of which depends explicitely on the model vector itself (heteroscedastic correlated noise). This is in contrast with the result of Dose *et al. [@Dose98]*, who obtain an expression for an effective variance, and do not consider the covariance part. - the mode of the posterior pdf depends on the actual value of the measured IRF $\mathbf{h}$. A bias in the optimal values for the model vector is thus to be expected, as a different realization of the IRF would lead to a different solution. In any case, a consistent uncertainty analysis should ensure that the exact value lies within confidence intervals. Application =========== An application of interest is for instance the lifetime estimation of unstable chemical species from their luminescence decays. Model ----- A mono-exponential decay signal with lifetime $\tau$ is generated over a regular time grid ($n_{h}=n=100$). The model is $m_{i}=\textrm{exp}(-t_{i}/\tau)$. The IRF is a gaussian function centered at $t_{0}$, and of FWHM $w_{h}$ $$h_{i}=\textrm{exp}\left(-4\ln(2)(t_{i}-t_{0})^{2}/w_{h}^{2}\right).$$ In order to keep a single parameters, the model after convolution is rescaled to the maximal value of the signal, and we can set $p(\tau|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})\equiv p(\mathbf{m}|\mathbf{s},\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})$. Homoscedastic noise is considered for both signal and IRF, i.e. $\mathbf{R}_{s}=\sigma_{s}^{2}*\mathbf{I}_{n}$ , $\mathbf{R}_{h}=\sigma_{h}^{2}*\mathbf{I}_{n_{h}}$. Finally, a uniform prior distribution for $\tau$ is used ($p(\tau)=cte$). Comparison of models of the posterior pdf ----------------------------------------- The exact expression for the posterior pdf (eq.\[result\]) is compared to approximate expressions : - no correction for the noisy IRF ($\sigma_{h}=0$, in our model), which is the most commonly used method; - the diagonal approximation of eq.\[result\], which implements some level of variance correction, but fails to encompass the correlation in the model’s noise; - the ”effective variance” method [@Dose98; @Toussaint99]. **Variance.** Fig. \[fig\_dist1\] represents the posterior pdf $p(\tau|\mathbf{s},\sigma_{s},\mathbf{h},\sigma_{h})$ computed by Monte Carlo simulation, and by the different methods in the case of a same measurement accuracy for the signal and the IRF ($\sigma_{h}=\sigma_{s}=0.01$). The Monte Carlo method consists in repeated analysis of randomly noised signal and IRF to build histograms of the maximum a posteriori (MAP) lifetime values (modes of the posterior pdf). All curves have been shifted to a common mode, in order to facilitate comparison. The exact expression is fully coherent with the histogram resulting of the simulation, i.e. it takes properly the variance of the signal and the variance of the IRF into account. It can be seen on this figure that the approximate methods all perform quite similarly and fail to recover the full variance of the lifetime. The ”effective variance” method is seen to be numerically equivalent to the diagonal approximation of our method, and it performs only slightly better than the totally uncorrected method. Correlation in the noise of the convolved model can thus have a major impact on uncertainty quantification. **Bias.** All methods perform similarly with regard to the bias on lifetime estimation (Fig. \[fig\_bias\]). In this figure, we reported the estimation by the ”effective variance” method as function of the estimation by the exact method. The biases of both methods are highly correlated and practically identical. However, underestimation of the confidence intervals by the approximate method results in inconsistent estimations, i.e. it fails significantly more than the exact method to include the exact value inside the confidence interval, and confidence intervals for different realizations of the noise are frequently disjoint. In this regard, Eq. \[result\] performs much better. **Accuracy of the IRF.** For a given lifetime, when the IRF is measured with a better accuracy ($\sigma_{h}<\sigma_{s}$), the differences observed between the various methods tend to vanish (Fig.\[cap:Ratio-sig\]). For instance, if the IRF is ten times more accurate than the signal, the uncorrected method provides exact results over all the practical range of lifetimes. It is also observed that longer lifetimes are relatively more affected than shorter ones, which is a pure effect of noise correlation (see next section). Structure of the correlation matrix ----------------------------------- The convolution of the mono-exponential model by the IRF is a vector $\tilde{\mathbf{m}}$ which elements obbey the following reccurence$$\tilde{\mathbf{m}}_{i}=\textrm{exp}(-\frac{\Delta t}{\tau})\tilde{\mathbf{m}}_{i-1}+\mathbf{\hat{h}}_{i}+\mathbf{e}_{h,i}.$$ As soon as the IRF vanishes the correlation between consecutive points is$$<\tilde{\mathbf{m}}_{i},\tilde{\mathbf{m}}_{i-1}>=\textrm{exp}(-\frac{\Delta t}{\tau})$$ As there is supposedly no correlation in the signal noise, the covariance matrix $\mathbf{K}$ preserves this correlation scheme. An approximation of the correlation matrix can thus been expressed as$$\mathbf{C}=\left(\begin{array}{ccccc} 1 & \rho & \rho^{2} & \cdots & \rho^{n}\\ \rho & 1 & \rho & \cdots & \rho^{n-1}\\ \vdots & \vdots & \ddots & \cdots & \vdots\\ \vdots & \vdots & \vdots & 1 & \rho\\ \rho^{n} & \rho^{n-1} & \cdots & \rho & 1\end{array}\right),$$ where $\rho=\textrm{exp}(-\frac{\Delta t}{\tau})$. The noise correlation decays thus exponentially with de delay between points of the model, the decay rate being the inverse of the theoretical lifetime. Support length of the IRF ------------------------- When the IRF is recorded on the same support as the signal, most of the its elements are pure noise. We saw above that these points contribute significantly to the correlation of the noise in the convolved model. Limiting the support of the IRF ($n_{h}<n$), or zeroing it’s purely noisy elements might thus enable to improve the correlation matrix. If we observe the standard deviation for the convolved model (Fig. \[cap:Standard-deviation\]), we see that the truncation of the support of the IRF contributes significantly to reduce the uncertainty at larger times. As shown on Fig.\[fig\_dist1\](c), this enables some uncertainty reduction for the lifetime estimation, but the effect of the correlated noise is still quite marked. ### Identifiability The behaviour of the present model with regard to the limits of detection of lifetimes due to the IRF has been tested by reconstructing the posterior pdf from synthetic signals generated with very small lifetimes. The posterior pdf displays explicitely the non identifiability of lifetimes that are too small (Fig. \[cap:Identif\]). When $\tau$ decreases, the pdf becomes asymmetric, defining an upper limit for the lifetime, but no lower limit, except the one imposed by the prior. Conclusion ========== The use of measured instrumental response functions for data deconvolution is a source of uncertainty. We derived a new expression of the likelihood within a bayesian framework to explicitely incorporate this effect and display it’s importance. Convolution of a noisy IRF with a model curve produces a noisy model curve with correlated noise. This has been illustrated on a luminescence lifetime measurement setup, for which it was shown that existing approximate treatments were markedly defficient. It was also shown that, in this case, the correlation length of the noise was directly related to the lifetime to be estimated. Longer lifetimes are thus counterintuitively more affected by IRF’s uncertainty that shorter ones. Although the most efficient way to reduce this effect is clearly to improve the IRF’s measurement accuracy, we have shown that an qualitative improvement can very simply be obtained by zeroing those parts of the IRF consisting of pure noise. The method has been applied to an homoscedastic noise pattern, but extension to cases where the noise is dependent on signal intensity (e.g. photon counting methods) is straightforward, as long as the normal noise distribution approximation is valid. Similarly, cases where the IRF is locally fluctuating due to minor modifications of the experimental setup can be easily treated by a careful modelling of the variance/covariance matrix. We are studying extension of this method to Poisson uncertainties, and to the evaluation of the resolution limits of a fluorescence TCSCP apparatus [@Livesey87]. The ultimate goal is to obtain consistent uncertainty estimation for lifetimes recovered from fluorescence spectra analysis. An alternative treatment is to model the IRF by a function, which parameters pdf’s are estimated by a bayesian analysis$$p(\mathbf{m}|s,\mathbf{R}_{s},\mathbf{h},\mathbf{R}_{h})=\frac{p(\mathbf{m})}{p(\mathbf{s})}\int d\mathbf{p}_{h}\: p(\mathbf{s}|\mathbf{m},\mathbf{R}_{s},\mathbf{p}_{h})p(\mathbf{p}_{h}|\mathbf{h},\mathbf{R}_{h}).$$ [1]{} V. Dose, R. Fischer, and W. [von der Linden]{}. Deconvolution based on experimentally determined apparatus functions. In G. Erickson, editor, [*Maximum Entropy and Bayesian Methods*]{}, pages 147–152. Kluwer Academic, Dordrecht, 1998. A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin. . Chapman & Hall, London, 1995. D. S. Sivia. . Clarendon (Oxford Univ. Press), Oxford, 1996. S. J. Press. . Wiley, New York, 1989. K. M. Hanson. A framework for assessing uncertainties in simulation predictions. , 133:179–188, 1999. U. v Toussaint, R. Fischer, K. Krieger, and V. Dose. Depth profile determination with confidence intervals from rutherford backscattering data. , 1:11.1–11.13, 1999. A. K. Livesey and J. C. Brochon. Analysing the distribution of decay constants in pulse-fluorimetry using the maximum entropy method. , 52:693–706, 1987.

ArXiv

--- author: - 'Juan Sebastián Arias-Valero[^1]' - 'Octavio Alberto Agustín-Aquino[^2]' - 'Emilio Lluis-Puebla[^3]' bibliography: - 'MySubmissionBibTexDatabase.bib' --- > **Abstract:** We generalize Mazzola’s first-species counterpoint model to arbitrary *rings* and obtain some *counting results* that enrich the theory of admitted successors, pointing to a *structural approach*, beyond computations. The original *motivations* of the model, as well as all *technical passages*, are carefully reviewed so as to provide a complete exposition of its *essential* features. Based on this exposition, we suggest several improvements and stress the value of a possible *simplification*. > **Keywords:** counterpoint; rings; modules; combinatorics > *2010 Mathematics Subject Classification:* 00A65; 13C99; 05A99 Introduction ============ This article has several purposes. First, it intends to provide a rigorous, but strongly motivated, generalized exposition of a model for *first-species counterpoint*, based on ring and module theory. This kind of counterpoint is the simplest one and the didactic basis of *Renaissance counterpoint* theory, as taught by Johann J. Fux in [@Fux]. The model was introduced by Mazzola in [@Inicount], where the discussion confined to the case when the ground ring is $\mathbb{Z}_{12}$; a ring that can be used to model the algebraic behavior of the twelve intervals between tones in the *chromatic scale* of Western musical tradition. Then, the model was re-exposed with some additional computational results by Hichert in [@MazzolaTopos Part VII]. Further generalizations to the case when the rings are of the form $\mathbb{Z}_{n}$ were considered in [@OAAAthesis; @Junod], and then included in a collaborative compendium of mathematical counterpoint theory and its computational aspects [@Octavio]. The motivation for such a generalization to $\mathbb{Z}_n$ was the existence of *microtonal scales* with more than twelve tones, which have been used for making real music [@octaviotod]. Second, at the same time, we generalize the model to noncommutative rings. However, though this generalization can sound difficult to the reader, the essential feature to bear in mind, while reading this paper, is the fact that two given elements of the ground ring need not commute. On the contrary, dropping the commutativity hypothesis offers the possibility of defining counterpoint notions on noncommutative rings of generalized intervals, which can be worth exploring in musical terms; see Example \[noncomex\] and Section \[handnc\]. From a more theoretic perspective, this generalization has the advantage that in some proofs, the essential arguments, as well as the necessity of the commutative hypothesis for the ground ring, are identified. For example, in Section \[commufor\] we obtain some counting formulas for the cardinality of deformed consonance sets, and a maximization criterion of these cardinals, thanks to the commutativity hypothesis. Anyway, the paper can be read by assuming that all rings are commutative, since there is no specialized result from noncommutative ring theory used here. A basic course on rings and modules is the only prerequisite. Third, the previous literature on this subject has a strong computational flavour. In this paper, we emphasize the necessity of a structural and conceptual approach to the model. The main goal would be to provide a theoretic proof of the Counterpoint Theorem [@Octavio Cf. Theorem 3.1]. This theorem gives the exact number of admitted successors of a counterpoint interval and was established by using Hichert’s algorithm. The aforementioned counting results could simplify this algorithm. Also, they allow us to compute the admitted successors, *by hand*, in the case of Renaissance counterpoint. Thus, we point to understand *why* the number of admissible successors of a counterpoint interval is that obtained by Hichert’s algorithm. We believe that a profound comprehension of Renaissance counterpoint and its generalizations pass through a structural understanding of their models, or rather, of counterpoint itself. Regarding other mathematical models of counterpoint, Mazzola’s school has an important counterpart in D. Tymoczko. Tymoczko’s model [@Tympolem Appendix] is based on orbifolds and is oriented towards voice leading. Moreover, it works by means of a geometric reading of the usual counterpoint rules, in contrast to the predictive character of Mazzola’s model, which follows the principle that these rules obey profound relations based on musical symmetries. There has been a polemic around Mazzola’s and Tymoczko’s models. Tymoczko’s initial critique can be found in [@Tympolem] and a response by Mazzola and the second author can be found in [@Mazzpolem]. We organize this paper as follows. We start with the basic definitions: symmetries (natural operations in music) in Section \[sec1\], hierarchy of dichotomies and polarities (consonance/dissonance partitions and their symmetries) in Section \[sec2\], and dual numbers (modelling two voices in counterpoint) in Section \[sec3\]. In particular, in Section \[sec3.4\], we prove that several dichotomies are strong, in contrast to previous approaches, where the authors advised the reader to check the strong dichotomy property with the computer. Then, in Section \[locsec\], we discuss global and local symmetries of consonances/dissonances, which, according to Mazzola, are the responsible for tension/resolution forces in counterpoint. This allows us to provide a formal definition of first-species counterpoint in Section \[sec5.3\]. In Section \[secsimp\], we simplify this definition, by reducing it to the case when the cantus firmus is the note c and use a small group of symmetries. Based on this simplification, we prove that, in any notion of counterpoint, it is always possible to find successors of a given consonant interval. This result is the *Little theorem of counterpoint*. A careful review of a crucial lemma (Lemma \[basiccount\]) in the proof of this theorem leads us to the *counting formulas* and the *maximization criterion* in Section \[seccountfor\] for deformed consonance sets. This section is the main contribution of this article. With these tools, in Section \[hand\], we obtain the admitted successors of consonant intervals in Renaissance counterpoint, without the computer. In Section \[handnc\], we explore a noncommutative notion of counterpoint. Finally, in Section \[final\], we compare the strengths and the weaknesses of Mazzola’s model, and subsequently suggest some improvements and, above all, a *simplification* of the model. The appendix includes the proof of a variation of the *rearrangement inequality*, which is crucial for the maximization criterion, and a brief discussion on the *categorical groupoid of intervals of a ring*, which is useful for deciding when a partition of a ring into two equipotent sets has an affine symmetry. Throughout this paper, $R$ denotes an arbitrary ring with unity, not necessarily commutative. When needed, commutativity hypothesis on $R$ is explicitly stated. Similarly, modules considered in this paper are right $R$-modules, unless we indicate otherwise. Ring symmetries {#sec1} =============== *Affine homomorphisms are the natural correspondences that occur in music*. They are the natural formalization that encompasses musical transformations like transposition and inversion in the ring $\mathbb{Z}_{12}$ of tones.[^4] On the other hand, they come from *affine geometry*, understood as the study of the *invariant* properties under the action of the group of affine automorphisms of the plane $\mathbb{R}^2$ regarded as a real vector space. However, in music, we are interested in the *variations* of musical material (scales, intervals) with symmetries. The basic variation in counterpoint is the *prohibition of parallelisms*, whereas the fundamental invariant of affine geometry is *parallelism* between lines. Let $R$ be a ring. Given two $R$-modules $M$ and $N$, an *affine homomorphism* from $M$ to $N$ is the composite $e^a\circ f$, where $f:M\longrightarrow N$ is an $R$-homomorphism, $a\in N$, and $e^a:N\longrightarrow N:x\mapsto x+a$ is the *translation* associated with $a$. We write a typical affine homomorphism $e^a\circ f$ as $e^af$, for short. Observe that the composition rule for affine homomorphisms is $(e^af)\circ (e^bg)=e^{f(b)+a}fg$. The monoid (with respect to the composition) $\operatorname{End}_R(M)$ of $R$-endomorphisms of an $R$-module $M$ is a ring with the usual sum of homomorphisms. However, the monoid of affine endomorphisms of $M$ is not a ring since the distributivity fails, except when $M$ is the trivial module. We denote by $\operatorname{Aut}_R(M)$ the group of $R$-automorphisms of an $R$-module $M$. An affine homomorphism $e^af:M\longrightarrow N$ is an isomorphism if and only if $f$ is an $R$-isomorphism.[^5] Thus, affine automorphisms of $R$ are of the form $e^af$, where $f\in \operatorname{Aut}_R(R)$ and $a\in R$. On the other hand, recall that there is a ring isomorphism $$\operatorname{End}_R(R)\cong R,$$ which identifies an element $r\in R$ with the $R$-endomorphism obtained by left multiplication with $r$. This isomorphism restricts to a group isomorphism $$\operatorname{Aut}_R(R)\cong R^*,$$ which establishes a bijective correspondence between $R$-automorphisms of $R$ and invertible elements of $R$. From this discussion we conclude that an affine automorphism of $R$ is the function associated with the linear polynomial $bx+a$, where $b\in R^*$ and $a\in R$. In what follows, we will call affine automorphisms of $R$ **symmetries** of $R$ and denote the group of symmetries of a ring $R$ by $\operatorname{\mathbf{Sym}}(R)$. We use the notation $e^ab$ for the symmetry associated with $bx+a$. Two symmetries $e^ab$ and $e^{a'}b'$ are equal if and only if $a=a'$ and $b=b'$. Observe that $e^01$ is the identity symmetry and that the composition of symmetries is given by the formula $$e^ab\circ e^{a'}b'=e^{ba'+a}bb'.\label{Eq:CompositionOfSymmetries}$$ The inverse of a symmetry $e^ab$ is $e^{-b^{-1}a}b^{-1}$. Basic motivations and definitions {#sec2} ================================= In Renaissance counterpoint we divide the ring $\mathbb{Z}_{12}$ of **intervals** into two disjoint subsets, namely the set $K$ of *consonances* and the set $D$ of *dissonances*. The consonances are unison, minor third, major third, perfect fifth, minor sixth, and major sixth. The dissonances are minor second, major second, perfect fourth, tritone, minor seventh, and major seventh. The corresponding mathematical definitions are $$K=\{0,3,4,7,8,9\}\text{ and }D=\{1,2,5,6,10,11\}.$$ A composition of *first-species counterpoint* consists of two voices, *cantus firmus* and *discantus*, whose notes have the same length and where each interval between the voices is a consonance. See [@Fux pp. 19-29] for details. The germ of Mazzola’s model is the observation that there is a *unique* symmetry $p$ of the ring $\mathbb{Z}_{12}$, namely $e^25$, that sends consonances to dissonances. In fact, $$e^25(0)=2,\ e^25(3)=5,\ e^25(4)=10,\ e^25(7)=1, \ e^25(8)=6, \ e^25(9)=11.$$ We postpone the proof of the uniqueness to Section \[sec3.4\]. Moreover, $e^25\circ e^25=e^01=id$, so $e^25$ is its own inverse. Two important structural implications of this uniqueness property, for mathematical counterpoint theory, can be appreciated in Theorem \[indxdich\] and Lemma \[commusimp\]. Up to now, we have the intervals, consonances, and dissonances of counterpoint, but we need to model the voices in a composition. We achieve this by means of the dual numbers ring[^6] $\mathbb{Z}_{12}[\epsilon]$, which is the ring of linear polynomials $c+d\epsilon$, with $c,d\in \mathbb{Z}_{12}$, in an indeterminate $\epsilon$ satisfying the relation $\epsilon^2=0$. The element $c+d\epsilon$, called a **contrapuntal interval**, represents a pitch class $c$, of the cantus firmus, together with the interval $d$ between $c$ and the pitch class $c+d$ from the superior discantus. As already said, in a piece of Renaissance counterpoint, it is mandatory that $d$ be a consonance. For instance, the dual number $2+7\epsilon$ in $\mathbb{Z}_{12}[\epsilon]$ comes from the following musical example.  Thus, we have a partition of the ring $\mathbb{Z}_{12}[\epsilon]$ into **contrapuntal consonances and dissonances**, namely $\{K[\epsilon],D[\epsilon]\}$, where $X[\epsilon]$ consists of all $c+d\epsilon$ with $d\in X$ for $X=K,D$. We would want properties for this partition analogous to those of $\{K,D\}$. Certainly, the symmetry $e^{2\epsilon} 5$ of $\mathbb{Z}_{12}[\epsilon]$ is a quite natural[^7] extension of $e^2 5$ sending $K[\epsilon]$ to $D[\epsilon]$ since $e^{2\epsilon}5$ is simple and acts on the interval part $d$ of a dual number $c+d\epsilon$ just as $e^25$. But in this case *it is not the unique* sending $K[\epsilon]$ to $D[\epsilon]$. For example, $e^{2\epsilon+1} 5$ also does. We thus have two kinds of partitions $\{K,D\}$ of a ring $R$ that are important for counterpoint: those for which there is a symmetry $p$ of $R$ that sends $K$ to $D$ (dual numbers case) and those for which such a symmetry exists and is unique (intervals case). We want to compose a piece of first-species counterpoint, which is a suitable finite sequence of dual numbers $\xi_1,\dots ,\xi_n$ in $K[\epsilon]$ with cantus firmus and discantus in a *mode*.[^8] However, not all sequences are valid since counterpoint has some *rules*. These rules establish when $\xi_{i+1}$ is an *admitted successor* of $\xi_i$ for $i=0,\dots ,n-1$. In the model to be exposed in the following sections, we aim to predict these rules and formalize the idea of an admitted successor so that, given a cantus firmus, we can compose a discantus with the aid of the rules and our own musicality. It is important to stress that the model is not an automatic or mechanic process for producing compositions, although the symmetries and the rules can provide a framework for such a mechanization [@Octavio; @Junod]. Dichotomies {#sec3.1} ----------- We start to develop our theory by formalizing the common properties of the rings of intervals $\mathbb{Z}_{12}$ and of dual numbers $\mathbb{Z}_{12}[\epsilon]$ that are relevant for counterpoint. Let $R$ be a ring. - A partition $\{K,D\}$ of $R$ is a **dichotomy** of $R$ if $|K|=|D|$. Note that a finite ring $R$ has dichotomies if and only if its cardinality is even. - A **self-complementary dichotomy**[^9] of $R$ is a triple $(K,D,p)$, where $\{K,D\}$ is a partition of $R$, $p$ is a symmetry of $R$, and $p(K)=D$. Note that necessarily $\{K,D\}$ is a dichotomy of $R$. The triples $(K,D,e^25)$ and $(K[\epsilon],D[\epsilon],e^{2\epsilon}5)$ from Renaissance counterpoint are self-complementary dichotomies. The dichotomy $\{\{0,1,3,6,8,11\},\{2,4,5,7,9,10\}\}$ of $\mathbb{Z}_{12}$ is not part of a self-complementary dichotomy of $\mathbb{Z}_{12}$ as proved in Appendix. We would like to construct self-complementary dichotomies other than the basic one of Renaissance counterpoint, so as to create new notions of counterpoint. Given an arbitrary finite ring $R$ of even cardinality, we could start with a symmetry $p$ of $R$ and aim to construct a self-complementary dichotomy $(K,D,p)$. We pick some $k_1\in R$ to be in $K$, and define $d_1:=p(k_1)$, with $d_1$ to be in $D$. Then, if $|R|>2$, we take some $k_2\in R\setminus \{k_1,d_1\}$ to be in $K$ and define $d_2:=p(k_2)$ to be in $D\setminus \{d_1\}$, and so on. Note that this process produces a dichotomy $\{K,D\}$ provided *$p$ has no fixed points* and is finite since $R$ is. We say that a symmetry $p$ without fixed points is a **derangement**. The following proposition contains some concluding properties of self-complementary dichotomies. \[propsymdich1\] Each self-complementary dichotomy $(K,D,p)$ of $R$ has the following properties: 1. The identity $p(D)=K$ holds, so $(D,K,p)$ is a self-complementary dichotomy. 2. The symmetry $p$ is a derangement. 1\. Since $p$ is a bijection, the properties of the inverse image imply that $p(D)=p(R\setminus K)=R\setminus p(K)=R\setminus D=K$. 2. Let $x\in R$. If $x\in K$, then $x\neq p(x)\in D$. If $x\in D$, then, by 1, $x\neq p(x)\in K$. Thus, *self-complementary dichotomies are strongly related to derangements*. Strong dichotomies and polarities {#sec3.2} --------------------------------- We formalize the additional uniqueness property of the Renaissance dichotomy $(K,D,e^25)$ with the following definitions. Let $R$ be a ring. A self-complementary dichotomy $(K,D,p)$ of $R$ is a **strong dichotomy** of $R$ if $p$ is the unique symmetry of $R$ such that $p(K)=D$. In such a case, we also say that $p$ is a **polarity**. As we will show in the next sections, once we have a strong dichotomy of a ring (made up of generalized intervals), we have an induced self-complementary dichotomy of the respective dual numbers ring (made up from contrapuntal intervals between two voices), and an associated theory of admitted successors. Thus, strong dichotomies lead to new *counterpoint worlds* for composing non-traditional counterpoint. We need to determine whether a given self-complementary dichotomy $(K,D,p)$ is strong. We approach this problem by determining the number of symmetries sending $K$ to $D$. Given another symmetry $q$ with $q(K)=D$, note that $$q^{-1}\circ p(K)=q^{-1}(D)=K$$ and hence $q^{-1}\circ p$ is in the stabilizer $\theta(K)$, where $$\theta(K):=\{g \in \operatorname{\mathbf{Sym}}(R)\ |\ g(K)=K\}.$$ This suggests the right action of $\theta(K)$, by composition, on the set $\operatorname{\mathbf{Sym}}(K,D)$ of symmetries of $R$ sending $K$ to $D$, as shown in the following diagram. $$\begin{array}{cccc} \circ :&\operatorname{\mathbf{Sym}}(K,D) \times \theta(K)& \longrightarrow& \operatorname{\mathbf{Sym}}(K,D)\\ &(q,g) &\longmapsto & q\circ g \end{array}$$ As we have observed, every $q\in \operatorname{\mathbf{Sym}}(K,D)$ is in the same orbit of $p$. Also, the stabilizer of $p$ under this action is the identity since $p\circ g=p$ implies $g=id_R$. Hence, the following proposition. Let $(K,D,p)$ be a self-complementary dichotomy of a ring $R$. There is a bijective correspondence between $\operatorname{\mathbf{Sym}}(K,D)$ and $\theta(K)$. The bijection sends an element $g\in \theta(K)$ to $ p\circ g$. The function $\theta(K)\longrightarrow \operatorname{\mathbf{Sym}}(K,D): g\mapsto p\circ g$ is surjective since the action is transitive and is injective since the stabilizer of $p$ is trivial. Thus, we have reduced the study of $\operatorname{\mathbf{Sym}}(K,D)$ to that of the stabilizer group $\theta(K)$ of $K$. Besides, we have determined when a self-complementary dichotomy is strong, as established in the following corollary. We say that a partition $\{K,D\}$ of $R$ is **rigid** if $\theta(K)$ is the trivial group. \[rig\] A self-complementary dichotomy $(K,D,p)$ is a strong dichotomy if and only if $\{K,D\}$ is rigid. Now, our objective is to prove a pair of lemmas that help us to decide whether a given dichotomy is rigid, by discarding non-identity symmetries that could be in $\theta(K)$. These lemmas could be a first approximation to a more conceptual argument.[^10] Consider the left action $\cdot$, by multiplication, of the group $R^*$ on $R$. \[userigid\] Let $\{K,D\}$ be a dichotomy of a ring $R$. Suppose that $a\in R$ and that $C$ is either $K$ or $D$. If there is $r\in R$ such that 1. the orbit $R^*r$ of $r$ under the action $\cdot$ is contained in $C$ and 2. $r+a\notin C$, then $e^ab\notin \theta(K)$ for each $b\in R^*$. If $r+a\notin C$ for $r$ as above, then for each $b\in R^*$, $b(b^{-1}\cdot r)+a\notin C$ with $b^{-1}\cdot r\in R^*r\subseteq C$, and hence $e^ab\notin \theta(C)=\theta(K)$. In particular, if $r=0$, we obtain the following result. \[disc0\] Suppose that $C$ is either $K$ or $D$ and that $0\in C$. Each symmetry of the form $e^ab$ with $a\in R\setminus C$ is not in $\theta(K)$. In particular, $e^ab\notin \theta(K)$. Before using these lemmas to check that some dichotomies are strong, we first study a possible way to construct self-complementary dichotomies that are good candidates to strong ones. Constructing strong dichotomies: quasipolarities {#sec3.3} ------------------------------------------------ A symmetry $p$ of a ring $R$ is **involutive** if $p\circ p=id$. If $p=e^ab$, by Equation , this condition is equivalent to $e^{ba+a} b^2=e^0 1$, that is, to $ba+a=0$ and $b^2=1$. In the next proposition we prove that all polarities are involutive, so if we want a polarity to open up a counterpoint world, we could start with an involutive derangement, then define a suitable self-complementary dichotomy following the strategy in Section \[sec3.1\], and finally check whether the uniqueness property holds. \[involut\] If a self-complementary dichotomy $(K,D,p)$ is strong, then $p$ is involutive. If $(K,D,p)$ is a self-complementary dichotomy and $p$ has the uniqueness property, then $p\circ p(K)=p(D)=K$ by 1 in Proposition \[propsymdich1\], and hence $p\circ p=id$ by Corollary \[rig\]. The converse of this proposition does not hold since, in the Renaissance dichotomy $(K[\epsilon],D[\epsilon],e^{2\epsilon}5)$, $e^{2\epsilon}5$ is involutive but the dichotomy is not strong, as already observed. We call involutive derangements **quasipolarities**. A **quasipolarization** of $R$ is a self-complementary dichotomy $(K,D,p)$ of $R$ where $p$ is involutive. Therefore, in view of Proposition \[propsymdich1\], $(K,D,p)$ is a quasipolarization if and only if $p$ is a quasipolarity. Note that not all self-complementary dichotomies are quasipolarizations. For example, the self-complementary dichotomy $(\{0,2,4,6,8,10\},\{1,3,5,7,9,11\},e^1)$ of $\mathbb{Z}_{12}$ is not a quasipolarization since $e^1\circ e^1=e^2\neq id$. Thus, the construction of quasipolarizations is a good beginning if we want a strong dichotomy, though quasipolarizations need not be strong dichotomies. Examples of strong dichotomies {#sec3.4} ------------------------------ In this section, we use Lemmas \[userigid\] and \[disc0\] to open up a series of counterpoint worlds. First, consider the case when $R$ is the commutative ring $\mathbb{Z}_{12}$ and different examples of quasipolarizations that we will check to be strong dichotomies. The set of orbits of the action of the group $\{1,5,7,11\}$ of invertible elements of $\mathbb{Z}_{12}$ on $\mathbb{Z}_{12}$ is $$\{\{0\}, \textcolor{green}{\{1,5,7,11\}}, \textcolor{blue}{\{2,10\}}, \textcolor{magenta}{\{3,9\}}, \textcolor{violet}{\{4,8\}}, \textcolor{red}{\{6\}}\}.$$ In the following example we prove the claim that the Renaissance quasipolarization is a strong dichotomy. Consider the dichotomy $(K,D,e^25)$, where $K=\{0,\textcolor{magenta}{3},\textcolor{violet}{4},\textcolor{green}{7},\textcolor{violet}{8},\textcolor{magenta}{9}\}$ and $D=\{\textcolor{green}{1},\textcolor{blue}{2},\textcolor{green}{5},\textcolor{red}{6},\textcolor{blue}{10},\textcolor{green}{11}\}$. Let us prove that the dichotomy is strong. According to Corollary \[rig\], it is enough to show that our dichotomy is rigid, or equivalently, that $\theta(K)=\{e^01\}$. In fact, by Lemma \[disc0\], $e^ab\notin \theta(K)$ for each symmetry $e^ab$ with $a\in D$. The cases when $a\in K$ remain. If $a=0$, then since $5\times 7=11\notin K$, $7\times 7=1\notin K$, and $11\times 7=5\notin K$, we conclude that the symmetries of the form $e^0b$ with $b=5,7,9$ are not in $\theta(K)$. Finally, since $$\textcolor{magenta}{3}+8=\textcolor{violet}{8}+3=11\notin{K},\ \textcolor{violet}{4}+9=\textcolor{magenta}{9}+4=1\notin{K}\text{, and } \textcolor{magenta}{3}+7\notin{K},$$ Lemma \[userigid\] implies that all symmetries $e^ab$ with $a\in \{3,4,7,8,9\}$ are not in $\theta(K)$. We have exhausted all possibilities, except the identity, and hence $\theta(K)$ is trivial.[$\Diamond$]{} Consider the dichotomy $(K,D,e^5 11)$, where $K=\{0,\textcolor{blue}{2},\textcolor{violet}{4},\textcolor{red}{6},\textcolor{magenta}{9},\textcolor{blue}{10}\}$ and $D=\{\textcolor{green}{1},\textcolor{magenta}{3},\textcolor{green}{5},\textcolor{green}{7},\textcolor{violet}{8},\textcolor{green}{11}\}$. Let us show that $e^5 11$ is a polarity by using the same strategy of the preceding example. As before, $e^ab\notin \theta(K)$ for each symmetry $e^ab$ with $a\in D$, and we need to discard the cases when $a\in K$. If $a=0$, then since $5\times 4=8\notin K$, $7\times 9=3\notin K$, and $11\times 4=8\notin K$, the symmetries of the form $e^0b$ with $b=5,7,11$ are not in $\theta(K)$. Also, since $$\textcolor{red}{6}+2=\textcolor{blue}{2}+6=8\notin{K},\ \textcolor{green}{5}+4=9\notin{D},\ \textcolor{blue}{2}+9=11\notin{K}\text{, and } \textcolor{blue}{10}+10=8\notin{K},$$ Lemma \[userigid\] implies that all symmetries $e^ab$ with $a\in \{2,4,6,9,10\}$ are not in $\theta(K)$. Hence, $\theta(K)$ is trivial. [$\Diamond$]{} The next computation establishes the existence of at least a strong dichotomy in $\mathbb{Z}_{2k}$ for each $k\geq 3$. These dichotomies are the beginning of a lot of microtonal counterpoint worlds, which were first studied in [@OAAAthesis]. Recall that all invertible elements of $\mathbb{Z}_{2k}$ are odd since they are coprime with $2k$. Let[^11] $k\geq 3$. In $\mathbb{Z}_{2k}$, consider the dichotomy $(K,D,e^{-1}(-1))$, where $$K=\{0,1,3,..., 2k-5,2k-3\} \text{ and } D=\{-1,2k-2,2k-4,...,4,2\}.$$ We already know (Lemma \[disc0\]) that $e^ab\notin \theta(K)$ for each symmetry $e^ab$ with $a\in D$. If $a=0$, then $b(-1)=-b\neq -1$ with $-b$ odd whenever $b\in\mathbb{Z}_{2k}^*\setminus \{1\}$ and hence $-b\notin D$, so the symmetries of the form $e^0b$ with $b\in\mathbb{Z}_{2k}^*\setminus \{1\}$ are not in $\theta(D)=\theta(K)$. Also, since all orbits of even numbers, except the orbit $\{0\}$ of $0$, under the action of $\mathbb{Z}_{2k}^*$ on $\mathbb{Z}_{2k}$ are contained in $D$, the equations $$2+1=3\notin{D}, ...,2+(2k-5)=2k-3\notin{D}\text{, and } 4+(2k-3)=1\notin{D},$$ imply (Lemma \[userigid\]) that all symmetries $e^ab$ with $a\in \{1,3,..., 2k-5,2k-3\}$ are not in $\theta(K)$. Hence, $\theta(K)$ is trivial. [$\Diamond$]{} Now, a noncommutative example. \[noncomex\] Let $R$ be the noncommutative ring of all upper triangular matrices $2\times 2$ with entries in $\mathbb{Z}_2$. The symmetry $e^I$, where $I$ is the identity matrix, is a quasipolarity, so we can construct self-complementary dichotomies with the procedure of Section \[sec3.1\]. A possible choice is the dichotomy with $K=\{\mathbf{0},A_1,A_2,A_3\}$, where $$A_1= \begin{bmatrix} 1 & 0 \\ 0 & 0 \end{bmatrix},\ A_2=\begin{bmatrix} 1 & 1 \\ 0 & 0 \end{bmatrix},\text{ and } A_3= \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix}.$$ In this case, $R^*=\{I,A_3\}$ and $A_1$ and $A_2$ are invariant under the left action of $R^*$. Let us show that $K$ is rigid. As usual, $e^AB\notin \theta(K)$ for each symmetry $e^AB$ with $A\in D$. The cases when $A\in K$ remain. If $A=0$, then since $A_3 A_3=I\notin K$, the symmetry $e^{\mathbf{0}}A_3$ is not in $\theta(K)$. Also, $$A_2+A_1=A_1+A_2\notin{K}\text{, and } A_1+A_3\notin{K},$$ so, by Lemma \[userigid\], all symmetries $e^AB$ with $A\in \{A_1,A_2,A_3\}$ are not in $\theta(K)$. Thus, $\theta(K)$ is trivial, and $(K,R\setminus K,e^I)$ is a strong dichotomy. The reader can obtain other examples of strong dichotomies from quasipolarities of $R$, whenever the chosen $K$ is not invariant under the left action of $A_3$, by using a similar argument. [$\Diamond$]{} Finally, note that there is no strong dichotomy of $\mathbb{Z}_4$ to define a counterpoint world with a four-tone scale (which can be thought as a diminished seventh chord). In fact, the possible dichotomies $\{\{0,1\},\{2,3\}\}$, $\{\{0,2\},\{1,3\}\}$, and $\{\{0,3\},\{1,2\}\}$ are invariant under $e^1(-1)$, $e^2$, and $e^{-1}(-1)$, respectively, so they are not rigid or strong (Corollary \[rig\]). The dual numbers ring {#sec3} ===================== Given an arbitrary ring $R$, we can construct the polynomial ring $R[x]$ and the two-sided ideal $\left\langle x^2\right\rangle$ consisting of all (left or right) multiples of $x^2$. Thus, we have the quotient ring[^12] $R[x]/\left\langle x^2\right\rangle$, which we call *the ring of dual numbers* associated with $R$. If we denote by $\epsilon$ the class of $x$, then each element of the ring of dual numbers can uniquely be written[^13] as $c+d\epsilon$, and hence it makes sense to denote the ring of dual numbers by $R[\epsilon]$. The ring $R[\epsilon]$ is noncommutative if $R$ is. An element $c+d\epsilon \in R[\epsilon]$ is invertible if and only if $c$ is invertible in $R$. In fact, if $(c'+d'\epsilon)(c+d\epsilon)=1$ and $(c+d\epsilon)(c'+d'\epsilon)=1$, then $c'c+(c'd+d'c)\epsilon=1$ and $cc'+(cd'+dc')\epsilon=1$, and hence $c'=c^{-1}$ ($c$ is invertible with two-sided inverse $c'$). Reciprocally, if $c$ is invertible, define $c'=c^{-1}$ and $d'=-c'dc'$. Thus, $c'+d'\epsilon$ is a two-sided inverse for $c+d\epsilon$. Recall, from Section \[sec2\], that an element $c+d\epsilon$ of $\mathbb{Z}_{12}[\epsilon]$ represents a generalized pitch class $c$, from the cantus firmus, together with the interval $d$ formed between $c$ and a pitch class from a superior discantus. However, we must ask why we choose the dual numbers ring structure on the elements $c+d\epsilon$ and not another. Mazzola requires [@Inicount p. 558] the characterizing relation $\epsilon^2=0$ because it entails two voices and it is consistent with two competing interpretations for three voices (cantus firmus, discantus, and countertenor), namely the *English theory*, which regards intervals only in relation to the cantus firmus, and the *continental theory*, which considers all intervallic relationships between the voices.[^14] We have the following proposition regarding dichotomies of the dual numbers ring $R[\epsilon]$ induced by dichotomies of a ring $R$. Given subsets $X,Y\subseteq R$, we define $$X+ Y \epsilon =\{c+d\epsilon\in R[\epsilon]\ |\ c\in X\text{ and } d\in Y\}.$$ Also, we denote by $X[\epsilon]$ the set $R+X\epsilon$. \[dualexts\] Let $\{K,D\}$ be a dichotomy of a ring $R$. 1. The pair $\{K[\epsilon],D[\epsilon]\}$ is a dichotomy of $R[\epsilon]$. 2. If $(K,D,e^ab)$ is a self-complementary dichotomy of $R$, then $(K[\epsilon],D[\epsilon],e^{a\epsilon}b)$ is a self-complementary dichotomy of $R[\epsilon]$. 3. If $(K,D,e^ab)$ is a quasipolarization of $R$, then $(K[\epsilon],D[\epsilon],e^{a\epsilon}b)$ is a quasipolarization of $R[\epsilon]$. As previously commented, strong dichotomies need not induce strong dichotomies on dual numbers. 1\. Exercise. 2. This follows from 1 and the equalities $$\begin{aligned} e^{a\epsilon}b(K[\epsilon])&=\{e^{a\epsilon}b(r+k\epsilon)\ |\ r\in R\text{ and }k\in K\}\\ &=\{br+(bk+a)\epsilon\ |\ r\in R\text{ and }k\in K\}\\ &=bR+(e^ab(K))\epsilon\\ &=R+D\epsilon\\ &=D[\epsilon].\end{aligned}$$ As to the fourth equality above, note that $bR=R$ since $b$ is invertible. 3. This follows from 2 and the equation $$e^{a\epsilon}b\circ e^{a\epsilon}b=e^{(ba+a)\epsilon}b^2=e^0 1,$$ which is a consequence of the identity $e^a b\circ e^a b=e^0 1$, equivalent to $ba+a=0$ and $b^2=1$. Global and local symmetries {#locsec} =========================== Once we have the dual numbers, which model the two voices of first-species counterpoint, the next step is the theory of admitted successors of a given contrapuntal consonance $\xi$, with $\xi=c+k\epsilon$ and $k$ consonance of a given strong dichotomy $(K,D,e^ab)$. The basic principle of the model is the *alternation between consonances and dissonances* as generator of the movement from $\xi$ to its successor. In fact, we introduce a dissonance character within consonances and apply the alternation principle as follows. We deform the induced dichotomy $\{K[\epsilon],D[\epsilon]\}$ of the dual numbers into a new one $\{g(K[\epsilon]),g(D[\epsilon])\}$, by applying a symmetry $g$ so that $\xi$ is a deformed dissonance, that is, $\xi\in g(D[\epsilon])$, which must move to a deformed consonance $\eta\in g(K[\epsilon])$ by alternation. Moreover, since in a counterpoint piece all intervals are consonances, we require that $\eta\in g(K[\epsilon])\cap K[\epsilon]$. See Figure \[counterpoint\]. This explains the first condition and part of the third one of contrapuntal symmetry on $g$. However, the model also establishes a *hierarchy of tensions* among the alternations between contrapuntal intervals and their successors, thanks to *global and local symmetries*. In contemporary physics, local symmetries are the responsible for the fundamental nature forces (gravitation, weak nuclear, electromagnetic, strong nuclear), via breaking of global symmetries of physical situations. The precise language for describing local (or gauge) symmetries is *differential geometry*, in particular, *fiber bundles* [@gauge Section 5.3]. We try to translate, *informally*, this situation to counterpoint as follows. We regard the dual numbers projection onto their cantus firmus part as a sort of *tangent bundle* over $R$ (see Figure \[bundle\]), the latter playing the role of a manifold.[^15] The number $\epsilon$ represents an infinitesimal tangent vector and $x+R\epsilon$ representing its associated **tangent space** at $x$, which we denote by $I_x$. In this way, Mazzola regards the discantus as a tangential approximation to the cantus firmus. This offers a more mathematical explanation of the use of dual numbers in counterpoint. To explain global and local symmetries in our context, let us consider two different families of partitions of the fibers of $R[\epsilon]$. First, take the partition $\{x+K\epsilon,x+D\epsilon\}$ of each $I_x$ into *real* consonances and dissonances. Second, consider the partition $$\{g(K[\epsilon])\cap I_x,g(D[\epsilon])\cap I_x\}$$ of each $I_x$ into consonances and dissonances *deformed* by a symmetry $g$ of $R[\epsilon]$. We will show that there is a unique symmetry $p^z[\epsilon]$, naturally induced by the original polarity $e^ab$ (Section \[sec5.1\]), that leaves invariant $I_z$ and interchanges $K[\epsilon]$ and $D[\epsilon]$; in particular, it interchanges consonances $z+K\epsilon$ and dissonances $z+D\epsilon$ within the fiber $I_z$. This symmetry acts by translation on fibers and sends consonances in a fiber to dissonances in the translated fiber. Moreover, if we require that $p^z[\epsilon]$ sends $g(K[\epsilon])$ to $g(D[\epsilon])$ (second condition of contrapuntal symmetry on $g$), then this symmetry also sends deformed consonances to translated deformed dissonances. According to Mazzola [@MazzolaTopos pp. 648-649], $p^z[\epsilon]$ is a **global symmetry** with respect to the first family of partitions since consonances can be translated, that is, $y+(x+K\epsilon)=(y+x)+K\epsilon$, and hence $p^z[\epsilon]$ acts on consonances by translation plus consonance/dissonance interchange. In contrast, $p^z[\epsilon]$ is usually[^16] a **local symmetry** with respect to the second family of partitions (deformed consonances/dissonances), since this translation property does not hold, and hence $p^z[\epsilon]$ does not act on deformed consonances by translation plus deformed consonance/dissonance interchange; see Figure \[globandloc\]. This means that local symmetries highly depend on the parameter $x$, whereas global symmetries are essentially constant with respect to $x$. The mathematical details of the global and the local character of $p^z[\epsilon]$ are in Section \[sec5.2\]. In Mazzola’s model, this breaking of regularity of local symmetries produces an extra tension between a contrapuntal interval and its successor beyond the mere alternation. However, the global or local character of contrapuntal symmetries is a further feature rather than a precondition in their definition. The second condition only paves the way for this feature, by means of a consistent interchange of deformed consonances and dissonances, between fibers. The crucial point is the third condition of contrapuntal symmetry on $g$, which establishes that the sets of successors $g(K[\epsilon])\cap K[\epsilon]$ must have the maximum number of elements. This ensures a maximum of alternations, which could be interpreted as a maximum independence of voices—an important principle of counterpoint. With these reflections at hand, we define first-species counterpoint in Section \[sec5.3\]. Fiber-invariant symmetries {#sec5.1} -------------------------- The interchange of contrapuntal consonances and dissonances and the invariance of the fiber $I_z$ are the basic properties of $p^z[\epsilon]$. In the case when $z=0$, $p^0[\epsilon]$ is the natural extension $e^{a\epsilon}b$ studied in Proposition \[dualexts\]. We next prove that $p^z[\epsilon]$ is uniquely determined, for strong dichotomies, and that $p^z[\epsilon]$ is the conjugate $e^z p^0[\epsilon] e^{-z}$. Let $g$ be a symmetry of $R[\epsilon]$ and $x\in R$. We say that $g$ is **$x$-invariant** if it leaves invariant the tangent space $I_x$, that is, if $g(I_x)=I_x$. [$\clubsuit$]{} We first compute some candidates to $p^x[\epsilon]$. \[charinv\] A symmetry $g$ of $R[\epsilon]$, where $g=e^{u+v\epsilon}(c+d\epsilon)$, is $x$-invariant if and only if $u=(1-c)x$. Note that $$\begin{aligned} g(x+R\epsilon) &= e^{u+v\epsilon}(c+d\epsilon)(x+R\epsilon) \\ & = cx+u+ (cR+dx+v)\epsilon = cx+u+ R\epsilon.\end{aligned}$$ Thus, $g$ is $x$-invariant if and only $cx+u=x$, the latter condition being equivalent to $u=(1-c)x$. We denote by $H_x$ the *group*[^17] of all $x$-invariant symmetries of $R[\epsilon]$. In the case when $x=0$, we simply write $H$ instead of $H_0$. The conjugation automorphism $e^x\circ (-)\circ e^{-x}$ of $\operatorname{\mathbf{Sym}}(R[\epsilon])$ restricts to an isomorphism $H\longrightarrow H_x$ for each $x$, as established in the following proposition. \[conj\] For each $x\in R$, the conjugation homomorphism $e^x\circ (-)\circ e^{-x}:H \longrightarrow H_x$ is an isomorphism of groups. First, let us prove that the conjugation homomorphism has its images in $H_x$. If $h\in H$, then $$e^{x}\circ h\circ e^{-x} (x+R\epsilon) = e^x\circ h(R\epsilon) = e^{x}(R\epsilon)=x+R\epsilon.$$ Moreover, the inverse of $e^x\circ (-)\circ e^{-x}$ is $e^{-x}\circ (-)\circ e^x$. In turn, this conjugation restricts to bijections $H(K[\epsilon],D[\epsilon])\longrightarrow H_x(K[\epsilon],D[\epsilon])$ between respective sets of symmetries sending $K[\epsilon]$ to $D[\epsilon]$ (check). The *uniqueness* in the following theorem offers an explanation of the structural role of the strong dichotomy $(K,D,p)$. \[indxdich\] Let $(K,D,e^ab)$ be a strong dichotomy of $R$. For each cantus firmus $x\in R$, there is a unique $x$-invariant symmetry $p^x[\epsilon]$, defined by $p^x[\epsilon]=e^x\circ e^{a\epsilon}b\circ e^{-x}=e^{(1-b)x+ a\epsilon}b$, such that $(K[\epsilon],D[\epsilon],p^x[\epsilon])$ is a self-complementary dichotomy of $R[\epsilon]$. In particular, by Lemma \[fiber\], $$p^x[\epsilon](x+K\epsilon)=x+D\epsilon.$$ Since self-complementary dichotomies $(K[\epsilon],D[\epsilon],p')$ with $p'$ $x$-invariant are in correspondence (conjugation) with self-complementary dichotomies $(K[\epsilon],D[\epsilon],p)$ where $p$ is $0$-invariant, the proof of the theorem reduces to the case when $x=0$ (Proposition \[unique0\]). In that case, $p^0[x]$ is the natural extension $e^{a\epsilon}b$, so $p^x[\epsilon]=e^x\circ e^{a\epsilon}b\circ e^{-x}=e^{(1-b)x+ a\epsilon}b$. \[unique0\] Let $(K,D,e^ab)$ be a strong dichotomy of $R$. The symmetry $e^{a\epsilon}b$ is the unique $0$-invariant symmetry such that $(K[\epsilon],D[\epsilon],e^{a\epsilon}b)$ is a self-complementary dichotomy of $R[\epsilon]$. Let $h$ be a $0$-invariant symmetry, where $h=e^{u+v\epsilon}(c+d\epsilon)$. By Proposition \[charinv\], $u=0$. Moreover, if $h(K[\epsilon])=D[\epsilon]$, from Lemma \[fiber\], we obtain that $cK+v=D$, so the condition that $p$ is a polarity implies that $c=b$ and $v=a$. Further, we claim that $d=0$. In fact, by the condition that $h(K[\epsilon])=D[\epsilon]$, $$e^{v\epsilon}(c+d\epsilon)(1+K\epsilon)=c+(cK+v+d)\epsilon=c+(D+d)\epsilon\subseteq D[\epsilon]$$ and $$e^{v\epsilon}(c+d\epsilon)(-1+K\epsilon)=-c+(cK+v-d)\epsilon=-c+(D-d)\epsilon\subseteq D[\epsilon],$$ so $D+d\subseteq D$ and $D-d\subseteq D$ and hence $D+d\subseteq D$ and $D\subseteq D+d$. Thus, $D=D+d$ and, since both $K$ and $D$ are rigid (Corollary \[rig\]), $d=0$. Up to now, we know that each $0$-invariant symmetry $h$ such that $h(K[\epsilon])=D[\epsilon]$ is necessarily $e^{a\epsilon}b$. It remains to show that $(K[\epsilon],D[\epsilon],e^{a\epsilon}b)$ is a self-complementary dichotomy. This was done in Proposition \[dualexts\]. \[fiber\] If $h$ is an $x$-invariant symmetry of $R[\epsilon]$ such that $h(K[\epsilon])=D[\epsilon]$, then $h(x+K\epsilon)=x+D\epsilon$. Under our assumptions, $$h(x+K\epsilon)=h(K[\epsilon]\cap(x+R\epsilon))=h(K[\epsilon])\cap(x+R\epsilon)= D[\epsilon]\cap (x+R\epsilon)=x+D\epsilon.$$ The global and the local {#sec5.2} ------------------------ Let us explain why $p^z[\epsilon]$ is a global symmetry of $R[\epsilon]$ with respect to the partition $\{x+K\epsilon,x+D\epsilon\}$ in each fiber. Given $x+r\epsilon$ in a fiber $I_x$, $$p^z[\epsilon](x+r\epsilon)=x+(b-1)(x-z)+(br+a)\epsilon.$$ Thus, $p^z[\epsilon]$ sends the fiber $I_x$ to $I_{x+(b-1)(x-z)}$, acts by translation by $(b-1)(x-z)$ units, and sends the set of consonances $x+K\epsilon$ to the set of dissonances $x+(b-1)(x-z)+D\epsilon$, corresponding to translation of $x+K\epsilon$ by $(b-1)(x-z)$ plus interchange of consonances and dissonances in $I_{x+(b-1)(x-z)}$. Note that this factorization (translation plus interchange) holds since *translation acts on sets of consonances* of the form $x+K\epsilon$, that is, $y+K\epsilon=(y-x)+x+K\epsilon$. Nevertheless, $p^z[\epsilon]$ is usually local with respect to deformed consonances/dissonances $\{g(K[\epsilon])\cap I_x, g(D[\epsilon])\cap I_x\}$. In this case, we can send deformed consonances in $I_x$ to deformed dissonances in $I_{x+(b-1)(x-z)}$ by requiring that $p^z[\epsilon]$ sends $g(K[\epsilon])$ to $g(D[\epsilon])$. Certainly, under this condition, $$\begin{aligned} p^z[\epsilon](g(K[\epsilon])\cap I_x) &= p^z[\epsilon](g(K[\epsilon]))\cap p^z[\epsilon](I_x) \\ &= g(D[\epsilon])\cap I_{x+(b-1)(x-z)}.\end{aligned}$$ But, unlike the case of real consonances, *translation does not act on sets of deformed consonances*. If $g=e^{u+v\epsilon}(c+d\epsilon)$, the consonances in $g(K[\epsilon])\cap I_x$ are those images $e^{u+v\epsilon}(c+d\epsilon)(r+k\epsilon)\in g(K[\epsilon])$ in $x+R\epsilon$, which satisfy $$e^{u+v\epsilon}(c+d\epsilon)(r+k\epsilon)=cr+u+(ck+dr+v)\epsilon$$ and $cr+u=x$. Thus, $$\label{defcons} g(K[\epsilon])\cap I_x=x+(cK+dc^{-1}(x-u)+v)\epsilon.$$ This means that, in general, $g(K[\epsilon])\cap I_y$ is not a translation of $g(K[\epsilon])\cap I_x$. We illustrate this fact with an example due to Mazzola. Let $R=\mathbb{Z}_{12}$, $(K,D,e^25)$ the Renaissance strong dichotomy, and $g=e^{8\epsilon}(5+4\epsilon)$. Here, $$p^0[\epsilon]\circ g=e^{2\epsilon}5 \circ e^{8\epsilon}(5+4\epsilon)=e^{6\epsilon}(1+4\epsilon)=e^{8\epsilon}(5+4\epsilon)\circ e^{2\epsilon}5= g\circ p^0[\epsilon],$$ and hence $p^0[\epsilon](g(K[\epsilon]))=g(D[\epsilon])$ (check). Moreover, in this case $$g(K[\epsilon])\cap (x+\mathbb{Z}_{12}\epsilon)=x+(5K+8x+8)\epsilon,$$ and hence $1+(5K+4)\epsilon$ (deformed consonances for $y=1$) is not a translation of $5+5K\epsilon$ (deformed consonances for $x=5$). [$\Diamond$]{} We also note a subtle fact from Equation : a symmetry $g$ makes translation an action on the partitions of the form $\{g(K[\epsilon])\cap I_x,g(D[\epsilon])\cap I_x\}$ if and only if $d=0$. In fact, $cK+dc^{-1}(x-u)+v=cK+dc^{-1}(y-u)+v$ for all $x,y\in R$ is equivalent to $K+c^{-1}dc^{-1}(x-u)=K+c^{-1}dc^{-1}(y-u)$ for all $x,y\in R$ and to $K+c^{-1}dc^{-1}(x-y)=K$ for all $x,y\in R$. In particular, if $x-y=1$ in the last equation, we obtain that $c^{-1}dc^{-1}=0$ by the rigidity of $K$ and hence $d=0$. On the other hand, if $d=0$, then $g(K[\epsilon])\cap I_x=x+(cK+v)\epsilon$ and we can translate the consonances through fibers. Thus, *the condition $d\neq 0$ on a deformation symmetry is equivalent to its local character*. Defining first-species counterpoint {#sec5.3} ----------------------------------- Now we can give the definition of admitted successor of a contrapuntal interval. We start with a contrapuntal consonance $\xi$. We require it to be a dissonance deformed by a symmetry $g$ (condition 1), which moves to a deformed consonance. Condition 3 ensures that this motion occurs within a maximum independence between cantus firmus and discantus. Condition 2 ensures that we have an appropriate environment for further determining the quality of the movement. In fact, local symmetries will detect tension forces. \[def\] Let $R$ be a ring, $(K,D,p)$ a *strong dichotomy* of $R$, and $(K[\epsilon],D[\epsilon],p^x[\epsilon])$, for each $x\in R$, the induced dichotomy of $R[\epsilon]$ (Theorem \[indxdich\]). - A **contrapuntal symmetry** for a consonance $\xi\in K[\epsilon]$, where $\xi=x+k\epsilon$, is a symmetry $g$ of $R[\epsilon]$ such that 1. $\xi\in g(D[\epsilon])$, 2. the triple $(g(K[\epsilon]),g(D[\epsilon]),p^x[\epsilon])$ is a self-complementary dichotomy, and 3. the cardinality of $g(K[\epsilon])\cap K[\epsilon]$ is maximum among all $g$ satisfying 1 and 2. Note that the contrapuntal symmetry for a given consonance is not required to be unique. - An **admitted successor** of a consonance $\xi\in K[\epsilon]$ is an element $\eta$ of $g(K[\epsilon])\cap K[\epsilon]$ for some contrapuntal symmetry $g$. See Figure \[counterpoint\]. - **A piece of first-species counterpoint** is a finite sequence $\xi_1,\xi_2,\dots, \xi_n$ of contrapuntal consonances, where $\xi_{i+1}$ is an admitted successor of $\xi_i$ for each $i=1,\dots ,n-1$. We require that cantus firmus and discantus belong to a certain subset $X_t$ of $R$ with a distinguished element $t\in X$, called **mode** and **tonic**, respectively. [$\clubsuit$]{} Simplified computation of admitted successors {#secsimp} ============================================= In this section we characterize admitted successors of a contrapuntal consonance $x+k\epsilon$ as the translations, by $x$, of the admitted successors of $k\epsilon$ associated with contrapuntal symmetries in $H$. This offers and important simplification of the original computation. First, we transfer symmetries satisfying conditions 1, 2, and 3, and admitted successors, between the consonances $x+k\epsilon$ and $k\epsilon$. The basic tool is the **translation permutation** $e^{x}\circ (-):\operatorname{\mathbf{Sym}}(R[\epsilon])\longrightarrow \operatorname{\mathbf{Sym}}(R[\epsilon])$, whose inverse is $e^{-x}\circ (-)$. Transfer of the first condition {#i} ------------------------------- Let $x+k\epsilon$ be a contrapuntal consonance. We define $S^{1}_{x+k\epsilon}$ as the set of all $g\in \operatorname{\mathbf{Sym}}(R[\epsilon])$ satisfying condition 1 in Definition \[def\], namely $x+k\epsilon \in g(D[\epsilon])$. Note that the translation $e^{x}\circ (-)$ restricts to a bijection $S^1_{k\epsilon}\longrightarrow S^1_{x+k\epsilon}$. This means that $k\epsilon\in g(D[\epsilon])$ if and only if $x+k\epsilon\in e^xg(D[\epsilon])$. Transfer of the second condition {#sec6.2} -------------------------------- We define $S^{2}_{x}$ as the set of all $g\in \operatorname{\mathbf{Sym}}(R[\epsilon])$ satisfying condition 2 in Definition \[def\], namely $p^x[\epsilon](g(K[\epsilon]))=g(K[\epsilon])$. The translation $e^{x}\circ (-)$ restricts to a bijection $S^2_0\longrightarrow S^2_x$, as shown by the following equivalences. Recall that $p^0[\epsilon]=e^{-x}\circ p^x[\epsilon]\circ e^{x}$ by Theorem \[indxdich\]. $$\begin{aligned} p^0[\epsilon](g(K[\epsilon]))=g(D[\epsilon]) & \Leftrightarrow\\ e^{-x}p^x[\epsilon]e^{x} g(K[\epsilon])=g(D[\epsilon]) & \Leftrightarrow\\ p^x[\epsilon](e^{x} g(K[\epsilon]))=e^{x}g(D[\epsilon])\\\end{aligned}$$ Transfer of the third condition and admitted successors {#sec6.3} ------------------------------------------------------- Let $x+k\epsilon$ be a contrapuntal consonance and $S_{x+k\epsilon}$ the set of all $g\in \operatorname{\mathbf{Sym}}(R[\epsilon])$ satisfying 1 and 2 in Definition \[def\]. Since the translation $e^{x}\circ (-)$ restricts to bijections between sets of symmetries satisfying 1 and 2, respectively (Sections \[i\] and \[sec6.2\]), then it restricts to a bijection $S_{k\epsilon}\longrightarrow S_{x+k\epsilon}$. On the other hand, note that $$\begin{aligned} e^{x}(g(K[\epsilon])\cap K[\epsilon]) & =e^{x}g(K[\epsilon])\cap e^{x}(K[\epsilon])\\ & =e^{x}g(K[\epsilon])\cap K[\epsilon].\end{aligned}$$ In fact, the first equality holds since $e^{x}$, as a bijection, commutes with intersections. Also, the second equation holds because $X[\epsilon]$ (for any $X$) is invariant under transformations of the form $e^{y}$. In particular, this means that $|g(K[\epsilon])\cap K[\epsilon]|=|e^{x}g(K[\epsilon])\cap K[\epsilon]|$ and hence, in that sense, the bijection $S_{k\epsilon}\longrightarrow S_{x+k\epsilon}$ preserves the cardinality of the sets of the form $g(K[\epsilon])\cap K[\epsilon]$. Thus, $|g(K[\epsilon])\cap K[\epsilon]|$ is maximum, among all $g$ satisfying 1 and 2 in Definition \[def\], for the consonance $k\epsilon$, if and only $|e^{x}g(K[\epsilon])\cap K[\epsilon]|$ is for $x+k\epsilon$. This means that $S_{k\epsilon}\longrightarrow S_{x+k\epsilon}$ restricts to a bijective correspondence between contrapuntal symmetries. In particular, $g(K[\epsilon])\cap K[\epsilon]$ is a set of admitted successors of $k\epsilon$ if and only if $e^xg(K[\epsilon])\cap K[\epsilon]$ is for $x+k\epsilon$. This implies the following lemma. \[transsuc\] The sets of admitted successors of $x+k\epsilon$ can be computed as the sets of the form $$e^x(g(K[\epsilon])\cap K[\epsilon]),$$ where $g$ is a contrapuntal symmetry for $k\epsilon$. Restricting symmetries ---------------------- Given a symmetry $g$, we next prove that $g(K[\epsilon])=h(K[\epsilon])$, where $h\in H$. Note that this immediately implies that $h$ is contrapuntal for $x+k\epsilon$ whenever $g$ is. This suggests to draw our attention to the contrapuntal symmetries in $H$. Recall that $H$ consists of all $0$-invariant symmetries of $R[\epsilon]$, that is, the symmetries of the form $e^{v\epsilon}(c+d\epsilon)$ with $c\in R^*$ (Proposition \[charinv\]). \[hache\] Let $g$ be a symmetry of $R[\epsilon]$. There is $h_g\in H$ such that $$g(X[\epsilon])=h_g(X[\epsilon])$$ for any nonempty subset $X$ of $R$. Concretely, if $g=e^{u+v\epsilon}(c+d\epsilon)$, then $h_g$ is $e^{(v-dc^{-1}u)\epsilon}(c+d\epsilon)$. If $g=e^{u+v\epsilon}(c+d\epsilon)$, then $$g(X[\epsilon])=\{cr+u+(cx+v+dr)\epsilon\ |\ r+x\epsilon\in X[\epsilon]\}.$$ Thus, since we can regard $cr+u$ as a parameter $r'$, ranging over $R$, of which $cx+v+dr$ depends because $r$ can be obtained as $c^{-1}r'-c^{-1}u$, we obtain that $$g(X[\epsilon])=\{r'+(cx+(v-dc^{-1}u)+dc^{-1}r')\epsilon\ |\ r'\in R\text{ and }x\in X\}.$$ The problem with this presentation is that it cannot be identified with the image of $X[\epsilon]$ under a symmetry in $H$, since $c$ in $cx$ should left multiply $r'$. For this reason, we must correct the parameter by defining a new one, say $s$, by means of the equation $cs=cr+u$. This choice yields $$\begin{aligned} g(X[\epsilon])=& \{cs+(cx+v+d(s-c^{-1}u))\epsilon\ |\ s+x\epsilon\in X[\epsilon]\}\\ =& \{cs+(cx+(v-dc^{-1}u)+ds)\epsilon\ |\ s+x\epsilon\in X[\epsilon]\}\\ =& \{e^{(v-dc^{-1}u)\epsilon}(c+d\epsilon)(s+x\epsilon)\ |\ s+x\epsilon\in X[\epsilon]\}\\ =& e^{(v-dc^{-1}u)\epsilon}(c+d\epsilon)(X[\epsilon]).\end{aligned}$$ This means that the desired $h_g$ is $e^{(v-dc^{-1}u)\epsilon}(c+d\epsilon)$. A first simplification ---------------------- Our previous observations lead to a considerably simpler characterization of contrapuntal symmetries and admitted successors, only involving symmetries in $H$. \[minisimp\] The admitted successors of a consonance $x+k\epsilon\in K[\epsilon]$ can be computed as the elements of the sets of the form $$e^x(h(K[\epsilon])\cap K[\epsilon]),$$ where $h\in H$ and - $k\epsilon\in h(D[\epsilon])$, - the triple $(h(K[\epsilon]),h(D[\epsilon]),p^0[\epsilon])$ is a self-complementary dichotomy, and - the cardinality of $h(K[\epsilon])\cap K[\epsilon]$ is maximum among all $h\in H$ satisfying a and b. By Lemma \[transsuc\], it is enough to show that the collection of all sets of the form $g(K[\epsilon])\cap K[\epsilon])$, with $g$ contrapuntal symmetry for $k\epsilon$, is equal to the set of all intersections $h(K[\epsilon])\cap K[\epsilon])$ with $h$ satisfying (a), (b), and (c) above. If $g$ is a contrapuntal symmetry for $k\epsilon$, then, since $g(K[\epsilon])=h_g(K[\epsilon])$, $h_g$ is a contrapuntal symmetry and, in particular, satisfies (a) and (b). Also, $h_g$ satisfies (c) because $H\subseteq \operatorname{\mathbf{Sym}}(R[\epsilon])$. Moreover, $g(K[\epsilon])\cap K[\epsilon]=h_g(K[\epsilon])\cap K[\epsilon]$. Conversely, if $h$ satisfies (a), (b), and (c), we claim that $h$ is a contrapuntal symmetry and hence $h(K[\epsilon])\cap K[\epsilon]$ is an usual admitted successors set. To prove the claim, take $g$ satisfying 1 and 2 in Definition \[def\] for the consonance $k\epsilon$. The symmetry $g$ satisfies (a), (b), and $g(K[\epsilon])=h_g(K[\epsilon])$, so $$|g(K[\epsilon])\cap K[\epsilon]|=|h_g(K[\epsilon])\cap K[\epsilon]|\leq |h(K[\epsilon])\cap K[\epsilon]|,$$ and $h$ is contrapuntal. Further simplification of the second condition ---------------------------------------------- Now we prove that the condition (b) in Proposition \[minisimp\] is just a commutativity one. Recall that the basic dichotomy $(K,D,e^ab)$ is *strong* by hypothesis. \[commusimp\] The triple $(h(K[\epsilon]),h(D[\epsilon]),p^0[\epsilon])$ is a self-complementary dichotomy, for $h\in H$, if and only if $p^0[\epsilon]\circ h=h\circ p^0[\epsilon]$. In fact, given $h\in H$, we have the following equivalences. $$\begin{aligned} p^0[\epsilon](h(K[\epsilon]))=h(D[\epsilon]) & \Leftrightarrow\\ p^0[\epsilon]\circ h(K[\epsilon])=h\circ p^0[\epsilon](K[\epsilon]) & \Leftrightarrow\\ p^0[\epsilon]\circ h =h\circ p^0[\epsilon]\end{aligned}$$ The first equivalence holds since $p^0[\epsilon](K[\epsilon])=D[\epsilon]$ (Theorem \[indxdich\] or Proposition \[dualexts\]). The second equivalence holds by Lemma \[rigid\]. \[rigid\] The stabilizer of $K[\epsilon]$ is $\{e^u\ |\ u\in R\}$. In particular, $K[\epsilon]$ is rigid with respect to $H$, and $f(K[\epsilon])=f'(K[\epsilon])$ implies $f=f'$ whenever $f,f'\in H$. On the one hand, note that $\{e^u\ |\ u\in R\}$ is contained in the stabilizer of $K[\epsilon]$. On the other hand, suppose that $e^{u+v\epsilon}(c+d\epsilon)(K[\epsilon])=K[\epsilon]$. Then, by intersecting both sides of the equation with the fiber $I_x$ we obtain $$e^{u+v\epsilon}(c+d\epsilon)(K[\epsilon])\cap (x+R\epsilon)=K[\epsilon]\cap (x+R\epsilon),$$ and hence $$\label{rarito} \{cr+u+(ck+dr+v)\epsilon\ | \ r\in R\text{ and }k\in K\}\cap (x+R\epsilon)=x+K\epsilon.$$ But $\{cr+u+(ck+dr+v)\epsilon\ | \ r\in R\text{ and }k\in K\}$ intersects $x+R\epsilon$ just when $cr+u=x$, that is, if and only if $r=c^{-1}(x-u)$. Thus, Equation  implies $$cK+dc^{-1}(x-u)+v=K$$ for each $x\in R$. By taking $x=u$, we obtain $cK+v=K$ and hence $v=0$ and $c=1$ since $K$ is rigid (Corollary \[rig\]). Then, by taking $x=c+u$, we obtain $K+d=K$, so $d=0$. This means that the initial symmetry is just $e^u$. Moreover, $K[\epsilon]$ is rigid with respect to $H$ since the unique element of $H$ of the form $e^u$ is the identity $e^0$. Finally, suppose that $f(K[\epsilon])=f'(K[\epsilon])$ with $f,f'$ in the *group* $H$. This implies that $f'^{-1}\circ f(K[\epsilon])=K[\epsilon]$, so $f'^{-1}\circ f=id$, and $f=f'$. Main theorem ------------ If we replace condition (b) in Proposition \[minisimp\] by its equivalent obtained in Lemma \[commusimp\], we immediately obtain the main theorem. \[redu\] Let $R$ be a ring and $(K,D,p)$ a strong dichotomy of $R$. The admitted successors of a consonance $x+k\epsilon\in K[\epsilon]$ can be computed as the elements of the sets of the form $$\label{transpose} e^x(h(K[\epsilon])\cap K[\epsilon]),$$ where $h\in H$ and 1. $k\epsilon\in h(D[\epsilon])$, 2. $p^0[\epsilon]\circ h=h\circ p^0[\epsilon]$, and 3. the cardinality of $h(K[\epsilon])\cap K[\epsilon]$ is maximum among all $h\in H$ satisfying 1 and 2. This theorem says that, to compute the admitted successors of a consonance $x+k\epsilon$, it is enough to do so for $k\epsilon$, and then apply the transposition $e^x$ (Equation ). In this simplification, we only use symmetries in $H$ and reduce the condition 2 to a suitable commutativity. The little theorem of first-species counterpoint ================================================ Let $R$ be a *finite* ring, $(K,D,p)$ a strong dichotomy of $R$ with $p=e^a b$, and $(K[\epsilon],D[\epsilon],p^0[\epsilon])$ the induced dichotomy of $R[\epsilon]$, with $p^0[\epsilon]=e^{a\epsilon}b$; see Proposition \[dualexts\] or Theorem \[indxdich\]. According to Theorem \[redu\], it is desirable to count the number of elements of the sets of the form $h(K[\epsilon])\cap K[\epsilon]$, for $h\in H$, so as to choose suitable maximum cardinals. However, this task is difficult in general. In the *Little theorem of counterpoint* we establish that each consonance has at least $|K|^2$ admitted successors. For example, in Renaissance counterpoint, each consonance has at least 36 admitted successors. Therefore, we know that there is plenty of choices to compose a counterpoint piece, though the present computation does not provide the admitted successors explicitly, which requires a greater effort. We obtain a more complete computation of cardinals in Section \[seccountfor\], and a maximization criterion in Section \[secmax\]. If $e^{v\epsilon}(c+d\epsilon)$, $c \in R^*$, belongs to $H$ and $r+k\epsilon \in K[\epsilon]$, then $$e^{v\epsilon}(c+d\epsilon)(r+k\epsilon)=cr+(ck+v+dr)\epsilon.$$ Thus, $$e^{v\epsilon}(c+d\epsilon)(K[\epsilon])=\bigsqcup\limits_{r\in R}cr+(cK+v+dr)\epsilon$$ and $$\label{presuc} e^{v\epsilon}(c+d\epsilon)(K[\epsilon])\cap K[\epsilon]=\bigsqcup\limits_{r\in R}cr+((cK+v+dr)\cap K)\epsilon$$ because $e^0 c$ is a permutation of $R$, and hence $$\label{counting} |e^{v\epsilon}(c+d\epsilon)(K[\epsilon])\cap K[\epsilon]|=\sum\limits_{r\in R}|(cK+v+dr)\cap K|.$$ The proof idea of the little theorem of counterpoint is the following. Equation  gives an expression for $|h(K[\epsilon])\cap K[\epsilon]|$, whenever $h\in H$. As we show in Section \[lower\], the exact value of the sum in Equation  is $|K|^2$ if $d=1$. Thus, given a consonance $k\epsilon$, if we prove that there is at least an $h\in H$, with $d=1$, satisfying the conditions 1 and 2 in Theorem \[redu\], then we deduce that the number of admissible successors is at least $|K|^2$ by the maximum property of admitted successors sets associated with contrapuntal symmetries. So as to prove the existence of such an $h$, we first need to give concrete criteria for deciding when an $h\in H$ satisfies 1 or 2. We do this in Sections \[1con\] and \[2con\]. Finally, we translate our result to any consonance $x+k\epsilon$, by using the bijective translation $e^x$. The first condition criterion {#1con} ----------------------------- Regarding the condition 1 in Theorem \[redu\], given $h=e^{v\epsilon}(c+d\epsilon)$, we have the following equivalences, where $p$ is the polarity $e^ab$ of $(K,D,e^ab)$. $$\begin{aligned} k\epsilon \in h(D[\epsilon]) & \Leftrightarrow k \in cD+v \\ & \Leftrightarrow k \in c\cdot p(K)+v\\ & \Leftrightarrow v=k-c\cdot p(s)\text{ for some }s\in K\end{aligned}$$ The second condition criterion {#2con} ------------------------------ As to 2 in Theorem \[redu\], given $h=e^{v\epsilon}(c+d\epsilon)$, we have the following equivalences. $$\begin{aligned} p^0[\epsilon]\circ h=h\circ p^0[\epsilon]& \Leftrightarrow e^{a\epsilon}b\circ e^{v\epsilon}(c+d\epsilon) =e^{v\epsilon}(c+d\epsilon)\circ e^{a\epsilon}b\\ & \Leftrightarrow e^{(bv+a)\epsilon}(bc+bd\epsilon)=e^{(ca+v)\epsilon}(cb+db\epsilon)\\ & \Leftrightarrow bv+a=ca+v\text{, }bc=cb\text{, and }bd=db\end{aligned}$$ The little theorem {#lower} ------------------ We can prove the existence of an $h\in H$ with $d=1$ that satisfies 1 and 2 in Theorem \[redu\], for the consonance $k\epsilon$. In fact, according to Sections \[1con\] and \[2con\], it is enough to show that the following system of equations, in the unknowns $c\in R^*$, $v\in R$, and $s\in K$, has at least a solution for $d=1$. $$\left\lbrace \begin{matrix} v=k-cbs-ca\\ bv+a = ca+v \\ bc = cb\\ bd = db \end{matrix} \right.$$ Certainly, it has the solution $s=k$, $c=b$ and $v=-ba$ (note that $b^2=1$ since $e^ab$ is involutive by Proposition \[involut\]). To conclude, $e^{-ba\epsilon}(b+\epsilon)$ is the desired $h$. We are almost done, but before, we need a lemma that allows to give the exact value of $|h(K[\epsilon])\cap K[\epsilon]|$ for the $h$ obtained. \[basiccount\] Let $K$ and $K'$ be subsets of a finite Abelian group $R$. The equation $$\sum\limits_{r\in R}|(K'+r)\cap K|=|K'||K|$$ holds. Let $\chi_K:R\longrightarrow \{0,1\}$ be the characteristic function of $K$. $$\begin{aligned} \sum\limits_{r\in R}|(K'+r)\cap K| &= \sum\limits_{r\in R} \sum\limits_{s\in K'} \chi_{K}(s+r)\\ &= \sum\limits_{s\in K'} \sum\limits_{r\in R} \chi_{K}(s+r)\\ &= \sum\limits_{s\in K'} \sum\limits_{r'\in R} \chi_{K}(r')\\ &= \sum\limits_{s\in K'} |K|=|K'||K|\end{aligned}$$ \[prelow\] Let $R$ be a finite ring and $(K,D,p)$ a strong dichotomy. Each consonance $k\epsilon$ has at least $|K|^2$ admitted successors. If $h\in H$ satisfies $d=1$, then, according to Lemma \[basiccount\], $$|h(K[\epsilon])\cap K[\epsilon]|=\sum\limits_{r\in R}|(cK+v+r)\cap K|=|cK+v||K|=|K|^2.$$ Hence, $h$, with $h=e^{-ba\epsilon}(b+\epsilon)$, satisfies 1, 2, and the previous equation. Now, let $N$ be the number of admitted successors of $k\epsilon$. If $h'$ satisfies the conditions of Theorem \[redu\], for $k\epsilon$, then $$N\geq |h'(K[\epsilon])\cap K[\epsilon]|\geq |h(K[\epsilon])\cap K[\epsilon]|=|K|^2.$$ An upper bound for a single contrapuntal symmetry can be obtained as well. First, suppose that $d$ is not $0$. Note that $$\sum_{r\in R}|(cK+v+dr)\cap K|=\rho \sum_{r\in \operatorname{Im}(d\cdot\underline{\ \ })}|(cK+v+r)\cap K|,$$ where $d\cdot\underline{\ \ }:R\longrightarrow R$ is the $R$-endomorphism that sends an element $r\in R$ to $dr$, and $\rho =| \operatorname{Ker}(d\cdot\underline{\ \ })|$. In the case when $e^{v+r}c$ is not the identity $e^01$, note that $|(cK+v+r)\cap K|$ is at most $|K|-1$ since $K$ is rigid. On the other hand, there is at most a value of $r$ in $\operatorname{Im}(d\cdot\underline{\ \ })$ that makes $e^{v+r}c$ the identity, and hence there is at most an $r$ such that $|(cK+v+r)\cap K|=|K|$. Thus, $$\begin{aligned} \rho \sum_{r\in \operatorname{Im}(d\cdot\underline{\ \ })}|(cK+v+r)\cap K| &\leq \rho [(|\operatorname{Im}(d\cdot\underline{\ \ })|-1)(|K|-1)+|K| ]\\ & = \rho \left[\left(\frac{|R|}{\rho}-1\right)(|K|-1)+|K|\right]\\ & =\rho \left[\left(\frac{2|K|}{\rho}-1\right)(|K|-1)+|K|\right]\\ & = 2|K|^2-2|K|+\rho\\ &\leq 2|K|^2-2|K|+|K|=2|K|^2-|K|.\end{aligned}$$ As to the last inequality note that $\rho$, as a divisor of $|R|$ that is not[^18] $|R|$, must be less than or equal to $|K|$ (which coincides with $|R|/2$), the greatest divisor of $|R|$ different from $|R|$. Now suppose that $d=0$. So as to find an upper bound for $$\sum_{r\in R}|(cK+v)\cap K|$$ we consider two cases. If $c\neq 1$, then $e^vc$ is not the identity $e^01$. If $c=1$, then, since we require $h$ (for $h=e^{v\epsilon}c$) to satisfy condition 1 in Theorem \[redu\], by Section \[1con\], $v=k- p(s)$ for some $s\in K$. This means that $v\neq 0$ since $p(s)\in D$, and hence $e^vc$ is not the identity. In both cases, $|(cK+v)\cap K|\leq |K|-1$ since $K$ is rigid. Thus, $$\sum_{r\in R}|(cK+v)\cap K|\leq |R|(|K|-1)=2|K|(|K|-1).$$ To sum up, collecting the results for $d\neq 0$ and $d=0$ and using Theorem \[redu\], we obtain the upper bound in the following theorem. \[low\] Let $R$ be a finite ring and $(K,D,p)$ a strong dichotomy. Each consonance $x+k\epsilon$ has at least $|K|^2$ admitted successors, and at most $2|K|^{2}-|K|$ for a single contrapuntal symmetry. By Theorem \[redu\], $e^x$ is a bijection between the sets of admitted successors of $k\epsilon$ and $x+k\epsilon$. Thus, the number of admitted successors of $x+k\epsilon$ is equal to the number of admitted successors of $k\epsilon$. The lower bound now follows from Lemma \[prelow\]. The upper bound corresponds to the preceding discussion. Counting formulas and maximization {#seccountfor} ================================== Now we generalize Lemma \[basiccount\] so as to have a better approximation to the exact value of the terms in Equation . \[gencount\] Let $K$ and $K'$ be subsets of a finite ring $R$ and $d\in R$. The equation $$\label{forgen} \sum\limits_{r\in R}|(K'+dr)\cap K|=\rho\sum\limits_{\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })}|K'_{\gamma}||K_{\gamma}|$$ holds, where $\operatorname{Coker}(d\cdot\underline{\ \ })=R/ \operatorname{Im}(d\cdot\underline{\ \ })$, $$X_{\gamma}=\{x\in X\ |\ [x]=\gamma\}$$ for each $X\subseteq R$, and $\rho=|\operatorname{Ker}(d\cdot\underline{\ \ })|$. $$\begin{aligned} \sum\limits_{r\in R}|(K'+dr)\cap K| &= \rho\sum\limits_{r\in \operatorname{Im}(d\cdot\underline{\ \ })}|(K'+r)\cap K|\\ &= \rho\sum\limits_{r\in \operatorname{Im}(d\cdot\underline{\ \ })} \sum\limits_{s\in K'} \chi_{K}(s+r)\\ &= \rho \sum\limits_{s\in K'} \sum\limits_{r\in \operatorname{Im}(d\cdot\underline{\ \ })} \chi_{K}(s+r)\\ &= \rho\sum\limits_{s\in K'} |K_{[s]}|\\ &= \rho\sum\limits_{\gamma\in K'/\operatorname{Im}(d\cdot\underline{\ \ })} |K'_{\gamma}||K_{\gamma}|=\rho\sum\limits_{\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })} |K'_{\gamma}||K_{\gamma}|.\end{aligned}$$ As to the last equality, note that if $\gamma$ is not in the image $K'/\operatorname{Im}(d\cdot\underline{\ \ })$, of $K'$ under the canonical projection onto the cokernel, then $|K'_{\gamma}|=0$. The following corollary illustrates the lemma in the important case when the ring is $\mathbb{Z}_{n}$. \[countint\] Let $K$ and $K'$ be subsets of $\mathbb{Z}_{n}$ and $d\in \mathbb{Z}_{n}$. The equation $$\label{forint} \sum\limits_{r=0}^{n-1}|(K'+dr)\cap K|=\rho\sum\limits_{i=0}^{\rho-1}|K'_{i}||K_{i}|$$ holds, where $$X_{i}=\{x\in X\ |\ x \equiv i\pmod{\rho}\}$$ for each $X\subseteq \mathbb{Z}_{12}$ and $\rho=\gcd (d,n)$. Note that (isomorphism and index theorems for groups) $$|\operatorname{Ker}(d\cdot\underline{\ \ })|=|R|/|\operatorname{Im}(d\cdot\underline{\ \ })|=|R/\operatorname{Im}(d\cdot\underline{\ \ })|=|\operatorname{Coker}(d\cdot\underline{\ \ })|.$$ Now, by [@Fraleigh Theorem 6.14], $R/\operatorname{Im}(d\cdot\underline{\ \ })=\mathbb{Z}_{12}/d\mathbb{Z}_{12}=\mathbb{Z}_{12}/\rho\mathbb{Z}_{12}\cong \mathbb{Z}_{\rho}$, where $\rho=\gcd(d,n)$, so $|\operatorname{Ker}(d\cdot\underline{\ \ })|=|\operatorname{Coker}(d\cdot\underline{\ \ })|=\rho$. Moreover, $[x]=\gamma$ for $\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })$ if and only if $x \equiv i\pmod{\rho}$ by identifying $\gamma$ and $i$ through the isomorphism $\operatorname{Coker}(d\cdot\underline{\ \ })\cong \mathbb{Z}_{\rho}$. Just two curiosities. In Lemma \[gencount\], the case $d=1$ corresponds essentially to Lemma \[basiccount\]. The case $d=0$ is just the formula $$\sum\limits_{r\in R}|K'\cap K|=|R||K'\cap K|.$$ Let $(K,D,e^2 5)$ be the Renaissance strong dichotomy of $\mathbb{Z}_{12}$, with $K=\{0,3,4,7,8,9\}$ and $D=\{1,2,5,6,10,11\}$. The following table contains the particular values of the right-hand side of Equation , by using Equation  with $K'=cK+v$, for $d$ ranging over $\mathbb{Z}_{12}$. After the table, we briefly justify each row. $d$ $\rho$ $\sum\limits_{r=0}^{11}|(K'+dr)\cap K|$ ------------- -------- ------------------------------------------------- 0 12 $12(|K'_0|+|K'_3|+|K'_4|+|K'_7|+|K'_8|+|K'_9|)$ 1, 5, 7, 11 1 36 2, 10 2 36 3, 9 3 $3(|K'_0|3+|K'_1|2+|K'_2|)$ 4, 8 4 $4(|K'_0|3+|K'_1|+|K'_3|2)$ 6 6 $6(|K'_0|+|K'_1|+|K'_2|+|K'_3|2+|K'_4|)$ *Case* $d=0$. The number of elements in $K$ congruent to $i$ modulo $12$ is just given by the characteristic function of $K$ as a subset of $\mathbb{Z}_{12}$. *Case* $d=1$, $5$, $7$, $11$. Since all integers are congruent modulo 1, we obtain the expression $|K'||K|$ for the sum, whose exact value is $6\times 6$. The *case* $d=2$, $10$ is quite interesting. The integers congruent to $0$ (respectively $1$) modulo $2$ are the even (respectively odd) ones. Now, there are three even numbers ($0$, $4$, and $8$) and three odd numbers ($3$, $7$, and $9$) in $K$. Thus, the counting formula becomes $2(|K'_0|3+|K'_1|3)$. But multiplying $K$ by $c$ (which is always odd because it is invertible) does not alter the parity of its elements, and adding $v$ to $cK$ does not alter the number of odd or even elements. For this reason, $|K'_0|=|K'_1|=3$ and the exact value of the counting formula is $2( (3\times 3)+ (3\times 3))$. *Case* $d=3$, $9$. In this case, $|K_0|=|\{0,3,9\}|=3$, $|K_1|=|\{4,7\}|=2$, $|K_2|=|\{8\}|=1$, and the remaining terms of the form $K_i$ are empty. *Case* $d=4$, $8$. Here $|K_0|=|\{0,4,8\}|=3$, $|K_1|=|\{9\}|=1$, $|K_3|=|\{3,7\}|=2$, and the remaining terms of the form $K_i$ are empty. *Case* $d=6$. Here $|K_0|=|\{0\}|=1$, $|K_1|=|\{7\}|=1$, $|K_2|=|\{8\}|=1$, $|K_3|=|\{3,9\}|=2$, $|K_4|=|\{4\}|=1$, and the remaining terms of the form $K_i$ are empty. [$\Diamond$]{} Another important consequence of Equation  $$\sum\limits_{r\in R}|(K'+dr)\cap K| = \rho_{d}\sum\limits_{r\in \operatorname{Im}(d\cdot\underline{\ \ })}|(K'+r)\cap K|$$ is that if $\operatorname{Im}(d\cdot\underline{\ \ })=\operatorname{Im}(d'\cdot\underline{\ \ })$, then $$\rho_{d}=|\operatorname{Ker}(d\cdot\underline{\ \ })|=|R|/|\operatorname{Im}(d\cdot\underline{\ \ })|=|R|/|\operatorname{Im}(d'\cdot\underline{\ \ })|=\rho_{d'}$$ and hence $$\sum\limits_{r\in R}|(K'+dr)\cap K| =\sum\limits_{r\in R}|(K'+d'r)\cap K|.$$ Moreover, in the case when the ring is $\mathbb{Z}_n$, we observe from Equation  that we can reduce the computation of the sums of the form $\sum\limits_{r\in R}|(K'+dr)\cap K|$ for all $d$ with the same $\rho$ (recall that $\rho=\gcd(d,n)$) to the computation of the sum $\sum\limits_{r\in R}|(K'+\rho r)\cap K|$, since these sums coincide. Main counting formulas {#commufor} ---------------------- Now, according to Equation  we will focus on the case when $K'=cK+v$. Let us assume that $R$ is commutative. For each $d$, the sequence $(|K'_{\gamma}|)_{\gamma}$ is a *rearrangement*[^19] of $(|K_{\gamma}|)_{\gamma}$, where $\gamma$ ranges over $\operatorname{Coker}(d\cdot\underline{\ \ })$. In fact, note first that $$\begin{aligned} |K'_{[r]}|&=|\{e^vc(k)\ |\ k\in K\text{ and }[e^vc(k)]=[r] \text{ in } \operatorname{Coker}(d\cdot\underline{\ \ })\}|\\ &=|\{e^vc(k)\ |\ k\in K\text{, }ck+v-r \in dR\}|\\ &=|\{e^vc(k)\ |\ k\in K\text{, }k-(c^{-1}r- c^{-1}v)\in dR\}|\\ &=|\{e^vc(k)\ |\ k\in K\text{, }k-(e^vc)^{-1}(r)\in dR\}|\\ &=|\{e^vc(k)\ |\ k\in K\text{, }[k]=[(e^vc)^{-1}(r)]\text{ in } \operatorname{Coker}(d\cdot\underline{\ \ })\}|\\ &=|\{k\in K\ |\ [k]=[(e^vc)^{-1}(r)]\}|=|K_{[(e^vc)^{-1}(r)]}|.\end{aligned}$$ Second, by the commutativity of $R$, the function (actually a ring symmetry) $$\begin{array}{cccc} [e^vc]:&\operatorname{Coker}(d\cdot\underline{\ \ }) & \longrightarrow & \operatorname{Coker}(d\cdot\underline{\ \ })\\ &{[r]} & \longmapsto & [e^vc(r)] \end{array}$$ is well defined (check) with inverse given by the function induced by $(e^vc)^{-1}$, so $[e^vc]$ and $[(e^vc)^{-1}]$ are permutations of the cokernel. Thus, regarding $(|K'_{\gamma}|)_{\gamma}$ and $(|K_{\gamma}|)_{\gamma}$ as functions from the cokernel to $\mathbb{N}$, the former is the composite of the latter with $[(e^vc)^{-1}]$. Also, note that $[(e^vc)^{-1}]=[e^{c^{-1}v}c^{-1}]=e^{[c]^{-1}[v]}[c]^{-1}$. This proves the main counting formula. \[mainform\] Let $R$ be a commutative ring, $K\subseteq R$, and $e^{v\epsilon}(c+d\epsilon)$ a symmetry in $H$. The equation $$\label{maincount} |e^{v\epsilon}(c+d\epsilon)(K[\epsilon])\cap K[\epsilon]|=\rho\sum\limits_{\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })}|K_{[(e^vc)^{-1}](\gamma)}||K_{\gamma}|=\rho\sum\limits_{\alpha\in \operatorname{Coker}(d\cdot\underline{\ \ })}|K_{\alpha}||K_{[e^vc](\alpha)}|$$ holds, where $K_{\gamma}=\{x\in K\ |\ [x]=\gamma\}$ and $\rho=|\operatorname{Ker}(d\cdot\underline{\ \ })|$. In particular, if $R=\mathbb{Z}_n$, the right-hand term of Equation  coincides with $$\label{maincountforint} \rho\sum\limits_{i=0}^{\rho-1}|K_{i}||K_{e^vc(i)}|,$$ where $K_{i}=\{x\in K\ |\ x \equiv i\pmod{\rho}\}$, $\rho=\gcd (d,n)$, and $e^vc$ is reduced modulo $\rho$. Combine the previous discussion with Equation , Lemma \[gencount\], and Corollary \[countint\]. The second equality in Equation  follows from the first one and the change of variable $\alpha=[(e^vc)^{-1}](\gamma)$. Maximization criterion {#secmax} ---------------------- The following maximization criterion helps to find symmetries $h$ such that $|h(K[\epsilon])\cap K[\epsilon]|$ is maximum among all symmetries with $h=e^{v\epsilon}(c+d\epsilon)$ and $d$ fixed. It is important to emphasize that it need not find symmetries with maximum values among all symmetries satisfying conditions 1 and 2 of contrapuntal symmetry, but the criterion is very useful to discard a number of symmetries whose values are not maximum. \[max\] Assume the hypotheses of Theorem \[mainform\]. The right-hand side of Equation  $$\rho\sum\limits_{\alpha\in \operatorname{Coker}(d\cdot\underline{\ \ })}|K_{\alpha}||K_{[e^vc](\alpha)}|$$ is maximum, for $e^vc$ ranging over all symmetries of $R$, if and only if $|K_{[e^vc](\alpha)}|=|K_{\alpha}|$ for each $\alpha\in \operatorname{Coker}(d\cdot\underline{\ \ })$. In particular, the sum is maximum if $e^vc\equiv e^01\pmod{d}$, that is,[^20] if $v\equiv 0\pmod{d}$ and $c\equiv 1\pmod{d}$. Further, the maximum value is $$\rho\sum\limits_{\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })}|K_{\gamma}|^2.$$ Apply the rearrangement inequality (Theorem \[r.i.\]) to the sequences $(|K_{\gamma}|)_{\gamma}$ and $(|K_{[e^vc](\gamma)}|)_{\gamma}$, the latter being a rearrangement of the former by Section \[mainform\]. Certainly, the criterion is useful because if a symmetry $e^{v\epsilon}(c+d\epsilon)$ satisfies the conditions 1 and 2 for a contrapuntal symmetry and $e^vc\equiv e^01\pmod{d}$, then we get rid of all symmetries with the same $d$ that induce a rearrangement of $(K_{\gamma})_{\gamma}$, which are usually all but those satisfying $e^vc\equiv e^01\pmod{d}$. Steps for the computation of contrapuntal symmetries ---------------------------------------------------- The steps for calculating the contrapuntal symmetries in $H$, of the form $e^{v\epsilon}(c+d\epsilon)$, for a consonance of the form $k\epsilon$ is the following. We start with a strong dichotomy $(K,D,e^ab)$. 1. Solve the following system of equations in the unknowns $c\in R^*$, $v$, and $d$. $$\left\lbrace \begin{matrix} bv+a = ca+v \\ bc = cb\\ bd = db \end{matrix} \right.$$ This corresponds to condition 2 of contrapuntal symmetry (Section \[2con\]). 2. For each consonance $k\epsilon$, among the symmetries of the form $e^{v\epsilon}(c+d\epsilon)$ obtained in 1, choose those with $v\in k-cD$. This corresponds to condition 1 of contrapuntal symmetry (Section \[1con\]). The reason for first computing the symmetries satisfying condition 2 is that those symmetries are usually less than those satisfying condition 1, so we perform less operations. 3. For each consonance $k\epsilon$, among the symmetries $e^{v\epsilon}(c+d\epsilon)$ obtained in 2, choose those such that the right-hand term of Equation  $$\rho\sum\limits_{\gamma\in \operatorname{Coker}(d\cdot\underline{\ \ })}|(cK+v)_{\gamma}||K_{\gamma}|$$ with $K'=cK+v$ is maximum. This ensures that the condition 3 holds by Equation . In the case when $R$ is commutative, maximize the right-hand side of Equation . In the case when $R=\mathbb{Z}_n$, maximize Equation . So as to discard a number of symmetries, if $R$ is commutative, Theorem \[max\] can be used. In the next section we apply the previous steps to computing the contrapuntal symmetries in the case of first-species Renaissance counterpoint. First-species Renaissance counterpoint {#hand} ====================================== Let $(K,D,e^2 5)$ be the Renaissance strong dichotomy of $\mathbb{Z}_{12}$, with $K=\{0,3,4,7,8,9\}$ and $D=\{1,2,5,6,10,11\}$. Condition 2 {#secondcond} ----------- By the commutativity of $\mathbb{Z}_{12}$, the condition 2 for symmetries $e^{v\epsilon}(c+d\epsilon)\in H$, $c\in R^*$, reduces to solve the equation $$5v+2=c2+v,$$ which is equivalent to $$4v=2(c-1).$$ If $c=1,7$, then the equation becomes $4v=0$, so $v=0,3,6,9$. If $c=5,11$, then the equation becomes $4(v-2)=0$ and hence $v=2,5, 8,11$. The solutions are summarized in the following table. $c$ $v$ ------- ------------- 1, 7 0, 3, 6, 9 5, 11 2, 5, 8, 11 Condition 1 {#firstcond} ----------- Now, among these solutions, we choose, for each consonance $k\epsilon$, those satisfying $$v\in k-cD.$$ In the following table, we organize the results of the operations involved. Specifically, we compute $3\mathbb{Z}_{12}\cap (k-cD)$ for $c=1,7$ and $(2+3\mathbb{Z}_{12})\cap (k-cD)$ for $c=5,11$. 0 3 4 7 8 9 ---- -------------- ------------- ------------ ----------- ------------- -------------- 1 $\{6\}$ $\{9\}$ $\{3,6\}$ $\{6,9\}$ $\{3,6,9\}$ $\{3\}$ 5 $\{2,5,11\}$ $\{2,5,8\}$ $\{2,11\}$ $\{2,5\}$ $\{2\}$ $\{2,8,11\}$ 7 $\{6\}$ $\{9\}$ $\{6,9\}$ $\{0,9\}$ $\{3,6,9\}$ $\{3\}$ 11 $\{2,5,11\}$ $\{2,5,8\}$ $\{2,5\}$ $\{5,8\}$ $\{2\}$ $\{2,8,11\}$ We can easily fill in the table as follows. There are six entries, namely those labelled by $(c,k)$ with $c=1,5$ and $k=0,4,8$, which we can start from, the other being obtained by using certain symmetries. In fact, note that $$7(3\mathbb{Z}_{12}\cap (k-D))=3\mathbb{Z}_{12}\cap (7k-7D)),$$ $$7((2+3\mathbb{Z}_{12})\cap (k-5D))=(2+3\mathbb{Z}_{12})\cap (7k-11D),$$ $$\pm 3+(3\mathbb{Z}_{12}\cap (k-cD))=3\mathbb{Z}_{12}\cap (\pm 3+k-cD),$$ and $$\pm 3+((2+3\mathbb{Z}_{12})\cap (k-cD))=(2+3\mathbb{Z}_{12})\cap (\pm 3+k-cD).$$ For example, the entry $(1,3)$ is obtained by adding $3$ to the entry $(1,0)$, and the entry $(7,0)$ is obtained by multiplying the entry $(1,0)$ by $7$. Maximization {#countsym} ------------ In the case of the Renaissance dichotomy, Theorem \[max\] allows us to obtain the following table of maximum values, without the restrictions of conditions 1 and 2 of contrapuntal symmetry, and their corresponding symmetries. We do not include the value $d=0$ since in that case the maximum is not useful, because the identity does not satisfy the condition 1. After the table we justify the results. $d$ $\rho$ maximum $\sum\limits_{r=0}^{11}|(cK+v+dr)\cap K|$ $e^vc$ ------------- -------- --------------------------------------------------- ----------------------- 1, 5, 7, 11 1 36 any 2, 10 2 36 any 3, 9 3 42 $\equiv e^01\pmod{3}$ 4, 8 4 56 $\equiv e^01\pmod{4}$ 6 6 48 $\equiv e^01\pmod{6}$ If $\rho=3$, since all entries of the vector $(|K_0|,|K_1|,|K_2|)$ are different, then the unique rearrangement that coincides with it is the composition with $e^01$, and hence the maximum is only taken for the identity modulo $3$. The same is true for $\rho=4$. If $\rho=6$, then $(|K_0|,|K_1|,|K_2|,|K_3|,|K_4|,|K_5|)=(1,1,1,2,1,0)$. Now, if a symmetry $e^vc$ modulo 6 induces a rearrangement that leaves the vector invariant, then the rearrangement leaves $K_3$ invariant, but $e^vc(3)=3+v=3$ and hence $v=0$. Moreover, $e^05(1)=5$ and hence it does not induce a rearrangement that leaves $K_1$ invariant, so $e^vc=e^01$. **Contrapuntal symmetries for $0\epsilon$.** According to the table in Section \[firstcond\], there are exactly two symmetries for $\rho=6$, namely $e^{6\epsilon}(1+6\epsilon)$ and $e^{6\epsilon}(7+6\epsilon)$, with $e^vc$ congruent to $e^01$ modulo $6$. This allows us to discard the cases $\rho=3,2,1$. Up to now the maximum is $48$, and it remains to examine the cases $\rho=4,12$. In the case when $\rho=4$, the residues modulo 4 of the candidates in the table produce the following sums, calculated with Equation . 1 2 3 --- ---- ---- ---- 1 36 16 36 3 24 20 48 Thus, we have find two new symmetries with sum 48 (the maximum up to now), namely $e^{11 \epsilon}(11+8\epsilon)$ and $e^{11 \epsilon}(11+4\epsilon)$. Here, $e^{11}11 \equiv e^33\pmod{4}$. It remains to examine the case $\rho=12$. We have the following sums. 6 2 5 11 ---- ---- ---- ---- ---- 1 24 7 36 5 0 36 36 11 12 24 48 This means that the maximum sum is $48$ and that there is yet another symmetry $e^{11\epsilon}11$ for our account. **Contrapuntal symmetries for $3\epsilon$** According to the second table in Section \[firstcond\], there are exactly two symmetries for $\rho=4$, namely $e^{8\epsilon}(5+4\epsilon)$ and $e^{8\epsilon}(5+8\epsilon)$, with $e^vc$ congruent to $e^01$ modulo $4$. We thus discard the cases $\rho=6,3,2,1$. Up to now the maximum is $56$. The remaining case is $\rho=12$. We have the following sums. 9 2 5 8 ---- ---- ---- ---- ---- 1 36 7 24 5 0 36 48 11 12 24 36 The maximum is 56 and there are no more contrapuntal symmetries. **Contrapuntal symmetries for $4\epsilon$** There are exactly two symmetries for $\rho=6$, namely $e^{6\epsilon}(1+6\epsilon)$ and $e^{6\epsilon}(7+6\epsilon)$, with $e^vc$ congruent to $e^01$ modulo $6$. This allows us to discard the cases $\rho=3,2,1$. Up to now the maximum is $48$. The sums are less than 48 for $\rho=4$ as shown in the following table. 1 2 3 --- ---- ---- ---- 1 16 36 3 24 20 The case $\rho=12$ yields the following sums. 3 6 2 11 9 5 ---- ---- ---- ---- ---- ---- ---- 1 36 24 5 0 36 7 36 24 11 12 24 Hence, the maximum is 48 and there are no more contrapuntal symmetries. **Contrapuntal symmetries for $7\epsilon$** The case $\rho=12$ yields the following sums. 0 6 2 8 9 5 ---- ---- ---- --- ---- ---- ---- 1 24 36 5 0 36 7 60 24 11 36 24 Thus, we discard all remaining cases for $\rho$ and the maximum is 60, corresponding to a unique contrapuntal symmetry $e^07$. **Contrapuntal symmetries for $8\epsilon$** There are exactly two symmetries for $\rho=6$, namely $e^{6\epsilon}(1+6\epsilon)$ and $e^{6\epsilon}(7+6\epsilon)$, with $e^vc$ congruent to $e^01$ modulo $6$. This allows us to discard the cases $\rho=3,2,1$. Up to now the maximum is $48$. The sums for $\rho=4$ are the same of the consonance $0\epsilon$. In the table from Section \[firstcond\], we observe that there are two additional symmetries, namely $e^{3\epsilon}(7+4\epsilon)$ and $e^{3\epsilon}(7+8\epsilon)$, with $e^vc$ congruent to $e^33$ modulo $4$. The case $\rho=12$ yields the following sums. 2 3 6 9 ---- ---- ---- ---- ---- 1 36 24 36 5 0 7 48 36 24 11 12 Hence, the maximum is 48 and there is an additional contrapuntal symmetry $e^{3\epsilon}7$. **Contrapuntal symmetries for $9\epsilon$** There are exactly two symmetries for $\rho=4$, namely $e^{8\epsilon}(5+4\epsilon)$ and $e^{8\epsilon}(5+8\epsilon)$, with $e^vc$ congruent to $e^01$ modulo $4$. We thus discard the cases $\rho=6,3,2,1$. Up to now the maximum is $56$. For $\rho=12$ we have the following sums. 3 2 11 8 ---- ---- ---- ---- ---- 1 36 7 48 5 0 36 48 11 12 48 36 The maximum is 56 and there are no more contrapuntal symmetries. Admitted successors {#secsuc} ------------------- We finally can obtain the list of admitted successors, by directly using Equation  for each contrapuntal symmetry $h$ obtained in Section \[countsym\]; see Table \[tab:adsu\]. In this table we can appreciate the utility of the maximization criterion (Theorem \[max\]). The contrapuntal symmetries $e^{6\epsilon}(1+6\epsilon)$ (three times) and $e^{6\epsilon}(7+6\epsilon)$ (three times) are equal to the identity modulo $6$, and the contrapuntal symmetries $e^{8\epsilon}(5+8\epsilon)$ (two times) and $e^{8\epsilon}(5+4\epsilon)$ (two times) are equal to the identity modulo $4$. In these cases, the criterion was used to discard many symmetries $h$ such that the cardinalities of the associated deformed consonances sets $h(K[\epsilon])\cap K[\epsilon]$ were not maximum. On the other hand, the **prohibition of parallel fifths** is an important conclusion that we can draw from the table. In fact, the consonance $7\epsilon$ has as set of admitted successors $\mathbb{Z}_{12}+(K\setminus\{7\})\epsilon$, that is, any interval can follow a fifth, except a fifth. In the following section, we show that the predictions of the model are congruent with the examples from [@Fux]. However, a rigorous comparison between the allowances/prohibitions of the model and Fux is to be eventually communicated, because the study mentioned in [@MazzolaTopos Section 31.4.1] was not published. $k$ $|h(K[\epsilon])\cap K[\epsilon]|$ $h$ Admitted successors of $k\epsilon$ ----- ------------------------------------ ------------------------------ --------------------------------------------- $r+\{3,9\}\epsilon$, $r$ even $r+K\epsilon$, $r$ odd $r+\{3,7,9\}\epsilon$, $r$ even $r+(K\setminus\{7\})\epsilon$, $r$ odd $\{0,3,6,9\}+\{3,4,7,8\}\epsilon$ $\{1,4,7,10\}+\{0,3,7,8\}\epsilon$ $\{2,5,8,11\}+\{0,3,4,7\}\epsilon$ $\{0,3,6,9\}+\{3,4,7,8\}\epsilon$ $\{1,4,7,10\}+\{0,3,4,7\}\epsilon$ $\{2,5,8,11\}+\{0,3,7,8\}\epsilon$ $e^{11\epsilon}11$ $\mathbb{Z}_{12}+\{3,4,7,8\}\epsilon$ $\{0,3,6,9\}+\{0,4,7,8\}\epsilon$ $\{1,4,7,10\}+(K\setminus\{7\})\epsilon$ $\{2,5,8,11\}+(K\setminus\{9\})\epsilon$ $\{0,3,6,9\}+\{0,4,7,8\}\epsilon$ $\{1,4,7,10\}+(K\setminus\{9\})\epsilon$ $\{2,5,8,11\}+(K\setminus\{7\})\epsilon$ $e^{6\epsilon}(1+6\epsilon)$ $e^{6\epsilon}(7+6\epsilon)$ $7$ $60$ $e^07$ $\mathbb{Z}_{12}+(K\setminus\{7\})\epsilon$ $e^{3\epsilon}7$ $\mathbb{Z}_{12}+\{0,3,4,7\}\epsilon$ $e^{6\epsilon}(1+6\epsilon)$ $e^{6\epsilon}(7+6\epsilon)$ $\{0,3,6,9\}+\{0,3,4,7\}\epsilon$ $\{1,4,7,10\}+\{3,4,7,8\}\epsilon$ $\{2,5,8,11\}+\{0,3,7,8\}\epsilon$ $\{0,3,6,9\}+\{0,3,4,7\}\epsilon$ $\{1,4,7,10\}+\{0,3,7,8\}\epsilon$ $\{2,5,8,11\}+\{3,4,7,8\}\epsilon$ $e^{8\epsilon}(5+8\epsilon)$ $e^{8\epsilon}(5+4\epsilon)$ : Contrapuntal symmetries and admitted successors for the cantus firmus $0$. We obtain the admitted successors of $x+k\epsilon$ by adding $x$ to the cantus firmus of the results.[]{data-label="tab:adsu"} Five musical examples from Fux’s Gradus ad Parnassum {#music} ---------------------------------------------------- Under the results of Table \[tab:adsu\], in this section we analyse the five examples of first-species counterpoint provided in [@Fux Chapter one] in which cantus firmus is the lower voice.[^21] In these examples, all modes are essentially the diatonic scale set, $X=\{0,2,4,5,7,9,11\}$, with the tonics $2$ (first example), $4$ (second example), $5$ (third example), $7$ (fourth example), and $9$ (fifth example). These tonics correspond to the notes d (Dorian mode), e (Phrygian), f (Lydian), g (Mixolydian), and a (Eolian), respectively. However, these modes can suffer alterations related to musical considerations. For instance, according to [@Fux p. 28], in the modes $X_2$ (Dorian), $X_7$ (Mixolydian), and $X_9$ (Eolian), the seventh degrees must be raised, especially in the last but one measure. The prohibition of the interval $6$ (tritone) in the melodies is also accepted in this context. It is important to stress that these musical features are not prerequisites of the model, and are congruent with its results. A direct verification shows that all successors in Examples 1 to 4, corresponding to Figures 5, 11, 13 and 15 in [@Fux] are predicted by the model. In Example 4, the model also predicts the *battuta* octave correction. Only the fifth example, corresponding to Figure 23 in [@Fux], has a successor that is not predicted by the model. It occurs in the bars 5-6. The interval $4\epsilon$ (major third) goes to $4+8\epsilon$ (minor sixth), but no contrapuntal symmetry for $4\epsilon$ predicts $4+8\epsilon$ as a successor. A non-commutative counterpoint {#handnc} ============================== Consider the same data of Example \[noncomex\]. There, we have the strong dichotomy $(K,D,e^I)$ of the ring $R$ of upper triangular matrices on $\mathbb{Z}_2$, with $K=\{\mathbf{0},A_1,A_2,A_3\}$ and $D=\{I,B_1,B_2,B_3\}$, where $B_i=A_i+I$ for $i=1,2,3$. In this case, $b=I$ and $a=I$. Condition 2 {#secondcondnc} ----------- We solve the equation $$v+I=c+v,$$ so $c=I$, and $v$ and $d$ range over $R$. Condition 1 {#firstcondnc} ----------- Now, among the previous solutions, we take, for each consonance $k\epsilon$, those satisfying $$v\in k-D=k+D.$$ The following table contains the results. $k$ $\mathbf{0}$ $A_1$ $A_2$ $A_3$ ------- -------------- --------------------- --------------------- --------------------- -- -- $k+D$ $D$ $\{B_1,I,A_3,A_2\}$ $\{B_2,A_3,I,A_1\}$ $\{B_3,A_2,A_1,I\}$ Maximization {#countsymnc} ------------ Since $c=1$, and $1$ commutes with all elements of $R$, the maximization criterion (Theorem \[max\]) remains valid in this case, and we use it. We start by computing $\operatorname{Im}(r\cdot (-))$, $\rho$, $K/\operatorname{Im}(r\cdot (-))$, and $(|K_{\gamma}|_{\gamma})$ in the following table. The ordering on $\gamma$ is that written for the classes in $K/\operatorname{Im}(r\cdot (-))$. As before, the maximization criterion is not useful for the case $d=\mathbf{0}$ since the identity does not satisfy the condition 1. $r$ $rR$ $\rho$ $ K/\operatorname{Im}(r\cdot (-))$ $(|K_{\gamma}|_{\gamma})$ --------------------- ------------------------------ -------- ---------------------------------------------------------------------- --------------------------- $A_1,A_2$ $\{\mathbf{0},A_1,A_2,B_3\}$ 2 $\{A_1R,A_1R+A_3\}$ $(3,1)$ $A_3,I$ $R$ 1 $\{R\}$ $(4)$ $B_i$ for $i=1,2,3$ $\{\mathbf{0},B_i\}$ 4 $\{\{\mathbf{0},B_i\}, \{A_i,I\},\{A_1,A_2,A_3\}\setminus \{A_i\}\}$ $(1,1,2,0)$ We have the following maximum values with the restriction $c=1$. In the case $\rho=2$, the unique $[e^v]$ inducing the identity rearrangement of $(3,1)$ is the identity modulo $A_1$. In the case $d=B_1$, if $[e^v]$ induces an identity rearrangement of $(1,1,3,0)$, then necessarily $[e^v]$ leaves invariant $K_{\{A_2,A_3\}}$, which is equal to $\{A_2,A_3\}$. The possibilities are either $v=\mathbf{0}$ or $v=B_1$, that is, $e^v$ is the identity modulo $B_1$. The cases $d=B_2$ and $d=B_3$ are similar. $d$ $\rho$ maximum $\sum\limits_{r\in R}|(K+v+dr)\cap K|$ $v$ ----------- -------- ------------------------------------------------ -------------------------- $A_1,A_2$ 2 20 $\mathbf{0},A_1,A_2,B_3$ $A_3,I$ 1 16 any $B_1$ 4 24 $\mathbf{0},B_1$ $B_2$ 4 24 $\mathbf{0},B_2$ $B_3$ 4 24 $\mathbf{0},B_3$ Now we must compute the cardinals for $d=0$, that is, those of the form $8|(K+v)\cap K|$ with $v\neq \mathbf{0}$ by the table in Section \[firstcondnc\]. We already have $K+I+k=D+k$ for $k\in K$, by the same table. Moreover, $K+A_i=\{\mathbf{0},A_i\}\cup (\{B_1,B_2,B_3\}\setminus \{B_i\})$. Thus, $8|(K+v)\cap K|=0$ if $v=I$ and $8|(K+v)\cap K|= 16$ if $v\notin \{id, \mathbf{0}\}$, so the maximum cardinals (subject to conditions 1 and 2) occur for the symmetries with maximum value $24$ in the preceding table. The following table shows the contrapuntal symmetries directly obtained. $k$ $h$ ------------------------ ------------------------------------------------- -- -- $\mathbf{0}$ $ e^{B_i\epsilon}(I+B_i\epsilon)$ for $i=1,2,3$ $A_i$ (with $i=1,2,3$) $e^{B_i\epsilon}(I+B_i\epsilon)$ Admitted successors {#secsucnc} ------------------- The direct use of Equation  for each contrapuntal symmetry $h$ obtained in Section \[countsymnc\] produces the following table. In this example, the maximization criterion is useful to determine all contrapuntal symmetries. $k$ $h$ admitted successors of $k\epsilon$ -------------- --------------------------------------------------- ------------------------------------------------------------------------- -- $\mathbf{0}$ $e^{B_i\epsilon}(I+B_i\epsilon)$, $1\leq i\leq 3$ see $k=A_i$ $\{A_3,I,B_1,B_2\}+K\epsilon$ $\{\mathbf{0},A_1,A_2,B_3\}+(\{A_1,A_2,A_3\}\setminus \{A_i\})\epsilon$ Note that some parallelisms of the consonance $A_i\epsilon$ are forbidden in this notion of counterpoint. The counterpoint notion that results for multiplication on the right is very similar.[^22] Final reflections on Mazzola’s model {#final} ==================================== The validity of the Little theorem of counterpoint for noncommutative rings stresses several important features of Mazzola’s model. First, the little theorem shows that the model is nontrivial in the noncommutative case (at least $|K|^2$ admitted successors always), so it is worth experimenting with new examples of counterpoint; for instance, we can perform counterpoint in noncommutative rings of square matrices. Second, the possibility of generalization also stresses a certain profundity and intrinsic generality in the little theorem. The present noncommutative generalization suggests a structural richness beyond algorithmic computations, which is hinted at by the counting formulas and the maximization criterion. Moreover, these two tools allow us to make a reasonable calculation of contrapuntal symmetries *by hand*—a task closer to craftsmanship of counterpoint. A lot of combinatorics is to be studied regarding the revelation of profound algebraic structures behind counterpoint. Other strengths of Mazzola’s model are the following. 1\. In contrast to psychological or descriptive approaches, it has a very simple but powerful idea as philosophy: counterpoint rules obey profound relations based on musical symmetries. 2\. Its astonishing agreement with Palestrina’s Missa Papae Marcellus, as studied in [@Nieto]. 3\. It predicts the prohibition of parallel fifths, among other rules (Section \[music\]), as a consequence of the partition of the twelve intervals into consonances and dissonances. 4\. It connects contrapuntal tension/resolution forces with fundamental nature forces, via a crucial idea in contemporary physics, namely local symmetries, which are formalized thanks to differential geometry, in particular, tangent bundles. Mazzola’s analogous algebraic language is based on dual numbers. 5\. It paves the way to many other forms of counterpoint, which opens up a lot of fields of musical experimentation. However, the previous mathematical discussions suggest the following weaknesses of the model. 1\. The model is not natural. Though the mathematical justification is interesting, it is not entirely clear why we use the structure of the dual numbers ring to model intervals and not another structure,[^23] even if the product in the dual numbers ring has a musical interpretation [@Octavio Section 2.2.1], though *incidentally*. Mazzola’s musical justification of dual numbers was discussed in Section \[sec3\]. Moreover, a mathematical model of counterpoint should clarify and simplify the classical theory in music treatises. We would expect a teachable model, which stresses the essential ideas behind counterpoint so that a consistent theory can be established, in contrast to the usual pedagogical model based on unjustified rules. 2\. The use of local symmetries is unclear. In Mazzola’s works, there is no formal definition of global and local symmetry. In this paper, we tried to provide such a definition based on Mazzola’s hints. However, we observed that local and global symmetries give further explanations on the nature of the tension between an interval and its successors but do not play an essential role in Definition \[def\]. For example, the contrapuntal symmetry $e^07$, for $7\epsilon$, obtained in Section \[secsuc\] is not local (definition in Section \[sec5.2\]) but it is responsible for the parallel fifths prohibition. This suggests a simplification of the condition 2, the latter being related to local/global symmetries according to \[sec5.2\]. For the same reason, we could question the use of dual numbers as a bundle analogue, since this analogy has a clear relation to local/global symmetries. 3\. It only deals with first-species counterpoint and a convincing theory of the remaining species, *and the cases of three or more voices*, are not at hand. Based on the previous remarks, we suggest a *simplification* of the first-species counterpoint model, sharing the strengths of Mazzola’s one and correcting their weaknesses. Appendix {#ap .unnumbered} ======== \[r.i.\] Suppose given an ordered list of real numbers $$a_1\geq a_2\geq \dots \geq a_n.$$ The inequality $$\sum\limits_{i=1}^n a_i^2\geq \sum\limits_{i=1}^n a_ia_{\sigma(i)}$$ holds for each permutation $\sigma$ of $\{1,\dots, n\}$. Moreover, the equality holds if and only if $(a_1,\dots,a_n)=(a_{\sigma(1)},\dots, a_{\sigma(n)})$. Induction on $n$. The case $n=1$ is just the equality $a_1^2=a_1^2$. Now suppose the result for $n-1$. We consider two cases. *First case*: $a_{\sigma(1)}=a_1$. We can assume $\sigma(1)=1$ (switch $\sigma(1)$ and $1$) since this do not alter the sum $\sum\limits_{i=1}^n a_ia_{\sigma(i)}$. In this case, $$\sum\limits_{i=1}^n a_ia_{\sigma(i)}=a_1^2+\sum\limits_{i=2}^n a_ia_{\sigma(i)}=a_1^2+\sum\limits_{i=1}^{n-1} a'_ia'_{\sigma'(i)},$$ where $a'_i=a_{i+1}$ and $\sigma'(i)=\sigma(i+1)-1$ for each $i=1, \dots , n-1$. Thus, by applying the induction hypothesis, $$\sum\limits_{i=2}^{n} a_ia_{\sigma(i)}\leq \sum\limits_{i=2}^{n} a_i^2$$ and the equality holds if and only if $(a_2,\dots ,a_n)=(a_{\sigma(2)},\dots, a_{\sigma(n)})$. Hence, by adding $a_1^2$ on both sides, $$\sum\limits_{i=1}^{n} a_ia_{\sigma(i)}\leq \sum\limits_{i=1}^{n} a_i^2$$ and the equality holds if and only if $(a_1,\dots ,a_n)=(a_{\sigma(1)},\dots, a_{\sigma(n)})$. *Second case*: $a_{\sigma(1)}\neq a_1$. Note that there is $k$ with $a_{\sigma(k)}=a_1$ such that $a_k< a_1$; otherwise $\{k\ |\ a_{\sigma(k)}=a_1\}\subseteq \{k\ |\ a_k=a_1\}$ and, since these sets have the same finite cardinality, they are equal, which is a contradiction since $1$ is in the latter but not in the former. Thus, $$\sum\limits_{i=1}^n a_ia_{\sigma(i)}=a_1a_{\sigma(1)}+a_k a_{\sigma(k)}+\sum\limits_{i\in\{1,\dots, n\}\setminus \{1,k\}} a_ia_{\sigma(i)}$$ and, if $\sigma'$ is obtained from $\sigma$ by switching $\sigma(1)$ and $\sigma(k)$, then $$\sum\limits_{i=1}^n a_ia_{\sigma'(i)}-\sum\limits_{i=1}^n a_ia_{\sigma(i)}=(a_1-a_k)(a_{\sigma(k)}-a_{\sigma(1)})=(a_1-a_k)(a_1-a_{\sigma(1)})>0.$$ This means that $$\sum\limits_{i=1}^n a_ia_{\sigma(i)}<\sum\limits_{i=1}^n a_ia_{\sigma'(i)}\leq \sum\limits_{i=1}^n a_i^2,$$ where the right-hand inequality corresponds to the first case. Finally, note that here the inequality is always strict and $(a_1,\dots ,a_n)\neq(a_{\sigma(1)},\dots, a_{\sigma(n)})$. The following is a short reflection about the problem of deciding when a dichotomy $\{K,D\}$ of a ring $R$ is not self-complementary. Let $X\subseteq R$. We associate with $X$ a categorical **groupoid of intervals** $\mathcal{G}(X)$ as follows. Its set of objects is $X$. For each $x,y\in X$, there is a unique morphism from $x$ to $y$, namely the triple $(x,y,y-x)$. We define the composition by $$(y,z,z-y)\circ (x,y,y-x)=(x,z,z-y+y-x)=(x,z,z-x).$$ The identities are of the form $(x,x,0)$. The inverse of $(x,y,y-x)$ is $(y,x,x-y)$. Given a self-complementary dichotomy $(K,D,e^ab)$ of $R$, we have an induced groupoid isomorphism $F:\mathcal{G}(K)\longrightarrow \mathcal{G}(D)$ sending[^24] $(x,y,y-x)$ to $(e^ab(x),e^ab(y),b(y-x))$, which acts linearly. Hence, *a self-complementary dichotomy induces a linear correspondence between the intervals of $K$ and $D$*. This fact can be used to prove that the dichotomy $\{K,D\}$ with $K=\{0,1,3,6,8,11\}$ and $D=\{2,4,5,7,9,11\}$ is not self-complementary. In fact, the sequence $(11,0,1)$ of elements of $K$ has the maximum length among sequences $(x_1,\dots ,x_n)$ such that $x_{i+1}-x_i=1$ for $i=0,\dots n-1$. If a symmetry $e^ab$ sends $K$ to $D$, then the groupoid isomorphism $F$ sends $(11,0,1)$ to a sequence $(d_1,d_2,d_3)$ in $D$ with maximum length among sequences $(y_1,\dots ,y_n)$ such that $y_{i+1}-y_{i}=b$. But the sequences of maximum length with $y_{i+1}-y_{i}=b$ are $(4,5)$ (for $b=1$), $(5,4)$ (for $b=11$), $(11,4,9,2,7)$ (for $b=5$), and $(7,2,9,4,11)$ (for $b=7$), whose length is not $3$; a contradiction. Funding {#funding .unnumbered} ======= This work was supported by Programa de Becas Posdoctorales en la UNAM 2019, which is coordinated by Dirección de Asuntos del Personal Académico (DGAPA) at Universidad Nacional Autónoma de México. Disclosure statement {#disclosure-statement .unnumbered} ==================== No potential conflict of interest was reported by the authors. [^1]: [jsariasv1@gmail.com](jsariasv1@gmail.com) [^2]: [octavioalberto@mixteco.utm.mx](octavioalberto@mixteco.utm.mx) [^3]: [lluisp@unam.mx](lluisp@unam.mx) [^4]: The ring $\mathbb{Z}_{12}$ models the twelve tones of the chromatic scale and, at the same time, the twelve intervals between these tones. The distinction is usually clear from the context. [^5]: First, suppose that $e^b g:N\longrightarrow M$ is the inverse of $e^a f$. Then $(e^af)(e^bg)=e^{f(b)+a}fg=id_N$ and $(e^bg)(e^a f)=e^{g(a)+b}gf=id_M$. Therefore, $fg=id_N$, $gf=id_M$, and $b=-g(a)$. Conversely, if $fg=id_N$ and $gf=id_M$, then $(e^af)(e^{-g(a)}g)=id_N$ and $(e^{-g(a)}g)(e^af)=id_M$. [^6]: We discuss the choice of dual numbers to model this situation in Section \[sec3\]. [^7]: A precise statement of this naturality claim is Proposition \[unique0\]. [^8]: A musical introduction to ecclesiastical modes is [@Jep pp. 59-82]. We consider modes from a more mathematical perspective in Definition \[def\]. [^9]: This definition is essentially the same of an **autocomplementary** dichotomy in [@MazzolaTopos Definition 92]. [^10]: This conceptual argument can be related to the groupoid of intervals of $K$, briefly introduced in Appendix. [^11]: Thanks to the hypothesis $k\geq 3$, the six distinct elements $0$, $1$, $3$, $-1$, $2$, and $4$ are in $\mathbb{Z}_{2k}$. [^12]: See [@Lam p. 3] for details. [^13]: In this representation we make use of the fact that $R$ can be regarded as a subring of $R[x]/\left\langle x^2\right\rangle$. [^14]: More specifically, in the continental theory we would consider the polynomial $a+\epsilon b+\epsilon^{2}c$, where $a$ is the cantus firmus and $b$ and $c$ the intervals between the cantus firmus and the discantus and between the discantus and the countertenor, respectively. On the other hand, we would have $a+\epsilon b$ and $a+\epsilon(b+c)$ in the English theory. However, we could disagree with this argument because it is not clear why we take the algebraically dependent variables $\epsilon$ and $\epsilon^2$ to separate two, in principle, independent intervals. Section \[locsec\] provides a more convincing argument for using dual numbers. They allow a certain analogy with tangent bundles and their local symmetries, the latter being responsible for nature forces, and, in particular, tension/resolution forces in counterpoint, according to Mazzola’s ideas. [^15]: There is a main reason for such an analogy of $R$ with a variety: the Abelian group $\mathbb{Z}_{12}$ can be regarded as a combinatorial torus via the decomposition $\mathbb{Z}_{12}\cong \mathbb{Z}_{4}\times \mathbb{Z}_3$. See [@MazzolaTopos; @Octavio] for details on the *torus of thirds*. [^16]: This is a subtle point: there can be non-identity symmetries $g$ of $R[\epsilon]$ such that $p^z[\epsilon]$ is global. Details in Section \[sec5.2\]. [^17]: In fact, this set is a group since it is an stabilizer one. [^18]: In fact $\rho$, which is by definition $|\operatorname{Ker}(d\cdot\underline{\ \ })|$, is equal to $|R|$ if and only if $d=0$–but we are assuming $d\neq 0$. [^19]: A rearrangement of a function (sequence) $f:S\longrightarrow X$ is a function of the form $f\circ \sigma$, where $\sigma$ is a permutation of $S$. [^20]: **Definition**: we say that $x\equiv y\pmod{d}$ if $[x]=[y]$ in $R/\operatorname{Im}(d\cdot\underline{\ \ })$. [^21]: We will omit the remaining cases, which are related with the problem of *hanging counterpoint* [@Octavio Section 2.2.1], in our exposition of the essential features of the model. [^22]: To be more exact, this notion is obtained by applying the previous theory to the opposite ring $R^{op}$ instead of $R$. In that case, all right $R$-modules considered become left $R$-modules. [^23]: Dual numbers are essentially pairs $(c,d)$ (cantus and interval with the discantus) of elements of $\mathbb{Z}_{12}$. A more natural structure on these pairs is that induced by the usual addition and product of tones in $\mathbb{Z}_{12}$, that is, the ring structure. Curiously, the induced structure is that of the quotient ring $\mathbb{Z}_{12}[x]/\left\langle x^2-x \right\rangle$, which is similar to the dual numbers one but has a slightly different product. [^24]: It is the unique possible definition if the correspondence on objects is $e^ab$.

ArXiv

--- abstract: 'We prove a structure theorem for triangulated Calabi-Yau categories: An algebraic $2$-Calabi-Yau triangulated category over an algebraically closed field is a cluster category iff it contains a cluster tilting subcategory whose quiver has no oriented cycles. We prove a similar characterization for higher cluster categories. As an application to commutative algebra, we show that the stable category of maximal Cohen-Macaulay modules over a certain isolated singularity of dimension three is a cluster category. This entails the classification of the rigid Cohen-Macaulay modules first obtained by Iyama-Yoshino. As an application to the combinatorics of quiver mutation, we prove the non-acyclicity of the quivers of endomorphism algebras of cluster-tilting objects in the stable categories of representation-infinite preprojective algebras. No direct combinatorial proof is known as yet. In the appendix, Michel Van den Bergh gives an alternative proof of the main theorem by appealing to the universal property of the triangulated orbit category.' address: - 'UFR de Mathématiques, Institut de Mathématiques de Jussieu, UMR 7586 du CNRS, Case 7012, Université Paris 7, 2 place Jussieu, 75251 Paris Cedex 05, France' - 'Institutt for matematiske fag, Norges Teknisk-naturvitenskapelige universitet, N-7491, Trondheim, Norway' author: - Bernhard Keller - | Idun Reiten\ \ with an appendix by Michel Van den Bergh title: 'Acyclic Calabi-Yau categories\' --- [^1] Introduction ============ Cluster algebras were introduced and studied by Fomin-Zelevinsky and Berenstein-Fomin-Zelevinsky in a series of articles [@FominZelevinsky02] [@FominZelevinsky03] [@BerensteinFominZelevinsky05] [@FominZelevinsky07]. It was the discovery of Marsh-Reineke-Zelevinsky [@MarshReinekeZelevinsky03] that they are closely connected to quiver representations. This link is similar to the one between quantum groups and quiver representations discovered by Ringel [@Ringel90] and investigated by Kashiwara, Lusztig, Nakajima and many others. The link between cluster algebras and quiver representations becomes especially beautiful if, instead of categories of quiver representations, one considers certain triangulated categories deduced from them: the so-called [*cluster categories*]{}. These were introduced in [@BuanMarshReinekeReitenTodorov04] and, for Dynkin quivers of type $A_n$, in [@CalderoChapotonSchiffler04]. If $k$ is a field and $Q$ a quiver without oriented cycles, the associated cluster category ${{\mathcal C}}_Q$ is the ‘largest’ $2$-Calabi-Yau category under the derived category of representations of $Q$ over $k$. It was shown [@BuanMarshReiten06] [@BuanMarshReitenTodorov05] [@CalderoChapoton06] [@CalderoKeller05a] [@CalderoKeller05b] that this category fully determines the combinatorics of the cluster algebra associated with $Q$ and carries considerably more information. This was used to prove significant new results on cluster algebras, [[*cf.*]{} ]{}[[*e.g.*]{} ]{}[@BuanReiten06] [@CalderoReineke06]. We refer to [@FominZelevinsky03a] [@Zelevinsky04] for more background on cluster algebras and to [@AssemBruestleSchifflerTodorov05] [@AssemBruestleSchiffler06] [@BuanMarshReiten04b] [@BuanReiten05b] [@BuanReitenSeven06] [@BaurMarsh06] [@GeissLeclercSchroeer06] [@GeissLeclercSchroeer05c] [@Iyama05a] [@IyamaReiten06] [@Ringel07] [@Thomas06] [@Zhu06] for recent developments in the study of their links with representations of quivers and finite-dimensional algebras. The question arises as to whether, for a given quiver $Q$ without oriented cycles, the cluster category is the ‘unique model’ of the associated cluster algebra. In other words, if we view the cluster algebra as a combinatorial invariant associated with the cluster category, is the category determined by this invariant ? In this paper, we show that surprisingly, this question has a positive answer. Namely, we prove that if $k$ is an algebraically closed field and ${{\mathcal C}}$ an algebraic $2$-Calabi-Yau category containing a cluster tilting object $T$ whose endomorphism algebra has a quiver $Q$ without oriented cycles, then ${{\mathcal C}}$ is triangle equivalent to the cluster category ${{\mathcal C}}_Q$. Notice that this result is ‘of Morita type’, but much stronger than typical Morita theorems, since we only need to know the [*quiver*]{} of the endomorphism algebra, not the algebra itself. We give several applications: First, we show on an example that cluster categories naturally appear as stable categories of Cohen-Macaulay modules over certain singularities. This yields an alternative proof of Iyama-Yoshino’s [@IyamaYoshino06] classification of rigid Cohen-Macaulay modules over a certain isolated singularity. More examples may be obtained from [@IyamaReiten06] and [@IyamaYoshino06], [[*cf.*]{} ]{}[@BurbanIyamaKellerReiten07]. Secondly, we show that the quivers associated in [@GeissLeclercSchroeer05b] with representation-infinite finite-dimensional preprojective algebras are not mutation-equivalent to quivers without oriented cycles. This last result was obtained independently by C. Geiss [@Geiss06]. It has been used in [@DerksenWeymanZelevinsky07], Example 8.7, to show that the class of rigid quivers with potential is strictly greater than the class of quivers with potential mutation-equivalent to acyclic ones. An application to the realization of cluster categories as stable categories of Frobenius categories with finite-dimensional morphism spaces is given in [@BuanIyamaReitenScott07] and [@GeissLeclercSchroeer06a]. In [@HolmJoergensen06] and [@HolmJoergensen06a], the authors use our results to determine which stable categories of representation-finite selfinjective algebras of type $A$ and $D$ are higher cluster categories. More generally, in [@Amiot07], the author obtains a classification of ‘most’ triangulated categories with finitely many indecomposables by methods similar to ours. The main difficulty in the proof is the construction of a triangle functor between the cluster category and the given Calabi-Yau category. Our construction is based on the description [@Keller05] of the cluster category as a stable derived category of a certain differential graded category. This approach leads to interesting connections between Calabi-Yau categories of dimensions $2$ and $3$, which have been further investigated in [@Tabuada07]: It turns out that each algebraic Calabi-Yau category of dimension $2$ containing a cluster-tilting subcategory is equivalent to a stable derived category of a differential graded category whose perfect derived category is Calabi-Yau of dimension $3$. A more direct approach, based on the universal property of the cluster category [@Keller05], has been discovered by Michel Van den Bergh, who has kindly accepted to include his proof as an appendix to this article. It turns out that the main theorem and its proofs can be generalized almost without effort to Calabi-Yau categories of any dimension $d\geq 2$. However, one has to take into account that in the $d$-cluster category, the selfextensions of the canonical cluster tilting object vanish in degrees $-d+2, \ldots, -1$. This condition therefore has to be added to the hypotheses of the generalized main theorem. Acknowledgments {#acknowledgments .unnumbered} --------------- This research started during a stay of the first-named author at the Norwegian University of Science and Technology (NTNU). He thanks the second-named author and the members of her group at the NTNU for their warm hospitality. Both authors thank Michel Van den Bergh for pointing out gaps and detours in the original proof and for agreeing to include his own proof as an appendix to this article. They are grateful to Carl Fredrik Berg for pointing out reference [@Benson98]. The main theorem and two applications ===================================== Statement {#subsection:statement} --------- Let $k$ be a perfect field. Let ${{\mathcal E}}$ be a $k$-linear Frobenius category with split idempotents. Suppose that its stable category ${{\mathcal C}}={\underline{{{\mathcal E}}}}$ has finite-dimensional ${{\operatorname{\mathsf{Hom}}}}$-spaces and is Calabi-Yau of CY-dimension $2$, [[*i.e.*]{} ]{}we have bifunctorial isomorphisms $$D{{\mathcal C}}(X,Y) {\stackrel{_\sim}{\rightarrow}}{{\mathcal C}}(Y,S^2X) {\: , \;}X,Y\in{{\mathcal C}}{\: , \;}$$ where $D$ is the duality functor ${{\operatorname{\mathsf{Hom}}}}_k(?,k)$ and $S$ the suspension of ${{\mathcal C}}$. Let ${{\mathcal T}}\subset{{\mathcal C}}$ be a cluster tilting subcategory. Recall from [@KellerReiten07] that this means that ${{\mathcal T}}$ is a $k$-linear subcategory which is functorially finite in ${{\mathcal C}}$ and such that an object $X$ of ${{\mathcal C}}$ belongs to ${{\mathcal T}}$ iff we have ${{\operatorname{\mathsf{Ext}}}}^1(T,X)=0$ for all objects $T$ of ${{\mathcal T}}$. As shown in [@KellerReiten07], the category ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ of finitely presented ${{\mathcal T}}$-modules is then abelian. If it is hereditary, the cluster category ${{\mathcal C}}_{{\mathcal T}}$, as defined in [@BuanMarshReinekeReitenTodorov04], is the orbit category of the bounded derived category ${{\mathcal D}}^b({{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}})$ under the action of the autoequivalence $S^2 \circ \Sigma^{-1}$ where $S$ is the suspension and $\Sigma$ the Serre functor of ${{\mathcal D}}^b({{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}})$. If ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ is hereditary, then ${{\mathcal C}}$ is triangle equivalent to the cluster category ${{\mathcal C}}_{{\mathcal T}}$. We will prove the theorem in section \[section:proofs\] below. Now assume that $k$ is algebraically closed. Let ${{\mathcal R}}$ be the radical of ${{\mathcal T}}$, [[*i.e.*]{} ]{}the ideal such that for two indecomposables $X,Y$, the space ${{\mathcal R}}(X,Y)$ is formed by the non isomorphisms from $X$ to $Y$. Let $Q$ be the quiver of ${{\mathcal T}}$: Its vertices are the isomorphism classes of indecomposables of ${{\mathcal T}}$ and the number of arrows from the class of an indecomposable $X$ to an indecomposable $Y$ is the dimension of the vector space ${{\mathcal R}}(X,Y)/{{\mathcal R}}^2(X,Y)$. If $k$ is algebraically closed and for each vertex $x$ of $Q$, only finitely many paths start in $x$ and only finitely many paths end in $x$, then ${{\mathcal C}}$ is triangle equivalent to the cluster category ${{\mathcal C}}_Q$. Note that under the assumptions of the corollary, the projective (right) $kQ$-module $kQ(?,x)$ and the injective $kQ$-module $D kQ(x,?)$ are of finite total dimension and that the category ${{\operatorname{\mathsf{mod}}}\nolimits}kQ$ of finitely presented $kQ$-modules coincides with the category of modules of finite total dimension. We will prove the corollary in section \[section:proofs\] below. Application: Cohen-Macaulay modules {#subsection:CM-modules} ----------------------------------- Suppose that $k$ is algebraically closed of characteristic $0$. Let the cyclic group $G={\mathbb{Z}}/3{\mathbb{Z}}$ act on the power series ring $S=k[[X,Y,Z]]$ such that a generator of $G$ multiplies each indeterminate by the same primitive third root of unity. Then the fixed point ring $R=S^G$ is a Gorenstein ring, [[*cf.*]{} ]{}[[*e.g.*]{} ]{}[@Watanabe74], and an isolated singularity of dimension $3$, [[*cf.*]{} ]{}[[*e.g.*]{} ]{}Corollary 8.2 of [@IyamaYoshino06]. The category ${{\operatorname{\mathsf{CM}}}}(R)$ of maximal Cohen-Macaulay modules is an exact Frobenius category. By Auslander’s results [@Auslander76], [[*cf.*]{} ]{}Lemma 3.10 of [@Yoshino90], its stable category ${{\mathcal C}}={\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$ is $2$-Calabi Yau. By work of Iyama [@Iyama05], the module $T=S$ is a cluster-tilting object in ${{\mathcal C}}$. The endomorphism ring of $T$ over $R$ is the skew group ring $S*G$. Under the action of $G$, the module $T$ decomposes into three indecomposable direct factors $T=T_1\oplus T_2\oplus S^G$ and we see that its endomorphism ring $S*G$ is isomorphic to the completed path algebra of the quiver $$\xymatrix{ & S^G \ar[rd] \ar@<1ex>[rd] \ar@<-1ex>[rd] & \\ T_1 \ar[ru] \ar@<1ex>[ru] \ar@<-1ex>[ru] & & T_2 \ar[ll] \ar@<1ex>[ll] \ar@<2ex>[ll] }$$ subject to all the ‘commutativity relations’ obtained by labelling the three arrows between any consecutive vertices by $X$, $Y$ and $Z$. The stable endomorphism ring of $T$ is thus isomorphic to the path algebra of the generalized Kronecker quiver $$Q : \xymatrix{ 1 \ar[r] \ar@<1ex>[r] \ar@<-1ex>[r] & 2}.$$ The theorem now shows that the stable category of Cohen-Macaulay modules ${\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$ is triangle equivalent to the cluster category ${{\mathcal C}}_Q$. As a further application, we give an alternative proof of a theorem from [@IyamaYoshino06], stating that the indecomposable nonprojective rigid modules in ${{\operatorname{\mathsf{CM}}}}(R)$ are exactly the modules of the form $\Omega^i(T_1)$ and $\Omega^i(T_2)$ for $i\in{\mathbb{Z}}$. For this, note that the indecomposable rigid objects in ${{\mathcal C}}_Q$ are exactly the images of the indeomposable rigid $kQ$-modules and the $SP$, for $P$ indecomposable projective $kQ$-module [@BuanMarshReinekeReitenTodorov04]. So they correspond to the vertices of the component of the AR-quiver of ${{\mathcal C}}_Q$ containing the indecomposable projective $kQ$-modules. The corresponding component of the AR-quiver of ${\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$ is the one containing $T_1$ and $T_2$. Hence the indecomposable rigid objects in ${\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$ are all $\tau$-shifts of these. Finally, we use that $\tau=\Omega^{-1}$ in this case [@Auslander76]. Note that this application does not need the full force of the main theorem. For we only use that the AR-quivers of ${{\mathcal C}}_Q$ and ${\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$ are isomorphic, with the component of the projective $kQ$-modules for ${{\mathcal C}}_Q$ corresponding to the component of $T_1$ and $T_2$ for ${\underline{{{\operatorname{\mathsf{CM}}}}}}(R)$, and it is easy to see that this follows from proposition 2.1 c) and lemma 3.5 of [@KellerReiten07], [[*cf.*]{} ]{}also [@BuanMarshReiten04]. Application: Non acyclicity --------------------------- Let $k$ be an algebraically closed field and $\Lambda$ the preprojective algebra of a simply laced Dynkin diagram $\Delta$. Then $\Lambda$ is a finite-dimensional selfinjective algebra and the stable category ${{\mathcal C}}$ of finite-dimensional $\Lambda$-modules is $2$-Calabi-Yau, [[*cf.*]{} ]{}[@CrawleyBoevey00], and admits a canonical cluster-tilting subcategory ${{\mathcal T}}'$ with finitely many indecomposables, [[*cf.*]{} ]{}[@GeissLeclercSchroeer05b]. Let $Q'$ be its quiver. For example, by \[loc. cit.\], the quivers $Q'$ corresponding to $\Delta=A_5$ and $\Delta=D_4$ are respectively $$\begin{xy} 0;<0.7pt,0pt>:<0pt,-0.63pt>:: (75,0) *+{\circ} ="0", (50,50) *+{\circ} ="1", (100,50) *+{\circ} ="2", (25,100) *+{\circ} ="3", (75,100) *+{\circ} ="4", (125,100) *+{\circ} ="5", (0,150) *+{\circ} ="6", (50,150) *+{\circ} ="7", (100,150) *+{\circ} ="8", (150,150) *+{\circ} ="9", "1", {\ar"0"}, "0", {\ar"2"}, "2", {\ar"1"}, "3", {\ar"1"}, "1", {\ar"4"}, "4", {\ar"2"}, "2", {\ar"5"}, "4", {\ar"3"}, "6", {\ar"3"}, "3", {\ar"7"}, "5", {\ar"4"}, "7", {\ar"4"}, "4", {\ar"8"}, "8", {\ar"5"}, "5", {\ar"9"}, "7", {\ar"6"}, "8", {\ar"7"}, "9", {\ar"8"}, \end{xy} \quad\quad \begin{xy} 0;<0.7pt,0pt>:<0pt,-0.63pt>:: (0,75) *+{\circ} ="0", (50,0) *+{\circ} ="1", (50,100) *+{\circ} ="2", (50,150) *+{\circ} ="3", (100,75) *+{\circ} ="4", (150,0) *+{\circ} ="5", (150,100) *+{\circ} ="6", (150,150) *+{\circ} ="7", "0", {\ar"1"}, "0", {\ar"2"}, "0", {\ar"3"}, "4", {\ar"0"}, "1", {\ar"4"}, "5", {\ar"1"}, "2", {\ar"4"}, "6", {\ar"2"}, "3", {\ar"4"}, "7", {\ar"3"}, "4", {\ar"5"}, "4", {\ar"6"}, "4", {\ar"7"}, \end{xy}$$ Part b) of the following proposition was obtained independently by C. Geiss [@Geiss06]. Suppose that $\Lambda$ is representation-infinite. - The stable category ${{\mathcal C}}={\underline{{{\operatorname{\mathsf{mod}}}\nolimits}}} \Lambda$ is not equivalent to the cluster category ${{\mathcal C}}_Q$ of a finite quiver $Q$ without oriented cycles. - The quiver $Q'$ of the canonical cluster-tilting subcategory of [@GeissLeclercSchroeer05b] is not mutation-equivalent to a quiver $Q$ without oriented cycles. In particular, it follows that the two above quivers are not mutation-equivalent to quivers without oriented cycles. In the proof of the proposition, we use the main theorem. Let us stress that, as at the end of \[subsection:CM-modules\], we do not need its full force but only use the isomorphism of AR-quivers. This is the variant of the proof also given by C. Geiss [@Geiss06]. a\) Recall first that the AR-translation $\tau$ is isomorphic to the suspension in any $2$-Calabi-Yau category, so that it is preserved under triangle equivalences. We know from [@AuslanderReiten96] that the AR-translation $\tau$ of ${{\mathcal C}}$ is periodic of period dividing $6$. In particular, we have $\tau^6(X){\stackrel{_\sim}{\rightarrow}}X$ for each indecomposable $X$ of ${{\mathcal C}}$. But in ${{\mathcal C}}_Q$, for each indecomposable $X$ which is the image of a preprojective $kQ$-module, the iterated translates $\tau^{-n}(X)$, $n\geq 0$, are all pairwise non isomorphic since $Q$ is representation-infinite. b\) Suppose that $Q'$ is mutation-equivalent to a quiver $Q$. By one of the main results of [@GeissLeclercSchroeer07], it follows that ${{\mathcal C}}$ contains a cluster-tilting subcategory ${{\mathcal T}}$ whose quiver is $Q$. If $Q$ does not have oriented cycles, it follows from the main theorem that ${{\mathcal C}}$ is triangle equivalent to ${{\mathcal C}}_Q$ in contradiction to a). Proofs {#section:proofs} ====== Proof of the corollary ---------------------- First recall from [@KellerReiten07] that the category ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ of finitely presented ${{\mathcal T}}$-modules is abelian and Gorenstein of dimension at most $1$. It follows from our hypothesis that each object of ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ has a finite composition series all of whose subquotients are simple modules $$S_M = {{\mathcal T}}(?,M)/{{\mathcal R}}(?,M)$$ associated with indecomposables $M$ of ${{\mathcal T}}$ and that each of these simple modules is of finite projective dimension. Thus each object of ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ is of finite projective dimension so that ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ has to be hereditary. Since $k$ is algebraically closed, it follows that ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ is equivalent to ${{\operatorname{\mathsf{mod}}}\nolimits}kQ$ and the claim of the corollary follows from the theorem. Plan of the proof of the theorem {#subsection:plan} -------------------------------- Our aim is to construct a triangle equivalence ${{\mathcal C}}\to{{\mathcal C}}_{{\mathcal T}}$ such that the triangle $$\xymatrix{{{\mathcal C}}\ar[r] \ar[rd] & {{\mathcal C}}_{{\mathcal T}}\ar[d]\\ & {{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}}$$ becomes commutative, where the diagonal functor takes $X$ to ${{\mathcal C}}(?,X)|{{\mathcal T}}$. To construct the triangle equivalence ${{\mathcal C}}\to {{\mathcal C}}_{{\mathcal T}}$, we use the construction of ${{\mathcal C}}_{{\mathcal T}}$ given in [@Keller05], namely, the category ${{\mathcal C}}_{{\mathcal T}}$ is the stable derived category ${{\operatorname{\mathsf{stab}}}}({{\mathcal T}}\oplus D{{\mathcal T}}[-3])$ of the differential graded (=dg) category whose objects are the objects of ${{\mathcal T}}$ and whose morphism complexes are given by the graded modules $${{\mathcal T}}(x,y)\oplus (D{{\mathcal T}}(y,x))[-3]$$ endowed with the vanishing differential (the construction of the stable derived category is recalled in section \[subsection:proof\] below). Thus, we have to construct an equivalence $${{\mathcal C}}\to {{\operatorname{\mathsf{stab}}}}({{\mathcal T}}\oplus (D{{\mathcal T}})[-3]).$$ We proceed in three steps: 1) We construct a dg category ${{\mathcal A}}$ and a triangle functor $${{\mathcal C}}\to {{\operatorname{\mathsf{stab}}}}({{\mathcal A}}).$$ We show moreover that the subcategory of indecomposables of the homology $H^*{{\mathcal A}}$ is isomorphic to ${{\mathcal T}}\oplus (D{{\mathcal T}})[-3]$. 2\) Using the fact that $k$ is perfect we show that the dg category ${{\mathcal A}}$ is formal, [[*i.e.*]{} ]{}linked to its homology by a chain of quasi-isomorphisms. This yields the required triangle functor $${{\mathcal C}}\to{{\operatorname{\mathsf{stab}}}}({{\mathcal A}}) {\stackrel{_\sim}{\rightarrow}}{{\operatorname{\mathsf{stab}}}}(H^*{{\mathcal A}}) {\stackrel{_\sim}{\rightarrow}}{{\operatorname{\mathsf{stab}}}}({{\mathcal T}}\oplus (D{{\mathcal T}})[-3])= {{\mathcal C}}_{{\mathcal T}}.$$ 3\) In a final step, we show that the composed functor ${{\mathcal C}}\to{{\mathcal C}}_{{\mathcal T}}$ is fully faithful and that its image generates ${{\mathcal C}}_{{\mathcal T}}$. The proof {#subsection:proof} --------- Let ${{\mathcal M}}\subset{{\mathcal E}}$ be the preimage of ${{\mathcal T}}$ under the projection functor. In particular, ${{\mathcal M}}$ contains the subcategory ${{\mathcal P}}$ of the projective-injective objects in ${{\mathcal M}}$. Note that ${{\mathcal T}}$ equals the quotient ${\underline{{{\mathcal M}}}}$ of ${{\mathcal M}}$ by the ideal of morphisms factoring through a projective-injective. For each object $M$ of ${{\mathcal T}}$, choose an ${{\mathcal E}}$-acyclic complex $A_M$ of the form $$0 \to M_1 \to M_0 \to P \to M \to 0 {\: , \;}$$ where $P$ is ${{\mathcal E}}$-projective and $M_0, M_1$ are in ${{\mathcal M}}$, [[*cf.*]{} ]{}[@KellerReiten07]. Note that if $\Omega M$ denotes the kernel of $P\to M$, the induced morphism $M_0 \to \Omega M$ is automatically a right ${{\mathcal M}}$-approximation of $\Omega M$. Let ${{\mathcal A}}$ be the dg (=differential graded) subcategory of the dg category ${{\mathcal C}}({{\mathcal E}})_{dg}$ of complexes over ${{\mathcal E}}$ whose objects are these acyclic complexes. Thus, for two objects $A_L$ and $A_M$ of ${{\mathcal A}}$, we have $$H^{n} {{\mathcal A}}(A_L,A_M) = {{\operatorname{\mathsf{Hom}}}}_{{{\mathcal H}}{{\mathcal E}}}(A_L, A_M[n]) {\: , \;}$$ where ${{\mathcal H}}{{\mathcal E}}$ denotes the homotopy category of complexes over ${{\mathcal E}}$. To compute this space, let $G_1$ be the functor ${{\mathcal E}}\to {{\operatorname{\mathsf{Mod}}}\nolimits}{{\mathcal M}}$ taking an object $X$ to ${{\mathcal E}}(?,X)|{{\mathcal M}}$. The image of $A_M$ under $G_1$ is a projective resolution of the ${{\mathcal M}}$-module ${\underline{M}}={\underline{{{\mathcal E}}}}(?,M)$. Thus we have $${{\operatorname{\mathsf{Hom}}}}_{{{\mathcal H}}{{\mathcal E}}}(A_L, A_M[n]) {\stackrel{_\sim}{\rightarrow}}({{\mathcal D}}{{\mathcal M}})({\underline{L}}, {\underline{M}}[n]) = {{\operatorname{\mathsf{Ext}}}}^n_{{\mathcal M}}({\underline{L}},{\underline{M}}) {\: , \;}$$ where ${{\mathcal D}}{{\mathcal M}}$ denotes the (unbounded) derived category of ${{\operatorname{\mathsf{Mod}}}\nolimits}{{\mathcal M}}$. Notice that by the Yoneda lemma, for each object $N$ of ${{\mathcal M}}$, we have a canonical isomorphism $${{\operatorname{\mathsf{Hom}}}}_{{\mathcal M}}(G_1 N, {\underline{M}}) {\stackrel{_\sim}{\rightarrow}}{\underline{M}}(N) = {\underline{{{\mathcal E}}}}(N,M)={{\mathcal C}}(N,M).$$ Using this we see that the vector space ${{\operatorname{\mathsf{Ext}}}}^n_{{\mathcal M}}({\underline{L}},{\underline{M}})$ is the homology in degree $n$ of the complex $$0 \to {{\mathcal C}}(L,M)\to 0 \to {{\mathcal C}}(L_0,M) \to {{\mathcal C}}(L_1, M) \to 0.$$ Clearly it is isomorphic to ${{\mathcal C}}(L,M)$ for $n=0$. Using the triangle $$S^{-2} L \to L_1 \to L_0 \to S^{-1} L$$ and the fact that ${{\mathcal C}}(S^{-1}L_0, M)=0$ and ${{\mathcal C}}(S^{-1}L,M)=0$, we see that the homology is isomorphic to ${{\mathcal C}}(S^{-2}L, M)=D{{\mathcal C}}(M,L)$ for $n=3$ and vanishes for all other $n\neq 0$. More precisely, we see that the map $M \mapsto A_M$ extends to an equivalence whose target is the (additive) graded category $H^*{{\mathcal A}}$ and whose source is the graded category ${{\mathcal T}}\oplus (D{{\mathcal T}})[-3]$ whose objects are those of ${{\mathcal T}}$ and whose morphisms are given by $${{\mathcal T}}(L,M) \oplus (D{{\mathcal T}}(M,L))[-3].$$ In particular, we have a faithful functor ${{\mathcal T}}\to H^*{{\mathcal A}}$ which yields an equivalence from ${{\mathcal T}}$ to $H^0{{\mathcal A}}$. We denote by ${{\mathcal D}}^b({{\mathcal A}})$ the full subcategory of the derived category ${{\mathcal D}}{{\mathcal A}}$ whose objects are the dg modules $X$ such that the restriction of the sum of the $H^nX$, $n\in{\mathbb{Z}}$, to ${{\mathcal T}}$ lies in the category ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ of finitely presented ${{\mathcal T}}$-modules (by Proposition 2.1 a) of [@KellerReiten07], this category is abelian). In particular, each representable ${{\mathcal A}}$-module lies in ${{\mathcal D}}^b({{\mathcal A}})$ (by Proposition 2.1 b) of [@KellerReiten07]) and thus the perfect derived category ${{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})$ is contained in ${{\mathcal D}}^b({{\mathcal A}})$. We denote by ${{\operatorname{\mathsf{stab}}}}({{\mathcal A}})$ the triangle quotient ${{\mathcal D}}^b({{\mathcal A}})/{{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})$. Recall from [@KellerVossieck87] [@Rickard89b] that we have a triangle equivalence $${\underline{{{\mathcal E}}}} {\stackrel{_\sim}{\rightarrow}}{{\mathcal D}}^b({{\mathcal E}})/{{\mathcal D}}^b({{\mathcal P}}).$$ Let $G: {{\mathcal H}}^b({{\mathcal E}}) \to {{\mathcal D}}{{\mathcal A}}$ be the functor which takes a bounded complex $X$ over ${{\mathcal E}}$ to the functor $$A_M \mapsto {{\operatorname{\mathsf{Hom^\bullet}}}}_{{\mathcal E}}(A_M, X) {\: , \;}$$ where ${{\operatorname{\mathsf{Hom^\bullet}}}}_{{\mathcal E}}$ is the complex whose $n$th component is formed by the morphisms of graded objects, homogeneous of degree $n$, and the differential is the supercommutator with the differentials of $A_M$ and $X$. We will show that $G$ takes ${{\mathcal D}}^b({{\mathcal P}})$ to zero, that it maps ${{\mathcal H}}^b({{\mathcal E}})$ to ${{\mathcal D}}^b({{\mathcal A}})$ and the subcategory of acyclic complexes to ${{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})$. Thus it will induce a triangle functor $${{\mathcal D}}^b({{\mathcal E}})/{{\mathcal D}}^b({{\mathcal P}}) \to {{\mathcal D}}^b({{\mathcal A}})/{{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})$$ and we will obtain the required functor as the composition $${\underline{{{\mathcal E}}}} {\stackrel{_\sim}{\rightarrow}}{{\mathcal D}}^b({{\mathcal E}})/{{\mathcal D}}^b({{\mathcal P}}) \to {{\mathcal D}}^b({{\mathcal A}})/{{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})={{\operatorname{\mathsf{stab}}}}({{\mathcal A}}).$$ First recall that if $A$ is an ayclic complex and $I$ a left bounded complex of injectives, then each morphism from $A$ to $I$ is nullhomotopic. In particular, the complex ${{\operatorname{\mathsf{Hom^\bullet}}}}_{{\mathcal E}}(A_M,P)$ is nullhomotopic for each $P$ in ${{\mathcal H}}^b({{\mathcal P}})$. Thus $G$ takes ${{\mathcal H}}^b({{\mathcal P}})$ to zero. Now, we would like to show that $G$ takes values in ${{\mathcal D}}^b({{\mathcal A}})$ and that the image of each bounded acyclic complex is in ${{\operatorname{\mathsf{per}}}\nolimits}({{\mathcal A}})$. For this, we need to compute $$(GX)(A_L)={{\operatorname{\mathsf{Hom^\bullet}}}}_{{{\mathcal E}}}(A_L, X)$$ for $L$ in ${{\mathcal M}}$ and $X$ in ${{\mathcal H}}^b({{\mathcal E}})$. To show that the restriction of the sum of the homologies of $GX$ lies in ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$, it suffices to show that this holds if $X$ is concentrated in one degree. Moreover, if we have a conflation $$0 \to M_1 \to M_0 \to X \to 0$$ of ${{\mathcal E}}$ with $M_i$ in ${{\mathcal M}}$, it induces a short exact sequence of complexes $$0 \to GM_1 \to GM_0 \to GX \to 0.$$ So we may suppose that $X$ is an object of ${{\mathcal M}}$ considered as a complex concentrated in degree $0$. Then one computes that the space $${{\operatorname{\mathsf{Hom}}}}_{{{\mathcal H}}{{\mathcal E}}}(A_L, X[n])$$ is isomorphic to the homology in degree $n$ of the complex $$0 \to {{\mathcal C}}(L,X) \to 0 \to {{\mathcal C}}(L_0,X) \to {{\mathcal C}}(L_1,X) \to 0 {\: , \;}$$ where ${{\mathcal C}}(L,X)$ is in degree $0$. For $n=0$, we find that the homology is ${{\mathcal C}}(L,X)$. Using the triangle $$S^{-2} L \to L_1 \to L_0 \to S^{-1}L$$ and the vanishing of ${{\mathcal C}}(S^{-1}L,X)$ and ${{\mathcal C}}(S^{-1}L_0,X)$ we see that the homology in degree $n$ is ${{\mathcal C}}(L,S^2 X)$ for $n=3$ and vanishes for all other $n\neq 0$. This shows that the restriction of the sum of the homologies of $GX$ to ${{\mathcal T}}$ lies in ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ since the restriction of ${\underline{{{\mathcal E}}}}(?,Y)$ to ${{\mathcal T}}$ lies in ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ for each $Y$ in ${\underline{{{\mathcal E}}}}$. Now we have to show that $G$ takes acyclic bounded complexes to perfect dg ${{\mathcal A}}$-modules. For this, we first observe that we have a factorization of $G$ as the composition $$\xymatrix{{{\mathcal H}}^b({{\mathcal E}}) \ar[r]^{G_1} & {{\mathcal D}}{{\mathcal M}}\ar[r]^{G_2} & {{\mathcal D}}{{\mathcal A}}}$$ where $G_1$ sends $X$ to ${{\operatorname{\mathsf{Hom^\bullet}}}}(?,X)|{{\mathcal M}}$ and $G_2$ sends $Y$ to the dg module $$A_L \mapsto {{\operatorname{\mathsf{Hom^\bullet}}}}(G_1 A_L, Y).$$ Clearly the functor $G_1$ sends ${{\mathcal E}}$-acyclic bounded complexes to bounded complexes whose homology modules are in ${{\operatorname{\mathsf{mod}}}\nolimits}{\underline{{{\mathcal M}}}}$. Since ${{\operatorname{\mathsf{mod}}}\nolimits}{\underline{{{\mathcal M}}}}$ lies in ${{\operatorname{\mathsf{per}}}\nolimits}{{\mathcal M}}$, it follows that $G_1$ sends bounded acyclic complexes to objects of ${{\operatorname{\mathsf{per}}}\nolimits}_{{\underline{{{\mathcal M}}}}}({{\mathcal M}})$. Under the functor $G_2$, the module ${\underline{{{\mathcal M}}}}(?,L)$ is sent to $A_L$ and $G_2$ restricted to the triangulated subcategory generated by the ${\underline{{{\mathcal M}}}}(?,L)$ is fully faithful. We claim that this subcategory equals ${{\operatorname{\mathsf{per}}}\nolimits}_{{\underline{{{\mathcal M}}}}}({{\mathcal M}})$. Indeed, each object in ${{\operatorname{\mathsf{per}}}\nolimits}_{{\underline{{{\mathcal M}}}}}({{\mathcal M}})$ is an iterated extension of its homology objects placed in their respective degrees. So it suffices to show that each object concentrated in degree $0$ is the cone over a morphism between objects ${\underline{{{\mathcal M}}}}(?,L)$, $L\in{{\mathcal M}}$. But this is clear since ${{\operatorname{\mathsf{mod}}}\nolimits}{\underline{{{\mathcal M}}}}$ is equivalent to ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$, which is hereditary. It follows that $G_2$ induces an equivalence from ${{\operatorname{\mathsf{per}}}\nolimits}_{{\underline{{{\mathcal M}}}}}({{\mathcal M}})$ to ${{\operatorname{\mathsf{per}}}\nolimits}{{\mathcal A}}$ and thus $G=G_2 G_1$ sends bounded acyclic complexes to ${{\operatorname{\mathsf{per}}}\nolimits}{{\mathcal A}}$. Thus, we obtain the required triangle functor $F: {{\mathcal C}}\to {{\operatorname{\mathsf{stab}}}}({{\mathcal A}})$. In section \[subsection:formality\] below, we will show that ${{\mathcal A}}$ is formal. Thus we get an isomorphism $${{\mathcal T}}\oplus (D{{\mathcal T}})[-3] {\stackrel{_\sim}{\rightarrow}}{{\mathcal A}}$$ in the homotopy category of small dg categories. This yields an equivalence $${{\mathcal C}}_{{\mathcal T}}= {{\operatorname{\mathsf{stab}}}}({{\mathcal T}}\oplus (D{{\mathcal T}})[-3]) {\stackrel{_\sim}{\rightarrow}}{{\operatorname{\mathsf{stab}}}}({{\mathcal A}}).$$ By construction, it takes each object $T$ of ${{\mathcal T}}$ to the module $T^\wedge={{\mathcal T}}(?,T)$ in ${{\mathcal C}}_{{\mathcal T}}$. Since ${{\mathcal T}}$ generates ${{\mathcal C}}$ and the $T^\wedge$, $T\in{{\mathcal T}}$, generate ${{\mathcal C}}_{{\mathcal T}}$, it is enough to show that $F$ is fully faithful. We thank Michel Van den Bergh for simplifying our original argument: For each object $X$ of ${{\mathcal C}}$, we have a triangle $$T_1 \to T_0 \to X \to ST_1$$ with $T_0, T_1$ in ${{\mathcal T}}$. Thus, to conclude that $F$ induces a bijection $${{\mathcal C}}(T,X) \to {{\mathcal C}}_{{\mathcal T}}(FT,FX)$$ for each $T\in{{\mathcal T}}$, it suffices to show that $F$ induces bijections $${{\mathcal C}}(T, T'[i]) {\stackrel{_\sim}{\rightarrow}}{{\mathcal C}}_{{\mathcal T}}(FT, FT'[i])$$ for $T,T'$ in ${{\mathcal T}}$ and $0\leq i\leq 1$. This is clear for $i=1$ and not hard to see for $i=0$. We conclude that for each $Y$ of ${{\mathcal C}}$, $F$ induces bijections $${{\mathcal C}}(T',Y[i]) {\stackrel{_\sim}{\rightarrow}}{{\mathcal C}}_{{\mathcal T}}(FT', FY[i])$$ for all $T'$ in ${{\mathcal T}}$ and all $Y$ in ${{\mathcal C}}$ and $i\in{\mathbb{Z}}$. By the above triangle, it follows that $F$ induces bijections $${{\mathcal C}}(X,Y) \to {{\mathcal C}}_{{\mathcal T}}(FX,FY)$$ for all $X,Y$ in ${{\mathcal C}}$. Formality {#subsection:formality} --------- For categories ${{\mathcal T}}$ given by ‘small enough’ quivers $Q$, one can use the argument of Lemma 4.21 of Seidel-Thomas’ [@SeidelThomas01] to show that the category ${{\mathcal T}}\oplus (D{{\mathcal T}})[-3]$ is intrinsically formal and thus ${{\mathcal A}}$ is formal. We thank Michel Van den Bergh for pointing out that for general categories ${{\mathcal T}}$ with hereditary module categories, Seidel-Thomas’ argument cannot be adapted. Instead, we show directly that ${{\mathcal A}}$ is formal (we do not know if ${{\mathcal T}}\oplus (D{{\mathcal T}})[-3]$ is intrinsically formal). Of course, it suffices to show that the full subcategory ${{\mathcal A}}'$ whose objects are the $A_M$ with indecomposable $M$ is formal. Since $k$ is perfect, the category of bimodules over a semi-simple $k$-category is still semisimple. From this, one deduces that the category ${{\mathcal T}}$ is equivalent to the tensor category of a bimodule over the semi-simplification of ${{\mathcal T}}$, [[*cf.*]{} ]{}Proposition 4.2.5 in [@Benson98]. Using this we can construct a lift of the functor ${{\operatorname{\mathsf{ind}}}\nolimits}{{\mathcal T}}\to H^0 ({{\mathcal A}}')$ to a functor ${{\operatorname{\mathsf{ind}}}\nolimits}{{\mathcal T}}\to Z^0({{\mathcal A}}')$, where ${{\operatorname{\mathsf{ind}}}\nolimits}{{\mathcal T}}$ denotes the full subcategory of ${{\mathcal T}}$ formed by a set of representatives of the isomorphism classes of the indecomposables. We define a ${{\mathcal T}}$-bimodule by $$X(L,M) = {{\operatorname{\mathsf{Hom^\bullet}}}}_{{\mathcal E}}(A_L, M) {\: , \;}L,M\in {{\mathcal M}}{\: , \;}$$ where we consider $M$ as a subcomplex of $A_M$. Note that $X$ is a right ideal in the category ${{\mathcal A}}'$, that it is a $kQ$-subbimodule of $(L,M) \mapsto {{\mathcal A}}'(A_L,A_M)$ and that we have $fg=0$ for all homogeneous elements $f$, $g$ of $X$ of degree $>0$. The computation made above in the proof that $G$ takes ${{\mathcal H}}^b({{\mathcal E}})$ to ${{\mathcal D}}^b({{\mathcal A}})$ shows that $X$ has homology only in degree $3$ and that we have a bimodule isomorphism $$D{{\mathcal T}}(M,L) {\stackrel{_\sim}{\rightarrow}}H^3 X(L,M) {\: , \;}L,M\in {{\mathcal M}}.$$ Thus we have an isomorphism $$D{{\mathcal T}}[-3] {\stackrel{_\sim}{\rightarrow}}X$$ in the derived category of ${{\mathcal T}}$-bimodules. We choose a projective bimodule resolution $P$ of $D{{\mathcal T}}[-3]$ whose non zero components are concentrated in degrees $1$, $2$ and $3$ (note that this is possible since the bimodule category is of global dimension $2$). We obtain a morphism of complexes of bimodules $$P \to X$$ inducing an isomorphism in homology. We compose it with the inclusion $X \to {{\mathcal A}}'$. All products of elements in the image of $P$ vanish since they all lie in components of degree $>0$ of $X$. Thus we obtain a morphism of dg categories $${{\mathcal T}}\oplus P \to {{\mathcal A}}'$$ inducing an isomorphism in homology. This clearly shows that ${{\mathcal A}}'$ is formal. A generalization to higher dimensions ===================================== Negative extension groups {#subsection:negative} ------------------------- Let $k$ be a field and $H$ a finite-dimensional hereditary $k$-algebra. We write $\nu$ for the Serre functor of the bounded derived category ${{\mathcal D}}={{\mathcal D}}^b({{\operatorname{\mathsf{mod}}}\nolimits}H)$ and $S$ for its suspension functor. Let $d\geq 2$ be an integer. Let ${{\mathcal C}}={{\mathcal C}}^{(d)}_H$ be the [*$d$-cluster category*]{}, [[*i.e.*]{} ]{}the orbit category of ${{\mathcal D}}$ under the action of the automorphism $\nu^{-1} S^d$, and $\pi : {{\mathcal D}}\to {{\mathcal C}}$ the canonical projection functor. We know from [@Keller05] that ${{\mathcal C}}$ is canonically triangulated and $d$-Calabi Yau and that $\pi$ is a triangle functor. Moreover, the image $\pi(H)$ of $H$ in ${{\mathcal C}}$ is a $d$-cluster tilting object, [[*cf.*]{} ]{}[[*e.g.*]{} ]{}[@KellerReiten07]. The fact that the module $H$ is projective and concentrated in degree $0$ yields vanishing properties for the negative selfextension groups of $\pi(H)$ if $d\geq 3$: We have $${{\operatorname{\mathsf{Hom}}}}(\pi(H), S^{-i}\pi(H))=0$$ for $1 \leq i \leq d-2$. Put $T=H$. For $p\in{\mathbb{Z}}$, let ${{\mathcal D}}_{\leq p}$ and ${{\mathcal D}}_{\geq p}$ be the $(-p)$th suspensions of the canonical left, respectively right, aisles of ${{\mathcal D}}$, [[*cf.*]{} ]{}[@KellerVossieck88]. We have to show that the groups $${{\operatorname{\mathsf{Hom}}}}(T, \nu^{-p} S^{pd-i} T)$$ vanish for all $p\in{\mathbb{Z}}$ and all $1\leq i \leq d-2$. Suppose that $p=-q$ for some $q\geq 0$. Then we have $${{\operatorname{\mathsf{Hom}}}}(T, \nu^{-p} S^{pd-i} T)) = {{\operatorname{\mathsf{Hom}}}}(T, \nu^{q} S^{-qd-i} T)$$ and the last group vanishes since $T$ lies in ${{\mathcal D}}_{\leq 0}$ and $\nu^{q} S^{-qd-i}T$ lies in ${{\mathcal D}}_{\geq q(d-1)+i}$ and we have $q(d-1)+i>0$. Now suppose that $p\geq 1$. Then we have $${{\operatorname{\mathsf{Hom}}}}(T, \nu^{-p} S^{pd-i} T) = {{\operatorname{\mathsf{Hom}}}}(\nu^{p} T , S^{pd-i} T) = {{\operatorname{\mathsf{Hom}}}}(\nu^{p-1}(\nu T), S^{pd-i} T)$$ and this group vanishes since we have $\nu^{p-1}(\nu T) \in {{\mathcal D}}_{\geq p-1}$ (because $\nu T=\nu H$ is in ${{\operatorname{\mathsf{mod}}}\nolimits}H$) and $S^{pd-i}T\in {{\mathcal D}}_{\leq -(pd-i)}$ and $$(pd-i)-(p-1) = p(d-1) -i+1 \geq d-1-i+1\geq d-i \geq 2.$$ A characterization of higher cluster categories ----------------------------------------------- Let $d\geq 2$ be an integer, $k$ an algebraically closed field and ${{\mathcal C}}$ a ${{\operatorname{\mathsf{Hom}}}}$-finite algebraic $d$-Calabi Yau category containing a $d$-cluster tilting object $T$. Suppose that ${{\operatorname{\mathsf{Hom}}}}(T, S^{-i}T)=0$ for $1\leq i\leq d-2$. If $H={{\operatorname{\mathsf{End}}}}(T)$ is hereditary, then, with the notations of section \[subsection:negative\], there is a triangle equivalence ${{\mathcal C}}{\stackrel{_\sim}{\rightarrow}}{{\mathcal C}}^{(d)}_H$ taking $T$ to $\pi(H)$. Notice that by the lemma above, the assumption on the vanishing of the negative extension groups is necessary. These assumptions imply that the endomorphism algebra is Gorenstein of dimension $\leq d-1$, as we show in lemma \[lemma:Gorenstein\] below. For $d\geq 3$, this does not, of course, imply that the endomorphism algebra is hereditary if its quiver does not have oriented cycles, but it implies that the global dimension is at most $d-1$. We will prove the theorem below in section \[subsection:proof2\]. In [@IyamaYoshino06], Theorem 1.3, the reader will find an example from the study of rigid Cohen-Macaulay modules which shows that the vanishing of the negative extension groups does not follow from the other hypotheses. The following simple example, based on an idea of M. Van den Bergh, is similar in spirit: Example ------- Let ${\widetilde{H}}$ be the path algebra of a quiver with underlying graph $A_6$ and alternating orientation. Put ${\widetilde{{{\mathcal D}}}}={{\mathcal D}}^b({{\operatorname{\mathsf{mod}}}\nolimits}{\widetilde{H}})$ and let ${{\mathcal C}}$ be the orbit category of ${\widetilde{{{\mathcal D}}}}$ under the automorphism $F=\tau^4$ (where $\tau=S^{-1}\nu$). Then ${{\mathcal C}}$ is $3$-Calabi Yau: Indeed, one checks that $F^2=\tau^{-1} S^2$ in ${\widetilde{{{\mathcal D}}}}$, which clearly yields $\nu=S^3$ in ${{\mathcal C}}$. The following diagram shows a piece of the Auslander-Reiten quiver of ${\widetilde{{{\mathcal D}}}}$ which is a ‘fundamental domain’ for $F$. To obtain the Auslander-Reiten quiver of ${{\mathcal C}}$, we identify the left and right borders. $$\begin{xy} 0;<1pt,0pt>:<0pt,-1pt>:: (0,0) *+{P_1} ="0", (22,22) *+{P_2} ="1", (0,44) *+{P_3} ="2", (22,67) *{\circ} ="3", (0,89) *{\circ} ="4", (22,111) *{\circ} ="5", (45,0) *{\circ} ="6", (67,22) *{\circ} ="7", (45,44) *{\circ} ="8", (67,67) *{\circ} ="9", (45,89) *{\circ} ="10", (67,111) *{\circ} ="11", (89,0) *{\circ} ="12", (112,22) *{\circ} ="13", (89,44) *{\circ} ="14", (112,67) *{\circ} ="15", (89,89) *{\circ} ="16", (112,111) *{\circ} ="17", (134,0) *{\circ} ="18", (156,22) *{\circ} ="19", (134,44) *{\circ} ="20", (156,67) *+{SP_3} ="21", (134,89) *{\circ} ="22", (156,111) *{\circ} ="23", (179,0) *+{FP_1} ="24", (201,22) *+{FP_2} ="25", (179,44) *+{FP_3} ="26", (201,67) *{\circ} ="27", (179,89) *{\circ} ="28", (201,111) *{\circ} ="29", "0", {\ar"1"}, "2", {\ar"1"}, "1", {\ar"6"}, "1", {\ar"8"}, "2", {\ar"3"}, "4", {\ar"3"}, "3", {\ar"8"}, "3", {\ar"10"}, "4", {\ar"5"}, "5", {\ar"10"}, "6", {\ar"7"}, "8", {\ar"7"}, "7", {\ar"12"}, "7", {\ar"14"}, "8", {\ar"9"}, "10", {\ar"9"}, "9", {\ar"14"}, "9", {\ar"16"}, "10", {\ar"11"}, "11", {\ar"16"}, "12", {\ar"13"}, "14", {\ar"13"}, "13", {\ar"18"}, "13", {\ar"20"}, "14", {\ar"15"}, "16", {\ar"15"}, "15", {\ar"20"}, "15", {\ar"22"}, "16", {\ar"17"}, "17", {\ar"22"}, "18", {\ar"19"}, "20", {\ar"19"}, "19", {\ar"24"}, "19", {\ar"26"}, "20", {\ar"21"}, "22", {\ar"21"}, "21", {\ar"26"}, "21", {\ar"28"}, "22", {\ar"23"}, "23", {\ar"28"}, "24", {\ar"25"}, "26", {\ar"25"}, "26", {\ar"27"}, "28", {\ar"27"}, "28", {\ar"29"}, \end{xy}$$ Using the mesh category of this quiver, it is not hard to check that the sum of the images of the indecomposable projectives $P_1, P_2, P_3$ in ${{\mathcal C}}$ is a $3$-cluster tilting object whose endomorphism ring $H$ is the path algebra on the full subquiver with the corresponding $3$ vertices. On the other hand, the image of $P_3$ in ${{\mathcal C}}$ has a one-dimensional space of $(-1)$-extensions. Note that ${{\mathcal C}}={\widetilde{{{\mathcal D}}}}/F$ is nevertheless an orbit category and admits the $3$-cluster category $${{\mathcal C}}^{(3)}_{{\widetilde{H}}}={\widetilde{{{\mathcal D}}}}/F^2$$ as a ‘2-sheeted covering’. Proof {#subsection:proof2} ----- The proof of the theorem follows the lines of the one in section \[subsection:proof\]: Let ${{\mathcal T}}$ be the full subcategory of ${{\mathcal C}}$ whose objects are the direct sums of direct factors of $T$. Let $M$ be an object of ${{\mathcal T}}$. We construct an ${{\mathcal E}}$-acyclic complex $A_M$ $$0 \to M_{d+1} \to M_d \to \ldots \to M_1 \to M_0 \to 0$$ which yields a resolution of the ${{\mathcal M}}$-module $${\underline{{{\mathcal E}}}}(?,M): {{\mathcal M}}^{op}\to {{\operatorname{\mathsf{Mod}}}\nolimits}k$$ as in part b) of Theorem 5.4 in [@KellerReiten07]. Thus we can take $M_0=M$ and the morphism $M_1 \to M_0$ is a deflation with projective $M_1$. Each morphism $$M_i \to Z_{i-1}={{\operatorname{\mathsf{ker}}}\nolimits}(M_{i-1} \to M_{i-2}) {\: , \;}i\geq 2{\: , \;}$$ yields a ${{\mathcal T}}$-approximation in ${{\mathcal C}}$. Our vanishing assumption then implies that $M_2$, …, $M_{d-1}$ are projective. As in section \[subsection:proof\], we let ${{\mathcal A}}$ be the dg subcategory of the dg category ${{\mathcal C}}({{\mathcal E}})_{dg}$ of complexes over ${{\mathcal E}}$ whose objects are these acyclic complexes. Thus, for two objects $A_L$ and $A_M$ of ${{\mathcal A}}$, we have $$H^{n} {{\mathcal A}}(A_L,A_M) = {{\operatorname{\mathsf{Hom}}}}_{{{\mathcal H}}{{\mathcal E}}}(A_L, A_M[n]).$$ One computes that this vector space is isomorphic to ${{\mathcal C}}(L,M)$ for $n=0$, to ${{\mathcal C}}(L,\Sigma M)=D{{\mathcal C}}(M,L)$ for $n=d+1$ and vanishes for all other $n$. Here we use again our vanishing hypothesis. We see that the map $M \mapsto A_M$ extends to an equivalence whose target is the (additive) graded category $H^*{{\mathcal A}}$ and whose source is the graded category ${{\mathcal T}}\oplus (D{{\mathcal T}})[-(d+1)]$ whose objects are those of ${{\mathcal T}}$ and whose morphisms are given by $${{\mathcal T}}(L,M) \oplus (D{{\mathcal T}}(M,L))[-(d+1)].$$ Now the proof proceeds as in \[subsection:proof\] and we obtain a triangle functor $F: {{\mathcal C}}\to {{\mathcal C}}_{{\mathcal T}}$ taking the subcategory ${{\mathcal T}}$ to ${{\operatorname{\mathsf{add}}}\nolimits}\pi(H)$ and whose restriction to ${{\mathcal T}}$ is an equivalence. By lemma \[lemma:equivalences\] below, $F$ is an equivalence. Equivalences between $d$-Calabi Yau categories ---------------------------------------------- Let $d\geq 2$ be an integer, $k$ a field and ${{\mathcal C}}$ and ${{\mathcal C}}'$ $k$-linear triangulated categories which are $d$-Calabi Yau. Let ${{\mathcal T}}\subset {{\mathcal C}}$ and ${{\mathcal T}}'\subset {{\mathcal C}}'$ be $d$-cluster tilting subcategories. Suppose that $F:{{\mathcal C}}\to{{\mathcal C}}'$ is a triangle functor taking ${{\mathcal T}}$ to ${{\mathcal T}}'$. \[lemma:equivalences\] $F$ is an equivalence iff the restriction of $F$ to ${{\mathcal T}}$ is an equivalence. It follows from Proposition 5.5 a) of [@KellerReiten07], [[*cf.*]{} ]{}also part (1) of Theorem 3.1 of [@IyamaYoshino06], that ${{\mathcal C}}$ equals its subcategory $${{\mathcal T}}*S{{\mathcal T}}* \ldots * S^{d-1}{{\mathcal T}}$$ and similarly for ${{\mathcal C}}'$. Suppose that the restriction of $F$ to ${{\mathcal T}}$ is an equivalence. Let $T\in{{\mathcal T}}$. By induction, we see that for each $1\leq i\leq d-1$, the map $${{\mathcal C}}(T,Y) \to {{\mathcal C}}'(FT, FY)$$ is bijective for each $Y\in {{\mathcal T}}*S{{\mathcal T}}*\ldots * S^i{{\mathcal T}}$. Thus the map $${{\mathcal C}}(S^j T, Y) \to {{\mathcal C}}'(S^j FT, FY)$$ is bijective for all $j\in{\mathbb{Z}}$ and $Y$ in ${{\mathcal C}}$. Then it follows that the map $${{\mathcal C}}(X,Y) \to {{\mathcal C}}'(FX,FY)$$ is bijective for all $X,Y$ in ${{\mathcal C}}$. Thus $F$ is fully faithful. Since ${{\mathcal T}}'$ generates ${{\mathcal C}}'$, the functor $F$ is an equivalence. Conversely, if $F$ is an equivalence and takes ${{\mathcal T}}$ to ${{\mathcal T}}'$, then the image of ${{\mathcal T}}$ has to be ${{\mathcal T}}'$ since $F{{\mathcal T}}$ is maximal $(d-1)$-orthogonal in ${{\mathcal C}}'$. The Gorenstein property for certain $d$-Calabi Yau categories {#subsection:Gorenstein} ------------------------------------------------------------- Let $d\geq 2$ be an integer, $k$ a field and ${{\mathcal C}}$ a $k$-linear triangulated category which is $d$-Calabi Yau. Let ${{\mathcal T}}\subset {{\mathcal C}}$ be a $d$-cluster tilting subcategory such that we have $${{\mathcal C}}(T, S^{-i}T')=0$$ for all $1\leq i \leq d-2$ and all $T$, $T'$ of ${{\mathcal T}}$. \[lemma:Gorenstein\] The category ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ is Gorenstein of dimension less than or equal to $d-1$. As in [@KellerReiten07], one sees that the functor ${{\operatorname{\mathsf{Hom}}}}(T,?)$ induces an equivalence from the category $S^{d}{{\mathcal T}}$ to the category of injectives of ${{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$. So we have to show that the ${{\mathcal T}}$-module ${{\operatorname{\mathsf{Hom}}}}(?,S^{d}T)$ is of projective dimension $\leq d-1$ for each $T$ in ${{\mathcal T}}$. Put $Y=S^{d}T$. We proceed as in section 5.5 of [@KellerReiten07]: Let $T_0 \to Y$ be a right ${{\mathcal T}}$-approximation of $Y$. We define an object $Z_0$ by the triangle $$Z_0 \to T_0 \to Y \to SZ_0.$$ Now we choose a right ${{\mathcal T}}$-approximation $T_1 \to Z_0$ and define $Z_1$ by the triangle $$Z_1 \to T_1 \to Z_0 \to SZ_1.$$ We continue inductively constructing triangles $$Z_i \to T_i \to Z_{i-1} \to SZ_i$$ for $1<i\leq d-2$. By proposition 5.5 of [@KellerReiten07], the object $Z_{d-2}$ belongs to ${{\mathcal T}}$. We obtain a complex $$0 \to Z_{d-2} \to T_{d-2} \to \ldots \to T_1 \to T_0 \to Y \to 0.$$ We claim that its image under the functor $F:{{\mathcal C}}\to {{\operatorname{\mathsf{mod}}}\nolimits}{{\mathcal T}}$ taking an object $X$ of ${{\mathcal C}}$ to ${{\mathcal C}}(?,X)|{{\mathcal T}}$ is a projective resolution of $FY$. Indeed, by induction one checks that the object $Z_i$ belongs to $$S^{d-i-1}{{\mathcal T}}* S^{-i}{{\mathcal T}}*S^{-i+1}{{\mathcal T}}* \ldots * S^{-1}{{\mathcal T}}* {{\mathcal T}}.$$ Thus, for each $T\in{{\mathcal T}}$, we have ${{\mathcal C}}(T, S^{-1} Z_i)=0$ by our vanishing assumptions. Moreover, the maps $FT_{i+1}\to FZ_i$ are surjective, by construction. Therefore, the triangle $$S^{-1} Z_{i-1} \to Z_i \to T_i \to Z_{i-1}$$ induces a short exact sequence $$0 \to FZ_{i} \to FT_i \to FZ_i \to 0$$ for each $0\leq i \leq d-2$, where $Z_{-1}=Y$. This implies the assertion. An alternative proof of the main theorem ======================================== <span style="font-variant:small-caps;">Michel Van den Bergh</span> In this appendix we give a proof of Theorem \[subsection:statement\] which is based on the universal property of orbit categories [@Keller05]. We use the same notations as in the main text, but for the purposes of exposition we will assume that $\Tscr$ consists of a single object $T$ such that $B=\Cscr(T,T)=kQ$ where $Q$ is a (necessarily finite) quiver. The extension to more general $\Tscr$ is routine. The dualizing module -------------------- For use below we recall a version of the Gorensteinness result from [@KellerReiten07]. Assume that $\Cscr$ is a two-dimensional ${{\operatorname{\mathsf{Ext}}}}$-finite Krull-Schmidt Calabi-Yau category with a cluster tilting object $T$. Let $B=\Cscr(T,T)$. For a finitely generated projective right $B$-module we define $P\otimes_B T$ in the obvious way. For any $M\in \Cscr$, there is a distinguished triangle ([[*e.g.*]{} ]{} [@KellerReiten07]) $$P''\otimes_B T\xrightarrow{\phi} Q''\otimes_B T\r M\r$$ Now we apply this with $M=T[2]$. Consider a distinguished triangle $$P''\otimes_B T\r Q''\otimes_B T\r T[2]\r$$ Applying the long exact sequence for ${{\operatorname{\mathsf{Hom}}}}_\Cscr(T,-)$ we obtain a corresponding projective resolution as right module of the dualizing module of $B$: $$\label{nice} 0\r P''\r Q''\r DB\r 0$$ If we choose any other right module resolution of $DB$ $$\label{nice1} 0\r P'\r Q'\xrightarrow{\alpha} DB\r 0$$ then it is equal to up to contractible summands. Hence we obtain a distinguished triangle $$\label{nice2} P'\otimes_B T\r Q'\otimes_B T\xrightarrow{\alpha'} T[2]\r$$ Changing, if necessary, $\alpha'$ by a unit in $B={{\operatorname{\mathsf{End}}}}(T[2])$ we may and we will assume that ${{\operatorname{\mathsf{Hom}}}}(T,\alpha')=\alpha$ (under the canonical identifications ${{\operatorname{\mathsf{Hom}}}}_\Cscr(T,Q'\otimes_BT)=Q'$ and ${{\operatorname{\mathsf{Hom}}}}_\Cscr(T,T[2])=DB$). The proof {#the-proof} --------- We now let $\Cscr$ be as in the main text. By [@Keller94 Thm.4.3] we may assume that $\Cscr$ is a strict ($=$ closed under isomorphism) triangulated subcategory of a derived category ${{\mathcal D}}(\Ascr)$ for some DG-category $\Ascr$. We denote by ${}_B\Cscr$ the full subcategory of ${{\mathcal D}}(B\otimes \Ascr)$ whose objects are differential graded $B\otimes \Ascr$-modules which are in $\Cscr$ when considered as $\Ascr$-modules. Clearly ${}_B\Cscr$ is triangulated. \[ABlemma\] Assume that $B=kQ$. Then the following holds. - $T$ may be lifted to an object in ${}_B\Cscr$, also denoted by $T$. - There is an isomorphism in ${}_B\Cscr$: $ DB\Lotimes_B T\cong T[2] $. We may assume that $T$ is a homotopy projective $\Ascr$-module containing a summand for each of the vertices of $Q$. Then we may lift the action of the arrows in $Q$ to an action of $kQ$ on $T$. Hence (a) holds. To prove (b), we choose a resolution of $B$-bimodules $$0\r P' \r Q' \xrightarrow{\alpha} DB\r 0$$ where $P'$, $Q'$ are projective on the right. Such a resolution may be obtained by suitably truncating a projective bimodule resolution of $DB$. Derived tensoring this resolution on the right by $T$ and comparing with we find an isomorphism in $\Cscr$ $$\label{isomorphism} c:DB\Lotimes_B T\cong T[2]$$ between objects in ${}_B\Cscr$. Note that $c$ satisfies $c\circ \alpha=\alpha'$. We claim that $c$ in is compatible with the left $B$-actions in $\Cscr$ on both sides. Let $b\in B$. Then we have a commutative diagram of right $B$-modules $$\label{tocomparewith} \begin{CD} 0 @>>> P' @>>> Q' @>>> DB @>>> 0\\ @. @Vb\cdot VV @V b\cdot VV @Vb \cdot VV @. \\ 0@>>> P' @>>> Q' @>>> DB @>>> 0 \end{CD}$$ Tensoring on the right by $T$ we obtain a morphism of triangles in $\Cscr$ $$\begin{CD} P'\otimes_B T @>>> Q'\otimes_B T @>\alpha'>> T[2] @>>> \\ @V b\cdot VV @V b\cdot VV @V b' VV\\ P'\otimes_B T @>>> Q'\otimes_B T @>\alpha'>> T[2] @>>> \end{CD}$$ where $b'=c\circ (b\otimes \Id_T)\circ c^{-1}$. We need to prove that $b'=b$ under the identification $B={{\operatorname{\mathsf{End}}}}_\Cscr(T[2])$. This follows easily by applying the functor ${{\operatorname{\mathsf{Hom}}}}_\Cscr(T,-)$ and comparing to (using the fact that ${{\operatorname{\mathsf{Hom}}}}(T,\alpha')=\alpha$). The proof of (b) can now be completed by invoking the following lemma. \[ll\] Assume that $B$ has Hochschild dimension one. Let $M,N\in {}_B\Cscr$. Then the map $${{\operatorname{\mathsf{Hom}}}}_{{}_B\Cscr}(M,N)\r {{\operatorname{\mathsf{Hom}}}}_{\Cscr}(M,N)^B$$ is surjective (where $(-)^B$ denotes the $B$-centralizer). Replacing $M$ by a homotopy projective and $N$ by a homotopy injective $B\otimes \Ascr$-module one easily obtains the following identity $${{\operatorname{\mathsf{RHom}}}}_{B\otimes\Ascr}(M,N)={{\operatorname{\mathsf{RHom}}}}_{B^e}(B, {{\operatorname{\mathsf{RHom}}}}_{\Ascr}(M,N))$$ which yields a spectral sequence $$E^{pq}_2: {{\operatorname{\mathsf{Ext}}}}^p_{B^e}(B, {{\operatorname{\mathsf{Ext}}}}^q_{\Cscr}(M,N))\Rightarrow {{\operatorname{\mathsf{Ext}}}}^n_{{}_B\Cscr}(M,N)$$ Using the fact that $B$ has projective dimension one as bimodule this yields a short exact sequence $$0\r {{\operatorname{\mathsf{Ext}}}}^1_{B^e}(B, {{\operatorname{\mathsf{Ext}}}}^{-1}_\Cscr(M,N))\r {{\operatorname{\mathsf{Hom}}}}_{{}_B\Cscr}(M,N)\r {{\operatorname{\mathsf{Hom}}}}_{B^e}(B, {{\operatorname{\mathsf{Hom}}}}_\Cscr(M,N))\r 0$$ which gives in particular what we wanted to show. We can now finish the proof of the main theorem. By Lemma \[ABlemma\](a) we have a functor $$F=-\Lotimes_B T:D^b(B)\r \Cscr$$ which by \[ABlemma\](b) satisfies $$F\circ \Sigma[-2]=F\circ (-\otimes_B DB[-2])\cong F$$ By the universal property of orbit categories [@Keller05], we obtain an exact functor $${{\mathcal D}}^b(B)/\Sigma[-2]\r \Cscr$$ which sends $B$ to $T$. This functor is then an equivalence by Lemma \[lemma:equivalences\] which finishes the proof. \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{} [10]{} Claire Amiot, *On the structure of triangulated categories with finitely many indecomposables*, arXiv:math/0612141v2 \[math.CT\], to appear in Bull. SMF. I. Assem, T. Br[ü]{}stle, R. Schiffler, and G. Todorov, *Cluster categories and duplicated algebras*, J. Algebra **305** (2006), no. 1, 548–561. Ibrahim Assem, Thomas Brüstle, and Ralf Schiffler, *[Cluster-tilted algebras as trivial extensions]{}*, arXiv:math.RT/0601537, to appear in J. London Math. Soc. Maurice Auslander, *Functors and morphisms determined by objects*, Representation theory of algebras (Proc. Conf., Temple Univ., Philadelphia, Pa., 1976), Dekker, New York, 1978, pp. 1–244. Lecture Notes in Pure Appl. Math., Vol. 37. Maurice Auslander and Idun Reiten, *[$D$]{}[T]{}r-periodic modules and functors*, Representation theory of algebras (Cocoyoc, 1994), CMS Conf. Proc., vol. 18, Amer. Math. Soc., Providence, RI, 1996, pp. 39–50. Aslak Bakke Buan, Osamu Iyama, Idun Reiten, and Jeanne Scott, *[Cluster structures for 2-Calabi-Yau categories and unipotent groups]{}*, arXiv:math.RT/0701557. Aslak Bakke Buan, Robert J. Marsh, Markus Reineke, Idun Reiten, and Gordana Todorov, *Tilting theory and cluster combinatorics*, Advances in Mathematics **204 (2)** (2006), 572–618. Aslak Bakke Buan, Robert J. Marsh, and Idun Reiten, *[Cluster mutation via quiver representations]{}*, arXiv:math.RT/0412077, to appear in Comm. Math. Helv. Aslak Bakke Buan, Robert J. Marsh, and Idun Reiten, *Cluster-tilted algebras of finite representation type*, J. Algebra **306** (2006), no. 2, 412–431. , *Cluster-tilted algebras*, Trans. Amer. Math. Soc., **359** (2007), no. 1, 323–332, electronic. Aslak Bakke Buan, Robert J. Marsh, Idun Reiten, and Gordana Todorov, *[Clusters and seeds in acyclic cluster algebras]{}*, arXiv:math.RT/0510359, to appear in Proc. AMS, with an appendix by Aslak Bakke Buan, Philippe Caldero, Bernhard Keller, Robert Marsh, Idun Reiten and Gordana Todorov. Aslak Bakke Buan and Idun Reiten, *[From tilted to cluster-tilted algebras of Dynkin type]{}*, arXiv:math.RT/0510445. Aslak Bakke Buan and Idun Reiten, *Acyclic quivers of finite mutation type*, Int. Math. Res. Not. (2006), Art. ID 12804, 10. Aslak Bakke Buan, Idun Reiten, and Ahmet Seven, *Tame concealed algebras and cluster quivers of minimal infinite type*, arXiv:math.RT/0512137, to appear in J. Pure Appl. Alg. Karin Baur and Robert J. Marsh, *[A geometric description of $m$-cluster categories]{}*, arXiv:math.RT/0607151, to appear in IMRN. D. J. Benson, *Representations and cohomology. [I]{}*, second ed., Cambridge Studies in Advanced Mathematics, vol. 30, Cambridge University Press, Cambridge, 1998, Basic representation theory of finite groups and associative algebras. Arkady Berenstein, Sergey Fomin, and Andrei Zelevinsky, *Cluster algebras. [III]{}. [U]{}pper bounds and double [B]{}ruhat cells*, Duke Math. J. **126** (2005), no. 1, 1–52. Igor Burban, Osamu Iyama, Bernhard Keller, and Idun Reiten, *Cluster tilting for one-dimensional hypersurface singularities*, arXiv:0704.1249v1 \[math.RT\]. Philippe Caldero and Fr[é]{}d[é]{}ric Chapoton, *Cluster algebras as [H]{}all algebras of quiver representations*, Comment. Math. Helv. **81** (2006), no. 3, 595–616. Philippe Caldero, Frédéric Chapoton, and Ralf Schiffler, *Quivers with relations arising from clusters (${A}_n$ case)*, Trans. Amer. Math. Soc. **358** (2006), no. 5, 1347–1364. Philippe Caldero and Bernhard Keller, *[From triangulated categories to cluster algebras]{}*, arXiv:math.RT/0506018, to appear in Inv. Math. , *[From triangulated categories to cluster algebras II]{}*, arXiv:math.RT/0510251, to appear in Ann. Scient. ENS. Philippe Caldero and Markus Reineke, *On the quiver grassmannian in the acyclic case*, arXiv:math/0611074. William Crawley-Boevey, *On the exceptional fibres of [K]{}leinian singularities*, Amer. J. Math. **122** (2000), no. 5, 1027–1037. Harm Derksen, Jerzy Weymann, and Andrei Zelevinsky, *Quivers with potentials and their representations [I]{}: [Mutations]{}*, arXiv:0704.0649v2. Sergey Fomin and Andrei Zelevinsky, *Cluster algebras. [I]{}. [F]{}oundations*, J. Amer. Math. Soc. **15** (2002), no. 2, 497–529 (electronic). , *Cluster algebras. [II]{}. [F]{}inite type classification*, Invent. Math. **154** (2003), no. 1, 63–121. , *Cluster algebras: notes for the [CDM]{}-03 conference*, Current developments in mathematics, 2003, Int. Press, Somerville, MA, 2003, pp. 1–34. , *Cluster algebras [IV]{}: Coefficients*, Compositio Mathematica **143** (2007), 112–164. Christof Geiß, *Private communication*, 2006. Christof Gei[ß]{}, Bernard Leclerc, and Jan Schröer, *[Auslander algebras and initial seeds for cluster algebras]{}*, arXiv:math.RT/0506405, to appear in Journal of the LMS. , *[Partial flag varieties and preprojective algebras]{}*, arXiv:math.RT/0609138. , *Rigid modules over preprojective algebras [II]{}: The [K]{}ac-[M]{}oody case*, arXiv:math.RT/0703039. , *[Semicanonical bases and preprojective algebras II: A multiplication formula]{}*, arXiv:math.RT/0509483, to appear in Compositio Mathematica. , *Rigid modules over preprojective algebras*, Invent. Math. **165** (2006), no. 3, 589–632. Thorsten Holm and Peter Jorgensen, *Cluster categories and selfinjective algebras: type [A]{}*, arXiv:math/0610728v1 \[math.RT\]. , *Cluster categories and selfinjective algebras: type [D]{}*, arXiv:math/0612451v1 \[math.RT\]. Osamu Iyama, *Maximal orthogonal subcategories of triangulated categories satisfying [S]{}erre duality*, Oberwolfach Report 6, 2005. Osamu Iyama, *Higher dimensional [A]{}uslander-[R]{}eiten theory on maximal orthogonal subcategories*, Proceedings of the 37th Symposium on Ring Theory and Representation Theory, Symp. Ring Theory Represent Theory Organ. Comm., Osaka, 2005, pp. 24–30. Osamu Iyama and Idun Reiten, *[Fomin-Zelevinsky mutation and tilting modules over Calabi-Yau algebras]{}*, arXiv:math.RT/0605136, to appear in Amer. J. Math. Osamu Iyama and Yuji Yoshino, *[Mutations in triangulated categories and rigid Cohen-Macaulay modules]{}*, arXiv:math.RT/0607736. Bernhard Keller, *Deriving [D]{}[G]{} categories*, Ann. Sci. [É]{}cole Norm. Sup. (4) **27** (1994), no. 1, 63–102. , *[On triangulated orbit categories]{}*, Doc. Math. **10** (2005), 551–581. Bernhard Keller and Idun Reiten, *[Cluster-tilted algebras are Gorenstein and stably Calabi-Yau]{}*, Advances in Mathematics **211** (2007), 123–151. Bernhard Keller and Dieter Vossieck, *Sous les cat[é]{}gories d[é]{}riv[é]{}es*, C. R. Acad. Sci. Paris S[é]{}r. I Math. **305** (1987), no. 6, 225–228. , *Aisles in derived categories*, Bull. Soc. Math. Belg. S[é]{}r. A **40** (1988), no. 2, 239–253. Robert Marsh, Markus Reineke, and Andrei Zelevinsky, *Generalized associahedra via quiver representations*, Trans. Amer. Math. Soc. **355** (2003), no. 10, 4171–4186 (electronic). Jeremy Rickard, *Derived categories and stable equivalence*, J. Pure and Appl. Algebra **61** (1989), 303–317. Claus Michael Ringel, *Hall algebras and quantum groups*, Invent. Math. **101** (1990), no. 3, 583–591. , *Some remarks concerning tilting modules and tilted algebras. [Origin. Relevance. Future.]{}*, Handbook of Tilting Theory, LMS Lecture Note Series, vol. 332, Cambridge Univ. Press, Cambridge, 2007, pp. 49–104. Paul Seidel and Richard Thomas, *Braid group actions on derived categories of coherent sheaves*, Duke Math. J. **108** (2001), no. 1, 37–108. Gonçalo Tabuada, *[On the structure of Calabi-Yau categories with a cluster tilting subcategory]{}*, Doc. Math. **12** (2007), 193–213. Hugh Thomas, *[Defining an $m$-cluster category]{}*, arXiv:math.RT/0607173. Keiichi Watanabe, *Certain invariant subrings are [G]{}orenstein. [I]{}, [II]{}*, Osaka J. Math. **11** (1974), 1–8; ibid. 11 (1974), 379–388. Yuji Yoshino, *Cohen-[M]{}acaulay modules over [C]{}ohen-[M]{}acaulay rings*, London Mathematical Society Lecture Note Series, vol. 146, Cambridge University Press, Cambridge, 1990. Andrei Zelevinsky, *[Cluster algebras: notes for 2004 IMCC (Chonju, Korea, August 2004)]{}*, arXiv:math.RT/0407414. Bin Zhu, *[Generalized cluster complexes via quiver representations]{}*, arXiv:math.RT/0607155. [^1]: I. R. supported by a grant from the Norwegian Research Council

ArXiv

--- abstract: '[In this paper, we study an age-structured model which has strong biological background about mosquito plasticity. Firstly, we prove the existence of solutions and the comparison principle for a generalized system. Then, we prove the existence of the optimal control for the best harvesting. Finally, we establish necessary optimality conditions.]{}' author: - | Lin Lin Li ^a^, Cláudia Pio Ferreira ^b^, Bedreddine Ainseba ^a^\ \ \ title: '**[Optimal control of an age-structured problem modelling mosquito plasticity]{}**' --- Introduction ============ Throughout the human history, people have always been combating against many infectious diseases, such as malaria, dengue, yellow and Chikungunya fever, encephalitis and the diseases have caused uncounted mortality of mankind. During the past decades, many researchers studied the pathology of these infectious diseases and tried to control the transmission of them. One of the most studied diseases is malaria, which is mainly transmitted by Anopheles gambiae and Anopheles funestus, the main vectors [@Bn]. As the statistical data show, malaria affects more than 100 tropical countries, placing 3.3 billion people at risk [@WHO1] and the life of one African child’s life is taken by malaria every minute [@WHO2]. To reduce human’s suffering from malaria, people have been seeking efficient ways to control the malaria transmission for many years. In the past decades, the control of malaria has made slow but steady progress and the overall mortality rate has dropped by more than $25\%$ since 2000 [@Ma]. The main strategies of controlling malaria are insecticide treated nets (ITNs) and indoor residual spraying (IRS) [@Bn; @Hb; @Ma; @WHO1]. However, the effectiveness of these strategies depends on the susceptibility of the vector species to insecticides and their behaviours, ecology and population genetics. ITNs and IRS are efficient ways against the main vectors of malaria in Africa. However, the resistance of mosquitoes to insecticides forces them to adapt their behaviours to ensure their survival and reproduction. Especially, they can adapt their bitting behaviour from night to daylight [@Ss]. This new behavioural patterns lead to a resurgence of malaria morbidity in several parts of Africa [@Tj]. Thus, new methods are desired to replace the traditional strategies. In this work, we are going to model mosquito population adaption and study the optimal control problem. We consider a linear model describing the dynamics of a single species population with age dependence and spatial structure as follows $$\label{eq:01} \left\{ \begin{array}{lll} Dp- \delta \Delta p+\mu (a)p = u(a,t,x)p,&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24), \end{array} \right.$$ where $Q_{{a_\dagger}}=(0,{a_\dagger} ) \times (0,T)\times (0,24)$ and $$Dp(a,t,x)={\lim_{\varepsilon \rightarrow 0} }\frac{p\left( a+\varepsilon ,t+\varepsilon, x \right) -p\left( a,t,x\right) }{\varepsilon }$$ is the directional derivative of $p$ with respect to direction $\left( 1,1,0\right) .$ For $p$ smooth enough, it is easy to know that $$Dp=\frac {\partial p}{\partial t}+\frac {\partial p}{\partial a}.$$ Here, $p(a,t,x)$ is the distribution of individuals of age $a\ge 0$ at time $t\ge 0$ and bitting at time $x \in [0,24] $, $a_{\dagger }$ means the life expectancy of an individual and $T$ is a positive constant. As we announced, the mosquitoes can adapt their bitting time. Thus, we set their adapting model to be a $\Delta$ diffusion with a diffusive coefficient $\delta$. Moreover, $\beta(a)$ and $\mu(a)$ denote the natural fertility-rate and the natural death-rate of individual of age $a$, respectively. In fact, the new generation is also able to adapt the bitting time in order to maximize its fitness. Let $\eta$ be the maximum bitting time difference which the new generation can reach and we model the adaption of the new generation by a kernel $K$ as defined as below $$\label{eq:02} K(x,s)= \begin{cases} {(x-s)^{2}}{e^{-(x-s)^{2}}}, &s \in (0,24), \\ 0, &\text{else}. \end{cases}$$ The control function $u(a,t,x)$ represents the insecticidal effort, such as the use of ITNs and RIs. In our paper, the main goal is to prove that there exists an optimal control $u$ in limited conditions, that is, $u$ is bounded by two functions $\varsigma_1$ and $\varsigma_2$ such that the insecticidal efficiency reaches the best. Since the control function $u$ is negative, it means that we can deal with the following optimal problem $$(OH)\ \ \ \ \ \ \ \ \ Maximize \left\{-\displaystyle \int_{Q_{{a_\dagger}}}u(a,t,x)p^{u}(a,t,x)dtdxda\right\},$$ subject to $u\in U$, $$U=\{u(a,t,x)\in L^{2}(Q_{{a_\dagger}})| \ \varsigma_{1}(a,t,x)\leq u(a,t,x)\leq \varsigma_{2}(a,t,x) \ a.e.\ in \ Q_{{a_\dagger}}\},$$ where $\varsigma_{1},\varsigma_{2}\in L^{\infty}(Q_{{a_\dagger}}),\ \varsigma_{1}(a,t,x)\leq \varsigma_{2}(a,t,x) \leq 0 \ a.e.\ in \ Q_{{a_\dagger}}$ and $p^{u}$ is the solution of system . Here, we say that the control $u^{*}\in U $ is optimal if $$\displaystyle \int_{Q_{{a_\dagger}}}u^{*}(a,t,x)p^{u^{*}}(a,t,x)dtdxda\leq \displaystyle \int_{Q_{{a_\dagger}}}u(a,t,x)p^{u}(a,t,x)dtdxda,$$ for any $u\in U$. The pair $(u^{*},p^{u^{*}})$ is an optimal pair and $\displaystyle \int_{Q_{{a_\dagger}}}u^{*}p^{u^{*}}dtdxda$ is the optimal value of the cost functional. Let us recall some history about the optimal control researches. Since 1985 when Brokate [@Bm] first proposed the optimal control of the population dynamical system with an age structure, it has been widely concerned and extensively studied by more and more researchers in the past few years. It is worth mentioning that the researches of Gurtin and Murphy [@Gm1; @Gm2] about the optimal harvesting of age-structured populations provide an important basis for subsequent researches of the optimal control problem. As is well known, the optimal harvesting problem governed by nonlinear age dependent population dynamics with diffusion was considered by Aniţa [@As], where he mainly discussed the impact of the control in homogeneous Neuman boundary conditions. For more rich results about the optimal control of an age structure with non-periodic boundary conditions, one can refer to [@As2; @As3; @Fk; @Zc] and references cited therein. Note that the above results are about nonperiodic boundary conditions. However, we have seen from the practical significance of biology that it is advantageous to consider age-structured models with periodic boundary conditions and nonlocal birth processes. We would like to refer to [@Al; @Pe] for some studies about the optimal control problem with periodic boundary conditions. We also refer to [@Am; @Kd; @Lf; @Lg] as reviewing references of the optimal control problem. Let us now mention some of our work about other aspects of system with periodic boundary conditions and nonlocal birth processes. In [@Ll1], large time behaviour of the solution for such age-structured population model was considered. Moreover, we considered the local exact controllability of such age-structured problem in [@Ll2]. In this work, we study the optimal control of system . From the biological point of view (one can refer to [@Ga; @Gu; @We]), we make the following hypotheses throughout this paper: (J1) : $\mu(a)\in L^{\infty}_{loc}((0,a_{\dagger}))$, $\displaystyle \int_0^{a_{\dagger}} {\mu}(a)d a=+\infty $ and $\mu(a)\geq 0 \ a.e. \ in \ (0,a_{\dagger})$; (J2) : $\beta(a) \in L^{\infty}((0,a_{\dagger}))$, $\beta(a)\geq0 \ a.e. \ in \ (0,a_{\dagger})$; (J3) : $p_{0}(a,x)\in L^{2}((0,{a_\dagger} ) \times (0,24))$, $p_{0}(a,x)\geq 0 \ a.e. \ in \ (0,{a_\dagger} ) \times (0,24)$. Now we state our main results. \[theorem1\] For any $u\in U$, there exists a unique solution $p^{u}(a,t,x)\in L^2(Q_{a_\dagger})$ of the system . \[theorem2\] Problem $(OH)$ admits at least one optimal pair $(u^*,p^*)$. \[theorem3\] Let $(u^{*}(a,t,x),p^{*}(a,t,x))$ be an optimal pair for $(OH)$ and $q(a,t,x)$ be the solution of the following system $$\label{eq:000009} \left\{ \begin{array}{lll} Dq+ \delta \Delta q-\mu (a)q+ \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)q(0,t,s)ds =- u^{*}q-u^{*},(a,t,x)\in Q_{a_\dagger}, \\ q(a,t,0) =q(a,t,24),\hspace{4.55cm}(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x q(a,t,0) =\partial_x q(a,t,24),\hspace{3.8cm}(a,t)\in (0,a_\dagger)\times (0,T), \\ q(a_\dagger,t,x)=0,\hspace{5.8cm}(t,x)\in (0,T)\times (0,24), \\ q(a,T,x)=0 , \hspace{5.85cm}(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Then, one has $$u^{*}(a,t,x)=\left\{ \begin{aligned} &\varsigma_{1}(a,t,x),\ \ \text{if}\ \ q(a,t,x)>-1,\\ &\varsigma_{2}(a,t,x),\ \ \text{if}\ \ q(a,t,x)<-1. \end{aligned} \right.$$ This paper is organized as follows. In Section 2, we prove the existence of solutions and the comparison result for a linear model which is in general settings. Section 3 is devoted to the proof of the existence of an optimal control of system by Mazur’s Theorem. Section 4 focuses on the necessary optimality conditions. Preliminaries ============= In this section, we study some properties of the following system, which is in general settings, $$\label{eq:2.1} \left\{ \begin{array}{lll} Dp- \delta \Delta p+\mu (a,t,x)p = f(a,t,x),&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24), \end{array} \right.$$ where $\beta$, $p_0$ are under the assumptions $(J2)$, $(J3)$, $\mu$ and $f$ satisfy $$\label{mu} \mu(a,t,x)\in L^{\infty}_{loc}([0,a_{\dagger})\times [0,T]\times [0,24]),\ \mu(a,t,x)\ge 0 \text{ a.e. in $Q_{a_{\dagger}}$},$$ $$f(a,t,x)\in L^{2}(Q_{a_{\dagger}}), \ \ f(a,t,x)\ge 0 \ \text{a.e. in $Q_{a_{\dagger}}$}.$$ Especially, we prove that there exists a unique solution of system and the comparison principle for system . Before going further, we need an auxiliary lemma, which can be proved by following the proof of [@As Lemma A2.7]. \[lemma2.1\] For any $y_0(x)\in L^2(0,24)$, $g(t,x)\in L^2((0,T)\times(0,24))$, there exists a unique solution $y(t,x)\in L^2((0,T);H^1(0,24))\cup L^2_{loc}((0,T); H^2(0,24))$ of the following system $$\left\{ \begin{array}{lll} \frac{\partial y}{\partial t}(t,x)- \delta\Delta y(t,x) = g(t,x),&(t,x)\in (0,T)\times (0,24), \\ y(t,0) =y(t,24),&t\in (0,T), \\ y'(t,0) =y'(t,24),&t\in (0,T), \\ y(0,x)=y_0(x) , &x\in (0,24). \end{array} \right.$$ It is known that there exists an orthogonal basis $\{\varphi_j\}_{j\in\mathbb{N}} \subset L^2(0,24)$ and $\{\lambda_j\}\subset \mathbb{R}^+$, $\lambda_0=0$, $\lambda_j\rightarrow +\infty$ as $j\rightarrow +\infty$ such that $$\left\{ \begin{array}{lll} -\Delta \varphi_j(x)=\lambda_j \varphi_j(x), \text{ in $(0,24)$},\\ \varphi_j(0)=\varphi(24),\\ \varphi'_j(0)=\varphi'_j(24). \end{array} \right.$$ We can replace the basis in the proof of [@As Lemma A2.7] by our $\{\varphi_j\}_{j\in\mathbb{N}}$ and follow the same proof to get Lemma \[lemma2.1\]. Let us first deal with the case when $\mu$ satisfies (A) : $\mu\in L^{\infty}(Q_{a_{\dagger}})$, $\mu(a,t,x)\ge 0$ a.e. in $Q_{a_{\dagger}}$. \[lemma1\] For any fixed $f(a,t,x)\in L^{2}(Q_{{a_\dagger}})$, $b(t,x)\in L^{2}((0,T)\times(0,24))$, there exists a unique solution $p_b(a,t,x)\in L^{2}(Q_{{a_\dagger}})$ of the following system $$\label{eq:111222401} \left\{ \begin{array}{lll} Dp- \delta \Delta p+\mu(a,t,x) p = f(a,t,x),&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=b(t,x),&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24), \end{array} \right.$$ where $\mu$ is under $(A)$. Fix any $q(a,t,x)\in L^{2}(Q_{{a_\dagger}})$, we first prove that the following system has a unique solution $p_{b,q}(a,t,x)$, $$\label{eq:1112224} \left\{ \begin{array}{lll} Dp- \delta \Delta p+\mu q = f,&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=b(t,x),&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Let $S$ be an arbitrary characteristic line of equation $$S=\{(a_{0}+s,t_{0}+s);s\in(0,\alpha)\},$$ where $(a_0,t_0)\in \{0\}\times(0,T)\cup (0,a_{\dagger})\times \{0\}$ and $(a_0+\alpha,t_0+\alpha) \in \{a_{\dagger}\}\times (0,T) \cup (0,a_{\dagger})\times \{T\}$ and define $$\label{eq:111} \left\{ \begin{array}{lll} \widetilde{p}(s,x) = p(a_{0}+s,t_{0}+s,x),&(s,x)\in (0,\alpha)\times(0,24), \\ \widetilde{q}(s,x) = q(a_{0}+s,t_{0}+s,x),&(s,x)\in (0,\alpha)\times(0,24), \\ \widetilde{f}(s,x) = f(a_{0}+s,t_{0}+s,x),&(s,x)\in (0,\alpha)\times(0,24), \\ \widetilde{\mu}(s,x) = \mu(a_{0}+s,t_{0}+s,x),&(s,x)\in (0,\alpha)\times(0,24). \end{array} \right.$$ According to Lemma \[lemma2.1\], the following system admits a unique solution $\widetilde{p}\in L^2((0,\alpha);H^1(0,24))\cap {L^{2}}_{loc}((0,\alpha);H^{2}(0,24))$, $$\label{eq:222} \begin{cases} \frac {\partial \widetilde{p}}{\partial s}- \delta \Delta \widetilde{p} =\widetilde{f}-\widetilde{\mu}\widetilde{q},&(s,x)\in (0,\alpha)\times(0,24), \\ \partial _{x} \widetilde{p}(s,0)=\partial _{x} \widetilde{p}(s,24),&s\in (0,\alpha), \\ \widetilde{p}(s,0)= \widetilde{p}(s,24),&s\in (0,\alpha), \\ \widetilde{p}(0,x)=\left\{ \begin{aligned} &b(t_{0},x),\ \ \ \ a_{0}=0,\ \ x \in (0,24),\\ &p_{0}(a_{0},x),\ \ \ t_{0}=0,\ \ x \in (0,24). \end{aligned} \right. \end{cases}$$ In fact, multiplying the first equation of system by $\widetilde{p}$ and integrating on $(0,s)\times(0,24)$, one has $$\|\widetilde{p}(s)\|^2_{L^{2}(0,24)}\leq \|\widetilde{p}(0)\|^2_{L^{2}(0,24)}+\|\widetilde{f}-\widetilde{\mu}\widetilde{q}\|^2_{L^{2}((0,\alpha)\times(0,24))}+\displaystyle \int_0^{s}\|\widetilde{p}(\tau)\|^2_{L^{2}(0,24)}d\tau.$$ Then by a lemma from Bellman (see in Appendix) we get $$\label{eq:2.6} \|\widetilde{p}(s)\|^2_{L^{2}(0,24)}\leq C(\|\widetilde{p}(0)\|^2_{L^{2}(0,24)}+\|\widetilde{f}-\widetilde{\mu}\widetilde{q}\|^2_{L^{2}((0,\alpha)(0,24))})e^{\alpha},\ \forall s \in[0,\alpha]$$ Now let us denote $$p_{b,q}(a_{0}+s,t_{0}+s,x)=\widetilde{p}(s,x),\ (s,x)\in (0,\alpha)\times(0,24)$$ for any characteristic line $S$. It follows from Lemma \[lemma2.1\] and that $p_{b,q}\in L^{2}(S;H^{1}(0,24))\cap {L^{2}}_{loc}(S;H^{2}(0,24))$ for almost any characteristic line $S$, and $p_{b,q}$ satisfies $$\label{eq:20202} \left\{ \begin{array}{lll} Dp_{b,q}- \delta \Delta p_{b,q}+\mu (a,t,x)q = f(a,t,x),&(a,t,x)\in Q_{a_\dagger}, \\ p_{b,q}(a,t,0) =p_{b,q}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p_{b,q}(a,t,0) =\partial_x p_{b,q}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p_{b,q}(0,t,x)=b(t,x),&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p_{b,q}(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Now we show that $p_{b,q}(a,t,x)\in L^{2}(Q_{a_\dagger})$. It is known that there exists an orthonormal basis $\{\varphi_j\}_{j\in\mathbb{N}} \subset L^2(0,24)$ and $\{\lambda_j\}\subset \mathbb{R}^+$, $\lambda_0=0$, $\lambda_j\rightarrow +\infty$ as $j\rightarrow +\infty$ such that $$\left\{ \begin{array}{lll} -\Delta \varphi_j(x)=\lambda_j \varphi_j(x), \text{ in $(0,24)$},\\ \varphi_j(0)=\varphi(24),\\ \varphi'_j(0)=\varphi'_j(24). \end{array} \right.$$ Then, one has that $$f(a,t,x)-\mu(a,t,x) q=\sum_{j=1}^{\infty} v^j(a,t)\varphi_j(x),\text{in $L^2(0,24)$, a.e. $(a,t)\in(0,a_{\dagger})\times(0,T)$},$$ $$b(t,x)=\sum_{j=1}^{\infty} b^j(t)\varphi_j(x), \text{ in $L^2(0,24)$, a.e. $t\in(0,T)$},$$ $$p_0(a,x)=\sum_{j=1}^{\infty} p_0^j(a)\varphi_j(x), \text{ in $L^2(0,24)$, a.e. $a\in(0,a_{\dagger})$}.$$ Furthermore, $p_{b,q}(a,t,x)$ has the following expression $$p_{b,q}(a,t,x):=\sum_{j=1}^{\infty} p_{b,q}^j(a,t)\varphi_j(x),\text{ in $L^2(0,24)$ a.e. $(a,t)\in(0,a_{\dagger})\times(0,T)$}.$$ By substituting it into , one gets that $p_{b,q}^j(a,t)$ satisfies $$\left\{ \begin{array}{lll} Dp_{b,q}^j+\lambda_j \delta p_{b,q}^j=v^j(a,t), &(a,t)\in (0,a_{\dagger})\times(0,T),\\ p_{b,q}^j(a,t)\varphi_j(0)=p_{b,q}^j(a,t)\varphi_j(24),&(a,t)\in (0,a_{\dagger})\times(0,T),\\ p_{b,q}^j(a,t)\varphi_j^{'}(0)=p_{b,q}^j(a,t)\varphi_j^{'}(24),&(a,t)\in (0,a_{\dagger})\times(0,T),\\ p_{b,q}^j(0,t)=b^j(t), &t\in(0,T)\\ p_{b,q}^j(a,0)=p_0^j(a), &a\in(0,a_{\dagger}). \end{array} \right.$$ One can follow the computation of Lemma 4.1 in Aniţa $\cite [p_{.} 113-114]{As}$ and get that $p_{b,q}(a,t,x)\in L^{2}(Q_{a_\dagger})$ satisfies $$\begin{aligned} \label{eq:0001167} \|p_{b,q}\|^{2}_{L^{2}(Q_{{a_\dagger}})}\leq e^{T}(\|p_{0}\|^{2}_{L^{2}((0,a_\dagger)\times (0,24))}+\|b\|^{2}_{L^{2}((0,T)\times (0,24))}+\|f-\mu q\|^{2}_{L^{2}(Q_{a_\dagger})}).\end{aligned}$$ For an arbitrary $q(a,t,x)\in L^{2}(Q_{{a_\dagger}})$, we have obtained that system has a solution $p_{b,q}\in L^2(Q_{a_{\dagger}})$. Let us set a mapping $\Pi: L^2(Q_{a_\dagger})\rightarrow L^2(Q_{a_\dagger})$ by $\Pi (q_{i}(a,t,x))={p_{b,q}}_{i}(a,t,x)$. Take any two functions $q_1$, $q_2\in L^2(Q_{a_\dagger})$ and then $p_{b,q_1}-p_{b,q_2}$ satisfies $$\label{eq:202022233} \left\{ \begin{array}{lll} D({p_{b,q}}_{1}-{p_{b,q}}_{2})- \delta \Delta({p_{b,q}}_{1}-{p_{b,q}}_{2})+\mu (q_{1}-q_{2}) = 0,&(a,t,x)\in Q_{a_\dagger}, \\ ({p_{b,q}}_{1}-{p_{b,q}}_{2})(a,t,0) =({p_{b,q}}_{1}-{p_{b,q}}_{2})(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x ({p_{b,q}}_{1}-{p_{b,q}}_{2})(a,t,0) =\partial_x ({p_{b,q}}_{1}-{p_{b,q}}_{2})(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ ({p_{b,q}}_{1}-{p_{b,q}}_{2})(0,t,x)=0,&(t,x)\in (0,T)\times (0,24), \\ ({p_{b,q}}_{1}-{p_{b,q}}_{2})(a,0,x)=0,&(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ By the result of , one has $$\begin{aligned} \|{p_{b,q}}_{1}-{p_{b,q}}_{2}\|^{2}_{L^{2}(Q_{{a_\dagger}})} \leq e^{T}(\|\mu (q_{1}-q_{2})\|^{2}_{L^{2}(Q_{a_\dagger})}), \ in \ L^{2}(Q_{a_\dagger}).\end{aligned}$$ Obviously, when $T$ is small enough, $p_{b,q}(a,t,x)$ is a contraction mapping with respect to $q(a,t,x)$. Consequently, there exists a unique solution $p_{b}$ of system for sufficient small $T$. However, one can extend $T$ by following previous steps for $t\in(T,2T)$. Thus, system has a unique solution $p_{b}\in L^{2}(Q_{a_\dagger})$. One can follow the same idea of the proof of [@As Lemma 4.1.2] to get the following Lemma. \[lemma2.2\] For any $b_1(t,x)$, $b_2(t,x)\in L^2((0,T)\times(0,24))$, $0\le b_1(t,x)\le b_2(t,x)$ a.e. in $(0,T)\times(0,24)$, one has $$0\le p_{b_1}(a,t,x)\le p_{b_2}(a,t,x), \text{ a.e. in $Q_{a_{\dagger}}$},$$ where $p_{b_1}(a,t,x)$ and $p_{b_2}(a,t,x)$ are the solutions of system under $(A)$ with $b_1(a,t,x)$ and $b_2(a,t,x)$ respectively. \[lemma2\] There exists a unique solution $p(a,t,x)\in L^2(Q_{a_\dagger})$ of system under $(A)$. Let us define an operator $\mathscr{F}: L^2((0,T)\times (0,24))\rightarrow L^2((0,T)\times (0,24))$ by $$(\mathscr{F}b)(t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p_{b}(a,t,s)ds da,\ \text{ a.e. in } (0,T)\times (0,24).$$ For any fixed $b_{i}\in L^2((0,T)\times (0,24))$ $(i=1,2)$, let $p_{b_{1}},p_{b_{2}}\in L^2(Q_{a_{\dagger}})$ be the solutions of system with $b_1(a,t,x)$, $b_2(a,t,x)$ respectively. Let $v(a,t,x)=p_{b_{1}}(a,t,x)-p_{b_{2}}(a,t,x)$ and then $v(a,t,x)$ satisfies $$\label{eq:102345} \left\{ \begin{array}{lll} Dv- \delta \Delta v+\mu (a,t,x)v = 0,&(a,t,x)\in Q_{a_\dagger}, \\ v(a,t,0) =v(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x v(a,t,0) =\partial_x v(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ v(0,t,x)=b_{1}(t,x)-b_{2}(t,x),&(t,x)\in (0,T)\times (0,24), \\ v(a,0,x)=0 , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Then it follows by the computation of Lemma 4.1 in Aniţa $\cite [p_{.} 116]{As}$ that $$\begin{aligned} \int_0^T e^{-\lambda t} \|v(t)\|^2_{L^2((0,a_{\dagger})\times(0,24))} dt \le \frac{1}{\lambda} \int_0^T e^{-\lambda t} \|b_1(t)-b_2(t)\|^2_{L^2(0,24)} dt\end{aligned}$$ for any $\lambda>0$. Consider $L^2((0,T)\times (0,24))$ with the norm $$\|b\|=\left(\int_0^T e^{-\lambda t} \|b(t)\|^2_{L^2(0,24)} dt \right)^2, \text{ for any $b\in L^2((0,T)\times(0,24))$}.$$ Then one has $$\begin{aligned} &\|\mathscr{F}b_{1}-\mathscr{F}b_{2}\|^{2}\\ =&\displaystyle \int_0^{T} e^{-\lambda t}\|\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)(p_{b_{1}}(a,t,s)-p_{b_{2}}(a,t,s))ds da\|^{2}_{L^{2}(0,24)}dt\\ \leq & C \displaystyle \int_0^{a_\dagger} \beta^{2}(a)da \displaystyle \int_0^{T} e^{-\lambda t}\|v(t)\|^{2}_{L^2((0,a_\dagger)\times (0,24))}dt\\ \leq & \frac{C}{\lambda} \displaystyle \int_0^{a_\dagger} \beta^{2}(a)da\|b_1-b_2\|^{2},\end{aligned}$$ where $C$ is an appropriate positive constant related to $K(x,s)$. One can choose $\lambda$ large such that $\lambda >C \displaystyle \int_0^{a_\dagger} \beta^{2}(a)da $ and then $\mathscr{F}$ is a contraction mapping on $L^2((0,T)\times (0,24))$ with the norm $\|\cdot\|$. This completes the proof. From Lemma \[lemma2\], one gets that the operator $\mathscr{F}$ is a contraction mapping. Moreover, combined with Lemma \[lemma2.2\], one can follow the rest of the proof of [@As Lemma 4.1.1] to get the following comparison principle for . \[lemma3\] If $p_{i}(i\in {1,2})$ are the solutions of the following systems $$\left\{ \begin{array}{lll} Dp_{i}- \delta \Delta p_{i}+\mu_{i} (a,t,x)p_{i} = f_{i},&(a,t,x)\in Q_{a_\dagger}, \\ p_{i}(a,t,0) =p_{i}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p_{i}(a,t,0) =\partial_x p_{i}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p_{i}(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta_{i}(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p_{i}(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p_{i}(a,0,x)=p_{0i}(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24), \end{array} \right.$$ where $\mu_{1}\geq\mu_{2}$, $f_{1}\leq f_{2}$, $\beta_{1}\leq\beta_{2}$, $p_{01}\leq p_{02}$ and $\mu_1$, $\mu_2$ satisfy $(A)$, then $$0\leq p_{1}(a,t,x)\leq p_{2}(a,t,x) \ \ a.e. \ in \ Q_{a_\dagger}.$$ By referring to the proof of [@As Theorem 4.1.3, Theorem 4.1.4] for the case when $\mu(a,t,x)$ satisfies , one can define $$\mu^N(a,t,x)=\min\{\mu(a,t,x),N\}, \text{ for any $N\in\mathbb{N}^+$},$$ and denote $p_N(a,t,x)$ to be the solution of system with $\mu_N$. Passing to the limit as $N\rightarrow +\infty$ for $p_N(a,t,x)$, one can get the solution of system . Then by the results of Lemma \[lemma2\] and Lemma \[lemma3\], we can get the following lemma. \[lemma2.7\] There is a unique solution $p(a,t,x)\in L^2(Q_{a_{\dagger}})$ of system with $\mu$ satisfying . If $p_{i}(i\in {1,2})$ are the solutions of system with $\mu_1$, $f_1$, $\beta_1$, $p_{01}$ and $\mu_2$, $f_2$, $\beta_2$, $p_{02}$ respectively ($\mu_1$, $\mu_2$ satisfy ) and $\mu_{1}\geq\mu_{2}$, $f_{1}\leq f_{2}$, $\beta_{1}\leq\beta_{2}$, $p_{01}\leq p_{02}$, then $$0\leq p_{1}(a,t,x)\leq p_{2}(a,t,x) \ \ a.e. \ in \ Q_{a_\dagger}.$$ According to Lemma \[lemma2.7\] , we obtain the result of Theorem \[theorem1\] directly. Existence of an optimal control =============================== In this section, our main job is to obtain the existence of an optimal control of by Mazur’s Theorem, that is, we prove Theorem \[theorem2\]. 0.3cm Let $\Psi: U\rightarrow \mathbf{R^{+}}$ be defined by $$\Psi(u)=\displaystyle \int_{Q_{{a_{\dagger}}}}u(a,t,x)p^{u}(a,t,x)dtdxda.$$ Then by the definition of $u(a,t,x)$, we have $$\displaystyle \int_{Q_{{a_{\dagger}}}}\varsigma_{1}(a,t,x)\overline{p}(a,t,x)dtdxda\leq\Psi(u)\leq 0,$$ where $\overline{p}(a,t,x)$ is a solution of system with $u\equiv0$, $\mu\equiv0$, $\beta=\|\beta(a)\|_{L^{\infty}(0,a_\dagger)}$, $p_{0}=\|p_{0}\|_{L^{\infty}((0,a_\dagger)\times (0,24))}$, that is, $$\label{eq:333} \left\{ \begin{array}{lll} Dp- \delta \Delta p = 0,&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0 , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Thus, we can assume that $$d=\inf_{u\in U}\Psi(u),$$ and there exists a sequence $\{u_{N}\} \in U, N \in N^{*}$ such that $$d\leq \Psi(u_{N})< d+\frac{1}{N},$$ $$\label{eq:555} \Psi(u_{N})\rightarrow d.$$ Since the result of Lemma \[lemma2.7\], one obtains $$0\leq p^{u_{N}}(a,t,x)\leq \overline{p}(a,t,x) \ \ \ a.e. \ in \ Q_{a_\dagger}.$$ Thus there exists a subsequence which still be denoted by $\{u_{N}\}$ such that $$p^{u_{N}}\rightarrow p^{*} \ \ \ weakly\ in \ L^{2}(Q_{{a_\dagger}}).$$ By Mazur’s Theorem, one has that $\forall \epsilon >0$, there exists $\lambda_{i}^{N}\geq0$, $\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}=1$ such that $$\|p^{*}-\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}p^{u_{i}}\|_{L^{2}(Q_{{a_\dagger}})}\leq \epsilon,$$ where $k_{N}\geq N+1$. Now we denote $$\widetilde{p}_{N}(a,t,x)=\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}p^{u_{i}}(a,t,x),$$ therefore, $$\widetilde{p}_{N}\rightarrow p^{*} \ \ \ in \ L^{2}(Q_{{a_\dagger}}).$$ Now we consider the sequence $\{\widetilde{u}_{N} \}$ of controls $\{{u}_{i} \}$. Here $\widetilde{u}_{N} (a,t,x)$ is defined by $$\widetilde{u}_{N} (a,t,x)=\left\{ \begin{aligned} &\frac{\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}{u}_{i}(a,t,x)p^{u_{i}}(a,t,x)}{\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}p^{u_{i}}}(a,t,x),\ \ if \sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}p^{u_{i}}\neq 0,\\ &\varsigma_{1}(a,t,x),\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ if \sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}p^{u_{i}}= 0. \end{aligned} \right.$$ It is easy to check that $\widetilde{u}_{N} \in U$. Thus, one learns that there exists a subsequence $\{\widetilde{u}_{N} \}_{N\in N^{*}}$ such that $$\widetilde{u}_{N}\rightarrow u^{*} \ \ \ weakly\ \ in \ \ L^{2}(Q_{{a_\dagger}}).$$ Obviously, $\widetilde{p}_{N}(a,t,x)$ is a solution of $$\label{eq:444} \left\{ \begin{array}{lll} Dp- \delta \Delta p+\mu (a)p = \widetilde{u}_{N}(a,t,x)p,&(a,t,x)\in Q_{a_\dagger}, \\ p(a,t,0) =p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p(a,t,0) =\partial_x p(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Passing to the limit in , we get $$\left\{ \begin{array}{lll} Dp^{*}- \delta \Delta p^{*}+\mu (a)p^{*} = u^{*} p^{*},&(a,t,x)\in Q_{a_\dagger}, \\ p^{*}(a,t,0) =p^{*}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x p^{*}(a,t,0) =\partial_x p^{*}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle p^{*}(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)p^{*}(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle p^{*}(a,0,x)=p_0(a,x) , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ It means that $p^{*}$ is the solution of system corresponding to $u^{*}$. Therefore, we have $$\begin{aligned} \sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}\Psi({u_{i}})=&\sum_{i=N+1}^{k_{N}} {\lambda_{i}^{N}}\displaystyle \int_{Q_{{a_{\dagger}}}}{u}_{i}(a,t,x)p^{u_{i}}(a,t,x)dx dt da \\ =&\displaystyle \int_{Q_{{a_{\dagger}}}}\widetilde{u}_{N}(a,t,x)\widetilde{p}_{N}(a,t,x)dx dt da \\ \rightarrow &\displaystyle \int_{Q_{{a_{\dagger}}}}{u}^{*}(a,t,x){p}^{*}(a,t,x)dx dt da \\ = &\Psi(u^{*}).\end{aligned}$$ Using and the last equation, we can conclude that $d=\Psi(u^{*})$. Necessary optimality conditions =============================== In this section, our goal is to obtain the necessary optimality conditions of $(OH)$ which is Theorem \[theorem3\]. First of all, we can get that system has a unique solution $q(a,t,x)\in L^2(Q_{a_\dagger})$ by the same method as in the proof of the existence and uniqueness of solutions of system in Section 2. Since $(u^{*},p^{*})$ is an optimal pair for $(OH)$, we have $$\displaystyle \int_{Q_{{a_{\dagger}}}}u^{*}p^{u^{*}}dtdxda\leq \displaystyle \int_{Q_{{a_{\dagger}}}}(u^{*}+\epsilon v)p^{u^{*}+\epsilon v}dtdxda$$ for any $\epsilon>0$ small enough, arbitrary $v(a,t,x)\in L^{\infty}(Q_{a_{\dag}})$ such that $$\left\{ \begin{array}{lll} v(a,t,x)\leq 0,\ \ \text{if}\ \ u^{*}(a,t,x)=\varsigma_{2}(a,t,x),\\ v(a,t,x)\geq 0,\ \ \text{if}\ \ u^{*}(a,t,x)=\varsigma_{1}(a,t,x), \end{array} \right.$$ which implies $$\begin{aligned} \label{eq:00023} \displaystyle \int_{Q_{{a_{\dagger}}}}u^{*}\frac{p^{u^{*}+\epsilon v}-p^{u^{*}}}{\epsilon}dtdxda+\displaystyle \int_{Q_{{a_{\dagger}}}} vp^{u^{*}+\epsilon v}dtdxda\geq0.\end{aligned}$$ Let $z^{\epsilon}(a,t,x)=\frac{p^{u^{*}+\epsilon v}(a,t,x)-p^{u^{*}}(a,t,x)}{\epsilon}$, $y^{\epsilon}(a,t,x)=\epsilon z^{\epsilon}(a,t,x)$, then $y^{\epsilon}(a,t,x)$ satisfies $$\left\{ \begin{array}{lll} Dy^{\epsilon}- \delta \Delta y^{\epsilon}+\mu (a)y^{\epsilon} = u^{*}y^{\epsilon}+\epsilon v p^{u^{*}+ \epsilon v},&(a,t,x)\in Q_{a_\dagger}, \\ y^{\epsilon}(a,t,0) =y^{\epsilon}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x y^{\epsilon}(a,t,0) =\partial_x y^{\epsilon}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle y^{\epsilon}(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)y^{\epsilon}(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle y^{\epsilon}(a,0,x)=0 , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Multiplying the first equation by $y^{\epsilon}$ and integrating on $Q_{t}=(0,a_\dagger)\times (0,t)\times (0,24)$, one obtains $$\begin{aligned} \|y^{\epsilon}(t)\|_{L^{2}((0,a_\dagger)\times (0,24))}^{2} \leq C \displaystyle \int_0^{t} \|y^{\epsilon}(s)\|_{L^{2}((0,a_\dagger)\times (0,24))}^{2}ds+ \epsilon \displaystyle \int_{Q_t} |v| p^{u^{*}+ \epsilon v}|y^{\epsilon}|dsdxda.\end{aligned}$$ Then by the result of Lemma \[lemma2.7\] and Bellman’s Lemma (see in Appendix), we get $$\begin{aligned} \|y^{\epsilon}(t)\|_{L^{2}((0,a_\dagger)\times (0,24))}^{2} &\leq \epsilon^{2} \displaystyle \int_{Q_{{a_{\dagger}}}} |v|^{2} \overline{p}^{2}dtdxda+(1+C) \displaystyle \int_0^{t} \|y^{\epsilon}(s)\|_{L^{2}((0,a_\dagger)\times (0,24))}^{2}ds\\ &\leq \epsilon^{2} e^{(1+C)t} \displaystyle \int_{Q_{{a_{\dagger}}}} |v|^{2} \overline{p}^{2}dtdxda\end{aligned}$$ where $\overline{p}(a,t,x)$ is a solution of system , $t\in[0,T]$ and $C$ is a positive constant. This implies that $$\begin{aligned} \label{eq:00011} y^{\epsilon}\rightarrow 0 \ \ in \ \ L^{\infty}(0,T; L^{2}((0,a_\dagger)\times (0,24)) )\ \ as \ \ \epsilon\rightarrow 0^{+}.\end{aligned}$$ So the following convergence holds $$p^{u^{*}+\epsilon v}\rightarrow p^{u^{*}}\ \ in \ \ L^{\infty}(0,T; L^{2}((0,a_\dagger)\times (0,24))) \ \ as \ \ \epsilon\rightarrow 0^{+}.$$ Recalling the definition of $z^{\epsilon}(a,t,x)$, one has that $z^{\epsilon}(a,t,x)$ satisfies $$\left\{ \begin{array}{lll} Dz^{\epsilon}- \delta \Delta z^{\epsilon}+\mu (a)z^{\epsilon} = u^{*}z^{\epsilon}+v p^{u^{*}+ \epsilon v},&(a,t,x)\in Q_{a_\dagger}, \\ z^{\epsilon}(a,t,0) =z^{\epsilon}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x z^{\epsilon}(a,t,0) =\partial_x z^{\epsilon}(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle z^{\epsilon}(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)z^{\epsilon}(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle z^{\epsilon}(a,0,x)=0 , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ Let $h^{\epsilon}(a,t,x)=z^{\epsilon}(a,t,x)-z(a,t,x)$, where $z(a,t,x)$ is a solution of the following system $$\left\{ \begin{array}{lll} Dz- \delta \Delta z+\mu (a)z = u^{*}z+v p^{u^{*}},&(a,t,x)\in Q_{a_\dagger}, \\ z(a,t,0) =z(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x z(a,t,0) =\partial_x z(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle z(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)z(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle z(a,0,x)=0 , &(a,x)\in (0,{a_\dagger}) \times (0,24). \end{array} \right.$$ Following the above proof step by step, we can get that $$\begin{aligned} \|h^{\epsilon}(t)\|_{L^{2}((0,a_\dagger)\times (0,24))}^{2}\leq e^{(1+C)t} \displaystyle \int_{Q_{{a_{\dagger}}}} |v|^{2} |y^{\epsilon}|^{2}dtdxda.\end{aligned}$$ Using , one obtains $$z^{\epsilon}\rightarrow z\ \ in \ \ L^{\infty}(0,T; L^{2}((0,a_\dagger)\times (0,24)))\ \ as \ \ \epsilon\rightarrow 0^{+}.$$ Passing to the limit in , it follows $$\begin{aligned} \label{eq:00024} \displaystyle \int_{Q_{{a_{\dagger}}}}u^{*}zdtdxda+\displaystyle \int_{Q_{{a_{\dagger}}}} vp^{u^{*}}dtdxda\geq0,\end{aligned}$$ for arbitrary $v(a,t,x)\in L^{\infty}(Q_{a_{\dag}})$ such that $$\left\{ \begin{array}{lll} v(a,t,x)\leq 0,\ \ \text{if}\ \ u^{*}(a,t,x)=\varsigma_{2}(a,t,x),\\ v(a,t,x)\geq 0,\ \ \text{if}\ \ u^{*}(a,t,x)=\varsigma_{1}(a,t,x). \end{array} \right.$$ Multiplying the first equation of system by $z(a,t,x)$ and integrating on $Q_{a_{\dagger}}$, we get $$\begin{aligned} \label{eq:00025} \displaystyle \int_{Q_{{a_{\dagger}}}}vp^{u^{*}}qdtdxda=\displaystyle \displaystyle \int_{Q_{{a_{\dagger}}}}u^{*}zdtdxda.\end{aligned}$$ Combining with , we learn that $$\begin{aligned} \displaystyle \int_{Q_{{a_{\dagger}}}}vp^{u^{*}}(q+1)dtdxda\geq 0,\end{aligned}$$ for arbitrary $v(a,t,x)\in L^{\infty}(Q_{a_{\dag}})$ such that $$\left\{ \begin{array}{lll} v(a,t,x)\leq 0,\ \ \text{if} \ \ u^{*}(a,t,x)=\varsigma_{2}(a,t,x),\\ v(a,t,x)\geq 0,\ \ \text{if} \ \ u^{*}(a,t,x)=\varsigma_{1}(a,t,x). \end{array} \right.$$ For any $(a,t,x)\in Q_{a_{\dag}}$, if $p^{u^{*}}(a,t,x)\neq 0$ holds, we can conclude that $$u^{*}(a,t,x)=\left\{ \begin{aligned} &\varsigma_{1}(a,t,x),\ \ \text{if}\ \ q(a,t,x)>-1,\\ &\varsigma_{2}(a,t,x),\ \ \text{if}\ \ q(a,t,x)<-1. \end{aligned} \right.$$ We now consider the set $B=\{(a,t,x)\in Q_{a_{\dag}} | p^{u^{*}}(a,t,x)= 0\}$. Take any function $w(a,t,x)\in L^{\infty}(Q_{a_{\dag}})$ such that $w(a,t,x)\neq 0$ for $(a,t,x) \in B$ and $w(a,t,x)\equiv 0$ for $(a,t,x) \in Q_{a_{\dag}}-B$ and $u^*+w\in U$. Let $z(a,t,x)=p^{u^*+w}-p^{u^*}$ and then it satisfies $$\left\{ \begin{array}{lll} Dz- \delta \Delta z+\mu (a)z = u^*(a,t,x)z+w(a,t,x)z,&(a,t,x)\in Q_{a_\dagger}, \\ z(a,t,0) =z(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \partial_x z(a,t,0) =\partial_x z(a,t,24),&(a,t)\in (0,a_\dagger)\times (0,T), \\ \displaystyle z(0,t,x)=\displaystyle \int_0^{a_\dagger} \beta(a)\displaystyle \int_{x-\eta}^{x+\eta}K(x,s)z(a,t,s)ds da,&(t,x)\in (0,T)\times (0,24), \\ \displaystyle z(a,0,x)=0 , &(a,x)\in (0,{a_\dagger} ) \times (0,24). \end{array} \right.$$ By the uniqueness result, one can infer that $z(a,t,x)\equiv 0$ a.e. in $Q_{a_{\dagger}}$. This implies that we can change $u^*$ in $B$ with arbitrary values in $[\varsigma_1(a,t,x),\varsigma_2(a,t,x)]$ and the value of the related cost functional of $(OH)$ remains the same. Then the conclusion is obvious and the proof is complete. Appendix ======== We present here a well-known result of Bellman in [@As]. \[lemma22\] If $x\in C([a,b])$, $\psi \in L^{1}(a,b)$, $\psi(t)\geq0 \ a.e. \ t \in(a,b)$, $M \in R$ and for each $t\in[a,b]$, $$x(t)\leq M+\displaystyle \int_a^{t}\psi(s)x(s)ds,$$ then $$x(t)\leq M \exp\left(\displaystyle \int_a^{t}\psi(s)d s\right).$$ [7]{} Aniţa L I, et al. Optimal harvesting for periodic age-dependent population dynamics with logistic term. Applied Mathematics and Computation, 2009, 215(7): 2701-2715. Athans M, Falb P L. Optimal control: an introduction to the theory and its applications. Courier Corporation, 2013. Aniţa S. Analysis and control of age-dependent population dynamics. Springer Science & Business Media, 2000. Aniţa S. Optimal control of a nonlinear population dynamics with diffusion. Journal of Mathematical Analysis and Applications, 1990, 152(1): 176-208. Aniţa S. Optimal harvesting for a nonlinear age-dependent population dynamics. Journal of Mathematical Analysis and Applications, 1998, 226(1): 6-22. Brokate M. Pontryagin’s principle for control problems in age-dependent population dynamics. Journal of Mathematical Biology, 1985, 23(1): 75-101. Becker N, Petric D, Boase C, et al. Mosquitoes and their control. New York: Springer, 2003. Fister K R, Lenhart S. Optimal control of a competitive system with age-structure. Journal of Mathematical Analysis and Applications, 2004, 291(2): 526-537. Garroni M G, Langlais M. Age-dependent population diffusion with external constraint. Journal of Mathematical Biology, 1982, 14(1): 77-94. Gurtin M E, Murphy L F. On the optimal harvesting of age-structured populations: some simple models. Mathematical Biosciences, 1981, 55(1-2): 115-136. Gurtin M E, Murphy L F. On the optimal harvesting of persistent age-structured populations. Journal of Mathematical Biology, 1981, 13(2): 131-148. Gurtin M E. A system of equations for age-dependent population diffusion. Journal of Theoretical Biology, 1973, 40(2): 389-392. Huho B, et al. Consistently high estimates for the proportion of human exposure to malaria vector populations occurring indoors in rural Africa. International journal of epidemiology, 2013, 42(1): 235. Kirk D E. Optimal control theory: an introduction. Courier Corporation, 2012. Lewis F L, Vrabie D, Syrmos V L. Optimal control. John Wiley & Sons, 2012. Leitmann G. The calculus of variations and optimal control: an introduction. Springer Science & Business Media, 2013. Li L L, Ferreira C P, Ainseba B E. Large time behaviour of the solution for an age-structured population model, to appear. Li L L, Ferreira C P, Ainseba B E. Local exact controllability of an age-structured problem modelling mosquito plasticity, to appear. Maxmen A. Malaria surge feared: the WHO releases action plan to tackle the spread of insecticide-resistant mosquitoes. Nature, 2012, 485(7398): 293-294. Park E J, et al. Optimal harvesting for periodic age-dependent population dynamics. SIAM Journal on Applied Mathematics, 1998, 58(5): 1648-1666. Sougoufara S, et al. Biting by Anopheles funestus in broad daylight after use of long-lasting insecticidal nets: a new challenge to malaria elimination. Malaria journal, 2014, 13(1): 125. Trape J F, et al. Malaria morbidity and pyrethroid resistance after the introduction of insecticide-treated bednets and artemisinin-based combination therapies: a longitudinal study. The Lancet infectious diseases, 2011, 11(12): 925-932. Webb G F. Theory of nonlinear age-dependent population dynamics. CRC Press, 1985. World Health Organization. World malaria report 2008. World Health Organization, 2008. World Health Organization. World malaria report 2012. World Health Organization, 2012. Zhao C, Wang M, Zhao P. Optimal control of harvesting for age-dependent predator-prey system. Mathematical and computer modelling, 2005, 42(5-6): 573-584.

ArXiv

--- abstract: 'We present time-resolved spectroscopy of the optical counterpart to the proposed ultra-compact binary system V407 Vul (=RX J1914.4+2456). Our Gemini spectra resolve the 9.48 minute periodicity that has previously been reported for this source. We find that the optical counterpart is dominated by a reddened late-type spectrum of type G9V and contains solely unresolved absorption features. No radial velocity signatures exceeding 10 km/s could be detected on periods from minutes to hours. Using interstellar extinction estimates, we derive a distance to the G9 star of $1.1-3.5$ kpc. In addition to this stellar spectrum, we detect a blue component that modulates solely on the 9.48 minute period, and peaks $\sim 0.15$ in phase ahead of the X-ray peak. This blue component which contributes up to 40% of the light shows no evidence for emission line features that are the usual hallmarks of an interacting binary. Good seeing images obtained with the Magellan telescopes indicate that the variable and the G-star are aligned to better than 0.1". Despite the low probability of a chance alignment of a field star along the line of sight, the G9 light cannot be directly associated with the 9.48 minute variable that powers the luminous ($\sim$10$^{35}$ erg/s) and highly variable X-ray source. The outlook for the detection of conclusive radial velocity measurements will remain challenging due to the extinction along the line of sight in conjunction with the contaminating effect of the G9V star.' author: - 'D.Steeghs' - 'T.R.Marsh, S.C.C.Barros' - 'G.Nelemans, P.J.Groot, G.H.A.Roelofs' - 'G.Ramsay, M.Cropper' title: 'GEMINI spectroscopy of the ultra-compact binary candidate V407 Vul[^1]' --- Introduction ============ ![Top: Average optical spectrum of V407 Vul revealing a reddened late G to early K star absorption line spectrum. Small gaps near 5000 and 5950Å are due to the gaps in between the individual CCD detectors. The y-axis represents the mean counts s$^{-1}$Å$^{-1}$ obtained in the 100s exposures, no flux calibration was performed. Bottom panels compare the spectrum of V407 Vul with that of a G9V star sampled at the same resolution revealing a remarkable match. Dotted lines indicate the location of two diffuse interstellar bands that reflect reddening towards the source. \[fig1\]](f1.eps){width="8.5cm"} Double white dwarf binaries form one of the largest populations of compact binaries with a few hundred million systems expected across the Galaxy (Napiwotzki et al. 2003). A substantial fraction of these are brought into contact at very short orbital periods (minutes) through angular momentum loss via gravitational wave emission. In the majority of cases, a rapid merger ensues due to unstable mass transfer (Webbink 1984, Iben 1990). If the combined mass of such double white dwarfs exceeds the Chandrasekhar mass, a Type Ia supernova may be produced. A subset of such double white dwarfs are expected to survive the onset of mass transfer and become stable semi-detached binaries where the more massive white dwarf accretes from its companion via Roche-lobe overflow. Such binaries are initially formed at orbital periods of a few minutes and will evolve towards longer orbital periods as mass transfer drives the two stars apart (Nelemans et al. 2001; Marsh, Nelemans & Steeghs 2004). Observationally a small group of such accreting white dwarfs are known, the so-called AM CVn stars. These are hydrogen-deficient cataclysmic variables, indicating that their mass donor stars are degenerate. Although such systems can also be produced using a low-mass helium star donor (Savonije et al. 1986) or evolved post main-sequence donors (Podsiadlowski et al. 2002), the double white dwarf formation channel is the only one that can produce systems with orbital periods shorter than 10 minutes. Recently, three variable stars have been identified as candidate accreting binary systems with ultra-short orbital periods; V407 Vul (Cropper et al. 1998; Ramsay et al. 2000), ES Cet (Warner & Woudt 2002) and RX J0806.3-0939 (Ramsay et al. 2002b; Israel et al. 2002). Optical and X-ray light curves reveal large amplitude variations on stable periods of 9.5, 10.3 and 5.3 [*minutes*]{} respectively, but on no other periods, suggesting that these may be their orbital periods. V407 Vul was initially discovered as a [*ROSAT*]{} X-ray source (RXJ1914.4+2456; Haberl & Motch 1995), displaying a stable X-ray periodicity of 569s which was suggested to represent the spin period of an accreting magnetic white dwarf in an intermediate polar (IP, Motch et al. 1996). Cropper et al. (1998) first suggested that V407 Vul might be a double-degenerate binary with the X-ray period representing its orbital period. This was motivated by the soft X-ray spectrum, the shape of the X-ray lightcurve and the lack of other periodicities which is very atypical for an IP. In this model, the accreting white dwarf is magnetic and synchronised with the binary orbit. When the optical counterpart was identified (Ramsay et al. 2000), the phasing of the optical modulation, the lack of strong emission lines and the absence of polarised light raised new questions concerning the structure of V407 Vul. Several alternative models were suggested which include a non-magnetic model where the accretion stream impacts the primary directly due to the compact nature of double-degenerate binaries (Marsh & Steeghs 2002, Ramsay et al. 2002a), a unipolar inductor model analogous to the Jupiter-Io system (Wu et al. 2002) or a stream-fed IP viewed nearly face-on (Norton, Haswell & Wynn 2004). Only in the IP interpretation would the 569s period not be the orbital period but instead be the spin-period of the primary. Strohmayer (2002,2004) analysed archival [*ROSAT*]{} and [*ASCA*]{} data as well as new [*Chandra*]{} X-ray observations and found that the X-ray period was slowly decreasing which appears at odds with an accreting ultra-compact binary model since the period change is of the wrong sign. Although this appeared to support the non-mass transfer model of Wu et al. (2002), we may simply be seeing short term period changes that commonly occur in accreting binaries (Marsh & Nelemans 2005). In order to clarify the nature of V407 Vul and compare its characteristics with the other AM CVn systems, we obtained high-time resolution optical spectroscopy using the [*Gemini-North*]{} telescope that resolves the 9.48 minute period. In Section 2, we discuss the observations and data reduction and present the key observational results in Section 3. In Section 4 we discuss the implications of our observations concerning the nature of V407 Vul. Observations ============ Gemini spectroscopy ------------------- We employed the [*GMOS*]{} spectrograph (Hook et al. 2004) mounted on the 8.1m Gemini-North telescope on Mauna Kea, Hawaii and obtained time-series spectroscopy of V407 Vul in queue observing mode. The detector array consisting of three 2k$\times$4k CCDs was operated in $4\times$4 binning mode to reduce readout overheads and sample the spectral resolution element achieved with our 0.75” slit across 2.3 binned pixels. The B600 grating delivered spectra covering 4058-6920Å at 1.8Å/pixel, with two small gaps near 5000Å and 5950Å at the interfaces between the individual CCDs. The slit was kept at a fixed position angle in order to accommodate both V047 Vul and a nearby comparison star in the slit. Weather conditions were non-photometric with a few exposures affected by thin cirrus while the seeing was between 0.75-0.8". Table 1 provides a log of observations. [lccccc]{} UT Date & exp. & readout &\# exp. & UT interval & cycles\ Aug 9 2002 & 100s & 14s & 84 & 06:30 - 09:19 & 17.8\ Sep 2 2002 & 100s & 11s & 84 & 05:54 - 08:34 & 17.8\ Sep 4 2002 & 100s & 12s & 84 & 06:14 - 08:57 & 17.2\ totals & & & 252 & & 52.8\ All frames were first de-biased using the overscan strips. Tungsten exposures obtained at each observing block with the telescope pointed at our target served to construct a flat-field frame that was used to perform pixel-to-pixel sensitivity corrections. The spectra of both our primary target V407 Vul as well as the comparison star were then optimally extracted from each frame. Science frames were bracketed by CuAr exposures that were used to wavelength calibrate the spectra by interpolating the arc-scale derived from the two nearest arcs. Finally, the simultaneously recorded comparison star spectra were used to correct the V407 Vul spectra for time and wavelength dependent slit-loss effects by dividing the individual comparison star spectra to the master comparison star spectrum obtained at lowest air-mass. A second order polynomial was then fitted to these ratio spectra and applied to the V407 Vul spectra. All relevant error contributions were properly propagated throughout the reduction process to produce final science spectra with accurate statistical errors. This reduction recipe delivered a total of 252 science spectra covering close to 53 cycles of the proposed 9.48 min orbital period. Analysis ======== The puzzling optical spectrum of V407 Vul ----------------------------------------- The grand average optical spectrum obtained by averaging all 252 exposures with equal weights is shown in Figure 1. It reveals a reddened late-type stellar spectrum with a forest of narrow absorption lines, including strong and narrow H$\alpha$ absorption. In line with the low S/N spectrum of Ramsay et al. (2002a), no emission line features can be discerned. The narrowness of the observed absorption lines, which we do not resolve at our spectral resolution of 4Å, is a surprise since an origin in a binary system would broaden spectral features considerably in an average spectrum due to binary motion. The optical spectrum of V407 Vul thus appears to be consistent with a normal stellar spectrum. By comparing the average spectrum with late-type templates, we estimate its spectral type to be close to G9/K0. Although we are somewhat limited in determining the luminosity class at our spectral resolution, the lack of giant features suggests a main-sequence classification. To illustrate, we compare the average spectrum to that of a G9V star in Figure 1 which indeed provides a very good match to all the observed features. It is clear that it is not possible to fit a G9 star into a 9.48 minute binary system, and this identification thus intially appears in conflict with an ultra-compact binary scenario. No emission features can be discerned at the locations of strong hydrogen and helium lines in our average spectrum, nor is there evidence that the Balmer absorption lines are filled in by weak emission. The signal to noise in our average spectrum, which represents a total of 7 hours worth of exposure time, is around 125 per pixel at $H\alpha$. This is therefore large enough to be sensitive to even very weak lines. In all respects, the optical spectrum of V407 Vul resembles that of a single star and appears at odds with the observed X-ray luminosity and variability that has been reported. Time variability ---------------- In order to look for the expected variability on the 9.48 minute period previously reported in X-rays and the optical, we constructed lightcurves by summing over our observed spectral range and calculated Scargle-normalised power spectra (Scargle 1982). These clearly reveal a coherent intensity modulation on the previously reported 9.48 min period with no significant signals detected at other periods. Although our uneven time-sampling limits a sensitive search for weak side-bands, we do not find any evidence for them near the strong 9.48 min peak. We thus find very similar optical variability behaviour as was discussed in Ramsay et al. (2002a). The variability does confirm that we have acquired spectra of the correct optical counterpart, ruling out a possible mis-identification. A close inspection of the GMOS acquisition images near the slit position revealed no contaminating field stars that could have contributed to the spectra. Although we detect a 5-15% amplitude intensity modulation in our data, there appears to be no corresponding radial velocity motions. Cross-correlating the individual 100s exposures reveals no significant radial velocity shift in any of the 252 spectra that were obtained in three hour long blocks over a baseline of 26 days. Since our time-series cover of order 53 cycles of the 9.48 min period, we phase-folded our data on the X-ray period using the ephemeris reported by Strohmayer (2004) in order to increase our sensitivity. The ephemeris was converted into the heliocentric frame and the UTC time-system and includes the measured period derivative term. ![Optical lightcurves after folding on the 9.48 min period using 15 phase bins. Two cycles are plotted obtained in three wavelength bands: red (6200-6700Å), middle (5300-5800Å) and blue (4100-4600Å). The y-scale is adjusted such that each panel spans $\pm$ 20 % around the mean. Bottom panel are radial velocity shifts derived by cross-correlating the absorption line spectrum. \[fig2\]](f2.eps){width="8.5cm"} We then constructed light-curves of various spectral segments after phase-folding, revealing that the peak-to peak amplitude increases from 9% in the red part of our spectral range up to 25% in the blue end (Figure 2). As remarked in previous studies, the optical peaks earlier in phase relative to the X-ray peak. In our case, this was 0.15 cycles in advance of the Strohmayer (2004) X-ray ephemeris. In order to quantify the lack of radial velocity motions in the absorption lines across the 9.48 minute period, we also cross-correlated each phase-folded spectrum with the mean spectrum. The derived radial velocities are also plotted in Figure 2 and reveal no motions phased with the 9.48 min period. All radial velocities are consistent with zero at all phases. In order to place an upper limit on any possible radial velocity amplitude, we shifted the spectra by various known amounts and tried to recover the shifts using the above procedures. We found that we can rule out any ’orbital’ radial velocity shifts across the 9.48 minute period with amplitudes larger than $\sim$ 10 km s$^{-1}$. Radial velocity limits on the period ------------------------------------ Our radial velocity time series revealed no detectable radial velocity motions in the G-star spectra whatsoever. We ran a period detection algorithm in order to evaluate our detection limits on any periodic motions in the G-star. We define a detection as occurring when the $\chi^2$ for a fit of a constant to the velocities is sufficiently large as to have a very small probability $p$ of occurring by chance (we took $p = 10^{-3}$). For a given orbital period, we then calculate the orbital speed of the G star induced by the variable taking the masses to be $\sim0.8 M_{\odot}$ for the G-star and $\sim0.6 M_{\odot}$ for the white dwarf primary. Orbital velocities for a 9.48 minute binary are of order 1000 km s$^{-1}$, but even for periods as long as 10 hours would still be larger than 100 km s$^{-1}$. This translates into a lower limit on the orbital inclination $i_c$ which will trigger a detection. Assuming randomly inclined orbits, the detection probability is then $cos(i_c)$. Averaging this over the zeropoint of the orbital phase assuming it is uniformly distributed from 0 to 1 gives us the overall detection probability for a particular orbital period. The observational uncertainties in the velocities are taken into account through Monte Carlo generation of orbital velocities plus noise, in the form of Gaussian white noise. Finally, we re-binned the resulting detection probabilities into period bins spanning 0.05 in $\log{P}$ for periods ranging between a few minutes and 100 days. These detection probabilities are plotted in Fig.3. Our detection probability is over 99.9% for any periods under $\sim3$ hours and remains better than 99% for periods up to $\sim14$ hours. This thus argues against identifying the G9-type star that dominates the optical light as the donor star in a compact binary system. We return to the implications of these constraints in Section 4. Decomposing the variable part ----------------------------- In order to decompose the spectrum and determine the spectral characteristics of the variable part of the light, we fitted a sine-curve to the time-series of each wavelength pixel while fixing the period to the known X-ray period of 9.48 min. We thus derived the semi-amplitude of the modulation as a function of wavelength. This modulation spectrum is plotted in Fig.4 and shows a relatively flat modulation semi-amplitude of 5% in the red, rising steeply and approaching 20% on the blue end, in line with the modulations displayed by the folded lightcurves in Fig.2. No spectral features, either in absorption or emission, are detected in this modulation spectrum with significance, apart from a few bright pixels that correspond to a small amount of contamination due to strong sky lines. In particular, near the location of potential Hydrogen and Helium features, we see no evidence for any excess variability. It thus appears that also the variable part of the light shows no evidence for emission lines that are the hallmarks of an accreting binary. Although our sensitivity to weak lines in the modulated spectrum is obviously worse than our ability to see them in the average spectra, we still have the sensitivity to pick up emission lines at the strengths observed in typical cataclysmic variables. For example, at H$\alpha$, we have the sensitivity to pick up emission lines with equivalent widths larger than $\sim$ 4Å($5\sigma$ limit). The above decomposition of a non-varying G-star spectrum plus a variable blue component appears to describe our time-series fully. We checked whether there was any evidence for a non-variable component in addition to the two components described above by studying the depth of the G-star absorption lines relative to the continuum. Any additional light source would dilute the absorption lines relative to that of an uncontaminated G-star. We find that this dilution is at most 10-15% at the red end of our spectral range, rising to 30-40% at the very blue end. This is consistent with the full amplitude of the variable component and indicates that any additional contribution to the light is rather small. Reddening and distance ---------------------- The line of sight toward V407 Vul suffers from considerable extinction although the estimates vary significantly. According to the dust maps of Schlegel et al. (1998), the integrated line of sight column is $N_H = 3.2 \times 10^{21}$ cm$^{-2}$ using the $N_H/E(B-V)$ scaling as derived by Predehl & Schmitt (1995). Joshi (2005) derived a mean extinction of $A_{V} \sim 1.5$ mag kpc$^{-1}$ ($\sim 2.7 \times 10^{21}$ cm$^{-2}$ kpc$^{-1}$) in the region towards V407 Vul though with considerable local scatter. Spectral fits to the X-ray spectra from various missions have also lead to a range of derived column densities. Haberl & Motch (1995) determined a column of $N_H = 1.0 \times 10^{22}$ cm$^{-2}$ using an absorbed black-body fit to the [*ROSAT*]{} spectrum, although Ramsay et al. (2000,2002a) remark that the $N_H$ is not closely constrained from fits to either [*ASCA*]{} or [*ROSAT*]{} data. More recently, Ramsay et al. (2005) find $N_H = 4.2 \times 10^{21}$ cm$^{-2}$ from fits to [*XMM-Newton*]{} data, and a re-analysis of the same data in Ramsay, Cropper & Hakala (2006) gives an even lower column of $N_H = 1.2 \times 10^{21}$ cm$^{-2}$ when applying a different plasma model. This range in reported extinction spans nearly an order of magnitude and has considerable impact on the derived X-ray luminosities, which range from $3 \times 10^{31}$ to $10^{35}$ ergs s$^{-1}$. In our data, a substantial amount of reddening is apparent through the steep slope of the optical spectrum shown in Fig.1. No flux standard data was available that would permit us to calibrate the Gemini spectra. However, our wavelength range includes several diffuse interstellar bands that can serve as reddening indicators (DIBs; Fig 1). The ratio between the two components in the prominent Na-D doublet near 5895Å was close to 1, indicating saturation. This doublet thus can only provide a lower limit of $E(B-V)>0.5$ ($N_H > 3 \times 10^{21}$ cm$^{-2}$). However, the 5780Å DIB is present with an $EW(5780)=0.7 \pm 0.1$Å. Using the $E(B-V)$ vs $EW$ relation as published in Herbig (1993), this translates to a $E(B-V) = 1.4 \pm 0.2$ ($N_H \sim 7 \pm 1 \times 10^{21} $cm$^{-2}$) towards V407 Vul. These values are subject to the accuracy of the calibrations used, for which significant variations along specific lines of sight are possible. The DIB features in our mean spectrum thus suggests that the reddening towards the G-star is substantial. We also determined the strengths of these DIBs in the comparison star that was present in our slit and find equivalent widths consistent with those of V407 Vul. As a final consistency check, we tried to reproduce the observed broad-band colors of V407 Vul as listed in Ramsay et al. (2002a). We were not able to fit all the colors satisfactorily with a model consisting of a reddened G9-star plus a blue component. Particularly puzzling are the near-infrared colors, which should be dominated by the G9-star yet cannot be described by reddened stellar models. A reasonable match to the optical colors can be achieved if we decompose the colors into a reddened G9-star with $N_H \sim 3-4 \times 10^{21}$ cm$^{-2}$ plus a blue component with contributions in line with Fig.4. Models with reddening values as large as those inferred from the 5780Å DIB would overpredict the contribution from the variable in the blue bands considerably. Given V407 Vul’s observed $R$ magnitude of $R=19.2 \pm 0.1$ (Ramsay et al. 2002a), this leads to a de-reddened $R$-band magnitude of $15.5-17.9$ using the Schlegel et al. (1998) extinction scaling and our derived range of $N_H=3-7 \times 10^{21}$ cm$^{-2}$. Since the G9 star light totally dominates the R-band, we can estimate the distance to this star using a G9 absolute magnitude of $M_R(G9V) \sim 5.2 \pm 0.3$ (Houk et al. 1997) and find $d= 1.1-3.5$ kpc. We remark that the uncertainty in the derived distance is completely dominated by the uncertainty in the assumed $N_H$. Since we cannot firmly rule out a (sub)giant classification, these are strict lower limits on the distance. According to Neckel & Klare (1980), the interstellar absorption ramps up rapidly within 1kpc in this region and our distance and hydrogen column from the DIBs are thus in that sense consistent, while it would be puzzling if the distance was much larger yet the reddening much lower. Whether this represents the distance to the variable optical/X-ray source is not certain since the observed properties of the G9 star make it impossible to associate it directly with either the X-ray source or a member in a compact binary. Our interstellar bands sample the column towards the G star, and the apparent disparity with the recent X-ray inferred columns could indicate that the X-ray source is not at the same distance as the G9 star, but significantly closer. The observed X-ray flux shows significant variability (Ramsay et al. 2000,2005,2006, Strohmayer 2004). If we assume that the X-ray source is at the same distance as the G9-star, the inferred bolometric X-ray luminosity is $10^{33}-10^{35}$ erg s$^{-1}$ depending on the spectral model used. This is clearly far beyond any chromospheric X-ray emission that could be powered by a G9 star, let alone be modulated at 9.48 minutes. Chance alignment constraints ---------------------------- Given the puzzling properties of the G9 starlight that dominates the optical spectrum of such a strong and variable X-ray source, we investigated the possibility whether the G9 star is merely a chance alignment along the line of sight, and is not actually associated with the 9.48 minute variable. On June 23 2003, we obtained $BVRI$ images of the field surrounding V407 Vul with the MAGIC camera on the Magellan-Clay telescope at Las Campanas Observatory. The pixel scale was 0.069“ and seeing conditions were very good, achieving a point source FWHM = 0.45”. The optical counterpart to V407 Vul was detected as a non-resolved point source in all images. We also compared the position of the counterpart relative to nearby field stars and found that these positions matched to better than 1 pixel when comparing B,V,R and I-band frames. We therefore conclude that the source dominating in the I-band (the G9 star), and the variable source that contributes $\sim$ 30% to the light in the B-band are aligned to within 0.1“. By counting all stars in our imaging data down to the brightness of V407 Vul, we derive a local field star density of $4.6 \times 10^{-3}$ stars arcsec$^{-2}$. The probability of chance alignment of a field star at the level of $<0.1$” as derived from our imaging analysis amounts to $\le 1.4 \times 10^{-4}$. These numbers are consistent with the even tighter alignment constraints derived in Barros et al. (2006, in prep) based on high-speed optical photometry of V407 Vul. While the lack of emission lines or radial velocity motions argues against relating the G9 dominated spectrum directly with the source of the variable light, this small probability suggests that a physical connection needs to be considered. Discussion ========== Our phase-resolved spectroscopy reveals that the optical light of V407 Vul is dominated by a G9V-type stellar spectrum, viewed through a column of $N_H= 3-7 \times 10^{21}$ cm$^{-2}$ and at a distance of $d= 1.1-3.5$ kpc. Good seeing images indicate that this G9 star is very closely aligned with the 9.48 minute variable source making a chance alignment with a field star unlikely. However, the complete lack of radial velocity signatures on periods from minutes to hours and the absence of any emission features appears to rule out that this G9 star is a member of a compact binary that powers a $10^{35}$ erg s$^{-1}$ variable X-ray source. Although the G9 totally dominates our spectra in the red end, the broad-band colors of V407 Vul cannot be modeled with a simple reddened G9 star model. Ramsay et al. (2000,2002a) also discuss this issue and find that the infrared colors are not consistent with reddened late-type stars. However, the magnitudes in the various colors were not obtained simultaneously and Ramsay et al. (2000) report a change in the K-band magnitude of 1.1 mag between two epochs. Such significant variability and the apparent mismatch of colors in the IR presents yet another puzzling aspect that does not fit obviously with the apparently non-varying G9 starlight that dominates our spectra. Perhaps another spectral component starts to contribute towards the red. A contemporaneous study of the optical/IR colors and variability appears to be warranted so that these characteristics can be reconciled. A triple system? ---------------- A possibility could be that the G9 star is part of a triple system, consisting of the G9 star in a relatively wide orbit around the compact binary that is responsible for the observed variability and X-ray emission. This would lead to no obvious radial velocities on short periods, nor would there be emission lines since the G9 star is relatively far removed from the compact binary. Although bringing in their own formation and stability concerns, triple star systems are relatively abundant and the low mass X-ray binary V1727 Cyg, for example, is thought to be a hierarchical triple (Garcia et al. 1989). An interesting consequence of a triple scenario is that the observed period derivative of V407 Vul (Strohmayer 2002, 2004, Ramsay et al. 2005,2006) may then be influenced or even dominated by orbital time-delay effects due to the G9 star. In that case, the observed period derivative cannot be interpreted as tracking the secular mass transfer rate of the ultra-compact binary. This makes the period derivative a less clear-cut test of the accretion geometry and the expected angular momentum loss rate from gravitational wave emission (see also Marsh & Nelemans 2005). If the observed period drift instead tracks the time-delays due to a much longer and wider orbit, we would expect cyclical period changes modulated with this long orbital period. Given that the current period drift of $\sim$400s over $\sim$4,000 days shows no sign yet of turning over (e.g Ramsay et al. 2006) we have only covered a modest part of the orbit during the 11 years of period studies. This in turn implies that the orbital period of the G9 star in this scenario would be larger than $\sim$25 years or so, well beyond what is needed for dynamical stability. Alternatively, despite the low probability, the G9 may be physically unrelated to the 9.48 minute variable but happens to be extremely well aligned. The X-ray source could then indeed be much closer, which would alleviate the issue of the low hydrogen columns that are inferred from the X-ray spectra. Either way, it is clear that the G9 star cannot be associated directly with the intriguing variable source that makes V407 Vul so fascinating. It unfortunately severely inhibits our ability to study the properties of this enigmatic variable. We remark that this also implies that care must be taken when interpreting observed magnitudes and modulation amplitudes since the observed light is a composite between the G9 star and the blue variable. Despite the dominance of the G9 light, we were able to clearly detect a blue component that modulates coherently on the previously reported 9.48 minute period, with no evidence of variability at other periods sampled by several efforts at different wave-bands to date. This component starts to make an appreciable contribution in the blue end of our spectral range and it appears that despite the significant reddening along the line of sight, it is at the blue end that we may be able to constrain the properties of V407 Vul. As discussed in Section 3.3., even this variable component to the light shows no evidence for emission line features that would be expected for a mass transferring compact binary. If, as the most recent X-ray analysis may suggest, the X-ray source actually suffers from much less reddening compared to the G9 star, deeper UV studies are feasible and could provide important constraints on the bolometric flux from the variable. Since the discovery of V407 Vul as a variable X-ray source, a number of models have been put forward for V407 Vul and its related short period cousins RXJ0806 and ES Cet. We do not wish to repeat all the pros and cons of the various models in detail, but will focus on those features relevant for the spectroscopy results presented here. A stream-fed IP? ---------------- Most scenarios identify the sole periodicity that has been detected in these systems as the orbital period of an ultra-compact binary. An alternative model put forward by Norton et al. (2004) suggests that these periods represent the beat period between spin and orbit of a magnetic white dwarf viewed at very low inclinations and accreting from a stream instead of an accretion disc (a so called stream-fed IP). The orbital period of the binary in the stream-fed IP scenario would be 1.5-2.5 hours for V407 Vul, which would lead to orbital velocities of $>$ 250 km s$^{-1}$. Our period analysis in Section 3 also illustrates that our detection probability for periods in the IP range is very high, yet we see no evidence for orbital motion. Given our measurement limits, the inclination would then have to be less than two degrees to explain the lack of measurable radial velocities. An additional issue is the total lack of emission features. Although Norton et al. argue that a narrow stream feeding the WD may be optically thick and this would not produce emission lines, this is not what is observed in other accreting systems. Despite the presence of optically thick disks in non-magnetic CVs and the presence of dense, optically thick streams in magnetic polars and IPs, all these systems invariably shows very strong emission lines from optically thin regions above those disks and streams. The situation would be very similar to that of high-state polars where accretion stream emission is visible out to the L1 point (e.g. Schwope et al., 1997). Indeed, the prototypical stream-fed IP V4200 Oph shows very strong hydrogen and helium emission features (Hellier & Beardmore 2002). We have the sensitivity to see such strong lines both in the grand average spectrum (Fig.1) as well as the variable part of the light (Fig.3) and see no such lines in either. In summary, we see no evidence for the two tell-tale signs that would be expected for a low-inclination IP (radial velocities on $P_{orb}$ and emission lines). The continued lack of evidence for additional periodicities in the now rather extensive number of observations at different wavebands and the atypical X-ray properties of V407 Vul are additional issues which reduce the appeal of an IP interpretation. An ultra-compact binary hiding behind a star? --------------------------------------------- All remaining scenarios have identified 9.48 minutes as the orbital period of the binary. As argued above, this would imply that the G9 star is not directly related to the ultra-compact binary and merely dilutes and complicates the analysis of the underlying variable. In the double-degenerate polar model of Cropper et al. (1998) and the direct-impact model (Nelemans et al. 2001, Marsh & Steeghs 2002, Ramsay et al. 2002a) the X-rays are powered by mass transfer onto the primary white dwarf. One would thus expect to see emission lines from the accretion flow in those cases. In the unipolar-inductor model of Wu et al. (2002), there would be no mass transfer and thus no emission lines from the accretion flow itself, though there could be emission from the irradiated mass donor star. The fact that the variable part of the light is very blue suggests an origin in a hot environment, which in the accreting ultra-compact models could naturally be associated with the accreting primary that is expected to be heated to temperatures in excess of 50,000K (Bildsten et al. 2006). Since our data resolve the 9.48 min period, we would also be able to pick up rapidly moving narrow components that would be expected if the emission lines originate along a stream or on the surface of the rapidly rotating donor star. We certainly do not see the kind of strong helium emission lines that are typically observed in the longer period disk-accreting AM CVn systems (e.g. Nather, Robinson & Stover 1981, Roelofs et al. 2005), including the 10.3minute system ES Ceti (Wegner et al. 1987; Warner & Woudt 2002). However, a very weak emission line spectrum, such as is observed in RX J0806.3-0939 (Israel et al. 2002) could actually be hidden in the seemingly featureless spectrum of the variable source (Fig.3). While our signal to noise is good enough to see lines of typical strengths such as displayed in all the other CVs and AM CVns, very weak lines would go undetected. The fact that such short-period systems may not accrete via a disk but instead via a (direct impact) stream could be one of the reasons for a reduced emission line contribution. The detection of emission line components showing radial velocity motions across the 9.48 min orbit would have been conclusive proof of the binary period, and indeed was the aim of our experiment. We are thus severely hampered by the diluting effect of the G9 starlight in conjunction with the reddening along the line of sight to make conclusive statements concerning the presence or absence of very weak helium lines. A significant improvement will be difficult to make given the faintness of the source at blue wavelengths. Outlook ------- The true nature of V407 Vul continues to puzzle us. The search for radial velocity signatures of its orbit remains elusive. Additional, high-resolution, spectroscopy in the red end will permit more stringent constraints to be placed on the properties of the G9 star that dominates the light. If this star is indeed in a longer orbit, accurate velocity measurements sampling longer base-lines could for example confirm the possibility of a triple. While the underlying variable is best studied in the blue, the reddening and intrinsic faintness of the source makes this a rather challenging avenue. After all, we were not able to detect features even with a considerable amount of 8m-class spectroscopy. However a rigorous analysis of the spectral energy distribution of the source including sensitive UV flux constraints should shed more light on the relative contributions of the G9 star and the variable powering the X-ray source. DS acknowledges a Smithsonian Astrophysical Observatory Clay Fellowship as well as support through NASA GO grants NNG04GG30G and NNG04GG96G. GN is supported by NWO-VENI grant 639.041.405. TRM acknowledges the support of a PPARC Senior Research Fellowship. SCCB is supported by FCT. We thank Josh Grindlay for usefull comments related to the triple scenario. We would like to thank the Gemini staff for their help in the planning and acquisition of the spectra discussed in this paper. This work uses observations obtained at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Particle Physics and Astronomy Research Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), CNPq (Brazil), and CONICET (Argentina). This work uses data obtained with the Magellan telescopes. The Magellan project is operated by Carnegie Observatories at Las Campanas Observatory, Chile on behalf of the Magellan consortium. Bildsten, L., Townsley, D., Deloye,C., Nelemans, G., 2006, ApJ, 640, 466 Cropper, M., Harrop-Allin, M. K., Mason, K. O., Mittaz, J. P. D., Potter, S. B., Ramsay, G., 1998, MNRAS, 293, L57 Garcia, M.R., Bailyn, C.D., Grindlay, J.E., Molnar, L.A., 1989, ApJ, 341, L87 Haberl, F., Motch, C., 1995, A&A, 297, L37 Herbig, G.H., 1993, ApJ, 407, 142 Hellier, C., Beardmore, A.P., 2002, MNRAS, 331, 407 Hook, I., Jørgensen, I., Allington-Smith, J. R., Davies, R. L., Metcalfe, N., Murowinski, R. G., Crampton, D., 2004, PASP, 116, 425 Houk, N., Swift, C. M., Murray, C. A., Penston, M. J., Binney, J. J., 1997, ESA SP-402, 279 Iben, I.J, 1990, ApJ, 353, 215 Israel, G. L., et al., 2002, A&A, 386, L13 Joshi, Y.C., MNRAS, 2005, 362, 1259 Marsh, T.R, Nelemans, G., MNRAS, 363, 581 Marsh, T. R., Nelemans, G., Steeghs, D., 2004, MNRAS, 350, 113 Marsh, T. R., Steeghs, D., 2002, MNRAS, 331, L7 Motch, C., Haberl, F., Guillout, P., Pakull, M., Reinsch, K., Krautter, J, 1996, A&A, 307, 459 Napiwotzki, R., et al., 2003, in: de Martino D., Silvotti R., Solheim J.-E., Kalytis R., eds, NATO Sci Series II, Volume 105, White Dwarfs, Kluwer Academic, Dordrecht Nather, R.E., Robinson, E. L., Stover, R. J., 1981, ApJ, 244, 269 Neckel, Th., Klare, G., 1980, A&ASS, 42, 251 Nelemans, G., Portegies Zwart, S. F., Verbunt, F., Yungelson, L. R., 2001, A&A, 368, 939 Norton, A. J., Haswell, C. A., Wynn, G. A., 2004, A&A, 419, 1025 Podsiadlowski, Ph., Han, Z., Rappaport, S, 2003, MNRAS, 340, 1214 Predehl, P., Schmitt, J. H. M. M., 1995, A&A, 293, 889 Ramsay, G., Cropper, M., Hakala, P., 2006, MNRAS, 367, L62 Ramsay, G., Hakala, P., Wu, K., Cropper, M., Mason, K. O., Córdova, F. A., Priedhorsky, W, 2005, MNRAS, 357, 49 Ramsay, G., Wu, K., Cropper, M., Schmidt, G., Sekiguchi, K., Iwamuro, F., Maihara, T., 2002a, MNRAS, 333, 575 Ramsay, G., Hakala, P., Cropper, M., 2002b, MNRAS, 332, L7 Ramsay, G., Cropper, M., Wu, K., Mason, K. O., Hakala, P., 2000, MNRAS, 311, 75 Roelofs, G. H. A., Groot, P. J., Marsh, T. R., Steeghs, D., Barros, S. C. C., Nelemans, G., 2005, MNRAS, 361, 487 Scargle, J.D., 1982, ApJ, 263, 835 Schlegel, D. J., Finkbeiner, D. P., Davis, M., 1998, ApJ, 500, 525 Strohmayer, T. E., 2002, ApJ, 581, 577 Strohmayer, T. E., 2004, ApJ, 610, 416 Schwope, A. D., Mantel, K.-H., Horne, K., 1997, A&A, 319, 894 Webbink, R.F., 1984, ApJ, 277, 355 Warner, B., Woudt, P. A., 2002, PASP, 114, 129 Wegner, G., McMahon, R. K., Boley, F. I. 1987, AJ, 94, 1271 Wu, K., Cropper, M., Ramsay, G., Sekiguchi, K., 2002, MNRAS, 331, 221 [^1]: This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile.

ArXiv

--- abstract: 'We introduce a fundamental lemma called the Poisson matching lemma, and apply it to prove one-shot achievability results for various settings, namely channels with state information at the encoder, lossy source coding with side information at the decoder, joint source-channel coding, broadcast channels, distributed lossy source coding, multiple access channels, channel resolvability and wiretap channels. Our one-shot bounds improve upon the best known one-shot bounds in most of the aforementioned settings (except multiple access channels, channel resolvability and wiretap channels, where we recover bounds comparable to the best known bounds), with shorter proofs in some settings even when compared to the conventional asymptotic approach using typicality. The Poisson matching lemma replaces both the packing and covering lemmas, greatly simplifying the error analysis. This paper extends the work of Li and El Gamal on Poisson functional representation, which mainly considered variable-length source coding settings, whereas this paper studies fixed-length settings, and is not limited to source coding, showing that the Poisson functional representation is a viable alternative to typicality for most problems in network information theory.' author: - | Cheuk Ting Li and Venkat Anantharam\ EECS, UC Berkeley, Berkeley, CA, USA\ Email: ctli@berkeley.edu, ananth@eecs.berkeley.edu bibliography: - 'ref.bib' title: 'A Unified Framework for One-shot Achievability via the Poisson Matching Lemma' --- Introduction ============ The Poisson functional representation was introduced by Li and El Gamal [@sfrl_trans] to prove the strong functional representation lemma: for any pair of random variables $(X,Y)$, there exists a random variable $Z$ independent of $X$ such that $Y$ is a function of $(X,Z)$, and $H(Y|Z)\le I(X;Y)+\log(I(X;Y)+1)+4$. The lemma is applied to show various one-shot variable-length lossy source coding results, and a simple proof of the asymptotic achievability in the Gelfand-Pinsker theorem [@gelfand1980coding]. In this paper, we introduce the Poisson matching lemma, which gives a bound on the probability of mismatch between the Poisson functional representations applied on different distributions, and use it to prove one-shot achievability results for various settings, namely channels with state information at the encoder, lossy source coding with side information at the decoder, joint source-channel coding, broadcast channels, distributed lossy source coding, multiple access channels, channel resolvability and wiretap channels. The Poisson matching lemma can replace both the packing and covering lemmas (and generalizations such as the mutual covering lemma) in asymptotic typicality-based proofs. The one-shot bounds in this paper subsume the corresponding asymptotic achievability results by straightforward applications of the law of large numbers. Various non-asymptotic alternatives to typicality have been proposed, e.g. one-shot packing and covering lemmas [@verdu2012nonasymp; @liu2015oneshotmutual], stochastic likelihood coder [@yassaee2013oneshot], likelihood encoder [@song2016likelihood] and random binning [@yassaee2013binning]. However, these non-asymptotic approaches generally require more complex proofs than their asymptotic counterparts, whereas proofs using the Poisson matching lemma can be even simpler than asymptotic proofs. Our approach is better than the conventional asymptotic approach using typicality (and previous one-shot results, e.g. [@verdu2012nonasymp; @yassaee2013oneshot]), in the following ways: 1. We can give one-shot bounds stronger than the best known one-shot bounds in many settings discussed in this paper, with the exception of channel coding, multiple access channels, channel resolvability and wiretap channels, which are included for demonstration purposes, where we recover bounds comparable to the best known bounds. 2. Our proofs work for random variables in general Polish spaces. 3. To the best of our knowledge, for the achievability in the Gelfand-Pinsker theorem [@gelfand1980coding] (for channels with state information at the encoder) and the Wyner-Ziv theorem [@wyner1976ratedistort; @wyner1978rate] (for lossy source coding with side information at the decoder), our proofs are significantly shorter than all previous proofs (another short proof of the achievability in the Gelfand-Pinsker theorem is given in [@sfrl_trans], though it is asymptotic). Using our approach, we can also greatly shorten the proof of the achievability of the dispersion in joint source-channel coding [@kostina2013joint]. 4. Our proofs only use the Poisson matching lemma introduced in this paper, which replaces both the packing and covering lemmas in proofs using typicality. The Poisson matching lemma can also be used to prove a soft covering lemma. Hence the Poisson matching lemma can be the only tool needed to prove a wide range of results in network information theory. 5. Our analyses usually involve fewer (or no) uses of sub-codebooks and binning. As a result, we can reduce the number of error events and give sharper second-order bounds. For example: 1. Conventional proofs of the Gelfand-Pinsker theorem involve one sub-codebook, giving an additional error event, whereas we do not use any sub-codebook. 2. Conventional proofs of the Wyner-Ziv theorem and the Berger-Tung inner bound [@berger1978multiterminal; @tung1978multiterminal] (for distributed lossy source coding) use binning, giving additional error events, whereas we do not require binning. 3. Conventional proofs of Marton’s inner bound [@marton1979broadcast] (for broadcast channels) involve two sub-codebooks, whereas we use only one. 6. In our approach, the encoders and decoders are characterized using a common framework (the Poisson functional representation), which is noteworthy since the roles of an encoder and a decoder in an operational setting are very different, and their constructions usually have little in common in conventional approaches. Notation {#notation .unnumbered} -------- Throughout this paper, we assume that $\log$ is to base 2 and the entropy $H$ is in bits. We write $\mathrm{exp}_{a}(b)$ for $a^{b}$. The set of positive integers is denoted as $\mathbb{N}=\{1,2,\ldots\}$. We use the notation: $X_{a}^{b}:=(X_{a},\ldots,X_{b})$, $X^{n}:=X_{1}^{n}$ and $[a:b]:=[a,b]\cap\mathbb{Z}$. The conditional information density is denoted as $$\iota_{X;Y|Z}(x;y|z):=\log\frac{dP_{XY|Z=z}}{d(P_{X|Z=z}\times P_{Y|Z=z})}(x,y).$$ We consider $\iota_{X;Y|Z}(x;y|z)$ to be defined only if $P_{XY|Z=z}\ll P_{X|Z=z}\times P_{Y|Z=z}$. For discrete $X$, we write the probability mass function as $p_{X}$. For continuous $X$, we write the probability density function as $f_{X}$. For a general random variable $X$ in a measurable space, we write its distribution as $P_{X}$. The uniform distribution over a finite set $S$ is denoted as $\mathrm{Unif}(S)$. The joint distribution of $X_{1},\ldots,X_{n}\stackrel{iid}{\sim}P_{X}$ is written as $P_{X}^{\otimes n}$. The degenerate distribution $\mathbf{P}\{X=a\}=1$ is denoted as $\delta_{a}$. The conditional independence of $X$ and $Z$ given $Y$ is denoted as $X\leftrightarrow Y\leftrightarrow Z$. The Q-function and its inverse are denoted as $\mathcal{Q}(x)$ and $\mathcal{Q}^{-1}(\epsilon)$ respectively. For $V\in\mathbb{R}^{n\times n}$ positive semidefinite, define $\mathcal{Q}^{-1}(V,\epsilon)=\{x\in\mathbb{R}^{n}:\,\mathbf{P}\{X\le x\}\ge1-\epsilon\}$ where $X\sim N(0,V)$ and $X\le x$ denotes entrywise comparison. We assume that every random variable mentioned in this paper lies in a Polish space with its Borel $\sigma$-algebra, and all functions mentioned (e.g. distortion measures, the function $x(u,s)$ in Theorem \[thm:channel\_state\]) are measurable. The Lebesgue measure over $\mathbb{R}$ is denoted as $\lambda$. The Lebesgue measure restricted to the set $S\subseteq\mathbb{R}$ is denoted as $\lambda_{S}$. For two measures $\mu,\nu$ over $\mathcal{X}$ (a Polish space with its Borel $\sigma$-algebra) such that $\nu$ is absolutely continuous with respect to $\mu$ (denoted as $\nu\ll\mu$), the Radon-Nikodym derivative is written as $$\frac{d\nu}{d\mu}:\,\mathcal{X}\to[0,\infty).$$ If $\nu_{1},\nu_{2}\ll\mu$ (but $\nu_{1}\ll\nu_{2}$ may not hold), we write $$\frac{d\nu_{1}}{d\nu_{2}}(x)=\frac{d\nu_{1}}{d\mu}(x)\left(\frac{d\nu_{2}}{d\mu}(x)\right)^{-1}\in[0,\infty],\label{eq:rnderiv}$$ which is $0$ if $(d\nu_{1}/d\mu)(x)=0$, and is $\infty$ if $(d\nu_{1}/d\mu)(x)>0$ and $(d\nu_{2}/d\mu)(x)=0$. The total variation distance between two distributions $P,Q$ over $\mathcal{X}$ is denoted as $\Vert P-Q\Vert_{\mathrm{TV}}=\sup_{A\subseteq\mathcal{X}\,\mathrm{measurable}}|P(A)-Q(A)|$. Poisson Matching Lemma ====================== We first state the definition of Poisson functional representation in [@sfrl_trans], with a different notation that allows the proofs to be written in a simpler and more intuitive manner. \[Poisson functional representation\]Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ on $\mathcal{U}\times\mathbb{R}_{\ge0}$ (where $\mathcal{U}$ is a Polish space with its Borel $\sigma$-algebra, and $\mu$ is $\sigma$-finite). For $P\ll\mu$ a probability measure over $\mathcal{U}$, define $$\tilde{U}_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}\right):=\bar{U}_{K_{P}(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}})},$$ where $$K_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}\right):=\underset{i:\,\frac{dP}{d\mu}(\bar{U}_{i})>0}{\arg\min}T_{i}\left(\frac{dP}{d\mu}(\bar{U}_{i})\right)^{-1},$$ with arbitrary tie-breaking (a tie occurs with probability 0). We omit $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ and only write $\tilde{U}_{P}$ if the Poisson process is clear from the context. If the Poisson process is $\{\bar{X}_{i},T_{i}\}_{i\in\mathbb{N}}$ instead of $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$, then the Poisson functional representation is likewise denoted as $\tilde{X}_{P}$. If $\bar{U}_{i}=(\bar{X}_{i},\bar{Y}_{i})$ is multivariate, and $P$ is a distribution over $\mathcal{X}\times\mathcal{Y}$, the Poisson functional representation is denoted as $(\tilde{X},\tilde{Y})_{P}$. We write its components as $(\tilde{X},\tilde{Y})_{P}=(\tilde{X}_{P},\tilde{Y}_{P})$. Note that while $dP/d\mu$ is only uniquely defined up to a $\mu$-null set, changing the value of $dP/d\mu$ on a $\mu$-null set will only affect the values of $\tilde{U}_{P}$ on a null set with respect to the distribution of $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$, since the probability that there exists $\bar{U}_{i}$ on that $\mu$-null set is zero. Therefore $\tilde{U}_{P}$ is uniquely defined up to a null set. By the mapping theorem [@kingman1992poisson; @last2017lectures] (also see Appendix A of [@sfrl_trans]), we have $\tilde{U}_{P}\sim P$. This is termed Poisson functional representation in [@sfrl_trans] since it can be regarded as a construction for the functional representation lemma [@elgamal2011network]. Consider the distribution $P_{U,X}$. Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{U}\times\lambda_{\mathbb{R}_{\ge0}}$, $X\sim P_{X}$ independent of the process, and $U:=\tilde{U}_{P_{U|X}(\cdot|X)}$. Then $(U,X)\sim P_{U,X}$. Hence we can express $U$ as a function of $X$ and $\{\bar{U}_{i},T_{i}\}$ (which is independent of $X$). This fact will be used repeatedly throughout the proofs in this paper. For two different distributions $P$ and $Q$, $\tilde{U}_{P}$ and $\tilde{U}_{Q}$ are coupled in such a way that $\tilde{U}_{P}=\tilde{U}_{Q}$ occurs with a probability that can be bounded in terms of $dP/dQ$. We now present the core lemma of this paper. The proof is given in Appendix \[subsec:pf\_phidiv\]. \[Poisson matching lemma\]\[lem:phidiv\]Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$, and $P,Q$ be probability measures on $\mathcal{U}$ with $P,Q\ll\mu$. Then we have the following almost surely: $$\mathbf{P}\left\{ \left.\tilde{U}_{Q}\neq\tilde{U}_{P}\,\right|\,\tilde{U}_{P}\right\} \le1-\left(1+\frac{dP}{dQ}(\tilde{U}_{P})\right)^{-1},\label{eq:phidiv}$$ where we write $(dP/dQ)(u)=(dP/d\mu)(u)/((dQ/d\mu)(u))$ as in (we do not require $P\ll Q$). The right hand side of is considered to be 1 if $(dP/d\mu)(\tilde{U}_{P})>0$ and $(dQ/d\mu)(\tilde{U}_{P})=0$. The exact expression for the left hand side of is in . We usually do not apply the Poisson matching lemma on fixed $P,Q$, but rather on conditional distributions. The following conditional version of the Poisson matching lemma follows directly from applying the lemma on $(P,Q)\leftarrow(P_{U|X}(\cdot|X),Q_{U|Y}(\cdot|Y))$. The proof is given in Appendix \[subsec:phidiv\_cond\] for the sake of completeness. \[Conditional Poisson matching lemma\]\[lem:cond\_phidiv\]Fix a distribution $P_{X,U,Y}$ and a probability kernel $Q_{U|Y}$ (that is not necessarily $P_{U|Y}$) satisfying $P_{U|X}(\cdot|X),Q_{U|Y}(\cdot|Y)\ll\mu$ almost surely. Let $X\sim P_{X}$, and $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $X$. Let $U=\tilde{U}_{P_{U|X}(\cdot|X)}$ and $Y|(X,U,\{\bar{U}_{i},T_{i}\}_{i})\sim P_{Y|X,U}(\cdot|X,U)$ (note that $(X,U,Y)\sim P_{X,U,Y}$ and $Y\leftrightarrow(X,U)\leftrightarrow\{\bar{U}_{i},T_{i}\}_{i}$). Then we have the following almost surely: $$\mathbf{P}\left\{ \left.\tilde{U}_{Q_{U|Y}(\cdot|Y)}\neq U\,\right|\,X,U,Y\right\} \le1-\left(1+\frac{dP_{U|X}(\cdot|X)}{dQ_{U|Y}(\cdot|Y)}(U)\right)^{-1}.$$ The condition that $P_{U|X}(\cdot|X),Q_{U|Y}(\cdot|Y)\ll\mu$ almost surely is satisfied, for example, when $\mu=P_{U}$, $Q_{U|Y}=P_{U|Y}$, $P_{UX}\ll P_{U}\times P_{X}$ and $P_{UY}\ll P_{U}\times P_{Y}$. Note that since $X{\perp\!\!\!\perp}\{\bar{U}_{i},T_{i}\}_{i}$, we have $\tilde{U}_{P_{U|X}(\cdot|X)}|X\sim P_{U|X}$, whereas $Y$ may not be independent of $\{\bar{U}_{i},T_{i}\}_{i}$, so $\tilde{U}_{Q_{U|Y}(\cdot|Y)}$ may not follow the conditional distribution $Q_{U|Y}$. One-shot Channel Coding\[sec:channel\] ====================================== To demonstrate the application of the Poisson matching lemma, we apply it to recover a bound for one-shot channel coding in [@yassaee2013oneshot] (with a slight penalty of having $\mathsf{L}$ instead of $\mathsf{L}-1$). Upon observing $M\sim\mathrm{Unif}[1:\mathsf{L}]$, the encoder produces $X$, which is sent through the channel $P_{Y|X}$. The decoder observes $Y$ and recovers $\hat{M}$ with error probability $P_{e}=\mathbf{P}\{M\neq\hat{M}\}$. \[prop:channel\]Fix any $P_{X}$. There exists a code for the channel $P_{Y|X}$, with message $M\sim\mathrm{Unif}[1:\mathsf{L}]$, with average error probability $$P_{e}\le\mathbf{E}\left[1-\left(1+\mathsf{L}2^{-\iota_{X;Y}(X;Y)}\right)^{-1}\right]$$ if $P_{XY}\ll P_{X}\times P_{Y}$. Let $\{(\bar{X}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{X}\times P_{M}\times\lambda_{\mathbb{R}_{\ge0}}$ (where $P_{M}$ is $\mathrm{Unif}[1:\mathsf{L}]$) independent of $M$. The encoding function is $m\mapsto\tilde{X}_{P_{X}\times\delta_{m}}$ (i.e., $X=\tilde{X}_{P_{X}\times\delta_{M}}$), and the decoding function is $y\mapsto\tilde{M}_{P_{X|Y}(\cdot|y)\times P_{M}}$ (i.e., $\hat{M}=\tilde{M}_{P_{X|Y}(\cdot|Y)\times P_{M}}$). Note that the encoding and decoding functions also depend on the common randomness $\{(\bar{X}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}$, which will be fixed later. We have $(M,X,Y)\sim P_{M}\times P_{X}P_{Y|X}$. $$\begin{aligned} & \mathbf{P}\left\{ M\neq\tilde{M}_{P_{X|Y}(\cdot|Y)\times P_{M}}\right\} \\ & \le\mathbf{P}\left\{ (X,M)\neq(\tilde{X},\tilde{M})_{P_{X|Y}(\cdot|Y)\times P_{M}}\right\} \\ & =\mathbf{E}\left[\mathbf{P}\left\{ \left.(X,M)\neq(\tilde{X},\tilde{M})_{P_{X|Y}(\cdot|Y)\times P_{M}}\,\right|\,M,X,Y\right\} \right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[1-\left(1+\frac{dP_{X}\times\delta_{M}}{dP_{X|Y}(\cdot|Y)\times P_{M}}(X,M)\right)^{-1}\right]\\ & =\mathbf{E}\left[1-(1+\mathsf{L}2^{-\iota_{X;Y}(X;Y)})^{-1}\right],\end{aligned}$$ where (a) is by the conditional Poisson matching lemma (Lemma \[lem:cond\_phidiv\]) on $(X,U,Y,Q_{U|Y})\leftarrow(M,(X,M),Y,P_{X|Y}\times P_{M})$ (note that $P_{X,M|M}=P_{X}\times\delta_{M}$). Therefore there exists a fixed $\{(\bar{x}_{i},\bar{m}_{i}),t_{i}\}_{i\in\mathbb{N}}$ such that conditioned on $\{(\bar{X}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}=\{(\bar{x}_{i},\bar{m}_{i}),t_{i}\}_{i\in\mathbb{N}}$, the average probability of error is bounded by $\mathbf{E}\left[1-(1+\mathsf{L}2^{-\iota_{X;Y}(X;Y)})^{-1}\right]$. Compared to the scheme in [@yassaee2013oneshot], we use the Poisson process $\{(\bar{X}_{i},\bar{M}_{i}),T_{i}\}$ to create a codebook, instead of the conventional i.i.d. random codebook in [@yassaee2013oneshot]. While the codewords for different $m$’s are still i.i.d., we attach a bias $T_{i}$ to each codeword. Our scheme does not use a stochastic decoder as in [@yassaee2013oneshot], but rather a biased maximum likelihood decoder $\tilde{M}_{P_{X|Y}(\cdot|y)\times P_{M}}=\bar{M}_{K}$ where $K=\arg\max_{i}T_{i}^{-1}(dP_{X|Y}(\cdot|y)/dP_{X})(\bar{X}_{i})$. In the following sections, we will demonstrate how our approach can lead to simpler proofs and sharper bounds compared to [@yassaee2013oneshot]. Using the generalized Poisson matching lemma that will be introduced in Section \[sec:gen\_pml\], we can prove the following bound. The proof is in Appendix \[subsec:channel2\]. \[thm:channel2\]Fix any $P_{X}$. There exists a code for the channel $P_{Y|X}$, with message $M\sim\mathrm{Unif}[1:\mathsf{L}]$, with average error probability $$P_{e}\le\mathbf{E}\left[1-\left(1-\min\left\{ 2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right)^{(\mathsf{L}+1)/2}\right]$$ if $P_{XY}\ll P_{X}\times P_{Y}$. Compare this to the dependence testing bound [@polyanskiy2010channel]: $$P_{e}\le\mathbf{E}\left[\min\left\{ \frac{\mathsf{L}-1}{2}\cdot2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right].$$ Theorem \[thm:channel2\] is at least as strong (with a slight penalty of having $(\mathsf{L}+1)/2$ instead of $(\mathsf{L}-1)/2$) since $$\begin{aligned} & \mathbf{E}\left[1-\left(1-\min\left\{ 2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right)^{(\mathsf{L}+1)/2}\right]\\ & \le\mathbf{E}\left[\min\left\{ \frac{\mathsf{L}+1}{2}\cdot2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right].\end{aligned}$$ Apart from the dependence testing bound [@polyanskiy2010channel], there are other one-shot bounds for channel coding such as the random-coding union (RCU) bound and the $\kappa\beta$ bound in [@polyanskiy2010channel], which are tighter in certain situations (e.g. the RCU bound is suitable for error exponent analysis). The technique introduced in this paper is suitable for first and second order analysis, but does not seem to give tight error exponent bounds. One-shot Coding for Channels with State Information at the Encoder\[sec:channel\_state\] ======================================================================================== The one-shot coding setting for a channel with state information at the encoder is described as follows. Upon observing $M\sim\mathrm{Unif}[1:\mathsf{L}]$ and $S\sim P_{S}$, the encoder produces $X$, which is sent through the channel $P_{Y|X,S}$ with state $S$. The decoder observes $Y$ and recovers $\hat{M}$ with error probability $P_{e}=\mathbf{P}\{M\neq\hat{M}\}$. We show a one-shot version of the Gelfand-Pinsker theorem [@gelfand1980coding]. This is the first one-shot bound attaining the best known second order result in [@scarlett2015dispersions] (which considers a finite-blocklength, not one-shot scenario). Our bound is stronger than the one-shot bounds in [@verdu2012nonasymp; @yassaee2013oneshot; @watanabe2015nonasymp] (in the second order), and significantly simpler to state and prove than all the aforementioned results. Unlike previous approaches, our proof does not require sub-codebooks. \[thm:channel\_state\]Fix any $P_{U|S}$ and function $x:\mathcal{U}\times\mathcal{S}\to\mathcal{X}$. There exists a code for the channel $P_{Y|X,S}$ with state distribution $P_{S}$ with message $M\sim\mathrm{Unif}[1:\mathsf{L}]$, with error probability $$P_{e}\le\mathbf{E}\left[1-(1+\mathsf{L}2^{\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)})^{-1}\right]$$ if $P_{US}\ll P_{U}\times P_{S}$ and $P_{UY}\ll P_{U}\times P_{Y}$, where $(S,U,X,Y)\sim P_{S}P_{U|S}\delta_{x(U,S)}P_{Y|X,S}$. Let $\{(\bar{U}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{U}\times P_{M}\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $M,S$. The encoding function is $(m,s)\mapsto x(\tilde{U}_{P_{U|S}(\cdot|s)\times\delta_{m}},s)$ (let $U=\tilde{U}_{P_{U|S}(\cdot|S)\times\delta_{M}}$, $X=x(U,S)$), and the decoding function is $y\mapsto\tilde{M}_{P_{U|Y}(\cdot|y)\times P_{M}}$ (i.e., $\hat{M}=\tilde{M}_{P_{U|Y}(\cdot|Y)\times P_{M}}$). Note that $(M,S,U,X,Y)\sim P_{M}\times P_{S}P_{U|S}\delta_{x(U,S)}P_{Y|X,S}$. We have $$\begin{aligned} & \mathbf{P}\{M\neq\tilde{M}_{P_{U|Y}(\cdot|Y)\times P_{M}}\}\\ & \le\mathbf{P}\{(U,M)\neq(\tilde{U},\tilde{M})_{P_{U|Y}(\cdot|Y)\times P_{M}}\}\\ & =\mathbf{E}\left[\mathbf{P}\left\{ \left.(U,M)\neq(\tilde{U},\tilde{M})_{P_{U|Y}(\cdot|Y)\times P_{M}}\,\right|\,M,S,U,Y\right\} \right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[1-\left(1+\frac{dP_{U|S}(\cdot|S)\times\delta_{M}}{dP_{U|Y}(\cdot|Y)\times P_{M}}(U,M)\right)^{-1}\right]\\ & =\mathbf{E}\left[1-(1+\mathsf{L}2^{\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)})^{-1}\right].\end{aligned}$$ where (a) is by the conditional Poisson matching lemma on $((M,S),\,(U,M),\,Y,\,P_{U|Y}\times P_{M})$ (note that $P_{U,M|M,S}=P_{U|S}\times\delta_{M}$). Therefore there exists a fixed $\{(\bar{u}_{i},\bar{m}_{i}),t_{i}\}_{i\in\mathbb{N}}$ attaining the desired bound. Compared to Theorem 3 in [@verdu2012nonasymp]: $$P_{e}\le\mathbf{P}\{\iota_{U;S}(U;S)>\log\mathsf{J}-\gamma\}+\mathbf{P}\{\iota_{U;Y}(U;Y)\le\log\mathsf{L}\mathsf{J}+\gamma\}+2^{-\gamma}+e^{-2^{\gamma}}$$ for any $\gamma>0$, $\mathsf{J}\in\mathbb{N}$, our result is strictly stronger since $$\begin{aligned} & \mathbf{E}\left[1-\left(1+\mathsf{L}2^{\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)}\right)^{-1}\right]\\ & \le\mathbf{P}\{\iota_{U;S}(U;S)>\log\mathsf{J}-\gamma\}+\mathbf{P}\{\iota_{U;Y}(U;Y)\le\log\mathsf{L}\mathsf{J}+\gamma\}\\ & \;\;\;\;+\mathbf{E}\left[1-\left(1+\mathsf{L}2^{\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)}\right)^{-1}\,|\,\iota_{U;S}(U;S)\le\log\mathsf{J}-\gamma,\,\iota_{U;Y}(U;Y)>\log\mathsf{L}\mathsf{J}+\gamma\right]\\ & \le\mathbf{P}\{\iota_{U;S}(U;S)>\log\mathsf{J}-\gamma\}+\mathbf{P}\{\iota_{U;Y}(U;Y)\le\log\mathsf{L}\mathsf{J}+\gamma\}+2^{-2\gamma}\\ & <\mathbf{P}\{\iota_{U;S}(U;S)>\log\mathsf{J}-\gamma\}+\mathbf{P}\{\iota_{U;Y}(U;Y)\le\log\mathsf{L}\mathsf{J}+\gamma\}+2^{-\gamma}+e^{-2^{\gamma}}.\end{aligned}$$ This is due to the fact that the Poisson matching lemma simultaneously replaces both the covering and the packing lemma, resulting in only one error event. Next, we prove a second-order result. Fix $\epsilon>0$. Let $C:=I(U;Y)-I(U;S)$, $V:=\mathrm{Var}[\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)]$. We apply Theorem \[thm:channel\_state\] on $n$ uses of the memoryless channel with i.i.d. state sequence $S^{n}=(S_{1},\ldots,S_{n})$, and $$\mathsf{L}:=\left\lfloor \mathrm{exp}_{2}\left(nC-\sqrt{nV}\mathcal{Q}^{-1}\left(\epsilon-\frac{\alpha}{\sqrt{n}}\right)-\frac{1}{2}\log n\right)\right\rfloor ,$$ where $\alpha$ is a constant that depends on $P_{S,U,Y}$. For $n>\alpha^{2}\epsilon^{-2}$, by the Berry-Esseen theorem [@berry1941accuracy; @esseen1942liapunov; @feller1971introduction], we have $$\begin{aligned} P_{e} & \le\mathbf{E}\left[\min\left\{ 2^{\log\mathsf{L}+\iota_{U^{n};S^{n}}(U^{n};S^{n})-\iota_{U^{n};Y^{n}}(U^{n};Y^{n})},\,1\right\} \right]\\ & \le\frac{1}{\sqrt{n}}+\mathbf{P}\left\{ 2^{\log\mathsf{L}+\iota_{U^{n};S^{n}}(U^{n};S^{n})-\iota_{U^{n};Y^{n}}(U^{n};Y^{n})}>\frac{1}{\sqrt{n}}\right\} \\ & \le\frac{1}{\sqrt{n}}+\mathbf{P}\left\{ \frac{1}{\sqrt{n}}\sum_{i=1}^{n}\left(\iota_{U;Y}(U_{i};Y_{i})-\iota_{U;S}(U_{i};S_{i})-C\right)<-\sqrt{V}\mathcal{Q}^{-1}\left(\epsilon-\frac{\alpha}{\sqrt{n}}\right)\right\} \\ & \le\frac{1}{\sqrt{n}}+\epsilon-\frac{\alpha}{\sqrt{n}}+\frac{\alpha-1}{\sqrt{n}}\\ & \le\epsilon\end{aligned}$$ if we let $\alpha-1$ be the constant given by the Berry-Esseen theorem. This coincides with the best known second order result in [@scarlett2015dispersions], which is stronger than the second order results implied by [@verdu2012nonasymp; @yassaee2013oneshot; @watanabe2015nonasymp]. We bound $\iota_{U;S}(U;S)-\iota_{U;Y}(U;Y)$ as a single quantity, instead of bounding the two terms separately as in [@verdu2012nonasymp; @yassaee2013oneshot; @watanabe2015nonasymp], resulting in a sharper second order bound. One-shot Lossy Source Coding with Side Information at the Decoder ================================================================= The one-shot lossy source coding setting with side information at the decoder is described as follows. Upon observing $X\sim P_{X}$, the encoder produces $M\in[1:\mathsf{L}]$. The decoder observes $M$ and $Y\sim P_{Y|X}$ and recovers $\hat{Z}\in\mathcal{Z}$ with probability of excess distortion $P_{e}=\mathbf{P}\{\mathsf{d}(X,\hat{Z})>\mathsf{D}\}$, where $\mathsf{d}:\mathcal{X}\times\mathcal{Z}\to\mathbb{R}_{\ge0}$ is a distortion measure. We show a one-shot version of the Wyner-Ziv theorem [@wyner1976ratedistort; @wyner1978rate]. Our bound is stronger than those in [@verdu2012nonasymp; @watanabe2015nonasymp], and significantly simpler to state and prove. Unlike previous approaches, our proof does not require binning. \[thm:lscsi\]Fix any $P_{U|X}$ and function $z:\mathcal{U}\times\mathcal{Y}\to\mathcal{Z}$. There exists a code for lossy source coding with source distribution $P_{X}$, side information at the decoder given by $P_{Y|X}$, and message size $\mathsf{L}$, with probability of excess distortion $$P_{e}\le\mathbf{E}\left[1-\mathbf{1}\{\mathsf{d}(X,Z)\le\mathsf{D}\}(1+\mathsf{L}^{-1}2^{\iota_{U;X}(U;X)-\iota_{U;Y}(U;Y)})^{-1}\right]$$ if $P_{UX}\ll P_{U}\times P_{X}$ and $P_{UY}\ll P_{U}\times P_{Y}$, where $(X,Y,U,Z)\sim P_{X}P_{Y|X}P_{U|X}\delta_{z(U,Y)}$. Let $\{(\bar{U}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{U}\times P_{M}\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $X$, where $P_{M}$ is $\mathrm{Unif}[1:\mathsf{L}]$. The encoding function is $x\mapsto\tilde{M}_{P_{U|X}(\cdot|x)\times P_{M}}$ (i.e., $M=\tilde{M}_{P_{U|X}(\cdot|X)\times P_{M}}$), and the decoding function is $(m,y)\mapsto z(\tilde{U}_{P_{U|Y}(\cdot|y)\times\delta_{m}},y)$ (let $\hat{U}=\tilde{U}_{P_{U|Y}(\cdot|Y)\times\delta_{M}}$, $\hat{Z}=z(\hat{U},Y)$). Also define $U=\tilde{U}_{P_{U|X}(\cdot|X)\times P_{M}}$, $Z=z(U,Y)$. Note that $(M,X,Y,U,Z)\sim P_{M}\times P_{X}P_{Y|X}P_{U|X}\delta_{z(U,Y)}$. We have $$\begin{aligned} & \mathbf{P}\{\mathsf{d}(X,\hat{Z})>\mathsf{D}\}\\ & \le1-\mathbf{P}\{\mathsf{d}(X,Z)\le\mathsf{D}\,\mathrm{and}\,U=\hat{U}\}\\ & \le\mathbf{E}\left[1-\mathbf{1}\{\mathsf{d}(X,Z)\le\mathsf{D}\}\mathbf{P}\{(U,M)=(\tilde{U},\tilde{M})_{P_{U|Y}(\cdot|Y)\times\delta_{M}}\,|\,M,X,Y,U\}\right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[1-\mathbf{1}\{\mathsf{d}(X,Z)\le\mathsf{D}\}\left(1+\frac{dP_{U|X}(\cdot|X)\times P_{M}}{dP_{U|Y}(\cdot|Y)\times\delta_{M}}(U,M)\right)^{-1}\right]\\ & \le\mathbf{E}\left[1-\mathbf{1}\{\mathsf{d}(X,Z)\le\mathsf{D}\}(1+\mathsf{L}^{-1}2^{\iota_{U;X}(U;X)-\iota_{U;Y}(U;Y)})^{-1}\right].\end{aligned}$$ where (a) is by the conditional Poisson matching lemma on $(X,\,(U,M),\,(M,Y),\,P_{U|Y}\times\delta_{M})$ (note that $P_{U,M|X}=P_{U|X}\times P_{M}$). Therefore there exists a fixed $\{(\bar{u}_{i},\bar{m}_{i}),t_{i}\}_{i\in\mathbb{N}}$ attaining the desired bound. This reduces to lossy source coding (without side information) when $Y=\emptyset$. Note that the encoder is designed in the same way with or without side information. An encoder for lossy source coding is sufficient to achieve the bound in Theorem \[thm:lscsi\] even when side information is present. Binning is not required at the encoder. Similar to the case in Section \[sec:channel\_state\], it can be checked that our bound is stronger than that in Theorem 2 in [@verdu2012nonasymp]. Compared to Corollary 9 in [@watanabe2015nonasymp]: $$\begin{aligned} P_{e} & \le\mathbf{P}\{\iota_{U;X}(U;X)>\gamma_{\mathrm{c}}\;\mathrm{or}\;\iota_{U;Y}(U;Y)<\gamma_{\mathrm{p}}\;\mathrm{or}\;\mathsf{d}(X,Z)>\mathsf{D}\}+\frac{\mathsf{J}}{2^{\gamma_{\mathrm{p}}}\mathsf{L}}+\frac{1}{2}\sqrt{\frac{2^{\gamma_{\mathrm{c}}}}{\mathsf{J}}}\label{eq:lscsi_watanabe}\end{aligned}$$ for any $\gamma_{\mathrm{p}},\gamma_{\mathrm{c}}>0$, $\mathsf{J}\in\mathbb{N}$, our result is stronger since $$\begin{aligned} & \mathbf{E}\left[1-\mathbf{1}\{\mathsf{d}(X,Z)\le\mathsf{D}\}(1+\mathsf{L}^{-1}2^{\iota_{U;X}(U;X)-\iota_{U;Y}(U;Y)})^{-1}\right]\\ & \le\mathbf{P}\{\iota_{U;X}(U;X)>\gamma_{\mathrm{c}}\;\mathrm{or}\;\iota_{U;Y}(U;Y)<\gamma_{\mathrm{p}}\;\mathrm{or}\;\mathsf{d}(X,Z)>\mathsf{D}\}+\mathsf{L}^{-1}2^{\gamma_{\mathrm{c}}-\gamma_{\mathrm{p}}}\\ & \le\mathbf{P}\{\iota_{U;X}(U;X)>\gamma_{\mathrm{c}}\;\mathrm{or}\;\iota_{U;Y}(U;Y)<\gamma_{\mathrm{p}}\;\mathrm{or}\;\mathsf{d}(X,Z)>\mathsf{D}\}+\frac{\mathsf{J}}{2^{\gamma_{\mathrm{p}}}\mathsf{L}}+\frac{1}{2}\sqrt{\frac{2^{\gamma_{\mathrm{c}}}}{\mathsf{J}}},\end{aligned}$$ where the last inequality is due to $$a+b\ge(a+b)^{3}=27\left(\frac{a+2(b/2)}{3}\right)^{3}\ge27a(b/2)^{2}\ge4ab^{2}$$ by the AM-GM inequality for $a,b\ge0$, $a+b\le1$ (since the right hand side of $\le1$ for it to be meaningful). We bound $\iota_{U;X}(U;X)-\iota_{U;Y}(U;Y)$ as a single quantity, instead of bounding the two terms separately, resulting in a sharper bound. One-shot Joint Source-Channel Coding ==================================== The one-shot joint source-channel coding setting is described as follows. Upon observing the source symbol $W\sim P_{W}$, the encoder produces $X\in\mathcal{X}$, which is sent through the channel $P_{Y|X}$. The decoder observes $Y$ and recovers $\hat{Z}\in\mathcal{Z}$ with probability of excess distortion $P_{e}=\mathbf{P}\{\mathsf{d}(W,\hat{Z})>\mathsf{D}\}$, where $\mathsf{d}:\mathcal{W}\times\mathcal{Z}\to\mathbb{R}_{\ge0}$ is a distortion measure. We show a one-shot joint source-channel coding result that achieves the optimal dispersion in [@kostina2013joint]. \[thm:jscc\]Fix any $P_{X}$ and $P_{Z}$. There exists a code for the source distribution $P_{W}$ and channel $P_{Y|X}$, with probability of excess distortion $$P_{e}\le\mathbf{E}\left[\left(1+P_{Z}(\mathcal{B}_{\mathsf{D}}(W))2^{\iota_{X;Y}(X;Y)}\right)^{-1}\right]$$ if $P_{XY}\ll P_{X}\times P_{Y}$, where $(W,X,Y)\sim P_{W}\times P_{X}P_{Y|X}$, and $\mathcal{B}_{\mathsf{D}}(w):=\{z:\,\mathsf{d}(w,z)\le\mathsf{D}\}$. Let $\{(\bar{X}_{i},\bar{Z}_{i}),T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{X}\times P_{Z}\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $W$. Let $\rho(w):=P_{Z}(\mathcal{B}_{\mathsf{D}}(w))$. Let $P_{\check{Z}|W}$ be defined as $$P_{\check{Z}|W}(A|w):=\begin{cases} P_{Z}(A\cap\mathcal{B}_{\mathsf{D}}(w))/\rho(w) & \mathrm{if}\,\rho(w)>0\\ P_{Z}(A) & \mathrm{if}\,\rho(w)=0. \end{cases}$$ The encoding function is $w\mapsto\tilde{X}_{P_{X}\times P_{\check{Z}|W}(\cdot|w)}$ (i.e., $X=\tilde{X}_{P_{X}\times P_{\check{Z}|W}(\cdot|W)}$). The decoding function is $y\mapsto\tilde{Z}_{P_{X|Y}(\cdot|y)\times P_{Z}}$ (i.e., $\hat{Z}=\tilde{Z}_{P_{X|Y}(\cdot|Y)\times P_{Z}}$). Also define $\check{Z}=\tilde{Z}_{P_{X}\times P_{\check{Z}|W}(\cdot|W)}$. We have $(X,Y,W,\check{Z})\sim P_{X}P_{Y|X}\times P_{W}P_{\check{Z}|W}$. $$\begin{aligned} & \mathbf{P}\{\mathsf{d}(W,\hat{Z})>\mathsf{D}\}\\ & \le\mathbf{P}\{\rho(W)=0\}+\mathbf{P}\{\rho(W)>0\;\mathrm{and}\;\check{Z}\neq\hat{Z}\}\\ & \le\mathbf{P}\{\rho(W)=0\}+\mathbf{E}\left[\mathbf{1}\{\rho(W)>0\}\mathbf{P}\{(X,\check{Z})\neq(\tilde{X},\tilde{Z})_{P_{X|Y}(\cdot|Y)\times P_{Z}}\,|\,X,Y,W,\check{Z}\}\right]\\ & \stackrel{(a)}{\le}\mathbf{P}\{\rho(W)=0\}+\mathbf{E}\left[\mathbf{1}\{\rho(W)>0\}\left(1-\left(1+\frac{dP_{X}\times P_{\check{Z}|W}(\cdot|W)}{dP_{X|Y}(\cdot|Y)\times P_{Z}}(X,\check{Z})\right)^{-1}\right)\right]\\ & =\mathbf{P}\{\rho(W)=0\}+\mathbf{E}\left[\mathbf{1}\{\rho(W)>0\}\left(1-\left(1+(\rho(W))^{-1}2^{-\iota_{X;Y}(X;Y)}\right)^{-1}\right)\right]\\ & =\mathbf{E}\left[\left(1+\rho(W)2^{\iota_{X;Y}(X;Y)}\right)^{-1}\right],\end{aligned}$$ where (a) is by the conditional Poisson matching lemma on $(W,\,(X,\check{Z}),\,Y,\,P_{X|Y}\times P_{Z})$ (note that $P_{X,\check{Z}|W}=P_{X}\times P_{\check{Z}|W}$). Therefore there exists a fixed $\{(\bar{x}_{i},\bar{z}_{i}),t_{i}\}_{i\in\mathbb{N}}$ attaining the desired bound. Compare this to Theorem 7 in [@kostina2013joint]: $$P_{e}\le\mathbf{E}\left[\min\left\{ J2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right]+\mathbf{E}\left[(1-P_{Z}(\mathcal{B}_{\mathsf{D}}(W)))^{J}\right]\label{eq:jscc_kostina}$$ for any $P_{J|W}$, $J\in\mathbb{N}$. While neither of the bounds implies the other, our bound is at least within a factor of 2 from , since $$\begin{aligned} & \mathbf{E}\left[\left(1+P_{Z}(\mathcal{B}_{\mathsf{D}}(W))2^{\iota_{X;Y}(X;Y)}\right)^{-1}\right]\\ & \le\mathbf{E}\left[\left(1+(2J)^{-1}2^{\iota_{X;Y}(X;Y)}\right)^{-1}\right]+\mathbf{P}\left\{ (2J)^{-1}\ge P_{Z}(\mathcal{B}_{\mathsf{D}}(W))\right\} \\ & \le\mathbf{E}\left[\min\left\{ 2J2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right]+2\mathbf{E}\left[\max\{1-JP_{Z}(\mathcal{B}_{\mathsf{D}}(W)),\,0\}\right]\\ & \le2\mathbf{E}\left[\min\left\{ J2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right]+2\mathbf{E}\left[(1-P_{Z}(\mathcal{B}_{\mathsf{D}}(W)))^{J}\right].\end{aligned}$$ However, does not imply a bound that is within a constant factor from our bound. Theorem 8 in [@kostina2013joint] is obtained by substituting $J=\lfloor\gamma/P_{Z}(\mathcal{B}_{\mathsf{D}}(W))\rfloor$ in : $$P_{e}\le\mathbf{E}\left[\min\left\{ \gamma P_{Z}(\mathcal{B}_{\mathsf{D}}(W))^{-1}2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right]+e^{1-\gamma},$$ which is strictly weaker than our bound with an unbounded multiplicative gap $\gamma$ (that tends to $\infty$ when the bound tends to 0). Hence our bound is stronger than Theorem 7 and 8 in [@kostina2013joint] (ignoring constant multiplicative gaps). Also our proof is significantly shorter than that of Theorem 7 in [@kostina2013joint]. Please refer to Appendix \[subsec:second\_jscc\] for the proof that Theorem \[thm:jscc\] achieves the optimal dispersion. Poisson Matching Lemma Beyond the First Index\[sec:gen\_pml\] ============================================================= The Poisson functional representation concerns the point with the smallest $T_{i}((dP/d\mu)(\bar{U}_{i}))^{-1}$. We can generalize it to obtain a sequence ordered in ascending order of $T_{i}((dP/d\mu)(\bar{U}_{i}))^{-1}$. \[Mapped Poisson process\] Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ on $\mathcal{U}\times\mathbb{R}_{\ge0}$ (where $\mathcal{U}$ is a Polish space with its Borel $\sigma$-algebra, and $\mu$ is $\sigma$-finite). For $P\ll\mu$ a probability measure over $\mathcal{U}$, let $i_{P,1},i_{P,2},\ldots\in\mathbb{N}$ be a sequence of distinct integers such that $\bigcup_{j=1}^{\infty}\{i_{P,j}\}=\{i:\,(dP/d\mu)(\bar{U}_{i})>0\}$ and $\{T_{i_{P,j}}((dP/d\mu)(\bar{U}_{i_{P,j}}))^{-1}\}_{j\in\mathbb{N}}$ is sorted in ascending order with arbitrary tie-breaking (a tie occurs with probability 0). For $j\in\mathbb{N},\,u\in\mathcal{U}$, define the *mapped Poisson process with respect to $P$* as $$\left\{ \tilde{U}_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}},\,j\right),\,\tilde{T}_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}},\,j\right)\right\} _{j\in\mathbb{N}},\label{eq:mappedpp}$$ where $$\tilde{T}_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}},\,j\right):=T_{i_{P,j}}\left(\frac{dP}{d\mu}(\bar{U}_{i_{P,j}})\right)^{-1},$$ $$\tilde{U}_{P}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}},\,j\right):=\bar{U}_{i_{P,j}}.$$ For $P,Q\ll\mu$ probability measures over $\mathcal{U}$, define $i_{P,1},i_{P,2},\ldots\in\mathbb{N}$ and $i_{Q,1},i_{Q,2},\ldots\in\mathbb{N}$ as above. Define $$\Upsilon_{P\Vert Q}\left(\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}},\,j\right):=\min\{k\in\mathbb{N}:\,i_{Q,k}=i_{P,j}\},$$ where the minimum is $\infty$ if such $k$ does not exist. We omit $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ and only write $\tilde{U}_{P}(j)$, $\tilde{T}_{P}(j)$, $\Upsilon_{P\Vert Q}(j)$ if the Poisson process is clear from the context. Note that, with probability 1, we have either $\tilde{U}_{Q}(\Upsilon_{P\Vert Q}(j))=\tilde{U}_{P}(j)$ or $\Upsilon_{P\Vert Q}(j)=\infty$. Also, for any $j,k\in\mathbb{N}$, $\Upsilon_{P\Vert Q}(j)=k\Leftrightarrow\Upsilon_{Q\Vert P}(k)=j$. Loosely speaking, $\Upsilon_{P\Vert Q}(j)$ can be regarded as “$\tilde{U}_{Q}^{-1}(\tilde{U}_{P}(j))$” (if there are no atoms in $\mu$), i.e., finding the $j$-th point in the mapped Poisson process w.r.t. $P$, then finding its index in the mapped Poisson process w.r.t. $Q$. While $dP/d\mu$ is only uniquely defined up to a $\mu$-null set, changing the value of $dP/d\mu$ on a $\mu$-null set will only affect the values of $\{\tilde{U}_{P}(j),\,\tilde{T}_{P}(j)\}_{j\in\mathbb{N}}$ on a null set with respect to the distribution of $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$, since the probability that there exists $\bar{U}_{i}$ in that $\mu$-null set is zero. Therefore $\{\tilde{U}_{P}(j),\,\tilde{T}_{P}(j)\}_{j\in\mathbb{N}}$ is uniquely defined up to a null set. The same is true for $\Upsilon_{P\Vert Q}(j)$. By the mapping theorem [@kingman1992poisson; @last2017lectures] (also see Appendix A of [@sfrl_trans]), $$\{\bar{U}_{i_{P,j}},\,T_{i_{P,j}}((dP/d\mu)(U_{i_{P,j}}))^{-1}\}_{j\in\mathbb{N}}=\{\tilde{U}_{P}(j),\tilde{T}_{P}(j)\}_{j\in\mathbb{N}}$$ is a Poisson process with intensity measure $P\times\lambda_{\mathbb{R}_{\ge0}}$. Hence $$\tilde{U}_{P}(1),\tilde{U}_{P}(2),\ldots\stackrel{iid}{\sim}P.$$ We present a generalized Poisson matching lemma concerning the indices beyond the first. The proof is given in Appendix \[subsec:pf\_phidiv\]. \[Generalized Poisson matching lemma\]\[lem:phidiv\_gen\]Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ on $\mathcal{U}\times\mathbb{R}_{\ge0}$, and $P,Q$ be probability measures over $\mathcal{U}$ with $P,Q\ll\mu$. Fix any $j\in\mathbb{N}$. Then we have the following almost surely: $$\mathbf{E}\left[\left.\Upsilon_{P\Vert Q}(j)\,\right|\,\tilde{U}_{P}(j)\right]\le j\frac{dP}{dQ}(\tilde{U}_{P}(j))+1,$$ where we write $(dP/dQ)(u)=(dP/d\mu)(u)/((dQ/d\mu)(u))$ as in (we do not require $P\ll Q$). As a result, we have the following almost surely: for all $k\in\mathbb{N}$, $$\begin{aligned} \mathbf{P}\left\{ \left.\tilde{U}_{P}(j)\notin\{\tilde{U}_{Q}(i)\}_{i\in[1:k]}\,\right|\,\tilde{U}_{P}(j)\right\} & \le\mathbf{P}\left\{ \left.\Upsilon_{P\Vert Q}(j)>k\,\right|\,\tilde{U}_{P}(j)\right\} \\ & \le\min\left\{ \frac{j}{k}\frac{dP}{dQ}(\tilde{U}_{P}(j)),\,1\right\} .\end{aligned}$$ For $k=1$, this can be slightly strengthened to $$\mathbf{P}\left\{ \left.\Upsilon_{P\Vert Q}(j)>1\,\right|\,\tilde{U}_{P}(j)\right\} \le1-\left(1-\min\left\{ \frac{dP}{dQ}(\tilde{U}_{P}(j)),\,1\right\} \right)^{j}.$$ For $j=1$, this can be slightly strengthened to: for all $k\in\mathbb{N}$, $$\begin{aligned} \mathbf{P}\left\{ \left.\Upsilon_{P\Vert Q}(1)>k\,\right|\,\tilde{U}_{P}(1)\right\} & \le\Bigl(1-\Bigl(1+\frac{dP}{dQ}(\tilde{U}_{P}(1))\Bigr)^{-1}\Bigr)^{k}\\ & \le1-\Bigl(1+k^{-1}\frac{dP}{dQ}(\tilde{U}_{P}(1))\Bigr)^{-1}.\end{aligned}$$ The exact distribution of $\Upsilon_{P\Vert Q}(j)$ is given in . Similar to Lemma \[lem:cond\_phidiv\], we can state a conditional version of the generalized Poisson matching lemma. The proof follows the same logic as Lemma \[lem:cond\_phidiv\] and is omitted. \[Conditional generalized Poisson matching lemma\]\[lem:cond\_phidiv\_gen\]Fix a distribution $P_{X,J,U,Y}$ and a probability kernel $Q_{U|Y}$, satisfying $J\in\mathbb{N}$ and $P_{U|X,J}(\cdot|X,J),Q_{U|Y}(\cdot|Y)\ll\mu$ almost surely. Let $(X,J)\sim P_{X,J}$, and $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $(X,J)$. Let $U=\tilde{U}_{P_{U|X,J}(\cdot|X,J)}(J)$ and $Y|(X,J,U,\{\bar{U}_{i},T_{i}\}_{i})\sim P_{Y|X,J,U}(\cdot|X,J,U)$ (note that $(X,J,U,Y)\sim P_{X,J,U,Y}$ and $Y\leftrightarrow(X,J,U)\leftrightarrow\{\bar{U}_{i},T_{i}\}_{i}$). Then we have the following almost surely: $$\mathbf{E}\left[\left.\Upsilon_{P_{U|X,J}(\cdot|X,J)\Vert Q_{U|Y}(\cdot|Y)}(J)\,\right|\,X,J,U,Y\right]\le J\frac{dP_{U|X,J}(\cdot|X,J)}{dQ_{U|Y}(\cdot|Y)}(U)+1,$$ and for all $k\in\mathbb{N}$, $$\mathbf{P}\left\{ \left.\Upsilon_{P_{U|X,J}(\cdot|X,J)\Vert Q_{U|Y}(\cdot|Y)}(J)>k\,\right|\,X,J,U,Y\right\} \le\min\left\{ \frac{J}{k}\frac{dP_{U|X,J}(\cdot|X,J)}{dQ_{U|Y}(\cdot|Y)}(U),\,1\right\} ,$$ and $$\mathbf{P}\left\{ \left.\Upsilon_{P_{U|X,J}(\cdot|X,J)\Vert Q_{U|Y}(\cdot|Y)}(J)>1\,\right|\,X,J,U,Y\right\} \le1-\left(1-\min\left\{ \frac{dP_{U|X,J}(\cdot|X,J)}{dQ_{U|Y}(\cdot|Y)}(U),\,1\right\} \right)^{J}.$$ If $J=1$ almost surely, then we also have the following almost surely: for all $k\in\mathbb{N}$, $$\begin{aligned} \mathbf{P}\left\{ \left.\Upsilon_{P_{U|X}(\cdot|X)\Vert Q_{U|Y}(\cdot|Y)}(1)>k\,\right|\,X,U,Y\right\} & \le\Bigl(1-\Bigl(1+\frac{dP_{U|X}(\cdot|X)}{dQ_{U|Y}(\cdot|Y)}(U)\Bigr)^{-1}\Bigr)^{k}\\ & \le1-\Bigl(1+k^{-1}\frac{dP_{U|X}(\cdot|X)}{dQ_{U|Y}(\cdot|Y)}(U)\Bigr)^{-1}.\end{aligned}$$ We can use the generalized Poisson matching lemma to extend Proposition \[prop:channel\] to the list decoding setting with fixed list size $\mathsf{J}$. The decoder outputs the list $\{\tilde{M}_{P_{X|Y}(\cdot|Y)\times P_{M}}(j)\}_{j\in[1:\mathsf{J}]}$. The error event becomes $(X,M)\notin\{(\tilde{X},\tilde{M})_{P_{X|Y}(\cdot|Y)\times P_{M}}(j)\}_{j\in[1:\mathsf{J}]}$. The probability of error is bounded by $\mathbf{E}\left[(1-(1+\mathsf{L}2^{-\iota_{X;Y}(X;Y)})^{-1})^{\mathsf{J}}\right]$. One-shot Coding for Broadcast Channels and Mutual Covering ========================================================== The one-shot coding setting for the broadcast channel with common message is described as follows. Upon observing three independent messages $M_{j}\sim\mathrm{Unif}[1:\mathsf{L}_{j}]$, $j=0,1,2$, the encoder produces $X$, which is sent through the broadcast channel $P_{Y_{1},Y_{2}|X}$. Decoder $j$ observes $Y_{j}$ and recovers $\hat{M}_{0j}$ and $\hat{M}_{j}$ ($j=1,2$). The error probability is $P_{e}=\mathbf{P}\{(M_{0},M_{0},M_{1},M_{2})\neq(\hat{M}_{01},\hat{M}_{02},\hat{M}_{1},\hat{M}_{2})\}$. We show a one-shot version of the inner bound in [@liang2007broadcast Theorem 5] (which is shown to be equivalent to [@gelfand1980capacity Theorem 1] in [@liang2011equivalence]). The proof is given in Appendix \[subsec:pf\_bc\_cm\]. \[thm:bc\_cm\]Fix any $P_{U_{0},U_{1},U_{2}}$ and function $x:\mathcal{U}_{0}\times\mathcal{U}_{1}\times\mathcal{U}_{2}\to\mathcal{X}$. For any $\mathsf{J},\mathsf{K}_{1},\mathsf{K}_{2}\in\mathbb{N}$, there exists a code for the broadcast channel $P_{Y_{1},Y_{2}|X}$ for independent messages $M_{j}\sim\mathrm{Unif}[1:\mathsf{L}_{j}]$, $j=0,1,2$, with the error probability bounded by $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{1}\mathsf{J}A2^{-\iota_{U_{0},U_{1};Y_{1}}(U_{0},U_{1};Y_{1})}+\tilde{\mathsf{L}}_{1}\mathsf{J}A2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\\ & \;\;\;\;\;\;+\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{2}\mathsf{J}^{-1}B2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})}+\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{2}(1-\mathsf{J}^{-1})B2^{-\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})}\\ & \;\;\;\;\;\;+\tilde{\mathsf{L}}_{2}\mathsf{J}^{-1}B2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}+\tilde{\mathsf{L}}_{2}(1-\mathsf{J}^{-1})B2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})},\,1\biggr\}\biggr]\end{aligned}$$ if all the information density terms are defined almost surely, where $$\begin{aligned} \tilde{\mathsf{L}}_{0} & :=\mathsf{L}_{0}\mathsf{K}_{1}\mathsf{K}_{2},\\ \tilde{\mathsf{L}}_{a} & :=\lceil\mathsf{L}_{a}/\mathsf{K}_{a}\rceil\;\mathrm{for}\,a=1,2,\\ A & :=(\log(\tilde{\mathsf{L}}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1)^{2},\\ B & :=\bigl(\log((\tilde{\mathsf{L}}_{2}\mathsf{J}^{-1}2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}\\ & \;\;\;\;+\tilde{\mathsf{L}}_{2}(1-\mathsf{J}^{-1})2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})})^{-1}+1)+1\bigr)^{2}.\end{aligned}$$ As a result, for $\gamma>0$, $$\begin{aligned} P_{e} & \le\mathbf{P}\biggl\{\log\tilde{\mathsf{L}}_{1}\mathsf{J}>\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})-\gamma\;\;\mathrm{or}\;\;\log\tilde{\mathsf{L}}_{2}>\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})-\gamma\nonumber \\ & \;\;\;\;\mathrm{or}\;\;\log\tilde{\mathsf{L}}_{2}\mathsf{J}^{-1}>\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})-\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\gamma\nonumber \\ & \;\;\;\;\mathrm{or}\;\;\log\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{1}\mathsf{J}>\iota_{U_{0},U_{1};Y_{1}}(U_{0},U_{1};Y_{1})-\gamma\;\;\mathrm{or}\;\;\log\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{2}>\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})-\gamma\nonumber \\ & \;\;\;\;\mathrm{or}\;\;\log\tilde{\mathsf{L}}_{0}\tilde{\mathsf{L}}_{2}\mathsf{J}^{-1}>\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})-\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\gamma\biggr\}\nonumber \\ & +2^{-\gamma}\left(8\mathbf{E}\left[(\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0}))^{2}+(\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0}))^{2}\right]+12\gamma^{2}+84\right).\label{eq:bc_pe2}\end{aligned}$$ The logarithmic terms $A$ and $B$ (or the last term in ) result in an $O(n^{-1}\log n)$ penalty on the rate in the finite blocklength regime, and do not affect the second order result. Ignoring the last term in , the error event in is a strict subset of those in [@yassaee2013oneshot eqn (32)] and [@liu2015oneshotmutual eqn (49)]. This is because the error event in [@yassaee2013oneshot] is a superset of by Fourier-Motzkin elimination on $\mathsf{J}_{2}$ in the error event in [@yassaee2013oneshot], but the reverse is not true since Fourier-Motzkin elimination only guarantees the existence of a random variable for $\mathsf{J}_{2}$ (that depends on the information density terms) satisfying the bounds, but $\mathsf{J}_{2}$ must be a constant since it is a parameter of the code construction in [@yassaee2013oneshot]. Theorem \[thm:bc\_cm\] gives the following second order bound. Consider $n$ independent channel uses. Let $\mathsf{L}_{a}=2^{nR_{a}}$ for $a=0,1,2$. By the multi-dimensional Berry-Esseen theorem [@bentkus2003dependence] (using the notation in [@yassaee2013oneshot]), we have $P_{e}\le\epsilon$ if there exists $\bar{R},\hat{R}_{1},\hat{R}_{2}\ge0$ such that $$\left[\begin{array}{c} \tilde{R}_{1}+\bar{R}\\ \tilde{R}_{2}\\ \tilde{R}_{2}-\bar{R}\\ \tilde{R}_{0}+\tilde{R}_{1}+\bar{R}\\ \tilde{R}_{0}+\tilde{R}_{2}\\ \tilde{R}_{0}+\tilde{R}_{2}-\bar{R} \end{array}\right]\in\mathbf{E}[I]-\frac{1}{\sqrt{n}}\mathcal{Q}^{-1}\left(\mathrm{Cov}[I],\,\epsilon-\frac{\beta}{\sqrt{n}}\right)-\frac{\beta\log n}{n}$$ if $n>\beta^{2}\epsilon^{-2}$, where $\beta$ is a constant that depends on $P_{U_{0},U_{1},U_{2},Y_{1},Y_{2}}$, and $\tilde{R}_{0}=R_{0}+\hat{R}_{1}+\hat{R}_{2}$, $\tilde{R}_{a}=R_{a}-\hat{R}_{a}$ for $a=1,2$, and $$I=\left[\begin{array}{c} \iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})\\ \iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})\\ \iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})-\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})\\ \iota_{U_{0},U_{1};Y_{1}}(U_{0},U_{1};Y_{1})\\ \iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})\\ \iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})-\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0}) \end{array}\right].$$ To demonstrate the use of the generalized Poisson matching lemma in place of the mutual covering lemma, we prove a one-shot version of Marton’s inner bound without common message [@marton1979broadcast] (i.e., $\mathsf{L}_{0}=1$). Our bound is stronger than that in [@verdu2012nonasymp] in the sense that our bound implies [@verdu2012nonasymp] (with a slight penalty of having $2^{1-\gamma}+2^{-2\gamma}$ instead of $2^{1-\gamma}+e^{-2^{\gamma}}$), but [@verdu2012nonasymp] does not imply our bound. We also note that a finite-blocklength bound is given in [@yassaee2013binning]. Nevertheless, the analysis in [@yassaee2013binning] only works for discrete auxiliary random variables $U_{1},U_{2}$, and does not appear to yield a one-shot bound due to the use of typical sequences. In the conventional mutual covering approach in [@yassaee2013oneshot; @liu2015oneshotmutual], sub-codebooks for both $U_{1}$ and $U_{2}$ are generated, whereas in our approach we generate a sub-codebook only for $U_{1}$, and the codebook of $U_{2}$ adapts to the sub-codebook automatically, eliminating the need for a sub-codebook for $U_{2}$. \[thm:bc\_ncm\]Fix any $P_{U_{1},U_{2}}$ and function $x:\mathcal{U}_{1}\times\mathcal{U}_{2}\to\mathcal{X}$. For any $\mathsf{J}\in\mathbb{N}$, there exists a code for the broadcast channel $P_{Y_{1},Y_{2}|X}$ for independent private messages $M_{j}\sim\mathrm{Unif}[1:\mathsf{L}_{j}]$, $j=1,2$, with the error probability bounded by $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}}(U_{1};Y_{1})}+\mathsf{L}_{2}(1-\mathsf{J}^{-1})2^{-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}+\mathsf{L}_{2}\mathsf{J}^{-1}2^{\iota_{U_{1};U_{2}}(U_{1};U_{2})-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})},\,1\biggr\}\biggr]\end{aligned}$$ if all the information density terms are defined, where $(U_{1},U_{2},X,Y_{1},Y_{2})\sim P_{U_{1}U_{2}}\delta_{x(U_{1},U_{2})}P_{Y_{1},Y_{2}|X}$. Let $\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$, $\{(\bar{U}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i\in\mathbb{N}}$ be two independent Poisson processes with intensity measures $P_{U_{1}}\times P_{M_{1}}\times\lambda_{\mathbb{R}_{\ge0}}$ and $P_{U_{2}}\times P_{M_{2}}\times\lambda_{\mathbb{R}_{\ge0}}$ respectively, independent of $M_{1},M_{2}$. The encoder would generate $X$ such that $$(M_{1},M_{2},K,\{\check{U}_{1j}\}_{j\in[1:\mathsf{J}]},U_{1},U_{2},X)\sim P_{M_{1}}\times P_{M_{2}}\times P_{K}P_{U_{1}}^{\otimes\mathsf{J}}\delta_{\check{U}_{1K}}P_{U_{2}|U_{1}}\delta_{x(U_{1},U_{2})},\label{eq:bc_dist}$$ where $P_{K}=\mathrm{Unif}[1:\mathsf{J}]$, and $\{\check{U}_{1j}\}_{j\in[1:\mathsf{J}]}\in\mathcal{U}_{1}^{\mathsf{J}}$ is an intermediate list (which can be regarded as a sub-codebook). The term $P_{U_{1}}^{\otimes\mathsf{J}}\delta_{\check{U}_{1K}}$ in means that $\{\check{U}_{1j}\}_{j}$ are i.i.d. $P_{U_{1}}$, and $U_{1}=\check{U}_{1K}$. To accomplish this, the encoder computes $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}}\times\delta_{M_{1}}}(j)$ for $j=1,\ldots,\mathsf{J}$ (which Poisson process we are referring to can be deduced from whether we are discussing $U_{1}$ or $U_{2}$), $U_{2}=(\tilde{U}_{2})_{\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}P_{U_{2}|U_{1}}(\cdot|\check{U}_{1j})\times\delta_{M_{2}}}$, and $(K,U_{1})|(\{\check{U}_{1j}\}_{j},U_{2})\sim P_{K,U_{1}|\{\check{U}_{1j}\}_{j},U_{2}}$ (where $P_{K,U_{1}|\{\check{U}_{1j}\}_{j},U_{2}}$ is derived from ), and outputs $X=x(U_{1},U_{2})$. It can be verified that is satisfied. The decoding functions are $\hat{M}_{1}=(\tilde{M}_{1})_{P_{U_{1}|Y_{1}}(\cdot|Y_{1})\times P_{M_{1}}}$, $\hat{M}_{2}=(\tilde{M}_{2})_{P_{U_{2}|Y_{2}}(\cdot|Y_{2})\times P_{M_{2}}}$. We have the following almost surely: $$\begin{aligned} & \mathbf{P}\biggl\{(\tilde{U}_{1},\tilde{M}_{1})_{P_{U_{1}|Y_{1}}(\cdot|Y_{1})\times P_{M_{1}}}\neq(U_{1},M_{1})\,\biggl|\,U_{1},U_{2},Y_{1},Y_{2},M_{1},K\biggr\}\\ & \stackrel{(a)}{=}\mathbf{P}\biggl\{(\tilde{U}_{1},\tilde{M}_{1})_{P_{U_{1}|Y_{1}}(\cdot|Y_{1})\times P_{M_{1}}}\neq(U_{1},M_{1})\,\biggl|\,U_{1},Y_{1},M_{1},K\biggr\}\\ & \stackrel{(b)}{\le}K\frac{dP_{U_{1}}\times\delta_{M_{1}}}{dP_{U_{1}|Y_{1}}(\cdot|Y_{1})\times P_{M_{1}}}(U_{1},M_{1})\\ & \le\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}}(U_{1};Y_{1})},\end{aligned}$$ where (a) is by $(U_{2},Y_{2})\leftrightarrow(U_{1},Y_{1},M_{1},K)\leftrightarrow\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$ (see Figure \[fig:bc\_bayes\] middle), and (b) is by the conditional generalized Poisson matching lemma on $(X,J,U,Y,Q_{U|Y})\leftarrow(M_{1},\,K,\,(U_{1},M_{1}),\,Y_{1},\,P_{U_{1}|Y_{1}}\times P_{M_{1}})$, since $P_{U_{1},M_{1}|M_{1},K}=P_{U_{1}}\times\delta_{M_{1}}$, $(M_{1},K)\perp\!\!\!\perp\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$, and $Y_{1}\leftrightarrow(U_{1},M_{1},K)\leftrightarrow\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$, which can be deduced from and $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}}\times\delta_{M_{1}}}(j)$ (see Figure \[fig:bc\_bayes\] middle). Also, almost surely, $$\begin{aligned} & \mathbf{P}\left\{ \left.(\tilde{U}_{2},\tilde{M}_{2})_{P_{U_{2}|Y_{2}}(\cdot|Y_{2})\times P_{M_{2}}}\neq(U_{2},M_{2})\,\right|\,U_{1},U_{2},Y_{1},Y_{2},M_{2}\right\} \\ & \stackrel{(a)}{=}\mathbf{P}\left\{ \left.(\tilde{U}_{2},\tilde{M}_{2})_{P_{U_{2}|Y_{2}}(\cdot|Y_{2})\times P_{M_{2}}}\neq(U_{2},M_{2})\,\right|\,U_{1},U_{2},Y_{2},M_{2}\right\} \\ & \stackrel{(b)}{\le}\mathbf{E}\left[\left.\frac{d(\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}P_{U_{2}|U_{1}}(\cdot|\check{U}_{1j}))\times\delta_{M_{2}}}{dP_{U_{2}|Y_{2}}(\cdot|Y_{2})\times P_{M_{2}}}(U_{2},M_{2})\,\right|\,U_{1},U_{2},Y_{2},M_{2}\right]\\ & =\mathbf{E}\left[\left.\mathsf{L}_{2}\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}2^{\iota_{U_{1};U_{2}}(\check{U}_{1j};U_{2})-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}\,\right|\,U_{1},U_{2},Y_{2},M_{2}\right]\\ & =\mathbf{E}\left[\left.\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}\Bigl(2^{\iota_{U_{1};U_{2}}(U_{1};U_{2})}+\sum_{j\in[1:\mathsf{J}]\backslash K}2^{\iota_{U_{1};U_{2}}(\check{U}_{1j};U_{2})}\Bigr)\,\right|\,U_{1},U_{2},Y_{2},M_{2}\right]\\ & =\mathbf{E}\left[\left.\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}\Bigl(2^{\iota_{U_{1};U_{2}}(U_{1};U_{2})}+\sum_{j=1}^{\mathsf{J}-1}2^{\iota_{U_{1};U_{2}}(\check{U}_{1,j+\mathbf{1}\{j\ge K\}};U_{2})}\Bigr)\,\right|\,U_{1},U_{2},Y_{2},M_{2}\right]\\ & \stackrel{(c)}{\le}\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}(2^{\iota_{U_{1};U_{2}}(U_{1};U_{2})}+\mathsf{J}-1),\end{aligned}$$ where (a) is by $Y_{1}\leftrightarrow(U_{1},U_{2},Y_{2},M_{2})\leftrightarrow\{(\bar{U}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i}$ (see Figure \[fig:bc\_bayes\] right), (b) is by the conditional Poisson matching lemma on $((\{\check{U}_{1j}\}_{j},M_{2}),\,(U_{2},M_{2}),\,Y_{2},\,P_{U_{2}|Y_{2}}\times P_{M_{2}})$, and (c) is because $\{\check{U}_{1,j+\mathbf{1}\{j\ge K\}}\}_{j\in[1:\mathsf{J}-1]}$ (the $\check{U}_{1j}$’s not selected as $U_{1}$) are independent of $(U_{1},U_{2},Y_{2},M_{2})$, $\mathbf{E}[2^{\iota_{U_{1};U_{2}}(\check{U}_{1,j+\mathbf{1}\{j\ge K\}};U_{2})}\,|\,U_{2}]=1$, and Jensen’s inequality. Hence, $$\begin{aligned} & \mathbf{P}\{(M_{1},M_{2})\neq(\hat{M}_{1},\hat{M}_{2})\}\\ & =\mathbf{E}\left[\mathbf{P}\left\{ \left.(M_{1},M_{2})\neq(\hat{M}_{1},\hat{M}_{2})\,\right|\,U_{1},U_{2},Y_{1},Y_{2}\right\} \right]\\ & \le\mathbf{E}\left[\min\left\{ \mathbf{P}\left\{ \left.M_{1}\neq\hat{M_{1}}\,\right|\,U_{1},U_{2},Y_{1},Y_{2}\right\} +\mathbf{P}\left\{ \left.M_{2}\neq\hat{M_{2}}\,\right|\,U_{1},U_{2},Y_{1},Y_{2}\right\} ,\,1\right\} \right]\\ & \le\mathbf{E}\biggl[\min\Bigl\{\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}}(U_{1};Y_{1})}+\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2};Y_{2}}(U_{2};Y_{2})}(2^{\iota_{U_{1};U_{2}}(U_{1};U_{2})}+\mathsf{J}-1),\,1\Bigr\}\biggr].\end{aligned}$$ Therefore there exist fixed realizations of the Poisson processes attaining the desired bound. ![\[fig:bc\_bayes\]Left: The Bayesian network described in . Middle: The Bayesian network deduced from and $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}}\times\delta_{M_{1}}}(j)$. Right: The Bayesian network describing the encoding scheme. Note that all three are valid Bayesian networks, and the desired conditional independence relations can be deduced using d-separation.](bc_bayes) One-shot Distributed Lossy Source Coding ======================================== The one-shot distributed lossy source coding setting is described as follows. Let $(X_{1},X_{2})\sim P_{X_{1},X_{2}}$. Upon observing $X_{j}$, encoder $j$ produces $M_{j}\in[1:\mathsf{L}_{j}]$, $j=1,2$. The decoder observes $M_{1},M_{2}$ and recovers $\hat{Z}_{1}\in\mathcal{Z}_{1}$, $\hat{Z}_{2}\in\mathcal{Z}_{2}$ with probability of excess distortion $P_{e}=\mathbf{P}\{\mathsf{d}_{1}(X_{1},\hat{Z}_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},\hat{Z}_{2})>\mathsf{D}_{2}\}$, where $\mathsf{d}_{j}:\mathcal{X}_{j}\times\mathcal{Z}_{j}\to\mathbb{R}_{\ge0}$ is a distortion measure for $j=1,2$. We show a one-shot version of the Berger-Tung inner bound [@berger1978multiterminal; @tung1978multiterminal]. \[thm:dlsc\]Fix any $P_{U_{1}|X_{1}}$, $P_{U_{2}|X_{2}}$ and functions $z_{j}:\mathcal{U}_{1}\times\mathcal{U}_{2}\to\mathcal{Z}_{j}$, $j=1,2$. There exists a code for distributed lossy source coding with sources $P_{X_{1}},P_{X_{2}}$ and message sizes $\mathsf{L}_{1},\mathsf{L}_{2}$, with probability of excess distortion $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathbf{1}\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\}+\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})}\nonumber \\ & \;\;\;\;+\left(\mathsf{L}_{1}^{-1}\mathsf{L}_{2}^{-1}2^{\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})}+\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\right)\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2},\,1\biggr\}\biggr]\label{eq:dlsc_pe1}\end{aligned}$$ if all the information density terms are defined, where $(X_{1},X_{2},U_{1},U_{2},Z_{1},Z_{2})\sim P_{X_{1},X_{2}}P_{U_{1}|X_{1}}P_{U_{2}|X_{2}}\delta_{z_{1}(U_{1},U_{2})}\delta_{z_{2}(U_{1},U_{2})}$. As a result, for $\gamma>0$, $$\begin{aligned} P_{e} & \le\mathbf{P}\biggl\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\;\;\mathrm{or}\;\;\log\mathsf{L}_{1}<\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})+\gamma\nonumber \\ & \;\;\;\;\mathrm{or}\;\;\log\mathsf{L}_{2}<\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})+\gamma\;\;\mathrm{or}\;\;\log\mathsf{L}_{1}\mathsf{L}_{2}<\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})+\gamma\biggr\}\nonumber \\ & \;\;\;+2^{-\gamma}\left(4\mathbf{E}[(\iota_{U_{1};U_{2}}(U_{1};U_{2}))^{2}]+4\gamma^{2}+29\right).\label{eq:dlsc_pe2}\end{aligned}$$ The logarithmic term in (or the last term in ) results in an $O(n^{-1}\log n)$ penalty on the rate in the finite blocklength regime, and does not affect the second order result. Ignoring the last term in , the error event in is a strict subset of that in [@yassaee2013oneshot eqn (47)]. This is because the error event in [@yassaee2013oneshot] is a superset of by Fourier-Motzkin elimination on $\mathsf{J}_{1},\mathsf{J}_{2}$ in the error event in [@yassaee2013oneshot], but the reverse is not true since Fourier-Motzkin elimination only guarantees the existence of random variables for $\mathsf{J}_{1},\mathsf{J}_{2}$ (that depend on the information density terms) satisfying the bounds, but $\mathsf{J}_{1},\mathsf{J}_{2}$ must be constants since they are parameters of the code construction in [@yassaee2013oneshot]. We now prove the result. Unlike previous approaches, our proof does not require binning. The encoders are the same as those for point-to-point lossy source coding. Let $\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$, $\{(\bar{U}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i\in\mathbb{N}}$ be two independent Poisson processes with intensity measures $P_{U_{1}}\times P_{M_{1}}\times\lambda_{\mathbb{R}_{\ge0}}$ and $P_{U_{2}}\times P_{M_{2}}\times\lambda_{\mathbb{R}_{\ge0}}$ respectively, independent of $X_{1},X_{2}$. The encoding functions are $M_{j}=(\tilde{M}_{j})_{P_{U_{j}|X_{j}}(\cdot|X_{j})\times P_{M_{j}}}$, $j=1,2$ (which Poisson process we are referring to can be deduced from whether we are discussing $M_{1}$ or $M_{2}$). Also define $U_{j}=(\tilde{U}_{j})_{P_{U_{j}|X_{j}}(\cdot|X_{j})\times P_{M_{j}}}$, $Z_{j}=z_{j}(U_{1},U_{2})$, $j=1,2$. For the decoding function, let $\check{U}_{1k}=(\tilde{U}_{1})_{P_{U_{1}}\times\delta_{M_{1}}}(k)$ for $k\in\mathbb{N}$, $\hat{U}_{2}=(\tilde{U}_{2})_{\sum_{k=1}^{\infty}\phi(k)P_{U_{2}|U_{1}}(\cdot|\check{U}_{1k})\times\delta_{M_{2}}}$ where $\phi(k)\propto k^{-1}(\log(k+2))^{-2}$ with $\sum_{k=1}^{\infty}\phi(k)=1$, and $\hat{U}_{1}=(\tilde{U}_{1})_{P_{U_{1}|U_{2}}(\cdot|\hat{U}_{2})\times\delta_{M_{1}}}$, $\hat{Z}_{j}=z_{j}(\hat{U}_{1},\hat{U}_{2})$, $j=1,2$. Note that $(M_{1},M_{2},X_{1},X_{2},U_{1},U_{2},Z_{1},Z_{2})\sim P_{M_{1}}\times P_{M_{2}}\times P_{X_{1},X_{2}}P_{U_{1}|X_{1}}P_{U_{2}|X_{2}}\delta_{z_{1}(U_{1},U_{2})}\delta_{z_{2}(U_{1},U_{2})}$. Let $K=\Upsilon_{P_{U_{1}|X_{1}}(\cdot|X_{1})\times P_{M_{1}}\Vert P_{U_{1}}\times\delta_{M_{1}}}(1)$ (using the Poisson process $\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$). By the conditional generalized Poisson matching lemma on $(X_{1},\,1,\,(U_{1},M_{1}),\,M_{1},\,P_{U_{1}}\times\delta_{M_{1}})$ (note that $P_{U_{1},M_{1}|X_{1}}=P_{U_{1}|X_{1}}\times P_{M_{1}}$), almost surely, $$\begin{aligned} \mathbf{E}\left[\left.K\,\right|\,X_{1},U_{1},M_{1}\right] & \le\frac{dP_{U_{1}|X_{1}}(\cdot|X_{1})\times P_{M_{1}}}{dP_{U_{1}}\times\delta_{M_{1}}}(U_{1},M_{1})+1\nonumber \\ & =\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}}(U_{1};X_{1})}+1.\label{eq:dlsc_ek}\end{aligned}$$ Since $\{\check{U}_{1k}\}_{k}$ is a function of $\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$ and $M_{1}$, we have $\{\check{U}_{1k}\}_{k}\leftrightarrow(X_{1},X_{2},U_{1},U_{2},M_{2})\leftrightarrow\{(\bar{U}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i}$. By the conditional Poisson matching lemma on $(X_{2},\,(U_{2},M_{2}),\,(\{\check{U}_{1k}\}_{k},M_{2}),\,\sum_{k=1}^{\infty}\phi(k)P_{U_{2}|U_{1}}(\cdot|\check{U}_{1k})\times\delta_{M_{2}})$ (note that $P_{U_{2},M_{2}|X_{2}}=P_{U_{2}|X_{2}}\times P_{M_{2}}$), almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{\left.(\tilde{U}_{2},\tilde{M}_{2})_{\sum_{k=1}^{\infty}\phi(k)P_{U_{2}|U_{1}}(\cdot|\check{U}_{1k})\times\delta_{M_{2}}}\neq(U_{2},M_{2})\,\right|\,X_{1},X_{2},U_{1},U_{2},M_{2}\biggr\}\\ & \le\mathbf{E}\left[\left.\min\biggl\{\frac{dP_{U_{2}|X_{2}}(\cdot|X_{2})\times P_{M_{2}}}{d(\sum_{k=1}^{\infty}\phi(k)P_{U_{2}|U_{1}}(\cdot|\check{U}_{1k}))\times\delta_{M_{2}}}(U_{2},M_{2}),\,1\biggr\}\,\right|\,X_{1},X_{2},U_{1},U_{2},M_{2}\right]\\ & \le\mathbf{E}\left[\left.\min\biggl\{\mathsf{L}_{2}^{-1}\frac{dP_{U_{2}|X_{2}}(\cdot|X_{2})}{\phi(K)dP_{U_{2}|U_{1}}(\cdot|U_{1})}(U_{2}),\,1\biggr\}\,\right|\,X_{1},X_{2},U_{1},U_{2},M_{2}\right]\\ & =\mathbf{E}\left[\left.\min\{\mathsf{L}_{2}^{-1}(\phi(K))^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})},\,1\}\,\right|\,X_{1},X_{2},U_{1},U_{2},M_{2}\right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[\left.K\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2}\,\right|\,X_{1},X_{2},U_{1},U_{2},M_{2}\right]\\ & \stackrel{(b)}{\le}\left(\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}}(U_{1};X_{1})}+1\right)\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2}\\ & =\left(\mathsf{L}_{1}^{-1}\mathsf{L}_{2}^{-1}2^{\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})}+\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\right)\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2},\end{aligned}$$ where (a) is by Proposition \[prop:phi\_ineq\], and (b) is by $K\leftrightarrow(U_{1},X_{1})\leftrightarrow(X_{2},U_{2},M_{2})$, and Jensen’s inequality. By the conditional Poisson matching lemma on $(X_{1},\,(U_{1},M_{1}),\,(U_{2},M_{1}),\,P_{U_{1}|U_{2}}\times\delta_{M_{1}})$ (note that $P_{U_{1},M_{1}|X_{1}}=P_{U_{1}|X_{1}}\times P_{M_{1}}$), and $X_{2}\leftrightarrow(X_{1},U_{1},U_{2},M_{1})\leftrightarrow\{(\bar{U}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$, almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{(\tilde{U}_{1},\tilde{M}_{1})_{P_{U_{1}|U_{2}}(\cdot|U_{2})\times\delta_{M_{1}}}\neq(U_{1},M_{1})\,\biggl|\,X_{1},X_{2},U_{1},U_{2},M_{1}\biggr\}\\ & \le\frac{dP_{U_{1}|X_{1}}(\cdot|X_{1})\times P_{M_{1}}}{dP_{U_{1}|U_{2}}(\cdot|U_{2})\times\delta_{M_{1}}}(U_{1},M_{1})\\ & =\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})}.\end{aligned}$$ We have $$\begin{aligned} & \mathbf{P}\{\mathsf{d}_{1}(X_{1},\hat{Z}_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},\hat{Z}_{2})>\mathsf{D}_{2}\}\\ & \le\mathbf{E}\biggl[\mathbf{P}\biggl\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\;\mathrm{or}\;\hat{U}_{2}\neq U_{2}\\ & \;\;\;\;\;\mathrm{or}\;(\hat{U}_{2}=U_{2}\;\mathrm{and}\;\hat{U}_{1}\neq U_{1})\,\biggl|\,X_{1},X_{2},U_{1},U_{2}\biggr\}\biggr]\\ & \le\mathbf{E}\biggl[\min\biggl\{\mathbf{1}\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\}+\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})}\\ & \;\;\;\;+\left(\mathsf{L}_{1}^{-1}\mathsf{L}_{2}^{-1}2^{\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})}+\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\right)\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2},\,1\biggr\}\biggr]\end{aligned}$$ Therefore there exist fixed values of the Poisson processes attaining the desired bound. For , if the event in does not occur, by Proposition \[prop:phi\_ineq\] with $\alpha=\gamma-1$, $\tilde{\alpha}=\gamma$, $\beta=\iota_{U_{1};U_{2}}(U_{1};U_{2})-\gamma$, $$\begin{aligned} & \mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})}\\ & \;\;\;+\left(\mathsf{L}_{1}^{-1}\mathsf{L}_{2}^{-1}2^{\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})}+\mathsf{L}_{2}^{-1}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}\right)\left(\log(\mathsf{L}_{2}2^{-\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+1)+1\right)^{2}\\ & \le2^{-\gamma}+2^{1-\gamma}\left(2(\iota_{U_{1};U_{2}}(U_{1};U_{2}))^{2}+2\gamma^{2}+14\right)\\ & =2^{-\gamma}\left(4(\iota_{U_{1};U_{2}}(U_{1};U_{2}))^{2}+4\gamma^{2}+29\right).\end{aligned}$$ The reason for the logarithmic term is that we want to translate a bound on $\mathbf{E}[K]$ (given by the generalized Poisson matching lemma) into a bound on $\mathbf{E}[(\phi(K))^{-1}]$ for some distribution $\phi$ over $\mathbb{N}$. Ideally, we wish $(\phi(k))^{-1}\propto k$, but this is impossible since the harmonic series diverges. Therefore we use a slow converging series $\phi(k)\propto k^{-1}(\log(k+2))^{-2}$ instead, resulting in a logarithmic penalty. If we use $\mathsf{J}^{-1}\mathbf{1}\{k\le\mathsf{J}\}$ instead of $\phi(k)$ in the proof, we can obtain the following bound for any $\mathsf{J}\in\mathbb{N}$: $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathbf{1}\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\}\\ & \;\;\;+\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};X_{1}}(U_{1};X_{1})}+\mathsf{L}_{2}^{-1}\mathsf{J}2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})},\,1\biggr\}\biggr].\end{aligned}$$ Compared to Theorem \[thm:dlsc\], this does not contain the logarithmic term, but requires optimizing over $\mathsf{J}$, and may give a worse second order result. Another choice is to use $g(k)\propto k^{-1}\mathbf{1}\{k\le\mathsf{J}\}$ instead of $\phi(k)$. We can obtain the following bound for any $\mathsf{J}\in\mathbb{N}$: $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathbf{1}\{\mathsf{d}_{1}(X_{1},Z_{1})>\mathsf{D}_{1}\;\mathrm{or}\;\mathsf{d}_{2}(X_{2},Z_{2})>\mathsf{D}_{2}\}+\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};X_{1}}(U_{1};X_{1})}\\ & \;\;+\mathsf{L}_{1}^{-1}\mathsf{L}_{2}^{-1}(\ln\mathsf{J}+1)2^{\iota_{U_{1},U_{2};X_{1},X_{2}}(U_{1},U_{2};X_{1},X_{2})}+\mathsf{L}_{2}^{-1}(\ln\mathsf{J}+1)2^{\iota_{U_{2};X_{2}|U_{1}}(U_{2};X_{2}|U_{1})}+\mathsf{L}_{1}^{-1}2^{\iota_{U_{1};X_{1}|U_{2}}(U_{1};X_{1}|U_{2})},\,1\biggr\}\biggr].\end{aligned}$$ which gives the same second order result as Theorem \[thm:dlsc\]. Nevertheless, we prefer using $\phi(k)$ which eliminates the need for a parameter $\mathsf{J}$ at the decoder. One-shot Coding for Multiple Access Channels ============================================ The one-shot coding setting for the multiple access channel is described as follows. Upon observing $M_{j}\sim\mathrm{Unif}[1:\mathsf{L}_{j}]$ ($M_{1},M_{2}$ independent), encoder $j$ produces $X_{j}$, $j=1,2$. The decoder observes the output $Y$ of the channel $P_{Y|X_{1},X_{2}}$ and recovers $(\hat{M}_{1},\hat{M}_{2})$. The error probability is $P_{e}=\mathbf{P}\{(M_{1},M_{2})\neq(\hat{M}_{1},\hat{M}_{2})\}$. We present a one-shot achievability result for the capacity region in [@ahlswede1971multi; @liao1972multiple; @ahlswede1974capacity]. While this result is slightly weaker than that in [@verdu2012nonasymp], we include it to illustrate the use of the generalized Poisson matching lemma in simultaneous decoding. Note that the logarithmic term results in an $O(n^{-1}\log n)$ penalty on the rate in the finite blocklength regime, and does not affect the second order result. \[thm:mac\]Fix any $P_{X_{1}},P_{X_{2}}$. There exists a code for the multiple access channel $P_{Y|X_{1},X_{2}}$ for messages $M_{j}\sim\mathrm{Unif}[1:\mathsf{L}_{j}]$, $j=1,2$, with the error probability bounded by $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\left(\mathsf{L}_{1}\mathsf{L}_{2}2^{-\iota_{X_{1},X_{2};Y}(X_{1},X_{2};Y)}+\mathsf{L}_{2}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\right)\left(\log(\mathsf{L}_{2}^{-1}2^{\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+1)+1\right)^{2}\\ & \;\;\;\;\;\;\;\;\;\;+\mathsf{L}_{1}2^{-\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)},\,1\biggr\}\biggr]\end{aligned}$$ if $P_{X_{1}X_{2}Y}\ll P_{X_{1}}\times P_{X_{2}}\times P_{Y}$, where $(X_{1},X_{2},Y)\sim P_{X_{1}}P_{X_{2}}P_{Y|X_{1},X_{2}}$. As a result, for $\gamma>0$, $$\begin{aligned} P_{e} & \le\mathbf{P}\biggl\{\log\mathsf{L}_{1}>\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)-\gamma\;\;\mathrm{or}\;\;\log\mathsf{L}_{2}>\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)-\gamma\nonumber \\ & \;\;\;\;\mathrm{or}\;\;\log\mathsf{L}_{1}\mathsf{L}_{2}>\iota_{X_{1},X_{2};Y}(X_{1},X_{2};Y)-\gamma\biggr\}+2^{-\gamma}\left(4\mathbf{E}[(\iota_{X_{1};X_{2}|Y}(X_{1};X_{2}|Y))^{2}]+4\gamma^{2}+29\right).\label{eq:mac_pe2}\end{aligned}$$ Let $\{(\bar{X}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$, $\{(\bar{X}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i\in\mathbb{N}}$ be two independent Poisson processes with intensity measures $P_{X_{1}}\times P_{M_{1}}\times\lambda_{\mathbb{R}_{\ge0}}$ and $P_{X_{2}}\times P_{M_{2}}\times\lambda_{\mathbb{R}_{\ge0}}$ respectively, independent of $M_{1},M_{2}$. The encoding functions are $X_{1}=(\tilde{X}_{1})_{P_{X_{1}}\times\delta_{M_{1}}}$, $X_{2}=(\tilde{X}_{2})_{P_{X_{2}}\times\delta_{M_{2}}}$ (which Poisson process we are referring to can be deduced from whether we are discussing $X_{1}$ or $X_{2}$). For the decoding function, let $\check{X}_{1k}=(\tilde{X}_{1})_{P_{X_{1}|Y}(\cdot|Y)\times P_{M_{1}}}(k)$ for $k\in\mathbb{N}$, $(\hat{X}_{2},\hat{M}_{2})=(\tilde{X}_{2},\tilde{M}_{2})_{\sum_{k=1}^{\infty}\phi(k)P_{X_{2}|X_{1},Y}(\cdot|\check{X}_{1k},Y)\times P_{M_{2}}}$ where $\phi(k)\propto k^{-1}(\log(k+2))^{-2}$ with $\sum_{k=1}^{\infty}\phi(k)=1$, and $\hat{M}_{1}=(\tilde{M}_{1})_{P_{X_{1}|X_{2},Y}(\cdot|\hat{X}_{2},Y)\times P_{M_{1}}}$. Let $K=\Upsilon_{P_{X_{1}}\times\delta_{M_{1}}\Vert P_{X_{1}|Y}(\cdot|Y)\times P_{M_{1}}}(1)$ (using the Poisson process $\{(\bar{X}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$). By the conditional generalized Poisson matching lemma on $(M_{1},\,1,\,(X_{1},M_{1}),\,Y,\,P_{X_{1}|Y}\times P_{M_{1}})$ (note that $P_{X_{1},M_{1}|M_{1}}=P_{X_{1}}\times\delta_{M_{1}}$), almost surely, $$\begin{aligned} \mathbf{E}\left[\left.K\,\right|\,X_{1},Y,M_{1}\right] & \le\frac{dP_{X_{1}}\times\delta_{M_{1}}}{dP_{X_{1}|Y}(\cdot|Y)\times P_{M_{1}}}(X_{1},M_{1})+1\nonumber \\ & =\mathsf{L}_{1}2^{-\iota_{X_{1};Y}(X_{1};Y)}+1.\label{eq:mac_ek}\end{aligned}$$ Since $\{\check{X}_{1k}\}_{k}$ is a function of $\{(\bar{X}_{1,i},\bar{M}_{1,i}),T_{1,i}\}_{i}$ and $Y$, we have $\{\check{X}_{1k}\}_{k}\leftrightarrow(X_{1},X_{2},Y,M_{2})\leftrightarrow\{(\bar{X}_{2,i},\bar{M}_{2,i}),T_{2,i}\}_{i}$. By the conditional Poisson matching lemma on $(M_{2},\,(X_{2},M_{2}),\,(\{\check{X}_{1k}\}_{k},Y),\,\sum_{k=1}^{\infty}\phi(k)P_{X_{2}|X_{1},Y}(\cdot|\check{X}_{1k},Y)\times P_{M_{2}})$ (note that $P_{X_{2},M_{2}|M_{2}}=P_{X_{2}}\times\delta_{M_{2}}$), almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{\left.(\tilde{X}_{2},\tilde{M}_{2})_{\sum_{k=1}^{\infty}\phi(k)P_{X_{2}|X_{1},Y}(\cdot|\check{X}_{1k},Y)\times P_{M_{2}}}\neq(X_{2},M_{2})\,\right|\,X_{1},X_{2},Y,M_{2}\biggr\}\\ & \le\mathbf{E}\left[\left.\min\biggl\{\frac{dP_{X_{2}}\times\delta_{M_{2}}}{d(\sum_{k=1}^{\infty}\phi(k)P_{X_{2}|X_{1},Y}(\cdot|\check{X}_{1k},Y))\times P_{M_{2}}}(X_{2},M_{2}),\,1\biggr\}\,\right|\,X_{1},X_{2},Y,M_{2}\right]\\ & \le\mathbf{E}\left[\left.\min\biggl\{\mathsf{L}_{2}\frac{dP_{X_{2}}}{\phi(K)dP_{X_{2}|X_{1},Y}(\cdot|X_{1},Y)}(X_{2}),\,1\biggr\}\,\right|\,X_{1},X_{2},Y,M_{2}\right]\\ & =\mathbf{E}\left[\left.\min\{\mathsf{L}_{2}(\phi(K))^{-1}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)},\,1\}\,\right|\,X_{1},X_{2},Y,M_{2}\right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[\left.K\mathsf{L}_{2}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\left(\log(\mathsf{L}_{2}^{-1}2^{\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+1)+1\right)^{2}\,\right|\,X_{1},X_{2},Y,M_{2}\right]\\ & \stackrel{(b)}{\le}\left(\mathsf{L}_{1}2^{-\iota_{X_{1};Y}(X_{1};Y)}+1\right)\mathsf{L}_{2}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\left(\log(\mathsf{L}_{2}^{-1}2^{\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+1)+1\right)^{2}\\ & =\left(\mathsf{L}_{1}\mathsf{L}_{2}2^{-\iota_{X_{1},X_{2};Y}(X_{1},X_{2};Y)}+\mathsf{L}_{2}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\right)\left(\log(\mathsf{L}_{2}^{-1}2^{\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+1)+1\right)^{2},\end{aligned}$$ where (a) is by Proposition \[prop:phi\_ineq\], and (b) is by $K\leftrightarrow(X_{1},Y)\leftrightarrow X_{2}$, and Jensen’s inequality. By the conditional Poisson matching lemma on $(M_{1},\,(X_{1},M_{1}),\,(X_{2},Y),\,P_{X_{1}|X_{2},Y}\times P_{M_{1}})$ (note that $P_{X_{1},M_{1}|M_{1}}=P_{X_{1}}\times\delta_{M_{1}}$), almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{(\tilde{X}_{1},\tilde{M}_{1})_{P_{X_{1}|X_{2},Y}(\cdot|X_{2},Y)\times P_{M_{1}}}\neq(X_{1},M_{1})\,\biggl|\,X_{1},X_{2},Y,M_{1}\biggr\}\\ & \le\frac{dP_{X_{1}}\times\delta_{M_{1}}}{dP_{X_{1}|X_{2},Y}(\cdot|X_{2},Y)\times P_{M_{1}}}(X_{1},M_{1})\\ & =\mathsf{L}_{1}2^{-\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)}.\end{aligned}$$ Therefore there exist fixed values of the Poisson processes attaining the desired bound. For , if the event in does not occur, by Proposition \[prop:phi\_ineq\] with $\alpha=\gamma-1$, $\tilde{\alpha}=\gamma$, $\beta=\iota_{X_{1};X_{2}|Y}(X_{1};X_{2}|Y)-\gamma$, $$\begin{aligned} & \left(\mathsf{L}_{1}\mathsf{L}_{2}2^{-\iota_{X_{1},X_{2};Y}(X_{1},X_{2};Y)}+\mathsf{L}_{2}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\right)\left(\log(\mathsf{L}_{2}^{-1}2^{\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+1)+1\right)^{2}+\mathsf{L}_{1}2^{-\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)}\\ & \le2^{1-\gamma}\left(2(\iota_{X_{1};X_{2}|Y}(X_{1};X_{2}|Y))^{2}+2\gamma^{2}+14\right)+2^{-\gamma}\\ & =2^{-\gamma}\left(4(\iota_{X_{1};X_{2}|Y}(X_{1};X_{2}|Y))^{2}+4\gamma^{2}+29\right).\end{aligned}$$ If we use $\mathsf{J}^{-1}\mathbf{1}\{k\le\mathsf{J}\}$ instead of $\phi(k)$ in the proof, we can obtain the following bound for any $\mathsf{J}\in\mathbb{N}$: $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathsf{L}_{1}\mathsf{J}^{-1}2^{-\iota_{X_{1};Y}(X_{1};Y)}+\mathsf{L}_{2}\mathsf{J}2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}+\mathsf{L}_{1}2^{-\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)},\,1\biggr\}\biggr].\end{aligned}$$ Compared to Theorem \[thm:mac\], this does not contain the logarithmic term, but requires optimizing over $\mathsf{J}$, and may give a worse second order result. Another choice is to use $g(k)\propto k^{-1}\mathbf{1}\{k\le\mathsf{J}\}$ instead of $\phi(k)$. We can obtain the following bound for any $\mathsf{J}\in\mathbb{N}$: $$\begin{aligned} P_{e} & \le\mathbf{E}\biggl[\min\biggl\{\mathsf{L}_{1}\mathsf{L}_{2}(\ln\mathsf{J}+1)2^{-\iota_{X_{1},X_{2};Y}(X_{1},X_{2};Y)}+\mathsf{L}_{2}(\ln\mathsf{J}+1)2^{-\iota_{X_{2};X_{1},Y}(X_{2};X_{1},Y)}\\ & \;\;\;\;\;\;\;\;\;\;+\mathsf{L}_{1}2^{-\iota_{X_{1};X_{2},Y}(X_{1};X_{2},Y)}+\mathsf{L}_{1}\mathsf{J}^{-1}2^{-\iota_{X_{1};Y}(X_{1};Y)},\,1\biggr\}\biggr],\end{aligned}$$ which gives the same second order result as Theorem \[thm:mac\]. Nevertheless, we prefer using $\phi(k)$ which eliminates the need for a parameter $\mathsf{J}$ at the decoder. One-shot Channel Resolvability and Soft Covering ================================================ The one-shot channel resolvability setting [@han1993approximation] is described as follows. Fix a channel $P_{Y|X}$ and input distribution $P_{X}$. Upon observing an integer $M\sim\mathrm{Unif}[1:\mathsf{L}]$, the encoder applies a deterministic mapping $g:[1:\mathsf{L}]\to\mathcal{X}$ on $M$ to produce $\hat{X}=g(M)$, which is sent through the channel $P_{Y|X}$ and gives the output $\hat{Y}$. The goal is to minimize the total variation distance between $P_{\hat{Y}}$ and $P_{Y}$ ($Y$-marginal of $P_{X}P_{Y|X}$), i.e., $\epsilon:=\Vert\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X}(\cdot|g(m))-P_{Y}(\cdot)\Vert_{\mathrm{TV}}$. We show a one-shot channel resolvability result using the the Poisson matching lemma. This result can also be regarded as a one-shot soft covering lemma [@cuff2013synthesis]. \[prop:resolve\]Given channel $P_{Y|X}$ and input distribution $P_{X}$ with $P_{XY}\ll P_{X}\times P_{Y}$. Let $\{\check{X}_{m}\}_{m\in[1:\mathsf{L}]}\stackrel{iid}{\sim}P_{X}$, then for any $\mathsf{J}\in\mathbb{N}$, $$\begin{aligned} & \mathbf{E}\left[\Bigl\Vert\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X}(\cdot|\check{X}_{m})-P_{Y}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]\nonumber \\ & \le\mathbf{E}\left[(1+2^{-\iota_{X;Y}(X;Y)})^{-\mathsf{J}}\right]+\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}.\label{eq:resolve_pe1}\end{aligned}$$ As a result, for any $0<\gamma\le\log\mathsf{L}$, $$\begin{aligned} & \mathbf{E}\left[\Bigl\Vert\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X}(\cdot|\check{X}_{m})-P_{Y}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]\nonumber \\ & \le\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\mathsf{L}-\gamma\right\} +2^{-\gamma/2}\left(1+\frac{1}{2}\sqrt{\gamma}\right)+\frac{1}{2}\sqrt{\mathsf{L}^{-1}}.\label{eq:resolve_pe2}\end{aligned}$$ Hence there exists a code for channel resolvability satisfying the above bounds. Let $\mathfrak{P}=\{\bar{Y}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{Y}\times\lambda_{\mathbb{R}_{\ge0}}$. Let $M\sim\mathrm{Unif}[1:\mathsf{L}]$, $\{\check{X}_{m}\}_{m\in[1:\mathsf{L}]}\stackrel{iid}{\sim}P_{X}$ ($M\perp\!\!\!\perp\{\check{X}_{j}\}_{j}\perp\!\!\!\perp\mathfrak{P}$), and $X=\check{X}_{M}$. Let $Y=\tilde{Y}_{P_{Y|X}(\cdot|X)}$, and $\hat{Y}_{j}=\tilde{Y}_{P_{Y}}(j)$ for $j\in\mathbb{N}$. We have $$\begin{aligned} & \mathbf{E}\left[\Vert P_{Y|\{\check{X}_{m}\}_{m}}(\cdot|\{\check{X}_{m}\}_{m})-P_{Y}(\cdot)\Vert_{\mathrm{TV}}\right]\\ & \stackrel{(a)}{\le}\mathbf{E}\left[\Vert P_{Y|\{\check{X}_{m}\}_{m},\mathfrak{P}}(\cdot|\{\check{X}_{m}\}_{m},\mathfrak{P})-P_{Y|\mathfrak{P}}(\cdot|\mathfrak{P})\Vert_{\mathrm{TV}}\right]\\ & \stackrel{(b)}{=}\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in\mathbb{N}}}\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-P_{Y|\{\check{X}_{m}\}_{m},\mathfrak{P}}(y|\{\check{X}_{m}\}_{m},\mathfrak{P})\right|\right]\\ & \le\frac{1}{2}\mathbf{E}\biggl[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-P_{Y|\{\check{X}_{m}\}_{m},\mathfrak{P}}(y|\{\check{X}_{m}\}_{m},\mathfrak{P})\right|\\ & \;\;\;\;\;\;\;\;+\sum_{y\in\{\hat{Y}_{j}\}_{j\in\mathbb{N}}\backslash\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\left(P_{Y|\mathfrak{P}}(y|\mathfrak{P})+P_{Y|\{\check{X}_{m}\}_{m},\mathfrak{P}}(y|\{\check{X}_{m}\}_{m},\mathfrak{P})\right)\biggr]\\ & =\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\right|\right]\\ & \;\;\;\;\;\;\;\;+\frac{1}{2}\mathbf{E}\left[P_{Y|\mathfrak{P}}(\mathcal{Y}\backslash\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}|\mathfrak{P})+P_{Y|\{\check{X}_{m}\}_{m},\mathfrak{P}}(\mathcal{Y}\backslash\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}\,|\,\{\check{X}_{m}\}_{m},\mathfrak{P})\right]\\ & =\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\right|\right]+\mathbf{P}\left\{ Y\notin\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}\right\} ,\end{aligned}$$ where (a) is by the convexity of the total variation distance, and (b) is because $Y\in\{\hat{Y}_{j}\}_{j\in\mathbb{N}}$ almost surely (note that the summation $\sum_{y\in\{\hat{Y}_{j}\}_{j\in\mathbb{N}}}$ ignores multiplicity of elements in $\{\hat{Y}_{j}\}_{j\in\mathbb{N}}$). For the first term, note that since $Y$ is a function of $(X,\mathfrak{P})$, we have $P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\in\{0,1\}$, and hence $$\left(\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\right)\,\biggl|\,\mathfrak{P}\sim\mathrm{Bin}(\mathsf{L},\,P_{Y|\mathfrak{P}}(y|\mathfrak{P})).$$ We have $$\begin{aligned} & \frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\right|\right]\\ & =\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\mathbf{E}\left[\left.\left|P_{Y|\mathfrak{P}}(y|\mathfrak{P})-\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\right|\,\right|\,\mathfrak{P}\right]\right]\\ & \le\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\sqrt{\mathrm{Var}\left[\left.\mathsf{L}^{-1}\sum_{m=1}^{\mathsf{L}}P_{Y|X,\mathfrak{P}}(y|\check{X}_{m},\mathfrak{P})\,\right|\,\mathfrak{P}\right]}\right]\\ & \le\frac{1}{2}\mathbf{E}\left[\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\sqrt{\mathsf{L}^{-1}P_{Y|\mathfrak{P}}(y|\mathfrak{P})}\right]\\ & \le\frac{1}{2}\mathbf{E}\left[\sqrt{\mathsf{J}\sum_{y\in\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}}\mathsf{L}^{-1}P_{Y|\mathfrak{P}}(y|\mathfrak{P})}\right]\\ & \le\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}.\end{aligned}$$ For the second term, by the conditional generalized Poisson matching lemma on $(X,\,1,\,Y,\,\emptyset,\,P_{Y})$, $$\begin{aligned} & \mathbf{P}\{Y\notin\{\hat{Y}_{j}\}_{j\in[1:\mathsf{J}]}\}\\ & \le\mathbf{E}\left[\left(1-\left(1+\frac{dP_{Y|X}(\cdot|X)}{dP_{Y}}(Y)\right)^{-1}\right)^{\mathsf{J}}\right]\\ & =\mathbf{E}\left[(1-(1+2^{\iota_{X;Y}(X;Y)})^{-1})^{\mathsf{J}}\right].\end{aligned}$$ Hence, $$\begin{aligned} & \mathbf{E}\left[\Vert P_{Y|\{\check{X}_{j}\}_{j}}(\cdot|\{\check{X}_{m}\}_{m})-P_{Y}(\cdot)\Vert_{\mathrm{TV}}\right]\\ & \le\mathbf{E}\left[(1-(1+2^{\iota_{X;Y}(X;Y)})^{-1})^{\mathsf{J}}\right]+\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}\\ & =\mathbf{E}\left[(1+2^{-\iota_{X;Y}(X;Y)})^{-\mathsf{J}}\right]+\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}.\end{aligned}$$ For , substitute $\mathsf{J}=\lceil\gamma2^{-\gamma}\mathsf{L}\rceil$, $$\begin{aligned} & \mathbf{E}\left[(1-(1+2^{\iota_{X;Y}(X;Y)})^{-1})^{\mathsf{J}}\right]+\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}\\ & \stackrel{(a)}{\le}\mathbf{E}\left[(1-(1+(2\mathsf{L}2^{-\gamma})^{-1}2^{\iota_{X;Y}(X;Y)})^{-1})^{\mathsf{J}(2\mathsf{L}2^{-\gamma})^{-1}}\right]+\frac{1}{2}\sqrt{\mathsf{J}\mathsf{L}^{-1}}\\ & \le\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\mathsf{L}-\gamma\right\} +2^{-\mathsf{J}(2\mathsf{L}2^{-\gamma})^{-1}}+\frac{1}{2}\sqrt{(\gamma2^{-\gamma}\mathsf{L}+1)\mathsf{L}^{-1}}\\ & \le\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\mathsf{L}-\gamma\right\} +2^{-\gamma/2}+\frac{1}{2}\sqrt{\gamma2^{-\gamma}}+\frac{1}{2}\sqrt{\mathsf{L}^{-1}}\\ & =\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\mathsf{L}-\gamma\right\} +2^{-\gamma/2}\left(1+\frac{1}{2}\sqrt{\gamma}\right)+\frac{1}{2}\sqrt{\mathsf{L}^{-1}},\end{aligned}$$ where (a) is because $\gamma\le\log\mathsf{L}$, $2\mathsf{L}2^{-\gamma}>1$ and $(1-(1+\alpha)^{-1})^{\beta}\le1-(1+\beta^{-1}\alpha)^{-1}$ for $\alpha\ge0$, $\beta\ge1$. Compare this to Theorem 2 in [@hayashi2006resolvability] (weakened by substituting $\delta'_{p,W,C}\le C$): for any $\alpha>0$, $$\epsilon\le\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\alpha\right\} +\frac{1}{2}\sqrt{\alpha\mathsf{L}^{-1}}.$$ If we assume $1\le\alpha\le\mathsf{L}$ and substitute $\gamma=\log(\mathsf{L}/\alpha)$ in , we obtain the following slightly weaker bound (within a logarithmic gap from that in [@hayashi2006resolvability]): $$\begin{aligned} \epsilon & \le\mathbf{P}\left\{ \iota_{X;Y}(X;Y)>\log\alpha\right\} +\sqrt{\alpha\mathsf{L}^{-1}}\left(1+\frac{1}{2}\sqrt{\log(\mathsf{L}/\alpha)}\right)+\frac{1}{2}\sqrt{\mathsf{L}^{-1}}.\end{aligned}$$ Nevertheless, the bound in [@hayashi2006resolvability] does not imply , so neither bound is stronger than the other. The channel resolvability or soft covering bound in Proposition \[prop:resolve\] can be applied to prove various secrecy and coordination results, e.g. one-shot coding for wiretap channels [@wyner1975wire], one-shot channel synthesis [@cuff2013synthesis], and one-shot distributed source simulation [@wyner1975common]. Hence these results can also be proved using the Poisson matching lemma alone. In the next section, we will prove a one-shot result for wiretap channels. One-shot Coding for Wiretap Channels ==================================== The one-shot version of the wiretap channel setting [@wyner1975wire] is described as follows. Upon observing $M\sim\mathrm{Unif}[1:\mathsf{L}]$, the encoder produces $X$, which is sent through the broadcast channel $P_{Y,Z|X}$. The legitimate decoder observes $Y$ and recovers $\hat{M}$ with error probability $P_{e}=\mathbf{P}\{M\neq\hat{M}\}$. The eavesdropper observes $Z$. Secrecy is measured by the total variation distance $\epsilon:=\Vert P_{M,Z}-P_{M}\times P_{Z}\Vert_{\mathrm{TV}}$. The following bound is a direct result of the generalized Poisson matching lemma and Proposition \[prop:resolve\]. It is included for demonstration purposes. See [@hayashi2006resolvability; @yassaee2015one; @liu2017resolvability] for other one-shot bounds (that are not strictly stronger or weaker than ours). \[prop:wire\]Fix any $P_{U,X}$. For any $\nu\ge0$, $\mathsf{K},\mathsf{J}\in\mathbb{N}$, there exists a code for the wiretap channel $P_{Y,Z|X}$, with message $M\sim\mathrm{Unif}[1:\mathsf{L}]$, with average error probability $P_{e}$ and secrecy measure $\epsilon$ satisfying $$\begin{aligned} P_{e}+\nu\epsilon & \le\mathbf{E}\left[\min\{\mathsf{L}\mathsf{K}2^{-\iota_{U;Y}(U;Y)},\,1\}\right]\\ & \;\;\;+\nu\left(2\mathbf{E}\left[(1+2^{-\iota_{U;Z}(U;Z)})^{-\mathsf{J}}\right]+\sqrt{\mathsf{J}\mathsf{K}^{-1}}\right)\end{aligned}$$ if $P_{UY}\ll P_{U}\times P_{Y}$ and $P_{UZ}\ll P_{U}\times P_{Z}$. Let $\mathfrak{P}=\{(\bar{U}_{i},\bar{M}_{i}),T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{U}\times P_{M}\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $M$. Let $K\sim\mathrm{Unif}[1:\mathsf{K}]$ independent of $(M,\mathfrak{P})$. The encoder computes $U=\tilde{U}_{P_{U}\times\delta_{M}}(K)$ and generates $X|U\sim P_{X|U}$. The decoder recovers $\hat{M}=\tilde{M}_{P_{U|Y}(\cdot|Y)\times P_{M}}$. We have $(M,K,U,X,Y,Z)\sim P_{M}\times P_{K}\times P_{U,X}P_{Y,Z|X}$. By the conditional generalized Poisson matching lemma on $(M,\,K,\,(U,M),\,Y,\,P_{U|Y}\times P_{M})$ (note that $P_{U,M|M,K}=P_{U}\times\delta_{M}$), $$\begin{aligned} & \mathbf{P}\left\{ M\neq\hat{M}\right\} \\ & \le\mathbf{E}\left[\mathbf{P}\left\{ (U,M)\neq(\tilde{U},\tilde{M})_{P_{U|Y}(\cdot|Y)\times P_{M}}\,|\,M,K,U,Y\right\} \right]\\ & \le\mathbf{E}\left[\min\left\{ \mathsf{K}\frac{dP_{U}\times\delta_{M}}{dP_{U|Y}(\cdot|Y)\times P_{M}}(U,M),\,1\right\} \right]\\ & =\mathbf{E}\left[\min\{\mathsf{L}\mathsf{K}2^{-\iota_{U;Y}(U;Y)},\,1\}\right].\end{aligned}$$ For the secrecy measure, $$\begin{aligned} & \mathbf{E}\left[\Bigl\Vert P_{M,Z|\mathfrak{P}}(\cdot,\cdot|\mathfrak{P})-P_{M}(\cdot)\times P_{Z|\mathfrak{P}}(\cdot|\mathfrak{P})\Bigr\Vert_{\mathrm{TV}}\right]\\ & =\mathbf{E}\left[\Bigl\Vert P_{Z|M,\mathfrak{P}}(\cdot|M,\mathfrak{P})-P_{Z|\mathfrak{P}}(\cdot|\mathfrak{P})\Bigr\Vert_{\mathrm{TV}}\right]\\ & \le\mathbf{E}\left[\Bigl\Vert P_{Z|M,\mathfrak{P}}(\cdot|M,\mathfrak{P})-P_{Z}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]+\mathbf{E}\left[\Bigl\Vert P_{Z|\mathfrak{P}}(\cdot|\mathfrak{P})-P_{Z}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]\\ & \stackrel{(a)}{\le}2\mathbf{E}\left[\Bigl\Vert P_{Z|M,\mathfrak{P}}(\cdot|M,\mathfrak{P})-P_{Z}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]\\ & =2\mathbf{E}\left[\Bigl\Vert\mathsf{K}^{-1}\sum_{k=1}^{\mathsf{K}}P_{Z|U}(\cdot|\tilde{U}_{P_{U}\times\delta_{M}}(k))-P_{Z}(\cdot)\Bigr\Vert_{\mathrm{TV}}\right]\\ & \stackrel{(b)}{\le}2\mathbf{E}\left[(1+2^{-\iota_{U;Z}(U;Z)})^{-\mathsf{J}}\right]+\sqrt{\mathsf{J}\mathsf{K}^{-1}},\end{aligned}$$ where (a) is by the convexity of total variation distance, and (b) is by Proposition \[prop:resolve\] since $\{\tilde{U}_{P_{U}\times\delta_{m}}(k)\}_{k\in[1:\mathsf{K}]}\stackrel{iid}{\sim}P_{U}$ for any $m$. Therefore there exists a fixed set of points for $\mathfrak{P}$ satisfying the desired bound. Strong Functional Representation Lemma and Noncausal Sampling ============================================================= The generalized Poisson matching lemma can be applied to give a slight improvement on the constant in the strong functional representation lemma in [@sfrl_trans], and hence improves on the variable-length channel simulation result in [@harsha2010communication], and the result on minimax remote prediction with a communication constraint in [@li2018minimax]. It also gives an achievability bound on the moments for the noncausal sampling setting in [@liu2018rejection]. \[prop:concave\]Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ over $\mathcal{U}\times\mathbb{R}_{\ge0}$, and $P,Q$ be probability measures over $\mathcal{U}$ with $P\ll Q\ll\mu$. For any $j\in\mathbb{N}$, $g:\mathbb{R}_{\ge0}\to\mathbb{R}$ concave nondecreasing, we have $$\mathbf{E}\left[g(\Upsilon_{P\Vert Q}(j)-1)\right]\le\mathbf{E}_{U\sim P}\left[g\left(j\frac{dP}{dQ}(U)\right)\right],$$ i.e., $j(dP/dQ)(U)$ dominates $\Upsilon_{P\Vert Q}(j)-1$ in the second order. As a result, let $$\mathfrak{C}[xg'(x)](y)=\inf\left\{ \alpha y+\beta:\,xg'(x)\le\alpha x+\beta\;\forall x\ge0\right\}$$ be the upper concave envelope of $xg'(x)$, then $$\mathbf{E}\left[g(\Upsilon_{P\Vert Q}(j))\right]\le\mathbf{E}_{U\sim P}\left[g\left(j\frac{dP}{dQ}(U)\right)\right]+j^{-1}\mathfrak{C}[xg'(x)](j).$$ In particular, $$\mathbf{E}\left[\log\Upsilon_{P\Vert Q}(j)\right]\le D(P\Vert Q)+\log j+j^{-1}\log e,$$ and for $\gamma\in(0,1)$, $$\begin{aligned} \mathbf{E}\left[(\Upsilon_{P\Vert Q}(j))^{\gamma}\right] & \le j^{\gamma}\mathbf{E}_{U\sim P}\left[\left(\frac{dP}{dQ}(U)\right)^{\gamma}\right]+\gamma j^{\gamma-1}\\ & =j^{\gamma}2^{\gamma D_{\gamma+1}(P\Vert Q)}+\gamma j^{\gamma-1},\end{aligned}$$ where $D_{\gamma+1}(P\Vert Q)=\gamma^{-1}\log\mathbf{E}_{U\sim P}\left[\left((dP/dQ)(U)\right)^{\gamma}\right]$ is the Rényi divergence. For $g:\mathbb{R}_{\ge0}\to\mathbb{R}$ concave nondecreasing, we have $$\begin{aligned} & \mathbf{E}\left[g(\Upsilon_{P\Vert Q}(j)-1)\right]\\ & =\int\mathbf{E}\left[\left.g(\Upsilon_{P\Vert Q}(j)-1)\,\right|\,\tilde{U}_{P}(j)=u\right]P(du)\\ & \stackrel{(a)}{\le}\int g\left(\mathbf{E}\left[\left.\Upsilon_{P\Vert Q}(j)\,\right|\,\tilde{U}_{P}(j)=u\right]-1\right)P(du)\\ & \stackrel{(b)}{\le}\int g\left(j\frac{dP}{dQ}(u)\right)P(du),\end{aligned}$$ where (a) is by Jensen’s inequality, and (b) is by the generalized Poisson matching lemma. For any $\alpha,\beta$ such that $xg'(x)\le\alpha x+\beta$ for $x\ge0$, $$\begin{aligned} & \mathbf{E}\left[g(\Upsilon_{P\Vert Q}(j))\right]\\ & \le\int g\left(j\frac{dP}{dQ}(u)+1\right)P(du)\\ & \le\int g\left(j\frac{dP}{dQ}(u)\right)P(du)+\int g'\left(j\frac{dP}{dQ}(u)\right)P(du)\\ & =\int g\left(j\frac{dP}{dQ}(u)\right)P(du)+j^{-1}\int g'\left(j\frac{dP}{dQ}(u)\right)j\frac{dP}{dQ}(u)Q(du)\\ & \le\int g\left(j\frac{dP}{dQ}(u)\right)P(du)+j^{-1}\int\left(\alpha j\frac{dP}{dQ}(u)+\beta\right)Q(du)\\ & =\int g\left(j\frac{dP}{dQ}(u)\right)P(du)+j^{-1}(\alpha j+\beta).\end{aligned}$$ For $g(x)=\log x$, $xg'(x)=\log e$, and hence $$\mathbf{E}\left[\log\Upsilon_{P\Vert Q}(j)\right]\le D(P\Vert Q)+\log j+j^{-1}\log e.$$ For $g(x)=x^{\gamma}$, $\gamma\in(0,1)$, $xg'(x)=\gamma x^{\gamma}$ is concave, and hence $$\begin{aligned} \mathbf{E}\left[(\Upsilon_{P\Vert Q}(j))^{\gamma}\right] & \le\mathbf{E}_{U\sim P}\left[\left(j\frac{dP}{dQ}(U)\right)^{\gamma}\right]+j^{-1}\gamma j^{\gamma}\\ & =j^{\gamma}\mathbf{E}_{U\sim P}\left[\left(\frac{dP}{dQ}(U)\right)^{\gamma}\right]+\gamma j^{\gamma-1}.\end{aligned}$$ Consider the setting in the strong functional representation lemma [@sfrl_trans]: given $(X,Y)$, we want to find a random variable $Z$ independent of $X$ such that $Y$ is a function of $(X,Z)$, and $H(Y|Z)$ is minimized. Take $Z=\{\bar{Y}_{i},T_{i}\}_{i\in\mathbb{N}}$. Applying Proposition \[prop:concave\] on $P=P_{Y|X}(\cdot|X)$, $Q=P_{Y}$, we obtain $$\begin{aligned} \mathbf{E}\left[\log\Upsilon_{P_{Y|X}(\cdot|X)\Vert P_{Y}}(1)\right] & \le\mathbf{E}\left[D(P_{Y|X}(\cdot|X)\Vert P_{Y})\right]\\ & =I(X;Y).\end{aligned}$$ Using Proposition 4 in [@sfrl_trans], $$\begin{aligned} & H(Y|Z)\\ & \le H(\Upsilon_{P_{Y|X}(\cdot|X)\Vert P_{Y}}(1))\\ & \le\mathbf{E}\left[\log\Upsilon_{P_{Y|X}(\cdot|X)\Vert P_{Y}}(1)\right]+\log\left(\mathbf{E}\left[\log\Upsilon_{P_{Y|X}(\cdot|X)\Vert P_{Y}}(1)\right]+1\right)+1\\ & \le I(X;Y)+\log e+\log\left(I(X;Y)+\log e+1\right)+1\\ & \le I(X;Y)+\log\left(I(X;Y)+1\right)+\log e+1+\log\left(\log e+1\right)\\ & \le I(X;Y)+\log\left(I(X;Y)+1\right)+3.732.\end{aligned}$$ The constant $3.732$ is smaller than that in [@sfrl_trans]: $$e^{-1}\log e+2+\log\left(e^{-1}\log e+2\right)\approx3.870.$$ Conclusions and Discussion ========================== In this paper, we introduced a simple yet versatile approach to achievability proofs via the Poisson matching lemma. By reducing the uses of sub-codebooks and binning, we improved upon existing one-shot bounds on channels with state information at the encoder, lossy source coding with side information at the decoder, broadcast channels, and distributed lossy source coding. The Poisson matching lemma can replace the packing lemma, covering lemma and soft covering lemma to be the only tool needed to prove a wide range of results in network information theory. In the proofs, random variables (e.g. the channel input and message in channel coding settings, the source and description in source coding settings, the channel output in channel resolvability) are regarded as points in a Poisson process. The Poisson functional representation is applied to map the Poisson process to give the correct conditional distribution. Viewing every random variable in the operational setting as a Poisson process gives a simple, unified and systematic approach to code constructions. A possible extension is to generalize the Poisson functional representation to the multivariate case. In the proof of Marton’s inner bound for broadcast channels, we had two independent Poisson processes for $U_{1}$ and $U_{2}$ respectively. We first used the process for $U_{1}$ to obtain a list of values for $U_{1}$, then used the list to index into the process for $U_{2}$. A more symmetric approach where we select $(U_{1},U_{2})$ together (similar to the conventional mutual covering approach) using a multivariate version of the Poisson functional representation may be possible. Similarly, for distributed lossy source coding and the multiple access channel, it may be possible to decode both sources/messages simultaneously. While it can be argued that the gain we obtained in broadcast channels and distributed lossy source coding over conventional approaches comes from the asymmetry of our construction (our bounds are asymmetric unlike previous bounds), a symmetric treatment that does not result in a looser bound may be developed in the future. Acknowledgements ================ The authors acknowledge support from the NSF grants CNS-1527846, CCF-1618145, the NSF Science & Technology Center grant CCF-0939370 (Science of Information), and the William and Flora Hewlett Foundation supported Center for Long Term Cybersecurity at Berkeley. Proof of Lemmas \[lem:phidiv\] and \[lem:phidiv\_gen\]\[subsec:pf\_phidiv\] --------------------------------------------------------------------------- We first prove Lemma \[lem:phidiv\_gen\]. For notational simplicity, we use $\{X_{i}\}_{i\in\mathbb{N}}\sim\mathfrak{P}(\mu)$ to denote that $\{X_{i}\}_{i\in\mathbb{N}}$ is the set of points of a Poisson process with intensity measure $\mu$ (the ordering of the points is ignored). Let $f(u)=(dP/d\mu)(u)$, $g(u)=(dQ/d\mu)(u)$. Let $\{\bar{U}_{i},T_{i}\}_{i\in\mathbb{N}}\sim\mathfrak{P}(\mu\times\lambda_{\mathbb{R}_{\ge0}})$. Let $\{\check{U}_{k},\check{T}_{k}\}_{k\in\mathbb{N}}$ be the points $(\bar{U}_{i},T_{i})$ where $f(\bar{U}_{i})=0$. By the mapping theorem [@kingman1992poisson; @last2017lectures] on the mapping $$\psi(u,t)=\begin{cases} (1,\,u,\,t/f(u)) & \mathrm{if}\;f(u)>0\\ (0,\,u,\,t) & \mathrm{if}\;f(u)=0, \end{cases}$$ we have $\{\psi(\bar{U}_{i},T_{i})\}_{i\in\mathbb{N}}\sim\mathfrak{P}(\delta_{1}\times P\times\lambda_{\mathbb{R}_{\ge0}}+\delta_{0}\times\mu_{\{f(u)=0\}}\times\lambda_{\mathbb{R}_{\ge0}})$ (where $\mu_{\{f(u)=0\}}$ denotes $\mu$ restricted to the set $\{u:\,f(u)=0\}$), and hence $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k\in\mathbb{N}}\sim\mathfrak{P}(P\times\lambda_{\mathbb{R}_{\ge0}})$ (the points in $\{\psi(\bar{U}_{i},T_{i})\}_{i\in\mathbb{N}}$ with $f(\bar{U}_{i})>0$) is independent of $\{\check{U}_{k},\check{T}_{k}\}_{k\in\mathbb{N}}\sim\mathfrak{P}(\mu_{\{f(u)=0\}}\times\lambda_{\mathbb{R}_{\ge0}})$ (the points in $\{\psi(\bar{U}_{i},T_{i})\}_{k\in\mathbb{N}}$ with $f(\bar{U}_{i})=0$). Condition on $\tilde{U}_{P}(j)=u$ and $\tilde{T}_{P}(j)=t$ unless otherwise stated. Assume $f(u)>0$ (which happens almost surely since $\tilde{U}_{P}(j)\sim P$) and $g(u)>0$ (otherwise the inequalities in the lemmas trivially hold). Recall that $\tilde{T}_{P}(1)\le\tilde{T}_{P}(2)\le\cdots$ by definition. It is straightforward to check that $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k>j}\sim\mathfrak{P}(P\times\lambda_{[t,\infty)})$ independent of $\{\tilde{U}_{P}(k)\}_{k<j}\stackrel{iid}{\sim}P$ independent of $\{\tilde{T}_{P}(k)\}_{k<j}\sim\mathrm{Unif}(t\Delta_{*}^{j-1})$, the uniform distribution over the ordered simplex $t\Delta_{*}^{j-1}=\{s^{j-1}:\,0\le s_{1}\le\cdots\le s_{j-1}\le t\}$ (i.e., $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k<j}$ has the same distribution as $j-1$ i.i.d. points following $P\times\mathrm{Unif}[0,t]$ sorted in ascending order of the second coordinate). We have $$\begin{aligned} & \Upsilon_{P\Vert Q}(j)-1\\ & =\left|\left\{ k:\,T_{k}/g(\bar{U}_{k})<tf(u)/g(u)\right\} \right|\\ & =\left|\left\{ k:\,f(\bar{U}_{k})=0\;\mathrm{and}\;T_{k}/g(\bar{U}_{k})<tf(u)/g(u)\right\} \right|\\ & \;\;\;\;+\left|\left\{ k:\,\,f(\bar{U}_{k})>0\;\mathrm{and}\;T_{k}/g(\bar{U}_{k})<tf(u)/g(u)\right\} \right|\\ & =\left|\left\{ k:\,\check{T}_{k}/g(\check{U}_{k})<tf(u)/g(u)\right\} \right|\\ & \;\;\;\;+\left|\left\{ k:\,\tilde{T}_{P}(k)f(\tilde{U}_{P}(k))/g(\tilde{U}_{P}(k))<tf(u)/g(u)\right\} \right|\\ & =A_{0}+A_{1}+B,\end{aligned}$$ where $$\begin{aligned} A_{0} & :=\left|\left\{ k:\,\check{T}_{k}/g(\check{U}_{k})<tf(u)/g(u)\right\} \right|,\\ A_{1} & :=\left|\left\{ k>j:\,\tilde{T}_{P}(k)f(\tilde{U}_{P}(k))/g(\tilde{U}_{P}(k))<tf(u)/g(u)\right\} \right|,\\ B & :=\left|\left\{ k<j:\,\tilde{T}_{P}(k)f(\tilde{U}_{P}(k))/g(\tilde{U}_{P}(k))<tf(u)/g(u)\right\} \right|.\end{aligned}$$ Due to the aforementioned independence between $\{\check{U}_{k},\check{T}_{k}\}_{k\in\mathbb{N}}$, $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k>j}$ and $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k<j}$, we have $A_{0}{\perp\!\!\!\perp}A_{1}{\perp\!\!\!\perp}B$. For $A_{0}$, since $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k}{\perp\!\!\!\perp}\{\check{U}_{k},\check{T}_{k}\}_{k}$, conditioning on $(\tilde{U}_{P}(j),\tilde{T}_{P}(j))=(u,t)$ does not affect the distribution of $\{\check{U}_{k},\check{T}_{k}\}_{k}$, and hence $A_{0}$ follows the Poisson distribution with rate $$\begin{aligned} & (\mu_{\{f(u)=0\}}\times\lambda)\left(\left\{ (v,s):\,s/g(v)<tf(u)/g(u)\right\} \right)\\ & =\int\mathbf{1}\{f(v)=0\}\frac{tg(v)f(u)}{g(u)}\mu(dv).\end{aligned}$$ For $A_{1}$, since $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k>j}\sim\mathfrak{P}(P\times\lambda_{[t,\infty)})$, $A_{1}$ follows the Poisson distribution with rate $$\begin{aligned} & (P\times\lambda_{[t,\infty)})\left(\left\{ (v,s):\,sf(v)/g(v)<tf(u)/g(u)\right\} \right)\\ & =\int\max\left\{ \frac{tg(v)f(u)}{f(v)g(u)}-t,\,0\right\} f(v)\mu(dv)\\ & =t\int\mathbf{1}\{f(v)>0\}\max\left\{ \frac{g(v)f(u)}{g(u)}-f(v),\,0\right\} \mu(dv).\end{aligned}$$ Hence $A:=A_{0}+A_{1}$ follows the Poisson distribution with rate $$\begin{aligned} & t\int\left(\mathbf{1}\{f(v)=0\}\frac{g(v)f(u)}{g(u)}+\mathbf{1}\{f(v)>0\}\max\left\{ \frac{g(v)f(u)}{g(u)}-f(v),\,0\right\} \right)\mu(dv)\\ & =t\int\max\left\{ \frac{g(v)f(u)}{g(u)}-f(v),\,0\right\} \mu(dv)\\ & =tf(u)\int\max\left\{ \frac{g(v)}{g(u)}-\frac{f(v)}{f(u)},\,0\right\} \mu(dv)\\ & =:\,t\alpha(u).\end{aligned}$$ For $B$, since $\{\tilde{U}_{P}(k),\tilde{T}_{P}(k)\}_{k<j}$ has the same distribution as $j-1$ i.i.d. points following $P\times\mathrm{Unif}[0,t]$ sorted in ascending order of the second coordinate, $B$ follows the binomial distribution with number of trials $j-1$ and success probability $$\begin{aligned} & (P\times\mathrm{Unif}[0,t])\left(\left\{ (v,s):\,sf(v)/g(v)<tf(u)/g(u)\right\} \right)\\ & =t^{-1}\int\min\left\{ \frac{tg(v)f(u)}{f(v)g(u)},\,t\right\} f(v)\mu(dv)\\ & =f(u)\int\min\left\{ \frac{g(v)}{g(u)},\,\frac{f(v)}{f(u)}\right\} \mu(dv)\\ & =:\,\beta(u).\end{aligned}$$ Conditioned on $\tilde{U}_{P}(j)=u$ (without conditioning on $\tilde{T}_{P}(j)$), we have $\tilde{T}_{P}(j)\sim\mathrm{Erlang}(j,1)$, and $(A,B)|\{\tilde{T}_{P}(j)=t\}\sim\mathrm{Poi}(t\alpha(u))\times\mathrm{Bin}(j-1,\beta(u))$. Hence, conditioned on $\tilde{U}_{P}(j)=u$, the distribution of $\Upsilon_{P\Vert Q}(j)-1=A+B$ is $$\mathrm{NegBin}\left(j,\,1-\frac{1}{1+\alpha(u)}\right)+\mathrm{Bin}(j-1,\beta(u)),\label{eq:dist_exact}$$ i.e., the sum of a negative binomial random variable and an independent binomial random variable. The mean is $$\begin{aligned} & \mathbf{E}\left[\left.\Upsilon_{P\Vert Q}(j)\,\right|\,\tilde{U}_{P}(j)=u\right]-1\\ & =j\alpha(u)+(j-1)\beta(u)\\ & =jf(u)\int\max\left\{ \frac{g(v)}{g(u)}-\frac{f(v)}{f(u)},\,0\right\} \mu(dv)+(j-1)f(u)\int\min\left\{ \frac{g(v)}{g(u)},\,\frac{f(v)}{f(u)}\right\} \mu(dv)\\ & =jf(u)\int\frac{g(v)}{g(u)}\mu(dv)-f(u)\int\min\left\{ \frac{g(v)}{g(u)},\,\frac{f(v)}{f(u)}\right\} \mu(dv)\\ & \le j\frac{f(u)}{g(u)}\\ & =j\frac{dP}{dQ}(u).\end{aligned}$$ Also, $$\begin{aligned} & \mathbf{P}\left\{ \left.\Upsilon_{P\Vert Q}(1)>1\,\right|\,\tilde{U}_{P}(j)=u\right\} \\ & =1-\mathbf{P}\left\{ \left.A=0\;\mathrm{and}\;B=0\,\right|\,\tilde{U}_{P}(j)=u\right\} \\ & =1-\frac{(1-\beta(u))^{j-1}}{(1+\alpha(u))^{j}}\\ & \le1-\left(\frac{1-\beta(u)}{1+\alpha(u)}\right)^{j}\\ & \le1-\left(1-\min\{\alpha(u)+\beta(u),1\}\right)^{j}\\ & =1-\left(1-\min\left\{ f(u)\int\frac{g(v)}{g(u)}\mu(dv),\,1\right\} \right)^{j}\\ & =1-\left(1-\min\left\{ \frac{f(u)}{g(u)},\,1\right\} \right)^{j}\\ & =1-\left(1-\min\left\{ \frac{dP}{dQ}(u),\,1\right\} \right)^{j}.\end{aligned}$$ For $j=1$, $$\begin{aligned} & \mathbf{P}\left\{ \left.\Upsilon_{P\Vert Q}(1)>k\,\right|\,\tilde{U}_{P}(1)=u\right\} \nonumber \\ & =\left(1-(1+\alpha(u))^{-1}\right)^{k}\nonumber \\ & =\left(1-\left(1+f(u)\int\max\left\{ \frac{g(v)}{g(u)}-\frac{f(v)}{f(u)},\,0\right\} \mu(dv)\right)^{-1}\right)^{k}\label{eq:prob_exact}\\ & \le\left(1-\left(1+f(u)\int\frac{g(v)}{g(u)}\mu(dv)\right)^{-1}\right)^{k}\nonumber \\ & =\left(1-(1+f(u)/g(u))^{-1}\right)^{k}\nonumber \\ & \le\Bigl(1-\Bigl(1+\frac{dP}{dQ}(u)\Bigr)^{-1}\Bigr)^{k}\nonumber \\ & =\exp\Bigl(-k\ln\Bigl(\Bigl(\frac{dP}{dQ}(u)\Bigr)^{-1}+1\Bigr)\Bigr)\nonumber \\ & \le\exp\Bigl(-\ln\Bigl(k\Bigl(\frac{dP}{dQ}(u)\Bigr)^{-1}+1\Bigr)\Bigr)\nonumber \\ & =1-\Bigl(1+k^{-1}\frac{dP}{dQ}(u)\Bigr)^{-1}.\nonumber \end{aligned}$$ Proof of the Conditional Poisson Matching Lemma\[subsec:phidiv\_cond\] ---------------------------------------------------------------------- The conditional Poisson matching lemma is intuitively obvious. The Poisson matching lemma can be equivalently stated as: for any probability measures $\nu,\xi\ll\mu$, the following holds for $\nu$-almost all $u$: $$\mathbf{P}_{\{\bar{U}_{i},T_{i}\}_{i}\sim P_{\{\bar{U}_{i},T_{i}\}_{i}\,|\,\tilde{U}_{\nu}=u}}\{\tilde{U}_{\xi}(\{\bar{U}_{i},T_{i}\}_{i})\neq u\}\le1-\left(1+\frac{d\nu}{d\xi}(u)\right)^{-1},$$ where $P_{\{\bar{U}_{i},T_{i}\}_{i}\,|\,\tilde{U}_{\nu}=u}$ is the conditional distribution of the Poisson process given $\tilde{U}_{\nu}=u$. Intuitively, we can consider the Poisson matching lemma to be a statement with 3 parameters $\nu,\xi,u$ (ignore the almost-all condition on $u$ for the moment). Since the statement holds for (almost) any $(\nu,\xi,u)$, it also holds for any random choice of $(\nu,\xi,u)$. In particular, it holds for $(\nu,\xi,u)=(P_{U|X}(\cdot|X),\,Q_{U|Y}(\cdot|Y),\,U)$, where $(X,U,Y)\sim P_{X,U,Y}$, which gives the conditional Poisson matching lemma. Note that the probability in the conditional Poisson matching lemma is conditional on $(X,U,Y)$, where $(X,U,Y)\leftrightarrow(\nu,\xi,u)\leftrightarrow\{\bar{U}_{i},T_{i}\}_{i}$, and hence conditioning on $(X,U,Y)$ has the same effect on $\{\bar{U}_{i},T_{i}\}_{i}$ as conditioning on the parameters $(\nu,\xi,u)$. We now prove the conditional Poisson matching lemma rigorously. Let $(\Omega,\mathcal{F},P_{\{\bar{U}_{i},T_{i}\}_{i}})$ be the probability space for $\{\bar{U}_{i},T_{i}\}_{i}$, the points of a Poisson process with intensity measure $\mu\times\lambda_{\mathbb{R}_{\ge0}}$ on $\mathcal{U}\times\mathbb{R}_{\ge0}$ (let $\mathcal{E}$ be the Borel $\sigma$-algebra of $\mathcal{U}$). The Poisson matching lemma can be equivalently stated as: for any probability measures $\nu,\xi\ll\mu$, and $\kappa:\mathcal{U}\times\mathcal{F}\to[0,1]$ a regular conditional probability distribution (RCPD) of $\{\bar{U}_{i},T_{i}\}_{i}$ conditioned on $\tilde{U}_{\nu}(\{\bar{U}_{i},T_{i}\}_{i})$ (i.e., $\kappa$ is a probability kernel, and $P_{\{\bar{U}_{i},T_{i}\}_{i}}(A\cap\tilde{U}_{\nu}^{-1}(B))=\int_{B}\kappa(u,A)\nu(du)$ for any $A\in\mathcal{F}$, $B\in\mathcal{E}$ , where $\tilde{U}_{\nu}^{-1}(B)$ denotes the preimage of $B$ under $\tilde{U}_{\nu}:\Omega\to\mathcal{U}$, note that $\tilde{U}_{\nu}(\{\bar{U}_{i},T_{i}\}_{i})\sim\nu$), then we have $$\int\mathbf{1}\{\tilde{U}_{\xi}(\{\bar{u}_{i},t_{i}\}_{i})\neq u\}\kappa(u,\,d\{\bar{u}_{i},t_{i}\}_{i})\le1-\left(1+\frac{d\nu}{d\xi}(u)\right)^{-1}\label{eq:phidiv_int}$$ for $\nu$-almost all $u$. Consider the conditional Poisson matching lemma. We have the following for $P_{X,U,Y}$-almost all $(x,u,y)$: $$\begin{aligned} & \mathbf{P}\left\{ \left.\tilde{U}_{Q_{U|Y}(\cdot|Y)}\neq U\,\right|\,X=x,U=u,Y=y\right\} \\ & =\int\mathbf{1}\{\tilde{U}_{Q_{U|Y}(\cdot|y)}(\{\bar{u}_{i},t_{i}\}_{i})\neq u\}P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U,Y}(d\{\bar{u}_{i},t_{i}\}_{i}|x,u,y)\\ & \stackrel{(a)}{=}\int\mathbf{1}\{\tilde{U}_{Q_{U|Y}(\cdot|y)}(\{\bar{u}_{i},t_{i}\}_{i})\neq u\}P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(d\{\bar{u}_{i},t_{i}\}_{i}|x,u)\\ & \stackrel{(b)}{\le}1-\left(1+\frac{dP_{U|X}(\cdot|x)}{dQ_{U|Y}(\cdot|y)}(u)\right)^{-1},\end{aligned}$$ where (a) holds for $P_{X,U,Y}$-almost all $(x,u,y)$ due to $Y\leftrightarrow(X,U)\leftrightarrow\{\bar{U}_{i},T_{i}\}_{i}$, and (b) is by with $(\nu,\xi,\kappa)\leftarrow(P_{U|X}(\cdot|x),\,Q_{U|Y}(\cdot|y),\,P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(\cdot|x,\cdot))$, which holds for $P_{U|X}(\cdot|x)$-almost all $u$, and hence holds for $P_{X,U,Y}$-almost all $(x,u,y)$. We now check that $P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(\cdot|x,\cdot)$ satisfies the RCPD condition for $P_{X}$-almost all $x$. Since $X{\perp\!\!\!\perp}\{\bar{U}_{i},T_{i}\}_{i}$, we have $P_{\{\bar{U}_{i},T_{i}\}_{i}}(\cdot)=P_{\{\bar{U}_{i},T_{i}\}_{i}|X}(\cdot|x)$ for $P_{X}$-almost all $x$. Since $U=\tilde{U}_{P_{U|X}(\cdot|X)}(\{\bar{U}_{i},T_{i}\}_{i})$, we have $P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(\{u\})\allowbreak\,|\,x,u)=1$ for $P_{X,U}$-almost all $(x,u)$. Hence the following conditions are satisfied for $P_{X}$-almost all $x$: $$P_{\{\bar{U}_{i},T_{i}\}_{i}}(\cdot)=P_{\{\bar{U}_{i},T_{i}\}_{i}|X}(\cdot|x),\label{eq:cond_phidiv_xindep}$$ $$P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(\{u\})|x,u)=1\;\text{for}\,P_{U|X}(\cdot|x)\text{-almost all}\,u.\label{eq:cond_phidiv_u}$$ For any $x$ satisfying and , we have the following: for all $A\in\mathcal{F}$, $B\in\mathcal{E}$, $$\begin{aligned} & P_{\{\bar{U}_{i},T_{i}\}_{i}}(A\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(B))\\ & \stackrel{(a)}{=}P_{\{\bar{U}_{i},T_{i}\}_{i}|X}(A\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(B)|x)\\ & =\int P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(A\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(B)\,|\,x,u)P_{U|X}(du|x)\\ & \stackrel{(b)}{=}\int P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(A\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(B)\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(\{u\})\,|\,x,u)P_{U|X}(du|x)\\ & =\int\mathbf{1}\{u\in B\}P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(A\cap\tilde{U}_{P_{U|X}(\cdot|x)}^{-1}(\{u\})\,|\,x,u)P_{U|X}(du|x)\\ & \stackrel{(c)}{=}\int_{B}P_{\{\bar{U}_{i},T_{i}\}_{i}|X,U}(A|x,u)P_{U|X}(du|x),\end{aligned}$$ where (a) is by , and (b), (c) are by . Proof of Theorem \[thm:channel2\]\[subsec:channel2\] ---------------------------------------------------- Let $\{\bar{X}_{i},T_{i}\}_{i\in\mathbb{N}}$ be the points of a Poisson process with intensity measure $P_{X}\times\lambda_{\mathbb{R}_{\ge0}}$ independent of $M$. The encoding function is $m\mapsto\tilde{X}_{P_{X}}(m)$ (i.e., $X=\tilde{X}_{P_{X}}(M)$), and the decoding function is $y\mapsto\Upsilon_{P_{X|Y}(\cdot|y)\Vert P_{X}}(1)$. We have $(M,X,Y)\sim P_{M}\times P_{X}P_{Y|X}$, $$\begin{aligned} & \mathbf{P}\{M\neq\Upsilon_{P_{X|Y}(\cdot|Y)\Vert P_{X}}(1)\}\\ & \stackrel{(a)}{=}\mathbf{P}\{\Upsilon_{P_{X}\Vert P_{X|Y}(\cdot|Y)}(M)>1\}\\ & =\mathbf{E}\left[\mathbf{P}\left\{ \left.\Upsilon_{P_{X}\Vert P_{X|Y}(\cdot|Y)}(M)>1\,\right|\,M,X,Y\right\} \right]\\ & \stackrel{(b)}{\le}\mathbf{E}\left[1-\left(1-\min\left\{ \frac{dP_{X}}{dP_{X|Y}(\cdot|Y)}(X),\,1\right\} \right)^{M}\right]\\ & =\mathbf{E}\left[1-\left(1-\min\left\{ 2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right)^{M}\right]\\ & \stackrel{(c)}{\le}\mathbf{E}\left[1-\left(1-\min\left\{ 2^{-\iota_{X;Y}(X;Y)},\,1\right\} \right)^{(\mathsf{L}+1)/2}\right],\end{aligned}$$ where (a) is by the definition of $\Upsilon$, (b) is by the conditional generalized Poisson matching lemma on $(\emptyset,M,X,Y,P_{X|Y})$, and (c) is by $M{\perp\!\!\!\perp}(X,Y)$ and Jensen’s inequality. Therefore there exists a fixed $\{\bar{x}_{i},t_{i}\}_{i\in\mathbb{N}}$ attaining the desired bound. $\blacksquare$ A noteworthy property of this construction is that both the encoder and the decoder do not require knowledge of $\mathsf{L}$. The code can transmit any integer $m\in\mathbb{N}$ with error probability $\mathbf{E}\left[1-(1-\min\{2^{-\iota_{X;Y}(X;Y)},\,1\})^{m}\right]$, assuming unlimited common randomness $\{\bar{X}_{i},T_{i}\}_{i\in\mathbb{N}}$ between the encoder and the decoder. Dispersion of Joint Source-Channel Coding\[subsec:second\_jscc\] ---------------------------------------------------------------- We show a second order result for joint source-channel coding using Theorem \[thm:jscc\] that coincides with the optimal dispersion in [@kostina2013joint]. Consider an i.i.d. source sequence $W^{k}$ of length $k$, separable distortion measure $\mathsf{d}(w^{k},\hat{z}^{k})=\frac{1}{k}\sum_{i=1}^{k}\mathsf{d}(w_{i},\hat{z}_{i})$, and $n$ uses of the memoryless channel $P_{Y|X}$. Let $P_{Z|W}$ attain the infimum of the rate-distortion function $$R(\mathsf{D}):=\inf_{P_{Z|W}:\,\mathbf{E}[\mathsf{d}(W,Z)]\le\mathsf{D}}I(W;Z).$$ The $\mathsf{D}$-tilted information [@kostina2012lossy] is defined as $$\jmath_{W}(w,\mathsf{D}):=-\log\mathbf{E}\left[2^{\nu^{*}(\mathsf{D}-\mathsf{d}(w,Z))}\right],$$ where $Z\sim P_{Z}$ (the unconditional $Z$-marginal of $P_{W}P_{Z|W}$), and $\nu^{*}=-R'(\mathsf{D})$ (the derivative exists if the infimum in $R(\mathsf{D})$ is achieved by a unique $P_{Z|W}$ [@kostina2012lossy]). We invoke a lemma in [@kostina2012lossy]: \[[@kostina2012lossy], Lemma 2\]\[lem:kostina\_lemma\]If the following conditions hold: - $\inf\{\tilde{\mathsf{D}}\ge0:\,R(\tilde{\mathsf{D}})<\infty\}<\mathsf{D}<\inf_{z\in\mathcal{Z}}\mathbf{E}[\mathsf{d}(W,z)]$, - the infimum in $R(\mathsf{D})$ is achieved by a unique $P_{Z|W}$, - there exists a finite set $\tilde{\mathcal{Z}}\subseteq\mathcal{Z}$ such that $\mathbf{E}[\min_{z\in\tilde{\mathcal{Z}}}\mathsf{d}(W,z)]<\infty$, and - $\mathbf{E}_{P_{W}\times P_{Z}}[(\mathsf{d}(W,Z))^{9}]<\infty$ (computed assuming $W,Z$ independent), then there exist constants $\alpha,\beta,\gamma,k_{0}>0$ such that for $k\ge k_{0}$, $$\mathbf{P}\left\{ -\log P_{Z^{k}}(\mathcal{B}_{\mathsf{D}}(W^{k}))\le\sum_{i=1}^{k}\jmath_{W}(W_{i},\mathsf{D})+\alpha\log k+\beta\right\} \ge1-\frac{\gamma}{\sqrt{k}},$$ where $W^{k}\stackrel{iid}{\sim}P_{W}$, and $P_{Z^{k}}=P_{Z}^{\otimes k}$. We now show a second order result. Fix $P_{X}$, $0<\epsilon<1$, $n,k\in\mathbb{N}$. We have $P_{e}=\mathbf{P}\{\mathsf{d}(W^{k},\hat{Z}^{k})>\mathsf{D}\}\le\epsilon$ if the conditions in Lemma \[lem:kostina\_lemma\] are satisfied, $k\ge k_{0}$, and $$nC-kR(\mathsf{D})\ge\sqrt{nV+k\mathcal{V}(\mathsf{D})}\mathcal{Q}^{-1}\left(\epsilon-\frac{\eta}{\sqrt{\min\{n,k\}}}\right)+\alpha\log k+\frac{1}{2}\log n+\beta,$$ where $C:=I(X;Y)$, $V:=\mathrm{Var}[\iota_{X;Y}(X;Y)]$, $\mathcal{V}(\mathsf{D}):=\mathrm{Var}[\jmath_{W}(W,\mathsf{D})]$, and $\eta>0$ is a constant that depends on $P_{X,Y}$ and the distribution of $\jmath_{W}(W,\mathsf{D})$. We have $$\begin{aligned} P_{e} & =\mathbf{P}\{\mathsf{d}(W^{k},\hat{Z}^{k})>\mathsf{D}\}\\ & \stackrel{(a)}{\le}\mathbf{P}\left\{ -\log P_{Z^{k}}(\mathcal{B}_{\mathsf{D}}(W^{k}))>\sum_{i=1}^{k}\jmath_{W}(W_{i},\mathsf{D})+\alpha\log k+\beta\right\} \\ & \;\;\;+\mathbf{E}\left[\left(1+2^{-\sum_{i=1}^{k}\jmath_{W}(W_{i},\mathsf{D})-\alpha\log k-\beta}2^{\iota_{X^{n};Y^{n}}(X^{n};Y^{n})}\right)^{-1}\right]\\ & \stackrel{(b)}{\le}\frac{\gamma}{\sqrt{k}}+\frac{1}{\sqrt{n}}+\mathbf{P}\left\{ 2^{\sum_{i=1}^{k}\jmath_{W}(W_{i},\mathsf{D})-\iota_{X^{n};Y^{n}}(X^{n};Y^{n})+\alpha\log k+\beta}>\frac{1}{\sqrt{n}}\right\} \\ & =\frac{\gamma}{\sqrt{k}}+\frac{1}{\sqrt{n}}+\mathbf{P}\left\{ \sum_{i=1}^{n}(\iota_{X;Y}(X_{i};Y_{i})-C)-\sum_{i=1}^{k}(\jmath_{W}(W_{i},\mathsf{D})-R(\mathsf{D}))<-nC+kR(\mathsf{D})+\alpha\log k+\frac{1}{2}\log n+\beta\right\} \\ & \le\frac{\gamma}{\sqrt{k}}+\frac{1}{\sqrt{n}}+\mathbf{P}\left\{ \sum_{i=1}^{n}(\iota_{X;Y}(X_{i};Y_{i})-C)-\sum_{i=1}^{k}(\jmath_{W}(W_{i},\mathsf{D})-R(\mathsf{D}))<-\sqrt{nV+k\mathcal{V}(\mathsf{D})}\mathcal{Q}^{-1}\left(\epsilon-\frac{\eta}{\sqrt{\min\{n,k\}}}\right)\right\} \\ & \stackrel{(c)}{\le}\frac{\gamma}{\sqrt{k}}+\frac{1}{\sqrt{n}}+\epsilon-\frac{\eta}{\sqrt{\min\{n,k\}}}+\frac{\eta-\gamma-1}{\sqrt{\min\{n,k\}}}\\ & \le\epsilon\end{aligned}$$ where (a) is by Theorem \[thm:jscc\], (b) is by Lemma \[lem:kostina\_lemma\], and (c) is by the Berry-Esseen theorem [@berry1941accuracy; @esseen1942liapunov; @feller1971introduction] if we let $\eta-\gamma-1$ be a constant given by the Berry-Esseen theorem. This coincides with the optimal dispersion in [@kostina2013joint]. Although this is not a self-contained proof (it requires the lemma in [@kostina2012lossy] for the dispersion of lossy source coding), it shows how we can obtain the achievability of the dispersion in joint source-channel coding from a result on the dispersion of lossy source coding with little additional effort, using the Poisson matching lemma. This proof is considerably simpler than that in [@kostina2013joint]. Properties of $\phi(t)$ ----------------------- Let $\phi:\mathbb{R}_{>0}\to\mathbb{R}_{>0}$, $\phi(t)=ct^{-1}(\log(t+2))^{-2}$, where $c>0$ such that $\sum_{j=1}^{\infty}\phi(j)=1$. Note that $(\phi(t))^{-1}$ is convex. It can be checked numerically that $1\le c\le2$. We prove a useful inequality about $\phi(t)$. \[prop:phi\_ineq\]For any $s>0$, $t\ge1$, we have $$\min\{s(\phi(t))^{-1},\,1\}\le\min\left\{ st\left(\log(s^{-1}+1)+1\right)^{2},\,1\right\} .$$ Moreover, if $st\le2^{-\alpha}$, $t-1\le2^{\beta}$, and $\tilde{\alpha}\ge\max\{\alpha,0\}$, then $$\min\left\{ st\left(\log(s^{-1}+1)+1\right)^{2},\,1\right\} \le2^{-\alpha}\left(2(\tilde{\alpha}+\beta)^{2}+2\tilde{\alpha}^{2}+14\right).$$ Write $\phi^{-1}(t)$ for the inverse function of $\phi$. Since $$\phi\left(\frac{c}{t\left(\log\left(c/t+2\right)\right)^{2}}\right)=\frac{t\left(\log\left(c/t+2\right)\right)^{2}}{\left(\log\left(\frac{c}{t\left(\log\left(c/t+2\right)\right)^{2}}+2\right)\right)^{2}}\ge t,$$ we have $$\phi^{-1}(t)\ge\frac{c}{t\left(\log\left(c/t+2\right)\right)^{2}}.$$ By the convexity of $(\phi(t))^{-1}$ , $$\begin{aligned} \min\left\{ \frac{s}{\phi(t)},\,1\right\} & \le\min\left\{ \frac{t}{\phi^{-1}(s)},\,1\right\} \\ & \le\min\left\{ tc^{-1}s\left(\log\left(c/s+2\right)\right)^{2},\,1\right\} \\ & \le\min\left\{ st\left(\log\left(2/s+2\right)\right)^{2},\,1\right\} \\ & =\min\left\{ st\left(\log(s^{-1}+1)+1\right)^{2},\,1\right\} .\end{aligned}$$ If $st\le2^{-\alpha}$, $t-1\le2^{\beta}$, and $\tilde{\alpha}\ge\max\{\alpha,0\}$, $$\begin{aligned} & \min\left\{ st\left(\log(s^{-1}+1)+1\right)^{2},\,1\right\} \\ & =\min\left\{ st\left(\log(t/(st)+1)+1\right)^{2},\,1\right\} \\ & \le2^{-\alpha}\left(\log((2^{\beta}+1)2^{\alpha}+1)+1\right)^{2}\\ & \le2^{-\alpha}\left(\log(2^{\tilde{\alpha}+\beta}+2^{\tilde{\alpha}}+1)+1\right)^{2}\\ & \le2^{-\alpha}\left(\max\{\tilde{\alpha}+\beta,\,\tilde{\alpha}\}+\log3+1\right)^{2}\\ & \le2^{-\alpha}\left(2(\tilde{\alpha}+\beta)^{2}+2\tilde{\alpha}^{2}+14\right),\end{aligned}$$ where the last inequality follows from considering whether $\beta$ is positive or negative, and the inequality $(x+y)^{2}\le2x^{2}+2y^{2}$. Proof of Theorem \[thm:bc\_cm\] for Broadcast Channel with Common Message\[subsec:pf\_bc\_cm\] ---------------------------------------------------------------------------------------------- The parameters $\mathsf{K}_{1},\mathsf{K}_{2}$ correspond to rate splitting. We can split $M_{1}\in[1:\mathsf{L}_{1}]$ into $M_{10}\in[1:\mathsf{K}_{1}]$ and $M_{11}\in[1:\lceil\mathsf{L}_{1}\mathsf{K}_{1}^{-1}\rceil]$, and treat $M_{10}$ as part of $M_{0}$ to be decoded by both decoders. Although $M_{10}$ and $M_{11}$ may not be uniformly distributed, we can apply a random cyclic shift to $M_{1}$ such that $M_{1}\sim\mathrm{Unif}[1:\mathsf{K}_{1}\lceil\mathsf{L}_{1}\mathsf{K}_{1}^{-1}\rceil]$ (and hence $M_{10},M_{11}$ are also uniform), and condition on a fixed shift at the end. Also $M_{2}$ can be split similarly. Therefore we assume $\mathsf{K}_{1}=\mathsf{K}_{2}=1$ without loss of generality. Let $\mathfrak{P}_{0}=\{(\bar{U}_{0,i},\bar{M}_{00,i}),T_{0,i}\}_{i\in\mathbb{N}}$, $\mathfrak{P}_{1}=\{(\bar{U}_{1,i},\bar{M}_{01,i},\bar{M}_{1,i}),T_{1,i}\}_{i\in\mathbb{N}}$, $\mathfrak{P}_{2}=\{(\bar{U}_{2,i},\bar{M}_{02,i},\bar{M}_{2,i}),T_{2,i}\}_{i\in\mathbb{N}}$ be three independent Poisson processes with intensity measures $P_{U_{0}}\times P_{M_{0}}\times\lambda_{\mathbb{R}_{\ge0}}$, $P_{U_{1}}\times P_{M_{0}}\times P_{M_{1}}\times\lambda_{\mathbb{R}_{\ge0}}$ and $P_{U_{2}}\times P_{M_{0}}\times P_{M_{2}}\times\lambda_{\mathbb{R}_{\ge0}}$ respectively, independent of $M_{0},M_{1},M_{2}$. The encoder would generate $X$ such that $$\begin{aligned} & (M_{0},M_{1},M_{2},U_{0},J,\{\check{U}_{1j}\}_{j\in[1:\mathsf{J}]},U_{1},U_{2},X)\nonumber \\ & \sim P_{M_{0}}\times P_{M_{1}}\times P_{M_{2}}\times P_{U_{0}}P_{J}P_{U_{1}|U_{0}}^{\otimes\mathsf{J}}\delta_{\check{U}_{1J}}P_{U_{2}|U_{0},U_{1}}\delta_{x(U_{0},U_{1},U_{2})},\label{eq:bc2_dist}\end{aligned}$$ where $P_{J}=\mathrm{Unif}[1:\mathsf{J}]$, and $\{\check{U}_{1j}\}_{j\in[1:\mathsf{J}]}\in\mathcal{U}_{1}^{\mathsf{J}}$ is an intermediate list (which can be regarded as a sub-codebook). The term $P_{U_{1}|U_{0}}^{\otimes\mathsf{J}}\delta_{\check{U}_{1J}}$ in means that $\{\check{U}_{1j}\}_{j}$ are conditionally i.i.d. $P_{U_{1}|U_{0}}$ given $U_{0}$, and $U_{1}=\check{U}_{1J}$. To accomplish this, the encoder computes $U_{0}=(\tilde{U}_{0})_{P_{U_{0}}\times\delta_{M_{0}}}$, $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}|U_{0}}(\cdot|U_{0})\times\delta_{M_{0}}\times\delta_{M_{1}}}(j)$ for $j=1,\ldots,\mathsf{J}$,\ $U_{2}=(\tilde{U}_{2})_{\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}P_{U_{2}|U_{0},U_{1}}(\cdot|U_{0},\check{U}_{1j})\times\delta_{M_{0}}\times\delta_{M_{2}}}$ (which Poisson process we are referring to can be deduced from whether we are discussing $U_{0}$, $U_{1}$ or $U_{2}$), $(J,U_{1})|(U_{0},\{\check{U}_{1j}\}_{j},U_{2})\sim P_{J,U_{1}|U_{0},\{\check{U}_{1j}\}_{j},U_{2}}$ (where $P_{J,U_{1}|U_{0},\{\check{U}_{1j}\}_{j},U_{2}}$ is derived from ), and outputs $X=x(U_{0},U_{1},U_{2})$. It can be verified that is satisfied. For the decoding function at the decoder $a\in[1:2]$, let $(\check{U}_{0aj},\check{M}_{0aj})=(\tilde{U}_{0},\tilde{M}_{00})_{P_{U_{0}|Y_{a}}(\cdot|Y_{a})\times P_{M_{0}}}(j)$ for $j\in\mathbb{N}$, $(\hat{U}_{a},\hat{M}_{0a},\hat{M}_{a})=(\tilde{U}_{a},\tilde{M}_{0a},\tilde{M}_{a})_{\sum_{j=1}^{\infty}\phi(j)(P_{U_{a}|U_{0},Y_{a}}(\cdot|\check{U}_{0aj},Y_{a})\times\delta_{\check{M}_{0aj}})\times P_{M_{a}}}$ where $\phi(j)\propto j^{-1}(\log(j+2))^{-2}$ with $\sum_{j=1}^{\infty}\phi(j)=1$. Let $K_{a}=\Upsilon_{P_{U_{0}}\times\delta_{M_{0}}\Vert P_{U_{0}|Y_{a}}(\cdot|Y_{a})\times P_{M_{0}}}(1)$ (using the Poisson process $\mathfrak{P}_{0}$). By the conditional generalized Poisson matching lemma on $(M_{0},\,1,\,(U_{0},M_{0}),\,Y_{a},\,P_{U_{0}|Y_{a}}\times P_{M_{0}})$, almost surely, $$\begin{aligned} \mathbf{E}\left[\left.K_{a}\,\right|\,U_{0},Y_{a},M_{0}\right] & \le\frac{dP_{U_{0}}\times\delta_{M_{0}}}{dP_{U_{0}|Y_{a}}(\cdot|Y_{a})\times P_{M_{0}}}(U_{0},M_{0})+1\nonumber \\ & =\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{a}}(U_{0};Y_{a})}+1.\label{eq:bc2_Ka}\end{aligned}$$ By , $U_{0}=(\tilde{U}_{0})_{P_{U_{0}}\times\delta_{M_{0}}}$, $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}|U_{0}}(\cdot|U_{0})\times\delta_{M_{0}}\times\delta_{M_{1}}}(j)$, and $(\check{U}_{01j},\check{M}_{01j})=(\tilde{U}_{0},\tilde{M}_{00})_{P_{U_{0}|Y_{1}}(\cdot|Y_{1})\times P_{M_{0}}}(j)$, we have $(M_{0},M_{1},U_{0},J)\perp\!\!\!\perp\mathfrak{P}_{1}$ and $(\{(\check{U}_{01j},\check{M}_{01j})\}_{j},Y_{1})\leftrightarrow(M_{0},M_{1},U_{0},J,U_{1})\leftrightarrow\mathfrak{P}_{1}$ (see Figure middle). Hence by the conditional generalized Poisson matching lemma on $((M_{0},M_{1},U_{0}),\,J,\,(U_{1},M_{0},M_{1}),\,(\{(\check{U}_{01j},\check{M}_{01j})\}_{j},Y_{1}),\,\allowbreak\sum_{j=1}^{\infty}\phi(j)\allowbreak(P_{U_{1}|U_{0},Y_{1}}(\cdot|\check{U}_{01j},Y_{1})\times\delta_{\check{M}_{01j}})\times P_{M_{1}})$, almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{(\tilde{U}_{1},\tilde{M}_{01},\tilde{M}_{1})_{\sum_{j=1}^{\infty}\phi(j)(P_{U_{1}|U_{0},Y_{1}}(\cdot|\check{U}_{01j},Y_{1})\times\delta_{\check{M}_{01j}})\times P_{M_{1}}}\neq(U_{1},M_{0},M_{1})\,\biggl|\,U_{0},U_{1},U_{2},J,Y_{1},Y_{2},M_{0},M_{1}\biggr\}\\ & \stackrel{(a)}{=}\mathbf{P}\biggl\{(\tilde{U}_{1},\tilde{M}_{01},\tilde{M}_{1})_{\sum_{j=1}^{\infty}\phi(j)(P_{U_{1}|U_{0},Y_{1}}(\cdot|\check{U}_{01j},Y_{1})\times\delta_{\check{M}_{01j}})\times P_{M_{1}}}\neq(U_{1},M_{0},M_{1})\,\biggl|\,U_{0},U_{1},J,Y_{1},M_{0},M_{1}\biggr\}\\ & \stackrel{(b)}{\le}\mathbf{E}\left[\left.\min\left\{ \mathsf{J}\frac{dP_{U_{1}|U_{0}}(\cdot|U_{0})\times\delta_{M_{0}}\times\delta_{M_{1}}}{d(\sum_{j=1}^{\infty}\phi(j)(P_{U_{1}|U_{0},Y_{1}}(\cdot|\check{U}_{01j},Y_{1})\times\delta_{\check{M}_{01j}}))\times P_{M_{1}}}(U_{1},M_{0},M_{1}),\,1\right\} \,\right|U_{0},U_{1},J,Y_{1},M_{0},M_{1}\right]\\ & \le\mathbf{E}\left[\left.\min\left\{ \frac{\mathsf{L}_{1}\mathsf{J}}{\phi(K_{1})}\frac{dP_{U_{1}|U_{0}}(\cdot|U_{0})\times\delta_{M_{0}}}{dP_{U_{1}|U_{0},Y_{1}}(\cdot|U_{0},Y_{1})\times\delta_{M_{0}}}(U_{1},M_{0}),\,1\right\} \,\right|U_{0},U_{1},J,Y_{1},M_{0},M_{1}\right]\\ & =\mathbf{E}\left[\left.\min\left\{ \frac{\mathsf{L}_{1}\mathsf{J}}{\phi(K_{1})}2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})},\,1\right\} \,\right|U_{0},U_{1},J,Y_{1},M_{0},M_{1}\right]\\ & \stackrel{(c)}{\le}\mathbf{E}\left[\left.K_{1}\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\left(\log(\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1\right)^{2}\,\right|U_{0},U_{1},J,Y_{1},M_{0},M_{1}\right]\\ & \stackrel{(d)}{\le}(\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{1}}(U_{0};Y_{1})}+1)\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\left(\log(\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1\right)^{2},\end{aligned}$$ where (a) is due to $(U_{2},Y_{2})\leftrightarrow(M_{0},M_{1},U_{0},J,U_{1},Y_{1})\leftrightarrow\mathfrak{P}_{1}$ (see Figure \[fig:bc2\_bayes\] middle), (b) is due to the aforementioned application of the conditional generalized Poisson matching lemma, (c) is by Proposition \[prop:phi\_ineq\], and (d) is due to and $K_{1}\leftrightarrow(U_{0},Y_{1},M_{0})\leftrightarrow(J,U_{1},M_{1})$ (see Figure \[fig:bc2\_bayes\] middle). Also, since $(M_{0},M_{2},U_{0},\{\check{U}_{1j}\}_{j})\perp\!\!\!\perp\mathfrak{P}_{2}$ and $(\{(\check{U}_{02j},\check{M}_{02j})\}_{j},Y_{2})\leftrightarrow(M_{0},M_{2},U_{0},\{\check{U}_{1j}\}_{j},U_{2})\leftrightarrow\mathfrak{P}_{2}$ (see Figure \[fig:bc2\_bayes\] right), by the conditional Poisson matching lemma on $((M_{0},M_{2},U_{0},\{\check{U}_{1j}\}_{j}),\,(U_{2},M_{0},M_{2}),\,(\{(\check{U}_{02j},\check{M}_{02j})\}_{j},Y_{2}),\,\allowbreak\sum_{j=1}^{\infty}\phi(j)\allowbreak(P_{U_{2}|U_{0},Y_{2}}\allowbreak(\cdot|\check{U}_{02j},\allowbreak Y_{2})\times\delta_{\check{M}_{02j}})\times P_{M_{2}})$, almost surely, $$\begin{aligned} & \mathbf{P}\biggl\{(\tilde{U}_{2},\tilde{M}_{02},\tilde{M}_{2})_{\sum_{j=1}^{\infty}\phi(j)(P_{U_{2}|U_{0},Y_{2}}(\cdot|\check{U}_{02j},Y_{2})\times\delta_{\check{M}_{02j}})\times P_{M_{2}}}\neq(U_{2},M_{0},M_{2})\,\biggl|\,U_{0},U_{1},U_{2},Y_{1},Y_{2},M_{0},M_{2}\biggr\}\\ & \stackrel{(a)}{=}\mathbf{P}\biggl\{(\tilde{U}_{2},\tilde{M}_{02},\tilde{M}_{2})_{\sum_{j=1}^{\infty}\phi(j)(P_{U_{2}|U_{0},Y_{2}}(\cdot|\check{U}_{02j},Y_{2})\times\delta_{\check{M}_{02j}})\times P_{M_{2}}}\neq(U_{2},M_{0},M_{2})\,\biggl|\,U_{0},U_{2},Y_{2},M_{0},M_{2}\biggr\}\\ & \stackrel{(b)}{\le}\mathbf{E}\Biggl[\mathbf{E}\Biggl[\min\Biggl\{\frac{d(\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}P_{U_{2}|U_{0},U_{1}}(\cdot|U_{0},\check{U}_{1j}))\times\delta_{M_{0}}\times\delta_{M_{2}}}{d(\sum_{j=1}^{\infty}\phi(j)(P_{U_{2}|U_{0},Y_{2}}(\cdot|\check{U}_{02j},Y_{2})\times\delta_{\check{M}_{02j}}))\times P_{M_{2}}}(U_{2},M_{0},M_{2}),\,1\Biggr\}\\ & \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\,\Biggl|\,\{\check{U}_{1j}\}_{j},U_{0},U_{2},Y_{2},M_{0},M_{2}\Biggr]\,\Biggl|\,U_{0},U_{2},Y_{2},M_{0},M_{2}\Biggr]\\ & \le\mathbf{E}\left[\left.\min\left\{ \frac{\mathsf{L}_{2}}{\phi(K_{2})}\frac{d(\mathsf{J}^{-1}\sum_{j=1}^{\mathsf{J}}P_{U_{2}|U_{0},U_{1}}(\cdot|U_{0},\check{U}_{1j}))\times\delta_{M_{0}}}{dP_{U_{2}|U_{0},Y_{2}}(\cdot|U_{0},Y_{2})\times\delta_{M_{0}}}(U_{2},M_{0}),\,1\right\} \,\right|U_{0},U_{2},Y_{2},M_{0},M_{2}\right]\\ & =\mathbf{E}\left[\left.\min\left\{ \frac{\mathsf{L}_{2}\mathsf{J}^{-1}}{\phi(K_{2})}\sum_{j=1}^{\mathsf{J}}2^{\iota_{U_{1};U_{2}|U_{0}}(\check{U}_{1j};U_{2}|U_{0})-\iota_{U_{2};Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})},\,1\right\} \,\right|U_{0},U_{2},Y_{2},M_{0},M_{2}\right]\\ & \stackrel{(c)}{\le}\mathbf{E}\left[\left.\min\left\{ \frac{\mathsf{L}_{2}\mathsf{J}^{-1}}{\phi(K_{2})}2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1),\,1\right\} \,\right|U_{0},U_{2},Y_{2},M_{0},M_{2}\right]\\ & \stackrel{(d)}{\le}\mathbf{E}\biggl[K_{2}\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)\\ & \;\;\;\;\;\left(\log(\mathsf{L}_{2}^{-1}\mathsf{J}2^{\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)^{-1}+1)+1\right)^{2}\,\biggl|\,U_{0},U_{2},Y_{2},M_{0},M_{2}\biggr]\\ & \stackrel{(e)}{\le}(\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{2}}(U_{0};Y_{2})}+1)\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)\\ & \;\;\;\;\;\left(\log(\mathsf{L}_{2}^{-1}\mathsf{J}2^{\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)^{-1}+1)+1\right)^{2},\end{aligned}$$ where (a) is due to $(U_{1},Y_{1})\leftrightarrow(U_{0},U_{2},Y_{2},M_{0},M_{2})\leftrightarrow\mathfrak{P}_{2}$ (see Figure \[fig:bc2\_bayes\] right), (b) is due to the aforementioned application of the conditional Poisson matching lemma, (c) is by the same arguments as in the proof of Theorem \[thm:bc\_ncm\], (d) is by Proposition \[prop:phi\_ineq\], and (e) is due to and $K_{2}\leftrightarrow(U_{0},Y_{2},M_{0})\leftrightarrow(U_{2},M_{2})$ (see Figure \[fig:bc2\_bayes\] right). Hence $$\begin{aligned} & \mathbf{P}\{(M_{0},M_{0},M_{1},M_{2})\neq(\hat{M}_{00},\hat{M}_{01},\hat{M}_{1},\hat{M}_{2})\}\\ & \le\mathbf{E}\biggl[\min\biggl\{(\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{1}}(U_{0};Y_{1})}+1)\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\left(\log(\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1\right)^{2}\\ & \;\;\;\;+(\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{2}}(U_{0};Y_{2})}+1)\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)\\ & \;\;\;\;\;\;\left(\log(\mathsf{L}_{2}^{-1}\mathsf{J}2^{\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)^{-1}+1)+1\right)^{2},\,1\biggr\}\biggr]\\ & \le\mathbf{E}\biggl[\min\biggl\{\mathsf{L}_{0}\mathsf{L}_{1}\mathsf{J}A2^{-\iota_{U_{0},U_{1};Y_{1}}(U_{0},U_{1};Y_{1})}+\mathsf{L}_{1}\mathsf{J}A2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\\ & \;\;\;\;+\mathsf{L}_{0}\mathsf{L}_{2}\mathsf{J}^{-1}B2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})}+\mathsf{L}_{0}\mathsf{L}_{2}(1-\mathsf{J}^{-1})B2^{-\iota_{U_{0},U_{2};Y_{2}}(U_{0},U_{2};Y_{2})}\\ & \;\;\;\;+\mathsf{L}_{2}\mathsf{J}^{-1}B2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}+\mathsf{L}_{2}(1-\mathsf{J}^{-1})B2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})},\,1\biggr\}\biggr],\end{aligned}$$ where $A=(\log(\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1)^{2}$, $B=\bigl(\log((\mathsf{L}_{2}\mathsf{J}^{-1}2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}+\mathsf{L}_{2}(1-\mathsf{J}^{-1})2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})})^{-1}+1)+1\bigr)^{2}$. For , if the event in does not occur, by Proposition \[prop:phi\_ineq\], $$\begin{aligned} & (\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{1}}(U_{0};Y_{1})}+1)\mathsf{L}_{1}\mathsf{J}2^{-\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}\left(\log(\mathsf{L}_{1}^{-1}\mathsf{J}^{-1}2^{\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0})}+1)+1\right)^{2}\\ & \;\;+(\mathsf{L}_{0}2^{-\iota_{U_{0};Y_{2}}(U_{0};Y_{2})}+1)\mathsf{L}_{2}\mathsf{J}^{-1}2^{-\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)\\ & \;\;\;\;\;\left(\log(\mathsf{L}_{2}^{-1}\mathsf{J}2^{\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0})}(2^{\iota_{U_{1};U_{2}|U_{0}}(U_{1};U_{2}|U_{0})}+\mathsf{J}-1)^{-1}+1)+1\right)^{2}\\ & \le2^{1-\gamma}\left(2(\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0}))^{2}+2\gamma^{2}+14\right)+2^{2-\gamma}\left(2(\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0}))^{2}+2\gamma^{2}+14\right)\\ & \le2^{-\gamma}\left(8(\iota_{U_{1};Y_{1}|U_{0}}(U_{1};Y_{1}|U_{0}))^{2}+8(\iota_{U_{2},Y_{2}|U_{0}}(U_{2};Y_{2}|U_{0}))^{2}+12\gamma^{2}+84\right).\end{aligned}$$ ![\[fig:bc2\_bayes\]Left: The Bayesian network described in . Middle: The Bayesian network deduced from , $U_{0}=(\tilde{U}_{0})_{P_{U_{0}}\times\delta_{M_{0}}}$, $\check{U}_{1j}=(\tilde{U}_{1})_{P_{U_{1}|U_{0}}(\cdot|U_{0})\times\delta_{M_{0}}\times\delta_{M_{1}}}(j)$, and $(\check{U}_{01j},\check{M}_{01j})=(\tilde{U}_{0},\tilde{M}_{00})_{P_{U_{0}|Y_{1}}(\cdot|Y_{1})\times P_{M_{0}}}(j)$. Right: The Bayesian network describing the scheme. Note that all three are valid Bayesian networks, and the desired conditional independence relations can be deduced using d-separation.](bc2_bayes)

ArXiv

--- abstract: 'A first-order percolation transition, called explosive percolation, was recently discovered in evolution networks with random edge selection under a certain restriction. However, the network percolation with more realistic evolution mechanisms such as preferential attachment has not yet been concerned. We propose a tunable network percolation model by introducing a hybrid mechanism of edge selection into the Bohman-Frieze-Wormald (BFW) model, in which a parameter adjusts the relative weights between random and preferential selections. A large number of simulations indicate that there exist crossover phenomena of percolation transition by adjusting the parameter in the evolution processes. When the strategy of selecting a candidate edge is dominated by random selection, a single discontinuous percolation transition occurs. When a candidate edge is selected more preferentially based on node‘s degree, the size of the largest component undergoes multiple discontinuous jumps, which exhibits a peculiar difference from the network percolation of random selection with a certain restriction. Besides, the percolation transition becomes continuous when the candidate edge is selected completely preferentially.' author: - Xiaolong Chen - Chun Yang - Linfeng Zhong - Ming Tang title: Crossover phenomena of percolation transition in evolution networks with hybrid attachment --- **Percolation is a pervasive concept in graph theory and statistics physics, which has a broad application in network science, epidemiology, and so on. Now it’s a useful theoretical tool to study the large-scale connectivity of complex networks. It was known as a continuous percolation process of random networks before the product rule was introduced into the network evolution process by Achlioptas in 2009. In the percolation model proposed by Achlioptas, the size of the largest component jumps abruptly at a transition point, which is called explosive percolation, and it is considered as a first-order transition. Stimulated by the work of Achlioptas, a great amount of attention has been given to the research of the explosive percolation in evolution networks with random edge selection under a certain restriction. However, the percolation transition in a more realistic evolution process of networks is still lack of studying. Based on the classical Bohman-Frieze-Wormald model, we propose a percolation model in evolution process of networks with hybrid attachment. In our model, a tunable parameter is taken into account to adjust the way of selecting a candidate edge. Through numeric simulations, we find that there are crossover phenomena of percolation transition in the evolution process, which exhibit a peculiar difference from the classical Bohman-Frieze-Wormald network percolation. Specifically, when the strategy of selecting a candidate edge is dominated by random selection, there is only a single discontinuous jump of the largest component size. When a candidate edge is selected more preferentially based on node‘s degree, the largest component grows with multiple discontinuous jumps. Besides, nearly completely preferential selection makes the size of the largest component increase continuously. Our work reveals the effect of hybrid attachment mechanism on the network percolation transition, and offer some insights into understanding the network percolation in a more realistic evolution process.** Introduction {#sec:intro} ============ Percolation is used to describe the movement and filtering of fluids through porous materials in materials science [@grimmett1999percolation]. During the last five decades, percolation theory has brought new understanding to a broad range of topics in networks science [@boccaletti2006complex; @castellano2009statistical; @newman2010networks] and epidemiology [@moore2000epidemics; @fu2008epidemic]. During 1959 to 1961, Erd[ö]{}s and R[é]{}nyi proposed the evolution model of random graphs, i.e., Erd[ö]{}s-R[é]{}nyi (ER) model [@erdds1959random; @erd6s1960evolution]. In this model, a randomly selected link is added to the network at each time step. Such random edge selection rule generates a second-order percolation transition that the largest connected component grows continuously. In 2009, Achlioptas introduced a product rule of edge selection [@achlioptas2009explosive]. In the model, two candidate edges are randomly selected, and the one merging the two components with smaller product of their sizes is added to the network. It postpones the emergence of the giant connected component and hence leads to a abrupt jump of the largest component size, which is called as explosive percolation. It is in striking contrast to the second-order percolation transition in the classical ER model, and a great deal of researches have been triggered [@ziff2009explosive; @radicchi2009explosive; @cho2009percolation; @friedman2009construction; @riordan2011explosive; @lee2011continuity; @araujo2010explosive; @cho2013avoiding; @chen2011explosive; @d2015anomalous]. In particular, refs. [@cho2009percolation; @radicchi2009explosive] explored the explosive percolation of the configuration scale-free networks in a cooperative Achlioptas growth process. Ref. [@cho2009percolation] found that there exists a critical degree-exponent parameter $\lambda_c\in(2.3 , 2.4)$. If $\lambda<\lambda_c$, the transition threshold tends to zero in the thermodynamic limit, and the transition is continuous; otherwise, a discontinuous percolation transition occurs near a finite threshold. Almost at the same time, ref. [@radicchi2009explosive] got similar results through numeric simulations. However, it is proven rigorously that the explosive percolations for all the Achlioptas processes are actually continuous by subsequent researches [@da2010explosive; @grassberger2011explosive; @lee2011continuity]. Besides the Achlioptas process, another percolation model was introduced by Bohman, Frieze and Wormald (BFW) [@bohman2004avoidance]. In this model, a single edge is randomly selected at each step. If the size of the resulting component is less than a specified value, this edge would be added to the network, or a critical function is used to decide whether to receive it or not. Later on, Chen *et al*. [@chen2011explosive] demonstrated that the transition of the BFW model is actually discontinuous. Although percolation transition in random networks has been studied extensively, its application to real-world networks is still scarcely investigated [@rozenfeld2010explosive; @pan2011using]. Actually, in the evolution processes of most real-world networks, the strategy of selecting a newly added edge follows the mechanism of preferential selection instead of random selection. The preferential attachment was introduced by Barab[á]{}si and Albert [@barabasi1999emergence], and it is considered as a vital mechanism to describe the evolution of the real-world networks with heterogeneous architecture [@newman2001clustering; @de2007preferential; @holme2002growing; @small2015growing]. In the model, the probability that a new node leads an edge to an exist node $i$ is in proportion to its degree $k_i$, that is $\prod_{i}\sim k_i$. However, the architectures of most real-world networks can not be described by strictly scale-free or random model [@newman2001random; @fararo1964study]. Thus a more realistic network evolution model with hybrid attachment was proposed [@liu2002connectivity; @liu2003propagation]. In the model, the probability that a node $i$ is attached by a new edge is in proportion to $\prod_{i}\sim(1-q)k_i+q$, where $q$ controls the relative weights between the heterogeneous and homogeneous components in the evolution process. In this paper, we propose a hybrid BFW model (HBFW) with mixture of preferential and random edge selection to study the network percolation in a more realistic case. When $q=0$, the model degenerates to the original BFW model, where each candidate edge is selected randomly, and the degrees thus satisfy exponential distribution. When $q=1$, the network is strictly scale free as the preferential attachment dominates network evolution process. Simulation results demonstrate that three separate parameter regions $\left[0,q^{\uppercase\expandafter{\romannumeral1}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$ exist in the model, where $q^{\uppercase\expandafter{\romannumeral1}}_{c}\in (0.84, 0.86)$ and $q^{\uppercase\expandafter{\romannumeral2}}_c\in (0.97, 0.98)$ respectively stand for two critical values. And crossover phenomena of percolation transition occur between these regions. Specifically, there is only a single discontinuous transition when $q\in\left[0, q^{\uppercase\expandafter{\romannumeral1}}_c \right]$. And when $q\in\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, multiple discontinuous jumps of the largest component size occur in the evolution process. Moreover, when $q\in\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$, the first transition of the HBFW model becomes continuous. Most extremely, there is single giant component all through the process, when $q$ approaches to 1. It shows totally different properties of percolation transitions in a network with more realistic evolution mechanism, compared with those in random networks or configuration networks. PERCOLATION MODEL {#sec:model} ================= In our model, we use the mechanism of the hybrid attachment to modify the rule of selecting a candidate edge in the BFW model. By introducing a tunable parameter *q*, we get a hybrid rule of preferential selection and random selection. Initially, the system is a network *G* with *N* isolated nodes. Denote *K* as the cap size and set $\emph{K=2}$. $u\left(l\right)$ and $L\left(l\right)$ respectively represent the number of candidate edges and accepted edges at the current step *l*. And $t=L(l)/N$ is the density of accepted edges at the current step. Then we consider the evolution process of the HBFW model. Namely, at each step *l*, the two nodes attaching to a candidate edge $e_u$ is selected according to the following two steps: $1)$ The first node is selected randomly from the system. $2)$ The second node is selected by the rule of hybrid attachment. The node $i$ is selected by the preferential (random) selection strategy with probability $q$ ($1-q$). According to the selection rule described above, a candidate edge $e_u$ is obtained, and then it is judged by the BFW algorithm. $3)$ Edge addition in the BFW algorithm. We can decide whether the candidate $e_u$ should be accepted or not according to the following criteria: (***i***) Denote $C_1$ as the size of the largest component if $e_u$ is added to the system; (***ii***) If $C_1\leq K$, we accept $e_u$ , and set $L(l+1)=L(l)+1$, $u(l+1)=u(l)+1$, $l=l+1$; (***iii***) Otherwise, if $L(l)/u(l)<g(K)=1/2+(2K)^{-1/2}$, set $K=K+1$, and go to (*ii*), until $C_1\leq K$ or $L(l)/u(l)>g(K)$. If $C_1\leq K$, $e_u$ is added to the network, and set $L(l+1)=L(l)+1$, $u(l+1)=u(l)+1$, $l=l+1$; or if $L(l)/u(l)>g(K)$, $e_u$ is rejected, and set $u(l+1)=u(l)+1$, $l=l+1$; Repeat steps $1) \sim 3)$. Note that initially all the nodes in the system are isolate, so the selection rule of preferential attachment can not be applied to any node. Thus without loss of generality, at the first step, the two nodes are randomly selected. SIMULATION RESULTS {#sec:simulation} ================== First of all, we explore the effect of the hybrid attachment on the architecture of the generated networks. In order to demonstrate the result clearly and without loss of generality, we set the generated networks of $\langle k \rangle =5$. We exhibit the degree distribution of the networks with five typical values of parameter $q$, as shown in Figs. \[digDist(mix)\] (a) and (b). When *q* is relatively small (e.g., $q=0.0$ and $q=0.6$), a pair of nodes connected by a candidate edge are selected almost by the strategy of random selection, thus degrees of all nodes satisfy an exponential distribution [@newman2001random] \[see fig.  \[digDist(mix)\] (a)\]. With the increase of $q$, it is more likely to select a large-degree node, then the heterogeneity of the networks grows gradually. And after about $q=0.86$, the degree distribution of the networks evolves asymptotically to power-law distribution [@de2007preferential], as shown in Fig. \[digDist(mix)\](b). Next, we discuss the properties of percolation transition in the HBFW model. For clarity, we denote $NC_i \thicksim o(N)$ if $\lim_{N\rightarrow\infty}C_{i}\rightarrow 0$, where $C_i$ is the fraction of nodes in the *ith* largest component, and $C_1$ is defined as the order parameter of network percolation. If $\lim_{N\rightarrow\infty}C_{i}\rightarrow c, c\in \left(0,1 \right]$, we denote $NC_i\thicksim O \left( N \right)$. In order to investigate the impact of hybrid attachment on percolation transition, we perform simulations with varieties of $q$. In the evolution process, edges are selected in strict accordance with the rules mentioned above. We first focus on the time evolutions of $C_1$ and $C_2$. Fig. \[orderparam\] illustrates the effect of the hybrid attachment on the percolation transition of the HBFW model by exhibiting the results for $q=0.0, q=0.6, q=0.96$ and $q=0.99$. Fig. \[orderparam\] shows that when $q=0.0$, it degenerates to the BFW model and there is only a unique discontinuous jump of the largest component at the percolation threshold. But for $q>0$, as shown in Figs. \[orderparam\](b)$\sim$(d), we can see that the transition point decreases gradually with the probability $q$ of preferential attachment. Most importantly, for a large value of $q$ multiple jumps of $C_1$ and $C_2$ occur after the first percolation transition \[see Figs. \[orderparam\](c) and (d)\]. When $q\rightarrow1$, as shown in Fig. \[orderparam\](d), the first transition of $C_1$ and $C_2$ is prone to be continuous. ![(Color online) Typical degree distribution of generated networks with $\langle k\rangle =5$. (a) the degree distribution for $q=0.0$ and $q=0.6$, (b) the degree distribution for $q=0.86$, $q=0.96$, $q=0.99$. System size is set sa $N=10^6$.[]{data-label="digDist(mix)"}](degDist.eps){height="80mm" width="58mm"} ![(Color online) Time evolutions of $C_1$ and $C_2$ for four typical values of $q$. (a) and (b) respectively depict the evolution processes for $q=0.0$ and $q=0.6$, which are both smaller than $q^{\uppercase\expandafter{\romannumeral1}}_c$; (c) and (d) respectively depict the processes for $q=0.96$ and $q=0.99$ , which are both larger than $q^{\uppercase\expandafter{\romannumeral1}}_c$. System size is set as $N=10^6$. []{data-label="orderparam"}](orderparam.eps){width="1\linewidth"} From large numbers of simulations, we find that there exists two critical values $q^{\uppercase\expandafter{\romannumeral1}}_c\in (0.84, 0.86)$ and $q^{\uppercase\expandafter{\romannumeral2}}_c\in (0.97, 0.0.98)$. For $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$, there is only one jump of $C_1$ and $C_2$ at the transition point. And for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$, multiple jumps of the $C_1$ and $C_2$ emerge. However, the first transition of $C_1$ and $C_2$ is prone to be continuous for $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c, 1]$. In the following section, we focus on the time evolution of the largest component, and demonstrate the above conclusions from two aspects. Firstly, by analyzing the jump size of the largest component $\Delta C_1$ [@nagler2011impact], and the time interval $\Delta t_i, i=1...3$ between two adjacent jumps of the largest component, we demonstrate the existence of multiple discontinuous jumps for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$. And we strengthen the above results by investigating the susceptibility $\chi$ of $C_1$. Secondly, by using the technology of “critical window” [@achlioptas2009explosive], we examine whether the first transition is discontinuous or not for different value of $q$. ![(Color online) Finite size effects of multiple jumps of the largest component $\Delta{C_1}$ and the time intervals between each two adjacent jumps $\Delta{t_1}\sim\Delta{t_3}$. The four largest jumps of order parameter $\Delta{C^{1}_{1}}\sim \Delta{C^{4}_{1}}$ for $q=0.6$ (a) and $q=0.96$ (b). Time delays $\Delta{t_1}\sim\Delta{t_3}$ for $q=0.6$ (c) and $q=0.96$ (d). Data are averaged over 100 network realizations.[]{data-label="jump"}](jump.eps){width="1\linewidth"} Firstly, we demonstrate that there is only one discontinuous jump of $C_1$ for $q\in[0, q^{\uppercase\expandafter{\romannumeral1}}_c]$, and multiple discontinuous jumps for $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c, q^{\uppercase\expandafter{\romannumeral2}}_c]$. In fact, Ref. [@nagler2011impact] pointed out that if the maximum increase of the order parameter $\Delta C^{max}_{1}$ caused by adding a single edge satisfies the condition $$\label{maxJump} \lim_{N\rightarrow \infty}\Delta C^{max}_{1}/N>0,$$ the transition is strongly discontinuous. We consider the finite size effects of top 4 largest jumps of the largest component for different values of $q$. Figs. \[jump\] (a) and (b) show the cases of $q=0.6$ and $q=0.96$, where $\Delta C^{i}_1$ stands for the $ith$ largest jump size of the largest component. Simulations results illustrate that the size of the jumps are independent of the system size $N$. Therefore, the top four largest jumps of the largest component is discontinuous for $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$ in the thermodynamic limit. Next, we investigate the time delays between two adjacent jumps of the largest component to judge how many discontinuous jumps occur in the evolution process. The $\Delta t_i, i=1,2,3$ stands for the time delay between the $ith$ jump and the$(i+1)th$ jump. Fig. \[jump\](c) shows that when $q=0.6$, $\Delta t_i(i=1,2,3)$ converge to zero in the thermodynamic limit. This means that these discontinuous jumps in the HBFW process with $q=0.6$ occur at almost the same time when *N* is large enough. Thus there is only one discontinuous jump of $C_1$ when $N\rightarrow\infty$. For $q=0.96$, the two time delays converge asymptotically to two positive constant values in the thermodynamic limit, i.e. $\lim_{N\rightarrow \infty}\Delta t_{1}=0.206\pm 0.002$ and $\lim_{N\rightarrow \infty}\Delta t_{2}=0.057\pm 0.004$ \[see Fig. \[jump\](d)\]. This means that three separate discontinuous jumps of $C_1$ occur during the percolation process. ![(Color online) Time evolution and finite size effects of the susceptibility $\chi$. Time evolution of the $\chi$ for $q=0.6$ (a) and $q=0.96$ (b). (c) The maximum value of susceptibility $\chi_{max}$ versus network size for $q=0.6$ and $q=0.96$. (d) The other two peak values of susceptibility $\chi_{2}$ and $\chi_{3}$ versus network size for $q=0.6$ and $q=0.96$. Data are averaged over 100 network realizations.[]{data-label="deviation"}](deviation.eps){width="1\linewidth"} To strengthen the above conclusions, we consider the susceptibility $$\label{chi} \chi=\sqrt{\langle C^2_1\rangle-\langle C_1\rangle^2},$$ which quantifies the amplitude of the fluctuations of the largest component size [@shu2015numerical]. When a transition occur in the evolution process, $\chi$ reaches to a peak value $\chi_{max}$. Further, a none zero value of $\chi_{max}$ in the thermodynamic limit indicates a discontinuous transition [@binder1981finite; @voss1984fractal]. In Fig. \[deviation\], we exhibit the susceptibility of the largest component sizes for $q=0.6$ and $q=0.96$ with corresponding to Fig. \[jump\]. Fig. \[deviation\](a) shows that there is only one peak of $\chi$, and Fig. \[deviation\](b) tells us that it has three obvious peaks of $\chi$ as time *t* increases. For the sake of clarity, we denote the second and third largest value of $\chi$ for $q=0.96$ as $\chi_{2}$ and $\chi_{3}$. Results of Figs. \[deviation\](a) and (b) show that for $q=0.96$, there are three jumps of the largest component size, but for $q=0.6$ only one jump of the largest component size occurs. Besides, Figs. \[deviation\](c) and (d) demonstrate that the three peak values of the susceptibility for $q=0.96$, and the maximum value of the susceptibility for $q=0.6$ tend to be nonzero values in the thermodynamic limit. Thus, it leads to the conclusion that the jump of the largest component for $q=0.6$ and the three jumps for $q=0.96$ are essentially discontinuous. ![(Color online) Relationship between the rescaled size of critical window $\Delta(\alpha,\beta)$ and network size *N*. (a) $\alpha=1/2,\beta=0.25$ and (b) $\alpha=1/2,\beta=0.1$ are chosen in $\Delta(\alpha,\beta)$ for $q=0.86,q=0.96$, and $q=0.99$. Data are averaged over 100 realizations.[]{data-label="cw"}](cw.eps){height="86mm" width="55mm"} At last, we demonstrate the continuity of the first transition for $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c, 1]$ in the HBFW model by applying the theory of “critical window” proposed in Ref. [@achlioptas2009explosive]. Let $L_0$ denote the number of added edges at the last moment for $$\label{L0} NC_1\leq N^\alpha,$$ and $L_1$ as the number of added edges at the first moment for $$\label{L1} NC_1\geq \beta N$$ where $\alpha,\beta \in(0,1)$. $\Delta (\alpha,\beta)=L_1-L_0$ denotes the number of added edges during the evolution process, which is called as critical window. Fig. \[cw\] shows the rescaled value of $\Delta(\alpha, \beta)/N$ as a function of system size *N* for $q=0.86$, $q=0.96$, and $q=0.99$. We first set $\alpha=1/2, \beta=0.25$. As shown in Fig. \[cw\](a), $\Delta(\alpha, \beta)/N$ is sublinear to $N$ and satisfies $\lim_{N\rightarrow \infty}\Delta(\alpha, \beta)/N=0$ for $q=0.86$ and $q=0.96$, while $\lim_{N\rightarrow \infty}\Delta(\alpha, \beta)/N=0.732\pm0.002$ for $q=0.99$ \[see Fig. \[cw\](a)\]. For more accurate, we lower the value of $\beta$. In Fig. \[cw\](b), we set $\beta=0.1$ and get the accordant result as Fig. \[cw\](a). Figs. \[cw\](a) and (b) reveal the fact that the first jump of the HBFW process is discontinuous for $q=0.86$ and $q=0.96$, but it is continuous for $q=0.99$. More simulations demonstrate the existence of the critical value $q^{\uppercase\expandafter{\romannumeral2}}_c\in(0.97,0.98)$, below which the first jump of the largest component in HBFW process is discontinuous, but above which it is continuous. ![(Color online) The component distributions at the vicinity of transition threshold for $q=0.0$ (a), $q=0.96$ (b) and $q=0.99$ (c). System size is $N=10^6$.[]{data-label="cluster"}](clusterDist.eps){height="120mm" width="65mm"} Finally, we give a qualitative analysis for the two crossover phenomena of percolation transitions in the HBFW model. In the BFW model, the evolution of the largest component is restricted by the cap size $K$. Thus at each time step, the small components are merged together by a newly added edge with high probability. Consequently, a mass of components, whose sizes are approximately equal to the largest component size, emerge in the subcritical regime. In other words, a “powder key” is formed, which is a necessary condition for the discontinuous percolation transition [@friedman2009construction]. As shown in Fig. \[cluster\] (a), there are a mass of components with size of about two hundred nodes at $t=0.96$, and with size of about seven hundred nodes at $t=0.97$. Subsequently, a vanishingly small number of added edges can merge these small components, and thus lead to an explosive percolation. And at $t=0.98$, two giant components formed. The hybrid attachment mechanism in the HBFW model makes evolution procedure much more complex than the classical BFW process, which leads to peculiar properties of the percolation transitions. Generally, there are two underlying competed mechanisms behind the percolation procedures in the HBFW model, which are the suppression and promotion of growth of the largest component, respectively. On the one hand, the cap size $K$ restricts the formation of large components. On the other hand, with the increase of $q$, the probability of selecting large-degree nodes rises, which accelerates the growth of large components. When $q<q^{\uppercase\expandafter{\romannumeral1}}_c$, the strategy of selecting a candidate node is predominated by the random selection \[see Fig. \[digDist(mix)\] (a)\], the suppression plays a main role, rendering the percolation type similar to BFW model \[see Figs. \[orderparam\] (a) and (b)\]. With the increase of $q$, edges are more likely to attach to large-degree nodes \[see Fig. \[digDist(mix)\] (b)\]. Owing to the suppression of cap size *K*, the large-degree nodes absorb more and more small components or isolate nodes into the components, to which these hub nodes belong. Consequently, a mass of hub-centric components with sizes being sublinear to the system size $N$ appear in the subcritical regime. As shown in Fig. \[cluster\] (b), there are many components with size of about one thousand nodes at $t=0.6$. In the subsequent procedure, each added edge would combine two of these hub-centric components with a very small probability, and thus leads to a discontinuous jump of the largest component. As shown in Fig. \[cluster\], three discontinuous jumps of the largest component size occur in the time duration $[0.6,0.65]$, $(0.65,0.7]$ and $(0.7,0.85]$. When $q$ is large enough, i.e., $q>q^{\uppercase\expandafter{\romannumeral2}}_c$, the cap value $K$ that constrains the growth of the largest component won’t work anymore. Affected by the mechanism of preferential attachment, the large-degree nodes absorb small components or isolate nodes continuously, and giant components form at the early stage of the evolution process. Thus, it leads to the continuous growth of the largest component. As shown in Fig. \[cluster\] (c), several components with size of about ten thousand nodes form at $t=0.5$. We see that the large components absorb more small components, thus several giant components form at $t=0.6$. In the following time steps, these giant components are merged by edges, and at last the two stable giant component exist in the supercritical regime, e.g., $t=0.7$ and $t=0.95$. If $q$ approaches to 1, at each time step, the probability of selecting the second node approximates to $\prod_{i}\sim k_i$. The isolate nodes or small components will be absorbed into the large components continuously by the large degree nodes. And when $q=1$, we can imagine that there will be a single giant component to which those large-degree nodes belong, and it grows continuously all through the process. Thus, there will be no jump of the largest component. Discussion {#sec:dis} ========== In this paper, we propose a network percolation model with the hybrid attachment based on the BFW process. In the model, the tune parameter $q$ controls the evolution strategy. When $q$ is relatively small, the strategy of random edge selection dominates the evolution process. When $q$ becomes larger, the strategy of preferential edge selection dominates the evolution process. A large number of numeric simulations indicate that there exist crossover phenomena of percolation transition between three separate regions $\left[0,q^{\uppercase\expandafter{\romannumeral1}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c \right]$, $\left(q^{\uppercase\expandafter{\romannumeral2}}_c, 1\right]$. In the region $q\in[0,q^{\uppercase\expandafter{\romannumeral1}}_c]$, the strategy of random edge selection dominates the evolution process, and with the constraint of forming the giant component, there is only one discontinuous jump of the largest component at the percolation threshold as the origin BFW model. In the region $q\in(q^{\uppercase\expandafter{\romannumeral1}}_c,q^{\uppercase\expandafter{\romannumeral2}}_c]$, the strategy of more preferential edge selection dominates the evolution process, and multiple discontinuous jumps occur as a result of competition between the mechanisms of constraining and accelerating the forming of the giant component, which are caused by the restricted of the cap size $K$, and the preferential selection of a candidate edge respectively. Further, in the region $q\in(q^{\uppercase\expandafter{\romannumeral2}}_c,1]$, with nearly complete preferential selection, the mechanism of constraining the giant connected component is no longer function, the largest component grows continuously at the first transition point. And if $q$ approaches to 1, a single giant component grows continuously throughout the evolution process. Percolation of networks has been one of the hottest topics in recent years, and it is considered as an important tool to study the robustness of networks [@parshani2010interdependent], spreading of epidemics [@newman2002spread] and so on. Although many important and interesting percolation phenomena are observed on both random networks and lattices, it is the first time to study the percolation in network evolution process with hybrid attachment, which will help us to understand the percolation transition in a more realistic process. It is unusual to find the crossover phenomena of percolation transition in the network evolution process, which have recently stimulated broad attention in other fields, such as the processes of social contagions [@wang2016dynamics]. Specially, a recent work studied the percolation transition in scale-free networks and the multiple discontinuous percolation transitions are observed [@chen2015multiple]. We study the percolation of networks in a more realistic evolution process from the aspect of numeric simulations, but its theoretical analysis need to be carried out. Besides, the percolation transition in real-world networks needs further works. This work was partially supported by the National Natural Science Foundation of China under Grants Nos. 11105025, 11575041 and 61433014, and Project No. 9140A06030614DZ02083. [100]{} G. Grimmett. *What is Percolation* (Springer, 1999). S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D.-U. Hwang, Phys. Rep. **4**, 424 (2006). C. Castellano, S. Fortunato, V. Loreto, Rev. Mod. Phys. **81**, 591 (2009). M. Newman. *Networks: an introduction* (OUP Oxford, 2010). C. Moore, M. E. Newman, Phys. Rev. E **61**, 5678 (2000). X. Fu, M. Small, D.M. Walker, H. Zhang, Phys. Rev. E **77**, 036113 (2008). P. Erd[ö]{}s, A. R[é]{}nyi, Publ. Math. Debrecen **6**, 290 (1959). P. Erd[ö]{}s, A. R[é]{}nyi, Publ. Math. Inst. Hungar. Acad. Sci **5**, 17 (1960). D. Achlioptas, R. M. D’Souza, J. Spencer, Science **323**, 1453 (2009). R. M. Ziff, Phys. Rev. Lett. **103**, 045701 (2009). Y. S. Cho, J. S. Kim, J. Park, B. Kahng, D. Kim, Phys. Rev. Lett. **103**, 135702 (2009). F. Radicchi, S. Fortunato, Phys. Rev. Lett. **103**, 168701 (2009). E. J. Friedman, A. S. Landsberg, Phys. Rev. Lett. **103**, 255701 (2009). O. Riordan, L. Warnke, Science **333**, 322 (2011). H. K. Lee, B. J. Kim, H. Park, Phys. Rev. E **84**, 020101 (2011). N. A. Araujo, H. J. Herrmann, Phys. Rev. . **105**, 035701 (2010) Y. S. Cho, S. Hwang, H. J. Herrmann, B. Kahng, Science **339**, 1185 (2013). W. Chen, R. M. D’Souza, Phys. Rev. Lett. **106**, 115701 (2011). R. M. D’Souza, J. Nagler, Nature Phys. **11**, 531 (2015). J. Nagler, A. Levina, M. Timme, Nature Phys. **7**, 265 (2011) R. A. da Costa, S. N. Dorogovtsev, A. V. Goltsev, J. F. F. Mendes, Phys. Rev. Lett. **105**, 255701 (2010). P. Grassberger, C. Christensen, G. Bizhani, S.-W. Son, M. Paczuski, Phys. Rev. Lett. **106**, 225701 (2011). T. Bohman, A. Frieze, N. C. Wormald, Random Struct. Algor. **25**, 432 (2004). H. D. Rozenfeld, L. K. Gallos, H. A. Makse, Eur. Phys. J. B **75**, 305 (2010). R. K. Pan, M. Kivel[ä]{}, J. Saram[ä]{}ki, K. Kaski, J. Kert[é]{}sz, Phys. Rev. E **83**, 046112 (2011). A.-L. Barab[á]{}si, R. Albert, Science **286**, 509 (1999). M. E. Newman, Phys. Rev. E **64**, 025102 (2001). B. F. de Blasio, [Å]{}. Svensson, F. Liljeros, Proc. Natl. Acad. Sci. USA **104**, 10762 (2007). P. Holme, B. J. Kim, Phys. Rev. E **65**, 026107 (2002). M. Small, Y. Li, T. Stemler, K. Judd, Phys. Rev. E **91**, 042801 (2015). M. E. Newman, S. H. Strogatz, D. J. Watts, Phys. Rev. E **64**, 026118 (2001). T. J. Fararo, M. H. Sunshine. *A study of a biased friendship net* (Youth Development Center, Syracuse University, Syracuse, 1964). Z. Liu, Y.-C. Lai, N. Ye, P. Dasgupta, Phys. Lett. A **303(5)**, 337 (2002). Z. Liu, Y.-C. Lai, N. Ye, Phys. Rev. E **67(3)**, 031911 (2003). P. Shu, W. Wang, M. Tang, Chaos **25**, 063104 (2015). K. Binder, Z. Phys. B: Conden. Matter **43**, 119 (1981). R. F. Voss, J. Phys. A: Math. Gen. **17**, L373 (1984). R. Parshani, S. V. Buldyrev, S. Havlin, Phys. Rev. Lett. **105**, 048701 (2010). M. E. J. Newman, Phys. Rev. E **66**, 016128 (2002). W. Wang, M. Tang, P. Shu, Z. Wang, New J. Phys. **18**, 013029 (2016). W. Chen, Z. M. Zheng, R. M. D’Souza, J. Stat. Mech.: Theory E, **2015**, P04011 (2015).

ArXiv

--- abstract: 'Economic systems, traditionally analyzed as almost independent national systems, are increasingly connected on a global scale. Only recently becoming available, the World Input-Output Database (WIOD) is one of the first efforts to construct the multi-regional input-output (MRIO) tables at the global level. By viewing the world input-output system as an interdependent network where the nodes are the individual industries in different economies and the edges are the monetary goods flows between industries, we study the network properties of the so-called world input-output network (WION) and document its evolution over time. We are able to quantify not only some global network properties such as assortativity, clustering coefficient, and degree and strength distributions, but also its subgraph structure and dynamics by using community detection techniques. Over time, we detect a marked increase in cross-country connectivity of the production system, only temporarily interrupted by the 2008-2009 crisis. Moreover, we find a growing input-output regional community in Europe led by Germany and the rise of China in the global production system. Finally, we use the network-based PageRank centrality and community coreness measure to identify the key industries and economies in the WION and the results are different from the one obtained by the traditional final-demand-weighted backward linkage measure.' author: - Federica Cerina - Zhen Zhu - Alessandro Chessa - Massimo Riccaboni bibliography: - 'WION.bib' title: 'World Input-Output Network' --- Introduction {#sec:intro} ============ As the global economy becomes increasingly integrated, an isolated view based on the national input-output table[^1] is no longer sufficient to assess an individual economy’s strength and weakness, not to mention finding solutions to global challenges such as climate change and financial crises. Hence, a multi-regional input-output (MRIO) framework is needed to draw a high-resolution representation of the global economy [@wiedmann2011quo]. In practice, however, due to the expensive process of collecting data and the variety of classifications used by different agencies, for a long time, the input-output tables have only been available for a limited number of countries and for discontinuous years. Fortunately, the fully-fledged MRIO databases started to become available in recent years[^2]. Unlike the national input-output table where exports and imports are aggregated and appended to final demand and country-specific value added respectively, for each individual economy, the MRIO table splits its exports into intermediate use and final use in every foreign economy and also traces its imports back to the industry origins in every foreign economy. As a result, the inter-industrial relationships in the MRIO table are recorded not only within the same economy but also across economies. The availability of the MRIO databases was soon followed by a wave of empirical studies of topics ranging from global value chains and trade fragmentation in economics [@baldwin2013supply; @timmer2013fragmentation; @koopman2014tracing] to global environmental accounting in ecology and resources management [@lenzen2012international; @lenzen2013international; @wiedmann2013material]. However, to the best of our knowledge, our paper is the first attempt to explore the MRIO tables from a networks perspective, even though there have been some networks studies of the input-output tables at the national level and for selected countries [@blochl2011vertex; @mcnerney2013network; @contreras2014]. Complex networks are a modern way to characterize mathematically a series of different systems in the shape of subunits (nodes) connected by their interaction (edges) [@ABRMP; @SIAM]. Such modeling has been proved to be fruitful for the description of a variety of different phenomena ranging from biology [@GCbook2] to economics [@kitsak2010; @pammolli2002; @riccaboni2010; @riccaboni2013; @chessa2013] and finance [@nature]. Here we move forward by considering the global MRIO system as a world input-output network (WION), where the nodes are the individual industries in different economies and the edges are the monetary goods flows[^3] between industries, similarly to what have been done recently by Acemoglu, Carvalho, Ozdaglar and Tahbaz-Salehi for the US economy only [@acemoglu2012network]. Different from many network systems observed in reality, the WION has the following features: 1) It is directed and weighted, i.e., an industry can act as both a seller and a buyer at the same time and the monetary goods flows between industries vary a lot; 2) It is much denser within the same economy than across economies, i.e., despite the continuously integrated global economy, most economic transactions still happen within the country border;[^4] 3) It is with strong self-loops, i.e., an industry can acquire a significant amount of inputs from itself[^5]. Taking into account the features above, we explore the WION by quantifying not only some global network properties such as assortativity but also some local network properties such as PageRank centrality. Furthermore, we apply community detection and community core detection techniques to examine the structure of the WION over time. This paper makes some significant contributions to the literature of input-output economics. First, it is the first attempt to quantify the network properties of the WION by taking into account its edge weights and directedness[^6]. By doing that, we detect a marked increase in cross-country connectivity, apart from a sharp drop in 2009 due to the financial crisis. Second, the community detection results reveal growing input-output international communities. Among them, we notice in particular the emergence of a large European community led by Germany. Third, we use the network-based PageRank centrality and community coreness measure to identify the key industries and economies in the WION and the results are different from the one obtained by the traditional final-demand-weighted backward linkage measure. The rest of the paper is structured as follows. Section \[sec:data\] describes the database used and the MRIO framework. We also conduct a basic MRIO analysis to identify the key industries at the global level in this section. Section \[sec:properties\] quantifies some global network properties of the WION and its subgraph structure and dynamics by using community detection techniques. Moreover, we use the network-based PageRank centrality and community coreness measure to identify the key industries in the WION. Finally, Section \[sec:conclusion\] concludes the paper. The Data Description and the Leontief-Inverse-Based Method of Identifying the Key Industries {#sec:data} ============================================================================================ The WIOD Data and the MRIO Framework ------------------------------------ We use the World Input-Output Database (WIOD) [@timmer2012world] to map out the WION. At the time of writing, the WIOD input-output tables cover 35 industries for each of the 40 economies (27 EU countries and 13 major economies in other regions) plus the rest of the world (RoW) and the years from 1995 to 2011[^7]. For each year, there is a harmonized global level input-output table recording the input-output relationships between any pair of industries in any pair of economies[^8]. The numbers in the WIOD are in current basic (producers’) prices and are expressed in millions of US dollars[^9]. Table \[table\_1\] shows an example of a MRIO table with two economies and two industries. The $4\times4$ inter-industry table is called the transactions matrix and is often denoted by $\boldsymbol{\mathrm{Z}}$. The rows of $\boldsymbol{\mathrm{Z}}$ record the distributions of the industry outputs throughout the two economies while the columns of $\boldsymbol{\mathrm{Z}}$ record the composition of inputs required by each industry. Notice that in this example all the industries buy inputs from themselves, which is often observed in real data. Besides intermediate industry use, the remaining outputs are absorbed by the additional columns of final demand, which includes household consumption, government expenditure, and so forth[^10]. Similarly, production necessitates not only inter-industry transactions but also labor, management, depreciation of capital, and taxes, which are summarized as the additional row of value added. The final demand matrix is often denoted by $\boldsymbol{\mathrm{F}}$ and the value added vector is often denoted by $\boldsymbol{\mathrm{v}}$. Finally, the last row and the last column record the total industry outputs and its vector is denoted by $\boldsymbol{\mathrm{x}}$. The Leontief-Inverse-Based Method of Identifying the Key Industries ------------------------------------------------------------------- If we use $\boldsymbol{\mathrm{i}}$ to denote a summation vector of conformable size, i.e., a vector of all 1’s with the length conformable to the multiplying matrix, and let $\boldsymbol{\mathrm{F}}\boldsymbol{\mathrm{i}}=\boldsymbol{\mathrm{f}}$, we then have $\boldsymbol{\mathrm{Z}}\boldsymbol{\mathrm{i}}+\boldsymbol{\mathrm{f}}=\boldsymbol{\mathrm{x}}$. Furthermore, if dividing each column of $\boldsymbol{\mathrm{Z}}$ by its corresponding total output in $\boldsymbol{\mathrm{x}}$, we get the so-called technical coefficients matrix $\boldsymbol{\mathrm{A}}$[^11]. Replacing $\boldsymbol{\mathrm{Z}}\boldsymbol{\mathrm{i}}$ with $\boldsymbol{\mathrm{A}}\boldsymbol{\mathrm{x}}$, we rewrite the above equation as $\boldsymbol{\mathrm{A}}\boldsymbol{\mathrm{x}}+\boldsymbol{\mathrm{f}}=\boldsymbol{\mathrm{x}}$. It can be rearranged as $(\boldsymbol{\mathrm{I}}-\boldsymbol{\mathrm{A}})\boldsymbol{\mathrm{x}}=\boldsymbol{\mathrm{f}}$. Finally, we can solve $\boldsymbol{\mathrm{x}}$ as follows: $$\boldsymbol{\mathrm{x}}=(\boldsymbol{\mathrm{I}}-\boldsymbol{\mathrm{A}})^{-1}\boldsymbol{\mathrm{f}}$$ where matrix $(\boldsymbol{\mathrm{I}}-\boldsymbol{\mathrm{A}})^{-1}$ is often denoted by $\boldsymbol{\mathrm{L}}$ and is called the Leontief inverse. The intuition behind the Leontief inverse is that an increase in the final demand of an industry’s output will induce not only more production from the industry itself but also more from other related industries because more inputs are required. Therefore, the Leontief inverse takes into account both the direct and indirect effects of a demand increase. For instance, $L_{ij}$ measures the total output produced in Industry $i$ given a one-unit increase in Industry $j$’s final demand[^12]. As a result, $\boldsymbol{\mathrm{i}}'\boldsymbol{\mathrm{L}}$ sums up each column of $\boldsymbol{\mathrm{L}}$ and each sum measures the total output of all the industries given a one-unit increase in the corresponding industry’s final demand. The vector $\boldsymbol{\mathrm{i}}'\boldsymbol{\mathrm{L}}$ is called the backward linkage measure[^13] and can be used to rank the industries and identify the key ones in the economy [@yotopoulos1973balanced]. However, as pointed out by Laumas [@laumas1976weighting], the key assumption embedded in the backward linkage measure is that every industry is assigned with the same weight (or unweighted), which is far from the reality. The problem with the unweighted backward linkage measure can be demonstrated by using the hypothetical data from Table \[table\_1\]. The calculated $\boldsymbol{\mathrm{i}}'\boldsymbol{\mathrm{L}}$ is $\begin{bmatrix} 2.0688 & 1.8377 & 1.2223 & 1.1854 \end{bmatrix}$, which considers the industries in Economy 1 more important than the ones in Economy 2, despite the fact that Economy 2 is a lot larger than Economy 1 in terms of total outputs. The industries of the 40 economies covered in the WIOD are very heterogeneous in terms of both total outputs and technical structure, which certainly makes the unweighted backward linkage measure not a good choice to identify the most central industries on a global scale. In order to identify the key industries in the WIOD, we hence follow Laumas [@laumas1976weighting] and use the final-demand-weighted backward linkage measure, which is denoted by $\boldsymbol{\mathrm{w}}$ and is defined here as the Hadamard (element-wise) product of the vector of the unweighted backward linkage measure and the vector of the percentage shares of the total final demand across industries, i.e., $$\boldsymbol{\mathrm{w}}=\boldsymbol{\mathrm{i}}'\boldsymbol{\mathrm{L}}\circ\frac{\boldsymbol{\mathrm{f}}'}{\boldsymbol{\mathrm{i}}'\boldsymbol{\mathrm{f}}}$$ where $\circ$ is the element-wise multiplication operator. Table \[table\_2\] shows the top 20 industries for the years 1995, 2003, and 2011, respectively. The first column of each year is produced by the final-demand-weighted backward linkage measure, i.e., $\boldsymbol{\mathrm{w}}$. For the selected years, only four large economies, China, Germany, Japan, and USA, ever qualified for the top 20. Another noticeable change over time is the rise of China, which topped the list in 2011 with its industry of construction. Tables \[table\_A3\] and \[table\_A4\] in the appendix provide an alternative way of viewing the key industries and economies over time identified by the final-demand-weighted backward linkage measure. In particular, Table \[table\_A3\] lists the most important economies by industry while Table \[table\_A4\] lists the most important industries by economy. The Network Properties of the WION and the Network-Based Methods of Identifying the Key Industries {#sec:properties} ================================================================================================== As mentioned in Section \[sec:intro\], the complex networks approach has been widely used in economics and finance in recent years [@kitsak2010; @pammolli2002; @riccaboni2010; @riccaboni2013; @chessa2013; @nature]. Designed to keep track of the inter-industrial relationships, the input-output system is an ideal test bed for network science. Particularly the MRIO system can be viewed as an interdependent complex network, i.e., the WION, where the nodes are the individual industries in different economies and the edges are the flows between industries. Figure \[table1Net\] provides a topological view of Table \[table\_1\]. The blue nodes are the individual industries. The red nodes are the value added sources from the two economies, whereas the green nodes are the final demand destinations in the two economies. The edges are with arrows indicating the directions of the monetary goods flows[^14] and with varying widths indicating the magnitudes of the flows. The color of the edge is set the same as the source node’s. Finally, because we are only concerned with the inter-industrial input-output relationships, when formulating the WION, we focus our attention on the network among the blue nodes. \[table\_2\] The visualization of the WION in 1995 (Figure \[WION1995\]) and 2011 (Figure \[WION2011\]) are shown in Figure \[WION9511\]. Each node represents a certain industry in a certain economy. The size of the node is proportional to its total degree. The edges are directed and only those with strength greater than one billion US dollars are present. Finally, different colors represent different economies. Clearly the WION has become denser over time and some countries like China have moved to the core of the network, thus confirming the results in Table \[table\_2\]. \ In this section, we first examine some global network properties of the WION such as assortativity, clustering coefficient, and degree and strength distributions. We also study the subgraph structure and dynamics of the WION by using community detection techniques. Finally, we use the network-based PageRank centrality and community coreness measure to identify the key industries and economies in the WION and the results are different from the one obtained by the above Leontief-inverse-based method. The Global Network Properties of the WION ----------------------------------------- Because the WION is directed, we can calculate the assortativity coefficient in three ways, namely, in-degree assortativity, out-degree assortativity, and total-degree assortativity. As shown in Figure \[assort\], they all behave similarly over time. First, they have all been negative throughout the whole period. Since assortativity measures the tendencies of nodes to connect with other nodes that have similar (or dissimilar) degrees as themselves, a negative coefficient means that dissimilar nodes are (slightly[^15]) more likely to be connected.[^16] One possible explanation of the negativity is that high-degree industries such as construction often take inputs (or supply outputs) from (or to) low-degree industries such as transport services. Moreover, the spatial constraints (each node has only few neighboring nodes in the same country) introduce degree-degree anticorrelations (disassortativity) since high degree sectors are in different countries and the probability to connect decays with distance [@havlin2014]. Second, all the coefficients show an increasing trend before 2007 and a significant decline after 2007. The behavior of the assortativity measures seems to be correlated with the trend of the foreign share in the inter-industrial transactions over time (Figure \[foreign\_all\]). That is, we can calculate a globalization indicator as the percentage of inputs from foreign origins (or equivalently, the percentage of outputs to foreign destinations) of the transactions matrix $\boldsymbol{\mathrm{Z}}$ of the 40 WIOD economies. Same as observed in assortativity, the foreign share of $\boldsymbol{\mathrm{Z}}$ had a steady growth (from 9.9% in 1995 to 12.8% in 2007) before 2007 and a sharp decrease after 2007[^17]. The increase in the foreign share implies more interactions across economies and hence tends to make the WION less dissortative. The opposite happens when the foreign share goes down as a result of the global financial crisis. Third, we notice that the in-degree assortativity tends to be lower than the out-degree assortativity, but there is a tendency to close the gap between the two measures. We interpret this evidence as a clear signal of the globalization of production chains, that is to say, both global buying and selling hubs have now a higher chance to be connected across borders. \ The hump-shaped behavior is also observed in the clustering coefficient. Figure \[cluster\_all\] shows that the average weighted clustering coefficient of the WION has been steadily increasing but was followed by a decline since 2007. Again, a possible explanation is that the booming economy before 2007 introduced more interactions between industries, hence higher clustering coefficient, and the financial crisis after 2007 stifled the excess relationships. \ We can also examine the global network properties of the WION by plotting its degree and strength distributions. Recall that the transactions matrix $\boldsymbol{\mathrm{Z}}$ is essentially a weighted adjacency matrix, which records the edge weights between any pair of nodes in the WION. If we denote the regular binary adjacency matrix as $\boldsymbol{\mathrm{D}}$, where $D_{ij}=D_{ji}=1$ if either $Z_{ij}>0$ or $Z_{ji}>0$, then we have the following definitions for a given node $i$: 1) In-degree: $D_i^{in}=\sum_{j\neq i}D_{ji}$; 2) Out-degree: $D_i^{out}=\sum_{j\neq i}D_{ij}$; 3) Total-degree: $D_i^{total}=D_i^{in}+D_i^{out}$; 4) In-strength: $S_i^{in}=\sum_{j\neq i}Z_{ji}$; 5) Out-strength: $S_i^{out}=\sum_{j\neq i}Z_{ij}$; 6) Total-strength: $S_i^{total}=S_i^{in}+S_i^{out}$. As shown in Figure \[degreeDist\], unlike other network systems such as the internet, where the degree distributions follow the power law, the WION is characterized by the highly left-skewed degree distributions. Most nodes enjoy high-degree connections in the WION because the industries are highly aggregated. That is, it is hard to find two completely disconnected industries given the high level of aggregation. Furthermore, the WION is almost complete, i.e., every node is connected with almost every node, if represented by unweighted edges[^18]. We can also take into account the edge weights and examine the strength distributions of the WION. Figure \[strengthDist\] shows the in-strength, out-strength, and total-strength distributions for the years 1995, 2003, and 2011. We perform Gabaix-Ibragimov test [@gabaix2009power; @gabaix2011rank] to see if the tails of the distributions are Pareto but find no significant power-law tails. Moreover, like the previous studies at the national level [@mcnerney2013network], the strength distributions can be well approximated by the log-normal distributions. As reasoned by Acemoglu et al. [@acemoglu2012network], this asymmetric and heavy-tailed distribution of strength in the WION may serve as the origin of economic fluctuations. The Community Detection in the WION ----------------------------------- Another main property of networks is the community structure, i.e. the partition of a network into clusters, with many edges connecting nodes in the same cluster and few connecting nodes between different ones [@fortunato2010]. In the following we use the modularity optimization method introduced by Newman and Girvan [@newman2004finding]. It is based on the idea that a random graph is not expected to have a community structure. Therefore, the possible existence of clusters is revealed by the comparison between the actual density of edges in a subgraph and the expected density if the nodes are attached randomly. The expected edge density depends on the chosen null model, i.e., a copy of the original graph keeping some of its structural properties but without community structure [@fortunato2010]. The most popular null model, introduced by Newman and Girvan [@newman2004finding], keeps the degree sequence and consists of a randomized version of the original graph, where edges are rewired at random, under the constraint that the expected degree of each node matches the degree of the node in the original graph. The modularity function to be optimized is, then, defined as: $$Q = \frac{1}{2m} \sum_{ij} (A_{ij}-P_{ij})\delta({C_i,C_j})$$ \[eq3\] where the summation operator runs over all the node pairs. $A$ is the adjacency matrix, and $m$ is the total number of edges. The $\delta$ function equals 1 if the two nodes $i$ and $j$ are in the same community and 0 otherwise. Finally, $P_{ij} = \frac{k_i k_j}{2m}$ is the probability of the presence of an edge between the two nodes $i$ and $j$ in the randomized null model. Figures \[fig3\], \[fig4\], and \[fig5\] report the community detection results for the selected years 1995, 2003, and 2011, respectively[^19]. The 40 countries in the WIOD are arranged by rows while the 35 industries are arranged by columns. Different colors indicate different communities detected. There are two interesting findings in our results. First, most communities were based on a single economy, i.e., the same color often goes through a single row. This echoes one of the features of the WION mentioned in Section \[sec:intro\], i.e., most of the inter-industrial activities are still restricted in the country border. Second, for all the three years selected, we always color the community involving Germany in red and put it on the top. As a result, our algorithm captures a growing Germany-centered[^20] input-output community. Since the WIOD monetary goods flows are based on undeflated current prices, one possible reason for the emergence of the German community is that the community members may have experienced significantly more inflation and/or exchange rate volatility than other regions in the world. Referring to the World Bank inflation data and the exchange rate data used in the WIOD, we show that this is hardly the case. Panel (a) of Figure \[inflation\_exchange\] in the appendix compares the average inflation rate, i.e., the annual GDP deflator, across all the WIOD economies[^21] with the average annual GDP deflator across the 9 major member economies in the German community detected in 2011, i.e., Germany, Austria, Belgium, Luxembourg, Hungary, Czech Republic, Slovakia, Slovenia, and Poland. During 1995-2011, the average inflation of the German community was almost always below that of all the WIOD economies. Panel (b) of Figure \[inflation\_exchange\] in the appendix compares the average exchange rate, i.e., US dollars per unit of local currency, across all the WIOD economies[^22] with the average exchange rate across the above 9 major economies in the German community detected in 2011. The average exchange rate of the German community was basically below that of all the WIOD economies before 2000. Only from 2001, the community average became slightly (no more than 16%) higher than the overall average. Therefore, the emergence of the German community cannot be attributed to inflation or exchange rate dynamics. Since most of the 40 economies in the WIOD are in Europe, we cannot rule out the possibility that similar regional input-output communities are emerging in other continents. Indeed, we find also an integrated NAFTA community in North America. However, since many Asian economies are not included in the WIOD, we cannot argue if a similar trend is ongoing in the Far East. Within each community, we also carry out the community core detection (see below for more technical details). In Figures \[fig3\]-\[fig5\], we identify the top 3 core economy-industry pairs for each community. The first place is with thick and solid border. The second place is with thick and dashed border. The third place is with border and texture. In general, the cores are mostly concentrated in the industries of agriculture (1), mining (2), food (3), metals (12), construction (18), and financial, business, and public services (28-31). Over time, while the services industries (28-31) have become the cores in more and more developed economies, the primary industries (1-3) have become less central in the developed economies and have only remained as the cores in a few emerging economies, which is consistent with the Kuznets facts [@kuznets1957quantitative; @kuznets1973modern]. Furthermore, for the growing community centered on Germany, the cores are always identified in Germany (that is why we simply call it the German community) for the three selected years. It is also worth noting that, the German industry of transport equipment (15) is identified as a core in 2011 and the car industry is the most integrated in the German community, which spans over 17 economies. The Network-Based Methods of Identifying the Key Industries ----------------------------------------------------------- Since on a global scale the traditional assumption of stable input-output technical coefficients is violated due to the dynamics of international trade, the traditional final-demand-weighted backward linkage measure alone is insufficient to evaluate the importance of any given industry on the global economy. However, the networks approach provides us a holistic view of the global production system and we can compute various centrality measures to compare the nodes in the network. Here we focus on two network-based methods of identifying the key industries in the WION, PageRank centrality[^23] and community coreness measure. ### PageRank Centrality Given a network, it is a problem of capital importance to bring order to its structure by ranking nodes according to their relevance. Among the many proposed, a successful and widely used centrality measure is PageRank [@page1999pagerank], a Google patented method. The idea is that the nodes are considered important if they are connected by other important nodes. Since the WION is weighted, we use a weighted version of PageRank, which is computed iteratively as follows: 1. At $t=0$, an initial probability distribution is assumed, usually $PR(i; 0) = \frac{1}{N}$ where $N$ is the total number of nodes; 2. At each time step, the PageRank of node $i$ is computed as: $$PR(i;t+1) = \frac{1-d}{N} + d \sum_{j \in M(i)} \frac{PR(j; t)w_{ij}}{S(j)}$$ where $M(i)$ are the in-neighbors of $i$, $w_{ij}$ is the weight of the link between the nodes $i$ and $j$, $S$ is the sum of the weights of the outgoing edges from $j$, and the damping factor $d$ is set to its default value, 0.85. In Table \[table\_2\], the second column of each year is produced by the PageRank centrality, which is denoted by $PR$.[^24] Unlike the final-demand-weighted backward linkage measure, where only 4 economies are among the top 20, the PageRank centrality recognizes 10 economies in the top 20 list for the three selected years. Tables \[table\_A5\] and \[table\_A6\] in the appendix provide an alternative way of viewing the key industries and economies over time identified by the PageRank centrality. In particular, Table \[table\_A5\] lists the most important economies by industry while Table \[table\_A6\] lists the most important industries by economy. ### Community Coreness Measure The other network-based method of identifying the key industries is the community coreness measure. Nodes of a community do not have the same importance for the community stability: the removal of a node in the core of the community affects the partition much more than the deletion of a node that stays on the periphery of the community [@deleo2013]. Therefore, in the following we define a novel way of detecting cores inside communities by using the properties of the modularity function \[eq3\]. By definition, if the modularity associated with a network has been optimized, every perturbation in the partition leads to a negative variation in the modularity, $\mathrm{d}Q$. If we move a node from its community, we have $M-1$ possible choices, with $M$ as the number of communities, as the node’s new host community. It is possible to define the $|\mathrm{d}Q|$ associated with each node as the smallest variation in absolute value (or the closest to 0 since $\mathrm{d}Q$ is always a negative number) of all the possible choices. We call $|\mathrm{d}Q|$ the community coreness measure. In the WION, once we have the $|\mathrm{d}Q|$ for each industry, we can consider the one with the biggest $|\mathrm{d}Q|$ the most important. We can also normalize the $|\mathrm{d}Q|$ to identify the most important nodes within each community. The results are shown in Figures \[fig3\], \[fig4\], and \[fig5\], where the first place in each community is with thick and solid border, the second place is with thick dashed border, and the third place is with both border and texture. In Table \[table\_2\], the third column of each year is produced by the community coreness measure, which is denoted again by $|\mathrm{d}Q|$. Interestingly, like the final-demand-weighted backward linkage measure, the community coreness measure also only includes China, Germany, Japan, and USA in the top 20 list for the selected years. Tables \[table\_A7\] and \[table\_A8\] in the appendix provide an alternative way of viewing the key industries and economies over time identified by the community coreness measure. In particular, Table \[table\_A7\] lists the most important economies by industry while Table \[table\_A8\] lists the most important industries by economy. Now we have totally three methods to identify the key industries in the WION, the traditional final-demand-weighted backward linkage measure, the PageRank centrality measure, and the community coreness measure. They have different results from each other. For instance, the industry of transport equipment in Germany is captured by the PageRank but not by the other two while the industry of other business activities in USA is more important by $|\mathrm{d}Q|$ than by the other two (see Table \[table\_2\]). Table \[table\_3\] reports the correlation coefficient matrix among the three methods for the selected years 1995, 2003, and 2011. We find that all the three methods are positively correlated, while $\boldsymbol{\mathrm{w}}$ and $|\mathrm{d}Q|$ are correlated even more. Therefore, the network-based $|\mathrm{d}Q|$ and especially $PR$ can be used to complement, if not to substitute, $\boldsymbol{\mathrm{w}}$ to identify the key industries in the WION. \[table\_3\] Concluding Remarks {#sec:conclusion} ================== This paper investigates a MRIO system characterized by the recently available WIOD database. By viewing the world input-output system as an interdependent network where the nodes are the individual industries in different economies and the edges are the monetary goods flows between industries, we study the network properties of the so-called world input-output network (WION) and document its evolution over time. We are able to quantify not only some global network properties such as assortativity, clustering coefficient, and degree and strength distributions, but also its subgraph structure and dynamics by using community detection techniques. Over time, we trace the effects of globalization and the 2008-2009 financial crisis. We notice that national economies are increasingly interconnected in global production chains. Moreover, we detect the emergence of regional input-output community. In particular we see the formation of a large European community led by Germany. Finally, because on a global scale the traditional assumption of stable input-output technical coefficients is violated due to the dynamics of international trade, we also use the network-based PageRank centrality and community coreness measure to identify the key industries in the WION and the results are different from the one obtained by the traditional final-demand-weighted backward linkage measure. As mentioned above, due to the limited coverage of the WIOD, we cannot argue if the input-output integration is also observed in other continents. Therefore, in our future work, we will utilize another MRIO database, EORA [@lenzen2012mapping; @lenzen2013building], which covers about 187 countries in the world and the years from 1990 to 2011. Moreover, since each of the three methods of identifying the key industries captures a different aspect of the importance of any given industry, future work is also needed to compare the methods so as to identify the systematically important industries for the global economy. Acknowledgments =============== Authors thank Michelangelo Puliga for insightful discussions. All authors acknowledge support from the FET projects MULTIPLEX 317532 and SIMPOL 610704 and the PNR project CRISIS Lab. MR and ZZ acknowledge funding from the MIUR (FIRB project RBFR12BA3Y). FC gratefully acknowledges Sardinia Regional Government for the financial support of her PhD scholarship (P.O.R. Sardegna F.S.E. Operational Programme of the Autonomous Region of Sardinia, European Social Fund 2007–2013 - Axis IV Human Resources, Objective l.3, Line of Activity l.3.1.). **APPENDIX** \[table\_A1\] \[table\_A2\] \ \[table\_A3\] \[table\_A4\] \[table\_A5\] \[table\_A6\] \[table\_A7\] \[table\_A8\] [^1]: Ever since Leontief [@leontief1936quantitative] formalized its structure, the input-output table has been used extensively by economists, environmentalists, and policy makers alike. By keeping track of the inter-industrial relationships, the input-output table offers a reasonably accurate measurement of the response of any given economy in the face of external shocks or policy interventions. [^2]: Tukker and Dietzenbacher [@tukker2013global] summarize the recent development of the MRIO databases. [^3]: More precisely, the edges are the monetary goods and services flows. The direction of the flows go from the seller industry to the buyer industry. They are monetary because they are denoted in current US dollars. [^4]: In contrast, due to the low-digit industry classification, the input-output networks at the national level are almost complete [@blochl2011vertex]. [^5]: This is also due to the aggregated industry classification. [^6]: Carvalho [@carvalho2013survey] also use a networks approach to study the WIOD data. But he only uses a single year (2006) and considers it as an unweighted network. [^7]: Tables \[table\_A1\] and \[table\_A2\] in the appendix have the lists of countries and industries covered in the WIOD. [^8]: Again, the relationship can also be an industry to itself and within the same economy. [^9]: The basic prices are also called the producers’ prices, which represent the amount receivable by the producers. An alternative is the purchases’ prices, which represent the amount paid by the purchases and often include trade and transport margins. The former is preferred by the WIOD because it better reflects the cost structures underlying the industries [@timmer2012world]. [^10]: Here we only show the aggregated final demand for the two economies [^11]: The ratios are called technical coefficients because they represent the technologies employed by the industries to transform inputs into outputs. [^12]: The Leontief inverse is demand-driven, i.e., a repercussion effect triggered by an increase in final demand. Another strand of the input-output economics literature is based on the supply-driven model, where a repercussion effect is triggered by an increase in value added (primary inputs) [@ghosh1958input; @miller2009input]. [^13]: It is backward because the linkage is identified by tracing back to the upstream industries. [^14]: Strictly speaking, the flows from the red nodes to the blue nodes are not goods but primary inputs in nature. [^15]: In the case of WION, all the coefficients are of very small magnitude less than 0.06. [^16]: Notice that when calculating the assortativity for in-degree, out-degree, and total-degree, respectively, we consider the nodes as the neighbors of a given node if they are connected with the given node by only incoming edges, by only outgoing edges, and by either incoming or outgoing edges, respectively. In contrast, Carvalho [@carvalho2013survey] defines the neighborhood solely on the basis of the incoming edges and finds a positive assortative relationship. [^17]: While the most severely depressed domestic edges during 2008-2009 in terms of the magnitude of the reduced flows are mostly within USA, the top 3 most impacted foreign edges are all from the mining industry to the coke and fuel industry and are from Canada to USA, from Netherlands to Belgium, and from Mexico to USA, respectively. [^18]: The same feature is also found in the input-output networks at the national level [@mcnerney2013network]. Using a single-year (2006) data of the WIOD, Carvalho [@carvalho2013survey] also reports the heavy-tailed but non-power-law degree distributions. [^19]: We perform the community detection for all available years (1995-2011). Results are available upon request. [^20]: It is centered on Germany because the community core detection results below show that the cores of this red community are all within Germany. [^21]: The data is unavailable for Taiwan. We also exclude Bulgaria because it had a hyperinflation in 1997, which will bias the average if included. The data source is the World Development Indicators, the World Bank, <http://data.worldbank.org/indicator/NY.GDP.DEFL.KD.ZG>. [^22]: The data source is the exchange rate data used in the WIOD, <http://www.wiod.org/protected3/data/update_sep12/EXR_WIOD_Sep12.xlsx>. [^23]: We choose PageRank over other centrality measures such as closeness and betweenness because the former systematically measures the influence of a given node and has been widely used in the previous literature to identify the key nodes [@acemoglu2012network; @carvalho2013survey]. [^24]: Our PageRank result differs from the one reported by Carvalho [@carvalho2013survey], where he uses an unweighted version of PageRank.

ArXiv

--- abstract: | We prove that the symmetric weak GARCH limit is a geometric mean-reverting stochastic volatility process with diffusion determined by kurtosis of physical log returns; this provides an improved fit to implied volatility surfaces. When log returns are normal the limit coincides with Nelson’s limit. The limit is unique, unlike strong GARCH limits, because assumptions about convergence of model parameters is unnecessary – parameter convergence is uniquely determined by time-aggregation of the weak GARCH process.\ author: - 'Carol Alexander[^1] $\,$ and Emese Lazar[^2]' date: title: On the Continuous Limit of Weak GARCH --- Keywords: GARCH, stochastic volatility, time aggregation, diffusion limit.\ JEL Classification Codes: C32, G13.\ Introduction ============ The symmetric weak GARCH process introduced by Drost and Nijman (1993) is characterised by the absence of parametric conditional distributions of the errors, with the familiar autoregressive equation being defined for the best linear predictor of the residuals rather than the conditional variance. It is the only class of GARCH process which satisfies the time-aggregation property (i.e. doubling or halving the sampling frequency doesn’t change the class, it remains a weak GARCH process). Models that do not satisfy the time-aggregation property may not have unique limit. Nelson (1990) derived a limit of the strong symmetric normal GARCH model as a stochastic variance process with independent Brownians. This is fundamental for limits of other GARCH processes – see Lindner (2009) for a brief overview. But, like other strong GARCH limits, Nelson’s limit is not unique because it is necessary to make assumptions about the convergence rate of the parameters, and different assumptions lead to different limits which may be either stochastic or deterministic – see Corradi (2000). Also, because strong GARCH is not time-aggregating, the discretized version of the continuous limit not only depends on the frequency, it may not even be a GARCH process. Drost and Werker (1996) introduce continuous-time symmetric GARCH diffusion and jump-diffusion processes that exhibit weak GARCH-type behaviour at all discrete frequencies, showing that their characterisation also depends on the kurtosis of the observed discrete time data. Building on this, here we show that the continuous limit of a symmetric weak GARCH process is a stochastic volatility model similar to Nelson’s limit but the diffusion coefficient is related to the kurtosis of the distribution of log returns. This endows the limit of weak GARCH with an additional parameter and simulations demonstrate how this can improve the fit to implied volatilities when calibrated in the risk-neutral measure. If log returns are normally distributed our limit reduces to Nelson’s strong GARCH diffusion. We also prove that there is no ambiguity about parameter convergence, it follows uniquely and directly from the definition of the weak GARCH process so that the limit is unique. Therefore, within the class of weak GARCH processes, where time aggregation prevails, knowledge of the discrete time GARCH parameters at only one frequency, and knowledge of the kurtosis, completely determines the coefficients of the continuous GARCH process. So, in estimating a continuous time GARCH process in the physical measure it suffices to estimate the discrete time GARCH parameters for the available data frequency. The Weak GARCH Process ====================== Following Engle (1982) and Bollerslev (1986) the GARCH(1,1) process for a log return ${y_t} $ can be written as: ${y_t} = \mu + {\varepsilon_t}\quad$with $E\left( {\left. {{\varepsilon _{t + 1}}} \right|{I_t}} \right) = 0, $ where ${I_t} $ is the ${\sigma}$-algebra generated by the residual vector $\left( {{\varepsilon_t}} \right) $. The classical or strong GARCH definition states: $$\label{eq1} E\left( {\left. {\varepsilon _{t + 1}^2} \right|{I_t}} \right) = {h_t},$$ where ${h_t} $ is the conditional variance. Now in the symmetric version of both strong and weak GARCH, we assume ${h_t} = \omega + \alpha \varepsilon_t^2 + \beta {h_{t - 1}} $. But in the weak GARCH process (Drost and Nijman, 1993) $h_t$ is the best linear predictor (BLP) of the squared residuals, not the conditional variance, replacing with: $$E\left( {{\varepsilon _{t + 1}}\varepsilon _{t - i}^r} \right) = 0\quad i \ge 0\quad r = 0,1,2; \qquad E\left( {\left( {\varepsilon _{t + 1}^2 - {h_t}} \right)\varepsilon _{t - i}^r} \right) = 0\quad i \ge 0\quad r = 0,1,2.$$ The assumption that $0$ and ${h_t} $ are the BLPs for the residuals and squared residuals respectively, guarantees that the BLP of the squared residuals aggregates in time, but only for symmetric processes. For a finite step-length $\Delta$ we consider the $\Delta$-step process for the residuals and the GARCH process. Time is indexed as $k\Delta$, with $k=1,2, \ldots$ and we use a pre-subscript for the time step and, to be able to compare variances for different step-lengths, we divide by the step-length. Thus ${}_\Delta {h_{k\Delta }} $ denotes the BLP for ${\Delta ^{ - 1}}{}_\Delta \varepsilon _{k\Delta }^2 $. Using ${}_\Delta \lambda = {}_\Delta \alpha + {}_\Delta \beta $, for $ i \ge 0$ and $r = 0,1,2$ the annualised weak GARCH process may be written: $$\begin{aligned} \label{eq2} {}_\Delta {y_{k\Delta }} = \Delta \mu + {}_\Delta {\varepsilon _{k\Delta }},\qquad {}_\Delta {h_{k\Delta }} = {}_\Delta \omega + {}_\Delta \alpha {\Delta ^{ - 1}}\varepsilon _{k\Delta }^2 + {}_\Delta \beta {}_\Delta {h_{\left( {k - 1} \right)\Delta }},\\ \nonumber E\left( {{}_\Delta {\varepsilon _{\left( {k + 1} \right)\Delta }}{}_\Delta \varepsilon _{\left( {k - i} \right)\Delta }^r} \right) = 0, \qquad E\left( {\left( {{\Delta ^{ - 1}}{}_\Delta \varepsilon _{\left( {k + 1} \right)\Delta }^2 - {}_\Delta {h_{k\Delta }}} \right){}_\Delta \varepsilon _{\left( {k - i} \right)\Delta }^r} \right) = 0.\end{aligned}$$ The first paper that discusses the continuous limit of GARCH is that of Nelson (1990). Under the conditions: $$\omega = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}{}_\Delta \omega } \right){\rm{;}}\quad \alpha = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1/2}}{}_\Delta \alpha } \right){\rm{;}}\quad \theta = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right)} \right){\rm{;}}\quad 0 < \omega ,\alpha ,\theta < \infty$$ the limit will be a stochastic volatility model with independent Brownians, i.e. $dS_t = \mu \,S_t dt + \sqrt{V_t}\,S_t dB_{1t}$ with $dV_t = \left( {\omega - \theta V_t} \right)dt + \sqrt 2 \alpha V_t dB_{2t}$ where $V_t$ is the continuous-time limit of $h_t$. On the other hand, Corradi (2000) proves that, if we assume the following convergence rates: $$\omega = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}{}_\Delta \omega } \right){\rm{;}}\quad \alpha = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}{}_\Delta \alpha } \right){\rm{;}}\quad \theta = \mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right)} \right){\rm{;}}\quad 0 < \omega ,\alpha ,\theta < \infty$$ then the continuous-time limit is a deterministic variance model with the same price dynamics but with $ dV_t = \left( {\omega - \theta V_t} \right)dt $. The difference between the two assumptions lies with the convergence of alpha (at rate $\sqrt \Delta $ versus rate $\Delta$). Which assumption is correct has been the subject of considerable debate. Here we argue that the asssumptions of Nelson are correct, but we promote a different continuous limit because it is best to use the time aggregating model. Without time aggregation we have a strong GARCH process for a given frequency, but for any other frequencies the process will not be a strong GARCH process anymore. For a weak GARCH using step-lengths $\Delta$ and $\delta$, $\delta < \Delta$, Drost and Nijman (1993) proved the following relationship between the parameters: $${}_\Delta \omega = {}_\delta \omega \left( {1 - {{\left( {{}_\delta \lambda } \right)}^{{\delta ^{ - 1}}\Delta }}} \right){\left( {1 - {}_\delta \lambda } \right)^{ - 1}} \quad \mbox{and} \quad {}_\Delta \alpha = {\left( {{}_\delta \lambda } \right)^{{\delta ^{ - 1}}\Delta }} - {}_\Delta \beta.$$ The relationship between the unconditional kurtosis coefficients, denoted ${}_\Delta \kappa $ and ${}_\delta \kappa $ respectively, is: $$\label{eqkurt} {}_\Delta \kappa = 3 + {\Delta ^{ - 1}}\delta \left( {{}_\delta \kappa - 3} \right) + 6\left( {{}_\delta \kappa - 1} \right)\frac{{\left( {{\delta ^{ - 1}}\Delta \left( {1 - {}_\delta \lambda } \right) - \left( {1 - {}_\delta {\lambda ^{{\delta ^{ - 1}}\Delta }}} \right)} \right){}_\delta \alpha \left( {1 - {}_\delta {\lambda ^2} + {}_\delta \alpha {}_\delta \lambda } \right)}}{{{{\left( {{\delta ^{ - 1}}\Delta } \right)}^2}{{\left( {1 - {}_\delta \lambda } \right)}^2}\left( {1 - {}_\delta {\lambda ^2} + {}_\delta {\alpha ^2}} \right)}}.$$ Drost and Nijman (1993) derive the following relationship between the low and high frequency parameters: $${}_\Delta \beta {\left( {1 + {}_\Delta {\beta ^2}} \right)^{ - 1}} = \left( {{}_{\Delta ,\delta }c {}_\delta \lambda {{\kern 1pt} ^{{\delta ^{ - 1}}\Delta }} - 1} \right){\left( {{}_{\Delta ,\delta }c\left( {1 + {}_\delta \lambda {{\kern 1pt} ^{2{\delta ^{ - 1}}\Delta }}} \right) - 2{\kern 1pt} } \right)^{ - 1}}$$ where $$\begin{aligned} \label{eqc} {}_{\Delta ,\delta }c = \left[ \begin{array}{l} {\delta ^{ - 1}}\Delta {\left( {1 - {}_\delta \beta } \right)^2} + 2{\delta ^{ - 1}}\Delta \left( {{\delta ^{ - 1}}\Delta - 1} \right)\left( {1 - {}_\delta \lambda } \right)\left( {1 - {}_\delta \lambda {{\kern 1pt} ^2} + {}_\delta {\alpha ^2}} \right){\left( {{}_\delta \kappa - 1} \right)^{ - 1}}{\left( {1 + {}_\delta \lambda {\kern 1pt} } \right)^{ - 1}}\\ \nonumber {\rm{ }} + 4\left( {{\delta ^{ - 1}}\Delta \left( {1 - {}_\delta \lambda {\kern 1pt} } \right) - \left( {1 - {}_\delta \lambda {{\kern 1pt} ^{{\delta ^{ - 1}}\Delta }}} \right)} \right){}_\delta \alpha \left( {1 - {}_\delta \beta {}_\delta \lambda {\kern 1pt} } \right){\left( {1 - {}_\delta \lambda {{\kern 1pt} ^2}} \right)^{ - 1}} \end{array} \right] \times \\ \quad \quad {\left[ {{}_\delta \alpha \left( {1 - {}_\delta \beta {}_\delta \lambda } \right)\left( {1 - {}_\delta \lambda {{\kern 1pt} ^{2{\delta ^{ - 1}}\Delta }}} \right){{\left( {1 - {}_\delta \lambda {{\kern 1pt} ^2}} \right)}^{ - 1}}} \right]^{ - 1}}\end{aligned}$$ To derive the continuous limit of this model we are interested in the inverse relationship: expressing the high frequency ($\delta$-step) parameters and their limit based on the low frequency ($\Delta$-step) parameters, for $\delta < \Delta$: $${}_\delta \omega = {}_\Delta \omega \left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right){\left( {1 - {}_\Delta \lambda } \right)^{ - 1}} \quad \mbox{and} \quad {}_\delta {\lambda ^{{\delta ^{ - 1}}}} = {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}}}$$ Also: $$\left( {2\left( {\frac{{{}_\Delta \beta }}{{1 + {}_\Delta {\beta ^2}}}} \right) - 1} \right){}_\delta \alpha \left( {1 - {}_\delta \beta {}_\delta \lambda } \right)\left( {\frac{{1 - {}_\Delta {\lambda ^2}}}{{1 - {}_\delta {\lambda ^2}}}} \right) =$$ $$\label{eq3} = \left( {\left( {\frac{{{}_\Delta \beta }}{{1 + {}_\Delta {\beta ^2}}}} \right)\left( {1 + {}_\Delta {\lambda ^2}} \right) - {}_\Delta \lambda } \right)\left( \begin{array}{l} \Delta {\delta ^{ - 1}}{\left( {1 - {}_\delta \beta } \right)^2} + 2\Delta {\delta ^{ - 1}}\left( {\Delta {\delta ^{ - 1}} - 1} \right){\left( {{}_\delta \kappa - 1} \right)^{ - 1}}\left( {\frac{{1 - {}_\delta \lambda }}{{1 + {}_\delta \lambda }}} \right)\left( {1 - {}_\delta {\lambda ^2} + {}_\delta {\alpha ^2}} \right)\\ + 4{\left( {1 - {}_\delta {\lambda ^2}} \right)^{ - 1}}{}_\delta \alpha \left( {\Delta {\delta ^{ - 1}}\left( {1 - {}_\delta \lambda } \right) - \left( {1 - {}_\Delta \lambda } \right)} \right)\left( {1 - {}_\delta \beta {}_\delta \lambda } \right) \end{array} \right)$$ and $$\label{eq4} {}_\delta \kappa = 1 + \left( {\frac{{\left( {2 + {\delta ^{ - 1}}\Delta \left( {{}_\Delta \kappa - 3} \right)} \right)\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{2{\Delta ^{ - 1}}\delta }} + {}_\delta {\alpha ^2}} \right)}}{{\left( {\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{2{\Delta ^{ - 1}}\delta }} + {}_\delta {\alpha ^2}} \right)} \right) + 6{}_\delta \alpha \left( {1 - {\Delta ^{ - 1}}\delta \left( {1 - {}_\Delta \lambda {\kern 1pt} } \right){{\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)}^{ - 1}}} \right)\left( {1 - {}_\delta \beta {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)}}} \right) .$$ Continuous Limit of Weak GARCH ============================== The continuous time limit of an econometric model may not offer equivalence with the discrete time model. For equivalence, it must be that the discretization of the continuous limit yields the same discrete time model as the original. Furthermore, the discretized model must be the same for all frequencies. Clearly, this cannot happen if the discrete model does not aggregate in time. Thus, it is only when (1) the original discrete time model is time aggregating, and (2) the model can be discretized at any frequency in the form of the original model, that we have an equivalence between discrete and continuous models. The first step for deriving the continuous limit of symmetric weak GARCH is to determine the limits and convergence speeds of the parameters. In contrast to the strong GARCH process, where there is some freedom to choose assumptions about parameter convergence speeds, we now find that it is not needed to make any assumption about parameter convergence. Instead, the time-aggregation property of weak GARCH implies unique convergence speeds for all parameters, as the following proposition shows:\ **Proposition 1:** The convergence rates for the parameters implied by the weak GARCH model are as follows: $$\omega = \mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1}}{}_\Delta \omega ;\quad \alpha = \mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1/2}}{}_\Delta \alpha ;\quad \theta = \mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right);\quad 0 < \omega ,\alpha ,\theta < \infty$$ Also, the unconditional kurtosis converges to $\kappa = \mathop {\lim }\limits_{\Delta \downarrow 0} {}_\Delta \kappa = 3{\left( {1 - {\theta ^{ - 1}}{\alpha ^2}} \right)^{ - 1}} .$\ **Proof:** We get ${}_\delta {\lambda ^{{\delta ^{ - 1}}}} = {}_\Delta {\lambda ^{{\Delta ^{ - 1}}}} $, which is a constant between 0 and 1 denoted $\exp \left( { - \theta } \right)$ with $\theta > 0.$ Thus $${}_\Delta \lambda = \exp \left( { - \theta \Delta } \right) \quad \mbox{and} \quad \mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right) = \mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1}}\left( {1 - \exp \left( { - \theta \Delta } \right)} \right) = \theta .$$ Also ${}_\delta \omega {\left( {1 - {}_\delta \lambda } \right)^{ - 1}} = {}_\Delta \omega {\left( {1 - {}_\Delta \lambda } \right)^{ - 1}} $ is a positive constant denoted $\omega {\theta ^{ - 1}} $, $\omega >0$ and ${}_\Delta \omega = \omega {\theta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right) $, so $$\mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{\Delta ^{ - 1}}{}_\Delta \omega } \right) = \omega {\theta ^{ - 1}}\mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1}}\left( {1 - \exp \left( { - \theta \Delta } \right)} \right) = \omega .$$ The formula for kurtosis may now be written: $$\label{eq5} {}_\delta \kappa = 1 + \left( {{}_\Delta \kappa - 3 + 2{\Delta ^{ - 1}}\delta } \right){\left( {{\Delta ^{ - 1}}\delta + 6\frac{{\left( {{\delta ^{ - 1}}\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right) - {\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda {\kern 1pt} } \right)} \right)}}{{\Delta {{\left( {{\delta ^{ - 1}}\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)} \right)}^2}}}{}_{\Delta ,\delta }A} \right)^{ - 1}} , $$ with $${}_{\Delta ,\delta }A = \frac{{{}_\delta \alpha {\delta ^{ - 1}}\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{2{\Delta ^{ - 1}}\delta }}} \right) + {\delta ^{ - 1}}{}_\delta {\alpha ^2} - {\delta ^{ - 1}}{}_\delta {\alpha ^2}\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{{\Delta ^{ - 1}}\delta }}} \right)}}{{{\delta ^{ - 1}}\left( {1 - {}_\Delta \lambda {{\kern 1pt} ^{2{\Delta ^{ - 1}}\delta }}} \right) + {\delta ^{ - 1}}{}_\delta {\alpha ^2}}} .$$ But $$\mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - 1}}\left( {1 - {}_\Delta {\lambda ^{{\Delta ^{ - 1}}\delta }}} \right) = \theta \quad \mbox{and} \quad \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - 1}}\left( {1 - {}_\Delta {\lambda ^{2{\Delta ^{ - 1}}\delta }}} \right) = 2\theta .$$ Thus, using $_\delta \alpha \downarrow 0,$ we have $$\mathop {\lim }\limits_{\delta \downarrow 0} {}_{\Delta ,\delta }A = {\left( {2\theta {{\left( {\mathop {\lim }\limits_{\delta \downarrow 0} \left( {{\delta ^{ - 1}}{}_\delta {\alpha ^2}} \right)} \right)}^{ - 1}} + 1} \right)^{ - 1}} .$$ Hence, taking the limit of as $\delta \downarrow 0$ and then $\Delta \downarrow 0$ yields $\kappa = 3{\left( {1 - {\theta ^{ - 1}}\mathop {\lim }\limits_{\delta \downarrow 0} \left( {{\delta ^{ - 1}}{}_\delta {\alpha ^2}} \right)} \right)^{ - 1}} $. The limit of the unconditional kurtosis is finite and positive, which forces $0 \le \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - 1}}{}_\delta {\alpha ^2} < \theta $, so the kurtosis will be higher than 3. To see the speed of convergence for $\alpha$ , we consider the limit $\alpha : = \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - w}}{}_\delta \alpha $ with $\alpha \in \left( {0,\infty } \right) $ with $w$ unknown. Since $\mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - 1}}{}_\delta {\alpha ^2} < \theta $, $w \ge 1/2$, for $y = \min \left( {w,1} \right)$ and $ z = \min \left( {2w,1} \right) $ we write $$\begin{aligned} \label{eq6} \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}\left( {1 - {}_\delta \lambda + {}_\delta \alpha } \right) = \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}\left( {1 - {}_\delta \lambda } \right) + \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}{}_\delta \alpha \in \left( {0,\infty } \right)\\ \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}\left( {1 - {}_\delta {\lambda ^2} - {}_\delta \alpha \left( {1 - {}_\delta \lambda } \right) + {}_\delta \alpha } \right) = \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}\left( {1 - {}_\delta {\lambda ^2}} \right) + \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - y}}{}_\delta \alpha \in \left( {0,\infty } \right)\\ \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - z}}\left( {1 - {}_\delta {\lambda ^2} + {}_\delta {\alpha ^2}} \right) = \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - z}}\left( {1 - {}_\delta {\lambda ^2}} \right) + \mathop {\lim }\limits_{\delta \downarrow 0} {\delta ^{ - z}}{}_\delta {\alpha ^2} \in \left( {0,\infty } \right)\end{aligned}$$ Also, using and noting that ${}_\delta \kappa \ne 1 $, since ${}_\delta {\alpha ^2} > 0 $, we can compute $$\left( {2\left( {\frac{{{}_\Delta \beta }}{{1 + {}_\Delta {\beta ^2}}}} \right) - 1} \right){}_\delta \alpha \left( {1 - {}_\delta {\lambda ^2} - {}_\delta \alpha \left( {1 - {}_\delta \lambda } \right) + {}_\delta \alpha } \right)\left( {\frac{{1 - {}_\Delta {\lambda ^2}}}{{1 - {}_\delta {\lambda ^2}}}} \right).$$ If $w > 1/2$, we can multiply the above expression by ${\delta ^{1 - w - y}} $ and then computing the limit as $\delta$ tends to zero leads to a contradiction in terms of limits. So we must have $w = 1/2$ and this sets the convergence of $\alpha$.\ Now consider the conditional variance and the conditional kurtosis of the residuals; where the conditional mean and skewness are equal with zero: $${}_\Delta \sigma _{k\Delta }^2 = E\left( {\left. {{\Delta ^{ - 1}}{{\left( {{}_\Delta {\varepsilon _{\left( {k + 1} \right)\Delta }} - \Delta {}_\Delta {\mu _{k\Delta }}} \right)}^2}} \right|{}_\Delta {I_{k\Delta }}} \right)$$ $${}_\Delta {\kappa _{k\Delta }} = E\left( {\left. {{\Delta ^{ - 2}}{}_\Delta \sigma _{k\Delta }^{ - 4}{{\left( {{}_\Delta {\varepsilon _{\left( {k + 1} \right)\Delta }} - \Delta {}_\Delta {\mu _{k\Delta }}} \right)}^4}} \right|{}_\Delta {I_{k\Delta }}} \right)$$ where ${}_\Delta {I_{k\Delta }} $ is the $\sigma$-algebra generated by the vector $\left( {{}_\Delta {\varepsilon _{k\Delta }}} \right).$ We divide by $\Delta$ when computing the conditional variance series so that the variance over $\Delta$ is comparable with $\Delta$ times the 1-step variance. The conditional expectation of the second moment and the kurtosis must be positive, and we shall assume that the following limits exist for $k\Delta\le t < (k+1)\Delta$: $V\left( t \right): = \mathop {\lim }\limits_{\Delta \downarrow 0} {_\Delta }{h_t} $ where $_\Delta {h_t}: = {}_\Delta {h_{k\Delta }} $, $\mu \left( t \right) = \mu $ and $\kappa \left( t \right): = \mathop {\lim }\limits_{\Delta \downarrow 0} {_\Delta }{\kappa_t} $ where $_\Delta {\kappa_t}: = {}_\Delta {\kappa _{k\Delta }}.$ Due to the symmetrical nature of the returns, we can write $$\label{eq13} E\left( {\left. {{\Delta ^{ - 1}}{}_\Delta \varepsilon _{\left( {k + 1} \right)\Delta }^2} \right|{I_{k\Delta }}} \right) = {}_\Delta \sigma _{k\Delta }^2.$$ Note that $ {}_\Delta \sigma _{k\Delta }^2 - {}_\Delta {h_{k\Delta }} $ has to be different from zero, otherwise the process will be a semi-strong GARCH, However, we assume that as the time step decreases, the difference between the conditional variance and the BLP of the squared residuals converges to zero at a speed of square root of the time step, i.e. $\mathop {\lim }\limits_{\Delta \downarrow 0} {\Delta ^{ - 1/2}}\left( {{}_\Delta \sigma_t^2 - {}_\Delta {h_t}} \right) = 0. $ In other words, the BLP of the squared residuals is ‘close’ to the conditional variance process. This is the only assumption we make and we consider that it is non-binding because as the time step decreases, the BLP process becomes more and more informative and so it converges fast to the conditional variance, i.e. $V\left( t \right) = \mathop {\lim }\limits_{\Delta \downarrow 0} {_\Delta }\sigma_t^2 $ where $_\Delta \sigma_t^2: = {}_\Delta \sigma _{k\Delta }^2 $ for $k\Delta\le t < (k+1)\Delta, $ so that $\mathop {\lim }\limits_{\Delta \downarrow 0} \left( {{}_\Delta \sigma_t^2 - {}_\Delta {h_t}} \right) = 0 $ as well.\ **Theorem 1:** The continuous time limit of the weak GARCH process defined in is the following stochastic volatility model, based on the limiting parameters given above and in Proposition 1: $$\begin{aligned} \frac{dS_t}{S_t} &=& \mu dt + \sqrt{V_t}\,dB_{1t},\\ dV_t &=& \left( \omega - \theta V_t \right)dt + \alpha \sqrt{\left( {\kappa_t - 1} \right)}\,V_t \, dB_{2t},\end{aligned}$$ where $B_{1t}$ and $B_{2t} $ are independent Brownian motions.\ **Proof:** We employ the convergence theorem for stochastic difference equations to stochastic differential equations given by Nelson (1990). For the returns process we have: $$E\left( {{\Delta ^{ - 1}}\left. {{y_{\left( {k + 1} \right)\Delta }}} \right|{I_{k\Delta }}} \right) = \mu + E\left( {{\Delta ^{ - 1}}{\varepsilon _{\left( {k + 1} \right)\Delta }}|{I_{k\Delta }}} \right) = \mu .$$ And, using it can be shown that: $E\left( {{\Delta ^{ - 1}}\left. {y_{\left( {k + 1} \right)\Delta }^2} \right|{I_{k\Delta }}} \right) = {h_{k\Delta }} + o\left( 1 \right), $ $$E\left( {{\Delta ^{ - 1}}\left. {\left( {{h_{\left( {k + 1} \right)\Delta }} - {h_{k\Delta }}} \right)} \right|{I_{k\Delta }}} \right) = {\Delta ^{ - 1}}{}_\Delta \omega - {\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right){h_{k\Delta }} + \left( {{\Delta ^{ - 1/2}}{}_\Delta \alpha } \right){\Delta ^{ - 1/2}}\left( {\sigma _{k\Delta }^2 - {h_{k\Delta }}} \right) + o\left( 1 \right)$$ and this converges to $\omega - \theta V_t $ by Proposition 1. The variance of the variance component is: $$E\left( {\left. {{\Delta ^{ - 1}}{{\left( {{h_{\left( {k + 1} \right)h}} - {h_{k\Delta }}} \right)}^2}} \right|{I_{k\Delta }}} \right) = {\Delta ^{ - 1}}{}_\Delta {\alpha ^2}\left( {E\left( {\left. {\left( {{}_\Delta \sigma _{k\Delta }^4} \right){{\left( {{\Delta ^2}{}_\Delta \sigma _{k\Delta }^4} \right)}^{ - 1}}\left( {\varepsilon _{\left( {k + 1} \right)\Delta }^4} \right) - h_{k\Delta }^2} \right|{I_{k\Delta }}} \right)} \right) + o\left( 1 \right).$$ The covariance between the returns and the changes in the variances converges as follows: $$E\left( {{\Delta ^{ - 1}}\left. {{y_{\left( {k + 1} \right)\Delta }}\left( {{h_{\left( {k + 1} \right)\Delta }} - {h_{k\Delta }}} \right)} \right|{I_{k\Delta }}} \right) = o\left( 1 \right).$$ Therefore, the limits of the expected squared terms and cross-product derived above define the following covariance matrix of the continuous process: $$A_t = \left( {\begin{array}{*{20}{c}} {V_t}&0\\ 0&{{\alpha ^2}\left( {\kappa_t - 1} \right)V{_t^2}} \end{array}} \right) .\qed$$ Discrete-time weak GARCH processes are characterized by (i) the existence of a long-term volatility; (ii) mean reversion in the variance process; (iii) the variance is stochastic, i.e. it has a non-zero variance; and (iv) the variance process is uncorrelated with the returns process, which is an implication of the symmetry of the returns’ distribution, being a requirement of weak GARCH processes. All these properties are also present in the continuous limit above; in addition, in our limit model the variance has a higher variance as compared to the limit of Nelson (1990), which results in extra kurtosis, and can be time-varying. These properties are intuitive and parallel the observed behaviour of implied volatilities in the risk neutral measure: see for example, Bates (1997, 2000) and Bakshi et al. (2003). Note that the limit process reduces to the diffusion derived by Nelson (1990) if $\kappa = 3 $, and in this case we obtain the smallest value of the volatility of the variance process, i.e. ${2^{1/2}}\alpha V_t $. Drost and Werker (1996) postulate that the conditional kurtosis is independent of $t$. However, in our limit this is allowed to be time-varying. Finally, we show that there is a discretization of the continuous limit under which the original GARCH model is returned when the series of returns and variances are discretized as follows:\ **Discretization Scheme:**\ $dt \mapsto \Delta $, $\frac{{dS_t}}{{S_t}} \mapsto {}_\Delta {y_{\left( {k + 1} \right)\Delta }}$, $V_t\mapsto {}_\Delta {V_{k\Delta }}$ and $dV_t \mapsto {}_\Delta {V_{\left( {k + 1} \right)\Delta }} - {}_\Delta {V_{k\Delta }} .$ The parameter $\mu$ stays unchanged during discretization. The rest of the parameters are discretized as: $ \omega \mapsto {\Delta ^{ - 1}}{}_\Delta \omega$, $\theta \mapsto {\Delta ^{ - 1}}\left( {1 - {}_\Delta \lambda } \right)$ and $\alpha \mapsto {\Delta ^{ - 1/2}}{}_\Delta \alpha $ where we specify the parameters $\left( {{}_\Delta \omega ,{}_\Delta \alpha ,{}_\Delta \beta } \right) $ in terms of the parameters $\left( {\omega ,\theta ,\alpha } \right) $ of the continuous model: ${}_\Delta \omega = \omega {\theta ^{ - 1}}\left( {1 - \exp \left( { - \theta \Delta } \right)} \right) $ , ${}_\Delta \lambda = \exp \left( { - \theta \Delta } \right) $ , ${}_\Delta \alpha = {}_\Delta \lambda - {}_\Delta \beta $ and $$\label{eqbeta} {}_\Delta \beta = \frac{1}{2} \frac{{{}_\Delta c\left( {1 + \exp \left( { - 2\theta \Delta } \right)} \right) - 2 + \left( {1 - \exp \left( { - \theta \Delta } \right)} \right){{\left( {{}_\Delta {c^2}{{\left( {1 - \exp \left( { - \theta \Delta } \right)} \right)}^2} - 4{}_\Delta c} \right)}^{1/2}}}}{{{}_\Delta c\exp \left( { - \theta \Delta } \right) - 1}}$$ where $$\label{eqc2} {}_\Delta c = \left[ \Delta \alpha ^2 + 2\alpha ^2\left( \Delta - \theta ^{- 1}\left( 1 - \exp \left( - \theta \Delta \right) \right) \right) + \Delta ^2\theta \left(\theta - \alpha ^2 \right) \right] \times \left[\frac{1}{2}\alpha ^2 \theta ^{-1} \left( 1 - \exp \left( - 2\theta \Delta \right)\right) \right]^{ - 1}.$$ The unconditional kurtosis is discretized as: $${}_\Delta \kappa = 3 + 6\left( {\kappa - 1} \right){\alpha ^2}\left( {\theta \Delta - \left( {1 - \exp \left( { - \theta \Delta } \right)} \right)} \right){\theta ^{ - 2}}{\Delta ^{ - 2}}{\left( {{\alpha ^2} + 2\theta } \right)^{ - 1}}$$ whilst the conditional kurtosis is discretized using $\kappa_t \to {}_\Delta {\kappa _{k\Delta }} = \kappa \left( {k\Delta } \right)\quad {\rm{for}}\quad k\Delta \le t < \left( {k + 1} \right)\Delta . $ The Brownian motions $B_1$ and $B_2$ that drive the price and variance equations are discretized by assuming a time step of length $\Delta$, and we can express the changes in the Brownian motions at time $t = k\Delta$ as: $$\begin{aligned} {B_1}\left( {\left( {k + 1} \right)\Delta } \right) - {B_1}\left( {k\Delta } \right) = {\Delta ^{1/2}}{}_\Delta {\xi _{\left( {k + 1} \right)\Delta }},\\ {B_2}\left( {\left( {k + 1} \right)\Delta } \right) - {B_2}\left( {k\Delta } \right) = {\Delta ^{1/2}}{}_\Delta {\eta _{\left( {k + 1} \right)\Delta }},\end{aligned}$$ where ${}_\Delta {\xi _{\left( {k + 1} \right)\Delta }} $is a standard normal variable, ${}_\Delta {\xi _{\left( {k + 1} \right)\Delta }} \, | \, {}_\Delta I_{k\Delta } \sim N\left( {0,1} \right)$ and ${}_\Delta {\eta _{\left( {k + 1} \right)\Delta }} $is defined as $$\label{eqeta} {}_\Delta {\eta _{\left( {k + 1} \right)\Delta }} = {2^{ - 1/2}}\left( {{}_\Delta \xi _{\left( {k + 1} \right)\Delta }^2 - 1} \right) .$$ Now define the normal variable ${}_\Delta {\varepsilon _{\left( {k + 1} \right)\Delta }} = {\Delta ^{1/2}}{}_\Delta V_{k\Delta }^{1/2}{}_\Delta {\xi _{\left( {k + 1} \right)\Delta }}$ and set ${}_\Delta {\tilde \varepsilon _{k\Delta }} = {G^{ - 1}}\left[ {F\left( {{}_\Delta {\varepsilon _{k\Delta }}} \right)} \right],$ where $F$ is the normal distribution and $G$ is the distribution for a variable ${}_\Delta {\tilde \varepsilon _{k\Delta }} $ that has zero mean and variance $\Delta {}_\Delta V_{k\Delta }^{} $, like ${}_\Delta {\varepsilon _{k\Delta }} $, but kurtosis equal to ${}_\Delta {\kappa _{k\Delta }}.$ This way, the errors of the discretized model have non-zero excess kurtosis.\ **Discussion:** The continuous model has two independent sources of randomness yet the discrete model has only one. That is, the discretization reduces the number of sources of randomness in the continuous model, via . There is no loss of generality using this discretization since the properties of the discretized Brownian motion (mean, variance and correlation) are maintained; ${}_\Delta {\eta _{\left( {k + 1} \right)\Delta }} $ is not exactly normal but it has a zero conditional mean, a unit conditional variance and zero correlation with ${}_\Delta {\xi _{\left( {k + 1} \right)\Delta }} $. We are bound to use such a method because, as argued by Lindner (2009, p. 482), the classic discretization does not work in this case. **Theorem 2:** Under the above discretization scheme the continuous limit in Theorem 1 returns the original weak GARCH model and the time aggregation property is preserved.\ **Proof:** The discretization of $\mu dt$ is obvious, and that for $\theta$ and $\omega$ will follow from the discretization of $\omega$ and $\lambda$ because: $$\omega \Delta \approx \omega \Delta \frac{{\left( {1 - \exp \left( { - \theta \Delta } \right)} \right)}}{{\theta \Delta }} = {}_\Delta \omega \quad \mbox{and} \, \, \theta \Delta \approx 1 - \exp \left( { - \theta \Delta } \right) = 1 - \left( {{}_\Delta \alpha + {}_\Delta \beta } \right) = 1 - {}_\Delta \lambda.$$ This gives: ${}_\Delta \omega = \omega {\theta ^{ - 1}}\left( {1 - \exp \left( { - \theta \Delta } \right)} \right) $ and ${}_\Delta \lambda = \exp \left( { - \theta \Delta } \right).$ As is clear from and , it is the discretization of $\beta$ that is most complex. From the aggregation results in Drost and Nijman (1993) we know that the unconditional kurtosis for a given frequency $\Delta$ may be expressed as a function of the parameters at an arbitrary higher frequency $\delta$ as: $${}_\Delta \kappa = 3 + {\Delta ^{ - 1}}\delta \left( {{}_\delta \kappa - 3} \right) + 6\left( {{}_\delta \kappa - 1} \right)\frac{{\left( {{\delta ^{ - 1}}\Delta \left( {1 - {}_\delta \lambda } \right) - \left( {1 - {}_\delta {\lambda ^{{\delta ^{ - 1}}\Delta }}} \right)} \right){}_\delta \alpha \left( {1 - {}_\delta {\lambda ^2} + {}_\delta \alpha {}_\delta \lambda } \right)}}{{{{\left( {{\delta ^{ - 1}}\Delta } \right)}^2}{{\left( {1 - {}_\delta \lambda } \right)}^2}\left( {1 - {}_\delta {\lambda ^2} + {}_\delta {\alpha ^2}} \right)}} .$$ Denoting the limit of the unconditional kurtosis by $\kappa : = \mathop {{\rm{lim}}}\limits_{\delta \downarrow 0} {}_\delta \kappa $, we obtain: $$\label{eqkurt2} {}_\Delta \kappa = 3 + 6\left( {\kappa - 1} \right)\frac{{{\alpha ^2}\left( {\theta \Delta - \left( {1 - \exp \left( { - \theta \Delta } \right)} \right)} \right)}}{{{\theta ^2}{\Delta ^2}\left( {{\alpha ^2} + 2\theta } \right)}}$$ where, by Proposition 1, the limit of the unconditional kurtosis is given by $\kappa = 3{\left( {1 - {\theta ^{ - 1}}{\alpha ^2}} \right)^{ - 1}} $. From the proof of Proposition 1, we know that for any two time steps $\Delta> \delta$, $_\Delta\beta$ is the solution to : $$\label{eqbeta2} \frac{{{}_\Delta \beta }}{{1 + {}_\Delta {\beta ^2}}} = \frac{{{}_{\Delta ,\delta }c{}_\delta {\lambda ^{{\delta ^{ - 1}}\Delta }} - 1}}{{{}_{\Delta ,\delta }c\left( {1 + {}_\delta {\lambda ^{2{\delta ^{ - 1}}\Delta }}} \right) - 2}}$$ where ${}_{\Delta ,\delta }c $ is given by . We want a discretization which ensures that will hold. Taking the limits of when $\delta$ goes to 0, we define: $${}_\Delta c: = \mathop {{\rm{lim}}}\limits_{\delta \downarrow 0} {}_{\Delta ,\delta }c = \left[ {\Delta {\alpha ^2} + {\Delta ^2}\theta \left( {\theta - {\alpha ^2}} \right) + 2{\alpha ^2}\left( {\Delta - {\theta ^{ - 1}}\left( {1 - \exp \left( { - \theta \Delta } \right)} \right)} \right)} \right]\left[ \frac{1}{2} \theta ^{ - 1}\alpha ^2\left( {1 - \exp \left( { - 2\theta \Delta } \right)} \right) \right]^{ - 1} .$$ This means that we can discretize the continuous model by solving the following equation: $$\frac{{{}_\Delta \beta }}{{1 + {}_\Delta {\beta ^2}}} = \frac{{{}_\Delta c\exp \left( { - \theta \Delta } \right) - 1}}{{{}_\Delta c\left( {1 + \exp \left( { - 2\theta \Delta } \right)} \right) - 2}} .$$ First, we have to make sure that this will have solutions, and then we have to show that there is a unique solution between zero and one. Let’s consider the function whose roots we want to find: $$f\left( x \right) = {x^2} - mx + 1,\quad m = \frac{{{}_\Delta c\left( {1 + \exp \left( { - 2\theta \Delta } \right)} \right) - 2}}{{{}_\Delta c\exp \left( { - \theta \Delta } \right) - 1}} .$$ This has two roots ${x_1}$ and ${x_2} $ where ${x_1}{x_2} = 1$ and ${x_1} + {x_2} = m $. If we show that $m$ is positive, then both roots are positive and one will be less than 1. For the existence we need that $m > 2$. If ${}_\Delta c\exp \left( { - \theta \Delta } \right) > 1 $ then $m > 2$ is equivalent to ${\left( {1 - \exp \left( { - \theta \Delta } \right)} \right)^2} > 0 $. Thus, all we need to show is that ${}_\Delta c\exp \left( { - \theta \Delta } \right) > 1 $, which is equivalent to: $6\theta \Delta + 2{\alpha ^{ - 2}}{\Delta ^2}{\theta ^3} + 5\exp \left( { - \theta \Delta } \right) > \exp \left( {\theta \Delta } \right) + 2{\Delta ^2}{\theta ^2} + 4 .$ Both sides of the above equation converge to 5 when $\Delta \rightarrow 0$, and it can be shown, using derivatives with respect to $\Delta $, that the left hand side converges faster. Thus ${}_\Delta c\exp \left( { - \theta \Delta } \right) > 1 $, so for any small step $\Delta$ close enough to zero there will always be a unique solution for $_\Delta \beta$ between zero and one that satisfies the above equation; this solution will be: $${}_\Delta \beta = \frac{1}{2} \left( {m - {{\left( {{m^2} - 4} \right)}^{1/2}}} \right) = \frac{{{}_\Delta c\left( {1 + \exp \left( { - 2\theta \Delta } \right)} \right) - 2 + \left( {1 - \exp \left( { - \theta \Delta } \right)} \right){{\left( {{}_\Delta {c^2}{{\left( {1 + \exp \left( { - \theta \Delta } \right)} \right)}^2} - 4{}_\Delta c} \right)}^{1/2}}}}{{2\left( {{}_\Delta c\exp \left( { - \theta \Delta } \right) - 1} \right)}}.$$ Also, we have that ${}_\Delta \alpha = \exp \left( { - \theta \Delta } \right) - {}_\Delta \beta $ . The discretization of the Brownian motions in our scheme is obvious, whilst there is no loss of generality in assuming . Now, we have ${y_{\left( {k + 1} \right)\Delta }} = \mu \Delta + {}_\Delta {\varepsilon _{\left( {k + 1} \right)\Delta }} $, hence $E\left( {{\Delta ^{ - 1}}{}_\Delta \varepsilon _{\left( {k + 1} \right)\Delta }^2| {{}_\Delta {I_{k\Delta }}} } \right) = {}_\Delta {V_{k\Delta }} $ with: $${}_\Delta {V_{\left( {k + 1} \right)\Delta }} = {}_\Delta \omega + {}_\Delta \alpha {\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k + 1} \right)\Delta }^2 + {}_\Delta \beta {}_\Delta {V_{k\Delta }} + {}_\Delta {u_{\left( {k + 1} \right)\Delta }},$$ $$\\{}_\Delta {u_{\left( {k + 1} \right)\Delta }}= {}_\Delta \alpha {}_\Delta {V_{k\Delta }}\left[ {1 - {{\left( {{2^{ - 1}}\left( {{}_\Delta {\kappa _{k\Delta }} - 1} \right)} \right)}^{1/2}} + \left( {{{\left( {{2^{ - 1}}\left( {{}_\Delta {\kappa _{k\Delta }} - 1} \right)} \right)}^{1/2}} - 1} \right){}_\Delta \tilde \xi _{\left( {k + 1} \right)\Delta }^2} \right]$$ where ${}_\Delta \tilde \xi _{\left( {k + 1} \right)\Delta }^{}$ has an unconditional kurtosis of ${}_\Delta {\kappa _{\left( {k + 1} \right)\Delta }} $, which can be approximated by ${}_\Delta {\kappa _{k\Delta }} $. So far we have considered the conditional variance; for the BLP of the squared residuals we have: $${}_\Delta {h_{k\Delta }} = {}_\Delta {V_{k\Delta }} - \sum\limits_{j = 0}^k {{}_\Delta {\beta ^j}{}_\Delta {u_{\left( {k - j} \right)\Delta }}} .$$ It is easy to see that this follows a GARCH process as ${}_\Delta {h_{k\Delta }} = {}_\Delta \omega + {}_\Delta \alpha \left( {{\Delta ^{ - 1}}{}_\Delta \varepsilon _{k\Delta }^2} \right) + {}_\Delta \beta {}_\Delta {h_{\left( {k - 1} \right)\Delta }} $. To have a weak GARCH we have to show that ${}_\Delta {h_{k\Delta }} $ is the BLP of ${\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k + 1} \right)\Delta }^2 $ , which requires showing that: $E\left( {\left( {{\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k + 1} \right)\Delta }^2 - {}_\Delta {h_{k\Delta }}} \right){}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^r} \right) = 0$ for $i \ge 0$, $r = 0,1,2 $. Since $E\left( {\left( {{}_\Delta {V_{\left( {k - j - 1} \right)\Delta }} - {\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k - j} \right)\Delta }^2} \right){}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^r} \right) = 0$, this reduces to showing $ E\left( {{}_\Delta {u_{\left( {k - j} \right)\Delta }}{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^r} \right) = 0$. This is satisfied for $i \ne j$. We now show the proof for $r = 1$ and $i = j$: We have to show that: $$E\left( {\left( {{}_\Delta {V_{\left( {k - i - 1} \right)\Delta }} - {\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^2} \right){}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^{}} \right) = 0,\quad i \ge 0$$ or $$E\left( {E\left( {{}_\Delta {V_{\left( {k - i - 1} \right)\Delta }}{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^{} - {\Delta ^{ - 1}}{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^3} \right)| {{}_\Delta {I_{\left( {k - i - 1} \right)\Delta }}} } \right) = 0,\quad i \ge 0,$$ which is clearly true. Also: $E\left( {\left. {{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^2} \right|{}_\Delta {I_{\left( {k - i - 1} \right)\Delta }}} \right) = \Delta {}_\Delta {V_{\left( {k - i - 1} \right)\Delta }} $, and $E\left( {\left. {{}_\Delta \tilde \varepsilon _{\left( {k - i} \right)\Delta }^4} \right|{}_\Delta {I_{\left( {k - i - 1} \right)\Delta }}} \right) = {\Delta ^2}{}_\Delta V_{\left( {k - i - 1} \right)\Delta }^2{}_\Delta {\kappa _{\left( {k - i - 1} \right)\Delta }} $. Thus, we have a weak GARCH specification; this means that the time aggregation is preserved by our discretization. It is easy to see that ${}_\Delta {\lambda ^{{\Delta ^{ - 1}}}} = {}_\delta {\lambda ^{{\delta ^{ - 1}}}} $and that ${}_\Delta \omega = {}_\delta \omega \left( {1 - {}_\delta {\lambda ^{{\delta ^{ - 1}}\Delta }}} \right)\left( {1 - {}_\delta \lambda } \right) $. We also have the relations for the kurtosis and for $\beta$. For the kurtosis, we need to prove , that is: $$\left( {\kappa - 1} \right){\alpha ^2}{\left( {1 - {}_\delta \lambda } \right)^2}\left( {1 - {}_\delta {\lambda ^2} + {}_\delta {\alpha ^2}} \right) = \left( {2{\delta ^2}{\theta ^2}\left( {{\alpha ^2} + 2\theta } \right) + 6\left( {\kappa - 1} \right){\alpha ^2}\left( {\theta \delta - \left( {1 - {}_\delta \lambda } \right)} \right)} \right){}_\delta \alpha \left( {1 - {}_\delta \lambda + {}_\delta \alpha {}_\delta \lambda } \right) .$$ After some algebra, this may be written as: $$\frac{{{}_\delta \beta }}{{1 + {}_\delta {\beta ^2}}} = \frac{{\left( {\delta {\alpha ^2} + {\delta ^2}\theta \left( {\theta - {\alpha ^2}} \right) + 2{\alpha ^2}\left( {\delta - \left( {1 - {}_\delta \lambda } \right)/\theta } \right)} \right){}_\delta \lambda - \raise.5ex\hbox{ $\scriptstyle 1 $}\kern-.1em/ \kern-.15em\lower.25ex\hbox{ $\scriptstyle 2 $} {\alpha ^2}{\theta ^{ - 1}}\left( {1 - {}_\delta {\lambda ^2}} \right)}}{{\left( {\delta {\alpha ^2} + {\delta ^2}\theta \left( {\theta - {\alpha ^2}} \right) + 2{\alpha ^2}\left( {\delta - \left( {1 - {}_\delta \lambda } \right)/\theta } \right)} \right)\left( {1 + {}_\delta {\lambda ^2}} \right) - {\alpha ^2}{\theta ^{ - 1}}\left( {1 - {}_\delta {\lambda ^2}} \right)}}.$$ Since the above expression holds, we have shown that the kurtosis is time aggregating. Simulation of the Model ======================= We simulate the continuous GARCH process 100,000 times, with 1000 steps, computing the price of a standard European option, and then finding the Black-Scholes (1973) implied volatility. Figure 1 (a) compares the volatility skew based on Nelson$'$s diffusion (solid line), with those from the weak GARCH diffusion, showing how different values of instantaneous kurtosis (assumed constant) can influence the shape of the implied volatility. When the instantaneous kurtosis is time dependent, in our case $\kappa \left( {T - t} \right) = {\rm{7}} - {\rm{2(}}T - t{\rm{)}} $, the steepness of the weak GARCH skew decreases with the option maturity, as in Figure 1 (b), showing that the GARCH limit has considerable flexibility to fit a volatility smile surface through a suitable parameterization of the instantaneous kurtosis.\ **Figure 1:** Comparison of volatility smiles generated by of the weak and strong GARCH diffusions: $\mbox{(a)} \, T-t =1; \, \mbox{(b)} \, T \, \mbox{varies}$ with $ \omega = 0.0045;\, \alpha = 0.1; \, \theta = 0.05;\, \mu = \tau = r = 0;\, \sigma_0 = 30\%$ {width="90.00000%"} Conclusions =========== We have presented several arguments which motivate the use of the weak rather than the strong version of the model for deriving a weak limit, i.e. a limit in distribution. There are four problems with the strong GARCH: First, it is not time aggregating: if we generate a GARCH process and then resample at another frequency the result is not a GARCH process. Second, the limit of strong GARCH may only be derived by making a specific assumption about the convergence of the parameters and different assumptions lead to different limits; Third, any discretization of the strong GARCH diffusion is not a GARCH model. And fourth, the variance of the variance is either zero or too small to fit the implied skew. This paper has derived the continuous limit of the weak GARCH by conjecturing only that the difference between the GARCH BLP process and the conditional variance converges to zero with the square root of the step-length. This GARCH model is time aggregating and it implies the convergence rates for all parameters (no need to make assumptions about these). Furthermore, the limit model derived is unique and a discretization that returns the original weak GARCH model is given. The weak GARCH diffusion is a stochastic variance process with independent Brownian motions in which the variance diffusion coefficient is related to the instantaneous kurtosis, and the limit reduces to Nelson’s GARCH diffusion when the excess kurtosis is zero.\ **References**\ \ Badescu A., Elliott R.J., Ortega J-P (2014) ‘Quadratic Hedging Schemes for Non-Gaussian GARCH Models’, Journal of Economic Dynamics and Control Vol. 42, 13-32.\ Badescu A., Elliott R.J., Ortega J-P (2015) ‘Non-Gaussian GARCH Option Pricing Models and their Diffusion Limits’, European Journal of Operational Research Vol. 247, 820-830.\ Badescu A., Cui. Z., Ortega J-P (2017) ‘Non-Gaussian GARCH Option Pricing Models, Variance-Dependent Kernels and Diffusion Limits’, Journal of Financial Econometrics, Vol. 15, 602–648.\ Bollerslev, T. (1986) ‘Generalized Autoregressive Conditional Heteroskedasticity’, Journal of Econometrics Vol. 31, 309-328.\ Brown, L. D., Y. Wang and L.H. Zhao (2002) ‘On the Statistical Equivalence at Suitable Frequencies of GARCH and Stochastic Volatility Models with Suitable Diffusion Models’, University of Pennsylvania Working paper.\ Buchmann, B. and Müller G. (2012) ‘Limit Experiments of GARCH’, Bernoulli Vol. 18, 64-99.\ Corradi, V. (2000) ‘Reconsidering the Continuous Time Limit of the GARCH(1,1) Process’, Journal of Econometrics Vol. 96, 145-153.\ Drost, F.C. and Nijman, T.E. (1993) ‘Temporal Aggregation of GARCH Processes’, Econometrica Vol. 61 (4), 909-927.\ Drost, F.C. and Werker, B.J.M. (1996) ‘Closing the GARCH Gap Continuous Time GARCH Modelling’, Journal of Econometrics Vol. 74, 31-57.\ Engle, R.F. (1982) ‘Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation’, Econometrica Vol. 50 (4), 987-1007.\ Fornari, F. and A. Mele (2005) ‘Approximating Volatility Diffusions with CEV-ARCH Models’, Journal of Economic Dynamics and Control Vol. 30 (6), 931-966.\ Kallsen J. and Vesenmayer B. (2009) ‘CO-GARCH as a Continuous-Time Limit of GARCH(1,1)’, Stochastic Processes and their Applications Vol. 119, 74-98.\ Kluppelberg, C., Lindner, A. and Maller, R. (2004) ‘A Continuous Time GARCH Process Driven by Levy Process Stationarity and Second Order Behaviour’, Journal of Applied Probability Vol. 43 (3), 601-622.\ Lindner A. (2009) ‘Continuous Time Approximations to GARCH and Stochastic Volatility Models’, Handbook of Financial Time Series, Edited by Andersen T., Davis R., Kreiss, J-P. and Mikosch, T. (Springer), 481-496.\ Maller, R., Miller, G. and Szimayer, A. (2008) ‘GARCH Modelling in Continuous Time for Irregularly Spaced Time Series Data’. Bernoulli, Vol. 14, 519-542.\ Meddahi, N. and Renault, E. (2004) ‘Temporal Aggregation of Volatility Models’, Journal of Econometrics 19, 355 - 379.\ Mele, A. and F. Fornari (2000) ‘Stochastic Volatility in Financial Markets Crossing the Bridge to Continuous Time’, Kluwer Academic Publishers.\ Muller, U.A., M.M. Dacorogna, R. Davé, R.B. Olsen, O.V. Pictett and J.E. Von Weizsäcker (1997) ‘Volatilities of Different Time Resolutions – Analyzing the Dynamics of Market Components’, Journal of Empirical Finance Vol. 4, 213-239.\ Nelson, D.B. (1990) ‘ARCH Models as Diffusion Approximations’, Journal of Econometrics Vol. 45, 7-38.\ Trifi, A. (2006) ‘Issues of Aggregation Over Time of Conditional Heteroscedastic Volatility Models What Kind of Diffusion Do We Recover?’, Studies in Nonlinear Dynamics and Econometrics Vol. 10 (4), 1314-1323.\ Wang, Y. (2002) ‘Asymptotic Non-Equivalence of GARCH Models and Diffusions’, The Annals of Statistics Vol. 30, 754-783.\ Zhang, R.-M. and Lin, Z.-Y. (2012) ‘Limit theory for a geneeral class of GARCH models with just barely infinite variance’, Journal of Time Series Analysis Vol. 33, 161-174\ Zheng, Z. (2005) ‘Re-crossing the Bridge from Discrete Time to Continuous Time Towards a Complete Model with Stochastic Volatility I’, available at SSRN http//ssrn.com/abstract=694261. [^1]: University of Sussex Business School, c.alexander@sussex.ac.uk [^2]: ICMA Centre, Henley Business School, Unibversity of Reading, e.lazar@icmacentre.ac.uk

ArXiv

**Characterization of several kinds** **of quantum analogues of relative entropy** Masahito Hayashi *ERATO Quantum Computation and Information Project, JST* *Hongo White Building, 5-28-3 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan* *Superrobust Computation Project* *Information Science and Technology Strategic Core (21st Century COE by MEXT)* *Graduate School of Information Science and Technology, The University of Tokyo* *7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan* Introduction ============ In the quantum information theory, we usually focus on the quantum relative entropy $D(\rho\|\sigma){\stackrel{\rm def}{=}}\Tr \rho (\log \rho- \log \sigma)$ as a quantum analogue of relative entropy (divergence). However, there are many kinds of quantum analogues of relative entropy. Some of them have been discussed from the viewpoint of operator algebra [@HP; @HP2]. In the classical information geometry, the divergence can be defined by using the integral along the autoparallel curve. Since the geometrical approach in classical information systems is very attractive, excellent insights for quantum information system can be expected through the consideration from geometrical viewpoints. By extending this definition to the quantum system, Nagaoka [@Nag94; @Na-pra] defined quantum analogues of divergence based on the integral along the parallel translation. $e$-parallel translation and $m$-parallel translation are known as most popular parallel translations in the quantum system as well as in the classical system. These divergences are called $e$-path-divergence and $e$-path-divergence, respectively. In particular, In the classical system, the path-divergences of both translations give the usual relative entropy. On the other hand, Fisher information is unique in the classical system. However, it is not unique in the quantum system. Petz[@Petz3] completely characterized its quantum analogues. As famous examples, SLD Fisher information, RLD Fisher information, and Bogoljubov Fisher information are known[@Helstrom:1967; @Petz3; @Petz4; @Na; @HolP]. Nagaoka showed that the quantum path-divergence concerning $e$ ($m$)-parallel translation coincides with the quantum relative entropy $D(\rho\|\sigma)$ when the quantum Fisher information of interest is Bogoljubov Fisher information[@Nag94; @Na-pra]. He also calculated the quantum path-divergence with the SLD Fisher information concerning $e$-parallel translation[@Nag94]. In this paper, we calculate the quantum path-divergence other than the above cases. Then, we succeeded in relating information geometrical path-divergence and an operator-algebraic divergence $\overline{D}(\rho\|\sigma) = \Tr \rho \log (\rho^{\frac{1}{2}}\sigma^{-1}\rho^{\frac{1}{2}})$, was introduced through operator-algebraic context by Belavkin and Staszewski [@BS]. Further, we proved the additivity of quantum path-divergence defined by $e$-parallel translation, and the monotonicity of quantum path-divergence defined by $m$-parallel translation. These two parallel translations are the dual parallel translations of each other. Since these two properties are fundamental, they are expected to be applied in the research field of quantum information. In the classical system, the divergence also can be defined from a convex function. Hence, divergence is closely related to convex analysis. Amari & Nagaoka [@Na] showed that only Bogoljubov Fisher information has zero-torsion. That is, the geometry of Bogoljubov inner product has the dual flat structure. They also proved the equivalence of the following two conditions. 1) The path-divergences of dual parallel translations can be given from potential function. 2) The dual parallel translation has the dual flat structure. Hence, in the quantum case, we can conclude that only path-divergences of Bogoljubov Fisher information is given by a potential function. This result indicates that the geometry of Bogoljubov Fisher information is closely related to optimization problem in quantum system. In fact, in their proof, the calculations concerning Christoffel symbols were essentially used. However, many quantum information scientists are not familiar to such analysis. In this paper, we give another proof of this argument without any use of Christoffel symbols. This paper can be expected to be a good guidance for quantum information geometry for quantum information scientist. This paper is organized as follows. In section \[2\], we review the information geometrical characterization of divergence $D(p\|q)$ in the classical system. we also review how the divergence can be defined by the convex function in the classical system. In section \[3\], we give a review of inner product in quantum systems, which is a fundamental of quantum information geometry. In section \[4\], two kinds of autoparallel translations and autoparallel curves are reviewed. In section \[5\], we treat quantum analogues of relative entropy from the operator-algebraic viewpoint. In section \[6\], we examine quantum path-divergences based on $e$-autoparallel translation, and consider their properties. In section \[7\], we examine quantum path-divergences based on $m$-autoparallel translation, and consider their properties. In particular, the relation between an operator-algebraic divergence and quantum path-divergences based on $e$ ($m$)-autoparallel translation are derived in section \[6\] (\[7\]). Divergence in Classical Systems {#2} =============================== First, we review the information geometrical characterization of divergence $D(p\|q)$ in the classical system [@Na]. Let $p(\omega)$ be a probability distribution, and $X(\omega)$ be a random variable. When the family $\{p_\theta | \theta \in \Theta\}$ has the form $$\begin{aligned} p_\theta(\omega)& = p(\omega) e^{\theta X(\omega)-\mu(\theta)} \label{4-10-11}\\ \mu(\theta) &{\stackrel{\rm def}{=}}\log \sum_{\omega} p(\omega) e^{\theta X(\omega)} \label{5-1-5},\end{aligned}$$ the logarithmic derivative at respective points equals the logarithmic derivative at a fixed point with the addition of a constant. In this case, the family is called an exponential family, and $\mu(\theta)$ is called the moment function of $X$. In particular, since the logarithmic derivative is closely related to exponential families, it is often called the exponential ($e$) representation of the derivative. Therefore, we use the superscript $(e)$ in the inner product $\langle~, ~\rangle^{(e)}_p$. The function $\mu(\theta)$ is often called a potential function in the context of information geometry. Since the second derivative $\mu''(\theta)$ is the Fisher information $J_\theta \ge 0$, the moment function $\mu(\theta)$ is a convex function. Therefore, the first derivative $\mu'(\theta) =\sum_\omega p_\theta (\omega) X(\omega)$ is monotone increasing. That is, we may regard it as another parameter identifying the distribution $p_\theta$, and denote it by $\eta$. The original parameter $\theta$ is called a natural parameter and the other parameter $\eta$ is an expectation parameter. For example, in the binomial distribution, the parameterization $p_\theta(0)= 1/(1+e^{\theta})$, $p_\theta(1)= e^{\theta}/(1+e^{\theta})$ is the natural parameter, and the parameterization $p_\eta(0)=\eta$, $p_\eta(1)=1-\eta$ is the expectation parameter. Hence, the binomial distribution is an exponential family. Further, let $X_1(\omega), \ldots, X_k(\omega)$ be $k$ random variables. We can define a $k$-parameter exponential family $$\begin{aligned} p_\theta(\omega) &{\stackrel{\rm def}{=}}p(\omega) e^{\sum_i \theta^i X_i(\omega)- \mu(\theta)} , \nonumber \\ \mu(\theta) &{\stackrel{\rm def}{=}}\log \sum_\omega p(\omega) e^{\sum_i \theta^i X_i(\omega)} .\label{7-1-1}\end{aligned}$$ The parameters $\theta^i$ are natural parameters, and the other parameters $\eta_i {\stackrel{\rm def}{=}}\frac{\partial \mu}{\partial \theta^i} =\sum_\omega p_\theta(\omega)X_i(\omega)$ are expectation parameters. Since the second derivative $\frac{\partial^2 \mu(\theta)} {\partial \theta^j \partial \theta^i}$ is equal to the Fisher Information matrix $J_{\theta:i,j}$, the moment function $\mu(\theta)$ is a convex function. Let $\mu(\theta)$ be a twice-differentiable and strictly convex function defined on a subset of the $d$-dimensional real vector space $\real^d$. The divergence concerning the convex function $\mu$ is defined by $$\begin{aligned} D^\mu(\bar{\theta}\|\theta) &{\stackrel{\rm def}{=}}\sum_i \eta_i(\bar{\theta}) (\bar{\theta}^i-\theta^i) - \mu(\bar{\theta})+\mu(\theta), \nonumber \\ \eta_i(\theta) &{\stackrel{\rm def}{=}}\frac{\partial \mu}{\partial \theta^i}(\theta)\label{7-3-9}.\end{aligned}$$ This quantity has the following two characterizations: $$\begin{aligned} D^\mu(\bar{\theta}\|\theta) =& \max_{\tilde{\theta}} \frac{\partial \mu}{\partial \theta^i}(\bar{\theta}) (\tilde{\theta}^i-\theta^i) - \mu(\tilde{\theta})+\mu(\theta) \nonumber \\ =& \int_0^1 \sum_{i,j} (\bar{\theta}^i-\theta^i ) (\bar{\theta}^j-\theta^j ) \frac{\partial^2 \mu}{\partial \theta^i\partial \theta^j} (\theta+ (\bar{\theta}-\theta)t) t d t.\label{7-4-4}\end{aligned}$$ In the one-parameter case, we obtain $$\begin{aligned} &D^\mu(\bar{\theta}\|\theta) = \mu'(\bar{\theta})(\bar{\theta}-\theta) - \mu(\bar{\theta})+\mu(\theta)\nonumber\\ =& \max_{\tilde{\theta}} \mu'(\bar{\theta})(\tilde{\theta}-\theta) - \mu(\tilde{\theta})+\mu(\theta) = \int_\theta^{\bar{\theta}} \mu''(\tilde{\theta}) (\tilde{\theta}- \theta) d \tilde{\theta}.\label{6-30-3}\end{aligned}$$ Since the function $\mu$ is strictly convex, the correspondence $\theta^i \leftrightarrow \eta_i = \frac{\partial \mu}{\partial \theta^i}$ is one-to-one. Hence, the divergence $D^\mu(\bar{\theta}\|\theta)$ can be expressed with the parameter $\eta$. For this purpose, we define the Legendre transform $\nu$ of $\mu$ $$\begin{aligned} \nu(\eta){\stackrel{\rm def}{=}}\max_{\tilde{\theta}} \sum_i \eta_i \tilde{\theta}^i - \mu(\tilde{\theta}).\end{aligned}$$ Then, the function $\nu$ is a convex function, and we can recover the function $\mu$ and $\theta$ as $$\begin{aligned} \mu(\theta) =\max_{\tilde{\eta}} \sum_i \theta_i \tilde{\eta}^i - \nu(\tilde{\eta}) , \quad \theta^i = \frac{\partial \nu}{\partial \eta_i}.\end{aligned}$$ The second derivative matrix $\frac{\partial^2 \nu}{\partial \eta_i \partial \eta_j}$ of $\nu$ is equal to the inverse of the matrix $\frac{\partial^2 \mu}{\partial \theta^i \partial \theta^j}$. In particular, when $\eta_i = \frac{\partial \mu}{\partial \theta^i}(\theta)$, $$\begin{aligned} \nu(\eta) &= \sum_i \eta_i {\theta}^i - \mu({\theta}) = D^\mu(\theta\|0)- \mu(0) ,\label{7-1-2}\\ \mu(\theta)&= \sum_i \theta_i {\eta}^i - \nu({\eta}) =D^\nu(\eta\|0)-\nu(0).\end{aligned}$$ Using this relation, we can characterize the divergence concerning the convex function $\mu$ by the divergence concerning the convex function $\nu$ as $$\begin{aligned} D^\mu(\bar{\theta}\|\theta) = D^\nu(\eta\|\bar{\eta}) = \sum_i \theta^i(\eta_i-\bar{\eta}_i) - \nu(\eta)+\nu(\bar{\eta}). \label{6-30-1}\end{aligned}$$ Now, we apply the discussion about the divergence to a multi-parametric exponential family $\{p_\theta| \theta \in \real \}$ defined in (\[7-1-1\]) [@Na]. Then, $$\begin{aligned} D(p_{\bar{\theta}}\|p_\theta) = D^\mu(\bar{\theta}\|\theta) = \sum_i \eta_i(\bar{\theta})(\bar{\theta}^i-\theta^i) - \mu(\bar{\theta})+\mu(\theta).\end{aligned}$$ In particular, applying (\[6-30-3\]) to a one-parameter exponential family (\[4-10-11\]), we have $$\begin{aligned} &D(p_{\bar{\theta}}\|p_{\theta}) = D(p_{\eta(\theta)+\epsilon}\|p_{\eta(\theta)}) = (\bar{\theta}-\theta)\eta(\bar{\theta}) - \mu(\bar{\theta})+\mu(\theta)\nonumber \\ = &\int_\theta^{\bar{\theta}} J_{\tilde{\theta}}(\tilde{\theta} - \theta) d \tilde{\theta} = \max_{\tilde{\theta} :\tilde{\theta} \ge \theta} (\tilde{\theta}-\theta) (\eta(\theta)+ \epsilon) - \mu(\tilde{\theta})+ \mu(\theta)\label{5-1-8-1}.\end{aligned}$$ In the following, we consider the case where $p$ is the uniform distribution $p_{\mix}$. Let the random variables $X_1(\omega), \ldots, X_k(\omega)$ be a CONS of the space of random variables with expectation $0$ under the uniform distribution $p_{\mix}$, and $Y^1(\omega), \ldots, Y^k(\omega)$ be its dual basis satisfying $\sum_\omega Y^i(\omega)X_j(\omega)= \delta^i_j$. Then, any distribution can be parameterized by the expectation parameter as $$\begin{aligned} p_{\eta(\theta)}(\omega) = p_{\mix}(\omega) +\sum_{i} \eta_i(\theta) Y^i(\omega).\end{aligned}$$ From (\[6-30-1\]) and (\[7-1-2\]), $$\begin{aligned} D(p_{\bar{\eta}}\|p_{\eta}) &= D^\nu(\eta\|\bar{\eta}) = \frac{\partial \nu}{\partial \eta_i} (\eta_i - \bar{\eta}_i) -\nu({\eta})+ \nu(\bar{\eta})\\ \nu(\eta) &=D(p_\eta\|p_{\mix})= - H(p_\eta)+ H(p_{\mix})\end{aligned}$$ because $\mu(0)=0$. The second derivative matrix of $\nu$ is the inverse of the second derivative matrix of $\mu$, [*i.e.*]{}, the Fisher information matrix concerning the natural parameter $\theta$. That is, the second derivative matrix of $\nu$ coincides with the Fisher information matrix concerning the expectation parameter $\eta$. Hence, applying (\[6-30-3\]) to the subspace $\{(1-t) p+ t q|0 \le t \le1 \}$, we have $$\begin{aligned} D(p\|q) =\int_0^1 J_t t dt ,\label{7-2-7}\end{aligned}$$ where $J_t$ is the Fisher information concerning the parameter $t$. Inner Products in Quantum Systems {#3} ================================= In this section, in order to define the quantum analogues of divergence, we define as inner products in quantum systems. There are at least three possible ways of defining the product corresponding to $X\rho$: $$\begin{aligned} E_{\rho,s}(X) & {\stackrel{\rm def}{=}}X \circ \rho {\stackrel{\rm def}{=}}\frac{1}{2}\left(\rho X + X \rho \right) \label{6-20-1},\\ E_{\rho,b}(X) & {\stackrel{\rm def}{=}}\int_0^1 \rho^{\lambda} X \rho^{1-\lambda} \,d \lambda , \nonumber\\ E_{\rho,r}(X) & {\stackrel{\rm def}{=}}\rho X .\label{6-20-3}\end{aligned}$$ Here, $X$ is not necessarily Hermitian. These extensions are unified in the general form [@Petz4] $$\begin{aligned} E_{\rho,p}(X) & {\stackrel{\rm def}{=}}\int_0^1 E_{\rho,\lambda}(X) p(d \lambda) , \label{6-20-4-1} \\ E_{\rho,\lambda}(X) & {\stackrel{\rm def}{=}}\rho^{\lambda} X \rho^{1-\lambda} , \label{7-2-1}\end{aligned}$$ where $p$ is an arbitrary probability distribution on $[0,1]$. The case (\[6-20-1\]) corresponds to the case (\[6-20-4-1\]) with $p(1)=p(0)=1/2$, and the case (\[6-20-3\]) does to the case (\[6-20-4-1\]) with $p(1)=1$. In particular, the map $E_{\rho,x}$ is symmetric, when $E_{\rho,x}(X)$ is Hermitian if and only if $X$ is Hermitian. Hence, when the distribution $p$ is symmetric, [*i.e.*]{}, $p(x)=p(1-x)$, the map $E_{\rho,p}$ is symmetric. When $\rho \,> 0$, these maps possess inverses. Accordingly, we may define these types of inner products $$\begin{aligned} \langle Y, X \rangle_{\rho,x}^{(e)} {\stackrel{\rm def}{=}}\Tr Y^* E_{\rho,x}(X) \quad x= s,b,r ,\lambda, p .\end{aligned}$$ If $X,Y,\rho$ all commute, these have the same value. These are called the SLD, Bogoljubov[^1], RLD, $\lambda$, and $p$ inner products[@Helstrom:1967; @Petz3; @Petz4; @Na; @HolP], respectively (reasons for this will be given in the next section). These inner products are positive semi-definite and Hermitian, [*i.e.*]{}, $$\begin{aligned} \left(\|X\|_{\rho,x}^{(e)}\right)^2 {\stackrel{\rm def}{=}}\langle X, X \rangle_{\rho,x}^{(e)} \ge 0,~ \langle Y, X \rangle_{\rho,x}^{(e)} = (\langle X, Y \rangle_{\rho,x}^{(e)} )^* .\end{aligned}$$ A dual inner product may be defined $\langle A, B \rangle_{\rho,x}^{(m)} {\stackrel{\rm def}{=}}\Tr (E_{\rho,x}^{- 1}(A))^* B$ with respect to the correspondence $A= E_{\rho,x}(X)$. Denote the norm of these inner products as $\left( \|A\|_{\rho,x}^{(m)} \right)^2 {\stackrel{\rm def}{=}}\langle A, A \rangle_{\rho,x}^{(m)}$. Hence, the inner product $\langle A, B \rangle^{(m)}_{\rho,x}$ is positive semi-definite and Hermitian. Using this inner product, we define quantum analogues of Fisher information as $$\begin{aligned} J_{\theta_0,x}{\stackrel{\rm def}{=}}\left(\left\|\frac{\,d \rho_\theta}{\,d \theta}(\theta_0) \right\|_{\rho_{\theta_0},x}^{(m)}\right)^2 \end{aligned}$$ for a one-parameter family $\{\rho_\theta\}$ and $x= s,r,b,\lambda,p$. Autoparallel Curves in Quantum Systems {#4} ====================================== Next, we define parallel transport and autoparallel curves in quantum systems according to Nagaoka [@Nag94] and Amari & Nagaoka[@Na]. To introduce the concept of a parallel transport, consider an infinitesimal displacement in a one-parameter quantum state family $\{\rho_\theta| \theta \in \real\}$. The difference between $\rho_{\theta+\epsilon}$ and $\rho_\theta$ approximately equals to $\frac{\,d \rho_\theta}{\,d \theta}(\theta) \epsilon$. Hence, the state $\rho_{\theta+\epsilon}$ can be regarded as the state transported from the state $\rho_\theta$ in the direction $\frac{\,d \rho_\theta}{\,d \theta}(\theta)$ by an amount $\epsilon$. However, if the state $\rho_{\theta+\epsilon}$ coincides precisely with the state displaced from the state $\rho_\theta$ by $\epsilon$ in the direction of $\frac{\,d \rho_\theta}{\,d \theta}(\theta)$, the infinitesimal displacement at the intermediate states $\rho_{\theta+\epsilon'}$ ($0\,<\epsilon'\,< \epsilon$) must equal the infinitesimal displacement $\frac{\,d \rho_\theta}{\,d \theta}(\theta)\Delta$ at $\theta$. Then, the problem is to ascertain which infinitesimal displacement at the point $\theta+\epsilon'$ corresponds to the given infinitesimal displacement $\frac{\,d \rho_\theta}{\,d \theta}(\theta)\Delta$ at the initial point $\theta$. The rule for matching the infinitesimal displacement at one point to the infinitesimal displacement at another point is called parallel transport. The coefficient $\frac{\,d \rho_\theta}{\,d \theta}(\theta)$ of the infinitesimal displacement at $\theta$ is called the tangent vector, as it represents the slope of the tangent line of the state family $\{\rho_\theta|\theta\in \real \}$ at $\theta$. Therefore, we can consider the parallel transport of a tangent vector instead of the parallel transport of an infinitesimal displacement. Commonly used parallel transports can be classified into those based on the $ m $ representation ($m$ parallel translation) and those based on the $ e $ representation ($e$ parallel translation). The $ m $ parallel translation $\Pi_{\rho_{\theta},\rho_{\theta'}}^{(m)}$ moves the tangent vector at one point $\rho_{\theta}$ to the tangent vector with the same $ m $ representation at another point $\rho_{\theta'}$. On the other hand, the $ e $ parallel translation $\Pi_{x,\rho_{\theta},\rho_{\theta'}}^{(e)}$ moves the tangent vector at one point $\rho_\theta$ with the $ e $ representation $ L $ to the tangent vector at another point $\rho_{\theta'}$ with the $ e $ representation $L - \Tr \rho_{\theta'} L$[@Na]. Of course, this definition requires the coincidence between the set of $ e $ representations at the point $\theta$ and that at another point $\theta'$. Hence, this type of $e$ parallel translation is defined only for the symmetric inner product $\langle X,Y\rangle^{(e)}_{\rho,x}$, and its definition depends on the choice of the metric. Indeed, the $ e $ parallel translation can be regarded as the dual parallel translation of the $ m $ parallel translation concerning the metric $\langle X,Y\rangle^{(e)}_{\rho,x}$ in the following sense: $$\begin{aligned} \Tr X^* \Pi_{\rho_{\theta},\rho_{\theta'}}^{(m)}(A)= \Tr \Pi_{x,\rho_{\theta'},\rho_{\theta}}^{(e)}(X)^* A,\end{aligned}$$ where $X$ is the $e$ representation of a tangent vector at $\rho_{\theta'}$ and $A$ is the $m$ representation of another tangent vector at $\rho_{\theta}$. Further, a one-parameter quantum state family is called a geodesic or an autoparallel curve when the tangent vector ([*i.e.*]{} the derivative) at each point is given as a parallel transport of a tangent vector at a fixed point. Especially, the $e$ geodesic is called a one-parameter exponential family. For example, in an $e$ geodesic with respect to SLD $\{\rho_\theta | \theta \in \real \}$, any state $\rho_{\theta}$ coincides with the state transported from the state $\rho_0$ along the autoparallel curve in the direction $L$ by an amount $\theta$, where $L$ denotes the SLD $e$ representation of the derivative at $\rho_0$. We shall henceforth denote the state as $\Pi^\theta_{L,s} \rho_0$. Similarly, $\Pi^\theta_{L,b} \rho_0$ denotes the state transported autoparallely with respect to the Bogoljubov $e$ representation from $\rho_0 $ in the direction $L$ by an amount $ \theta$. When the given metric is not symmetric, the $e$ parallel translation moves the tangent vector at one point $\theta$ under the $ e $ representation $\tilde{L}$ to the tangent vector at another point $\theta'$ with the $ e $ representation $\tilde{L}' - \Tr \rho_{\theta'} \tilde{L}'$ with the condition $\tilde{L} + \tilde{L}^* = \tilde{L}' + (\tilde{L}')^*$. That is, we require the same Hermitian part in the $e$ representation. Hence, the $ e $ parallel translation $\Pi_{x,\rho_{\theta},\rho_{\theta'}}^{(e)}$ coincides with the $ e $ parallel translation $\Pi_{s(x),\rho_{\theta},\rho_{\theta'}}^{(e)}$ with regard to its symmetrized inner product. Therefore, we can define the state transported from the state $\rho_0$ along the autoparallel curve in the direction with the Hermitian part $L$ by an amount $\theta$ with respect to RLD ($\lambda$, $p$), and denote them by $\Pi^\theta_{L,r} \rho_0$ ($\Pi^\theta_{L,\lambda} \rho_0$, $\Pi^\theta_{L,p} \rho_0$), respectively. However, only the SLD one-parameter exponential family $\{\Pi^\theta_{L,s} \rho_0| s \in \real \}$ plays an important role in quantum estimation examined in the next section. \[3-20t3\] $\Pi^\theta_{L,s} \sigma$, $\Pi^\theta_{L,b} \sigma$, $\Pi^\theta_{L,r} \sigma$ and $\Pi^\theta_{L,\frac{1}{2}} \sigma$ may be written in the following form[@Nagaoka:1989:2; @Nag94; @Na]: $$\begin{aligned} \Pi^\theta_{L,s} \sigma &= e^{-\mu_s(\theta)} e^{\frac{\theta}{2} L} \sigma e^{\frac{\theta}{2} L} , \label{3-19-5} \\ \Pi^\theta_{L,b} \sigma &= e^{-\mu_b(\theta)} e^{ \log \sigma + \theta L} \label{3-19-5.1},\\ \Pi^\theta_{L,r} \sigma &= e^{-\mu_r(\theta)} \sqrt{\sigma} e^{ \theta L_r} \sqrt{\sigma} \label{3-19-5.2},\\ \Pi^\theta_{L,\frac{1}{2}} \sigma &= e^{-\mu_{\frac{1}{2}}(\theta)} \sigma^{\frac{1}{4}} e^{ \frac{\theta}{2} L_{\frac{1}{2}}} \sigma^{\frac{1}{2}} e^{ \frac{\theta}{2} L_{\frac{1}{2}}} \sigma^{\frac{1}{4}} \label{3-19-5.3},\end{aligned}$$ where we choose Hermitian matrices $L_r$ and $L_{\frac{1}{2}}$ as $L= \frac{1}{2} (\sigma^{-\frac{1}{2}} L_r \sigma^{\frac{1}{2}}+ \sigma^{\frac{1}{2}} L_r \sigma^{-\frac{1}{2}} ) $ and $L= \frac{1}{2}(\sigma^{-\frac{1}{4}}L_{\frac{1}{2}} \sigma^{\frac{1}{4}} + \sigma^{\frac{1}{4}}L_{\frac{1}{2}} \sigma^{-\frac{1}{4}})$, respectively, and $$\begin{aligned} \mu_s(\theta) &{\stackrel{\rm def}{=}}\log \Tr e^{\frac{\theta}{2} L} \sigma e^{\frac{\theta}{2} L} \nonumber \\ \mu_b(\theta) & {\stackrel{\rm def}{=}}\log \Tr e^{ \log \sigma + \theta L},\label{9-18-6} \\ \mu_r(\theta) &{\stackrel{\rm def}{=}}\log \Tr \sqrt{\sigma}e^{ \theta L_r} \sqrt{\sigma},\nonumber \\ \mu_{1/2}(\theta) &{\stackrel{\rm def}{=}}\log \Tr \sigma^{\frac{1}{4}} e^{ \frac{\theta}{2} L_{\frac{1}{2}}} \sigma^{\frac{1}{2}} e^{ \frac{\theta}{2} L_{\frac{1}{2}}} \sigma^{\frac{1}{4}}. \nonumber\end{aligned}$$ Non-Geometrical Characterization of Divergences in Quantum Systems {#5} ================================================================== First, we briefly characterize quantum analogues of divergence from the non-geometrical viewpoint. A quantity $\tilde{D}(\rho\|\sigma)$ can be regarded as a quantum version of divergence if any commutative states $\rho$ and $\sigma$ satisfy $$\begin{aligned} \tilde{D}(\rho\|\sigma) = D(p\|\bar{p}),\label{7-5-4}\end{aligned}$$ where $p$ and $\bar{p}$ is the probability distribution consisting of the eigenvalues of $\rho$ and $\sigma$. If a relative entropy $\tilde{D}(\rho\|\sigma)$ satisfies the monotonicity for a POVM $\bM= \{M_i\}$: $$\begin{aligned} \tilde{D}(\rho\|\sigma) \ge D(\rP_\rho^{\bM}\|\rP_\sigma^{\bM}) ,~ \rP_\rho^{\bM}(i){\stackrel{\rm def}{=}}\Tr \rho M_i \label{7-5-2}\end{aligned}$$ and the additivity $$\begin{aligned} \tilde{D}(\rho_1 \otimes \rho_2\| \sigma_1 \otimes \sigma_2 )= \tilde{D}(\rho_1\|\sigma_1) +\tilde{D}(\rho_2\|\sigma_2),\label{7-5-10}\end{aligned}$$ then Hiai & Petz [@HP]’s result yields the relation $$\begin{aligned} \tilde{D}(\rho\|\sigma) = \lim \frac{\tilde{D}({\rho^{\otimes n}}\|{\sigma^{\otimes n}}) }{n} \ge \lim \sup_{\bM} \frac{D(\rP_{{\rho^{\otimes n}}}^{\bM}\|\rP_{{\sigma^{\otimes n}}}^{\bM})}{n} = D(\rho\|\sigma)\label{7-5-1}.\end{aligned}$$ That is, the quantum relative entropy $D(\rho\|\sigma)$ is the minimum quantum analogue of relative entropy with the monotonicity for measurement and the additivity. Further, Hiai & Petz [@HP] showed the inequality $$\begin{aligned} D(\rho\|\sigma) \le \overline{D}(\rho\|\sigma) {\stackrel{\rm def}{=}}\Tr \rho \log (\rho^{\frac{1}{2}}\sigma^{-1}\rho^{\frac{1}{2}}) . \label{9-26-22}\end{aligned}$$ Quantum Path-divergences Based on $e$-Parallel Translation {#6} ========================================================== Now, using the concept of the exponential family, we extend the path-divergence based on the first equation in (\[5-1-8-1\]). For any two states $\rho$ and $\sigma$, we choose the Hermitian matrix $L$ such that the exponential family $\{\Pi^\theta_{L,x} \sigma \}_{\theta \in [0,1]}$ concerning the inner product $J_{\theta,x}$ satisfies $$\begin{aligned} \Pi^{1}_{L,x} \sigma= \rho.\label{7-2-5}\end{aligned}$$ Then, we define the $x$-$e$-divergence as follows: $$\begin{aligned} D^{(e)}_x(\rho\|\sigma) = \int_0^1 J_{\theta,x} \theta d \theta,\end{aligned}$$ where $J_{\theta,x}$ is the Fisher information concerning the exponential family $\Pi^\theta_{L,x} \sigma$. Since $\Pi^\theta_{L^1 \otimes I + I \otimes L^2,x} (\sigma_1 \otimes \sigma_2) $ equals $(\Pi^\theta_{L^1,x} \sigma_1)\otimes (\Pi^\theta_{L^2,x} \sigma_2) $, $$\begin{aligned} D^{(e)}_x(\rho_1\otimes \rho_2\|\sigma_1\otimes \sigma_2)= D^{(e)}_x(\rho_1\|\sigma_1)+ D^{(e)}_x(\rho_2\|\sigma_2),\label{7-5-11}\end{aligned}$$ [*i.e.*]{}, the $e$-divergence satisfies the additivity for any inner product. When $$\begin{aligned} L= \left\{ \begin{array}{ll} 2\log \sigma^{-\frac{1}{2}} (\sigma^{\frac{1}{2}}\rho\sigma^{\frac{1}{2}})^{\frac{1}{2}} \sigma^{-\frac{1}{2}} & \hbox{ for }x= s\\ \log \rho -\log \sigma & \hbox{ for }x= b\\ \frac{1}{2}\bigl[ \sigma^{-\frac{1}{2}} \log (\sigma^{-\frac{1}{2}}\rho\sigma^{-\frac{1}{2}}) \sigma^{\frac{1}{2}} & \hbox{ for }x= r\\ \quad + \sigma^{\frac{1}{2}} \log (\sigma^{-\frac{1}{2}}\rho\sigma^{-\frac{1}{2}}) \sigma^{-\frac{1}{2}}\bigr] & \\ \sigma^{-\frac{1}{4}} \log (\sigma^{-\frac{1}{4}}\rho^{\frac{1}{2}}\sigma^{-\frac{1}{4}}) \sigma^{\frac{1}{4}} &\hbox{ for }x= \frac{1}{2}\\ \quad + \sigma^{\frac{1}{4}} \log (\sigma^{-\frac{1}{4}}\rho^{\frac{1}{2}}\sigma^{-\frac{1}{4}}) \sigma^{-\frac{1}{4}}, & \end{array} \right.\label{7-3-1}\end{aligned}$$ the condition (\[7-2-5\]) holds. Hence, we obtain $$\begin{aligned} D^{(e)}_s(\rho\|\sigma)&= 2 \Tr \rho \log \sigma^{-\frac{1}{2}} (\sigma^{\frac{1}{2}}\rho\sigma^{\frac{1}{2}})^{\frac{1}{2}} \sigma^{-\frac{1}{2}} \label{7-3-2.s}\\ D^{(e)}_b(\rho\|\sigma)&= \Tr \rho (\log \rho -\log \sigma) =D(\rho\|\sigma) \label{7-3-2.b}\\ D^{(e)}_r(\rho\|\sigma)&= \Tr \rho \log (\rho^{\frac{1}{2}} \sigma^{-1}\rho^{\frac{1}{2}}) =\overline{D}(\rho\|\sigma) \label{7-3-2.r}\\ D^{(e)}_{\frac{1}{2}}(\rho\|\sigma)&= 2 \Tr (\sigma^{\frac{1}{4}} \rho^{\frac{1}{2}} \sigma^{\frac{1}{4}}) (\sigma^{-\frac{1}{4}}\rho^{\frac{1}{2}}\sigma^{-\frac{1}{4}}) \log (\sigma^{-\frac{1}{4}}\rho^{\frac{1}{2}}\sigma^{-\frac{1}{4}}). \label{7-3-2.1/2}\end{aligned}$$ Nagaoka [@Nag94] obtained the above results for $x=s,b$. Now, we compare these quantum analogues of relative entropy given in (\[7-3-2.s\])–(\[7-3-2.1/2\]). As is easily checked, these satisfy the condition (\[7-5-4\]) for quantum analogues of relative entropy. Let $\bM$ be a measurement corresponding to the spectral decomposition of $\sigma^{-1/2} (\sigma^{1/2}\rho\sigma^{1/2})^{1/2}\sigma^{-1/2}$. This PVM $\bM$ satisfies that $D_s^{(e)} (\rho\|\sigma)= D(\rP_\rho^{\bM}\|\rP_\sigma^{\bM})$. Thus, from the monotonicity for measurement concerning the quantum relative entropy $D(\rho\|\sigma)$, $$\begin{aligned} D(\rho\|\sigma)\ge D^{(e)}_s(\rho\|\sigma) = 2 \Tr \rho \log \sigma^{-\frac{1}{2}} (\sigma^{\frac{1}{2}}\rho\sigma^{\frac{1}{2}})^{\frac{1}{2}} \sigma^{-\frac{1}{2}} .\label{7-5-6}\end{aligned}$$ From (\[9-26-22\]), $$\begin{aligned} D(\rho\|\sigma)\le D^{(e)}_{r}(\rho\|\sigma) =\Tr \rho \log (\rho^{\frac{1}{2}}\sigma^{-1}\rho^{\frac{1}{2}}) .\label{7-5-7}\end{aligned}$$ Hence, from the inequality (\[7-5-1\]) and the additivity (\[7-5-11\]), $D^{(e)}_s(\rho\|\sigma)$ and $D^{(e)}_r(\rho\|\sigma)$ do not satisfy the monotonicity even for measurements because the equality in (\[7-5-6\]) and (\[7-5-7\]) does not always hold. Quantum Path-divergences Based on $m$-Parallel Translation {#7} ========================================================== Further, we can extend the path-divergence based on the equation (\[7-2-7\]). For any two states $\rho$ and $\sigma$, the family $\{(1-t)\rho + t\sigma | 0 \le t \le 1\}$ is the $m$ geodesic joining $\rho$ and $\sigma$. Hence, as an extension of (\[7-2-7\]), we can define the $x$-$m$ divergence as $$\begin{aligned} D^{(m)}_x(\rho\|\sigma){\stackrel{\rm def}{=}}\int_0^1 J_{t,x} t d t.\end{aligned}$$ Since the family $\{(1-t)\kappa(\rho) + t\kappa(\sigma) | 0 \le t \le 1\}$ is the $m$ geodesic joining $\kappa(\rho)$ and $\kappa(\sigma)$ for any TP-CP map $\kappa$, we have $$\begin{aligned} D^{(m)}_x(\rho\|\sigma) \ge D^{(m)}_x(\kappa(\rho)\|\kappa(\sigma)),\label{7-5-14}\end{aligned}$$ [*i.e.*]{}, the $m$ divergence satisfies the monotonicity. Since the RLD is the largest inner product, $$\begin{aligned} D^{(m)}_r(\rho\|\sigma) \ge D^{(m)}_x(\rho\|\sigma).\label{7-5-19}\end{aligned}$$ We can calculate the $m$ divergence as $$\begin{aligned} D^{(m)}_b(\rho\|\sigma)&= \Tr \rho (\log \rho -\log \sigma) =D(\rho\|\sigma) \label{7-4-7.b}\\ D^{(m)}_r(\rho\|\sigma)&= \Tr \rho \log (\sqrt{\rho}\sigma^{-1}\sqrt{\rho}) = \overline{D}(\rho\|\sigma) . \label{7-4-7.r}\end{aligned}$$ In fact, The Bogoljubov case (\[7-4-7.b\]) has been obtained by Nagaoka [@Na-pra], and follows from Theorem \[7-4-5\]. Hence, $\Tr \rho \log (\sqrt{\rho}\sigma^{-1}\sqrt{\rho})= D^{(m)}_r(\rho\|\sigma)$ satisfies the monotonicity for TP-CP maps. Also, from (\[7-5-19\]), we obtain $\Tr \rho \log (\sqrt{\rho}\sigma^{-1}\sqrt{\rho}) \ge D(\rho\|\sigma)$[@HP]. Further, all of $x$-$m$ divergences do not necessarily satisfy the additivity (\[7-5-11\]). At least, when the inner product $J_{x,\theta}$ is smaller than the Bogoljubov inner product $J_{b,\theta}$, [*i.e.*]{}, $J_{\theta,x} \le J_{\theta,b}$, we have $D(\rho\|\sigma)\ge D^{(m)}_x(\rho\|\sigma)$. From (\[7-5-1\]) and the monotonicity (\[7-5-14\]), $D^{(m)}_x(\rho\|\sigma)$ does not satisfy the additivity (\[7-5-11\]). For example, SLD $m$ divergence does not satisfy the additivity (\[7-5-11\]). We can now verify whether it is possible in two-parameter state families to have states that are $ e $ autoparallel transported in the direction of $L_1$ by $\theta^1$, and in the direction $L_2$ by $\theta^2$. In order to define such a state, we require that the state that is $ e $ autoparallel transported first in the $L_1$ direction by $\theta^1$ from $\rho_0$, then further $ e $ autoparallel transported in the $L_2$ direction by $\theta^2$ coincides with the state that is $ e $ autoparallel transported in the $L_2$ direction by $\theta^2$ from $\rho_0$, then $ e $ autoparallel transported in the $L_1$ direction by $\theta^1$. That is, if such a state would be defined, the relation $$\begin{aligned} \Pi^{\theta^2}_{L_2,x} \Pi^{\theta^1}_{L_1,x} \sigma = \Pi^{\theta^1}_{L_1,x} \Pi^{\theta^2}_{L_2,x} \sigma \label{7-3-10}\end{aligned}$$ should hold. Concerning this condition, we have the following theorem. \[7-4-5\] The following conditions for the inner product $J_{\theta,x}$ are equivalent 1. $J_{\theta,x}$ is the Bogoljubov inner product, [*i.e.*]{}, $x=b$. 2. The condition (\[7-3-10\]) holds for any two Hermitian matrices $L_1$ and $L_2$ and any state $\rho_0$. 3. $D^{(e)}_x(\rho_{\bar{\theta}}\|\rho_\theta) = D^\mu(\bar{\theta}\|\theta)$. 4. $D^{(e)}_x(\rho\|\sigma) =D(\rho\|\sigma)$. 5. $D^{(m)}_x(\rho_{\bar{\eta}}\|\rho_\eta) = D^\nu(\eta\|\bar{\eta})$. 6. $D^{(m)}_x(\rho\|\sigma)=D(\rho\|\sigma)$. Here, the convex functions $\mu(\theta)$, $\nu(\eta)$ and the states $\rho_\theta$, $\rho_\eta$ are defined by $$\begin{aligned} \rho_\theta {\stackrel{\rm def}{=}}&\exp(\sum_i \theta^i X_i-\mu(\theta)),\nonumber \\ \mu(\theta){\stackrel{\rm def}{=}}&\log \Tr \exp(\sum_i \theta^i X_i),\label{7-3-4}\\ \rho_\eta {\stackrel{\rm def}{=}}&\rho_{\mix}+ \sum_j \eta_j Y^j, \nonumber \\ \nu(\eta){\stackrel{\rm def}{=}}& D_x^{(m)}(\rho_0\|\rho_\eta) = -H(\rho_\eta)+H(\rho_{\mix}), \nonumber\end{aligned}$$ where $X_1, \ldots, X_k$ is a basis of the set of traceless Hermitian matrices, and $Y^1, \ldots, Y^k$ is its dual basis. This theorem implies that only the quantum path-divergence based on the Bogoljubov Fisher information can be characterized by the convex function among quantum path-divergence based on $m$-parallel translation. Concluding Remark ================= In this paper, we proved the additivity of $e$-divergences and the monotonicity of $m$-divergences. We also found interesting relations between geometrical path-divergences and an operator-algebraic divergence as $$\begin{aligned} D^{(e)}_r(\rho\|\sigma)= D^{(m)}_r(\rho\|\sigma)=\overline{D}(\rho\|\sigma) .\end{aligned}$$ In addition, we obtained the characterization of Bogoljubov inner product as Theorem \[7-4-5\], which is a generalization of Amari & Nagaoka [@Na]’s characterization. It is expected that these characterizations are applied to quantum information. [99]{} S. Amari and H. Nagaoka, [*Methods of Information Geometry*]{}, (AMS & Oxford University Press, 2000). H. Nagaoka, “Differential Geometrical Aspects of Quantum State Estimation and Relative Entropy,” in [*Quantum Communications and Measurement,*]{} edited by V. P. Belavkin, O. Hirota and R. L Hudson 449-452 (Plenum, New York, 1995). H. Nagaoka, “On the Parameter Estimation Problem for Quantum Statistical Models,” [*Proc. 12th Symposium on Information Theory and Its Applications (SITA)*]{}, 577–582 (1989). It is also appeared as Chapter 10 of [*Asymptotic Theory of Quantum Statistical Inference,*]{} edited by M. Hayashi. F. Hiai and D. Petz, “The Golden-Thompson Trace Inequality is Complemented,” [*Linear Algebra and its Applications*]{}, [**181**]{}, 153-185 (1993). F. Hiai and D. Petz, “The proper formula for relative entropy and its asymptotics in quantum probability,” [*Com. Math. Phys.*]{}, [**143**]{}, 99–114, (1991). D. Petz and G. Toth, [*Lett. Math. Phys.*]{}, [**27**]{}, 205, (1993). C. W. Helstrom, “Minimum mean-square error estimation in quantum statistics,” [*Phys. Lett.*]{}, [**25A**]{}, 101-102 (1976). D. Petz, “Monotone Metrics on Matrix Spaces,” [*Linear Algebra and its Applications*]{}, [**224**]{}, 81-96 (1996). A. S. Holevo, [*Probabilistic and Statistical Aspects of Quantum Theory*]{}, (North-Holland, 1982); Originally in Russian (1980). H. Nagaoka: Private communication to A. Fujiwara (1991). V. P. Balavkin and P. Staszewski, “C$^*$-algebraic generalization of relative entropy and entropy,” [*Ann. Inst. H. Poincaré Sect.*]{} [**A 37**]{}, 51–58 (1982). [^1]: The Bogoljubov inner product is also called the canonical correlation in statistical mechanics. In linear response theory, it is often used to give an approximate correlation between two different physical quantities.

ArXiv

--- abstract: 'High-resolution ground-based optical speckle and near-infrared adaptive optics images are taken to search for stars in close angular proximity to host stars of candidate planets identified by the NASA [[*Kepler*]{} ]{}Mission. Neighboring stars are a potential source of false positive signals. These stars also blend into [[*Kepler*]{} ]{}light curves, affecting estimated planet properties, and are important for an understanding of planets in multiple star systems. Deep images with high angular resolution help to validate candidate planets by excluding potential background eclipsing binaries as the source of the transit signals. A study of 18 [[*Kepler*]{} ]{}Object of Interest stars hosting a total of 28 candidate and validated planets is presented. Validation levels are determined for 18 planets against the likelihood of a false positive from a background eclipsing binary. Most of these are validated at the 99% level or higher, including 5 newly-validated planets in two systems: Kepler-430 and Kepler-431. The stellar properties of the candidate host stars are determined by supplementing existing literature values with new spectroscopic characterizations. Close neighbors of 7 of these stars are examined using multi-wavelength photometry to determine their nature and influence on the candidate planet properties. Most of the close neighbors appear to be gravitationally-bound secondaries, while a few are best explained as closely co-aligned field stars. Revised planet properties are derived for each candidate and validated planet, including cases where the close neighbors are the potential host stars.' author: - 'Mark E. Everett, Thomas Barclay, David R. Ciardi, Elliott P. Horch, Steve B. Howell, Justin R. Crepp, David R. Silva' title: 'High-resolution Multi-band Imaging for Validation and Characterization of Small [[*Kepler*]{} ]{}Planets' --- INTRODUCTION {#sec:introduction} ============ The NASA [[*Kepler*]{} ]{}Mission employed a 0.95 m aperture Schmidt telescope in solar orbit for a total of 4 years (May 2009 – May 2013). [ *Kepler’s*]{} focal plane was filled with 42 CCDs to collect time series photometry on selected targets in a 115 square degree field. [[*Kepler*]{} ]{}detected transiting exoplanets from a sample of over 150,000 target stars, most of which fell in the [[*Kepler*]{} ]{}magnitude range $Kp=8-16$ ($Kp$, the [[*Kepler*]{} ]{}bandpass, spans roughly $430-900$ nm). The mission was designed to detect and quantify the population of small planets orbiting within or near the habitable zones (HZs) of Sun-like stars [@boruckietal10]. [[*Kepler*]{} ]{}has produced thousands of candidate planets, dozens of which are good HZ or near-HZ candidates [@batalhaetal13]. To help confirm the candidates as true exoplanets, the mission has relied on ground-based follow-up observations of the candidate host stars. The process of producing a list of transiting planets from [[*Kepler*]{} ]{}data is a long one. First, raw pixel fluxes are calibrated [@quintanaetal10], and light curves are extracted from apertures and reduced, correcting the flux time series by way of “cotrending” to remove variations correlated with ancillary spacecraft data [@twickenetal10]. At the same time, nearby stars identified in the [[*Kepler*]{} ]{}Input Catalog [KIC; @brownetal11] are used to estimate blended (excess) flux in the light curves, and this excess flux is removed. Following reduction, the light curves are searched for any significant periodic events similar to those of transiting planets [@jenkinsetal10b]. These “threshold crossing events” (TCEs) consist of true planet transits and false positives (events appearing much like planet transits, but attributable to other phenomena). False positives include astrophysical sources like eclipsing binary stars, planets transiting nearby, fainter stars blended with the KOI star, and instrumental artifacts occurring (quasi-)periodically in the time series and which coincide when a light curve is searched on a certain period. Here and after, “KOI star” refers to the brightest star near the center of the [[*Kepler*]{} ]{}aperture as measured in the [[*Kepler*]{} ]{}bandpass (an unambiguous definition for the sample in this study). Cases where the TCEs are due to planets transiting stars other than the KOI are treated as false positives because their planetary properties will have been miscalculated based on adoption of the KOI star properties. Such false positives should be removed from the KOI list, if possible, to maintain it as a well-defined statistical sample. The TCEs are subjected to data validation through a series of automated tests [@wuetal10] and human inspection to weed out obvious false positives. Those TCEs passing data validation are deemed [[*Kepler*]{} ]{}Objects of Interest (KOIs) and given a disposition that identifies some as planet candidates. The term KOI can refer to planet candidates as well as their host stars. The KOI list forms a large and relatively clean sample with respect to instrumental false positives, but still contains a significant number of astrophysical false positives. The false positive rate is uncertain, but is likely to be about $10\%$ [@fressinetal13; @santerneetal13]. Follow-up observations may be used to identify the false positives and more accurately characterize the host star properties from which planet properties are derived. For example, the high resolution follow-up imaging described here is used to determine the location and brightness of each star that contributes flux to planet candidate light curves because [*Kepler*]{} imaging is optimized only for photometry (having $4\arcsec$ wide pixels, typical stellar profiles of $\sim6\arcsec$ FWHM and variably-sized photometric apertures that are typically several pixels across). The number of confirmed or validated [[*Kepler*]{} ]{}planets currently stands at 965, which is 23% of the total number of both confirmed and candidate [[*Kepler*]{} ]{}planets (4233). This relatively small fraction partially reflects the challenges to follow-up observing and analysis needed to confirm planets with a low level of false positive probability. This study presents the analysis of high spatial resolution observations of 18 KOI stars and the 28 validated and candidate planets they harbor. The stellar sample is listed in Table \[table:spklobservations\]. Each KOI star has been observed using high resolution optical speckle imaging techniques to search for or put limits on the brightness of previously-unresolved neighboring stars. Many have also been observed in the near-infrared (near-IR) with adaptive optics imaging with the same goals in mind. Most of the host stars have been observed spectroscopically to define their stellar properties, while the others have stellar properties available in the literature. The high-resolution imaging is used to calculate a validation level for 18 planets around 12 of these stars by constraining the non-detection of nearby sources. Two new validated planetary systems containing 5 planets are designated Kepler-430 and Kepler-431. The effects of blending by neighboring stars are examined and quantified for planets orbiting the 7 affected stars and tests are performed that help to distinguish whether these neighboring stars are gravitationally-bound companions or field stars. These high resolution imaging and single epoch spectral observations prove to be an efficient follow-up method for planet validations and refinement of the planet and host star sample. Such observations lead to a better understanding of the sample of small [[*Kepler*]{} ]{}planets. CANDIDATE PLANET SAMPLE {#sec:sample} ======================= The sample analyzed here is a set of KOI host stars observed with optical speckle imaging at Gemini North during July 2013. These targets were selected from the KOI list at the time on the basis of two main considerations: (1) they were not previously observed with high resolution optical imaging at an 8 m or larger telescope and (2) they hosted a candidate planet having an estimated radius less than $1.5R_{\bigoplus}$ and/or a predicted planet equilibrium temperature $T_{eq}<320$ K. At the time of target selection there was a total of 750 stars hosting at least one planet meeting this size constraint and 20 stars hosting at least one planet meeting the temperature constraint (temperatures low enough to be considered HZ candidates). Since that time, planets have been validated for 140 of these 750 host stars, primarily as part of a validation study of planets in multiple planet systems [@roweetal14], although most of these are lacking the high resolution imaging needed to thoroughly investigate their possible stellar multiplicity. A total of 25 of the brightest of these 750 stars was observed (selected to include some with low equilibrium temperature), but 5 of the stars were subsequently found by the mission to be false positive events (mostly cases where the variable was not the KOI, but another star in the aperture). The results for two stars of the sample are discussed separately in the literature: KOI 571 (Kepler-186) by @quintanaetal14 and KOI 2626 by @ciardietal14. The remaining 18 stars discussed here (Table \[table:spklobservations\]) hosted a total of 28 candidates (although some have been subsequently validated). Along with new observations, analysis of these candidates began by inspecting ground-based data and [[*Kepler*]{} ]{}data products available from the web site of the [[*Kepler*]{} ]{}Community Follow-up Observing Program (CFOP)[^1]. This included the $J$-band survey taken at UKIRT (by Phil Lucas) that covers the entire [[*Kepler*]{} ]{}field under relatively good seeing conditions ($0.8-0.9\arcsec$ FWHM). The $J$ images were examined to locate stars nearby each KOI. Sources as close as $\sim1\arcsec$ (corresponding to 408 AU at the mean distance of the stellar sample) could be readily seen in these images, but more importantly they covered areas outside of the relatively small fields of the follow-up high resolution images. Another data product used were the [[*Kepler*]{} ]{}Mission’s data validation reports that show light curves and statistical tests on such things as the motion of stellar centroids in and out of transit, comparison of the depths of odd versus even numbered transits, offsets of the transit relative to predicted positions for the star, and in-transit versus out-of-transit pixel flux differences. The statistics in the validation reports help determine if any of the candidates are particularly suspect as false positives [@brysonetal13]. The candidates discussed hereafter are “good” candidates in that the inspection uncovered nothing especially indicative of false positives. OBSERVATIONS AND DATA REDUCTION {#sec:observations} =============================== Speckle Imaging at Gemini North ------------------------------- Speckle imaging observations were obtained at Gemini North during the interval UT 25$-$31 July 2013. The Differential Speckle Survey Instrument (DSSI), a dual-channel speckle imaging system described by @horchetal09, was configured with the 692 nm filter (40 nm FWHM) on the first port and the 880 nm filter (50 nm FWHM) on the second port. While the mounting of the camera went smoothly, there was a light-leak problem in the 880 nm channel on the night of 25 July UT, so the data from that channel was of significantly lower quality and will not be reported here. The problem was identified and eliminated by the start of 26 July UT. The pixel scale and orientation were measured by observing two well-known binary systems, HU 1176 (ie. HIP 83838 or HR 6377) and STT 535 (ie. HIP 104858 or HR 8123). The known orbital elements from the Sixth Orbit Catalog[^2] were used to calculate the position angle and separation at the time of the observation, and then compared with the raw pixel coordinates, thereby deriving the scale. Each camera has a slightly different value. The final values were determined to be $0.01076\arcsec~{\rm pixel}^{-1}$ for the 692 nm camera and $0.01142\arcsec~{\rm pixel}^{-1}$ for the 880 nm camera. The position angle difference between pixel axes and celestial coordinates was determined to be $5.69\arcdeg$. Previous experiences and similar observations were taken during 2012 at Gemini North and are described by @horchetal12. Images were acquired simultaneously in both cameras. The raw data file for each camera consists of 1000 frames (which is called an “exposure”); at least three exposures were taken for each of the objects and were examined individually and then co-added to achieve the best possible final result. While the objects were acquired and centered on the two detectors with real-time full-frame readout ($512\times512$ pixels), the science exposures consisted of frames that were $256\times256$-pixel subarrays, centered on the target. Each frame was 60 ms in duration, meaning each exposure represented 1 minute of integration time. The choice for the number of exposures taken generally followed the magnitude of the target, as one would expect, with the fainter objects receiving more time, but also modified at the telescope depending on seeing, airmass and other factors. Table \[table:spklobservations\] gives the number of exposures and the [[*Kepler*]{} ]{}magnitudes for the systems under study. The seeing for the run varied between approximately FWHM$=0.5\arcsec-0.8\arcsec$, with substantial changes from exposure to exposure for some objects due to weather systems that were in the area at the time (including Tropical Storm Flossie, which grazed the Hawaiian Islands on UT 29 and 30 July). Overall, the data from 27 July represents the bulk of what we present here. This was a relatively calm night with slightly better seeing than the run as a whole. The basic methodology for speckle data reduction has been described in previous papers, e.g. @howelletal11 and @horchetal12. The latter deals specifically with Gemini data taken in 2012. It is based on Fourier analysis of correlation functions made from the raw speckle data frames. The autocorrelation is used to estimate the modulus of the object’s Fourier transform. A point source observation is required to deconvolve the point spread function (which amounts to a division in the Fourier domain). The triple correlation function can be used to generate the phase of the object in the Fourier plane. Combining these two functions, an estimate of the Fourier transform of the object is obtained. This is then low-pass filtered with a Gaussian function and inverse transformed to arrive at the final reconstructed image with a diffraction-limited resolution of FWHM$\simeq0.02\arcsec$. Example reconstructed speckle images centered on the double source KOI 1964 are shown in Fig. \[Fig:KOI1964\_4filters\]. For [[*Kepler*]{} ]{}follow-up observations, we use the reconstructed images to measure the limiting magnitude difference of each observation as a function of distance from the primary star, that is, it is an estimate of the brightest star that could be missed as a function of separation from the primary. As shown in the previous papers, these curves are generally monotonically increasing as a function of separation, meaning that the limiting magnitude near the central star is lower than farther away from the star. Up to the present, we have published $5\sigma$ confidence limits as a function of separation, using all peaks in the reconstructed image to generate a mean and standard deviation of the mean of the peak values. A detectable companion star then must have a peak value larger than the mean plus 5$\sigma$. For Gemini data, the results on fainter targets from our run in July 2012 generally showed two image artifacts that were undesirable in the final reconstructed image: a faint cross pattern centered on the target, and correlated noise patterns over length scales of $\sim0.05-0.10\arcsec$. These effects can combine to give a detectability curve with non-Gaussian distribution of peak heights and/or a non-monotonic nature as a function of separation. We have studied these two effects in the Fourier plane and developed two strategies to reduce their appearance in reconstructed images. First, the cross pattern on the image plane maps to a cross on the Fourier plane, which can be cleanly seen in a region beyond the diffraction limit and removed by replacing the pixel values in the cross with an average of pixel values on either side. Second, the correlated noise appears to be reduced when the point source used to do the deconvolution step is a better match to the point spread function of the target star. Therefore, we have developed an algorithm to “fine-tune” the shape of our point source observation in the Fourier plane based on estimating the difference in dispersion expected for the point source observation and the science target (which is a function of observation time and sky position), and calibrating out the point source dispersion accordingly. These techniques appear to yield reconstructed images which are free of the cross and whose noise peaks have a more Gaussian distribution. Near-IR AO Imaging ------------------ Ten of the KOIs were observed with near-IR adaptive optics (AO) in the $J$, $K^\prime$ and $Ks$ filters either at the Lick Observatory Shane 3.5 m, the Palomar Observatory Hale 5 m or the 10m Keck-II Telescope (see Table \[table:AOobservations\]), as part of a general infrared AO survey of KOIs [e.g., @roweetal14; @marcyetal14; @adamsetal12]. Targets observed with the Lick, Palomar, or Keck AO systems utilized the IRCAL [@lloydetal00], PHARO [@haywardetal01], or NIRC2 [@wizinowichetal04; @johanssonetal08] instruments respectively. The observations were made in the $J$ filter for the Lick observations, the $J$ and $Ks$ filters for the Lick and Palomar observations, and the $K^\prime$ filter for the Keck observations. The targets themselves served as natural guide stars and the observations were obtained in a 5-point quincunx dither pattern at Lick and Palomar, and a 3-point dither pattern at Keck to avoid the lower left quadrant of the NIRC2 array. Five images were collected per dither pattern position, each shifted $1\arcsec$ from the previous dither pattern position to enable the use of the source frames for creating the sky image. The IRCAL array is $256\times256$ with 75 mas pixels and a field of view of $19.2\arcsec\times19.2\arcsec$, the PHARO array is $1024\times1014$ with 25 mas pixels and a field of view of $25.6\arcsec\times25.6\arcsec$, and the NIRC2 array is $1024\times1024$ with 10 mas pixels and a field of view of $10.1\arcsec\times10.1\arcsec$. Each frame was dark subtracted and flat fielded and the sky frames were constructed for each target from the target frames themselves by median filtering and coadding the 15 or 25 dithered frames. Individual exposure times varied depending on the brightness of the target but typically were $10-30$ seconds per frame. Data reduction was performed with a custom set of IDL routines. Aperture photometry was used to obtain the relative magnitudes of stars for those fields with multiple sources. Point source detection limits were estimated in a series of concentric annuli drawn around the star. The separation and widths of the annuli were set to the FWHM of the primary target point spread function. The standard deviation of the background counts is calculated for each annulus, and the $5\sigma$ limits are determined within annular rings [see also @adamsetal12]. The PSF widths for the Lick, Palomar, and Keck images were typically found to be 4 pixels for the three instruments corresponding to $0.3\arcsec$, $0.1\arcsec$, and $0.04\arcsec$ FWHM respectively. Typical contrast levels are $2-3$ magnitudes at a separation of 1 FWHM and $7-8$ magnitudes at $>5$ FWHM with potentially deeper limits past 10 FHWM. An example of AO imaging done at Palomar toward KOI 1964 is shown in Figure \[Fig:KOI1964\_4filters\]. This study includes observations in both $K^\prime$ and $Ks$ filters. The $K^\prime$ filter differs only slightly from $Ks$ (with central wavelengths of $2.12~{\mu{\rm m}}$ and $2.15~{\mu{\rm m}}$ respectively). Because of this, the differential magnitudes of stars measured in either filter are treated as equivalent since any differences are expected to be slight. For calculations and modeling, the $Ks$ bandpass is used. Spectroscopy at NOAO Mayall 4m {#subsec:spectroscopy} ------------------------------ Most of the KOI host stars (16 of 18) were observed spectroscopically at the National Optical Astronomy Observatory (NOAO) Mayall 4m telescope at Kitt Peak during the 2010 and 2013 observing seasons. Table \[table:spectraobs\] lists the 11 spectra actually used to determine stellar properties (other stars were too cool, too hot, or have published asteroseismology measurements of stellar properties as discussed in §\[subsec:literatureproperties\]). The stars were observed with integration times of $5-15$ minutes using the long-slit spectrograph RCSpec setup to disperse the spectra with 0.072 nm pixel$^{-1}$ at a nominal resolution of $\delta\lambda = 0.17$ nm. The wavelength coverage with the best calibrated fluxes was approximately $380-490$ nm. More details of this observing program are discussed in @everettetal13. Spectral frames are reduced in the manner described by @everettetal13. Briefly, the overscan bias is subtracted and trimmed off each frame. Bias frames and flat field frames are then combined, with outlier rejection, to form a master residual bias image and flat. These master frames are applied to each observation in the usual manner. Stellar spectra are extracted using an aperture that traces the stellar image across the CCD and sky-subtracted using night sky spectra extracted from areas of the slit containing sky. Wavelength calibration is provided by an arc lamp exposure at each pointing and flux calibration is done using an observation of a spectrophotometric standard star along with a Kitt Peak extinction curve scaled to the airmass of each observation. Since focus changes significantly across the CCD, only the best focused portion of the spectrum is used for analysis ($\lambda=460-489$ nm where the focus is tight and important spectral features like H$\beta$ are found). PROPERTIES OF THE KOI STARS {#sec:stellarproperties} =========================== The properties of the candidate host stars are estimated in a number of ways. For most stars, a newly acquired spectrum, taken at the Mayall 4m telescope, is available as discussed in §\[subsec:spectroscopy\]. In other cases, values are obtained from the literature and are variously based on asteroseismology, photometry, or spectral analysis used in conjunction with light curve fits. Of all stellar properties, the radius is the most fundamental for characterizing transiting exoplanets because it is used to derive the planet radius. It is worth noting that a number of the candidate host stars have neighboring stars closeby. When the apparent separations are small enough, the neighbors can affect both the follow-up photometry and spectroscopy as flux from the neighbor is introduced into the data. However, in most cases the neighbors are at least several magnitudes fainter and so the contamination is slight. To determine the properties of both the KOI star and its neighbors, we take a two-step approach: First, the properties of the KOI star are established from asteroseismology, if available, otherwise spectroscopy or, lastly, photometry when that is the only available source. Second, once the properties of the KOI star are established, the properties of the neighbors are estimated photometrically as will be discussed in § \[subsec:neighborproperties\]. In most cases, the photometry of the neighbors is measured relative to the KOI star, so determining the properties of the neighbors depends on first characterizing the KOI star. New Spectroscopic Properties ---------------------------- In the case of the KOIs observed spectroscopically at the Mayall 4m telescope, an estimate for [$T_{\rm eff}$]{}, [$log(g)$]{} and [$\rm [Fe/H]$]{} is made in the manner described in detail by @everettetal13. Very briefly, each spectrum is iteratively fit to a grid of synthetic model spectra taken from @coelhoetal05, who parameterized their models using these three properties. The spectral models of @coelhoetal05 are based on the stellar atmosphere models of @castellikurucz03, and were chosen by @everettetal13 from among the publicly available model spectra for their well-sampled grid in parameter values. The model fitting method is calibrated using a set of similar spectra taken of test stars whose properties were well known [*a priori*]{}. Parameter uncertainties for this method are based on the degree to which the fitted properties of the test star set matched their [*a priori*]{} values. The [$T_{\rm eff}$]{}, [$log(g)$]{} and [$\rm [Fe/H]$]{} values from these spectra are listed in Table \[table:stellarproperties\] and marked as coming from Reference 1 or 3. A mass, radius and luminosity is determined later for these stars based on isochrone fits (see §\[subsec:isochronefits\]). Properties from the Literature {#subsec:literatureproperties} ------------------------------ For some KOIs, we have no 4m spectrum or the star was such that it could not be fit (these spectral fits were reliable only within the effective temperature range $4750K<T_{\rm eff}<7200K$). For these stars, values of [$T_{\rm eff}$]{}, [$log(g)$]{}, and [$\rm [Fe/H]$]{} are taken from the literature. The values adopted (in Table \[table:stellarproperties\]) are those listed in the stellar properties catalog of @huberetal14 which contains “best available” properties for almost all of the stars targeted by [[*Kepler*]{} ]{}. It includes properties of very well characterized stars alongside those based on photometry alone (generally the least reliable method of characterization). For those stars with only photometry, like the hot star KOI 3204, @huberetal14 derive new stellar properties, first by identifying any giants using asteroseismology, then finding [$T_{\rm eff}$]{} from the available photometry. They determine other parameters with a Bayesian statistical analysis that includes empirically-motivated priors on [$\rm [Fe/H]$]{} and [$log(g)$]{} that help constrain photometric fits of model spectra to optical and near-infrared colors. For other stars, @huberetal14 rely on existing data as inputs to the Bayesian analysis. The properties of the cool stars KOI 3255 and KOI 3284 are calculated based on photometrically-derived properties from @pinsonneaultetal12 and @dressingcharbonneau13 respectively. For KOI 1964, the constraints are provided by @batalhaetal13 and are based on the light curve and spectroscopic fitting techniques described by @buchhaveetal12. Three of the KOI stars (KOIs 268, 274, and 1537) have been analyzed both asteroseismologically and spectroscopically by @huberetal13 who provide quite accurate and precise values for [$T_{\rm eff}$]{}, [$\rm [Fe/H]$]{}, [$log(g)$]{}, $R_\star$ and $M_\star$. For these stars, the mass and radius are the literature values. For all other stars the radii and masses are determined from new isochrone fits described next. Properties from Isochrone Fits {#subsec:isochronefits} ------------------------------ A new isochrone fitting procedure has been developed for this study to determine the stellar properties for both KOI stars and any potentially bound secondaries (see §\[subsec:neighborsbound\] for a discussion of neighboring stars’ properties). For the purpose of isochrone fitting, each KOI star is described by the set of most probable values for the same three properties ([$T_{\rm eff}$]{}, [$log(g)$]{}, [$\rm [Fe/H]$]{}) with a probability distribution described by half of a normal distribution each for the positive and negative uncertainties (which may differ). A set of Dartmouth isochrones [@girardietal05] is constructed using the interpolation software provided with their distribution. The isochrones span an age range between $1-13$ Gyr at 0.5 Gyr intervals and metallicity ([$\rm [Fe/H]$]{}) range between $-0.4$ and $+0.4$ with steps of 0.02 dex with no $\alpha$-element enhancements. To obtain a finely-sampled set of stellar mass points defining each isochrone (where the original isochrones had some large gaps), new points are created using linear interpolation such that the final intervals between successive stellar masses never exceeds 0.02$M_\odot$. To find the properties of the primary star, a probability level is assigned to each mass point in the set of isochrones based on its location in the ([$T_{\rm eff}$]{}, [$log(g)$]{}, [$\rm [Fe/H]$]{}) probability distribution. The mass point with highest probability and the extent of the parameter space mapped out by those points whose probabilities fall inside a certain threshold level define the central values and $1\sigma$ uncertainties in the other stellar properties (e.g., $R_\star$ and $M_\star$ as listed in Table \[table:stellarproperties\] and absolute magnitudes as discussed later). Exceptions to this (for Table \[table:stellarproperties\]) were made for stars with asteroseismology, whose masses and radii are supplied in the available literature. The precision to which stellar radius is estimated varies between the different techniques. Table \[table:stellarproperties\] lists 12 KOIs with stellar radii derived from spectra without the input of asteroseismology. The mean uncertainty in stellar radius for these stars is 16.9% (averaging all plus and minus uncertainties together). There are three stars with stellar radii based solely on photometric colors with a mean uncertainty of 22.8% (with uncertainties varying greatly among the sample). The three stars with properties based on asteroseismology have a much lower mean radius uncertainty of 2.3%, illustrating the impact of this technique. Magnitudes in the 692nm and 880nm Filters {#subsec:dssimags} ----------------------------------------- The Dartmouth isochrones already predict absolute $Kp$, $B$, $V$, SDSS $griz$, $J$, and $Ks$ magnitudes, but not magnitudes for the specialized 692nm and 880nm filters used in speckle imaging. To add absolute magnitudes for the 692nm and 880nm filters to the isochrone data, color$-$[$T_{\rm eff}$]{} relationships are derived that relate these magnitudes to SDSS magnitudes. These color$-$[$T_{\rm eff}$]{} relationships are calculated based on solar metallicity model spectra published by @munarietal05, the filter transmission curves, the QE curve of the DSSI CCDs, an atmospheric extinction curve for Mauna Kea at the typical observing airmass of 1.3, and the AB magnitude system. The color$-$[$T_{\rm eff}$]{} relationships between the SDSS magnitudes and speckle imaging filters are shown in Figure \[Fig:color\_vs\_Teff\]. Because the lowest [$T_{\rm eff}$]{} in the model spectra of @munarietal05 was 3500 K, the color$-$[$T_{\rm eff}$]{} relationship is linearly-extrapolated down to 2750 K (although the lowest [$T_{\rm eff}$]{} actually found in the isochrones is $\sim3000$ K). Additionally, to obtain magnitudes for stars with $log(g)>5$, a $log(g)=5.5$ curve is found by linear extrapolation of the colors predicted at $log(g)=4.5$ and 5.0. These extrapolations are indicated in Figure \[Fig:color\_vs\_Teff\] with light grey lines. For any star defined by [$log(g)$]{} and $T_{\rm eff}$, the absolute magnitudes in the speckle bandpasses can now be found by interpolating between the two bounding curves in the color$-$[$T_{\rm eff}$]{} relationships given their absolute $g$ or $z$ magnitudes. These calculations are done assuming solar metallicity models for each star. Metallicity has a noticeable, but small effect on colors for stars cooler than 4000 K which increases with decreasing effective temperature. At [$T_{\rm eff}$]{}$=3000$ K, there are color differences of $\sim0.02$ in $880 - z$ and $\sim0.15$ in $692 - r$ when comparing [$\rm [Fe/H]$]{}$=-0.5$ models to solar metallicity models. Thus, there is additional uncertainty in modeling fluxes in the speckle filters among cool stars. This mainly impacts a few of the faint neighbor stars discussed in §\[sec:companions\]. FALSE POSITIVE PROBABILITY ANALYSIS {#sec:fpp} =================================== A false positive probability is determined for planet candidates orbiting each KOI star that is sufficiently isolated from detectable neighbors such that the KOI star is the only [*detected*]{} star near the source of the candidate signal. This analysis compares the probability of the KOI being a planet host to the probability that a nearly co-aligned and fainter field star is the source of a false positive signal (ie. an eclipsing binary or transiting planet). The scenario of an unresolved triple KOI in which two components form an eclipsing pair is not considered here because for these it is difficult to calculate some of the complex scenarios for a given system. Scenarios such as these have been considered by @fressinetal13 who found that the incidence of false positives attributable to an eclipsing secondary component in a hierarchical triple stellar system are quite low, especially among candidates of Neptune and smaller planets like in the sample considered here. We estimate the false positive probability for each planet candidate by integrating the parameter space not excluded by [[*Kepler*]{} ]{}data or follow-up observations with respect to a Galactic model. This method is based on the approach described in @barclayetal13 and @wangetal13 [@wangetal14]. The information used to restrict the parameter space is the transit depth, the [[*Kepler*]{} ]{}out-of-transit pixel response function (PRF) centroid statistic and the 692 nm Gemini speckle and any near-IR AO observations of the star. The PRF is the observed appearance of point sources and depends on the PSF produced by the [[*Kepler*]{} ]{}optics, spacecraft jitter, focus, and spectral class of a given point source (although the latter effect is not considered significant enough to treat individually). The measured PRF and its centroid statistic [@brysonetal13], the quarter-by-quarter standard deviation between a stellar centroid in an out-of-transit image and the difference image between in-transit and out-of-transit light curve points are products of the [[*Kepler*]{} ]{}pipeline [@tenenbaumetal13; @tenenbaumetal14]. The transit depth provides a limit on the faintest star that could produce a false positive signal matching the light curve. This comes from assuming a total eclipse by a background eclipsing binary star of identical components, which would produce a 50% eclipse depth. This maximum eclipse depth is adopted under the expectation that for more general binaries with unequal mass components, the larger star will be brighter in the [[*Kepler*]{} ]{}bandpass. For a maximum eclipse depth, the background star, outside of eclipse, would be ${\Delta}Kp_{max}$ magnitudes fainter than the KOI and the observed transit depth can be expressed in terms of $\delta$, the KOI’s fractional transit depth: ${\Delta}Kp_{max}=-2.5{\times}log_{10}(2\delta)$. For example, if the observed transit depth were 100 ppm, this could be induced by a eclipsing binary of at most $Kp=9.25$ magnitudes fainter than the KOI. Our estimates of the transit depth are taken from the [[*Kepler*]{} ]{}data analysis pipeline [@jenkinsetal10a; @tenenbaumetal13; @tenenbaumetal14]. Values for ${\Delta}Kp_{max}$ are listed in Table \[table:validations\]. The [[*Kepler*]{} ]{}out-of-transit PRF centroid statistic is used to set an exclusion radius for each planet candidate. Any star outside this exclusion radius is excluded from being the source of the candidate transit signal because any such source outside this radius would produce a larger centroid statistic. To find the exclusion radius, we use a $3\sigma$ threshold where $\sigma$ is the PRF statistic discussed above. These exclusion radii establish which KOI stars are sufficiently isolated for the analysis as well as restrict the area inside of which false positives are modeled. Values for the exclusion radius, $r_{ex}$, are listed in Table \[table:validations\]. In the case of KOI 2311, no exclusion radius could be determined due to the lack of PRF centroid data on this star. As discussed in §\[sec:companions\], there are 8 candidate planets transiting 5 KOI stars that show a neighboring star within the exclusion radius. These candidates, plus the two of KOI 2311 are excluded from the validation calculation. We then include both AO data and DSSI speckle data – we convert the $K$-band AO data to ${\Delta}Kp$ using the equations of @howelletal12 while we utilize the 692 nm DSSI data and assume no difference between this bandpass and $Kp$. This provides a brightness-dependent limit on the maximum separation between a target star and a false positive inducing star. The relative brightnesses and angular separations of potential background stars excluded by the photometry, centroid statistics and transit depths for 4 KOIs are shown in Figure \[Fig:validations\]. There are 18 candidate planets orbiting 12 stars that qualified for the validation tests. We use the TRILEGAL galactic simulation [@girardietal12] to first estimate the stellar population within $1\arcdeg$ around the target star. We then integrate the region of parameter space not excluded by observations with respect to the population model. This provides a number of false positive stars, which is usually much less than 1. We then estimate what fraction of these are likely to be either background eclipsing binaries or background planet hosts [@slawsonetal11; @burkeetal14]. Finally, we compare the number of false positives like this in the entire [[*Kepler*]{} ]{}data set (i.e. we multiply the number of false positives for this star by the [[*Kepler*]{} ]{}sample size of 150,000 [@kochetal10] with the predicted number of planets like this in the [[*Kepler*]{} ]{}data set [@fressinetal13]). The ratio of the total planets to the total number of planets plus false positives yields the probability that a candidate is a planet. If the planet candidate is in a multi-planet system we boost the odds that the candidate is a planet by a factor of $\sim30$ for two-candidate systems. This multiplicity boost is justified on the basis of statistics done on the [[*Kepler*]{} ]{}sample [@lissaueretal12; @lissaueretal14]. Assuming false positives are randomly distributed among targets, multiple planet systems should not have a higher false positive rate than other targets. It is found that a large fraction of KOIs with at least one planet candidate are found to have at least two candidates, meaning a larger fraction of the planets in these systems are real. Table \[table:validations\] lists the total validation level for each candidate around an isolated star for which PRF centroid data is available. Two of the systems are designated Kepler-430 and Kepler-431 as newly-validated multi-planet host stars hosting a total of 5 planets validated at $\geq99.8$%. The letter designations for these planets are given in the table. Several of the planets around these KOIs have previous validation calculations done or have been deemed validated. @xie14 showed that the two current planet candidates of KOI 274 exhibited anti-correlated transit timing variations, validating both as interacting planets. @wangetal14 used the native seeing UKIRT $J$-band survey images of the [[*Kepler*]{} ]{}field to help calculate validation percentages for the multi-planet KOIs 115, 274, 284, 369, 2311 and 3097, including validation boosting due to their planetary multiplicities. For KOI 115.01 they report a 99.8% validation (considering it part of a 3-planet system in contrast to our study that excludes KOI 115.03 due to low detection significance). Wang et al. also reported validation levels above 90% for KOIs 115.02 and 284.01, along with lower levels for the other candidates. @roweetal14 validated a set of multi-planet KOIs including KOIs 115, 274, 284, and 369 by incorporating various tests on the [[*Kepler*]{} ]{}data and external data products to arrive at $>99$% validation levels for their hosted planets. NEIGHBORING STARS {#sec:companions} ================= Observed Properties of Neighboring Stars {#subsec:neighborproperties} ---------------------------------------- The neighboring point sources (hereafter assumed to be stars) detected around 7 of the KOIs in the high resolution images as well as native seeing survey images are listed in Table \[table:secondaries\]. The table provides the relative separations ($\rho$), position angles ($\theta$), and magnitudes fainter than the KOI (${\Delta}$mag). Here, each neighbor star is given a designation of “B” or “C” as an identifier. These stars could be foreground or background stars that are closely aligned by mere chance with the KOI star or may be gravitationally bound secondaries. The closer and brighter these neighboring stars are, the more likely they are to be gravitationally bound to the KOI, as discussed below. Table \[table:secondaries\] shows that 7 KOI stars have a neighbor $\leq4\arcsec$ away detected in high resolution images. In the case of 5 of these KOIs, one neighbor lies within the exclusion radius for all planet candidates (KOI 268B, 284B, 1964B, 3255B, and 3284B). Such neighbors are potential sources of a false positive (ie. an eclipsing or transiting system that is blended with the KOI). Even very faint neighbors, with magnitudes relative to the KOI star of ${\Delta}Kp=9-10$ can produce a transit-like signal at the level expected for an Earth-sized planet transiting a Sun-like star [@mortonjohnson11]. The more distant neighbors should not be considered possible sources of a false positive. These stars are nonetheless important to consider as possible members in a stellar binary with the KOI and for their dilution of the KOI light curves. To correct for dilution, a search was made for any star that could possibly dilute the light curve at a 1% level or greater. Both the high resolution images and other ground-based imaging surveys such as the UKIRT $J$-band survey and a catalog of $UBV$ photometry by @everettetal12 were used. Some neighbors are left out of this search, namely any listed in the KIC, because any excess flux blended into the light curve by KIC stars will have already been estimated and removed by the [[*Kepler*]{} ]{}pipeline. Two significant KIC stars were noted near the KOIs in this sample: KIC 11560901, about $8.5\arcsec$ away from KOI 2365 (Kepler-430), and KIC 8212005, about $12.9\arcsec$ away from KOI 2593. Previous High Resolution Imaging -------------------------------- Several of these stars have been previously observed with high-resolution imaging by @adamsetal12 [@adamsetal13], @lawetal14 and @lilloboxetal14. KOI 1537 was reported by @adamsetal13 to have a close neighbor with a separation of $0.13\arcsec$ in $Ks$ AO images. The same KOI was observed by @lawetal14 in an optical AO (RoboAO) survey of KOIs, but no neighbor was found. Their non-detection would be consistent with the close separation and RoboAO imaging resolution. No neighbor of KOI 1537 is detected in our speckle data. This is surprising given the high spatial resolution, which would easily resolve the separation, and the small magnitude difference ${\Delta}Ks=0.15$ reported by @adamsetal13. The limiting contrast for detecting neighbors at a separation of $0.13\arcsec$ in the 692 nm and 880 nm images is 5.37 and 4.69 magnitudes respectively. These apparently discrepant observations cannot be attributed to a hypothetical very red neighbor. The color $880-Ks=1.08$ is found for KOI 1537 from its isochrone fit, meaning any line-of-sight neighbor would need to be at least as red as $880-Ks\simeq5.6$, or redder than the reddest stars ($880-Ks=3.9$) found in the isochrones of §\[subsec:isochronefits\]. In light of this observation, KOI 1537 is treated as a single star here. @adamsetal12 observed KOIs 268 and 284 and detected the neighbors of both stars with $J$ and $K$ magnitude differences in agreement with the values published here. (Note that the position angles they reported for KOI 268 are apparently erroneously flipped.) KOI 115 was observed by both Law et al. and Lillo-Box et al. It was seen as single by Law et al., in agreement with our data. Lillo-Box et al., who observed with a larger field-of-view, reported a neighbor to KOI 115 about $4\arcsec$ away and 8 magnitudes fainter in $i$. Law et al. and Lillo-Box et al. also observed KOI 2593. Both reported it as a single source. In addition to KOIs 115, 1537 and 2593, three other stars (KOIs 268, 1964, and 2365) were observed by Law et al. in the optical and, in the case of KOI 1964, in a near-IR $Ks$ image. They reported the closest neighbor to KOI 268 to have an optical magnitude difference of $3.82\pm0.27$, which is consistent with our near-IR magnitude differences for a neighbor redder than the KOI. Their $Ks$ observations of KOI 1964 were in good agreement with ours. They found KOI 2365 (Kepler-430) to be single, again in agreement with our results. Distinguishing Bound Companions from Field Stars {#subsec:bg_vs_field} ------------------------------------------------ Various evidence may be used to determine if neighboring stars are gravitationally-bound secondaries or unrelated, line-of-sight field stars. A full simulation of the properties and frequency of secondary stars and field stars could be used. Instead, in this study, a series of simpler tests are applied. These tests consider the brightness of the neighbor, its angular separation from the KOI star, and the stellar colors for those stars observed at multiple wavelengths. Other approaches to using multi-color photometry to investigate the possible physical association of neighbor stars with KOI stars may be found in @gillilandetal14 and @lilloboxetal14. ### Angular Separation and Apparent Brightness {#subsubsec:bgprobabilities} First, the angular separation from the KOI star and magnitude of the neighbors are examined relative to a random distribution of field stars at the location of the KOI. In doing so, an initial assumption is made that KOI stars are not preferentially co-aligned with unrelated field stars. A randomly-generated set of stars representing a 1 square degree field is produced at the location of the KOI using the TRILEGAL Galaxy model [@girardietal05] Version 1.6 web form. The model predicts apparent magnitudes in various passbands including $J$ and $Ks$. The number of stars in the TRILEGAL model brighter than the neighbor star is found and multiplied by the ratio of the circular area inside the neighbor’s separation ($\rho$) to the 1 square degree model field to get a “background” probability, $P_{BG}$. $P_{BG}$ is the likelihood that a field star of the same brightness or brighter than the neighbor would lie by chance at the same or a smaller angular separation from the KOI. In most cases the $Ks$ bandpass is used for this calculation because it yields the lowest probabilities. To find approximate $Ks$ magnitudes for the neighbors, the differential photometry of the AO images is used along with the $Ks$ magnitude of the associated 2MASS point source [@skrutskieetal06]. For neighbor “C” of KOI 3284, the $J$ magnitude of the UKIRT survey was used instead due to unavailable $K$-band data. The background probabilities along with the apparent $Ks$ magnitudes derived for each source are listed in Table \[table:isofits\]. As an empirical check on this method, the calculation was also run using [[*Kepler*]{} ]{}magnitudes of sources extracted from 1 square degree of the KIC at the location of each KOI. The probabilities determined from the KIC agreed with those of the TRILEGAL model to within a factor of 2 (some were higher, others lower). Note that in many cases $P_{BG}$ is quite low, a promising indication that the neighboring stars are gravitationally-bound companions. The expectation is that gravitationally-bound secondaries outnumber co-aligned neighbors in high resolution images such as these, especially within separations of $\sim1.2\arcsec$ [@horchetal14]. For this reason, these close neighbors are likely dominated by bound companions. However, false positive scenarios for small planet candidate KOIs include the case of large planets transiting background (or field) stars that are closely co-aligned with the KOIs. These cases can appear much like the observed double KOI sources we have detected in terms of relative magnitude and separation [@fressinetal13]. For these cases, the assumption that background stars are randomly-distributed in the sky is invalid and the low values of $P_{BG}$ are best treated as just one of several indicators that help distinguish between field stars and bound companions. These probabilities are most applicable for neighbor stars that lie outside the exclusion radius. On the other hand, a low value of $P_{BG}$ for a neighbor star inside the exclusion radius can be explained as either a bound companion or a false positive. ### Colors and Relative Brightness of Neighbors {#subsubsec:colormagcomparison} Both stars of gravitationally-bound pairs should lie on the same isochrone and this can be tested for KOIs that have been observed at multiple wavelengths. The test relies on the relative brightnesses and colors of the two stars. To determine these, the isochrone fits are used to find the colors of the KOI stars while the differential photometry of the imaging provides relative colors and brightnesses of the neighbor. Table \[table:neighborcolors\] lists the relative brightnesses and colors for the double KOI sources that have been observed in more than one filter. The magnitudes in the table are absolute magnitudes for the KOI stars and likewise for their neighbor stars if they are gravitationally bound. Figure \[Fig:isochrones\] compares the magnitudes and colors of 6 KOI stars and their close neighbors alongside the isochrones describing the KOI star properties. The colors and magnitudes of each KOI star (within its $1\sigma$ uncertainty range) are indicated by dark grey regions in each panel. The light grey regions show the set of isochrones that pass through these ranges of uncertainty. The magnitudes in each plot represent absolute magnitude as predicted by the Dartmouth isochrones or, in the case of the 692 nm and 880 nm filters, calculated from the isochrones’ SDSS magnitudes as described in § \[subsec:dssimags\]. The relative magnitude and colors of the neighbors with respect to the KOI stars are calculated using the relative photometry provided by the speckle imaging analysis and near-IR AO images. The neighboring stars’ colors and magnitudes are shown as rectangular boxes that indicate the $1\sigma$ photometric uncertainties. In most panels of Figure \[Fig:isochrones\] the relative colors and brightnesses of the close neighbor stars are consistent with the isochrones describing the KOI stars. In other words, most of the neighbors are consistent with being gravitationally-bound secondaries. However, in five panels ($g$, $h$, $l$, $m$ and $n$) the neighbor falls quite far from the isochrones as would be expected in the case of most field stars. The neighbors of KOI 1964 (panel $g$) and KOI 2311 (panel $h$) are fainter and/or bluer in these colors than main sequence stars at the distance of the KOI, but this situation is not seen in other plots for the same KOIs (panels $e$, $f$ and $i$). Neighbor C to KOI 3284 is too faint or blue relative to the Main Sequence in both colors examined. In the case of neighbor B to KOI 3284 (panel $l$), the neighbor is too red relative to the Main Sequence. It is also too faint to be a Milky Way giant. In this case, the neighbor is so red and faint relative to KOI 3284 (itself a M dwarf) that, if a bound companion, its luminosity would place it at an extremely low mass where the model colors are most uncertain. This case should be treated with some caution. ### Color Relative to Background Population To compare the colors of the close neighbors of 6 KOI stars to the colors of field stars of the same apparent brightness, the TRILEGAL Galaxy models are used again (the same 1 square degree field populations used in §\[subsubsec:bgprobabilities\]). This time the field star populations are restricted to those stars within one magnitude of the apparent magnitude of the neighbor star in the bluer of two filters being considered. The number of field stars is plotted as a function of colors in Figure \[Fig:backgroundcolors\]. The colors of the neighboring stars are indicated by vertical lines (solid lines represent the central value and dotted lines the $1\sigma$ uncertainty interval). The color distributions of the field stars show several features. In $J - Ks$, the peak near 0.25 is due to large numbers of upper main sequence plus turn-off stars, a second peak near 0.6 is due to giants, and some plots show a peak at 0.75 due to lower main sequence dwarfs. The same features are seen in $692 - Ks$ at 1.1, 1.8, and 2.25 respectively. The upper Main Sequence plus turn-off stars and giants show up as peaks near 0 and 0.15 respectively in $692 - 880$. The red tails of the $692 - 880$ and $692 - Ks$ colors are comprised of the lowest mass dwarfs. The field star color distributions are affected by reddening while the colors derived for the secondary stars are intrinsic colors (zero reddening effects). However, extinction is quite small in the TRILEGAL models where $A_V=0.03-0.04$ to distant lines of sight in the [[*Kepler*]{} ]{}field and so reddening corrections in the $692 - Ks$ color would be only $0.02-0.03$ magnitudes using the extinction curve of @cardellietal89. For this reason, no adjustment for the effects of reddening has been made. In each panel of Figure \[Fig:backgroundcolors\], the colors of the neighbor stars are either consistent with the bulk of the field star distribution or redder than it. Assuming field stars are distributed randomly on the sky around each KOI, a first order expectation is that their colors will be drawn from this same distribution. For this reason, the relatively red colors of the neighbors seen in panels $e$, $f$, $i$, $k$ and $l$ (ie. KOI 1964B, 2311B, 3255B and 3284B) are best explained by low-mass gravitationally-bound secondaries (although the extreme color measured for KOI 3284B was difficult to explain as discussed earlier). Two other neighbor stars, KOI 268 B and C (panels $a$ and $b$) are also quite red, but so too are more field stars in $J - Ks$. ### Assessing the Nature of Each Neighbor Star None of the observations definitively distinguishes between gravitationally-bound or field star neighbors, but in most cases the evidence points toward the close neighbor being a gravitationally-bound secondary. Values of $P_{BG}$ are quite small for all but the three most distant neighbors (KOI 3255C, KOI 3284C and KOI 4407C), which means these have a reasonable likelihood of being nearby field stars. KOI 1964B and KOI 2311B also show some evidence of being field stars based on some colors ($J - Ks$ in the case of KOI 1964B and $692 - Ks$ in the case of KOI 2311B). However, KOI 1964B is less conclusive because in the other colors examined it appears to be a relatively low-mass, red, bound companion. The photometry for KOI 3284C is inconsistent with that of a binary companion, but is internally consistent with that of a background dwarf (see § \[subsubsec:fieldstarprops\]). The other 5 neighbors with multi-band photometry (KOIs 268B, 268C, 284B, 3255B and 3284B) are the most consistent with being bound companions. KOI3255C has photometry in $J$ only and KOI 4407B and C in $Ks$ only, so their natures remain indeterminate. Overall, among the 11 neighbor stars in this study 8 have multi-band photometry. Of those 8, five are deemed likely bound companions, one a likely field star, and the two others remain too ambiguous to classify. Based on this, the fraction of likely bound companions may be as high as 87.5% or as low as 62.5%. This can be compared to the lucky imaging survey of 174 candidate or confirmed KOI host stars by @lilloboxetal14. Among their targets observed in both $i$ and $z$ filters, five were found with close companions, but they considered only one of them to be a bound companion. The significance of the lower fraction of bound companions is difficult to quantify, but given the larger mean separations for the companions in the Lillo-Box et al. survey, a greater fraction of field stars is reasonable. In a study of 23 KOIs observed in two filters with [ *HST/WFC3*]{}, @gillilandetal14 quantified the odds for neighboring stars to be the bound companions of KOIs. They found 8 neighboring stars were physically associated with the target KOI, and 6 of these had relatively close separations of $<1\arcsec$. Clearly, high resolution imaging inside of $\sim1\arcsec$ is needed to find most of the wide binary companions to KOIs. Blending Corrections for Crowded KOIs {#subsec:neighborsbound} ------------------------------------- In order to correct the light curves for the effects of blending, neighbors’ stellar properties must be estimated. Two separate scenarios are considered for the status of these neighbor stars: bound companions and unrelated field stars. For completeness, and because most of the neighbor stars are consistent with being gravitationally-bound secondaries, a calculation is made for each neighbor assuming it is gravitationally bound. In the case of binaries, the secondary star properties are more easily constrained by the observations because each component shares a common distance, extinction, composition and age. The second case, that of neighbor stars being unrelated field stars, is considered for a few cases where evidence suggests or allows this scenario. Constraining the properties of field stars can prove more difficult. To find the stellar properties of assumed secondary stars, the relative photometry is used for isochrone fits as described in § \[subsubsec:secondaryprops\]. To find properties of field star neighbors, the most likely types of stars are identified in Galaxy models as discussed in § \[subsubsec:fieldstarprops\]. For either case, an aperture correction is found for each star in § \[subsubsec:apcorrections\]. Finally, a blending correction, is formulated in § \[subsubsec:blendcorrections\] based on the $Kp$ magnitudes and aperture corrections for each star in the blend. These corrections are used to reevaluate the planet properties for crowded KOIs as described in § \[sec:planetproperties\]. ### Properties Assuming Neighbors are Secondary Stars {#subsubsec:secondaryprops} Each image taken in a different filter provides an independent measurement of the relative brightnesses of the secondary and primary stars. The magnitude differences, along with their observational uncertainties, map the probability distribution of the primary star properties along a set of isochrones into a distribution of secondary star properties. Figure \[Fig:KOI1964isofit\] shows the distribution of primary and secondary star properties in [$log(g)$]{} and [$T_{\rm eff}$]{} for the double source KOI 1964 observed in four different filters. Curves outline the range of isochrones used in the model fits. When relative magnitudes are available in multiple filters, the secondary properties derived from each are combined in a weighted average to obtain a final estimate. One of the most important secondary properties is $Kp$ because it helps to determine the excess flux contributed to the light curve by the secondary. Table \[table:isofits\] lists both apparent and absolute $Kp$ magnitude from the isochrone fits for each multiple KOI (apparent $Kp$ are mean values calculated from multiple filters and the absolute $Kp$ values are individual values for each filter). [[*Kepler*]{} ]{}magnitudes are reported in the KIC for each of the blended KOIs considered here. In cases where neighbors lie $\leq1.5\arcsec$ from the KOI, the KIC magnitude is assumed to be the blend of each component whereas more distant neighbors are assumed to not contribute to the cataloged magnitude of the KOI. Table \[table:isofits\] also contains the number of standard deviations each individual filter’s results are from the mean $Kp$. These numbers indicate how well data from different passbands match expectations for a bound companion. The photometry of both neighbors to KOI 268 and the neighbors of KOIs 284 and 3255 agree well with expectations for bound secondaries. The photometry for the neighbor of KOI 1964 agrees less well, but is plausibly consistent with a bound companion. The photometry for the neighbors of KOI 3284 and KOI 2311 is inconsistent with a bound companion. This result is similar to the previous analysis indicating these neighbors are likely to be field stars. Table \[table:blendedstars\] gives the averaged values of ${\Delta}Kp$ for each KOI and neighbor based on the individual photometric measurements of each system. This table lists separate values for different planets in multi-candidate systems since the blending situation will later be considered separately for each planet. ### Properties Assuming Neighbors are Field Stars {#subsubsec:fieldstarprops} As discussed in § \[subsec:bg\_vs\_field\], there are three neighboring stars with multi-color photometry that show colors plausibly inconsistent with those of a gravitationally-bound companion (KOI 1964B by its $J-Ks$ color, 2311B by its $692-Ks$ color and 3284C in all colors). To determine what types of field stars match the observed brightness and colors, the TRILEGAL Galaxy models are used once again. This time the area covered by the Galaxy models is increased to 4 square degrees for KOI 1964 and 10 square degrees for KOI 2311 to ensure a rich sample of model stars. The subset of model stars whose apparent magnitudes lie within 1 magnitude of the neighbor star and whose colors fall within the observed uncertainty intervals listed in Table \[table:neighborcolors\] are extracted. Here, the apparent magnitude adopted is that of the bluer filter and each subset contains $\sim200$ stars or more. The field stars are comprised of either dwarfs, evolved stars or both in various cases. For KOI 1964B, evolved stars (with [$log(g)$]{}$<4.0$) are dominant with dwarfs comprising just 13 out of the 284 total model stars (4.6%). For KOI 2311B the situation is much different with dwarfs accounting for 11771 out of 11825 model stars (99.5%) and for KOI 3284C the model contained no evolved stars among the 264 stars matching $B-V$ or the 195 stars matching in $V-J$. A similar analysis is done on the three neighbor stars with photometry available in only one filter. For these stars, KOI 3255C and KOI 4407B and C, the subset of stars drawn from the TRILEGAL model is unconstrained by color and representative of all field stars matching the brightness of the neighbor. For KOI 3255C, potential background stars are almost entirely dwarfs (32158 out of 32724 stars or 98.3%). Evolved stars are more likely as background cases for KOI 4407B (only 620 out of 1786 or 34.7% are dwarfs) and for KOI 4407C (12363 out of 14706 or 84.1% are dwarfs). While the stellar properties of the potential background stars can vary greatly (e.g., different luminosity classes or stellar radii), the most important property to examine is the relative [[*Kepler*]{} ]{}magnitude, which determines the amount of blending. ${\Delta}Kp$ is largely a function of effective temperatures and the TRILEGAL stars of all luminosity classes are combined for its determination. For cases where a presumed field star fell within the exclusion radius (field stars that could be false positive sources), a careful consideration of its other stellar parameters would be needed, but this situation was not encountered in our sample. Figure \[Fig:fieldstarmags\] shows the distributions in ${\Delta}Kp$ for the 6 potential background neighbors and Table \[table:blendedBGstars\] gives the mean differences in [[*Kepler*]{} ]{}magnitude for each case. Note that the table lists the status of two neighbors as bound companions (KOI 3255B and KOI 3284B) and for these the [[*Kepler*]{} ]{}magnitudes are the same as in Table \[table:blendedstars\]. It is the second neighbors (C) of these KOIs that are treated as field stars. Figure \[Fig:fieldstarmags\] shows the obvious differences in the distributions of possible ${\Delta}Kp$ values between those stars with color information and those without it. With constraints on the color, the uncertainty on ${\Delta}Kp$ is as low as $0.05-0.10$ magnitudes. This is true even for KOI 1964B where the color comes from near-IR measurements. For those neighbors observed in only one filter, the uncertainty in ${\Delta}Kp$ is at least 2 magnitudes. This uncertainty is partly due to the near-IR wavelengths of these observations; a single photometric measurement in the 692 nm filter, for example, would yield an uncertainty in ${\Delta}Kp$ of $\sim0.2$ magnitudes due to its closer match with the [[*Kepler*]{} ]{}bandpass. ### Aperture Corrections {#subsubsec:apcorrections} The fraction of each star’s flux that falls in the light curve aperture is determined using the [[*Kepler*]{} ]{}Mission data analysis tools of PyKE [@stillbarclay12], slightly modified for these purposes. The set of pixels in [*Kepler*]{}’s pixel light curve files is analyzed at epochs coinciding with transits calculated from each candidate’s ephemeris. The average of the effects over all epochs should represent the effects of blending in the folded light curves analyzed for planet properties. For KOIs with orbital periods longer than 15 days, each transit time is examined. For shorter orbital periods, fewer transit times are examined for purposes of efficiency (e.g., every 2nd or 3rd transit). [[*Kepler*]{} ]{}targets shift pixel location with time, however the most significant differences occur between different [[*Kepler*]{} ]{}quarters as the spacecraft rolls by $90\arcdeg$ and new light curve apertures are used on different CCDs. Because the KOI stars are the brightest star in each aperture, and sampled both in the core and the wings, the PyKE tool [*kepprf*]{} is used to fit its PRF to the [[*Kepler*]{} ]{}data and report back a source center and the fraction of the flux that falls inside the aperture. For secondary stars, pixel coordinates are fixed relative to the KOI based on the ground-based astrometry. The percentage of the pixel response function of each secondary that falls inside the aperture is determined and reported in Tables \[table:blendedstars\] and \[table:blendedBGstars\]. Where these numbers are reported for multiple planet candidates of the same KOI, the quite small differences in flux may be seen. ### Corrections for Blended Light Curves {#subsubsec:blendcorrections} A blending correction must be made to properly interpret the light curves of blended KOIs and revise the planet properties derived from them. Furthermore, because for some blended KOIs, more than one star could be the source of the transit-like variations (ie. be the host star), the effects of blending are considered with respect to each star. The goal is to describe an intrinsic light curve for each possible host star. The amount of dilution in these light curves and its effect on a key [[*Kepler*]{} ]{}measurement, the intrinsic fractional depth of a transit signal, $\delta_{true}$, may be written as: $$\label{Eq:AA} \delta_{true,i} = \delta_{obs} \left( 1+\frac{1}{F_i}\sum_{\stackrel{j=1}{j\neq i}}^{N}F_j \right).$$ Here, the transit (or eclipse) is assumed to occur to the $i$th star, and the observed fractional transit depth is $\delta_{obs}$. The intrinsic fractional transit depth of the $i$th star, $\delta_{true,i}$ is found from $\delta_{obs}$ using a blending correction factor dependent on the flux of all stars in the blend. Similar treatments were adopted by @lawetal14 and @lilloboxetal14. Tables \[table:blendedstars\] and \[table:blendedBGstars\] list the blend corrections calculated from Eq. \[Eq:AA\] ($\delta_{true}/\delta_{obs}$) using the $Kp$ magnitude for each component converted to relative flux and with the fluxes modified by the aperture corrections of § \[subsubsec:apcorrections\]. It should be noted that of course these blend corrections are based on the mean predicted values of ${\Delta}Kp$ and are subject to its uncertainties which, for cases of field star neighbors observed in a single filter, can be quite large. PLANET PROPERTIES {#sec:planetproperties} ================= Revised planet properties are derived based on the new stellar properties and corrections for blending in the light curves of crowded KOIs. All of the KOIs have “best” current values for various stellar and planetary properties, which can be found in the Cumulative KOI database at the NASA Exoplanet Archive[^3] or the stellar properties catalog of @huberetal14, which includes the stellar masses. A first order revised planet radius may be found by scaling the current value of the planet radius by a factor equal to the ratio of the new stellar radius to the current stellar radius (one used by the mission for its current light curve fit). Such a scaling is ideal for cases where the revisions to stellar radius are small, although here it is applied to all cases for illustrative purposes. For KOIs with close secondaries, multiple scenarios are considered wherein the transiting body may orbit the secondary. In this case, “close” secondaries are only those stars falling inside the exclusion radius around each KOI (so could be the planet host; see §\[sec:fpp\]). Table \[table:planetproperties\] lists the isochrone fit values for stellar radii, masses and effective temperatures for each KOI star as discussed in §\[sec:stellarproperties\]. In this table, the same properties listed for the neighbor stars represent those derived from isochrone fits under the assumption that the neighbors are gravitationally bound as described in (§ \[subsubsec:secondaryprops\]). Also listed are transit (ie. planet orbital) periods, $P_p$, and the planet-to-star radius ratio, $R_p/R_\star$, a parameter found from light curve fits and taken from the cumulative KOI database. This radius ratio, when multiplied by the new stellar radius, scales the planet radius in accordance with the revised stellar radius. The other scaling done to the planet radii accounts for the blends; here, the value $\sqrt{\delta_{true}/\delta_{obs}}$, derived from Equation \[Eq:AA\] (see Tables \[table:blendedstars\] and \[table:blendedBGstars\]), is found and applied as a multiplicative factor that increases planet radii. Two (generally) different values of the revised planet radii are given in Table \[table:planetproperties\] representing the case where each neighbor is assumed to be a binary companion star ($R_{p{\rm BIN}}$) and the case where at least one neighbor is assumed to be a background star ($R_{p{\rm BG}}$) as listed in Table \[table:blendedBGstars\]. Cases where the KOI arises due to transits of an object orbiting a neighboring field star are not examined because such situations are complicated by large uncertainties in determining the [[*Kepler*]{} ]{}magnitude of some neighbors and the often wide, bimodal distributions in background star radii (ie. they may be either dwarfs or evolved stars). Note too that a complete reevaluation of the planet radii could be performed from new light curve fits, but such reanalysis lies beyond the scope of this study. The magnitude of the “deblending” factor varies by KOI. Consider, first of all, the KOI star as the host. For KOI 2311, the secondary is relatively faint so its effects on the planetary radii are negligible. In KOIs 268, 1964, and 4407, estimates for new planet radii are $1-5$% higher after deblending. For KOI 3284, the radius is 8% higher and for KOIs 284 and 3255, the planet radii need adjustment upwards by about 30%. In cases where a secondary star is the potential host star (making the KOI a false positive under our definition), the effects on planet radius can be even larger and perhaps be large enough to rule it out as a planet. This can be seen in a comparison between the derived radius of the same planet assuming the KOI as the host star versus the secondary star as the host. In the case of KOI 268.01, the radius is 3 times larger if it orbits the secondary (it is a giant planet rather than sub-Neptune size). For KOI 2311.01 and 2311.02, planet sizes change from Earth-like to more like that of Neptune. Finally, it is notable that in none of the cases would a candidate orbiting one of these secondary stars require a radius exceeding that of a giant planet (ie. require a stellar eclipse). These examples make it clear that understanding the role of light curve blends alongside uncertainties in stellar properties is vital for understanding the [[*Kepler*]{} ]{}planet sample. New stellar masses and effective temperatures invite a recalculation of planet equilibrium temperatures, an indicator of planet habitability. Various assumptions are made in the calculation: circular orbits and planets with an albedo of 0.3 and uniform surface temperature. The planet transit (ie. orbital) periods, $P_P$ are used to find the semi-major axis, $a$, of each orbit (given the stellar mass) and this establishes the equilibrium temperatures, $T_{eq}$, listed in Table \[table:planetproperties\]. Examination of the equilibrium temperatures reveals the low-$T_{eq}$ candidates selected for this study. It also shows that lower values of $T_{eq}$ are expected for planets orbiting the potential secondaries as opposed to the KOI stars. Initial and revised planet radii and equilibrium temperatures are plotted in Figure \[Fig:planetchanges\]. In this case, each planet is assumed to orbit the KOI star. Corrections for the revised stellar properties and those for deblending are shown separately. For this plot, “initial” stellar properties are adopted from the catalog of @huberetal14, in the case of those stars with KPNO 4m spectra taken as part of this study, or are the literature values cited in Table \[table:stellarproperties\] otherwise. The plot shows the considerable corrections due for some KOIs. Mostly, the corrections for stellar properties have resulted in hotter and larger candidates. Deblending corrections increase planetary size estimates. INDIVIDUAL KOIs {#sec:individualkois} =============== An overview of new findings for individual planet candidates and validated or confirmed planets is summarized here. A complete list of the host star properties, neighboring stars not listed in the KIC and planet properties may be found in Tables \[table:stellarproperties\], \[table:secondaries\] and \[table:planetproperties\] respectively. #### KOI 115 (Kepler-105): Two previously-validated planets orbit this isolated solar-type star with orbital periods of 5.412 and 7.126 days. These planets are validated in our study as well and given new radii estimates of 2.54 and 1.44 $R_{\bigoplus}$ respectively. #### KOI 265: This isolated, perhaps slightly-evolved, solar-type star hosts a single 1.71 $R_{\bigoplus}$ planet candidate with a 3.568-day orbital period. We present new $J$ and $Ks$ AO imaging in addition to the speckle imaging for this star to yield a 94% validation level. #### KOI 268: This is a 6343 K dwarf star hosting a single planet candidate in a 110-day orbit. New $J$ and $K^\prime$ AO imaging reveals two neighboring stars (B and C) several magnitudes fainter. Since neighbor B lies within the exclusion radius, this planet cannot be validated using our methods. The proximity and relative magnitudes in two filters provide evidence that both of these neighbors are bound companions, but this is less certain than for other KOIs in our study. This KOI is subject to slight blending by its neighbors. If the candidate orbits the KOI star it is estimated to have a radius of 3.04 $R_{\bigoplus}$. If, on the other hand, it orbits the assumed 4007 K dwarf secondary star (neighbor B), the radius would be 9.33 $R_{\bigoplus}$ and its equilibrium temperature would be rather low (217 K). #### KOI 274 (Kepler-128): Two previously-validated planets orbit this slightly-evolved isolated solar-type star with 15 and 22.8-day orbital periods. We validate the planets again with speckle imaging (finding the host star to be single). Both planets have radii near 1.2 $R_{\bigoplus}$. #### KOI 284 (Kepler-132): There is a system of three previously-validated planets and one planet candidate orbiting this solar type star, whose stellar properties we characterize with new spectroscopy. The [[*Kepler*]{} ]{}light curve is significantly blended by flux from a neighboring star, which falls within the exclusion radius for each planet (so re-validation of this KOI is not attempted here). In addition to speckle imaging, we publish new $J$ and $Ks$ AO photometry here. It is difficult to use the multi-band photometry to determine if the neighbor is a bound companion or field star. Its close angular proximity to the KOI and photometry consistent with that of a bound companion argue that this is the most likely case. If a bound companion star, the planets and candidate planet radii all fall within a $2-3R_{\bigoplus}$ range and have virtually identical radius estimates whether they orbit the KOI or secondary star. #### KOI 369 (Kepler-144): This 6157 K dwarf harbors two previously-validated planets. We provide new spectroscopic characterization and $J$-band AO imaging for this star. No neighboring stars are found, enabling us to re-validate the planets based on the speckle and AO imaging. Kepler-144b and c are found to have radii of 1.78 and 1.69 $R_{\bigoplus}$ respectively. #### KOI 1537: One candidate planet orbits this 6260 K dwarf with a period of 10 days. The speckle imaging presented here shows no neighboring stars for this KOI in contrast to published AO observations in $Ks$ [@adamsetal13]. The two photometric studies are apparently irreconcilable assuming a very red star. Validation is attempted for this candidate, but the resulting level is fairly low (86.7%). No blending correction is performed for this KOI, resulting in a planet radius of 1.35 $R_{\bigoplus}$. #### KOI 1964: This 5547 K dwarf is observed with both speckle imaging and $J$ and $Ks$ AO imaging. A neighboring star lies about $0.4\arcsec$ to the north and within the exclusion radius of its single 2.2-day orbital period planet candidate. Relative photometry of the neighboring star results in some ambiguity: near-IR colors suggest the neighbor is a background field star while the optical colors are more consistent with those of a bound companion. Both cases are examined and the slight blending effects are evaluated and corrected to conclude that the candidate has a nominal radius of $0.764$ or $0.785R_{\bigoplus}$ if it orbits the KOI star (the two cases represent different blending corrections assuming the neighbor is alternatively a bound companion or field star). If it orbits the neighbor and that star is a 3892 K dwarf secondary, the planetary radius would need to be increased to 2.03 $R_{\bigoplus}$. #### KOI 2311: This solar type dwarf hosts two candidate planets with orbital periods of approximately 192 and 14 days. A new spectral characterization is presented for this star. New optical speckle and AO imaging in $J$ and $K^\prime$ reveal a faint neighboring star about $1\arcsec$ away. Since the [[*Kepler*]{} ]{}data pipeline did not report the astrometry needed to define an exclusion radius, the neighbor is also assumed to be the potential host star. Since the multi-band photometry is ambiguous in terms of determining if the neighbor star is a bound companion or not, both scenarios are examined. The neighbor is relatively faint, so while a deblending calculation is performed, it has no significant impact on interpreting the light curve. The inner planet candidate is found to be 0.932 $R_{\bigoplus}$ if it orbits the KOI and 4.16 $R_{\bigoplus}$ if it orbits the neighbor (here assuming the bound companion scenario). The outer planet candidate is cool (337 K) and small (1.15 $R_{\bigoplus}$) if it orbits the KOI and if it orbits the neighboring star (again assuming a bound companion), it is cold (117 K) and larger than Neptune (5.14 $R_{\bigoplus}$). #### KOI 2365 (Kepler-430): This is a solar type dwarf that hosts two planets and is found to be an isolated star in the speckle imaging and characterized with a new spectrum. Both planets are newly-validated at the 99.9% level using the speckle data and a planet multiplicity boost. The inner planet, Kepler-430b, orbits with a 36-day period and is found to be 3.25 $R_{\bigoplus}$ with $T_{eq}=667$ K. The outer planet, Kepler-430c, orbits in a 111-day period orbit and is found to be 1.75 $R_{\bigoplus}$ with $T_{eq}=458$ K. #### KOI 2593: This is an isolated star hosting a single candidate. New spectroscopy is presented and shows the KOI to be a 6119 K dwarf. The planet is validated at a 90.6% level using speckle imaging along with a $K^\prime$ AO image and the planet candidate is found to have a $1.10~R_{\bigoplus}$ radius with an equilibrium temperature of 915 K. #### KOI 2755: This is an isolated star hosting a single candidate. New spectroscopy is presented that shows the KOI to be a 5792 K dwarf. The planet is validated at a 82.7% level using speckle imaging and the planetary properties are found to be $R_p=1.06R_{\bigoplus}$ and $T_{eq}=974$ K. #### KOI 3097 (Kepler-431): This is a solar type dwarf that hosts three planets. A new spectrum is used to characterize the star as a 6004 K dwarf. Speckle imaging shows that the star is single and is used to validate these planets for the first time at the 99.8% level. Each is a small planet orbiting close to the parent star. Kepler-431b (KOI 3071.02) orbits with a 6.8-day period and is found with $R_p=0.764$ and $T_{eq}=1032~K$. Kepler-431c (KOI 3097.03) orbits with a 8.7-day period and is found with $R_p=0.668$ and $T_{eq}=951~K$. Kepler-431d (KOI 3097.01) orbits with a 11.9-day period and is found to have $R_p=1.11$ and $T_{eq}=856~K$. #### KOI 3204: This is a hot (7338 K) dwarf star with a single planet candidate in a 0.57-day orbital period. Speckle imaging shows this KOI to be a single star, helping confirm the hot planet candidate’s properties, which are calculated here to be $R_p=1.01R_{\bigoplus}$ and $T_{eq}=3268$ K. The planet is validated at a level of 98.5% based on the speckle images. #### KOI 3224: This is a dwarf slightly cooler than the Sun (5382 K) as shown by the new spectroscopy in this study. Speckle imaging shows it is a single star and validates the planet at a level of 90.5%. Its single planet candidate is sub-Earth size ($0.667~R_{\bigoplus}$) and hot ($T_{eq}=1129~K$). #### KOI 3255: This is a somewhat faint, cool dwarf (4427 K) that is observed using a combination of optical speckle and $K^\prime$ AO imaging. It harbors a single, cool planet candidate in a 66.7-day orbital period. KOI 3255 has two neighbors: B, a closeby and relatively bright star at $0.18\arcsec$ separation and a fainter companion, C, about $3\arcsec$ away that is identified in the UKIRT $J$-band survey by P. Lucas. Since neighbor B lies within the exclusion radius, no validation calculation is done. Blending effects are quite significant for this KOI due to the relatively bright neighbor B. The multi-band photometry of neighbor B is in agreement with that of a bound companion as also suggested by its close separation. A background star is also a possibility, so both scenarios are examined. Neighbor C is observed only in $J$ and has a relatively wide separation, so its nature is ambiguous. Given its relative faintness, however, the lack of color information for neighbor C is of relatively minor concern. The candidate planet radius is found to be $2.11~R_{\bigoplus}$ if it orbits the KOI and $2.42~R_{\bigoplus}$ if it orbits neighbor B. Considering the KOI and (an assumed bound) neighbor B in turn as the potential host star, the equilibrium temperatures for KOI 3255.01 are quite low, 294 K and 276 K respectively, making it a prime HZ candidate. #### KOI 3284: This is the lowest mass KOI star in the study, a 3688 K dwarf. It harbors a single planet candidate with a 35-day orbital period. It is observed using $K^\prime$ AO along with optical speckle imaging and found to have a nearby companion, B, at $0.44\arcsec$ separation. Another neighbor, C, is found in $UBV$ and $J$-band surveys of the [[*Kepler*]{} ]{}field. Neighbor B falls within the exclusion radius so no validation is attempted. Blending effects are fairly significant for this KOI. Once corrected, the planet candidate is found to have nominally $R_p=0.99$ (assuming background neighbors) or $1.00R_\bigoplus$ (assuming bound neighbors) and $T_{eq}=272~K$ if it orbits the KOI star. If it orbits neighbor B, and this neighbor is a bound companion, one finds $R_p=1.46R_{\bigoplus}$ and $T_{eq}=184~K$. The photometry for neighbor C is best explained as that of a background dwarf. KOI 3284.01 is a small HZ candidate. #### KOI 4407: This is a 6408 K dwarf as found in the new spectral characterization of this study. It hosts a single small planet candidate with a 1.34-day orbital period. The star was observed using both speckle and $Ks$-band AO imaging and found to have two neighbors as observed only within the wider field of the AO images. These neighbors both lie outside of the exclusion radius and validation is possible at a 19.2% level (the 17 ppm light curve transit depth is exceptionally shallow making validation difficult). Since the neighboring stars are observed in a single passband, their nature (whether secondary/tertiary or field stars) is ambiguous. There is some light curve blending due to the neighbors. If both neighbors are gravitationally bound to the KOI, the planetary properties would be $R_p=0.65R_\bigoplus$ and $T_{eq}=2121~K$. Nominally, the radius would be very similar if the neighbors are field stars, but with additional uncertainty because ${\Delta}Kp$ is poorly constrained. While nominally, ${\Delta}Kp = 2.70$ and 7.64 for field star neighbors B and C respectively, their relative ${\Delta}Kp$ magnitudes could be as bright as 1.0 and 6.6 or as faint as 4.0 and 9.4 respectively (see Figure \[Fig:fieldstarmags\]). In such eventualities, the blending-corrected planet radius could range from $R_p=0.64$ (for the case of faintest possible neighbors) to 0.75 (assuming each neighbor is as bright as possible). CONCLUSION {#sec:conclusion} ========== A high resolution speckle imaging survey was done on 18 KOI stars that host a total of 28 planets and candidate planets. This was supplemented by near-IR adaptive optics imagery of 10 and new spectroscopic characterizations for 11 of these stars. Validations (planet status confidence levels) are calculated for 18 of the planets or candidate planets that orbit the 12 host stars that are sufficiently isolated from detectable neighbors. There are 12 of the 18 planets validated at levels $>98$%. Five of these planets are first time validations with levels of at least $99.8$% (validating the two-planet system KOI 2365 as Kepler-430 and three-planet system KOI 3097 as Kepler-431). The high resolution imaging helped discover and then provide multi-color photometry to characterize close neighbor stars to 7 of the KOIs. These data, along with stellar characterization of the primary stars, were used to examine the relationship of the neighbors to the KOIs (gravitationally bound vs. field stars), and “deblend” the light curves by removing the excess light curve dilution due to neighbor stars. A reevaluation of the planet properties was done for the KOIs, accounting for revised host star properties and blending effects. Potential cases where neighboring stars could be the source of false positive planet signals were also evaluated. Further observations can help to solve some of the unresolved questions surrounding KOIs with neighboring stars. As shown, for the example targets, when double KOIs are observed in two or more filters, it is much easier to characterize both stars. Because much data is already published or publicly available, including good stellar characteristics for KOIs and imaging at multiple wavelengths (e.g., wide-field surveys of the [[*Kepler*]{} ]{}field in addition to targeted surveys of KOIs), it should be possible to apply many of the methods from this study to larger numbers of stars and determine, statistically, how blended KOIs bias the [[*Kepler*]{} ]{}planet sample. By quantifying the biases, appropriate corrections can be applied. There are uses for compiling a list of validated or candidate planets harbored by binary stars. Many stars are binary, but how does that environment affect planet occurrence and orbital properties? To better discriminate binary companions from unrelated, closely co-aligned field stars, repeated speckle astrometry can be used to find common proper motions pairs. For many KOIs, this will be a straightforward test. For example, the KIC lists proper motion measurements for 6 of the 18 stars in our sample with values ranging from $4-24$ mas yr$^{-1}$. Two of the KOIs found to be double are among this group and have proper motions of 4 and 6 mas yr$^{-1}$. As @horchetal12 have shown, and with some more analysis of recent data, speckle imaging at Gemini yields astrometric precision between $1-1.5$ mas for targets in the brightness range of KOIs. At the observed rates, relative proper motions (or common proper motions) can be detected using a pair of observations spaced 1 or 2 years apart. Recent work by @benedictetal14 shows that proper motions can now be derived for KOIs based on [[*Kepler*]{} ]{}pixel data, yielding precision 3 times better than in the existing catalogs. This level of astrometry could prove very helpful in identifying common proper motions. Additionally, ESA’s [*Gaia*]{} Mission [@perrymanetal01] promises to deliver a revolutionary astrometric dataset, impacting many fields. Once available, [*Gaia*]{} data should result in the reevaluation of [[*Kepler*]{} ]{}data, including the definitive detection of many astrometric binaries. Most of the data presented here is made available to the community for download at the [[*Kepler*]{} ]{}Community Follow-up Observing Program website (CFOP)$^8$, a service of the NASA Exoplanet Archive. These data include tabulated sensitivity curves for each of the speckle observations. The authors acknowledge the support of many people and programs that made this work possible. This paper includes data collected by the [[*Kepler*]{} ]{}Mission. Funding for the mission is provided by the NASA Science Mission directorate. M. E. Everett received support through NASA Agreement NNX13AB60A. T. Barclay was partially supported by a NASA Keck PI Data Award, administered by the NASA Exoplanet Science Institute. Comments to improve upon a draft of this paper were received from an anonymous referee, who we thank for the help. Data for this paper were obtained from numerous sources, including: (1) The NASA Exoplanet Archive, which is operated by the California Institute of Technology, under contract with the National Aeronautics and Space Administration under the Exoplanet Exploration Program; (2) the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX13AC07G and by other grants and contracts; and (3) the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. This work also made use of PyKE [@stillbarclay12], a software package for the reduction and analysis of Kepler data. This open source software project is developed and distributed by the NASA Kepler Guest Observer Office. The speckle imaging observations were obtained as part of the program GN-2013B-Q-87 at the Gemini Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), Ministério da Ciência, Tecnologia e Inovação (Brazil) and Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina). We are very grateful for the excellent support of the Gemini administration and support staff who helped make the visiting instrument program possible and the DSSI observing run a great success. Some of the data presented herein were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation. Finally, the authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain. Facilities: , , , , , Adams, E. R. et al 2012, , 144, 42 Adams, E. R., Dupree, A. K., Kulesa, C. & McCarthy, D. 2013, , 146, 9 Barclay, T. et al. 2013, , 768, 101 Batalha, N. M. et al. 2013, , 204, 24B Benedict, G. F., Tanner, A. M., Cargile, P. A. & Ciardi, D. R. 2014, arXiv:1408.4054 Borucki, W. J. et al. 2010, Science, 327, 977 Brown, T. M., Latham, D. W., Everett, M. E. & Esquerdo, G. A. 2011, , 142, 112 Bryson, S. T. et al. 2013, , 125, 889 Buchhave, L. A. et al. 2012, Nature, 486, 375 Burke, C. J. et al. 2014, , 210, 19 Cardelli, J. A., Clayton, G. C. & Mathis, J. S. 1989, , 345, 245 Castelli, F. & Kurucz, R. L. 2003, in Proc. of the 210th Symposium of the IAU at Uppsala University, Uppsala, Sweden, 17-21 June, 2002. ed. by N. Piskunov, W. W. Weiss, & D. F. Gray. Published on behalf of the IAU by the Astronomical Society of the Pacific, A20 Ciardi, D. R. et al. 2014, in prep. Coelho, P., Barbuy, B., Meléndez, J., Schiavon, R. P., Castilho, B. V. 2005, , 443, 735 Dressing, C. D. & Charbonneau, D. 2013, , 767, 95D Everett, M. E., Howell, S. B., & Kinemuchi, K. 2012, , 124, 316 Everett, M. E., Howell, S. B., Silva, D. R., & Szkody, P. 2013, ApJ, 771, 107E Fressin, F. et al. 2013, , 766, 81 Gilliland, R. L., Cartier, K. M. S., Adams, E. R., Ciardi, D. R., Kalas, P. & Wright, J. T. 2014, arXiv:1407.1009v2 Girardi, L., Groenewegen, M. A. T., Hatziminaoglou, E. & da Costa, L. 2005, , 436, 895G Girardi, L. et al. 2012, in Astrophysics and Space Science Proceedings, Red Giants as Probes of the Structure and Evolution of the Milky Way (Berlin: Springer-Verlag), 165 Hayward, T. L. et al. 2001, , 113, 105 Horch, E. P., Howell, S. B., Everett, M. E., & Ciardi, D. R. 2012, AJ, 144, 165 Horch, E. P., Howell, S. B., Everett, M. E., & Ciardi, D. R. 2014, , 795, 60 Horch, E. P., Veillette, D. R., Baena Gallé, R., Shah, S. C., O’Rielly, G. V., & van Altena, W. F. 2009, AJ, 137, 5057 Howell, S. B., Everett, M. E., Sherry, W., Horch, E. & Ciardi, D. R. 2011, , 142, 19H Howell, S. B. et al. 2012, , 746, 123 Huber, D. et al. 2013, , 767, 127 Huber, D. et al. 2014, ApJS, 211, 2 Jenkins, J. M. et al. 2010, , 713, 87 Jenkins, J. M. et al. 2010, Proc. SPIE, 7740, 10 Johansson, E. M. et al. 2008, SPIE, 7015, 91 Koch, D. G. et al. 2010, , 713, L79 Kolbl, R. et al. 2014, in prep. Law, N. M. et al. 2014, , 791, 35 Lillo-Box, J., Barrado, D. & Bouy, H. 2014, , 566, 103 Lissauer, J. J. et al. 2012, , 750, 112 Lissauer, J. J. et al. 2014, , 784, 44 Lloyd, J. P. et al. 2000, SPIE, 4008, 814 Marcy, G. W. et al. 2014, , 210, 20 Morton, T. D. & Johnson, J. A. 2011, , 738, 170 Munari, U., Sordo, R., Castelli, F. & Zwitter, T. 2005, , 442, 1127 Perryman, M. A. C. et al. 2001, , 369, 339 Pinsonneault, M. H., An, D., Molenda-Żakowicz, J., Chaplin, W. J., Metcalfe, T. S., & Bruntt, H. 2012, , 199, 30 Quintana, E. V. et al. 2010, Proc. SPIE, Vol. 7740, 64Q Quintana, E. V. et al. 2014, Science, 344, 277 Rowe, J. F. et al. 2014, , 784, 45 Santerne, A. et al. 2013, , 557, 139 Skrutskie, M. F. et al. 2006, , 131, 1163 Slawson, R. W. et al. 2011, , 142, 160 Still, M. & Barclay, T. 2012, Astrophysics Source Code Library, record ascl:1208.004 Tenenbaum, P. et al. 2013, , 206, 5 Tenenbaum, P. et al. 2014, , 211, 6 Twicken, J. D. et al. 2010 Proc. SPIE 7740, 62T Wang, J. et al. 2013, , 776, 10 Wang, J. et al. 2014, , 783, 4 Wizinowich, P. L. et al. 2004, SPIE, 5490, 1 Wu, H. et al. 2010, Proc. SPIE, Vol. 7740, 42W Xie, J.-W. 2014, , 210, 25 [clrccc]{} 115 & Kepler-105 & 9579641 & 12.791 & 2013 July 25 & 12\ 265 & & 12024120 & 11.994 & 2013 July 28 & 9\ 268 & & 3425851 & 10.560 & 2013 July 25 & 3\ 274 & Kepler-128 & 8077137 & 11.390 & 2013 July 27 & 6\ 284 & Kepler-132 & 6021275 & 11.818 & 2013 July 25 & 6\ 369 & Kepler-144 & 7175184 & 11.992 & 2013 July 27 & 9\ 1537 & & 9872292 & 11.740 & 2013 July 27 & 6\ 1964 & & 7887791 & 10.687 & 2013 July 27 & 3\ 2311 & & 4247991 & 12.570 & 2013 July 25 & 9\ 2365 & Kepler-430 & 11560897 & 13.848 & 2013 July 25 & 18\ 2593 & & 8212002 & 11.714 & 2013 July 27 & 6\ 2755 & & 3545135 & 12.147 & 2013 July 27 & 9\ 3097 & Kepler-431 & 7582689 & 11.973 & 2013 July 27 & 6\ 3204 & & 11456279 & 11.825 & 2013 July 27 & 6\ 3224 & & 10384298 & 12.192 & 2013 July 27 & 9\ 3255 & & 8183288 & 14.352 & 2013 July 27 & 21\ 3284 & & 6497146 & 14.467 & 2013 July 25 & 25\ 4407 & & 8396660 & 11.179 & 2013 July 28 & 6\ [crclcc]{} 265 & & 11.994 & 2010 July 2 & Palomar/Pharo & $J,Ks$\ 268 & & 10.560 & 2012 July 4 & Keck/NIRC2 & $J,K^\prime$\ 284 & Kepler-132 & 11.818 & 2010 July 1-2 & Palomar/Pharo & $J,Ks$\ 369 & Kepler-144 & 11.992 & 2011 Sept. 10 & Lick/IRCAL & $J$\ 1964 & & 10.687 & 2013 June 26 & Palomar/Pharo & $J,Ks$\ 2311 & & 12.570 & 2012 Aug. 25 & Keck/NIRC2 & $J,K^\prime$\ 2593 & & 11.714 & 2013 July 7 & Keck/NIRC2 & $K^\prime$\ 3255 & & 14.352 & 2012 Aug. 25 & Keck/NIRC2 & $K^\prime$\ 3284 & & 14.467 & 2013 July 6 & Keck/NIRC2 & $K^\prime$\ 4407 & & 11.179 & 2013 June 27 & Palomar/Pharo & $Ks$\ [cccc]{} 115 & Kepler-105 & 12.791 & 2010 May 24\ 265 & & 11.994 & 2010 Sept. 14\ 284 & Kepler-132 & 11.818 & 2013 Sept. 1\ 369 & Kepler-144 & 11.992 & 2013 Sept. 1\ 2311 & & 12.570 & 2013 Sept. 1\ 2365 & Kepler-430 & 13.848 & 2013 Sept. 1\ 2593 & & 11.714 & 2013 Sept. 1\ 2755 & & 12.147 & 2013 Sept. 1\ 3097 & Kepler-431 & 11.973 & 2013 Sept. 1\ 3224 & & 12.192 & 2013 Sept. 1\ 4407 & & 11.179 & 2013 Sept. 1\ [clcccccc]{} 115 & Kepler-105 & $6065\pm75$ & $4.43\pm0.15$ & $-0.10\pm0.10$ & $1.015\begin{array}{c}+0.189\\-0.071\end{array}$ & $1.027\begin{array}{c}+0.057\\-0.035\end{array}$ & 1\ 265 & & $5915\pm75$ & $4.07\pm0.15$ & $0.06\pm0.10$ & $1.564\begin{array}{c}+0.456\\-0.252\end{array}$ & $1.097\begin{array}{c}+0.122\\-0.054\end{array}$ & 1\ 268 & & $6343\pm85$ & $4.259\pm0.010$ & $-0.040\pm0.101$ & $1.366\pm0.026$ & $1.230\pm0.058$ & 2\ 274 & Kepler-128 & $6072\pm75$ & $4.070\pm0.011$ & $-0.090\pm0.101$ & $1.659\pm0.038$ & $1.184\pm0.074$ & 2\ 284 & Kepler-132 & $5879\pm75$ & $4.15\pm0.15$ & $-0.04\pm0.10$ & $1.408\begin{array}{c}+0.284\\-0.240\end{array}$ & $1.023\begin{array}{c}+0.080\\-0.055\end{array}$ & 3\ 369 & Kepler-144 & $6157\pm75$ & $4.14\pm0.15$ & $-0.02\pm0.10$ & $1.491\begin{array}{c}+0.288\\-0.247\end{array}$ & $1.126\begin{array}{c}+0.108\\-0.049\end{array}$ & 3\ 1537 & & $6260\pm116$ & $4.047\pm0.014$ & $0.100\pm0.109$ & $1.824\pm0.049$ & $1.366\pm0.101$ & 2\ 1964 & & $5547\begin{array}{c}+109\\-91\end{array}$ & $4.388\begin{array}{c}+0.107\\-0.126\end{array}$ & $-0.040\begin{array}{c}+0.160\\-0.140\end{array}$ & $0.989\begin{array}{c}+0.177\\-0.109\end{array}$ & $0.871\begin{array}{c}+0.068\\-0.038\end{array}$ & 4\ 2311 & & $5657\pm75$ & $4.29\pm0.15$ & $0.15\pm0.10$ & $1.182\begin{array}{c}+0.220\\-0.195\end{array}$ & $0.975\begin{array}{c}+0.046\\-0.039\end{array}$ & 3\ 2365 & Kepler-430 & $5884\pm75$ & $4.15\pm0.15$ & $0.20\pm0.10$ & $1.485\begin{array}{c}+0.266\\-0.234\end{array}$ & $1.166\begin{array}{c}+0.134\\-0.095\end{array}$ & 3\ 2593 & & $6119\pm75$ & $4.21\pm0.15$ & $0.16\pm0.10$ & $1.453\begin{array}{c}+0.304\\-0.287\end{array}$ & $1.230\begin{array}{c}+0.091\\-0.093\end{array}$ & 3\ 2755 & & $5792\pm75$ & $4.29\pm0.15$ & $0.01\pm0.10$ & $1.172\begin{array}{c}+0.236\\-0.173\end{array}$ & $0.973\begin{array}{c}+0.047\\-0.037\end{array}$ & 3\ 3097 & Kepler-431 & $6004\pm75$ & $4.40\pm0.15$ & $0.07\pm0.10$ & $1.092\begin{array}{c}+0.191\\-0.109\end{array}$ & $1.071\begin{array}{c}+0.059\\-0.037\end{array}$ & 3\ 3204 & & $7338\begin{array}{c}+226\\-336\end{array}$ & $4.225\begin{array}{c}+0.060\\-0.445\end{array}$ & $0.070\begin{array}{c}+0.170\\-0.390\end{array}$ & $1.593\begin{array}{c}+1.273\\-0.202\end{array}$ & $1.553\begin{array}{c}+0.375\\-0.225\end{array}$ & 5\ 3224 & & $5382\pm75$ & $4.30\pm0.15$ & $0.10\pm0.10$ & $0.962\begin{array}{c}+0.100\\-0.091\end{array}$ & $0.866\begin{array}{c}+0.040\\-0.021\end{array}$ & 3\ 3255 & & $4427\begin{array}{c}+133\\-129\end{array}$ & $4.639\begin{array}{c}+0.055\\-0.033\end{array}$ & $-0.320\begin{array}{c}+0.340\\-0.320\end{array}$ & $0.622\begin{array}{c}+0.056\\-0.060\end{array}$ & $0.615\begin{array}{c}+0.066\\-0.049\end{array}$ & 6\ 3284 & & $3688\begin{array}{c}+73\\-50\end{array}$ & $4.788\begin{array}{c}+0.060\\-0.080\end{array}$ & $-0.100\pm0.100$ & $0.463\begin{array}{c}+0.070\\-0.050\end{array}$ & $0.479\begin{array}{c}+0.060\\-0.050\end{array}$ & 7\ 4407 & & $6408\pm75$ & $4.22\pm0.15$ & $0.01\pm0.10$ & $1.435\begin{array}{c}+0.329\\-0.265\end{array}$ & $1.234\begin{array}{c}+0.102\\-0.065\end{array}$ & 3\ [rlrll]{} 115.01 & Kepler-105b & 7.31 & 0.195 & 0.99996\ 115.02 & Kepler-105c & 8.58 & 0.72 & 0.9997\ 265.01 & & 9.29 & 0.99 & 0.940\ 268.01 & & 7.51 & 2.31 &\ 274.01 & Kepler-128b & 10.00 & 1.56 & 0.998\ 274.02 & Kepler-128c & 9.96 & 2.1 & 0.998\ 284.01 & Kepler-132d & 8.70 & 1.08 &\ 284.02 & Kepler-132c & 9.03 & 1.44 &\ 284.03 & Kepler-132b & 9.15 & 1.53 &\ 284.04 & & 8.95 & 7.8 &\ 369.01 & Kepler-144b & 9.01 & 1.29 & 0.989\ 369.02 & Kepler-144c & 9.10 & 0.9 & 0.994\ 1537.01 & & 9.72 & 1.38 & 0.867\ 1964.01 & & 9.96 & 1.86 &\ 2311.01 & & 9.18 & &\ 2311.02 & & 10.27 & &\ 2365.01 & Kepler-430b & 8.18 & 0.84 & 0.9993\ 2365.02 & Kepler-430c & 8.88 & 2.64 & 0.999\ 2593.01 & & 9.84 & 3.3 & 0.906\ 2755.01 & & 9.56 & 0.96 & 0.827\ 3097.01 & Kepler-431d & 9.90 & 1.23 & 0.9994\ 3097.02 & Kepler-431b & 10.48 & 1.32 & 0.998\ 3097.03 & Kepler-431c & 10.55 & 2.25 & 0.998\ 3204.01 & & 10.42 & 0.36 & 0.985\ 3224.01 & & 10.03 & 2.67 & 0.905\ 3255.01 & & 7.45 & 0.57 &\ 3284.01 & & 7.79 & 1.41 &\ 4407.01 & & 11.20 & 0.78 & 0.192\ [cccccccc]{} 268 & 3425851 & NIRC2 & $J$ & B & $267.69\pm0.02$ & $1.7591\pm0.0002$ & $3.11\pm0.05$\ & & NIRC2 & $K^\prime$ & B & $267.69\pm0.02$ & $1.7591\pm0.0002$ & $2.54\pm0.03$\ & & NIRC2 & $J$ & C & $310.19\pm0.02$ & $2.5243\pm0.0006$ & $4.33\pm0.05$\ & & NIRC2 & $K^\prime$ & C & $310.19\pm0.02$ & $2.5243\pm0.0006$ & $3.79\pm0.04$\ 284 & 6021275 & DSSI & 692nm & B & 97.44 & 0.8672 & $0.66\pm0.15$\ & & DSSI & 880nm & B & 97.25 & 0.8681 &\ & & Pharo & $J$ & B & & & 0.26\ & & Pharo & $Ks$ & B & & & 0.26\ 1964 & 7887791 & DSSI & 692nm & B & 1.72 & 0.3916 & $3.54\pm0.15$\ & & DSSI & 880nm & B & 2.81 & 0.4039 & $2.85\pm0.15$\ & & Pharo & $J$ & B & & & 1.96\ & & Pharo & $Ks$ & B & & & 1.78\ 2311 & 4247991 & DSSI & 692nm & B & 69.03 & 1.0295 & $5.47\pm0.15$\ & & NIRC2 & $J$ & B & $70.19\pm0.04$ & $1.0264\pm0.0003$ & $5.38\pm0.13$\ & & NIRC2 & $K^\prime$ & B & $70.19\pm0.04$ & $1.0264\pm0.0003$ & $4.74\pm0.06$\ 3255 & 8183288 & DSSI & 692nm & B & 336.41 & 0.1812 & $0.52\pm0.15$\ & & DSSI & 880nm & B & 337.99 & 0.1852 & $0.40\pm0.15$\ & & NIRC2 & $K^\prime$ & B & $336\pm3$ & $0.175\pm0.015$ & $0.11\pm0.04$\ & & UKIRT & $J$ & C & 45.0 & 3.05 & $4.761\pm0.063$\ 3284 & 6497146 & DSSI & 692nm & B & 193.06 & 0.4380 & $3.56\pm0.15$\ & & NIRC2 & $K^\prime$ & B & & & $2.01\pm0.15$\ & & WIYN & $B$ & C & & & $1.802\pm0.046$\ & & WIYN & $V$ & C & 3.2 & 3.98 & $2.013\pm0.035$\ & & UKIRT & $J$ & C & 3.2 & 4.01 & $2.904\pm0.008$\ 4407 & 8396660 & Pharo & $Ks$ & B & 299.8 & 2.45 & $1.988\pm0.005$\ & & Pharo & $Ks$ & C & 311.0 & 2.65 & $4.972\pm0.022$\ [ccccccccc]{} 268 & KOI & & & 10.56 & & & $3.56\pm0.10$ &\ 268 & B & 267.69 & 1.7591 & 14.88 & $5.1\times10^{-4}$ & $Ks$ & $7.87\begin{array}{c}+0.13\\-0.11\end{array}$ & -0.10\ 268 & B & 267.69 & 1.7591 & 14.88 & & $J$ & $7.89\begin{array}{c}+0.14\\-0.10\end{array}$ & 0.07\ 268 & C & 310.19 & 2.5243 & 16.59 & $2.5\times10^{-3}$ & $Ks$ & $9.59\begin{array}{c}+0.09\\-0.07\end{array}$ & 0.0\ 268 & C & 310.19 & 2.5243 & 16.59 & & $J$ & $9.59\begin{array}{c}+0.11\\-0.09\end{array}$ & 0.0\ 284 & KOI & & & 12.38 & & & $3.75\begin{array}{c}+0.41\\-0.40\end{array}$ &\ 284 & B & 97.44 & 0.867 & 12.80 & $4.8\times10^{-5}$ & $Ks$ & $4.02\begin{array}{c}+0.45\\-0.46\end{array}$ & -0.32\ 284 & B & 97.44 & 0.867 & 12.80 & & $J$ & $4.02\begin{array}{c}+0.42\\-0.44\end{array}$ & -0.34\ 284 & B & 97.44 & 0.867 & 12.80 & & 692nm & $4.41\begin{array}{c}+0.44\\-0.40\end{array}$ & 0.61\ 1964 & KOI & & & 10.73 & & & $4.78\begin{array}{c}+0.23\\-0.26\end{array}$ &\ 1964 & B & 2.26 & 0.3978 & 14.10 & $1.2\times10^{-5}$ & $Ks$ & $8.00\begin{array}{c}+0.33\\-0.34\end{array}$ & -0.46\ 1964 & B & 2.26 & 0.3978 & 14.10 & & $J$ & $7.64\begin{array}{c}+0.30\\-0.36\end{array}$ & -1.74\ 1964 & B & 2.26 & 0.3978 & 14.10 & & 880nm & $8.36\begin{array}{c}+0.30\\-0.29\end{array}$ & 0.70\ 1964 & B & 2.26 & 0.3978 & 14.10 & & 692nm & $8.38\begin{array}{c}+0.23\\-0.22\end{array}$ & 1.05\ 2311 & KOI & & & 12.57 & & & $4.29\begin{array}{c}+0.39\\-0.37\end{array}$ &\ 2311 & B & 70.19 & 1.0264 & 19.02 & $2.1\times10^{-3}$ & $Ks$ & $11.32\begin{array}{c}+0.37\\-0.46\end{array}$ & 1.26\ 2311 & B & 70.19 & 1.0264 & 19.02 & & $J$ & $11.59\begin{array}{c}+0.40\\-0.44\end{array}$ & 1.92\ 2311 & B & 70.19 & 1.0264 & 19.02 & & 692nm & $9.72\begin{array}{c}+0.38\\-0.36\end{array}$ & -2.70\ 3255 & KOI & & & 14.92 & & & $7.04\begin{array}{c}+0.22\\-0.23\end{array}$ &\ 3255 & B & 337.20 & 0.1832 & 15.33 & $1.3\times10^{-5}$ & $Ks$ & $7.25\pm0.21$ & -0.90\ 3255 & B & 337.20 & 0.1832 & 15.33 & & 880nm & $7.54\begin{array}{c}+0.27\\-0.22\end{array}$ & 0.46\ 3255 & B & 337.20 & 0.1832 & 15.33 & & 692nm & $7.54\begin{array}{c}+0.26\\-0.22\end{array}$ & 0.46\ 3255 & C & 45.0 & 3.05 & 19.77 & $6.2\times10^{-2}$ & $J$ & $12.73\begin{array}{c}+0.24\\-0.13\end{array}$ &\ 3284 & KOI & & & 14.55 & & & $8.89\begin{array}{c}+0.17\\-0.22\end{array}$ &\ 3284 & B & 193.06 & 0.4380 & 17.32 & $4.6\times10^{-5}$ & $Ks$ & $11.28\pm0.19$ & -2.07\ 3284 & B & 193.06 & 0.4380 & 17.32 & & 692nm & $12.29\begin{array}{c}+0.19\\-0.24\end{array}$ & 2.62\ 3284 & C & 3.2 & 4.01 & 16.73 & $9.2\times10^{-3}$ & $J$ & $12.35\begin{array}{c}+0.13\\-0.18\end{array}$ & 6.53\ 3284 & C & 3.2 & 4.01 & 16.73 & & $V$ & $10.71\begin{array}{c}+0.19\\-0.24\end{array}$ & -2.33\ 3284 & C & 3.2 & 4.01 & 16.73 & & $B$ & $10.45\begin{array}{c}+0.17\\-0.24\end{array}$ & -4.02\ 4407 & KOI & & & 11.18 & & & $3.34\begin{array}{c}+0.44\\-0.45\end{array}$ &\ 4407 & B & 299.8 & 2.45 & 14.36 & $1.9\times10^{-3}$ & $Ks$ & $6.52\begin{array}{c}+0.89\\-1.01\end{array}$ &\ 4407 & C & 311.0 & 2.65 & 18.64 & $2.5\times10^{-2}$ & $Ks$ & $10.80\begin{array}{c}+0.49\\-0.54\end{array}$ &\ [llccccccc]{} 268 & KOI & $3.58\begin{array}{c}+0.10\\-0.09\end{array}$ & $3.63\begin{array}{c}+0.10\\-0.08\end{array}$ & $2.82\begin{array}{c}+0.09\\-0.07\end{array}$ & $2.55\begin{array}{c}+0.09\\-0.06\end{array}$ & $0.05\begin{array}{c}+0.00\\-0.01\end{array}$ & $1.03\begin{array}{c}+0.02\\-0.04\end{array}$ & $0.27\pm0.01$\ 268 & B & & & $5.93\begin{array}{c}+0.10\\-0.09\end{array}$ & $5.09\begin{array}{c}+0.09\\-0.07\end{array}$ & & & $0.84\pm0.06$\ 268 & C & & & $7.15\begin{array}{c}+0.10\\-0.09\end{array}$ & $6.34\begin{array}{c}+0.10\\-0.07\end{array}$ & & & $0.81\begin{array}{c}+0.06\\-0.07\end{array}$\ 284 & KOI & $3.73\begin{array}{c}+0.41\\-0.39\end{array}$ & $3.73\begin{array}{c}+0.41\\-0.40\end{array}$ & $2.87\begin{array}{c}+0.41\\-0.40\end{array}$ & $2.53\begin{array}{c}+0.41\\-0.40\end{array}$ & $0.01\pm0.01$ & $1.20\pm0.03$ & $0.34\pm0.01$\ 284 & B & $4.39\begin{array}{c}+0.44\\-0.42\end{array}$ & & $3.13\begin{array}{c}+0.41\\-0.40\end{array}$ & $2.79\begin{array}{c}+0.41\\-0.40\end{array}$ & & $1.60\pm0.16$ & $0.34\pm0.04$\ 1964 & KOI & $4.73\begin{array}{c}+0.23\\-0.26\end{array}$ & $4.69\begin{array}{c}+0.23\\-0.25\end{array}$ & $3.77\begin{array}{c}+0.23\\-0.24\end{array}$ & $3.37\pm0.23$ & $0.04\pm0.01$ & $1.36\begin{array}{c}+0.04\\-0.05\end{array}$ & $0.41\pm0.02$\ 1964 & B & $8.27\begin{array}{c}+0.27\\-0.30\end{array}$ & $7.54\begin{array}{c}+0.27\\-0.29\end{array}$ & $5.73\begin{array}{c}+0.23\\-0.24\end{array}$ & $5.15\pm0.23$ & $0.73\pm0.21$ & $3.12\pm0.16$ & $0.59\pm0.04$\ 2311 & KOI & $4.24\begin{array}{c}+0.39\\-0.37\end{array}$ & $4.23\begin{array}{c}+0.39\\-0.37\end{array}$ & $3.33\begin{array}{c}+0.39\\-0.37\end{array}$ & $2.95\begin{array}{c}+0.39\\-0.37\end{array}$ & $0.02\pm0.01$ & $1.29\pm0.04$ & $0.37\pm0.02$\ 2311 & B & $9.71\begin{array}{c}+0.41\\-0.40\end{array}$ & & $8.71\begin{array}{c}+0.41\\-0.39\end{array}$ & $7.69\begin{array}{c}+0.39\\-0.37\end{array}$ & & $2.02\pm0.17$ & $1.01\pm0.14$\ 3255 & KOI & $6.85\begin{array}{c}+0.22\\-0.21\end{array}$ & $6.61\begin{array}{c}+0.17\\-0.18\end{array}$ & $5.43\begin{array}{c}+0.16\\-0.17\end{array}$ & $4.69\begin{array}{c}+0.13\\-0.15\end{array}$ & $0.24\begin{array}{c}+0.06\\-0.04\end{array}$ & $2.16\begin{array}{c}+0.10\\-0.11\end{array}$ & $0.75\begin{array}{c}+0.03\\-0.04\end{array}$\ 3255 & B & $7.37\pm0.26$ & $7.01\begin{array}{c}+0.23\\-0.24\end{array}$ & & $4.80\begin{array}{c}+0.14\\-0.16\end{array}$ & $0.36\pm0.22$ & $2.57\begin{array}{c}+0.18\\-0.19\end{array}$ &\ 3284 & KOI & $8.79\begin{array}{c}+0.18\\-0.24\end{array}$ & $7.99\begin{array}{c}+0.14\\-0.18\end{array}$ & $6.64\begin{array}{c}+0.13\\-0.16\end{array}$ & $5.83\begin{array}{c}+0.13\\-0.17\end{array}$ & $0.80\begin{array}{c}+0.06\\-0.07\end{array}$ & $2.97\begin{array}{c}+0.07\\-0.09\end{array}$ & $0.82\pm0.01$\ 3284 & B & $12.35\begin{array}{c}+0.23\\-0.28\end{array}$ & & & $7.84\begin{array}{c}+0.14\\-0.17\end{array}$ & & $4.52\begin{array}{c}+0.17\\-0.18\end{array}$ &\ [ccccccc]{} 268.01 & KOI 268 & & & 0.00 & 98.358 & 1.022\ 268.01 & B & 267.69 & 1.7591 & 4.32 & 98.331 & 55.28\ 268.01 & C & 310.19 & 2.5243 & 6.03 & 98.076 & 264.9\ 284.01 & Kepler-132 & & & 0.00 & 95.270 & 1.685\ 284.01 & B & 97.44 & 0.867 & 0.41 & 95.148 & 2.461\ 284.02 & Kepler-132 & & & 0.00 & 95.336 & 1.685\ 284.02 & B & 97.44 & 0.867 & 0.41 & 95.206 & 2.461\ 284.03 & Kepler-132 & & & 0.00 & 95.364 & 1.684\ 284.03 & B & 97.44 & 0.867 & 0.41 & 95.183 & 2.462\ 284.04 & Kepler-132 & & & 0.00 & 95.193 & 1.684\ 284.04 & B & 97.44 & 0.867 & 0.41 & 94.974 & 2.462\ 1964.01 & KOI 1964 & & & 0.00 & 98.202 & 1.045\ 1964.01 & B & 2.26 & 0.3978 & 3.37 & 98.070 & 23.30\ 2311.01 & KOI 2311 & & & 0.00 & 94.821 & 1.003\ 2311.01 & B & 70.19 & 1.0264 & 6.45 & 94.149 & 384.0\ 2311.02 & KOI 2311 & & & 0.00 & 94.741 & 1.003\ 2311.02 & B & 70.19 & 1.0264 & 6.45 & 94.275 & 383.1\ 3255.01 & KOI 3255 & & & 0.00 & 87.770 & 1.700\ 3255.01 & B & 337.20 & 0.1832 & 0.41 & 87.768 & 2.465\ 3255.01 & C & 45.0 & 3.05 & 5.69 & 69.447 & 189.0\ 3284.01 & KOI 3284 & & & 0.00 & 68.748 & 1.160\ 3284.01 & B & 193.06 & 0.4380 & 2.77 & 68.832 & 14.86\ 3284.01 & C & 3.2 & 4.01 & 2.26 & 45.582 & 16.02\ 4407.01 & KOI 4407 & & & 0.00 & 96.491 & 1.054\ 4407.01 & B & 299.8 & 2.45 & 3.18 & 96.426 & 19.74\ 4407.01 & C & 311.0 & 2.65 & 7.46 & 96.034 & 1021\ [cccccccc]{} 1964.01 & KOI 1964 & & & & 0.00 & 98.202 & 1.106\ 1964.01 & B & background & 2.26 & 0.3978 & 2.44 & 98.070 & 10.47\ 2311.01 & KOI 2311 & & & & 0.00 & 94.821 & 1.005\ 2311.01 & B & background & 70.19 & 1.0264 & 5.70 & 94.149 & 192.9\ 2311.02 & KOI 2311 & & & & 0.00 & 94.741 & 1.005\ 2311.02 & B & background & 70.19 & 1.0264 & 5.70 & 94.275 & 192.5\ 3255.01 & KOI 3255 & & & & 0.00 & 87.770 & 1.698\ 3255.01 & B & bound & 337.20 & 0.1832 & 0.41 & 87.768 & 2.466\ 3255.01 & C & background & 45.0 & 3.05 & 4.81 & 69.447 & 182.7\ 3284.01 & KOI 3284 & & & & 0.00 & 68.748 & 1.145\ 3284.01 & B & bound & 193.06 & 0.4380 & 2.77 & 68.832 & 14.66\ 3284.01 & C & background & 3.2 & 4.01 & 2.48 & 45.582 & 19.42\ 4407.01 & KOI 4407 & & & & 0.00 & 96.491 & 1.084\ 4407.01 & B & background & 299.8 & 2.45 & 2.70 & 96.426 & 13.04\ 4407.01 & C & background & 311.0 & 2.65 & 7.64 & 96.034 & 1239\ [ccccccccccccc]{} 115 & KOI & 115.01 & Kepler-105b & 1.015 & 1.027 & 6065 & 5.412 & $2.292\times10^{-2}$ & 0.0609 & 1092 & 2.54 &\ 115 & KOI & 115.02 & Kepler-105c & 1.015 & 1.027 & 6065 & 7.126 & $1.296\times10^{-2}$ & 0.0731 & 997 & 1.44 &\ 265 & KOI & 265.01 & & 1.564 & 1.097 & 5915 & 3.568 & $1.001\times10^{-2}$ & 0.0471 & 1503 & 1.71 &\ 268 & KOI & 268.01 & & 1.366 & 1.230 & 6343 & 110.379 & $2.011\times10^{-2}$ & 0.4825 & 470 & 3.04 &\ 268 & B & 268.01 & & 0.571 & 0.596 & 4007 & 110.379 & $2.011\times10^{-2}$ & 0.3789 & 217 & 9.33 &\ 274 & KOI & 274.01 & Kepler-128b & 1.659 & 1.184 & 6072 & 15.092 & $6.630\times10^{-3}$ & 0.1264 & 970 & 1.20 &\ 274 & KOI & 274.02 & Kepler-128c & 1.659 & 1.184 & 6072 & 22.795 & $6.670\times10^{-3}$ & 0.1664 & 845 & 1.21 &\ 284 & KOI & 284.01 & Kepler-132d & 1.408 & 1.023 & 5879 & 18.010 & $1.490\times10^{-2}$ & 0.1355 & 836 & 2.98 &\ 284 & B & 284.01 & Kepler-132d & 1.169 & 0.987 & 5850 & 18.010 & $1.490\times10^{-2}$ & 0.1339 & 762 & 2.99 &\ 284 & KOI & 284.02 & Kepler-132c & 1.408 & 1.023 & 5879 & 6.415 & $1.135\times10^{-2}$ & 0.0681 & 1179 & 2.27 &\ 284 & B & 284.02 & Kepler-132c & 1.169 & 0.987 & 5850 & 6.415 & $1.135\times10^{-2}$ & 0.0673 & 1075 & 2.27 &\ 284 & KOI & 284.03 & Kepler-132b & 1.408 & 1.023 & 5879 & 6.178 & $1.087\times10^{-2}$ & 0.0664 & 1194 & 2.17 &\ 284 & B & 284.03 & Kepler-132b & 1.169 & 0.987 & 5850 & 6.178 & $1.087\times10^{-2}$ & 0.0656 & 1089 & 2.18 &\ 284 & KOI & 284.04 & & 1.408 & 1.023 & 5879 & 110.287 & $1.125\times10^{-2}$ & 0.4535 & 457 & 2.25 &\ 284 & B & 284.04 & & 1.169 & 0.987 & 5850 & 110.287 & $1.125\times10^{-2}$ & 0.4481 & 416 & 2.25 &\ 369 & KOI & 369.01 & Kepler-144b & 1.491 & 1.126 & 6157 & 5.885 & $1.090\times10^{-2}$ & 0.0664 & 1287 & 1.78 &\ 369 & KOI & 369.02 & Kepler-144c & 1.491 & 1.126 & 6157 & 10.105 & $1.037\times10^{-2}$ & 0.0952 & 1075 & 1.69 &\ 1537 & KOI & 1537.01 & & 1.824 & 1.366 & 6260 & 10.191 & $6.750\times10^{-3}$ & 0.1021 & 1167 & 1.35 &\ 1964 & KOI & 1964.01 & & 0.989 & 0.871 & 5547 & 2.229 & $6.910\times10^{-3}$ & 0.0319 & 1362 & 0.764 & 0.785\ 1964 & B & 1964.01 & & 0.556 & 0.569 & 3892 & 2.229 & $6.910\times10^{-3}$ & 0.0277 & 769 & 2.03 &\ 2311 & KOI & 2311.01 & & 1.182 & 0.975 & 5657 & 191.864 & $8.900\times10^{-3}$ & 0.6456 & 337 & 1.15 & 1.15\ 2311 & B & 2311.01 & & 0.270 & 0.256 & 3285 & 191.864 & $8.900\times10^{-3}$ & 0.4135 & 117 & 5.14 &\ 2311 & KOI & 2311.02 & & 1.182 & 0.975 & 5657 & 13.726 & $7.200\times10^{-3}$ & 0.1112 & 813 & 0.932 & 0.932\ 2311 & B & 2311.02 & & 0.270 & 0.256 & 3285 & 13.726 & $7.200\times10^{-3}$ & 0.0713 & 282 & 4.16 &\ 2365 & KOI & 2365.01 & Kepler-430b & 1.485 & 1.166 & 5884 & 35.968 & $2.003\times10^{-2}$ & 0.2244 & 667 & 3.25 &\ 2365 & KOI & 2365.02 & Kepler-430c & 1.485 & 1.166 & 5884 & 110.979 & $1.080\times10^{-2}$ & 0.4757 & 458 & 1.75 &\ 2593 & KOI & 2593.01 & & 1.453 & 1.230 & 6119 & 14.798 & $6.910\times10^{-3}$ & 0.1264 & 915 & 1.10 &\ 2755 & KOI & 2755.01 & & 1.172 & 0.973 & 5792 & 8.483 & $8.300\times10^{-3}$ & 0.0807 & 974 & 1.06 &\ 3097 & KOI & 3097.01 & Kepler-431d & 1.092 & 1.071 & 6004 & 11.922 & $9.300\times10^{-3}$ & 0.1045 & 856 & 1.11 &\ 3097 & KOI & 3097.02 & Kepler-431b & 1.092 & 1.071 & 6004 & 6.803 & $6.400\times10^{-3}$ & 0.0719 & 1032 & 0.764 &\ 3097 & KOI & 3097.03 & Kepler-431c & 1.092 & 1.071 & 6004 & 8.703 & $5.600\times10^{-3}$ & 0.0847 & 951 & 0.668 &\ 3204 & KOI & 3204.01 & & 1.593 & 1.553 & 7338 & 0.573 & $5.800\times10^{-3}$ & 0.0156 & 3268 & 1.01 &\ 3224 & KOI & 3224.01 & & 0.962 & 0.866 & 5382 & 3.439 & $6.340\times10^{-3}$ & 0.0425 & 1129 & 0.667 &\ 3255 & KOI & 3255.01 & & 0.622 & 0.615 & 4427 & 66.651 & $2.385\times10^{-2}$ & 0.2736 & 294 & 2.11 & 2.11\ 3255 & B & 3255.01 & & 0.592 & 0.593 & 4227 & 66.651 & $2.385\times10^{-2}$ & 0.2703 & 276 & 2.42 &\ 3284 & KOI & 3284.01 & & 0.463 & 0.479 & 3688 & 35.233 & $1.830\times10^{-2}$ & 0.1646 & 272 & 0.997 & 0.991\ 3284 & B & 3284.01 & & 0.189 & 0.163 & 3255 & 35.233 & $1.830\times10^{-2}$ & 0.1148 & 184 & 1.46 &\ 4407 & KOI & 4407.01 & & 1.435 & 1.234 & 6408 & 1.338 & $4.040\times10^{-3}$ & 0.0255 & 2121 & 0.650 & 0.660\ [^1]: https://cfop.ipac.caltech.edu/ [^2]: http://ad.usno.navy.mil/wds/orb6.html [^3]: http://exoplanetarchive.ipac.caltech.edu/

ArXiv

--- author: - Frits Beukers date: 'December 5, 2008' title: 'Algebraic A-hypergeometric functions' --- Introduction ============ The classically known hypergeometric functions of Euler-Gauss ($_2F_1$), its one-variable generalisations $_{p+1}F_p$ and the many variable generalisations, such as Appell’s functions, the Lauricella functions and Horn series are all examples of the so-called $A$-hypergeometric functions introduced by Gel’fand, Kapranov, Zelevinsky in [@1; @2; @3]. We like to add that completely independently B.Dwork developed a theory of generalised hypergeometric functions in [@13] which is in many aspects parallel to the theory of A-hypergeometric functions. The connection between the theories has been investigated in [@5] and [@14]. The definition of A-hypergeometric functions begins with a finite subset $A\subset\bbbz^r$ (hence their name) consisting of $N$ vectors $\v a_1,\ldots,\v a_N$ such that - The $\bbbz$-span of $\v a_1,\ldots,\v a_N$ equals $\bbbz^r$. - There exists a linear form $h$ on $\bbbr^r$ such that $h(\v a_i)=1$ for all $i$. The second condition ensures that we shall be working in the case of so-called Fuchsian systems. In a number of papers, eg [@5], this condition is dropped to include the case of so-called confluent hypergeometric equations. We are also given a vector of parameters $\alpha=(\alpha_1,\ldots,\alpha_r)$ which could be chosen in $\bbbc^r$, but we shall restrict to $\alpha\in\bbbr^r$. The lattice $L\subset\bbbz^N$ of relations consists of all $(l_1,\ldots,l_N)\in\bbbz^N$ such that $\sum_{i=1}^Nl_i\v a_i=0$. The A-hypergeometric equations are a set of partial differential equations with independent variables $v_1,\ldots,v_N$. This set consists of two groups. The first are the structure equations $$\Box_{\v l}\Phi:=\prod_{l_i>0}\partial_i^{l_i}\Phi-\prod_{l_i<0}\partial_i^{|l_i|}\Phi=0\eqno{{\rm (A1)}}$$ for all $\v l=(l_1,\ldots,l_N)\in L$. The operators $\Box_{\v l}$ are called the box-operators. The second group consists of the homogeneity or Euler equations. $$Z_i\Phi:=\left(a_{1,i}v_1\partial_1+a_{2,i}v_2\partial_2+\cdots +a_{N,i}v_N\partial_N-\alpha_i\right)\Phi =0,\ i=1,2,\ldots,r \eqno{{\rm (A2)}}$$ where $a_{k,i}$ denotes the $i$-th coordinate of $\v a_k$. In general the A-hypergeometric system is a holonomic system of dimension equal to the $r-1$-dimensional volume of the so-called A-polytope $Q(A)$. This polytope is the convex hull of the endpoints of the $\v a_i$. The volume-measure is normalised to $1$ for a $r-1$-simplex of lattice-points in the plane $h(\v x)=1$ having no other lattice points in its interior. In the first days of the theory of A-hypergeometric systems there was some confusion as to what ’general’ means, see [@5]. To avoid these difficulties we make an additional assumption, which ensures that the dimension of the A-hypergeometric system indeed equals the volume of $Q(A)$ - The $\bbbr_{\ge0}$-span of $A$ intersected with $\bbbz^r$ equals the $\bbbz_{\ge0}$-span of $A$. Under this condition we have the following Theorem. Let notations be as above. If condition [(iii)]{} is satisfied then the system of A-hypergeometric differential equations is holonomic of rank equal to the volume of the convex hull $Q(A)$ of $A$. For a complete story on the dimension of the solution space we refer to [@18]. In the present paper we shall use condition (iii) in the proof of the important Proposition \[modpSolutions1\]. To describe the standard hypergeometric solution of the A-hypergeometric system we define the projection map $\psi_L:\bbbr^N\to\bbbr^r$ given by $\psi_L:\v e_i\mapsto\v a_i$ for $i=1,\ldots,N$. Here $\v e_i$ denotes the $i$-th vector in the standardbasis of $\bbbr^N$. Clearly the kernel of $\psi_L$ is the space $L\otimes\bbbr$. Choose a point $\gamma=(\gamma_1,\ldots,\gamma_N)$ in $\psi_L^{-1}(\alpha)$, in other words choose $\gamma_1, \ldots,\gamma_N$ such that $\alpha=\gamma_1\v a_1+\cdots+\gamma_N\v a_N$. Then a formal solution of the A-hypergeometric system can be given by $$\Phi_{L,\gamma}(v_1,\ldots,v_N)=\sum_{\v l\in L}{\v v^{\v l}\over\Gamma(\v l+\gamma+\v 1)}$$ where we use the short-hand notation $${\v v^{\v l+\gamma}\over\Gamma(\v l+\gamma+\v 1)}={v_1^{l_1+\gamma_1}\cdots v_N^{l_N+\gamma_N} \over \Gamma(l_1+\gamma_1+1)\cdots\Gamma(l_N+\gamma_N+1)}.$$ By a proper choice of $\gamma\in\psi_L^{-1}(\alpha)$ this formal solution gives rise to actual powerseries solutions with a non-trivial region of convergence. The real positive cone generated by the vectors $\v a_i$ is denoted by $C(A)$. This is a polyhedral cone with a finite number of faces. We recall the following Theorem. \[irreducible\] The A-hypergeometric system is irreducible if and only if $\alpha+\bbbz^r$ contains no points in any face of $C(A)$. This Theorem is proved in [@4], Theorem 2.11 using perverse sheaves. It would be nice to have a more elementary proof however. Let us now assume that $\alpha\in\bbbq^r$. We shall be interested in those irreducible A-hypergeometric system that have a complete set of solutions algebraic over $\bbbc(v_1,\ldots,v_N)$. This question was first raised in the case of Euler-Gauss hypergeometric functions and the answer is provided by the famous list of H.A.Schwarz, see [@6]. In 1989 this list was extended to general one-variable $_{p+1}F_p$ by Beukers and Heckman, see [@7]. For the several variable cases, a characterization for Appell-Lauricella $F_D$ was provided by Sasaki [@8] in 1976 and Wolfart, Cohen [@9] in 1992. The Appell systems $F_4$ and $F_2$ were classified by M.Kato in [@10] (1997) and [@11] (2000). In the case of one-variable hypergeometric functions there is a simple combinatorial criterion to decide if they are algebraic or not. Consider the hypergeometric function $$_pF_{p-1}(\alpha_1,\ldots,\alpha_p;\beta_1,\ldots,\beta_{p-1}|z).$$ Define $\beta_p=1$. We assume $\alpha_i-\beta_j\not\in\bbbz$ for all $i,j=1,\ldots,p$, which ensures that the corresponding hypergeometric differential equation is irreducible. We shall say that the sets $\alpha_i$ and $\beta_j$ interlace modulo 1 if the points of the sets $e^{2\pi i\alpha_j}$ and $e^{2\pi i\beta_k}$ occur alternatingly when running along the unit circle. The following Theorem is proved in [@7]. Suppose the one-variable hypergeometric equation with parameters $\alpha_i,\beta_j\in\bbbq$ ($i,j=1,2,\ldots,p$) with $\beta_p=1$ is irreducible. Let $D$ be the common denominator of the parameters. Then the solution set of the hypergeometric equation consists of algebraic functions (over $\bbbc(z)$) if and only if the sets $k\alpha_i$ and $k\beta_j$ interlace modulo 1 for every integer $k$ with $1\le k<D$ and $\gcd(k,D)=1$. It is the purpose of this paper to generalize the interlacing condition to a similar condition for A-hypergeometric systems. We assume that $\alpha\in\bbbr^r$ and define $K_{\alpha}=(\alpha+\bbbz^r)\cap C(A)$. A point $\v p\in K_{\alpha}$ is called an [*apexpoint*]{} if $\v p\not\in\v q+C(A)$ for every $\v q\in K_{\alpha}$ with $\v q\ne\v p$. We call the number of apexpoints the [*signature*]{} of the polytope $A$ and parameters $\alpha$. Notation: $\sigma(A,\alpha)$. \[apexbound\] Let $\alpha\in\bbbr^r$. Then $\sigma(A,\alpha)$ is less than or equal to the volume of the $A$-polytope $Q(A)$. We say that the signature is [*maximal*]{} if it equals the volume of $Q(A)$. \[algebraic\] Let $\alpha\in\bbbq^r$ and suppose the A-hypergeometric system is irreducible. Let $D$ be the common denominator of the coordinates of $\alpha$. Then the solution set of the A-hypergeometric system consists of algebraic solutions (over $\bbbc(v_1,\ldots,v_N)$) if and only if $\sigma(A,k\alpha)$ is maximal for all integers $k$ with $1\le k<D$ and $\gcd(k,D)=1$. To compare this result with the one-variable interlacing condition for $_2F_1$ we illustrate a connection. In the case of Euler-Gauss hypergeometric function we have $r=3,N=4$ and $$\v a_1=(1,0,0),\ \v a_2=(0,1,0),\ \v a_3=(0,0,1),\ \v a_4=(1,1,-1).$$ The faces of the cone generated by $\v a_i$ ($i=1,\ldots,4$) are given by $x=0,y=0,x+z=0,y+z=0$ (we use the coordinates $x,y,z$ in $\bbbr^3$. We define $\alpha=(-a,-b,c-1)$. Theorem \[irreducible\] implies that irreducibility comes down to the inequalities $-a,-b,-a+c,-b+c\not\in\bbbz$. These are the familiar irreducibility conditions for the Euler-Gauss hypergeometric functions. The lattice of relations has rank one and is generated by $(-1,-1,1,1)$. We choose $\gamma=(-a,-b,c-1,0)$. Then the formal solution $\Phi_{L,\gamma}$ reads $$v_1^{-a} v_2^{-b} v_3^{c-1}\sum_{k\in\bbbz} {v_1^{-k} v_2^{-k} v_3^k v_4^k\over\Gamma(-k-a+1) \Gamma(-k-b+1)\Gamma(c+k)\Gamma(k+1)}.$$ Clearly $1/\Gamma(k+1)$ vanishes for $k\in\bbbz_{<0}$, so our summation actually runs over $k\in\bbbz_{\ge0}$. Apply the identity $1/\Gamma(1-z)=\sin(\pi z)\Gamma(z)/\pi$ to $z=k+a$ and $z=k+b$ to obtain $$\Phi_{L,\gamma}=v_1^{-a} v_2^{-b} v_3^{c-1}{\sin(\pi a)\sin(\pi b)\over\pi^2} \sum_{k=0}^{\infty}{\Gamma(a+k)\Gamma(b+k)\over\Gamma(c+k)k!}\left( {v_3v_4\over v_1v_2}\right)^k.$$ Setting $v_1=v_2=v_3=1$ and $v_4=z$ we recognize the Euler-Gaussian hypergeometric series $_2F_1(a,b,c|z)$. By shifting over $\bbbz$ if necessary we can see to it that $a,b,c$ are in the interval $(-1,0)$. Suppose that the sets $\{a,b\}$ and $\{0,c\}$ interlace modulo $1$. By interchange of $a,b$ if necessary we can restrict ourselves to the case $-1<a<c<b<0$. It is straightforward to verify that $(-a,1-b,c)$ and $(-a,-b,1+c)$ are apexpoints of $K_{\alpha}$. If the sets do not interlace then one checks that $(-a,-b,c)$ is the unique apexpoint if $a,b<c$ and $(-a,-b,1+c)$ is the unique apexpoint if $c<a,b$. A second example is Appell’s hypergeometric equation $F_2$. The Appell $F_2$ hypergeometric function reads $$F_2(a,b,b',c,c'|x,y)=\sum_{m=0}^{\infty}\sum_{n=0}^{\infty} {(a)_{m+n}(b)_m(b')_n\over (c)_m(c')_n m! n!}\ x^m y^n,$$ where $(x)_n$ denotes the Pochhammer symbol defined by $\Gamma(x+n)/\Gamma(x)=x(x+1)\cdots(x+n-1)$. The function $F_2$ satisfies a system of partial differential equations of rank 4. Algebraicity of these functions is completely described in [@11]. The A-parameters are as follows. We have $r=5$ and $N=7$. The set $A$ consists of the standard basisvectors $\v a_1,\ldots,\v a_5$ in $\bbbr^5$ and $\v a_6=(1,1,0,-1,0),\ \v a_7=(1,0,1,0,-1)$. We take $\alpha=(-a,-b,-b',c,c')$. The lattice $L$ of relations is generated by $(-1,-1,0,1,0,1,0)$ and $(-1,0,-1,0,1,0,1)$. Take $\gamma=(-a,-b,-b',c,c',0,0)$. In a similar way as we did for the Euler-Gauss functions we can now go from the formal expansion $\Phi_{L,\gamma}$ to the explicit Appell function $F_2$. One can compute that the cone $C(A)$ has 8 faces and they are given by $x_1=0, x_2=0, x_3=0$, $x_2+x_4=0, x_3+x_5=0, x_1+x_4=0, x_1+x_5=0$ and $x_1+x_4+x_5=0$. Using Theorem \[irreducible\] it follows that the A-hypergeometric system is irreducible if and only if none of the following numbers is an integer, $$a,b,b',-b+c,-b'+c',-a+c,-a+c',-a+c+c'.$$ These are precisely the irreducibility condtions for the $F_2$-system given in [@11]. In that paper it is shown for example that with the choice $$\alpha=(-a,-b,-b',c,c')=(1/10, 7/10, 9/10, 3/5, 1/5)$$ the solutions of the rank 4 Appell system are all algebraic. A small computer calculation shows that the apex points of $K_{\alpha}$ are given by $$(21/10, 7/10, 9/10, -2/5,-4/5),\quad (11/10, 7/10, 9/10, 3/5, -4/5)$$ $$11/10, 7/10, 9/10, -2/5, 1/5),\quad (1/10, 7/10, 9/10, 3/5, 1/5).$$ Similarly there are four apexpoints for the conjugate parameter 5-tuples $3\alpha,7\alpha,9\alpha$. A simple example ================ In this section we show a more elaborate example of algebraic hypergeometric functions of Horn-type G3 which, to our knowledge, has not been dealt with before. The corresponding series with parameters $a,b$ is given by $$G_3(a,b,x,y)=\sum_{m\ge0,n\ge0}{(a)_{2m-n}(b)_{2n-m}\over m!n!}x^my^n.$$ Here again, $(x)_n$ denotes the Pochhammer symbol defined by $\Gamma(x+n)/\Gamma(x)$. However, now the index $n$ may be negative, in which case the definition explicitly reads $\Gamma(x+n)/\Gamma(x)=1/(x-1)\cdots(x-|n|)$. The system of differential equations is a rank 3 system. The set $A$ can be chosen in $\bbbz^2$ for example as $$\v a_1=(-1,2),\quad \v a_2=(0,1),\quad \v a_3=(1,0),\qquad \v a_4=(2,-1).$$ Below we show a picture of the cone $C(A)$ spanned by the elements of $A$, together with the points from $A$. In addition, the dark grey area indicates the set of apexpoints with respect to $A$. The parameter vector of the corresponding A-hypergeometric system is given by $(-a,-b)$. \[G3\] Consider the A-hypergeometric system corresponding to the Horn G3 equations. In the following cases the system is irreducible and has only algebraic solutions. 1. $a+b\in\bbbz$ and $a,b\not\in\bbbz$. 2. $a\is1/2\mod{\bbbz}, b\is 1/3,2/3\mod{\bbbz}$ 3. $a\is1/3,2/3\mod{\bbbz}, b\is1/2\mod{\bbbz}$ [**Proof**]{} This is an application of Theorem \[algebraic\]. In all cases we need only be interested in $a,b\mod{\bbbz}$. In the following picture the light grey area is the cone $C(A)$, the dark grey area indicates the location of the apexpoints. If $a+b\in\bbbz$ then we note that there are precisely 3 points of $(-a,-b)+\bbbz^2$ in the dark grey area, all lying on the line $x+y=1$. Hence three apexpoints. {height="4.5cm"} If $a+b\in\bbbz$ then also $ka+kb\in\bbbz$ for any integer. Irreducibility of the systems is ensured by Theorem \[irreducible\] and the fact that $a,b\not\in\bbbz$. Therefore, in the first case all conditions of Theorem \[algebraic\] are fulfilled. In the following picture we have drawn the sets $(1/2,1/3)+\bbbz^2$ and $(1/2,2/3)+\bbbz^2$ intersected with $C(A)$. {height="4.5cm"} Clearly each set has three apexpoints and Theorem \[algebraic\] can be applied to prove the second case. The third case runs similarly. It has been verified by J.Schipper, an Utrecht graduate student, that Theorem \[G3\] gives the characterisation of all irreducible Horn G3-systems with algebraic solutions. A fairly involved calculation reveals that a formula for $G_3(a,1-a,x,y)$ can be given as follows $$G_3(a,1-a,x,y)=f(x,y)^a\sqrt{{g(x,y)\over \Delta}}$$ where $$\Delta=1+4x+4y+18xy-27x^2y^2$$ and $$xf^3-y=f-f^2,\qquad g(g-1-3x)^2=x^2\Delta.$$ For reference we display the series expansions of $f$ and $g$. $$\begin{aligned} f&=&1+(y-x) +\left(2 x^2-y x-y^2\right) +\left(-5 x^3+3 y x^2+2 y^3\right)\\ &&+\left(14 x^4-10 y x^3+y^3 x-5y^4\right)+O(x,y)^5 \end{aligned}$$ $$g=1+2 x -x(x+2 y)+2x(x-y)^2 -x(x-y)^2(5 x+4 y) +O(x,y)^5$$ Moreover, $g=1+4x-2xf-3x^2f^2$. In particular, $f$ and $g$ generate the same cubic extension of $\bbbq(x,y)$. The signature ============= [**Proof**]{} of Proposition \[apexbound\]. We use the following property. Let $\v b_1,\ldots,\v b_r\in\bbbz^r$ be independent vectors. Let $\beta\in\bbbr^r$. Then the number of points of $\beta+\bbbz^r$ inside the fundamental block $\{\sum_{i=1}^r\lambda_i\v b_i\ |\ 0\le\lambda_i<1\}$ is equal to $|\det(\v b_1,\ldots,\v b_r)|$. Write $Q(A)$ as a union of $r-1$-simplices $\cup_{i=1}^m\sigma_i$ (a so-called triangulation of $Q(A)$). Every simplex $\sigma_i$ is spanned by $r$ independent vectors $\v a_{i_j}$ ($j=1,\ldots,r$). Let $B_i$ be the fundamental block spanned by these vectors. Let $\v a$ be an apexpoint of $K_{\alpha}$. Then $\v a$ is contained in a positive cone spanned by one of the simplices $\sigma_i$. For every choice of $\v a_{i_j}$ ($j=1,\ldots,r$) the point $\v a-\v a_{i_j}$ falls outside this cone. If not, then $\v a$ would be contained in $\v a_{i_j}+C(A)$. Hence $\v a\in B_i$. Since $B_i$ contains at most $|\det(B_i)|$ point from $\alpha+\bbbz^r$, we see that the number of apexpoints is bounded above by $\sum_{i=1}^m|\det(B_i)|$. This equals precisely the $r-1$-dimensional volume of the $A$-polytope $Q(A)$. For any $\v k=(k_1,\ldots,k_N)\in\bbbz^N$ we define the hypercube $$F(\v k)=\{(x_1,\ldots,x_N)| k_i\le x_i<k_i+1,\ (i=1,2,\ldots,N)\}$$ The space $\bbbr^N$ can be seen as the disjoint union of cells $F(\v k)$ with $\v k$ running over $\bbbz^N$. Let us denote $L(\bbbr)=L\otimes\bbbr$. We intersect the union $\cup_{\v k}F(\v k)$ with the translated space $\gamma+L(\bbbr)$. Each hypercube intersects $\gamma+L(\bbbr)$ in a cell which may, or may not be closed in $\gamma+L(\bbbr)$. Let us denote $V(\v k)= F(\v k)\cap(\gamma+L)$. We call $V(\v k)$ a compact cell if it is closed and non-empty. Of course, if we shift a compact cell over a point of $L$, we get another compact cell. In the following Proposition we denote the shifted hyperquadrant $\v x+\bbbr_{\ge0}^N$ by $P(\v x)$. \[compactcell\] With the notations as above, $V(\v k)$ is a compact cell if and only if $P(\v k)$ has non-trivial intersection with $\gamma+L$ and $P(\v k+\v e_i)$ has empty intersection with $\gamma+L$ for $i=1,2,\ldots,N$. In particular, $V(\v k)=P(\v k)\cap(\gamma+L(\bbbr))$. [**Proof**]{}. Let us denote the intersection of $\cup_{i=1}^N P(\v k+\v e_i)$ with $\gamma+L$ by $W$. Notice that $W$ is the set-theoretic difference between $P(\v k)\cap(\gamma+L)$ and $V(\v k)$. In particular $W$ is a closed set. Suppose that $V(\v k)$ is a compact cell. The only way that the difference $W$ of the two non-empty closed convex sets $P(\v k)\cap(\gamma+L)$ and $V(\v k)$ can be closed is when $W$ is empty. Hence $P(\v k+\v e_i)\cap(\gamma+L)$ is empty for all $i$. Suppose conversely that $W$ is empty. Then $V(\v k)=P(\v k)$. Since $P(\v k)$ is closed, the same should hold for $V(\v k)$. Since $V(\v k)$ is also bounded, we conclude that $V(\v k)$ is compact. In the following recall the map $\psi:\bbbr^N\to\bbbr^r$ given by $\psi:\v e_i\mapsto \v a_i$ for $i=1,\ldots,N$. \[cell\_apex\] The compact cells in $\gamma+L(\bbbr)$, modulo $L$, are in 1-1 correspondence with the apex-points of $K_{\alpha}$. The correspondence is given by $V(\v k)\mapsto\alpha-\psi(\v k)$. Let $\v a,\v a'$ be two different apexpoints. Then $\psi^{-1}(\v a)\cap P(\v 0)$ and $\psi^{-1}(\v a')\cap P(\v 0)$ are disjoint and contained in the unit cube $0\le x_i<1$ for $i=1,2,\ldots,N$. [**Proof**]{} An apexpoint $\v a\in\alpha+\bbbz^r$ is characterized by the fact that $\v a\in C(A)$ and $\v a-\v a_i\not\in C(A)$ for $i=1,\ldots, N$. Notice that $\psi:\bbbr^N\to\bbbr^r$ is actually the quotient map $\bbbr^N\to\bbbr^N/L(\bbbr)$. Let $\v k\in\bbbz^N$. First recall that $\psi(\gamma)=\alpha$ and observe that $\psi(P(\v k))=\psi(\v k)+C(A)$. As a result we see that $P(\v k)\cap(\gamma+L(\bbbr))$ is non-empty if and only if $\alpha-\psi(\v k)\in C(A)$. Let us assume that $V(\v k)$ is compact and apply Proposition \[compactcell\]. Then $\alpha-\psi(\v k)\in C(A)$ and $\alpha-\psi(\v k)-\v a_i\not\in C(A)$ for $i=1,\ldots,N$. Hence $\alpha-\psi(\v k)$ is an apexpoint. Conversely, when $\v a$ is an apexpoint, find $\v k\in\bbbz^N$ such that $\v a=\alpha-\psi(\v k)$. Then we find that $(\gamma+L(\bbbr))\cap P(\v k)$ is non empty and the sets $\psi^{-1}(\alpha)\cap P(\v k+\v e_i)$ for $i=1,2,\ldots,N$ are empty. Therefore, by application of Proposition \[compactcell\], $V(\v k)$ is a compact cell. Let $\v a$ be an apexpoint and choose $\v k\in\bbbz^N$ such that $\alpha-\psi(\v k)=\v a$. Then $\psi^{-1}(\alpha)\cap P(\v k)$ is a compact cell. Consequently, after shifting over $\v k$, the set $\psi^{-1}(\v a)\cap P(\v 0)$ is a compact cell in $\gamma-\v k+L(\bbbr)$. Hence $\psi^{-1}(\v a)\cap P(\v 0)$ is contained in the unit cube in $\bbbz^N$. Two sets $\psi^{-1}(\v a)$ and $\psi^{-1}(\v a')$ are obviously distinct whenever $\v a$ and $\v a'$ are distinct. Mod $p$ solutions ================= Let us assume that $\alpha\in\bbbz^r$ and let $p$ be a prime. We describe the polynomial solutions in $\bbbf_p[v_1,\ldots,v_N]$ of the A-hypergeometric system with parameters $A$ and $\alpha$ considered modulo $p$. Let $\beta_i/p$ ($i=1,\ldots,\sigma$) be the set of apexpoints of $(\alpha/p+\bbbz^r)\cap C(A)$. To any apexpoint we associate the set of lattice points $$\Gamma_i=\psi^{-1}(\beta_i)\cap\bbbz_{\ge0}^N$$ where $\psi:\bbbr^N\to \bbbr^N/L(\bbbr)$ is defined as in the previous section. For any $i=1,2,\ldots,\sigma$ we define $$\Psi_i:=\sum_{\v l\in\Gamma_i}{{\v v}^{\v l}\over\Gamma(\v l+1)}.$$ This is a polynomial solution to the A-hypergeometric system with parameters $A,\beta_i$. Since $\beta_i/p$ is an apexpoint of $(\alpha/p+\bbbz^r)\cap C(A)$, the preimage $\psi^{-1}(\beta_i/p)\cap P(\v 0)$ is contained in the unit cube in $\bbbr^N$ according to Proposition \[cell\_apex\]. Hence the points of $\Gamma_i=\psi^{-1}(\beta_i)\cap\bbbz_{\ge0}^N$ are contained in the cube $0\le x_i<p$ for $i=1,\ldots,N$. In particular none of the positive coordinates of any $\v l\in\Gamma_i$ is divisible by $p$, hence $\Gamma(\v l+\v 1)\not\is0\mod{p}$ for all $\v l\in\Gamma_i$. This means that $\Psi_i$ can be reduced modulo $p$. Furthermore, since $\alpha\is\beta_i\mod{p}$, the polynomial $\Psi_i$ is a polynomial solution modulo $p$ to the A-hypergeometric system with parameters $A,\alpha$. Since each of the sets $\Gamma_i$ is contained in the cube $0\le x_i<p$ any two shifts $\Gamma_i+p\v k_i$ and $\Gamma_j+p\v k_j$ for different $i,j$ and $\v k_i,\v k_j\in\bbbz^N$ are disjoint. In particular the polynomials $\Psi_i$ are independent over $\bbbf_p[v_1^p,\ldots,v_N^p]$. \[modpSolutions1\] Every mod $p$ polynomial solution of the A-hypergeometric system with parameters $A,\alpha\in\bbbz^r$ is an $\bbbf_p[v_1^p,\ldots,v_N^p]$-linear combination of the polynomials $\Psi_i$. Let $P=\sum_{\v m}p_{\v m}\v v^{\v m}$ be a polynomial solution of the A-hypergeometric system with parameters $A,\alpha$. For formal reasons we extend the summation over all of $\bbbz^N$ but it should be understood that the set of multi-indices $\v m$ with $p_{\v m}\not\is0\mod{p}$ is finite and contained in $\bbbz_{\ge0}^N$. Let us substitute this in the system (A1). Any $\v l\in L$ ca be decomposed as $\v l=\v l_+-\v l_-$ where $\v l_+,\v l_-\in\bbbz_{\ge0}^N$ and we assume they have disjoint support. Denoting $[\v m]_{\v r}=\prod_{i=1}^N m_i(m_i-1)\cdots(m_i-r_i+1)$ for any $\v r\in\bbbz_{\ge0}^N$ we obtain $$\begin{aligned} 0&\is&\sum_{\v m}\left([\v m]_{l_+}p_{\v m}\v v^{\v m-l_+}- [\v m]_{l_-}p_{\v m}\v v^{\v m-l_-}\right)\mod{p}\\ &\is&\v v^{-l_+}\sum_{\v m}\left([\v m]_{l_+}p_{\v m}\v -[\v m-\v l]_{l_-}p_{\v m-\v l} \right)v^{\v m}\mod{p}\end{aligned}$$ Hence $$\label{recursion} [\v m]_{l_+}p_{\v m} -[\v m-\v l]_{l_-}p_{\v m-\v l}\is0\mod{p}$$ for every $\v m\in \bbbz^N$ and every $\v l\in L$. Substitution in (A2) gives $$\sum_{\v m}(-\alpha_i+a_{1i}m_1+\cdots+a_{Ni}m_N)\v p_{\v m}v^{\v m}\is0\mod{p}$$ for $i=1,2,\ldots,r$. Hence $-\alpha_i+a_{1i}m_1+\cdots+a_{Ni}m_N\is0\mod{p}$ for every $\v m$ with $p_{\v m}\not\is0\mod{p}$. Note that the system (A2) gives no extra relations between different $p_{\v m}$. They only require that $\psi(\v m)\is\alpha\mod{p}$. The system (A1) relates only those coefficients $p_{\v m}$ for which the multi-indexes $\v m$ differ by an element of $L$. Hence we can split $P$ as a sum of terms of the form $P_{\beta}=\sum_{\psi(\v m)=\beta}p_{\v m}\v v^{\v m}$ with $\beta\in\bbbz^r$ and each summand $P_{\beta}$ satisfies modulo $p$ the A-hypergeometric system with parameters $A,\alpha$. The equations (A2) applied to $P_{\beta}$ tell us that $\beta\is\alpha\mod{p}$. We define a partial ordering on $\bbbr^N$. We say that $\v y\ge \v x$ if all components of $\v y$ are larger or equal than the corresponding component of $\v x$. In particular, when $\v y\ge \v x$ and $\v y\ne\v x$ we write $\v y>\v x$. When $\v m=(m_1,\ldots,m_N)\in\bbbz^N$ we denote by $\lfloor \v m/p\rfloor$ the vector $(\lfloor m_1/p\rfloor,\ldots,\lfloor m_N/p\rfloor)$. Consider the recursion (\[recursion\]). We claim that $[\v m]_{l_+}\is0\mod{p}$ if and only if $\lfloor \v m/p\rfloor-\lfloor (\v m-\v l)/p\rfloor$ has at least one positive component. This can be seen through the following sequence of equivalences, $$\begin{aligned} [\v m]_{\v l_+}\is0\mod{p} &\iff& \exists i:\ m_i(m_i-1)\cdots(m_i-l_i+1)\is0\mod{p}\\ &\iff& \exists i,\lambda:\ 0\le\lambda<l_i,\ m_i-\lambda\is0\mod{p}\\ &\iff& \exists i,\lambda:\ 0\le\lambda<l_i,\ (m_i-\lambda)/p\in\bbbz\\ &\iff& \exists i:\ \lfloor m_i/p\rfloor -\lfloor(m_i-l_i)/p\rfloor>0\end{aligned}$$ Similarly we see that $[\v m-\v l]_{l_-}\is0\mod{p}$ if and only if $\lfloor \v m/p\rfloor-\lfloor (\v m-\v l)/p\rfloor$ has at least one negative component. In terms of our partial ordering this implies that $[\v m]_{l_+}\is0\mod{p}$ and $[\v m-\v l]_{l_-}\is0\mod{p}$ if and only if neither $\lfloor \v m/p\rfloor\ge \lfloor (\v m-\v l)/p\rfloor$ nor $\lfloor \v m/p\rfloor\le \lfloor (\v m-\v l)/p\rfloor$, i.e $\lfloor \v m/p\rfloor$ and $\lfloor (\v m-\v l)/p\rfloor$ are unrelated. Write $\v m'=\v m-\v l$, then $p_{\v m}$ and $p_{\v m'}$ are related through (\[recursion\]) if and only if $\psi(\v m)=\psi(\v m')$ and $\lfloor \v m/p\rfloor$ and $\lfloor \v m'/p\rfloor$ are related. Now suppose that $p_{\v m}\not\is0\mod{p}$. We assert that for any $\lambda\in L(\bbbr)$ the inequality $\lfloor \v m/p\rfloor \le \lfloor (\v m-\lambda)/p\rfloor$ implies equality. First we deal with the case when $\lambda=\v l\in L$. Suppose $\lfloor \v m/p\rfloor < \lfloor (\v m-\v l)/p\rfloor$. Then $[\v m]_{l_+}\not\is 0\mod{p}$ and $[\v m-\v l]_{l_-}\is0\mod{p}$. This gives a contradiction with relation (\[recursion\]). Hence $\lfloor \v m/p\rfloor \le \lfloor (\v m-\v l)/p\rfloor$ implies equality. Now, in general, suppose that there exists $\lambda\in L(\bbbr)$ such that $\lfloor \v m/p\rfloor < \lfloor (\v m-\lambda)/p\rfloor$. The vector $\v m-\lambda-p\lfloor(\v m-\lambda)/p\rfloor$ has non-negative coefficients. Hence its image under $\psi$ is contained in the cone $C(A)$. Moreover, since $\psi(\lambda)=0$, the image has integer coordinates. Choose a vector $\v k\in\bbbz_{\ge0}^N$ such that $\psi(\v k)=\psi(\v m- p\lfloor(\v m-\lambda)/p\rfloor)$. Notice that this is only possible because of Assumption iii) which we made in the introduction. Hence there exists $\v l\in L$ such that $\v k=\v m-\v l-p\lfloor(\v m-\lambda)/p\rfloor$. In particular, $\lfloor(\v m-\v l)/p\rfloor \ge \lfloor(\v m-\lambda)/p\rfloor$. Since, by assumption, the latter vector is strictly larger than $\lfloor\v m/p\rfloor$ we again get a contradiction. Hence we conclude that $$\label{equality} \lfloor(\v m-\lambda)/p\rfloor\ge \lfloor\v m/p\rfloor \Rightarrow \lfloor(\v m-\lambda)/p\rfloor= \lfloor\v m/p\rfloor.$$ Another way of phrasing property (\[equality\]) is to say that $\v m/p$ is contained in a compact cell of the affine space $\v m/p+L(\bbbr)$. To see this consider the cell $V(\lfloor m/p\rfloor)$. Of course it contains $\v m/p$. Let now $(\v m-\lambda)/p$ be any other point in $P(\lfloor \v m/p\rfloor)\cap \v m/p+L(\bbbr)$. Then $\lfloor(\v m-\lambda)/p\rfloor \ge\lfloor\v m/p\rfloor$ and we have seen that this implies equality. Therefore $(\v m-\lambda)/p$ is contained in $V(\lfloor\v m/p\rfloor)$. Hence the latter cell is compact by Proposition \[compactcell\]. Let $\beta$ be as in the polynomial $P_{\beta}$ above and $\gamma\in\bbbr^N$ such that $\psi(\gamma)=\beta$. Let $\v m,\v m'\in\bbbz_{\ge0}^N$ be such that $\psi(\v m)=\psi(\v m')=\beta$ and such that $\v m/p,\v m'/p$ are in a compact cell of $\gamma/p+L(\bbbr)$. Let $\v l=\v m-\v m'$. If $\v m/p,\v m'/p$ belong to different compact cells we have neither $\lfloor \v m/p\rfloor \le \lfloor\v m'/p\rfloor$ nor $\lfloor \v m/p\rfloor \ge \lfloor\v m'/p\rfloor$. Hence $p_{\v m}$ and $p_{\v m'}$ are unrelated by relation (\[recursion\]). As a consequence of this all, the polynomial $P_{\beta}$ splits as a sum of terms of the form $\sum_{\lfloor \v m/p\rfloor=\v k}p_{\v m}\v v^{\v m}$ and each such sum is a solution of (A1) and (A2). The latter summation can be rewritten as $\v v^{p\v k}\sum_{\lfloor\v m/p\rfloor=\v k}p_{\v m}\v v^{\v m-p\v k}.$ The multi-indices $\v m$ in $\sum_{\lfloor\v m/p\rfloor=\v k}p_{\v m}\v v^{\v m-p\v k}$ should in addition satisfy $\psi(\v m)=\beta$. Replace $\v m$ by $\v n+p\v k$ and we obtain the solution $$\label{reducedsolution} \sum_{\lfloor\v n/p\rfloor=\v 0}b_{\v n}\v v^{\v n},$$ where we put $b_{\v n}=p_{n+p\v k}$. We now know that all multi-indices $\v n$ are contained in the cube $0\le x_i<p$ for $i=1,\ldots,N$. Furthermore, in the recursion relation $$[\v n]_{l_+}b_{\v n} -[\v n-\v l]_{l_-}b_{\v n-\v l}\is0\mod{p}$$ both coefficients are non-zero whenever $\v n\ge\v0$ and $\v n-\v l\ge\v 0$. Hence the space of solutions of the form (\[reducedsolution\]) has dimension at most one. On the other hand we do have such a solution, namely $\Psi_i$ where $i$ is chosen such that the apexpoint $\beta_i$ is equal to the apexpoint $\beta-\psi(\v k)$. We now consider polynomial mod $p$ solutions for A-hypergeometric systems with parameters $\alpha\in\bbbq^r$. \[modpSolutions2\] Let $\alpha\in\bbbq^r$ and let $D$ the common denominator of the coordinates of $\alpha$. Let $p$ be a prime not dividing $D$. Let $\rho\is -p^{-1}\mod{D}$ if $D>1$ and $\rho=1$ if $D=1$. Let $s$ be the signature of $A$ and $\rho\alpha$. Suppose that the A-hypergeometric system we consider is irreducible. Then, when $p$ is sufficiently large, the polynomial mod $p$ solutions of the A-hypergeometric system with parameters $A,\alpha$ is a free $\bbbf_p[v_i^{p}]$-module of rank $s$. [**Proof**]{}. Let $\v k=(1+p\rho)\alpha$. Notice that $\v k\in\bbbz^r$ and $\v k\is \alpha\mod{p}$. So it suffices to look at the mod $p$ A-hypergeometric system with parameters $A,\v k$. In Proposition \[modpSolutions1\] we saw that these solutions form a free module of rank $s'$ where $s'$ is the signature of $A$ and $\v k/p$. Let $\delta$ be the minimal distance of the points of $\rho\alpha+\bbbz^r$ to the faces of $C(A)$. Suppose $\delta=0$. Then there is a point $\rho\alpha+\v k$ with $\v k\in\bbbz^r$ contained in a face of $C(A)$. Choose $\mu\in\bbbz$ such that $\mu\rho\is1\mod{D}$. Then $\mu(\rho\alpha+\v k)=\alpha+\mu\v k+(\mu\rho-1)\alpha$ is on a face of $C(A)$. This contradicts the irreducibility of our A-hypergeometric system by Theorem \[irreducible\]. So $\delta>0$. Let us assume that $p$ is so large that $|\alpha/p|<\delta$. Then the points of $(\rho\alpha+\bbbz^r)\cap C(A)$ and $(\v k/p+\bbbz^r)\cap C(A)$ are in one-to-one correspondence given by $\v x\sim\v y\iff |\v x-\v y|<\delta$. In particular the number of apexpoints of both sets is equal, hence $s=s'$. This proves our assertion. Proof of the main theorem ========================= This section is devoted to a proof of Theorem \[algebraic\]. Let notations be as in Theorem \[algebraic\] and suppose we consider an irreducible A-hypergeometric system with parameters $\alpha\in\bbbq^r$. Let $p$ be a prime which is large enough in the sense of Proposition \[modpSolutions2\]. Let $D$ be the common denominator of the elements of $\alpha$ and $\rho\is-p^{-1}\mod{D}$. Then the statement that $\sigma(A,\rho\alpha)$ is maximal is equivalent to the statement that the A-hypergeometric system modulo $p$ has a maximal $\bbbf(\v v^p)$-independent set of polynomial solutions. A fortiori the following two statements are equivalent: i\) $\sigma(A,k\alpha)$ is maximal for every $k$ with $1\le k<D$ and $\gcd(k,D)=1$\ ii) modulo almost every prime $p$ the A-hypergeometric system modulo $p$ has a maximal set of polynomial solutions modulo $p$. A famous conjecture, attributed to Grothendieck implies that statement (ii) is equivalent to the following statement, iii\) The A-hypergeometric system has a complete set of algebraic solutions. If Grothendieck’s conjecture were proven we would be done here. Fortunately, in two papers by N.M.Katz ([@12] and [@15]) Grothendieck’s conjecture is proven in the case when the system of differential equations is (a factor of) a Picard-Fuchs system, i.e. a system of differential equations satisfied by the period integral on families of algebraic varieties. More precisely we refer to Theorem 8.1(5) of [@15], which states Suppose we have a system of partial linear differential equations, as sketched above, whose $p$-curvature vanishes for almost all $p$. Then, if the system is a subsystem of a Picard-Fuchs system, the solution space consists of algebraic functions. The above theorem is formulated in terms of vanishing $p$-curvature for almost all $p$, but according to a Lemma by Cartier (Theorem 7.1 of [@15]) this is equivalent to the system having a maximal set of independent polynomial solutions modulo $p$ for almost all $p$. To finish the proof of Theorem \[algebraic\] it remains to show that the A-hypergeometric equations for $\alpha\in\bbbq^r$ do arise from algebraic geometry. We shall do so in Sections \[pochhammer\] and \[eulerintegral\], where we construct Euler type integrals for the solutions of the A-hypergeometric system. In an attempt to maintain the lowtech nature of this paper we finish this Section with a proof of the (easier) implication (iii)$\Rightarrow$ (ii). Before doing so we need a few introductory concepts from the theory of linear differential equations. Let $k$ be a field which, in our case, is usually $\bbbq$ or $\bbbf_p$. Consider the differential field $K=k(v_1,\ldots,v_N)=k(\v v)$ with derivations $\partial_i={\partial\over\partial v_i}$ for $i=1,\ldots,N$. The subfield $C_K\subset K$ of elements all of whose derivatives are zero, is called the field of constants. When the characteristic of $k$ is zero we have $C_K=k$, when the characteristic is $p>0$ we have $C_K=k(\v v^p)$. Throughout this section we let ${\cal L}$ be a finite set of linear partial differential operators with coefficients in $K$, like the A-hypergeometric system operators when $k=\bbbq$. Consider the differential ring $K[\partial_1,\ldots, \partial_N]$ and let $({\cal L})$ be the left ideal generated by the differential operators of the system. We assume that the quotient $K[\partial_i]/({\cal L})$ is a $K$-vector space of finite dimension $d$. Throughout this section we also fix a monomial $K$-basis $\partial^{\v b}=\partial_1^{b_1}\cdots\partial_N^{b_N}$ with $\v b\in B$ and where $B$ is a finite set of $N$-tuples in $\bbbz_{\ge0}^N$ of cardinality $d$. \[wronskian\] Let ${\cal K}$ be some differential extension of $K$ with field of constants $C_K$. Let $f_1,\ldots,f_m\in{\cal K}$ be a set of $C_K$-linear independent solutions of the system $L(f)=0,\ L\in{\cal L}$. Then $m\le d$. Moreover, if $m=d$ the determinant $$W_B(f_1,\ldots,f_d)=\det(\partial^{\v b}f_i)_{\v b\in B;i=1,\ldots,d}$$ is nonzero. In case we have $d$ independent solutions we call $W_B$ the [*Wronskian matrix*]{} with respect to $B$ and $f_1,\ldots,f_d$. Obviously, if $g_1,\ldots,g_d$ are $C_K$-linear dependent solutions then $W_B(g_1,\ldots,g_d)=0$. [**Proof**]{}. Suppose that either $m>d$ or $m=d$ and $W_B=0$. In both cases there exists a ${\cal K}$-linear relation between the vectors $df_i:=(\partial^{\v b}f_i)_{\v b\in B}$ for $i=1,2,\ldots,m$. Choose $\mu<m$ maximal such that $df_i,\ i=1,\ldots,\mu$ are ${\cal K}$-linear independent. Then, up to a factor, the vectors $df_i,\ i=1,\ldots,\mu+1$ satisfy a unique dependence relation $\sum_{i=1}^{\mu+1}A_idf_i=0$ with $A_i\in{\cal K}$ not all zero. For any $j$ we can apply the operator $\partial_j$ to this relation to obtain $$\sum_{i=1}^{\mu+1}\partial_j(A_i)df_i+A_i\partial_j(df_i)=0.$$ Since $\partial_j\partial^{\v b}$ is a $K$-linear combination of the elements $\partial^{\v b},\v b\in B$ in $K[\partial_i]/({\cal L})$ there exists a $d\times d$-matrix $M_j$ with elements in $K$ such that $\partial_j(df_i)=df_i\cdot M_j$. Consequently $\sum_{i=1}^{\mu+1}A_i\partial_j(df_i)=\sum_{i=1}^{\mu+1}A_if_i\cdot M_j=0$ and so we are left with $\sum_{i=1}^{\mu+1}\partial_j(A_i)df_i=0$. Since the relation between $df_i,i=1,\ldots,\mu+1$ is unique up to factor there exists $\lambda_j\in K$ such that $\partial_j(A_i)=\lambda_jA_i$ for all $i$. Suppose $A_1\ne0$. Then this implies that $\partial_j(A_i/A_1)=0$ for all $i$ and all $j$. We conclude that $A_i/A_1\in C_K$ for all $i$. Hence there is a relation between the $df_i$ with coefficients in $C_K$. A fortiori there is a $C_K$-linear relation between the $f_i$. This contradicts our assumption of independence of $f_1,\ldots,f_m$. So we conclude that $m\le d$ and if $m=d$ then $W_B\ne0$. \[algebraic2modP\] Suppose the system of equations $L(y)=0,\ L\in{\cal L}$ has only algebraic solutions and that they form a vector space of dimension $d$. Then for almost all $p$ the system of equations modulo $p$ has a $\bbbf(\v v^p)$-basis of $d$ polynomial solutions in $\bbbf(\v v)$. [**Proof**]{}. Let $f_1,\ldots,f_d$ be a basis of algebraic solutions. Choose a point $\v q\in\bbbq^N$ such that $f_i$ are all analytic near the point $\v q$. Then $f_1,\ldots,f_d$ can be considered as power series expansions in $\v v-\v q$. According to Eisenstein’s theorem for powerseries of algebraic functions we have that the coefficients of the $f_i$ can be reduced modulo $p$ for almost all $p$. Moreover, let $\partial^{\v b},\v b\in B$ be a monomial basis of $K[\partial_i]/({\cal L})$. Then the Wronskian determinant $W_B(f_1,\ldots,f_d)$ is non-zero. So for almost all $p$ the powerseries $f_i$ can be reduced modulo $p$ and moreover, $W_B(f_1,\ldots,f_d) \not\is0\mod{p}$. Hence, for almost all $p$ the powerseries $f_i\mod{p}$ are linearly independent over the quotient field of $\bbbf[[(\v v-\v q)^p]]$, the power series in $(\v v-\v q)^p$. Fix one such prime $p$. Let $P$ be the set $\{(b_1,\ldots,b_N)\in\bbbz^N\ |\ 0\le b_i<p\ {\rm for} \ i=1,\ldots,N\}$. Every solution $f$ can be written in the form $$f\is\sum_{\v b\in P}a_{\v b}(\v v-\v q)^{\v b}\mod{p},$$ where $a_{\v b}\in \bbbf[[(\v v-\v q)^p]]$. For every $L\in{\cal L}$ we have that $$\sum_{\v b\in P}a_{\v b}L(\v v-\v q)^{\v b}\is0\pmod{p}.$$ Let $Q$ be the quotient field of $\bbbf[[(\v v-\v q)^p]]$. The $Q$-linear relations between the polynomials $L(\v v-\v q)^{\v b}$ for every $L$ form a vector space of dimension $d$ since the space of solutions mod $p$ has this dimension. Moreover the space of $Q$-linear relations between the polynomials $L(\v v-\v q)^{\v b}$ is generated by $\bbbf((\v v-\v q)^p)$-linear relations or, what amounts to the same, $\bbbfa(\v v^p)$-linear relations. Pochhammer cycles {#pochhammer} ================= In the construction of Euler integrals one often uses so-called twisted homology cycles. In [@4] this is done on an abstract level, in [@19] it is done more explicitly. In this paper we prefer to follow a more concrete approach by constructing a closed cycle of integration such that the (multivalued) integrand can be chosen in a continuous manner and the resulting integral is non-zero. For the ordinary Euler-Gauss function this is realised by integration over the so-called Pochhammer contour. Here we construct its $n$-dimensional generalisation. In Section \[eulerintegral\] we use it to define an Euler integral for A-hypergeometric functions. Consider the hyperplane $H$ given by $t_0+t_1+\cdots+t_n=1$ in $\bbbc^{n+1}$ and the affine subspaces $H_i$ given by $t_i=0$ for $(i=0,1,2,\ldots,n)$. Let $H^o$ be the complement in $H$ of all $H_i$. We construct an $n$-dimensional real cycle $P_n$ in $H^o$ which is a generalisation of the ordinary 1-dimensional Pochhammer cycle (the case $n=1$). When $n>1$ it has the property that its homotopy class in $H^o$ is non-trivial, but that its fundamental group is trivial. One can find a sketchy discussion of such cycles in [@16 Section 3.5]. Let $\epsilon$ be a positive but sufficiently small real number. We start with a polytope $F$ in $\bbbr^{n+1}$ given by the inequalities $$|x_{i_1}|+|x_{i_2}|+\cdots+|x_{i_k}|\le 1-(n+1-k)\epsilon$$ for all $k=1,\ldots,n+1$ and all $0\le i_1<i_2<\cdots<i_k\le n$. Geometrically this is an $n+1$-dimensional octahedron with the faces of codimension $\ge2$ sheared off. For example in the case $n=2$ it looks like {height="6cm"} The faces of $F$ can be enumerated by vectors $\mu=(\mu_0,\mu_1,\ldots,\mu_n)\in\{0,\pm1\}^{n+1}$, not all $\mu_i$ equal to $0$, as follows. Denote $|\mu|=\sum_{i=0}^n|\mu_i|$. The face corresponding to $\mu$ is defined by $$\begin{aligned} F_{\mu}&:& \mu_0 x_0+\mu_1 x_1+\cdots+\mu_n x_n=1-(n+1-|\mu|)\epsilon, \quad \mu_j x_j\ge\epsilon \ {\rm whenever}\ \mu_j\ne0\\ &&|x_j|\le \epsilon\ {\rm whenever}\ \mu_j=0.\end{aligned}$$ Notice that as a polytope $F_{\mu}$ is isomorphic to $\Delta_{|\mu|-1}\times I^{n+1-|\mu|}$ where $\Delta_r$ is the standard $r$-dimensional simplex and $I$ the unit real interval. Notice in particular that we have $3^n-1$ faces. The $n-1$-dimensional side-cells of $F_{\mu}$ are easily described. Choose an index $j$ with $0\le j\le n$. If $\mu_j\ne0$ we set $\mu_j x_j=\epsilon$, if $\mu_j=0$ we set either $x_j=\epsilon$ or $x_j=-\epsilon$. As a corollary we see that two faces $F_{\mu}$ and $F_{\mu'}$ meet in an $n-1$-cell if and only if there exists an index $j$ such that $|\mu_j|\ne|\mu'_j|$ and $\mu_i=\mu'_i$ for all $i\ne j$. The vertices of $F$ are the points with one coordinate equal to $\pm(1-n\epsilon)$ and all other coordinates $\pm\epsilon$. We now define a continuous and piecewise smooth map $P:\cup_{\mu}F_{\mu}\to H$ as follows. Suppose the point $(x_0,x_1,\ldots,x_n)$ is in $F_{\mu}$. Then its image under $P$ is defined as $$\label{intermediate} {1\over y_0+y_1+\cdots+y_n}(y_0,y_1,\ldots,y_n)$$ where $y_j=\mu_j x_j$ if $\mu_j\ne0$ and $y_j=E_{\epsilon}(x_j)$ if $\mu_j=0$. Here $E_{\epsilon}(x)=\epsilon e^{\pi i (1-x/\epsilon)}$. When $\epsilon$ is sufficiently small we easily check that $P$ is injective. We define our $n$-dimensional Pochhammer cycle $P_n$ to be its image. \[generalbeta\] Let $\beta_0,\beta_1,\ldots,\beta_n$ be complex numbers. Consider the integral $$B(\beta_0,\beta_1,\ldots,\beta_n)= \int_{P_n}\omega(\beta_0,\ldots,\beta_n)$$ where $$\omega(\beta_0,\ldots,\beta_n)=t_0^{\beta_0-1}t_1^{\beta_1-1}\cdots t_n^{\beta_n-1} \ dt_1\wedge dt_2 \wedge\cdots\wedge dt_n.$$ Then, for a suitable choice of the multivalued integrand, we have $$B(\beta_0,\ldots,\beta_n)={1\over \Gamma(\beta_0+\beta_1+\cdots+\beta_n)} \prod_{j=0}^n(1-e^{-2\pi i\beta_j})\Gamma(\beta_j).$$ [**Proof**]{} The problem with $\omega$ is its multivaluedness. This is precisely the reason for constructing the Pochhammer cycle $P_n$. Now that we have our cycle we solve the problem by making a choice for the pulled back differential form $P^*\omega$ and integrating it over $\partial F$. Furthermore, the integral will not depend on the choice of $\epsilon$. Therefore we let $\epsilon\to0$. In doing so we assume that the real parts of all $\beta_i$ are positive. The Proposition then follows by analytic continuation of the $\beta_j$. On the face $F_{\mu}$ we define $T:F_{\mu}\to\bbbc$ by $$T:(x_0,x_1,\ldots,x_n)=\prod_{\mu_j\ne0} |x_j|^{\beta_j-1}e^{\pi i(\mu_j-1)\beta_j}\prod_{\mu_k=0} \epsilon^{\beta_j-1} e^{\pi i(x_j/\epsilon-1)(\beta_j-1)}.$$ This gives us a continuous function on $\partial F$. For real positive $\lambda$ we define the complex power $\lambda^z$ by $\exp(z\log\lambda)$. With the notations as in (\[intermediate\]) we have $t_i=y_i/(y_0+\cdots+y_n)$ and, as a result, $$dt_1\wedge dt_2 \wedge\cdots\wedge dt_n =\sum_{j=0}^n (-1)^j y_jdy_0\wedge\cdots\wedge \check{dy_j}\wedge\cdots dy_n$$ where $\check{dy_j}$ denotes suppression of $dy_j$. It is straightforward to see that integration of $T(x_0,\ldots,x_n)$ over $F_{\mu}$ with $|\mu|<n+1$ gives us an integral of order $O(\epsilon^{\beta})$ where $\beta$ is the minimum of the real parts of all $\beta_j$. Hence they tend to $0$ as $\epsilon\to0$. It remains to consider the cases $|\mu|=n+1$. Notice that $T$ restricted to such an $F_{\mu}$ has the form $$T(x_0,\ldots,x_n)= \prod_{j=0}^n e^{\pi i(\mu_j-1)\beta_j}|x_j|^{\beta_j-1}.$$ Furthermore, restricted to $F_{\mu}$ we have $$\sum_{j=0}^n (-1)^j y_jdy_0\wedge\cdots\wedge \check{dy_j}\wedge\cdots dy_n =dy_1\wedge dy_2 \wedge\cdots\wedge dy_n$$ and $y_0+y_1+\cdots+y_n=1$. Our integral over $F_{\mu}$ now reads $$\prod_{j=0}^n\mu_j e^{\pi i(\mu_j-1)\beta_j}\int_{\Delta}(1-y_1-\ldots-y_n)^{\beta_0-1}y_1^{\beta_1-1} \cdots y_n^{\beta_n-1}dy_1\wedge \cdots\wedge dy_n$$ where $\Delta$ is the domain given by the inequalities $y_i\ge\epsilon$ for $i=1,2,\ldots,n$ and $y_1+\cdots+y_n\le 1-\epsilon$. The extra factor $\prod_j\mu_j$ accounts for the orientation of the integration domains. The latter integral is a generalisation of the Euler beta-function integral. Its value is $\Gamma(\beta_0)\cdots\Gamma(\beta_n)/\Gamma(\beta_0+\cdots+\beta_n)$. Adding these evaluation over all $F_{\mu}$ gives us our assertion. For the next section we notice that if $\beta_0=0$ the subfactor $(1-e^{-2\pi i\beta_0}) \Gamma(\beta_0)$ becomes $2\pi i$. An Euler integral for A-hypergeometric functions {#eulerintegral} ================================================ We now adopt the usual notation from A-hypergeometric functions. Define $$I(A,\alpha,v_1,\ldots,v_N)=\int_{\Gamma}{\v t^{\alpha}\over 1-\sum_{i=1}^Nv_i\v t^{\v a_i}} \ {dt_1\over t_1}\wedge{dt_2\over t_2}\wedge\cdots\wedge{dt_r\over t_r},$$ where $\Gamma$ is an $r$-cycle which doesn’t intersect the hyperplane $1-\sum_{i=1}^Nv_i\v t^{\v a_i}=0$ for an open subset of $\v v\in\bbbc^N$ and such that the multivalued integrand can be defined on $\Gamma$ continuously and such that the integral is not identically zero. We shall specify $\Gamma$ in the course of this section. First note that an integral such as this satisfies the A-hypergeometric equations easily. The substitution $t_i\to \lambda_i t_i$ shows that $$I(A,\alpha,\lambda^{\v a_1}v_1,\ldots,\lambda^{\v a_n}v_N)=\lambda^{\alpha}I(A,\alpha, v_1,\ldots,v_N).$$ This accounts for the homogeneity equations. For the “box”-equations, write $\v l\in L$ as $\v u-\v w$ where $\v u,\v w\in\bbbz_{\ge0}^N$ have disjoint supports. Then $$\Box_{\v l}I(A,\alpha,\v v)=|\v u|!\int_{\Gamma}{\v t^{\alpha+\v \sum_i u_i\v a_i}- t^{\alpha+\v \sum_i w_i\v a_i}\over (1-\sum_{i=1}^Nv_i\v t^{\v a_i})^{|\v u|+1}} \ {dt_1\over t_1}\wedge{dt_2\over t_2}\wedge\cdots\wedge{dt_r\over t_r}$$ where $|\v u|$ is the sum of the coordinates of $\v u$, which is equal to $|\v w|$ since $|\v u|-|\v w|=|\v l|=\sum_{i=1}^Nl_ih(\v a_i)=h(\sum_i l_i\v a_i)=0$. Notice that the numerator in the last integrand vanishes because $\sum_iu_i\v a_i =\sum_iw_i\v a_i$. So $\Box_{\v l}I(A,\alpha,\v v)$ vanishes. We now specify our cycle of integration $\Gamma$. Choose $r$ vectors in $A$ such that their determinent is $1$. After permutation of indices and change of coordinates if necessary we can assume that $\v a_i=\v e_i$ for $i=1,\ldots,r$ (the standard basis of $\bbbr^r$). Our integral now acquires the form $$\int_{\Gamma}{\v t^{\alpha}\over 1-v_1t_1-\cdots-v_rt_r-\sum_{i=r+1}^Nv_i\v t^{\v a_i}} \ {dt_1\over t_1}\wedge{dt_2\over t_2}\wedge\cdots\wedge{dt_r\over t_r}.$$ Perform the change of variables $t_i\to t_i/v_i$ for $i=1,\ldots,r$. Up to a factor $v_1^{\alpha_1} \cdots v_r^{\alpha_r}$ we get the integral $$\int_{\Gamma}{\v t^{\alpha}\over 1-t_1-\cdots-t_r-\sum_{i=r+1}^Nu_i\v t^{\v a_i}} \ {dt_1\over t_1}\wedge{dt_2\over t_2}\wedge\cdots\wedge{dt_r\over t_r},$$ where the $u_i$ are Laurent monomials in $v_1,\ldots,v_N$. Without loss of generality we might as well assume that $v_1=\ldots=v_r=1$ so that we get the integral $$\int_{\Gamma}{\v t^{\alpha}\over 1-t_1-\cdots-t_r-\sum_{i=r+1}^Nv_i\v t^{\v a_i}} \ {dt_1\over t_1}\wedge{dt_2\over t_2}\wedge\cdots\wedge{dt_r\over t_r}.$$ For the $r$-cycle $\Gamma$ we choose the projection of the Pochhammer $r$-cycle on $t_0+t_1+\cdots+t_r=1$ to $t_1,\ldots,t_r$ space. Denote it by $\Gamma_r$. By keeping the $v_i$ sufficiently small the hypersurface $1-t_1-\cdots-t_r-\sum_{i=r+1}^Nv_i\v t^{\v a_i}=0$ does not intersect $\Gamma_r$. To show that we get a non-zero integral we set $\v v=\v 0$ and use the evaluation in Proposition \[generalbeta\]. We see that it is non-zero if all $\alpha_i$ have non-integral values. When one of the $\alpha_i$ is integral we need to proceed with more care. We develop the integrand in a geometric series and integrate it over $\Gamma_r$. We have $$\begin{aligned} &&{\v t^{\alpha}\over 1-t_1-\cdots-t_r-\sum_{i=r+1}^Nv_i\v t^{\v a_i}}\\ &&=\sum_{m_{r+1},\ldots,m_N\ge0}{|m|\choose m_{r+1},\ldots,m_N} {\v t^{\alpha+m_{r+1}\v a_{r+1}+\cdots+m_N\v a_N}\over (1-t_1-\cdots-t_r)^{|m|+1}} \ v_{r+1}^{m_{r+1}}\cdots v_N^{m_N}\end{aligned}$$ where $|m|=m_{r+1}+\cdots+m_N$. We now integrate over $\Gamma_r$ term by term. For this we use Proposition \[generalbeta\]. We infer that all terms are zero if and only if there exists $i$ such that the $i$-th coordinate of $\alpha$ is integral and positive and the $i$-th coordinate of each of $\v a_{r+1},\ldots,\v a_N$ is non-negative. In particular this means that the cone $C(A)$ is contained in the halfspace $x_i\ge0$. Moreover, the points $\v a_j=\v e_j$ with $j\ne i$ and $1\le j\le r$ are contained in the subspace $x_i=0$, so they span (part of) a face of $C(A)$. The set $\alpha+\bbbz^r$ has non-trivial intersection with this face because $\alpha_i\in\bbbz$. From Theorem \[irreducible\] it follows that our system is reducible, contradicting our assumption of irreducibility. So in all cases we have that the Euler integral is non-trivial. By irreducibility of the A-hypergeometric system all solutions of the hypergeometric system can be given by linear combinations of period integrals of the type $I(A,\alpha,\v v)$ (but with different integration cycles). [99]{} A.Adolphson, Hypergeometric functions and rings generated by monomials. Duke Math. J. [**73**]{} (1994), 269-290. P. Beazley Cohen, J.Wolfart, Algebraic Appell-Lauricella functions. Analysis [**12**]{} (1992), 359-376. F. Beukers, G.Heckman, Monodromy for the hypergeometric function $\sb nF\sb {n-1}$. Invent. Math. [**95**]{} (1989), 325-354 B.Dwork, [*Generalized hypergeometric functions*]{}. Oxford Mathematical Monographs, Oxford University Press, New York, 1990. B.Dwork, F.Loeser, Hypergeometric series. Japan. J. Math. (N.S.) [**19**]{} (1993), 81-129. I.M.Gelfand, M.I.Graev, A.V.Zelevinsky, Holonomic systems of equations and series of hypergeometric type, Doklady Akad. Nauk SSSR [**295**]{} (1987), 14-19 (in Russian). I.M.Gelfand, A.V.Zelevinsky, M.M.Kapranov, Equations of hypergeometric type and Newton polytopes, Doklady Akad. Nauk SSSR [**300**]{} (1988), 529-534 (in Russian) I.M.Gelfand, A.V.Zelevinsky, M.M.Kapranov, Hypergeometric functions and toric varieties, Funktsional. Anal. i Prilozhen. [**23**]{} (1989), 12-26 (in Russian) I.M.Gelfand, M.M.Kapranov, A.V.Zelevinsky, Generalized Euler integrals and A-hypergeometric functions, Adv. in Math [**84**]{} (1990), 255-271. M.Kato, Appell’s $F\sb 4$ with finite irreducible monodromy group. Kyushu J. Math. [**51**]{} (1997), 125-147. M.Kato, Appell’s hypergeometric systems $F\sb 2$ with finite irreducible monodromy groups. Kyushu J. Math. [**54**]{} (2000), 279-305. N.M.Katz, Algebraic solutions of differential equations ($p$-curvature and the Hodge filtration), Invent. Math. [**18**]{} (1972), 1-118 N.M.Katz, A conjecture in the arithmetic theory of differential equations, Bull. de la SMF 110 (1982), 203-239, corrections on p347-348. M.Kita, On hypergeometric functions in several variables. I. New integral representations of Euler type. Japan. J. Math. (N.S.) [**18**]{} (1992), 25-74. L.F.Matusevich, E.Miller, U.Walther, Homological methods for hypergeometric families, J. Amer. Math. Soc. [**18**]{} (2005), 919-941. T.Sasaki, On the finiteness of the monodromy group of the system of hypergeometric differential equations $(F\sb{D})$. J. Fac. Sci. Univ. Tokyo Sect. IA Math. [**24**]{} (1977) 565-573. H. A. Schwarz, Über diejenigen Fälle, in welchen die Gaussische hypergeometrische reihe eine algebraische Funktion ihres vierten Elements darstellt, J. Reine Angew. Math. [**75**]{} (1873), 292–335. M.Yoshida, [*Hypergeometric Functions, My Love*]{}, Aspects of Mathematics 32, Vieweg 1997 Department of Mathematics\ Universiteit Utrecht\ P.O. Box 80010, NL-3508 TA\ Utrecht, The Netherlands\ email: f.beukers@uu.nl

ArXiv

--- abstract: 'A rational knot or link can be put into a standard alternating format which has horizontal and vertical twist sites (double helices). The number and type of these twist sites are determined by terms of next-to-highest $z$-degree in Kauffman’s regular isotopy invariant $\Lambda (a,z)$. In particular, for a knot or link with $c$ crossings, the coefficient of the $z^{c-2}$ term is equal to the number of twist sites in its standard diagram. Furthermore, the coefficients of the $a^{-2}z^{c-2}$ and $a^2z^{c-2}$ terms count the number of left-turning and right-turning twist sites, respectively.' author: - 'Mark E. Kidwell' - 'Kerry M. Luse' title: Predicting the number and type of twist sites in a rational knot or link --- Introduction ============ In attempting to extend some results of Abe, [@Abe], about Kauffman’s clock moves, [@FKT], it became necessary to distinguish rational knots and links from more complicated types. Using Conway’s continued fraction rule [@Conway], it is possible to put a rational tangle into a standard format with a certain number of twist sites (double helices) that alternate between horizontal and vertical. (Twist sites with a single crossing are somewhat troublesome, but can be classified as horizontal or vertical by their relative position in the tangle.) As is well known to pipe threaders, there are two types of helices, those with positive torsion and those with negative torsion. See Figure \[fig:twist type\]. This fact, which also applies to double helices, is independent of any orientation assigned to the strands. A given twist site can contribute positive or negative writhe to its tangle depending on how the strands are oriented. In an alternating diagram of a rational tangle, the twist sites must alternate between double helices of positive and negative torsion. [0.35]{} $$\xygraph{ !{0;/r2.0pc/:} !{\hunder} !{\hcrossneg} !{\hcrossneg} !{\hunder-} }$$ [0.35]{} $$\xygraph{ !{0;/r2.0pc/:} !{\hover} !{\hcross} !{\hcross} !{\hover-} }$$ We show in this paper that certain terms in Kauffman’s two-variable, regular isotopy invariant $\Lambda(a,z)$ predict the number and type of twist sites in a rational knot or link in standard alternating form. It has been known since the 1980’s [@Thi] that, for a reduced, alternating diagram with $c$ crossings, the terms of $\Lambda$ of highest$z$-degree have $z$-exponent $c-1$ and depend only on the underlying Conway basic polyhedron of the diagram. In particular, for an algebraic knot or link (and rationals are a subclass), these terms are $a^{-1}z^{c-1}+az^{c-1}$. Most of our attention will be focused on the next-highest $z$-terms $a^{-2}z^{c-2}$, $z^{c-2}$, and $a^2z^{c-2}$. We shall call the coefficients of these three terms $u_-$, $u_0$, and $u_+$. According to Thistlethwaite, no term in $\Lambda$ can have the sum of its $z$-exponent and the absolute value of its $a$-exponent greater than $c$ [@Thi]. Thistlethwaite also proves that $u_-$, $u_0$, and $u_+$ are non-negative. We will prove that the coefficient $u_0$ of $z^{c-2}$ for a rational knot or link equals the number of twist sites in its standard diagram. Before we can discuss $u_-$ and $u_+$, we must mention that the Hopf link is ambiguous as to whether its twist sites are left-turning or right-turning (see Figure \[fig: right and left hopf\] ). In this case, $c=2$, there is one twist site, and the only term in $\Lambda$ with $z$-exponent zero is 1. That is, $u_-=u_+=0$. In all other cases involving alternating knots, $u_-+u_+=u_0$, as proved in [@Thi]. We will show that $u_+$ and $u_-$ are equal to the number of right-turning and left-turning twist sites for a rational knot or link with three or more crossings in standard format. [0.25]{} $$\xygraph{ !{0;/r2.0pc/:} !{\vunder} !{\vunder-} [uur]!{\hcap[2]} [l]!{\hcap[-2]} }$$ [0.25]{} $$\xygraph{ !{0;/r2.0pc/:} !{\vover} !{\vover-} [uur]!{\hcap[2]} [l]!{\hcap[-2]} }$$ Standard format for a rational tangle ===================================== We will draw our rational tangles in a herringbone pattern starting at the northwest (NW) quadrant of the tangle and ending at the southeast (SE) quadrant (Figure \[fig:standard knot\]). In addition: 1. The first and last twist sites have at least two crossings.\[tangle cond 1\] 2. The twist sites alternate between horizontal and vertical. This rule defines whether a twist site with a single crossing is horizontal or vertical. \[tangle cond 2\] 3. All horizontal twist sites are left-turning and all vertical twist sites are right-turning. \[tangle cond 3\] 4. The last (SE) twist site is horizontal. \[tangle cond 4\] ![A standard diagram of a rational knot.[]{data-label="fig:standard knot"}](standardknot){width=".65\textwidth"} These rules have a number of simple consequences. Rule \[tangle cond 3\] ensures that our tangle diagrams are alternating. At every crossing, the overcrossing segment has positive slope. In keeping with Rule \[tangle cond 4\], we will always use the “numerator closure,” $N(T)$, (NW strand joined to NE, SW strand joined to SE) when making our tangles into rational knots or links. By rules \[tangle cond 2\] and \[tangle cond 3\], the number of left-turning twist sites will never differ from the number of right-turning twist sites by more than one. By \[tangle cond 4\], left-turning twist sites will never be in the minority. We will write the Conway notation [@Conway] for a rational tangle with three or more twist sites as $pq_1q_2\ldots q_kr$, where $p\geq2$, $r\geq2$, and $q_i\geq1$ for $i=1,\ldots,k$. There are $k+2$ twist sites in such a tangle. Rational tangles with one or two twist sites are exceptional, and will be treated separately. Properties of Kauffman’s $\Lambda$ polynomial ============================================= We follow the conventions of Kauffman’s defining paper [@RegIs] and of [@Thi], namely: 1. A simple closed curve in the plane has $\Lambda$ invariant 1. \[simple\] 2. (The four term skein relation) \[skein\] $$\Lambda_{D_+}+\Lambda_{D_-}=z(\Lambda_{D_0}+\Lambda_{D_\infty})$$ See Figure \[fig:skein relation\]. 3. (The loop relation)\[loop\] - $\Lambda\bigl(\hspace{.5cm} \xy (0,0)*{\xygraph{ *\xycircle<10pt>{{.}}}}; (-1,2)*{\xygraph{ !{0;/r1.0pc/:} !{\hcross} !{\hcap} }} \endxy \hspace{1.25cm} \bigr) =a\Lambda\bigl(\hspace{.5cm} \xy (0,0)*{\xygraph{ *\xycircle<10pt>{{.}}}}; (-1,2)*{\xygraph{ !{0;/r1.0pc/:} !{\huntwist} !{\hcap} }} \endxy \hspace{1.25cm} \bigr)$ - $\Lambda\bigl(\hspace{.5cm} \xy (0,0)*{\xygraph{ *\xycircle<10pt>{{.}}}}; (-1,2)*{\xygraph{ !{0;/r1.0pc/:} !{\hcrossneg} !{\hcap} }} \endxy \hspace{1.25cm} \bigr)=a^{-1}\Lambda\bigl(\hspace{.5cm} \xy (0,0)*{\xygraph{ *\xycircle<10pt>{{.}}}}; (-1,2)*{\xygraph{ !{0;/r1.0pc/:} !{\huntwist} !{\hcap} }} \endxy \hspace{1.25cm} \bigr)$ [0.15]{} $\xygraph{ !{0;/r2.0pc/:} !{\xunderh} } $ [0.15]{} $ \xygraph{ !{0;/r2.0pc/:} !{\xoverh} }$ [0.15]{} $ \xygraph{ !{0;/r2.0pc/:} !{\xunoverv} }$ [0.15]{} $ \xygraph{ !{0;/r2.0pc/:} !{\xunoverh} }$ Rule \[loop\] shows that $\Lambda$ is changed by type I Reidemeister moves, so it can only be a regular isotopy (preserved by Reidemeister II and III) invariant. The symmetry of $D_+$ and $D_-$ , and of $D_0$ and $D_{\infty}$, makes it difficult to distinguish them. By our conventions for standard format, all crossings are of the form $D_+$. If $D_+$ is an alternating diagram, then $D_-$ will have a bridge that overcrosses at three consecutive crossings. According to [@Thi] and [@Kid], the largest exponent of $z$ in $\Lambda_{D_-}$ is at most $c-3$, so this polynomial will not contribute to the $z^{c-2}$ or higher terms of $D_+$. In a slight deviation from [@Thi], we define $\tilde{\Lambda}(a,z)=u_-a^{-2}z^{c-2}+u_0z^{c-2}+u_+a^2z^{c-2}+a^{-1}z^{c-1}+az^{c-1}$. This truncated invariant satisfies the skein relation $$\tilde{\Lambda}_{D_+}=z(\tilde{\Lambda}_{D_0}+\tilde{\Lambda}_{D_{\infty}})\label{eq: truncated skein}$$ It is possible and necessary to distinguish the two types of smoothings at a crossing along a twist site with at least two crossings. We say that a smoothing is ***[axial]{}*** if it reduces a twist site of $n$ crossings to a twist site of $n-1$ crossings. We say that a smoothing is ***[cross-sectional]{}*** if it cuts across the twist site and creates a loop. See Figure \[smoothings\]. It then follows that in a horizontal twist site, $D_0$ is a cross-sectional smoothing and $D_{\infty}$ is an axial smoothing, while for a vertical twist site, $D_0$ is an axial smoothing and $D_{\infty}$ is a cross-sectional smoothing. This last observation can be used to classify smoothings of one-crossing twist sites as axial or cross-sectional. However, we will avoid smoothing a one-crossing twist site as much as possible. A large part of our argument will involve specifying when a given rational knot or link has the same $\tilde{\Lambda}$ invariant (up to a power of $z$) as an axial smoothing at some crossing and when the cross-sectional smoothing makes a contribution to $\tilde{\Lambda}$. $$\xy (-5,30)*{ \xygraph{ !{0;/r2.0pc/:} !{\hover} !{\hcross} !{\hcross} !{\hover-} }}="x"; (12,20)*{\xygraph{}}="a"; (28,20)*{\xygraph{}}="b"; (-35,0)*{ \xygraph{ !{0;/r2.0pc/:} !{\hover} !{\hcross} !{\hcross} !{\hunover-} }}="y"; (5,5)*{\xygraph{}}="c"; (33,5)*{\xygraph{}}="d"; (20,0)* {\xygraph{ !{0;/r2.0pc/:} !{\hover} !{\hcross} !{\hcross} !{\huntwist} }}="z"; {\ar@2{>} "a";"c"}; {\ar@2{>} "b";"d"}; \endxy$$ Main results ============ We will say that a rational tangle $pq_1q_2\ldots q_kr$ is ***[minimal]{}*** if $p=2$, $r=2$, and $q_i=1$ for $i=1,2,\ldots k$. Otherwise, we say that the tangle (and a particular twist site) has ***[extra crossings]{}***. \[lem: T to T’\] If a rational tangle $T$ in standard format with $c$ crossings has extra crossings, then there is a tangle $T'$ with $c-1$ crossings, still in standard format with the same number of twist sites as $T$, such that $T'$ is obtained from $T$ by an axial smoothing. Simply perform the axial smoothing at a twist site with extra crossings. We point out, in anticipation of future research, that Lemma \[lem: T to T’\] applies not only to rational knots and links, but to any knot or link that is decomposed into rational tangles. Rational tangles with one or two twist sites -------------------------------------------- We begin with the $\Lambda$-invariant of the Hopf link, as given in [@knotinfo]: $$\begin{array}{lccc} \Lambda(a,z)=&-az^{-1}& & +az \\ &&+1&\\ &-a^{-1}z^{-1}&&+a^{-1}z \end{array}$$ and thus $$\begin{array}{lcc} \tilde{\Lambda}(a,z)=&&az\\ &+1&\\ &&+a^{-1}z. \end{array}$$ The middle term of $z^{c-2}=z^0$ is 1, and, in standard format, the link has one twist site. As noted earlier, the handedness of this twist site is ambiguous. We move on to a left-turning trefoil. The axial smoothing of one of its crossings is the Hopf link, and the cross-sectional smoothing is an unknot with two crossings that give it a $\Lambda$-invariant of $a^2$ (see Figure \[fig:trefoil smoothing\]).By , we have $\tilde{\Lambda}_{\textrm{tref}}(a,z)=z(1+a^{-1}z+az+a^2)$, or $$\begin{array}{lcc} \tilde{\Lambda}_{\textrm{tref}}(a,z)=&a^2z&\\ &&+az^2\\ &+z&\\ &&+a^{-1}z^2 \end{array}$$ ![The axial and cross-sectional smoothings of a trefoil knot.[]{data-label="fig:trefoil smoothing"}](trefoilsmoothing){width=".75\textwidth"} Let $K$ be a knot or link with a single left-turning twist site and $c\geq4$ crossings. Then $\tilde{\Lambda}$ is identical, up to powers of $z$, to $\tilde{\Lambda}$ for the left-turning trefoil. In this case, $\tilde{\Lambda}$ will have terms of degree $c-1$ and $c-2\geq 2$. The cross-sectional smoothing is again an unknot with $\Lambda=a^{c-1}$. Even multiplied by $z$ in , this term makes no contributions to the $\tilde{\Lambda}$-invariant of the $c$-crossing knot. We can thus smooth our way back to the trefoil with no change in $\tilde{\Lambda}$ except added powers of $z$. That is: when we have a single left-turning twist site with at least three crossings, $u_+=1$ (the number of left turning twist sites), $u_0=1$ (the total number of twist sites), and $u_-=0$ (the number of right-turning twist sites). We move on to rational tangles of the form $pr$ (two twist sites). The simplest such tangle satisfying our rules is 22, which closes to the figure-eight knot. According to [@knotinfo]: $$\begin{array}{lcc} \tilde{\Lambda}_{\textrm{fig 8}}(a,z)=&a^2z^2&\\ &&+az^3\\ &+2z^2&\\ &&+a^{-1}z^3\\ &a^{-2}z^2 \end{array}$$ The following lemma shows that this pattern holds for the numerator closure of any tangle $pr$. Let $K$ be a knot or link corresponding to the numerator closure of the rational tangle $pr$ with $p\geq2$ and $r\geq2$. Then $\tilde{\Lambda}_K$ is identical up to powers of $z$ to $\tilde{\Lambda}_{\textrm{fig 8}}$. The case where $p=2$ and $r=2$ is just the case of the figure eight knot. The remaining cases have either $p>2$ or $r>2$, or both. The invariant $\tilde{\Lambda}$ of such a knot or link will have non-zero $z$-exponents at the powers $p+r-1$ and $p+r-2$. If we perform the cross-sectional smoothing of a $p$-crossing with $p\geq3$, we get a diagram of $r$ with extra crossings (see Figure \[fig:N43\] for example). The highest $z$-power in this smoothed knot or link will be $z^{r-1}$, but when we multiply by $z$ as in we get $z^r$. Since $3\leq p$, $r+3\leq p+r$, $r+1\leq p+r-2$ and so $r<p+r-2$ and the cross-sectional smoothing makes no contribution to the $\tilde{\Lambda}$-invariant of $N(pr)$. A similar analysis holds if we perform a smoothing at an $r$-site with $r\geq3$. ![The cross-sectional and axial smoothings of the knot $N(43)$.[]{data-label="fig:N43"}](N43){width=".75\textwidth"} In all cases, $u_+=1$ (the number of left-turning twist sites), $u_0=2$ (the total number of twist sites), and $u_-=1$ (the number of right turning twist sites). Rational tangles with three or more twist sites ----------------------------------------------- \[lem: T’ connect sum\] Let $T=pq_1q_2\ldots q_kr$ be a tangle in standard format with at least three twist sites. Let $T'$ be a tangle obtained by performing a cross-sectional smoothing on one crossing in an interior twist site of $T$. Then $N(T')$ is the connected sum of two knots or links. See Figure \[fig:general smoothing\]. Suppose the cross-sectional smoothing is performed on the $q_i$ twist site. There is a simple closed curve in the plane that intersects $T$ in four points and separates $pq_1\ldots q_i$ from $q_{i+1}\ldots q_kr$. (If $i=k$, the $r$-twist site will stand alone.) After the cross-sectional smoothing, this simple closed curve can be made to intersect $T$ in two points, defining the connected sum. ![The cross-sectional smoothing of a rational knot in standard form.[]{data-label="fig:general smoothing"}](generalsmoothing){width=".75\textwidth"} Recall that the $\Lambda$-invariant of a connected sum $K' \# K''$ is the product of the $\Lambda$-invariants of its factors. Suppose $K'$ has $c_1$ crossings and $K''$ has $c_2$ crossings, both reduced and alternating. For $\tilde{\Lambda}$ we have: $$\left( \begin{array}{cc} u'_+a^2z^{c_1-2}&\\ &az^{c_1-1}\\ +u'_0z^{c_1-2}&\\ &+a^{-1}z^{c_1-1}\\ +u'_-a^{-2}z^{c_1-2}& \end{array} \right) * \left( \begin{array}{cc} u''_+a^2z^{c_2-2}&\\ &az^{c_2-1}\\ +u''_0z^{c_2-2}&\\ &+a^{-1}z^{c_2-1}\\ +u''_-a^{-2}z^{c_2-2}& \end{array} \right) =$$ $$\cdots \begin{array}{cc} +(u'_++u_+^{''})a^3z^{c_1+c_2-3}&\\ &a^2z^{c_1+c_2-2}\\ +(u'_++u'_0+u''_++u''_0)az^{c_1+c_2-3}&\\ &+2z^{c_1+c_2-2}\\ +(u'_0+u'_-+u''_0+u''_-)a^{-1}z^{c_1+c_2-3}&\\ &a^{-2}z^{c_1+c_2-2}\\ +(u'_-+u''_-)a^{-3}z^{c_1+c_2-3}&\\ \end{array}$$ In particular, if the connected sum has $c$ crossings, then the highest $z$-degree of its $\tilde{\Lambda}$-invariant is $c-2$. Let the tangle $T$ be in the standard format $pq_1\ldots q_kr$. We first want to show that if $T$ has $j$ extra crossings, then $\tilde{\Lambda}_T(a,z)=z^j\tilde{\Lambda}_m(a,z)$ where $m$ is the minimal tangle $21^k2$. The cases $p\geq3$ and $r\geq3$ proceed as in the case of tangles with two twist sites. Now suppose $q_i\geq2$. We wish to show that $\tilde{\Lambda}_T(a,z)=z\tilde{\Lambda}_{T'}(a,z)$ where $T'=pq_1\ldots (q_i-1)\ldots q_kr$, which is equivalent to saying that the cross-sectional smoothing $T''$ at a $q_i$ twist site makes no contribution to $\tilde{\Lambda}_T$ if $q_i\geq2$. If $T$ has $c$ crossings, then $T''$ loses one crossing at the smoothing and at least one other ($q_i\geq2$) when any loops along the $q_i$ twist site are un-looped. We also know that $T''$ is a connected sum, so that $\tilde{\Lambda}_{T''}$ has terms of $z$-degree at most $c-4$. But the terms of $\tilde{\Lambda}_T$ have $z$-degree $c-2$ and $c-1$ so that, even multiplied by $z$ in , $\tilde{\Lambda}_{T''}$ makes no contribution to $\tilde{\Lambda}_T$. We iterate this argument until we are down to the minimal tangle $21^k2$. A rational knot or link, $N(T)$, in standard format with $k+2$ twist sites and $c$ crossings has: $$\begin{array}{lcc} \tilde{\Lambda}(a,z)=&\left\lceil\frac{k+2}{2}\right\rceil a^2z^{c-2}&\\ & & +az^{c-1}\\ &+(k+2)z^{c-2}\\ & & +a^{-1}z^{c-1}\\ &+\left\lfloor\frac{k+2}{2}\right\rfloor a^{-2}z^{c-2}& \end{array}$$ The critical case is $N(21^k2)$. We will smooth the last singleton vertex, which counts as a vertical twist site. Figure \[fig:21112\] illustrates that we lose two twist sites and one crossing when we perform the axial smoothing. Call the resulting tangle $T'$. As a result of this type of smoothing, 212 becomes 4, 2112 becomes 23, and for $k\geq 3$, $21^k2$ becomes $21^{k-2}3$. $T'$ has $2+(k-2)+3=k+3$ crossings. In all cases, the cross-sectional smoothings are connected sums with one component a Hopf link. Call this tangle $T''$. It has no loops, an alternating diagram, and $(k-1)+2+2=k+3$ crossings. As explained above, the terms of highest $z$-degree in $\tilde{\Lambda}_{T''}$ are $(a^2+2+a^{-2})z^{k+1}$ (connected sum powers are reduced by 1). ![The axial smoothing of $N(21112)$ results in the knot $N(213)$. The cross-sectional smoothing of $N(21112)$ results in the knot $N(22)\#N(2)$.[]{data-label="fig:21112"}](21112example.png){width=".75\textwidth"} Meanwhile, we can assume by induction that: $$\begin{array}{lcc} \tilde{\Lambda}_{T'}(a,z)=&\left\lceil\frac{k}{2}\right\rceil a^2z^{k+1}&\\ & & az^{k+2}\\ &+kz^{k+1}\\ & & +a^{-1}z^{k+2}\\ &\left\lfloor\frac{k}{2}\right\rfloor a^{-2}z^{k+1}& \end{array}$$ Thus the highest $z$-powers of $\tilde{\Lambda}_{T''}$ are on the same level as the next-to-highest $z$-powers of $\tilde{\Lambda}_{T'}$. By , we have for $\Lambda_T$ that $u_+=\left\lceil\frac{k}{2}\right\rceil+1$, $u_0=k+2$, and $u_-=\left\lfloor\frac{k}{2}\right\rfloor+1$. If $k$ is even, say $k=2l$, then $\left\lceil\frac{k}{2}\right\rceil+1=l+1=\frac{k}{2}+\frac{2}{2}=\frac{k+2}{2}=\left\lceil\frac{k+2}{2}\right\rceil$. If $k$ is odd, say $k=2l+1$, then $\left\lceil\frac{k}{2}\right\rceil+1=\left\lceil\frac{2l+1}{2}\right\rceil+1=\left\lceil l+\frac{1}{2}\right\rceil+1=l+1+1=\frac{k-1}{2}+2=\frac{k+3}{2}=\left\lceil\frac{k+3}{2}\right\rceil=\left\lceil\frac{k+2}{2}\right\rceil$. Similarly, $u_-=\left\lfloor\frac{k}{2}\right\rfloor+1=\left\lfloor\frac{k+2}{2}\right\rfloor$. The formula given in the theorem is verified for a minimal tangle. Any tangle with extra crossings can be reduced to this case. We call a knot or link ***[top-heavy]{}*** if $u_+\geq u_-$, ***[bottom-heavy]{}*** if $u_-\geq u_+$, and ***[balanced]{}*** if $u_+=u_-$. It is a consequence of the choices we made when defining standard format that we get the top-heavy form of $\tilde{\Lambda}$ for knots with an odd number of twist sites. The mirror image knot will have $\tilde{\Lambda}$ bottom-heavy. In a rational knot or link with an even number of twist sites, $\tilde{\Lambda}$ will be balanced. (When taking a mirror image, $a$ is replaced by $a^{-1}$ in $\Lambda$.) We could not have written this paper without the data provided by Livingston and Cha’s excellent programs Knot Info and Link Info [@knotinfo],[@linkinfo]. We are uncertain, however, how one form of the knot diagram was chosen as “the” form and the other was labeled as “mirror.” The situation for links is even more complicated. In every case that we have examined, the Kauffman polynomial given corresponds to “the” knot diagram, but there is a slight preference for the bottom-heavy form. Let us give a simple geometric consequence of our calculations: Let $T$ be a rational tangle in standard format with an odd number of twist sites. Then $N(T)$ cannot be an amphicheiral knot or link. Since $u_+=u_-+1$, the invariant $\tilde{\Lambda}$ does not have the necessary symmetry under $a\leftrightarrow a^{-1}$. More complicated cases ====================== We hope to extend our work to algebraic knots and links that are not rational. Our searches in Knot Info and Link Info have revealed at least two differences in the $\tilde{\Lambda}$ invariants: 1. The coefficient $u_0$ can exceed the number of twist sites by 1 (for example, $8_5$ and $8_{10}$ in Rolfsen notation) or 2 (for example, $10_{79}-10_{81}$). We have not observed an excess of 3 or more, but cannot rule it out at this point. 2. In rational knots and links in standard format, we have $u_+=u_-$ or $u_+=u_-+1$. This near-balance does not necessarily hold in non-rational knots or links. For example, the non-rational link with smallest crossing number $6^3_1=L6a5=[2;2;2]$ (see Figure \[fig:6\^3\_1\]), has $u_+=3$, $u_0=4$, and $u_-=1$. The coefficient $u_0$ exceeds the number of twist sites by 1, and $u_-$ exceeds the number of right-turning twist sites by 1, while $u_+$ accurately predicts the number of left-turning twist sites. ![The three component link with Rolfsen notation $6^3_1$.[]{data-label="fig:6^3_1"}](6_3_1){width=".5\textwidth"} [100]{} Y. Abe, *The Clock Number of a Knot*, arXiv:1103.0072v1 \[math.GT\] 1 Mar 2011. J.C. Cha and C. Livingston, *Knot Info: Table of Knot Invariants*, http://www.indiana.edu/ knotinfo, August 28, 2014. J.C. Cha and C. Livingston, *Link Info: Table of Knot Invariants*, http://www.indiana.edu/ linkinfo, August 28, 2014. J.H. Conway, *An enumeration of knots and links, and some of their algebraic properties*, Computational problems in abstract algebra (ed. Leech), Pergamon Press, 1969, 329-358. L.H. Kauffman *Formal Knot Theory*, Dover Publications, 2006. L.H. Kauffman *An invariant of regular isotopy*, Transactions of the American Mathematical Society, Vol 318, 2, (1990), 417-471. M.E. Kidwell, *On the degree of the Brandt-Lickorish-Millet-Ho polynomial of a link*, Proceedings of the American Mathematical Society, 100 (4), 1987: 755-762. D. Rolfsen, *Knots and Links*, AMS Chelsea Publishing, Providence, RI, 2003. M.B. Thistlethwaite, *Kauffman’s polynomial and alternating links*, Topology, 27 (3), 1988: 311-318.

ArXiv

--- abstract: 'In this paper we study the cosmic acceleration for five dynamical dark energy models whose equation of state varies with redshift. The cosmological parameters of these models are constrained by performing a MCMC analysis using mainly gas mass fraction, ${f_{\mathrm{gas}\,}}$, measurements in two samples of galaxy clusters: one reported by Allen et al. (2004), which consists of $42$ points spanning the redshift range $0.05<z<1.1$, and the other by Hasselfield et al. (2013) from the Atacama Cosmology Telescope survey, which consists of $91$ data points in the redshift range $0.118 < \mathrm{z} < 1.36$. In addition, we perform a joint analysis with the measurements of the Hubble parameter $H(z)$, baryon acoustic oscillations and the cosmic microwave background radiation from WMAP and Planck measurements to estimate the equation of state parameters. We obtained that both ${f_{\mathrm{gas}\,}}$ samples provide consistent constraints on the cosmological parameters. We found that the ${f_{\mathrm{gas}\,}}$ data is consistent at the $2\sigma$ confidence level with a cosmic slowing down of the acceleration at late times for most of the parameterizations. The constraints of the joint analysis using WMAP and Planck measurements show that this trend disappears. We have confirmed that the ${f_{\mathrm{gas}\,}}$ probe provides competitive constraints on the dark energy parameters when a $w(z)$ is assumed.' author: - | Juan Magaña$^{1}$[^1], V. Motta$^{1}$[^2], Víctor H. Cárdenas$^{1}$[^3], and G. Foëx$^{1,2}$[^4]\ $^{1}$Instituto de Física y Astronomía, Facultad de Ciencias, Universidad de Valparaíso, Avda. Gran Bretaña 1111, Valparaíso, Chile.\ $^{2}$Max Planck Institute for Extraterrestrial Physics, Giessenbachstrasse, 85748 Garching, Germany title: 'Testing cosmic acceleration for $w(z)$ parameterizations using ${f_{\mathrm{gas}\,}}$ measurements in galaxy clusters' --- dark energy, cosmological parameters, galaxy clusters Introduction ============ The standard cosmological paradigm states that the Universe evolves from a decelerated to an accelerated phase at late times. The evidence of this cosmic acceleration (CA) comes not only from type Ia supernova (SNIa) [@Perlmutter:1999; @Riess:1998], but also from several other cosmological observations [@Mortonson:2014; @Davis:2014; @Li:2013; @Albrecht:2006]. There are two approaches to explain this feature of the Universe: one is to modify the right hand side of Einstein’s equation by assuming the existence of an energy source, dubbed dark energy (DE), that produces these effects, and the other is to modify the left hand side by modifying the gravity theory [@Joyce:2016]. The main property of the DE source that generates the CA is an equation of state (EoS) parameter $w$, which is the ratio between its pressure and energy density. When $w=-1$ we obtain the well-known cosmological constant, $\Lambda$. The $\Lambda$ cold dark matter ($\Lambda$CDM) model satisfactorily explains the CA. However it presents several unsolved problems, i.e., the ∼120 orders of magnitude between the quantum field theory prediction and the cosmological measurements and why the DE density is similar to that of dark matter (DM) today [@Weinberg:89; @cop06]. Additionally, the increase in number and precision of cosmological observations has allowed confronting and explore beyond the standard $\Lambda$CDM [@DiValentino:2015ola; @Gong:2013; @magana:2015; @Marra:2013rba; @PlanckXVI]. One of the methods used to explore departures from the $\Lambda$CDM model is to assume a redshift dependent EoS parameter $w(z)$, given by a suitable parameterization. The most used parameterization for $w(z)$ is the Chevallier-Polarski-Linder (CPL) ansatz [@Chevallier:2000qy; @Linder:2003nc] that expands $w(z)=w_{0}+ w_{1}z/(1+z)$, where $w_{0}$ is the present value of the EoS and $w_{1}$ is its derivative with respect to redshift. An intriguing feature emerges once the CPL function for $w(z)$ is used with SNIa. @shafi2009 found evidence of a low redshift transition of the reconstructed deceleration parameter $q(z)$, the so called slowing down of the CA, that shows a turn around at $z\sim0.3$ [see also @Guimaraes:2011]. We have found that such a feature does not only appear with SNIa but also with gas mass fraction (${f_{\mathrm{gas}\,}}$) data in clusters [@victor_fgas]. Furthermore, another curious behavior emerges when the high redshift data sets are taken into account (as cosmic microwave background radiation, CMB, or baryon acoustic oscillations, BAO): the transition feature at low redshift disappears, making it consistent with the $\Lambda$CDM model [@shafi2009; @Li:2011; @victor_rivera; @victor_fgas]. Then, the key point is to separate the low redshift and high redshift data to study this effect. @magana:2014 investigated this trend in $q(z)$ for five different EoS parameterizations using also four SNIa data sets; the Constitution [@Hicken:2009dk], Union $2$ [@Union2], Union $2.1$ [@Susuki:2012] and Lick Observatory Supernova Search (LOSS) [@Ganeshalingam:2013mia] SNIa samples. They found that, using only SNIa data, the acceleration of the Universe seems to have already peaked and it is evolving towards lower rates of acceleration. This result is independent of the EoS parameterization used and it is observed in the Constitution, Union $2$ and LOSS data sets. However, this behavior disappears when the Union $2.1$ sample is used. As in previous studies, if the cosmological observations at high redshift are included in the analysis, the results change $q(z)$ from a decelerated to an accelerated phase, very similar to the case of the cosmological constant. Recently, @Hu:2015 performed a similar analysis for several $w(z)$ functions using the latest cosmological data from SNIa, CMB and BAO. As in @magana:2014, the authors confirm that the evolution of CA is independent of the EoS parameterization. In addition, they found that the Legacy Survey [SNLS$3$, @Conley:2011] SNIa data favors the slowing down of the CA while the Joint Light-Curve Analysis [JLA, @Betoule:2014] sample prefers an eternal CA. They also shown that the effects of different BAO data on CA are negligible. By revisiting the role of the spatial curvature, @Hu:2015 found that a non-flat Universe prefers a slowing down of the CA. This result is consistent with those obtained by @victor_rivera. An interesting result is that the Planck 2015 data [@Planck2015:XIII; @Planck2015:XIV] favored the slowing down of the CA, while in previous works, this trend disappears when the CMB data is taken into account in the analysis. Here we are interested in testing the transition in the cosmic acceleration as well as the possibility of the slowing down in the acceleration for dynamical DE with a $w(z)$ parameterization using galaxy clusters observations. Several authors have addressed that the gas mass fraction measurements in galaxy clusters can be used to put constraints on cosmological models [@Sasaki:1996; @Allen:2004cd; @Bonamente:2006; @Allen:2007ue; @LaRoque:2006; @Ettori:2009; @Mantz:2014]. Therefore, ${f_{\mathrm{gas}\,}}$ data constitute a complementary geometric method to constrain the parameter space of the same EoS parameterizations studied in @magana:2014. The paper is organized as follows: in the next section we introduce the cosmological framework for a flat Universe case. In section §\[sec:models\] we present the $w(z)$ parameterizations. In section §\[sec:data\] we describe the methodology and data used to constrain the parameters of the models. In section §\[sec:results\] we present and discuss the results. Finally, we present our conclusions in section §\[sec:conclusions\]. Cosmological framework ====================== We consider a flat Friedmann-Lemaître-Robertson-Walker (FLRW) Universe with scale factor $a$ and whose DE component has a dynamical EoS $w(z)$. The dimensionless Hubble parameter $E(z)$ for this Universe is given by $$E^{2}(z)\equiv H^{2}(z)/H^{2}_{0}=\Omega_{m}(1+z)^{3} + \Omega_{r}(1+z)^{4} +\Omega_{de}X(z),\quad \label{eq:Ez}$$ where $H_{0}=H(0)=h\times 100\,\mathrm{km s}^{-1}\mathrm{Mpc}^{-1}$, $\Omega_{m}$ and $\Omega_{r}$ are the density parameter at present day for matter and radiation respectively. We compute $\Omega_{r}=2.469\times10^{-5}h^{-2}(1+0.2271 N_{eff})$, where $N_{eff}=3.04$ is the standard number of relativistic species [@Komatsu:2011]. The density parameter for DE is written as $\Omega_{de}=1-\Omega_{m}-\Omega_{r}$, and the function $X(z)$ reads as $$X(z)\equiv \frac{\rho_{de}(z)}{\rho_{de}(0)}= \mathrm{exp}\left(3\int^{z}_{0}\frac{1+w(z)}{1+z}\mathrm{dz}\right), \label{eq:fz}$$ where $\rho_{de}(z)$ is the energy density of DE at redshift $z$, and $\rho_{de}(0)$ its present value. The comoving distance from the observer to redshift $z$ is given by $$r(z)=\frac{c}{H_0}\int_0^z \frac{dz'}{E(z')}, \label{eq:rz}$$ and it is related to the angular diameter distance as $$D_{A} (z) = \frac{r(z)}{(1+z)}. \label{eq:dA}$$ Since we are interested in testing the CA for dynamical DE models, we study the deceleration parameter *q(z)* defined as $$q(z) = - \frac{\ddot{a}(z)a(z)}{\dot{a}^{2}(z)}, \label{eq:qa}$$ where the dot stands for the derivative with respect to the cosmic time. Using eq. (\[eq:Ez\]), this expression can be rewritten as $$q(z) = \frac{(1+z)}{E(z)} \frac{dE(z)}{dz}-1. \label{eq:qz}$$ parameterizations {#sec:models} ================== [lll]{}\ Parameterization & $w(\infty)$ & $X(z)$\ \ JBP & $w_{0}$ & $(1+z)^{3(1+w_{0})}\mathrm{exp}\left[\frac{3}{2}\frac{w_{1}z^{2}}{(1+z)^{2}}\right]$\ BA & $w_{0}+w_{1}$ & $(1+z)^{3(1+w_0)}(1+z^2)^{\frac{3}{2}w_{1}}$\ FSLL I & $w_{0}$ & $(1+z)^{3(1+w_{0})}\mathrm{exp}\left[\frac{3w_{1}}{2}\mathrm{arctan(z)}\right] \left(1+z^{2}\right)^{\frac{3}{4}w_{1}}\left(1+z\right)^{-\frac{3}{2}w_{1}}$\ FSLL II& $w_{0}+w_{1}$ & $(1+z)^{3(1+w_{0})}\mathrm{exp}\left[- \frac{3w_{1}}{2}\mathrm{arctan(z)}\right] \left(1+z^{2}\right)^{\frac{3}{4}w_{1}}\left(1+z\right)^{\frac{3}{2}w_{1}}$\ SL & $\frac{1}{2} (-1-8 w_{0}+9 w_{0.5})$ & $(1+z)^{\frac{3}{2}(1-8w_{0}+9w_{0.5})} \mathrm{exp}\left[\frac{3z\left\{w_{0}(52z+40)-9w_{0.5}(5z+4)+7z+4\right\}}{8(1+z)^2}\right]$\ One natural extension to the cosmological constant is to explore dynamical DE where the EoS parameter $w(z)$ varies with redshift through an explicit parameterization. Here we consider the following five parameterizations: Jassal-Bagla-Padmanabhan [JBP, @Jassal:2005a; @Jassal:2005b], Barbosa-Alcaniz [BA, @Barboza:2008rh], Feng-Shen-Li-Li [FSLL, @Feng:2012gf], and Sendra-Lazkoz [SL, @Sendra:2011pt] $$\begin{aligned} w(z)&=&w_{0} + w_{1}\frac{z}{\left(1+z\right)^{2}} \quad \mbox{JBP},\\ w(z)&=&w_0 + w_1 \frac{z(1+z)}{1+z^2} \quad \mbox{BA},\\ w(z)&=&w_{0} + w_{1}\frac{z}{1+z^{2}} \qquad \mbox{FSLL I},\\ w(z)&=&w_{0} + w_{1}\frac{z^{2}}{1+z^{2}} \qquad \mbox{FSLL II},\\ w(z)&=&-1 + c_{1}\left(\frac{1+2z}{1+z}\right) + c_{2}\left(\frac{1+2z}{1+z}\right)^{2} \quad \mbox{SL},\end{aligned}$$ where $w_{0}=w(0)$ for all parameterizations, $w_{1}=w'(0)$ ($'$ denotes derivative with respect to redshift), the constant $c_{1}=(16w_{0}-9w_{0.5}+7)/4$, and $c_{2}=-3w_{0}+ (9w_{0.5}-3)/4$, being $w_{0.5}$ the value of the EoS at $z=0.5$. Table \[tab:functions\] shows the amplitude of the EoS when $z\rightarrow \infty$ and the function $X(z)$ given by Eq. (\[eq:fz\]) for each parameterization. Methodology and data {#sec:data} ==================== The EoS parameters for each $w(z)$ function are constrained by performing a Markov Chain Monte Carlo (MCMC) analysis using the following cosmological observations: the ${f_{\mathrm{gas}\,}}$ measurements in galaxy clusters, $H(z)$ estimations, BAO data, and the CMB information. The gas mass fraction in galaxy clusters ---------------------------------------- ### ${f_{\mathrm{gas}\,}}$ modeling data The gas mass fraction is defined as ${f_{\mathrm{gas}\,}}\equiv M_{gas}/M_{tot}$, where $M_{gas}$ is the X-ray gas mass and $M_{tot}$ is the dynamical total mass of the galaxy cluster. The ${f_{\mathrm{gas}\,}}$ is useful as a cosmological probe when a scaling relation between the baryonic gas mass fraction in galaxy clusters and the global fraction of baryonic matter and dark matter of the Universe $\Omega_{b}/\Omega_{0m}$ is assumed. The ${f_{\mathrm{gas}\,}}$ estimations for any cosmological model fitted to the reference $\Lambda$CDM data can be obtained by writing $M_{gas}$ and $M_{tot}$ in terms of $D_{A}(z)$ [@Sasaki:1996; @Nesseris:2006er]: $$\begin{aligned} {f_{\mathrm{gas}\,}}(z) \equiv \frac{b}{1+\alpha}\frac{\Omega_b}{\Omega_{m}}\left(\frac{D_{A}^{\Lambda CDM} (z)}{D_A(z)}\right)^{3/2}, \label{eq:one}\end{aligned}$$ where ${D_{A}}(z)$ is the angular diameter distance for any cosmological model, $D_{A}^{\Lambda CDM}(z)$ is the angular diameter distance for the $\Lambda$CDM reference model, $\Omega_{b}$ is the baryonic density parameter, and $\Omega_{m}$ is the current DM density parameter. Here $b$ is a bias factor which relates the baryonic fraction in clusters with the one in the Universe. The constant $\alpha$ relates the baryonic luminous mass and the baryonic gas mass. This constant is $\alpha\approx 0.19 \sqrt{h}$ [@Allen:2004cd]. Hereafter, we refer to the fitting function given by the Eq. (\[eq:one\]) as A04. From here on, we assume that the gas mass fraction measurements are uncorrelated. To constrain the parameters of the $w(z)$ functions introduced in §$3$, we use the following chi-square in the MCMC analysis $$\begin{aligned} \chi_{{f_{\mathrm{gas}\,}}}^{2} &=& \left(\sum_{i=1}^{N}\frac{[{f_{\mathrm{gas}\,}}(z_i)-{f_{\mathrm{gas}\,}}^{obs}(z_i)]^2}{\sigma_{f_{gas,i}}^2}\right) +\left(\frac{h-h_{,prior}}{\sigma_{h,prior}}\right)^{2}\nonumber\\ &+& \left(\frac{\Omega_b h^{2}-\Omega_b h^{2}_{,prior}}{\sigma_{\Omega_b h^{2},prior}}\right)^{2} +\left(\frac{b-0.824}{0.089}\right)^{2}, \label{eq:chifgasa04}\end{aligned}$$ where ${f_{\mathrm{gas}\,}}(z)$ is calculated using A04, ${f_{\mathrm{gas}\,}}^{obs}$ is the observed gas mass fraction, and $\sigma_{{f_{\mathrm{gas}\,}}}$ is the error in the measurements. $N=42$ ($91$) for the Allen (ACT) data set. Table \[tab:priors\] gives the adopted priors for $h$ and $\Omega_b h^{2}$ from WMAP [@Hinshaw:2013] and Planck measurements [@PlanckXVI]. [|cc|]{} Parameter&Allowance\ \ $h$& $0.73\pm 0.0175$ (Gaussian)\ $\Omega_b h^{2}$ & $0.02202\pm 0.00046$ (Gaussian)\ $\Omega_{m}$ & $[0.2,0.4]$ (Uniform)\ $w_{0}$ & $[-2,0]$ (Uniform)\ $w_{1} (w_{0.5})$ & $[-10,2]$ (Uniform)\ \ $b$&$0.824\pm0.089$ (Gaussian)\ \ $\eta$&$0.214\pm0.022$ (Gaussian)\ $\gamma$&$1.0<\gamma<1.1$ (Uniform)\ $s_{0}/h_{70}^{0.5}$&$0.16\pm0.048$ (Gaussian)\ $s_{1}$&$-0.2<s_{1}<0.2$ (Uniform)\ $b_{0}$&$0.65<b_{0}<1.0$ (Uniform)\ $b_{1}$&$-0.1<b_{1}<0.1$ (Uniform)\ $K$ & $1.0\pm0.1$ (Gaussian)\ @Allen:2007ue modified the Eq. (\[eq:one\]) to include more corrections related to the cluster physics and the cosmological model as: $${f_{\mathrm{gas}\,}}(z) = \frac{ K A \gamma b(z)} {1+s(z) } \left( \frac{\Omega_{\rm b}}{\Omega_{\rm m}} \right) \left[ \frac{D_{\rm A}^{\rm \Lambda CDM}(z)}{D_{\rm A}(z)} \right]^{1.5}. \label{eq:fgasA08}$$ The factor $A$ in Eq. (\[eq:fgasA08\]) is given by $$A= \left( \frac{ \theta_{2500}^{\rm \Lambda CDM}}{\theta_{2500}} \right)^\eta \approx \left( \frac{ H(z) D_{\rm A}(z)} { \left[ H(z) D_{\rm A}(z)\right] ^{\rm \Lambda CDM}} \right)^\eta,$$ which accounts for the change in the angle subtended by $r_{2500}$ [^5] as the underlying cosmology is varied. Here, $\eta$ is the slope of the ${f_{\mathrm{gas}\,}}$ data in the region of $r_{2500}$ as measured for the reference $\Lambda$CDM cosmology. The parameter $\gamma$ takes into account the non-thermal pressure support in the clusters. The parameter $s(z)=s_{0}(1+s_{1}z$) models the baryonic mass fraction in stars as function of redshift. The bias factor also depends on the redshift as $b(z)=b_{0}(1+b_{1}z)$. $K$ parameterizes the residual uncertainty in the accuracy of the instrument calibration and X-ray modelling. Hereafter, we refer to ${f_{\mathrm{gas}\,}}$ function given by the Eq. (\[eq:fgasA08\]) as A08. We also consider the fitting function $A08$ to estimate (Table \[tab:priors\] lists the adopted priors for $A08$) the EoS parameters using $$\begin{aligned} \chi_{{f_{\mathrm{gas}\,}}}^{2} &=& \left(\sum_{i=1}^{N}\frac{[{f_{\mathrm{gas}\,}}(z_i)-{f_{\mathrm{gas}\,}}^{obs}(z_i)]^2}{\sigma_{f_{gas,i}}^2}\right) +\left(\frac{h-h_{,prior}}{\sigma_{h,prior}}\right)^{2}\nonumber\\ &+& \left(\frac{\Omega_b h^{2}-\Omega_b h^{2}_{,prior}}{\sigma_{\Omega_b h^{2},prior}}\right)^{2} +\left(\frac{\eta-0.214}{0.022}\right)^{2}\nonumber\\ &+&\left(\frac{s_{0}/\sqrt{(100/70)h}-0.16}{0.048} \right)^{2} +\left(\frac{K-1.0}{0.1} \right)^{2}.\qquad \label{eq:chifgasa08}\end{aligned}$$ ### ${f_{\mathrm{gas}\,}}$ samples We use two galaxy cluster gas mass fraction samples: the sample by @Allen:2004cd, ${f_{\mathrm{gas}\,}}^{Allen}$, which consists of $42$ points spanning the redshift range $0.05<z<1.1$ and the ACT sample, ${f_{\mathrm{gas}\,}}^{ACT}$, which consists of $91$ data points in the redshift range $0.118 < \mbox{z} < 1.36$. While @Allen:2004cd provide the ${f_{\mathrm{gas}\,}}$ measurements, the ACT sample gives the mass [^6] $M_{500}$ calculated using the one-parameter family of universal pressure profiles for the cluster gas [@Hasselfield:2013]. The relation between $M_{500}$ and ${f_{\mathrm{gas}\,}}$ is [see @Goncalves; @Vikhlinin:2009 for more details]. $${f_{\mathrm{gas}\,}}=(0.0764\pm 0.004)\,h^{-1.5} + (0.037\pm0.006) \log{M_{15}}, \label{eq:fgasVik}$$ where $M_{15}$ is the cluster total mass, $M_{500}$, in units of $10^{15}h^{-1}M_{\odot}$. Notice the computed ${f_{\mathrm{gas}\,}}$ data points could be biased due to the Eq. (\[eq:fgasVik\]), since it depends on a fiducial cosmology. In addition, the constraints also could be deteriorated due to the intrinsic scatter in the scaling relation. In Appendix A we study the impact of considering $15\%$ error ($3\sigma$) in each parameter of the Eq. \[eq:fgasVik\] for the JBP model. We show that the best fits obtained are within the $1\sigma$ confidence level. $H(z)$ measurements ------------------- Our Bayesian analysis also considers the measurements of the Hubble parameter which is directly related to the expansion history of the Universe. We used $34$ points in the redshift range $0.07\le z \le2.3$ compiled by @Sharov:2014 (see also Farooq & Ratra 2013 and Chen et al. 2013). The data set is shown in Table \[tab:Hz\] of the Appendix \[Ap:Hz\]. It is worth to notice that some $H(z)$ points might be correlated to BAO measurements, i.e. BAO points together with the best fit value of the sound horizon from CMB data could be used to estimate the $H(z)$ value. Although some of the $H(z)$ points in the Table 3 were obtained in this manner, through this work we assume there is no correlation between $H(z)$ and BAO. Then, the $\chi^2_{H}$ can be written as $$\chi_{H}^2 = \sum_{i=1}^{34} \frac{ \left[ H(z_{i}) -H_{obs}(z_{i})\right]^2 }{ \sigma_{H_i}^{2} },$$ where $H_{obs}(z_{i})$ is the observational Hubble parameter at $z_{i}$, $\sigma_{H_i}$ its error, and $H(z_{i})$ is the theoretical value for a model. BAOs measurements ----------------- Another independent test to put constraints on cosmological parameters is to use the BAO signature. The BAO measurements considered in our analysis are obtained from the Six-degree-Field Galaxy Survey (6dFGS) BAO data [@Beutler2011:6dF], the WiggleZ experiment [@Blake2011:wigglez], the Sloan Digital Sky Survey (SDSS) Data Release 7 (DR7) BAO distance measurements [@Percival2010:dr7], the Baryon Oscillation Spectroscopic Survey (BOSS) SDSS Data Release 9 BAO distance measurements (SDSS DR9) [@Anderson2012:dr9] and the most recent BAO distance estimations from Data release $11$ (DR11) of the BOSS (quasars) of SDSS. WiggleZ points are obtained using the low redshift data from SDSS and 6dFGS, thus they might be correlated. Nevertheless, in this work we assume that they are not correlated. The 6dFGS BAO estimated the distance ratio $d_{z}=0.336\pm0.015$ at $z=0.106$ [@Beutler2011:6dF], where $$d_{z} = \frac{r_{d}}{D_V(z)}, \label{eq:dz}$$ where $r_{d}=r_{s}(z_{d})$ and the distance scale $D_V$ is defined as $$D_V(z)=\frac{1}{H_0}\left[(1+z)^2D_A(z)^2\frac{cz}{E(z)}\right]^{1/3}.$$ In the standard scenario $D_A(z)$ relates to the luminosity distance through $D_A(z)=D_L(z)/(1+z)^2$. The comoving sound horizon, $r_s(z)$, is defined as $$r_s(z) = c \int_z^\infty \frac{c_s(z')}{H(z')}dz',$$ where the sound speed $c_s(z) = 1/\sqrt{3\left(1+\bar{R_b}/\left(1+z\right)\right)}$, with $\bar{R_b} = 31500\, \Omega_{b}h^2(T_{CMB}/2.7\rm{K})^{-4}$, and $T_{CMB}$ is the CMB temperature. The redshift $z_d$ at the baryon drag epoch is well fitted with the formula proposed by @Eisenstein98, $$z_d =\frac{1291(\Omega_{m}h^2)^{0.251}}{1+0.659\,(\Omega_{m}h^2)^{0.828}}[1+b_1(\Omega_b h^2)^{b_2}],$$ where $$\begin{aligned} b_1 &=& 0.313\left(\Omega_{m}\,h^2\right)^{-0.419}\left[1+0.607\left(\Omega_{m}\,h^2\right)^{0.674}\right], \\ b_2 &=& 0.238\left(\Omega_{m}\,h^2\right)^{0.223}.\end{aligned}$$ It is worth to note that these equations were calculated for the standard cosmology, here we assume that they are valid for dynamical DE models. Therefore, the chi-square for the 6dFGS BAO measurement is given by $$\chi^2_{\mathrm{6dFGS}} = \left(\frac{d_z-0.336}{0.015}\right)^2.$$ From WiggleZ, @Kazin:2014 [see also Gong et al. 2015] estimated the following three points for $d_{z}= 0.0870\pm0.0042$, $0.0672\pm0.0031$, and $0.0593\pm0.0020$ at redshifts $z=0.44,0.6$ and $0.73$ respectively. Thus, the $\chi^2$ for the WiggleZ BAO data is given by $$\begin{aligned} \chi^2_{\mathrm{WiggleZ}} &=& \left(\frac{d_z(0.44)-0.0870}{0.0042}\right)^2\nonumber\\ &+&\left(\frac{d_z(0.6)-0.0672}{0.0031}\right)^2\nonumber\\ &+&\left(\frac{d_z(0.73)-0.0593}{0.0020}\right)^2.\end{aligned}$$ Similarly, using the clustering of galaxies from SDSS DR7, @Ross:2015 [see also Percival et al. 2010] obtained a consensus measurement $D_{\mathrm{V}}=(664\pm25)(r_{d}/r_{d,fid})$ Mpc of the BAO scale at $z=0.15$. By adopting the value of the fiducial cosmology $r_{d,fid}$ and inverting, @Gong:2015 estimate $d_{z}$ and the $\chi^2$ can be expressed as $$\chi^2_{\mathrm{DR7}}=\left(\frac{d_z(0.15)-0.2239}{0.0084}\right)^2,$$ @Anderson2014b [see also Anderson et. al. 2012] measure the BAO signature from the SDSS-III BOSS DR11 and give the consensus estimations for the distance $D_{v}=(1264\pm25\mathrm{Mpc})(r_{d}/r_{d,fid})$ at redshift $z=0.32$ and $D_{v}=(2056\pm20\mathrm{Mpc})(r_{d}/r_{d,fid})$ at redshift $z=0.57$. Using the $r_{d,fid}$ value given by the authors and inverting to obtain $d_{z}$, the $\chi^2$ is written as $$\chi^2_{\mathrm{DR11a}}=\left(\frac{d_z(0.32)-0.1181}{0.0023}\right)^2+\left(\frac{d_z(0.57)-0.0726}{0.0007}\right)^2,$$ The most recent measured position of the BAO peak from SDSS DR11 determines $D_{H}/r_{d}=9.18\pm0.28$ at $z=2.34$ [@Delubac2014], and $D_{H}/r_{d}=9.0\pm0.3$ at $z=2.36$ [@Font-Ribera:2014] where $D_{H}(z)=c/H(z)$. Thus, we compute the $\chi^2$ for these points as $$\chi^2_{\mathrm{DR11b}}=\left(\frac{\frac{D_H(2.34)}{r_{d}}-9.18}{0.28}\right)^2+ \left(\frac{\frac{D_H(2.36)}{r_{d}}-9.00}{0.3}\right)^2.$$ The total $\chi^2$ for all the BAO data points can be written as $$\chi^2_{\mathrm{BAO}} = \chi^2_{\mathrm{6dFGS}}+\chi^2_{\mathrm{ WiggleZ}} + \chi^2_{\mathrm{DR7}} + \chi^2_{\mathrm{DR11a}} + \chi^2_{\mathrm{DR11b}}.$$ CMB --- We also include CMB information by using the following distance posteriors: the acoustic scale, $l_{A}$, the shift parameter, $R$, and the decoupling redshift, $z_{*}$. These quantities can be used to constrain cosmological parameters without considering the full CMB measurements [@Komatsu:2009; @Komatsu:2011; @wang:2006; @Wright:2007]. Although these distance posterior are computed assuming an underlying cosmology, several authors have proved that these quantities are almost independent on the input DE models [@Li:2008; @Mukherjee:2008; @wang:2012; @Planck2015:XIV]. The acoustic scale is defined as $$l_A = \frac{\pi r(z_*)}{r_s(z_*)}, \label{eq:lA}$$ where the redshift of decoupling $z_*$ is given by [@Hu:1996], $$z_* = 1048[1+0.00124(\Omega_b h^2)^{-0.738}] [1+g_1(\Omega_{m}h^2)^{g_2}],$$ and $$\begin{aligned} g_1 &=& \frac{0.0783(\Omega_b h^2)^{-0.238}}{1+39.5(\Omega_b h^2)^{0.763}},\nonumber\\ g_2 &=& \frac{0.560}{1+21.1(\Omega_b h^2)^{1.81}}.\end{aligned}$$ It is worth to note that we assume that these fitting formulae are valid in dynamical DE models. The shift parameter is defined as [@Bond:1997] $$R = \frac{\sqrt{\Omega_{m}H_{0}^2}}{c} r(z_{*}).$$ Thus, the $\chi^2$ for the CMB data is constructed as $$\label{cmbchi} \chi^2_{\mathrm{CMB}} = X^T\,\mathrm{Cov}_{\mathrm{CMB}}^{-1}\,X,$$ where $\mathrm{Cov}_{\mathrm{CMB}}^{-1}$ is the inverse covariance matrix and $$X =\left( \begin{array}{c} l_A^{th} - l_A^{obs} \\ R^{th} - R^{obs}\\ z_*^{th} - z_{*}^{obs} \end{array}\right),$$ the superscripts $th$ and $obs$ refer to the theoretical and observational estimations respectively. @Hinshaw:2013 obtained for a flat $\Lambda$CDM Universe, using the WMAP 9-yr temperature power spectrum, the following distance posteriors $l_{A}^{obs}=302.40$, $R^{obs}=1.7246$, $z_{*}^{obs}=1090.88$, and the inverse covariance matrix $$\mathrm{Cov}^{-1}_{\mathrm{WMAP9}} = \left( \begin{array}{ccc} 3.182 & 18.253 & -1.419\\ 18.253 & 11887.879 & -193.808\\ -1.429 & -193.808 & 4.556 \end{array}\right). \label{eq:Cwmap9}$$ On the other hand, for a flat $w$CDM, @Neveu:2016 estimated from Planck measurements: $l_{A}^{obs}=301.787\pm0.089$, $R^{obs}=1.7492\pm0.0049$, $z_{*}^{obs}=1089.99\pm0.29$. They also provide the following the inverse covariance matrix, $\mathrm{Cov}^{-1}_{Pl}$, of these quantities $$\mathrm{Cov}^{-1}_{\mathrm{Pl}}= \left( \begin{array}{ccc} 162.48 & -1529.4 & 2.0688 \\ -1529.4 & 207232 & -2866.8 \\ 2.0688 & -2866.8 & 53.572 \\ \end{array} \right).$$ Results and discussion {#sec:results} ====================== In all our estimations, we use Gaussian priors on $h$ [@Riess:2016], and $\Omega_{b}h^{2}$ [@Cooke:2014], which are derived from measurements independent from the CMB and useful in dynamical DE models, thus leaving $\Omega_{m}$, $w_{0}$, and $w_{1}$ ($w_{0.5}$ for the SL model) as the only free parameters of the analysis (see Table \[tab:priors\]). To calculate these constraints we assume a Gaussian likelihood $\mathcal{L}\propto \exp(-\chi^{2}/2)$. First, we obtain the constraints using the measurements of ${f_{\mathrm{gas}\,}}$ alone. As mentioned before, we consider the ${f_{\mathrm{gas}\,}}^{Allen}$ or the ${f_{\mathrm{gas}\,}}^{ACT}$ sample. We also estimated the parameters from the combination of $H(z)+$BAO$+$CMB (WMAP or Planck) and the joint analysis of ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB. Thus, with the $\chi^{2}$ functions defined in section §\[sec:data\] for each data, we construct the total $\chi^{2}$-function for these three cases: $\chi^{2}=\chi^{2}_{{f_{\mathrm{gas}\,}}}$, $\chi^{2}=\chi^{2}_{H}+\chi^{2}_{\mathrm{BAO}}+\chi^{2}_{\mathrm{CMB}}$, and $\chi^{2}=\chi^{2}_{{f_{\mathrm{gas}\,}}}+\chi^{2}_{H}+\chi^{2}_{\mathrm{BAO}}+\chi^{2}_{\mathrm{CMB}}$. We perform our Bayesian analysis using the *emcee* Python module [@emcee]. In all our computations, we consider a total of $6500$ steps with $500$ walkers and $1000$ steps were removed which correspond to the burn-in phase to stabilize the estimations. We start the walkers in a small ball around the expected points of maximum probability estimated with a differential evolution method. To judge the convergence of the sampler we ask that the acceptance fraction is in the $[0.2-0.5]$ range and check the autocorrelation time which is found to be $\mathcal{O}(60)$ and $\mathcal{O}(80)$ for the A04 and A08 fitting respectively. Impact of ${f_{\mathrm{gas}\,}}$ fitting function ------------------------------------------------- For all $w(z)$ parameterizations we estimated the cosmological parameters using the chi-squares functions for both A$04$ and A$08$ ${f_{\mathrm{gas}\,}}$ fitting equations. The first one (Eq. \[eq:one\]) assumes that the baryonic fraction from clusters does not depend on redshift, while the second one (Eq. \[eq:fgasA08\]) assumes a linear dependence for $s$ and $b$ parameters with redshift. In addition, the A$08$ equation considers more correction factors due to the underlying cosmological model and the cluster physics. Fig. \[fig:comparison\] shows that there is no significative difference in the $w_{0}$ and $w_{1}$ constraints for the JBP parameterization derived from ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ sample when A$04$ or A$08$ function are used. When the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample is used a similar result is found, and the $q(z)$ parameter trend is the same for both A$04$ or A$08$ functions (see Appendix B for more details). In addition, we found the same result for the other $w(z)$ functions. Thus, in what follows, we only present the results obtained using the A$04$ equation. For completeness, Tables \[tab:jbp\]-C5 (Tables C4-C5 are available online only) give the best fits for $\Omega_{m}$, $w_{0}$ and $w_{1}$, by considering the A$08$ fitting function (Eq. \[eq:fgasA08\]) in the parameter estimation. ![Comparison of the $w_{0}-w_{1}$ confidence contours for the JBP parameterization using the ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ data alone with the A$04$ and A$08$ equations. The blue filled contours show the $68\%$, $95\%$ and $99\%$ confidence levels obtained by considering Eq. (\[eq:one\]), while the orange dashed contours show the same by considering Eq. (\[eq:fgasA08\]). The star and the dot marks represent the best fit obtained using each one of these ${f_{\mathrm{gas}\,}}$ fitting function respectively. Observe that there is no significative statistically difference in the $w_{0}$ and $w_{1}$ bounds. We found the same result for the other $w(z)$ functions[]{data-label="fig:comparison"}](jbp_contours_comparison.pdf){width="8cm" height="8cm"} Impact of ${f_{\mathrm{gas}\,}}$ measurements on the parameter estimation ------------------------------------------------------------------------- Tables and Figures \[tab:jbp\]-C5 (Tables and Figures C4-C5 are available online only) summarize our results for $\Omega_{m}$ and the EoS parameters for each $w(z)$ parameterization estimated with the different tests. Notice that, in general, the estimations on $\Omega_{m}$ derived from the ${f_{\mathrm{gas}\,}}$ data are in very good agreement to those obtained from the combination of $H(z)+$BAO$+$WMAP and of the joint analysis of the all data sets (see priors in Table \[tab:priors\]). Thus we confirm that ${f_{\mathrm{gas}\,}}$ measurements in galaxy clusters is a suitable probe to put bounds on the content of dark matter in the Universe. For the EoS parameters, $w_{0}$ and $w_{1}$, the best fits show different estimations when using ${f_{\mathrm{gas}\,}}$, $H(z)+$BAO$+$WMAP, and the joint analysis of all data sets. However, these bounds are consistent at the $3\sigma$ confidence level (see Fig. \[fig:comparison\_data\] for the comparison of the JBP estimations). On the other hand, the bounds estimated with the joint analysis of the data are consistent, at $68\%$ confidence level, with the cosmological constant, i.e., $w_{0}\approx -1$ and $w_{1}\approx 0$. Since our interest is to focus on the gas mass fraction probe, in what follows, we only present the results obtained using the ${f_{\mathrm{gas}\,}}$ data and the joint analysis ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB. ![Comparison of the $w_{0}-w_{1}$ $68\%$, $95\%$ and $99\%$ confidence levels, for the JBP model using ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ (no-filled dotted blue contours), $H(z)+$BAO$+$WMAP (no-filled green contours), and the joint analysis ${f_{\mathrm{gas}\,}}^{Allen}+H(z)+$BAO$+$WMAP (filled dashed red contours) data. The star, triangle, and dot represent the best fit obtained using each one of these data set respectively. The best fit obtained using $f_{gas}$ is consistent at $3\sigma$ level with those obtained from $H(z)+$BAO$+$WMAP and the joint analysis.[]{data-label="fig:comparison_data"}](comparison_data.pdf){width="8cm" height="8cm"} Cosmic acceleration for the $w(z)$ parameterizations ---------------------------------------------------- Tables \[tab:jbp\]-C5 give the estimated redshift of the transition, $z_{t}$, from a decelerated to an accelerated phase. If the slowing down of the CA emerges, we also give the redshift when it occurs. Both redshifts were calculated using the following Monte Carlo approach. We randomly selected a 1:20 subsample of the posterior sample and calculated for each point the $q(z)$ on a grid in $z$. Then, we obtained the $z_{t}$ and $z_{sd}$ distributions which are well-approximated by a Gaussian funtion. Therefore, by producing a histrogram, we estimated the central values and its errors. We found that the $z_{t}$ value estimated using only the ${f_{\mathrm{gas}\,}}^{Allen}$ data is lower than those obtained with the joint analysis of ${f_{\mathrm{gas}\,}}^{Allen}+H(z)+$BAO$+$CMB data. We reconstructed the cosmological evolution of the $q(z)$ parameter and propagated its error as follows. By considering the subsample, we produced a well-approximated Gaussian distribution of $q(z_{i})$ for each $z_{i}$ on the grid in $z$ and determined the central value, $1$ and $2\sigma$ levels. The behaviour of the $q(z)$ parameter for each parameterization using the ${f_{\mathrm{gas}\,}}^{Allen}$, $H(z)+$BAO$+$WMAP, and the joint analysis of ${f_{\mathrm{gas}\,}}^{Allen}+H(z)+$BAO$+$WMAP data are shown in the left panel of Figure \[fig:qz\]. We found that for all parameterizations (except for FSLLII, which presents a behavior slightly oscillatory) the CA has two transitions when the ${f_{\mathrm{gas}\,}}$ constraints are used. At redshifts $0.67<z_{t}<0.72$, the Universe begins its accelerated expansion, however, it reaches a peak of acceleration at $0.19<z_{sd}<0.24$ and then the acceleration slows down. Moreover, this trend is supported at $2\sigma$ confidence level (dashed lines in Fig. \[fig:qz\]). Our results are in good agreement with those obtained by @Hu:2015. They found that in the JBP, BA and FSLL parameterizations the cosmic acceleration presents a slowing down at $z\sim0.24$, $0.26$, and $0.28$ respectively within the $2\sigma$ confidence level. Nevertheless, we calculated the Akaike information criteria [AIC, @Akaike:1974] and Bayesian information criteria [BIC, @Schwarz:1978] for the $\Lambda$CDM and the five dynamical DE models. By comparing with the $\Lambda$CDM, we obtained that the dynamical models lead larger AICs ($\sim3$) and BICs ($\sim12$) values, indicating that a dynamical DE model is not necessary to explain the ${f_{\mathrm{gas}\,}}$ data. [this result is similar to that obtained by @Hu:2015]. Since that the slowing down of cosmic acceleration depend on the $w_{0}-w_{1}$ values, which are completely degenerated, it is crucial to provide narrow EoS constraints to prevent misleading results. A future increase in ${f_{\mathrm{gas}\,}}$ points (i.e. decrease in the systematic uncertainties) will produce a better parameter estimation for $w(z)$ parameterizations. For instance, @Mantz:2014 simulated the ${f_{\mathrm{gas}\,}}$ measurements for $400$ galaxy clusters and obtained for an evolving DE model an increase in the Figure-of-Merit (FoM) by a factor $\sim15$ in an optimistic scenario of future surveys (this implies a reduction of the area enclosed by the $w_{0}-w_{1}$ $95\%$ confidence contour and the EoS constraints are very close to the cosmological constant). Therefore the slowing down of cosmic acceleration obtained from ${f_{\mathrm{gas}\,}}$ constraints could be only a statistical fluctuation. On the other hand, if the $H(z)+$BAO$+$WMAP constraints are used to reconstruct $q(z)$, we found that it evolves similarly to that of the cosmological constant. When we used the estimations obtained with the joint analysis of the all data, the transition from a decelerated to an accelerated phase occurs at $0.64<z_{t}<0.70$ and $q(z)\sim -0.7$ at $z\rightarrow0$, i.e., the slowing down of the CA disappears, and it evolves as the cosmological constant. Furthermore, we found that the evolution of the CA is independent of the EoS parameterization (see the different panels of Fig. \[fig:qz\]). This same scenario was found by @magana:2014 when several SNIa data sets are considered [see also @Hu:2015; @Shi:2011]. Although to discern which model is the preferred one by observations is not the scope of this paper, it can be determined by comparing the $\chi_{min}$, given in Tables \[tab:jbp\]-C5, among data sets. Any $w(z)$ parametrization could be plausible to model a dynamical dark energy. ------------------------------------------------- --------------------------------------------------- {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} ------------------------------------------------- --------------------------------------------------- ----------------------------------------------------- ------------------------------------------------------- {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} {width="40.00000%"} ----------------------------------------------------- ------------------------------------------------------- Impact of CMB data on $q(z)$ ---------------------------- As shown in the previous section, when the CMB data from WMAP measurements are included, the slowing down of the CA, which emerges when only low-redshift data are used, disappears [see also @shafi2009; @Li:2011; @victor_rivera; @victor_fgas; @magana:2014]. Nevertheless, in a recent paper, @Hu:2015 found that for the CPL parameterization, when the Planck data [from the $2013$ and $2015$ measurements, @PlanckXVI; @Planck2015:XIII; @Planck2015:XIV] in combination with the BAO and SNIa (JLA sample) data are considered in the analysis, the slowing down of the CA is still present at the $1\sigma$ and $2\sigma$ levels. To prove this result, we also consider the Planck $2015$ data in the ${f_{\mathrm{gas}\,}}$ and ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB analysis. We found that there are no significative differences between the best fits on $\Omega_{m}$ obtained using Planck with those from WMAP (see tables \[tab:jbp\]-C5 and Figures \[fig:JBP\]-C5). Additionally, the $w_{0}$ and $w_{1}$ constraints obtained using Planck data in the joint analysis are consistent at $1\sigma$ confidence level to those estimated including WMAP measurements. When $q(z)$ is reconstructed for each parameterization using the constraints derived of the combination of all data, we obtain that the universe evolves from a decelerated phase to one accelerated phase at $0.63<z_{t}<0.68$, and at $z\rightarrow0$ the cosmic acceleration is consistent with that of the standard model (see right panels of Fig. \[fig:qz\]). This result is in agreement with those obtained using WMAP data where the slowing down of CA obtained using the gas mass fraction disappears by adding the CMB (high-redshift) information. Nevertheless, our result is in tension with that found by @Hu:2015 and we check that the Planck 2015 measurements do not favor a slowing down of the CA. Impact of the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample on the constraints ----------------------------------------------------------------------------- In the previous sections, we have presented the results using the ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ in the $w(z)$ fitting. Here we investigate whether the slowing down of the CA appears when the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample, with about twice the data points by ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ set, is used in the parameter estimation. As before, we estimate the $\Omega_{m}$, $w_{0}$, and $w_{1}$ constraints using only the ${f_{\mathrm{gas}\,}}$ data and the joint analysis ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB (WMAP or Planck). We give the best fits for each parameterization in Tables \[tab:jbp\]-C5. Notice that the $\Omega_{m}$ constraints from the different data sets are very similar to those obtained when the ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ sample was used. Although there is a slightly tension in the $w_{0}$ and $w_{1}$ best fits obtained from both gas mass fraction samples, they are consistent within their $68\%$ confidence regions. When the $q(z)$ deceleration parameter for each $w(z)$ is reconstructed using the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ constraints, the transition from a decelerated phase to an accelerated phase occurs at $\simeq0.75$. We also found that the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ is consistent at the $2\sigma$ confidence level with the slowing down of the CA at $0.20<z_{sd}<0.25$ for the JBP, BA, FSLLI, and SL parameterizations (see left panels of Fig. \[fig:qz\_ACT\]). Notice that this trend is softened compared with those obtained using the ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ sample. As in the case of SNIa data, we confirm that this trend occurs for several ${f_{\mathrm{gas}\,}}$ samples. By reconstructing $q(z)$ using the ${f_{\mathrm{gas}\,}}^{ACT}+H(z)+$BAO$+$WMAP (Planck) constraints, the CA evolves as the cosmological constant and the Universe passes from a decelerated phase to an accelerated phase at $0.68<z_{t}<0.72$ ($0.66<z_{t}<0.69$). It is worth to note that the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample could yield biased constraints due to the cluster gas physics and underlying cosmology assumed in the ${f_{\mathrm{gas}\,}}-M_{500}$ relation (Eq. \[eq:fgasVik\]) which was used to compute the data point. To take into account its bias we performed a test using large ${f_{\mathrm{gas}\,}}$ uncertainties (see Appendix A) and we found no significative differences on the $\Omega_{m}$, $w_{0}$, and $w_{1}$ constraints, leading to similar evolutions of $q(z)$ as those obtained with normal errors. Conclusions {#sec:conclusions} =========== We investigated the evolution of the deceleration parameter $q(z)$ in the following models with the EoS as a function of redshift using gas mass fraction measurements in galaxy clusters: JBP, BA, FSLLI, FSLLII, and SL. We used two ${f_{\mathrm{gas}\,}}$ data sets: the sample by @Allen:2004cd which consists of $42$ points spanning the redshift range $0.05<z<1.1$ and the ACT sample which consists of $91$ data points in the redshift range $0.118 < \mathrm{z} < 1.36$ [@Hasselfield:2013]. These ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ data points were computed using a ${f_{\mathrm{gas}\,}}-M_{500}$ relation [@Vikhlinin:2009]. In spite of the difference between the samples, they both provide consistent estimations for the cosmological parameters. We validated that the gas mass fraction test is a qualified tool to constrain the cosmological parameters $\Omega_{m}, w_{0}, w_{1}$, and then the CA in these models. We also considered different fitting function for the ${f_{\mathrm{gas}\,}}$ data points: A$04$ (Eq. \[eq:one\]) and A$08$ (Eq. \[eq:fgasA08\]). The main difference between both functions is that the second one considers several corrections due to the underlying cosmology and the cluster physics. We found that the A$08$ fitting function does not offer significantly better cosmological constraints than the simplest A$04$ function (see Fig.\[fig:comparison\]). A MCMC analysis was performed using the ${f_{\mathrm{gas}\,}}$ data and the joint ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB (WMAP and Planck) measurements. We used these data sets to reconstruct the $q(z)$ parameter. For all parameterizations (except for FSLLII), a slowing down of CA emerges at late times (at the $2\sigma$ confidence level) when using the ${f_{\mathrm{gas}\,}}$ data alone, i.e. different $w(z)$ functions does not influence the $q(z)$ cosmic evolution. This is in agreement with results obtained by @magana:2014 [see also Shi et al. 2011] using several SNIa samples. Our results are also in agreement with those obtained by @Hu:2015, who performed a comprehensive analysis with several cosmological data. Furthermore, our results suggest that this CA trend is not a systematic effect produced by SNIa or ${f_{\mathrm{gas}\,}}$ samples (as suggested previously by @shafi2009 [@victor_fgas; @magana:2014; @Hu:2015]). Nevertheless, the emergence of the slowing down of cosmic acceleration depend on the $w_{0}-w_{1}$ values. Therefore, to prevent misleading results, it is crucial to provide narrow EoS constraints. Here, we used two fgas samples, by Allen and ACT, which have poor statistics. A future increase in ${f_{\mathrm{gas}\,}}$ points (i.e. decrease in the systematic uncertainties) will produce a better parameter estimation for $w(z)$ parameterizations. For instance, @Mantz:2014 simulated the ${f_{\mathrm{gas}\,}}$ measurements for $400$ galaxy clusters and obtained for an evolving DE model an increase in the FoM by a factor $\sim15$ in an optimistic scenario of future surveys (this implies a reduction of the area enclosed by the $w_{0}-w_{1}$ $95\%$ confidence contour and the EoS constraints are very close to the cosmological constant). Therefore the slowing down of cosmic acceleration obtained from ${f_{\mathrm{gas}\,}}$ constraints could be only a statistical fluctuation. On the other hand, the CA behavior changes when the $H(z)+$BAO$+$CMB constraints are used. We obtained for each parameterization that these data favor an accelerated expansion at late times. The $q(z)$ reconstructed from the joint analysis ${f_{\mathrm{gas}\,}}+H(z)+$BAO$+$CMB (WMAP and Planck) shows that the CA behaves as the cosmological constant. Our result is in tension with those obtained by @Hu:2015, who found a CA slowing down for the CPL parameterization using the Planck data. We investigated the impact to consider the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample and we confirmed that, in general, it is consistent at the $2\sigma$ confidence level with the slowing down CA (except for the FSLLII parameterization). As in the ${f_{\mathrm{gas}\,}}^{\mathrm{Allen}}$ case, this trend disappears by adding the $H(z)$, BAO and WMAP measurements in the parameter estimation. Finally, our results suggest that several low-redshift cosmological data are consistent with a CA slowing down when $w(z)$ is parameterized. Although this trend does not emerge in other approaches to study the evolution of $q(z)$ [see for example @Rani:2015; @Zhang2016] we encourage to the community to further investigate this phenomena in DE models. To improve our constraints, a large high precision ${f_{\mathrm{gas}\,}}$ sample is needed. Acknowledgments =============== We thank the anonymous referee for thoughtful remarks and suggestions. J.M. acknowledges support from Gemini $32130024$ and FONDECYT 3160674. V.M. acknowledges support from ECOS-CONICYT C12U02 and Centro de Astrofísica de Valparaíso. V.C. acknowledges support from DIUV 50/2013, [99]{} Akaike, H. 1974, ITAC, 19, 716 Albrecht et al., 2006, Report of the Dark Energy Task Force, arXiv:astro-ph/0609591 Allen S.W., Schmidt R.W., Ebeling H., Fabian A.C. and Speybroeck L. van, 2004, MNRAS 353, 457. Allen S.A., Rapetti D.A., Schmidt R.W., Ebeling H., Morris G. and Fabian A.C., 2008, MNRAS 383, 879. Amanullah R. et al., 2010, ApJ 716, 712 Anderson L., Aubourg E., Bailey S., et al., 2014, MNRAS, 439, 83. Anderson L., Aubourg E., Bailey S., Beutler, F., Bhardwaj, V., et al., 2014, MNRAS, 441, 24. Anderson L., Aubourg E., Bailey S., et al., 2012, MNRAS, 427, 3435. Barboza E. M. Jr. and Alcaniz J. S., 2008, PLB, 666, 415. Beutler F., Blake C., Colless M., et al., 2011, MNRAS, 416, 3017. Betoule, M., Kessler, R., Guy, J., Mosher, J., Hardin, D., et al., 2014, A&A, 568, A22 Blake C., Kazin E. A., Beutler F., et al., 2011, MNRAS, 418, 1707. Blake C., Brough S., Colless M., et al., 2012, MNRAS, 425, 405. Bonamente M., Joy M. K., LaRoque S. J., Carlstrom J. E., Reese E. D., Dawson K. S., 2006, ApJ, 647, 25. Bond J. R., Efstathiou G., Tegmark M., 1997, MNRAS, 291, L33. Busca N. G., Delubac T., Rich J., et al., 2013, A&A 552, 18. Cárdenas V. H., Rivera M., 2012, PLB, 710, 251. Cárdenas V. H., Bernal C., Bonilla A., 2013, MNRAS, 433, 3534 Chen Y., Geng C.-Q., Cao S., H. Y.-M., Zhu Z.-H., 2013, arXiv:1312.1443. Chevallier M., Polarski D., 2001, IJMPD, 10, 213. Chuang C.-H., and Wang Y., 2013, MNRAS, 435, 255. Cooke, R. J., Pettini, M., Jorgenson, R. A., Murphy, M. T., Steidel, C. C. et al. 2014, ApJ, 781, 31 Conley, A., Guy, J., Sullivan, M., et al. 2011, ApJS, 192, 1 Copeland E. J., Sami M., Tsujikawa S., 2006, IJMPD, 15, 1753, D’Agostini, G., 2004, arXiv:physics/0403086. Davis, T. M., 2014, GReGr, 46, 1731 Delubac T., Bautista J. E., Busca N. G., et al., 2015, A&A, 574, A59 Di Valentino E., Melchiorri A. and Silk J., 2015, PRD, 92, 121302. Eisenstein D. J. and Hu W., 1998, ApJ, 496, 605. Eisenstein D. J. et al., 2005, ApJ, 633, 560. Ettori S., Morandi A., Tozzi P., et al., 2009, A&A 501, 61. Farooq O., and Ratra B., 2013, ApJ, 766, L7. Feng C. -J., Shen X. -Y., Li P. and Li X. -Z., 2012, JCAP, 209, 023. Font-Ribera, A., Kirkby, D., Busca, N., Miralda-Escudé, J., Ross, N. P., et al., 2014, JCAP, 05, 027 Foreman-Mackey, D., Hogg, D. W., Lang, D., Goodman, J., 2013, PASP, 125, 306 Ganeshalingam M., Li W. and Filippenko A. V., 2013, MNRAS 433, 2240. Gaztañaga E., Cabr[é]{} A., and Hui L., 2009, MNRAS, 399, 1663. Gonçalves R. S., Bernui A., Holanda R. F. L., Alcaniz J. S., 2015, A&A, 573, A88 Gong, Y., Gao, Q., Zhu, Z.-H., 2013, MNRAS, 430, 3142 Gong, Y., Ma, Y.-Z., Zhang, S.-N., Chen, X., 2015, PRD, 92, 063523. Guimarães, A.C.C., Lima, J.A.S., 2011, CQG, 28, 125026 Hasselfield M., Hilton M., Marriage T. A., et al., 2013, JCAP, 07, 008. Hicken, M., Wood-Vasey, W. M.,Blondin, S., Challis, P., Jha, S., et al., 2009, ApJ, 700,1097 Hinshaw G., et al., 2013, ApJS, 208, 19. Hu W. and Sugiyama N., 1996, ApJ, 471, 542. Jassal H. K., Bagla J. S., and Padmanabhan T., 2005, MNRAS, 356, L11. Jassal H. K., Bagla J. S., and Padmanabhan T., 2005, PRD, 72, 103503. Joyce, A., Lombriser, L., & Schmidt, F. 2016, arXiv:1601.06133 Kazin, E. A., Koda, J., Blake, C., Padmanabhan, N., Brough, S., et al., 2014, MNRAS, 441, 3524 Komatsu, E., et al., 2009, ApJS, 180, 330 Komatsu, E., et al., 2011, ApJS, 192, 18 LaRoque S. J., Bonamente M., Carlstrom J. E., et al., 2006, ApJ 652, 917. Li, H., Xia, J.-Q., Zhao, G.-B., Fan, Z., and Zhang, X., 2008, ApJ, 683, L1 Li Z., Wu P., Yu H., 2011, PLB, 695, 1. Li, M., Li, X.-D., Wang, S., Wang, Y., 2013, Frontiers of Physics, 8, 828 Linder E. V., 2003, PRL, 90, 091301. Magaña J., Cárdenas V. H., Motta V., 2014, JCAP, 10, 017 Magaña J., Motta V., Cárdenas V. H., Verdugo, T., Jullo, E., 2015, ApJ, 813, 69 Mantz A. B., Allen S. W., Morris R. G., et al., 2014, MNRAS, 440, 2077. Marra V., Amendola L., Sawicki I. and Valkenburg W., 2013, PRL, 110, 24, 241305 Moresco M., Cimatti M., Jimenez R., et al., 2012, JCAP, 08, 006. Mortonson, M. J.; Weinberg, D. H., White, M., 2014, Chapter 25 of Particle Data Group 2014 Review of Particle Physics, arXiv:1401.0046. Mukherjee, P., Kunz, M., Parkinson, D., & Wang, Y., 2008, PRD 78, 083529 Nesseris S. and Perivolaropoulos L., 2007, JCAP 01, 018. Neveu, J., Ruhlmann-Kleider, V., Astier, P., Besançon, M., Guy, J., et al., 2016, arXiv:1605.02627 Oka, A., Saito, S., Nishimichi, T., Taruya, A., and Yamamoto, K., 2014, MNRAS, 439, 2515 Percival W. J., et al., 2010, MNRAS, 401, 2148. Perlmutter, S., Aldering, G., Goldhaber, G., Knop, R. A., Nugent, P., et al., 1999, ApJ, 517, 565 Planck collaboration: Ade P. A. R., et al., 2014, A & A, 571, A16 Planck collaboration: Ade P. A. R., et al., 2015, arXiv:1502.01589 Planck collaboration: Ade P. A. R., et al., 2015, arXiv:1502.01590 Rani, N., Jain, D., Mahajan, S., Mukherjee, A., Pires, N., 2015, JCAP, 12, 045 Riess, A. G., Filippenko, A. V.; Challis, P., Clocchiatti, A., Diercks, A., et al., AJ, 116, 1009 Riess, A. G., Macri, L. M., Hoffmann, S. L., Scolnic, D., Casertano, et al., 2016, Accepted in ApJ, arXiv:1604.01424 Ross, A. J.; Samushia, L., Howlett, C., Percival, W. J., Burden, et al., 2015, MNRAS, 449, 835 Sasaki S., 1996, PASJ, 48, L119. Sendra I., and Lazkoz R., 2012, MNRAS, 422, 776. Shafieloo A., Sahni V. and Starobinsky A.A., 2009, PRD 80, 101301 Sharov, G. S. and Vorontsova, E. G., 2014, JCAP, 10, 057 Shi, K., Huang, Y.-F., Lu, T., 2011, RAA, 11, 1403 Stern D., Jimenez R., Verde L., Kamionkowski M., and Stanford S. A., 2010, JCAP, 02, 008. Susuki, N., Rubin, D., Lidman, C., Aldering, G., Amanullah, R. [*et al.*]{}, 2012, ApJ 746, 85. Schwarz, G. 1978, AnSta, 6, 461 Vikhlinin A., 2009, ApJ, 692, 1033. Wang Y., Mukherjee P., 2006, ApJ, 650, 1 Wang, Y., Chuang, C.-H-, Mukherjee, P., 2012, PRD, 85, 023517 Wang, S., Hu, Y., Li, M., Li, N., 2016, ApJ, 821, 60. Weinberg S., 1989, RMP, 61, 1. Wetterich C., 1988, Nucl. Phys. B, 302, 668. Wright, E. L., 2007, ApJ, 664, 633 Zhang C., Zhang H., Yuan S., Zhang T.-J. and Sun Y.-C., 2014, RAA, 14, 1221 Zhang M.-J. and Xia J.-Q., 2016, JCAP, 12, 005 Impact of the uncertainties on the ${f_{\mathrm{gas}\,}}-M_{500}$ relation ========================================================================== The cosmological constraints in the models could be biased due to the uncertainties on the scaling relation parameters in the Eq. \[eq:fgasVik\]. To test the impact of these uncertainties, we constrain, for instance, the JBP model using $15\%$ error in each parameter of the equation. Figure \[fig:sampling\] shows that the best fit using $15\%$ errors is within the $1\sigma$ confidence level of the previous best fit (see Figure C1). Figure \[fig:sampling\] also shows that the $q(z)$ parameter exhibits the slowing down feature at $1\sigma$ level as presented before (Figure 3). The scatter in the $f_{gas}-M_{500}$ scaling relation has little or no significant impact in the cosmological parameter estimation up to $1\sigma$ confidence level. ------------------------------------------------ -------------------------------------------------------- [{width="3in"}]{} [{width="3in"}]{} [{width="3in"}]{} [{width="3in"}]{} ------------------------------------------------ -------------------------------------------------------- Impact of the A08 fitting function when the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ is used ========================================================================================= To test whether the constraints derived from the ${f_{\mathrm{gas}\,}}^{\mathrm{ACT}}$ sample are improved using the A$08$ fitting function instead of the A$04$ we constrain the JBP model using both fitting functions. The left panel of the Fig. \[fig:comparison2\] shows that there is no significative difference in the $w_{0}$ and $w_{1}$ constraints when A$04$ or A$08$ functions are used. In addition, the right panel of the Fig. \[fig:comparison2\] shows that the $q(z)$ parameter behaviour is essentially the same when the best fits derived from both functions are used. We found the same result for the other $w(z)$ functions. -------------------------------------------------------- -------------------------------------------------- {width="3in"} {width="3in"} -------------------------------------------------------- -------------------------------------------------- Tables and confidence contours for each model using and $+H(z)+$BAO$+$CMB {#Ap:tables_contours} ========================================================================= [|lcccccc|]{}\ Data set& $\chi^{2}_{min}$&$\Omega_{m}$& $w_{0}$&$w_{1}$&$z_{t}$&$z_{sd}$\ \ H(z)+BAO+W9 & $49.32$ & $0.28^{+0.01}_{-0.01}$ & $-1.25^{+0.22}_{-0.16}$ & $ 0.51^{+0.99}_{-1.47}$&$0.66^{+0.04}_{-0.04}$& No\ H(z)+BAO+Pl & $67.22$ & $0.29^{+0.01}_{-0.01}$ & $-1.22^{+0.21}_{-0.16}$ & $0.55^{+0.91}_{-1.18}$&$0.65^{+0.04}_{-0.04}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.80$ & $0.28^{+0.03}_{-0.03}$& $-0.80^{+0.45}_{-0.45}$ & $-3.78^{+3.73}_{-3.76}$&$0.71^{+0.09}_{-0.09}$& $0.20^{+0.06}_{-0.09}\,(2.6)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $160.71$ & $0.31^{+0.02}_{-0.03}$& $-0.59^{+0.34}_{-0.39}$ & $-3.74^{+2.50}_{-2.19}$&$0.75^{+0.08}_{-0.07}$& $0.23^{+0.05}_{-0.09}\, (3.2)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.40$ & $0.28^{+0.01}_{-0.01}$ & $-1.26^{+0.21}_{-0.15}$ & $0.53^{+0.97}_{-1.41}$&$0.66^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.29$ & $0.29^{+0.01}_{-0.01}$ & $-1.24^{+0.19}_{-0.15}$ & $0.65^{+0.85}_{-1.08}$&$0.64^{+0.04}_{-0.03}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.64$ & $0.29^{+0.01}_{-0.01}$ & $-1.12^{+0.23}_{-0.20}$ & $-0.31^{+1.30}_{-1.53}$&$0.68^{+0.03}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.30$ & $0.29^{+0.01}_{-0.01}$ & $-1.15^{+0.20}_{-0.18}$ & $0.05^{+1.05}_{-1.17}$&$0.67^{+0.04}_{-0.04}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.93$ & $0.29^{+0.03}_{-0.04}$& $-0.81^{+0.47}_{-0.48}$ & $-3.33^{+3.44}_{-3.69}$&$0.71^{+0.13}_{-0.11}$& $0.20^{+0.07}_{-0.10}\,(2.3)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $174.41$ & $0.31^{+0.02}_{-0.03}$& $-0.59^{+0.36}_{-0.40}$ & $-3.23^{+2.93}_{-2.76}$&$0.75^{+0.10}_{-0.10}$& $0.24^{+0.05}_{-0.10}\,(3.2)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.32$ & $0.28^{+0.01}_{-0.01}$ & $-1.24^{+0.21}_{-0.16}$ & $0.44^{+1.03}_{-1.42}$&$0.66^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.01$ & $0.29^{+0.01}_{-0.01}$ & $-1.22^{+0.19}_{-0.16}$ & $0.54^{+0.90}_{-1.12}$&$0.65^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.44$ & $0.29^{+0.01}_{-0.01}$ & $-1.12^{+0.23}_{-0.19}$ & $-0.11^{+1.25}_{-1.58}$&$0.67^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.08$ & $0.29^{+0.01}_{-0.01}$ & $-1.11^{+0.20}_{-0.18}$ & $0.00^{+1.04}_{-1.17}$&$0.66^{+0.04}_{-0.04}$& No\ [|lcccccc|]{}\ Data set& $\chi^{2}_{min}$&$\Omega_{m}$& $w_{0}$&$w_{1}$&$z_{t}$&$z_{sd}$\ \ H(z)+BAO+W9 & $49.85$ & $0.28^{+0.01}_{-0.01}$ & $-1.16^{+0.14}_{-0.14}$ & $-0.04^{+0.30}_{-0.37}$&$0.68^{+0.05}_{-0.05}$& No\ H(z)+BAO+Pl & $67.46$ & $0.29^{+0.01}_{-0.01}$ & $-1.12^{+0.13}_{-0.13}$ & $0.007^{+0.22}_{-0.24}$&$0.67^{+0.05}_{-0.04}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.71$ & $0.28^{+0.03}_{-0.03}$& $-0.80^{+0.48}_{-0.50}$ & $-2.18^{+2.32}_{-2.54}$&$0.68^{+0.11}_{-0.12}$& $0.19^{+0.06}_{-0.09}\,(2.5)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $161.52$ & $0.31^{+0.02}_{-0.03}$& $-0.85^{+0.25}_{-0.25}$ & $-1.07^{+0.79}_{-0.83}$&$0.75^{+0.08}_{-0.09}$& $0.20^{-0.06}_{-0.10}\,(2.5)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.92$ & $0.28^{+0.01}_{-0.01}$ & $-1.17^{+0.13}_{-0.12}$ & $-0.03^{+0.28}_{-0.35}$&$0.68^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.78$ & $0.29^{+0.01}_{-0.01}$ & $-1.15^{+0.12}_{-0.11}$ & $0.04^{+0.20}_{-0.22}$&$0.66^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.49$ & $0.29^{+0.01}_{-0.01}$ & $-1.10^{+0.12}_{-0.11}$ & $-0.20^{+0.30}_{-0.37}$&$0.70^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.30$ & $0.29^{+0.008}_{-0.008}$ & $-1.13^{+0.10}_{-0.10}$ & $-0.02^{+0.20}_{-0.22}$&$0.68^{+0.04}_{-0.04}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.85$ & $0.29^{+0.03}_{-0.04}$& $-0.86^{+0.48}_{-0.49}$ & $-1.72^{+1.98}_{-2.31}$&$0.67^{+0.14}_{-0.16}$& $0.18^{+0.07}_{-0.09}\,(2.2)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $174.33$ & $0.31^{+0.02}_{-0.03}$& $-0.83^{+0.29}_{-0.26}$ & $-0.92^{+1.01}_{-1.28}$&$0.73^{+0.10}_{-0.15}$& $0.19^{+0.06}_{-0.12}\,(2.1)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.75$ & $0.28^{+0.01}_{-0.01}$ & $-1.16^{+0.13}_{-0.13}$ & $-0.06^{+0.29}_{-0.36}$&$0.68^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.39$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.12}_{-0.12}$ & $0.01^{+0.21}_{-0.23}$&$0.67^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.35$ & $0.29^{+0.01}_{-0.01}$ & $-1.07^{+0.12}_{-0.12}$ & $-0.20^{+0.29}_{-0.36}$&$0.70^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $229.97$ & $0.30^{+0.01}_{-0.01}$ & $-1.06^{+0.11}_{-0.11}$ & $0.08^{+0.20}_{-0.22}$&$0.68^{+0.05}_{-0.05}$& No\ [|lcccccc|]{}\ Data set& $\chi^{2}_{min}$&$\Omega_{m}$& $w_{0}$&$w_{1}$&$z_{t}$&$z_{sd}$\ \ H(z)+BAO+W9 & $49.24$ & $0.28^{+0.01}_{-0.01}$ & $-1.27^{+0.20}_{-0.19}$ & $0.40^{+0.68}_{-0.78}$&$0.64^{+0.06}_{-0.06}$& No\ H(z)+BAO+Pl & $67.01$ & $0.29^{+0.01}_{-0.01}$ & $-1.22^{+0.18}_{-0.17}$ & $0.32^{+0.54}_{-0.57}$&$0.63^{+0.05}_{-0.05}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.75$ & $0.28^{+0.03}_{-0.03}$& $-0.79^{+0.48}_{-0.49}$ & $-2.77^{+2.90}_{-3.07}$&$0.69^{+0.10}_{-0.10}$& $0.21^{+0.07}_{-0.09}\,(2.6)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $161.59$ & $0.31^{+0.02}_{-0.03}$& $-0.74^{+0.30}_{-0.31}$ & $-1.83^{+1.32}_{-1.30}$&$0.75^{+0.08}_{-0.08}$& $0.24^{+0.08}_{-0.11}\,(2.5)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.35$ & $0.28^{+0.01}_{-0.01}$ & $-1.27^{+0.18}_{-0.17}$ & $0.36^{+0.64}_{-0.73}$&$0.64^{+0.06}_{-0.06}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.22$ & $0.28^{+0.01}_{-0.01}$ & $-1.23^{+0.16}_{-0.15}$ & $0.36^{+0.50}_{-0.52}$&$0.63^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.60$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.18}_{-0.17}$ & $-0.15^{+0.67}_{-0.74}$&$0.68^{+0.04}_{-0.06}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.23$ & $0.29^{+0.009}_{-0.009}$ & $-1.16^{+0.15}_{-0.14}$ & $0.07^{+0.49}_{-0.52}$&$0.66^{+0.05}_{-0.05}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.86$ & $0.29^{+0.03}_{-0.04}$& $-0.83^{+0.48}_{-0.50}$ & $-2.29^{+2.56}_{-2.87}$&$0.69^{+0.14}_{-0.13}$& $0.23^{+0.12}_{-0.10}\,(2.0)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $174.33$ & $0.31^{+0.02}_{-0.03}$& $-0.83^{+0.29}_{-0.26}$ & $-0.92^{+1.01}_{-1.28}$&$0.73^{+0.10}_{-0.13}$& $0.26^{+0.08}_{-0.11}\,(2.7)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.33$ & $0.28^{+0.01}_{-0.01}$ & $-1.25^{+0.18}_{-0.17}$ & $0.31^{+0.66}_{-0.74}$&$0.64^{+0.06}_{-0.06}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.01$ & $0.28^{+0.01}_{-0.01}$ & $-1.22^{+0.16}_{-0.16}$ & $0.31^{+0.52}_{-0.54}$&$0.64^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.35$ & $0.29^{+0.01}_{-0.01}$ & $-1.07^{+0.12}_{-0.12}$ & $-0.20^{+0.29}_{-0.36}$&$0.67^{+0.05}_{-0.06}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $229.97$ & $0.30^{+0.01}_{-0.01}$ & $-1.06^{+0.11}_{-0.11}$ & $-0.08^{+0.20}_{-0.22}$&$0.66^{+0.05}_{-0.05}$& No\ [|lcccccc|]{}\ Data set& $\chi^{2}_{min}$&$\Omega_{m}$& $w_{0}$&$w_{1}$&$z_{t}$&$z_{sd}$\ \ H(z)+BAO+W9 & $49.82$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.10}_{-0.10}$ & $-0.36^{+0.58}_{-0.74}$&$0.71^{+0.06}_{-0.05}$& No\ H(z)+BAO+Pl & $67.61$ & $0.29^{+0.01}_{-0.01}$ & $-1.09^{+0.10}_{-0.10}$ & $-0.13^{+0.39}_{-0.44}$&$0.69^{+0.05}_{-0.04}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.50$ & $0.28^{+0.02}_{-0.02}$& $-1.05^{+0.25}_{-0.27}$ & $-4.58^{+4.07}_{-3.64}$&$0.72^{+0.11}_{-0.13}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $161.64$ & $0.31^{+0.02}_{-0.03}$& $-0.99^{+0.17}_{-0.18}$ & $-2.54^{+1.95}_{-2.41}$&$0.75^{+0.10}_{-0.10}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.95$ & $0.29^{+0.01}_{-0.01}$ & $-1.15^{+0.09}_{-0.09}$ & $-0.30^{+0.56}_{-0.72}$&$0.70^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.98$ & $0.29^{+0.01}_{-0.01}$ & $-1.12^{+0.09}_{-0.09}$ & $-0.05^{+0.36}_{-0.40}$&$0.68^{+0.04}_{-0.03}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.28$ & $0.29^{+0.01}_{-0.01}$ & $-1.12^{+0.08}_{-0.07}$ & $-0.55^{+0.58}_{-0.76}$&$0.72^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.34$ & $0.29^{+0.007}_{-0.007}$ & $-1.13^{+0.07}_{-0.07}$ & $-0.09^{+0.34}_{-0.39}$&$0.69^{+0.04}_{+0.03}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.60$ & $0.29^{+0.03}_{-0.04}$& $-1.04^{+0.29}_{-0.32}$ & $-4.30^{+3.87}_{-3.72}$&$0.69^{+0.14}_{-0.16}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $170.02$ & $0.31^{+0.02}_{-0.03}$& $-0.95^{+0.16}_{-0.17}$ & $-1.94^{+2.25}_{-3.38}$&$0.73^{+0.12}_{-0.17}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.64$ & $0.29^{+0.01}_{-0.01}$ & $-1.14^{+0.10}_{-0.09}$ & $-0.36^{+0.57}_{-0.76}$&$0.71^{+0.06}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $109.53$ & $0.29^{+0.01}_{-0.01}$ & $-1.11^{+0.09}_{-0.09}$ & $-0.08^{+0.36}_{-0.40}$&$0.68^{+0.05}_{-0.03}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $211.99$ & $0.29^{+0.01}_{-0.01}$ & $-1.08^{+0.08}_{-0.08}$ & $-0.57^{+0.55}_{-0.73}$&$0.73^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $229.97$ & $0.30^{+0.01}_{-0.01}$ & $-1.06^{+0.08}_{-0.08}$ & $-0.25^{+0.35}_{-0.38}$&$0.69^{+0.05}_{-0.04}$& No\ [|lcccccc|]{}\ Data set& $\chi^{2}_{min}$&$\Omega_{m}$& $w_{0}$&$w_{1}$&$z_{t}$&$z_{sd}$\ \ H(z)+BAO+W9 & $50.01$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.15}_{-0.14}$ & $ -1.20^{+0.09}_{-0.12}$&$0.69^{+0.05}_{-0.04}$& No\ H(z)+BAO+Pl & $68.52$ & $0.29^{+0.01}_{-0.01}$ & $-1.10^{+0.13}_{-0.13}$ & $-1.13^{+0.05}_{-0.05}$&$0.68^{+0.04}_{-0.03}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.71$ & $0.28^{+0.03}_{-0.03}$& $-0.83^{+0.50}_{-0.56}$ & $-1.92^{+0.93}_{-1.02}$&$0.67^{+0.12}_{-0.15}$& $0.24^{+0.12}_{-0.11}\,(2.0)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $161.64$ & $0.31^{+0.02}_{-0.03}$& $-0.81^{+0.27}_{-0.27}$ & $-1.48^{+0.28}_{-0.29}$&$0.76^{+0.09}_{-0.09}$& $0.25^{+0.09}_{+0.13}\,(2.2)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $92.11$ & $0.28^{+0.01}_{-0.01}$ & $-1.15^{+0.13}_{-0.12}$ & $-1.21^{+0.09}_{-0.12}$&$0.69^{+0.05}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $110.97$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.12}_{-0.12}$ & $-1.13^{+0.04}_{-0.50}$&$0.67^{+0.04}_{-0.03}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.46$ & $0.29^{+0.01}_{-0.01}$ & $-1.08^{+0.13}_{-0.11}$ & $-1.23^{+0.09}_{-0.11}$&$0.71^{+0.04}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $231.91$ & $0.29^{+0.008}_{-0.008}$ & $-1.12^{+0.10}_{-0.10}$ & $-1.15^{+0.04}_{-0.05}$&$0.68^{+0.04}_{-0.03}$& No\ \ \ ${f_{\mathrm{gas}\,}}^{Allen}$ & $41.76$ & $0.29^{+0.03}_{-0.04}$& $-0.87^{+0.50}_{-0.54}$ & $-1.76^{+0.77}_{-0.93}$&$0.66^{+0.15}_{-0.18}$& $0.25^{+0.13}_{-0.12}\,(2.0)$\ ${f_{\mathrm{gas}\,}}^{ACT}$ & $179.20$ & $0.31^{+0.02}_{-0.04}$& $-0.79^{+0.32}_{-0.29}$ & $-1.39^{+0.42}_{-0.49}$&$0.74^{+0.11}_{-0.15}$& $0.27^{+0.09}_{-0.13}\,(2.4)$\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+W9 & $91.87$ & $0.29^{+0.01}_{-0.01}$ & $-1.13^{+0.13}_{-0.13}$ & $-1.21^{+0.09}_{-0.12}$&$0.69^{+0.05}_{-0.04}$& No\ ${f_{\mathrm{gas}\,}}^{Allen}$+H(z)+BAO+Pl & $110.27$ & $0.29^{+0.01}_{-0.01}$ & $-1.11^{+0.12}_{-0.12}$ & $-1.13^{+0.05}_{-0.05}$&$0.67^{+0.04}_{-0.03}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+W9 & $212.29$ & $0.29^{+0.01}_{-0.01}$ & $-1.05^{+0.13}_{-0.12}$ & $-1.20^{+0.10}_{-0.12}$&$0.71^{+0.05}_{-0.05}$& No\ ${f_{\mathrm{gas}\,}}^{ACT}$+H(z)+BAO+Pl & $230.83$ & $0.30^{+0.01}_{-0.01}$ & $-1.04^{+0.11}_{-0.11}$ & $-1.12^{+0.05}_{-0.05}$&$0.69^{+0.04}_{-0.04}$& No\ JBP parameterization -- -- -- -- -- -- BA parameterization -- -- -- -- -- -- FSLLI parameterization -- -- -- -- -- -- FSLLII parameterization -- -- -- -- -- -- SL parameterization -- -- -- -- -- -- The H(z) data set {#Ap:Hz} ================= $z$ $H(z)$ $\sigma_{H}$ Reference Method ------- -------- -------------- --------------------- ------------ 0.07 69 19.6 [@Zhang2012] DA 0.1 69 12 [@Stern2010] DA 0.12 68.6 26.2 [@Zhang2012] DA 0.17 83 8 [@Stern2010] DA 0.179 75 4 [@Moresco2012] DA 0.199 75 5 [@Moresco2012] DA 0.2 72.9 29.6 [@Zhang2012] DA 0.24 79.69 2.65 [@Gaztanaga2009] Clustering 0.27 77 14 [@Stern2010] DA 0.28 88.8 36.6 [@Zhang2012] DA 0.3 81.7 6.22 [@Oka:2014] Clustering 0.35 82.7 8.4 [@Chuang2013] Clustering 0.352 83 14 [@Moresco2012] DA 0.4 95 17 [@Stern2010] DA 0.43 86.45 3.68 [@Gaztanaga2009] Clustering 0.44 82.6 7.8 [@Blake2012] Clustering 0.48 97 62 [@Stern2010] DA 0.57 92.9 7.8 [@Anderson2014] Clustering 0.593 104 13 [@Moresco2012] DA 0.600 87.9 6.1 [@Blake2012] Clustering 0.68 92 8 [@Moresco2012] DA 0.73 97.3 7 [@Blake2012] Clustering 0.781 105 12 [@Moresco2012] DA 0.875 125 17 [@Moresco2012] DA 0.88 90 40 [@Stern2010] DA 0.9 117 23 [@Stern2010] DA 1.037 154 20 [@Moresco2012] DA 1.3 168 17 [@Stern2010] DA 1.43 177 18 [@Stern2010] DA 1.53 140 14 [@Stern2010] DA 1.75 202 40 [@Stern2010] DA 2.3 224 8 [@Busca2013] Clustering 2.34 222 7 [@Delubac2014] Clustering 2.36 226 8 [@Font-Ribera:2014] Clustering [^1]: juan.magana@uv.cl [^2]: veronica.motta@uv.cl [^3]: victor.cardenas@uv.cl [^4]: gfoex@mpe.mpg.de [^5]: the radius within which the cluster average density is $2500$ times the critical density. [^6]: The mass $M_{500}$ is defined as the one measured within the radius $R_{500}$ at which the enclosed mean density is $500$ times the critical density at the cluster redshift.

ArXiv

6.5mm 1.3cm 22.4cm 16.5cm [\ [**and Financial Markets**]{} ]{} F. M. Ramos[^1], C. Rodrigues Neto and R. R. Rosa Laboratório Associado de Computação e Matemática Aplicada (LAC)\ Instituto Nacional de Pesquisas Espaciais (INPE)\ São José dos Campos - SP, Brazil PACS.02.50-r - Probability theory, stochastic processes, and statistics.\ PACS.47.27Eq - Turbulence simulation and modeling.\ PACS.89.90+n - Other areas of general interest to physicists. [ - We present a new framework for modeling the statistical behavior of both fully developed turbulence and short-term dynamics of financial markets based on the nonextensive thermostatistics proposed by Tsallis. We also show that intermittency – strong bursts in the energy dissipation or clusters of high price volatility – and nonextensivity – anomalous scaling of usually extensive properties like entropy – are naturally linked by a single parameter $q$, from the nonextensive thermostatistics.]{} Scaling invariance plays a fundamental role in many natural phenomena and frequently emerges from some sort of underlying cascade process. A classical example is fully developed homogeneous isotropic three-dimensional turbulence, which is characterized by a cascade of kinetic energy from large forcing scales to smaller and smaller ones through a hierarchy of eddies. At the end of the cascade, the energy dissipates by viscosity, turning into heat. Recently, some authors$^{1-3}$ have studied the phenomenological relationship among financial market dynamics, scaling behavior and hydrodynamic turbulence. Particularly, Ghashghaie [*et al.*]{}$^{2}$ conjectured the existence of a temporal information cascade similar to the spatial energy cascade found in fully developed turbulence. Traditionally, the properties of turbulent flows are studied from the statistics of velocity differences $v_r(x) = v(x) - v(x+r)$ at different scales $r$. As with other physical systems that depend on the dynamical evolution of a large number of nonlinearly coupled subsystems, the energy cascade in turbulence generates a spatial scaling behavior – power-law behavior with $r$ – of the moments $\langle v_r^n \rangle$ of the probability distribution function (PDF) of $v_r$ (the angle brackets $\langle \rangle$ denote the mean value of the enclosed quantity). For large values of the Reynolds number, which measures the ratio of nonlinear inertial forces to the linear dissipative forces within the fluid, there is a wide separation between the scale of energy input (integral scale $L$) and the viscous dissipation scale (Kolmogorov scale $\eta$). Though at large scales ($\sim L$) the PDFs are normally distributed, far from the integral scale they are strongly non-Gaussian and display wings fatter than expected for a normal process. This is the striking signature of the intermittency phenomenon. After publication of the Kolmogorov K62 refined similarity hypotheses$^{8}$, the problem of small scale intermittency became one of the central questions on isotropic turbulence. Over the past years several papers$^{9-16}$ have discussed intermittency and the so-called ‘PDF problem’. Similar attempts$^{1,2,17}$ have been made to explain the same peculiar shape observed in the PDF of price changes $z_{\tau} = z(t) - z(t + \tau)$ at small time intervals. Based on the scaling properties of multifractals, Tsallis$^{4-7}$ has proposed a generalization of Boltzmann-Gibbs thermostatistics by introducing a family of generalized nonextensive entropy functionals $S_q[p]$ with a single parameter $q$. These functionals reduce to the classical, extensive Boltzmann-Gibbs form as $q \rightarrow 1$. Optimizing $S_q[p]$ subject to appropriate constraints$^{6}$, we obtain the distribution $$\label{pdf} p_q(x) = [1 - \beta (1-q) x^2]^{1/(1-q)}/Z_q~~.$$ The appropriate normalization factor, for $1 < q < 3$, is given by $$Z_q \equiv \left[ \frac{\beta (q-1)}{\pi} \right]^{1/2} \frac{\Gamma(1/(q-1))}{\Gamma((3-q)/2(q-1))}~~.$$ In the limit of $q \rightarrow 1$, we recover the Gaussian distribution. The above distribution, we claim, provides the simplest and most accurate model for handling the PDF problem. To show this, we stay in the context of fully developed turbulence ($x \equiv v_r$). From equation (\[pdf\]), we can easily obtain the second moment $$\label{sec} \langle v_r^2 \rangle = \frac{1}{\beta (5 - 3 q)}~~,$$ and the flatness coefficient (kurtosis) $$\label{kur} K_r = \frac{\langle v_r^4 \rangle}{\langle v_r^2 \rangle^2} = \frac{3 \, (5 - 3 q)}{(7 - 5 q)}~~.$$ We remark that the flatness coefficient, which is directly related to the occurrence of intermittency, is solely determined by the parameter $q$. Also, we note that the positiveness of $K_r$ sets an upper bound on the value of $q$, namely $q < 7/5$. This bound coincides with the one obtained by Boghossian$^{18}$ through a $q$-generalization of Navier-Stokes equations. Moreover, this limit implies that the second moment of distribution (\[pdf\]) will always remain finite, which is empirically expected from the phenomena here analyzed. At this point, if we assume$^{2,9,10}$ a scaling of the moments $\langle v_r^n \rangle$ of $v_r$ as $r^{\zeta_n}$, the variation with $r$ of the PDF of the velocity differences and of its related moments can be completely determined. Under the assumptions of Kolmogorov log-normal model$^{8}$, $\zeta_n = n/3 - \frac{1}{18} \mu n (n-3)$, where $\mu$ is the intermittency exponent. Then, we immediately obtain the functional forms of the flatness coefficient and the parameter $q$, respectively $$\label{kur2} K_r = K_L (r/L)^{\alpha}$$ and $$\label{qq} q =\frac{5 - 7 (r/L)^{\alpha}}{3 - 5 (r/L)^{\alpha}}~~,$$ with $K_L =3$, the expected value for a Gaussian process, and $\alpha = - 4 \mu/9$. The correspondent expression for $\beta$ can be derived similarly from equation (\[sec\]). However, to account for the well known asymmetry of the velocity distributions we may consider $\beta=\beta^+$, for $v_r \geq 0$, and $\beta=\beta^-$, for $v_r < 0$. In this case, both the second and third moments of the modified PDF shall be used to determine $\beta^-$ and $\beta^+$. Equations (\[kur\]) and (\[qq\]) remain unchanged, as far as the asymmetry of the PDF is small. We checked our model with turbulence statistics data taken from reference 2, provided by Chabaud [*et al.*]{}$^{13}$. Firstly, we adjusted by least-squares to the experimental data, the parameters of our PDF, considering both the symmetric and asymmetric forms of equation (\[pdf\]). Then, we compared the estimated values of $q$ with the predictions of equation (\[qq\]), for a intermittency coefficient of $\mu =0.25$ (the best estimate currently available is$^{12}$ $0.25 \pm 0.05$). The results are displayed in Figs. 1a and 2. A good agreement is observed through all spatial scales and for all orders of normalized velocity differences. Note that the solid line in Fig. 2 is not simply a fit to the data, since $\mu$ is obtained independently. We also computed value of $q$ at the Kolmogorov scale and obtained $q_{\eta} \simeq 9/7$. The same approach adopted in turbulence can be straightforwardly applied (with $x \equiv z_{\tau}$) to model the statistics of price differences in financial markets, as far as the relevant parameters at the integral time scale – time span for which a convergence to a Gaussian process is found – are available. We tested our model with price changes data taken from reference 2, provided by Olsen & Associates. The results are displayed in Figs. 1b and 2. Since we do not have an independent estimate of the intermittency coefficient of the information cascade, we also adjusted equation (\[qq\]) to the data and obtained $\alpha = -0.17$ and $\tau_L \simeq 2.2$ days, the corresponding integral scale of the process. We observe that the proposed model reproduces with good accuracy the statistics of price differences over all temporal scales. However, we remark that the integral scale value of 2.2 days strongly disagrees with other estimates$^{1}$ of $\tau_L$ (roughly 1 month). Nonextensivity, a matter of speculation in some areas$^{19}$, is an essential feature of Tsallis generalized thermostatistics. If we suppose a scenario of a cascade of bifurcations with $n$ levels, and scale the generalized entropy $S_q$, averaged over a volume of size $V=\eta^3$ and normalized by Boltzmann constant, we have at the first level $$S_q(2V) = 2 S_q(V) + (1-q) S_q^2(V)$$ and at the top of the cascade $$\label{casca1} S_q(2^n V) = 2 S_q(2^{n-1}V) + (1-q) S_q^2(2^{n-1}V) \simeq 2^n S_q(V) + (1-q) 2^{n-1} S_q(V)~~,$$ where the higher order terms in $S_q$ have been discharged. Cascade processes are also described in terms of fractal or multifractals models$^{11,20-23}$. Within these frameworks, in high Reynolds number turbulence, the energy dissipation is not uniformly distributed within the fluid but rather concentrated on subsets of non-integer fractal $D_F$ dimension. This picture leads to a scaling behavior with dimensionality not equal to that of the embedding space. In this case, if we consider the cascade of bifurcations described above, we find $$\label{casca2} S_q(2^n V) = 2^{n D_F/3} S_q(V)~~.$$ It follows immediately from equations (\[casca1\]) and (\[casca2\]) that $$\label{casca3} D_F \simeq \frac{3}{n} \left[ \frac{log(3 - q)}{log(2)} + n - 1 \right]~~.$$ Note that the parameter $q$, through equation (\[casca3\]), offers a quantitative picture of the transition from small-scale intermittent, nonextensive, fractal behavior to large-scale Gaussian, extensive homogeneity. For higher values of $n$ and $q=1$, at the top of the cascade, we have $D_F = 3$. At the bottom of the cascade ($n =1$), using the value $q_{\eta} \simeq 9/7$, estimated previously, we get $D_F \simeq 2.33$, a good approximation of the so-called ‘magic’ value of 2.35, often measured in different experimental contexts$^{24-27}$. This result and the divergence of the correlation function of the energy dissipation in the fluid$^{9}$ $$\label{dis} \langle \epsilon(x)\epsilon(x+r) \rangle \sim \frac{\langle v_r^6 \rangle}{r^2} = \frac{1}{r^2} \left[ \frac{15 \beta^{-2}}{(5 - 3 q)(7 - 5 q)(9 - 7 q)} \right]$$ suggest a more stringent bound on $q$, namely $q < 9/7$. If true, this new bound immediately implies that, in fully developed turbulence, relation $(\eta/L)^{\alpha} = 2$ is invariant, regardless the Reynolds number. Though only qualitatively, the above picture may be also applied to the information cascade. However, since there is nothing equivalent to viscous damping in the dynamics of speculative markets, the information cascade depth is only limited by the minimum time necessary to perform a trading transaction. [**References**]{} 1\. Mantegna, R.N. & Stanley, H.E. Scaling behaviour in the dynamics of an economic index. [*Nature*]{} [**376**]{}, 46-49 (1995). 2\. Ghashghaie, S., Breymann, W., Peinke, J., Talkner, P. & Dodge, Y. Turbulent cascades in foreign exchange markets. [*Nature*]{} [**381**]{}, 767-770 (1996). 3\. Mantegna, R.N. & Stanley, H.E. Turbulence and financial markets. [*Nature*]{} [**383**]{}, 587-588 (1996). 4\. Tsallis, C. Possible generalization of Boltzmann-Gibbs statistics. [*J. Stat. Phys.*]{} [**52**]{}, 479-487 (1988). 5\. Tsallis, C., Sá Barreto, F.C. & Loh, E.D. Generalization of the Planck radiation law and application to the cosmic microwave background radiation. [*Phys. Rev. E*]{} [**52**]{}, 1447-1451 (1995). 6\. Tsallis, C., Levy, S. V. F., Souza, A. M. C. & Maynard, R. Statistical-mechanical foundation of the ubiquity of Lévy distributions in nature. [*Phys. Rev. Lett.*]{} [**75**]{}, 3589-3593 (1995). 7\. Lyra, M. L. & Tsallis, C. Nonextensivity and multifractality in low-dimensional dissipative systems [*Phys. Rev. Lett.*]{} [**80**]{}, 53-56 (1998). 8\. Kolmogorov, A. M. A refinement of previous hypotheses concerning the local structure of turbulence of a viscous incompressible fluid at high Reynolds number. [*J. Fluid Mech.*]{} [**13**]{}, 82-85 (1962). 9\. Anselmet, F., Gagne, Y., Hopfinger, E. J. & Antonia, R. A. High-order velocity structure functions in turbulent shear flows. [*J. Fluid Mech.*]{} [**140**]{}, 63-89 (1984). 10\. Castaing, B., Gagne, Y. & Hopfinger, E. J. Velocity probability density functions of high Reynolds number turbulence. [*Physica D*]{} [**46**]{}, 177-200 (1990). 11\. Benzi, R., Biferale, L., Paladin, G., Vulpiani, A. & Vergassola, M. Multifractality in the statistics of the velocity gradients in turbulence. [*Phys. Rev. Lett.*]{} [**67**]{}, 2299-2302 (1991). 12\. Sreenivasan, K. R. & Kailasnath, P. An update on the intermittency exponent in turbulence. [*Phys. Fluids*]{} [**5**]{}, 512-514 (1992). 13\. Chabaud et al. Transition towards developed turbulence. [*Phys. Rev. Lett.*]{} [**73**]{}, 3227-3230, (1994). 14\. Praskovsky, A. & Oncley, S. Measurements of the Kolmogorov constant and intermittency exponent at very high Reynolds number. [*Phys. Fluids*]{} [**6**]{}, 2886-2888 (1994) 15\. Vassilicos, J.C. Turbulence and intermittency [*Nature*]{} [**374**]{}, 408-409 (1995). 16\. Naert, A., Castaing, B., Chabaud, B., B. Hébral & Peinke, J. Conditional statistics of velocity fluctuations in turbulence. [*Physica D*]{} [**113**]{}, 73-78 (1998). 17\. Taylor, S. J. [*Math. Fin.*]{} Modeling stochastic volatility: a review and comparative study. [**4**]{}, 183-204 (1994). 18\. Boghossian, B. M., Navier-Stokes equations for generalized thermostatistics. [*Braz. J. Phys.*]{} (in the press). 19\. Maddox, J., When entropy does not seem extensive. [*Nature*]{} [**365**]{}, 103- (1993). 20\. Mandelbrot, B. B. Intermittent turbulence in self-similar cascades: divergence of high moments and dimension of the carrier. [*J. Fluid Mech.*]{} [**62**]{}, 331-358 (1974). 21\. Parisi, G. & Frisch, U. in [*Turbulence and Predictability in Geophysical Fluid Dynamics and Climatic Dynamics*]{} (ed. Ghil, M., Benzi, R., & Parisi, G.) 84- (North-Holland, Amsterdam, 1985). 22\. Paladin, G. & Vulpiani, A. Anomalous scaling laws in multifractal objects. [*Phys. Rep.*]{} [**156**]{}, 147-225 (1987). 23\. Meneveau, C. & Sreenivasan, K. R. The multifractal nature of turbulent energy dissipation. [*J. Fluid Mech.*]{} [**224**]{}, 429-484 (1991). 24\. Hentschel, H. G. E. & Procaccia, I. Relative diffusion in turbulent media: the fractal dimension of clouds. [*Phys. Rev. A*]{} [**29**]{}, 1461-1470, (1984). 25\. Sreenivasan, K. R. & Meneveau, C. Fractal facets of turbulence. [*J. Fluid Mech.*]{} [**173**]{}, 357-386 (1986). 26\. Sreenivasan, K. R., Ramshankar, R. & Meneveau, C. Mixing, entrainment and fractal dimensions of surfaces in turbulent flows. [*Proc. R. Soc. Lond. A*]{} [**421**]{}, 79-108 (1989). 27\. Vassilicos, J. C. & Hunt, J. C. R. Fractal dimensions and spectra of interfaces with application to turbulence. [*Proc. R. Soc. Lond. A*]{} [**435**]{}, 505-534 (1991). [**Acknowledgments:**]{} We thank C. Tsallis for helpful discussions. This work was supported by FAPESP-Brazil and CNPq-Brazil. [**Captions**]{} [**Figure 1**]{} (a) Data points: standardized probability distribution $p_q(v_r)$ of velocity differences $v_r(x) = v(x) - v(x+r)$ for spatial scales $r = 0.0073L$, $0.0407L$, $0.3036L$, $0.7150L$, with $L/\eta=454$ and $L$ and $\eta$ being, respectively, the integral and Kolmogorov scales (see text); data taken from ref. 2, provided by Chabaud [*et al.*]{}$^{13}$; Solid lines: least-squares fits of modified PDF (\[pdf\]); from top to bottom: $q= 1.26$, 1.20, 1.11, 1.08; $\beta^- = 0.69$, 0.66, 0.55, 0.62; $\beta^+ = 0.88$, 0.82, 0.76, 0.70 (for better visibility the curves have been vertically shifted with respect to each other). [**Figure 1**]{} (b) Data points: standardized probability distribution $p_q(z_{\tau})$ of price differences $z_{\tau} = z(t) - z(t + \tau)$ for temporal scales $\tau = 0.0035\tau_L$, $0.0276\tau_L$, $0.2210\tau_L$, $0.8838\tau_L$, with $\tau_L=186265~s$ being the integral scale (see text); data taken from ref. 2, provided by Olsen & Associates; Solid lines: least-squares fits of modified PDF (\[pdf\]); from top to bottom: $q= 1.35$, 1.26, 1.16, 1.11; $\beta^- = 1.12$, 0.83, 0.75, 0.75; $\beta^+ = 0.98$, 0.72, 0.61, 0.77.(for better visibility the curves have been vertically shifted with respect to each other). [**Figure 2**]{} Dependence of the parameter $q$ on normalized spatial ($r/L$) and temporal scales ($\tau/\tau_L$); turbulence data: solid line (model), open squares (data, asymmetric adjust), solid squares (data, symmetric adjust); financial data: dotted line (model), open triangles (data, asymmetric adjust), solid triangles (data, symmetric adjust). [^1]: e-mail:fernando@lac.inpe.br

ArXiv

--- abstract: 'We investigate by Monte Carlo simulations the structure, energetics and superfluid properties of thin $^4$He films (up to four layers) on a glass substrate, at low temperature. The first adsorbed layer is found to be solid and “inert", i.e., atoms are localized and do not participate to quantum exchanges. Additional layers are liquid, with no clear layer separation above the second one. It is found that a single $^3$He impurity resides on the outmost layer, not significantly further away from the substrate than $^4$He atoms on the same layer.' author: - Massimo Boninsegni title: Thin helium film on a glass substrate --- Introduction {#intro} ============ Adsorption of thin films of the most abundant isotope of He ($^4$He) on various substrates, is a subject of longstanding interest, motivated by the unusual, often intriguing physical properties that these films display [@bruch97]. A chief example is the rich phase diagram of Helium on graphite, which has been the subject of an intense investigation over the past three decades. Significant attention has also been devoted to adsorption on weakly attractive substrates, e.g., those of alkali metals, following original suggestions that, at low temperature, the weakness of the physisorption potential ought to result in unusual wetting [@cheng91; @cheng93] and superfluid [@boninsegni99] behavior. Less is known quantitatively, both theoretically and experimentally, about adsorption over a glass substrate, despite its obvious relevance to experimental studies of the superfluid properties of liquid (and solid) helium embedded in such porous materials such as vycor or aerogel [@reppy92]. Of particular interest are issues such as the interplay between the one or more solid, “inert" layers that are expected to form over the substrate, and the liquid-like behaviour that should progressively emerge as successive layers are adsorbed. Glass is the archetypal disordered substrate, and that renders its theoretical modelling particularly challenging. Microscopic theoretical studies of adsorption of helium (both pure [@apaja03b; @buffoni05] and in isotopic mixtures [@pricaupenko95]) on glass, based on realistic interaction potentials, have typically made use of a vary simple model of the substrate, described as a flat surface on account of its lack of a definite crystal structure. As mentioned above, the substrate is in reality irregular, disordered, and this can be expected to promote the appearance of a stable, low-coverage insulating film, in which atoms are localized at randomly located wells of the adsorption potential; such a phase is obviously lost if a flat substrate model is assumed. As the coverage is increased, however, a first-order transition to a regular (monolayer) crystal is expected, and as further layers are adsorbed the corrugation of the substrate ought to become less relevant, i.e., a flat substrate model should capture the essential physics of the problem (this is what observed, e.g., in numerical simulations of helium films on regular crystalline substrates, such as graphite [@corboz08]). In this work we study the adsorption of $^4$He on a glass substrate, modelled as in all previous studies by others, by means of numerical simulations based on the continuous-space Worm Algorithm (WA). The purpose of this investigation is that of determining quantitatively structural and superfluid properties of the adsorbed $^4$He film at low temperature. Specifically, we present results for $T$=1 K and $T$=0.5 K. We considered films up to four layers thick, and also looked at a single $^3$He impurity diluted in the film, for different values of coverage. There is considerable uncertainty as to the actual values of the parameters that are used in the potential describing the interaction of a helium atom with the substrate, and we attempted to pick values that overlap with most previous studies. Our main findings are the following: 1. [The first helium adlayer is solid, whereas the successive ones are liquid. In the solid “inert" layer, atoms are highly localized, the layer is fairly well separated from the second, and quantum exchanges (both intra- and inter-layer) are virtually absent. On the other hand, successive layers are liquid-like, with no sharp separation between layers above the second. Atomic exchanges take place, mostly within the layer for the second layer, as well as inter-layer above the second one. Layers above the first are all superfluid in the low temperature limit. Thus, one can conceptually distinguish the first adlayer from the rest of the system. Previous studies [@pricaupenko95] have often assumed that two inert layers would form, but the results presented here show that the model potential utilized is not sufficiently strong to stabilize more than one such layer.]{} 2. [A single $^3$He impurity diluted in the film, for different values of $^4$He coverage, always positions itself on the top layer. Its wave function is not significantly more spread out than that of a $^4$He atom on the same layer, i.e., the $^3$He atom does not spend a significantly fraction of time at greater distances from the substrate. In other words, $^3$He atoms do not “float" on top of the $^4$He film, as one might have expected.]{} The remainder of this manuscript is organized as follows: in the next section, the model Hamiltonian is briefly introduced, and some details of the computational method utilized are presented (for a thorough illustration of the methodology, the reader is referred to Refs. ); we then discuss our results and outline our conclusions in the two following sections. Model ===== Our system of interest is modelled as an ensemble of $N$ $^4$He atoms, regarded as point particles, moving in the presence of an infinite, smooth planar substrate (positioned at $z$=0). The system is enclosed in a vessel shaped as a parallelepiped, with periodic boundary conditions in all directions [@note]. Let $A$ be the area of the substrate; correspondingly, the nominal $^4$He coverage is $\theta=N/A$. The quantum-mechanical many-body Hamiltonian is the following: $$\label{one} \hat H = -{\lambda}\sum_{i=1}^N \nabla_i^2 + \sum_{i<j} V(r_{ij}) +\sum_{i=1}^N U(z_i)\ .$$ Here, $\lambda$=6.0596 KÅ$^2$, $V$ is the potential describing the interaction between two helium atoms, only depending on their relative distance, whereas $U$ is the potential describing the interaction of a helium atom with the substrate, also depending only on the distance of the atom from the substrate. We use the accepted Aziz potential [@aziz79] to describe the interaction of two Helium atoms, which has been used in almost all previous studies, and has also been shown [@worm2] to provide an accurate description of the energetics and structural properties of liquid $^4$He in the superfluid phase. More delicate is the issue of which potential to use to describe the interaction of a helium atom with the smooth substrate (i.e., the $U$ term in (\[one\])). The simplest potential is the so-called “3-9": $$\label{39} U_{3-9}(z) = \frac{D}{2}\biggl ( \frac {a^9}{z^9}-3\frac{a^3}{z^3}\biggr )$$ which is a functional form obtained by integrating the Lennard-Jones potential over a semi-infinite, continuous slab. The parameters $a$ and $D$ are normally adjusted to fit the results of some [*ab initio*]{} electronic structure calculations for the specific adatom-substrate system of interest. $D$ is the characteristic depth of the attractive well of the potential, whereas $a$ is essentially the location of the minimum of such a well. There appears to be some ambiguity as to the values of the coefficients that would make this potential appropriate for the system of interest here. Choices made in some of the previous studies are summarized in Table \[tab:1\]. Although these studies were aimed at describing the environment experienced by a helium atom in the confines of various porous materials, such as vycor, the “building block" for all of them is the interaction potential felt by a helium atom close to an infinite, flat glass (SiO$_2$) substrate. As is seen in the table, there seems to be a reasonable agreement over the value of $D$, which is $\sim$ 100 K in all studies. On the other hand, while $a$ has been taken to be slightly above 2 Å  in Refs. , in Ref. it was set to a much larger value, namely 3.6 Å. As shown in Fig. \[fig:1\], the ensuing potentials are quantitatively very different, the parameters utilized in Ref. giving rise to a potential with a significantly more attractive long-range tail (in turn likely to give rise to quantitatively different adsorption characteristics).\ The claim is made in Ref. that $a$=3.6 Å applies to the interaction of helium atoms with a substrate of MgO (regarded as sufficiently similar to SiO$_2$); but the most reliable estimates reported in the literature (see, e.g., Ref. ) place $a$ for such a system in the 2.2-2.5 Å range. In this work, we utilize the same values of $a$ and $D$ taken in Ref. , which are close to those of Ref. . [lll]{} Ref. & $a$ (Å) & $D$ (K)\ & 2.18 & 128\ & 3.6 & 87\ & 2.05 & 100\ Methodology =========== We performed Quantum Monte Carlo (QMC) simulations of the system described by Eq. (\[one\]) using the continuous-space Worm Algorithm. This methodology has emerged over the past few years as a far superior alternative to conventional Path Integral Monte Carlo (PIMC), to compute accurate thermodynamic properties of Bose systems at finite temperature. Specifically, the WA allows one to simulate systems comprising up to two orders of magnitude more particles than PIMC, allowing for a reliable extrapolation of values of physical observables to the thermodynamic limit.\ The specific implementation utilized in this project is [*canonical*]{}, i.e., we keep the number $N$ of particles fixed [@fabio]. Other technical aspects of the calculations are common to any other QMC simulation scheme. Results ======= Computational details --------------------- Results were obtained by simulating systems comprising a number $N$ of particles ranging from 36 to 108, spanning a range of $^4$He coverage $\theta$ from 0.050 to 0.300 Å$^{-2}$. Estimates presented here are obtained at two temperatures, namely $T$=1 K and $T$=0.5 K; based on the wealth of accumulated theoretical and experimental data for liquid $^4$He, $T$=1 K is low enough that structural and energetic properties are essentially those of the ground state. The usual fourth-order high-temperature propagator utilized in all previous studies based on the WA was adopted here; convergence of the estimates was observed for a time step $\tau\approx 3.125\times 10^{-3}$ K$^{-1}$, which corresponds to $P$=320 imaginary time “slices" at $T$=1 K. Because the computational cost was negligible, all estimates reported here were obtained using twice as small a time step, in order to be on the safe side. Energetics ---------- Fig. \[fig:2\] shows the energy per particle $e(\theta)$ as a function of coverage, computed by QMC simulations at $T$=1 K. Assuming these to be essentially ground state results, based on a piecewise polynomial fit to the estimates for $e(\theta)$ and the relation $$\mu(\theta) = e(\theta) + \theta \frac {de}{d\theta}$$ we obtained the $T$=0 chemical potential, also shown in Fig. \[fig:2\].\ The monolayer equilibrium coverage $\theta_e$, for which $\mu(\theta_e) = e(\theta_e)$, is found to be close to 0.052 Å$^{-2}$, to be compared to 0.043 Å$^{-2}$ for the purely two-dimensional system [@giorgini]. This corresponds to a superfluid monolayer, but it is important to restate that the model of substrate considered here lacks the corrugation and disorder of the surface, which will stabilize a disordered, “glassy" insulating monolayer in the real system. On increasing coverage, the system forms a two-dimensional (triangular) crystal at $\theta_{1C}\approx$ 0.08 Å$^{-2}$. The crystal can be compressed up to $\theta_{1P}$=0.100 Å$^{-2}$, at which second layer promotion is observed. The system forms a stable liquid second layer at $\theta_{2L}$=0.153 Å$^{-2}$, whereas in the coverage region $\theta_{1P} \le \theta\le \theta_{2L}$ puddles of liquid form on top of the solid first layer.\ Above $\theta_{2L}$, the chemical potential is a smooth, monotonically increasing function of coverage, approaching asymptotically the value $\sim$ -7.2 K for bulk liquid $^4$He. This signals continuous growth of the adsorbed film, without well-defined layer formation. This is confirmed by direct observation of the structure of the film, as well as by its superfluid properties. Structure --------- Fig. \[fig:3\] shows the computed $^4$He density profiles $n(z)$, in the direction perpendicular to the substrate, for four different coverages. The first observation is that there exists a fairly sharp demarcation between the innermost (first) layer and the second, with essentially no overlap between the two, and no interlayer atomic exchange. On the other hand, second and third layers already overlap significantly, and essentially no separation exists between third and fourth layers. The fourth layer is broad and ill-defined, its average density very close to that of equilibrium liquid $^4$He (0.022 Å$^{-3}$). Additional evidence of the ill-defined character of layers in this system is provided in Fig. \[integ\], where the integrated coverage \[inte\] (z)dx\^ dy\^ \_0\^z dz\^(x\^,y\^,z\^) which depends on the distance from the substrate, is shown for a film of total coverage $\theta$=0.299 Å$^{-2}$. Steps can be seen in correspondence of the appearance of the second layer (at approximately 0.100 Å$^{-2}$) and of the third at roughly 0.165 Å$^{-2}$ can be clearly seen. However, the boundary between third and fourth layer is already scarcely visible. This is not qualitatively dissimilar from what is observed on other substrates [@pavloff]. or The wealth of theoretical data accumulated over a wide variety of substrates, generally suggests that although the details of the first layer clearly reflect the properties of the underlying substrate (strength and corrugation), beyond the first layer what one observes seems to be essentially substrate-independent. Atoms in the first layer are arranged on a regular triangular crystal and are not involved in permutational exchanges, either within the layer itself or with atoms of the second layer. Both the first and second layers are slightly compressed as further layers form on top. At its highest compression, the second layer has a half-width at half maximum around 1 Å, and corresponds to an effective coverage of approximately 0.050 Å$^{-2}$, i.e., the second layer is already liquid. This contradicts the assumption of two solid adsorbed layers of helium on the substrate, modelled as in this work (with the same choice of parameters), made in Ref. , as the potential (\[39\]) with the choice of parameters made here, only leads to the formation of a single solid layer. Superfluidity ------------- Unlike structure and energetics, the superfluid properties of a systems studied by computer simulation are strongly affected by its finite size, i.e., the number of particles. In particular, phase transitions are smeared, and one will typically observe a finite value of the superfluid fraction $\rho_S$ for temperatures above the critical temperature $T_c$. In this study, given the crudeness of the model utilized we are not interested in providing accurate values of the transition temperature, but rather to assess reliably whether an adsorbed film of a given coverage can be expected to display superfluidity below a reasonably well-defined temperature. Because of the two-dimensional (2D) nature of the system studied here, the superfluid transition is expected to fall in the Kosterlitz-Thouless universality class [@kt78], with a sharp universal “step" at $T_c$. Finite-size effects affect mostly estimates obtained in the vicinity of $T_c$, whereas for $T < T_c$, the results for a finite system are quantitatively representative of the thermodynamic limit, where $\rho_S \sim \rho_S(T=0)$ \[unity for a translationally invariant system\]. The most recent estimate for $T_c^{2D}$ is 0.655 K, based on the Aziz pair potential utilized here, which affords fairly closed agreement with experiment for bulk liquid $^4$He [@worm; @worm2]. Computer simulations of an adsorbed $^4$He monolayer on Lithium [@boninsegni99], yielded evidence that such a film essentially mimics the physics of 2D $^4$He, and because the substrate considered here is considerably more attractive than Lithium, we can expect the same to be true here. Indeed, the superfluid fraction of the equilibrium stable monolayer film at $T$=0.5 K, is close to 100%, consistently with its behaviour being fairly close to that of purely two-dimensional $^4$He (such a monolayer is considerably narrower than on a Li substrate, see Ref. ). As mentioned above, as the coverage is increased the first layer crystallizes, and superfluidity disappears, to reappear only when a stable liquid second layer forms, at a coverage $\theta_{2L}$. As shown in Fig. \[fig:4\], the superfluid fraction at $\theta=\theta_{2L}$ at $T$=0.5 K is close to 33%, i.e., consistent with an underlying non-superfluid crystalline layer, and an entirely superfluid second layer on top. That particles in the first layer are not involved in quantum-mechanical exchanges is inferred by histograms of permutation cycles, as well as from the lack of overlap between the first and second layer, as shown in Figs. \[fig:3\] and \[fig:snap\].\ Given that no evidence is observed of crystallization for layers above the second, and that the superfluid transition temperature is expected to be [*generally*]{} an increasing function of the film thickness (the superfluid transition temperature should approach the bulk value $T$=2.177 K in the thick film limit), it seems reasonable to posit that $T$=0.5 K should be sufficiently low to render finite-size effects essentially unnoticeable. At coverages above $\theta_{2L}$, quantum exchanges involving particles in the same layer (other than the first) occur at $T$=1 K already, interlayer exchanges being more frequent between the third and fourth layer, which overlap most substantially (as shown in Fig. \[fig:snap\]). Consequently, the superfluid fraction is finite for any stable film consisting of more than one layer, confirming that all layers above the first are liquid. Results for the $^4$He superfluid fraction $\rho_S$ at $T$=0.5 K are shown in Fig. \[fig:4\]. The behaviour as a function of $\theta$, for the coverages above $\theta_{2L}$ is oscillatory, with local minima indicating layer “completion". However, $\rho_S$ stays finite at all coverages. As the coverage increases, the superfluid fraction is expected to approach 100% in the thick film limit. Single $^3$He impurity. ----------------------- ![Probability density of position in the direction perpendicular to the substrate, for a $^3$He atom embedded in a $^4$He film of coverage $\theta$=0.100 Å$^{-2}$ (top, single layer), 0.178 Å$^{-2}$ (middle, two layers) and 0.244 Å$^{-2}$ (bottom, three layers). Dashed lines show the corresponding probability density for a $^4$He atom in the same layer where the $^3$He atom resides, which is always the top one. Units on the ordinate axis are arbitrary.[]{data-label="fig:5"}](f5){width="40.00000%"} The physics of a single $^3$He atom dissolved in an adsorbed $^4$He film is of interest for a number of reasons. First of all, $^4$He is never isotopically pure, as contaminating $^3$He impurities are always present and their effect on the behaviour of $^4$He in confinement, or restricted geometries, is not completely understood. Moreover, isotopic liquid helium mixtures are of considerable fundamental interest, as one of the simplest Fermi-Bose mixtures accessible in nature. Their phase diagram has been extensively investigated experimentally, including in the confines of porous glasses such as aerogel [@kim93]. The most basic question that one can ask, is in what region of an adsorbed $^4$He film of varying thickness will a $^3$He atom position itself. The wave function of a single $^3$He atom embedded in an adsorbed $^4$He monolayer is expected to be more “spread out", i.e., the atom is expected spend more time further away from the substrate than a $^4$He, because of its lighter mass. In the case of a few adsorbed $^4$He layers, it is presumed that the $^3$He atom will reside in the top layer, near the free surface of the film, where the three-dimensional density approaches that of bulk liquid $^4$He. Conceivably, the $^3$He atom might even “float" on top of the outmost $^4$He layer. In order to assess these issues quantitatively, we have carried out simulations of a single $^4$He impurity at the same helium coverages considered for the case of a pure $^4$He film. At all coverages explored here, the $^3$He atom is found to reside on the top layer of the film. Fig. \[fig:5\] shows the probability density of position in the direction perpendicular to the substrate, for a single $^3$He atom embedded in adsorbed $^4$He films of various coverages, corresponding to one, two and three adsorbed layers.\ In all cases shown, the $^3$He and $^4$He density profiles overlap substantially, to indicate that the $^3$He impurity is essentially embedded in the $^4$He layer; it does not, e.g., float atop. On average, the lighter atom lays only very slightly further away from the substrate than a $^4$He atom on the same layer, although this displacement becomes more pronounced as the number of adlayers increases. This is attributable to the weakening of the attraction exerted by the substrate on the $^3$He atom (laying near the surface) as the film thickens, and the enhanced quantum delocalization of $^3$He with respect to atoms of the heavier isotope. In the limit of a thick $^4$He film, the $^3$He impurity is expected to remain near the surface, where the presence of a bound state has been predicted theoretically [@saam; @cohen73; @epstein] and verified experimentally [@guo; @gasparini]. Conclusions =========== We have studied by accurate, first principles Quantum Monte Carlo simulations the physics of an adsorbed film of $^4$He over a model glass substrate, at low temperature. We utilized the most commonly adopted model of a glass substrate, neglecting the disorder and irregularity of its surface. We considered films up to four layers thick, and also looked at a single $^3$He impurity diluted in the film, for different values of coverage.\ It is found that only the first adlayer is solid, whereas the successive ones are liquid. The solid “inert" layer layer is fairly well separated from the second, atoms are highly localized and no quantum exchanges (both intra- and inter-layer) take place. On the other hand, successive layers are liquid-like, with no sharp separation between layers above the third. In the second layers, quantum exchanges are mainly confined to the layer itself, whereas interlayer exchanges occur between third and fourth layers.\ A single $^3$He impurity is always observed on the top layer, not near the substrate. This appears to be in contrast with the conclusion of Ref. [@pricaupenko95], where the existence of a homogeneous $^3$He-$^4$He mixture was predicted near the substrate, based on a density functional calculation. On the other hand, the results shown here lend credence to the scenario proposed in Ref. [@boninsegni95], according to which the heavier isotope in the mixture, $^4$He, would be prevalently found in the vicinity of the surface of, e.g., aerogel strands. The wave function of the $^3$He atom, however, is not significantly more spread out than that of a $^4$He atom on the same layer, i.e., the $^3$He atom does not spend a significantly fraction of time at greater distances from the substrate. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported in part by the Natural Science and Engineering Research Council of Canada under research grant 121210893, and by the Alberta Informatics Circle of Research Excellence (iCore). See, for instance, L. W. Bruch, M. W. Cole and E. Zaremba, Physical Adsorption: Forces and Phenomena. Clarendon Press, Oxford (1997). E. Cheng, M. W. Cole, W. F. Saam and J. Treiner, Phys. Rev. Lett. [**67**]{}, 1007 (1991). E. Cheng, M. W. Cole and J. Dupont-Roc, Rev. Mod. Phys. [**65**]{}, 557 (1993). M. Boninsegni, M. W. Cole and F. Toigo, [Phys. Rev. Lett.]{} [**83**]{}, 2002 (1999). See, for instance, J. D. Reppy, J. Low Temp. Phys. [**87**]{}, 205 (1992). V. Apaja and E. Krotscheck, Phys. Rev. B [**67**]{}, 184304 (2003). S. A. Khairallah and D. M. Ceperley, Phys. Rev. Lett. [**95**]{}, 185301 (2005). L. Pricaupenko and J. Treiner, Phys. Rev. Lett. [**74**]{}, 430 (1995). See, for instance, P. Corboz, M. Boninsegni, L. Pollet and M. Troyer, Phys. Rev. B [**78**]{}, 245414 (2008). M. Boninsegni, N. Prokof’ev and B. Svistunov, Phys. Rev. Lett. [**96**]{}, 070601 (2006). M. Boninsegni, N. Prokof’ev and B. Svistunov, Phys. Rev. E [**74**]{}, 036701 (2007). R. A. Aziz, V. P. S. Nain, J. S. Carley, W. L. Taylor and G. T. McConville, [J. Chem. Phys.]{} [**70**]{}, 4330 (1979). G. Vidali, G. Ihm, H.-Y. Kim and M. W. Cole, Surf. Sci. Rep. [**12**]{}, 133 (1991). See, for instance, F. Mezzacapo and M. Boninsegni, Phys. Rev. Lett. [**97**]{}, 045301 (2006). S. Giorgini, J. Boronat and J. Casulleras, Phys. Rev. B [**54**]{}, 6099 (1996). See, for instance, N. Pavloff and J. Treiner. J. Low Temp. Phys. [**83**]{}, 331 (1991). J. M. Kosterlitz and D. J. Thouless, Prog. Low Temp. Phys. [**VIIB**]{}, 371 (1978). See, for instance, S. B. Kim, J. Ma and M. H. W. Chan, Phys. Rev. Lett. [**71**]{}, 2268 (1993). W. F. Saam, Phys. Rev. A [**4**]{}, 1278 (1971). C. C. Chang and M. Cohen, Phys. Rev. A [**8**]{}, 3131 (1973). E. Krotscheck, M. Saarela, and J. L. Epstein, Phys. Rev. B [**38**]{}, 111 (1988). H. M. Guo, D. O. Edwards, R. E. Sarwinski and J. T. Tough, Phys. Rev. Lett. [**27**]{}, 1259 (1971). J. R. Eckardt, D. O. Edwards, P. P. Fatouros, F. M. Gasparini, and S. Y. Shen, Phys. Rev. Lett. [**32**]{}, 706 (1974). M. Boninsegni and D. M. Ceperley, Phys. Rev. Lett. [**74**]{}, 2288 (1995).

ArXiv

--- abstract: 'Robust, or model-independent properties of the variance swap are well-known, and date back to @Dupire:93 and @Neuberger:94, who showed that, given the price of co-terminal call options, the price of a variance swap was exactly specified under the assumption that the price process is continuous. In @CoxWang:11 we showed that a lower bound on the price of a variance call could be established using a solution to the Skorokhod embedding problem due to @Root:69. In this paper, we provide a construction, and a proof of optimality of the upper bound, using results of @Rost:71 and @Chacon:85, and show how this proof can be used to determine a super-hedging strategy which is model-independent. In addition, we outline how the hedging strategy may be computed numerically. Using these methods, we also show that the Heston-Nandi model is ‘asymptotically extreme’ in the sense that, for large maturities, the Heston-Nandi model gives prices for variance call options which are approximately the lowest values consistent with the same call price data.' author: - 'Alexander M. G. Cox[^1]' - 'Jiajie Wang[^2]' bibliography: - 'VarOptionsNew.bib' title: Optimal robust bounds for variance options --- Introduction {#sec:Introduction} ============ The classical approach to derivative pricing problems is to hypothesise a certain model for the underlying asset, and to invoke the Fundamental Theorem of Asset Pricing to identify arbitrage-free prices with discounted expectations under risk-neutral measures. In this setting, market information, for example in the form of traded ‘vanilla’ options, may be incorporated by choosing a parametrised class of models, and determining the parameters by finding the best fit to the observed prices. An alternative approach to incorporating market information is to use the traded options as part of a hedging strategy. This approach can be particularly beneficial in the presence of model-risk, since, if carefully chosen, the hedging properties of the strategy may still hold under a wide class of models. The archetypal example of this is the hedging of a variance swap using a $\log$-contract due to @Dupire:93 and @Neuberger:94. Suppose a (discounted) asset price has dynamics under the risk-neutral measure: $$\label{eq:BasicModel} \frac{{\,\mathrm{d}}S_t}{S_t} = \sigma_t \, {\,\mathrm{d}}W_t,$$ where the process $\sigma_t$ is not necessarily known. A variance swap is a contract where the payoff depends on the realised quadratic variation of the log-price process: $\langle \ln S\rangle_T = \int_0^T \sigma^2_t \, {\mathrm{d}t}$. Dupire and Neuberger observed that $${\,\mathrm{d}}(\ln S_t) = \sigma_t \, {\,\mathrm{d}}W_t - {\dfrac{1}{2}}\sigma_t^2 \, {\mathrm{d}t}$$ from which we conclude that $$\int_0^T \sigma^2_t \, {\mathrm{d}t}= 2\ln(S_0) - 2\ln(S_T) + 2 \int \frac{1}{S_t} \, {\,\mathrm{d}}S_t.$$ It follows that one can replicate (up to a constant) the payoff of a variance swap by shorting two log-contracts (that is, contracts which pay $\ln(S_T)$ at maturity), and dynamically trading in the asset so as to always hold $2/S_t$ units of the asset. The resulting portfolio will hedge the variance swap under essentially [*any*]{} model of the form (subject to very mild measurability and integrability conditions on $\sigma_t$). It follows that the price of a variance swap is essentially determined once one observes the price of a log-contract. In practice, call options are more liquidly traded, but the log contract can be statically replicated given a continuum of call options, and so this information is more commonly used to determine the price. In this paper, we consider what can be said about the price of options which pay the holder a function of the realised variance in the presence of co-terminal call options. Two important examples of these contracts are the variance call, which has payoff $\left(\langle \ln S\rangle_T -K\right)_+$ and the volatility swap, which has payoff $\sqrt{\langle \ln S\rangle_T} - K$. Unlike in the case of the variance swap, it is no longer possible to provide a trading strategy which exactly replicates the payoff in any model, however, we are able to provide both super- and sub-hedging strategies which work for a large class of models, and therefore provide model-independent bounds on the prices of such options. Our methods rely on techniques from the theory of Skorokhod embeddings, and in particular, we need a novel proof of optimality of some existing constructions. In one direction, the bounds have been established in a preceding paper, @CoxWang:11, and some of the methods used in this paper for the other direction follow similar approaches, although there remain substantial technical differences. In addition to establishing theoretical bounds on the prices of these options, we provide some numerical investigation of the bounds we obtain, and a justification of our numerical techniques. Finally, we are also able to establish, using the optimality techniques of [@CoxWang:11], a result relating to the extremality of the Heston-Nandi model (the classical stochastic volatility model of Heston, where the volatility and asset processes are perfectly anti-correlated). Essentially, we show that for large maturities, under the Heston-Nandi model, the price of a variance call option will closely approximate the lowest price possible in the class of models which produce identical call prices. Given that the Heston model is commonly used for pricing options on variance, and that calibration can often lead to values of the correlation parameter close to $-1$, this suggests that using this model for pricing variance options amounts to taking a strong ‘bet’ on which model most accurately reflects reality. The theme of model-independent, or robust, pricing is one that has received a great deal of attention in recent years. The approach we take in this paper can be traced back to @Hobson:98, and more recent work, closely related to variance options, includes @Dupire:05 [@CarrLee:10; @DavisOblojRaval:10; @CoxWang:11] and @OberhauserDosReis:13. In addition, @HobsonKlimmek:11 consider variance swaps where the model may include jumps — a case which we exclude. An alternative, related approach is based on the [*uncertain volatility*]{} models of @Avellaneda:1995aa. Recent papers which take this approach include @Galichon:2011aa [@possamai_robust_2013] and @Neufeld:2013aa. We explain how our results may be interpreted in this framework in Remark \[rmk:uncertainvolatility\]. Other recent connected work in this direction includes @Beiglbock:2013aa and @Dolinsky:2013aa, where connections with optimal transport are established. We proceed as follows: in Section \[sec:financial-motivation\] we introduce our financial setup, and explain why the financial problem of interest can be related to the Skorokhod embedding problem. This motivates Section \[sec:constr-rosts-barr\], where we provide a characterisation of the solution of @Rost:71 and @Chacon:85 to the Skorokhod embedding problem. In Section \[sec:optRost\] we give a novel proof of the optimality of these barriers, and explain how these constructions may be used to derive superhedges for certain options. In Section \[sec:numerical-results\] we show how the solutions may be computed numerically, and provide some graphical evidence of the behaviour of the hedging strategies. In Section \[sec:extr-hest-nandi\] we prove our optimality result for the Heston-Nandi model. Financial Motivation {#sec:financial-motivation} ==================== To motivate our financial models, we begin with a fairly classical setup: we suppose that there is a market which consists of a traded asset, with price $S_t$ defined on a probability space $\left(\Omega, {\mathcal{F}}, \left( {\mathcal{F}}_t\right)_{t \ge 0}, {\mathbb{P}}\right)$, satisfying the usual conditions, with: $$\label{eq:Sdefn} \frac{dS_t}{S_t} = r_t {\,\mathrm{d}}t + \sigma_t {\,\mathrm{d}}W_t$$ under some probability measure ${\mathbb{Q}}\sim {\mathbb{P}}$, where ${\mathbb{P}}$ is the objective probability measure, and $W_t$ a ${\mathbb{Q}}$-Brownian motion. In addition, we suppose $r_t$ is the risk-free rate which we require to be known, but which need not be constant. In particular, let $r_t, \sigma_t$ be locally bounded, progressively measurable processes so that the integral in is well defined, and so $S_t$ is an Itô process. We suppose that the process $\sigma_t$ is not known (or more specifically, we aim to produce conclusions which hold for all $\sigma_t$ in the class described). Specifically, we shall, at least initially, suppose: \[ass:Price\] The asset price process, under some probability measure ${\mathbb{Q}}\sim {\mathbb{P}}$, is the solution to , where $r_t$ and $\sigma_t$ are locally bounded, predictable processes. We shall later see that some relaxation of this condition is possible, using pathwise approaches to stochastic integration. In addition, we need to make the following assumptions regarding the set of call options which are initially traded: \[ass:Calls\] We suppose that call options with maturity $T$, and at all strikes $\left\{K: K \ge 0\right\}$ are traded at time $0$, and the prices, $C(K)$, are assumed to be known. In addition, we suppose call-put parity holds, so that the price of a put option with strike $K$ is $P(K) = e^{-\int_0^T r_s {\,\mathrm{d}}s}K -S_0 + C(K)$. We make the additional assumptions that $C(K)$ is a continuous, decreasing and convex function, with $C(0)= S_0$, $C_+'(0) = -e^{-\int_0^T r_s {\,\mathrm{d}}s}$ and $C(K) \to 0$ as $K \to \infty$. Many of these notions can be motivated by arbitrage concerns (see [e.g.]{} @CoxObloj:11). That there are plausible situations in which these assumptions do not hold can be seen by considering models with bubbles ([e.g.]{} [@CoxHobson:05]), in which call-put parity fails, and $C(K) \not\to 0$ as $K \to \infty$. Let us define $B_t = e^{\int_0^t r_s {\,\mathrm{d}}s}$, and make the assumptions above. Since (classically) prices correspond to expectations under ${\mathbb{Q}}$, the implied law of $B_T^{-1}S_T$ (which we will denote $\mu$) can be recovered by the Breeden-Litzenberger formula [@BreedenLitzenberger:78]: $$\label{eq:BL} \mu((K,\infty)) = {\mathbb{Q}}^*(B_T^{-1}S_T \in (K,\infty)) = - B_T C_+'(B_TK).$$ Here we have used ${\mathbb{Q}}^*$ to emphasise the fact that this is only an [*implied*]{} probability, and not necessarily the distribution under the actual measure ${\mathbb{Q}}$. It can now also be seen that the assumption that $C_+'(0) = -B_T^{-1}$ is equivalent to assuming that there is no atom at 0 — [i.e.]{} $\mu((0,\infty)) = 1$. In general, the equality here could be replaced with an inequality ($C_+'(0) \ge -B_T^{-1}$) if one wished to consider models with a positive probability of the asset being worthless at time $T$. We do not impose the condition that the law of $B_T^{-1} S_T$ under ${\mathbb{Q}}$ is $\mu$, we merely note that this is the law implied by the traded options. We also do not assume anything about the price paths of the call options: our only assumptions are their initial prices, and that they return the usual payoff at maturity. Finally, it follows from the assumptions that $\mu$ is an integrable measure with mean $S_0$. Our goal is to now to use the knowledge of the call prices to find a lower or upper bound on the price of an option which has payoff $$F\left( \int_0^T \sigma_t^2 \, {\mathrm{d}t}\right) = F\left( \left< \ln S\right>_T \right).$$ The term $\int_0^T \sigma_t^2 \, {\mathrm{d}t}$ is commonly referred to as the [*realised variance*]{}. There are a number of pertinent examples which motivate us: the most common case, and where the answer is well known, is the case of a variance swap, where the payoff of the option is $(\left< \ln S\right>_T - K)$. An obvious modification of this is the [*variance call*]{} which has payoff $(\left< \ln S\right>_T - K)_+$, and the corresponding variance put. In addition, [ *volatility swaps*]{} are traded, where the payoff is $\sqrt{\left< \ln S\right>_T} - K$. As well as options written on the realised variance, there are classes of options which trade on various forms of weighted realised variance: define $$RV^w_T = \int_0^T w(S_t) \, {\mathrm{d}}\left< \ln S\right>_t = \int_0^T w(S_t) \sigma_t^2 \, {\mathrm{d}t}$$ then many of the above options can be recast in terms of their weighted versions. Common examples of these include options on corridor variance, where $w(x) = {\boldsymbol{1}_{\{\ensuremath{x \in [a,b]}\}}}$, and the gamma swap [@Lee:2010aa], where $w(x)= x$. In fact, for a simplified exposition, we will assume that the weight depends not on the spot price, but rather on the discounted spot price (or equivalently, the forward price). In the case of most interest, where $w(x) = 1$, this makes no difference, as it would, for example in an equity setting where the dividend yield and the interest rate were the same; we also refer to @Lee:2010ac, where it is indicated how such an approximation may be accounted for using a model-independent hedge involving calls of [*all*]{} maturities, although this is beyond the general methodology described in the article, where we will generally assume that only calls of one maturity are observed. Note we assume that our underlying price process, $S_t$, has continuous paths. This is an important assumption, and our conclusions will not generally hold otherwise. @HobsonKlimmek:11 consider related questions in the case where the underlying asset may jump. We also assume that the payoff of the option is exactly the realised quadratic variation, wheras in reality, financial contracts will be written on a discretised version of the quadratic variation (for example, the sum of squared daily log-returns); the effects of this approximation are also considered by, for example, @HobsonKlimmek:11. Our approach is motivated by the following heuristics. Consider the discounted stock price: $$X_t = e^{-\int_0^t r_s {\,\mathrm{d}}s} S_t = B_t^{-1} S_t.$$ Under Assumption \[ass:Price\], $X_t$ satisfies the SDE: $${\mathrm{d}}X_t = X_t \sigma_t {\,\mathrm{d}}W_t.$$ Let $\lambda(x)$ be a strictly positive, continuous function, and define a time change $\tau_t = \int_0^t \lambda(X_s) \sigma_s^2 {\,\mathrm{d}}s$. Writing $A_t$ for the right-continuous inverse, so that $\tau_{A_t} = t$, we note that ${\widetilde}{W}_t = \int_0^{A_t} \sigma_s \lambda(X_s)^{1/2} {\,\mathrm{d}}W_s$ is a Brownian motion with respect to the filtration ${\widetilde}{{\mathcal{F}}}_t = {\mathcal{F}}_{A_t}$, and if we set ${\widetilde}{X}_t = X_{A_t}$, we have: $$\label{eq:Xdiffusiondef} {\mathrm{d}}{\widetilde}{X}_t = {\widetilde}{X}_t \lambda({\widetilde}{X}_t)^{-1/2} {\,\mathrm{d}}{\widetilde}{W}_t.$$ In particular, under mild assumptions on $\lambda$, ${\widetilde}{X}_t$ is now a diffusion on natural scale, and we note also that ${\widetilde}{X}_0 = S_0$ and ${\widetilde}{X}_{\tau_T} = X_T = B_T^{-1} S_T$. It follows that $({\widetilde}{X}_{\tau_T},\tau_T) = (B_T^{-1} S_T,\int_0^T \lambda(X_s) \sigma_s^2 \, {\mathrm{d}s})$, and therefore that , which implies knowledge of the law of $B_T^{-1} S_T$, also tells us the law of ${\widetilde}{X}_{\tau_T}$. The key observation is that there is now a correspondence between the possible joint laws of a stopped diffusion and its stopping time, and the joint laws of the (discounted) asset price at a fixed time, and the weighted realised variance at that time. Since we wish to find the extremal possible prices of options whose payoff is $F\left(\int_0^T w(X_t)\sigma^2_t \, {\mathrm{d}t}\right)$, if we take $\lambda(x) \equiv w(x)$, the problem would appear to be equivalent to that of finding a stopping time which maximises or minimises ${\mathbb{E}}F(\tau)$ subject to $\mathcal{L}({\widetilde}{X}_\tau) = \mu$, where $\mu$ is the law of $B_T^{-1}S_T$ inferred by the market call prices. The general problem of finding a stopping time for a process which has a given distribution is known as the Skorokhod Embedding problem, and solutions with given optimality properties have been well studied in recent years [@AzemaYor:79; @Hobson:98; @CoxHobsonObloj:08; @CoxObloj:11b] — for a survey of these results, we refer the reader to @HobsonSurvey. In @CoxWang:11, we established that a construction of a Skorokhod embedding due to @Root:69 corresponded to minimising payoffs of the above form where $F(\cdot)$ is a convex increasing function. In this paper, we show that a related construction, which can be traced back to work of @Rost:71, and @Chacon:85, maximises such payoffs. In addition, we shall show how these constructions may be used to derive trading strategies, which will super- or sub-hedge in any of the models under consideration, and are a hedge in the extremal model. Finally, we will also briefly consider a similar problem where the option is [*forward starting*]{} — so the payoff depends on the realised variance accumulated between future dates $T_0$ and $T_1$. In this case, we assume that we observe call prices at times $T_0$ and $T_1$, which correspond to distributions at times $T_0$ and $T_1$. The implied distribution at time $T_0$, $\nu$ say, can be interpreted as the law of $X_{T_0}$. Taking ${\widetilde}{X}_{t} = X_{T_0+\tau_t}$, we get a problem similar to above, but with ${\widetilde}{X}_{0} \sim \nu$ for some new probability measure $\nu$. This setting will be considered further in Remark \[rmk:ForwardStart\]. Construction of Rost’s Barrier {#sec:constr-rosts-barr} ============================== Background ---------- In this section, we recall some important results due to Rost and Chacon concerning the construction of solutions to the Skorokhod embedding problem. These results are established under fairly general assumptions on the underlying process ${\widetilde}{X}_t$. One of the main goals of this section is to provide conditions under which we can apply their results in our setting. We begin by recalling the notion of a reversed barrier. \[defn:rostbarrier\] A closed subset $B$ of $[-\infty,+\infty]\times[0,+\infty]$ is a *reversed barrier* if 1. \[rostbarrier1\] $(x,0)\in B$ for all $x\in[-\infty,+\infty]$; 2. \[rostbarrier2\] $(\pm\infty,t)\in B$ for all $t\in[0,\infty]$; 3. \[rostbarrier3\] if $(x,t)\in B$ then $(x,s)\in B$ whenever $s<t$. Given a reversed barrier, we can construct a stopping time of a process ${\widetilde}{X}_t$ as $\tau = \inf\{t > 0: ({\widetilde}{X}_t,t) \in B\}$. Then it is known that, given a measure $\mu$ satisfying certain conditions, there exists a reversed barrier which embeds the law $\mu$, that is, such that ${\widetilde}{X}_\tau \sim \mu$. Moreover, in the case where ${\widetilde}{X}_t$ is a diffusion, the reversed barrier has the property that the corresponding stopping time minimises the [ *capped expectation*]{} ${\mathbb{E}}\left[ \tau \wedge t\right]$ over solutions to the Skorokhod embedding problem, for all $t \ge 0$. Using the observation that $$F(\tau) = F(0) + \tau F'(0) + \int_0^\infty F''(t) \left( \tau -t \right)_+ \, {\mathrm{d}t}$$ and $(\tau-t)_+ = \tau - \tau \wedge t$, and the fact that ${\mathbb{E}}\sigma$ depends only on the law of $X_{\sigma}$ for ‘nice’ stopping times (a point we will elaborate on shortly), then it is immediate that ${\mathbb{E}}F(\tau)$ is maximised over such stopping times for all convex functions with $F'(0) = 0$. These observations are essentially due to work of Rost and Chacon. In @Rost:71, the notion of a [*filling scheme*]{} stopping time was introduced for a general Markov process, and this was shown to embed and have the optimality property described. Later, @Chacon:85[^3] proved that in many cases, the filling scheme stopping time would actually be almost-surely equal to a stopping time generated by a reversed barrier. More recent results concerning these constructions can be found in [@CoxPeskir:12]. Before proving our results characterising the reversed barriers, we recall some important background. Given a probability distribution $\mu$ and a Markov process ${\widetilde}{X}$, the Skorokhod embedding problem is to find a stopping time $\tau$ such that ${\widetilde}{X}_\tau\sim\mu$. Motivated by the financial setting, we consider the case that $$\label{eq:positivemu} \mu\text{ is a probability distribution with } \mu((0,\infty))=1,$$ and ${\widetilde}{X}$ is a regular (see @RogersWilliams:00b for terminology relating to one-dimensional diffusions) diffusion on $I = (0,\infty)$, which is a solution to , with initial distribution ${\widetilde}{X}_0 \sim \nu$, for some given distribution $\nu$, and a continuous function $\lambda$ on $I$ which is strictly positive. Since $0 \not\in I$, $0$ is inaccessible for ${\widetilde}{X}_t$. Recall that we wish to include the case of forward-starting options, in which case ${\widetilde}{X}_0$ is assumed to have the law inferred from the call prices at the start date of the contract. In the theory of Skorokhod embeddings, it is usually natural to restrict to the class of [*minimal*]{} stopping times, however since ${\widetilde}{X}_t$ is transient, any embedding will be minimal. Moreover, for example by considering ${\widetilde}{X}_t$ as a time change of a Brownian motion stopped on hitting $0$, if we restrict ourselves to laws $\mu$ and $\nu$ which have the same mean, then any embedding of $\mu$ must have $({\widetilde}{X}_{t \wedge \tau})_{t \ge 0}$ a uniformly integrable (UI) process, and moreover a necessary and sufficient condition for the existence of an embedding is that $$\label{eq:potcond} {\mathrm{U}_{\nu}}(x) := - \int_{{\mathbb{R}}} |y-x| \, \nu({\mathrm{d}}y) \ge - \int_{{\mathbb{R}}} |y-x| \, \mu({\mathrm{d}}y) =: {\mathrm{U}_{\mu}}(x) > -\infty,$$ for all $x\in{\mathbb{R}}$. By Jensen’s inequality, such a constraint is clearly necessary for the existence of an embedding; further, using time-change arguments, and reducing to the Brownian case, it is the only restriction that is required. To understand this notion in the financial setting, note that from we deduce that ${\mathrm{U}_{\mu}}(x) = S_0 - 2C(B_Tx) - x$, giving an affine mapping between the function ${\mathrm{U}_{\mu}}(x)$ and the call prices. For any given reversed barrier, we define $D:=({\mathbb{R}}\times I)\backslash B$, and then one can show that there exists a unique upper semi-continuous function $R:I\rightarrow[0,\infty]$ such that $$D = \{(x,t): t> R(x)\}\ \ \text{ and }\ \ B = \{(x,t): 0 \leq t\leq R(x)\}.$$ Then the stopping time of interest is: $$\tau_D = \inf\{t>0:\,({\widetilde}{X}_t,t)\notin D\} = \inf\{t>0: t \leq R({\widetilde}{X}_t)\},$$ and we will call such an embedding a [*Rost stopping time*]{}, or a [*Rost embedding*]{}. We may also refer to $D$ as a reversed barrier, with the meaning intended to be inferred from the notation. Note that multiple barriers may solve the same stopping problem: for example, if the target distribution contains atoms, the barrier between atoms may be unspecified away from the starting point, provided it is never beyond the ‘spikes’ of the atoms. We can put together the work of Rost and Chacon in the specific case that the underlying process is a regular diffusion on $I$, satisfying a certain regularity condition needed to ensure Chacon’s result holds. Specifically, we introduce the set: $$\label{eq:LambdaCond} \mathcal{D} = \left\{ \lambda \in C(I;{\mathbb{R}}) : \parbox{8cm}{$\lambda$ is strictly positive and the solution to \eqref{eq:Xdiffusiondef} defines a regular diffusion on $I$, with transition density $p(t,x,y)$ with respect to Lebesgue such that, for any $x_0 \in I$, $c>0$, open set $A$ containing $x_0$ and ${\varepsilon}>0$, there exists $\delta >0$ such that $|\left(p(x,x_0,t)-p(x,x_0,s)\right)x_0^2 \lambda(x_0)^{-1}|<{\varepsilon}$ whenever $|s-t| < \delta$ and either $x_0 \not\in A$ or $t>c$.}\right\}$$ Then we can prove the following result: \[thm:RostChacon\] Suppose $\mu$ and $\nu$ are probability measures on $I$ such that holds, and suppose ${\widetilde}{X}_t$ solves for some $\lambda \in \mathcal{D}$ with ${\widetilde}{X}_0 \sim \nu$: 1. if $\mu$ and $\nu$ have disjoint support, then there exists a reversed barrier $D$ such that ${\widetilde}{X}_{\tau_D} \sim \mu$; 2. if $\mu$ and $\nu$ do not have disjoint support, then (on a possibly enlarged probability space) there exists a random variable $S \in \{0,\infty\}$, and reversed barrier $D$, such that ${\widetilde}{X}_{\tau_D \wedge S} \sim \mu$. Moreover, in both cases, the resulting embedding maximises ${\mathbb{E}}F(\sigma)$ over all stopping times $\sigma$ with ${\widetilde}{X}_\sigma \sim \mu$ and ${\mathbb{E}}\sigma = {\mathbb{E}}\tau_D < \infty$, for any convex function $F$ on $[0,\infty)$. That the condition ${\mathbb{E}}\sigma = {\mathbb{E}}\tau_D$ is reasonable can be seen by considering $$\label{eq:Qvers1} Q({\widetilde}{X}_t) := \int_{x_0}^{{\widetilde}{X}_t} \int_{x_0}^y \lambda(z) z^{-2} \, {\mathrm{d}z}{\mathrm{d}y}= \int_0^t \int_{x_0}^{x_s} \lambda(y) y^{-2} \, {\mathrm{d}y}{\mathrm{d}}{\widetilde}{X}_s + {\dfrac{1}{2}}t + Q({\widetilde}{X}_0).$$ Noting that $Q(x)$ is convex, we can take expectations along a localising sequence and apply Fatou’s Lemma to see that, for any stopping time $\tau$, ${\mathbb{E}}\tau \ge {\mathbb{E}}Q({\widetilde}{X}_\tau)$, the second term depending only on the law of ${\widetilde}{X}_\tau$. For well behaved stopping times (specifically, where ${\mathbb{E}}Q({\widetilde}{X}_{\tau\wedge N}) \to {\mathbb{E}}Q({\widetilde}{X}_{\tau})$), we would in fact expect equality here. We also observe that there is a trivial extension of this result when $F(\cdot)$ is a concave function, then $-F(\cdot)$ is a convex function, and so the resulting stopping time minimises ${\mathbb{E}}F(\sigma)$ over the same class of stopping times. We first show that, under the conditions above, there exists a filling scheme stopping time. Standard time-change arguments, and reduction to the Brownian case show that when $\mu$ and $\nu$ have the same mean, then is both necessary and sufficient for the existence of a Skorokhod embedding which solves the problem. In addition, by @Rost:71 [Theorem 4], this is sufficient to deduce the existence of a filling scheme stopping time: from the proof of this result, it is clear that whenever an embedding exists, it can be taken as a filling scheme stopping time. Now consider the case where $\mu$ and $\nu$ have disjoint support. Then @Chacon:85 [Theorem 3.24] states that a filling scheme stopping time is a reversed barrier stopping time provided: 1. ${\widetilde}{X}_t$ is a standard Markov process, in duality with a standard Markov process $\hat{X}_t$ (we refer the reader to [e.g.]{} @BlumenthalGetoor:68 [Definition VI.1.2]); 2. the transition measures relating to ${\widetilde}{X}_t$ and $\hat{X}_t$ have densities; 3. \[item:1\] the transition density $p(x,y,t)$ for ${\widetilde}{X}_t$ satisfies an equicontinuity property: for any $x_0 \in {\mathbb{R}}$, $c>0$ and open set $A$ containing $x_0$, then given ${\varepsilon}>0$, there exists $\delta >0$ such that $|p(x,x_0,t)-p(x,x_0,s)|<{\varepsilon}$ whenever $|s-t| < \delta$ and either $x_0 \not\in A$ or $t>c$. Since ${\widetilde}{X}_t$ is a regular diffusion with inaccessible endpoints, it is its own dual process, with respect to the speed measure (which is $\lambda(x) x^{-2} {\mathrm{d}x}$), see @Fitzsimmons:98 [Remark 1.15]. Moreover, under these conditions, the transition density exists ([e.g.]{} @RogersWilliams:00b [Theorem V.50.11]), and guarantees that \[item:1\] holds. Finally, suppose $\mu$ and $\nu$ are not disjointly supported. Then, using the Hahn-Jordan decomposition, we can find disjoint, non-negative measures $\nu_0$ and $\mu_0$, such that $\nu_0(A) \le \nu(A)$, $\mu_0(A) \le \mu(A)$, for all $A \in \mathcal{B}(0,\infty)$, and $\mu-\nu = \mu_0-\nu_0$. Since $\mu$ and $\nu$ are not disjoint, $\mu_0$ and $\nu_0$ are non-trivial. We write in addition $\nu \wedge \mu = \nu - \nu_0$, observing that also then $\nu \wedge \mu = \mu - \mu_0$. By enlarging the probability space if necessary, let $Z$ be a uniform random variable on $(0,1)$, independent of the process, and define the Radon-Nikodym derivative $f = {\,\mathrm{d}}(\nu \wedge \mu)/ {\,\mathrm{d}}\nu$. Then if we set $$S =\begin{cases} 0,\ &\text{if }\ Z\leq f({\widetilde}{X}_0);\\ \infty,\ &\text{if }\ Z> f({\widetilde}{X}_0), \end{cases}$$ it follows that ${\mathbb{P}}(X \in A, S = 0) = (\nu \wedge \mu) (A)$. Now define the normalised measures, $\nu^*(A) = \nu_0(A) / \nu_0((0,\infty))$ and $\mu^*(A) = \mu_0(A) / \mu_0((0,\infty))$, and construct the reversed barriers for the initial distribution $\nu^*$ and target distribution $\mu^*$. (It is straightforward to check that $U_{\nu_0}(x) - U_{\mu_0}(x) = U_{\nu}(x) - U_{\mu}(x)$, and therefore that the construction is possible.) However, it is now clear that this embeds and is exactly the stopping time described in the statement of the theorem. Moreover, this description of the first step of the construction is exactly the first step described in the construction of the filling scheme by Rost, so it follows that this stopping time is a filling scheme stopping time. The final statement is @Chacon:85 [Proposition 2.2]. It is clear that the above conditions include the main case of interest — the case where $\lambda(x) = 1$, which is the case corresponding to options on realised variance. For the point $0$ to be inaccessible, we require $\int_{0+} \lambda(x) x^{-1} \, {\mathrm{d}x}= \infty$ [@RogersWilliams:00b Theorem V.51.2]. In principle, this would exclude, for example, the case where $\lambda(x) = x$, the Gamma swap, however in practice, this case could be approximated by taking $\lambda(x) = x \vee {\varepsilon}$ for some small ${\varepsilon}>0$ . Note however that in the case where $\lambda(x) = x$, ${\widetilde}{X}_t$ is a Bessel process of dimension $0$, and so in particular, the process will hit zero with positive probability; this is not crucial, since we assume our target measure has no atom of mass at $0$, and so we would expect the reversed barrier to stop the process before hitting zero, and therefore the exact behaviour near zero should not affect the barrier substantially. Construction of reversed barriers --------------------------------- In this section, we show how the reversed barrier determined in Theorem \[thm:RostChacon\] can be constructed. As above, we suppose that we have a time-homogeneous diffusion, ${\widetilde}{X}_t$, such that: $$\label{eq:Xdefn} \begin{split} {\mathrm{d}}{\widetilde}{X}_t & = \sigma({\widetilde}{X}_t) {\mathrm{d}}W_t,\\ {\widetilde}{X}_0 & \sim \nu, \end{split}$$ so $\sigma(x) = x \lambda(x)^{-1/2}$. In addition, we suppose that the diffusion coefficient, $\sigma: I \to (0,\infty)$ is a continuously differentiable function such that: $$\label{eq:SigmaCond} x^2 \sigma(x)^{-2} \in \mathcal{D}, \quad |\sigma(x) x^{-1}| \text{ and } |\sigma'(x) \sigma(x) x^{-1}| \text{ are bounded on } (0,\infty),$$ or equivalently, that $\lambda: I \to (0,\infty)$ is continuously differentiable, and $$\label{eq:LambdaCond2} \lambda(x) \in \mathcal{D}, \quad |\lambda(x)^{-1}| \text{ and } |\lambda'(x) \lambda(x)^{-2} x| \text{ are bounded on } (0,\infty).$$ Then our general problem is: -------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{}: Find an upper-semicontinuous function $R(x)$ such that the domain $D= \{(x,t): t > R(x)\}$ has ${\widetilde}{X}_{\tau_D\wedge S} \sim \mu$, where ${\widetilde}{X}_t$ is given by , $\tau_D = \inf\{t>0: ({\widetilde}{X}_t,t) \not\in D\} = \inf\{t>0: t \le R({\widetilde}{X}_t)\}$, and $S \in \{0,\infty\}$ is an ${\mathcal{F}}_0$-measurable random variable such that ${\widetilde}{X}_0 \sim \nu \wedge \mu$ on $\{S=0\}$. -------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Here, the measure $\nu \wedge \mu$ is as defined in the proof of Theorem \[thm:RostChacon\]. We restrict ourselves to the case where ${\mathrm{U}_{\nu}}(x) \ge {\mathrm{U}_{\mu}}(x)$. We will also introduce the notation ${\bar{\tau}_D}= \tau_D \wedge S$. Then we have the following result: Suppose and hold. Assume $D$ solves [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{}. Then $u(x,t) = {\mathrm{U}_{\mu}}(x) + {\mathbb{E}}^\nu \left| x-{\widetilde}{X}_{t\wedge{\bar{\tau}_D}}\right|$ is the unique bounded viscosity solution to: \[eq:ViscSoln\] $$\begin{aligned} \label{eq:ViscSolnMain} {\frac{\partial u}{\partial t}}(x,t) & = \left( \frac{\sigma(x)^2}{2} {\frac{\partial ^2 u}{\partial x^2}}(x,t)\right)_+ \\ u(0,x) & = {\mathrm{U}_{\mu}}(x) - {\mathrm{U}_{\nu}}(x). \label{eq:ViscSolnBC} \end{aligned}$$ Moreover, given the solution $u$ to , a reversed barrier $D$ which solves [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{} can be recovered by $D = \{ (x,t) : u(x,t) > u(0,t)\}$. In a recent paper, @OberhauserDosReis:13 make a very similar observation, and they also use a viscosity solution approach to derive existence and uniqueness in the Rost setting (unlike [@CoxWang:11] where a variational inequality-based approach is taken). They also work in the slightly more general setting where the diffusion coefficient $\sigma$ may depend both on time and space; it seems very likely that the results should extend to this setting, but we observe that the optimality of such a construction is no longer easily determined; given that we are interested in the optimality properties of such processes, we restrict ourselves to the time-homogenous case. We wish to use standard results on viscosity solutions from @FlemingSoner:06. The equation is really a forward equation, rather than a backward equation, which is the setting in [@FlemingSoner:06]; however we can apply their results to the function $v(x,T-t) = -u(x,t)$, for a fixed $T>0$. Since $T$ is arbitrary, the extension to an infinite horizon is straightforward. We first note that since $\mu$ and $\nu$ are integrable, the function $u$ is bounded, both $-{\mathrm{U}_{\mu}}(x)$ and $h(x,t) = {\mathbb{E}}^{\nu}|x-{\widetilde}{X}_{t \wedge {\bar{\tau}_D}}|$ are convex, continuous functions, and so their second derivatives (in $x$) exist as positive measures. Moreover, $-{\mathrm{U}_{\mu}}''(x) = 2 \mu({\mathrm{d}x})$, and we can also decompose $h_{xx}(x,t)/2$ into two measures $\mu_t^1$ and $\mu_t^2$ defined by: $$\begin{aligned} \int f(x) \, \mu_t^1 ({\mathrm{d}x}) & = {\mathbb{E}}^\nu \left[f({\widetilde}{X}_t); t < {\bar{\tau}_D}\right]\\ \int f(x) \, \mu_t^2 ({\mathrm{d}x}) & = {\mathbb{E}}^\nu \left[f({\widetilde}{X}_{{\bar{\tau}_D}}); t \ge {\bar{\tau}_D}\right]. \end{aligned}$$ In particular, since ${\bar{\tau}_D}$ embeds $\mu$, we must have $\mu_t^2(A) \le \mu(A)$ for all $A$, and $\mu_t^2(A) = \mu(A)$ for all $A \subseteq \{x | t \ge R(x)\}$. In addition, $\mu_t^1$ will be dominated by the transition density of a diffusion (which exists by ) started with distribution $\nu$, and so will have a density $f^1(x,t)$ with respect to Lebesgue for all $t>0$, and since $D$ is open, by a slight modification of @CoxWang:11 [Lemma 3.3], it is easily checked that $u(x,t)$ satisfies ${\frac{\partial u}{\partial t}}(x,t) =\frac{\sigma(x)^2}{2} {\frac{\partial ^2 u}{\partial x^2}}(x,t)$ in $D$. We observe also that $f^1(x_n,t_n) \to 0$ whenever $(x_n,t_n) \to (x,t) \in B$, since $f^1$ is dominated by the density of a diffusion with initial law $\nu$, killed if it hits $x$ before $t$, which also has this property. The result now follows from @FlemingSoner:06 [Proposition V.4.1], when we observe that for $t\le R(x)$, any smooth function $w$ such that $w-u$ is a local minimum at $(x,t)$ must have ${\frac{\partial w}{\partial t}} = 0$. For $t < R(x)$ this follows from observing that ${\mathbb{E}}^\nu \left| x-{\widetilde}{X}_{t\wedge{\bar{\tau}_D}}\right|$ is constant in $t$ whenever $t<R(x)$. For $t=R(x)$, we observe that $f^1(x,t)$ can be made arbitrarily small in a ball near $(x,t)$, and ${\frac{\partial ^2u}{\partial x^2}} = f^1(x,t)$ in $D$. Hence, for such $w$, we have $-w_t + \left(\sigma^2(x) w_{xx}\right)_+ \ge 0$. For smooth $w$ such that $w-u$ has a local maximum at $(x,t)$, we first observe that for $t \le R(x)$, the argument above implies that $u_{xx}(x,t) \le 0$, and so $w_{xx}(x,t) \le 0$ also. In addition, by Jensen’s inequality, $u(x,t)$ is non-decreasing, so $w_{t}(x,t) \ge 0$. It follows that $-w_t + \left(\sigma^2(x) w_{xx}\right)_+ \le 0$. It follows that $u$ is indeed a bounded viscosity solution. To see that it is unique, we apply @FlemingSoner:06 [Theorem V.9.1] to $u(t,{\mathrm{e}}^y)$ (noting also the comment immediately preceding this proof). It is now routine to check that is sufficient to ensure that there is a unique bounded viscosity solution to . To conclude that a reversed barrier can be recovered from a solution $u$ to , we observe that Theorem \[thm:RostChacon\] guarantees the existence of a reversed barrier $D$, and by Jensen’s inequality, we see that $D^* = \{(x,t): u(x,t) > u(t,0)\}$ does indeed define a reversed barrier. From the arguments above, we also conclude that ${{\mathbb{P}}}({\bar{\tau}_D}= \bar{\tau}_{D^*}) = 1$, since ${\widetilde}{X}_{{\bar{\tau}_D}}$ is supported on the set where $u(x,t) = u(x,0)$, and $f(x,t) = 0$ on this set also. We observe also that the solution $u(x,t)$ has an interpretation in terms of an optimal stopping problem [c.f. @CoxWang:11 Remark 4.4]. Fix $T>0$ and set $$v(x,t) = \sup_{\tau \in [t,T]} {\mathbb{E}}^{(x,t)}\left[{\mathrm{U}_{\mu}}({\widetilde}{X}_\tau) - {\mathrm{U}_{\nu}}({\widetilde}{X}_{\tau})\right],$$ where the supremum is taken over stopping times $\tau$, and the expectation is taken conditional on ${\widetilde}{X}_t = x$. Then standard results for optimal stopping problems suggest that $v(x,t)$ is the solution to the viscosity equation: $$\label{eq:OptStopVisc} \begin{split} & \max\left\{{\dfrac{1}{2}}\sigma(x)^2 {\frac{\partial ^2 v}{\partial x^2}}(x,t) + {\frac{\partial v}{\partial t}}(x,t), v(x,t)-({\mathrm{U}_{\mu}}(x) - {\mathrm{U}_{\nu}}(x))\right\} = 0,\\ & v(x,T) = {\mathrm{U}_{\mu}}(x) - {\mathrm{U}_{\nu}}(x). \end{split}$$ Now observe from the problem formulation that $v(x,t)$ is certainly decreasing in $t$ (a stopping time which is feasible for $t_1$ is also feasible for $t_0<t_1$, with the same reward). Using the fact that both the solution to and are monotone in $t$, it is possible to deduce that $v(x,T-t)$ solves , and $u(x,T-t)$ solves , so that they must be the same function. Optimality of Rost’s Barrier, and superhedging strategies {#sec:optRost} ========================================================= Optimality via pathwise inequalities {#sec:optim-via-pathw} ------------------------------------ For a given distribution $\mu$, Theorem \[thm:RostChacon\] says that Rost’s solution is the “maximal variance” embedding. A slight generalisation of this result leads us to consider the following problem: -------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ [[**[OPT]{}${}^*(\sigma,\nu,\mu)$**]{}]{}: Suppose and hold, and ${\widetilde}{X}$ solves . Find a stopping time $\tau$ such that ${\widetilde}{X}_\tau \sim \mu$ and, for a given increasing convex function $F$ with $F(0)=0$, $${\mathbb{E}}\big[ F(\tau)\big] \ =\ \sup_{\rho: {\widetilde}{X}_{\rho} \sim \mu} {\mathbb{E}}\big[ F(\rho)\big].$$ -------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ Our aim in this section is to find the super-replicating hedging strategy for call-type payoffs on variance options, however it is not immediately obvious how to recover such an identity directly from the proofs of the optimality criterion given in @Chacon:85. Rather, we shall provide a ‘pathwise inequality’ which encodes the optimality in the sense that we can find a supermartingale $G_t$, and a function $H(x)$ such that $$\label{eq:pathwiseineq} F(t) \le G_t + H({\widetilde}{X}_t)$$ and such that, for Rost’s embedding ${\bar{\tau}_D}$, equality holds in and $G_{t \wedge {\bar{\tau}_D}}$ is a UI martingale. It then follows that ${\bar{\tau}_D}$ does indeed maximise ${\mathbb{E}}F(\tau)$ among all solutions to the Skorokhod embedding problem, and further, using , we can super-replicate call-type payoffs on variance options by a dynamic trading strategy. Suppose that we have already found the solution to [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{}, ${\bar{\tau}_D}$. Define the function $$\label{eq:Mdefn} M(x,t) = {\mathbb{E}}^{(x,t)}\left[f({\bar{\tau}_D})\right],$$ where $f$ is the left derivative of $F$ and ${\bar{\tau}_D}$ is the corresponding hitting time of $B$. Specifically, observe that ${\widetilde}{X}_t$ is a Markov process, and we interpret the expectation as the average of $f({\bar{\tau}_D})$ given that we start at ${\widetilde}{X}_t = x$, with ${\bar{\tau}_D}\ge t$. At zero, given the possibility of stopping at time $0$, it is not immediately clear how to interpret the conditioning — it will turn out not to matter, but a natural choice would be to replace ${\bar{\tau}_D}$ with $\tau_D$. In the following, we shall assume: $$\label{eq:Massumption} M(x, t)\text{ is locally bounded on }{\mathbb{R}}\times {\mathbb{R}}_+.$$ Obviously, $M(x,t)=f(t)$ whenever $0\leq t<R(x)$. Now given a fixed time $T>0$, and choosing $S_0^* \in (\inf \operatorname{supp}(\mu),\sup \operatorname{supp}(\mu))$ (which is non-empty if $\mu$ is non-trivial), we define $$Z_{T}(x)\ =\ 2\int_{S_0^*}^{x}\int_{S_0^*}^{y} \dfrac{M(z,T)}{\sigma(z)^2}{\,\mathrm{d}}z{\,\mathrm{d}}y,$$ and in particular, $Z_{T}''(x)=2M(x,T)/\sigma^2(x)$, and $Z_T$ is a convex function. Define also $$\begin{aligned} G_{T}(x,t) &= F(T)-\int_{t}^{T}M(x,s){\,\mathrm{d}}s-Z_{T}(x) \nonumber\\ H_{T}(x) &= \int_{R(x)}^{T}\Big[M(x,s)-f(s)\Big]{\,\mathrm{d}}s+Z_{T}(x) \nonumber\\ &= \int_{R(x)\wedge T}^{T}\Big[M(x,s)-f(s)\Big]{\,\mathrm{d}}s+Z_{T}(x).\nonumber\\ Q(x) & = \int_{S_0^*}^x\int_{S_0^*}^y\dfrac{2}{\sigma(z)^2}{\,\mathrm{d}}z{\,\mathrm{d}}y. \label{eq:Qdefn}\end{aligned}$$ Then we have the following results \[prop:GTineq\] For all $(x,t,T)\in{\mathbb{R}}_+\times{\mathbb{R}}_+\times{\mathbb{R}}_+$, we have, $$\label{eq:GTineq} \begin{cases} \ G_{T}(x,t)\,+\,H_{T}(x)\ \geq\ F(t),\ \ \ &\text{if }\ t>R(x)\,;\\ \ G_{T}(x,t)\,+\,H_{T}(x)\ =\ F(t),&\text{if }\ t\leq R(x)\,. \end{cases}$$ \[lem:GTmartsubmart\] Under the setting – , suppose that the stopping time ${\bar{\tau}_D}$ is the solution to [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{}. Moreover, assume $f$ is bounded and $$\label{eq:qvTbddcond} \text{for any }\ T>0,\ \ \left(Q({\widetilde}{X}_t);\,0\le t\le T\right) \ \text{ is a uniformly integrable family. }$$ Then for any $T>0$, the process $$\label{eq:GTmart} \left(G_T({\widetilde}{X}_{t\wedge{\bar{\tau}_D}},t\wedge{\bar{\tau}_D});\,0\le t\le T\right) \ \text{ is a martingale,}$$ and $$\label{eq:GTsupmart} \left(G_T({\widetilde}{X}_{t},t);\,0\le t\le T\right) \ \text{ is a supermartingale.}$$ We note that when $\sigma(x) = x$, so ${\widetilde}{X}_t$ is geometric Brownian motion, then it is straightforward to check that, for all $T>0$, $\sup_{t\le T}{\mathbb{E}}[Q({\widetilde}{X}_t)^2]<\infty$, and so is trivially satisfied provided ${\mathbb{E}}[Q({\widetilde}{X}_0)^2] < \infty$. Note also that since ${\widetilde}{X}_t$ is a local martingale bounded below, for any embedding $\tau$ which embeds $\mu$ we have ${\mathbb{E}}[{\widetilde}{X}_\tau]= {\mathbb{E}}[{\widetilde}{X}_{{\bar{\tau}_D}}]$. It follows that if ${\mathbb{E}}[{\widetilde}{X}_{{\bar{\tau}_D}}] ={\mathbb{E}}[{\widetilde}{X}_0]$, any embedding of $\mu$ is a martingale, and not just a local-martingale. Then the main result of this section follows. \[thm:optRost\] Suppose that ${\bar{\tau}_D}$ is the solution to [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{}, and holds, then ${\bar{\tau}_D}$ solves [[**[OPT]{}${}^*(\sigma,\nu,\mu)$**]{}]{}. We first consider the case where ${\mathbb{E}}[{\bar{\tau}_D}]=\infty$. Since $F(t)\geq\alpha+\beta t$ for some constants $\alpha\in{\mathbb{R}}$ and $\beta\in{\mathbb{R}}_+$, we must have ${\mathbb{E}}[F({\bar{\tau}_D})]=\infty$. The result is trivial. So we always assume ${\mathbb{E}}[{\bar{\tau}_D}]<\infty$. Under the assumption ${\mathbb{E}}[{\bar{\tau}_D}]<\infty$, consider $Q(\cdot)$ given by . We have (recall ) ${\mathbb{E}}[Q({\widetilde}{X}_{{\bar{\tau}_D}})]={\mathbb{E}}[{\bar{\tau}_D}] + {\mathbb{E}}[Q({\widetilde}{X}_0)]<\infty$. Therefore, for all $\tau$ embedding $\mu$, ${\mathbb{E}}[\tau] = {\mathbb{E}}[Q({\widetilde}{X}_\tau)] ={\mathbb{E}}[Q({\widetilde}{X}_{{\bar{\tau}_D}})]={\mathbb{E}}[{\bar{\tau}_D}]+ {\mathbb{E}}[Q({\widetilde}{X}_0)]<\infty $. In the remainder of this proof, we always assume ${\mathbb{E}}[\tau]={\mathbb{E}}[{\bar{\tau}_D}]<\infty$. We first assume $f$ is bounded, since $f$ is increasing, $$\label{eq:fassumption} \text{there exists }C<\infty,\text{ such that }\lim_{t \to \infty} f(t) = C.$$ For $T>0$, since $M(\cdot,T)$ is also bounded by $C$, then $${\mathbb{E}}[Z_T({\widetilde}{X}_{t\wedge\tau})]\leq C{\mathbb{E}}[Q({\widetilde}{X}_{t\wedge\tau})] = C ({\mathbb{E}}[t\wedge\tau]+{\mathbb{E}}[Q({\widetilde}{X}_0)])<\infty,$$ and the same argument implies ${\mathbb{E}}[Z_T({\widetilde}{X}_\tau)]<\infty$. So ${\mathbb{E}}[Z_T({\widetilde}{X}_\tau)|\mathcal{F}_t]$ is a uniformly integrable martingale, and by convexity, $Z_T({\widetilde}{X}_{t\wedge\tau})\leq{\mathbb{E}}[Z_T({\widetilde}{X}_\tau)|\mathcal{F}_t]$. Therefore, $$-C|T-(t\wedge\tau)|\ \leq\ F(T)-G_T({\widetilde}{X}_{t\wedge\tau},t\wedge\tau) \ \leq\ C|T-(t\wedge\tau)|+{\mathbb{E}}[Z_T({\widetilde}{X}_\tau)|{\mathcal{F}}_t].$$ It follows that ${\mathbb{E}}[G_T({\widetilde}{X}_{t\wedge\tau},t\wedge\tau)] \to{\mathbb{E}}[G_T({\widetilde}{X}_{\tau},\tau)]$ as $t\to\infty$. On the other hand, $${\mathbb{E}}\left[ H_T({\widetilde}{X}_{\tau})\right] \ =\ {\mathbb{E}}\left[\int^{T}_{T\wedge R({\widetilde}{X}_{\tau})}\Big[ M({\widetilde}{X}_{\tau},s)-f(s)\Big]{\,\mathrm{d}}s\right] + {\mathbb{E}}\left[Z_T({\widetilde}{X}_{\tau})\right]\ <\ \infty.$$ The same arguments hold when $\tau$ is replaced by ${\bar{\tau}_D}$, and then we have $${\mathbb{E}}\left[ H_T({\widetilde}{X}_{\tau})\right]\ =\ {\mathbb{E}}\left[ H_T({\widetilde}{X}_{{\bar{\tau}_D}})\right] \hspace{15pt}\text{ and }\hspace{15pt} {\mathbb{E}}\left[ Z_T({\widetilde}{X}_{\tau})\right]\ =\ {\mathbb{E}}\left[ Z_T({\widetilde}{X}_{{\bar{\tau}_D}})\right].$$ In addition, by Lemma \[lem:GTmartsubmart\], we have, $${\mathbb{E}}\left[G_T({\widetilde}{X}_{T\wedge{\bar{\tau}_D}},T\wedge{\bar{\tau}_D})\right] \geq\ {\mathbb{E}}\left[G_T({\widetilde}{X}_{T\wedge\tau},T\wedge\tau)\right].$$ Combining the results above with , we have $$\label{eq:Fineq} \begin{split} {\mathbb{E}}F(\tau)& \leq {\mathbb{E}}\big[G_T({\widetilde}{X}_\tau,\tau)+H_T({\widetilde}{X}_\tau)\big]\\ &= {\mathbb{E}}\big[G_T({\widetilde}{X}_{T\wedge\tau},T\wedge\tau) +H_T({\widetilde}{X}_\tau)\big]+{\mathbb{E}}\big[G_T({\widetilde}{X}_\tau,\tau) -G_T({\widetilde}{X}_{T\wedge\tau},T\wedge\tau)\big]\\ &\leq {\mathbb{E}}\big[G_T({\widetilde}{X}_{T\wedge{\bar{\tau}_D}},T\wedge{\bar{\tau}_D}) +H_T({\widetilde}{X}_{{\bar{\tau}_D}})\big]\\ & \hspace{30pt} {}+{\mathbb{E}}\big[G_T({\widetilde}{X}_\tau,\tau) -G_T({\widetilde}{X}_{T\wedge\tau},T\wedge\tau) \big]\\ &= {\mathbb{E}}\big[G_T({\widetilde}{X}_{{\bar{\tau}_D}},{\bar{\tau}_D})+H_T({\widetilde}{X}_{{\bar{\tau}_D}})\big] +{\mathbb{E}}\big[G_T({\widetilde}{X}_\tau,\tau)-G_T({\widetilde}{X}_{T\wedge\tau}, T\wedge\tau)\big]\\ &\hspace{30pt} {}-{\mathbb{E}}\big[G_T({\widetilde}{X}_{{\bar{\tau}_D}},{\bar{\tau}_D})- G_T({\widetilde}{X}_{T\wedge{\bar{\tau}_D}},T\wedge{\bar{\tau}_D})\big]\\ &= {\mathbb{E}}\big[F({\bar{\tau}_D})\big] +{\mathbb{E}}\left[\int_{T\wedge\tau}^{T}M({\widetilde}{X}_{T\wedge\tau},s){\,\mathrm{d}}s -\int_{\tau}^{T}M({\widetilde}{X}_{\tau},s){\,\mathrm{d}}s +Z_T({\widetilde}{X}_{T\wedge\tau})\right]\\ &\hspace{30pt} -{\mathbb{E}}\left[\int_{T\wedge{\bar{\tau}_D}}^{T}M({\widetilde}{X}_{T\wedge{\bar{\tau}_D}},s){\,\mathrm{d}}s -\int_{{\bar{\tau}_D}}^{T}M({\widetilde}{X}_{{\bar{\tau}_D}},s){\,\mathrm{d}}s +Z_T({\widetilde}{X}_{T\wedge{\bar{\tau}_D}})\right]\\ &= {\mathbb{E}}\big[F({\bar{\tau}_D})\big] +{\mathbb{E}}\left[{\boldsymbol{1}}_{[\tau>T]}\int_{T}^{\tau}M({\widetilde}{X}_{\tau},s){\,\mathrm{d}}s \right]\\ &\hspace{30pt} {}-{\mathbb{E}}\left[{\boldsymbol{1}}_{[{\bar{\tau}_D}>T]} \int_{T}^{{\bar{\tau}_D}}M({\widetilde}{X}_{{\bar{\tau}_D}},s){\,\mathrm{d}}s \right]+{\mathbb{E}}\Big[Z_T({\widetilde}{X}_{T\wedge\tau})-Z_T({\widetilde}{X}_{T\wedge{\bar{\tau}_D}})\Big]. \end{split}$$ Since $f\leq C$, we have $$\begin{split} 0 \leq {\mathbb{E}}\left[{\boldsymbol{1}}_{[\tau>T]} \int_{T}^{\tau}M({\widetilde}{X}_{\tau},s){\,\mathrm{d}}s\right] &\leq C{\mathbb{E}}\big[{\boldsymbol{1}}_{[\tau>T]}(\tau-T)\big]\\ &= C{\mathbb{E}}\big[\tau-T\wedge\tau\big]\ \longrightarrow\ 0, \ \ \ \text{ as }\ T\rightarrow\infty. \end{split}$$ Similarly, $$\lim_{T\rightarrow\infty} {\mathbb{E}}\left[{\boldsymbol{1}}_{[{\bar{\tau}_D}>T]}\int_{T}^{{\bar{\tau}_D}} M({\widetilde}{X}_{{\bar{\tau}_D}},s){\,\mathrm{d}}s\right]\ =\ 0.$$ Now, by the fact that ${\mathbb{E}}[Q({\widetilde}{X}_{T\wedge\tau})]={\mathbb{E}}[T\wedge\tau] + {\mathbb{E}}[Q({\widetilde}{X}_0)]$ and the convexity of $Q$, $Q({\widetilde}{X}_{T\wedge\tau})\le {\mathbb{E}}[Q({\widetilde}{X}_\tau)|\mathcal{F}_T]$, hence, $Q({\widetilde}{X}_{t\wedge\tau})\rightarrow Q({\widetilde}{X}_\tau)$ in $L^1$. Noting that $Z_T({\widetilde}{X}_{T\wedge\tau})\leq CQ({\widetilde}{X}_{T\wedge\tau})$ and $Z_T({\widetilde}{X}_{T\wedge\tau})\to CQ({\widetilde}{X}_{\tau})$ a.s. as $T\to\infty$, we have $$\lim_{T\to\infty}{\mathbb{E}}\left[Z_T({\widetilde}{X}_{T\wedge\tau})\right] \ =\ C{\mathbb{E}}\left[Q({\widetilde}{X}_\tau)\right]\ <\infty.$$ The same arguments hold when $\tau$ is replaced by ${\bar{\tau}_D}$, and moreover, ${\mathbb{E}}[Q({\widetilde}{X}_\tau)]={\mathbb{E}}[Q({\widetilde}{X}_{{\bar{\tau}_D}})]$. Now, let $T$ go to infinity in , and we have $${\mathbb{E}}\big[F(\tau)\big]\ \leq\ {\mathbb{E}}\big[F({\bar{\tau}_D})\big].$$ To observe that the result still holds when $f$ is unbounded, observe that we can apply the above argument to $f(t) \wedge N$, and $F_N(t) = \int_0^t f(s) \wedge N {\,\mathrm{d}}s$ to get ${\mathbb{E}}\left[ F_N({\bar{\tau}_D}) \right] \ge {\mathbb{E}}\left[ F_N(\tau)\right]$, and the conclusion follows on letting $N\to \infty$. Now we turn to the proofs of Proposition \[prop:GTineq\] and Lemma \[lem:GTmartsubmart\]. If $(x,t)\in D$, i.e. $t>R(x)$, $$\begin{split} G_{T}(x,t)+H_T(x) & =\int_{R(x)}^{t}M(x,s){\,\mathrm{d}}s +F(R(x)) \\ & \geq\int_{R(x)}^{t}f(s){\,\mathrm{d}}s+F(R(x))=F(t). \end{split}$$ If $(x,t)\notin D$, i.e. $t\leq R(x)$, $$\begin{split} G_{T}(x,t)+H_T(x) & =-\int_{t}^{R(x)}M(x,s){\,\mathrm{d}}s+F(R(x))\\ & =-\int_{t}^{R(x)}f(s){\,\mathrm{d}}s+F(R(x))=F(t). \end{split}$$ For $s\leq t\leq T$, by , the Meyer-Itô formula gives, $$Z_T({\widetilde}{X}_t)-Z_T({\widetilde}{X}_s) \ =\ \int_s^t Z'_T({\widetilde}{X}_u){\,\mathrm{d}}{\widetilde}{X}_u+\int_{s}^{t}M({\widetilde}{X}_u,T){\,\mathrm{d}}u.$$ By and the fact $f$ is bounded, it is easy to see that the family $(Z_T({\widetilde}{X}_t);0\leq t\leq T)$ is uniformly integrable. By the Doob-Meyer decomposition theorem (e.g. @KaratzasShreve:91 [Theorem 4.10, Chapter 1]), the first term on the right-hand side is a uniformly integrable martingale, $${\mathbb{E}}\big[Z_T({\widetilde}{X}_t)-Z_T({\widetilde}{X}_s)\,\big|\,\mathcal{F}_s\big] \ =\ \int_{s}^{t}{\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big|\,\mathcal{F}_s\big] {\,\mathrm{d}}u.$$ Then we have, $$\begin{split} &G_{T}({\widetilde}{X}_s,s)-{\mathbb{E}}\left[G_T({\widetilde}{X}_t,t)\big| \mathcal{F}_s\right]\\ & \hspace{10pt} {}= \int_{t}^{T}{\mathbb{E}}\big[M({\widetilde}{X}_t,u)\big| \mathcal{F}_s\big]{\,\mathrm{d}}u +\int_{s}^{t}{\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big| \,\mathcal{F}_s\big]{\,\mathrm{d}}u-\int_{s}^{T}M({\widetilde}{X}_s,u){\,\mathrm{d}}u\\ & \hspace{10pt} {}= \int_{t}^{T}{\mathbb{E}}\big[M({\widetilde}{X}_t,u)\big| \mathcal{F}_s\big]{\,\mathrm{d}}u-\int_{s}^{T-t+s}M({\widetilde}{X}_s,u){\,\mathrm{d}}u\\ &\hspace{90pt} {} +\int_{s}^{t}{\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big|\,\mathcal{F}_s\big]{\,\mathrm{d}}u -\int_{T-t+s}^{T}M({\widetilde}{X}_s,u){\,\mathrm{d}}u\\ & \hspace{10pt} {}= \int_{t}^{T}\Big\{{\mathbb{E}}\big[M({\widetilde}{X}_t,u)\,\big| \,\mathcal{F}_s\big] -M({\widetilde}{X}_{s},u-(t-s))\Big\}{\,\mathrm{d}}u\\ &\hspace{90pt} {}+\int_{s}^{t}\Big\{{\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big| \,\mathcal{F}_s\big]-M({\widetilde}{X}_{s},T-(t-u))\Big\}{\,\mathrm{d}}u. \end{split}$$ Now, observe that ${\widetilde}{X}_t$ is a Markov process, so we can write $Y_t$ for an independent copy of ${\widetilde}{X}$, and $\bar{\sigma}_D$ for the corresponding hitting time of the reversed barrier, and write ${\widetilde}{X}^x$ for ${\widetilde}{X}$ started at $x$. We have, for $u\in(t,T]$: $$\begin{aligned} {\mathbb{E}}^{(x,u-(t-s))}\big[f({\bar{\tau}_D})|\mathcal{F}_u\big] \leq & {\boldsymbol{1}}_{[{\bar{\tau}_D}\leq u]}f(u)+{\boldsymbol{1}}_{[{\bar{\tau}_D}>u]} {\mathbb{E}}^{(x,u-(t-s))}\big[f({\bar{\tau}_D})|\mathcal{F}_u\big]\nonumber\\ = & {\boldsymbol{1}}_{[{\bar{\tau}_D}\leq u]}f(u)+{\boldsymbol{1}}_{[{\bar{\tau}_D}>u]} {\mathbb{E}}^{({\widetilde}{X}_{t-s}^{x},u)}\big[f(\bar{\sigma}_D)\big] \nonumber\\ \leq & M({\widetilde}{X}_{t-s}^{x},u). \label{eq:Rprep} \end{aligned}$$ Hence, $$\label{eq:R1} \begin{split} {\mathbb{E}}\big[M({\widetilde}{X}_t,u)\,\big|\,\mathcal{F}_s\big] \ &=\ {\mathbb{E}}^{{\widetilde}{X}_s}M({\widetilde}{X}_{t-s},u)\\ &\geq\ {\mathbb{E}}^{({\widetilde}{X}_{s},u-(t-s))} \big[f({\widetilde}{\tau}_D)\big]\ =\ M({\widetilde}{X}_s,u-(t-s)). \end{split}$$ For $u\in(s,T)$, replacing $u$ by $T$ and $t$ by $u$ in gives that $${\mathbb{E}}^{(x,T-(u-s))}\big[f({\bar{\tau}_D})\,|\,\mathcal{F}_T\big] \ \leq\ M({\widetilde}{X}^{x}_{u-s}, T),$$ and hence, $$\label{eq:R2} {\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big|\,\mathcal{F}_s\big] \ \geq\ M({\widetilde}{X}_s,T-(u-s)).$$ It follows that $$\begin{split} \int_{s}^{t}\Big\{ & {\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big| \,\mathcal{F}_s\big]-M({\widetilde}{X}_{s},T-(t-u))\Big\}{\,\mathrm{d}}u\\ & = \int_{s}^{t}\Big\{{\mathbb{E}}\big[M({\widetilde}{X}_u,T)\,\big| \,\mathcal{F}_s\big]-M({\widetilde}{X}_{s},T-(u-s))\Big\}{\,\mathrm{d}}u\ \geq 0. \end{split}$$ Therefore, $$G_{T}({\widetilde}{X}_s,s)-{\mathbb{E}}\left[G_T({\widetilde}{X}_t,t)|\mathcal{F}_s\right]\ \geq\ 0,$$ which implies . On the other hand, as a part of , $${\boldsymbol{1}}_{[{\bar{\tau}_D}>u]}{\mathbb{E}}^{(x,u-(t-s))}\big[f({\bar{\tau}_D})\,| \,\mathcal{F}_u\big]\ =\ {\boldsymbol{1}}_{[{\bar{\tau}_D}>u]} {\mathbb{E}}^{({\widetilde}{X}_{t-s}^{x},u)}\big[f(\bar{\sigma}_D)\big],$$ and on $\{u<{\bar{\tau}_D}\}$ equality holds in the inequalities and . Thus, follows. For bounded $f$, although the pathwise inequality in this section $G_T({\widetilde}{X}_t,t)+H_T({\widetilde}{X}_t)$ $\geq F(t)$ holds for all $T,t>0$, $G_T({\widetilde}{X}_t,t)$ is a supermartingale only on $[0,T]$. For hedging purposes, we would really like to know: can we find a *global* pathwise inequality $G^\ast_t+H^\ast({\widetilde}{X}_t)\geq F(t)$, such that $G^\ast_t$ is a supermartingale on $[0,\infty]$ and a martingale on $[0,{\bar{\tau}_D}]$? We now provide conditions where we can find such $G^\ast$ and $H^\ast$. We replace by a stronger assumption: there exists some $\alpha>1$, such that $$\label{eq:sfassumption} \text{for $t$ sufficiently large, } \ C\ \geq\ f(t)\ \geq\ C-O(t^{-\alpha}).$$ Under this assumption, it is easy to check there exists a $J(x,t)$ such that $$\label{eq:Jdefn} J(x,t)\ =\ \lim_{T\to\infty}\int_t^T \left[M(x,s)-f(s)\right]{\,\mathrm{d}}s,$$ then we define $$\label{eq:GHlim} \left\{ \begin{aligned} G(x,t)\ &=\ \lim_{T\to\infty}G_T(x,t) \ =\ F(t)-J(x,t)-CQ(x);\\ H(x)\ &=\ \lim_{T\to\infty}H_T(x)\ =\ J(x,R(x))+CQ(x). \end{aligned}\right.$$ Letting $T\to\infty$ in , $$\label{eq:sGTineq} \begin{cases} \ G(x,t)+H(x)\ >\ F(t),\ \ \ &\text{if }\ t>R(x);\\ \ G(x,t)+H(x)\ =\ F(t),&\text{if }\ t\leq R(x). \end{cases}$$ By the monotone convergence theorem, for all $t>0$, ${\mathbb{E}}[\int^T_t[M({\widetilde}{X}_t,s)-f(s)]{\,\mathrm{d}}s]\to {\mathbb{E}}[J({\widetilde}{X}_t,t)]$ as $T\to\infty$, and then by Scheffé’s Lemma, $\int^T_t[M({\widetilde}{X}_t,s)-f(s)]{\,\mathrm{d}}s \to J({\widetilde}{X}_t,t)$ in $L^1$. On the other hand, since $Z_T({\widetilde}{X}_t)\to CQ({\widetilde}{X}_t)$ in $L^1$, $$G_T({\widetilde}{X}_t,t)\ \xrightarrow{L^1}\ G({\widetilde}{X}_t,t) \ \text{ and } \ H_T({\widetilde}{X}_t)\ \xrightarrow{L^1}\ H({\widetilde}{X}_t).$$ It follows that the process $\left(G({\widetilde}{X}_t,t);t\geq0\right)$ is a supermartingale and the process $\left(G({\widetilde}{X}_{t\wedge{\bar{\tau}_D}},t\wedge{\bar{\tau}_D});t\ge 0\right)$ is a martingale (since the conditional expectation, as an operator, is continuous in $L^p$ for $p\geq1$). We then can show as before that (if $\tau$, ${\bar{\tau}_D}$ are integrable), $$\begin{split} {\mathbb{E}}\left[F(\tau)\right]\ &\leq\ {\mathbb{E}}\left[G({\widetilde}{X}_\tau,\tau) +H({\widetilde}{X}_\tau)\right]\\ &\leq\ {\mathbb{E}}\left[ G({\widetilde}{X}_{{\bar{\tau}_D}},{\bar{\tau}_D}) +H({\widetilde}{X}_{{\bar{\tau}_D}})\right]\ =\ {\mathbb{E}}\left[F({\bar{\tau}_D})\right]. \end{split}$$ An example where holds is the call-type payoff: $F(t)=(t-K)_+$. We see that for $t>K$, the left derivative $f(t) = 1$, and hence $$J(x,t) = \int_t^K [M(x,s)-f(s)]{\,\mathrm{d}}s,$$ we then repeat all arguments above to obtain the pathwise inequality and the optimality result. On the other hand, the condition will fail if [e.g.]{} $F(t) = t^2$. Recall that a volatility swap corresponds to the choice of $F(t) = \sqrt{t}$, and that we can consider concave functions by taking $-F(t)$. This causes difficulties since $-F$ is not increasing, nor can we make it increasing by considering $-F(t) + \alpha t$, for some $\alpha >0$, since $F'(t) \to \infty$ as $t \to \infty$. However $F(t)$ can be approximated from both above and below by functions $F^1(t)$ and $F^2(t)$ which are concave, and have bounded derivatives. (In the case of the upper approximation, we have $F^1(0) >0$, but this can be made arbitrarily small). An optimality result will then follow in this setting — note also that the condition will be satisfied in this case. Superhedging options on weighted realised variance -------------------------------------------------- We now return to the financial context described in Section \[sec:financial-motivation\]. Our aim is to use the construction we produced for the proof of optimality in Section \[sec:optim-via-pathw\] to provide a model-independent hedging strategy for derivatives which are convex functions of weighted realised variance. We will suppose initially that our options are not forward starting, so $\nu = \delta_{S_0}$. We now define ${\bar{\tau}_D}$ as the embedding of $\mu$ for the diffusion ${\widetilde}{X}$, and define functions: $G, H, J$, and $Q$ as in the previous section (so holds). Our aim is to use , which now reads: $$\label{eq:GHineq} G(X_{A_t},t)+H(X_{A_t})\ =\ G({\widetilde}{X}_t,t)+H({\widetilde}{X}_t) \ \geq\ F(t)\ =\ F\left(\int_{0}^{A_t}w(X_s)\sigma_s^2{\,\mathrm{d}}s\right),$$ to construct a super-replicating portfolio. We shall first show that we can construct a trading strategy that super-replicates the $G({\widetilde}{X}_t,t)$ portion of the portfolio. Then we argue that we are able, using a portfolio of calls, puts, cash and the underlying, to replicate the payoff $H(X_T)$. Since $(G({\widetilde}{X}_t,t))_{t\geq0}$ is a supermartingale, we do not expect to be able to replicate this in a completely self-financing manner. However, by the Doob-Meyer decomposition theorem, and the martingale representation theorem, we can certainly find some process $({\widetilde}{\phi}_t)_{t\geq0}$ such that: $$G({\widetilde}{X}_t,t)\ \leq\ G({\widetilde}{X}_0,0) +\int_0^t{\widetilde}{\phi}_s{\,\mathrm{d}}{\widetilde}{X}_s$$ and such that there is equality at $t={\bar{\tau}_D}$. Moreover, since $(G({\widetilde}{X}_{t\wedge{\bar{\tau}_D}},t\wedge{\bar{\tau}_D}))_{t\geq0}$ is a martingale, and $G$ is of $\mathbb{C}^{2,1}$ class in $D$ (since $M(x,t)$ is), we have: $$G({\widetilde}{X}_{t\wedge{\bar{\tau}_D}},t\wedge{\bar{\tau}_D}) = G({\widetilde}{X}_0,0)+\int_{0}^{t\wedge{\bar{\tau}_D}} {\dfrac{\partial G}{\partial x}}({\widetilde}{X}_{s\wedge{\bar{\tau}_D}},s\wedge{\bar{\tau}_D}){\,\mathrm{d}}{\widetilde}{X}_s.$$ More generally, we would not expect $\partial G/\partial x$ to exist everywhere in $D^\complement$, however, if for example left and right derivatives exist, then we could choose $${\widetilde}{\phi}_t\ \in\ \left[{\dfrac{\partial G}{\partial x}}(x-,t), {\dfrac{\partial G}{\partial x}}(x+,t)\right]$$ as our holding of the risky asset. It follows then that we can identify a process $\left({\widetilde}{\phi}_t;t\geq0\right)$ with $$G({\widetilde}{X}_{\tau_t},\tau_t)\ \leq\ G({\widetilde}{X}_0,0) +\int_{0}^{\tau_t}{\widetilde}{\phi}_s{\,\mathrm{d}}{\widetilde}{X}_s \ =\ G(X_0,0)+\int_0^t{\widetilde}{\phi}_{\tau_s}{\,\mathrm{d}}X_s,$$ where we have used e.g. @RevuzYor:99 [Proposition V.1.4]. Finally, writing $\phi_t={\widetilde}{\phi}_{\tau_t}$, then $$\label{eq:Gineq} G(X_{t},\tau_t)\ \leq\ G(X_0,0)+\int_0^t\phi_s{\,\mathrm{d}}X_s \ =\ G(X_0,0)+\int_0^t\phi_s {\,\mathrm{d}}\left(B^{-1}_s S_s\right).$$ If we consider the self-financing portfolio which consists of holding $\phi_s B^{-1}_T$ units of the risky asset, and an initial investment of $G(X_0,0)B^{-1}_T-\phi_0S_0B^{-1}_T$ in the risk-free asset, this has value $V_t$ at time $t$, where ${\,\mathrm{d}}\left(B_{t}^{-1}V_t\right) =B^{-1}_T\phi_t\,{\,\mathrm{d}}\left(B^{-1}_tS_t\right)$ and $V_0\ =\ G(X_0,0)B_T^{-1}$, and therefore $$V_T\ =\ B_T\left(V_0B^{-1}_0+\int_0^T B_T^{-1}\phi_s {\,\mathrm{d}}\left(B_{s}^{-1}S_s\right)\right) \ =\ G(X_0,0)+\int_0^T\phi_s{\,\mathrm{d}}X_s.$$ We now turn to the $H(X_T)$ component in . If $H(x)$ can be written as the difference of two convex functions (so in particular, $H''({\,\mathrm{d}}K)$ is a well defined signed measure) we can write: $$\begin{split} H(x) & = H(S_0)+H'_+( S_0)(x-S_0)+\int_{(S_0,\infty)}(x-K)_+ H''({\,\mathrm{d}}K)\\ & \qquad {}+\int_{(0,S_0]}(K-x)_+ H''({\,\mathrm{d}}K). \end{split}$$ Taking $x=X_T=B_{T}^{-1}S_T$ we get: $$\begin{split} H(X_T) & = H(S_0)+H'_+( S_0)(B_{T}^{-1}S_T-S_0) +B_{T}^{-1}\int_{(S_0,\infty)}(S_T-B_T K)_+ H''({\,\mathrm{d}}K)\\ & \qquad {}+B_{T}^{-1}\int_{(0,S_0]}(B_T K-S_T)_+ H''({\,\mathrm{d}}K). \end{split}$$ This implies that the payoff $H(X_T)$ can be replicated at time $T$ by ‘holding’ a portfolio of: $$\label{eq:Hport} \begin{split} &B_{T}^{-1}\big[H(S_0)-S_0H'_+(S_0)\big]\ \text{ in cash;}\\ &B_{T}^{-1}H'_+(S_0)\ \,\,\text{ units of the asset;}\\ &B_{T}^{-1}H''({\,\mathrm{d}}K)\ \text{ units of the call with strike } B_T K\text{ for }K\in(S_0,\infty);\\ &B_{T}^{-1}H''({\,\mathrm{d}}K)\ \text{ units of the put with strike } B_T K\text{ for }K\in(0,S_0],\\ \end{split}$$ where the final two terms should be interpreted appropriately. In practice, the function $H(\cdot)$ can typically be approximated by a piecewise linear function, where the ‘kinks’ in the function correspond to traded strikes of calls or puts, in which case the number of units of each option to hold is determined by the change in the gradient at the relevant strike. The initial cost of setting up such a portfolio is well defined provided the integrability condition: $$\label{eq:Hddcond} \int_{(0,S_0]}P(B_T K)|H''|({\,\mathrm{d}}K) +\int_{(S_0,\infty)} C(B_T K)|H''|({\,\mathrm{d}}K)\ <\ \infty,$$ holds, where $|H''|({\,\mathrm{d}}K)$ is the total variation of the signed measure $H''({\,\mathrm{d}}K)$. We therefore shall make the following assumption: \[ass:Hass\] The payoff $H(X_T)$ can be replicated using a suitable portfolio of call and put options, cash and the underlying, with a finite price at time $0$. We can therefore combine these to get the following theorem: \[thm:rHedge\] Suppose Assumptions \[ass:Price\], \[ass:Calls\] and \[ass:Hass\] hold, and suppose $F(\cdot)$ is a convex, increasing function with $F(0)=0$ and the left derivative $f(t):=F'_-(t)$ satisfies . Let $M(x,t)$ and $J(x,t)$ be given by and , and are determined by the solution to [**[SEP]{}**]{}${}^*(x w(x)^{-1/2},\delta_{S_0},\mu)$, where $\mu$ is determined by and $w \in \mathcal{D}$. We also define $Q$ after , such that holds and then the functions $G$ and $H$ are given by . Then there exists an arbitrage if the price of an option with payoff $F(RV^w_T)$ is strictly greater than $$\label{eq:Fub} \begin{split} B_T^{-1}\Big[G(S_0,0)+H(S_0) & +\int_{(S_0,\infty)}C(B_T K)H''({\,\mathrm{d}}K)\\ & \qquad {}+\int_{(0,S_0]}P(B_T K)H''({\,\mathrm{d}}K)\Big]. \end{split}$$ Moreover, this bound is optimal in the sense that there exists a model which is free of arbitrage, under which the bound can be attained, and the arbitrage strategy can be chosen independent of the model. According to the arguments above, our superhedge of the variance option can be described as the combination of a static portfolio and a self-financing dynamic portfolio which consists of an additional $B^{-1}_T\psi_t$ units of the risky asset and an additional initial cash holding of $B^{-1}_T\left(G(S_0,0)-\psi_0S_0\right)$. In the case where $G(x,t)$ is sufficiently differentiable, we can identify the process $\psi_t={\widetilde}{\psi}_{\tau_t}$ by $$\psi_t = {\dfrac{\partial G}{\partial x}}(X_{t},\tau_t).$$ We observe that this strategy is independent of the true model. It is easy to see that the total initial investment of this superhedge is given be . In the case where $G$ is not sufficiently differentiable, we first observe that $G(x,t)$ is continuous: note that $Q(x)$ and $F(t)$ are trivially so by , and $M(x,t)$ is continuous in $D$, and in $D^\complement$, and additionally at jumps of $R(x)$: it follows that $M(x_n,t) \to M(x,t)$ as $x_n \to x$ except possibly at a set of Lebesgue measure zero, and hence $G(x,t)$ is continuous. Now consider $G(x,t)$ on a bounded open set of the form $\mathcal{O} = (y_0,y_1) \times (t_0,t_1)$. By continuity, $G$ can be approximated uniformly on the boundary $\partial \mathcal{O}$ (or more relevantly, on the boundary where $t>t_0$, $\partial \mathcal{O}_+$) by a smooth function. Specifically, for fixed ${\varepsilon}>0$, there exists a function $G_{\varepsilon}(x,t)$ such that $G(x,t) + {\varepsilon}\ge G_{{\varepsilon}}(x,t) \ge G(x,t)$ on $\partial \mathcal{O}_+$. Moreover, the function $$\label{eq:Gepsdefn} G_{{\varepsilon}}(x,t) = {\mathbb{E}}^{(x,t)} \left[ G_{{\varepsilon}}({\widetilde}{X}_{\tau_{\partial \mathcal{O}_+}}, \tau_{\partial \mathcal{O}_+}) \right]$$ is $C^{2,1}$ and a martingale on $\bar{\mathcal{O}}$, and so $$G_{{\varepsilon}}({\widetilde}{X}_{\tau_{\partial \mathcal{O}_+}},\tau_{\partial \mathcal{O}_+}) = G_{{\varepsilon}}({\widetilde}{X}_0,t_0) +\int_{t_0}^{\tau_{\partial \mathcal{O}_+}}{\frac{\partial G_{{\varepsilon}}}{\partial x}}{\,\mathrm{d}}{\widetilde}{X}_s.$$ Since $G$ is a supermartingale, for $(x,t) \in \mathcal{O}$, from we have $G(x,t) \ge G_{{\varepsilon}}(x,t) - {\varepsilon}$. Now observe that we can choose a countable sequence of such sets $\mathcal{O}_1, \mathcal{O}_2, \ldots$ with each set centred at the exit point of the previous set, and such that any continuous path is guaranteed to pass through only finitely many such sets on a finite time interval. For any fixed $\delta > 0$, we can take a sequence of strictly positive ${\varepsilon}_1, {\varepsilon}_2, \ldots$ such that $\sum_{i} {\varepsilon}_i = \delta$, and apply the arguments above to generate a sequence of functions $G_{{\varepsilon}_i}(x,t)$ on $\mathcal{O}_i$. It follows that, given $\delta > 0$, we can always find a function ${\widetilde}{\psi}_t$ such that $$\delta + G({\widetilde}{X}_0,0) + \int_0^t {\widetilde}{\psi}_t \, {\,\mathrm{d}}{\widetilde}{X}_s \ge G({\widetilde}{X}_t,t).$$ Since $\delta$ was arbitrary, whenever the price of an option is strictly greater than , we can choose $\delta$ sufficiently small that the arbitrage still works. Finally, we observe that at any time $t \in [0,T]$, the arbitrage strategy is worth at least $F(\tau_t) \ge F(0)$, so the strategy is bounded below, and hence admissible. To see that this is the best possible upper bound, we need to show that there is a model which satisfies Assumption \[ass:Price\], has law $\mu$ under ${\mathbb{Q}}$ at time $T$, and such that the superhedge is actually a hedge. But consider the stopping time ${\bar{\tau}_D}$ for the process ${\widetilde}{X}_t$. Define the process $\left(X_t;0\leq t\leq T\right)$ by $$X_t= {\widetilde}{X}_{\frac{t}{T-t}\wedge{\bar{\tau}_D}}, \ \ \text{for }\ t\in[0,T],$$ and then $X_t$ satisfies the stochastic differential equation $${\,\mathrm{d}}X_s=\hat{\sigma}_s X_s w(X_s)^{-1/2} {\,\mathrm{d}}W_s = \sigma_s^2 X_s {\,\mathrm{d}}W_s$$ with the choice of $$\label{eq:choicesigma} \hat{\sigma}^2_s = \dfrac{T}{(T-s)^2}{\boldsymbol{1}}_{[\frac{s}{T-s}<{\bar{\tau}_D}]}, \quad \sigma^2_s = \hat{\sigma}_s^2 w(X_s)^{-1/2}.$$ Since ${\bar{\tau}_D}<\infty$, a.s., then $X_T={\widetilde}{X}_{{\bar{\tau}_D}}$, and $$\tau_T= \int_0^T w(X_s) \sigma_s^2 {\mathrm{d}s}= \int_0^T \dfrac{T}{(T-s)^2}{\boldsymbol{1}}_{[\frac{s}{T-s}<{\bar{\tau}_D}]} = {\bar{\tau}_D}.$$ Hence $S_t=X_t B_t$ is a price process satisfying Assumption \[ass:Price\] with $$F\left(\int_0^Tw(X_s) \sigma^2_t{\,\mathrm{d}}t\right)\ =\ F({\bar{\tau}_D}).$$ Finally, it follows that at time $T$, the value of the hedging portfolio exactly equals the payoff of the option. \[rmk:ForwardStart\] The above result assumes that the option payoff depends on the realised weighted variation computed between time $0$ and a fixed time $T$. In some situations, [*forward-starting*]{} versions of these derivatives may be traded. Here, one is interested in the payoff of an option written on the variation observed between a fixed time $T_0 >0$ and the maturity date $T_1$: $\int_{T_0}^{T_1} w(X_t)\sigma^2_t \, {\mathrm{d}t}$. If one observes traded options at both $T_0$ and $T_1$, these again imply the (hypothesised, risk-neutral) distributions at times $T_0$, and $T_1$, and it is reasonable to suppose that the upper bound on the price of an option (for suitable, convex $F(\cdot)$) should correspond to the solution of [[**[SEP]{}${}^*(\sigma,\nu,\mu)$**]{}]{} determined above. Let $G$ and $H$ be the functions derived above. The question remains as to how one includes the additional information at time $T_0$ in the hedging strategy. (For clarity, we suppose $B_t = 1$ for all $t\ge 0$.) In order to have the correct hedge for $t \in [T_0,T_1]$, we need a portfolio of call options maturing at time $T_1$ with payoff $H(X_{T_1})$. In addition to the payoff at maturity, we need a dynamic portfolio worth (at least) $G(X_t,\tau_t)$, where now $\tau_t = \int_{T_0}^t \lambda(X_s) \sigma_s^2 \, {\mathrm{d}s}$ — specifically, recalling $F(0)=0$, and Proposition \[prop:GTineq\], we should have $G(X_{T_0},0) + H(X_{T_0}) = 0$. This implies that we need a portfolio of call options with maturity $T_0$ and with payoff $-H(X_{T_0})$. Under a similar assumption to Assumption \[ass:Hass\], this is possible, and the resulting strategy will give a superhedge which is a hedge under the optimal model corresponding to the Rost embedding. Strictly speaking, Theorem \[thm:rHedge\] is model-dependent: our arbitrage strategy is specified in a way that is independent of the exact model, but some of the underlying concepts — specifically the quadratic variation in the option payoff, and the stochastic integral term that is implemented in the hedge both depend on an underlying probability space, and it could therefore be argued that the strategies are not truly model-independent. In the following remarks, we briefly outline how one might relax this assumption. In a similar manner to recent work of @DavisOblojRaval:10, we can formulate this result without any need for a probabilistic framework. The difficulty in treating the previous arguments on a purely pathwise basis is that we need to make sense of the stochastic integral term in , and the quadratic variation in the option payoff. However, under mild assumptions on the paths of $S_t=B_t^{-1} X_t$, and a stronger assumption on $G$ (specifically, that $G$ is $\mathbb{C}^{2,1}$)[^4], we can recover a pathwise result, based on a version of Itô’s formula due to @Follmer:81. Suppose we fix a sequence of partitions $\pi_n = \{0 = t_0^n \le t_1^n \le t_2^n \le \cdots \le t_{n}^n=T\}$ of $[0,T]$, such that $\sup_{i \le n,}|t_{i}^n-t_{i-1}^n| \to 0$ as $n \to \infty$. Then we define the class $\mathcal{QV}$ of continuous, strictly positive paths $X_t$ such that $$\label{eq:QVdefn} \sum_{i =1}^n \left(\frac{ X_{t_{i}^n}-X_{t_{i-1}}^n}{X_{t_{i-1}}^n}\right)^2 \delta_{t_{i-1}^n} = \mu^n \to \mu \text{ where } \mu([0,t]) = \int_0^t \sigma_s^2 \, {\mathrm{d}s}$$ for some bounded measurable function $\sigma_s:[0,T] \to {\mathbb{R}}_+$. Here $\delta_t$ is the Dirac measure at $t$, and the convergence is in the sense of weak convergence of measures as $n \to \infty$, possibly down a subsequence. Then, following the proof of the main theorem in @Follmer:81, an application of Taylor’s Theorem to the terms $G(X_{t_{i}^n},\tau_{t_i^n})-G(X_{t_{i-1}^n},\tau_{t_{i-1}^n})$, where $\tau_t = \int_0^t \lambda(X_s) \sigma_s^2 \, {\mathrm{d}s}$, gives $$\begin{aligned} G(X_T,\tau_T) - G(X_0,0) = {} & \sum_{i =1}^n {\frac{\partial G}{\partial x}} \left( X_{t_i^n}-X_{t_{i-1}^n}\right) + \sum_{i =1}^n {\frac{\partial G}{\partial t}} \left( \tau_{t_i^n}-\tau_{t_{i-1}^n}\right) \\& {}+ {\dfrac{1}{2}}\sum_{i =1}^n {\frac{\partial^{2} G}{\partial x^{2}}} \left( X_{t_i^n}-X_{t_{i-1}^n}\right)^2. \end{aligned}$$ It follows that, whenever $X_t$ is a path in $\mathcal{QV}$, then: $$\begin{aligned} \sum_{i =1}^n {\frac{\partial G}{\partial x}} \left( X_{t_i^n}-X_{t_{i-1}^n}\right) \to {} & G(X_T,\tau_T) - G(X_0,0) \\ & {} - \int_0^T \sigma_s^2 \left( \lambda(X_s){\frac{\partial G}{\partial t}} + {\dfrac{1}{2}}{\frac{\partial^{2} G}{\partial x^{2}}} X_s^{2}\right) \, {\mathrm{d}s}. \end{aligned}$$ Recall that $\sigma(x) = x \lambda^{-1/2}(x)$, and since $G \in \mathbb{C}^{2,1}$ and a supermartingale for $X_t$ where $X_t$ solves , the final integrand will be negative. We conclude that, in the limit as we trade more often, *for any* $X_t \in \mathcal{QV}$, we will have a portfolio which superhedges. One could then recover the statement in the probabilistic setting by observing that, almost surely, a path from a model of the form described by Assumption \[ass:Price\] lies in $\mathcal{QV}$. \[rmk:uncertainvolatility\] An alternative approach that is still within a more general, model-independent setting, but where we do not need to assume strong differentiability conditions on $G$, can be constructed using the uncertain volatility approach, originally introduced by @Avellaneda:1995aa. We base our presentation on the paper of @possamai_robust_2013. Let $\Omega = \left\{\omega \in C([0,T];(0,\infty)), \omega(0) = X_0 \right\}$ be a path space, equipped with the uniform norm, $||\omega|| = \sup_{t \in [0,T]}|\omega(t)|$, and let $X_t(\omega) = \omega(t)$ be the canonical process. Let ${{\mathbb{P}}}_0$ be the probability measure on $\Omega$ such that $X_t$ is a standard geometric Brownian motion ([i.e.]{} $\log{X_t}$ has quadratic variation $t$). Let $\mathbb{F}$ be the filtration generated by $X$. Let $\mathbb{H}_{loc}^e({{\mathbb{P}}}_0,\mathbb{F})$ be the set of non-negative, $\mathbb{F}$-progressively measurable processes $\alpha_t$ such that $\exp\left\{{\dfrac{1}{2}}\int_0^\cdot \alpha_s \, {\mathrm{d}s}\right\}$ is ${{\mathbb{P}}}_0$-locally integrable. Then for $\alpha \in \mathbb{H}_{loc}^e({{\mathbb{P}}}_0,\mathbb{F})$ we can define $$X_t^\alpha = \exp \left\{ \int_0^t \alpha_s^{1/2} \, {\,\mathrm{d}}\log(X_s) - {\dfrac{1}{2}}\int_0^t \alpha_s \, {\mathrm{d}s}\right\}.$$ In particular, under ${{\mathbb{P}}}_0$, $X_t^\alpha$ has $\left< \log X\right>_t = \int_0^t \alpha_s \, {\mathrm{d}s}$. Then we can define a probability measure on $\Omega$ by ${{\mathbb{P}}}^\alpha(X_t \in A) = {{\mathbb{P}}}_0(X_t^\alpha \in A)$, or equivalently, ${{\mathbb{P}}}^\alpha = {{\mathbb{P}}}_0 \circ (X_\cdot^\alpha)^{-1}$. It follows that there is a class of probability measures $\mathcal{P} = \left\{{{\mathbb{P}}}^\alpha: \alpha \in \mathbb{H}_{loc}^e({{\mathbb{P}}}_0,\mathbb{F})\right\}$ on the space $(\Omega,\mathbb{F})$. We aim to produce conclusions which hold for all ${{\mathbb{P}}}^\alpha$, and we say that something holds $\mathcal{P}$-quasi surely (q.s.) if it holds ${{\mathbb{P}}}$-[a.s.]{} for all ${{\mathbb{P}}}\in \mathcal{P}$. We now have a filtered space $(\Omega,\mathbb{F})$, and a class of (non-dominated) probability measures $\mathcal{P}$ under which we can discuss trading strategies simultaneously. Observe that the variance process $\langle \log X \rangle_t$ can be defined pathwise on $\Omega$ using the results of @karandikar_pathwise_1995: set $a_0^n = 0$, and $a_{i+1}^n = \inf\{t \ge a_i^n : |\log(\omega(t)) - \log(\omega(a_i^n))| \ge 2^{-n}\}$, and consider the process $V_t(\omega)$ defined by $$\label{eq:Vnolimit} V_t(\omega) = \lim_{n \to \infty} \sum_{i=0}^n \left( \log(\omega(a_{i+1}^n\wedge t))-\log(\omega(a_{i}^n\wedge t)) \right)$$ if the limit exists, where the limit is taken in the sense of uniform convergence on $[0,T]$ and defined to be zero otherwise. The limit exists ${{\mathbb{P}}}^\alpha$-[a.s.]{} for each $\alpha$, and when the limit exists, the limit is $\mathcal{F}_t$ measurable and ${{\mathbb{P}}}^\alpha$-[a.s.]{} equal to $\langle \log X \rangle_t$. If we write $$\mathcal{N}^\mathcal{P} = \{ E \subset \Omega: \exists {\widetilde}{E} \in \mathcal{F} \ s.t.{}\ E \subseteq {\widetilde}{E},\ {{\mathbb{P}}}(E) = 0 \ \forall \ {{\mathbb{P}}}\in \mathcal{P}\},$$ then the set of $\omega$ for which the limit in fails is an element of $\mathcal{N}^\mathcal{P}$. As a result, we can make the process $V_t$ adapted by considering the augmented filtration: $$\mathcal{F}_t' = \mathcal{F}_t \vee \mathcal{N}^\mathcal{P}, \quad \mathbb{F}' = \{\mathcal{F}_t', t \ge 0\}.$$ Under this augmented filtration, the process $V_t$ remains $\mathbb{F}'$-progressively measurable, and indeed is continuous; it follows that the trading strategy described in the proof of Theorem \[thm:rHedge\] can be constructed, giving a càdlàg process $\psi_t$ which is continuous except at the times when the process $(\omega(t),V_t(\omega))$ exits the sets $\mathcal{O}_1, \mathcal{O}_2, \ldots$. Using the construction of @karandikar_pathwise_1995, we can again define pathwise a process $I_t$, which agrees ${{\mathbb{P}}}^\alpha$-[a.s.]{} with the classical stochastic integral $J_t^{{\mathbb{P}}}= \int_0^t \psi_s {\,\mathrm{d}}X_s$. (Observe however that we may need to work in a ${{\mathbb{P}}}^\alpha$-augmented filtration for this latter object to be defined). Since by Theorem \[thm:rHedge\] we have $J_T^{{\mathbb{P}}}\ge G(X_T,\langle \log X \rangle_T)-{\varepsilon}$, it follows that $I_T \ge G(X_T,V_T) - {\varepsilon}$ $\mathcal{P}$-q.s., and therefore the strategy we describe makes sense in the uncertain volatility setting. The fact that we have a concrete characterisation of $\psi_t$ enables us to avoid much of the technical difficulties that arise in [@possamai_robust_2013] and related papers. However, our results are in one sense also not quite so strong: we only obtain a strategy which superhedges our payoff less some ${\varepsilon}>0$. The results in [@possamai_robust_2013] suggest that this is unnecessary. However, our results are stronger in another direction: we do not require any integrability restriction on the payoff of the option under the class of models we consider — this constraint is already embedded in our restriction to non-negative price processes. Numerical Results {#sec:numerical-results} ================= Numerical solution of the viscosity equation -------------------------------------------- An important goal is to use the results of the previous sections to find numerical bounds, and their associated option prices and hedging strategies, corresponding to the solutions of Rost and Root. The hardest aspect of this is finding the numerical solution to the viscosity equation , and its equivalent for the Root solution. The solution to the Rost viscosity equation is roughly equivalent to solving a parabolic PDE inside the continuation region, while outside this region we know the solution will be equal to the initial boundary condition. The numerical solution is made harder by the fact that, particularly in the case of the Rost solution, we expect the behaviour of the barrier near the initial starting point to be very sensitive to any discretisation: in the case where the starting measure is a point mass at $X_0$, and the target measure also places mass continuously (say) near $X_0$, then we are looking for a barrier function $R(x)$ with $R(X_0) = 0$, and a positive, but non-zero probability that $R(X_t)>t$ for some small time $t$. According to the law of the iterated logarithm, the behaviour of the stopped process will be very sensitive to small changes in $R(\cdot)$. As a result, a numerical method that can concentrate on this initial region would be beneficial. On the other hand, the behaviour of the barrier at large times is also of interest, although here we expect the numerics are likely to be less sensitive to discretisation. A second question concerns the convergence and stability of our numerical methods. The theory behind the numerical approximation of viscosity equations is fairly well understood — dating back to the methods of @Barles:1991aa. In this paper, we use the results of @Barles:2007aa, which are suited to our purposes. Since we wish to use a large range of time steps, and we look to have non-equal grid point spacings, we will look to use an implicit method, in order to provide unconditional stability of the numerical regime. The results of [@Barles:2007aa] provide us with the necessary justification. To outline the numerical method used, we consider a standard numerical scheme, with ${\mathbf{u}}^n$ the (vector valued) approximation to $u(x,t)$ evaluated at $t=t^n$, and at spatial positions ${\mathbf{x}}$. We approximate $\mathcal{L}u = \frac{\sigma(x)^2}{2} {\frac{\partial ^2 u}{\partial x^2}}(x,t)$ using the Kushner approximation described by [@Barles:2007aa], which ensures that the finite difference operator $L$ can be written in the form $(L{\mathbf{u}})_i = \sum_{j} c_{j}(t^n,x_{i+j}) (u_{i+j}-u_i)$, where the $c_j$’s are non-negative and zero except on some finite subset of $\mathbb{Z} \setminus \{0\}$. We also need to assume that the measures $\mu$ and $\nu$ both have compact support and the same mean, in which case ${\mathbf{u}}$ is constant and zero at the endpoints of ${\mathbf{x}}$. Then an implicit numerical scheme to solve will take ${\mathbf{u}}_0= u({\mathbf{x}},0)$, and solve iteratively $$\label{eq:DiscretePDE} \frac{{\mathbf{u}}_{n+1}-{\mathbf{u}}_n}{t^{n+1}-t^n} = \max\{L{\mathbf{u}}_{n+1},0\}.$$ The difficulty here arising from the fact that the maximisation depends on the unknown ${\mathbf{u}}_{n+1}$. We can rearrange this expression, writing ${\mathbf{z}} = {\mathbf{u}}_{n+1}-{\mathbf{u}}_{n}$, and $\Delta t^n = t^{n+1}-t^n$, to see that this is equivalent to the problem of finding ${\mathbf{z}} \ge 0$ such that: $(I-\Delta t^n L) {\mathbf{z}} - \Delta t^n L {\mathbf{u}}^n \ge 0$, and ${\mathbf{z}}^\intercal \left((I-\Delta t^n L) {\mathbf{z}} - \Delta t^n L {\mathbf{u}}^n\right) = 0$. This is a classical linear complementarity problem (LCP), and may be hard to solve (or at least, may involve many evaluations of the matrix multiplication inside the maximisation), however, at this point we can exploit the fact that the structure of the solution implies that ${\mathbf{z}}$ will be zero at exactly the points where we are in the barrier. Since the barrier should generally change relatively slowly, as an initial supposition, it is likely that the spatial values where ${\mathbf{z}} = 0$ for the previous time-step are likely to be roughly the same at the next step. It follows that a numerical scheme for solving LCPs which involves pivoting on a set of basis variables may be very efficient at solving . The algorithm we will use for this purpose is the Complementary Pivot (or Lemke’s) algorithm. We refer to @Murty:1988aa for details on the numerical implementation of the Complementary Pivot algorithm. We note also that a similar method can be used to justify implicit methods for the Root solution (the case of explicit solutions being justified directly by the results of [@OberhauserDosReis:13]). Analysis of numerical evidence ------------------------------ Using the methods outlined above, we can analyse the solutions of Root and Rost numerically. In general, we consider $\nu = \delta_{S_0}$ and $\mu$ will be determined by assuming that we observe prices of call options which are consistent with a Heston market model. In general we will consider features of barriers under Heston models since they permit relatively straightforward computation of both call prices, and prices of variance options. In what follows, we take our given prices to come from a Heston model with parameters: $\rho = -0.65, v_0 = 0.04, \theta = 0.035, \kappa = 1.2, r = 0, \xi = 0.5, S_0 = 2$ (see for the meaning of the parameters). ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Plots of the function $u(x,t)$ (top) and the corresponding barriers (bottom) for the Rost (left) and Root (right) barriers.[]{data-label="fig:basicbarriers"}](Heston_Rost_u.eps "fig:"){width=".48\textwidth"} ![Plots of the function $u(x,t)$ (top) and the corresponding barriers (bottom) for the Rost (left) and Root (right) barriers.[]{data-label="fig:basicbarriers"}](Heston_Root_u.eps "fig:"){width=".48\textwidth"} ![Plots of the function $u(x,t)$ (top) and the corresponding barriers (bottom) for the Rost (left) and Root (right) barriers.[]{data-label="fig:basicbarriers"}](Heston_Rost_Barrier.eps "fig:"){width=".48\textwidth"} ![Plots of the function $u(x,t)$ (top) and the corresponding barriers (bottom) for the Rost (left) and Root (right) barriers.[]{data-label="fig:basicbarriers"}](Heston_Root_Barrier.eps "fig:"){width=".48\textwidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![The original call prices from which we obtained our barrier, and the empirical call prices obtained by simulation for the Rost and Root barriers (left); The implied volatility of the call prices (right). Note that numerically the Rost barrier proves harder to correctly compute/simulate.[]{data-label="fig:EmpiricalImpliedVol"}](Heston_Empirical_Calls.eps "fig:"){width=".48\textwidth"} ![The original call prices from which we obtained our barrier, and the empirical call prices obtained by simulation for the Rost and Root barriers (left); The implied volatility of the call prices (right). Note that numerically the Rost barrier proves harder to correctly compute/simulate.[]{data-label="fig:EmpiricalImpliedVol"}](Heston_Empirical_ImpVol.eps "fig:"){width=".48\textwidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Figure \[fig:basicbarriers\] shows the functions $u(x,t)$ in both the Rost and Root solutions, and their corresponding barrier functions. We can confirm that these functions do indeed embed the correct distributions by simulation: we compute the distribution of a process stopped on exit from the barrier and compute the corresponding call prices empirically. In fact, it is more informative to plot the implied volatility of the empirically obtained call prices. This is done in Figure \[fig:EmpiricalImpliedVol\]. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![We compare the barriers for multiple maturities. In this figure we compute the barriers at equal spaced maturities of the underlying Heston model (the last barrier corresponding to $T=1$) for the Rost (left) and Root (right) cases.[]{data-label="fig:BarriersInMaturity"}](Heston_Rost_Barriers_in_Maturity.eps "fig:"){width=".48\textwidth"} ![We compare the barriers for multiple maturities. In this figure we compute the barriers at equal spaced maturities of the underlying Heston model (the last barrier corresponding to $T=1$) for the Rost (left) and Root (right) cases.[]{data-label="fig:BarriersInMaturity"}](Heston_Root_Barriers_in_Maturity.eps "fig:"){width=".48\textwidth"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- One can also consider the behaviour of the barriers in time. In Figure \[fig:BarriersInMaturity\] we plot the barriers for a sequence of call prices with increasing maturity. ![A plot of the upper bound on the price of a Variance call for different maturities (equally spaced, up to $T=1$). For comparison, we also plot the actual prices of the variance calls under the Heston model corresponding to the shortest and longest maturities.[]{data-label="fig:VarCallInMaturity"}](Heston_Var_Call_Upper_Bound_in_Maturity.eps){width=".68\textwidth"} Of course, our interest lies in the implied bounds of options on variance. We first consider the case of a variance call. In Figure \[fig:VarCallInMaturity\] we display the upper bound on the price of a variance call derived in Section \[sec:optRost\]. As might be expected, there is a substantial difference between the upper bound and the model-implied price. ![A plot of a realisation of the intrinsic value of the option ($F(\langle \ln S \rangle_t)$, and the value of the super-hedging and sub-hedging portfolios at time $t$, for $t \in [0,T]$. Here $F(t) = t(t-v_K)$, where $v_K = 0.01875$.[]{data-label="fig:HedgeSimulate"}](Heston_Hedge_True_Model.eps){width=".68\textwidth"} To see how the hedges constructed in perform in a given realisation, we can simulate a path, and compute the values of the super- and sub-hedging strategies along the realisation. In this example, we consider an option on variance with payoff $F(\langle \ln S \rangle_T)$, where $F(t) = t (t \wedge v_K)$. This is shown in Figure \[fig:HedgeSimulate\]. ![A plot of a realisation of the intrinsic value of the option, as in Figure \[fig:HedgeSimulate\], and the corresponding sub- and super-hedges, where the realisation is taken from a different model to that under which the original hedge was constructed.[]{data-label="fig:HedgeWrongModel"}](Heston_Hedge_Wrong_Model.eps){width=".68\textwidth"} The main attraction of these hedging portfolios is that they remain super/sub-hedges under a different model. For example, in Figure \[fig:HedgeWrongModel\] we show how these hedges behave if the path realisation comes from a Heston model with different parameters. Here we set: $\rho' = 0.5, \theta' = 0.07$ and $\kappa' = 2.4$. To conclude, we show that the sub- and super-hedges provide good model-robustness by computing (empirically) the difference between the payoff of an option on variance, and the corresponding super- or sub-hedge. This is shown in Figure \[fig:HedgeHistogram\], which also shows the effect of model-misspecification on the distribution of the hedging error. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![The (empirically computed) distribution of the hedging error for the super-hedge (left) and sub-hedge (right). In the former case, a positive error represents a surplus in the hedge, while in the latter case, a positive error represents an underhedge. We also compare the tracking error with the tracking error for the same strategy when our realised path is determined by the mis-specified model of Figure \[fig:HedgeWrongModel\].[]{data-label="fig:HedgeHistogram"}](Heston_Superhedge_Error_Compare.eps "fig:"){width=".48\textwidth"} ![The (empirically computed) distribution of the hedging error for the super-hedge (left) and sub-hedge (right). In the former case, a positive error represents a surplus in the hedge, while in the latter case, a positive error represents an underhedge. We also compare the tracking error with the tracking error for the same strategy when our realised path is determined by the mis-specified model of Figure \[fig:HedgeWrongModel\].[]{data-label="fig:HedgeHistogram"}](Heston_Subhedge_Error_Compare.eps "fig:"){width=".48\textwidth"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Extremality and the Heston-Nandi model {#sec:extr-hest-nandi} ====================================== In this section, we consider a particular, commonly used model for asset prices — the Heston-Nandi model — and show that it can have particularly bad implications for the pricing of variance options. The Heston-Nandi model [@HestonNandi:98] is the common Heston stochastic volatility model [@Heston:93], where the correlation $\rho$ between the Brownian motions driving the asset and the volatility processes is taken to be $-1$. Since in some asset classes, $\rho \approx -1$ is not abnormal ([e.g.]{} a Heston model calibrated to the S&P 500 typically has $\rho \approx -0.9 \pm 0.1$, for example, see @GuillaumeSchoutens:12), and the pricing of options on variance in Heston models is also common practice [@JavaheriWilmottHaug:04], this may substantially bias the prices towards the extreme models. The Heston model is given (under the risk-neutral measure) by: $$\label{eq:Heston} \begin{aligned} {\mathrm{d}}S_t & = r S_t \, {\mathrm{d}t}+ \sqrt{v_t} S_t \, {\mathrm{d}}B_t,\\ {\mathrm{d}}v_t & = \kappa (\theta - v_t) \, {\mathrm{d}t}+ \xi \sqrt{v_t} \, {\mathrm{d}}\widetilde{B}_t, \end{aligned}$$ where $B_t$ and $\widetilde{B}_t$ are Brownian motions with correlation $\rho$. The Heston-Nandi model is the restricted case where $\rho = -1$, and so $\widetilde{B}_t = - B_t$. Note that $v_t = \sigma_t^2$ in our previous notation, so we are interested in options on $\int_0^T v_t \, {\mathrm{d}t}$ The simplification $\rho = -1$ allows for the following observation: using Itô’s Lemma, we know $$\begin{aligned} {\mathrm{d}}(\log({\mathrm{e}}^{-rt}S_t)) & = - {\dfrac{1}{2}}v_t {\,\mathrm{d}}t + \sqrt{v_t} {\,\mathrm{d}}B_t \\ & = \left( \frac{\kappa \theta}{\xi} - \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right)v_t \right) {\,\mathrm{d}}t -\frac{1}{\xi} {\mathrm{d}}v_t.\end{aligned}$$ Solving, we see that $$\label{eq:HNProc} \log \left( \frac{{\mathrm{e}}^{-rT}S_T}{S_0} \right) = \frac{1}{\xi}(v_0 - v_T) + \frac{\kappa \theta}{\xi} T - \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right) \int_0^T v_t {\,\mathrm{d}}t.$$ If we assume that the maturity time of our option, $T$, is sufficiently large, since $v_T$ is mean reverting, $(v_T - v_0)\approx (\theta - v_0)$ will be small in relation to the other terms on the right-hand side. If we temporarily ignore the $v_T-v_0$ term, tells us that, at time $T$, we have $$\int_0^T v_t \, {\mathrm{d}t}\approx \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right)^{-1} \left[ \log\left( \frac{S_0}{S_T}\right) + \frac{\kappa \theta}{\xi} T \right].$$ Writing $$\label{eq:RTdefn} R_T(x) = \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right)^{-1} \left[ \log\left( \frac{S_0}{x}\right) + \frac{\kappa \theta}{\xi} T \right],$$ then we have $$T \approx \inf\left\{ s \ge 0 : \int_0^s v_t \, {\mathrm{d}t}\ge R_T(X_s)\right\}.$$ This describes a barrier stopping time, corresponding to a Root stopping time, with $$\label{eq:HNDdefn} D_T = \left\{ (x,t) \in {\mathbb{R}}\times {\mathbb{R}}_+ : t < R_T(x)\right\}.$$ So, ignoring the term in $v_T$, we might conjecture that the corresponding model minimises the value of a derivative which is a convex, increasing function of $v_T$ over all models with the same law at time $T$. This leads to the following result: \[thm:HNThm\] Let $M>0$ and suppose $\xi, \theta, \kappa, r>0, \xi \neq 2 \kappa$ are given parameters of a Heston-Nandi model, ${\mathbb{Q}}^{HN}$. Suppose $\mathcal{Q}_T$ is the class of models ${\mathbb{Q}}$ satisfying Assumption \[ass:Price\] and ${\mathbb{E}}^{{\mathbb{Q}}^{HN}}(S_T-K)_+ = {\mathbb{E}}^{\mathbb{Q}}(S_T -K)_+$ for all $K\ge 0$. Then there exists a constant $\kappa$, depending only on $M$ and the parameters of the Heston-Nandi model, such that for all convex, increasing functions $F(t)$ with suitably smooth derivative $f(t) = F'(t)$ such that $f(t), f'(t) \le M^*$, and for all $T \ge 0$ $$\label{eq:HNThm} {\mathbb{E}}^{{\mathbb{Q}}^{HN}} F\left(\langle \log S\rangle_T\right) \le \inf_{{\mathbb{Q}}\in \mathcal{Q}_T} {\mathbb{E}}^{\mathbb{Q}}F\left(\langle \log S\rangle_T\right) + \kappa.$$ Note that the strength of the result depends on the fact that the constant $\kappa$ is independent of both $T$ and $F$. In particular, $\langle \log S\rangle_T$ should be both growing in $T$ and increasing in variance as $T$ increases. That this does not appear in the bound leads us to claim that ${\mathbb{Q}}^{HN}$ is [*asymptotically optimal*]{}. In fact, the continuity assumptions can be trivially relaxed, and this leads to the simple corollary: The conclusions of Theorem \[thm:HNThm\] hold where the class of functions $F$ considered is the set of variance call payoffs: $F_K(t) = (t-K)_+$ for all $K \in {\mathbb{R}}^+$ and all maturity dates $T>0$. For fixed $M^*$, the function $F_K(t)$ can be approximated uniformly from above and below by a suitably smooth function satisfying the conditions of Theorem \[thm:HNThm\], independent of $K$. The result follows. The above result demonstrates that the seemingly strong assumptions on the function $F(t)$ required in Theorem \[thm:HNThm\] are not a big restriction: by allowing a slightly larger constant, we can consider the class of functions which can be approximated by such functions uniformly. As we will see, the exact smoothness requirements on $f(t)$ are that $f(t)$ has a Hölder continuous second derivative (although we believe that this assumption could be relaxed). Our arguments rely on the construction of a barrier, and the proof of optimality described in [@CoxWang:11]. We recall some important definitions here. We suppose that we are given a barrier function $R_T(x)$ as defined in , and consider the geometric Brownian motion $({\widetilde}{X}_t)$ on this domain, with corresponding hitting time $\tau_{D_T}$, where $D_T$ is as defined in . Then we define the function $$\label{eq:M2defn} M(x,t) = {\mathbb{E}}^{(x,t)}f(\tau_{D_T})$$ and observe that (under the assumptions of Theorem \[thm:HNThm\]), we have $M(x,t)$ bounded. Since we consider the case where ${\widetilde}{X}$ is geometric Brownian motion, we can assume $\sigma(x) = x$ in the formulae from [@CoxWang:11]. Now define a function $Z(x)$ by: $$\label{eq:Zdefn} Z(x) = 2 \int_{S_0}^x \int_{S_0}^y \frac{M(z,0)}{z^2} {\,\mathrm{d}}z {\,\mathrm{d}}y.$$ So in particular, we have $Z''(x) = 2\frac{M(x,0)}{x^2}$ and $Z(x)$ is a convex function. Define also: $$\label{eq:Gdefn} G(x,t) = \int_0^t M(x,s) {\,\mathrm{d}}s - Z(x),$$ and $$\label{eq:Hdefn} H(x) = \int_0^{R_T(x)} (f(s) - M(x,s)) \, {\mathrm{d}s}+ Z(x).$$ Then (@CoxWang:11 [Proposition 5.1]) for all $(x,t) \in {\mathbb{R}}_+ \times {\mathbb{R}}_+$: $$\label{eq:GFineq} G(x,t) + H(x) \le F(t)$$ with equality when $t = R_T(x)$. In addition, if for any $T > 0$: $$\label{eq:qvbddcond} {\mathbb{E}}\left[ \int_0^T Z'({\widetilde}{X}_s)^2 \sigma({\widetilde}{X}_s)^2 {\,\mathrm{d}}s \right] < \infty, \quad {\mathbb{E}}Z({\widetilde}{X}_0) < \infty,$$ then the process $$\label{eq:Gmart} G({\widetilde}{X}_{t\wedge \tau_{D_T}},t \wedge \tau_{D_T}) \text{ is a martingale,}$$ and $$\label{eq:Gsubmart} G({\widetilde}{X}_{t},t) \text{ is a submartingale.}$$ We collect some useful properties of these functions in the following lemma: \[lem:PropofFn\] Under the assumptions of Theorem \[thm:HNThm\], the functions $Z(x), H(x)$ and $G(x,t)$ as defined above have the following properties: 1. \[item:2\] $|{\frac{\partial G}{\partial t}}(x,t)| \le M^*$ for all $(x,t) \in {\mathbb{R}}_+\times {\mathbb{R}}_+$. 2. \[item:3\] $G({\widetilde}{X}_t,t)$ is a submartingale, with decomposition: $$\begin{aligned} G({\widetilde}{X}_t,t) & = G({\widetilde}{X}_0,0) + \int_0^t {\widetilde}{X}_s \left(\int_0^{R_T({\widetilde}{X}_s)\wedge t}{\frac{\partial M}{\partial x}}({\widetilde}{X}_s,r) \, {\,\mathrm{d}}r -Z'({\widetilde}{X}_s)\right) \, {\,\mathrm{d}}{\widetilde}{X}_s \\ & \quad \quad {} - \int_0^t \gamma({\widetilde}{X}_s) {\boldsymbol{1}_{\{\ensuremath{s>R_T({\widetilde}{X}_s)}\}}} \, {\mathrm{d}s}, \end{aligned}$$ where $$\gamma(x) = f(R_T(x))-{\dfrac{1}{2}}x^2 {\frac{\partial M}{\partial x}}(x,R_T(x)-) R_T'(x) \ge 0$$ is a bounded function. Let $v_t$ be the squared volatility process for the Heston-Nandi price process $S_t$, and suppose we fix $T>0$ (although we will want our constants to be independent of $T$). Define the time-change process $\tau_T = \int_0^T v_t \, {\mathrm{d}t}$, and let $A_t$ be the right-inverse of $\tau_t$. In particular, if we define as usual ${\widetilde}{X_t} = {\mathrm{e}}^{-rA_t} S_{A_t}$, then ${\widetilde}{X}_t$ is a geometric Brownian motion with fixed law $\mu_T$ at time $\tau_T$. Using in we get: $$\label{eq:RTformula} R_T({\widetilde}{X}_{\tau_T}) = \int_0^T v_s \, {\mathrm{d}s}+\frac{1}{\xi} \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right)^{-1} (v_T-v_0).$$ Since the variance process $v_s$ is mean reverting ${\mathbb{E}}|v_T-v_0|$, can be bounded uniformly for all $T$ by some constant depending only on the parameters of the model and so in particular, there exists a constant $\kappa_1$ such that $${\mathbb{E}}|R_T({\widetilde}{X}_{\tau_T}) - \tau_T| < \kappa_1.$$ From the bound on $f(t)$, it then follows that: $$\label{eq:Fbound} {\mathbb{E}}[|F(\tau_T)-F(R_T({\widetilde}{X}_{\tau_T}))|] \le M^* \kappa_1.$$ Similarly, using Lemma \[lem:PropofFn\].\[item:2\], we get $${\mathbb{E}}[|G({\widetilde}{X}_{\tau_T},\tau_T)-G({\widetilde}{X}_{\tau_T}, R_T({\widetilde}{X}_{\tau_T}))|] \le M^* \kappa_1.$$ In addition, using the decomposition from Lemma \[lem:PropofFn\].\[item:3\], and noting that $\gamma(x)$ is bounded above by a constant, $\kappa_2$ say, we have: $${\mathbb{E}}G({\widetilde}{X}_{\tau_T},\tau_T) \le G({\widetilde}{X}_0,0) + \kappa_2{\mathbb{E}}\left[\int_0^{\tau_T} {\boldsymbol{1}_{\{\ensuremath{s > R_T({\widetilde}{X}_s)}\}}} \, {\mathrm{d}s}\right].$$ Observe from the definition of $\tau_t$, and evaluated at a general time $t = A_s$: $$\begin{aligned} \{ R_T({\widetilde}{X}_s) < s \} & = \left\{ \left( \frac{\kappa}{\xi} + {\dfrac{1}{2}}\right)^{-1} \left[ \frac{\kappa\theta}{\xi} (T-A_s) + \frac{1}{\xi} \left(v_{A_s}-v_0\right)\right] \le 0 \right\} \\ & = \left\{A_s \ge \frac{\kappa\theta T + v_{A_s} - v_0}{\kappa\theta}\right\} \subseteq \left\{A_s \ge \frac{\kappa\theta T - v_0}{\kappa\theta}\right\}. \end{aligned}$$ Hence $$\begin{aligned} {\mathbb{E}}\left[\int_0^{\tau_T} {\boldsymbol{1}_{\{\ensuremath{s > R_T({\widetilde}{X}_s)}\}}} \, {\mathrm{d}s}\right] & \le {\mathbb{E}}\left[ \int_0^{\tau_T} {\boldsymbol{1}_{\{\ensuremath{A_s \ge T-\frac{v_0}{\kappa\theta}}\}}} \, {\mathrm{d}s}\right] = {\mathbb{E}}\left[ \int_{\tau_{T-\frac{v_0}{\kappa\theta}}}^{\tau_T} \, {\mathrm{d}s}\right] \\ & = {\mathbb{E}}\left[ \tau_T - \tau_{T-\frac{v_0}{\kappa\theta}} \right] = {\mathbb{E}}\left[ \int_{T-\frac{v_0}{\kappa\theta}}^{T} v_s \, {\mathrm{d}s}\right]. \end{aligned}$$ Again, since $v_s$ is mean reverting, the right-hand-side can be bounded independently of $T$, and so $$\label{eq:Gbound} {{\mathbb{E}}\left[G({\widetilde}{X}_{\tau_T},\tau_T)\right]} \le G({\widetilde}{X}_0,0) + \kappa_2$$ for some constant $\kappa_2$. Now, using and the fact that holds with equality when $t=R_T(x)$, we have: $$\begin{aligned} {\mathbb{E}}\left[ F(\tau_T) \right] & \le {\mathbb{E}}\left[ F(R_T({\widetilde}{X}_{\tau_T}))\right] + M^*\kappa_1\\ & = {\mathbb{E}}\left[G({\widetilde}{X}_{\tau_T},R_T({\widetilde}{X}_{\tau_T}))\right] + {{\mathbb{E}}\left[H({\widetilde}{X}_{\tau_T})\right]} + M^*\kappa_1\\ & = {\mathbb{E}}\left[G({\widetilde}{X}_{\tau_T},\tau_T)\right] + {{\mathbb{E}}\left[H({\widetilde}{X}_{\tau_T})\right]} + 2M^*\kappa_1\\ & \le G({\widetilde}{X}_0,0) + {{\mathbb{E}}\left[H({\widetilde}{X}_{\tau_T})\right]} + 2M^*\kappa_1+\kappa_2 \end{aligned}$$ It remains for us to show that ${{\mathbb{E}}\left[F({\widetilde}{X}_\sigma)\right]} \ge G({\widetilde}{X}_0,0) + {{\mathbb{E}}\left[H({\widetilde}{X}_{\tau_T})\right]} = G({\widetilde}{X}_0,0) + {{\mathbb{E}}\left[H({\widetilde}{X}_{\sigma})\right]}$ for any stopping time $\sigma$ with ${\widetilde}{X}_\sigma \sim {\widetilde}{X}_{\tau_T}$. We consider a localising sequence, $\sigma_N \uparrow \sigma$, and note that we then have: ${{\mathbb{E}}\left[G({\widetilde}{X}_{\sigma_N},\sigma_N)\right]} \ge G({\widetilde}{X}_0,0)$, since $G$ is a submartingale, and in addition, ${{\mathbb{E}}\left[F(\sigma_N)\right]} \uparrow {{\mathbb{E}}\left[F(\sigma)\right]}$ since $F(\cdot)$ is increasing. On account of , it remains only to show ${{\mathbb{E}}\left[H({\widetilde}{X}_{\sigma})\right]} \le {{\mathbb{E}}\left[H({\widetilde}{X}_{\sigma_N})\right]}$. We first observe that $f$ is an increasing and bounded function, and if $f(t) = f(\infty)$ for all $t \ge t_0$, for some $t_0\in {\mathbb{R}}_+$, $M(x,t) = f(t)$ for all $t \ge t_0$. Since also $Z(x) \ge 0$ ($Z(x)$ is convex with $Z(S_0) = Z'(S_0) = 0$), we must have $H(x)$ bounded below. We can therefore apply Fatou’s Lemma to deduce ${{\mathbb{E}}\left[H({\widetilde}{X}_{\sigma})\right]} \le {{\mathbb{E}}\left[H({\widetilde}{X}_{\sigma_N})\right]}$. To remove the assumption on $f(t)$, we observe that, by smoothly truncating $f$, we can approximate $F$ from below by an increasing sequence $F_N$ of functions which each have constant derivative, and such that ${{\mathbb{E}}\left[F_N({\widetilde}{X}_{\tau_T})\right]} \uparrow {{\mathbb{E}}\left[F({\widetilde}{X}_{\tau_T})\right]}$. Since each approximation satisfies the bound, the same must be true in the limit. We first show \[item:2\]. Observe that $M(x,R_T(x)) = f(R_T(x))$, so in particular, $M$ is continuous, and $f(t) \le M(x,t) \le M^*$ since $M(x,t)$ is increasing in $t$. It follows immediately that ${\frac{\partial G}{\partial t}}(x,t) = M(x,t)$ is continuous and bounded, and in fact, is non-negative. For \[item:3\] we aim to use @Peskir:05 [Theorem 3.1]. We note that $M(x,t)$ is $C^{2,1}$ in ${D_T}$ since it is a martingale (in particular, $M$ is the (unique bounded) solution to a parabolic initial-value boundary problem). In fact, by @Lieberman:96 [Theorem 5.14], if we assume that $f''(t)$ is bounded in a Hölder norm, it follows that $M(x,t)$ has Hölder-bounded first and second spatial derivatives, and first time derivative. It is easy to check that $G(x,t) = \int_0^t M(x,s) \, {\mathrm{d}s}- Z(x)$ is also $C^{2,1}$ in ${D_T}$ as a consequence. Moreover, computing explicitly, we see that ${\frac{\partial G}{\partial x}}(x,t) = \int_0^t {\frac{\partial M}{\partial x}}(x,s) \, {\mathrm{d}s}- Z'(x)$ is also continuous on $\bar{D}_T$. If we write $C = \{(x,t) : R_T(x) < t\}$, so $\bar{C} \cup \bar{D}_T = {\mathbb{R}}_+ \times {\mathbb{R}}_+$ with boundary $R_T(x) = t$, we see that on $C$, $$G(x,t) = \int_0^{R_T(x)} M(x,s) \, {\mathrm{d}s}+ \int_{R_T(x)}^t f(s) \, {\mathrm{d}s}- Z(x)$$ and again, $G$ is $C^{2,1}$ in $C$, and ${\frac{\partial G}{\partial x}}(x,t)$ is continuous on $\bar{C}$. Considering the conditions required for Theorem 3.1 of [@Peskir:05], we observe that (3.18), (3.19), (3.26), (3.30) and (3.33) of [@Peskir:05] have now been shown, and so the theorem holds. Moreover, since the first spatial derivative of $G$ is continuous across the boundary, we do not get a local-time term on the boundary. Computing ${\frac{\partial G}{\partial t}}(x,t) + {\dfrac{1}{2}}\sigma(x)^2 {\frac{\partial ^2G}{\partial x^2}}(x,t)$ results in the expression for $\gamma(x)$ stated, and we observe that the boundary function $R_T(x)$ is a decreasing function of $x$, which implies in turn that ${\frac{\partial M}{\partial x}}(x,t)$ is positive at the boundary (since $f$ is increasing), so $\gamma(x)\ge0$. We finally show that $\gamma(x)$ is bounded. By assumption, $f$ is bounded, so we need only consider the second term. We have $M(x,R_T(x)) = f(R_T(x))$, and differentiating (recall that the derivatives of $M(x,t)$ on ${D_T}$ are Hölder continuous, and so extend continuously to the boundary) and rearranging we get: $${\frac{\partial M}{\partial x}}(x,R_T(x)) = R_T'(x) \left( f'(R_T(x))- {\frac{\partial M}{\partial t}}(x,R_T(x)) \right)$$ Observing that (via a standard coupling argument) $${{\mathbb{P}}}^{(x,t)}(\tau_{D_T} > t) \ge {{\mathbb{P}}}^{(x,t+\delta t)}(\tau_{D_T} > t+\delta t)$$ whenever $(x,t) \in {D_T}$ (the later path sees a ‘bigger’ stopping region). It follows that ${\mathbb{E}}^{(x,t+\delta t)} \left[ f(\tau_{D_T}) \right] \le {\mathbb{E}}^{(x,t)} \left[ f(\tau_{D_T}+\delta t) \right]$, and therefore that ${\frac{\partial M}{\partial t}}(x,t) \le \sup_{t \in {\mathbb{R}}_+} f'(t) \le M^*$. Recalling finally that $\sigma(x) = x$, and observing that $R_T'(x) = - ( \kappa/\xi + 1/2)^{-1} x^{-1}$, we conclude that $\gamma(x)$ is indeed bounded. ![The upper and lower model-independent bounds on the price of a variance call, plotted as functions of the correlation between the asset and volatility processes. The constant line represents the price of the variance call under the Heston model — this is constant, since the price is unaffected by the choice of $\rho$.[]{data-label="fig:RhoBehaviour"}](Heston_Var_Call_Bounds_in_rho.eps){width=".68\textwidth"} We finish with some numerical evidence to support our conjecture. Figure \[fig:RhoBehaviour\] shows the upper and lower bounds on the price of a variance call, as determined using the numerical methods of Section \[sec:numerical-results\], seen as a function of the parameter $\rho$. In this example, we use the same parameters as before, but with $T=4$. It is notable that the lower bound and the price arising from the Heston model are certainly close. It is also interesting to observe that it is not only the Heston-Nandi model that seems to be close to extremal; rather this seems to be a more general property of the Heston model. A good explanation of this fact eludes us, but better understanding of this behaviour would appear to be both practically relevant, and theoretically interesting. [^1]: Department of Mathematical Sciences, University of Bath, Bath, U. K.. e-mail: `a.m.g.cox@bath.ac.uk`, web: `http://www.maths.bath.ac.uk/\simmapamgc/` [^2]: Dipartimento di Metodi e Modelli per l’Economia, il Territorio e la Finanza, Università di Roma ‘La Sapienza’, Italy. e-mail: `jiajie.wang@uniroma1.it` [^3]: We note that Chacon calls what we refer to a reversed barrier as a barrier. We follow the terminology established in @Obloj:04. [^4]: This would appear to be a very strong assumption on $G$. However, along the lines of [@DavisOblojRaval:10], it seems reasonable that the conclusions would hold in a milder sense; what seems harder is to both provide a set of conditions under which these conclusions hold, and which can be verified under relatively natural constraints on our modelling setup.

ArXiv

--- author: - 'L. Foschini, E. Pian, L. Maraschi, C.M. Raiteri, F. Tavecchio, G. Ghisellini, G. Tosti, G. Malaguti, G. Di Cocco' date: 'Received 22 November 2005; Accepted 5 January 2006' title: 'A Short Hard X-ray Flare from the Blazar NRAO 530 Observed by $INTEGRAL$[^1]' --- Introduction ============ Blazars are, among active galactic nuclei (AGN), the most luminous and most dramatically variable. These extreme properties are likely due to relativistic aberration in a kilo-parsec jet oriented at a small angle with respect to our line of sight, where plasma moving with a Lorentz factor of $\sim$10-20 causes radiation boosting and time foreshortening (see reviews by Urry & Padovani 1995, Wagner & Witzel 1995, Ulrich et al. 1997). In these sources, the maximum power output and the largest variability amplitudes are observed at high-energies, from X- to $\gamma-$rays, therefore blazars are interesting targets for monitoring with satellites like *INTEGRAL* (Pian et al. 1999). During its first three years of activity *INTEGRAL* detected two blazars in outburst, with observations triggered by ground telescopes (S5 $0716+714$ Pian et al. 2005, 3C$454.3$ Foschini et al. 2005, Pian et al., in preparation), and one in high state (S5 $0836+710$), serendipitously detected during the observation of S5 $0716+714$ (Pian et al. 2005). Indeed, thanks to the large field-of-view of the IBIS instrument onboard INTEGRAL, many blazars, possibly in active state, may be detected serendipitously by the satellite during its long pointings. In order to exploit this advantage, we have systematically searched the *INTEGRAL* public archive for possible detection of flaring blazars. We report here about the interesting detection of a rapid flare apparently associated with the blazar NRAO $530$. NRAO $530$ ($z=0.902$) is a blazar belonging to the optically violent variable (OVV) quasar subclass, with emission lines in the optical spectrum, particularly H$\beta$ with a rest-frame equivalent width of $16\AA$, smaller than usually observed in OVVs (Junkkarinen 1984; cf also Veron-Cetty & Veron 2000). It exhibits strong outbursts in radio (Bower et al. 1997), optical (up to $3$ magnitudes in $1977$, Pollock et al. 1979, Webb et al. 1988), X-ray (*HEAO-1*, Marscher et al. 1979), and $\gamma-$ray bands (EGRET/CGRO, Mukherjee et al. 1997), with significant emission above $1$ GeV (Lamb & Macomb 1997). The flare detected by the IBIS/ISGRI instrument on board *INTEGRAL* occurred in February $2004$ with a timescale ($\approx 2000$ s) much shorter than any variability reported before. In the following, we present in detail the *INTEGRAL* data analysis and results (Section 2), data obtained quasisimultaneously in the radio (Section 3) and a discussion of the relevance of the detected flare for blazars.  $INTEGRAL$ data analysis ======================== The *INTEGRAL* satellite (Winkler et al. 2003) was launched on $17$ October $2002$ and it carries two main instruments for the $\gamma-$ray astrophysics, the imager IBIS ($0.02-10$ MeV, Ubertini et al. 2003) and the spectrometer SPI ($0.02-8$ MeV, Vedrenne et al. 2003), plus two monitors, JEM-X ($3-35$ keV, Lund et al. 2003) and OMC (V filter, Mas-Hesse et al. 2003). Since NRAO $530$ is located close to the Galactic Centre ($l=12^{\circ}.0$, $b=+10^{\circ}.8$), a very crowded region, we do not use the data from the spectrometer SPI, since it has an angular resolution of about $2^{\circ}.5$. Also the OMC is strongly limited by the source confusion: its point-spread function (PSF) is $25''$ (FWHM). The US Naval Observatory Catalog (USNO B1, Monet et al. 2003) reports $12$ sources of brightness comparable to that of NRAO $530$ inside a circular region of $25''$ radius centered around the position of the blazar. Thus, the OMC observations are presumably highly contaminated, and therefore not used. We also checked the JEM-X data for possible detection, however the source was never optimally located in the monitor field-of-view: in only one of the 11 pointings the source angular distance from the FOV centre was $2^{\circ}.5$, less than the limit of $<3^{\circ}$ required for good JEM-X performance. Using the INTEGRAL pointing[^2] where the source is detected by IBIS (see below), which also corresponds to the minimum distance from the JEM-X FOV center, we determine an upper limit of $30$ mCrab in the $5-15$ keV energy band ($5\sigma$ level, as suggested in Westergaard et al. 2005). Therefore, in the following we will present only the analysis of the data acquired by IBIS, that is the *INTEGRAL* instrument with best sensitivity in the hard X-ray range. The analysis has been performed using the Offline Scientific Analysis (OSA) software package[^3], whose algorithms for the reduction and treatment of the IBIS instrument data are described in Goldwurm et al. (2003). The imager IBIS (Ubertini et al. 2003) is composed by two detectors, ISGRI, sensitive to the radiation between $20$ keV and $1$ MeV (Lebrun et al. 2003), and PICsIT ($175$ keV to $10$ MeV, Di Cocco et al. 2003), both coupled with the same tungsten coded mask. The latter has the pattern of a modified uniformly redundant array (MURA, Gottesman & Fenimore 1989), with a basic configuration of $53\times 53$ square pixels $11.2$ mm sized. The mask is separated from the ISGRI layer by about $3200$ mm, so that the angular distance of ghosts in the system point spread function (SPSF) of ISGRI is about $10^{\circ}.5$ (Gros et al. 2003). This means that sources with reciprocal angular distances of that size can be confused and specific care is necessary in order to disentangle the different contributions. On February 17, $2004$ (orbit $164$), while *INTEGRAL* was executing the Galactic Centre Deep Exposure (GCDE), ISGRI detected a signal during the time interval between $00^{\rm h}:40^{\rm m}:05^{\rm s}$ and $01^{\rm h}:09^{\rm m}:18^{\rm s}$ UTC at $\alpha=17^{\rm h}:33^{\rm m}:04^{\rm s}$ and $\delta=-13^{\rm h}:02^{\rm m}:50^{\rm s}$ (J2000). The $3'-$radius uncertainty region associated with this centroid is consistent with the blazar NRAO $530$ ($\alpha=17^{\rm h}33^{\rm m}02^{\rm s}.71$, $\delta=-13^{\circ}04'49''.5$, J2000) and one ghost of the SPSF of the neutron star GX $17+2$ ($\alpha=18^{\rm h}16^{\rm m}01^{\rm s}.40$, $\delta=-14^{\circ}02'11''.0$, J2000). The angular distance between the two sources is precisely $10^{\circ}.5$, so that the two SPSF overlap (Fig. \[isgrimap\]). However, one can disentangle the contribution of the two sources by analysing their time evolution: if one SPSF is contaminated by the other, both sources should show the same time behaviour.   Since the latest version of `OSA 5.0` for IBIS/ISGRI removes the sources located at the potential positions of the ghosts of brightest sources, we used an older version (`OSA 3.0`), where this feature was not yet implemented. We also cross-checked the consistency of the results obtained with `OSA 3.0` with `OSA 5.0`, however, in the latter case, we adopted an input source catalog containing only NRAO $530$ as a source, to by-pass the above protection. Obviously, we also checked that the results on GX $17+2$ were not affected when adopting this solution, by performing a run with `OSA 5.0` and the full input catalog. We compared the behaviour of the background and of the 2 sources by analyzing the 5 pointings before and after the event, corresponding to the time interval between February $16^{\rm th}$, $2004$ at $21^{\rm h}:59^{\rm m}:08^{\rm s}$, and February $17^{\rm th}$, $2004$ at $03^{\rm h}:49^{\rm m}:44^{\rm s}$ UTC. We normalized the count rates of the background and of GX $17+2$ in the $20-40$ keV energy band to their weighted averages. For NRAO $530$, which has one detection only, we left the count rates unchanged. Fig. \[fig1\] shows the lightcurves of the background, GX $17+2$, and NRAO $530$. While GX $17+2$ follows the background time evolution (Fig. \[fig1\], left), this is not the case for NRAO $530$ (Fig. \[fig1\], right). Were the putative detection of NRAO $530$ only a ghost of GX $17+2$, then it would follow the same temporal behaviour, which would contradict the derived upper limits (see also Fig. \[isgrimap\]). No signal was detected in any other pointing at that position during the GCDE, nor during any other INTEGRAL observation of this sky region. The difference in the upper limits on the NRAO $530$ count rate in the various pointings (Fig. \[fig1\], right) is due to the different location of NRAO $530$ with respect to the ISGRI FOV: the farther the source from the FOV center, the more shallow its upper limit, due to the radially decreasing sensitivity of the ISGRI detector. Note that the upper limits obtained in the $\sim$2 hours bracketing the NRAO 530 detection indicate that the flux change is highly significant. Therefore, the hypothesis that the source was constant and could only be detected when sufficiently close to the ISGRI FOV center is ruled out. We conclude that the detection of the rapid variation is genuine. The total time on source is $1745$ s, that can be considered the minimum time scale of the event. It is worth mentioning that *INTEGRAL* performs two slews, one before and the other after the pointing, each $\approx 120$ s long. It is not possible to detect the source during a slew in a coded-mask instrument (except for very bright cases, like GRB) and therefore it is not possible to say if NRAO $530$ was flaring already during the slews. Taking into account also the marginal detection in the pointing following the central one the maximum time scale of the event is $\approx 1 $ h. We searched the SIMBAD catalog for Galactic hard X-ray sources within a $6'$ radius (twice the ISGRI error radius for a source with this significance, see Gros et al. 2003) of the position of NRAO $530$, that could possibly contaminate our ISGRI detection, but found none. Therefore, we attribute the flux detection entirely to NRAO $530$. The count rates of NRAO $530$ in the $20-40$ keV energy band, already corrected for systematics, are: $2.7\pm 0.5$ c/s when using `OSA 3`, that correspond to $26\pm 5$ mCrab[^4]. When using `OSA 5`, the rate is $3.2\pm 0.6$ c/s, corresponding to $30\pm 6$ mCrab, consistent with the measurement with `OSA 3`. We take a weighted average as reference: $28\pm 3$ mCrab. By adopting the standard spectrum of the Crab as reported by Toor & Seward (1974)[^5], the converted flux is $(2.1\pm 0.2)\times 10^{-10}$ erg cm$^{-2}$ s$^{-1}$. No detection has been found in PICsIT, the high-energy layer of IBIS. Radio data ========== We searched also for quasi-simultaneous data at other wavelengths and found only a radio observation at $2$ cm (VLBA), made on $11$ February $2004$, a few days before the *INTEGRAL* detection, within the MOJAVE project[^6] (Lister & Homan 2005). The flux density was $4.1$ Jy, with $3.1$% linear polarization. The previous observation dates back to $9$ October $2002$, with a higher flux ($5.7$ Jy) but lower polarization ($0.5$%). In 2005, lower fluxes and linear polarization have been observed. These data can be compared with the historical radio observations by Bower et al. (1997): as noted already by those authors there is no apparent correlation between the radio flux density and the optical activity in $1977-1978$. NRAO $530$ generally shows fluxes from a few up to more than $10$ Jy, depending on the radio frequency. Therefore, the MOJAVE 2004 measurement of $4.1$ Jy could indicate a low activity phase. ![Spectral energy distribution of NRAO $530$ updated with the *INTEGRAL* data of the present work. Radio data from NED and Bower et al. (1997). Reference value of optical observations from Hewitt & Burbidge (1993); historical maximum from Pollock et al. (1979), Webb et al. (1988). X-rays: *ROSAT* (Comastri et al. 1997), *Einstein* (Marscher & Broderick 1981), *HEAO-1* (Marscher & Broderick 1981). $\gamma-$rays: *CGRO*/EGRET (Hartman et al. 1999). Blue cross: IBIS/ISGRI, present work.[]{data-label="fig2"}](4804_f4.ps) The polarization measurements reported by Bower et al. (1997) show an increase (up to about $8$% at $8$ GHz) during the two years preceding the optical outburst in $1977-1978$. In our case, the MOJAVE measurement of a moderately high polarization value some days before the hard-X flare could be linked to the latter. We note that the present episode is a rapid event, not a long outburst like in $1977-1978$, therefore the fact that only moderate polarization was observed is not inconsistent with the fast flare in hard X-rays. Discussion ========== Hard X-ray emission from the blazar NRAO $530$ was detected serendipitously by *INTEGRAL* with IBIS/ISGRI during a long exposure of the Galactic Center. During a time interval lasting $\approx 2000$ s the source flux in the $20-40$ keV range rose above the detection threshold by about a factor of $\sim 2$ and then faded again. The spectral energy distribution of NRAO $530$ constructed with historical data and with the present *INTEGRAL* flux is reported in Fig. \[fig2\]. Interestingly, although very high, the IBIS/ISGRI point is consistent with the extrapolation to lower energies of the average spectrum observed in the EGRET energy range. The variability amplitude measured by EGRET was a factor $6$ and rather erratic, still on timescales of weeks (Mukherjee et al. 1997). On the other hand, timescales as short as implied here would not have been accessible for EGRET due to the low counting rate. The hard X-ray flux measured during the flare for any reasonable spectral shape would imply a much larger X-ray flux than observed previously. Large variations have been observed from this object also in the optical (more than a factor 10 over timescales of years), thus overall the amplitude of variability is not unprecedented. What is truly exceptional in this event is the short timescale, less than an hour in an object of extremely high luminosity. This may cast doubts on the association with NRAO $530$. The event could then be attributed to some still unknown galactic source inside the $3'$ radius circle of positional uncertainty in the ISGRI detection. In this case, only future satellites, like SIMBOL-X (Ferrando et al. 2005), will have the necessary spatial resolution at hard X-rays to confirm the association with NRAO $530$. For the moment, with the present data, instrument performances, and catalogs, it is interesting to consider the possibility that this highly significant variation is associated with the blazar. Rapid flares, with timescales of thousands of seconds in the X-rays, are often observed in BL Lac sources, especially in the TeV sources (e.g. Brinkmann et al. 2005), but they are rarely seen in blazars belonging to the quasar class, where typical timescales, even in hard X-ray–$\gamma$-ray band, are in the range of several hours–days (e.g. Wagner & Witzel 1995). It is worth noting that FSRQ are generally faint in X-rays and therefore the above assertion could be biased in the past by the lack of high-energy astrophysics satellites with the necessary sensitivity. The present and the next generation of X- and $\gamma-$ray satellites should remove this bias. The interpretation of such a rapid flare in the context of the standard synchrotron-inverse Compton models for blazars (e.g. Ghisellini et al. 1998) requires quite extreme conditions for the emitting source. A flare lasting less than $\Delta t=3600$ s must be emitted by a region with a size $R < 1.1\times 10^{14} \delta$ cm, where $\delta$ is the Doppler factor. Adopting the value of $\delta=15$, suggested by VLBI observations (Bower & Backer 1998), the size of the source is $R< 1.6 \times 10^{15}$ cm, an order of magnitude less than dimensions typically estimated for blazars ($10^{16}-10^{17}$ cm). In the case of a mass of the central supermassive black hole of $10^{8}M_{\odot}$, the above distance would correspond to only about $50$ times the radius of the innermost stable orbit. A direct alternative is to admit a large value of the Doppler factor, $\delta \sim 100$. Such large values of $\delta$ are sometimes invoked to explain the rapid intra day variability observed in the radio band (e.g. Wagner & Witzel 1995). In this case, it would be possible to explain the short flare in terms of unsteadyness of the jet flow, due to a single non stationary shock (Hughes et al. 1985) possibly induced by a collision of two relativistic plasma shells in the jet (internal shock, Spada et al. 2001). The event discussed here could also result from an internal shock developping very close to the origin of the jet with less extreme values of $\delta$. The moderate polarization observed at radio wavelengths is consistent with these scenarios, since it arises from regions much further out in the jet. Although the data available are few and sparse, it is possible that the present observation represents the “tip of an iceberg” leading the way to discover more short time scale events. We plan to continue our search of variability of blazars in the *INTEGRAL* field of view. LF wishes to thank A. Domingo Garau for useful discussion about OMC data. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration, of data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC), provided by NASA’s Goddard Space Flight Center, and of data from the Monitoring of Jets in AGN with VLBA Experiments (MOJAVE) Project. This work was partly supported by the European Community’s Human Potential Programme under contract HPRN-CT-2002-00321 (ENIGMA). *Note added in proof*. After acceptance of the manuscript, the Swift satellite performed two observations of NRAO530 (on 10 and 14 February 2006). During both observations, the X-Ray Telescope (XRT) detected only the blazar NRAO 530 within the ISGRI error circle of $3'$. This result strengthens our conclusion that the exceptional flare detected by INTEGRAL is very likely due to NRAO 530. Brinkmann W., Papadakis I.E., Raeth C., Mimica P., Haberl F., 2005, A&A 443, 397 Bower G.C., Donald C.B., Wright M., et al., 1997, ApJ 484, 118 Bower G.C. & Backer D.C., 1998, ApJ 507, L117 Comastri A., Fossati G., Ghisellini G., Molendi S., 1997, ApJ 480, 534 Di Cocco G., Caroli E., Celesti E., et al., 2003, A&A 411, L189 Ferrando P., Goldwurm A., Laurent P., et al., 2005, Proceedings of SPIE 5900, 195 Foschini L., Di Cocco G., Malaguti G., et al., 2005, ATel 497 Ghisellini G., Celotti A., Fossati G., Maraschi L., Comastri A., 1998, MNRAS 301, 451 Goldwurm A., David P., Foschini L., et al., 2003, A&A 411, L223 Gottesman S.R., Fenimore E.E., 1989, Appl. Opt. 28, 4344 Gros A., Goldwurm A., Cadolle-Bel M., et al. 2003, A&A 411, L179 Hartman R.C., Bertsch D.L., Bloom S.D., et al., 1999, ApJS 123, 79 Hewitt A. & Burbidge G., 1993, ApJS 87, 451 Hughes P.A., Aller H.D., Aller M.F., 1985, ApJ 298, 301 Jorstad S.V., Marscher A.P., Mattox J.R., et al., 2001, ApJ 556, 738 Junkkarinen V., 1984, PASP 96, 539 Lamb R.C. & Macomb D.J., 1997, ApJ 488, 872 Lebrun F., Leray J.P., Lavocat P., et al., 2003, A&A 411, L141 Lister M.L. & Homan D.C., 2005, AJ 130, 1389 Lund L., Budtz-Jørgensen C., Westergaard N.J., et al., 2003, A&A 411, L231 Marscher A.P., Marshall F.E., Mushotzky R.F. et al., 1979, ApJ 233, 498 Marscher A.P. & Broderick J.J., 1981, ApJ 249, 406 Mas-Hesse J.M., Giménez A., Culhane J.L., et al., 2003, A&A 411, L261 Monet D.G., Levine S.E., Casian B., et al., 2003, AJ 125, 984 Mukherjee R., Bertch D.L., Bloom S.D., et al., 1997, ApJ 490, 116 Pian E., Malaguti G., Maraschi L., Ghisellini G., Palumbo G.G.C., 1999, ApL&C 39, 137 Pian E., Foschini L., Beckmann V., et al., 2005, A&A 429, 427 Pollock J.T., Pica A.J., Smith A.G., et al., 1979, AJ 84, 1658 Spada M., Ghisellini G., Lazzati D., Celotti A., 2001, MNRAS 325, 1559 Toor A. & Seward F.D., 1974, ApJ 79, 995 Ubertini P., Lebrun F., Di Cocco G., et al., 2003, A&A 411, L131 Ulrich M.-H., Maraschi L., Urry C.M., 1997, ARA&A 35, 445 Urry, C.M. & Padovani, P. 1995, PASP, 107, 803 Vedrenne G., Roques J.-P., Schönfelder V., et al., 2003, A&A 411, L63 Veron-Cetty M.P. & Veron P., 2000, A&A Rev. 10, 81 Wagner S.J., Witzel A., 1995, ARA&A 33, 163 Webb J.R., Smith A.G., Leacock R.J., et al., 1988, AJ 95, 374 Westergaard N.J., Oxborrow C.A., Chevenez J., et al., 2005, JEM-X Science Analysis: Scientific Validation Report v 5.0, Danish National Space Center Winkler C., Courvoisier T.J.-L., Di Cocco G., et al., 2003, A&A 411, L1 [^1]: Based on observations obtained with *INTEGRAL*, an ESA mission with instruments and science data centre funded by ESA member states (especially the PI countries: Denmark, France, Germany, Italy, Switzerland, Spain), Czech Republic and Poland, and with the participation of Russia and the USA. [^2]: Every single pointing is called “Science Window” (ScW) in the *INTEGRAL* jargon. A ScW has a duration of about 2000 seconds. [^3]: Available at `http://isdc.unige.ch/`. [^4]: We assumed $1~{\rm Crab} = 105$ c/s in the $20-40$ keV energy band of the IBIS/ISGRI detector. [^5]: Power Law model with $\Gamma=2.1$ and normalization $9.7$ ph cm$^{-2}$ s$^{-1}$ keV$^{-1}$ at $1$ keV. [^6]: `http://www.physics.purdue.edu/astro/MOJAVE/`

ArXiv

--- abstract: | Valgrind [@Valgrind], and specifically the included tool Memcheck, offers an easy and reliable way for checking the correctness of memory operations in programs. This works in an unintrusive way where Valgrind translates the program into intermediate code and executes it on an emulated CPU. The heavy weight tool Memcheck uses this to keep a full shadow copy of the memory used by a program and tracking accesses to it. This allows the detection of memory leaks and checking the validity of accesses. Though suited for a wide variety of programs, this approach still fails when accelerator based programming models are involved. The code running on these devices is separate from the code running on the host. Access to memory on the device and starting of kernels is being handled by an API provided by the driver being used. Hence Valgrind is unable to understand and instrument operations being run on the device. To circumvent this limitation a new set of wrapper functions have been introduced. These wrap a subset of the CUDA Driver API function that is responsible for (de-)allocation memory regions on the device and the respective memory copy operations. This allows to check whether memory is fully allocated during a transfer and, through the functionality provided by Valgrind, whether the memory transfered to the device from the host is defined and addressable. Through this technique it is possible to detect a number of common programming mistakes, which are very difficult to debug by other means. The combination of these wrappers together with the Valgrind tool Memcheck is being called Cudagrind. address: 'High Performance Computing Center Stuttgart, Nobelstr. 19, 70565 Stuttgart' author: - - bibliography: - 'ParCo-2013.bib' title: 'Cudagrind: A Valgrind Extension for CUDA' --- , CUDA, memcheck, Valgrind, Cudagrind, H4H Introduction {#introduction .unnumbered} ============ Checking of memory operations at runtime is an important tool in ensuring a programs correct behavior. Tools based on frameworks like Valgrind [@Valgrind] or Intel Pin [@IntelPin] and debuggers are valuable assets. A solid understanding of their correct usage can potentially save hundreds of hours spent on locating subtle problems like a misaligned pointer or prematurely freed memory regions. The problems arising from these type of errors include erratic and undefined behavior due to reading of undefined memory regions and crashes due to writing into forbidden memory areas. In the worst case this can lead to serious security concerns, for the user and the operating system alike, when the spurious access happens on memory that has been allocated by different parts of the system. But despite the importance of these type of runtime checks the existing tools do not apply when working with accelerators like CUDA based GPGPUs from NVIDIA. In this case the kernels running on the device operate on dedicated device memory. This device memory features a very high bandwidth that allows exceptional performance boosts for the right problems. But at the same time it is very difficult to debug kernels running on the device, since the memory is not accessible from the host part of the program. In case of CUDA this memory can only be accessed through a certain set of functions offered by the CUDA driver and provided through it’s driver API [^1] or from inside a kernel. Due to this indirect access pattern classical tools, like Valgrind, are unable to work with and on device memory. Tools working on the device, on the other hand, often do not handle the memory transactions between host and device. This is where [**Cudagrind**]{}, the tool presented in this paper, comes into play. It aims at establishing the missing link between host and device memory checking and is powered by the feature rich Valgrind framework. Memory Checking =============== There’s a wide variety of tools available when it comes to ensuring the correctness of a given program. These range from static analysis tools like Lint [@Lint], over tools working at runtime like Valgrind [@Using_Valgrind], to post-mortem analysis offered by debuggers which are invoked after a problem occurs. While these tools can’t offer a full protection against memory errors of any kind, else they would solve the halting problem, they do offer an easy way to increase the confidence into ones own code. Especially with ever growing projects that feature hundreds of thousand lines of code which are impossible to check manually. Valgrind: Heavy Weight Dynamic Binary Instrumentation ----------------------------------------------------- The Valgrind framework offers a novel way for dynamic instrumentation of arbitrary programs. When executing a program with Valgrind a just-in-time ([**JIT**]{}) compiler translates its binary representation into an intermediate representation ([**IR**]{}). This architecture-neutral IR, called VEX, is designed to be executed by a generic x86 processor which is provided and emulated by the Valgrind runtime. The JIT compiler ensures that each machine instruction in the original binary is translated into an equivalent sequence of VEX/IR instructions to be executed on the virtual processor. This approach comes with the downside of needing a working JIT compiler for each architecture Valgrind is being executed on. But it also allows dynamic instrumentation of any program at runtime. Tools build on top of the Valgrind framework, like the well-known memory checker Memcheck, utilize this dynamic instrumentation for various runtime checks. In the case of Memcheck the tool keeps a full shadow copy of the executed program’s memory and registers. This copy is updated every time the program accesses memory regions or (de-)allocates memory. The information being tracked in this shadow copy allows Memcheck to perform checks on every access to any memory region or register. If there is an arbitrary access to unallocated memory or reading access to memory, that has been allocated but never defined, the tool can now print error messages and warnings at runtime. Due to it being a full copy the implemented checks are performed with bitwise precision. The downside of this approach lies in the sever overhead incurred at runtime. The emulation performed by the Valgrind runtime slows a program by a factor of 4 compared to its normal execution. Using a tool like Memcheck can raise this factor up to 100 due to the additional work performed by the tool. Additionally, depending on the type of instrumentation, the memory available to a program can be much smaller as well. For Memcheck the available memory is at least cut in half due to the full shadow copy being generated. Cudagrind: The Missing Link --------------------------- The way Valgrind handles the translation of machine code into the VEX IR poses another problem. The JIT compiler has to be adapted to every new architecture Valgrind is meant to work with. This is true for instructions introduced by new processor architectures, e.g. the AVX instructions introduced with Intel’s Sandy Bridge architecture. But also accelerator based programing models that rely on dedicated devices, like NVDIA’s CUDA GPGPUs or also Intel’s Xeon Phi, are affected. In these cases a precompiled binary, e.g. a CUBIN file, or dedicated intermediate code, e.g. PTX binary code, is being transfered to the accelerator. The execution of these binaries or kernels is then controlled through a set of functions offered by the underlying driver or runtime API. This prevents Valgrind from handling and instrumenting the code inside these kernels. The Valgrind JIT does not understand the CUBIN or PTX code and it does not have access to code running solely on the device. This gap is meant to be closed by the memory checker `cuda-memcheck`, provided with NVIDIA’s CUDA SDK. But despite working well with the actual code running on the device, there’s still a missing link. `Cuda-memcheck` is unable to check the memory transfers happening between host and device, as these are handled through a set of functions accessible only through the CUDA Driver API. Valgrind on the other hand is unable to check the correctness of these operations since the Valgrind runtime does not know about the memory located on the card and whether it’s been previously allocated or not. This is where Cudagrind comes into play. A set of wrappers, based on Valgrind’s wrapping functionality, is provided. It covers all CUDA Driver API functions related to allocation, deallocation and transfer of memory between host and device. The wrappers perform internal bookkeeping of allocated memory on the device, check the CUDA Driver API function parameters at runtime. Additionally definedness checks, provided by Valgrind, are utilized in order to locate errors happening during memory transfers between host and CUDA devices. Cudagrind Internals =================== In case of the CUDA programming model any access to a CUDA enabled device is handled by the CUDA driver. This is especially true for programs written fully on the driver level, but also holds when working on a higher layer in the CUDA stack (Figure \[fig:CUDA-Stack\]). The CUDA runtime, which offers a more streamlined and easier to use access to the device, relies directly on the functions provided by the CUDA driver. So a call to `cudaMemcpy(..)` implicitly leads to a call of any of the `cuMemcpyXtoY(..)` functions provided by the driver. Moving one abstraction layer higher offers the same picture. Any program utilizing a pragma based approach like OpenACC will rely on the functionality offered by the CUDA driver or, implementation dependent, also the CUDA runtime. Finally any program relying on a CUDA library will implicitly rely on the underlying implementation, which in turn relies on the CUDA driver to provide the needed functionality. ![The placement of Cudagrind in the CUDA stack. There is a wrapper for every memory related function on the CUDA driver level. Functions calls from an application on a higher level, e.g. a CUDA library function, are handled when the respective driver function is called from the library or runtime level.[]{data-label="fig:CUDA-Stack"}](Images/Cudagrind_Wrapper_Stack){width=".75\textwidth"} So no matter the programing model or abstraction layer used, every program relies on the CUDA driver. Since this is also true for any device memory related operation, the CUDA driver is a natural choice for the wrappers provided by Cudagrind. Anytime a program handles memory on the device or transfers data between the device and/or the host the respective Cudagrind wrapper will be called instead. Depending on the requested operation the internal list of memory regions allocated on the device is being updated (Figure \[fig:CG-List\]) and the function parameters are cross checked against this list as well as the knowledge of Valgrind about the device on the host memory. This allows detection of various errors like - data being copied into or from unallocated memory regions. - undefined data being copied from host onto the device[^2]. - copying more data than there’s allocated memory. - certain concurrent accesses when several threads are used. ![Internal list of memory regions that have been allocated on the device. Each allocation generates a new entry into the list of linear memory or device arrays of the respective CUDA context ctx. Both lists contain information about the device pointer returned by the allocation and either the size of the linear memory or device array descriptor.[]{data-label="fig:CG-List"}](Images/Cudagrind_Internal_List){width=".75\textwidth"} Software Requirements {#sec:req} ===================== To compile the Cudagrind wrappers the following components need to be installed and available - Valgrind 3.6.0 or newer. - CUDA SDK 4.0 or newer. - CUDA driver that works with installed SDK. Additionally the location of the respective driver’s `libcuda.so` and Valgrind’s `valgrind.h` needs to be known and passed to the linker when compiling Cudagrind. To execute a program with the Cudagrind wrappers the same Valgrind version used during the compilation needs to be present. The generated `libcudagrind.so` and the `libcuda.so` from the respective CUDA driver needs to be added to `LD_PRELOAD` in this order. The addition of `libcuda.so` to `LD_PRELOAD` is only needed in cases where the original program has not been fully linked against it. This is the case when using certain OpenACC compilers, where running the program with the Cudagrind wrappers would lead to unknown references to functions called only from within the wrappers but not the main program itself. Now the executable that is to be run with Cudagrind can simply be called with `valgrind <executable>`. Every time one of the wrapped functions in the CUDA driver is executed Valgrind will replace it with the respective Cudagrind wrapper and perform all needed checks at runtime. Optionally a suppression file can be generated and passed to Valgrind at runtime. This suppresses certain spurious errors reported by Valgrind in the CUDA driver[^3]. In certain rare cases the lib path of the Valgrind installation `/path/to/valgrind/lib/valgrind` needs to be added to `LD_LIBRARY_PATH`. Usage Example ============= This section presents a brief introduction to how Cudagrind can assist in detecting and solving memory related errors in a CUDA program. Let ’vecsum’ in Listing \[lst:vecsum\] be a kernel that performs a basic vector addition of two input vectors `a` and `b` and writes the result into the vector `c`. __global__ void vecsum(double *c, double *a, double *b) { start = threadIdx.x+(blockIdx.x*blockSize.x); size = blockSize.x*grindSize.x; for (int i = start ; i < size ; i += blockSize.x) { c[i] = a[i]+b[i]; } } Further assume the following excerpt from a program utilizing this kernel in Listing \[lst:vecsum\], where a wrong amount of memory is allocated for the result pointed at by `c`. In this case the programmer only allocated enough memory to hold `size` values of the type float instead of the twice as big type double expected by the kernel. ... double *c, *a, *b; double *c_host; int size = 1000000; ... cudaMalloc((void**)&c, size*sizeof(float)); cudaMalloc((void**)&a, size*sizeof(double)); cudaMalloc((void**)&b, size*sizeof(double)); ... vecsum<<<gridSize, blockSize>>>(c, a, b); ... cudaMemcpy(c_host, c, size*sizeof(double), cudaMemcpyDeviceToHost); ... The result of executing this code with different types of error checking is now examined in the following sections. No Tools, No API Error Value Checking ------------------------------------- If the code is being run, without external tools or utilization of error values returned by the CUDA runtime, this error might go unnoticed. First the kernel’s result will be undefined due to the write into unallocated memory regions. Then the call to `cudaMemcpy` will fail due to the wrong amount of memory being scheduled for copying. In the best case this will lead to a failure detected in the remainder of the program. But more likely the bug will go unnoticed, leading to the program continuing with undefined values stored in the host memory pointed at by `c_host`. On the bright side, this method of error detection does not ask for recompilation nor debug information being present in the binary! No Tools, API Error Value Checking ---------------------------------- The simplest, yet still often not utilized, method of error checking within any program relying on the CUDA runtime is the error value returned by the runtime. This can be retrieved either directly as return value from each function call or, especially in case of asynchronous operations, by a call to `cudaGetLastError` or `cudaPeekAtLastError`. For asynchronous operation a preprocessor macro like the one in Listing \[lst:API-check\] can be used to simplify the process. To get the same functionality for the asynchronous case a synchronization would have to be added to get the exact error message generated by the operation. #define cudaVerify(x) do { \ cudaError_t __cu_result = x; \ if (__cu_result!=cudaSuccess) { \ fprintf(stderr, \ "%s:%i: error: cuda function call failed:\n" \ " %s;\nmessage: %s\n",__FILE__,__LINE__, \ #x,cudaGetErrorString(__cu_result)); \ exit(1); \ } \ } while(0) Using `cudaVerify` on the call to `cudaMemcpy` in Listing \[lst:snippet\] produces the output shown in Listing \[lst:API-check\_out\]. This method shows the location in the code where the error is happening, even if the binary contains no debug information, and causes nearly no overhead. But the error message ’invalid argument’ reported by the CUDA runtime is very undescriptive and the method is error prone since the `cudaVerify` macro has to be used with every call to any CUDA runtime function. Additional care has to be taken for asynchronous functions and when the code also utilizes the CUDA driver API or other CUDA based programming models. hpc43598 n093302 306$./a.out example01.cu:60: error: cuda function call failed: cudaMemcpy(c_host, c,(vector_size)*sizeof(double), cudaMemcpyDeviceToHost); message: invalid argument Valgrind/Memcheck ----------------- Utilizing only the Valgrind tool Memcheck produces no warnings or error messages, except possibly spurious errors inside the CUDA driver library `libcuda.so`. As mentioned in Section \[sec:req\] these are not connected to actual errors in the program being run and can be either ignored or suppressed with a Valgrind suppression that’s tailored to CUDA driver being used. In this case Valgrind is unable to detect the error since the problem only occurs on the device and inside the kernel being executed. This part of the program is effectively invisible to the Valgrind framework. At the current time it is unable to translate and run the PTX or CUBIN code that is being transfered to and run on the device by the CUDA driver. cuda-memcheck ------------- NVIDIA’s memory checker `cuda-memcheck` can also be used to check for memory related problems at runtime. In this case the program is passed to `cuda-memcheck` as parameter and the tool prints every error located to `stdout`. Another plus point is the tight integration with `cuda-gdb`. Using ’`set cuda memcheck on`’ inside the `gdb` based debugger, which also comes bundled with the CUDA SDK, will enable the same checking abilities when debugging a program. Listing \[lst:cuda-memcheck\_out\] shows the output produced by `cuda-memcheck` when being used on the code from Listing \[lst:snippet\]. In this case `cuda-memcheck` manages to detect location of the error and reports the internal error number reported by the CUDA runtime. But the display of the actual error name or description is missing as is the actual location in the source code of the program. The location is only reported as address in the binary, not in the source code, even if the program has been compiled with `-g -G` for full debug informations. hpc43598 n093302 307$cuda-memcheck ./a.out CUDA-MEMCHECK Program hit error 11 on CUDA API call to cudaMemcpy Saved host backtrace up to driver entry point at error Host Frame:/../libcuda.so [0x26a180] Host Frame:/../libcudart.so.5.0 (cudaMemcpy + 0x28c) [0x3305c] Host Frame:./a.out [0xc55] Host Frame:/lib64/libc.so.6 (__libc_start_main + 0xfd) [0x1ecdd] Host Frame:./a.out [0x9f9] ERROR SUMMARY: 1 error Valgrind/Memcheck with Cudagrind Wrappers ----------------------------------------- Running the aforementioned code snippet in Valgrind with enabled Cudagrind wrappers will produce the output shown in Listing \[lst:Cudagrind\_out\]. The output shown gives a detailed error message. It tells the user that the allocated device memory during a device to host transfer has been too small and that there are only half of the needed amount of bytes available. This already gives a hint at the actual error in Listing \[lst:snippet\], where the result vector `c` has been allocated for floats instead of doubles. Additionally, if the program has been compiled with debug information, the exact location in the source code is being shown as in Listing \[lst:Cudagrind\_out\]. Otherwise, similar to running the program with `cuda-memcheck`, the location inside the binary is being shown. hpc43598 n093302 307$valgrind ./a.out Error: Allocated device memory too small for device->host copy. Expected 8000000 allocated bytes but only found 4000000. at 0x4C18E79: VALGRIND_PRINTF_BACKTRACE (valgrind.h:4477) by 0x4C19261: cuMemcpyDtoH_v2 (cuMemcpyDtoH.c:58) by 0x5A1E531: ??? (in /../libcudart.so.5.0.35) by 0x5A40E43: cudaMemcpy (in /../libcudart.so.5.0.35) by 0x400C54: main (example01.cu:60) Summary ======= Cudagrind offers a novel approach to close the missing link between host side memory checking, e.g. through Valgrind based tools, and device side memory checking with `cuda-memcheck`. Utilizing the provided set of wrappers it is now possible to track the allocation status and definedness of memory during host/device, device/host and device/device memory transactions at runtime. With Cudagrind being dynamically executed by the Valgrind core and depending on only the CUDA driver there is no recompilation needed, except for the precise localization of potential errors in the original source code. Due to the location of the Cudagrind wrappers in the CUDA stack they work with ’everything CUDA’. This includes programs directly accessing the CUDA driver or the CUDA runtime as well as applications relying on CUDA libraries or based on pragma based programming models like OpenACC. The only requirement is for the chosen method to be built on top of the CUDA driver. Additionally with how Valgrind and its wrappers work there is no overhead when running the program on its own, as the original program does not need to be changed or recompiled/-linked in any way. At the time of writing a beta Version of Cudagrind has been released[^4]. Outlook ------- With the work presented in this paper a solid foundation for memory transaction checking for CUDA based programs has been laid out. Future version might provide additional wrappers that, in addition to the actual memory transactions, check the arguments of kernels running on the device. This involves providing a wrapper for `cuLaunchKernel` and unpacking the parameter list passed to this function in order to link these parameters to the actual variables in the program. Another possible approach would be a thorough integration of CUDA within the Valgrind framework. Akin to how Valgrind provides a virtual x86 CPU it would be possible to extend it’s core with a virtual device running PTX or CUBIN code produced by NVIDIA’s CUDA compiler. While allowing full control and complete memory checking capabilities, even for code running on the device, this is no trivial task though. One way to achieve this goal might lie in the extension and integration of a dynamic CUDA compilation framework like ’GPU Ocelot’ [@Ocelot]. This would compile CUDA code into host based code, which in turn could be handled by a modified Valgrind kernel which would have to be aware of the properties specific to code running on a device. Another approach, as proposed in [@Ocelot_Instrumentation], would be to include the instrumentation in Ocelot. But in this case a solution for memory checking on the host side would’ve to be added to Ocelot. [^1]: http://docs.nvidia.com/cuda/cuda-driver-api/index.html [^2]: This might not be an error in cases where the host memory is accessed in a strided way. [^3]: See the Valgrind manual that comes with your installation or at the Valgrind homepage at http://valgrind.org/ for more information about suppression files. [^4]: https://www.hlrs.de/organization/av/spmt/research/cudagrind/

ArXiv

--- abstract: 'In this paper, we investigate how much of the numerical artefacts introduced by finite system size and choice of boundary conditions can be removed by finite size scaling, for strongly-correlated systems with quasi-long-range order. Starting from the exact ground-state wave functions of hardcore bosons and spinless fermions with infinite nearest-neighbor repulsion on finite periodic chains and finite open chains, we compute the two-point, density-density, and pair-pair correlation functions, and fit these to various asymptotic power laws. Comparing the finite-periodic-chain and finite-open-chain correlations with their infinite-chain counterparts, we find reasonable agreement among them for the power-law amplitudes and exponents, but poor agreement for the phase shifts. More importantly, for chain lengths on the order of 100, we find our finite-open-chain calculation overestimates some infinite-chain exponents (as did a recent density-matrix renormalization-group (DMRG) calculation on finite smooth chains), whereas our finite-periodic-chain calculation underestimates these exponents. We attribute this systematic difference to the different choice of boundary conditions. Eventually, both finite-chain exponents approach the infinite-chain limit: by a chain length of 1000 for periodic chains, and $> 2000$ for open chains. There is, however, a misleading apparent finite size scaling convergence at shorter chain lengths, for both our finite-chain exponents, as well as the finite-smooth-chain exponents. Implications of this observation are discussed.' address: 'Division of Physics and Applied Physics, School of Physical and Mathematical Sciences, Nanyang Technological University, 21 Nanyang Link, Singapore 637371, Republic of Singapore' author: - Sisi Tan - Siew Ann Cheong title: 'Understanding finite size effects in quasi-long-range orders for exactly solvable chain models' --- finite size effects ,exact solution ,hardcore bosons ,spinless fermions ,boundary conditions. Introduction ============ Long-range and quasi-long-range orders in strongly-correlated systems are the object of numerous theoretical [@Daniel1991PA; @Tian2002PRB66e224408] and numerical [@Hirsch1989PRB40e4769; @Minoru1989PRB40e2494; @Millis2000PRB61e12496] studies. While it is fairly straightforward to ascertain the presence or absence of true long-range order [@Xavier2007PRB76e014422; @Hohenberg1976PR158; @mermin1966PRL17; @karl1992PRB45e7229; @Hirsch1989PRL62e591], it is much harder to identify the dominant correlations from amongst several competing quasi-long-range orders in the quantum-mechanical ground-state wave function. Much of this difficulty lies with the fact that few models that describe strongly-correlated systems can be solved analytically, and we have to resort to approximate or numerical methods to solve for the ground-state wave function. In particular, numerical solutions can only be obtained for a finite system, when our real interest is in the dominant quasi-long-range order of the infinite thermodynamic system. An obvious solution to this finite size problem would be to harness as much computational power as we can, to simulate extremely large systems [@Liebsch2009PRB80e165126; @Hager2003jcp194; @Yuki2009Jcp130e234114]. Even so, our state-of-the-art simulations are still many orders of magnitude smaller than systems probed experimentally. Alternatively, we can simulate systems of different sizes, and thereafter perform finite size scaling [@fisch2009PRB79e214429; @Lee1991PRB43e1268; @Chayes1986PRL57], to extrapolate the results to infinite system size. In finite size scaling, the computational condensed matter community is essentially guided by general results from statistical mechanics [@Hofstetter1993EL21e993; @Marcelo1988PRB38e9172; @Li1996PRE53e2940; @Toral2007CCP2e177], as well as exact solutions [@Bernu1992PRL69e2590; @Bethe1931ZP71e1312; @CN1967PRL19e1312]. However, finite size scaling is still very much an art, with no rigorous theorem proving that it will always work. Besides the finite number of sites, our choice of boundary conditions might also affect the convergence properties of finite size scaling, and hence the reliability of our numerical solutions. In two recent studies, various quasi-long-range correlation functions were calculated using exact diagonalization [@Cheong2009PRB79e212402] and DMRG [@Munder2009NJP0910]. In particular, in Ref. [@Munder2009NJP0910], the power-law exponents obtained after finite size scaling do not agree with those derived from exact solutions in Ref. [@Cheong2009PRB80e165124]. In this paper, we investigate the relative importance of finite system size and choice of boundary conditions, by comparing the power-law exponents of various correlation functions in a finite periodic and open chain of hardcore bosons or spinless fermions with infinite nearest-neighbor repulsion, against those of the infinite chain,[@Cheong2009PRB80e165124] as well as exponents of finite smooth chain obtained from the DMRG calculation [@Munder2009NJP0910]. In Sec. \[sect:methods\], we will describe how the ground state of finite periodic/open chain of particles with infinite nearest-neighbor repulsion (excluded chain) can be mapped to the ground state of finite periodic/open chain of particles without nearest-neighbor repulsion (included chain). Based on this correspondence between ground states, we then explain how correlation functions in the excluded chain can be written in terms of corresponding correlation functions in the included chain, through an intervening-particle expansion. Thereafter, we present in Sec. \[sect:results\] our results for the two-point, density-density, and pair-pair correlation functions for finite excluded chains of hardcore bosons and spinless fermions. We fit these to different asymptotic combinations of simple and oscillatory power laws, to find generally good agreement between the amplitudes and exponents, and generally poor agreement between the phase shifts for finite and infinite chains. More importantly, we find for finite chains of the same intermediate lengths, reasonable agreement between our open-chain exponents and the finite-smooth-chain exponents in DMRG study. Both show strong systematic differences with the finite-periodic-chain exponents. By repeating these exponent calculations for very long chains, we established that these differences at intermediate chain lengths to be due to the different choice of boundary conditions. These findings are then summarized in Sec. \[sect:conclusions\]. Models and Methods {#sect:methods} ================== This section is organized into two subsections. In Sec. \[sect:chainmodels\], we define our chain models of hardcore bosons and spinless fermions with infinite nearest-neighbor repulsion, and show how the ground states of these finite periodic/open excluded chains can be mapped to the ground states of finite periodic/open included chains of hardcore bosons and spinless fermions without nearest-neighbor repulsion through a right-exclusion map. Then in Sec. \[sect:interpartexp\], we describe the intervening-particle expansion method, which allows us to write the ground-state expectation of a given excluded-chain observable, as a sum over conditional ground-state expectations of included chain observables. Chain models {#sect:chainmodels} ------------ The excluded chain models for hardcore bosons and spinless fermions are given by the Hamiltonians $$\begin{aligned} H^{(e,b)} &= -t \sum_j \left({B_{j}^{\dagger}}{B_{j+1}} + {B_{j+1}^{\dagger}}{B_{j}}\right) + V \sum_j N_j N_{j+1} + {} \\ &\quad\ U \sum_j N_j (N_j - 1), \\ H^{(e,f)} &= -t \sum_j \left({C_{j}^{\dagger}}{C_{j+1}} + {C_{j+1}^{\dagger}}{C_{j}}\right) + V \sum_j N_j N_{j+1}, \end{aligned}$$ respectively, where $t$ is the hopping matrix element, $V$ is the nearest-neighbor repulsion, and $U$ is the on-site repulsion for hardcore bosons. Exact solutions for these models can be obtained in the limit $U \to \infty$ and $V \to \infty$, such that each site can only be singly occupied, and no adjacent sites can be simultaneously occupied. The excluded chains can be mapped to included chains of hardcore bosons and spinless fermions, with Hamiltonians $$\begin{aligned} H^{(i,b)} &= -t \sum_j \left({b_{j}^{\dagger}}{b_{j+1}} + {b_{j+1}^{\dagger}}{b_{j}}\right) + U \sum_j n_j (n_j - 1), \\ H^{(i,f)} &= -t \sum_j \left({c_{j}^{\dagger}}{c_{j+1}} + {c_{j+1}^{\dagger}}{c_{j}}\right). \end{aligned}$$ respectively, using a right-exclusion map first used by Fendley to map a supersymmetric chain of spinless fermions to the $XXZ$ chain [@fendley03]. As shown in Fig. \[fig:rightexclusionmap\], the right-exclusion map deletes an empty site to the right of each occupied site, to map an excluded configuration for a chain of length $L$ to an included configuration for a chain of length $L - P$. For finite open chains, this mapping is one-to-one, as shown in Ref. [@Cheong2009PRB80e165124]. For finite periodic chains, the right-exclusion map is not one-to-one, as illustrated in Fig. 1. Nevertheless, it is possible to construct a one-to-one correspondence between the Bloch states of the excluded chain and Bloch states of the included chain. Details for this construction can be found in Ref. [@Cheong2006thesis]. ![Schematic diagram illustrating the many-to-one mapping from $P$-particle configurations of an excluded periodic chain of length $L$ to $P$-particle configurations of an included periodic chain of length $L' = L - P$, by deleting a single empty site on the right of a particle.[]{data-label="fig:rightexclusionmap"}](rightexclusionPBC) For the rest of this paper, we will use $L$ to denote the length of the periodic/open excluded chain, whose sites are indexed by $j = 1, \dots, L$. We will also use $P$ to denote the total number of particles on the chain, and ${A_{j}} = {B_{j}}, {C_{j}}$ as a common notation for hardcore boson and spinless fermion operators. Intervening-particle expansion {#sect:interpartexp} ------------------------------ Because of the one-to-one correspondence between Bloch states, the Bloch-state amplitudes in the excluded chain ground state are identical to those in the included chain ground state. Hardcore bosons or spinless fermions, the included-chain ground state can be ultimately written in terms of a one-dimensional Fermi sea with discrete wave numbers. Hence the ground-state expectations of all observables on the excluded chain can ultimately be expressed in terms of such Fermi-sea expectations. In this subsection, we briefly describe how the correlations $\braket{O_1 O_2}$ of two separated local operators of the excluded chain can be calculated using the method of *intervening-particle expansion*. First, for every observable $O$ on the excluded chain, we note that it is possible to define a *corresponding observable* $O'$ on the included chain, such that $$\label{eqn:OOp} \frac{1}{{\overline{N}}} \braket{ J^{\alpha} | O | J^{\beta} } = \frac{1}{{\overline{n}}} \braket{ j^{\alpha} | O' | j^{\beta} },$$ for all excluded configurations $\ket{J^{\alpha}}$ and $\ket{J^{\beta}}$ that are mapped to the included configurations $\ket{j^{\alpha}}$ and $\ket{j^{\beta}}$ by the right-exclusion map. Eq.  then tells us that $$\frac{1}{{\overline{N}}} \braket{O} = \frac{1}{{\overline{n}}} \braket{O'}$$ between the ground-state expectations of $O$ and $O'$. Next, we expand the ground-state expectation $$\braket{O_j O_{j+r}} = \sum_{\{p\}} \braket{O_j O_p O_{j+r}}$$ as a sum over conditional ground-state expectations, where $O_p$ is a product of $p$ particle-occupation number operators $N_j$ and $r - p$ hole-occupation number operators $(1 - N_j)$. We call this sum over all possible ways to insert particles between $O_j$ and $O_{j+r}$ the *intervening-particle expansion*. Finally, we rewrite each conditional ground-state expectation $\braket{O_j O_p O_{j+r}}$ on the excluded chain in terms of their corresponding conditional ground-state expectation $\braket{O'_j O'_p O'_{j+r-p}}$ on the included chain. In terms of these conditional expectations, which can in turn be written in terms of expectations of the one-dimensional Fermi sea, the intervening-particle expansion becomes $$\braket{O_j O_{j+r}} = \frac{{\overline{N}}}{{\overline{n}}} \sum_{\{p\}} \braket{O'_j O'_p O'_{j+r-p}},$$ which can always be evaluated numerically for separations $r$ that are not too large. Results and discussions {#sect:results} ======================= In Ref. [@Cheong2009PRB80e165124], three correlation functions: (i) the two-point function $\braket{{A_{j}^{\dagger}}{A_{j+r}}}$; (ii) the density-density correlation function $\braket{N_j N_{j+r}}$; and (iii) the pair-pair correlation function $\braket{{A_{j}^{\dagger}}{A_{j+2}^{\dagger}}{A_{j+r}}{A_{j+r+2}}}$ were systematically examined for infinite chains of hardcore bosons and spinless fermions with infinite nearest-neighbor repulsion, and also in three limiting cases for an infinite ladder. In this section, we calculate these three correlation functions for the finite periodic/open chain, for comparison against those of the infinite chain [@Cheong2009PRB80e165124], as well as against those of finite smooth chain computed in the recent DMRG study [@Munder2009NJP0910]. For the finite open chain, we analyze the numerical correlation functions only for $\bar{N} \leq 0.25$. We believe the results are not reliable for $\bar{N} > 0.25$, because of phase separation in the finite open chain ground states (the symptom of which can be seen in Fig. \[fig:cdwpofp\]). To understand the relative importance of finite chain length and choice of boundary conditions, we fit all the slowly-decaying correlation functions to the asymptotic form $$A_0 + A_1\, r^{-A_2} + A_3\, r^{-A_4}\, \cos(k r + \phi),$$ where $A_0$ is a constant term, $A_1$ and $A_2$ are the amplitude and exponent of a simple power law, while $A_3$ and $A_4$ are the amplitude and exponent of an oscillatory power law with wave number $k$ and phase shift $\phi$. We then check how these finite-chain exponents depend on chain length and boundary conditions, and thus the effectiveness of finite size scaling. Two-point functions ------------------- The hardcore boson two-point function $\braket{{B_{j}^{\dagger}}{B_{j+r}}}$ obtained using the intervening-particle expansion appears to be an oscillatory power law sitting on top of a simple power law. As shown in Fig. \[fig:flbpo\], we fit the hardcore boson two-point function to the asymptotic form $\braket{{B_{j}^{\dagger}}{B_{j+r}}} = A_1\, r^{-A_2} + A_3\, r^{-A_4}\, \cos(k r + \phi)$. As expected, the fitted wave number is $k = \pi{\overline{N}}$. For the infinite chain, the leading exponent (that of the simple power law) was found to be very close to $A_2 = \frac{1}{2}$, as predicted by Efetov and Larkin for an included chain of hardcore bosons [@Efetov1976SP42e390]. For our finite periodic chain, this leading exponent was found to approach $\frac{1}{2}$ with increasing density from below, as shown in Fig. \[fig:flbpofp\]. In comparison, the finite-smooth-chain exponent $A_2 = 2{\overline{N}}+ \frac{1}{2}$ obtained in Ref. [@Munder2009NJP0910] approaches $\frac{1}{2}$ with decreasing density from above. The finite-open-chain exponent $A_2$ agrees very closely with the infinite-chain results. Otherwise, we find reasonable agreement between the amplitudes and exponents of the finite and infinite chains, but poor agreement between their phase shifts. ![The fitted (top) amplitudes $A_1$ (black) and $A_3$ (red), (middle) exponents $A_2$ (black) and $A_4$ (red), and (bottom) phase shift $\phi$ of the simple power law and oscillatory power law as a function of density ${\overline{N}}$, for the two-point function of a $L = 100$ periodic(dashed lines) and open(dot-dashed lines) chain of hardcore bosons with infinite nearest-neighbor repulsion. The corresponding parameters of the infinite chain from Ref. [@Cheong2009PRB80e165124] are shown as solid lines, whereas finite-smooth-chain results from Ref. [@Munder2009NJP0910] are shown as dotted lines.[]{data-label="fig:flbpofp"}](flbfp) Similarly, for the spinless fermion two-point function, which is fitted to a single oscillatory power law $\braket{{C_{j}^{\dagger}}{C_{j+r}}} = A_3\, r^{-A_4}\, \cos(k r + \phi)$ (see Fig. \[fig:flfpo\]), we find reasonable agreement between the amplitudes and exponents of the finite and infinite chains, but poor agreement between the phase shifts of the finite and infinite chains (see Fig. \[fig:flfpofp\]). As expected, the fitted wave number is $k = \pi{\overline{N}}$. For noninteracting spinless fermions, the exponent of the oscillatory power law is $A_4 = 1$. We see from Fig. \[fig:flfpofp\] that for both finite and infinite chains, $A_4$ starts close to 1 at ${\overline{N}}= 0$, and decreases with increasing density. In particular, the finite-chain exponent is smaller than the infinite-chain exponent, telling us that two-point correlations are stronger on finite chains. ![The fitted amplitude $A_3$ (top), exponent $A_4$ (middle) and phase shift $\phi$ (bottom) of the oscillatory power-law fit to the two-point function of a $L = 100$ periodic(dashed lines) and open(dot-dashed lines) chain of spinless fermions with infinite nearest-neighbor repulsion, as a function of density ${\overline{N}}$. The corresponding parameters of the infinite chain from Ref. [@Cheong2009PRB80e165124] are shown as solid lines.[]{data-label="fig:flfpofp"}](flffp) Density-density correlations ---------------------------- The density-density correlation function, which is identical for hardcore bosons and spinless fermions on a chain, can be fitted very well to the asymptotic form $\braket{N_j N_j+r} = A_0 + A_3\, r^{-A_4}\, \cos(k r + \phi)$ (see Fig. \[fig:cdwpo\]). Again, the fitted wave number $k = 2\pi{\overline{N}}$ agrees with our expectations. Also, as with the two-point functions, we see in Fig. \[fig:cdwpofp\] that there is reasonable agreement between the finite and infinite chains for the fitted amplitudes and exponents, but poor agreement for the phase shifts. For both finite chains, we find $A_0 > \braket{N_j}\braket{N_j+r} = {\overline{N}}^2$, the limit attained in the infinite excluded chain. This excess density-density correlation compared to the infinite chain persists at $L = 200$ and $L = 1000$. Also shown in Fig. \[fig:cdwpofp\] are the oscillatory power-law exponent $A_4$ for the finite periodic/open chain, the infinite chain, and the finite smooth chain. In Fig. \[fig:cdwpofp\](a), while we see the exponents in all three cases decaying with increasing ${\overline{N}}$, we again find the finite-periodic-chain exponent (dashed line) to be generally smaller than the infinite-chain exponent (solid line), the finite-open-chain exponent to be slightly larger than the infinite-chain exponent, and the finite-smooth-chain exponent (dotted line) to be larger than the infinite-chain exponent. Since the same chain lengths were used, we suggest two possible explanations for these systematic differences. As we shall later show for these intermediate chain lengths, the weaker density-density correlation of finite smooth chain and finite open chain is not an artefact resulting from spatial averaging done to ‘restore’ translational symmetry. Instead, it is a genuine finite size effect related to the choice of smooth [@Munder2009NJP0910] and open boundary conditions, as opposed to periodic boundary conditions used on our finite periodic chains. ![The fitted constant $A_0$ (top), and amplitude $A_3$ (second from top), exponent $A_4$ (second from bottom), and phase shift $\phi$ of the oscillatory power law as a function of density ${\overline{N}}$, for the density-density correlation function of a $L = 100$ periodic(dashed lines) and open(dot-dashed lines) chain of particles with infinite nearest-neighbor repulsion. The corresponding parameters of the infinite chain from Ref. [@Cheong2009PRB80e165124] are shown as solid lines ($A_4 = \frac{1}{2} + \frac{5}{2}(\frac{1}{2} - {\overline{N}})$ for the infinite chain), while the finite-smooth-chain exponent $A_4 = -2.96\, {\overline{N}}+ 2.23$ is shown as the dotted line.[]{data-label="fig:cdwpofp"}](cdwfp) Pair-pair correlations ---------------------- The pair-pair correlation function $\braket{{A_{j-2}^{\dagger}}{A_{j}^{\dagger}}{A_{j+r}}{A_{j+r+2}}}$ is also identical for both spinless fermions and hardcore bosons. For spinless fermions, $\braket{{A_{j-2}^{\dagger}}{A_{j}^{\dagger}}{A_{j+r}}{A_{j+r+2}}}$ can be interpreted as the superconducting correlations. As shown in Fig. \[fig:scpo\], the pair-pair correlation function can be fitted very well to the asymptotic form $A_1 r^{-A_2} + A_3 r^{-A_4} \cos (kr + \phi)$. As expected, the wave number is $k = 2\pi\bar{N}$ if we allow $k$ to be a fitting parameter. Again, from Fig. \[fig:scpofp\], we see that there is reasonable agreement on the amplitudes and exponents between the finite and infinite chains, but poor agreement between their phase shifts. ![The fitted amplitudes $A_1$ and $A_3$ (top), exponents $A_2$ and $A_4$ (middle), and phase shifts $\phi$ (bottom) of the pair-pair correlation function of a $L = 100$ periodic(dashed lines) and open(dot-dashed lines) chain of particles with infinite nearest-neighbor repulsion, as a function of the density ${\overline{N}}$. The corresponding parameters of the infinite chain from Ref. [@Cheong2009PRB80e165124] are shown as solid lines.[]{data-label="fig:scpofp"}](scffp) Comparison between different chain lengths ------------------------------------------ In the DMRG study of finite smooth chain [@Munder2009NJP0910], a systematic study on the effects of chain lengths was undertaken, even though there was no rigorous attempts at performing finite size scaling. They found, for the oscillatory power-law exponent $A_4$ of the density-density correlation on ladders of lengths $L = 100, 150, 200$, the fitted parameter appears to have converged onto the straight line $A_4 = -2.96\,{\overline{N}}+ 2.23$. In Fig. \[fig:cdwfss\], we show the fitted exponent of the density-density correlations on finite periodic chains with lengths $L = 50, 100, 200, 1000$ and finite open chains with lengths $L=100, 150, 200, 300, 1000, 2000$. For the finite periodic chain, we also find an apparent convergence at a chain length of approximately $L = 200$. However, this ‘limiting’ behaviour of $A_4$ does not agree with the exponent determined from the infinite chain, and is also systematically different from the ‘limiting’ finite-smooth-chain exponent. In fact, unlike the infinite-chain and finite-smooth-chain exponents, the finite-periodic-chain exponent appears to be a strongly nonlinear function of $\bar{N}$. When we go to a $L = 1000$ periodic chain, the exponent $A_4$ is still strongly $\bar{N}$-dependent. However, we can see from Fig. \[fig:cdwfss\] that it is closer to the infinite-chain limit. For finite open chains, however, the chain-length-dependence of the exponent $A_4$ is rather different. Starting from being close to the infinite-chain limit at $L = 100$, we see the exponent moving *away* from the infinite-chain limit as the chain length goes to $L = 150$ to $L = 200$ to $L = 300$ to $L = 1000$. In fact, there is an apparent convergence at $L = 200$ to the finite-smooth-chain exponent. This suggests that a $L = 100$ chain with smoothed open boundary conditions behaves like a longer chain with hard open boundary conditions, which is what they desired when White introduced the smoothed boundary conditions [@Vekic1993PRL71e4283; @Vekic1996PRB53e14552]. What they, and DMRG group would not have guessed, is the finite open chain exponent deviating from the infinite chain exponent initially as we go to longer and longer chains. From the open chain results of $L = 1000$ and $L = 2000$, we find that the convergence to infinite-chain results is slower than for finite periodic chains. With this, we demonstrated that the choice of periodic/open boundary conditions resulted in the finite-chain exponents being systematically lower/higher than the infinite-chain exponent, for chain lengths on the order of $L = 100$ to $L = 200$. This observation is important, because DMRG calculations on finite smooth chains are currently limited to such chain lengths. Furthermore, we showed that the finite-chain exponents do converge onto the infinite-chain limit, but only when the chain length is on the order of $L = 1000$. Moreover, this convergence is not uniform for finite chains. This poses an important challenge for the use of finite size scaling to determine the infinite-chain limits. ![The fitted oscillatory power-law exponent $A_4$ for the density-density correlation function of hardcore particles with infinite nearest-neighbor repulsion as a function of density ${\overline{N}}$, for different periodic chain lengths, $L = 50$ (red), $L = 100$ (blue), and $L = 200$ (black). The dashed line is the infinite-chain exponent from Ref. [@Cheong2009PRB80e165124], whereas the dotted line is the best-fit straight line to the finite-smooth-chain exponent from Ref. [@Munder2009NJP0910].[]{data-label="fig:cdwfss"}](cdwfss) Finally, we compare the density-density correlation functions of different chain lengths, against the infinite-chain density-density correlation function, at density $\bar{N} = 0.10$. As shown in Fig. \[fig:cdwfcfss\], oscillatory power-law decay in the density-density correlation function is seen for finite and infinite chains. The finite periodic chain and infinite chain amplitudes are very similiar. The rates at which the density-density correlations decay on the finite periodic chain and infinite chain are also very similar. This explains the reasonable agreement seen in the fitted amplitudes and exponents. However, the phase shifts in the finite periodic chain and the infinite chain are rather different, and increasing the chain length from $L = 50$ to $L = 100$ to $L = 200$ does not make the finite-chain phase shift approach the infinite-chain phase shift. This is true even at $L = 1000$. As pointed out by the referee, this is expected as the phase shift is a parameter that is adjusted globally to minimize the effects of the boundaries, and thus will not converge smoothly with increasing chain length. More importantly, the amplitude of the oscillatory power law is very small for finite open chains, because of the spatial averaging performed. This same spatial averaging is done in the finite-smooth-chain calculations. ![The density-density correlation functions of infinite, finite periodic, finite open chains with ${\overline{N}}= 0.10$. The solid line is the infinite-chain exponent from Ref. [@Cheong2009PRB80e165124].[]{data-label="fig:cdwfcfss"}](cdwfcfss) Conclusions {#sect:conclusions} =========== In conclusion, we employed the exact mapping between the ground states of the excluded chains and included chains of hardcore bosons and spinless fermions, along with the intervening-particle expansion method, both developed in Ref. [@Cheong2009PRB80e165124], to investigate how the finite chain length and the choice of the boundary conditions manifest themselves in the two-point, density-density, and pairing correlation functions. We fitted these finite-chain correlations to the generic asymptotic form $A_0 + A_1\, r^{-A_2} + A_3\, r^{-A_4}\, \cos(kr + \phi)$, and compared the fitted parameters to those obtained from an infinite chain, and a finite smooth chain. In general, we find reasonable agreement between fitted amplitudes and exponents for the finite and infinite chains, but poor agreement for the fitted phases. We also find the phase shifts depend on both boundary conditions and system length. Comparing the finite-periodic-chain, finite-open-chain, finite-smooth-chain and infinite-chain exponents for correlation functions at the same length, we find that the finite-periodic-chain exponents are closer to the infinite-chain results and finite-open-chain exponents are closer to the finite-smooth-chain ones. We attribute this systematic difference to the finite size effects and the different choices of boundary conditions. In a nutshell, finite-periodic-chain correlations are stronger than they are with open or smooth boundary conditions (or any variants thereof). More importantly, we observed at intermediate chain lengths an apparent ‘convergence’ of our finite-chain exponents to functions of the density ${\overline{N}}$ different from the infinite-chain exponents. A straightforward finite size scaling may thus lead us to the wrong physics. This same phenomenon was observed in the DMRG study of finite smooth chain as well, suggesting in general that uniform convergence cannot be expected for chain lengths of $L \sim 10^2$, and perhaps not even for chain lengths of $L \sim 10^3$. A more positive take on these general results would that that numerical studies using more than one set of boundary conditions offer us a better sense of how far we might be away from the infinite-system limits, even when there is no guarantee that the results from different boundary conditions will bound the true infinite-chain limit. Acknowledgments {#acknowledgments .unnumbered} =============== This research is supported by startup grant SUG 19/07 from the Nanyang Technological University. We thank Pinaki Sengupta for discussions, and the anonymous referee for suggestions on how to improve the paper. [99]{} D. Boyanovsky and D. Jasnow, Physica A **177**, 537 (1991). G.-S. Tian and H.-Q. Lin, Phys. Rev. B **66**, 224408 (2002). J. E. Hirsch and S. Tang, Phys. Rev. B **40**, 4769 (1989). M. Takahashi, Phys. Rev. B **40**, 2494 (1989). A. J. Millis, Phys. Rev. B **61**, 12496 (2000). N. D. Mermin and H. Wagner, Phys. Rev. Lett. **17**, 1133 (1966). P. C. Hohenberg, Phys. Rev. **158**, 383 (1967). J. E. Hirsch and S. Tang, Phys. Rev. Lett. **62**, 591 (1989). K. J. Runge, Phys. Rev. B **45**, 7229 (1992). J. C. Xavier and A. L. Malvezzi, Phys. Rev. B **76**, 014422 (2007). G. Hager, E. Jeckelmann, H. Fehske, and G. Wellein, J. Comp. Phys. **194**, 795 (2004). A. Liebsch and N.-H. Tong, Phys. Rev. B **80**, 165126 (2009). Y. Kurashige and T. Yanai, J. Chem. Phys. **130**, 234114 (2009). J. T. Chayes, L. Chayes, D. S. Fisher, and T. Spencer, Phys. Rev. Lett. **57**, 2999 (1986). J. Lee and J. M. Kosterlitz, Phys. Rev. B **43**, 1268 (1991). R. Fisch, Phys. Rev. B **79**, 214429 (2009). Marcelo D. Grynberg, Phys. Rev. B **38**, 9172 (1988). E. Hofstetter and M. Schreiber, Europhys. Lett. **21**, 993 (1993). Z. Li and L. Schülke, Phys. Rev. E **53**, 2940 (1996). R. Toral and Claudio J. Tessone, Commun. Comput. Phys. **2**, 177-195 (2007). H. A. Bethe, Z. Phys. **71**, 205 (1931). C. N. Yang, Phys. Rev. B **19**, 1312 (1967). B. Bernu, C. Lhuillier, and L. Pierre, Phys. Rev. Lett. **69**, 2590 (1992). S.-A. Cheong and C. L. Henley, Phys. Rev. B **79**, 212402 (2009). W. Münder, A. Weichselbaum, A. Holzner, J. von Delft, and C. L. Henley, New J. Phys. (to appear). S.-A. Cheong and C. L. Henley, Phys. Rev. B **80**, 165124 (2009). M. Vekić and S. R. White, Phys. Rev. Lett. **71**, 2483-2486 (1993) M. Vekić and S. R. White, Phys. Rev. B. **53**, 14552-14557 (1996) K. B. Efetov and A. I. Larkin, Sov. Phys.-JETP **42**, 390 (1976). P. Fendley, B. Nienhuis, and K. Schoutens, J. Phys. A: Math. Gen. **36**, 12399 (2003). S.-A. Cheong, *Many-Body Fermion Density Matrices*, Ph.D. thesis, Cornell University, 2006. Available at URL: <http://people.ccmr.cornell.edu/~clh/Theses/cheong-habis.pdf>.

ArXiv

--- abstract: 'Under the action of an alternating perpendicular magnetic field the polarity of the vortex state nanodisk can be efficiently switched. We predict the regular and *chaotic* dynamics of the vortex polarity and propose a simple analytical description in terms of a *reduced vortex core* model. Conditions for the controllable polarity switching are analyzed.' author: - 'Oleksandr V. Pylypovskyi' - 'Denis D. Sheka' - 'Volodymyr P. Kravchuk' - 'Franz G. Mertens' - Yuri Gaididei date: 'June 17, 2013' title: Regular and chaotic vortex core reversal by a resonant perpendicular magnetic field --- Introduction ============ Investigation of the magnetization dynamics at the nanoscale is a key task of the modern nanomagnetism.[@Braun12] One of the typical topologically nontrivial magnetic configurations of a nanoscaled magnet is a magnetic vortex, which can form a ground state configuration of submicron–sized magnetic disk–shaped particles (nanodots). Such a vortex is characterized by a curling divergent–free in–plane configuration with magnetization tangential to the edge surface of the nanoparticle. [@Hubert98] The out–of–plane magnetization appears only in a very thin region around the vortex core with about the size of an exchange length (typically about 10 nm for magnetically soft materials [@Wachowiak02]). The vortex state is degenerated with respect to the upward or downward magnetization of the vortex core (the vortex polarity $p=\pm1$), hence the vortex polarity can be considered as a bit of information in nonvolatile magnetic vortex random–access memories (VRAM). [@Kim08; @Yu11a] That is why one needs to control the vortex polarity switching process in a very fast way. The vortex polarity switching phenomena were predicted originally for the Heisenberg 2D magnets. [@Gaididei99; @Gaididei00] The interest to this problem was renewed after an experimental detection of the vortex core reversal in nanodots by an excitation with short bursts of an alternating field, [@Waeyenberge06] which opened a possibility to use the vortex state dots as the VRAM. Moreover, this motivated numerous fundamental studies of the vortex core switching mechanism itself.[@Braun12] There are two basic scenarios of the vortex polarity switching. In the first, axially–symmetric (or punch–through) scenario, the vortex polarity is switched due to the direct pumping of axially–symmetric magnon modes. Such a switching occurs, e.g., under the influence of a DC transversal field. [@Okuno02; @Thiaville03; @Kravchuk07a; @Vila09] In the second, axially–asymmetric scenario, the switching occurs due to a nonlinear resonance in the system of certain magnon modes with nonlinear coupling. [@Kravchuk09; @Gaididei10b] Such a scenario is accompanied by the temporary creation and annihilation of vortex–antivortex pairs. [@Waeyenberge06] The axially–asymmetric switching occurs, e.g., under the action of different in–plane AC magnetic fields or by a spin polarized current, see Ref.  and references therein. Recently the interest to the axially–symmetric switching was renewed: using the micromagnetic simulations @Wang12 [@Yoo12] demonstrated that the vortex polarity reversal can be realized under the action of an alternating perpendicular magnetic field. In this case the resonant pumping of the radial magnon modes initiates the switching at much lower field intensities than by the DC fields. We have very recently predicted the possibility of the chaotic dynamics of the vortex polarity under the action of the homogeneous transversal AC magnetic field in the 10 GHz range:[@Pylypovskyi13b] $$\label{eq:sim:Field} {\bm{B}}(t) = {\bm{e}}_z B_0 \sin \left(2\pi f t \right).$$ In order to describe the switching behavior we proposed in Ref.  the analytical two–parameter cutoff model, which gave us a possibility to describe both deterministic and chaotic behavior of the vortex polarity. The goal of the current work is to study in detail the vortex dynamics under the action of a perpendicular AC magnetic field: we found a rich vortex polarity dynamical behavior, including the *regular* and *chaotic* regimes of magnetization reversal. In order to analyze the complicated temporal evolution of the vortex polarity we used here the discrete *reduced core model*, [@Gaididei99; @Gaididei00; @Zagorodny03] which allows us to describe different regimes of vortex polarity dynamics, including the resonant behavior, the weakly nonlinear regimes, the reversal dynamics, and the chaotic regime. The reduced core model is another way to treat the discretness effects. As opposed to the cutoff model, the core model is simpler, hence it allows to go further in analytics. The paper is organized as follows: The full–scale micromagnetic simulations are detailed in Sec. \[sec:sim\]. Our diagram of switching events demonstrates regimes of the regular reversal (single, multiple and periodic ones), intermittent and chaotic regimes. In Sec. \[sec:core\] we describe the comprehensive vortex core dynamics using a simple collective coordinate model, which provides all features of the full–scale simulations. We propose a way of a unidirectional switching controlled switching in Sec. \[sec:control\]. In Sec. \[sec:conclusion\] we state our main conclusions. In Appendix \[app:core\] we derive the reduced core mode. We use the method of multiple scales to perform a weakly nonlinear analysis of the analytical model in Appendix \[app:wna\]. Micromagnetic Simulations of Regular and Chaotic Dynamics {#sec:sim} ========================================================= at (0,2.5) [{width="\textwidth"}]{}; [(pointer) at (8.7,0.92); (leftCorner) at (0,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer1) at (6.35,0.75); (pointer2) at (7.5,0.75); (leftCorner) at (3.025,4.475); (leftCorner) – ($(leftCorner) + (2.925/3-0.1,0)$) – (pointer1) – ($(leftCorner) + (2.925/3+0.1,0)$) – ($(leftCorner) + (2.925/3*2-0.1,0)$) – (pointer2) –($(leftCorner) + (2.925/3*2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925, 2.925 )$) –($(leftCorner) + (0, 2.925 )$) – cycle; at ($(leftCorner) + (2.925/2-0.01, 2.925 /2)$)[]{};]{}; [(pointer) at (8.7,1.2); (leftCorner) at (6.05,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer) at (9.96,1.2); (leftCorner) at (9.075,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; (1.7,3) node\[left\] [$f/3$]{} – (2.,3.4); (2.4,2.1) node\[left\] [$f/2$]{} – (2.8,2.3); Nowadays the micromagnetic simulations are the inherent tools for the nanomagnetic research.[@Kim10] Namely using the numerical simulations it was shown in Refs.  that the resonant perpendicular field forces the vortex core to reverse. Here we perform full–scale micromagnetic simulations to study the complicated vortex core dynamics in details. We consider a disk–shaped nanoparticle (198 nm in diameter and 21 nm in thickness) under the action of the vertical oscillating field using an <span style="font-variant:small-caps;">OOMMF</span>[@OOMMF] micromagnetic simulator with integration method RK5(4)7FC. The material parameters correspond to Permalloy ($\mathrm{Ni}_{81}\mathrm{Fe}_{19}$) with exchange constant $A = 13$pJ, saturation magnetization $M_s = 860$kA/m, zero anisotropy coefficient and the Gilbert damping coefficient $\alpha = 0.01$. The mesh cell was chosen to be $3\times 3\times 21$nm (the three–dimensional mesh will be discussed at the end of this section). For all <span style="font-variant:small-caps;">OOMMF</span> simulations we use as initial state the relaxed vortex with the polarity directed upward and counter–clockwise in–plane magnetization direction. We also simulated the dynamics of the vortices for the samples with other geometrical parameters: as we expected the qualitative behavior of the system remains the same. [^1] First of all we examined the eigenfrequencies of the lower axially–symmetric spin waves by applying a rectangular pulse with the strength of 30mT during 100ps perpendicular to the nanodisk in the vortex state in the same way as in Ref. . Under the action of such a pulse, the magnetization starts to oscillate: a set of symmetrical magnon modes $f_{m=0}^{n}$ is excited. Using the fast Fourier transformation (FFT) of the $z$–component of the total magnetization, typically during $t\in[100\text{\,ps}; 20\text{\,ns}]$, we identified the eigenfrequency of the lowest symmetrical mode $f_{m=0}^{n=1} = 13.98$ GHz. This value defines the lowest threshold for the polarity switching [@Yoo12]. The next nearest peaks in the FFT spectrum correspond to 16.75 GHz and 27.93 GHz. It is already known[@Yoo12] that the vortex polarity switching under the action of the AC field occurs in a wide range of field parameters (the field intensities $B_0$ and field frequencies $f$). The minimal field intensity is reached at about the resonance frequency $f_0^1$. In the current study we are interested in the long–time vortex dynamics, which is accompanied by the axially–symmetric polarity reversal mechanism. In all numerical experiments we calculated the polarity and the position of the vortex as functions of time: The vortex position ${\bm{R}}(t)$ is determined as cross–section of isosurfaces $M_x({\bm{R}})=0$ and $M_y({\bm{R}})=0$,[@Hertel07] and the vortex polarity $p(t)$ is determined as the average $z$–magnetization of four neighbor cells to ${\bm{R}}(t)$. To study in details the temporal evolution of the vortex polarity, we simulated the long–time system dynamics with the time step of 1ps for a wide range of the field parameters (the field intensity $B_0$ varies from $10$ to $500$mT, and the field frequency $f$ changes from $3$ to $21$ GHz).[^2] The results can be summarized in the diagram of dynamical regimes, see Fig. \[fig:sim:homega\]. Depending on the field parameters ($B_0$, $f$), one can separate several different dynamical regimes: (i) the absence of the vortex polarity switching, (ii) the chaotic polarity oscillations, (iii) the regular switchings with frequencies depending on the field frequency, (iv) the intermittent switchings, and (v) the complex vortex–magnon dynamics, where the vortex escapes from the origin. \(i) We start from a weak field: the field intensity is not strong enough to switch the vortex polarity; this regime corresponds to the linear or weakly nonlinear oscillations of the vortex polarity (marked as open boxes in Fig. \[fig:sim:homega\]). The weak pumping of the system (field intensities $B_0 \lesssim 5$mT, see Fig. \[fig:sim:nonlinAmp\]) causes the resonance at the frequency $f_0^1$. The increase of the field intensity leads to the nonlinear dynamical regime. However, if the field intensity is not strong enough, one has a weakly nonlinear regime, which corresponds to the nonlinear resonance. Apart from the nonlinear resonance behavior, the strong pumping causes the vortex polarity instability,[^3] it also causes the shift of the main peak in the FFT spectrum (see Fig. \[fig:sim:nonlinAmp\]b) and the beats in the polarity oscillations (see Fig. \[fig:sim:nonlinAmp\]a). ![(Color online) Nonlinear resonance curves from micromagnetic simulations. Insets: (a) The temporal evolution of the polarity, (b) the FFT spectrum of the vortex polarity for $B_0=9$mT, $f=13.2$GHz during 15ns. Arrow indicates the pumping frequency. []{data-label="fig:sim:nonlinAmp"}](nonlinAmpOOMMF){width="\columnwidth"} Let us consider the case when the magnetization reversal occurs. The switching diagram (see Fig. \[fig:sim:homega\]) has two well–defined minima. The first one corresponds to the resonant excitation of the radially symmetrical mode $f_0^1$. The second minimum near 18GHz, probably, corresponds to the dynamics near the higher resonances[@Yoo12]. \(ii) The open circles [$\textcolor{yellow}{\bullet}$-0.81em$\textcolor{brown}{\circ}$]{} on the switching diagram (see Fig. \[fig:sim:homega\]) correspond to the chaotic polarity reversal process. The typical temporal evolution is shown in Fig. \[fig:sim:homega\]c. To draw a conclusion about chaotic behavior of the vortex polarity, or more accurately, to make quantitative measures of chaotic dynamics, we use two standard ways: the autocorrelation function for the temporal evolution of the vortex polarity and the Fourier distribution of its frequency spectra.[@Moon04] First, we define the autocorrelation function of the vortex polarity signal $$\label{eq:autocorrelation} C(t_i) = \frac{1}{N}\sum_{j=1}^N p(t_{i+j}) p(t_j),\quad i=\overline{1,N}$$ for the discretized time $t_j = j t_0$ with the step $t_0=1$ps, with the boundary values assumed as zero. It is well known[@Moon04] from the correlational analysis, when a signal is chaotic, information about its past origins is lost, i.e. the signal is only correlated with its recent part: the autocorrelation function decays very rapidly, $C(t)\to0$ as $t\to\infty$.[@Moon04] For a periodical signal, the autocorrelation function is a periodic too. A typical example is presented in Fig. \[fig:sim:autocor\]: the autocorrelation function $C(t)$ is aperiodic and sharply decays for the applied magnetic field 70mT with the frequency 14GHz, which corresponds to the chaotic dynamics. The autocorrelation for the regular dynamics demonstrates the oscillations under the action of $B_0=110$mT with $f=18$GHz. The second way is to calculate the Fourier spectrum of a chaotic signal. A typical FFT signal is presented in the Fig. \[fig:sim:fourier\]. It is distinctive for the chaotic regime that the continuous spectrum dominates the discrete spikes (one can identify in the Fig. \[fig:sim:fourier\] only one discrete spike at the pumping frequency). The fitting of such a signal demonstrates typical pink noise behavior with a power law decay of the spectrum, $\mathcal{F}(f) \propto 1/f^\beta$ with $\beta\approx 0.77$.[^4] \(iii) The regular oscillations of the vortex polarity appear in the high frequency regime, see the black diamonds on the switching diagram (see Figs. \[fig:sim:homega\], \[fig:sim:homega\]d). We have detected the periodical motion of the vortex polarity using the pumping frequency 18GHz with the field intensities higher than 100mT. The main peak in the FFT spectrum corresponds to 6GHz, *i. e.* it occurs at $f/3$ of the pumping. Other spikes with decaying intensities appear with steps of 6GHz. We compare the autocorrelation functions for the regular and chaotic oscillations, see Fig. \[fig:sim:autocor\]. In contrast to the chaotic regime, $C(t)$ for periodic oscillations exhibits a high periodicity with a slowly decaying amplitude due to the finite observation time. ![(Color online) Autocorrelation function for $B_0=110$mT and $f=18$GHz (blue dashed line) and $B_0=70$mT and $f=14$GHz (red solid line). The first process demonstrates a periodical process with periodical $C(t)$ and the second one demonstrates a chaotic behaviour with rapidly decaying $C(t)$. []{data-label="fig:sim:autocor"}](autocorrelation01){width="\columnwidth"} In order to compare the temporal dynamics of the polarity in chaotic and regular regimes, we calculate the pseudo–phase trajectories. The method of the pseudo–phase space is usually used when only one variable \[the discretized vortex polarity $p(t_i)$ in our case\] is measured:[@Moon04] the pseudo–phase–space plot can be made using $p(t_i)$ and its future value $p_{t_{i+1}}$, where the absolute value of the time step $t_{i+1}-t_i$ affects only the shape of the trajectory. In the case of chaotic dynamics, one has open trajectories in pseudo–phase–space $\bigl( p(t_i), p(t_{i+1}) \bigr)$, see the Fig. \[fig:sim:pseudofield\](a). In the regular case, pseudo–phase trajectories are closed, see the Fig. \[fig:sim:pseudofield\](b). Both trajectories are shown for the first 10ns of the dynamics: in the first case the trajectory every time makes a new loop in a different place and in the second case all loops coincide. Below in Sec. \[sec:core\] we construct the phase trajectories for the theoretical model of our system (see Figs. \[fig:core:homega\]a, \[fig:core:homega\]d). \(iv) The green triangles on the switching diagram (see Fig. \[fig:sim:homega\]) correspond to an intermittent process. The typical example of the temporal dynamics in such a regime is plotted in the Fig. \[fig:sim:homega\]a: the vortex state can retain its polarity for a relatively long time of a few nanoseconds; after that multiple reversal processess occur during $50-100$ns. Note that in the vicinity of other regimes in the switching diagram we observed that the vortex polarity, after a few switches, can be ‘frozen’ for the rest of the observation time. For example, two switching events occur during the first 1.2ns ($B_0 = 30$mT and $f=14$GHz, see Fig. \[fig:sim:homega\]a), after that the dynamical polarity has only weak oscillations. A similar picture occurs for $B_0 = 70$mT and $f=17$GHz, where after three reversals during the first nanosecond the resulting polarity remains negative. Since the reversals occur only at the beginning, one can conclude that this occurs because the field is not switched on smoothly. \(v) The last regime corresponds to the field parameters ($B_0$, $f$), where the vortex escapes from the system origin on a long time scale (see the red circles in Fig. \[fig:sim:homega\]). Typically, the vortex starts to move during the first 10ns. The detailed analysis shows that the switching scenario differs essentially from the above mentioned one: the magnetization reversal is mediated by the transient creation and annihilation of a vortex–antivortex pair [@Waeyenberge06] (for details of the axially–asymmetric switching mechanism see Ref.  and references therein). Two examples of the possible trajectories are shown in Fig. \[fig:sim:homega\]b: The trajectory b$_1$ corresponds to the chaotic motion, which is accompanied by numerous reversal events. In the regular regime the vortex trajectory has a smooth shape (b$_2$). When the vortex stays in the center of the sample, polarity switching is accompanied by generation of the radially symmetrical modes. After some time of observation, a new 4-fold symmetry occurs around the vortex, which was mentioned in the Ref.  and linked with the square mesh symmetry used in the OOMMF. When the vortex moves from the center, the switching scenario is changed: a pair antivortex–new vortex is created and the antivortex annihilates with the old vortex. The further magnon dynamics becomes unpredictable. We performed very long–time simulations (up to 30ns) for all parameters from the switching diagram, where the vortex does not leave the disk center (see Fig. \[fig:sim:homega\]): the vortex motion was found for all parameters with $f< 17$GHz. For higher frequencies (e.g., for $f=18$GHz and $B_0=100$mT) the small oscillations of the vortex position were observed only for $t>29$ns. In the prolonged simulations (iv) the vortex polarity does not change its value during the time of observation in agreement with the conclusion made above that the field is sharply switched on. ![(Color online) FFT spectrum of the vortex polarity ($B_0=70$mT, $f=14$GHz): solid line corresponds to the numerical data, the dashed line is the fitting to the pink noise. []{data-label="fig:sim:fourier"}](fourier){width="\columnwidth"} \[page:features\] The switching diagram for the low–frequency range has several new features. One can identify from the plot two local minima (4.5GHz and 6GHz), which correspond to resonances for fractional frequencies ($\frac13f_0^1$ and $\frac12f_0^1$). The strong field causes the vortex polarity reversal, which corresponds to the quasi–static regime and the lower fields cause the escape of the vortex from the system origin. We checked the idea about the quasi–static regime by computing the threshold value for the static field, which is in our case 611mT, cf. Refs. . ![(Color online) (a) The pseudo–phase diagram for the chaotic vortex polarity dynamics (the time step 1ns, the filled circles mark polarities $p=\pm 1$); (b) the same diagram for the regular dynamics.[]{data-label="fig:sim:pseudofield"}](pseudofields "fig:"){width="\columnwidth"}\ (a) $B_0 = 100$mT, $f=16$GHz (b) $B_0 = 100$mT, $f=18$GHz It should be noted that in all simulations discussed above we used the effective 2D mesh $3\times 3\times 21$nm. In order to check our assumption about the uniform magnetization distribution along the thickness z–coordinate, we also performed 3D OOMMF simulations with the mesh size $3 \times 3 \times 3$nm. One can see that eigenfrequencies and boundaries of dynamical regimes are slightly influenced by the nonhomogeneous distribution along the z–coordinate, see Fig. \[fig:sim:homegaDiagram3D\]. It is known that the vortex reversal under the action of a perpendicular static field is accompanied by the temporal creation and annihilation of a Bloch point: the switching process, as a rule, is mediated by the creation of two Bloch points, however, the single Bloch point scenario was also mentioned. [@Thiaville03] The Bloch point propagation during the polarity reversal under the ac perpendicular field was also mentioned by @Yoo12. It should be noted that the Bloch point is a 3D micromagnetic singularity, hence it does not exist in 2D simulations. The dynamics of the Bloch point is not well studied to the moment. During the switching process we observed a new 3D picture of the switching, where some switching events are not completed: the vortex near one face surface rapidly reverses its polarity to opposite and returns back, while the vortex near the second face surface saves its polarity during this time. Description of Different Dynamical Regimes {#sec:core} ========================================== ![(Color online) Comparison of 2D and 3D simulations. Symbols correspond to 3D simulations: The red circles describe the switching process, the open blue boxes describe the dynamics without switching. The dotted line represents the border of the switching region for 3D simulations, colored domains correspond to 2D simulations, see Fig. \[fig:sim:homega\].[]{data-label="fig:sim:homegaDiagram3D"}](homegaDiagramFull3D){width="\columnwidth"} To gain some insight to the switching mechanism, we need a model which allows the magnetization reversal process. It is worth reminding that in the continuum limit the vortex states with different polarities are separated by an infinite barrier. In the spin lattice the barrier becomes finite [@Wysin94] and the reversal can occur. It is already known from our previous paper[@Pylypovskyi13b] that the dominating contribution to the switching mechanism is caused by the exchange interaction inside the vortex core. That is why to describe the polarity reversal process we use here the discrete *reduced core model*, which was initially introduced by @Wysin94 for the vortex instability phenomenon. Later the vortex core model was developed to analyze the vortex polarity switching in Heisenberg magnets.[@Gaididei99; @Gaididei00; @Zagorodny03] One has to note that the reduced core model does not pretend a quantitative agreement with simulations. It is the simplest model which allows to describe a rich variety of different regimes of vortex polarity dynamics, including the resonant behavior, the weakly nonlinear regimes, the reversal dynamics, and the chaotic regime. We consider the anisotropic classical Heisenberg disk shaped system with thickness $L_z$ and the radius $L$, assuming that the magnetization of the magnet is uniform along the thickness. In terms of the normalized magnetic moment $$\label{eq:m} {\bm{m}}_{{\bm{n}}} = \left(\sqrt{1-m_{{\bm{n}}}^2}\cos\phi_{{\bm{n}}}, \sqrt{1-m_{{\bm{n}}}^2} \sin\phi_{{\bm{n}}}, m_{{\bm{n}}}\right)$$ the energy of such a magnet with the account of the interaction with magnetic field reads $$\label{eq:Energy} \begin{split} E &= - \frac{A L_z}{2} \sum_{({\bm{n}},{\bm{\delta}})} \left[ {\bm{m}}_{{\bm{n}}}\cdot {\bm{m}}_{{\bm{n}}+{\bm{\delta}}} - (1-\lambda) m_{{\bm{n}}}^z m_{{\bm{n}}+{\bm{\delta}}}^z \right]\\ &- a_0^2 M_s L_z \sum_{{\bm{n}}} {\bm{m}}_{{\bm{n}}}\cdot {\bm{B}}(t), \end{split}$$ where the vector ${\bm{\delta}}$ connects nearest neighbors of the three–dimensional cubic lattice with the lattice constant $a_0$, $A$ is the exchange constant, the parameter $\lambda\in(0,1)$ is the effective anisotropy constant, and $M_s$ is the saturation magnetization. According to this model the planar vortex is stable when $\lambda < \lambda_c$, where $\lambda_c \approx 0.72$ for the square lattice.[@Wysin94] In a cylindrical frame of reference $(r,\chi,z)$ the planar vortex distribution is described by \[eq:vortex\] $$\label{eq:planar-vortex} m_v = 0, \qquad \phi_v = \chi + \mathcal{C},$$ where $\mathcal{C} = \pm \pi/2$ is a vortex chirality (we use here the positive sign in calculations below). When $\lambda>\lambda_c$, the nonplanar vortex appears, which is characterized by the well–localized out–of–plane magnetization $m_v\neq0$.[@Wysin94] (-0.5,-0.5) grid (3.5,3.5); (0,0) – (0.9, 0.9); (0,0) – (1.20344867377942, 0.4143806095590807); (1,1) circle (3pt); (1,2) circle (3pt); (2,1) circle (3pt); (2,2) circle (3pt); (0,0) circle (2pt); (0,3) circle (2pt); (3,0) circle (2pt); (3,3) circle (2pt); (1,0) circle (2pt); (2,0) circle (2pt); (3,1) circle (2pt); (3,2) circle (2pt); (2,3) circle (2pt); (1,3) circle (2pt); (0,2) circle (2pt); (0,1) circle (2pt); (0,0) – (0.75895,-0.25298); (0,0) – (0.75895,0.25298); (0,0) – (0.25298, 0.75895); (0,0) – (-0.25298, 0.75895); (0,0) – (-0.75895,0.25298); (0,0) – (-0.75895,-0.25298); (0,0) – (0.25298, -0.75895); (0,0) – (-0.25298, -0.75895); (0,0) – (0.56569, 0.56569); (0,0) – (-0.56569, 0.56569); (0,0) – (-0.56569, -0.56569); (0,0) – (0.56569, -0.56569); (0,0) – (0.7564148604794535, 0.26045452356572535); (0,0) – (-0.26045452356572535, 0.7564148604794535); (0,0) – (-0.7564148604794536, -0.2604545235657251); (0,0) – (0.26045452356572507, -0.7564148604794536); (0,0) arc (20:42:8mm); at (0,0) [$\psi$]{}; (0,0) arc (100:170:3mm); (0,0) arc (190:260:3mm); (0,0) arc (280:350:3mm); (0,0) arc (10:80:3mm); at (0,0) [4]{}; at (0,0) [1]{}; at (0,0) [2]{}; at (0,0) [3]{}; at (0,0) [12]{}; at (0,0) [5]{}; at (0,0) [6]{}; at (0,0) [7]{}; at (0,0) [8]{}; at (0,0) [9]{}; at (0,0) [10]{}; at (0,0) [11]{}; In the reduced core approach, we suppose that only the four magnetic moments of the first coordinate shell can vary, forming the vortex core; all the other moments are fixed in the sample plane in a vortex configuration , see Fig. \[fig:core-schematic\]. By symmetry, all four moments are characterized by the same out–of–plane magnetization $\mu$ and equal in–plane phase $\psi$, which is determined as a deviation from the vortex configuration. Therefore, the magnetization distribution of the first coordinate shell is describe as follows: $$\label{eq:magnMomTurnPh} m_{i}^z = \mu,\quad \phi_{i} = \chi + \mathcal{C} + \psi, \quad i=\overline{1,4}.$$ We consider the core magnetization $\mu$, which has the meaning of the *dynamical vortex polarity*, and the in–plane *turning phase* $\psi$ as two collective variables. The energy of the model, normalized by $\epsilon=8AL_z\lambda$ has the form (see Appendix \[app:core\] for details): $$\label{eq:E-core} \mathscr{E} = - \frac{\mu^2}{2} - \Lambda\sqrt{1-\mu^2}\cos\psi - \mu h \sin\omega \tau,$$ where we introduced the reduced anisotropy parameter $\Lambda = 2/(\lambda\sqrt{5})$, the reduced field intensity $h = a_0^2M_sB_0/(2A\lambda)$, the reduced field frequency $\omega = 2\pi f M_s/(\epsilon\gamma)$, and the rescaled time $\tau = \epsilon\gamma t/M_s$. We use $\Lambda = 0.9415$ ($\lambda=0.95$) and $\eta = 0.002$ in a majority of numerical calculations below. Note that such a value of the $\Lambda$–parameter is chosen for illustrative purposes, it does not fit to the correct material parameters from simulations. The magnetization dynamics in the reduced core model can be described by the following equations (see Appendix \[app:core\] for details): $$\label{eq:core:motion} \begin{split} \dot \mu &= \Lambda\sqrt{1-\mu^2}\sin\psi + \eta \Bigl[ \mu(1-\mu^2)\\ & - \Lambda \mu\sqrt{1-\mu^2} \cos\psi + h(1-\mu^2) \sin \omega\tau \Bigr],\\ \dot \psi &= \mu - \frac{\Lambda \mu \cos\psi}{\sqrt{1-\mu^2}} + h\sin \omega\tau - \frac{\eta \Lambda \sin\psi}{\sqrt{1-m^2}}, \end{split}$$ where the overdot means the derivative with respect to $\tau$ and $\eta$ is a Gilbert damping coefficient. The ground state of the model corresponds to $$\label{eq:ground-state} \mu_0 = \pm \sqrt{1-\Lambda^2}, \qquad \psi_0 = 0.$$ In terms of the core model two opposite values of $\mu_0$ describe vortices with opposite polarities $\mu_0$. ![(Color online) Nonlinear resonance curve: The numerical solution of Eqs.  (the dashed line) and the analytical solution (the solid line) for the following parameters: $\Lambda = 0.9415$, $h=0.0002$, $\eta = 0.01$, $\mu(0) = 0.337$, $\psi(0) = 0$. The effective double–well potential $\mathscr{U}$, see , is plotted in the inset. []{data-label="fig:core:nonlinAmp"}](nonlinAmp){width="\columnwidth"} at (0,2.5) [{width="\textwidth"}]{}; [(pointer) at (1.9,4.12); (leftCorner) at (0,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer) at (1.85,4.17); (leftCorner) at (3.025,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer) at (7.05,4.1); (leftCorner) at (6.05,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer) at (7.,4.15); (leftCorner) at (9.075,5.5); (leftCorner) – ($(leftCorner) + (2.925/2-0.1,0)$) – (pointer) – ($(leftCorner) + (2.925/2+0.1,0)$) – ($(leftCorner) + (2.925,0)$) – ($(leftCorner) + (2.925,1.9)$) – ($(leftCorner) + (0,1.9)$) – cycle; at ($(leftCorner) + (2.925/2-0.01,1.9/2)$)[]{};]{}; [(pointer) at (2.77,3.7); (leftCorner) at (0.75,0.8); (leftCorner) – ($(leftCorner) + (3/2-0.1,0)$) – (pointer) – ($(leftCorner) + (3/2+0.1,0)$) – ($(leftCorner) + (3,0)$) – ($(leftCorner) + (3,-1.9)$) –($(leftCorner) + (0,-1.9)$) – cycle; at ($(leftCorner) + (3/2-0.01,-1.9/2)$)[]{};]{}; [(pointer) at (4.6,0.87); (leftCorner) at (8.9,2.21); (leftCorner) – ($(leftCorner) + (3,0)$) – ($(leftCorner) + (3,-1.9)$) – ($(leftCorner) + (0,-1.9)$) – ($(leftCorner) + (0,-1.9/2-0.1)$) – (pointer) – ($(leftCorner) + (0,-1.9/2+0.1)$) – cycle; at ($(leftCorner) + (3/2-0.01,-1.9/2)$)[]{};]{}; Let us start our analysis with a system without damping, $\eta=0$. Supposing that the turning phase is small enough, $|\psi|\ll1$, one can easily exclude $\psi$ from the consideration. In this case the Eqs.  correspond to the effective Lagrangian $$\label{eq:eff-L} \begin{split} \mathscr L = & \frac{\mathscr M \dot\mu^2}{2} - \mathscr{U}(\mu) + \mu h \sin \omega\tau, \\ \mathscr M = & \frac{1}{\Lambda \sqrt{1-\mu^2}},\qquad \mathscr{U}(\mu) = - \frac{\mu^2}{2} - \Lambda\sqrt{1-\mu^2}. \end{split}$$ This simplification allows us to interpret the complicated dynamics as the motion of a particle with variable mass $\mathscr M$ in the double–well potential $\mathscr U(\mu)$ under a periodic pumping, see the inset in the Fig. \[fig:core:nonlinAmp\]. The linear oscillations near the equilibrium state correspond to the harmonic oscillations of the effective particle in one of the wells; the eigenfrequency of such oscillations is $$\label{omega0} \omega_0 = \sqrt{1-\Lambda^2}.$$ Let us describe the nonlinear regime of the dynamics. In spite of the small damping in the system, its value can be comparable with the pumping intensity. Therefore we consider below the full set of the model equations Eqs. . To analyze the weakly nonlinear regime, we use the multiscale perturbation method[@Kevorkian81; @Nayfeh85; @Nayfeh08]. When the field intensity is much less than the frequency detuning ($h \ll |\omega - \omega_0|$), we can limit ourselves to a three–scale expansion in the form $$\label{eq:core:smallAmp} \begin{split} \mu = \mu_0 + \sum_{n = 1}^3 \varepsilon^n \mu_n(T_0, T_1, T_2),\quad T_n = \varepsilon^n\tau,\\ \psi = \sum_{n = 1}^3 \varepsilon^n \psi_n(T_0, T_1, T_2),\quad \omega = \omega_0 + \omega_\pm,\\ \omega_\pm = \varepsilon^2 \omega_2,\quad \eta = \varepsilon^2 \eta_2, \quad h = \varepsilon^3h_3. \end{split}$$ Using such the expansion, one can derive from Eqs.  the resonance curve $\omega_\pm(h)$ as the solution of the following equation $$\label{eq:core:resonanceCurve} \begin{split} h^2\Lambda^4 & =a^2\left(2\sqrt{1-\Lambda ^2}\omega_{\pm}+\frac{2+\Lambda ^2}{2\Lambda ^2}a^2\right)^2\\ & +{\eta ^2\left(1-\Lambda ^2\right)\left(2-\Lambda ^2\right)^2}a^2, \end{split}$$ with $|a|$ being the amplitude of oscillations, see Appendix \[app:wna\] for details. The typical nonlinear resonance curve is shown in the Fig. \[fig:core:nonlinAmp\], cf. Fig. \[fig:sim:nonlinAmp\]. We analyze the strongly nonlinear regimes solving numerically Eqs.  in a wide range of parameters ($\omega,h$), see the diagram of switching events in the Fig. \[fig:core:homega\]. The absolute minimum in the diagram corresponds to the switching in the range near the resonance frequency $\omega_0$. Other local minima correspond to resonances at double frequency, $2\omega_0$, and sub–harmonics, $\omega_0/2$ and $\omega_0/3$. Note that all resonance frequencies are shifted in the low–frequency direction due to the nonlinear resonance. We classified the dynamical regimes by using the method of Poincaré maps ($15\cdot 10^3$ points per map). We constructed such maps for each pair ($\omega$, $h$) where the switching takes place. One can separate four oscillation regimes related to the corresponding regimes in the OOMMF simulations (Sec. \[sec:sim\]) with vortex dynamics in the center of the sample: (i) the absence of switching, (ii) the chaotic dynamics, (iii) the regular polarity oscillations between two polarities $\pm \mu_0$, (iv) the switching with final oscillations around one of the points $(\pm \mu_0, 0)$ in the coordinates $(\mu,\psi)$. \(i) The border between the switching region and the region, where field or frequency are not enough for jumps between $(\pm \mu_0,0)$, shows a few well-defined resonance minima corresponding to resonances at $\omega_0/3$, $\omega_0/2$, $\omega_0$ and $2\omega_0$. [0.32]{} {width="\columnwidth"} [0.32]{} {width="\columnwidth"} [0.32]{} {width="\columnwidth"} \(ii) The chaotic dynamics occurs for the low-frequency part of the Fig. \[fig:core:homega\] and in the stretched region between resonances at $\omega_0$ and $2\omega_0$. An example of the temporal evolution is shown in the Fig. \[fig:core:homega\]c. The projection of the phase diagram on the $(\mu,\psi)$ plane (see Fig. \[fig:core:homega\]d) looks similarly to the pseudo–phase diagram in the Fig. \[fig:sim:pseudofield\](a): the projection of trajectory is not closed and representation point makes a lot of windings around both $\pm \mu_0$. The shape of the chaotic Poincaré maps depends on the frequency. They show the shape of strange attractors, see Fig. \[fig:core:poincareEvolution\]. Their Cantor structure is ill–defined due to using low damping. Note, that they are similar to strange attractors for the Duffing oscillator[@Moon04] (nonlinear oscillator with quadratic and cubic nonlinearities in the double-well potential). However, the reduced core model has a more complicated nonlinearity term; using the mechanical analogy one can speak about the motion of a particle with a variable mass (7) in a double–well potential. \(iii) The main part of the diagram of switching events \[fig:core:homega\] is occupied by the region of regular dynamics. The most frequently observed Poincaré maps for this case contain some number of stable focuses. The observed numbers are $\overline{1,12}$, 15, 16, 18, 21, 24, 30 and 96. The most frequently observed ones are 1 (the grey region in the Fig. \[fig:core:homega\]) and 3 (included into the blue region in the Fig. \[fig:core:homega\]). Some of the points with a higher number of focuses demonstrate a complicated regular dynamics in phase space, see Fig. \[fig:core:homega\]e. The analogue of quasi–rectangular regular polarity oscillations in OOMMF simulations is found in the $\omega_0/3$ region, see phase diagram in the Fig. \[fig:core:homega\]a and temporal evolution in the Fig. \[fig:core:homega\]b (compare with the pseudo–phase diagram shown in the Fig. \[fig:sim:pseudofield\](b) and the temporal evolution in the Fig. \[fig:sim:homega\]d). \(iv) The analogue of the intermittent switching linked with perturbation by the applying of the external field are shown by two dark–grey color intensities in the Fig. \[fig:core:homega\]. The final dynamics is an oscillation around upward or downward polarity (points $(\mu_0,0)$ and $(-\mu_0,0)$ in the phase space projection, respectively). An example of the temporal evolution is shown in the Fig. \[fig:core:homega\]f. As in OOMMF simulations, such oscillations typically occur near the border of the switching region. As in case of the chaotic dynamics, the resulting polarity is highly dependent on field, frequency and integration conditions. Controlled switching {#sec:control} ==================== [cc]{} \[41pt\] [0.48]{} {width="\columnwidth"} & [0.48]{} {width="\columnwidth"} \ & [0.48]{} {width="\columnwidth"} \ As it is shown by analysis of the diagram of switchings events (Fig. \[fig:sim:homega\]), the vortex polarity switching under the action of the perpendicular resonant field produces multiple switchings during a short time comparable with one period of the acting field. Such a situation is not unique among different polarity switching methods. So, for the axially–asymmetric scenario, sufficient strength of gaussian pulse in the sample’s plane produces more than one sequential vortex–antivortex pair creation and annihilations.[@Hertel07] Using an in–plane rotating field with frequency $\omega_r$ codirectional with the vortex polarity $p$ ($\omega_r p > 0$) stabilizes the vortex in the center of the sample. But when $\omega_r p < 0$ the reversal occurs in a specific range of the field intensities and frequencies, above which multiple switching was observed. [@Kravchuk07c] A similar picture was reported in Ref.  for the current–induced switching. However for the further application in contrast to multiple switching a unidirectional vortex polarity reversal, is needed. Because the pumping  does not select any direction of the vortex polarity, the most natural way to avoid the multiple reversal consists in limiting the pulse duration. We test an influence of a short wave train in the form $$\label{eq:sim:trainField} {\bm{B}} = \begin{cases} B_0{\bm{e}}_z \sin \left( 2\pi f t\right),& t \in \left[0,\dfrac{N}{f}\right], \\ 0, & \text{otherwise}, \end{cases}$$ where $N\in\mathbb{N}$ is the number of periods of the sinusoidal magnetic field in the wave train. We investigate the vortex dynamics under the action of the field on the resonant frequency $f = f_0$. The vortex polarity is observed during a long time $10N/f$ in order to damp magnons. The response of the magnetization to the field shows the nonlinear dependency on $B_0$ and $N$, see Fig. \[fig:train\]. When field intensity and period numbers are small, there are only small oscillations of the magnetization inside the vortex core. When $N$ becomes larger than some critical value, a typical system behaviour looks like a few switchings, which also occurs when the field is already turned off. A unidirectional switching from upward polarity to downward is observed for $B_0 = 30$mT and $N=6,7$ and some higher fields. We check the controllability of the discussed switching method by applying of series of wave trains. The series of wave trains is applied to the relaxed vortex. The time interval between trains is varied with steps of $0.5/f_0$ in different series. For $B_0=30$mT and $N=6$ the first switching occurs in 627ps and the vortex starts to relax (field is turned off at the time 429ps). The sequence containing 3 wave trains allows to get controllable unidirectional vortex polarity reversal with minimal interval between wave trains of $3.2$ns, see Fig. \[fig:unidirectional\]. These time intervals correspond speed of changing state of such memory cell about 250MHz. The core model also gives the same qualitative result, see Fig. \[fig:unidirectionalCore\]. Discussion {#sec:conclusion} ========== The axially–symmetric switching of the vortex core under the action of periodic pumping was very recenltly predicted in Refs. . @Wang12 were concerned with switching events: for the typical nanodot size the switching at the resonant frequency occurs during 600ps. @Yoo12 computed the diagram of switching events where they noticed the existence of resonances on double and triple harmonics and shown that the exchange energy becomes higher than the threshold value for the vortex core reversal. In this work we study the long– and short–time vortex dynamics and propose an analytical model which describes the phenomena of full–scale simulations. In order to explain the complicated vortex dynamics, we use here the reduced core model. [@Wysin94; @Gaididei99; @Gaididei00; @Zagorodny03] It should be noted that this model does not pretend a quantitative agreement with simulations. In particular, it does not provide even the eigenfrequencies of the radially symmetric magnon modes, which is rather complicated task. [@Zaspel09; @Galkin09] Nevertheless, the model we use is the simplest one which allows to describe a rich variety of different regimes of vortex polarity dynamics. This model provides a simple physical picture of the switching phenomenon in terms of the nonlinear resonance in a double–well potential. Such a potential arises mainly from the exchange interaction: the presence of two wells corresponds to the energy degeneracy with respect to the direction of the vortex polarity (up or down); the energy barrier between the wells becomes higher as the discreteness effects become less important. One has to stress that the switching process is forbidden in the continuum theory. In a real magnet the magnetization reversal is possible due to the discreteness of the lattice. That is why to describe the switching analytically we use the discrete core model: the switching can be considered as the motion of an effective mechanical particle with a variable mass in the double–well potential. Under the action of periodical pumping the particle starts to oscillate near the bottom of one of the wells. When the pumping increases, there appear nonlinear oscillations of the particle; under a further forcing the particle overcomes the barrier, which corresponds to the magnetization reversal process. The chaotic dynamics of the magnetization is studied for domains walls[@Shutyi07] and current–induced phenomena in monodomain nanoparticles.[@Lee04a; @Berkov05; @Yang07a] Very recently the existence of incommensurate chaotic vortex dynamics in spin valves was reported.[@Petit-Watelot12] In our case the chaos enters in the vortex polarity switching process due to the periodical pumping of the system with two equivalent equilibrium states as it happens in a Duffing oscillator.[@Moon04] The periodic pumping does not select the preferable vortex polarity direction which causes multiple switchings under the action of sufficiently high fields and frequencies. However, accurate fitting of the pulse duration and the time interval between sequential pulses allows to obtain a controlled unidirectional core reversal. Thus, the chaotic, regular and controlled vortex polarity dynamics could find applications in physical layer data encryption[@Kocarev95; @Schoell08] and memory devices.[@Kim08; @Yu11a] In the current study we do not consider thermal effects on the vortex dynamics. The influence of the temperature was found to be not essential for the current–induced motion of an individual vortex in Py nanodisks [@Ishida06]. Nevertheless it should be noted while magnetic vortices are stable up to very high temperature,[@Muxworthy03] the heating can influence the gyroscopical vortex dynamics [@Kamionka11; @Depondt12]. The heat can induce the vortex dynamics in the system. [@Machado12] The temperature–induced vortex dynamics also can influence the critical fields for vortex nucleation and annihilation.[@Mihajlovic10] We expect that the physical picture discussed in the paper with a variety of different dynamical regimes survives with the temperature. The thermal effects will cause the shift of boundaries between different regimes. O.V.P. and D.D.S. thank the University of Bayreuth and Computing Center of the University of Bayreuth, where a part of this work was performed, for kind hospitality. O.V.P. acknowledges the support from the BAYHOST project. D.D.S. acknowledges the support from the Alexander von Humboldt Foundation. F.G.M. acknowledges the support by MICINN through FIS2011-24540, and by Junta de Andalucia under Project No. FQM207. All simulations results presented in the work were obtained using the computing cluster of Kiev University [@unicc] and Bayreuth University [@btrzx]. Reduced Core Model {#app:core} ================== Taking into account the explicit form of the magnetic field, , the energy reads[@Zagorodny03] $$\label{eq:Energy-1} \begin{split} E &= - \frac{A L_z}{2}\!\!\!\sum_{({\bm{n}},{\bm{\delta}})}\! \Biggl[\! \sqrt{(1-m_{{\bm{n}}}^2)(1-m_{{\bm{n}} +{\bm{\delta}}}^2)}\cos(\phi_{{\bm{n}}}\! -\phi_{{\bm{n}}+{\bm{\delta}}})\\ & + \lambda m_{{\bm{n}}} m_{{\bm{n}}+{\bm{\delta}}} \Bigr]- a_0^2 M_s L_z B_0 \sin\left(2\pi f t\right) \sum_{{\bm{n}}} m_{{\bm{n}}}. \end{split}$$ Now we incorporate here the reduce core Ansatz . Then the energy reads $$\label{eq:Energy-2} \begin{split} E &= -4a_0^2M_sL_z \mu B_0\sin(2\pi f t) - 4AL_z\lambda \mu^2 - \\ & - \frac{16}{\sqrt{5}}AL_z\sqrt{1-\mu^2}\cos\psi. \end{split}$$ After the renormalization the Eq. \[eq:Energy-2\] takes the form , where $ \mathscr{E} = E/\epsilon$, $\epsilon = 8AL_z\lambda$. The magnetization dynamics follows the Landau–Lifshitz–Gilbert equations $$\label{eq:LL} \frac{\mathrm d{\bm{m}}_{{\bm{n}}}}{\mathrm d\tau} = {\bm{m}}_{{\bm{n}}}\times \frac{\partial \mathscr{E}}{\partial {\bm{m}}_{{\bm{n}}}} + \eta {\bm{m}}_{{\bm{n}}} \times \frac{\mathrm d{\bm{m}}_{{\bm{n}}}}{\mathrm d\tau},$$ with $\eta$ being a Gilbert damping coefficient and the rescaled time $\tau = \epsilon\gamma t/M_s$. Substituting Eqs.  into the Eq.  we obtain the equations for the $(\mu,\psi)$: $$\label{eq:LLGnew} \begin{aligned} \frac{\mathrm d\mu}{\mathrm d\tau} & = \frac{\partial \mathscr E}{\partial \psi} - \eta (1-\mu^2) \frac{\partial \mathscr E}{\partial \mu},\\ \frac{\mathrm d\psi}{\mathrm d\tau} & = - \frac{\partial \mathscr E}{\partial \mu} - \frac{\eta}{1-\mu^2} \frac{\partial \mathscr E}{\partial \psi}, \end{aligned}$$ Substituting Eq.  into we get the Eq. . Weakly nonlinear analysis {#app:wna} ========================= Let us consider the weakly nonlinear case for Eqs. . Using the series expansion , the time derivative reads $$\frac{\mathrm{d}}{\mathrm{d}t} = \sum_{n=0}^2\varepsilon^n D_n,\quad D_n = \frac{\mathrm{d}}{\mathrm{d}T_n},$$ and the equations of motion  break into three pairs of equations for the different orders in $\varepsilon$: $$\begin{aligned} \label{eq:app:firm} D_0 \mu_1 & = \Lambda^2 \psi_1,\\ \label{eq:app:firpsi} D_0 \psi_1 & = \left(1-\frac{1}{\Lambda ^2}\right) \mu_1,\\ \label{eq:app:secm} D_1 \mu_1 & + D_0 \mu_2 = - \sqrt{1-\Lambda ^2} \mu_1 \psi_1 + \Lambda ^2 \psi_2,\\ \label{eq:app:secpsi} D_1 \psi_1 & + D_0 \psi_2 = \left(1-\frac{1}{\Lambda ^2}\right) \mu_2 - \frac{3 \sqrt{1-\Lambda ^2}}{2 \Lambda ^4} \mu_1^2 \\ & + \frac{\sqrt{1-\Lambda ^2}}{2} \psi_1^2, \nonumber \\ \label{eq:app:thirm} D_2 \mu_1 & + D_1 \mu_2 + D_0 \mu_3 = \Lambda^2 \psi_3 - \frac{1}{2 \Lambda ^2} \mu_1^2\psi_1\\ & - \sqrt{1-\Lambda ^2} (\mu_2\psi_1 + \mu_1 \psi_2) - \frac{\Lambda ^2}{6} \psi_1^3 \nonumber\\ & -\left(1-\Lambda ^2\right) \eta_2 \mu_1,\nonumber \\ \label{eq:app:thirpsi} D_2 \psi_1 & + D_1 \psi_2 + D_0 \psi_3 = \left(1-\frac{1}{\Lambda ^2}\right) \mu_3\\ & - \frac{5-4 \Lambda ^2}{2 \Lambda ^6} \mu_1^3 - \frac{3 \sqrt{1-\Lambda ^2}}{\Lambda ^4} \mu_1 \mu_2 + \frac{1}{2 \Lambda ^2} \mu_1 \psi_1^2 \nonumber \\ & +\sqrt{1-\Lambda ^2} \psi_1\psi_2 + h_3\sin \omega t - \eta_2 \psi_1. \nonumber\end{aligned}$$ The solution of Eqs.  and reads $$\mu_1(T_0, T_1, T_2) = A(T_1, T_2) e^{i \omega_0 T_0} + A^*(T_1, T_2) e^{-i \omega_0 T_0}.$$ Following the Floquet theory [@Nayfeh85], one needs to omit all secular terms. Thus, Eqs.  and show $A(T_1,T_2) \equiv A(T_2)$ and Eqs.  and gives the equation for $A(T_2)$: $$\label{eq:app:lasteq} \begin{split} 2 i \sqrt{1-\Lambda ^2} A' & + i \eta_2 \sqrt{1-\Lambda ^2} \left(2-\Lambda ^2\right) A \\ & - 2\frac{2+\Lambda^2}{\Lambda^2} A^2 A^* = \Lambda^2 \frac{h_3}{2i}e^{i\omega_2 T_2}. \end{split}$$ By solving the Eq.  one obtains the Eq. . [51]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****,  ()](\doibase 10.1080/00018732.2012.663070) @noop [**]{} (, , ) [****,  ()](\doibase 10.1126/science.1075302) [****,  ()](\doibase 10.1063/1.2807274) [****,  ()](\doibase 10.1063/1.3551524) [****,  ()](http://link.aps.org/abstract/PRB/v59/p7010) [****,  ()](\doibase 10.1103/PhysRevB.61.9449) [****,  ()](http://dx.doi.org/10.1038/nature05240) [****,  ()](http://www.sciencedirect.com/science/article/B6TJJ-447DCMV-3/2/84e9fdfbf9e0cab3d3182fc4db4b4032) [****, ()](http://link.aps.org/abstract/PRB/v67/e094410) [****,  ()](http://dx.doi.org/10.1134/S1063783407100186) [****,  ()](\doibase 10.1103/PhysRevB.79.172410) [****,  ()](\doibase 10.1103/PhysRevB.80.100405) [****,  ()](\doibase 10.1103/PhysRevB.81.094431) [****, ()](\doibase 10.1063/1.2957013) [****,  ()](\doibase 10.1063/1.3687909) [****,  ()](\doibase 10.1063/1.4705690) [****,  ()](http://www.ujp.bitp.kiev.ua/files/journals/58/6/580611p.pdf) [****,  ()](\doibase 10.1140/epjb/e2003-00057-y) [****,  ()](http://stacks.iop.org/0022-3727/43/i=26/a=264004) [“,” ](http://math.nist.gov/oommf/) [****, ()](http://link.aps.org/abstract/PRL/v98/e117201) @noop [**]{} (, ) p.  [****, ()](\doibase 10.1103/PhysRevB.49.8780) [**](http://books.google.com.ua/books?id=pT3vAAAAMAAJ), Applied mathematical sciences (, ) [**](http://books.google.com.ua/books?id=3zbvAAAAMAAJ), A Wiley Interscience publication (, ) [**](http://books.google.com.ua/books?id=eh6RmWZ51NIC), Physics textbook (, ) [****,  ()](\doibase 10.1063/1.2770819) [****,  ()](\doibase 10.1063/1.2780107) [****,  ()](\doibase 10.1103/PhysRevB.80.094415) [****, ()](http://dx.doi.org/10.1134/S1063776109070103) [****,  ()](http://dx.doi.org/10.1134/S106377610705010X),  [****,  ()](http://dx.doi.org/10.1038/nmat1237) [****,  ()](\doibase 10.1103/PhysRevB.71.052403) [****,  ()](\doibase 10.1103/PhysRevLett.99.134101) [****,  ()](http://dx.doi.org/10.1038/nphys2362) [****,  ()](\doibase 10.1103/PhysRevLett.74.5028) , ed., @noop [**]{},  ed. (, ) p.  [****, ()](http://link.aps.org/abstract/PRB/v74/e014424) [****,  ()](http://dx.doi.org/10.1029/2002JB002195) [****,  ()](\doibase 10.1103/PhysRevB.83.224424) [****,  ()](\doibase 10.1016/j.physleta.2012.09.039) [****,  ()](\doibase 10.1063/1.3694757) [****,  ()](\doibase 10.1063/1.3360841) [](http://cluster.univ.kiev.ua/eng/) [](http://www.rz.uni-bayreuth.de/) @noop [**]{}, edited by  (, ) [^1]: We also check the polarity dynamics for the samples with the same diameter and height 33nm, with diameter 360nm and heights 21nm and 33nm under the action of five different pairs of field intensity and frequency: (60mT, 10GHz), (60mT, 13GHz), (80mT, 13GHz), (100mT, 13GHz), (100mT, 18GHz). The qualitative behavior of the system described in this section remains the same. Nevertheless the geometrical parameters influence the frequencies of magnon modes and the threshold value of the switching field. Therefore, we expect that the change of the nanodot size shifts the characteristic frequencies of the minimal threshold fields and the threshold field amplitudes, while the qualitative behavior of the system remains the same. [^2]: The field and frequency increments are varied depending on the position on the diagram of switching events: 10mT and 1GHz in the range $9\div19$GHz and $10\div180$mT, 10mT and 0.5GHz in the range $3\div8.5$GHz and $100\div500$mT along the border of the diagram, 50mT and 1GHz for other ranges. For the frequencies $f > 8.5$GHz the full time of simulations was 10ns \[extended till 30ns for some particular points $(f,B_0)$\] and for frequencies $f \le 8.5$GHz the full time of simulations was 5ns. [^3]: Note that the vortex polarity instability which correspond to the nonlinear resonance was predicted in our previous paper, see Ref. . [^4]: As opposed frequency–independent white noise, the pink noise is characterized by the power law decay with $1/f^\beta$ spectrum, where $\beta\in(0,2)$, see Ref. .

ArXiv

--- abstract: 'The GIT chamber decomposition arising from a subtorus action on a quasiprojective toric variety is a polyhedral complex. Denote by $\Sigma$ the fan that is the cone over the polyhedral complex. In this paper we show that the toric variety defined by the fan $\Sigma$ is the normalization of the toric Chow quotient of a closely related affine toric variety by a complementary torus.' address: - 'Department of Mathematics, Stony Brook University, Stony Brook, NY 11794' - 'Department of Mathematics, Hill Center-Busch Campus, Rutgers University, 110 Frelinghuysen Rd, Piscataway, NJ 08854' author: - Alastair Craw - Diane Maclagan title: Fiber fans and toric quotients --- Introduction ============ Let $X_P$ be the quasiprojective toric variety defined by a full-dimensional polyhedron $P\subseteq \Q^n$, and let $T_L = \Spec \K[L]$ be a $d$-dimensional subtorus of the dense torus $(\K^*)^n$ of $X_P$. The inclusion of $T_L$ into $(\K^*)^n$ induces a linear map $\pi\colon \Q^n\rightarrow \Q^d \cong L\otimes_{\Z} \Q$ and hence a linear projection of polyhedra $\pi\colon P\rightarrow Q$, where $Q = \pi(P)$. Thaddeus [@Thaddeus] proved that $\pi$ induces a cell decomposition of $Q$ that is the GIT chamber decomposition arising from the action of $T_L$ on $X_P$. See Hu [@Hu] for the most general results about toric GIT, including quotients of toric varieties that do not necessarily arise from polyhedra. By taking the cone over each cell in this GIT chamber decomposition, we obtain a fan $\Sigma$ in $\Q^d\oplus \Q$. It is natural to ask for an explicit description of the toric variety defined by this fan. The main result of this paper provides an answer to this question as follows (see Theorem \[thm:main\] and Section \[sec:GeneralDuality\] for relevant notation): \[t:main\] Let $T_L$ be a subtorus acting on a toric variety $X_P$, and let $\Sigma$ be the fan arising from the GIT chamber decomposition associated to the $T_L$-action on $X_P$ as above. Then there is a polyhedral cone $P':= (\widetilde{P})^\vee$ and a lattice $L':=\image(\widetilde{\pi}^\vee)$ such that $\Sigma$ is the fan of the toric Chow quotient $X_{P'}/T_{L'}$. The Chow quotient $X_{P'}/T_{L'}$ of a *projective* toric variety $X_{P'}$ by a subtorus $T_{L'}$ was studied by Kapranov–Sturmfels–Zelevinsky [@KSZ]. We generalize this by constructing the *toric Chow quotient* $X_{P}/T_{L}$ of an arbitrary quasiprojective toric variety $X_{P}$ by a subtorus $T_{L}$ (see Definition \[defn:toricchow\]). In order to prove Theorem \[t:main\], we consider the *fiber fan* associated to a linear projection of polyhedra $\pi\colon P\rightarrow Q$, denoted $\mathcal{N}(P,Q)$, generalizing the normal fan of a fiber polytope (see Billera–Sturmfels [@BilleraSturmfels]). We show that the toric Chow quotient $X_P/T_L$ is a not-necessarily normal toric variety whose fan is the fiber fan $\mathcal{N}(P,Q)$. In the special case where $P$ is a polytope, so $X_P$ is projective, this result is due to Kapranov–Sturmfels–Zelevinsky [@KSZ]. Note however that the statement of Theorem \[t:main\] is new even in the case where $P$ is a polytope. If the original polyhedron $P$ is a polyhedral cone, the statement of Theorem \[t:main\] can be streamlined, and a converse can be added as follows (see Theorem \[thm:affineduality\]). \[t:maintwo\] Let $P\subseteq \Q^n$ be a polyhedral cone and $\pi\colon P\rightarrow Q$ a linear projection. Set $P' := P^\vee$ and $L':= \image(\pi^\vee)$. Then 1. the fan of the toric Chow quotient $X_{P}/T_{L}$ is equal to the GIT chamber decomposition arising from the action of $T_{L'}$ on $X_{P'}$; and 2. the GIT chamber decomposition arising from the $T_L$-action on $X_P$ is the fan of the toric Chow quotient $X_{P'}/T_{L'}$. The statement of Theorem \[t:maintwo\], part , was partially known to Altmann–Hausen [@AltmannHausen] in the case where $P$ is a simplicial cone, though no proof was given. The main theorems are proved in Section \[s:mainproof\]. Section \[s:polyhedral\] contains the necessary polyhedral preliminaries, while Sections \[sec:toricGIT\] and \[sec:toricChow\] contain background on toric GIT, and introduce the toric Chow quotient, respectively. **Acknowledgements** Special thanks are due to Rekha Thomas for providing a great deal of help. This project arose out of our joint work in [@CMT1]. We also thank Klaus Altmann and Mark Haiman for useful conversations. The second author was partially supported by NSF grant DMS-0500386. Polyhedral Geometry {#s:polyhedral} =================== Polyhedral Conventions ---------------------- Let $P \subseteq \mathbb Q^n$ be an $n$-dimensional polyhedron. For $\mathbf{w} \in (\mathbb Q^n)^*$ we denote by $\face_{\mathbf{w}}(P)$ the face of $P$ minimizing $\mathbf{w}$. Given a face $F$ of $P$, the *inner normal cone* $\mathcal N_{P}(F)$ is the set of ${\textbf w} \in (\Q^n)^{\ast}$ such that $\face_{\mathbf{w}}(P)=F$. The *inner normal fan* $\mathcal{N}(P)$ of $P$ is the polyhedral fan whose cells are the inner normal cones $\{\mathcal N_{P}(F)\}$ as $F$ varies over the faces of $P$. Two polyhedra $P$ and $P'$ are [ *normally equivalent*]{} if they have the same normal fan. The [*recession cone*]{} $\rec(P)$ of a polyhedron $P$ is $\{ {\bf u}\in \Q^n : {\bf x} + {\bf u} \in P \text{ for all } {\bf x} \in P \}$. If $P=\{ {\bf x} \in \mathbb Q^n : A{\bf x} \geq {\bf b} \}$ for some $r \times n$ matrix $A$ and vector ${\bf b} \in \mathbb Q^r$, then $\rec(P)=\{ {\bf x} \in \mathbb Q^n : A{\bf x} \geq 0 \}$. The [*support*]{} of a polyhedral fan is the set of vectors lying in some cone of the fan. The fan $\mathcal{N}(P)$ is supported on the dual cone $\rec(P)^\vee = \{ {\bf y} \in (\Q^k)^* : {\bf y} \cdot {\bf x} \geq 0 \,\, \forall \,\, {\bf x} \in \rec(P) \}$ of the recession cone $\rec(P)$ of $P$. Indeed, ${\bf y}\not\in \rec(P)^\vee$ if and only if there exists ${\bf x}\in \rec(P)$ such that ${\bf y}\cdot {\bf x}< 0$, in which case ${\bf y}\cdot {\bf x}'$ is unbounded below for ${\bf x}':= \lambda {\bf x}\in \rec(P)$ with $\lambda \in \Q_{\geq 0}$. Let $\pi \colon \mathbb Q^n \rightarrow \mathbb Q^d$ be a linear map, and set $Q:=\pi(P)$. We may assume that $\pi$ is the projection onto the last $d$ coordinates. We denote by $\pi_{n-d}$ the projection onto the first $n-d$ coordinates $\Q^{n-d}$, and for $\mathbf{q} \in Q$ we set $P_\mathbf{q}:= \pi_{n-d}(\pi^{-1}(\mathbf{q}) \cap P)$. For ${\bf w}\in (\Q^{n-d})^*$, we write $\widetilde{\face}_{\bf{w}}(P_\mathbf{q})$ for the unique smallest face of $P$ containing the preimage of $\face_{\mathbf{w}}(P_{\mathbf{q}})$ under $\pi_{n-d}$. Fiber fans ---------- In this section we define the notion of a fiber fan $\mathcal N(P,Q)$ for a linear projection $\pi: P \rightarrow Q$ of polyhedra. This is the common refinement of the normal fans of the fibers $\{\pi^{-1}({\bf q}): {\bf q} \in Q\}$ of the map $\pi$. We first show that this definition makes sense. For $\mathbf{q}, \mathbf{q}' \in Q$, set $\mathbf{q} \sim \mathbf{q}$ if $P_\mathbf{q}$ is normally equivalent to $P_{\mathbf{q}'}$. Then there are only finitely many equivalence classes for $\sim$. Write $P= \{ {\bf x} : A{\bf x} \geq {\bf b} \}$, where $A$ is an $r \times n$ matrix for some $r$, and ${\bf b} \in \mathbb Q^r$. Then for $\mathbf{q} \in Q$, $\pi^{-1}(\mathbf{q}) \cap P = \{ ({\bf x}',\mathbf{q}) : {\bf x}' \in \mathbb Q^{n-d}, A({\bf x}',{\bf q}) \geq {\bf b} \}$. Write $A$ in block form as $(A'|A'')$, where $A'$ is an $r \times (n-d)$ matrix, and $A''$ is an $r \times d$ matrix. Then $P_{\mathbf{q}}= \{ {\bf x}' \in \mathbb Q^{n-d} : A' {\bf x}' \geq {\bf b}-A''\mathbf{q}\}$. Now for a given matrix $A'$ there are only finitely many normal equivalence classes of polyhedra of the form $\{{\bf x}'\in \mathbb Q^{n-d} : A'{\bf x}' \geq {\bf c} \}$ as $\mathbf{c}$ varies, so it follows that there are only finitely many normal equivalence classes of fibers of $\pi$. Pick $\mathbf{q}_1,\dots,\mathbf{q}_r \in Q$ such that the $P_{\mathbf{q}_i}$ are representatives of the different normal equivalence classes of fibers of $\pi$. Let $F$ be the Minkowski sum of the $P_{\mathbf{q}_i}$. The [*fiber fan*]{} $\mathcal N(P,Q)$ is the inner normal fan of $F$. The recession cone of $P_\mathbf{q}$ is the same for all $\mathbf{q} \in Q$. Write $P=\{{\bf x} \in \mathbb Q^n : A{\bf x} \geq {\bf b} \}$ for some $r \times n$ matrix $A$ and ${\bf b} \in \mathbb Q^r$. Decompose $A=(A'|A'')$, where $A'$ is an $r \times (n-d)$ matrix, and $A''$ is an $r \times d$ matrix. Translating the polyhedron $P$ by ${\bf a} \in \mathbb Q^n$ translates $Q$ by $\pi({\bf a})$, and $P_\mathbf{q} = (P+{\bf a})_{\mathbf{q}+\pi({\bf a})}$, so we may assume ${\bf 0}\in P$. Note that $P_{\bf 0} = \{ {\bf x} \in \mathbb Q^{n-d} : A' {\bf x} \geq {\bf b} \}$. We will show that $\rec(P_{\mathbf{q}}) = \rec(P_{\bf 0})$ for all ${\bf q}\in Q$. Let $\mathbf{v} \in \rec(P_{\mathbf{q}})$. Then ${\bf u}+\lambda \mathbf{v} \in P_\mathbf{q}$ for all ${\bf u} \in P_\mathbf{q}$ and $\lambda>0$. Since ${\bf 0} \in P$, we have $b_i\leq 0$ for all $i$, where $b_i$ is the $i$th component of ${\bf b}$. Suppose that $(\mathbf{v},{\bf 0}) \not \in P$ for ${\bf 0}\in \Q^d$. Then there is some $i$ with ${\bf a_i} \cdot (\mathbf{v},{\bf 0}) < b_i\leq 0$, where ${\bf a_i}$ is the $i$th row of $A$. But then ${\bf a_i} \cdot ({\bf u}+\lambda \mathbf{v}, \mathbf{q})={\bf a_i} \cdot ({\bf u},\mathbf{q}) + \lambda {\bf a_i} \cdot(\mathbf{v},0) < b_i$ for $\lambda$ sufficiently large, which means that $({\bf u}+\lambda \mathbf{v},\mathbf{q}) \not \in P$, so ${\bf u} +\lambda \mathbf{v} \not \in P_\mathbf{q}$. Therefore we have $\{ (\mathbf{v},{\bf 0}) \in \Q^n: \mathbf{v} \in \rec(P_{\mathbf{q}}) \} \subseteq P$ after all. The above argument shows that ${\bf a_i} \cdot (\mathbf{v},{\bf 0}) = {\bf a'_i} \cdot \mathbf{v} \geq 0$ for each row ${\bf a'_i}$ of $A'$, so $\mathbf{v}\in\rec(P_0)$. For the opposite inclusion, note that since $P_{\bf 0} = \{ {\bf x} \in \mathbb Q^{n-d} : A' {\bf x} \geq {\bf b} \}$, the set $\{ (\mathbf{v},0) : \mathbf{v} \in \rec(P_0) \}$ lies in $\rec(P)$. Thus if $\mathbf{v} \in \rec(P_0)$, then $({\bf u},\mathbf{q})+\lambda(\mathbf{v},{\bf 0}) \in P$ for all $\lambda >0$ and ${\bf u} \in P_\mathbf{q}$. This means $({\bf u}+\lambda \mathbf{v},\mathbf{q}) \in P$ for all $\lambda>0$ and ${\bf u} \in P_\mathbf{q}$, so ${\bf u}+\lambda \mathbf{v} \in P_\mathbf{q}$ for all $\lambda>0$ and ${\bf u} \in P_\mathbf{q}$, giving $\mathbf{v} \in \rec(P_{\mathbf{q}})$ as required. If $\mathcal F_1,\dots,\mathcal F_r$ are polyhedral fans with the same support, then their [*common refinement*]{} is the fan $\mathcal F$ whose cones are the intersections $\cap_{i=1}^r \sigma_i$, where $\sigma_i$ is a cone in $\mathcal F_i$. If $P$ and $P'$ are two polyhedra with the same recession cone, then the normal fan of the Minkowski sum of $P$ and $P'$ is the common refinement of the normal fans of $P$ and $P'$. Thus the fiber fan $\mathcal N(P,Q)$ is the common refinement of the normal fans of the fibers. In the case that $P$ is a polytope, the fiber fan $\mathcal N (P, Q)$ is the normal fan of the [*fiber polytope*]{} introduced by Billera and Sturmfels [@BilleraSturmfels]. While the fiber fan of a linear projection of polyhedra is still the normal fan of a polyhedron, there is not a canonical choice of a polyhedron with that normal fan if the polyhedron being projected is not a polytope. This motivates our definition of the fiber fan, instead of a more general fiber polyhedron. Note that for applications to toric varieties, a fan suffices to define the variety. Duality for polyhedral cones {#sec:coneduality} ---------------------------- This section establishes a duality result for polyhedral cones that is the key to Theorem \[t:maintwo\]. Versions of Lemmas \[lemma:faceequality\] and \[l:duallemma\] are known to the experts, though we know of no proofs in the literature. Let $C\subset \Q^n$ be a full-dimensional polyhedral cone. For $d < n$, let $\pi\colon \Q^n\rightarrow \Q^d$ be the projection onto the last $d$ coordinates. For $\mathbf{v}\in \pi(C)$, consider the polyhedral slice $C_{\mathbf{v}}\subseteq \mathbb Q^{n-d}$. As $\mathbb Q^{n-d}$ is canonically isomorphic to $\ker(\pi)$, we may regard the normal fan $\mathcal{N}(C_\mathbf{v})$ as lying in the dual vector space $\ker(\pi)^* \cong (\mathbb Q^{n-d})^*$. Let $\pi^\vee\colon (\Q^n)^*\rightarrow \ker(\pi)^*$ denote the map dual to the inclusion of $\ker(\pi)$ in $\Q^n$. In the basis of $(\Q^n)^*$ dual to the standard basis of $\Q^n$ this is projection on the first $n-d$ coordinates. For ${\textbf w}\in \pi^{\vee}(C^\vee)$ we consider the polyhedral slice $C^{\vee} \cap (\pi^{\vee})^{-1}(\textbf w)$, and write $C^\vee_{\bf w}:= \pi_d(C^\vee\cap (\pi^{\vee})^{-1}({\bf w})) \subseteq (\Q^d)^*$ for the isomorphic image of $C^{\vee} \cap (\pi^{\vee})^{-1}(\textbf w)$ under the projection $\pi_d\colon (\Q^n)^*\rightarrow (\Q^d)^*$ onto the last $d$ coordinates. There is a canonical isomorphism $\ker(\pi^\vee)\cong (\Q^d)^*$, so the normal fan $\mathcal{N}(C^\vee_{\textbf w})$ can be regarded as living in either $\Q^d$ or in $(\ker(\pi^{\vee}))^*$ for all ${\bf w}\in (\Q^{n-d})^*$. Let $\relint(F)$ denote the relative interior of a set $F$, which is the interior of $F$ in its affine span. \[lemma:faceequality\] Let $F\subset C$ be a face. If $\mathbf{v} \in \pi(\relint(F))$ then we have $\pi^{\vee}( \mathcal{N}_C(F)) = \mathcal{N}_{C_\mathbf{v}}(F_\mathbf{v})$. Write $C = \{{\bf x}\in \Q^n : A{\bf x}\geq {\bf 0}\}$ where $A$ is the $r\times n$ matrix whose rows ${\bf a}_1,\dots ,{\bf a}_r\in (\Q^n)^*$ form the facet normals of $C$. We may assume that $F$ is cut out by the first $p$ facet inequalities defining $C$, so $\mathcal{N}_{C}(F) = \Q_{\geq 0}\langle{\bf a}_1, \ldots, {\bf a}_p\rangle$ and $\pi^{\vee}(\mathcal{N}_{C}(F)) = \Q_{\geq 0}\langle \pi^{\vee}({\bf a}_1), \dots, \pi^{\vee}({\bf a}_p)\rangle$. Note that $$F\cap \pi^{-1}(\mathbf{v}) = \Big{\{} ({\bf u},\mathbf{v}) : \begin{array}{ll} {\bf a}_i \cdot ({\bf u} , \mathbf{v}) = 0 & \text{for } i = 1, \ldots, p; \\ {\bf a}_i \cdot ({\bf u}, \mathbf{v}) \geq 0 & \text{for } i = p+1, \ldots, r. \end{array}\Big{\}}$$ since $F$ is cut out by the first $p$ facet inequalities defining $C$. The assumption that $\mathbf{v} \in \pi(\relint(F))$ ensures that each $\geq$ above is a $>$, so $$F_{\mathbf{v}} = \big{\{}{\bf u} \in \Q^{n-d} : \pi^{\vee}({\bf a}_i) \cdot {\bf u} = c_i \text{ for }i = 1, \dots,p; \; \pi^{\vee}({\bf a}_i) \cdot {\bf u} > c_i \text{ for }i = p+1,\dots, r \big{\}}$$ where $c_i=-\mathbf{a}_i \cdot \mathbf{v}$ for $i=1, \ldots, r$. This implies that the rational cone $\mathcal{N}_{C_{\mathbf{v}}}(F_\mathbf{v})$ is also generated by $\{\pi^{\vee}({\bf a}_1), \dots, \pi^{\vee}({\bf a}_p)\}$ as claimed. \[l:duallemma\] Fix $\mathbf{v}\in \pi(C)$ and ${\bf w}\in \pi^{\vee}(C^\vee)$. If $F\subset C$ is a face then $$F = \widetilde{\face}_{\bf w}(C_{\mathbf{v}})\iff \mathbf{v} \in\pi(\relint(F))\text{ and }{\bf w} \in\pi^{\vee}\big{(}\relint(\mathcal{N}_{C}(F))\big{)}.$$ Suppose $F = \widetilde{\face}_{\bf w}(C_{\mathbf{v}})$. Then the preimage of $\face_{\bf w}(C_{\mathbf{v}})$ under $\pi_{n-d}$ lies in no proper face of $F$, so the intersection of $\relint(F)$ with this preimage is nonempty. This implies the set $\pi^{-1}(\mathbf{v}) \cap \relint(F)$ is nonempty, so $\mathbf{v} \in \pi(\relint(F))$. We have ${\bf w}\in \relint(\mathcal{N}_{C_\mathbf{v}}(\face_{\bf w}(C_\mathbf{v})))$ by definition. Applying Lemma \[lemma:faceequality\] to the face $F_\mathbf{v} = \pi_n(F\cap \pi^{-1}(\mathbf{v})) = \face_{\bf w}(C_\mathbf{v})$ gives ${\bf w}\in \relint\pi^{\vee}(\mathcal{N}_C(F)) = \pi^{\vee}(\relint\mathcal{N}_C(F))$. Conversely, suppose $\mathbf{v} \in \pi(\relint(F))$ and ${\bf w} \in\pi^{\vee}(\relint(\mathcal{N}_{C}(F)))$. The first assumption ensures that the face $F_\mathbf{v}\subseteq C_\mathbf{v}$ is nonempty and its preimage under $\pi_{n-d}$ lies in no proper subface of $F$. Also, ${\bf w} \in\pi^{\vee}(\relint(\mathcal{N}_{C}(F))) = \relint\pi^{\vee}(\mathcal{N}_C(F)) = \relint \mathcal{N}_{C_\mathbf{v}}(F_\mathbf{v})$ by Lemma \[lemma:faceequality\], hence $F_\mathbf{v} = \face_{\bf w}(C_{\mathbf{v}})$. Since the preimage of $F_\mathbf{v}$ under $\pi_{n-d}$ lies in no proper subface of $F$, we have $F = \widetilde{\face}_{\bf w}(C_{\mathbf{v}})$. We now present the main result of this section. See Figure \[theorem2.3\] for an illustration. \[p:fiberduality\] For $\mathbf{v}\in \pi(C)$ and ${\bf w}\in \pi^{\vee}(C^\vee)$ we have $$\label{eqn:fiberduality} \mathcal{N}_C\big{(}\widetilde{\face}_{\bf w}(C_{\mathbf{v}})\big{)} = \widetilde{\face}_{\mathbf{v}}(C^\vee_{\bf w}).$$ Applying Lemma \[l:duallemma\] to the face $F:=\widetilde{\face}_{\bf w}(C_{\mathbf{v}})$ of $C$ gives $\mathbf{v} \in \pi(\relint (F))$ and ${\bf w}\in \pi^{\vee}(\relint(\mathcal{N}_{C}(F)))$. The face $\mathcal{N}_C(F)$ of $C^{\vee}$ satisfies $\mathcal{N}_{C^{\vee}}(\mathcal{N}_C(F)) = F$, so ${\bf w} \in \pi^{\vee}(\relint(\mathcal{N}_C(F)))$ and $\mathbf{v} \in \pi(\relint(\mathcal{N}_{C^{\vee}}(\mathcal{N}_C(F))))$. Now apply Lemma \[l:duallemma\] to the face $\mathcal{N}_C(F)$ of $C^{\vee}$ to deduce that $\mathcal{N}_C(F) = \widetilde{\face}_{\mathbf{v}}(C^\vee_{\bf w})$ as required. Proposition \[p:fiberduality\] is the key duality result that is required in the proof of Theorem 6.2 of Craw–Maclagan–Thomas [@CMT1]. The following definition is a variant for polyhedral cones of a notion for polytopes due to Billera–Sturmfels [@BilleraSturmfels § 2]. For $\mathbf{v} \in \pi(C)$, the subdivision of ${\pi^\vee}(C^\vee)$ consisting of the cells $\big\{ \pi^\vee\big( \widetilde{\face}_{\mathbf{v}} (C^\vee_{\bf w})\big{)} : {\bf w} \in {\pi^\vee}(C)\big\}$ is the ${\mathbf{v}}$*-induced ${\pi^\vee}$-coherent subdivision* of ${\pi^\vee}(C^\vee)$. The duality result of Proposition \[p:fiberduality\] leads immediately to the following description of the normal fan of the polyhedral slice $C_\mathbf{v}$ for $\mathbf{v}\in \pi(C)$. \[cor:normalfan\] For $\mathbf{v}\in \pi(C)$, the inner normal fan of the polyhedron $C_{\mathbf{v}}$ is the $\mathbf{v}$-induced $\pi^{\vee}$-coherent subdivision of $\pi^{\vee}(C^\vee)$. In other words, $$\mathcal{N}(C_\mathbf{v}) = \big{\{}\pi^{\vee}\big{(}\widetilde{\face}_{\mathbf{v}}(C^\vee_{\bf w})\big{)} : {\bf w} \in \pi^{\vee}(C^\vee) \big{\}}.$$ In particular, for $\mathbf{v} = \emph{\textbf{0}}$ we have $\mathcal{N}(C_{\emph{\textbf{0}}}) = \pi^{\vee}(C^\vee)$. By Lemma \[lemma:faceequality\] and Proposition \[p:fiberduality\], we have $$\mathcal{N}_{C_{\mathbf{v}}}\big{(}\face_{\bf w}(C_{\mathbf{v}})\big{)} = \pi^{\vee}\big{(}\mathcal{N}_{C}\big{(}\widetilde{\face}_{\bf w}(C_{\mathbf{v}})\big{)}\big{)} = \pi^{\vee}\big{(}\widetilde{\face}_{\mathbf{v}}(C^\vee_{\bf w})\big{)}.$$ The set of cones $\mathcal{N}_{C_{\mathbf{v}}}(\face_{\bf w}(C_{\mathbf{v}}))$ taken over all vectors ${\bf w} \in \pi^{\vee}(C^\vee)$ equals the inner normal fan $\mathcal{N}(C_\mathbf{v})$ as required. For the second statement, observe that $\face_{\textbf{0}}(C^\vee_{\bf w}) = C^\vee_{\bf w}$ for all ${\bf w} \in \pi^{\vee}(C^\vee)$, and the smallest face of the cone $C^\vee$ containing the set $C^\vee_{\bf w}$ is $C^\vee$ itself. \[d:subdivisioninduced\] Given a linear map $\pi\colon \Q^n\rightarrow \Q^d$ and a fan $\Sigma\subseteq \Q^n$, the *fan induced by $\pi(\Sigma)$* is the fan obtained by taking the common refinement of the cones $\{\pi(\sigma) : \sigma\in \Sigma\}$. If $\Sigma$ consists of the faces of a cone $P\subseteq \Q^n$, we simply write $\pi(P)$. \[cor:coneduality\] The fiber fan $\mathcal N(C,\pi(C))$ is the fan induced by $\pi^{\vee}(C^{\vee})$. The fan $\mathcal N(C, \pi(C))$ is the common refinement of the normal fans of the fibers $C_{\mathbf{v}}$ for $\mathbf{v} \in \pi(C)$. As $\mathbf{v}$ varies in $\pi(C)$, $\widetilde{\face}_{\mathbf{w}}(C^{\vee}_{\mathbf{v}})$ varies over all faces of $C^{\vee}$. Thus $\mathcal N(C,\pi(C))$ is the common refinement of all $\pi^{\vee}(F)$ for $F$ a face of $C^{\vee}$. Duality for general polyhedra {#sec:GeneralDuality} ----------------------------- We now generalize the results of the previous subsection from polyhedral cones to general polyhedra. \[d:tildeP\] For any polyhedron $P \subseteq \mathbb Q^n$, let $\widetilde{P} \subseteq \Q^n\oplus \Q$ be the polyhedral cone obtained as the closure in $\Q^{n}\oplus \Q$ of the set $\{ (\lambda {\bf x},\lambda) : {\bf x} \in P, \lambda \in \Q_{>0} \}$. Note the $\widetilde{P} \cap (\Q^n \oplus \{0\}) \cong \rec(P)$. Given $\pi\colon \Q^n\rightarrow \Q^d$, we write $\widetilde{\pi}\colon \Q^n\oplus \Q\rightarrow \Q^d\oplus \Q$ for the map sending $({\bf x},\lambda)$ to $(\pi({\bf x}),\lambda)$. The first projection $p_1\colon \Q^{n}\oplus \Q\rightarrow \Q^n$ fits into the diagram $$\label{eqn:liftdiagram} \begin{CD} 0 @>>> \ker(\widetilde{\pi}) @>\widetilde{\iota}>> \Q^{n}\oplus \Q @>\widetilde{\pi}>> \Q^{d}\oplus \Q @>>> 0 \\ @. @V\cong VV @Vp_1VV @VVV @. \\ 0 @>>> \ker(\pi) @>\iota>> \Q^{n} @>\pi>> \Q^d @>>> 0 \\ \end{CD},$$ and canonically identifies $\ker(\pi)$ with $\ker(\widetilde{\pi})$. Write $\pi^\vee \colon (\Q^n)^*\rightarrow \ker(\pi)^*$ and $\widetilde{\pi}^\vee \colon (\Q^{n+1})^*\rightarrow \ker(\widetilde{\pi})^*$ for the maps obtained by pullback via $\iota$ and $\widetilde{\iota}$ respectively. These maps fit into the dual diagram $$\label{eqn:dualliftdiagram} \begin{CD} 0 @<<< \ker(\widetilde{\pi})^* @<\widetilde{\pi}^\vee<< (\Q^n\oplus\Q)^* @<<< (\Q^{d}\oplus \Q)^* @<<< 0 \\ @. @Ap_1^* A\cong A @Ap_1^* AA @AAA @. \\ 0 @<<< \ker(\pi)^* @<\pi^\vee<< (\Q^{n})^* @<<< (\Q^d)^* @<<< 0 \\ \end{CD}.$$ Since $P\subseteq \Q^n$ and $\widetilde{P}\subseteq \Q^n\oplus\Q$, the normal fans satisfy $\mathcal{N}(P)\subseteq (\Q^n)^*$ and $\mathcal{N}\big{(}\widetilde{P}\big{)}\subseteq (\Q^n\oplus\Q)^*$ respectively. After identifying $\ker(\pi)^*$ with $\ker(\widetilde{\pi})^*$ via the vertical map $p_1^*$ from diagram (\[eqn:dualliftdiagram\]), the following fans coincide: 1. the fiber fan $\mathcal{N}(P,\pi(P))$; 2. the fiber fan $\mathcal{N}(\widetilde{P},\widetilde{\pi}(\widetilde{P}))$; 3. the fan induced by $\widetilde{\pi}^\vee\big{(}\mathcal{N}\big{(}\widetilde{P})\big{)}$; 4. the fan induced by $\pi^\vee(\mathcal{N}(P))$. We first show that fans (1) and (2) coincide. For ${\bf q} \in \mathbb Q^{d}$, set $\widetilde{\bf q}:= ({\bf q},1)\in \Q^d\oplus \Q$. Note that $(\widetilde{P})_{\widetilde{\bf q}} = P_{\bf q}$, since $P$ is the intersection of $\widetilde{P}$ with $\Q^d\oplus \{1\}$. Since $\widetilde{P}$ is a cone, the polyhedron $(\widetilde{P})_{({\bf q},j)}$ is normally equivalent to $(\widetilde{P})_{({\bf q}/j,1)}$ for $j>0$. It follows that the fan $\mathcal N (P,\pi(P))$ is equal to the common refinement of (the images under $p_1^*$ of) the normal fans of fibers $(\widetilde{P})_{({\bf q},j)}$ for $j>0$ and ${\bf q} \in \mathbb Q^d$. Since $(\widetilde{P})_{({\bf q},j)} = \emptyset$ for $j<0$, the statement that the fiber fans (1) and (2) coincide follows once we show that $\mathcal N (P , \pi(P))$ is a refinement of the normal fan of $(\widetilde{P})_{({\bf q},0)}$ for all $({\bf q},0) \in \widetilde{\pi}(\widetilde{P})$. Since $(\widetilde{P})_{({\bf q},0)}=\rec(P)_{\bf q}$, we have $({\bf q},0) \in \widetilde{\pi}(\widetilde{P})$ if and only if ${\bf q} \in \pi(\rec(P))$. Thus, it is enough to show that $\mathcal N (P , \pi(P))$ refines $\mathcal N(\rec(P)_{\bf q})$ for all ${\bf q} \in \pi(\rec(P))$. We show below that for all ${\bf q} \in \pi(\rec(P))$ there is $\lambda > 0$ such that $\lambda{\bf q} \in \pi(P)$, and $\mathcal N(P_{\lambda{\bf q}})$ refines $\mathcal N(\rec(P)_{\lambda{\bf q}})$. Since these normal fans equal the corresponding ones with $\lambda{\bf q}$ replaced by ${\bf q}$, this shows that $\mathcal N(P_{{\bf q}})$ and hence $\mathcal N(P, \pi(P))$ is a refinement of the normal fan of $\rec(P)_{\bf q}$. Fix ${\bf q} \in \pi(\rec(P))$. The condition that $\lambda \mathbf{q} \in \pi(P)$ is satisfied for all $\lambda \gg 0$, or all $\lambda $ if $\mathbf{0} \in P$. Write $P=\{ \mathbf{x} \in \mathbb Q^n : A\mathbf{x} \geq \mathbf{b} \}$, where $A$ is an $r \times n$ matrix, and $\mathbf{b} \in \mathbb Q^r$. Write $A$ in block form as $A=(A'|A'')$, where $A'$ is a $r \times (n-d)$ matrix, and $A''$ is a $r \times d$ matrix. Then $P_{\lambda \mathbf{q}}=\{ \mathbf{x} \in \mathbb Q^{n-d} : A'\mathbf{x} \geq \mathbf{b} - A''\lambda \mathbf{q}\}$, while $\rec(P)_{\lambda \mathbf{q}}=\{ \mathbf{x} \in \mathbb Q^{n-d} : A'\mathbf{x} \geq - A''\lambda \mathbf{q}\}$. We first consider the case that $A''\mathbf{q} \neq \mathbf{0}$. Since $\lambda \mathbf{q} \in \pi(P)$, $P_{\lambda \mathbf{q}}$ and $\rec(P)_{\lambda \mathbf{q}}$ are nonempty for all $\lambda$, so $\mathbf{b}-A''\lambda \mathbf{q}$ and $-A''\lambda \mathbf{q}$ lie in some chambers of the chamber complex of the Gale dual of $A'$ (see [@BFS]). For most generic choices of $\mathbf{q}$ the vector $-A''\mathbf{q}$ lies in the interior of a top-dimensional chamber, so for $\lambda \gg 0$ the vectors $-A''\mathbf{q}$ and $\mathbf{b}-A''\lambda \mathbf{q}$ lie in the same chamber, and thus $P_{\lambda \mathbf{q}}$ and $\rec(P)_{\lambda \mathbf{q}}$ are normally equivalent. If $\mathbf{q}$ is not generic, then for $\lambda \gg 0$ the vector $\mathbf{b}-A''\lambda \mathbf{q}$ either lies in the same chamber as $-A''\lambda \mathbf{q}$, or in a larger dimensional chamber that has the chamber of $-A''\lambda \mathbf{q}$ as a face. In the first case $P_{\lambda \mathbf{q}}$ and $\rec(P)_{\lambda \mathbf{q}}$ are normally equivalent, while in the second the normal fan of $P_{\lambda \mathbf{q}}$ is a refinement of that of $\rec(P)_{\lambda \mathbf{q}}$. Finally, if $A''\mathbf{q} = \mathbf{0}$, then $\rec(P)_{\lambda \mathbf{q}}=\rec(P_{\lambda \mathbf{q}})$ is the recession cone of $P_{\lambda \mathbf{q}}$, so the refinement is automatic in this case. To see that fan (2) equals fan (3), apply Corollary \[cor:coneduality\] with $C=\widetilde{P}$ and $\widetilde{\pi}$ in place of $\pi$. Thus, it remains to show that fans (3) and (4) coincide. The proof is a simple trick. The normal fan $\mathcal{N}(P)$ equals the fiber fan $\mathcal{N}(P,g(P))$ of the trivial linear map $g\colon \Q^n\rightarrow \Q^0$, for which the lift $\widetilde{g}$ is the second projection $p_2\colon \Q^n\oplus \Q\rightarrow \Q$. The equality of the fans (1) and (2) established earlier in this proof holds for any surjective linear map, and in particular it holds the trivial map $g$. Thus, $p_1^*$ identifies $\mathcal{N}(P) = \mathcal{N}(P,g(P))$ with $\mathcal{N}(\widetilde{P},p_2(\widetilde{P}))$. The equality of fans (2) and (3) then implies that $p_1^*$ identifies $\mathcal{N}(P)$ with $p_2^\vee(\mathcal{N}(\widetilde{P}))$. Finally, apply $\pi^\vee$ to see that $p_1^*$ identifies $\pi^\vee(\mathcal{N}(P))$ with $(\pi^\vee \circ p_2^\vee)\big{(}\mathcal{N}(\widetilde{P})\big{)}$. If the upward-pointing arrow in diagram (\[eqn:dualliftdiagram\]) representing $p_1^*\colon (\Q^n)^*\rightarrow (\Q^n\oplus \Q)^*$ is replaced by the downward-pointing arrow representing $p_2^\vee\colon (\Q^n\oplus \Q)^*\rightarrow (\Q^n)^*$, the resulting diagram commutes giving $\widetilde{\pi}^\vee = p_1^*\circ \pi^\vee \circ p_2^\vee$. This proves that $p_1^*$ identifies fans (3) and (4) as required. \[l:heightone\] Let $\pi\colon P\rightarrow Q$ be a linear projection of polyhedra. Then $\widetilde{\pi}(\widetilde{P}) = \widetilde{Q}$, so $\widetilde{\pi}(\widetilde{P})$ is the cone over $Q$. Moreover, the fan induced by $\widetilde{\pi}(\widetilde{P})$ is the cone over the subdivision of $Q$ induced by $\pi(P)$, so it is obtained by taking the cone over each cell in that subdivision. It is easy to see that $\pi(\rec(P)) = \rec(\pi(P)) = \rec(Q)$, so the equality of sets $\widetilde{\pi}\big{(}\widetilde{P}\big{)} = \widetilde{Q}$ follows from the fact that $\widetilde{\pi}(\lambda {\bf x},\lambda) = (\lambda \pi({\bf x}),\lambda)$ for ${\bf x}\in P$. For the second statement, note first that the faces of $\widetilde{P}$ divide naturally into two disjoint sets: those arising from faces of the recession cone of $P$; and those of the form $\widetilde{E}$ for some face $E\subseteq P$. The faces in the former set lie in $\Q^n\oplus \{0\}\subset \Q^n\oplus \Q$, so it is enough to show that the given decomposition of $Q$ is the intersection of the slice $\Q^d\oplus \{1\}\subset \Q^d\oplus \Q$ with the fan obtained as the common refinement of the cones $\big{\{}\widetilde{\pi}(\widetilde{E})\subseteq \widetilde{Q} : E\subseteq P \text{ is a face}\big{\}}$. The result follows since for every face $E\subseteq P$, the intersection of $\Q^d\oplus \{1\}$ with the cone $\widetilde{\pi}(\widetilde{E})\subseteq \widetilde{Q}$ is the cell $\pi(E)\subseteq Q$. \[p:PveeFiberfan\] Let $\pi\colon P\rightarrow Q$ be a linear projection of polyhedra, and set $C:= (\widetilde{P})^{\vee}$. Then the fiber fan $\mathcal{N}( C, \widetilde{\pi}^{\vee}(C))$ is the fan induced by $\widetilde{\pi}(\widetilde{P})$, and thus is the cone over the subdivision of $Q$ induced by $\pi(P)$. Applying Corollary \[cor:coneduality\] to the map $\widetilde{\pi}^\vee\colon C\rightarrow \widetilde{\pi}^\vee(C)$ gives $\mathcal{N}( C, \widetilde{\pi}^{\vee}(C)) = (\widetilde{\pi}^\vee)^\vee(C^\vee)$. Since $\widetilde{P}$ is a polyhedral cone, we have $C^\vee = ((\widetilde{P})^\vee)^\vee = \widetilde{P}$. The first statement follows from the fact that $(\widetilde{\pi}^\vee)^\vee = \widetilde{\pi}$. The second statement follows from Lemma \[l:heightone\]. On quotients of toric varieties by subtori {#s:toricquotients} ========================================== GIT quotients of quasiprojective toric varieties {#sec:toricGIT} ------------------------------------------------ We first recall the polyhedral construction of quasiprojective toric varieties that are projective over affine toric varieties (see Thaddeus [@Thaddeus §2.8]). We work over an algebraically closed field $\K$. Let $P\subseteq \Q^n$ be an $n$-dimensional lattice polyhedron and let $\widetilde{P}$ be the closure in $\Q^{n}\oplus \Q$ of the cone over $P$ as given in Definition \[d:tildeP\]. The semigroup algebra $R:= \K[\widetilde{P}\cap (\Z^{n}\oplus \Z)]$ defines the affine toric variety $\Spec R$. The second projection determines a grading of the ring $R$. Let $R_j$ be the $\K$-vector subspace spanned by the elements of $R$ of degree $j\in \N$. Then $X_P := \Proj \bigoplus_{j\geq 0} R_j$ is the $n$-dimensional quasiprojective toric variety whose fan is $\mathcal{N}(P)$, the inner normal fan of the polyhedron $P$. Note that $X_{\widetilde{P}}\cong \Spec R$ is the affine cone over $X_P$. Also, since $R_{0}$ is isomorphic to the semigroup algebra $\K[\rec(P)\cap \Z^n]$ determined by the recession cone $\rec(P)\subseteq \Q^n$, we see that $X_P$ is projective over the affine toric variety $\Spec R_0$. Thus, $X_P$ is projective if $P$ is a polytope. Any quasiprojective toric variety $X$ that is projective over an affine toric variety arises in this way once a relatively ample torus-invariant divisor $D$ is chosen on $X$, in which case $R_j \cong H^0(X,\owe_X(jD))$ for all $j\in \N$. Let $(\K^*)^n = \Spec \K[\Z^n]$ denote the dense torus of $X_P$, and let $T_L:= \Spec \K[L]$ be a $d$-dimensional algebraic subtorus of $(\K^*)^n$. Choose generators for $L \cong \Z^d$ and write $\pi_{\Z}\colon \Z^n \rightarrow \Z^d$ for the map of character lattices induced by the inclusion $T_L\hookrightarrow (\K^*)^n$. Composing $\pi_{\Z}$ with the $\Z^n$-grading of the ring R arising from the action of $(\K^*)^n$ on $X_P$ determines a $\Z^d$-grading of $R$. Given a character ${\mathbf{v}}\in \Z^d$ of the torus $T_L$, the GIT quotient of $X_P$ by the action of $T_L$ linearized by $\mathbf{v}$ is the quotient of the affine variety $\Spec(R)$ by the lift of $T_L$. Specifically, begin by defining $\widetilde{\pi_{\Z}} \colon \Z^{n}\oplus \Z\rightarrow \Z^{d}\oplus \Z$ by sending $(\mathbf{x},\lambda)$ to $(\pi_{\Z}(\mathbf{x}),\lambda)$. This lattice map induces an inclusion of the $(d+1)$-dimensional torus $T_{\widetilde{L}} :=\Spec \K[(\widetilde{L}]$ into the dense torus $(\K^*)^{n+1} = \Spec \K[\Z^n\oplus\Z]$ of $X_{\widetilde{P}}$, where $\widetilde{L}= L \oplus \Z$. Characters ${\bf v}\in \Z^d$ of $T_L$ give rise to characters $\widetilde{{\bf v}}:= ({\bf v},1)\in \Z^{d}\oplus\Z$ of $T_{\widetilde{L}}$. Let $R_{j\widetilde{{\bf v}}}$ denote the $j\widetilde{{\bf v}}$-graded piece of the ring $R$ with respect to the $\Z^{d}\oplus\Z$-grading of $R$. Then $$\label{eqn:GITdef} X_P \git_{\mathbf{v}} T_L= X_{\widetilde{P}} \git_{\widetilde{\bf v}}T_{\widetilde{L}}:= \textstyle{\Proj \bigoplus_{j\geq 0} R_{j{\widetilde{\bf v}}}}.$$ We now adopt the notation of Section \[sec:GeneralDuality\]. Thus, we write $\pi\colon \Q^n\rightarrow \Q^d$ and $\widetilde{\pi}\colon \Q^{n}\oplus\Q\rightarrow \Q^{d}\oplus\Q$ for the linear maps obtained from $\pi_{\Z}$ and $\widetilde{\pi_\Z}$ respectively by extending scalars. The following result is due to Kapranov–Sturmfels–Zelevinsky [@KSZ] when $P$ is a polytope, and to Thaddeus [@Thaddeus] in general. \[lemma:XgitT\] For ${\bf v}\in \pi(P)$, the GIT quotient $X_P\git_{\bf v}T_L$ is the toric variety with fan $\mathcal{N}(P_{\bf v})$. We know that $X_P\git_{\bf v}T\cong X_{\widetilde{P}}\git_{\widetilde{\bf v}}T_{\widetilde{L}}$ from above. The left-most vertical arrow in the diagram (\[eqn:liftdiagram\]) sends the polyhedron $\widetilde{P}_{\widetilde{\bf v}}$ isomorphically onto $P_{\bf v}$, so the left-most vertical map in the dual diagram (\[eqn:dualliftdiagram\]) identifies the normal fans $\mathcal{N}(\widetilde{P}_{\widetilde{\bf v}})$ and $\mathcal{N}(P_{\bf v})$. The result follows from the fact that for $j\in \N$, the graded piece $R_{j{\widetilde{\bf v}}}$ is isomorphic to the $\K$-vector space spanned by the lattice points of the slice $\widetilde{P} \cap \widetilde{\pi}^{-1}(j\widetilde{{\bf v}})$. For ${\bf v}\in \pi(P)$, the fan $\mathcal{N}(P_{\bf v})$ of $X_{P}\git_{\bf v}T$ is isomorphic to the $\widetilde{{\bf v}}$-induced $\widetilde{\pi}^{\vee}$-coherent subdivision of the cone $\widetilde{\pi}^\vee(\widetilde{P}^{\vee})\subseteq \ker(\widetilde{\pi})^*$. Corollary \[cor:normalfan\] shows that $\mathcal{N}(\widetilde{P}_{\widetilde{{\bf v}}})$ is the $\widetilde{{\bf v}}$-induced $\widetilde{\pi}^\vee$-coherent subdivision of the cone $\widetilde{\pi}^\vee(\widetilde{P}^{\vee})$. The result follows since $\mathcal{N}(P_{\bf v})$ is isomorphic to $\mathcal{N}(\widetilde{P}_{\widetilde{\bf v}})$. \[d:GITchamber\] A character ${\bf v}\in \pi(P)$ is *generic* if every ${\bf v}$-semistable point of $X_P$ is ${\bf v}$-stable in the sense of GIT (see Dolgachev–Hu [@DolgachevHu (0.2.2)]). For generic characters ${\bf v}, {\bf v}'\in \pi(P)$, we set $\mathbf{v} \sim \mathbf{v}'$ if every ${\bf v}$-stable point of $X_P$ is ${\bf v}'$-stable and vice-versa. This equivalence relation gives a polyhedral decomposition of $\pi(P)$, called the [*GIT chamber decomposition*]{} associated to the action of $T_L$ on $X_P$. \[r:GITchamber\] Thaddeus [@Thaddeus] showed that the GIT chamber decomposition is the subdivision of $Q$ induced by $\pi(P)$ in the sense of Definition \[d:subdivisioninduced\]. For analogous results for general GIT quotients, see Thaddeus [@ThaddeusGeneral] or Dolgachev–Hu [@DolgachevHu]. The toric Chow quotient {#sec:toricChow} ----------------------- In this section we construct the toric Chow quotient of a quasiprojective toric variety by a subtorus. We first introduce a subcategory of the category of $\K$-schemes. Fix $T := (\K^*)^n$ for $n\in \N$. Consider the category $\mathcal{C}_T$ whose objects are Noetherian $\K$-schemes $X$ with a $T$-action $\sigma_X \colon T\times X\rightarrow X$ and an irreducible $T$-invariant component $X_0\subseteq X$, such that the restriction of the $T$-action to $X_0$, denoted $\sigma_{X_0} \colon T\times X_0\rightarrow X_0$, gives $X_0$ the structure of a not-necessarily normal toric variety with dense torus $T$. The morphisms of $\mathcal{C}_T$ are $T$-equivariant proper morphisms $f\colon X\rightarrow X'$ over $\K$ that induce a birational morphism $f\vert_{X_0}\colon X_0\rightarrow X'_0$ between the toric components. \[l:fiberprodsexist\] Fiber products exist in the category $\mathcal{C}_T$. Moreover, the fiber product in $\mathcal{C}_T$ coincides with the fiber product in the category of $\K$-schemes. For an object $S$ in $\mathcal{C}_T$, let $f\colon X\rightarrow S$ and $f'\colon X'\rightarrow S$ be morphisms in $\mathcal{C}_T$. We must show that the fiber product in the category of $\K$-schemes $Z:=X\times_S X'$ is an object of $\mathcal{C}_T$, and that the canonical projections $p\colon Z\rightarrow X$ and $q\colon Z\rightarrow X'$ are morphisms in $\mathcal{C}_T$. To construct the component $Z_0\subseteq Z$, note that the schemes $X$, $X'$ and $S$ contain open subvarieties $T_X$, $T_{X'}$ and $T_S$ respectively, each isomorphic to the torus $T$. The fiber product $T_Z:= T_X\times_{T_S} T_{X'}$ is canonically isomorphic to the subscheme $p^{-1}(T_X)\cap q^{-1}(T_{X'})$ of $Z$ by [@EGA Chap. I, Coro 3.2.3], so it is open in $Z$. Furthermore, the $T$-equivariant birational morphisms $f\vert_{X_0}\colon X_0\rightarrow S_0$ and $f'\vert_{X'_0}\colon X'_0\rightarrow S_0$ restrict to give isomorphisms $T_X\cong T_S$ and $T_{X'}\cong T_S$, so $T_Z$ is isomorphic to $T\times_T T\cong T$ and hence is irreducible. In particular, the $n$-torus $T_Z\subseteq Z$ is dense in some component of $Z$ that we denote $Z_0$. We next show that $Z_0$ is reduced. Since this is a local question, and $T_Z \subseteq Z_0$ is reduced, we can reduce to the case where $Z_0=\Spec(A)$ is an irreducible Noetherian $\K$-scheme with a nonempty open subscheme $W$ for which the induced scheme structure on $W$ is reduced. We may assume that $W=\Spec(A_f)$ for some $f \in A$. Writing $A=\K[x_1,\dots,x_m]/I$ for some $I$, we note that $I$ is primary since $Z_0$ is irreducible, and since $W$ is nonempty $f \not \in \sqrt I$. Suppose $Z_0$ is not reduced, so there exists $g \not \in I$ with $g^k \in I$ for some $I$. Then $(g/1)^k \in I_f$, so since $W$ is reduced we have $g/1 \in I_f$. But then $g/1=y/f^l$ for some $l$ and some $y\in I$. This means that $gf^l \in I$, which contradicts $I$ being primary, $g \not \in I$, and $f \not \in \sqrt I$. We thus conclude that $Z_0$ is reduced. We now describe the torus action. The fiber product $Z$ embeds as a closed subscheme of the product $X\times_{\K} X'$. Since the morphisms $f$ and $f'$ are $T$-equivariant, it can be shown that the product $T$-action on $X\times_{\K} X'$ restricts to give an action $\sigma_Z\colon T\times Z\rightarrow Z$. This restricts to give an action $T\times T_Z\rightarrow T_Z$ that coincides with the multiplicative structure on the torus $T_Z$ after identifying $T_Z\cong T$. Thus, $\sigma_Z\colon T\times Z\rightarrow Z$ extends the natural multiplicative structure of $T_Z$. To see that $Z_0$ is $T$-invariant we argue by continuity as follows. The component $Z_0\subseteq Z$ is closed and contains $T_Z$, so $\sigma_Z^{-1}(Z)$ is a closed subscheme of $T\times Z$ that contains $\sigma_Z^{-1}(T_Z) = T\times T_Z$. In particular, $\sigma_Z^{-1}(Z)$ contains the closure $T\times Z_0$ of $T\times T_Z$, so the image of $T\times Z_0$ under $\sigma_Z$ lies in $Z_0$. The resulting action $\sigma_{Z_0}\colon T\times Z_0\rightarrow Z_0$ extends the multiplicative structure of $T_Z$ on itself by the above, so $Z_0$ is a not-necessarily-normal toric variety. It remain to prove that the projections $p$ and $q$ are morphisms in $\mathcal{C}_T$. Properness of $p$ and $q$ follows from properness of $f$ and $f'$ by base extension. The $T$-action on $Z$ was constructed to ensure that $p$ and $q$ are $T$-equivariant. Their restrictions $p\vert_{T_Z} \colon T_Z \rightarrow T_X$ and $p'\vert_{T_Z} \colon T_Z \rightarrow T_{X'}$ are isomorphisms, so both $p$ and $q$ are birational on the toric components as required. We now return to the GIT set-up from Section \[sec:toricGIT\], where $\pi\colon P\rightarrow Q$ is a linear projection of polyhedra, with $L \cong \mathbb Z^d$ the image of the corresponding map $\pi_{\mathbb Z} \colon \mathbb Z^n \rightarrow \mathbb Z^d$. Set $M:=\ker(\pi_{\Z})$. For ${\bf v}\in Q$, the GIT quotient $X_P\git_{\bf v}T_L = X_{P_{\bf v}}$ is the toric variety with dense torus $T_M$ defined by the fan $\mathcal{N}(P_{\bf v})$. Let $\poset$ be the face poset of the subdivision of $Q$ induced by $\pi(P)$, with the faces on the boundary of $Q$ removed. Thus $\tau \prec \sigma$ if $\tau$ is a face of $\sigma$. To each $\sigma \in \poset$ we associate the toric variety $X_{\sigma}:=X_{P_{\bf v}}$ for any $\mathbf{v} \in \sigma$. If $\tau$ is a face of $\sigma$, then there is a proper, birational toric morphism from $X_{\sigma}$ to $X_{\tau}$ by Thaddeus [@Thaddeus Theorem 3.11 and Corollary 3.12]. Therefore $\poset$ gives a directed system of $\K$-schemes. We are interested in the inverse limit of this system in the category of $\K$-schemes. \[prop:invlimit\] Let $T_M$ be the dense torus in $X_{\sigma}$ for $\sigma \in \poset$. Then the inverse limit in the category of $\K$-schemes $$\label{eqn:invlimit} Z:= \varprojlim_{\sigma\in \poset} X_{\sigma}$$ exists and is an object of the category $\mathcal{C}_{T_M}$. In particular, $Z$ has an irreducible component $Z_0$ that is a not-necessarily-normal toric variety with dense torus $T_M$. For each $\sigma \in \poset$, the toric variety $X_{\sigma}$ is an object of the category $\mathcal{C}_{T_M}$. The proof is by induction on the number of elements of $\poset$. The base case is when $\poset$ has only one maximal element $\sigma$. Then $Z=X_{\sigma}$ exists and is an object of $\mathcal{C}_{T_M}$. Suppose now that the inverse limit exists, and is an object in $\mathcal{C}_{T_M}$, for any such poset with fewer elements than $\poset$ whose elements are toric varieties with dense torus $T_M$, and whose maps are proper birational toric morphisms. Fix a maximal element $\sigma$ of $\poset$, and let $\poset'=\poset \setminus \sigma$, and let $\poset''$ be the subposet of all $\tau \in \poset$ with $\tau \prec \sigma$. Let $Z_{\poset'}$ and $Z_{\poset''}$ be the corresponding inverse limits. Define $$Z':=Z_{\poset'} \times_{Z_{\poset''}} X_{\sigma}.$$ Lemma \[l:fiberprodsexist\] and induction show that $Z'$ is an object in $\mathcal C_{T_M}$. It remains to show that $Z'$ is the inverse limit (\[eqn:invlimit\]) in the category of $\K$-schemes. Let $Y$ be any $\K$-scheme with maps $f_{\tau}$ to each $X_{\tau}$ for $\tau \in \poset$ that commute appropriately with the maps in $\poset$. By the universal property of the inverse limits $Z_{\poset'}$ and $Z_{\poset''}$ we get a unique map from $Z_{\poset'}$ to $Z_{\poset''}$ that commutes with the maps in $\poset$, and a unique map from $Y$ to $Z_{\poset'}$ and $Z_{\poset''}$ whose compositions with the maps from $Z_{\poset'}$ to $Z_{\poset''}$ and from $X_{\sigma}$ to $Z_{\poset''}$ agree. Then the universal property of fiber products gives a unique map $\phi$ from $Y$ to $Z'$ such that the composition of $\phi$ with the maps from $Z'$ to each $X_{\tau}$ equals the corresponding $f_{\tau}$, so $Z'$ is the inverse limit. \[defn:toricchow\] The *toric Chow quotient* of the toric variety $X_P$ by the subtorus $T_L$ is the not-necessarily-normal toric variety $X_P/T_L:= Z_0$ arising as the toric component of the inverse limit $Z$ as in Proposition \[prop:invlimit\]. Corollary \[coro:chow\] below justifies the choice of terminology; see also [@HaimanSturmfels]. First, we explain the link between the toric Chow quotient and fiber fans. \[p:toricChow\] Let $\overline{X_P/T_L}$ be the normalization of the toric Chow quotient. Then $\overline{X_P/T_L}$ is the toric variety with fan $\mathcal{N}(P,\pi(P))$. Write $\overline{Z_0} = \overline{X_P/T_L}$. For each $\sigma \in \poset$ there is a proper, birational toric morphism $\overline{Z_0}\rightarrow X_{\sigma}$, so the fan of $\overline{Z_0}$ refines the fan of $X_{\sigma}$ by [@Fulton Proposition 2.4]. In particular, the fan of $\overline{Z_0}$ refines the common refinement $\mathcal{N}(P,\pi(P))$ of these fans. The proposition follows once we show that $\mathcal{N}(P,\pi(P))$ refines the fan of $\overline{Z_0}$. It is enough to construct a proper, birational toric morphism $X_{CR}\rightarrow \overline{Z_0}$, where $X_{CR}$ is the toric variety with fan $\mathcal{N}(P,\pi(P))$. Since $\mathcal{N}(P,\pi(P))$ refines the fan of each $X_{\sigma}$, there is a proper, birational toric morphism $X_{CR}\rightarrow X_{\sigma}$. The universal property of the inverse limit $Z$ in the category $\mathcal{C}_{T_M}$ then gives a proper $T_M$-equivariant morphism $\phi\colon X_{CR}\rightarrow Z$ that is birational onto $Z_0$. Since $X_{CR}$ is normal, the universal property of normalization gives a morphism $\overline{\phi}\colon X_{CR}\rightarrow \overline{Z_0}$ that is both proper (since $\phi$ is proper and the normalization is separated) and birational (since $\phi$ is birational onto $Z_0$ and the normalization is finite). It remains to show that $\overline{\phi}$ is a toric morphism. Since $\phi$ is a dominant birational morphism of not-necessarily-normal toric varieties, it is given locally by an inclusion of subsemigroups $S\rightarrow S'$ of $M$, where $\Spec \K[S]$ is a chart on $Z_0$, $\Spec \K[S']$ is a chart on $X_{CR}$ and $T_M = \Spec \K[M]$. Note that $S'$ is saturated since $X_{CR}$ is normal. The normalization $\overline{Z_0}\rightarrow Z_0$ is given locally by the inclusion of $S$ in its saturation $S_{\sat}$. Since $S_{\sat}$ is the smallest saturated subsemigroup of $M$, the inclusion $S\hookrightarrow S'$ factors through $S_{\sat}$. The induced semigroup morphism $S_{\sat}\rightarrow S'$ gives the local description of the induced map $\overline{\phi}\colon X_{CR} \rightarrow \overline{Z_0}$, so $\overline{\phi}$ is a toric morphism. \[coro:chow\] Let $P$ be a polytope. Then the toric Chow quotient $X_P / T_L$ is equal to the Chow quotient in the sense of Kapranov [@Kapranov Section 0.1]. Write $\Chow$ for the Chow quotient. For a polytope $P$, Kapranov-Sturmfels-Zelevinsky [@KSZ Proposition 2.3] proved that the normalization of the Chow quotient is the toric variety defined by the fan $\mathcal{N}(P,\pi(P))$, so $X_P / T_L$ and $\Chow$ share the same normalization. To see that they coincide, recall from [@KSZ Corollary 4.3] that $\Chow$ is a subvariety of the inverse limit $Z$ defined in equation (\[eqn:invlimit\]). Since $\Chow$ is irreducible and contains the dense torus $T_Z$, it is a subvariety of the component $Z_0 = X_P/T_L$ of $Z$. If this inclusion $\Chow\subseteq X_P/T_L$ were strict, $\Chow$ and $X_P/T_L$ could not share the same normalization, hence $\Chow = X_P / T_L$ as claimed. The statement of Corollary \[coro:chow\] was known to the authors of [@KSZ Section 4], though they did not supply a proof. Thus, for a polyhedron $P$, the toric Chow quotient $X_P/T_L$ is the appropriate generalization of the Chow quotient $\Chow$. Proof of the main theorems {#s:mainproof} -------------------------- We now prove the main theorems. Recall that the fan of a not-necessarily-normal toric variety, such as the toric Chow quotient, is the fan of its normalization. \[thm:main\] Let $\pi\colon P\rightarrow Q$ be a linear projection of polyhedra. Set $P':= (\widetilde{P})^{\vee}$ and $L':= \image(\widetilde{\pi}^\vee)$. The cone over the GIT chamber decomposition arising from the $T_L$-action on $X_P$ is the fan of the toric Chow quotient $X_{P'}/T_{L'}$. The fan defining the toric Chow quotient $X_{P'}/T_{L'}$ is $\mathcal{N}(P',\widetilde{\pi}^\vee(P'))$ by Proposition \[p:toricChow\]. Now Proposition \[p:PveeFiberfan\] implies that the fan defining the toric Chow quotient $X_{P'}/T_{L'}$ is equal to the cone over the subdivision of $Q$ induced by $\pi(P)$. The result follows from Remark \[r:GITchamber\], since the subdivision of $Q$ induced by $\pi(P)$ is the GIT chamber decomposition arising from the $T_L$-action on $X_P$. The case when $P$ is a polyhedral cone (so $X_P$ is affine) is of particular interest. For example, Cox [@Cox] showed that every simplicial, projective toric variety can be constructed as a GIT quotient of affine space by a subtorus. In the affine case, Theorem \[thm:main\] can be strengthened as follows: \[thm:affineduality\] Let $P\subseteq \Q^n$ be a polyhedral cone and $\pi\colon P\rightarrow Q$ a linear projection. Set $P' := P^\vee$ and $L':= \image(\pi^\vee)$. Then 1. the fan of the toric Chow quotient $X_{P}/T_{L}$ is equal to the GIT chamber decomposition arising from the action of $T_{L'}$ on $X_{P'}$; and 2. the GIT chamber decomposition arising from the $T_L$-action on $X_P$ is the fan of the toric Chow quotient $X_{P'}/T_{L'}$. For part , note that the fan of $X_{P}/T_{L}$ is the fiber fan $\mathcal{N}(P,\pi(P))$ by Proposition \[p:toricChow\]. Corollary \[cor:coneduality\] shows that $\mathcal{N}(P,\pi(P))$ equals the fan induced by $\pi^\vee(P^\vee)$. This is precisely the GIT chamber decomposition arising from the action of $T_{L'}$ on $X_{P^\vee}$ according by Remark \[r:GITchamber\]. For part , note that the fan of the toric Chow quotient $X_{P'}/T_{L'}$ is the fiber fan $\mathcal{N}(P^\vee,\pi^\vee(P^\vee))$. This equals $(\pi^\vee)^\vee((P^\vee)^\vee) = \pi(P)$ by applying Corollary \[cor:coneduality\] with $C = P^\vee$ and $\pi^\vee$ playing the role of $\pi$. If $X$ is a nonnormal quasiprojective toric variety then one can give similar combinatorial descriptions of both GIT quotients by subtori $X \git_{\mathbf{v}}T$, and of the normalization of the toric Chow quotient; normality played no essential role in either construction. However the description of the GIT chamber complex in Definition \[d:GITchamber\] is now a priori only a coarsening of the true GIT chamber complex, as the quotient toric varieties are no longer determined solely by their fans. \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{} [CMT05]{} K. Altmann and J. Hausen, *[Polyhedral divisors and algebraic torus actions]{}*, (2003), To appear in Mathematische Annalen. L. Billera, P. Filliman, and B. Sturmfels, *Constructions and complexity of secondary polytopes*, Adv. Math. **83** (1990), no. 2, 155–179. L. Billera and B. Sturmfels, *Fiber polytopes*, Ann. of Math. (2) **135** (1992), no. 3, 527–549. A. Craw, D. Maclagan, and R. R. Thomas, *[The coherent component of the moduli of McKay quiver representations for abelian groups]{}*, (2005), arXiv:math.AG/0505115. D. Cox, *The homogeneous coordinate ring of a toric variety*, J. Algebraic Geom. **4** (1995), no. 1, 17–50. I. Dolgachev and Y. Hu, *Variation of geometric invariant theory quotients*, Inst. Hautes Études Sci. Publ. Math. (1998), no. 87, 5–56, With an appendix by Nicolas Ressayre. W. Fulton, *Introduction to toric varieties*, Annals of Mathematics Studies, vol. 131, Princeton University Press, Princeton, NJ, 1993, The William H. Roever Lectures in Geometry. A Grothendieck and J Dieudonné, *Éléments de géométrie algébrique. [I]{}. [L]{}e langage des schémas*, Inst. Hautes Études Sci. Publ. Math. (1960), no. 4, 228. M. Haiman and B. Sturmfels, *Multigraded [H]{}ilbert schemes*, J. Algebraic Geom. **13** (2004), no. 4, 725–769. Y. Hu, *Combinatorics and quotients of toric varieties*, Discrete Comput. Geom. **28** (2002), no. 2, 151–174. M. Kapranov, *Chow quotients of [G]{}rassmannians. [I]{}*, I. M. Gel fand Seminar, Adv. Soviet Math., vol. 16, Amer. Math. Soc., Providence, RI, 1993, pp. 29–110. M. Kapranov, B. Sturmfels, and A Zelevinsky, *Quotients of toric varieties*, Math. Ann. **135** (1992), no. 3, 527–549. M. Thaddeus, *Toric quotients and flips*, Topology, geometry and field theory, World Sci. Publishing, River Edge, NJ, 1994, pp. 193–213. , *Geometric invariant theory and flips*, J. Amer. Math. Soc. **9** (1996), no. 3, 691–723.

ArXiv

Standard Model in adS slice with UV-localized Higgs field Sergey Mironov$^{a,b}$, Mikhail Osipov$^{c,d}$, Sabir Ramazanov$^{a,c}$ *$^{a}$Moscow State University, Department of Physics,\ Vorobjevy Gory, 119991, Moscow, Russia\ $^{b}$Institute of Theoretical and Experimental Physics,\ B. Cheremushkinskaya ul. 25, 117259 Moscow, Russia\ $^{c}$Institute for Nuclear Research of the Russian Academy of Sciences,\ 60th October Anniversary prospect 7a, Moscow, 117312, Russia\ $^{d}$Moscow Institute of Physics and Technology,\ Institutskii per., 9, Dolgoprudny, 141700 Moscow Region, Russia* Abstract We discuss five-dimensional Standard Model in a slice of adS space-time with the Higgs field residing near or on the UV brane. Allowing fermion fields to propagate in the bulk, we obtain the hierarchy of their masses and quark mixings without introducing large or small Yukawa couplings. However, the interaction of fermions with the Higgs and gauge boson KK excitations gives rise to FCNC with no built-in suppression mechanism. This strongly constrains the scale of KK masses. We also discuss neutrino mass generation via KK excitations of the Higgs field. We find that this mechanism is subdominant in the scenarios of spontaneous symmetry breaking we consider. Introduction and summary ======================== Five-dimensional theories with the Standard Model fields living in adS slice attract considerable interest, especially due to their possible connection to adS/CFT correspondence [@Maldacena; @Gubser; @Witten]. Originally, theories of this sort focused on the problem of the hierarchy between the Planck and electroweak scales [@Sundrum]. Later on, it has been understood that they can explain the hierarchy of fermion masses and quark mixings without introducing large or small parameters into the original action [@Gherghetta; @Huber; @Stephan; @Grossman; @Chang]. AdS slice is a solution to the Einstein equations in 5-dimensional (5D) space-time with two gravitating branes. Upon fine-tuning the brane tension and 5D cosmological constant, one obtains the metric $$ds^2=e^{-2k|y|}\eta_{\mu \nu}dx^{\mu}dx^{\nu}-dy^2 \; , \label{metrics}$$ where $k$ is the adS curvature, $y$ denotes the coordinate of the fifth warped dimension, which is $S^1/Z_2$ orbifold of size $R$. Two branes are placed at $y=0$ and $y=\pi R$. These are ultraviolet (UV, $y=0$) and infrared (IR, $y=\pi R$) branes, respectively. We choose the 4-dimensional (4D) Minkowski metric as $\eta_{\mu \nu}=(1,-1,-1,-1)$. This setup is known as the Randall-Sundum model of type one (RS1) [@Sundrum]. In the original RS1 model, only gravity is supposed to propagate in the 5D space-time, while the SM fields reside on the IR brane. However, this is not the only possibility, as one can allow all particles or some of them propagate in the bulk. In that case, the 5D fields can be expanded in the tower of the Kaluza–Klein (KK) modes, and the zero modes are associated with the SM fields. The hierarchy of fermion masses and quark mixings is then obtained by an appropriate choice of the fermion localization in the bulk. It is most common to assume that the Higgs field is localized on or near the IR brane, for reviews see, e.g., Refs. [@Casagrande; @Neubert]. In that case, the zero modes of light fermions are localized near the UV brane, while zero modes of heavy quarks are localized towards the IR brane. The smallness of the SM Yukawa couplings is then due to the small overlaps of the zero modes of the light fermions and the Higgs field. Masses of KK modes of the gauge and Higgs fields (if the latter lives in the bulk) are constrained by the requirement of FCNC suppression. In models with the Higgs field localized on or near the IR brane, this constraint is fairly weak: excited KK modes can have masses of order 10 TeV. This is due to the so called RS–GIM mechanism [@Perez; @Agashe], which is built-in: since the zero modes of light fermions are localized near the UV brane, their overlaps with the KK modes of the gauge and Higgs fields are exponentially small. Moreover, introducing additional symmetries, it is possible to relax the constraints on KK masses down to about $3$ TeV [@Csaki; @Blanke; @Randall; @Santiago]. Thus, effects of new physics can be observed in experiments at LHC. According to the holographic picture [@Contino; @lesh; @Arkani], every bulk zero mode field corresponds to an eigenstate in the dual 4D theory, which is a mixture of elementary source and composite CFT fields. If the bulk zero mode is localized towards the UV brane, the massless eigenstate of the dual theory is predominantly the source field. Conversely, the dual interpretation of a bulk zero mode localized towards the IR brane is a state that is predominantly a CFT bound state. If the Higgs field is confined to the IR brane, it is interpreted as a pure CFT bound state in the dual theory. The top-quark is mostly a CFT bound state, while light fermions are mostly elementary. In this paper we turn this picture upside down and consider a scenario with the bulk Higgs field localized near or on the UV brane. Without introducing the hierarchy in the parameters of the original action, we show that the realistic pattern of fermion masses and quark mixings can be obtained in this case as well. However, the overall picture is quite different. Namely, light quarks and right leptons are localized near the IR brane to have small overlaps with the UV-localized Higgs field and, consequently, small Yukawa couplings. Hence, many light SM fields are mostly CFT composites in the dual picture, while heavy fields are predominantly elementary. The IR localization of light fermions introduces, however, the FCNC problem. Indeed, the KK excitations of the Higgs field and bulk gauge bosons also live near the IR brane. Thus, their wave functions have large overlaps with the wave functions of light fermions. From the point of view of the effective 4D theory, this means that the corresponding couplings are of order one. Then the only parameter one can use for suppressing FCNC is the mass scale of KK excitations. The latter must be very high to satisfy the existing constraints coming from kaon mixing. We will see that the constraint on the KK scale is particularly strong for the Higgs field living in the bulk. A way to avoid FCNC mediated by the Higgs KK excitations is to localize the Higgs field on the UV brane. Then there are no Higgs KK excitations at all. In that case the dominant source of FCNC is the exchange by the KK excitations of the gauge fields. Although the constraints here are less severe, the allowed scale of the KK excitations is still beyond the experimental reach. We also discuss neutrino masses of the Dirac type. It is straightforward to obtain them via the interaction with the zero mode of the Higgs field. An alternative possibility, which would probably be more interesting, would be the neutrino mass generation via the interaction with the KK excitations of the Higgs field. If it worked, the smallness of the neutrino masses would be due to the suppression of the vacuum expectation values of the heavy KK Higgs modes, rather than due to small effective 4D Yukawa couplings. In the particular scenarios we consider, this mechanism is subdominant, however: the neutrino interactions with the Higgs zero mode are always strong enough to generate the main contribution to the neutrino masses. This paper is organized as follows. In Section 2 we discuss possible scenarios of spontaneous symmetry breaking in the 5D Standard Model with the Higgs field localized towards the UV brane. In Section 3 we show that realistic 4D fermion masses and quark mixings can be obtained without introducing small or large parameters in the 5D action. There we also discuss neutrino masses of the Dirac type. We consider the FCNC problem in Section 4. We find that the constraints on the mass scale is $m_{KK}\gtrsim 5\times 10^{5}$ TeV for the Higgs field living in the bulk and $m_{KK} \gtrsim 700$ TeV for the Higgs field localized on the UV brane. This reiterates the power of the FCNC constraints in models without a built-in mechanism of the FCNC suppression. Scenarios of electroweak symmetry breaking ========================================== We consider the Standard Model in the slice $y\in (0,\pi R)$ of 5D adS space-time with the metric . We will see in Section 3 that the fermion mass hierarchy problem is naturally solved provided that $$\label{large} e^{k\pi R}\gg 1\; .$$ We treat $e^{k\pi R}$ as a large parameter in what follows. The action for the Higgs field living in the 5D bulk is $$\label{spsym1} S_{5}= \int d^{4}x dy \sqrt{g}\Bigl(\frac{1}{2}g^{M N}{\partial}_{M} H {\partial}_N H -\frac{1}{2}m^2_{H} {H}^2-V(H)\Bigr)+S_{b}\; ,$$ where $m_H$ is the bulk Higgs mass, $V(H)$ is the symmetry breaking potential. The brane term $S_{b}$ is added to have the zero mode in the absence of the potential $V(H)$ [@Gherghetta], $$\label{braneterm} \nonumber S_{b}=\Bigl(1-\frac{\alpha}{2}\Bigr)k\int d^4xdy\sqrt{g}(\delta (y-\pi R)-\delta(y)){H}^2\; .$$ The constant $\alpha$ is tuned to $\alpha =\sqrt{4+\frac{m^2_H}{k^2}}$, so that the zero mode exists. We will momentarily see that with the negative sign in front of $\alpha$ chosen in and $\alpha >1$, the Higgs zero mode is localized near the UV brane. This is the case we study in what follows. Let us first switch off the potential $V(H)$, i.e., set $V(H)=0$, and consider the free scalar field. One derives from the 5D action with the brane term the following equations of motion and boundary conditions, $$\partial_{\mu} \partial^{\mu} H+e^{2ky}\partial_5(e^{-4ky}\partial_5 H)+m^2_He^{-2ky}H=0\; ,$$ $$\label{bounds} \partial_5 H-(2-\alpha)k H|_{0,~\pi R}=0\; .$$ Following the standard procedure, we expand the field $H$ in the infinite sum: $$\label{scKK} H (x,y)=\sum_{n=0}^{\infty} {h}_{n} (x) {H}_{n} (y)\; ,$$ where $h_{n}(x)$ are KK modes with masses $m_{n}$, while $H_{n}(y)$ are their bulk profiles. The zero mode is given by [@Gherghetta] $$\label{scprofiles} H_{0}(y)=N_{0}e^{(2-\alpha) ky}\; .$$ It is clear from that the effective profile is actually $e^{-ky}H_{0}(y)$. Hence, the zero mode is UV-localized for $\alpha >1$. The normalization constant $N_{0}$ ensures the standard form of the kinetic term in the effective 4D action. The latter condition reads $$\label{norm} \int^{\pi R}_0 dye^{-2ky}H_{n}^2(y)=1\; ,$$ so that $$\label{scnorm} N_{0}=\sqrt {\frac{2k(\alpha-1)}{1-e^{2(1-\alpha)k\pi R}}} \approx \sqrt{2k(\alpha -1)}\; .$$ The profiles of the excited KK modes are given by [@Gherghetta] $$\label{scprofn} H_{n}(y)=N_{n} e^{2ky}\left[J_{\alpha}\Bigl(\frac{m_{n}}{k}e^{ky}\Bigr)+ \frac{J_{\alpha-1}(\frac{m_{n}}{k})}{J_{-\alpha+1}(\frac{m_{n}}{k})}J_{-\alpha} \Bigl(\frac{m_{n}}{k}e^{ky}\Bigr)\right]\; ,$$ with the normalization constants $$\label{scnormn} N_{n}\approx \frac{m_{n}}{\sqrt{k}} \frac{1}{\sqrt{{\int}^{\beta_n}_0 sJ^2_{\alpha}(s)ds}}\; .$$ Here $$\beta_n=\frac{m_n}{k}e^{k\pi R}\; .$$ The boundary conditions determine the eigenvalues $\beta_n$ and hence the masses of the KK excitations; these are found from $$J_{\alpha-1}(\beta_n)=0\; .$$ Clearly, the lowest KK modes have $\beta_n\sim 1$ and hence $$m_n\sim ke^{-k\pi R}\; .$$ Note that $\frac{m_n}{k}$ is a small parameter for not too large values of $n$ in the regime we consider. To obtain the Higgs VEV, we turn on the potential $V(H)$. Let us begin with the choice $$V(H)=-\frac{{\mu}^2}{2} {H}^2+\lambda {H}^4 \; ,$$ so that symmetry breaking occurs due to the bulk mass term. By inserting the KK decomposition into the action and integrating over the fifth coordinate, one obtains the effective 4D action. Assuming that the KK excitations are small, we treat the interaction between the zero modes and excited KK modes in the linear approximation in $h_{n}$ and write $$\label{spsymeff1} \begin{split} S_{eff}&=\int d^4 x \Bigl(\frac{1}{2}({\partial h_0})^2 +\sum_{n=1}^{\infty} \frac{1}{2}({\partial {h}_n})^2 -\sum_{n=1}^{\infty} \frac{1}{2}m^2_{n}{h}^2_n+ \frac{1}{2} {\mu}^2(c_0 h_0^2+ \sum_{n=1}^{\infty} 2 c_n {h}_0{h}_n)\\ &\quad -\lambda (a_0 {h}^4_0+\sum_{n=1}^{\infty} 4a_n {h}^3_0 {h}_n)\Bigr)\; , \end{split}$$ where the constants $a_{0}$, $a_{n}$, $c_0$ and $c_n$ are the overlap integrals $$\label{a0c0} a_0=\int_{0}^{\pi R} dy \sqrt{g} {H}^4_0\; , \quad c_0=\int_{0}^{\pi R} dy \sqrt{g}{H}^2_0\; ,$$ $$\label{ancn} a_n=\int_{0}^{\pi R} dy \sqrt{g} {H}^3_0 {H}_n \; , \quad c_n=\int_0^{\pi R} dy \sqrt{g} {H}_0 {H}_n\; .$$ Making use of the effective action , we derive VEVs of the zero and excited KK modes: $$\label{spsymvev0} v_0= \sqrt{\frac{c_0{\mu}^2}{4a_0 \lambda}}\; ,$$ $$\label{spsymvevn} v_n= \frac{c_n{\mu}^2v_0-4a_n\lambda v_0^3}{m^2_n}\; .$$ As the zero mode represents the standard Higgs field, we have: $v_0=v_{SM}= 247$ GeV. We evaluate the integrals and obtain $$\label{barhan} a_0\approx \frac{N^4_0}{4(\alpha -1)k}\; , \quad c_0\approx \frac{N^2_0}{2\alpha k}\; .$$ We see that the constants $a_0$ and $c_0$ are estimated as $a_0\sim k$ and $c_0 \sim 1$. Then the mass parameter $\mu$ responsible for symmetry breaking is of the order of the SM Higgs VEV, while $\lambda\lesssim k^{-1}$ in order that the effective 4D coupling $\lambda_4=\lambda a_0$ be small. Making use of in , we write the KK VEVs as follows $$v_{n}= (a_{0}c_{n}-a_{n}c_{0})\frac{v_0\mu^2}{a_0m^2_n}\; .$$ In what follows, we need the integrals to the subleading order in $\frac{m_{n}}{k}.$ The constants $a_{n}$ are different at $\alpha >2$ and $\alpha <2$, $$\label{an} \begin{split} a_n&=-\frac{N^3_0 N_n}{2^{\alpha} \Gamma (\alpha) m_n}\left(\frac{m_n}{k}\right)^{\alpha -1}\left[1-\frac{(\alpha -1)}{2(\alpha -2)(2\alpha -3)}\left(\frac{m_n}{k}\right)^2 \right] \qquad \alpha >2 \; ,\\ a_n&=-\frac{N^3_0 N_n}{2^{\alpha} \Gamma (\alpha) m_n}\left(\frac{m_n}{k}\right)^{\alpha -1}\left[1-2^{\alpha}\Gamma (\alpha)\left(\frac{m_n}{k}\right)^{2(\alpha -1)} \int^{\beta_n}_0s^{3(1-\alpha)}J_{\alpha}(s)ds \right] \qquad \alpha <2\; . \end{split}$$ The constants $c_{n}$ are given by $$\label{cn} \begin{split} c_n=-\frac{N_0 N_n}{\alpha 2^{\alpha} \Gamma (\alpha) m_n}\left(\frac{m_n}{k}\right)^{\alpha -1}\left[2(\alpha -1)+\left(\frac{m_n}{k}\right)^2 \ln \frac{m_n}{k}\right]\; . \end{split}$$ The constants $a_{n}$ and $c_{n}$ are estimated as $a_{n}\sim k\left(\frac{m_n}{k}\right)^{\alpha -1}$ and $c_{n}\sim \left(\frac{m_n}{k}\right)^{\alpha -1}$. Then a naive estimate of the KK VEVs would be $v_{n}\sim \left(\frac{v_{SM}}{m_n}\right)^2\left(\frac{m_n}{k}\right)^{\alpha -1}v_{SM}$. However, this is not the case. Indeed, to the leading order in $\frac{m_n}{k}$ the constants satisfy $$\frac{a_0}{c_0}=\frac{a_n}{c_n}\; .$$ Taking into account the subleading terms in Eqs. , , we obtain the following expressions[^1] for the KK VEVs at $\alpha >2$ and $\alpha <2$: $$\begin{aligned} \label{vevnn} v_n&= \frac{1}{2^{\alpha} \Gamma{(\alpha +1)}} \sqrt{\frac{2(\alpha -1)}{\int^{\beta_n}_0 sJ^2_{\alpha}(s)ds}} \left(\frac{\mu}{m_n}\right)^2 \left(\frac{m_{n}}{k}\right)^{\alpha +1}\ln \frac{k}{m_n}v_0 \qquad \alpha >2 \; ,\\ v_n&= \sqrt{\frac{8(\alpha -1)^3}{\int^{\beta_n}_0 sJ^2_{\alpha}(s)ds}} \left(\int^{\beta_{n}}_{0}s^{3(1-\alpha)} J_{\alpha}(s) ds\right)\left(\frac{m_{n}}{k}\right)^{3(\alpha -1)}\frac{{\mu}^2}{\alpha m^2_n}v_0 \qquad \alpha <2\; .\end{aligned}$$ So, modulo factors of order one, the estimates are $v_{n}\sim \left(\frac{v_{SM}}{m_n}\right)^2 \left(\frac{m_{n}}{k}\right)^{\alpha +1}v_{SM}$ in the case of the Higgs field localized with the parameter $\alpha>2$ and $v_n\sim\left(\frac{v_{SM}}{m_n}\right)^2\left(\frac{m_{n}}{k}\right)^{3(\alpha -1)}v_{SM}$ for $\alpha <2$. So strong suppression is absent in a model with another mechanism of spontaneous symmetry breaking. Instead of the potential , one introduces $$\label{spsym2} V(H)=-\frac{1}{2}M\delta (y)H^2+\lambda H^4\; ,$$ so that the mass term resides on the UV brane. Proceeding as before, we arrive at the following effective 4D action: $$\label{spsymmeff2} \begin{split} S_{eff}&=\int d^4 x \Bigl(\frac{1}{2}({\partial h_0})^2 +\sum_{n=1}^{\infty} \frac{1}{2}({\partial h_n})^2 -\sum_{n=1}^{\infty} \frac{1}{2}m^2_{n}h^2_n+ \frac{1}{2} M( h_0^2 H_0^2(0)+\\ &\quad +\sum_{n=1}^{\infty} 2h_0h_n H_0(0) H_n(0)) -\lambda(a_0 h^4_0+\sum_{n=1}^{\infty} 4a_n h^3_0 {h}_n)\Bigr)\; , \end{split}$$ where $a_0$ and $a_n$ are again the overlap integrals , . In this case, the VEVs are $$\label{vev0} v_0= \sqrt{\frac{Mk}{4a_0 \lambda}}\; ,$$ $$\label{vevn} v_n= \frac{M {H}_0 (0){H}_n (0)v_0-4a_{n}\lambda v_0^3}{m^2_n}\; .$$ We see from that the mass $M$ must be small, $M\sim \frac{v_{SM}^2}{k} \; .$ Using this estimate as well as Eqs. —, we find, modulo a factor of order one $$\label{estimate} v_n\sim \frac{v_{SM}^3}{m^2_n}\left(\frac{m_n}{k}\right)^{\alpha -1}\; .$$ It is straightforward to show that the cancellation between the leading order terms does not occur in Eq. , so the estimate is indeed valid. However, this value is still very small from the viewpoint of physical applications discussed in Section 3.2. Mass pattern of fermions ======================== Quarks ------ The action for free bulk fermions is $$\label{freedom} S^{\Psi}_5=\int d^4 x dy\sqrt{g} \left(ig^{MN}\bar{\Psi}\Gamma_{M}\nabla_N \Psi -m_{\Psi} \bar{\Psi} \Psi \right)\; ,$$ where $\Gamma_{M}$ are the 5D gamma matrices in adS space-time, $\nabla_{M}$ is the covariant derivative, $m_{\Psi}$ is the fermion bulk mass. One chooses the fermions transforming as $\Psi (-y) =\pm \gamma_5 \Psi (y)$ under the orbifold $Z_2$ symmetry, where the lower sign refers to $SU(2)_{L}$-doublets $Q$ and the upper one to singlets $u$ and $d$. As a result, there are no left zero modes of singlet quarks and right zero modes of doublets [@Gherghetta; @Grossman] and one arrives at the SM chiral structure. Hereafter we consider zero modes of fermions. Their profiles are given by [@Gherghetta; @Grossman] $$\label{prferm} Q_{0}(y)=N_{L}e^{(2-c_{L})ky}\; ,\quad u_{0}(y)=N^{u}_{R}e^{(2-c^{u}_{R})ky}\; , \quad d_{0}(y)=N^{d}_{R}e^{(2-c^{d})ky}\; .$$ The constants $c_{L,R}$ are related to the fermion bulk masses, $c_{R}=\frac{m_{\Psi}}{k}\;, ~c_{L}=-\frac{m_{\Psi}}{k} \; ,$ and the normalization constants are $$\label{fermnorm} N_{L}=\sqrt{\frac{(1-2c_{L})k}{e^{(1-2c_{L})k\pi R}-1}}\; , \quad N^{u,d}_{R}=\sqrt{\frac{(1-2c^{u,d}_{R})k}{e^{(1-2c^{u,d}_{R})k\pi R}-1}}\; .$$ It is worth noting that with the account of the warp factor in , the effective profiles of the zero modes are $$\label{aralka} \Psi_0 =N e^{(1/2-c)ky}\; .$$ Hence, the zero modes are localized towards the IR and UV branes for $c<1/2$ and $c>1/2$, respectively. Now, assuming that the Higgs field lives in the bulk, we introduce its interaction with fermions, $$\label{quarks5D} \begin{split} S^{q}_{5}=\int d^4xdy \sqrt{g} \Bigl(\lambda_{ij}^{d}\bar{Q_i}Hd_j+\lambda_{ij}^{u}\bar{Q_i}\widetilde{H}u_j+h.c.\Bigr)\; . \end{split}$$ Neglecting the excited KK modes of the Higgs field for the time being and integrating Eq.  over the extra dimensional coordinate, we derive the effective 4D action: $$\label{quarkseff} S_{eff}^{q}=\int d^4x \Bigl(\lambda_{ij}^{d}I^{d}_{0ij}\bar{d}_{Li}(x)d_{Rj}(x)h_{0}(x)+\lambda_{ij}^{u}I^{u}_{0ij}\bar{u}_{Li}(x) u_{Rj}(x)h_{0}(x)+h.c.\Bigr)\; ,$$ where $I^{u,d}_{0ij}$ are the overlap integrals of the zero modes of the Higgs and quark fields, $$\label{overlap0} I^{u,d}_{0ij}=\int^{\pi R}_{0} dy \sqrt{g}H_0(y)Q_{0i}(y)d_{0j}(y)\; .$$ Explicitly, $$\label{overlap0expl} I^{u,d}_{0ij}=N_{Li}N^{u,d}_{Rj}N_0\frac{1-e^{(2-\alpha-c_{Li}-c^{u,d}_{Rj})k\pi R}}{(\alpha+c_{Li}+c^{u,d}_{Rj}-2)k}\; .$$ In the low-energy theory, the action leads to the quark mass matrix, $$\label{massquarks1} M^{u,d}_{ij}={\lambda}_{ij}^{u,d}I_{0ij}^{u,d}v_{SM}\; .$$ In our calculations we assume that the condition $2-\alpha-c_{Li}-c_{Rj} <0$ is satisfied, so that the exponential factor in Eq.  can be neglected. Furthermore, we do not introduce the hierarchy between the 5D Yukawa couplings and set $\lambda^{u,d}_{i,j} \sim k^{-1/2}$. Then, using Eq. , we estimate the elements of the mass matrix as follows, $$\label{massquarkssim} M_{ij}^{u,d}\sim N_{Li} N_{Rj}^{u,d} \frac{v_{SM}}{k}\; .$$ The hierarchy between quark masses and mixings is generated by the hierarchy between the normalization constants $N_{Li}$, $N_{Rj}$, which in turn is due to zero mode profiles. Aiming at diagonalizing the mass term, we perform the unitary transformations of left and right quarks (up- and down-quarks independently) with corresponding matrices $A_{L}$ and $A_{R}$. The latter satisfy the conditions $$\label{krolik} A_{L}MM^{\dagger}A_{L}^{-1}=M_{diag}^2\; , \quad A_{R}{M}^{\dagger}{M} {A}_{R}^{-1}=M_{diag}^2\; .$$ These ensure that the quark mass matrix $$\label{sabirman} m=A_{L}MA^{-1}_{R}$$ is diagonal. With no hierarchy between the Yukawa couplings, the Hermitean matrix $MM^{\dagger}$ has the following structure, $$\label{mm} (MM^{\dagger})_{ij} \sim N_{Li}N_{Lj}\sum_{k=1}^{3}N^2_{Rk} \frac{v^2_{SM}}{k^2}\; .$$ So, only the normalization constants of the doublets $N_{Li}$ are responsible for the hierarchy in Eq. . We order them as follows: $N_{L1}\ll N_{L2}\ll N_{L3}$. Then the elements of the matrix $A_{L}$ are estimated by the the ratios of the constants $N_{Li}$: $$\label{AL} A_L \sim\left( \begin{array}{cccc} 1 & \frac{N_{L1}}{N_{L2}} & \frac{N_{L1}}{N_{L3}} \\ \frac{N_{L1}}{N_{L2}} & 1 & \frac{N_{L2}}{N_{L3}}\\ \frac{N_{L1}}{N_{L3}} & \frac{N_{L2}}{N_{L3}} & 1 \end{array} \right)\; .$$ By analogy, the Hermitean matrix $M^{\dagger}M$ is estimated as $$M^{\dagger}M\sim N_{Ri}N_{Rj}\sum_{k=1}^{3}N^{2}_{Lk}\frac{v^2_{SM}}{k^2}\; .$$ Then, assuming the hierarchy of the normalization constants, $N_{R1}\ll N_{R2}\ll N_{R3}$, we obtain $$\label{AR} A_R \sim\left( \begin{array}{cccc} 1 & \frac{N_{R1}}{N_{R2}} & \frac{N_{R1}}{N_{R3}} \\ \frac{N_{R1}}{N_{R2}} & 1 & \frac{N_{R2}}{N_{R3}}\\ \frac{N_{R1}}{N_{R3}} & \frac{N_{R2}}{N_{R3}} & 1 \end{array} \right)\; .$$ Formally, this estimate is valid also for $N_{R1}\sim N_{R2}\sim N_{R3}$, as is the case for down-quarks (see below). Finally, using the estimates , and , we estimate the quark masses: $$\label{sim} m^{u,d}_{i}\sim N_{Li}N^{u,d}_{Ri}\frac{v_{SM}}{k}\; .$$ Now, let us consider flavor mixing in the quark sector, which is described by the Cabibbo–Kobayashi–Maskawa matrix. The latter is given by $$\nonumber C=A_{L}^d ({A^u_L})^{-1}\; .$$ Since the matrices $A_{L}^d$ and $A^u_L$ have one and the same general structure given by Eq. , the CKM matrix is also estimated by the right hand side of Eq. . We now recall the entries of the CKM matrix, $$\label{CKM} |C|=\left( \begin{array}{cccc} 0.97 & 0.23 & 0.0040\\ 0.23 & 0.97 & 0.042\\ 0.0081& 0.041& 0.99 \end{array} \right)\; ,$$ and compare them with Eq. . We see that the right pattern is obtained for $$\label{NL} \frac{N_{L1}}{N_{L2}}\approx \frac{1}{10}\; ,\quad \frac{N_{L2}}{N_{L3}}\approx \frac{1}{25}\; .$$ (50,105)(15,40) (62.5,153)[(0,-1)[77]{}]{} (140.5,153)[(0,-1)[77]{}]{} (62.5,81)(137.5,81.375)(140.5,122.25) (140.5,81.25)(68.5,80.875)(62.5,100) (62.5,81)(137,81.125)(140.5,97.75) (62.5,81)(132.5,81.375)(140.5,136.25) (59.25,150)[(0,0)\[cc\][UV]{}]{} (144,150)[(0,0)\[cc\][IR]{}]{} (58.75,80)[(0,0)\[cc\][0]{}]{} (127,144)[$(u,d)_{L}$]{} (51,128)[$(t,b)_{L}$]{} (67,95)[$t_{R}$]{} (141,105)[$(c,s)_{L}$]{} (141,96)[$c_{R}$]{} (141,118.5)[$b_{R},s_{R}$]{} (141,133)[$u_{R},d_{R}$]{} (67,133)[$H_{0}$]{} (62.5,80.75)(132.5,81)(140.25,107.25) (62.5,80.75)(126.75,80.75)(140.25,150.75) (140.5,81)(71.5,82.375)(62.5,129.25)(140.5,81)(72,81.250)(62.5,150) The estimate is consistent with known experimental values of the quark masses [@pdg] $m_{u}\approx 2.6$ MeV, $m_d\approx 5.0$ MeV, $m_c\approx 1.3$ GeV, $m_s\approx 100$ MeV, $m_b\approx 4.2$ GeV, $m_t\approx 171$ GeV, provided that $$\frac{N^{u}_{R1}}{N^{u}_{R2}}\approx \frac{1}{30}\; ,\quad \frac{N^{u}_{R2}}{N^{u}_{R3}}\approx \frac{1}{5}\; ,\quad \frac{N^{d}_{R1}}{N^{d}_{R2}}\approx \frac{1}{2}\; ,\quad \frac{N^{d}_{R2}}{N^{d}_{R3}}\approx 1\; ,\quad \frac{N^{d}_{R3}}{N^{u}_{R3}}\approx \frac{1}{40}\; .$$ The overall scale of the normalization constants is obtained by requiring that the mass of the top quark has the correct value. This gives $N_{L3} N^{u}_{R3} \simeq k$. As an example, we choose $$N^{u}_{L3}=1.3 \sqrt{k}\; , \quad N^{u}_{R3}=0.6 \sqrt{k}\; .$$ Finally, using formulas , one estimates the dimensionless bulk masses of quarks, which we denoted by $c$. Their values for the warp factor $k\pi R =10$ are given in Table \[muchacha\]. We see explicitly that $c_{L,R}<1/2$ for the lightest four quarks. As we noticed above, this means that they live near the IR brane, as expected. Qualitative picture of the quark localization is shown in Fig. \[Klara\]. We have still to choose the set of 5D Yukawa couplings $\lambda^{u,d}_{ij} \sqrt{k}$. Allowing them to vary within the interval $(1/3,~3)$, one can adjust masses and flavor mixings. Let us set $$\label{lambdau} \lambda^{u}_{ij} \sqrt{k} = \left( \begin{array}{cccc} 1.2 & 0.4 & -1.9 \\ 1.7 & 1.1 & -0.9\\ -0.8 & 0.6 & 1.3 \end{array} \right)\; ,$$ $$\label{lambdad} \lambda^{d}_{ij} \sqrt{k} = \left( \begin{array}{cccc} 0.9 & -0.4 & 1.3 \\ -1.3 & 1.3 & -0.5\\ 1.8 & 0.3 & 1.1 \end{array} \right)\; ,$$ where we ignore phases for the sake of simplicity. With these 5D Yukawa couplings one obtains the mixing matrix $$\label{CKM} |C|= \left( \begin{array}{cccc} 0.97 & 0.25 & 0.0040\\ 0.25 & 0.97 & 0.040\\ 0.0010& 0.040& 0.998 \end{array} \right)$$ and quark masses $m_{u}\approx 3$ MeV, $m_{d}\approx 9$ MeV, $m_{c}\approx 1.5$ GeV, $m_{s}\approx 90$ MeV, $m_{b}\approx 4.5$ GeV, $m_{t}\approx 170$ GeV. Obviously, all these values are in a reasonable agreement with the experimental data. $c$ $Q$ $u$ $d$ ---------- ------ ----- ----- $c_{L1}$ -0.1 - - $c_{L2}$ 0.2 - - $c_{L3}$ 1.3 - - $c_{R1}$ - 0.0 0.0 $c_{R2}$ - 0.4 0.1 $c_{R3}$ - 0.7 0.1 : \[muchacha\] Quark parameters in the 5D SM with warp factor $k\pi R=10$. Leptons ------- The interaction of leptons with the Higgs field in the 5D bulk is $$\label{leptons5D} \begin{split} S^{l}_5&=\int d^4xdy \sqrt{g} \Bigl(\lambda^{l}_{ij}\bar{L_i}l_{j}+{\lambda}^{\nu}_{ij}\bar{L_i}\widetilde{H} \nu_{j}+h.c.\Bigr)\; , \end{split}$$ where $L_{i}(x,y)$ are the lepton $SU(2)_{L}$-doublets, $l_{j}(x,y)$ and $\nu_j(x,y)$ are singlet charged and neutral leptons, respectively, and $\lambda^{l}_{ij}$ and $\lambda^{\nu}_{ij}$ are their 5D Yukawa couplings. This interaction generates the neutrino masses of the Dirac type. An interesting possibility here would be that neutrinos obtain their masses predominantly via the interaction with excited KK modes of the Higgs field. Then the smallness of the neutrino masses would be due to the suppression of VEVs of the Higgs KK excitations. Let us see, however, that this mechanism does not work in the model we discuss. To this end, we keep all modes in the decomposition of the Higgs field . Inserting the latter into Eq. , we derive the neutrino mass matrix in the low-energy limit, $$\label{leptonsmasses} M^{\nu}_{ij}= \lambda^{\nu}_{ij} \Bigl(v_0I^{\nu}_{0ij} +\sum_{n=1}^{\infty} v_nI^{\nu}_{nij}\Bigr)\; ,$$ where $v_0$ and $v_n$ are VEVs of the zero and excited modes of the Higgs field, as described in Section 2; $I^{\nu}_{0ij}$ and $I^{\nu}_{nij}$ are the overlap integrals of appropriate wave functions. The first integral is given by the Eq.  with the substitution $u \rightarrow \nu$, while the second one is $$I^{\nu}_{nij}=\int_{0}^{\pi R} dyL_{io}(y)H_{n}(y)\nu_{j}(y)\; ,$$ where $H_{n}(y)$ is given by Eq. . Explicitly, $$\label{lol} I^{\nu}_{nij}=N_{n}N_{Li}N^{\nu}_{Rj} \frac{1}{m_n}\left(\frac{m_n}{k}\right)^{c_{Li}+c^{\nu}_{Rj}-1} \int^{\beta_n}_{\frac{m_n}{k}} s^{1-c_{Li} -c^{\nu}_{Rj}} J_{\alpha}(s)ds\; .$$ Hereafter we assume that $c_{L}+c^{\nu}_{R} <2+\alpha$, so that the integral here is of order $1$. One can show that our basic conclusion is valid in the opposite case as well. The expression shows that neutrinos obtain their masses due to their interactions with both zero mode and excited modes of the Higgs field. The excited mode contribution would dominate for $$\label{inequal1} v_0 I^{\nu}_{0} \ll \sum_{n=1}^{\infty}v_nI^{\nu}_{n}\; .$$ Omitting summation and using and , we rewrite Eq.  as follows, $$\label{inequal} |e^{(2-\alpha -c_L -c^{\nu}_R)k\pi R} -1|\ll e^{(1-c_L -c^{\nu}_R)k\pi R} \frac{v_n}{v_{0}}\; .$$ Here we use the fact that $\frac{m_{n}}{k}e^{k\pi R}\sim 1$ for the lightest KK modes. The condition is equivalent to the following two, $$e^{(c_L +c^{\nu}_R -1)k\pi R}\ll \frac{v_n}{v_0}$$ and $$\label{inequality} e^{(1-\alpha)k\pi R}\ll \frac{v_n}{v_0}\; ,$$ which must be satisfied simultaneously. We now recall the expressions and for VEVs of the excited Higgs modes, and find that in both scenarios of spontaneous symmetry breaking considered in Section 2, the inequality is not satisfied. In the best case, the contribution of the Higgs KK excitations is suppressed by the small factor $\frac{v^{2}_{SM}}{m^2_n}$ as compared to the zero mode. (50,105)(15,40) (62.5,153)[(0,-1)[77]{}]{} (140.5,153)[(0,-1)[77]{}]{} (59.25,150)[(0,0)\[cc\][UV]{}]{} (144,150)[(0,0)\[cc\][IR]{}]{} (58.75,80)[(0,0)\[cc\][0]{}]{} (49,105)[$(\nu_{e}, e)_{L}$]{} (68,130)[$(\nu_{\tau}, \tau)_{L}$]{} (57,133)[$H_{0}$]{} (48.5,120)[$(\nu_{\mu}, \mu)_{L}$]{} (132.5,139.5)[$\nu_{eR}$]{} (141,108)[$e_{R}$]{} (141,119)[$\nu_{\tau R}$]{} (141,89)[$\tau_{R}$]{} (141,134)[$\nu_{\mu R}$]{} (141,99)[$\mu_{R}$]{} (140.5, 81)(72,81.250)(62.5,150) (62.5,81)(137.5,81.375)(140.5,122.25) (62.5,81)(132.5,81.375)(140.5,136.25) (62.5,81)(124.75,81.25)(140.5,147.5) (62.5,81)(131.5,80.75)(140.5,109.5) (62.5,81)(133.5,81.75)(140.5,100.5) (62.5,81)(137.5,81.125)(140.5,89.75) (140.5,81)(74.5,80.875)(62.5,120) (140.5,81)(72,81.125)(62.5,105) (140.5,81)(72,82.250)(62.5,150)(140.5,81)(72,81.500)(62.5,135) Thus, all lepton masses are obtained via the interaction with the zero mode of the Higgs field. Still, the picture here is rather different as compared to the quark sector. Indeed, the lepton mixing matrix does not exhibit strong hierarchy [@pdg], $$|C| \approx \left( \begin{array}{cccc} 0.79-0.88 & 0.47-0.61 & <0.18\\ 0.19-0.52 & 0.42-0.73 & 0.52-0.82\\ 0.20-0.53 & 0.44-0.74 & 0.56-0.81 \end{array} \right)\; .$$ Similarly to the case of quarks, we estimate it by the right hand side of Eq. . Then we conclude that the normalization constants $N_{L}$ are of one and the same order. It is therefore natural to assume that all $N_{L}\sim \sqrt{k}$, so that the dimensionless bulk masses of lepton doublets $c_L >1/2$, i.e., the doublets reside near the UV brane. Otherwise, we would need to fine tune the parameters $c_{L}$ to be very close to each other. Obviously, up to the change of notations $u\rightarrow \nu$ and $d\rightarrow l$, the estimate Eq.  remains valid for leptons. By choosing the normalization constants of singlet fermions $N_{R}$ and thus the parameters $c^{l}_R$ and $c^{\nu}_R$ in an appropriate way, one adjusts 4D masses of leptons. Since we assume that $N_{L}\sim \sqrt{k}$, the constants $N_{R}$ must be small, $N_{R}\ll \sqrt{k}$, in order that the lepton masses be small compared to the Higgs VEV. Hence, all $c^{l}_R$ and $c^{\nu}_R$ must be smaller than $1/2$. So, we come to the assignment that all singlet leptons reside towards the IR brane. Finally, in Table \[iducha\] we present the set of the 5D parameters $c$ leading to the correct hierarchy of lepton masses. We choose the normal hierarchy of neutrino masses and assume no degeneracy, i.e. $m_{1}\ll m_{2}$, $m_{2}\approx \sqrt{\Delta m^{2}_{sol}}\approx 0.008$ eV and $m_{3}\approx \sqrt{\Delta m^{2}_{atm}}\approx 0.05$ eV. The qualitative picture of lepton localization in the 5D bulk is shown in Fig. \[Klarushka\]. $c$ $L$ $\nu$ $e$ ---------- ----- --------- ------ $c_{L1}$ 1.0 - - $c_{L2}$ 2.0 - - $c_{L3}$ 3.0 - - $c_{R1}$ - $<-2.1$ -0.8 $c_{R2}$ - -2.1 -0.3 $c_{R3}$ - -1.8 0.1 : \[iducha\] Lepton parameters in the 5D SM with warp factor $k\pi R=10$. To conclude, masses and mixings in both quark and lepton sectors are reproduced in the model we discuss without introducing large or small parameters. The profiles of the fermion and Higgs zero modes are naturally steep in the warped fifth dimension, which translates into the strong hierarchies of masses and quark mixings in the 4D world. The non-hierarchical pattern of neutrino mixings is also natural with our choice of the localization of the left lepton doublets. Kaon mixing =========== Kaon mixing mediated by excited Higgs field ------------------------------------------- Unlike in the case of the Higgs field localized towards the IR brane, the FCNC suppression is not at all automatic in models we consider. The main source of FCNC in the model with the bulk Higgs field, whose zero mode is localized near the UV brane, is the interaction of light quarks with the KK excitations of the Higgs field. This interaction is fairly strong because both zero modes of light quarks and the Higgs KK modes are large near the IR brane, so they overlap substantially. Let us consider in detail the interaction of light down-quarks with the KK excitations of the Higgs field. Since this interaction is not diagonal in the flavor space, it leads to kaon mixing, which is severely constrained by experiment. Integrating the relevant terms in the action over the fifth coordinate, we arrive at the effective 4D action, $$\label{effex} S^{eff}_{quark}=\int d^4x \Bigl(\sum_{n=1}^{\infty} y^{d}_{nij}\bar{d}_{Li}(x)d_{Rj}(x)h_{n}(x)+h.c.\Bigr)\; .$$ Here $y^{u,d}_{nij}$ are the effective 4D Yukawa couplings. They are given by $$\label{Yukawa} y_{nij}=\lambda^{d}_{ij}I^{d}_{nij}\; ,$$ where $I^{d}_{nij}$ are the overlap integrals of the Higgs KK excitations and zero modes of the singlet down-quarks and quark doublets, $$\label{overlapn} I^{d}_{nij}=\int dy \sqrt{g}Q_{0i}(y)H_n(y)d_{0j}(y)\; .$$ Explicitly, $$\label{karas} I^{d}_{nij}=N_{Li}N^{u,d}_{Rj} N_{n}\frac{1}{m_n}\left(\frac{m_n}{k}\right)^{c_{Li}+c^{u,d}_{Rj}-1} \int^{\beta_n}_{0} s^{1-c_{Li} -c^{u,d}_{Rj}} J_{\alpha}(s)ds\; .$$ As shown in Section 3.1, the lightest down-quarks reside towards the IR brane. Thus, their normalization constants are $$N_{Li}\approx \sqrt{(1-2c_{Li})k}e^{(c_{Li}-1/2)k\pi R}, \quad N^{d}_{Rj}\approx \sqrt{(1-2c^{d}_{Rj})k}e^{(c^{d}_{Rj}-1/2)k\pi R}\; .$$ In this way we obtain the 4D Yukawa couplings of s- and d-quarks, $$\label{baku} y_{nij}\approx \lambda^{d}_{ij}\sqrt{\frac{(1-2c_{Li})(1-2c^{d}_{Rj})k}{\int^{\beta_n}_0sJ^2_{\alpha}(s)ds}} \int^{\beta_n}_0 \Bigl(\frac{s}{\beta_n}\Bigr)^{1-c_{Li}-c^{d}_{Rj}}J_{\alpha}(s) ds\; .$$ Here the flavor indices are $i,j=1,2$, and the integrals are of order 1. Hence, the Yukawa couplings are unsuppressed, $y_{nij}\sim 1$. To obtain the Yukawa couplings of the physical quark states, we perform the rotation of the quark fields with the matrices $A_{L}$ and $A_{R}$. We obtain in the physical basis $$\begin{split} y'_{n12}& \approx a_{L11}(y_{n11}a^{-1}_{R12}+y_{n12}a^{-1}_{R22})\; ,\\ y'_{21}& \approx a_{L22}(y_{n21}a^{-1}_{R11}+y_{n22}a^{-1}_{R21})\; , \end{split}$$ where we neglect the $b$-quark contribution; the constants $a$ are the entries of the matrices $A_{L}$ and $A_{R}$ estimated by Eqs. , . Obviously, the physical Yukawa couplings are also unsuppressed, $y'_{n12} \sim y'_{n21} \sim 1$. This precisely means that the RS-GIM mechanism does not work in the case of the Higgs field residing near the UV brane. Consequently, the only way to suppress dangerous FCNC is to assume that the Higgs KK excitations have very large masses. (50,85)(52,40) (81.558,92.389)[(1,0)[.9797]{}]{} (83.518,92.376)[(1,0)[.9797]{}]{} (85.477,92.362)[(1,0)[.9797]{}]{} (87.436,92.349)[(1,0)[.9797]{}]{} (89.396,92.335)[(1,0)[.9797]{}]{} (91.355,92.322)[(1,0)[.9797]{}]{} (93.315,92.308)[(1,0)[.9797]{}]{} (95.274,92.295)[(1,0)[.9797]{}]{} (97.234,92.281)[(1,0)[.9797]{}]{} (99.193,92.268)[(1,0)[.9797]{}]{} (101.153,92.254)[(1,0)[.9797]{}]{} (69,103.25)[(1,-1)[.07]{}]{}(56.5,114.25)(.0382848392,-.0336906585)[653]{}[(1,0)[.0382848392]{}]{} (70.80,80.5)[(-1,-1)[.07]{}]{}(81.5,92.25)(-.0336990596,-.0368338558)[638]{}[(0,-1)[.0368338558]{}]{} (115.50,103.375)[(-1,-1)[.07]{}]{}(128,114.75)(-.03777777778,-.0337037037)[675]{}[(-1,0)[.03777777778]{}]{} (115.15,79.625)[(1,-1)[.07]{}]{}(102.25,92)(.03440054496,-.03371934605)[734]{}[(1,0)[.03440054496]{}]{} (90,95)[$h_{n}$]{} (61.5,112)[$s_{L}$]{} (61.5,77)[$\overline{d}_{R}$]{} (116,110)[$\overline{s}_{R}$]{} (120,77)[$d_{L}$]{} (50,85)(15,40) (72.68,92.43)[(1,0)[.9797]{}]{} (74.639,92.416)[(1,0)[.9797]{}]{} (76.599,92.403)[(1,0)[.9797]{}]{} (78.558,92.389)[(1,0)[.9797]{}]{} (80.518,92.376)[(1,0)[.9797]{}]{} (82.477,92.362)[(1,0)[.9797]{}]{} (84.436,92.349)[(1,0)[.9797]{}]{} (86.396,92.335)[(1,0)[.9797]{}]{} (88.355,92.322)[(1,0)[.9797]{}]{} (90.315,92.308)[(1,0)[.9797]{}]{} (92.274,92.295)[(1,0)[.9797]{}]{} (60.5,103.25)[(1,-1)[.07]{}]{}(48,114.25)(.0382848392,-.0336906585)[653]{}[(1,0)[.0382848392]{}]{} (62.25,80.5)[(-1,-1)[.07]{}]{}(73,92.25)(-.0336990596,-.0368338558)[638]{}[(0,-1)[.0368338558]{}]{} (105.85,103.375)[(-1,-1)[.07]{}]{}(118.5,114.75)(-.03777777778,-.0337037037)[675]{}[(-1,0)[.03777777778]{}]{} (106,79.625)[(1,-1)[.07]{}]{}(93.5,92)(.03440054496,-.03371934605)[734]{}[(1,0)[.03440054496]{}]{} (81,95)[$h_{n}$]{} (54,112)[$s_{L}$]{} (53,77)[$\overline{d}_{R}$]{} (109.5,112)[$\overline{s}_{L}$]{} (111.5,77)[$d_{R}$]{} (-12,75)[a)]{} (69,75)[b)]{} Generally, $\Delta F=2$ processes are described by the following Hamiltonian: $$\label{Wilson} H^{\Delta F=2}_{eff}=\sum_{a=1}^{5} C_a Q_a^{q_i q_j}+ \sum_{a=1}^{3} \widetilde{C}_a \widetilde{Q}_a^{q_i q_j}\; .$$ The four-fermion operators $Q_{a}$ are given by $$\label{Wilsons} Q_1^{q_i q_j}=\bar{q}^{\alpha}_{jL}\gamma_{\mu}q^{\alpha}_{iL} \bar{q}^{\beta}_{jL}\gamma^{\mu}q^{\beta}_{iL}, \quad Q_2^{q_i q_j}=\bar{q}^{\alpha}_{jR}q^{\alpha}_{iL} \bar{q}^{\beta}_{jR}q^{\beta}_{iL}, \quad Q_3^{q_i q_j}=\bar{q}^{\alpha}_{jR}q^{\beta}_{iL} \bar{q}^{\beta}_{jL}q^{\alpha}_{iR}\; ,$$ $$\nonumber Q_4^{q_i q_j}=\bar{q}^{\alpha}_{jR}q^{\alpha}_{iL} \bar{q}^{\beta}_{jL}q^{\beta}_{iR}, \quad Q_5^{q_i q_j}=\bar{q}^{\alpha}_{jR}q^{\beta}_{iL} \bar{q}^{\beta}_{jL}q^{\alpha}_{iR}\; .$$ The operators $\widetilde{Q}_a$ are obtained by the interchange $L\leftrightarrow R$. Choosing $q_{j},q_{i} =d,s$, we focus on kaon mixing. The exchange by the KK excitations of the Higgs field contributes to the coefficients $C_2$, $\widetilde{C}_2$ and $C_4$, $$\label{row24} C_{2}=\sum^{\infty}_{n=1} \frac{({y'}_{n21}^{\star})^2}{m^2_n}\; , \quad \widetilde{C}_2=\sum^{\infty}_{n=1} \frac{(y'_{n12})^2}{m^2_n}\; , \quad C_{4}=\sum^{\infty}_{n=1} \frac{2{y'}_{n12}{{y'}^{\star}_{n21}}}{m^2_n}\; .$$ This is shown in Fig. \[aralk\], where Figs. \[aralk\]a and  \[aralk\]b correspond to $C_{2}$ and $C_{4}$, respectively. The imaginary parts of the coefficients $C$ are responsible for CP-violating mixing of kaons $K^{0}_1$ and $K^{0}_2$. Within our model, there is no natural way to suppress the phases of the coefficients $C$. Thus, we assume that $\mbox{Im}C\sim C$. The experimental constraints on the imaginary parts are [@Bona] $$\label{bon24} -5.1 \times 10^{-17} ~\mbox{GeV}^{-2} \lesssim ~\mbox{Im} C_{2}, ~\mbox{Im} \widetilde{C}_2 \lesssim 9.3 \times 10^{-17} ~\mbox{GeV}^{-2}\; ,$$ $$\label{bona24} -1.8 \times 10^{-17} ~\mbox{GeV}^{-2}\lesssim ~\mbox{Im} C_{4} \lesssim 0.9 \times 10^{-17} ~\mbox{GeV}^{-2}\; .$$ By comparing Eq.  with Eq.  and recalling that $y'_{nij} \sim 1$, we see that the masses of the Higgs KK excitations must be very large: $$\label{constraint!} m_n\gtrsim 5\times 10^{5} ~\mbox{TeV}\; .$$ The real parts of the coefficients $C$ contribute to the kaon mass difference. The corresponding constraints are three orders of magnitude weaker than and . Using these constraints, we find that irrespectively of the above assumption $\mbox{Im} C\sim C$, the masses of the Higgs KK excitations must obey $m_{n} \gtrsim 10^{4}~\mbox{TeV}$. Other sources of kaon mixing ---------------------------- A possible way to avoid the constraint is to assume that the Higgs field is localized on the UV brane. Hence, it does not have KK excitations at all, and the analysis of Section 4.1 does not apply. In this case, the major source of kaon mixing is the interaction of down-quarks with the KK excitations of the bulk gauge fields. For simplicity, let us consider the interaction with the bulk photons; exchange by the KK Z-bosons is treated in a similar way and yields analogous results. The relevant part of the 5D action is given by $$\label{actel} S_{5}^{\gamma}=e_5\int d^4 x dy \sqrt{g} \Bigl(\bar{Q}_{i}g^{MN}\Gamma_M A_N Q_{i}+\bar{d}_{i}g^{MN}\Gamma_M A_N d_{i}\Bigr)\; .$$ As usual, we expand the gauge field in the tower of KK modes, $$\label{klara} A^{\mu}(x,y)=\sum_{n=0}^{\infty} a^{\mu}_{n}(x)A_{n}(y)\; .$$ The zero mode of the bulk electromagnetic field is flat [@Rizzo; @Pomarol], $$\label{el0} A_{0}(y)=\frac{1}{\sqrt{\pi R}}\; .$$ Therefore, the 4D electric charge is $e=\frac{e_5}{\sqrt{\pi R}}$. The profiles of the KK excitations are [@Rizzo; @Pomarol] $$\label{eln} A_{n}(y)=N^{\gamma}_{n}e^{ky}\left[J_{1} \Bigl(\frac{m^{\gamma}_n}{k}e^{ky}\Bigr) +C_{n}Y_{1}\Bigl(\frac{m^{\gamma}_n}{k}e^{ky}\Bigr)\right]\; ,$$ where $m^{\gamma}_n$ denote the masses of the KK excitations. The normalization factor $N^{\gamma}_n$ is given by $$\label{gaugenormn} N^{\gamma}_{n}\approx \frac{m^{\gamma}_{n}}{\sqrt{k}} \frac{1}{\sqrt{{\int}^{\gamma_n}_0 sJ^2_{1}(s)ds}}\; ,$$ while the constant $C_n$ is $$C_{n}=-\frac{J_{1}\Bigl(\frac{m^{\gamma}_n}{k}\Bigr)+\frac{m^{\gamma}_n}{k}J'_{1}\Bigl(\frac{m^{\gamma}_n}{k}\Bigr)}{Y_{1}\Bigl(\frac{m^{\gamma}_n}{k}\Bigr)+ \frac{m^{\gamma}_n}{k}Y'_{1}\Bigl(\frac{m^{\gamma}_{n}}{k}\Bigr)}\; .$$ The masses of the KK excitations of the electromagnetic field are determined from the following eigenvalue equation [@Rizzo; @Pomarol] $$\frac{m^{\gamma}_{n}e^{k\pi R}}{k} =\gamma_{n}, \quad J_{1}\left(\frac{m^{\gamma}_{n}}{k}e^{k\pi R}\right)=0 \; .$$ Hereafter we omit the zero mode in the decomposition , since the interaction with the zero mode is universal in the flavor space and does not give rise to flavor violating processes. By inserting Eq.  into the 5D action and integrating the latter over the fifth coordinate, we arrive at the following effective 4D action: $$\label{effactel} S_{eff}^{\gamma}=\int d^4 x \Bigl(\sum_{n=1}^{\infty} b^{L}_{nij}\bar{d}_{Li} (x)\gamma_{\mu}d_{Lj}(x)a^{\mu}_n(x)+(L\leftrightarrow R)\Bigr)\; .$$ The constants $b^{L(R)}_{nij}$ are the effective 4D couplings of left (right) down-quarks with the $n$-th KK excitation. These couplings are obtained from the initial 5D theory, $$b^{L(R)}_{nij}=e_5 W^{L(R)}_{nij}\label{koshka}\; ,$$ where the constants $W^{L(R)}_{nij}$ are the overlap integrals of the appropriate wave functions, $$\label{klarik} W^{L}_{nij}=\delta_{ij} \int dy \sqrt{g}e^{ky}A_{n}(y)Q_{Li}(y)Q_{Lj}(y)$$ and the analogous expression for the constants $W^{R}_{nij}$. By performing the integration in Eq.  and using Eq. , we obtain $$b^{L}_{nij}=e_5 \delta_{ij}N_{Li}N_{Lj}N_{n}\frac{1}{m^{\gamma}_n}\Bigl(\frac{m^{\gamma}_n}{k}\Bigr)^{c_{Li}+c_{Lj}-1}\int^{\gamma_n}_0s^{1-c_{Li}-c_{Lj}}J_1(s)ds\; .$$ The expression for the couplings $b^{R}_{nij}$ is obtained by the replacing $N_{Li}\rightarrow N^{d}_{Ri}$ and $c_{Li}\rightarrow c^{d}_{Rj}$. As in the previous case, all RS suppression factors disappear for $i,j=1,2$. Thus, the couplings $b^{L(R)}_{n11}$ and $b^{L(R)}_{n22}$ are of the order of the 4D electromagnetic coupling $e$. The effective action is diagonal in the quark fields. The flavor violating terms appear upon the rotation of the quark fields by the matrices $A_{L}$ and $A_{R}$, $$S^{eff}_{K-\widetilde{K}}=\int d^{4}x \Bigl( \sum_{n=1}^{\infty} {b'}^{L}_{n12} \bar{d}_{L}(x)\gamma_{\mu}s_L(x)a^{\mu}_{n}(x)+ \sum_{n=1}^{\infty} {b'}^{L}_{n21} \bar{s}_{L}(x)\gamma_{\mu}d_L(x)a^{\mu}_{n}(x)+(L\leftrightarrow R)\Bigr)\; ,$$ where, according to Eqs. , , the couplings ${b'}^{L(R)}_{n12}$ and ${b'}^{L(R)}_{n21}$ are estimated as $$|{b'}^{L}_{n12}|\sim |{b'}^{L}_{n21}|\sim |b^{L}_{n11}-b^{L}_{n22}|\frac{N_{L1}}{N_{L2}}$$ and $$|{b'}^{R}_{n12}|\sim |{b'}^{R}_{n21}|\sim |b^{R}_{n11}-b^{R}_{n22}|\frac{N^{d}_{R1}}{N^{d}_{R2}}\; .$$ Note that the constants $b'^{L}_{n12}$ and $b'^{L}_{n21}$ are suppressed as compared to the initial ones $b^{L}_{n11}$ and $b^{L}_{n22}$, which is a consequence of the smallness of the factor $\frac{N_{L1}}{N_{L2}}$. Note also that the profiles of $d_{R}$ and $s_{R}$ can be chosen very similar to each other, so that $b^{R}_{n11}\approx b^{R}_{n22}$ and hence $b'_{n12}$ and $b'_{n21}$ can be made very small. For the warp factor $k\pi R=10$ and the parameters $c$ listed in Table \[muchacha\], we obtain $|{b'}^{L}_{n21}|\sim |{b'}^{L}_{n12}|\sim \frac{1}{50}$ and $|{b'}^{R}_{n12}|\sim |{b'}^{R}_{n21}|\sim \frac{1}{30}$. The exchange by the KK excitations of the electromagnetic field gives contribution to the coefficients $C_{1}$ and $\widetilde{C}_1$, $$\label{bon1} C_1=\sum_{n=1}^{\infty} \frac{({b'}^{L}_{n12})^2}{{m^{\gamma}_n}^2}\; ,\quad \widetilde{C}_1=\sum_{n=1}^{\infty} \frac{({b'}^{R}_{n12})^2}{{m^{\gamma}_n}^2}\; .$$ Their imaginary parts are constrained as follows [@Bona], $$-4.4 \times 10^{-15} ~\mbox{GeV}^{-2} \lesssim ~\mbox{Im} C_1, ~\mbox{Im} \widetilde{C}_{1}\lesssim 2.8 \times 10^{-15} ~\mbox{GeV}^{-2}\; .$$ These bounds imply the following constraint on the masses of the photon KK excitations: $$\label{constraint!!} m^{\gamma}_{n}\gtrsim 700~\mbox{TeV}\; .$$ We see that the masses of the KK excitations can be three orders of magnitude smaller than the ones in the bulk Higgs scenario. However, their values are still out of reach of future experiments. Besides the exchange by the KK modes, there are other sources of FCNC. They are important in the case of the IR-localized Higgs field, but subdominant in our scenario. One of these sources is the interaction of down-quarks with the zero mode of the $Z$-boson. The profile of the latter is not exactly flat [@Rizzo; @Pomarol]. However, the deviation from the flatness is of the order of the ratio $\frac{m^{2}_{Z}}{m^{2}_{KK}}$, where $m_{KK}$ is the typical mass scale of the KK excitations. Accordingly, flavor-violating vertices are suppressed by $\frac{m^{2}_{Z}}{m^{2}_{KK}}$. With $m_{KK}$ constrained by Eqs.  or , the contribution to the coefficient $C_1$ coming from the interaction with the zero mode of the Z-boson is negligibly small, $C_1\sim \frac{m^2_{Z}}{M^4_{KK}}$. As described in [@Azatov], there are also flavor violating processes mediated by the zero mode of the Higgs field. They occur due to the interaction with the KK excitations of fermions. These processes give negligibly small contribution to the coefficients $C$ for the same reason as above. Acknowledgements ================ The authors thank Valery Rubakov for useful discussions. This work was supported by Federal Agency for Science and Innovation of Russian Federation under state contracts 02.740.11.5194 (SM) and 02.740.11.0244 (MO and SR), by the Russian Foundation of Basic Research grant 08-02-00287 (SM), by the grant NS-5525.2010.2 (MO and SR), by the grants of the President of the Russian Federation MK-4317.2009.2 (MO) and MK-7748.2010.2 (SR), by Federal Agency for Education under state contract P520 (SR), by the Dynasty Foundation (SR). [99]{} J. M. Maldacena, Adv. Theor. Math. Phys. **2** (1998) 231 \[Int. J. Theor. Phys. **38** (1999) 1113\] \[arXiv:hep-th/9711200\].\[Maldacena\] S. S. Gubser, I. R. Klebanov, A. M. Polyakov, Phys. Lett. B **428** (1998) 105; \[arXiv:hep-th/9802109\].\[Gubser\] E. Witten, Adv. Theor. Math. Phys. **2** (1998) 253; \[arXiv:hep-th/9802150\].\[Witten\] L. Randall, R. Sundrum, Phys. Rev. Lett., **83** (1999) 3370; \[arXiv:hep-ph/9905221\].\[Sundrum\] T. Gherghetta, A. Pomarol, Nucl. Phys. B **586** (2000) 141; \[arXiv:hep-ph/0003129\].\[Gherghetta\] S. J. Huber and Q. Shafi, Phys. Lett. B **498** (2001); \[arXiv:hep-ph/0010195\].\[Huber\] S. Huber, Nucl. Phys. B **666** (2003) 269; \[arXiv:hep-ph/0303183\].\[Stephan\] Y. Grossman and M. Neubert, Phys. Lett. B\[Grossman\] **474** (2000) (361); \[arXiv:hep-ph/9913408\]. S. Chang, C. S. Kim, M. Yamaguchi, Phys. Rev. D (2006) 033002; \[arXiv:hep-ph/0511099\].\[Chang\] S. Casagrande, F. Goertz, U. Haisch, M. Neubert, T. Pfoh, JHEP **0810** (2008) (094); \[arXiv:hep-ph/0807.4937\].\[Casagrande\] M. Bauer, S. Casagrande, U. Haish, M. Neubert; \[arXiv:hep-ph/0912.1625\].\[Neubert\] K. Agashe, G. Perez, A. Soni, Phys. Rev. Lett. **93** (2004) 201804; \[arXiv:hep-ph/0406101\].\[Perez\] K. Agashe, G. Perez and A. Soni, Phys. Rev. D **71** (2005) 016002; \[arXiv:hep-ph/0408134\].\[Agashe\] C. Csaki, A. Falkowski and A. Weiler, Phys. Rev. D **80** (2009) 0160001; \[arXiv:hep-ph/0806.3757\].\[Csaki\] M. Blanke, A. J. Buras, B. Duling, S. Gori and A. Weiler, JHEP **0903** (2009) (001); \[arXiv:hep-ph/0809.1073\].\[Blanke\] A. Fitzpatrick, G. Perez, L. Randall, Phys. Rev. Lett. **100** (2008) 171604; \[arXiv:hep-ph/0710.1869\].\[Randall\] J. Santiago, JHEP **0812** (2008) (046); \[arXiv:hep-ph/0806.1230\].\[Santiago\] R. Contino and A. Pomarol, JHEP **0411** (2004) 058; \[arXiv:hep-th/0406257\].\[Contino\] T. Gherghetta, \[arXiv:hep-ph/0601213\]; Les Houches 2005 Session LXXXIV lectures. N. Arkani-Hamed, M. Porrati, L. Randall, JHEP **0108** (2001) (017); \[arXiv:hep-th/0012148\].\[Arkani\] C. Amsler et al. \[Particle Data Group\], Phys. Lett. B **667** (2008) 1.\[pdg\] M. Bona et al., JHEP **0803** (2008) (049); \[arXiv:hep-ph/0707.0636\].\[Bona\] H. Davoudiasl, J. L. Hewett, T. G. Rizzo, Phys. Lett. B **473** (2000) 43; \[arXiv:hep-ph/9911262\].\[Rizzo\] A. Pomarol, Phys. Lett. B **486** (2000) (153); \[arXiv:hep-ph/9911294\].\[Pomarol\] A. Azatov, M. Toharia, L. Zhu, Phys. Rev. D **80** (2009) 035016; \[arXiv:hep-ph/0906.1990\].\[Azatov\] [^1]: One can show that the terms omitted in are negligible.

ArXiv

--- abstract: 'We study the dynamics of dissipative spins for general spin-spin coupling. We investigate the population dynamics and relaxation of the purity in the white noise regime, in which exact results are available. Inter alia, we find distinct reduction of decoherence and slowdown of purity decay around degeneracy points. We also determine in analytic form the one-phonon exchange contribution to decoherence and relaxation in the ohmic quantum noise regime valid down to zero temperature.' address: 'II. Institut für Theoretische Physik, Universität Stuttgart, D-70550 Stuttgart, Germany.' author: - Peter Nägele - Ulrich Weiss title: 'Dynamics of coupled spins in the white- and quantum-noise regime' --- and Dissipative Coupled Spins ,Purity ,Generalized Spin-Boson Model ,White Noise ,Quantum Noise Introduction ============ The spin-boson model is a key model since the 80’s for the quantitative study of decoherence, relaxation and energy dissipation [@leggett; @bookweiss]. In view of the substantial progress in fabrication of coupled qubit devices [@pashkin_quantum_2003; @clarke] and major advances in quantum state manipulation towards quantum computation [@nielsen-chuang], there is growing interest in the accurate calculation of the dynamics of coupled spins for realistic environmental couplings. Here we communicate new analytical results for the dynamics of two mutually interacting spins each liable to independent white noise or to quantum noise forces. Previous work [@naegele] is extended in three different directions. First, we determine the dynamics of the purity in the white noise regime (WNR). Second, we study the dynamics near two different degeneracy points and find striking reduction of purity decay and decoherence. Third, we calculate the exact one-phonon contribution to dephasing and relaxation in the quantum noise regime in analytic form. The resulting expressions hold down to zero temperature. The major results are obtained within the real-time path sum method for the 16 states of the two-spin density matrix. The environmental couplings are included via the Feynman-Vernon method. Here we omit methodical and technical aspects. We rather put emphasis on results in analytic form and their physical implications In section \[smodel\] the model of two interacting spins each coupled to its own dissipative environment is introduced. An analytic analysis of the white noise regime, inter alia study of purity and degeneracy points, is given in section \[swnr\], while section \[sqnr\] deals with the effects of weak quantum noise. Model {#smodel} ===== We consider two two-state systems or spins which are mutually coupled via Ising, XY, and/or Heisenberg coupling. In addition, each of them is coupled to its own heat bath. In pseudospin representation, the two-spin-boson Hamiltonian reads (we put $\hbar =k_{\rm B}^{} =1$) $$\label{ham1} H = H_\mathrm{SS} + H_\mathrm{SR} \; .$$ Here, $H_\mathrm{SS}$ represents the interacting two spins, $$\label{ham2} \begin{array}{rcl} H_\mathrm{SS} &=& - {\textstyle \frac{1}{2}}\Delta_1\, \sigma_x - {\textstyle \frac{1}{2}} \Delta_2 \, \tau_x - {\textstyle \frac{1}{2}} \epsilon_1 \,\sigma_z - {\textstyle \frac{1}{2}} \epsilon_2\, \tau_z \, \\ && -\; {\textstyle \frac{1}{2}} v_x \,\sigma_x \tau_x - {\textstyle \frac{1}{2}} v_y \,\sigma_y \tau_y - {\textstyle \frac{1}{2}} v_z\,\sigma_z \tau_z \; . \end{array}$$ In the basis formed by the localized states $|R\!\!>$ and $|L\!\! >$, the parameters $\epsilon_{1,2}^{}$ and $\Delta_{1,2}$ represent the bias energies and tunneling couplings of the $\sigma$- and $\tau$-spin, and $v_{x,y,z}$ are the interaction parameters. The term $H_\mathrm{SR}$ describes the spin-reservoir couplings and the reservoirs, $$H_\mathrm{SR} = -\,\frac{1}{2}\sigma_z X_1 \,-\,\frac{1}{2}\tau_z X_2 + \sum_{\zeta=1,2}\sum_\alpha \omega_{\zeta,\alpha}^{} b_{\zeta,\alpha}^\dag b_{\zeta,\alpha}^{} \; .$$ Here, $X_\zeta(t) = \sum_{\alpha} c_{\zeta,\alpha}^{} [\,b_{\zeta,\alpha}^{}(t)\,+\,b_{\zeta,\alpha}^\dag(t)\,] \;(\zeta = 1,2)$ is a collective reservoir mode. All effects of the environment are carried by the power spectrum of the collective bath modes. We have $S_{\zeta,\zeta'}^{}(\omega) = \delta_{\zeta,\zeta'}^{} S_\zeta^{}(\omega)$, where $$\label{powerspec} \begin{split} S_{\zeta}^{}(\omega) \,&=\, {\rm Re}\! \int_{-\infty}^\infty\!\!\!{\rm d}t\, {{\rm e}}^{{{\rm i}}\omega t}_{} \! \left\langle X_{\zeta}^{}(t)X_{\zeta}^{}(0)\,\right\rangle_{\!\beta} \\ & =\, \pi G_\zeta^{}(\omega)\coth\Big(\!\frac{\omega}{2 T}\!\Big) \; . \end{split}$$ The spectral density of the coupling is [@leggett; @bookweiss] $$G_\zeta(\omega)\;=\; \sum_\alpha c^2_{\zeta\!,\,\alpha} \delta(\omega - \omega_{\zeta\!,\,\alpha} ) \;=\; 2 K_\zeta\, \omega \,{\rm e}^{-|\omega| /\omega_{\rm c}^{} }_{} \; .$$ The second form describes ohmic coupling with high-frequency cut-off $\omega_{\rm c}^{}$ and dimensionless damping constant $K_\zeta$. Instead of the independent baths, one might also choose a common bath for the two spins [@wilhelm]. In the sequel, we confine ourselves to the case of $v_y^{}$- and $v_z^{}$-coupling of the two spins. This case is most interesting concerning application to coupled Josephson junctions. In addition, we disregard the bias terms. The Hamiltonian $H_\mathrm{SS}$ is diagonalized with the unitary matrix $$U=\frac{1}{2}\left( \begin{array}{llll} \frac{\cos \left(\phi _1\right)}{f_+(\phi_1)} & f_+(\phi_1) & f_+(\phi_1) & \frac{\cos \left(\phi _1\right)}{f_+(\phi_1)} \\[2mm] -\frac{\cos \left(\phi _2\right)}{f_-(\phi_2)} & f_-(\phi_2) & - f_-(\phi_2) & \frac{\cos \left(\phi _2\right)}{f_-(\phi_2)} \\[2mm] -\frac{\cos \left(\phi _2\right)}{f_+(\phi_2)} & -f_+(\phi_2) & f_+(\phi_2) & \frac{\cos \left(\phi _2\right)}{f_+(\phi_2)} \\[2mm] \frac{\cos \left(\phi _1\right)}{f_-(\phi_1)} & -f_-(\phi_1) & -f_-(\phi_1) & \frac{\cos \left(\phi _1\right)}{f_-(\phi_1)} \end{array} \right) ,$$ with $f_{\pm}(\phi) = \sqrt{1 \pm \sin \phi}\; $ and mixing angles $\phi_{1,2} = \arctan\left[ (v_y \mp v_z)/(\Delta_1\pm\Delta_2) \right]$. We then get $$\widetilde H = U\,H\,U^{-1} = -\frac{\Omega}{2} \,(\sigma_z \otimes \mathds{1}) - \frac{\delta}{2} \, (\mathds{1} \otimes \tau_z) ,$$ with the eigen frequencies $$\label{barefreq} \begin{split} \Omega &= {\textstyle \frac{1}{2}}( \Omega_{+} + \Omega_{-} )\, , \qquad \, \delta = {\textstyle \frac{1}{2}} ( \Omega_{+} - \Omega_{-} ) \, , \\[1mm] \Omega_{\pm} &= \sqrt{(\Delta_1 \pm \Delta_2)^2 + (v_y \mp v_z)^2}\; . \end{split}$$ We have the Vieta relations $$\begin{split} \Omega_{+}^{2} + \Omega_{-}^{2} &= 2\,( \Delta_1^2 + \Delta_2^2 + v_y^2 + v_z^2 ) \; , \\ \Omega^2+\delta^2 &= \Delta_1^2 + \Delta_2^2 + v_y^2 + v_z^2 \; ,\\ \Omega^2 \, \delta^2 &= (v_y\,v_z - \Delta_1\,\Delta_2)^2\; . \end{split}$$ The two-spin density matrix has 16 matrix elements. They can be expressed as linear combinations of the unit matrix and the following 15 expectation values, $\langle \sigma_i\otimes \mathbf{1} \rangle_t \equiv \langle \sigma_i\rangle_t$, $\langle \mathbf{1} \otimes \tau_i \rangle_t \equiv \langle \tau_i\rangle_t $, and $\langle \sigma_i \otimes \tau_j \rangle_t \equiv \langle \sigma_i \tau_j \rangle_t$ ($i=x,\,y,\,z$ and $j=x,\,y,\,z$). These quantities, denoted by $W_j(t)$ ($j=1,\cdots,\,15$), obey the equations of motion $\dot W_j(t) = \,\mathrm{i}\, [\,H_\mathrm{SS},\,W_j(t)\,]$ ($j=1,\cdots,15$). The set of coupled equations are conveniently solved in Laplace space $\lambda$. Throughout we choose the initial state $\langle \sigma_z\rangle_0 =1$, $\langle \tau_z\rangle_0^{}=1$, and $\langle \sigma_z \tau_z \rangle_0^{}=1$, and all other expectations zero. The resulting expressions may be written as $$\label{fracexpr} W_j(\lambda) = N_j^{}(\lambda)/D_j^{}(\lambda) \; ,\quad j=1,\cdots,\,15 \; ,$$ where the $N_j^{}(\lambda)$ are different for the individual $W_j(\lambda)$, while the $D_j^{}(\lambda)$ fall into the following two categories, $$\begin{array}{rcl} D_{\Omega,\delta}(\lambda) &=& (\lambda^2 +\Omega^2)(\lambda^2+\delta^2) \;,\\ D_{\Omega_\pm}(\lambda) &=& \lambda(\lambda^2 +\Omega^2_+)(\lambda^2+\Omega^2_-) \; . \end{array}$$ White-noise regime {#swnr} ================== General features and qualitative behavior {#ssgf} ----------------------------------------- The exact formal solution of the dissipative two-spin dynamics has been discussed in Ref. [@naegele]. Explicit expressions for the $W_j(\lambda)$ have been given in the white noise regime (WNR), in which (\[powerspec\]) reduces to $$\label{wnl} S_\zeta(\omega \ll T ) = 2\vartheta_\zeta\;, \quad\mbox{where}\quad \vartheta_\zeta = 2 \pi K_\zeta T \; .$$ It was found that the form (\[wnl\]) is expedient in the regime $K_\zeta{\,\raisebox{-.5ex} {$\stackrel{\raisebox{-.5pt}{$\textstyle <$}}{\sim}$}\,}0.3$ and $\Omega_\pm{\,\raisebox{-.5ex} {$\stackrel{\raisebox{-.5pt}{$\textstyle <$}}{\sim}$}\,}T\ll\omega_{\rm c}$, which covers not only the incoherent regime but also a sizeable domain of the coherent regime. The analysis shows that reservoir modes in the range $2\pi T < \omega <\omega_{\rm c}^{}$ give rise to an adiabatic (Franck-Condon-type) renormalization of the tunneling coupling, $\Delta_\zeta^2 \to\bar\Delta_\zeta^2 = (2\pi T/\Delta_{\zeta,{\rm r}})^{2K_\zeta}_{}\Delta_{\zeta,{\rm r}}^2$ with $\Delta_{\zeta,{\rm r}}^{1-K_\zeta}=\Delta_\zeta/\omega_{\rm c}^{K_\zeta}$ [@bookweiss; @naegele]. In the reminder of this section, we assume that the $\Delta_{1,2}$ are the renormalized ones. The modes with $\omega<2\pi T$ lead to decoherence and relaxation. In the WNR, they are accounted for by an appropriate shift of the Laplace variable $\lambda$ in the time interval, in which spin $\zeta$ dwells in an off-diagonal state. We have ($\zeta=1,2$) $$\lambda\to\lambda_\zeta = \lambda + \vartheta_\zeta\;,\quad\mbox{and}\quad \lambda \to\lambda_{12} = \lambda + \vartheta_1 + \vartheta_2 \; .$$ The resulting 15 coupled equations are $$\begin{array}{rcl} \lambda\, \langle\sigma _z\rangle &=& 1 - \Delta _1 \langle\sigma _y\rangle - v_x \langle\tau _x \sigma _y\rangle + v_y \langle\sigma _x \tau _y\rangle \, , \\ \lambda_1\, \langle\sigma _y\rangle &=& \Delta _1 \langle\sigma _z\rangle + v_x \langle\sigma _z \tau _x\rangle - v_z \langle\tau _z \sigma_x\rangle \, , \\ \lambda_1\, \langle\sigma _x\rangle &=& - v_y \langle\sigma _z \tau _y\rangle + v_z \langle\sigma _y \tau _z\rangle \, , \\ \lambda\, \langle\tau_z\rangle &=& 1 - \Delta _2 \langle\tau _y\rangle - v_x \langle\sigma _x \tau _y\rangle + v_y \langle\sigma _y \tau _x\rangle \, , \\ \lambda_2\, \langle\tau _y\rangle &=& \Delta _2 \langle\tau _z\rangle + v_x \langle\sigma _x \tau_z\rangle - v_z \langle\sigma _z \tau _x\rangle \, , \\ \lambda_2\, \langle\tau _x\rangle &=& - v_y \langle\sigma _y \tau _z\rangle + v_z \langle\sigma _z \tau _y\rangle \, , \\ \lambda_{12}\, \langle\sigma _x \tau _y\rangle &=& \Delta _2 \langle\sigma _x \tau _z\rangle +\, v_x \langle\tau _z\rangle - v_y \langle\sigma _z\rangle \, , \\ \lambda_{12}\, \langle\sigma _y \tau_x\rangle &=& \Delta _1 \langle\tau _x \sigma _z\rangle +\, v_x \langle\sigma _z\rangle - v_y \langle\tau _z\rangle \, , \\ \lambda_2\langle\sigma _z \tau _y\rangle &=& \Delta _2 \langle\sigma _z \tau _z\rangle \!-\! \Delta _1 \langle\sigma _y \tau_y\rangle \!+\! v_y \langle\sigma _x\rangle \!-\! v_z \langle\tau _x\rangle \, , \\ \lambda_1 \langle\sigma _y \tau _z\rangle &=& \Delta _1 \langle\sigma _z \tau _z\rangle \!-\! \Delta _2 \langle\sigma _y \tau _y\rangle \!+\! v_y \langle\tau_x\rangle \! -\! v_z \langle\sigma _x\rangle \, , \\ \lambda_1\, \langle\sigma _x \tau _z\rangle &=& - \Delta _2 \langle\sigma _x \tau _y\rangle - v_x \langle\tau _y\rangle + v_z \langle\sigma _y\rangle \, , \\ \lambda_2\, \langle\sigma _z \tau _x\rangle &=& - \Delta _1 \langle\tau_x \sigma _y\rangle - v_x \langle\sigma _y\rangle + v_z \langle\tau _y\rangle \, , \\ \lambda_{12}\, \langle\sigma _y \tau _y\rangle &=& \Delta _1 \langle\sigma _z \tau _y\rangle + \Delta _2 \langle\sigma_y \tau _z\rangle\, , \\ \lambda\, \langle\sigma _z \tau _z\rangle &=& 1 - \Delta _1 \langle\sigma _y \tau _z\rangle - \Delta _2 \langle\sigma _z \tau _y\rangle\, , \\ \lambda_{12}\, \langle\sigma _x \tau _x\rangle &=& 0 \, . \end{array}$$ The solutions for all $W_j(\lambda)$ are again in the form (\[fracexpr\]). There arise only three different denominators, namely $$\begin{split} D_1(\lambda) &= (\lambda\lambda_1+\Delta_1^2)(\lambda_{1}\lambda_{12} +\Delta_2^2) + v_y^2(\lambda_1^2+v_z^2 ) \\ &\quad+ \lambda \lambda_{12} v_z^2 - 2 v_y v_z \Delta_1\Delta_2 \; , \\[1mm] D_2(\lambda) &= (\lambda\lambda_2+\Delta_2^2)(\lambda_{2}\lambda_{12} +\Delta_1^2) + v_y^2(\lambda_2^2+v_z^2 ) \\ &\quad+ \lambda \lambda_{12} v_z^2 - 2 v_y v_z \Delta_1\Delta_2 \; , \\[1mm] D_3(\lambda) &= \lambda\lambda_{12} [\, \lambda_1^2\lambda_2^2 + v_z^2(\lambda_1^2+\lambda_2^2 ) + 2 v_y^2 \lambda_1\lambda_2 \, ] \\ &+ (v_y^2 \Delta_2^2 + v_z^2\Delta_1^2 + \Delta_2^2 \lambda_1\lambda_2 ) (\lambda_1\lambda_{12}+ \lambda\lambda_2 ) \\ &+ (v_y^2 \Delta_1^2 + v_z^2\Delta_2^2 + \Delta_1^2 \lambda_1\lambda_2 ) (\lambda_2\lambda_{12}+ \lambda\lambda_1 ) \\ &+ (v_y^2 - v_z^2)^2 \lambda\lambda_{12} + (\Delta_1^2-\Delta_2^2)^2 \lambda_1\lambda_2 \\ &+ 2v_y v_z \Delta_1\Delta_2 (\lambda+\lambda_{12})(\lambda_1+\lambda_2 ) \, . \end{split}$$ For instance, we obtain $$\label{undamp} \begin{split} \langle \sigma_z (\lambda) \rangle &= \frac{\lambda_1 \left( \Delta_2^2 + \lambda_1 \lambda_{12} \right) + v_z^2 \lambda_{12}}{D_1(\lambda)} \;, \\[1mm] \langle \tau_z (\lambda) \rangle &= \frac{\lambda_2 \left( \Delta_1^2 + \lambda_2 \lambda_{12} \right) + v_z^2 \lambda_{12}}{D_2(\lambda)} \; ,\\[1mm] \langle \sigma_y\tau_y (\lambda) \rangle &= \big[\,\Delta_1\Delta_2(v_y^2\!+\! v_z^2+\lambda_1\lambda_2)\\ &\quad \;\;+ v_y v_z(\Delta_1^2\!+\! \Delta_2^2)\, \big] \,\frac{\lambda_1+\lambda_2}{ D_3(\lambda)} \; . \end{split}$$ For vanishing bath coupling, both $D_{1}(\lambda)$ and $D_{2}(\lambda)$ go to $D_{\Omega,\delta}(\lambda)$, and $D_3(\lambda)\to \lambda$ reduces to $D_{\Omega_\pm}(\lambda)$. The dynamics of the expectations $A_j(t)$ is mainly determined by the zeros $\lambda_i$ of $D_k(\lambda)$ $(k=1,\,2,\,3)$. They appear in complex conjugate or real pairs. We get $$A_j(t) \;=\; \sum_{i=1}^n B_i\,{\rm e}^{\lambda_i t}_{}\;,$$ where $n$ is either 4 or 6. The behaviors of the four $\lambda_j$ of $\langle\sigma_z(\lambda)\rangle$ (and of $\langle\tau_z(\lambda)\rangle$) and the six $\lambda_j$ of $\langle\sigma_y\tau_y(\lambda)\rangle$, and the respective amplitudes $B_j$, are quite diversified. In Fig. \[fig:realim\] we show plots of the real parts (rates) and imaginary parts (oscillation frequencies) of the four $\lambda_j$ of $\langle\sigma_z(\lambda)\rangle$ as functions of $\vartheta$ for a particular set of parameters (identical spins, $\Delta_{1,2} =\Delta,\,\vartheta_{1,2}=\vartheta$). In the coupling range $v_z < v_z^\ast$ there are three crossover temperatures $\vartheta_1^\ast$, $\vartheta_2^\ast$ and $\vartheta_3^\ast$ at which the discriminant of $D_1(\lambda)$ is zero. For $v_z > v_z^\ast$ there is only a single crossover temperature $\vartheta_3^\ast$. The critical coupling strength is $$v_z^\ast \;=\; {\textstyle \frac{1}{2}}\,\big[\, (2\,\Delta^2+v_y^2)^{1/2} -v_y \,\big] \; .$$ A plot of the crossover temperatures is shown in Fig. \[fig:sz\_crosstemp\]. The particular case $\Delta_1=\Delta_2$ and $v_z= -v_y$ is a degeneracy point, $\Omega=\delta$ (cf. subsection \[sdegp\]). In this case, the crossover curves $\vartheta_1^\ast(v_z)$ and $\vartheta_2^\ast(v_z)$ coincide. In the regime $\vartheta < \vartheta_2^\ast$ the dynamics is coherent, described by a superposition of two damped oscillations with amplitudes of comparable size. For $\vartheta < \vartheta_1^\ast$, the two oscillations have different frequency, $\Omega > \delta$, but the same damping rate. In the range $\vartheta_1^\ast < \vartheta < \vartheta_2^\ast$, they have the same frequency, but different damping rates. In the range $\vartheta > \vartheta_2^\ast$, the dynamics is incoherent with 4 different relaxation rates. The two smallest rates have sizeable amplitudes and thus dominate the relaxation dynamics. In the so-called Kondo regime $\vartheta >\vartheta_3^\ast$, three of the four $\lambda_j$-contributions have negligibly small amplitudes. The only relevant contribution is that with the smallest rate. The Kondo characteristics is that this rate decreases with increasing temperature, $\gamma_{\rm K} =\Delta^2/\vartheta$ (cf. Fig. \[fig:realim\]). This counter-intuitive feature is already well-known from the ohmic single-spin model [@bookweiss]. ![$\langle\sigma_z\rangle$: Plots of ${\rm Re}\,\lambda_j$ and ${\rm Im}\,\lambda_j$ against $\vartheta=\vartheta_{1,2}$. The parameters are $\Delta_{1,2}=1$, $v_z=0.4$, $v_y=0.1$. []{data-label="fig:realim"}](re.eps "fig:"){width="5.2" height="3.5"} ![$\langle\sigma_z\rangle$: Plots of ${\rm Re}\,\lambda_j$ and ${\rm Im}\,\lambda_j$ against $\vartheta=\vartheta_{1,2}$. The parameters are $\Delta_{1,2}=1$, $v_z=0.4$, $v_y=0.1$. []{data-label="fig:realim"}](im.eps "fig:"){width="5.4" height="3.5"} ![$\langle\sigma_z^{}\rangle$: Crossover temperatures against $v_z$ for $v_y=0.1$ and identical spins, $\Delta_{1,2}=1$, and $\vartheta_{1,2}=\vartheta$.[]{data-label="fig:sz_crosstemp"}](cross_temp.eps){width="5.0" height="3.5"} Low temperature WN regime {#sys_weak_vy_vz} ------------------------- The WN regime has a low temperature bound roughly given by $T \approx\Omega_{\pm}$. Above this bound and below the first crossover temperature, $\Omega_\pm {\,\raisebox{-.5ex} {$\stackrel{\raisebox{-.5pt}{$\textstyle <$}}{\sim}$}\,}T <T_1^\ast$, the real parts of the $\{\lambda_j\}$ vary linearly with temperature. In this regime, systematic low-temperature expansion of the zeros of $D_j(\lambda)$ ($j=1,\,2,\,3$) is straightforward. The results are as follows:\ $\langle \sigma_z\rangle_t$: There is a superposition of two damped oscillations, $\lambda_{1,2}^{} = \pm{\rm i}\,\Omega -\gamma_{\Omega}^{}$ and $\lambda_{3,4}^{} = \pm{\rm i}\,\delta -\gamma_{\delta}^{}$. The frequencies $\Omega$ and $\delta$ are close to their bare values given in Eq. (\[barefreq\]) near $T=\Omega_\pm$. As $T$ is increased, they approach each other and coincide at $T=T_1^\ast$. The respective damping rates $\gamma_\Omega^{}$ and $\gamma_\delta^{}$ in the regime $\Omega_\pm<T < T_1^\ast$ read $$\label{ratesz} \begin{split} \gamma_\Omega^{} &= \frac{1}{2}\vartheta_1 +\frac{\Omega^2 - \Delta_1^2- v_y^2 }{2(\Omega^2_{}-\delta^2_{})}\vartheta_1 + \frac{\Omega^2-\Delta_1^2-v_z^2}{2(\Omega^2_{}-\delta^2_{})}\vartheta_2 \, , \\[2mm] \gamma_\delta^{} &= \frac{1}{2}\vartheta_1 + \,\frac{\Delta_1^2 + v^2_{y} - \delta^2}{2(\Omega^2_{}-\delta^2_{})}\vartheta_1 + \, \frac{\Delta_1^2 + v_z^2 - \delta_{}^2}{2(\Omega^2_{}-\delta^2_{})} \vartheta_2 \, . \end{split}$$ The amplitudes of the oscillatory contributions (in zeroth order in $\vartheta_{1,2}$) read[^1] $$B_\Omega = \frac{\Omega^2 -v_z^2-\Delta_2^2}{ 2(\Omega^2-\delta^2)}\;,\quad B_\delta = \frac{\Delta_2^2 +v_z^2-\delta^2}{ 2 (\Omega^2-\delta^2)} \;.$$ Mutual exchanges $\Delta_{1}\leftrightarrow\Delta_{2}$ and $\vartheta_{1}\leftrightarrow\vartheta_{2}$ yield the corresponding rates and amplitudes for $\langle\tau_z^{}\rangle_t$.\ $\langle \sigma_z\tau_z \rangle_t$: According to the zeros of $D_3(\lambda)$, $\lambda_{1,2} = \pm{\rm i}\,\Omega_+ - \gamma_{\Omega_+}^{}$, $\lambda_{3,4} = \pm{\rm i}\,\Omega_- - \gamma_{\Omega_-}^{}$, $\lambda_{5,6}^{} = -\gamma_{5,6}^{}$, there are two damped oscillatory and two relaxation contributions. The frequencies $\Omega_\pm$ are close to their bare values given in Eq. (\[barefreq\]) near $T=\Omega_\pm$, and they coincide at the first crossover temperature $T_1^\ast$. The damping rates $\gamma_{\Omega_\pm}^{}$ and amplitudes $B_{\Omega_\pm}$ of the oscillations are $$\label{ratesztz} \begin{split} \gamma_{\Omega_+}^{} &\;=\; \gamma_{\Omega_-}^{} = {\textstyle\frac{1}{2} }(\vartheta_1+\vartheta_2)\, , \\[1mm] B_{\Omega_+} &\;=\; \frac{(\Delta_1+\Delta_2)^2}{4\Omega_+^2} \;,\quad B_{\Omega_-} = \frac{(\Delta_1 - \Delta_2)^2}{4\Omega_-^2} \;. \end{split}$$ The relaxation rates $\gamma_{5,6}^{}$ are determined by a quadratic equation, which is obtained by truncation of $D_3(\lambda)$. The resulting expressions for the relaxation rates and associated amplitudes are $$\label{ratesztz1} \begin{split} \gamma_{5}^{} &\;=\; \frac{\Omega^2 -\Delta_1^2 -v_y^2}{\Omega^2-\delta^2}\,\vartheta_1 + \frac{\Omega^2 -\Delta_2^2 -v_y^2}{\Omega^2-\delta^2}\,\vartheta_2 \, , \\[1mm] \gamma_{6}^{} &\;=\; \frac{\Delta_1^2 +v_y^2-\delta^2}{\Omega^2-\delta^2}\,\vartheta_1 \;+\; \frac{\Delta_2^2 +v_y^2 - \delta^2}{\Omega^2-\delta^2}\,\vartheta_2 \, , \end{split}$$ $$B_{\gamma_5^{}} = \frac{(v_y\Omega + v_z\delta)^2}{(\Omega^2-\delta^2)^2}\;, \quad\;\; B_{\gamma_6^{}} = \frac{(v_y\delta + v_z\Omega)^2}{(\Omega^2-\delta^2)^2}\; .$$ ### The limit $\Delta_2\to 0$ and $v_y\to 0$ {#sdegen} In the limit $\Delta_2\! \to \! 0$ and $v_y\!\!\to\!\! 0$, the transition frequencies $\Omega_\pm$ become degenerate, and hence $\Omega=\sqrt{\Delta_1^2 +v_z^2}$ and $\delta =0$. Accordingly, the expression (\[ratesz\]) for $\gamma_\delta^{}$ is not valid anymore. To cope with this limiting case, we must determine $\lambda_{3,4}$ from a quadratic equation, which is found from $D_1(\lambda) = 0 $ by reduction. The respective complex eigenvalues for slight detuning are found as $$\lambda\; = \; \pm\, \mathrm{i} \,\sqrt{\delta^2- \frac{1}{4}\big( \frac{v_z^2}{\Omega^2}\vartheta_1 + \vartheta_2\big)^2\,} - \frac{\Omega^2+\Delta_1^2}{2\Omega^2}\vartheta_1 - \frac{\vartheta_2}{2}\; .$$ From this form we see that the two complex conjugate eigenvalues turn into two real eigenvalues, when $\delta$ is sufficiently small. At $\delta=0$, the rate expressions are $$\gamma_+^{} = \vartheta_1+\vartheta_2\, ,\quad \mbox{and}\quad \gamma_-^{} = \gamma_{\rm r}^{} := \frac{\Delta_1^2}{\Omega^2}\,\vartheta_1 \, .$$ In addition, the analysis shows that the residuum associated with the pole at $\lambda=-\gamma_+^{}$ is zero, while the other yields the amplitude $B_{\gamma_-^{}}^{} = \frac{1}{2}v_z^2/\Omega^2$. At this point, we remark that, in the limit $\Delta_2\to 0$, the coupling $v_z$ takes the role of a biasing energy for spin $\sigma$. Thus, the dissipative two-spin model reduces to the standard biased spin-boson model. The rate $\gamma_{\rm r}^{}$ is just the relaxation rate of this model. Furthermore, the rate $\gamma_\Omega^{}$ in Eq. (\[ratesz\]) reduces to the form $$\gamma_\Omega^{} \;=\; {\textstyle \frac{1}{2}}\gamma_{\rm r}^{}\,+\, (v_z^2/\Omega^2)\,\vartheta_1 \; .$$ This expression coincides indeed with the decoherence rate of the biased spin-boson model in the WNR [@bookweiss]. Purity {#spur} ------ For a system described by the density matrix $\rho(t)$, the purity $P(t) :=\mathrm{Tr}\rho^2(t)$ tells us whether the system is in a pure state or in a mixture. For a pure state, there is $P=1$ while for a fully mixed state $P=1/N$. Here $N$ ist the number of the system’s accessible states. In the low temperature WN regime discusssed in the preceeding subsection, the purity $P(t)$ is found as (the index $\sigma,\tau$ refers to the respective spin) $$\begin{split} &P(t) = {\textstyle \frac{1}{4} } + {\textstyle \frac{1}{8}}(1-C_\sigma )\,{\rm e}^{- 2\gamma_{\delta,\sigma}^{} t}_{} + {\textstyle \frac{1}{8}}(1+C_\sigma )\,{\rm e}^{- 2\gamma_{\Omega,\sigma}^{} t}_{} \\ &\qquad + {\textstyle \frac{1}{8}}(1-C_\tau )\,{\rm e}^{- 2\gamma_{\delta,\tau}^{} t}_{} + {\textstyle \frac{1}{8}}(1+C_\tau )\,{\rm e}^{- 2\gamma_{\Omega,\tau}^{} t}_{} \\ &\qquad + {\textstyle \frac{1}{8}}(1-C_{\Omega_-}^{} )\,{\rm e}^{- 2\gamma_{\Omega_-}^{} t}_{} + {\textstyle \frac{1}{8}}(1-C_{\Omega_+}^{} )\,{\rm e}^{- 2\gamma_{\Omega_+}^{} t}_{} \\ &\qquad + {\textstyle \frac{1}{4} }\frac{(v_y\Omega + v_z\delta)^2}{(\Omega^2-\delta^2)^2}\,{\rm e}^{-2\gamma_5^{}t}_{} + {\textstyle \frac{1}{4} }\frac{(v_y\delta + v_z\Omega)^2}{(\Omega^2-\delta^2)^2}\,{\rm e}^{-2\gamma_6^{}t}_{} \end{split}$$ with the amplitudes $$C_{\sigma,\tau}^{} = \frac{\Delta_{1,2}^2-\Delta_{2,1}^2 + v_y^2-v_z^2}{\Omega^2 -\delta^2}\, ,\quad C_{\Omega_\pm^{}}^{} = \frac{(v_y \mp v_z)^2}{\Omega_\pm^{2}} \, .$$ This function smoothly drops on the time-scale given by the system’s damping and relaxation rates from the initial value $P(0)=1$ to the fully mixed thermal equilibrium state, $P(t\to\infty) =\frac{1}{4}$. Observe that all dephasing and relaxation rates relevant to the decay of the expectation values $W_j(t)$ ($j=1,\,\cdots,\,15$) contribute to the decay of the purity. Decoherence dip near degeneracy points {#sdegp} -------------------------------------- Of particular interest are degeneracy points of the two-spin system. There are two different cases: $$\begin{split} \Delta_1=\Delta_2\, ,\quad v_y = -v_z\, \!\quad\! &\Longrightarrow \!\quad\! \;\;\;\, \Omega=\delta\; \;\;\;\,\quad\mbox{(case $I$)}\, ,\\[2mm] v_y v_z = \Delta_1\Delta_2\, \;\;\qquad &\Longrightarrow\quad \Omega_+ =\Omega_- \, \quad\mbox{(case $II$)}\, . \end{split}$$ For comparison, we also study the nondegenerate point conjugate to case $I$ $$\Delta_1=\Delta_2\, , \!\quad \!v_y = v_z\,\quad\! \Longrightarrow\quad\! \Omega,\delta=\Delta \pm v\,\quad(\mbox{case}\;\; I^\ast) \, .$$ Consider first $\langle\sigma_z\rangle_t$ in case $I$. We find from Eq. (\[ratesz\]) upon taking the limits $\Delta_{1,2} = \lim_{\eta\to 0} \Delta\pm\frac{1}{2}\eta$ and $ v_{y,z} = \lim_{\kappa\to 0}\pm \;v +\frac{1}{2}\kappa$ the rate expressions $$\gamma_{\Omega,\delta}^{} = \frac{3}{4}\vartheta_1 + \frac{1}{4}\vartheta_2\mp \frac{\Delta}{4\Omega}(\vartheta_1+\vartheta_2) \mp \frac{v}{4\Omega}(\vartheta_1-\vartheta_2) \; ,$$ where $\Omega=\sqrt{\Delta^2+v^2}$. These forms reduce for equal bath coupling $\vartheta=\vartheta_{1,2}$ to $$\label{sz1} \gamma_{\Omega,\delta}^{} \;=\; \vartheta \;\mp\; \frac{\Delta}{2\Omega}\,\vartheta \; .$$ Thus, the one rate is smaller and the other larger than $\vartheta$. The amplitudes associated with (\[sz1\]) are found as $$\label{sz2} B_{\Omega,\delta} \;=\; \frac{1}{4} \left(1 \,\pm\, \frac{\Delta +v}{\Omega } \right) \; ,$$ Hence in the regime $v\ll\Delta$, the amplitude of the smaller rate is maximal, while that of the larger rate is negligibly small. The results (\[sz1\]) and (\[sz2\]) may be compared with those of the nondegenerate case $I^\ast$. They are $$\gamma_{\Omega} = \gamma_\delta^{} = {\textstyle \frac{3}{4}}\,\vartheta_1 + {\textstyle \frac{1}{4}}\,\vartheta_2 \, , \quad \mbox{and} \quad B_{\Omega}=B_\delta ={\textstyle\frac{1}{4}} \; .$$ The decline of decoherence at the degeneracy point $I$ compared to point $I^\ast$ is clearly visible in Fig. \[fig:deg1\]. The decoherence minimum follows from competition of the two equally preferred ground states. Due to the $v_y$- and $v_z$-couplings, the system could relax either to parallel or to antiparallel alignment in $y$- or $z$-direction. Hence the respective second spin–spin coupling gives rise to partial suppression of decoherence. Reduction of decoherence at point $I$ may be looked upon as a new type of frustration of decoherence [@novais]. Here, the phenomenon is due to the non-commutative spin-spin couplings. ![Plots of $\langle\sigma_z^{}\rangle_t$ against $t$ for identical spins at the degeneracy point $I$ (full curve) ($v_y=-v_z$) and at the nondegenerate point $I^\ast$, $v_y=v_z$ (dashed curve). The parameters are $v_y = 0.2$, $\Delta=1$, $K=0.01$, $T=2$.[]{data-label="fig:deg1"}](qu_fr_t1.eps){width="5.9" height="3.5"} For $\langle \sigma_z\tau_z \rangle_t$ the picture is similar. In case $I$, we have $\Omega_+ = 2 \Omega$, $\Omega_- = 0$, and the rates and amplitudes read $$\begin{split} \gamma_{\Omega_+}^{} &= \gamma_{\Omega_-}^{}\;=\;{\textstyle \frac{1}{2}}\,(\vartheta_1+\vartheta_2) \\ \gamma_{5,6}^{} &= \frac{\vartheta_1 + \vartheta_2}{2} \mp \frac{\Delta}{2\Omega}(\vartheta_1-\vartheta_2) \mp \frac{v}{2\Omega}(\vartheta_1+\vartheta_2) \, , \\ B_{\Omega_+}^{} &= \frac{\Delta^2}{4\Omega^2}\, ,\quad B_{\Omega_-}^{} = \frac{1}{4}\, , \quad B_{\gamma_5^{}} = B_{\gamma_6^{}} = \frac{v^2}{4\Omega^2} \; . \end{split}$$ Thus we find for equal bath couplings $\vartheta_{1,2} =\vartheta$ $$\gamma_{5,6}^{}\; =\; \vartheta \,\mp \frac{v}{\Omega}\,\vartheta \; .$$ In contrast, in the non-degenerate case $I^\ast$ we have $\Omega_+=2\Delta$, $\Omega_- = 2 v$, and the rates and amplitudes are $$\begin{split} \gamma_{\Omega_+}^{} &= \gamma_{\Omega_-}^{} = {\textstyle \frac{1}{2}}(\vartheta_1+\vartheta_2)\, ,\quad \gamma_{5,6}^{} = {\textstyle \frac{1}{2}}(\vartheta_1+\vartheta_2)\, ,\\ B_{\Omega_+} &= {\textstyle \frac{1}{4}} \, , \quad B_{\Omega_-}^{} =0 \, , \quad B_{\gamma_5^{}} =B_{\gamma_6^{}} = {\textstyle \frac{1}{4} } \, . \end{split}$$ Thus we have decline of relaxation of $\langle\sigma_z\tau_z\rangle_t$ at the degeneracy point $I$. In Fig. \[fig:purimin\] we show a plot of $- \frac{1}{P(t)}\frac{\mathrm{d}P(t)}{\mathrm{d}t} $, which is a form of an effective transition rate from a pure to the fully mixed state. At the degeneracy point $\zeta =-1$ (case $I$), there is a distinct slowdown of the extinction of the pure intial state. ![The quantity $- \frac{1}{P(t)} \frac{\mathrm{d}P(t)}{\mathrm{d}t}$ is plotted against $\zeta = v_y/v_z$ for identical spins. There is a distinct lowering of the effective damping rate at the degeneracy point $\zeta =-1$ (case $I$). The parameters are $\Delta=1$, K=0.01 and $T=3$.[]{data-label="fig:purimin"}](purity_inset.eps){width="5.9" height="3.5"} Case $II$ is another set of parameters for which the spectrum is degenerate, $\Omega_+ =\Omega_-$ [@grigorenko:040506]. An expedient parametrization for identical spins, $\Delta_{1,2}=\Delta$, is $$\begin{split} v_y^2 &= {\textstyle \frac{1}{2}}\, (v^2 - v_-^2)\, , \quad v_z^2 = {\textstyle \frac{1}{2}}\, (v^2 + v_-^2)\, ,\\ v^2 &= v_y^2 +v_z^2 \, , \quad\quad\;\; v_-^2 =\sqrt{v^4-4\Delta^4} \;>\; 0 \, . \end{split}$$ Then we have $$\Omega^2 \;=\; 2 \Delta^2 + v^2\;, \qquad \delta\,=\,0 \;.$$ The undamped dynamics of $\langle\sigma_z\rangle_t$ is $$\langle\sigma_z\rangle_t = \cos(\Omega t)\,+\, \frac{\Omega^2 + v_-^2}{2\Omega^2}\,\big[1\,-\,\cos(\Omega t)\big] \;,$$ as follows from Eq. (\[undamp\]). At the degeneracy point $II$, the rate expressions (\[ratesz\]) would yield $$\gamma_{\Omega,\delta}^{} = \frac{1}{4}\,(3\vartheta_1+\vartheta_2) \pm \frac{v_-^2}{4\Omega^2}(\vartheta_1-\vartheta_2)\, .$$ For identical bath couplings, these would reduce to $$\gamma_\Omega^{} = \gamma_\delta^{} = \vartheta \, .$$ Now, as we have already argued in subsection \[sdegen\], the expressions for $\gamma_\delta^{}$ are not correct near to and at the degeneracy point. In fact, then the poles of $\langle\sigma_z(\lambda)\rangle$ at $\lambda = \pm \,{\rm i}\,\delta -\gamma_\delta^{}$ are determined by a quadratic equation in $\lambda$. As one approaches the degeneracy point, the complex conjugate roots turn into two real ones. We find in the limit $\delta\to 0$ for $\vartheta_{1,2}=\vartheta$ $$\mp \,{\rm i}\,\delta +\gamma_\delta \;\;\to\;\; \gamma_\pm^{} = \vartheta \pm \sqrt{\frac{v^2+v_-^2}{2\Omega^2}}\vartheta = \vartheta \pm \frac{v_z}{\Omega} \vartheta \, .$$ Furthermore, the respective amplitudes are $$B_\Omega = \frac{1}{4}\,\Big( 1 - \frac{v_-^2}{\Omega^2 }\Big) \, ,\quad B_{\gamma_\pm^{}} = \frac{1}{4}\,\Big( 1 + \frac{v_-^2}{\Omega^2} \Big) \mp \frac{1}{2} \frac{v_z}{\Omega} \, .$$ Now, since the rate $\gamma_-^{}$ is smaller than $\vartheta$ and the respective amplitude is nonzero, we have again, now at point $II$, reduction of the decay of the purity $P(t)$. Similar behavior occurs also in $\langle\sigma_z\tau_z\rangle_t$. While $\gamma_{\Omega_+}^{}$ and $\gamma_{\Omega_-}^{}$ are as in Eq. (\[ratesztz\]), we now have for $\vartheta_{1,2} =\vartheta$ $$\gamma_5^{} = \vartheta + \frac{v_-^2}{\Omega^2}\,\vartheta\, ,\quad \gamma_6^{} = \vartheta - \frac{v_-^2}{\Omega^2}\,\vartheta\, ,$$ and the amplitudes read $$B_{\Omega_+} = \frac{\Delta^2}{\Omega^2}\,,\quad B_{\Omega_-} =0\, , \quad B_{\gamma_{5,6}^{}} = \frac{v^2\mp v_-^2}{2 \Omega^2} \;.$$ The decrease of the rate $\gamma_{6}^{}$ for $v_-^2>0$ and of $\gamma_5^{}$ for $v_-^2>0$ leads again to a slowdown of the decay of $P(t)$ around the degeneracy point $II$. We have verified numerically that in the quantum noise regime $T\ll \Delta$ considered below, in addition to the rate $\gamma_6^{}$ or $\gamma_5$, also the dephasing rates $\gamma_{\Omega_\pm}^{}$ and $\gamma_{\Omega,\delta}^{}$ exhibit a pronounced minimum at the degeneracy point $II$. As a result, besides the indentation in the purity characteristics (cf. Fig. \[fig:purimin\]), also the dephasing of the two-spin dynamics is considerably slowed down at the degeneracy point $II$. A numerical analysis of this phenomenon at the degeneracy point $II$ is reported in Ref. [@grigorenko:040506]. Quantum noise regime {#sqnr} ==================== Consider next the extension of the analysis to the colored quantum noise regime (QNR) relevant at $T{\,\raisebox{-.5ex} {$\stackrel{\raisebox{-.5pt}{$\textstyle <$}}{\sim}$}\,}\Omega_\pm$. Since at low $T$ quantum noise prevails, Eq. (\[wnl\]) is not valid anymore. Rather we have to revert to the expression (\[powerspec\]). In the ohmic case, we have $$\label{powerspec1} S_{\zeta}(\omega) \;=\; 2\pi K_\zeta \,\omega \coth\Big(\frac{\omega}{2T}\Big) \; .$$ We have studied the effect of the one-phonon exchange contribution to the dynamics of the two-spin model using both the perturbative Redfield approach [@blum] and the self-energy method within the path sum method [@bookweiss]. The latter amounts to systematic calculation of the self-energy to linear order in the bath correlations, ${\rm Re}\,\langle X_\zeta(t) X_\zeta(0)\rangle_T^{}= Q_\zeta(t)$ ($\zeta=1,\,2$). We have $$\label{bcorr} Q_\zeta(t) \,=\, \frac{1}{\pi} \int_0^\infty \!\!{\rm d}\omega\,\,\frac{S_\zeta(\omega)}{\omega^2}\, [\,1-\cos(\omega t)\,] \, .$$ In the standard notion of sojourns and blips [@leggett], the one-phonon self-energy, say $\Sigma_1(\lambda)$, receives contributions from the intra-blip correlation and from the four inter-blip correlations induced by bath 1 between a pair of blips of the $\sigma$-spin. The inter-blip correlations vanish in the WNR. In the usual charge picture, the former correlation is a charge-charge interaction and the latter correlation corresponds to a dipole-dipole interaction. In the time intervall between the correlated blips, the $\sigma$-spin may perform any number of uncorrelated jumps between its two blip and two sojourn states. In addition, we must take into account all transitions which spin $\tau$ can make during the dwell time of spin $\sigma$ in sojourn and blip states which are spanned by bath correlations. The succession of flips of the $\sigma$- and $\tau$-spin is dictated by the Hamiltonian (\[ham1\]) with (\[ham2\]). Following the lines expounded for the single spin-boson model [@bookweiss], it is straightforward, but tedious, to calculate the self-energy $\Sigma_1(\lambda)$. Interchange of the two spins and reservoirs then yields $\Sigma_2(\lambda)$. The self-energies $\Sigma_{1,2}(\lambda)$ lead to shifts of the poles of the $W_j(\lambda)$. It is advantageous to measure the resulting shifts in terms of generalized scaled temperatures $\Theta_{1,2}$. These depend on the power spectra (\[powerspec\]) and are normalized such that they reduce to the previously introduced scaled temperatures $\vartheta_{1,2}$, Eq. (\[wnl\]), in the white-noise limit. For lack of space, we now put $v_y=0$.\ $\langle\sigma_z\rangle_t$: The damping rates of the two oscillations with frequencies $\Omega$ and $\delta$ are found to read $$\label{sig_z_rates_low_temperature} \begin{split} \gamma_\Omega^{} &= \frac{2\,\Omega^2 - \Delta_1^2- \delta^2}{2\,(\Omega^2-\delta^2)}\,\Theta_1^{(\Omega)} + \frac{\Delta_2^2 -\delta^2}{2\,(\Omega^2-\delta^2)}\,\Theta_2^{(\Omega)} \, , \\ \gamma_\delta^{} &= \frac{\Omega^2+\Delta_1^2-2\delta^2}{2\,(\Omega^2-\delta^2)}\,\Theta_1^{(\delta)} + \frac{\Omega^2 -\Delta_2^2}{2\,(\Omega^2-\delta^2)}\,\Theta_2^{(\delta)} \, , \end{split}$$ where $$\begin{array}{rcl} \Theta_1^{(\Omega)} &=& \frac{\pi}{2}\, \frac{2\,(\Omega^2 - \Delta_1^2) S_1(\delta) + (\Delta_1^2 - \delta^2) S_1(\Omega)}{ 2\Omega^2 - \delta^2 - \Delta_1^2} \; , \\[2mm] \Theta_1^{(\delta)} &=& \frac{\pi}{2}\, \frac{ (\Omega^2 - \Delta_1^2) S_1(\delta) + 2(\Delta_1^2 - \delta^2) S_1(\Omega)}{\Omega^2 + \Delta_1^2 - 2\delta^2} \; . \end{array}$$ $$\label{sz_gen_temp_2} \Theta_2^{(\Omega)} = {\textstyle\frac{1}{2}}\,\pi\,S_2(\Omega) \, , \qquad\; \; \;\; \Theta_2^{(\delta)} = {\textstyle\frac{1}{2}}\, \pi\,S_2(\delta) \, .$$ The amplitudes asociated with the complex frequencies $\lambda=\mp\mathrm{i}\,\Omega-\gamma_\Omega^{}$ and $\lambda=\mp\mathrm{i}\,\delta-\gamma_\delta^{}$ are $$B_\Omega = \frac{\Omega^2 - \Delta_2^2 -v_z^2}{2(\Omega^2-\delta^2)} \;, \quad B_\delta = \frac{v_z^2 +\Delta_2^2 -\delta^2}{2(\Omega^2-\delta^2)}\; .$$ These one-phonon rate expressions hold in the QNR down to $T=0$ and they smoothly map on the WNR results (\[ratesz\]) at elevated temperatures. In the corresponding expressions for $\langle\tau_z\rangle_t$, the indices 1 and 2 are interchanged. Following the lines expounded in subsection \[sdegen\], we may also consider the limit $\Delta_2\to 0$. In this limit, the characteristics of the pole trajectories is as in subsection \[sdegen\]. The resulting forms for $\gamma_{\rm r}^{}$ and $\gamma_\Omega^{}$ are those of the biased spin-boson model in the one-phonon QNR [@bookweiss], $$\label{sbrate} \gamma_{\mathrm{r}} \,= \,\frac{\pi}{2}\,\frac{\Delta_1^2}{\Omega^2} \,S_1(\Omega)\, , \quad \gamma = \frac{\gamma_{\mathrm{r}}}{2} \,+ \,\frac{\pi}{2}\, \frac{v_z^2}{\Omega^2}\,S_1(0)\, .$$ Consider next the limit $v_z \rightarrow \infty$ for the symmetric case $\Delta_{1,2}=\Delta$ and $K_{1,2} =K$. For large coupling, the two spins are locked together and behave like a single spin with oscillation frequency $\bar \delta = \Delta^2/v_z$. Since the amplitude $B_\Omega$ becomes neglibly small, the dynamics is $\langle\sigma_z^{}\rangle_t =\cos(\bar\delta t)\exp(-\gamma_{\bar\delta}^{} t)$ with the dephasing rate $$\label{sbrate1} \gamma_{\bar\delta}^{} \;= \; \pi\, K\, \bar\delta \coth \Big( \frac{\bar\delta}{2\,T} \Big)\, ,$$ as follows from Eq. (\[sig\_z\_rates\_low\_temperature\]). Since the effective spin is unbiased there is no relaxation term. In the WNR limit, the rate (\[sbrate1\]) reduces to $\gamma_{\bar\delta}^{} = \vartheta$, which is twice the dephasing rate of the spin-boson model, [Eq. (\[sbrate\])]{}, in this limit. The additional factor two is because the effective spin is coupled to two identical reservoirs. The corresponding expressions for $\langle\tau_z\rangle_t$ follow from these forms by interchange of the indices 1 and 2.\ $\langle\sigma_z\tau_z\rangle_t$: As temperature is lowered from the WNR to the QNR, the damping rates of the oscillations with frequencies $\Omega_\pm$ change from $\gamma_{\Omega_\pm}^{} =\frac{1}{2}(\vartheta_1+\vartheta_2)$, Eq. (\[ratesztz\]), to the one phonon expression $$\begin{split} \gamma_{\Omega_\pm}^{} &\;=\; \frac{1}{2}\Big( \Theta_1^{(\Omega_{\pm})} + \Theta_2^{\Omega_{\pm})} \Big)\, , \\[1mm] \Theta_1^{(\Omega_{\pm})} &\;=\; \frac{\pi}{2}\,\frac{\Omega^2 - \Delta_1^2}{\Omega^2-\delta^2} \,S_1(\delta) \;+\; \frac{\pi}{2}\,\frac{\Delta_1^2-\delta^2}{\Omega^2-\delta^2}\,S_1(\Omega) \, , \\ \Theta_2^{(\Omega_{\pm})} &\;=\; \frac{\pi}{2}\,\frac{\Omega^2 - \Delta_2^2}{\Omega^2-\delta^2} S_2(\delta) \;+\; \frac{\pi}{2}\,\frac{\Delta_2^2-\delta^2}{\Omega^2-\delta^2}\,S_2(\Omega) \,. \end{split}$$ As regards the relaxation rates $\gamma_5^{}$ and $\gamma_6^{}$ of $\langle\sigma_z\tau_z\rangle_t$, the situation is more subtle, because they are determined by a quadratic equation in $\lambda$ which involves the self-energy $\Sigma_{1,2}(\lambda)$ in linear and second order in $K_{1,2}$. The calculation is most easily performed within the Redfield approach. The resulting rate expressions are $$\label{ratesztz2} \begin{split} \gamma_{5}^{} &\;=\; \frac{\Omega^2 -\Delta_1^2}{\Omega^2-\delta^2}\,\Theta_1^{(0)} + \frac{\Omega^2 -\Delta_2^2}{\Omega^2-\delta^2}\,\Theta_2^{(0)} \, , \\[1mm] \gamma_{6}^{} &\;=\; \frac{\Delta_1^2 -\delta^2}{\Omega^2-\delta^2}\,\Theta_1^{(0)} \;+\; \frac{\Delta_2^2 - \delta^2}{\Omega^2-\delta^2}\,\Theta_2^{(0)} \, , \end{split}$$ and the amplitudes read $$B_{\gamma_5^{}} = \frac{v_z^2\delta^2}{(\Omega^2-\delta^2)^2}\;, \quad B_{\gamma_6^{}} = \frac{v_z^2\Omega^2}{(\Omega^2-\delta^2)^2}\; .$$ The functions $\Theta_{1,2}^{(0)}$ depend on the power spectra at the transition frequencies $\Omega$ and $\delta$. With the abbreviaton $\Delta_1^2 \pm \Delta_2^2 = \Delta_{\pm}^2$, we find the explicit form $$\begin{split} \Theta_{1,2}^{(0)} &= \frac{\pi}{4} \Big\{ S_{1,2}(\Omega) + S_{1,2}(\delta) \\ &\qquad \pm \, \frac{\Delta_-^4 + v_z^2\Delta_+^2}{(\Omega^2-\delta^2)\Delta_-^2}\,[S_{1,2}(\Omega)-S_{1,2}(\delta)] \\ & \qquad \pm \, \frac{2v_z^2\Delta_{2,1}^2}{(\Omega^2-\delta^2)\Delta_-^2}\, [ S_{2,1}(\Omega)- S_{2,1}(\delta) ] \Big\} \end{split}$$ These expressions hold under assumption $\Delta_1\neq\Delta_2$. Summary ======= We have studied the dynamics of a spin or qubit coupled to another spin. The latter could be another qubit, a bistable impurity, or a measuring device. We have given the dynamical equations in the WNR for general spin-spin coupling and we have discussed the rich features of the coupled dynamics. Analytic expressions for dephasing and relaxation rates and for the decay of the purity have been given in the one-phonon WNR limit. Furthermore, the corresponding generalization to the quantum noise regime, which is based on a systematic calculation of the self-energy, has been presented. Our results smoothly match with those of the perturbative Redfield approach. Financial support by the DFG through SFB/TR 21 is gratefully acknowledged. [10]{} A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher, A. Garg, W. Zwerger, Rev. Mod. Phys. 59 (1987) 1; ibid. 67 (1995) 725 (E). U. Weiss, [*Quantum Dissipative Systems*]{} (Series in Condensed Matter Physics, vol 13) 3rd edn, World Scientific, Singapore, 2008. Yu. A. Pashkin, T. Yamamoto, O. Astafiev, Y. Nakamura, D. V. Averin, J. S. Tsai, Nature 421 (2003) 823. J. Clarke, F.K. Wilhelm, Nature 453 (2008) 1031. M. A. Nielsen, U. A. Chuang, Quantum Computation and Quantum Information, Cambridge University Press, Cambridge, 2000. P. Nägele, G. Campagnano, U. Weiss, New. J. Phys. 10 (2008) 115010. M. J. Storcz, F. K. Wilhelm, Phys. Rev. A 67 (2003) 042319; M. J. Storcz, F. Hellmann, C. Hrelescu, F. K. Wilhelm, Phys. Rev. A 72 (2005) 052314. E. Novais, A. H. C. Neto, L. Borda, I. Affleck, G. Zarand, Phys. Rev. B 72 (2005) 014417. I. A. Grigorenko, D. V. Khveshchenko, Phys. Rev. Lett. 94 (2005) 040506. K. Blum, [*Density Matrix Theory and Applications*]{},2nd edn, Plenum Press, New York, 1996. [^1]: The amplitudes $B_{\Omega,\delta}$ and $B_{\Omega_\pm}$ are the residues of each of the corresponding two complex poles.

ArXiv

--- abstract: 'We point out that the polarization state of radiation from a photonic crystal slab is strongly constrained by the direct non-resonant scattering process. The phase difference between the two linearly-polarized components in the far field can be predicted analytically and is largely independent of the periodic pattern. We verify the prediction with full-field electromagnetic simulations.' author: - Chia Wei Hsu - Bo Zhen - Marin Soljačić - 'A. Douglas Stone' title: Polarization state of radiation from a photonic crystal slab --- Photonic crystal (PhC) slabs are structures periodic in the transverse directions and with a finite thickness in the third dimension [@JJ_book; @1996_Krauss_Nature; @1999_Johnson_PRB; @2014_Zhou_PQE]. The photonic modes are index guided at frequencies below the light line [@JJ_book; @1996_Krauss_Nature; @1999_Johnson_PRB], while they couple to external radiations at frequencies above the light line [@2014_Zhou_PQE; @2002_Fan_PRB; @2002_Tikhodeev_PRB]. PhC slabs are attractive for their ease of fabrication and integration, tailorability, and their wide range of functionalities. They have enabled high-power high-beam-quality lasers [@2014_Hirose_NatPhoton] and many other devices [@2014_Zhou_PQE], and in recent years also became a versatile platform for studying emerging phenomena such as bound states in the continuum [@2013_Hsu_Nature; @2016_Gansch_LSA; @2017_Kodigala_Nature] and non-Hermitian exceptional points [@2015_Zhen_Nature]. Polarization is one of the most fundamental properties of electromagnetic waves. Controlling the polarization state is critical for optical communications, microscopy, nonlinear optics, and quantum optics, to name only a few. One can have the radiation from a PhC slab be in one of the two orthogonal linear polarizations by reducing the symmetry of the unit cell [@2001_Noda_Science; @2002_Bristow_JQE; @2004_Lousse_OE; @2005_Altug_OL; @2007_Fedotov_PRL; @2008_Kilic_JOSAA; @2013_Huang_OE], but it is limited to one of the two linear polarizations. One can also control the beam pattern of radiation [@2006_Miyai_Nature] or create vector beams [@2006_Miyai_Nature; @2011_Iwahashi_OE; @2012_Kitamura_OL], but at each outgoing angle the light is still mostly linearly polarized. In fact, among the large body of literature on PhC slabs and related structures, only a few recent works indicate that the radiation can deviate from linear polarization [@2011_Konishi_PRL; @2013_Shitrit_Science; @2014_Maksimov_PRB; @2015_Lobanov_PRB]. To our knowledge, there has been no analytical treatment on the polarization state of resonant radiation from PhC slabs. Here we show that part of the polarization state—specifically the phase difference between the two linear components—can be predicted with simple analytic theory and is largely independent of the actual periodic pattern. Our results provide a useful tool for designing PhC slabs that emit with the desired polarizations. {width="\columnwidth"} We start by considering a PhC slab that is periodic in $x$ and $y$ and has up-down symmetry ($\sigma_z$) and 180$^\circ$ rotational symmetry around the $z$-axis ($C_2$). The dynamics of a resonance on an isolated band at in-plane wavevector ${\bf k}_{\parallel} = (k_x, k_y)$ can be described with temporal coupled-mode theory (TCMT) [@Haus_book; @2003_Fan_JOSAA; @2004_Suh_JQE] via $$\begin{aligned} \label{eq:CMT_1} {dA}/{dt} &= \left( -i \omega_0 - \gamma \right) A + K^{\rm T} s^+, \\ \label{eq:CMT_2} s^- &= C s^+ + D A, \end{aligned}$$ where $A$ is the resonance amplitude, $\omega_0$ its resonant frequency, and $\gamma$ the decay rate due to radiation. Here we assume there is no gain or absorption loss. The column vectors $s^{\pm}$ contain amplitudes of the incident ($+$) and outgoing ($-$) planewaves at ${\bf k}_{\parallel}$, in both $s$ and $p$ polarizations: $s^{\pm} = (s^{u\pm}_s, s^{d\pm}_s, s^{u\pm}_p, s^{d\pm}_p)^{\rm T}$; the superscripts $u$ and $d$ indicate above (up) and below (down) the slab respectively, and we consider frequencies where there is no higher-order diffractions. This is schematically illustrated in Fig. 1(a). The amplitudes are normalized such that $|A|^2$ is the energy in the guided resonance, and $|s_{n}^{\pm}|^2$ is the longitudinal flux carried by the $n$-th planewave component. Column vectors $K$ and $D$ contain coupling coefficients from the resonance to the four input/output channels: $K = (k^{u}_s, k^{d}_s, k^{u}_p, k^{d}_p)^{\rm T}$, $D = (d^{u}_s, d^{d}_s, d^{u}_p, d^{d}_p)^{\rm T}$. The mirror symmetry $\sigma_z$ requires the field profile of the resonance to be either even or odd in $z$, so the coupling coefficients need to satisfy $$\label{eq:sigma} (d^{u}_s, d^{d}_s) = d_s (1, \sigma), \quad (d^{u}_p, d^{d}_p) = d_p (1, \sigma), \quad \sigma = \pm 1.$$ When the transverse electric field $(E_x, E_y)$ is used to determine the phases of $A$ and $s^{\pm}$, we have $\sigma = 1$ and $-1$ for TE-like and TM-like resonances respectively. Lastly, $C$ is a scattering matrix describing the direct (non-resonant) response and can be modeled as that of a homogeneous slab [@2002_Fan_PRB; @2003_Fan_JOSAA]. Since a homogeneous slab does not couple the two polarizations, $C$ is block diagonal with $$\label{eq:Cs_Cp} C = \begin{pmatrix} C_s & 0\\ 0 & C_p \end{pmatrix}, \quad C_s = \begin{pmatrix} r_s & t_s\\ t_s & r_s \end{pmatrix}, \quad C_p = \begin{pmatrix} r_p & t_p\\ t_p & r_p \end{pmatrix}.$$ When there is a finite amplitude $A$ in the resonance but no incident field ($s^+ = 0$), the outgoing radiation is $s^- = DA$; this can be the case when the resonance is excited via near field or fluorescence. Since $A$ is an overall amplitude, the polarization state of the far-field radiation is uniquely defined by the coupling coefficients $D$. For a high-$Q$ resonance, the mode profile does not change substantially when moderate gain is introduced, so the polarization state of the resulting laser emission is also given by $D$. Such polarization state cannot be arbitrary because the the coupling coefficients are constrained by time-reversal symmetry. Given $C_2$ symmetry, time reversal of the process $s^- = DA$ has incident field $(DA)^*$ with no outgoing field, so the directly scattered light $C(DA)^*$ must cancel the resonant radiation $DA^*$, meaning [@2003_Fan_JOSAA; @2004_Suh_JQE] $$\label{eq:CDD} CD^* = -D.$$ Eq.  may also be derived through energy conservation: unitarity of the resonant scattering process means that the resonant radiation must have certain phase relationship with the directly scattered waves. Eq.  imposes a universal constraint on the possible polarization state of the resonant radiation. Inserting Eqs.  and into Eq.  yields $$\label{eq:d_phase} {d_s^2}/{|d_s|^2} = - ( r_s + \sigma t_s)$$ when $|d_s| \neq 0$, and similarly for the $p$ component. Note that the right-hand side of Eq.  has unit magnitude because of the unitarity of $C$. Thus we have $$\label{eq:arg_cs_over_cp} \arg\left(\frac{d_s}{d_p}\right) = \frac{1}{2} \arg \left( \frac{r_s + \sigma t_s}{r_p + \sigma t_p} \right) + N \pi,$$ where $N$ is some integer. That is, the phase difference between the two polarization components in the resonant radiation is determined uniquely (up to a possible $\pi$ phase difference) by the direct-scattering coefficients. The direct-scattering coefficients do not depend on the actual pattern of the PhC slab, aside from some weak dependence on the effective index of the homogeneous slab and the resonant frequency. Therefore, Eq.  is universal. Even though Eq.  has been known in the literature, its implication on the polarization state seems to be largely unnoticed until now.w Given Eq. , the only free variable in the polarization state is $|d_s/d_p|$. When the direct-scattering coefficients are the same for the two polarizations ($r_s=r_p$, $t_s=t_p$), $d_s/d_p$ needs to be real, so the resonant radiation is linearly polarized; for isotropic materials, this is the case at the normal angle $(k_x=k_y=0)$ or at Fabry–Pérot resonances (where $t_s=t_p=\pm1$). Away from these points, there will be a phase difference, and the resonant radiation is generally elliptically polarized unless $d_s=0$ or $d_p=0$ ([*e.g.*]{} along high-symmetry lines). Approaching the grazing angle, the phase shift approaches $\pm \pi/2$ as the Fresnel reflection coefficients approach $r_p = -r_s = 1$. Fig. 2(a) verifies the TCMT prediction of Eq.  against finite-difference time-domain (FDTD) simulations [@Taflove_book], at different outgoing angles $\theta = \cos^{-1}(\hat{k}\cdot\hat{z})$ where $\hat{k}$ is the upward-pointing direction of radiation. The simulated structure uses parameters from Ref.  and is evaluated along a line $k_y = 0.5 k_x$ in the Brillouin zone; avoiding high-symmetry lines ensures that the radiating field generally contains both polarizations. We calculate the two components of the far field, $d_s \propto \hat{s}\cdot{\bf E}_{\rm rad}$, $d_p \propto \hat{p}\cdot{\bf E}_{\rm rad}$, with $\hat{s} = \hat{z}\times\hat{k}_\parallel$, $\hat{p} = \hat{s}\times\hat{k}$, and ${\bf E}_{\rm rad}$ being the radiative field. When the nodal line of $d_p$ is crossed as seem from the amplitudes in Fig. 2(b), we observe a $\pi$ phase jump in $\arg(d_s/d_p)$. The only input for the TCMT prediction are the analytically known transmission and reflection coefficients of a homogeneous slab [@2002_Fan_PRB; @2003_Fan_JOSAA], yet the prediction matches the full-field simulations quantitatively. Fig. 2(c) shows the degree of circular polarization, $\rho_c \equiv (|d_{\rm R}|^2-|d_{\rm L}|^2)/(|d_{\rm R}|^2+|d_{\rm L}|^2)$, where $d_{\rm R} = (d_p + i d_s)/\sqrt{2}$ and $d_{\rm L} = (d_p - i d_s)/\sqrt{2}$ are the far-field amplitudes in the circular polarization basis. {width="\columnwidth"} Even though the far-field vector $\vec{d} = d_s \hat{s} + d_p \hat{p}$ may be elliptically polarized, it is still possible to define a linearly polarized vector $\vec{d'} = d_s' \hat{s} + d_p' \hat{p}$ via a smooth and deterministic phase projection, $$d_s' \equiv d_s, \quad d_p' \equiv d_p \sqrt{\frac{r_s + \sigma t_s}{r_p + \sigma t_p}}.$$ Strictly speaking, the vector fields discussed in Ref.  should have been defined via this linearly polarized vector $\vec{d'}$ rather than the possibly elliptically polarized vector $\vec{d}$. With a slight modification, the above analysis can be generalized to PhC slabs without up-down mirror symmetry $\sigma_z$, but still in a uniform background ([*i.e.*]{}, same material above and below the slab). The direct scattering matrix still takes on the form of Eq. , but we no longer assume Eq. . By defining $d^{\pm}_s \equiv d^{u}_s \pm d^{d}_s$, $d^{\pm}_p \equiv d^{u}_p \pm d^{d}_p$, we get two equations $$\label{eq:arg_cs_over_cp_C2_or_P} \arg\left(\frac{d^{\pm}_s}{d^{\pm}_p}\right) = \frac{1}{2} \arg \left( \frac{r_s \pm t_s}{r_p \pm t_p} \right) + N^\pm \pi$$ with integers $N^\pm$. With $\sigma_z$ symmetry, one of the two equations vanishes, and the other one reduces to Eq. . This formalism also generalizes to PhC slabs without a 180$^\circ$ rotational symmetry $C_2$ (but still in a uniform background). Since time reversal flips the propagation direction, both the resonance at ${\bf k}_\parallel$ and the resonance at $-{\bf k}_\parallel$ need to be considered [@2013_Hsu_LSA; @2016_Zhou_Optica]; the two can have different properties in the absence of $C_2$. This yields $C_{{\bf k}_\parallel} D_{-{\bf k}_\parallel}^* = -D_{{\bf k}_\parallel}$, and $$\arg\left[\frac{d_s^{\pm}({\bf k}_\parallel) d_s^{\pm}(-{\bf k}_\parallel)}{d_p^{\pm}({\bf k}_\parallel) d_p^{\pm}(-{\bf k}_\parallel)}\right] = \arg \left( \frac{r_s \pm t_s}{r_p \pm t_p} \right),$$ which jointly constrains the polarization state of radiation at ${\bf k}_\parallel$ and at $-{\bf k}_\parallel$. The present formalism can also treat a PhC slab on a mirror substrate ([*e.g.*]{}, metal or distributed Bragg reflector) as shown schematically in Fig. 1(b); we only need to set $d^d_{s,p}=0$ and $t_{s,p}=0$. For example, we have $\arg\left({d_s}/{d_p}\right) = [ \arg \left( {r_s}/{r_p} \right)/2] + N \pi$ when the structure is $C_2$ symmetric. Note that the resonances of a homogeneous slab ([*i.e.*]{}, Fabry–Pérot resonances) must emit linearly polarized light because the two polarizations are decoupled; inhomogeneity is necessary for polarization mixing. Meanwhile, for modes in a cylindrical fiber with non-zero propagation constant and non-zero angular momentum, the two polarizations are coupled at the circular dielectric boundary, so resonances in a fiber may emit elliptically polarized light even without inhomogeneity [@2005_Schwefel_JOSAB]. The present formalism provides a useful tool for predicting and engineering the polarization state of radiation from a photonic crystal slab. It can be generalized to treat more complex geometries ([*e.g.*]{}, multiple layers, differing materials above and below, birefringent materials) and to multiple resonances including the case of degeneracy. One may also use it to study the polarization state of radiation from other types of resonators [@2012_Ruan_PRA; @2014_Hsu_NL; @2012_Verslegers_PRL; @2014_Peng_nphys]. We thank Lei Shi, Ling Lu, and Yu Guo for helpful discussions. C.W.H. and A.D.S. are partially supported by NSF via grants DMR-1307632 and ECCS-1068642. B.Z. and M.S. are partially supported by the Army Research Office through the Institute for Soldier Nanotechnologies under contract no. W911NF-13-D-0001 and by the Materials Research Science and Engineering Centers program of the NSF under award no. DMR-1419807. [37]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**]{},  ed. (, , ) @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevB.60.5751) [****,  ()](\doibase http://dx.doi.org/10.1016/j.pquantelec.2014.01.001) [****,  ()](\doibase 10.1103/PhysRevB.65.235112) [****,  ()](\doibase 10.1103/PhysRevB.66.045102) [****,  ()](\doibase 10.1038/nphoton.2014.75) [****,  ()](\doibase 10.1038/nature12289) @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1126/science.1061738) [****,  ()](\doibase 10.1109/JQE.2002.1017601) [****,  ()](\doibase 10.1364/OPEX.12.001575) [****,  ()](\doibase 10.1364/OL.30.000982) [****,  ()](\doibase 10.1103/PhysRevLett.99.147401) [****,  ()](\doibase 10.1364/JOSAA.25.002680) [****,  ()](\doibase 10.1364/OE.21.020675) @noop [****, ()]{} [****,  ()](\doibase 10.1364/OE.19.011963) [****,  ()](\doibase 10.1364/OL.37.002421) [****,  ()](\doibase 10.1103/PhysRevLett.106.057402) [****,  ()](\doibase 10.1126/science.1234892) [****,  ()](\doibase 10.1103/PhysRevB.89.045316) [****,  ()](\doibase 10.1103/PhysRevB.92.205309) @noop [**]{} (, , ) [****,  ()](\doibase 10.1364/JOSAA.20.000569) [****,  ()](\doibase 10.1109/JQE.2004.834773) @noop [**]{},  ed. (, , ) [****,  ()](\doibase 10.1103/PhysRevLett.113.257401) [****,  ()](\doibase 10.1038/lsa.2013.40) @noop [****,  ()]{} [****,  ()](\doibase 10.1364/JOSAB.22.002295) [****,  ()](\doibase 10.1103/PhysRevA.85.043828) [****,  ()](\doibase 10.1021/nl500340n) [****,  ()](\doibase 10.1103/PhysRevLett.108.083902) @noop [****,  ()]{}

ArXiv

--- author: - | Grant Keady and Benchawan Wiwatanapataphee\ Department of Mathematics and Statistics\ Curtin University, Bentley, 6102\ Western Australia title: Functions with constant Laplacian satisfying Robin boundary conditions on an ellipse --- Abstract {#abstract .unnumbered} ======== We study the problem of finding functions, defined within and on an ellipse, whose Laplacian is -1 and which satisfy a homogeneous Robin boundary condition on the ellipse. The parameter in the Robin condition is denoted by $\beta$. The integral of the solution over the ellipse, denoted by $Q$, is a quantity of interest in some physical applications. The dependence of $Q$ on $\beta$ and the ellipse’s geometry is found. Several methods are used. - To find the general solution the boundary value problem is formulated in elliptic cylindrical coordinates. A Fourier series solution is then derived. Results concerning the difference equation which the Fourier coefficients satisfy are presented. - Variational methods have given simple and accurate lower bounds. - Various asymptotic approximations are found directly from the pde formulations, this being far easier than from our series solution. The pde for large $\beta$ asymptotics again leads to difference equations. It is intended that this arXiv preprint will be referenced by the journal version, which will be submitted soon, as the arXiv contains material, e.g. codes for calculating $Q$, not in the very much shorter journal version. Introduction {#sec:Intro} ============ The pde problem {#subsec:GovEq} --------------- With $\Omega$ a plane domain and $\beta\ge{0}$ given, we seek the solution of $$- \frac{\partial^2 u}{\partial x^2}- \frac{\partial^2 u}{\partial y^2} = 1 \qquad {\rm in\ \ }\Omega ,\qquad u + \beta \frac{\partial u}{\partial n} = 0 \qquad {\rm on\ \ }\partial\Omega . \eqno({\rm P}(\beta))$$ Under reasonable conditions on the boundary $\partial\Omega$ (classical) solutions exist and, by Maximum Principle arguments, are unique and positive in $\Omega$. [@KM93] surveys some of the easier results. In the context of flows, this boundary condition is called a [*slip boundary condition*]{} or [*Navier’s boundary condition*]{}; in the wider mathematical literature it is called a [*Robin boundary condition*]{}. The $\beta=0$ case is, in the fluid mechanics context, called a ‘no-slip boundary condition’: in elasticity the $\beta=0$ case is called ‘the elastic torsion problem’. A functional of interest, in the context of flows, is the volume flow rate $${Q}:= \int_\Omega u . \label{eq:Qdef}$$ We remark that the same pde problem arises in contexts other than slip flow, for example, heat-flow with ‘Newton’s law of cooling’, e.g. [@KM93; @MK94]. While there are other applications of this pde problem we will, on occasions, use the fluid flow terminology for $u$, $Q$ and $\beta$. There are many inequalities established for Problem (P($\beta$)). There are isoperimetric inequalities, reviewed in [@KWK18] and others, for example, from [@KM93] $${\rm max}\left( \beta\, \frac{|\Omega|^2}{|\partial\Omega|} +{Q}_0,\ \beta\, \frac{|\Omega|^2}{|\partial\Omega|} + \Sigma_\infty +\frac{\Sigma_1}{\beta}\right) \le {Q}(\beta) \le \beta\, \frac{|\Omega|^2}{|\partial\Omega|} + \Sigma_\infty = U(\beta) , \label{in:vBnds}$$ where $\Sigma_\infty>0$ and $\Sigma_1\le{0}$ are defined in equation (\[eq:SigDef\]) at the beginning of §\[sec:Varl\] and used associated with asymptotics for $\beta$ large. $U(\beta)$ is defined as the expression immediately to its left. These old inequalities (\[in:vBnds\]) are established using the variational formulation of §\[sec:Varl\] in [@KM93] : see inequalities (4.4) and (4.9). The inequality (4.4) of [@KM93] is well-known, (4.9) less so. The same methods lead to an improvement given in inequality (\[in:lbbL\]). Except in this subsection and a very small number of introductory subsections, e.g. §\[subsec:inS0Sinf\], §\[subsec:VarlGen\] and §\[subsec:betaSmallGenOm\] where we allow $\Omega$ to be more general, we will be treating $\Omega$ as the elliptical domains within ellipses described by $$\frac{x^2}{a^2}+\frac{y^2}{b^2} \leq 1\qquad{\rm with\ \ } a\geq b . \label{eq:ellab}$$ Outside the main paper, in a lengthy appendix, we use the variational techniques of this paper to obtain similar results for rectangular domains. In this paper, we derive, for elliptical domains, a Fourier series form of the analytical solution of the problem and provide many checks on it. Others, e.g. [@DM07], have used similar separation of variables approaches but have chosen to approximate the Fourier series of the function $g$ occuring in the boundary condition, equation (\[bc:4\]) by its first two terms, whereas we treat this more carefully. Some recent papers treat the problem in the context of slip flows in elliptic microchannels, e.g. [@DT16; @DM07; @Duan07]. (Again in the context of flows in microchannels, time-dependent and other aspects are treated in various papers. [@Gupt; @Hasl; @Suh; @Suh18] study pulsatile flows. Papers [@Day55; @Tr73] are relevant to studies of the decay of transients and starting flows. Other physical effects are studied in [@Ghos; @Hsu]. The steady flow case is, for each of these, a special case when a parameter is set to zero.) We make considerable effort to check our results. For example, there are a number of rigorously proved bounds. An old variational result, a lower bound on ${Q}$ is equation (4.4) of  [@KM93], $$\begin{aligned} {Q}(\beta) &\geq& {Q}(\beta=0) + \frac{\beta\, |\Omega|^2}{|\partial\Omega|}\quad {\rm \ for\ general\ }\Omega, \nonumber\\ &=& \frac{\pi a^3 b^3}{4 (a^2 + b^2)} + \frac{\pi^2 \beta\, a^2 b^2}{|\partial\Omega|}\quad\quad {\rm \ for\ ellipse\ }\Omega . \label{KM93eq4p4}\end{aligned}$$ Inequality (\[KM93eq4p4\]) and the more general form preceding it are very easy to establish: see §\[sec:Varl\]. Inequality (\[KM93eq4p4\]) becomes an equality when the ellipse is a circle, i.e. $b=a$. The left-hand term in the right-hand side of inequality (\[KM93eq4p4\]) gives the dominant term in the asymptotic expansion of $Q$ as $\beta$ tends to 0. The right-hand term in inequality (\[KM93eq4p4\]) gives the dominant term in the asymptotic expansion of $Q$ as $\beta$ tends to infinity. However (except when $b=a$), in both asymptotic limits, the other term is not the next term in the asymptotic expansion. We sometimes denote the area $|\Omega|$ by $A$ and the perimeter $|\partial\Omega|$ by $L$. For an ellipse, the formula for $L$ is given in equation (\[eq:perimE\]). A much more recent inequality, a generalization of the St Venant Inequality (see [@PoS51]), is given in [@BG15] (with related work in [@BB13]): $${Q}\leq Q_\odot(\beta)\, := \frac{\pi (a\, b)^{3/2}}{8}\left(\sqrt{a b} + 4\beta\right) . \label{eq:isoperQ}$$ In words: amongst all ellipses with a given area that which has the greatest ${Q}$ is the circular disk. See equation (\[eq:circleQinf\]). (This, and other isoperimetric inequalities, are reviewed, in the context of microchannels in [@KW16; @KWK18].) The structure of this paper {#subsec:structure} --------------------------- The rest of this paper is organized as follows. - In §\[sec:explicit\] we summarize the very well-known elementary solutions, polynomial in Cartesian coordinates, for the simplest special cases against which to check our general solution. Each of them forms the lowest order term in asymptotics developed later in this paper. - In §\[sec:GovElliptic\], the boundary value problem is formulated in elliptic cylindrical coordinates, and then the series solution is derived. - The next sections provide many checks. To reduce the number of parameters, in these sections, we take $b=1/a$, so the area of the ellipse is $\pi$. - In §\[sec:Varl\] we present a variational formulation and associated bounds. - In §\[sec:nearCircQs\] we present asymptotics for the situation where the ellipse is nearly circular. - In §\[sec:betaSmall\] we give asymptotics for $\beta$ small. - In §\[sec:betaLarge\] we study asymptotics for $\beta$ large.\ - Finally, a conclusion is given in §\[sec:Conclusion\]. - Appendix A treats various geometric matters. Appendix B has some notes on elliptic integrals. These arise in the Fourier series for a functions $g$ and $\hat g$ used at several sections of the paper, notably §\[sec:GovElliptic\] and §\[sec:betaLarge\]. Appendix \[app:Furtherg\] collects information on these functions primarily in connection with possible further work mentioned in §\[sec:Conclusion\]. Appendix D gives some numerical examples, its first parts comparing our numerics with some previously published examples. We conclude this appendix with an attempt to suggest numbers which might arise in connection with possible applications involving blood flow. Appendix \[app:Rect\] surveys some similar results for rectangular domains. Some simple explicit solutions:\ $u$ quadratic polynomial in $x$, $y$ {#sec:explicit} ==================================== Circular cross-section with $\beta\ge{0}$ {#subsec:Circular} ----------------------------------------- When $\Omega$ is circular, radius $a$, in polar coordinates, $r=\sqrt{x^2+y^2}$, the solution is $$u_{\beta\odot}= {u}= \frac{1}{4}\left(a^2-r^2\right) +\frac{{\beta}a}{2} . \label{eq:circuSteady}$$ In the context of fluid flows, when ${\beta}=0$ this is Poiseuille flow. $$Q_{\beta\odot}={Q}({\beta}) = \frac{\pi a^3}{8} (a+4{\beta}) . \label{eq:circleQinf}$$ See also [@La32]§[331]{}, p586. In later sections the circle will have radius $a=1$, area $\pi$. Elliptic cross-section with $\beta={0}$ {#subsec:beta0} --------------------------------------- In this subsection (and in sections §\[sec:Varl\] to §\[sec:betaLarge\]) we consider ellipses with $b=1/a$ and $a\ge{1}$. The $\beta=0$ solution dates back at least to St Venant. In polar coordinates with the origin at the centroid of the ellipse the velocity $u_e$ is, in Cartesian coordinates, a quadratic polynomial in $x$ and $y$ while, in polar coordinates, it is $$u_{0e} =\frac{1}{4}\left( \sqrt {1-\epsilon^{2}} -r^2 +\epsilon \, r^2 \, \cos(2\theta)\, \right) .$$ Here $$\epsilon = \frac{a^2 - a^{-2}}{a^2 +a^{-2}} .$$ In Cartesian coordinates $$u_0 =\frac{1- (x/a)^2 - (y/b)^2}{2/a^2 + 2/b^2} . \label{eq:u0Cart}$$ Also $$Q_{0e}={Q}({\rm ellipse}, \beta=0) =\frac{\pi}{4 (a^2 + a^{-2})} = \frac{\pi}{8} \, \sqrt {1-\epsilon^{2}} . \label{eq:Qellb0}$$ Also relevant is $e$ the eccentricity as defined in equation (\[eq:eDef\]). ![$\beta=0$: plot of Qratio=$(Q_\odot-Q(e))/Q_\odot$ against eccentricity, $e$.[]{data-label="fig:QratioBeta0jpg"}](\grpath/QratioBeta0jpg.jpg){height="7cm" width="13cm"} There are various interesting alternative expressions for ${Q}$. The ellipse’s moments of inertia are $$I_{xx}=\frac{\pi a^2}{4} , \quad I_{yy}=\frac{\pi}{4 a^2}\ {\rm so\ \ } {Q}= J:= \frac{I_{xx} I_{yy}}{I_{xx} +I_{yy}} .$$ See [@PoS51] p112. For any domain (and $\beta=0$) it is known that ${Q}\le{J}$. Fourier series solution {#sec:GovElliptic} ======================= Using elliptic coordinates {#subsec:EllCoords} -------------------------- ### Formulae when $a\ge{b}>0$ {#formulae-when-ageb0 .unnumbered} For the geometry of elliptic domains specified as in equation (\[eq:ellab\]) with, as always, $a>b$, it is convenient here to use elliptic coordinates $(\eta, \psi)$ which are related to the rectangular coordinates (as in [@DT16] equation (18)) by $$\label{eq:5} x = c \cosh(\eta) \cos(\psi); \;\;\;\;\;y = c \sinh(\eta) \sin(\psi); ,$$ where $0 \leq \eta \leq \infty,\;0 \leq \psi \leq 2 \pi$, and $c$ and $-c$ are two common foci of the ellipse. The origin of coordinates in the Cartesian system (the centre of our ellipse) is at $\eta=0$, $\psi=\pi/2$. The upper side of the positive semi-major axis is $\eta=0$, $0<\psi<\pi/2$: the lower side is $\eta=0$, $0>\psi>-\pi/2$. Let $a$ and $b$ denote lengths of the semi-axes of the ellipse and $a > b > 0$. Except in this subsection, and elsewhere when so specified, we take $b=1/a$. Returning to the general situation, $$a = c \cosh(\eta) {\rm\ \ and\ \ } b = c \sinh(\eta), {\rm \ \ and\ \ } c=\sqrt{a^2-b^2}.$$ Defining $\eta_0$ (as in [@DT16] equation (19)) by $$\eta_0 = \ln\frac{1+b/a}{\sqrt{1-(b/a)^2}} = {\rm arctanh}(\frac{b}{a}),$$ the boundary of the ellipse can be represented by $$x= c \cosh(\eta_0) \cos(\psi), \qquad y= \sinh(\eta_0) \sin(\psi),$$ so that $$c=\frac{a}{\cosh \eta_0}=\frac{b}{\sinh \eta_0}.$$ The eccentricity is $$e = \sqrt{1- \frac{b^2}{a^2}} = \sqrt{1-\tanh^2(\eta_0)} =\frac{1}{\cosh(\eta_0)}.$$ (Nearly circular ellipses have $\eta_0$ large and $c$ near 0.) The Jacobian $J$ of the transformation from $(x,y)$-coordinates to $(\eta,\psi)$ coordinates is $$J(\eta,\psi)= c^2 \left( \cosh^2(\eta) -\cos^2(\psi) \right) = \frac{c^2}{2} \left( \cosh(2\eta) -\cos(2\psi) \right) .$$ This is needed in the derivation of equation (\[eq:QJac\]). The Jacobian will be used in other places in the paper, e.g. §\[subsec:Qinf\]. ### Formulae when $b=1/a$ {#formulae-when-b1a .unnumbered} In all major calculations henceforth we scale distances so that we have $b=1/a$. Formulae used later include the following: $$\begin{aligned} c^2 =\frac{2}{\sinh(2\eta_0)} &=& a^2-\frac{1}{a^2}=\frac{e^2}{\sqrt{1-e^2}}, \label{eq:cae}\\ \tanh(\eta_0) &=& \frac{1}{a^2} , \label{eq:tanha}\\ \cosh(\eta_0) &=& \frac{1}{e} , \label{eq:coshe}\\ e^2 &=& 1- \frac{1}{a^4}. \label{eq:eaDef}\end{aligned}$$ The pde and boundary conditions ------------------------------- The pde of Problem (P($\beta$)) is recast in these coordinates as $$\begin{aligned} \label{eq:7} \frac{\partial^2 u}{\partial \eta^2} + \frac{\partial^2 u}{\partial \psi^2} &=&-c^2 (\cosh^2(\eta)-\cos^2(\psi) ) ,\nonumber \\ &=& -\frac{c^2}{2}(\cosh(2\eta)-\cos(2\psi)) .\end{aligned}$$ We will solve the pde (\[eq:7\]) subject to boundary conditions (\[bc:1\]), (\[bc:2\]) and (\[bc:3\]). The solution is assumed to be symmetric about the mid-plane for $\psi\;=\;0,\;\frac{\pi}{2}$. Therefore, symmetric boundary condition is applied, i.e., $$\label{bc:1} \frac{\partial u}{\partial \psi}\; = 0\; \mathrm{on}\;\;\psi\;=\;0,\;\frac{\pi}{2}.$$ In addition, we set $$\label{bc:2} \frac{\partial u}{\partial \eta} = 0\;\;\mathrm{on}\;\;\eta=0.$$ The boundary condition on the ellipse, i.e. at those points when $\eta=\eta_0$ is $$\label{bc:3} u(\eta_0,\psi)+\frac{\beta}{c\sqrt{\cosh^2\eta_0 - \cos^2 \psi}} \frac{\partial u}{\partial \eta}(\eta_0,\psi) \;=\;0.$$ Define $$\label{eq:gdef} g(\psi)= \frac{\cosh(\eta_0)}{ (\cosh^2(\eta_0) - \cos^2(\psi))^{1/2}} .$$ Boundary condition (\[bc:3\]) becomes $$\label{bc:4} u(\eta_0,\psi)+\frac{\beta}{c \cosh(\eta_0)}g(\psi) \frac{\partial u}{\partial \eta}(\eta_0,\psi) \;=\;0.$$ In summary, the problem in elliptic coordinates is given as the following boundary value problem (BVP).\ [**BVP**]{}: Find $u(\eta,\psi)$ such that the above governing equation (\[eq:7\]) and associated boundary conditions (\[bc:1\]), (\[bc:2\]) and (\[bc:4\]) are satisfied. The Fourier series for $g$ {#subsec:gFS} -------------------------- ### Ellipses in general {#ellipses-in-general .unnumbered} The definition of $g$ from (\[eq:gdef\]) can be re-cast in several ways: $$\begin{aligned} g(\psi) &=& \left(1-\frac{\cos^2(\psi)}{\cosh^2(\eta_0)}\right)^{-1/2} ,\\ &=& C\, g_*(\psi) \ {\rm with \ \ } g_*(\psi) = \left(1-\frac{\cos(2\psi)}{\cosh(2\eta_0)}\right)^{-1/2}\ {\rm and}\ \ C= \left(\frac{1+\cosh(2\eta_0)}{\cosh(2\eta_0)} \right)^{1/2}.\end{aligned}$$ In an appendix we will use a more economic notation using $$q=\cosh^2(\eta_0) , \label{eq:qEta0}$$ and the form for $g_*$ is consistent with the occurence, in Fourier coefficients, of polynomials in $\cosh(2\eta_0)=2 q-1$. The Fourier series for $g$ is used. Define $$g_n = \frac{2}{\pi}\int_0^\pi g(\psi)\cos(2 n\psi)\, d\psi , \label{eq:bgnDef}$$ so $$g(\psi) = \frac{g_0}{2} + \sum_{n=1}^\infty g_n\, \cos(2 n\psi) ,$$ In various approximations the first few terms in the Fourier series are used. For these define $${\rm EllipticE}_0= {\rm EllipticE}(\frac{1}{\cosh(\eta_0)}) \qquad{\rm and\ \ } {\rm EllipticK}_0= {\rm EllipticK}(\frac{1}{\cosh(\eta_0)}). \label{eq:Ellnotn}$$ We have $$\begin{aligned} g_0 &=& \frac{4}{\pi}\, {\rm EllipticK}_0 , \label{eq:g0Ell}\\ g_1 &=& -\frac{8\cosh(\eta_0)^2}{\pi}\, {\rm EllipticE}_0 + \frac{4(2\cosh(\eta_0)^2-1)}{\pi}\, {\rm EllipticK}_0 \ . \label{eq:g1Ell}\end{aligned}$$ All the terms $g_n$ can be written in terms of Legendre $Q$ functions though we first encountered them in the form presented by maple, $$g_n =\frac{4}{\pi}\left( E_n \, {\rm EllipticE}_0 + K_n\, {\rm EllipticK}_0 \right), \label{eq:gnEK}$$ where $E_n$ and $K_n$ are polynomials of degree $n$ in $\cosh(\eta_0)^2$ with rational number coefficients. Should future work require more terms we remark that a three term recurrence relation exists to determine the polynomials $E_n$ and $K_n$. See Appendix \[app:Furtherg\]. In the notation of equation (\[eq:qEta0\]), the $K_n$ sequence of polynomials starts with $$K_0=1 , \qquad K_1= 2q-1 = q_2 .$$ The $E_n$ sequence of polynomials starts with $$E_0=0 , \qquad E_1= 2q = q_2-1 .$$ There are also other representations: see Appendix \[app:Furtherg\]. There are several very elementary facts which have some use. These include that $$\cos(2 k\psi)= {\rm ChebyshevT}(\cos(2\psi)) .$$ Also, the Taylor series $$(1-X)^{-1/2}= \sum_{k=0}^\infty \frac{(2 k)!}{2^{2k}( k!)^2} X^k ,$$ can be applied to the preceding expressions for $g$ to obtain series which are sums of powers of $\cos(\psi)^2$ or of $\cos(2\psi)$. To date, our only use of these has been in checking the asymptotic approximations, at small eccentricity, to the Fourier coefficients of $g$ and of $1/g$. The reciprocal of $g$ also occurs in later calculations (and also in [@DT16] equation (61) though they use a Taylor series rather than Fourier series used here). ### Approximating the Fourier series for nearly circular ellipses {#approximating-the-fourier-series-for-nearly-circular-ellipses .unnumbered} Nearly circular ellipses have $\eta_0$ large. Then the lowest approximation is $$g(\psi)\sim g_\epsilon(\psi) = 1+\frac{\cos^2(\psi)}{2\cosh^2(\eta_0)} =1+\frac{1}{4\cosh^2(\eta_0)}+\frac{\cos(2\psi)}{4\cosh^2(\eta_0)} ,$$ and $$\frac{1}{g(\psi)}\sim 1-\frac{1}{4\cosh^2(\eta_0)}-\frac{\cos(2\psi)}{4\cosh^2(\eta_0)} . \label{eq:gRecipFSnearCirc}$$ We will use the next approximation after this, and will use $e=1/\cosh(\eta_0)$: $$\frac{1}{g(\psi)}\sim 1-\frac{1}{4}e^2 - \frac{3}{64} e^4 - \left( \frac{1}{4}e^2+ \frac{1}{16}e^4 \right) \cos(2\psi) - \frac{1}{64}e^4 \, \cos(4\psi) . \label{eq:gRecipFSnearCirc2}$$ We have, for $e$ small the following Fourier coefficients of ${\hat g}=1/g$: $$\begin{aligned} {\hat g}_0 &\sim& 2-\frac{1}{2} e^2-\frac{3}{32} e^4 \label{eq:B0e} \\ {\hat g}_1 &\sim& -\frac{1}{4} e^2 -\frac{1}{16} e^4 \label{eq:B1e} \\ {\hat g}_2 &\sim& -\frac{1}{64} e^4 \label{eq:B2e}\end{aligned}$$ (We will need these in the treatment of $u_\infty$. See equation (\[eq:rgFSB\]).) Completing the Fourier series solution {#subsec:Analytical} -------------------------------------- The general solution of equation (\[eq:7\]) can be expressed in the form of $$\label{us} u(\eta, \psi)\;=\;v_h+v_p,$$ where $v_h$ and $v_p$ denote the homogeneous and the particular solution of the pde (\[eq:7\]) subjecting to the boundary conditions (\[bc:1\]),(\[bc:2\]) and(\[bc:4\]).\ For the particular solution $v_p$, from the right hand side of the equation (\[eq:7\]), we can expressed $v_p$ in the form of $$\label{eq:9b} v_p\;=\;{\hat g}_1\cosh(2\eta)+{\hat g}_2\sinh(2\eta)+{\hat g}_3\cos(2\psi)+{\hat g}_4\sin(2\psi),$$ which gives $$\label{eq:10b} \frac{\partial^2 v_p}{\partial \eta^2}\;=\;2 {\hat g}_1\sinh(2\eta)+2 {\hat g}_2\cosh(2\eta);$$ $$\label{eq:11b} \frac{\partial^2 v_p}{\partial \psi^2}\;=-2 {\hat g}_3\sin(2\psi)+2 {\hat g}_4\cos(2\psi).$$ Substituting $v_p$ into the pde (\[eq:7\]) and comparing the coefficients of $\cosh(2\eta)$ and $\cos(2\psi)$ yields $$\label{eq:12b} v_p\;=\;-\frac{c^2}{8} (\cosh(2\eta)+\cos(2\psi)) \ \ \left( = -\frac{1}{4}\, (x^2+y^2)\ \right).$$ By separation of variables, the solution $v_h$ of the Laplace equation subject to above boundary conditions (\[bc:1\]) and (\[bc:2\]) is $$\label{eq:8b} v_h\;=\;\sum_{n=0}^{\infty}A_n \, \cosh(2n\eta)\cos(2n\psi).$$ Hence, the general solution can be expressed in the form of $$\label{eq:13} u(\eta,\psi)\;=\;v_h+v_p\;=\;\sum_{n=0}^\infty A_n \, \cosh(2n\eta)\cos(2n\psi) - \frac{c^2}{8}(\cosh(2\eta)+\cos(2\psi)),$$ where term $A_n$ can be determined by using the boundary condition (\[bc:4\]). We then obtain $$\begin{aligned} \label{eq:14} &\sum_{n=0}^\infty A_n \, \cosh(2 n\eta_0)\cos(2 n\psi)- \frac{c^2}{8}\, (\cosh(2\eta_0)+\cos(2\psi)) + \frac{\beta g(\psi)}{c\cosh(\eta_0)} \nonumber \\ &\sum_{n=0}^\infty n 2 A_n \sinh(2 n \eta_0)\cos(2 n \psi)-\frac{c^2 \beta g(\psi)}{4 c \cosh(\eta_0)}\sinh(2 \eta_0) = 0.\end{aligned}$$ or $$\begin{aligned} \label{eq:18} &\sum_{n=0}^\infty A_n \, \left( \cosh(2 n\eta_0) + \frac{2 n \beta g(\psi)}{c\cosh(\eta_0)} \sinh(2 n \eta_0) \right)\cos(2 n\psi) \nonumber\\ &=\frac{c^2}{8}(\cosh(2\eta_0)+\cos(2\psi)) + \frac{c^2 \beta g(\psi)}{4 c \cosh(\eta_0)}\sinh(2 \eta_0)\;=\;f(\psi).\end{aligned}$$ As $g$ is an even function so is $f$ and $h_n$ where $$h_n(\psi)=\cosh(2 n\eta_0) + \frac{2 n \beta g(\psi)}{c\cosh(\eta_0)} \sinh(2 n \eta_0).$$ (The Fourier coefficients of $f$ and of $h_n$ are very simply related to the Fourier coefficients of $g$.) We then have $$\sum_{n=1}^\infty A_n \int_{-\pi}^{\pi} h_n(\psi) \cos(2n\psi)\cos(2m\psi) d\psi = - \int_{-\pi}^{\pi} f(\psi) \cos(2m \psi) d\psi.$$ Hence if we truncate the summation at $n_{\rm max}$ we have a system of linear equations for the $A_n$ for $n$ up to $n_{\rm max}$. Using the $A_n$ so found, we obtain an approximation to the Laplace solution $v_h$ and then steady-state fluid flow, $u(\eta,\psi)=v_h+v_p$ in the elliptical channel. Finally, the volume flow rate of fluid passing through a elliptical cross sectional area, $Q$, can determined by $${Q}\;=\; {c^2} \int_0^{2\pi} \int_0^{\eta_0} u(\eta,s) \left( \cosh^2(\eta) -\cos^2(s) \right) d \eta ds. \label{eq:QJac}$$ ${Q}$ from the Fourier series of $v_h$ -------------------------------------- We will substiute $u=v_p+v_h$ with $v_h$ given by its Fourier series into the formula for $Q$ given in (\[eq:QJac\]). Note that all the Fourier components with $n>1$ contribute nothing to the integral. First $$\begin{aligned} Q_p &=& c^2 \int_0^{2\pi} \int_0^{\eta_0} v_p \, \left( \cosh^2(\eta) -\cos^2(s) \right) d \eta ds , \\ &=& -\frac{1}{4} I_2 = -\frac{1}{4}\, \frac{\pi}{4} \, (a^2+\frac{1}{a^2}) = - \frac{\pi}{16} \, \frac{2-e^2}{\sqrt{1-e^2}} ,\end{aligned}$$ $I_2$ being the polar area moment of inertia about the centroid. On using $c^2=2/\sinh(2\eta_0)$, $$\begin{aligned} Q &=& Q_p+ c^2 \int_0^{2\pi} \int_0^{\eta_0} ( A_0 + A_1 \cosh(2\eta)\cos(2 s)) \left( \cosh^2(\eta) -\cos^2(s) \right) d \eta ds , \nonumber \\ &=& - \frac{\pi}{8} \, \frac{\cosh(2\eta_0)}{\sinh(2\eta_0)}+ \pi\, A_0 - \frac{1}{2}\,\pi\, A_1 . \label{eq:QFS}\end{aligned}$$ Inequalities from a variational formulation {#sec:Varl} =========================================== Variational formulations are available. For example, in [@KM93] equation (4.1), the functional $\cal J$ is defined as $${\cal J}(v) ={\cal A}(v) - \frac{1}{\beta}{\cal B}(v) \quad{\rm where\ \ } {\cal A}(v)= \int_\Omega(2 v -|\nabla v|^2), \ \ {\cal B}(v) =\int_{\partial\Omega} v^2 . \label{eq:Varl}$$ For any simply-connected domain with sufficiently smooth boundary, the maximiser $u$ of this functional solves Problem (P($\beta$)). Furthermore, ${\cal J}(u)$ is ${Q}={\cal Q}(u)$ where $${\cal Q}(v) = \int_\Omega v .$$ Thus we can find lower bounds for ${Q}$ by evaluating ${\cal J}(v)$ for particular choices of $v$. In this section, we have two distinct subsections.\ $\bullet$ The first of these is included as motivation for further work on Problem (P($\infty$)) for the ellipse. We have a series solution for $\Sigma_\infty$ for the ellipse involving Legendre functions but have yet to find $\Sigma_1$. The lower bound is useful in situations, as with the rectangle treated in the final appendix, where $u_\infty$ is available analytically.\ $\bullet$ The second of these is good when the solutions of Problem (P($\beta$)) have nearly elliptical level curves. This is the case for the ellipse, and also, for large $\beta$ for a number of other domains, including rectangles. A development from [@KM93] {#subsec:inS0Sinf} -------------------------- Our notation is as in [@KM93]. For large $\beta$ $$u \sim \beta\, \frac{|\Omega|}{|\partial\Omega|} + u_\infty + o(1) ,$$ where $u_\infty$ solves $$- \frac{\partial^2 u_\infty}{\partial x^2}- \frac{\partial^2 u_\infty}{\partial y^2} = 1\, {\rm in}\,\Omega ,\, \frac{\partial u_\infty}{\partial n} = -\frac{|\Omega|}{|\partial\Omega|} \, {\rm on}\,\partial\Omega \, {\rm and}\, \int_{\partial\Omega} u_\infty = 0 . \eqno({\rm P}(\infty))$$ Define $$\Sigma_\infty = \int_\Omega u_\infty , \qquad {\rm and\ \ } \Sigma_1= - \int_{\partial\Omega} u_\infty^2 , \label{eq:SigDef}$$ with $u_\infty$ satisfying Problem (P($\infty$)). The terms $\Sigma_\infty>0$ and $\Sigma_1\le{0}$ (both obviously independent of $\beta$) give coefficients in the next terms in the asymptotic expansion $${Q}(\beta) \sim \frac{\beta\, |\Omega|^2}{|\partial\Omega|} + \Sigma_\infty +\frac{\Sigma_1}{\beta} \qquad {\rm for\ \ } \beta\rightarrow\infty .$$ Inequality (4.10) of [@KM93] (also (4.12)) states $$\frac{\beta\, |\Omega|^2}{|\partial\Omega|} + \Sigma_\infty +\frac{\Sigma_1}{\beta} \le {Q}(\beta) . \label{in:KM4p10}$$ We now improve on the lower bounds on ${Q}(\beta)$ given in (\[in:vBnds\]) and in (\[in:KM4p10\]). A reasonable choice for test functions $v$ to insert into ${\cal J}(v)$ is $$v = c_0 + t_0 u_0 + t_\infty u_\infty . \label{eq:vTest}$$ In substituting the functions as in (\[eq:vTest\]) into ${\cal J}(v)$ it is useful to note that simple applications of the Divergence Theorem give $${Q}_0 =\int_\Omega |\nabla u_0|^2 \qquad {\rm and\ \ \ } \Sigma_\infty =\int_\Omega |\nabla u_\infty|^2 .$$ The Divergence Theorem also gives $$\int_\Omega {\rm div} (u_0\nabla{u_\infty)}= 0 \qquad{\rm whence \ \ } {Q}_0 = \int_\Omega \nabla u_0 \cdot \nabla u_\infty .$$ Now $$\begin{aligned} {\cal A}(v) &=& 2c_0|\Omega| + t_0(2-t_0)\,{Q}_0 + t_\infty (2-t_\infty)\,\Sigma_\infty -2 t_0\, t_\infty\, \int_\Omega \nabla u_0 \cdot \nabla u_\infty ,\\ &=& 2c_0|\Omega|+ t_0\,(2-t_0-2t_\infty)\, {Q}_0 +t_\infty (2-t_\infty)\,\Sigma_\infty ,\\ {\cal B}(v) &=& c_0^2 | \partial\Omega| - t_\infty^2\,\Sigma_1 \ \ .\end{aligned}$$ Next we notice that we have quadratics in $c_0$ and separately in $(t_0, \ t_\infty)$ to maximize. Maximizing over $c_0$ gives $$c_{0,{\rm max}} = \beta\, \frac{|\Omega|}{ | \partial\Omega| } .$$ Substituting for $c_0$ and considering the function so formed, quadratic in the $t$ variables, we find that its hessian is negative semidefinite, and that there is a unique maximum, at which $t_0+t_\infty=1$. We calculate the maximizing $t$ and denoting them by $t_{\rm max}$, then use $ c_{0,{\rm max}}$ and $t_{0,\rm max}$, $t_{\infty,\rm max}$ to define, by equation (\[eq:vTest\]), a function $v_{\rm max}$. Then ${\cal J}(v_{\rm max})$ is a rational function of $\beta$. Written in a form appropriate for $\beta$ not close to 0, $$\begin{aligned} {Q}(\beta)\ge {\cal J}(v_{\rm max}) &=& R( \frac{|\Omega|^2}{|\partial\Omega|} ,\Sigma_\infty,\Sigma_1,{Q}_0;\beta), \label{in:lbbL}\\ R( \frac{|\Omega|^2}{|\partial\Omega|} ,\Sigma_\infty,\Sigma_1,{Q}_0;\beta) &=& \left(\beta\, \frac{|\Omega|^2}{|\partial\Omega|} + \Sigma_\infty +\frac{1}{\beta} \Sigma_1\right) + \frac{\Sigma_1^2}{\beta^2(\Sigma_\infty - {Q}_0) -\beta\Sigma_1} . \nonumber \end{aligned}$$ (On using $\Sigma_1\le{0}$ we see that this is an improvement on the left-hand side of [@KM93] (4.12).) In a form appropriate for $\beta$ small, the rational function $R$ can be rewritten to give: $${Q}(\beta)\ge {\cal J}(v_{\rm max}) = {Q}_0 + \beta\, \frac{|\Omega|^2}{|\partial\Omega|} + \frac{\beta(\Sigma_\infty-{Q}_0)^2}{\beta(\Sigma_\infty-{Q}_0)-\Sigma_1} . \label{in:lbbS}$$ This inequality improves on (4.4) of  [@KM93]. The function ${\cal J}(v_{\rm max})$ is increasing in $\beta$, a property which it shares with ${Q}(\beta)$. ${\cal J}(v_{\rm max})$ is concave in $\beta$: $$\frac{\partial^2{\cal J}(v_{\rm max})}{\partial\beta^2} = \frac{2\Sigma_1 (\Sigma_\infty - {Q}_0)^3}{(\beta (\Sigma_\infty -{Q}_0)-\Sigma_1)^3} .$$ Quadratic test functions, and a simple first case {#subsec:VarlGen} ------------------------------------------------- For domains like ellipses symmetric about both the $x$-axis and $y$-axis one nice choice for $v$ are the quadratic functions $$\leqno{(i)}\qquad\qquad\qquad\qquad\qquad v = c_0+c_2\left( (\frac{x}{a})^2 +a^2 y^2\right) ,$$ and, more generally (and with a different $c_0$) $$\leqno{(ii)}\qquad\qquad\qquad\qquad\qquad v = c_0+c_{xx} x^2 + c_{yy} y^2 .$$ (Case (ii) is motivated by the likelihood that for a range of $\beta$ and a range of $a$ the level curves of $u$ look to be like ellipses. See, for example, Figure 4 of [@SV12] and our own contour plots shown in Appendix \[sec:Numerical\].) We briefly return to general domains. An application of the Divergence Theorem give $${\rm if\ \ } -\Delta v=1\ \ {\rm then\ \ } {\cal J}(v)-{\cal Q}(v) = \int_{\partial\Omega} v (v+\beta \frac{\partial v}{\partial n} ) \label{eq:JmQDgce}$$ [**Case (i)**]{} We begin with case (i). While the beginning of the study with case (ii) will allow $\Omega$ to be more general than merely an ellipse, for case (i) we restrict to the ellipse. The final result, lower bound, will give the same result as we reported in §\[subsec:GovEq\] from equation (4.4) of [@KM93]. Write $J$ or $J(c_0,c_2)$ for the value of $\cal J$ evaluated at the quadratic function (i) $$J = 2 c_0 A+ 2 c_2\left( \frac{1}{a^2} (1-\frac{2 c_2}{a^2}) I_{XX} +a^2 (1- 2 a^2 c_2)I_{YY}\right) -(c0+c2)^2 \frac{L}{\beta} .$$ Finding the gradient of $J$ with respect to $[c_0,c_2]$ we find the maximum of $J$ will occur when the gradient is 0, i.e. $$\begin{aligned} L c_0 + L c_2 &=& \pi \beta ,\\ 2L c_0 + c_2\left( 2L +\pi\beta(a^2+\frac{1}{a^2})\right) &=& \pi\beta .\end{aligned}$$ This is readily solved, and on integrating, yields the result given before at inequality (\[KM93eq4p4\]). It is straightforward to show that the quadratic $u$ which is the variational winner in Case (i) is such that the result in (\[eq:JmQDgce\]) can be applied to establish ${\cal J}(u)={\cal Q}(u)$ (a fact which can be established by other means). The application of (\[eq:JmQDgce\]) is facilitated by the fact that the quadratic $u=u_0 +\beta|\Omega|/{|\partial\Omega|}$ is constant on $\partial\Omega$. [**Case (ii), an introduction**]{} Next we begin to treat case (ii). Write $J$ or $J(c_0,c_{xx},c_{yy})$ for the value of $\cal J$ evaluated at the quadratic function (ii). $$\begin{aligned} J &=& 2\left( c_0 A + c_{xx}(1-2 c_{xx}) I_{xx} + c_{yy} (1-2 c_{yy}) I_{yy} \right)-\\ &\ &\frac{1}{\beta}\left( c_0^2 L + 2 c_0 c_{xx} i_{xx}+ 2 c_0 c_{yy} i_{yy} + 2 c_{xx} c_{yy} i_{xxyy} + c_{xx}^2 i_{xxxx} + c_{yy}^2 i_{yyyy} \right) \ .\end{aligned}$$ where $A$ is the area, $L$ the perimeter and the $I$ are area moments and $i$ boundary moments $$I_{xx}=\int_\Omega x^2, \ \ I_{yy}=\int_\Omega y^2 ,$$ and $$i_{xx}= \int_{\partial\Omega} x^2 , \ \ i_{yy}= \int_{\partial\Omega} y^2 , \ \ i_{xxyy}= \int_{\partial\Omega} x^2 y^2 , \ \ i_{xxxx}= \int_{\partial\Omega} x^4 , \ \ i_{yyyy}= \int_{\partial\Omega} y^4 .$$ Define also $$Q_q = c_0 A + c_{xx}I_{xx} + c_{yy}I_{yy} , \label{eq:Qq}$$ the subscript $q$ reminding us of the quadratic approximation to the velocity field. The gradient of $J$, here denoted $\mathbf g$, is $${\mathbf g} = -\frac{2}{\beta} \left( \begin{array}{c} c_0 L + c_{xx} i_{xx} + c_{yy} i_{yy} - \beta A\\ c_0 i_{xx} +c_{xx} (4\beta I_{xx} + i_{xxxx}) + c_{yy} i_{xxyy} - \beta I_{xx}\\ c_0 i_{yy} +c_{xx} i_{xxyy} +c_{yy} (4\beta I_{yy} + i_{yyyy}) - \beta I_{yy} \end{array} \right)\ .$$ We remark that $$2(J -Q_q) = {\mathbf g}[1] c_0 + {\mathbf g}[2] c_{xx} + {\mathbf g}[3] c_{yy} . \label{eq:JminusQ}$$ This gradient $\mathbf g$ is zero when $$\begin{aligned} c_0 L + c_{xx} i_{xx} + c_{yy} i_{yy} &=& \beta A , \\ c_0 i_{xx} +c_{xx} (4\beta I_{xx} + i_{xxxx}) + c_{yy} i_{xxyy} &=& \beta I_{xx} , \\ c_0 i_{yy} +c_{xx} i_{xxyy} +c_{yy} (4\beta I_{yy} + i_{yyyy}) &=& \beta I_{yy} .\end{aligned}$$ Up until now the $c_0$, $c_{xx}$, $c_{yy}$ have been general. Henceforth we use the same letters to denote the $c$ obtained by solving ${\mathbf g}=0$, the gradient of $J$ equals zero. We remark that after finding $c_0$, $c_{xx}$, $c_{yy}$ there are two methods of estimating ${Q}$. On the one hand, we can estimate ${Q}$ from $$Q_c := c_0 A + c_{xx} I_{xx} + c_{yy} I_{yy} , \label{eq:Qc}$$ while, on the other hand, we have the lower bound $$J_c := J(c_0,c_{xx},c_{yy})\leq{{Q}} .$$ For general $\Omega$ inequalities on moments (such as some in references in [@Ke06]) may lead to lower bounds for $Q$ in terms of simpler geometric quantities. Also for some domains, polygons for example, the various moments could be calculated. In general the restriction of the test functions to quadratics isn’t likely to give very tight lower bounds though there may be exceptions. Rectangles at large $\beta$ might be one such as $u_\infty$ (see [@KM93] or equations (\[eq:uInf\])) is, for rectangles, quadratic. For an ellipse with $a>1$ (and $b=1/a$) the area moments are elementary, but the perimeter and other boundary moments involve elliptic integrals: we present these in the next subsection. For a circle, all the moments are elementary. For a circle of radius 1, the moments are: $$\begin{aligned} A &=&\pi,\ \ I_{xx}=\frac{\pi}{4}, \ \ I_{yy}=\frac{\pi}{4} , \\ L &=&2\pi, \ \ i_{xx}=\pi=i_{yy}, \ \ i_{xxxx}=\frac{3\pi}{4}=i_{yyyy}, \ \ i_{xxyy}=\frac{\pi}{4} .\end{aligned}$$ For the circle, the variational winner has $c_0$, $c_{xx}=c_{yy}$ giving the exact solution previously presented in equation (\[eq:circuSteady\]): both $Q_c$ and $J_c$ evaluate to the formula for ${Q}$ given in equation (\[eq:circleQinf\]). A note of caution concerning the behaviour when $\beta=0$ is appropriate here. If one naively sets $\beta=0$ in the system of equations ${\rm grad}(J)=0$ the right-hand sides are all zero. The solution with all the $c$ equal to zero incontravertably gives a lower bound on ${Q}$: the lower bound being zero. Of course if $\beta=0$ the term ${\cal B}(v)/\beta$ is the difficulty for functions $v$ not vanishing identically over the whole boundary. As a quadratic function vanishing on the boundary provides the minimizer for ellipses with $\beta=0$ it seems appropriate, at this stage, to return to the specifics for an ellipse. Variational methods for the ellipse, Case (ii) continued {#subsec:VarlEll} -------------------------------------------------------- For the ellipse we can write $i_{xx}$ and $i_{yy}$ in terms of the polar second moment about the centre, $i_2=i_{xx}+i_{yy}$, and the perimeter $L$ Similarly the 4-th moments can all be written in terms of the polar fourth moment about the centre, $i_4$ and $i_2$ and $L$. Each of $L$, $i_2$ and $i_4$ can be integrated in terms of elliptic integrals. The variational winner over the quadratics gives lower bounds on $Q$ as in the following code. ${Q}$ for nearly circular ellipses {#sec:nearCircQs} ================================== ${Q}$ at $\beta=0$ {#subsec:nearCircQbeta0} ------------------ Now, at $\beta=0$, the explicit solution as given in §\[sec:explicit\], yields asymptotics for $\epsilon\rightarrow{0}$: $$\begin{aligned} {Q}&= \frac{\pi}{4(a^2+a^{-2})} = \frac{\pi}{8}\, \sqrt{1-\epsilon^2} , \nonumber \\ &\sim \frac{\pi}{8}\left(1 - 2(a-1)^2+ O( (a-1)^3)\ \right) \quad {\rm as\ }\ a\rightarrow 1 , \nonumber \\ &\sim \frac{\pi}{8}\left( 1 - \frac{1}{2}\epsilon^2+ O( \epsilon^3) \ \right) \quad {\rm as\ }\ \epsilon\rightarrow 0, \label{eq:PexactEps} $$ Nearly circular ellipses with $\beta\ge{0}$ {#subsec:nearCircbetaGt0} ------------------------------------------- The explicit solution available for when $\beta=0$ is presented above, and is useful to check against our asymptotics for the solution when $\beta\ge{0}$. When $\epsilon\rightarrow{0}$, equation (\[eq:Qellb0\]) agrees with the asymptotics given earlier in equation (\[eq:PexactEps\]). Concerning the ellipse when $\beta>0$, when $\epsilon$ is small, $u$ can be approximated in the form $$u = \frac{1}{4}\left( 1-r^2 \right) + \frac{\beta}{2} +\epsilon^2\, t_{02} + \epsilon \, t_{11} r^2 \cos(2\theta) + \epsilon^2 \, t_{22} \, r^4\, \cos(4\theta) . \label{eq:ueps}$$ Details of the perturbation analysis are in [@KW16]. The result of the perturbation analysis is $$\begin{aligned} t_{11} &= \frac{1}{4}\, \frac{1+\beta}{1+2\beta} , \\ t_{02} &= -\frac{1}{32}\, \frac{4 + 5\beta +6\beta^2}{1+2\beta} = -\frac{1}{32}\,\left( 1+3\beta + \frac{3}{1+2\beta}\right) , \\ t_{22} &= -\frac{1}{32}\, \frac{\beta (1-2\beta)}{(1+4\beta)(1+2\beta)} .\end{aligned}$$ Integrating $u$ with these parameters over the ellipse gives the expansion for ${Q}({\rm ellipse})$: $$\begin{aligned} {Q}({\rm ellipse}) &\sim& \frac{\pi}{8} (1+4\beta) + q_1\epsilon^2 \label{eq:QsepsSmall}\\ q_1 &=& - \frac{\pi}{16} \left(1 - 8 t_{11} - 16 t_{02}\right) = - \frac{\pi}{16} \left(1 +{\frac {\beta\, \left( 1+6\,\beta \right) }{2(2\,\beta+1)}}\right) . \label{eq:q1Def}\end{aligned}$$ Maple code for the near-circular approximation is as follows: ### Near circular, $\beta$ small {#near-circular-beta-small .unnumbered} When $\beta={0}$, this agrees with the asymptotics given in equation (\[eq:PexactEps\]). When $\beta$ is small but non-zero, asymptotics given in §\[subsec:betaSmallEll\] check: see equation (\[eq:Q1bSmallNC\]). ### Near circular, $\beta$ large {#near-circular-beta-large .unnumbered} When we first take $\epsilon$ small, then let $\beta$ tend to infinity we have the approximation $$Q\sim \frac{\pi}{2}\beta +\frac{\pi}{8} -\frac{3\pi}{32}\beta\epsilon^2 -\frac{\pi}{32}\epsilon^2 .$$ This is used as a check in §\[subsec:Qinf\]. ### Numerical results {#numerical-results .unnumbered} Table \[tbl:nearCirc\] below compares the previous asymptotics (column A) with variational lower bound (column V) and with the Fourier series solution (column F). ----------- ------------- ------------- ------------- $a=5/4$ $\beta$ A V F 1/64 .3825119944 .3807427556 .3807440330 1/16 .4551128646 .4530404184 .4530434724 1/4 .7437766793 .7406928885 .7406944966 1 1.888863782 1.882275712 1.882277797 4 6.450075877 6.429038519 6.429053564 16 24.68100694 24.60087273 24.60089638 64 97.59955268 97.28234644 97.28237296 $a=17/16$ $\beta$ A V F 1/64 .4143605368 .4143493072 .4143493804 1/16 .4879061200 .4878930227 .4878930894 1/4 .7819440804 .7819247200 .7819248140 1 1.957301880 1.957260968 1.957261120 4 6.657144997 6.657014853 6.657015098 16 25.45536249 25.45486728 25.45486742 64 100.6478027 100.6458431 100.6458432 $a=65/64$ $\beta$ A V F 1/64 .4170525380 .4170525625 .4170524896 1/16 .4906779729 .4906782880 .4909655446 1/4 .7851701838 .7851700411 .8350137728 1 1.963086617 1.963085377 1.963085489 4 6.674647535 6.674645198 6.675381808 16 25.52081497 25.52081128 25.51131048 64 100.9054563 100.9054466 100.9054478 ----------- ------------- ------------- ------------- \[tbl:nearCirc\] [**A simpler, but less accurate, variational approach**]{} ![§\[subsec:nearCircbetaGt0\]. Plot of Qratio at various values of $\beta$ for small eccentricity.[]{data-label="fig:QratioEpsSmalljpg"}](\grpath/QratioEpsSmalljpg.jpg){height="7cm" width="13cm"} ![§\[subsec:nearCircbetaGt0\]. Plot of small beta approximation term $Q_1$ in green and the variational lower bound.[]{data-label="fig:Q1bndBetaSmalljpg"}](\grpath/Q1bndBetaSmalljpg.jpg){height="7cm" width="13cm"} The lower bound shown in Figure \[fig:Q1bndBetaSmalljpg\] comes from a trial function which is just the $\beta=0$ quadratic in Cartesians plus the (best) constant. Of course this is exact for eccentricity zero. Of course we do better with a more general quadratic but the integrals are more ugly. Flows with $\beta$ small {#sec:betaSmall} ======================== General $\Omega$ {#subsec:betaSmallGenOm} ---------------- The asymptotics are described in [@KM93]. The problem is $$\begin{aligned} - \Delta u &=& 1 \qquad {\rm in}\ \ \Omega ,\\ \beta \frac{\partial u}{\partial n} + u &=& 0 \qquad {\rm on}\ \ \partial\Omega .\end{aligned}$$ When $\beta$ is small, the solution $u_\beta$ is asymptotically $$u_\beta \sim u_{\beta=0} + \sum_{j=1}^n \beta^j u_j ,$$ with each $u_j$ ($j\ge{1}$) harmonic (and independent of $\beta$). Our interest is in ${Q}$, here denoted $Q_\beta$, $$Q_\beta = \int_\Omega u_\beta ,$$ so $$Q_\beta \sim Q_0 +\beta Q_1\qquad{\rm where}\ \ Q_1=\int_\Omega u_1 ,$$ and the integral of $u_1$ over $\Omega$ can be found explicitly (at least up to a single-variable integral) even when $u_1$ cannot be found explicitly. Now $$\begin{aligned} - \Delta u_1 &=& 0 \qquad {\rm in}\ \ \Omega ,\\ \beta u_1 &=& -\beta \frac{\partial u_0}{\partial n} \qquad {\rm on}\ \ \partial\Omega .\end{aligned}$$ The Divergence Theorem gives $$\begin{aligned} Q_1=\int_\Omega u_1 &=& \int_\Omega\left( u_0 \Delta u_1 - u_1\Delta u_0\right) = \int_\Omega {\rm div}\left( u_0\nabla u_1 - u_1\nabla u_0\right) ,\\ &=& \int_{\partial\Omega}\left( u_0 \frac{\partial u_1}{\partial n}. - u_1 \frac{\partial u_0}{\partial n}\right) = \int_{\partial\Omega}\left( \frac{\partial u_0}{\partial n}\right)^2 \, ds .\end{aligned}$$ We remark $$\int_{\partial\Omega}\, ds=|\partial\Omega|, \qquad \int_{\partial\Omega}\left( \frac{\partial u_0}{\partial n}\right) \, ds= -|\Omega| .$$ This yields, via Cauchy-Schwarz, $$|\Omega|^2 \le |\partial\Omega|\ Q_1 ,$$ an inequality which becomes more accurate the closer the ellipse is to a circle. $\Omega$ an ellipse {#subsec:betaSmallEll} ------------------- In general, and in the notation of the preceding subsection, $$Q_1 = \int_{\partial\Omega}\left( \frac{\partial u_0}{\partial n}\right)^2 \, ds = \int_{\partial\Omega}\, |\nabla u_0|^2 \, ds .$$ For an ellipse $u_0=u_{0e}$ with $u_{0e}$ given in §\[sec:explicit\]: thus $$Q_{1,e} = \int_{\partial\Omega}\, |\nabla u_{0e}|^2 \, ds .$$ Although the calculation can be done in polar coordinates, perhaps Cartesians lead to simpler integrals: $$|\nabla u_{0e}|^2 = \frac{ {a^{-4}}x^2 + a^4 y^2}{(a^{-2}+a^2)^2} .$$ The Cartesian equation for that part of the ellipse in the first quadrant is $$Y(x)= \frac{1}{a}\, \sqrt{1-(\frac{x}{a})^2} , \qquad{\rm for\ \ } 0<x<a .$$ The perimeter is $$\begin{aligned} |\partial\Omega| &=& 4 \int_0^a \sqrt{1+ Y'(x)^2}\, dx , \nonumber\\ &=& 4 a {\rm EllipticE}\left(\frac{\sqrt{a^4-1}}{a^2}\right) \quad{\rm for\ \ } a\ge{1} , \nonumber\\ &=& 4 a {\rm EllipticE}(e) \quad{\rm for\ \ } a\ge{1} , \label{eq:perimE}\\ &=& 4 (1-e^2)^{-1/4}\, {\rm EllipticE}(e) , \nonumber\end{aligned}$$ where $e$ is the eccentricity as defined in equation (\[eq:eaDef\]).\ We now calculate $Q_1$, beginning with $$|\nabla u_{0e}|^2 = \frac{ a^2 - a^{-2} (a^2 -a^{-2}) x^2}{(a^{-2}+a^2)^2} \qquad{\rm on\ \ } y=Y(x) .$$ On integrating this $Q_1$ is given by $$Q_1 = \frac{4}{3}\, \frac{a^3}{(1+a^4)^2}\left( 2(1+a^4){\rm EllipticE}(e)- {\rm EllipticK}(e)\right) \quad {\rm for\ \ } a\ge{1} . \label{eq:Q1bSmall}$$ ### $\beta$ small, near circular {#beta-small-near-circular .unnumbered} The asymptotic expansion for nearly circular ellipse, $a$ near 1, has $$Q_1 \sim \frac{\pi}{2} - \frac{\pi}{8} (a-1)^2\qquad {\rm for\ \ } a\rightarrow{1} . \label{eq:Q1bSmallNC}$$ This agrees with the earlier expansion given in equations (\[eq:QsepsSmall\],\[eq:q1Def\]) where, for $\beta$ small $${Q}\sim \frac{\pi}{8}(1+ 4\beta)- \frac{\pi}{4}(1+ \frac{\beta}{2}) (a-1)^2 \ \quad{\rm for }\ \ a\rightarrow{1}, \ \beta\rightarrow{0} .$$ ### Numerical results {#numerical-results-1 .unnumbered} In Table \[tbl:betaSmall\] below we compare with the numerical values given by the lower-bound variational winner amongst quadratics. -------------- ---------------- ---------------- ---------------- $\beta=1/4$ $a$ A V F 65/64 .7851858133 .7851702110 .8350137728 17/16 .7821607512 .7819248090 .7819248140 5/4 .7432200201 .7406928855 .7406944966 2 .4953609194 .4909982744 .4910897996 4 .2147581278 .2139371008 .2142092926 16 .4473378672e-1 .4463047672e-1 .4473097574e-1 $\beta=1/16$ $a$ A V F 65/64 .4906792259 .4906780838 .4909655446 17/16 .4879127252 .4878930761 .4878930894 5/4 .4532504638 .4530404192 .4530434724 2 .2624399058 .2620398525 .2621170972 4 .9036181975e-1 .9023065471e-1 .9032360830e-1 16 .1348438275e-1 .1345885965e-1 .1348419675e-1 $\beta=1/64$ $a$ A V F 65/64 .4170525792 .4170526698 .4170524896 17/16 .4143507188 .4143493767 .4143493804 5/4 .3807580746 .3807427560 .3807440330 2 .2042096524 .2041611381 .2041881644 4 .5926274272e-1 .5923472697e-1 .5926025316e-1 16 .5672031762e-2 .5665670014e-2 .5671955380e-2 -------------- ---------------- ---------------- ---------------- \[tbl:betaSmall\] $u_1$ for $\beta$ small: Fourier series? ---------------------------------------- We begin by recalling (ref[eq:12b]{}) $$v_p\;=\; -\frac{c^2}{8} (\cosh(2\eta)+\cos(2\psi)) = -\frac{1}{4} (x^2+y^2) .$$ Several items relating to $v_p$ are used, in particular its normal derivative at the boundary, the $\eta$-derivative being closely related. The diagonal matrix $M_0$ has as its $(n,n)$ entry $$m_{0,0}= 2\pi , \quad{\rm and\ for\ }n\geq{1}\quad m_{0,n}=\pi\,\cosh(2n\eta_0).$$ The first subscript, here $0$, reminds us that the entries come from $M_0$. The vector $F_0$ has nonzero entries only at $n=0$ and $n=1$ when $$f_{0,0}= \frac{c^2\pi}{4}\, \cosh(2\eta_0) \quad{\rm and\ } f_{0,1}= \frac{c^2\pi}{8} .$$ Again, the first subscript, here $0$, reminds us that the entries come from $f_0$. The symmetric matrix $M_1$ is full with entries involving EllipticK and EllipticE. The vector $f_1$, related to the Fourier series of $g$, similarly involves EllipticK and EllipticE. For $\beta$ positive but small, we seek a representation of the form $$u \sim u_0 + \beta \, u_1\qquad{\rm for}\ \ \beta\rightarrow{0} .$$ Our main interest is checking with an asymptotic result obtained using Cartesian coordinates: $$Q \sim Q_0 + \beta \, Q_1\qquad{\rm for}\ \ \beta\rightarrow{0} .$$ Our calculation in Cartesian coordinates was completed a year earlier than the elliptic coordinate work here. ### $\beta=0$ re-visited {#beta0-re-visited .unnumbered} As one small check on the equations determining the $A_n$, we note that when $\beta=0$ the equations are very easy to solve. The matrix $M=M_0$ is diagonal. The only nonzero $A_n$ are those corresponding to $n=0$ and $n=1$.. When $\beta=0$, and subscripting with the first 0 in a pair merely reminding us that the values are for $\beta=0$, $u_0= v_p +A_{0,0} +A_{0,1}\cosh(2\eta)\cos(2\psi)$ and the equation $M_0 A= f_0$ gives $$\begin{aligned} A_{0,0} &=& \frac{f_0}{m_{0,0}}= \frac{c^2\, \cosh(2\eta_0) }{8}=\frac{\cosh(2\eta_0)}{4\sinh(2\eta_0)} \\ A_{0,1} &=& \frac{f_1}{m_{0,1}}= \frac{c^2}{8\cosh(2\eta_0) }= \frac{1}{4\cosh(2\eta_0)\sinh(2\eta_0)} .\end{aligned}$$ As we have already calculated $Q_0$ from other methods, the calculation here merely serves as a check on $A_{0,0}$ and $A_{0,1}$. ### $\beta>0$ small {#beta0-small .unnumbered} Consider now approximations when $\beta$ is small. In §[\[sec:betaSmall\]]{} we find $$Q\sim Q_0 +\beta\, Q_1\qquad{\rm as}\ \ \beta\rightarrow{0} ,$$ with $Q_0$ as in equation (\[eq:Qellb0\]). The approach began with assuming $$u\sim u_0 +\beta\, u_1\qquad{\rm as}\ \ \beta\rightarrow{0} ,$$ but did not require finding $u_1$. If one wished to find $u_1$, a Fourier series approach as in this section might be appropriate. Approximate the vector $A$ of Fourier coefficients by $$A \sim A_0 + \beta \, A_1 ,$$ where the vector $A_0$ is that already calculated in the preceding subsection concerning $\beta=0$. Then we require, to order $\beta$, $$(M_0+\beta M_1)(A_0 + \beta A_1) = (f_0+ \beta f_1) ,$$ i.e., on approximating to order $\beta$, $$M_0 A_1 = f_1 -M_1 A_0 .$$ Equivalently, using, as noted previously, that $M_0$ is diagonal, $$\begin{aligned} m_{0,0} A_{1,0} &=& f_{1,0} - m_{1,0,0} A_{0,0} - m_{1,0,1} A_{0,1},\\ m_{0,1} A_{1,1} &=&f_{1,1} -m_{1,1,0} A_{0,0} - m_{1,1,1} A_{0,1} \\ \vdots &=& \vdots \\ m_{0,n} A_{1,n} &=&f_{1,n} -m_{1,n,0} A_{0,0} - m_{1,n,1} A_{0,1}\end{aligned}$$ Again the first subscript denotes the appropriate order in $\beta$. While the calculations are routine, it may be that the solution $A_1$ is ugly, and, in any event, the human effort at keeping all the details correct is considerable. We suspect that, at this stage in the study and use of microchannels, the greater accuracy for $u$ is not needed. A calculation which would provide an additional check is to finding $A_{1,0}$ and $A_{1,1}$ which is all that is needed to check against the result for $Q_1$ given in equation (\[eq:Q1bSmall\]).\ $\beta$ large {#sec:betaLarge} ============= The dominant term ----------------- For general $\Omega$ $$Q(\beta)\sim \beta\frac{|\Omega|^2}{|\partial\Omega|} \qquad{\rm for\ \ } \beta\rightarrow\infty .$$ For our ellipses of area $\pi$ the perimeter is given by equation (\[eq:perimE\]), Hence $$Q(\beta) \sim \beta\frac{\pi^2}{4 a {\rm EllipticE}(e)} \qquad{\rm for\ \ } \beta\rightarrow\infty .$$ Maple code for this is: betaLargeQfn:= (a,beta) -> beta*Pi^2/(4*a*EllipticE(sqrt(1-1/a^4))); ### $\beta$ large, nearly circular ellipses {#beta-large-nearly-circular-ellipses .unnumbered} Further asymptotic approximation of that above for nearly circular ellipses is: $$Q(\beta) \sim \beta\left(\frac{\pi}{2} - \frac{3\pi}{8} (a-1)^2 + O((a-1)^3)\quad \right) \qquad{\rm for\ \ } \beta\rightarrow\infty, \ a\rightarrow 1 .$$ This may be compared with the asymptotics found in equations (\[eq:QsepsSmall\],\[eq:q1Def\]) further approximated for $\beta$ large. Then $q_1\sim -\pi\beta/32$ so $${Q}\sim \frac{\pi\beta}{2}+q_1\epsilon^2 \sim \frac{\pi\beta}{2} - \frac{3\pi\beta}{32}\epsilon^2 \sim \frac{\pi\beta}{2} - \frac{3\pi\beta}{8}(a-1)^2 ,$$ which agrees with the result of the preceding paragraph. We have yet to find higher order approximations. The variational approximation using quadratic test functions agrees with the lowest order term. This first term $O(\beta)$ is, at fixed $\beta$ a decreasing function of $a$ on $a>1$. Note that the quadratic lower bound in column V in the following table, Table \[tbl:betaLarge\], is a better approximation than $\beta |\Omega|^2/|\partial\Omega|$ which is less than the lower bound.\ ### Numerical results, Table \[tbl:betaLarge\] {#numerical-results-tabletblbetalarge .unnumbered} ------------ ------------- ------------- ------------- $\beta=4$ $a$ A V F 65/64 6.282052744 6.674647024 6.675381808 17/16 6.265911941 6.657014849 6.657015098 5/4 6.056752387 6.429038511 6.429053564 2 4.602060688 4.869625399 4.870154656 4 2.449872080 2.627235185 2.627747294 16 .6168200091 .6676825774 .6691074764 $\beta=16$ $a$ A V F 65/64 25.12821097 25.52081287 25.51131048 17/16 25.06364776 25.45486727 25.45486742 5/4 24.22700955 24.60087275 24.60089638 2 18.40824275 18.69310262 18.69427704 4 9.799488320 10.09964112 10.10542728 16 2.467280036 2.656743920 2.661123514 $\beta=64$ $a$ A V F 65/64 100.5128439 100.9054480 100.9054478 17/16 100.2545911 100.6458431 100.6458432 5/4 96.90803820 97.28234641 97.28237296 2 73.63297100 73.92334640 73.92489450 4 39.19795328 39.57699513 39.59760516 16 9.869120146 10.54995058 10.55945235 ------------ ------------- ------------- ------------- : \[tbl:betaLarge\] The next term, and $u_\infty$, for ellipses in general {#subsec:Qinf} ------------------------------------------------------ We seek the term $\Sigma_\infty$ (independent of $\beta$) in the asymptotic expansion $$Q \sim \frac{\beta\, |\Omega|^2}{|\partial\Omega|} + \Sigma_\infty \qquad {\rm for\ \ } \beta\rightarrow\infty .$$ The notation here is, as in [@KM93], $$\Sigma_\infty = \int_\Omega u_\infty ,$$ with $u_\infty$ satisfying equations (\[eq:uInf\]). One result which follows as a limiting case of the isoperimetric result of [@BG15] is the following.\ [*Amongst all ellipses with a given area, that which maximizes $\Sigma_\infty$ is the circular disk.*]{}\ We have yet to investigate the possibility of a proof of this from the series we have for $\Sigma_\infty$. (In the case of a rectangle, we have proved that $\Sigma_\infty$ is maximized by the square: see [@KWK18].) The approach here is to seek a Fourier series solution for $u_\infty$, investigate its contour plot, and after this to integrate $u_\infty$ thereby finding the next term in the expansion for $Q$. The integration task only requires the constant ($n=0$) and the next ($n=1$) Fourier coefficients of $u_\infty$. The tasks have been completed for nearly circular ellipses, but the general case, specifically determining the $n=0$ constant coefficient of $u_\infty$, remains ‘work in progress’. We recast the elliptic coordinates Fourier series approach so that one uses the Fourier series of $1/g$ rather than $g$. Write the Fourier series of $1/g$ as $$\frac{1}{g(s)} = \frac{{\hat g}_0}{2} + \sum_{n=1}^\infty {\hat g}_n \cos(2 n s) . \label{eq:rgFSB}$$ As with the earlier calculation of the Fourier series of $g$, the coefficients ${\hat g}_n$ can be found in terms of EllipticE and EllipticK functions. In particular, with the notation of (\[eq:Ellnotn\]), we have $$\begin{aligned} |\partial\Omega| &=& 4 c\, \cosh(\eta_0)\, {\rm EllipticE}_0 , \nonumber\\ {\hat g}_0 &=& \frac{4}{\pi} \, {\rm EllipticE}_0 = \frac{1}{\pi\, c\, \cosh(\eta_0)} \, |\partial\Omega| \label{eq:gBar0Ell}\\ {\hat g}_1 &=& \frac{4}{3\pi} \, \left( -\cosh(2\eta_0) {\rm EllipticE}_0 + (\cosh(2\eta_0) -1) {\rm EllipticK}_0 \right) , \label{eq:gBar1Ell}\end{aligned}$$ and all the higher ${\hat g}_n$ can also be expressed similarly in terms of ${\rm EllipticE}_0$ and ${\rm EllipticK}_0$: $${\hat g}_n =\frac{4}{\pi}\left( {\hat E}_n \, {\rm EllipticE}_0 + {\hat K}_n\, {\rm EllipticK}_0 \right), \label{eq:gBarnEK}$$ where ${\hat E}_n$ and ${\hat K}_n$ are polynomials of degree $n$ in $q=\cosh(\eta_0)^2$ with rational number coefficients. Again there is a three term recurrence relation exists to determine the polynomials ${\hat E}_n$ and ${\hat K}_n$. In the notation of equation (\[eq:qEta0\]), the ${\hat K}_n$ sequence of polynomials starts with $${\hat K}_0=0 , \qquad {\hat K}_1= \frac{2}{3} q =\frac{1}{3} ( q_2 -1).$$ The ${\hat E}_n$ sequence of polynomials starts with $${\hat E}_0=1 , \qquad {\hat E}_1= -\frac{1}{3} (2 q-1)=-\frac{1}{3} q_2 .$$ See Appendix \[app:Furtherg\], and the recurrence (\[eq:gBarnrec\]). As in [@KM93], $u\sim~\beta|\Omega|/|\partial\Omega|+u_\infty$ where $$-\Delta u_\infty = 1\ {\rm and\ } \frac{\partial u_\infty}{\partial n}= -\frac{|\Omega|}{|\partial\Omega|}, \ \ \int_{\partial\Omega} u_\infty = 0 . \label{eq:uInf}$$ Set $$u_\infty = v_\infty + v_p ,$$ with $v_p$ as in (\[eq:12b\]) before. Once again the harmonic function $v_\infty$ can be represented as a Fourier series $$v_\infty = \sum_{n=0}^\infty V_n \cosh(2n\eta)\cos(2n\psi) . \label{eq:vInfDef}$$ To determine $V_n$ for $n\ge{1}$ we will need $$\frac{\partial v_\infty}{\partial \eta}(\eta_0,\psi) = \sum_{n=1}^\infty 2n\,V_n \sinh(2n\eta_0)\cos(2n\psi) \qquad{\rm on\ } \eta=\eta_0.$$ The Neumann boundary condition is $$\frac{g(\psi)}{c\cosh(\eta_0)} \frac{\partial u_\infty}{\partial \eta}(\eta_0,\psi) = \frac{1}{c\sqrt{\cosh^2\eta_0 - \cos^2 \psi}} \frac{\partial u_\infty}{\partial \eta}(\eta_0,\psi) \;=\; -\frac{|\Omega|}{|\partial\Omega|} .$$ In terms of the harmonic function $v_\infty$ this is $$\frac{\partial v_\infty}{\partial \eta} = -\frac{\partial v_p}{\partial \eta} - \frac{|\Omega|}{|\partial\Omega|}\, \frac{c\cosh(\eta_0)}{g(\psi)} \qquad{\rm on\ } \eta=\eta_0. \label{eq:vInfNormal}$$ Now one finds $$\frac{\partial v_p}{\partial \eta}= -\frac{1}{2} \qquad{\rm on\ } \eta=\eta_0. \label{eq:vpeta}$$ (In Cartesian coordinates we have $\nabla{v_p}\cdot{\nabla{u_0}}$ is constant around $\partial\Omega$, but we seek $u_\infty$ such that $\nabla{u_\infty}\cdot{\nabla{u_0}}$ is a constant multiple of $|{\nabla{u_0}}|$ there.) The integral around the boundary of the normal derivative of a harmonic function must be 0, and this accords with the value of ${\hat g}_0$, the constant term in the Fourier series of $1/g$ to ensure this as combining (\[eq:vpeta\]) and $$\frac{|\Omega|}{|\partial\Omega|}\, {c\cosh(\eta_0)}{{\hat g}_0} = \frac{\pi}{|\partial\Omega|} \, \frac{1}{2\pi} \, |\partial\Omega| =\frac{1}{2}\\$$ in (\[eq:vInfNormal\]) we have the required result. From the Fourier series of $1/g(\psi)$ we can find all the Fourier coefficients (\[eq:vInfDef\]), except the constant term, in the series for $v_\infty$. In particular the term $V_1$ is found from (\[eq:vInfNormal\]) (on using (\[eq:vpeta\]), contributing 0) $$2 V_1 \sinh(2\eta_0) = 0 - \frac{|\Omega|}{|\partial\Omega|}\, {c\cosh(\eta_0)}{{\hat g}_1} = -\frac{\pi\, {\hat g}_1}{4{\rm EllipticE}_0} .$$ Hence $$V_1= - \frac{\pi\, {\hat g}_1}{8{\rm EllipticE}_0\, \sinh(2\eta_0)} . \label{eq:V1inf}$$ Similarly $$V_2= - \frac{\pi\, {\hat g}_2}{16\,{\rm EllipticE}_0\, \sinh(4\eta_0)} , \label{eq:V2inf}$$and, more generally, for $n>0$, $$V_n= - \frac{\pi\, {\hat g}_n} {8\, n\,{\rm EllipticE}_0\, \sinh(2n\eta_0)} . \label{eq:Vninf}$$ [**Contour plots for $u_\infty$**]{} As we have all the Fourier coefficients $V_n$ except $V_0$ we can produce contour plots. We show one of these, that for $a=2$. The contourplot (at $a=2$) is visually indistinguishable from that produced from the quadratic found using the $c_0$, $c_XX$, $c_YY$ values found in §\[subsec:VarlEll\]. Except when $\eta_0$ is small, the $V_n$ will decrease rapidly with $n$, and the plots are essentially those from truncating the Fourier series so that it only includes the $V_1$ term. The numerical solutions given in Figure \[fig:ContourPlotBetajpg\] of Appendix D are, for $\beta\ge{4}$, very much like that of Figure \[fig:contoursuInfaEq2jpg\]. ![$a=2$. Contours of $u_\infty$[]{data-label="fig:contoursuInfaEq2jpg"}](\grpath/contoursuInfaEq2.jpg){height="6cm" width="10cm"} [**Finding $V_0$ and $Q$?**]{} The calculation of $V_0$ remains. We will outline how it can be found as an infinite series, but shall only complete the details for nearly circular ellipses. The condition that the boundary integral of $u_\infty$ is zero remains to be used. Now, with, as usual, $s$ for arc length around the boundary, $\eta=\eta_0$, $$(\frac{d s}{d\psi})^2 ={(\frac{d x}{d\psi})^2+(\frac{d y}{d\psi})^2} = c^2\left(\cosh(\eta_0)^2-\cos(\psi)^2\right) = J(\eta_0,\psi) .$$ The square root of the Jacobian is related to the function $g$: $$\sqrt{J(\eta_0,\psi)}= \frac{c\cosh(\eta_0)}{g(\psi)} = \frac{a}{g(\psi)}.$$ We need to satisfy $$0 = \int_{\partial\Omega} u_\infty = 4 \int_{0}^{\pi/2} (v_p+v_\infty) \, \sqrt{J(\eta_0,\psi)}\, d\psi .$$ The integral $I_{p,\partial}$ around the boundary of $v_p$ is $-i_2/4$ where $i_2$ is the boundary moment of inertia and can also be evaluated directly, $$\begin{aligned} I_{p,\partial} &=& 4 \int_{0}^{\pi/2} v_p \, \sqrt{J(\eta_0,\psi)}\, d\psi = 4a \int_{0}^{\pi/2} v_p \,{\hat g}(\psi)\, d\psi ,\\ &=& - \frac{a^3}{3}\left( (2-e^2) {\rm EllipticE}_0 + (1-e^2) {\rm EllipticK}_0 \right).\end{aligned}$$ When evaluating the boundary integral of $v_\infty$, the square root brings in all the $V_n$. - It may be that there are useful Mean Value results applying to ellipses, and harmonic functions in them, of $$\int_{\partial\Omega} v_\infty - v_\infty(0)\, |\partial\Omega|\ ?$$ It is, of course, zero for a circle. In general it wou’t be zero, as, for example, $$\int_{\partial\Omega} (x^2-y^2) = i_X - i_Y \ne{0} {\rm \ \ when\ \ } a>1 .$$ Nevertheless, the quantity can be related to an integral along the major axis of the ellipse, and [@Ro64] is given as a reference in papers by Symeonidis. - The integral $$I_{\infty,\partial} = 4 \int_{0}^{\pi/2} v_\infty \, \sqrt{J(\eta_0,\psi)}\, d\psi = 4a \int_{0}^{\pi/2} v_\infty \,{\hat g}(\psi)\, d\psi ,$$ can be evaluated, with $V_0$ left as an unknown using Parseval’s Theorem. We could find $V_0$ if we could evaluate $$\sum_{n=1}^\infty \frac{{\hat g}_n^2}{n\tanh(2n\eta_0)} ,$$ which can be recast as a series involving Legendre functions $Q_{n-1/2}^{-1}$: see \[app:Furtherg\]. At each fixed $a>1$ (or equivalently $\eta_0$ or eccentricity $e$) this would be a simple float numerical task. However, we have not yet found a simple formula for $V_0$ and hence $\Sigma_\infty$ as simple as that we have for $Q_1$, in equation (\[eq:Q1bSmall\]). Finally to determine $\Sigma_\infty$ using $u_\infty= v_p + v_\infty$ and recalling that we will only need, as in (\[eq:QFS\]), the $n=0$ and $n=1$ Fourier coefficients of $v_\infty$: $$\Sigma_\infty = - \frac{\pi}{8} \, \frac{\cosh(2\eta_0)}{\sinh(2\eta_0)}+ \pi\, V_0 - \frac{1}{2}\,\pi\, V_1 . \label{eq:QinfFS}$$ Combining this with the dominant term we have $$Q\sim \frac{\beta |\Omega|^2}{|\partial\Omega|} + \Sigma_\infty \qquad{\rm as\ } \beta\rightarrow\infty .$$ We remark that for $a$ large $\Sigma_1$ becomes large. The asymptotics for $Q(a,\beta)$ for $\beta$ large should only include the $\Sigma_1$ term when $\Sigma_1/\beta$ is smaller than $\Sigma_\infty$. In practice this means that, for $a$ large, the large $\beta$ asymptotics are good only for $\beta$ very large. (When the domain is a rectangle, rather than an ellipse, there is the same behaviour and this is very easy to demonstrate as $u_\infty$ is, in Cartesian coordinates, a simple quadratic.) ### $\beta$ large for nearly circular ellipses {#beta-large-for-nearly-circular-ellipses .unnumbered} Recall that the eccentricity $e=1/\cosh(\eta_0)$. From equation (\[eq:gRecipFSnearCirc\]) $${\hat g}_0 = 1-\frac{e^2}{4},\qquad {\hat g}_1 = -\frac{e^2}{4} .$$ Also $${\rm EllipticE}_0 \sim \frac{\pi}{2} - \frac{\pi}{8}\, e^2 - \frac{3\pi}{128}\, e^4 , \qquad {\rm EllipticK}_0\sim \frac{\pi}{2} + \frac{\pi}{8}\, e^2 + \frac{9\pi}{128}\, e^4 ,$$ and $$a=\left(\frac{1}{1-e^2}\right)^{1/4} \sim 1+ \frac{1}{4}\, e^2+ \frac{5}{32}\, e^4 ,\quad c=a\,e ,$$ and $$|\partial\Omega|=4a\, {\rm EllipticE}_0 \sim 2\pi + \frac{3\pi}{32} e^4 .$$ (The terms in $e^4$ in the series for ${\rm EllipticE}_0$ and $a$ are only there to check against the series for $|\partial\Omega|$.) We also note $$\sinh(2\eta_0)=2 \frac{1}{e}\,\sqrt{\frac{1}{e^2}-1} \sim \frac{2}{e^2} -1 - \frac{1}{4}\, e^2 .$$ The Fourier coefficients of $v_\infty$ are needed. From equation (\[eq:V1inf\]) $$V_1= -\frac{\pi\, {\hat g}_1}{8{\rm EllipticE}_0\, \sinh(2\eta_0)} \sim \frac{1}{32} e^4 .$$ [**Comparison with earlier near-circular results**]{} Because, for nearly circular ellipses, the asymptotics for all the $V_n$, $n\geq{1}$ are uncomplicated, and the terms get rapidly smaller in $n$, we can find $V_0$. Maple code gives, for $\beta\rightarrow\infty$ and $e\rightarrow{0}$, $$Q\sim \left( \frac{1}{2}\,\pi -{\frac {3}{128}}\,\pi \,{e}^{4} \right) \beta+\frac{1}{8}\, \pi -{\frac {1}{128}}\,\pi \,{e}^{4}-{\frac {1}{256}}\,{\frac {\pi \,{ e}^{4}}{\beta}} .$$ Another quick look at the $\epsilon$ small solution of §\[subsec:nearCircbetaGt0\], in $\eta$, $\psi$ coordinates, produces the same result. $u_\infty$ for other domains ---------------------------- It seems rarly to happen that both $u_0$ and $u_\infty$ have simple explicit formulae representing them. An exception is the equilateral triangle: see [@KM93; @MK94]. And, of course, the circular disk has $u_0=u_\infty$ and is the only domain for which there is equality. (There may well be other less trivial domains, for example, the semicircle.) For our ellipse, $u_0$ is simply, in Cartesians, a quadratic polynomial, or equivalently has a Fourier series (in elliptic coordinates) with just 2 terms, while $u_\infty$ has a Fourier series with all terms nonzero. The situation with a rectangle is as follows. There is an elaborate Fourier series in Cartesian coordinates for $u_0$. However $u_\infty$ is simply a quadratic in the Cartesian coordinates. Indeed our function $u_0$, given in equation (\[eq:u0Cart\]), is $u_\infty$ for the rectangle $(-a^2,a^2)\times(-b^2/b^2)$. The methods in §\[subsec:VarlGen\] would give good lower bounds for $Q({\rm rectangle})$ when $\beta$ is large. See Appendix \[subsec:VarlRectq\]. (For large $\beta$ asymptotics for a rectangle, see [@KWK18].) Conclusion and Open Problems {#sec:Conclusion} ============================ The solution to the Robin boundary problem has been approached from several directions: Fourier series, variational bounds, and asymptotic approximations. There are several questions that remain. 1. What can be said about asymptotics when the eccentricity tends to 1? The small $\beta$ and the large $\beta$ approximations yield some information but more should be possible. (We remark that at large $\beta$, small eccentricity will cause the Fourier coefficients $V_n$ to decay more slowly.) Matched asymptotic approximations may be appropriate. 2. Can $u_\infty$ be found explicitly, or at least its integral over $\Omega$, $\Sigma_\infty$?\ Given that there are 3-term recurrence relations for the Fourier coefficients of both $g$ and $1/g$, might there be a recurrence relation for the coefficients $V_n$? One might combine equation (\[eq:Vninf\]) with items from Appendix \[app:Furtherg\]. 3. In the asymptotics for $\beta$ tending to zero, we found how $Q$ changed, but can one find tidy formulae giving the departure of $u$ from $u_0$? 4. Might it be possible to make better use of the Fourier series for $g$ and $1/g$? In connection with items 2 and 3 above, we remark that there is a ‘Poisson Integral Formula’ for solving the Dirichlet problem for Laplace’s equation in an ellipse. This follows from the conformal map between ellipse and disk: see [@Kob], p177. See also [@Ro64; @Mi90; @LK14; @SD16]. Appendix: Geometry of ellipses {#sec:GeometryI} ============================== General ellipses {#subsec:GeomGen} ---------------- ### Polar coordinates Consider the ellipse $\frac{x^2}{a^2}+a^2 y^2 \le{1}$. In polar coordinates relative to the centre the boundary curve is $$\begin{aligned} r &= \frac{1}{\sqrt{\frac{\cos(\theta)^2}{a^2} + a^2 \sin(\theta)^2}} = \sqrt{\frac{2}{{a}^{2}+{a}^{-2}-\left( {a}^{2}-{a}^{-2} \right) {\cos(2\theta)}}} \nonumber \\ &= r\left(\frac{\pi}{4}\right) \left( 1 - \frac{a^2-a^{-2}}{a^2+a^{-2}} \cos(2\theta) \right)^{-1/2}\ \ {\rm where}\ \ r\left(\frac{\pi}{4}\right) = \sqrt{\frac{2}{{a}^{2}+{a}^{-2}}} \label{eq:rEllipsec2} \\ &= \sqrt{\frac{1 +\tan^2(\theta)}{a^{-2} + a^2 \tan^2(\theta)} } \ . \label{eq:rEllipset}\end{aligned}$$ In our computations we take $a\ge{1}$.\ Without any assumptions on $a$, the coefficients in the Fourier series for $r(\theta)$ involve elliptic integrals. Our interest in some later sections will be in $a$ near 1. However $(a-1)$ might not be the best perturbation parameter. We have also used, in computations, $$\epsilon =\frac{a^2-a^{-2}}{a^2+a^{-2}} \sim 2(a-1) \qquad {\rm\ as\ }\ a\rightarrow{1}\ ,$$ The ellipse is $$r^2= \frac{\sqrt{1-\epsilon^2}}{1- \epsilon\cos(2\theta)} .$$ We have also used (as have others, e.g. [@Day55]), the eccentricity, as in equation (\[eq:eaDef\]), $$e=\sqrt{1-\frac{1}{a^4}} \sim 2(a-1)^{1/2} \qquad {\rm\ as\ }\ a\rightarrow{1} . \label{eq:eDef}$$ As both $e^2$ and $\epsilon$ are relatively similar low-order rational functions of $a^2$, one readily finds that $$\frac{\epsilon}{e^2}= \frac{1}{2}\, (1+\epsilon) .$$ We remark that the polar equation in terms of the eccentricity is $$r= \frac{1}{a\sqrt{1- (e\cos(\theta))^2}} = \frac{g(\theta)}{a}.$$ Including an $a$ dependence in $r(a,\theta)$ we remark that $r(1/a,\theta)=r(a,\theta+\pi/2)$ and $\epsilon(1/a)=-\epsilon(a)$. The binomial expansions $$\begin{aligned} \frac{r(0)}{r(\pi/4)} &= (1-\epsilon)^{-1/2} = 1+ \sum_{k=1}^\infty \frac{ (2k)! }{2^{2k} (k!)^2}\ \epsilon^k , \nonumber \\ \frac{r(\theta)}{r(\pi/4)} &= (1-\epsilon\cos(2\theta))^{-1/2} = 1+ \sum_{k=1}^\infty \frac{ (2k)! }{2^{2k} (k!)^2}\ \cos(2\theta))^k \epsilon^k , \label{eq:rCosk}\end{aligned}$$ may be useful in finding higher terms in the perturbation expansions (for $a$ near 1) of some domain functionals. The symmetries of the ellipse explain the form of the expansion in (\[eq:ueps\]): - It is symmetric about $\theta=0$, hence only cosine terms. - It is symmetric about $\theta=\pi/2$ and hence only the even order cosine terms $\cos(2 m\theta)$. - When $\epsilon$ is replaced by $-\epsilon$ and $\theta$ by $\theta+\pi/2$ the expression is unchanged and hence the form of the polynomial coefficients in $\epsilon$ forming the Fourier coefficients. For $m$ is odd, only odd powers of $\epsilon$ appear: for $m$ is even, only even powers of $\epsilon$ appear. We see these symmetries in connection the solutions of our pde problem. Returning to the study of the boundary curve, we also need the expansion for $r(\pi/4)$: $$a^2=\sqrt{\frac{1+\epsilon}{1-\epsilon}} , \qquad \frac{2}{r(\pi/4)}= a^2 + a^{-2} =\frac{2}{\sqrt{1-\epsilon^2}},\qquad r\left(\frac{\pi}{4}\right) = \left(1 -\epsilon^2 \right)^{1/4} .$$ There are a few items of undergraduate calculus that are used. Let $s$ denote arclength measured around the curve. Then $$\begin{aligned} \frac{d s}{d\theta} &=& \sqrt{ r(\theta)^2 +\left(\frac{d r(\theta)}{d\theta}\right)^2}\qquad {\rm in \ general}, \\ &=& 2 a\sqrt{\frac{ (1+a^8)-(a^8-1)\cos(2\theta)} {((1+a^4)-(a^4 -1)\cos(2\theta))^3}}\qquad {\rm for \ an \ ellipse} , \\ &=& 2 a\sqrt{\frac{ (1+\epsilon)(1+\epsilon^2-2\epsilon\cos(2\theta))} {(1-\epsilon\cos(2\theta))^3}}\qquad \ \ \ {\rm for \ an \ ellipse} .\end{aligned}$$ We remark that, for the ellipse the normal can be found from the gradient of the ellastic torsion function, $\nabla u_{0e} = \nabla u({\rm ellipse},\beta=0)$. ### The usual parametric description of an ellipse The boundary of the ellipse can be described by $$x=a\cos(\psi),\qquad y=\frac{1}{a}\sin(\psi) . \label{eq:param}$$ (The relation between the parameter $\psi$ and the polar angle $\theta$ is $\tan(\theta)=\tan(\psi)/a^2$.) $$\begin{aligned} (\frac{ds}{d\psi})^2 &=& (\frac{dx}{d\psi})^2 + (\frac{dy}{d\psi})^2 ,\\ &=& a^2 \sin(\psi)^2 + \frac{1}{a^2} \cos(\psi)^2 ,\\ &=& a^2\left( 1 - e^2 \cos(\psi)^2 \right)\end{aligned}$$ From this the perimeter is calculated: $$\begin{aligned} |\partial\Omega| &=& 4 a \int_0^{\pi/2} \sqrt{1- (e\cos(\psi))^2}\, d\psi ,\\ &=& 4 a \int_0^{\pi/2} \sqrt{1- (e\sin({\hat\psi}))^2}\, d{\hat\psi} ,\\ &=& 4 a\, {\rm EllipticE}(e) .\end{aligned}$$ [**The support function, $h$**]{} There are several equivalent definitions of the support function $h$ which we consider as a function of points on the boundary. Thus at the point associated with the value $\psi$ in definition (\[eq:param\]) $$h =\frac{1}{a\sqrt{1+e^2\cos^2(\psi)}} .$$ As noted in [@Kim07], the curvature is $h^3$. Once again, the functions $g$ and $\hat g$ make an appearance. ### Elliptic coordinates The elliptic coordinates are related to Cartesians by $$x= c\cosh(\eta)\cos(\psi), \qquad y= c\sinh(\eta)\sin(\psi),$$ and we will set the parameter $c$ by $$c= \sqrt{a^2-a^{-2}} .$$ The boundary of the ellipse can be represented, with fixed $\eta_0$, $$x= c\cosh(\eta_0)\cos(\psi), \qquad y= c\sinh(\eta_0)\sin(\psi).$$ In this $$a= c\cosh(\eta_0), \qquad a^{-1}= c\sinh(\eta_0) ,$$ so that $$\tanh(\eta_0) = a^{-2},\quad (c^2=a^2-a^{-2}),\quad \epsilon= \frac{1}{\cosh(2\eta_0)} .$$ Nearly circular ellipses will have $\eta_0$ large. The perimeter is calculated: $$\begin{aligned} |\partial\Omega| &=& 4 \int_0^{\pi/2} \sqrt{ (\frac{d x}{d\psi})^2 +(\frac{d y}{d\psi})^2}\, d\psi ,\\ &=& 4 c \int_0^{\pi/2} \sqrt{ (\cosh(\eta_0))^2 - (\cos(\psi))^2} \, d\psi , \\ &=& 4 c \cosh(\eta_0) \, {\rm EllipticE}\left(\frac{1}{\cosh(\eta_0)}\right) , \\ &=& 4 a\, {\rm EllipticE}(e) .\end{aligned}$$ We remark that toroidal functions occur elsewhere in this study and that [@Ab09] gives a formula, symmetric in $a$ and $b$, $$|\partial\Omega| = 2\pi \sqrt{a\,b}\, {\rm LegendreP}_{1/2}\left(\frac{a^2+b^2}{2 a\, b}\right) . \label{eq:ab09P}$$ Nearly circular ellipses, polar coordinates {#subsec:GeomNearCirc} ------------------------------------------- The first few terms in the expansion of $r(\theta)$ are $$r(\theta) = 1+\frac{1}{2}\,\cos \left( 2\,\theta \right) \epsilon+ \left( \frac{3}{16}\,\cos \left( 4\,\theta \right) -\frac{1}{16} \right) {\epsilon}^{2} + O(\epsilon^3) .$$ Alternatively, we can consider asymptotics as $a\rightarrow{1}$. Then $\rho=r_{\rm ellipse}-1$ satisfies [$$\begin{aligned} \rho(\theta) &\sim (a-1) \cos(2\theta) - (\frac{1}{4} +\frac{1}{2} \cos(2\theta)-\frac{3}{4} \cos(4\theta)) (a-1)^2 + O( (a-1)^3) \nonumber \\ &= -\frac{1}{4} (a-1)^2 + \left((a-1)-\frac{1}{2} (a-1)^2\right) \cos(2\theta)+\frac{3}{4} (a-1)^2\cos(4\theta) + O( (a-1)^3) . \label{eq:ellipseRho}\end{aligned}$$ ]{} For calculations extending the use of higher order terms in the Fourier series (\[eq:ellipseRho\]). one might need to replace $x^{2n}$ terms using formulae like $$x^{2n}=2^{1-2n}\left(\frac{1}{2} \binom{2n}{n}+\sum_{j=1}^n \binom{2n}{n-j}T_{2j}(x) \right),$$ and, at some future date, we may for elliptical $\Omega$ use this but have yet to do so. Moments of inertia {#subsec:Ic} ------------------ ### General $\Omega$ {#general-omega} [**Area moments**]{} The polar moment of inertia, taking the origin at the centroid, is $$I_c = \int_\Omega \left( (x-x_c)^2 + (y-y_c)^2 \right) ,$$ where $z_c=(x_c,y_c)$ is the centroid of $\Omega$. When the boundaries are given in polar coordinates, this is $$I_c =\frac{1}{4} \int_0^{2\pi} r(\theta)^4\ d\theta .$$ Higher order moments arise in connection with calculations based on polynomial test functions in the variational approach of §\[sec:Varl\]. [**Boundary moments**]{} See §\[sec:Varl\]. ### Ellipse [**Area moments**]{} For our disk and ellipse these are $$I_c({\rm disk})= \frac{\pi}{2} a^4, \qquad I_c({\rm ellipse})= \frac{\pi}{4} (a^2 + a^{-2}) =\frac{\pi}{2\sqrt{1-\epsilon^2}} . \label{eq:icEllipseEps}$$ The asymptotics below check with the entries in the table of [@PoS51] treating domain functionals for nearly circular domains: $$\begin{aligned} \frac{2 I_c({\rm ellipse})}{\pi} &\sim 1 + 2(a-1)^2 \qquad{\rm as\ } a\rightarrow 1 ,\\ \left(\frac{2 I_c({\rm ellipse})}{\pi}\right)^{1/4} & \sim 1 + \frac{1}{2}(a-1)^2 + o(a-1)^2 ) \qquad{\rm as\ } a\rightarrow 1 ,\\ & \sim 1 + a_0 + \frac{3}{4}a_2^2 + o(a-1)^2 )\qquad{\rm as\ } a\rightarrow 1 .\end{aligned}$$ [**Boundary moments**]{} For doubly-symmetric domains like our ellipse, our notation in §\[sec:Varl\] is $$i_{2n} = \int_{\partial\Omega} r^{2n}\, ds .$$ For the ellipse, these can be evaluated in terms of elliptic integrals. Ellipse geometry: miscellaneous {#subsec:ellipseMisc} ------------------------------- The modulus of asymmetry for an ellipse is calculated in [@Fi14]. The calculation involves calculating ${\hat g}_0$. In the case $\beta=0$ there are improvements to the St Venant inequality in terms of the modulus of asymmetry. (For the St Venant inequality, see [@PoS51], and, for ellispes, trivially the $\beta=0$ case in inequality (\[eq:isoperQ\]).) Appendix: Elliptic integrals {#appx:EllipticInts} ============================ The functions denoted by EllipticE and EllipticK in this paper are, in this appendix, written as $E(k)$ and $K(k)$. Our notation follows that of many (but not all!) authors, in particular [@Law89]§3.8 and [@GR]: $$\begin{aligned} E(k) &=& \int_0^{\pi/2} \sqrt{1-k^2\,\sin(\theta)^2}\, d\theta,\\ K(k) &=& \int_0^{\pi/2} \frac{1}{\sqrt{1-k^2\,\sin(\theta)^2}}\, d\theta .\end{aligned}$$ For $0\le{z}<1$ $$K(k)=\int_0^1 \frac{1}{\sqrt{(1-t^2)(1-k^2 t^2)}}\, dt ,$$ $$E(k)_=\int_0^1\sqrt{ \frac{1-k^2 t^2}{1-t^2}}\, dt .$$ The Cauchy-Schwarz inequality yields $$\frac{\pi^2}{4} \le K(k)\, E(k) .$$ The functions can be written as hypergeometric functions: $$\begin{aligned} K(k) &=&\frac{\pi}{2} {\ }_2{F}_1(\frac{1}{2},\frac{1}{2};1; k^2) , \\ E(k) &=& \frac{\pi}{2} {\ }_2F_1(-\frac{1}{2},\frac{1}{2};1; k^2) .\end{aligned}$$ In showing, in Appendix \[app:Furtherg\] equation (\[eq:gHathy\]), that two different representations of ${\hat g}_n$ are equivalent, we needed to tell maple that, as in [@GR] 8.126, with $k'=\sqrt{1-k^2}$, (and, in a different notation in [@AS]17.3.29) $$\begin{aligned} K\left(\frac{1-k'}{1+k'}\right) &=& \frac{1+k'}{2}\, K(k) , \\ E\left(\frac{1-k'}{1+k'}\right) &=& \frac{1}{1+k'}\, \left(E(k)+k'\, K(k)\right) .\end{aligned}$$ (The first of these is given, in a different notation, on [functions.wolfram.com]{}.) In our application $k=1/\sqrt{q}=e$. Another identity (which we will see in connection with toroidal functions equation (\[eq:niceW0\])) is ‘Legendre’s relation’ ([@AS]17.3.13): $$K(k) E(k') - K(k) K(k') +K(k') E(k) = \frac{\pi}{2} . \label{eq:LegRel}$$ Complex k --------- The EllipticK function has a mirror symmetry: $$K({\bar k}) = {\bar K(k)} .$$ The real and imaginary parts of $K(\exp(i\phi))$ are given, for example, at the ‘Complex Characteristics’ section of\ [functions.wolfram.com]{}. From this one finds $$|K(\exp(i\phi)|^2 =\frac{1}{4}\left(K(\cos(\phi))^2 + K(\sin(\phi))^2\right) . \label{eq:Kdisk}$$ See equation (\[eq:weightK\]). Appendix: A difference equation – and toroidal functions {#app:Furtherg} ======================================================== The Fourier coefficients $g_n$ and ${\hat g}_n$ are central to our separation of variables solution of Problems (P($\beta$)) and (P($\infty$)) respectively. On using maple to find these, maple returned (instances of) them as the sum of two terms: the first term is the product of a polynomial with an EllipticE function, and the second term is the product of a polynomial with an EllipticK function. The dependence on $n$ is only in the polynomials. See equations (\[eq:gnEK\]), (\[eq:gBarnEK\]), repeated at equations (\[eq:gnEll\]), (\[eq:gBarnEll\]). The polynomials satisfy the same recurrence relations as $g_n$ and ${\hat g}_n$: see (RE($\alpha$)) below where $\alpha=1/2$ for $g_n$ and $\alpha=3/2$ for ${\hat g}_n$. It happens that $g_n$ and ${\hat g}_n$ can also be expressed in terms of LegendreQ functions – specifically toroidal functions. Again sequences of polynomials satisfying (RE(1/2)) and (RE(3/2)) arise: see equations (\[eq:Q1pol\]), (\[eq:Q3pol\]). More generally, for any $\alpha$, these polynomial sequences are present in the entries on Legendre functions at functions.wolfram.com. There is also some literature concerning the polynomials in families of recurrence relations which include our case $\alpha=1/2$: see [@BD67; @Gr85] and subsection \[subapp:orthog\] below. It isn’t obvious what, for general $\alpha$, the polynomials should be called, though for $\alpha=0$ and $\alpha=1$ they are Chebyshev polynomials. In the case $\alpha=1/2$, the sequence of polynomials (particularly those labelled $p_{01}$) might be called ‘associated toroidal Legendre polynomials’, and we now refer to all the sequences of polynomials $p_{01}(n)$, $p_{10}(n)$ of equation (\[eq:Q1pol\])) and ${\hat p}_{01}(n)$, ${\hat p}_{10}(n)$ of equation (\[eq:Q3pol\])) (or linear combinations thereof, e.g. $p_{10}+x\,p_{01}$ and, more generally, $\alpha$ half an odd integer) by this name. Though others have used the adjective ‘associated’ in this context, we remark that the adjective is used in several different ways. While the previously cited [@BD67; @Gr85] on ‘associated Legendre functions’ includes our case of $\alpha=1/2$ it doesn’t include $\alpha=3/2$. ‘Associated Gegenbauer polynomials’ as in [@BI82]§3 or, more generally, the family of Pollaczek polynomials introduced in [@Poll50] (see [@HTF2]p220) cover general $\alpha$ but have more parameters than our simpler cases. (See also our §\[subapp:orthog\].) Questions which arise include the following. - Are there simple formulae for these sequences of polynomials? E.g. (i) Can each sequence be expressed as a single hypergeometric function (perhaps a ${}_{3}F_2$)? (ii) What are the formulae for coefficients? - Finally, how might the information on the polynomials be best applied to progress the tasks associated with the pde questions which we have listed below in a subsection concerning ‘Hopes’. A related problem, also involving sequences of polynomials satisfying (RE(1/2)) or (RE(3/2)) is to discover formulae, for general $n$, agreeing with maple’s output for the $g_n$ and ${\hat g}_n$ involving EllipticE and EllipticK functions. This also seems to be an old problem: finding formulae for the toroidal functions $Q_{n-1/2}$ involving EllipticE and EllipticK functions. We will do relatively little on this last problem. Notation {#subapp:Notation} -------- We denote the reciprocal of $g$ by $\hat g$: $${\hat g}(\psi)=\frac{1}{g(\psi)}=\sqrt{1-\frac{\cos^2(\psi)}{q}} .$$ As described earlier, the Fourier cosine coefficients of $g$ are denoted $g_n$ (see equation (\[eq:bgnDef\])), and those of ${\hat g}$ by ${\hat g}_n$: $$g_n = \frac{2}{\pi}\int_0^\pi g(\psi)\cos(2 n\psi)\, d\psi , \label{eq:agnDef}$$ $${\hat g}_n = \frac{2}{\pi}\int_0^\pi {\hat g}(\psi)\cos(2 n\psi)\, d\psi . \label{eq:bgBarnDef}$$ Both $g_n$ and ${\hat g}_n$ can be written concisely in terms of toroidal functions, Legendre functions: see equations (\[eq:gnQ\]), (\[eq:gBarnQ\]). The sequences $g_n$ and ${\hat g}_n$ are used, so far, just in numerics. In both sequences our interest is in $q>1$ but, the integral for ${\hat g}_n$ evaluates at other values of $q$, for example at $q=1$ to $${\hat g}_n(q=1) =\frac{2}{\pi} \int_0^\pi \sin(\psi) \cos(2n\psi) \, d\psi = \frac{1}{\pi}\, \frac{1}{n^2-1/4} . \label{eq:gBarnq1}$$ Though the integrals are complex, one finds $$\frac{g_{n+1}(q=1/2)}{g_{n-1}(q=1/2)} = -\, \frac{n-1/2}{n+1/2} ,$$ and a similar formula for the ${\hat g}_n$ sequence. Hopes for application to the pde problems {#subapp:Hopes} ----------------------------------------- We have various hopes for future results. - One hope is that further results on ${\hat g}_n$ may ultimately lead to further analytical results concerning $u_\infty$ and, in particular, its values on the boundary, and thence to analytical expressions for $\Sigma_\infty$ and $\Sigma_1$. - A more modest goal, also as yet not implemented, is that it may be possible to improve the efficiency of the numerical calculations of $g_n$ and of ${\hat g}_n$ as the present codes call on maple to evaluate each of the integrals separately for each $n$ and uses maple’s results given in terms of EllipticE and EllipticK functions. - There also remains the prospect, should there be any demand for codes in languages used by engineers, e.g. matlab, to provide such codes. (We remark that as at 2019a matlab’s legendre function does not compute toroidal functions. We have yet to check the matlab File Exchange site.) We remark that there are numerical stability issues with recurrence relations when performed purely within fixed precision float computation but these can be circumvented using the Symbolic Toolbox. It isn’t essential for the code to be entirely within matlab as it would also be possible to use existing Fortran or C codes (such as described in [@GST00]) mex-ed into matlab. (Matlab codes and related algorithms for toroidal functions are mentioned by several authors, Majic is one, Gil and Segura others, e.g [@GST00]. There is a library, POLPAK, implementing algorithms in various languages, including matlab.)\ Noting the present limitations of matlab’s legendre function, our present suggestion is to compute $Q_{-1/2}$, $Q_{1/2}$ and $Q_{-1/2}^{-1}$, $Q_{1/2}^{-1}$ from their expressions in terms of EllipticE and EllipticK functions, and then the toroidal $Q$ functions using the polynomials $p_{01}$, $p_{10}$ and ${\hat p}_{01}$, ${\hat p}_{10}$ defined below. (Matlab’s Symbolic Toolbox might have a role in connection with the polynomials.) Introducing the difference equation (RE($\alpha$)) {#subapp:REalpha} -------------------------------------------------- In the body of this paper we noted, at equations (\[eq:gnEK\]) and (\[eq:gBarnEK\]), that maple, when asked to perform the integrals defining $g_n$ and ${\hat g}_n$ returned a result in which the terms multiplying elliptic integrals are polynomials in $q$ or rather $q_2=2q-1$. We denoted the polynomials by $E_n(q_2)$, $K_n(q_2)$ (these satisfying the recurrence (RE(1/2)) defined below) and by ${\hat E}_n(q_2)$, ${\hat K}_n(q_2)$ (satisfying the recurrence (RE(3/2))). While we have yet to make any significant use of the items in this appendix for our Robin boundary condition problems, we have discovered that the Fourier series of $g$ and of $\hat g$, and related studies of toroidal functions, have been studied in connection with various different problems since at least the 1880s. In spite of the bewilderingly large literature both on Legendre functions and on orthogonal polynomials, we have not seen a systematic account of the few results we present concerning sequences of polynomials satisfying the simple recurrences satisfied by each of $g_n$ and of ${\hat g}_n$. Our sequences of Fourier coefficients satisfy difference equations of the form $$(n+\alpha) u_{n+1} =2 q_2 n u_n - (n-\alpha) u_{n-1} , \eqno{({\rm RE}(\alpha))}$$ where $q_2=2q-1$, and where conformity to conventions in the orthogonal polynomial literature seems appropriate, we write $x$ rather than $q_2$. The $g_n$ satisfy this with $\alpha=1/2$ while the ${\hat g}_n$ satisfy it with $\alpha=3/2$. The recurrence relation (RE($\alpha$)) satisfied by Legendre functions is given in standard references such as [@AS] 8.5.3 and [@HTF1] §3.8 (2). The order $\mu$ of the Legendre functions, $Q^\mu_\nu(q_2)$ is determined by $\alpha=\frac{1}{2}-\mu$: the degrees in the sequence are of the form $\nu=n-1/2$. Our interest is in $\alpha=1/2$ (for which we give later polynomial sequences $p_{01}(n)$ and $p_{10}(n)$) and $\alpha=3/2$ (for which we give later polynomial sequences ${\hat p}_{01}(n)$ and ${\hat p}_{10}(n)$) but results at other $\alpha$ might be useful in contexts other than the pde problem of this paper. An aside: (RE(0)) and (RE(1)): Chebyshev {#subapp:RE0} ---------------------------------------- At $\alpha=0$ the $n$ cancels and the recurrence is constant coefficient, so solvable by $u_n=r^n$ for appropriate $r$. The general solution is a linear combination of the Chebyshev polynomial solutions $$u_n(\alpha=0)= c_T(q_2) T_n(q_2) + c_U(q_2) U_n(q_2),$$ where $c_T$ and $c_U$ are independent of $n$. As $n$ no longer occurs in the coefficients in the recurrence (RE(0)), equally good as a polynomial solution is $u_n=T_{n-m}$ for integer $m$. Chebyshev polynomials also arise in solving the recurrence when $\alpha=1$: solutions are $$u_n(\alpha=1)=\frac{u_n(\alpha=0)}{n} .$$ As remarked before, solutions can be expressed in terms of Legendre functions. For the $\alpha=0$, $\mu=1/2$ case, see [@HTF1]3.6.1(12). Also we have $$\begin{aligned} P^{1/2}_{n-1/2}(\cos(\theta)) &=& \sqrt{\frac{2}{\pi\sin(\theta)}} \cos(n\theta) = \sqrt{\frac{2}{\pi\sin(\theta)}} T_n(\cos(\theta)) , \\ Q^{1/2}_{n-1/2}(\cos(\theta)) &=& -\sqrt{\frac{\pi}{2\sin(\theta)}} \sin(n\theta) , \\ P^{1/2}_{n-1/2}(\cosh(\eta)) &=& \sqrt{\frac{2}{\pi\sinh(\eta)}} \cosh(n\eta) = \sqrt{\frac{2}{\pi\sinh(\eta)}} T_n(\cosh(\eta)), \\ Q^{1/2}_{n-1/2}(\cosh(\eta)) &=& -\sqrt{\frac{\pi}{2\sinh(\eta)}} \exp(-n\eta) .\end{aligned}$$ We remark that $$\begin{aligned} U_n(x) &=& \frac{x\, T_{n+1}(x) - T_{n+2}(x)}{1-x^2} ,\\ U_n(\cosh(\eta)) &=& \frac{\sinh((n+1)\eta)}{\sinh(\eta)} ,\\ \exp(-n\eta) &=&T_n(\cosh(\eta))-\sinh(\eta)U_{n-1}(\cosh(\eta)),\end{aligned}$$ and $u_n=\exp(-n\cosh^{-1}(q_2))$ solves (RE(0)). Similar formulae are available for the case $\alpha=1$, order $\mu=-1/2$. Maple’s [rsolve]{} solves (RE($m$)) for other integer $m$. None of the items in this elementary subsubsection are relevant to the $\alpha=1/2$ and $\alpha=3/2$ cases which arise in our partial differential equation problem. However, relations between families of polynomials and various connections between Legendre functions will arise in the context of the $\alpha=1/2$ and $\alpha=3/2$ cases and the easier situation sometimes suggests possible results. Deriving (RE($\alpha$)) from integrals defining Fourier coefficients {#subapp:fromFS} -------------------------------------------------------------------- While the expressions for $g_n$ and ${\hat g}_n$ in terms of toroidal functions have value, much can be derived, simply from their definitions, including the recurrences (RE($\alpha$)) and inter-relationships. The simple relationships between $g$ and ${\hat g}$ lead to relationships between their Fourier coefficients. - Since $${\hat g}(\psi)=\left(1- \frac{\cos^2(\psi)}{q}\right) \, g(\psi) = \frac{1}{2 q} \left( (2q-1)-\cos(2\psi)\right)\, g(\psi) ,$$ we have by multiplying both sides by $\cos(2n\psi)$, using cosine formulae, and integrating $$4 q {\hat g}_n = 2 (2q-1) g_n - g_{n+1}- g_{n-1} . \label{eq:cross0}$$ - Since $$2 q \frac{ d {\hat g}(\psi)}{d q} + {\hat g}(\psi) - g(\psi) = 0,$$ the Fourier coefficients $g_n$ can be found from those of $ {\hat g}$: $$g_n = {\hat g}_n + 2 q \frac{ d {\hat g}_n}{d q} = 2\sqrt{q} \frac{d}{dq} (\sqrt{q}{\hat g}_n ) . \label{eq:gDq}$$ - Since $$\frac{ d {\hat g}(\psi)}{d q} =\frac{\cos(\psi)^2 \, g}{2 q^2} = \frac{(\cos(2\psi)+1 \, g}{4 q^2} ,$$ $$\frac{ d {\hat g}_n}{d q} = \frac{1}{8q^2} (2g_n +g_{n-1}+g_{n+1}) . \label{eq:ghDq}$$ (Equations (\[eq:gDq\]) and (\[eq:ghDq\]) check against (\[eq:cross0\]).) Define also $$q_2 = 2 q -1 .$$ There are three-term recurrence relations between the coefficients: $$\begin{aligned} (n-\frac{1}{2}) g_n &=& 2 q_2 (n-1) \, g_{n-1} - (n-\frac{3}{2}) \, g_{n-2} , \label{eq:gnrec}\\ (n+\frac{1}{2}) {\hat g}_n &=& 2 q_2 (n-1) \, {\hat g}_{n-1} - (n-\frac{5}{2} )\, {\hat g}_{n-2} . \label{eq:gBarnrec}\end{aligned}$$ The derivation is treated below, where we derive the following (RE($1/2$)) equivalent to (\[eq:gnrec\]): $$(n+\frac{1}{2}) g_{n+1} = 2 q_2 n \, g_{n} - (n-\frac{1}{2}) \, g_{n-1} . \eqno{({\rm RE}(1/2))}$$ Corresponding to (\[eq:gBarnrec\]) we have (RE($3/2$)) $$(n+\frac{3}{2}) {\hat g}_{n+1} = 2 q_2 n \, {\hat g}_{n} - (n-\frac{3}{2}) \, {\hat g}_{n-1} . \eqno{({\rm RE}(3/2))}$$ We now repeat equations given in the main part of the paper. In these the argument of the elliptic integrals is $1/\sqrt{q}$ and we wrote, $$\begin{aligned} g_n &=& \frac{4}{\pi}\, \left( E_n {\rm EllipticE}_0 + K_n {\rm EllipticK}_0 \right), \label{eq:gnEll}\\ {\hat g}_n &=& \frac{4}{\pi}\, \left( {\hat E}_n {\rm EllipticE}_0 + {\hat K}_n {\rm EllipticK}_0 \right) , \label{eq:gBarnEll}\end{aligned}$$ with the $E_n$ , $K_n$, ${\hat E}_n$, $ {\hat K}_n$ polynomials in $q_2$ of degree at most $n$. The polynomials $E_n$ , $K_n$ satisfy the recurrence (\[eq:gnrec\]), (RE(1/2)): the polynomials ${\hat E}_n$ , ${\hat K}_n$ satisfy the recurrence (\[eq:gBarnrec\]), (RE(3/2)). The starting values for the iterations to determine the polynomials are given, for the $g_n$ iteration by (\[eq:g0Ell\]), (\[eq:g1Ell\]) and for the ${\hat g}_n$ iteration by (\[eq:gBar0Ell\]), (\[eq:gBar1Ell\]). Maple’s rsolve will solve (RE($\alpha$)) in the special cases when $q_2=0$ and $q_2=1$. When $q_2=1$ there are constant solutions ($g_n=1$ for all $n$), and reduction of order gives the other. The method to establish (\[eq:gnrec\]), (\[eq:gBarnrec\]) involves a further relation between the coefficients $g_n$, ${\hat g}_n$ of the two series, and is suggested in a posting by Jack D’Aurizio on stackexchange. See\ [`math.stackexchange.com/questions/930003/fourier-series-of-sqrt1-k2-sin2t`\ ]{} We have $$g_{m+1}= (2 q-1) g_m + (4 m-2) q \, {\hat g}_m , \label{eq:cross1}$$ and also, exactly as in D’Aurizio’s post, $$8 m q {\hat g}_m = g_{m+1}- g_{m-1} . \label{eq:cross2}$$ Eliminating ${\hat g}_m$ between equations (\[eq:cross1\]), (\[eq:cross2\]) we find the recurrence (RE(1/2)) with $n$ there replaced by $m$. Also, equations (\[eq:cross1\]) and (\[eq:cross2\]) together yield $$g_{m-1}= (2 q-1) g_m - (4 m+2) q \, {\hat g}_m , \label{eq:cross1a2}$$ It may be that, for $q_2>1$, all the $g_m$ are positive, and the sequence is decreasing.\ We have ${\hat g}_0>0$ and it may be that, for $q_2>1$ and $m\geq{1}$ that the terms ${\hat g}_m$ are all negative and form an increasing sequence. ### Generating functions? Correcting a misprint in [@Ch78] p202, equation (12.5), (which reports [@BD67]) for $\alpha=1/2$, a generating function is $$\begin{aligned} G(x,w) &=& \frac{1}{2}\,\frac{1}{\sqrt{w(1-2xw+w^2)}}\, \int_0^w \frac{1}{\sqrt{t(1-2xt+t^2)}} \, dt , \label{eq:GenDef}\\ G(x,w) &=& \sum_{n=0}^\infty p_{01}(n+1,x) w^n . \label{eq:Genp01}\end{aligned}$$ The $p_{01}(n)$ are polynomials of degree $n-1$ satisfying (RE(1/2)) starting from $p_{01}(0)=0$, $p_{01}(1)=1$. These are studied extensively in later parts of this appendix. (Equation (\[eq:p01Leg\]) follows from this: see [@Ch78] equation (12.6).) This subsection is ‘incomplete’, with everything after this paragraph merely an early personal attempt predating finding the result reported in equation (\[eq:GenDef\]) above. It may be that, for general $\alpha$, elaborate special functions, e.g. Apell hypergeometrics, may be needed as they were in the more general situation in [@BI82]. At least for our special case of $\alpha=1/2$ incomplete elliptic integrals suffice: see also equations (1.10), (1.11) of  [@VZ07]. The stackexchange posting suggests there might be reasonably neat expressions for generating functions, $$G(X) = g_0 + g_1\, X +\sum_{n=2}^\infty g_n X^n, \qquad {\hat G}(X) = {\hat g}_0 + {\hat g}_1\, X +\sum_{n=2}^\infty {\hat g_n} X^n ,$$ and that these functions satisfy first order linear differential equations.\ The differential equation satisfied by $G$ is $${\cal L} G =(X-2 q_2 X^2 +X^3)\frac{d G}{d X} -\frac{1}{2} (1-X^2) G = \frac{1}{2} (-g_0 +g_1 X) .$$ $X=0$ is a singular point. The homogeneous de ${\cal L}G_h=0$ is solved by a function $G_h$ with $G_h(0)=0$, $$G_h(X) = \sqrt{\frac{X}{1-2q_2 X + X^2}} .$$ We would be content with representations of $G$ valid for $q_2>1$ and $0\leq{X}<1/(2q_2)$. Applying the ‘variation of parameters’ formula to the nonhomogenous de above, using $G_h$, leads to a messy particular solution in terms of incomplete elliptic integrals. (Though possibly unrelated, we will see elliptic integrals in other contexts: see §\[subapp:orthog\].)\ The corresponding de for ${\hat G}$ is: $${\hat{\cal L}}{\hat G} =(X-2 q_2 X^2 +X^3)\frac{d {\hat G}}{d X} + \frac{1}{2} (1-X^2) {\hat G} = \frac{1}{2} ({\hat g}_0 +3{\hat g}_1 X) .$$ $X=0$ is a singular point. The homogeneous de ${\cal L}{\hat G}_h=0$ is solved by a function ${\hat G}_h$ with ${\hat G}_h(0)=0$, $${\hat G}_h(X) = \sqrt{\frac{1-2q_2 X + X^2}{X}} ,$$ (which is the reciprocal of $G_h$). The relations (\[eq:cross0\],\[eq:gDq\]) lead to relations between $G$ and $\hat G$, for example $$G= 2\sqrt{q}\frac{d}{dq} (\sqrt{q} {\hat G}) .$$ One can derive a differential equation for ${\hat G}(q)$ involving derivatives with respect to $q$. It may be that if one starts from a closed form for ${\hat G}(1)$, i.e. at $q=1$, solving the initial value problem might be useful. ### Hypergeometric representations Also presented on stackexchange is a hypergeometric formula for ${\hat g}_n$. See also [@Cv09] at Lemma 1 which gives a slightly different representation. Define $$\lambda= \frac{1}{\sqrt{q}+\sqrt{q-1}} = \sqrt{q}-\sqrt{q-1} ,$$ from which $$\lambda^2 = \frac{1}{q_2 +\sqrt{q_2^2-1}} .$$ We have $${\hat g}_n = - \frac{2}{\lambda\sqrt{q}}\, \frac{(2n)!}{(2n-1)\,2^{2n+1}\, (n!)^2}\, (\lambda^2)^n\, {\ }_2{F}_1(-\frac{1}{2},n-\frac{1}{2};n+1;\lambda^4) . \label{eq:gHathy}$$ (Quick check, ${\hat g}_0>0$ and, for $n>0$, ${\hat g}_n<0$, though we only have numeric evidence for this.) ### Negative $n$ We remark that from the $\cos(2n\psi)$ in the defining integrals we have $$g_{-n}= g_n, \qquad {\hat g}_{-n}={\hat g}_n .$$ See also equations (\[eq:AS822\]), (\[eq:Pminus\]). It is also more generally true that if $u_n$ is a solution of (RE($\alpha$)) and $\alpha\ne{0}$ then $u_{-n}=u_n$. Toroidal functions, LegendreQ functions {#subapp:Toroidal} --------------------------------------- ### Main facts concerning $Q$ functions We defer to a later subsection items on LegendreP functions. LegendreP functions are equally appropriate as solutions of (RE(1/2)) and (RE(3/2)) but our initial concern is with the Fourier coefficients $g_n$ and ${\hat g}_n$ and these are LegendreQ functions. Also in this subsection we are interested in the functions defined on the interval $(1,\infty)$ though in some later sections we are also concerned with the interval $(-1,1)$. (Branch points will then need to be considered even though our application involves real-valued functions of a real argument.) Toroidal functions are special cases of Legendre functions, specifically $Q^\mu_{n-1/2}$ and $P^\mu_{n-1/2}$ with $n$ integer, and, sufficient for our purposes also $\mu$ integer. We have, as in [@HTF1][§]{}3.10(3) and other references such as [@Co07] $$\begin{aligned} g_n &=& \frac{4\sqrt{q}}{\pi}\, Q_{n-1/2}(q_2) , \label{eq:gnQ}\\ {\hat g}_n &=& \frac{2\sqrt{q-1}}{\pi}\, Q^{-1}_{n-1/2}(q_2) . \label{eq:gBarnQ}\end{aligned}$$ (See equations (43) and (44) of [@Co07].) Toroidal functions can be related to Elliptic Integrals. In particular $$\begin{aligned} Q_{-1/2}(q_2) &=&\frac{1}{\sqrt{q}}\, {\rm EllipticK}(\frac{1}{\sqrt{q}}) , \label{eq:Qmh} \\ Q_{1/2}(q_2) &=& q_2 \frac{1}{\sqrt{q}}\, {\rm EllipticK}(\frac{1}{\sqrt{q}} )- 2\sqrt{q} \, {\rm EllipticE}(\frac{1}{\sqrt{q}} ) . \label{eq:Qph} \end{aligned}$$ That for $Q_{-1/2}$ is [@AS] 8.13.3: that for $Q_{1/2}$ is [@AS] 8.13.7. The corresponding formulae for the Legendre $P$ functions, $P_{-1/2}(q_2)$ and $P_{1/2}(q_2)$, though not immediately relevant to the $g_n$ and ${\hat g}_n$ are relevant to solving (RE($\alpha$)) and are treated in a later section. We also remark, as in [@CD10] p339, [$$\begin{aligned} Q^{1}_{-1/2}(q_2) &=& - \frac{1}{2\sqrt{q-1}}\, {\rm EllipticE}(\frac{1}{\sqrt{q}}) , \label{eq:Qp1mh} \\ Q^{1}_{1/2}(q_2) &=& - \frac{q_2}{2\sqrt{q-1}}\, {\rm EllipticE}(\frac{1}{\sqrt{q}}) +\sqrt{q-1} {\rm EllipticK}(\frac{1}{\sqrt{q}} ) . \label{eq:Qp1ph}\end{aligned}$$ ]{} and, looking ahead, from (\[eq:Qp1m1\]) $$\begin{aligned} Q^{-1}_{-1/2}(q_2) &=& -4\, Q^{1}_{-1/2}(q_2) , \label{eq:Qm1mh} \\ Q^{-1}_{1/2}(q_2) &=& \frac{4}{3} Q^{1}_{1/2}(q_2) . \label{eq:Qm1ph}\end{aligned}$$ [@AS] 8.2.2 gives, for integer $\mu$, $$Q_{-n-1/2}^\mu(z) = Q_{n-1/2}^\mu(z) . \label{eq:AS822}$$ [@HTF1][§]{}3.8(9) gives $$\frac{d Q_{n-1/2}(z)}{dz} = \frac{n^2-1/4}{\sqrt{z^2-1}} \, Q^{-1}_{n-1/2} . \label{eq:HTF1Qp3p8p9}$$ As noted above, the $g_n$ are given by a function of $q_2$ times $Q_{n-1/2}(q_2)$, the latter functions solving (RE($1/2$)). The ${\hat g}_n$ are given by a function of $q_2$ times $Q^{-1}_{n-1/2}(q_2)$, the latter functions solving (RE($3/2$)). Where it seems appropriate we prefer to arrange our formulae so that the $Q^{-1}_{n-1/2}$ are written in terms of $Q_{m-1/2}$, for example from [@HTF1]§3.8(5), $$Q^{-1}_{n-1/2}(z) = \frac{z\,Q_{n-1/2}(z)-Q_{n-3/2}(z)}{(n+1/2)\sqrt{z^2-1}} . \label{eq:HTF385}$$ (This agrees with (\[eq:cross1a2\]).) Also, as in [@HTF1][§]{}3.8(3) $$Q^{-1}_{n-1/2}(z) = \frac{Q_{n+1/2}(z)-Q_{n-3/2}(z)}{2n\sqrt{z^2-1}} . \label{eq:HTFQ3p8p3}$$ (See also Theorem \[thm:thm1\] below.) Particular cases of equation (\[eq:HTF385\]), at $n=0$ and $n=1$, are: $$\begin{aligned} Q^{-1}_{-1/2}(z) &=& \frac{z\,Q_{-1/2}(z)-Q_{1/2}(z)}{(1/2)\sqrt{z^2-1}} , \label{eq:HTF0Q}\\ Q^{-1}_{1/2}(z) &=& \frac{z\,Q_{1/2}(z)-Q_{-1/2}(z)}{(3/2)\sqrt{z^2-1}} . \label{eq:HTF1Q}\end{aligned}$$ We note, as in [@AS] 8.2.6, $$Q^{1}_{n-1/2} (z) = (n^2-1/4) Q^{-1}_{n-1/2} (z) \label{eq:Qp1m1}$$ which can be deduced from [@HTF1] §3.8 equations (5) and (7). (This accords with Theorem \[thm:thm3a\] given in the next section.) As noted above Conway [@Co07] and various other authors give formulae for $g_n$ (see [@Co07] equation (43)) and for ${\hat g}_n$ in terms of Legendre $Q$ functions. Concerning the latter, see [@Co07] equation (44), $$\begin{aligned} {\hat g}_n &=& \frac{2(\sqrt{q-1})}{\pi} Q^{-1}_{n-1/2}(q_2) = \frac{2(\sqrt{q-1})}{\pi} \frac{1}{n^2-1/4} Q^1_{n-1/2}(q_2) , \label{eq:ConQh1}\\ &=& \frac{2(\sqrt{q-1})}{\pi}\left( \frac{q_2\,Q_{n-1/2}(q_2)-Q_{n-3/2}(q_2)}{(n+1/2)\sqrt{q_2^2-1}} \right) . \label{eq:ConQh2} \end{aligned}$$ (See also [@CD10] p339. It also accords with the corollary to Theorem \[thm:thm1\].) As a check we recall (\[eq:gBarnq1\]) $${\hat g}_n(q_2=1)= - \frac{1}{\pi(n^2-1/4)} .$$ [@HTF1] §3.9(6) gives asymptotics for $q_2$ tending down to 1: $$Q^{-1}_{n-1/2}(q_2) \sim -\frac{1}{\sqrt{2}}\frac{\Gamma(n-1/2)}{\Gamma(n+3/2)}\frac{1}{\sqrt{q_2-1}} = -\frac{1}{\sqrt{2}}\frac{1}{n^2-1/4} \frac{1}{\sqrt{2q-2}}.$$ Using this last expression in (\[eq:ConQh1\]) checks with the elementary result that ${\hat g}_n(q_2=1)= - {1}/({\pi(n^2-1/4)})$. ### Integrals of sums of squares of our Fourier coefficients Using Parseval’s Theorem it is easy to find $\sum_0^\infty{\hat g}_n^2$, etc. See also [@Co07]§[5]{}. (RE($\alpha$)) and polynomial solutions {#subapp:REalphaPoly} --------------------------------------- The earlier subsections have had a narrow focus on the $g_n$ and ${\hat g}_n$ Fourier coefficients. In this subsection we will focus on polynomial solutions of (RE($\alpha$)). Since the general solution of the (RE($\alpha$)) difference equation can be formed from a linear combination of $Q^\mu_{n-1/2}$ and $P^\mu_{n-1/2}$ we will, later in this section, be led to consider the Legendre $P$ functions. ### Easy results While too elementary to suggest immediate generalization, we note that when we set $x=q_2=1$ in (RE($\alpha$)), the recurrence is easy to solve. For all $\alpha$ one solution is $u_n=1$ for all $n$. Reduction of order can be used to find another solution. In particular (RE(1/2)) is then solved in terms of the digamma function $\psi$, specifically by $u_n=\psi(n+1/2)$. We have $$\psi(n+\frac{3}{2}) = \psi(n+\frac{1}{2}) +\frac{1}{n+1/2} .$$ Forming the corresponding equation with $n$ replaced by $n-1$, and subtracting the two equations we find a solution. With $$\begin{aligned} u_n &=& (\psi(n+1/2)-\psi(1/2))/2, \qquad u_0=0,\ u_1=1 ,\\ u_n &=& 1-(\psi(n+1/2)-\psi(1/2))/2, \qquad u_0=1,\ u_1=0 ,\end{aligned}$$ we have solutions to the case $x=q_2=1$ of (RE(1/2)) When $\alpha=3/2$ the formulae are simpler. Then we have that the sequence ${\hat u}_n=(n^2-1)/(4n^2-1)$ satisfies (RE(3/2)) and has ${\hat u}_0=1$ and ${\hat u}_1=0$. A simple relationship between the ${\hat u}_n$ of this paragraph and the $u_n$ of the preceding is $$\frac{2}{4n^2-1} = \frac{\psi(n+3/2)-\psi(n-1/2)}{n} .$$ This can be regarded as a simple lead-in to Theorem \[thm:thm1\]. Next some results follow from very routine algebraic manipulation. Define $${\cal L}(\alpha,u,n):= 2 n q_2 u_n - (n+\alpha) u_{n+1} - (n-\alpha) u_{n-1} .$$ \[thm:thm1\] If $u_n$ solves [[(RE(1/2))]{}]{} and $${\hat u}_n = \frac{1}{n} \left( u_{n+1}- u_{n-1}\right ) ,$$ then $ {\hat u}_n $ solves [(RE(3/2))]{}. [*Proof.*]{} $$\begin{aligned} {\cal L}(\frac{3}{2},{\hat u},n) &=& 2 n q_2 \frac{u_{n+1}-u_{n-1}}{n} - (n+\frac{3}{2}) \frac{u_{n+2}-u_{n}}{n+1} - (n-\frac{3}{2}) \frac{u_{n}-u_{n-2}}{n-1} , \\ &=& \frac{1}{n^2-1}\left( (n-1) {\cal L}(\frac{1}{2},{ u},n+1) + (n+1) {\cal L}(\frac{1}{2},{ u},n-1)\right) ,\\ &=& 0 .\end{aligned}$$ This establishes the theorem. [**Corollary.**]{} [*If $u_n$ solves [[(RE(1/2))]{}]{} and $${\hat u}_n =\frac{x u_{n}- u_{n-1}}{2n+1} ,$$ then $ {\hat u}_n $ solves [(RE(3/2))]{}, and so too does $${\hat u}_n =\frac{x u_{n}- u_{n+1}}{2n-1} .$$*]{} (The latter is suggested by [@HTF2]3.8(6).) [Remarks.]{} - We have already seen in (\[eq:cross2\]) that $${\hat g}_n = \frac{1}{4 n (q_2+1)} \left( g_{n+1}- g_{n-1}\right ) .$$ See also [@HTF1] 3.8.3 - With the definitions in the later subsection on ‘Polynomial sequences’ we have the further examples $$\begin{aligned} -4{\hat p}_{01}(n)+12 x{\hat p}_{10}(n) &=& \frac{3}{n}\left(p_{10}(n+1)-p_{10}(n-1)\right) , \label{eq:L1ppBarn} \\ 4 x{\hat p}_{01}(n)-12{\hat p}_{10}(n) &=& \frac{3}{n}\left(p_{01}(n+1)-p_{01}(n-1)\right) . \label{eq:L2ppBarn}\end{aligned}$$ (Substituting $x=1$ and adding the equations has each side summing to 0 which provides one simple check.) \[thm:thm2\] If $u_n$ solves [(RE(1/2))]{} and $${\hat u}_n = \left( 2q_2 u_n - u_{n+1}- u_{n-1}\right ) ,$$ then $ {\hat u}_n $ solves [(RE(3/2))]{}. [Remark.]{} We have already seen in (\[eq:cross0\]) that $${\hat g}_n = \frac{1}{2(q_2+1)} \left( 2q_2 g_n - g_{n+1}- g_{n-1}\right ) .$$ There are many miscellaneous results. For example a routine calculation shows that \[thm:thm3a\] If ${\hat u}_n$ satisfies [(RE(3/2))]{}, then $(n^2-1/4){\hat u}_n$ satisfies [(RE(-1/2))]{}. (See also [@HTF2]3.8(1) with $\mu=-1$.) ### Equivalent recurrence relations There are many equivalent forms of recurrence relations. - Defining, from a sequence satisfying (RE(3/2)), the $a_n$ sequence via $${\hat u}_{n+1} = a_{n+1} \, (2 q_2)^n \, \frac{\Gamma(n+1)\Gamma(5/2)}{\Gamma(n+5/2)} = a_{n+1} \, (2 q_2)^n \, \frac{3\, 2^n\, n!}{(2n+3)!!} ,$$ we have, for $n>1$ $$a_{n+1}= a_n + \frac{(n-3/2)(n+1/2)}{4\, q_2^2\, n(n-1)} a_{n-1} . \eqno{\rm{(RE11(3/2))}}$$ (See also [@VZ07] equations (5.12) and (5.17) for (RE(1/2)).) While ${\hat u}_n$ might, for some initial conditions be polynomial in $q_2$ degree $n-1$, the $a_n$ are no longer polynomial, merely rational. Also, we cannot set $n=1$ in the last equation. - Some studies of recurrence equations for sequences of polynomials involve recurrences for monic polynomials, of the form $$v_{n+1}= x\,v_n - f(n)\, v_{n-1} . \eqno{{\rm (RE1xf)}}$$ Starting with a sequence of polynomials $u_n$ satisfying (RE($\alpha$)), one sets $v_n=u_n/g(n)$ and proceeds as follows: $$g(n+1) (n+\alpha) v_{n+1} = 2 n x g(n) v_n - g(n-1) (n-\alpha) v_{n-1} .$$ For this to be of the required form we must have $$g(n+1) (n+\alpha) = 2 n g(n) ,$$ which solves to $$g(n) = c\, \frac{n!}{\Gamma(n+\alpha)} .$$ From this $$f(n)=\frac{g(n-1) (n-\alpha)}{g(n+1) (n+\alpha)} = \frac{(n-\alpha)(n-1+\alpha)}{4(n-1)n} .$$ Thus, for $n>1$ $$v_{n+1}= x\, v_n + = \frac{(n-\alpha)(n-1+\alpha)}{4(n-1)n}\, v_{n-1} . \eqno{(\rm{RE1x}(\alpha))}$$ Again we cannot set $n=0$ or $n=1$ in the preceding (except, taking appropriate limits, in the cases $\alpha=0$ and $\alpha=1$). For the $\alpha=1/2$ case see [@Ch78] Chapter 6, equation (12.9). See also [@VZ07] equations (5.12) and (5.17), but note, as occurs with some other references, the indexing is different from ours, in this case their $S_n$ is our $v_{n+1}$ (monic form of $p_{01}$). There is another well-known approach to the calculation of $f$ for (RE1xf). Let $(v_n)$ be a sequence of monic polynomials, mutually orthogonal. We do not, at this stage, need to know the weight function but we will assume it is positive, even in $x$ and the interval is symmetric about $x=0$. Taking inner products of each side of (RE1xf) with $v_{n-1}$ gives $$0 = \langle x v_n, v_{n-1}\rangle - f_n \langle v_{n-1}, v_{n-1}\rangle .$$ Shifting the subscripts of the terms in (RE1xf) down by 1, i.e. $n$ replaced by $n-1$, and then taking the inner product with $v_n$ gives $$\langle v_n, v_{n}\rangle = \langle x v_{n-1}, v_{n}\rangle .$$ The preceding two equations yield $$f_n = \frac{\langle v_n, v_{n}\rangle} {\langle v_{n-1}, v_{n-1}\rangle} .$$ We can apply this at $\alpha=1/2$ to our polynomial sequence $p_{01}(n)$ defined later, see §\[subapp:PolyQ\] and (\[eq:p01PQ\]). The weight function is known explicitly and calculations (from [@BD67], reported in various places including  [@Ch78]) give $$\frac{\langle p_{01}(n), p_{01}(n)\rangle} {\langle p_{01}(n-1), p_{01}(n-1)\rangle} =\frac{n-1}{n} .$$ We need monic polynomials. So $$f_n = \frac{n-1}{n}\, \left(\frac{{\rm lcoeff}(p_{01}(n-1)} {{\rm lcoeff}(p_{01}(n)}\right)^2 .$$ Using the result on the leading coefficient given in equation (\[eq:lc01\]) we find the formula gives $$f_n =\frac{n-1}{n}\, \left( \frac{2n-1}{4(n-1} \right)^2 ,$$ which agrees with the $\alpha=1/2$ case given in (RE1x($\alpha$) above. - A different transformation $v_n=\psi(n,\alpha)\sqrt{n}\, u_n$, with $\psi$ specified later, but with $\psi(n,1/2)=1$, leads to another family of recurrence relations, of the form $$a_n v_{n+1} = x v_n - a_{n-1} v_{n-1} . \eqno{({\rm RE}a_n{\rm x}(\alpha))}$$ The recurrence $({\rm RE}a_n{\rm x}(\alpha))$ can be written $$\begin{pmatrix} v_{n+1}\\ a_n v_n \end{pmatrix} =A(n) \, \begin{pmatrix} v_{n}\\ a_{n-1} v_{n-1} \end{pmatrix} \ {\rm where\ } A(n)=\begin{bmatrix} \frac{x}{a_n}& -\frac{1}{a_n}\\ a_n& 0 \end{bmatrix} . \label{eq:Amat}$$ (We remark that $({\rm RE}a_n{\rm x}(\alpha))$ can be transformed to (RE1xf), the latter with a different $v$ of course, with $f(n)=a_{n-1}^2$.) The fact that ${\rm det}(A(n))=1$ leads to generalizations of the result we have at equation (\[eq:niceph\]). Specialise now to our (RE($\alpha$)). For (RE(1/2)) we have $$a_n = \frac{n+1/2}{2\sqrt{n(n+1)}} . \label{eq:anDef}$$ More generally, for (RE($\alpha$)) we have $$\psi(n,\alpha) =\sqrt{\frac{\Gamma(n+\alpha)}{\Gamma(n+1-\alpha}},\quad a_n = \frac{\sqrt{\Gamma(n+1+\alpha)\Gamma(n+2-\alpha)}} {\sqrt{n(n+1)\Gamma(n+1-\alpha)\Gamma(n+\alpha)}} .$$ Returning to the case $\alpha=1/2$, on using $a_n$ of equation (\[eq:anDef\]) in $A(n)$ of equation (\[eq:Amat\]), we find that the product of matrices $$A(n)\, A(n-1)\, \ldots A(2)\, A(1) = \begin{bmatrix} \sqrt{n+1}\,p_{01}(n+1)& 4\, \sqrt{n+1}\,{ p}_{10}(n+1)\\ \sqrt{n}\, a(n)\,p_{01}(n)& 4\, \sqrt{n}\,a(n)\, {p}_{10}(n) \end{bmatrix} ,$$ where $p_{01}$ and $p_{10}$ are the polynomials satisfying (RE(1/2)) starting 0,1 and 1,0 respectively. - It is possible that further work related to the items in this paragraph might be productive. We record them in the hope that some reader might suggest how it might be developed. - Next a definition. See [@Ch78]§5.5. A sequence of positive numbers $(f_n)$ is called a [*chain sequence*]{} if there exists another sequence $(g_n)$ such that $$f_n = g_n\, (1-g_{n-1}) ,\ {\rm with\ } 0\le g_0<1,\ 0<g_n<1 {\ \rm for\ } n=1,2,\ldots \ \ .$$ The $f(n,\alpha)=a_{n-1}^2$ above is a chain sequence. Use $$g_n=\frac{1}{2}-\frac{1}{4n} \ {\rm for\ \ } \alpha=\frac{1}{2} .$$ - In some papers the $a_n$ is written as a ratio $a_n=\gamma(n)/\gamma(n-1)$. For $\alpha=1/2$ the $a_n$ of equation (\[eq:anDef\]) has $$a_n=\frac{\gamma(n)}{\gamma(n-1)}\ \ {\rm with\ \ } \gamma(n) = \frac{2^{n-1}\, n!}{\sqrt{n}}\, \frac{\Gamma(1/2)}{\Gamma(n+1/2)} .$$ ### Legendre $P_{n-1/2}$ functions Formulae. involving Legendre $P_{n-1/2}$ functions, similar to those in (\[eq:Qmh\],\[eq:Qph\],\[eq:Qp1mh\],\[eq:Qp1ph\]) can be found. [$$\begin{aligned} P_{-1/2}(\cos(\theta)) &=& \frac{2}{\pi} {\rm EllipticK}(\sin(\frac{\theta}{2})) , \label{eq:PmhDL}\\ P_{1/2}(\cos(\theta)) &=& \frac{2}{\pi}\left( 2 {\rm EllipticE}(\sin(\frac{\theta}{2})) - {\rm EllipticK}(\sin(\frac{\theta}{2}))\right) , \label{eq:PphDL}\\ Q_{-1/2}(\cos(\theta)) &=& {\rm EllipticK}(\cos(\frac{\theta}{2})) , \label{eq:QmhDL}\\ Q_{1/2}(\cos(\theta)) &=& {\rm EllipticK}(\cos(\frac{\theta}{2})) - 2 {\rm EllipticE}(\cos(\frac{\theta}{2})) , \label{eq:QphDL}\\ P_{-1/2}(\cosh(\xi)) &=& \frac{2}{\pi\cosh(\frac{\xi}{2})} {\rm EllipticK}(\tanh(\frac{\xi}{2})) , \label{eq:PmhDLx}\\ P_{1/2}(\cosh(\xi)) &=& \frac{2\exp(\xi/2)}{\pi} {\rm EllipticE}(\sqrt{1-\exp(-2\xi)}) , \label{eq:PphDLx}\\ Q_{-1/2}(\cosh(\xi)) &=& \frac{2}{\sqrt{\pi}}\exp(-\frac{\xi}{2}) {\rm EllipticK}(\exp(-{\xi})) , \label{eq:QmhDLx}\\ Q_{1/2}(\cosh(\xi)) &=& \frac{2}{\sqrt{\pi}} \frac{\cosh(\xi)}{\cosh(\frac{\xi}{2})} {\rm EllipticK}(\frac{1}{\cosh(\frac{\xi}{2})}) -\\ &\ &\qquad\qquad \frac{4}{\sqrt{\pi}}\cosh(\frac{\xi}{2}) {\rm EllipticE}(\frac{1}{\cosh(\frac{\xi}{2})}) .\label{eq:QphDLx}\end{aligned}$$ ]{} See [@AS]8.13. There are many equivalent ways of expressing the formulae, and we will use, in a later section, [$$P_{-1/2}(x)=\frac{2}{\pi} {\rm EllipticK}\left( \sqrt{\frac{1-x}{2}}\right), \quad Q_{-1/2}(x)= {\rm EllipticK}\left( \sqrt{\frac{1+x}{2}}\right) \quad {\rm for\ } |x|<1 .$$ ]{} Though we don’t explicitly use the following it does relate to the weight $w$ in a later section being even, so we note $$\frac{\pi}{2}\,P_{-1/2}(-x)=Q_{-1/2}(x),\qquad {\rm and\ similarly\ \ } \frac{\pi}{2}\,P_{1/2}(-x)=-Q_{1/2}(x) .$$ The functions with order $\mu=-1$ arise in the case $\alpha=3/2$. The left-hand parts of the following formulae, corresponding to $\mu=1$ in [@AS]8.2.5, parallel those given earlier, (\[eq:Qm1mh\],\[eq:Qm1ph\]), $$\begin{aligned} P^{-1}_{-1/2}(q_2) &=& -4\, P^{1}_{-1/2}(q_2) , \label{eq:Pm1mh} \\ P^{-1}_{1/2}(q_2) &=& \frac{4}{3} P^{1}_{1/2}(q_2) . \label{eq:Pm1ph} \\ Q^{-1}_{-1/2}(q_2) &=& -4\, Q^{1}_{-1/2}(q_2) = \frac{4}{\sqrt{2(q_2-1)}}\, {\rm EllipticE}\left(\sqrt{\frac{2}{q_2+1}}\right) , \label{eq:Qm1mhC} \\ Q^{-1}_{1/2}(q_2) &=& \frac{4}{3} \, \frac{-q_2}{\sqrt{2(q_2-1)}}\, {\rm EllipticE}\left(\sqrt{\frac{2}{q_2+1}}\right)+\\ &\ & \qquad\qquad \frac{4}{3}\,\sqrt{\frac{q_2-1}{2}} {\rm EllipticK}\left(\sqrt{\frac{2}{q_2+1}}\right) . \label{eq:Qm1phC} \end{aligned}$$ There are many remarkable formulae involving both $P$ and $Q$ functions, including Whipple’s, which for integer $n$ and $\mu$ is, for $z>1$, $$\begin{aligned} P_{n-{\frac {1}{2}}}^{\mu}(z ) &=&{\frac {(-1)^{n}}{\Gamma (n-\mu+{\frac {1}{2}})}} \sqrt{\frac {2}{\pi}}\, \left(z^2-1\right)^{-1/4} Q_{\mu-{\frac {1}{2}}}^{n}(\frac{z}{\sqrt{z^2-1}} )\\ Q_{n-{\frac {1}{2}}}^{\mu}(z ) &=& {\frac {(-1)^{n}\pi }{\Gamma (n-\mu+{\frac {1}{2}})}} \sqrt{\frac {\pi}{2}}\, \left(z^2-1\right)^{-1/4} P_{\mu-{\frac {1}{2}}}^{n}(\frac{z}{\sqrt{z^2-1}} ) .\end{aligned}$$ When both $n$ and $\mu$ are zero, we have $$\begin{aligned} P_{-{\frac{1}{2}}}(z ) &=&\frac{\sqrt{2}}{\pi} \left(z^2-1\right)^{-1/4} Q_{-{\frac{1}{2}}}(\frac{z}{\sqrt{z^2-1}} ),\\ P_{-{\frac{1}{2}}}(\cosh(\eta)) &=& \frac{\sqrt{2}}{\pi} \left(\sinh(\eta)\right)^{-1/2} Q_{-{\frac{1}{2}}}(\coth(\eta)) ,\\ Q_{-{\frac{1}{2}}}(z ) &=& {\frac{\pi}{\sqrt{2}}}\, \left(z^2-1\right)^{-1/4} P_{-{\frac{1}{2}}}(\frac{z}{\sqrt{z^2-1}} ) ,\\ Q_{-{\frac{1}{2}}}(\cosh(\eta)) &=& \frac{\pi}{\sqrt{2}} \left(\sinh(\eta)\right)^{-1/2} P_{-{\frac{1}{2}}}(\coth(\eta)) .\end{aligned}$$ Another notable formula is from [@HTF2]3.3.1(1) $$P_{-n-1/2}^\mu(z) = P_{n-1/2}^\mu(z) . \label{eq:Pminus}$$ The corresponding formula for $Q$ is (\[eq:AS822\]) given in [@HTF2]3.3.1(3). There are formulae involving Wronskians and similar determinantal forms, e.g. for $z>0$ $$P_{1/2}(z) Q_{-1/2}(z)- Q_{1/2}(z) P_{-1/2}(z) = 2 . \label{eq:niceW0}$$ On substituting from equations (\[eq:PmhDL\])–(\[eq:QphDL\]) one finds that equation (\[eq:niceW0\]) is equivalent to Legendre’s relation (\[eq:LegRel\]) for the EllipticE and EllipticK functions. [@AS]8.1.8 gives, when $\mu=0$, $${\rm Wronskian}(P_\nu(z),Q_\nu(z)) =\frac{-1}{z^2-1} .$$ (See also  [@HTF1]3.4(25).) We also have from [@HTF1] 3.8 (9) and [@AS]8.5.2 $$\frac{d P_{n-1/2}(z)}{dz} = \frac{n^2-1/4}{\sqrt{z^2-1}} \, P^{-1}_{n-1/2} . \label{eq:HTF1Pp3p8p9}$$ Using this and the corresponding formula for the $Q$, namely (\[eq:HTF1Qp3p8p9\]) $$P_{n-1/2}(z) Q^{-1}_{n-1/2}(z) - P^{-1}_{n-1/2}(z) Q_{n-1/2}(z) = -\frac{1}{(n^2-1/4)\, \sqrt{z^2-1}} . $$ (See also [@GST00] equation (7), as in (\[eq:GST007\]) below.) The Wronskian expression and [@AS] 8.5.4 $$P_{n+1/2}(z) Q_{n-1/2}(z)- Q_{n+1/2}(z) P_{n-1/2}(z) = \frac{1}{n+1/2} . \label{eq:niceWn}$$ (See also equation (\[eq:niceph\]).) Equation (\[eq:niceWn\]) is the $m=0$ case of equation (6) of [@GST00]: $$P_{n+1/2}^m(z) Q_{n-1/2}^m(z)- Q_{n+1/2}^m(z) P_{n-1/2}^m(z) = \frac{\Gamma(n+\frac{1}{2}+m)}{\Gamma(n+\frac{3}{2}-m)} \, (-1)^m . \label{eq:GST006}$$ Equation (7) of [@GST00] allows for orders to be changed by 1 $$P_{n-1/2}^m(z) Q_{n-1/2}^{m+1}(z)- Q_{n-1/2}^m(z) P_{n-1/2}^{m+1}(z) = \frac{\Gamma(n+\frac{1}{2}+m)}{\Gamma(n+\frac{1}{2}-m)} \, \frac{(-1)^m}{\sqrt{z^2-1}} . \label{eq:GST007}$$ Writing $z$ where formerly we wrote $q_2$, the general solution of (RE(1/2)) is $$u_n(z) = f_P(z) P_{n-1/2}(z)+f_Q(z) Q_{n-1/2}(z) ,$$ where $f_P$ and $f_Q$ are independent of $n$. We can prescribe $u_0$ and $u_1$ and determine $f_P$ and $f_Q$. Using equation (\[eq:niceW0\]) we have $$f_P= -\frac{1}{2} (u_0 Q_{1/2}- u_1 Q_{-1/2}),\ f_Q= \frac{1}{2} (u_0 P_{1/2}- u_1 P_{-1/2}).$$ In particular, the sequence starting with $u_0=0$, $u_1=1$, and denoted by $p_{01}(n)$ in the next subsection is $$p_{01}(n)=\frac{1}{2} (Q_{-1/2} P_{n-1/2} -P_{-1/2} Q_{n-1/2} ). \label{eq:p01PQ}$$ Similarly $$p_{10}(n)= -\frac{1}{2} (Q_{1/2} P_{n-1/2} -P_{1/2} Q_{n-1/2} ). \label{eq:p10PQ}$$ While these must be the polynomial sequences tabulated in the next subsection, it isn’t immediately obvious how to extract information about the polynomials from the formula above. (For general $\alpha$, see [@BI82] equation (3.5).) One item that follows readily is that the above with equations (\[eq:AS822\]) and (\[eq:Pminus\]) imply $p_{01}(-n)=p_{01}(n)$, etc. ### Reduction of order We have already noted that reduction of order can be used when $x=q_2=1$ producing from the obvious constant solution to (RE($\alpha$)) a second linearly independent solution. This subsection fails to help so far in our quest to find more information on the polynomial solutions. However reduction of order gives equation (\[eq:wSol\])and this can be used to give a relation between toroidal $P$ and $Q$ functions. Assume $u_n$ is one solution of (RE($\alpha$)) and seek a second linearly independent solution $v_n$ with $$v_n = w_n \, u_n\qquad {\rm with\ \ } w_1\ne w_0 \quad {\rm and\ with\ all\ } u_n\ne{0-}{\rm function}.$$ Writing out the equation stating that $w_n u_n$ satisfies (RE($\alpha$)) and then eliminating $u_n$ from the fact that $u_n$ also satisfies (RE($\alpha$)) gives $$(n+\alpha) u_{n+1} (w_{n+1}-w_n)+ (n-\alpha) u_{n-1} (w_{n-1}-w_n)= 0.$$ Let $\theta_n=w_{n+1}-w_n$. Then $$(n+\alpha) u_{n+1}\theta_n -(n-\alpha) u_{n-1}\theta_{n-1} = 0.$$ Hence $$\frac{\theta_n}{\theta_{n-1}} =\frac{(n-\alpha)u_{n-1}}{(n+\alpha)u_{n+1}} ,$$ and hence $$\frac{\theta_n}{\theta_{n-2}} =\frac{(n-\alpha)(n-1-\alpha) u_{n-1} u_{n-2}} {(n+\alpha)(n-1+\alpha) u_{n+1} u_n} .$$ Now set $\alpha=1/2$. The last equation becomes $$\frac{\theta_n}{\theta_{n-2}} =\frac{(n-2+\frac{1}{2}) u_{n-1} u_{n-2}} {(n+\frac{1}{2}) u_{n+1} u_n} .$$ Continuing to multiply the ratios of $\theta$ we get $$\frac{\theta_n}{\theta_{0}} =\frac{(\frac{1}{2}) u_{1} u_{0}} {(n+\frac{1}{2}) u_{n+1} u_n} .$$ Our concern is with $w$ rather than $\theta$ and $$w_{n+1}-w_0 = \sum_{k=0}^n \theta_k .$$ Now $\theta_0$ is known and nonzero as $\theta_0=w_1-w_0$. Hence $$w_{n+1} = w_0 + \theta_0 u_0 u_1 \sum_{k=0}^n \frac{1}{(2k+1) u_{k+1} u_k} . \label{eq:wSol}$$ The preceding equation concludes the general reduction of order. However, a quick check is to return to the case $x=q_2=1$ and the known solution $u_k=1$ for all $k$. The reduction of order formula for $w_{n+1}$ involves the sum of the reciprocals of odd integers, and this is consistent with our previous formula involving the digamma function $\psi$. We remark that one could view one instance as $w_n=P_{n-1/2}/Q_{n-1/2}$ and then one has an identity involving Legendre functions. It presumably is a much disguised consequence of the recurrence relations (RE(1/2)) satisfied by the $P$ and $Q$. We remark also that if the input $u_n$ to the reduction of order is polynomial in $x=q_2$, the output $v_n$ is rational but not necessarily polynomial.\ For a formula similar to equation (\[eq:wSol\]) see [@Gr85] equation (14). Polynomial sequences relating to LegendreQ functions {#subapp:PolyQ} ---------------------------------------------------- Clearly we can write $$Q_{n-1/2}(z) = p_{10}(\alpha=1/2,n)\, Q_{-1/2}(z) + p_{01}(\alpha=1/2,n)\, Q_{1/2}(z) , \label{eq:Q1pol}$$ where the polynomials $p$ satisfy (RE(1/2)) and $$p_{10}(\alpha=1/2,0)=1,\ p_{10}(\alpha=1/2,1)=0,\qquad p_{01}(\alpha=1/2,0)=0,\ p_{01}(\alpha=1/2,1)=1 .$$ When space is too limited, we drop from the argument list the $\alpha=1/2$, leaving $n$ as the sole argument. ----------------------------------------------------------------------------------------------------------------------- $n$ $p_{01}(n)$ $p_{10}(n)$ ----- ----------------------------------------------------------- ----------------------------------------------------- $0$ $0$ $1$ $1$ $1$ $0$ $2$ $\frac{4}{3}\,x$ $-\frac{1}{3}$ $3$ ${\frac {32}{15}}\,{x}^{2}-\frac{3}{5}$ $-{\frac {8}{15}}\,x$ $4$ ${\frac {128}{35}}\,{x}^{3}-{\frac {208}{105}}\,x$ $-{\frac {32}{35}}\,{x}^{2}+{\frac {5}{21}}$ $5$ ${\frac {2048}{315}}\,{x}^{4}-{\frac {544}{105}}\,{x}^{2} $-{\frac {512}{315}}\,{x}^{3}+{\frac {88}{105}}\,x$ +{\frac {7}{15}}$ ----------------------------------------------------------------------------------------------------------------------- Plots of the $p_{01}$ polynomials are given in Figure \[fig:p01plot\]. ![Plots of $p_{01}(n)$ for $n=1$ to 5.[]{data-label="fig:p01plot"}](p01plot.jpg "fig:"){height="6cm" width="8cm"}\ [ ]{} Similarly, we write $$Q_{n-1/2}^{-1}(z) = {\hat p}_{10}(\alpha=1/2,n)\, Q_{-1/2}(z) + {\hat p}_{01}(\alpha=1/2,n)\, Q_{1/2}(z) , \label{eq:Q3pol}$$ where the polynomials ${\hat p}$ satisfy (RE(3/2)) and $${\hat p}_{10}(\alpha=3/2,0)=1,\ {\hat p}_{10}(\alpha=3/2,1)=0,\qquad {\hat p}_{01}(\alpha=3/2,0)=0,\ {\hat p}_{01}(\alpha=3/2,1)=1 .$$ Thus, $p_{01}$ and $p_{10}$ are the polynomials corresponding to $\alpha=1/2$ and ${\hat p}_{01}$ and ${\hat p}_{10}$ are those corresponding to $\alpha=3/2$. -------------------------------------------------------------------------------------------------------------------------- $n$ ${\hat p}_{01}(n)$ ${\hat p}_{10}(n)$ ----- ------------------------------------------------------------ ------------------------------------------------------- $0$ $0$ $1$ $1$ $1$ $0$ $2$ $\frac{4}{5}\,x$ $\frac{1}{5}$ $3$ ${\frac {32}{35}}\,{x}^{2}-\frac{1}{7}$ ${\frac {8}{35}}\,x$ $4$ ${\frac {128}{105}}\,{x}^{3}-{\frac {16}{35}}\,x$ ${\frac {32}{105}}\,{x}^{2}-{\frac {1}{15}}$ $5$ ${\frac {2048}{1155}}\,{x}^{4}-{\frac {416}{385}}\,{x}^{2} ${\frac {512}{1155}}\,{x}^{3}-{\frac {232}{1155}}\,x$ +{\frac {5}{77}}$ -------------------------------------------------------------------------------------------------------------------------- We have seen in earlier sections that there are many connections between the polynomials at $\alpha=1/2$ and $\alpha=3/2$. For example, defining $$p_{1x}(n) = p_{10}(n) + x\, p_{01}(n) ,$$ Theorem \[thm:thm1\] suggests relations with the ${\hat p}$. We have $$\frac{p_{1x}(n+1)-p_{1x}(n-1)}{x^2-1} = \frac{3\, {\hat p}_{01}(n)}{4\, n} .$$ This also follows from equations (\[eq:L1ppBarn\]) (\[eq:L2ppBarn\]).) Various quantities, whose notation has been standardised, as in [@HTF2]§[10]{}, are associated with sequences of orthogonal polynomials. These include the leading coefficient, often denoted $k_n$, moments, often denoted $c_n$ and discussed in the later subsection with the weight function, and the norm squared of the polynomial, denoted $h_n$. (The quantities $k_n$ and $h_n$ occur in [@HTF2]10.3(11) in connection with a Christoffel-Darboux formula, a formula we will present in a later subsection.) For the family $p_{01}$ we have $${\rm lcoeff}(p_{01}(n)) = k_n = \frac{2^{2n-2} (n-1)!}{(2n-1)!!} . \label{eq:lc01}$$ The leading coefficients of the other sequences of polynomials satisfy, for $n\ge{2}$, $$\begin{aligned} {\rm lcoeff}(p_{01}(n)) &=& -4\, {\rm lcoeff}(p_{10}(n)) ,\\ {\rm lcoeff}({\hat p}_{01}(n)) &=& 4\, {\rm lcoeff}({\hat p}_{10}(n)) ,\\ {\rm lcoeff}({\hat p}_{01}(n)) &=& \frac{3}{2n+1}\, {\rm lcoeff}(p_{01}(n)) ,\\ {\rm lcoeff}({\hat p}_{10}(n)) &=& -\frac{3}{2n+1} {\rm lcoeff}(p_{10}(n)) .\end{aligned}$$ [@HTF2]10.3(7,8) writes the recurrence relation in the form $$p_{n+1}= A_n x\, p_n - C_n p_{n-1} ,$$ and notes that $$A_n = \frac{k_{n+1}}{k_n}, \qquad C_n= \frac{k_{n+1}\, k_{n-1}}{k_n^2}\,\frac{h_n}{h_{n-1}} .$$ For us at $\alpha=1/2$, $$A_n=\frac{4n}{2n+1},\ \frac{h_n}{h_m}=\frac{m}{n},\ C_n=\frac{2n-1}{2n+1} .$$ Our interest is in $g_n$ and ${\hat g}_n$, so $\alpha=1/2$ and $\alpha=3/2$. There remains the possibility that investigating other values of $\alpha$ might ultimately be useful, and, perhaps lead to the discovery of further ‘contiguous relations’ such as those in $\mu$ for the Legendre functions. In this connection we record some results on $\alpha=-1/2$. --------------------------------------------------------------------------------------------------------------------- $n$ ${ p}_{01}(\alpha=-\frac{1}{2},n)$ ${ p}_{10}(\alpha=-\frac{1}{2},n)$ ----- ---------------------------------------------------------- ---------------------------------------------------- $0$ $0$ $1$ $1$ $1$ $0$ $2$ $4\,x$ $-3$ $3$ ${\frac {32}{3}}\,{x}^{2}-\frac{5}{3}$ $-8\,x$ $4$ ${\frac {128}{5}}\,{x}^{3}-{\frac {48}{5}}\,x$ $-{\frac {96}{5}}\,{x}^{2}+{\frac {21}{5}}$ $5$ ${\frac {2048}{35}}\,{x}^{4}-{\frac {1248}{35}}\,{x}^{2} ${-\frac{1536}{35}}\,{x}^{3}+{\frac {696}{35}}\,x$ +{\frac {15}{7}}$ --------------------------------------------------------------------------------------------------------------------- We focus on connections between the solutions of (RE(-1/2)) and those of (RE(1/2)) and (RE(3/2)). As noted in Theorem \[thm:thm3a\], $$p_{01}(\alpha=-1/2,n) = \frac{4 n^2 -1}{3}\, {\hat p}_{01}(n) .$$ ### Derivatives of $p_{01}$ and of $p_{10}$ [**Ordinary differential equations**]{} The Legendre differential equation satisfied by $ Q_{n-1/2}(z)$ leads to coupled d.e.s for $p_{01}(n)$ and $p_{10}(n)$ (where we drop the reminder $\alpha=1/2$). [$$\begin{aligned} (1-x^2) {\frac {d^{2}p_{10}(n)}{d{x}^{2}}}- x{\frac {d p_{10}(n)}{dx}}+ p_{10}(n){n}^{2} + {\frac {d p_{01}(n)}{dx}} &=& 0 \label{eq:p10de} ,\\ (1-x^2) {\frac {d^{2}p_{01}(n)}{d{x}^{2}}}- 3\, x\, {\frac {d p_{01}(n)}{dx}} + ({n}^{2}-1)\,p_{01}(n)- {\frac {d p_{10}(n)}{dx}} &=& 0 . \label{eq:p01de} $$ ]{} From these one can find 4-th order differential equations satisfied by the polynomials. For example $p_{01}(n)$, satisfies, with $$\begin{aligned} P &=& (1-x^2)^2 , \\ Q &=& 10 x (1-x^2) , \\ R &=& 2n^2 (1-x^2)-9 +25 x^2 , \\ S &=& -3 x (2n^2 -5) , \end{aligned}$$ the differential equation $$P\frac{d^4 p_{01}(n)}{d x^4} + Q\frac{d^3 p_{01}(n)}{d x^3} + R\frac{d^2 p_{01}(n)}{d x^2} + S\frac{d p_{01}(n)}{d x} + (n^2-1)^2 p_{01}(n) = 0.$$ The notation is as in [@Me66]. The conditions in the theorems in [@Me66] are [*not*]{} satisfied, but some equations similar to his are. We have, with $W_M=\sqrt{1-x^2}$ which is [*not*]{} the weight function for us, $$\begin{aligned} W_M Q &=& 2\frac{d}{dx}\left(W_M P\right) ,\\ W_M S &=&\frac{d}{dx}\left(W_M(R-1)\right) - \frac{d^3}{dx^3}\left(W_M P\right) . \end{aligned}$$ The remainder of this subsection is speculative (and won’t be included in any paper). We also don’t see any use for the main goals of the paper. We have yet to write the d.e. in the form $${\cal L}_4 u = f(x,n) {\cal L}_2 u ,$$ with, in some appropriate sense, the 4th order ${\cal L}_4$ and the 2nd order ${\cal L}_2$ being somehow ‘self-adjoint’. Associated with this effort, we would like to be able to express derivatives of the $w(x)$ defined in equation (\[eq:weightK\]) in terms of a rational function of $w(x)$, $x$ and $\sqrt{1-x^2}$ (or something similarly simple). We believe there is much more to be discovered, e.g. in connection with variational characterisations of the (generalized eigen)solutions of the fourth order operator. We expect that the qualitative behaviour would be akin to the generalized matrix eigenvalue problem $$A u = \lambda B u ,$$ with matrices $A$ and $B$ self-adjoint (real-symmetric). The function spaces involved might include the space of all polynomials. Having found (eigen)solutions polynomial of degree $n$, Rayleigh Ritz might be appropriate to produce solutions of degree $n+1$. Different inner products may be appropriate in different circumstances. [**The Wronskian of $p_{01}$ and $p_{10}$**]{} We find, at any fixed $n$, using equations (\[eq:dp10\]), (\[eq:dp01\]) to eliminate the derivative terms and then equation (\[eq:niceph\]) to eliminate the polynomials indexed by $n-1$, $$\begin{aligned} {\rm Wronskian}(p_{01},p_{10}) &=& (p_{01}\, p_{10}' - p_{10}\, p_{01}') , \nonumber \\ &=& \frac{(p_{01}+ p_{10})^2 +2\, (x-1)p_{01}\, p_{10} -1} {2(x^2-1)} . \label{eq:Wn}\end{aligned}$$ ### Contiguous relations We have already seen several instances of ‘contiguous relations’. - Some of these involve only the polynomials at a fixed value of $\alpha$ as in (RE($\alpha$)) and, in this category, we mention (\[eq:niceph\]) and (\[eq:niceph3\]) below. - Some of these involve polynomials at a fixed $n$ and the different sequences $p$ at $\alpha=1/2$ and $\hat p$ at $\alpha=3/2$. One such, namely (\[eq:L1ppBarn\]), (\[eq:L2ppBarn\]), was a simple consequence of Theorem \[thm:thm1\]. - Others involve not merely the polynomials, but also their first derivatives. We now see an identity (\[eq:niceph\]) for the polynomial solutions of (RE(1/2)) which is similar to (\[eq:niceWn\]) applying to LegendreQ. At $\alpha=1/2$ we have $$p_{10}(n)\, p_{01}(n+1)- p_{10}(n+1)\, p_{01}(n) = \frac{1}{2n+1} . \label{eq:niceph}$$ (See also the preceding subsection on alternative formulations of the recurrence relation, specifically (RE1x$a_n$).) At $\alpha=3/2$ we have $${\hat p}_{10}(n)\, {\hat p}_{01}(n+1)- {\hat p}_{10}(n+1)\, {\hat p}_{01}(n) = \frac{-3}{(2n-1)(2n+1)(2n+3)} . \label{eq:niceph3}$$ (There is a similar result for $\alpha=0$: $$T_n(x)\, U_{n+1}(x) - T_{n+1}(x)\, U_n(x) = 1.$$ And, again, another similar result for $\alpha=1$.) These are easily proved by induction, as given in the following generalization which applies not just to polynomial solutions of the difference equations. \[thm:induct\] Let $u_n$ and $U_n$ be two sequences satisfying [(RE($\alpha$))]{}. Then $$u_n\, U_{n+1} - u_{n+1}\, U_n = c(x)\, \frac{\Gamma(n+1-\alpha)}{\Gamma(n+1+\alpha)} ,$$ where $$c(x) = \left(u_0\, U_{1} - u_{1}\, U_0\right)\, \frac{\Gamma(1+\alpha)}{\Gamma(1-\alpha)} .$$ [*Proof.*]{} The proof is by induction. Define $${\cal S}(n) = u_n\, U_{n+1} - u_{n+1}\, U_n - c(x)\, \frac{\Gamma(n+1-\alpha)}{\Gamma(n+1+\alpha)} ,$$ with $c(x)$ as in the statement of the theorem. We have, from the definition of $c(x)$, that ${\cal S}(0)=0$. We will next show that ${\cal S}(n-1)=0$ implies that ${\cal S}(n)=0$. Let ${\cal S}^{*}(n)$ be the expression obtained from ${\cal S}(n)$ on using (RE($\alpha$)) to write $u_{n+1}$ in terms of $u_n$ and $u_{n-1}$ and, similarly, to write $U_{n+1}$ in terms of $U_n$ and $U_{n-1}$. One finds $${\cal S}^{*}(n) = \frac{n-\alpha}{n+\alpha}\, {\cal S}(n-1) .$$ Thus ${\cal S}(n-1)=0$ implies ${\cal S}(n)=0$ so, as asserted, the theorem is established by induction. Formulae involving the derivatives of the polynomials can also be found. The starting point in deriving (\[eq:dp10\]) and (\[eq:dp01\]) below is $$(x^2-1)\, \frac{d Q_\nu}{d x} = \nu\, (x Q_\nu - Q_{\nu-1}) .$$ See [@HTF1] 3.8.10. Also $$\begin{aligned} (x^2-1)\, \frac{d Q_{1/2}}{d x} &=& \frac{1}{2}\, (x Q_{1/2}-Q_{-1/2}) ,\\ (x^2-1)\, \frac{d Q_{-1/2}}{d x} &=& -\frac{1}{2}\, (x Q_{-1/2}-Q_{1/2}) .\end{aligned}$$ This leads to the somewhat lengthy identities [$$\begin{aligned} 2(x^2-1)\, \frac{d p_{10}(j)}{d x} &=& 2 j x p_{10}(j)+p_{01}(j)-(2j-1)p_{10}(j-1) , \label{eq:dp10}\\ 2(x^2-1)\, \frac{d p_{01}(j)}{d x} &=& (2j-2)x p_{01}(j)-p_{10}(j)-(2j-1)p_{01}(j-1) . \label{eq:dp01}\end{aligned}$$ ]{} There are also formulae involving the four sequences of polynomials, both the $p$ sequences and the $\hat p$ sequences. The starting point for equations (\[eq:pB01n\]) and (\[eq:pB10n\]) below is that derivatives with respect to $z$ of $Q_{n-1/2}^\mu(z)$ can be expressed in terms of the undifferentiated function at a different value of $\mu$. In particular, begin with equation (\[eq:HTF1Qp3p8p9\]). After some calculation one finds $$\begin{aligned} (4n^2-1)\,{\hat p}_{01}(n) &=& 6\frac{d}{dx}\left(x\,p_{01}(n)+p_{10}(n)\right) - 3p_{01}(n) , \label{eq:pB01n} \\ (4n^2-1)\,{\hat p}_{10}(n) , &=& 2\frac{d}{dx}\left(x\,p_{10}(n)+p_{01}(n)\right) - 3p_{10}(n) . \label{eq:pB10n}\end{aligned}$$ ### Christoffel-Darboux (CD) As in [@HTF2]10.3(11) we have $$\sum_{j=1}^n j p_{01}(j)^2 = \frac{2n+1}{4}\, \left( p_{01}(n)\, \frac{d p_{01}(n+1)}{d x} - \frac{d p_{01}(n)}{d x} \, p_{01}(n+1) \right) .$$ The CD kernel, defined by $$K(x,y) = \sum_{j=1}^n j p_{01}(j,x)\, p_{01}(j,y) ,$$ has the reproducing property $$K(x,y)= \langle K(x,\cdot),K(\cdot,y) \rangle .$$ ### A convolution identity involving Legendre polynomials A special case of [@BI82] equation (3.13) is $$p_{10}(n) = -\sum_{k=0}^n\frac{{\rm LegendreP}(k,x)\, {\rm LegendreP}(n-k,x)}{2k-1} . \label{eq:p10Leg}$$ Another identity (also in [@Ch78] (12.6) is $$p_{01}(n) = -\sum_{k=0}^{n-1}\frac{{\rm LegendreP}(k,x)\, {\rm LegendreP}(n-1-k,x)}{2k+1} . \label{eq:p01Leg}$$ (Equation (\[eq:p01Leg\]) can also be found using the generating function given in (\[eq:GenDef\]).) We have yet to find similar concise identities for $\alpha=3/2$ and the $\hat p$ polynomials. Of course, using equations (\[eq:p01Leg\]), (\[eq:p10Leg\]) in equation (\[eq:L2ppBarn\]) gives the $4x{\hat p}(n)-12{\hat p}(n)$ in terms of Legendre polynomials. Presumably related to $\alpha=0$ or $\alpha=1$ we have the well-known identity ([@Ch78] (12.7)): $${\rm ChebyshevU}(n,x) = \sum_{k=0}^n {\rm LegendreP}(k,x)\, {\rm LegendreP}(n-k,x) .$$ We have yet to find a similar identity for ChebyshevT. There are, of course, many other ways the identities can be written as, if $f$ is any real-valued function odd about 0 (and any functions of $(k,x)$ not just LegendreP), $$0 = \sum_{k=0}^n f(k-\frac{n}{2}) {\rm LegendreP}(k,x)\, {\rm LegendreP}(n-k,x) .$$ A typical example would be $f(k-\frac{n}{2})=k-\frac{n}{2}$. Allowing other Gegenbauer polynomials than the Legendre $P$ polynomials above, we remark that one obtains convolution identities on using $\gamma=-\beta=\alpha-1$ in [@BI82] equation (3.13). $\alpha=1/2$: orthogonality and related results {#subapp:orthog} ----------------------------------------------- Results from [@BD67] are reported in [@Ch78] and [@Gr85]. Equation (12.3) of [@Ch78] and equation(1) of [@Gr85] defines for the interval $(-1,1)$, a weight function $w$ (in the case $\nu=1/2$ appropriate to our problem) by $$w(x) = \frac{4}{Q_{-1/2}(x)^2 +(\pi P_{-1/2}(x)/2)^2} . \label{eq:wPQ}$$ Using representations of the toroidal functions in terms of complete elliptic integrals, EllipticK, this rewrites to $$w(x) = \frac{4}{K\left(\sqrt{\frac{1-x}{2}}\right)^2+ K\left(\sqrt{\frac{1+x}{2}}\right)^2} . \label{eq:weightK}$$ (See also [@VZ07] where, with $A_n$ satisfying their (1.5), $k^n A_n$ satisfies (RE(1/2)). Also their $\psi_n$ satisfying their (4.10) satisfy our (RE(1/2)), except for a shift. Weight functions involving EllipticK occur in their (3.7), (3.9), (5.19) and, close to the form we report in (5.20) and (5.21). See also [@Rees45].) We have yet to investigate (RE(1/2)) in connection with study of examples of ‘polynomials orthogonal on the (boundary of the) unit disk’: see equation (\[eq:Kdisk\]). Define the inner product $$\langle u,v\rangle = \int_{-1}^1 u(x)\, v(x)\, w(x)\, dx .$$ Direct calculation (via wolframalpha) for several $(m,n)$ pairs confirms $$\langle p_{01}(m),p_{01}(n)\rangle = \frac{\delta_{mn}}{n} . \label{eq:p01ort}$$ Shortly we will give an elementary derivation of this without reference to the details of the inner product. The moments are defined by $$c_n = \langle 1, x^n\rangle ,$$ and are obviously 0 when $n$ is odd. We have $$c_0=1, \ c_2=\frac{9}{2^5},\ c_4=\frac{39}{2^8},\ c_6=\frac{6633}{2^{16}} ,$$ but have yet to find the formula for $c_{2n}$. Once all the $c_n$ are calculated, [@HTF2] 10.3(4) gives a determinantal formula for the orthogonal polynomials. Our derivation of equation (\[eq:p01ort\]) involves considering $$\phi(n)= \langle x^{n-1}, p_{01}(n) \rangle.$$ (See also [@Ch78] Theorems 3.2 and 4.2.) Multiplying the recurrence $$2n x p_{01}(n) = (n+\frac{1}{2})\, p_{01}(n+1) + (n-\frac{1}{2}) \, p_{01}(n-1) ,$$ by $x^{n-2}$ and taking the inner product gives $$2 n \phi(n) = (n-\frac{1}{2})\, \phi(n-1) .$$ If one agrees that $\phi(1)=\langle 1,1\rangle =1$, this first order recurrence for $\phi$ solves to $$\phi(n)=\frac{(2n-1)!!}{4^{n-1}\, n!} =\frac{1}{n\, {\rm lcoeff}(p_{01}(n))} ,$$ where the leading coefficient lcoeff is given by equation (\[eq:lc01\]). Next consider $\langle p_{01}(n), p_{01}(n) \rangle$ beginning with considering $$p_{01}(n) ={\rm lcoeff}(p_{01}(n))\, x^{n-1} +\ {\rm lower\ order} ,$$ and noting that the inner product of the lower order monomials with $p_{01}(n)$ is 0. Thus $$\langle p_{01}(n), p_{01}(n) \rangle = {\rm lcoeff}(p_{01}(n)) \, \langle x^{n-1}, p_{01}(n) \rangle = {\rm lcoeff}(p_{01}(n)) \, \phi(n) .$$ From the preceding calculated value of $\phi(n)$ we find the value stated in (\[eq:p01ort\]). ### $\alpha$ generally The orthogonality result is also available, in a more general form in [@HTF2],p220,§10.21 where one of the Pollaczek sequences is treated. The weight function, for our $\alpha=1/2$ case is given in [@HTF2]10.21(14) when one sets $$a=b=0=t, \qquad c=\lambda= \frac{1}{2} .$$ These values in (13) yield our (RE(1/2)) while, in (14), the weight function involves, as before, EllipticK functions as the hypergeometric expression in (14) is $${}_2{F}_1(\frac{1}{2},\frac{1}{2};1;z) =\frac{2}{\pi} K(\sqrt{z}) .$$ For $\alpha=3/2$ we have,in both (13) and (14), $$a=b=0=t, \qquad c=\frac{3}{2},\ \lambda= -\frac{1}{2} .$$ This time the weight function given in (14) is $ {}_2{F}_1(\frac{3}{2},\frac{3}{2};1;z)$ which, again, can be expressed in terms of EllipticE and EllipticK. We comment further on the $\alpha=3/2$ case below. ### More concerning $\alpha=3/2$ We noted, in connection with Pollaczek polynomials above, that the weight function for our $\alpha=3/2$ case is available. However, the setting is considerably more general than appropriate for our very special cases. Gegenbauer polynomials (themselves special cases of Jacobi polynomials) satisfy the recurrence $$(n+1)\, C_{n+1}^\lambda{(x)} = 2(n+\lambda)x\, C_n^\lambda{(x)}- (n+2\lambda-1) C_{n-1}^\lambda{(x)} . \eqno({\rm REG})$$ - Our (RE(1)), satisfied by scaled Chebyshev polynomials, is the $\lambda=0$ case of this. - (RE(0)) is the $\lambda=1$ case of (REG). However (REG) does not include the (RE($\alpha$)) needed in our paper. The case $\lambda=1/2$ in (REG) gives Legendre polynomials. [@BI82]§3, equation (3.1) modify this to $$(n+\lambda+1)\, C_{n+1}^\lambda{(x,\beta)} = 2x\,(n+\beta+\lambda)x\, C_n^\lambda{(x,\beta)}- (2\beta + n+\lambda-1) C_{n-1}^\lambda{(x,\beta)} . \eqno({\rm BI3.2})$$ With $\lambda=\alpha-1=-\beta$ we have our (RE($\alpha$)). In particular - $\lambda=\beta=0$ is our (RE(1)). - $\lambda=-1=-\beta$ is our (RE(0)). - $\lambda=-1/2=-\beta$ is our (RE(1/2)). - $\lambda=1/2=-\beta$ is our (RE(3/2)). ### An aside on other cases in [@Gr85] This is an aside as the only case which is amongst our (RE($\alpha$)) is the case $\alpha=1/2$. The recurrence relation at the beginning of [@Gr85] is, with a shift so as to accord with our indexing, $$(n+\nu) p_{n+1} = (2n+2\nu-1)x\, p_n - (n+\nu-1)\, p_{n-1} .$$ We have already used that when $\nu=1/2$ this is our (RE(1/2)). The weight function for orthogonality is $$w(\nu,x) =\frac{1}{(\nu Q_{\nu-1}(x))^2 +(2\nu P_{\nu-1}(x)/\pi)^2 } .$$ Now, the limit of $w(\nu,x)$ as $\nu$ tends to 0 is 1, so we will write $$w(0,x) =1 .$$ The recurrence is solved by the Legendre polynomials and these are orthogonal with respect to the weight $w(0,x)$. We also have $$\begin{aligned} w(1,x) &=& \frac{1}{Q_0(x)^2 + (\pi P_0(x)/2)^2} ,\\ &=& \frac{1}{\left(\frac{1}{2}\log|\frac{1+x}{1-x}|\right)^2 + \frac{\pi^2}{4}} .\end{aligned}$$ Repeating work in [@Gr85], the usual notation for the LegendreQ functions is $$Q_n(x) = P_n(x) Q_0(x) - W_{n-1}(x) .$$ In [@Gr85] there is the demonstration that the polynomials $W_n$ are orthogonal with respect to the weight $w(1,x)$. Bessel functions {#subapp:Bessel} ---------------- In view of one of our goals being to sum series such as that for $u_\infty$ there is a case for investigating representations of the toroidal functions. Connections with Bessel functions are available. See [@Sonn59] where the reference is to Watson’s classic book on Bessel functions. See also [@CV12] equation (3) and references there. The conditions on the parameters are not those we need here. The formulae below are sometimes approached via Laplace transforms. We have, for $q>1$, $$\sqrt{q}\int_0^\infty \exp(-t\sqrt{q})\, {\rm BesselI}(0,t\cos(\psi))\, dt = g(\psi) =\frac{1}{\sqrt{1- \frac{\cos(\psi)^2}{q}}} . \label{eq:besselg}$$ The Fourier coefficients of $g$, $Q_{n-1/2}(q_2)$ can also be written in terms of integrals of Bessel functions. We have, for $s>1$, $$\int_0^\infty \exp(-s\, t)\, {\rm BesselI}(n,\frac{t}{2})^2\, dt = \frac{2}{\pi} Q_{n-1/2}(2 s^2- 1) ,$$ and its application here would have $s=\sqrt{q}$. See [@GR] 6.612.3. Boundary moments revisited {#subapp:BndryMom} -------------------------- We remark that the perimeter can be expressed in terms of a toroidal function: see equation (\[eq:ab09P\]). Having introduced the $g_n$ and ${\hat g}_n$, we might note that boundary moments can be expressed in terms of them. For example, as $$r^2= x^2+ y^2 = a^2 (1-e^2)\left( 1-\frac{e^2}{2} + \frac{e^2}{2} \cos(2\psi) \right) ,$$ we have $$\begin{aligned} i_2 &=& a \int_{-\pi}^\pi r^2\, {\hat g}(\psi)\, d\psi\\ &=& a^3 (1-e^2)\left( ( 1-\frac{e^2}{2})\frac{{\hat g}_0}{2} + \frac{e^2}{2} {\hat g}_1 \right) .\end{aligned}$$ There may be useful recurrence relations involving the $i_n$. It may be that there are neat expressions for any moment, not merely the ‘polar’ moments of the preceding paragraph, involving toroidal functions, both of the LegendreP and LegendreQ kinds. Appendix: Result Validation and Numerical Examples {#sec:Numerical} ================================================== Slip flow in Elliptic Channel with $a=2,\;b=\frac{1}{2}$ {#subsec:Spiga} -------------------------------------------------------- Figure \[fig:ContourPlotBetajpg\] gives, at left, contour plots of $u$ for $\beta=\frac{1}{4},\;\frac{1}{16}$ and $\frac{1}{64}$, and, at right, those for $\beta=4,~16,~64$. The contours are very similar to those in Figure 4 of [@SV12]. It is noted that when $\beta>0$ on the boundary (as well as in the interior) $u>0$. At any given fixed position $z$ on the boundary, $u(z)$ increases when $\beta$ value increases. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta1p4.PNG "fig:"){height="2cm" width="7cm"} ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta4.PNG "fig:"){height="2cm" width="7cm"} $\beta=\frac{1}{4}$ $\beta=4$ ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta1p16.PNG "fig:"){height="2cm" width="7cm"} ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta16.PNG "fig:"){height="2cm" width="7cm"} $\beta=\frac{1}{16}$ $\beta=16$ ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta1p64.PNG "fig:"){height="2cm" width="7cm"} ![$a=2$. Contour plot of axial velocity on the ellipse for six different $\beta$ values.[]{data-label="fig:ContourPlotBetajpg"}](\grpath/ContourBeta64.PNG "fig:"){height="2cm" width="7cm"} $\beta=\frac{1}{64}$ $\beta=64$ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Comparing the results with Ritz Method for Slip Flow {#subsec:Wang} ---------------------------------------------------- Here we compare with results given in [@Wa14]. $\lambda$ $c$ $a=\frac{1}{\sqrt{c}}$ $\beta=\frac{\lambda}{\sqrt{c}}$ A V F Ritz method ----------- ----- ------------------------ ---------------------------------- ----------- ----------- ------------ ------------- -- .25 2.000000 .200000 .027080 .026893 .0268994 .0268994 0.1 .5 1.414213 .141421 .131628 .131224 .1312286 .1312286 .75 1.154700 .115470 .313497 .313320 .3133204 .3133204 .25 2.000000 .400000 .042611 .041996 .0419990 .0419990 0.2 .5 1.414213 .282843 .184716 .183372 .1833741 .1833741 .75 1.154700 .230940 .414937 .414338 .4143388 .4143388 .25 2.000000 1.000000 .071907 .086515 .0865169 .0865167 0.5 .5 1.414214 .707107 .343981 .338284 .3382847 .3382847 .75 1.154700 .577350 .719257 .716699 .7166986 .7166986 .25 2.000000 2.000000 .143814 .159568 .1595802 .1595797 1.0 .5 1.414214 1.414214 .509349 .594535 .5945397 .5945397 .75 1.154701 1.154701 1.004665 1.219750 1.2197505 1.2197505 .25 2.000000 4.000000 .287629 .304352 .3043838 .3043838 2.0 .5 1.414214 2.828427 1.018698 1.105106 1.1051201 1.1051201 .75 1.154700 2.309401 2.009330 2.224974 2.2249748 2.2249748 .25 2.000000 10.000000 .719072 .736625 .7366869 .7366851 5.0 .5 1.414214 7.071068 2.546745 2.634130 2.6341541 2.6341541 .75 1.154701 5.773503 5.023326 5.239416 5.2394175 5.2394175 .25 2.000000 20.000000 1.438144 1.456036 1.4561139 1.4561106 10 .5 1.414214 14.142136 5.093491 5.181257 5.1812861 5.1812861 .75 1.154701 11.547005 10.046652 10.262916 10.2629186 10.2629186 Blood flow problem {#subsec:Numerical} ------------------ Most of the veins and arteries in our bodies can be taken as having circular cross-section. However in places where the vein or artery has to go through a region which is squeezed in by muscle or bone one might expect the cross-section of the vein or artery to depart somewhat from circular. In this appendix, we take fluid density,$\rho$, of 1.05 $g/mL$ and fluid viscosity, $\mu$, of 0.04 $Poise$. See Table 2 in [@Wblood]. The geometry of elliptical cross-section of the channel is described by coordinates $(x, y)$ with $$x=\overline{r}(\psi) \cos(\psi);\;\;\; y=\overline{r}(\psi) \sin(\psi),$$ where $$\overline{r}(\psi)=\lambda k(1+\kappa(k, \psi)),\;\;\;0 \leq \psi \leq 2\pi$$ and $$\kappa = -\frac{1}{4}(k-1)^2+(k-1-\frac{1}{2}(k-1)^2)\cos(2\psi)+\frac{3}{4}(k-1)^2\cos(k \psi).$$ The tube ellipticity $\varepsilon=\sqrt{1-b^2/a^2}$ with the lengths of the half-axes $a=\bar{r}(0)\cos(0)$ and $b=\bar{r}(\pi/2)\sin(\pi/2)$ is determined by setting $\lambda=0.005$ and $k=1.15$. Thus, the tube has the ellipticity of 0.6720 with $a=6.6\; \mu m$, $b=4.9\; \mu m$, $\eta_0\;=\;0.9518$ and $c=4.4 \;\mu m$. The boundary of the elliptical cross-section is described by $$\partial \Omega: (x,y)=(c \cosh \eta_0 \cos \psi, c \sinh \eta_0 \sin \psi).$$ ![The cross sections of elliptical shapes with the same length of minor axis $b = 4.9 \;\mu m.$ and various lengths of major axis $a=6.6,\; 7.1,\; 7.6,\; 8.1$ and $8.5 \;\mu m.$ []{data-label="fig2"}](\grpath/ellipse.PNG "fig:")\ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Axial velocity profiles in $x$ and $y$ directions obtained from the model with no slip lengths: $\ell =0.0$, and various shapes of cross section.[]{data-label="fig3"}](\grpath/axialVx.PNG "fig:") ![Axial velocity profiles in $x$ and $y$ directions obtained from the model with no slip lengths: $\ell =0.0$, and various shapes of cross section.[]{data-label="fig3"}](\grpath/axialVy.PNG "fig:") ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Axial velocity profiles obtained from the model with $a=6.6 \;\mu m.$ and $b := 4.9 \;\mu m.$[]{data-label="fig4"}](\grpath/AxialVxy.PNG "fig:")\ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Axial velocity profiles in $x$ and $y$ directions obtained from the model having $a=6.5 \;\mu m.$ and $b = 4.9\;\mu m$ and various slip lengths $\ell (\mu m.)$ of 0.0 (solid line), 6.25(dotted line), 25 (dashed line), 37.5 (dash-dotted line) and 50 (long dashed line).[]{data-label="fig5"}](\grpath/axialVx2.PNG "fig:") ![Axial velocity profiles in $x$ and $y$ directions obtained from the model having $a=6.5 \;\mu m.$ and $b = 4.9\;\mu m$ and various slip lengths $\ell (\mu m.)$ of 0.0 (solid line), 6.25(dotted line), 25 (dashed line), 37.5 (dash-dotted line) and 50 (long dashed line).[]{data-label="fig5"}](\grpath/axialVy2.PNG "fig:") --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- The investigate the influence of elliptical shapes on constant pressure-driven flow of fluid, various sizes of the major axis $a$ are chosen to be vary between $6.6 \;\mu m.$ to $8.6 \;\mu m.$ while the length of minor axis is fixed as $b = 4.9 \;\mu m$.\ To demonstrate the impact of the slip length $\ell$ on the constant pressure-driven flow of fluid, values of slip length $\ell$ are chosen to be vary from $0.0$ to $50 \;\mu m.$. The results show that the slip length has a direct influence on the axial velocity. Larger slip length provides higher velocity as shown in Figures \[fig3\], \[fig4\] and \[fig5\]. Appendix: Rectangular domains {#app:Rect} ============================= Rectangles: introduction {#subsec:Rect} ------------------------ In this subsection we will consider the rectangle $$\Omega=(-a,a)\times(b,b)\qquad a=rh, \ b=\frac{h}{r} ,$$ and usually treat the case $h=1$ when, of course we have the area of the rectangle as 4. (Outside the context of flows with slip, the corresponding problem for box shapes in $R^n$ can be treated in exactly the same way. See [@MK94].) We believe the value of our approximation, good lower bound (\[in:lbbS\],\[in:lbbS\]), is that it applies to all simply-connected domains. We expect that it will be relatively uncommon for all five functionals involved in the inequality to be explicitly available. A more common situation would be that some would involve computation. In this respect rectangles, besides being a geometry that is relatively common in practice, is reasonably typical. ### The functionals ${Q}_0$, $\Sigma_\infty$ and $\Sigma_1$, etc. Before presenting the functionals which depend on solving pde problems we note the geometric functionals which we use: $$|\Omega|=4a\, b = 4\, h^2,\ \ |\partial\Omega|=4 (a+b ) = 4\, h\, (r + \frac{1}{r} ) .$$ $B$ defined in [@PoS51] is $$B=4(a/b + b/a) = 4 (r^2 + \frac{1}{r^2} .)$$ We use this expression for $B$ below in deriving a very simple lower bound for ${Q}(\beta)$. The calculation of ${Q}_0$ is an undergraduate exercise in separation of variables: see [@LSU79] Problem 133, p74. The quantity ${Q}_0$ is available only as an infinite series, $${Q}_0 = \frac{4 a^3b}{3}\left( 1-\frac{192}{\pi^5}\,\frac{a}{b}\sum_{n=1}^\infty \frac{\tanh(\pi(2n-1)b/(2a))}{(2n-1)^5})\right) .$$ We don’t know a closed form for its sum. Also, indicating dependence on the sides as ${Q}_0(a,b)$, it is ‘physically’ evident that ${Q}_0(a,b)={Q}_0(b,a)$. Also it is known (by different arguments in [@PoS51]) that ${Q}_0(r,1/r)$ attains its maximum for a square, $r=1$. Neither of the last two sentences are immediate from the series. However computation of ${Q}_0(a,b)$ isn’t difficult, just one line of code. The other quantities in the inequality (\[in:lbbS\],\[in:lbbS\]) are extremely easy to calculate, and have very simple formulae, just rational functions of the rectangle’s side lengths. Once again it is clear that ${\cal J}(v_{\rm max})(a,b)$ is unchanged if we swap $a$ and $b$ as this is true of all five functionals. There are also series solutions for the case $\beta>0$ in $R^n$ as well as the $R^2$ case relevant to flow in ducts: see [@MK94; @DM07; @DG14; @WWS14; @KW16]. The series are very substantially messier than the $\beta=0$ case, and require about 20 lines of code. Once again ${Q}_\beta(a,b)-{Q}_\beta(b,a)$ should be zero and we have found for rectangles with largish aspect ratios and truncated sums there can be noticible differences. If, however, an approximation is adequate, it is clear that the rational function of $\beta$ is far easier code. Having said this, nevertheless, the value of our approximation is that it is simple, and general, not merely for rectangular ducts. Let’s start the rectangle calculation. The well-known series solution for ${Q}_0$ is easily coded: Much easier are the calculations associated with Problem (P($\infty$)). It is easily verified that $$u_\infty = \,{\frac {ab \left( {a}^{2}+6\,ab+{b}^{2} \right) } {6 \left( a+b \right) ^{2}}}-{\frac {b{x}^{2}+a{y}^{2}}{2\,(a+b)}} ,$$ and hence $$\Sigma_\infty = {\frac {8 {a}^{3}{b}^{3}}{ 3\left( a+b \right) ^{2}}} .$$ From [@KWK18] we note the easy result:\ [*Amongst rectangles with a given area, that which maximizes $\Sigma_\infty$ is the square.*]{}\ These results (and further terms in the expansion for ${Q}_r(\beta)$ at large $\beta$) are presented in [@KWK18]. Also $$\Sigma_1 = - {\frac {4}{45}}\,{\frac {{a}^{2}{b}^{2} \left( {a}^{4}+6\,b{a}^{3}-10\,{a}^{2}{b}^{2}+6\,a{b}^{3}+{b}^{4} \right) } { \left( a+b \right) ^{3}}}$$ We remark that, both for ellipses and rectangles, when $\beta$ is moderately large but the aspect ratio $a/b$ is very large, the $\Sigma_1/\beta$ term can be relatively large compared to $\Sigma_\infty$. Asymptotics at large $\beta$ require care when the aspect ratio is large. ### Comparison of the lower bound with the series We now compare the lower bound with the elaborate series solution: see Figure \[fig:r5o4ratiojpg\]. ![A plot at $a=5/4$, $b=4/5$ of the ratio $(S(\beta)-{\cal J}(v_{\rm max}))/{\cal J}(v_{\rm max})$ as a function of $\beta$. This tends to 0 both as $\beta\rightarrow{0}$ and as $\beta\rightarrow\infty$. \[fig:r5o4ratiojpg\]](r5o4ratiojpg.jpg){height="7cm" width="13cm"} Using quadratic test functions {#subsec:VarlRectq} ------------------------------ As we did with the ellipse, once again choose quadratic test functions. This time we cannot expect useful results for $\beta$ very small. Indeed, at $\beta=0$ we obtain just the trivial lower bound of 0. However the results are good for $\beta$ large. $$Q_{r,LB}= \,{\frac {4{a}^{2}{b}^{2}\beta\, \left( 15\,a\beta+{b}^{2}+5\,ab\right) \left( 5\,ab+{a}^{2}+15\,b\beta \right) } {3(30\,a{b}^{3}\beta+5\,{b}^{4}\beta+30\,{a}^{3}b\beta+5\,{a}^{4}\beta+ 75\,{a}^{2}b{\beta}^{2}+75\,a{b}^{2}{\beta}^{2}+ 2\,{a}^{2}{b}^{3}+2\,{b}^{2}{a}^{3})}} \label{eq:QrLB}$$ With $b=1/a$ when one plots both this lower bound and the series solution, the curves are very close for all $\beta>1$. [00]{} P. Abbott, On the perimeter of an ellipse, [*The Mathematica Journal*]{}, [**11**]{}(2), (2009). M. Abramowitz and I.A. Stegun, [*Handbook of Mathematical Functions*]{}, (Dover: 1965). P. Barracund and D. Dickinson, On the associated Legendre polynomials. In [*Orthogonal expansions and their continuous analogues*]{} (Southern Illinois University, Carbondale, 1967), 43–50. M. van den Berg and D. Bucur, On the torsion function with Robin or Dirichlet boundary conditions, arXiv 1305.2137 (2013). R.H. Boyer, Discrete Bessel functions, [*J. Math. Analysis and Applications*]{}, [**2**]{} (1961) 509-524. D. Bucur and A. Giacomini, The St Venant inequality for the Laplace operator with Robin boundary conditions, [*Milan J. Math*]{} [**83**]{} (2015), 327-343.\ doi:10.1007/s00032-015-0243-0 J. Bustiz and M.E.H. Ismail, The associated ultraspherical polynomials and their $q$-analogues, [*Canadian J. Math*]{} [**34**]{}(3) (1982), 718-736 C.S. Chihara, [*An Introduction to Orthogonal Polynomials*]{}, (Gordon and Breach: 1978). H. S. Cohl and D. E. Dominici, Generalized Heine’s identity for complex Fourier series of binomials, [*Proc. Roy. Soc. A*]{} [**467**]{} (2010)\ `https://doi.org/10.1098/rspa.2010.0222` H. S. Cohl and H. Volkmer, Definite integrals using orthogonality and integral transforms. [*SIGMA*]{} [**8**]{} (2012) J. T. Conway, Fourier and other series containing associated Legendre functions for incomplete Epstein-Hubbell integrals and functions related to elliptic integrals, [*Integral Transforms and Special Functions*]{}, [**18:3**]{} (2007), 179-191.\ `doi: 10.1080/10652460701210284` D. Cvijovic, The Fourier series expansions of the Legendre incomplete elliptic integrals of the first and second kind, [*Integral Transforms and Special Functions*]{}, [**21**]{}(3) (2009), 235-242.\ `http://dx.doi.org/10.1080/10652460903178552` Y. Damiano and G.C. Georgiou, Viscoplastic Poiseuille flow in a rectangular duct with wall slip, [*Journal of NonNewtonian Fluid Mech.*]{} [**214**]{} (2014).\ `doi:10.1016/j.jnnfm.2014.10.002` S.K. Das and F. Tahmouresi, Analytical solution of fully developed gaseous slip flow in elliptic microchannel, [*Int. J. Adv. Appl. Math. and Mech.*]{} [**3i**]{} (2016) 1 - 15. S. D. Daymond, The principal frequencies of vibrating systems with elliptic boundaries, [*Quart. J. Mech. Appl. Math.*]{}, [**8**]{}, 1955, pp361-372. Z Duan and Y.S. Muzychka, Slip flow in non-circular microchannels, [*Microfluidics and Nanofluidics*]{} [**3**]{}, (2007), 473-484. Z. Duan, Y.S. Muzychka, Slip flow in elliptic micro-channels, [*International Journal of Thermal Sciences*]{} [**46**]{}, 1104–1111 (2007). A. Erdelyi, W. Magnus, F. Oberhettinger and F.G. Tricomi, [*Higher Transcendental Functions, Vol 1*]{}, (McGraw-Hill: 1953) A. Erdelyi, W. Magnus, F. Oberhettinger and F.G. Tricomi, [*Higher Transcendental Functions, Vol 2*]{}, (McGraw-Hill: 1953) S. Finch, Fraenkel asymmetry, (2014). A.K. Ghosh, A. R. Khan, L. Debnath, On a pulsatile flow of a two phase viscous fluid in a tube of elliptic cross section, [*Internat. J. Math. & Math. Sci.*]{} [**13**]{}(4), 669-676 (1990). A. Gil, J. Segura, and N.M. Temme, Computing Toroidal Functions for Wide Ranges of the Parameters, [*Journal of Computational Physics*]{} [**161**]{} (2000), 204–217.\ `doi:10.1006/jcph.2000.6498` I.S. Gradshteyn and I.W. Ryzhik, [*Tables of Integrals, Series and Products*]{}, (Academic Press: 1965). G.C. Grosjean, Theory of recursive generation of systems of orthogonal polynomials: An illustrative example. [*Journal of Computational and Applied Mathematics*]{} [**12-13**]{}, 299-318 (1985). S. Gupta, D. Poulikakos and V. Kurtcuoglua. Analytical solution for pulsatile viscous flow in a straight elliptic annulus and application to the motion of the cerebrospinal fluid, [*Phys. Fluids*]{} [**20**]{}, 093607 (2008);\ `http://dx.doi.org/10.1063/1.2988858` H. Haslam and M. Zamir. Pulsatile flow in tubes of elliptic cross sections, [*Annals of Biomedical Eng.*]{} [**26**]{}, 780–787 (1998). W.M. Hicks, On toroidal functions. XIV [*Proc. Roy. Soc. London*]{} (1881).\ `https://royalsocietypublishing.org/doi/pdf/10.1098/rstl.1881.0014` Jyh-Ping Hsu, Chen-Yuan Kao, Shiojenn Tseng, and Chur-Jen Chen, Electrokinetic Flow through an Elliptical Microchannel: Effects of Aspect Ratio and Electrical Boundary Conditions, [*Journal of Colloid and Interface Science*]{} [**248**]{}, 176–184 (2002). M. Kaewbumrung, S. Orankitjaroen, P. Boonkrong, B. Nuntadilok and B. Wiwatanapataphee. Numerical Simulation of Dispersed Particle-Blood Flow in the Stenosed Coronary Arteries, [*International Journal of Differential Equations*]{} (2018), Article ID 2593425, 16 pages\ `https://doi.org/10.1155/2018/2593425` G. Keady, On Hadwiger’s results concerning Minkowski sums and isoperimetric inequalities for moments of inertia, [*RGMIA Research Report Collection*]{} [**9**]{}(4) (2006). G. Keady and A. McNabb, Functions with constant Laplacian satisfying homogeneous Robin boundary conditions, [*I.M.A. Jnl of Applied Mathematics*]{}, [**50**]{} (1993), 205-224.\ doi:10.1093/imamat/50.3.205 KW16 G. Keady and B. Wiwatanapataphee, 2016, Some isoperimetric results concerning unidirectional flows in microchannels, arXiv. See also URL\ `http://arxiv.org/abs/1604.03394` G. Keady, B. Wiwatanapataphee and N. Khajohnsaksumeth, On Unidirectional Flows in Microchannels with Slip at the Boundary In [*Proceedings of The International Conference on Engineering and Applied Sciences*]{}, 2018 (TICEAS). (Most conveniently available via ResearchGate.) D.S. Kim and Y.H. Kim, A characterization of ellipses, [*American Math. Monthly*]{} [**144**]{}(1) (2007), 66-70. H. Kober, [*Dictionary of Conformal Representations*]{}, (Dover: 1957) H. Lamb, [*Hydrodynamics*]{}, (Dover printing of 1932 edition.) D.F. Lawden, [*Elliptic Functions and Applications*]{}, (Springer, 1989). N.N. Lebedev, L.P. Skalskaya and Y.S. Uflyand, [*Worked Problems in Applied Mathematics*]{}, (Dover: 1979) A. Liemert and A. Kienle, Exact solution of Poisson’s equation with an elliptical boundary, [*Appl. Math. and Computation*]{} [**238**]{} (2014), 123-131. A. McNabb and G.Keady, Diffusion and the torsion parameter, [*J. Australian Math. Soc.*]{}, [**35B**]{}, (1994), 289-301.\ doi:10.1017/S0334270000009309 J.W. Meux, Ordinary differential equations of the fourth order with orthogonal polynomial solutions, [*Amer. Math. Monthly*]{} [**73**]{}(4), (1966), 104-110. R.F. Millar, An integral equation solution to the Dirichlet problem for Laplace’s equation in an ellipse, [*J. Math. Anal. Applns*]{} [**147**]{} (1990), 154-170. F. Pollaczek, Sur une famille de polynomes orthogonaux a quatre parametres. [*C R Acad Sci. Paris*]{}. (1950) [**230**]{}2254–2256. G. Polya and G.Szego. [*Isoperimetric Inequalities in Mathematical Physics*]{}, Ann. of Math. Studies no. 27, (Princeton University Press, Princeton, 1951). C.J. Rees, Elliptic orthogonal polynomials, [*Duke Math. J.*]{} [**12**]{} (1945), 173-187. W.C. Royster, A Poisson integral formula for the ellipse and some applications, [*Proc. Amer. Math. Soc.*]{} [**15**]{}(4) (1964), 661-670.\ `DOI:10.2307.2034769` N.B. Savasaneril and H. Delibas, Analytic solution for two-dimensional heat equation for an ellipse region, [*New Trends in Math. Sciences, NTMSci*]{} [**4**]{}(1) (2016), 65-70.\ `doi:http://dx.doi.org/10.20852/ntmsci.2016115614` B.O.U. Sonnerup, Expression as a Legendre function, of an elliptic integral occurring in wing theory, TAR-TN 59-1, November 1959. S. Suharsono, Analytical study of fluid flows with slip boundary, Ph.D. Thesis, (2012)\ ` https://espace.curtin.edu.au/handle/20.500.11937/2281` S. Suharsono, A study slip flow through elliptic microchannels, [*Sci.Int(Lahore))*]{}, [**30**]{}(3),423-426 (2018). Q. Sun, Y. Wu, L. Liu, B. Wiwatanapataphee. Study of a Newtonain fluid through circular channels with slip boundary taking into account electrokinetic effect, [*Abstract and Applied Analysis*]{}, 2013. M. Spiga and P. Vocale, Slip flow in elliptic microducts with constant heat flux, [*Advances in Mechanical Engineering*]{} (2012), Article ID 481280.\ doi:10.1155/2012/481280 L. Vinet and A. Zhedanov, Elliptic biorthogonal polynomials connected with Hermite’s continued fraction, [*Symmetry, Integrability and Geometry: Methods and Applications*]{} (SIGMA) [**3**]{} (2007) B.A. Troesch and H.R. Troesch, Eigenfrequencies of an elliptic membrane, [*Mathematics of Computation*]{} [**27**]{}(124):755-755 (1973).\ doi: 10.1090/S0025-5718-1973-0421276-2 C.Y. Wang, Ritz method for slip flow in super-elliptic ducts, [*European J. of Mechanics B/Fluids*]{} [**43**]{}, (2014), 85-89. B. Wiwatanapataphee, Y.H. Wu, and S. Suharnso, Transient flows of Newtonian fluid through a rectangular microchannel with slip, [*Abstract and Applied Analysis*]{} paper 530605 (2014).

ArXiv

--- abstract: | We report the detection of CO molecular line emission in the z=4.5 millimeter-detected galaxy COSMOS\_J100054+023436 (hereafter: [J1000+0234]{}) using the IRAM Plateau de Bure interferometer (PdBI) and NRAO’s Very Large Array (VLA). The $^{12}$CO(4-3) line as observed with PdBI has a full line width of $\sim1000\,{\>{\rm km}\,{\rm s}^{-1}}$, an integrated line flux of 0.66 Jy${\>{\rm km}\,{\rm s}^{-1}}$, and a CO luminosity of 3.2$\times10^{10}\,{\>{\rm L_{\odot}}}$. Comparison to the 3.3$\sigma$ detection of the CO(2-1) line emission with the VLA suggests that the molecular gas is likely thermalized to the J=4-3 transition level. The corresponding molecular gas mass is $\rm 2.6\times10^{10}\,{\>{\rm M_{\odot}}}$ assuming an ULIRG-like conversion factor. From the spatial offset of the red- and blue-shifted line peaks and the line width a dynamical mass of $\rm 1.1\times10^{11}\,{\>{\rm M_{\odot}}}$ is estimated assuming a merging scenario. The molecular gas distribution coincides with the rest-frame optical and radio position of the object while being offset by 0.5” from the previously detected Ly$\alpha$ emission. [J1000+0234  ]{}exhibits very typical properties for lower redshift (z$\sim$2) sub-millimeter galaxies (SMGs) and thus is very likely one of the long sought after high redshift (z$>$4) objects of this population. The large CO(4-3) line width taken together with its highly disturbed rest-frame UV geometry suggest an ongoing major merger about a billion years after the Big Bang. Given its large star formation rate (SFR) of $>1000{\>{\rm M_{\odot}}}\,yr^{-1}$ and molecular gas content this object could be the precursor of a ’red-and-dead’ elliptical observed at a redshift of z=2. author: - 'E. Schinnerer, C.L. Carilli, P. Capak, A. Martinez-Sansigre, N.Z. Scoville, V. Smolčić, Y. Taniguchi M.S. Yun, F. Bertoldi O. Le Fevre, L. de Ravel' title: | Molecular Gas in a Submillimeter Galaxy at z=4.5:\ Evidence for a Major Merger at 1 Billion Years after the Big Bang --- Introduction ============ Wide field blank sky surveys at millimeter and sub-millimeter wavelengths have established a population of active star forming galaxies at high redshift [e.g. @bla02]. These sources or so-called submillimeter galaxies (SMGs) dominate the (sub-)millimeter background, and represent (50-75)% of the star formation at high redshift causing a significant revision to the optically derived star formation history of the Universe [@sma97; @hug98; @bar98; @ber00]. The bulk of this population lies at redshifts below z=3, with far-infrared (FIR) luminosities of $\ge10^{13}$L$_\sun$, which (if powered by star-formation) imply star-formation rates in excess of 1000M$_\sun$ yr$^{-1}$. This is sufficient to build up the stellar mass of a giant elliptical galaxy in about 1 Gyr [@cha05]. At redshifts z$\le$3 about 100 SMGs are now known, many of which have been studied in detail, including CO observations that indicate that they are very massive systems [@gre05]. Recent high resolution studies of the molecular gas in z$\sim$2-3 SMGs [@tac06; @tac08] showed that the star forming regions are fairly compact and that the SMGs resemble “scaled-up and more gas-rich versions” of the local Ultra-Luminous Infra-Red Galaxies [ULIRGs; e.g. @dow98]. Due to their derived central densities, which are close to those of ellipticals or massive bulges, they appear to form stars at the maximal rate over very short time scales (”maximum starburst”). Thus the SMGs phase appears to last for about 100Myr [@tac08]. These massive starburst galaxies place tight constraints on galaxy formation models [e.g. @bau05]. One major focus of current and future (sub-)mm surveys is to identify these massive starbursts at z$>$3, a redshift range for which SMGs could place tight constraints on hierarchical merger vs. monolithic collapse models in galaxy formation scenarios. The discovery of such an SMG above a redshift of z=4 indicates that massive galaxy formation is already well under-way when the universe is only 1.5 Gyr old. The recently discovered object originally selected as a V band drop-out with a weak radio counterpart [@car08] and with a stellar mass of $\rm M_*>10^{10}\,{\>{\rm M_{\odot}}}$, a young starburst age of 2-8Myr and a star formation rate of SFR $>1000\,{\>{\rm M_{\odot}}}yr^{-1}$ is a candidate for such a z$>$4 SMG, as it was detected in its FIR continuum [@cap08]. Throughout this paper we assume a standard concordance cosmology ($\rm H_{\circ}=70, \Omega_M=0.3, \Omega_{\Lambda}=0.7$). Observations ============ The bolometer camera MAMBO-2 at the IRAM 30m telescope was used in on-off mode in December 2007 and January 2008 to measure the 1.2mm flux density of [J1000+0234]{}. The weather conditions during the observations were good, the reduction was performed using MOPSIC (written by R. Zylka, IRAM). The achieved rms for a total on-sky integration of 70min is 0.67mJy. [J1000+0234  ]{}was detected at a 5$\sigma$ level with a total 1.2mm continuum flux density of 3.4mJy. The $^{12}$CO(4-3) line tracing cold molecular gas in [J1000+0234  ]{}was observed with the IRAM Plateau de Bure interferometer (PdBI) between January and April 2008 in the special C and C configurations. Both receivers were tuned to the redshifted line frequency of 83.3257GHz covering a total bandwidth of $\sim$ 0.9GHz. For calibration and mapping, we used the standard IRAM GILDAS software packages CLIC and MAPPING [@gui00]. The quasar 1005+058 was used for atmospheric calibration while standard calibrators were used for flux calibration. The 12hr of total integration time result in a data cube of 2.3”$\times$1.9” resolution (using robust weighting) with an rms of $\rm 0.42\,mJy\,beam^{-1}$ for 20MHz ($\sim72\,{\>{\rm km}\,{\rm s}^{-1}}$) wide channels. We used the VLA in D configuration to observe the $^{12}$CO(2-1) line in June and July 2008 for a total of 21hr. The two 50 MHz wide correlator units (each corresponding to a velocity width of $\sim360\,{\>{\rm km}\,{\rm s}^{-1}}$) were tuned to 41.6352GHz (on-line) and 41.7351GHz (off-line). The observations were done in the fast-switching mode using the nearby quasar 1018+055, at a distance of 5.3$^{\circ}$ from our source, for atmospheric amplitude and phase calibration. 20 antennas were available during the observations and the phase stability was typically $\sim10^{\circ}$. The final images using natural weighting have a resolution of 1.9”$\times$1.4” and an rms of 49$\rm \mu Jy\,beam^{-1}$. All observed and derived properties of [J1000+0234  ]{}are summarized in Tab. \[tab:prop\]. The Molecular Gas Properties ============================ The CO(4-3) line was detected at a 5.5$\sigma$ level (Fig. \[fig:co\]) in a 280MHz wide channel, which corresponds to a full line width at zero intensity (FWZI) of $\sim1000\,{\>{\rm km}\,{\rm s}^{-1}}$, i.e. with an integrated flux of 0.66Jy$\rm {\>{\rm km}\,{\rm s}^{-1}}$. The peak position of the CO(4-3) emission (see Fig. \[fig:comp\]) agrees very well within the positional uncertainties with the rest-frame near-IR and radio position derived by [@cap08]. The line emission of CO(4-3) is centered at an observed frequency of 83.1857GHz, which corresponds to a redshift of z=4.5423 (assuming a rest frequency of 461.040768GHz for $^{12}$CO(4-3)). Given the large line width this is in agreement with the optically derived redshift of z=4.547$\pm$0.002 [@cap08]. The CO(4-3) emission coincides with the rest-frame optical counterpart (as traced by the IRAC data) of [J1000+0234  ]{}(Fig. \[fig:comp\]). Averaging the remaining channels blue- and redwards of the detected line emission results in a line-free image of 620MHz width ($\sim2200\,{\>{\rm km}\,{\rm s}^{-1}}$) with a tentative 2.2$\sigma$ detection of the continuum at the rest-frame optical/radio position of the source (Fig. \[fig:co\]). The VLA observations with a channel width of $\sim360\,{\>{\rm km}\,{\rm s}^{-1}}$ are centered at the middle of the line detected in CO(4-3). CO(2-1) line emission is detected at the 3.3$\sigma$ level and coincides spatially with the CO(4-3) emission within its positional uncertainties of FWHM$_{beam}$/(S/N)$\sim$0.5”. Only about a third of the total line width was covered during the VLA observations, so the detected line flux of 0.057Jy${\>{\rm km}\,{\rm s}^{-1}}$ could be higher by a factor of $\sim$3, taking the missed blue- and redshifted emission into account by assuming a box-car line shape. The ratio of the CO(4-3) line flux of 0.66Jy${\>{\rm km}\,{\rm s}^{-1}}$ to the thus corrected CO(2-1) line flux of $\sim$0.16Jy${\>{\rm km}\,{\rm s}^{-1}}$ is about 4, implying that the molecular gas is still thermalized at the J=4-3 transition, as the flux still increases with $\nu^2$ at this transition. Continuum emission from [J1000+0234  ]{}was detected at 1.2mm at (3.4$\pm$0.67)mJy, while the off-line channels in the interferometric observations resulted in a 3$\sigma$ upper limit of 150$\mu$Jy at 41.7GHz (VLA) and a tentative 2.2$\sigma$ detection of (0.17$\pm$0.08)mJy at 83.3GHz (PdBI). All these values are consistent with the SED fits presented by [@cap08][^1] and a dust temperature of 30 - 50 K, typically found in z$\sim$2 SMGs [@pop06; @kov06]. Masses and Typical SMG Properties ================================= As the molecular gas appears to be thermalized at least up to the CO(4-3) transition as inferred above, its line flux can be used to estimate the CO luminosity \[setting $\rm L'_{CO}(J=4-3)=L'_{CO}(J=1-0)$\] and hence the molecular gas mass present in [J1000+0234]{}. To derive the CO luminosity $\rm L'_{CO}$ and molecular gas mass $\rm M_{gas}$ we use equations 3 and 4 from [@sol05] with a conversion factor for $\rm L'_{CO}$ to $\rm M_{gas}$ of $\alpha$=0.8$\rm M_{\odot}(K\,{\>{\rm km}\,{\rm s}^{-1}}\,pc^2)^{-1}$ derived for local ULIRGs [@dow98] and used for high redshift objects [@sol05]. The observed line flux of $\rm S_{CO(4-3)}$=0.66Jy${\>{\rm km}\,{\rm s}^{-1}}$ corresponds to a CO luminosity of $\rm L'_{CO}\sim3.2\times10^{10}\,L_{\odot}$ or a molecular gas mass of $\rm M_{gas}\sim2.6\times10^{10}\,{\>{\rm M_{\odot}}}$. We imaged the red and blue ($\approx$ 500${\>{\rm km}\,{\rm s}^{-1}}$ wide) halves of the line emission separately, resulting in a 4.7$\sigma$ (red) and 3.8$\sigma$ (blue) emission peak, respectively. The two peaks (Fig. \[fig:velo\]) show a spatial shift of $\sim$ 1” roughly from northwest to southeast (with a positional uncertainty of $\sim$0.5”). This offset corresponds to $\sim$ 6.6kpc at the redshift of [J1000+0234]{}, suggesting that the CO emission might be fairly extended. Using this spatial offset between the red- and blue-shifted half of the line emission and the line width of $\sim$1000${\>{\rm km}\,{\rm s}^{-1}}$, we can estimate the dynamical mass of [J1000+0234]{}. We use the relation between dynamical mass $M_{dyn}$, velocity width $\rm \Delta\,v_{FWHM}=v_{rot}(r)\times sin\,{\it i}/2.4$ and radial extent $r$ of the CO emitting region: $\rm M_{dyn}\times sin^2{\it i} = 4\times10^4\,r\,\Delta\,v^2_{FWHM}$ [@ner03]. As the system is likely merging (as discussed below) we include also a factor of 2 for a merging system assuming that the gas has already settled into the new potential [@gen03]. The estimated dynamical mass of [J1000+0234  ]{}is about 1$\times$10$^{11}{\>{\rm M_{\odot}}}$ for values of $\rm v_{rot}(r)\times sin\,{\it i} \sim 1000\,{\>{\rm km}\,{\rm s}^{-1}}$ and a radius $r$ of 3.3kpc. Thus [J1000+0234  ]{}has a gas fraction of $\rm M_{gas}/M_{dyn} \sim 25\%$, typical for SMGs [@gre05; @tac06]. We would like to caution that the derived offset of $\sim$1” has significant uncertainties and thus the derived dynamical mass could be overestimated if the separation were smaller. The properties of [J1000+0234  ]{}are very similar to the global properties derived for z$\sim$2 SMGs in terms of molecular gas mass ($\rm <M_{gas}> \sim 3\times10^{10}{\>{\rm M_{\odot}}}$), extent ($\rm r_{CO} \sim 2\,kpc$), CO line width ($\rm <FWHM_{CO}> \sim 780\,{\>{\rm km}\,{\rm s}^{-1}}$), dynamical mass ($\rm <M_{dyn}> \sim 1.2\times10^{11}\,{\>{\rm M_{\odot}}}$) and molecular gas mass fraction ($\sim$ 25%) [@gre05]. Its estimated current gas consumption time scale of $\rm \tau=M_{gas}/SFR \approx 30\,Myr$ assuming our gas mass and a SFR of $\rm \ge 1000\,{\>{\rm M_{\odot}}}\,yr^{-1}$ [@cap08] is similar to the gas depletion rates of $\sim$ 100Myr found for the z$\sim$2 SMGs [@tac08]. Taking together the stellar mass of $\rm \ge 10^{10}\,{\>{\rm M_{\odot}}}$ produced in a recent burst $\sim$ 7Myr ago [@cap08] and the available gas reservoir and gas consumption time scale derived here, a significant fraction of the stellar mass of a massive elliptical could be produced on a relatively short time scale. Assuming that these recently formed stars will be passively evolving since then, [J1000+0234  ]{}could turn into a ’red-and-dead’ elliptical – like those that are found at z=2 when about 2Gyr have passed since z=4.5. Evidence for an Ongoing Major Merger ==================================== The molecular gas emission arises from a region that is highly obscured at the observed optical wavelengths [see Fig. 1 of @cap08 Fig. \[fig:comp\] and \[fig:velo\]]. However, it is coincident with the position of the rest-frame optical and radio emission (Fig. \[fig:comp\]). The HST ACS image of [J1000+0234  ]{}[@cap08] was adaptively smoothed [@sco00] to enhance the structure present in the blue continuum covered by the F814W filter (Fig. \[fig:velo\]). The molecular gas appears to lie next to bright blue continuum emission that is at its eastern side. Ly$\alpha$ emission has also been associated with this bright rim in blue continuum [@cap08]. The motion of the gas is roughly along a region of higher extinction (compared to the blue western component) running roughly from north to south and shows a larger spatial extent than the blue continuum emission. Note the object west of the gas emission is a foreground galaxy at z=1.41. As already mentioned by [@cap08] this geometry is very reminiscent of an ongoing merger. As the redshift of the molecular gas is very close to the redshift derived from the Ly$\alpha$ line (z=4.542 vs. z=4.547), it is very likely that both components belong to the same (highly disturbed) system. Both the large FWZI of the CO(4-3) line and the appearance of [J1000+0234  ]{}in Ly$\alpha$ make an ongoing merger scenario very likely. We derive a conservative estimate of the merging time for [J1000+0234  ]{}of significantly less than a billion years using the prescription of [@kit08] for close galaxy pairs with velocities of $v < 3000\,{\>{\rm km}\,{\rm s}^{-1}}$. This time scale is consistent with having a relaxed system at a redshift of z=2. The expected number of z=4-5 SMGs from the model of [@bau05] yields of the order of 10-20 such sources within the 2deg$^2$ COSMOS field. As [J1000+0234  ]{}was originally identified as a Lyman Break Galaxy via its V-band drop [@cap08; @car08], this suggests that UV emission can be detectable from ongoing major mergers at these redshifts. We have identified 4 additional V-band dropout LBGs with 4$\sigma$ radio detections [@car08] that could also be SMGs at this redshift. An inspection of their optical appearance in the HST ACS data shows that two of them also have a disturbed geometry. This suggests that there could be more objects sharing properties of both the SMG and LBG population at z$>$4 and [J1000+0234  ]{}might therefore provide an important link between the SMG and LBG populations at high redshifts. The case of [J1000+0234  ]{}shows that (at least some) z$>$4 SMGs can have very similar properties to their well-studied low redshift (z$\sim$2) counterparts, that major mergers with very large SFRs are likely present at z=4.5 and that these systems might be the precursors of elliptical galaxies found at the red sequence at a redshift of z=2. ES would like to thank J.M. Winter and R. Neri for their great help during the visit of IRAM, Grenoble. ES thanks A. Weiß and E.F. Bell for helpful discussions. ES acknowledges the hospitality of the Aspen Center for Physics where the manuscript was prepared. CC thanks the Max-Planck-Gesellschaft and the Humboldt-Stiftung for support through the Max-Planck-Forschungspreis. The National Radio Astronomy Observatory (NRAO) is operated by Associated Universities, Inc., under cooperative agreement with the National Science Foundation. [*Facilities:*]{} Barger, A. J., Cowie, L. L., Sanders, D. B., Fulton, E., Taniguchi, Y., Sato, Y., Kawara, K., & Okuda, H. 1998, , 394, 248 Baugh, C. M., Lacey, C. G., Frenk, C. S., Granato, G. L., Silva, L., Bressan, A., Benson, A. J., & Cole, S. 2005, , 356, 1191 Bertoldi, F., et al. 2000, , 360, 92 Blain, A. W., Smail, I., Ivison, R. J., Kneib, J.-P., & Frayer, D. T. 2002, , 369, 111 Carilli, C.L., et al.  2008, , in press Capak, P., et al. 2008, , 681, L53 Chapman, S. C., Blain, A. W., Smail, I., & Ivison, R. J.  2005, , 622, 772 Downes, D., & Solomon, P. M. 1998, , 507, 615 Genzel, R., Baker, A. J., Tacconi, L. J., Lutz, D., Cox, P., Guilloteau, S., & Omont, A. 2003, , 584, 633 Greve, T. R., et al. 2005, , 359, 1165 Guilloteau, S., & Lucas, R. 2000, Imaging at Radio through Submillimeter Wavelengths, 217, 299 Hughes, D. H., et al. 1998, , 394, 241 Kitzbichler, M.G., & White, S.D.M. 2008, , subm., astro-ph/0804.1965 Kov[á]{}cs, A., Chapman, S. C., Dowell, C. D., Blain, A. W., Ivison, R. J., Smail, I., & Phillips, T. G. 2006, , 650, 592 Neri, R., et al. 2003, , 597, L113 Pope, A., et al. 2006, , 370, 1185 Schinnerer, E., et al. 2007, , 172, 46 Scoville, N. Z., et al. 2000, , 119, 991 Smail, I., Ivison, R. J., & Blain, A. W. 1997, , 490, L5 Solomon, P. M., & Vanden Bout, P. A. 2005, , 43, 677 Tacconi, L. J., et al. 2006, , 640, 228 Tacconi, L. J., et al. 2008, , 680, 246 Wei[ß]{}, A., Downes, D., Walter, F., & Henkel, C. 2005, , 440, L45 Younger, J. D., et al. 2007, , 671, 1531 [lrl]{} $\rm S_{250\,GHz}$ \[mJy\] & 3.4$\pm$0.67 & MAMBO\ $\rm S_{83\,GHz}$ \[mJy\] & 0.2$\pm$0.09 & off-line channels ($\sim$2$\sigma$, PdBI)\ $\rm S_{42\,GHz}$ \[mJy\] & $\le$0.15 & off-line channel ($\le$3$\sigma$, VLA)\ $\rm z_{CO(4-3)}$ & 4.542 & center of CO(4-3) line\ $\rm FWZI_{CO(4-3)}$ \[${\>{\rm km}\,{\rm s}^{-1}}$\] & $\sim$ 1000 & CO(4-3) line\ $\rm S_{CO(4-3)}$ \[Jy${\>{\rm km}\,{\rm s}^{-1}}$\] & 0.66$\pm$0.12 & 5.5$\sigma$ detection (PdBI)\ $\rm S_{CO(2-1)}$ \[Jy${\>{\rm km}\,{\rm s}^{-1}}$\] & 0.057$\pm$0.017$^{\dagger}$ & 3.3$\sigma$ detection$^*$ (VLA)\ $\rm R.A._{CO(4-3)}$ (2000)& 10:00:54.484 & line peak\ $\rm Dec._{CO(4-3)}$ (2000) & +02:34:35.73 & line peak\ $\rm L'_{CO} [L_{\odot}]$ & 3.2$\times$10$^{10}$ & from CO(4-3)\ $\rm M_{gas} [M_{\odot}]$ & 2.6$\times$10$^{10}$ & from CO(4-3)\ $\rm M_{dyn}\times sin^2{\it i} [M_{\odot}]$ & 1.3$\times$10$^{11}$ & using FWZI & r=0.5”\ [^1]: The scaling of the SEDs presented in Fig. 4 of [@cap08] is overestimated by a factor of 10$^9$ (P. Capak, priv. comm.).

ArXiv

--- abstract: 'In this work we undertake the problem of a transition metal impurity in an oxide. We present an [*ab initio*]{} study of the relaxations introduced in TiO$_2$ when a Cd impurity replaces substitutionally a Ti atom. Using the Full-Potential Linearized-Augmented-Plane-Wave method we obtain relaxed structures for different charge states of the impurity and computed the electric-field gradients (EFGs) at the Cd site. We find that EFGs, and also relaxations, are dependent on the charge state of the impurity. This dependence is very remarkable in the case of the EFG and is explained analyzing the electronic structure of the studied system. We predict fairly anisotropic relaxations for the nearest oxygen neighbors of the Cd impurity. The experimental confirmation of this prediction and a brief report of these calculations have recently been presented \[P.R.L. 89, 55503 (2002)\]. Our results for relaxations and EFGs are in clear contradiction with previous studies of this system that assumed isotropic relaxations and point out that no simple model is viable to describe relaxations and the EFG at Cd in TiO$_2$ even approximately.' author: - 'Leonardo A. Errico' - 'Gabriel Fabricius$^\dag$' - Mario Rentería title: '**The problem of a metal impurity in an oxide: [*ab initio*]{} study of electronic and structural properties of Cd in Rutile TiO$_2$**' --- \[sec:1\]INTRODUCTION ===================== The problem of metal impurities in oxides is a challenge in solid-state physics both from a fundamental and applied point of view. Description of impurity levels in oxide semiconductors has a great technological interest, but a complete theoretical description of the problem is very difficult. In effect, impurities introduce atomic relaxations in the host and modify the electronic structure of the system, being the interplay between these two effects one of the principal difficulties in the theoretical approach. Some experimental techniques used in material science introduce probe-atoms into the systems to be studied,[@techniques1; @techniques2] giving valuable physical information [*seen*]{} by the probe-atom, usually an impurity in the system. In particular, TDPAC spectroscopy [@TDPAC] is a hyperfine technique that enables a high resolution determination of the EFG tensor at impurity sites. Due to the r$^{-3}$ dependence of the EFG operator from the charge sources, the EFG is mostly originated in the electronic charge density close to the impurity nucleus, which in turn reflects the probe chemistry with its neighborhood. Thus, it would be important that the theoretical approach to the metal-impurity problem would be able to predict the EFG tensor with the required accuracy. Since the EFG is very sensitive to the anisotropic charge distribution close to the nucleus, for its accurate calculation the entire electronic configuration of the host, perturbed by the presence of the impurity, has to be determined. This kind of calculations should be performed at the [*ab initio*]{} level and can be done in the frame of the Density Functional Theory (DFT). A well proven method to solve the all-electron DFT equations in solids is the Full-Potential Linearized-Augmented-Plane-Wave (FLAPW) method.[@flapw] This method was able to predict with high accuracy and precision the EFG at undisturbed lattice sites in a large variety of pure systems.[@blaha; @efg] However, on systems with impurities, very few calculations have been performed and the state-of-the-art is far from being routinely in this field. [@state-of-the-art; @state2; @state3; @state4; @state5; @state6; @state7; @state8; @state9] The first fully self-consistent [*ab initio*]{} determination of the EFG tensor at an impurity site (Cd) in an oxide (TiO$_2$) has recently been reported [@prl]. In that work we performed a FLAPW calculation of the relaxations introduced by the impurity and studied their interplay with the electronic structure of the system, predicting highly anisotropic relaxations of the nearest neighbors of the impurity and a drastic change in the orientation of the principal component of the EFG tensor. This prediction was confirmed in the same work by a key PAC experiment. Due to the complexity of the physical situation we want to describe, even in the framework of [*ab initio*]{} calculations some variables have to be determined through comparison with experiments. For example, we performed the calculations for two charge states of the impurity and one of them could be discarded because it gave a value for the EFG asymmetry parameter, $\eta$, incompatible with experience. The central purpose of this work is to present and discuss the electronic structure of the TiO[$_2$]{}:Cd impurity system with more detail that we could do in Ref. . Here, we study the relationship between the calculated EFG at the impurity site - the key experimental quantity - and the interplaying physical quantities: the structural relaxations and the character and filling of the impurity state. We also discuss in detail the precision of our calculations studying its dependence on the different parameters that control the precision (impurity dilution, base dimension, average in k-space, etc.) and its dependence on the approximation used for the exchange-correlation potential. Another important question to be addressed concerning all experimental techniques that introduce impurity-tracers in solids is at what extent simple models usually used to predict the measured quantities are acceptable in view of the modifications introduced by the presence of the impurity. Concerning the EFG, simple approximations like the point-charge model [@techniques2] with antishielding factors[@stern; @lieb] and isotropic relaxations of nearest-neighbor atoms [@lieb; @akai] are commonly used. In this work we give arguments against the applicability of these simple models for the studied system. This paper is organized as follows. In section II we describe the method of calculation used in the present work and include some preliminary calculations for an unrelaxed cell in order to discuss methodologically the point of the charge state of the impurity. In section III.(A-C) we present and discuss the results of performing relaxations of the Cd nearest-neighbors (NN) in a 72-atoms super-cell for two different states of charge of the impurity, while in section III.D we study the accuracy of these results performing several additional calculations. Even if section III.D could be found rather [*technical*]{} by people not specially interested in first-principles calculations, we think it will be useful for the [*ab initio*]{} community interested in the problem of impurities in solids. In section III.E we compare our results with previous studies of this system. Finally, in section IV we present our conclusions. \[sec:2\]METHOD OF CALCULATION ============================== \[sec:2.A\]A general scheme --------------------------- Our aim is to calculate from first principles the structural relaxations produced in the TiO$_2$ lattice when a Cd impurity replaces a Ti atom and the electronic structure of the resulting system. In particular we are interested in the EFG tensor at the Cd site. The general scheme we adopted to deal with this problem was the following: we considered a super-cell (SC), containing a single Cd impurity, repeated periodically and performed first-principles calculations We studied the relaxation introduced by the impurity computing the forces on the Cd neighbors and moving them until the forces vanished. The calculations were performed with the [wien97]{} implementation, developed by Blaha [*et al.*]{},[@wien97] of the Full-Potential Linearized-Augmented-Plane-Wave (FLAPW) method and we worked in the LDA approximation using the Perdew and Wang parametrization for the exchange-correlation potential.[@LDA] In this method the unit cell is divided into non-overlapping spheres with radii R$_i$ and an interstitial region. The atomic spheres radii used for Cd, Ti and O were 1.11, 0.95 and 0.85 [Å]{}, respectively. We took for the parameter RK$_{MAX}$, which controls the size of the basis-set in these calculations, the value of 6 that gives basis-set consisting in 4500 LAPW functions for the SC described in section \[sec:2.B\]. We also introduced local orbitals (LO) to include Ti-3$s$ and 3$p$, O-2$s$ and Cd-4$p$ orbitals. Integration in reciprocal space was performed using the tetrahedron method taking an adequate number of [*k*]{}-points in the first Brillouin zone. Once self-consistency of the potential was achieved, quantum-mechanical-derived forces were obtained according to Yu [*et al.*]{}[@forces], the ions were displaced according to a Newton damped scheme [@newton] and new positions for Cd neighbors were obtained. The procedure is repeated until the forces on the ions are below a tolerance value taken as . At the relaxed structure, the $V_{ij}$ elements of the EFG tensor are obtained directly from the $V_{2M}$ components of the lattice harmonic expansion of the self-consistent potential (a more detailed description of the formalism used to compute the EFG tensor may be found in Ref. ). There is still an important point to take into account concerning the calculation of the electronic structure of an impurity system: the charge state of the impurity. We let this point to the end of this section as it will be easily handled after the analysis of some preliminary calculations. \[sec:2.B\] Cell and super-cell ------------------------------- The rutile (TiO$_2$) structure is tetragonal ($a=b=$ 4.5845$_{1}$ [Å]{}, $c=$ 2.9533$_{1}$ [Å]{}). The unit cell (shown in Fig. \[cell\]) contains 2 metal atoms (Ti) at positions 2$a$ (0, 0, 0) and (1/2, 1/2, 1/2) and 4 anions (O) at positions 4$f$ $\pm$(u, u, 0; u+1/2, 1/2-u, 1/2), with u = 0.30493$_7$.[@tio2] The super-cell considered in the present work consists of twelve unit cells of TiO$_2$ where one Ti atom has been replaced by a Cd atom. The resulting 72-atom SC (called 72A-SC in the future) has dimensions $a'=$2$a$, $b'=$2$b$, $c'=$3$c$ and is also tetragonal with $c'/a'=$0.97 giving an almost cubic lattice. This SC keeps Cd atoms as far as possible from each other for the given cell volume. For checking purposes we have also considered SCs containing eight and sixteen unit cells. We assume that relaxations to be performed preserve point group symmetry of the cell in this initial configuration. Symmetry restricts O1 and O2 displacements to $yz$ plane and $x$ direction respectively. In order to check this assumption and the stability of the obtained solution, at the end of the relaxation process we performed new calculations with O1 and O2 atoms displaced from their symmetry positions and verified that these solutions are a minimum for each system studied. The self-consistent calculations were performed taking 8 $k$-points in the irreducible Brillouin zone for the metallic system (situation $(i)$), and 2 $k$-points for the non-metallic ones (situation $(ii)$). In order to plot the DOS we calculate eigenvalues at a denser mesh of 36 $k$-points. \[sec:2.C\] Preliminary study: Charge state of the impurity ----------------------------------------------------------- In order to discuss the problem of the charge state of the impurity we have done some preliminary calculations. We have calculated the self-consistent electronic structure of the 72A-SC with all atoms in their initial unrelaxed positions. We want to analyze the changes in the electronic structure of the system caused by the presence of the impurity, neglecting for the moment the problem of structural relaxations. The density of states of this system is compared with the one of pure TiO$_2$ in Fig. \[DOS1\]. Pure Ti$^{+4}$O$_2 ^{-2}$ is a wide band-gap semiconductor with the O-$p$ band filled and the Ti-$d$ band empty. Since Cd valence is 2+, when a Cd atom replaces a Ti atom in the SC the resulting system is metallic because of the lack of two electrons necessary to fill up the O-$p$ band. The question that arises at this point is if the real system we want to describe provides the lacking two electrons (via an oxygen vacancy, for example) or not, and if this point is relevant for the present calculation. We will show first that this point is absolutely relevant for the present calculation. Comparison of Fig. \[DOS1\](a) and (b) shows that the presence of the Cd impurity in the SC introduces the appearance of Cd-$d$ levels and impurity states at the top and the bottom of the valence band in the corresponding DOS. The ones at the top of the valence band can be better seen looking at the band structure in this energy range shown in Fig. \[bands\_c12n-sr\]. The two bands that are immediately above and below the Fermi level when crossing A-point correspond to impurity anti-bonding states that are spatially located at Cd and at their O1 and O2 nearest-neighbor atoms. In particular, the wave function of the impurity state that remains almost completely unoccupied has character [Cd-$d_{yz}$]{}, [O1-$p_y$]{}, and [O1-$p_z$]{}. Then, providing two electrons to the system implies a drastic change in the symmetry of the electronic charge distribution in the neighborhood of the impurity. Therefore, different nearest-neighbors relaxations and also different EFG could be expected for different charge states of the impurity. In the present work, we present calculations for two charge states of the impurity system, corresponding to two different physical situations: $(i)$ Cd$^0$ (neutral impurity state), corresponding, e.g., to an extremely pure crystal at low temperatures, and $(ii)$ Cd$^-$ (charged acceptor state), the system provides two electrons via an oxygen vacancy, donor defects, etc. To study case $(i)$ we used the 72A-SC already described, to study case $(ii)$ we also used the 72A-SC but performed self-consistent calculations adding two electrons to the SC that we compensate with an homogeneous positive background in order to have a neutral cell to compute total energy and forces (this procedure is implemented since version [wien97.9]{} of the FLAPW package). We have also simulated situation $(ii)$ with an alternative procedure: we replaced in the 72A-SC the two most distant oxygen atoms by two fluorine atoms. In this way we provide two electrons to fill up the O-$p$ band without introducing any artificial background. We expected that the difference between flourine and oxygen potentials should modify only slightly the results. In summary, we performed self-consistent FLAPW calculations for the following systems: $ (i)$ 0.3cm = Cd$^0$, 1.8cm = (TiO$_2$)$_{23}$CdO$_2$\ $ (ii)$ Cd$^-$(2e), (TiO$_2$)$_{23}$CdO$_2 + 2e^-$\ Cd$^-$(fluorine), (TiO$_2$)$_{23}$CdF$_2$ RESULTS AND DISCUSSION ====================== \[sec:3.A\] Structural relaxations ---------------------------------- Let us first consider the relaxation of only the 6 nearest oxygen neighbors of the Cd impurity (O1 and O2 in Fig \[cell\].) In Table \[tab:distances\] we compared the results of the relaxation of these oxygen atoms for the different systems studied. We see that for both charge-states of the impurity the relaxations are quite anisotropic, with the Cd-O1 distance larger than Cd-O2 distance, opposite to the initial unrelaxed structure. This result is opposite to what other authors have assumed in previous studies of this system [@akai; @lieb] and confirms the tendency predicted in our previous calculation with a much smaller SC.[@our_nqi] As it can be seen in Table \[tab:distances\] the difference in the charge state of the impurity affects essentially the relaxation of O1 atoms that present a slightly larger relaxation for the charged impurity. Relaxation for the two ways of simulating the charged state of the impurity (case Cd$^-$) are very similar indicating that both approaches are well suited to deal with this problem. $d$(Cd-O1) $d$(Cd-O2) $\theta _1$ ------------------ -- ------------ ------------ ------------- TiO$_2$ 1.944 1.977 40.47 Cd$^0$ 2.153 2.108 39.59 Cd$^-$(2e) 2.185 2.111 39.55 Cd$^-$(fluorine) 2.191 2.121 39.73 : \[tab:distances\]Final coordinates of the Cd nearest oxygen neighbors for the different calculations performed with the 72A-SC compared with the ones of pure TiO$_2$. $d$(Cd-O1) and $d$(Cd-O2) are the distances (in Å) from Cd to O1 and O2 atoms, respectively. $\theta _1$ is the angle (in degrees) between $z$ axis and $\vec r_{O1} - \vec r_{Cd}$. Anisotropy in the relaxations of the nearest oxygen neighbors of the Cd impurity can be understood by inspection of Fig. \[stretching\]. Stretching of Cd-O2 bond implies a considerable shortening in Ti-O2 bonds. However, stretching of Cd-O1 bond doesn’t affect so much Ti-O1 bonds since the structure is more open in this direction. So, at the end of the relaxation, Cd-O1 bond stretches almost twice than Cd-O2 bond. \[sec:3.B\] Electronic structure -------------------------------- In Fig. \[bands\] we show the bands for the 72A-SC for the two states of charge considered for the impurity. Comparison of Fig. \[bands\](a) with Fig. \[bands\_c12n-sr\] shows that as a consequence of relaxation the outermost impurity level falls down in energy and goes into the O-$p$ band becoming half-occupied. Relaxation increases the Cd-O1 distance and this produces a softening of the Cd-$d -$O1-$p$ interaction and the fall down in energy of the anti-bonding impurity states. Figure \[bands\] shows that the band structure of the neutral and charged relaxed structures are very similar, but the outermost impurity state is a slightly raised in energy when it is completely filled (case $(ii)$) as a consequence of the larger Coulomb repulsion. Comparison of Fig. \[DOS2\](a) and Fig. \[DOS2\](c) shows the little fall down in energy mentioned for the anti-bonding impurity states an also a little rise for the bonding ones at the bottom of the O-$p$ band. A shift upwards of about 1.5 eV of Cd-$d$ levels from the un-relaxed  to the relaxed one is also present. To look at the orbital composition of the impurity states we plot in Fig. \[PDOSNyC\](a) and (b) the PDOS for Cd-$d$, O1-$p$ and O2-$p$ for the two charge states of the impurity. The impurity state near the top of the valence band has Cd-$d_{yz}$ , O1-$p_y$ and O1-$p_z$ character and, as we mention looking at the bands, it is shifted upwards in energy when it is completely filled (case Cd$^-$). In Fig. \[PDOSNyC\](a)(Cd atom) and (b)(Cd atom) it can also be seen the presence of other impurity state with components Cd-$d_{x^2-y^2}$ and Cd-$d_{3z^2-r^2}$ which is completely filled in both cases. This impurity state is located mainly at Cd-$d_{x^2-y^2}$ and O2-$p_x$ but involves also contributions from O1-$p_z$, Cd-$d_{3z^2-r^2}$ and O3-$p_x$ (O3 is the NN of O2 atom in the $x$ direction). From the present calculations the outermost impurity state has an occupation of around 1.3 e for the 72A-SC(Cd$^0$) and 2 e for the 72A-SC(Cd$^-$). We identify this 0.7 additional electrons in this state as the driving force that produce the slightly larger relaxation for O1 atoms in the charged cell with respect to the neutral cell. We want to mention that the fact that the impurity level falls at the Fermi energy in the Cd$^0$ case is not fortuitous. Due to Coulomb repulsion, the impurity level would fall below the Fermi energy if it was empty an above if it was filled being the only self-consistent solution to be half-occupied at exactly the Fermi energy. The occupation of the impurity level in the 72A-SC(Cd$^0$) is therefore a constant number and fairly independent of small fluctuations of charge in the cell. \[sec:3.C\] Electric-Field Gradients ------------------------------------ [lcccccc]{} & $V_{XX}$ & $V_{YY}$ & $V_{ZZ}$ & $V_{33}$ &$V_{33}$-direction & $\eta$\ & -7.16 & +6.82 & +0.34 & -7.16 & **X** & 0.91\ & -2.87 & +4.55 & -1.68 & +4.55 & **Y** & 0.26\ & -2.46 & +4.10 & -1.63 & +4.10 & **Y** & 0.20\ Exp.[@lieb] & & & & 5.23(5) & ... & 0.18(1)\ Exp.[@exp2] & & & & 5.34(1) & ... & 0.18(1)\ Exp. (singlecrystal)[@prl]& & & & 5.34(1) & **X** or **Y** & 0.18(1)\ Calc.[@akai]& +1.54 & +3.56 & -5.09 & -5.09 & **Z** &0.39\ Exp.(pure TiO$_2$)[@nmr_tio2] & & & & 2.2(1) & **Z** &0.19(1)\ In Table \[tab:efgs\] we show the results for the V$_{ii}$ principal components of the EFG tensor for the three systems studied. The resulting EFGs for the two approaches used to simulate the charged impurity are very similar, the difference of 0.4 x  in components $V_{XX}$ and $V_{YY}$ is within one could expect for the small difference found in the oxygen positions, since EFG is very sensitive to small structural changes. These results agree very well (in magnitude, symmetry and orientation, see Table \[tab:efgs\]) with the experimental results obtained for the EFG at Cd impurities substitutionally located at cationic sites in rutile TiO$_2$.[@prl] The difference between the EFGs obtained for the charged and neutral cells is very remarkable: the sign, direction and absolute value of the largest $V_{ii}$ component ($V_{33}$) are different in both situations, and also the value of the asymmetry parameter $\eta$. The high $\eta$ value obtained for shows that the electron availability present in the sample leads the impurity to be in a charged state. In order to investigate the origin of the difference in the EFG for the two charge states of the impurity we concentrate in the valence contribution to the EFG which originates in the asymmetry of the valence charge distribution inside the muffin-tin sphere. The valence contribution is usually dominant in FLAPW calculations and can be split in the different orbital symmetries.[@efg] In Table \[tab:efgs comp\] we show the total valence contribution to $V_{ii}$ and its components arising from $p$ and $d$ orbital symmetries. We see that the largest differences correspond to $d$-components of $V_{ii}$. This difference is originated in the filling of the impurity state at the Fermi level that has an important component of Cd-$d_{yz}$ symmetry as can be seen in Figs. \[PDOSNyC\](a) and (b)(Cd atom). A simple analysis in terms of partial charges [@dnd] shows that the effect of adding $\delta n$ electrons to an orbital $d_{yz}$ is to produce a change in $V_{ii}$ components given by: $\delta V_{XX}$= $I_d \delta n$, $\delta V_{YY}$= - $I_d \delta n /2$, $\delta V_{ZZ}$= - $I_d \delta n /2$, where $I_d$ is proportional to $\langle 1/r^3 \rangle$ for $d$ orbitals inside the muffin tin sphere. Integration of unoccupied [Cd-$d_{yz}$]{} PDOS from Fig. \[PDOSNyC\](a)(Cd atom) gives $\delta n$= 0.074. Inspection of Table \[tab:efgs comp\] shows that the changes $\delta V_{ii}$ are quite well described by this estimation giving for $I_d$ a value around 83 $\times$ 10$^{21}$V/m$^2$. Another interesting point is the presence of $d$ contributions to $V_{ii}$ in the 72A-SC\[Cd$^-$(2e)\] although the Cd-$d$ band is completely filled (see Table \[tab:efgs comp\]). This curious point may be explained because radial $d$ wave functions are energy dependent, so, electrons in anti-bonding impurity states (at the Fermi level) contribute to $V_{ii}$ with a different $\langle 1/r^3 \rangle$ factor than the ones of the same symmetry in Cd-$d$ levels (between 6 and 7 eV below the Fermi level). We have found an increment of 30$\%$ in $\langle 1/r^3 \rangle$ when going from an energy of -7 eV to the Fermi level. This dependence of $\langle 1/r^3 \rangle$ with energy was also mentioned and verified by Blaha [*et al.*]{} in Ref.  for Cu in Cu$_2$O. Finally, it has to be mentioned that the difference in $p$ contributions to $V_{ii}$ between the 72A-SC(Cd$^0$) and the 72A-SC\[Cd$^-$(2e)\] is not negligible and is caused by the different positions of the Cd nearest oxygen neighbors in each case. ---------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- $p$ $d$ TOT $V_{XX}$ $V_{YY}$ $V_{ZZ}$ $V_{XX}$ $V_{YY}$ $V_{ZZ}$ $V_{XX}$ $V_{YY}$ $V_{ZZ}$ 72A-SC(Cd$^0$) -1.62 +3.17 -1.55 -6.44 +4.29 +2.15 -7.97 +7.31 +0.66 72A-SC\[Cd$^-$(2e)\] -2.62 +3.55 -0.93 -0.33 +1.18 -0.85 -2.86 +4.58 -1.72 diff. -1.00 +0.38 +0.62 +6.11 -3.11 -3.00 +5.13 -2.73 -2.38 ---------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- \[sec:3.D\] Accuracy of the present study: Further relaxations and other tests ------------------------------------------------------------------------------ The main sources of error of the present study in order to compare with experiment are the size of the SC considered, the size of the basis, the LDA approximation used for the exchange-correlation potential and the relaxation process that has been restricted to the six nearest oxygen neighbors of the Cd atom. To check the accuracy of the present study we have performed several additional calculations. ### Charged impurity: Cd$^-$ We first focus on the case of the charged impurity where we have performed the most intensive tests. #### Charge of the cell The differences found between the results for the 72A-SC\[Cd$^-$(2e)\] and 72A-SC\[Cd$^-$(fluorine)\] (Tables \[tab:distances\] and \[tab:efgs\]) could be taken as a measure of the error performed in simulating the charged impurity state. But in fact, we expect the procedure using the 72A-SC\[Cd$^-$(2e)\] to give better results that the one using the 72A-SC\[Cd$^-$(fluorine)\] since in the former the global properties of the system are only smoothly affected (background density is 0.003 e/Å$^3$ for this SC). Moreover, when considering relaxations of atoms more distant from Cd than its NN, the presence of fluorine atoms will spuriously influence the results, so, in what follows we refer to the . #### Size effects In order to check how appropriate are the dimensions of the 72A-SC ($2a \times 2b \times 3c$) used in the present work we have performed self-consistent electronic structure calculations for a 48A-SC ($2a \times 2b \times 2c$) and a 96A-SC ($2a \times 2b \times 4c$). In these calculations we put the nearest oxygen neighbors of Cd at relative positions from Cd that correspond to equilibrium in the 72A-SC. The size of the forces that oxygen atoms O1 and O2 experiment in the 48A-SC and 96A-SC is a measure of the convergence of the present calculation. We obtained that in the 96A-SC the forces on O1 and O2 atoms point outwards (respect to Cd) and are around . Forces of this size produce, during the relaxation process in the 72A-SC, changes in distances smaller than and changes in EFG of about , so, similar changes would be expected if relaxation in the 96A-SC was performed. In the case of the 48A-SC forces point inwards and are of the same magnitude for O2 atoms but they are about in O1 atoms. We see that there is a size effect that makes relaxations to be larger for larger SCs, but for the 72A-SC the effect is quite small and no significant variations should be expected if relaxations in bigger SCs were considered. #### Basis size We compute the self-consistent electronic structure of the 72A-SC\[Cd$^-$(2e)\] for the relaxed positions of Table \[tab:distances\] increasing the size of the basis to 7100 LAPW functions (RK$_{MAX}$=7). Forces on O1 and O2 atoms are below the tolerance value of indicating that the result of the relaxation performed is unaltered for a substantial increase of the basis size. In fact, the forces on all the other atoms in the cell have the same values than using RK$_{MAX}$=6 (within the tolerance) except for the Ti neighbors of Cd in the $z$ direction where the forces differ in . Changes obtained in $V_{ii}$ components of the EFG tensor are smaller than 0.05 $\times$ 10$^{21}$V/m$^2$. We have also considered the inclusion of a Cd-4$d$ LO to improve linearization. When a Cd-4$d$ LO is introduced with energy at the Fermi level in order to improve the description of the impurity states no influence in the forces is detected. The change in $V_{ii}$ components of the EFG is smaller than 0.1 $\times$ . #### Exchange-correlation potential We performed electronic self-consistent calculations and relaxation of Cd nearest oxygen neighbors for the system 72A-SC\[Cd$^-$(2e)\] using the generalized gradient approximation (GGA) [@gga] instead of LDA. With the use of this parametrization for the exchange-correlation potential we obtained 2.18 [Å]{} and 2.10 [Å]{} for Cd-O1 and Cd-O2 distances, respectively (i.e. only a small change of 0.01 [Å]{} in Cd-O2 distance). For the EFG we obtained $V_{33}$=$V_{YY}$=+4.94 $\times$  and $\eta$=0.37. #### Further relaxations In order to study the effect of relaxing the coordinates of atoms beyond the nearest neighbors of the Cd atom we have performed the following two additional relaxations: (a) we allow to relax the coordinates of all atoms within a cutoff radius $R_C$=4 [Å]{} centered at Cd (this involves 24 atoms) and (b) idem for $R_C$=4.6 [Å]{} (this involves 42 atoms). For radius larger than 4.6 [Å]{}, atoms to be relaxed would be nearer to the images of Cd atom from neighboring cells than to the Cd itself, so we consider this radius as a limit for the present SC. In Table \[tab:further relax\] we compare the results obtained for both relaxations with the ones corresponding to the NN relaxation. We observe that there is not any qualitative change in the results already discuss for the NN relaxation. There are, however, some small variations in the values predicted for the EFG. The differences between results from relaxations (a) and (b) (that are as larger as the differences between relaxations (a) and NN) are in part caused because in (b) the relaxation of O3 atom is allowed. Atom O3 is the NN of O2 atom in the $X$ direction and its relaxation of about 0.04 [Å]{} allows a further relaxation of about 0.012 [Å]{} of the O2 atom. The fact that atom O3 is the one that experiments the largest relaxation (not considering O1 and O2) shows that directional bonding plays an important role in this structure. R$_C$ N$_A$ d(Cd-O1) d(Cd-O2) $V_{XX}$ $V_{YY}$ $V_{ZZ}$ $\eta$ ----- ------- ------- ---------- ---------- ---------- ---------- ---------- -------- NN 2.5 6 2.185 2.111 -2.87 +4.55 -1.68 0.26 (a) 4.0 24 2.176 2.104 -3.25 +4.99 -1.74 0.30 (b) 4.6 42 2.187 2.116 -3.17 +4.86 -1.69 0.30 : \[tab:further relax\]Results of the different relaxations performed for the 72A-SC\[Cd$^-$(2e)\]. In each case all the coordinates of the N$_A$ atoms within a radius $R_C$ (in Å) are relaxed until forces on them are below 0.025 eV/Å. All units as in Table I and II. In summary, we observed that none of the factors considered influence qualitatively the results, but their effect is neither negligible. We therefore confirm our prediction about sign (positive) and direction ($Y$) of $V_{33}$, but it is not possible to perform an exact prediction about its magnitude. Ours checks shows that $d$(Cd-O1) and $d$(Cd-O2) are converged within 0.01 [Å]{} and [V$_{33}$]{} and $\eta$ within 0.5 $\times$ and 0.1, respectively. Taking the values from the last row of Table \[tab:further relax\] as our predicted values for $V_{ii}$ we obtain discrepancies with experiment: [@exp2; @prl] of 0.37 $\times$  for $|V_{33}|$ and 0.12 for $\eta$ that could be attributed to precision errors of the present calculations. ### Neutral impurity: Cd$^0$ Let us briefly discuss the case of the 72A-SC(Cd$^0$). Due to the factor four that exists because of $k$-sampling, calculations are much more time consuming in this case. Hence, we have checked the size of the SC performing calculations only on a 48A-SC, and also checked the size of the basis repeating the calculation for the 48A-SC with RK$_{MAX}$=7. Comparison of the values obtained for EFG and forces let us conclude that errors in the 72A-SC(Cd$^0$) are expected to be of the same magnitude than the ones obtained for the 72A-SC\[Cd$^-$(2e)\]. Only the effect of adding a Cd-4$d$ LO and the use of GGA instead LDA produce a larger variation of the EFG in this case than in the charged cell, and is understandable because of the larger $d$ contribution to $V_{ii}$ in the case of the neutral state of the impurity. Calculations in the 72A-SC(Cd$^0$) with this parametrization for the exchange-correlation potential and introducing Cd-4$d$ LO predict values of $V_{33}$=$V_{XX}$=-7.90$\times$  and $\eta$=0.80. In addition, we found a force of inwards at O1 atoms, so a small refinement in O1 position would be expected. About further relaxations, we have only performed relaxation (a) ($R_C$=4 Å) for the 72A-SC(Cd$^0$) and also obtain variations of the same order than in 72A-SC\[Cd$^-$(2e)\], in particular we obtained $V_{33}$=$V_{XX}$=-6.8$\times$ and $\eta$=0.97. Then, the same conclusions about accuracy than in the 72A-SC\[Cd$^-$(2e)\] hold for the 72A-SC(Cd$^0$) but note that, due to the high value of $\eta$ in this case, the sign and direction of $eq$ could change because of precision. \[sec:3.E\] Comparison with other calculations ---------------------------------------------- The simplest and most widely used approximation for the calculation of the EFG at a probe-atom is the point-charge model (PCM).[@techniques2] In this approximation the EFG tensor at the probe site is $(1-\gamma _{\infty}) V_{ij}^{latt.}$, where $V_{ij}^{latt.}$ is the EFG tensor produced by valence nominal charges located at the ion positions in the lattice, and $\gamma _{\infty}$ is the Sternheimer antishielding factor[@stern] that depends only on the probe-atom. In this way, this model assumes that the symmetry and orientation of the EFG tensor at impurity sites are unaltered by its presence. PCM gives for TiO$_2$(Cd): $V_{33}$=$V_{XX}$=-2.27 $\times$  and $\eta$=0.40 when a value of -29.27 is used for $\gamma _{\infty}$.[@feiock] In Ref.  the authors assume that relaxation is responsible for the disagreement of PCM predictions with experiment and speculate that an isotropic relaxation of 0.04 Å outwards of all the Cd nearest oxygen neighbors would produce the desired result for $|V_{33}|$ and $\eta$. Our results indicate that relaxations are not isotropic and they are so much larger than this, but even if the relaxed coordinates from our calculation (see Table I) were used, the PCM would fail in the description of the EFG giving $V_{33}$=$V_{XX}$= -8.20 $\times$  in clear contradiction with $V_{33}$=$V_{YY}$= +4.55 $\times$  that we obtain from the self-consistent FLAPW calculation. The disagreement of PCM with the FLAPW prediction could not either be attributed in this case to a change in the value of $\gamma _{\infty}$ since the sign and directions of both predictions are different. Thus, it is clear that the problem of EFG at cationic sites in TiO$_2$ is too complicated to be described even approximately by simple PCM calculations. In our previous work [@our_nqi] we performed self-consistent electronic structure calculations of this system with a twelve-atom super-cell (12A-SC: $a \times b \times 2c$). Relaxations were performed only for the neutral charge state of the impurity and at the end of the relaxation process two electrons were added to compute EFG without computing the self-consistent potential of the charged cell. The structural relaxation obtained by this procedure were smaller than in the present work but account for the inversion of Cd-O1 and Cd-O2 distances \[$d$(Cd-O1)=2.12 [Å]{}, $d$(Cd-O2)=2.07 [Å]{}\] with respect to the unrelaxed structure (see Table \[tab:distances\]). The description obtained for the EFG was very similar to the one of the 72A-SC\[Cd$^-$(2e)\], but we understand that this agreement is somehow fortuitous because usage of 12A-SC relaxed coordinates to compute the EFG in the 72A-SC\[Cd$^-$(2e)\] gives very different results. In Ref. , the authors have used a very similar approach as that we used in Ref.  using a 12A-SC, but they assumed that the relaxations of the nearest oxygen neighbors of the Cd atom were isotropic. As a consequence, they obtained $d$(Cd-O1)=2.04 [Å]{}, $d$(Cd-O2)=2.08 [Å]{}, and also a very different result for EFG (see Table I), $V_{33}$=$V_{ZZ}$=-5.09 $\times$ and $\eta$=0.39 (for a carefully comparison of these two calculations see Ref.). A question that arises at this point is if self-consistent electronic FLAPW calculations performed with a [*converged*]{} SC give EFG compatible with experience or not in the case that relaxations of the oxygen NN of the Cd atom are constrained to be isotropic. This is an interesting point in order to know if available experimental data is enough to refute the assumption that the oxygen NN relax isotropically. We have performed self-consistent calculations for the system 72A-SC\[Cd$^-$(2e)\] for different positions of O1 and O2 atoms but moving them outwards, keeping the relation $d$(Cd-O1)/$d$(Cd-O2) constant. We determined the equilibrium position of oxygen atoms as the one that produce a minimum in the energy \[see Fig. \[relaj\](a)\]. We obtain that Cd-O distances are 2.12 [Å]{} and 2.16 [Å]{} for O1 and O2 atoms, respectively \[a relaxation of 9% of the unrelaxed distances, see Fig. \[relaj\](a)\]. If we compare Figs. \[relaj\](a) and (b), we can note that due to the assumption of isotropic relaxations, there is no inversion in the Cd-O distances and, in consequence, the strong change in the EFG components does not take place. In particular, there is not change in sign and orientation of $V_{33}$ as in the case of our free relaxation. At the equilibrium position we obtained for $V_{33}$ a value of -4.46 $\times$  (pointing in \[001\]-direction) and a high $\eta$ value of 0.91, confirming that an isotropic relaxation is not consistent with the experimental data. CONCLUSIONS =========== In this work we have studied through a series of first-principles calculations the problem of a Cd impurity substitutionally-located at the cationic site in rutile TiO$_2$. The main result of our work, i.e. that Cd introduces in the host fairly anisotropic relaxations of its nearest oxygen neighbors and that this produces a change of orientation of V$_{33}$ from the \[001\] to the \[110\] direction when a Ti atom is replaced by a Cd atom in pure TiO$_2$, was briefly presented in a recent work with the experimental confirmation of the last prediction [@prl]. In this work we have presented details about the electronic structure of the different impurity systems considered. We have considered atomic relaxations and electronic structure self-consistently and have obtained that both aspects of the problem interact with each other. We obtained that atomic relaxations are different for the charged and neutral state of the impurity and that, on the other side, the relaxation process produces a drastic variation in the asymmetry of the charge distribution near the probe-atom for a given charge state of the impurity, which is detected in the strong variation of the EFG tensor. We have shown that the huge difference in the values of the asymmetry parameter $\eta$ between the charged and neutral state of the impurity arises because of the filling of the impurity level at the Fermi energy. This difference in the $\eta$ value determines, through comparison with experiment, that Cd is in a charged state when it is introduced as impurity in TiO$_2$ at room temperature. From these results we have confirmed that the EFG tensor is a very useful magnitude because it is sensible to subtle details of the electronic structure and it can be determined experimentally with high resolution. We have performed a series of checks of the accuracy of the present calculations in order to show that all the predictions of this work are the same if an increment in the basis size, k-mesh or size of the SC are considered or if a different exchange-correlation potential is used. We have also shown that considering relaxations beyond nearest neighbors does not produce any qualitative change in our results. Finally, we have checked that the hypothesis of isotropic relaxations and the use of PCM approximations give results incompatible with experiments and with our calculations. From our results it is clear that the problem of the EFG at Cd impurities in TiO$_2$ is too complicated to be described (even approximately) by simple models like PCM, antishielding factors and isotropic relaxations. We can conclude that a proper theoretical description of electronic properties of metal impurities in oxide semiconductors should consider self-consistently the charge-state of the impurity and the impurity-induced distortions in the host, specially in the first shell of neighbors of the impurity. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== We are indebted to Prof. Dr. Mariana Weissmann for fruitful discussions and critical reading of the manuscript. We gratefully acknowledge the support of Prof. Dr. A.G. Bibiloni on this project. We are grateful to Dr. M. Cervera for helpful suggestions. This work was partially supported by CONICET, Agencia Nacional de Promoción Científica y Tecnológica (ANPCyT) under PICT98 03-03727, and Fundación Antorchas, Argentina, and The Third World Academy of Sciences (TWAS), Italy (RGA97-057). See, e.g., G. Schatz and A. Weidinger, in [*Nuclear Condensed Matter Physics - Nuclear Methods and Applications*]{}, edited by John Wiley & Sons (Chichester, England, 1996) p.63; Proceedings of the 12th International Conference on Hyperfine Interactions, Park City, Utah, 2001, edited by W.E. Evenson, H. Jaeger, and M.O. Zacate, Hyp. Interactions [**136/137**]{}, 2001. E.N. Kaufmann and R.J. Vianden, Rev. Mod. Phys. [**51**]{}, 161 (1979). H. Frauenfelder and R. Steffen, in [*$\alpha$-, $\beta$-, and $\gamma$-Ray Spectroscopy*]{}, edited by K. Siegbahn (North-Holland, Amsterdam, 1968), Vol. 2, p. 917. S.H. Wei and H. Krakauer, Phys. Rev. Lett. [**55**]{}, 1200 (1985). P. Dufek, P. Blaha, and K. Schwarz, Phys. Rev. Lett. [**75**]{}, 3545 (1995). P. Blaha, K. Schwarz, and P.H.Dederichs, Phys. Rev. B [**37**]{}, 2792 (1988); K. Schwarz, C. Ambrosch-Draxl, and P. Blaha, Phys. Rev. B [**42**]{}, 2051 (1990). C. Verdozzi, D.R. Jennison, P.A. Schultz, M.P. Sears, J.C. Barbour, and B.G. Potter, Phys. Rev. Lett. [**80**]{}, 5615 (1998). A. Settels, T. Korhonen, N. Papanikolaou, R. Zeller, and P.H. Dederichs, Phys. Rev. Lett. [**83**]{}, 4369 (1999). C. Stampfl, C.G. Van de Walle, D. Vogel, P. Krüger, and J. Pollmann, Phys. Rev. B [**61**]{}, R7846 (2000). S. Lany, P. Blaha, J. Hamann, V. Ostheimer, H.Wolf, and T. Wichert, Phys. Rev. B [**62**]{}, R2259 (2000). S. Jeong and A. Oshiyama, Phys. Rev. Lett. [**86**]{}, 3574 (2001). C.O. Rodriguez, M. V. Ganduglia-Pirovano, E.L. Peltzer y Blancá, and M. Petersen, Phys. Rev. B [**64**]{}, 144419 (2001). L.A. Terrazos, H.M. Petrilli, M. Marszalek, H. Saitovitch, P.R.J. Silva, P. Blaha, K. Schwarz, Solid State Commun. [**121**]{}, 525 (2002). A.N. Baranov, V. S. Stepanyuk, W. Hergert, A.A. Katsnelson, A. Settels, R. Zeller, and P.H. Dederichs, Phys. Rev. B [**66**]{}, 155117 (2002). Su-Huai Wei and S.B. Zhang, Phys. Rev. B [**66**]{}, 155211 (2002). L.A. Errico, G. Fabricius, M. Rentería, P. de la Presa, and M. Forker, Phys. Rev. Lett. [**89**]{}, 55503 (2002). R.M. Sternheimer, Phys. Rev. [**95**]{}, 736 (1954) , and references therein. T. Wenzel, A. Bartos, K.P. Lieb, M. Uhrmacher, and D. Wiarda, Ann. Physik. [**1**]{}, 155 (1992). K. Sato, H. Akai, and T. Minamisono, Z. Naturforsch. [**53a**]{}, 396 (1998). P. Blaha, K. Schwarz, P. Dufek, and J. Luitz, [wien97]{}, Vienna University of Technology, 1997. Improved and updated Unix version of the original copyrighted [Wien]{}-code, which was published by P. Blaha, K. Schwarz, P.I. Sorantin, and S. B. Trickey, in Comput. Phys. Commun. [**59**]{}, 399 (1990). J.P. Perdew and Y. Wang, Phys. Rev. B [**45**]{}, 13244 (1992). R. Yu, D. Singh, and H. Krakauer, Phys. Rev. B [**43**]{}, 6411 (1991). B. Kohler, S. Wilker, M. Scheffler, R. Kouba, and C. Ambrosch-Draxl, Comput. Phys. Commun. [**94**]{}, 31 (1996). R.J. Hill and C.J. Howard, J. Appl. Cryst. [**20**]{}, 467 (1987). L. Errico, G. Fabricius, and M. Rentería, Z. Naturforsch. [**55a**]{}, 267 (2000); [**55**]{}, 983(E)(2000). J.M. Adams and G.L. Catchen, Phys. Rev. B [**50**]{}, 1264 (1994). O. Kanert and H. Kolem, J. Phys. C [**21**]{}, 3909 (1988). When axial symmetry is present (this is the case of the variation in the $d$ contribution that we try to explain) it is useful to define the asymmetry count, which for $d$ orbitals is: $\Delta n_d = n_{d_{yz}} - 1/2 (n_{d_{xy}}+n_{d_{xz}}) + 1/2 (n_{d_{z^2}} - n_{d_{x^2-y^2}})$, where $n_i$ is the charge of $i$-character (in electrons) inside the MT sphere and $x$ is the axial axis. With this definition $V_{XX}(d)$=$I_d \Delta n_d$ and $V_{YY}$=$V_{ZZ}$=-$V_{XX}$/2. Note that this way of defining $\Delta n_d$ differs from the usual one in FLAPW [@efg] only because of the choice of the axial axis. P. Blaha and K. Schwarz, Hyp. Interactions [**52**]{}, 153 (1989). J.P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. [**77**]{}, 3865 (1996). F.D. Feiock and W.R. Johnson, Phys. Rev. [**187**]{}, 39 (1969). P. Blaha, D.J. Singh, P.I. Sorantin, and K. Schwarz, Phys. Rev. B [**46**]{}, 1321 (1992).

ArXiv

--- abstract: 'We derive path-integral expressions for the second and third virial coefficients of monatomic quantum gases. Unlike previous work that considered only Boltzmann statistics, we include exchange effects (Bose–Einstein or Fermi–Dirac statistics). We use state-of-the-art pair and three-body potentials to calculate the third virial coefficient of ${}^3$He and ${}^4$He in the temperature range $2.6 - 24.5561$ K. We obtain uncertainties smaller than those of the limited experimental data. Inclusion of exchange effects is necessary to obtain accurate results below about 7 K.' author: - Giovanni Garberoglio - 'Allan H. Harvey' bibliography: - 'virial-He.bib' title: 'Path-integral calculation of the third virial coefficient of quantum gases at low temperatures [^1] ' --- Introduction ============ Thermodynamic properties of fluids at very low temperatures are of significant interest. For example, the current International Temperature Scale [@its90] makes use of volumetric properties and vapor pressures of helium isotopes below the triple point of neon (24.5561 K); below the triple point of hydrogen (13.8033 K), the scale is based entirely on properties of ${}^3$He and ${}^4$He. The theoretical analysis of relevant properties at these conditions, such as the virial coefficients that describe the fluid’s departure from ideal-gas behavior, is complicated by the presence of quantum effects. The inclusion of quantum effects in the calculation of virial coefficients was one of the first numerical applications of the Path-Integral Monte Carlo (PIMC) method. [@FH] In a series of pioneering works published in the 1960’s, Fosdick and Jordan showed how to calculate the second and third virial coefficient of a monatomic gas using computer simulations. [@fosdick-jordan66; @jordan-fosdick68; @fosdick68] Given the limited computational resources available at that time, they were able to calculate the third virial coefficient only in the case of two-body interactions, using a model potential of the Lennard-Jones form and assuming distinguishable particles (Boltzmann statistics). They argued that their method could be extended to include the proper quantum statistics, but they were able to compute exchange effects only in the case of the second virial coefficient. Recently, the exponential increase in computational power has enabled use of the path-integral method to calculate the properties of quantum degenerate systems, notably superfluid helium. [@Ceperley] At the same time, progress in the computation of [*ab initio*]{} electronic properties of interacting atoms resulted in the availability of very precise two- and three-body interparticle potentials, at least for the lightest particles such as helium atoms [@Jez07; @HBV07; @He-3body; @CPS09; @He2-2010] or hydrogen molecules. [@diep-johnson00; @virial-H2] A natural application for these potentials is the calculation of virial coefficients. As is well known, the second virial coefficient depends only on the two-body potential, the third virial coefficient depends only on two-body and three-body interactions, etc. The second virial coefficient for a monatomic gas can be rigorously obtained at the fully quantum level from the calculation of the phase shifts due to the pair potential, and previous work has shown that a completely [*ab initio*]{} calculation of second virial coefficients for helium can have uncertainties comparable to and in many cases smaller than those of the most precise experiments. [@phi07; @hur00; @BHV07; @Mehl09; @Cencek11] In the case of the third virial coefficient, no closed-form solution of the quantum statistical mechanics problem is known. First-order semiclassical approaches have been derived [@yokota60; @ram-singh73] and show that, in the case of helium, quantum diffraction effects result in significant modifications of the classical result, even at room temperature. However, there is no rigorous way to evaluate the accuracy or uncertainty of the semiclassical result, especially at low temperatures. In recent work, [@Garberoglio2009b] we extended the methodology pioneered by Fosdick and Jordan, deriving a set of formulae allowing a path-integral calculation of the third virial coefficient $C(T)$ of monatomic species for arbitrary two- and three-body potentials. Our results were limited to Boltzmann statistics (i.e., distinguishable particles) and we did not present results for temperatures lower than the triple point of neon ($24.5561$ K), which we deemed to be a reasonable lower bound so that exchange effects could be neglected. Nevertheless, we were able to compute the value of the third virial coefficient of ${}^4$He with an uncertainty one order of magnitude smaller than that of the best experiments. Recent experimental results overlapping with our temperature range, [@Gaiser09; @Gaiser10] although mostly consistent with our calculations, seemed to indicate a systematic deviation which the authors speculated could originate from our neglect of the proper quantum statistics of helium atoms. In this paper, we extend our computational methodology to calculate the quantum statistical contributions to the third virial coefficient, and compute $C(T)$ for both isotopes of helium in the temperature range $2.6 - 24.5561$ K, extending the temperature range considered in our previous work down into the range where exchange effects are important. We show that quantum statistical effects are significant only for temperatures smaller than about $7$ K, and compare our results to low-temperature experimental data. In a subsequent publication [@paper2], we will present results covering the entire temperature range (improving on our previous results for ${}^4$He at 24.5561 K and above) with rigorously derived uncertainties. We will also extend our methodology to include acoustic virial coefficients, and compare those calculations to available data. In the present work, our focus is on low temperatures and specifically on the effect of non-Boltzmann statistics. Path-integral calculation of the virial coefficients ==================================================== The second and third virial coefficients, $B(T)$ and $C(T)$ respectively, are given by [@hcb] $$\begin{aligned} B(T) &=& -\frac{1}{2 V} \left( Z_2 - Z_1^2 \right) \label{eq:B} \\ C(T) &=& 4 B^2(T) - \frac{1}{3 V} \left[ Z_3 - 3 Z_2 Z_1 + 2 Z_1^3 \right], \label{eq:C}\end{aligned}$$ where $V$ is the integration volume (with the limit $V \rightarrow \infty$ taken at the end of the calculations), and the functions $Z_N$ are given by: $$\begin{aligned} Z_3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \langle 1 2 3 | \mathrm e^{-\beta \hat H_3} \sum_{\pi_3} {\cal P}_{\pi_3} | 1 2 3 \rangle \label{eq:Z3} \\ Z_2 &=& \Lambda^6 \int \mathrm d1 \mathrm d2 ~ \langle 1 2 | \mathrm e^{-\beta \hat H_2 } \sum_{\pi_2} {\cal P}_{\pi_2} | 1 2 \rangle \label{eq:Z2} \\ Z_1 &=& \Lambda^3 \int \mathrm d1 ~ \langle 1 | \mathrm e^{-\beta \hat H_1} | 1 \rangle = V, \label{eq:Z1}\end{aligned}$$ where $\hat H_N$ is the $N$-body Hamiltonian, $\beta = 1/(k_\mathrm B T)$, ${\cal P}$ is a permutation operator (multiplied by the sign of the permutation in the case of Fermi–Dirac statistics), the index $\pi_3$ runs over the 6 permutations of 3 objects (i.e., $123$, $132$, $213$, $321$, $231$ and $312$), and $\pi_2$ runs over the 2 permutations of 2 objects (i.e., $12$ and $21$). $\Lambda = h / \sqrt{2 \pi m k_\mathrm B T}$ is the thermal de Broglie wavelength of a particle of mass $m$ at temperature $T$. For the sake of conciseness, we denote by $| i \rangle$ an eigenvector of the position operator relative to particle $i$ and by $\mathrm d i$ ($i=1,2,3$) the integration volume relative to the Cartesian coordinates of the $i$-th particle. Note that, in order to produce the molar units used by experimenters and in our subsequent comparisons with data, the right side of Eq. (\[eq:B\]) and the second term in the right side of Eq. (\[eq:C\]) must be multiplied by Avogadro’s number and its square, respectively. In the following, we will derive a path-integral expression for the calculation of the virial coefficients with Eqs. (\[eq:B\]) and (\[eq:C\]). We perform the derivation in detail in the case of $B(T)$ to establish the notation, and then extend the results to the more interesting case of $C(T)$. Second virial coefficient ------------------------- In this paper, we adopt Cartesian coordinates to describe the atomic positions. This differs from the approach developed in Refs.  and , where Jacobi coordinates were used. This choice allows the exchange contribution to be computed in a much simpler manner than would be the case if Jacobi coordinates were used, especially in the case of three or more particles. From Eqs. (\[eq:B\]) and (\[eq:Z2\]), it can be seen that there are two contributions to $B(T)$. The first one comes from considering the identity permutation only, and takes into account only quantum diffraction effects. This is the only contribution that gives a nonzero result at high temperatures, where the particles can be treated as distinguishable (Boltzmann statistics). The second contribution to $B(T)$, which we will call [*exchange*]{} (xc), comes from the only other permutation involved in the definition of the quantity $Z_2$ above. The expression of these two contributions in Cartesian coordinates is: $$\begin{aligned} B_\mathrm{Boltzmann}(T) &=& -\frac{\Lambda^6}{2 V} \int \mathrm d \boldsymbol{r}_1 \mathrm d \boldsymbol{r}_2 ~ \langle \boldsymbol{r}_1 \boldsymbol{r}_2 | \exp\left[ -\beta (\hat K_2 + \hat U_2(|\boldsymbol{r}_2 - \boldsymbol{r}_1|) ) \right] - \exp\left[ -\beta \hat K_2 \right] | \boldsymbol{r}_1 \boldsymbol{r}_2 \rangle \label{eq:Bdir} \\ B_\mathrm{xc}(T) &=& \mp \frac{\Lambda^6}{2 V} \int \mathrm d \boldsymbol{r}_1 \mathrm d \boldsymbol{r}_2 ~ \langle \boldsymbol{r}_1 \boldsymbol{r}_2 | \exp\left[ -\beta (\hat K_2 + \hat U_2(|\boldsymbol{r}_2 - \boldsymbol{r}_1|) ) \right]| \boldsymbol{r}_2 \boldsymbol{r}_1 \rangle, \label{eq:Bx} \end{aligned}$$ where we denote by $K_N$ the total kinetic energy of $N$ bodies and by $\hat U_2(r)$ the two-body potential energy operator. The upper (lower) sign in Eq. (\[eq:Bx\]) corresponds to Bose–Einstein (Fermi–Dirac) statistics. Equations (\[eq:Bdir\]) and (\[eq:Bx\]) can be rewritten by using the Trotter identity $$\mathrm e^{\hat K_2 + \hat U_2} = \lim_{P \rightarrow \infty} \left( \mathrm e^{\hat K_2/P} \mathrm e^{\hat U_2/P} \right)^P \label{eq:Trotter}$$ with a positive integer value of the Trotter index $P$. Following the procedure outlined in Ref. , one can then write $B_\mathrm{Boltzmann}(T)$ as $$\begin{aligned} B_\mathrm{Boltzmann}(T) = - 2 \pi^2 \int_0^\infty r^2 \mathrm dr ~ \left(\exp\left[-\beta \overline{U}_2(r)\right] - 1 \right), \label{eq:BBoltz}\end{aligned}$$ where the two-body effective potential $\overline U_2(r)$ is given by $$\begin{aligned} \exp\left[-\beta \overline{U}_2(r)\right] &=& \int \prod_{i=1}^{P-1} \mathrm d\Delta\boldsymbol{x}_1^{(i)} \mathrm d\Delta\boldsymbol{x}_2^{(i)} ~ \exp\left[ -\frac{\beta}{P} \sum_{i=1}^{P} U_2(|\boldsymbol{r} + \boldsymbol{x}_2^{(i)} - \boldsymbol{x}_1^{(i)}|) \right] \times \nonumber \\ & & F_\mathrm{ring}(\Delta\boldsymbol{x}_1^{(1)},\ldots,\Delta\boldsymbol{x}_1^{(P)}) F_\mathrm{ring}(\Delta\boldsymbol{x}_2^{(1)},\ldots,\Delta\boldsymbol{x}_2^{(P)}) \label{eq:U2_eff} \\ &\underset{P \rightarrow \infty}{=}& \oint {\cal D} \boldsymbol x_1 {\cal D} \boldsymbol x_2 ~ \exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{m}{2} \left| \frac{\mathrm d \boldsymbol x_1(\tau)}{\mathrm d \tau} \right|^2 + \frac{m}{2} \left| \frac{\mathrm d \boldsymbol x_2(\tau)}{\mathrm d \tau} \right|^2 ~ + U_2(|\boldsymbol r + \boldsymbol x_1(\tau) - \boldsymbol x_2(\tau)|) \mathrm d \tau \right], \label{eq:U2_pi}\end{aligned}$$ where $$F_\mathrm{ring} = \Lambda^3 \left( \frac{P^{3/2}}{\Lambda^3} \right)^P \exp\left[ -\frac{\pi P}{\Lambda^2} \sum_{i=1}^P \left| \Delta\boldsymbol{x}_1^{(i)} \right|^2 \right]. \label{eq:Fring}$$ In the previous equations, we have defined $\Delta\boldsymbol{x}_k^{(i)} = \boldsymbol{r}_k^{(i+1)} - \boldsymbol{r}_k^{(i)}$, where $\boldsymbol{r}_k^{(i)}$ is the coordinate of particle $k$ ($k=1,2$) in the $i$-th “imaginary time slice”. These “slices” are obtained by inserting $P$ completeness relations of the form $$1 = \int \mathrm d\boldsymbol{r}_1^{(i)} \mathrm d\boldsymbol{r}_2^{(i)} ~ | \boldsymbol{r}_1^{(i)} \boldsymbol{r}_2^{(i)} \rangle \langle \boldsymbol{r}_1^{(i)} \boldsymbol{r}_2^{(i)} |$$ between the factors $\mathrm e^{\hat K_2/P}$ and $\mathrm e^{\hat U_2/P}$ of the Trotter expansion of Eq. (\[eq:Trotter\]). We used the overall translation invariance of the system to remove the factor $V$ in Eq. (\[eq:Bdir\]) and fix the $\tau=0$ slice of particle 2 at the origin of the coordinate system. We also denoted by $\boldsymbol{x}_1^{(i)}$ and $\boldsymbol{x}_2^{(i)}$ the coordinates of two ring polymers having one of their endpoints fixed at the origin ($\boldsymbol{x}_1^{(1)} = \boldsymbol{x}_2^{(1)} = \mathbf 0$), and we introduced the variable $\boldsymbol{r}$ denoting the distance between the $\tau = 0$ time slice of the two ring polymers. In the classical limit, where the paths $\boldsymbol x_1(\tau)$ and $\boldsymbol x_2(\tau)$ shrink to a point, the coordinate $r$ reduces to the distance between the particles and one has $\overline U_2(r) = U_2(r)$. Note that the effect of the identity permutation is to set $\boldsymbol{r}_k^{(P+1)} = \boldsymbol{r}_k^{(1)}$. The path-integral formalism allows one to map the quantum statistical properties of a system with $N$ distinguishable particles (Boltzmann statistics) onto the classical statistical properties of a system of $N$ ring polymers, each having $P$ beads (sometimes called imaginary-time slices), which are distributed according to the function $F_\mathrm{ring}$ of Eq. (\[eq:Fring\]). [@boltzmann-bias] The mapping is exact in the $P \rightarrow \infty$ limit, although convergence is usually reached with a finite (albeit large) value of $P$. In the calculation of the second virial coefficient, Eq. (\[eq:BBoltz\]) shows that the second virial coefficient at the level of Boltzmann statistics is obtained from an expression similar to that for the classical second virial coefficient, using an effective two-body potential. This effective potential, $\overline{U}_2(r)$, is obtained by averaging the intermolecular potential $U_2(r)$ over the coordinates of two ring polymers, corresponding to the two interacting particles entering the definition of $B(T)$. Equation (\[eq:BBoltz\]) is equivalent to Eq. (19) of Ref. . The only difference is that the current approach uses Cartesian coordinates, and therefore we are left with an average over two ring polymers of mass $m$ instead of one ring polymer of mass $\mu = m/2$, corresponding to the relative coordinate of the two-particle system. The two approaches are of course equivalent, and in fact it can be shown that Eqs. (\[eq:BBoltz\]) and (\[eq:U2\_eff\]) reduce to the form derived in Refs.  and . Equation (\[eq:BBoltz\]) is the same expression previously derived by Diep and Johnson for spherically symmetric potentials on the basis of heuristic arguments, [@diep-johnson00] and later generalized by Schenter to the case of rigid bodies and applied to a model for water. [@Schenter02] Equation (\[eq:U2\_eff\]) is actually the discretized version of a path integral, as shown in Eq. (\[eq:U2\_pi\]). The circled integral is defined as $$\begin{aligned} \oint {\cal D} \boldsymbol x ~ \exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{m}{2} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2 ~ \mathrm d \tau \right] &\equiv& \nonumber \\ \lim_{P \rightarrow \infty} \int \prod_{i=1}^{P-1} \mathrm d\Delta\boldsymbol{x}^{(i)} F_\mathrm{ring}(\Delta\boldsymbol{x}^{(1)},\ldots,\Delta\boldsymbol{x}^{(P)}) &=& 1, \label{eq:pi}\end{aligned}$$ and it indicates that one has to consider all the cyclic paths with ending points at the origin, that is $\boldsymbol x(0) = \boldsymbol x(\beta \hbar) = \boldsymbol 0$. The normalization of the path integral is also indicated in Eq. (\[eq:pi\]). We can perform on Eq. (\[eq:Bx\]), describing the exchange contribution to the second virial coefficient, the same steps leading from Eq. (\[eq:Bdir\]) to Eq. (\[eq:BBoltz\]). The only difference is the presence of the permutation operator, whose main consequence is fact that $\boldsymbol r^{(P+1)}_1 = \boldsymbol r^{(1)}_2$ and $\boldsymbol r^{(P+1)}_2 = \boldsymbol r^{(1)}_1$. In this case, defining $\boldsymbol X^{(i)} = \boldsymbol r^{(i)}_1$ and $\boldsymbol X^{(i+P)} = \boldsymbol r^{(i)}_2$, one obtains $$\begin{aligned} B_\mathrm{xc}(T) &=& \mp \frac{\Lambda^6}{2 V} \int \mathrm d \boldsymbol X^{(1)} \ldots \mathrm d \boldsymbol X^{(2P)} ~ \exp\left[-\frac{\beta}{P} \sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right] \times \nonumber \\ & & \left( \frac{P^{3/2}}{\Lambda^3} \right)^{2P} \exp\left[-\frac{\pi P}{\Lambda^2} \sum_{i=1}^{2P} \left( \boldsymbol X^{(i+1)} - \boldsymbol X^{(i)}\right)^2 \right] \label{eq:Bxc} \\ &=& \mp \frac{\Lambda^3}{2 V} \int \mathrm d \boldsymbol X^{(1)} \ldots \mathrm d \boldsymbol X^{(2P)} ~ \exp\left[-\frac{\beta}{P} \sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right] \times \nonumber \\ & & \frac{\Lambda_\mu^3}{2^{3/2}} \left( \frac{(2P)^{3/2}}{\Lambda_\mu^3} \right)^{2P} \exp\left[-\frac{\pi ~ 2 P}{\Lambda_\mu^2} \sum_{i=1}^{2P} \left( \boldsymbol X^{(i+1)} - \boldsymbol X^{(i)}\right)^2 \right] \\ &=& \mp \frac{1}{2} \frac{\Lambda^3}{2^{3/2}} \left\langle \exp\left[-\frac{\beta}{P} \sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right] \right\rangle \label{eq:Bexchange} \\ &\underset{P \rightarrow \infty}{=}& \mp \frac{\Lambda^3}{2^{5/2}} ~ \oint {\cal D} \boldsymbol X ~ \exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{\mu}{2} \left| \frac{\mathrm d \boldsymbol X(\tau)}{\mathrm d \tau} \right|^2 + U_2(|\boldsymbol X(\tau + \beta \hbar /2) - \boldsymbol X(\tau)|) ~ \mathrm d\tau \right], \label{eq:Bexch_pi}\end{aligned}$$ where we have defined $\Lambda_\mu = \sqrt{2}\Lambda$. The exchange contribution to the second virial coefficient is given simply as an average of the two-body potential taken on ring polymers corresponding to particles of mass $\mu = m/2$. In the discretized version of the path integral, one has to consider $2P$ beads. In Eq. (\[eq:Bexchange\]), we have used the overall translation invariance of the integral to remove the factor of $V$ in the denominator. The effect of the various permutations can be visualized as generating paths with a larger number of beads, which are obtained by coalescing the ring polymers corresponding to the particles that are exchanged by the permutation operator. Third virial coefficient ------------------------ We now discuss the third virial coefficient, starting from the expression given in Eq. (\[eq:C\]). Since $4 B^2(T)$ can be calculated by the methods of the previous section, we concentrate on the second term, whose summands can be written as follows: $$\begin{aligned} Z_3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \langle 1 2 3 | \mathrm e^{-\beta \hat H_3} \sum_{\pi_3} {\cal P}_{\pi_3} | 1 2 3 \rangle \\ Z_2 Z_1 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \langle 1 2 3 | \mathrm e^{-\beta (\hat H_2 + \hat T_3)} \sum_{\pi_2} {\cal P}_{\pi_2} | 1 2 3 \rangle \label{eq:Z2Z1} \\ Z_1^3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \langle 1 2 3 | \mathrm e^{-\beta \hat K_3} | 1 2 3 \rangle, \label{eq:Z13}\end{aligned}$$ where $\hat T_i = -\frac{\hbar^2}{2m} \nabla_i^2$ is the kinetic energy operator of particle $i$. We can simplify the expression in square brackets on the right-hand side of Eq. (\[eq:C\]) by writing the three $Z_2 Z_1$ terms choosing each time a different particle for $Z_1$ (in Eq. (\[eq:Z2Z1\]) we have chosen particle 3 as coming from $Z_1$). After considering all the permutations of two and three particles, we end up with $6 + 3 \times 2 + 1 = 13$ terms building the term in square brackets of Eq. (\[eq:C\]). It is useful to collect these 13 terms as follows: 1. [[*Term 1 (identity term)*]{}: we sum together permutation $123$ from $Z_3$, the identity permutations from the three $Z_2 Z_1$ and the whole $2 Z_1^3$ term. Adding $4 B^2_{\rm Boltzmann}(T)$, one obtains the Boltzmann expression for $C(T)$, already discussed in Ref. . In the present formulation based on Cartesian coordinates, the value $C(T)$ in the case of Boltzmann statistics involves an average over three independent ring polymers, which correspond to the three particles. In the following, this contribution to $C(T)$ will be referred to as $C_{\rm Boltzmann}(T)$ and is made by $1 + 3 + 1 = 5$ of the 13 terms described above.]{} 2. [[*Term 2 (odd term)*]{}: we take permutations $132$, $213$ and $321$ from $Z_3$ and the three exchange permutations from the $Z_2 Z_1$ terms. These permutations are all odd, and we consider them with a positive sign (Bose–Einstein statistics). In the case of Fermi–Dirac statistics, this term has to be multiplied by an overall minus sign. All of these permutations correspond to configurations where two of the three particles are exchanged. The sum of these 6 terms will be referred to as $C_{\rm odd}(T)$.]{} 3. [[*Term 3 (even term)*]{}: we take the permutations $231$ and $312$ from $Z_3$. These are the remaining two terms from the 13, and are both even permutations, hence the name. Both of these terms correspond to a cyclic exchange of the three particles, and their sum will be referred to as $C_{\rm even}(T)$.]{} Using these definitions, the full $C(T)$, including quantum statistical effects, can be written as $$C(T) = C_{\rm Boltzmann}(T) \pm C_{\rm odd}(T) + C_{\rm even}(T) + C_{\rm B}(T), \label{eq:C_full}$$ where the last term in the right-hand sum is given by $$C_{\rm B}(T) = \pm 8 B_{\rm Boltzmann}(T) B_\mathrm{xc}(T) + 4 B^2_\mathrm{xc}(T), \label{eq:CB}$$ since the contribution of $4 B^2_\mathrm{Boltzmann}(T)$ to $C(T)$ is already included in $C_{\rm Boltzmann}(T)$. In Eqs. (\[eq:C\_full\]) and (\[eq:CB\]), the upper (lower) sign corresponds to Bose–Einstein (Fermi–Dirac) statistics. Using the same procedure outlined above in the case of $B(T)$, one can write the Boltzmann contribution to the third virial coefficient as $$\begin{aligned} C_\mathrm{Boltzmann}(T) &=& 4 B^2_\mathrm{Boltzmann}(T) - \nonumber \\ & & \frac{1}{3} \int \mathrm d\boldsymbol{r}_1 \mathrm d\boldsymbol{r}_2 \left[ \mathrm e^{-\beta \overline{V}_3(\boldsymbol{r}_1,\boldsymbol{r}_2)} - \mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_1|)} - \mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_2|)} - \mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_1 - \boldsymbol{r}_2|)} + 2 \right], \label{eq:CBoltz} \\ \exp\left[-\beta \overline{V}_3(\boldsymbol{r}_1,\boldsymbol{r}_2)\right] &=& \int \prod_{i=1}^{P-1} \Delta \boldsymbol x_1^{(i)} \Delta \boldsymbol x_2^{(i)} \Delta \boldsymbol x_3^{(i)} ~ F_\mathrm{ring}^{(1)} F_\mathrm{ring}^{(2)} F_\mathrm{ring}^{(3)} ~ \exp\left[-\beta \overline{V}_3^{\mathrm B}(\boldsymbol{r}_1,\boldsymbol{r}_2) \right] \label{eq:V3disc} \\ &\underset{P \rightarrow \infty}{=}& \oint {\cal D} \boldsymbol x_1 {\cal D} \boldsymbol x_2 {\cal D} \boldsymbol x_3 ~ \exp\left[-\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{m}{2} \left( \left| \frac{\mathrm d \boldsymbol x_1(\tau)}{\mathrm d \tau} \right|^2 + \left| \frac{\mathrm d \boldsymbol x_2(\tau)}{\mathrm d \tau} \right|^2 + \left| \frac{\mathrm d \boldsymbol x_3(\tau)}{\mathrm d \tau} \right|^2 \right) + \right. \nonumber \\ & & \left. V_3(\boldsymbol r_1 + \boldsymbol x_1(\tau), \boldsymbol r_2 + \boldsymbol x_2(\tau), \boldsymbol x_3(\tau)) ~ \mathrm d \tau \right], \label{eq:V3}\end{aligned}$$ where $F_\mathrm{ring}^{(k)}$ denotes the probability distribution of the path relative to particle $k$, as defined in Eq. (\[eq:Fring\]). In Eq. (\[eq:V3disc\]), the three-body effective potential energy $\overline{V}_3$ is obtained as an average performed over [*three*]{} independent ring polymers of the total three-body potential energy: $$\begin{aligned} V_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z}) &=& U_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z}) + U_2(|\boldsymbol x - \boldsymbol y|) + \nonumber \\ & & U_2(|\boldsymbol x - \boldsymbol z|) + U_2(|\boldsymbol y - \boldsymbol z|),\end{aligned}$$ where $U_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z})$ is the non-additive three-body potential of three atoms. In Eq. (\[eq:V3disc\]) the total three-body potential energy for the Boltzmann contribution to the third virial coefficient is $$\begin{aligned} \overline{V}_3^{\mathrm B}(\boldsymbol{r}_1,\boldsymbol{r}_2) &=& \frac{1}{P} \sum_{i=1}^P U_3(\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)}, \boldsymbol{r}_2 + \boldsymbol{x}_2^{(i)}, \boldsymbol{x}_3^{(i)}) + \nonumber \nonumber \\ & & U_2(|\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)} - \boldsymbol{r}_2 - \boldsymbol{x}_2^{(i)}|) + \nonumber \nonumber \\ & & U_2(|\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)} - \boldsymbol{x}_3^{(i)}|) +\nonumber \nonumber \\ & & U_2(|\boldsymbol{r}_2 + \boldsymbol{x}_2^{(i)} - \boldsymbol{x}_3^{(i)}|),\end{aligned}$$ where the variables with superscript $(i)$ denote the coordinates of three ring polymers with one of the beads at the origin. Notice that in passing from Eq. (\[eq:C\]) to Eq. (\[eq:CBoltz\]) we have used the translation invariance of the integrand to perform the integration over $\boldsymbol r_3$, which removed the factor of $V$ in the denominator. As a consequence, the paths corresponding to particle 3 have their endpoints at the origin of the coordinate system (or, equivalently, the third particle is fixed at the origin when the classical limit is performed.) In the same limit, the variables $\boldsymbol r_1$ and $\boldsymbol r_2$ appearing in Eq. (\[eq:V3\]) reduce to the positions of particles 1 and 2, respectively, and one has $\overline V_3(\boldsymbol r_1,\boldsymbol r_2) = V_3(\boldsymbol r_1,\boldsymbol r_2)$. The term $C_\mathrm{odd}(T)$ is obtained by exchanging the positions of two particles. This operation reduces the number of ring polymers to two: one having $2P$ beads, corresponding to the exchanged particles, and the other having $P$ beads, corresponding to the remaining one. The odd contribution is given by $$\begin{aligned} C_\mathrm{odd}(T) &=& - \frac{\Lambda^9}V \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \langle 1 2 3 | \exp\left[-\beta \hat H_3 \right] - \exp\left[-\beta (\hat K_2 + \hat U_2(\boldsymbol{r}_2 - \boldsymbol{r}_1)) \right] | 2 1 3 \rangle \label{eq:Codd1} \\ &=& - \frac{\Lambda^3}{2^{3/2}} \int \mathrm d \boldsymbol r_3 ~ \left\langle \exp\left[ -\beta \overline{V}^\mathrm{odd}_3 \right] - \exp\left[ -\beta \overline{U}^\mathrm{odd}_2 \right] \right\rangle \label{eq:Codd} \\ &=& - \frac{\Lambda^3}{2^{3/2}} \int \mathrm d \boldsymbol r_3 ~ \left\{ \oint {\cal D}\boldsymbol x {\cal D}\boldsymbol y ~ \exp\left[-\frac{1}{\hbar} \int_{0}^{\beta \hbar} \frac{m}{4} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2 + \frac{m}{2} \left| \frac{\mathrm d \boldsymbol y(\tau)}{\mathrm d \tau} \right|^2 + V_3\left(\boldsymbol x\left(\tau + {\beta \hbar}/{2}\right), \boldsymbol x(\tau), \boldsymbol r_3 + \boldsymbol y(\tau)\right) \mathrm d \tau \right] \right. \nonumber \\ & & \left. - \oint {\cal D}\boldsymbol x ~ \exp\left[-\frac{1}{\hbar} \int_{0}^{\beta \hbar} \frac{m}{4} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2 + U_2(|\boldsymbol x(\tau + \beta \hbar /2) - \boldsymbol x(\tau)|) \mathrm d \tau \right] \right\},\end{aligned}$$ where we have defined $$\begin{aligned} \overline{V}^\mathrm{odd}_3(\boldsymbol r_3) &=& \frac{1}{P} \sum_{i=1}^{P} U_3(\boldsymbol X^{(i)}, \boldsymbol X^{(i+P)}, \boldsymbol r_3 + \boldsymbol x_3^{(i)}) + \nonumber \\ & & U_2(|\boldsymbol X^{(i)} - \boldsymbol X^{(i+P)}|) + \nonumber \\ & & U_2(|\boldsymbol X^{(i)} - \boldsymbol r_3 - \boldsymbol x_3^{(i)}|) + \nonumber \\ & & U_2(|\boldsymbol X^{(i+P)} - \boldsymbol r_3 - \boldsymbol x_3^{(i)}|) \label{eq:V_odd} \\ \overline{U}^\mathrm{odd}_2 &=& \frac{1}{P} \sum_{i=1}^{P} U_2(|\boldsymbol X^{(i)} - \boldsymbol X^{(i+P)}|). \label{eq:U2_odd}\end{aligned}$$ The $2P$ variables $\boldsymbol X^{(i)}$ have been defined analogously to what has been done in Eq. (\[eq:Bxc\]). Notice that in the discretized version, the average defining the odd exchange term in Eq. (\[eq:Codd\]) is performed over two different kinds of ring polymers: the first has $2P$ beads of mass $m/2$ and connects particles 1 and 2 whose coordinates are exchanged by the permutation operator, whereas the second – corresponding to the third particle of mass $m$ – has $P$ beads. A similar derivation holds for the even contribution to the third virial coefficient, which is given by $$\begin{aligned} C_\mathrm{even}(T) &=& -\frac{2 \Lambda^9}{3 V} \int \mathrm d1 \mathrm d2 \mathrm d3 ~ \left\langle 1 2 3 \left| \exp\left( -\beta \hat H_3 \right) \right| 3 1 2 \right\rangle = -\frac{2}{3} \frac{\Lambda^6}{3^{3/2}} \left\langle \exp\left( - \beta \overline{V}^\mathrm{even}_3 \right) \right\rangle \label{eq:Ceven} \\ &=& -\frac{2}{3} \frac{\Lambda^6}{3^{3/2}} \oint {\cal D}\boldsymbol{x} ~ \exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{m}{6} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2 + V_3(\boldsymbol x(\tau + 2 \beta \hbar /3), \boldsymbol x(\tau + \beta \hbar /3), \boldsymbol x(\tau)) \mathrm d \tau \right],\end{aligned}$$ where we have defined $$\begin{aligned} \overline{V}^\mathrm{even}_3 &=& \frac{1}{P} \sum_{i=1}^{P} U_3(\boldsymbol Y^{(i)}, \boldsymbol Y^{(i+P)}, \boldsymbol Y^{(i+2P)}) + \nonumber \\ & & U_2(|\boldsymbol Y^{(i)} - \boldsymbol Y^{(i+P)}|) + \nonumber \\ & & U_2(|\boldsymbol Y^{(i)} - \boldsymbol Y^{(i+2P)}|) + \nonumber \\ & & U_2(|\boldsymbol Y^{(i+P)} - \boldsymbol Y^{(i+2P)}|), \label{eq:V_even}\end{aligned}$$ together with $\boldsymbol Y^{(i)} = \boldsymbol r_1^{(i)}$, $\boldsymbol Y^{(i+P)} = \boldsymbol r_2^{(i)}$, and $\boldsymbol Y^{(i+2P)} = \boldsymbol r_3^{(i)}$. In the discretized version, the even contribution to the third virial coefficient is an average over the coordinates of the $3P$ beads of a single ring polymer corresponding to a particle of mass $m/3$. Notice that, from a computational point of view, the evaluation of the exchange contributions to the third virial coefficient is much less demanding than the calculation of the Boltzmann part, which is given as an integral over the positions of two particles. In fact, the odd contribution is calculated as an integration over the position of one particle only, whereas the even contribution is given by a simple average over ideal-gas ring-polymer configurations. In particular, the full calculation of $C(T)$ at the lowest temperature with 2.5 GHz processors required $\sim 2400$ CPU hours, only 15% of which was needed to calculate the exchange contributions. Results and discussion ====================== Details of the calculation -------------------------- We have calculated $C(T)$ for both isotopes of helium with the path-integral method described above. We used the highly accurate two-body potential of Przybytek [*et al.*]{}, [@He2-2010] which includes the most significant corrections (adiabatic, relativistic, and quantum electrodynamics) to the Born–Oppenheimer result. We also used the three-body [*ab initio*]{} potential of Cencek [*et al.*]{}, [@CPS09] which was derived at the Full Configuration Interaction level and has an uncertainty approximately one-fifth that of the three-body potential [@He-3body] used in our previous work. [@Garberoglio2009b] We generated ring-polymer configurations using the interpolation formula of Levy. [@levy54; @fosdick-jordan66] The number of beads was chosen as a function of the temperature $T$ according to the formulae $P = \mathrm{int}[(1200~\mathrm{K})/T] + 7$ for ${}^4$He and $P = \mathrm{int}[(1800~\mathrm{K})/T] + 7$ for ${}^3$He, where $\mathrm{int}[x]$ indicates the integer closest to $x$. These values of $P$ were enough to reach convergence in the path-integral results at all the temperatures considered in the present study. The spatial integrations were performed with the VEGAS algorithm [@vegas1], as implemented in the GNU Scientific Library, [@gsl] with 1 million integration points and cutting off the interactions at 4 nm. The three-body interaction was pre-calculated on a three-dimensional grid and interpolated with cubic splines. The values of the virial coefficient and their statistical uncertainty were obtained by averaging over the results of 256 independent runs. First of all, we checked that our methodology was able to reproduce well-converged fully quantum $B(T)$ calculations for helium, which were obtained using the same pair potential as the present work. [@Cencek11] Our results agree within mutual uncertainties with these independent calculations, and confirm the observation, already made when analyzing theoretical $B(T)$ calculations performed using Lennard-Jones potentials, that exchange effects are significant only for temperatures lower than about 7 K. [@boyd69] The exchange contribution to the second virial coefficient is negative in the case of Bose–Einstein statistics and positive in the case of Fermi–Dirac statistics, as one would expect. The third virial coefficient of ${}^4$He ---------------------------------------- [d|dd|dd|dd|dd|dd]{} & & & & &\ (K) & & & & &\ 2.6 & 266 & 21 & 245 & 21 & -863 & 3 & -88.1 & 0.5 & 972 & 1\ 2.8 & 631 & 21 & 607 & 20 & -504 & 2 & -49.0 & 0.3 & 577.5 & 0.7\ 3 & 848 & 17 & 828 & 17 & -301.7 & 1.4 & -28.0 & 0.2 & 349.67 & 0.5\ 3.2 & 937 & 14 & 923 & 14 & -184.9 & 0.8 & -16.29 & 0.12 & 215.0 & 0.3\ 3.5 & 1061 & 10 & 1050 & 10 & -88.9 & 0.5 & -7.35 & 0.06 & 106.89 & 0.15\ 3.7 & 1070 & 9 & 1062 & 9 & -55.5 & 0.4 & -4.50 & 0.05 & 67.97 & 0.12\ 4 & 1082 & 8 & 1077 & 8 & -27.9 & 0.2 & -2.14 & 0.02 & 35.14 & 0.07\ 4.2 & 1074 & 7 & 1070 & 7 & -17.79 & 0.16 & -1.352 & 0.017 & 22.83 & 0.05\ 4.5 & 1049 & 6 & 1047 & 6 & -9.156 & 0.09 & -0.663 & 0.008 & 12.18 & 0.03\ 5 & 986 & 5 & 985 & 5 & -3.28 & 0.05 & -0.227 & 0.004 & 4.50 & 0.02\ 6 & 861 & 3 & 861 & 3 & -0.361 & 0.014 & -0.027 & 0.001 & 0.682 & 0.004\ 7 & 746& 2 & 746 & 2 & -0.029 & 0.003 & -0.004 & 0.0003 & 0.115 & 0.001\ 8.5 & 620.7 & 1.6 & 620.7 & 1.6 & & & & & &\ 10 & 532.3 & 0.8 & 532.3 & 0.8 & & & & & &\ 12 & 449.7 & 0.8 & 449.7 & 0.8 & & & & & &\ 13.8033 & 401.0 & 0.4 & 401.0 & 0.4 & & & & & &\ 15 & 375.1 & 0.5 & 375.1 & 0.5 & & & & & &\ 17 & 342.2 & 0.4 & 342.2 & 0.4 & & & & & &\ 18.689 & 321.2 & 0.2 & 321.2 & 0.2 & & & & & &\ 20 & 307.7 & 0.3 & 307.7 & 0.3 & & & & & &\ 24.5561 & 274.2 & 0.2 & 274.2 & 0.2 & & & & & & ![The third virial coefficient of ${}^4$He. The black circles are the results of the present calculations, with error bars representing expanded uncertainties with coverage factor $k=2$. The gray area shows the results of the recent low-temperature experiments by Gaiser and collaborators. [@Gaiser09; @Gaiser10][]{data-label="fig:CHe4"}](C-He4-lowT){width="0.95\linewidth"} We report in Table \[tab:CHe4\] the values of the third virial coefficient of ${}^4$He, together with the various contributions of Eq. (\[eq:C\_full\]), for temperatures in the range from $2.6$ K to $24.5561$ K, which is the lowest temperature studied in our previous work. [@Garberoglio2009b] The same data are plotted in Figure \[fig:CHe4\], where they are compared with the recent experimental measurements by Gaiser and collaborators. [@Gaiser09; @Gaiser10] More extensive comparison with available data over a wide range of temperatures will be presented elsewhere. [@paper2] In Fig. \[fig:CHe4\], our results are plotted with expanded uncertainties with coverage factor $k=2$ as derived in Ref. ; the uncertainty at the same expanded level for the experimental results was estimated from a figure in Ref. . First, we notice that exchange effects are completely negligible in the calculation of the third virial coefficient for temperatures larger than 7 K, where their contribution to the overall value is close to one thousandth of that of the Boltzmann part. This is analogous to what has already been observed for the second virial coefficient. When the temperature is lower than 7 K, the various exchange terms have contributions of similar magnitude and opposite sign, but their overall contribution to $C(T)$ is positive at all the temperatures that have been investigated. The exchange contribution to $C(T)$ is comparable to the statistical uncertainty of the calculation, which progressively increases as the temperature is lowered. In Fig. \[fig:CHe4\], it can be seen that our theoretical values of $C(T)$ are compatible with those of recent experiments [@Gaiser09; @Gaiser10] down to the temperature of 10 K. For lower temperatures, the experimental results are somewhat larger than the calculated values, even though agreement is found again for temperatures below 4 K, where $C(T)$ passes through a maximum. The third virial coefficient of ${}^3$He ---------------------------------------- Similar behavior is observed in the case of the third virial coefficient for ${}^3$He, whose calculated values are reported in Table \[tab:CHe3\]. Also in this case the exchange contributions are of opposite signs, but their combined effect is to reduce the value obtained with Boltzmann statistics, which is the opposite trend to that observed for ${}^4$He. ![The magnitude and sign of the various contributions to $C(T)$ at $T=3$ K.[]{data-label="fig:BarChart"}](BarChart){width="0.95\linewidth"} The effects of the various contributions to the third virial coefficient, in both the Bose–Einstein and Fermi–Dirac case, are summarized in Fig. \[fig:BarChart\] for the representative temperature of $T=3$ K. First, we notice that the largest contribution to the third virial coefficient comes from the Boltzmann term. The even exchange term has only a minor contribution, whereas the two remaining terms ($C_\mathrm{odd}$ and $C_\mathrm B$) have almost equal magnitudes and opposite signs. In the case of Bose–Einstein statistics, the contribution to $C$ from $C_\mathrm{odd}$ is negative, while that from $C_\mathrm B$ is positive; the opposite situation is observed in the case of Fermi–Dirac statistics. The overall sum of the exchange contributions is positive for ${}^4$He and negative in the case of ${}^3$He. The magnitude of each exchange contribution at a given temperature is significantly greater for ${}^3$He; this reflects the larger de Broglie wavelength, which not only appears directly in the exchange terms but also affects the range of space sampled by the ring polymers. In the case of ${}^3$He, the exchange contribution is significantly larger than the uncertainty of our calculations, at least at the lowest temperatures that we have investigated. Similarly to the case of ${}^4$He, quantum statistical effects on $C(T)$ contribute less than one part in a thousand for temperatures higher than 7 K. Even in the case of ${}^3$He, we observe $C(T)$ pass through a maximum, at a temperature around 3 K, which is 1 K lower than the temperature where $C(T)$ reaches a maximum for the ${}^4$He isotope. [d|dd|dd|dd|dd|dd]{} & & & & &\ (K) & & & & &\ 2.6 & 1338 &29 & 1857 &28 & -1803 &4 & -274.8 &0.8 & -2047 &2\ 2.8 & 1477 &24 & 1817 &23 & -1164 &3 & -167.9 &0.6 & -1336.6 &1.3\ 3 & 1480 &17 & 1712 &17 & -760.2 &2.2 & -105.7 &0.4 & -886.4 &0.9\ 3.2 & 1463 &17 & 1621 &17 & -503.5 &1.7 & -66.8 &0.3 & -594.2 &0.6\ 3.5 & 1395 &13 & 1487 &13 & -277.2 &1.2 & -34.89 &0.15 & -333.7 &0.4\ 3.7 & 1376 &11 & 1439 &11 & -189.2 &0.8 & -23.12 &0.11 & -229.6 &0.3\ 4 & 1303 &9 & 1342 &9 & -107.0 &0.5 & -12.69 & 0.07 & -133.43 &0.16\ 4.2 & 1245 &9 & 1273 &9 & -73.9 &0.5 & -8.65 &0.05 & -93.82 &0.13\ 4.5 & 1173 &7 & 1190 &7 & -43.7 &0.3 & -4.86 &0.03 & -55.94 &0.09\ 5 & 1071 &6 & 1079 &6 & -18.41 &0.16 & -1.963 &0.016 & -24.42 &0.05\ 6 & 895 &4 & 897 &4 & -3.56 &0.06 & -0.353 &0.005 & -5.143 &0.013\ 7 & 772 &3 & 773 &3 & -0.78 &0.02 & -0.0784 &0.002 & -1.196 &0.005\ 8.5 & 645 &2 & 645 &2 & -0.059 &0.006 & -0.0087 &0.0003 & -0.155 &0.001\ 10 & 558.3 &1.6 & 558.3 &1.6 & & & & & &\ 12 & 475.5 &1.1 & 475.5 &1.1 & & & & & &\ 13.8033 & 426.2 &0.8 & 426.2 &0.8 & & & & & &\ 15 & 402.0 &0.7 & 402.0 &0.7 & & & & & &\ 17 & 369.6 &0.5 & 369.6 &0.5 & & & & & &\ 18.689 & 347.8 &0.4 & 347.8 &0.4 & & & & & &\ 20 & 333.4 &0.4 & 333.4 &0.4 & & & & & &\ 24.5561 & 297.8 &0.3 & 297.8 &0.3 & & & & & & ![The third virial coefficient of ${}^3$He.[]{data-label="fig:CHe3"}](Helium-3){width="0.95\linewidth"} There are only a few sources of experimental data for $C(T)$ for ${}^3$He. Keller [@Keller55] measured five pressure-volume isotherms at temperatures below 4 K; these were later reanalyzed by Roberts [*et al.*]{} [@Roberts64] and meaningful values of $C$ were obtained only for the two highest temperatures. A later analysis of the Keller data was performed by Steur (unpublished), whose equation for temperatures below 3.8 K was reported by Fellmuth and Schuster [@Fellmuth92]. Some points were also extracted from volumetric data by Karnatsevich [*et al.*]{} [@Karna88] Recently, Gaiser and Fellmuth [@Gaiser08; @Gaiser_pc] extracted virial coefficients from their measurements of two isotherms for ${}^3$He with dielectric-constant gas thermometry. Figure \[fig:CHe3\] compares our calculated values to the available experimental data, where the error bars represent expanded uncertainties with coverage factor $k=2$. Error bars are not drawn for our values above 5 K because they would be smaller than the size of the symbol. As was the case in our previous work, [@Garberoglio2009b] the uncertainty of our values of $C(T)$ is determined by the statistical uncertainty of our Monte Carlo calculations (shown in Tables \[tab:CHe4\] and \[tab:CHe3\]) and by the uncertainty in the two- and three-body potentials. At the temperatures considered here, the statistical uncertainty is the dominant contribution to the overall uncertainty. The full uncertainty analysis is presented elsewhere. [@paper2] For the experimental points, these expanded uncertainties were taken as reported in the original sources; we note that in some cases (notably Ref. ) this appears to be merely the scatter of a fit and therefore underestimates the total uncertainty. Our results are qualitatively similar to the rather scattered experimental data. We are quantitatively consistent with the values based on analysis of the data of Keller, but values from the other experimental sources are more positive than our results. We note that a similar comparison for ${}^4$He [@paper2], where the experimental data situation is much better, shows the $C(T)$ values of Ref.  for ${}^4$He to deviate in a very similar way not only from our results but from other experimental data we consider to be reliable. Conclusions =========== We used path-integral methods to derive an expression for the third virial coefficient of monatomic gases, including the effect of quantum statistics. We applied this formalism to the case of helium isotopes, using state-of-the-art two- and three-body potentials. We showed that exchange effects make no significant contribution to the third virial coefficient above a temperature of approximately 7 K for both the fermionic and bosonic isotope. This is the same behavior observed in the calculation of the second virial coefficient. For temperatures lower than $7$ K, the sign of the contribution to $C(T)$ from exchange effects depends on the bosonic or fermionic nature of the atom. In the case of ${}^4$He, the exchange contribution to $C(T)$ increases its value compared to the value obtained with Boltzmann statistics, although in our simulations the total exchange contribution has the same order of magnitude as the statistical uncertainty of the PIMC integration. In the case of ${}^3$He, the exchange contribution is negative, and its magnitude is much larger than the statistical uncertainty. The range of temperatures that we have investigated covers the low-temperature maximum of $C(T)$ for both isotopes. The third virial coefficient of ${}^4$He reaches its maximum close to $4$ K, whereas in the case of ${}^3$He the maximum is attained at a lower temperature. For both helium isotopes, the uncertainty in our calculated third virial coefficients is much smaller than that of the limited and sometimes inconsistent experimental data. For ${}^4$He, we obtain good agreement with the most recent experimental results, except for some temperatures below 10 K. A full comparison with available experimental data for ${}^4$He, including the higher temperatures of importance for metrology, will be presented elsewhere.[@paper2] For ${}^3$He, we are qualitatively consistent with the sparse and scattered experimental values; in this case especially our calculations provide results that are much less uncertain than experiment. In both cases, at the temperatures considered here, the uncertainty is dominated by the statistical uncertainty of the Monte Carlo integration, meaning that the uncertainty of $C(T)$ could be reduced somewhat with greater expenditure of computer resources. We note two directions in which extension of the present work could be fruitful. One is the calculation of higher-order virial coefficients, which is a straightforward extension of the method presented here. This would be much more computationally demanding, but the fourth virial coefficient $D(T)$ may be feasible, at least at higher temperatures where the number of beads in the ring polymers would not be large. Second, the method can be extended to calculate temperature derivatives such as $\mathrm dC/ \mathrm dT$; such derivatives are of interest in interpreting acoustic measurements. Work on the evaluation of acoustic virial coefficients is in progress. [@paper2] We thank C. Gaiser for providing information on low-temperature data for $C(T)$ of helium isotopes, and M. R. Moldover and J. B. Mehl for helpful discussions on various aspects of this work. The calculations were performed on the KORE computing cluster at Fondazione Bruno Kessler. [^1]: Partial contribution of the National Institute of Standards and Technology, not subject to copyright in the United States.

ArXiv

--- abstract: 'We compute analytic expressions for the non-linearity parameters characterizing the bi- and tri-spectrum of primordial curvature perturbations generated during an inflationary epoch of the early universe driven by an arbitrary number fields. We assume neither slow roll nor a separable potential; instead, to compute Non-Gaussianities, we assume a separable Hubble parameter. We apply the formalism to an exact solvable toy-model and show under which conditions observably large non-Gaussianities are produced.' author: - 'Diana Battefeld$^{1,2)}$' - 'Thorsten Battefeld$^{2)}$' title: | On Non-Gaussianities in Multi-Field Inflation (N fields):\ Bi- and Tri-spectra beyond Slow-Roll --- Introduction ============ The spectrum of temperature fluctuations in the cosmic microwave background radiation (CMBR) is commonly accepted to have been generated by the amplification of vacuum fluctuations in light fields during an inflationary epoch of the early universe. Observations of the CMBR [@COBE; @Spergel:2006hy; @Maxima; @Boomerang; @Komatsu:2008hk] are in agreement with the prediction based on the simplest, single-field inflationary models of a nearly scale invariant, highly Gaussian spectrum of adiabatic cosmological perturbations, see i.e. [@Lyth:1998xn; @Bassett:2005xm] for reviews; the validity of the inflationary framework has withstood the tests of time since its inception in $1980$. Our interest in density perturbations, commonly characterized by the comoving curvature perturbation $\zeta$, stems from the fact that one can differentiate between models by considering higher order correlation functions, which will be probed in upcoming experiments [@Komatsu:2009kd], such as PLANCK [@Planck]. The non linearity parameters $f_{NL}$ [@Komatsu:2001rj], characterizing the three-point correlation function, and $\tau_{NL}$ [@Seery:2006js; @Byrnes:2006vq], characterizing the four-point correlation function, are widely used to estimate non-Gaussianities (NG). Even though primordial NG, i.e. a non-zero three-point function, have not yet been observed, analysis of the 5-year WMAP data provides strong indications for their presence, $-9<f_{NL}^{local}<111$ at $95\%$CL [@Komatsu:2008hk] or $-4 < f_{NL}^{local} < 80$ at $95\%$ CL in the re-analysis by K. Smith et.al. [@Smith:2009jr]. The recently launched PLANCK satellite [@Planck] will be sensitive enough to detect NG in the next years with $\Delta f_{NL}\sim \pm 5$, if they are indeed present at the currently expected level; further, bounds on $\tau_{NL}$ will be considerably improved. Simple single-field models have typically tiny non-Gaussianities [@Maldacena:2002vr; @Acquaviva:2002ud; @Creminelli:2003iq; @Babich:2004gb; @Bartolo:2004if; @Seery:2005wm; @Seery:2008ax]. However, more complicated models warrant the possibility of allowing for larger NG within the inflationary framework; examples are features in the potential [@Chen:2006xjb; @Chen:2008wn], preheating [@Enqvist:2004ey; @Jokinen:2005by; @Barnaby:2006cq; @Barnaby:2006km], modulated preheating [@Dvali:2003em; @Battefeld:2007st; @Suyama:2007bg; @Byrnes:2008zz], the presence of additional light fields, and thus isocurvature perturbations, as in the curvaton-scenario [@Linde:1996gt; @Bartolo:2001cw; @Bernardeau:2002jf; @Linde:2005yw; @Boubekeur:2005fj; @Sasaki:2006kq; @Malik:2006pm; @Kawasaki:2008sn; @Huang:2008zj; @Hikage:2008sk; @Langlois:2008vk; @Li:2008fma; @Ichikawa:2008iq; @Dutta:2008if], trapped fields [@Suyama:2008nt], DBI-inflation/k-inflation [@Alishahiha:2004eh; @Chen:2006nt; @Huang:2006eh; @Arroja:2008ga; @Arroja:2008yy; @Gao:2009gd; @Cai:2009hw; @Khoury:2008wj; @Arroja:2009pd; @Chen:2009bc; @Mizuno:2009mv; @Gao:2009at; @Mizuno:2009cv; @Langlois:2009ej; @Langlois:2008qf; @Langlois:2008wt; @RenauxPetel:2009sj], nonlocal inflation [@Barnaby:2007yb; @Barnaby:2008fk], among others. Alternative proposals, such as the new-ekpyrotic scenario [@Buchbinder:2007ad; @Creminelli:2007aq; @Lehners:2007ac], have usually quite strong NG signals too [@Koyama:2007if; @Buchbinder:2007at; @Khoury:2008wj; @Lehners:2009ja; @Lehners:2007wc; @Lehners:2008my], see [@Lehners:2008vx] for a review. In this article, we further investigate the possibility to generate NG in multi-field inflationary models. The generation of NG in multi-field models was considered in [@Bernardeau:2002jy; @Alabidi:2005qi; @Lyth:2005fi; @Lyth:2005du; @Seery:2005gb; @Battefeld:2006sz; @Seery:2006vu; @Byrnes:2007tm; @Byrnes:2006vq; @Yokoyama:2007uu; @Yokoyama:2007dw] (see also [@Salem:2005nd; @Lyth:2005qk; @Alabidi:2006wa; @Alabidi:2006hg; @Sasaki:2008uc; @Byrnes:2008zy; @Byrnes:2008wi; @Naruko:2008sq] for hybrid/multi-brid models, and [@Huang:2009xa; @Gao:2008dt; @Battefeld:2007en; @Choi:2007su; @Misra:2007cq; @Misra:2008tx] for related work), where it was found that the non-linearity parameters are usually slow roll suppressed[^1]. The presence of isocurvature modes in multi-field models [@Gordon:2000hv; @Malik:2005cy; @Malik:2006ir] can cause Fourier-modes of the curvature perturbation $\zeta_k$ to evolve even after horizon exit, provided that the trajectory in field space is curved. Hence, NG can be sourced [@Bernardeau:2002jy], but a sharp turn, and thus violation of the slow-roll conditions, is usually required to give rise to large NG. A computation of $f_{NL}$ involving two fields, a separable potential, and the slow-roll approximation was made by Vernizzi and Wands [@Vernizzi:2006ve], which was later extended to an arbitrary number of fields in [@Battefeld:2006sz] and the tri-spectrum ($\tau_{NL}$) in [@Seery:2006vu]. However, slow roll can be temporarily violated during inflation, i.e. if a bump in the potential is encountered, if fields start to decay during inflation as in staggered/cascade inflation [@Becker:2005sg; @Ashoorioon:2006wc; @Battefeld:2008py; @Battefeld:2008ur; @Battefeld:2008qg], and it is necessarily violated at the end of inflation and during (p)re-heating. Unfortunately, almost all analytic studies based on the (non-linear) $\delta N$-formalism [@Lyth:2004gb; @Lyth:2005fi][^2] are based on the slow roll approximation, rendering them unapplicable in these cases (see however [@Yokoyama:2007dw]); other computational techniques that do not involve the slow roll approximation are much more involved and analytic results for higher order correlation functions become unfeasible. Recently, Byrnes and Tasinato [@Byrnes:2009qy] found a class of inflationary models that are amendable to an analytic computation of NG within the $\delta N$-formalism, without imposing the slow-roll or the horizon-crossing approximation, but requiring a separable Hubble parameter $H=\sum_k H_k(\varphi_k)$. They focused on two-field models with canonical kinetic terms and computed an expression for $f_{NL}$. In this paper, we generalize this framework to an arbitrary number of fields and provide expressions for $f_{NL}$ and $\tau_{NL}$. We demonstrate the applicability of the formalism in a concrete toy model with up to six inflaton fields that allows for the generation of large NG towards the end of inflation. Due to the assumption of a separable Hubble parameter, the potential necessarily contains cross couplings between the inflatons. Thus, simple models of assisted inflation [@Liddle:1998jc; @Malik:1998gy; @Kanti:1999vt; @Kanti:1999ie] do not fall into the class of models amendable to our treatment. However, the absence of cross couplings does in fact constitute fine tuning, and more realistic models of multi-field inflation in string theory are closer to the ones we consider. Furthermore, the presence of cross couplings should render the framework at hand applicable to an analytic computation of NG during (p)reheating, a project we plan to come back to in the future. The concrete outline of this article is as follows: After introducing the setup in Sec. \[sec:setup\], we review briefly the $\delta N$-formalism in Sec. \[sec:deltaN\] and show how expressions for the power-spectrum, the bi-spectrum and the tri-spectrum may be obtained, Sec. \[sec:power\]-\[sec:taunl\]. The needed derivatives of the volume expansion rate are computed in Sec. \[sec:derivativesofn\], which are then used to provide analytic expressions for the power-spectrum, $f_{NL}$ and $\tau_{NL}$ in Sec. \[sec:analyticpower\]-\[sec:analytictaunl\]. These are the main results of this paper. After a short discussion in Sec. \[sec:discussion\], we provide a concrete example in Sec. \[sec:examples\]: a two-field model as introduced in [@Byrnes:2009qy] (Sec. \[caseN=2\]), and a case with up to six inflatons (Sec. \[casengeneral\]). We conclude in Sec. \[sec:conclusion\]. The Setup \[sec:setup\] ======================= The statistical information about primordial perturbations can be extracted from correlation functions of the curvature perturbation $\zeta$, which are imprinted onto the temperature fluctuations of the CMBR after inflation. Our goal is to derive expressions for the nonlinearity parameters $f_{NL}$ and $\tau_{NL}$ that characterize the magnitude of the three- and four-point correlation functions (bi- and tri-spectrum) of $\zeta$ by use of the (non-linear) $\delta N$-formalism [@Lyth:2004gb; @Lyth:2005fi] in multi-field inflationary models, with an arbitrary number of fields, without imposing slow roll, but assuming a separable Hubble parameter [@Salopek:1990jq] $$\begin{aligned} H=\sum_{k=1}^{\mathcal{N}}H_k(\varphi_{k})\,. \label{seperableH}\end{aligned}$$ As such, our treatment extends the work of Byrnes and Tasinato [@Byrnes:2009qy] who computed $f_{NL}$ under the same assumption in two-field models. Technically, [@Byrnes:2009qy] parallels the work of Vernizzi and Wands [@Vernizzi:2006ve], who computed a general expression for $f_{NL}$ for two fields by assuming slow roll and a separable potential $V=\sum_k V_k(\varphi_k)$. The latter work was extended in [@Battefeld:2006sz] to an arbitrary number of fields and in [@Seery:2005gb] to the tri-spectrum. Here, we follow [@Battefeld:2006sz; @Seery:2005gb] closely. We focus on inflation driven by $\mathcal N$ scalar fields with the action $$\begin{aligned} S=\frac{1}{2}\int d^4x\sqrt{-g}\left(\frac{1}{2}\sum_{k=1}^{\mathcal{N}}\partial^{\mu}\varphi_k\partial_{\mu}\varphi_k+V(\varphi_1,\varphi_2,...)\right)\,.\end{aligned}$$ Here and in the following we set the reduced Planck mass equal to one, $m_p=(8\pi G)^{-1/2}\equiv 1$; all sums run from $1$ to $\mathcal{N}$ unless specified otherwise. Focussing on a homogeneous universe, the Friedmann and Klein-Gordon equations can be written as a first order Hamilton-Jacobi system [@Salopek:1990jq] $$\begin{aligned} H^2&=&\frac{1}{3}V+\frac{2}{3}\sum_k \left(\frac{\partial H}{\partial \varphi_k}\right)^2\,,\label{HJ1}\\ \dot{\varphi}_k&=&-2\frac{\partial H}{\partial \varphi_k}\,,\label{HJ2}\end{aligned}$$ where a dot denotes a derivative with respect to cosmic time and $H=\dot{a}/a$. These equations are exact, namely, they do not rely on the slow roll approximation. Because (\[HJ2\]) decouples if the Hubble parameter is separable, we specifically focus on this case. Inflation takes place as long as the Hubble slow-evolution parameter remains small, $$\begin{aligned} \epsilon\equiv -\frac{\dot{H}}{H^2} \ll 1\,.\end{aligned}$$ In analogy to the slow roll parameters, it is useful to define $$\begin{aligned} \delta_k&\equiv &\left(\frac{H_{,k}}{H}\right)^2\,,\label{correspondingslowroll1}\label{deltak}\\ \gamma_k&\equiv &\frac{H_{,kk}}{H}\,,\label{correspondingslowroll2}\label{gammak}\\ \beta_k&\equiv &\frac{H_{,kkk}}{H}\frac{H_{,k}}{H}\,,\label{correspondingslowroll3}\label{betak}\end{aligned}$$ where we used the short hand notation $H_{,k}\equiv \partial H/\partial\varphi_k$. Note the absence of any mixed derivatives due to the Ansatz (\[seperableH\]). The universe inflates as long as $$\begin{aligned} \delta\equiv \sum_{k}\delta_k \ll \frac{1}{2}\,,\label{correspondingslowroll4}\end{aligned}$$ because $\epsilon =2\delta$; however, neither $\gamma_k$ nor $\beta_k$ are required to be small for inflation to take place. The $\delta N$-formalism \[sec:deltaN\] ======================================= The $\delta N$-formalism goes back to Starobinski [@Starobinski], was extended by Sasaki and Stewart in [@Sasaki:1995aw] and generalized to higher orders in [@Lyth:2004gb; @Lyth:2005fi] (see also [@Vernizzi:2006ve; @Seery:2005gb; @Allen:2005ye] for related work [^3]). To employ the $\delta N$-formalism and compute correlation functions of $\zeta$, we need to evaluate the unperturbed volume expansion rate from an initially flat hypersurface at $t_*$ to a final uniform density hypersurface at $t_c$ $$\begin{aligned} N(t_c,t_*)\equiv\int^{t_c}_{t_*} H\,dt\,.\end{aligned}$$ In the following, all integrals are assumed to run over values from $t_*$ to $t_c$ if not specified otherwise. Since different trajectories in field space can provide the same homogeneous expansion rate, we define, in analogy to the $\mathcal{N}-1$ integrals of motion in slow roll inflation [@Battefeld:2006sz], $$\begin{aligned} C_k\equiv -\int\frac{d\varphi_k}{H_{,k}}+\int{\frac{d\varphi_{k+1}}{H_{,{k+1}}}}\,, \label{C}\end{aligned}$$ for $k=1\dots \mathcal{N}-1$. These quantities are conserved during inflation and can be used to discriminate between different trajectories [^4]. The perturbation of the expansion rate, $\delta N$, is identical to the curvature perturbation $\zeta$, $$\begin{aligned} \zeta&=&\delta N =\sum_kN_{,k}\delta\varphi_{k}^*+\frac{1}{2}\sum_{kl}N_{,kl}\delta\varphi_{,k}^*\delta\varphi_{l}^*+...\,,\label{zeta} \end{aligned}$$ which is valid to any order in perturbation theory [@Lyth:2004gb; @Lyth:2005fi]. If a quantity is to be evaluated at $t_*$ or $t_c$ we denote it by a superscript, i.e. $H_{,k}^*=\partial H/\partial\varphi_k|_{t_*}$ [^5]. For simplicity we use the short hand notation $$\begin{aligned} N_{,k}&\equiv&\frac{\partial N}{\partial\varphi_{k}^*}\,, \label{def:derivativeofN}\\ N_{,kl}&\equiv&\frac{\partial^2 N}{\partial\varphi_{k}^*\partial\varphi_{l}^*}\,,\\ \nonumber &\vdots&\end{aligned}$$ With the Ansatz (\[seperableH\]), we can simplify the volume expansion rate to [@Byrnes:2009qy] $$\begin{aligned} N(t_c,t_*)=-\frac{1}{2}\sum_{k=1}^{\mathcal N}\int_{\varphi_k^*}^{\varphi_k^c}\frac{H_k}{H_{,k}}d\varphi_k\,. \label{N}\end{aligned}$$ Note that we did not assume slow roll, but $H$ needs to be separable in order for (\[N\]) to hold [^6]. The power-spectrum \[sec:power\] -------------------------------- The power-spectrum of $\zeta$ is defined in terms of the two-point correlation function as $$\begin{aligned} {<\zeta_{\bf{k_1}}\zeta_{\bf{k_2}}>}&\equiv&{(2\pi)^3\delta^3(\bf{k_1}+\bf{k_2})}\frac{2\pi^2}{k_1^3}\mathcal P_{\zeta}(k_{1})\,,\end{aligned}$$ where $\zeta_{\bf{k}}$ denotes a Fourier mode of the curvature perturbation. Using (\[N\]), the power-spectrum can be related to the derivatives of the volume expansion rate, $$\begin{aligned} \mathcal{P}_{\zeta}=\sum_{k} N^2_{,k}\mathcal P^*\,,\label{powersp}\end{aligned}$$ and the scalar spectral index and the tensor to scalar ratio become [@Vernizzi:2006ve] $$\begin{aligned} n_{s}-1&\equiv& \frac{\partial \ln \mathcal{P}_\zeta}{\partial \ln k}=-2\epsilon^*+\frac{2}{H^*}\frac{\sum_{kl=1}^{\mathcal N}\dot{\varphi}_{k}^*N_{,kl}N_{,l}}{\sum_{m=1}^{\mathcal N}N^2_{,m}}\,,\label{ns}\\ r&\equiv& \frac{\mathcal{P}_T}{\mathcal{P}_\zeta}=\frac{8\mathcal{P}^*}{\mathcal{P}_\zeta}\,,\end{aligned}$$ where $\epsilon^*\equiv-(\dot{H}/H^2)^*$ and $\mathcal{P}^*=k^3P^*(k)/(2\pi^2)=(H^*)^{2}/(4\pi^2)$; these have to conform to the COBE normalization $\mathcal{P}_\zeta=(2.41\pm 0.11) \times 10^{-9}$ [@Komatsu:2008hk] as well as $n_s= 0.960 \pm 0.013$ from the WMAP $5$-year data analysis [@Komatsu:2008hk] and $r<0.22\; (95\%\;\mbox{CL})$ combining WMAP5, BAO and SN data [@Komatsu:2008hk]. In order to compute these observable quantities, we need to evaluate the derivatives of $N$ with respect to $\varphi_k^*$, see Sec. \[sec:derivativesofn\]. The bi-spectrum and $f_{NL}$ \[sec:fnl\] ---------------------------------------- The bi-spectrum of $\zeta$ is defined in terms of the three-point correlation function as $$\begin{aligned} {<\zeta_{\bf{k_1}}\zeta_{\bf{k_2}}\zeta_{\bf{k_3}}>}&\equiv&{(2\pi)^3\delta^3(\bf{k_1}+\bf{k_2}+\bf{k_3})}\mathcal B_{\zeta}(k_{1},k_{2},k_{3})\,.\end{aligned}$$ A measure of its magnitude is the non-linearity parameter $f_{NL}$, defined as[^7] $$\begin{aligned} -\frac{6}{5}f_{NL}&\equiv& \frac{k_1^3k_2^3k_3^3}{k_1^3+k_2^3+k_3^3}\frac{\mathcal{B}_\zeta}{4\pi^2 \mathcal{P}_\zeta^2} \,.\end{aligned}$$ Using (\[N\]), one can relate the non-linearity parameter $f_{NL}$ to the derivatives of the expansion rate $N$ with respect to the field values as [@Vernizzi:2006ve; @Seery:2005gb], $$\begin{aligned} -\frac{6}{5}f_{NL}&=&\frac{r}{16}(1+f)+\frac{\sum_{kl}N_{,k}N_{,l}N_{,kl}}{(\sum_{k}N^2_{,k})^2}\label{fnl}\\ &\equiv & \frac{r}{16}(1+f)-\frac{6}{5}f_{NL}^{(4)}\,.\label{r}\end{aligned}$$ Here, $f$ includes the contingency of $f_{NL}$ on the shape of the momentum triangle, $0\leq f\leq 5/6$ [@Maldacena:2002vr; @Vernizzi:2006ve]; $f$ is maximal for an equilateral triangle and minimal for a triangle where two sides are much longer than the third [@Maldacena:2002vr]. Since the tensor:scalar ratio $r$ is smaller than one, it is evident that the first term in (\[fnl\]) is suppressed. The WMAP$5$ data [@Komatsu:2008hk] currently yields the bound $-9<f_{NL}^{local}<111$ at $95\%$CL or $-4 < f_{NL}^{local} < 80$ at $95\%$ CL in the re-analysis by K. Smith et.al. [@Smith:2009jr], which will be improved considerably in the near future by PLANCK [@Planck]. Nevertheless, unless $f_{NL}>1$, non-Gaussianities are unlikely to be ever detected. For this reason, from here on, we focus on the second term in (\[fnl\]), which provides a momentum independent contribution to $f_{NL}$. The tri-spectrum and $\tau_{NL}$\[sec:taunl\] --------------------------------------------- The tri-spectrum of $\zeta$ is defined in terms of the four-point correlation function as $$\begin{aligned} {<\zeta_{\bf{k_1}}\zeta_{\bf{k_2}}\zeta_{\bf{k_3}}\zeta_{\bf{k_4}}>}&\equiv&{(2\pi)^3\delta^3(\bf{k_1}+\bf{k_2}+\bf{k_3}+\bf{k_4})}T_{\zeta}({\bf{k}}_1,{\bf{k}}_2,{\bf{k}}_3,{\bf{k}}_4)\,,\end{aligned}$$ where only the connected part of the correlator is considered [@Okamoto:2002ik; @Kogo:2006kh]. To estimate the magnitude of $T_{\zeta}$, it is common to define $\tau_{NL}$ by $$\begin{aligned} T_\zeta\equiv\frac{1}{2}\tau_{NL}\bigg(\sum_{\alpha} N^2_{,\alpha}\bigg)^3[P^*(k_1)P^*(k_2)P^*(k_{14})+23 \mbox{ permutations}]\,,\label{T}\end{aligned}$$ where ${\bf{k}}_{ij}={\bf{k}}_{i}+{\bf{k}}_{j}$. Defining $$\begin{aligned} T\equiv\sum_{i<j}\sum_{s\neq ij}k_i^3k_j^3\left(k_{is}^{-3}+k_{js}^{-3}\right)\,,\end{aligned}$$ the momentum dependence in (\[T\]) can be made explicit [@Seery:2006vu] $$\begin{aligned} T_{\zeta}=\frac{4\pi^6}{\Pi_ik^3_i}\tau_{NL}\bigg(\sum_{k}(N^*_{,k })^2\bigg)^3(\mathcal{P}^*)^3 T\,.\end{aligned}$$ $\tau_{NL}$ contains four tree-level components $\Delta \tau_{NL}^{(i)}$, $i=1\dots 4$. One component can be shown to be bounded from above by the tensor:scalar ratio [@Seery:2006vu], $|\Delta\tau_{NL}^{(1)}|\lesssim r/50$, which is too small to be observable and thus of no interest to us. The remaining three contributions are [^8] [@Seery:2005gb] (see also [@Byrnes:2006vq; @Alabidi:2005qi]) $$\begin{aligned} \Delta \tau_{NL}^{(2)}&=&\frac{\sum_{kl}\dot{\varphi}_{k}^*N_{,kl}N_{,l}}{\left(\sum_lN^2_{,l}\right)^2}\frac{\mathcal K}{4H^*}\,,\label{tau2}\\ \Delta \tau_{NL}^{(3)}&=&\frac{\sum_{klm}N_{,kl}N_{,ml}N_{,k}N_{,m}}{\left(\sum_l N^2_{,l}\right)^3}\,,\label{tau3}\\ \Delta \tau_{NL}^{(4)}&=&2\frac{\sum_{klm}N_{,klm}N_{,k}N_{,l}N_{,m}}{\left(\sum_lN^2_{,l}\right)^3}\frac{\sum_ik^3_i}{T}\,,\label{tau4}\end{aligned}$$ where $\mathcal K$ is defined as $$\begin{aligned} \mathcal K=\frac{1}{T}\sum_{perms}\frac{k_1^3}{k^3_{12}}\mathcal M(k_{12},k_{3},k_{4})\,,\end{aligned}$$ and $$\begin{aligned} \mathcal M(k_1,k_2,k_3)\equiv-k_1k_2^2-4\frac{k^2_2k^2_3}{k_1+k_2+k_3+k_4}+\frac{1}{2}k_1^3+\frac{k_2^2k_3^2(k_2-k_3)}{(k_1+k_2+k_3+k_4)^2}\,.\end{aligned}$$ The Derivatives of $N$ \[sec:derivativesofn\] ============================================= To evaluate the non-linearty parameters $f_{NL}$ and $\tau_{NL}$ we need to compute the derivatives of the volume expansion rate with respect to the fields. In this section, we follow [@Battefeld:2006sz] and [@Seery:2006vu] closely. Based on (\[N\]), the total differential of $N$ reads $$\begin{aligned} dN=\sum_{k=1}^{\mathcal N}\frac{1}{2}\left[\left(\frac{H_{k}}{H_{,{k}}}\right)^*-\sum_{l=1}^{\mathcal{N}}\frac{\partial\varphi_{l}^c}{\partial\varphi_{k}^*}\left(\frac{H_{l}}{H_{,{l}}}\right)^c\right]d\varphi_{k}^*\,.\label{dN}\end{aligned}$$ Using the integrals of motion $C_k$ in (\[C\]), we can relate $d\varphi_{k}^c$ and $d\varphi_{k}^*$ $$\begin{aligned} d\varphi_{m}^c=\sum^{\mathcal N-1}_{l=1}\frac{\partial\varphi_{m}^c}{\partial C_l}\left(\sum_{k=1}^{\mathcal N}\frac{\partial C_l}{\partial\varphi_{k}^*}d\varphi_{k}^*\right)\,,\label{dphij}\end{aligned}$$ where $$\begin{aligned} \frac{\partial C_l}{\partial\varphi_{k}^*}=\frac{1}{H_{,{k}}^*}\left(\delta_{lk-1}-\delta_{lk}\right)\,.\label{delta}\end{aligned}$$ We want to evaluate NG at $t_c$ where the condition $\rho=\mbox{const}$ holds; in order to evaluate this condition, we first eliminate $\partial\varphi_{j}^c/\partial C_i$ in favor of $\partial\varphi_{1}^c/\partial C_i$. Hence we define $$\begin{aligned} \tilde{C_k}&\equiv&\sum^{k-1}_{l=1}C_l\\ &=&-\int\frac{d\varphi_1}{H_{,1}}+\int\frac{d\varphi_k}{H_{,k}}\,,\end{aligned}$$ and differentiate with respect to $C_k$, $$\begin{aligned} \frac{\partial\tilde C_l}{\partial C_k}=-\frac{\partial\varphi_{1}^c}{\partial C_k}\frac{1}{H_{\prime{1}}^c}+\frac{\partial\varphi_{l}^c}{\partial C_k}\frac{1}{H_{,{l}}^c}\,,\end{aligned}$$ so that $$\begin{aligned} \frac{\partial\varphi_{l}^c}{\partial C_k}=\frac{H_{,l}^c}{H_{,{1}}^c}\frac{\partial\varphi_{1}^c}{\partial C_k}+H_{,{1}}^c\Theta_{kl}\,,\end{aligned}$$ where we defined $$\begin{aligned} \mathbf{\Theta}_{kl}= \left\{\begin{array}{ll} 1,&\textrm{if k} \leq l-1\\ 0,&\textrm{if k} >l-1\,. \end{array}\right.\end{aligned}$$ Inserting this into the derivative (with respect to $C_k$) of the $\rho=\mbox{const}$ condition, that is into $$\begin{aligned} 0=\sum_{l=1}^{\mathcal N}H_{,{l}}^c\frac{\partial \varphi_{l}^c}{\partial C_k}\,,\end{aligned}$$ we obtain $$\begin{aligned} \frac{\partial\varphi_{l}^c}{\partial C_k}=-\left[H_{,l}\frac{\sum^{\mathcal N}_{m=k+1}H_{,m}^2}{\sum_{m=1}^{\mathcal N}H_{,{m}}^2}\right]^c+H_{,l}^c\Theta_{kl}\,.\label{thetaeqn}\end{aligned}$$ Using (\[thetaeqn\]), (\[dphij\]) and (\[delta\]) we arrive at $$\begin{aligned} \frac{\partial\varphi_{l}^c}{\partial\varphi^*_{k}}&=&-\frac{H_{,{l}}^c}{H_{,{k}}^*}\left[\frac{H_{,{k}}^2}{\sum_{m=1}^{\mathcal N}H_{,m}^2}-\delta_{lk}\right]^c\\ &=&-\frac{H^c}{H^*}\sqrt{\frac{\delta_{l}^c}{\delta_{k}^*}}\left[\frac{\delta_{k}^c}{\delta^c}-\delta_{lk}\right]\,. \label{diffphicphi*}\end{aligned}$$ To write the derivatives of $N$, it is convenient to define $$\begin{aligned} E_{km}&\equiv&\frac{\delta_k^c}{\delta^c}-\delta_{mk}\,,\label{E}\\ Y_k&\equiv&\delta_k^c\left(1-\frac{2\gamma_k^c}{\delta^c}\right)\,,\label{Y}\\ X_k&\equiv&2\left[\gamma_k^cY_k-\frac{\delta_k^c\beta_k^c}{\delta^c}\right]\,,\label{X}\end{aligned}$$ with $\delta_k$, $\gamma_k$ and $\beta_k$ from (\[deltak\])-(\[betak\]). Differentiating $\delta_k$ and $\beta_k$ yields $$\begin{aligned} \frac{\partial\delta_{l}^c}{\partial\varphi_{m}^c}&=&2\left(\gamma_{l}\delta_{lm}-\sqrt{\delta_{l}\delta_{m}}\right)^c\sqrt{\delta_{l}^c}\,,\label{der1}\\ \frac{\partial\delta^c}{\partial\varphi_{m}^c}&=&2\sqrt{\delta_{m}^c}\left(\gamma_{m}-\delta\right)^c\,,\label{der2}\\ \frac{\partial\gamma_{m}^c}{\partial\varphi_{l}^c}&=&\sqrt{\delta_{l}^c}\left(\delta_{ml}\frac{\beta_{m}}{\delta_{m}}-\gamma_{m}\right)^c\,.\label{der3}\end{aligned}$$ Further, the derivatives of $E_{km}$ and $Y_{k}$ read $$\begin{aligned} \frac{\partial E_{km}}{\partial \varphi_{l}^c}&=&-2\left(\frac{\sqrt{\delta_{l}}\gamma_{l}}{\delta}\right)^c E_{kl}\,,\\ \frac{\partial Y_{k}}{\partial \varphi_{l}^c}&=&\frac{\delta_{kl}}{\sqrt{\delta_{l}^c}}X_{l}-\frac{2}{\sqrt{\delta_{l}^c}}\left(\gamma_{k}\frac{\delta_{k}}{\delta}Y_{l}+\delta_{l}Y_{k}\right)^c\,,\end{aligned}$$ where we have used equations (\[E\])-(\[X\]) and (\[der1\])-(\[der3\]). By means of equation (\[dN\]), along with (\[diffphicphi\*\]) and the above, we obtain the first derivatives of the expansion rate with respect to $\varphi_k^*$ as $$\begin{aligned} N_{,k}=\frac{1}{2}\frac{u_k}{\sqrt{\delta_{k}^*}}\,, \label{firstN}\end{aligned}$$ where we introduced $$\begin{aligned} u_k&\equiv&\frac{H_{k}^*+Z_{k}^c}{H^*}\,,\\ Z_{k}^c&\equiv& H^c\frac{\delta_k^c}{\delta^c}-H_k^c\,.\end{aligned}$$ The second derivatives of $N$ become $$\begin{aligned} N_{,kl}=\frac{1}{2}\left[\delta_{kl}\left(1-\frac{\gamma_{k}^*}{\delta_{k}^*}u_k\right)+\frac{1}{\sqrt{\delta_{k}^*}}\frac{1}{H^*}\frac{\partial Z_{k}^c}{\partial \varphi_{l}^*}\right]\,,\label{secondN}\end{aligned}$$ where $$\begin{aligned} \frac{\partial Z_{k}^c}{\partial\varphi_{l}^*}&=&\frac{H^*}{\sqrt{\delta_{l}^*}}A_{kl}\,,\label{derZkc}\end{aligned}$$ and we have defined the symmetric matrix, $$\begin{aligned} A_{kl}&\equiv &-\left(\frac{H^c}{H^{*}}\right)^2\bigg(\sum_{m}E_{km}E_{lm}Y_{m}\bigg)^c\,.\end{aligned}$$ The third derivatives become, $$\begin{aligned} N_{,klm}&=&\frac{\delta_{klm}}{2\sqrt{\delta_{k}^*}}\left[\frac{2(\gamma^*_k)^{2}-\beta_{k}^*}{\delta_{k}^*}u_k-\gamma_{k}^*\right]-\frac{\gamma_{k}^*}{2\delta_{k}^*H^*}\left(\delta_{km}\frac{\partial Z_{k}^c}{\partial\varphi_{l}^*}+\delta_{kl}\frac{\partial Z_{k}^c}{\partial\varphi_{m}^*}\right)\\ \nonumber &&+\frac{1}{2}\frac{1}{\sqrt{\delta_{k}^*}H^*}\frac{\partial^2 Z_{k}^c}{\partial\varphi_{l}^*\partial\varphi_{m}^*}\,,\label{thirdN}\end{aligned}$$ where $$\begin{aligned} \frac{\partial^2Z_{k}^c}{\partial\varphi_{l}^*\partial\varphi_{m}^*}&=&-\delta_{lm}\frac{\gamma_{m}^*}{\delta_{m}^*}H^*A_{kl}+\frac{H^*}{\sqrt{\delta_{m}^*\delta_{l}^*}}A_{klm}\,,\end{aligned}$$ and we defined the totally symmetric tensor $$\begin{aligned} A_{klm}&\equiv&\left(\frac{H^{c}}{H^*}\right)^3\sum_{\alpha=1}^{\mathcal{N}}\Bigg[X_{\alpha}E_{k\alpha}E_{l\alpha}E_{m\alpha}+\sum_{\beta=1}^{\mathcal{N}}Y_{\beta}(Y_{\alpha}-\delta_{\alpha}^c)\\ \nonumber &&\times(E_{k\beta}E_{l\alpha}E_{m\alpha}+E_{k\alpha}E_{l\beta}E_{m\alpha}+E_{k\alpha}E_{l\alpha}E_{m\beta})\Bigg]\,.\end{aligned}$$ Analytic Expressions \[sec:analyticexpressions\] ================================================ Equipped with $N_{,k}$ from (\[firstN\]), $N_{,kl}$ from (\[secondN\]) and $N_{,klm}$ from (\[thirdN\]), we are now ready to compute observables such as the scalar spectral index as well as the momentum independent components of the non-linearity parameters $f_{NL}$ and $\tau_{NL}$. The power-spectrum \[sec:analyticpower\] ---------------------------------------- Using the derivatives of the expansion rate $N_{,k}$ and $N_{,kl}$ we can simplify the power-spectrum in (\[powersp\]) to $$\begin{aligned} \mathcal P_{\zeta}=\frac{1}{4}\sum_k\frac{u_k^2}{\delta_{k}^*}\mathcal{P}^*\,,\end{aligned}$$ and the scalar spectral index in (\[ns\]) to $$\begin{aligned} n_{s}-1=4\delta^{*}-4\frac{\sum_k\left(1-\frac{\gamma_{k}^*u_k}{\delta_{k}^*}\right)u_k+\sum_{kl}\frac{u_l}{\delta_{l}^*}A_{kl}}{\sum_ n\frac{u_n^2}{\delta_{n}^*}}\,.\label{ns2}\end{aligned}$$ This expression reduces to the ones in [@Byrnes:2009qy] for $\mathcal{N}=2$ and mimics a similar expression in the slow roll case with a separable potential [@Battefeld:2006sz]; note however that the meaning of quantities such as $u_k$ as well as numerical factors differ from the slow roll case [^9]. The bi-spectrum: $f_{NL}^{(4)}$\[sec:analyticfnl\] -------------------------------------------------- The general expression for $f_{NL}^{(4)}$ in (\[fnl\]) can be simplified to $$\begin{aligned} -\frac{6}{5}f^{(4)}_{NL}=2\frac{\sum_{k}\frac{u_k^2}{\delta_{k}^*}\left(1-\frac{u_k\gamma_{k}^*}{\delta_{k}^*}\right)+\sum_{kl}\frac{u_ku_l}{\delta_{k}^*\delta_{l}^*}A_{kl}}{\left(\sum_{m}\frac{u_m^2}{\delta_{m}^*}\right)^2}\,. \label{fnle}\end{aligned}$$ Again, this result reduces to the one found in [@Byrnes:2009qy] for two fields and mimics formally the slow roll result of [@Battefeld:2006sz], except for different numerical factors. The tri-spectrum: $\Delta \tau_{NL}^{(i)}$, $i=2,3,4$ \[sec:analytictaunl\] --------------------------------------------------------------------------- The three parameters in (\[tau2\])-(\[tau4\]) can be written as $$\begin{aligned} \Delta\tau^{(2)}_{NL}&=&\frac{\mathcal K}{4}\frac{-2^3}{\left(\sum_n\frac{u_n^2}{\delta_{n}^*}\right)^2}\Bigg[\sum_k\left(1-\frac{\gamma_{k}^*}{\delta_{k}^*}{u_k}\right)u_k+\sum_{kl}A_{kl}\frac{u_l}{\delta_{l}^*}\Bigg]\,,\label{tau2e}\\ \Delta\tau^{(3)}_{NL}&=&\frac{2^2}{\left(\sum_n\frac{u_n^2}{\delta_{n}^*}\right)^3}\Bigg[\sum_m\left(1-\frac{\gamma_{m}^*}{\delta_{m}^*}{u_m}\right)^2\frac{u_m^2}{\delta_{m}^*}+2\sum_{lm}\left(1-\frac{\gamma_{m}^*}{\delta_{m}^*}{u_m}\right)A_{lm}\frac{u_lu_m}{\delta_{l}^*{\delta_{m}^*}}\label{tau3e}\\ \nonumber &&+\sum_{klm}\frac{A_{kl}A_{km}}{\delta_{k}^*}\frac{u_lu_m}{\delta_{l}^*\delta_{m}^*}\Bigg]\,,\\ \Delta\tau^{(4)}_{NL}&=&\frac{\sum_ik_i^3}{T}\frac{2^3}{\left(\sum_n\frac{u_n^2}{\delta_{n}^*}\right)^3}\Bigg[\sum_m\frac{u_m^3}{(\delta_{m}^*)^2}\left(\frac{2(\gamma_{m}^*)^2-\beta_m^*}{\delta_m^*}u_m-\gamma_{m}^*\right)-3\sum_{lm}\frac{u_lu_m^2}{\delta_{l}^*(\delta_{m}^*)^2}\gamma_{m}^*A_{ml}\label{tau4e}\\ \nonumber &&+\sum_{klm}\frac{u_ku_lu_m}{\delta_{k}^*\delta_{l}^*\delta_{m}^*}A_{klm}\Bigg]\,.\end{aligned}$$ As expected, these expressions mimic formally the corresponding ones in [@Seery:2006vu]; however, we would like to emphasize, once again, that numerical factors as well as the definition and meaning of quantities differ. Discussion \[sec:discussion\] ----------------------------- The expressions for $n_s$ in (\[ns2\]), $f_{NL}^{(4)}$ in (\[fnle\]) and $\tau_{NL}^{(i)}$, $i=2,3,4$ in (\[tau2e\])-(\[tau4e\]) are our main results. They are applicable for any set of scalar fields with canonical kinetic terms and potentials such that a separable Hubble parameter results. Since fields can evolve fast and couplings between the fields are included, an application to the end phase of inflation, including preheating, appears feasible. Because $A_{kl}$ and $A_{klm}$ are symmetric in their indices, one can readily check that these terms only contribute to the non-linearity parameters if the fields are not multiples of each other, that is whenever the trajectory in field space is curved. This is expected, since in these instances isocurvature perturbations can influence the adiabatic mode. The expressions for $f_{NL}$ and $\tau_{NL}^{(i)}$, $i=2,3,4$ mimic formally the corresponding ones in slow roll models in [@Battefeld:2006sz; @Seery:2006vu], but numerical factors as well as the definition of terms differ. In the case of two fields, we recover the results of [@Byrnes:2009qy] for $f_{NL}^{(4)}$. Example \[sec:examples\] ======================== We would like to show the applicability of the formalism using a simple example which allows for the generation of large NG. Consider $H=\sum_k H_k$ with $$\begin{aligned} H_k(\varphi_k)=\frac{H_0}{\mathcal{N}}\left(1-A_k e^{-\alpha_k\varphi_k}\right)\,,\end{aligned}$$ with $A_k>0$. For $\mathcal{N}=2$, this example reduces to the second one of [@Byrnes:2009qy], so we can easily compare results. We first note that any $A_k\neq 1$ can be reabsorbed into a rescaling of $\varphi_k$. Second, if $\alpha_k<0$, we can simply redefine $\varphi_k\rightarrow -\varphi_k$ and $\alpha_k\rightarrow -\alpha_k$. Hence, we can set $A_k\equiv 1$ and $\alpha_k>0$ for all $k$ without loss of generality, so that $$\begin{aligned} H=H_0\left(1-\sum_k\frac{1}{\mathcal{N}}e^{-\alpha_k\varphi_k}\right)\,. \label{Hexample1}\end{aligned}$$ The Hamilton-Jacobi equations of motion in (\[HJ1\]) and (\[HJ2\]) are solved by $$\begin{aligned} a(\tau)&=&a_0e^{\tau}\prod_k\left(e^{\alpha_k\varphi_k^*}-\frac{2}{\mathcal{N}}\alpha_k^2\tau\right)^{1/(2\alpha_k^2)}\,,\\ \varphi_{k}(\tau)&=&\frac{1}{\alpha_k}\ln\left(e^{\alpha_k\varphi_k^*}-\frac{2}{\mathcal{N}}\alpha_k^2\tau\right)\,,\end{aligned}$$ where we defined the dimensionless time $\tau\equiv H_0t$. We set $t_*=0$, so that $\varphi_k(0)=\varphi_k^*$. The corresponding potential of the scalar fields is $$\begin{aligned} V=3H^2-2\sum_k H_{,k}^2\,.\end{aligned}$$ The potential has a wide plateau with $V\approx 3H_0^2$ that drops off steeply for $\varphi_k\sim \ln(2)/\alpha_k$. If the fields encounter this drop, inflation ends quickly and they roll to $-\infty$ in finite time; thus, the potential is unsuitable to describe the universe after inflation, but may be taken as a toy model during inflation, satisfying our demand for a separable Hubble parameter. Given $H$ from (\[Hexample1\]) we can compute $\delta_k$, $\gamma_k$ and $\beta_k$ from (\[deltak\])-(\[betak\]) to $$\begin{aligned} \delta_k&=&\left(\frac{H_0}{H(\tau)}\right)^2\frac{\alpha_k^2}{\mathcal{N}^2}e^{-2\alpha_k\varphi_k(\tau)}\,,\\ \gamma_k&=&-\alpha_k\sqrt{\delta_k}\,,\\ \beta_k&=&\alpha_k^2\delta_k=\gamma_k^2\,.\end{aligned}$$ Since inflation requires $\delta=\sum_k\delta_k\ll 1/2$, we necessarily have $\delta_k \ll 1/2$ during inflation. However, for suitable large $\alpha_k$, that is for sharp cliffs in $V$, we can still achieve large $\gamma_k$ and $\beta_k$. As a consequence, large NG are at least in principle possible, as we shall see below. We choose $t_c$ at the end of inflation, that is we determine $t_c$ by solving $$\begin{aligned} 2\delta(t_c)=1\end{aligned}$$ for $t_c$. The volume expansion rate in (\[N\]) from $t_c$ to $t_*$ can be integrated to $$\begin{aligned} N&=&\sum_k\left(\frac{\varphi_k^c-\varphi_k^*}{2\alpha_k}-\frac{e^{\alpha_k\varphi_k^c}-e^{\alpha_k\varphi_k^*}}{2\alpha_k^2}\right)\\ &=&\sum_k\left(\frac{1}{2\alpha_k^2}\ln\left(\frac{H^c\sqrt{\delta_k^c}}{H^*\sqrt{\delta_k^*}}\right)+\frac{H_0}{2\mathcal{N}H^*}\left(\frac{H^*}{H^c\gamma_k^c}-\frac{1}{\gamma_k^*}\right)\right)\,. \label{Nexample1}\end{aligned}$$ In order to conform with current observations we need $N\sim 60$ and $\delta^c\sim 10^{-2}$, to give a scalar spectral index close to $1$. So far, we have not specified the initial conditions, except for imposing one constraint, $N\sim 60$. In the two field case of [@Byrnes:2009qy] a set of initial conditions was found that yields a large $f_{NL}$, which can be generalized to $$\begin{aligned} \varphi_k^*=\frac{1}{\alpha_k}\ln\left(\left(N-\sqrt{\frac{m_k}{2\alpha_k^2}}\right)\frac{2\alpha_k^2}{\mathcal{N}}\right)\,, \label{init}\end{aligned}$$ where $m_k$ determines $\delta_k^c=1/(2m_k)$. If $\alpha_k>R\gg 1$ while $m_k<R^2$ with some large $R$, the logarithm in (\[Nexample1\]) can be neglected and, using $H_0\approx H^*$, one can write $$\begin{aligned} N&\approx&\sum_k\frac{1}{2\mathcal{N}}\left(\frac{H_0}{H^c\gamma_k^c}-\frac{1}{\gamma_k^*}\right)\,.\end{aligned}$$ It is further possible [@Byrnes:2009qy] to provide a simple analytic estimate in the two field case, leading to the expectation that $f_{NL}=\mathcal{O}(\mathcal{N} \alpha_k/\sqrt{m_k})$ and $\tau_{NL}=\mathcal{O}(f_{NL}^2)$, see [@Byrnes:2009qy] for details (the full expressions for $\tau_{NL}$ where not computed there). We use the full analytic expression in (\[ns2\]) for $n_s$, (\[fnle\]) for $f_{Nl}^{(4)}$ and (\[tau2e\])-(\[tau4e\]) for $\Delta \tau_{NL}^{(i)}$. Case-study $\mathcal{N}=2$\[caseN=2\] ------------------------------------- {width="\textwidth"} Choosing the initial conditions according to (\[init\]) with $\alpha_1=100$, $\alpha_2=20$, $m_1=6$ and $m_2=6/5$, we recover the result of [@Byrnes:2009qy], that is $n_s-1\approx -0.033$ and $f_{NL}$ from Fig. \[pic:NG\_Byrnes\] (a); we can also compute $\Delta \tau^{(i)}_{NL}$, see Fig. \[pic:NG\_Byrnes\] (b)-(d). As expected, not only does $f_{NL}$ become large once $2\delta=\epsilon$ approaches one, but also $\Delta\tau_{NL}^{(3)}$ and $\Delta\tau_{NL}^{(4)}$, both of which roughly scale as $\left(f_{NL}^{(4)}\right)^2$, as emphasized in [@Byrnes:2009qy], while $\Delta\tau_{NL}^{(2)}$ remains small. However, we would like to stress that the actual expressions differ by factors of order one from the simple estimates in [@Byrnes:2009qy] ($\Delta\tau_{NL}^{(3)}$ and $\Delta\tau_{NL}^{(4)}$ are roughly a factor of $3$ bigger than the estimate in [@Byrnes:2009qy] for this example [^10]). Since NG are only produced shortly before the end of inflation, we have chosen $\epsilon=-\dot{H}/H^2$ as a time variable ($N(t_{c})\approx 59.999$ when $\epsilon(t_{c})=1$ and $N(\tilde{t})\approx 59.922$ when $\epsilon(\tilde{t})=0.1$). {width="\textwidth"} It is instructive to alter the initial conditions slightly, see Fig. \[pic:NG\_Byrnes\_inichanges\], where we vary $\varphi_2^*$: as soon as we deviate from (\[init\]) in about $1$ part in $1000$, NG at the end of inflation become suppressed. The trajectories leading to large NG are the ones close to the line differentiating between trajectories bending towards the $\varphi_2$-axis or the $\varphi_1$-axis respectively, see Fig. \[pic:NG\_Byrnes\_inichanges\] (a). Close to this repeller, isocurvature perturbations are unstable and grow; once the trajectory bends, they influence the adiabatic modes which become non-Gaussian to some extent. This is similar to the origin of non-Gaussianities in the new-ekpyrotic scenario, where two fields are supposed to roll along a crest in the potential; if the ekpyrotic phase is terminated by rolling off the crest, as in [@Koyama:2007if], isocurvature modes are converted to adiabatic ones and as a byproduct NG are produced. Interestingly, it can be seen in Fig. \[pic:NG\_Byrnes\_inichanges\] (b)-(d) that an increase in the non-linearity parameters appears to be a transient effect, which may or may not be imprinted onto radiation, depending on the nature of reheating which can take anywhere from a fraction of an e-fold in efficient preheating models [^11] to many e-folds in the case of the old theory or reheating (see [@Bassett:2005xm] for a review of (p)reheating). Since the enhancement and decay in Fig. \[pic:NG\_Byrnes\_inichanges\] takes place in a fraction of the last e-fold, it becomes crucial to follow the evolution of the non-linearity parameters through the decay of the inflatons to make unambiguous predictions. Note that the sign of $\Delta \tau^{(4)}_{NL}$ can change depending on the choice of $t_c$, see i.e. Fig. \[pic:NG\_Byrnes\_inichanges\] (d), while $\Delta \tau^{(3)}_{NL}$ is always positive, in agreement with expectations from slow roll models, see i.e. [@Huang:2009xa]. Unfortunately, the potential we focused on is unsuitable for $\epsilon> 1$. To apply our formalism to a toy model of (p)reheating, one needs to find an $H_k(\varphi_k)$ such that $V(\varphi_1,\dots,\varphi_{\mathcal{N}})$ possesses a global minimum after inflation. We postpone this interesting study to a future publication. We conclude that in this toy model NG are only produced for fine tuned initial conditions. Case-study $\mathcal{N}=2,\dots,6$\[casengeneral\] -------------------------------------------------- {width="\textwidth"} The $\mathcal{N}=2$-case is special in that symmetries in $A_{kl}$ and $A_{klm}$ cause cancellations in the non-linearity parameters that lead to the simple analytic expressions in [@Byrnes:2009qy], such as the one for $f_{NL}^{(4)}$ in (38) of [@Byrnes:2009qy]. Here, we stay close to the previous section’s example, that is, we once again use initial conditions according to (\[fnle\]), but we increase the number of fields from $\mathcal{N}=2\dots 6$. We further set $\alpha_k=k \times 100/\mathcal{N}$ and $m_k=\mathcal{N}$. As in Sec. \[caseN=2\], the initial values are chosen to ensure observably large NG, albeit at the cost of fine tuning. In Fig. \[pic:NG\_Multi\], we plot $f_{NL}^{(4)}$ and $\Delta \tau^{(i)}_{NL}$ for $i=2,3,4$ using the full expressions in (\[fnle\]) and (\[tau2e\])-(\[tau4e\]). The scalar spectral index ($n_s-1\approx -0.033$) is unaffected by increasing the number of fields. The level of NG varies mildly if we increase the number of fields, but could be tuned to coincide at $\epsilon=1$ by small adjustments of the initial conditions. The largest difference in Fig. \[pic:NG\_Multi\] is caused by raising $\mathcal{N}=2$ to $3$, simply because the change in the exponent of $H_k$ is largest ($\mathcal N=2$ has $\alpha_1=50$ and $\alpha_2=100$, while $\mathcal N=3$ has $\alpha_1\approx 33.3$, $\alpha_2\approx 66.6$ and $\alpha_3=100$), which directly impacts NG. Fig. \[pic:NG\_Byrnes\_inichanges\] and \[pic:NG\_Multi\] illustrate the ambiguity problem of multi-field inflationary models. Whereas in the single-field case, the initial value, and thus all observables, are determined by the required e-folding number $N\sim 60$, there is already ambiguity in the two field case if only the two-point correlation function is considered; this may be lifted if NG are observed, since NG are usually undetectable in single field models, but can be larger if more fields are involved [^12], see Fig. \[pic:NG\_Byrnes\_inichanges\]. However, discriminating between $\mathcal{N}=2$ and $\mathcal{N}>2$ appears to be hopeless in the absence of a compelling reason for choosing one set of initial conditions over the other even if NG are observed. If one can choose starting points for the fields freely, any level of NG of a two-field case can be mimicked by a model with more fields. For instance, modest changes of the initial values in Fig. \[pic:NG\_Multi\] similar to the ones in Fig. \[pic:NG\_Byrnes\_inichanges\] would allow for a tuning of the non-linearity parameters at $\epsilon=1$. Discussion and Conclusions \[sec:conclusion\] ============================================= Based on the assumption of a separable Hubble parameter $H=\sum_kH_{k}(\varphi_k)$, we computed analytically Non-Gaussianities (NG), that is the non-linearity parameters characterizing the bi- and tri-spectrum, in multi-field inflationary models with an arbitrary number of fields and without using the slow roll approximation, the horizon crossing approximation, or a separable potential, extending the work of Byrnes and Tasinato [@Byrnes:2009qy]. Based on the $\delta N$-formalism, we derived analytic expression for $n_s$, $f_{NL}^{(4)}$ and $\Delta \tau^{(i)}_{NL}$ for $i=2,3,4$, which are easily computed once the background evolution of the fields is known. To show the applicability of the formalism, we considered a simple exponential dependence of the Hubble parameter on the fields. This model has the advantage of being analytically solvable at the background level, but it comes at the price of being unrealistic once inflation comes to an end. If initial conditions are fine tuned, observably large NG can be produced towards the end of inflation, but the general prediction of this class of models are negligible NG. Further, even if NG are produced when $\epsilon=-\dot{H}/H^2\sim 1$, it remains to be seen if they are imprinted onto radiation after inflation, because the sudden increase of the non-linearity parameters appears to be a transient phenomenon. We note that the formalism should be well suited for a computation of NG during (p)reheating, because the number of fields is not restricted, the slow roll approximation is not required and cross couplings between the fields are allowed. Indeed, in light of the example studied here, it appears mandatory to follow the non-linearity parameters through the era of inflaton decay, until fluctuations are imprinted onto radiation. We postpone a study of NG from (p)reheating to a forthcoming publication. We would like to thank C. Byrnes and G. Tasinato for comments. T.B. is supported by the Council on Science and Technology at Princeton University and is thankful for hospitaliy at the HIP and the APC. D. B was supported by the EU EP6 Marie Curie Research and Training Network ’UniverseNet’ (MRTN-CT-2006-035863) and is thankful for hospitality at Princeton University, the APC, and thanks Kari Enqvist for support. This paper is dedicated to the memory of Erlinda Balmaceda (1927-2009). D. N. Spergel [*et al.*]{}, arXiv:astro-ph/0603449. G. F. Smoot [*et al.*]{}, Astrophys. J.  [**396**]{}, L1 (1992). S. Hanany [*et al.*]{}, Astrophys. J.  [**545**]{}, L5 (2000) \[arXiv:astro-ph/0005123\]. C. B. Netterfield [*et al.*]{} \[Boomerang Collaboration\], Astrophys. J.  [**571**]{}, 604 (2002) \[arXiv:astro-ph/0104460\]. E. Komatsu [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl.  [**180**]{}, 330 (2009) \[arXiv:0803.0547 \[astro-ph\]\]. D. H. Lyth and A. Riotto, Phys. Rept.  [**314**]{}, 1 (1999) \[arXiv:hep-ph/9807278\]. B. A. Bassett, S. Tsujikawa and D. Wands, Rev. Mod. Phys.  [**78**]{}, 537 (2006) \[arXiv:astro-ph/0507632\]. E. Komatsu [*et al.*]{}, arXiv:0902.4759 \[astro-ph.CO\]. See the official ESA page http://www.rssd.esa.int/index.php?project=Planck E. Komatsu and D. N. Spergel, Phys. Rev.  D [**63**]{}, 063002 (2001) \[arXiv:astro-ph/0005036\]. D. Seery and J. E. Lidsey, JCAP [**0701**]{}, 008 (2007) \[arXiv:astro-ph/0611034\]. C. T. Byrnes, K. Koyama, M. Sasaki and D. Wands, JCAP [**0711**]{}, 027 (2007) \[arXiv:0705.4096 \[hep-th\]\]. C. T. Byrnes, M. Sasaki and D. Wands, Phys. Rev.  D [**74**]{}, 123519 (2006) \[arXiv:astro-ph/0611075\]. S. Yokoyama, T. Suyama and T. Tanaka, JCAP [**0707**]{}, 013 (2007) \[arXiv:0705.3178 \[astro-ph\]\]. S. Yokoyama, T. Suyama and T. Tanaka, Phys. Rev.  D [**77**]{}, 083511 (2008) \[arXiv:0711.2920 \[astro-ph\]\]. K. M. Smith, L. Senatore and M. Zaldarriaga, arXiv:0901.2572 \[astro-ph\]. J. M. Maldacena, JHEP [**0305**]{}, 013 (2003) \[arXiv:astro-ph/0210603\]. V. Acquaviva, N. Bartolo, S. Matarrese and A. Riotto, Nucl. Phys. B [**667**]{}, 119 (2003) \[arXiv:astro-ph/0209156\]. D. Babich, P. Creminelli and M. Zaldarriaga, JCAP [**0408**]{}, 009 (2004) \[arXiv:astro-ph/0405356\]. D. Seery and J. E. Lidsey, JCAP [**0506**]{}, 003 (2005) \[arXiv:astro-ph/0503692\]. N. Bartolo, E. Komatsu, S. Matarrese and A. Riotto, Phys. Rept.  [**402**]{}, 103 (2004) \[arXiv:astro-ph/0406398\]. D. Seery, M. S. Sloth and F. Vernizzi, JCAP [**0903**]{}, 018 (2009) \[arXiv:0811.3934 \[astro-ph\]\]. P. Creminelli, JCAP [**0310**]{}, 003 (2003) \[arXiv:astro-ph/0306122\]. X. Chen, R. Easther and E. A. Lim, JCAP [**0706**]{}, 023 (2007) \[arXiv:astro-ph/0611645\]. X. Chen, R. Easther and E. A. Lim, JCAP [**0804**]{}, 010 (2008) \[arXiv:0801.3295 \[astro-ph\]\]. K. Enqvist, A. Jokinen, A. Mazumdar, T. Multamaki and A. Vaihkonen, Phys. Rev. Lett.  [**94**]{}, 161301 (2005) \[arXiv:astro-ph/0411394\]. A. Jokinen and A. Mazumdar, JCAP [**0604**]{}, 003 (2006) \[arXiv:astro-ph/0512368\]. N. Barnaby and J. M. Cline, Phys. Rev.  D [**73**]{}, 106012 (2006) \[arXiv:astro-ph/0601481\]. N. Barnaby and J. M. Cline, Phys. Rev.  D [**75**]{}, 086004 (2007) \[arXiv:astro-ph/0611750\]. G. Dvali, A. Gruzinov and M. Zaldarriaga, Phys. Rev.  D [**69**]{}, 023505 (2004) \[arXiv:astro-ph/0303591\]. T. Suyama and M. Yamaguchi, Phys. Rev.  D [**77**]{}, 023505 (2008) \[arXiv:0709.2545 \[astro-ph\]\]. C. T. Byrnes, JCAP [**0901**]{}, 011 (2009) \[arXiv:0810.3913 \[astro-ph\]\]. T. Battefeld, Phys. Rev.  D [**77**]{}, 063503 (2008) \[arXiv:0710.2540 \[hep-th\]\]. M. Sasaki, J. Valiviita and D. Wands, Phys. Rev.  D [**74**]{}, 103003 (2006) \[arXiv:astro-ph/0607627\]. L. Boubekeur and D. H. Lyth, Phys. Rev.  D [**73**]{}, 021301 (2006) \[arXiv:astro-ph/0504046\]. A. D. Linde and V. F. Mukhanov, Phys. Rev.  D [**56**]{}, 535 (1997) \[arXiv:astro-ph/9610219\]. A. Linde and V. Mukhanov, JCAP [**0604**]{}, 009 (2006) \[arXiv:astro-ph/0511736\]. N. Bartolo, S. Matarrese and A. Riotto, Phys. Rev.  D [**65**]{}, 103505 (2002) \[arXiv:hep-ph/0112261\]. F. Bernardeau and J. P. Uzan, Phys. Rev.  D [**67**]{}, 121301 (2003) \[arXiv:astro-ph/0209330\]. K. A. Malik and D. H. Lyth, JCAP [**0609**]{}, 008 (2006) \[arXiv:astro-ph/0604387\]. B. Dutta, L. Leblond and J. Kumar, Phys. Rev.  D [**78**]{}, 083522 (2008) \[arXiv:0805.1229 \[hep-th\]\]. K. Ichikawa, T. Suyama, T. Takahashi and M. Yamaguchi, Phys. Rev.  D [**78**]{}, 023513 (2008) \[arXiv:0802.4138 \[astro-ph\]\]. S. Li, Y. F. Cai and Y. S. Piao, Phys. Lett.  B [**671**]{}, 423 (2009) \[arXiv:0806.2363 \[hep-ph\]\]. D. Langlois, F. Vernizzi and D. Wands, JCAP [**0812**]{}, 004 (2008) \[arXiv:0809.4646 \[astro-ph\]\]. C. Hikage, K. Koyama, T. Matsubara, T. Takahashi and M. Yamaguchi, arXiv:0812.3500 \[astro-ph\]. M. Kawasaki, K. Nakayama, T. Sekiguchi, T. Suyama and F. Takahashi, JCAP [**0811**]{}, 019 (2008) \[arXiv:0808.0009 \[astro-ph\]\]. Q. G. Huang, JCAP [**0811**]{}, 005 (2008) \[arXiv:0808.1793 \[hep-th\]\]. T. Suyama and F. Takahashi, JCAP [**0809**]{}, 007 (2008) \[arXiv:0804.0425 \[astro-ph\]\]. S. Mizuno, F. Arroja and K. Koyama, arXiv:0907.2439 \[hep-th\]. M. Alishahiha, E. Silverstein and D. Tong, Phys. Rev.  D [**70**]{}, 123505 (2004) \[arXiv:hep-th/0404084\]. X. Chen, M. x. Huang, S. Kachru and G. Shiu, JCAP [**0701**]{} (2007) 002 \[arXiv:hep-th/0605045\]. X. Chen, M. x. Huang and G. Shiu, Phys. Rev.  D [**74**]{} (2006) 121301 \[arXiv:hep-th/0610235\]. F. Arroja and K. Koyama, Phys. Rev.  D [**77**]{}, 083517 (2008) \[arXiv:0802.1167 \[hep-th\]\]. F. Arroja, S. Mizuno and K. Koyama, JCAP [**0808**]{}, 015 (2008) \[arXiv:0806.0619 \[astro-ph\]\]. D. Langlois, S. Renaux-Petel, D. A. Steer and T. Tanaka, Phys. Rev. Lett.  [**101**]{}, 061301 (2008) \[arXiv:0804.3139 \[hep-th\]\]. D. Langlois, S. Renaux-Petel, D. A. Steer and T. Tanaka, Phys. Rev.  D [**78**]{}, 063523 (2008) \[arXiv:0806.0336 \[hep-th\]\]. X. Gao and B. Hu, JCAP [**0908**]{}, 012 (2009) \[arXiv:0903.1920 \[astro-ph.CO\]\]. Y. F. Cai and H. Y. Xia, Phys. Lett.  B [**677**]{}, 226 (2009) \[arXiv:0904.0062 \[hep-th\]\]. D. Langlois, S. Renaux-Petel and D. A. Steer, JCAP [**0904**]{}, 021 (2009) \[arXiv:0902.2941 \[hep-th\]\]. J. Khoury and F. Piazza, arXiv:0811.3633 \[hep-th\]. F. Arroja, S. Mizuno, K. Koyama and T. Tanaka, arXiv:0905.3641 \[hep-th\]. X. Chen, B. Hu, M. x. Huang, G. Shiu and Y. Wang, arXiv:0905.3494 \[astro-ph.CO\]. S. Mizuno, F. Arroja, K. Koyama and T. Tanaka, arXiv:0905.4557 \[hep-th\]. X. Gao, M. Li and C. Lin, arXiv:0906.1345 \[astro-ph.CO\]. S. Renaux-Petel, arXiv:0907.2476 \[hep-th\]. N. Barnaby and J. M. Cline, JCAP [**0707**]{}, 017 (2007) \[arXiv:0704.3426 \[hep-th\]\]. N. Barnaby and J. M. Cline, JCAP [**0806**]{}, 030 (2008) \[arXiv:0802.3218 \[hep-th\]\]. P. Creminelli and L. Senatore, JCAP [**0711**]{}, 010 (2007) \[arXiv:hep-th/0702165\]. J. L. Lehners, P. McFadden, N. Turok and P. J. Steinhardt, Phys. Rev.  D [**76**]{}, 103501 (2007) \[arXiv:hep-th/0702153\]. E. I. Buchbinder, J. Khoury and B. A. Ovrut, Phys. Rev.  D [**76**]{}, 123503 (2007) \[arXiv:hep-th/0702154\]. K. Koyama, S. Mizuno, F. Vernizzi and D. Wands, JCAP [**0711**]{}, 024 (2007) \[arXiv:0708.4321 \[hep-th\]\]. E. I. Buchbinder, J. Khoury and B. A. Ovrut, Phys. Rev. Lett.  [**100**]{}, 171302 (2008) \[arXiv:0710.5172 \[hep-th\]\]. J. L. Lehners and P. J. Steinhardt, Phys. Rev.  D [**78**]{}, 023506 (2008) \[Erratum-ibid.  D [**79**]{}, 129902 (2009)\] \[arXiv:0804.1293 \[hep-th\]\]. J. L. Lehners and P. J. Steinhardt, Phys. Rev.  D [**77**]{}, 063533 (2008) \[Erratum-ibid.  D [**79**]{}, 129903 (2009)\] \[arXiv:0712.3779 \[hep-th\]\]. J. L. Lehners and S. Renaux-Petel, arXiv:0906.0530 \[hep-th\]. J. L. Lehners, Phys. Rept.  [**465**]{}, 223 (2008) \[arXiv:0806.1245 \[astro-ph\]\]. T. Battefeld and R. Easther, JCAP [**0703**]{}, 020 (2007) \[arXiv:astro-ph/0610296\]. D. Seery and J. E. Lidsey, JCAP [**0509**]{}, 011 (2005) \[arXiv:astro-ph/0506056\]. D. H. Lyth and Y. Rodriguez, Phys. Rev. Lett.  [**95**]{}, 121302 (2005) \[arXiv:astro-ph/0504045\]. L. Alabidi and D. H. Lyth, JCAP [**0605**]{}, 016 (2006) \[arXiv:astro-ph/0510441\]. D. Seery, J. E. Lidsey and M. S. Sloth, JCAP [**0701**]{}, 027 (2007) \[arXiv:astro-ph/0610210\]. F. Bernardeau and J. P. Uzan, Phys. Rev.  D [**66**]{}, 103506 (2002) \[arXiv:hep-ph/0207295\]. D. H. Lyth and Y. Rodriguez, Phys. Rev.  D [**71**]{}, 123508 (2005) \[arXiv:astro-ph/0502578\]. D. H. Lyth, JCAP [**0511**]{}, 006 (2005) \[arXiv:astro-ph/0510443\]. M. P. Salem, Phys. Rev.  D [**72**]{}, 123516 (2005) \[arXiv:astro-ph/0511146\]. L. Alabidi and D. Lyth, JCAP [**0608**]{}, 006 (2006) \[arXiv:astro-ph/0604569\]. L. Alabidi, JCAP [**0610**]{}, 015 (2006) \[arXiv:astro-ph/0604611\]. M. Sasaki, Prog. Theor. Phys.  [**120**]{}, 159 (2008) \[arXiv:0805.0974 \[astro-ph\]\]. C. T. Byrnes, K. Y. Choi and L. M. H. Hall, JCAP [**0810**]{}, 008 (2008) \[arXiv:0807.1101 \[astro-ph\]\]. C. T. Byrnes, K. Y. Choi and L. M. H. Hall, JCAP [**0902**]{}, 017 (2009) \[arXiv:0812.0807 \[astro-ph\]\]. A. Naruko and M. Sasaki, Prog. Theor. Phys.  [**121**]{}, 193 (2009) \[arXiv:0807.0180 \[astro-ph\]\]. X. Gao, JCAP [**0806**]{}, 029 (2008) \[arXiv:0804.1055 \[astro-ph\]\]. Q. G. Huang, JCAP [**0905**]{}, 005 (2009) \[arXiv:0903.1542 \[hep-th\]\]. D. Battefeld and T. Battefeld, JCAP [**0705**]{}, 012 (2007) \[arXiv:hep-th/0703012\]. K. Y. Choi, L. M. H. Hall and C. van de Bruck, JCAP [**0702**]{}, 029 (2007) \[arXiv:astro-ph/0701247\]. A. Misra and P. Shukla, Nucl. Phys.  B [**800**]{}, 384 (2008) \[arXiv:0712.1260 \[hep-th\]\]. A. Misra and P. Shukla, Nucl. Phys.  B [**810**]{}, 174 (2009) \[arXiv:0807.0996 \[hep-th\]\]. H. R. S. Cogollo, Y. Rodriguez and C. A. Valenzuela-Toledo, JCAP [**0808**]{}, 029 (2008) \[arXiv:0806.1546 \[astro-ph\]\]. Y. Rodriguez and C. A. Valenzuela-Toledo, arXiv:0811.4092 \[astro-ph\]. Q. G. Huang, JCAP [**0906**]{}, 035 (2009) \[arXiv:0904.2649 \[hep-th\]\]. K. A. Malik, JCAP [**0511**]{}, 005 (2005) \[arXiv:astro-ph/0506532\]. K. A. Malik, JCAP [**0703**]{}, 004 (2007) \[arXiv:astro-ph/0610864\]. C. Gordon, D. Wands, B. A. Bassett and R. Maartens, Phys. Rev.  D [**63**]{}, 023506 (2001) \[arXiv:astro-ph/0009131\]. F. Vernizzi and D. Wands, JCAP [**0605**]{}, 019 (2006) \[arXiv:astro-ph/0603799\]. D. Battefeld, T. Battefeld and A. C. Davis, JCAP [**0810**]{}, 032 (2008) \[arXiv:0806.1953 \[hep-th\]\]. T. Battefeld, arXiv:0809.3242 \[astro-ph\]. D. Battefeld and T. Battefeld, JCAP [**0903**]{}, 027 (2009) \[arXiv:0812.0367 \[hep-th\]\]. K. Becker, M. Becker and A. Krause, Nucl. Phys.  B [**715**]{}, 349 (2005) \[arXiv:hep-th/0501130\]. A. Ashoorioon and A. Krause, arXiv:hep-th/0607001. D. H. Lyth, K. A. Malik and M. Sasaki, JCAP [**0505**]{}, 004 (2005) \[arXiv:astro-ph/0411220\]. A. A. Starobinsky, JETP Lett. [**42**]{}, 152 (1985) \[Pis. Hz. Esp. Tor. Fizz. [42]{}, 124 (1985)\]. M. Sasaki and E. D. Stewart, Prog. Theor. Phys.  [**95**]{}, 71 (1996) \[arXiv:astro-ph/9507001\]. C. T. Byrnes and G. Tasinato, arXiv:0906.0767 \[astro-ph.CO\]. A. R. Liddle, A. Mazumdar and F. E. Schunck, Phys. Rev. D [**58**]{}, 061301 (1998) \[arXiv:astro-ph/9804177\]. P. Kanti and K. A. Olive, Phys. Rev. D [**60**]{}, 043502 (1999) \[arXiv:hep-ph/9903524\]. P. Kanti and K. A. Olive, Phys. Lett.  B [**464**]{}, 192 (1999) \[arXiv:hep-ph/9906331\]. K. A. Malik and D. Wands, Phys. Rev.  D [**59**]{}, 123501 (1999) \[arXiv:astro-ph/9812204\]. D. S. Salopek and J. R. Bond, Phys. Rev.  D [**42**]{}, 3936 (1990). L. E. Allen, S. Gupta and D. Wands, JCAP [**0601**]{}, 006 (2006) \[arXiv:astro-ph/0509719\]. G. I. Rigopoulos and E. P. S. Shellard, Phys. Rev.  D [**68**]{}, 123518 (2003) \[arXiv:astro-ph/0306620\]. T. Okamoto and W. Hu, Phys. Rev.  D [**66**]{}, 063008 (2002) \[arXiv:astro-ph/0206155\]. N. Kogo and E. Komatsu, Phys. Rev.  D [**73**]{}, 083007 (2006) \[arXiv:astro-ph/0602099\]. D. Battefeld, T. Battefeld and J. T. Giblin, Phys. Rev.  D [**79**]{}, 123510 (2009) \[arXiv:0904.2778 \[astro-ph.CO\]\]. D. Battefeld, arXiv:0809.3455 \[astro-ph\]. D. Battefeld and S. Kawai, Phys. Rev.  D [**77**]{}, 123507 (2008) \[arXiv:0803.0321 \[astro-ph\]\]. [^1]: In some cases large NG are possible during slow roll even in simple models, see i.e. [@Byrnes:2008wi] for fine tuning initial conditions, [@Rodriguez:2008hy; @Cogollo:2008bi] for contributions of loops or [@Huang:2009vk] for effects related to geometric quantities of the hyper-surface at which inflation ends. [^2]: This formalism is a crucial extension of the linear $\delta N$-formalism [@Starobinski; @Sasaki:1995aw], and needed to deal with higher order correlation functions. [^3]: The separate universe formalism put forward by Rigopoulos and Shellard in e.g. [@Rigopoulos:2003ak] is equivalent to the $\delta N$-formalism. [^4]: The definition in (\[C\]) is not unique and sets of (independent) linear combinations of the $C_k$’s may be used alternatively [@Seery:2006js]. [^5]: The derivatives of the volume expansion rate with respect to $\varphi_k^*$ depend on both, $t_c$ and $t_*$, so no superscript is used in (\[def:derivativeofN\]). To avoid confusion with summation indices, we use $k,l,m,\dots$ exclusively for indices that run from $1$ to $\mathcal{N}$ and $i,j$ for indices that run from $1$ to $4$. [^6]: If the slow roll approximation is made and a separable potential is used, one can derive [@Lyth:1998xn] $N(t_c,t_*)=-\int_*^c \sum_{k=1}^{\mathcal N}(V_k/ V_k^\prime) d\varphi_k$ which also enables the analytic computation of non-linearity parameters within the $\delta N$-formalism. [^7]: We use the sign convention of [@Vernizzi:2006ve] for $f_{NL}$, which is opposite to the one by Komatsu and Spergel [@Komatsu:2001rj; @Komatsu:2008hk]. [^8]: In the notation of Byrnes [@Byrnes:2006vq; @Byrnes:2009qy] $\Delta\tau_{NL}^{(3)}=\tau_{NL}^{(\mbox{\tiny Byrnes})}$ and $\Delta\tau_{NL}^{(4)}=g_{NL}^{(\mbox{\tiny Byrnes})} 27{T}/(25\sum_ik_i^3)$. [^9]: In general, our assumption of a separable Hubble parameter excludes cases with a separable potential; hence neither case can be derived as the limit of the other. [^10]: The approximation in [@Byrnes:2009qy] is valid if one of the fields gives the dominant contribution to $\zeta$ at all required orders. In the example above, the approximation in [@Byrnes:2009qy] gets more accurate if $\alpha_2$ is decreased, approaching the $1\%$ level for $\alpha_2\sim 1$. We thank C. Byrnes for discussions. [^11]: Note that preheating via parametric resonance is generically suppressed in multi-field models of inflation [@Battefeld:2008bu; @Battefeld:2008rd; @Battefeld:2009xw], but tachyonic instabilities can still lead to a fast decay of the inflatons [@Battefeld:2009xw]. [^12]: Not all multi-field potentials offer the possibility to generate large NG; see i.e. the first example in [@Byrnes:2009qy] for a case with unobservably small NG, (regardless of initial conditions), $\mathcal{N}$-flation [@Battefeld:2007en], the examples in [@Battefeld:2006sz], or the examples in this paper if the initial conditions are not fine tunned, among others.

ArXiv

Introduction {#sec:Introduction} ============ Back in 1978, Fang and Fronsdal, in a paper discussing the deformation theoretic approach to deriving a self-interacting massless spin-2 field theory, proposed what they called a “generalized Gupta program” to the effect of deriving self-interactions for higher spin gauge fields [@FangFronsdal1979]. Fronsdal apparently soon realized that such a program was not likely to succeed for any single spin $s$. In a conference proceedings from 1979 he suggested that, what we now refer to as an [*infinite tower*]{} of higher spin gauge fields, is needed. Indeed he wrote [@Fronsdal1979conf]: “[*For spins exceeding 2 it would seem to be very difficult to find a non-Abelian gauge algebra without including all spins or at least all integer spins. Thus the question calls for a single, unifying gauge algebra for all integer spins.*]{}” That something like this could be the case can surmised from an unsophisticated generalization from spin 1 and 2. For spin 1, commuting two “gauge generators” $\xi_a T^a$ and $\eta_a T^a$ yields a new transformation of the same form. Likewise, by analogy, for spin 2, commuting $\xi^\mu\partial_\mu$ and $\eta^\mu\partial_\mu$, yields an object of the same form. The obvious generalization to spin 3 would be commuting objects of the form $\xi^{\mu\nu}_aT^a\partial_\mu\partial_\nu$. However, this operation does not close [@AKHB1985]. In the language of deformation theory, we discover an obstruction. The form of the non-closure terms in the commutator suggests the introduction of spin 5 gauge transformations [@Burgers1985thesis]. The need for a “single unifying gauge algebra for all integer spins” can therefore be discerned. Furthermore, the step to thinking in terms of infinite component objects $\Phi$ in some way maintaining fields of all spin, is now short. In the paper referred to above, Fronsdal outlined one such approach which will be described below. Thus, adding in the fact that the massless representations of the Poincar[é]{} group can be realized on symmetric tensor fields (possibly modulo certain traces), these considerations together strongly hint at some kind of expansion $$\begin{gathered} \Phi=\phi B+\phi^\mu B_\mu+\phi^{\mu\nu} B_{\mu\nu}+\phi^{\mu\nu\rho} B_{\mu\nu\rho}+\cdots,\end{gathered}$$ where the set $\{B_{\mu_1\mu_2\ldots\mu_s}\}_{s=0}^\infty=\{B_{(s)}\}_{s=0}^\infty$ can be thought of as an infinite basis or an infinite set of generators. Many such proposals can be found in the literature, indeed, this simple idea seems to be reinvented ever and anon. The question arises as to whether the $B$’s obey an algebra, for instance $$\begin{gathered} [B_{(s)},B_{(t)}]=\sum_{i,j}^{s,t}c^{ij}B_{(i,j)}\end{gathered}$$ in a vague, but hopefully suggestive notation, or if they simply commute. Of course, higher order brackets might occur, or be necessary, at least the Jacobiator must be considered. This opens up the whole field of abstract algebra. Clearly, without guiding principles, we’re groping about in the dark. However, after thirty years of groping about by many authors, the territory is glowing feebly. The purpose of the present paper is certainly not to attempt a full review, rather we will try to further illuminate some parts of the terrain. Lacking guiding principles, our work will very much be playing around with equations. Since in my opinion, not very much is gained by working in higher dimensions, I will work in four spacetime dimensions unless otherwise stated. Higher spin bases and generators {#sec:HSBG} ================================ In this section, without any claims of being exhaustive, we will look at a few examples of higher spin bases/generators. Perhaps the simplest one, and certainly one that obviously might come to mind upon trying to generalize spin 1 and 2, is a derivative basis already alluded to in the introduction. Back to zero ------------ When two infinitesimal transformations $\delta_\xi$ and $\delta_\eta$ in some space are performed consecutively, the result is in general different depending on the order in which the transformations are done. This is one reason why it is interesting to study the commutator $[\delta_\xi,\delta_\eta]$ of the transformations. This could be anything, but it seems that the interesting cases arise when this difference can itself be expressed as a new transformation of the same type as the ones that are commuted. In any case, once the commutator is singled out as an interesting object of study, the Jacobi identities are forced upon the theory. As they are simply syntactical consequences of writing out $[[\delta_{\xi_1},\delta_{\xi_2}],\delta_{\xi_3}]+\mbox{cyc.\,perm.}$ and assuming that the infinitesimal operations $\delta_\xi$ are associative, there is no way to get around them. Thus the semantics of the theory must obey them, i.e. whatever results from explicitly calculating the commutator brackets, it must obey the Jacobi identities. In general, brackets $[\cdot,\cdot]$ are used as a notation for operations that are not commutators defined in terms of an underlying associative multiplication. In such cases the Jacobiator $[[\cdot,\cdot],\cdot]+\mbox{cyc.\,perm.}$ might only be zero up to higher order brackets, or [*homotopies*]{}. In any case, independent axioms are needed for the Jacobiator. Derivative basis ---------------- Even on the weak inductive base step of comparing the sequence of natural numbers $s=1,2,3,\ldots$ to the known sequence of spin 1 and 2 gauge parameters $\xi^a$, $\xi^\mu$, it could have been guessed that the higher spin continuation of this sequence should be something like $\xi^{a\mu_1\mu_2}$, $\xi^{\mu_1\mu_2\mu_3},\ldots$ with some internal index $a\in\{1,\ldots,n\}$ for odd spin and $s-1$ spacetime indices $\mu_i$ for spin $s$. The construction in 1983 of the cubic covariant vertex in Minkowski spacetime [@BerendsBurgersvanDam1984] for spin 3 and the light-front arbitrary $s$ cubic vertices [@BBB1983a] would have corroborated such a guess. Now let us see what, if anything, can be surmised from taking a derivative basis seriously. To begin with, assume that a spin 3 gauge generator is actually given by an object of the form $\xi^{\mu\nu}_aT^a\partial_\mu\partial_\nu$. Commuting two of these yield $$\begin{gathered} [\xi^{\mu\nu}_{1a}T^a\partial_\mu\partial_\nu,\xi^{\rho\sigma}_{2b}T^b\partial_\rho\partial_\sigma]= \xi^{\mu\nu}_{1a}\xi^{\rho\sigma}_{2b}[T^a,T^b]\partial_\mu\partial_\nu\partial_\rho\partial_\sigma+ 2(\xi^{\mu\nu}_{1a}\partial_\mu\xi^{\rho\sigma}_{2b}-\xi^{\mu\nu}_{2a}\partial_\mu\xi^{\rho\sigma}_{1b}) T^aT^b\partial_\nu\partial_\rho\partial_\sigma\nonumber\\ \phantom{[\xi^{\mu\nu}_{1a}T^a\partial_\mu\partial_\nu,\xi^{\rho\sigma}_{2b}T^b\partial_\rho\partial_\sigma]=}{} +(\xi^{\mu\nu}_{1a}\partial_\mu\partial_\nu\xi^{\rho\sigma}_{2b} -\xi^{\mu\nu}_{2a}\partial_\mu\partial_\nu\xi^{\rho\sigma}_{1b})T^aT^b\partial_\rho\partial_\sigma.\label{eq:CommutingSpin3Generators}\end{gathered}$$ The first term is easily interpreted as a spin 5 transformation by taking $[T_a,T_b]=f_{ab}^{\,\,\,\,c}T_c$ as we would be prejudiced to require of the matrices $T$. However, then the last term becomes problematic since it clearly must be a new spin 3 transformation. It the seems (provisionally) that we need a weaker set of equations for products of matrices, namely $$\begin{gathered} \label{eq:MatrixEquations} T_aT_b=g_{ab}^{\,\,\,\,c}T_c,\end{gathered}$$ with no conditions on the coefficients $g_{ab}^{\,\,\,\,c}$. Next, the second term of (\[eq:CommutingSpin3Generators\]) which corresponds to a spin 4 transformation, implies that also the even higher spin ($s>2$) fields must carry internal indices. However, the light-front cubic interaction terms for even higher spin require no anti-symmetric internal indexing, indeed in that case the interactions would vanish. On the other hand, commuting two spin 4 generators of the form $\xi^{\mu\nu\rho}$, yield spin 3 generators without internal matrices $T$. These considerations show that in order to make a simpleminded scheme like the one described here work, every higher spin field must be expanded over a set of matrices $T$ with a product satisfying equation (\[eq:MatrixEquations\]) as well as having a component along a unit matrix $E$, or $$\begin{gathered} \phi=\varphi^aT_a+\varphi E.\end{gathered}$$ Taking $E$ to be $T_0$ we let the gauge fields be valued in a Clifford algebra with basis elements $T_a$, $a\in\{0,\ldots,N\}$. The equation (\[eq:MatrixEquations\]) can then be satisfied with suitable choices for the $g_{ab}^{\,\,\,\,c}$ (such bases can be built using [*matrix units*]{} [@FuchsScweigertBook]). Tentatively, the would-be gauge algebra is generated by operators $$\begin{gathered} \Xi=\xi^a T_a+\sum_{k=1}^\infty\xi^{a\mu_{i_1}\ldots\mu_{i_k}}T_a\partial_{\mu_{i_1}}\cdots\partial_{\mu_{i_k}},\end{gathered}$$ acting as $$\begin{gathered} \delta_\xi\Phi=[\Xi,\Phi].\end{gathered}$$ The crucial question is of course whether operators of this form generate a structure that can be considered to be a gauge algebra? This was studied in [@BB1986]. There it was shown that the operators $\Xi$ obey the Jacobi identity (which they do since they associate as a consequence of the associativity of the matrices $T$ and the Leibniz rule for the derivatives $\partial$) and that structure functions can be extracted according to the following scheme. Using the Jacobi identity $$\begin{gathered} [\delta_\lambda,\delta_\xi]\Phi=[\Lambda,[\Xi,\Phi]]-[\Xi,[\Lambda,\Phi]]=[[\Lambda,\Xi],\Phi],\end{gathered}$$ we find that the commutator of two transformations can be written $$\begin{gathered} [\delta_\lambda,\delta_\xi]\Phi=\delta_{[\lambda,\xi]}\Phi=\delta_\omega\Phi,\end{gathered}$$ with structure thus extracted as $\Omega=[\Lambda,\Xi]$. There are however some problems with this approach, the most serious that it does not seem to reproduce the lowest order spin 3 structure equations $[\lambda,\xi]^a_{\mu\nu}$ of [@BerendsBurgersvanDam1984; @Burgers1985thesis] (BBvD). I write [*seem*]{}, because as far as I know, this has not really been investigated. Reconsidering the whole setup, it is clear that, had the derivative basis scheme worked, we would have had a field independent higher spin gauge algebra. According to the analysis of (BBvD), they find field dependence once one goes beyond spin 2. They then perform a general analysis of the gauge algebra problem and find that the spin 3 gauge algebra cannot close on spin 3 fields. The non-closure terms however have a form that suggests that they correspond to spin 5 gauge transformations. Unfortunately it is not known whether in this way introducing a tower of higher spin gauge fields, the full algebra would become field independent. Clearly, there must be more interesting structure to extract here. There is a recent paper that alludes to this [@Bekaert2007]. Fronsdal’s symplectic basis --------------------------- In the conference proceedings referred to in the introduction [@Fronsdal1979conf], Fronsdal briefly outlines an attempt to set up a non-Abelian higher spin gauge algebra. He considers a cotangent phase space over spacetime with the usual symplectic structure and coordinates $(x^\mu,\pi_\nu)$. Then he considers a set of traceless symmetric gauge parameters $\{\xi^{\mu_1\dots\mu_{s-1}}\}_{s=0}^\infty$ and the corresponding set of gauge fields $\{\phi^{\mu_1\dots\mu_{s}}\}_{s=1}^\infty$. These are collected into formal power series $$\begin{gathered} \Xi(\pi,x)=\sum_{s=1}^\infty(\pi^2)^{1-s/2}\pi_{\mu_1}\cdots\pi_{\mu_{s-1}}\xi^{\mu_1\dots\mu_{s-1}},\nonumber\\ \Phi(\pi,x)=\pi^2+\sum_{s=1}^\infty(\pi^2)^{1-s/2}\pi_{\mu_1}\cdots\pi_{\mu_{s}}\phi^{\mu_1\dots\mu_{s}}. $$ Transformations can now be calculated using the Poisson bracket $\{x_\mu,\pi_\nu\}=\eta_{\mu\nu}$ between $x_\mu$ and $\pi_\nu$ $$\begin{gathered} \delta_\Xi\Phi=\{\Xi,\Phi\}.\end{gathered}$$ The presence of the $\pi^2$ term in $H$ is needed in order to get the free theory transformations $$\begin{gathered} \{\Xi,\pi^2\}=2\sum_{s=1}^\infty(\pi^2)^{1-s/2}\pi_{\mu_1}\cdots\pi_{\mu_{s}}\partial^{\mu_1}\xi^{\mu_2\dots\mu_{s}}.\end{gathered}$$ Again, the Jacobi identity allows us to extract structure equations $$\begin{gathered} [\delta_\Lambda,\delta_\Xi]\Phi=\{\Lambda,\{\Xi,\Phi\}\}-\{\Xi,\{\Lambda,\Phi\}\}=\delta_{\{\Lambda,\Xi\}}\Phi.\end{gathered}$$ And once again we would get a field independent gauge algebra. Note that this construction is a form of Schouten brackets. Oscillator basis {#subsec:OscBas} ---------------- Formally introducing a covariant harmonic oscillator pair $(\alpha_\mu,\alpha_\mu^\dagger)$ with the usual commutator $[\alpha_\mu,\alpha_\mu^\dagger]=\eta_{\mu\nu}$ and a vacuum $|0\rangle$ satisfying $\alpha_\mu|0\rangle=0$, one could consider collecting all higher spin gauge fields in the expansion $$\begin{gathered} \label{eq:OscillatorExpansionFields} |\Phi\rangle=(\phi +\phi^\mu\alpha_\mu^\dagger+\phi^{\mu\nu}\alpha_\mu^\dagger\alpha_\nu^\dagger+ \phi^{\mu\nu\rho}\alpha_\mu^\dagger\alpha_\nu^\dagger\alpha_\rho^\dagger+\cdots)|0\rangle.\end{gathered}$$ This yields a concrete calculational scheme and one could hope to be able to rig some kind of kinetic operator $K$ which when acting on the field $|\Phi\rangle$, generate field equations for the component fields. This can indeed be done as is well known. In practice it is done using BRST methods. Several equivalent such formulations were discovered in the mid 1980’s [@SiegelZwiebach1987; @OuvryStern1987a; @AKHB1987a; @Meurice1988]. Formulations of this form has subsequently been rediscovered, developed and extended by a few research groups during the last ten years [@Labastida1989; @PashnevTsulaia1997; @PashnevTsulaia1998a; @FranciaSagnotti2002a]. Of course, expansions like (\[eq:OscillatorExpansionFields\]) first occurred in the bosonic string theory, and the BRST formulations was partly inspired by string field theory and by taking the limit $\alpha '\rightarrow\infty$ (or the zero tension limit) [@AKHB1987a]. The zero tension limit has been much studied lately (see for example [@Sundborg2001a; @Bonelli2003a; @SagnottiTsulaia2004a]). For some recent reviews of these developments as well as more complete sets of references see for example [@FranciaSagnotti2006rw; @Bianchi2004a] and also [@SagnottiSezginSundell2005; @FranciaMouradSagnotti2007a] which contain many references as well interesting discussions of these topics. This is closely tangled up with ideas about holography and the AdS/CFT duality [@Maldacena1997a] in a way that seems not fully understood. It is thus obvious that higher spin gauge fields occur in very many contexts. They are a prevalent feature of any model of extended and/or composite relativistic systems in various background geometries and dimensions. For discussions of these topics, see [@SezginSundell2002a; @BeisertBianchiMoralesSamtleben2004a]. But rather than study them in these contexts, I will continue to discuss higher spin symmetry and higher spin gauge theory as a free-standing theoretical construct. Thus let us return to the oscillator-BRST formulation and briefly review the simplest model. Consider therefore a phase space spanned by bosonic variables $(x_\mu,p_\mu)$ and $(\alpha_\mu,\alpha^{\dagger}_\mu)$ and ghost variables $(c^+,b_+)$, $(c^-,b_-)$ and $(c^0,b_0)$ with commutation relations $$\begin{gathered} [x_{\mu},p_{\nu}]=i\eta_{\mu\nu},\qquad[\alpha_{\mu},\alpha_{\nu}^{\dagger}]=\eta_{\mu\nu}, \qquad\{c^+,b_+\}=\{c^-,b_-\}=\{c^0,b_0\}=1.\end{gathered}$$ The ghosts have the following properties under Hermitian conjugation $$\begin{gathered} (c^-)^\dagger=c^+,\qquad(b_-)^\dagger=b_+,\qquad(c^0)^\dagger=c^0,\qquad(b_0)^\dagger=b_0.\end{gathered}$$ The doubly degenerate vacuum states $|+\rangle$, $|-\rangle$ are annihilated by the operators $\alpha_\mu$, $c^-$, $b^+$ while the degeneracy is given by $$\begin{gathered} b_0|+\rangle=0,\qquad b_0|-\rangle=|+\rangle,\qquad c^0|-\rangle=0,\qquad c^0|+\rangle=|-\rangle.\end{gathered}$$ These equations relating the vacua, then implies that either one of the two vacua must be odd. I will choose $|-\rangle$ Grassmann odd. The higher spin fields are collected into the ket $|\Phi\rangle$ with expansion $$\begin{gathered} |\Phi\rangle=\Phi(p)|+\rangle+F(p)c^+b_{-}|+\rangle+H(p)b_-)|-\rangle,\end{gathered}$$ where $\Phi(p)$ contains the symmetric higher spin gauge fields, and $F(p)$ and $H(p)$ are certain auxiliary fields. These fields are further expanded in terms of the oscillators $$\begin{gathered} \Phi=\Phi_0+i\Phi^\mu\alpha^\dagger_\mu+\Phi^{\mu\nu}\alpha^\dagger_\mu\alpha^\dagger_\nu+\cdots,\\ F=F_0+iF^\mu\alpha^\dagger_\mu+F^{\mu\nu}\alpha^\dagger_\mu\alpha^\dagger_\nu+\cdots, \\ H=H_0+iH^\mu\alpha^\dagger_\mu+H^{\mu\nu}\alpha^\dagger_\mu\alpha^\dagger_\nu+\cdots. $$ The gauge parameters are collected in $$\begin{gathered} |\Xi\rangle=(\xi_0-i\xi^\mu\alpha^\dagger_\mu+\xi^{\mu\nu}\alpha^\dagger_\mu\alpha^\dagger_\nu+\cdots)b_-|+\rangle.\end{gathered}$$ The BRST operator $Q$ is expressed in terms of the generators $$\begin{gathered} G_0=\tfrac 12 p^2,\qquad G_-=\alpha\cdot p,\qquad G_+=\alpha^{\dagger}\cdot p,\end{gathered}$$ spanning the simple algebra $$\begin{gathered} [G_-,G_+]=2G_0,\end{gathered}$$ with all other commutators zero. In terms of these generators, the BRST operator reads $$\begin{gathered} Q=c^0G_0-c^+G_+ -c^-G_- -2c^+c^-b_0.\end{gathered}$$ The action $$\begin{gathered} \label{eq:FreeAction} A=\langle\Phi|Q|\Phi\rangle,\end{gathered}$$ is invariant under the gauge transformations, $$\begin{gathered} \label{eq:FreeGaugeTransformation} \delta_\Xi|\Phi\rangle=Q|\Xi\rangle,\end{gathered}$$ as is the field equation $$\begin{gathered} \label{eq:FreeFieldEquation} Q|\Phi\rangle=0.\end{gathered}$$ When expanding the equations (\[eq:FreeAction\]), (\[eq:FreeGaugeTransformation\]) and (\[eq:FreeFieldEquation\]), everything works out nicely for the component fields, except for the fact that the theory contains auxiliary fields which cannot be solved for without introducing a further constraint. This constraint is applied to both the field and the gauge parameter $$\begin{gathered} T|\Phi\rangle=0,\qquad T|\Xi\rangle=0,\end{gathered}$$ where $T$ is the operator $$\begin{gathered} T=\tfrac 12 \alpha\cdot\alpha+b_+c^-.\end{gathered}$$ When expanded, the constraint equations yield the double tracelessness constraint for component fields of spin $s\geq 4$ and the tracelessness constraint for the corresponding component gauge parameters. The constraints are needed in order to get the correct number of physical degrees of freedom. Including the constraints, this formulation precisely reproduces the Fronsdal equations [@Fronsdal1978] for higher spin gauge fields. The free field theory is however still gauge invariant without imposing these constraints, and as discussed in [@AKHB2007a], the constraints can actually be discarded with. The presence of the extra fields are then instead controlled by two global symmetries of the theory [@AKHB1988]. Other approaches to circumventing the tracelessness constraints have been proposed [@FranciaSagnotti2003a]. With some extra work this theory can be formulated in terms the Batalin–Vilkovisky field/antifield formalism. We will return to this issue in Section \[sec:RotBRST-BVMA\]. Perhaps just as interesting as the technical reconstruction of Fronsdal’s free higher spin theory in terms of the BRST formalism, is the hint at an underlying mechanical model. In the first two schemes outlined above, it is not clear what the bases [*are*]{}. In the first case, generalizing from spin 1 $\xi_aT^a$ and spin 2 $\xi^\mu\partial_\mu$, (where for spin 2 we really should consider a full Poincar[é]{} generator $\xi^\mu\partial_\mu+{1\over 2}\xi^{ij}M_{ij}$) it is clear that we have no conceptual understanding what the higher derivative operators actually do. Perhaps somewhat vaguely we could say that we are re-using spacetime tangent space, not introducing anything new. The same goes for Fronsdal’s approach. Although clever, it is clear that the scheme again is just re-using spacetime, but now its cotangent superstructure. In contrast to this, introducing oscillators do add a something new, indeed new dimensions apart from spacetime itself. It is possible construct mechanical models that produces the first class constraints behind the above BRST construction of the field theory [@AKHB1987b; @HenneauxTeitelboim1989a]. During the 1960’s and 70’s there were other, parallel approaches to strong interactions apart from string theory (before QCD and asymptotic freedom swept the table). These also involved infinite towers of massless particles/fields and ideas about underlying composite mechanical models [@EighthNobelSymposium; @CasalbuoniLonghi1975], though at that time the interest was not primarily in the higher spin gauge fields. Deformation vs. gauging {#sec:DvsG} ======================= There are two main approaches to introducing interactions into a massless free spin 1 theory. Either gauge a global but non-Abelian Lie algebra, or deform local but Abelian gauge symmetries. It is of some interest to reconsider these approaches in order to get a hold on what is involved in generalizing to all spins. One problem with the approaches described so far, apart from the fact that very few solid positive results on interactions have been achieved, is that they suffer from a weak conceptual underpinning. In my opinion this is also true for the much more successful Vasiliev approach, although in this case there are quite a few technical circumstances that lend some basic strength to the approach. We will now try to understand this in a simpleminded straightforward way, and in the process clarify some of the connections between the AdS and the Minkowski space (MiS) formulations of higher spin theory. The Vasiliev approach is well described in the literature so I will not attempt yet another review, but instead focus some relevant points, perhaps in a somewhat idiosyncratic way. What are the higher spin algebras? ---------------------------------- The Vasiliev approach to higher spins seems simple enough. Just take a higher spin algebra and gauge it. Now, from where do we get the higher spin algebras? That’s also easy, just take the algebra $so(3,2)$ or any of its higher dimensional and/or supersymmetric versions and form powers of its generators to any order. What results is an infinite dimensional associative algebra. The rest is technicalities. These are amply covered in the literature, what we want to do here is something more basic. The question is: why would this work? And why would it work in AdS and not in MiS. Let $T^a$ be the generators of an $N$-dimensional Lie algebra. Forming free powers of these generators and collecting these into the set $$\begin{gathered} \label{eq:InfiniteDimAssociativeAlgebra} {\bf A}=\{T^{a_1}T^{a_2}\cdots T^{a_{i-1}}T^{a_i}\}_{i=1}^\infty\end{gathered}$$ we can promote [**A**]{} to an associative algebra through the universal enveloping algebra construction. This involves forming equivalence classes of generators by modding out by the elements of the ideal generated by elements of the form $[T^a,T^b]$ and then taking care of automorphisms. Non-compact Lie algebras have oscillator representations and these are easily seen to be infinite dimensional. Thus when we use a non-compact Lie algebra such as $so(3,2)$ to promote it to an infinite dimensional associative algebra using the construction (\[eq:InfiniteDimAssociativeAlgebra\]), what we are really doing is to use its infinite-dimensional representations as a kind of basis for an infinite dimensional algebra. It is worthwhile to do this in some detail for the basic case of $so(3,2)$. ### The higher spin algebra $\boldsymbol{hso(3,2)}$ {#the-higher-spin-algebra-boldsymbolhso32 .unnumbered} Let us then consider a pair of independent oscillators $(a,a^\dagger)$ and $(b,b^\dagger)$ with $[a,a^{\dagger}]=1$ and $[b,b ^{\dagger}]=1$. Using these we can build low dimensional Lie algebras. Thus for the compact algebra $su(2)$ we define the generators $$\begin{gathered} J_+= b^{\dagger}a,\qquad J_-=a^{\dagger}b,\qquad J_3=\tfrac 12 (b^{\dagger}b-a^{\dagger}a).\end{gathered}$$ spanning $$\begin{gathered} [J_3,J_+]=J_+,\qquad [J_3,J_-]=-J_-,\qquad [J_+,J_-]=2J_3.\end{gathered}$$ To the $su(2)$ generators we can add the number operator $$\begin{gathered} E=\tfrac 12 (b^{\dagger}b+a^{\dagger}a+1),\end{gathered}$$ which commutes with the rest of the generators. This turns $su(2)$ into $u(2)$. Next, using just one oscillator, we can build the non-compact $sp(2)$ algebra $$\begin{gathered} S_+ =\tfrac 12 a^{\dagger}a^{\dagger},\qquad S_- = \tfrac 12 aa,\qquad S_3 = \tfrac 12 \big(a^{\dagger}a+\tfrac 12\big).\end{gathered}$$ The algebra is $$\begin{gathered} [S_3,S_+]=S_+,\qquad [S_3,S_-]=-S_-,\qquad [S_+,S_-]=-2S_3.\end{gathered}$$ Starting with $sp(2)$ it is clear that we get an infinite dimensional representation over an oscillator ground state $|0\rangle$ with $a|0\rangle=0$. Also, forming powers of the generators $S_+$, $S_-$, $S_3$ we get an infinite dimensional algebra represented over this same representation space. This is all well known including the necessary brush up details. Now what happens in the $su(2)$ case? Define the double ground state $|0_a,0_b\rangle\doteq|0_a\rangle|0_b\rangle$ with the obvious properties. For short we just write $|0,0\rangle$. Excited states are written $|n_a,n_b\rangle$. The Fock space ${\bf F}_{su(2)}$ built upon the spin 0 ground state $|0,0\rangle$ gets a natural grading by spin. Using standard notation $|jm\rangle$ for angular momentum states we get $$\begin{gathered} {\bf F}_{su(2)} = \{|0,0\rangle\}\cup\{(|0,1\rangle,|1,0\rangle\}\cup\{|0,2\rangle,|1,1\rangle,|2,0\rangle)\}\cup\cdots\nonumber\\ \phantom{{\bf F}_{su(2)}}{} =\bigcup_{n=0,1,2,\ldots}\big(\cup_{n_a+n_b=n}\{|n_a,n_b\rangle\}\big) =\bigcup_{j=0,{1\over2},1,\ldots}(\oplus_{m=-j}^{m=j}|jm\rangle).\label{eq:SU2FockSpace}\end{gathered}$$ We can now consider powers of the $su(2)$ generators, say $(J_+)^p$ and $(J_-)^q$. Then acting with these operators on spin ${n\over 2}$ states $|n_a,n_b\rangle$ we either get zero or a new spin ${n\over 2}$ state (when $p,q\leq n$), i.e. we stay within the same subspace, or we get zero (when $p,q>n$). Furthermore, since all states are eigenstates of $J_3$, powers of $J_3$ also stay within the same spin ${n\over 2}$ subspace. So in this way we cannot build an infinite dimensional algebra based on the compact $su(2)$ algebra that acts transitively on the full Fock space of states. Let us now turn to the non-compact algebra $so(3,2)$. Its ten generators splits into three components according to $so(3,2)\mapsto g^{-1}\oplus g^{0}\oplus g^{+1}$ where $$\begin{gathered} g^{-1} = \{L_-^-,L_+^-,L_3^-\},\qquad g^0= \{E,J_+,J_-,J_3\},\qquad g^{+1}= \{L_+^+,L_-^+,L_3^+\}.$$ The $g^0$ generators are precisely the already defined $u(2)$ generators. The rest are expressed in terms of the oscillators as $$\begin{gathered} L_-^-=aa,\qquad L_+^-=bb,\qquad L_3^-=ab,\qquad L_+^+=a^\dagger a^\dagger,\qquad L_-^+=b^\dagger b^\dagger,\qquad L_3^+=a^\dagger b^\dagger.$$ Clearly, the $g^{-1}$ are lowering operators and the $g^{+1}$ are raising operators, and the overall structure of the algebra is $$\begin{gathered} [g^m,g^m]\subseteq g^{m+n},\qquad m,n=\{\pm1,0\}.\end{gathered}$$ Using the raising and lowering operators we can now step up and down in the full Fock space ${\bf F}_{su(2)}$. This Fock space is actually the ${\rm Di}\oplus{\rm Rac}$ Fock space of states $|e,j\rangle$ with ground state $|{1\over 2},0\rangle$ and with the dispersion equation $e=j+{1\over 2}$ [@Dirac1963; @FlatoFronsdal1978a; @FlatoFronsdal1980a; @GunaydinSacliogu1982a; @GunaydinPreprint1989]. The subset of states $\{|n_a,n_b\rangle:n_a+n_b=n\}$ has $e={1\over 2}(n_a+n_b+1)$ and $m={1\over 2}(n_a-n_b)$, where as in (\[eq:SU2FockSpace\]), $m$ is the $J_3$ quantum number ranging between $-j$ and $j$. Finally, modulo technical details, the higher spin algebra $hso(3,2)$ is built by simply taking all positive powers of the $so(3,2)$ generators. It becomes an infinite-dimensional Lie algebra transitively represented on the ${\rm Di}\oplus{\rm Rac}$ weight space (alternatively the Fock space ${\bf F}_{su(2)}$). So what are the higher spin algebras? It seems that in four spacetime dimensions, the simplest higher spin algebra is that of an infinite-dimensional transformation algebra acting on the direct sum of all representations of the 3D angular momentum algebra. The crucial point is that the algebra connects states of different spin. What we get is a concrete realization of the universal enveloping algebra of $so(3,2)$. A question is now, can this be done for the Poincar[é]{} algebra? Analyzing how the Poincar[é]{} algebra is embedded in $so(3,2)$ it becomes clear that the $so(3,2)$ raising and lowering operators ${L^{\pm}_a}$ are linear combinations of Poincar[é]{} translations and boosts with the deformation parameter $\epsilon={1\over\sqrt\Lambda}$ as coefficient ($a$ is a space index in the set $\{1,2,3\}$ or $\{+,-,3\}$) $$\begin{gathered} L^{\pm}_a=\epsilon P_a\pm i L_{a0}.\end{gathered}$$ Thus when the Wigner–Inönü contraction $\Lambda\rightarrow 0$ is performed, the raising and lowering operators break up, and we lose much of the interesting AdS structure. Still there are remnants of the structure in MiS as is apparent from in the BRST oscillator expansion (\[eq:OscillatorExpansionFields\]) of higher spin gauge fields. An interesting line of research would be to extend the MiS Lorentz algebra $so(3,1)$ to an infinite algebra. Gauging ------- Whereas the gauge theory approach to spin 1 interactions is fairly straightforward, indeed it is the paradigmatic example, gauging approaches to interacting spin 2 has always been plagued by conceptual (and technical) difficulties. As a backdrop, let us briefly run through the spin 1 case. We have a non-Abelian Lie algebra represented by matrices $T_a$ $$\begin{gathered} [T_a,T_b]=f_{ab}^{\,\,\,\,c}T_c,\end{gathered}$$ and a vector of matter fields $\varphi$ transforming in some representation $$\begin{gathered} \delta\varphi=\epsilon\varphi=\epsilon^aT_a\varphi,\end{gathered}$$ and a matter Lagrangian which is invariant under these transformations. Making the parameters $\epsilon$ local: $\epsilon(x)$, we find the problem that spacetime derivatives of the fields transform in-homogeneously $$\begin{gathered} \delta(\partial_\mu\varphi)=\partial_\mu(\delta\varphi)=\epsilon^aT_a\partial_\mu\varphi+(\partial_\mu\epsilon^a)T_a\varphi.\end{gathered}$$ The remedy is introducing new gauge fields $A_\mu=A_\mu^aT_a$ transforming as $$\begin{gathered} \delta A_\mu=[\epsilon,A_\mu]-\partial_\mu\epsilon.\end{gathered}$$ Then the covariant derivative $D_\mu=\partial_\mu+A_\mu$ transforms homogeneously. Replacing ordinary derivatives with covariant ones then restores invariance to the matter Lagrangian. However, in our present context, the matter Lagrangian is just a crutch, what we are after is non-linear dynamics for the gauge field itself. The solution is quite simple and beautiful. Commuting covariant derivatives we find the field strengths $F_{\mu\nu}$ according to $$\begin{gathered} \label{eq:FieldStrengths} F_{\mu\nu}=[D_\mu,D_\nu]=\partial_\mu A_\nu-\partial_\nu A_\mu+[A_\mu,A_\nu].\end{gathered}$$ These transform as $$\begin{gathered} \delta F_{\mu\nu}=[\epsilon,F_{\mu\nu}],\end{gathered}$$ and it finally transpires that a non-linear, gauge-invariant Lagrangian can be written $$\begin{gathered} {\cal L}=-{1\over 4g^2} F_{\mu\nu}^aF^{\mu\nu}_a,\end{gathered}$$ from which we read off that the interacting theory is a deformation of free spin 1 gauge theory running up to cubic and quartic order in the interaction. (Note that in the above brief outline the gauge coupling constant $g$ has been absorbed into the fields and parameters, it can be restored by the substitutions $A\mapsto gA,\epsilon\mapsto g\epsilon$.) What is the strong point of this? Clearly the fact that the correct non-linearity is forced upon us through the equation (\[eq:FieldStrengths\]). It is like following a recipe. The weak point is finding the correct spin 1 Lagrangian, it looks simple enough here, but this step does not generalize easily to higher spin. In the case of spin 2, the trouble starts at the very first step, choosing the gauge group. Having General Relativity in the back of our minds, we think about the local tangent spaces. In a frame, or vierbein, formulation this becomes very concrete as local Poincar[é]{} transformations. But we are running somewhat ahead of ourselves. Start with an (active) infinitesimal Poincar[é]{} transformation $$\begin{gathered} \delta x^\mu=\epsilon^\mu_{\,\,\,\nu} x^\nu+\epsilon^\mu,\end{gathered}$$ and correspondingly, a set of matter fields $\varphi$ transforming as $$\begin{gathered} \label{eq:PoincareOnMatterField} \delta\varphi=\tfrac 12 \epsilon^{ij}S_{ij}\varphi-\epsilon^\mu\partial_\mu\varphi.\end{gathered}$$ Many things have slipped in here. In the first term we have changed indices from (curved) spacetime indices $\mu$, $\nu$ to tangent space Lorentz indices $i$, $j$. So far we can see that we are really just generalizing the spin 1 recipe in a straightforward manner. The role of the $T_a$ matrices are now played by the Lorentz $so(3,1)$ matrices $S_{ij}$. In the second term we see that corresponding to the translation part of the Poincar[é]{} transformations, we get generators that are spacetime derivatives. Anyone who has played around with these equations for a while is likely to suffer from some confusion, and it gets worse. As soon as the parameters $\epsilon^{ij}$ and $\epsilon^\mu$ are made into local functions of the coordinates $x^\mu$ the distinction between local translations and local Lorentz transformations becomes blurred. Translations with a local $\epsilon^\mu(x)$ already contains all local coordinate transformations generated by the vector field $\epsilon^\mu(x)\partial_\mu$. However, if the local Lorentz transformations $\epsilon^{ij}S_{ij}$ are discarded, the transformations on the fields (\[eq:PoincareOnMatterField\]) become ambiguous. Correspondingly, there are various approaches in the literature, gauging the Lorentz group [@Kibble], gauging the translations only, or gauging the full Poincar[é]{} group. A few more papers from the heydays of this approach to gravity are [@Utiyama; @MacDowellMansouri1977; @MansouriChang; @Grensing; @HehlHeydeKerlick]. Following the review article [@KibbleStelle1986] and proceeding to gauge the full Poincar[é]{} group, we again find that the derivate of a matter field transforms inhomogeneously $$\begin{gathered} \delta\partial_\mu\varphi=\tfrac 12 \epsilon^{ij}S_{ij}\partial_\mu\varphi-\epsilon^\nu\partial_\nu\partial_\mu\varphi+ \tfrac 12 (\partial_\mu\epsilon^{ij})S_{ij}\varphi-(\partial_\mu\epsilon^\nu)\partial_\nu\varphi.\end{gathered}$$ The third term can be taken care of by introducing a gauge field $\omega^{ij}_\mu$ with an inhomogeneous transformation term $-\partial_\mu\epsilon^{ij}$, and correspondingly we have a covariant derivative $$\begin{gathered} \nabla_\mu=\partial_\mu+\tfrac 12 \omega^{\,\,\,ij}_\mu.\end{gathered}$$ The last term must be treated in a different way since it involves derivatives $\partial_\mu$ instead of matrices. Here we do a multiplicative gauging introducing the vierbeins $e^{\,\,\mu}_i$ $$\begin{gathered} D_i=e^{\,\,\mu}_i\nabla_\mu.\end{gathered}$$ The rest of the story from here on involves commuting the covariant derivatives to find curvature $R_{ij}^{\;\;\;kl}$ and torsion tensors $T_{ij}^{\;\;k}$. Due to the multiplicative gauging, the curvature tensor is second order in spacetime derivatives (in contrast to the spin 1 field strengths which are first order in derivatives), and this is reflected in the gravity Lagrangian being expressed as $$\begin{gathered} {\cal L}\simeq R_{ij}^{\;\;\;ij}.\end{gathered}$$ The theory so obtained is almost Einstein gravity, the difference is the presence of torsion [@Kibble; @HehlHeydeKerlick]. This very (indeed) brief review of the gauging approaches to spin 1 and 2, shows disturbing differences between the two cases. There certainly isn’t any standard recipe to generalize. In a very loose language we could say that Yang–Mills theory is basically very algebraic whereas Gravity is very geometric. One way to iron out the differences would be to geometrize Yang–Mills, this is precisely what one does in the modern fiber bundle approach. Another way would be to make Gravity more algebraic. As we saw, one source to the problems came from the translation part of the Poincar[é]{} group which gives rise to derivatives as gauge generators. However in the AdS group $SO(3,2)$ the translation generators $P_\mu$ are just a subset of the $so(3,2)$ generators according to $P_\mu=\sqrt \Lambda M_{\mu 4}$ with $[P_\mu,P_\nu]=-i\Lambda M_{\mu\nu}$. So one could attempt to set up an $SO(3,2)$ gauge theory of gravity. This was done by Kibble and Stelle [@StelleWest1980; @KibbleStelle1986] (see also [@MacDowellMansouri1977]). Standard Einstein gravity is recovered upon spontaneously breaking $SO(3,2)$ to Poincar[é]{} corresponding to a Wigner–Inönü contraction $\Lambda\rightarrow0$. This could be thought of as making gravity more algebraic. The Vasiliev approach to higher spin follow this road. The core of the Vasiliev theory ------------------------------- What the Vasiliev theory of higher spins does is basically that it extends the gauge algebra $so(3,2)$ to the higher spin algebra $hso(3,2)$ (and correspondingly in other dimensions with or without supersymmetry) and then by generalizing the techniques of MacDowell and Mansouri [@MacDowellMansouri1977]) and Stelle and West [@StelleWest1980], it manages to derive field equations for interacting higher spin gauge fields coupled to gravity and Yang–Mills. This is a very impressive achievement and it involves a lot of technical details. Still it has so far not been possible to write down a unifying Lagrangian for this theory, and partly for this reason it must be said that the approach is still not entirely understood. The field equations of the Vasiliev approach are written as a generalization of the Maurer–Cartan equations for a Lie algebra (for good reviews, see [@BekaertCnockaertIazeollaVasiliev2005; @JohanEngqvist] which we follow here) $$\begin{gathered} \label{eq:MaurerCartan1} R^a\doteq d\sigma^a+\tfrac 12 f_{bc}^{\,\,\,\,a}\sigma^b\wedge\sigma^c=0,\end{gathered}$$ and where the $\sigma^a$ are 1-forms on the Lie group manifold. The Jacobi identity follows from the integrability conditions $dR^a=0$ using $dd=0$. This formulation is completely equivalent to the more common formulation in terms of vector fields (or generators) $T_a$ satisfying the usual Lie brackets $[T_a,T_b]=f_{ab}^{\;\;\;c}T_c$ and with the inner product $\langle\sigma^a,T_b\rangle=\delta^a_ b$. The Maurer–Cartan equations are invariant under gauge transformations $$\begin{gathered} \delta\sigma^a=D\epsilon^a\doteq d\epsilon^a+f_{bc}^{\,\,\,\,a}\sigma^b\wedge\epsilon^c.\end{gathered}$$ This is then generalized by introducing, a possibly infinite, collection of $(p+1)$-form curvatures $R^\alpha$ defined as $$\begin{gathered} \label{eq:MaurerCartan2} R^\alpha\doteq d\Sigma^a+F^\alpha(\Sigma^\beta)=0,\end{gathered}$$ where $$\begin{gathered} \label{eq:DefineStructureMC} F^\alpha(\Sigma^\beta)\doteq\sum_{k=1}^\infty f_{\;\;\beta_1\dots\beta_k}^{\alpha}\Sigma^{\beta_1}\wedge\cdots\wedge\Sigma^{\beta_k},\end{gathered}$$ where the $\Sigma^\beta$ are $p$-forms on the Lie group manifold. Here $f_{\beta_1\dots\beta_k}^{\alpha}$ are the structure constants of a, possibly infinite-dimensional, algebra. The form degrees must match, i.e. $\sum\limits_{i=1}^k p_{\beta_i}=p_\alpha+1$. Now the integrability conditions read $dF^\alpha(\Sigma^\beta)=0$, or $$\begin{gathered} \label{eq:IntegrabilityConditionMC} F^\beta\wedge{\partial F^\alpha\over\partial\Sigma^\beta}=0,\end{gathered}$$ yielding generalized Jacobi identities for the structure constants $f_{\beta_1\dots\beta_k}^{\alpha}$. The gauge transformations also generalize to $$\begin{gathered} \delta\Sigma^\alpha=d\epsilon^\alpha-\epsilon^\beta\wedge{\partial F^\alpha\over\partial\Sigma^\beta},\end{gathered}$$ under which the generalized curvatures of (\[eq:MaurerCartan2\]) stays invariant. This kind of structure was initially developed in the context supergravity and termed Cartan Integrable Systems [@DAuriaFre1982a], later renamed as Free Differential Algebra in accordance with the corresponding mathematical constructs [@Sullivan1977]. The actual field equations are formulated in terms of two differential forms, one zero-form $\Phi=\Sigma^0$ and one one-form $A=\Sigma^1$ which in their turn are infinite expansions in terms of oscillators, much like the expansions above, but with important differences. Vasiliev’s expansions are in terms of both creation and annihilation operators and will therefore span infinite-dimensional algebras (given some restrictions), indeed the higher spin algebras. This is the abstract scheme employed in the Vasiliev approach. Of course, we haven’t even touched on the mass of technicalities involved in actually setting up higher spin gauge theory in this framework. Some helpful papers are [@BekaertCnockaertIazeollaVasiliev2005; @JohanEngqvist; @Vasiliev2005a] where also many further references can be found. Here we will restrict ourselves to pointing out how strongly homotopy Lie algebras emerge. Consider the operator $$\begin{gathered} \label{eq:KOperator} K=F^\alpha(\Sigma){\partial\over \partial\Sigma^\alpha}.\end{gathered}$$ Then calculating the square of this operator we get $$\begin{gathered} K^2=\tfrac 12 K\wedge K=\left(F^\beta\wedge{\partial F^\alpha\over\partial\Sigma^\beta}\right){\partial\over\partial\Sigma^a}=0,\end{gathered}$$ which is zero by (\[eq:IntegrabilityConditionMC\]). From here the step to the algebraic structure of a $L_\infty$ algebra is short. Already in the generalization of (\[eq:MaurerCartan1\]) to (\[eq:MaurerCartan2\]) and the definition (\[eq:DefineStructureMC\]) we see a hint of the higher order brackets of an sh-Lie algebra. This becomes more clear if we write the Maurer–Cartan equations in an abstract, coordinate-free way as $d\sigma=-{1\over 2}[\sigma,\sigma]$. Generalizing this recklessly, we would consider multibrackets $[\Sigma,\Sigma,\ldots,\Sigma]$ and the Jacobi identities for the ordinary Lie bracket $[\cdot,\cdot]$ would generalize to the sh-Lie identities. Let us step back and do this in some more detail. Strongly homotopy Lie algebras and higher order brackets -------------------------------------------------------- There are a few variants of the basic definitions of strongly homotopy Lie algebras in the literature (see for example [@LadaStasheff1993a; @LadaMarkl1995a]), but the following, mildly technical, is sufficient for our purpose to bring out the connection to both the Vasiliev formalism and to the product identities of the BRST approach described below. ### Definition {#definition .unnumbered} Consider a ${\bf Z}_2$ graded vector space $V=V_0\oplus V_1$ over some number field, and denote the elements by $x$. The grading is given by $\varrho$ with $\varrho(x)=0$ if $x\in V_0$ and $\varrho(x)=1$ if $x\in V_1$. $V$ is supposed to carry a sequence of $n$-linear products denoted by brackets. The graded $n$-linearity is expressed by $$\begin{gathered} [x_1,\ldots,x_n,x_{n+1},\dots,x_m]=(-)^{\varrho(x_n)\varrho(x_{n+1})}[x_1,\ldots,x_{n+1},x_n,\dots,x_m],\\ \lbrack x_1,\ldots,a_n x_n+b_n x'_n\,\ldots,x_m\rbrack\nonumber\\ \qquad{}= a_n(-)^{\iota(a_n,n)}\lbrack x_1,\ldots,x_n,\ldots,x_m\rbrack+b_n(-)^{\iota(b_n,n)}\lbrack x_1,\ldots,x'_n,\ldots,x_m\rbrack, $$ where $\iota(a_n,n)={\varrho(a_n)(\varrho(x_1)+\cdots +\varrho(x_{n-1})}$. The defining identities (or “main” identities) for the algebra are, for all $n\in{\bf N}$ $$\begin{gathered} \label{eq:SHDefiningEquations} \sum_{{k=0}\atop{l=0}}^{k+l=n}\sum_{\pi(k,l)}\epsilon(\pi(k,l))\lbrack\lbrack x_{\pi(1)},\ldots,x_{\pi(k)}],x_{\pi(k+1)},\ldots,x_{\pi(k+l)}\rbrack=0,\end{gathered}$$ where $\pi(k,l)$ stands for $(k,l)$-shuffles. A $(k,l)$-shuffle is a permutation $\pi$ of the indices $1,2,\dots$, $k+l$ such that $\pi(1)<\cdots<\pi(k)$ and $\pi(k+1)<\cdots<\pi(k+l)$. $\epsilon(\pi(k,l))$ is the sign picked up during the shuffle as the points $x_i$ with indices $0\leq i\leq k$ are taken through the points $x_j$ with indices $k+1\leq j\leq l$. This is just the normal procedure in graded algebras. The bracket (or braces) notation is common. Note that the $n$-ary brackets are all independent and abstract, not to be thought of as deriving from some underlying product (just as the abstract Lie bracket $[\cdot,\cdot]$ need not derive from a product). As discussed in [@Voronov2003a], other types of gradings can be considered. The ${\bf Z}_2$ grading employed here can be thought as providing room for a BRST-type algebra with even fields and odd gauge parameters. In the BRST-BV reformulation [@AKHB2007a] (see Section \[sec:RotBRST-BVMA\] of the present paper) the mechanical and field theory ghosts conspire to make all objects even (fields, anti-fields, ghosts and anti-ghosts) so in that case we can in fact do with an ungraded algebra. In order to accommodate the FDA of Vasiliev, a grading with respect to form degree should be defined. Running the risk of pointing out something that is obvious, let us note that even in the case where we have an underlying associative product, such as when we work with matrix algebras, nothing prevents us from calculating higher order brackets and see what we get. What we would get is roughly linear combinations of elements in the vector space that the algebra is built upon, and these linear combinations would define for us generalized structure coefficients. Such higher order Lie algebras has been investigated in [@AzcrragaBueno1997a]. They obey generalized Jacobi identities similar in structure to the sh-Lie identities (of which they are a special case), but then as a consequence of the underlying associative product. Next following [@Voronov2003a] we can introduce generating functions for the sh-Lie structure. As pointed out there, due to the symmetry and linearity properties, the full sh-Lie structure is fixed by the algebra of even and coinciding elements, $\Sigma$ say. We can indeed think of the $\Sigma$ as vectors in one huge graded vector space, for example graded by form degree, corresponding to the forms of the Vasiliev theory. Next we write equation (\[eq:DefineStructureMC\]) in terms of brackets $$\begin{gathered} K=\sum_{k\geq 0}{1\over k!}\underbrace{[\Sigma,\Sigma,\ldots,\Sigma]}_{k}.\end{gathered}$$ $K$ is itself an element of the vector space and we can consider it as a formal vector field $$\begin{gathered} K=F^\alpha(\Sigma){\partial\over\partial\Sigma^\alpha},\end{gathered}$$ which is what we had before (\[eq:KOperator\]). But now computing the square $K^2$ using its definition in terms of the brackets, we get Jacobiators $$\begin{gathered} J^n(\Sigma,\ldots,\Sigma)=\sum_{l=0}^{n}{n!\over l!(n-l)!}[\,\underbrace{[\Sigma,\ldots,\Sigma]}_{n-l},\underbrace{\Sigma,\ldots,\Sigma}_{l}\,],\end{gathered}$$ for even elements $\Sigma$. In terms of these Jacobiators the “generalized Jacobi identities” or simply “main” identities from the definition of an sh-Lie algebra become $J^n=0$ for all $n$, or $$\begin{gathered} J=\sum_{n\geq 0}{1\over n!}J^n=K^2=0.\end{gathered}$$ This, admittedly hand-waving, argument shows that there is a connection between sh-Lie algebras and FDA’s. This is often alluded to in the literature (see for example [@BekaertCnockaertIazeollaVasiliev2005] but I haven’t seen any rigorous proof. But there cannot really be any doubt that the technical details can be supplied to make the connection precise. According to [@Vasiliev2005a], any FDA can be obtained from an sh-Lie algebra by fixing the form degree $p_\alpha$ of the $\Sigma^\alpha$. Implicitly, this is what we have done in the argument above. The same formal structure can also be seen in abstract approaches to BRST-BV theory. As in [@BarnichGrigoriev2005a] we can write an odd vector field $K$ $$\begin{gathered} K=\Omega^b_{\;a}\psi^a{\partial\over\partial\psi^b}+U^c_{\;ab}\psi^a\psi^b {\partial\over\partial\psi^c}+U^d_{\;abc}\psi^a\psi^b\psi^c{\partial\over\partial\psi^d}+\cdots,\end{gathered}$$ where $K$ could be either a BRST-operator $Q$ or a BRST-BV operator $s$ depending on the context. In both case, the equation $K^2=0$ generates an sh-Lie algebra. Contemplating all this one might wonder whether there are more abstract common algebraic structures underlying both the Vasiliev approach and for instance the constructions outlined in Section \[sec:HSBG\], and in particular the BRST-BV approach of [@AKHB2007a] (see Section \[sec:RotBRST-BVMA\] in the present paper) which superficially look very different. It seems that the language of category theory, and in particular the language of operads, could furnish us with such an abstract framework. Some preliminary remarks on this can be found in Section \[sec:US\]. For more connections between the Vasiliev the BRST formalism, see [@BarnichGrigorievSemikhatovTipunin2004], where it is shown that both the free Vasiliev theory and the BRST theory can be derived from a common “parent field theory”. Deforming --------- The deformation approach starts out with a free gauge theory. We have the action and the Abelian gauge transformations and the object is to deform the action and the gauge transformations by non linear terms, still retaining gauge invariance. This works nicely for spin 1 as was shown a long time ago in [@OgievetskyPolubarinov1963]. The analogous deformation approach to spin 2 is considerable more complicated, both technically and conceptually, and there is a long list of classical papers treating this problem [@Gupta1952; @Gupta1954; @Kraichnan1955; @Wyss1965; @Thirring1961; @Feynman1962a; @Deser1970; @BoulwareDeserKay1979; @FangFronsdal1979; @Deser1987]. No wonder then, perhaps, that the higher spin interaction problem is even more convoluted. The list of references pertaining to this problem is quite long, for a reasonable subset, see [@AKHB2005a]. Review of the BRST-BV Minkowski approach {#sec:RotBRST-BVMA} ======================================== As a modern example of the deformation method, let us describe the concrete Fock complex vertex implementation of the BRST-BV approach to higher spin gauge self interactions in MiS in a top-down fashion. The exposition here will be brief, a thorough description can be found in the recent paper [@AKHB2007a], which also contains references to the relevant theoretical background. For a general discussion of the BV-deformation approach to interactions see [@BarnichHenneaux1993a; @Henneaux1997a]. Thus we start by writing out the master action to all orders in formal perturbation theory $$\begin{gathered} \label{eq:MasterAction} S=\langle\Psi|Q|\Psi\rangle|_{{\rm gh_m}=0}+\sum_{n=3}^\infty g^{n-2}\langle\Psi|^{\otimes n}|{\cal V}_n\rangle|_{{\rm gh_m}=0}.\end{gathered}$$ Here $|\Psi\rangle$ denotes a formal sum of ghost, field, antifield, antighost components $$\begin{gathered} |\Psi\rangle=|{\cal C}\rangle\oplus|\Phi\rangle\oplus|\Phi^\#\rangle\oplus|{\cal C}^\#\rangle.\end{gathered}$$ The components $|{\cal C}\rangle$, $|\Phi\rangle$, $|\Phi^\#\rangle$, $|{\cal C}^\#\rangle$ live in the Fock complex of mechanics oscillators and ghosts described in Section \[subsec:OscBas\]. Actually, the field $|\Phi\rangle$ is precisely the higher spin field whereas the field theory ghost field $|{\cal C}\rangle$ replaces the gauge parameter $|\Xi\rangle$. The $\#$-decorated components are the corresponding anti-field and anti-ghost. When we want to refer to the fields without having this concrete realization in mind, we just write $\Psi$. In this formulation there are no trace constraint operators $T$. Instead, the theory is subject to two global symmetries $$\begin{gathered} \delta_{\cal P}|\Psi\rangle=i\epsilon{\cal P}|\Psi\rangle,\qquad \delta_{\cal T}|\Psi\rangle=i\epsilon{\cal T}|\Psi\rangle,$$ where the definition of the operators ${\cal P}$ and ${\cal T}$ can be found in [@AKHB2007a]. It then follows that $\delta_{\cal P}\langle\Psi|Q|\Psi\rangle=\delta_{\cal T}\langle\Psi|Q|\Psi\rangle=0$. In this formalism there are, for a given primary spin $s$ field, a further spin $s-2$ field (corresponding to the trace of the spin $s$ field) that is not solved for by the trace constraint. For example, the primary spin 5 field will be accompanied by a new secondary spin 3 field. It is interesting to note that cubic spin 3 interactions with non-minimal number of derivatives have been found [@BekaertBoulangerCnockaert2006a], i.e. with five derivatives instead of three. This is expected for a secondary spin 3 field accompanying the primary spin 5 field. The vertex operators $|{\cal V}_n\rangle$, encode all interaction data, i.e. all $n$-point vertex data as well as all higher order structure function data. The BRST invariance of the classical theory is now expressed by the master equation where $(\cdot,\cdot)$ is the anti-bracket of BV theory. Explicitly calculating $(S,S)=0$ order by order in vertex order $n$ yields $$\begin{gathered} \label{eq:QV3} \sum_{r=1}^3 Q_r|{\cal V}_{3}\rangle=0, \\ \label{eq:QVn} \sum_{r=1}^{n+1}Q_r|{\cal V}_{n+1}\rangle=-\sum_{p=0}^{\lfloor(n-3)/2\rfloor}|{\cal V}_{p+3}\rangle\diamond|{\cal V}_{n-p}\rangle,\end{gathered}$$ where $\diamond$ denotes a certain symmetrized contraction in the mechanics Fock complex. These equations encode the structure of a $L_\infty$ algebra as will be shown in Section \[sec:US\]. Extending the global symmetries to the interacting theory yield the following conditions on the vertices $$\begin{gathered} \sum_{r=0}^n{\cal P}_r|{\cal V}_n\rangle=0,\qquad \sum_{r=0}^n{\cal T}_r|{\cal V}_n\rangle=0.$$ In order to take into account field redefinitions, we introduce field redefinition vertices $|{\cal R}_n\rangle$ and write $$\begin{gathered} |\Psi\rangle\rightarrow|\Psi_r\rangle =|\Psi\rangle +\sum_{n=2}^\infty g^{(n-1)}\langle\Psi|^{\otimes n}|{\cal R}_{n+1}\rangle.\end{gathered}$$ Performing field transformations of this form on the free action produce fake interaction terms of the form $$\begin{gathered} \sum_{n=3}^\infty g^{(n-2)}\langle\Psi|^{\otimes n}\sum_{r=1}^nQ_r|{\cal R}_n\rangle.\end{gathered}$$ Comparing to the general form of the interactions given in equation (\[eq:MasterAction\]), we see that fake interactions can be characterized by $$\begin{gathered} |{\cal V}_n\rangle_{\rm fake}=\sum_{r=1}^nQ_r|{\cal R}_n\rangle,\end{gathered}$$ which in the language of homology is to say that fake interaction vertices are exact. Let us clarify one potentially confusing issue. The vertices $|{\cal V}_n\rangle$ provide us with products of fields, i.e. the product of $n$ fields $\Psi_1,\ldots,\Psi_n$, abstractly denoted by ${\bf pr}(\Psi_1,\ldots,\Psi_n)$, can be calculated as $$\begin{gathered} {\bf pr}(\Psi_1,\ldots,\Psi_n)\hookrightarrow\langle\Psi_1|\cdots\langle\Psi_n|{\cal V}_{n+1}\rangle\rightarrow|\Psi_{n+1}\rangle,\end{gathered}$$ i.e. contracting the $(n+1)$ vertex with $n$ bra fields produces a new ket field. The index numbers are precisely the numbering of the mechanics Fock spaces including the field momenta. The vertices enforce momentum conservation. In this context it is natural to chose the abstract products ${\bf pr}$ to be fully symmetric in the abstract fields $\Psi$, and indeed, the Fock fields $|\Psi\rangle$ can be chosen to be Grassmann even. Now, these products corresponds to what in the mathematics literature are generally denoted by $n$-ary brackets $[\cdot,\cdot,\ldots,\cdot]$, generalizing the two-bracket $[\cdot,\cdot]$ of a differential graded Lie algebra. Such a bracket is skew symmetric, rather than symmetric like our 2-products ${\bf pr}(\cdot,\cdot)$ implemented by the three-vertex $|{\cal V}_3\rangle$. This issue is resolved by noting that our product is in fact alternatingly Grassmann odd/even in the mechanics Fock complex due to the presence of $n$ odd vacua $|-\rangle$. Thus the two-product is intrinsically odd, and instead of the conventional grading $$\begin{gathered} {\bf pr}(\Psi_1,\Psi_2) = -(-)^{\epsilon(\Psi_1)\cdot\epsilon(\Psi_2)}{\bf pr}(\Psi_2,\Psi_1),\qquad \epsilon([\Psi_1,\Psi_2]) = \epsilon(\Psi_1)+\epsilon(\Psi_2),$$ we have $$\begin{gathered} {\bf pr}(\Psi_1,\Psi_2) = (-)^{\epsilon(\Psi_1)\cdot\epsilon(\Psi_2)}{\bf pr}(\Psi_2,\Psi_1),\qquad \epsilon([\Psi_1,\Psi_2]) = \epsilon(\Psi_1)+\epsilon(\Psi_2)+1,$$ and correspondingly for higher order products. This is also consistent with the recursive equations (\[eq:QVn\]) since $\epsilon(Q)=1$. Unifying structures {#sec:US} =================== It seems quite clear that the algebraic structure of strongly homotopy Lie algebras naturally crop up in various formulations of higher spin gauge theory. How can this be understood from a formalism independent way? According to the literature on the subject [@Stasheff1997a], $L_\infty$ structures was first spotted by Stasheff in the BBvD analysis [@BerendsBurgersvanDam1985] of the general higher spin interaction problem. Subsequently it was proved that given that the BBvD higher spin gauge algebra exists, it must be sh-Lie [@FulpLadaStasheff2002]. Influenced by this, and inspired by computer science thinking, I proved that the sh-Lie structure is grounded already in the syntax of any formal power series formulation of gauge theory. Could this be a hint that sh-Lie algebras, rather than being deep features of higher spins, are just superficial aspects of the formalism? As will be argued in the next section, category theory can throw light on this. In category theory we do not seek structure by peeking into the objects, but instead build structure on the outside, so to speak. Which is of course precisely what we do when we build models of physical systems. The complex objects so formed can then be re-analyzed, i.e. peeked into. Abstraction, categories and operads ----------------------------------- In my opinion, and emphasized in [@AKHB2005a], one of the main problems in higher spin theory is to control the inherent complexity. In that paper, we took a rather simpleminded syntax-semantics approach to the problem and were then able to see that the somewhat elusive connections between gauge invariance, BRST-BV formulations and strongly homotopy algebras that has often been referred to in the literature, is grounded already in the syntax of the theory. Now the theory of categories in general and the theory of operads in particular offer a solid mathematical framework for precisely this kind of situation. This should perhaps not come as a surprise since category theory is routinely applied to the syntax-semantics duality in theoretical computer science. The first problem one is confronted with upon trying to apply category theory to field theory [@Baez2004a] is: what are the objects and what are the morphisms? As briefly discussed in [@AKHB2007a], there is no one unique answer to that question, so let us start in another place. Any formulation of interacting higher spin gauge field theory will involve multi-field interaction vertices of some sort. The natural categorical correspondence then ought to be the $n$-ary multi-operations ${\cal P}(n)$ of an operad [@May1997]. Now the operad in itself only provides the axioms for the ${\cal P}(n)$’s, that is, they provide a syntax for the interactions. To get a concrete model we need a semantics, i.e. in computer science terminology: an evaluation, or in category theory language: an algebra for the operad. Indeed an algebra for an operad precisely furnishes an evaluation of the operad. Let us make this more exact. Operads and algebras for operads -------------------------------- The concept of an operad captures the idea of abstract $n$-ary operations ${\cal P}(n)$ with $n$ input lines and one output line. The ${\cal P}(n)$’s (one set for each $n\geq 1$) can be taken to be vector spaces which is just to say that they can be added and multiplied with numbers from some ground field $\bf k$. It is then natural to string these $n$-operations together to from new $n$-operations. There are then some issues to contemplate such as associativity of the compositions and permutations of input lines. This is controlled by the axioms of operads (see for example [@Leinster]). A potentially confusing issue is the type of objects figuring at the inputs of the $n$-operations. Having field theory application in mind, there are two basic options. On the one hand, aiming at modeling configuration space fields, we have just one single object $\Phi$. This then provide input to the $n$-operations and we have the rudiments of a non-polynomial field theory. On the other hand, aiming at modeling momentum space fields, the objects come labeled by an index $i$ (or $p_i$ in concrete field theory). Then we have a countable set $\{\Phi_i\}_{i=1}^\infty$ of objects. In this case we have a slight generalization of operads into multi-categories where the $n$-operations take inputs from the set of objects $\{\Phi_i\}$. Specializing to unary operations ${\cal P}(1)$ we have an ordinary category with a countable set of objects and with the ${\cal P}$’s playing the role of arrows. Another confusing issue is how an abstraction like this can really capture the details of field theory? How are the n-operations related to the concrete higher spin vertices of the BRST-BV approach? Consider the vector space ${\cal P}(n)$ of $n$-operations. Using the axioms of the multi-category (or the operad) these can be deconstructed into terms of simpler constituents, eventually coming down to a set of basic operations (classically, at the tree level). So, in the set of ${\cal P}(n)$ there is one special element (one for each $n$) which is the abstraction of the concrete vertex operator $|{\cal V}_n\rangle$. Let us denote the corresponding abstract element by ${\cal V}(n)$. Thus, the objects of the multi-category are abstract fields $\Phi_i$. We then consider the vertex operators $|{\cal V}_n\rangle$ as maps $V_n$ providing evaluations $$\begin{gathered} \label{eq:Evaluation} V_n:{\cal V}(n)\otimes \Phi^{\otimes n}\rightarrow\Phi.\end{gathered}$$ On general non-elementary operations ${\cal P}(n)$, the action of the maps $V_n$ are defined recursively in the standard way of defining maps on recursive data structures. Semantic mapping of the sh-Lie structure ---------------------------------------- To see the strength of this simple formalism let us see how the main identities of sh-Lie structure directly maps to the concrete vertex equations (\[eq:QV3\]), (\[eq:QVn\]) of Section \[sec:RotBRST-BVMA\]. Let us first write down the product identities of the sh-Lie algebra as they were derived in [@AKHB2005a] in the case of even abstract fields $\Phi$. Taking into account that the product maps are fully symmetric in all arguments, we can write the product identities as $$\begin{gathered} \sum_{{k=0,l=0}\atop \rm{cycl.\, perm.}}^{k+l=n}{\bf pr}(\Phi^k,{\bf pr}(\Phi^l))=0,\end{gathered}$$ where we write the multi-categorical morphism as ${\bf pr}(\Phi^n)$ which we think of as the abstract representation of the product of $n$ fields mapping to a new field $$\begin{gathered} \label{eq:ProductMap} \Phi^{\otimes n}\rightarrow\Phi:{\bf pr}(\Phi_1,\ldots,\Phi_n)\rightarrow\Phi_{n+1}.\end{gathered}$$ These identities are just special cases of the sh-Lie identities (\[eq:SHDefiningEquations\]), now written in terms of products instead of brackets. It will be convenient to introduce a special provision to deal with the one-product ${\bf pr}(\Phi)$ which is naturally interpreted as a linear transformation $K\Phi$. Straining the formalism a little, the one-product can be made to conform to (\[eq:ProductMap\]), if we write $$\begin{gathered} K_2\Phi_2=K_2\Phi_1\delta_{12}={\bf pr}(\Phi_1)\rightarrow\Phi_2\end{gathered}$$ by which we mean that in whatever way we compute the one-product (or linear transformation), we change index on route. This is practical when we do the same in the Fock complex $$\begin{gathered} Q_2|\Phi_2\rangle=Q_1|\Phi_2\rangle\delta_{12}=\langle\Phi_1|{\cal V}_2\rangle.\end{gathered}$$ In this way we can think of the action of the BRST operator $Q$ in terms of a 2-vertex $|{\cal V}_2\rangle$. The semantic map can now be defined. The abstract fields $\Phi$ are simply mapped to Fock complex fibers $|\Phi\rangle$, and the products ${\bf pr}_n$ are mapped to vertices $|{\cal V}_{n+1}\rangle$. Indeed using an arrow $\hookrightarrow$ to denote the semantic map, we have $$\begin{gathered} \Phi_k\hookrightarrow|\Phi_k\rangle \qquad \mbox{for}\quad k\in N,\\ {\bf pr}(\Phi_1,\ldots,\Phi_n)\hookrightarrow\langle\Phi_1|\cdots\langle\Phi_n|{\cal V}_{n+1}\rangle \qquad \mbox{for}\quad n\geq 1 .\end{gathered}$$ This is our concrete realization of the [*evaluation*]{} (\[eq:Evaluation\]) (in computer science vernacular), or [*algebra of the operad*]{} as would be the operadic notion. Note that the last equation precisely utilizes our special provisions for the one-product, i.e. the case $n=1$.We are bit lenient with the formalism here, as the $(n+1)$-th Fock space in the vertex $|{\cal V}_{n+1}\rangle$ should really be switched to a bra. As it stands, the last equation produces a ket. It could be fixed at the cost of a more cumbersome notation. With the ground so prepared we can finally apply the semantic map $\hookrightarrow$ to the left hand side of the product identities to get $$\begin{gathered} \sum_{{k=0,l=0}\atop \rm{cycl.\, perm.}}^{k+l=n}{\bf pr}(\Phi^k,{\bf pr}(\Phi^l)) \hookrightarrow \sum_{{k=0,l=0}\atop \rm{cycl.\, perm.}}^{k+l=n}\langle\Phi|^{\otimes k}(\langle\Phi|^{\otimes l}|{\cal V}_{l+1}\rangle)\cdot|{\cal V}_{k+2}\rangle.\end{gathered}$$ In writing this equation, we are freely switching bra $\leftrightarrow$ ket Fock spaces as need arise to do the contractions. With $l=n-k$ we now have $$\begin{gathered} \label{eq:ProductIdentityMapped2} \sum_{{k=0}\atop \rm{cycl. perm.}}^{n-1}\langle\Phi|^{\otimes k}\langle\Phi|^{\otimes n-k}|{\cal V}_{k+2}\rangle\cdot|{\cal V}_{n-k+1}\rangle=0.\end{gathered}$$ The sum stops at $k=n-1$ since the last term $k=n$ is zero, in the abstract product identities corresponding to ${\bf pr}(\,)=0$, which in the implementation would be $|{\cal V}_1\rangle=0$, i.e. there is no 1-vertex. Then focusing on the first ($k=0$) and next to last ($k=n-1$) terms, we see, using the conventions introduced for the 2-vertex, that they simply give us $$\begin{gathered} \sum_{r=1}^{n+1}Q_r|{\cal V}_{n+1}\rangle.\end{gathered}$$ Then the rest of the terms in (\[eq:ProductIdentityMapped2\]) pair off nicely in a similar way. Thus the ${\bf pr}(\Phi^k,({\bf pr}\Phi^{n-k}))$ term for $k\geq 1$ maps to precisely ${n\choose k}$ terms containing the vertex combination $|{\cal V}_{k+2}\rangle\cdot|{\cal V}_{n-k+1}\rangle$, while the ${\bf pr}(\Phi^{n-k-1},({\bf pr}\Phi^{k+1}))$ term maps to precisely ${n\choose n-k-1}$ terms also containing the vertex combination $|{\cal V}_{k+2}\rangle\cdot|{\cal V}_{n-k+1}\rangle$. Since ${n\choose k}+{n\choose n-k-1}={n+1\choose n-k}$ we see that we get precisely the correct number of terms to fully symmetrize $|{\cal V}_{k+2}\rangle\cdot|{\cal V}_{n-k+1}\rangle$. This is so because this contraction of vertices has $n+1$ free non-contracted indices (two of the $n+3$ being contracted). Doing the algebra carefully yields $$\begin{gathered} \sum_{k=1}^{\lfloor(n-1)/2\rfloor}|{\cal V}_{k+2}\rangle\diamond|{\cal V}_{n-k+1}\rangle,\end{gathered}$$ or, upon re-indexing the sum with $p=k-1$ $$\begin{gathered} \sum_{p=0}^{\lfloor(n-3)/2\rfloor}|{\cal V}_{p+3}\rangle\diamond|{\cal V}_{n-p}\rangle.\end{gathered}$$ Hence, collecting all the terms, we get $$\begin{gathered} \sum_{r=1}^{n+1}Q_r|{\cal V}_{n+1}\rangle=-\sum_{p=0}^{\lfloor(n-3)/2\rfloor}|{\cal V}_{p+3}\rangle\diamond|{\cal V}_{n-p}\rangle,\end{gathered}$$ which is exactly what we got before from the explicit $(S,S)=0$ calculation to all orders in $g$ and antighost number. In conclusion, the syntactically derived product identities of the sh-Lie algebra maps semantically to equations for the vertices in the Fock complex implementation. This result lends considerable strength to our framework. If higher spin gauge fields have anything to do with physical reality, then we would eventually have to do numerical calculations. Given the complexity of the theory discerned so far, these calculations would undoubtedly have to be computerized. In that case, all “objects” of the theory will have to be mapped to countable infinite data structures (truncated to finite data structures in practice). One might speculate whether in that case formulations that more easily translate into recursive and algebraic data structures of, for example, a functional programming language, might be more useful than the ordinary “pen-and-paper” mathematical formalism of field theory? A note on references ==================== The literature on higher spin field theory is enormous and growing rapidly. The present paper is not intended to be a full review and leaves out many interesting developments or just mentions them in passing. The reason for this is at least twofold: (i) the limits of my own knowledge, (ii) my wish to put forth a certain point of view, perhaps stressing what has gone un-noticed in other works. Thus I hope the present paper can serve as a complement to other excellent reviews of higher spin gauge theory. My referencing necessarily reflects these limitations and choices. I do apologize for any inadvertent omissions. Acknowledgements {#acknowledgements .unnumbered} ---------------- Work partly supported by the Research and Education Board at the University College of Bor[å]{}s. [99]{} Fang J., Fronsdal C., Deformations of gauge groups: gravitation, [*J. Math. Phys.*]{} [**20**]{} (1979), 2264–2271. Fronsdal C., Some open problems with higher spins, in Supergravity, Editors P. van Nieuwenhuizen and D.Z. Freedman, North-Holland Publishing Company, 1979, 245–249. Bengtsson A.K.H., Gauge invariance for spin-3 fields, [*Phys. Rev. D*]{} [**32**]{} (1985), 2031–2036. Burgers G.J.H., On the construction interactions of field theories for higher spin massless particles, PhD Thesis, Rijksuniversteit, Leiden, 1985. Berends F.A., Burgers G.J.H., van Dam H., On spin three self interactions, [*Z. Phys. C*]{} [**24**]{} (1984), 247–254. Bengtsson A.K.H., Bengtsson I., Brink L., Cubic interaction terms for arbitrary spin, [*Nuclear Phys. B*]{} [**227**]{} (1983), 31–40. Fuchs J., Scweigert C., Symmetries, Lie algebras and representations, [*Cambridge Monographs on Mathematical Physics*]{}, Cambridge University Press, 1997. Bengtsson A.K.H., Bengtsson I., Massless higher-spin fields revisited, [*Classical Quantum Gravity*]{} [**3**]{} (1986), 927–936. Bekaert X., Higher spin algebras as higher symmetries, [arXiv:0704.0898](http://arxiv.org/abs/0704.0898). Siegel W., Zwiebach B., Gauge string fields from the light-cone, [*Nuclear Phys. B*]{} [**282**]{} (1987), 125–141. Ouvry S., Stern J., Gauge fields of any spin and symmetry, [*Phys. Lett. B*]{} [**177**]{} (1987), 335–340. Bengtsson A.K.H., A unified action for higher spin gauge bosons from covariant string theory, [*Phys. Lett. B*]{} [**182**]{} (1987), 321–325. Meurice Y., From points to gauge fields, [*Phys. Lett. B*]{} [**186**]{} (1988), 189–194. Labastida J.M.F., Massless particles in arbitrary representations of the [L]{}orentz group, [*Nuclear Phys. B*]{} [**322**]{} (1989), 185–209. Pashnev A., Tsulaia M.M., Dimensional reduction and BRST approach to the description of a Regge trajectory, [*Modern Phys. Lett. A*]{} [**12**]{} (1997), 861–870, [hep-th/9703010](http://arxiv.org/abs/hep-th/9703010). Pashnev A., Tsulaia M., Description of the higher massless irreducible integer spins in the [BRST]{} approach, [*Modern Phys. Lett. A*]{} [**13**]{} (1998), 1853–1864, [hep-th/9803207](http://arxiv.org/abs/hep-th/9803207). Francia D., Sagnotti A., Free geometric equations for higher spins, [*Phys. Lett. B*]{} [**543**]{} (2002), 303–310, [hep-th/0207002](http://arxiv.org/abs/hep-th/0207002). Sundborg B., Stringy gravity, interacting tensionless strings and massless higher spins, [*Nuclear Phys. Proc. Suppl.*]{} [**102**]{} (2001), 113–119, [hep-th/0103247](http://arxiv.org/abs/hep-th/0103247). Bonelli G., On the tensionless limit of bosonic strings, infinite symmetries and higher spins, [*Nuclear Phys. B*]{} [**669**]{} (2003), 159–172, [hep-th/0305155](http://arxiv.org/abs/hep-th/0305155). Sagnotti A., Tsulaia M., On higher spins and the tensionless limit of string theory, [*Nuclear Phys. B*]{} [**682**]{} (2004), 83–116. [hep-th/0311257](http://arxiv.org/abs/hep-th/0311257). Francia D., Sagnotti A., Higher-spin geometry and string theory, [*J. Phys. Conf. Ser.*]{} [**33**]{} (2006), 57–72, [hep-th/0601199](http://arxiv.org/abs/hep-th/0601199). Bianchi M., Higher spins and stringy ${\rm AdS}_5 \times S_5$, [*Fortsch. Phys.*]{} [**53**]{} (2005), 665–691, [hep-th/0409304](http://arxiv.org/abs/hep-th/0409304). Sezgin E., Sagnotti A., Sundell P., On higher spins with a strong $sp(2,r)$ condition, in Proceedings of First Solvay Workshop on Higher-Spin Gauge Theories (May 12–14, 2004, Brussels), Editors G. Bonelli, R. Argurio, G. Barnich and M. Grigoriev, Université Libre de Bruxelles, International Solvay Institutes for Physics and Chemistry, 2004, 100–121, [hep-th/0501156](http://arxiv.org/abs/hep-th/0501156). Mourad J., Francia D., Sagnotti A., Current exchanges and unconstrained higher spins, [*Nuclear Phys. B*]{} [**773**]{} (2007), 203–237, [hep-th/0701163](http://arxiv.org/abs/hep-th/0701163). Maldacena J., The large $N$ limit of superconformal field theories and supergravity, [*Adv. Theor. Math. Phys.*]{} [**2**]{} (1998), 231–252, [hep-th/9711200](http://arxiv.org/abs/hep-th/9711200). Sezgin E., Sundell P., Massless higher spins and holography, [*Nuclear Phys. B*]{} [**644**]{} (2002), 303–370, Erratum, [*Nuclear Phys. B*]{} [**660**]{} (2003), 403, [hep-th/0205131](http://arxiv.org/abs/hep-th/0205131). Morales J.F., Beisert N., Bianchi M., Samtleben H., Higher spin symmetry and $N = 4$ [SYM]{}, [*J. High Energy Phys.*]{} [**2004**]{} (2004), no. 07, 058, 35 pages, [hep-th/0405057](http://arxiv.org/abs/hep-th/0405057). Fronsdal C., Massless fields with integer spin, [*Phys. Rev. D*]{} [**18**]{} (1978), 3624–3629. Bengtsson A.K.H., Structure of higher spin gauge interactions, [*J. Math. Phys.*]{} [**48**]{} (2007), 072302, 35 pages, [hep-th/0611067](http://arxiv.org/abs/hep-th/0611067). Bengtsson A.K.H., [BRST]{} approach to interacting higher-spin gauge fields, [*Classical Quantum Gravity*]{} [**5**]{} (1988), 437–451. Francia D., Sagnotti A., On the geometry of higher-spin gauge fields, [*Classical Quantum Gravity*]{} [**20**]{} (2003), 473–486, [hep-th/0212185](http://arxiv.org/abs/hep-th/0212185). Bengtsson A.K.H., A one-dimensional invariance principle for gauge fields of integer spin, [*Phys. Lett. B*]{} [**189**]{} (1987), 337–340. Henneaux M., Teitelboim C., First and second quantized point particles of any spin, in Quantum Mechanics of Fundamental Systems 2, Editors C. Teitelboim and J. Zanelli, [*Series of the Centro de Estudios Cient[í]{}ficos de Santiago*]{}, Plenum Press, New York, 1989, 113–152. Svartholm N. (Editor), Elementary Particle theory. Relativistic groups and analyticity, Almqvist-Wiksell, Wiley Interscience Division, 1968. Casalbuoni R., Longhi G., A geometrical model for nonhadrons and its implications for hadrons, [*Nouvo Cimento A*]{} [**25**]{} (1975), 482–502. Dirac P.A.M., A remarkable represenation of the 3 + 2 de[S]{}itter group, [*J. Math. Phys.*]{} [**4**]{} (1963), 901–909. Flato M., Fronsdal C., One massless particle equals two [D]{}irac singletons. [*Lett. Math. Phys.*]{} [**2**]{} (1978), 421–426. Flato M., Fronsdal C., On Dis and Racs, [*Phys. Lett. B*]{} [**97**]{} (1980), 236–240. G[ü]{}naydin M., Sa[ç]{}lio[= g]{}lu C., Oscillator-like unitary representations of non-compact groups with a [J]{}ordan structure and the non-compact groups of supergravity, [*Comm. Math. Phys.*]{} [**87**]{} (1982), 159–179. G[ü]{}naydin M., Singleton and doubleton supermultiplets of space-time supergroups and infinite spin superalgebras, Preprint CERN-TH-5500/89, HU-TFT-89-35, 1989. Kibble T.W.B., Lorentz invariance and the gravitational field, [*J. Math. Phys.*]{} [**2**]{} (1961), 212–221. Utiyama R., Invariant theoretical interpretation of interaction, [*Phys. Rev.*]{} [**101**]{} (1956), 1597–1607. MacDowell S.W., Mansouri F., Unified geometric theory of gravity and supergravity, [*Phys. Rev. Lett.*]{} [**38**]{} (1977), 739–742. Mansouri F., Chang L.N., Gravitation as a gauge theory, [*Phys. Rev. D*]{} [**13**]{} (1976), 3192–3200. Grensing D., Grensing G., General relativity as a gauge theory of the [P]{}oincar[é]{} group, [*Phys. Rev. D*]{} [**28**]{} (1983), 286–296. Hehl F.W., von der Heyde P., Kerlick G.D., Nester J.M., General relativity with spin and torsion: foundations and prospects, [*Rev. Modern Phys.*]{} [**48**]{} (1976), 393–416. Kibble T.W.B., Stelle K.S., Gauge theories of gravity and supergravity, in Progress in Quantum Field Theory: In Honour of Professor H. Umewaza, Editors H. Ezawa and S. Kamefuchi, North-Holland, Amsterdam, 1986, 57–81. Stelle K.S., West P.C., Spontaneously broken de [S]{}itter symmetry and the gravitational holonomy group, [*Phys. Rev. D*]{} [**21**]{} (1980), 1466–1488. Iazeolla C., Bekaert X., Cnockaert S., Vasiliev M.A., Nonlinear higher spin theories in various dimensions, in Proceedings of First Solvay Workshop on Higher-Spin Gauge Theories (May 12–14, 2004, Brussels), Editors G. Bonelli, R. Argurio, G. Barnich and M. Grigoriev, Université Libre de Bruxelles, International Solvay Institutes for Physics and Chemistry, 2004, 132–197, [hep-th/0503128](http://arxiv.org/abs/hep-th/0503128). Engqvist J., Dualities, symmetries and unbroken phases in string theory, PhD Thesis, Uppsala University, 2005. D’Auria R., Fr[é]{} P., Geometric supergravity in $D=11$ and its hidden supergroup, [*Nuclear Phys. B*]{} [**201**]{} (1982), 101–140, Errata, [*Nuclear Phys. B*]{} [**206**]{} (1982), 496. Sullivan D., Infinitesimal computations in topology, [*Inst. Hautes Études Sci. Publ. Math.*]{} (1977), no. 47, 269–331. Vasiliev M.A., Actions, charges and off-shell fields in the unfolded dynamics approach, [*Int. J. Geom. Methods Mod. Phys.*]{} [**3**]{} (2006), 37–80, [hep-th/0504090](http://arxiv.org/abs/hep-th/0504090). Lada T., Stasheff J., Introduction to [sh-Lie]{} algebras for physicists, [*Internat. J. Theoret. Phys.*]{} [**32**]{} (1993), 1087–1104, [hep-th/9209099](http://arxiv.org/abs/hep-th/9209099). Lada T., Markl M., Strongly homotopy [L]{}ie algebras, [*Comm. Algebra*]{} [**23**]{} (1995), 2147–2161, [hep-th/9406095](http://arxiv.org/abs/hep-th/9406095). Voronov T., Higher derived brackets and homotopy algebras, [*J. Pure Appl. Algebra*]{} [**202**]{} (2005), 133–153, [math.QA/0304038](http://arxiv.org/abs/math.QA/0304038). de Azc[á]{}rraga J.A., P[é]{}rez Bueno J.C., Higher-order simple [L]{}ie algbras, [*Comm. Math. Phys.*]{} [**184**]{} (1997), 669–681, [hep-th/9605213](http://arxiv.org/abs/hep-th/9605213). Barnich G., Grigoriev M., BRST extension of the non-linear unfolded formalism, [hep-th/0504119](http://arxiv.org/abs/hep-th/0504119). Semikhatov A., Barnich G., Grigoriev M., Tipunin I., Parent field theory and unfolding in [BRST]{} first-quantized terms, [*Comm. Math. Phys.*]{} [**260**]{} (2005), 147–181, [hep-th/0406192](http://arxiv.org/abs/hep-th/0406192). Ogievetski V.I., Polubarinov I.V., Interacting fields of spin 1 and symmetry properties, [*Ann. Phys.*]{} [**25**]{} (1963), 358–386. Gupta S.N., Quantization of [E]{}instein’s gravitational field: linear approximation, [*Proc. Phys. Soc. Sect. A*]{} [**65**]{} (1952), 161–169. Gupta S.N., Gravitation and electromagnetism, [*Phys. Rev. (2)*]{} [**96**]{} (1954), 1693–1685. Kraichnan R.H., Special-relativistic derivation of generally covariant gravitation theory, [*Phys. Rev.*]{} [**98**]{} (1955), 1118–1122. Wyss W., Zur unizit[ä]{}t der [G]{}ravitationstheorie, [*Helv. Phys. Acta*]{} [**38**]{} (1965), 469–480. Thirring W.E., An alternative approach to the theory of gravitation, [*Ann. Phys.*]{} [**16**]{} (1961), 96–117. Feynman R.P., Feynman lectures on gravitation, Westview Press, 2002. Deser S., Self-interaction and gauge invariance, [*Gen. Relativity Gravitation*]{} [**1**]{} (1970), 9–18. Boulware D.G., Deser S., Kay J.H., Supergravity from self-interaction, [*Phys. A*]{} [**96**]{} (1979), 141–162. Deser S., Gravity from self-interaction in a curved background, [*Classical Quantum Gravity*]{} [**4**]{} (1987), L99–L105. Bengtsson A.K.H., An abstract interface to higher spin gauge field theory, [*J. Math. Phys.*]{} [**46**]{} (2005), 042312, 23 pages, [hep-th/0403267](http://arxiv.org/abs/hep-th/0403267). Barnich G., Henneaux M., Consistent couplings between fields with a gauge freedom and deformations of the master equation, [*Phys. Lett. B*]{} [**311**]{} (1993), 123–129, [hep-th/9304057](http://arxiv.org/abs/hep-th/9304057). Henneaux M., Consistent interactions between gauge fields: the cohomological approach, in Secondary Calculus and Cohomological Physics, Editors M. Henneaux, J. Krasil’shchik and A. Vinogradov, [*Contemp. Math.*]{} [**219**]{} (1998), 93–109, [hep-th/9712226](http://arxiv.org/abs/hep-th/9712226). Boulanger N., Bekaert X., Cnockaert S., Spin three gauge theory revisited, [*J. High Energy Phys.*]{} [**2006**]{} (2006), no. 01, 052, 34 pages, [hep-th/0508048](http://arxiv.org/abs/hep-th/0508048). Stasheff J., The (secret?) homological algebra of the [B]{}atalin–[V]{}ilkovisky approach, in Secondary Calculus and Cohomological Physics, Editors M. Henneaux, J. Krasil’shchik and A. Vinogradov, [*Contemp. Math.*]{} [**219**]{} (1998), 195–210, [hep-th/9712157](http://arxiv.org/abs/hep-th/9712157). Berends F.A., Burgers G.J.H., van Dam H., On the theoretical problems in constructing interactions involving higher-spin massless particles, [*Nuclear Phys. B*]{} [**260**]{} (1985), 295–322. Fulp R., Lada T., Stasheff J., [Sh-Lie]{} algebras induced by gauge transformations, [*Comm. Math. Phys.*]{} [**231**]{} (2002), 25–43, [math.QA/0012106](http://arxiv.org/abs/math.QA/0012106). Baez J., Quantum quandaries: a category-theoretic perspective, in Structural Foundations of Quantum Gravity, Editors D.P. Rickles, S.R.D. French and J. Saatsi, Oxford University Press, 2006, 240–265, . May J.P., Operads, algebras and modules, in Operads: Proceedings of Renaissance Conferences (1995, Hartford, CT/Luminy), [*Contemp. Math.*]{} [**202**]{} (1997), 15–31. Leinster T., Higher operads, higher categories, Cambridge University Press, 1989.

ArXiv

--- bibliography: - 'Doc.bib' --- > 0.5 cm 0.5 cm > Bright moon light in front of my bed,\ > Maybe... frost on the ground.\ > I raise my head, behold the bright moon,\ > I bow my head— home-sick.\ > \[Lĭ Bái, Tāng Dynasty poet\] Abstract for HEP-TH {#abstract-for-hep-th .unnumbered} =================== Modified version of Ph.D. thesis {#modified-version-of-ph.d.-thesis .unnumbered} -------------------------------- Herein we propose a new numerical technique for solving field theories: the large momentum frame (LMF). This technique combines several advantages of lattice gauge theory with the simplicity of front form quantisation. We apply the LMF on QED(1+1) and on the $\phi^4(3+1)$ theory. We demonstrate both analytically and in practical examples (1) that the LMF does neither correspond to the [*infinite*]{} momentum frame (IMF) nor to the front-form (FF) (2) that the LMF is not equivalent to the IMF (3) that the IMF is unphysical since it violates the lattice scaling window and (4) that the FF is even more unphysical because FF propagators violate micro-causality, causality and the finiteness of the speed of light. We argue that distribution functions measured in deep inelastic scattering should be interpreted in the LMF (preferably in the Breit frame) rather than in the FF formalism. In particular, we argue that deep inelastic scattering probes [*space*]{}-like distribution functions. Résumé I {#résumé-i .unnumbered} ======== Signé par (Norbert Scheu) (Helmut Kröger) Résumé II {#résumé-ii .unnumbered} ========= Signé par (Norbert Scheu) (Helmut Kröger) Avant-propos {#avant-propos .unnumbered} ============

ArXiv

--- abstract: 'It is shown that all vacuum solutions of Einstein field equation with a positive cosmological constant are the solutions of a model of dS gauge theory of gravity. Therefore, the model is expected to pass the observational tests on the scale of solar system and explain the indirect evidence of gravitational wave from the binary pulsars PSR1913+16.' author: - 'Chao-Guang Huang$^{a,e}$[^1], Yu Tian$^{b,e}$[^2], Xiaoning Wu$^{c}$[^3] Han-Ying Guo$^{d,e}$[^4]' date: November 2007 title: 'On Torsion-free Vacuum Solutions of the Model of de Sitter Gauge Theory of Gravity' --- The astronomical observations show that our universe is probably an asymptotically de Sitter(dS) one [@SN; @WMAP]. It arises the interests on dS gauge theories of gravity. There is a model of the dS gauge theory of gravity, which was first proposed in the 1970’s [@dSG; @T77]. The model can be stimulated from dS invariant special relativity [@dSSR; @meetings; @dSSR2] and the principle of localization [@Guo2], just like that the Poincaré gauge theory of gravity may be stimulated from the Einstein special relativity and the localization of Poincaré symmetry [@PGT]. The principle of localization is that the full symmetry of the special relativity as well as the laws of dynamics are both localized. The gravitational action of the model takes the Yang-Mills form of [@dSG; @T77; @Guo2] S\_[GYM]{}=1 [4g\^2]{} \_[M]{}d\^4x e [**Tr**]{}\_[dS]{}([F]{}\_[F]{}\^), where $e=\det(e^a_\mu)$ is the determinant of the tetrad $e^a_\mu$, $g$ is a dimensionless coupling constant introduced as usual in the gauge theory to describe the self-interaction of the gauge field, \[dScurv\] [F]{}\_=([F]{}\^[AB]{}\_[   ]{} )= ( [cc]{} F\^[ab]{}\_[  ]{}+R\^[-2]{} e\^[ab]{}\_[  ]{} & R\^[-1]{}T\^a\_[ ]{}\ -R\^[-1]{}T\^b\_[ ]{} & 0 ) is the curvature of dS connection[^5] \_=([B]{}\^[AB]{}\_[   ]{})=( [cc]{} B\^[ab]{}\_[  ]{} & R\^[-1]{}e\^a\_\ -R\^[-1]{}e\^b\_& 0 ) so(1,4), and ${\bf Tr}_{\rm dS}$ is the trace for the $so(1,4)$ indices $A, B$. In Eq.(\[dScurv\]), $F^{ab}_{~~\mu\nu}$ and $T^a_{~\mu\nu}$ are the curvature and torsion tensors of the Lorentz connection $B^a_{\ b\mu}\in so(1,3)$, respectively, $R$ is the dS radius, and $e_{ab}^{~\mu\nu}=e_a^\mu e_b^\nu-e_a^\nu e_b^\mu$. In terms of $F^{ab}_{~~\,\mu\nu}$ and $T^{a}_{~\, \mu\nu}$, the gravitational action can be rewritten as S\_[GYM]{}= -\_[M]{}d\^4x e , \[GYM\] where $F= \frac{1}{2} F^{ab}_{~\ \mu\nu}e_{ab}^{~\ \mu\nu}$ the scalar curvature, the same as the action in the Einstein-Cartan theory, $\chi=1/({ 16}\pi G)$ is a dimensional coupling constant, $\La = 3/R^2=3\chi g^2$ is the cosmological constant. The gravitational field equations, obtained by the variation of the total action \[S\_t\]S\_[T]{}=S\_[GYM]{}+S\_[M]{}with respect to $e^a_{~\mu},B^{ab}_{~~\mu}$, are \[Geq2\]T\_[a   ||]{}\^[ ]{} &-& F\_[ a]{}\^+F e\_a\^- e\_a\^= 8G( T\_[[M]{}a]{}\^[  ]{}+T\_[[G]{}a ]{}\^[  ]{}),\ \[Geq2’\]F\_[ab   ||]{}\^[  ]{} &=& R\^[-2]{}(16G S\^\_[[M]{}ab]{}+S\^\_[[G]{}ab]{}).Here, $S_{\rm M}$ is the action of the matter source with minimum coupling, $||$ represents the covariant derivative using both Christoffel symbol $\{^\mu_{\nu\ka}\}$ and connection $B^a_{\ b\mu}$, $F_a^{~\mu}=-F_{ab}^{~~\mu\nu}e^b_\nu$, T\_[[M]{}a]{}\^[  ]{}&:=&-1 e\ \[emG\]T\_[[G]{}a]{}\^[  ]{}&:=&g\^[-2]{} T\_[[F]{}a]{}\^[  ]{}+2T\_[[T]{}a]{}\^[  ]{} are the tetrad form of the stress-energy tensor for matter and gravity, respectively, where \[emF\]T\_[[F]{}a]{}\^[  ]{}&:=&- d\^4x e [Tr]{}(F\_F\^)\ &=& e\_[a]{}\^(F\^F\_)-e\_a\^(F\^ F\_) is the tetrad form of the stress-energy tensor for curvature and \[emT\]T\_[[T]{}a]{}\^[  ]{}&:=&- d\^4x e T\^b\_[ ]{}T\_b\^[ ]{}+T\_[a||]{}\^[ ]{}\ &=& e\_a\^T\_b\^[ ]{}T\^[b]{}\_[ ]{}-e\_a\^T\_b\^[ ]{}T\^b\_[ ]{} the tetrad form of the stress-energy tensor for torsion, and S\_[[M]{}ab]{}\^[ ]{} =1 [2]{} and $S_{{\rm G}ab}^{\quad \,\mu }$ are spin currents for matter and gravity, respectively. Especially, the spin current for gravity can be divided into two parts, \[spG\]S\_[[G]{}ab]{}\^[ ]{}&=&S\_[[F]{}ab]{}\^[ ]{}+2S\_[[T]{}ab]{}\^, where S\_[[F]{}ab]{}\^[ ]{}&:=&1 [2]{}d\^4 x F = -e\^[  ]{}\_[ab  [||]{}]{} = Y\^\_[  ]{} e\_[ab]{}\^[  ]{}+Y \^\_[  ]{} e\_[ab]{}\^[  ]{}\ S\_[[T]{}ab]{}\^&:=& 1 [2]{}1 4 d\^4 x T\^c\_[ ]{}T\_c\^[ ]{} =T\_[\[a]{}\^[ ]{}e\_[b\]]{}\^ are the spin current for curvature $F_{ab}^{\ \ \mu\nu}$ and torsion $T_a^{~\mu\nu}$, respectively. In Ref.[@wzc], it is shown that all vacuum solutions of Einstein field equation without cosmological constant are the solutions of Eq.(\[Geq2\]) and Eq.(\[Geq2’\]) for the case of sourceless, torsion-free, and vanishing cosmological constant. However, a positive cosmological constant is vitally important for the dS gauge theories of gravity. Without a positive cosmological constant, the gravity should be a Poincaré or AdS one. Therefore, in order to see whether the model of dS gauge theory of gravity can pass the observational tests on the scale of solar system, it should be important to explore if the vacuum solutions of Einstein field equation with a positive cosmological constant do satisfy the equations of the model. The purpose of the present Note is to show that it is just the case. Namely, all vacuum solutions of Einstein field equation with a positive cosmological constant are the solutions of the torsion-free vacuum equations of the model of dS gauge theory of gravity. For the sourceless case, the torsion-free gravitational field equations of the model reduce to && [R]{}\_[ a]{}\^- e\_a\^+ e\_a\^= -8G( T\_[[M]{}a]{}\^[  ]{}+T\_[[R]{}a ]{}\^[  ]{}), \[ElEq\]\ &&[R]{}\_[ab  ;]{}\^[  ]{} = 0, \[YangEq\] where $T_{{\rm R}a}^{~~ \mu}=e_{a}^{\nu}T_{{\rm R}\ \nu}^{~\mu}$ the tetrad form of the stress-energy tensor of Riemann curvature ${\cal R}_{ab}^{~~\mu\nu}{}$, and a semicolon $;$ is the covariant derivative using both the Christoffel and Ricci rotation coefficients. Eq. (\[ElEq\]) is the Einstein-like equation, while Eq.(\[YangEq\]) is the Yang equation [@Yang]. It can be shown [@wzc] that T\_[[R]{} ]{}\^[  ]{} &=&[R]{}\_[ab ]{}[R]{}\^[ab]{} - \_\^([R]{}\_[ab]{}[R]{}\^[ab]{})\ &=& 1 2 ([R]{}\_[R]{}\^+[R]{}\^\*\_[ ]{} [R]{}\^[\*]{})\ &=&2C\_\^[  ]{}[R]{}\^\_+([R]{}\_\^-\_\^), \[emR\] where ${\cal R}_{\ka\si\mu\la}$ is the Riemann curvature tensor, ${\cal R}^*_{\ka\si\mu\la} =\frac 1 2 {\cal R}_{\ka\si \tau \rho}\eps^{\tau \rho}_{~~\mu\la}$ is the right dual of the Riemann curvature tensor, $C_{\la\mu\ka\nu}$ is the Weyl tensor. In the last step in (\[emR\]), the Géhéniau-Debever decomposition for the Riemann curvature, \_=C\_ + E\_ + G\_, is used [[@GD]]{}, where E\_ &=& 1 2 (g\_ S\_+g\_ S\_-g\_ S\_-g\_ S\_),\ G\_ &=& (g\_g\_-g\_g\_),\ S\_ &=& [R]{}\_-1 4 [R]{}g\_. On the other hand, the vacuum Einstein field equation with a (positive) cosmological constant reads \_[ ]{}\^- \_\^+ \_\^= 0. It results in =4, \_[ ]{}\^= \_\^,\[Ricci\] and thus S\_=0. Since the Weyl tensor is totally traceless, the stress-energy tenor for Riemann curvature vanishes, [*i.e.*]{}, T\_[[R]{} ]{}\^[  ]{} =0. Therefore, all vacuum solutions of Einstein field equation with a cosmological constant are solutions of Eq.(\[ElEq\]). In addition, the Bianchi identity \^\_[  ;]{} +[R]{}\^\_[  ;]{}+[R]{}\^\_[  ;]{}=0 leads to 0=[R]{}\^\_[  ;]{} -[R]{}\^\_[ ;]{}+[R]{}\^\_[ ;]{}=[R]{}\^[  ]{}\_[  ;]{}. \[rYEq\] Namely, Yang equation (\[YangEq\]) is also satisfied. (The last step of Eq.(\[rYEq\]) is valid because of Eq.(\[Ricci\]).) Therefore, we come to the conclusion that all vacuum solutions of the Einstein field equation with a positive cosmological constant are the torsion-free vacuum solutions of the model of dS gauge theory of gravity. In particular, the dS, Schwarzschild-dS, and Kerr-de Sitter metrics satisfy the Eqs.(\[Geq2\]) and (\[Geq2’\]). Note that the Birkhoff theorem has been proved for the gravitational theory (\[GYM\]) without a cosmological constant [@RauchNieh]. So, the model is expected to pass the observational tests on the scale of solar system. In addition, the model has the same metric waves as general relativity and thus is expected to explain the indirect evidence of the existence of gravitational wave from the observation data on the binary pulsar PSR1913+16. One might think that the above results are trivial because the Yang equation does not appear at all if the torsion-free condition is assumed in the action, in which case the tetrad and connection are not independent. However, the torsion-free manifold is just the specific situation of the the model. There is no reason to set the torsion to be zero before the variation. In fact, it can be shown that all solutions of vacuum Einstein field equation with a positive cosmological constant are also the vacuum, torsion-free solutions of the field equations when the terms $$F_a^{\ \,\mu}F^a_{\ \,\mu}, \ e^a_\nu e^b_\mu F_a^{\ \,\mu}F_b^{\ \,\nu} , \ e^{ab}_{\ \ \, \la\si}e^{cd}_{\ \ \, \mu\nu}F_{ab}^{\ \ \,\mu\nu}F_{cd}^{\ \ \,\la\si} , \ e^b_{\si} e^{c}_{\mu}F_{ab\ \, \nu}^{\ \ \mu}F_{ac}^{\ \ \,\nu\si} , \ e_a^\la e_b^{\si} T^a_{\ \mu\la}T^{b\mu}_{\ \ \si} , \ e_a^\si e_b^{\mu} T^a_{\ \mu\la}T^{b\la}_{\ \ \si} \nno$$ are added in the gravitational Lagrangian. Obviously, the last two terms have no contribution to the vacuum, torsion-free field equations, while the middle two terms contribute the same as the term $F_{ab}^{\ \ \mu\nu}F^{ab}_{\ \ \mu\nu}$ does thus only alter the unimportant coefficients. The first two terms add the term $(R_{[a}^\mu e_{b]}^\nu)_{;\nu}$ in Yang equation and the stress-energy tensor $R_{\mu \la} R^{\nu \la} -\frac 1 4 \dl_\mu^\nu R_{\si \la} R^{\si \la}$ in Einstein equation. Both of them vanish for the solutions of the vacuum Einstein equation with a positive cosmological constant. Obviously, the conclusion is still valid if the integral of the second Chern form of the dS connection over the manifold is added in the action. Finally, the similar discussions can be applied to the AdS case as well. -4mm We thank Z. Xu, B. Zhou and H.-Q. Zhang for useful discussions. This work is supported by NSFC under Grant Nos. 90503002, 10605005, 10775140, 10705048, 10731080 and Knowledge Innovation Funds of CAS (KJCX3-SYW-S03). [07]{} A. G. Riess *et al*., Astron. J. [**116**]{} (1998), 1009, astro-ph/9805201; S. Perlmutter *et al*., Astrophys. J. [**517**]{} (1999), 565, astro-ph/9812133. A. G. Riess *et al*., Astrophys. J. [**536**]{} (2000), 62, astro-ph/0001384. A. G. Riess *et al*., Astrophys. J. [**560**]{} (2001), 49, astro-ph/0104455. C. L. Bennett *et al*., Astrophys. J. Suppl. [**148**]{} (2003), 1, astro-ph/0302207; D. N. Spergel *et al*., Astrophys. J. Suppl. [**148**]{} (2003), 175, astro-ph/0302209. Y.-S. Wu, G.-D. Li and H.-Y. Guo, Kexue Tongbao (Chi. Sci. Bull.) [**19**]{} (1974), 509; Y. An, S, Chen, Z.-L. Zou and H.-Y. Guo, [*ibid*]{}, 379; H.-Y. Guo, [*ibid*]{} [**21**]{} (1976) 31; Z. L. Zou, et al, Sci. Sinica [**XXII**]{} (1979), 628; M.-L. Yan, B.-H. Zhao and H.-Y. Guo, Kexue Tongbao (Chi. Sci. Bull.) [**24**]{} (1979), 587; [*Acta Physica Sinica*]{} [**33**]{} (1984), 1377; 1386 (all in Chinese). P. K. Townsend, 15 (1977), 2795; A. A. Tseytsin, 26 (1982), 3327. K.-H. Look (Q.-K. Lu) 1970, Why the Minkowski metric must be used? unpublished; K.-H. Look, C.-L. Tsou (Z.-L. Zou) and H.-Y. Kuo (H.-Y. Guo), Acta Phys. Sin. [**23**]{} (1974), 225; Nature (Shanghai, Suppl.); H.-Y. Guo, Kexue Tongbao (Chinese Sci. Bull.) [**22**]{} (1977), 487 (all in Chinese). H.-Y. Guo, in [*Proceedings of the 2nd Marcel Grossmann Meeting on General Relativity*]{}, ed. by R. Ruffini, (North-Holland, 1982), 801; [*Nucl. Phys. B*]{} (Proc. Suppl.) [**6**]{} (1989), 381; C.-G. Huang and H.-Y. Guo, in [*Gravitation and Astrophysics — On the Occasion of the 90th Year of General Relativity, Proceddings of the VII Asia-Pacific International Conference*]{}, ed. by J. M. Nester, C.-M Chen, and J.-P. Hsu (World Scientific, 2007, Singapore), 260. H.-Y. Guo, C.-G. Huang, Z. Xu and B. Zhou, [*Mod. Phys. Lett.*]{} [**A19**]{} (2004), 1701, hep-th/0403013; 331 (2004), 1, hep-th/0403171; H.-Y. Guo, C.-G. Huang, Z. Xu and B. Zhou, [*Chin. Phys. Lett.*]{} [**22**]{} (2005), 2477, hep-th/0508094; H.-Y. Guo, C.-G. Huang, Y. Tian, Z. Xu and B. Zhou, [*Acta Phys. Sin.*]{} [**54**]{} (2005), 2494 (in Chinese); H.-Y. Guo, C.-G. Huang and B. Zhou, [*Europhys. Lett.*]{} [**72**]{} (2005), 1045, hep-th/0404010. H.-Y. Guo, C.-G. Huang, Y. Tian, H.-T. Wu, B. Zhou, Class. Quant. Grav. [**24**]{} (2007) 4009; H.-Y. Guo, C.-G. Huang, Y. Tian, B. Zhou, Front. Phys. China, [**2**]{} (2007), 358; H.-Y. Guo, [*Phys. Lett.*]{} [**B653**]{} (2007) 88; H.-Y. Guo, Special Relativity and Theory of Gravity via Maximum Symmetry and Localization – In Honor of the 80th Birthday of Professor Qikeng Lu, arXiv:0707.3855. T. W. B. Kibble, [*J. Math. Phys.*]{} [**2**]{} (1961), 212; F. W. Held, P. von der Heyde, G. D. Kerlick, and J. M. Nester, [*Rev Mod. Phys.*]{} [**48**]{} (1976), 393 and references therein. S. W. MacDowell and F. Mansouri,  [**38**]{} (1977), 739; Erratum-ibid. [**38**]{} (1977), 1376; K. S. Stelle and P. C. West, 21 (1980), 1466; F. Wilczek,  [**80**]{} (1998), 4951; L. Freidel and A. Starodubtsev, Quantum gravity in terms of topological observables, arXiv: hep-th/0501191; M. Leclerc, Annals of Physics, [**321**]{} (2006), 708; E. Witten, Three-Dimensional Gravity Reconsidered, arXiv:0706.3359. C. N. Yang, Phys. Rev. Lett. [**33**]{} (1974), 445. Y.-S. Wu, Z.-L. Zou and S. Chen, [*Kexue Tongbao (Chin. Sci. Bull.)*]{} [**18**]{} (1973), 119 (in Chinese). H. Stephani, D. Kramer, M. MacCallum, C. Hoenselaers, E. Herlt, Exact Solutions of Einstein’s Field Equations, (Cambridge University Press, Cambridge, 1980). R. Rauch and N. T. Nieh, Phys. Rev. D[**24**]{} (1981), 2029. [^1]: Email: huangcg@ihep.ac.cn [^2]: Email: ytian@bit.edu.cn [^3]: Email: wuxn@amss.ac.cn [^4]: Email: hyguo@ihep.ac.cn [^5]: The same -connection with different dynamics has also been explored in Ref. [@dSconnection].

ArXiv

--- abstract: 'We study the stability of singly- and doubly-quantized vortex states of harmonically trapped dipolar Bose-Einstein Condensates (BECs) by calculating the low-lying excitations of these condensates. We map the dynamical stability of these vortices as functions of the dipole-dipole interaction strength and trap geometry by finding where their excitations have purely real energy eigenvalues. In contrast to BECs with purely contact interactions, we find that dipolar BECs in singly-quantized vortex states go unstable to modes with an increasing number of angular and radial nodes for more oblate trap aspect ratios, corresponding to *local* collapse that occurs on a characteristic length scale. Additionally, we find that dipolar BECs in doubly-quantized vortex states are unstable to decay into a different topological state (with two singly-quantized vortices) for all interaction strengths when the trap geometry is sufficiently prolate to make the dipoles attractive, and in windows of interaction strength when the trap geometry is sufficiently oblate to make the dipoles repulsive.' author: - 'Ryan M. Wilson' - Shai Ronen - 'John L. Bohn' title: 'Stability and Excitations of a Dipolar Bose-Einstein Condensate with a Vortex' --- Introduction ============ The observation of strong dipolar effects in a Bose-Einstein condensate (BEC) of harmonically trapped $^{52}\mathrm{Cr}$ [@KochNature08a; @Lahaye08] atoms marks encouraging progress towards understanding these novel systems. In contrast to the isotropic contact interaction present in condensates of alkali atoms, the dipole-dipole interaction is long ranged, anisotropic, and is predicted to induce interesting ground state structures [@Ronen07; @Baranov08PR] and excitation spectra in both fully and partially trapped systems. For example, an excitation spectrum much like the roton-maxon spectrum in superfluid $\mathrm{He}$ is predicted in dipolar BECs (DBECs) with both three-dimensional (3D) [@Ronen07; @Wilson08] and quasi-two-dimensional (2D) [@Santos03a] geometries. Additionally, the presence of dipolar effects has recently been shown to be critical in explaining the rich behaviors of spinor BEC systems [@Kawaguchi06; @Vengalattore08], the effects of which are strongly conditioned by the attractive part of the dipole-dipole interaction. For this reason, it is instructive to compare the dipolar system to a BEC with attractive contact interactions. In this article, we find the properties of the two systems to be in stark contrast. To see this contrast, we consider the effects of the dipole-dipole interaction on a condensate with a single vortex core [@Pu06]. The conditions for the generation of such a DBEC vortex state are studied in Ref. [@Bijnen07]. First, consider a trapped BEC with attractive contact interactions, characterized by the $s-$wave scattering length $a_s$ of the constituent particles and with *no* vortex. For such a system, there always exists a critical particle number above which the condensate goes unstable, with preference to collapse in the region of maximum density at the center of the trap [@Dodd96; @Bradley97]. Stirring the condensate into a vortex state serves to stabilize the system by introducing a kinetic energy component due to angular momemtum that offsets the interparticle attraction. So, in general, the vortex will sustain a larger number of particles than the non-vortex state, and is more stable. The case for a DBEC, however, is quite different. Consider a DBEC with its dipolar entities polarized in the $z$-direction. Because the dipole-dipole interaction is anisotropic, the structure and stability of a DBEC depends strongly on the trap aspect ratio $\lambda = \omega_z / \omega_\rho$, where $\omega_z$ and $\omega_\rho$ are the axial and radial trap frequencies, respectively. For smaller $\lambda$, the dipole-dipole interaction distends the condensate into a prolate shape, where macroscopic collapse can occur due to long-range attraction in the direction of polarization. As $\lambda$ is increased, the condensate is stabilized since the dipole-dipole interaction is predominitely repulsive in this more oblate geometry. For a moderate number of particles, the energy cost of stacking the dipoles in the direction of polarization is outweighed by the tight harmonic confinement in this direction. However, for a sufficiently large number of particles, a DBEC in an oblate trap is subject to an instability due to local density fluctuations, which are foreign to the contact potential case. The attractive part of the dipole-dipole interaction dominates in regions where the higher-density fluctuations occur, inititating a local collapse of the condensate. As will be discussed below, this instability is intimitely connected with an excitation that goes soft at a critical number of dipoles, and that has been dubbed the “discrete-roton.”  [@Ronen07] Signatures of this local collapse have been articulated in the simulations of Ref. [@Parker08], where the manifestation of the roton in the collapse of a DBEC is shown for the early stages of collapse. In the presence of a singly-quantized vortex, the region of high density is forced away from the center of the trap due to the zero-density of the vortex core. Depending on the aspect ratio $\lambda$ of the trap, this either serves to stabilize (for smaller $\lambda$) or destabilize (for larger $\lambda$) the DBEC. For smaller $\lambda$, the vortex core simply breaks the prolate shape of the condensate along the direction of polarization, eliminating much of the attractive dipole-dipole interaction in this direction and thus increasing the energy due to interactions. Conversely, for larger $\lambda$ the vortex increases the density in the periphery of the core and thus encourages local collapse. Just as the roton wavelength is set by the the confinement length in the direction of polarization ($z$), the local density fluctuations occur at the same length scale regardless of the trap geometry. Widening the trap radially while keeping the axial trapping frequency fixed makes more room for regions of density fluctuations instead of enlarging the existing regions. This marks a clear and important distinction between the dipole-dipole and contact interactions. An additional consideration relevant to the stability of DBECs with a vortex are the excitations of the vortex core itself [@Klawunn08]. As we will see below, these excitations are unlikely to play a role in destabilizing the vortex for oblate, or even mildly prolate, traps. Our paper is organized as follows: In section \[sec:methods\] we describe the model that we use to study the ground state with a vortex, including the novel algorithm that we employ to carry out our calculations. In section \[sec:BdG\] we discuss the Bogoliubov de Gennes (BdG) formalism and present our calculations of the BdG spectrum in reference to the stability of the system. Finally, in section \[sec:s2\] we calculate the BdG spectrum of a DBEC with a doubly-quantized vortex. Interestingly, in a more prolate trap geometry where the dipoles are mostly attractive we find that the BdG spectrum looks very similar to the BEC with purely attractive contact interactions, while in a more oblate trap geometry where the dipoles are mostly repulsive we find that the BdG spectrum looks very similar to that of the BEC with purely repulsive contact interactions, having windows of dynamical stability for certain dipole-dipole interaction strengths [@Lundh06; @Pu99]. Methods {#sec:methods} ======= At ultracold temperatures, a condensate of $N$ bosons trapped in an external potential $U({\mathbf{r}})$ may be described within mean field theory [@Bortolotti06] by the nonlocal Gross-Pitaevskii Equation (GPE): $$\begin{aligned} \label{GPE} \left[ -\frac{\hbar^2}{2M} \nabla^2 + U({\mathbf{r}}) + (N-1) \right. \nonumber \\ \left. \times \int d{\mathbf{r}}^\prime \Psi_0^*({\mathbf{r}}^\prime) V({\mathbf{r}}-{\mathbf{r}}^\prime) \Psi_0({\mathbf{r}}^\prime)\right] \Psi_0({\mathbf{r}}) = \mu \Psi_0({\mathbf{r}})\end{aligned}$$ where $\Psi_0({\mathbf{r}})$ is the condensate wave function with unit norm, ${\mathbf{r}}$ is the distance from the trap center, $M$ is the boson’s mass and $V({\mathbf{r}}- {\mathbf{r}}^\prime)$ is the two-particle interaction potential. For a cylindrically symmetric harmonic trap, the external potential is given by $U({\mathbf{r}}) = \frac{1}{2}M\omega_\rho^2(\rho^2+\lambda^2 z^2)$, where $\lambda = \omega_z/\omega_\rho$ is the trap aspect ratio. The interaction potential for a dipolar system has the form [@Yi00] $$\label{intpot} V({\mathbf{r}}-{\mathbf{r}}^\prime) = \frac{4\pi\hbar^2 a_s}{M} \delta({\mathbf{r}}- {\mathbf{r}}^\prime) + d^2\frac{1-3\cos^2{\theta}}{|{\mathbf{r}}-{\mathbf{r}}^\prime|^3}$$ where $d$ is the dipole moment and $\theta$ is the angle between the vector ${\mathbf{r}}-{\mathbf{r}}^\prime$ and the dipole axis. To illuminate purely dipolar effects, we set $a_s=0$ in this work, a limit that has been achieved experimentally in a BEC of atomic $^{52}\mathrm{Cr}$ by employing a Fano-Feshbach resonance [@Werner05]. To characterize the structure and stability of a DBEC with a vortex, we introduce the dimensionless characteristic dipole strength, $$\label{D} D = (N-1)\frac{Md^2}{\hbar^2 a_\mathrm{ho}}$$ where $a_\mathrm{ho} = \sqrt{\hbar/M\omega_\rho}$ is the radial harmonic oscillator length. Notice that increasing (decreasing) $D$ corresponds to either increasing (decreasing) the number of particles in the condensate or the square of the dipole moment of the particles. So for a DBEC of, say, $^{52}\mathrm{Cr}$, one must change the number of atoms in the condensate in order to change $D$, since the magnetic dipole moment of $^{52}\mathrm{Cr}$ (6 $\mu_B$) is effectively fixed. The second term in Eq. (\[intpot\]) describes the two-body dipole-dipole interaction for dipoles that are polarized along the trap axis ($z$-axis) [@JacksonEM], as may be achieved by applying a strong external field to the condensate. This term is long ranged ($\propto 1/r^3$), anisotropic, and has a coordinate-space singularity as the distance between the particles goes to zero. These concerns are handled by treating the mean field dipole term in the GPE, $d^2\int d{\mathbf{r}}^\prime \Psi_0^*({\mathbf{r}}^\prime)\frac{1-3\cos^2{\theta}}{|{\mathbf{r}}-{\mathbf{r}}^\prime|^3}\Psi_0({\mathbf{r}}^\prime)$, by means of convolution. This method eliminates the singularity in coordinate space since the dipolar mean field in momentum space, $\tilde{V}_D({\mathbf{k}})$, is given by [@Goral03] $$\label{Vk} \tilde{V}_D({\mathbf{k}}) = \frac{4\pi}{3}(3 k_z^2/k^2-1).$$ The dipolar mean field in coordinate space may then be calculated in terms of $\tilde{V}_D({\mathbf{k}})$, $$\label{Vr} \Phi_D({\mathbf{r}}) = \mathcal{F}^{-1}\left[\tilde{V}_D({\mathbf{k}})\tilde{n}({\mathbf{k}})\right],$$ where $\mathcal{F}^{-1}$ is the inverse Fourier transform operator and $\tilde{n}({\mathbf{k}})$ is the Fourier transform of the condensate density. In general, these transforms must be computed in three dimensions to capture the three-dimensional (3D) nature of the system. However, the system that we are considering possesses cylindrical symmetry in both the trapping and dipolar interaction potentials. With such a symmetry, the condensate states may be written in cylindrical coordinates as eigenstates of the orbital angular momentum projection $s$, $\Psi_0({\mathbf{r}}) = \psi_0(\rho,z)e^{is\varphi}$ [@BEC2003]. The $s=0$ state corresponds to a rotationless condensate, while the $s>0$ states correspond to condensates with vortices of charge $s$, or equivalently, with $\hbar s$ units of orbital angular momentum per particle. This formulation allows for a distinctive computational algorithm to be applied to the problem, reducing a fully 3D calculation to a 2D one by working in cylindrical coordinates and integrating out the simple $\varphi$ dependence of the state. Specifically, the algorithm uses a one-dimensional (1D) Hankel transform of order $s$ in the $\rho$-coordinate and a 1D Fourier transform in the $z$-coordinate to transform a function with the angular dependence $e^{is\varphi}$ into momentum space. For example, the transform of the wave function $\psi_0(\rho,z)e^{is\varphi}$ is $$\begin{aligned} \label{transform} \tilde{\psi_0}({\mathbf{k}}) = \sqrt{2\pi} i^{-s} e^{-isk_\varphi} \nonumber \\ \times \int_{-\infty}^\infty dz \, e^{-ik_z z} \int_0^\infty \rho \, d\rho \, \psi_0(\rho,z) J_s(k_\rho \rho)\end{aligned}$$ where $J_s(k_\rho \rho)$ is the Bessel function of order $s$. In practice, the Fourier transform and the Hankel transform of order $s$ are performed on discrete grids. The application of this algorithm to solving the GPE is detailed in Ref. [@Ronen06a]. In addition to calculating the ground state with a specified vorticity using this algorithm, we also extended it to calculate the BdG excitation spectrum in the presence of a vortex ground state (see section \[sec:BdG\]), which describes the low-lying excitations and reveals critical information regarding the dynamical stability of the system. This extension of our algorithm is described in Appendix \[app:numerics\]. Excited States and Stability {#sec:BdG} ============================ To study the elementary excitation spectrum of a DBEC with a given projection of angular momentum $s$ in the ground state, we use the BdG *ansatz* and write a wave function of the time-dependent GPE of the form (with $\hbar = 1$) $$\label{BdGansatz} \psi({\mathbf{r}},t) = \left[ \psi_0(\rho,z) + \vartheta({\mathbf{r}},t) \right]e^{i(s\varphi -\mu t)}$$ where $\mu$ is the chemical potential of the ground state, $\psi_0(\rho,z)e^{is\varphi}$, and $\vartheta({\mathbf{r}},t)$ is given by $$\label{vartheta} \vartheta({\mathbf{r}},t) = \delta\left[ u(\rho,z)e^{i(m\varphi -\omega t)} + v^*(\rho,z)e^{-i(m\varphi - \omega t)} \right]$$ where $\delta \ll 1$ to ensure that the excitations have small amplitudes, $\omega$ is the frequency of the excitation and the modes $u(\rho,z)e^{im\varphi}$ and $v^*(\rho,z)e^{-im\varphi}$ are normalized by [@BEC2003] $$\label{uvnorm} \int d{\mathbf{r}}\left[ |u(\rho,z)|^2 - |v(\rho,z)|^2 \right] = 1.$$ The BdG modes are characterized by the quantum number $m$, being their projection of orbital angular momentum onto the $z$-axis. This *ansatz* represents a vortex with angular momentum $\hbar s$ per particle giving rise to excitations with angular momentum $\hbar (s \pm m)$. By inserting Eq. (\[BdGansatz\]) into the time-dependent GPE (Eq. (\[GPE\]) with $\mu$ on the right hand side replaced by $i\hbar \frac{\partial}{\partial t}$) and linearizing about $\delta$, the coupled BdG equations are derived by collecting terms evolving in time like $e^{-i\omega t}$ and $e^{i\omega t}$, respectively, $$\begin{aligned} \label{BdG1} \omega \, u(\rho,z)e^{im\varphi} = \left[ H_0 - \mu \right. \nonumber \\ \left. + \int d{\mathbf{r}}^\prime \psi^*_0(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) \psi_0(\rho^\prime,z^\prime)\right] u(\rho,z)e^{im\varphi} \nonumber \\ + \int d{\mathbf{r}}^\prime \psi_0^*(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) u(\rho^\prime,z^\prime)e^{im\varphi^\prime}\psi_0(\rho,z) \nonumber \\ + \int d{\mathbf{r}}^\prime v(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) \psi_0(\rho^\prime,z^\prime)e^{im\varphi^\prime}\psi_0(\rho,z) \nonumber \\\end{aligned}$$ $$\begin{aligned} \label{BdG2} -\omega \, v(\rho,z)e^{im\varphi} = \left[ H_0 - \mu \right. \nonumber \\ \left. + \int d{\mathbf{r}}^\prime \psi^*_0(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) \psi_0(\rho^\prime,z^\prime)\right] v(\rho,z)e^{im\varphi} \nonumber \\ + \int d{\mathbf{r}}^\prime \psi_0(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) v(\rho^\prime,z^\prime)e^{im\varphi^\prime}\psi_0^*(\rho,z) \nonumber \\ + \int d{\mathbf{r}}^\prime u(\rho^\prime,z^\prime) V_N({\mathbf{r}}-{\mathbf{r}}^\prime) \psi_0^*(\rho^\prime,z^\prime)e^{im\varphi^\prime}\psi_0^*(\rho,z), \nonumber \\\end{aligned}$$ where $H_0 = -\frac{\hbar^2}{2M} \nabla^2 + U({\mathbf{r}})$ and $V_N({\mathbf{r}}-{\mathbf{r}}^\prime) = (N-1)V({\mathbf{r}}-{\mathbf{r}}^\prime)$. By using Hankel transforms to compute the interaction terms in Eqs. (\[BdG1\]) and (\[BdG2\]), we are able to account for the angular dependence of the integrands by using a Bessel function of the appropriate order, as described in Appendix \[app:numerics\]. To solve these equations, we write them in matrix form, as in [@Ronen06a; @Lundh06], and diagonalize them numerically to find the eigenvectors $(u_i,v_i^*)$ and the eigenvalues $\omega_i$. We point out that while there are solutions of the BdG equations of the form $(u_i,v_i^*)$, there are always solutions of the form $(v_i^*,u_i)$ with $\omega_i \rightarrow -\omega_i$ and $m\rightarrow -m$. For the case of $s=0$ BECs, there is a solution of the original form $(u_i,v_i^*)$ with the same $\omega_i$ but with $m \rightarrow -m$. This simply expresses the fact that counter-rotating excitations are degenerate due to the reflection symmetry of the $s=0$ ground state. The presence of a vortex breaks this double degeneracy. We shall say that the excitation $(u_i,v_i^*)$ has a positive norm when $\int d{\mathbf{r}}\left[ |u(\rho,z)|^2-|v(\rho,z)|^2 \right] > 0$. It can then be normalized such that $\int d{\mathbf{r}}\left[ |u(\rho,z)|^2-|v(\rho,z)|^2 \right]=1$. The solution $(v_i^*,u_i)$ with $\omega_i \rightarrow -\omega_i$ and $m \rightarrow -m$ will then have a negative norm, obeying $\int d{\mathbf{r}}\left[ |u(\rho,z)|^2-|v(\rho,z)|^2 \right]=-1$. A positive norm mode with a negative energy eigenvalue signifies that there exists a lower energy solution of the GPE; the same situation is represented by a negative norm mode with a positive energy eigenvalue [@Svidzinsky98]. The solutions of the BdG equations characterize the stability of $s=1$ DBECs. The global thermodynamical instability of $s=1$ DBECs is seen as a negative norm BdG mode with $m=1$ and positive energy for all trap aspect ratios and dipolar interaction strengths. This mode corresponds to the system’s decay into the energetically favored rotationless ground state, just as for BECs with purely contact interactions. The component of the mode with angular dependence $e^{-i(m-s)\varphi}=1$ is in this case rotationless, capturing the symmetry of the $s=0$ ground state. Since this mode populates the core of the vortex, it is refered to as a core mode. However, at ultracold temperatures, thermodynamical stability is less relevant in characterizing the stability of a condensate since there needs to be some thermal processes acting on the system to dissipitavely drive it into a lower energy state. We therefore disregard thermodynamical instability in the following. Instead, we focus on studying the dynamical stability of $s=1$ DBECs. The emergence of a BdG energy eigenvalue with a nonzero imaginary part corresponds to a dynamical instability in the system [@BEC2003]. For example, suppose that a BdG mode $(u_i,v^*_i)$ has energy $\omega_i = \omega_R + i \omega_I$ with $\omega_I \neq 0$; then the mode will have the time dependences $e^{(\omega_I-i\omega_R)t}$ and $e^{-(\omega_I - i\omega_R)t}$, either exponentially growing or decaying in time. Consequently, $\omega_I$ determines the rate of decay of the condensate, given by $\tau = 1/\omega_I$. ![\[fig:stab\] (color online). The black dashed line marks the maximum dipole strength $D$, for a given trap aspect ratio $\lambda$, above which the Gross-Pitaevskii Energy functional has no minimum corresponding to the $s=1$ DBEC. The solid line marks a more restrictive stability line, determined by the onset of dynamical instability, signaled by the emergence of an imaginary energy in an excitation mode. The pink (darker) region represents where oscillations with local minima are observed on dynamically stable states. The inset is an isodensity surface plot of an $s=1$ DBEC at the point in parameter space indicated by the arrow. The ripples in the density are explained in Ref [@Wilson08]. ](stab.eps){width="\columnwidth"} We determine where $s=1$ DBECs are dynamically stable by finding the region in parameter space where all of the BdG modes have purely real energy eigenvalues. This region is shown by the colored portion of Figure \[fig:stab\]. The dashed line in this figure marks, for a given $\lambda$, the $D$ below which we find a local minimum of the GP energy functional [@Ronen06a] by using our reduced 2D algorithm. We find that, for all $\lambda$, the $D$ above which the GP energy functional has no minimum corresponding to an $s=1$ ground state and the $D$ at which the BdG spectrum begins to possess imaginary energy, denoted $D_\mathrm{crit}$, are never equal. Indeed, dynamical instability occurs for values of $D$ at which the GPE has a solution. This is because in the 2D minimization of the vortex-state energy, perturbations that break the $s=1$ angular momentum are not allowed, these are only examined later with the BdG equations. Using a fully 3D calculation, we check the accuracy of $D_\mathrm{crit}$ for various trap aspect ratios by time evolving the condensate wave function with an initial random perturbation. The $D_\mathrm{crit}$ that we calculate using the 3D algorithm, corresponding to the $D$ at which we observe collapse under time evolution, agrees with the $D_\mathrm{crit}$ that we calculate by finding imaginary energy eigenvalues in the BdG spectrum using our 2D algorithm. The pink (darker) region in Figure \[fig:stab\] represents the region where we find dynamically stable $s=1$ ground states having radial ripples with local minima, as illustrated by the inset. This feature has been explained in detail in Ref. [@Wilson08]. We find that $s=1$ DBECs possess imaginary energy in their BdG spectrum only when two modes of opposite norm are degenerate with each other, just as is the case for a BEC with contact interactions. This circumstance was recently studied in Ref. [@Nakamura08], where it is confirmed perturbatively for BECs with contact interactions. Ref. [@Lundh06] also confirmed this claim using a two-mode approximation. Indeed, we find that the same holds true for DBECs, where the only difference between the two systems is the shape of the mean field potential. At all aspect ratios, we observe, for some finite value of $D$, two modes with opposite norm approach and then go degenerate with each other at $D_\mathrm{crit}$. At the point of degeneracy, the modes develop equal and opposite imaginary energies, signifying dynamical instability. If two modes that have the same norm approach each other, they undergo an avoided crossing instead of becoming degenerate. For a BEC with pure contact interactions in the $s=1$ vortex state, the mode that defines the onset of dynamical instability is independent of aspect ratio $\lambda$. Positive contact interactions ensure dynamical stability while negative contact interactions (for $\lambda \gtrsim 0.3$) bring about a dynamical instability due to an $m=2$ mode [@Saito02]. Additionally, the $s=2$ state is dynamically unstable due to an $m=2$ mode for negative contact interactions, while an $m=2$ mode defines windows of dynamical stability for positive contact interactions. This holds true for these systems no matter how oblate the trap. The case for a DBEC, however, is quite different. Figure \[fig:exDimag\] illustrates the imaginary parts of the BdG spectrum for $m=1-5$ for DBECs in traps with aspect ratios $\lambda = 2$ and $\lambda = 15$. Where these imaginary energies are zero, from $D=0$ to $D_\mathrm{crit}$, the condensates are dynamically stable. Notice that for $\lambda = 2$, an $m=2$ mode develops imaginary energy at a $D$ well below the other modes, defining $D_\mathrm{crit}$ for this aspect ratio. However, an $m=4$ mode serves to define $D_\mathrm{crit}$ for $\lambda = 15$. Indeed, unlike BECs with contact interactions, modes with different $m$ quantum numbers serve to define $D_\mathrm{crit}$ at different aspect ratios for DBECs. For moderate trap aspect ratios (such as $\lambda=2$), an $m=2$ mode defines $D_\mathrm{crit}$ for the DBEC, similar to the case for contact interactions. However, as the trap aspect ratio is increased to more oblate shapes, modes with larger $m$ quantum numbers develop imaginary energy eigenvalues at smaller values of $D$ than the $m=2$ mode. Figure \[fig:exm\](a) illustrates this by plotting the *differences* between the $D$’s at which the BdG modes with different angular symmetries first develop imaginary energy eigenvalues, and $D_\mathrm{crit}$, as a function of $\lambda$. Thus, for a given $\lambda$ the lowest curve identifies the symmetry of the unstable mode. For $6 \lesssim \lambda \lesssim 12$, an $m=3$ mode defines $D_\mathrm{crit}$ while for larger aspect ratios, an $m=4$ mode defines $D_\mathrm{crit}$. Although it is not shown here, we find that at even larger aspect ratios the vortex decays into still higher $m$-modes. ![\[fig:exDimag\] (color online). a) The imaginary part of the BdG excitation spectrum for a number of $m$ values for a DBEC in a trap with aspect ratio a) $\lambda = 2$ and b) $\lambda=15$. For $\lambda=2$, the $m=2$ modes clearly develop imaginary energy eigenvalues at a $D$ smaller than the other modes, defining $D_\mathrm{crit}\simeq 8$ for this aspect ratio. For $\lambda=15$, the $m=4$ modes develop imaginary energy at a $D$ smaller than the other modes, defining $D_\mathrm{crit} \simeq 108$ for this aspect ratio.](exDimag.eps){width="\columnwidth"} The relevance of the $m$-dependent dynamical instability is that the dipole-dipole interaction leads a BEC to instability locally and at a fixed length scale, the wavelength of which is determined by the axial harmonic oscillator length. We find that, at the onset of imaginary energy, these modes have radial nodal spacings very similar to that of the roton on the rotationless DBEC, namely $\lambda / 2 \sim \pi l_z$, where $\l_z = \sqrt{\hbar / M \omega_z}$ is the axial harmonic oscillator length. The angular dependence of these modes behaves in the same way. Increasing $\lambda$ decreases the ratio $l_z / a_\mathrm{ho}$, so more radial nodes, fixed by $l_z$, can fit into the condensate for larger $\lambda$. In the same way, more angular nodes can fit into the condensate, therefore bringing about dynamical instability by modes with larger $m$ quantum number, and hence more angular nodes. All of the previously discussed BdG modes that we identify as being responsible for dynamical instability are axially symmetric and nodeless in $z$. Modes that break this axial symmetry can correspond to vortex excitations, where the vortex core itself may tilt or bend, and have been termed “kelvon” modes. Ref. [@Klawunn08] reports that, for a singly-quantized vortex in a DBEC that is otherwise spatially homogeneous, the condensate is dynamically unstable to a kelvon mode when an external periodic potential is applied along the direction of the vortex. We find that, in a harmonically trapped DBEC, a mode with a single node at $z=0$ determines $D_\mathrm{crit}$ for $\lambda \lesssim 0.28$. Modes of this type might therefore correspond to a kelvon-instability in prolate traps, but we leave these considerations for future work. As was done in Ref. [@Saito02] for self-attractive BECs in the singly-quantized vortex state, we perform time-dependent simulations of a DBEC where $D$ is chosen to be just above $D_\mathrm{crit}$, enabling us to go beyond the small deviations from the stationary vortex state and see the actual process of collapse. Initializing the simulations with random noise, we observe collapse, at all aspect ratios, with an angular symmetry corresponding to the $m$ quantum number of the mode that first develops an imaginary energy eigenvalue. ![\[fig:exm\] (color online). a) The difference between the $D$ at which BdG modes with different $m$ quantum numbers first develop imaginary energy eigenvalues and $D_\mathrm{crit}$, the smallest $D$ at which any mode develops an imaginary energy eigenvalue. Modes with larger $m$ serve to define $D_\mathrm{crit}$ for more oblate traps. b) The same, but for negative contact interactions instead of dipole-dipole interactions. An $m=2$ BdG mode is always the first to develop an imaginary energy eigenvalue for this case, except for trap aspect ratios $\lambda \lesssim 0.3$, for which an $m=1$ mode plays this role.](exm.eps){width="\columnwidth"} Stability of Vortices with $s=2$ {#sec:s2} ================================ The dynamical instability of condensates with doubly-quantized vortices and contact interactions has been studied extensively [@Pu99; @Carr06; @Mihalache06; @Lundh06]. These studies report windows of positive scattering length where the BECs are dynamically unstable to an $m=2$ BdG core mode, as well as dynamical instability for all values of negative scattering length due to an $m=2$ mode. Knowing that the dipolar mean field in a DBEC can be engineered to be more attractive or repulsive for smaller or larger trap aspect ratios, respectively, we investigate the presence of these features in $s=2$ DBECs. When the harmonic trap is more spherical, the dipoles are free to stack vertically, creating an attractive mean field in this direction. However, in pancake shaped traps the dipoles create a more repulsive mean field. Thus, for larger trap aspect ratios DBECs are more self-repulsive than for smaller aspect ratios, mimicking the mean field of condensates with repulsive contact interactions. As an example, we calculate the contribution of the dipolar mean field to the energy of a DBEC in a trap with aspect ratio $\lambda = 2$ and with $\lambda = 15$ for a fixed $D=10$. In the $\lambda = 15$ trap, we find that this contribution is about five times larger than in the $\lambda = 2$ trap. Indeed, for a DBEC in a trap with aspect ratio $\lambda = 2$ we find that there exists an $m=2$ mode with a complex energy eigenvalue for all values of $D$. However, for $\lambda = 15$ we find that there are windows in $D$ where an $m=2$ BdG mode has a complex energy eigenvalue, while this same mode has purely real energy outside of these windows, as illustrated in Figure \[fig:k2\]. For trap aspect ratios $\lambda \lesssim 7.5$, there are no windows of dynamical stability and the condensate is dynamically unstable for all $D$. However, windows of dynamical stability appear for aspect ratios $\lambda \gtrsim 7.5$ and continue for larger $\lambda$. As is reported in Ref. [@Lundh06], we find that there is an $m=2$ core mode with negative norm and positive real energy that increases monotonically as it goes successively degenerate with positive norm modes as $D$ is increased to create the windows of dynamical instability. This mode represents the $s=2$ condensate’s instability to splitting into a condensate with two singly quantized vortices. The core mode is thermodynamically unstable for all values of $D$ and is only dynamically unstable for the windows shown in Figure \[fig:k2\]. ![\[fig:k2\] (color online). The positive imaginary part of the energy eigenvalues for a doubly quantized ($s=2$) DBEC in a trap with aspect ratio a) $\lambda=2$ and b) $\lambda=15$. For $\lambda=2$, an $m=2$ mode possesses imaginary energy for all values of $D$, signifying a dynamical instability for all $D$. However, for $\lambda=15$ there are windows of dynamical stability as the $m=2$ mode alternates having and not having imaginary energy for different values of $D$, as is the case for a purely repulsive $s=2$ BEC.](exDimagk2.eps){width="\columnwidth"} Conclusion ========== We have implemented a novel 2D algorithm to study the stability and excitations of harmonically trapped DBECs with single vortices by taking advantage of the cylindrical symmetry of the system. By solving the BdG equations for the $s=1$ DBEC, we systematically map its dynamical stability as a function of trap aspect ratio and dipole-dipole interaction strength $D$. The BdG spectrum reveals a dynamical instability in the form of a complex energy eigenvalue. The value of $D$ at which this imaginary energy appears marks the threshold of dynamical stability, $D_\mathrm{crit}$, for the given trap aspect ratio. By inspecting the BdG spectrum for all $m$ quantum numbers, we determine the symmetry of the mode that is responsible for the dynamical instability in the condensate. We find, in contrast to BECs with purely contact interactions, that DBECs with a singly-quantized vortex go unstable to modes with larger $m$ quantum numbers for larger trap aspect ratios, signifying a type of local collapse of these condensates. We have checked the accuracy of $D_\mathrm{crit}$ for various trap aspect ratios by performing fully 3D simulations. Additionally, we find that, in analogy to a self-repulsive BEC with a doubly-quantized vortex, at larger trap aspect ratios there are successive regions in $D$ where the $s=2$ DBECs are dynamically unstable due to an $m=2$ core mode, while the condensates are dynamically stable outside of these regions. The authors acknowledge the financial support of the U.S. Department of Energy and of the National Science Foundation. Hankel Transforms and Interpolation {#app:numerics} =================================== Consider the Hankel transform of a function $f(\rho)$, $\tilde{f}(k)$, where $f(\rho) \simeq 0$ for $\rho > R$ and $\tilde{f}(k) \simeq 0$ for $k > K$, and define $S \equiv RK$. The discrete Hankel transform (DHT) of order $m$ of $f(\rho)$ is then given by [@Guizar04] $$\label{DHT} \tilde{f}(k_{mi}) = \frac{2}{K^2}\sum_{j=1}^{N}\frac{f(\rho_{mj})}{J^2_{m+1}(\alpha_{mj})}J_m\left( \frac{\alpha_{mj}\alpha_{mi}}{S} \right)$$ where $i,j = 1,\ldots,N$, $\rho_{mj} = \alpha_{mj} / K$, $k_{mi} = \alpha_{mi} / R$ and $\alpha_{mi}$ is the $i^\mathrm{th}$ root of $J_m(\rho)$. Conversely, the inverse DHT of order $m$ of $\tilde{f}(k)$ is given by $$\label{iDHT} f(\rho_{mi}) = \frac{2}{R^2}\sum_{j=1}^{N} \frac{\tilde{f}(k_{mj})}{J^2_{m+1}(\alpha_{mj})} J_m\left( \frac{\alpha_{mj}\alpha_{mi}}{S} \right)_.$$ Eqs. (\[DHT\]) and (\[iDHT\]) show that in order to perform a DHT of order $m$ on a function, the function should be defined on a grid proportional to the zeros of the $m$-order Bessel function, $\rho_{mi}$. For the problem we are considering, the functions $f(\rho)$ and $\tilde{f}(k)$ have the angular dependence $e^{im\varphi}$ and $e^{-imk_\varphi}$, respectively, which is why the DHTs above are written with Bessel functions of order $m$. When we calculate the BdG modes for the vortex states of a DBEC (see section \[sec:BdG\]), we take DHTs of functions like $\psi_0^*(\rho,z)e^{-is\varphi}u(\rho,z)e^{i(m+s)\varphi}$ where $\psi_0^*(\rho,z)$ is defined on a grid proportional to the zeros of the $s$-order Bessel function and $u(\rho,z)$ is defined on a grid proportional the the zeros of the Bessel function of order $m+s$. To perform a DHT on the $\rho$-coordinate of the function $\psi_0^*(\rho,z)u(\rho,z)$, the function must be defined on a grid proportional to the zeros of the Bessel function of order $m$, so $\psi_0^*(\rho,z)$ and $u(\rho,z)$ must both be interpolated onto this grid. To accomplish this, we have developed an accurate interpolation scheme based on the DHT itself. As explained in Ref. [@ArfkenWeber], the function $f(\rho)$ may be expanded in an $m^\mathrm{th}$ order Bessel series, $$\label{B1} f(\rho) = \sum_{i=1}^N c_{mi}J_m\left( \alpha_{mi}\frac{\rho}{R} \right)$$ where the coefficients $c_{mi}$ are given by $$\label{B2} c_{mi} = \frac{2}{R^2\,[J_{m+1}(\alpha_{mi})]^2}\int_0^R f(\rho)J_m\left( \alpha_{mi}\frac{\rho}{R} \right)\rho \, d\rho.$$ The integral in Eq. (\[B2\]) is just the Hankel transform (\[transform\]) with $\alpha_{mi}/R = k_{mi}$, giving the transformed function $\tilde{f}(k_i)$. If $\rho$ is discretized in Eq. (\[B1\]), then this prescription gives exactly Eq. (\[iDHT\]). We wish to consider the case where our function is defined on the grid $\rho_{mi}$, proportional to the zeros of the Bessel function of order $m$, but it needs to be defined on the grid $\rho_{ni}$, with $n\neq m$. To do this, we expand $f(\rho_{ni})$ in a Bessel series, $$\label{int1} f(\rho_{ni}) = \sum_{j=1}^N c_{mj} J_m\left( \alpha_{mj} \frac{\rho_{ni}}{R} \right)$$ where the coefficients $c_{mj}$ are given by Eq. (\[B2\]) and are computed in terms of the zeros of the Bessel function of order $m$. However, in Eq. (\[int1\]), the function is expanded on the grid $\rho_{ni}$, proportional to the zeros of the Bessel function of order $n$. The interpolation algorithm then simply follows by inserting the expression for the coefficients, $$\label{int2} f(\rho_{ni}) = \frac{2}{R^2} \sum_{j=1}^N \frac{\tilde{f}(k_{mj})}{J^2_{m+1}(\alpha_{mj})} J_m\left( \frac{\alpha_{mj}\alpha_{ni}}{S} \right)_,$$ where $\tilde{f}(k_{mj})$ is the discrete Hankel transform of $f(\rho_{mj})$, Eq. (\[DHT\]). [30]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , ****, (). , , , , , , , , ****, (). , , , ****, (). , ****, (). , , , , ****, (). , , , ****, (). , , , ****, (). , , , , ****, (). , ****, (). , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , , , , . , , , , ****, (). , ****, (). , , , , ****, (). , , , , ****, (). , ****, (). , , , , , , , ****, (). , ** (, ), ed. , ****, (). , ** (, ). , , , ****, (). , ****, (). , , , , ****, (). , ****, (). , ****, (). , , , , ****, (). , ****, (). , ** (, ), ed.

ArXiv

--- abstract: 'The purpose of this work is to study the 3D focusing inhomogeneous nonlinear Schrödinger equation $$i u_t +\Delta u+|x|^{-b}|u|^2 u = 0,$$ where $0<b<1/2$. Let $Q$ be the ground state solution of $-Q+\Delta Q+ |x|^{-b}|Q|^{2}Q=0$ and $s_c=(1+b)/2$. We show that if the radial initial data $u_0$ belongs to $H^1(\mathbb{R}^3)$ and satisfies $E(u_0)^{s_c}M(u_0)^{1-s_c}<E(Q)^{s_c}M(Q)^{1-s_c}$ and $\| \nabla u_0 \|_{L^2}^{s_c} \|u_0\|_{L^2}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c}$, then the corresponding solution is global and scatters in $H^1(\mathbb{R}^3)$. Our proof is based in the ideas introduced by Kenig-Merle [@KENIG] in their study of the energy-critical NLS and Holmer-Roudenko [@HOLROU] for the radial 3D cubic NLS.' address: - 'LUIZ G. FARAH Department of Mathematics, Federal University of Minas Gerais, BRAZIL' - 'CARLOS M. GUZMÁN Department of Mathematics, Federal University of Minas Gerais, BRAZIL' author: - 'LUIZ G. FARAH AND CARLOS M. GUZMÁN' bibliography: - 'bibguzman.bib' title: SCATTERING FOR THE RADIAL 3D CUBIC FOCUSING INHOMOGENEOUS NONLINEAR SCHRÖDINGER EQUATION --- \[section\] \[theorem\][Proposition]{} \[theorem\][Lemma]{} \[theorem\][Corollary]{} \[theorem\][Remark]{} \[theorem\][Definition]{} Introduction ============ In this paper, we consider the Cauchy problem, also called the initial value problem (IVP), for the focusing inhomogenous nonlinear Schrödinger (INLS) equation on $\mathbb{R}^3$, that is $$\label{INLS} \left\{\begin{array}{cl} i\partial_tu +\Delta u + |x|^{-b} |u|^2 u =0, & \;\;\;t\in \mathbb{R} ,\;x\in \mathbb{R}^3,\\ u(0,x)=u_0(x), & \end{array}\right.$$ where $u = u(t,x)$ is a complex-valued function in space-time $\mathbb{R}\times\mathbb{R}^3$ and $0<b<1/2$.  Before review some results about the Cauchy problem , let us recall the critical Sobolev index. For a fixed $\delta >0$, the rescaled function $u_\delta(t,x)=\delta^{\frac{2-b}{2}}u(\delta^2 t,\delta x)$ is solution of if only if $u(t,x)$ is a solution. This scaling property gives rise to a scale-invariant norm. Indeed, computing the homogeneus Sobolev norm of $u_\delta(0,x)$ we get $$\|u_\delta(0,.)\|_{\dot{H^s}}=\delta^{s-\frac{3}{2}+\frac{2-b}{2}}\|u_0\|_{\dot{H^s}}.$$ Thus, the scale invariant Sobolev space is $H^{s_c}(\mathbb{R}^3)$, where $s_c=\frac{1+b}{2}$ (the critical Sobolev index). Note that, the restriction $0<b< 1/2$ implies $0<s_c<1$ and therefore we are in the mass-supercritical and energy-subcritical case. In addition, we recall that the INLS equation has the following conserved quantities $$\label{mass} M[u_0]=M[u(t)]=\int_{\mathbb{R}^3}|u(t,x)|^2dx$$ and $$\label{energy} E[u_0]=E[u(t)]=\frac{1}{2}\int_{\mathbb{R}^3}| \nabla u(t,x)|^2dx-\frac{1}{4}\left\| |x|^{-b}|u|^{4}\right\|_{L^1_x},$$ which are calling Mass and Energy, respectively.  Next, we briefly review recent developments on the well-posedness theory for the general INLS equation $$\label{GINLS} \left\{\begin{array}{cl} i\partial_tu +\Delta u + |x|^{-b} |u|^\alpha u =0, & \;\;\;x\in \mathbb{R}^N,\\ u(0,x)=u_0(x). & \end{array}\right.$$ Genoud and Stuart [@GENOUD]-[@GENSTU], using the abstract theory developed by Cazenave [@CAZENAVEBOOK] and some sharp Gagliardo-Nirenberg inequalities, showed that is well-posed in $H^1(\mathbb{R}^N)$ - locally if $0<\alpha <2^*, $ - globally for small initial condition if $\frac{4-2b}{N}<\alpha <\frac{4-2b}{N-2}$, - globally for any initial condition if $ 0 < \alpha < \frac{4-2b}{N}$, - globally if $\alpha = \frac{4-2b}{N}$, assuming $\|u_0\|_{L^2}<\|Q\|_{L^2}$, where $Q$ is the ground state of the equation $-Q+\Delta Q+ |x|^{-b}|Q|^{\frac{4-2b}{N}}Q=0$ and $2^*=\frac{4-2b}{N-2}$, if $N\geq 3$ or $2^* =\infty$, if $N=1,2$. Also, Combet and Genoud [@COMGEN] established the classification of minimal mass blow-up solutions of with $L^2$ critical nonlinearity, that is, $\alpha = \frac{4-2b}{N}$.\ Recently, the second author in [@CARLOS], using the contraction mapping principle based on the Strichartz estimates, proved that the IVP is locally well-posed in $H^1(\mathbb{R}^N)$, for $0<\alpha<2^*$. Moreover, for $N\geq 2$, $\frac{4-2b}{N}<\alpha <2^*$ these solutions are global in $H^1(\mathbb{R}^N)$ for small initial data. It worth mentioning that Genoud and Stuart [@GENOUD]-[@GENSTU] consider $0<b<\min\{2,N\}$, and second author in [@CARLOS] assume $0<b<\widetilde{2}$, where $\widetilde{2}=N/3$ if $N=1,2,3$ and $\widetilde{2}=2$ if $N\geq 4$. This new restriction on $b$ is needed to estimate the nonlinear part of the equation in order to use the well known Strichartz estimates associated to the linear flow.  On the other hand, since $$\label{QI1} \|u_\delta\|_{L^2_x}=\delta^{-s_c}\|u\|_{L^2_x},\;\;\;\;\|\nabla u_\delta\|_{L^2_x}=\delta^{1-s_c}\|\nabla u\|_{L^2_x}$$ and $$\left \| |x|^{-b}|u_\delta|^{4} \right\|_{L^1_x}=\delta^{2(1-s_c)}\left \| |x|^{-b}|u|^{4} \right\|_{L^1_x},$$ the following quantities enjoy a scaling invariant property $$\label{QI2} E[u_\delta]^{s_c}M[u_\delta]^{1-s_c}=E[u]^{s_c}M[u]^{1-s_c},\;\;\|\nabla u_\delta\|^{s_c}_{L^2_x}\|u_\delta\|^{1-s_c}_{L^2_x}=\|\nabla u\|^{s_c}_{L^2_x}\|u\|^{1-s_c}_{L^2_x}.$$ These quantities were introduced in Holmer-Roudenko [@HOLROU] in the context of mass-supercritical and energy-subcritical nonlinear Schrödinger equation (NLS), which is equation with $b=0$, and they were used to understand the dichotomy between blowup/global regularity. Indeed, in [@HOLROU], the authors consider the $3D$ cubic NLS and proved that if the initial data $u_0\in H^1(\mathbb{R}^3)$ is radial and satisfy $$\label{ENLS} E(u_0)M(u_0)<E(Q)M(Q)$$ and $$\label{MNLS} \| \nabla u_0 \|_{L^2} \|u_0\|_{L^2}<\|\nabla Q \|_{L^2} \|Q\|_{L^2},$$ then the corresponding solution $u(t)$ of the Cauchy problem (with $b=0$) is globally defined and scatters[^1] in $H^1(\mathbb{R}^3)$ where $Q$ is the ground state solution of the nonlinear elliptic equation $-Q+\Delta Q+|Q|^{2}Q=0$. The subsequent work Duyckaerts-Holmer-Roudenko [@DUCHOLROU] has removed the radial assumption on the initial data. In both papers, they used the method of the concentration-compactness and rigidity technique employed by Kenig-Merle [@KENIG] in their study of the energy critical NLS. Inspired by these works, we investigate same problem for the IVP .\ The results in Holmer-Roudenko [@HOLROU] and Duyckaerts-Holmer-Roudenko [@DUCHOLROU] have been generalized for the general NLS equation (with $b=0$) in the mass-supercritical and energy-subcritical case, by Fang-Xie-Cazenave [@JIANCAZENAVE] and Guevara [@GUEVARA]. Moreover, the recent works of Hong [@HONG] and Killip-Murphy-Visan-Zheng [@KMVZ2016] also obtained analogous result for the cubic focusing NLS equation perturbed by a potential. It’s worth mentioning that global well-posedness and scattering for the mass critical and energy critical NLS has also received a lot of attention in the literature and we refer to Dodson [@D12]-[@D15]-[@D16], Tao-Visan-Zhang [@TVZ07], Killip-Tao-Visan [@KTV09], Killip-Visan-Zhang [@KVZ08], Colliander-Keel-Staffilani-Takaoka-Tao [@CKSTT08], Ryckman-Visan [@RV07], Visan [@V07] and Killip-Visan [@KV10] for the results in these directions.  In a recent work of the first author in [@LG] showed global well-posedness for the $L^2$-supercritical and $H^1$-subcritical inhomogeneous nonlinear Schrödinger equation under assumptions similar to -. Below we state his result for the $3D$ cubic INLS, since this is the case we are interested in the present work. \[TG\]Let $0<b<1$. Suppose that $u(t)$ is the solution of with initial data $u_0\in H^1(\mathbb{R}^3)$ satisfying $$\label{EMC} E[u_0]^{s_c}M[u_0]^{1-s_c}<E[Q]^{s_c}M[Q]^{1-s_c}$$ and $$\label{GFC}\| \nabla u_0 \|_{L^2}^{s_c} \|u_0\|_{L^2}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c},$$ then $u(t)$ is a global solution in $H^1(\mathbb{R}^3)$. Furthermore, for any $t\in \mathbb{R}$ we have $$\label{GR}\| \nabla u(t) \|_{L^2}^{s_c} \|u(t)\|_{L^2}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c},$$ where $Q$ is unique positive solution of the elliptic equation $$\label{GSE} -Q+\Delta Q+ |x|^{-b}|Q|^{2}Q=0.$$ In [@LG Teorema $1.6$] the author also considers the case $$\| \nabla u_0 \|_{L^2}^{s_c} \|u_0\|_{L^2}^{1-s_c}>\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c}.$$ Indeed assuming the last relation and then the solution blows-up in finite time if the initial data $u_0$ has finite variance, i.e., $|x|u_0\in L^2(\mathbb{R}^3)$. This is the extension to the INLS model of the result proved by Holmer-Roudenko [@HOLROU2] for the NLS equation.  Our aim in this paper is to show that the global solutions obtained in Theorem \[TG\] also scatters (in the radial case) according to the following definition \[SCATTER\] A global solution $u(t)$ to the Cauchy problem scatters forward in time in $H^1(\mathbb{R}^3)$, if there exists $\phi^+\in H^1(\mathbb{R}^3)$ such that $$\lim_{t\rightarrow +\infty}\|u(t)-U(t)\phi^+\|_{H^1}=0.$$ Also, we say that $u(t)$ scatters backward in time if there exist $\phi^-\in H^1(\mathbb{R}^3)$ such that $$\lim_{t\rightarrow -\infty}\|u(t)-U(t)\phi^-\|_{H^1}=0.$$ Here, $U(t)$ denotes unitary group associated to the linear equation $i\partial_tu +\Delta u=0$, with initial data $u_0$. The precise statement of our main theorem is the following. \[SCATTERING\] Let $u_0\in H^1(\mathbb{R}^3)$ be radial and $0<b<1/2$. Suppose that and are satisfied then the solution $u$ of is global in $H^1(\mathbb{R}^3)$ and scatters both forward and backward in time. The above theorem extends the result obtained by Holmer-Roudenko [@HOLROU] to the INLS model. On the other hand, since the solutions of the INLS equation do not enjoy conservation of Momentum, we were not able to use the same ideas introduced by Duyckaerts-Holmer-Roudenko [@DUCHOLROU] to remove the radial assumption.  The plan of this work is as follows: in the next section we introduce some notations and estimates. In Section $3$, we sketch the proof of our main result (Theorem \[SCATTERING\]), assuming all the technical points. In Section $4$, we collect some preliminary results about the Cauchy problem . Next in Section $5$, we recall some properties of ground state and show the existence of wave operator. In Section $6$, we construct a critical solution denoted by $u_c$ and show some of its properties (the key ingredient in this step is a profile decomposition result related to the linear flow). Finally, Section $7$ is devoted to the rigidity theorem. Notation and preliminaries ==========================  Let us start this section by introducing the notation used throughout the paper. We use $c$ to denote various constants that may vary line by line. Given any positive numbers $a$ and $b$, the notation $a\lesssim b$ means that there exists a positive constant $c$ that $a\leq cb$, with $c$ uniform with respect to the set where a and b vary. Let a set $A\subset \mathbb{R}^3$, $A^C=\mathbb{R}^N \backslash A$ denotes the complement of $A$. Given $x,y \in \mathbb{R}^3$, $x\cdot y$ denotes the inner product of $x$ and $y$ in $\mathbb{R}^3$.  We use $\|.\|_{L^p}$ to denote the $L^p(\mathbb{R}^3)$ norm with $p\geq 1$. If necessary, we use subscript to inform with variable we are concerned with. The mixed norms in the spaces $L^q_tL^r_x$ and $L^q_TL^r_x$ of $f(x,t)$ are defined, respectively, as $$\|f\|_{L^q_tL^r_x}=\left(\int_{\mathbb{R}}\|f(t,.)\|^q_{L^r_x}dt\right)^{\frac{1}{q}}$$ and $$\|f\|_{L^q_TL^r_x}=\left(\int_T^\infty\|f(t,.)\|^q_{L^r_x}dt\right)^{\frac{1}{q}}$$ with the usual modifications when $q=\infty$ or $r=\infty$.  For $s\in \mathbb{R}$, $J^s$ and $D^s$ denote the Bessel and the Riesz potentials of order $s$, given via Fourier transform by the formulas $$\widehat{J^s f}=(1+|y|^2)^{\frac{s}{2}}\widehat{f}\;\;\;\textnormal{and} \;\;\;\;\widehat{D^sf}=|y|^s\widehat{f},$$ where the Fourier transform of $f(x)$ is given by $$\widehat{f}(y)=\int_{\mathbb{R}^3}e^{ix.y}f(x)dx.$$ On the other hand, we define the norm of the Sobolev spaces $H^{s,r}(\mathbb{R}^3)$ and $\dot{H}^{s,r}(\mathbb{R}^3)$, respectively, by $$\|f\|_{H^{s,r}}:=\|J^sf\|_{L^r}\;\;\;\;\textnormal{and}\;\;\;\;\|f\|_{\dot{H}^{s,r}}:=\|D^sf\|_{L^r}.$$ If $r=2$ we denote $H^{s,2}=H^s$ and $\dot{H}^{s,2}=\dot{H}^s$.  Next, we recall some Strichartz type estimates associated to the linear Schrödinger propagator.\ **Strichartz type estimates.** We say the pair $(q,r)$ is $L^2$-admissible or simply admissible par if they satisfy the condition $$\label{L2admissivel} \frac{2}{q}=\frac{3}{2}-\frac{3}{r},$$ where $2\leq r \leq 6$. We also called the pair $\dot{H}^s$-admissible if[^2] $$\label{Hsadmissivel}\frac{2}{q}=\frac{3}{2}-\frac{3}{r}-s,$$ where $\frac{6}{3-2s}\leq r \leq 6^{-}$. Here, $a^-$ is a fixed number slightly smaller than a ($a^-=a-\varepsilon$ with $\varepsilon>0$ small enough) and, in a similar way, we define $a^+$. Finally we say that $(q,r)$ is $\dot{H}^{-s}$-admissible if $$\frac{2}{q}=\frac{3}{2}-\frac{3}{r}+s,$$ where $\left(\frac{6}{3-2s}\right)^{+}\leq r \leq 6^{-}$.   Given $s\in \mathbb{R}$, we use the set $\mathcal{A}_s=\{(q,r);\; (q,r)\; \textnormal{is} \;\dot{H}^s\textnormal{-admissible}\}$ to define the Strichartz norm $$\|u\|_{S(\dot{H}^{s})}=\sup_{(q,r)\in \mathcal{A}_{s}}\|u\|_{L^q_tL^r_x}.$$ In the same way, the dual Strichartz norm is given by $$\|u\|_{S'(\dot{H}^{-s})}=\inf_{(q,r)\in \mathcal{A}_{-s}}\|u\|_{L^{q'}_tL^{r'}_x},$$ where $(q',r')$ is such that $\frac{1}{q}+\frac{1}{q'}=1$ and $\frac{1}{r}+\frac{1}{r'}=1$ for $(q,r)\in \mathcal{A}_s$.  Note that, if $s=0$ then $\mathcal{A}_0$ is the set of all $L^2$-admissible pairs. Moreover, if $s=0$, $S(\dot{H}^0)=S(L^2)$ and $S'(\dot{H}^{0})=S'(L^2)$. We write $S(\dot{H}^s)$ or $S'(\dot{H}^{-s})$ if the mixed norm is evaluated over $\mathbb{R}\times\mathbb{R}^3$. To indicate a restriction to a time interval $I\subset (-\infty,\infty)$ and a subset $A$ of $\mathbb{R}^3$, we use the notations $S(\dot{H}^s(A);I)$ and $S'(\dot{H}^{-s}(A);I)$.  The next lemmas provide some inequalities that will be useful in our work. \[ILE\] If $t\neq 0$, $\frac{1}{p}+\frac{1}{p'}=1$ and $p'\in[1,2]$, then $U(t):L^{p'}(\mathbb{R}^3)\rightarrow L^p(\mathbb{R}^3)$ is continuous and $$\|U(t)f\|_{L^p_x}\lesssim|t|^{-\frac{3}{2}\left(\frac{1}{p'}-\frac{1}{p}\right)}\|f\|_{L^{p'}}.$$ See Linares-Ponce [@FELGUS Lemma $4.1$]. **(Sobolev embedding)**\[SI\] Let $1\leq p<+\infty$. If $s\in \left(0,\frac{3}{2}\right)$ then $H^{s}(\mathbb{R}^3)$ is continuously embedded in $L^r(\mathbb{R}^3)$ where $s=\frac{3}{p}-\frac{3}{r}$. Moreover, $$\label{SEI} \|f\|_{L^r}\leq c\|D^sf\|_{L^{2}}.$$ See Linares-Ponce [@FELGUS Theorem $3.3$]. \[SEI2\] Using Lemma \[SI\] we have that $H^s(\mathbb{R}^3)$ is continuously embedded in $L^r(\mathbb{R}^3)$ and $$\label{SEI22} \|f\|_{L^r}\leq c\|f\|_{H^{s}},$$ where $r\in[2,\frac{6}{3-2s}]$.  Next we list the well-known Strichartz estimates we are going to use in this work. We refer the reader to Linares-Ponce [@FELGUS] and Kato [@KATO] for detailed proofs of what follows (see also Holmer-Roudenko [@HOLROU] and Guevara [@GUEVARA]). \[Lemma-Str\] The following statements hold. - (Linear estimates). $$\label{SE1} \| U(t)f \|_{S(L^2)} \leq c\|f\|_{L^2},$$ $$\label{SE2} \| U(t)f \|_{S(\dot{H}^s)} \leq c\|f\|_{\dot{H}^s}.$$ - (Inhomogeneous estimates). $$\label{SE3} \left \| \int_{\mathbb{R}} U(t-t')g(.,t') dt' \right\|_{S(L^2)}\;+\; \left \| \int_{0}^t U(t-t')g(.,t') dt' \right \|_{S(L^2) } \leq c\|g\|_{S'(L^2)},$$ $$\label{SE5} \left \| \int_{0}^t U(t-t')g(.,t') dt' \right \|_{S(\dot{H}^s) } \leq c\|g\|_{S'(\dot{H}^{-s})}.$$  We end this section with three important remarks. \[nonlinerity\] Let $F(x,z)=|x|^{-b}|z|^2 z$, and $f(z)=|z|^2 z$. The complex derivative of $f$ is $f_z(z)=2|z|^2$ and $f_{\bar{z}}(z)=z^2$. For $z,w\in \mathbb{C}$, we have $$f(z)-f(w)=\int_{0}^{1}\left[f_z(w+\theta(z-w))(z-w)+f_{\bar{z}}(w+\theta(z-w))\overline{(z-w)}\right]d\theta.$$ Thus, $$\label{FEI} |F(x,z)-F(x,w)|\lesssim |x|^{-b}\left( |z|^2+ |w|^2 \right)|z-w|.$$  Now we are interested in estimating $\nabla \left( F(x,z)-F(x,w) \right)$. A simple computation gives $$\label{NONLI11} \nabla F(x,z)=\nabla(|x|^{-b})f(z)+|x|^{-b} \nabla f(z)$$ where $\nabla f(z)=f'(z)\nabla z=f_z(z)\nabla z+f_{\bar{z}}(z) \overline{\nabla z}$.\ First we estimate $|\nabla (f(z)-f(w))|$. Note that $$\label{NONLI55} \nabla (f(z)-f(w))=f'(z)(\nabla z-\nabla w)+(f'(z)-f'(w))\nabla w.$$ So, since $$\label{NONLI66} |f_z(z)-f_z(w)|\;,\;| f_{\bar{z}}(z)-f_{\bar{z}}(w)|\lesssim (|z|+|w|)|z-w|$$ we get, by $$\label{NONLI88} |\nabla (f(z)-f(w))|\lesssim |z|^2|\nabla (z- w)|+(|z|+|w|)|\nabla w||z-w|.$$ Therefore, by , and the last two inequalities we obtain $$\label{SECONDEI} \left|\nabla \left(F(x,z)-F(x,w)\right)\right|\lesssim |x|^{-b-1}(|z|^{2}+|w|^{2})|z-w|+|x|^{-b}|z|^2|\nabla (z- w)|+M,$$ where $M \lesssim |x|^{-b}(|z|+|w|)|\nabla w||z-w|$. \[RB\] Let $B=B(0,1)=\{ x\in \mathbb{R}^3;|x|\leq 1\}$ and $b>0$. If $x\in B^C$ then $|x|^{-b}<1$ and so $$\left \||x|^{-b}f \right\|_{L^r_x}\leq \|f\|_{L_x^r(B^C)}+\left\||x|^{-b}f\right\|_{L_x^r(B)}.$$  The next remark provides a condition for the integrability of $|x|^{-b}$ on $B$ and $B^C$. \[RIxb\] Note that if $\frac{3}{\gamma}-b>0$ then $\||x|^{-b}\|_{L^\gamma(B)}<+\infty$. Indeed $$\int_{B}|x|^{-\gamma b}dx=c\int_{0}^{1}r^{-\gamma b}r^{2}dr=c_1 \left. r^{3-\gamma b} \right |_0^1<+\infty\;\;\textnormal{if}\;\;\frac{3}{\gamma} - b>0.$$ Similarly, we have that $\||x|^{-b}\|_{L^\gamma(B^C)}$ is finite if $\frac{3}{\gamma}- b<0$. Sketch of the proof of Theorem \[SCATTERING\] {#SPMR} =============================================  Similarly as in the NLS model, we have the following scattering criteria for global solution in $H^1(\mathbb{R}^3)$ (the proof will be given after Proposition \[GWPH1\] below). \[SCATTERSH1\][**($H^1$ scattering)**]{} Let $0<b<1/2$. If $u(t)$ be a global solution of with initial data $u_0 \in H^1(\mathbb{R}^3)$. If $\|u\|_{S(\dot{H}^{s_c})}< +\infty$ and $\sup\limits_{t\in \mathbb{R}}\|u(t)\|_{H^1_x}\leq B$, then $u(t)$ scatters in $H^1(\mathbb{R}^3)$ as $t \rightarrow \pm\infty$.  Let $u(t)$ be the corresponding $H^1$ solution for the Cauchy problem with radial data $u_0\in H^1(\mathbb{R}^3)$ satisfying and . We already know by Theorem \[TG\] that the solution is globally defined and $\sup\limits_{t\in \mathbb{R}}\|u(t)\|_{H^1}< \infty$. So, in view of Proposition \[SCATTERSH1\], our goal is to show that (recalling $s_c=\frac{1+b}{2}$) $$\label{HsFINITE} \|u\|_{S(\dot{H}^{s_c})}<+\infty.$$ The technique employed here to achieve the scattering property combines the concentration-compactness and rigidity ideas introduced by Kenig-Merle [@KENIG]. It is also based on the works of Holmer-Roudenko [@HOLROU] and Duyckaerts-Holmer-Roudenko [@DUCHOLROU]. We describe it in the sequel, but first we need some preliminary definitions. We shall say that SC($u_0$) holds if the solution $u(t)$ with initial data $u_0\in H^1(\mathbb{R}^3)$ is global and holds. For each $\delta > 0$ define the set $A_\delta$ to be the collection of all initial data in $H^1(\mathbb{R}^3)$ satisfying $$\begin{aligned} A_\delta=\{u_0\in H^1:E[u_0]^{s_c}M[u_0]^{1-s_c}<\delta\;\textnormal{and}\;\|\nabla u_0\|^{s_c}_{L^2}\| u_0\|^{1-s_c}_{L^2}<\|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2} \}\end{aligned}$$ and define $$\label{deltac} \delta_c=\sup \{\; \delta>0:\; u_0\; \in A_\delta\; \Longrightarrow SC(u_0)\; \textnormal{holds} \}=\sup_{\delta>0} B_\delta.$$ Note that $B_\delta \neq \emptyset$. In fact, applying the Strichartz estimate , interpolation and Lemma \[LGS\] (i) below, we obtain $$\begin{aligned} \|U(t)u_0\|_{S(\dot{H}^{s_c})}&\leq& c\|u_0\|_{\dot{H}^{s_c}}\leq c\|\nabla u_0\|^{s_c}_{L^2}\|u_0\|^{1-s_c}_{L^2}\\ &\leq& c\left(\frac{3+b}{ s_c}\right)^{\frac{s_c}{2}}E[u_0]^{\frac{s_c}{2}}M[u_0]^{\frac{1-s_c}{2}}. \end{aligned}$$ So if $u_0\in A_\delta$ then $E[u_0]^{s_c}M[u_0]^{1-s_c}<\left(\frac{ s_c}{3+2b}\right)^{s_c}\delta'^2$, which implies $\|U(t)u_0\|_{S(\dot{H}^{s_c})}\leq c\delta'$. Then, by the small data theory (Proposition \[GWPH1\] below) we have that $SC(u_0)$ holds for $\delta'>0$ small enough.  Next, we sketch the proof of Theorem \[SCATTERING\]. If $\delta_c\geq E[Q]^{s_c}M[Q]^{1-s_c}$ then we are done. Assume now, by contradiction, that $\delta_c<E[Q]^{s_c}M[Q]^{1-s_c}$. Therefore, there exists a sequence of radial solutions $u_n$ to with $H^1$ initial data $u_{n,0}$ (rescale all of them to have $\|u_{n,0}\|_{L^2} = 1$ for all $n$) such that[^3] $$\label{CC0} \|\nabla u_{n,0}\|^{s_c}_{L^2} < \|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2}$$ and $$E[u_n]^{s_c} \searrow \delta_c\; \textnormal{as}\; n \rightarrow +\infty,$$ for which SC($u_{n,0}$) does not hold for any $n\in \mathbb{R}^3$. However, we already know by Theorem \[TG\] that $u_n$ is globally defined. Hence, we must have $\|u_n\|_{S(\dot{H}^{s_c})}=+\infty$. Then using a profile decomposition result (see Proposition \[LPD\] below) on the sequence $\{u_{n,0}\}_{n\in \mathbb{N}}$ we can construct a critical solution of , denoted by $u_c$, that lies exactly at the threshold $\delta_c$, satisfies (therefore $u_c$ is globally defined again by Theorem \[TG\]) and $\|u_c\|_{S(\dot{H}^{s_c})}=+\infty$ (see Proposition \[ECS\] below). On the other hand, we prove that the critical solution $u_c$ has the property that $K=\{u_c(t):t\in[0,+\infty)\}$ is precompact in $H^1(\mathbb{R}^3)$ (see Proposition \[PSC\] below). Finally, the rigidity theorem (Theorem \[RT\] below) will imply that such critical solution is identically zero, which contradicts the fact that $\|u_c\|_{S(\dot{H}^{s_c})}=+\infty$. Cauchy Problem ==============  In this section we show a miscellaneous of results for the Cauchy problem . These results will be useful in the next sections. We start stating the following two lemmas. To this end, we use the following numbers $$\label{PA1} \widehat{q}=\frac{4(4-\theta)}{6+2b-\theta(1+b)},\;\;\;\widehat{r}\;=\;\frac{6(4-\theta)}{2(3-b)-\theta(2-b)},$$ and $$\label{PA2} \widetilde{a}\;=\;\frac{2(4-\theta)}{(7+2b-3\theta)-(2-b)(1-\theta)},\;\;\; \widehat{a}=\frac{2(4-\theta)}{1-b}.$$ It is easy to see that $(\widehat{q},\widehat{r})$ is $L^2$-admissible, $(\widehat{a},\widehat{r})$ is $\dot{H}^{s_c}$-admissible and $(\widetilde{a},\widehat{r})$ is $\dot{H}^{-s_c}$-admissible. \[LG1\] Let $0<b<1$, then there exist $c>0$ and a positive number $\theta<2$ such that - $\left \||x|^{-b}|u|^2 v \right\|_{S'(\dot{H}^{-s_c})} \leq c \| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c})} \|v\|_{S(\dot{H}^{s_c})}$, - $\left\||x|^{-b}|u|^2 v \right\|_{S'(L^2)}\leq c\| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c})} \| v\|_{S(L^2)}$. \(i) We divide the estimate in $B$ and $B^C$, indeed $$\left \||x|^{-b}|u|^2 v \right\|_{S'(\dot{H}^{-s_c})} \leq \left \||x|^{-b}|u|^2 v \right\|_{S'\left(\dot{H}^{-s_c}(B)\right)} + \left \||x|^{-b}|u|^2 v \right\|_{S'\left(\dot{H}^{-s_c}(B^C)\right)}.$$  We first consider the estimate on $B$. By the Hölder inequality we deduce $$\begin{aligned} \label{LG1Hs1} \left \| |x|^{-b}|u|^2 v \right\|_{L^{\widehat{r}'}_x(B)} &\leq& \||x|^{-b}\|_{L^\gamma(B)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{(2-\theta)r_2}} \|v\|_{L^{\widehat{r}}_x} \nonumber \\ &=&\||x|^{-b}\|_{L^\gamma(B)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|v\|_{L^{\widehat{r}}_x},\end{aligned}$$ where $$\label{LG1Hs2} \frac{1}{\widehat{r}'}=\frac{1}{\gamma}+\frac{1}{r_1}+\frac{1}{r_2}+\frac{1}{\widehat{r}}\;\;\textnormal{and}\;\;\widehat{r}=(2-\theta)r_2.$$ In order to have the norm $\||x|^{-b}\|_{L^\gamma(B)}$ bounded we need $\frac{3}{\gamma}>b$ (see Remark \[RIxb\]). In fact, observe that implies $$\frac{3}{\gamma}=3-\frac{3(4-\theta)}{\widehat{r}}-\frac{3}{r_1},$$ and from it follows that $$\label{LG1Hs3} \frac{3}{\gamma}-b=\frac{\theta(2-b)}{2}-\frac{3}{r_1}.$$ Choosing $r_1>1$ such that $\theta r_1=6$ we obtain $\frac{3}{\gamma}-b=\theta(1-b)>0$ since $b<1$, that is, $|x|^{-b}\in L^\gamma (B)$. Moreover, using the Sobolev embedding (with $s=1$) and we get $$\label{LG1Hs4} \left \| |x|^{-b}|u|^2 v \right\|_{L^{\widehat{r}'}_x(B)} \leq c\|u\|^{\theta}_{H^1_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|v\|_{L^{\widehat{r}}_x}.$$ On the other hand, we claim that $$\label{LG1Hs41} \left \| |x|^{-b}|u|^2 v \right\|_{L^{\widehat{r}'}_x(B^C)} \leq c\|u\|^{\theta}_{H^1_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|v\|_{L^{\widehat{r}}_x}.$$ Indeed, Arguing in the same way as before we deduce $$\begin{aligned} \left \| |x|^{-b}|u|^2 v \right\|_{L^{\widehat{r}'}_x(B^C)} \leq \||x|^{-b}\|_{L^\gamma(B^C)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|v\|_{L^{\widehat{r}}_x},\end{aligned}$$ where the relation holds. By Remark \[RIxb\], to show that $\||x|^{-b}\|_{L^\gamma(B^C)}$ is finite we need to verify that $\frac{3}{\gamma}-b<0$. Indeed, choosing $r_1>1$ such that $\theta r_1=2$ and using $\eqref{LG1Hs3}$ we have $\frac{3}{\gamma}-b=-\frac{\theta(1+b)}{2}$, which is negative. Therefore the Sobolev inequality implies . This completes the proof of the claim.  Now, inequalities and yield $$\label{LG1Hs5} \left \| |x|^{-b}|u|^2 v \right\|_{L^{\widehat{r}'}_x} \leq c\|u\|^{\theta}_{H^1_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|v\|_{L^{\widehat{r}}_x}$$ and the Hölder inequality in the time variable leads to $$\begin{aligned} \left \| |x|^{-b}|u|^2 v \right\|_{L_t^{\widetilde{a}'}L^{\widehat{r}'}_x}&\leq& c \|u\|^{\theta}_{L^\infty_tH^1_x} \|u\|^{2-\theta}_{ L_t^{(2-\theta)a_1} L_x^{\widehat{r}}} \|v\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x} \nonumber \\ &=& c \|u\|^{\theta}_{L^\infty_tH^1_x} \|u\|^{2-\theta}_{ L_t^{\widehat{a}} L_x^{\widehat{r}}} \|v\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x},\end{aligned}$$ where $$\label{L4i} \frac{1}{\widetilde{a}'}=\frac{2-\theta}{\widehat{a}}+\frac{1}{\widehat{a}}.$$ Since $\widehat{a}$ and $\widetilde{a}$ defined in satisfy we conclude the proof of item[^4] (i).  (ii) In the previous item we already have , then applying Hölder’s inequality in the time variable we obtain $$\label{LGHsii} \left\| |x|^{-b}|u|^2 v\right \|_{L_t^{\widehat{q}'}L^{\widehat{r}'}_x}\leq c \|u\|^{\theta}_{L^\infty_tH^s_x}\|u\|^{2-\theta}_{L_t^{\widehat{a}} L_x^{\widehat{r}}} \|v\|_{L^{\widehat{q}}_tL^{\widehat{r}}_x},$$ since $$\label{LG2Hs1} \frac{1}{\widehat{q}'}=\frac{2-\theta}{\widehat{a}}+\frac{1}{\widehat{q}}$$ by and . The proof is finished since $(\widehat{q},\widehat{r})$ is $L^2$-admissible. \[RGP\] In the perturbation theory we use the following estimate $$\left \||x|^{-b}|u| v w\right\|_{S'(L^2)} \leq c \| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{1-\theta}_{S(\dot{H}^{s_c})} \|v\|_{S(\dot{H}^{s_c})}\| w\|_{S(L^2)},$$ where $\theta\in(0,1)$ is a sufficiently small number. Its proof follows from the ideas of Lemma \[LG1\] (ii), that is, we can repeat all the computations replacing $|u|^2 v$ by $|u|vw$ or, to be more precise, replacing $|u|^2 v=|u|^\theta |u|^{2-\theta }v$ by $|u| vw=|u|^\theta |u|^{1-\theta }vw$. \[LG3\] Let $0<b<1/2$. There exist $c>0$ and $\theta\in (0,2)$ sufficiently small such that $$\left\|\nabla (|x|^{-b}|u|^2 u)\right\|_{S'(L^2)}\leq c\| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c})} \|\nabla u\|_{S(L^2)}.$$ Since $(2,6)$ is $L^2$-admissible in 3D and applying the product rule for derivatives we have $$\begin{aligned} \left\|\nabla\left(|x|^{-b}|u|^2 u \right)\right\|_{S'(L^2)}&\leq& \left\||x|^{-b}\nabla\left(|u|^2 u \right)\right\|_{S'(L^2)}+\left\|\nabla\left(|x|^{-b}\right)|u|^2 u \right\|_{S'(L^2)}\\ &\leq & \left\||x|^{-b}\nabla\left(|u|^2 u \right)\right\|_{L_t^{\widehat{q}'}L^{\widehat{r}'}_x} + \left\|\nabla\left(|x|^{-b}\right)|u|^2 u \right\|_{L_t^{2'}L^{6'}_x}\\ &\leq & N_1+N2.\end{aligned}$$  First, we estimate $N_1$ (dividing in $B$ and $B^C$). It follows from Hölder’s inequality that $$\begin{aligned} \label{LG3Hs3} \left\||x|^{-b}\nabla\left(|u|^2 u \right)\right\|_{L^{\widehat{r}'}_x(B)} &\leq& \||x|^{-b}\|_{L^\gamma(B)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{(2-\theta)r_2}} \|\nabla u\|_{ L^{\widehat{r}}_x}\nonumber \\ &=& \||x|^{-b}\|_{L^\gamma(B)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|\nabla u\|_{ L^{\widehat{r}}_x},\end{aligned}$$ where $$\label{LG3Hs4} \frac{1}{\widehat{r}'}=\frac{1}{\gamma}+\frac{1}{r_1}+\frac{1}{r_2}+\frac{1}{\widehat{r}}\;\;\;\;\textnormal{and}\;\;\;\; \widehat{r}=(2-\theta)r_2.$$ Notice that the right hand side of is the same as the right hand side of , with $v=\nabla u$. Thus, arguing in the same way as in Lemma \[LG1\] (i) we obtain $$\left \| |x|^{-b}\nabla\left(|u|^2 u \right) \right\|_{L^{\widehat{r}'}_x(B)} \leq c\|u\|^{\theta}_{H^1_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|\nabla u\|_{L^{\widehat{r}}_x}.$$ We also obtain, by Lemma \[LG1\] (i) $$\left \| |x|^{-b}\nabla\left(|u|^2 u \right) \right\|_{L^{\widehat{r}'}_x(B^C)} \leq c\|u\|^{\theta}_{H^1_x} \|u\|^{2-\theta}_{L_x^{\widehat{r}}} \|\nabla u\|_{L^{\widehat{r}}_x}.$$ Moreover, the Hölder inequality in the time variable leads to (since $\frac{1}{\widetilde{q}'}=\frac{2-\theta}{\widehat{a}}+\frac{1}{\widehat{q}}$) $$\begin{aligned} \label{SLG0} N_1=\left \| |x|^{-b}|u|^2 \nabla u \right\|_{L_t^{\widetilde{q}'}L^{\widehat{r}'}_x}&\leq& c \|u\|^{\theta}_{L^\infty_tH^1_x} \|u\|^{2-\theta}_{ L_t^{(2-\theta)a_1} L_x^{\widehat{r}}} \|\nabla u\|_{L^{\widehat{q}}_tL^{\widehat{r}}_x} \nonumber \\ &=& c \|u\|^{\theta}_{L^\infty_tH^1_x} \|u\|^{2-\theta}_{ L_t^{\widehat{a}} L_x^{\widehat{r}}} \|\nabla u\|_{L^{\widehat{q}}_tL^{\widehat{r}}_x}.\end{aligned}$$  To estimate $N_2$ we use the pairs $(\bar{a},\bar{r})=\left(8(1-\theta), \frac{12(1-\theta)}{3-2b-\theta(4-2b)}\right)$ $\dot{H}^{s_c}$-admissible and $(q,r)=\left(\frac{8(1-\theta)}{2-3\theta},\frac{12(1-\theta)}{4-3\theta}\right)$ $L^2$-admissible.[^5] . Let $A\subset \mathbb{R}^N$ such that $A=B$ or $A=B^C$. The Hölder inequality and the Sobolev embedding , with $s=1$ imply $$\begin{aligned} \label{SLG32} \left\| \nabla \left(|x|^{-b}\right)|u|^2 u \right \|_{L^{6'}_x(A)} & \leq& c\left\||x|^{-b-1} \right\|_{L^\gamma(A)} \|u\|^{\theta}_{L^{\theta r_1}_x} \|u\|^{2-\theta}_{L_x^{(2-\theta)r_2}} \| u\|_{L_x^{ r_3}} \nonumber \\ &\leq & c \left\||x|^{-b-1}\right\|_{L^\gamma(A)} \|u\|^{\theta}_{L^{\theta r_1}_x} \| u\|^{2-\theta}_{L_x^{\bar{r}}} \|\nabla u\|_{L_x^r},\end{aligned}$$ where $$\label{SLG3} \frac{1}{6'}=\frac{1}{\gamma}+\frac{1}{r_1}+\frac{1}{r_2}+\frac{1}{r_3};\;\;\;\;1=\frac{3}{r}-\frac{3}{r_3};\;\;\;\bar{r}=(2-\theta)r_2.$$ Note that the second equation in is valid since $r<3$. On the other hand, in order to show that $\||x|^{-b-1}\|_{L^d(A)}$ is bounded, we need $\frac{3}{d}-b-1>0$ when $A$ is the ball $B$ and $\frac{3}{d}-b-1<0$ when $A=B^C$, by Remark \[RIxb\]. Indeed, using and the values of $q$, $r$, $\bar{q}$ and $\bar{r}$ defined above one has $$\begin{aligned} \label{SLG33} \frac{3}{\gamma}-b-1&=& \frac{5}{2}-b-\frac{3}{r_1}-\frac{3(2-\theta)}{\bar{r}}-\frac{3}{r}=\frac{\theta(2-b)}{2}-\frac{3}{r_1}. \end{aligned}$$ Now choosing $r_1$ such that $$\theta r_1>\frac{6}{2-b} \textrm{ when } A=B \quad \textrm{and} \quad \theta r_1<\frac{6}{2-b} \textrm{ when } A=B^C,$$ we get $\frac{3}{d}-b-1>0$ when $A=B$ and $\frac{3}{d}-b-1<0$ when $A=B^C$, so $|x|^{-b-1}\in L^d(A)$. In addition, we have by the Sobolev embedding (since $2<\frac{6}{2-b}<6$) and $$\left\| \nabla \left(|x|^{-b}\right)|u|^2 u \right \|_{L^{6'}_x(A)} \leq c \|u\|^{\theta}_{H^1_x}\|u\|^{2-\theta}_{L^{\bar{r}}_x} \|\nabla u\|_{L_x^r}.$$ Finally, by Hölder’s inequality in the time variable and the fact that $\frac{1}{2^{'}}=\frac{2-\theta}{\bar{a}}+\frac{1}{q}$, we conclude $$\label{SLG35} N_2= \left\| \nabla \left(|x|^{-b}\right)|u|^2 u \right\|_{{L^{2'}_t L^{6'}_x}} \leq c \|u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{L^{\bar{a}}_tL^{\bar{r}}_x} \|\nabla u\|_{L^q_tL_x^r}.$$ The proof is completed combining and . We notice that in Lemma \[LG1\] and Remark \[RGP\] we assume $0<b<1$. On the other hand, in Lemma \[LG3\] the required assumption is $0<b<1/2$ (see footnote $5$). For this reason in our main result, Theorem \[SCATTERING\]), the restriction on $b$ is different than the one in Theorem \[TG\]. \[RSglobal\] A consequence of the previous lemma is the following estimate $$\left\| |x|^{-b-1}|u|^2 v \right\|_{S'(L^2)}\lesssim \| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c})} \|\nabla v\|_{S(L^2)}.$$  Our first result in this section concerning the IVP is the following \[GWPH1\][**(Small data global theory in $H^1$)**]{} Let $0<b<1/2$ and $u_0 \in H^1(\mathbb{R}^3)$. Assume $\|u_0\|_{H^1}\leq A$. There there exists $\delta=\delta(A)>0$ such that if $\|U(t)u_0\|_{S(\dot{H}^{s_c})}<\delta$, then there exists a unique global solution $u$ of such that $$\label{NGWP3} \|u\|_{S(\dot{H}^{s_c})}\leq 2\|U(t)u_0\|_{S(\dot{H}^{s_c})}$$ and $$\label{NGWP4} \|u\|_{S\left(L^2\right)}+\|\nabla u\|_{S\left(L^2\right)}\leq 2c\|u_0\|_{H^1}.$$ To this end, we use the contraction mapping principle. Define $$B=\{ u:\;\|u\|_{S(\dot{H}^{s_c})}\leq 2\|U(t)u_0\|_{S(\dot{H}^{s_c})}\;\textnormal{and}\;\|u\|_{S(L^2)}+\|\nabla u\|_{S(L^2)}\leq 2c\|u_0\|_{H^1}\}.$$ We prove that $G$ defined below $$\label{OPERATOR} G(u)(t)=U(t)u_0+i \int_0^t U(t-t')F(x,u)(t')dt',$$ where $F(x,u)=|x|^{-b}|u|^2 u$ is a contraction on $B$ equipped with the metric $$d(u,v)=\|u-v\|_{S(L^2)}+\|u-v\|_{S(\dot{H}^{s_c})}.$$  Indeed, we deduce by the Strichartz inequalities (\[SE1\]), (\[SE2\]), and $$\label{GHs1} \|G(u)\|_{S(\dot{H}^{s_c})}\leq \|U(t)u_0\|_{S(\dot{H}^{s_c})}+ c\| F \|_{S'(\dot{H}^{-s_c})}$$ $$\label{GHs2} \|G(u)\|_{S(L^2)}\leq c\|u_0\|_{L^2}+ c\| F \|_{S'(L^2)}$$ $$\label{GHs3} \|\nabla G(u)\|_{S(L^2)}\leq c \|\nabla u_0\|_{L^2}+ c\|\nabla F\|_{S'(L^2)}.$$ On the other hand, it follows from Lemmas \[LG1\] and \[LG3\] that $$\begin{aligned} \|F\|_{S'(\dot{H}^{-s_c})}&\leq & c\| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c})}\| u \|_{S(\dot{H}^{s_c})}\\ \|F\|_{S'(L^2)}&\leq& c\| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c})}\| u \|_{S(L^2)}\end{aligned}$$ $$\begin{aligned} \|\nabla F\|_{S'(L^2)}&\leq& c\| u\|^{\theta}_{L^\infty_tH^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c})} \|\nabla u\|_{S(L^2)}.\end{aligned}$$ Combining - and the last inequalities, we get for $u\in B$ $$\begin{aligned} \label{TGHS} \|G(u)\|_{S(\dot{H}^{s_c})}\leq& \|U(t)u_0\|_{S(\dot{H}^{s_c})} +c\| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c})}\| u \|_{S(\dot{H}^{s_c})} \nonumber \\ \leq & \|U(t)u_0\|_{S(\dot{H}^{s_c})}+8c^{\theta+1}\|u_0\|^\theta_{H^1}\| U(t)u_0 \|^{3-\theta}_{S(\dot{H}^{s_c})}.\end{aligned}$$ In addition, setting $X=\|\nabla u\|_{S(L^2)}+\| u\|_{S(L^2)}$ then $$\begin{aligned} \|G(u)\|_{S(L^2)}+\|\nabla G(u)\|_{S(L^2)}&\leq & c\|u_0\|_{H^1}+c\| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c})}X \nonumber\\ &\leq & c\|u_0\|_{H^1}+16c^{\theta+2}\|u_0\|_{H^1}^{\theta+1}\| U(t)u_0 \|^{2-\theta}_{S(\dot{H}^{s_c})},\end{aligned}$$ where we have have used the fact that $X\leq 2^2c\|u_0\|_{H^1}$ since $u\in B$.  Now if $\| U(t)u_0 \|_{S(\dot{H}^{s_c})}<\delta$ with $$\label{WD} \delta\leq \min\left\{\sqrt[2-\theta]{\frac{1}{16c^{\theta+1}A^\theta}} , \sqrt[2-\theta]{ \frac{1}{32c^{\theta+1}A^\theta}}\right\},$$ where $A>0$ is a number such that $\|u_0\|_{H^1}\leq A$, we get $$\|G(u)\|_{S(\dot{H}^{s_c})}\leq 2\|U(t)u_0\|_{S(\dot{H}^{s_c})}$$ and $$\|G(u)\|_{S(L^2)}+\|\nabla G(u)\|_{S(L^2)}\leq 2c\|u_0\|_{H^1},$$ that is $G(u)\in B$. The contraction property can be obtained by similar arguments. Therefore, by the Banach Fixed Point Theorem, $G$ has a unique fixed point $u\in B$, which is a global solution of .  We now show Proposition \[SCATTERSH1\] (this result gives us the criterion to establish scattering). First, we claim that $$\label{SCATTER1} \|u\|_{S(L^2)}+\|\nabla u\|_{S(L^2)}<+\infty.$$  Indeed, since $\|u\|_{S(\dot{H}^{s_c})}<+\infty$, given $\delta>0$ we can decompose $[0,\infty)$ into $n$ many intervals $I_j=[t_j,t_{j+1})$ such that $\|u\|_{S(\dot{H}^{s_c};I_j)}<\delta$ for all $j=1,...,n$. On the time interval $I_j$ we consider the integral equation $$\label{SCATTER2} u(t)=U(t-t_j)u(t_j)+i\int_{t_j}^{t_{j+1}}U(t-s)(|x|^{-b}|u|^2 u)(s)ds.$$ It follows from the Strichartz estimates and that $$\label{SCATTER3} \|u\|_{S(L^2;I_j)}\leq c\|u(t_j)\|_{L^2_x}+c\left\||x|^{-b}|u|^2 u \right\|_{S'(L^2;I_j)}$$ $$\label{SCATTER4} \|\nabla u\|_{S(L^2;I_j)}\leq c\|\nabla u(t_j)\|_{L^2_x}+c\left\|\nabla(|x|^{-b}|u|^2 u) \right\|_{S'(L^2;I_j)}.$$ From Lemmas \[LG1\] (ii) and \[LG3\] we have $$\begin{aligned} \left\||x|^{-b}|u|^2 u \right\|_{S'(L^2;I_j)} &\leq& c \| u \|^\theta_{L^\infty_{I_j}H^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c};I_j)}\|u\|_{S(L^2;I_j)},\end{aligned}$$ $$\|\nabla(|x|^{-b}|u|^2 u)\|_{S'(L^2;I_j)}\leq c\| u\|^{\theta}_{L^\infty_{I_j}H^1_x}\|u\|^{2-\theta}_{S(\dot{H}^{s_c};I_j)} \|\nabla u\|_{S(L^2;I_j)} .$$ Thus, using , and the last two estimates we get $$\|u\|_{S(L^2;I_j)}\leq c B+cB^\theta\delta^{2-\theta}\|u\|_{S(L^2;I_j)}$$ and $$\label{nablau} \|\nabla u\|_{S(L^2;I_j)}\leq c B+ cB^{\theta+1}\delta^{2-\theta}+cB^\theta\delta^{2-\theta}\|\nabla u\|_{S(L^2;I_j)},$$ where we have used the assumption $\sup\limits_{t\in \mathbb{R}}\|u(t)\|_{H^1}\leq B$. Taking $\delta>0$ such that $c B^\theta\delta^{2-\theta}<\frac{1}{2}$ we obtain $\| u\|_{S(L^2;I_j)}+ \|\nabla u\|_{S(L^2;I_j)} \leq cB$, and by summing over the $n$ intervals, we conclude the proof of .  Returning to the proof of the proposition, let $$\phi^+=u_0+i\int\limits_{0}^{+\infty}U(-s)|x|^{-b}(|u|^2 u)(s)ds,$$ Note that, $\phi^+ \in H^1(\mathbb{R}^3)$. Indeed, by the same arguments as ones used before we deduce $$\|\phi^+\|_{L^2}+\|\nabla\phi^+\|_{L^2}\leq c\|u_0\|_{H^1}+c \| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c})}\left( \|u\|_{S(L^2)}+\| \nabla u\|_{S(L^2)} \right).$$ Therefore, yields $\|\phi\|_{H^1}<+\infty$.  On the other hand, since $u$ is a solution of we get $$u(t)-U(t)\phi^+=-i\int\limits_{t}^{+\infty}U(t-s)|x|^{-b}(|u|^2 u)(s)ds.$$ Similarly as before, we have $$\|u(t)-U(t)\phi\|_{H^1_x} \leq c \| u \|^\theta_{L^\infty_tH^1_x}\| u \|^{2-\theta}_{S(\dot{H}^{s_c};[t,\infty))} \left( \|u\|_{S(L^2)}+ \| \nabla u\|_{S(L^2)} \right)$$ The proof is completed after using and $\|u\|_{S(\dot{H}^{s_c};[t,\infty))}\rightarrow 0$ as $t \rightarrow +\infty$. In the same way we define $$\phi^-=u_0+i\int_0^{-\infty}U(-s)|x|^{-b}(|u|^2 u)(s)ds,$$ and using the same argument as before we have $\phi^-\in H^1$ and $$\|u(t)-U(t)\phi^-\|_{H^1_x}\rightarrow 0 \,\,\textnormal{as}\,\,t\rightarrow -\infty.$$  Next, we study the perturbation theory for the IVP following the exposition in Killip-Kwon-Shao-Visan [@KKSV Theorem $3.1$]. We first obtain a short-time perturbation which can be iterated to obtain a long-time perturbation result. \[STP\][**(Short-time perturbation theory for the INLS)**]{} Let $I\subseteq \mathbb{R}$ be a time interval containing zero and let $\widetilde{u}$ defined on $I\times \mathbb{R}^3$ be a solution (in the sense of the appropriated integral equation) to $$\label{PE}i\partial_t \widetilde{u} +\Delta \widetilde{u} + |x|^{-b} |\widetilde{u}|^2 \widetilde{u} =e,$$ with initial data $\widetilde{u}_0\in H^1(\mathbb{R}^3)$, satisfying $$\label{PC11} \sup_{t\in I} \|\widetilde{u}(t)\|_{H^1_x}\leq M \;\; \textnormal{and}\;\; \|\widetilde{u}\|_{S(\dot{H}^{s_c}; I)}\leq \varepsilon,$$ for some positive constant $M$ and some small $\varepsilon>0$. Let $u_0\in H^1(\mathbb{R}^3)$ such that $$\label{PC22} \|u_0-\widetilde{u}_0\|_{H^1}\leq M'\;\; \textnormal{and}\;\; \|U(t)(u_0-\widetilde{u}_0)\|_{S(\dot{H}^{s_c}; I)}\leq \varepsilon,\;\;\textnormal{for }\; M'>0.$$ In addition, assume the following conditions $$\label{PC33} \|e\|_{S'(L^2; I)}+\|\nabla e\|_{S'(L^2; I)}+ \|e\|_{S'(\dot{H}^{-s_c}; I)}\leq \varepsilon.$$ There exists $\varepsilon_0(M,M')>0$ such that if $\varepsilon<\varepsilon_0$, then there is a unique solution $u$ to on $I\times \mathbb{R}^3$ with initial data $u_0$, at the time $t=0$, satisfying $$\label{C} \|u\|_{S(\dot{H}^{s_c}; I)}\lesssim \varepsilon $$ and $$\label{C1} \|u\|_{S(L^2; I)}+\|\nabla u\|_{S(L^2; I)}\lesssim c(M,M').$$ We use the following claim (we will show it later): there exists $\varepsilon_0>0$ sufficiently small such that, if $\|\widetilde{u}\|_{S(\dot{H}^{s_c};I)}\leq \varepsilon_0$ then $$\label{ST1} \|\widetilde{u}\|_{S(L^2;I)} \lesssim M\;\;\;\textnormal{and}\;\;\;\;\|\nabla\widetilde{u}\|_{S(L^2;I)} \lesssim M.$$  We may assume, without loss of generality, that $0=\inf I$. Let us first prove the existence of a solution $w$ for the following initial value problem $$\label{IVPP} \left\{\begin{array}{cl} i\partial_tw +\Delta w + H(x,\widetilde{u},w)+e= 0,& \\ w(0,x)= u_0(x)-\widetilde{u}_0(x),& \end{array}\right.$$ where $H(x,\widetilde{u},w)=|x|^{-b} \left(|\widetilde{u}+w|^2 (\widetilde{u}+w)-|\widetilde{u}|^2 \widetilde{u}\right)$.  To this end, let $$\label{IEP} G (w)(t):=U(t)w_0+i \int_0^t U(t-s)(H(x,\widetilde{u},w)+e)(s)ds$$ and define $$B_{\rho,K}=\{ w\in C(I;H^1(\mathbb{R}^3)):\;\|w\|_{S(\dot{H}^{s_c};I)}\leq \rho\;\textnormal{and}\;\|w\|_{S(L^2;I)}+\|\nabla w\|_{S(L^2;I)}\leq K \}.$$ For a suitable choice of the parameters $\rho>0$ and $K>0$, we need to show that $G$ in defines a contraction on $B_{\rho,K}$. Indeed, applying Strichartz inequalities (\[SE1\]), (\[SE2\]), (\[SE3\]) and we have $$\label{SP1}\|G(w)\|_{S(\dot{H}^{s_c};I)}\lesssim \|U(t)w_0\|_{S(\dot{H}^{s_c};I)}+ \| H(\cdot,\widetilde{u},w) \|_{S'(\dot{H}^{-s_c};I)}+\|e \|_{S'(\dot{H}^{-s_c};I)}$$ $$\label{SP2} \|G(w)\|_{S(L^2;I)}\lesssim \|w_0\|_{L^2}+ \| H(\cdot,\widetilde{u},w) \|_{S'(L^2;I)}+\|e\|_{S'(L^2;I)}$$ $$\label{SP3} \|\nabla G(w)\|_{S(L^2;I)}\lesssim \|\nabla w_0\|_{L^2}+ \| \nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)}+\|\nabla e\|_{S'(L^2;I)}.$$ On the other hand, since $$\label{EI} \left| |\widetilde{u}+w|^2(\widetilde{u}+w)-|\widetilde{u}|^2\widetilde{u}\right|\lesssim |\widetilde{u}|^2|w|+|w|^{3}$$ by , we get $$\|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c};I)}\leq \||x|^{-b}|\widetilde{u}|^2 w\|_{S'(\dot{H}^{-s_c};I)}+\||x|^{-b}|w|^2 w\|_{S'(\dot{H}^{-s_c};I)},$$ which implies using Lemma \[LG1\] (i) that $$\begin{aligned} \label{SP4} \|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c};I)}\lesssim \left(\| \widetilde{u} \|^\theta_{L^\infty_tH^1_x}\| \widetilde{u} \|^{2-\theta}_{S(\dot{H}^{s_c};I)}+ \| w \|^\theta_{L^\infty_tH^1_x} \| w\|^{2-\theta}_{S(\dot{H}^{s_c};I)} \right) \| w \|_{S(\dot{H}^{s_c};I)}.\end{aligned}$$ The same argument and Lemma \[LG1\] (ii) also yield $$\begin{aligned} \label{SP5} \|H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)}\lesssim \left(\| \widetilde{u} \|^\theta_{L^\infty_tH^1_x}\| \widetilde{u} \|^{2-\theta}_{S(\dot{H}^{s_c};I)} + \| w \|^\theta_{L^\infty_tH^1_x}\| w\|^{2-\theta}_{S(\dot{H}^{s_c};I)} \right)\| w \|_{S(L^2;I)}.\end{aligned}$$ Now, we estimate $\|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)}$. It follows from and that $$ |\nabla H(x,\widetilde{u},w)| \lesssim |x|^{-b-1}(|\widetilde{u}|^{2}+|w|^{2})|w|+|x|^{-b}(|\widetilde{u}|^2+|w|^2) |\nabla w| +E,$$ where $E \lesssim |x|^{-b}\left(|\widetilde{u}|+|w|\right)|w||\nabla \widetilde{u}|. $ Thus, Lemma \[LG1\] (ii), Remark \[RSglobal\] and Remark \[RGP\] lead to $$\|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)} \lesssim \left(\| \widetilde{u} \|^\theta_{L^\infty_tH^1_x}\| \widetilde{u} \|^{2-\theta}_{S(\dot{H}^{s_c};I)} + \| w \|^\theta_{L^\infty_tH^1_x}\| w\|^{2-\theta}_{S(\dot{H}^{s_c};I)} \right)\|\nabla w \|_{S(L^2;I)}$$ $$\begin{aligned} \label{SP6} \hspace{1.5cm}+\left(\| \widetilde{u} \|^\theta_{L^\infty_tH^1_x} \| \widetilde{u} \|^{1-\theta}_{S(\dot{H}^{s_c};I)} + \| w \|^\theta_{L^\infty_tH^1_x} \| w \|^{1-\theta}_{S(\dot{H}^{s_c};I)} \right) \| w \|_{S(\dot{H}^{s_c};I)} \|\nabla \widetilde{u} \|_{S(L^2;I)}\end{aligned}$$  Hence, combining , and if $u\in B(\rho,K)$, we have $$\begin{aligned} \label{SP7} \|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c};I)} \lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)\rho\end{aligned}$$ $$\begin{aligned} \label{SP8} \|H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)}\lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)K.\end{aligned}$$ Furthermore, and imply $$\begin{aligned} \label{SP9} \|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2;I)} \lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)K +\left( M^\theta \varepsilon^{1-\theta} + K^\theta \rho^{1-\theta} \right) \rho M. $$ Therefore, we deduce by - together with - that $$\|G(w)\|_{S(\dot{H}^{s_c};I)}\leq c\varepsilon+ cA\rho$$ $$\|G(w)\|_{S(L^2;I)}\leq cM'+c\varepsilon +cAK,$$ where we also used the hypothesis - and $A=M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}$. We also have, using , $$\|\nabla G(w)\|_{S(L^2;I)}\leq cM'+c\varepsilon +cAK+cB \rho M,$$ where $B= M^\theta \varepsilon^{1-\theta} + K^\theta \rho^{1-\theta}$.\ Choosing $\rho=2c\varepsilon$, $K=3cM'$ and $\varepsilon_0$ sufficiently small such that $$cA<\frac{1}{3}\;\;\;\;\textnormal{and}\;\;\;c(\varepsilon+B \rho M+K^\theta \rho^{2-\theta} M)<\frac{K}{3},$$ we obtain $$\|G(w)\|_{S(\dot{H}^{s_c};I)}\leq \rho\;\;\;\textnormal{and}\;\;\;\|G(w)\|_{S(L^2;I)}+\|\nabla G(w)\|_{S(L^2;I)}\leq K.$$ The above calculations establish that $G$ is well defined on $B(\rho,K)$. The contraction property can be obtained by similar arguments. Hence, by the Banach Fixed Point Theorem we obtain a unique solution $w$ on $I\times \mathbb{R}^N$ such that $$\|w\|_{S(\dot{H}^{s_c};I)}\lesssim \varepsilon \;\;\;\textnormal{and}\;\;\;\|w\|_{S(L^2;I)}+\|w\|_{S(L^2;I)} \lesssim M'.$$ Finally, it is easy to see that $u=\widetilde{u}+w$ is a solution to satisfying and .  To complete the proof we now show . Indeed, we first show that $$\label{widetilde{u}} \|\nabla\widetilde{u}\|_{S(L^2;I)}\lesssim M.$$ Using the same arguments as before, we have $$\|\nabla\widetilde{u}\|_{S(L^2;I)}\lesssim \|\nabla\widetilde{u}_0\|_{L^2}+ \left\|\nabla(|x|^{-b}|\widetilde{u}|^2\widetilde{u})\right\|_{S'(L^2;I)}+\|\nabla e\|_{S'(L^2;I)}.$$ Lemma \[LG3\] implies $$\begin{aligned} \|\nabla\widetilde{u}\|_{S(L^2;I)}&\lesssim& M+ \| \widetilde{u} \|^\theta_{L^\infty_tH^1_x}\| \widetilde{u} \|^{2-\theta}_{S(\dot{H}^{s_c};I)} \|\nabla \widetilde{u} \|_{S(L^2;I)} +\varepsilon\\ &\lesssim& M+\varepsilon+ M^\theta \varepsilon_0^{2-\theta}\|\nabla \widetilde{u} \|_{S(L^2;I)}.\end{aligned}$$ Therefore, choosing $\varepsilon_0$ sufficiently small the linear term $M^\theta \varepsilon_0^{2-\theta}\|\nabla \widetilde{u} \|_{S(L^2;I)}$ may be absorbed by the left-hand term and we conclude the proof of . Similar estimates also imply $\|\widetilde{u}\|_{S(L^2;I)}\lesssim M$. \[RSP\] From Proposition \[STP\], we also have the following estimates: $$\label{RSP1} \|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c}; I)}\leq C(M,M') \varepsilon$$ and $$\label{RSP2} \|H(\cdot,\widetilde{u},w)\|_{S'(L^2; I)}+\|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2; I)}\leq C(M,M')\varepsilon^{2-\theta},$$ with $\theta>0$ small enough. Indeed, the relations , and imply $$\begin{aligned} \|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c}; I)} \lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)\rho,\end{aligned}$$ $$\begin{aligned} \|H(\cdot,\widetilde{u},w)\|_{S'(L^2; I)}\lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)K \end{aligned}$$ and $$\begin{aligned} \|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2; I)} \lesssim \left(M^\theta\varepsilon^{2-\theta}+K^\theta \rho^{2-\theta}\right)K+\left( M^\theta \varepsilon^{1-\theta} + K^\theta \rho^{1-\theta} \right) \rho M.\end{aligned}$$ Therefore, the choice $\rho=2c\varepsilon$ and $K=3cM'$ in Proposition \[STP\] yield and .  In the sequel, we prove the long-time perturbation result. \[LTP\][**(Long-time perturbation theory for the INLS)**]{} Let $I\subseteq \mathbb{R}$ be a time interval containing zero and let $\widetilde{u}$ defined on $I\times \mathbb{R}^3$ be a solution (in the sense of the appropriated integral equation) to $$i\partial_t \widetilde{u} +\Delta \widetilde{u} + |x|^{-b} |\widetilde{u}|^2 \widetilde{u} =e,$$ with initial data $\widetilde{u}_0\in H^1(\mathbb{R}^3)$, satisfying $$\label{HLP1} \sup_{t\in I} \|\widetilde{u}\|_{H^1_x}\leq M \;\; \textnormal{and}\;\; \|\widetilde{u}\|_{S(\dot{H}^{s_c}; I)}\leq L,$$ for some positive constants $M,L$. Let $u_0\in H^1(\mathbb{R}^3)$ such that $$\label{HLP2} \|u_0-\widetilde{u}_0\|_{H^1}\leq M'\;\; \textnormal{and}\;\; \|U(t)(u_0-\widetilde{u}_0)\|_{S(\dot{H}^{s_c}; I)}\leq \varepsilon,$$ for some positive constant $M'$ and some $0<\varepsilon<\varepsilon_1=\varepsilon_1(M,M',L)$. Moreover, assume also the following conditions $$\|e\|_{S'(L^2; I)}+\|\nabla e\|_{S'(L^2; I)}+ \|e\|_{S'(\dot{H}^{-s_c}; I)}\leq \varepsilon.$$ Then, there exists a unique solution $u$ to on $I\times \mathbb{R}^3$ with initial data $u_0$ at the time $t=0$ satisfying $$\label{CLP} \|u-\widetilde{u}\|_{S(\dot{H}^{s_c}; I)}\leq C(M,M',L)\varepsilon\;\;\;\;\;\;\;\textnormal{and}$$ $$\label{CLP1} \|u\|_{S(\dot{H}^{s_c}; I)} +\|u\|_{S(L^2; I)}+\|\nabla u\|_{S(L^2; I)}\leq C(M,M',L).$$ First observe that since $\|\widetilde{u}\|_{S(\dot{H}^{s_c}; I)}\leq L$, given[^6] $\varepsilon<\varepsilon_0(M,2M')$ we can partition $I$ into $n = n(L,\varepsilon)$ intervals $I_j = [t_j ,t_{j+1})$ such that for each $j$, the quantity $\|\widetilde{u}\|_{S(\dot{H}^{s_c};I_j)}\leq \varepsilon$. Note that $M'$ is being replaced by $2M'$, as the $H^1$-norm of the difference of two different initial data may increase in each iteration.  Again, we may assume, without loss of generality, that $0=\inf I$. Let $w$ be defined by $u = \widetilde{u} + w$, then $w$ solves IVP with initial time $t_j$. Thus, the integral equation in the interval $I_j = [t_j ,t_{j+1})$ reads as follows $$w(t)=U(t-t_j)w(t_j)+i\int_{t_j}^{t}U(t-s)(H(x,\widetilde{u},w)+e)(s)ds,$$ where $H(x,\widetilde{u},w)=|x|^{-b} \left(|\widetilde{u}+w|^2 (\widetilde{u}+w)-|\widetilde{u}|^2 \widetilde{u}\right)$.  Thus, choosing $\varepsilon_1$ sufficiently small (depending on $n$, $M$, and $M'$), we may apply Proposition \[STP\] (Short-time perturbation theory) to obtain for each $0\leq j<n$ and all $\varepsilon<\varepsilon_1$, $$\label{LP1} \|u-\widetilde{u}\|_{S(\dot{H}^{s_c};I_j)}\leq C(M,M',j)\varepsilon$$ and $$\label{LP2} \|w\|_{S(\dot{H}^{s_c};I_j)}+\|w\|_{S'(L^2;I_j)}+\|\nabla w\|_{S'(L^2;I_j)}\leq C(M,M',j)$$ provided we can show $$\label{LP3} \|U(t-t_j)(u(t_j)-\widetilde{u}(t_j))\|_{S(\dot{H}^{s_c};I_j)}\leq C(M,M',j)\varepsilon\leq \varepsilon_0$$ and $$\label{LP4} \|u(t_j)-\widetilde{u}(t_j)\|_{H^1_x}\leq 2M',$$ For each $0\leq j<n$.  Indeed, by the Strichartz estimates and , we have $$\begin{aligned} \|U(t-t_j)w(t_j)\|_{S(\dot{H}^{s_c};I_j)}&\lesssim& \|U(t)w_0\|_{S(\dot{H}^{s_c}; I)}+\|H(\cdot,\widetilde{u},w)\|_{S'(\dot{H}^{-s_c};[0,t_j])}\\ &&+\|e\|_{S'(\dot{H}^{-s_c};I)}, \end{aligned}$$ which implies by that $$\|U(t-t_j)(u(t_j)-\widetilde{u}(t_j))\|_{S(\dot{H}^{s_c}; I_j)}\lesssim \varepsilon+\sum_{k=0}^{j-1}C(k,M,M')\varepsilon.$$  Similarly, it follows from Strichartz estimates , and that $$\begin{aligned} \|u(t_j)-\widetilde{u}(t_j)\|_{H^1_x}&\lesssim & \|u_0-\widetilde{u}_0\|_{H^1}+\|e\|_{S'(L^2; I)}+\|\nabla e\|_{S'(L^2;I)}\\ &&+\| H(\cdot,\widetilde{u},w)\|_{S'(L^2;[0,t_j])}+\|\nabla H(\cdot,\widetilde{u},w)\|_{S'(L^2;[0,t_j])}\\ &\lesssim& M'+\varepsilon+\sum_{k=0}^{j-1}C(k,M,M')\varepsilon^{2-\theta}. \end{aligned}$$ Taking $\varepsilon_1=\varepsilon(n,M,M')$ sufficiently small, we see that and hold and so, it implies and .  Finally, summing this over all subintervals $I_j$ we obtain $$\|u-\widetilde{u}\|_{S(\dot{H}^{s_c}; I)}\leq C(M,M',L)\varepsilon$$ and $$\|w\|_{S(\dot{H}^{s_c}; I)}+\|w\|_{S'(L^2; I)}+\|\nabla w\|_{S'(L^2; I)}\leq C(M,M',L).$$ This completes the proof. Properties of the ground state, energy bounds and wave operator ===============================================================  In this section, we recall some properties that are related to our problem. In [@LG] the first author proved the following Gagliardo-Nirenberg inequality $$\label{GNI} \left\||x|^{-b}|u|^{4} \right\|_{L^1_x}\leq C_{GN}\|\nabla u\|^{3+b}_{L^2_x}\|u\|^{1-b}_{L^2_x},$$ with the sharp constant (recalling $s_c=\frac{1+b}{2}$) $$\label{GNI1}C_{GN}=\frac{4}{3+b}\left(\frac{1-b}{3+b}\right)^{s_c}\frac{1}{\|Q\|^{2}_{L^2}}$$ where $Q$ is the ground state solution of . Moreover, $Q$ satisfies the following relations $$\label{GS1} \|\nabla Q\|^2_{L^2}=\frac{3+b}{1-b}\|Q\|^2_{L^2}$$ and $$\label{GS2} \left\||x|^{-b}|Q|^{4} \right\|_{L^1}=\frac{4}{3+b}\|\nabla Q\|^2_{L^2}.$$ Note that, combining , and one has $$\label{GNI2}C_{GN}=\frac{4}{(3+b)\|\nabla Q\|^{2 s_c}_{L^2}\|Q\|^{2(1-s_c)}_{L^2}}.$$ On the other hand, we also have $$\label{EGS} E[Q]=\frac{1}{2}\|\nabla Q\|^2_{L^2}-\frac{1}{4}\left\||x|^{-b}|Q|^{4}\right\|_{L^1}=\frac{s_c}{3+b}\|\nabla Q\|^2_{L^2}.$$  The next lemma provides some estimates that will be needed for the compactness and rigidity results. \[LGS\] Let $v \in H^1(\mathbb{R}^3)$ such that $$\label{LGS1} \|\nabla v\|^{s_c}_{L^2}\|v\|_{L^2}^{1-s_c}\leq \|\nabla Q\|^{s_c}_{L^2}\|Q\|_{L^2}^{1-s_c}.$$ Then, the following statements hold - $\frac{s_c}{3 +b}\|\nabla v\|^2_{L^2}\leq E(v)\leq \frac{1}{2} \|\nabla v\|^{2}_{L^2}$, - $\|\nabla v\|^{s_c}_{L^2}\|v\|^{1-s_c}_{L^2}\leq w^{\frac{1}{2}} \|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2}$, - $16A E[v]\leq 8A \|\nabla v\|_{L^2}^2\leq 8 \|\nabla v\|^2_{L^2}-2(3+b)\left\||x|^{-b}|v|^{4}\right\|_{L^1}$, where $w=\frac{E[v]^{s_c}M[v]^{1-s_c}}{E[Q]^{s_c}M[Q]^{1-s_c}}$ and $A=(1-w)$. \(i) The second inequality is immediate from the definition of Energy . The first one is obtained by observing that $$\begin{aligned} E[v] &\geq& \frac{1}{2}\|\nabla v\|^2_{L^2} - \frac{C_{GN}}{4}\|\nabla v\|^{3+b}_{L^2}\|v\|^{1-b}_{L^2}\\ &=&\frac{1}{2}\|\nabla v\|^2_{L^2}\left(1- \frac{C_{GN}}{2} \|\nabla v\|^{2 s_c}_{L^2}\|v\|^{2(1-s_c)}_{L^2} \right)\\ &\geq& \frac{1}{2}\|\nabla v\|^2_{L^2}\left(1- \frac{C_{GN}}{2} \|\nabla Q\|^{2 s_c}_{L^2}\|Q\|^{2(1-s_c)}_{L^2} \right)\\ &=&\frac{1+b}{2(3+b)}\|\nabla v\|^2_{L^2}=\frac{ s_c}{3+b}\|\nabla v\|^2_{L^2},\end{aligned}$$ where we have used , and . \(ii) The first inequality in (i) yields $\|\nabla v\|^2_{L^2}\leq \frac{3 +b}{ s_c}E(v)$, multiplying it by $M[v]^\sigma=\|v\|_{L^2}^{2\sigma}$, where $\sigma=\frac{1-s_c}{s_c}$, we have $$\begin{aligned} \|\nabla v\|^2_{L^2}\| v\|^{2\sigma}_{L^2}&\leq& \frac{3 +b}{s_c}E[v]M[v]^\sigma\\ & = & \frac{3 +b}{ s_c}\frac{E[v]M[v]^\sigma}{E[Q]M[Q]^\sigma} E[Q]M[Q]^\sigma\\ &=& w \|\nabla Q\|^2\|Q\|^{2\sigma}_{L^2},\end{aligned}$$ where we have used . \(iii) The first inequality obviously holds. Next, let $B=8 \|\nabla v\|^2_{L^2}-2(3+b)\left\||x|^{-b}|v|^{4}\right\|_{L^1}$. Applying the Gagliardo-Niremberg inequality and item (ii) we deduce $$\begin{aligned} B&\geq& 8\|\nabla v\|^2_{L^2}- 2(3+b)C_{GN}\|\nabla v\|^{3+b}_{L^2}\|v\|^{1-b}_{L^2} \\ &\geq& \|\nabla v\|^2_{L^2}\left(8- 2(3+b)C_{GN}w \|\nabla Q\|^{2 s_c}_{L^2}\|Q\|^{2(1-s_c)}_{L^2} \right)\\ &=& \|\nabla v\|^2_{L^2}8(1-w),\end{aligned}$$ where in the last equality, we have used .  Now, applying the ideas introduced by Côte [@COTE] for the KdV equation (see also Guevara [@GUEVARA] Proposition $2.18$, with $(N,\alpha)=(3,2)$), we show the existence of the Wave Operator. Before stating our result, we prove the following lemma. \[LEWO\]Let $0<b<1$. If $f$ and $g\in H^1(\mathbb{R}^3)$ then - $\left\| |x|^{-b} |f|^{3}g \right\|_{L^1} \leq c \|f\|^{3}_{L^{4}} \|g\|_{L^{4}} + c\|f\|^{3}_{L^{r}}\|g\|_{L^{r}}$ - $\left\| |x|^{-b} |f|^{3}g \right\|_{L^1} \leq c \|f\|^{3}_{{H^1}} \|g\|_{H^1}$ - $ \lim\limits_{|t|\rightarrow +\infty} \left\| |x|^{-b} |U(t) f|^{3}g \right\|_{L^1_x}=0.$ where $\frac{12}{3-b}<r<6$. \(i) We divide the estimate in $B^C$ and $B$. Applying the Hölder inequality, since $1=\frac{3}{4}+\frac{1}{4}$, one has $$\begin{aligned} \label{LEWO1} \left\| |x|^{-b} |f|^{3}g \right\|_{L^1} &\leq &\left\| |x|^{-b} |f|^{3}g \right\|_{L^1(B^C)} +\left\| |x|^{-b} |f|^{3}g \right\|_{L^1(B)} \nonumber\\ &\leq& \|f\|^{3}_{L^{4}}\|g\|_{L^{4}}+\|x|^{-b}|\|_{L^\gamma(B)}\|f\|^{3}_{L^{3\beta}}\|g\|_{L^r} \nonumber\\ &=& \|f\|^{3}_{L^{4}}\|g\|_{L^{4}}+\|x|^{-b}|\|_{L^\gamma(B)} \|f\|^{3}_{L^{r}} \|g\|_{L^r},\end{aligned}$$ where $$\label{LEWO2} 1=\frac{1}{\gamma}+\frac{1}{\beta}+\frac{1}{r}\;\;\;\;\;\textnormal{and}\;\;\;\;\;\;r=3\beta.$$ To complete the proof we need to check that $\||x|^{-b}\|_{L^\gamma(B)}$ is bounded, i.e., $\frac{3}{\gamma}>b$ (see Remark \[RIxb\]). In fact, we deduce from $$\frac{3}{\gamma}=3-\frac{12}{r},$$ and thus, since $r>\frac{12}{3-b}$ we obtain the desired result ($\frac{3}{\gamma}-b>0$).  (ii) By the Sobolev inequality , it is easy to see that $H^1 \hookrightarrow L^{4} $ and $H^1 \hookrightarrow L^{r}$ (where $2<\frac{12}{3-b}<r<6$), then using we get (ii).  (iii) Similarly as (i) and (ii), we get $$\begin{aligned} \label{lematecnico} \left\| |x|^{-b} |U(t) f|^{3}g \right\|_{L^1_x}\leq c\|U(t)f\|^{\alpha+1}_{L^{4}} \|g\|_{H^1}+ c\|U(t)f\|^{3}_{L^{r}}\|g\|_{H^1},\end{aligned}$$ for $\frac{12}{3-b}<r<6$. We now show that $\|U(t)f\|_{L^{r}_x}$ and $\|U(t)f\|_{L^{4}_x}$ $\rightarrow 0$ as $|t|\rightarrow +\infty$. Indeed, since $r$ and $4$ belong to $(2,6)$ then it suffices to show $$\label{LEWO3} \lim\limits_{|t|\rightarrow +\infty}\|U(t)f\|_{L^p_x}=0,$$ where $2<p<6$. Let $\widetilde{f}\in H^1\cap L^{p'}$, the Sobolev embedding and Lemma \[ILE\] yield $$\|U(t)f\|_{L^{p}_x}\leq c \|f-\widetilde{f}\|_{{H}^1}+c|t|^{-\frac{3(p-2)}{2p}}\|\widetilde{f}\|_{L^{p'}}.$$ Since $p>2$ then the exponent of $|t|$ is negative and so approximating $f$ by $\widetilde{f}\in C^\infty_0$ in $H^1$, we deduce . [**(Existence of Wave Operator)**]{}\[PEWO\] Suppose $\phi \in H^1(\mathbb{R}^3)$ and, for some[^7] $0<\lambda \leq (\frac{2 s_c}{3+b})^{\frac{s_c}{2}}$, $$\label{HEWO} \|\nabla \phi\|^{2s_c}_{L^2}\|\phi\|^{2(1-s_c)}_{L^2}<\lambda^2\left(\frac{3+b}{ s_c}\right)^{s_c}E[Q]^{s_c}M[Q]^{1-s_c}.$$ Then, there exists $u^+_0\in H^1(\mathbb{R}^3)$ such that $u$ solving with initial data $u^+_0$ is global in $H^1(\mathbb{R}^3)$ with - $M[u]=M[\phi]$, - $E[u]=\frac{1}{2}\|\nabla \phi\|^2_{L^2}$, - $\lim\limits_{t\rightarrow +\infty} \|u(t)-U(t)\phi\|_{H^1_x}=0$, - $\|\nabla u(t)\|^{s_c}_{L^2}\| u(t)\|^{1-s_c}_{L^2}\leq \lambda\|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2}$. We will divide the proof in two parts. First, we construct the wave operator for large time. Indeed, let $I_T=[T,+\infty)$ for $T\gg 1$ and define $$\label{IEWO1} G(w)(t)=- i \int_t^{+\infty} U(t-s)(|x|^{-b}|w+U(t)\phi|^2 (w+U(t)\phi)(s)ds,\;\;t\in I_T$$ and $$B(T,\rho)=\{w\in C\left(I_T;H^1(\mathbb{R}^3)\right): \|w\|_{T}\leq \rho \},$$ where $$\|w\|_T=\|w\|_{S(\dot{H}^{s_c};I_T)}+\|w\|_{S(L^2;I_T)}+\|\nabla w\|_{S(L^2;I_T)}.$$ Our goal is to find a fixed point for $G$ on $B(T,\rho)$.  Applying the Strichartz estimates and Lemmas \[LG1\]-\[LG3\], we deduce $$\begin{aligned} \label{EWO1} \| G(w) \|_{S(\dot{H}^{s_c};I_T)} \lesssim & \| w+U(t)\phi\|^\theta_{L^\infty_{T}H^1_x}\| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}\| w+U(t)\phi \|_{S(\dot{H}^{s_c};I_T)}\end{aligned}$$ $$\begin{aligned} \label{EWO2} \| G(w) \|_{S(L^2;I_T)} \lesssim & \| w+U(t)\phi\|^\theta_{L^\infty_{T}H^1_x}\| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}\| w+U(t)\phi \|_{S(L^2;I_T)}\end{aligned}$$ and $$\begin{aligned} \label{EWO3} \|\nabla G(w) \|_{S(L^2;I_T)} \lesssim & \| w+U(t)\phi\|^\theta_{L^\infty_{T}H^1_x}\| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}\| \nabla(w+U(t)\phi) \|_{S(L^2;I_T)} \end{aligned}$$ Thus, $$\begin{aligned} \label{EWO4} \|G(w)\|_{T} &\lesssim & \| w+U(t)\phi\|^\theta_{L^\infty_{T}H^1_x}\| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}\| w+U(t)\phi \|_{T}. \nonumber \\\end{aligned}$$  Since[^8] $$\label{U(t)phi} \| U(t)\phi \|_{S(\dot{H}^{s_c};I_T)}\rightarrow 0$$ as $T\rightarrow +\infty$, we can find $T_0>0$ large enough and $\rho>0$ small enough such that $G$ is well defined on $B(T_0,\rho)$. The same computations show that $G$ is a contraction on $B(T_0,\rho)$. Therefore, $G$ has a unique fixed point, which we denote by $w$.  On the other hand, from and since $$\| w+U(t)\phi\|_{L^\infty_{T}H^1_x}\leq \|w\|_{H^1} +\|\phi\|_{H^1}<+\infty,$$ one has (recalling $G(w)=w$) $$\begin{aligned} \| w \|_{S(\dot{H}^{s_c};I_T)} &\lesssim & \| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}\| w+U(t)\phi \|_{S(\dot{H}^{s_c};I_T)}\\ &\lesssim & A\| w\|_{S(\dot{H}^{s_c};I_T)} +A\|U(t)\phi \|_{S(\dot{H}^{s_c};I_T)}\end{aligned}$$ where $A=\| w+U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}$. In addition, if $\rho$ has been chosen small enough and since $\|U(t)\phi\|_{S(\dot{H}^{s_c};I_T)}$ is also sufficiently small for $T$ large, we deduce $$A\leq c\| w\|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}+c\| U(t)\phi \|^{2-\theta}_{S(\dot{H}^{s_c};I_T)}<\frac{1}{2},$$ and so (using the last two inequalities) $$\frac{1}{2} \| w \|_{S(\dot{H}^{s_c};I_T)} \lesssim A \|U(t)\phi \|_{S(\dot{H}^{s_c};I_T)},$$ which implies, $$\label{EWO5} \| w \|_{S(\dot{H}^{s_c};I_T)}\rightarrow 0\;\;\;\;\textnormal{as}\;\;\;\; T\rightarrow +\infty.$$ Hence, , and $\eqref{EWO5}$ also yield that[^9] $$\| w \|_{S(L^{2};I_T)}\;,\,\|\nabla w \|_{S(L^{2};I_T)}\rightarrow 0\;\;\;\;\textnormal{as}\;\;\;\; T\rightarrow +\infty,$$ and finally $$\label{EWO6} \|w\|_{T}\rightarrow 0 \;\; \textnormal{as}\;\; T\rightarrow +\infty.$$  Next, we claim that $u(t)=U(t)\phi+w(t)$ satisfies in the time interval $[T_0,\infty)$. To do this, we need to show that $$\label{CWO} u(t)=U(t-T_0)u(T_0)+i\int_{T_0}^{t}U(t-s)(|x|^{-b}|u|^2 u)(s)ds,$$ for all $t\in [T_0,\infty)$. Indeed, since $$w(t)=- i \int_t^\infty U(t-s)|x|^{-b}|w+U(t)\phi|^2 (w+U(t)\phi)(s)ds,$$ then $$\begin{aligned} U(T_0-t)w(t)&=&- i \int_t^\infty U(T_0-s)|x|^{-b}|w+U(t)\phi|^2 (w+U(t)\phi)(s)ds\\ &=& i\int_{T_0}^t U(T_0-s)|x|^{-b}|w+U(t)\phi|^2 (w+U(t)\phi)(s)ds+w(T_0), \end{aligned}$$ and so applying $U(t-T_0)$ on both sides, we get $$w(t)=U(t-T_0)w(T_0)+ i \int_{T_0}^t U(t-s)|x|^{-b}|w+U(t)\phi|^2 (w+U(t)\phi)(s)ds.$$ Finally, adding $U(t)\phi$ in both sides of the last equation, we deduce .  Now we show relations (i)-(iv). Since $u(t)=U(t)\phi+w$ then $$\begin{aligned} \label{EWO7} \|u(t)-U(t)\phi\|_{L^\infty_TH^1_x}=\|w\|_{L^\infty_TH^1_x}\leq c\|w\|_{S(L^2;I_T)}+c\|\nabla w\|_{S(L^2;I_T)}\leq c\|w\|_{T} \end{aligned}$$ and so from $\eqref{EWO2}$ we obtain (iii). Furthermore, using it is clear that $$\label{EWO81} \lim_{t\rightarrow\infty}\|u(t)\|_{L^2_x}=\| \phi\|_{L^2}.$$ and $$\label{EWO8} \lim_{t\rightarrow\infty}\|\nabla u(t)\|_{L^2_x}=\|\nabla \phi\|_{L^2}.$$ By the mass conservation we have $\|u(t)\|_{L^2}=\|u(T_0)\|_{L^2}$ for all $t$, so from we deduce $\|u(T_0)\|_{L^2}=\|\phi\|_{L^2}$, i.e., item (i) holds. On the other hand, it follows from Lemma \[LEWO\] (ii) $$\begin{aligned} \left\| |x|^{-b}|u(t)|^{4} \right\|_{L^1_x}&\leq& c\left\| |x|^{-b}|u(t)-U(t)\phi|^{4} \right\|_{L^1_x}+c\left\| |x|^{-b}|U(t)\phi|^{4} \right\|_{L^1_x}\\ &\leq & c\left\|u(t)-U(t)\phi| \right\|^{4}_{H^1_x}+c\left\| |x|^{-b}|U(t)\phi|^{4} \right\|_{L^1_x},\end{aligned}$$ which goes to zero as $t\rightarrow +\infty$, by item (iii) and Lemma \[LEWO\] (iii), i.e. $$\label{EWO9} \lim_{t\rightarrow\infty}\left\| |x|^{-b}|u(t)|^{4} \right\|_{L^1_x}=0.$$ Combining and , it is easy to deduce (ii).  Next, in view of , (i) and (ii) we have $$E[u]^{s_c}M[u]^{1-s_c}=\frac{1}{2^{s_c}}\|\nabla \phi\|^{2s_c}_{L^2}\|\phi\|^{2(1-s_c)}_{L^2}<\lambda^2\left(\frac{3+b}{2 s_c}\right)^{s_c}E[Q]^{s_c}M[Q]^{1-s_c}$$ and by our choice of $\lambda$ we conclude $$\label{EWO10} E[u]^{s_c}M[u]^{1-s_c}<E[Q]^{s_c}M[Q]^{1-s_c}.$$ Moreover, from , and $$\begin{aligned} \lim_{t\rightarrow \infty}\|\nabla u(t)\|^{2s_c}_{L^2_x}\|u(t)\|^{2(1-s_c)}_{L^2_x}&=&\|\nabla \phi\|^{2s_c}_{L^2}\|\phi\|^{2(1-s_c)}_{L^2}\\ &<& \lambda^2\left(\frac{3+b}{ s_c}\right)^{s_c}E[Q]^{s_c}M[Q]^{1-s_c}\\ &=&\lambda^2\|\nabla Q\|^{2s_c}_{L^2}\|Q\|^{2(1-s_c)}_{L^2},\end{aligned}$$ where we have used . Thus, one can take $T_1>0$ sufficiently large such that $$\label{EWO11} \|\nabla u(T_1)\|^{s_c}_{L^2_x}\|u(T_1)\|^{1-s_c}_{L^2_x}<\lambda \|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2}.$$ Therefore, since $\lambda<1$, we deduce that relations and hold with $u_0=u(T_1)$ and so, by Theorem \[TG\], we have in fact that $u(t)$ constructed above is a global solution of . \[backward\] A similar Wave Operator construction also holds when the time limit is taken as $t\rightarrow -\infty$ (backward in time). Existence and compactness of a critical solution ================================================  The goal of this section is to construct a critical solution (denoted by $u_c$) of . We divide the study in two parts, first we establish a profile decomposition result and also an Energy Pythagorean expansion for such decomposition. In the sequel, using the results of the first part we construct $u_c$ and discuss some of its properties.  We start this section recalling some elementary inequalities (see Gérard [@Ge98] inequality (1.10) and Guevara [@GUEVARA] page 217). Let $(z_j)\subset\mathbb{C}^M$ with $M\geq 2$. For all $q>1$ there exists $C_{q,M}>0$ such that $$\label{FI} \left|\;\left | \sum_{j=1}^M z_j \right|^q-\sum_{j=1}^M|z_j|^q \right| \leq C_{q,M}\sum_{j\neq k}^{M} |z_j| |z_k|^{q-1},$$ and for $\beta>0$ there exists a constant $C_{\beta,M}>0$ such that $$\label{EIerror} \left| \left|\sum_{j=1}^{M}z_j\right|^\beta\sum_{j=1}^{M}z_j-\sum_{j=1}^{M} |z_j|^\beta z_j\right|\leq C_{\beta,M}\sum_{j=1}^{M}\sum_{1\leq j\neq k\leq M}|z_j|^\beta|z_k|.$$ Profile expansion ----------------- This subsection contains a profile decomposition and an energy Pythagorean expansion results. We use similar arguments as the ones in Holmer-Roudenko [@HOLROU Lemma $5.2$] (see also Fang-Xie-Cazenave [@JIANCAZENAVE Theorem 5.1], with $(N,\alpha)=(3,2)$) and, for the sake of completeness, we provide the details here. \[LPD\] [**(Profile decomposition)**]{} Let $\phi_n(x)$ be a radial uniformly bounded sequence in $H^1(\mathbb{R}^3)$. Then for each $M\in \mathbb{N}$ there exists a subsequence of $\phi_n$ (also denoted by $\phi_n$), such that, for each $1\leq j\leq M$, there exist a profile $\psi^j$ in $H^1(\mathbb{R}^3)$, a sequence $t_n^j$ of time shifts and a sequence $W_n^M$ of remainders in $H^1(\mathbb{R}^3)$, such that $$\label{Aproximation} \phi_n(x)=\sum_{j=1}^{M}U(-t_n^j)\psi^j(x)+W_n^M(x)$$ with the properties: - *Pairwise divergence* for the time sequences. For $1\leq k\neq j\leq M$, $$\label{PD} \lim\limits_{n \rightarrow +\infty}|t_n^j-t_n^k|=+\infty.$$ - *Asymptotic smallness* for the remainder sequence[^10] $$\label{AS} \lim\limits_{M \rightarrow +\infty}\left(\lim\limits_{n \rightarrow +\infty}\|U(t)W_n^M\|_{S(\dot{H}^{s_c})}\right)=0.$$ - *Asymptotic Pythagoream expansion*. For fixed $M\in \mathbb{N}$ and any $s\in [0,1]$, we have $$\label{PDNHs}\|\phi_n\|^2_{\dot{H}^s}=\sum_{j=1}^{M}\|\psi^j\|^2_{\dot{H}^s}+\|W_n^M\|^2_{\dot{H}^s}+o_n(1)$$ where $o_n(1) \rightarrow 0$ as $n\rightarrow +\infty$. Let $C_1>0$ such that $\|\phi_n\|_{H^1}\leq C_1$. For every $(a,r)$ $\dot{H}^{s_c}$-admissible we can define $r_1=2r$ and $a_1=\frac{4r}{r(3-2s_c)-3}$. Note that $(a_1,r_1)$ is also $\dot{H}^{s_c}$-admissible, then combining the interpolation inequality with $\eta=\frac{3}{r(3-2s_c)-3}\in (0,1)$ and the Strichartz estimate , we have $$\begin{aligned} \label{LPD1} \|U(t)W_n^M\|_{L_t^aL^r_x}&\leq&\|U(t)W_n^M\|^{1-\eta}_{L_t^{a_1}L^{r_1}_x}\|U(t)W_n^M\|^\eta_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}\nonumber\\ & \leq & \|W_n^M\|^{1-\eta}_{\dot{H}^{s_c}}\|U(t)W_n^M\|^\eta_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}.\end{aligned}$$ Since we will have $\|W_n^M\|_{\dot{H}^{s_c}}\leq C_1$, then we need to show that the second norm in the right hand side of goes to zero as $n$ and $M$ go to infinite, that is $$\label{LPD2} \lim\limits_{M \rightarrow +\infty}\left(\limsup\limits_{n \rightarrow +\infty}\|U(t)W_n^M\|_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}\right)=0.$$  First we construct $\psi^1_n$, $t_n^1$ and $W_n^1$. Let $$A_1=\limsup \limits_{n\rightarrow +\infty} \|U(t)\phi_n\|_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}.$$ If $A_1=0$, the proof is complete with $\psi^j=0$ for all $j=1,\dots,M$. Assume that $A_1>0$. Passing to a subsequence, we may consider $A_1=\lim \limits_{n\rightarrow +\infty} \|U(t)\phi_n\|_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}$. We claim that there exist a time sequence $t_n^1$ and $\psi^1$ such that $U(t_n^1)\phi_n \rightharpoonup \psi^1$ and $$\label{LPD22} \beta C_1^ { \frac{3-2s_c}{2s_c(1-s_c)} }\|\psi^1\|_{\dot{H}^{s_c}}\geq A_1^{\frac{3-2s_c^2}{2s_c(1-s_c)}},$$ where $\beta>0$ is independent of $C_1$, $A_1$ and $\phi_n$. Indeed, let $\zeta\in C^\infty_0(\mathbb{R}^3)$ a real-valued and radially symmetric function such that $0\leq \zeta \leq 1$, $\zeta(\xi)=1$ for $|\xi|\leq 1$ and $\zeta(\xi)=0$ for $|\xi|\geq 2$. Given $r>0$, define $\chi_r$ by $\widehat{\chi_r}(\xi)=\zeta(\frac{\xi}{r})$. From the Sobolev embedding and since the operator $U(t)$ is an isometry in $H^{s_c}$, we deduce (recalling $0<s_c<1$) $$\begin{aligned} \label{LPD23} \|U(t)\phi_n -\chi_r*U(t)\phi_n\|^2_{L^\infty_tL_x^{\frac{6}{3-2s_c}}} &\leq c\|U(t)\phi_n -\chi_r*U(t)\phi_n\|^2_{L^\infty_tH_x^{s_c} } \nonumber \\ &\leq c \int |\xi|^{2s_c}|(1-\widehat{\chi_r})^2|\widehat{\phi}_n(\xi)|^2d\xi \nonumber \\ &\leq c\int_{|\xi|>r} |\xi|^{-2(1-s_c)} |\xi|^2 |\widehat{\phi}_n(\xi)|^2d\xi \nonumber \\ &\leq c r^{-2(1-s_c)}\|\phi\|^2_{\dot{H}^1} \leq c r^{-2(1-s_c)}C_1^2. \end{aligned}$$ Choosing $$\label{LPD24} r=\left(\frac{4\sqrt{c}C_1}{A_1}\right)^{\frac{1}{1-s_c}}$$ and for $n$ large enough we have $$\label{LPD3} \|\chi_r*U(t)\phi_n\|_{L^\infty_tL_x^{\frac{6}{3-2s_c}}}\geq \frac{A_1}{2}.$$ Note that, from the standard interpolation in Lebesgue spaces $$\begin{aligned} \label{LPD4} \|\chi_r*U(t)\phi_n\|^3_{L^\infty_tL_x^{\frac{6}{3-2s_c}}}&\leq&\|\chi_r*U(t)\phi_n\|^{3-2s_c}_{L^\infty_tL_x^2} \|\chi_r*U(t)\phi_n\|^{2s_c}_{L^\infty_tL_x^\infty}\nonumber \\ &\leq & C_1^{3-2s_c}\|\chi_r*U(t)\phi_n\|^{2s_c}_{L^\infty_tL_x^\infty},\end{aligned}$$ thus inequalities and lead to $$\|\chi_r*U(t)\phi_n\|_{L^\infty_tL_x^\infty}\geq \left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}}.$$ It follows from the radial Sobolev Gagliardo-Nirenberg inequality (since all $\phi_n$ are radial functions and so are $\chi_r*U(t)\phi_n$) that $$\begin{aligned} \|\chi_r*U(t)\phi_n\|_{L^\infty_tL_x^\infty(|x|\geq R)}&\leq &\frac{1}{R} \|\chi_r*U(t)\phi_n\|^{\frac{1}{2}}_{L^2_x}\|\nabla(\chi_r*U(t)\phi_n)\|^{\frac{1}{2}}_{L^2_x}\leq \frac{C_1}{R},\end{aligned}$$ which implies for $R>0$ sufficiently large $$\|\chi_r*U(t)\phi_n\|_{L^\infty_tL_x^\infty(|x|\leq R)}\geq \frac{1}{2} \left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}},$$ where we have used the two last inequalities. Now, let $t_n^1$ and $x_n^1$, with $|x_n^1|\leq R$, be sequences such that for each $n\in \mathbb{N}$ $$\left|\chi_r*U(t_n^1)\phi_n(x_n^1)\right|\geq \frac{1}{4}\left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}}$$ or $$\label{A_1} \frac{1}{4}\left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}}\leq \left|\int \chi_r(x_n^1-y)U(t_n^1)\phi_n(y)dy\right|.$$ On the other hand, since $\|U(t_n^1)\phi_n\|_{H^1}=\|\phi_n\|_{H^1}\leq C_1$ then $U(t^1_n)\phi_n$ converges weakly in $H^1$, i.e., there exists $\psi^1$ a radial function such that (up to a subsequence) $U(t_n^1)\phi_n \rightharpoonup \psi^1$ in $H^1$ and $\|\psi^1\|_{H^1}\leq \limsup \limits_{n\rightarrow +\infty}\|\phi_n\|_{H^1}\leq C_1$. In addition, $x_n^1\rightarrow x^1$ (also up to a subsequence) since $x_n^1$ is bounded. Hence the inequality , the Plancherel formula and the Cauchy-Schwarz inequality yield $$\frac{1}{8}\left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}}\leq \left|\int \chi_r(x^1-y) \psi^1(y)dy \right|\leq \|\chi_r\|_{\dot{H}^{-s_c}}\|\psi^1\|_{\dot{H}^{s_c}},$$ which implies $\frac{1}{8}\left(\frac{A_1}{2C_1^{\frac{3-2s_c}{3}}} \right)^{\frac{3}{2s_c}}\leq cr^{\frac{3-2s_c}{2}}\|\psi^1\|_{\dot{H}^{s_c}}$, where we have used $$\|\chi_r\|_{\dot{H}^{-s_c}}=\left(\int_{0<|\xi|<2r}|\xi|^{-2s_c}|\widehat{\chi_r}(\xi)|^2d\xi\right)^\frac{1}{2}\leq c\left(\int_{0}^{2r}\rho^{-2s_c}\rho^{2}d\rho\right)^\frac{1}{2}\leq cr^{\frac{3-2s_c}{2}}.$$ Therefore in view of our choice of $r$ (see ) we obtain , concluding the claim.  Next, define $W^1_n =\phi_n-U(-t_n^1)\psi^1$. It is easy to see that, for any $0\leq s\leq 1$, - $U(t_n^1)W^1_n \rightharpoonup 0$ in $H^1$ (since $U(t_n^1)\phi_n \rightharpoonup \psi^1$), - $\langle \phi_n,U(-t^1_n)\psi^1 \rangle_{\dot{H}^s}=\langle U(t^1_n)\phi_n,\psi^1 \rangle_{\dot{H}^s}\rightarrow \|\psi^1\|^2_{\dot{H}^s}$, - $\|W_n^1\|^2_{\dot{H}^s}=\|\phi_n\|^2_{\dot{H}^s}-\|\psi^1\|^2_{\dot{H}^s}+o_n(1)$. The last item, with $s=0$ and $s=1$, implies $\|W_n^1\|_{H^1}\leq C_1$.  Let $A_2=\limsup \limits_{n\rightarrow +\infty}\|U(t)W_n^1\|_{L_t^\infty L^{\frac{6}{3-2s}}_x}$. If $A_2=0$ the result follows taking $\psi^j=0$ for all $j=2,\dots,M$.. Let $A_2>0$, repeating the above argument with $\phi_n$ replaced by $W_n^1$ we obtain a sequence $t_n^2$ and a function $\psi^2$ such that $U(t_n^2)W_n^1\rightharpoonup \psi^2$ in $H^1$ and $\beta C_1^ { \frac{3-2s_c}{2s_c(1-s_c)} }\|\psi^2\|_{\dot{H}^{s_c}}\geq A_2^{\frac{3-2s_c^2}{2s_c(1-s_c)}}. $  We now prove that $|t_n^2-t_n^1|\rightarrow +\infty$. In fact, if we suppose (up to a subsequence) $t_n^2-t_n^1\rightarrow t^*$ finite, then $$U(t_n^2-t_n^1)\left(U(t_n^1)\phi_n-\psi^1 \right)=U(t_n^2)\left(\phi_n-U(-t_n^1)\psi^1 \right)=U(t_n^2)W_n^1\rightharpoonup \psi^2.$$ On the other hand, since $U(t_n^1)\phi_n\rightharpoonup \psi^1$, the left side of the above expression converges weakly to $0$, and thus $\psi^2=0$, a contradiction. Define $W_n^2=W_n^1-U(-t_n^2)\psi^2$. For any $0\leq s\leq 1$, since $|t_n^1-t_n^2|\rightarrow +\infty$, we deduce $$\begin{aligned} \langle \phi_n,U(-t_n^2)\psi^2 \rangle_{\dot{H}^{s}}&=&\langle U(t_n^2)\phi_n,\psi^2 \rangle_{\dot{H}^{s}}\\ &=&\langle U(t_n^2)\left(W_n^1+U(-t_n^1)\psi^1\right),\psi^2 \rangle_{\dot{H}^{s}}\\&=&\langle U(t_n^2)W_n^1,\psi^2 \rangle_{\dot{H}^{s}}+\langle U(t_n^2-t_n^1)\psi^1,\psi^2 \rangle_{\dot{H}^{s}} \\ &\rightarrow& \|\psi^2\|^2_{\dot{H}^{s}}.\end{aligned}$$ In addition, the definition of $W_n^2$ implies that $$\|W_n^2\|^2_{\dot{H}^s}=\|W_n^1\|^2_{\dot{H}^{s_c}}-\|\psi^2\|^2_{\dot{H}^s}+o_n(1)$$ and $\|W_n^2\|_{H^1}\leq C_1$.  By induction we can construct $\psi^M$, $t_n^M$ and $W_n^M$ such that $U(t_n^M)W_n^{M-1}\rightharpoonup \psi^M$ in $H^1$ and $$\label{LPD5} \beta C_1^ { \frac{3-2s_c}{2s_c(1-s_c)} }\|\psi^M\|_{\dot{H}^{s_c}}\geq A_M^{\frac{3-2s_c^2}{2s_c(1-s_c)}},$$ where $A_M=\lim \limits_{n\rightarrow +\infty}\|U(t)W_n^{M-1}\|_{L_t^\infty L^{\frac{6}{3-2s_c}}_x}$.  Next, we show . Suppose $1\leq j<M$, we prove that $|t^M_n-t_n^j|\rightarrow +\infty$ by induction assuming $|t^M_n-t_n^k|\rightarrow +\infty$ for $k=j+1, \dots, M-1$. Indeed, let $t^M_n-t_n^j\rightarrow t_0$ finite (up to a subsequence) then it is easy to see $$U(t_n^M-t_n^j)\left(U(t_n^j)W_n^{j-1}-\psi^j\right)-U(t_n^M-t_n^{j+1})\psi^{j+1}-...-U(t_n^M-t_n^{M-1})\psi^{M-1}$$ $$=U(t_n^M)W_n^{M-1}\rightharpoonup \psi^M.$$ Since the left side converges weakly to $0$, we have $\psi^M=0$, a contradiction. We now consider $$W_n^M=\phi_n-U(-t_n^1)\psi^1-U(-t_n^2)\psi^2-...-U(-t_n^M)\psi^M.$$ Similarly as before, by we get for any $0\leq s\leq 1$ $$\langle \phi_n,U(-t_n^M)\psi^M \rangle_{\dot{H}^s}=\langle U(t_n^M)W_n^{M-1},\psi^M \rangle_{\dot{H}^s}+o_n(1),$$ and so $\langle \phi_n,U(-t_n^M)\psi^M \rangle_{\dot{H}^s}\rightarrow \|\psi^M\|^2_{\dot{H}^s}$. Thus expanding $\|W_n^M\|^2_{\dot{H}^s}$ we deduce that also holds.  Finally, the inequality together with the relation yield $$\sum_{M\geq 1} \left(\frac{A_M^{\frac{3-2s_c^2}{s_c(1-s_c)}}}{\beta^2C_1^{ \frac{3-2s_c}{s_c(1-s_c)} }}\right)\leq \lim_{n\rightarrow+\infty}\|\phi_n\|^2_{\dot{H}^{s_c}}<+\infty,$$ which implies that $A_M\rightarrow 0$ as $M\rightarrow +\infty$ i.e., holds[^11]. Therefore, from we get . This completes the proof. \[RLPD\] It follows from the proof of Proposition \[LPD\] that $$\label{RLPD1} \lim\limits_{M,n\rightarrow \infty} \|W_n^{M}\|_{L^{p}} =0,$$ where $2<p<6$. Indeed, first we show $$\label{RLPD2} \lim\limits_{M \rightarrow +\infty}\left(\lim\limits_{n \rightarrow +\infty}\|U(t)W_n^M\|_{L_t^\infty L^p_x}\right)=0.$$ Note that, $\dot{H}^{s}\hookrightarrow L^{p}$ where $s=\frac{3}{2}-\frac{3}{p}$ (see inequality ). Since $2<p<6$ then $0<s<1$, thus repeating the argument used for showing with $\frac{6}{3-2s_c}$ replaced by $p$ and $s_c$ by $s$, we obtain . On the other hand, follows directly from and the inequality $$\|W_n^{M}\|_{L_x^{p}}\leq \|U(t)W_n^{M}\|_{L^\infty_tL_x^{p}},$$ since $W_n^{M}=U(0)W_n^{M}$. \[EPE\][**(Energy Pythagoream Expansion)**]{} Under the hypothesis of Proposition \[LPD\] we obtain $$\label{PDE} E[\phi_n]=\sum_{j=1}^{M}E[U(-t_n^j)\psi^j]+E[W_n^M]+o_n(1).$$ By definition of $E[u]$ and with $s=1$, we have $$E[\phi_n]-\sum_{j=1}^{M}E[U(-t_n^j)\psi^j]-E[W_n^M]=-\frac{A_n}{\alpha+2}+o_n(1),$$ where $$A_n=\left\| |x|^{-b}|\phi_n|^{4} \right\|_{L^1}-\sum_{j=1}^{M}\left\| |x|^{-b} |U(-t_n^j)\psi^j|^{4} \right\|_{L_x^1}-\left\| |x|^{-b}|W_n^M|^{4} \right\|_{L^1}.$$  For a fixed $M\in \mathbb{N}$, if $A_n \rightarrow 0$ as $n\rightarrow +\infty$ then holds. To prove this fact, pick $M_1\geq M$ and rewrite the last expression as $$\begin{aligned} A_n&=& \int \left( |x|^{-b}|\phi_n|^{4} - \sum_{j=1}^{M} |x|^{-b} |U(-t_n^j)\psi^j|^{4} - |x|^{-b} |W_n^M|^{4} \right)dx \\ &=& I^1_n+I^2_n+I^3_n, \end{aligned}$$ where $$\begin{aligned} I^1_n&=& \int |x|^{-b} \left[ |\phi_n|^{4}-|\phi_n-W_n^{M_1}|^{4}\right]dx, \\ I^2_n&=& \int |x|^{-b} \left[ |W_n^{M_1}-W_n^M|^{4} -|W_n^M|^{4}\right]dx, \end{aligned}$$ $$\begin{aligned} I^3_n&=& \int |x|^{-b}\left[ |\phi_n-W_n^{M_1}|^{4}- \sum_{j=1}^{M} |U(-t_n^j)\psi^j|^{4}- |W_n^{M_1}-W_n^M|^{4} \right] dx .\end{aligned}$$  We first estimate $I^1_n$. Combining and Lemma \[LEWO\] (i)-(ii) we have $$\begin{aligned} |I^1_n|&\lesssim& \int |x|^{-b}\left( |\phi_n|^{3}|W_n^{M_1}|+|\phi_n||W_n^{M_1}|^{3} + |W_n^{M_1}|^{4} \right)dx\\ &\lesssim& \left( \|\phi_n\|^{3}_{L^r}\|W_n^{M_1}\|_{L^r}+ \|\phi_n\|_{L^r}\|W_n^{M_1}\|^{3}_{L^{r}} + \|W_n^{M_1}\|^{4}_{L^{r}} \right)+\\ & & \left( \|\phi_n\|^{3}_{L^{4}}\|W_n^{M_1}\|_{L^{4}}+ \|\phi_n\|_{L^{4}}\|W_n^{M_1}\|^{3}_{L^{4}} + \|W_n^{M_1}\|^{4}_{{L^{4}}} \right)\\ &\lesssim& \|\phi_n\|^{3}_{H^1}\|W_n^{M_1}\|_{L^r}+ \|\phi_n\|_{H^1}\|W_n^{M_1}\|^{3}_{L^r} + \|W_n^{M_1}\|^{3}_{L^r} +\\ & & \|\phi_n\|^{3}_{H^1}\|W_n^{M_1}\|_{L^{4}}+ \|\phi_n\|_{H^1}\|W_n^{M_1}\|^{3}_{L^{4}} + \|W_n^{M_1}\|^{4}_{L^{4}} ,\end{aligned}$$ where $\frac{12}{3-b}<r<6$. In view of inequality and since $\{\phi_n\}$ is uniformly bounded in $H^1$, we conclude that[^12] $$I^1_n\rightarrow +\infty\;\;\textnormal{as}\;\;n, M_1\rightarrow +\infty.$$  Also, by similar arguments (replacing $\phi_n$ by $W_n^{M}$) we have $$I^2_n\rightarrow +\infty\;\;\textnormal{as}\;\;n, M_1\rightarrow +\infty,$$ where we have used that $W_n^{M}$ is uniformly bounded by .  Finaly we consider the term $I^3_n$. Since, $$\phi_n-W_n^{M_1}=\sum\limits_{j=1}^{M_1}U(-t_n^j)\psi^j\,\,\,\textnormal{and}\,\,\,W_n^{M}-W_n^{M_1}=\sum\limits_{j=M+1}^{M_1}U(-t_n^j)\psi^j,$$ we can rewrite $I^3_n$ as $$\begin{aligned} I^3_n=\int |x|^{-b} \left( \left| \sum\limits_{j=1}^{M_1} U(-t_n^j)\psi^j \right|^{4}- \sum\limits_{j=1}^{M_1}| U(-t_n^j)\psi^j |^{4}\right)dx\end{aligned}$$ $$-\int |x|^{-b} \left( \left| \sum\limits_{j=M+1}^{M_1} U(-t_n^j)\psi^j \right|^{4}- \sum\limits_{j=M+1}^{M_1}| U(-t_n^j)\psi^j |^{4}\right)dx.$$ To complete the prove we make use of the following claim. *Claim.* For a fixed $M_1\in \mathbb{N}$ and for some $j_0\in \mathbb{N}$ ($j_0< M_1$), we get $$D_n= \left\||x|^{-b} \left|\sum_{j=j_0}^{M_1} U(-t_n^j)\psi \right|^{4} \right\|_{L^1_x} - \sum_{j=j_0}^{M_1} \left\| |x|^{-b} |U(-t_n^j)\psi^j|^{4} \right\|_{L^1_x}\rightarrow 0\;,\;\;\textnormal{as}\;\;\;n\rightarrow +\infty.$$  Indeed, it is clear that the last limit implies that $I^3_n\rightarrow 0\;\textnormal{as}\;n\rightarrow +\infty$ completing the proof of relation . To prove the claim note that implies $$D_n\leq \sum_{j\neq k}^{M_1}\int |x|^{-b}|U(-t_n^j)\psi^j||U(-t_n^k)\psi^k|^{3}dx.$$ Thus, from Lemma \[LEWO\] (i) one has $$\label{lematecnico1} E^{j,k}_n\leq c \| U(-t_n^k)\psi^k\|^{3}_{L^{4}_x}\| U(-t_n^j)\psi^j\|_{L^{4}_x}+c\| U(-t_n^k)\psi^k\|^{3}_{L^r_x}\| U(-t_n^j)\psi^j\|_{L^r_x},$$ where $2<\frac{12}{3-b}< r<6$ and $E^{j,k}_n=\int |x|^{-b}|U(-t_n^j)\psi^j||U(-t_n^k)\psi^k|^{3}dx$. In view of we can consider that $t_n^k$, $t_n^j$ or both go to infinite as $n$ goes to infinite. If $t_n^j\rightarrow +\infty$ as $n\rightarrow +\infty$, so it follow from the last inequality and since $H^{1}\hookrightarrow L^{4}$ and $H^{1}\hookrightarrow L^{r}$ that $$\begin{aligned} E^{j,k}_n & \leq& c \|\psi^k\|^{3}_{H^1}\| U(-t_n^j)\psi^j\|_{L^{4}_x} +c\| \psi^k\|^{3}_{H^1} \| U(-t_n^j)\psi^j\|_{L^r_x}\\ & \leq& c \| U(-t_n^j)\psi^j\|_{L^{4}_x} +c\| U(-t_n^j)\psi^j\|_{L^r_x},\end{aligned}$$ where in the last inequality we have used that $(\psi^k)_{k\in\mathbb{N}}$ is a uniformly bounded sequence in $H^1$. Hence, if $n\rightarrow +\infty$ we have $t_n^j\rightarrow +\infty$ and using with $t=t_n^j$ and $f=\psi^j$ we conclude that $E^{j,k}_n\rightarrow 0$ as $n\rightarrow +\infty$. Similarly for the case $t^k_n\rightarrow +\infty$ as $n\rightarrow +\infty$, we have $$\begin{aligned} E^{j,k}_n & \leq& c \| U(-t_n^k)\psi^k\|^{3}_{L^{4}_x} \|\psi^j\|_{H^1}+c\| U(-t_n^k)\psi^k\|^{3}_{L^r_x}\| \psi^j\|_{H^1}\\ & \leq& c \| U(-t_n^k)\psi^k\|^{3}_{L^{4}_x}+c\| U(-t_n^k)\psi^k\|^{3}_{L^r_x},\end{aligned}$$ which implies that $E^{j,k}_n\rightarrow 0$ as $n\rightarrow +\infty$ by with $t=t_n^k$ and $f=\psi^k$. Finally, since $D_n$ is a finite sum of terms in the form of $E^{j,k}$ we deduce $D_n\rightarrow 0$ as $n\rightarrow +\infty$. Critical solution {#CCS} ----------------- In this subsection, assuming that $\delta_c<E[u]^{s_c}M[u]^{1-s_c}$ (see ), we construct a global solution, denoted by $u_c$, of with infinite Strichartz norm $\|\cdot\|_{S(\dot{H}^{s_c})}$ and satisfying $$E[u_c]^{s_c}M[u_c]^{1-s_c}=\delta_c.$$ Next, we show that the flow associated to this critical solution is precompact in $H^1(\mathbb{R}^3)$. \[ECS\][**(Existence of a critical solution)**]{} If $\delta_c<E[Q]^{s_c}M[Q]^{1-s_c},$ then there exists a radial function $u_{c,0}\in H^1(\mathbb{R}^3)$ such that the corresponding solution $u_c$ of the IVP is global in $H^1(\mathbb{R}^3)$. Moreover the following properties hold - $M[u_c]=1$, - $E[u_c]^{s_c}=\delta_c$, - $\| \nabla u_{c,0} \|_{L^2}^{s_c} \|u_{c,0}\|_{L^2}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c}$, - $\|u_{c}\|_{S(\dot{H}^{s_c})}=+\infty$.   Recall from Subsection \[SPMR\] that there exists a sequence of solutions $u_n$ to with $H^1$ initial data $u_{n,0}$, with $\|u_n\|_{L^2} = 1$ for all $n\in \mathbb{N}$, such that $$\label{PCS1} \|\nabla u_{n,0}\|^{s_c}_{L^2} < \|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2}$$ and $$\label{PCS2} E[u_n] \searrow \delta_c^{\frac{1}{s_c}}\;\; \textnormal{as}\;\; n \rightarrow +\infty.$$ Moreover $$\label{un} \|u_n\|_{S(\dot{H}^{s_c})}=+\infty$$ for every $n\in \mathbb{N}$. Note that, in view of the assumption $\delta_c<E[Q]^{s_c}M[Q]^{1-s_c}$, there exists $a \in (0,1)$ such that $$\label{PCS21} E[u_n]\leq a E[Q]M[Q]^{\sigma},$$ where $\sigma=\frac{1-s_c}{s_c}$. Furthermore, implies by Lemma \[LGS\] (ii) that $$\|\nabla u_{n,0}\|^{2}_{L^2} \leq w^{\frac{1}{s_c}} \|\nabla Q\|^{2}_{L^2}\|Q\|^{2\sigma}_{L^2},$$ where $w=\frac{E[u_n]^{s_c}M[u_n]^{1-s_c}}{E[Q]^{s_c}M[Q]^{1-s_c}}$, thus we deduce from and $\|u_n\|_{L^2} = 1$ that $w^{\frac{1}{s_c}}\leq a$ which implies $$\label{PCS22} \|\nabla u_{n,0}\|^{2}_{L^2} \leq a \|\nabla Q\|^{2}_{L^2}\|Q\|^{2\sigma}_{L^2}.$$  On the other hand, the linear profile decomposition (Proposition \[LPD\]) applied to $u_{n,0}$, which is a uniformly bounded sequence in $H^1(\mathbb{R}^3)$ by , yields $$\label{PCS3} u_{n,0}(x)=\sum_{j=1}^{M}U(-t_n^j)\psi^j(x)+W_n^M(x),$$ where $M$ will be taken large later. It follows from the Pythagorean expansion , with $s=0$, that for all $M\in \mathbb{N}$ $$\label{PCS4} \sum_{j=1}^{M}\|\psi^j\|^2_{L^2}+\lim_{n\rightarrow +\infty}\|W_n^M\|^2_{L^2}\leq \lim_{n\rightarrow +\infty}\|u_{n,0}\|^2_{L^2}= 1,$$ this implies that $$\label{PCS41} \sum_{j=1}^{M}\|\psi^j\|^2_{L^2}\leq 1.$$ In addition, another application of , with $s=1$, and lead to $$\label{Sumpsij} \sum_{j=1}^{M}\|\nabla \psi^j\|^2_{L^2}+\lim_{n\rightarrow +\infty}\|\nabla W_n^M\|^2_{L^2}\leq \lim_{n\rightarrow +\infty}\|\nabla u_{n,0}\|^2_{L^2}\leq a\|\nabla Q\|^2_{L^2}\|Q\|^{2\sigma}_{L^2},$$ and so $$\label{PCS5} \|\nabla \psi^j\|^{s_c}_{L^2}\leq a^{\frac{s_c}{2}}\|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2},\;\;j=1,\dots,M.$$  Let $\{t^j_n\}_{n\in \mathbb{N}}$ be the sequence given by Proposition \[LPD\]. From , and the fact that $U(t)$ is an isometry in $L^2(\mathbb{R}^3)$ and $\dot{H}^1(\mathbb{R}^3)$ we deduce $$\label{U(-t_n^j)} \|U(-t_n^j)\psi^j\|^{1-s_c}_{L^2_x}\|\nabla U(-t_n^j)\psi^j\|^{s_c}_{L^2_x}\leq a^{\frac{s_c}{2}} \|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2}.$$ Now, Lemma \[LGS\] (i) yields $$\label{PCS6} E[U(-t_n^j)\psi^j]\geq c(b)\|\nabla \psi^j\|_{L^2}\geq 0$$  A complete similar analysis also gives, for all $M\in \mathbb{N}$, $$\lim_{n\rightarrow +\infty}\|W_n^M\|^2_{L^2}\leq 1,$$ $$\lim_{n\rightarrow +\infty}\|\nabla W_n^M\|^{s_c}_{L^2}\leq a^{\frac{s_c}{2}}\|\nabla Q\|^{s_c}_{L^2}\|Q\|^{1-s_c}_{L^2},$$ and for $n$ large enough (depending on $M$) $$\label{PCS7} E[W_n^M]\geq 0.$$  The energy Pythagorean expansion (Proposition \[EPE\]) allows us to deduce $$\sum_{j=1}^{M}\lim_{n\rightarrow+\infty}E[U(-t_n^j)\psi^j]+\lim_{n\rightarrow+\infty}E[W_n^M]=\lim_{n\rightarrow+\infty}E[u_{n,0}]=\delta_c^{\frac{1}{s_c}},$$ which implies, by and , that $$\label{PCS8} \lim_{n\rightarrow\infty}E[U(-t_n^j)\psi^j]\leq \delta_c^{\frac{1}{s_c}},\;\textnormal{for all}\;\;j=1,...,M.$$  Now, if more than one $\psi^j\neq 0$, we show a contradiction and thus the profile expansion given by is reduced to the case that only one profile is nonzero. In fact, if more than one $\psi^j\neq 0$, then by we must have $M[\psi^j]<1$ for each $j$. Passing to a subsequence, if necessary, we have two cases to consider:  [**Case $1$**]{}. If for a given $j$, $t^j_n\rightarrow t^*$ finite (at most only one such $j$ exists by ), then the continuity of the linear flow in $H^1(\mathbb{R}^3)$ yields $$\label{widetildepsi} U(-t_n^j)\psi^j\rightarrow U(-t^*)\psi^j\;\;\;\;\textnormal{strongly in}\;H^1.$$ Let us denote the solution of with initial data $\psi$ by INLS$(t)\psi$. Set $\widetilde{\psi}^j=\textnormal{INLS}(t^*)(U(-t^*)\psi^j)$ so that $\mbox{INLS}(-t^*)\widetilde{\psi}^j=U(-t^*)\psi^j$. Since the set $$\mathcal{K}:=\left\{u_0\in H^1(\mathbb{R}^3):\; \textrm{relations} \; \eqref{EMC} \; \textrm{and} \; \eqref{GFC} \; \textrm{hold}\;\right\}$$ is closed in $H^1(\mathbb{R}^3)$ then $\widetilde{\psi}^j\in \mathcal{K}$ and therefore INLS$(t)\widetilde{\psi}^j$ is a global solution by Theorem \[TG\]. Moreover from , and the fact that $M[\psi^j]<1$ we have $$\|\widetilde{\psi}^j\|^{1-s_c}_{L^2_x}\|\nabla \widetilde{\psi}^j\|^{s_c}_{L^2_x}\leq \|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2}$$ and $$E[\widetilde{\psi}^j]^{s_c}M[\widetilde{\psi}^j]^{1-s_c}< \delta_c.$$ So, the definition of $\delta_c$ (see ) implies $$\label{CSCP} \|\textnormal{INLS}(t)\widetilde{\psi}^j\|_{S(\dot{H}^{s_c})}<+\infty.$$ Finally, from it is easy to see $$\label{CSWO3} \lim_{n\rightarrow+\infty}\|\textnormal{INLS}(-t_n^j)\widetilde{\psi}^j-U(-t_n^j)\psi^j\|_{H^1_x}=0.$$  [**Case $2$.**]{} If $|t^j_n|\rightarrow+\infty$ then by Lemma \[LEWO\] (iii), $\left\||x|^{-b}|U(-t_n^j)\psi^j|^{4}\right\|_{L^1_x}\rightarrow 0$. Thus, by the definition of Energy and the fact that $U(t)$ is an isometry in $\dot{H}^1(\mathbb{R}^3)$, we deduce $$\label{CS0} \left( \frac{1}{2}\|\nabla \psi^j\|^2_{L^2}\right)^{s_c}= \lim_{n \rightarrow \infty} E [ U(-t_n^j) \psi^j]^{s_c} \leq \delta_c < E[Q]^{s_c}M[Q]^{1-s_c},$$ where we have used . Therefore, by the existence of wave operator, Proposition \[PEWO\] with $\lambda=(\frac{2 s_c}{3+b})^{\frac{s_c}{2}}<1$ (see also Remark \[backward\]), there exists $\widetilde{\psi}^j\in H^1(\mathbb{R}^3)$ such that $$\label{CSWO1} M[\widetilde{\psi}^j]=M[\psi^j]\;\;\;\textrm{ and }\;\;\;\;E[\widetilde{\psi}^j]=\frac{1}{2}\|\nabla \psi^j\|^2_{L^2},$$ $$\label{CSWO2} \|\nabla \textnormal{INLS}(t)\widetilde{\psi}^j\|^{s_c}_{L^2_x}\|\widetilde{\psi}^j\|^{1-s_c}_{L^2}<\|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2}$$ and also holds in this case. Since $M[{\psi}^j]<1$ and using -, we get $E[\widetilde{\psi}^j]^{s_c}M[\widetilde{\psi}^j]^{1-s_c}<\delta_c$. Hence, the definition of $\delta_c$ together with also lead to . To sum up, in either case, we obtain a new profile $\widetilde{\psi}^j$ for the given $\psi^j$ such that hold.\ Next, we define $u_n(t)=\textnormal{INLS}(t)u_{n,0}$; $v^j(t)=\textnormal{INLS}(t)\widetilde{\psi}^j$; $\widetilde{u}_n(t)=\sum_{j=1}^{M}v^j(t-t_n^j)$ and $$\label{CSR}\widetilde{W}_n^M=\sum_{j=1}^{M}\left[ U(-t_n^j)\psi^j-\textnormal{INLS}(-t_n^j)\widetilde{\psi}^j \right]+W_n^M.$$ Then $\widetilde{u}_n(t)$ solves the following equation $$\label{widetildeun} i\partial_t\widetilde{u}_n+\Delta \widetilde{u}_n+|x|^{-b}|\widetilde{u}_n|^{2}\widetilde{u}_n=e_n^M,$$ where $$\label{CSR1} e_n^M=|x|^{-b}\left( |\widetilde{u}_n|^{2}\widetilde{u}_n-\sum_{j=1}^{M}|v^j(t-t_n^j)|^{2}v^j(t-t_n^j) \right).$$ Also note that by definition of $\widetilde{W}_n^M$ in and we can write $$\label{aproximation1} u_{n,0}=\sum_{j=1}^{M}\textnormal{INLS}(-t_n^j)\widetilde{\psi}^j+\widetilde{W}_n^M,$$ so it is easy to see $u_{n,0}-\widetilde{u}_n(0)=\widetilde{W}_n^M$, then combining and the Strichartz inequality , we estimate $$\|U(t)\widetilde{W}_n^M\|_{S(\dot{H}^{s_c})}\leq c\sum_{j=1}^{M}\|\textnormal{INLS}(-t_n^j)\widetilde{\psi}^j-U(-t_n^j)\psi^j\|_{H^1}+\|U(t)W_n^M\|_{S(\dot{H}^{s_c})},$$ which implies $$\label{CSR2} \lim_{M\rightarrow +\infty} \left[\lim_{n\rightarrow +\infty} \|U(t)(u_{n,0}-\widetilde{u}_{n,0})\|_{S(\dot{H}^{s_c})}\right]=0,$$ where we used and .  The idea now is to approximate $u_n$ by $\widetilde{u}_n$. Therefore, from the long time perturbation theory (Proposition \[LTP\]) and we conclude $\|u_n\|_{S(\dot{H}^{s_c})}<+\infty$, for $n$ large enough, which is a contradiction with . Indeed, we assume the following two claims for a moment to conclude the proof.\ [**Claim $1$.**]{} For each $M$ and $\varepsilon>0$, there exists $n_0=n_0(M,\varepsilon)$ such that $$\label{claim2} n>n_0\;\; \Rightarrow\;\; \|e_n^M\|_{S'(\dot{H}^{-s_c})}+\|e_n^M\|_{S'(L^2)}+\|\nabla e_n^M\|_{S'(L^2)}\leq\varepsilon.$$  [**Claim $2$.**]{} There exist $L>0$ and $S>0$ independent of $M$ such that for any $M$, there exists $n_1=n_1(M)$ such that $$\label{claim1} n>n_1\;\; \Rightarrow\;\; \|\widetilde{u}_n\|_{S(\dot{H}^{s_c})}\leq L\;\;\textnormal{and}\;\;\|\widetilde{u}_n\|_{L^\infty_tH^1_x}\leq S.$$  Note that by , there exists $M_1=M_1(\varepsilon)$ such that for each $M>M_1$ there exists $n_2=n_2(M)$ such that $$n>n_2\;\; \Rightarrow\;\; \|U(t)(u_{n,0}-\widetilde{u}_{n,0})\|_{S(\dot{H}^{s_c})}\leq \varepsilon,$$ with $\varepsilon<\varepsilon_1$ as in Proposition \[LTP\]. Thus, if the two claims hold true, by Proposition \[LTP\], for $M$ large enough and $n>\max\{n_0,n_1,n_2\}$, we obtain $\|u_n\|_{S(\dot{H}^{s_c})}<+\infty$, reaching the desired contradiction .  Up to now, we have reduced the profile expansion to the case where $\psi^1\neq 0$ and $\psi^j= 0$ for all $j\geq 2$. We now begin to show the existence of a critical solution. From the same arguments as the ones in the previous case (the case when more than one $\psi^j \neq 0 $), we can find $\widetilde{\psi}^1$ such that $$u_{n,0}=\textnormal{INLS}(-t_n^1)\widetilde{\psi}^1+\widetilde{W}_n^M,$$ with $$\label{CSWO11} M[\widetilde{\psi}^1]=M[\psi^1]\leq 1$$ $$\label{CSWO221} E[\widetilde{\psi}^1]^{s_c}=\left(\frac{1}{2} \|\nabla \psi^1 \|^2_{L^2}\right)^{s_c}\leq \delta_c$$ $$\label{CSWO22} \|\nabla \textnormal{INLS}(t)\widetilde{\psi}^1\|^{s_c}_{L^2_x}\|\widetilde{\psi}^1\|^{1-s_c}_{L^2}< \|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2}$$ and $$\label{CSCP1} \lim_{n\rightarrow +\infty} \|U(t)(u_{n,0}-\widetilde{u}_{n,0})\|_{S(\dot{H}^{s_c})}=\lim_{n\rightarrow +\infty}\|U(t)\widetilde{W}_{n}^M\|_{S(\dot{H}^{s_c})}=0.$$  Let $\widetilde{\psi}^1=u_{c,0}$ and $u_c$ be the global solution to (in view of Theorem \[TG\] and inequalities -) with initial data $\widetilde{\psi}^1$, that is, $u_c(t)=\textnormal{INLS}(t)\widetilde{\psi}^1$. We claim that $$\label{claimfinal} \|u_c\|_{S(\dot{H}^{s_c})}=+\infty.$$  Indeed, suppose, by contradiction, that $\|u_c\|_{S(\dot{H}^{s_c})}<+\infty$. Let, $$\widetilde{u}_n(t)=\textnormal{INLS}(t-t_n^j)\widetilde{\psi}^1,$$ then $$\|\widetilde{u}_n(t)\|_{S(\dot{H}^{s_c})}=\|\textnormal{INLS}(t-t_n^j)\widetilde{\psi}^1\|_{S(\dot{H}^{s_c})}=\|\textnormal{INLS}(t)\widetilde{\psi}^1\|_{S(\dot{H}^{s_c})}=\|u_c\|_{S(\dot{H}^{s_c})}<+\infty.$$ Furthermore, it follows from - that $$\sup_{t\in \mathbb{R}}\|\widetilde{u}_n\|_{H^1_x}=\sup_{t\in \mathbb{R}}\|u_c\|_{H^1_x}<+\infty.$$ and $$\|U(t)(u_{n,0}-\widetilde{u}_{n,0})\|_{S(\dot{H}^{s_c})}\leq \varepsilon,$$ for $n$ large enough. Hence, by the long time perturbation theory (Proposition \[LTP\]) with $e=0$, we obtain $\|u_n\|_{S(\dot{H}^{s_c})}<+\infty$, which is a contradiction with .  On the other hand, the relation implies $E[u_c]^{s_c}M[u_c]^{1-s_c}=\delta_c$ (see ). Thus, we conclude from and $$M[u_c]=1\;\;\;\;\textnormal{and}\;\;\;\;E[u_c]^{s_c}=\delta_c.$$ Also note that implies (iii) in the statement of the Proposition \[ECS\].  To complete the proof it remains to establish Claims $1$ and $2$ (see and ).  [**Proof of Claim $1$.**]{} First, we show that for each $M$ and $\varepsilon>0$, there exists $n_0=n_0(M,\varepsilon)$ such that $\|e_n^M\|_{S'(\dot{H}^{-s_c})}< \frac{\varepsilon}{3}$. From and (with $\beta=2$), we deduce $$\label{ec21} \|e_n^M\|_{S'(\dot{H}^{-s_c})}\leq C_{\alpha,M}\sum_{j=1}^{M}\sum_{1\leq j\neq k\leq M} \left\||x|^{-b}|v^k|^2|v^j|\right\|_{L^{\widetilde{a}'}_tL^{\widehat{r}'}_x}.$$ We claim that the norm in the right hand side of goes to $0$ as $n\rightarrow +\infty$. Indeed, by the relation one has $$\begin{aligned} \label{ec22} \left\||x|^{-b}|v^k|^2|v^j|\right\|_{L^{\widetilde{a}'}_tL^{\widehat{r}'}_x} \leq& c \|v^k\|^{\theta}_{L^\infty_tH^1_x} \left\|\|v^k(t-t_n^k)\|^{2-\theta}_{L_x^{\widehat{r}}} \|v^j(t-t_n^j)\|_{L^{\widehat{r}}_x}\right\|_{L^{\widetilde{a}'}_t}. \end{aligned}$$ Fix $1\leq j\neq k\leq M$. Note that, $\|v^k\|_{H^1_x}<+\infty$ (see - ) and by $v^j$, $v^k\in S(\dot{H^{s_c}})$ and , so we can approximate $v^j$ by functions of $C_0^\infty(\mathbb{R}^{3+1})$. Hence, defining $$g_n(t)= \|v^k(t)\|^{2-\theta}_{L_x^{\widehat{r}}} \|v^j(t-(t_n^j-t_n^k))\|_{L^{\widehat{r}}_x},$$ we deduce - $g_n\in L^{\widetilde{a}'}_t$. Indeed, applying the Hölder inequality since $\frac{1}{\widetilde{a}'}=\frac{\alpha-\theta}{\widehat{a}}+\frac{1}{\widehat{a}}$ we get $$\|g_n\|_{L^{\widetilde{a}'}_t}\leq \|v^k\|^{2-\theta}_{L^{\widehat{a}}_t L_x^{\widehat{r}}} \|v^j\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x}\leq \|v^k\|^{2-\theta}_{S(\dot{H}^{s_c})} \|v^j\|_{S(\dot{H}^{s_c})}<+\infty.$$ Furthermore, implies that $g_n(t)\rightarrow 0$ as $n\rightarrow +\infty$. - $|g_n(t)|\leq KI_{supp(v^j)}\|v^k(t)\|^{2-\theta}_{L_x^{\widehat{r}}}\equiv g(t)$ for all $n$, where $K>0$ and $I_{supp(v^j)}$ is the characteristic function of $supp(v^j)$. Similarly as (i), we obtain $$\|g\|_{L^{\widetilde{a}'}_t}\leq \|v^k\|^{2-\theta}_{L^{\widehat{a}}_t L_x^{\widehat{r}}}\|I_{supp(v^j)}\|_{L^{\widehat{a}}_t L_x^{\widehat{r}}}<+\infty.$$ That is, $g\in L^{\widetilde{a}'}_t$. Then, the Dominated Convergence Theorem yields $\|g_n\|_{L^{\widetilde{a}'}_t}\rightarrow 0$ as $n\rightarrow +\infty$, and so combining this result with we conclude the proof of the first estimate.  Next, we prove $\|e_n^M\|_{S'(L^2)}<\frac{\varepsilon}{3}$. Using again the elementary inequality we estimate $$\|e_n^M\|_{S'(L^2)}\leq C_{\alpha,M}\sum_{j=1}^{M}\sum_{1\leq j\neq k\leq M} \left\||x|^{-b}|v^k|^2|v^j|\right\|_{L^{\widehat{q}'}_tL^{\widehat{r}'}_x}.$$ On the other hand, we have (see proof of Lemma \[LG1\] (ii)) $$\begin{aligned} \left\| |x|^{-b}|v^k|^2|v^j \right \|_{L_t^{\widehat{q}'}L^{\widehat{r}'}_x}&\leq& c \|v^k\|^{\theta}_{L^\infty_tH^1_x}\left \| \|v^k(t-t_n^k)\|^{2-\theta}_{ L_x^{\widehat{r}}} \|v^j(t-t_n^j)\|_{L^{\widehat{r}}_x}\right\|_{L_t^{\widehat{q}'}}\\ &\leq & c\|v^k\|^{\theta}_{L^\infty_tH^1_x} \|v^k\|^{2-\theta}_{L^{\widehat{a}}_t L_x^{\widehat{r}}} \|v^j\|_{L_t^{\widehat{q}}L^{\widehat{r}}_x}\\ &\leq & c\|v^k\|^{\theta}_{L^\infty_tH^1_x} \|v^k\|^{2-\theta}_{S(\dot{H}^{s_c})} \|v^j\|_{S(L^2)}.\end{aligned}$$ Since $v^j\in S(\dot{H}^{s_c})$ then by the norms $\|v^j\|_{S(L^2)}$ and $\|\nabla v^j\|_{S(L^2)}$ are bounded quantities. This implies that the right hand side of the last inequality is finite. Therefore, using the same argument as in the previous case we get $$\left \| \|v^k(t-t_n^k) \|^{2-\theta}_{ L_x^{\widehat{r}}} \|v^j(t-t_n^j) \|_{L^{\widehat{r}}_x}\right\|_{L_t^{\widehat{q}'}}\rightarrow 0,$$ as $n\rightarrow +\infty$, which lead to $ \left\| |x|^{-b}|v^k|^2|v^j \right \|_{L_t^{\widehat{q}'}L^{\widehat{r}'}_x}\rightarrow 0. $  Finally, we prove $\|\nabla e_n^M\|_{S'(L^2)}<\frac{\varepsilon}{3}$. Note that $$\begin{aligned} \label{ec23} \nabla e_n^M&=&\nabla (|x|^{-b})\left( f(\widetilde{u}_n)-\sum_{j=1}^M f(v^j) \right)+|x|^{-b}\nabla \left( f(\widetilde{u}_n)-\sum_{j=1}^M f(v^j)\right)\nonumber \\ &\equiv& R^1_n+R^2_n,\end{aligned}$$ where $f(v)=|v|^2 v$. First, we consider $R^1_n$. The estimate yields $$\| R^1_n\|_{S'(L^2)}\leq c\; C_{\alpha,M}\sum_{j=1}^{M}\sum_{1\leq j\neq k\leq M} \left\||x|^{-b-1}|v^k|^2|v^j|\right\|_{L^{\widehat{q}'}_tL^{\widehat{r}'}_x}$$ and by Remark \[RSglobal\] we deduce that $\left\||x|^{-b-1}|v^k|^2|v^j|\right\|_{L^{\widehat{q}'}_tL^{\widehat{r}'}_x}$ is finite, then by the same argument as before we have $$\left\||x|^{-b-1}|v^k(t-t_n^k)|^2|v^j(t-t_n^j)|\right\|_{L^{\widehat{q}'}_tL^{\widehat{r}'}_x}\rightarrow 0\;\;\textnormal{as}\;\;n\rightarrow+\infty.$$ Therefore, the last two relations yield $\|R^1_n\|_{S'(L^2)}\rightarrow 0$ as $n\rightarrow+\infty$.  On the other hand, observe that $$\begin{aligned} \label{ec231} \nabla ( f(\widetilde{u}_n)-\sum_{j=1}^M f(v^j))&=&f'(\widetilde{u}_n)\nabla \widetilde{u}_n-\sum_{j=1}^M f'(v^j)\nabla v^j\nonumber\\ &=&\sum_{j=1}^M (f'(\widetilde{u}_n)- f'(v^j))\nabla v^j.\end{aligned}$$ Since $|f'(\widetilde{u}_n)- f'(v^j)|\leq C_{\alpha,M}\sum_{1\leq k\neq j\leq M}|v^k|(|v^j|+|v^k|)$, we deduce using the last two relations together with and $$\|R_n^2\|_{S'(L^2)}\lesssim \sum_{j=1}^M\sum_{1\leq k\neq j\leq M} \left\||x|^{-b} |v^k|(|v^j|+|v^k|)|\nabla v^j|\right\|_{S'(L^2)}.$$ Therefore, from Lemma \[LG1\] (ii) (see also Remark \[RGP\]) we have that the right hand side of the last two inequalities are finite quantities and, by an analogous argument as before, we conclude that $$\|R_n^2\|_{S'(L^2)}\rightarrow 0\;\;\;\textnormal{as}\;\;\;n\rightarrow+\infty.$$ This completes the proof of Claim $1$.  [**Proof of Claim $2.$**]{} First, we show that $\|\widetilde{u}_n\|_{L^\infty_tH^1_x}$ and $\|\widetilde{u}_n\|_{L^\gamma_tL^\gamma_x}$ are bounded quantities where $\gamma=\frac{10}{3}$. Indeed, we already know (see and ) that there exists $C_0$ such that $$\sum_{j=1}^{\infty}\|\psi^j\|^2_{H^1_x}\leq C_0,$$ then we can choose $M_0\in \mathbb{N}$ large enough such that $$\label{SP} \sum_{j=M_0}^{\infty}\|\psi^j\|^2_{H^1_x}\leq \frac{\delta}{2},$$ where $\delta>0$ is a sufficiently small.\ Fix $M\geq M_0$. From , there exists $n_1(M)\in \mathbb{N}$ where for all $n> n_1(M)$, we obtain $$\sum_{j=M_0}^{M}\|\textnormal{INLS}(-t_n^j)\widetilde{\psi}^j\|^2_{H^1_x}\leq \delta,$$ where we have used . This is equivalent to $$\label{claim11} \sum_{j=M_0}^{M}\|v^j(-t_n^j)\|^2_{H^1_x}\leq \delta.$$ Therefore, by the Small Data Theory (Proposition \[GWPH1\])[^13] $$\sum_{j=M_0}^{M}\|v^j(t-t_n^j)\|^2_{L_t^{\infty}H^1_x}\leq c\delta\;\;\textnormal{for}\;n\geq n_1(M).$$ Note that, $$\left\|\sum_{j=M_0}^{M} v^j(t-t_n^j)\right\|^2_{H^1_x}=\sum_{j=M_0}^{M}\|v^j(t-t_n^j)\|^2_{H_x^1}+2\sum_{M_0\leq l \neq k\leq M}\langle v^l(t-t_n^l),v^k(t-t_n^k)\rangle_{H^1_x},$$ so, for $l\neq k$ we deduce from $\eqref{PD}$ that (see [@JIANCAZENAVE Corollary $4.4$] for more details) $$\sup_{t\in\mathbb{R}} |\langle v^l(t-t_n^l),v^k(t-t_n^k)\rangle_{H^1_x}|\rightarrow 0\;\;\textnormal{as}\;\;n\rightarrow +\infty.$$ Hence, since $\|v^j\|_{L^{\infty}_tH_x^1}$ is bounded (see - ), by definition of $\widetilde{u}_n$ there exists $S>0$ (independent of $M$) such that $$\label{claim12} \sup_{t\in\mathbb{R}}\|\widetilde{u}_n\|^2_{H^1_x}\leq S \;\,\textnormal{for}\;\;n>n_1(M).$$  We now show $\|\widetilde{u}_n\|_{L^\gamma_tL^\gamma_x}\leq L_1$. Using again with $\delta$ small enough and the Small Data Theory (noting that $(\gamma,\gamma)$ is $L^2$-admissible and $\gamma >2$), we have $$\label{claim111} \sum_{j=M_0}^{M}\|v^j(t-t_n^j)\|^{\gamma}_{L^\gamma_tL^\gamma_x}\leq c \sum_{j=M_0}^{M}\|v^j(-t_n^j)\|^{\gamma}_{H^1_x}\leq c \sum_{j=M_0}^{M}\|v^j(-t_n^j)\|^2_{H^1_x}\leq c\delta,$$ for $n\geq n_1(M)$. On the other hand, in view of $$\left\|\sum_{j=M_0}^{M} v^j(t-t_n^j)\right\|^\gamma_{L^\gamma_tL^\gamma_x}\leq \sum_{j=M_0}^{M}\|v^j\|^\gamma_{L^\gamma_tL^\gamma_x}+C_M\sum_{M_0\leq j \neq k\leq M}\int_{\mathbb{R}^{3+1}} |v^j||v^k||v^k|^{\gamma-2}$$ for all $M>M_0$. Observe that, given $j$ such that $M_0\leq j\neq k\leq M$, the Hölder inequality yields $$\begin{aligned} \label{claim112} \int_{\mathbb{R}^{3+1}} |v^j||v^k||v^k|^{\gamma-2}& \leq &\|v^k(t-t_n^k)\|_{L^\gamma_{t}L^\gamma_x}\left( \int_{\mathbb{R}^{3+1}} |v^j|^{\frac{\gamma}{2}}|v^k|^{\frac{\gamma}{2}} \right)^{\frac{2}{\gamma}} \nonumber \\ &\leq& c \|v^j(-t_n^j)\|_{H^1_x}\left( \int_{\mathbb{R}^{3+1}} |v^j|^{\frac{\gamma}{2}}|v^k|^{\frac{\gamma}{2}} \right)^{\frac{2}{\gamma}}.\end{aligned}$$ Since $v^j$ and $v^k\in L^\gamma_tL^\gamma_x$ we have that the right hand side of is bounded and so by similar arguments as in the previous claim, we deduce from that the integral in the right hand side of the previous inequality goes to $0$ as $n\rightarrow +\infty$ (another proof of this fact can be found in [@JIANCAZENAVE Lemma $4.5$]). This implies that there exists $L_1$ (independent of $M$) such that $$\label{claim113} \|\widetilde{u}_n\|_{L^\gamma_tL^\gamma_x} \leq \sum_{j=1}^{M_0}\|v^j\|_{L^\gamma_tL^\gamma_x}+\left\|\sum_{j=M_0}^{M} v^j\right\|_{L^\gamma_tL^\gamma_x}\leq L_1\;\;\;\textnormal{for}\;n\geq n_1(M),$$ where we have used .  To complete the proof of the Claim $2$ we will show the following inequality $$\label{ineq2} \|\widetilde{u}_n\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x}\leq \|\widetilde{u}_n\|^{1-\frac{\gamma}{\widehat{a}}}_{L^\infty_tH^1_x}\|\widetilde{u}_n\|^{\frac{\gamma}{\widehat{a}}}_{L^\gamma_tL^\gamma_x},$$ where $\widehat{a}$ and $\widehat{r}$ are defined in -.  Assuming the last inequality for a moment let us conclude the proof of the Claim $2$. Indeed combining and we deduce from that $$\|\widetilde{u}_n\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x}\leq S^{1-\frac{\gamma}{\widehat{a}}}L_1^{\frac{\gamma}{\widehat{a}}}=L_2,\;\;\;\textnormal{for}\;n\geq n_1(M).$$ Then, since $\widetilde{u}_n$ satisfies the perturbed equation we can apply the Strichartz estimates to the integral formulation and conclude (using also Claim $1$) $$\begin{aligned} \|\widetilde{u}_n\|_{S(\dot{H}^{s_c})}&\leq &c\|\widetilde{u}_{n,0}\|_{H^1_x}+c\left\| |x|^{-b} |\widetilde{u}_n|^2 \widetilde{u}_n \right\|_{L^{\widetilde{a}'}_tL_x^{\bar{\widehat{r}}'}}+\|e^M_n\|_{S'(\dot{H}^{-s_c})}\\ &\leq & cS+cL_2+\varepsilon = L,\end{aligned}$$ for $n\geq n_1(M)$, which completes the proof of the Claim $2$.  We now prove the inequality . Indeed, the interpolation inequality implies $$\|\widetilde{u}_n\|_{L^{\widehat{a}}_tL^{\widehat{r}}_x}\leq \|\widetilde{u}_n\|^{1-\frac{\gamma}{\widehat{a}}}_{L^\infty_tL^p_x}\|\widetilde{u}_n\|^{\frac{\gamma}{\widehat{a}}}_{L^\gamma_tL^\gamma_x},$$ where $\frac{1}{\widehat{r}}=\left(1-\frac{\gamma}{\widehat{a}}\right)\left(\frac{1}{p}\right)+\frac{1}{\widehat{a}}$. Using the values of $\widehat{r}$, $\widehat{a}$ and $\gamma$ we obtain $$p=\frac{14-6\theta+10b}{3+b-\theta(2-b)}.$$ Choosing $0<\theta <2/3$ and $b<1$ then it is easy to see that $2<p<6$. Thus by the Sobolev embedding $H^1 \hookrightarrow L^p$ the inequality holds.  In the next proposition, we prove the precompactness of the flow associated to the critical solution $u_c$. The argument is very similar to Holmer-Roudenko [@HOLROU Proposition $5.5$]. \[PSC\][**(Precompactness of the flow of the critical solution)**]{} Let $u_c$ be as in Proposition \[ECS\] and define $$K=\{u_c(t)\;:\;t\in[0,+\infty)\}\subset H^1.$$ Then $K$ is precompact in $H^1(\mathbb{R}^3)$. Let $\{t_n\}\subseteq [0,+\infty )$ a sequence of times and $\phi_n=u_c(t_n)$ be a uniformly bounded sequence in $H^1(\mathbb{R}^3)$. We need to show that $u_c(t_n)$ has a subsequence converging in $H^1(\mathbb{R}^3)$. If $\{t_n\}$ is bounded, we can assume $t_n \rightarrow t^*$ finite, so by the continuity of the solution in $H^1(\mathbb{R}^3)$ the result is clear. Next, assume that $t_n\rightarrow +\infty$.  The linear profile expansion (Proposition \[LPD\]) implies the existence of profiles $\psi^j$ and a remainder $W_n^M$ such that $$u_c(t_n)=\sum_{j=1}^{M}U(-t_n^j)\psi^j+W_n^M,$$ with $|t_n^j-t_n^k|\rightarrow +\infty$ as $n\rightarrow +\infty$ for any $j\neq k$. Then, by the energy Pythagorean expansion (Proposition \[EPE\]), we get $$\label{ECCS} \sum_{j=1}^{M}\lim_{n\rightarrow +\infty}E[U(-t_n^j)\psi^j]+\lim_{n\rightarrow +\infty} E[W_n^M]=E[u_c]=\delta_c,$$ where we have used Proposition \[ECS\] (ii). This implies that $$\lim_{n\rightarrow +\infty}E[U(-t_n^j)\psi^j]\leq \delta_c\;\;\;\;\forall\;j,$$ since each energy in is nonnegative by Lemma (i).\ Moreover, by with $s=0$ we obtain $$\label{MCCS} \sum_{j=1}^{M}M[\psi^j]+\lim_{n\rightarrow +\infty}M[W_n^M]=M[u_c]=1,$$ by Proposition \[ECS\] (i).  If more than one $\psi^j \neq 0$, similar to the proof in Proposition \[ECS\], we have a contradiction with the fact that $\|u_c\|_{S(\dot{H}^{s_c})}=+\infty$. Thus, we address the case that only $\psi^j=0$ for all $j\geq 2$, and so $$\label{CCS1} u_c(t_n)=U(-t_n^1)\psi^1+W_n^M.$$ Also as in the proof of Proposition \[ECS\], we obtain that $$\label{MECS} M[\psi^1] =M[u_c]=1\;\;\;\textnormal{and}\;\;\;\lim_{n\rightarrow +\infty}E[U(-t_n^1)\psi^1]=\delta_c,$$ and using , together with , we deduce that $$\lim_{n\rightarrow +\infty}M[W_n^M]=0\;\;\;\textnormal{and}\;\;\;\lim_{n\rightarrow +\infty}E[W_n^M]=0.$$ Thus, Lemma \[LGS\] (i) yields $$\label{ERCS} \lim_{n\rightarrow +\infty}\|W_n^M\|_{H^1}=0.$$  We claim now that $t^1_n$ converges to some finite $t^*$ (up to a subsequence). In this case, since $U(-t_n^1)\psi^1\rightarrow U(-t^*)\psi^1$ in $H^1(\mathbb{R}^3)$ and $\eqref{ERCS}$ holds, the relation implies that $u_c(t_n)$ converges in $H^1(\mathbb{R}^3)$, concluding the proof.  Assume by contradiction that $|t^1_n|\rightarrow +\infty$, then we have two cases to consider. If $t^1_n\rightarrow -\infty$, by $$\|U(t)u_c(t_n)\|_{S(\dot{H}^{s_c};[0,+\infty))}\leq\|U(t-t_n^1)\psi^1\|_{S(\dot{H}^{s_c};[0,+\infty))}+\|U(t)W_n^M\|_{S(\dot{H}^{s_c};[0,+\infty))}.$$ Next, note that since $t^1_n\rightarrow -\infty$ we obtain $$\label{PPCS1} \|U(t-t_n^1)\psi^1\|_{S(\dot{H}^{s_c};[0,+\infty))}\leq \|U(t)\psi^1\|_{S(\dot{H}^{s_c};[-t_n^j,+\infty))}\leq \frac{1}{2}\delta,$$ and also $$\label{PPCS2} \|U(t)W_n^M\|_{S(\dot{H}^{s_c})}\leq \frac{1}{2}\delta,$$ given $\delta>0$ for $n, M$ sufficiently large, where in the last inequality we have used and . Hence, $$\|U(t)u_c(t_n)\|_{S(\dot{H}^{s_c};[0,+\infty))}\leq \delta.$$ Therefore, choosing $\delta>0$ sufficiently small, by the small data theory (Proposition \[GWPH1\]) we get that $ \|u_c\|_{S(\dot{H}^{s_c})}\leq 2\delta,$ which is a contradiction with Proposition \[ECS\] (iv).  On the other hand, if $t^1_n\rightarrow +\infty$, the same arguments also give that for $n$ large, $$\|U(t)u_c(t_n)\|_{S(\dot{H}^{s_c};(-\infty,0])}\leq \delta,$$ and again the small data theory (Proposition \[GWPH1\]) implies $ \|u_c\|_{S(\dot{H}^{s_c};(-\infty,t_n])}\leq 2\delta. $ Since $t_n\rightarrow +\infty$ as $n\rightarrow +\infty$, from the last inequality we get $\|u_c\|_{S(\dot{H}^{s_c})}\leq 2\delta$, which is also a contradiction. Thus, $t_n^1$ must converge to some finite $t^*$, completing the proof of Proposition \[PSC\]. Rigidity theorem ================  The main result of this section is a rigidity theorem, which implies that the critical solution $u_c$ constructed in Section \[CCS\] must be identically zero and so reaching a contradiction in view of Proposition \[ECS\] (iv). Before proving this result, we begin showing some preliminaries that will help us in the proof. \[PFIUL\][**(Precompactness of the flow implies uniform localization)**]{} Let $u$ be a solution of such that $$K=\{u(t)\;:\;t\in[0,+\infty)\}$$ is precompact in $H^1(\mathbb{R}^3)$. Then for each $\varepsilon>0$, there exists $R>0$ so that $$\label{PFIUL1} \int_{|x|>R}|\nabla u(t,x)|^2dx\leq \varepsilon,\;\textnormal{for all}\;0\leq t<+\infty.$$ The proof follows the same steps as in Holmer-Roudenko [@HOLROU Lemma $5.6$]. So we omit the details We will also need the following local virial identity. \[VI\] [**(Virial identity)**]{} Let $\phi\in C^\infty_0(\mathbb{R}^3)$, $\phi \geq 0$ and $T>0$. For $R>0$ and $t\in [0,T]$ define $$\label{DFZ}z_R(t)=\int_{\mathbb{R}^3} R^2 \phi\left(\frac{x}{R}\right)|u(t,x)|^2dx,$$ where $u$ is a solution of . Then we have $$\label{FD}z'_R(t)=2R Im\int_{\mathbb{R}^3} \nabla\phi\left(\frac{x}{R}\right)\cdot\nabla u \bar{u}dx$$ and $$\begin{aligned} \label{SD} z''_R(t)&= 4\sum_{j,k} Re\int \frac{\partial u}{\partial_{x_k}} \frac{\partial \bar{u}}{\partial_{x_j}}\frac{\partial^2\phi}{\partial x_k\partial x_j}\left(\frac{x}{R}\right)dx-\frac{1}{R^2}\int |u|^2\Delta^2 \phi\left(\frac{x}{R}\right) dx \nonumber \\ & -\int |x|^{-b}|u|^{4}\Delta \phi\left(\frac{x}{R}\right) dx+R\int \nabla (|x|^{-b})\cdot\nabla \phi\left(\frac{x}{R}\right) |u|^{4}dx. \end{aligned}$$ We first compute $z'_R$. Note that $$\partial_t|u|^2=2Re( u_t\bar{u})=2Im(iu_t\bar{u}).$$ Since $u$ satisfies and using integration by parts, we have $$\begin{aligned} z'_R(t)&=&2Im\int R^2\phi\left(\frac{x}{R}\right)iu_t \bar{u}dx\\ &=&-2Im\int R^2\phi\left(\frac{x}{R}\right)\left(\Delta u\bar{u}+|x|^{-b}|u|^{4}\right)dx\\ &=& -2Im\int R^2\phi\left(\frac{x}{R}\right)\nabla \cdot (\nabla u \bar{u})dx\\ &=&2RIm\int \nabla \phi\left(\frac{x}{R}\right)\cdot\nabla u \bar{u}dx.\end{aligned}$$ On the other hand, using again integration by parts and the fact that $z-\bar{z}=2iIm z$, we obtain $$\begin{aligned} z''_R(t)&=& 2RIm\int \nabla \phi\left(\frac{x}{R}\right)\cdot \left( \bar{u}_t\nabla u +\bar{u}\nabla u_t\right)dx\\ &=& 2RIm\left\{\sum_{j}\int \bar{u}_t\partial_{x_j}u\partial_{x_j}\phi\left(\frac{x}{R}\right)dx- u_t\partial_{x_j}\left(\bar{u}\partial_{x_j}\phi\left(\frac{x}{R}\right)\right) dx \right\}\\ &=& 2RIm\left\{\sum_{j}2i Im \int \bar{u}_t\partial_{x_j}u\partial_{x_j}\phi\left(\frac{x}{R}\right) dx-\int\frac{1}{R}u_t \bar{u}\partial^2_{x_j}\phi\left(\frac{x}{R}\right) dx \right\}\\ &=&4R I_1+2I_2,\end{aligned}$$ where $$I_1=Im \sum_{j}\int \bar{u}_t\partial_{x_j}u\partial_{x_j}\phi\left(\frac{x}{R}\right)\;\;\textnormal{and}\;\;I_2=-Im \sum_{j}\int u_t\bar{u} \partial^2_{x_j}\phi\left(\frac{x}{R}\right) dx.$$ We start considering $I_2$. Since $u$ is a solution of we get $$\begin{aligned} I_2&=&-Im\left\{\sum_{j,k} \int i \partial^2_{x_k}u \bar{u}\partial^2_{x_j} \phi\left(\frac{x}{R}\right) dx \right\}-\sum_{j}\int |x|^{-b}|u|^{4}\partial^2_{x_j}\phi\left(\frac{x}{R}\right) dx\\ &= & Im \left\{\sum_{j,k} \int i \left( |\partial_{x_k}u|^2 \partial^2_{x_j} \phi\left(\frac{x}{R}\right)+ \frac{1}{R}\partial_{x_k}u\bar{u}\frac{\partial^3\phi}{\partial x_k \partial x^2_j}\left(\frac{x}{R}\right) \right)dx\right\} \\ & & - \int |x|^{-b}|u|^{4}\Delta \phi\left(\frac{x}{R}\right) dx\\ &=&\int \left(|\nabla u|^2-|x|^{-b}|u|^{4}\right) \Delta \phi\left(\frac{x}{R}\right) dx+\frac{1}{R} \sum_{j,k} Re\int \partial_{x_k}u\bar{u} \frac{\partial^3\phi}{\partial x_k \partial x^2_j}\left(\frac{x}{R}\right)dx, \end{aligned}$$ where we have used integration by parts and the fact that $Im (iz)=Re (z)$. Furthermore, since $\partial_{x_k}|u|^2=2 Re \left(\partial_{x_k}u\bar{u}\right)$ another integration by parts yields $$\begin{aligned} \label{VI1} I_2=&\int \left(|\nabla u|^2-|x|^{-b}|u|^{4}\right) \Delta \phi\left(\frac{x}{R}\right) dx-\frac{1}{2R^2}\sum_{j,k}\int |u|^2\frac{\partial^4\phi}{\partial x^2_k\partial x^2_j}\left(\frac{x}{R}\right)dx \nonumber \\ =&\int \left(|\nabla u|^2-|x|^{-b}|u|^{4}\right) \Delta \phi\left(\frac{x}{R}\right) dx-\frac{1}{2R^2}\int |u|^2\Delta^2\phi\left(\frac{x}{R}\right) dx.\end{aligned}$$ Next, we deduce using the equation and $Im (z)=-Im (\bar{z})$ that $$\begin{aligned} I_1&=-Im \sum_{j}u_t \partial_{x_j}\bar{u}\partial_{x_j} \phi\left(\frac{x}{R}\right) dx\\ &=-Im i\sum_{j}\left\{\int \left( \Delta u + |x|^{-b}|u|^{2}u \right)\partial_{x_j}\bar{u}\partial_{x_j} \phi\left(\frac{x}{R}\right) dx \right\}\\ &= -Re \sum_{j,k}\int \partial^2_{x_k} u\partial_{x_j} \bar{u} \partial_{x_j} \phi\left(\frac{x}{R}\right) dx-\sum_j \int |x|^{-b}\partial_{x_j} \phi\left(\frac{x}{R}\right)|u|^{2} Re (\partial_{x_j}\bar{u} u) dx\\ &=-Re \sum_{j,k}\int \partial^2_{x_k} u\partial_{x_j} \bar{u} \partial_{x_j} \phi\left(\frac{x}{R}\right)dx-\frac{1}{4}\sum_j \int |x|^{-b}\partial_{x_j}\phi\left(\frac{x}{R}\right)\partial_{x_j}(|u|^{4}) dx\\ &\equiv A+B,\end{aligned}$$ where we have used $Im (iz)=Re (z)$ and $\partial_{x_j}(|u|^{4})=4|u|^2 Re (\partial_{x_j}\bar{u} u) $. Moreover, since $\partial_{x_j}|\partial_{x_k}u|^2=2Re \left(\partial_{x_k}u \frac{\partial^2\bar{u}}{ \partial x_k \partial x_j}\right) $ and using integration by parts twice, we get $$\begin{aligned} A&=& Re \sum_{j,k} \left\{ \int \left(\partial_{x_j}\phi\left(\frac{x}{R}\right) \partial_{x_k}u \frac{\partial^2\bar{u}}{ \partial x_k \partial x_j}+ \frac{1}{R} \partial_{x_k}u \partial_{x_j}\bar{u} \frac{ \partial^2\phi}{\partial x_j\partial x_k}\left(\frac{x}{R}\right) \right)dx \right\}\\ &=&-\sum_{j,k}\frac{1}{2R} \int |\partial_{x_k}u|^2\partial^2_{x_j}\phi\left(\frac{x}{R}\right) dx+\frac{1}{R} \sum_{i,j} Re \int \partial_{x_k}u \partial_{x_j}\bar{u} \frac{ \partial^2\phi}{\partial x_j\partial x_k}\left(\frac{x}{R}\right) dx \\ &=&-\frac{1}{2R} \int |\nabla u|^2 \Delta \phi\left(\frac{x}{R}\right) dx+\frac{1}{R} \sum_{i,j} Re \int \partial_{x_k}u \partial_{x_j}\bar{u} \frac{ \partial^2\phi}{\partial x_j\partial x_k}\left(\frac{x}{R}\right) dx.\end{aligned}$$ Similarly, integrating by parts $$\begin{aligned} B&=\frac{1}{4} \sum_j \left( \int \partial_{x_j}\phi\left(\frac{x}{R}\right)\partial_{x_j}(|x|^{-b}) |u|^{4} dx+ \frac{1}{R} \int \partial^2_{x_j}\phi\left(\frac{x}{R}\right)|x|^{-b} |u|^{4} dx \right) \\ &=\frac{1}{4} \int \nabla \phi\left(\frac{x}{R}\right)\cdot \nabla (|x|^{-b}) |u|^{4} dx+ \frac{1}{4R} \int \Delta \phi\left(\frac{x}{R}\right)|x|^{-b} |u|^{4} dx. \\\end{aligned}$$ Therefore, $$\begin{aligned} \label{VI2} I_1&=-\frac{1}{2R} \int |\nabla u|^2 \Delta \phi\left(\frac{x}{R}\right) dx+\frac{1}{R} \sum_{i,j}Re \int \partial_{x_k}u \partial_{x_j}\bar{u} \frac{ \partial^2\phi}{\partial x_j\partial x_k}\left(\frac{x}{R}\right) dx\nonumber \\ & +\frac{1}{4} \int \nabla \phi\left(\frac{x}{R}\right)\cdot \nabla (|x|^{-b}) |u|^{4} dx+ \frac{1}{4R} \int \Delta \phi\left(\frac{x}{R}\right)|x|^{-b} |u|^{4} dx.\end{aligned}$$ Finally it is easy to check that combining and we obatin , which complete the proof. Finally, we apply the previous results to proof the rigidity theorem. \[RT\][**(Rigidity)**]{} Let $u_0\in H^1(\mathbb{R}^3)$ satisfying $$E[u_0]^{s_c}M[u_0]^{1-s_c} <E[Q]^{s_c}M[Q]^{1-s_c}$$ and $$\| \nabla u_{0} \|_{L^2}^{s_c} \|u_{0}\|_{L^2}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c}.$$ If the global $H^1(\mathbb{R}^3)$-solution $u$ with initial data $u_0$ satisfies $$K=\{u(t)\;:\;t\in[0,+\infty)\}\; \textnormal{is precompact in}\; H^1(\mathbb{R}^3)$$ then $u_0$ must vanishes, i.e., $u_0=0$. By Theorem \[TG\] we have that $u$ is global in $H^1(\mathbb{R}^3)$ and $$\label{TR1} \| \nabla u(t) \|_{L^2_x}^{s_c} \|u(t)\|_{L^2_x}^{1-s_c}<\|\nabla Q \|_{L^2}^{s_c} \|Q\|_{L^2}^{1-s_c}.$$  On the other hand, let $\phi \in C_0^\infty$ be radial, with $$\phi(x)=\left\{\begin{array}{cl} |x|^2&\textnormal{for}\;|x|\leq 1\\ 0&\textnormal{for}\;|x|\geq 2. \end{array}\right.$$ Then, using , the Hölder inequality and we obtain $$\begin{aligned} |z'_R(t)| &\leq & cR\int_{|x|<2R}|\nabla u(t)||u(t)|dx\leq cR\|\nabla u(t)\|_{L^2}\|u(t)\|_{L^2}\lesssim cR.\end{aligned}$$ Hence, $$\begin{aligned} \label{FDI}|z'_R(t)-z'_R(0)|\leq |z'_R(t)|+|z'_R(0)|\leq 2cR,\;\;\textnormal{for all }\;t>0. $$  The idea now is to obtain a lower bound for $z''_R(t)$ strictly greater than zero and reach a contradiction. Indeed, from the local virial identity $$\begin{aligned} \label{SDz} z''_R(t)&= 4\sum_{j,k} Re\int \partial_{x_k} u \partial_{x_j}\bar{u}\frac{\partial^2\phi}{\partial x_k\partial x_j}\left(\frac{x}{R}\right)dx-\frac{1}{R^2}\int |u|^2\Delta^2 \phi\left(\frac{x}{R}\right) dx \nonumber \\ & -\int |x|^{-b}|u|^{4}\Delta \phi\left(\frac{x}{R}\right) dx+R\int \nabla (|x|^{-b})\cdot\nabla \phi\left(\frac{x}{R}\right) |u|^{4}dx \nonumber \\ & = 8 \| \nabla u \|^2_{L^2_x}- 2(3+b) \left\||x|^{-b}|u|^{4}\right\|_{L^1_x}+R(u(t)), \end{aligned}$$ where $$\begin{aligned} R(u(t))&= 4 \sum\limits_{j}Re\int \left( \partial^2_{x_j} \phi\left(\frac{x}{R}\right)-2\right )|\partial_{x_j}u|^2 + 4 \sum\limits_{j\neq k}Re\int \frac{\partial^2\phi}{\partial x_k\partial x_j}\left(\frac{x}{R}\right)\partial_{x_k}u\partial_{x_j}\bar{u}\\ &- \frac{1}{R^2}\int |u|^2\Delta^2\phi\left(\frac{x}{R}\right)+R\int \nabla(|x|^{-b})\cdot\nabla \phi\left(\frac{x}{R}\right)|u|^{4}\\ &+\int \left( -\left( \Delta\phi\left(\frac{x}{R}\right)-6 \right)+2b\right) |x|^{-b}|u|^{4}. \end{aligned}$$ Since $\phi(x)$ is radial and $\phi(x)=|x|^2$ if $|x|\leq 1$, the sum of all terms in the definition of $R(u(t))$ integrating over $|x|\leq R$ is zero. Indeed, for the first three terms this is clear by the definition of $\phi(x)$. In the fourth term we have $$2\int_{|x|\leq R} \nabla(|x|^{-b})\cdot x|u|^4dx=2\int_{|x|\leq R} -b|x|^{-b}|u|^4dx,$$ and adding the last term (also integrating over $|x|\leq R$) we get zero since $\Delta \phi =6$, if $|x|\leq R$. Therefore, for the integration on the region $|x|> R$, we have the following bound $$\begin{aligned} |R(u(t))|&\leq& c\int_{|x|>R} \left(|\nabla u(t)|^2+\frac{1}{R^2}|u(t)|^2+|x|^{-b}|u(t)|^{4}\right)dx\end{aligned}$$ $$\begin{aligned} \label{RESTO} &\leq & c\int_{|x|>R} \left(|\nabla u(t)|^2+\frac{1}{R^2}|u(t)|^2+\frac{1}{R^b}|u(t)|^{4}\right)dx,\end{aligned}$$ where we have used that all derivatives of $\phi$ are bounded and $|R\partial_{x_j}(|x|^{-b})|\leq c|x|^{-b}$ if $|x|>R$.  Next we use that $K$ is precompact in $H^1(\mathbb{R}^3)$. By Proposition \[PFIUL\], given $\varepsilon>0$ there exists $R_1>0$ such that $\int_{|x|>R_1} |\nabla u(t)|^2\leq\varepsilon$. Furthermore, by Mass conservation , there exists $R_2>0$ such that $\frac{1}{R^2_2}\int_{|x|>R_2} | u(t)|^2\leq \varepsilon$. Finally, by the Sobolev embedding $H^1\hookrightarrow L^{4}$, there exists $R_3$ such that $\frac{1}{R_3^b}\int_{|x|>R_3} |u(t)|^{4}\leq c\varepsilon$ (recall that $\|u(t)\|_{H^1_x}$ is uniformly bounded for all $t>0$ by and Mass conservation ). Taking $R=\max\{R_1,R_2,R_3\}$ the inequality implies $$\label{RESTO1} |R(u(t))|\leq c\int_{|x|>R} \left(|\nabla u(t)|^2+\frac{1}{R^2}|u(t)|^2+\frac{1}{R^b}|u(t)|^{4}\right)dx\leq c\varepsilon.$$ On the other hand, Lemma \[LGS\] (iii), and yield $$\label{RESTO2} z''_R(t)\geq 16A E[u]-|R(u(t))|\geq 16AE[u]-c\varepsilon,$$ where $A=1-w$ and $w=\frac{E[v]^{s_c}M[v]^{1-s_c}}{E[Q]^{s_c}M[Q]^{1-s_c}}$.\ Now, choosing $\varepsilon=\frac{8A}{c}E[u]$, with $c$ as in we have $$z''_R(t)\geq 8AE[u].$$ Thus, integrating the last inequality from $0$ to $t$ we deduce $$\label{SDI} z'_R(t)-z'_R(0)\geq 8AE[u]t.$$ Now sending $t\rightarrow \infty$ the left hand of also goes to $+\infty$, however from it must be bounded. Therefore, we have a contradiction unless $E[u]=0$ which implies $u\equiv 0$ by Lemma \[LGS\] (i). **Acknowledgment**   L.G.F. was partially supported by CNPq and FAPEMIG/Brazil. C.G. was partially supported by Capes/Brazil. The authors also thank Svetlana Roudenko for fruitful conversations concerning this work.\ [^1]: Notice that, in this case the critical Sobolev index is $s_c=1/2$. [^2]: It worth mentioning that, the pair $\left(\infty,\frac{6}{3-2s}\right)$ also satisfies the relation , however, in our work we will not make use of this pair when we estimate the nonlinearity $|x|^{-b} |u|^2 u$. [^3]: We can rescale $u_{n,0}$ such that $\|u_{n,0}\|_{L^2} = 1$. Indeed, if $u^\lambda_{n,0}(x)=\lambda^{\frac{2-b}{2}}u_{n,0}(\lambda x)$ then by we have $E[u^\lambda_{n,0}]^{s_c}M[u^\lambda_{n,0}]^{1-s_c}<E[Q]^{s_c}M[Q]^{1-s_c}$ and $\|\nabla u^\lambda_{n,0}\|^{s_c}_{L^2}\| u^\lambda_{n,0}\|^{1-s_c}_{L^2}<\|\nabla Q\|^{s_c}_{L^2}\| Q\|^{1-s_c}_{L^2} $. Moreover, since $\|u^\lambda_{n,0}\|_{L^2} = \lambda^{-s_c}\|u_{n,0}\|_{L^2}$ by , setting $\lambda^{s_c}=\|u_{n,0}\|_{L^2}$ we have $\|u^\lambda_{n,0}\|_{L^2} = 1$. [^4]: Recall that $(\widehat{a},\widehat{r})$ is $\dot{H}^{s_c}$-admissible and $(\widetilde{a},\widehat{r})$ is $\dot{H}^{-s_c}$-admissible. [^5]: Note that $\frac{6}{2-b}=\frac{6}{3-2s_c}<\bar{r}<6$ (condition of $H^s$-admissible pair ). Indeed, it is easy to check that $\bar{r}>\frac{6}{2-b}$. On the other hand, $\bar{r}<6\Leftrightarrow \theta(2-2b)<1-2b$, which is true by the assumption $b<1/2$ and $\theta>0$ is a small number. Moreover it is easy to see that $2<r<6$, i.e., $r$ satisfies the condition of admissible pair . [^6]: $\varepsilon_0$ is given by the previous result and $\varepsilon$ to be determined later. [^7]: Note that $(\frac{2 s_c}{3+b})^{\frac{s_c}{2}}<1$. [^8]: Note that is possible not true using the norm $L^{\infty}_{I_T}L^{\frac{6}{3-2s_c}}_x$ and for this reason we remove the pair $\left(\infty,\frac{6}{3-2s_c}\right)$ in the definition of $\dot{H}^{s}$-admissible pair. [^9]: Observe that $\| w+U(t)\phi \|_{S(\dot{H}^{s_c};I_T)}\leq \| w \|_{S(\dot{H}^{s_c};I_T)}+\| U(t)\phi \|_{S(\dot{H}^{s_c};I_T)} \rightarrow 0$ as $T\rightarrow +\infty$ by and $\| w+U(t)\phi\|^\theta_{L^\infty_{T}H^1_x}, \| w+U(t)\phi \|_{S(L^2;I_T)}, \| \nabla(w+U(t)\phi) \|_{S(L^2;I_T)} <\infty$ since $w\in B(T,\rho)$ and $\phi \in H^1(\mathbb{R}^3)$. [^10]: Recalling that $s_c=\frac{1+b}{2}$. [^11]: Note that $3-2s_c^2>0$ since $s_c=\frac{1+b}{2}<1$. [^12]: We can apply Remark \[RLPD\] since $r$ and $4 \in (2,6)$. [^13]: Recall that the pair $(\infty,2)$ is $L^2$-admissible.

ArXiv

--- abstract: 'We address Standard Model predictions for flavor-changing radiative transitions of the pseudoscalar charm mesons. Short-distance contributions in $D$ radiative transitions are contrasted with those in $B$ decays. A full analysis is presented of the $c\to u+\gamma$ electromagnetic penguin amplitude with QCD radiative corrections included. Given the importance of long-range effects for the charm sector, special attention is paid to such contributions as the vector dominance and pole amplitudes. A number of two-body final states in exclusive charm radiative decays is considered and the corresponding branching ratio predictions are given.' --- \ 2.5cm [**Radiative Weak Decays of Charm Mesons**]{}\ 1.6cm [Gustavo Burdman$^{(a)}$, Eugene Golowich$^{(b)}$]{}\ [JoAnne L. Hewett$^{(c)}$ and Sandip Pakvasa$^{(d)}$]{}\ .3cm $^{(a)}$Fermilab\ Batavia, IL 60510, USA\ $^{(b)}$Department of Physics and Astronomy\ University of Massachusetts, Amherst MA 01003, USA\ $^{(c)}$Stanford Linear Accelerator Center\ Stanford CA 94309, USA\ $^{(d)}$Department of Physics and Astronomy\ University of Hawaii, Honolulu HI 96822, USA\ .3cm 2truecm **Introduction** ================ Important milestones in the study of the $b$-quark system were reached with the recent observations of both the exclusive decay $B \to K^* \gamma$$^{\cite{cleo1}}$ \_[B K\^\* ]{} = (4.5 1.5 0.9 )10\^[-5]{}  , \[excl\] and of the inclusive transition$^{\cite{cleoinc}}$ \_[b s ]{} = (2.32 0.57 0.35 )10\^[-4]{}   . \[incl\] To first approximation, these flavor-changing radiative decays can be interpreted at the quark level in terms of the $b \to s \gamma$ transition. The Standard Model allows for such a process by means of a one-loop penguin-type amplitude. Within errors, agreement of the measured branching ratios and Standard Model predictions appears to be reasonable. The small magnitudes of these branching ratios indicate just how sensitive experimental probes of $b$-quark hadrons have become. However, an outstanding question is the size of the long distance contributions[@gp; @soni] to such radiative $B$ decays relative to the short distance penguin amplitude. This issue must be addressed in order to establish the viability of determining the value of the ratio of CKM matrix elements, $|V_{td}|/|V_{ts}|$, from a measurement of the ratio of exclusive branching fractions $B_{B\to\rho\gamma}/B_{B\to K^*\gamma}$. Important as they are, the above measurements by no means exhaust the set of interesting problems. It has become increasingly evident that the database for charm hadrons is also in a state of rapid expansion, and that physically important levels of sensitivity are being achieved. Perhaps the most impressive example of this to-date is the recent observation of the nonleptonic decay $D^0 \to K^+ \pi^-$, with branching ratio$^{\cite{cleo2}}$ = 0.0077 0.0025 0.0025   . \[dcs\] This transition has been interpreted as evidence of a doubly Cabibbo suppressed transition rather than of $D^0$-${\bar D}^0$ mixing. The discussion in this paper will be directed towards a somewhat different aspect of charm physics, the flavor-changing radiative decays. These transitions require the joint occurrence of weak and electromagnetic interactions. From Table 1$^{\cite{pdg},\cite{selen},\cite{fte}}$, we see that no such events (involving emission of real or virtual photons) have yet been observed. However, these decays are an active area of study, and data gathered in ongoing fixed-target experiments are establishing markedly improved bounds. Our objective in the analysis to follow will be to provide up-to-date predictions for flavor-changing radiative transitions of charm systems. Since the experimental situation for charm mesons is at present more favorable than for charm baryons, we shall restrict our attention to the former. Even with this restriction, it is a tall order to supply accurate theoretical values. It has become evident over a long period of time that theoretical calculations of $D$-meson weak decays are not particularly trustworthy, due in part to the absence of a rapidly convergent approximation scheme and also to the presence of significant hadron dynamical effects in the $D$ meson mass region. Despite this, we feel that one can make some definite statements, such as the relative importance of long-range and short-range effects and of the various types of final states which can reasonably be anticipated. We shall base our analysis on a variety of theoretical techniques, from operator-product expansion and renormalization-group methods to more phenomenological approaches like vector-meson-dominance (VMD). Measurement of radiative charm decays would probe the long distance contributions and thus provide further insight in the extrapolation of calculational techniques to the $B$ sector. [lc]{}\ Mode & Branching Ratio\ $D^0 \to \rho^0 \gamma$ & $<1.4\times 10^{-4}$\ $D^0 \to \phi^0 \gamma$ & $<2.0\times 10^{-4}$\ $D^0 \to {\bar K}^{0} e^+ e^-$ & $<1.7\times 10^{-3}$\ $D^0 \to \rho^0 e^+ e^-$ & $< 4.5 \times 10^{-4}$\ $D^0 \to \rho^0 \mu^+ \mu^-$ & $< 8.1 \times 10^{-4}$\ $D^+ \to \pi^+ e^+ e^-$ & $< 2.5 \times 10^{-3}$\ $D^+ \to \pi^+ \mu^+ \mu^-$ & $< 2.9 \times 10^{-3}$\ $D^+ \to K^+ e^+ e^-$ & $< 4.8 \times 10^{-3}$\ $D^+ \to K^+ \mu^+ \mu^-$ & $< 9.2 \times 10^{-3}$\ Let us summarize the contents to follow. In Section 2, we consider the short-range component in radiative charm decays, primarily the charm counterpart of the penguin amplitude which dominates the radiative $B$-meson decays. In addition to addressing $c$-quark physics, our analysis contains a purely theoretical advance by removing an unnecessary assumption made in earlier studies involving $b$-quark applications. Section 3 begins our analysis of the so-called ‘long-distance’ contributions with an analysis of pole diagrams, which are induced by the weak mixing of pseudoscalar and/or vector charm mesons with noncharm states. In Section 4, we continue our study of long distance effects by turning our attention to a study of VMD amplitudes. Our conclusions and recommendations for future studies are given in Section 5. There is also an Appendix in which the applicability of VMD to certain light-meson decays is commented on. **Short Distance Contributions** ================================ Examples of diagrams which mediate the short-distance transition amplitudes for radiative charm decay are depicted in Fig. 1. They have in common that the photon emission occurs in a region of spacetime determined by the propagator of the $W$-boson. In view of the large $W$-mass $M_{\rm W}$, this region has a very limited extent compared to the length scale of the strong interactions, hence the name ‘short-distance’. Looking ahead, our conclusion regarding such short-distance amplitudes will be that in radiative decays of charm mesons they are small relative to long-distance effects, even though they receive large enhancements from QCD corrections. As described earlier, this is of course in stark contrast to $B$ decay. [c]{}\ \ \ \ \ \ \ \ [Figure 1. Short-distance Effects]{}\ In recognition of the importance attached to the electromagnetic penguin transition in radiative $B$ decays, we give a brief pedagogical comparison between the role played by this effect for $B$ and for $D$ decay. The two transitions in question are given at the quark-level by b ([**p**]{},) && s ([**p**]{}’,’) + ([**q**]{},)\ c ([**p**]{},) && u ([**p**]{}’,’) + ([**q**]{},) \[cu\]   . To highlight the crucial role played by the quark masses and CKM matrix elements, let us at first ignore the effect of QCD radiative corrections. The relevant Feynman diagrams are then depicted in Figs. 1(a),1(c) and the penguin amplitude for the transition of a heavy quark $Q$ to a much lighter quark $q$ and an on-shell photon is given by$^{\cite{il}}$ \[il1\]\ & & [eG\_[F]{}4\^2]{} \_i \_i F\_2(x\_i)  [|u]{}\_q (p’,’ )\^ (q,) \_q\^u\_Q (p ,)  , where $P_{R(L)}$ are the right(left)-handed helicity projection operators, $x_i \equiv {m_i^2 / M_{\rm W}^2}$, $\lambda_i\equiv {V_{\rm is}}^*V_{\rm ib}$ for ${\cal A}_{b\to s\gamma}$ and $\lambda_i\equiv {V_{\rm ci}}^*V_{\rm ui}$ for ${\cal A}_{c\to u\gamma}$. The function $F_2$ gives the contribution of each internal quark to the electromagnetic penguin loop,\ & & Q + [2x\^3+5x\^2-x4(x-1)\^3]{} -[3x\^3x2(x-1)\^4]{} , \[il2\] with $Q$ being the charge of the internal quark. For $b\to s \gamma$ the sum is carried out over the quarks $u,c,t$ and the term proportional to the $s$-quark mass in Eq. (\[il1\]) is generally neglected, whereas for $c \to u \gamma$, one sums over the quarks $d , s , b$ and ignores the corresponding term proportional to the $u$-quark mass. Let us get acquainted with some of the numerical values. In Table 2, we first display the magnitude of the function $F_2$ and then fold in the CKM dependence for the $b\to s \gamma$ transition (we take $m_u=5$ MeV, $m_c=1.5$ GeV, $m_t=174$ GeV, and the central values of the CKM matrix elements as given in Ref. [@pdg]). [ccc]{}\ Quark & $F_2$ & $|V_{\rm ib}^{\phantom{x}} V_{\rm is}^{~*}| F_2$\ $u$ & $2.27\times 10^{-9}$ & $1.29\times 10^{-12}$\ $c$ & $2.03\times 10^{-4}$ & $7.34\times 10^{-6}$\ $t$ & $0.39$ & $1.56\times 10^{-2}$\ Dominance of the $t$-quark intermediate state is evident, even upon including the CKM factors. Its effect is so large that the other intermediate states are numerically negligible and hence are typically omitted. The corresponding situation is given for $c \to u \gamma$ in Table 3 (with $m_d=11$ MeV, $m_s=150$ MeV, and $m_b=4.9$ GeV). [ccc]{}\ Quark & $F_2$ & $|V_{\rm ci}^{~*} V_{\rm ui}| F_2$\ $d$ & $1.57\times 10^{-9}$ & $3.36\times 10^{-10}$\ $s$ & $2.92\times 10^{-7}$ & $6.26\times 10^{-8}$\ $b$ & $3.31\times 10^{-4}$ & $3.17\times 10^{-8}$\ The amplitude for $c\to u\gamma$ differs from that of $b\to s\gamma$ in two important respects, (i) there is no single intermediate state which dominates, and (ii) the overall magnitude is much smaller. Neglecting the final state fermion mass, the QCD uncorrected decay rate $\Gamma^{(0)}_{Q\to q\gamma}$ is given by \^[(0)]{}\_[Qq]{} = [G\_F\^2128 \^4]{}m\_Q\^5 | \_i \_i F\_2(x\_i)|\^2 . \[il4\] To obtain the branching fraction, the inclusive rate is scaled to that of the semi-leptonic decay $Q\to q'\ell\nu$. This procedure removes uncertainties in the calculation due to the overall factor of $m_Q^5$ which appears in both expressions, and reduces the ambiguities involved with the imprecisely determined CKM factors. Taking the above numerical values for the internal quark masses, and using the values of the semi-leptonic branching ratios as given in Ref. [@pdg], this yields \_[bs]{} & = & [32]{} \_[BX]{} ,\ & = & 1.2910\^[-4]{} ,\ [B]{}\_[cu]{} & = & [32]{} \_[D\^+X\^+]{} ,\ & = & 1.3910\^[-17]{} . \[br0\] Here, the function $g(x)$ is the usual phase space factor in semi-leptonic meson decay, where constituent values of the final-state quark masses have been used.$^{\cite{cab}}$ The QCD uncorrected $c \to u\gamma$ transition is seen to have an unobservably small branching fraction. We next examine the impact of the QCD radiative corrections on the above branching ratios. We begin by reviewing the calculation for the $b\to s \gamma$ transition, which will serve as the foundation of our subsequent discussion of $c\to u\gamma$. The QCD corrections are calculated$^{\cite{{qcd},{gsw}}}$ via an operator product expansion based on the effective hamiltonian H\^[|b|=1]{}\_[eff]{}=-[4G\_F2]{}\_t \_[k=1]{}\^8c\_k()O\_k() , \[effh\] where the $\{ O_k\}$ are a complete set of renormalized dimension-six operators involving light fields which govern the $b\to s$ transitions. They consist of two current-current operators $O_{1,2}$, four strong penguin operators $O_{3-6}$, and electro- and chromo-magnetic dipole operators $O_{7}$ and $O_8$, O\_[1]{} & = & (|c\_\_P\_L b\_)(|s\_\^P\_L c\_) ,\ O\_[2]{} & = & (|c\_\_P\_L b\_)(|s\_\^P\_L c\_) ,\ O\_3 & = & (|s\_\_P\_L b\_)\_q(|q\_\^P\_L q\_) ,\ O\_4 & = & (|s\_\_P\_L b\_)\_q(|q\_\^P\_L q\_) ,\ O\_5 & = & (|s\_\_P\_L b\_)\_q(|q\_\^P\_R q\_) , \[effh1\]\ O\_6 & = & (|s\_\_P\_L b\_)\_q(|q\_\^P\_R q\_) ,\ O\_7 & = & [e16\^2]{}m\_b(|s\_\_P\_R b\_) F\^ ,\ O\_8 & = & [g\_s16\^2]{}m\_b(|s\_\_T\^a\_ P\_R b\_)G\^[a]{}  . The above effective hamiltonian is then evolved from the electroweak scale down to $\mu\sim m_b$ by the Renormalization Group Equations (RGE). In the RG analysis, the Wilson coefficients are to be evaluated perturbatively at the $W$ scale where the matching conditions are imposed and then evolved down to the renormalization scale $\mu$. The expressions for the $\{ c_k\}$ at the $W$ scale are ------------------ --- ------------------------- $c_{1,3-6}(M_W)$ = $0$ , $c_7(M_W)$ = $-{1\over 2}F_2(x_t)$ , ------------------ --- ------------------------- ------------ --- ------------------------- $c_2(M_W)$ = $1$  , $c_8(M_W)$ = $-{1\over 2}D(x_t)$   . ------------ --- ------------------------- \[cimw\] with D(x)=[x\^3-5x\^2-2x4(x-1)\^3]{}+[3x\^2x2(x-1)\^4]{}  . \[c8mw\] The solution to the RGE at the leading logarithmic order is given by c\_k\^[eff]{}()=U\^5\_[k]{}(, M\_W)c\_(M\_W) , \[ceff\] where $U^5_{k\ell}$ denotes the evolution matrix in a five-flavor context and is determined by U\^5(m\_1,m\_2)\_[kn]{}=O\_[k]{}O\^[-1]{}\_[n]{} . \[u5\] In the above we define $\eta\equiv\alpha_s(m_2)/\alpha_s(m_1)$ and $\vec{a}_\ell\equiv\gamma^D_{\ell\ell}/2\beta_0$ ([*not*]{} summed on $\ell$), where $\beta_0=11-2n_f/3$ and $\gamma^D=O^{-1}\, \gamma^{(eff)\, T}\, O$ is the diagonalized form of the $8\times 8$ anomalous dimension matrix. We use the scheme-independent form of the matrix $\gamma^{eff}$, which is given explicitly in Ref. [@bmmp] in terms of the number of $Q=+2/3$ and $Q=-1/3$ quarks present in the effective theory. Scaling again to the semi-leptonic decay, the branching fraction is now given by \_[bs]{} = [6]{}| [V\_[tb]{}V\^\*\_[ts]{}V\_[cb]{}]{}|\^2 \_[BX]{} , \[br\] The numerical values of the separate contributions to $c_7^{eff}(\mu)$ are, with $m_t = 174$ GeV (for illustration purposes), $\mu=m_b=4.87$ GeV, and $\alpha_s(M_Z)=0.124$ as determined by LEP$^{\cite{lepalpha}}$, c\_7\^[eff]{}() & = & 0.670 c\_7(M\_W) + 0.091 c\_8(M\_W) - 0.172 c\_2(M\_W) ,\ & = & 0.670(-0.195) + 0.091(-0.097) - 0.172  . \[c7eff\] Taking the overall CKM factor in the branching fraction to be unity, and $|V_{ub}|/|V_{cb}|=0.08$, this procedure yields $${\rm B}_{b\to s\gamma} = \left( 2.92^{+0.77}_{-0.59}\right) \times 10^{-4} \,. \label{qcdbr}$$ The central value corresponds to $\mu=m_b$, while the upper and lower errors represent the deviation due to assuming $\mu=m_b/2$ and $\mu=2m_b$, respectively. We see that this value compares favorably to the recent CLEO measurement$^{\cite{cleoinc}}$ of the inclusive rate cited earlier in Eq. (\[incl\]). When compared with the uncorrected result of Eq. (\[br0\]), the QCD corrections are seen to increase the branching ratio by roughly a factor of $2$. We take this opportunity to reflect further on the size of the QCD corrections. Earlier estimates$^{\cite{desh}}$ of these corrections found that the enhancements to the $b\to s\gamma$ branching fraction were more than an order of magnitude for $m_t<M_W$. This is because the effect of the QCD radiative correction to the weak vertex is to replace the GIM power suppression in Eq. (\[il1\]) by a logarithmic suppression. We explicitly illustrate this effect in Fig. 2, where we show the dependence of the $c_7$ Wilson coefficient on the mass of a single internal quark using the calculational procedure described above. In the lower of the two curves, the dependence of $c_7$ determined at scale $\mu=m_W$ is displayed, while the upper curve corresponds to the evolved $c_7$ evaluated at $\mu=m_b$. We see that $c_7(\mu=m_b)$ is a reasonably flat function of the intermediate quark mass, and that the corrections are substantial for light internal quarks, with an increase of 3-4 orders of magnitude in the rate for $m_q=5 - 10$ GeV. For the case of one heavy internal quark, [*e.g.*]{}, $b\to s\gamma$ with $m_t>M_W$, we see that the GIM mechanism no longer plays such a crucial role and the QCD enhancements are not as dramatic. [c]{}\ \ \ \ \ \ \ \ \ \ \ \ \ [Figure 2. Dependence of $c_7$ on Intermediate-quark Mass]{}\ We now consider the case of radiative charm transitions. The $|\Delta c|=1$ effective hamiltonian can be written as H\^[|c|=1]{}\_[eff]{}=[-4G\_F2]{}\_b \_[k=1]{}\^[10]{}c\_k()O\_k() , \[cheff\] with $\lambda_i=V^*_{ci}V_{ui}$ as defined previously. The CKM structure of the operators differs dramatically from the $b\to s\gamma$ case. Here, both $O_1$ and $O_2$ have [*two*]{} contributions which have approximately equal CKM weighting since $\lambda_s\simeq\lambda_d$. We stress that extreme caution must be exercised in order to correctly incorporate these terms. To be precise we explicitly separate $O_{1}$ and $O_{2}$ into two operators according to their CKM structure, O\_[1a]{} & = & (|u\_\_P\_Ls\_)(|s\_\^P\_L c\_), O\_[1b]{} = (|u\_\_P\_Ld\_) (|d\_\^P\_Lc\_) ,\ O\_[2a]{} & = & (|u\_\_P\_Ls\_)(|s\_\^P\_Lc\_), O\_[2b]{} = (|u\_\_P\_Ld\_) (|d\_\^P\_Lc\_) ,\ \[newop\] and write the remaining $|\Delta c| = 1$ operators in a form analogous to their $|\Delta b| = 1$ counterparts, O\_3 & = & (|u\_\_P\_Lc\_)\_q(|q\_\^P\_Lq\_) ,\ O\_4 & = & (|u\_\_P\_Lc\_)\_q(|q\_\^P\_Lq\_) ,\ O\_5 & = & (|u\_\_P\_Lc\_)\_q(|q\_\^P\_Rq\_) , \[effop\]\ O\_6 & = & (|u\_\_P\_Lc\_)\_q(|q\_\^P\_Rq\_) ,\ O\_7 & = & [e16\^2]{}m\_c(|u\_\_P\_Rc\_) F\^ ,\ O\_8 & = & [g\_s16\^2]{}m\_c(|u\_\_T\^a\_ P\_Rc\_)G\^[a]{} , where the terms proportional to $m_u$ in $O_{7,8}$ have again been neglected. Since the quantity $-\lambda_b$ has been factorized in Eq. (\[cheff\]) above, the values of the corresponding Wilson coefficients at the matching scale are now --------------- --- ----------------------------- $c_{1a}(M_W)$ = $0$ , $c_{2a}(M_W)$ = $ -\lambda_s / \lambda_b$ , --------------- --- ----------------------------- --------------- --- ------------------------------- $c_{1b}(M_W)$ = $0$  , $c_{2b}(M_W)$ = $ -\lambda_d / \lambda_b$   . --------------- --- ------------------------------- \[coeffs\] The values of the Wilson coefficients for $c_{3-6}(M_W)$ are the same as in Eq. (\[cimw\]), and the coefficients $c_{7,8}(M_W)$ are modified to c\_7(M\_W) & = & -[12]{},\ c\_8(M\_W) & = & -[12]{} , \[c7mwc\] with each containing intermediate $s$-quark and $b$-quark contributions. Due to the CKM dependence, $c_{7,8}(M_W)$ now contain both real and imaginary terms which in principle must be evolved separately. We note that the real parts of the $x_s$-dependent terms are numerically the same order of magnitude as the $x_b$ terms. Now we evolve the effective theory down to the scale $\mu\sim m_c$. This takes place in two successive steps; first, we go from the electroweak scale down to $m_b$ working in an effective 5 flavor theory, and then to $\mu<m_b$ in an effective 4 flavor theory. This procedure is similar to what is performed for the $|\Delta s| = 1$ kaon transitions$^{\cite{kaon}}$, where the effective theory is evolved to $\mu\sim 1$ GeV in 3 successive steps. We then have c\_k\^[eff]{}() & = & U\^4\_[k]{}(, m\_b)U\^5\_[n]{}(m\_b, M\_W) c\_n(M\_W) ,\ c\_k\^[eff]{}() & = & U\^4\_[k]{}(, m\_b)U\^5\_[n]{}(m\_b, M\_W) c\_n(M\_W) , \[creim\] and |c\_7\^[eff]{}()|\^2=|c\_7\^[eff]{}()|\^2 + |c\_7\^[eff]{}()|\^2 . \[c7sq\] The renormalization group evolution matrices $U^4$ and $U^5$ are determined as in Eq. (\[u5\]) now using a $10\times 10$ anomalous dimension matrix $\gamma^{eff}$. We take the anomalous dimensions of the split operators $O_{1a-b}$ and $O_{2a-b}$ to be exactly those for $O_1$ and $O_2$, respectively, as the anomalous dimensions do not depend on the CKM structure of the operator. We use the form of $\gamma^{eff}$ as given in Ref. [@bmmp], taking care to keep $n_f=4$ and $5$ as needed. The relative numerical values of the contributions to $c_7^{eff}(\mu)$ with $\mu=m_c=1.5$ GeV are\ & = & 0.458 c\_7(M\_W) + 0.125 c\_8(M\_W) - 0.312\[c\_[2a]{}(M\_W)+c\_[2b]{}(M\_W)\] ,\ & = & 0.458(-0.24110\^[-6]{}) + 0.125(-0.13910\^[-5]{}) -0.312 ( [-\_s-\_d\_b]{}) ,\ & = & 0.458(-0.24110\^[-6]{}) + 0.125(-0.13910\^[-5]{}) -0.312 , \[cnumb\] where the CKM unitarity condition $\lambda_s + \lambda_d = - \lambda_b$ has been used to simplify the final term. Incidentally, it should be stressed that the choice of $-\lambda_b$ as a prefactor in Eq. (\[cheff\]) was quite arbitrary, and we could have pulled out some other factor, say $\lambda_s$ (or $\lambda_d$). This would have affected the Wilson coefficients at the matching scale, but the final result would have remained, as it must, unchanged. We now compute the branching fraction. We evaluate $\alpha_s$ in the $\overline {MS}$ scheme (using $\alpha_s(M_Z)=0.124$ as before) and extend the range down to the charm scale using the Bernreuther matching conditions$^{\cite{bern}}$ at the threshold $\mu=m_b=4.87$ GeV. Note that we have also taken the CKM matrix elements to be real and have neglected any possible imaginary components. Given the small values of $c_{7,8}(M_W)$, this approximation is well justified. It is clear from the above that the $c_2(M_W)$ term completely dominates, due to the small contributions to $c_{7,8}(M_W)$ from the light internal quark masses. This is in stark contrast to $b\to s$ transitions (and likewise to $s\to d$), where the heavy internal $t$-quark forces the the magnetic dipole coefficients to be competitive with $c_2(M_W)$. This can be seen explicitly by comparing the above with Eq. (\[c7eff\]). The QCD-corrected branching fraction is then $$\begin{aligned} {\rm B}_{c\to u\gamma} & = & {6\alpha\over \pi}\cdot \left| {V^*_{cb}V_{ub}\over V_{cs}}\right|^2 \cdot {|c_7^{eff}(\mu)|^2\over g(m_s/m_c)+{ \mbox{$|V_{cd}|^2$}\over \mbox{$|V_{cs}|^2$}} g(m_d/m_c) } \cdot {\rm B}_{D^+\to X\ell^+\nu} \,, \nonumber \\ & = & \left( 4.21 - 7.94 \right) \times 10^{-12} \,, \label{qcdchbr}\end{aligned}$$ where the lower (upper) value in the numerical range corresponds to the scale $\mu=2m_c (m_c)$. We see that the effects of the QCD corrections are quite dramatic in charm radiative decays, and that the rate is given almost entirely as a consequence of operator mixing. The stability of this result can be tested once the complete next-to-leading order corrections to the magnetic dipole transitions are known. Finally, we wish to comment further on the CKM dependence of the $|\Delta b| = 1$ and $|\Delta c| = 1$ effective hamiltonians. We consider each case separately: [*$\ (i)\ |\Delta b| = 1$ transition*]{}: Here, the the $t$-quark contribution is seen to dominate in every respect. Thus, for the dipole operators $O_{7}$ and $O_8$, the $u$-quark and $c$-quark loops are omitted because they are numerically tiny ([*e.g.*]{}, see Table 2). Likewise, due to the smallness of the $u$-quark CKM factors, the approximation is made in the literature$^{\cite{gsw}}$ to omit any current-current operators containing $u$-quark fields. This explains why only the $c$-quark dependent operators $O_{1,2}$ appear in the $|\Delta b| = 1$ operator basis of Eq. (\[effh1\]). This assumption also explains another aspect of the analysis. Ordinarily one would expect $O_{1,2}$ to be accompanied by the CKM factor $-\lambda_c$, yet it is the prefactor $\lambda_t$ which appears in the effective hamiltonian of Eq. (\[effh\]). This is because the tiny value of $\lambda_u$ has allowed one to write the CKM unitarity relation as $\lambda_c \simeq - \lambda_t$ and thus remove dependence upon $\lambda_u$. [*$(ii)\ |\Delta c| = 1$ transition*]{}: In this case, the CKM dependence is more complicated since no single quark-loop is dominant. One must expand the operator basis as we did in Eq. (\[newop\]). However, we wish to take note of a seemingly remarkable feature which occurs upon carrying out the RG analysis. The operators $O_{2a}$ and $O_{2b}$ turn out to have equal anomalous dimensions and thus $c_{2a}$ and $c_{2b}$ have the [*same*]{} numerical coefficient in Eq. (\[cnumb\]). The most elegant way to understand this result is to exploit the $U$-spin symmetry present in the system of operators $O_{1a,1b}$ and $O_{2a,2b}$. Thus, suppose instead of proceeding as we did, we replaced the operators $O_{2a}$ and $O_{2b}$ of Eq. (\[newop\]) with the equivalent set $O_{2a}'$ and $O_{2b}'$, where O\_[2a]{} O\_[2a]{}’ && [12]{}(O\_[2a]{} + O\_[2b]{})\ &=& [12]{},\ O\_[2b]{} O\_[2b]{}’ && [12]{}(O\_[2a]{} - O\_[2b]{}) \[sp1\]\ &=& [12]{} , along with a corresponding replacement of coefficients, c\_[2a]{} c\_[2a]{}’ c\_[2b]{} c\_[2b]{}’   . \[sp2\] The matching conditions for the modified coefficients would then have become c\_[2a]{}’ (M\_W) = [\_s + \_d -\_b]{} = 1 c\_[2b]{}’ (M\_W) = [\_s - \_d -\_b]{}   , \[sp3\] with analogous replacements made also for $O_{1a}$ and $O_{1b}$. Since the mixing of $O_{2a}$ and $O_{2b}$ with $O_7$ has no dependence on the mass of the internal $s$ and $d$ quarks, the operator $O_{2b}'$ does not contribute to the process $c \ra u+ \gamma$ due to cancellation between the $s$-quark and $d$-quark contributions. This cancellation is in fact just the manifestation of an underlying $U$-spin symmetry. That is, in the limit of neglecting quark mass, $O_{2b}'$ carries $U$-spin $1$ and thus cannot couple to a photon. This decoupling occurs via the very $s$-quark and $d$-quark cancellation under discussion. As a consequence, the other operator $O_{2a}'$ must have the the same anomalous dimension as $O_{2a}$ under RG flow, and we obtain the result cited above. As a corollary, it is clear that in the $|\Delta b| = 1$ transition, the approximation made of omitting the $u$-quark current-current operators is quite unnecessary. One could just as easily deal with an expanded operator basis containing $u$-quark fields analogous to that of Eq. (\[newop\]) or by invoking an $SU(4)$ version of $u$-quark/$c$-quark $U$-spin symmetry and proceeding as above. **Long Distance Pole Contributions** ==================================== As seen in the previous section, the short distance contributions to charm radiative decays give very small branching ratios, even when the large QCD enhancement is taken into account. There will, however, be additional contributions that appear as exclusive modes for which the momentum scale of the intermediate quarks is a strong interaction scale and not the short distance scale $M_W$. This forces us to view the intermediate states as hadrons rather than quarks. These long distance contributions can be partitioned into two basic classes. The first corresponds, at the quark level, to annihilation diagrams $c\bar{q}_1\to~q_2\bar{q}_3$ where a photon line is attached to any of the four quark lines. In terms of hadronic degrees of freedom, these give rise to the set of contributions which include the [*pole diagrams*]{}. The second type of contribution corresponds to the underlying quark processes $c\to~q_1\bar{q}_2q$, followed by $\bar{q}_2q\to~\gamma$. At the hadronic level, this is the so-called [*vector meson dominance*]{} mechanism. We shall discuss the pole amplitudes in this section, leaving consideration of the VMD mechanism for Section 4. The pole amplitudes are but a subset of an entire class of long-distance contributions, including the two-particle intermediate states and proceeding to all higher $n$-particle intermediate states. However, of these the most phenomenologically accessible are the single-particle or pole terms. The relevant diagrams, appearing in Figs. 3(a,b), are seen to fall into either of two basic classes. We shall refer to transitions as type I if weak-mixing occurs before photon emission, [*i.e.*]{} if the incoming $D$ meson experiences weak-mixing, and as type II if photon emission occurs before weak-mixing, [*i.e.*]{} if the final state meson is created via weak-mixing. In principle, the intermediate states occurring in the type I and type II amplitudes consist respectively of all possible virtual spin-zero and spin-one particles. We shall find it practicable, however, to take into account only the lightest such virtual particles. [c]{}\ \ \ \ \ \ \ \ [Figure 3. Pole Contributions]{}\ In analyzing long range effects for flavor-changing $D$ decays, we shall employ the effective weak hamiltonian of Bauer, Stech and Wirbel$^{\cite{bsw}}$ (BSW), $${\cal H}_{\rm w} = -{G_{\rm F} \over \sqrt{2}} \left[ :a_1 ({\bar u} d')({\bar s}' c) + a_2 ({\bar s}' d')({\bar u} c) : \right] \ \ , \label{py1}$$ where the colons denote normal-ordering and $d'$, $s'$ are the CKM-mixed fields d’ &=& V\_[ud]{} d + V\_[us]{} s  ,\ s’ &=& V\_[cs]{} s + V\_[cd]{} d  . \[py2\] We shall work in the $2\times 2$ basis of quark flavors, = ( [cc]{} V\_[ud]{} & V\_[us]{}\ V\_[cd]{} & V\_[cs]{}\ ) ( [cc]{} 0.975 & 0.222\ -0.222 & 0.975\ )   . \[py3\] Specific forms for the Cabibbo-favored and Cabibbo-suppressed hamiltonians will be given shortly. The quark fields occur in left-handed combinations, denoted by ([|q]{}\_1 q\_2) \_1 \_(1 + \_5) q\_2   , \[py4\] and $a_1, a_2$ are free parameters whose values will generally depend on the mass scale being probed. Here, they are determined by fitting to $D\to {\bar K}\pi$ data$^{\cite{wir}}$, a\_1 (m\_c\^2) = 1.2 0.1  , a\_2 (m\_c\^2) = -0.5 0.1   . \[py5\] [**Pole Amplitudes of Type I**]{} Among the possible exclusive $D$ decays, the most promising for experimental detection occur in the class of vector meson–photon ($V\gamma$) final states, D([**p**]{}) V ([**k**]{}, ) + ([**q**]{}, )   . \[py6\] For these, the transition amplitude has the gauge invariant form \_[DV ]{} = \_\^(k ,) \_\^ (q, )   . \[py7\] The parity-violating and parity-conserving amplitudes are denoted by ${\cal A}^{\rm pv}$ and ${\cal A}^{\rm pc}$ respectively, and each carries the dimension of inverse energy. Both amplitudes are generally required because the weak interaction does not respect parity invariance. The $D\to V \gamma$ decay rate is given by \_[D V ]{} = [|[**q**]{}|\^3 4]{}   (|[A]{}\^[pv]{}|\^2 + |[A]{}\^[pc]{}|\^2 )   , \[py8\] where ${\bf q}$ is the decay momentum in the $D$ rest frame, |[**q**]{}| = [m\_[D]{}\^2 - m\_V\^2 2 m\_[D]{}]{}   . \[py10\] Which particular combination of the parity-conserving and parity-violating amplitudes contributes to the decay process will depend upon the weak-mixing amplitude. In principle, a charm meson can mix with a sequence of either scalar $\{ S_n \}$ or pseudoscalar $\{ P_n \}$ mesons. Although some work on scalar-mixing has been done,$^{\cite{eg}}$ the outcome is rather model-dependent because detailed experimental and theoretical understanding about scalar states is lacking. In this paper, we shall therefore consider only the weak-mixing of charm mesons with light pseudoscalar mesons and thus work with only parity-conserving (PC) pole amplitudes. It is appropriate at this point to comment on the notation to be employed from this point on in both Section 3 and Section 4. We shall denote $f_P$ as the decay constant of pseudoscalar meson $P$ and define $h_{V\gamma P}$ as the coupling constant for the EM interaction vertex of the photon $\gamma$ with the mesons $V$, $P$. Also, the decay constant of vector meson $V$ is given in terms of the $V$-to-vacuum matrix element of the vector current, 0 | V\_\^a | V\^b ([**q**]{},) = \^[ab]{} [m\_V\^2 f\_V]{} \_\^\* ([**q**]{},) \^[ab]{} g\_V \_\^\* ([**q**]{},)   . \[vec\] Note that we define two equivalent parameterizations, $g_V$ (with units of GeV$^2$) and $f_V$ (dimensionless), for the vector decay constant. We have found that employing $g_V$ in the discussion of pole amplitudes alleviates notational confusion which would otherwise occur between the vector and pseudoscalar decay constants $f_V$ and $f_P$. However, it is traditional to use $f_V$ in discussing VMD amplitudes, and we do so in Section 4. The constants $\{ f_V\}$ are obtained from $\Gamma_{V\to l^+l^-}$ data and have recently been compiled in Table 1 of Ref. [@gp]. Now, a pseudoscalar state $P$ which is created by weak mixing will propagate virtually until it eventually decays into the final state. This latter transition is electromagnetic and hence parity-conserving. It has the amplitude \_[V P]{} = h\_[VP]{}\_\^(k ,) \_\^ (q, ) \^ k\_ p\_  . \[py11\] The absolute value of the coupling constant $h_{V\gamma P}$ can be inferred phenomenologically by using |h\_[VP]{}|\^2 = { [ll]{} 12\_[V P ]{}/ |[**q**]{}|\^3 & (M\_V &gt; M\_P)\ 4\_[P V ]{}/ |[**q**]{}|\^3 & (M\_P &gt; M\_V) . . \[py13\] The general type I decay amplitude ${\cal A}_I$ for $D\to V\gamma$ is then given by \_I\^[pc]{}(DV) = \_n  h\_[VP\_n]{} P\_n | [H]{}\^[(eff)]{}\_[w]{} | D   . \[py14\] With Fig. 3(a) as a guide, the notation should be self-evident. Predictions for $D\to V\gamma$ decay amplitudes will be obtained below in terms of both type I and type II pole amplitudes, and in the next section we shall do the same by using VMD amplitudes. We can, however, accomplish somewhat more. In principle, the discussion for $V\gamma$ final states extends to a larger set of meson-photon final states $M\gamma$, where the only restriction on meson $M$ is that it have spin greater than zero. For each different type of $M\gamma$ final state, there will be a gauge invariant $D$-decay amplitude like Eq. (\[py7\]) and an $M\gamma P$ interaction vertex like Eq. (\[py11\]). However, the generic form of Eq. (\[py14\]) continues to hold, except that $h_{V\gamma P_n}$ is replaced by $h_{M\gamma P_n}$. Of course, to have predictive power requires knowledge of the $h_{M\gamma P_n}$ coupling constant. Fortunately, much has been learned about radiative decays in light meson systems over the years. In particular, there are varying amounts of experimental evidence for $17$ such transitions in the listing of Ref. [@pdg]. Of these, $10$ involve $1^- \to 0^-$ mesonic transitions, $3$ involve $2^+ \to 0^-$, $2$ involve $1^+ \to 0^-$ and $2$ involve $0^- \to 1^-$. This information allows us to extend the analysis of type I amplitudes from just $V\gamma$ final states to include both $A\gamma$ and $T\gamma$ configurations as well, where ‘$A$’ and ‘$T$’ stand for axialvector and tensor mesons respectively. The $A\gamma$ final states are very analogous to the $V\gamma$ decays in that the coupling constant $h_{A\gamma P_n}$ is found via Eq. (\[py13\]) and the $D\to A\gamma$ decay amplitude has the same form as Eq. (\[py8\]). For the $T\gamma$ final states, one uses instead |h\_[TP]{}|\^2 = [40\_[T P ]{}|[**q**]{}|\^5]{} \[p16\] as well as \_[D T ]{} = [|[**q**]{}|\^5 4]{}  |[A]{}\_I\^[pc]{}|\^2   . \[py8a\] To summarize, we shall include in our study of type I amplitudes certain $D\to A\gamma$ and $D\to T\gamma$ transitions. For the type II or VMD amplitudes, however, we shall limit our calculations to just the $V\gamma$ final states. [*Cabibbo-favored (CF) transitions*]{}: In this case, the BSW hamiltonian becomes $${\cal H}^{\rm (CF)}_{\rm w} = - V_{ud} {V_{cs}}^*{G_{\rm F} \over \sqrt{2}} \left[ :a_1 ({\bar u} d)({\bar s} c) + a_2 ({\bar s} d)({\bar u} c) : \right] \ \ . \label{p7}$$ The calculation of weak-mixing matrix elements of $D$’s with the light pseudoscalar mesons is straightforward and results are tabulated in Table 4. The fact that these mixing amplitudes are evaluated in vacuum saturation makes the forms in Table 4 easy to interpret. Thus, for example, in Cabibbo-favored $D_s^+$ decay, it is the term in the BSW hamiltonian with coefficient $a_1$ which contributes, and as such, the weak-mixing matrix element is naturally proportional to the decay constants $f_\pi$ and $f_{D_s^+}$. [lc]{}\ Mixing & Matrix Element\ $D_s^+ \to \pi^+$ & $a_1 V_{ud} {V_{cs}}^* f_\pi f_{D_s^+} m_{D_s^+}^2 G_F/\sqrt{2}$\ $D^0 \to {\bar K}^0$ & $a_2 V_{ud} {V_{cs}}^* f_K f_{D} m_{D^0}^2 G_F/\sqrt{2}$\ $D^+ \to \pi^+$ & $0$\ For the decay constants of the light mesons we use f\_= 131 [MeV]{} f\_K = 161 [MeV]{}   . \[p8\] The present situation for the decay constants $f_D$ and $f_{D_s}$ of the charm mesons is somewhat problematic. Experiment provides the upper limit for $f_D$, $$f_D < 290~{\rm MeV} \ \ , \label{p9}$$ as obtained from the branching ratio determination $Br_{D^+ \to \mu^+ \nu_\mu} < 7.2 \times 10^{-4}$ at $90 \%$ confidence-level$^{\cite{mark3}}$. Thus only theoretical estimates exist for $f_D$. These occur in three categories, lattice theoretic$^{\cite{cb}}$, QCD sum rule$^{\cite{dom}}$ and quark model fits to color-hyperfine mass splittings$^{\cite{ros}}$. Estimates fall in the range $185 < f_D ({\rm MeV}) < 262$. We shall adopt the value f\_D\^[latt.]{} 216 [MeV]{}   , \[p10\] which is an average over the lattice estimates$^{\cite{shig}}$ and falls between the other two types of determinations. Recently, the following experimental results (in units of MeV) for $f_{D_s}$ were announced by CLEO$^{\cite{cleo6}}$, WA75$^{\cite{wa75}}$ and BES$^{\cite{bes}}$, f\_[D\_s]{} = { [ll]{} 344 37 67 & ([CLEO]{})\ 232 45 52 & ([WA75]{})\ 434\^[+153 +35]{}\_[-133 -33]{} & ([BES]{})   , . \[p11\] where the CLEO value is inferred from the ratio $\Gamma_{D_s^+ \to \mu^+ \nu_\mu}/ \Gamma_{D_s^+ \to \phi \pi^+}$ along with the branching ratio $B_{D_s^+ \to \phi \pi^+}$. In our numerical analysis, we shall use the following weighted average of the above decay constants, f\_[D\_s]{}\^[expt.]{} = 299 [MeV]{}   . \[p13\] For the sake of comparison, we note that this value is somewhat larger than the central value of a weighted average taken from a compilation of existing lattice estimates,$^{\cite{shig}}$ f\_[D\_s]{}\^[latt.]{} = 242 [MeV]{}   . \[p14\] The only other ingredients needed are the radiative coupling constants $h_{M\gamma P}$, which were defined earlier. Putting together all the necessary ingredients and ranging over the set of final state mesons $M = \rho(770)$, $K^*(892)$, $b_1 (1235)$, $a_1 (1270)$, $a_2 (1320)$ and $K_2^*(1430)$ yields the magnitudes of type I pole-model amplitudes (in units of GeV$^{-1}$) given in Table 5. [lc]{}\ Mode & $|{\cal A}_I^{\rm pc}|~({\rm GeV}^{-1})$\ $D_s^+ \to \rho^+\gamma$ & $8.2\times 10^{-8}$\ $D_s^+ \to b_1^+ (1230)\gamma$ & $7.2\times 10^{-8}$\ $D_s^+ \to a_1^+ (1270)\gamma$ & $1.2\times 10^{-7}$\ $D_s^+ \to a_2^+ (1320)\gamma$ & $2.1\times 10^{-7}$\ $D^0 \to {\bar K}^{*0}\gamma$ & $5.6\times 10^{-8}$\ These values should be considered as upper bounds for the following reason. We have considered the lightest possible intermediate states, pions and kaons, because only for these particles is there sufficient data for determining coupling constants. However, the pion and kaon intermediate states propagate far off-shell. Instead of having a squared momentum near the mass-shell value $q^2 = m_\pi^2$, the virtual pion must carry $q^2 = m_{D_s}^2 \gg m_\pi^2$ and similarly for the kaon. This effect could well suppress the transition amplitude. In principle, one is to sum over all pion-like and kaon-like intermediate states. Other possible contributions should be heavier and thus less affected by this suppression effect. For pion-like intermediate states, the next state in order of increasing mass would be $\pi (1300)$ and beyond that the unconfirmed state $\pi (1770)$. Although there is not sufficient data to make a numerical estimate of their effect, we can anticipate for such states that   (i) the propagator contribution will indeed be larger,  (ii) the weak-mixing between a ground state $D$ meson and a radially excited meson $P_n$ will be wave-function suppressed, \(iii) the radiative coupling constant $h_{M \gamma P_n}$ might well be relatively smaller due to phase space competition with other decay modes of the massive meson $P_n$. We would expect the net result of these effects to decrease the overall contribution from the excited states. [*Cabibbo-suppressed (CS) transitions*]{}: The weak-mixing now proceeds according to the weak hamiltonian \^[(CS)]{}\_[w]{} &=& - [G\_[F]{} ]{}   , \[p17\] The action of the $({\bar d} d)$ and $({\bar s} s)$ operators on the vacuum when expressed in terms of the pseudoscalar meson states becomes ([|d]{} d) &=& -0.7071 \^0 + 0.58 + 0.40 ’\ ([|s]{} s) &=& -0.57 + 0.82 ’   , \[p18\] where an $\eta - \eta'$ mixing angle $\theta_{\rm P} = -20^o$ is adopted.$^{\cite{dgh}}$ In addition, we take$^{\cite{dgh}}$ f\_f\_[’]{} f\_  . \[eta\] The mixing amplitudes which are relevant for Cabibbo-suppressed decay appear in Table 6. Observe that we have simplified the notation for $D^0$ transitions by expressing $V_{ud}{V_{cd}}^*$ in terms of $V_{us}{V_{cs}}^*$. [lc]{}\ Mixing & Matrix Element\ $D^+ \to \pi^+$ & $a_1 V_{ud} V_{cd}^* f_\pi f_{D^+} m_{D^+}^2 G_F/\sqrt{2}$\ $D_s^+ \to K^+$ & $a_1 V_{us} V_{cs}^* f_K f_{D_s} m_{D_s}^2 G_F/\sqrt{2}$\ $D^0 \to \pi^0$ & $0.7071 a_2 V_{us} V^*_{cs} f_{D^0} f_{\pi} m_{D^0}^2 G_F/2$\ $D^0 \to \eta$ & -$1.15 a_2 V_{us} V^*_{cs} f_{D^0} f_{\eta} m_{D^0}^2 G_F/2$\ $D^0 \to \eta'$ & $0.42 a_2 V_{us} V^*_{cs} f_{D^0} f_{\eta'} m_{D^0}^2 G_F/2$\ The analysis for Cabibbo-suppressed decays proceeds analogous to that for Cabibbo-favored decays, with one significant complication. For each of the Cabibbo-favored transitions, only one amplitude contributes. For $D^0$ decay, however, all the Cabibbo-suppressed pole amplitudes contain a sum over $\pi^0$, $\eta$ and $\eta'$ intermediate states. It is important to get the relative phases of the interfering amplitudes correct. We have therefore performed an analysis of the nine $V^0 \to P^0 \gamma$ couplings in light of the most recent data, where $V^0 = \rho^0$, $\omega^0$, $\phi^0$ and $P^0 = \pi^0$, $\eta$ and $\eta'$.$^{\cite{kjm}}$ The magnitudes of the Cabibbo-suppressed amplitudes are displayed in Table 7. [lc]{}\ Mode & $|{\cal A}_I^{\rm pc}|~({\rm GeV}^{-1})$\ $D^+ \to \rho^+\gamma$ & $1.3\times 10^{-8}$\ $D^+ \to b_1^+(1230)\gamma$ & $1.2\times 10^{-8}$\ $D^+ \to a_1^+(1270)\gamma$ & $4.9\times 10^{-9}$\ $D^+ \to a_2^+(1320)\gamma$ & $3.4\times 10^{-8}$\ $D_s^+ \to K^{*+}\gamma$ & $2.8\times 10^{-8}$\ $D_s^+ \to K_2^{*+}(1430)\gamma$ & $6.0\times 10^{-8}$\ $D^0 \to \rho^0\gamma$ & $4.8\times 10^{-9}$\ $D^0 \to \omega^0\gamma$ & $6.1\times 10^{-9}$\ $D^0 \to \phi^0\gamma$ & $7.4\times 10^{-9}$\ [**Pole Amplitudes of Type II**]{} Analogous to the type I $D\to V\gamma$ decay amplitude of Eq. (\[py14\]) we have \_[II]{}\^[pc]{}(DV) = \_n  V | [H]{}\^[(eff)]{}\_[w]{} | D\^\*\_n h\_[D\^\*\_n D]{} \[tpii\] for the corresponding type II transition. From the viewpoint of phenomenology, the type II transitions are more problematic than are those of type I because less experimental input is available. Thus, we shall need to rely a bit more heavily on theoretical predictions. The first difficulty is that the couplings $h_{D^{*0}\gamma D^0}$, $h_{D^{*+}\gamma D^+}$, and $h_{D_s^{*+}\gamma D_s^+}$ have not yet been experimentally measured. This is because, although the relevant photonic branching ratios have been measured, only upper bounds exist for the full widths of the associated spin-one exited states, $D^{*0}$, $D^{*+}$ and $D_s^{*+}$, \_[D\^[\*0]{}]{} &lt; 2100 [keV]{} && \_[D\^[\*0]{}D\^0]{} &lt; 764 [keV]{}\ \_[D\^[\*+]{}]{} &lt; 131 [keV]{} && \_[D\^[\*+]{}D\^+]{} &lt; 1.44 [keV]{} \[p20\]\ \_[D\_s\^[\*+]{}]{} &lt; 4500 [keV]{} && \_[D\_s\^[\*+]{}D\_s\^+]{} &lt; 4500 [keV]{}  . Fortunately, predictions for the $\Gamma_{D^{*0}\to D^0\gamma}$, $\Gamma_{D^{*+}\to D^+\gamma}$ and $\Gamma_{D_s^{*+}\to D_s^+\gamma}$ transitions have appeared in the literature recently.$^{\cite{{aip},{cdn},{ox},{fr}}}$ There is some spread in predictions, and so we choose the representative values, \_[D\^[\*0]{}D\^0]{} = 20 [keV]{} , \_[D\^[\*+]{}D\^+]{} = 0.5 [keV]{}  , \_[D\_s\^[\*+]{}D\_s\^+]{} = 0.3 [keV]{}  , \[p21\] which implies h\_[D\^[\*0]{}D\^0]{} &=& 0.542 [GeV]{}\^[-1]{} ,\ h\_[D\^[\*+]{}D\^+]{} &=& -0.087 [GeV]{}\^[-1]{}  , \[p21a\]\ h\_[D\_s\^[\*+]{}D\_s\^+]{} &=& -0.066 [GeV]{}\^[-1]{}   , where we have adopted the phases implied by the quark model. A rough check on whether the above values are reasonable is afforded by the nonrelativistic quark model, in which h\_[D\^[\*0]{}D\^0]{} &=& 2e ,\ h\_[D\^[\*+]{}D\^+]{} &=& e , \[p21b\]\ h\_[D\_s\^[\*+]{}D\_s\^+]{} &=& e , where the $\{ M_k \}$ are constituent quark masses, distinct from the current masses $\{ m_k\}$ of Section 2. If we take $M_c \simeq 1.64$ GeV, as implied by a fit to $D$ and $D^*$ masses, then the relations in Eqs. (\[p21\]-\[p21b\]) yield $M_u \simeq M_d = 0.48$ GeV and $M_s \simeq 0.53$ GeV. The other of the difficulties concerns the weak-mixing matrix elements. For type II transition amplitudes, the mixing occurs between charm and light vector mesons, as in \^+ | [H]{}\^[(eff)]{}\_[w]{} | D\_s\^[\*+]{} a\_1 V\_[ud]{} V\_[cs]{} g\_[\^+]{} g\_[[D]{}\_s\^[\*+]{}]{} G\_F/   . \[p22\] In the above, the $g_V$ are the vector meson decay constants defined in Eq. (\[vec\]) and whose determination we shall discuss shortly. As with the type I amplitudes, we have employed vacuum saturation. To determine the action of the $({\bar d} d)$ and $({\bar s} s)$ operators upon the vacuum we employ the ideally-mixed vector meson states, so that ([|d]{} d) = [- \^0 ]{} ([|s]{} s) =   . \[p22a\] For the light $1^-$ mesons, the collection $\{ g_V \}$ of vector decay constants can be determined by referring to the vacuum-to-meson matrix elements of $J^\mu_{\rm em}$ given in Table 1 of Ref. [@gp]. Together with isospin and $SU(3)$ relations along with quark model insights, these generate all the needed values, [*e.g.*]{} g\_[\^+]{} 0.17 [GeV]{}\^2  , g\_[K\^\*]{} g\_[\^+]{} 0.22 [GeV]{}\^2  , …  . \[p23\] To estimate the $D_s^{*+}$ and $D^{*0}$ decay constants, we invoke the heavy-quark-symmetry relations, g\_[D\_s\^[\*]{}]{} &=& m\_[D\_s]{} f\_[D\_s]{} 0.588 [GeV]{}\^2  ,\ g\_[D\^[\*]{}]{} &=& m\_[D]{} f\_[D]{} 0.403 [GeV]{}\^2   . \[p24\] The magnitudes of the type II amplitudes thus calculated are given in Table 8. [lc]{}\ Mode & $|{\cal A}_{II}^{\rm pc}|~({\rm GeV}^{-1})$\ $D_s^+ \to \rho^+\gamma$ & $1.9\times 10^{-8}$\ $D^0 \to {\bar K}^{*0}\gamma$ & $5.9\times 10^{-8}$\ $D^+ \to \rho^+\gamma$ & $3.6\times 10^{-9}$\ $D_s^+ \to K^{*+}\gamma$ & $5.1\times 10^{-9}$\ $D^0 \to \rho^0\gamma$ & $4.7\times 10^{-9}$\ $D^0 \to \omega^0\gamma$ & $6.9\times 10^{-9}$\ $D^0 \to \phi^0\gamma$ & $1.6\times 10^{-8}$\ $D^*$ excitations with spins not equal to one will not contribute to type II amplitudes if we adhere strictly to the hamiltonian of Eq. (\[py1\])$^{\cite{gp}}$ and continue to work within the vacuum saturation framework. The reason is that mesons with $J>1$ cannot have a nonzero matrix element with the vacuum via the current ${\bar q}\gamma_\mu (1 + \gamma_5) c$. The possibility of an intermediate charm meson with $J=0$ is disallowed since it could only mix with a final-state $J=0$ particle and the decay of a spinless particle to another spinless particle plus a photon is forbidden. Although we have considered just final state vector mesons in Table 8, it should be obvious that in principle the spin-one intermediate $D^*$ states can also mix weakly with axialvector final state mesons. Unfortunately, one knows less about the decay constants of axialvector mesons than one does of the vector mesons. **Long Distance VMD Contributions** =================================== The VMD contribution to charm meson radiative decay is depicted in Fig. 4, where a $D$ meson is seen to (i) decay weakly into a final state of a vector meson $V$ and a meson $M$ of nonzero spin, followed by (ii) an electromagnetic VMD conversion of $V$ into a photon. Roughly speaking, in the VMD approach the $D\to M\gamma$ amplitude is obtained by multiplying the $D\to MV$ amplitude by the factor $e / f_V$ where $e$ is the electric charge and $f_V$ is the dimensionless version of the vector meson decay constant defined in Eq. (\[vec\]). It is important to keep in mind that in the VMD process $D\to MV$, the vector meson $V$ is off-shell. Thus, to obtain the VMD amplitude for $D\to M\gamma$ will require an extrapolation from $p_V^2=m_{V}^{2}$ to $p_V^2=0$ for both the $V\to\gamma$ vertex and the $D\to MV$ transition. For our considerations, the main intermediate states will involve virtual rho and phi mesons. We shall employ the observation made in Ref. [[@paul]]{} that the rho-gamma vertex seems to be unaffected by the extrapolation whereas the phi-gamma vertex is reduced by a factor of $\eta_\phi \simeq \sqrt{2}$. In the following, we will consider a number of examples for the case $M = V$, and so we shall be working with VMD chains which begin with the process $D\to VV$. Since the $VV$ final states have $L=0,1,2$ as allowed orbital angular momentum values, the VMD amplitude will in general have a parity-conserving part ${\cal A}_{\rm VMD}^{\rm pc}$ corresponding to the $VV$ $P$-wave and a parity-violating part ${\cal A}_{\rm VMD}^{\rm pv}$ corresponding to the $VV$ $S$-wave and/or $D$-wave. [c]{}\ \ \ \ \ \ \ \ [Figure 4. VMD Contribution]{}\ In practice, there are two means for determining the $D\to MV$ part of a VMD amplitude for $D\to M\gamma$: 1. One can input $D\to MV$ experimental data directly in order to phenomenologically determine the $D\to MV$ amplitude. In this approach, it is crucial to maintain gauge invariance. A careful discussion of how to construct a gauge invariant amplitude was recently given in Ref. [@gp] (which considered this type of empirical VMD contribution to $B\to K^*\gamma$), so we need not detail this procedure here. Since the database for $D\to MV$ transitions is unfortunately small, the ability to generate VMD amplitudes using this phenomenological method is limited. 2. One can employ some theoretical description to model the $D\to MV$ amplitude. Since the models currently available do not always reliably reproduce branching ratios and polarizations of final-state vector mesons in decays of heavy mesons,$^{\cite{dtp}}$ this method is also not beyond criticism. For definiteness, we shall continue to employ the BSW model$^{\cite{bsw}}$ introduced in Section 3.$^{\cite{k}}$ Within this approach, the squared VMD amplitude for the important case where $M$ is a vector meson becomes & & |[A]{}\_[VMD]{}|\^2=a\_[i]{}\^[2]{}(m\_[c]{}\^[2]{})f\^[2]{}\_[X]{} I\ & & ()  , \[bsw\_rad\] where $|{\bf k}|$ is the photon spatial momentum, $q$ represents either of the $d$ or $s$ light quarks, and $I$ is a process-dependent isospin coefficient. The BSW coefficients $a_{1}(m_{c}^{2})$ and $a_{2}(m_{c}^{2})$ which correspond to the color-favored and color-suppressed operators are given in Eq. (\[py5\]). The remaining notation is explained by noting that in the factorization approximation for $D\to MV$, one of the final state particles, which we call $X$ (either $M~{\rm or}~V$), couples directly to the vacuum and the other, which we call $Y$ (either $V~{\rm or}~M$), appears in the $D$-to-$Y$ matrix element of the charged weak current $J^\mu_{\rm ch}$. Thus the quantity $f_X$ is the decay constant of $X$, and $A_1(q^2)$ and $V(q^2)$ are the semileptonic form factors defined by $$\begin{aligned} \lefteqn{\langle Y(p_Y)|J^\mu_{\rm ch}|D(P)\rangle =} \nonumber \\ & &\frac{2V(q^2)}{m_D+m_Y}\epsilon^{\mu\nu\rho\sigma}\epsilon^*_\nu P_\rho p_{Y\sigma} + 2m_Y iA_0(q^2)\frac{\epsilon^*.q}{q^2}q^\mu \nonumber \\ & & +i\left[(m_D+m_Y)A_1(q^2) \epsilon^{*\mu} - \frac{\epsilon^*.qA_2(q^2)}{m_D+m_Y}(P+p_y)^\mu \right. \nonumber \\ & &\left. - 2m_YA_3(q^2)\frac{\epsilon^*.q}{q^2}q^\mu\right] \ . \label{ffs_def}\end{aligned}$$ In the VMD amplitude, the form factors are to be evaluated at $q_{0}^{2}=0$ if $X=V$ and at $q_{0}^{2}=m_{M}^2$ if $X=M$. Throughout, we shall make use of the form factors as measured$^{\cite{form}}$ in $D\to K^* l\nu$ and also employ $SU(3)$ relations as needed. This should provide a good estimate of the form factors appearing in the $D$-to-$\rho$ and $D$-to-$\phi$ matrix elements. Whenever the form factors are to be evaluated at momentum transfers other than at $q^2 = 0$, we shall use a monopole form to extrapolate from $q^2=0$. This amounts simply to dividing the form factors at $q^2=0$ by the quantity $1-q^2/m_{\rm pole}^{2}$. In the following, we shall give VMD predictions for a number of specific $D\to M\gamma$ decays, grouped as Cabibbo favored, singly suppressed or doubly suppressed. In the few cases where we can employ both the above approaches, we shall refer to them respectively as ‘Meth. 1’ and ‘Meth. 2’. Given the lack of abundant $D\to MV$ data, however, we shall be forced to adopt the theoretical approach of Meth. 2 in most cases. Before we can proceed, there is another topic which must be addressed, the dynamical complication of significant Final State Interactions (FSI). Although presumably not a problem in $B\to MV$ decays, detectable FSI are known to exist in the $D$-meson mass region. This can produce an ambiguity in the VMD analysis because FSI will inherently be part of any VMD amplitude obtained from $D\to MV$ data, but will not be present in the BSW construction. It is difficult to remove the effect of FSI from the phenomenological VMD amplitude because the vector meson $V$ is to be taken off-shell, and FSI might have an important kinematic dependence, [*i.e.*]{} the $p_{V}^{2}$ dependence of the FSI has also to be taken into account. Consequently, any FSI effects entering in data may not be present to the same extent in the VMD amplitudes. As regards the factorization construction (Meth. 2 above), the exclusion of any FSI effects in the BSW amplitude amounts [*de facto*]{} to a specific prescription for the $p_V^2$-dependence of the FSI. There is some information on the $p_V^2$-dependence of the $\gamma - V$ couplings and of certain matrix elements, but it is not possible at this time to separate the two effects. As we show in the Appendix, the effect for $\rho$ emission in $A_2 (1320)$ decay can be as much as a factor of two. By contrast, no such suppression is seen in $\rho$-photoproduction, although in $\phi$-photoproduction an effective reduction of about $\sqrt{2}$ in amplitude is observed and a somewhat smaller effect of $\sqrt{1.5}$ is seen in $\omega$-photoproduction. [**Cabibbo-favored Modes**]{} #### $D^0\to {\bar K}{^{0*}}\gamma$: This is an instance in which the phenomenological approach is applicable since experimental information on the $D\to MV$ intermediate state is available. There is a branching ratio determination$^{\cite{pdg}}$ $B_{D^0 \ra \bar{K}^{*0} \rho^0}=(1.6 \pm 0.4)\%$ and the amplitude is known to be (i) almost all transverse and (ii) almost all $S$-wave.[^1] This allows us to write the VMD contribution to $D^0 \ra \bar{K}^{*0} \gamma$ as $${\cal A}^{\rm pv}_{\rm VMD} = {e \over f_\rho} \cdot \frac{a_{D^0\to \bar{K}^{*0} \rho^0}}{m_D E_\gamma}\ , \qquad \qquad {\cal A}^{\rm pc}_{\rm VMD} \simeq 0 \ , \qquad ({\rm Meth.~1})$$ where we follow the notation of Ref. [@gp] and denote $a_{D^0\to \bar{K}^{*0} \rho^0}$ as the phenomenological $S$-wave amplitude for $D^0\to \bar{K}^{*0} \rho^0$. With $\Gamma_{D^0 \to \bar{K}^{*0} \rho^0} = 2.53\times 10^{-14}$ GeV and $a_{D^0\to \bar{K}^{*0} \rho^0} = 1.63\times 10^{-6}$ GeV, this yields ${\cal A}^{\rm pv}_{\rm VMD} (D^0 \ra \bar{K}^{*0} \gamma)$ of about $6.8\times 10^{-8}~{\rm GeV}^{-1}$. The data on $D^0 \to {\bar K}^{*0}\rho^0$ is consistent with no parity-conserving ($P$-wave) contribution. Alternatively, the factorization approach of Eq. (\[bsw\_rad\]) predicts both amplitudes. In this case, we take $a_i=a_2$, and the vector meson to be mixed with the photon is the $\rho^0$, so that $V=\rho^0$ and $X=\bar{K^{0*}}$. The form factors needed are those entering in $D\to\rho$ semileptonic transitions. Making use of the measured $D\to K^*$ form factors implies $I=1/2$. To extrapolate the form factors from $q^2=0$ to $q^2=m_{K^*}^{2}$, we use a monopole form where the $D^*$ is the nearest singularity. The parity-violating and parity-conserving amplitudes are given in Eq. (\[bsw\_rad\]) by the terms involving the $A_1$ and $V$ form factors respectively. Using $f_{K^*}=0.2~{\rm GeV}^2$ we obtain $${\cal A}_{\rm VMD}^{\rm pv}=5.1\times 10^{-8}~{\rm GeV}^{-1} \ , \qquad {\cal A}_{\rm VMD}^{\rm pc} = 3.8\times 10^{-8}~{\rm GeV}^{-1} \ . \qquad ({\rm Meth.~2})$$ We notice that ${\cal A}_{\rm VMD}^{\rm pv}$ is in reasonable agreement with the one obtained from the use of data from the nonleptonic mode, given the large uncertainties involved in these predictions. Indeed, the factorization estimate for the $D^0\to\bar{K^{0*}}\rho^0$ $S$-wave amplitude gives $a_{D^0\to \bar{K}^{*0} \rho^0} =1.3\times 10^{-6}~{\rm GeV}^{-1}$ which is within 20% of the experimental value. It also predicts a $P$-wave branching fraction of $0.15\%$ for $D^0 \to \bar{K^{0*}}\rho^0$, which is below the current upper limit of $0.30\%$. #### $D^+_s\to\rho^+\gamma$: The VMD amplitude for this decay proceeds via $D^+_s\to\phi\rho^+$ followed by $\phi$-$\gamma$ mixing. Although the branching ratio for $D^+_s \ra \phi \rho^+$ is known to be ($6.5 \pm_{1.8}^{1.6}\%$), no information on helicities or partial waves exists, so we cannot apply the phenomenological method here. Turning instead to the factorization approach of Eq. (\[bsw\_rad\]), we have $X=V=\phi$ and $Y=M=\rho^+$. Therefore we require the $D^+_s\to\phi$ semileptonic form factors evaluated at $q_{0}^{2}=m_{\rho}^{2}$. Although there is experimental information on these decays, the branching fraction and the form factors depend strongly on $B_{D^+_s\to\phi\pi^+}$, which is still very uncertain. Thus, again making use of $D\to K^*$ data, taking $I=1$ and with a decay constant of $g_\rho\simeq 0.17~{\rm GeV}^2$, we find $${\cal A}^{\rm pv}_{\rm VMD} = 3.2\times10^{-8}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}^{\rm pc}_{\rm VMD} = 2.8\times10^{-8}~{\rm GeV}^{-1} \ \ .$$ The Cabibbo-favored VMD amplitudes are summarized in Table 9. [l|cc]{}\ Mode &\ & Parity-conserving & Parity-violating\ $D^0 \to {\bar K}^{*0}\gamma$ & $3.8$ & 5.1-6.8\ $D_s^+ \to \rho^+\gamma$ & $3.2$ & $2.8$\ [**Singly Cabibbo-suppressed Modes**]{} #### $D^0\to\rho^0\gamma$: This process can proceed via two different intermediate states, namely $\rho^0\phi$ and $\rho^0\rho^0$. There is one known branching ratio $B_{D^0 \ra \rho^0\phi}\ = (1.9 \pm 0.5)\times 10^{-3}$ with no helicity (or partial wave) information. Letting $\eta_T$ be the transverse fraction of the observed branching ratio, $\eta_S$ the $S$-wave fraction in the transverse mode, and $\eta_P$ the $P$-wave fraction, we then obtain for the $S$-wave amplitude of $D^0 \ra \rho^0\phi$, $$a_{D^0 \to \rho^0\phi} = m_D\sqrt{ 4\pi \Gamma_{D^0 \ra \rho^0\phi} \eta_T \eta_S \over |{\bf k}|} = 7\times 10^{-7} \sqrt{\eta_T \eta_S}~{\rm GeV}\ \ ,$$ and for the corresponding $P$-wave, $${b_{D^0\to \rho^0 \phi}\over m_\phi m_\rho} = \sqrt{4 \pi \Gamma_{D^0 \ra \rho^0\phi} \eta_T \eta_P \over |{\bf k}|^3} = 1.46\times 10^{-6} \sqrt{\eta_T \eta_P}~{\rm GeV}^{-1}\ \ ,$$ where again we employ the notation of Ref. [@gp] in denoting $b_{D^0\to \rho^0\phi}$ as the phenomenological $P$-wave amplitude for $D^0\to \rho^0\phi$. Then, multiplying by the VMD factor $e/f_{\phi}$, we obtain the Meth. 1 estimate $${\cal A}^{\rm pv}_{\rm VMD} = 0.60\times 10^{-8}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}^{\rm pc}_{\rm VMD} = 1.0\times 10^{-8}~{\rm GeV} \label{rhogam_1}$$ for $\eta_T \sim 0.5$ and $\eta_S \sim 0.66$. On the other hand, there is no available experimental information for $D^0\to\rho^0\rho^0$, other than $B_{D^0\to\pi^+\pi^-\pi^+\pi^-} = (8.3\pm0.9)\times 10^{-3}$ which can be taken as an upper limit. Let us also estimate the $D^0 \to \rho^0 \gamma$ mode in the factorization approach, which can be used to predict both the off-shell amplitudes, $D^0\to \rho^0\phi$ and $D^0\to \rho^0\rho^0$. In both cases we need the $D^0\to\rho^0$ form factors, for which $I=1/2$. Using Eq. (\[bsw\_rad\]) we obtain $$\begin{aligned} {\cal A}_{\rm VMD}^{\rm pv}(\rho^0\phi \to \rho^0\gamma )&=&0.22\times 10^{-8}~{\rm Gev}^{-1}\nonumber\\ {\cal A}_{\rm VMD}^{\rm pc}(\rho^0\phi \to \rho^0\gamma )&=&0.18\times 10^{-8}~{\rm Gev}^{-1}\nonumber\\ {\cal A}_{\rm VMD}^{\rm pv}(\rho^0\rho^0 \to \rho^0\gamma )&=&0.75\times 10^{-8}~{\rm Gev}^{-1}\nonumber\\ {\cal A}_{\rm VMD}^{\rm pc}(\rho^0\rho^0 \to \rho^0\gamma ) &\simeq& 0 \ \ . \label{rhogam_2}\end{aligned}$$ Our estimate for the parity-conserving $\rho^0\rho^0 \to \rho^0\gamma$ transition is based on the observation that an on-shell $P$-wave $\rho^0\rho^0$ state is forbidden by Bose statistics and hence the associated off-shell amplitude will be suppressed. First, let us compare the first two rows in Eq. (\[rhogam\_2\]) with the results obtained in Eq. (\[rhogam\_1\]) by making use of the $\rho^0\phi$ data. We can see that the factorization amplitudes are lower, as caused by smaller predictions for the nonleptonic intermediate modes. In general, factorization predictions will be modified by FSI. For instance, in the case at hand there could be a large enhancement due to $K^*\bar{K}^*\to\rho^0\phi$ rescattering effects.$^{\cite{jfd}}$ If this is the case, this effect strongly depends on the kinematics and it is different in the off-shell nonleptonic amplitudes entering in the calculation of the VMD diagrams. The factorization approach provides a prediction which is free from FSI effects. In these cases we will take these two estimates as the allowed range. On the other hand, factorization predicts that the $\rho^0\rho^0$ intermediate state provides most of the VMD amplitude. When both intermediate states are taken into account in the factorization estimate, the predictions of Eq. (\[rhogam\_1\]) and Eq. (\[rhogam\_2\]) roughly agree. This will not be the case for the following mode. #### $D^0\to\phi\gamma$: Now, there is only one nonleptonic intermediate state, $\phi\rho^0$. The amplitudes as extracted from $D^0\to\phi\rho^0$ data are $$\begin{aligned} {\cal A}^{\rm pv}_{\rm VMD} & = & \frac{e/f_\rho}{e/f_\phi} ~ {\cal A}_{\rm VMD}^{\rm pv}\big|_{D^0 \to \rho^0 \gamma} = 2.1 \times 10^{-8}~{\rm GeV}^{-1} \ \ , \nonumber\\ {\cal A}^{\rm pc}_{\rm VMD} & = & 3.5 \times 10^{-8}~{\rm GeV}^{-1} \label{phigam_1}\ \ .\end{aligned}$$ On the other hand, factorization predicts the much smaller amplitudes $${\cal A}^{\rm pv}_{\rm VMD} = 0.7\times 10^{-8}~{\rm GeV}^{-1} \quad {\rm and} \quad {\cal A}^{\rm pc}_{\rm VMD} = 0.6\times 10^{-8}~{\rm GeV}^{-1} \label{phigam_2}.$$ Part of the difference between the predictions in Eq. (\[phigam\_1\]) and Eq. (\[phigam\_2\]) may be due to the presence of FSI effects in the on-shell amplitude measured and used in Eq. (\[phigam\_1\]) and the assumed absence of FSI in Eq. (\[phigam\_2\]). In Table 10 we include both predictions as the allowed range. #### $D^0\to\omega\gamma$: This mode is very similar to $D^0\to\rho^0\gamma$ and we obtain from the factorization approach of Meth. 2 $${\cal A}^{\rm pv}_{\rm VMD} = 0.7\times 10^{-8}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}^{\rm pc}_{\rm VMD} = 0.6\times 10^{-8}~{\rm GeV}^{-1} \ \ .$$ #### $D^+ \ra \rho^+ \gamma$: Here, the mode $D^+ \to \rho^+ \rho^0$ should give the dominant contribution to the VMD amplitude, implying the Meth. 2 amplitudes $${\cal A}^{\rm pc}_{\rm VMD} = 1.9\times 10^{-8}~{\rm GeV}^{-1} \quad {\rm and} \quad {\cal A}^{\rm pv}_{\rm VMD} = 1.6\times 10^{-8}~{\rm GeV}^{-1}\ .$$ Incidentally, the expectation for $D^+ \ra \rho^+ \rho^0$ is that its branching ratio should be at least 0.4%. #### $D^+_s \ra K^{*+} \gamma$: Proceding analogously, we use the factorization estimate of $D^+_s \ra K^{*+} \rho^0$ to express the VMD amplitudes for $D^+_s \ra K^{*+} \gamma$ as $$\begin{aligned} {\cal A}^{\rm pv}_{\rm VMD} &=& \frac{e}{f_\rho} \frac{G_F}{\sqrt{2}} \frac{a_2}{\sqrt{2}} (m_{D_s} + m_K*) {m_\rho f_\rho \over A_1} {m_{D_s} E_\gamma} \ ,\\ \nonumber {\cal A}^{\rm pc}_{\rm VMD} & = & \frac{m^3_{D_{s}} E_\gamma }{m_{K^*} m_\rho} {\cal A}^{\rm pv}_{\rm VMD} \cdot \frac{2}{(m_{D_{s}}+ m_{K^*})^2} {V\over A_1} \ \ .\end{aligned}$$ Upon taking the form factors to be identical to those in $D^+ \ra \bar{K}^{*0}$ in the SU(3) limit, we have $${\cal A}^{\rm pv}_{\rm VMD} = 1.0\times 10^{-8}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}^{\rm pc}_{\rm VMD} = 0.9\times 10^{-8}~{\rm GeV}^{-1} \ \ .$$ Our VMD predictions for the magnitudes of the Cabibbo-suppressed transition amplitudes are summarized in Table 10. [l|cc]{}\ Mode &\ & Parity-conserving & Parity-violating\ $D^+ \to \rho^+\gamma$ & $1.6$ & $1.9$\ $D_s^+ \to K^{*+}\gamma$ & $0.9$ & $1.0$\ $D^0 \to \rho^0\gamma$ & $ 0.2-1.0 $ & $0.5-1.0$\ $D^0 \to \omega^0\gamma$ & $ 0.6 $ & $0.7$\ $D^0 \to \phi^0\gamma$ & $0.6-3.5$ & $0.9-2.1$\ [**Doubly Cabibbo-Suppressed Modes**]{} Finally, to estimate the size of the doubly Cabibbo-suppressed modes, we consider the $D^+ \ra K^{* +} \gamma$ transition. Upon computing the amplitudes using the factorization expression of Eq. (\[bsw\_rad\]), we obtain $${\cal A}_{\rm VMD}^{\rm pv}= 4.2\times 10^{-9}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}_{\rm VMD}^{\rm pc}= 4.4\times 10^{-9}~{\rm GeV}^{-1} \ . \label{dcsd1}$$ Similarly, we find for the mode $D^0 \ra K^{* 0} \gamma$ $${\cal A}_{\rm VMD}^{\rm pv}= 1.75 \times 10^{-9}~{\rm GeV}^{-1} \qquad {\rm and} \qquad {\cal A}_{\rm VMD}^{\rm pc}= 1.83 \times 10^{-9}~{\rm GeV}^{-1} \ . \label{dcsd2}$$ **Summary and Conclusions** =========================== The existing database for direct evidence of radiative $D$ decays is meagre, as has been shown in Table 1. However, interesting levels of experimental sensitivity are currently being attained and we can anticipate the detection of radiative signals in the not-too-distant future. Our motivation in undertaking the study reported here has been to stimulate such experimental efforts. As shown here, weak radiative decays of charmed mesons are not dominated by the short distance penguin diagrams of Figs. 1(a),1(c), but rather by long distance processes involving nonperturbative strong interaction dynamics. From the standpoint of probing the inner workings of the Standard Model, one might have naively hoped to use radiative charm decays in order to observe the short distance process $c\to u\gamma$. This would be in analogy with radiative $B$-meson decay, where the amplitude is dominated by the penguin transition $b\to s\gamma$ and receives a large enhancement from QCD radiative corrections. By contrast, the corresponding $c\to u\gamma$ charm transition is minuscule at lowest order and would require an unexpectedly large QCD enhancement to become detectable. To our knowledge, the calculation performed here of the QCD radiative correction to $c\to u\gamma$ is the first explicit and detailed analysis of this system given in the literature. In addition, we were able to employ $U$-spin arguments to clarify the role played by neutral, flavor-changing operators such as $\bar{u}c\bar{q}q$ which contribute to the expanded operator basis in the $RG$ analysis. Our conclusion that the $c\to u\gamma$ QCD radiative corrections are substantially larger than for $b\to s\gamma$ is due in part to the large operator mixing at the lower renormalization scale associated with the $c$-quark and in part to the disparate sizes of the Wilson coefficients at the matching scale of the contributing operators. Nevertheless, the radiatively-corrected $c\to u\gamma$ penguin transition remains extremely small. The main sources of suppression are the small quark masses and also the CKM factors $|V_{cb}^{*}V_{ub}|^2$ occurring in the numerator of the $c\to u\gamma$ branching fraction in Eq. (\[qcdchbr\]). In the Wolfenstein parameterization, with $\sin\theta_c=\lambda$, this CKM dependence amounts to a $\lambda^{10}$ suppression in decay rate. On the other hand, we have shown in Sections 3 and 4 that long distance contributions are several orders of magnitude larger. A very rough estimate of the typical branching ratio to be expected is B\_[DM]{}\~\_[em]{} B\_[DM]{}\^[non-lept.]{}  , \[esti\] where $B_{D\to M}^{\rm non-lept.}$ is the branching ratio for the nonleptonic $D$ decay to some final state $M$. Thus, typical branching ratios of order $B_{D\to M}^{\rm non-lept.}\sim (0.001-0.05)$ would induce radiative branching ratios in the range $B_{D\to M\gamma}\sim (7\times 10^{-6}\to 4\times 10^{-4})$. In Sections 3 and 4 we have performed a more detailed analysis by modeling the nonperturbative dynamics. Inspecting the long distance contributions to the set of exclusive processes $D^0\to\rho\gamma$, $D^0\to\omega\gamma$, $D_{s}\to K^{*+}\gamma$ and $D^+\to\rho^+\gamma$ for which $c\to u\gamma$ is the underlying transition, we see that the VMD and pole amplitudes carry a single factor $\lambda$, therefore representing an enhancement of $\lambda^4$ over the penguin amplitude. As can be seen in Table 11 the expected branching fractions for these modes are in the $10^{-6}-10^{-4}$ range, whereas we estimate $B_{c\to u\gamma}\sim 10^{-12}$. As a consequence, $c\to u\gamma$ is not a good process to test the validity of the Standard Model. That is, a hypothetical contribution from new physics would have to be extremely large in order to overcome the long distance physics. The situation is very different in radiative $B$ decays. The short distance transition $b\to s\gamma$ has the same CKM structure as the corresponding long distance contributions. For instance, the mode $B\to K^*\gamma$ might conceivably have long distance contamination of the order of $20\%$ in the rate $^{\cite{gp}}$. Although this is small compared to the charm case, it would be desirable to reduce the uncertainty in the calculation of these effects in order to subtract them from the measured signal. Moreover, long distance effects could also be affecting the inclusive $b\to s\gamma$ branching ratio, therefore limiting the precision with which the Standard Model can be tested in these decays. The various amplitudes are summarized in Table 11 and are given there in units of $10^{-8}~{\rm GeV}^{-1}$. In principle, the most conservative attitude is to take all relative signs as unknown, which would render the calculation of branching ratios highly uncertain. Fortunately, with the aid of the quark model we can reduce this ambiguity. The relative sign of Pole-II to Pole-I contributions is affected by (i) a minus sign difference in the pole denominators, (ii) an extra minus sign in type-II amplitudes due to the vector meson propagator, and (iii) minus sign differences in the $VP\gamma$ couplings between the $c$-quark and light-quark EM sectors. In a quark description of a $q_1 {\bar q}_2$ meson, this latter sign is inferred by studying $$h_{VP\gamma}= e \left(\frac{q_1}{m_1} + \frac{q_2}{m_2} \right) \ \ . \label{mu_qm}$$ Although this line of reasoning narrows down the range of predictions significantly, experimental data will be needed to obtain information regarding the relative phase between the pole and VMD contributions. In this regard, it will be helpful to note that, at least in our approach, the parity-violating amplitudes arise solely from the VMD process. [l||ccc|c|c]{}\ Mode & & ${\cal A}^{\rm pv}$ & $B_{D\to M\gamma}$ ($10^{-5}$)\ & P-I & P-II & VMD & VMD &\ $D^{+}_{s} \rho^+\gamma$ & $8.2$ & $-1.9$ & $\pm 3.2$ & $\pm 2.8$ & $6 - 38$\ $D^0 {\bar K}^{*0}\gamma$ & $5.6$ & $-5.9$ & $\pm 3.8$ & $\pm (5.1-6.8)$ & $7 - 12$\ $D^{+}_{s} b_1^+\gamma$ & $7.2$ & — & — & — & $\sim 6.3$\ $D^{+}_{s} a_1^+\gamma$ & $1.2$ & — & — & — & $\sim 0.2$\ $D^{+}_{s} a_2^+\gamma$ & $2.1$ & — & — & — & $\sim 0.01$\ $D^+ \rho^+\gamma$ & $1.3$ & $-0.4$ & $\pm 1.6$ & $\pm 1.9$ & $2 - 6$\ $D^{+} b_1^+\gamma$ & $1.2$ & — & — & — & $\sim 3.5$\ $D^{+} a_1^+\gamma$ & $0.5$ & — & — & — & $\sim 0.04$\ $D^{+} a_2^+\gamma$ & $3.4$ & — & — & — & $\sim 0.03$\ $D_s^+ K^{*+}\gamma$ & $2.8$ & $-0.5$ & $\pm 0.9$ & $\pm 1.0$ & $0.8 - 3$\ $D^{+}_{s} K_2^{*+}\gamma$ & $6.0$ & — & — & — & $\sim 0.2$\ $D^0 \rho^0\gamma$ & $0.5$ & $-0.5$ & $\pm (0.2 - 1.0)$ & $\pm (0.6 - 1.0)$ & $0.1-0.5$\ $D^0 \omega^0\gamma$ & $0.6$ & $-0.7$ & $\pm 0.6$ & $\pm 0.7$ & $\simeq 0.2$\ $D^0 \phi^0\gamma$ & $0.7$ & $-1.6$ & $\pm ( 0.6 - 3.5)$ & $\pm (0.9 - 2.1)$ & $0.1 - 3.4$\ $D^{+} K^{*+}\gamma$ & $0.4$ & $-0.1$ & $\pm 0.4$ & $\pm 0.4$ & $0.1 - 0.3$\ $D^{0} K^{*0}\gamma$ & $0.2$ & $-0.3$ & $\pm 0.2$ & $\pm 0.2$ & $\simeq 0.01$\ Finally, let us comment on the inclusive photon spectrum. In the $B$ system, the quark transition $b\to s\gamma$ provides a useful framework for predicting properties of the hadronic inclusivedecay $B\to X_s \gamma$. Thus, one estimates the $B\to X_s \gamma$ decay rate by computing the $b\to s\gamma$ decay rate and normalizing relative to the semileptonic decays to eliminate undue dependence on the mass $m_b$. Likewise, one predicts the photon energy spectrum in $B\to X_s \gamma$ decay by referring to the underlying two-body $b\to s\gamma$ decay.$^{\cite{fnl1},\cite{fnl2}}$ If quarks were free, there would be a monochromatic photon spike at $E_\gamma = (m_b^2 - m_s^2 )/2m_b \simeq m_b/2$. In reality, the photon spectrum becomes broadened via hadronization of the $s$-quark jet. The individual strange mesons ($K^*(892)$, $K_1(1270)$, [*etc*]{}) which populate the inclusive final state $X_s$ originate predominantly from the $s$-quark jet hadronization. These explanations of $B\to X_s \gamma$ inclusive decay are in accordance with the spectator model, and so isospin symmetry should manifest itself event-by-event. For example, the rates for isospin-related modes such as $B^0 \to K^{*0}\gamma$ and $B^- \to K^{*-}\gamma$ should be equal. A deviation from this pattern would constitute evidence for either non-spectator or new-physics contributions. In a heavy-quark effective theory description, such non-spectator effects would occur at subleading level. The theoretical description of charm inclusive decay could hardly be more different. Now, there is no emergent light-quark jet which hadronizes to form the set of final states. Instead, the ‘black box’ of long-range effects such as pole-amplitudes, VMD-amplitudes, [*etc*]{} dominates the physics. Thus, to determine the photon energy spectrum in $D\to X_u \gamma$, one would sum over the most important of the exclusive radiative modes. Presumably this would yield a reasonable description at least over the part of phase space where the photon energy is largest. It would be prudent to be on the lookout for the unexpected. For example, exclusive modes in light meson radiative decay are known to exhibit rather large isospin-violating effects, as in = 2.27 0.30 = 1.76 0.49  . \[c1\] If this effect were to be maintained mode by mode in the exclusive $D$ decays, it would lead to interesting levels of isospin violation in the inclusive decay. Of our results, Table 11 indicates that the likeliest possibility for isospin violation would appear to be in $D^0\to \rho^0\gamma /D^+\to \rho^+\gamma$. Charm radiative decays give us the opportunity to study various aspects of long distance dynamics. We have seen that the theoretical predictions of the branching fractions are, in some cases, rather uncertain due to model dependence. Experimental information will therefore be needed to complete the theoretical picture of these decays. It is an interesting irony that the understanding gained from future observation of different $D$ radiative decays can then be used to predict more confidently the size of such effects in $B$ decays. The research described in this paper was supported in part by the National Science Foundation and the Department of Energy. We gratefully acknowledge useful conversations with T.E. Browder, A. Datta, T. Rizzo, X. Tata and S. Willenbrock. We also wish to thank K.J. Moriarity for his assistance with numerical calculations. [99]{} \#1 \#2 \#3 [Mod. Phys. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Nucl. Phys. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rep. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rev. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Phys. Rev. Lett. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Rev. Mod. Phys. [**\#1**]{}, \#2 (\#3)]{} \#1 \#2 \#3 [Z. Phys. [**\#1**]{}, \#2 (\#3)]{} $CLEO$ collaboration, Phys. Rev. Lett. [**71**]{}, 674 (1993). M.S. Alam [*et al*]{} ($CLEO$ collaboration), ‘First Measurement of the Rate for the Inclusive Radiative Penguin Decay $b \to s \gamma$’, Cornell preprint CLNS-94-1314 (1994). E. Golowich and S. Pakvasa, ‘Uncertainties from Long Range Effects in $B\to K^* \gamma$’, Phys. Rev. D (to be published). D. Atwood and A. Soni, SLAC preprint SLAC-PUB-6635 (1994); E. Golowich and S. Pakvasa, B205 393 1988 ; H-Y. Cheng, Taipei preprint IP-ASTP-23-94 (1994). $CLEO$ collaboration, Phys. Rev. Lett. [**72**]{}, 1406 (1993). Particle Data Group, L. Montanet [*et al.*]{}, D50 1173 1994 . M. Selen, talk presented at APS Meeting, Washington DC, April 1994. J. Cumalet, in [*Proceedings of the Tau-Charm Factory in the Era of B Factories and CESR*]{}, SLAC-Report-451, ed. L.V. Beers and M.L. Perl, Stanford CA (1995) T. Inami and C.S. Lim, Prog. Theor. Phys. [**65**]{}, 297 (1981). N. Cabibbo and L. Maiani, 79B 109 1978 . M. Ciuchini [*et al.*]{}, CERN preprint CERN-TH.7283 (1994); Nucl. Phys. [**B415**]{}, 403 (1994); Phys. Lett. [**B301**]{}, 263 (1993); G. Cella [*et al.*]{}, B325 227 1994 ; M. Misiak, B393 23 1993 ; B269 161 1991 ; R. Grigjanis [*et al.*]{}, B224 209 1989 . B. Grinstein, R. Springer and M. Wise, B202 128 1988 ; B202 128 1988 . A.J. Buras [*et al*]{}, B424 374 1994 . B. Pietrzyk, ‘LEP Asymmetries and Fits of the Standard Model,’ talk presented at [*29th Rencontres de Moriond: Electroweak Interactions and Unified Theories*]{}, Meribel les Allues, France, March 1994, hep-ex/9406001. N.G. Deshpande, [*et al.*]{}, 59 183 1987 ; S. Bertolini, F. Borzumati, and A. Masiero, [*ibid.*]{} [**59**]{}, 180 (1987). F.J. Gilman and M.B. Wise, D20 2392 1979 ; C. Dib, I. Dunietz, and F.J. Gilman, A6 3573 1991 ; C.O. Dib, Ph.D. thesis, SLAC-Report-364 (1990); A.J. Buras, [*et al.*]{}, B370 69 1992 ; [*ibid.*]{} [**B375**]{}, 501 (1991). W. Bernreuther, C20 331 1983 . E. Golowich, D24 676 1981 . M. Wirbel, B. Stech and M. Bauer, Z. Phys. [**C29**]{}, 637 (1985); [*ibid.*]{} [**C34**]{}, 103 (1987). M. Wirbel, Prog. in Part. and Nucl. Phys. [**21**]{}, 33 (1988). MARK III Collaboration, Phys. Rev. Lett. [**60**]{}, 1375 (1988). C.W. Bernard, ‘Heavy-light and light-light weak matrix elements on the lattice’, review presented at $Lattice\ '93$, Dallas, Oct. 12-16 (1993). C.A. Dominguez, ‘Leptonic Decay Constants of Charm and Beauty Mesons in QCD: An Update’, invited talk presented at [*Third Workshop on the Tau-Charm Factory*]{}, Marbella, Spain (1993). J.F. Amundson [*et al*]{}, Phys. Rev. [**D47**]{}, 3059 (1993). J. Shigemitsu, ’Lattice Gauge Theory: Status Report 1994’, Ohio State preprint (1994), hep-ph/9408328. D. Acosta [*et al*]{} (CLEO Collaboration), Phys. Rev. [**49**]{}, 5690 (1994). S. Aoki [*et al*]{} (WA75 Collaboration), Prog. Theor. Phys. [**89**]{}, 131 (1993). J.Z. Bai [*et al*]{} (BES Collaboration), ‘A Direct Measurement of the $D_s$ branching Fraction to $\phi\pi$’, SLAC-PUB-95-6747 (1995). See Chap. VII of J.F. Donoghue, E. Golowich and B.R. Holstein, [*Dynamics of the Standard Model*]{}, Cambridge University Press (1992). Kenneth J. Moriarity, private communication. T.M. Aliev, E. Iltan and N.K. Pak, B334 169 1994 . P. Colangelo, F. De Fazio and G. Nardulli, B334 175 1994 . P.J. O’Donnell and Q.P. Xu, B336 113 1994 . Fayyazuddin and Riazuddin, B337 189 1994 . E. Paul, [*Proc. 1981 Intl. Symp. on Lepton/Photon Interactions at High Energies*]{}, ed. W. Pfiel (Bonn, 1981) pp. 301-333. For example, see R. Aleksan, A. Le Yaouanc, L. Oliver, O. Pène and J.-C. Raynal, ‘Critical Analysis of Theoretical Estimates for $B$ to Light Meson Form Factors and the $B \to \Psi K (K^*)$ Data’, Orsay preprint LPTHE-Orsay 94/13 (1994). Examples of analysis of this type appear in A.N. Kamal, R.C. Verma and N. Sinha, Phys. Rev. [**D43**]{}, 843, (1991); M. Gourdin, A.N. Kamal, Y.Y. Keum and X.Y. Pham, Preprint PAR/LPTHE/94-30. , see p. 1569 of Ref. [@pdg]. See the mechanism described in J.F. Donoghue, Phys. Rev. [**D33**]{}, 1516, (1986). A. Ali and C. Greub, Phys. Lett. [**B259**]{}, 182 (1991); Z. Phys. [**C49**]{}, 431 (1991); [*ibid.*]{} [**C60**]{}, 433 (1993). M. Neubert, Phys. Rev. [**D49**]{}, 4623 (1994). M. Gell-Mann, D. Sharp and W.G. Wagner, Phys. Rev. Lett. [**8**]{}, 261 (1962). The original application of VMD for analyzing hadronic radiative decays occurred in the light meson sector.$^{\cite{gmsw}}$ In order to test the VMD method using an up-to-date database, we too shall consider (briefly) light meson radiative decays in this Appendix. As we shall see from our study of two particularly clean examples, the situation is encouraging but not uniformly so. First, we shall revisit the original arena for testing VMD, the $\rho$ and $\omega$ decays into pion-photon final states. Then we shall analyze decays of a higher mass state, the tensor meson $A_2^+ (1320)$. We stress that in each of these cases the transition is purely electromagnetic, unlike the more complicated electroweak decays treated in the main body of the paper. Therefore, the ‘pole’ amplitudes do not occur here since there is no weak mixing, so one obtains a clean look at the VMD contribution. [**Radiative Decays of the Vector Mesons $\rho$ and $\omega$**]{} There are three electromagnetic $P$-wave decays in the $\rho$ – $\omega$ system, \^0 , \^+ \^+ \^0 \^0  . \[a10\] In the VMD approach, these are described in terms of two electromagnetic mixing amplitudes, $\omega$–$\gamma$ and $\rho$–$\gamma$, and one strong interaction vertex, $g_{\omega\rho\pi}$. Due to the off-shell nature of the VMD amplitudes, different momentum regions occur in the $\omega\rho\pi$ vertex for the transitions of Eq. (\[a10\]). In $\omega \to \pi\gamma$, the intermediate $\rho$ propagates at $q^2 = 0$ whereas for $\rho \to \pi\gamma$ it is the intermediate $\omega$ which propagates at $q^2 = 0$. Part of the VMD folklore built up over the years is that extrapolation of the light vector meson squared-momenta from the meson mass-shell to the photon mass-shell does not strongly affect the decay amplitude. The ratio of $\rho$ and $\omega$ decay widths can be used to test this as follows. Recall that for the VMD description of $1^- \to 0^- \gamma$ transitions, the strong vertex $g_{\omega\rho\pi}$ is related to the decay width $\Gamma$ via g\_ = \^[1/2]{}   , \[a11\] where $f \to f_\rho$ for $\omega$ decay and $f \to f_\omega$ for $\rho$ decay. Noting that the decay momenta in $\omega \to \pi\gamma$ and $\rho \to \pi\gamma$ are almost equal, one has |[f\_f\_]{}| = = 3.24 0.19  , \[a12\] provided the [*same*]{} strong vertex is used in each decay. In the above, we have used the charged-$\rho$ decay width in view of its superior accuracy. The value appearing in Eq. (\[a12\]) is seen to be in accord with that inferred from vector meson decay into lepton pairs ([*cf*]{} Table 1 of Ref. [@gp]), |[f\_f\_]{}| = 3.39 0.10   . \[a13\] Alternatively, one can use each of these radiative decays to extract determinations of $g_{\omega\rho\pi}$ as in Eq. (\[a11\]), and one finds g\_= { [ll]{} (11.73 0.35) [GeV]{}\^[-1]{} & (\^0)\ (12.40 0.64) [GeV]{}\^[-1]{} & (\^+ \^+ )\ (16.40 2.1) [GeV]{}\^[-1]{} & (\^0 \^0 )   . . \[a14\] The $\omega \to \pi^0\gamma$ and $\rho^+ \to \pi^+ \gamma$ determinations are seen to be consistent within experimental error. This is significant because these decays involve different momentum extrapolations as discussed above. The larger coupling obtained from $\rho^0 \to \pi^0 \gamma$ decay has substantially larger errors. We now turn to a different transition in which, if one accepts the data at face value, a non-negligible momentum dependence is present. [**Decays of the Tensor Meson $A_2^+ (1320)$**]{} The meson $A_2^+ (1320)$ has been observed to decay into both the $\pi \rho$ and $\pi \gamma$ modes, with branching ratios \_[A\_2 ]{} = 0.701 0.027 \_[A\_2 ]{} = (2.8 0.6)10\^[-3]{}   . \[a1\] These data turn out to provide a particularly clean test of the VMD method in two respects. First, there is just a single partial wave in the final state. As a consequence, the decay rates alone can be used to test VMD without any need for polarization information of the final state particles. The occurrence of a single orbital angular momentum in the final state follows from conservation of parity and of angular momentum. Thus we have :+ &=& (-)\^2 (-)\^L   L = 0,2,4,…\  [J]{}:|[**2**]{}| &=& |[**1**]{} + [**L**]{}|       L = 1,2,3 which implies that $L=2$. In addition, of the three light vector mesons $\rho, \omega, \phi$, only the $\rho$ can appear together with a pion in a final state of $A_2$ decay. The reason is that the decay $A_2 \to \pi ~V$ ($V$ is a vector meson) proceeds through the strong interactions and conservation of $G$-parity forbids the $\pi\omega$ and $\pi\phi$ modes. Thus, the rho is the only light vector meson involved in the VMD determination and interference with $\omega$ or $\phi$ mediated processes is absent. The amplitude for the transition $A_2^+ (p)\to \pi(q)^+ \rho^0(k)$ can be written as \_ = [g\_m\_A\^2]{}\^ p\_\^\_(k) q\_h\_(p) q\^ , \[a3\] where $g_{\pi\rho}$ is a dimensionless quantity and $h_{\beta\sigma}(p)$ is the spin-two polarization tensor of the $A_2$. From the decay rate relation, \_[A\_2\^+ \^+\^0]{} = [g\^2\_40]{} [[**q**]{}\_\^5 m\_A\^4]{}   , \[a4\] one determines a magnitude for the coupling $g_{\pi\rho}$. This can be used, in turn, to predict the radiative coupling $g_{\pi\gamma}$ via the VMD formula g\^[VMD]{}\_ = [ef\_]{} g\_   , \[a5\] and we find g\^[VMD]{}\_ = 1.99 0.06   . \[a6\] Alternatively, it is possible to determine the pion-photon coupling [*directly*]{}. Analogous to Eq. (\[a3\]), we can write down a gauge-invariant photon-emission transition, \_ = [g\_m\_A\^2]{}\^p\_\^\_(k) q\_h\_q\^ . \[a7\] Fixing the coupling $g_{\pi\gamma}$ in terms of the decay rate \_[A\_2\^+ \^+]{} = [g\^2\_40]{} [[**q**]{}\_\^5 m\_A\^4]{}   , \[a8\] yields the value g\^[expt]{}\_ = 0.98 0.11   . \[a9\] Thus, one obtains a factor-of-2 discrepancy between the empirical amplitude and the VMD prediction, with the VMD value being the larger. Several possible explanations for the lack of agreement come to mind. Although the radiative branching ratio given in Eq. (\[a1\]) has reasonably small error bars, the signal is based on only one experiment. Alternatively, there may be unexpectedly large momentum dependence in the $A_2\pi\rho$ vertex. Thus, as one proceeds from the rho mass-shell ($k^2 = m_\rho^2$) to the photon mass-shell ($k^2 = 0$), a ’softening’ might occur in the VMD estimate. However, to our knowledge there is no previous evidence for such momentum dependence for the $\rho$ extrapolation. [^1]: The Particle Data Group also lists branching ratios of $(3.0 \pm 0.6) \%$ and $(2.1 \pm 0.6)\%$ for $S$-wave and D-wave respectively.$^{\cite{pdg}}$ These values are completely consistent with the fact that the total transverse mode (which must be entirely $S$-wave by the absence of any $P$-wave) is $(1.6 \pm 0.5)\%$ and that the $S$-wave (longitudinal) must cancel with the D-wave to produce the net zero longitudinal branching ratio.

ArXiv

--- abstract: 'We present an updated optical and mechanical design of NEWS: the Near-infrared Echelle for Wide-band Spectroscopy (formerly called HiJaK: the High-resolution J, H and K spectrometer), a compact, high-resolution, near-infrared spectrometer for 5-meter class telescopes. NEWS provides a spectral resolution of 60,000 and covers the full 0.8–2.5 $\mu$m range in 5 modes. We adopt a compact, lightweight, monolithic design and developed NEWS to be mounted to the instrument cube at the Cassegrain focus of the the new 4.3-meter Discovery Channel Telescope.' author: - 'Mark J. Veyette' - 'Philip S. Muirhead' - 'Zachary J. Hall' - Brian Taylor - Jimmy Ye bibliography: - 'report.bib' title: 'NEWS: the near-infrared Echelle for wideband spectroscopy' --- Introduction {#sec:intro} ============ High-resolution, near-infrared (NIR) spectroscopy enables an enormously broad range of scientific studies (see Ref.  and references therein). However, relatively few facility-class, high-resolution, NIR spectrometers currently exist. Most are only available on large, heavily subscribed 8-to-10 meter class telescopes, such as NIRSPEC on the 10-meter Keck II Telescope[@McLean1998], CRIRES on the 8.2-meter VLT UT 1 Telescope[@Kaeufl2004], and IRCS on the 8.2-meter Subaru Telescope[@Tokunaga1998; @Kobayashi2000]. High-resolution, NIR spectrometers for 3–5-meter class telescopes like Lowell Observatory’s new 4.3-meter Discovery Channel Telescope[@Levine2012] (DCT) in Happy Jack, Arizona would provide greater accessibility for this powerful yet under-utilized tool for astronomy. Offering continuous, wide-band coverage has been an obstacle for high-resolution, NIR spectrometers. The number of resolution elements ($\Delta\lambda$) within the free spectral range (FSR) of a single order of a grating-based spectrometer is given by $$\frac{\mathrm{FSR}}{\Delta\lambda} = \frac{\lambda N}{\phi D},$$ where $\lambda$ is the wavelength at the center of the order, $N$ is the number of illuminated grooves, $\phi$ is the angular width of the slit, and $D$ is the diameter of the telescope. Traditionally, higher resolution is achieved by increasing $N$ by using a larger grating or a grating with more closely spaced grooves. In the IR, the number of resolution elements per order quickly becomes too large to fit a full order across a single detector with at least two pixels per resolution element to fully sample the spectrum. Immersion gratings provide one path to high resolution while maintaining small free spectral ranges. Spectral resolution increases linearly with the index of refraction of the medium the grating is immersed in. IGRINS[@Yuk2010] and iSHELL[@Rayner2012] both achieve high-resolution ($R >$ 40,000) across the NIR through the use of a silicon ($n=3.4$) immersion grating. However, silicon is not transmissive below 1.2 $\mu$m. At constant resolution, the 0.8 to 1.2 $\mu$m range accounts for over 35% of the information content in the 0.8 to 2.5 $\mu$m range. As we discuss in Section \[sec:science\], Y-band around 1 $\mu$m is a requirement for our primary science goal. Resolution also increases linearly with $\tan(\delta)$, the tangent of the blaze angle. With a high-blaze Echelle grating, high resolution can be achieved in a format that can be imaged by a single 2k$\times$2k detector. Here we present an optical design for a high-resolution NIR spectrograph called NEWS: the Near-infrared Echelle for Wide-band Spectroscopy. The design is based on a high-blaze R6 ($\tan(\delta) = 6$) Echelle grating and achieves a resolution of 60,000 over the full 0.8–2.5 $\mu$m range. The photometric z, Y, J, H, and K bands can be observed in their entirety without gaps. Scientific Motivation {#sec:science} ===================== The primary science goal of NEWS is to measure the abundances of individual elements in M dwarf stars. One application of this capability is to measure the ages of field M dwarfs. M dwarf stars are the most common class of star in the Galaxy, accounting for 70% of all stars [@Bochanski2010]. Their small radii and low effective temperatures ($2500 \mathrm{K} < T_\mathrm{Eff} < 3800 \mathrm{K}$) make nearby M dwarfs well suited for the detection of small, potentially habitable planets. Results from NASA’s *Kepler* Mission suggest M dwarfs are teeming with planets with $\sim$1 rocky planet per M dwarf with a period $<$150 days [@Morton2014; @Dressing2015]. NASA’s *Transiting Exoplanet Survey Satellite* (*TESS*) is expected to discover over 400 Earth-sized planets around nearby M dwarfs, including $\sim$50 orbiting within their host stars’ habitable zones [@Sullivan2015]. M dwarfs’ low effective temperatures make them difficult to characterize due to the formation of molecules throughout their atmospheres. Their visible and NIR spectra are dominated by millions of molecular lines that blend together even at high resolution. These molecular features render useless the standard methods developed for inferring fundamental parameters like $T_\mathrm{Eff}$, surface gravity, and chemical composition of Sun-like stars. The age of an M dwarf is perhaps the most challenging fundamental parameter to measure but enables a wide variety of stellar and exoplanet science. M dwarfs are known to host a wide range of planetary architectures: single short-period gaseous planets (e.g. GJ 1214 b [@Charbonneau2009]), compact multiple systems (e.g. *Kepler*-42, *Kepler*-445, and *Kepler*-446 [@Muirhead2012a; @Muirhead2015]), single rocky planets, and multi-planet systems with a wide range of planet masses. Ages of the host M dwarfs in these systems would answer outstanding questions on planet formation and evolution. It is an open question whether the scarcity of short-period gaseous planets is due to lower disk surface density around M dwarfs [@Dressing2013] or if they are a short-lived evolutionary state, evaporated by high UV flux [@Lopez2014]. Theorists disagree on the timescales for orbital evolution of short-period planets orbiting M dwarfs. Ref.  argue the tidally-induced eccentricity-damping timescale for short-period, low-mass planets is small, such that they should be circularized by 1 Gyr. However, Ref.  argue that eccentricity-damping is coupled to semi-major axis damping, extending the timescale to many billions of years. These competing hypotheses are testable given a means to measure M dwarf ages. Unlike solar-type stars, main-sequence M dwarfs move imperceptibly on a color-magnitude diagram. Gyrochronology, or the study of stellar spin-down versus age, holds some promise for measuring M dwarf ages. However, recent studies of M dwarfs with age-dated white dwarf companions suggest that M dwarfs do not spin down efficiently [@Morgan2012], and can hold onto their rapid rotation for billions of years [@West2015; @Newton2016]. ![\[fig:afe\_feh\_age\] The [\[$\alpha$/Fe\]]{}-\[Fe/H\]-age relation based on nearby red giants (data from Ref. ). This relation can provide ages for field M dwarfs given a method to measure [\[$\alpha$/Fe\]]{} in M dwarfs.](fig1.pdf){width="3.5in"} Our knowledge of the chemical evolution of the Galaxy provides a novel tool for estimating the ages of field M dwarfs. Early in the life of the Galaxy, core-collapse supernovae enrich the interstellar medium with the [$\alpha$]{}-elements O, Ne, Mg, Si, S, Ca, and Ti. Subsequently, type Ia supernovae contribute large amounts of Fe and the relative abundance of [$\alpha$]{}-elements to Fe ([\[$\alpha$/Fe\]]{}) decreases. Surveys of nearby Sun-like stars and red giants find an empirical relation between [\[$\alpha$/Fe\]]{}, \[Fe/H\], and age [@Haywood2013; @Bensby2014; @Feuillet2016]. Figure \[fig:afe\_feh\_age\] shows the [\[$\alpha$/Fe\]]{}-\[Fe/H\]-age relation for nearby red giants. We determined [\[$\alpha$/Fe\]]{} and \[Fe/H\] can be used to measure ages to an accuracy of $\pm1$ Gyr root-mean-square. This relation provides a new, powerful tool for estimating ages of field M dwarfs, given methods to measure their [\[$\alpha$/Fe\]]{} and \[Fe/H\]. The correlation between [\[$\alpha$/Fe\]]{}, \[Fe/H\], and age is not perfect given the chaotic nature of star formation and availability of pristine gas even in the late stages of galaxy evolution. Ref.  found that 6% of nearby red giants with [\[$\alpha$/Fe\]]{} $>$ 0.13 are younger than 6 Gyr. Nevertheless, measurements of [\[$\alpha$/Fe\]]{} and \[Fe/H\] can be combined into a powerful statistical tool for estimating ages of field M dwarfs. Recently multiple methods for measuring metallicity of M dwarfs have been empirically calibrated via widely-separated binary systems composed of an M dwarf with an FGK companion. The two stars are assumed to have formed at the same time, from the same material and, therefore, share a common chemical composition. Metal-sensitive indicators in M dwarf spectra can be calibrated on metallicities measured from the FGK companion. Methods have been developed to measure M dwarf metallicity from high-resolution NIR spectra [@Onehag2012; @Lindgren2016], high-resolution optical spectra [@Pineda2013; @Neves2014; @Maldonado2015], moderate-resolution NIR spectra [@Rojas2010; @Rojas2012; @Terrien2012; @Mann2013a; @Newton2014], and optical-NIR photometry [@Bonfils2005; @Casagrande2008; @Johnson2009; @Schlaufman2010; @Neves2012; @Johnson2012; @Hejazi2015]. Precise methods for measuring both [\[$\alpha$/Fe\]]{} and \[Fe/H\] are needed to estimate ages of field M dwarfs. Currently, no method exists to measure [$\alpha$]{}-enhancement. Based on `PHOENIX` BT-Settl synthetic spectra, we found that Y-band around 1 $\mu$m contains numerous [$\alpha$]{}-sensitive lines including many deep lines. is a rapid decay product of , a product of the [$\alpha$]{} process during core collapse supernovae. Figure \[fig:alphalines\] shows synthetic spectra of M dwarfs with varied abundance of the major [$\alpha$]{} tracers Mg, Si, and Ti. Y-band contains numerous isolated lines whose strength correlate with [\[$\alpha$/Fe\]]{}. However, these atomic lines are embedded within molecular absorption bands of TiO and FeH. High spectral resolution is required to sufficiently isolate atomic lines to measure [\[$\alpha$/Fe\]]{} in M dwarfs. ![\[fig:alphalines\] Psuedo-continuum normalized synthetic M dwarf spectra ($T_\mathrm{eff}$=3000 K, log($g$)=5.0, \[M/H\]=0.0) smoothed to $R$ = 60,000. Top: A representative sample of deep, [\[$\alpha$/Fe\]]{}-sensitive lines in Y-band for models with varied [\[$\alpha$/Fe\]]{}. Bottom: K-band metal lines and the CO 2-0 bandhead for models with varied C/O. High-resolution Y-band and K-band observations can be used to measure [$\alpha$]{}-abundance in M dwarf stars and calibrate out any effects of C and O abundance.](fig2.pdf){width="\linewidth"} It has recently been discovered that the relative abundances of carbon and oxygen strongly affect the psuedo-continuum level throughout M dwarf spectra [@Veyette2016]. In order to accurately measure [\[$\alpha$/Fe\]]{} in M dwarfs, we must first calibrate out any effect of C and O abundances (or their ratio[^1] C/O) on the equivalent widths (EWs) of lines in Y-band. Ref.  found that high-resolution observations of and CO lines in K-band can be used measure C and O abundances in M dwarfs. Figure \[fig:alphalines\] shows synthetic spectra of M dwarfs with varied C/O. In the atmosphere of an M dwarf, nearly all the C is locked away in energetically favorable CO, along with an equal amount of O. At $T_\mathrm{eff} < 3300$ K, the majority of the remaining O is found in . The strength of the 2.3 $\mu$m CO 2-0 bandhead and the 1.9 $\mu$m band can be used to measure C/O. The science case above led us to the following design requirements. NEWS must achieve high-resolution ($R >$ 30,000) and cover a large portion of the NIR window, from Y-band to K-band. We find that a resolution of $R$ = 60,000 provides a good balance between isolating atomic lines in M dwarf spectra but still allowing broadband observations in a single exposure. The design must also offer high throughput ($>$ 10%) in order to achieve high enough signal-to-noise to measure small changes in the EWs of atomic lines in M dwarf spectra. This requirement led us to use a slit-fed design mounted directly to the telescope as opposed to a bench-mounted, fiber-fed design which can be limited by modal noise (e.g. GIANO[@Origlia2014]). Although the ability to measure individual elemental abundances and ages of M dwarfs has dictated many of the design requirements for NEWS, we adopted an overall facility-class philosophy. Specifically, we designed NEWS to cover the full 0.8–2.5 $\mu$m range without gaps and offer multiple slit widths and lengths. We also developed NEWS to be extremely compact and lightweight so that the design can be implemented at nearly any 3–5 meter class telescope. Optical Design {#optics} ============== From Telescope to Detector -------------------------- We show the full NEWS optical layout in Figure \[fig:optics\] and list the key properties of the design in Table \[tab:news\]. The converging f/6.1 beam delivered from the telescope enters the cryostat through a fused silica window. The beam comes to a focus at one of six slits accessible by a slit wheel before passing through one of five order-selecting filters accessible by a filter wheel. The beam is then folded onto an Offner relay that serves as a cold pupil stop. After exiting the Offner relay, the beam goes through a second focus and expands to a diameter of 4 cm before being collimated by a parabolic mirror used off-axis. The collimated beam is reflected by an R6 Echelle grating that is tilted from Littrow by an in-plane angle of $\theta$ = 2 and by an off-plane angle of $\gamma$ = 1.5. The dispersed beam is refocused by the same parabolic mirror and is folded by a Mangin mirror to remove aberrations introduced by dispersing onto the parabolic mirror and direct the beam to the conjugate point of the same parabolic mirror. The re-collimated beam is cross-dispersed by a grating before being focused by an 8-element, all-spherical camera onto a Hawaii 2-RG detector. ![\[fig:optics\] CAD rendering of the NEWS optical layout.](fig3.png){width="3.5in"} ### Slit Selection The NEWS design equips a motorized wheel to cycle through six different slit options. The width options for the slits are either 05, 10, or 15 in the dispersion direction to enable observations during nights with poor seeing at reduced resolution. The silts are either 5 or 9 long. In order to fully separate adjacent orders, only the 5 slit can be used with the 6th and 7th orders of the cross-disperser (corresponding to z- and Y-band). Slits will be laser-cut into 2-inch diameter gold-coated silicon wafers. Tilted slit substrates direct a $\sim$6 arcmin$^2$ field back out of the cryostat where a slit-viewing camera images it onto an InGaAs detector. ### Collimation A monolithic parabolic mirror collimates the beam for both the Echelle and cross-disperser. The NEWS design employs a Mangin mirror to remove aberrations introduced by dispersing onto the parabolic mirror and to steer the beam returned by the Echelle to the conjugate point of the paraboloid for collimation onto the cross-disperser. The Mangin reflector is a spherical lens with curvature on both sides and a reflective coating on one side. ### Primary Dispersion We designed NEWS around a new R6 Echelle recently developed by Richardson Gratings with a coarse groove spacing and high blaze angle (13.3 grooves/mm, blazed at 80.5). The high blaze angle and coarse grooves enable high spectral resolution with a manageable number of resolution elements per order so that the entire echellogram can be imaged by a single square detector with sufficient sampling. The Echelle is tilted in-plane by $\theta$ = 2 in order return the full 0.8–2.5 $\mu$m range without gaps. The in-plane tilt reduces the peak efficiency by $\sim$30%. Th Echelle is also titled off-plane by $\gamma$ = 1.5 to separate the incoming and outgoing beams at the intermediate focus. ### Cross Dispersion For cross-dispersion, we utilize the same grating that is used as the primary dispersive grating in TripleSpec[@Herter2008] and GNIRS[@Elias1998]. The grating has 110.5 groves/mm and is blazed at 22. Orders 7, 6, 5, 4, and 3 of the cross-disperser correspond roughly to the z, Y, J, H, and K atmospheric windows, respectively. Only a single cross-dispersion order is accessible at a time and is selected by order-blocking filters. ### Spectrograph Camera We designed an 8-element, all-spherical f/2.6 camera to image the spectrum onto a 2k$\times$2k Teledyne Hawaii 2-RG detector with a 2.5 $\mu$m cutoff. The camera consists of two , one Infrasil, one Fused Silica, and 4 ZnSe lenses. Figure \[fig:camera\] shows the optical layout of the camera. ![\[fig:camera\] Camera Optical layout.](fig4.png){width="\linewidth"} [ | p[1.8in]{} | p[=4.25in]{} | ]{}\ Resolution ($\lambda / \Delta \lambda$) & 60,000 for 05 slit width, 30,000 for 10 slit width\ Wavelength coverage & 5 bands: Either 0.80-1.0 ($z$), 1.00-1.20 ($Y$), 1.20-1.45 ($J$), 1.45-1.85 ($H$) or 1.85-2.5 $\mu$m ($K$), selectable by filter wheel\ Slit sizes on sky & 6 unique sizes: either 05, 10 or 15 wide (dispersion directions), and either 90 or 50 long, laser cut into Au-coated Si substrates, selectable by slit wheel. Optimal slit length choice depends on band due to x-dispersion. One substrate lacks a slit for acquiring cold darks.\ Beam size on Echelle & 1.57 inches (4 cm)\ End-to-end throughput & 5% requirement, 10% goal\ DCT Properties & 4.3-meter diameter, Alt/az, Ritchey–Chretien, f/6.1 at Cassegrain, 127 $\mu$m per arcsecond\ Field Rotation & Instrument mount rotation built into DCT Cassegrain cube[@Bida2012]\ Atm. Disp. Correction & None (slit can be rotated to parallactic angle via the Cassegrain cube)\ Slit Viewer Camera & COTS InGaAs detector and short-wave infrared (SWIR) lens system external to cryostat operated in $J$ band (similar to NIHTS[@Bida2014])\ Guiding & Native DCT Off-axis guider\ Cold stop & Offner relay (identical to NIHTS[@Bida2014])\ Echelle & Newport/Richardson 53-\*-182E operated in quasi-Littrow (13.33 l/mm, 80.54Blaze)\ Cross-disperser & Newport/Richardson 53-\*-138R operated with in-plane tilt (110.5 l/mm, 22Blaze, identical to TSPEC[@Herter2008] and GNIRS[@Elias1998])\ Spec. Camera & 8-element, all spherical, f/2.6, , ZnSe, Infrasil, and fused silica lenses\ Spec. Detector & 2.5-$\mu$m-cutoff Teledyne Hawaii-2RG (2048x2048 18-$\mu$m pixels)\ Spec. Detector Electronics & Teledyne SIDECAR or Leach and custom detector interface board\ Sampling & 2.0 pixels per 05 resolution element (at R=60,000)\ Cryostat & Box-style, aluminium (e.g. Atlas Ultrahigh Vacuum)\ Optical Bench & 39$\times$25$\times$12 in (99$\times$64$\times$30 cm) custom honey-combed aluminum\ Optical Bench Temperature & Less than 100 Kelvin via a CTI-1050 cryodyne refrigeration system, 65 W max load, similar design to Mimir [@Clemens2007]\ Detector Temperature & Less than 78 Kelvin via a CTI-1050 cryodyne refrigeration system, similar design to Mimir [@Clemens2007]\ Total weight & $<$200 kg, within limit for the Cassegrain cube\ Optical Performance {#perf} ------------------- We optimized the camera lenses and Mangin fold mirror to minimize the RMS spot size across the detector. The largest source of aberration is wavelength-dependent field curvature introduced by dispersing onto the paraboloid, much of which is removed by the Mangin mirror. As a result, tolerances on the optical surfaces are quite large ($\sim$ 10 $\mu$m). In Figure \[fig:echellogram\], we show the full simulated echellogram and corner spot diagrams. RMS spot sizes are on the order of one 18 $\mu$m square pixel. ![\[fig:echellogram\] Ray-traced echellograms and corner spot diagrams for each observing mode of NEWS, selectable entirely by filter wheel. The filter selects an order of the cross-dispersing grating (orders 3 through 7), resulting in wavelength coverage from 0.8 to 2.5 $\mu$m. The large boxes represent the size of the full Hawaii-2RG detector (36.9 x 36.9 mm) and the small boxes represent the size of each pixel (18 x 18 $\mu$m). By using a Mangin reflector in combination with a conjugate paraboloid, aberrations introduced by the collimating paraboloid are reduced to roughly a pixel. The aberrations from the paraboloid dominate the RMS spot sizes so that the tolerances on the optical surfaces are large (typically 10 $\mu$m), except for the Offner relay (1 $\mu$m).](fig5.pdf){width="\linewidth"} Mechanical Design {#mech} ================= ![\[fig:mech\] CAD rendering of the full NEWS optomechanical design. The optical bench and all mounts will be custom machined from 6061-T6 Al.](fig6.png){width="4in"} Optomechanical design --------------------- Figure \[fig:mech\] shows the full NEWS optomechanical design. ### Optical bench To simplify mounting and maintain a compact design, we designed all the optical components of NEWS to lie in a single plane. The mounts for each optical component all attach directly to a monolithic optical bench milled from a single block of 6061-T6 aluminum. To minimize weight, but maintain rigidity, the optical bench is light-weighted with a honeycomb structure. The hexagons of the honeycomb are milled 2 inches deep with 1 inch sides (inner), leaving 0.3 inch thick walls and 1 inch of solid aluminum for the mounting surface. Figure \[fig:bench\] shows the light-weighted honeycomb structure. The optical bench will be surrounded by a low-emissivity radiation shield and mounted to an enclosing cryostat via G10 tabs. We applied finite element analysis (FEA) on the optical bench to ensure it meets mechanical tolerances ($<$ 10$\mu$m flex) under under the weight of the optical elements subject to varied gravity vectors. ![\[fig:bench\] CAD rendering of the NEWS custom honeycomb optical bench (the bottom side as shown in Figure \[fig:mech\]). An inset on the left shows a cross section of one of the honeycomb hexagons. The hexagons are 1 inch on a side (inner) with 0.3 inch thick walls and are milled 2 inches deep, leaving 1 inch of solid aluminum for the mounting surface.](fig7.pdf){width="4in"} ### Optics Mounts We designed all optical mounts to be machined out of 6061-T6 Al. To maintain optical positioning tolerances, we employ flexure supports on all optics to compensate for thermal contraction stress and to preload against a changing gravity vector. For example, the Offner relay primary mirror is constrained by a radial flexure ring with 16 EDM wire-cut spring restraints (see Figure \[fig:flexure\]). The design is identical to that used in NIHTS[@Bida2014]. ![\[fig:flexure\] CAD rendering of radial flexure ring mount for the Offner relay primary mirror. Wire-cut spring restraints compensate for thermal contraction stress and preload against a changing gravity vector.](fig8.png){width="2in"} Mounting to the DCT ------------------- We designed NEWS within the size and weight restrictions of mounting to one of the large ports on the DCT’s instrument cube at the Cassegrain focus of the telescope. The instrument cube rotates to maintain constant field alignment. Instruments mounted to the cube must clear the telescope mount supports at all rotation angles and the mount platform at all altitudes. Mounting directly to the telescope entails strict weight constraints. The total payload capability for the instrument cube is 1500 kg[@Smith2010b]. The large instrument ports each can support up to 360 kg[@Bida2012]. Our current design fits within a 99$\times$64$\times$30 cm in volume and weighs less than 100 kg (excluding the cryostat). ### Software System {#soft} Software control and interfacing to the TCS will be done with the Lowell Observatory LOIS[@Taylor2000]/LOUI systems which handle the current instrument suit on the DCT. This software system is a mature well defined control system that already has interfaces to the TCS, Guider, and AOS subsystems. If we decide that the ARC controller is the best match for NEWS then most of the interfacing to the detector has already been done for the NIHTS instrument, an H2RG NIR instrument and Mimir, an Aladdin III instrument. ![\[fig:ondct\] CAD rendering of of the NEWS mechanical design as it would be positioned on the DCT instrument cube. A cryostat containing the instrument would be mounted directly to one of the large ports on the instrument cube. The interference-free volume for the instrument cube is shown in transparent red. NEWS fits within the size constraints of mounting to the DCT. A six-foot-tall human figure is shown for scale.](fig9.png){width="5in"} Thermal Management {#therm} ------------------ Thermal modeling of NEWS shows that the predictive radiative load will be around 63 W at 300 K, assuming a polished aluminum cold shield wrapped in MLI and based on the current surface area of the instrument cold shield. Average high temperatures are 301 K for the hottest month of the year in Happy Jack, Arizona where DCT is located. This load can be easily handled by a CTI 1050 single stage cryodyne refrigeration unit which is capable of delivering 80 W of cooling at 78 K with enough overhead to handle any parasitic load from the wiring and the intermittent load of the motors. We will utilize thick copper strapping to the cold bench and shield with a much small copper line to the detector for the conductive paths for heat transfer. This line to the detector will have a small sapphire disk spacer to minimize the electrical noise transferred to the detector. We will monitor and control the instrument stability, cool down and warm up rates via a Lakeshore model 331 temperature controller. Summary ======= We have designed a high-resolution NIR spectrograph for the 4.3-meter Discovery Channel Telescope. NEWS achieves a resolution of $R$ = 60,000 over the full 0.8–2.5 $\mu$m range. Our design offers continuous coverage within each of its five observing modes corresponding to the photometric z-, Y-, J-, H-, and K-bands. If built, NEWS will be uniquely capable of measuring the composition and ages of field M dwarfs, including those who host planets detected by *TESS*. Support for this work was provided by the Department of Astronomy and the Institute for Astrophysical Research at Boston University. This research made use of the Massachusetts Green High Performance Computing Center in Holyoke, MA. [^1]: C/O is defined as $N_\mathrm{C}/N_\mathrm{O}$, where $N_\mathrm{C}$ and $N_\mathrm{O}$ are the number densities of carbon and oxygen, respectively.

ArXiv

--- abstract: 'Water cycling between Earth’s mantle and surface has previously been modelled and extrapolated to rocky exoplanets, but these studies neglected the host star. M-dwarf stars are more common than Sun-like stars and at least as likely to host temperate rocky planets (M-Earths). However, M dwarfs are active throughout their lifetimes; specifically, X-ray and extreme ultraviolet (XUV) radiation during their early evolution can cause rapid atmospheric loss on orbiting planets. The increased bolometric flux reaching M-Earths leads to warmer, moister upper atmospheres, while XUV radiation can photodissociate water molecules and drive hydrogen and oxygen escape to space. Here, we present a coupled model of deep-water cycling and water loss to space on M-Earths to explore whether these planets can remain habitable despite their volatile evolution. We use a cycling parameterization accounting for the dependence of mantle degassing on seafloor pressure, the dependence of regassing on mantle temperature, and the effect of water on mantle viscosity and thermal evolution. We assume the M dwarf’s XUV radiation decreases exponentially with time, and energy-limited water loss with 30% efficiency. We explore the effects of cycling and loss to space on planetary water inventories and water partitioning. Planet surfaces desiccated by loss can be rehydrated, provided there is sufficient water sequestered in the mantle to degas once loss rates diminish at later times. For a given water loss rate, the key parameter is the mantle overturn timescale at early times: if the mantle overturn timescale is longer than the loss timescale, then the planet is likely to keep some of its water.' author: - | Keavin Moore,$^{1,2}$[^1] Nicolas B. Cowan$^{1,2,3}$\ $^{1}$Department of Earth & Planetary Sciences, McGill University, 3450 rue University, Montréal, QC H3A 0E8, Canada\ $^{2}$McGill Space Institute, McGill University, 3550 rue University, Montréal, QC H3A 2A7, Canada\ $^{3}$Department of Physics, McGill University, 3600 rue University, Montréal, QC H3A 2T8, Canada date: 'Accepted XXX. Received YYY; in original form ZZZ' title: 'Keeping M-Earths Habitable in the Face of Atmospheric Loss by Sequestering Water in the Mantle' --- \[firstpage\] planets and satellites: atmospheres – planets and satellites: interiors – planets and satellites: tectonics – planets and satellites: terrestrial planets – planets and satellites: oceans – stars: low-mass Introduction {#sec:intro} ============ Habitability critically depends on the presence of liquid water on the surface of a planet. Earth is the only planet in the Universe with confirmed surface oceans and life as we know it, and as such, it is our template for a habitable planet [@langmuir12]. The habitable zone (HZ) around a star is usually defined by the stellar flux at a given orbital distance, which influences the surface temperature [@kasting93] — planets at the hot inner edge vaporize their oceans, while the oceans of a planet at the cold outer edge will freeze. Water is not only important for biological processes; it also influences planetary climate through the silicate weathering thermostat [@walker81; @sleep01; @abbot12; @alibert14]. It must be noted that the presence of liquid surface water may constitute habitability in the classical definition [@kasting93], but this does not necessarily mean the planet is hospitable and conducive to the origin of life. Rather, liquid surface water is a first step towards life as we know it, and our results only support the presence of liquid surface water and thus the classical definition of habitability. Recent studies find that the orbital environment around M dwarfs may lack the necessary levels of UV radiation to form RNA monomers, which are critical to the development of Earth-like biology [@ranjan17; @rimmer18]; however, the scarcity of UV may be overcome by transient flaring events. We must be careful in assuming liquid surface water means life due to the multitude of factors that allowed life to originate on the early Earth. Water Cycling ------------- It is speculated that there is at least as much water sequestered in the mantle as is present on the surface of Earth [@hirschmann06]. Current estimates based on experimental data put the water capacity of Earth’s mantle at 12 terrestrial oceans (TO, where 1 TO $\approx 1.4 \times 10^{21}$ kg; @hauri06; @cowan14). Water is exchanged between the surface and mantle reservoirs of Earth on geological timescales. This water cycle is mediated by plate tectonics, which depend not only on the cool, brittle lithospheric plates, but also on the viscous, flowing mantle (see, e.g., @hirschmann06; @langmuir12). Water dissolved in the mantle decreases its viscosity and allows it to flow more readily, a requirement for plate tectonics [@hauri06]. As plates separate from one another at mid-ocean ridges, the mantle below ascends to fill the gap and melts due to depressurization, and volatiles are released into the ocean by degassing as the new ocean crust solidifies. The ocean crust then spreads away towards subduction zones, and its minerals become hydrated due to hydrothermal interactions with seawater. This volatile-rich slab is then subducted back into the mantle. While most of the volatiles will contribute to water-rich magmas at convergent margin volcanoes, some water will continue into the deep mantle, regassing water into the interior. Water Loss to Space ------------------- Roughly 70% of stars in the Galaxy are M dwarfs [@henry04], and about 30% are host to at least one rocky exoplanet in the HZ [@dressing15]. We should then expect that 90-99% of temperate terrestrial planets orbit an M dwarf instead of a Sun-like star; henceforth, we call such planets M-Earths. M dwarfs are more active than Sun-like stars [@scalo07], specifically in the X-ray and extreme ultraviolet, collectively known as XUV. An M dwarf emits more XUV radiation during the first billion years of its lifetime, as the star evolves onto the main sequence. Young planets orbiting within the HZ may lose multiple oceans of water to space as they are bombarded by XUV radiation. Water molecules are photodissociated high in the planetary atmosphere [@wordsworth13; @wordsworth14; @luger15; @bolmont16; @schaefer16; @wordsworth18; @fleming20]; the lighter hydrogen is lost to space, while the heavier oxygen remains behind, either reacting with the surface or accumulating in the atmosphere and creating a biosignature false positive (see, e.g., @wordsworth13 [@wordsworth14; @wordsworth18]). Some oxygen may also hydrodynamically escape, dragged to space by the escaping hydrogen (e.g., @hamano13 [@luger15]). Studies also indicate that factors of 5–10 more XUV irradiation than modern-day Earth can lead to runaway atmospheric loss [@tian08], and that several Gyr-old planet-hosting M dwarfs may output 5–100 times more XUV radiation than the Sun today (e.g., @ribas16 [@ribas17]; @youngblood16). Moreover, it has recently been indicated that M dwarfs remain more active in the extreme-UV (EUV) for a given age than solar-type stars (see, e.g., Fig. 6 of @france18). We hypothesize that a planet whose surface becomes desiccated by loss of water to space can recover an ocean through the degassing of water sequestered in the mantle. This will depend on the initial amount of water partitioned between the surface and mantle, the mantle overturn timescale, and the XUV-driven water loss rate and timescale. While the deep-water cycle and atmospheric loss have been separately modelled in previous work, we seek to couple these two phenomena, combining aspects of geophysics, astronomy, and space physics. The paper continues as follows. We describe the cycling and loss equations of our model in Section \[sec:equations\], present our cycling results for various initial water inventories and loss rates in Section \[sec:results\], and discuss the results of our study in Section \[sec:discussion\]. Water Cycling & Loss Model {#sec:equations} ========================== Previous Work ------------- The deep-water cycle of Earth has previously been represented using two-box models. These models account for regassing of water from surface to mantle through subduction of hydrated basaltic oceanic crust, and degassing from mantle to surface by mid-ocean ridge volcanism. @mcgovern89 incorporated reduction of mantle viscosity through the addition of regassed water, while parameterizing degassing and regassing rates as dependent on the amount of volatiles present in the mantle and basaltic oceanic crust, respectively, along with mid-ocean ridge spreading rate and subduction efficiency. The mantle-temperature-dependent model of @schaefer15, based on the model of @sandu11, also included mantle viscosity and two convection regimes: single layer and boundary layer. @komacek16 simplified the mantle-temperature-dependent model of @schaefer15, and replaced the degassing rate with the seafloor-pressure-dependent degassing parameterization of @cowan14 to create a hybrid model. There are various water loss rates throughout the literature; for example, @wordsworth14 note that an N~2~-poor planet could lose up to 0.07 TO/Gyr, while loss rates from @luger15 range from 0.02 TO/Gyr to about 2 TO/Gyr, depending on initial water inventory and orbital distance within the HZ. Cycling & Loss Equations ------------------------ We use the time-dependent hybrid cycling model of @komacek16 and parameterize the water loss of @luger15 to represent the cycling and loss to space of water on an M-Earth. Our model accounts for the fact that hydration depth of ocean crust is likely affected by mantle temperature, $T$, more than seafloor pressure, $P$, and that degassing would shut off at late times when the mantle is cool. Meanwhile, it has been shown that degassing should be $P$-dependent [@kite09]. Our two-box $+$ sink model is shown schematically in Fig. \[fig:twoboxsink\], including surface and mantle reservoirs, exchange between the two, and water loss to space directly from the surface reservoir, for simplicity. Any changes or additions to the relevant thermal evolution and cycling equations from @komacek16 are described here and in Appendices \[sec:thermaleqns\] and \[sec:improvements\]. The thermal evolution and cycling equations were non-dimensionalized by @komacek16 to emphasize the physical processes over the control variables themselves. While we use the non-dimensionalized code for our simulations, we present the dimensionful equations here. Appendix \[sec:params\] contains a cheat sheet of all the model variables and parameters. ![Two-box model of water cycling between surface and mantle reservoirs on Earth, adapted from @cowan14 to include water loss to space (bolded). Water is degassed from the mantle to the surface through mid-ocean ridge volcanism, and regassed from the surface to the mantle through subduction of hydrated basaltic oceanic crust. Water is lost to space directly from the surface reservoir for simplicity, and is driven by XUV radiation from the host M dwarf, which decreases exponentially with time [@luger15].[]{data-label="fig:twoboxsink"}](2Box_CyclingCartoon_LosstoSpace.pdf){width="45.00000%"} The model developed by @komacek16 incorporates $P$-dependent degassing [@cowan14] and $T$-dependent regassing [@schaefer15]. The authors note that this hybrid model may be the most realistic deep-water cycling model of their study; for this reason, we use it as our representative water cycling parameterization. We restore the piecewise degassing limit of @cowan14; while this makes the equation harder to manipulate analytically, it ensures that a parcel of mantle cannot degas more water than it contains. The change in mantle water mass $W_{\mathrm{m}}$ with time $t$ is, $$\label{eqn:dx/dt_hyb} \begin{split} \frac{d W_{\mathrm{m}}}{dt} & = L_{\mathrm{MOR}} S(T) \Biggl[x_{\mathrm{h}} \rho_{\mathrm{c}} \chi_{\mathrm{r}} d_{\mathrm{h}}(T) \\ & - x \rho_{\mathrm{m}} d_{\mathrm{melt}} \min \biggl[ f_{\mathrm{degas},\oplus} \left(\frac{P}{P_\oplus} \right)^{-1}, ~1 \biggr] \Biggr], \end{split}$$ where the first term on the right-hand side is the regassing rate, $w_{\downarrow}$, and the second is the degassing rate, $w_{\uparrow}$. The other variables are as follows: $L_{\mathrm{MOR}}=3 \pi R_\mathrm{p}$ is the mid-ocean ridge length (with $R_{\mathrm{p}}$ the planetary radius), and $S(T)$ is the $T$-dependent spreading rate. The mass fraction of water in the hydrated crust is $x_{\mathrm{h}}$, $\rho_{\mathrm{c}}$ is the density of the crust, $\chi_{\mathrm{r}}$ is the subduction efficiency, and the hydrated layer depth is a function of mantle temperature, $d_{\mathrm{h}}(T)$. The mantle water mass fraction is $x$. The density of the upper mantle is $\rho_{\mathrm{m}}$, $d_{\mathrm{melt}}$ is the mid-ocean ridge melting depth, $f_{\mathrm{degas},\oplus} = 0.9$ is the nominal value of melt degassing for present-day Earth, $P$ is the seafloor pressure, and $P_\oplus$ is the seafloor pressure of Earth. The pressure dependence is defined as a power law, using the nominal value from @cowan14 for the exponent. The equivalent cycling equation for the surface water mass $W_{\mathrm{s}}$ is, $$\label{eqn:ds/dt_hyb} \begin{split} \frac{d W_{\mathrm{s}}}{dt} & = L_{\mathrm{MOR}} S(T) \Biggl[x \rho_{\mathrm{m}} d_{\mathrm{melt}} \min \biggl[ f_{\mathrm{degas},\oplus} \left(\frac{P}{P_\oplus} \right)^{-1}, ~1 \biggr] \\ & - x_{\mathrm{h}} \rho_{\mathrm{c}} \chi_{\mathrm{r}} d_{\mathrm{h}}(T) \Biggr] - \min \left[\phi_{\mathrm{loss}} \exp \left(\frac{-t}{\tau_{\mathrm{loss}}}\right),~ \frac{W_{\mathrm{s}}}{\tau_{\mathrm{step}}} \right]. \end{split}$$ Here, the first term on the right-hand side is now the degassing rate, $w_{\uparrow}$, and the second is the regassing rate, $w_{\downarrow}$. The third term on the right-hand side is the water loss rate, $w_{\mathrm{loss}}$, a decreasing exponential based on the M dwarf XUV evolution with time in Fig. 1 of @luger15, which utilizes the stellar models of @ribas05; note that we assume the XUV radiation simply decreases exponentially from its initial value, with the loss of water to space linearly correlated to this evolution. Note also we do not directly model the stellar evolution, nor do we account for a planetary magnetic field; as a result, we do not include flare- or stellar-wind-driven loss in this study. Recent studies support stellar-wind-driven ion pick-up escape leading to rapid, complete atmospheric erosion for planets orbiting ‘old’ (i.e., several Gyr) M dwarfs like Proxima Centauri b [@airapetian17; @garcia17], in the absence of a source of replenishment. Although ion escape is likely important for M-Earths, we do not include it in our current study. Instead, we solely focus on the energy-limited escape of @luger15, adopting the same efficiency of 30% to test similar loss rates. Our loss parameterization is piecewise-defined so that we do not lose more water than present on the surface in a given timestep. The exponential definition of loss to space stems from the exponential decrease of the M dwarf’s XUV luminosity, and includes a loss factor, $\phi_{\mathrm{loss}}$, and loss timescale, $\tau_{\mathrm{loss}}$. The former accounts for the range of water loss rates in the literature, and represents the energy-limited loss rate in a single variable, $\phi_{\mathrm{loss}}$; the latter represents the e-folding timescale of water loss to space, i.e., water loss is reduced by $1/e$ after $\tau_{\mathrm{loss}}$. For simplicity, we model loss of water directly to space from the surface. This approximation should be valid if the atmosphere is hot – and hence moist – in the era of high XUV. We include the hydrated layer check from @schaefer15 to ensure that the hydrated layer holds no more water than the surface itself. The model explicitly depends on mantle temperature via the mid-ocean ridge spreading rate, $S(T)$. Moreover, we stop degassing if the mantle cools below the solidus temperature, since no more melt will be present in the boundary layer. We calculate the wet solidus temperature using the parameterization of @katz03, since water in the mantle depresses the solidus of silicate minerals. Simulation Results {#sec:results} ================== We run simulations for various initial total water inventory, $W_{\mathrm{m},0} + W_{\mathrm{s},0}$, loss factor, $\phi_{\mathrm{loss}}$, and loss timescale, $\tau_{\mathrm{loss}}$. All simulations are run for 15 Gyr to allow our model to reach a steady state, if possible. Our parameter exploration is shown in Table \[tab:param\_space\]. We test four orders of magnitude for both $\phi_{\mathrm{loss}}$ and $\tau_{\mathrm{loss}}$, due to the range of loss rates in the literature, and because of the large uncertainties in observations and models of M dwarfs. We also test various initial water inventories, since planets are expected to form with different volatile inventories due to stochastic delivery and accretion [@raymond04; @raymond09]. Note that all simulations begin with an initial mantle temperature of $T_0 = 3200$ K, i.e., $T_0 = 2 T_{\mathrm{ref}}$, where $T_{\mathrm{ref}}$ is the reference temperature used in our thermal evolution calculations (detailed in Appendix \[sec:thermaleqns\]). Name Parameter Values Tested -------------------- ---------------------------------------------- -------------------------------------- Total water mass $W_{\mathrm{m},0} + W_{\mathrm{s},0}$ \[TO\] 0.1, 1, 10, 25 Mantle temperature $T_0$ \[K\] 3200 Loss factor $\phi_{\mathrm{loss}}$ \[TO/Gyr\] 0.1, 1, 10, 100 Loss timescale $\tau_{\mathrm{loss}}$ \[Gyr\] $10^{-3}$, $10^{-2}$, $10^{-1}$, $1$ : Parameter space for initial total water inventory, $W_{\mathrm{m},0} + W_{\mathrm{s},0}$, loss factor, $\phi_{\mathrm{loss}}$, and loss timescale, $\tau_{\mathrm{loss}}$. We initiate all simulations with the same mantle temperature, $T_0 = 2 T_{\mathrm{ref}}.$ Water mass is expressed in units of terrestrial oceans, where 1 TO $\approx 1.4 \times 10^{21}$ kg.[]{data-label="tab:param_space"} Each of the total water inventories from Table \[tab:param\_space\] is first run to steady-state partitioning between the mantle and surface reservoirs (Fig. \[fig:steadystate\_temps\]), without loss to space and at a constant mantle temperature of $T=3200$ K. To visualize the evolution of steady-state water partitioning as the mantle cools with time, we run simulations for various initial water inventories ($W_{\mathrm{m},0}+W_{\mathrm{s},0} = $ 0.1, 1, 5, 10, 25, 50, 100 TO) at three constant mantle temperatures, $T = $ 3200 K, 2500 K, 2000 K. As the mantle cools, the steady-state water partitioning moves towards the bottom right in Fig. \[fig:steadystate\_temps\], sequestering water in the mantle at the expense of surface water. ![\[fig:steadystate\_temps\] Steady-state water partitioning between surface, $W_{\mathrm{s}}$, and mantle, $W_{\mathrm{m}}$, reservoirs in units of terrestrial oceans (TO), for different mantle temperatures. Since the mantle cools with time, we would expect the steady state to change as well. Indeed, the steady-state curves shift towards the lower right; cooler temperatures lead to less surface water and more mantle water. For a given total water inventory and mantle temperature, there is a unique steady-state partitioning of water, and that is precisely the partitioning we use when initializing simulations in Figs. \[fig:hyb\_exploss\] & \[fig:Wloss\_tloss\_water\].](TotalWater_SteadyState_AllTemps_loglog.pdf){width="45.00000%"} @schaefer16 modelled atmosphere-interior exchange on a hot M-Earth, GJ 1132b, to determine atmospheric composition, beginning with a magma ocean and allowing solidification, and including loss to space, but the authors only simulate the first few 100 Myr. Our simulations begin after magma ocean solidification, once the steam atmosphere has mostly condensed onto the surface, and plate tectonics permit cycling. Nonetheless, our initial partitioning is qualitatively consistent with @schaefer16, with the majority of water on the surface. We are concerned with the surface water as it directly impacts habitability. We define four surface water regimes: a Dune planet, $10^{-5}$ TO $ \leq W_{\mathrm{s}} < 10^{-2}$ TO [@abe11]; an Earth-like regime, $10^{-2}$ TO $ \leq W_{\mathrm{s}} < 10$ TO; and a waterworld, where the surface is completely inundated, $W_{\mathrm{s}} \geq 10$ TO [@abbot12][^2]. We designate planets with $\lesssim 0.1$% of the surface water of a Dune planet as desiccated. This is the amount of water currently in the atmosphere of Earth ($1.29 \times 10^{16}$ kg, or ${\sim}10^{-5}$ TO; @gleick93); this amount of water is similar to Lake Superior. If precipitated onto the surface, it would produce a global ocean of depth ${\sim}2.5$ cm [@graham10]. While Dune planets, Earth-like planets, and waterworlds all have at least some liquid surface water and are thus habitable, only Earth-like planets are likely to have a silicate weathering thermostat. If surface desiccation occurs, then regassing stops. Degassing will continue if the mantle is still warm; once the degassing rate surpasses the loss rate, surface water will increase. The surface is rehydrated when it exceeds our desiccation limit of ${\sim}10^{-5}$ TO. Individual Cycling Results -------------------------- We first show the time-dependent cycling and loss results for two simulations from the parameter exploration (Fig. \[fig:hyb\_exploss\]). Both begin with $W_{\mathrm{m},0}+W_{\mathrm{s},0}$ = 1 TO of water, and loss factor $\phi_{\mathrm{loss}} = 10$ TO/Gyr. The “short loss” simulation uses a loss timescale of $\tau_{\mathrm{loss}} = 10^{-2}$ Gyr, while the “extreme loss” uses a longer loss timescale, $\tau_{\mathrm{loss}} = 10^{-1}$ Gyr. Since loss to space is occurring over a longer period for the latter, we expect a stronger reduction in surface water. Water cycling with short loss is shown in Fig. \[fig:hyb\_exploss\](a). Degassing, $w_{\uparrow}$, and regassing, $w_{\downarrow}$, are initially equal since we begin our cycling simulation from steady-state water partitioning. As a result, there is much more water on the surface, $W_{\mathrm{s}}$, than in the mantle, $W_{\mathrm{m}}$ (top panel). The loss rate to space, $w_{\mathrm{loss}}$, is initially higher than the cycling rates. Loss to space initially dominates, reducing the surface water inventory, which reduces the amount available to sequester back into the mantle at late times. The cycling rates surpass the loss rate at $t \approx \tau_{\mathrm{loss}} = 10^{-2}$ Gyr, and as loss slows, the water partitioning seeks a new steady state for the current total water inventory and cooler mantle temperature, $T$. {width="45.00000%"} {width="45.00000%"} The results for the extreme water loss simulation are shown in Fig. \[fig:hyb\_exploss\](b). The cycling begins similarly to the previous simulation, but the surface is rapidly desiccated (grey region). Degassing still provides water to the surface, where some is lost to space and a small amount is regassed, gradually reducing the mantle water inventory, while the surface water complement approaches zero. Eventually, degassing from the mantle surpasses loss to space, and the surface recovers enough water to once again become a Dune planet (brown region). Even with continued cycling, there is not enough total water remaining on the planet to recover Earth-like surface conditions. Nonetheless, Fig. \[fig:hyb\_exploss\](b) demonstrates that water sequestered in the mantle can rehydrate the surface once loss to space diminishes. Parameter Exploration --------------------- We now perform an exploration of the parameter space in Table \[tab:param\_space\]. We focus on the final surface water inventories to determine what surface conditions to expect after 15 Gyr of water cycling and loss to space. Our results for the 64 simulations are illustrated in Fig. \[fig:Wloss\_tloss\_water\]. We choose initial water inventories up to $W_{\mathrm{m},0}+W_{\mathrm{s},0}=25$ TO, far below the high-pressure ice limit of $W_{\mathrm{s,max}} = 100$ TO [@nakayama19], to allow plate tectonic-driven cycling to continue uninhibited. As shown in Fig. \[fig:Wloss\_tloss\_water\], planets can evolve between surface water regimes. Certain rates of water loss to space cause waterworlds (blue) to lose sufficient water to expose continents and become Earth-like (similar to the “waterworld self-arrest” of @abbot12), or Earth-like planets (green) to become dry Dune planets (brown) with little surface water, or even develop a completely desiccated, uninhabitable surface. Water loss is limited by the amount of water on the surface (top left panel, Fig. \[fig:Wloss\_tloss\_water\]). A loss rate of 10 TO/Gyr and a loss timescale of 0.1 Gyr might in principle desiccate a planet with a 0.1 TO inventory, but since the loss predominantly happens early in the evolution, it only removes the surface water present at that time. {width="90.00000%"} Fig. \[fig:Wloss\_tloss\_water\] only shows the initial and final surface water contents, but we check for mid-simulation desiccation, as shown in Fig. \[fig:hyb\_exploss\](b). Ten of the 64 simulated planets recover from desiccation into either the Dune or Earth-like regime. This further supports our mechanism of sequestering water in the mantle and degassing it once atmospheric loss has decreased appreciably to restore habitable surface conditions. Discussion & Conclusions {#sec:discussion} ======================== Our simulations show that sequestering water in the mantle and subsequent degassing enhances the likelihood of habitable M-Earths in the face of atmospheric loss, provided they have an Earth-like deep-water cycle. Model Timescales ---------------- There are three relevant timescales in our model that permit a more thorough interpretation of our results (Figs. \[fig:hyb\_exploss\] &  \[fig:Wloss\_tloss\_water\]). These timescales are the time to reach steady state, $\tau_{\mathrm{ss}}$, the loss timescale, $\tau_{\mathrm{loss}}$, and the mantle overturn timescale, $\tau_{\mathrm{overturn}}$. The surface water content and mantle temperature in our model change with time, not only due to loss but also water degassed and regassed from and to the mantle, respectively. The planet will therefore be approaching a changing steady state with time (Fig. \[fig:steadystate\_temps\]). This steady state will not be reached until loss diminishes and mantle cooling slows at late times, allowing degassing and regassing rates to equilibrate. Indeed, we only see steady state achieved late in our simulations, and only as long as the mantle remains warm. If the loss timescale is much longer than the mantle overturn timescale, $\tau_{\mathrm{loss}} \gg \tau_{\mathrm{overturn}}$, the water lost to space will be roughly $\phi_{\mathrm{loss}} \tau_{\mathrm{loss}}$, provided there is sufficient total water on the planet. This explains the different results for the same $\phi_{\mathrm{loss}} \tau_{\mathrm{loss}}$ in Fig. \[fig:Wloss\_tloss\_water\]. Since the time to reach steady-state is closely related to the mantle overturn timescale, $\tau_{\mathrm{loss}} \gg \tau_{\mathrm{overturn}}$ also means that the planet is always at or near a steady state (equal degassing and regassing), but that steady state is a moving target due to atmospheric loss. If $\tau_{\mathrm{loss}} \ll \tau_{\mathrm{overturn}}$, however, the total water lost is now limited by the initial surface water on the planet, $W_{\mathrm{s},0}$. Since most loss happens early on and the loss rate diminishes with time, the surface can eventually be rehydrated, provided there is sufficient water sequestered within the mantle. This explains the similar results seen in each panel of Fig. \[fig:Wloss\_tloss\_water\], on the left and bottom-left. The greater a planet’s initial water inventory, the farther towards the upper-right corner this plateau extends. In summary, the total amount of water lost, $W_{\mathrm{lost}}$, is, $$W_{\mathrm{lost}} = \begin{cases} \min[\phi_{\mathrm{loss}} \tau_{\mathrm{loss}}, ~W_{\mathrm{s},0}+W_{\mathrm{m},0}] & \tau_{\mathrm{loss}} \gg \tau_{\mathrm{overturn}} \\ \min[\phi_{\mathrm{loss}} \tau_{\mathrm{loss}}, ~W_{\mathrm{s},0}] & \tau_{\mathrm{loss}} \ll \tau_{\mathrm{overturn}}. \end{cases}$$ Thermal Evolution & Tectonic Mode --------------------------------- The model can approach a steady state once loss has diminished significantly, as long as the mantle remains above the solidus temperature. Once the mantle cools below the solidus temperature, degassing stops due to the absence of melt below mid-ocean ridges. This leads to regassing-dominated evolution, eventually trapping all water in the mantle [@schaefer15]. It has been postulated, however, that when the mantle cools below the solidus temperature or becomes desiccated, convection may stop, along with plate tectonics, transitioning to a “stagnant lid” regime [@noack14; @lenardic18]. As noted by @schaefer15, transitioning to a stagnant lid would stop both degassing and regassing, preserving the water inventories in surface and mantle reservoirs at that time. A stagnant lid would greatly affect our cycling parameterizations, but volatiles can still be cycled in a stagnant-lid regime, albeit at a much slower rate [@honing19]. We leave this complication for future work; however, since our current simulations sometimes regas all water into the mantle, presumably accounting for a stagnant lid would merely result in more surface water at late times. Observational Prospects ----------------------- Observationally characterizing M-Earth atmospheres in the near future is viable [@cowan15; @shields16; @gillon20], but direct detection of surface water on an exoplanet is probably still many years away [@cowan09; @robinson10; @lustig18]. To zeroth-order, our conclusions support continued observations of M dwarf systems in the search for habitability. Our results will be useful in interpreting observations, allowing inference of the cycling & loss history of M-Earths based on, e.g., the presence of H~2~O in transit spectra. Connecting surface water to climate, atmospheric structure, and transit spectroscopy will be the subject of a future study. Increasing the fidelity of our M-Earth water cycling & loss model will narrow the gap between predictions and observations. The key variables in our model include the initial water inventory, the initial water partitioning, the mantle overturn timescale, the loss rate and the loss timescale. The initial water inventory may be difficult to determine due to the stochastic nature of volatile delivery during planet formation [@raymond04; @raymond09]; however, studies of volatiles in protoplanetary disks (e.g., using ALMA; @harsono20; @loomis20) and studies of polluted white dwarfs (e.g., @farihi16; @veras17a [@veras17b]; @doyle19) may provide constraints. The geophysical processes that determine both the planetary water partitioning and mantle overturn timescale in our M-Earth model are based on present-day Earth. Determining the tectonic mode of an observed exoplanet will be difficult in the near future, but in principle, may be possible with LUVOIR (e.g., @cowan09); nonetheless, modelling can allow exploration of the potential geophysics on distant planets. Many uncertainties in our model arise due to our treatment of stellar evolution, but we may be able to constrain the loss rate (i.e., the loss factor and timescale) through a combination of M dwarf observations and modelling to better represent the governing loss processes on an M-Earth. Acknowledgements {#acknowledgements .unnumbered} ================ We thank the anonymous referee for a beneficial referee report that strengthened this manuscript. We thank Tad Komacek for providing his hybrid model and for valuable correspondence while re-coding the model. We also acknowledge thesis committee members Yajing Liu, Vincent van Hinsberg, Don Baker, Galen Halverson, and Timothy Merlis, as well as insightful conversations with Christie Rowe and Mark Jellinek. We thank Dylan Keating and Lisa Dang for comments on a draft of this manuscript. K.M. acknowledges support from a McGill University Dr. Richard H. Tomlinson Doctoral Fellowship, and from the Natural Sciences and Engineering Research Council of Canada (NSERC) Postgraduate Scholarships-Doctoral Fellowship. Data Availability {#data-availability .unnumbered} ================= The model and data presented within this manuscript are available from the corresponding author at reasonable request. [99]{} Abbot, D. S., Cowan, N. B., & Ciesla, F. J. 2012, , 756, 178 Abe, Y., Abe-Ouchi, A., Sleep, N. H., & Zahnle, K. J. 2011, Astrobiology, 11, 443 Airapetian, V. S., Glocer, A., Khazanov, G. V., Loyd, R. O. P., France, K., Sojka, J., Danchi, W. C., & Liemohn, M. W. 2017, , 836, L3 Alibert, Y. 2014, , 561, A41 Bolmont, E., Selsis, F., Owen, J. E., Ribas, I., Raymond, S. N., Leconte, J., & Gillon, M. 2016, , 464, 3728 Cowan, N. B., Agol, E., Meadows, V. S., Robinson, T., et al. 2009, , 700, 915 Cowan, N. B., & Abbot, D. S. 2014, , 781, 27 Cowan, N. B., Greene, T., Angerhausen, T., Batalha, N. E., Clampin, M., Colón, Crossfield, I. J. M., et al. 2015, Publications of the Astronomical Society of the Pacific, 127, 311 Doyle, A. E., Young, E. D., Klein, B., Zuckerman, B., Schlichting, H. E. 2019, Science, 366, 356 Dressing, C. D., & Charbonneau, D. 2015, , 807, 45 Farihi, J. 2016, New Astronomy Reviews, 71, 9 Fleming, D. P., Barnes, R., Luger, R., VanderPlas, J. T. 2020, , 891, 155 France, K., Arulanantham, N., Fossati, L., Lanza, A. F., Loyd, R. O. P., Redfield, S., & Schneider, P. C. 2018, , 239, 16 Garcia-Sage, K., Glocer, A., Drake, J. J., Gronoff, G., & Cohen, O. 2017, , 844, L13 Gillon, M., Meadows, V., Agol, E., et al. 2020, arXiv:2002.04798 Gleick, P. H. 1993, Water in Crisis: A Guide to the World’s Fresh Water Resources (New York: Oxford University Press) Graham, S., Parkinson, C., & Chahine, M. 2010, The Water Cycle, NASA Earth Observatory, Greenbelt, MD Hamano, K., Abe, Y., & Genda, H. 2013, Nature, 497, 607 Harsono, D., Persson, M. V., Ramos, A., Murillo, N. M., Maud, L. T., et al. 2020, , 636, A26 Hauri, E. H., Gaetani, G. A., & Green, T. H. 2006, E&PSL, 248, 715 Henry, T. J. 2004, ASP Conference Series 318: Spectroscopically and Spatially Resolving the Components of the Close Binary Stars, 159 Hirschmann, M. M. 2006, Annu. Rev. Earth Planet. Sci., 34:629–53 Höning, D., Tosi, N., & Spohn, T. 2019, , 627, A48 Kasting, J. F., Whitmire, D. P., & Reynolds, R. T. 1993, Icarus, 101, 108 Katz, R. F., Spiegelman, M., & Langmuir, C. H. 2003, GGG, 4, 1073 Kite, E. S., Manga, M., & Gaidos, E. 2009, , 700, 1732 Komacek, T. D., & Abbot, D. S. 2016, , 832, 54 Langmuir, C. H., & Broecker, W. 2012, How to Build a Habitable Planet (Princeton: Princeton Univ. Press) Lenardic, A. 2018, Phil. Trans. R. Soc. A, 376: 20170416. Li, Z. X. A., Lee, C.-T. A., Peslier, A. H., Lenardic, A., & Mackwell, S. J. 2008, Journal of Geophysical Research, 113, B09210 Loomis, R. A., [Ö]{}berg, K. I., Andrews, S. M., Bergin, E., Bergner, J., Blake, G. A., et al. 2020, , 893, 101 Luger, R., & Barnes, R. 2015, Astrobiology, 15, 119 Lustig-Yaeger, J., Meadows, V. S., Tovar Mendoza, G., et al. 2018, , 156, 301 McGovern, P. J., & Schubert, G. 1989, Earth and Planetary Science Letters, 96, 27 Nakayama, A., Kodama, T., Ikoma, M., & Abe, Y. 2019, , 488, 1580 Noack, L., & Breuer, D. 2014, , 98, 41 Ranjan, S., Wordsworth, R., & Sasselov, D. D. 2017, , 843, 110 Raymond, S. N., Quinn, T., & Lunine, J. I. 2004, , 168, 1 Raymond, S. N., O’Brien, D. P., Morbidelli, A., Kaib, N. A. 2009, , 203, 644 Ribas, I., Guinan, E. F., G[ü]{}del, M., & Audard, M. 2005, ApJ, 622, 680 Ribas, I., Bolmont, E., Selsis, F., Reiners, A., Leconte, J., Raymond, S. N., et al. 2016, , 596, A111 Ribas, I., Gregg, M. D., Boyajian, T. S., & Bolmont, E. 2017, , 603, A58 Rimmer, P. B., Xu, J., Thompson, S. J., Gillen, E., Sutherland, J. D., Queloz, D. 2018, Science Advances, 4, eaar3302 Robinson, T. D., Meadows, V. S., & Crisp, D. 2010, , 721, L67 Sandu, C., Lenardic, A., & McGovern, P. 2011, Journal of Geophysical Research, 116, B12404 Scalo, J., Kaltenegger, L., Segura, A. G., Fridlund, M. Ribas, I., Kulikov, Y. N., Grenfell, J. L., Rauer, H., Odert, P., Leitzinger, M., Selsis, F., Khodachenko, M. L., Eiroa, C., Kasting, J., & Lammer, H. 2007, Astrobiology, 7, 85 Schaefer, L., & Sasselov, D. 2015, , 801, 40 Schaefer, L., Wordsworth, R. D., Berta-Thompson, Z., & Sasselov, D. 2016, , 829, 63 Shields, A. L., Ballard, S., & Johnson, J. A. 2016, Physics Reports, 663, 1 Sleep, N. H., & Zahnle, K. 2001, Journal of Geophysical Research, 106, 1373 Tian, F., Kasting, J. F., Liu, H.-L., & Roble, R. G. 2008, Journal of Geophysical Research, 113, E05008 Turcotte, D. L., & Schubert, G. 2002, Geodynamics (New York, NY: Cambridge Univ. Press) Valencia, D., O’Connell, R. J., & Sasselov, D. 2006, , 181, 545 Veras, D., Carter, P. J., Leinhardt, Z. M., & G[ä]{}nsicke, B. T. 2017, , 465, 1008 Veras, D., Georgakarakos, N., Dobbs-Dixon, I., & G[ä]{}nsicke, B. T., 2017, , 465, 2053 Walker, J. C. G., Hays, P. B., & Kasting, J. F. 1981, Journal of Geophysical Research, 86, 9776 Wordsworth, R. D., & Pierrehumbert, R. T. 2013, , 778, 154 Wordsworth, R., & Pierrehumbert, R. 2014, , 785, L20 Wordsworth, R. D., Schaefer, L. K., & Fischer, R. A. 2018, , 155, 195 Youngblood, A., France, K., Parke Loyd, R. O., Linsky, J. L., Redfield, S., Christian Schneider, P., et al. 2016, , 824, 101 Thermal Evolution Equations {#sec:thermaleqns} =========================== The thermal evolution of the mantle in our model, which incorporates parameterized convection, is a simplified version of the thermal evolution presented within @schaefer15, itself based on the model of @sandu11. For simplicity, and to reproduce the low-viscosity model which reaches an analytic steady state (the goal of @komacek16), the convection is constrained to a boundary layer in the upper mantle. The evolution of the mantle temperature, $T$, with time, $t$, in our model is dependent on mantle water mass fraction $x = W_{\mathrm{m}}/ M_{\mathrm{m}}$, where $M_{\mathrm{m}}$ is the mass of the mantle. The thermal evolution equation is: $$\label{eqn:dT/dt} \begin{split} \rho_{\mathrm{m}} c_{\mathrm{p}} \frac{dT}{dt} & = Q(t) - \frac{A}{V} \frac{k \Delta T}{h} \left(\frac{\mathrm{Ra}}{\mathrm{Ra}_{\mathrm{crit}}} \right)^\beta \\ & = Q(t) - \frac{k A \Delta T}{h V} \left(\frac{\alpha \rho_{\mathrm{m}} g h^3 \Delta T}{\mathrm{Ra}_{\mathrm{crit}} \kappa \eta(T,x)} \right)^\beta. \end{split}$$ The value of $\beta = 0.3$ in this equation was determined empirically [@mcgovern89]. The density of the upper mantle is $\rho_{\mathrm{m}}$, and $c_{\mathrm{p}}$ is the mantle’s specific heat capacity. The heating rate from radionuclides is $Q(t) = Q_0 \exp^{-t/\tau_{\mathrm{decay}}}$. The decay timescale $\tau_{\mathrm{decay}} = 2$ Gyr was nominally chosen by @komacek16, based on the abundance and half-lives of radiogenic elements in the Earth’s mantle from @turcotte02. For reference, this value falls between the half-lives of ^40^K (1.3 Gyr) and ^238^U (4.5 Gyr). The thermal conductivity of the upper mantle is $k$, and $\Delta T = T - T_{\mathrm{s}}$ is the temperature contrast across the boundary layer, with $T_{\mathrm{s}}$ the surface temperature. The scaling laws of terrestrial planets from @valencia06 allow us to calculate the planet’s mantle thickness, $h$, the planet’s surface area, $A$, the mantle volume, $V$, and surface gravity, $g$. Planetary radius, $R$, and core radius, $R_{\mathrm{c}}$, are related to planetary mass, $M$, by $$\label{eqn:R} R = R_{\oplus} \left(\frac{M}{M_{\oplus}} \right)^p,$$ $$\label{eqn:R_c} R_{\mathrm{c}} = c R_{\oplus} \left(\frac{M}{M_{\oplus}} \right)^{p_{\mathrm{c}}},$$ where $p=0.27$, $c=0.547$, and $p_{\mathrm{c}}=0.25$. The remaining planet and mantle properties are: $$\label{eqn:h} h = R - R_{\mathrm{c}},$$ $$\label{eqn:A} A = 4 \pi R^2,$$ $$\label{eqn:V} V = \frac{4 \pi}{3} (R^3 - R_{\mathrm{c}}^3),$$ $$\label{eqn:g} g = \frac{G M}{R^2}.$$ The Rayleigh number of the mantle, $\mathrm{Ra} = \alpha \rho_{\mathrm{m}} g h^3 \Delta T/ \kappa \eta(T,x)$, can be calculated using the upper mantle density, $\rho_{\mathrm{m}}$, mantle thickness, $h$, and temperature contrast, $\Delta T$, along with the characteristic thermal expansivity, $\alpha$, the planet’s gravity, $g$, and the thermal diffusivity of the boundary layer, $\kappa$. The critical Rayleigh number for convection to occur in the upper mantle is $\mathrm{Ra}_{\mathrm{crit}} = 1100$ [@mcgovern89]. Due to the dependence of temperature $T$ on mantle water mass fraction $x$, the thermal evolution and cycling equations are integrated simultaneously. For the mantle viscosity, $\eta(T,x)$, we use the same parameterization as @komacek16, which in turn is a simplified version of that from the models of @sandu11 and @schaefer15 (i.e., without the pressure-dependence, since we are restricted to the upper mantle): $$\label{eqn:eta} \eta \approx \eta_0 f_{\mathrm{w}}^{-r} \exp \left[\frac{E_\mathrm{a}}{R_{\mathrm{gas}}} \left( \frac{1}{T} - \frac{1}{T_{\mathrm{ref}}} \right) \right].$$ Here, $\eta_0$ is the viscosity scale (chosen so that $\eta(x=x_{\oplus}, T=T_{\mathrm{ref}}) = 10^{21}$ Pa$\cdot$s to reproduce the viscosities of Earth’s mantle), $E_{\mathrm{a}}$ is the activation energy, $R_{\mathrm{gas}}$ is the universal gas constant, $T_{\mathrm{ref}}=1600$ K is the reference mantle temperature, $f_{\mathrm{w}}$ is water fugacity (see Eqn. \[eqn:f\_w\] below), and $r=1$ is the nominal value chosen by @schaefer15, based on measurements of wet olivine diffusion. The water fugacity, $f_{\mathrm{w}}$, can be calculated using experimental data from @li08, $$\label{eqn:f_w} \begin{split} \ln f_{\mathrm{w}} & = c_0 + c_1 \ln \left( \frac{B x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}}{1 - x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}} \right) + c_2 \ln^2 \left( \frac{B x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}}{1 - x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}} \right) \\ & + c_3 \ln^3 \left( \frac{B x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}}{1 - x \mu_{\mathrm{oliv}}/\mu_{\mathrm{w}}} \right), \end{split}$$ where $c_0 = -7.9859$, $c_1 = 4.3559$, $c_2 = -0.5742$, $c_3 = 0.0337$, $B=2 \times 10^6$ (which converts to number concentration of H atoms per $10^6$ Si atoms), $\mu_{\mathrm{oliv}}$ is the molecular weight of olivine, and $\mu_{\mathrm{w}}$ is the molecular weight of water. Model Improvements & Constraints {#sec:improvements} ================================ Our improvements to the hybrid model of @komacek16 include restoring the degassing limit from @cowan14, and capacity limits for mantle water (to account for a saturated mantle, $W_{\mathrm{m}} \leq 12$ TO; @hauri06 [@cowan14]) and surface water ($W_{\mathrm{s}} \leq 100$ TO, above which high-pressure ices will form at the ocean floor and significantly alter or hinder the degassing/regassing rates; @nakayama19). The addition of a simultaneous loss term in the cycling equations brings the model closer to predictions for XUV-driven water loss to space on M-Earths (e.g. @luger15). Our current loss factors, $\phi_{\mathrm{loss}}$, and timescales, $\tau_{\mathrm{loss}}$, allow for a phenomenological exploration of parameter space to show the effect of water loss rather than being directly calculated from, for example, stellar evolution models of XUV flux from M dwarfs (e.g., @ribas05). While coupled integrations of the thermal evolution in Eqn. \[eqn:dT/dt\] and the cycling of Eqns. \[eqn:dx/dt\_hyb\] and \[eqn:ds/dt\_hyb\] can be performed using the `scipy.integrate` package in Python, restrictions must be placed to ensure we do not obtain meaningless, unphysical results. Our hybrid model is robust to either or both reservoirs going to zero, an improvement over @komacek16. At each timestep, surface and mantle water inventories are checked, and the cycling equations are adjusted accordingly: 1. If mantle water mass $W_{\mathrm{m}}$ and surface water mass $W_{\mathrm{s}}$ are both greater than zero, then normal cycling and loss occurs, and the cycling equations are integrated as they appear in §\[sec:equations\]. 2. If $W_{\mathrm{m}} = 0$ but $W_{\mathrm{s}} > 0$, the degassing rate, $w_{\uparrow}$, is set to zero (i.e., it shuts off since there is no water in the mantle) in Eqns. \[eqn:dx/dt\_hyb\] and \[eqn:ds/dt\_hyb\], and the integration is performed. 3. If $W_{\mathrm{m}} > 0$ but $W_{\mathrm{s}} = 0$, the regassing rate, $w_{\downarrow}$, and loss rate, $w_{\mathrm{loss}}$, are set to zero (since there is no water on the surface) in Eqns. \[eqn:dx/dt\_hyb\] and \[eqn:ds/dt\_hyb\], and the integration is performed. We also set the fraction of water in melt that is degassed $f_{\mathrm{degas}}(P)=1$, due to its piecewise definition. We do this because the $P$-dependent degassing rate depends on the overlying surface water, $W_{\mathrm{s}}$; if there is no water on the surface, degassing should neither go to zero and shut off (or else water would stay in the mantle indefinitely), nor go to infinity (or all water would be instantaneously degassed from the mantle). 4. If both $W_{\mathrm{m}} = 0$ and $W_{\mathrm{s}} = 0$, the degassing, regassing, and loss rates are all set to zero, since the planet is completely desiccated. The integration continues so we can observe the thermal evolution of the mantle (i.e., its cooling with time), but there is no cycling or loss since there is no more water present on the surface or within the mantle of the planet. Our piecewise definition of loss in Eqn. \[eqn:ds/dt\_hyb\] ensures that the amount of surface water that is regassed and lost at a given timestep does not exceed the amount present on the surface, and the regassing/loss rates are adjusted accordingly based on the surface water mass, $W_{\mathrm{s}}$. Although it is a result we have yet to encounter, to account for complete mantle desiccation (a scenario proposed in the literature; see, e.g., @hamano13), we choose a minimum water fugacity to avoid $f_{\mathrm{w}} \rightarrow 0$ and mantle viscosity $\eta \rightarrow \infty$. We can then define a piecewise mantle water fugacity, $f_{\mathrm{w,eff}}$, represented by the equation, $$\label{f_w_limit} f_{\mathrm{w,eff}} = \max [10^{-5} f_{\mathrm{w}}(\tilde{x}=1), ~f_{\mathrm{w}}]$$ where $f_{\mathrm{w}}(\tilde{x}=1)$ is used to define the non-dimensional fugacity within our model code, $\tilde{f}_\mathrm{w} = f_\mathrm{w} / f_{\mathrm{w}}(\tilde{x}=1)$. This definition requires the non-dimensional water mass fraction $\tilde{x} = x f_\mathrm{m} / (\omega_0 \tilde{f}_\mathrm{b})$, where $f_{\mathrm{m}}$ is the mantle fraction, $\omega_0$ is the surface water mass fraction of Earth, and $\tilde{f}_\mathrm{b} = f_\mathrm{b} / f_{\mathrm{b},\oplus} = 1.3$ is the non-dimensional ocean basin covering fraction, with $f_{\mathrm{b},\oplus}=0.7$ and $f_\mathrm{b} = 0.9$ (i.e., 90% of planet covered in water). The value for $f_{\mathrm{b}}$ was chosen by @cowan14, which we also optimistically adopt for an Earth-like planet. The minimum value, $f_{\mathrm{w,eff}} = 10^{-5} f_\mathrm{w}(\tilde{x}=1)$, assumes that even in the case of a completely desiccated mantle, there will be a small amount of water trapped in the minerals (e.g., within the transition zone; @hirschmann06). This allows the mantle to continue convecting and our plate-tectonics-dependent cycling to proceed. Note that, throughout the thermal evolution equations presented above, $f_{\mathrm{w}}$ is used in place of $f_{\mathrm{w,eff}}$ for consistency with the literature. Finally, we note that the water fugacity was calculated incorrectly in many places in the original hybrid cycling model due to a misplaced bracket (Komacek 2019, priv. comm.). While this error does not significantly impact the final results, it does slightly change the time-dependent cycling results, and has been fixed in the model presented here. Model Parameters {#sec:params} ================ There are many variables throughout this paper. As such, we detail them all in Table \[tab:modelparams\], including their nominal values. [c|c|c]{} Name & Parameter & Value\ Mantle water mass$^{ab}$ & $W_{\mathrm{m}}$ & $W_{\mathrm{m},\oplus} = 2.36 \times 10^{21}$ kg $\approx 1.7$ TO\ Surface water mass$^{ab}$ & $W_{\mathrm{s}}$ & $W_{\mathrm{s},\oplus} \approx 1.4 \times 10^{21}$ kg $ = 1$ TO\ Mantle temperature$^a$ & $T$ & $T_{\mathrm{ref}} = 1600$ K\ Loss factor$^b$ & $\phi_{\mathrm{loss}}$ & $10$ TO Gyr$^{-1}$\ Loss timescale$^b$ & $\tau_{\mathrm{loss}}$ & $10^8$ yr\ Mid-ocean ridge length & $L_{\mathrm{MOR}}$ & $L_{\mathrm{MOR},\oplus} = 60 \times 10^6$ m\ Spreading rate$^a$ & $S$ & $S_{\oplus,\mathrm{avg}} \approx 0.1$ m yr$^{-1}$\ Water mass fraction in hydrated crust & $x_{\mathrm{h}}$ & 0.05\ Crust density & $\rho_{\mathrm{c}}$ & $3.0 \times 10^3$ kg m$^{-3}$\ Regassing efficiency & $\chi_{\mathrm{r}}$ & 0.03\ Hydrated layer depth$^a$ & $d_{\mathrm{h}}$ & $d_{\mathrm{h},\oplus} = 3.0 \times 10^3$ m\ Mantle water mass fraction & $x$ & $x_{\oplus} = 5.8 \times 10^{-4}$\ Mantle density & $\rho_{\mathrm{m}}$ & $3.3 \times 10^3$ kg m$^{-3}$\ Mid-ocean ridge melting depth & $d_{\mathrm{melt}}$ & $60 \times 10^3$ m\ Melt degassing efficiency of Earth & $f_{\mathrm{degas},\oplus}$ & 0.9\ Seafloor pressure$^a$ & $P$ & $P_{\oplus} = 4 \times 10^7$ Pa\ Timestep & $\tau_{\mathrm{step}}$ & ${\sim}28700$ yr\ Mantle overturn timescale & $\tau_{\mathrm{overturn}}$ & $\tau_{\mathrm{overturn},\oplus} \approx 6 \times 10^6$ yr\ Mantle specific heat capacity & $c_{\mathrm{p}}$ & 1200 J kg$^{-1}$ K$^{-1}$\ Radionuclide decay factor & $Q_0$ & $5 \times 10^{-8}$ J m$^{-3}$ s$^{-1}$\ Decay timescale & $\tau_{\mathrm{decay}}$ & 2 Gyr\ Mantle thermal conductivity & $k$ & 4.2 W m$^{-1}$ K$^{-1}$\ Surface temperature & $T_{\mathrm{s}}$ & 280 K\ Mantle critical Rayleigh number & $\mathrm{Ra}_{\mathrm{crit}}$ & 1100\ Heat flux exponent & $\beta$ & 0.3\ Mantle characteristic thermal expansivity & $\alpha$ & $2 \times 10^{-5}$ K$^{-1}$\ Mantle thermal diffusivity & $\kappa$ & $10^{-6}$ m$^2$ s$^{-1}$\ Planet radius & $R$ & $R_{\oplus} = 6.371 \times 10^6$ m\ Planet mass & $M$ & $M_{\oplus} = 5.972 \times 10^{24}$ kg\ Gravitational constant & $G$ & $6.67 \times 10^{-11}$ m$^3$ kg$^{-1}$ s$^{-2}$\ Mantle viscosity$^a$ & $\eta$ & $\eta(x=x_{\oplus}, T=T_{\mathrm{ref}}) = 10^{21}$ Pa s\ Water fugacity$^a$ & $f_{\mathrm{w}}$ & $f_{\mathrm{w}}(x_{\oplus}) \approx 17 \times 10^3$ Pa\ Fugacity exponent & $r$ & 1\ Activation energy & $E_{\mathrm{a}}$ & $335 \times 10^3$ J mol$^{-1}$\ Universal gas constant & $R_{\mathrm{gas}}$ & 8.314 J mol$^{-1}$ K$^{-1}$\ Molecular weight of olivine & $\mu_{\mathrm{oliv}}$ & 153.31 g mol$^{-1}$\ Molecular weight of water & $\mu_{\mathrm{w}}$ & 18.02 g mol$^{-1}$\ Planetary mantle fraction & $f_{\mathrm{m}}$ & 0.68\ Ocean basin covering fraction & $f_{\mathrm{b}}$ & 0.9\ Earth ocean basin covering fraction & $f_{\mathrm{b},\oplus}$ & 0.7\ \ \ \ \[lastpage\] [^1]: E-mail: keavin.moore@mail.mcgill.ca [^2]: The waterworld definition of @wordsworth13 is similar — all land is covered by water, but this does not completely inhibit degassing from the interior. We assume this as well for M-Earths in the waterworld regime.

ArXiv

--- abstract: | We compute Haar ensemble averages of ratios of random characteristic polynomials for the classical Lie groups $K = \mathrm{O}_N\,$, $\mathrm{SO}_N\,$, and $\mathrm{USp}_N\,$. To that end, we start from the Clifford-Weyl algebra in its canonical realization on the complex $\mathcal{A}_V$ of holomorphic differential forms for a $\mathbb{C}$-vector space $V_0\,$. From it we construct the Fock representation of an orthosymplectic Lie superalgebra $\mathfrak{osp}$ associated to $V_0\,$. Particular attention is paid to defining Howe’s oscillator semigroup and the representation that partially exponentiates the Lie algebra representation of $\mathfrak{sp} \subset \mathfrak{osp}$. In the process, by pushing the semigroup representation to its boundary and arguing by continuity, we provide a construction of the Shale-Weil-Segal representation of the metaplectic group. To deal with a product of $n$ ratios of characteristic polynomials, we let $V_0 = \mathbb{C}^n \otimes \mathbb{C}^N$ where $\mathbb {C}^N$ is equipped with the standard $K$-representation, and focus on the subspace $\mathcal{A}_V^K$ of $K$-equivariant forms. By Howe duality, this is a highest-weight irreducible representation of the centralizer $\mathfrak{g}$ of $\mathrm{Lie}(K)$ in $\mathfrak{osp}$. We identify the $K$-Haar expectation of $n$ ratios with the character of this $\mathfrak{g} $-representation, which we show to be uniquely determined by analyticity, Weyl group invariance, certain weight constraints and a system of differential equations coming from the Laplace-Casimir invariants of $\mathfrak{g}$. We find an explicit solution to the problem posed by all these conditions. In this way, we prove that the said Haar expectations are expressed by a Weyl-type character formula for *all* integers $N \ge 1$. This completes earlier work of Conrey, Farmer, and Zirnbauer for the case of $\mathrm{U}_N\,$. author: - 'A. Huckleberry, A. Püttmann and M.R. Zirnbauer' date: 'September 8, 2007' title: | Haar expectations of ratios of\ random characteristic polynomials --- Introduction ============ \[varying\] In this article we derive an explicit formula for the average $$I(t) := \int_K Z(t,k)\, dk$$ where $K$ is one of the classical compact Lie groups $\mathrm{O}_N\,$, $\mathrm{SO}_N\,$, or $\mathrm{USp}_N$ equipped with Haar measure $dk$ of unit mass $\int_K dk = 1$ and $$Z(t,k) := \prod_{j=1}^n \frac{\mathrm{Det}( \mathrm{e}^{\frac{\mathrm{i}}{2}\psi_j} \mathrm{Id}_N - \mathrm{e}^{-\frac{\mathrm{i}}{2}\psi_j} \,k)} {\mathrm{Det}(\mathrm{e}^{\frac{1}{2}\phi_j}\mathrm{Id}_N - \mathrm{e}^{- \frac{1}{2}\phi_j} \, k)}$$ depends on a set of complex parameters $t := (\mathrm{e}^{ \mathrm{i}\psi_1} , \ldots , \mathrm{e}^{\mathrm{i}\psi_n} , \mathrm{e}^{\phi_1} , \ldots, \mathrm{e}^{\phi_n})$, which satisfy $\mathfrak{Re}\, \phi_j > 0$ for all $j = 1, \ldots, n\,$. The case of $K = \mathrm{U}_N$ is handled in [@CFZ]. Note that $$Z(t,k) = \mathrm{e}^{\lambda_N} \prod_{j=1}^n \frac{\mathrm{Det}(\mathrm{Id}_N - \mathrm{e}^{-\mathrm{i} \psi_j}\, k)} {\mathrm{Det}(\mathrm{Id}_N - \mathrm{e}^{- \phi_j}\, k)}$$ with $\lambda_N = \frac{N}{2} \sum_{j=1}^n (\mathrm{i} \psi_j - \phi_j)$. This means that $Z(t,k)$ is a product of ratios of characteristic polynomials, which explains the title of the article. The Haar average $I(t)$ can be regarded as the (numerical part of the) character of an irreducible representation of a Lie supergroup $(\mathfrak{g}\, ,G)$ restricted to a suitable subset of a maximal torus of $G$. The Lie superalgebra $\mathfrak{g}$ is the Howe dual partner of the compact group $K$ in an orthosymplectic Lie superalgebra $\mathfrak{osp} \,$. It is naturally represented on a certain infinite-dimensional spinor-oscillator module $\mathfrak{a} (V)$ – more concretely, the complex of holomorphic differential forms on the vector space $\mathbb{C}^n \otimes \mathbb{C}^N$ – and the irreducible representation is that on the subspace $\mathfrak{a}(V)^K$ of $K$-equivariant forms. To even define the character, we must exponentiate the representation of the Lie algebra part of $\mathfrak{osp}$ on $\mathfrak{a}(V)$. This requires going to a completion $\mathcal{A}_V$ of $\mathfrak {a}(V)$, and can only be done partially. Nevertheless, the represented semigroup contains enough structure to derive Laplace-Casimir differential equations for its character. Our explicit formula for $I(t)$ looks exactly like a classical Weyl formula and is derived in terms of the roots of the Lie superalgebra $\mathfrak{g}$ and the Weyl group $\mathrm{W}$. Let us state this formula for $K = \mathrm{O}_N\,$, $\mathrm{USp}_N$ without going into the details of the $\lambda$-positive even and odd roots $\Delta ^+_{\lambda,0}$ and $\Delta ^+_{\lambda ,1}$ and the Weyl group $\mathrm{W}$ (see $\S$\[sect:WeylGroup\] for precise formulas). If $\mathrm{W}_\lambda$ is the isotropy subgroup of $\mathrm{W}$ fixing the highest weight $\lambda = \lambda_N\,$, then $$\label{eq:WeylCharacter} I(t)=\sum_{[w] \in \mathrm{W}/\mathrm{W}_\lambda} \mathrm{e}^{ w(\lambda_N)} \frac{\prod_{\beta \in \Delta_{\lambda,1}^+}(1 - \mathrm{e}^{-w(\beta)})} {\prod_{\alpha \in \Delta_{\lambda,0}^+} (1 - \mathrm{e}^{-w(\alpha)})} (\ln t) \;.$$ To prove this formula we establish certain properties of $I(t)$ which uniquely characterize it and are satisfied by the right-hand side. These are a weight expansion of $I(t)$ (see Corollary \[cor:weightexpansion\]), restrictions on the set of weights (see Corollary \[cor:weights\]), and the fact that $I(t)$ is annihilated by certain invariant differential operators (see Corollary \[cor:DlJchi\]). As was stated above, the case $K = \mathrm{U}_N$ is treated in [@CFZ]. Here, we restrict to the compact groups $K = \mathrm{O}_N \,$, $K = \mathrm{USp}_N\,$, and $K = \mathrm{SO}_N\,$. The cases $K = \mathrm{O}_N$ and $K = \mathrm{USp}_N$ can be treated simultaneously. Having established formula (\[eq:WeylCharacter\]) for $K = \mathrm{O}_N\,$, the following argument gives a similar result for $K = \mathrm{SO}_N\,$. Let $dk_\mathrm{O}$ and $dk_\mathrm{SO}$ be the unit mass Haar measures for $\mathrm{O}_N$ and $\mathrm{SO}_N\,$, respectively. The $\mathrm{O}_N$-measure $(1 + \mathrm{Det}\, k)\, dk_\mathrm{O}$ has unit mass on $\mathrm{SO}_N$ and zero mass on $\mathrm{O}_N^- = \mathrm{O}_N \setminus \mathrm{SO}_N\,$. It is $\mathrm{SO}_N$-invariant. Now, $$\begin{aligned} &&I_{\mathrm{SO}_N}(t) = \int_{\mathrm{SO}_N} Z(t,k)\, dk_\mathrm{SO} = \int_{\mathrm{O}_N} Z(t,k) (1+\mathrm{Det}\, k)\, dk_\mathrm{O} \\ = &&\int_{\mathrm{O}_N} Z(t,k)\, dk_\mathrm{O} + \int_{\mathrm{O}_N} Z(t,k) \mathrm{Det}(k) \, dk_\mathrm{O} = I_{\mathrm{O}_N}(t) + (-1)^N I_{\mathrm{O}_N}(t')\end{aligned}$$ with $t' = (\mathrm{e}^{- \mathrm{i}\psi_1}, \mathrm{e}^{ \mathrm{i} \psi_2} , \ldots, \mathrm{e}^{\mathrm{i}\psi_n}, \mathrm{e}^{\phi_1}, \ldots, \mathrm{e}^{\phi_n})$, since $\mathrm{Det}(k)\, Z(t,k) = (-1)^N Z(t',k)$. Comparison with results of other approaches ------------------------------------------- To facilitate the comparison with related work, we now present our final results in the following explicit form. Let $x_j := \mathrm{e}^{- \mathrm{i}\psi_j}$ and $y_l := \mathrm{e}^{-\phi_l}$. Consider first the case of the unitary symplectic group $K = \mathrm{USp}_N$ (where $N \in 2\mathbb{N}$). Then for any pair of non-negative integers $p,q$ in the range $q - p \le N+1$ one directly infers from (\[eq:WeylCharacter\]) the formula $$\int\limits_{\mathrm{USp}_N} \frac{\prod_{k=1}^p \mathrm{Det}(1 - x_k \, u)} {\prod_{l=1}^q \mathrm{Det} (1 - y_l \, u)} \, du = \sum_{\epsilon \in \{\pm 1\}^p} \, \frac{\prod_{k=1}^p x_k^{\frac{N}{2}(1-\epsilon_k)} \prod_{l=1}^q (1 - x_k^{\epsilon_k} y_l)}{\prod_{k\le k^\prime} (1 - x_k^{\epsilon_k} x_{k^\prime}^{\epsilon_{k^\prime}}) \prod_{l < l^\prime}(1 - y_l y_{l^\prime})} \;.$$ The sum on the right-hand side is over sign configurations $\epsilon \equiv (\epsilon_1, \ldots, \epsilon_p) \in \{ \pm 1 \}^p$. The proof proceeds by induction in $p\,$, starting from the result (\[eq:WeylCharacter\]) for $p = q$ and sending $x_p \to 0$ to pass from $p$ to $p-1$. In published recent work [@cfs; @bg] the same formula was derived under the more restrictive condition $q \le N/2\,$. In [@bg] this unwanted restriction on the parameter range came about because the numerator and denominator on the left-hand side were expanded *separately*, ignoring the supersymmetric Howe duality (see $\S$\[sect:osp-howe\] of the present paper) of the problem at hand. For $K = \mathrm{SO}_N$ the same induction process starting from (\[eq:WeylCharacter\]) yields the result $$\int\limits_{\mathrm{SO}_N} \frac{\prod_{k=1}^p \mathrm{Det}(1 - x_k \, u)} {\prod_{l=1}^q \mathrm{Det} (1 - y_l \, u)} \, du = \sum_{\epsilon \in \{ \pm 1\}^p} \frac{\prod_{k=1}^p (\epsilon_k x_k)^{\frac{N}{2}(1-\epsilon_k)} \prod_{l=1}^q (1 - x_k^{\epsilon_k} y_l)}{\prod_{k < k^\prime} (1 - x_k^{\epsilon_k} x_{k^\prime}^{\epsilon_{k^\prime}}) \prod_{l \le l^\prime}(1 - y_l y_{l^\prime})}$$ as long as $q-p \le N-1\,$. Please note that this includes even the case of the trivial group $K = \mathrm{SO}_1 = \{ \mathrm{Id} \}$ with any $p = q > 0\,$. For $K = \mathrm{O}_N$ one has an analogous result where the sum on the right-hand side is over $\epsilon$ with an *even* number of sign reversals. The very same formulas for $\mathrm{SO}_N$ and $\mathrm{O}_N$ were derived in the recent literature [@cfs; @bg] but, again, only in the much narrower range $q \le \mathrm{Int}[N/2]$. Howe duality and weight expansion --------------------------------- To find an explicit expression for the integral $I(t)$, we first of all observe that the integrand $Z(t,k)$ is the supertrace of a representation $\rho$ of a semigroup $(T_1 \times T_+) \times K$ on the spinor-oscillator module $\mathfrak{a}(V)$ (cf. Lemma \[lem:STrAV\]). More precisely, we start with the standard $K$-representation space $\mathbb{C}^N$, the $\mathbb{Z}_2$-graded vector space $U = U_0 \oplus U_1$ with $U_s \simeq \mathbb{C}^n$, and the abelian semigroup $$T_1 \times T_+ := \{ (\mathrm{diag}(\mathrm{e}^{\mathrm{i}\psi_1}, \ldots, \mathrm{e}^{\mathrm{i} \psi_n}), \mathrm{diag} (\mathrm{e}^{\phi_1}, \ldots, \mathrm{e}^{\phi_n})) \mid \mathfrak{Re}\, \phi_j > 0 \;, j = 1, \ldots, n \}$$ of diagonal transformations in $\mathrm{GL}(U_1) \times \mathrm{GL} (U_0)$. We then consider the vector space $V := U \otimes \mathbb{C}^N$ which is $\mathbb{Z}_2$-graded by $V_s = U_s \otimes \mathbb{C}^N$, the infinite-dimensional spinor-oscillator module $\mathfrak{a}(V) := \wedge (V_1^*) \otimes \mathrm{S} (V_0^*)$, and a representation $\rho$ of $(T_1 \times T_+) \times K$ on $\mathfrak{a}(V)$. We also let $V \oplus V^\ast =: W = W_0 \oplus W_1$ (not the Weyl group). Averaging the product of ratios $Z(t,k)$ with respect to the compact group $K$ corresponds to the projection from $\mathfrak{a}(V)$ onto the vector space $\mathfrak{a}(V)^K$ of $K$-invariants (Corollary \[cor:weightexpansion\]). Now, Howe duality (Proposition \[prop:Howeduality\]) implies that $\mathfrak{a}(V)^K$ is the representation space for an irreducible highest-weight representation $\rho_\ast$ of the Howe dual partner $\mathfrak{g}$ of $K$ in the orthosymplectic Lie superalgebra $\mathfrak{osp}(W)$. This representation $\rho_\ast$ is constructed by realizing $\mathfrak{g} \subset \mathfrak{osp}(W)$ as a subalgebra in the space of degree-two elements of the Clifford-Weyl algebra $\mathfrak{q}(W)$. Precise definitions of these objects, their relationships, and the Howe duality statement can be found in $\S$\[sect:osp-howe\]. Using the decomposition $\mathfrak{a}(V)^K = \oplus_{\gamma \in \Gamma}\, V_\gamma$ into weight spaces, Howe duality leads to the weight expansion $I(\mathrm{e}^H) = \mathrm{STr}_{\mathfrak{a}(V)^K} \, \mathrm{e}^{\rho_*(H)} = \sum_{\gamma \in \Gamma} B_\gamma \, \mathrm{e}^{\gamma(H)}$ for $t = \mathrm{e}^H \in T_1 \times T_+\,$. Here $\mathrm{STr}$ denotes the supertrace. There are strong restrictions on the set of weights $\Gamma$. Namely, if $\gamma \in \Gamma$, then $\gamma = \sum_{j=1}^n (\mathrm{i} m_j \psi_j - n_j \phi_j)$ and $- \frac{N}{2}\le m_j \le \frac{N}{2} \le n_j$ for all $j\,$. The coefficients $B_\gamma = \mathrm{STr}_{V_\gamma} (\mathrm{Id})$ are the dimensions of the weight spaces (multiplied by parity). Note that the set of weights of the representation $\rho_*$ of $\mathfrak{g}$ on $\mathfrak{a}(V)^K$ is infinite. Group representation and differential equations ----------------------------------------------- Before outlining the strategy for computing our character in the infinite-dimensional setting of representations of Lie superalgebras and groups, we recall the classical situation where $\rho_*$ is an irreducible finite-dimensional representation of a reductive Lie algebra $\mathfrak{g}$ and $\rho$ is the corresponding Lie group representation of the complex reductive group $G$. In that case the character $\chi $ of $\rho$, which automatically exists, is the trace $\mathrm{Tr}\, \rho\,$, which is a radial eigenfunction of every Laplace-Casimir operator. These differential equations can be completely understood by their behavior on a maximal torus of $G$. In our case we must consider the infinite-dimensional irreducible representation $\rho_*$ of the Lie superalgebra $\mathfrak{g} = \mathfrak{osp}$ on $\mathfrak{a}(V)^K$. Casimir elements, Laplace-Casimir operators of $\mathfrak{osp}$, and their radial parts have been described by Berezin [@B]. In the situation $U_0 \simeq U_1$ at hand we have the additional feature that every $\mathfrak{osp}$-Casimir element $I$ can be expressed as a bracket $I = [ \partial , F ]$ where $\partial$ is the holomorphic exterior derivative when we view $\mathfrak{a} (V)^K$ as the complex of $K$-equivariant holomorphic differential forms on $V_0\,$. To benefit from Berezin’s theory of radial parts, we construct a radial superfunction $\chi$ which is defined on an open set containing the torus $T_1 \times T_+$ such that its numerical part satisfies $\mathrm{num} \chi(t) = I(t)$ for all $t \in T_1\times T_+\,$. If we had a representation $(\rho_*,\rho)$ of a Lie supergroup $(\mathfrak{osp},G)$ at our disposal, we could define $\chi$ to be its character, i.e. $$\chi(g) \stackrel{?}{=} \mathrm{STr}_{\mathfrak{a}(V)^K} \rho(g)\, \mathrm{e}^{\sum_j \xi_j \, \rho_*(\Xi_j)}$$ (see $\S$\[supergroup characters\]). Since we don’t have such a representation, our idea is to define $\chi$ as a character on a totally real submanifold $M$ of maximal dimension which contains a real form of $T_1 \times T_+$ and is invariant with respect to conjugation by a real form $G_\mathbb{R}$ of $G\,$, and then to extend $\chi$ by analytic continuation. Thinking classically we consider the even part of the Lie superalgebra $\mathfrak{osp}(W_0 \oplus W_1)$, which is the Lie algebra $\mathfrak{o}(W_1) \oplus \mathfrak{sp}(W_0)$. The real structures at the Lie supergroup level come from a real form $W_\mathbb{R}$ of $W$. The associated real forms of $\mathfrak{o} (W_1)$ and $\mathfrak{sp}(W_0)$ are the real orthogonal Lie algebra $\mathfrak{o}(W_{1,\mathbb{R}})$ and the real symplectic Lie algebra $\mathfrak{sp}(W_{0,\mathbb{R}})$. These are defined in such a way that the elements in $\mathfrak{o}(W_{1,\mathbb{R}}) \oplus \mathfrak{sp} (W_{0,\mathbb{R}})$ and $\mathrm{i} W_\mathbb{R}$ are mapped as elements of the Clifford-Weyl algebra via the spinor-oscillator representation to anti-Hermitian operators on $\mathfrak{a}(V)$ with respect to a compatibly defined unitary structure. In this context we frequently use the unitary representation of the real Heisenberg group $\exp( \mathrm{i} W_{0, \mathbb{R}}) \times \mathrm{U}_1$ on the completion $\mathcal{A}_V$ of the module $\mathfrak{a}(V)$. Since $\wedge(V^*_1)$ has finite dimension, exponentiating the spinor representation of $\mathfrak{o}(W_{1,\mathbb{R}})$ causes no difficulties. This results in the spinor representation of $\mathrm{Spin}(W_{1,\mathbb{R}})$, a 2:1 covering of the compact group $\mathrm{SO}(W_{1,\mathbb{R}})$. So in this case one easily constructs a representation $R_1 : \, \mathrm{Spin} (W_{1, \mathbb{R}}) \to \mathrm{U} (\mathfrak{a}(V))$ which is compatible with $\rho_*|_{\mathfrak{o} (W_{1,\mathbb{R}})}$. Exponentiating the oscillator representation of $\mathfrak{sp} (W_{0,\mathbb{R}})$ on the infinite-dimensional vector space $\mathrm{S}(V_0^*)$ requires more effort. In $\S$\[metaplectic rep\], following Howe [@H2], we construct the Shale-Weil-Segal representation $R' : \, \mathrm{Mp}(W_{0, \mathbb{R}}) \to \mathrm{U} (\mathcal{A}_V)$ of the metaplectic group $\mathrm{Mp}(W_{0, \mathbb{R}})$ which is the 2:1 covering group of the real symplectic group $\mathrm{Sp} (W_{ 0,\mathbb{R}})$. This is compatible with $\rho_\ast \vert_{\mathfrak{sp} (W_{0, \mathbb{R}})}$. Altogether we see that the even part of the Lie superalgebra representation integrates to $G_\mathbb{R} = \mathrm{Spin} (W_{1 ,\mathbb{R}}) \times_{\mathbb{Z}_2} \mathrm{Mp}(W_{0,\mathbb{R}})\,$. The construction of $R^\prime$ uses a limiting process coming from the oscillator semigroup $\widetilde{\mathrm{H}}(W_0^s)$, which is the double covering of the contraction semigroup $\mathrm{H}(W_0^s) \subset \mathrm{Sp}(W_0)$ and has $\mathrm{Mp}(W_{0,\mathbb{R}})$ in its boundary. Furthermore, we have $\widetilde{ \mathrm{H}} (W_0^s) = \mathrm{Mp}(W_{0,\mathbb{R}}) \times M$ where $M$ is an analytic totally real submanifold of maximal dimension which contains a real form of the torus $T_+$ (see $\S$\[oscillator semigroup\]). The representation $R_0 : \,\widetilde{\mathrm{H}}(W_0^s)\to \mathrm{End} (\mathcal{A}_V)$ constructed in $\S$\[sec:semigrouprep\] facilitates the definition of the representation $R^\prime$ and of the character $\chi$ in $\S$\[subsec:existence\]. It should be underlined that Proposition \[holomorphicity\] ensures convergence of the superfunction $\chi(h)$, which is defined as a supertrace and exists for all $h \in \widetilde{\mathrm{H}}(W_0^s)$. On that basis, the key idea of our approach is to exploit the fact that every Casimir invariant $I \in \mathsf{U}(\mathfrak{g})$ is exact in the sense that $I = [ \partial , F]$. By a standard argument, this exactness property implies that every such invariant $I$ vanishes in the spinor-oscillator representation. This result in turn implies for our character $\chi$ the differential equations $D(I) \chi = 0$ where $D(I)$ is the Laplace-Casimir operator representing $I$. By drawing on Berezin’s theory of radial parts, we derive a system of differential equations which in combination with certain other properties ultimately determines $\chi$. In the case of $K = \mathrm{O}_N$ the Lie group associated to the even part of the real form of the Howe partner $\mathfrak{g}$ is embedded in a simple way in the full group $G_\mathbb{R}$ described above. It is itself just a lower-dimensional group of the same form. In the case of $K = \mathrm{USp}_N$ a sort of reversing procedure takes places and the analogous real form is $\mathrm{USp}_{2n} \times \mathrm{SO}_{2n}^\ast$. Nevertheless, the precise data which are used as input into the series developments, the uniqueness theorem and the final calculations of $\chi$ are essentially the same in the two cases. Therefore there is no difficulty handling them simultaneously. Howe dual pairs in the orthosymplectic Lie superalgebra {#sect:osp-howe} ======================================================= In this chapter we collect some foundational information from representation theory. Basic to our work is the orthosymplectic Lie superalgebra, $\mathfrak{osp}\,$, in its realization as the space spanned by supersymmetrized terms of degree two in the Clifford-Weyl algebra. Representing the latter by its fundamental representation on the spinor-oscillator module, one gets a representation of $\mathfrak{osp}$ and of all Howe dual pairs inside of $\mathfrak{osp} \,$. Roots and weights of the relevant representations are described in detail. Notion of Lie superalgebra -------------------------- A [*$\mathbb{Z}_2$-grading*]{} of a vector space $V$ over $\mathbb{K} = \mathbb{R}$ or $\mathbb{C}$ is a decomposition $V = V_0\oplus V_1$ of $V$ into the direct sum of two $\mathbb{K}$-vector spaces $V_0$ and $V_1\,$. The elements in $(V_0 \cup V_1) \setminus\{ 0\}$ are called [*homogeneous*]{}. The [*parity function*]{} $|\, |:(V_0 \cup V_1) \setminus \{ 0\} \to \mathbb{Z}_2\,$, $v\in V_s \mapsto |v| = s\,$, assigns to a homogeneous element its parity. We write $V \simeq \mathbb{K}^{p| q}$ if $\dim_\mathbb{K} V_0 = p$ and $\dim_\mathbb{K} V_1 = q\,$. A [*Lie superalgebra*]{} over $\mathbb{K}$ is a $\mathbb{Z}_2$-graded $\mathbb{K}$-vector space $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1$ equipped with a bilinear map $[\, , \,]: \, \mathfrak{g} \times \mathfrak{g} \to \mathfrak{g}$ satisfying 1. $[\mathfrak{g}_s , \mathfrak{g}_{s^\prime} ] \subset \mathfrak{g}_{s + s^\prime}\,$, i.e., $|[X,Y]|= |X| + |Y|$ (mod 2) for homogeneous elements $X,Y$. 2. Skew symmetry: $[X,Y] = -(-1)^{|X||Y|}[Y,X]$ for homogeneous $X,Y$. 3. Jacobi identity, which means that $\mathrm{ad}(X) = [X, \; ] : \, \mathfrak{g} \to \mathfrak{g}$ is a (super-)derivation: $$\mathrm{ad}(X)\, [Y,Z] = [\mathrm{ad}(X)Y , Z] + (-1)^{|X||Y|}[Y,\mathrm{ad}(X)Z] \;.$$ \[$\mathfrak{gl}(V)$\] Let $V = V_0 \oplus V_1$ be a $\mathbb{Z}_2$-graded $\mathbb{K}$-vector space. There is a canonical $\mathbb{Z}_2$-grading $\mathrm{End}(V) = \mathrm{End}(V)_0 \oplus \mathrm{End}(V)_1$ induced by the grading of $V$: $$\mathrm{End}(V)_s := \{X \in \mathrm{End}(V) \mid \forall s^\prime \in \mathbb{Z}_2 : \, X(V_{s^\prime}) \subset V_{s + s^\prime}\} \;.$$ The bilinear extension of $[X,Y] := XY - (-1)^{|X||Y|}YX$ for homogeneous elements $X , Y \in \mathrm{End}(V)$ to a bilinear map $[\, , \,] : \, \mathrm{End}(V) \times \mathrm{End}(V) \to \mathrm{End}(V)$ gives $\mathrm{End}(V)$ the structure of a Lie superalgebra, namely $\mathfrak{gl}(V)$. The Jacobi identity in this case is a direct consequence of the associativity $(XY)Z = X (YZ)$ and the definition of $[X,Y]$. In fact, for every $\mathbb{Z}_2$-graded associative algebra $\mathcal{A}$ the bracket $[\, , \,] : \, \mathcal{A} \times \mathcal{A} \to \mathcal{A}$ defined by $[X,Y] = XY - (-1)^{|X| |Y|} YX$ satisfies the Jacobi identity. Given a complex $\mathbb{Z}_2$-graded vector space $V = V_0 \oplus V_1\,$, consider the complex Lie superalgebra $\mathrm{End}(V) = \mathfrak{gl}(V)$. Equip $V$ with a unitary structure so that $V_0 \perp V_1\,$. For $X \in \mathrm{End}(V)$ we have the unique decomposition $X = \sum_{s=0}^1 X_s$ with $X_s \in \mathfrak{gl} (V)_s\,$. Let $\sigma : \, \mathfrak{gl}(V) \to \mathfrak{gl}(V)$ be the complex anti-linear involution defined by $\sigma(X)_s = - \mathrm{i}^s \overline{X}_s^\mathrm{\,t}\,$, where $X \mapsto \overline{X}^\mathrm{\,t}$ (complex conjugation and transpose) is defined by the unitary structure. $\sigma$ is an automorphism of the Lie superalgebra $\mathfrak{gl}(V)$. Indeed, $$\begin{aligned} \sigma([X,Y])&=&\sum\nolimits_{s,\, s^\prime} \sigma(X_s Y_{s^\prime})-\sum\nolimits_{s,\, s^\prime}(-1)^{s s^\prime} \sigma(Y_{s^\prime}X_s) \\&=&-\sum\nolimits_{s,\, s^\prime} \mathrm{i}^{(s-s^\prime)^2} \big(\overline{Y}_{s^\prime}^\mathrm{\,t} \overline{X}_s^\mathrm{\,t} - (-1)^{s s^\prime} \overline{X}_s^ \mathrm{\,t} \overline{Y}_{s^\prime}^\mathrm{\,t} \big) \;,\end{aligned}$$ which is the same as $$\begin{aligned} &&\sum\nolimits_{s,\, s^\prime} \mathrm{i}^{s^2 + {s^\prime}^2} \big(\overline{X}_s^\mathrm{\,t}\overline{Y}_{s^\prime}^\mathrm{ \,t}-(-1)^{s \,s^\prime}\overline{Y}_{s^\prime}^\mathrm{\,t} \overline{X}_s^\mathrm{\,t} \big) \\ &=& \sum\nolimits_{s,\, s^\prime}\sigma(X)_s \sigma(Y)_{ s^\prime} - \sum\nolimits_{s,\, s^\prime} (-1)^{s \, s^\prime} \sigma(Y)_{s^\prime} \sigma(X)_s = [\sigma(X),\sigma(Y)] \;.\end{aligned}$$ The set of fixed points $\mathrm{Fix} (\sigma) = \{X \in \mathfrak{gl}(V) \mid \sigma(X) = X \}$ is a real Lie superalgebra, a real form of $\mathfrak{gl} (V)$ called the unitary Lie superalgebra $\mathfrak{u} : = \mathfrak{u} (V)$. Note that $\mathfrak{u}_0$ is the compact real form $\mathfrak{u}(V_0) \oplus \mathfrak{u}(V_1)$ of the complex Lie algebra $\mathfrak{gl}(V_0) \oplus \mathfrak{gl} (V_1)$. \[exa:osp\] Let $V = V_0 \oplus V_1$ be a $\mathbb{Z}_2$-graded $\mathbb{K}$-vector space and put $W := V \oplus V^*$. The $\mathbb{Z}_2$-grading of $V$ induces a $\mathbb{Z}_2$-grading $W = W_0 \oplus W_1$ in the obvious manner: $W_0^{\vphantom{\ast}} = V_0^{\vphantom{\ast}} \oplus V_0^*$ and $W_1^{\vphantom{\ast}} = V_1^{\vphantom{\ast}} \oplus V_1^*\,$. Then consider the canonical alternating bilinear form $A$ on $W_0\,$, $$A : \,\, W_0 \times W_0 \to \mathbb{K} \;, \quad (v + \varphi \, , v' + \varphi') \mapsto \varphi'(v) - \varphi(v') \;,$$ and the canonical symmetric bilinear form $S$ on $W_1\,$, $$S : \,\, W_1 \times W_1 \to \mathbb{K} \;, \quad (v + \varphi \, , v' + \varphi') \mapsto \varphi'(v) + \varphi(v') \;.$$ The *orthosymplectic form* of $W$ is the non-degenerate bilinear form $Q : \, W \times W \to \mathbb{K}$ defined as the orthogonal sum $Q = A + S:$ $$Q(w_0^{\vphantom{\prime}} + w_1^{\vphantom{\prime}} , w'_0 + w'_1) = A(w_0^{\vphantom{\prime}} , w'_0) + S(w_1^{\vphantom{\prime}} , w'_1) \quad (w_s^{\vphantom{\prime}} \, , w_s^\prime \in W_s^{\vphantom{\prime}}) \;.$$ Note the exchange symmetry $Q(w,w^\prime) = -(-1)^{|w| |w^\prime|} Q(w^\prime, w)$ for $w, w^\prime \in W_0 \cup W_1\,$. Given $Q\,$, define a complex linear bijection $\tau : \, \mathrm{End}(W) \to \mathrm{End}(W)$ by the equation $$\label{eq:def-osp} Q(\tau(X) w , w') + (-1)^{|X||w|} Q(w, X w^\prime) = 0$$ for all $w, w^\prime \in W_0 \cup W_1\,$. It is easy to check that $\tau$ has the property $$\tau(XY) = - (-1)^{|X| |Y|} \tau(Y) \tau(X) \;,$$ which implies that $\tau$ is an involutory automorphism of the Lie superalgebra $\mathfrak{gl}(W)$ with bracket $[X,Y] = XY - (-1)^{|X| |Y|} YX\,$. Hence the subspace $\mathfrak{osp}(W) \subset \mathrm{End}(W)$ of $\tau$-fixed points is closed w.r.t. that bracket; it is called the (complex) [*orthosymplectic Lie superalgebra*]{} of $W$. \[exa:JH\] Using the notation of Example \[exa:osp\], consider the vector space $\widetilde{W} := W \oplus \mathbb{K}$ and take it to be $\mathbb{Z}_2$-graded by $\widetilde{W}_0 = W_0 \oplus \mathbb{K}$ and $\widetilde{W}_1 = W_1\,$. Define a bilinear mapping $[ \, , \,] : \, \widetilde{W} \times \widetilde{W} \to \widetilde{W}$ by $$[\mathbb{K}\, ,\widetilde{W}] = [\widetilde{W},\mathbb{K} ] = 0 \;, \quad [W , W] \subset \mathbb{K} \;, \quad [w , w^\prime] = Q(w,w^\prime) \quad (w,w^\prime \in W) \;.$$ By the basic properties of the orthosymplectic form $Q\,$, the vector space $\widetilde{W}$ equipped with this bracket is a Lie superalgebra – the so-called Jordan-Heisenberg algebra. Note that $\widetilde{W}$ is two-step nilpotent, i.e., $[\widetilde{W} , [ \widetilde{W} , \widetilde{W} ] ] = 0\,$. ### Supertrace Let $V = V_0\oplus V_1$ be a $\mathbb{Z}_2$-graded $\mathbb K$-vector space, and recall the decomposition $\mathrm{End}(V) = \bigoplus_{ s,\, t} \mathrm{Hom}(V_s \, , V_t)$. For $X \in \mathrm{End}(V)$, we denote by $X = \sum_{s,\, t} X_{\, t s}$ the corresponding decomposition of an operator. The [*supertrace*]{} on $V$ is the linear function $$\mathrm{STr} : \,\, \mathrm{End}(V) \to \mathbb{K} \;, \quad X \mapsto \mathrm{Tr}\, X_{00} - \mathrm{Tr}\, X_{11} = \sum\nolimits_s (-1)^s \mathrm{Tr}\, X_{s s} \;.$$ (If $\mathrm{dim}\, V = \infty\,$, then usually the domain of definition of $\mathrm{STr}$ must be restricted.) An $\mathrm{ad}$-*invariant bilinear form* on a Lie superalgebra $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1$ over $\mathbb{K}$ is a bilinear mapping $B : \, \mathfrak{g} \times \mathfrak{g} \to \mathbb{K}$ with the properties 1. $\mathfrak{g}_0$ and $\mathfrak{g}_1$ are $B$-orthogonal to each other ; 2. $B$ is symmetric on $\mathfrak{g}_0$ and skew on $\mathfrak{g}_1\,$; 3. $B([X,Y],Z) = B(X,[Y,Z])$ for all $X, Y, Z \in \mathfrak{g}\,$. We will repeatedly use the following direct consequences of the definition of $\mathrm{STr}$. \[lem:STr\] If $\mathfrak{g}$ is a Lie superalgebra in $\mathrm{End}(V)$, the trace form $B(X,Y) = \mathrm{STr}\, (XY)$ is an $\mathrm{ad} $-invariant bilinear form. One has $\mathrm{STr} \, [X,Y] = 0\,$. Recalling the setting of Example \[exa:osp\], note that the supertrace for $W = V \oplus V^\ast$ is odd under the $\mathfrak{gl}$-automorphism $\tau$ fixing $\mathfrak{osp}(W)$, i.e., $\mathrm{STr}_W \circ \tau = - \mathrm{STr}_W\,$. It follows that $\mathrm{STr}_W X = 0$ for any $X \in \mathfrak{osp}(W)$. Moreover, $\mathrm{STr}_W (X_1 X_2 \cdots X_{2n+1}) = 0$ for any product of an odd number of $\mathfrak{osp}$-elements. ### Universal enveloping algebra {#sect:UEA} Let $\mathfrak{g}$ be a Lie superalgebra with bracket $[ \, , \, ]$. The universal enveloping algebra $\mathsf{U}(\mathfrak{g})$ is defined as the quotient of the tensor algebra $\mathsf{T}(\mathfrak{g}) = \oplus_{n = 0}^\infty \mathsf{T}^n (\mathfrak{g})$ by the two-sided ideal $\mathsf{J} (\mathfrak{g})$ generated by all combinations $$X \otimes Y - (-1)^{|X| |Y|} Y \otimes X - [X,Y]$$ for homogeneous $X , Y \in \mathsf{T}^1( \mathfrak{g}) \equiv \mathfrak{g}\,$. If $\mathsf{U}_n(\mathfrak{g})$ is the image of $\mathsf{T}_n(\mathfrak{g}) := \oplus_{k = 0}^n \mathsf{T}^k (\mathfrak{g})$ under the projection $\mathsf{T} (\mathfrak{g}) \to \mathsf{U}(\mathfrak{g}) = \mathsf{T}(\mathfrak{g}) / \mathsf{J} (\mathfrak{g})$, the algebra $\mathsf{U}(\mathfrak{g})$ is filtered by $\mathsf{U}( \mathfrak{g}) = \cup_{n=0}^\infty \mathsf{U}_n (\mathfrak{g})$. The $\mathbb{Z}_2$-grading $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1$ gives rise to a $\mathbb{Z}_2$-grading of $\mathsf{T}(\mathfrak{g})$ by $$| X_1 \otimes X_2 \otimes \cdots \otimes X_n | = \sum\nolimits_{i = 1}^n | X_i | \quad (\text{for homogenous } X_i \in \mathfrak{g})\;,$$ and this in turn induces a canonical $\mathbb{Z}_2$-grading of $\mathsf{U} (\mathfrak{g})$. One might imagine introducing various bracket operations on $\mathsf {T}(\mathfrak{g})$ and/or $\mathsf{U}(\mathfrak{g})$. However, in view of the canonical $\mathbb{Z}_2$-grading, the natural bracket operation to use is the *supercommutator*, which is the bilinear map $\mathsf{T}(\mathfrak{g}) \times \mathsf{T}(\mathfrak{g}) \to \mathsf{T}(\mathfrak{g})$ defined by $\{ a , b \} := a b - (-1)^{|a||b|} b a$ for homogeneous elements $a, \, b \in \mathsf{T}(\mathfrak{g})$. (For the time being, we use a different symbol $\{ \, , \, \}$ for better distinction from the bracket $[\, , \,]$ on $\mathfrak{g}\,$.) Since by the definition of $\mathsf{J} (\mathfrak{g})$ one has $$\{ \mathsf{T}(\mathfrak{g}),\mathsf{J}(\mathfrak{g}) \} = \{ \mathsf{J}(\mathfrak{g}), \mathsf{T}(\mathfrak{g}) \} \subset \mathsf{J}(\mathfrak{g}) \;,$$ the supercommutator descends to a well-defined map $\{\,,\,\} : \, \mathsf{U}(\mathfrak{g})\times\mathsf{U}(\mathfrak{g}) \to \mathsf{U}(\mathfrak{g})$. \[lem:2.2\] If $\mathfrak{g}$ is a Lie superalgebra, the supercommutator $\{\,,\,\}$ gives $\mathsf{U}(\mathfrak{g})$ the structure of another Lie superalgebra in which $\{ \mathsf{U}_n(\mathfrak{g}) , \mathsf{U}_{n^\prime}(\mathfrak{g})\} \subset \mathsf{U}_{n + n^\prime - 1} (\mathfrak{g})$. Compatibility with the $\mathbb{Z}_2$-grading, skew symmetry, and Jacobi identity are properties of $\{ \, , \, \}$ that are immediate at the level of the tensor algebra $\mathsf{T}(\mathfrak{g})$. They descend to the corresponding properties at the level of $\mathsf{U} (\mathfrak{g})$ by the definition of the two-sided ideal $\mathsf{J} (\mathfrak{g})$. Thus $\mathsf{U}(\mathfrak{g})$ with the bracket $\{ \, , \, \}$ is a Lie superalgebra. To see that $\{\mathsf{U}_n(\mathfrak{g}),\mathsf{U}_{n^\prime} (\mathfrak{g})\}$ is contained in $\mathsf{U}_{n + n^\prime - 1} (\mathfrak{g})$, notice that this property holds true for $n = n^\prime = 1$ by the defining relations $\mathsf{J}(\mathfrak{g}) \equiv 0$ of $\mathsf{U} (\mathfrak{g})$. Then use the associative law for $\mathsf{U}(\mathfrak{g})$ to verify the formula $$\{ a , bc \} = a b c - (-1)^{|a| (|b| + |c|)} b c a = \{ a , b \} c + (-1)^{|a| |b|} b \{ a , c \}$$ for homogeneous $a, b, c \in \mathsf{U}(\mathfrak{g})$. The claim now follows by induction on the degree of the filtration $\mathsf{U} (\mathfrak{g}) = \cup_{n = 0}^\infty \mathsf{U}_n (\mathfrak{g})$. By definition, the supercommutator of $\mathsf{U}(\mathfrak{g})$ and the bracket of $\mathfrak{g}$ agree at the linear level: $\{X,Y\} \equiv [X,Y]$ for $X, Y \in \mathfrak{g}\,$. It is therefore reasonable to drop the distinction in notation and simply write $[ \, , \,]$ for both of these product operations. This we now do. For future use, note the following variant of the preceding formula: if $Y_1, \ldots, Y_k\, , X$ are any homogeneous elements of $\mathfrak{g}\,$, then $$\label{eq:brackets} [Y_1 \cdots Y_k \, , \, X] = \sum_{i=1}^k (-1)^{|X|\sum_{j = i+1}^k |Y_j|}\, Y_1 \cdots Y_{i-1}\, [Y_i\, , X] \, Y_{i+1} \cdots Y_k \;,$$ which expresses the supercommutator in $\mathsf{U}(\mathfrak{g})$ by the bracket in $\mathfrak{g}\,$. Structure of $\mathfrak{osp}(W)$ {#subsec:osp} -------------------------------- For a $\mathbb{Z}_2$-graded $\mathbb{K}$-vector space $V = V_0 \oplus V_1$ let $W = V \oplus V^\ast = W_0 \oplus W_1$ as in Example \[exa:osp\]. The [*orthogonal Lie algebra*]{} $\mathfrak{o}(W_1)$ is the Lie algebra of the Lie group $\mathrm{O}(W_1)$ of $\mathbb{K}$-linear transformations of $W_1$ that leave the non-degenerate symmetric bilinear form $S$ invariant. This means that $X \in \mathrm{End}(W_1)$ is in $\mathfrak{o}(W_1)$ if and only if $$\forall w, w^\prime \in W_1 : \,\, S(Xw,w') + S(w,Xw') = 0 \;.$$ Similarly, the [*symplectic Lie algebra*]{} $\mathfrak{sp}(W_0)$ is the Lie algebra of the automorphism group $\mathrm{Sp}(W_0)$ of $W_0$ equipped with the non-degenerate alternating bilinear form $A:$ $$\mathfrak{sp}(W_0) = \{ X\in\mathrm{End}(W_0) \mid \forall w ,w' \in W_0 : \,\, A(Xw ,w') + A(w,Xw') = 0 \} \;.$$ For the next statement, recall the definition of the orthosymplectic Lie superalgebra $\mathfrak{osp}(W)$ and the decomposition $\mathfrak{osp}(W) = \mathfrak{osp}(W)_0 \oplus \mathfrak{osp}(W)_1\,$. \[lem:iso-osp1\] As Lie algebras resp. vector spaces, $$\mathfrak{osp}(W)_0 \simeq \mathfrak{o}(W_1)\oplus\mathfrak{sp}(W_0) \;, \quad \mathfrak{osp}(W)_1 \simeq W_1^{\vphantom{\ast}} \otimes W_0^* \;.$$ The first isomorphism follows directly from the definitions. For the second isomorphism, decompose $X \in \mathfrak{osp}(W)_1$ as $X = X_{01}+ X_{10}$ where $X_{01} \in \mathrm{Hom}(W_1,W_0)$ and $X_{10} \in \mathrm{Hom}(W_0\, ,W_1)$. Then $$\mathfrak{osp}(W)_1 = \{ X_{10} + X_{01} \in \mathrm{End}(W)_1 \mid \forall w_s \in W_s : \, S(X_{10}w_0\, ,w_1) + A(w_0\, ,X_{01}w_1) = 0 \}\;.$$ Since both $S$ and $A$ are non-degenerate, the component $X_{01}$ is determined by the component $X_{10}\,$, and one therefore has $\mathfrak{osp}(W)_1 \simeq \mathrm{Hom}(W_0\, , W_1) \simeq W_1^{\vphantom{\ast}} \otimes W_0^*\,$. We now review how $\mathfrak{sp}(W_0)$ and $\mathfrak{o}(W_1)$ decompose for our case $W_s = V_s^{\vphantom{\ast}} \oplus V_s^{\ast} \,$. For that purpose, if $U$ is a vector space with dual vector space $U^\ast$, let $\mathrm{Sym}(U,U^\ast)$ and $\mathrm{Alt}(U, U^\ast)$ denote the symmetric resp. alternating linear maps from $U$ to $U^\ast$. \[lem:iso-osp2\] As vector spaces, $$\begin{aligned} \mathfrak{o}(W_1) &\simeq \mathrm{End}(V_1) \oplus \mathrm{Alt} (V_1^{\vphantom{\ast}}, V_1^\ast) \oplus \mathrm{Alt} (V_1^\ast , V_1^{\vphantom{\ast}}) \;, \\ \mathfrak{sp}(W_0) &\simeq \mathrm{End}(V_0) \oplus \mathrm{Sym}(V_0^{\vphantom{\ast}}\, , V_0^\ast)\oplus \mathrm{Sym}(V_0^\ast , V_0^{\vphantom{\ast}})\;.\end{aligned}$$ There is a canonical decomposition $$\mathrm{End}(W_s) = \mathrm{End}(V_s^{\vphantom{\ast}}) \oplus \mathrm{Hom}(V_s^\ast, V_s^{\vphantom{\ast}}) \oplus \mathrm{Hom} (V_s^{\vphantom{*}}\,,V_s^*)\oplus \mathrm{End}(V_s^*)$$ for $s = 0, 1$. Let $s = 1$ and write the corresponding decomposition of $X \in \mathrm{End}(W_1)$ as $$X = \mathsf{A} \oplus \mathsf{B} \oplus \mathsf{C} \oplus \mathsf{D} \;.$$ Substituting $w = v + \varphi$ and $w^\prime = v^\prime + \varphi^\prime$, the defining condition $S(Xw,w^\prime) = - S(w,X w^\prime)$ for $X \in \mathfrak{o}(W_1)$ then transcribes to $$\varphi^\prime(\mathsf{A}v) = - (\mathsf{D}\varphi^\prime)(v) \;, \quad (\mathsf{C}v)(v^\prime) = - (\mathsf{C}v^\prime)(v) \;, \quad \varphi^\prime(\mathsf{B}\varphi) = - \varphi(\mathsf{B}\varphi^\prime) \;,$$ for all $v, v^\prime \in V_1$ and $\varphi, \varphi^\prime \in V_1^\ast$. Thus $\mathsf{D} = - \mathsf{A}^\mathrm{t}$, and the maps $\mathsf{B}, \mathsf{C}$ are alternating. This already proves the statement for the case of $\mathfrak{o}(W_1)$. The situation for $\mathfrak{sp}(W_0)$ is identical but for a sign change: the symmetric form $S$ is replaced by the alternating form $A\,$, and this causes the parity of $\mathsf{B}, \mathsf{C}$ to be reversed. By adding up dimensions, Lemmas \[lem:iso-osp1\] and \[lem:iso-osp2\] entail the following consequence. \[cor:dimosp\] As a $\mathbb{Z}_2$-graded vector space, $\mathfrak{osp}(V \oplus V^\ast)$ is isomorphic to $\mathbb{K}^{p|q}$ where $p = d_0 (2d_0 + 1) + d_1 (2d_1 - 1)$, $q = 4 d_0 d_1\,$, and $d_s = \dim V_s \,$. There exists another way of thinking about $\mathfrak{osp}(W)$, which will play a key role in the sequel. To define it and keep the sign factors consistent and transparent, we need to be meticulous about our ordering conventions. Hence, if $v \in V$ is a vector and $\varphi \in V^\ast$ is a linear function, we write the value of $\varphi$ on $v$ as $$\varphi(v) \equiv \langle v , \varphi \rangle \;.$$ Based on this notational convention, if $V$ is a $\mathbb {Z}_2$-graded vector space and $X \in \mathrm{End}(V)$ is a homogeneous operator, we define the *supertranspose* $X^\mathrm{st} \in \mathrm{End}(V^\ast)$ of $X$ by $$\langle v , X^\mathrm{st} \varphi \rangle := (-1)^{|X| |v|} \langle Xv , \varphi \rangle \quad (v \in V_0 \cup V_1 , \,\, \varphi \in V^\ast) \;.$$ This definition differs from the usual transpose by a change of sign in the case when $X$ has a component in $\mathrm{Hom}(V_1 , V_0)$. From it, it follows directly that the negative supertranspose $\mathfrak{gl}(V) \to \mathfrak{gl}(V^\ast)$, $X \mapsto -X^\mathrm{st}$ is an isomorphism of Lie superalgebras: $$- [X,Y]^\mathrm{st} = [-X^\mathrm{st} , -Y^\mathrm{st}] \;.$$ The modified notion of transpose goes hand in hand with a modified notion of what it means for an operator in $\mathrm{Hom}(V,V^\ast)$ or $\mathrm{Hom}(V^\ast , V)$ to be symmetric. Thus, define the subspace $\mathrm{Sym}(V^\ast,V) \subset \mathrm{Hom}(V^\ast,V)$ to consist of the elements, say $\mathsf{B}$, which are symmetric in the $\mathbb{Z}_2$-graded sense: $$\label{eq:sym-B} \forall \, \varphi, \varphi^\prime \in V_0^\ast \cup V_1^\ast : \quad \langle \mathsf{B} \varphi , \varphi^\prime \rangle = \langle \mathsf{B} \varphi^\prime , \varphi \rangle \, (-1)^{|\varphi| |\varphi^\prime|} \;.$$ By the same principle, define $\mathrm{Sym}(V,V^\ast) \subset \mathrm{Hom}(V,V^\ast)$ as the set of solutions $\mathsf{C}$ of $$\label{eq:sym-C} \forall \, v, v^\prime \in V_0 \cup V_1 : \quad \langle v , \mathsf{C}v^\prime \rangle= \langle v^\prime, \mathsf{C} v\rangle \, (-1)^{|v| |v^\prime| + |v| + |v^\prime|} \;.$$ To make the connection with the decomposition of Lemma \[lem:iso-osp1\] and \[lem:iso-osp2\], notice that $$\mathrm{Sym}(V,V^\ast)\cap \mathrm{Hom}(V_s^{\vphantom{\ast}} \,, V_s^\ast) = \left\{\begin{array}{ll}\mathrm{Sym} (V_0^{\vphantom{\ast}}\,, V_0^\ast) &s = 0 \;,\\ \mathrm{Alt}(V_1^{\vphantom{\ast}} , V_1^\ast) &s = 1 \;, \end{array} \right.$$ and similar for the corresponding intersections involving $\mathrm{Sym}(V^\ast,V)$. Next, expressing the orthosymplectic form $Q$ of $W = V \oplus V^\ast$ as $$Q(v + \varphi , v^\prime + \varphi^\prime) = \langle v , \varphi^\prime \rangle - (-1)^{|v^\prime| |\varphi|} \langle v^\prime , \varphi \rangle \;,$$ and writing out the conditions resulting from $Q(Xw,w^\prime) + (-1)^ {|X| |w|} Q(w,Xw^\prime) = 0$ for the case of $X \equiv \mathsf{B} \in \mathrm{Hom}(V^\ast,V)$ and $X \equiv \mathsf{C} \in \mathrm{Hom}(V,V^\ast)$, one sees that $$\mathfrak{osp}(W) \cap \mathrm{Hom}(V,V^\ast) = \mathrm{Sym}(V,V^\ast)\;, \quad \mathfrak{osp}(W) \cap \mathrm{Hom}(V^\ast,V) = \mathrm{Sym}(V^\ast,V) \;.$$ This situation is summarized in the next statement. \[lem:osp-iso3\] The orthosymplectic Lie superalgebra of $\, W = V \oplus V^\ast$ decomposes as $$\mathfrak{osp}(W) = \mathfrak{g}^{(-2)}\oplus \mathfrak{g}^{(0)}\oplus\mathfrak{g}^{(+2)}\;,$$ where $\mathfrak{g}^{(+2)} := \mathrm{Sym}(V,V^\ast)$, and $\mathfrak{g}^{(-2)} := \mathrm{Sym}(V^\ast,V)$, and $$\mathfrak{g}^{(0)} := (\mathrm{End}(V)\oplus \mathrm{End}(V^\ast))\cap \mathfrak{osp}(W)\;.$$ The decomposition of Lemma \[lem:osp-iso3\] can be regarded as a $\mathbb{Z}$-grading of $\mathfrak{osp}(W)$. By the ‘block’ structure inherited from $W = V \oplus V^\ast$, this decomposition is compatible with the bracket $[\, , \,]:$ $$[ \mathfrak{g}^{(m)} , \, \mathfrak{g}^{(m^\prime)} ] \subset \mathfrak{g}^{(m + m^\prime)} \;,$$ where $\mathfrak{g}^{(m + m^\prime)} \equiv 0$ if $m + m^\prime \notin \{ \pm 2, 0\}$. It follows that each of the three subspaces $\mathfrak{g}^{( +2)}$, $\mathfrak{g}^{(-2)}$, and $\mathfrak{g}^{ (0)}$ is a Lie superalgebra, the first two with vanishing bracket. \[lem:iso-osp4\] The embedding $\mathrm{End}(V)\to \mathrm{End}(V) \oplus \mathrm{End} (V^\ast)$ by $\mathsf{A} \mapsto \mathsf{A} \oplus (-\mathsf{A}^ \mathrm{st})$ projected to $\mathfrak{osp}(W)$ is an isomorphism of Lie superalgebras $\mathfrak{gl}(V) \to \mathfrak{g}^{(0)}$. Since the negative supertranspose $\mathsf{A} \mapsto - \mathsf{A} ^\mathrm{st}$ is a homomorphism of Lie superalgebras, so is our embedding $\mathsf{A} \mapsto \mathsf{A} \oplus (- \mathsf{A} ^\mathrm{st})$. This map is clearly injective. To see that it is surjective, consider any homogeneous $X = \mathsf{A} \oplus \mathsf{D} \in \mathrm{End}(V) \oplus \mathrm{End}(V^\ast)$ viewed as an operator in $\mathrm{End}(W)$. The condition for $X$ to be in $\mathfrak{osp}(W)$ is (\[eq:def-osp\]). To get a non-trivial condition, choose $(w,w^\prime) = (v,\varphi)$ or $(w,w^\prime) = (\varphi,v)$. The first choice gives $$Q(Xv,\varphi) + (-1)^{|X||v|} Q(v,X\varphi) = \langle \mathsf{A} v , \varphi \rangle + (-1)^{|\mathsf{A}||v|} \langle v , \mathsf{D}\varphi \rangle = 0 \;.$$ Valid for all $v \in V_0 \cup V_1$ and $\varphi \in V^\ast$, this implies that $\mathsf{D} = - \mathsf{A}^\mathrm{st}$. The second choice leads to the same conclusion. Thus $X = \mathsf{A} \oplus \mathsf{D}$ is in $\mathfrak{osp}(W)$ if and only if $\mathsf{D} = - \mathsf{A}^\mathrm{st}$. In the following subsections we will often write $\mathfrak{osp}(W) \equiv \mathfrak{osp}$ for short. ### Roots and root spaces {#sect:osp-roots} A Cartan subalgebra of a Lie algebra $\mathfrak{g}_0$ is a maximal commutative subalgebra $\mathfrak{h} \subset \mathfrak{g}_0$ such that $\mathfrak{g}_0$ (or its complexification if $\mathfrak{g}_0$ is a real Lie algebra) has a basis consisting of eigenvectors of $\mathrm{ad}(H)$ for all $H \in \mathfrak{h} \,$. Recall that $|[X,Y]|=|X|+|Y|$ for homogeneous elements $X,Y$ of a Lie superalgebra $\mathfrak{g}\,$. From the vantage point of decomposing $\mathfrak{g}$ by eigenvectors or root spaces, it is therefore reasonable to call a Cartan subalgebra of $\mathfrak{g}_0$ a Cartan subalgebra of $\mathfrak{g}\,$. We will see that $X \in \mathfrak{osp}_1$ and $[X,H] = 0$ for all $H \in \mathfrak{h} \subset \mathfrak{osp}_0$ imply $X = 0\,$, i.e., there exists no commutative subalgebra of $\mathfrak{osp}$ that properly contains a Cartan subalgebra. Lie superalgebras with this property are called of type I in [@B]. Let us determine a Cartan subalgebra and the corresponding root space decomposition of $\mathfrak{osp}\,$. For $s, t = 0, 1$ choose bases $\{ e_{s, 1}\, , \ldots, \, e_{s,\,d_s} \}$ of $V_s$ and associated dual bases $\{ f_{t,1} , \ldots, f_{t,\, d_t} \}$ of $V_t^\ast\,$. Then for $j = 1, \ldots, d_s$ and $k = 1, \ldots, d_t$ define rank-one operators $E_{s, j \, ; \, t ,\, k}$ by the equation $E_{s,j\, ; \, t,\, k} (e_{u,l}) = e_{s,j}\, \delta_{t,u} \, \delta_{k,l}$ for all $u = 0, 1$ and $l = 1, \ldots, d_u \,$. These form a basis of $\mathrm{End}(V)$, and by Lemma \[lem:iso-osp4\] the operators $$X_{s j, \, t k}^{(0)} := E_{s , j\,; \, t ,\, k} \oplus (- E_{s,j\,;\,t,\,k})^\mathrm{st}$$ form a basis of $\mathfrak{g}^{(0)}$. Similarly, let bases of $\mathrm{Hom}(V^\ast,V)$ and $\mathrm{Hom}(V,V^\ast)$ be defined by $$F_{s,j\, ; \,t,\,k} (f_{u,l}) = e_{s,j} \, \delta_{t,u} \, \delta_{k,l} \;, \qquad \tilde{F}_{s,j\,;\,t,\,k} (e_{u,l}) = f_{s,j} \,\delta_{t,u}\, \delta_{k,l} \;,$$ for index pairs in the appropriate range. Then by Lemma \[lem:osp-iso3\] and equations (\[eq:sym-B\], \[eq:sym-C\]) the subalgebras $\mathfrak{g}^{(-2)}$ and $\mathfrak{g}^{(2)}$ are generated by the sets of operators $$\begin{aligned} &&X_{s j, \, t k}^{(-2)} := F_{s,j\,;\,t,\,k} + F_{t,\,k\,;\,s,j} \, (-1)^{|s| |t|} \;, \\ &&X_{s j, \, t k}^{(2)} := \tilde{F}_{s,j\,;\,t,\,k} + \tilde{F}_{t,\,k\,;\,s,j} \, (-1)^{|s| |t| + |s| + |t|}\;.\end{aligned}$$ Since $\mathfrak{osp}_0 \simeq \mathfrak{o}(W_1) \oplus \mathfrak{sp}(W_0)$, a Cartan subalgebra of $\mathfrak{osp}$ is the direct sum of a Cartan subalgebra of $\mathfrak{o}(W_1)$ and a Cartan subalgebra of $\mathfrak{sp}(W_0)$. Letting $\mathfrak{h}$ be the span of the diagonal operators $$H_{s j} := X_{s j ,\, s j}^{(0)}\quad (s=0,1;\, j= 1,\ldots,d_s)\;,$$ one has that $\mathfrak{h}$ is a Cartan subalgebra of $\mathfrak{osp} \,$. Indeed, if $\{ \vartheta_{s j} \}$ denotes the basis of $\mathfrak{h}^\ast$ dual to $\{ H_{s j} \}\,$, inspection of the adjoint action of $\mathfrak{h}$ on $\mathfrak{osp}$ gives the following result. \[lem:rootsosp\] The operators $X_{s j\, , \, t k}^{(m)}$ are eigenvectors of $\mathrm{ad}(H)$ for all $H \in \mathfrak{h}:$ $$[H\, ,X^{(m)}_{s j\, ,\, t k}] = \begin{cases} (\vartheta_{s j} - \vartheta_{t k})(H) \, X^{(m)}_{s j\, ,\, t k} & m = 0 \;, \\ (\vartheta_{s j} + \vartheta_{t k})(H) \, X^{(m)}_{s j\, , \, t k} & m = -2 \;, \\ (- \vartheta_{s j} - \vartheta_{t k})(H)\, X^{(m)}_{s j\,, \,t k} & m = 2\;.\end{cases}$$ A root of a Lie superalgebra $\mathfrak{g}$ is called [*even*]{} if its root space is in $\mathfrak{g}_0\,$, it is called [*odd*]{} if its root space is in $\mathfrak{g}_1\,$. We denote by $\Delta_0$ and $\Delta_1$ the set of even roots and the set of odd roots, respectively. For $\mathfrak{g} = \mathfrak{osp}$ we have $$\Delta_0 =\{ \pm \vartheta_{1j} \pm \vartheta_{1k}\, , \, \pm \vartheta_{0 j} \pm \vartheta_{0 l} \mid j < k , \, j\leq l\} , \quad \Delta_1 = \{ \pm \vartheta_{1 j} \pm \vartheta_{0 k} \}.$$ ### Casimir elements {#sect:osp-cas} As before, let $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1$ be a Lie superalgebra, and let $\mathsf{U}(\mathfrak{g}) = \cup_{n = 0}^\infty \mathsf{U}_n (\mathfrak{g})$ be its universal enveloping algebra. Denote the symmetric algebra of $\mathfrak{g}_0$ by $\mathrm{S}(\mathfrak{g}_0)$ and the exterior algebra of $\mathfrak{g}_1$ by $\wedge(\mathfrak{g}_1)$. The Poincaré-Birkhoff-Witt theorem for Lie superalgebras states that for each $n$ there is a bijective correspondence $$\mathsf{U}_n (\mathfrak{g}) / \mathsf{U}_{n-1} (\mathfrak{g}) \stackrel{\sim}{\to} \sum\nolimits_{k +l = n} \wedge^k( \mathfrak{g}_1 ) \otimes \mathrm{S}^l (\mathfrak{g}_0) \;.$$ The collection of inverse maps lift to a vector-space isomorphism, $$ \wedge(\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{g}_0) \stackrel{\sim}{\to} \mathsf{U}(\mathfrak{g}) \;,$$ called the super-symmetrization mapping. In other words, given a homogeneous basis $\{ e_1, \ldots, e_d \}$ of $\mathfrak{g} \,$, each element $x \in \mathsf{U}(\mathfrak{g})$ can be uniquely represented in the form $x = \sum_n \sum_{i_1, \ldots, \, i_n} x_{i_1, \ldots, \, i_n}\, e_{i_1} \cdots \, e_{i_n}$ with super-symmetrized coefficients, i.e., $$x_{i_1,\ldots,\, i_l,\, i_{l+1}, \ldots, \, i_n} = (-1)^{|e_{i_l}| |e_{i_{l+1}}|} x_{i_1,\, \ldots,\, i_{l+1},\, i_l,\, \ldots,\, i_n} \quad (1 \le l < n) \;.$$ The isomorphism $\wedge(\mathfrak{g}_1) \otimes \mathrm{S} (\mathfrak{g}_0) \simeq \mathsf{U}(\mathfrak{g})$ gives $\mathsf{U} (\mathfrak{g})$ a $\mathbb{Z}$-grading (by the degree $n$). Now recall that $\mathsf{U}(\mathfrak{g})$ comes with a canonical bracket operation, the supercommutator $[\, , \,] : \, \mathsf{U} (\mathfrak{g}) \times \mathsf{U}(\mathfrak{g}) \to \mathsf{U} (\mathfrak{g})$. An element $X \in \mathsf{U}(\mathfrak{g})$ is said to lie in the center of $\mathsf{U} (\mathfrak{g})$, and is called a *Casimir element*, iff $[X,Y]= 0$ for all $Y \in \mathsf{U} (\mathfrak{g})$. By the formula (\[eq:brackets\]), a necessary and sufficient condition for that is $[X,Y] = 0$ for all $Y \in \mathfrak{g}\,$. In the case of $\mathfrak{g} = \mathfrak{osp}\,$, for every $\ell \in \mathbb{N}$ there is a Casimir element $I_\ell$ of degree $2\ell$, which is constructed as follows. Consider the bilinear form $B : \, \mathfrak{osp} \times \mathfrak{osp} \to \mathbb{K}$ given by the supertrace (in some representation), $B(X,Y) : = \mathrm{STr}\, (XY)$. Recall that this form is $\mathrm{ad}$-invariant, which is to say that $B([X,Y],Z) = B(X,[Y,Z])$ for all $X, Y, Z \in \mathfrak{g}\,$. Taking the supertrace in the fundamental representation of $\mathfrak{osp} \,$, the form $B$ is non-degenerate, and therefore, if $e_1,\, \ldots, \, e_d$ is a homogeneous basis of $\mathfrak{osp}\,$, there is another homogeneous basis $\widetilde{e}_1 , \, \ldots, \, \widetilde{e}_d$ of $\mathfrak{osp}$ so that $B(\widetilde{e}_i \, , \, e_j) = \delta_{ij}\,$. Note $|\widetilde{e}_i | = |e_i|$ and put $$\label{eq:CasimirElement} I_\ell := \sum_{i_1,\,\ldots,\,i_{2\ell}=1}^d \widetilde{e}_ {i_1}\cdots\, \widetilde{e}_{i_{2\ell}}\,\mathrm{STr}\, (e_{i_{2\ell}}\cdots\, e_{i_1})\in \mathsf{U}(\mathfrak{osp}).$$ (Notice that, in view of the remark following Lemma \[lem:STr\], there is no point in making the same construction with an odd number of factors.) \[lem:IlCasimir\] For all $\ell \in \mathbb N$ the element $I_\ell$ is Casimir, and $|I_\ell| = 0$. By specializing the formula (\[eq:brackets\]) to the present case, $$[I_\ell\, , \,X] = \sum \sum_{k=1}^{2\ell} (-1)^{|X| (|\widetilde{e}_{i_{k+1}}| + \ldots +|\widetilde{e}_{i_{2\ell}}|)} \widetilde{e}_{i_1} \cdots \, \widetilde{e}_{i_{k-1}} \, [ \widetilde{e}_{i_k} , X] \, \widetilde{e}_{i_{k+1}} \cdots \, \widetilde{e}_{i_{2\ell}}\,\mathrm{STr}\,(e_{i_{2\ell}}\cdots\, e_{i_1})\;.$$ Now if $[\widetilde{e}_i \, , X] = \sum_j X_{ij}\, \widetilde{e}_j$ then from $\mathrm{ad}$-invariance, $B([\widetilde{e}_i \, , X], e_j) = X_{ij} = B(\widetilde{e}_i \, , [X,e_j])$, one has $[X,e_j] = \sum_i e_i\, X_{ij}\,$. Using this relation to transfer the $\mathrm{ad}(X)$-action from $\widetilde{e}_{i_k}$ to $e_{i_k}\,$, and reading the formula (\[eq:brackets\]) backwards, one obtains $$[I_\ell\,,\,X] = \sum \widetilde{e}_{i_1}\cdots \,\widetilde{e}_{ i_{2\ell}}\, \mathrm{STr}\,([X,e_{i_{2\ell}}\cdots\, e_{i_1}]) \;.$$ Since the supertrace of any bracket vanishes, one concludes that $[I_\ell\, , X ] = 0$. The other statement, $|I_\ell| = 0\,$, follows from $|\widetilde {e}_i | = |e_i|$, the additivity of the $\mathbb{Z}_2$-degree and the fact that $\mathrm{STr}\,(a) = 0$ for any odd element $a \in \mathsf{U}(\mathfrak{g})$. We now describe a useful property enjoyed by the Casimir elements $I_\ell$ of $\mathfrak{osp}(V \oplus V^\ast)$ in the special case of isomorphic components $V_0 \simeq V_1\,$. Recalling the notation of $\S$\[sect:osp-roots\], let $\partial := \sum_j X^{(0)}_{0j,1j}$ and $\widetilde{\partial} := - \sum_j X_{1j ,0j}^{(0)}\,$. These are odd elements of $\mathfrak{osp}\,$. (The reason for using the symbols $\partial, \widetilde{\partial}$ will become clear later). Notice that the bracket $\Lambda := [\partial , \widetilde{\partial}] = - \sum_{s,j} H_{s j}$ is in the Cartan algebra of $\mathfrak{osp}\,$. From $[\partial , \partial] = 2 \, \partial^2 = 0$ and the Jacobi identity one infers that $$[\partial,\Lambda] = [[\partial,\partial],\widetilde{\partial}] - [\partial , [\partial , \widetilde{\partial}]] = - [\partial , \Lambda] = 0 \;.$$ By the same argument, $[\widetilde{\partial},\Lambda] = 0\,$. One also sees that $\Lambda^2 = \mathrm{Id}\,$. Now define $F_\ell$ to be the following odd element of $\mathsf{U}(\mathfrak{osp})$: $$F_\ell = - \sum_{i_1,\,\ldots,\,i_{2\ell}=1}^d \widetilde{e}_{i_1} \cdots \, \widetilde{e}_{i_{2\ell}}\, \mathrm{STr}\,(e_{i_{2\ell}} \cdots \, e_{i_1} \widetilde{\partial} \Lambda) \;.$$ \[lem:Cas-exact\] Let $\mathfrak{osp}(V \oplus V^\ast)$ be the orthosymplectic Lie superalgebra for a $\mathbb{Z}_2$-graded vector space $V$ with isomorphic components $V_0 \simeq V_1\,$. Then for all $\ell \in \mathbb{N}$ the Casimir element $I_\ell$ is expressible as a bracket: $I_\ell = [\partial , F_\ell ]\,$. By the same argument as in the proof of Lemma \[lem:IlCasimir\], $$[\partial , F_\ell ] = - \sum \widetilde{e}_{i_1}\cdots \, \widetilde{e}_{i_{2\ell}}\,\mathrm{STr}\,([\partial ,e_{i_{2\ell}} \cdots e_{i_1}] \, \widetilde{\partial} \Lambda) \;.$$ Using the relations $[\partial ,\Lambda] = 0$ and $\Lambda^2 = \mathrm{Id}\,$, one has for any $a\in \mathsf{U} (\mathfrak{osp})$ that $$- \mathrm{STr}\, ([\partial, \, a]\, \widetilde{\partial} \Lambda) = \mathrm{STr}\, (\widetilde{\partial} \Lambda \, [\partial , \, a]) = \mathrm{STr}\, ([\widetilde{\partial} , \partial] \Lambda \, a) = \mathrm{STr}\,(\Lambda^2 a) = \mathrm{STr}\, (a) \;,$$ where the second equality sign is from $\mathrm{STr}\, (c,[b,a]) = \mathrm{STr}\, ([c,b]\, a)$. The statement of the lemma now follows on setting $a = e_{i_{2\ell}} \cdots\, e_{i_1}\,$. As we shall see, Lemma \[lem:Cas-exact\] has the drastic consequence that all $\mathfrak{osp}$-Casimir elements $I_\ell$ are zero in the spinor-oscillator representation of $\mathfrak{osp}(V \oplus V^\ast)$ for $V_0 \simeq V_1\,$. Howe pairs in $\mathfrak{osp}$ {#sect:howe-pairs} ------------------------------ In the present context, a pair $(\mathfrak{h},\mathfrak{h}')$ of subalgebras $\mathfrak{h}, \mathfrak{h}' \subset \mathfrak{g}$ of a Lie superalgebra $\mathfrak{g}$ is called a *dual pair* whenever $\mathfrak{h}'$ is the centralizer of $\mathfrak{h}$ in $\mathfrak{g}$ and vice versa. In this subsection, let $\mathbb{K} = \mathbb{C}\,$. Given a $\mathbb{Z}_2$-graded complex vector space $U = U_0 \oplus U_1$ we let $V := U \otimes \mathbb{C}^N$, where $\mathbb{C}^N$ is equipped with the standard representation of $\mathrm{GL} (\mathbb{C}^N)$, $\mathrm{O}(\mathbb{C}^N)\,$, or $\mathrm{Sp} (\mathbb{C}^N)\,$, as the case may be. As a result, the Lie algebra $\mathfrak{k}$ of whichever group is represented on $\mathbb{C}^N$ is embedded in $\mathfrak{osp}(V \oplus V^\ast)$. We will now describe the dual pairs $(\mathfrak{h},\mathfrak{k})$ in $\mathfrak{osp} (W)$ for $W = V \oplus V^*$. These are known as dual pairs in the sense of R. Howe. Let us begin by recalling that for any representation $\rho :\, K \to \mathrm{GL}(E)$ of a group $K$ on a vector space $E\,$, the dual representation $\rho^\ast : \, K \to \mathrm{GL}(E^\ast)$ on the linear forms on $E$ is given by $(\rho(k)\varphi)(x) = \varphi(\rho(k)^{-1}x)$. By this token, every representation $\rho :\, K \to \mathrm{GL}(\mathbb{C}^N)$ induces a representation $\rho_W = (\mathrm{Id} \otimes \rho) \times (\mathrm{Id} \otimes \rho^\ast)$ of $K$ on $W = V \oplus V^\ast$. Let $\rho: \, K \to \mathrm{GL}(\mathbb{C}^N)$ be any representation of a Lie group $K\,$. If $V = U \otimes \mathbb{C}^N$ for a $\mathbb{Z}_2$-graded complex vector space $U = U_0 \oplus U_1\,$, the induced representation ${\rho_W}_\ast (\mathfrak{k})$ of the Lie algebra $\mathfrak{k}$ of $K$ on $W = V \oplus V^\ast$ is a subalgebra of $\mathfrak{osp}(W)_0\,$. The $K$-action on $\mathbb{C}^N \otimes (\mathbb{C}^N)^\ast$ by $z \otimes \zeta \mapsto \rho(k)z \otimes \rho^\ast(k)\zeta$ preserves the canonical pairing $z \otimes \zeta \mapsto \zeta(z)$ between $\mathbb{C}^N$ and $(\mathbb{C}^N)^\ast$. Consequently, the $K$-action on $V \otimes V^\ast$ by $(\mathrm{Id} \otimes \rho) \otimes (\mathrm{Id} \otimes \rho^\ast)$ preserves the canonical pairing $V \otimes V^\ast \to \mathbb{C}\,$. Since the orthosymplectic form $Q : \, W \times W \to \mathbb{C}$ uses nothing but that pairing, it follows that $$Q( \rho_W(k) w \, , \, \rho_W(k) w^\prime) = Q(w,w^\prime) \quad (\text{for all } w,w^\prime \in W)\;.$$ Passing to the Lie algebra level one obtains ${\rho_W}_\ast( \mathfrak{k}) \subset \mathfrak{osp}(W)$. The operator $\rho_W(k)$ preserves the $\mathbb{Z}_2$-grading of $W$; therefore one actually has ${\rho_W}_\ast( \mathfrak{k}) \subset \mathfrak{osp}(W)_0\,$. Let us now assume that the complex Lie group $K$ is defined by a non-degenerate bilinear form $B :\, \mathbb{C}^N \times \mathbb{C}^N \to \mathbb{C}$ in the sense that $$K = \{ k \in \mathrm{GL}(\mathbb{C}^N) \mid \forall z\, , z^\prime \in \mathbb{C}^N : \, B(k z \, ,\, k z^\prime) = B(z\,,z^\prime)\}\;.$$ We then have a canonical isomorphism $\psi : \, \mathbb{C}^N \to (\mathbb{C}^N)^\ast$ by $z \mapsto B(z\, ,\,)$, and an isomorphism $\Psi : \, (U \oplus U^\ast) \otimes \mathbb{C}^N \to W$ by $(u + \varphi) \otimes z \mapsto u \otimes z + \varphi \otimes \psi(z)$. \[lem:2.11\] $\rho_W(k) = \Psi \circ (\mathrm{Id} \otimes k) \circ \Psi^{-1}$ for all $k \in K\,$. If $u \in U$, $\varphi \in U^\ast$, and $z \in \mathbb{C}^N$, then by the definition of $\Psi$ and $\rho_W(k)$, $$\rho_W(k) \Psi((u+\varphi) \otimes z) = u \otimes k z + \varphi \otimes \psi(z) k^{-1} \;.$$ Since $B$ is $K$-invariant, one has $\psi(z) k^{-1} = \psi(kz)$, and therefore $$u \otimes k z + \varphi \otimes \psi(z) k^{-1} = u \otimes k z + \varphi \otimes \psi(k z) = \Psi ((\mathrm{Id} \otimes k)((u+\varphi)\otimes z)) \;.$$ Thus $\rho_W(k) \circ \Psi = \Psi \circ (\mathrm{Id} \otimes k)$. Let us now examine what happens to the orthosymplectic form $Q$ on $W$ when it is pulled back by the isomorphism $\Psi$ to a bilinear form $\Psi^*Q$ on $(U\oplus U^*) \otimes \mathbb{C}^N:$ $$\Psi^*Q((u + \varphi) \otimes z\, , (u' + \varphi') \otimes z') = \varphi'(u)\, \psi(z')(z) - (-1)^{|u'||\varphi|} \varphi(u') \, \psi(z)(z') \;.$$ By definition, $\psi(z)(z^\prime) = B(z\, ,z^\prime)$, and writing $B(z\, , z^\prime) = (-1)^\delta B(z^\prime,z)$ where $\delta = 0$ if $B$ is symmetric and $\delta = 1$ if $B$ is alternating, we obtain $$\label{eq:pull-back} \Psi^*Q((u + \varphi) \otimes z \, , (u^\prime + \varphi^\prime) \otimes z^\prime) = (\varphi^\prime (u) - (-1)^{|u'||\varphi| + \delta} \varphi(u')) B(z',z) \;.$$ In view of this, let $\widetilde{U}$ denote the vector space $U = U_0 \oplus U_1$ with the twisted $\mathbb{Z}_2$-grading, i.e.$\widetilde{U}_s := U_{s+1}$ ($s \in \mathbb{Z}_2$). Moreover, notice that $\Psi$ determines an embedding $$\mathrm{End}(U \oplus U^\ast) \otimes \mathrm{End}(\mathbb{C}^N) \to \mathrm{End}(W) \;, \quad X \otimes k \mapsto \Psi \circ (X \otimes k) \circ \Psi^{-1} \;,$$ whose restriction to $\mathrm{End}(U \oplus U^*) \otimes \{ \mathrm{Id} \} \to \mathrm{End}(W)$ is an injective homomorphism. In the following we often write $\mathrm{O}(\mathbb{C}^N) \equiv \mathrm{O}_N$ and $\mathrm{Sp}(\mathbb{C}^N) \equiv \mathrm{Sp}_N$ for short. \[cor:2.2\] For $K = \mathrm{O}_N$ and $K = \mathrm{Sp}_N\,$, the map $X \mapsto \Psi \circ (X \otimes \mathrm{Id}) \circ \Psi^{-1}$ defines a Lie superalgebra embedding into $\mathfrak{osp}(W)$ of $\, \mathfrak{osp}(U \oplus U^*)$ resp. $\mathfrak{osp} (\widetilde{U} \oplus \widetilde{U}^\ast)$. For $K = \mathrm{O}_N$ the bilinear form $B$ of $\mathbb{C}^N$ is symmetric and the bilinear form $Q$ of $W$ pulls back – see (\[eq:pull-back\]) – to the standard orthosymplectic form of $U \oplus U^\ast$. For $K = \mathrm{Sp}_N\,$, the form $B$ is alternating. Its pullback, the orthosymplectic form of $U \oplus U^\ast$ twisted by the sign factor $(-1)^\delta$, is restored to standard form by switching to the $\mathbb{Z}_2$-graded vector space $\widetilde{U} \oplus \widetilde{U}^\ast$ with the twisted $\mathbb{Z}_2$-grading. To go further, we need a statement concerning $\mathrm{Hom}_G(V_1 ,V_2)$, the space of $G$-equivariant homomorphisms between two modules $V_1$ and $V_2$ for a group $G\,$. \[lem:HomG\] Let $X_1\, , X_2 \, , Y_1\, , Y_2$ be finite-dimensional vector spaces all of which are representation spaces for a group $G\,$. If the $G$-action on $X_1$ and $X_2$ is trivial, then $$\mathrm{Hom}_G(X_1 \otimes Y_1\, , X_2 \otimes Y_2) \simeq \mathrm{Hom}(X_1\, ,X_2) \otimes \mathrm{Hom}_G(Y_1\, ,Y_2) \;.$$ $\mathrm{Hom}(X_1 \otimes Y_1\, , X_2 \otimes Y_2)$ is canonically isomorphic to $X_1^\ast \otimes Y_1^\ast \otimes X_2 \otimes Y_2$ as a $G$-representation space, with $G$-equivariant maps corresponding to $G$-invariant tensors. Since the $G$-action on $X_1^\ast \otimes X_2^{\vphantom{\ast}}$ is trivial, one sees that $\mathrm{Hom}_G(X_1 \otimes Y_1\, , X_2 \otimes Y_2)$ is isomorphic to the tensor product of $X_1^\ast\otimes X_2^{\vphantom{\ast}} \simeq \mathrm{Hom}(X_1, X_2)$ with the space of $G$-invariants in $Y_1^\ast\otimes Y_2^{\vphantom{\ast}} \,$. The latter in turn is isomorphic to $\mathrm{Hom}_G(Y_1, Y_2)$. \[prop:dualpairs\] Writing $\mathfrak{g}_N \equiv \mathfrak{g}(\mathbb{C}^N)$ for $\mathfrak{g} = \mathfrak{gl}\,$, $\mathfrak{o}\,$, $\mathfrak{sp}\,$, the following pairs are dual pairs in $\mathfrak{osp}(W):$ $(\mathfrak{gl}(U),\mathfrak{gl}_N)$, $(\mathfrak{osp}(U\oplus U^*), \mathfrak{o}_N)$, $(\mathfrak{osp}(\widetilde{U}\oplus \widetilde{U}^*), \mathfrak{sp}_N)$. Here we calculate the centralizer of $\mathfrak{k}$ in $\mathfrak{osp}(W)$ for each of the three cases $\mathfrak{k} = \mathfrak{gl}_N\,$, $\mathfrak{o}_N\,$, $\mathfrak{sp}_N$ and refer the reader to [@H1] for the remaining details. Since both $V \subset W$ and $V^* \subset W$ are $K$-invariant subspaces, $\mathrm{End}_K(W)$ decomposes as $$\mathrm{End}_K(W) = \mathrm{End}_K(V) \oplus \mathrm{Hom}_K(V^*,V) \oplus \mathrm{Hom}_K(V,V^*) \oplus \mathrm{End}_K(V^*) \;.$$ By Schur’s lemma, $\mathrm{End}_K(\mathbb{C}^N) \simeq \mathbb{C}\,$, and therefore Lemma \[lem:HomG\] implies $$\mathrm{End}_K(V) = \mathrm{End}_K(U \otimes \mathbb{C}^N) \simeq \mathrm{End}(U) \otimes \mathrm{End}_K(\mathbb{C}^N) = \mathrm{End}(U)\;.$$ By the same reasoning, $\mathrm{End}_K(V^*) = \mathrm{End}(U^*)$. Applying Lemma \[lem:HomG\] to the two remaining summands, we obtain $$\mathrm{Hom}_K(V,V^*) \simeq \mathrm{Hom}(U,U^*) \otimes \mathrm{Hom}_K(\mathbb{C}^N,{\mathbb{C}^N}^*),$$ plus the same statement where each vector space is replaced by its dual. If $K = \mathrm{GL}(\mathbb{C}^N) \equiv \mathrm{GL}_N\,$, then $\mathrm{Hom}_K(\mathbb{C}^N ,{\mathbb{C}^N}^*) = \mathrm{Hom}_K ({\mathbb{C}^N}^* , \mathbb{C}^N) = \{ 0\}$. Hence, $$\Phi : \,\, \mathrm{End}(U) \oplus \mathrm{End}(U^\ast) \to \mathrm{End}_{\mathrm{GL}_N}(W)\;, \quad X \oplus Y \mapsto (X \otimes \mathrm{Id}) \times (Y \otimes \mathrm{Id})\;,$$ for $W = U \otimes \mathbb{C}^N \oplus U^\ast \otimes {\mathbb{C}^N}^\ast$ is an isomorphism. This means that the centralizer of $\mathfrak{gl}_N$ in $\mathfrak{osp}(W)$ is the intersection $\Phi(\mathrm{End}(U)\oplus\mathrm{End}(U^*)) \cap \mathfrak{osp}(W)$, which can be identified with $\mathrm{End}(U) = \mathfrak{gl}(U)$. Thus we have the first dual pair, $(\mathfrak{gl} (U), \mathfrak{gl}_N)$. In the case of $K = \mathrm{O}_N\,$, the discussion is shortened by recalling Lemma \[lem:2.11\] and the $K$-equivariant isomorphism $\Psi : \, (U \oplus U^\ast) \otimes \mathbb{C}^N \to W$. By Schur’s lemma, these imply $\mathrm{End}_K(W) \simeq \mathrm{End}(U \oplus U^\ast)$. From Corollary \[cor:2.2\] it then follows that the intersection $\mathfrak{osp}(W) \cap \mathrm{End}_K(W)$ is isomorphic as a Lie superalgebra to $\mathfrak{osp}(U \oplus U^\ast)$. Passing to the Lie algebra level for $K$, we get the second dual pair, $(\mathfrak{osp}(U \oplus U^\ast), \mathfrak{o}_N)$. Finally, if $K = \mathrm{Sp}_N\,$, the situation is identical except that Corollary \[cor:2.2\] compels us to switch to the $\mathbb{Z}_2$-twisted structure of orthosymplectic Lie superalgebra in $\mathrm{End}_K(W) \simeq \mathrm{End}(U \oplus U^\ast)$. This gives us the third dual pair, $(\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast), \mathfrak{sp}_N)$. Clifford-Weyl algebra $\mathfrak{q}(W)$ --------------------------------------- Let $\mathbb{K} = \mathbb{C}$ or $\mathbb{K} = \mathbb{R}$ (in this subsection the choice of number field again is immaterial) and recall from Example \[exa:JH\] the definition of the Jordan-Heisenberg Lie superalgebra $\widetilde{W} = W \oplus \mathbb{K}\,$, where $W = W_0 \oplus W_1$ is a $\mathbb{Z}_2$-graded vector space with components $W_1^{\vphantom{ \ast}} = V_1^{\vphantom{\ast}} \oplus V_1^\ast$ and $W_0^{\vphantom{ \ast}} = V_0^{\vphantom{\ast}} \oplus V_0^\ast$. The universal enveloping algebra of the Jordan-Heisenberg Lie superalgebra is called the *Clifford-Weyl algebra* (or quantum algebra). We denote it by $\mathfrak{q}(W) \equiv \mathsf{U}(\widetilde{W})$. Equivalently, one defines the Clifford-Weyl algebra $\mathfrak{q}(W)$ as the associative algebra generated by $\widetilde{W} = W \oplus \mathbb{K}$ subject to the following relations for all $w, w^\prime \in W_0 \cup W_1:$ $$w w' - (-1)^{|w||w'|} w' w = Q(w , w') \;.$$ In particular, $w_0\, w_1 = w_1\, w_0$ for all $w_0 \in W_0$ and $w_1 \in W_1\,$. Reordering by this commutation relation defines an isomorphism of associative algebras $\mathfrak{q}(W) \simeq \mathfrak{c}(W_1) \otimes \mathfrak{w}(W_0)$, where the Clifford algebra $\mathfrak{c}(W_1)$ is generated by $W_1 \oplus \mathbb{K}$ with the relations $w w^\prime + w^\prime w = S(w,w^\prime)$ for $w,w^\prime \in W_1\,$, and the Weyl algebra $\mathfrak{w}(W_0)$ is generated by $W_0 \oplus \mathbb{K}$ with the relations $w w^\prime - w^\prime w = A(w,w^\prime)$ for $w, w^\prime \in W_0\,$. As a universal enveloping algebra the Clifford-Weyl algebra $\mathfrak{q} (W)$ is filtered, $$\mathfrak{q}_0(W) := \mathbb{K} \subset \mathfrak{q}_1(W) := W \oplus \mathbb{K} \subset \ldots \subset \mathfrak{q}_n(W) \ldots \;,$$ and it inherits from the Jordan-Heisenberg algebra $\widetilde{W}$ a canonical $\mathbb{Z}_2$-grading and a canonical structure of Lie superalgebra by the supercommutator – see $\S$\[sect:UEA\] for the definitions. The next statement is a sharpened version of Lemma \[lem:2.2\]. \[lem:n+n’-2\] $[\mathfrak{q}_n(W) , \mathfrak{q}_{n^\prime}(W)] \subset \mathfrak{q}_{n + n^\prime - 2}(W)$. Lemma \[lem:2.2\] asserts the commutation relation $[\mathsf{U}_n (\mathfrak{g}), \mathsf{U}_{n^\prime} (\mathfrak{g}) ] \subset \mathsf{U}_{n + n^\prime-1} (\mathfrak{g})$ for the general case of a Lie superalgebra $\mathfrak{g}$ with bracket $[ \mathfrak{g}, \mathfrak{g}] \subset \mathfrak{g}\,$. For the specific case at hand, where the fundamental bracket $[W , W] \subset \mathbb{K}$ has zero component in $W$, the degree $n + n^\prime - 1$ is lowered to $n + n^\prime -2$ by the very argument proving that lemma. It now follows that each of the subspaces $\mathfrak{q}_n(W)$ for $n \le 2$ is a Lie superalgebra. Since $[\mathfrak{q}_2(W), \mathfrak{q}_1(W)] \subset \mathfrak{q}_1(W)$, the quotient space $\mathfrak{q}_2(W)/\mathfrak{q}_1(W)$ is also a Lie superalgebra. By the Poincaré-Birkhoff-Witt theorem, there exists a vector-space isomorphism $$\mathfrak{q}_2(W) / \mathfrak{q}_1(W) \stackrel{\sim}{\to} \mathfrak{s} \;,$$ sending $\mathfrak{q}_2(W) / \mathfrak{q}_1(W)$ to $\mathfrak{s}\,$, the space of super-symmetrized degree-two elements in $\mathfrak{q}_2 (W)$. Hence $\mathfrak{q}_2(W)$ has a direct-sum decomposition $\mathfrak{q}_2(W) = \mathfrak{q}_1(W) \oplus \mathfrak{s}\,$. If $\{ e_i \}$ is a homogeneous basis of $W$, every $a \in \mathfrak{s}$ is uniquely expressed as $$\label{eq:2.8} a = \sum\nolimits_{i,j} a_{ij}\, e_i \, e_j \;, \quad a_{ij} = (-1)^{|e_i| |e_j|} a_{ji} \;.$$ By adding and subtracting terms, $$2 w w^\prime = (w w^\prime + (-1)^{|w||w^\prime|} w^\prime w) + (w w^\prime - (-1)^{|w||w^\prime|} w^\prime w) \;,$$ one sees that the product $w w^\prime$ for $w,w^\prime \in W$ has scalar part $(w w^\prime)_\mathbb{K} = \frac{1}{2}[w,w^\prime] = \frac{1}{2}Q(w,w^\prime)$ with respect to the decomposition $\mathfrak{q}_2(W) = \mathbb{K} \oplus W \oplus \mathfrak{s}\,$. $[\mathfrak{s} , \mathfrak{s}] \subset \mathfrak{s}\,$. From the definition of $\mathfrak{s}$ and $[W,W] \subset \mathbb{K}$ it is clear that $[\mathfrak{s}, \mathfrak{s}] \subset \mathbb{K} \oplus \mathfrak{s}\,$. The statement to be proved, then, is that $[a,b]$ for $a,\, b \in \mathfrak{s}$ has zero scalar part. By the linearity of the supercommutator, it suffices to consider a single term of the sum (\[eq:2.8\]). Thus we put $a = w w^\prime + (-1)^{|w| |w^\prime|} w^\prime w$, and have $${\textstyle{\frac{1}{2}}} [a,b] = [w w^\prime,b] = w\, [w^\prime,b] + [w,b] \, w^\prime (-1)^{|w^\prime| |b|} \;.$$ Now we compute the scalar part of the right-hand side. Using the Jacobi identity for the Lie superalgebra $\mathfrak{q}(W)$ we obtain $$[a,b]_\mathbb{K} = [w,[w^\prime,b]] + [[w,b],w^\prime] (-1)^{|w^\prime| |b|} = [[w,w^\prime],b] \;.$$ The last expression vanishes because $[w,w^\prime] \subset \mathbb{K}$ lies in the center of $\mathfrak{q}(W)$. $\mathfrak{osp}(W)$ inside $\mathfrak{q}(W)$ -------------------------------------------- As a subspace of $\mathfrak{q}(W)$ which closes w.r.t. the supercommutator $[\, , \, ]$, $\mathfrak{s}$ is a Lie superalgebra. Now from Lemma \[lem:n+n’-2\] and the Jacobi identity for $\mathfrak{q}(W)$, one sees that $\mathfrak{s} \subset \mathfrak{q}_2(W)$ acts on each of the quotient spaces $\mathfrak{q}_n(W) / \mathfrak{q}_{n - 1}(W)$ for $n \ge 1$ by $a \mapsto [a,\,\,]\,$. In particular, $\mathfrak{s}$ acts on $\mathfrak{q}_1(W) / \mathfrak{q}_0(W) = W$ by $a \mapsto [a,w]$, which defines a homomorphism of Lie superalgebras $$\tau : \, \mathfrak{s} \to \mathfrak{gl}(W) \;, \quad a \mapsto \tau(a) = [a , \,\, ] \;.$$ The mapping $\tau$ is actually into $\mathfrak{osp}(W) \subset \mathfrak{gl}(W)$. Indeed, for $w,w^\prime \in W$ one has $$\begin{aligned} Q(\tau(a)w,w')+(-1)^{|\tau(a)||w|}Q(w,\tau(a)w') = [[a,w],w']+(-1)^{|a||w|}[w,[a,w']] \;,\end{aligned}$$ and since $[a,[w,w^\prime]] = 0\,$, this vanishes by the Jacobi identity. \[lem:tau-iso\] The map $\tau :\, \mathfrak{s} \to \mathfrak{osp}(W)$ is an isomorphism of Lie superalgebras. Being a homomorphism of Lie superalgebras, the linear mapping $\tau$ is an isomorphism of such algebras if it is bijective. We first show that $\tau$ is injective. So, let $a \in \mathfrak{s}$ be any element of the kernel of $\tau\,$. The equation $\tau (a) = 0$ means that $[[a,w],w^\prime] = [\tau(a)w,w^\prime]$ vanishes for all $w, w^\prime \in W$. To fathom the consequences of this, let $\{e_i\}$ and $\{ \widetilde{e}_i \}$ be two homogeneous bases of $W$ so that $Q(e_i\,,\widetilde{e}_j) = \delta_{ij} \,$. Using that $a \in \mathfrak{s}$ has a uniquely determined expansion $a = \sum a_{ij}\, e_i \, e_j$ with supersymmetric coefficients $a_{ij} = (-1)^{|e_i| |e_j|} a_{j\,i}\,$, one computes $$[ [ a , \widetilde{e}_j ] , \widetilde{e}_i ] = a_{ij} + (-1)^{|e_i| |e_j|} a_{j\,i} = 2 a_{ij} \;.$$ Thus the condition $[[a,w],w^\prime] = 0$ for all $w, w^\prime \in W$ implies $a = 0\,$. Hence $\tau$ is injective. By the Poincaré-Birkhoff-Witt isomorphism $$\mathfrak{s} \simeq \mathfrak{q}_2(W) / \mathfrak{q}_1(W) \simeq \sum\nolimits_{k+l = 2}\wedge^k (W_1) \otimes \mathrm{S}^l(W_0) \;,$$ the dimensions of the $\mathbb{Z}_2$-graded vector space $\mathfrak{s} = \mathfrak{s}_0 \oplus \mathfrak{s}_1$ are $$\dim\, \mathfrak{s}_0 = \dim\, \wedge^2 (W_1) + \dim \mathrm{S}^2 (W_0) \;, \quad \dim\, \mathfrak{s}_1 = \dim W_1 \, \dim W_0 \;.$$ These agree with those of $\mathfrak{osp}(W)$ as recorded in Corollary \[cor:dimosp\]. Hence our injective linear map $\tau : \, \mathfrak{s} \to \mathfrak{osp}(W)$ is in fact a bijection. By the isomorphism $\tau$ every representation $\rho$ of $\mathfrak{s} \subset \mathfrak{q}(W)$ induces a representation $\rho \circ \tau^{-1}$ of $\mathfrak{osp}(W)$. Let us conclude this subsection by writing down an explicit formula for $\tau^{-1}$. To do so, let $\{ e_i \}$ and $\{ \widetilde{e}_j \}$ be homogeneous bases of $W$ with $Q(e_i \, , \, \widetilde{e}_j) = \delta_{ij}$ as before. For $X \in \mathfrak{osp}(W)$ notice that the coefficients $a_{ij} := Q(e_i \, , X e_j) (-1)^{|e_j|}$ are supersymmetric: $$a_{ij} = (-1)^{|X||e_i| + 1 + |e_j|} Q(X e_i \, ,\, e_j) = (-1)^{|X||e_i|} Q(e_j \, , X e_i) = (-1)^{|e_i| |e_j|} a_{j\,i} \;,$$ where the last equality sign uses $(-1)^{|e_i| |e_j|}\, a_{j\,i} = (-1)^{|e_i| |X e_i|}\, a_{j\,i} = (-1)^{|e_i| |X| + |e_i|}\, a_{j\,i}\,$. The inverse map $\tau^{-1} : \, \mathfrak{osp}(W) \to \mathfrak{s}$ is now expressed as $$\label{eq:tau-inv} \tau^{-1}(X) = {\textstyle{\frac{1}{2}}} \sum\nolimits_{i,j} Q(e_i \, , X e_j) (-1)^{|e_j| + 1} \, \widetilde{e}_i \, \widetilde{e}_j \;.$$ To verify this formula, one calculates the double supercommutator $[ e_i\, , [\tau^{-1}(X) , e_j] ]$ and shows that the result is equal to $[e_i \, , X e_j] = Q(e_i\, , X e_j)$, which is precisely what is required from the definition of $\tau$ by $[\tau^{-1}(X),\, e_j] = X e_j\,$. Spinor-oscillator representation -------------------------------- As before, starting from a $\mathbb{Z}_2$-graded $\mathbb{K}$-vector space $V = V_0 \oplus V_1\,$, let the direct sum $W = V\oplus V^*$ be equipped with the orthosymplectic form $Q$ and denote by $\mathfrak{q}(W)$ the Clifford-Weyl algebra of $W$. Consider now the following tensor product of exterior and symmetric algebras: $$\mathfrak{a}(V) := \wedge(V_1^*) \otimes \mathrm{S}(V_0^*) \;.$$ Following R. Howe we call it the *spinor-oscillator module* of $\mathfrak{q}(W)$. Notice that $\mathfrak{a}(V)$ can be identified with the graded-commutative subalgebra in $\mathfrak{q}(W)$ which is generated by $V \oplus \mathbb{K}\,$. As such, $\mathfrak{a}(V)$ comes with a canonical $\mathbb{Z}_2$-grading and its space of endomorphisms carries a structure of Lie superalgebra, $\mathfrak{gl} (\mathfrak{a}(V)) \equiv \mathrm{End}(\mathfrak{a}(V))$. The algebra $\mathfrak{a}(V)$ now is to become a representation space for $\mathfrak{q}(W)$. Four operations are needed for this: the operator $\varepsilon(\varphi_1) : \, \wedge^k(V_1^\ast) \to \wedge^{k+1} (V_1^\ast)$ of exterior multiplication by a linear form $\varphi_1 \in V_1^\ast\,$; the operator $\iota(v_1) : \, \wedge^k( V_1^\ast) \to \wedge^{k-1}(V_1^\ast)$ of alternating contraction with a vector $v_1\in V_1\,$; the operator $\mu(\varphi_0):\, \mathrm{S}^l (V_0^\ast) \to \mathrm{S}^{l+1}(V_0^\ast)$ of multiplication with a linear function $\varphi_0 \in V_0^\ast\,$; and the operator $\delta(v_0) : \, \mathrm{S}^l (V_0^\ast) \to \mathrm{S}^{l-1} (V_0^\ast)$ of taking the directional derivative by a vector $v_0 \in V_0\,$. The operators $\varepsilon$ and $\iota$ obey the *canonical anti-commutation relations* (CAR), which is to say that $\varepsilon (\varphi)$ and $\varepsilon(\varphi^\prime)$ anti-commute, $\iota(v)$ and $\iota( v^\prime)$ do as well, and one has $$\iota(v) \varepsilon(\varphi) + \varepsilon(\varphi) \iota(v) = \varphi(v) \, \mathrm{Id}_{\wedge(V_1^\ast)} \;.$$ The operators $\mu$ and $\delta$ obey the *canonical commutation relations* (CCR), i.e., $\mu(\varphi)$ and $\mu(\varphi^\prime)$ commute, so do $\delta(v)$ and $\delta(v^\prime)$, and one has $$\delta(v) \mu(\varphi) - \mu(\varphi) \delta(v) = \varphi(v) \, \mathrm{Id}_{\mathrm{S}(V_0^\ast)} \;.$$ Given all these operations, one defines a linear mapping $q : \, W \to \mathrm{End}(\mathfrak{a}(V))$ by $$q(v_1 + \varphi_1 + v_0 + \varphi_0) = \iota(v_1) + \varepsilon(\varphi_1) + \delta(v_0) + \mu(\varphi_0) \quad (v_s \in V_s\, , \, \varphi_s \in V^*_s)\;,$$ with $\iota(v_1)$, $\varepsilon(\varphi_1)$ operating on the first factor of the tensor product $\wedge(V_1^\ast) \otimes \mathrm{S} (V_0^\ast)$, and $\delta(v_0)$, $\mu(\varphi_0)$ on the second factor. Of course the two sets $\varepsilon, \iota$ and $\mu, \delta$ commute with each other. In terms of $q\,$, the relations CAR and CCR are succinctly summarized as $$\label{eq:2.9} [q(w)\,,\,q(w^\prime)] = Q(w,w^\prime)\, \mathrm{Id}_{\mathfrak{a}(V)}\;,$$ where $[\, , \,]$ denotes the usual supercommutator of the Lie superalgebra $\mathfrak{gl}(\mathfrak{a}(V))$. By the relation (\[eq:2.9\]) the linear map $q$ extends to a representation of the Jordan-Heisenberg Lie superalgebra $\widetilde{W} = W \oplus \mathbb{K}\,$, with the constants of $\widetilde{W}$ acting as multiples of $\mathrm{Id}_{\mathfrak{a}(V)}\,$. Moreover, being a representation of $\widetilde{W}$, the map $q$ yields a representation of the universal enveloping algebra $\mathsf{U}( \widetilde{W}) \equiv \mathfrak{q}(W)$. This representation is referred to as the *spinor-oscillator representation* of $\mathfrak{q}(W)$. In the sequel we will be interested in the $\mathfrak{osp} (W)$-representation induced from it by the isomorphism $\tau^{-1}$. There is a natural $\mathbb{Z}$-grading $\mathfrak{a}(V) = \bigoplus_{m \ge 0} \, \mathfrak{a}^m(V)\,$, $$\mathfrak{a}^m(V) = {\textstyle\bigoplus\nolimits}_{k + l = m} \wedge^k(V_1^*)\otimes \mathrm{S}^l(V_0^\ast) \;.$$ Note that the operators $\varepsilon(\varphi_1)$ and $\mu(\varphi_0)$ increase the $\mathbb{Z}$-degree of $\mathfrak{a}(V)$ by one, while the operators $\iota(v_1)$ and $\delta(v_0)$ decrease it by one. Note also if $\Lambda = (- \mathrm{Id}_V) \oplus \mathrm{Id}_{V^\ast}$ is the $\mathfrak{osp}$-element introduced in $\S$\[sect:osp-cas\], then a direct computation using the formula (\[eq:tau-inv\]) shows that $\mathfrak{a}^m(V)$ is an eigenspace of the operator $(q \circ \tau^{-1})(\Lambda)$ with eigenvalue $m\,$. Thus $\Lambda \in \mathfrak{osp}$ is represented on the spinor-oscillator module $\mathfrak{a}(V)$ by the *degree*. ### Weight constraints {#subsubsec:WoHdP} We now specialize to the situation of $V = U \otimes \mathbb{C}^N$ with $U = U_0 \oplus U_1$ a $\mathbb{Z}_2$-graded vector space as in $\S$\[sect:howe-pairs\], and we require $U_0$ and $U_1$ to be isomorphic with dimension $\dim U_0 = \dim U_1 = n\,$. Recall that $$(\mathfrak{osp}(U\oplus U^\ast),\mathfrak{o}_N)\;, \quad (\mathfrak{osp} (\widetilde{U} \oplus \widetilde{U}^*),\mathfrak{sp}_N)\;,$$ are Howe dual pairs in $\mathfrak{osp}(W)$ which we denote by $(\mathfrak{g}, \mathfrak{k})$. There is a decomposition $$\begin{aligned} &&\mathfrak{g} = \mathfrak{g}^{(-2)} \oplus \mathfrak{g}^{(0)} \oplus \mathfrak{g}^{(2)}\;, \quad \mathfrak{g}^{(0)} = \mathfrak{g} \cap (\mathrm{End}(U)\oplus\mathrm{End}(U^*)) \;, \\ &&\mathfrak{g}^{(-2)} = \mathfrak{g} \cap \mathrm{Hom}(U^*,U)\;, \quad \mathfrak{g}^{(2)} = \mathfrak{g} \cap \mathrm{Hom}(U,U^*) \;,\end{aligned}$$ in both cases. The notation highlights the fact that the operators in $\mathfrak{g}^{(m)} \hookrightarrow \mathfrak{osp}(V \oplus V^\ast)$ change the degree of elements in $\mathfrak{a}(V)$ by $m\,$. Note that the Cartan subalgebra $\mathfrak{h}$ of diagonal operators in $\mathfrak{g}$ is contained in $\mathfrak{g}^{(0)}$ but $\mathfrak{h}\ne\mathfrak{g}^{(0)} $. Since the Lie algebra $\mathfrak{k}$ is defined on $\mathbb{C}^N$, the $\mathfrak{k}$-action on $\mathfrak{a}(V)$ preserves the degree. This action exponentiates to an action of the complex Lie group $K$ on $\mathfrak{a}(V)$. \[prop:Howeduality\] The subalgebra $\mathfrak{a}(V)^K$ of $K$-invariants in $\mathfrak{a}(V)$ is an irreducible module for $\mathfrak{g}\,$. The vacuum $1 \in \mathfrak{a}(V)^K$ is contained in it as a cyclic vector such that $$\mathfrak{g}^{(-2)}.1 = 0 \;, \quad \mathfrak{g}^{(0)}.1 = \langle 1 \rangle_{\mathbb{C}} \;, \quad \langle \mathfrak{g}^{(2)}.1 \rangle_{\mathbb{C}} = \mathfrak{a}(V)^K \;.$$ This is a restatement of Theorems 8 and 9 of [@H1]. In the case of $(\mathfrak{g}, \mathfrak{k}) = (\mathfrak{osp}(U \oplus U^\ast), \mathfrak{o}_N )$ it matters that $K = \mathrm{O}_N\,$, as the connected Lie group $K = \mathrm{SO}_N$ has invariants in $\mathfrak{a}(V)$ not contained in $\langle \mathfrak{g}^{(2)}.1 \rangle_\mathbb{C} \,$. Proposition \[prop:Howeduality\] has immediate consequences for the weights of the $\mathfrak{g}$-representation on $\mathfrak{a}(V)^K$. Using the notation of $\S$\[sect:osp-roots\], let $\{ H_{sj} \}$ be a standard basis of $\mathfrak{h}$ and $\{ \vartheta_{sj} \}$ the corresponding dual basis. We now write $\vartheta_{0j} =: \phi_j\,$ and $\vartheta_{1j} =: \mathrm{i}\psi_j$ ($j = 1, \ldots, n$). \[cor:weights\] The representations of $\mathfrak{osp}(U\oplus U^*)$ on $\mathfrak{a}(V)^{\mathrm{O}_N}$ and $\mathfrak{osp}(\widetilde{U} \oplus \widetilde {U}^*)$ on $\mathfrak{a}(V)^{ \mathrm{Sp}_N}$ each have highest weight $\lambda_N = \frac{N}{2}\sum_{j=1}^n (\mathrm{i}\psi_j - \phi_j)$. Every weight of these representations is of the form $\sum_{j=1}^n (\mathrm{i} m_j \psi_j - n_j \phi_j)$ with $-\frac{N}{2}\le m_j \le \frac{N}{2} \le n_j\,$. Recall from $\S$\[sect:howe-pairs\] the embedding of $\mathfrak{osp} (U\oplus U^*)$ and $\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^*)$ in $\mathfrak{osp}(W)$, and from Lemma \[lem:tau-iso\] the isomorphism $\tau^{-1} : \, \mathfrak{osp}(W) \to \mathfrak{s}$ where $\mathfrak{s}$ is the Lie superalgebra of supersymmetrized degree-two elements in $\mathfrak{q}(W)$. Specializing formula (\[eq:tau-inv\]) to the case of a Cartan algebra generator $H_{sj} \in \mathfrak{h} \subset \mathfrak{g}$ one gets $$\tau^{-1}(H_{s j}) = - \tfrac{1}{2} \sum\nolimits_{a = 1}^N ((f_{s,j}\otimes f_a)(e_{s,j} \otimes e_a) + (-1)^s (e_{s,j}\otimes e_a )(f_{s,j}\otimes f_a) ) \;,$$ where $\{e_a \}$ is a basis of $\mathbb{C}^N$ and $\{ f_a \}$ the dual basis of $(\mathbb{C}^N)^\ast$. Now let $\tau^{-1}(H_{sj}) \in \mathfrak{s}$ act by the corresponding operator, say $\hat{H}_{sj} := (q \circ \tau^{-1})(H_{sj})$, in the spinor-oscillator representation $q$ of $\mathfrak{s} \subset \mathfrak{q} (W)$. Application of that operator to the highest-weight vector $1 \in \mathbb{C} \equiv \wedge^0(V_1^\ast) \otimes \mathrm{S}^0 (V_0^\ast) \subset \mathfrak{a}(V)^K$ yields $$\begin{aligned} \hat{H}_{1 j}\, 1 &= \tfrac{1}{2}\sum\nolimits_a \iota(e_{1,j} \otimes e_a)\varepsilon(f_{1,j}\otimes f_a) 1 = \tfrac{N}{2}\;,\\ \hat{H}_{0 j}\, 1 &= - \tfrac{1}{2}\sum\nolimits_a \delta(e_{0,j} \otimes e_a) \mu(f_{0,j}\otimes f_a) 1 = - \tfrac{N}{2} \;.\end{aligned}$$ Altogether this means that $\hat{H}\, 1 = \lambda_N(H) 1$ where $\lambda_N(H)= \frac{N}{2} \sum_j (\mathrm{i}\psi_j(H) - \phi_j(H))$. From Lemma \[lem:rootsosp\] the roots $\alpha$ corresponding to root spaces $\mathfrak{g}_\alpha \subset \mathfrak{g}^{(2)}$ are of the form $$-\phi_j -\phi_{j^{\,\prime}} \;, \quad -\mathrm{i}\psi_j -\mathrm{i} \psi_{j^{\,\prime}}\;,\quad -\phi_j-\mathrm{i}\psi_{j^{\,\prime}}\;,$$ where the indices $j, j^{\,\prime}$ are subject to restrictions that depend on $\mathfrak{g}$ being $\mathfrak{osp} (U\oplus U^*)$ or $\mathfrak{osp}( \widetilde{U}\oplus \widetilde{U}^*)$. From $\mathfrak{a}(V)^K = \mathfrak{g}^{ (2)}.1$ one then has $m_j \le \frac{N}{2} \le n_j$ for every weight $\gamma = \sum (\mathrm{i} m_j \psi_j - n_j \phi_j)$ of the $\mathfrak{g}$-representation on $\mathfrak{a}(V)^K$. The restriction $m_j \ge \frac{N}{2} - N$ results from $\wedge( V_1^\ast) = \wedge(U_1^\ast \otimes (\mathbb{C}^N)^\ast)$ being isomorphic to $\otimes_{j=1}^n \wedge(\mathbb{C}^N)^\ast$ and the vanishing of $\wedge^k (\mathbb{C}^N)^\ast = 0$ for $k > N\,$. \[cor:degree\] For each of our two cases $\mathfrak{g} = \mathfrak{osp}(U \oplus U^\ast)$ and $\mathfrak{g} = \mathfrak{osp}(\widetilde{U}\oplus \widetilde{U}^\ast)$ the element $\Lambda = - \sum_{s,j} H_{s j} \subset \mathfrak{g}$ is represented on $\mathfrak{a}(V)^K$ by the degree operator. Since the $K$-action on $\mathfrak{a}(V)$ preserves the degree, the subalgebra $\mathfrak{a}(V)^K$ is still $\mathbb{Z}$-graded by the same degree. Summing the above expressions for $(q \circ \tau^{-1}) (H_{s j})$ over $s,j$ and using CAR and CCR to combine terms, we obtain $$(q \circ \tau^{-1})(\Lambda) = \sum_{j=1}^n \sum_{a=1}^N \big( \mu(f_{0,j}\otimes f_a) \delta(e_{0,j} \otimes e_a) + \varepsilon(f_{1,j}\otimes f_a) \iota(e_{1,j}\otimes e_a) \big) \;,$$ which is in fact the operator for the degree of the $\mathbb{Z}$-graded module $\mathfrak{a}(V)^K$. ### Positive and simple roots {#sect:simple-roots} We here record the systems of simple positive roots that we will use later (in $\S$\[unicity theorem\]). In the case of $\mathfrak{osp}(U \oplus U^\ast)$ this will be $$\phi_1 - \phi_2 \;, \ldots, \phi_{n-1} - \phi_n \;, \phi_n - \mathrm{i}\psi_1 \;, \mathrm{i}\psi_1 - \mathrm{i} \psi_2 \;, \ldots, \mathrm{i}\psi_{n-1} - \mathrm{i}\psi_n \;, \mathrm{i} \psi_{n-1} + \mathrm{i} \psi_n \;.$$ The corresponding system of positive roots for $\mathfrak{osp}(U \oplus U^\ast)$ is $$\phi_j \pm \phi_k \;, \, \mathrm{i}\psi_j \pm \mathrm{i}\psi_k \,\, (j < k) \;, \,\, 2\phi_j \,, \,\, \phi_j \pm \mathrm{i}\psi_k \,\, (j\, ,k =1, \ldots, n) \;.$$ In the case of $\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast)$ we choose the system of simple positive roots $$\phi_1 - \phi_2 \;, \ldots, \phi_{n-1} - \phi_n \;, \phi_n - \mathrm{i}\psi_1 \;, \mathrm{i}\psi_1 - \mathrm{i}\psi_2 \;, \ldots, \mathrm{i}\psi_{n-1} - \mathrm{i}\psi_n \;, 2\mathrm{i}\psi_n \;.$$ The corresponding positive root system then is $$\phi_j \pm \phi_k \;, \, \mathrm{i}\psi_j \pm \mathrm{i}\psi_k \,\, (j < k) \;, \,\, 2\mathrm{i}\psi_j \,, \,\, \phi_j \pm \mathrm{i}\psi_k \,\, (j\, ,k =1, \ldots, n) \;.$$ In both cases the roots $$\phi_j - \phi_k \;, \, \mathrm{i}\psi_j - \mathrm{i}\psi_k \,\, (j < k) \;, \quad \phi_j - \mathrm{i}\psi_k \,\, (j\, , k = 1, \ldots, n) \;,$$ form a system of positive roots for $\mathfrak{gl}(U) \simeq \mathfrak{g}^{(0)} \subset \mathfrak{osp}$. ### Unitary structure {#sect:2.6.3} We now equip the spinor-oscillator module $\mathfrak{a}(V)$ for $V = V_0 \oplus V_1$ with a unitary structure. The idea is to think of the algebra $\mathfrak{a}(V)$ as a subset of $\mathcal{O}(V_0\, , \, \wedge V_1^\ast)$, the holomorphic functions $V_0 \to \wedge(V_1^\ast)$. For such functions a Hermitian scalar product is defined via Berezin’s notion of superintegration as follows. For present purposes, it is imperative that $V$ be defined over $\mathbb{R}$, i.e., $V = V_\mathbb{R} \otimes \mathbb{C}$, and that $V$ be re-interpreted as a *real* vector space $V^\prime := V_\mathbb{R} \oplus J\, V_\mathbb{R}$ with complex structure $J \simeq \mathrm{i}\,$. Needless to say, this is done in a manner consistent with the $\mathbb{Z}_2$-grading, so that $V^\prime = V_0^\prime \oplus V_1^\prime$ and $V_s^\prime = V_{s,\mathbb{R}} \oplus J\, V_{s,\mathbb{R}} \simeq V_{s,\mathbb{R}} \otimes \mathbb{C} = V_s\,$. From $V_s = U_s \otimes \mathbb{C}^N$ and $U_1 \simeq U_0$ we are given an isomorphism $V_1 \simeq V_0\,$. This induces a canonical isomorphism $\wedge ({V_1^\prime}^\ast) \simeq \wedge( {V_0^\prime}^\ast)$, which gives rise to a bundle isomorphism $\Omega$ sending $\Gamma(V_0^\prime \, , \, \wedge {V_1^\prime}^\ast)$, the algebra of real-analytic functions on $V_0^\prime$ with values in $\wedge ({V_1^\prime}^\ast)$, to $\Gamma({V_0^\prime} \,,\, \wedge T^\ast V_0^\prime)$, the complex of real-analytic differential forms on $V_0^\prime\,$. Fixing some orientation of $V_0^\prime\,$, the Berezin (super-)integral for the $\mathbb{Z}_2$-graded vector space $V^\prime = V_0^\prime \oplus V_1^\prime$ is then defined as the composite map $$\Gamma(V_0^\prime \,,\, \wedge {V_1^\prime}^\ast) \stackrel{\Omega} {\longrightarrow} \Gamma(V_0^\prime \,,\,\wedge T^\ast V_0^\prime) \ \stackrel{\int}{\longrightarrow} \mathbb{C}\;,\quad \Phi\mapsto \Omega[\Phi]\mapsto \int_{V_0^\prime}\Omega[\Phi]\;,$$ whenever the integral over $V_0^\prime$ exists. Thus the Berezin integral is a two-step process: first the section $\Phi$ is converted into a differential form, then the form $\Omega[\Phi]$ is integrated in the usual sense to produce a complex number. Of course, by the rules of integration of differential forms only the top-degree component of $\Omega[\Phi]$ contributes to the integral. The subspace $V_\mathbb{R} \subset V^\prime$ has played no role so far, but now we use it to decompose the complexification $V^\prime \otimes \mathbb{C}$ into holomorphic and anti-holomorphic parts: $V^\prime \otimes \mathbb{C} = V \oplus \overline{V}$ and determine an operation of complex conjugation $V^\ast \to \overline{V^\ast}$. We also fix on $V = V_0 \oplus V_1$ a Hermitian scalar product (a.k.a. unitary structure) $\langle \, , \, \rangle$ so that $V_0 \perp V_1\,$. This scalar product determines a parity-preserving complex anti-linear bijection $c : \, V \to V^\ast$ by $v \mapsto cv = \langle v , \, \rangle$. Composing $c$ with complex conjugation $V^\ast \to \overline{V^\ast}$ we get a $\mathbb{C}$-linear isomorphism $V \to \overline{V^\ast}$, $v \mapsto \overline{cv}\,$. In this setting there is a distinguished Gaussian section $\gamma \in \Gamma(V_0^\prime \, , \wedge {V_1^\prime}^\ast \otimes \mathbb{C})$ singled out by the conditions $$\label{eq:def-gamma} \forall v_0 \in V_0\, , v_1 \in V_1 : \quad \delta(v_0) \gamma = - \mu(\overline{c v_0}) \gamma \;, \quad \iota(v_1) \gamma = - \varepsilon(\overline{c v_1}) \gamma \;.$$ To get a close-up view of $\gamma\,$, let $\{ e_{0,j }\}$ and $\{ e_{1,j} \}$ be orthonormal bases of $V_0$ resp. $V_1\,$, and let $z_j = c e_{0,j}$ and $\zeta_j = c e_{1,j}$ be the corresponding coordinate functions, with complex conjugates $\overline{z}_j$ and $\overline{\zeta}_j\,$. Viewing $\zeta_j\,$, $\overline{\zeta}_j$ as generators of $\wedge({V_1^\prime}^\ast \otimes \mathbb{C})$, our section $\gamma \in \Gamma(V_0^\prime \,,\, \wedge {V_1^\prime}^\ast \otimes \mathbb{C})$ is the standard Gaussian $$\gamma = \mathrm{const} \times \mathrm{e}^{-\sum_j (z_j \overline{z}_j + \zeta_j \overline{\zeta}_j )} \;.$$ We fix the normalization of $\gamma$ by the condition $\int_{V_0^\prime} \Omega[\gamma] = 1$. A unitary structure on the spinor-oscillator module $\mathfrak{a}(V)$ is now defined as follows. Let complex conjugation $V^\ast \to \overline{V^\ast}$ be extended to an algebra anti-homomorphism $\mathfrak{a}(V) \to \mathfrak{a}(\overline{V})$ by the convention $\overline{\Phi_1 \Phi_2} = \overline{\Phi}_2 \, \overline{\Phi}_1$ (without any sign factors). Then, if $\Phi_1\, , \Phi_2$ are any two elements of $\mathfrak{a}(V)$, we view them as holomorphic maps $V_0 \to \wedge (V_1^\ast)$, multiply $\overline{\Phi}_1$ with $\Phi_2$ to form $\overline{\Phi}_1 \Phi_2 \in \Gamma(V_0^\prime \,, \wedge {V_1^\prime}^\ast \otimes \mathbb{C})$, and define their Hermitian scalar product by $$\label{eq:HSP} \left\langle \Phi_1 \, , \Phi_2 \right\rangle_{\mathfrak{a}(V)} := \int_{V_0^\prime} \Omega[ \gamma \, \overline{\Phi}_1 \Phi_2 ]\;.$$ Let us mention in passing that (\[eq:HSP\]) coincides with the unitary structure of $\mathfrak{a}(V)$ used in the Hamiltonian formulation of quantum field theories and in the Fock space description of many-particle systems composed of fermions and bosons. The elements $$\label{eq:occ-no} \bigwedge\nolimits_j \zeta_j^{m_j} \otimes \prod\nolimits_j z_j^{n_j} / \sqrt{n_j \, !}$$ for $m_j \in \{ 0, 1 \}$ and $n_j \in \{ 0, 1, \ldots \}$ form an orthonormal set in $\mathfrak{a}(V)$, which in physics is called the occupation number basis of $\mathfrak{a}(V)$. \[lem:h.c.\] For all $v_0 \in V_0$ and $v_1 \in V_1$ the pairs of operators $\delta (v_0)$, $\mu(cv_0)$ and $\iota(v_1)$, $\varepsilon(cv_1)$ in $\mathrm{End} (\mathfrak{a}(V))$ obey the relations $$\delta(v_0)^\dagger = \mu(c v_0) \;, \quad \iota(v_1)^\dagger = \varepsilon(c v_1) \;,$$ i.e., they are mutual adjoints with respect to the unitary structure of $\mathfrak{a} (V)$. Let $v \in V_0\,$. Since $\overline{\Phi}_1 \in \mathfrak{a} (\overline{V})$ is anti-holomorphic, we have $\delta(v) \overline{\Phi}_1 = 0\,$. By the first defining property of $\gamma$ in (\[eq:def-gamma\]) and the fact that $\delta(v)$ is a derivation, $$\gamma\, \overline{\Phi}_1 \delta(v) \Phi_2 = \delta(v) \left( \gamma\, \overline{\Phi}_1 \Phi_2 \right) + \mu(\overline{cv}) \gamma\, \overline{\Phi}_1 \Phi_2 \;,$$ and passing to the Hermitian scalar product by the Berezin integral we obtain $$\left\langle \Phi_1 , \delta(v) \Phi_2 \right\rangle_{\mathfrak{a}(V)} = \int_{V_0} \Omega[\gamma\,\overline{\Phi}_1 \overline{\mu(cv)}\Phi_2] = \left\langle \mu(cv)\Phi_1,\Phi_2\right\rangle_{\mathfrak{a}(V)}\;.$$ By the definition of the $\dagger$-operation this means that $\delta(v)^\dagger = \mu(cv)$. In the case of $v \in V_1$ the argument is similar but for a few sign changes. Our starting relation changes to $$\gamma\, \overline{\Phi}_1 \iota(v) \Phi_2 = (-1)^{|\Phi_1|} \iota(v) \left( \gamma\, \overline{\Phi}_1 \Phi_2 \right) + (-1)^{|\Phi_1|} \varepsilon(\overline{cv})\gamma\, \overline{\Phi}_1 \Phi_2\;,$$ since the operator $\iota(v)$ is an *anti*-derivation. If $v \mapsto \tilde{v}$ denotes the isomorphism $V_1 \to V_0\,$, then $\Omega \circ \iota(v) = \iota(\tilde{v}) \circ \Omega$ and the first term on the right-hand side Berezin-integrates to zero because $\iota(\tilde{v})$ lowers the degree in $\wedge T^\ast V_0^\prime\,$. Therefore, $$\left\langle \Phi_1 , \iota(v) \Phi_2 \right\rangle_{\mathfrak{a}(V)} = \int_{V_0^\prime} \Omega[\gamma\,\overline{\Phi}_1 \overline{ \varepsilon(cv)} \Phi_2] = \left\langle \varepsilon(cv) \Phi_1 , \Phi_2 \right\rangle_{\mathfrak{a}(V)}\;,$$ which is the statement $\iota(v)^\dagger = \varepsilon(cv)$. By the Hermitian scalar product (\[eq:HSP\]) and the corresponding $L^2$-norm, the spinor-oscillator module $\mathfrak{a}(V)$ is completed to a Hilbert space, $\mathcal{A}_V\,$. A nice feature here is that, as an immediate consequence of the factors $1 / \sqrt{n_j \, !}$ in the orthonormal basis (\[eq:occ-no\]), the $L^2$-condition $\langle \Phi , \Phi \rangle_{\mathfrak{a}(V)} < \infty$ implies absolute convergence of the power series for $\Phi \in \mathcal{A}_V\,$. Hence $\mathcal{A}_V$ can be viewed as a subspace of $\mathcal{O}(V_0\, , \wedge V_1^\ast)$: $$\mathcal{A}_V = \{ \Phi \in \mathcal{O}(V_0 \, , \wedge V_1^\ast) \mid \langle \Phi , \Phi \rangle_{\mathfrak{a}(V)} < \infty \} \;.$$ In the important case of isomorphic components $V_0 \simeq V_1\,$, we may regard $\mathcal{A}_V$ as the Hilbert space of square-integrable holomorphic differential forms on $V_0\,$. Note that although $\delta(v)$ and $\mu(\varphi)$ do not exist as operators on the Hilbert space $\mathcal{A}_V\,$, they do extend to linear operators on $\mathcal{O}(V_0 \, , \wedge V_1^\ast)$ for all $v \in V_0$ and $\varphi \in V_0^\ast\,$. Real structures {#sect:2.7} --------------- In this subsection we define a real structure for the complex vector space $W = V\oplus V^\ast$ and describe, in particular, the resulting real forms of the ($\mathbb{Z}_2$-even components of the) Howe dual partners introduced above. Recalling the map $c:\, V \to V^\ast$, $v \mapsto \langle v, \, \rangle\,$, let $W_\mathbb{R} \simeq V$ be the vector space $$W_{\mathbb{R}} = \{ v+cv \mid v \in V \} \subset V \oplus V^* = W \;.$$ Note that $W_\mathbb{R}$ can be viewed as the fixed point set $W_\mathbb{R} = \mathrm{Fix}(C)$ of the involution $$C : \,\, W \to W \;, \quad v + \varphi \mapsto c^{-1}\varphi + cv \;.$$ By the orthogonality assumption, $W_\mathbb{R} = W_{0,\mathbb{R}} \oplus W_{1,\mathbb{R}}$ where $W_{s,\mathbb{R}} = W_s \cap W_\mathbb{R}\,$. The symmetric bilinear form $S$ on $W_1^{\vphantom{\ast}} = V_1^{\vphantom{\ast}} \oplus V_1^\ast$ restricts to a Euclidean structure $$S\, :\,\, W_{1,\mathbb{R}}\times W_{1,\mathbb{R}}\to \mathbb{R} \;, \quad (v + cv \, , v'+cv' ) \mapsto 2\, \mathfrak{Re}\langle v,v'\rangle \;,$$ whereas the alternating form $A$ on $W_0^{\vphantom{\ast}} = V_0^{\vphantom{\ast}} \oplus V_0^\ast$ induces a real-valued symplectic form $$\omega = \mathrm{i}A \, : \,\, W_{0,\mathbb{R}} \times W_{0,\mathbb{R}} \to \mathbb{R} \;, \quad (v+cv\,, v'+cv') \mapsto 2\, \mathfrak{Im}\langle v,v'\rangle \;.$$ Please be warned that, since $Q = S + A$ fails to be real-valued on $W_\mathbb{R}\,$, the intersection $\mathfrak{osp}(W) \cap \mathrm{End}(W_\mathbb{R})$ is *not* a real form of the complex Lie superalgebra $\mathfrak{osp}(W)$. The connected classical real Lie groups associated to the bilinear forms $S$ and $\omega$ are $$\begin{aligned} \mathrm{SO}(W_{1,\mathbb{R}}) & :=\{ g \in \mathrm{SL}(W_{1,\mathbb{R}}) \mid \forall w, w^\prime \in W_{1,\mathbb{R}} : \, S(gw,gw') = S(w,w')\} \;, \\ \mathrm{Sp}(W_{0,\mathbb{R}})& := \{ g\in\mathrm{GL}(W_{0,\mathbb{R}}) \mid \forall w,w^\prime \in W_{0,\mathbb{R}} : \, \omega(gw,gw') = \omega(w,w') \} \;.\end{aligned}$$ They have Lie algebras denoted by $\mathfrak{o}(W_{1,\mathbb{R}})$ and $\mathfrak{sp} (W_{0,\mathbb{R}})$. By construction we have $\mathfrak{osp}(W)_0 \cap \mathrm{End}(W_\mathbb{R}) \simeq \mathfrak{o}(W_{1,\mathbb{R}}) \oplus \mathfrak{sp} (W_{0,\mathbb{R}})$, and this in fact is a real form of the complex Lie algebra $\mathfrak{osp}(W)_0 \simeq \mathfrak{o} (W_1) \oplus \mathfrak{sp}(W_0)$. The elements of $\mathfrak{o}(W_{1,\mathbb{R}}) \oplus \mathfrak{sp}(W_{0,\mathbb{R}}) \subset \mathfrak{osp}(W)$ are mapped via $\tau^{-1}$ and the spinor-oscillator representation to anti-Hermitian operators in $\mathrm{End}(\mathfrak{a}(V))$. Let $X \in \mathfrak{o}(W_{1,\mathbb{R}}) \oplus \mathfrak{sp} (W_{0,\mathbb{R}})$. We know from Lemma \[lem:tau-iso\] that $\tau^{-1}(X)$ is a super-symmetrized element of degree two in the Clifford-Weyl algebra $\mathfrak{q}(W)$. To see the explicit form of such an element, recall the definition $\tau(a) w = [a , w]\,$. Since $Q = S + A\,$, and $A$ restricts to $\mathrm{i}\omega\,$, the fundamental bracket $[\, , \,] : \, W_\mathbb{R} \times W_\mathbb{R} \to \mathbb{C}$ given by $[w , w^\prime] = Q(w,w^\prime)$ is real-valued on $W_{1,\mathbb{R}}$ but imaginary-valued on $W_{0,\mathbb{R}}$. Therefore, $$\begin{aligned} \tau^{-1}(\mathfrak{o}(W_{1,\mathbb{R}})) &= \mathrm{span}_\mathbb{R} \{ w w^\prime - w^\prime w \} \quad (w, w^\prime \in W_{1,\mathbb{R}})\;, \\ \tau^{-1} (\mathfrak{sp}(W_{0,\mathbb{R}})) &= \mathrm{span}_\mathbb{R} \{\mathrm{i} w w^\prime + \mathrm{i} w^\prime w \} \quad (w, w^\prime \in W_{0,\mathbb{R}}) \;.\end{aligned}$$ The proposed statement $X^\dagger = -X$ now follows under the assumption that the spinor-oscillator representation maps every $w \in W_\mathbb{R}$ to a self-adjoint operator in $\mathrm{End}( \mathfrak{a}(V))$. But every element $w \in W_\mathbb{R}$ is of the form $v_1 + c v_1 + v_0 + c v_0$ and this maps to the operator $\iota(v_1) + \varepsilon(c v_1) + \delta(v_0) + \mu(c v_0)$, which is self-adjoint by Lemma \[lem:h.c.\]. Given the real structure $W_\mathbb{R}$ of $W$, we now ask how $\mathrm{End}( W_\mathbb{R})$ intersects with the Howe pairs $(\mathfrak{osp}(U \oplus U^\ast), \mathfrak{o}_N)$ and $(\mathfrak{osp}(\widetilde{U} \oplus \widetilde {U}^\ast), \mathfrak{sp}_N)$ embedded in $\mathfrak{osp}(W)$. By the observation that $Q$ restricted to $W_\mathbb{R}$ is not real-valued, $\mathfrak{osp}(U \oplus U^\ast) \cap \mathrm{End}( W_\mathbb{R})$ fails to be a real form of the complex Lie superalgebra $\mathfrak{osp}(U \oplus U^\ast)$, and the same goes for $\mathfrak{ osp} (\widetilde{U} \oplus \widetilde{U}^\ast)$. Nevertheless, it is still true that the even components of these intersections are real forms of the complex Lie algebras $\mathfrak{osp}(U\oplus U^\ast)_0$ and $\mathfrak{osp} (\widetilde{U}\oplus \widetilde{U}^\ast)_0\,$. The real forms of interest are best understood by expressing them via blocks with respect to the decomposition $W = V \oplus V^\ast$. Since $W_\mathbb{R} = \mathrm{Fix}(C)$, the complex linear endomorphisms of $W$ stabilizing $W_\mathbb{R}$ are given by $$\mathrm{End}(W_\mathbb{R}) \simeq \{ X \in \mathrm{End}(W) \mid X = C X C^{-1} \} \;.$$ Writing $X$ in block-decomposed form $$X = \mathsf{A} \oplus \mathsf{B} \oplus \mathsf{C} \oplus \mathsf{D} \equiv \begin{pmatrix} \mathsf{A} &\mathsf{B}\\ \mathsf{C} &\mathsf{D} \end{pmatrix} \;,$$ where $\mathsf{A} \in \mathrm{End}(V)$, $\mathsf{B} \in \mathrm{Hom}(V^\ast,V)$, $\mathsf{C} \in \mathrm{Hom}(V,V^\ast)$, and $\mathsf{D} \in \mathrm{End}(V^\ast)$, the condition $X = C X C^{-1}$ becomes $$\mathsf{C} = \overline{\mathsf{B}} \;, \quad \mathsf{D} = \overline{\mathsf{A}} \;.$$ The bar here is a short-hand notation for the complex anti-linear maps $$\begin{aligned} \mathrm{Hom}(V^\ast,V) \to \mathrm{Hom}(V,V^\ast)\;, \quad &\mathsf{B} \mapsto \overline{\mathsf{B}} :=c \mathsf{B} c \;,\\ \mathrm{End}(V) \to \mathrm{End}(V^\ast)\;, \quad &\mathsf{A} \mapsto \overline{\mathsf{A}} := c \mathsf{A} c^{-1} \;.\end{aligned}$$ When expressed with respect to compatible bases of $V$ and $V^\ast$, these maps are just the standard operation of taking the complex conjugate of the matrices of $\mathsf{A}$ and $\mathsf{B}$. Now, to get an understanding of the intersections $\mathfrak{o}_N \cap \mathrm{End}(W_\mathbb{R})$ and $\mathfrak{sp}_N \cap \mathrm{End}(W_\mathbb{R})$, recall the relation $\mathsf{D} = - \mathsf{A}^\mathrm{t}$ for $X \in \mathfrak{osp}(W)_0$ and the fact that the action of the complex Lie algebras $\mathfrak{o}_N = \mathfrak{o}(\mathbb{C}^N)$ and $\mathfrak{sp}_N = \mathfrak{sp} (\mathbb{C}^N)$ on $W$ stabilizes the decomposition $W = V \oplus V^\ast$, with the implication that $\mathsf{B} = \mathsf{C} = 0$ in both cases. Combining $\mathsf{D} = - \mathsf{A}^ \mathrm{t}$ with $\mathsf{D} = \overline{\mathsf{A}}$ one gets the anti-Hermitian property $\mathsf{A} = - \overline{\mathsf{A}}^ \mathrm{t}$, which means that $\mathfrak{o}_N \cap \mathrm{End}(W_\mathbb{R})$ and $\mathfrak{sp}_N \cap \mathrm{End}(W_\mathbb{R})$ are compact real forms of $\mathfrak{o}_N$ and $\mathfrak{sp}_N\,$. Turning to the Howe dual partners of $\mathfrak{o}_N$ and $\mathfrak{sp}_N\,$, recall from $\S$\[sect:howe-pairs\] the isomorphism $\psi : \, \mathbb{C}^N \to (\mathbb{C}^N)^\ast$ and arrange for it to be an isometry, $\overline{\psi^{-1}} = \psi^\mathrm{t}$, of the unitary structures of $\mathbb{C}^N$ and $(\mathbb{C}^N)^\ast$. Recall also the embedding of the two Lie superalgebras $\mathfrak{osp}(U \oplus U^\ast)$ and $\mathfrak{osp} (\widetilde{U} \oplus \widetilde{U}^\ast)$ into $\mathfrak{osp}(W) = \mathfrak{osp}(U \otimes \mathbb{C}^N \oplus U^\ast \otimes (\mathbb{C}^N)^\ast)$ by $$\begin{pmatrix} \mathsf{a} &\mathsf{b} \\ \mathsf{c} &\mathsf{d} \end{pmatrix} \mapsto \begin{pmatrix} \mathsf{a} \otimes \mathrm{Id} &\mathsf{b} \otimes \psi^{-1} \\ \mathsf{c} \otimes \psi &\mathsf{d} \otimes \mathrm{Id} \end{pmatrix} = \begin{pmatrix} \mathsf{A} &\mathsf{B} \\ \mathsf{C} &\mathsf{D} \end{pmatrix} \;.$$ Here the notation still means the same, i.e., $\mathsf{a} \in \mathrm{End}(U)$, $\mathsf{b} \in \mathrm{Hom}(U^\ast,U)$, and so on. Let a real structure $(U \oplus U^\ast)_\mathbb{R}$ of $U \oplus U^\ast$ be defined in the same way as the real structure $W_\mathbb{R} = (V \oplus V^\ast)_\mathbb{R}$ of $W = V \oplus V^\ast$. \[prop:2.4\] $\mathfrak{osp}(U \oplus U^\ast)_0 \cap \mathrm{End}(W_\mathbb{R}) \simeq \mathfrak{o}( (U_1^{\vphantom{\ast}} \oplus U_1^\ast)_\mathbb{R} ) \oplus \mathfrak{sp} ((U_0^{\vphantom{\ast}} \oplus U_0^\ast)_\mathbb{R})$. The intersection is computed by transferring the conditions $\mathsf{D} = \overline{\mathsf{A}}$ and $\mathsf{C} = \overline{ \mathsf{B}}$ to the level of $\mathfrak{osp}(U \oplus U^\ast)_0\,$. Of course $\mathsf{D} = \overline{\mathsf{A}}$ just reduces to the corresponding condition $\mathsf{d} = \overline{\mathsf{a}}\,$. Because the isometry $\psi : \, \mathbb{C}^N \to (\mathbb{C}^N)^\ast$ in the present case is symmetric one has $\overline{\psi^{-1}} = \psi^\mathrm{t} = + \psi$, so the condition $\mathsf{C} = \overline{\mathsf{B}}$ transfers to $\mathsf{c} = \overline{\mathsf{b}}\,$. For the same reason, the parity of the maps $\mathsf{b}, \mathsf{c}$ is identical to that of $\mathsf{B}, \mathsf{C}$, i.e., $\mathsf{b} \vert_{U_0^\ast \to U_0^{\vphantom{\ast}}}$ is symmetric, $\mathsf{b} \vert_{U_1^\ast \to U_1^{\vphantom{\ast}}}$ is skew, and similar for $\mathsf{c}\,$. Hence, computing the intersection $\mathfrak{osp}(U \oplus U^\ast)_0 \cap \mathrm{End}(W_\mathbb{R})$ amounts to the same as computing $\mathfrak{osp}(V \oplus V^\ast)_0 \cap \mathrm{End}(W_\mathbb{R})$, and the statement follows from our previous discussion of the latter case. In the case of the Howe pair $(\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast), \mathfrak{sp}_N)$ the isometry $\psi : \, \mathbb{C}^N \to (\mathbb{C}^N)^\ast$ is skew, so that $\overline{ \psi^{-1}} = \psi^\mathrm{t} = - \psi\,$. At the same time, the parity of $\mathsf{b}, \mathsf{c}$ is reversed as compared to $\mathsf{B}, \mathsf{C}$: now the map $\mathsf{b} \vert_{U_0^\ast \to U_0^{\vphantom{\ast}}}$ is skew and $\mathsf{b} \vert_{U_1^\ast \to U_1^{\vphantom{\ast}}}$ is symmetric (and similar for $\mathsf{c})$. Therefore, $$\mathfrak{osp}(\widetilde{U}\oplus\widetilde{U}^\ast)_0\cap \mathrm{End}(W_{1,\mathbb{R}})\simeq \{ \begin{pmatrix} \mathsf{a} &\mathsf{b}\\ - \overline{\mathsf{b}} &-\mathsf{a}^\mathrm{t} \end{pmatrix} \in \mathrm{End} (U_1^{\vphantom{\ast}} \oplus U_1^\ast) \mid \mathsf{a} = - \overline{\mathsf{a}}^\mathrm{t} \;, \, \mathsf{b} = + \mathsf{b}^\mathrm{t} \} \;,$$ which is a compact real form $\mathfrak{usp}(U_1^{\vphantom{\ast}} \oplus U_1^\ast)$ of $\mathfrak{sp}(U_1 \oplus U_1^\ast)$; and $$\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast)_0 \cap \mathrm{End}(W_{0,\mathbb{R}}) \simeq \{ \begin{pmatrix} \mathsf{a} &\mathsf{b}\\ - \overline{\mathsf{b}} &-\mathsf{a}^\mathrm{t} \end{pmatrix} \in \mathrm{End}(U_0^{\vphantom{\ast}} \oplus U_0^\ast) \mid \mathsf{a} = -\overline{\mathsf{a}}^\mathrm{t} \;, \, \mathsf{b} = -\mathsf{b}^\mathrm{t} \} \;,$$ which is a non-compact real form of $\mathfrak{o} (U_0^{\vphantom{\ast}} \oplus U_0^\ast)$ known as $\mathfrak{so}^\ast(U_0^{\vphantom{\ast}} \oplus U_0^\ast)$. Let us summarize this result. \[prop:2.5\] $\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast)_0 \cap \mathrm{End}(W_\mathbb{R}) \simeq \mathfrak{usp} (U_1^{\vphantom{\ast}} \oplus U_1^\ast) \oplus \mathfrak{so}^\ast(U_0^{\vphantom{\ast}} \oplus U_0^\ast)$. Semigroup representation {#integrating} ======================== As before, we identify the complex Lie superalgebra $\mathfrak{g} := \mathfrak{osp}(W)$ with the space of super-symmetrized degree-two elements in $\mathfrak{q}_2(W)$, so that $$\mathfrak{q}_2(W) = \mathfrak{g} \oplus \mathfrak{q}_1(W) \;, \quad \mathfrak{q}_1(W) = W \oplus \mathbb{C}\;.$$ The adjoint representation of $\mathfrak{g}$ on $\mathfrak{q}(W)$ restricts to the Lie algebra representation of $\mathfrak{g}_0 = \mathfrak{o}(W_1) \oplus \mathfrak{sp}(W_0)$ on $W = W_1\oplus W_0$ which is just the direct sum of the fundamental representations of $\mathfrak{o}(W_1)$ and $\mathfrak{sp}(W_0)$. These are integrated by the fundamental representations of the complex Lie groups $\mathrm{SO}(W_1)$ and $\mathrm{Sp}(W_0)$ respectively. Since the Clifford-Weyl algebra $\mathfrak{q}(W)$ is an associative algebra, one can ask if, given $x \in \mathfrak{g}_0 \subset \mathfrak{q}(W)$, the exponential series $\mathrm{e}^x$ makes sense. The existence of a one-parameter group $\mathrm{e}^{tx}$ for $x \in \mathfrak{g}_0$ would of course imply that $$\label{conjugation} \frac{d}{dt} \, \mathrm{e}^{\, tx} \, w \, \mathrm{e}^{-tx} \, \Big\vert_{t=0} = \mathrm{ad}(x)\, w \qquad (w \in W)\,.$$ Now $\mathfrak{q}(W) = \mathfrak{c}(W_1) \otimes \mathfrak{w}(W_0)$. Since the Clifford algebra $\mathfrak{c}(W_1)$ is finite-dimensional, the series $\mathrm{e}^x$ for $x \in \mathfrak{o}(W_1) \hookrightarrow \mathfrak{g}_0$ does make immediate sense. In this way one is able to exponentiate the Lie algebra $\mathfrak{o}(W_1)$ in $\mathfrak{c}(W_1)$. The associated complex Lie group, which is then embedded in $\mathfrak{c}(W_1)$, is $\mathrm{Spin}(W_1)$. This is a 2:1 cover of the complex orthogonal group $\mathrm{SO}(W_1)$. Its conjugation representation on $W_1$ as in (\[conjugation\]) realizes the covering map as a homomorphism $\mathrm{Spin}(W_1) \to \mathrm{SO}(W_1)$. Viewing the other summand $\mathfrak{sp}(W_0)$ of $\mathfrak{g}_0$ as being in the infinite-dimensional Weyl algebra $\mathfrak{w}(W_0)$, it is definitely not possibly to exponentiate it in such a naive way. This is in particular due to the fact that for most $x \in \mathfrak{sp}(W_0)$ the formal series $\mathrm{e}^x$ is not contained in any space $\mathfrak{w}_n (W_0)$ of the filtration of $\mathfrak{w}(W_0)$. As a first step toward remedying this situation, we consider $\mathfrak{q}(W)$ as a space of densely defined operators on the completion $\mathcal{A}_V$ (cf. $\S$\[sect:2.6.3\]) of the spinor-oscillator module $\mathfrak{a}(V)$. Since all difficulties are on the $W_0$ side, for the remainder of this chapter we simplify the notation by letting $W := W_0$ and discussing only the oscillator representation of $\mathfrak{w}(W)$. Recall that this representation on $\mathfrak{a}(V)$ is defined by multiplication $\mu( \varphi)$ for $\varphi \in V^*$ and the directional derivative $\delta(v)$ for $v \in V$. For $x \in \mathfrak{w}(W)$ there is at least no formal obstruction to the exponential series of $x$ existing in $\mathrm{End}(\mathcal A_V)$. However, direct inspection shows that convergence cannot be expected unless some restrictions are imposed on $x\,$. This is done by introducing a notion of unitarity and an associated semigroup of contraction operators. The oscillator semigroup ------------------------ Here we introduce the basic semigroup in the complex symplectic group. Various structures are lifted to its canonical 2:1 covering. Actions of the real symplectic and metaplectic groups are discussed along with the role played by the cone of elliptic elements. ### Contraction semigroup: definitions, basic properties Letting $\langle \, , \, \rangle$ be the unitary structure on $V$ which was fixed in the previous chapter, we recall the complex anti-linear bijection $c : \, V \to V^\ast$, $v \mapsto \langle v \, , \, \rangle\,$. There is an induced map $C : \, W \to W$ on $W = V \oplus V^\ast$ by $C(v + \varphi) = c^{-1}\varphi + cv\,$. As before, we put $W_\mathbb{R} := \mathrm{Fix}(C) \subset W$. Since we have restricted our attention to the symplectic side, the vector spaces $W$ and $W_\mathbb{R}$ are now equipped with the standard complex symplectic structure $A$ and real symplectic form $\omega = \mathrm{i}A$ respectively. From here on in this chapter we abbreviate the notation by writing $\mathrm{Sp} := \mathrm{Sp}(W)$ and $\mathfrak{sp} := \mathfrak{sp} (W)$. Let an anti-unitary involution $\sigma : \, \mathrm{Sp} \to \mathrm{Sp}$ be defined by $g \mapsto C g \, C^{-1}$. Its fixed point group $\mathrm{Fix} (\sigma)$ is the real form $\mathrm{Sp} (W_\mathbb{R}$) of main interest. We here denote it by $\mathrm{Sp}_\mathbb{R}$ and let $\mathfrak{sp}_\mathbb{R}$ stand for its Lie algebra. Given $A$ and $C$, consider the mixed-signature Hermitian structure $$W \times W \to \mathbb{C} \;, \quad (w\, ,w^\prime) \mapsto A(Cw\, ,w^\prime) \;,$$ which we denote by $A(Cw\, ,w^\prime) =: \langle w \, , w^\prime \rangle_s\,$, with subscript $s$ to distinguish it from the canonical Hermitian structure of $W$ given by $\langle v + \varphi , v^\prime + \varphi^\prime \rangle := \langle v \, , v^\prime \rangle + \langle c^{-1} \varphi^\prime , c^{-1} \varphi \rangle$. The relation between the two is $$\langle w \, , w^\prime \rangle_s = \langle w \, , s w^\prime \rangle \;, \qquad s = (- \mathrm{Id}_V) \oplus \mathrm{Id}_{V^*} \;.$$ Note also the relation $$\sigma(g) = C g\, C^{-1} = s\, (g^{-1})^\dagger s \qquad (g \in \mathrm{Sp}) \;.$$ Now observe that the real form $\mathrm{Sp}_\mathbb{R}$ is the subgroup of $\langle \, , \, \rangle_s$-isometries in $\mathrm{Sp}:$ $$\mathrm{Sp}_\mathbb{R} = \{g \in \mathrm{Sp} \mid \forall w \in W : \; \langle g w , g w \rangle_s = \langle w , w \rangle_s \} \;.$$ Then define a semigroup $\mathrm{H}(W^s)$ in $\mathrm{Sp}$ by $$\mathrm{H}(W^s) := \{ h \in \mathrm{Sp} \mid \forall w \in W ,\; w \not= 0 : \; \langle h w , h w \rangle_s < \langle w , w \rangle_s \} \,.$$ Note that the operation $g \mapsto g^\dagger$ of Hermitian conjugation with respect to $\langle \, ,\, \rangle $ stabilizes $\mathrm{Sp}$ and that $\mathrm{Sp}_\mathbb{R}$ is defined by the condition $g^\dagger s g = s\,$. The semigroup $\mathrm{H}(W^s)$ is defined by $h^\dagger s h < s\,$, or equivalently, $s - h^\dagger s h$ is positive definite. We will see later that $\mathrm{H}(W^s)$ (or, rather, a $2:1$ cover thereof) acts by contraction operators on the Hilbert space $\mathcal{A}_V$. It is immediate that $\mathrm{H}(W^s)$ is an open semigroup in $\mathrm{Sp}$ with $\mathrm{Sp}_\mathbb{R}$ on its boundary. Furthermore, $\mathrm{H}(W^s)$ is stabilized by the action of $\mathrm{Sp}_\mathbb{R} \times \mathrm{Sp}_\mathbb{R}$ by $h\mapsto g_1 h \, g_2^{-1}$. The map $\pi :\, \mathrm{Sp} \to \mathrm{Sp}\,$, $h \mapsto h \sigma(h^{-1})$, will play an important role in our considerations. It is invariant under the $\mathrm{Sp}_\mathbb{R}$-action by right multiplication, $\pi(hg^{-1}) = \pi(h)$, and is equivariant with respect to the action defined by left multiplication on its domain of definition and conjugation on its image space, $\pi(gh) = g \pi(h) g^{-1}$. Direct calculation shows that in fact the $\pi$-fibers are exactly the orbits of the $\mathrm{Sp}_\mathbb{R}$-action by right multiplication. Observe that if $h = \exp(\mathrm{i}X)$ for $X \in \mathfrak{sp}_\mathbb{R}\,$, then $\sigma(h) = h^{-1}$ and $\pi(h) = h^2$. In particular, if $\mathfrak{t}$ is a Cartan subalgebra of $\mathfrak{sp}$ which is defined over $\mathbb{R}\,$, then $\pi \vert_{\exp(\mathrm{i} \mathfrak{t}_\mathbb{R})}$ is just the squaring map $t \mapsto t^2$. ### Actions of $\mathrm{Sp}_\mathbb{R}\,$ We now fix a Cartan subalgebra $\mathfrak{t}$ having the property that $T_\mathbb{R} = \exp (\mathfrak{t}_\mathbb{R})$ is contained in the unitary maximal compact subgroup defined by $\langle \, ,\, \rangle$ of $\mathrm{Sp} (W_\mathbb{R})$. This means that $T$ acts diagonally on the decomposition $W = V \oplus V^*$ and there is a (unique up to order) orthogonal decomposition $V = E_1 \oplus \ldots \oplus E_d$ into one-dimensional subspaces so that if $F_j := c(E_j)$, then $T$ acts via characters $\chi_j$ on the vector spaces $P_j := E_j \oplus F_j$ by $t(e_j\, ,f_j) = (\chi_j(t) e_j \, , \chi_j(t)^{-1} f_j)$. In other words, we may choose $\{ e_j \}_{j = 1, \ldots, d}$ to be an orthonormal basis of $V$ and equip $V^*$ with the dual basis so that the elements $t \in T$ are in diagonal form: $$t = \mathrm{diag}(\lambda_1^{\vphantom{-1}},\ldots,\lambda_d^{ \vphantom{-1}}\, ,\lambda_1^{-1}, \ldots , \lambda_d^{-1})\;, \ \qquad \lambda_j = \chi_j(t) \;.$$ Observe that, conversely, the elements of $\mathrm{Sp}$ that stabilize the decomposition $W = P_1 \oplus \ldots \oplus P_d$ and act diagonally in the above sense, are exactly the elements of $T$. Moreover, $\exp (\mathrm{i} \mathfrak{t}_\mathbb{R})$ is the subgroup of elements $t \in T$ with $\chi_j(t) \in \mathbb{R}_+$ for all $j\,$. Note that the complex symplectic planes $P_j$ are $A$-orthogonal and defined over $\mathbb{R}$. We now wish to analyze $\mathrm{H}(W^s)$ via the map $\pi :\, h \mapsto h \sigma(h^{-1})$. However, for a technical reason related to the proof of Proposition \[M-slice\] below, we must begin with the opposite map, $\pi^\prime : \, h \mapsto \sigma(h^{-1}) h\,$. Thus let $M := \pi^\prime (\mathrm{H}(W^s))$ and write $\pi^\prime:\, \mathrm{H} (W^s)\to M\,$. The toral semigroup $T_+ := \exp(\mathrm{i} \mathfrak{t}_\mathbb{R})\,\cap \, \mathrm{H}(W^s)$ consists of those elements $t \in \exp( \mathrm{i} \mathfrak{t}_\mathbb{R})$ that act as contractions on $V^\ast$, i.e., $0 < \chi_j(t)^{-1} < 1$ for all $j\,$. The restriction $\pi^\prime \vert_{T_+} = \pi \vert_{T_+}$ is, as indicated above, the squaring map $t \mapsto t^2$; in particular we have $T_+ \subset M$ and the set $\{ g\, t g^{-1} \mid t \in T_+ \, , \, g \in \mathrm{Sp}_\mathbb{R} \}$ is likewise contained in $M\,$. In the sequel, we will often encounter the action of $\mathrm{Sp}_\mathbb{R}$ on $T_+$ and $M$ by conjugation. We therefore denote this action by a special name, $\mathrm{Int}(g)\, t := g\, t g^{-1}$. \[M-slice\] $M = \mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+ \,$. For $g \in \mathrm{Sp}$ one has $\sigma(g^{-1}) = C g^{-1} C^{-1} = s g^\dagger s\,$. Hence if $M \ni m = \sigma(h^{-1})h$ with $h \in \mathrm{H}(W^s)$, then $m = s h^\dagger s h\,$. Consequently, $\langle w, m w \rangle_s = \langle h w , h w \rangle_s < \langle w , w \rangle_s$ for all $w \in W \setminus \{ 0 \}$. In particular, $\langle w , m w \rangle_s\in \mathbb{R}$ and if $w \not= 0$ is an $m$-eigenvector with eigenvalue $\lambda$, it follows that $\lambda \in \mathbb{R}$ and $\lambda \langle w , w \rangle_s < \langle w , w \rangle_s \not = 0\,$. Now we have $C h C^{-1} = \sigma(h)$ and hence $C m C^{-1} = m^{-1}$. As a result, if $w \not= 0$ is an $m$-eigenvector with eigenvalue $\lambda$, then so is $C w$ with eigenvalue $\lambda^{-1}$. Since $C s C^{-1} = - s$, the product of $\langle w , w \rangle_s$ with $\langle Cw , Cw \rangle_s$ is always negative. If $\langle w , w \rangle_s > 0$ it follows that $\lambda < 1$ and $\lambda^{-1} > 1$; if $\langle Cw , Cw \rangle_s > 0$ then $\lambda^{-1} < 1$ and $\lambda > 1$. In both cases $0 < \lambda \not= 1$. Since $m$ does indeed have at least one eigenvector, we have constructed a complex 2-plane $Q_1$ as the span of the linearly independent vectors $w$ and $Cw$. The plane $Q_1$ is defined over $\mathbb{R}$ and, because $0 \not= \langle w , w \rangle_s = A(Cw,w)$, it is $A$-nondegenerate. Its $A$-orthogonal complement $Q_1^\perp$ is therefore also nondegenerate and defined over $\mathbb{R}$. The transformation $m \in \mathrm{Sp}$ stabilizes the decomposition $W = Q_1 \oplus Q_1^\perp\,$. Hence, proceeding by induction we obtain an $A$-orthogonal decomposition $W = Q_1 \oplus \ldots \oplus Q_d\,$. Since the $Q_j$ are $m$-invariant symplectic planes defined over $\mathbb{R}$, there exists $g \in \mathrm{Sp}_\mathbb{R}$ so that $t := g m g^{-1}$ stabilizes the above $T$-invariant decomposition $W = P_1 \oplus \ldots \oplus P_d\,$. Exchanging $w$ with $C w$ if necessary, we may assume that $t$ acts diagonally on $P_j = E_j \oplus F_j$ by $(e_j\, ,f_j) \mapsto (\lambda_j \, e_j \, , \lambda_j^{-1} f_j)$ with $\lambda_j > 1$. In other words, $t \in T_+\,$. If we let $$\mathrm{Sp}_\mathbb{R} T_+ \mathrm{Sp}_\mathbb{R} := \{g_1 t g_2^{-1} \mid g_1 , g_2 \in \mathrm{Sp}_\mathbb{R} \,, \, t \in T_+ \}\;,$$ then we now have the following analog of the $KAK$-decomposition. \[left/right transitive\] The semigroup $\mathrm{H}(W^s)$ decomposes as $\mathrm{H}(W^s) = \mathrm{Sp}_\mathbb{R} T_+ \mathrm{Sp}_\mathbb{R}\,$. In particular, $\mathrm{H}(W^s)$ is connected. By definition, $\mathrm{H}(W^s) = {\pi^\prime}^{-1}(M)$, and from Proposition \[M-slice\] one has $\mathrm{H}(W^s) = {\pi^\prime}^{-1}(\mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+)$. Now the map $\pi^\prime \vert_{T_+} : \, T_+ \to T_+ \,$, $t\mapsto t^2$ is surjective. Therefore ${\pi^\prime}^{-1}(T_+) = \mathrm{Sp}_\mathbb{R} T_+\,$, which is to say that each point in the fiber of $\pi^\prime$ over $t \in T_+$ lies in the orbit of $\sqrt{t} \in T_+$ generated by left multiplication with $\mathrm{Sp}_\mathbb{R}\,$. On the other hand, by the property $g \pi^\prime(h) g^{-1}= \pi^\prime(hg^{-1})$ of $\mathrm{Sp}_\mathbb{R}$-equivariance we have $$\mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+ = \mathrm{Int} (\mathrm{Sp}_\mathbb{R}) \pi^\prime ( {\pi^\prime}^{-1}(T_+) ) = \pi^\prime ( {\pi^\prime}^{-1}(T_+)\mathrm{Sp}_\mathbb{R} )$$ and hence $\mathrm{H}(W^s) = {\pi^\prime}^{-1}( \mathrm{Int} (\mathrm{Sp}_\mathbb{R}) T_+) = {\pi^\prime}^{-1} (T_+) \mathrm {Sp}_\mathbb{R} = \mathrm{Sp}_\mathbb{R} T_+ \mathrm{Sp}_\mathbb{R} \,$. Because $\mathrm{Sp}_\mathbb{R}$ and $T_+$ are connected, so is $\mathrm{H}(W^s) = \mathrm{Sp}_\mathbb{R} T_+ \mathrm{Sp}_\mathbb{R} \,$. It is clear that $M = \mathrm{Int}( \mathrm{Sp}_\mathbb{R} ) T_+ \subset \mathrm{H}(W^s)$. Furthermore, since both $T_+$ and $\mathrm{Sp}_\mathbb{R}$ are invariant under the operation of Hermitian conjugation $h\mapsto h^\dagger$ and under the involution $h \mapsto s h s$, we have the following consequences. \[cor:3.2\] $\mathrm{H}(W^s)$ is invariant under $h \mapsto h^\dagger$ and also under $h \mapsto s h s$. In particular, $\mathrm{H}(W^s)$ is stabilized by the map $h \mapsto \sigma(h^{-1}) = s h^\dagger s$. Letting $h^\prime := \sigma(h)^{-1}$ one has $\pi^\prime(h) = \sigma (h)^{-1} h = h^\prime \sigma(h^\prime)^{-1} = \pi(h^\prime)$ and hence $M = \pi^\prime(\mathrm{H}(W^s)) = \pi(\mathrm{H}(W^s))$. The stability of $\mathrm{H}(W^s)$ under $h \mapsto \sigma(h)^{-1}$ was not immediate from our definition of $\mathrm{H}(W^s)$, which is why we have been working from the viewpoint of $\mathrm{H}(W^s) = {\pi^\prime}^{-1}(M)$ so far. Now that we have it, we may regard $\mathrm{H}(W^s)$ as the total space of an $\mathrm{Sp}_\mathbb{R}$-principal bundle $\pi:\, \mathrm{H}(W^s) \to M$. We are going to see in Corollary \[bijective\] that this principal bundle is trivial. Next observe that, since $\sigma(m) = m^{-1}$ for $m = \sigma(h)^{-1} h = h^\prime \sigma(h^\prime)^{-1} \in M\,$, the maps $\pi : \, M \to M$ and $\pi^\prime : \, M \to M$ coincide and are just the square $m \mapsto m^2$. Thus the claim that the elements of $M$ have a unique square root in $M$ can be formulated as follows. \[squaring\] The squaring map $\pi = \pi^\prime : \, M \to M$ is bijective. Recall from Proposition \[M-slice\] that every $m \in M$ is diagonalizable in the sense that $M = \mathrm{Int}( \mathrm{Sp}_\mathbb{R}) T_+\,$. Since $\pi : \, T_+\to T_+$ is surjective, the surjectivity of $\pi : \, M \to M$ is immediate. For the injectivity of $\pi$ we note that the $m$-eigenspace with eigenvalue $\lambda$ is contained in the $m^2$-eigenspace with eigenvalue $\lambda^2$. The result then follows from the fact that all eigenvalues of $m$ are positive real numbers. \[bijective\] Each of the two $\mathrm{Sp}_\mathbb{R}$-equivariant maps $\mathrm{Sp}_\mathbb{R} \times M \to \mathrm{H} (W^s)$ defined by $(g,m) \mapsto gm$ and by $(g,m) \mapsto m g^{-1}$, is a bijection. Consider the map $(g,m) \mapsto gm\,$. Surjectivity is evident from Corollary \[left/right transitive\] and $M = \mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+\,$. For the injectivity it suffices to prove that if $m_1 , m_2 \in M$ and $g \in \mathrm{Sp}_\mathbb{R}$ with $g m_1 = m_2\,$, then $m_1 = m_2\,$. But this follows directly from $\pi^\prime(m_1) = \pi^\prime(g m_1) = \pi^\prime (m_2)$ and the fact that $\pi^\prime\vert_M$ is the bijective squaring map. The proof for the map $(g,m) \mapsto m g^{-1}$ is similar, with $\pi$ replacing $\pi^\prime$. ### Cone realization of $M$ Let us look more carefully at $M$ as a geometric object. First, as we have seen, the elements $m$ of $M$ satisfy the condition $m = \sigma (m^{-1})$. We regard $\psi :\, \mathrm{H}(W^s) \to \mathrm{H}(W^s) \,$, $h \mapsto \sigma(h^{-1})$, as an anti-holomorphic involution and reformulate this condition as $M \subset \mathrm{Fix} (\psi)$. In the present section we are going to show that $M$ is a closed, connected, real-analytic submanifold of $\mathrm{H}(W^s)$ which locally agrees with $\mathrm{Fix} (\psi)$. This implies in particular that $M$ is totally real in $\mathrm{H}(W^s)$ with $\dim_\mathbb{R} M = \dim_\mathbb{C} \mathrm{H}(W^s)$. We will also show that the exponential map identifies $M$ with a precisely defined open cone in $\mathrm{i} \mathfrak{sp}_\mathbb{R}\,$. We begin with the following statement. \[lem:M-closed\] The image $M$ of $\pi$ is closed as a subset of $\mathrm{H}(W^s)$. Let $h \in \mathrm{cl}(M) \subset \mathrm{H} (W^s)$. By the definition of $M$, we still have $h \sigma(h)^{-1} =: m \in M\,$. If $h_n$ is any sequence from $M$ with $h_n \to h\,$, then $h_n \sigma(h_n)^{-1} \to m\,$. But $m$ has a unique square root $\sqrt{m} \in M$ and $h_n = \sigma(h_n)^{-1} \to \sqrt{m} \,$. Hence $h = \sqrt{m} \in M\,$. $M$ of course fails to be closed as a subset of $\mathrm{Sp}\,$. For example, $g = \mathrm{Id}$ is in the closure of $M \subset \mathrm{Sp}$ but is not in $M\,$. \[lem:max-rank\] The exponential map $\exp : \, \mathfrak{sp} \to \mathrm{Sp}$ has maximal rank along $\mathfrak{t}_+\,$. We are going to use the fact that the squaring map $S : \, \mathrm{Sp} \to \mathrm{Sp}\,$, $g \mapsto g^2$, is a local diffeomorphism of $\mathrm{Sp}$ at any point $t \in T_+\,$. To show this, we compute the differential of $S$ at $t$ and obtain $$D_t S = dL_{\,t^2} \circ (\mathrm{Id}_\mathfrak{sp} + \mathrm{Ad}(t^{-1})) \circ dL_{\,t^{-1}} \;,$$ where $dL_g$ denotes the differential of the left translation $L_g : \,\mathrm{Sp}\to \mathrm{Sp}\,$, $g_1 \mapsto g g_1\,$. The middle map $\mathrm{Id}_\mathfrak{sp} + \mathrm{Ad}(t^{-1}) : \, \mathfrak{sp} \to \mathfrak{sp}$ is regular because all of the eigenvalues of $t \in T_+$ are positive real numbers. Since $dL_{\, t^{-1}} : \, T_t \mathrm{Sp} \to \mathfrak{sp}$ and $dL_{\,t^2} : \, \mathfrak{sp} \to T_{t^2} \mathrm{Sp}$ are isomorphisms, it follows that $D_t S :\, T_t \mathrm{Sp} \to T_{t^2} \mathrm{Sp}$ is an isomorphism. Turning to the proof of the lemma, given $\xi \in \mathfrak{t}_+$ we now choose $n \in \mathbb{N}$ so that $2^{-n} \xi$ is in a neighborhood of $0 \in \mathfrak{sp}$ where $\exp$ is a diffeomorphism. It follows that for $U$ a sufficiently small neighborhood of $\xi\,$, the exponential map expressed as $$U \ni \eta \mapsto 2^{-n} \eta \mapsto \exp(2^{-n}\eta) \stackrel{S^n}{\mapsto}(\exp(2^{-n}\eta))^{2^n} = \exp(\eta)$$ is a diffeomorphism of $U$ onto its image. Now recall $M = \mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+$ and consider the cone $$\mathcal{C} := \mathrm{Ad}(\mathrm{Sp}_\mathbb{R}) \mathfrak{t}_+ \subset \mathrm{i}\mathfrak{sp}_\mathbb{R} \;.$$ It follows from the equivariance of $\exp$, i.e., $\exp(\mathrm{Ad}(g) \xi) = \mathrm{Int}(g) \exp(\xi)$, that $\exp : \, \mathcal{C} \to \mathrm{Int}(\mathrm{Sp}_\mathbb{R}) T_+ = M$ is surjective. Furthermore, $\exp \vert_{\mathfrak{t}_+} : \, \mathfrak{t}_+ \to T_+$ is injective and for every $\xi \in \mathfrak{t}_+$ the isotropy groups of the $\mathrm{Sp}_\mathbb{R}$-actions at $\xi$ and $\exp(\xi)$ are the same. Therefore $\exp : \, \mathcal{C} \to M$ is also injective. In fact, much stronger regularity holds. For the statement of this result we recall the anti-holomorphic involution $\psi:\, \mathrm{H} (W^s) \to \mathrm{H}(W^s)$ defined by $h \mapsto \sigma(h^{-1})$ and let $\mathrm{Fix}(\psi)^0$ denote the connected component of $\mathrm{Fix} (\psi)$ that contains $M$. The image $M \subset \mathrm{H}(W^s)$ of $\pi : \, h \mapsto h \sigma(h^{-1})$ is the closed, connected, totally real submanifold $\mathrm{Fix}(\psi)^0$, which is half-dimensional in the sense that $\dim_\mathbb{R} M = \dim_\mathbb{C} \mathrm{H} (W^s)$. The set $\mathcal{C} = \mathrm{Ad}(\mathrm{Sp}_\mathbb{R}) \mathfrak{t}_+\,$, which is an open positive cone in $\mathrm{i} \mathfrak{sp}_\mathbb{R}\,$, is in bijection with $M$ by the real-analytic diffeomorphism $\exp : \, \mathcal{C} \to M$. Lemma \[lem:max-rank\] implies that $\mathfrak{t}_+$ possesses an open neighborhood $U$ in $\mathrm{i} \mathfrak{sp}_\mathbb{R}$ so that $\exp \vert_U$ is everywhere of maximal rank. Because $T_+$ lies in $\mathrm{H}( W^s)$ and $\mathrm{H}(W^s)$ is open in $\mathrm{Sp}\,$, by choosing $U$ small enough we may assume that $\exp(\frac{1}{2} U) \subset \mathrm {H} (W^s)$ and therefore that $\exp(U) \subset M = \exp (\mathcal{C})$. Since $\exp$ is a local diffeomorphism on $U$, we may also assume that $U \subset \mathcal{C} = \mathrm{Ad}(\mathrm{Sp}_\mathbb{R}) \mathfrak{t}_+\,$, and it then follows that $\mathcal{C} = \mathrm{Ad}(\mathrm{Sp}_\mathbb{R}) U$. In particular, this shows that $\mathcal{C}$ is open in $\mathrm{i}\mathfrak{sp}_\mathbb{R}\,$. In summary, $$\mathcal{C} = \mathrm{Ad}(\mathrm{Sp}_\mathbb{R}) U \stackrel{\exp}{\longrightarrow} \mathrm{Int}(\mathrm{Sp}_\mathbb{R}) \exp(U) = M \subset \mathrm{Fix}(\psi)^0 \;.$$ By the equivariance of $\exp\,$, we also know that it is everywhere of maximal rank on $\mathcal{C}\,$. Now $\psi$ is an anti-holomorphic involution. Therefore, $\mathrm {Fix}(\psi)^0$ is a totally real, half-dimensional closed submanifold of $\mathrm{H}(W^s)$, and since $\mathcal{C}$ is open in $\mathrm{i} \mathfrak{sp}_\mathbb{R}\,$, we also know that $\dim_\mathbb{C} \mathcal{C} = \dim_\mathbb{R} \mathrm{Fix}(\psi)^0$. The maximal rank property of $\exp$ then implies that $M = \mathrm{im}(\exp : \mathcal{C} \to \mathrm{Fix} (\psi)^0)$ is open in $\mathrm{Fix} (\psi)^0$. In Lemma \[lem:M-closed\] it was shown that $M$ is closed in $\mathrm{H} (W^s)$. Thus it is open and closed in the connected manifold $\mathrm{Fix}(\psi)^0$, and consequently $\exp : \, \mathcal{C} \to M = \mathrm{Fix}(\psi)^0$ is a local diffeomorphism of manifolds. Since we already know that $\exp : \, \mathcal{C} \to M$ is bijective, the desired result follows. \[semigroup product structure\] The two identifications $\mathrm{Sp}_\mathbb{R} \times M = \mathrm{H}(W^s)$ defined by $(g,m) \mapsto gm$ and $(g,m) \mapsto m g^{-1}$ are real-analytic diffeomorphisms. The fundamental group of $\mathrm{H}(W^s)$ is isomorphic to $\pi_1(\mathrm{Sp}_\mathbb{R}) \simeq \mathbb{Z}\,$. The first statement is proved by explicitly constructing a smooth inverse to each of the two maps. For this let $g^\prime m^\prime = m g^{-1} = h \in \mathrm{H}(W^s)$ and note that $m = \sqrt{\pi(h)}$ and $m^\prime = \sqrt{\pi^\prime (h)}$. Since the square root is a smooth map on $M\,$, a smooth inverse in the two cases is defined by $$h \mapsto ( h \sqrt{\pi^\prime(h)}^{\ -1} , \sqrt{\pi^\prime(h)}) \;, \quad \text{resp.} \quad h \mapsto ( h^{-1} \sqrt{\pi(h)} , \sqrt{\pi(h)} ) \;.$$ The second statement follows from $\mathcal{C} \simeq M$ by $\exp$ and the fact that $\mathrm{Sp}_\mathbb{R}$ is a product of a cell and a maximal compact subgroup $K\,$. We choose $K$ to be the unitary group $\mathrm{U} = \mathrm{U}(V, \langle \, ,\,\rangle)$ acting diagonally on $W = V \oplus V^*$ and recall that $\pi_1(\mathrm{U}) \simeq \mathbb{Z}\,$. Oscillator semigroup and metaplectic group {#oscillator semigroup} ------------------------------------------ Recall that we are concerned with the Lie algebra representation of $\mathfrak{sp}_\mathbb{R} \subset \mathfrak{sp}$ which is defined by the identification of $\mathfrak{sp}$ with the set of symmetrized elements of degree two in the Weyl algebra $\mathfrak{w}(W)$ and the representation of $\mathfrak{w}(W)$ on $\mathfrak{a}(V)$. In $\S$\[metaplectic rep\] we construct the oscillator representation of the metaplectic group $\mathrm{Mp}\,$, a 2:1 cover of $\mathrm{Sp}_\mathbb{R}\,$, which integrates this Lie algebra representation. Observe that since $\pi_1(\mathrm{Sp}_\mathbb{R}) \simeq \mathbb{Z}$ and $\mathbb{Z}$ has only one subgroup of index two, there is a unique such covering $\tau : \, \mathrm{Mp} \to \mathrm{Sp}_\mathbb{R}\,$. The method of construction [@H2] we use first yields a representation of the 2:1 covering space $\widetilde{\mathrm{H}} (W^s)$ of $\mathrm{H}(W^s)$ and then realizes the oscillator representation of $\mathrm{Mp}$ by taking limits that correspond to going to $\mathrm{Sp}_\mathbb{R}$ in the boundary of $\mathrm{H}(W^s)$. This representation of the *oscillator semigroup* $\widetilde{\mathrm{H}}(W^s)$ is for our purposes at least as important as the representation of $\mathrm{Mp}\,$. The goal of the present section is to lift all essential structures on $\mathrm{H}(W^s)$ to $\widetilde{\mathrm{H}}(W^s)$. ### Lifting the semigroup We begin by recalling a few basic facts about covering spaces. If $G$ is a connected Lie group, its universal covering space $U$ carries a canonical group structure: an element $u \in U$ in the fiber over $g \in G$ is a homotopy class $u \equiv [\alpha_g]$ of paths $\alpha_g : \, [0,1] \to G$ connecting $g$ with the neutral element $e \in G\,$; and an associative product $U \times U \to U$, $(u_1,u_2) \mapsto u_1 u_2\,$, is defined by taking $u_1 u_2$ to be the unique homotopy class which is given by pointwise multiplication of any two paths representing the homotopy classes $u_1, u_2\,$. The fundamental group $\pi_1(G) \equiv \pi_1(G,e)$ acts on $U$ by monodromy, i.e., if $[\alpha_g] = u \in U$ and $[c] = \gamma \in \pi_1(G)$, then one sets $\gamma(u) := [ \alpha_g \ast c] \in U$ where $\alpha_g \ast c$ is the path from $g$ to $e$ which is obtained by composing the path $\alpha_g$ with the loop $c$ based at $e\,$. This $\pi_1(G)$-action satisfies the compatibility condition $\gamma_1 (u_1) \gamma_2(u_2) = (\gamma_1 \gamma_2)(u_1 u_2)$ and in that sense is central. The situation for our semigroup $\mathrm{H} (W^s)$ is analogous except for the minor complication that the distinguished point $e = \mathrm{Id}$ does not lie in $\mathrm{H}(W^s)$ but lies in the closure of $\mathrm{H}(W^s)$. Hence, by the same principles, the universal cover $U$ of $\mathrm{H}(W^s)$ comes with a product operation and there is a central action of $\pi_1(\mathrm{H} (W^s))$ on $U$. Moreover, the product $U \times U \to U$ still is associative. To see this, first notice that the subsemigroup $T_+ \subset \mathrm{H}(W^s)$ is simply connected and as such is canonically embedded in $U$. Then for $u_1, u_2, u_3 \in U$ observe that $u_1 (u_2 u_3) = \gamma ((u_1 u_2) u_3)$ where $\gamma \in \pi_1 (\mathrm{H}(W^s))$ could theoretically depend on the $u_j\,$. However, any such dependence has to be continuous and the fundamental group is discrete, so in fact $\gamma$ is independent of the $u_j$ and, since $\gamma$ is the identity when the $u_j$ are in $T_+$ (lifted to $U$), the associativity follows. Let now $\Gamma \simeq 2\mathbb{Z}$ denote the subgroup of index two in $\pi_1 (\mathrm{H}(W^s)) \simeq \mathbb{Z}$ and consider $\widetilde{ \mathrm{H}} (W^s) := U/\Gamma$, which is our object of interest. Since the $\Gamma$-action on $U$ is central, i.e., $\gamma_1 (u_1) \gamma_2 (u_2) = (\gamma_1 \gamma_2) (u_1 u_2)$ for all $\gamma_1 , \gamma_2 \in \Gamma$ and $u_1 , u_2 \in U$, the product $U \times U \to U$ descends to a product $U/\Gamma \times U/\Gamma \to U/\Gamma \,$. Thus $U/ \Gamma = \widetilde{\mathrm{H}} (W^s)$ is a semigroup, and the situation at hand is summarized by the following statement. The 2:1 covering $\tau_H : \, U / \Gamma = \widetilde{\mathrm{H}} (W^s) \to \mathrm{H}(W^s)$, $[\alpha_h]\, \Gamma \mapsto h\,$, is a homomorphism of semigroups. ### Actions of the metaplectic group Recall that we have two 2:1 coverings: a homomorphism of groups $\tau:\, \mathrm{Mp} \to \mathrm{Sp}_\mathbb{R}\,$, along with a homomorphism of semigroups $\tau_H : \, \widetilde{\mathrm{H}}(W^s) \to \mathrm{H}(W^s)$. Now, a pair of elements $(g^\prime,g) \in \mathrm{Sp}_\mathbb{R} \times \mathrm{Sp}_\mathbb{R}$ determines a transformation $h \mapsto g^\prime h g^{-1}$ of $\mathrm{H} (W^s)$, and by the homotopy lifting property of covering maps a corresponding action of $\mathrm{Mp} \times \mathrm{Mp}$ on $\widetilde{ \mathrm{H}}(W^s)$ is obtained as follows. Consider the canonical mapping $\mathrm{Mp} \times M \to \mathrm{Sp}_\mathbb{R} \times M$ given by $\tau\,$. By the real-analytic diffeomorphism $\mathrm{Sp}_\mathbb{R} \times M \to \mathrm{H} (W^s)$, $(g,m) \mapsto g m\,$, this map can be regarded as a 2:1 covering of $\mathrm{H}(W^s)$, and since any two 2:1 coverings are isomorphic we get an identification of the covering space $\widetilde{\mathrm{H}}(W^s)$ with $\mathrm{Mp} \times M\,$. Moreover, the action of the group $\mathrm{Mp}$ on itself by left translations induces on $\mathrm{Mp} \times M \simeq \widetilde{\mathrm{ H}} (W^s)$ an $\mathrm{Mp}$-action which, by construction, satisfies the relation $$\tau_H(g \cdot h) = \tau(g) \tau_H(h) \quad (g \in \mathrm{Mp} \, , \,\, h \in \widetilde{\mathrm{H}}(W^s) ) \;.$$ This can be viewed as a statement of $\mathrm{Mp}$-equivariance of the covering map $\tau_H\,$. Now, we have another real-analytic diffeomorphism $\mathrm{Sp}_\mathbb{R} \times M \to \mathrm{H}(W^s)$ by $(g,m) \mapsto m g^{-1}$, which transfers left translation in $\mathrm{Sp}_\mathbb{R}$ to right multiplication on $\mathrm{H}(W^s)$, and by using it we can repeat the above construction. The result is another identification $\widetilde{ \mathrm{H}}(W^s) \simeq \mathrm{Mp} \times M$ and another $\mathrm{Mp}$-action on $\widetilde{\mathrm{H}}(W^s)$. Altogether we then have two actions of $\mathrm{Mp}$ on $\widetilde{\mathrm{H}} (W^s)$. The essence of the next statement is that they commute. \[lifting the product action\] There is a real-analytic action $(g_1,g_2) \mapsto (g_1,g_2) \cdot h$ of $\, \mathrm{Mp} \times \mathrm{Mp}$ on $\widetilde{\mathrm{H}} (W^s)$ such that the covering $\tau_H : \, \widetilde{ \mathrm{H}}(W^s) \to \mathrm{H} (W^s)$ is $(\mathrm{Mp} \times \mathrm{Mp})$-equivariant: $$\tau_H( (g_1,g_2) \cdot h) = \tau(g_1) \tau_H(h) \tau(g_2)^{-1}\;.$$ By construction, the stated equivariance property of $\tau_H$ holds for each of the two actions of $\mathrm{Mp}$ separately. It then follows that it holds for all $(g_1,g_2) \in \mathrm{Mp} \times \mathrm{Mp}$ if the two actions commute. But by $\tau_H( (g_1,e) \cdot h) = \tau(g_1) \tau_H(h)$ and $\tau_H( (e,g_2) \cdot h) = \tau_H(h) \tau(g_2)^{-1}$ the commutator $$g := (g_1,e) (e,g_2) (g_1,e)^{-1} (e,g_2)^{-1}$$ acts trivially on $\mathrm{H}(W^s)$ by $\tau_H\,$, i.e., $\tau_H(g \cdot h) = \tau_H(h)$. Therefore $g$ can be regarded as being in the covering group $\tau^{-1}(\mathrm{Id}) = \mathbb{Z}_2$ of the covering $\tau : \, \mathrm{Mp} \to \mathrm{Sp}_\mathbb{R}\,$. Since we can connect both $g_1$ and $g_2$ to the identity $e \in \mathrm{Mp}$ by a continuous curve, it follows from the discreteness of $\mathbb{Z}_2$ that $g \in \mathrm{Mp} \times \mathrm{Mp}$ acts trivially on $\widetilde{\mathrm{H}}(W^s)$. Notice that since the submanifold $M \subset \mathrm {H} (W^s)$ is simply connected, there exists a canonical lifting of $M$ (which we still denote by $M$) to the cover $\widetilde{\mathrm{H}}(W^s)$; this is the unique lifting by which $T_+ \subset M$ is embedded as a subsemigroup in $\widetilde{\mathrm{H}}(W^s)$. Proposition \[lifting the product action\] then allows us to write $\widetilde{ \mathrm{H}}(W^s) = \mathrm{Mp}.M.\mathrm{Mp}$. ### Lifting involutions Let us now turn to the issue of lifting the various involutions at hand. As a first remark, we observe that any Lie group automorphism $\varphi: \, \mathrm{Sp}_\mathbb{R} \to \mathrm{Sp}_\mathbb{R}$ uniquely lifts to a Lie group automorphism $\widetilde{\varphi}$ of the universal covering group $\widetilde{\mathrm{Sp}}_\mathbb{R}\,$, and the latter induces an automorphism of the fundamental group $\pi_1(\mathrm{Sp}_\mathbb{R}) \simeq \mathbb{Z}$ viewed as a subgroup of the center of $\widetilde{\mathrm{Sp}}_\mathbb{R}\,$. Now $\mathrm{Aut}( \pi_1( \mathrm{Sp}_\mathbb{R} )) \simeq \mathrm{Aut}(\mathbb{Z}) \simeq \mathbb{Z}_2$ and both elements of this automorphism group stabilize the subgroup $\Gamma \simeq 2\mathbb{Z}$ in $\pi_1(\mathrm{Sp}_\mathbb{R})$. Therefore $\widetilde{ \varphi}$ induces an automorphism of $\mathrm{Mp} = \widetilde{\mathrm{Sp}}_\mathbb{R} / \Gamma\,$. Since the operation $h\mapsto h^{-1}$ canonically lifts from $\mathrm{Sp}_\mathbb{R}$ to $\mathrm{Mp}$ and $h \mapsto (h^{-1})^\dagger$ is a Lie group automorphism of $\mathrm{Sp}_\mathbb{R}\,$, it follows that Hermitian conjugation $h\mapsto h^\dagger$ has a natural lift to $\mathrm{Mp}\,$. The same goes for the Lie group automorphism $h \mapsto s h s$ of $\mathrm{Sp}_\mathbb{R}\,$. Hermitian conjugation $h \mapsto h^\dagger$ and the involution $h \mapsto s h s$ lift to unique maps with the property that they stabilize the lifted manifold $M$. In particular, the basic anti-holomorphic map $\psi:\, \mathrm{H}(W^s) \to \mathrm{H}(W^s)$, $h \mapsto \sigma (h^{-1})= s h^\dagger s$ lifts to an anti-holomorphic map $\widetilde{\psi}: \, \widetilde{\mathrm{H}} (W^s)\to \widetilde{\mathrm {H}}(W^s)$ which is the identity on $M$ and $\mathrm{Mp} \times \mathrm{Mp}$-equivariant in that $\widetilde{\psi} (g_1 x\, g_2^{-1}) = g_2 \widetilde{\psi}(x) g_1^{-1}$ for all $g_1, g_2\in \mathrm{Mp}$ and $x \in \widetilde{\mathrm{H}}(W^s)$. Recall that the simply connected space $M \subset \mathrm{H} (W^s)$ has a canonical lifting (still denoted by $M$) to $\widetilde{ \mathrm{H}}(W^s)$. Since all of our involutions stabilize $M$ as a submanifold of $\mathrm{H}(W^s)$, they are canonically defined on the lifted manifold $M\,$. In particular, the involution $\psi$ on $M$ is the identity map, and therefore so is the lifted involution $\widetilde{\psi}$. Note furthermore that the involution defined by $h \mapsto s h s$ is holomorphic on $\mathrm{H}(W^s)$ and that the other two are anti-holomorphic. Now $\widetilde{\mathrm{H}}(W^s)$ is connected and the lifted version of $M$ is a totally real submanifold of $\widetilde{\mathrm{H}}(W^s)$ with $\dim_\mathbb{R} M = \dim_\mathbb{C} \widetilde{ \mathrm{H}}(W^s)$. In such a situation the identity principle of complex analysis implies that there exists at most one extension (holomorphic or anti-holomorphic) of an involution from $M$ to $\widetilde{ \mathrm{H}}(W^s)$. Therefore, it is enough to prove the existence of extensions. Since $h \in \widetilde{\mathrm{H}}(W^s)$ is uniquely representable as $h = g m$ with $g \in \mathrm{Mp}$ and $m \in M$, the involution $h \mapsto h^\dagger$ is extended by $gm \mapsto (gm)^\dagger = m^\dagger g^\dagger$. Similarly, $h \mapsto s h s$ extends by $gm \mapsto (sgs)(sms)$, and $h \mapsto s h^\dagger s$ does so by the composition of the other two. The equivariance property of $\widetilde{\psi}$ follows from the fact that $g \mapsto s g^\dagger s$ on $\mathrm{Mp}$ coincides with the operation of taking the inverse, $g \mapsto g^{-1}$. Oscillator semigroup representation {#sec:semigrouprep} ----------------------------------- Here we construct the fundamental representation of the semigroup $\widetilde{\mathrm{H}}(W^s)$ on the Hilbert space $\mathcal{A}_V$, which in the present context we call Fock space. Our approach is parallel to that of Howe [@H2]: the Fock space we use is related to the $L^2$-space of Howe’s work by the Bargmann transform [@Folland]. (Using the language of physics, one would say that Howe works with the position wave function while our treatment relies on the phase space wave function.) In particular, following Howe we take advantage of a realization of $\mathrm{H} (W^s)$ as the complement of a certain determinantal variety in the Siegel upper half plane. ### Cayley transformation Let us begin with some background information on the Cayley transformation, which is defined to be the meromorphic mapping $$C : \,\, \mathrm{End}(W) \to \mathrm{End}(W), \quad g \mapsto \frac{\mathrm{Id}_W + g}{\mathrm{Id}_W - g} \,.$$ If $g \in \mathrm{Sp}\,$, then from $A(gw , gw^\prime) = A(w , w^\prime)$ we have $$A((\mathrm{Id}_W + g)w \, , (\mathrm{Id}_W - g)w') + A((\mathrm{Id}_W - g) w \, , (\mathrm{Id}_W + g)w') = 0$$ for all $w , w' \in W$. By assuming that $(\mathrm{Id}_W - g)$ is invertible and then replacing $w$ and $w'$ by $(\mathrm{Id}_W - g)^{-1}w$ resp. $(\mathrm{Id}_W - g)^{-1} w'$, we see that $C$ maps the complement of the determinantal variety $\{ g \in \mathrm{Sp} \mid \mathrm{Det}(\mathrm{Id}_W - g) = 0\}$ into $\mathfrak{sp}\,$. The inverse of the Cayley transformation is given by $$g = C^{-1}(X) = \frac{X - \mathrm{Id}_W}{X + \mathrm{Id}_W} \,.$$ Reversing the above argument, one shows that if $(X + \mathrm{Id}_W)$ is invertible and $X \in \mathfrak{sp}\,$, then $C^{-1}(X) \in \mathrm{Sp}\,$. Moreover, by the relation $X + \mathrm{Id}_W = 2 (\mathrm{Id}_W - g)^{-1}$ for $C(g) = X\,$, if $\mathrm{Id}_W - g$ is regular, then so is $X + \mathrm{Id}_W\,$, and vice versa. Thus if we introduce the sets $$D_{\mathrm{Sp}} := \{g \in \mathrm{Sp} \mid \mathrm{Det}( \mathrm{Id}_W - g) = 0 \}\;, \quad D_{\mathfrak{sp}} := \{X \in \mathfrak{sp} \mid \mathrm{Det}(X + \mathrm{Id}_W) = 0 \}\;,$$ the following is immediate. \[Cayley biholo\] The Cayley transformation defines a bi-holomorphic map $$C:\,\, \mathrm{Sp} \setminus D_{\mathrm{Sp}}\to \mathfrak{sp} \setminus D_{\mathfrak{sp}}\;.$$ Now we consider the restriction of $C$ to the semigroup $\mathrm{H} (W^s)$. Letting $\dagger$ be the Hermitian conjugation of the previous section, denote by $\mathfrak{Re}(X) = \frac{1}{2}(X+ X^\dagger)$ the real part of an operator $X \in \mathrm{End}(W)$ and define the associated Siegel upper half space $\mathfrak{S}$ to be the subset of elements $X \in \mathrm{End}(W)$ which are symmetric with respect to the canonical symmetric bilinear form $S$ on $W = V \oplus V^*$ with $\mathfrak{Re} (X) > 0\,$. Notice the relations $S(w,w^\prime) = A(w,sw^\prime)$ and $A(sw,sw^\prime) = - A(w,w^\prime)$, from which it is seen that $X$ is symmetric if and only if $sX \in \mathfrak{sp}\,$. Define $D_\mathfrak{S} := \{X\in \mathfrak{S}\mid \mathrm{Det}(sX + \mathrm{Id}_W) = 0 \}$, let $$\zeta^+_s := \mathfrak{S} \setminus D_\mathfrak{S} \,,$$ and define a slightly modified Cayley transformation by $$a(g) := s \, \frac{\mathrm{Id}_W + g}{\mathrm{Id}_W - g}\,.$$ Translating Proposition \[Cayley biholo\], it follows that $a$ defines a bi-holomorphic map from $\mathrm{Sp} \setminus D_{\mathrm{Sp}}$ onto the set of $S$-symmetric operators with $D_\mathfrak{S}$ removed. The modified Cayley transformation $a : \, \mathrm{Sp} \setminus D_{\mathrm{Sp}} \to \mathrm{End}(W)$ given by $g \mapsto s \, (\mathrm{Id}_W + g) (\mathrm{Id}_W - g)^{-1}$ restricts to a bi-holomorphic map $$a : \,\, \mathrm{H}(W^s) \to \zeta^+_s\,.$$ This result is an immediate consequence of the following identity. For $g \in \mathrm{Sp} \setminus D_{\mathrm{Sp}}$ let $a(g) = X$ and define $Y := (sX + \mathrm{Id}_W)^{-1}$. Then $${\textstyle{\frac{1}{2}}} ( s - g^\dagger s g ) = Y^\dagger (X + X^\dagger) Y\,.$$ In particular, one has the following equivalence: $$\big( \mathfrak{Re}(X) > 0 \,\, \text{and} \,\, \mathrm{Det}(sX + \mathrm{Id}_W) \not = 0 \big) ~ \Leftrightarrow ~ s - g^\dagger s g > 0 \;.$$ It is convenient to rewrite $s - g^\dagger s g$ as $$s - g^\dagger s g = {\textstyle{\frac{1}{2}}} (\mathrm{Id}_W - g^\dagger) s (\mathrm{Id}_W + g)+ {\textstyle{\frac{1}{2}}}(\mathrm{Id}_W + g^\dagger) s (\mathrm{Id}_W - g)\,.$$ Using $a(g) = X$ one directly computes that $${\textstyle{\frac{1}{2}}}(\mathrm{Id}_W - g) = (sX + \mathrm{Id}_W)^{-1} \quad \text{and} \quad {\textstyle{\frac{1}{2}}}(\mathrm{Id}_W + g) = (sX + \mathrm{Id}_W)^{-1} sX \,.$$ The desired identity follows by inserting these relations in the previous equation. \[rem:twine-inv\] The modified Cayley transformation intertwines the anti-holomorphic involution $\psi : \, h \mapsto s h^\dagger s$ with the operation of taking the Hermitian conjugate $X \mapsto X^\dagger$: $$a(\psi(h)) = a(h)^\dagger \;.$$ Since $\zeta_{\,s}^+$ is obviously stable under Hermitian conjugation, this is another proof of the stability of $\mathrm{H}(W^s)$ under the involution $\psi\,$; cf. Corollary \[cor:3.2\]. ### Construction of the semigroup representation Let us now turn to the main goal of this section. Recall that we have a Lie algebra representation of $\mathfrak{sp}$ on $\mathfrak{a}(V) = \mathrm{S}(V^*)$ which is defined by its canonical embedding in $\mathfrak{w}_2(W)$. We shall now construct the corresponding representation of the semigroup $\widetilde{\mathrm{H}}(W^s)$ on the Fock space $\mathcal{A}_V$. It will be seen later that the character of this representation on the lifted toral semi-group $T_+$ is $\mathrm{Det}^{-\frac{1}{2}}(s - sh)$. This extends to $M = \mathrm{Int}(\mathrm{Mp}) T_+$ by the invariance of the character with respect to the conjugation action of $\mathrm{Mp}\,$. Since $\widetilde{\mathrm{H}}(W^s)$ is connected and $M$ is totally real of maximal dimension in $\widetilde{\mathrm{H}}(W^s)$, the identity principle then implies that if a semigroup representation of $\widetilde{\mathrm{H}}(W^s)$ can be constructed with a holomorphic character, this character must be given by the square root function $h \mapsto \mathrm{Det}^{-\frac{1}{2}}(s - sh)$. There is no difficulty discussing the square root on the simply connected submanifold $M$. However, in order to make sense of the square root function on the full semigroup, we must lift all considerations to $\widetilde{\mathrm{H}}(W^s)$. For convenience of notation, given $h \in \mathrm{H}(W^s)$ we let $a_h := a(h)$, and for $x \in \widetilde{\mathrm{H}}(W^s)$ we simply write $a_x \equiv a(\tau_H(x))$ where $\tau_H: \, \widetilde{\mathrm{H}}(W^s) \to \mathrm{H}(W^s)$ is the canonical covering map. Then we put $$f(h) := \mathrm{Det}(a_h + s) = \mathrm{Det}(2s (\mathrm{Id}_W - h)^{-1}) \;,$$ and wish to define $\phi :\, \widetilde{\mathrm{H}}(W^s)\to \mathbb{C}$ to be the square root of $f$ which agrees with the positive square root on $T_+\,$. (Here we regard $T_+$ as being in $\widetilde{\mathrm{H}} (W^s)$ by its canonical lifting as a subsemigroup of $\mathrm{H} (W^s)$ as in the previous section). This is possible because $\phi$ is naturally defined on $\{(\xi , \eta) \in \mathrm{H}(W^s)\times \mathbb{C} \mid f(\xi) = \eta^2 \}$ which is itself a 2:1 cover of $\mathrm{H} (W^s)$. Since up to equivariant equivalence there is only one such covering, namely $\widetilde{\mathrm{H}}(W^s) \to \mathrm{H}(W^s)$, it follows that we may define $\phi $ on $\widetilde{\mathrm{H}} (W^s)$ as desired. For the construction of the oscillator representation it is useful to observe that $\phi$ can be extended to a slightly larger space. This extension is constructed as follows. Regard the complex symplectic group $\mathrm{Sp}$ as the total space of an $\mathrm{Sp}_\mathbb{R}$-principal bundle $\pi : \, \mathrm{Sp} \to \pi(\mathrm{Sp})$, $g \mapsto g \sigma(g^{-1})$. Recall that the restricted map $\pi : \, M \to M$ is a diffeomorphism, and that $M$ contains the neutral element $\mathrm{Id} \in \mathrm{Sp}$ in its boundary. We choose a small ball $B$ centered at $\mathrm{Id}$ in the base $\pi(\mathrm{Sp})$, and using the fact that $M$ can be identified with a cone in $\mathrm{i}\mathfrak{sp}_\mathbb{R}$ we observe that $A := B \, \cup \, M \subset \pi(\mathrm{Sp})$ is contractible. Now $U := \pi^{-1}(A)$ is diffeomorphic to a product $\mathrm{Sp}_\mathbb{R} \times A$ and thus comes with a 2:1 covering $\widetilde{U} \to U$ defined by $\tau: \, \mathrm{Mp} \to \mathrm{Sp}_\mathbb{R}\,$. The covering space $\widetilde{U}$ contains $\widetilde{ \mathrm{H}}(W^s)$, and is invariant under the $\mathrm{Mp}$-action by right multiplication. By construction it also contains the metaplectic group $\mathrm{Mp}\,$, which covers the group $\mathrm{Sp}_\mathbb{R}$ in $\mathrm{Sp}\,$. Recall the definition of the determinant variety $D_{\mathrm{Sp}} = \{g\in \mathrm{Sp} \mid \mathrm{Det}(\mathrm{Id}_W - g) = 0 \}$. Let $\widetilde{D}_{\mathrm{Sp}}$ denote the set of points in $\widetilde{U}$ which lie over $D_{\mathrm{Sp}} \cap U$ by the covering $\widetilde{U} \to U$. \[Mp in the boundary\] There is a unique continuous extension of $\phi $ from $\widetilde{\mathrm{H}}(W^s)$ to its closure in $\widetilde{U}$ so that $\phi^2$ agrees with the lift of $f$ from $U$. The intersection of $\widetilde{D}_{\mathrm{Sp}}$ with any $\mathrm{Mp}$-orbit in $\widetilde{U}$ is nowhere dense in that orbit and the restriction of the extended function $\phi$ to the complement of that intersection is real-analytic. Before beginning the proof, it should be clarified that at the points of the lifted determinant variety, i.e., where the lifted square root $\phi$ of the function $f(g) = \mathrm{Det}(2s\,(\mathrm{Id}_W - g)^{-1})$ has a pole, *continuity of the extension* means that the reciprocal of $\phi$ extends to a continuous function which vanishes on that set. The intersection of $D_{\mathrm{Sp}}$ with any $\mathrm{Sp}_\mathbb{R}$-orbit in $U$ is nowhere dense in that orbit; therefore the same holds for the intersection of $\widetilde{D}_{\mathrm{Sp}}$ with any $\mathrm{Mp}$-orbit in $\widetilde{U}$. Let $x \in \widetilde{U} \setminus \widetilde{D}_{\mathrm{Sp}}$ be a point of the boundary of $\widetilde{\mathrm{H}}(W^s)$. Choose a local contractible section $\Sigma \subset \widetilde{U}$ of $\widetilde{U} \to A$ with $x \in \Sigma$ and a neighborhood $\Delta$ of the identity in $\mathrm{Mp}$ so that the map $\Delta \times \Sigma \to \widetilde{U}\,$, $(g,s) \mapsto s g^{-1}$, realizes $\Delta \times \Sigma$ as a neighborhood $\widetilde{V}$ of $x$ which has empty intersection with $\widetilde{D}_{\mathrm{Sp}}\,$. By construction $\widetilde{V} \cap \widetilde{\mathrm{H}}(W^s)$ is connected and is itself simply connected. Thus the desired unique extension of $\phi$ exists on $\widetilde{V}$. At $x$ this extension is simply defined by taking limits of $\phi$ along arbitrary sequences $\{ x_n \}$ from $\widetilde{\mathrm{H}}(W^s)$. Thus the extended function (still called $\phi$) is well-defined on the closure of $\widetilde{\mathrm {H}}(W^s)$ and is real-analytic on the complement of $\widetilde{D} _{\mathrm{Sp}}$ in every $\mathrm{Mp}$-orbit in that closure. It extends as a continuous function on the full closure of $\widetilde{ \mathrm{H}} (W^s)$ by defining it to be identically $\infty$ on $\widetilde{D}_{ \mathrm{Sp}}\,$, i.e., its reciprocal vanishes identically at these points. Now let us proceed with our main objective of defining the semigroup representation on $\widetilde{\mathrm{H}}(W^s)$. Recall the involution $\psi : \, \mathrm{H}(W^s) \to \mathrm{H}(W^s)$, $h \mapsto \sigma(h)^{-1}$, and its lift $\widetilde{\psi}$ to $\widetilde {\mathrm{H}}(W^s)$. The following will be of use at several points in the sequel. \[psi equivariance\] $\phi \circ \widetilde{\psi} = \overline{\phi}\,$. By direct calculation, $f \circ \psi = \overline{f}\,$. Thus, since $f = \phi^2$, we have either $\phi \circ \widetilde{\psi} = \overline{\phi}$ or $\phi \circ \widetilde{\psi} = - \overline{ \phi}\,$. The latter is not the case, as $\phi$ is not purely imaginary on the nonempty set $\mathrm{Fix} (\widetilde{\psi})$. The semigroup representation $R :\, \widetilde{\mathrm{H}}(W^s) \to \mathrm{End}(\mathcal{A}_V)$ will be given by a certain averaging process involving the standard representation of the Heisenberg group. The latter representation is defined as follows. For elements $w = v + cv$ of the real vector space $W_\mathbb{R}\,$, the operator $\delta (v) + \mu(cv)$ is self-adjoint and its exponential $$T_{v + cv} := \mathrm{e}^{\mathrm{i}\delta(v) + \mathrm{i}\mu(cv)}$$ converges and is unitary (see, e.g., [@RS]). These operators satisfy the relation $$\label{Heisenberg multiplication} T_{w^{\vphantom{\prime}}} T_{w'} = T_{w + w^\prime} \, \mathrm{e}^{\frac{\mathrm{i}}{2} \omega (w,w')} \quad (w, w^\prime \in W_\mathbb{R}) \;,$$ where $\omega := \mathrm{i} A \vert_{W_\mathbb{R}}$ is the induced real symplectic structure. If $T \mapsto T^\dagger$ denotes the adjoint operation in $\mathrm{End}(\mathcal{A}_V)$, it follows from $\delta(v)^\dagger = \mu(cv)$ that $$T_w^\dagger = T_{-w}^{\vphantom{\dagger}} = T_w^{-1} \quad (w \in W_\mathbb{R}) \;.$$ Thus if $H := W_\mathbb{R} \times \mathrm{U}_1$ is equipped with the Heisenberg group multiplication law, $$(w,z)\,(w',z') := (w + w', z \, z' \mathrm{e}^{ \frac{\mathrm{i}}{2} \omega(w,w')}) \;,$$ then $(w,z)\mapsto z\, T_w$ is an irreducible unitary representation of $H$ on $\mathcal A_V$. It is well known that up to equivariant isomorphisms there is only one such representation. The oscillator representation $x \mapsto R(x)$ of $\widetilde{ \mathrm{H}}(W^s)$ is now defined by $$R(x) = \int_{W_\mathbb{R}} \gamma_x(w) T_w \, \mathrm{dvol}(w) \;, \quad \gamma_x(w) := \phi(x) \, \mathrm{e}^{-\frac{1}{4} \langle w ,\, a_x w \rangle} \,.$$ Here $\mathrm{dvol}$ is the Euclidean volume element on $W_\mathbb{R}$ which we normalize so that $$\int_{W_\mathbb{R}} \mathrm{e}^{-\frac{1}{4} \langle w , w \rangle} \, \mathrm{dvol}(w) = 1 \;.$$ It should be stressed that we often parameterize $W_\mathbb{R} \simeq V$ by the map $v \mapsto v + cv = w\,$. Notice that by the positivity of $\mathfrak{Re}(a_x)$ the Gaussian function $w \mapsto \gamma_x(w)$ decreases rapidly, so that all integrals involved in the discussion above and below are easily seen to converge. In particular, since the unitary operator $T_w$ (for $w \in W_\mathbb{R}$) has $L^2$-norm $\Vert T_w \Vert = 1$, it follows for any $x \in \widetilde{\mathrm{H}} (W^s)$ that $$\Vert R(x) \Vert \le \vert \phi(x) \vert \int_{W_\mathbb{R}} \mathrm{e}^{-\frac{1}{4}\langle w , \, \mathfrak{Re}(a_{x}) w\rangle}\, \mathrm{dvol}(w) =: C(x) \;,$$ where the bound $C(x)$ by direct computation of the integral is a finite number: $$\label{eq:Schranke} C(x) = \vert \phi(x) \vert \, \mathrm{Det}^{-1/2} \left( \mathfrak{Re}(a_x) \right) = 2^{\, \dim_\mathbb{C} V} \left| \frac{\mathrm{Det} (\mathrm{Id}_W - h)} {\mathrm{Det} (s - h^\dagger s h)} \right|^{1/2}\;, \quad h = \tau_H(x) \;.$$ Thus $R(x)$ is a bounded linear operator on $\mathcal{A}_V$. In Proposition \[operator bound\] we will establish the uniform bound $\Vert R(x) \Vert \le C(x) < 1$ for all $x \in M$. It is also clear that the operator $R(x)$ depends continuously on $x \in \widetilde{ \mathrm{H}}(W^s)$. The main point now is to prove the semigroup multiplication rule $R(xy) = R(x)R(y)$. For this we apply the Heisenberg multiplication formula (\[Heisenberg multiplication\]) to the inner integral of $$R(x)R(y) = \int_{W_\mathbb{R}} \left( \int_{W_\mathbb{R}} \gamma_x(w-w') \gamma_y(w') T_{w-w'} T_{w'}\, \mathrm{dvol}(w') \right)\, \mathrm{dvol}(w)$$ to see that $R(xy) = R(x)R(y)$ is equivalent to the multiplication rule $\gamma_{xy} = \gamma_x \sharp \gamma_y$ where the right-hand side means the twisted convolution $$\label{twisted convolution} \gamma_x \sharp \gamma_y (w) := \int _{W_\mathbb{R}} \gamma_x(w-w') \gamma_y(w') \, \mathrm{e}^{\frac{\mathrm{i}}{2}\omega (w,w')} \, \mathrm{dvol}(w') \;.$$ For the proof of the formula $\gamma_x \sharp \gamma_y = \gamma_{xy} \,$, we will need to know that $\phi$ transforms as $$\label{transformation rule} \phi(xy) = \phi(x)\phi(y)\mathrm{Det}^{-\frac{1}{2}}(a_x + a_y)\;.$$ This transformation behavior, in turn, follows directly from the expression for the semigroup multiplication rule transferred to the upper half-space $\zeta ^+_{\,s}$; we record this expression in the following statement and refer to [@H2], p. 78, for the calculation. \[prop:semigroup-law\] Identifying $\mathrm{H}(W^s)$ with $\zeta_{\,s}^+$ by the modified Cayley transformation and denoting by $(X,Y) \mapsto X \circ Y$ the semigroup multiplication on $\zeta^+_{\,s}$, one has $$\label{transferred multiplication} X \circ Y + s = (Y+s)(X+Y)^{-1}(X+s) = X + s - (X-s)(X+Y)^{-1}(X+s) \,.$$ Given Proposition \[prop:semigroup-law\], to prove the transformation rule (\[transformation rule\]) just set $X = a_h$ and $Y = a_{h'}$ and note that, since the semigroup multiplication law for $\widetilde{\mathrm{H}}(W^s)$ by definition yields $a_h \circ a_{h'} = a_{hh'}\,$, the first expression in (\[transferred multiplication\]) implies $$\label{squared transformation law} f(hh') = f(h) f(h') \mathrm{Det}^{-1} (a_h + a_{h'}) \;,$$ where $f(h) = \mathrm{Det}(a_h + s)$ as above. The transformation rule for $\phi$ follows by taking the square root of (\[squared transformation law\]). As usual, the sign of the square root is determined by taking the positive sign at points of the lift of $T_+$ in $\widetilde{\mathrm{H}}(W^s)$. Now we come to the main point. The twisted convolution for $x , y \in \widetilde{\mathrm{H}}(W^s)$ satisfies $\gamma_x \sharp \gamma_y = \gamma_{xy}$. Observe that the phase factor for $w \, , w^\prime \in W_\mathbb{R}$ can be reorganized as $$\mathrm{e}^{\frac{\mathrm{i}}{2} \omega(w,w^\prime)} = \mathrm{e}^{-\frac{1}{2} A(w,w^\prime)} = \mathrm{e}^{\frac{1}{4} \langle w^\prime , \, s w \rangle - \frac{1}{4} \langle w , \, s w^\prime \rangle}\;.$$ Inserting the definitions of $\gamma_x$ and $\gamma_y$ in $\gamma_x \sharp \gamma_y$ we then have $$\gamma_x \sharp \gamma_y (w) = \phi(x)\phi(y) \, \mathrm{e}^{ - \frac{1}{4}\langle w, \, a_x w\rangle}\int \mathrm{e}^{- \frac{1}{4}\langle w^\prime , (a_x + a_y) w^\prime \rangle + \frac{1}{4}\langle w^\prime , (a_x + s) w \rangle + \frac{1}{4} \langle w,(a_x - s)w^\prime \rangle}\,\mathrm{dvol}(w^\prime)\;.$$ Since $\mathfrak{Re} (a_x + a_y) > 0\,$, the integrand is a rapidly decreasing function of $w^\prime \in W_\mathbb{R}$ and convergence of the integral over the domain $W_\mathbb{R}$ is guaranteed. We now wish to explicitly compute the integral by completing the square and shifting variables. For this it is a useful preparation to write $$\langle w^\prime , a_x w \rangle = A(w^\prime , s a_x w) \quad (w^\prime \in W_\mathbb{R}) \;,$$ and similarly for the other terms. We then holomorphically extend the right-hand side to $w^\prime$ in $W$, and by making a shift of integration variables $$w^\prime \to w^\prime + (a_x + a_y)^{-1} (a_x + s) w \;,$$ we bring the convolution integral into the form $$\gamma_x \sharp \gamma_y (w) = \mathrm{e}^{-\frac{1}{4} A(w,\, s (a_x \circ a_y) w)} \phi(x) \phi(y) \int_{W_\mathbb{R}} \mathrm{e}^{-\frac{1}{4} A(w^\prime , \, (s a_x + s a_y) w^\prime)} \, \mathrm{dvol}(w^\prime) \;,$$ where $a_x \circ a_y = -s + (a_y + s)(a_x + a_y)^{-1} (a_x + s) = a_x - (a_x - s)(a_x + a_y)^{-1}(a_x+s)$ is the semigroup multiplication on $\zeta_{\,s}^+$. Using $A(w,s w^\prime) = \langle w , w^\prime \rangle$ for $w \in W_\mathbb{R}$ and the defining relation $a_x \circ a_y = a_{xy}\,$, we see that the first factor on the right-hand side is $\mathrm{e}^{- \frac{1}{4} \langle w , \, a_{xy} w \rangle}$. By the transformation rule (\[transformation rule\]) the integral is evaluated as $$\int_{W_\mathbb{R}} \mathrm{e}^{-\frac{1}{4} \langle w^\prime ,\, (a_x + a_y) w^\prime \rangle}\, \mathrm{dvol}(w^\prime) = \mathrm{Det}^{-1/2}(a_x+a_y) = \frac{\phi(xy)}{\phi(x)\phi(y)}\;,$$ and multiplying factors it follows that $$\gamma_x \sharp \gamma_y(w) = \phi(xy) \, \mathrm{e}^{- \frac{1}{4} \langle w ,\, a_{xy} w \rangle} = \gamma_{xy}(w)\;,$$ which is the desired semigroup property. The mapping $R : \, \widetilde{\mathrm{H}}(W^s) \to \mathrm{End} (\mathcal{A}_V)$ defined by $$R(x) = \int_{W_\mathbb{R}} \gamma_x(w) T_w \,\mathrm{dvol}(w)$$ is a representation of the semigroup $\widetilde{\mathrm{H}}(W^s)$. We conclude this section by deriving a formula for the adjoint. The adjoint of $R(x)$ is computed as $R(x)^\dagger = R(\widetilde{ \psi}(x))$. In particular, $R(x)R(x)^\dagger = R(x\, \widetilde{ \psi}(x))$. We recall the relations $\overline{\phi} = \phi \circ \widetilde{\psi}$ from Proposition \[psi equivariance\] and $(a_h)^\dagger = a_{\psi(h)}$ from Remark \[rem:twine-inv\]. Since $\overline{\langle w , a_h w \rangle} = \langle w , (a_h)^\dagger w \rangle\,$, it follows that $$\overline{\gamma_x} = \gamma_{\widetilde{\psi}(x)} \,.$$ The desired formula, $R(x)^\dagger = R(\widetilde{\psi}(x))$, now results from this equation and the identities $T_w^\dagger = T_{-w}$ and $\gamma_x(-w) = \gamma_x(w)$. With this in hand, the second statement $R(x) R(x)^\dagger = R(x\, \widetilde{\psi} (x))$ is a consequence of the semigroup property. ### Basic conjugation formula Here we compute the effect of conjugating (in the semigroup sense) operators of the form $q(w)$, $w \in W$, with operators $R(x)$ coming from the semigroup. This is an immediate consequence of an analogous result for the operators $T_w\,$. For this we first allow $T_w$ to be defined for $w = v + \varphi \in W$ by $$T_w := \mathrm{e}^{\mathrm{i} q(w)} = \mathrm{e}^{\mathrm{i} \delta(v) + \mathrm{i} \mu(\varphi)} \;.$$ These operators are no longer defined on Fock space, but are defined on $\mathcal O(V)$. They satisfy $$\label{Heisenberg transformation rule} T_w T_{\tilde w} = T_{w + \tilde w} \, \mathrm{e}^{- \frac{1}{2} {A(w,\tilde w)}}\,.$$ Note that for $x \in \widetilde{\mathrm{H}}(W^s)$ and $w \in W$ the operators $R(x) T_w$ and $T_{\tau_H(x) w} R(x)$ are bounded on $\mathcal{A}_V$. Thus we interpret the following as a statement about operators on that space. \[Heisenberg conjugation\] For $w \in W$ and $x \in \widetilde{\mathrm{H}}(W^s)$ one has the relation $$R(x) T_w = T_{\tau_H(x) w} R(x)\,.$$ For convenience of notation we write $$R(x) = \phi(x) \int_{W_\mathbb{R}} \mathrm{e}^{-\frac{1}{4} A(\tilde{w} , \, s a_x \tilde{w})} T_{\tilde{w}}\, \mathrm{dvol} (\tilde{w})\,.$$ Thus $$R(x) T_w = \phi(x) \int_{W_\mathbb{R}}\mathrm{e}^{-\frac{1}{4} A(\tilde{w} ,\,s a_x \tilde{w}) - \frac{1}{2} A(\tilde{w},w)} T_{\tilde{w} + w}\, \mathrm{dvol}(\tilde{w})\;.$$ Now let $h := \tau_H(x)$ and change variables by the translation $\tilde{w} \mapsto \tilde{w} - w + h w\,$. Using the definition $s a_x = (\mathrm{Id}_W + h)(\mathrm{Id}_W - h)^{-1}$ and the relation $$A((\mathrm{Id}_W - h) w_1 ,(\mathrm{Id}_W + h)w_2) = - A((\mathrm{Id}_W + h)w_1, (\mathrm{Id}_W - h)w_2)$$ for all $w_1 , w_2 \in W$, one simplifies the exponent to obtain $$R(x) T_{w} = \phi(x) \int_{W_\mathbb{R}}\mathrm{e}^{-\frac{1}{4} A(\tilde{w} ,\,s a_x \tilde{w}) -\frac{1}{2} A(h w , \tilde{w})} T_{h w + \tilde{w}}\, \mathrm{dvol}(\tilde{w})\;.$$ Reading (\[Heisenberg transformation rule\]) backwards one sees that this expression equals $T_{hw} R(x)$. The basic conjugation rule now follows immediately. \[basic conjugation formula\] For every $x \in \widetilde{\mathrm{H}}(W^s)$ and $w \in W$ it follows that $$R(x) q(w) = q(\tau_H(x)w) R(x)\,.$$ Apply Proposition \[Heisenberg conjugation\] for $w$ replaced by $tw$ and differentiate both sides of the resulting formula at $t=0$. ### Spectral decomposition and operator bounds Numerous properties of $R$ are derived from a precise description of the spectral decomposition of $R(x)$ for $x \in M\,$. Since every orbit of $\mathrm{Sp}_\mathbb{R}$ acting by conjugation on $M$ has nonempty intersection with $T_+\,$, it is important to understand this decomposition when $x \in T_+\,$. For this we begin with the case where $V$ is one-dimensional. \[1-d eigenvalues\] Suppose that $V$ is one-dimensional and that the $T_+$-action on $W = V \oplus V^*$ is given by $x \cdot (v+\varphi ) = \lambda v + \lambda ^{-1}\varphi$ where $\lambda > 1\,$. If $f$ is a basis vector of $V^\ast$ then the monomials $\{f^m \}_{m \in \mathbb{N} \cup \{0\}}$ form a basis of $\mathcal A_V$ and one has $$R(x)f^m = \lambda^{-m-1/2} f^m \;.$$ First note that if $w = v + cv\,$, then $$\label{eigenvalue} a_x w = s \frac{1 + x}{1 - x} \cdot (v+cv) = - \frac{1+\lambda}{1-\lambda}\,v +\frac{1+\lambda^{-1}}{1 - \lambda^{-1}}\,cv = \frac{\lambda+1} {\lambda -1}(v+cv) \;.$$ Thus the Gaussian function $\gamma_x(w)$ in the present case is $$\gamma_x (v + cv) = \phi(x)\, \mathrm{e}^{-\frac{1}{2} \frac{\lambda +1} {\lambda-1} |v|^2} \;.$$ To apply the operator $T_{v+cv}$ to $f^m$ we use the description $$T_{v + cv} = \mathrm{e}^{\mathrm{i}\delta(v) + \mathrm{i}\mu(cv)} = \mathrm{e}^{\mathrm{i}\mu(cv)} \mathrm{e}^{-\frac{1}{2} |v|^2} \mathrm{e}^{\mathrm{i}\delta (v)}\,.$$ Decomposing $T_{v+cv}$ in this way is not allowed on the Fock space, but is allowed if we regard $T_{v+cv}$ as an operator on the full space $\mathcal{O}(V)$ of holomorphic functions. The calculations are now carried out on this larger space. Recall that $\delta(v) f^m = m f(v) f^{m-1}$. From this we obtain the explicit expression $$T_{v+cv} f^{m} = \mathrm{e}^{-\frac{1}{2}\vert v\vert ^2} \mathrm{e}^{\mathrm{i} \mu(cv)} \sum_{l=0}^m \frac{\mathrm{i}^l} {l!}m(m-1)\cdots (m-l+1) f(v)^l f^{m - l}\,.$$ Our goal is to compute $$I := \int_V \mathrm{e}^{-\frac{1}{2}\frac{\lambda+1}{\lambda-1} |v|^2} T_{v+cv} f^m \, \mathrm{dvol}(v) \;,$$ where $\mathrm{dvol}(v)$ corresponds to $\mathrm{dvol}(w)$ by the isomorphism $V \simeq W_\mathbb{R}\,$. Expanding the exponential $\mathrm{e}^{ \mathrm{i} \mu(cv)}$ and using $\mu(cv) f^m = |v|^2 f(v)^{-1} f^{m+1}$, the integral $I$ is a sum of Gaussian expected values of terms of the form $|v|^{2k} f(v)^{-k} f(v)^{l}$. The only terms which survive are those with $k = l\,$. Thus $$\begin{aligned} I&= f^m \sum_{k =0}^m (-1)^k \binom{m}{k} \int_V \frac{|v|^{2k}}{k!} \, \mathrm{e}^{-\frac{\lambda}{\lambda-1} |v|^2}\, \mathrm{dvol}(v) \\ &= 2^{-1} \sum_{k=0}^m (-1)^k \binom{m}{k} \bigg(\frac{\lambda -1}{\lambda}\bigg)^{k+1} f^m = 2^{-1} (1 - \lambda^{-1})\lambda^{-m} f^m \;. \end{aligned}$$ Now $\phi(x)^2 = \mathrm{Det}(a_x+s) = \mathrm{Det}(2s (\mathrm{Id}_W-\tau_H(x))^{-1})= (-2/ (1-\lambda)) (2/(1 - \lambda^{-1}))$, and $\phi(x) = 2 \lambda^{-1/2} (1 - \lambda^{-1})^{-1}$, since we are to take the positive square root at points $x \in T_+\,$. Hence, $R(x) f^m = \phi(x) 2^{-1} (1- \lambda^{-1}) \lambda^{-m} f^m = \lambda^{-m-1/2} f^m$ as claimed. Note that as $x \in T_+$ goes to the unit element (or, equivalently, $\lambda \to 1$), the expression $R(x) f^m$ converges to $f^m$ in the strong sense for all $m \in \mathbb{N} \cup \{0\}\,$. Now let $V$ be of arbitrary dimension and assume that $x \in T_+$ is diagonalized on $W = V \oplus V^*$ in a basis $\{ e_1 , \ldots, e_d , c e_1,\ldots ,c e_d \}$ with eigenvalues $\lambda_1, \ldots \lambda_d , \lambda_1^{-1}, \ldots ,\lambda_d^{-1}$ respectively. Since $x \in T_+$, we have $\lambda_i > 1$ for all $i$. For $f_i := c e_i$ and $m := (m_1,\ldots , m_d)$ we employ the standard multi-index notation $f^m := f_1^{m_1} \cdots f_d^{m_d}$ and $\lambda^m := \lambda_1^{m_1} \cdots \lambda_d^{m_d}$. In this case the multi-dimensional integrals split up into products of one-dimensional integrals. Thus, the following is an immediate consequence of the above. \[eigenvalues in T\_+\] Let $x \in T_+$ be diagonal in a basis $\{ e_i \}$ of $V$ with eigenvalues $\lambda_i$ ($i = 1, \ldots, d$). If $f^m$ is a monomial $f^m \equiv \prod_i (c e_i)^{m_i}$, then $R(x)f^m = \lambda^{-m-1/2} f^m$. One would expect the same result for the spectrum to hold for every conjugate $g T_+ g^{-1}$, and this expectation is indeed borne out. However, in the approach we are going to take here, we first need the existence and basic properties of the oscillator representation of the metaplectic group. The following is a first step in this direction. \[operator bound\] The operator norm function $\mathrm{Mp} \times T_+ \to \mathbb{R}^{>0}$, $(g,t) \mapsto \Vert R(g\,t g^{-1}) \Vert$ is bounded by a continuous $\mathrm{Mp}$-independent function $C(t) < 1\,$. For any $x \in \widetilde{\mathrm{H}}(W^s)$ we already have the bound $\Vert R(x) \Vert \le C(x)$ where $C(x)$ was computed in (\[eq:Schranke\]). That function $C(x)$ clearly is invariant under conjugation $x \mapsto g x g^{-1}$ by $g \in \mathrm{Mp}\,$. Evaluating it for the case of an element $x \equiv t \in T_+$ with eigenvalues $\lambda_i$ one obtains $$C(t) = 2^{\, \dim_\mathbb{C} V} \prod\nolimits_i \left( \lambda_i^{1/2}+\lambda_i^{-1/2} \right)^{-1} \;.$$ The inequality $C(t) < 1$ now follows from the fact that $\lambda_i > 1$ for all $i\,$. Since $R(x)^\dagger = R(\widetilde{\psi}(x))$ and $\Vert R(t g) \Vert^2 = \Vert R(t g)^\dagger R(t g) \Vert = \Vert R(g^{-1} t^2 g) \Vert$, we infer the following estimates. \[general operator bound\] For all $t \in T_+$ and $g \in \mathrm{Mp}$ one has $\Vert R(t g) \Vert < 1$ and $\Vert R(gt) \Vert < 1\,$. Representation of the metaplectic group {#metaplectic rep} --------------------------------------- Recall that we have realized the metaplectic group $\mathrm{Mp}$ in the boundary of the oscillator semigroup $\widetilde{\mathrm{H}}(W^s)$ and that $\widetilde{\mathrm{H}}(W^s)$ contains the lifted manifold $T_+$ in such a way that the neutral element $\mathrm{Id} \in \mathrm{Mp}$ is in its boundary. Here we show that for $x \in T_+$ and $g \in \mathrm{Mp}$ the limit $\lim_{x \to \mathrm{Id}} R(gx)$ is a well-defined unitary operator $R'(g)$ on Fock space and $R' :\, \mathrm{Mp} \to \mathrm{U} (\mathcal{A}_V)$ is a unitary representation. The basic properties of this *oscillator representation* are then used to derive important facts about the semigroup representation $R\,$. Convergence will eventually be discussed in the so-called *bounded strong\* topology* (see [@H2], p. 71). For the moment, however, we shall work with the slightly weaker notion of bounded strong topology where one only requires uniform boundedness and pointwise convergence of the operators themselves (with no mention made of their adjoints). Note that since $\Vert R(gx) \Vert < 1$ by Corollary \[general operator bound\], we need only prove the convergence of $R(gx) f$ on a dense set of functions $f \in \mathcal{A}_V$. Let us begin with $g = \mathrm{Id}\,$. If a sequence $x_n \in T_+$ converges to $\mathrm{Id} \in \mathrm{Mp}\,$, then the sequence $R(x_n)$ converges in the bounded strong topology to the identity operator on Fock space. If $f$ is any $T_+$-eigenfunction, the sequence $R(x_n) f$ converges to $f$ by the explicit description of the spectrum given in Corollary \[eigenvalues in T\_+\]. The statement then follows because the subspace generated by these functions is dense. Using this lemma along with the semigroup property, we now show that the limiting operators exist and are well-defined. \[well–defined\] If $x_n \in T_+$ converges to $\mathrm{Id} \in \mathrm{Mp}\,$, then for every $g \in \mathrm{Mp}$ the sequence of operators $R(g x_n)$ converges pointwise, i.e., $R(g x_n) f \to R^\prime(g) f$ for all $f$ in $\mathcal{A}_V$. The limiting operator $R'(g)$ is independent of the sequence $\{ x_n \}\,$. For any $m , n \in \mathbb{N}$ there exists some $t = t(m,n) \in T_+$ sufficiently near the identity so that $\tilde{x}_m = t^{-1} x_m$ and $\tilde{x}_n = t^{-1} x_n$ are still in $T_+\,$. By the semigroup property $R(g x_n) = R(g t) R(\tilde{x}_n)$ we then have $$\Vert R(g x_m)f - R(g x_n)f \Vert \le \Vert R(g t) \Vert \, \Vert R(\tilde{x}_m)f - R(\tilde{x}_n)f \Vert\,.$$ Letting $t = t(m,n)\to\mathrm{Id}$ it follows from Corollary \[general operator bound\] that $$\Vert R(g x_m)f - R(g x_n)f \Vert \le \Vert R(x_m)f - R(x_n)f \Vert\,.$$ Thus the Cauchy property of $R(x_n)f$ is passed on to $R(g x_n)f$ and therefore the sequence $R(g x_n)f$ converges in the Hilbert space $\mathcal{A}_V$. Let $\lim_{n \to \infty} R(g x_n)f =: R^\prime(g) f$. To show that the limit is well-defined, pick from $T_+$ another sequence $y_n \to \mathrm{Id}\,$, let $\lim_{n \to \infty} R(g y_n)f =: R^{\prime\prime}(g) f\,$, and notice that $\Vert R^\prime(g) f - R^{\prime\prime}(g)f \Vert$ is no bigger than $$\Vert R^\prime(g) f - R(g x_n)f \Vert + \Vert R(g x_n) f - R(g y_n)f \Vert + \Vert R(g y_n)f - R^{\prime\prime}(g) f \Vert \;.$$ Using the same reasoning as above, the middle term is estimated as $$\Vert R(g x_n) f - R(g y_n) f \Vert \le \Vert R(x_n) f - R(y_n) f \Vert \le \Vert R(x_n) f - f \Vert + \Vert R(y_n) f - f \Vert \;.$$ In the limit $n \to \infty$ this yields the desired result $R^\prime (g) = R^{\prime\prime}(g)$. Since $\Vert R(g x_n) \Vert < 1$ the sequence $R(g x_n)$ converges to $R^\prime(g)$ in the bounded strong topology. Such convergence preserves the product of operators, which is to say that if $A_n \to A$ and $B_n \to B$, then $A_n B_n \to AB\,$. Indeed, $$\Vert (A_n B_n - AB) f \Vert \le \Vert A_n (B_n - B)f \Vert + \Vert (A_n - A) B f \Vert \;,$$ and convergence follows from $\Vert A_n \Vert < 1$ and $A_n \to A$, $B_n \to B$. Note in particular that if $R(g x_n) \to R^\prime(g)$ and $R(g^{-1} x_n) \to R^\prime(g^{-1})$ then $R(g x_n) R(g^{-1} x_n) \to R^\prime(g) R^\prime(g^{-1})$. The bounded strong$^*$ topology also requires pointwise convergence of the sequence of adjoint operators. Therefore we must also consider sequences of the form $R(g x_n)^\dagger$. For this (see the proof of Theorem \[thm:Weil-repn\] below) we will use the following fact. Let $\{ A_n \}$ and $\{ B_n \}$ be sequences of bounded operators and let $C_n := A_n B_n\,$. If $C_n$ and $B_n$ converge pointwise with $B_n \to B$ and the sequence $\{ A_n \}$ is uniformly bounded, then $A_n$ converges pointwise on the image of $B\,$. If $f \in \mathrm{im}\, B\,$ then $f = \lim f_n$ where $f_n = B_n h$ for some Hilbert vector $h$. Write $$(A_m - A_n)f = A_m (f - f_m) + (A_m B_m - A_n B_n)h + A_n (f_n - f)$$ and use the uniform boundedness of $A_n$ to show that $A_n f$ converges. Applying this with $A_n = R(g x_n g^{-1})$, $B_n = R(x_n)$ and $C_n = A_n B_n = R(g x_n)R(g^{-1} x_n)$, we have the following statement about convergence along the conjugate $g T_+ g^{-1}$. For $g \in \mathrm{Mp}$ and $\{ x_n \}$ any sequence in $T_+$ with $x_n \to \mathrm{Id} \in \mathrm{Mp}\,$, it follows that $R(g x_n g^{-1})$ converges pointwise to $R'(g) R'(g^{-1})$. Next, if we take three sequences in $T_+$ and write $$\label {other side convergence} R(x_n y_n z_n) = R(x_n g^{-1}) R(g y_n g^{-1}) R(g z_n) \to \mathrm{Id}_{\mathcal{A}_V} \;,$$ then it follows that the sequence $R(x_n g^{-1})$ converges to an operator $B(g^{-1})$ on the image of $R'(g) R'(g^{-1}) R'(g)$ with $B(g^{-1}) R'(g) R'(g^{-1}) R'(g) = \mathrm{Id}_{\mathcal{A}_V}\,$. In particular, the operator $R'(g)$ is injective for all $g \in \mathrm{Mp}\,$. Finally, we define $y_n$ by $y_n^2 = x_n$ and write $R(g x_n) = R(g y_n g^{-1}) R(g y_n)$. Taking the limit of both sides of this equation entails that $$\label{last formal step} R'(g) = R'(g) R'(g^{-1}) R'(g) \;,$$ and since $R'(g)$ is injective, this now allows us to reach the main goal of this section. \[thm:Weil-repn\] For every $g \in \mathrm{Mp}$ and every sequence $\{ x_n \} \subset T_+$ with $x_n \to \mathrm{Id}$ the sequence $\{R( g x_n) \}$ converges in the bounded strong$^*$ topology. The limit $R'(g)$ is independent of the sequence and defines a unitary representation $R':\,\mathrm{Mp} \to \mathrm{U} (\mathcal A_V)$. From (\[last formal step\]) we have $R'(g)(\mathrm{Id}_{ \mathcal{A}_V} - R'(g^{-1}) R'(g)) = 0$ and, since $R'(g)$ is injective, $R'(g^{-1}) R'(g) = \mathrm{Id}_{\mathcal{A}_V}\,$. Hence $R'(g^{-1})$ is surjective, and thus $R'(g) \in \mathrm{GL}(\mathcal{A}_V)$ by exchanging $g \leftrightarrow g^{-1}$. For the homomorphism property we write $R(g_1 x_n) R(g_2 y_n) = R(g_1 x_n g_2 y_n) = R(g_1 x_n g_1^{-1}) R(g_1 g_2 y_n)$ and take limits to obtain $R'(g_1) R'(g_2) = R'(g_1 g_2)$. Convergence in the bounded strong$^*$ topology also requires convergence of the adjoint. This property follows from $R(g x_n)^\dagger = R(\widetilde{\psi}(g x_n)) = R(x_n g^{-1})$ and the discussion after (\[other side convergence\]), since $R'(g)$ is now known to be an isomorphism. Unitarity of the representation is then immediate from $R(g x_n)^\dagger \to R'(g)^\dagger$ and $R(x_n g^{-1}) \to B(g^{-1}) = R'(g)^{-1}$. Finally, we must show that $R' : \, \mathrm{Mp} \to \mathrm{U} (\mathcal{A}_V)$ is continuous. This amounts to showing that if $\{ g_k \}$ is a sequence in $\mathrm{Mp}$ which converges to $g\,$, then $R'(g_k) f \to R'(g) f$ for any $f \in \mathcal{A}_V$. Hence, we let $\{ x_n \}$ be a sequence in $T_+$ with $x_n \to \mathrm{Id}$ and choose $t = t(m,n)$ as in the proof of Proposition \[well–defined\] so that $$\Vert R(g_k x_m) - R(g_k x_n) \Vert \le \Vert R(g_k t) \Vert\, \Vert R(\tilde x_m) - R(\tilde x_n) \Vert \;,$$ and then let $t \to \mathrm{Id}\,$. Using the uniform boundedness of $R(g_k t)$ as $t \to \mathrm{Id}\,$, this shows that the convergence $R(g_k x_n) \to R'(g_k)$ is uniform in $g_k\,$. Since we have $g_k x_n \to g x_n$ for every fixed $n\,$, the continuity of $x \mapsto R(x) f$ then implies that $R'(g_k)f \to R'(g)f$. Let us underline two important consequences. For $g_1 , g_2 \in \mathrm{Mp}$ and $x \in \widetilde{\mathrm{H}} (W^s)$ it follows that $$R(g_1 x\, g_2) = R'(g_1) R(x) R'(g_2) \;.$$ If $y_m$ and $z_n$ are sequences in $T_+$ which converge to $\mathrm{Id}\,$, then, since $x \mapsto R(x) f$ is continuous for all $f$ in Fock space, $R(g_1 y_m x\, g_2 z_n)$ converges pointwise to $R(g_1 x \, g_2)$. On the other hand, we have $R(g_1 y_m x\, g_2 z_n) = R(g_1 y_m) R(x) R(g_2 z_n)$ by the semigroup property, and the right-hand side converges pointwise to $R'(g_1) R(x) R'(g_2)$. We refer to $R' :\, \mathrm{Mp} \to \mathrm{U}(\mathcal{A}_V)$ as the *oscillator* representation of the metaplectic group. It has the following fundamental conjugation property. Let $\mathrm{Mp} \times W \to W$, $(g,w) \mapsto \tau(g) w$ denote the representation of $\mathrm{Mp}$ on $W$ defined by first applying the covering map $\mathrm{Mp} \to \mathrm{Sp}_\mathbb{R}$ and then the standard representation of $\mathrm{Sp}\,$. If we let $W$ act on $\mathfrak{a}(V)$ by the Weyl representation $q$ then $$R'(g) q(w) R'(g)^{-1} = q(\tau(g) w)\,.$$ Since the inverse operator $R'(g)^{-1}$ is now available, this follows from the conjugation property at the semigroup level (see Proposition \[basic conjugation formula\]). Note that analogously we have the classical conjugation formula for the representation of the Heisenberg group on the Fock space $\mathcal{A}_V$, i.e., $$R'(g) T_w R'(g)^{-1} = T_{\tau(g) w}$$ for all $g \in \mathrm{Mp}$ and $w \in W_\mathbb{R}$ (see Proposition \[Heisenberg conjugation\]). ### The trace-class property The concrete formula for the eigenvalues of $R(x)$ which is given in Proposition \[eigenvalues in T\_+\] shows that if $x \in T_+$, then $R(x)$ is of trace class. Using the conjugation property proved above, we now show that this holds for all $x \in \widetilde{\mathrm {H}}(W^s)$. For every $x \in \widetilde{\mathrm{H}}(W^s)$ the operator $R(x)$ is of trace class. We must show that the operator $\sqrt{R(x)R(x)^\dagger}$ has finite trace. To verify this property observe that $R(x) R(x)^\dagger = R(y)$ with $y := x \widetilde{\psi}(x) \in M\,$. Since $y = g t^2 g^{-1}$ for some $t \in T_+$ and $\sqrt{R(g t^2 g^{-1})} = R'(g) R(t) R'(g)^{-1}$, the desired result follows from the explicit formula in Proposition \[eigenvalues in T\_+\] for the eigenvalues of $t\,$. \[prop:osc-char\] For every $x \in \widetilde{\mathrm{H}}(W^s)$ one has $$\mathrm{Tr}\, R(x) = 2^{-\dim_\mathbb{C} V} \phi(x) \;.$$ Since the representation of the Heisenberg group $H$ on $\mathcal{A}_V$ is irreducible, the space of bounded linear operators $\mathrm{End} (\mathcal{A}_V)^H$ that commute with the $H$-action is just $\mathbb{C} \mathrm{Id}_{\mathcal{A}_V}\,$. Now the dual of the space of trace-class operators is the space of bounded linear operators itself, where a bounded linear operator $A$ is realized as a functional on the trace-class operators by $F_A(B) := \mathrm{Tr} \,(AB)$ (see [@RS]). Invariance of $F_A$ by the $H$-representation is equivalent to the invariance of $A$ as an operator. Thus, an $H$-invariant bounded linear functional on the space of trace-class operators is just a multiple of $\mathrm{Tr} \,$. We construct such a functional, $\Phi\,$, by letting $1$ be the vacuum in $\mathcal A_V$ and averaging the expectation in $T_w 1 \in \mathcal{A}_V$ over all $w \in W_\mathbb{R} \,$: $$\label{general trace} \Phi(L) := \int_{W_\mathbb{R}} \left\langle T_w 1 , L \, T_w 1 \right\rangle_{\mathcal{A}_V} \mathrm{dvol}(w) \;.$$ Indeed, the representation of $H$ is unitary and by using $T_w^\dagger = T_w^{-1}$ along with the basic property $T_w T_{w'} = T_{w+w'}\, \mathrm{e}^{\frac{\mathrm{i}} {2} \omega(w,w')}$, one shows that $\Phi$ is invariant, i.e., $\Phi(T_w L T_w^{-1}) = \Phi(L)$ for all $w\in W_\mathbb{R}\,$. If $L:= 1 \langle 1,\cdot \rangle_{\mathcal{A}_V}$ is the projector on the vacuum, then it is a simple matter to show that $$\label{normalization} \mathrm{Tr}_{\mathcal A_V}(L) = \sqrt{2}^{\, \dim W_\mathbb{R}} \Phi(L)\;.$$ Thus, this formula holds for any bounded linear operator $L$ of trace class. Now $\Phi(R(x))$ is computed as $$\Phi(R(x))= \int_{W_\mathbb{R}}\left\langle T_w 1, \int_{W_\mathbb{R}}\gamma_x(w') T_{w'}\, \mathrm{dvol}(w') T_w 1 \right\rangle_{\mathcal{A}_V} \mathrm{dvol}(w)\,.$$ Due to the fact that $\gamma_x(w')$ is rapidly decreasing, we may take the expectation inside the inner integral: $$\Phi(R(x)) = \int_{W_\mathbb{R}} \int_{W_\mathbb{R}} \gamma_x(w^\prime) \left\langle T_w 1, T_{w'} T_w 1 \right\rangle_{\mathcal{A}_V} \mathrm{dvol}(w^\prime) \, \mathrm{dvol}(w)\;.$$ Again using the basic properties of $T_w\,$, we see that $$\left\langle T_w 1 , T_{w'} T_w 1 \right\rangle = \mathrm{e}^{ \mathrm{i}\omega(w^\prime,w)} \langle 1 , T_{w^\prime} 1 \rangle = \mathrm{e}^{\mathrm{i}\omega (w^\prime,w) - \frac{1}{4} \langle w^\prime , w^\prime \rangle} \;.$$ Thus the inner integral over $w^\prime$ is a Gaussian integral, and evaluating it we obtain $$\int_{W_\mathbb{R}} \gamma_x(w^\prime)\, \mathrm{e}^{-\frac{1}{4}\langle w^\prime, w^\prime \rangle + \mathrm{i} \omega(w^\prime,w)}\, \mathrm{dvol}(w^\prime) = \phi(x) \mathrm{Det}^{- \frac{1}{2}}(a_x + \mathrm{Id}_W)\, \mathrm{e}^{-\langle sw,(a_x + \mathrm{Id}_W)^{-1} s w\rangle}\,.$$ Integrating this quantity with respect to $\mathrm{dvol}(w)$ we obtain $$\Phi(R(x)) = 2^{-\dim W_\mathbb{R}} \phi (x) \;,$$ and the desired result for the trace follows from (\[normalization\]) and $\dim W_\mathbb{R} = 2 \dim_\mathbb{C} V$. \[holomorphicity\] For every $P_1 , P_2$ in the Weyl algebra and every $x \in \widetilde{ \mathrm{H}}(W^s)$ the operator $q(P_1) R(x) q(P_2)$ is of trace class on the Fock space $\mathcal A_V$. Furthermore, the function $\widetilde{\mathrm{H}}(W^s)\to \mathbb{C}\,$, $x \mapsto \mathrm{Tr} \, q(P_1) R(x) q(P_2)$, is holomorphic. (Sketch) It is enough to show that $q(P) R(x)$ is of trace class for the case that $P$ is a monomial operator, i.e., $q(P)= \mu( cv)^k \delta(v')^\ell $. We must then compute the trace of the square root of $C(y) = q(P) R(y) q(P)^\dagger$ for $y = x \tilde{\psi}(x) \in M\,$. For this, recall that the $\dagger $-operation interchanges multiplication and differentiation. Direct computation of $| \sqrt{C(y)} f^m |$ shows that this number is of the order of magnitude of $\lambda^{-m} m^{k + \ell}$ where $\lambda$ is the torus element corresponding to $\sqrt{y}\,$. Thus the conjugated operator has effectively the same trace as $R(\sqrt{y})$ itself and in particular $\mathrm{Tr}\, \sqrt{C(y)} < \infty\,$. Turning to holomorphicity, if we insert the equation $L(x) = q(P_1) R(x) q(P_2)$ in the formula (\[general trace\]) for the trace and use the definitions to compute the various integrals, then the function $F(x) = \Phi (L(x))$ is of the form $$F(x) = \int_{W_\mathbb{R}} g(x\, , w)\, \mathrm{dvol}(w) \;,$$ where $g(x\, , \cdot)$ is a rapidly decreasing function on $W_\mathbb{R}$ and $g(\cdot ,w)$ is holomorphic on the semigroup $\widetilde{ \mathrm{H}}(W^s)$. Since the $\bar\partial_x$–operator can be exchanged with the integral, it is immediate that $F$ is holomorphic and therefore so is $\mathrm{Tr}\, q(P_1) R(x) q(P_2)$. Compatibility with Lie algebra representation --------------------------------------------- We now show that the semigroup representation $R : \, \widetilde{ \mathrm{H}}(W^s) \to \mathrm{End}(\mathcal{A}_V)$ is compatible with the $\mathfrak{sp}$-representation $$\mathfrak{sp} \stackrel{\tau^{-1}} {\longrightarrow} \mathfrak{w}(W) \stackrel{q}{\longrightarrow} \mathfrak{gl} (\mathfrak{a}(V)) \;, \quad \mathfrak{a}(V) = \mathrm{S}(V^\ast) \;.$$ Let $h \in \mathrm{H}(W^s)$ and $Y \in \mathfrak{sp}\,$. Then, since the semigroup $\mathrm{H}(W^s)$ is open in $\mathrm{Sp}\,$, there exists some $\varepsilon > 0$ so that the curve $[-\varepsilon , \varepsilon] \ni t \mapsto \mathrm{e}^{tY} h$ lies in $\mathrm{H} (W^s)$. Fix some point $x \in \tau_H^{-1}(h)$ and let $t \mapsto \mathrm{e}^{tY} \cdot x$ denote the lifted curve in $\widetilde{\mathrm{H}}(W^s)$. For all $x \in \widetilde{\mathrm{H}}(W^s)$ and all $Y \in \mathfrak{sp}$ it follows that $$\frac{d}{dt}\bigg\vert_{t = 0} R(\mathrm{e}^{tY} \cdot x) = q (\tau^{-1}(Y)) R(x) \;.$$ Recall that the operator $R(x)$ is the result of integrating the Heisenberg translations $T_w$ against the Gaussian density $\gamma_x(w)\, \mathrm{dvol}(w)$. Thus $$\label{eq:3.12} \frac{d}{dt}\bigg\vert_{t = 0} R(\mathrm{e}^{tY} \cdot x) = \int_{W_\mathbb{R}} \frac{d}{dt}\bigg\vert_{t = 0} \gamma_{\, \mathrm{e}^{tY} \cdot x}(w) T_w \, \mathrm{dvol}(w) \;.$$ For $w_1 , w_2 \in W$ the linear transformation $Y: \, w \mapsto w_1 A(w_2 , w) + w_2 A(w_1 , w)$ is in $\mathfrak{sp}\,$, and $\mathfrak{sp}$ is spanned by such transformations. It is therefore sufficient to prove the statement of the lemma for $Y$ of this form. Hence let $Y := w_1 A(w_2 ,\cdot) + w_2 A(w_1 ,\cdot)$ and observe that the corresponding element in the Weyl algebra is $$\tau^{-1}(Y) = {\textstyle{\frac{1}{2}}} (w_1 w_2 + w_2 w_1) \;.$$ Now, defining $T_w$ for $w \in W$ by $T_w = \mathrm{e}^{\mathrm{i} q(w)}$ as before, we have $$\label{eq:3.13} q(\tau^{-1}(Y)) R(x) = - \frac{d^2}{dt_1 dt_2}\bigg\vert_{ t_1 = t_2 = 0} T_{t_1 w_1 + t_2 w_2} R(x) \;.$$ Therefore, for $\tilde{w} := t_1 w_1 + t_2 w_2$ consider the expression $$T_{\tilde{w}} R(x) = \int_{W_\mathbb{R}} \gamma_x(w) T_{\tilde{w}} T_w \, \mathrm{dvol}(w) \;.$$ Using $T_{\tilde{w}} T_w = \mathrm{e}^{-\frac{1}{2} A(\tilde{w},w)} T_{\tilde{w} + w}$ and shifting integration variables $w \to w - \tilde{w}$ we obtain $$\label{eq:3.14} T_{\tilde{w}} R(x) = \int_{W_\mathbb{R}} \gamma_x(w - \tilde{w})\, \mathrm{e}^{-\frac{1}{2} A(\tilde{w},w)} T_w \, \mathrm{dvol}(w)\;.$$ Comparing Eqs. (\[eq:3.13\],\[eq:3.14\]) with (\[eq:3.12\]) we see that the formula of the lemma is true if $$\frac{d}{dt} \bigg\vert_{t = 0} \gamma_{\,\mathrm{e}^{tY} \cdot x}(w) = - \frac{d^2}{dt_1 dt_2} \bigg\vert_{t_1 = t_2 = 0} \gamma_x(w - t_1 w_1 - t_2 w_2) \, \mathrm{e}^{-\frac{1}{2} A(t_1 w_1 + t_2 w_2 , w)} \;.$$ But checking this equation is just a simple matter of computing derivatives. Recall that $\gamma_x(w) = \phi(x)\, \mathrm{e}^{- \frac{1}{4} A(w,\, sa_x w)}$ and $\phi(x) = \mathrm{Det}^{1/2}(a_x + s)$. Writing $h := \tau_H (x)$ and using $\mathrm{Tr}\, Y=0$ one computes the left-hand side to be $$\frac{d}{dt} \bigg\vert_{t = 0} \gamma_{\,\mathrm{e}^{tY} \cdot x}(w) = \gamma_x(w) \left( {\textstyle{\frac{1}{4}}} \mathrm{Tr}\, (Y s a_x) + {\textstyle{\frac{1}{2}}} A(h(1 - h)^{-1} w , Y h (1 - h)^{-1} w) \right) \;.$$ On substituting $Y = w_1 A(w_2 , \cdot) + w_2 A(w_1 , \cdot)$, this expression immediately agrees with the result of taking the two derivatives on the right-hand side. The extended character ====================== Having prepared the algebraic foundations ($\S$\[sect:osp-howe\]) and the necessary representation-theoretic tools ($\S\ref{integrating})$, we now turn to the main part of our work, which is to prove the formula (\[eq:WeylCharacter\]) for the $K$-Haar expectation value $I(t)$ of a product of ratios of characteristic polynomials. We will achieve this goal by exploiting the fact that $I(t)$ is the same as the character $\chi(t)$ of the irreducible $\mathfrak{g}$-representation on $\mathcal{A}_V^K$. The key property determining $\chi$ is a system of differential equations coming from the Casimir invariants of the Lie superalgebra $\mathfrak{g}\,$: acting as differential operators on a certain supermanifold $\mathcal{F}$ of Lie supergroup type, these invariants annihilate $\chi$ as a section of $\mathcal{F}$. Generalities on Lie supergroup representations ---------------------------------------------- ### Grassmann envelopes For a $\mathbb{Z}_2$-graded complex vector space $V = V_0 \oplus V_1$ let $\mathrm{End}(V) \equiv \mathfrak{gl}(V)$ be the Lie superalgebra with bracket $[X , X^\prime] = X X^\prime - (-1)^{|X| |X^\prime|} X^\prime X$. Then, given any finite-dimensional complex vector space $\mathbb{C}^d$, define the *Grassmann envelope* of $\mathrm{End} (V)$ by the Grassmann algebra $\Omega = \wedge(\mathbb{C}^d) = \Omega_0 \oplus \Omega_1$ as $$\mathrm{End}_\Omega(V) = \oplus_s \left( \Omega_s \otimes \mathrm{End}(V)_s \right) \;, \quad \Omega_s = \oplus_{k \ge 0} \wedge^{2k+s} (\mathbb{C}^d) \;.$$ The Grassmann envelope $\mathrm{End}_\Omega(V)$ is given the structure of an associative algebra by $$(\omega \otimes X)\, (\omega^\prime \otimes X^\prime) := (-1)^{|\omega| |\omega^\prime|} \omega \omega^\prime \otimes X X^\prime \;,$$ and it also has the structure of a Lie algebra by the usual commutator: $$[ (\omega \otimes X) , (\omega^\prime \otimes X^\prime) ] = \omega^\prime \omega \otimes [ X , X^\prime ] \;.$$ More generally, if $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1 \subset \mathrm{End}(V)$ is a complex Lie (sub)superalgebra, the Grassmann envelope $\tilde{\mathfrak{g}} (\Omega)$ of $\mathfrak{g}$ by $\Omega$ is still defined in the same way: $$\tilde{\mathfrak{g}}(\Omega) := \oplus_s (\Omega_s \otimes \mathfrak{g}_s)\;,$$ and it still carries the same structure of associative algebra and Lie algebra. The supertrace $\mathrm{STr} : \, \tilde{\mathfrak{g}} (\Omega) \to \Omega_0$ is defined by $\omega \otimes X \mapsto \omega \, \mathrm{STr}\, X$. Note that this satisfies $\mathrm{STr}\, [\xi , \eta] = 0$ for all $\xi , \eta \in \tilde{\mathfrak{g}}(\Omega)$. The Grassmann envelope $\tilde{\mathfrak{g}}(\Omega)$ is $\mathbb{Z}_2$-graded by $\tilde{\mathfrak{g}} (\Omega) = \tilde{\mathfrak{g}}_0 \oplus \tilde{\mathfrak{g}}_1$ where $\tilde{\mathfrak{g}}_s = \Omega_s \otimes \mathfrak{g}_s\,$. There also exists a $\mathbb{Z}$-grading, which is induced by that of $\Omega$; this is $\tilde{\mathfrak{g}} (\Omega) = \bigoplus_k \tilde{\mathfrak{g}}^{(k)}$ with $\tilde{\mathfrak{g}}^{(2l + s)} = \wedge^{2l+s}(\mathbb{C}^d) \otimes \mathfrak{g}_s\,$. Note that $[\tilde{\mathfrak{g}}^{(k)} , \tilde{\mathfrak{g}}^{(l)}] \subset \tilde{\mathfrak{g} }^{(k + l)}$. Hence, $\mathfrak{n} := \bigoplus_{ k\geq 1} \tilde{\mathfrak{g} }^{(k)}$ is a nilpotent ideal. The subspace $\tilde{\mathfrak{g}}_0$ is a Lie subalgebra and we have $\mathfrak{g}_0 = \tilde{\mathfrak{g}}^{(0)}\subset \tilde{\mathfrak{g}}_0 \,$. The following is an elementary statement found, e.g., in [@B]. Let $x \in \mathrm{End}_\Omega(V)$ and consider the decompositions $x = x_0 + x_1$ and $x = x^{(0)} + \ldots + x^{(d)}$ with respect to the $\mathbb{Z}_2$- and $\mathbb{Z}$-gradings of $\mathrm{End}_\Omega (V)$. The element $x$ is invertible in $\mathrm{End}_\Omega(V)$ if and only if $x_0$ is invertible, and the latter is the case if and only if $x^{(0)} \in \mathrm{GL}(V_0)\times \mathrm{GL}(V_1)$. Let $G(\Omega)$ be any Lie group such that $\mathrm{Lie}(G(\Omega)) = \tilde{\mathfrak{g}}(\Omega)$. Following [@B] we write $$\begin{aligned} &&G = \exp (\mathfrak{g}_0) \subset G(\Omega)\;, \quad G(\Omega)_0 = \exp (\tilde{\mathfrak{g}}_0) \subset G(\Omega) \;, \\ &&G(\Omega)_1 = \exp (\tilde{\mathfrak{g}}_1)\;, \quad N(\Omega) = \exp(\mathfrak{n}) \;, \quad N(\Omega)_0 = \exp(\mathfrak{n}\cap \tilde{\mathfrak{g}}_0) \;.\end{aligned}$$ The objects $G\,$, $G(\Omega)_0\,$, $N(\Omega)$, and $N(\Omega)_0$ are subgroups of $G(\Omega)$. The group $N(\Omega)$ is a normal subgroup of $G(\Omega)$. Here is another elementary result from [@B]: Each element $\tilde{g} \in G(\Omega)$ has a unique factorization of the form $\tilde{g} = g \theta_1 = \theta_2\, g$ with $g \in G(\Omega)_0$ and $\theta_1 , \theta_2 \in G(\Omega)_1\,$. Each element $g \in G(\Omega)_0$ can be uniquely represented in the form $g = g_0\, n_1 = n_2\, g_0$ with $g_0 \in G$ and $n_1\, , n_2 \in N(\Omega)_0\,$. ### Lie supergroups A complex [*Lie supergroup*]{} is a pair $(\mathfrak{g}\, , G)$ consisting of a complex Lie superalgebra $\mathfrak{g} = \mathfrak{g}_0 \oplus \mathfrak{g}_1$ and a complex Lie group $G$ that exponentiates $\mathfrak{g}_0\,$, i.e., $\mathrm{Lie}(G) = \mathfrak{g}_0\,$. Given a Lie superalgebra $\mathfrak{g}$ there exist, in general, several choices of $G$. If $\mathfrak{g} \to \mathrm{End}(V)$ is a faithful finite-dimensional representation of $\mathfrak{g}\,$, i.e., $\mathfrak{g} \subset \mathrm{End}(V)$, then one choice is $G := \exp({\mathfrak{g}_0}) \subset \mathrm{GL}(V_0) \times \mathrm{GL}(V_1) \subset \mathrm{End}(V)_0\,$. A [*superfunction*]{} on a complex Lie supergroup $(\mathfrak{g}\, ,G)$ is a section in the sheaf $\mathcal{F} \equiv \mathcal{F}(G ,\mathfrak{g}_1)$ of germs of holomorphic functions on $G$ with values in $\wedge (\mathfrak{g}_1^*)$. This sheaf $\mathcal{F}$ is a locally free sheaf of $\mathcal{O}_G$-modules where $\mathcal{O}_G$ is the sheaf of germs of holomorphic functions on $G$. The gradings on $\wedge (\mathfrak{g}_1^*)$ give a $\mathbb{Z}$-grading $\mathcal{F} =\bigoplus_k \mathcal{F}^{(k)}$ and a $\mathbb{Z}_2$-grading $\mathcal{F} = \mathcal{F}_0 \oplus \mathcal{F}_1$. If $f = f^{(0)} + \ldots + f^{(d_1)}$ is the decomposition of a superfunction $f$ with respect to the former, then $\mathrm{num}(f) := f^{(0)}$ is called the [*numerical part*]{} of $f\,$. A [*superderivation*]{} $D \in \mathrm{Der}\, \mathcal{F}$ is a linear sheaf morphism $D : \, \mathcal{F} \to \mathcal{F}$ satisfying the graded Leibniz rule $D(fh) = (Df)h + (-1)^{|D| |f|}f(Dh)$ for $f , h \in \mathcal{F}$. The $\mathbb{Z}_2$-grading of $\mathcal{F}$ defines a $\mathbb{Z}_2$-grading on the vector space $\mathrm{Der}\, \mathcal{F}$ by $|D| = s$ if $D \mathcal{F}_{t} \subset \mathcal{F}_{s+t}\,$. ### Representation of $\mathfrak{g}$ by superderivations In the theory of ordinary Lie groups one has a realization of the Lie algebra by left- or right-invariant vector fields acting as derivations of the algebra of differentiable functions on the Lie group. In the same spirit, one wants to construct a representation of the Lie superalgebra $\mathfrak{g}$ by superderivations on the sheaf of algebras $\mathcal{F}$. The basic tool for this construction is a correspondence between superfunctions $f \in \mathcal{F}$ and functions on $G(\Omega)$ with values in $\Omega\,$. There exists no canonical correspondence of such kind; one possible choice is as follows. Let $\{E_1 , \ldots, E_{d_0}\}$ be a basis of $\mathfrak{g}_0\,$, and let $\{ F_1 , \ldots, F_{d_1} \}$ be a basis of $\mathfrak{g}_1$ with dual basis $\xi_1 , \ldots, \xi_{d_1}$ of $\mathfrak{g}_1^\ast$. Then if $X \in \mathfrak{g}_0$ and $\hat{X}^R$ denotes the corresponding left-invariant vector field on functions $\varphi$ on $G$, $$(\hat{X}^R \varphi)(g) := \frac{d}{dt} \varphi(g\, \mathrm{e}^{tX}) \big\vert_{t=0} \;,$$ we assign to a section $f = \sum_J f_J\, \xi_{j_1} \wedge \cdots \wedge\xi_{j_k} \in \mathcal{F}$ a function $\Phi_f : \, G(\Omega) \to \Omega$ by $$\Phi_f(g\, \mathrm{e}^{\sum \alpha_i \otimes E_i} \mathrm{e}^{\sum \beta_j \otimes F_j}) = \sum\nolimits_J (\mathrm{e}^{\sum \alpha_i \hat{E}_i^R} f_J)(g)\, \beta_{j_1} \wedge \cdots \wedge \beta_{j_k} \;,$$ where $\alpha_i \in \Omega_0$ and $\beta_j \in \Omega_1\,$. The assignment $f\mapsto \Phi_f$ is injective if $\Omega = \wedge ({\mathbb C}^d)$ and $d \geq d_1\,$; and is so in particular if $\mathbb{C}^d = \mathfrak{g}_1^*\,$. In that case, the inverse mapping $\Phi_f \mapsto f$ is given by restriction to elements $\tilde{g} = g \, \mathrm{e}^{\sum \xi_j \otimes F_j}$ with $g \in G\,$. The advantage of passing to the right-hand side of the correspondence $f \leftrightarrow \Phi_f$ is that functions on $G(\Omega)$ can be shifted by elements of $G(\Omega)$ by multiplication from the left or right. This possibility now puts us in a position to construct the desired representation of the Lie superalgebra $\mathfrak{g}$ by superderivations on the sheaf $\mathcal{F}(G,\mathfrak{g}_1)$. Consider first the case $X \in \mathfrak{g}_0\,$. Making the choice $\Omega := \wedge( \mathfrak{g}_1^\ast)$ and linearizing the function $\mathbb{R} \times G(\Omega) \to \Omega\,$, $(t,\tilde{g}) \mapsto \Psi(\mathrm {e}^{-tX} \tilde{g})$ with respect to the parameter of the one-parameter group $t \mapsto \mathrm{e}^{tX}$, we obtain a first-order differential operator $\tilde{X}^L$: $$\Psi(\mathrm{e}^{-tX} \tilde{g}) = \Psi(\tilde{g}) + (t\, \tilde{X}^L \Psi)(\tilde{g}) + \mathcal{O}(t^2) \;.$$ Similarly, $\tilde{X}^R$ is defined by $\Psi(\tilde{g}\, \mathrm {e}^{tX}) = \Psi(\tilde{g}) + (t\,\tilde{X}^R \Psi)(\tilde{g}) + \mathcal{O}(t^2)$. Derivations $\hat{X}^i$ on $\mathcal{F}$ are then given by $\hat{X}^i := \Phi^{-1} \circ \tilde{X}^i \circ \Phi$, i.e., by the equation $\Phi_{\hat{X}^i f} = \tilde{X}^i \Phi_f$ ($i = L\, , R$). To handle the case $X \in \mathfrak{g}_1\,$, we need one extra anti-commuting parameter for the purpose of differentiation. Denoting this parameter by $\tau$ we choose $\Omega := \wedge(\mathfrak{g}_1 ^\ast \oplus \mathbb{C}\tau)$ and define $\tilde{X}^{L,R}$ by $\Psi( \mathrm{e}^{-\tau \otimes X} \tilde{g}) = \Psi(\tilde{g}) + (\tau\, \tilde{X}^L \Psi)(\tilde{g})$ and $\Psi(\tilde{g}\, \mathrm{e}^{ \tau \otimes X}) = \Psi(\tilde{g}) + (\tau\,\tilde{X}^R \Psi)(\tilde{g})$. Again, the corresponding derivations on $\mathcal{F}$ are $\hat{X}^{L ,R} := \Phi^{-1} \circ \tilde{X}^{L,R} \circ \Phi$. On basic grounds [@B] one then has the following statement. The assignments $X\mapsto\hat{X}^L$ and $X\mapsto\hat{X}^R$ are representations of the Lie superalgebra $\mathfrak{g}$ on the sheaf $\mathcal{F}(G,\mathfrak{g}_1)$ of superfunctions. The two representations commute in the graded-commutative sense, i.e., $\hat{X}^L \hat{Y}^R = (-1)^{|X| |Y|} \hat{Y}^R \hat{X}^L$. ### Representations and characters of Lie supergroups {#supergroup characters} A *representation* of a Lie supergroup $(\mathfrak{g},G)$ on a $\mathbb{Z}_2$-graded vector space $V = V_0 \oplus V_1$ is a homomorphism of Lie superalgebras $\rho_* :\, \mathfrak{g} \to \mathrm{End}(V)$ and a homomorphism of Lie groups $\rho :\, G \to \mathrm{GL}(V_0) \times \mathrm{GL}(V_1) \subset \mathrm{GL}(V)$ which are compatible in the sense that $(d\rho)_e = \rho_* |_{\mathfrak{g}_0}\,$. \[example:superrepGL(U)\] Given a $\mathbb{Z}_2$-graded vector space $U = U_0 \oplus U_1\,$, consider the Lie superalgebra $\mathfrak{gl}(U)$ and the Lie supergroup $(\mathfrak{gl}(U) , \mathrm{GL}(U_0) \times \mathrm{GL}(U_1))$. For any integer $m$ the two maps $\rho :\, \mathrm{GL}(U_0) \times \mathrm{GL}(U_1) \to \mathbb{C}^\times$, $(g_0\, , g_1) \mapsto \mathrm{Det}^m(g_0) \, \mathrm{Det}^{-m} (g_1)$, and $\rho_* : \mathfrak{gl}(U) \to \mathbb{C}\,$, $X \mapsto m\, \mathrm{STr} X\,$, define a representation of $(\mathfrak{gl}(U), \mathrm{GL}(U_0) \times \mathrm{GL}(U_1))$ on the $\mathbb{Z}_2$-graded one-dimensional vector space $V_0 \oplus V_1$ with $V_0 = \mathbb{C}$ and $V_1 = \{0\}\,$. The [*character*]{} $\chi$ associated with the representation $(\rho_*,\rho;V)$ of a Lie supergroup $(\mathfrak{g},G)$ is the superfunction $$\chi : \, G \to \wedge(\mathfrak{g}_1^*)\;, \quad g \mapsto \mathrm{STr}_V\, \rho(g)\, \mathrm{e}^{\sum \xi_j\,\rho_*(F_j)}\;,$$ provided that the supertrace exists. Note that the numerical part $\mathrm{num}(\chi)$ of the character is just the supertrace of the Lie group representation $\rho : \, G \to \mathrm{GL}(V_0) \times \mathrm{GL}(V_1)$. An element $f \in \mathcal{F}(G,\mathfrak{g}_1)$ is called [*radial*]{} if the associated function $\Phi_f : \, G(\Omega) \to \Omega$ satisfies $\Phi_f (g^{-1} h g) = \Phi_f (h)$ for all $g , h \in G(\Omega)$ and if $(\hat{X}^L + \hat{X}^R) f = 0$ for all $X \in \mathfrak{g}_1\,$. The character of a Lie supergroup representation is radial. Let $\chi$ be the character associated with the representation ($\rho_\ast , \rho ; V)$ of a Lie supergroup $(\mathfrak{g} , G)$. Then by the definition of the correspondence $\chi \leftrightarrow \Phi_\chi$ one has $$\Phi_\chi(g_0\, \mathrm{e}^{\sum \alpha_i \otimes E_i} \mathrm{e}^{\sum \beta_j \otimes F_j}) = \mathrm{STr}_V \, \rho(g_0)\, \mathrm{e}^{\sum \alpha_i\, \rho_*(E_i)} \mathrm{e}^{\sum \beta_j\, \rho_*(F_j)} \;,$$ where the notation above ($g_0 \in G$, $\alpha_i \in \Omega_0$, $\beta_j \in \Omega_1$) is being employed. This satisfies $\Phi_\chi(h) = \Phi_\chi(g^{-1} h g)$, since $\mathrm{STr}\, (XY) = \mathrm{STr}\, (YX)$ for $X, Y \in \mathrm{End}_\Omega(V)$ and the representations $\rho_\ast$ and $\rho$ are compatible. For $X \in \mathfrak{g}_1$ the infinitesimal (or linearized) version of the same argument gives $(\tilde{X}^L + \tilde{X}^R ) \Phi_\chi = 0$ and hence $(\hat{X}^L + \hat{X}^R) \chi = 0\,$. Character of the spinor-oscillator representation {#subsec:existence} ------------------------------------------------- In Chapter \[integrating\] we constructed a semigroup representation $R_0 :\, \widetilde{\mathrm{H}}(W_0^s) \to \mathrm{End}( \mathcal{A}_V)$ and a group representation $R' :\, \mathrm{Mp}\to \mathrm{U} (\mathcal{A}_V)$ which exponentiate the restriction of the spinor-oscillator representation $\rho_*$ to $\mathfrak{sp} \subset \mathfrak{osp}_0\,$. These are compatible in the sense that $\widetilde{ \mathrm{H}}(W_0^s) = \mathrm{Mp} \times M$, $\mathrm{Mp}$ acts on $M$ by conjugation, and $R_0(g_1 x \, g_2) = R'(g_1) R_0(x) R'(g_2)$ for all $g_1 , g_2 \in \mathrm{Mp}$ and $x \in \widetilde{\mathrm{H}}(W_0^s)$. The representation $\rho_*|_{\mathfrak{o}} : \, \mathfrak{o} \subset \mathfrak{osp}_0 \to \mathfrak{gl}({\mathcal A}_V)$ exponentiates to a Lie group representation $R_1 :\, \mathrm{Spin}(W_1) \to \mathrm{GL}({\mathcal A}_V)$. Note that $R_1$ and $R_0$ commute, as they act on the two different factors of the tensor product $\mathfrak{a}(V) = \wedge (V_1^*) \otimes \mathrm{S}(V_0^*)$. Note also that the group $\mathbb{Z}_2$ acts on $\widetilde{\mathrm {H}}(W_0^s)$ and $\mathrm{Spin}(W_1)$ by deck transformations of the $2:1$ coverings $\widetilde{\mathrm{H}}(W_0^s) \to \mathrm{H}(W_0^s)$ and $\mathrm{Spin}(W_1) \to \mathrm{SO}(W_1)$. The non-trivial element of $\mathbb{Z}_2$ is represented by a sign change, $R_0 \to - R_0$ and $R_1 \to - R_1$. ### Spinor-oscillator character as a radial superfunction {#sect:4.2.1} Given the representations $R_1$ and $R_0\,$, we form the semigroup representation $$R\,:\,\, \mathrm{Spin}(W_1)\times_{\mathbb{Z}_2}\widetilde{\mathrm{H}} (W_0^s) \to \mathrm{End}(\mathcal{A}_V)\;, \quad (g_1 , g_0) \mapsto R_1(g_1) R_0(g_0) \;.$$ We already know the representation $R$ to be compatible with $\rho_* : \, \mathfrak{osp} \to \mathfrak{gl} (\mathcal{A}_V)$. Now we define a superfunction $\gamma$ on $\mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{ \mathrm{H}} (W_0^s)$ with values in $\wedge (\mathfrak{osp}_1^\ast)$ by $$\gamma(g_1,g_0) = \mathrm{STr}_{\mathcal{A}_V}\, R(g_1,g_0)\, \mathrm{e}^{\sum \xi_j\, \rho_*(F_j)} \;,$$ where $\{ F_1, \ldots, F_{d_1} \}$ is a basis of $\mathfrak{osp}_1$ and $\{ \xi_1, \ldots, \xi_{d_1}\}$ is its dual basis. We refer to $\gamma$ as the spinor-oscillator character. By the circumstance that $R$ and $\rho_*$ are representations on the infinite-dimensional vector space $\mathcal{A}_V$ we are obliged to discuss the domain of definition of $\gamma\,$. For this, notice that $\mathrm{Spin}(W_1)$ acts non-trivially only on the finite-dimensional (or spinor) part of $\mathcal{A}_V\,$. Expanding $\mathrm{e}^{\sum \xi_j\, \rho_* (F_j)}$ we obtain a finite sum $$\mathrm{e}^{\sum \xi_j\, \rho_*(F_j)} = \sum\nolimits_J \rho_*(P_J) \xi_{j_1} \cdots \xi_{j_k}$$ with $P_J \in \mathfrak{q}(W)$. (Here we recall that $\mathfrak{osp}_1$ can be viewed as part of the Clifford-Weyl algebra $\mathfrak{q}(W)$ by the isomorphism $\tau^{-1}:\, \mathfrak{osp} \to \mathfrak{s} \subset \mathfrak{q}(W)$.) By Proposition \[holomorphicity\] the operator $R(g) \rho_* (P_J)$ is of trace class. Thus the character $\gamma$ is defined on the full domain $\mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{ \mathrm{H}} (W_0^s)$. Moreover, $\gamma(g)$ depends analytically on $g\,$. Does there exist any good sense in which to think of the spinor-oscillator character $\gamma$ as a radial superfunction? There is no question that for every $X \in \mathfrak{osp}$ we do have superderivations $\hat{X}^L$ and $\hat{X}^R$ on $\mathcal{F}(\mathrm{Spin} \times \widetilde{\mathrm{H}})$, and the function $\gamma$ by its definition as a character does satisfy $(\hat{X}^L + \hat{X}^R)\chi = 0\,$. However, our semigroup elements $g \in \mathrm{Spin}(W_1) \times_{ \mathbb{Z}_2} \widetilde{ \mathrm{H}} (W_0^s)$ do not possess an inverse in the spinor-oscillator representation and we therefore shouldn’t expect such a relation as $\Phi_\chi(g^{-1} h g) = \Phi_\chi(h)$. A substitute is this. Let $\mathrm{Spin}_\mathbb{R} \subset \mathrm{Spin}(W_1)$ be the compact real subgroup which is obtained by exponentiating the skew-symmetric degree-two elements of the Clifford algebra generated by a Euclidean vector space $W_{1,\mathbb{R}} \subset W_1\,$. Let $G_\mathbb{R} := \mathrm{Spin}_\mathbb{R} \times_{\mathbb{Z}_2} \mathrm{Mp}\,$. Then on the real submanifold $\mathrm{Spin}_\mathbb{R} \times M_\mathrm{Sp}$ (with $M_\mathrm{Sp}$ formerly denoted by $M$) the superfunction $\gamma$ is $G_\mathbb{R}$-radial, since the representations $R'$ of $\mathrm{Mp}$ and $R_0$ on $M_\mathrm{Sp} \subset \widetilde{\mathrm{H}}(W_0^s)$ are compatible. As a result, we will later be able to apply Berezin’s theory of radial parts of Laplace-Casimir operators. ### Numerical part of the character Here we show that the restriction of the numerical part of $\gamma$ to a toral set $T_1 \times_{\mathbb{Z}_2} T_+$ in $\mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{ \mathrm{H}} (W_0^s)$ gives the autocorrelation function described in $\S$\[introduction\]. That is, we show that the ratios of characteristic polynomials $\mathrm{Det} (\mathrm{Id}_N - \mathrm{e}^{-\phi} k)$ and their averages with respect to a compact group $K$ (with $k \in K \subset \mathrm{U}_N$ and $\phi$ being a parameter), can be expressed as supertraces of operators on the spinor-oscillator module $\mathfrak{a}(V)$ and the submodule of $K$-invariants $\mathfrak{a}(V)^K$. We begin with a summary of the relevant facts. From Proposition \[basic conjugation formula\] we recall the basic conjugation rule for the oscillator representation: $$R_0(x) q(w) = q(\tau_H(x) w) R_0(x) \;,$$ where $x \in \widetilde{ \mathrm{H}}(W_0^s)$ , $w \in W_0\,$, and $\tau_H(x) \in \mathrm{H}(W_0^s) \subset \mathrm{Sp}(W_0)$. On the side of the spinor representation, the corresponding conjugation formula is $$R_1(y) q(w) = q(\tau_S(y)\, w) R_1(y) \;, \quad w \in W_1 \;, \ y \in \mathrm{Spin}(W_1) \;.$$ This defines the 2:1 covering homomorphism $\mathrm{Spin}(W_1) \to \mathrm{SO}(W_1)$, $y \mapsto \tau_S(y)$, which exponentiates the isomorphism of Lie algebras $\tau :\, \mathfrak{s} \cap \mathfrak{c}_2(W_1) \to \mathfrak{o}(W_1)$. Recall also (from Proposition \[prop:osc-char\]) the formula for the oscillator character: $$\mathrm{Tr}\, R_0(x) = \phi(x) \;, \quad \phi(x)^2 = \mathrm{Det}^{-1}(\mathrm{Id}_{W_0} - \tau_H(x)) \;.$$ An analogous result is known for the case of the spinor representation; see, e.g., the textbook [@BGV]. Defining the spinor character as the supertrace with respect to the canonical $\mathbb{Z}_2$-grading of the spinor module, one has $$\mathrm{STr}\, R_1(y) = \psi(y) \;, \quad \psi(y)^2 = \mathrm{Det}(\mathrm{Id}_{W_1} - \tau_S(y)) \;.$$ Thus the spinor character, just like the oscillator character, is a square root. By taking the supertrace over the total Fock representation space, we obtain the formula $$\label{eq:spin-osc-char} \mathrm{STr}_{\mathcal{A}_V} R_1(y) R_0(x) = \phi(x) \psi(y) =: \sqrt{ \frac{\mathrm{Det}(\mathrm{Id}_{W_1} - \tau_S(y))} {\mathrm{Det}(\mathrm{Id}_{W_0} - \tau_H(x))}} \;.$$ For $W_0 = W_1\,$, the case of our interest, $\mathrm{Spin} (W_0)$ intersects with $\widetilde{\mathrm{H}} (W_0^s)$ and the square root $\psi(y)$ is defined in such a way that $\psi(x) = \phi(x)^{-1}$ for $x \in \mathrm{Spin}(W_0) \cap \widetilde{\mathrm{H}} (W_0^s)$. Let now $U = U_0 \oplus U_1$ be a $\mathbb{Z}_2$-graded vector space, and let $V = U \otimes \mathbb{C}^N$. The Lie group $\mathrm{GL}(V_1) \times \mathrm{GL}(V_0)$ acts on $W = (V_1^{\vphantom{\ast}} \oplus V_1^\ast) \oplus (V_0^{\vphantom{\ast}} \oplus V_0^\ast)$ by $$(g_1,g_0) . (v_1 \oplus \varphi_1 \oplus v_0 \oplus \varphi_0) = (g_1 v_1) \oplus (\varphi_1 \circ g_1^{-1}) \oplus (g_0 v_0) \oplus (\varphi_0 \circ g_0^{-1}) \;.$$ This action serves to realize the group $G := (\mathrm{GL}(U_1) \times \mathrm{GL}(U_0)) \times_{\mathbb{C}^\times} \mathrm{GL} (\mathbb{C}^N)$ as a subgroup of $\mathrm {GL}(V_1) \times \mathrm{GL}(V_0) \subset \mathrm{SO}(W_1) \times \mathrm{Sp}(W_0)$. For the purpose of letting $G$ act in the spinor-oscillator module $\mathfrak{a}(V)$, let this representation be lifted to that of a double covering $\tilde{G} \hookrightarrow \mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{\mathrm{H}} (W_0^s)$. The following statement gives the value of the spinor-oscillator character on $(t_1,t_0;g) \in G$ where $g \in \mathrm{GL}(\mathbb{C}^N)$ and $t_s = \mathrm{diag}(t_{s,1}, \ldots, t_{s,\,n})$ are diagonal matrices in $\mathrm{GL}(U_s)$ (for $s = 0, 1$). \[lem:STrAV\] If $\dim U_0 = \dim U_1 = n$ and $| t_{0,j} | > 1$ for all $j = 1, \ldots, n\,$, then $$\mathrm{STr}_{\mathfrak{a}(V)} R(t_1,t_0;g) = \prod_{j=1}^n \sqrt{ \frac{t_{1,j}}{t_{0,j}}}^{\,N}\frac{\mathrm{Det} (\mathrm{Id}_N -(t_{1,j}\, g)^{-1})}{\mathrm{Det}(\mathrm{Id}_N - (t_{0,j}\, g)^{-1})} \;.$$ Since $t_1$ and $t_0$ are assumed to be of diagonal form, the statement holds true for the case of general $n$ if it does so for the special case $n = 1$. Hence let $n = 1$. In that case $t_1$ and $t_0$ are single numbers and $t_s g$ acts on $W_s = V_s^{\vphantom{\ast}} \oplus V_s^\ast \simeq \mathbb{C}^N \oplus (\mathbb{C}^N)^\ast$ as $(t_s g).(v \oplus \varphi) = (t_s g v) \oplus \varphi \circ (t_s g)^{-1}$ for $s = 0,1$. From equation (\[eq:spin-osc-char\]) we then have $$\mathrm{STr}_{\mathfrak{a}(V)} R(t_1,t_0;g) = \sqrt{\frac{ \mathrm{Det} (\mathrm{Id}_N - t_1 g)\,\mathrm{Det}(\mathrm{Id}_N - (t_1 g)^{-1})} {\mathrm{Det}(\mathrm{Id}_N-t_0 g)\,\mathrm{Det} (\mathrm{Id}_N-(t_0 g)^{-1})}} \;,$$ which turns into the stated formula on pulling out a factor of $\mathrm{Det}(-t_1 g) / \mathrm{Det}(- t_0 g)$ from under the square root. (Of course, the double covering of $\mathrm{GL}(U_1) \times \mathrm{GL}(U_0)$ is to be used in order to define this square root globally.) In the formula of Lemma \[lem:STrAV\] we now set $t_{1,j} = \mathrm {e}^{\mathrm{i}\psi_j}$ and $t_{0,j} = \mathrm{e}^{\phi_j}$. We then put $g^{-1} \equiv k \in K$ and integrate against Haar measure $dk$ of unit mass on $K$. This integral and the summation that defines the supertrace can be interchanged, as $\mathrm{STr}_{\mathfrak{a}(V)} R(t_1,t_0; k^{-1})$ is a finite sum of power series and the conditions $\mathfrak{Re} \, \phi_j > 0$ ensure uniform and absolute convergence. The representation of $K$ on $\mathfrak{a}(V)$ is induced by the representation of $K$ on $V$. Therefore, averaging over $K$ with respect to Haar measure has the effect of projecting from $\mathfrak {a}(V)$ to the $K$-trivial isotypic component $\mathfrak{a}(V)^K$, and we arrive at $$\label{eq:4.3mrz} \mathrm{STr}_{\mathfrak{a}(V)^K} R(t_1,t_0;\mathrm{Id}) = \mathrm{e}^{ (N/2) \sum_j (\mathrm{i}\psi_j - \phi_j)}\, \int\limits_K \prod_{j=1}^n \frac{\mathrm{Det}(\mathrm{Id}_N - \mathrm{e}^{- \mathrm{i}\psi_j}\, k)}{\mathrm{Det}( \mathrm{Id}_N - \mathrm{e}^{- \phi_j}\, k)} \, dk \;.$$ In the case of an even dimension $N$, the domain of definition of this formula is a complex torus $T := T_1 \times T_+$ where $T_1 = (\mathbb{C}^\times)^n$ and $T_+ \subset (\mathbb{C}^\times)^n$ is the open subset determined by the conditions $|t_{0,j} | = \mathrm{e}^{\mathfrak{Re}\, \phi_j} > 1$ for all $j\,$. For odd $N$ we must continue to work with a double cover (also denoted by $T$) to take the square root $\mathrm{e}^{ (N/2) \sum_j (\mathrm{i}\psi_j - \phi_j)}$. Let $\mathfrak{g}$ be the Howe dual partner of $\mathrm{Lie}(K)$ in $\mathfrak{osp}(W)$. We know from Proposition \[prop:dualpairs\] that $\mathfrak{g} = \mathfrak{osp}(U \oplus U^\ast)$ for $K = \mathrm{O}_N$ and $\mathfrak{g} = \mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast)$ for $K = \mathrm{USp}_N\,$. Recall also from $\S$\[subsubsec:WoHdP\] that the $\mathfrak{g}$-representation on $\mathfrak{a}(V)^K$ is irreducible and of highest weight $\lambda_N = (N/2) \sum_j (\mathrm{i}\psi_j - \phi_j)$. Denote by $\Gamma_\lambda$ the set of weights of this representation. Let $B_\gamma = (-1)^{ |\gamma|} \dim {\mathfrak{a}(V)^K}_\gamma$ be the dimension of the weight space ${\mathfrak{a}(V)^K}_\gamma$ multiplied with the correct sign to form the supertrace. \[cor:weightexpansion\] On the torus $T$ we have $$\sum_{\gamma \in \Gamma_\lambda} B_\gamma \, \mathrm{e}^\gamma = \mathrm{e}^{\lambda_N} \int\limits_K \prod_{j=1}^n \frac{ \mathrm{Det}(\mathrm{Id}_N - \mathrm{e}^{-\mathrm{i}\psi_j} \, k)} {\mathrm{Det}(\mathrm{Id}_N - \mathrm{e}^{-\phi_j}\, k)}\, dk \;.$$ On the right-hand side we recognize the correlation function (see $\S$\[introduction\]) which is the object of our study and, as we have explained, is related to the character of the irreducible $\mathfrak{g}$-representation on $\mathfrak{a}(V)^K$. The left-hand side gives this character (restricted to the toral set $T$) in the form of a weight expansion, some information about which has already been provided by Corollary \[cor:weights\] of $\S$\[subsubsec:WoHdP\]. Formula for the character ------------------------- Being an eigenfunction of (the radial parts of) the Laplace-Casimir operators that represent the center of the universal enveloping algebra $\mathsf{U}(\mathfrak{g})$, our character $\chi = \sum_\gamma B_\gamma\, \mathrm{e}^\gamma$ satisfies a certain system of differential equations. Here we first describe the origin and explicit form of these differential equations. We then prove that $\chi$ is determined uniquely by these in combination with the weight constraints for $\gamma \in \Gamma_\lambda\,$. Finally, we provide the explicit function with these properties. ### Extended character Formula (\[eq:4.3mrz\]) and Corollary \[cor:weightexpansion\] express the character $\chi$ as a function on the toral set $T$. Our next step is to describe a supermanifold with a $\mathfrak{g}$-action where $\chi$ exists as a $G_\mathbb{R}$-radial superfunction and Laplace-Casimir operators can be applied. Here we give only a rough sketch, leaving the details to the reader. First of all, the symmetry group $G_\mathbb {R}$ for $\chi$ has to be identified. Recall from the end of $\S$\[sect:4.2.1\] that the good real group acting in the spinor-oscillator representation $\mathcal{A}_V$ is $$\mathrm{Spin}(W_{1,\mathbb{R}}) \times_{\mathbb{Z}_2} \mathrm{Mp}(W_{0,\mathbb{R}}) =: G^\prime \;,$$ which contains $K = \mathrm{O}_N$ and $K = \mathrm{USp}_N$ as subgroups. Since we are studying the character $\chi$ of the $\mathfrak{g}$-representation on the subspace $\mathcal{A}_V^K$ of $K$-invariants, we now seek the subgroup $G_\mathbb{R} \subset G^\prime$ which centralizes $K$; this means that we are asking the exponentiated version of a question which was answered at the Lie algebra level in $\S$\[sect:2.7\]. Here, restricting the group $G^\prime$ to the centralizer of $K$ we find $$G_\mathbb{R} = \left\{ \begin{array}{ll} \mathrm{Spin}( (U_1^{ \vphantom{\ast}}\oplus U_1^\ast)_\mathbb{R})\times_{\mathbb{Z}_2} \mathrm{Mp}( (U_0^{\vphantom{\ast}} \oplus U_0^\ast)_\mathbb{R}) &\quad K = \mathrm{O}_N \;, \\ \mathrm{USp}(U_1^{\vphantom{\ast}} \oplus U_1^\ast) \times \mathrm{SO}^\ast(U_0^{\vphantom{\ast}} \oplus U_0^\ast) &\quad K = \mathrm{USp}_N\;.\end{array}\right.$$ We observe that $G_\mathbb{R}$ for the case of $K = \mathrm{O}_N$ is just the lower-dimensional copy of $G^\prime$ which corresponds to $U_s$ taking the role of $V_s\,$. We also see immediately that the Lie algebras $\mathrm{Lie}(G_\mathbb{R})$ coincide with the real forms described in Propositions \[prop:2.4\] and \[prop:2.5\]. The second object to construct is a real domain $M_\mathbb{R} \subset \mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{ \mathrm{H}} (W_0^s)$ with a $G_\mathbb{R}$-action by conjugation such that $M_\mathbb{R} = G_\mathbb{R}.T_\mathbb{R}$ where $T_\mathbb{R} \subset T$ is a real sub-torus. The choice we make for $T_\mathbb{R}$ is the one singled out by the parametrization $t_{1,j} = \mathrm{e}^{ \mathrm{i} \psi_j}$ and $t_{0,j} = \mathrm{e}^{\phi_j}$ with real-valued coordinates $\psi_j$ and $\phi_j\,$. We also want the elements of $M_\mathbb{R}$ to commute with those of $K$ as endomorphisms of $W$. By these requirements, the good real domain $M_\mathbb{R}$ to consider in the case of $K = \mathrm{O}_N$ is the lower-dimensional copy of $\mathrm{Spin}_\mathbb{R} \times M_\mathrm{Sp}$ which, again, corresponds to $U_s$ replacing $V_s\,$. By the detailed analysis of $\S$\[integrating\] (where $M_\mathrm{Sp}$ was simply denoted by $M$) and the fact that the diagonal elements constitute a maximal torus in $\mathrm{Spin} _\mathbb{R}\,$, we infer the desired property $M_\mathbb{R} = G_\mathbb{R} . T_\mathbb{R}\,$. In the case of $K = \mathrm{USp}_N$ the same requirements lead to the choice $$M_\mathbb{R} = \mathrm{USp}(U_1^{\vphantom{\ast}} \oplus U_1^\ast) \times M_\mathrm{SO} \subset \mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{ \mathrm{H}}(W_0^s) \;,$$ where the construction of $M_\mathrm{SO}$ is fully parallel to that of $M_\mathrm{Sp}\,$: if $\mathrm{SO}$ denotes the complex orthogonal group of the vector space $U_0^{\vphantom{\ast}} \oplus U_0^\ast\,$, we introduce the semigroup $\mathrm{H} \subset \mathrm{SO}$ which is defined by the inequality $g^\dagger s g < g$ for $s = - \mathrm{Id} _{U_0} \oplus \mathrm{Id}_{U_0^\ast}$ and then take $M_\mathrm{SO} \subset \mathrm{H}$ to be the totally real submanifold $M_\mathrm{SO} \subset \mathrm{H}$ of pseudo-Hermitian elements $m = s m^\dagger s\,$. For this choice we easily check that $M_\mathbb{R} = G_\mathbb{R} . T_\mathbb{R}\,$. Having constructed a manifold $M_\mathbb{R}$ with an action of $G_\mathbb{R}$ by conjugation, we now consider the supermanifold $\mathcal{F}(M_\mathbb{R}\, , \mathfrak{g}_1)$ which is the sheaf of algebras of analytic functions on $M_\mathbb{R}$ with values in $\wedge(\mathfrak{g}_1^\ast)$ where, once again, we remind the reader that $\mathfrak{g} = \mathfrak{osp}(U \oplus U^\ast)$ for $K = \mathrm{O}_N$ and $\mathfrak{g} = \mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^\ast)$ for $K = \mathrm{USp}_N\,$. By construction, for every point $x \in M_\mathbb{R}$ we have $T_x M_\mathbb{R} \otimes \mathbb{C} \simeq \mathfrak{g}_0\,$, the even part of the Lie superalgebra $\mathfrak{g}\,$. Hence by the basic principles reviewed at the beginning of this Chapter, the supermanifold $\mathcal{F} (M_\mathbb{R}\, , \mathfrak{g}_1)$ carries two representations of $\mathfrak{g}$ by superderivations ($X \mapsto \hat{X}^L$ and $X \mapsto \hat{X}^R$). The benefit from all this is that for every Casimir invariant $I \in \mathsf{U}(\mathfrak{g})$ we now get a Laplace-Casimir operator $D(I)$ on $\mathcal{F}$ by replacing each $X \in \mathfrak{g}$ in the polynomial expression for $I$ by the corresponding differential operator $\hat{X}^L$ (or $\hat{X}^R)$. Utilizing the present setting the character $\chi$, which was given in (\[eq:4.3mrz\]) as a function on $T$, will now be extended to a section of $\mathcal{F}(M_\mathbb{R},\mathfrak{g}_1)$. By construction, the direct product $M_\mathbb{R} \times K$ is contained as a subspace in $\mathrm{Spin}(W_1) \times_{\mathbb{Z}_2} \widetilde{\mathrm{H}}(W_0^s)$, and by restriction of the representation $R$ we get a mapping $R : \, M_\mathbb{R} \times K \to \mathrm{End}(\mathcal{A}_V)$ whose image still lies in the subspace of trace-class operators. A good definition of $\chi \in \mathcal{F}$ therefore is $$\chi(m) := \mathrm{STr}_{\mathcal{A}_V^K}\, R(m;\mathrm{Id}) \, \mathrm{e}^{\sum \xi_j \, \rho_\ast(F_j)} \;,$$ where $\{ F_j \}$ now is a basis of $\mathfrak{g}_1$ and $\{ \xi_j \}$ is its dual basis. The symbol $\rho_\ast$ here denotes the $\mathfrak{g}$-representation on $\mathcal{A}_V^K$. It is clear that by restricting the numerical part of $\chi(m)$ to the toral set $T_\mathbb{R} \subset M_\mathbb{R}$ we recover the function described in (\[eq:4.3mrz\]). ### Weyl group {#sect:WeylGroup} Consider now the numerical part of the section $\chi \in \mathcal{F} $. Being a $G_\mathbb{R}$-radial function, $\mathrm{num}(\chi)$ is invariant under the action of the Weyl group $W$ which normalizes $T_\mathbb{R}$ with respect to the $G_\mathbb{R}$-action on $M_\mathbb{R}$ by conjugation. Since the latter group action decomposes as a direct product of two factors, so does $W$. For both cases ($K = \mathrm{O}_N$, $\mathrm{USp}_N$) the second factor of the Weyl group $W$ is just the permutation group $\mathrm{S}_n\,$. As a matter of fact, conjugation of a diagonal element $t_0 \in M_\mathrm{Sp}$ or $t_0 \in M_\mathrm{SO}$ by $g \in \mathrm{Mp}((U_0^{\vphantom{\ast}} \oplus U_0^\ast)_\mathbb{R})$ or $g \in \mathrm{SO}^\ast( U_0^{ \vphantom{\ast}} \oplus U_0^\ast)$ can return another diagonal element only by permutation of the eigenvalues $\mathrm{e}^{ \phi_1} , \ldots, \mathrm{e}^{\phi_n}$ of $t_0\,$. (No inversion $\mathrm{e}^{ \phi_j} \to \mathrm{e}^{- \phi_j}$ is possible, as this would mean transgressing the oscillator semigroup.) This factor $\mathrm{S}_n$ of $W$ will play no important role in the following, as the expressions we will encounter are automatically invariant under such permutations. The first factors of $W$ are of greater significance. For the two cases of $K = \mathrm{O}_N$ and $K = \mathrm{USp}_N$ these are the Weyl groups $W_{\mathrm{SO}_{2n}}$ and $W_{\mathrm{Sp}_{2n}}$ respectively. An explicit description of these groups is as follows. Let $\{ e_1 , \ldots , e_n \}$ be an orthonormal basis of $U$ and decompose $U \oplus U^*$ into a direct sum of 2-planes, $$U \oplus U^* = P_1 \oplus \ldots \oplus P_n \;,$$ where $P_j$ is spanned by the vector $e_j$ and the linear form $c e_j = \langle e_j \, , \cdot \rangle$ ($j = 1, \ldots,n$). In both cases at hand, i.e., for the symmetric form $S$ as well the alternating form $A$, this is an orthogonal decomposition. The real torus under consideration is parameterized by $(\mathrm{e}^{\mathrm{i}\psi_1} , \ldots , \mathrm{e}^{\mathrm{i}\psi_n}) \in (\mathrm{U}_1)^n$ acting by $\mathrm{e}^{\mathrm{i}\psi_j} . (e_j) = \mathrm{e}^{\mathrm{i} \psi_j} \, e_j$ and $\mathrm{e}^{\mathrm{i} \psi_j}. (c e_j) = \mathrm{e}^{-\mathrm{i}\psi_j} ce_j\,$. The Weyl group $W_{\mathrm{Sp}}$ is generated by the permutations of these planes and the involutions which are defined by conjugation by the mapping that sends $e_j \mapsto c e_j$ and $c e_j \mapsto - e_j\,$. The Weyl group $W_\mathrm{SO}$ is generated by the permutations together with the involutions which are induced by the mappings that simply exchange $e_j$ with $c e_j\,$. Since we are in the special orthogonal group and the determinant for a single exchange $e_j \leftrightarrow c e_j$ is $-1$, the number of involutions in any word in $W_\mathrm{SO}$ has to be even. In summary, the $W$-action on our standard bases of linear functions, $\{ \mathrm{i}\psi_j\}$ and $\{\phi_j \}$, is given by the respective permutations together with the action of the involutions defined by sign change, $\mathrm{i} \psi_j \mapsto - \mathrm{i} \psi_j\,$. In the sequel, the Weyl group action will be understood to be either this standard action or alternatively, depending on the context, the corresponding action on the exponentiated functions $\{ \mathrm{e}^{\mathrm{i} \psi_j} \}$ and $\{ \mathrm{e}^{\phi_j} \}$. As a final remark on the subject, let us note that the Weyl group symmetries of the function $\chi(t)$ can also be read off directly from the explicit expression (\[eq:4.3mrz\]). In particular, the absence of reflections $\phi_j \to -\phi_j$ is clear from the conditions $\mathfrak{Re}\, \phi_j > 0\,$. ### Laplace-Casimir eigenvalues {#sect:CasEigenvalues} We are now ready to start deriving a system of differential equations for $\chi$. Recall that every Casimir element $I \in \mathsf{U} (\mathfrak{g})$ determines an invariant differential operator $D(I)$, called a Laplace-Casimir operator, by replacing each element $X \in \mathfrak{g}$ in the polynomial expression of $I$ by $\hat{X}^L$ or $\hat{X}^R$. Let $\chi$ be the character of an irreducible representation $(\rho_* , \rho)$ of a Lie supergroup $(\mathfrak{g}\, ,G)$ on a complex vector space. Then for any Casimir element $I \in \mathsf{U} (\mathfrak{g})$ the character $\chi$ is an eigenfunction of the differential operator $D(I)$. We have $(D(I)\chi)(g) = \mathrm{STr} \, \rho_*(I) \rho(g)\, \mathrm{e}^{ \sum_j \xi_j\, \rho_*(F_j)}$. Since $I$ is a Casimir element, $\rho_*(I)$ commutes with $\rho_*(X)$ for all $X \in \mathfrak{g}\,$. By Schur’s lemma for trace-class operators, $\rho_*(I) = \lambda(I) \mathrm{Id}$ with $\lambda(I) \in \mathbb{C}\,$. Hence, $D(I) \chi = \lambda(I) \chi\,$. Recall now from $\S$\[sect:osp-cas\] that for every $\ell \in \mathbb{N}$ we have a Casimir element $I_\ell \in \mathsf{U} (\mathfrak {osp})$ of degree $2\ell\,$. Recall also that under the assumption $V_0 \simeq V_1$ we introduced $\partial , \widetilde{\partial} \in \mathfrak{osp}_1\, $, $\Lambda = [ \partial , \widetilde{\partial} ] \in \mathfrak{osp}_0 \,$, and $F_\ell \in \mathsf{U} (\mathfrak{osp})$ such that $I_\ell = [\partial , F_\ell]$ and $[\partial , \Lambda] = 0\,$. Let us insert here the following comment. While $\partial , \widetilde{\partial}, \Lambda$ were abstractly defined as $\mathfrak{osp}$-generators, they acquire a transparent meaning when represented as operators on the spinor-oscillator module $\mathfrak{a}(V)$. In fact, using formula (\[eq:tau-inv\]) one finds $$\partial = \sum\nolimits_j \varepsilon(f_{1,j})\, \delta(e_{0,j}) \;, \quad \widetilde{\partial} = \sum\nolimits_j \mu(f_{0,j})\, \iota(e_{1,j}) \;.$$ Here, for notational brevity, we make no distinction between $\mathfrak{osp}$-elements $\partial$, $\widetilde{\partial}$, etc.and the operators representing them on $\mathfrak{a}(V)$. It is now natural to identify the spinor-oscillator module $\mathfrak{a}(V)$ (or rather, a suitable completion thereof) with the complex of holomorphic differential forms on $V_0\,$. Writing $\varepsilon (f_{1,j}) \equiv dz_j$ and $\delta(e_{0,j}) \equiv \partial / \partial z_j$ we then see that $\partial$ is the holomorphic exterior derivative: $$\partial = \sum\nolimits_j dz_j\,\frac{\partial}{\partial z_j}\;.$$ $\widetilde{\partial} = \iota(v)$ becomes the operator of contraction with the vector field $v = \sum z_j \, \partial / \partial z_j$ generating scale transformations $z_j \to \mathrm{e}^t z_j\,$, while $\Lambda = [\partial , \widetilde{\partial}] = \partial \circ \iota(v) + \iota(v) \circ \partial$ is interpreted as the Lie derivative with respect to that vector field. Consider now any irreducible $\mathfrak{osp}$-representation on a $\mathbb{Z}_2$-graded vector space $\mathcal{V}$ with the property that the $\mathcal{V}$-supertrace of $\mathrm{e}^{-t \Lambda}$ ($t > 0$) exists. Let $\lambda(I_\ell)$ be the scalar value of the Casimir invariant $I_\ell$ in the representation $\mathcal{V}$. Then a short computation using $I_\ell = [ \partial , F_\ell ]$ and $[\partial , \Lambda] = 0$ shows that $\lambda(I_\ell)$ multiplied by $\mathrm{STr}_\mathcal{V}\, \mathrm{e}^{-t \Lambda}$ vanishes: $$\lambda(I_\ell)\, \mathrm{STr}_\mathcal{V}\, \mathrm{e}^{-t \Lambda} = \mathrm{STr}_\mathcal{V}\, \mathrm{e}^{-t \Lambda} I_\ell = \mathrm{STr}_\mathcal{V}\, \mathrm{e}^{-t \Lambda} [\partial , F_\ell] = \mathrm{STr}_\mathcal{V}\, [\partial , \mathrm{e}^{-t \Lambda} F_\ell] = 0 \;,$$ since the supertrace of any bracket is zero. Thus we are facing a dichotomy: either we have $\mathrm{STr}_\mathcal{V}\, \mathrm {e}^{ -t \Lambda} = 0\,$, or else $\lambda(I_\ell) = 0$ for all $\ell \in \mathbb{N}\,$. Now our representation $\mathfrak{a} (V)^K$ realizes the latter alternative, which leads to the following consequence. \[lem:D(Il)chi=0\] Assuming that $U = U_0 \oplus U_1$ is a $\mathbb{Z}_2$-graded vector space with $U_0 \simeq U_1\,$, let $\chi \in \mathcal{F} (M_\mathbb{R}\, , \mathfrak{g}_1)$ be the character of the irreducible $\mathfrak{g} $-representation on $\mathfrak{a}(V)^K$ for $V = U \otimes \mathbb{C}^N$. Then $D(I_\ell)\chi = 0$ for all $\ell \in \mathbb{N}\,$. The condition $U_0 \simeq U_1$ is needed in order for the formula $I_\ell = [ \partial , F_\ell]$ of Lemma \[lem:Cas-exact\] to be available. For any real parameter $t > 0$ the supertrace of the operator $\mathrm{e}^{- t \Lambda}$ on $\mathfrak{a}(V)^K$ certainly exists and is non-zero. In fact, using formula (\[eq:4.3mrz\]) one computes the value as $$\mathrm{STr}_{\mathfrak{a}(V)^K}\, \mathrm{e}^{-t \Lambda} = \mathrm{STr}\, R(\mathrm{e}^t \mathrm{Id}_{n}\,, \mathrm{e}^t \mathrm{Id}_n\,; \mathrm{Id}_N)= \int_K \frac{\mathrm{Det}^n (\mathrm{Id}_N-\mathrm{e}^{-t}k)} {\mathrm{Det}^n (\mathrm{Id}_N - \mathrm{e}^{-t}k)}\, dk = 1 \not= 0 \;.$$ The dichotomy of $\lambda(I_\ell)\, \mathrm{STr}_{\mathfrak{a} (V)^K}\, \mathrm{e}^{-t\Lambda} = 0$ therefore gives $D(I_\ell) \chi = \lambda(I_\ell) \chi = 0\,$. ### Differential equations for the character {#differential equations} Since the Casimir elements commute with all elements of the Lie superalgebra, the Laplace-Casimir operators leave the set of radial superfunctions invariant. We denote by $\dot{D}(I_\ell)$ the radial parts of the Laplace-Casimir operators $D(I_\ell)$ for $\ell \in \mathbb{N}\,$. These operators, which arise by restricting the Laplace-Casimir operators to the space of radial functions, are given by differential operators on the torus $T_\mathbb{R}\,$. They have been described by Berezin [@B], and his results will now be stated in a form well adapted to our purposes. Recall from $\S$\[sect:osp-roots\] that $\mathfrak{osp}$-roots are of the form $\pm \vartheta_{s j} \pm \vartheta_{t k}\,$. Recall also the relations $\vartheta_{0j} = \phi_j$ and $\vartheta_{1j} = \mathrm{i}\psi_j$ ($j = 1, \ldots, n$). In the following we regard the variables $\phi_j$ and $\psi_j$ as real (local) coordinates for the real torus $T_\mathbb{R}\,$. For $\ell \in \mathbb{N}$ let $D_\ell$ be the degree-$2\ell$ differential operator $$D_\ell = \sum_{j=1}^n \frac{\partial^{2\ell}}{\partial \psi_j^{2\ell}}- (-1)^\ell \sum_{j=1}^n \frac{\partial^{2\ell}} {\partial \phi_j^{2\ell}} \;.$$ Let $J$ be the function defined on a dense open subset of $T_\mathbb{R}$ by $$J(t) = \frac{\prod_{\alpha\in\Delta_0^+} 2\sinh\frac{\alpha(\ln t)} {2}} {\prod_{\beta\in\Delta_1^+} 2 \sinh\frac{\beta(\ln t)}{2}} \;,$$ where $\Delta^+ = \Delta_0^+ \cup \Delta_1^+$ is a system of even and odd positive roots (cf. $\S$\[sect:simple-roots\]). Up to a constant multiple the radial part of the differential operator $D(I_\ell)$ is of the form $\dot{D}(I_\ell) = J^{-1}( D_\ell + Q_{\ell-1})\circ J$ where $Q_{\ell-1}$ is some polynomial (with constant coefficients) in the operators $D_k$ of total degree less than that of $D_\ell\,$. The first statement is a direct consequence of Theorem 3.2 of [@B] using the explicit form of the Casimir elements $I_\ell$ defined in equation (\[eq:CasimirElement\]) and the relations $\vartheta_{1j} = \mathrm{i}\psi_j$ and $\vartheta_{0j} = \phi_j\,$. The statement about the lower-order terms $Q_{\ell-1}$ is from Theorem 4.5 of [@B]. \[cor:DlJchi\] The function $\chi : \, T_\mathbb{R} \to \mathbb{C}$ given by $\chi(t) = \mathrm{STr}_{\mathfrak{a}(V)^K}\, R(t;\mathrm{Id}_N)$ satisfies the system of differential equations $D_\ell (J \chi) = 0$ for all $\ell \in \mathbb{N}\,$. From Lemma \[lem:D(Il)chi=0\] we have $D(I_\ell) \chi = 0$ and hence $\dot{D}(I_\ell) \chi |_{T_\mathbb{R}} = 0$ for all $\ell \in \mathbb{N}\,$. It follows by induction that $D_\ell (J\chi) = 0$ for all $\ell \in \mathbb{N}\,$. The methods of this section can be used to derive differential equations for the character of a certain class of irreducible representations of $\mathfrak{gl}(U) \simeq \mathfrak{g}^{(0)}$. Define $$J_0 = \frac{\prod_{j < k} 4 \sinh\frac{\mathrm{i}(\psi_j - \psi_k)}{2} \, \sinh\frac{\phi_j - \phi_k}{2}}{\prod_{j,\,k} 2 \sinh\frac{\phi_j - \mathrm{i} \psi_k}{2}} \;.$$ Here, $\{ \mathrm{i}(\psi_j - \psi_k), \phi_j - \phi_k \mid j < k \}$ and $\{ \phi_j - \mathrm{i} \psi_k \}$ are the sets of even and odd positive roots of $\mathfrak{g}^{(0)}$. The following statement is Corollary 4.12 of [@CFZ] adapted to the present context and notation. The idea of the proof is the same as that of Lemma \[lem:D(Il)chi=0\]. \[cor:4.4\] Let $\chi$ be the character of an irreducible representation of the Lie supergroup $(\mathfrak{gl}(U),\mathrm{GL}(U_0) \times \mathrm{GL}(U_1))$ on a finite-dimensional $\mathbb{Z}_2$-graded vector space $V = V_0 \oplus V_1\,$. If $U_0 \simeq U_1$ but $\mathrm{dim}(V_0) \not= \mathrm{dim}(V_1)$, then $D_\ell (J_0\chi) = 0$ for all $\ell \in \mathbb{N}\,$. ### Uniqueness theorem {#unicity theorem} Recall that the main goal of this paper can be stated as that of explicitly computing the character $\chi$ of a local representation defined on the space of invariants $\mathfrak{a}(V)^K$ in the spinor-oscillator module. We have restricted ourselves to the cases where $K$ is either $\mathrm{O}_N$ or $\mathrm{USp}_N\,$. The representation on $\mathfrak{a}(V)^K$ is defined at the infinitesimal level on the full complex Lie superalgebra $\mathfrak{g}$ which is the Howe partner of $K$ in the canonical realization of $\mathfrak{osp}$ in the Clifford-Weyl algebra of $V \oplus V^\ast$. It has been shown that $\chi : \, T_\mathbb{R} \to \mathbb{C}$ satisfies the differential equations $D_\ell (J \chi) = 0$. By analytic continuation it satisfies the same equations on the complex torus $T$. Recall that $\Gamma_\lambda$ denotes the set of weights of the $\mathfrak{g}$-representation on $\mathfrak{a}(V)^K$. Recall also from Corollary \[cor:weights\] that the weights $\gamma = \sum_{j = 1}^n (\mathrm{i} m_j \psi_j - n_j \phi_j) \in \Gamma_\lambda$ satisfy the weight constraints $- \frac{N}{2} \le m_j \le \frac{N}{2} \le n_j\,$. The highest weight is $\lambda = \frac{N}{2} \sum (\mathrm{i} \psi_j - \phi_j)$. By the definition of the torus $T$ the weights $\gamma \in \Gamma_\lambda$ are analytically integrable and we now view $\mathrm{e}^\gamma$ as a function on $T$. \[unicity\] The character $\chi : \, T \to \mathbb{C}$ is annihilated by the differential operators $D_\ell \circ J$ for all $\ell \in \mathbb{N}\,$, and it has a convergent expansion $\chi = \sum B_\gamma \, \mathrm{e}^{\gamma}$ where the sum runs over weights $\gamma = \sum_{j = 1}^n (\mathrm{i} m_j \psi_j - n_j \phi_j)$ satisfying the constraints $- \frac{N}{2} \le m_j \le \frac{N}{2} \le n_j\,$. For the case of $K = \mathrm{USp}_N$ it is the unique function on $T$ with these two properties and $B_\lambda = 1$. For $K = \mathrm{O}_N$ it is the unique $W$-invariant function on $T$ with these two properties and $B_\lambda = 1\,$, $B_{\lambda - \mathrm{i}N\psi_n} = 0\,$. To verify the property $B_{\lambda - \mathrm{i}N\psi_n} = 0$ which holds for the case of $K = \mathrm{O}_N\,$, look at the right-hand side of the formula of Corollary \[cor:weightexpansion\]: in order to generate a term $\mathrm{e}^\gamma = \mathrm{e}^{\lambda - \mathrm{i}N\psi_n}$ in the weight expansion, you must pick the term $\mathrm{e}^{- \mathrm{i} N\psi_n}$ in the expansion of the determinant for $j = n$ in the numerator; but the latter term depends on $k$ as $\mathrm{Det}(-k)$ which vanishes upon taking the Haar average for $K = \mathrm{O}_N \,$. By $W$-invariance the property $B_{\lambda - \mathrm{i}N\psi_n} = 0$ is equivalent to $B_{\lambda - \mathrm{i} N\psi_j} = 0$ for all $j\,$. In view of this Remark and Corollaries \[cor:weights\] and \[cor:DlJchi\], it is only the uniqueness statement of Theorem \[unicity\] that remains to be proved here. This requires a bit of preparation, in particular to appropriately formulate the condition $D_\ell (J\chi) = 0\,$. For that we develop $J \chi$ in a series $J \chi = \sum_\tau a_\tau f_\tau$ where the $f_\tau$ are $D_\ell$-eigenfunctions for *every* $\ell \in \mathbb{N}\,$. The first step is to determine an appropriate expansion for $J$. Recall that $$J = \frac{\prod_{\alpha \in \Delta ^+_0} (\mathrm{e}^{ \frac{\alpha}{2}}- \mathrm{e}^{- \frac{\alpha}{2}})} {\prod_{\beta \in \Delta ^+_1} (\mathrm{e}^{\frac{\beta}{2}} - \mathrm{e}^{-\frac{\beta}{2}})} \;.$$ Given a factor in the denominator of this representation, we wish to factor out, e.g., $\mathrm{e}^{-\frac{ \beta}{2}}$ to obtain a term $(1 - \mathrm{e}^{-\beta})^{-1}$ which we will attempt to develop in a geometric series. In order for this to converge uniformly on compact subsets of $T$ it is necessary and sufficient for $\mathfrak{Re}\, \beta$ to be positive on $\mathfrak{t}\,$. This of course depends on the root $\beta\,$. Fortunately, the sets of odd positive roots for our two cases of $K= \mathrm{O}_N$ and $K= \mathrm{USp}_N$ are the same (see $\S$\[sect:simple-roots\]): $$\Delta_1^+ =\{\phi_j\pm\mathrm{i}\psi_k \mid j\,,k=1,\ldots,n\}\;.$$ So indeed, if we factor out $\mathrm{e}^{-\frac{\beta}{2}}$ from each term in the denominator and do the same in the numerator, we obtain the expression $$J = \mathrm{e}^{\delta}\, \frac{\prod_{\alpha \in \Delta^+_0} (1 - \mathrm{e}^{-\alpha})} {\prod_{\beta \in \Delta ^+_1} {(1 - \mathrm{e}^{-\beta})}} \;,$$ and it is possible to expand each term of the denominator in a geometric series. Here $$\delta = \frac{1}{2}\sum_{\alpha \in \Delta ^+_0}\alpha - \frac{1}{2} \sum_{\beta \in \Delta^+_1} \beta$$ is half the *supersum* of the positive roots. Now let $\{\sigma_1, \ldots , \sigma _r\}$ be a basis of simple positive roots (cf. $\S$\[sect:simple-roots\]) and expand the terms $(1- \mathrm{e}^{-\beta})^{-1}$ in geometric series to obtain $$J = \mathrm{e}^{\delta} \, \sum_{b\ge 0} A_b \, \mathrm{e}^{b\sigma} \;,$$ which converges uniformly on compact subsets of $T$. In this expression $b$ and $\sigma$ denote the vectors $b = (b_1, \ldots , b_r)$ and $\sigma = (\sigma_1 , \ldots , \sigma_r)$, respectively, and $b \sigma := \sum b_i \sigma_i\,$. Following the usual multi-index notation, $b\ge 0$ means $b_i \ge 0$ for all $i\,$. Note $A_0 = 1$. Now we know that the character has a convergent series representation $$\chi = \sum _{\gamma \in \Gamma _\lambda} B_\gamma \, \mathrm{e}^{\gamma} \;.$$ Thus we may write $$J\chi = \sum_{\gamma \in \Gamma_\lambda} B_\gamma \sum_{b\ge 0} A_b \,\mathrm{e}^{\delta + \gamma + b\sigma}\;.$$ For convenience of the discussion we let $\tilde{\gamma} := \gamma + b \sigma$ and reorganize the sums as $$\label{ready for recursion} J\chi =\sum _{\tilde \gamma} \Big( \sum A_b\, B_{\tilde{\gamma} -b \sigma} \Big)\, \mathrm{e}^{\delta + \tilde\gamma} \;,$$ where the inner sum is a finite sum which runs over all $b\ge 0$ such that $\tilde{\gamma}- b\sigma \in \Gamma_\lambda\,$. We are now in a position to explain the recursion procedure which shows that $\chi$ is unique. Start by applying $D_\ell $ to $J\chi $ as represented in the expression (\[ready for recursion\]). Since $\delta +\tilde\gamma$ is of the form $\sum (\mathrm{i} m_k \psi _k - n_k \phi_k)$, we immediately see that it is an eigenfunction with eigenvalue $E(\ell ,\tilde \gamma) := (-1)^\ell \sum (m_k^{2\ell} - n_k^{2\ell})$. The functions $\mathrm{e}^{\delta +\tilde \gamma}$ in the sum are independent eigenfunctions. Hence it follows that $$\label{recursion equations} 0 = E(\ell ,\tilde\gamma) \sum A_b \, B_{\tilde\gamma - b\sigma}$$ for all $\tilde\gamma$ fixed and then for all $\ell\in\mathbb{N}\,$. From now on we consider the equations (\[recursion equations\]) only in those cases where $\tilde\gamma$ is itself a weight of our representation. (We have license to do so as only the uniqueness part of Theorem \[unicity\] remains to be proved.) In this case we have the following key fact. \[vanishing eigenvalues\] If $\gamma \in \Gamma_\lambda$ and the eigenvalue $E(\ell , \gamma)$ vanishes for all $\ell \in \mathbb{N}\,$, then $\gamma$ is the highest weight $\lambda$. Our first job is to compute $\delta$. For the list of even and odd positive roots we refer the reader to $\S\ref{sect:simple-roots}$. Direct computation shows that if $K = \mathrm{O}_N\,$, then $$\delta = \sum_{k=1}^n (k-1) (\mathrm{i} \psi_{n-k+1} - \phi_k)\;.$$ The same computation for the case of $K = \mathrm{USp}_N$ shows that $$\delta = \sum_{k=1}^n k\, (\mathrm{i} \psi_k - \phi_{n-k+1}) \;.$$ Now we write $\gamma = \sum_k (\mathrm{i} m_k \psi_k - n_k \phi_k)$ with the weight constraints $-\frac{N}{2} \le m_k \le \frac{N}{2} \le n_k\,$. The assumption that $E(\ell, \gamma)$ vanishes for all $\ell$ means that $$\sum _k (m_{n-k+1} + k - 1)^{2\ell} = \sum_k (n_k + k - 1)^{2\ell} \quad \text{for all}\ \ell$$ in the case of $K = \mathrm{O}_N\,$. In the case of $K = \mathrm{USp}_N$ it means that $$\sum_k (m_{n-k+1} + k)^{2\ell} = \sum_k (n_k + k)^{2\ell} \quad \text{for all}\ \ell \;.$$ In the second case the only solution for $m_k$ and $n_k$ satisfying the weight constraints is the highest weight $\lambda$ itself. In the first case there is one other solution, namely that which is obtained from the highest weight by replacing $m_n = \frac{N}{2}$ by $m_n = - \frac{N}{2}\,$. However, one directly checks that in the $\mathrm{O}_N$ case, where $2\mathrm{i}\psi_n$ is not a root, it is not possible to obtain such a $\gamma$ by adding some combination of roots from $\mathfrak{g}^{(2)}$ to $\lambda\,$. We are now able to give the proof of the uniqueness statement of Theorem \[unicity\]. We will determine $B_\gamma$ recursively, starting from $B_\lambda = 1$. Let $\gamma \not= \lambda$ be a weight that satisfies the weight constraints. Then if $K = \mathrm{USp}_N$ we know that $E(\ell , \gamma)$ is non-zero for some $\ell\,$. It therefore follows from equation (\[recursion equations\]) and $A_0 = 1$ that $$\label{basic equation} 0 = B_\gamma + \sum A_b \, B_{\gamma - b\sigma}\;,$$ where the sum runs over all $b \ne 0$ (recall that $b \ge 0$ is always the case) such that $\gamma - b\sigma \in \Gamma_\lambda\,$. Since the weights $\gamma - b\sigma $ involved in the sum are smaller than $\gamma$ in the natural partial order defined by the basis of simple roots, equation (\[basic equation\]) defines a recursion procedure for determining all coefficients $B_\gamma\,$. In the case of $K = \mathrm{O}_N$ we are confronted with the fact that the weight $\gamma = \lambda - \mathrm{i} N \psi_n$ satisfies the weight constraints and yet gives $E(\ell,\gamma) = 0$ for all $\ell$. However, in this exceptional situation the conditions of Theorem \[unicity\] provide that $B_{\lambda - \mathrm{i}N\psi_n} = 0$. Thus the expansion coefficients $B_\gamma$ are still uniquely determined by our recursion procedure. ### Explicit solution of the differential equations {#explicit solution} As before let $\Delta^+$ be a set of positive roots of $\mathfrak{g} = \mathfrak{osp} (U\oplus U^*)$ or $\mathfrak{osp}(\widetilde{U} \oplus \widetilde{U}^*)$. We now decompose these sets as $$\Delta^+ = \Delta_\lambda^+ \cup (\Delta^+ \setminus \Delta_\lambda^+)\;,\quad \Delta_\lambda^+ := \{\alpha\in\Delta^+ \mid \mathfrak{g}_\alpha \subset \mathfrak{g}^{(-2)}\} \;,$$ which means that $\Delta^+ \setminus \Delta_\lambda^+$ is a set of positive roots of $\mathfrak{g}^{(0)}$. Let $\Delta_\lambda^+$ be further decomposed as $\Delta_\lambda^+ = \Delta_{\lambda,0}^+ \cup \Delta_{\lambda,1}^+$ where $\Delta_{\lambda,0}^+$ and $\Delta_{\lambda,1}^+$ are the subsets of even and odd $\lambda$-positive roots. Then the function $J$ has a factorization as $J = J_0\, Z^{-1} \mathrm{e}^{\delta^\prime}$ with $$J_0 = \frac{\prod_{\alpha \in \Delta_0^+ \setminus \Delta_{\lambda,0}^+} 2 \sinh\frac{\alpha}{2}}{\prod_{\beta \in \Delta_1^+ \setminus \Delta_{\lambda,1}^+} 2 \sinh \frac{\beta}{2}}\ ,\quad Z = \frac{\prod_{\beta \in \Delta_{\lambda,1}^+} (1 - \mathrm{e}^{-\beta})}{\prod_{ \alpha\in\Delta_{\lambda,0}^+} (1-\mathrm{e}^{-\alpha})}\;,$$ and $\delta^\prime = \frac{1}{2} ( \sum \alpha - \sum \beta)$ is half the supersum of $\lambda$-positive roots. For the case of $K = \mathrm{O}_N$ one finds $\delta^\prime = - \frac{1}{2} \sum (\mathrm{i}\psi_j - \phi_j) = - \lambda_1\,$, while for $K = \mathrm{USp}_N$ one has $\delta^\prime = \lambda_1\,$. The Weyl group $W$ acts on $T$ and therefore on functions on $T$. Let $W_\lambda \subset W$ be the subgroup which stabilizes the highest weight $\lambda = \lambda_N$ and thus the corresponding function $\mathrm{e}^\lambda$ on $T$. Note that $W_\lambda$ is the direct product of the permutations of the set $\{ \mathrm{e}^{\phi_j} \}$ and the permutations of the set $\{ \mathrm{e}^{\mathrm{i}\psi_j}\}$. The symmetrizing operator $S_W$ from $W_\lambda$-invariant analytic functions to $W$-invariant analytic functions on $T$ is given by $$S_W (f) := \sum_{[w] \in W/W_\lambda} w(f) \;.$$ Notice that the function $\mathrm{e}^\lambda Z$ is $W_\lambda $-invariant; the symmetrized function $S_W( \mathrm{e}^\lambda Z)$ then is $W$-invariant. We now wish to show that this function coincides with our character $\chi$. In this endeavor, an obstacle appears to be that $\mathrm{e}^\lambda \chi$ by Corollary \[cor:weightexpansion\] is a polynomial in the variables $\mathrm {e}^{\mathrm{i}\psi_1}, \ldots, \mathrm{e}^{ \mathrm{i} \psi_n }$, whereas the function $Z$ has poles at $\mathrm{e}^{ \mathrm{i}(\psi_j + \psi_k)} = 1$. Hence our next step is to show that these poles are actually cancelled by the process of $W$-symmetrization. The function $S_W (\mathrm{e}^{\lambda} Z)$ is holomorphic on $\cap_{j=1}^n \{ \mathfrak{Re}\, \phi_j > 0 \}$. An even root $\alpha \in \Delta_0^+$ is some linear combination of either the functions $\phi_j$ or the functions $\psi_j\,$. Denoting the latter subset of even roots by $\Delta_0^+(\psi) \subset \Delta_0^+ \,$, let $\Sigma_\alpha \subset T$ for $\alpha \in \Delta_0^+(\psi)$ be the complex submanifold $$\Sigma_\alpha := \{ t \in T \mid \mathrm{e}^{\alpha}(t) = 1 \}.$$ By definition, the function $S_W (\mathrm{e}^\lambda Z)$ is holomorphic on $$\Big( \cap_{j=1}^n\{\mathfrak{Re}\, \phi_j > 0 \} \Big) \setminus \left( \cup_{\alpha \in \Delta_0^+(\psi)} \Sigma_\alpha \right) \;.$$ Now it is a theorem of complex analysis that if a function is holomorphic outside an analytic set of complex codimension at least two, then this function is everywhere holomorphic. Therefore, since the intersection of two or more of the submanifolds $\Sigma_\alpha$ is of codimension at least two in $T$, it suffices to show that for any $\alpha \in \Delta_0^+(\psi)$ our function $S_W (\mathrm{e} ^\lambda Z)$ extends holomorphically to $$D_\alpha := \Sigma_\alpha \setminus \left(\cup_{\Delta_0^+(\psi) \ni \alpha^\prime \not=\alpha}\,\Sigma_{\alpha^\prime}\right)\;.$$ Hence let $\alpha$ be some fixed root in $\Delta_0^+(\psi)$. There exists a Weyl group element $w \in W$ and a $w$-invariant neighborhood $U$ of $D_\alpha$ such that $w : \, U \to U$ is a reflection fixing the points of $D_\alpha\,$. Let $z_\alpha : \, U \to \mathbb{C}$ be a complex coordinate which is transverse to $D_\alpha$ in the sense that $w(z_\alpha) = - z_\alpha\,$. Because the root $\alpha$ occurs at most once in the product $Z$, the function $S_W(\mathrm{e}^\lambda Z)$ has at most a simple pole in $z_\alpha\,$. We may choose $U$ in such a way that $S_W(\mathrm{e} ^\lambda Z)$ is holomorphic on $U \setminus D_\alpha\,$. Doing so we have a unique decomposition $$S_W (\mathrm{e}^\lambda Z) = \frac{A}{z_\alpha} + B$$ where $A$ and $B$ are holomorphic in $U$. Since $S_W(\mathrm{e} ^\lambda Z)$ is $W$-invariant, we conclude that $w(A) = - A$ and hence $A = 0$ along $D_\alpha\,$. For all $\ell\, , N \in \mathbb{N}$ the function $\varphi :\, T \to \mathbb{C}$ defined by $\varphi = S_W(\mathrm{e}^{\lambda_N} Z)$ is a solution of the differential equation $D_\ell (J \varphi) = 0\,$. Using $Z = \mathrm{e}^{\delta^\prime} J_0 / J$ we write $\varphi = S_W(\mathrm{e}^{\lambda_N + \delta^\prime} J_0/J)$. Then, lifting the sum over cosets $[w] \in W/W_\lambda$ to a sum over Weyl group elements $w \in W$ we obtain $$\mathrm{ord}(W_\lambda) \, J \varphi = J \sum_{w \in W} w(\mathrm{e}^{\lambda_N + \delta^\prime} J_0 / J) =\sum_{w \in W} \mathrm{sgn}(w)\, w(J_0\,\mathrm{e}^{\lambda_N+\delta^\prime})\;,$$ where $w \mapsto \mathrm{sgn}(w) \in \mathbb{Z}_2 = \{ \pm 1 \}$ is the determinant of $w \in W \subset \mathrm{O}(\mathfrak{t}) = \mathrm{O}_{2n}\,$. The factor $\mathrm{e}^{\lambda_N + \delta^\prime}$ is the character of the representation $(\frac{N \mp 1}{2} \mathrm{STr}\, , \mathrm{SDet}^{\frac{N \mp 1}{2}})$ of (a double cover of) the Lie supergroup $(\mathfrak{g}^{(0)} , \mathrm{GL}(U_0) \times \mathrm{GL}(U_1))$. This representation is one-dimensional, and from Corollary \[cor:4.4\] we have $D_\ell (J_0\, \mathrm{e}^{ \lambda_N + \delta^\prime}) = 0$ for all $\ell , N \in \mathbb{N}\,$. The statement of the lemma now follows by applying the $W$-invariant differential operator $D_\ell$ to the formula for $\mathrm{ord}(W_\lambda) J \varphi$ above. ### Weight constraints {#weight-constraints} Here we carry out the final step in proving the explicit formula for the character $\chi$ of our representation. Since the formula in the case of $K = \mathrm{SO}_N$ follows directly from that for $K = \mathrm{O}_N$ (see $\S\ref{varying}$) and the case of $K = \mathrm{U}_N$ has been handled in [@CFZ], we need only discuss the cases of $K = \mathrm{O}_N$ and $K = \mathrm{USp}_N\,$. In order to show that the character is indeed given by $\chi = \varphi$ with $\varphi = S_W(\mathrm{e}^\lambda Z)$, it remains to prove that in the series development $\varphi = \sum B_{\gamma}\, \mathrm{e}^\gamma$ of the function defined by $$\label{key expression} \varphi = \sum_{[w] \in W/W_\lambda} \varphi_{[w]} \;, \quad \varphi_{[w]} := \mathrm{e}^{w(\lambda _N)} \frac{\prod_{\beta \in \Delta _{\lambda,1}^+} (1 - \mathrm{e}^{ -w(\beta)})} {\prod_{\alpha \in \Delta_{\lambda,0}^+} (1 - \mathrm{e}^{-w(\alpha)})} \;,$$ the only non-zero coefficients $B_\gamma$ are those where the linear functions $\gamma$ are of the form $\gamma = \sum (\mathrm{i} m_k \psi_k - n_k \phi_k)$ with $- \frac{N}{2} \le m_k \le \frac{N}{2} \le n_k\, $. We also have to show that $B_\gamma = 0$ in the case of the exceptional weight $\gamma = \lambda - \mathrm{i}N\psi_n$ occurring for $K = \mathrm{O}_N\,$. We have shown above that $\varphi = S_W(\mathrm{e}^\lambda Z)$ is holomorphic on the product of the full complex torus of the variables $\mathrm{e}^{\mathrm{i} \psi_k}$ with the domain defined by $\mathfrak{Re}\, \phi_k > \,0\,$. Although the individual terms $\varphi_{ [w]}$ in the representation of $\varphi$ have poles (which cancel in the Weyl group averaging process) we may still develop each term of $\varphi$ in a series expansion; this will in fact yield the desired weight constraints. We begin with the situation where $K = \mathrm{O}_N\,$. In this case $\Delta ^+_{\lambda,0}$ consists of the roots $\mathrm{i} \psi_j + \mathrm{i} \psi_k\,$ ($j<k$) and $\phi_j + \phi_k$ ($j \le k)$, and $\Delta^+_{ \lambda,1}$ is the set of roots of the form $\mathrm{i} \psi_j + \phi_k$ ($j,\,k = 1, \ldots, n$). Let us first consider the term of $\varphi$ where $[w] = W_\lambda\,$. Its denominator can be developed in a geometric series on the region corresponding to $\mathfrak{Re}\, \phi_k > 0$ for all $k\,$. There we may write this term as $$\varphi_{[\mathrm{Id}]}=\mathrm{e}^{\lambda_N}\prod_{\beta} (1-\mathrm{e}^{-\beta}) \prod_\alpha \sum_{n \ge 0} \mathrm{e}^{-n\alpha} \;.$$ Here and for the remainder of this paragraph $\alpha$ runs through the $\lambda$-positive even roots and $\beta$ through the $\lambda$-positive odd roots. Recall that $\lambda_N = \frac{N}{2}\sum (\mathrm{i} \psi_k - \phi_k )$, and note that all powers of $\mathrm{e}^{\mathrm{i} \psi_k}$ and $\mathrm{e}^{\phi_k}$ occurring in the series expansion of $$\prod_\beta (1 - \mathrm{e}^{-\beta}) \prod_\alpha \sum_{n \ge 0} \mathrm{e}^{- n\alpha}$$ are non-positive. Thus, if $\gamma = \sum (\mathrm{i} m_k \psi_k - n_k \phi_k)$ is a weight which arises in $\varphi_{[\mathrm{Id}]}\,$, then $n_k \ge \frac{N}{2}$ and $m_k \le \frac{N}{2}$. In the case of the $m_k$ this is a statement only about the term $\varphi_{ [\mathrm{Id}]}\,$, but, since the action of the Weyl group on the variables $\phi_k$ is just by permutation of the indices, it follows that $n_k \ge \frac{N}{2}$ holds always, independent of the term $\varphi_{[w]}$ under consideration. Hence, we neglect the $\phi_k$ in our further discussion and only analyze the powers of the exponentials $\mathrm{e}^{\mathrm{i} \psi_k}$ which arise in the other terms $\varphi_{[w]}\,$. Given a fixed index $k \in \{1, \ldots, n\}$ we will develop every term $\varphi_{[w]}$ on a region $R = R(k)$ defined by certain inequalities which in the case of $k = 1$ are $$\mathfrak{Re}(\mathfrak{i} \psi_1) > \ldots > \mathfrak{Re} (\mathrm{i} \psi_n) > 0 \;.$$ We now discuss this case in detail. Recall that $\mathrm{i} \psi_1$ occurs in the denominator in factors of the form $$(1-\mathrm{e}^{-w(\alpha)})^{-1} = (1 - \mathrm{e}^{- w(\mathrm{i}\psi_1) - w(\mathrm{i} \psi_j)})^{-1}$$ for $j > 1\,$. If $w(\mathrm{i} \psi_1) = \mathrm{i} \psi_1\,$, then we expand these factors just as in the case of $\varphi_{ [\mathrm{Id}]}\,$. Convergence of the resulting series is guaranteed no matter what $w$ does to $\psi_j\,$. In the situation where $w(\mathrm{i} \psi_1) = - \mathrm{i} \psi_1$ we rewrite the factors in the denominator as $(1-\mathrm{e}^{-w (\alpha)})^{-1} = -\mathrm{e}^{w(\alpha)} (1 - \mathrm{e}^{w(\alpha) })^{-1}$ and expand, and convergence in $R$ is again guaranteed. Adding these series we obtain a series representation $$\varphi = \sum_{[w]} \varphi_{[w]} = \sum_\gamma B_\gamma \, \mathrm{e}^\gamma \;,$$ which is convergent on $R\,$. \[m\_1 estimate\] If $\gamma = \sum (\mathrm{i} m_k \psi_k - n_k \phi_k)$ and $B_\gamma \not= 0\,$, then $m_1 \le \frac{N}{2}\, $. If $w(\mathrm{i} \psi_1) = \mathrm{i} \psi_1\,$, then by the same argument as in the case of $[w] = [\mathrm{Id}]$ we see that $\mathrm{e}^{ \mathrm{i} \psi_1}$ occurs in the series development of $\varphi_{[w]}$ with a power $m_1$ of at most $\frac{N}{2}\,$. Now suppose that $w(\mathrm{i}\psi_1) = - \mathrm{i} \psi_1\,$. Then, following the prescription above we rewrite the $\psi_1$-dependent factors in $\varphi_{[w]}$ as $$\label{eq:4.7mrz} \mathrm{e}^{\frac{N}{2} w(\mathrm{i}\psi_1)} \frac{\prod_{j \ge 1} (1 - \mathrm{e}^{-w(\mathrm{i}\psi_1 + \phi_j)})} {\prod_{j \ge 2} (1 - \mathrm{e}^{-w(\mathrm{i}\psi_1 + \mathrm{i}\psi_j)})} = \mathrm{e}^{(-\frac{N}{2}+1) \mathrm{i}\psi_1}\frac{\prod_{j\ge 1} (\mathrm{e}^{-\mathrm{i}\psi_1} - \mathrm{e}^{-\phi_j})} {\prod_{j \ge 2}(\mathrm{e}^{- \mathrm{i}\psi_1} - \mathrm{e}^{-w(\mathrm{i}\psi_j)})} \;,$$ and expand the r.h.s. in powers of $\mathrm{e}^{- \mathrm{i} \psi_1}$. It follows that in this case $m_1 \le - \frac{N}{2} + 1\,$, which for any positive integer $N$ implies that $m_1 \le \frac{N}{2}\,$. Using Weyl group invariance, this estimate for $m_1$ will now yield the desired result. \[constraints fulfilled\] Suppose that $K = \mathrm{O}_N$ and let $\varphi = \sum B_\gamma\, \mathrm{e}^\gamma$ be the globally convergent series expansion of the proposed character $\varphi = S_W(\mathrm{e}^\lambda Z)$. Then for every weight $\gamma = \sum (\mathrm{i} m_k \psi_k - n_k \phi_k)$ with $B_\gamma \not= 0$ it follows that $- \frac{N}{2} \le m_k \le \frac{N}{2} \le n_k\,$. The inequality $n_k \ge \frac{N}{2}$ was proved above as an immediate consequence of the fact that the Weyl group $W$ effectively acts only on the $\psi_j\,$. Above we showed that on the region $R$ the proposed character $\varphi$ has a series development where in every $\gamma$ the coefficient $m_1$ of $\mathrm{i} \psi_1$ is at most $\frac{N}{2}\,$. Recalling the fact that the function $\varphi$ is holomorphic on $T$, we infer that $m_1 \le \frac{N}{2}$ also holds true for the globally convergent series development $\sum B_\gamma\, \mathrm{e}^\gamma$. To get the same statement for $\mathrm{i} \psi_k$ with $k \not= 1$ we just change the definition of $R$ to $R(k)$ defined by the inequalities $\mathfrak{Re} (\mathrm{i} \psi_k) > \mathfrak{Re} (\mathrm{i} \psi_1) > \ldots > 0 \, $. Arguing for general $k$ as we did for $k = 1$ in the above Lemma, we show that the coefficient $m_k$ of $\mathrm{i} \psi_k$ in every $\gamma$ in the series expansion of every $\varphi_{[w]}$ on $R(k)$ is at most $\frac{N}{2}\,$. By the holomorphic property, the same is true for the global series expansion of the proposed character $\varphi$. Hence, to complete the proof we need only show the inequality $m_k \ge -\frac{N}{2}\,$. But for this it suffices to note that for every $k$ there is an element $w$ of the Weyl group with $w(\mathrm{i} \psi_k) = - \mathrm{i}\psi_k\,$. Indeed, using the Weyl invariance of $\varphi$, if there was some $\gamma$ where $m_k < - \frac{N}{2}\,$, then the coefficient of $\mathrm{i} \psi_k$ in $w(\gamma )$ would be larger than $\frac{N}{2}\,$. To complete our work, we must prove Lemma \[constraints fulfilled\] for the case $K = \mathrm{USp}_N\,$. For this we use the same notation as above for the basic linear functions, namely $\mathrm{i} \psi_k$ and $\phi_k\,$. Here, compared to the $\mathrm{O}_N$ case, there are only slight differences in the $\lambda$-positive roots and the Weyl group. The only difference in the roots is in $\Delta^+_{\lambda,0}$ where $\mathrm{i} \psi_j + \mathrm{i} \psi_k$ occurs in the larger range $j \le k$ and $\phi_j + \phi_k$ in the smaller range $j < k\,$. The Weyl group acts by permutation of indices on both the $\mathrm{i} \psi_j$ and $\phi_j$ and by sign reversal on the $\mathrm{i} \psi_j\, $. In this case, as opposed to the case above where only an even number of sign reversals were allowed, every sign reversal transformation is in the Weyl group. In order to prove Lemma \[constraints fulfilled\] in this case, we need only go through the argument in the $\mathrm{O}_N$ case and make minor adjustments. In fact, the main step is to prove Lemma \[m\_1 estimate\] and, there, the only change is that the range of $j$ for the factor $1 - \mathrm{e}^{-\mathrm{i}(\psi_1 + \psi_j)}$ is larger. This is only relevant in the case $w(\mathrm{i} \psi_1) = - \mathrm{i} \psi_1\,$, where we rewrite the additional denominator term $(1 - \mathrm{e}^{-w(2\mathrm{i} \psi_1)})^{-1}$ as $- \mathrm{e}^{-2\mathrm{i} \psi_1}(1 - \mathrm{e}^{- 2\mathrm{i} \psi _1})^{-1}$. Hence the factor in front of the ratio of products on the r.h.s. of equation (\[eq:4.7mrz\]) gets an additional factor of $\mathrm{e}^{-2\mathrm{i}\psi_1}$ and now is $\mathrm{e}^{-\mathrm{i} (\frac{N}{2} + 1)\psi_1}$. Thus $m_1 \le -\frac{N}{2} -1$ which certainly implies $m_1 \le \frac{N}{2}\,$. Let us summarize this discussion. For both $K = \mathrm{O}_N$ and $K = \mathrm{USp}_N$ every weight $\gamma = \sum (\mathrm{i} m_k \psi_k - n_k \phi_k)$ occurring in the series expansion $S_W(\mathrm{e}^\lambda Z) = \sum B_\gamma\, \mathrm{e}^\gamma $ obeys the weight constraints $$- \frac{N}{2} \le m_k \le \frac{N}{2} \le n_k \quad (k = 1, \ldots, n)\;.$$ Moreover, using the fact that the Weyl group transformations for $K = \mathrm{O}_N$ always involve an even number of sign changes, one sees that $B_{\lambda - \mathrm{i}N\psi_n} = 0$ in that case. As a consequence of the uniqueness theorem (Theorem \[unicity\]) we therefore have $$\chi = S_W(\mathrm{e}^\lambda Z) = \sum_{[w] \in W/W_\lambda} \mathrm{e}^{w(\lambda_N)}\frac{\prod_{\beta\in\Delta_{\lambda,1}^+} (1 - \mathrm{e}^{-w(\beta)})}{\prod_{\alpha\in\Delta_{\lambda,0}^+} (1 - \mathrm{e}^{-w(\alpha)})}$$ in both the $\mathrm{O}_N$ and $\mathrm{USp}_N$ cases. Since the $\mathrm{SO}_N$ case has been handled as a consequence of the result for $\mathrm{O}_N \,$, our work is now complete. [1]{} Berezin, F.A.: *Introduction to superanalysis*, Reidel Publishing Comp., Dordrecht, 1987 Berline, N., Getzler, E., Vergne, M.: *Heat kernels and Dirac operators*, Berlin, Springer Verlag, 1992 Bump, D., Gamburd, A.: On the averages of characteristic polynomials from classical groups, Commun. Math.Phys. [**265**]{} (2006) 227-274 Conrey, J.B., Farmer, D.W., Zirnbauer, M.R.: Howe pairs, supersymmetry and ratios of random characteristic polynomials for the unitary group (math-ph/0511024) Conrey, J.B., Forrester, P., Snaith, N.C.: Averages of ratios of characteristic polynomials for the compact classical groups, Int. Math. Res. Notices [**7**]{} (2005) 397-431 Folland, G.B.: *Harmonic analysis on phase space*, Annals of Mathematics Studies 122, Princeton, Princeton University Press, 1989 Howe, R.: Remarks on classical invariant theory, Trans. AMS [**313**]{} (1989) 539-570 Howe, R.: The oscillator semigroup, Proc. of Symposia in Pure Mathematics [**48**]{}, AMS (1987) Reed, M., Simon, B.: *Methods of modern mathematical physics. I: Functional analysis*, New York, Academic Press, 1972

ArXiv

--- abstract: 'We introduce the concept of monogamy deficit for quantum correlation by combining together two types of monogamy inequalities depending on different measurement sides. For tripartite pure state, we demonstrate a relation which connects two types of monogamy inequalities for quantum discord and provide the difference between them. By using this relation, we obtain an unified physical interpretation for these two monogamy deficit. In addition, we find an interesting fact that there is a general monogamy condition for several quantum correlations for tripartite pure states. We then provide a necessary and sufficient condition for the establishment of one kind of monogamy inequality for tripartite mixed state and generalize it to multipartite quantum state.' author: - 'Si-Yuan Liu' - Bo Li - 'Wen-Li Yang' - Heng Fan title: Monogamy deficit for quantum correlations in multipartite quantum system --- Introduction ============ Quantum correlations, such as entanglement and quantum discord, are assumed to be resources in quantum information processing and are different from classical correlations. On the other hand, in general, entanglement and discord are different from each other. Previous studies focus on entanglement which is a special quantum correlation enabling fascinating quantum information tasks such as super-dense coding[@CH], teleportation[@CH2], quantum cryptography[@AK], remote-state preparation[@AK2] and so on. However, some quantum applications superior than their classical counterparts are found with vanishing or negligible entanglement [@CH3; @EK; @AD]. In this sense, entanglement seems not capture all the quantum features of quantum correlations. So other measures of quantum correlations are proposed. Among those measures that in general go beyond entanglement, quantum discord is a widely accepted one in recent years [@Zurek; @Vedral; @key-15]. The analytic results of quantum discord and its physical meaning are studied extensively, for example, in Refs. [@AD; @PG; @AS; @AS2; @DC]. The experiments about quantum discord are implemented [@RA; @RA2]. Quantum discord can also be generalized to the multipartite situation [@WO; @KM; @IC; @ccm11], for more results, see a recent review paper [@key-15]. There are many fundamental differences between classical correlation and quantum correlations. One of them is the shareability of correlation among many parties. Generally speaking, classical correlation can be freely shared among many parties, while quantum ones do not have this property. For example, for tripartite pure state, if two parties are highly entangled, they cannot have a large amount of entanglement shared with a third one. The limits on the shareability of quantum correlations are described by monogamy inequalities. Much progresses have already been made about the monogamy properties of various quantum correlations [@JO1; @KOa; @SA; @key-12; @key-13; @key-14]. As one application, the monogamy property of quantum correlations also play a fundamental role for the security of the quantum key distributions[@LoChau; @MP]. Some known monogamy properties of entanglement measure are, for example, concurrence and squashed entanglement [@FE; @key-1; @key-14; @KO]. It is shown that the monogamy relation does not always be satisfied by the quantum correlations [@JO1]. So it is necessary to know when a specified quantum correlation can satisfy this property. Concerning about quantum discord, in general, it does not satisfy this nature [@MA]. However, it may have some interesting applications in case the monogamy condition is satisfied [@ma; @hufan]. We should note that there are two types of monogamy inequalities for quantum discord since it is asymmetric depending on the measurement side for a bipartite state [@XI]. A necessary and sufficient condition for one type of monogamy relation satisfying is given where only one side of measurement is studied [@MA]. A natural question is then that does there exist an analogous property for another class of monogamy with measurement taken on a different side? In this paper, we first demonstrate a relation between those two types of monogamy conditions for tripartite pure state and provide the difference between them. By using this relation, we provide an unified physical interpretation for these two monogamy deficit and generalize it to the $N$-partite pure state. Then we give a necessary and sufficient condition for the holding of the second type of monogamy relation and further generalize the result to the $N$-partite system. In particular, those two types of monogamy relations are generally studied independently. Our result that two monogamy inequalities can be combined together by introducing monogamy deficit provides a new, in general, more complete viewpoint. This can enlighten much research both on quantum correlation and monogamy property. This paper is organized as follows. In the next section, after introducing the definition of monogamy deficit, we derive a relation between these two types of monogamy deficit for tripartite pure state. Using thisn unified physical interpretation for these two monogamy deficit. In addition, we find an interesting fact that there is a general monogamy condition for several quantum correlations for tripartite pure states. Then, we provide a necessary and sufficient condition for the non-negative second monogamy deficit for tripartite mixed state. We also use a figure to illustrate the monogamy deficit under specific pure state. By considering the equivalent expression of the necessary and sufficient condition for tripartite pure state, we obtain a new physical interpretation of entanglement of formation. Finally, as an further application of the above necessary and sufficient condition, we extend our result to the $N$ partite system. The monogamy deficit for pure state {#connection1} =================================== The connection of two types of monogamy deficit {#connection} ----------------------------------------------- Quantum discord is defined as the difference between mutual information, which is accepted to be the total correlation, and maximum classical mutual information [@Zurek; @Vedral] $$\begin{aligned} D^{\rightarrow }(\rho _{AB}) &=&\widetilde{I}(\rho _{AB})-I^{\rightarrow }(\rho _{AB}) \notag \\ &=&S(\rho _{A})-S(\rho _{AB})+\min_{\{\mathcal{M} _{i}\}}\sum_{i}p_{i}S(\rho _{B|i}), \label{discord}\end{aligned}$$ where arrow “$\rightarrow$” means measurement on $`A'$ and “$\leftarrow$” means measurement on $`B'$, ${{M} _{i}}$ represent POVM measurement performed on $A$ for a bipartite state $\rho _{AB}$. So quantum discord is considered describing the quantumness of correlations. In this paper, we mainly use the same notations as those in Ref.[@MA]. We use the left arrow (“$\leftarrow$”) and the right arrow (“$\rightarrow$”) to distinguish the side of the measurement. Also we have notations, $\widetilde{I}(\rho _{AB})=S(\rho _{A})-\widetilde{S}(\rho _{A|B})$, and $\widetilde{S}(\rho _{A|B})=S(\rho _{AB})-S(\rho _{B})$, here $S(\sigma)=-{\rm tr}(\sigma\log_2\sigma)$ is the von Neumann entropy of a density matrix $\sigma$. By those definitions presented above, we know that quantum discord in general should be asymmetric and depend on the measurement side, which can be either $A$ or $B$. It is understandable that those two definitions possess different fundamental properties. Recently, two kinds of monogamy inequalities have been studied in Refs.[@MA; @giorgi] and [@XI]. For a tripartite state $\rho _{ABC}$, by combining two monogamy inequalities together, we define two kinds of monogamy deficit of quantum discord, $$\begin{aligned} \bigtriangleup_{D_A}^{\leftarrow }=D^{\leftarrow }(\rho_{A|BC})-D^{\leftarrow }(\rho_{AB})-D^{\leftarrow }(\rho_{AC}), \label{monogam1}\end{aligned}$$ $$\begin{aligned} \bigtriangleup_{D_A}^{\rightarrow }=D^{\rightarrow }(\rho_{A|BC})-D^{\rightarrow }(\rho_{AB})-D^{\rightarrow }(\rho_{AC}). \label{monogamyde}\end{aligned}$$ It is worth noting that a similar quantity as defined in Eq. (2) and (3) has also been introduced by Bera *et al.* [@MN]. In Eq. (\[monogam1\]), the first term $D^{\leftarrow }(\rho_{A|BC})$ involves a positive operator valued measurement (POVM) performed on $B$ and $C$, and the other involve measurements only on $A$. Because of the asymmetry of quantum discord, the above two monogamy deficit are apparently quite different. In this paper, however, we find that there is a relation between them, which means that the monogamy relations on one of them provide some limits on another. We first have a following observation. For two kinds of monogamy deficit $\bigtriangleup_{D_A}^{\leftarrow }$, $\bigtriangleup_{D_A}^{\rightarrow }$ of an arbitrary tripartite pure state $\rho _{ABC}$, we find, $$\begin{aligned} \bigtriangleup_{D_A}^{\leftarrow }&=&\frac{1}{2}(\bigtriangleup_{D_B}^{\rightarrow }+\bigtriangleup_{D_C}^{\rightarrow }), \label{1deficit} \\ \bigtriangleup_{D_A}^{\rightarrow }&=&\bigtriangleup_{D_B}^{\leftarrow }+\bigtriangleup_{D_C}^{\leftarrow }-\bigtriangleup_{D_A}^{\leftarrow }. \label{relation}\end{aligned}$$ The proof of those two relations can be the following. For simplicity, denote $S(\rho_{B|A})$ as the optimal conditional entropy of $I^{\rightarrow }(\rho _{AB})$ after the measurement which defined as $\min_{\{\mathcal{M} _{i}\}}\sum_{i}p_{i}S(\rho _{B|i})$, that is $I^{\rightarrow }(\rho _{AB})=S(\rho _{B})-\min_{\{\mathcal{M} _{i}\}}\sum_{i}p_{i}S(\rho _{B|i})=S(\rho _{B})-S(\rho_{B|A})$. Using the Koashi-Winter formula [@KO], we have $S(\rho_{B|A})=E(\rho_{BC})$, where $E(\rho_{BC})$ means the entanglement of formation for a bipartite state $\rho _{BC}$. Generally, for any tripartite pure state $|\psi\rangle_{A'A''A'''}$, we have $S(\rho_{A'|A''})=E(\rho_{A'A'''})$, where $A', A'', A'''$ correspond to any permutations of $A, B, C$. Further more, we find, $$\begin{aligned} D^{\leftarrow }(\rho _{A'A''}) &=&\widetilde{I}(\rho _{A'A''})-(S(\rho _{A'})-S(\rho_{A'|A''})) \notag \\ &=&S(\rho _{A''})-S(\rho _{A'''})+E(\rho _{A'A'''}), \notag \\ D^{\rightarrow }(\rho _{A'A''}) &=&\widetilde{I}(\rho _{A'A''})-(S(\rho _{A''})-S(\rho_{A''|A'})) \notag \\ &=&S(\rho _{A'})-S(\rho _{A'''})+E(\rho _{A''A'''}). \label{LEFt}\end{aligned}$$ Inserting (\[LEFt\]) into (\[monogam1\]),(\[monogamyde\]), we have that $$\begin{aligned} \bigtriangleup_{D_A}^{\leftarrow }&=&S(\rho _{A})-E(\rho _{AC})-E(\rho _{AB}), \notag \\ \bigtriangleup_{D_A}^{\rightarrow }&=&S(\rho _{B})+S(\rho _{C})-S(\rho _{A})-2E(\rho _{BC}). \label{monogam2}\end{aligned}$$ Similarly, $\bigtriangleup_{D_B}^{\leftarrow }, \bigtriangleup_{D_C}^{\leftarrow }$ and $\bigtriangleup_{D_B}^{\rightarrow }, \bigtriangleup_{D_C}^{\rightarrow }$ can be obtained by permutating the indices of (\[monogam1\]) and (\[monogamyde\]). Combining those results, we have (\[1deficit\]), (\[relation\]), which completes the proof. The above relations are interesting. They tell us that the two kinds of monogamy inequalities which was studied previously [@MA; @giorgi; @XI] actually are not independent. We find that $\bigtriangleup_{D_A}^{\leftarrow }$, which is the defined monogamy deficit having a coherent measurement taken on two parties $B$ and $C$, is precisely equal to the arithmetic mean of $\bigtriangleup_{D_B}^{\rightarrow }$ and $\bigtriangleup_{D_C}^{\rightarrow }$ in which the measurements are only performed individually on $B$ and $C$. To be explicit, the measurement for left hand side (l.h.s.) of equality (\[1deficit\]) is a coherent measurement on “$BC$” while on right hand side (r.h.s.), local measurements on “$B$” and “$C$” are performed. In Ref. [@giorgi], a transition from satisfying the monogamy inequality to violation of monogamy inequality is given, where the positive or negative of $\bigtriangleup_{D_A}^{\leftarrow }$ are studied. By the definition of monogamy deficit $\bigtriangleup_{D_A}^{\leftarrow }=D^{\leftarrow }(\rho_{A|BC})-D^{\leftarrow }(\rho_{AB})-D^{\leftarrow }(\rho_{AC})$, it is apparent that a coherent measurement on “$BC$” is necessary. Here our result (\[1deficit\]) shows that instead of a coherent measurement, local measurements individually on “$B$” and “$C$” can be performed to find this conclusion. We remark that local operation is much easier to be implemented than coherent measurement. Those results reveal the hidden relationship in monogamy deficit for quantum discord where the coherent measurement is replaced by local measurements. In the previous paragraphs, we have already discussed the relationship between these two monogamy deficit. Now let us consider the difference between them. The difference between the two monogamy deficit can be expressed as the following form $$\begin{aligned} E\left(\rho_{AB}\right) -\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}=\frac{1}{2} ( \triangle_{D_{C}}^{\leftarrow}- \triangle_{D_{C}}^{\rightarrow}). \label{deiference}\end{aligned}$$ Where the $\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}$ is the average of $D^{\rightarrow}\left(\rho_{BA}\right)$ and $D^{\rightarrow}\left(\rho_{AB}\right)$. Here we give a simple proof of the above formula. By using the Koashi-Winter relation, we have $$\begin{aligned} \triangle_{D_{B}}^{\rightarrow}=I^{\rightarrow}\left(\rho_{BA}\right) -D^{\rightarrow}\left(\rho_{BA}\right),\end{aligned}$$ $$\begin{aligned} \triangle_{D_{A}}^{\leftarrow}=I^{\rightarrow}\left(\rho_{BA}\right) -E\left(\rho_{AB}\right).\end{aligned}$$ By subtracting the above equalities, we have $$\begin{aligned} \triangle_{D_{B}}^{\rightarrow} -\triangle_{D_{A}}^{\leftarrow}=E\left(\rho_{AB}\right) - D^{\rightarrow}\left(\rho_{BA}\right),\end{aligned}$$ by exchanging the symbols of $A$ and $B$, we have the similar equation $$\begin{aligned} \triangle_{D_{A}}^{\rightarrow} - \triangle_{D_{B}}^{\leftarrow}=E\left(\rho_{BA}\right)- D^{\rightarrow}\left(\rho_{AB}\right).\end{aligned}$$ Combining the above formulas, the $E\left(\rho_{AB}\right)$ - $\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}$ can be expressed as follows $$\begin{aligned} E\left(\rho_{AB}\right) - \bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}= \frac{1}{2} ( \triangle_{D_{A}}^{\rightarrow}+ \triangle_{D_{B}}^{\rightarrow}) - \frac{1}{2} (\triangle_{D_{A}}^{\leftarrow} + \triangle_{D_{B}}^{\leftarrow}).\end{aligned}$$ Substituting (\[1deficit\]) and (\[relation\]) into the above equation, we have Eq.(\[deiference\]), which completes the proof. This formula is meaningful, it shows that the difference between these two monogamy deficit depends on the balance of entanglement of formation (EOF) and the average of discord. In other words, if the EOF between $A$ and $B$ is greater than or equal to the average of discord, the first monogamy deficit which contains measurements on $A$ and $B$ must be greater than or equal to the second monogamy deficit which only contains local measurement on $C$. Especially when $E\left(\rho_{AB}\right)$ is equal to the $\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}$, these two monogamy deficit is equivalent. In other words, in this case, when we consider the monogamy property of quantum discord, we only need to know one of them. Since the second monogamy deficit only contains local measurement on one party, it is easier to calculate and to be used in practical applications. Furthermore, the above formula provides a new physical interpretation of the difference between EOF and discord in the average sense. As an application of the relationship between the two monogamy deficit, for tripartite pure states, we provide an unified physical significance for these two monogamy deficit. To see this, we first consider the equivalent expression of the second monogamy deficit. According to the results in [@FF] and [@XI], we have $I^{\rightarrow}\left(\rho_{BA}\right)-D^{\rightarrow}\left(\rho_{BA}\right)=I^{\rightarrow}\left(\rho_{BC}\right)-D^{\rightarrow}\left(\rho_{BC}\right)=\widetilde{I}(\rho _{AC})-2E(\rho _{AC})$ and $\triangle_{D_{B}}^{\rightarrow}=\widetilde{I}(\rho _{AC})-2E(\rho _{AC})$. Combing the two equations, we have that $$\begin{aligned} \triangle_{D_{B}}^{\rightarrow}=I^{\rightarrow}\left(\rho_{BA}\right)-D^{\rightarrow}\left(\rho_{BA}\right)= I^{\rightarrow}\left(\rho_{BC}\right)-D^{\rightarrow}\left(\rho_{BC}\right).\label{monogam221}\end{aligned}$$ Where $I^{\rightarrow}\left(\rho_{BA}\right)$ represents the classical correlation, $D^{\rightarrow}\left(\rho_{BA}\right)$ represents quantum discord. By exchanging the subscript, we have $\triangle_{D_{C}}^{\rightarrow}=I^{\rightarrow}\left(\rho_{CA}\right)-D^{\rightarrow}\left(\rho_{CA}\right)=I^{\rightarrow}\left(\rho_{CB}\right)-D^{\rightarrow}\left(\rho_{CB}\right)$. It tells us that this monogamy deficit is equivalent to the difference between classical correlation and quantum correlation. Since the classical correlation can be regarded as locally accessible mutual information (LAMI)[@FFF], while the quantum correlation can be seen as locally inaccessible mutual information (LIMI). In this sense, this monogamy deficit tells us that how much mutual information can be extracted from a tripartite pure state by using local measurement on one party. To be more explicit, the monogamy inequality holds if and only if more than half of the mutual information between $AC$ or $BC$ can be accessed through local measurement performed on $C$. The above result provides a interesting relationship between the second monogamy deficit and the difference between LAMI and LIMI for arbitrary tripartite pure states. In the following, we generalize the relationship to arbitrary $N$-partite pure states. For arbitrary $N$-partite pure states $\rho_{A_1\cdots A_N}$, we have $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}- \triangle_{D_{A_{1\left(N-1\right)}}}^{\rightarrow}=I^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right) - D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right).\end{aligned}$$ Where $\triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}$ represent the second monogamy deficit for arbitrary $N$-partite pure state and is given by $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}=D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right) -\sum_{i=2}^{N} D^{\rightarrow}\left(\rho_{A_{1}A_{i}}\right),\end{aligned}$$ similarly, the second monogamy deficit $\triangle_{D_{A_{1\left(N-1\right)}}}^{\rightarrow}$ for its $(N-1)$-partite subsystem is $$\begin{aligned} \triangle_{D_{A_{1\left(N-1\right)}}}^{\rightarrow} =D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right) -\sum_{i=2}^{N-1} D^{\rightarrow}\left(\rho_{A_{1}A_{i}}\right).\end{aligned}$$ So we have $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}- \triangle_{D_{A_{1\left(N-1\right)}}}^{\rightarrow}&=&D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right)-D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right). \notag\\ & &- D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right).\end{aligned}$$ Using the Koashi-Winter relationship and considering the property of pure states, it is easy to show $$\begin{aligned} D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right) + I^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right) = S\left(\rho_{A_{1}}\right).\end{aligned}$$ Since $S\left(\rho_{A_{1}}\right)=D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right)$ for pure state, the above results can be rewritten as follows $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}- \triangle_{D_{A_{1\left(N-1\right)}}}^{\rightarrow} &=&I^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right) - D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right)\\ &=&I^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right) - D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right).\end{aligned}$$ Which completes the proof. This equation tells us that the difference between the second monogamy deficit of $N$-partite system and its $(N-1)$-partite subsystem is equivalent to the difference between classical correlation and quantum discord. That is to say, the difference between LAMI and LIMI can tell us that which system is more monogamous, the $N$-partite system or its $(N-1)$-partite subsystem. In other words, if we can extract more than half of the mutual information between $A_{1}$ and $A_{N}$ through local measurements on $A_{1}$, the $N$-partite system must be more monogamous than its $(N-1)$-partite subsystem. As we all know, in studying entanglement of a tripartite system, the monogamy deficit of entanglement can be seen as a tripartite correlation which is called tangle, or genuine entanglement. Similarly, for discord of an $N$-partite system, the monogamy deficit can also be seen as a type of multipartite correlation which beyond the usual bipartite correlations. In this sense, the $N$-partite system contains more multipartite correlation than its $(N-1)$-partite subsystem if and only if we can acquire at least half of the mutual information between $A_{1}$ and $A_{N}$ through local measurements on $A_{1}$. When $N=3$, the above result goes back to formula (\[monogam221\]). Now we can give a similar equivalent expression of $\triangle_{D_{C}}^{\leftarrow}$. In order to achieve this purpose, we first define the average of classical correlation and discord. As we all know, the discord and classical correlation are asymmetry quantities. By using the asymmetry of $I^{\rightarrow}\left(\rho_{BA}\right)$ and $I^{\rightarrow}\left(\rho_{AB}\right)$, we define the average of classical correlation $\bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}$ and the average discord $\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}$ [@FFF], $$\begin{aligned} \bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}=\frac{1}{2}(I^{\rightarrow}\left(\rho_{BA}\right)+I^{\rightarrow}\left(\rho_{AB}\right)),\nonumber\\ \bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}=\frac{1}{2}(D^{\rightarrow}\left(\rho_{BA}\right)+D^{\rightarrow}\left(\rho_{AB}\right)).\end{aligned}$$ From this definition, combing Eq. (\[1deficit\]) and (\[monogam221\]) , we have $$\begin{aligned} \triangle_{D_{C}}^{\leftarrow}&=&\frac{1}{2}(\triangle_{D_{A}}^{\rightarrow}+\triangle_{D_{B}}^{\rightarrow})\nonumber\\ &=&\frac{1}{2}(I^{\rightarrow}\left(\rho_{AB}\right)-D^{\rightarrow}\left(\rho_{AB}\right)+I^{\rightarrow}\left(\rho_{BA}\right)-D^{\rightarrow}\left(\rho_{BA}\right))\nonumber\\ &=&\bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}-\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}.\label{monogam22121}\end{aligned}$$ This formula means that the monogamy deficit which needs a coherent measurement performed on two parties $A$ and $B$ is equivalent to $\bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}-\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}$, which is the difference between the average of classical correlations and quantum correlations where local measurements are made individually on $A$ and $B$. In other words, according to previous view, this monogamy deficit represents our ability to extract the mutual information by performing local measurements on these two parties. According to the above definition, simply we have the relation, $\bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}+\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)}=\widetilde{I}\left(\rho_{AB}\right)$. Which means that the monogamy inequality holds if and only if we can acquire at least half of the mutual information through local measurements in the average sense. The Eq. (\[monogam22121\]) presented above can also be used as a criterion to check whether a given tripartite pure state belongs to GHZ class or W class state under stochastic local operations and classical communication (SLOCC). According to the results in Ref. [@MA], we have that a tripartite pure state belongs to GHZ class state if and only if $\triangle_{D_{C}}^{\leftarrow}\geq 0$, otherwise it belongs to W class state. In other words, we can say that a tripartite pure state belongs to GHZ class state if and only if the LAMI is always greater than or equal to the LIMI in the average sense when local measurements are performed on $A$ and $B$. While a tripartite pure state belongs to W class state if and only if LAMI is less than LIMI in the average sense when local measurements are performed on $A$ and $B$. In addition, a tripartite pure state belongs to GHZ class or W class depends on whether one can acquire no less than half of the mutual information through local measurements in the average sense. The above result tells us that there is a interesting relationship between the first monogamy deficit and the difference between the average of LAMI and LIMI for arbitrary tripartite pure states. In fact, we can generalize the relationship to arbitrary $N$-partite pure states. For arbitrary $N$-partite pure states $\rho_{A_1\cdots A_N}$, we have $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}-\triangle_{D_{A_{1\left(N-1\right)}}}^{\leftarrow} = \bar{\omega}_{\left(A_{2\cdots}A_{N-1}\right)\mid A_{N}}^{+}{}_{\left(I\right)} - \bar{\omega}_{\left(A_{2\cdots}A_{N-1}\right)\mid A_{N}}^{+}{}_{\left(D\right)}.\end{aligned}$$ Where $\triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}$ represent the first monogamy deficit for arbitrary $N$-partite pure state and is given by $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}=D^{\leftarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right) - \sum_{i=2}^{N}D^{\leftarrow}\left(\rho_{A_{1}A_{i}}\right),\end{aligned}$$ similarly, the first monogamy deficit $\triangle_{D_{A_{1\left(N-1\right)}}}^{\leftarrow}$ for its $(N-1)$-partite subsystem is $$\begin{aligned} \triangle_{D_{A_{1\left(N-1\right)}}}^{\leftarrow}=D^{\leftarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right) - \sum_{i=2}^{N-1}D^{\leftarrow}\left(\rho_{A_{1}A_{i}}\right).\end{aligned}$$ So we have $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}-\triangle_{D_{A_{1\left(N-1\right)}}}^{\leftarrow} &=&D^{\leftarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right) - D^{\leftarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N-1}}\right) \notag\\ & &- D^{\leftarrow}\left(\rho_{A_{1}A_{N}}\right).\end{aligned}$$ Let’s regard the $N$-partite pure state $\rho_{A_{1}\cdots A_{N}}$ as the tripartite pure state $\rho_{A_{1}\left(A_{2}\cdots A_{N-1}\right)A_{N}}$. In this sense, the right-hand side of the above formula can be considered to be an first monogamy deficit for this equivalent tripartite pure state. By using the previous equation (\[monogam22121\]), the above formula can be expressed as follows $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}-\triangle_{D_{A_{1\left(N-1\right)}}}^{\leftarrow} = \bar{\omega}_{\left(A_{2\cdots}A_{N-1}\right)\mid A_{N}}^{+}{}_{\left(I\right)} - \bar{\omega}_{\left(A_{2\cdots}A_{N-1}\right)\mid A_{N}}^{+}{}_{\left(D\right)}.\end{aligned}$$ Which completes the proof. This equation provides that the difference between the monogamy deficit of $N$-partite system and its $(N-1)$-partite subsystem is equivalent to the difference between the average of classical correlations and quantum correlations. That is to say, the right hand side of this equation can tell us that which system is more monogamous, the $N$-partite system or its $(N-1)$-partite subsystem. Then we can say that the $N$-partite system must be more monogamous than its $(N-1)$-partite subsystem if and only if the LAMI is always greater than or equal to the LIMI in the average sense. Additionally, similar as the explanation of the formula (10), the $N$-partite system contains more multipartite correlation than its $(N-1)$-partite subsystem if and only if we can acquire at least half of the mutual information through measurements performed on $\left(A_{2\cdots}A_{N-1}\right)$ and $A_{N}$ in the average sense. When $N=3$, the above result returns to formula (\[monogam22121\]). As a short summary for previous discussion, for tripartite pure state, we demonstrate a relation which connects two types of monogamy inequalities for quantum discord and provide the difference between them. By using this relation, we get an unified view for these two monogamy inequalities. That is, for arbitrary tripartite pure states, both of the monogamy inequalities hold only if one can extract more than half of mutual information by using local measurement. Monogamy deficit for discord and other quantum correlations {#squash} ----------------------------------------------------------- The squashed entanglement is an entanglement monotone for bipartite quantum states introduced by Christandl and Winter [@FE]. For bipartite state $\rho_{AB}$, the squashed entanglement is given by $$\begin{aligned} E_{sq}(\rho_{AB})\equiv \frac{1}{2}inf\widetilde{I}(\rho_{A:B\mid C}),\end{aligned}$$ here $\widetilde{I}(\rho_{A:B\mid C})$ is the conditional mutual information of $\rho_{ABC}$ with respect to particle $C$ (see Eq. (\[conditionm\])), $\rho_{ABC}$ is the extension of $\rho_{AB}$ and the infimum is taken over all extensions of $\rho_{AB}$ such that $\rho _{AB}={\rm Tr}_C\rho _{ABC}$. The squashed entanglement has many important properties [@KO]. For example, *(1) The squashed entanglement is upper bounded by entanglement of formation. (2) For any tripartite state $\rho_{ABC}$, we have $E_{sq}(\rho_{AB})+E_{sq}(\rho_{AC})\leq E_{sq}(\rho_{A(BC)})$.* By using the property (2), we can generalize the concept of monogamy deficit for squashed entanglement. We define the monogamy deficit for squashed entanglement as $$\begin{aligned} \triangle_{(E_{sq(C)})}=E_{sq}(\rho_{C(AB)})-E_{sq}(\rho_{CA})-E_{sq}(\rho_{CB}). \label{squuud}\end{aligned}$$ Similarly, one can also define the monogamy deficit $\triangle_{E_C}=E(\rho_{C(AB)})-E(\rho_{CA})-E(\rho_{CB})$ for the entanglement of formation. For the monogamy deficit for squashed entanglement, we can have the following result. For any tripartite pure state $\rho_{ABC}$, we may observe, $$\begin{aligned} \triangle_{(E_{sq(C)})}\geq\max \{ \triangle_{D_{C}}^{\leftarrow},0\}=\max\left\{ \bar{\omega}_{A\mid B}^{+}{}_{\left(I\right)}-\bar{\omega}_{A\mid B}^{+}{}_{\left(D\right)},0\right\}. \label{relati}\end{aligned}$$ The proof of this relation is presented below. By the above property (2) of monogamy deficit for squashed entanglement, we have $\triangle_{(E_{sq(C)})}\geq 0$, and the second equality is given by Eq.(\[monogam22121\]). We now only need to prove $\triangle_{(E_{sq(C)})}\geq \triangle_{E_{C}}$ and $\triangle_{E_{C}}=\triangle_{D_{C}}^{\leftarrow}$. For tripartite pure state, $E_{sq}(\rho_{C(AB)})=E(\rho_{C(AB)})$, thus we have $$\begin{aligned} \triangle_{(E_{sq(C)})}&=& E_{sq}(\rho_{C(AB)})-E_{sq}(\rho_{CA})-E_{sq}(\rho_{CB})\nonumber\\ &=& E(\rho_{C(AB)})-E_{sq}(\rho_{CA})-E_{sq}(\rho_{CB})\nonumber\\ &\geq & E(\rho_{C(AB)})-E(\rho_{CA})-E(\rho_{CB}).\nonumber\\ &=& \triangle_{E_{C}}.\end{aligned}$$ For tripartite pure state, we have $E(\rho_{C(AB)})=D^{\leftarrow }(\rho_{C|AB})=S(\rho_{C})$, combing $E(\rho_{CA})+E(\rho_{CB})=D^{\leftarrow}\left(\rho_{CA}\right)+D^{\leftarrow}\left(\rho_{CB}\right)$ in Ref. [@FF1], thus we have $\triangle_{E_{C}}=\triangle_{D_{C}}^{\leftarrow}$. That is $\triangle_{(E_{sq(C)})}\geq \triangle_{E_{C}}=\triangle_{D_{C}}^{\leftarrow}$. Now we know that Eq. (\[relati\]) is true. The quantum work deficit is an important information- theoretic measure of quantum correlation introduced by Oppenheim *et al.* [@JO]. For an arbitrary bipartite state $\rho_{AB}$, the quantum work-deficit is defined as $$\begin{aligned} \Delta(\rho_{AB})=I_G(\rho_{AB})-I_L(\rho_{AB}), \label{workdefi}\end{aligned}$$ where $I_G(\rho_{AB})$ represents the thermodynamic “work” that can be extracted from $\rho_{AB}$ by “closed global operations”, $I_L(\rho_{AB})$ represents the thermodynamic “work” that can be extracted from $\rho_{AB}$ by closed local operation and classical communication (CLOCC) [@key-12]. Further more, the one side work deficit $\Delta^\rightarrow(\rho_{AB})$ ($\Delta^\leftarrow(\rho_{AB})$) means that CLOCC is restricted on projection measurements at one particle $A$ ($B$). According to Ref.[@key-12], the one side work deficit is lower bounded by quantum discord, that is $$\begin{aligned} D^\rightarrow(\rho_{AB})\leq \Delta^\rightarrow(\rho_{AB}),\nonumber\\ D^\leftarrow(\rho_{AB}) \leq \Delta^\leftarrow(\rho_{AB}).\end{aligned}$$ Similar to quantum discord and squashed entanglement, we provide the definition of the monogamy deficit for work deficit. The two kinds of monogamy deficit for work deficit are given as, $$\begin{aligned} \bigtriangleup_{\Delta_A}^{\leftarrow }=\Delta^{\leftarrow }(\rho_{A:BC})-\Delta^{\leftarrow }(\rho_{AB})-\Delta^{\leftarrow }(\rho_{AC}), \label{monogam11}\end{aligned}$$ $$\begin{aligned} \bigtriangleup_{\Delta_A}^{\rightarrow }=\Delta^{\rightarrow }(\rho_{A:BC})-\Delta^{\rightarrow }(\rho_{AB})-\Delta^{\rightarrow }(\rho_{AC}). \label{monogamyde2}\end{aligned}$$ The first definition involves a POVM coherently performed on $B$ and $C$ together, and the other involve measurements only on $A$. For the monogamy deficit for work deficit, we present the following observations. For any tripartite pure state $\rho_{ABC}$, we have that $$\begin{aligned} & &\bigtriangleup_{\Delta_A}^{\leftarrow }\leq \bigtriangleup_{D_A}^{\leftarrow }\leq \triangle_{(E_{sq(A)})}, \label{relati12}\\ & & \bigtriangleup_{\Delta_A}^{\rightarrow }\leq \bigtriangleup_{D_A}^{\rightarrow }. \label{relati123}\end{aligned}$$ The correctness of these observations are presented below. The inequality $\bigtriangleup_{D_A}^{\leftarrow }\leq \triangle_{(E_{sq(A)})}$ is from (\[relati\]). For tripartite pure state $\rho_{ABC}$, we have $\Delta^{\leftarrow }(\rho_{A:BC})=D^{\leftarrow }(\rho_{A|BC})=S(\rho_A)$, $D^{\leftarrow }(\rho_{AB})\leq \Delta^{\leftarrow }(\rho_{AB})$, $D^{\leftarrow }(\rho_{AC})\leq \Delta^{\leftarrow }(\rho_{AC})$[@key-12]. Which implies that $\bigtriangleup_{\Delta_A}^{\leftarrow }\leq \bigtriangleup_{D_A}^{\leftarrow }$. Similarly, we can prove (\[relati123\]). The physical interpretation of Eq. (\[relati12\]) can be like the following. The monogamy property for work deficit implies the monogamy property for quantum discord, entanglement of formation and squashed entanglement for any pure state. In this case, we have $\bigtriangleup_{\Delta_A}^{\leftarrow }\leq \bar{\omega}_{B\mid C}^{+}{}_{\left(I\right)}-\bar{\omega}_{B\mid C}^{+}{}_{\left(D\right)} \leq \triangle_{(E_{sq(A)})} $. In addition, we can extract more than half of the mutual information between $BC$ in the average sense through local measurements. Combing (\[monogam221\]) and (\[relati123\]), we have $\bigtriangleup_{\Delta_A}^{\rightarrow }\geq 0$ which implies one can extract at least half of the mutual information between $AB$ or $AC$ by using local measurement of $A$. Necessary and sufficient criteria for non-negative monogamy deficit $\bigtriangleup_{D_A}^\rightarrow$ {#two} ====================================================================================================== As is shown in [@MA], a necessary and sufficient condition for discord to be monogamous is $\bigtriangleup_{D_A}^{\leftarrow }\ge 0$, see Eq. (\[monogam1\]) for definition. Since we present two kinds of monogamy deficit, an interesting question is what does it means if the measurement is taken on another side, $\bigtriangleup_{D_A}^\rightarrow \ge 0$? In this section, we will consider this question and prove a similar necessary and sufficient condition for the second kinds of monogamy inequality. We first present some definitions about mutual information, conditional mutual information with respect to a single particle $A$. For a tripartite state $\rho_{ABC}$, the unmeasured conditional mutual information with respect to particle $A$ is given as, $$\begin{aligned} \widetilde{I}_A(\rho_{B:C|A})\equiv \widetilde{S}(\rho_{B|A})+\widetilde{S}(\rho_{C|A})-\widetilde{S}(\rho_{BC|A}), \label{conditionm}\end{aligned}$$ and the interrogated conditional mutual information with respect to particle $A$ is, $$\begin{aligned} I_A(\rho_{B:C|A})\equiv S(\rho_{B|A})+S(\rho_{C|A})-S(\rho_{BC|A}).\end{aligned}$$ By the strong subadditivity of von Neumann entropy, we know, $\widetilde{I}_A(\rho_{B:C|A})\ge 0$, $I_A(\rho_{B:C|A})\ge 0$, both are non-negative. We next propose the concept of interaction information. The (unmeasured) interaction information $\widetilde{I}_A(\rho_{ABC})=\widetilde{I}_A(\rho_{B:C|A})-\widetilde{I}(\rho_{BC})$. By simple calculating, one may observe that $\widetilde{I}_A(\rho_{ABC})=S(\rho_{AB})+S(\rho_{AC})+S(\rho_{BC}) -(S(\rho_{A})+S(\rho_{B})+S(\rho_{C}))-S(\rho_{ABC})= \widetilde{I}(\rho_{ABC})$, this is the interaction information defined in Ref. [@MA]. For the state $\rho_{ABC}$ and a given measurement $\{\mathcal{M}_i^A\}$, an interrogated interaction information with respect to $A$ is given as, $$\begin{aligned} I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}\equiv I_A(\rho_{B:C|A})-I_A(\rho_{BC})_{\{\mathcal{M}_i^A\}}. \label{interaction-inf}\end{aligned}$$ Since $I_A(\rho_{BC})_{\{\mathcal{M}_i^A\}}$ do not have particle $A$, we have $I_A(\rho_{BC})_{\{\mathcal{M}_i^A\}}=\widetilde{I}(\rho_{BC})$, which does not involve any measurement. Given a tripartite quantum state $\rho_{ABC}$, $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ represents the interaction information with respect to $A$, which is defined in (\[interaction-inf\]). For this definition (\[interaction-inf\]), the first term of the right hand side is the conditional mutual information of $B$, $C$ when $A$ is present and measured, the second term is the mutual information of $BC$ where $A$ is absent. Here, $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ measures the effect on the amount of correlation shared between $B$ and $C$ by measuring $A$. A positive interaction information with respect to $A$ means the presentation of $A$ can enhance the total correlation between $B$ and $C$, while negative interaction information with respect to $A$ means the presentation of $A$ inhibits the total correlation between $B$ and $C$. $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ has the similar property as $I(\rho_{ABC})$ proposed in Ref. [@MA] and can be read as a necessary and sufficient criteria for a monogamy inequality. We next have the following theorem. For any state $\rho_{ABC}$, $D^\rightarrow(\rho_{AB})+D^\rightarrow(\rho_{AC})\leq D^\rightarrow(\rho_{A|BC})$ if and only if the interrogated interaction information with respect to $A$ is less than or equal to the unmeasured interaction information with respect to $A$. We only need to calculate the monogamy deficit $\bigtriangleup_{D_A}^\rightarrow$, $$\begin{aligned} \bigtriangleup_{D_A}^{\rightarrow } &=&D^{\rightarrow }(\rho _{A|BC})-D^{\rightarrow }(\rho _{AB})-D^{\rightarrow }(\rho _{AC}) \notag \\ &=&S(\rho _{BC|A})-\widetilde{S}(\rho _{BC|A})-(S(\rho _{B|A})-\widetilde{S}(\rho _{B|A})) \notag \\ & &-(S(\rho _{C|A})-\widetilde{S}(\rho _{C|A})) \notag \\ &=&S(\rho _{BC|A})-S(\rho _{B|A})-S(\rho _{C|A})+I(\rho _{BC}) \notag \\ & &-(\widetilde{S}(\rho _{BC|A})-\widetilde{S}(\rho _{B|A})-\widetilde{S}(\rho _{C|A})+\widetilde{I}(\rho _{BC})) \notag \\ &=&\widetilde{I}_A(\rho_{ABC})-I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}. \label{disefi}\end{aligned}$$ From (\[disefi\]), we have $\bigtriangleup_{D_A}^{\rightarrow }\geq 0$ if and only if $\widetilde{I}_A(\rho_{ABC})\geq I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ which completes the proof. For pure state, we have $\widetilde{I}_A(\rho_{ABC})=0$, and the monogamy deficit of quantum discord is equivalent to the non-positivity of the interrogated information with respect to $A$. To see a transition from violation to observation of monogamy, we consider a family of states [@giorgi], $$\begin{aligned} |\widetilde{\psi}\left(p,\varepsilon\right)\rangle&=&\sqrt{p\varepsilon}|000\rangle+\sqrt{p(1-\varepsilon)}|111\rangle \nonumber \\ &&+\sqrt{\frac{1-p}{2}}(|101\rangle+|110\rangle).\end{aligned}$$ Note that $|\widetilde{\psi}(\frac{1}{3},1)\rangle$ is the maximally entangled W state $\sqrt{\frac{1}{3}}(|000\rangle+|101\rangle+|110\rangle)$, while $|\widetilde{\psi}(1,\frac{1}{2})\rangle$ is the GHZ state, $\sqrt{\frac{1}{2}}(|000\rangle+|111\rangle)$. In Fig.1, $\triangle_{D_{A}}^{\rightarrow}=D^{\rightarrow}(\rho_{A|BC})-D^{\rightarrow}(\rho_{AB})-D^{\rightarrow}(\rho_{AC})$ is plotted as a function of $p$ for different values of $\varepsilon$. From this figure, we can show that the interrogated information with respect to $A$ can be positive or negative for tripartite pure state. The $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ is increasing with the increasing of $\varepsilon$. All of the three lines are very close to each other when $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ is positive and the critical point from positive to negative is almost identical for them. Especially for the W state, when p approaches to 1, the $I_A(\rho_{ABC})_{\{\mathcal{M}_i^A\}}$ approaches to zero. ![ (Color online) the monogamy deficit $\triangle_{D_{A}}^{\rightarrow}$ for $|\widetilde{\psi}(p,\varepsilon)\rangle$, $\triangle_{D_{A}}^{\rightarrow}$ as a function of $p$ for different values of $\varepsilon$ (see the main text). States are monogamy when the respective curves are positive. Red dashed line is for $\varepsilon=0.5$, blue dotted line is for $\epsilon=0.75$, blue solid line is for $\varepsilon=1$. The monogamy deficit is decreasing with the increasing of $\varepsilon$. All of the three lines are very close to each other when $\triangle_{D_{A}}^{\rightarrow}$ is polygamy and the critical point from polygamy to monogamy is almost identical for them. When $\varepsilon\rightarrow1$, $|\widetilde{\psi}(p,\varepsilon)\rangle$ approach to W states, in this case, when $p\rightarrow1$, $\triangle_{D_{A}}^{\rightarrow}\rightarrow0$. The blue solid line is increased first and then decreased. []{data-label="lct"}](tuf.eps){width="8cm"} As an application of the necessary and sufficient conditions, we find an interesting equivalent expression of entanglement of formation for tripartite pure states. For tripartite pure states, it is shown that [@XI], $\triangle_{D_{A}}^{\rightarrow}=\widetilde I(\rho_{BC})-2E(\rho_{BC})$. From the previous discussion, we have the formula, $\triangle_{D_{A}}^{\rightarrow}=\widetilde{I_{A}}\left(\rho_{B:C\mid A}\right)-I_{A}\left(\rho_{B:C\mid A}\right)$, which holds for general tripartite mixed states. When we consider the case of pure states, the above two expressions should be equal. That is to say, in this case, $\widetilde I(\rho_{BC})-2E(\rho_{BC})=\widetilde{I_{A}}\left(\rho_{B:C\mid A}\right)-I_{A}\left(\rho_{B:C\mid A}\right)$. At the same time, it is easy to show that $\widetilde{I_{A}}\left(\rho_{B:C\mid A}\right)=S\left(\rho_{BA}\right)+S\left(\rho_{CA}\right)-S\left(\rho_{BCA}\right)-S\left(\rho_{A}\right)=S\left(\rho_{C}\right)+S\left(\rho_{B}\right)-S\left(\rho_{BC}\right)=\widetilde I(\rho_{BC})$. So we have $E(\rho_{BC})=\frac{1}{2}I_{A}\left(\rho_{B:C\mid A}\right)$. Thus the interrogated conditional mutual information with respect to $A$ is twice of the entanglement of formation for state $\rho _{BC}$. Necessary and sufficient criteria for non-negative monogamy deficit for multipartite system {#necess} =========================================================================================== In this section, we generalize our result to multipartite system. We give a necessary and sufficient condition for $\triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}\ge 0$, where the monogamy deficit is defined for multipartite state, $\triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}=D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right)-D^{\rightarrow}\left(\rho_{A_{1}A_{2}}\right)-D^{\rightarrow}\left(\rho_{A_{1}A_{3}}\right)-\cdots-D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right)$. In order to consider this question, similar as the tripartite state, we next present some definitions about mutual information, conditional mutual information with respect to a single particle $A_{1}$. For a $N$-partite state $\rho_{A_{1}\cdots A_{N}}$, the unmeasured conditional mutual information with respect to particle $A_{1}$ is given as $\widetilde{I_{A_{1}}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)=\widetilde{S}\left(\rho_{A_{K}\mid A_{1}}\right)+\widetilde{S}\left(\rho_{\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)-\widetilde{S}\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)$, where $K=2,\ldots, N-1$. The interrogated conditional mutual information with respect to particle $A_{1}$ is $I_{A_{1}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)=S\left(\rho_{A_{K}\mid A_{1}}\right)+S\left(\rho_{\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)-S\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)$. By the strong subadditivity of von Neumann entropy, we have $\widetilde{I_{A_{1}}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)$ and $I_{A_{1}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)$ are both non-negative. We define the concept of interaction information with respect to $A_{1}$. The (unmeasured) interaction information is defined as, $\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)=\widetilde{I_{A_{1}}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)-\widetilde{I}\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)$. For the state $\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}$ and a given measurement $\left\{ M_{i}^{A_{1}}\right\} $, an interrogated interaction information with respect to $A_{1}$ is given as $I_{A_{1}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }=I_{A_{1}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)-I_{A_{1}}\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }$, where the suffix $I_{A_{1}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }$ is used to indicate the measurements on $A_{1}$. Similar as the above calculating, we find that $\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)=S\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)+S\left(\rho_{A_{K}A_{1}}\right)+S\left(\rho_{\left(A_{K+1}\cdots A_{N}\right)A_{1}}\right)-(S\left(\rho_{A_{K}}\right)+S\left(\rho_{\left(A_{K+1}\cdots A_{N}\right)}\right)+S\left(\rho_{A_{1}}\right))-S\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)A_{1}}\right)=\widetilde{I}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)$, which is the interaction information we have defined. Since $I_{A_{1}}\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }$ do not have particle $A_{1}$, we have $I_{A_{1}}\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }=\widetilde I\left(\rho_{A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)$, which also does not involve any measurement as in tripartite case. Now we can give the necessary and sufficient condition for $\triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}$ is no less than zero. We have the following theorem. For any $\rho_{A_{1}\cdots A_{N}}, \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}\geq0$ if and only if the interrogated interaction information with respect to $A_{1}$ being less than or equal to the unmeasured interaction information with respect to $A_{1}$. Here, we only need to calculate the monogamy deficit, $$\begin{aligned} \triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}&=& D^{\rightarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right) -D^{\rightarrow}\left(\rho_{A_{1}A_{2}}\right)-D^{\rightarrow}\left(\rho_{A_{1}A_{3}}\right)\notag \\ & &-\cdots-D^{\rightarrow}\left(\rho_{A_{1}A_{N}}\right)\notag \\ &=&[\widetilde{I_{A_{1}}}\left(\rho_{A_{2}:\left(A_{3}\cdots A_{N}\right)\mid A_{1}}\right)-I_{A_{1}}\left(\rho_{A_{2}:\left(A_{3}\cdots A_{N}\right)\mid A_{1}}\right)]\notag \\ & & +[\widetilde{I_{A_{1}}}\left(\rho_{A_{3}:\left(A_{4}\cdots A_{N}\right)\mid A_{1}}\right)-I_{A_{1}}\left(\rho_{A_{3}:\left(A_{4}\cdots A_{N}\right)\mid A_{1}}\right)]\notag \\ & & +\cdots+[\widetilde{I_{A_{1}}}\left(\rho_{A_{N-1}:A_{N}\mid A_{1}}\right)-I_{A_{1}}\left(\rho_{A_{N-1}:A_{N}\mid A_{1}}\right)]\notag \\ &=&\sum_{K=2}^{N-1}[\widetilde{I_{A_{1}}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)-I_{A_{1}}\left(\rho_{A_{K}:\left(A_{K+1}\cdots A_{N}\right)\mid A_{1}}\right)]\notag \\ &=&\sum_{K=2}^{N-1}\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right) \nonumber \\ &&-\sum_{K=2}^{N-1}I_{A_{1}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }. \label{disefi123}\end{aligned}$$ From (\[disefi123\]), we have $\triangle_{D_{A_{1\left(N\right)}}}^{\rightarrow}\geq 0$ if and only if $\sum_{K=2}^{N-1}\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)\geq \sum_{K=2}^{N-1}I_{A_{1}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)_{\left\{ M_{i}^{A_{1}}\right\} }$. Similarly, we can also get a necessary and sufficient condition for $$\begin{aligned} & &\triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}=\notag \\ & &D^{\leftarrow}\left(\rho_{A_{1}\mid A_{2}\cdots A_{N}}\right)-D^{\leftarrow}\left(\rho_{A_{1}A_{2}}\right)-D^{\leftarrow}\left(\rho_{A_{1}A_{3}}\right)\notag \\ & &-\cdots-D^{\leftarrow}\left(\rho_{A_{1}A_{N}}\right)\notag \\ &=&\sum_{K=2}^{N-1}[\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)-I\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)]. \label{disefi1234}\end{aligned}$$ Where there is no measurement contained in $\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)$, while local measurements $\{\mathcal{M}_i^{A_m}\}$ ($m=2,\ldots, N$) and coherent measurements $\{\mathcal{M}_i^{(A_k\ldots A_N)}\}$ ($k=2,\ldots, N-1$) contained in $I\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)]$. From the above proof, we have $\triangle_{D_{A_{1\left(N\right)}}}^{\leftarrow}\geq0$ if and only if $\sum_{K=2}^{N-1}\widetilde{I_{A_{1}}}\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)\geq \sum_{K=2}^{N-1} I\left(\rho_{A_{1}A_{K}\left(A_{K+1}\cdots A_{N}\right)}\right)$. Summary and discussion {#summ} ====================== We have introduced the concept of monogamy deficit by combining together the monogamy inequalities of quantum correlation for multipartite quantum system. Although two types of monogamy inequalities seem very different on their measurement sides, based on the concept of monogamy deficit, we have observed a relation and provided the difference between them. Using this relation, we obtain a unified physical interpretation for these two monogamy deficit. In addition, we find an interesting fact that there exists a general monogamy condition for several quantum correlations for tripartite pure states. By using the concept of interaction information with respect to one particle, we have proved that the necessary and sufficient condition for the quantum correlation being monogamous is that the interrogated interaction information with respect to one particle is less than or equal to the unmeasured interaction information. Our result can be generalized to $N$-partite system and may have applications in quantum information processing. We thank L. Chen for useful comments. This work is supported by “973” program (2010CB922904) and NSFC (11075126, 11031005, 11175248). [99]{} C.H. Bennett and S. Wiesner, Phys. Rev. Lett. 69, 2881(1992). C.H. Bennett, G. Brassard, Crépeau, R. Jozsa, A. Peres, and W. K. Wootters, Phys. Rev. Lett. 70, 1895 (1993). A. K. Ekert, Phys. Rev. Lett. 67, 661 (1991). A. K. Pati, Phys. Rev. A 63, 014302 (2000); C. H. Bennett, D. P. DiVincenzo, P. W. Shor, J. A. Smolin, B. M. Terhal, and W. K. Wootters, Phys. Rev. Lett. 87, 077902 (2001). C. H. Bennett, D. P. DiVincenzo, C. A. Fuchs, T. Mor, E. Rains, P. W. Shor, J. A. Smolin, and W. K. Wootters, Phys. Rev. A 59, 1070 (1999). E. Knill and R. Laflamme, Phys. Rev. Lett. 81, 5672 (1998). A. Datta, A. Shaji, and C.M. Caves, Phys. Rev. Lett. 100, 050502 (2008). H. Ollivier, and W. H. Zurek, Phys. Rev. Lett. [**88**]{}, 017901 (2001). L. Henderson, and V. Vedral, J. Phys. A **34**, 6899 (2001). K. Modi, A. Brodutch, H. Cable, T. Paterek, V. Vedral, Rev. Mod. Phys. **84**, 1655-1707 (2012); R. L. Franco, B. Bellomo, S. Maniscalco, and G. Compagno, Int. J. Mod. Phys. B 27, 1245053 (2013). P. Giorda, and M. G. A. Paris, Phys. Rev. Lett. 105, 020503 (2010); G. Adesso, A. Datta, Phys. Rev. Lett. 105, 030501 (2010). A. Streltsov, H. Kampermann, and D. Bru[ß]{}, Phys. Rev. Lett. 106, 160401 (2011). A. Shabani and D. A. Lidar, Phys. Rev. Lett. 102, 100402 (2009); C. A. Rodríguez-Rosario, G. Kimura, H. Imai, and A. Aspuru-Guzik, Phys. Rev. Lett. 106, 050403 (2011). D. Cavalcanti, L. Aolita, S. Boixo, K. Modi, M. Piani, and A. Winter, Phys. Rev. A 83, 032324 (2011); V. Madhok, A. Datta, Phys. Rev. A 83, 032323 (2011); B. Bellomo, R. L. Franco, and G. Compagno, Phys. Rev. A 86, 012312 (2012); B. Bellomo, G. Compagno, R. L. Franco, A. Ridolfo, S. Savasta, Int. J. Quant. Inf. 9, 1665 (2011). R. Auccaise, L. C. Cleri, D. O. Soares-Pinto, E. R. deAzevedo, J. Maziero, A. M. Souza, T. J. Bonagamba, R. S. Sarthour, I. S. Oliveira, and R. M. Serra, Phys. Rev. Lett. 107, 140403 (2011). R. Auccaise, J. Maziero, L. C. Cleri, D. O. Soares-Pinto, E. R. deAzevedo, T. J. Bonagamba, R. S. Sarthour, I. S. Oliveira, and R. M. Serra, Phys. Rev. Lett. 107, 070501 (2011). C. C. Rulli and M. S. Sarandy, Phys. Rev. A 84, 042109 (2011); M. Okrasa and Z. Walczak, EPL, 96 (2011) 60003. K. Modi, T. Paterek, W. Son, V. Vedral, and M. Williamson, Phys. Rev. Lett. 104, 080501 (2010). I. Chakrabarty, P. Agrawal, A. K. Pati , Eur. Phys. J. D. 65, 605 (2011). L. Chen, E. Chitambar, K. Modi, G. Vacanti, Phys. Rev. A 83, 020101(R) (2011). A. Streltsov, G. Adesso, M. Piani, and D. Bru[ß]{}, Phys. Rev. Lett. **109**, 050503. H. C. Braga, C. C. Rulli, T. R. de Oliveira, M. S. Sarandy, Phys. Rev. A 86, 062106 (2012). Sudha, A. R. U. Devi, and A. K. Rajagopal, Phys. Rev. A **85**, 012103 (2012). K. Salini, R. Prabhu, A. Sen De, U. Sen, arXiv:1206.4029. Y. K. Bai, N. Zhang, M. Y. Ye, Z. D. Wang, arXiv:1206.2096. T. J. Osborne, and F. Verstraete, Phys. Rev. Lett. **96**, 220503 (2006). M. Pawłowski, Phys. Rev. A 82, 032313 (2010). R. Prabhu, A. K. Pati, A. Sen De, U. Sen, Phys. Rev. A **86**, 052337 (2012). M. L. Hu and H. Fan, Phys. Rev. A (accepted), arXiv:1212.0139. F. F. Fanchini, M. F. Cornelio, M. C. de Oliveira, and A. O. Caldeira, Phys. Rev. A **84**, 012313 (2011). M. Christandl and A. Winter, J. Math. Phys. **45**, 829 (2004). G. L. Giorgi, Phys. Rev. A **84**, 054301 (2011). X. J. Ren, H. Fan, Quant. Inf. Comput. [**13**]{}, 0469 (2013). M. Koashi and A. Winter, Phys. Rev. A **69**, 022309 (2004) V. Coffman, J. Kundu, and W. K. Wootters, Phys. Rev. A 61, 052306 (2000). R. Prabhu, A. K. Pati, A. Sen De, U. Sen, Phys. Rev. A **85**, 040102 (2012). H. K. Lo and H. F. Chau, Science [**283**]{}, 2050 (1999). M. N. Bera, R. Prabhu, A. Sen De, U. Sen, Phys. Rev. A 86, 012319 (2012). F. F. Fanchini, M. C. de Oliveira, L. K. Castelano, M. F. Cornelio, arXiv:1110.1054v2. F. F. Fanchini, L. K. Castelano, M. F. Cornelio and M. C. de Oliveira, New J. Phys. [**14**]{}, 013027. J. Oppenheim, M. Horodecki, P. Horodecki, and R. Horodecki, Phys. Rev. Lett. 89, 180402 (2002); M. Horodecki, K. Horodecki, P. Horodecki, R. Horodecki, J. Oppenheim, A. Sen(De), and U. Sen, ibid. 90, 100402 (2003); M. Horodecki, P. Horodecki, R. Horodecki, J. Oppenheim, A. Sen(De), U. Sen, and B. Synak-Radtke, Phys. Rev. A 71, 062307 (2005).

ArXiv

--- author: - 'I. Balestra, S. Bianchi, G. Matt' bibliography: - 'sbs.bib' date: 'Received 19 August 2003/ Accepted 21 October 2003' title: 'The reprocessing features in the X–ray spectrum of the NELG MCG –5-23-16' --- Introduction ============ X–ray spectra of type 1 Active Galactic Nuclei (AGNs) have been extensively studied by many satellites in the last 15 years. After $Ginga$ it has become clear that typical X–ray broad band spectra of Seyfert galaxies result from the superposition of different components. The most common features arising over the primary power law emission are two reprocessed components: the so-called Compton reflection component at high energies (above $\sim$10 keV) and the Fe K$\alpha$ emission line at 6.4 keV from neutral iron. Both are due to reflection of the primary radiation by optically thick, neutral or low ionized, circumnuclear matter. At low energies, typically below about 1 keV, a further component may arise (soft excess), the origin of which is still rather unclear. The study of the reprocessed components and in particular of the iron line profile provides important information about the region from which they originate. If the line is produced in the innermost regions of the accretion disc, its profile must be broad, asymmetric and double-peaked due to kinematic and relativistic effects [@Fabian00 and references therein], while a narrow profile indicates an origin far from the nucleus, either in the torus envisaged in the Unification model [@Antonucci93] or in the Broad Line Regions. While ASCA observations indicate that a relativistic line is a common feature in Seyfert 1s [@Nandra97], XMM-$Newton$ have so far found unambiguous evidence for relativistic lines in an handful of objects only [@Wilms01; @Turner02], while finding routinely a narrow line [@Reeves02; @Bianchi03]. It is therefore very important for understanding the properties of the innermost regions of the accretion disc to search for more relativistic lines, to assess their frequency and strength. The fact that the relativistic line is broad, and therefore not easy to be separated from the continuum, and the confusing presence of the narrow component, implies that, unless the line equivalent width (EW) is very large, it should necessarily be searched for in bright sources. MCG –5-23-16 is a nearby (z = 0.0083) X–ray bright narrow emission line galaxy. Its 2–10 keV flux has varied by a factor of 4 within about 10 years: from $\sim8\times10^{-11}$ erg cm$^{-2}$ s$^{-2}$ in 1978 [@tennant83], to $\sim2\times10^{-11}$ erg cm$^{-2}$ s$^{-2}$ in 1989 [@np94], and again to a high state of $\sim9\times10^{-11}$ erg cm$^{-2}$ s$^{-2}$ in 1996 [@weaver98]. The nucleus is obscured by neutral matter with a column density of $\sim10^{22}$ cm$^{-2}$; no signature of a warm absorber is detected. A soft excess is present in the spectrum below about 1 keV. $RossiXTE$ detected a Compton reflection signature in this Seyfert galaxy for the first time [@weaver98]; $ASCA$ showed a strong (EW$\sim300$ eV) and broad Fe K$\alpha$ line with a possible complex profile [@weaver97]; BeppoSAX confirmed the presence of the Compton reflection component and measured an equivalent width of $\sim100$ eV for the iron line when fitted with a single gaussian [@per02]. More recently $Chandra$ HETGS spectra, briefly described in @weaver01, showed a narrow (FWHM $<$4000 km s$^{-1}$) line with EW$\sim90$ eV, while XMM-$Newton$ revealed a complex Fe K$\alpha$ line constituted by a narrow unresolved component (EW$\sim40$ eV) and a broad (FWHM $\sim40\,000$ km s$^{-1}$) component with an equivalent width of $\sim120$ eV [@dew03]. In this paper we present a reanalysis of the $ASCA$, BeppoSAX, $Chandra$ and XMM-$Newton$ observations based on a model which includes Compton reflection and two distinct components, a narrow and a broad one, for the iron line. We will exploit the different capabilities of the various satellites (energy resolution for $Chandra$, sensitivity for XMM-$Newton$, hard X–ray coverage for BeppoSAX) to obtain informations on the location and physical properties of the emitting regions. **Date** **Mission** **Instr.** **T$\mathrm{_{exp}}$ (ks)** ------------ -------------- ----------------------------------------------------------- ----------------------------- 11/05/1994 $ASCA$ <span style="font-variant:small-caps;">sis0-1</span> 34 29/11/1996 $ASCA$ <span style="font-variant:small-caps;">sis0-1</span> 27 <span style="font-variant:small-caps;">lecs</span> 36 24/04/1998 BeppoSAX <span style="font-variant:small-caps;">mecs</span> 77 <span style="font-variant:small-caps;">pds</span> 33 14/11/2000 $Chandra$ <span style="font-variant:small-caps;">acis-s hetg</span> 76 13/05/2001 XMM-$Newton$ <span style="font-variant:small-caps;">epic-pn</span> 38 (6.0) 01/12/2001 XMM-$Newton$ <span style="font-variant:small-caps;">epic-pn</span> 25 (19.1) : \[log\]The log of all analysed observations, and exposure times. Values in brackets show net exposure times after filtering for particle induced flares. \[data\]Observations and data reduction ======================================= ASCA ---- The source was observed by $ASCA$ twice, on May 1994 and on November 1996 (Table \[log\]). Data relative to the first observation were published by @weaver97, those relative to the second one were briefly mentioned by @weaver98. The spectra were extracted from the screened event files taken from the Goddard archive. In this paper we will only deal with the SIS0-1 spectra ($0.5-10$ keV). Data were analysed with HEAsoft 5.1 and <span style="font-variant:small-caps;">Xspec</span> 11.2.0. BeppoSAX -------- BeppoSAX observed this source on April 1998 (Table \[log\]). Results were published by @per02. Data reduction followed the standard procedure presented by @guainazzisax99, using HEAsoft 5.1. Spectra were extracted from regions of radius $8\arcmin$ for the LECS and $4\arcmin$ for the MECS, and analysed with <span style="font-variant:small-caps;">Xspec</span> 11.2.0. A normalization factor of 0.8 was adopted between the PDS and the MECS, appropriate for PDS spectra extracted with variable rise time threshold. The normalization of the LECS to the MECS was left as a free parameter to account for the different time coverage. \[chandra\]Chandra ------------------ The $Chandra$ observation was performed on November 2000 (Table \[log\]) with the Advanced CCD Imaging Spectrometer (ACIS-S) and the High-Energy Transmission Grating Spectrometer (HETGS) in place. The high flux of the source together with a default frame time of 3.2 s resulted in a 0th order spectrum which is strongly affected by pileup (80% according to WebPIMMS[^1]). Therefore, we will only use the 1st order co-added MEG ($0.4-5$ keV) and HEG ($0.8-10$ keV) spectra in this paper. Data were reduced with the Chandra Interactive Analysis of Observations software (CIAO 2.2.1), using the Chandra Calibration Database (CALDB 2.10). Grating spectra were analysed with $Sherpa$ 2.2.1. XMM-$Newton$ ------------ XMM-$Newton$ observed MCG –5-23-16 twice on 13 May 2001 and 1 December 2001 for 38 ks and 25 ks respectively (Table \[log\]). Both observations were performed in the Full Frame Mode using the blocking optical Medium Filter with the EPIC PN and MOS simultaneously operating. We will not deal with the RGS spectra in this paper because they offer poor statistics. Data were reduced with SAS 5.3.3. Events corresponding to pattern 0 were used for the PN and 0-12 for the MOS in both observations. The extracted background high energy ($\mathrm{E}>10$ keV) lightcurves clearly showed particle-induced flares, which were filtered adopting a 3 $\sigma$ threshold above the quiescent state background count rate. This resulted in net exposure times for the PN of 6.0 ks and 19.1 ks, respectively for the observations of May and December. In both observations the MOS count rates (3.5 s$^{-1}$ and 3.2 s$^{-1}$ for MOS1) are higher than the 1% pileup limit for this instrument (0.7 s$^{-1}$, see table 3 of the XMM-$Newton$ Users’ Handbook[^2]), therefore we will not deal with the MOS spectra in this paper. On the other hand, the PN count rates (7.5 s$^{-1}$ and 8.0 s$^{-1}$) are just within the 1% pileup limit (8 s$^{-1}$ for this instrument). EPIC-PN spectra ($0.5-10$ keV) and lightcurves were extracted from a radius of 40, and analysed with <span style="font-variant:small-caps;">Xspec</span> 11.2.0.\ In the following, errors are at the 90% confidence level for one interesting parameter ($\Delta \chi^2=2.71$). Data analysis ============= Flux and spectral variability ----------------------------- We analysed six X–ray observations of MCG –5-23-16, covering a history of seven years (see Table \[log\]). The $2-10$ keV flux (averaged over an entire observation) of the source has varied by about 25% in about 1 year (see Table \[flux\]). Large short-term variability is also present, as best illustrated by the BeppoSAX observation, which is the longest available (see Fig. \[hrmecs\]). Despite the remarkable flux variability, however, the hardness ratio ($4-10.5$ keV)/($1.5-4$ keV) is basically constant, implying that no spectral variability is present at these energies. Similar temporal behaviours are also found in the other observations. ----------------- ----------------------------------------------- **Observation** **Flux 2-10 keV** in units of $10^{-11}$ erg cm$^{-2}$ s$^{-2}$ $ASCA$ 1994 $9.24\pm0.26$ $ASCA$ 1996 $7.45\pm0.21$ BeppoSAX 1998 $9.30\pm0.07$ $Chandra$ 2000 $9.66\pm0.11$ XMM 05/2001 $8.08\pm0.01$ XMM 12/2001 $7.12\pm0.02$ ----------------- ----------------------------------------------- : \[flux\]The total flux ($2-10$ keV) as measured adopting the baseline model (described further in the text). Spectral analysis ----------------- We will only deal with data above 2.5 keV, because we are mainly interested in studying the reprocessing features, i.e. the iron lines and the Compton reflection component. Our baseline model consists of an exponentially cut-off power law and a reflection component from a neutral Compton-thick slab, isotropically illuminated by the primary radiation, subtending the solid angle $R=\Omega/2\pi$ and with an inclination angle $i$ with respect to the line of sight [<span style="font-variant:small-caps;">pexrav</span> model in <span style="font-variant:small-caps;">Xspec</span>: @mz95]. We kept the abundances of the heavy elements $Z$ fixed to the solar value. We also included in the model two components of the iron line to account for its complex profile (Figure \[iron\_xmm\]), as suggested both by $ASCA$ [@weaver97] and XMM-$Newton$ results [@dew03]: a narrow gaussian centred at 6.4 keV and a relativistically broadened iron line, produced by the accretion disc surrounding a Schwarzschild black hole [@fab89]. This choice will be further justified below. The <span style="font-variant:small-caps;">diskline</span> model parameterizes the radial line emissivity as a power law, i.e. $\propto r^{-q}$. We fixed the line rest energy to 6.4 keV, $q=-2$ and the inner radius $r\mathrm{_{in}}$ to 6 $r\mathrm{_{g}}$ (last stable orbit), while we kept the outer radius $r\mathrm{_{out}}$ as a free parameter. The inclination angle of the disc $i$ has been always linked to the same value of the <span style="font-variant:small-caps;">pexrav</span> model. This is equivalent to assume that both the broad line and the reflection component are emitted in the disc. (For this reason we also tried to fit the reflection component with the model <span style="font-variant:small-caps;">refsch</span> instead of <span style="font-variant:small-caps;">pexrav</span>, as the former includes relativistic effects. Both for the BeppoSAX and XMM-$Newton$ spectra, however, no significant differences are found in the values of the parameters and of the $\chi^{2}$ between the two models. Therefore, and for the sake of simplicity, we will adopt the <span style="font-variant:small-caps;">pexrav</span> model throughout this work). Finally we included a uniform cold absorber with column density N$\mathrm{_{H}}$ to take into account local absorption, in addition to the Galactic one N$\mathrm{_{g}}=8.12\times10^{20}$ cm$^{-2}$ (HEASARC W3nH[^3]). The analysis is made a bit complicated by the fact that, while the XMM-$Newton$ observation is definitely the best one to constrain the iron line properties, the BeppoSAX observations is the only capable to measure accurately enough the reflection component. We therefore decided to adopt an iterative procedure, consisting first in estimating the inclination angle from the relativistic line in the XMM-$Newton$ second, and longer, observation, then deriving $R$, with $i$ fixed, from the BeppoSAX spectrum, and finally coming back to the XMM-$Newton$ spectrum to check if, with the so found value of $R$, the inclination angle is still the same. Once that was verified, with $R=0.45$ and $i=42^{\circ}$ obtained from such procedure, we fitted all spectra with $R$ and $i$ fixed to these values. In performing the described procedure we found $r\mathrm{_{out}}$ to be very loosely constrained. We therefore fixed it in all fits to the value giving the lowest $\chi^2$, i.e. 400 $r\mathrm{_{g}}$. On the other hand, if $r\mathrm{_{in}}$ is left free to vary (with $r\mathrm{_{out}}$ fixed to 400 $r\mathrm{_{g}}$), no improvement in the $\chi^2$ is found, further justifying our choice of keeping $r\mathrm{_{in}}$ fixed to the innermost stable orbit (6 $r\mathrm{_{g}}$). The significance of the relativistic disc line is discussed in @dew03. Here, suffice it to say that fitting the spectrum with a narrow gaussian line results in $\chi^2/d.o.f.=243.4/158$; leaving the width of the line free to vary, the fit improves giving $\chi^2/d.o.f.=198.4/157$ (significant at more than 99.99% confidence level according to the F-test), with $\sigma=0.28\pm0.07$ keV (corresponding to a FWHM of $30\,800\pm7\,700$ km s$^{-1}$). Such a large width is hard to explain other than being due to relativistic and kinematic effects in the innermost regions of the accretion disc [see e.g. @Fabian00]. Indeed, a better fit ($\chi^2/d.o.f.=192.8/157$) is obtained using a relativistic line model, with all parameters fixed to the abovementioned values but $i$, left free to vary as said above. No significant further improvement is found adding a narrow line ($\chi^2/d.o.f.=191.5/156$). However, $Chandra$ HETGS, thanks to their excellent spectral resolution, clearly reveal the narrow iron line component at E$\mathrm{_{N}}=6.38\pm0.02$ keV (Figure \[iron\_chandra\]), which is unresolved even at the gratings resolution (FWHM$\lesssim6\,500$ km s$^{-1}$ at 99% confidence level). Therefore, given the $Chandra$ detection of a narrow component, we decided to add it to all fits. To illustrate the presence of the broad line, in Figure \[iron\_xmm\] the EPIC-PN residuals around the iron line energy for the second XMM-$Newton$ observation are shown, when fitting with a simple absorbed power law (upper panel), and with a power law plus a narrow Gaussian whose flux is fixed to $4.5\times10^{-5}$ ph cm$^{-2}$ s$^{-1}$ (lower panel; see below for the choice of this value). The broad line appears to be slightly asymmetric, consistently with the significant but not dramatic improvement in the $\chi^2$ when fitting with a relativistic line model rather than a broad Gaussian. The detected 2–10 keV flux variations between the different observations of only 25% at most (Table \[flux\]) should not result in any measurable variability on the amount of the reflection component. This is why we assumed $R$ to be constant and we fixed it to the BeppoSAX best fit value for all fits. In this respect, it is reassuring that the values of $\Gamma$ so obtained are all consistent with one another within the error bars. The cut-off energy E$\mathrm{_{c}}$ has also been assumed to be constant (E$\mathrm{_{c}}=110$ keV). Table \[PX\] shows the results obtained with the baseline model applied to the six observations considered. **ASCA 1994** **ASCA 1996** **BeppoSAX 1998** **Chandra 2000** **XMM 05/2001** **XMM 12/2001** --------------------------------------------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ --------------------- N$\mathrm{_H}$ ($10^{22}$ cm$^{-2}$) $0.89^{+0.36}_{-0.35}$ $1.21^{+0.46}_{-0.44}$ $1.57^{+0.29}_{-0.31}$ $1.25^{+0.29}_{-0.18}$ $1.94^{+0.38}_{-0.40}$ $1.80\pm0.23$ $\Gamma$ $1.65^{+0.07}_{-0.09}$ $1.69\pm0.10$ $1.73\pm0.07$ $1.70^{+0.06}_{-0.09}$ $1.77\pm0.07$ $1.74\pm0.04$ $R$ $0.45^*$ $0.45^*$ $0.45^{+0.22}_{-0.17}$ $0.45^*$ $0.45^*$ $0.45^*$ E$\mathrm{_{c}}$ (keV) $110^*$ $110^*$ $110^{+43}_{-21}$ $110^*$ $110^*$ $110^*$ I$\mathrm{_{N}}$ ($10^{-5}$ph cm$^{-2}$ s$^{-1}$) $6.8^{+1.9}_{-3.4}$ $4.8^{+3.5}_{-3.1}$ $7.7^{+4.3}_{-5.1}$ $8.6\pm3.5$ $4.1^{+2.3}_{-1.5}$ $4.3^{+0.9}_{-1.0}$ EW$\mathrm{_{N}}$ (eV) $58^{+16}_{-29}$ $50^{+36}_{-32}$ $65^{+36}_{-43}$ $70\pm28$ $38^{+21}_{-14}$ $45^{+9}_{-10}$ I$\mathrm{_{D}}$ ($10^{-5}$ph cm$^{-2}$ s$^{-1}$) $12.6^{+9.6}_{-7.6}$ $9.5^{+7.9}_{-5.2}$ $<15.4$ $<17.6$ $11.1^{+4.6}_{-6.2}$ $9.9^{+2.8}_{-2.5}$ EW$\mathrm{_{D}}$ (eV) $114^{+87}_{-69}$ $107^{+89}_{-59}$ $<134$ $<153$ $114^{+47}_{-64}$ $116^{+33}_{-29}$ $\chi^2$/dof 355.4/371 359.1/341 87.8/83 252.6/428 165.9/140 199.5/157 NOTE.–The following parameters are fixed: iron line energy E$\mathrm{_{\alpha}}=6.4$ keV; disc inclination angle $i=42\degr$ (both in <span style="font-variant:small-caps;">pexrav</span> and disc line model); disc line emissivity $q=-2$; inner and outer disc radius $r\mathrm{_{in}}=6~r\mathrm{_{g}}$ and $r\mathrm{_{out}}=400~r\mathrm{_{g}}$ respectively.\ $^*$ denotes fixed parameters.\ **ASCA 1994** **ASCA 1996** **BeppoSAX 1998** **Chandra 2000** **XMM 05/2001** **XMM 12/2001** --------------------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- --------------------- --------------------- I$\mathrm{_{D}}$ ($10^{-5}$ph cm$^{-2}$ s$^{-1}$) $13.5^{+6.8}_{-5.5}$ $10.2^{+7.7}_{-5.4}$ $10.8^{+4.4}_{-3.4}$ $10.3^{+9.6}_{-7.7}$ $9.8^{+5.5}_{-4.9}$ $9.8^{+2.6}_{-2.7}$ EW$\mathrm{_{D}}$ (eV) $127^{+64}_{-52}$ $116^{+88}_{-61}$ $97^{+40}_{-31}$ $91^{+85}_{-68}$ $101^{+57}_{-50}$ $114^{+30}_{-31}$ $\chi^2$/dof 356.1/372 359.1/342 88.4/84 255.4/429 165.9/141 199.6/158 In Figures \[flux64\] and \[broad\] the flux history of the narrow and broad line components, respectively, is shown. Both components are consistent with being constant within the, admittedly large, errors. In both cases this is not surprising, given the small flux variations in the continuum. For the narrow component, which is expected to originate in distant matter, this is also expected on theoretical grounds. To reduce the error bars on the broad component, we refitted all observations with the flux of the narrow component fixed to its weighted mean, i.e. $4.5\times10^{-5}$ ph cm$^{-2}$ s$^{-1}$. New results are summarized in Table \[PXF\] and illustrated in Figure \[broad2\]. No significant changes are found for the other parameters. \[discussion\]Discussion ======================== We have reanalysed the XMM-$Newton$ observations of MCG –5-23-16, as well as previous ASCA, BeppoSAX and $Chandra$ ones, confirming the presence of a significantly broad line already reported by @dew03. A fit with a relativistic disc profile is somewhat better than that with a broad Gaussian. Moreover, a so broad line ($\sigma=0.28$ keV) is hard to form other than in a relativistic disc [see e.g. @Fabian00], which we therefore consider the most likely explanation. A narrow line is also present, as clearly shown by $Chandra$, as well as a Compton reflection component, measured by BeppoSAX. Exploiting the modest long term flux variability, we could confidently fit the XMM-$Newton$ spectrum adding these two further components (with the parameter $R$ describing the relative amount of reflection component kept fixed to the value found by BeppoSAX). The detection of a broad, likely relativistic line in this source is important. Predicted on theoretical grounds, relativistic lines were found to be common in ASCA spectra of AGN [@Nandra97]. However, @Lub01 showed that at least part of the lines could be ascribed to a narrow component from distant matter, also to be expected on the basis of Unification Models [e.g. @Ghi94]. $Chandra$ and XMM-$Newton$ have indeed confirmed the almost ubiquity of the narrow component, while relativistic lines have been unambiguously detected so far only in a few cases [e.g. @Wilms01; @Turner02]. The discovery of a relativistic line in MCG –5-23-16 adds therefore one more case to a still short list. It is worth remarking that the relativistic lines have been observed so far in sources in which either the EW of the line is very large, as in the case of MCG –6-30-15 and NGC 3516 [@Turner02], or that are very bright, like MCG –5-23-16. The presence of the narrow component, in fact, could make difficult to detect the broad component in not very bright sources, if its EW is modest. Even if in some cases, e.g. NGC 5506 [@Matt01; @Bianchi03], NGC 7213 [@b03], NGC 5548 [@Pounds03] and NGC 4151 [@Schurch03], relativistic lines with significant equivalent widths are definitely absent in high signal-to-noise XMM-$Newton$ spectra, the actual frequency of relativistic lines is still to be assessed. The amount of reflection component, $R=0.45$, is low if compared with the total EW of the lines, which is about 170 eV [e.g. @matt91; @gf91]. One possible explanation is iron overabundance [@Matt97]; alternatively, one of the two line components may originate in Compton-thin matter. The obvious candidate is the narrow component (the broad one being likely emitted by the accretion disc), which therefore could be associated with the BLR rather than the molecular torus. (It is worth noting that the column density of the absorber is too low to make it a good candidate for producing the line [e.g. @Matt03]). This hypothesis cannot be directly tested, because MCG –5-23-16 is a NELG and therefore the BLR is supposed to be obscured in the optical band, and therefore no BLR line width, to be compared with the upper limit derived from $Chandra$, is available. Alternatively, the line can be emitted in a Compton-thin torus. If indeed the narrow line is emitted in Compton-thin matter, no evidence for a Compton-thick torus is present in this source, similar to the case of NGC 7213 [@b03]. We thank G.C. Perola for useful discussions on BeppoSAX data analysis, Kazushi Iwasawa for his help on reducing $ASCA$ data and the anonymous referee for helpful comments and suggestions. We acknowledge financial support from ASI. [^1]: http://heasarc.gsfc.nasa.gov/Tools/w3pimms.html [^2]: http://xmm.vilspa.esa.es/external/xmm\_user\_support/\ documentation/uhb\_2.1/XMM\_UHB.html [^3]: http://heasarc.gsfc.nasa.gov/cgi-bin/Tools/w3nh/w3nh.pl

ArXiv

--- author: - '[^1]' title: | Bare Demo of IEEEtran.cls for\ IEEE <span style="font-variant:small-caps;">Transactions on Magnetics</span> --- [Shell : Bare Demo of IEEEtran.cls for IEEE Transactions on Magnetics Journals]{} Introduction ============ demo file is intended to serve as a “starter file” for IEEE <span style="font-variant:small-caps;">Transactions on Magnetics</span> journal papers produced under LaTeX using IEEEtran.cls version 1.8b and later. I wish you the best of success. mds August 26, 2015 Subsection Heading Here ----------------------- Subsection text here. ### Subsubsection Heading Here Subsubsection text here. Conclusion ========== The conclusion goes here. Proof of the First Zonklar Equation =================================== Appendix one text goes here. Appendix two text goes here. Acknowledgment {#acknowledgment .unnumbered} ============== The authors would like to thank... [1]{} H. Kopka and P. W. Daly, *A Guide to LaTeX*, 3rd ed.1em plus 0.5em minus 0.4emHarlow, England: Addison-Wesley, 1999. [Michael Shell]{} Biography text here. [John Doe]{} Biography text here. [Jane Doe]{} Biography text here. [^1]: Manuscript received December 1, 2012; revised August 26, 2015. Corresponding author: M. Shell (email: http://www.michaelshell.org/contact.html).

ArXiv

--- author: - | S Chandra Mouli\ Department of Computer Science\ Purdue University\ `chandr@purdue.edu`\ Leonardo Teixeira\ Department of Computer Science\ Purdue University\ `lteixeir@purdue.edu` Jennifer Neville\ Department of Computer Science\ Purdue University\ `neville@cs.purdue.edu`\ Bruno Ribeiro\ Department of Computer Science\ Purdue University\ `ribeiro@cs.purdue.edu` bibliography: - 'paper.bib' title: Deep Lifetime Clustering ---

ArXiv

--- address: 'ITEP, B.Cheremushkinskaya,25, Moscow, Russia' author: - 'K.G. SELIVANOV' title: 'POST-CLASSICISM IN TREE AMPLITUDES' --- \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{} I would like to present here work done in collaboration with Alexei Rosly. This is devoted to solutions of field equations which describe tree amplitudes with any number of external legs. In works [@KS1], [@RS1], [@KS2] such solutions were constructed in Yang-Mills (see also a closely related work [@Ba]). In works [@RS2], [@KS3], [@KS4] such solutions were constructed in gravity and in Yang-Mills interacting with gravity. Work [@RS3] is about the sin(h)-Gordon case and work [@KS5] deals with the case of SUSY Yang-Mills. Unfortunately, because of lack of the time and because of the interest of the audience, I shall not be able to explain the constructions itself - you will have to trust me that they are beautiful. I shall only explain what type of solutions we construct. So to say, I shall only formulate the problem, a solution for which can be found in the references above. First of all, I, perhaps, need to explain that multi-leg amplitudes are non-trivial objects even in the tree approximation. The problem is that, although every given diagram gives quite a simple contribution, an algebraic function of external momenta, the number of different diagrams growths enormously with the number of legs, and all these algebraic functions are different, so that the whole thing becomes untreatable when the number of legs becomes bigger than, say, 10, to say nothing about arbitrary number of legs. In principle, one could take this as that the nature is such that the multi-leg amplitudes are complicated objects and nothing to do about it. However, there are known cases when the final expression for the amplitudes - if it is available at all - turns out to be much simpler than intermediate steps. For instance, in [@V], [@S], [@A] it was noticed that in ${\phi}^{4}$ theory all amplitudes $2 {\rightarrow} n$ for $n$ bigger than 4 vanish at threshold. Notice that every given diagram gives a nontrivial contribution. That is only sum of them what vanishes. Other example of the case when the final expression is much simpler than intermediate steps are the so-called Parke-Taylor, or “maximally helicity violating” amplitudes in Yang-Mills [@PT], [@BG]. These are amplitudes with two positive-helicity gluons in the initial state and any number of the positive-helicity gluons in the final state. Kinematics is arbitrary. The explicit expressions for these amplitudes were conjectured in [@PT] and were proven in [@BG]. I am not giving here those nice expressions, I just notice that they contain only the pairwise collinear singularities, those of the type of $\frac{1}{(p_{i},p_{j})}$, where $p_{i}$ stands for momentum of $i$-th particle, while a separate diagram definitely gives much more singularities. These remarkable cancellations above motivated us for the work presented here. Our initial point is that tree amplitudes are related to a solution of the classical field equation of the model. To make the relation precise we need to consider not the amplitudes themself, but form-factors, that is, would be amplitudes with one unamputated off-shell leg and a number of on-shell ones, which are amputated as they should be according to the LSZ rules. It is convenient to take the off-shell leg in the coordinate $(x)$ representation. Of course, the amplitudes can be obtained from the form-factors applying the LSZ rules to the off-shell leg. The form-factors can also be written as matrix elements of the field operator between vacuum and $n$-particle state, $$\label{ff} <p_{1}, \ldots , p_{l}|{\phi}(x)|0>_{tree}$$ where the subscript indicates that we are interested only in the tree contributions. At tree level, when analytical structure of the form-factor is very simple (it is just an algebraic function of the momenta), we can afford not to make a difference between in- and out- states. Notice that $x$-dependence of the form-factors is very simple, it is just a product of the plane waves corresponding to the momenta of external on-shell legs. This follows from the fact that in momentum representation, the dependence on the momentum of the off-shell leg is given by the overall conservation-law ${\delta}$-function: $$\label{harmonics} <p_{1}, \ldots , p_{l}|{\phi}(x)|0>_{tree}= {\int}dp' e^{ip'x} {\delta}(p'-{\sum}_{j}p_{j})( \ldots )= e^{i{\sum}_{j}p_{j}x}(\;a\;function\;of\;\{p_{j}\})$$ Obviously, an individual form-factor does not obey field equations. What obeys field equations is a generating function for form-factors which we define next. To do this we\ 1)fix a set of momenta $\{p_{j}\}: \; p_{1}, \ldots , p_{N}$;\ 2)introduce corresponding set of parameters $\{a_{j}\}:\; a_{1}, \ldots , a_{N}$;\ 3)define a function of $x$, $\{p_{j}\}$ and $\{a_{j}\}$ such that the individual form-factors appears as coefficients in the Taylor expansion of this function in powers of $a$’s: $$\label{ptb} {\Phi}(x, \{p\}, \{a\})=\sum_{l=1}^{L}\sum_{\{J\}}a_{J_{1}} \ldots a_{J_{l}}<p_{J_{1}}, \ldots , p_{J_{l}}|{\phi}(x)|0>_{tree}$$ where the one-particle form-factor is clearly $$\label{norm} <p|{\phi}(x)|0>=( \ldots ) e^{ipx}$$ $( \ldots )$ in Eq.(\[norm\]) stands for a polarization factors in a non-scalar case, as well as for a color matrix in the Yang-Mills case, etc. So defined generating function can be seen to obey field equations of the model. Indeed, applying the inverse propagator to the off-shell leg of an individual form-factor one obtains a sum of products of two, of three and so on - corresponding to the types of vertices in the theory - form-factors with smaller amount of legs. This relation can be used as a recursion relation. On the other hand, this relation is seen to be equivalent to field equations on the generating function: $$\label{eq} \left( \frac{{\partial}^{2}}{{\partial}x^{2}} + m^{2} \right) {\Phi}(x,\{p\},\{a\})={\lambda}_{3}{\Phi}^{2}+{\lambda}_{4}{\Phi}^{3}+ \ldots$$ Indeed, expanding Eq.(\[eq\]) in powers of $a$’s one obtains the recursion relation among form-factors. (Eq.(\[eq\]) is written for a scalar theory just for the brevity.) Having started with the form-factors we have proved that the generating function obeys field equations. We can reverse the logic and use the field equations to find ${\Phi}(x,\{p\},\{a\})$. To do this we must specify what solution to pick up, and according to the definition above we should pick up a solution which is a power series in the set of variables $$\label{harmonics2} {\cal E}_{j}=a_{j}e^{ip_{j}x}, \; j=1, \ldots, N$$ starting with first order terms of the type of $$\label{expansion} {\Phi}(x,\{p\},\{a\})={\sum}_{j=1}^{N}( \ldots){\cal E}_{j} + \; higher\; order\; terms\; in\; \{{\cal E}_{j}\}$$ $( \ldots)$ here is the same as in Eq.(\[norm\]). Clearly, solution of this type exists and is unique (in gauge theories it is unique after a gauge fixing, or, equivalently, it is unique modulo gauge transformations), provided the operator on r.h.s. of Eq.(\[eq\]) is invertible on the power series in the variables ${\cal E}_{j}$, which is true when the [*nonresonantness condition*]{} is satisfied, that is, none of linear combinations with integer coefficients of the momenta $p_{j}$ from the given set gets to mass shell, $$\label{nr} ({{\sum}_{j}n_{j}p_{j}})^{2}{\neq}m^{2}$$ This is simply a condition that there are no internal lines on-shell in the Feynman diagrams. It is also convenient to impose the [*nilpotency*]{} condition on the parameters $a_{j}$, $$\label{nil} a_{j}^{2}=0, \; j=1, \ldots, N$$ (note that in bosonic case still $a_{i}a_{j}=a_{j}a_{i}$) which is equivalent to excluding form-factors with identical particles from the generating function. In the massive case this condition is very convenient technically, while in the massless case it is even problematic to proceed without the nilpotency, because in the massless case identical particles necessary break nonresonantness. I would like to stress at this point that we have introduced a class of solutions of field equations which are as universal as, say, solitonic solutions. We called these solutions [*perturbiners*]{}. In the classical text-books on quantum field theory, e.g. [@FS],[@IZ], one can also find a discussion of solutions generating tree amplitudes, but they are defined differently, with use of asymptotic Feynman-type conditions. That definition is not very convenient, and it is not strange that no explicit examples were given in those books. In the papers cited at the beginning of this talk we constructed perturbiners in various theories. Unfortunately, in Yang-Mills and in gravity the generic perturbiner is unavailable so far. What allows one to proceed is the restriction on polarizations of the external on-shell particles in the form-factors. Restricted in this way perturbiner obeys not the full Yang-Mill (Einstein) equations, but just the self-duality equations instead. It generates just the same-helicity form-factors. Having obtained the specific perturbiner with only same-helicity form-factors included, one can add opposite helicity particles perturbatively. This has been done in [@RS1] in order to describe the Parke-Taylor amplitudes and in [@KS4] to generalize the Parke-Taylor amplitudes to include any number of the same helicity gravitons in addition to the gluons. Our main tool was the so-called zero-curvature representation of the field equations. For the self-duality equations (in Yang-Mills and in gravity) and for the sin(h)-Gordon the zero-curvature representation is one-dimensional and this the case when perturbiner is constructed very efficiently. Remarkably, same-helicity form-factors in Yang-Mills and generic (tree) form-factors in sin(h)-Gordon are described in the same way. Perhaps, this opens a perspective to find Yang-Mills avatars of the many known quantum exact results in sin(h)-Gordon. In [@KS5] a progress has been achieved about generic perturbiner in $N=3$ SUSY Yang-Mills (which, of course, contains complete information about all tree form-factors of non-supersymmetric Yang-Mills). However, the zero-curvature representation is two-dimensional in that case, the construction is much more involved, and the complete solution has not been obtained yet. In conclusion, I would like to stress that perturbiners provide probably the simplest way of describing tree amplitudes, involve nice mathematical constructions and reveal similar structures underlying different theories. I hope they will find their place in future developments. Acknowledgments {#acknowledgments .unnumbered} =============== I would like to acknowledge Russian Ministry of Sciences and organizers of the conference for a financial support. References {#references .unnumbered} ========== [99]{} K.G.Selivanov, ITEP-21-96, hep-ph/9604206 A.A.Rosly and K.G.Selivanov, K.G.Selivanov, Talk given at International Europhysics Conference on High-Energy Physics (HEP 97), Jerusalem, Israel, 19-26 Aug 1997. W.Bardeen, [*Prog.Theor.Phys.Suppl.*]{} 123 (1996) 1 A.A.Rosly and K.G.Selivanov, ITEP-TH-56-97, hep-th/9710196 K.G.Selivanov, K.G.Selivanov, [*Mod.Phys.Lett.A*]{} 12 [1997]{} 3087 A.A.Rosly and K.G.Selivanov, K.G.Selivanov, ITEP-TH-47-98, hep-th/9809046 M.B.Voloshin, B.Smith, E.Argyres, R.Kleiss and C.Papadopoulos, S.Parke and T.Taylor, F.Berends and W.Giele, L.D.Faddeev and A.A.Slavnov, Introduction to the Theory\ of Quantum Gauge Fields, Nauka, Moscow, 1978 C.Itzykson and J.B.Zuber, Quantum Field Theory\ New York, Usa: Mcgraw-hill (1980)

ArXiv

--- abstract: 'We characterize the validity of the Whitney extension theorem in the ultradifferentiable Roumieu setting with controlled loss of regularity. Specifically, we show that in the main Theorem 1.3 of [@RainerSchindl17] condition (1.3) can be dropped. Moreover, we clarify some questions that remained open in [@RainerSchindl17].' address: - 'A. Rainer: Fakultät für Mathematik, Universität Wien, Oskar-Morgenstern-Platz 1, A-1090 Wien, Austria' - 'G. Schindl: Departamento de Álgebra, Análisis Matemático, Geometría y Topología, Facultad de Ciencias, Universidad de Valladolid, Paseo de Belén 7, 47011 Valladolid, Spain' author: - Armin Rainer - Gerhard Schindl title: 'On the extension of Whitney ultrajets, II' --- [^1] Introduction ============ The main goal of this paper is to prove: \[mainadd\] Let ${\omega}$ be a non-quasianalytic concave weight function. Let ${\sigma}$ be a weight function satisfying ${\sigma}(t) = o(t)$ as $t \to \infty$. Then the following conditions are equivalent: 1. For every compact $E \subseteq {\mathbb{R}}^n$ we have $j^\infty_E({\mathcal{B}}^{\{{\omega}\}}({\mathbb{R}}^n)) \supseteq {\mathcal{B}}^{\{{\sigma}\}}(E)$, where $j_E^\infty$ assigns to each $f \in C^\infty({\mathbb{R}}^n)$ its infinite jet $(f^{({\alpha})}|_E)_{{\alpha}\in {\mathbb{N}}^n}$ on $E$. 2. There is $C>0$ such that $\int_{1}^\infty \frac{{\omega}(tu)}{u^2}\,du \le C{\sigma}(t) + C$ for all $t>0$. (Here ${\mathcal{B}}^{\{{\omega}\}}$ denotes the Roumieu class defined by the weight function ${\omega}$; we use the symbol ${\mathcal{B}}$ to emphasize that the defining estimates are global, cf. [@RainerSchindl17 2.2 and 2.6].) It means that Theorem 1.3 of [@RainerSchindl17] holds without the assumption (1.3) that the associated weight matrix ${\mathfrak{S}}$ of ${\sigma}$ satisfies $$\label{intro:good} {\;\forall}S \in {\mathfrak{S}}{\;\exists}T \in {\mathfrak{S}}{\;\exists}C \ge 1 {\;\forall}1\le j \le k : \frac{S_j}{jS_{j-1}} \le C\, \frac{T_k}{k T_{k-1}}.$$ is proved in . In we clarify some questions that remained open in [@RainerSchindl17] and obtain several characterizations of concave weight functions. For an overview of the background of we refer to the introduction in [@RainerSchindl17]. We use the notation and the definitions of said paper; the concept of *weight matrices* is recalled in the appendix at the end of this paper. Note that in the special case that ${\omega}$ and ${\sigma}$ coincide we recover the result of [@BBMT91]: \[thm:preservingclass\] Let ${\omega}$ be a weight function. The following conditions are equivalent: 1. For every compact $E \subseteq {\mathbb{R}}^n$ we have $j^\infty_E({\mathcal{B}}^{\{{\omega}\}}({\mathbb{R}}^n)) = {\mathcal{B}}^{\{{\omega}\}}(E)$. 2. There is $C>0$ such that $\int_{1}^\infty \frac{{\omega}(tu)}{u^2}\,du \le C{\omega}(t) + C$ for all $t>0$. Indeed, if ${\omega}$ satisfies (ii’) then it is non-quasianalytic, equivalent to a concave weight function [@MeiseTaylor88 Proposition 1.3], and ${\omega}(t) = o(t)$ as $t \to \infty$ [@BonetMeiseTaylor92 Remark 3.20]. That (ii’) is a necessary condition for (i’) is well-known. Note that also (i’) implies that ${\omega}$ is non-quasianalytic. Indeed, if ${\omega}$ is quasianalytic, then the Borel map $j^\infty_{\{0\}} : {\mathcal{B}}^{\{{\omega}\}}({\mathbb{R}}^n) \to {\mathcal{B}}^{\{{\omega}\}}(\{0\})$ is never surjective. For $t \ne O({\omega}(t))$ as $t \to \infty$ this follows from [@RainerSchindl15], for $t = O({\omega}(t))$ as $t \to \infty$ consider e.g. the formal series $\sum_{k=0}^\infty x^k$ which converges to the unbounded real analytic function $1/(1-x)$ function for $|x|<1$. Proof of Theorem 1 {#sec:proof} ================== Preparations {#preparations .unnumbered} ------------ First we recall a few definitions and facts. Let $m =(m_k)$ be a positive sequence satisfying $m_0 = 1$ and $m_k^{1/k} \to \infty$. The *log-convex minorant* of $m$ is given by $${\underline}m_k := \sup_{t>0} \frac{t^k}{\exp({\omega}_m(t))}, \quad k \in {\mathbb{N}},$$ where $${\omega}_m (t) := \sup_{k \in {\mathbb{N}}} \log \Big(\frac{t^k}{m_k}\Big), \quad t>0.$$ The function ${\omega}_m$ is increasing, convex in $\log t$, and zero for sufficiently small $t>0$. Related is the function $h_m(t) := \inf_{k \in {\mathbb{N}}} m_k t^k$, for $t>0$, and $h_m(0):=0$. It is increasing, continuous, positive for $t>0$, and equals $1$ for large $t$. Let $m=(m_k)$ be a positive *log-convex* sequence (i.e., $m = {\underline}m$) such that $m_0 =1$ and $m_k^{1/k} \to \infty$. Then the functions ${\overline}{\Gamma}_m$ and ${\underline}{\Gamma}_m$ defined in [@RainerSchindl17 Definition 3.1] coincide, we simply write ${\Gamma}_m$ in this case; note that log-convexity and $m_k^{1/k} \to \infty$ imply $m_{k}/m_{k-1} \to \infty$. Thus $$\label{eq:Gamma} {\Gamma}_m(t) = \min\{ k : h_m(t) = m_k t^k\} = \min \Big\{k : \frac{m_{k+1}}{m_k} \ge \frac1t\Big\}, \quad t>0.$$ By [@RainerSchindl17 Lemma 3.2], ${\Gamma}_m$ is decreasing, tending to $\infty$ as $t \to 0$, and $$\label{eq:Gamma1} k \mapsto m_k t^k \text{ is decreasing for } k \le {\Gamma}_m(t).$$ Recall that with every weight function ${\sigma}$ (always understood as defined in [@RainerSchindl17 Section 2.1]) is associated a weight matrix ${\mathfrak{S}}= \{S^\xi\}_{\xi>0}$, where $$S^\xi_k := \exp\big(\tfrac1\xi {\varphi}^*(\xi k\big)), \quad \text{ (here ${\varphi}= {\sigma}{\circ}\exp$ and ${\varphi}^*$ is its Young conjugate), }$$ such that ${\mathcal{B}}^{\{{\sigma}\}} = {\mathcal{B}}^{\{{\mathfrak{S}}\}}$ and ${\mathcal{B}}^{({\sigma})} = {\mathcal{B}}^{({\mathfrak{S}})}$ algebraically and topologically; cf. [@RainerSchindl17 2.5] and [@RainerSchindl12]. In the following we set $s^\xi_k := S^\xi_k/k!$. The next proposition shows that for a weight function ${\sigma}$ which is equivalent to a concave weight function and satisfies ${\sigma}(t) = o(t)$ as $t \to \infty$ we additionally have ${\mathcal{B}}^{\{{\sigma}\}} = {\mathcal{B}}^{\{{\underline}{\mathfrak{S}}\}}$ and ${\mathcal{B}}^{({\sigma})} = {\mathcal{B}}^{({\underline}{\mathfrak{S}})}$, where ${\underline}{\mathfrak{S}}= \{{\underline}S^\xi\}_{\xi>0}$ and $${\underline}S^\xi_k:= k!\, {\underline}s^\xi_k.$$ In particular, ${\underline}{\mathfrak{S}}$ satisfies . We say that ${\underline}S^\xi$ is *strongly log-convex* meaning that ${\underline}s^\xi_k={\underline}S^\xi_k/k!$ is log-convex. (Note the abuse of notation: ${\underline}S^\xi$ is *not* necessarily the log-convex minorant of $S^\xi$; this will cause no confusion.) Recall that two weight functions ${\omega}$ and ${\sigma}$ are called *equivalent* if ${\omega}(t)= O({\sigma}(t))$ and ${\sigma}(t) = O ({\omega}(t))$ as $t \to \infty$; this means that they define the same ultradifferentiable class. \[prop:strongmatrix\] Let ${\sigma}$ be a weight function satisfying ${\sigma}(t) = o(t)$ as $t \to \infty$ which is equivalent to a concave weight function. For each $\xi>0$ there exist constants $A,B,C >0$ such that $$\label{eq:strong} A^{-1} s^{\xi/B}_k \le {\underline}s^\xi_k \le s^\xi_k \le C^k {\underline}s^{B\xi}_k \quad \text{ for all } k \in {\mathbb{N}}.$$ Moreover, there is a constant $H\ge 1$ such that ${\underline}s^\xi_{j+k} \le H^{j+k} {\underline}s^{2\xi}_{j} {\underline}s^{2\xi}_{k}$, for all $\xi>0$ and all $j,k \in {\mathbb{N}}$, and thus $h_{{\underline}s^{\xi}}(t) \le h_{{\underline}s^{2\xi}}(Ht)^2$, for all $\xi>0$ and all $t>0$. Clearly, ${\underline}s^\xi \le s^\xi$. Let ${\underline}S^\xi_k:= k!\, {\underline}s^\xi_k$. By [@Jimenez-GarridoSanzSchindl17 Lemma 3.6], ${\omega}_{S^\xi}$ and ${\omega}_{{\underline}S^\xi}$ are equivalent, in particular, there exists $C\ge 1$ such that $$\label{eq:equi} {\omega}_{{\underline}S^\xi} \le C {\omega}_{S^\xi} + C.$$ By [@RainerSchindl17 Lemma 2.4(3)] and [@RainerSchindl16a Remark 2.5], we have $$ 2 {\omega}_{S^{2\xi}} \le {\omega}_{S^\xi}, \quad \text{ for all } \xi>0.$$ If $n$ is an integer such that $B:= 2^n \ge C$, then ${\omega}_{{\underline}S^\xi} \le {\omega}_{S^{\xi/B}} + C$ and hence $${\underline}S^\xi_k = \sup_{t>0} \frac{t^k}{\exp({\omega}_{{\underline}S^\xi}(t))} \ge e^{-C} \sup_{t>0} \frac{t^k}{\exp({\omega}_{S^{\xi/B}}(t))} = e^{-C} S^{\xi/B}_k.$$ This shows the first inequality in . By [@RainerSchindl17 Lemma 3.13], there exists $D\ge 1$ such that for all $\xi>0$, $$ 2 {\omega}_{s^{2\xi}}(t) \le {\omega}_{s^\xi}(Dt), \quad \text{ for } t>0$$ and therefore $$\begin{aligned} {\underline}s^\xi_{2k} = \sup_{t>0} \frac{(Dt)^{2k}}{\exp({\omega}_{s^\xi}(Dt))} \le D^{2k} \sup_{t>0} \frac{t^{2k}}{\exp(2{\omega}_{s^{2\xi}}(t))} = D^{2k} ({\underline}s^{2\xi}_{k})^2. \end{aligned}$$ Thus, by [@Schindl14 Theorem 9.5.1] (which is a generalization of [@Matsumoto84]), there exists a constant $H\ge 1$ such that ${\underline}s^\xi_{j+k} \le H^{j+k} {\underline}s^{2\xi}_{j} {\underline}s^{2\xi}_{k}$, for all $j,k$. That $h_{{\underline}s^{\xi}}(t) \le h_{{\underline}s^{2\xi}}(Ht)^2$, for all $\xi>0$ and all $t>0$, follows from [@RainerSchindl17 Lemma 3.12]. By [@Schindl15 Proposition 3.6], $$2 {\omega}_{{\underline}S^{2\xi}}(t) \le {\omega}_{{\underline}S^\xi}(Ht), \quad \text{ for } t>0,$$ for some (possibly different) $H\ge 1$. As above, using , we find ${\omega}_{{\underline}S^{B \xi}}(b t) \le {\omega}_{S^\xi}(t) + 1$ for some constant $0< b \le 1$. Then $$\begin{aligned} {\underline}S^{B \xi}_k = \sup_{t>0} \frac{(bt)^k}{\exp({\omega}_{{\underline}S^{B\xi}}(bt))} \ge e^{-1} b^k \sup_{t>0} \frac{t^k}{\exp({\omega}_{S^{\xi}}(t))} = e^{-1}b^k S^\xi_k. \end{aligned}$$ The last inequality of follows. alone is not enough to get rid of the assumption . It is not clear that ${\underline}{\mathfrak{S}}$ has the property that for all $S \in {\underline}{\mathfrak{S}}$ there is a $T \in {\underline}{\mathfrak{S}}$ such that $S_{2k}/S_{2k-1} \lesssim T_{k}/T_{k-1}$. Note that ${\mathfrak{S}}$ has this property (see [@RainerSchindl17 Lemma 2.4(4)]) and it enters crucially in Lemma 3.4 and Proposition 3.7 of [@RainerSchindl17]. \[reasongood\] We deal with this problem by introducing another intimately related weight matrix ${\mathfrak{V}}:= \{V^\xi\}_{\xi>0}$. For each $\xi>0$ we define $V^\xi_k := k!\, v^\xi_k$ by setting $$\label{eq:save-1} v^\xi_k := \min_{0 \le j \le k} {\underline}s^{2\xi}_{j} {\underline}s^{2\xi}_{k-j}, \quad k \in {\mathbb{N}}.$$ That means that for the sequence of quotients $v^\xi_k/v^\xi_{k-1}$ we have (cf. [@Komatsu73 Lemma 3.5]) $$\Big(\frac{v^\xi_1}{v^\xi_{0}},\frac{v^\xi_2}{v^\xi_{1}},\frac{v^\xi_3}{v^\xi_{2}},\frac{v^\xi_4}{v^\xi_{3}}, \ldots\Big) = \Big(\frac{{\underline}s^{2\xi}_1}{{\underline}s^{2\xi}_{0}},\frac{{\underline}s^{2\xi}_1}{{\underline}s^{2\xi}_{0}},\frac{{\underline}s^{2\xi}_2}{{\underline}s^{2\xi}_{1}}, \frac{{\underline}s^{2\xi}_2}{{\underline}s^{2\xi}_{1}},\frac{{\underline}s^{2\xi}_3}{{\underline}s^{2\xi}_{2}}, \frac{{\underline}s^{2\xi}_3}{{\underline}s^{2\xi}_{2}}, \ldots\Big).$$ Thus the sequence $v^\xi = (v^\xi_k)$ is log-convex and satisfies $$\label{eq:save} \frac{v^\xi_{2k-1}}{v^\xi_{2k-2}} = \frac{v^\xi_{2k}}{v^\xi_{2k-1}} = \frac{{\underline}s^{2\xi}_{k}}{{\underline}s^{2\xi}_{k-1}}, \quad \text{ for all } k\ge 1.$$ So, in view of , $$\label{eq:save2} 2 {\Gamma}_{{\underline}s^{2\xi}}(t) = {\Gamma}_{v^\xi}(t), \quad \text{ for all } t>0.$$ By , there is $H\ge 1$ such that for all $\xi>0$ $$\label{eq:save1} {\underline}s^\xi_k \le H^k v^\xi_k \le H^k {\underline}s_k^{2\xi}, \quad \text{ for all } k \in {\mathbb{N}}.$$ Thus, we also have ${\mathcal{B}}^{\{{\sigma}\}} = {\mathcal{B}}^{\{{\mathfrak{V}}\}}$ and ${\mathcal{B}}^{({\sigma})} = {\mathcal{B}}^{({\mathfrak{V}})}$ algebraically and topologically. Proof of Theorem 1 {#proof-of-theorem-1 .unnumbered} ------------------ The implication (i) $\Rightarrow$ (ii) follows from [@BonetMeiseTaylor92]. So we only prove the converse implication. Condition (ii) means that the weight function $$\label{kappa} {\kappa}(t) := \int_1^\infty \frac{{\omega}(tu)}{u^2}\, du$$ satisfies ${\kappa}(t) = O({\sigma}(t))$ as $t\to \infty$, i.e., ${\mathcal{B}}^{\{{\sigma}\}} \subseteq {\mathcal{B}}^{\{{\kappa}\}}$. Now ${\kappa}$ is concave and ${\kappa}(t)= o(t)$ as $t \to \infty$, see [@MeiseTaylor88 Proposition 1.3]. We will show that Whitney ultrajets of class ${\mathcal{B}}^{\{{\kappa}\}}$ admit extensions of class ${\mathcal{B}}^{\{{\omega}\}}$. Thus from now on we assume without loss of generality that ${\sigma}= {\kappa}$ is concave. Since ${\omega}$ is increasing we have ${\sigma}={\kappa}\ge {\omega}$ and hence, if ${\mathfrak{W}}= \{W^\xi\}_{\xi>0}$ denotes the weight matrix associated with ${\omega}$, $$\label{eq:order} {\underline}S^{\xi} \le S^{\xi} \le W^\xi, \quad \text{ for all } \xi>0.$$ Moreover, as well as and apply. Let us now indicate the necessary changes in the proof of [@RainerSchindl17 Theorem 1.3]. The changes also lead to some simplifications. We provide details in the hope that this contributes to a better understanding. $\bullet$ Every Whitney ultrajet $F=(F^{\alpha})$ of class ${\mathcal{B}}^{\{{\sigma}\}}$ on the compact set $E \subseteq {\mathbb{R}}^n$ is an element of ${\mathcal{B}}^{\{V^\xi\}}(E)$ for some $\xi>0$, i.e., there exist $C>0$ and ${\rho}\ge 1$ such that $$\begin{gathered} |F^{\alpha}(a)| \le C {\rho}^{|{\alpha}|} \, V^\xi_{|{\alpha}|}, \quad {\alpha}\in {\mathbb{N}}^n,~ a \in E, \label{jets1} \\ |(R^k_a F)^{\alpha}(b)| \le C {\rho}^{k+1} \, |{\alpha}|!\, v^\xi_{k+1}\, |b-a|^{k+1-|{\alpha}|}, \quad k \in {\mathbb{N}},\, |{\alpha}| \le k,~ a,b \in E. \label{jets2}\end{gathered}$$ Let $p\in {\mathbb{N}}$ be fixed (and to be specified later). Let $\{{\varphi}_{i,p}\}_{i\in {\mathbb{N}}}$ be the partition of unity provided by [@RainerSchindl17 Proposition 4.9], relative to the family of cubes $\{Q_i\}_{i \in {\mathbb{N}}}$ from [@RainerSchindl17 Lemma 4.7], and let $r_0 = r_0(p)$ be the constant appearing in this proposition. The center of $Q_i$ is denoted by $x_i$. We claim that an extension of class ${\mathcal{B}}^{\{{\omega}\}}$ of $F$ to a suitable neighborhood of $E$ in ${\mathbb{R}}^n$ is provided by $$f(x) := \begin{cases} \sum_{i\in {\mathbb{N}}} {\varphi}_{i,p}(x) \, T_{\hat x_i}^{p(x_i) } F(x), & \text{ if } x \in {\mathbb{R}}^n \setminus E, \\ F^0(x), & \text{ if } x \in E, \end{cases}$$ where, given $x \in {\mathbb{R}}^n \setminus E$, $\hat x$ is any point in $E$ with $d(x) := d(x,E) = |x-\hat x|$ and $$p(x):= \max\{2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x)) -1,0\}.$$ Here $L$ is a positive constant to be specified below. Recall that $Q^*_i$ is the closed cube with the same center as $Q_i$ expanded by the factor $9/8$. By [@RainerSchindl17 Corollary 4.8], $$\label{cor48} \frac12 d(x) \le d(x_i) \le 3 d(x), \quad \text{ for all } x \in Q_i^*.$$ Then $d(x)<1/(3L{\underline}s^{2\xi}_1)$ guarantees that both ${\Gamma}_{{\underline}s^{2\xi}}(L d(x_i))$ and ${\Gamma}_{{\underline}s^{2\xi}}(L d(x))$ are $\ge 1$, by , thus $p(x_i)= 2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x_i)) -1$ and $p(x)= 2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x)) -1$. $\bullet$ Replace [@RainerSchindl17 Lemma 5.2] by the following lemma. The only difference in the proof is that one uses instead of [@RainerSchindl17 (5.4)]. \[proposition9\] There is a constant $C_0 = C_0(n) >1$ such that, for all Whitney ultrajets $F= (F^{\alpha})_{{\alpha}}$ of class ${\mathcal{B}}^{\{V^\xi\}}$ that satisfy and , all $L \ge C_0 {\rho}$, all $x \in {\mathbb{R}}^n$, and all ${\alpha}\in {\mathbb{N}}^n$, $$\begin{aligned} |(T_{\hat x}^{p(x)} F)^{({\alpha})}(x)| &\le C (2L)^{|{\alpha}|+1} V^\xi_{|{\alpha}|}, \label{prop91} \intertext{and, if $|{\alpha}| < p(x)$,} |(T_{\hat x}^{p(x)}F)^{({\alpha})}(x)-F^{\alpha}(\hat x)| &\le C (2L)^{|{\alpha}|+1} |{\alpha}|!\, v^\xi_{|{\alpha}|+1} d(x). \label{prop92}\end{aligned}$$ We remark that (here and below) by $(T_{\hat x}^{p(x)} F)^{({\alpha})}(x)$ we mean the ${\alpha}$-th partial derivative of the polynomial $y \mapsto T_{\hat x}^{p(x)} F(y)$ evaluated at $y =x$. $\bullet$ Replace [@RainerSchindl17 Lemma 5.3] by: \[lem:H1\] There is a constant $C_1 = C_1(n)>0$ such that for all $L>C_1 {\rho}$, all ${\beta}\in {\mathbb{N}}^n$, and all $x \in Q_i^*$ with $d(x)<1/(3L{\underline}s^{2\xi}_1)$, $$\begin{aligned} \label{H1} |{\partial}^{\beta}(T_{\hat x_i}^{p(x_i) } F - T_{\hat x}^{p(x_i)) } F) (x)| &\le C L^{|{\beta}|+1} {\underline}S^{2\xi}_{|{\beta}|} \, h_{{\underline}s^{2\xi}}(L d(x_i)).\end{aligned}$$ It suffices to consider $|{\beta}| \le p(x_i) = 2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x_i)) -1 =: 2q-1$. Let $H_1$ denote the left-hand side of . By [@RainerSchindl17 Lemma 5.1 and Corollary 4.8] and , $$\begin{aligned} H_1 \le C (2n^2 {\rho})^{2q} |{\beta}|! \, v^\xi_{2q} (6d(x_i))^{2q-|{\beta}|} \le C (2n^2 {\rho})^{2q} |{\beta}|! \, ({\underline}s^{2\xi}_{q})^2 (6 d(x_i))^{2q-|{\beta}|}.\end{aligned}$$ By the definition of $q$, $h_{{\underline}s^{2\xi}}(L d(x_i)) = {\underline}s^{2\xi}_{q} (L d(x_i))^{q} \le {\underline}s^{2\xi}_{k} (L d(x_i))^{k}$ for all $k$. Thus $$\begin{aligned} H_1 &\le C \Big(\frac{12n^2 {\rho}}{ L}\Big)^{2q} \, L^{|{\beta}|}\, |{\beta}|!\,{\underline}s^{2\xi}_{|{\beta}|} \, h_{{\underline}s^{2\xi}}(L d(x_i)).\end{aligned}$$ If $L > 12 n^2 \, {\rho}$, then follows. $\bullet$ Replace [@RainerSchindl17 Lemma 5.4] by: \[lem:H2\] There is a constant $C_2 = C_2(n)>0$ such that for all $L>C_2 {\rho}$, all ${\beta}\in {\mathbb{N}}^n$, and all $x \in Q_i^*$ with $d(x)<1/(3L{\underline}s^{2\xi}_1)$, $$\begin{aligned} \label{H2} |{\partial}^{\beta}(T_{\hat x}^{p(x_i)} F - T_{\hat x}^{p(x) } F) (x)| &\le C \Big(\frac{3L }{n}\Big)^{|{\beta}|+1} {\underline}S^{2\xi}_{|{\beta}|} h_{{\underline}s^{2\xi}}(3 L d(x)). \end{aligned}$$ Both $p(x_i)$ and $p(x)$ are majorized by ${\Gamma}_{v^\xi}(Ld(x)/2)$, indeed, by , , and since ${\Gamma}_{v^\xi}$ is decreasing, $$p(x_i) = 2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x_i)) -1 \le 2 {\Gamma}_{{\underline}s^{2\xi}}(L d(x_i)) = {\Gamma}_{v^\xi}(L d(x_i)) \le {\Gamma}_{v^\xi}(L d(x)/2).$$ So the degree of the polynomial $T_{\hat x}^{p(x_i) } F - T_{\hat x}^{p(x) } F$ is at most ${\Gamma}_{v^\xi}(Ld(x)/2)$. The valuation of the polynomial is equal to $\min\{p(x_i),p(x)\}+1$ (unless $p(x_i) = p(x)$ in which case is trivial) and so at least $2 {\Gamma}_{{\underline}s^{2\xi}}(3 L d( x)) =: 2 q$, by . So if $H_2$ denotes the left-hand side of , then (see the calculation in [@RainerSchindl17 (5.7)]) $$\begin{aligned} H_2 &\le \frac{C |{\beta}|!}{(n d(x))^{|{\beta}|}} \sum_{j =2 q }^{{\Gamma}_{v^\xi}(Ld(x)/2)} (2n^2 {\rho}d(x))^{j} v^\xi_{j}. \end{aligned}$$ By , $v^\xi_j (L d(x)/2)^j \le v^\xi_{2 q} (L d(x)/2)^{2 q}$ for $2q \le j \le {\Gamma}_{v^\xi}(L d(x)/2)$. By the definition of $q$, $h_{{\underline}s^{2\xi}}(3L d(x)) = {\underline}s^{2\xi}_{q} (3L d(x))^{q} \le {\underline}s^{2\xi}_{k} (3L d(x))^{k}$ for all $k$. With this leads to $$\begin{aligned} H_2 &\le \frac{C |{\beta}|!}{(n d(x))^{|{\beta}|}} \sum_{j =2 q }^{{\Gamma}_{v^\xi}(L d(x)/2)} \Big(\frac{4 n^2 {\rho}}{L} \Big)^{j} v^\xi_{2 q} \Big(\frac{L d(x)}2\Big)^{2 q} \\ &\le \frac{C |{\beta}|!}{(n d(x))^{|{\beta}|}} \sum_{j =2 q }^{{\Gamma}_{v^\xi}(L d(x)/2)} \Big(\frac{4n^2 {\rho}}{L } \Big)^{j} ({\underline}s^{2\xi}_{q})^2 \Big(\frac{L d(x)}2\Big)^{2 q} \\ &\le C \Big(\frac{3L}{n}\Big)^{|{\beta}|} |{\beta}|!\, {\underline}s^{2\xi}_{|{\beta}|} h_{{\underline}s^{2\xi}}(3L d(x)) \sum_{j =2 q }^{{\Gamma}_{v^\xi}(L d(x)/2)} \Big(\frac{4n^2 {\rho}}{L } \Big)^{j}. \end{aligned}$$ If we choose $L \ge 8n^2 {\rho}$, then the sum is bounded by $2$, and follows. $\bullet$ Assume that $L$ is chosen such that $$\label{eq:L} L > \max\{C_0,C_1,C_2\} \, {\rho}$$ so that , , , and are valid. Recall that ${\mathfrak{W}}$ denotes the weight matrix associated with ${\omega}$. The next lemma is a substitute for the claim in the proof of Theorem 5.5 in [@RainerSchindl17]. There exist constants $K_j=K_j(n,{\omega})$, $j = 1,2,3$, such that the following holds. If $p = K_1 L$ and $L>K_2 {\rho}$, then there exist a weight sequence $W \in {\mathfrak{W}}$ and a constant $M_1= M_1(n,{\omega},L)>0$ such that for all $x \in {\mathbb{R}}^n \setminus E$ with $d(x) < \min \{r_0/(3B_1), 1/(3L{\underline}s^{2\xi}_1)\}$ and all ${\alpha}\in {\mathbb{N}}^n$, $$\label{eqclaim2} |{\partial}^{\alpha}(f - T_{\hat x}^{p(x) } F) (x)| \le C M_1^{|{\alpha}|+1} W_{|{\alpha}|} h_{{\underline}s^{4\xi}}(K_3 L d(x)),$$ where $C$ and ${\rho}$ are the constants from and (and $B_1$ is the universal constant from [@RainerSchindl17 Lemma 4.7]). By the Leibniz rule, $$\begin{aligned} {\partial}^{\alpha}& (f - T_{\hat x}^{p(x) } F) (x) = \sum_{{\beta}\le {\alpha}} \binom{{\alpha}}{{\beta}} \sum_i {\varphi}_{i,p}^{({\alpha}-{\beta})}(x) \, {\partial}^{\beta}(T_{\hat x_i}^{p(x_i) } F - T_{\hat x}^{p(x) } F) (x). \label{Leibniz}\end{aligned}$$ Now and imply, that for $x \in Q_i^*$ with $d(x)<1/(3L{\underline}s^{2\xi}_1)$, $$\begin{aligned} \label{eq:claim1} |{\partial}^{\beta}(T_{\hat x_i}^{p(x_i)} F - T_{\hat x}^{p(x)} F) (x)| \le C (6 L)^{|{\beta}|+1} {\underline}S^{2\xi}_{|{\beta}|} \, h_{{\underline}s^{2\xi} }(3L d(x)).\end{aligned}$$ As in [@RainerSchindl17] we conclude (using [@RainerSchindl17 Proposition 4.9]) that there exist $W = W(p) \in {\mathfrak{W}}$ and $M = M(p)>0$ such that for all $i \in {\mathbb{N}}$, all $x \in {\mathbb{R}}^n \setminus E$ with $d(x) < r_0/(3B_1)$, and all ${\beta}\in {\mathbb{N}}^n$, $$\label{eq:claim2} |\varphi^{({\beta})}_{i,p}(x)|\le M W_{|{\beta}|}\, \Pi(p,x)$$ where, by [@RainerSchindl17 Corollary 3.11], $$\begin{aligned} \label{eq:Pi} \Pi(p,x)&=\exp\Big(\frac{A_1(n)}{p}{\sigma}^{\star}\Big(\frac{b_1 p}{9 A_2(n)} d(x)\Big)\Big) \notag \\ &\le \Big(\frac{e}{h_{{\underline}s^{{\eta}}}(\frac{b_1 p d(x)}{9 A_2(n) B})} \Big)^{\frac{A_1(n) B}{p}}, \quad \text{ for some $B\ge 1$ and all } {\eta}>0. \end{aligned}$$ ($b_1$ is the universal constant from [@RainerSchindl17 Lemma 4.7] and $A_1(n) \le A_2(n)$ are constants depending only on $n$.) By , we may assume that ${\underline}S^{2\xi} \le W$. Then, by , , , and [@RainerSchindl17 Lemma 4.7], for $x \in {\mathbb{R}}^n \setminus E$ with $d(x) < \min \{r_0/(3B_1), 1/(3L{\underline}s^{2\xi}_1)\}$, $$\begin{aligned} |{\partial}^{\alpha}& (f - T_{\hat x}^{p(x) } F) (x)| \\ &\le \sum_{{\beta}\le {\alpha}} \frac{{\alpha}!}{{\beta}!({\alpha}-{\beta})!} \cdot 12^{2n} \cdot M W_{|{\alpha}|-|{\beta}|}\Pi(p,x) \cdot C (6L)^{|{\beta}|+1} {\underline}S^{2\xi}_{|{\beta}|} \, h_{{\underline}s^{2\xi}}(3L d(x)) \\ &\le 12^{2n}C M \Big(\sum_{j=0}^{|{\alpha}|} \frac{|{\alpha}|!\, n^{|{\alpha}|+j}}{j!(|{\alpha}|-j)!} (6L)^{j+1} W_{|{\alpha}|-j} {\underline}S^{2\xi}_{j}\Big) \, \Pi(p,x) \, h_{{\underline}s^{2\xi}}(3L d(x)) \\ &\le 6\cdot 12^{2n} L C M n^{|{\alpha}|} W_{|{\alpha}|} \Big(\sum_{j=0}^{|{\alpha}|} \frac{|{\alpha}|!\, }{j!(|{\alpha}|-j)!} (6 L n)^{j}\Big) \, \Pi(p,x) \, h_{{\underline}s^{2\xi}}(3L d(x)) \\ &= 6\cdot 12^{2n} L C M (n (1 + 6L n ))^{|{\alpha}|} W_{|{\alpha}|} \Pi(p,x) \, h_{{\underline}s^{2\xi}}(3L d(x)). \end{aligned}$$ By , there is $H\ge 1$ (independent of $\xi$) such that $h_{{\underline}s^{2\xi}}(t) \le h_{{\underline}s^{4\xi}}(Ht)^2$ for $t>0$. Let us choose $L$ according to and such that $p := 27 \,A_2(n) B H L/b_1 \ge A_1(n) B$ is an integer. Then, by and since $h_{{\underline}s^{4\xi}} \le 1$, $$\begin{aligned} \Pi(p,x) \, h_{{\underline}s^{2\xi}}(3L d(x)) \le \frac{e\, h_{{\underline}s^{2\xi}}(3L d(x))}{h_{{\underline}s^{4\xi}}(3 H L d(x))} \le e\, h_{{\underline}s^{4\xi}}(3 H L d(x))\end{aligned}$$ and we obtain . (Note that $M$ depends on $p$, and hence on $L$, which results in the non-explicit dependence of $M_1$.) $\bullet$ Let us finish the proof of . By and , for all $x \in {\mathbb{R}}^n \setminus E$ with $d(x) < \min \{r_0/(3B_1), 1/(3L{\underline}s^{2\xi}_1)\}$ and all ${\alpha}\in {\mathbb{N}}^n$, $$\begin{aligned} |f^{({\alpha})}(x)| &\le |(T_{\hat x}^{p(x)} F)^{({\alpha})}(x)| + |{\partial}^{\alpha}(f - T_{\hat x}^{p(x) } F) (x)| \label{final} \le C M^{|{\alpha}|+1} W_{|{\alpha}|} \end{aligned}$$ for a suitable constant $M=M(n,{\omega},L)$. Let us fix a point $a \in E$ and ${\alpha}\in {\mathbb{N}}^n$. Since ${\Gamma}_{{\underline}s^{2\xi}}(t) \to \infty$ as $t \to 0$, we have $|{\alpha}| < p(x)$ if $x \in {\mathbb{R}}^n \setminus E$ is sufficiently close to $a$. Thus, as $x \to a$, $$\begin{aligned} &|f^{({\alpha})}(x) - F^{{\alpha}}(a)| \\ &\le |{\partial}^{\alpha}(f - T_{\hat x}^{p(x) } F) (x)| + |(T_{\hat x}^{p(x)}F)^{({\alpha})}(x)-F^{\alpha}(\hat x)| + |F^{\alpha}(\hat x) - F^{\alpha}(a)| \\ & = O(h_{{\underline}s^{4\xi}}(K_3 L d(x))) + O(d(x)) + O(|\hat x - a|), \end{aligned}$$ by , , and . Hence $f^{({\alpha})}(x) \to F^{{\alpha}}(a)$ as $x \to a$. We may conclude that $f \in C^\infty({\mathbb{R}}^n)$ and extends $F$. After multiplication with a suitable cut-off function of class ${\mathcal{B}}^{\{{\omega}\}}$ with support in $\{x : d(x) < \min \{r_0/(3B_1), 1/(3L{\underline}s^{2\xi}_1)\}\}$, we find that $f \in {\mathcal{B}}^{\{{\omega}\}}({\mathbb{R}}^n)$ thanks to , , and [@RainerSchindl17 Lemma 2.4(5)]. The proof of is complete. Concave, good, and strong weight functions {#sec:questions} ========================================== In [@RainerSchindl17 Definition 3.5] we called a weight function ${\sigma}$ *good* if its associated weight matrix ${\mathfrak{S}}$ satisfies . A non-quasianalytic weight function ${\omega}$ is called *strong* if there is a constant $C>0$ such that $$\int_1^\infty \frac{{\omega}(tu)}{u^2}\,du \le C {\omega}(t) + C, \quad \text{ for all } t>0.$$ Otherwise put, ${\omega}$ is strong if and only if it is equivalent to the concave weight function $\kappa=\kappa({\omega})$ defined in . In [@RainerSchindl17] we asked the following questions: Question 3.21 : *Is every concave weight function equivalent to a good one?* Question 5.11 : *Is every strong weight function equivalent to a good one?* We will give partial answers to these questions and reveal some related connections in below. In [@RainerSchindl17] it was important that the associated weight matrix *itself* satisfies as explained after the proof of . Since we could overcome this problem (by introducing ${\mathfrak{V}}= \{V^\xi\}$), it is more natural to allow for a wider concept of goodness. For completeness we will also treat the Beurling case. A weight function ${\omega}$ is called *R-good* if there exists a weight matrix ${\mathfrak{M}}$ satisfying $$\label{eq:goodR} {\;\forall}M \in {\mathfrak{M}}{\;\exists}N \in {\mathfrak{M}}{\;\exists}C\ge 1 {\;\forall}1 \le j \le k : \frac{\mu_j}{j} \le C \frac{\nu_k}{k}$$ such that ${\mathcal{B}}^{\{{\omega}\}} = {\mathcal{B}}^{\{{\mathfrak{M}}\}}$. Recall that $\mu_k := M_k/M_{k-1}$ and $\nu_k := N_k/N_{k-1}$. Similarly, ${\omega}$ is called *B-good* if there exists a weight matrix ${\mathfrak{M}}$ satisfying $$\label{eq:goodB} {\;\forall}N \in {\mathfrak{M}}{\;\exists}M \in {\mathfrak{M}}{\;\exists}C\ge 1 {\;\forall}1 \le j \le k : \frac{\mu_j}{j} \le C \frac{\nu_k}{k}$$ such that ${\mathcal{B}}^{({\omega})} = {\mathcal{B}}^{({\mathfrak{M}})}$. The next lemma, which is inspired by [@Jimenez-GarridoSanz2016 Proposition 4.15], implies that for any weight matrix ${\mathfrak{M}}$ satisfying (resp. ) there is a weight matrix ${\mathfrak{S}}$ consisting of strongly log-convex weight sequences such that ${\mathcal{B}}^{\{{\mathfrak{M}}\}} = {\mathcal{B}}^{\{{\mathfrak{S}}\}}$ (resp. ${\mathcal{B}}^{({\mathfrak{M}})} = {\mathcal{B}}^{({\mathfrak{S}})}$). \[lem:good\] Assume that $1= \mu_0 \le \mu_1 \le \cdots$ and $1= \nu_0 \le \nu_1 \le \cdots$ satisfy $${\;\exists}C>0 : \frac{\mu_j}{j} \le C \frac{\nu_k}k, \quad \text{ for all } j \le k.$$ Then the sequence $\tilde \nu$ defined by $$\frac{\tilde \nu_k}k := \inf_{\ell \ge k} \frac{\nu_\ell}\ell, \quad \tilde \nu_0 :=1,$$ is such that $\tilde \nu_k/k$ is increasing and $C^{-1} \mu \le \tilde \nu \le \nu$. The next two corollaries are immediate from and results of [@RainerSchindl12], [@RainerSchindl14], and [@RainerSchindl16a]. Let ${\mathfrak{M}}$ be a weight matrix with the property that for all $M \in {\mathfrak{M}}$ there is $N \in {\mathfrak{M}}$ such that $(M_{k+1}/N_k)^{1/k}$ is bounded. Consider the following conditions: 1. ${\mathfrak{M}}$ satisfies . 2. There is a weight matrix ${\mathfrak{S}}$ consisting of strongly log-convex weight sequences such that ${\mathcal{B}}^{\{{\mathfrak{M}}\}} = {\mathcal{B}}^{\{{\mathfrak{S}}\}}$. 3. ${\mathcal{B}}^{\{{\mathfrak{M}}\}}$ is stable under composition. 4. $\forall M \in {\mathfrak{M}}{\;\exists}N \in {\mathfrak{M}}{\;\exists}C>0 {\;\forall}j\le k : m_j^{1/j} \le C\, n_k^{1/k}$. Then $\Leftrightarrow$ $\Rightarrow$ $\Leftrightarrow$ . If additionally ${\mathfrak{M}}$ satisfies $$\label{mgR} \forall M \in {\mathfrak{M}}{\;\exists}N \in {\mathfrak{M}}: \mu_k \lesssim N_k^{1/k},$$ then all four conditions are equivalent. Let ${\mathfrak{M}}$ be a weight matrix with the property that for all $N \in {\mathfrak{M}}$ there is $M \in {\mathfrak{M}}$ such that $(M_{k+1}/N_k)^{1/k}$ is bounded. Consider the following conditions: 1. ${\mathfrak{M}}$ satisfies . 2. There is a weight matrix ${\mathfrak{S}}$ consisting of strongly log-convex weight sequences such that ${\mathcal{B}}^{({\mathfrak{M}})} = {\mathcal{B}}^{({\mathfrak{S}})}$. 3. ${\mathcal{B}}^{({\mathfrak{M}})}$ is stable under composition. 4. $\forall N \in {\mathfrak{M}}{\;\exists}M \in {\mathfrak{M}}{\;\exists}C>0 {\;\forall}j\le k : m_j^{1/j} \le C\, n_k^{1/k}$. Then $\Leftrightarrow$ $\Rightarrow$ $\Leftrightarrow$ . If additionally ${\mathfrak{M}}$ satisfies $$\label{mgB} \forall N \in {\mathfrak{M}}{\;\exists}M \in {\mathfrak{M}}: \mu_k \lesssim N_k^{1/k},$$ then all four conditions are equivalent. In general, (c) $\not\Rightarrow$ (b) in neither of the corollaries which follows from [@RainerSchindl12 Example 3.6]. Note that if $M=N$ then and reduce to a condition which is usually called *moderate growth* or $M$. For weight functions ${\omega}$ we get a full characterization. \[thm:omegachar\] Let ${\omega}$ be a weight function satisfying ${\omega}(t) = o(t)$ as $t \to \infty$. Then the following are equivalent. 1. ${\omega}$ is equivalent to a concave weight function. 2. $\exists C>0 {\;\exists}t_0 >0 {\;\forall}{\lambda}\ge 1 {\;\forall}t \ge t_0 : {\omega}({\lambda}t)\le C {\lambda}\, {\omega}(t)$. 3. ${\mathcal{B}}^{\{{\omega}\}}$ is stable under composition. 4. ${\mathcal{B}}^{({\omega})}$ is stable under composition. 5. There is a weight matrix ${\mathfrak{S}}$ consisting of strongly log-convex weight sequences such that ${\mathcal{B}}^{\{{\omega}\}} = {\mathcal{B}}^{\{{\mathfrak{S}}\}}$. 6. There is a weight matrix ${\mathfrak{S}}$ consisting of strongly log-convex weight sequences such that ${\mathcal{B}}^{({\omega})} = {\mathcal{B}}^{({\mathfrak{S}})}$. 7. ${\omega}$ is R-good. 8. ${\omega}$ is B-good. Notice that the conditions in the theorem are furthermore equivalent to the classes ${\mathcal{B}}^{\{{\omega}\}}$ and ${\mathcal{B}}^{({\omega})}$ to be stable under inverse/implicit functions and solving ODEs, and, in terms of the associated weight matrix ${\mathfrak{W}}= \{W^\xi\}_{\xi>0}$, to $${\;\forall}\xi>0 {\;\exists}{\eta}>0 : (w^\xi_j)^{1/j} \le C\, (w^{\eta}_k)^{1/k} \quad \text{ for } j \le k,$$ as well as $${\;\forall}{\eta}>0 {\;\exists}\xi>0 : (w^\xi_j)^{1/j} \le C\, (w^{\eta}_k)^{1/k} \quad \text{ for } j \le k,$$ see [@RainerSchindl14]. In the forthcoming paper [@FurdosNenningRainer] we shall see that they are also equivalent to the property that ${\mathcal{B}}^{\{{\omega}\}}$, resp. ${\mathcal{B}}^{({\omega})}$, can be described by almost analytic extensions; see also [@PetzscheVogt84]. The equivalence of the first four conditions – is well-known, see e.g. [@RainerSchindl14], which is based on [@Peetre70 Lemma 1] and [@FernandezGalbis06]. That (a) implies (e) and (f) follows from . (e) $\Rightarrow$ (c) and (f) $\Rightarrow$ (d) are clear; cf. [@RainerSchindl12]. The equivalences (e) $\Leftrightarrow$ (g) and (f) $\Leftrightarrow$ (h) follow from . Weight matrices =============== By a *weight matrix* we mean a family ${\mathfrak{M}}$ of weight sequences $M \ge (k!)_k$ which is totally ordered with respect to the pointwise order relation on sequences, i.e., - ${\mathfrak{M}}\subseteq {\mathbb{R}}^{\mathbb{N}}$, - each $M \in {\mathfrak{M}}$ is a weight sequence, which means that $M_0 = 1$, $M_k^{1/k} \to \infty$, and $M$ is log-convex, - each $M \in {\mathfrak{M}}$ satisfies $k! \le M_k$ for all $k$, - for all $M,N \in {\mathfrak{M}}$ we have $M \le N$ or $M \ge N$. For a weight matrix ${\mathfrak{M}}$ and an open $U \subseteq {\mathbb{R}}^n$ we consider the Roumieu class $$\begin{aligned} {\mathcal{B}}^{\{{\mathfrak{M}}\}}(U) &:= {\operatorname}{ind}_{M \in {\mathfrak{M}}} {\mathcal{B}}^{\{M\}}(U),\end{aligned}$$ and the Beurling class $$\begin{aligned} {\mathcal{B}}^{({\mathfrak{M}})}(U) &:= {\operatorname}{proj}_{M \in {\mathfrak{M}}} {\mathcal{B}}^{(M)}(U).\end{aligned}$$ For weight matrices ${\mathfrak{M}}$, ${\mathfrak{N}}$ we have (cf. [@RainerSchindl12]) $$\begin{aligned} {\mathcal{B}}^{\{{\mathfrak{M}}\}} \subseteq {\mathcal{B}}^{\{{\mathfrak{N}}\}} \quad &\Leftrightarrow \quad {\;\forall}M\in {\mathfrak{M}}{\;\exists}N \in {\mathfrak{N}}: (M_k/N_k)^{1/k} \text{ is bounded}, \\ {\mathcal{B}}^{({\mathfrak{M}})} \subseteq {\mathcal{B}}^{({\mathfrak{N}})} \quad &\Leftrightarrow \quad {\;\forall}N\in {\mathfrak{N}}{\;\exists}M \in {\mathfrak{M}}: (M_k/N_k)^{1/k} \text{ is bounded}. $$ Analogous equivalences hold for the *local* classes $${\mathcal{E}}^{\{{\mathfrak{M}}\}}(U) := {\operatorname}{proj}_{V \Subset U} {\mathcal{B}}^{\{{\mathfrak{M}}\}}(V) \quad \text{ and } \quad {\mathcal{E}}^{({\mathfrak{M}})}(U) := {\operatorname}{proj}_{V \Subset U} {\mathcal{B}}^{({\mathfrak{M}})}(V).$$ \[2\][ [\#2](http://www.ams.org/mathscinet-getitem?mr=#1) ]{} \[2\][\#2]{} [10]{} J. Bonet, R. W. Braun, R. Meise, and B. A. Taylor, *Whitney’s extension theorem for nonquasianalytic classes of ultradifferentiable functions*, Studia Math. **99** (1991), no. 2, 155–184. J. Bonet, R. Meise, and B. A. Taylor, *On the range of the [B]{}orel map for classes of nonquasianalytic functions*, Progress in functional analysis ([P]{}eñiscola, 1990), North-Holland Math. Stud., vol. 170, North-Holland, Amsterdam, 1992, pp. 97–111. C. Fern[á]{}ndez and A. Galbis, *Superposition in classes of ultradifferentiable functions*, Publ. Res. Inst. Math. Sci. **42** (2006), no. 2, 399–419. S. F[[ü]{}]{}rd[[ö]{}]{}s, D. N. Nenning, A. Rainer, and G. Schindl, *Almost analytic extensions of ultradifferentiable functions*, in preparation. J. Jiménez-Garrido and J. Sanz, *Strongly regular sequences and proximate orders*, Journal of Mathematical Analysis and Applications **438** (2016), no. 2, 920–945. J. Jim[é]{}nez-Garrido, J. Sanz, and G. Schindl, *Sectorial extensions, via [L]{}aplace transforms, in ultraholomorphic classes defined by weight functions*, (2017), ar[X]{}iv:1710.10081. H. Komatsu, *Ultradistributions. [I]{}. [S]{}tructure theorems and a characterization*, J. Fac. Sci. Univ. Tokyo Sect. IA Math. **20** (1973), 25–105. W. Matsumoto, *Characterization of the separativity of ultradifferentiable classes*, J. Math. Kyoto Univ. **24** (1984), no. 4, 667–678. R. Meise and B. A. Taylor, *Whitney’s extension theorem for ultradifferentiable functions of [B]{}eurling type*, Ark. Mat. **26** (1988), no. 2, 265–287. J. Peetre, *Concave majorants of positive functions*, Acta Math. Acad. Sci. Hungar. **21** (1970), 327–333. H.-J. Petzsche and D. Vogt, *Almost analytic extension of ultradifferentiable functions and the boundary values of holomorphic functions*, Math. Ann. **267** (1984), no. 1, 17–35. A. Rainer and G. Schindl, *Composition in ultradifferentiable classes*, Studia Math. **224** (2014), no. 2, 97–131. , *Equivalence of stability properties for ultradifferentiable function classes*, Rev. R. Acad. Cienc. Exactas Fis. Nat. Ser. A Math. RACSAM. **110** (2016), no. 1, 17–32. , *Extension of [W]{}hitney jets of controlled growth*, Math. Nachr. **290** (2017), no. 14-15, 2356–2374, doi:10.1002/mana.201600321. , *On the [B]{}orel mapping in the quasianalytic setting*, Math. Scand. **121** (2017), 293–310. A. Rainer and G. Schindl, *On the extension of [W]{}hitney ultrajets*, [Studia Math.]{} **245** (2019), no. 3, 255–287, doi:10.4064/sm170906-23-11. G. Schindl, *Exponential laws for classes of [D]{}enjoy–[C]{}arleman differentiable mappings*, Ph.D. thesis, Universit[ä]{}t Wien, 2014, <http://othes.univie.ac.at/32755/1/2014-01-26_0304518.pdf>. , *Characterization of ultradifferentiable test functions defined by weight matrices in terms of their [F]{}ourier transform*, Note di Matematica **36** (2016), no. 2, 1–35, doi:10.1285/i15900932v36n2p1. [^1]: The first author was supported by the Austrian Science Fund (FWF) Project P 26735-N25. The second author was supported by FWF-Project J 3948-N35; within this project, he is an external researcher at the Universidad de Valladolid (Spain) for the period October 2016 – September 2018

ArXiv

--- author: - 'T. Gastine' - 'L. Duarte' - 'J. Wicht' title: 'Dipolar versus multipolar dynamos: the influence of the background density stratification' --- =1 [Dynamo action in giant planets and rapidly rotating stars leads to a broad variety of magnetic field geometries including small scale multipolar and large scale dipole-dominated topologies. Previous dynamo models suggest that solutions become multipolar once inertia becomes influential. Being tailored for terrestrial planets, most of these models neglected the background density stratification.]{} [We investigate the influence of the density stratification on convection-driven dynamo models.]{} [Three-dimensional nonlinear simulations of rapidly rotating spherical shells are employed using the anelastic approximation to incorporate density stratification. A systematic parametric study for various density stratifications and Rayleigh numbers at two different aspect ratios allows to explore the dependence of the magnetic field topology on these parameters.]{} [Anelastic dynamo models tend to produce a broad range of magnetic field geometries that fall on two distinct branches with either strong dipole-dominated or weak multipolar fields. As long as inertia is weak, both branches can coexist but the dipolar branch vanishes once inertia becomes influential. The dipolar branch also vanishes for stronger density stratifications. The reason is the concentration of the convective columns in a narrow region close to the outer boundary equator, a configuration that favors non-axisymmetric solutions. In multipolar solutions, zonal flows can become significant and participate in the toroidal field generation. Parker dynamo waves may then play an important role close to onset of dynamo action leading to a cyclic magnetic field behavior.]{} [These results are compatible with the magnetic field of gas planets that are likely generated in their deeper conducting envelopes where the density stratification is only mild. Our simulations also suggest that the fact that late M dwarfs have dipolar or multipolar magnetic fields can be explained in two ways. They may differ either by the relative influence of inertia or fall into the regime where both types of solutions coexist. ]{} Introduction ============ The magnetic fields of planets and rapidly rotating low-mass stars are maintained by magnetohydrodynamic dynamos operating in their interiors. Scaling laws for convection-driven dynamos successfully predict the magnetic field strengths for both types of objects indicating that similar mechanisms are at work [@Christensen09]. Recent observations show a broad variety of magnetic field geometries, ranging from large-scale dipole-dominated topologies [as on Earth, Jupiter and some rapidly rotating M dwarfs, see e.g. @Donati06; @midM] to small-scale more complex magnetic structures [e.g. @earlyM; @lateM]. Explaining the different field geometries remains a prime goal of dynamo theory. Unfortunately, the extreme parameters of planetary and stellar dynamo regions can not directly be adopted in the numerical models so that scaling laws are of prime importance [@Christensen10]. Global simulations of rapidly rotating convection successfully reproduce many properties of planetary dynamos [e.g. @Christensen06]. They show that the ordering influence of the Coriolis force is responsible for producing a dominant large scale dipolar magnetic field [for a stellar application, see also @Brown10] provided the Rayleigh number is not too large. Multipolar geometries are assumed once the Rayleigh number is increased beyond a value where inertial forces become important [@Kutzner02; @Sreenivasan06]. [@Christensen06] suggest that the ratio of (nonlinear) inertial to Coriolis forces can be quantified with what they call the “local Rossby number” $\text{Ro}_\ell=u_{\text{rms}}/\Omega \ell$, where $\ell$ is the typical lengthscale of the convective flow. Independently of the other system parameters, the transition between dipole-dominated and multipolar magnetic fields always happens around $\text{Ro}_\ell \simeq 0.1$. Scaling laws then, for example, predict a dipole-dominated field for Jupiter and a multipolar field for Mercury [@Olson06]. The latter may explain the strong quadrupolar component in the planet’s magnetic field [@Christensen06nat]. Earth lies at the boundary where the field is dipolar most of the time but occasionally forays into the multipolar regime allowing for magnetic field reversals. The scaling laws based on Boussinesq simulations are geared to model the dynamo in Earth’s liquid metal core where the background density and temperature variations can be neglected. Their application to gas giants and stars where both quantities vary by orders of magnitude is therefore questionable [@Chabrier06; @Nettelmann12]. Recent compressible dynamo models of fully-convective stars suggest a strong influence of the density stratification on the geometry of the magnetic field: while the fully compressible models of [@Dobler06] have a significant dipolar component, the strongly stratified anelastic models of [@Browning08] tend to produce multipolar dynamos [see also @Bessolaz11]. These differences stress the need of more systematic parameter studies to clarify the influence of density stratification. In addition, most of the previous Boussinesq studies have employed rigid flow boundary conditions for modelling terrestrial dynamo models. Stress-free boundary conditions, more appropriate for gas planets and stars, show a richer dynamical behaviour, including hemispherical dynamos and bistability where dipole-dominated and multipolar dynamos coexist at the same parameters [e.g. @Grote00; @Busse06; @Goudard08; @Simitev09; @Sasaki11]. This challenges the usefulness of the $\text{Ro}_\ell$ criterion since bistable cases exist significantly below $\text{Ro}_\ell\simeq 0.1$ [e.g. @Schrinner12]. Here we adopt the anelastic approximation [e.g. @Glatz1; @Brag95; @Lantz99] to explore the effect of a background density stratification on the dynamo process while filtering out fast acoustic waves. We conduct an extensive parameter study where we vary the degree of stratification and the Rayleigh number to determine the parameter range where dipole-dominated fields can be expected. The anelastic approximation and the numerical methods are introduced in section \[sec:model\]. In section \[sec:results\], we present the results of the parametric study and describe the different dynamo regimes before relating our results to observations in the concluding section \[sec:discussion\]. The dynamo model {#sec:model} ================ An anelastic formulation ------------------------ We consider MHD simulations of a conducting ideal gas in a spherical shell rotating at a constant frequency $\Omega$ about the $z$ axis. Convective motions are driven by a fixed entropy contrast $\Delta s$ between the inner radius $r_i$ and the outer radius $r_o$. Following the previous parametric studies of [@Christensen06], we use a dimensionless formulation, where the shell thickness $d=r_o-r_i$ is the reference lengthscale, the viscous diffusion time $d^2/\nu$ is the reference timescale and $\sqrt{\rho_{\text{top}} \mu \lambda \Omega}$ is the magnetic scale. Density and temperature are non-dimensionalised using their values at the outer boundary $\rho_{\text{top}}$ and $T_{\text{top}}$ and $\Delta s$ serves as the entropy scale. The kinematic viscosity $\nu$, magnetic diffusivity $\lambda$, magnetic permeability $\mu$, thermal diffusivity $\kappa$ and heat capacity $c_p$ are assumed to be constant. Following the anelastic formulation by [@Glatz1], [@Brag95] and [@Lantz99] we adopt a non-magnetic, hydrostatic and adiabatic background reference state that we denote with overbars in the following. It is defined by the temperature profile $d{\bar{T}}/dr = -g/c_p$ and the assumption of a polytropic gas ${\bar{\rho}}(r) = {\bar{T}}^m$, where $m$ is the polytropic index. Gravity typically increases linearly with radius in Boussinesq models where the density is approximately constant. Many anelastic models on the other hand assume that the density is predominantly concentrated below the dynamo zone so that $g\propto 1/r^2$ provides a better approximation [e.g. @Glatz84; @Jones11]. We adopt the form $$g(r) = g_1 \dfrac{r}{r_o} + g_2 \dfrac{r_o^2}{r^2},$$ with constants $g_1$ and $g_2$ to allow for both alternatives which then leads to the following background temperature profile [see also @Jones11; @Gastine12] $${\bar{T}}(r) = c_o{\left(}g_2\dfrac{r_o}{r}-\dfrac{g_1}{2} \dfrac{r^2}{r_o^2}{\right)}+ 1-c_o {\left(}g_2 -\dfrac{g_1}{2} {\right)},$$ with $$c_o = \dfrac{\eta {\left(}\exp\frac{N_\rho}{m}-1 {\right)}}{\frac{g_1}{2}(1-\eta^2)\eta+g_2(1-\eta)} \text{~and~} N_\rho = \ln\dfrac{{\bar{\rho}}(r_i)}{{\bar{\rho}}(r_o)}.$$ Here, $\eta=r_i/r_o$ is the aspect ratio of the spherical shell and $N_\rho$ is the number of density scale heights covered by the density background. Under the anelastic approximation the dimensionless equations governing convective motions and magnetic field generation are then given by $$\begin{aligned} \text{E}{\left(}\dfrac{\partial \vec{u}}{\partial t}+\vec{u}\cdot\vec{\nabla}\vec{u}{\right)}= & -\vec{\nabla}{\dfrac{p}{{\bar{\rho}}}}+\dfrac{\text{Ra}\,\text{E}}{\text{Pr} }g(r) s\,\vec{e_r} -2\vec{e_z}\times\vec{u} \\ & + \dfrac{1}{\text{Pm}\ {\bar{\rho}}}{\left(}\vec{\nabla}\times \vec{B}{\right)}\times \vec{B} + \dfrac{\text{E}}{{\bar{\rho}}} \vec{\nabla}\cdot\tens{S}, \end{aligned} \label{eq:NS}$$ $$\dfrac{\partial \vec{B}}{\partial t} = \vec{\nabla}\times {\left(}\vec{u}\times\vec{B}{\right)}+\dfrac{1}{\text{Pm}}\vec{\Delta}\vec{B}, \label{eq:induction}$$ $$\begin{aligned} {\bar{\rho}}{\bar{T}}{\left(}\dfrac{\partial s}{\partial t} + \vec{u}\cdot\vec{\nabla} s{\right)}= & \dfrac{1}{\text{Pr}}\vec{\nabla}\cdot{\left(}{\bar{\rho}}{\bar{T}}\vec{\nabla} s{\right)}+ \dfrac{\text{Pr}}{\text{Ra}}\,\dfrac{c_o}{2{\bar{\rho}}} (1-\eta)\,\tens{S}^2 \\ & + \dfrac{\text{Pr}}{\text{Ra\,E\,Pm}^2}\, c_o (1-\eta) {\left(}\vec{\nabla}\times\vec{B}{\right)}^2, \end{aligned} \label{eq:entropy}$$ $$\vec{\nabla}\cdot {\left(}{\bar{\rho}}\vec{u} {\right)}= 0 \quad,\quad \vec{\nabla}\cdot \vec{B} = 0 \label{eq:anel},$$ where $\vec{u}$, $\vec{B}$, $p$ and $s$ are velocity, magnetic field, pressure and entropy disturbance of the background state, respectively. $\tens{S}$ is the traceless rate-of-strain tensor with a constant kinematic viscosity given by $$\tens{S}={\bar{\rho}}\left(\frac{\partial u_i}{\partial x_j} + \frac{\partial u_j}{\partial x_i}-\frac{2}{3} \delta_{ij} \vec{\nabla}\cdot\vec{u}\right), \label{eq:tenseur}$$ where $\delta_{ij}$ is the identity matrix. In addition to the aspect ratio of the spherical shell $\eta$ and the two parameters involved in the description of the reference state ($N_\rho$ and $m$), the system of equations (\[eq:NS\]-\[eq:anel\]) is governed by four dimensionless parameters, namely the Ekman number E, the Prandtl number Pr, the magnetic Prandtl number Pm and the Rayleigh number Ra defined by $$\text{E} = \dfrac{\nu}{\Omega d^2} \ ;\ \text{Pr} = \dfrac{\nu}{\kappa} \ ;\ \text{Pm} = \dfrac{\nu}{\lambda} \ ;\ \text{Ra} = \dfrac{g_{\text{top}} d^3 \Delta s}{c_p \nu\kappa}, \label{eq:params}$$ where $g_{\text{top}}$ is the reference gravity at the outer boundary $r_o$. The numerical method -------------------- The numerical simulations in this parameter study have been computed with the anelastic version of the code MagIC [@Wicht02; @Gastine12], which has been validated in an anelastic dynamo benchmark [@Jones11]. To solve the system of equations (\[eq:NS\]-\[eq:anel\]), ${\bar{\rho}}\vec{u}$ and $\vec{B}$ are decomposed into poloidal and toroidal contributions $${\bar{\rho}}\vec{u} = \vec{\nabla}\times(\vec{\nabla}\times W \vec{e_r}) + \vec{\nabla}\times Z \vec{e_r} \ ,\ \vec{B}= \vec{\nabla}\times(\vec{\nabla}\times C \vec{e_r}) + \vec{\nabla}\times A \vec{e_r}.$$ The unknowns $W$, $Z$, $A$, $C$, $s$ and $p$ are then expanded in spherical harmonic functions up to degree $\ell_{\text{max}}$ in colatitude $\theta$ and longitude $\phi$ and in Chebyshev polynomials up to degree $N_r$ in the radial direction. An exhaustive description of the complete numerical method and the associated spectral transforms can be found in [@Glatz1]. Typical numerical resolutions employed in this study range from ($N_r=61,\,\ell_{\text{max}}=85$) for Boussinesq simulations close to onset to ($N_r=81,\,\ell_{\text{max}}=133$) for the more demanding cases with a significant density contrast. In all the simulations presented in this study, we have assumed constant entropy boundary conditions at $r_i$ and $r_o$. The mechanical boundary conditions are either no slip at the inner and stress-free at the outer boundary or stress-free at both limits (see Tab. \[tab:results\]). The magnetic field is match to a potential field at both boundaries. Non-dimensional diagnostic parameters ------------------------------------- To quantify the impact of the different control parameters on magnetic field and flow, we analyse several diagnostic non-dimensional properties. The typical rms flow amplitude in the shell is either given as the magnetic Reynolds number $\text{Rm} = u_{\text{rms}} d/\lambda$ or the Rossby number $\text{Ro}=u_{\text{rms}}/\Omega d$. Rm is a measure for the ratio of magnetic field generation to Ohmic dissipation and thus an important quantity for any dynamo. The Rossby number quantifies the ratio between inertia and Coriolis forces but @Christensen06 demonstrated that the local Rossby number $\text{Ro}_\ell=u_{\text{rms}}/\Omega \ell$ is a more appropriate measure, at least concerning the impact of inertia on the magnetic field geometry. The typical flow lengthscale $\ell$ can be calculated based on the spherical harmonic flow contributions $$\ell = \pi/\bar{\ell}_u,$$ with $$\bar{\ell}_u = \sum_\ell \dfrac{\ell\ \langle \vec{u}_\ell\cdot\vec{u}_\ell \rangle}{\langle \vec{u}\cdot \vec{u} \rangle}.$$ Here, $\vec{u}_\ell$ is the velocity field at a given spherical harmonic degree $\ell$ and $\langle \cdots \rangle$ corresponds to an average over time and radius. The magnetic field strength can, for example, be measured by the Elsasser number $\Lambda = B_{\text{rms}}^2/\rho\mu_o\lambda \Omega$, which is supposed to provide an estimate for the ratio of Lorentz to Coriolis forces. This estimate may be rather far from the true force balance [e.g. @Wicht10]. The modified Elsasser number suggested by [@King12] and defined by $$\Lambda_{\ell} = \dfrac{\Lambda}{\text{Rm}} \dfrac{d}{\delta_B} \quad\text{with}\quad \delta_B = \dfrac{\pi d}{2\sqrt{\bar{\ell}_B^2+\bar{m}_B^2} }, \label{eq:elsloc}$$ provides a more appropriate measure, being more directly related to the ratio of the two respective terms in the Navier-Stokes equation (\[eq:NS\]) when approximating the lengthscale entering in the curl in the Lorentz force by $\delta_B^{-1}$. In addition to the typical spherical harmonic degree $\bar{\ell}_B$, the typical order $\bar{m}_B$ is also used since both seem to contribute [@King12]: $$\bar{\ell}_B = \sum_\ell \dfrac{\ell\ \langle \vec{B}_\ell\cdot\vec{B}_\ell \rangle}{\langle \vec{B}\cdot \vec{B} \rangle} \ ,\ \bar{m}_B = \sum_m \dfrac{m\ \langle \vec{B}_m\cdot\vec{B}_m \rangle}{\langle \vec{B}\cdot \vec{B} \rangle}.$$ Finally, the geometry of the magnetic field is quantified by its dipolarity $$f_{\text{dip}} = \dfrac{\oiint\vec{B}^2_{\ell=1,m=0} \sin\theta\,d\theta\,d\phi} {\oiint\vec{B}^2\sin \theta\,d\theta\,d\phi},$$ that measures the ratio of the magnetic energy of the dipole to the total magnetic energy at the outer boundary $r_o$. A parameter study ----------------- Aspect ratio $N_\rho$ $\text{Ra}_c$ $m_{\text{crit}}$ -------------- ---------- ---------------------- ------------------- 0.2 0 $8.706\times 10^{5}$ 4 0.2 0.5 $1.378\times 10^{6}$ 5 0.2 1 $1.935\times 10^{6}$ 5 0.2 1.5 $2.617\times 10^{6}$ 5 0.2 1.7 $2.929\times 10^{6}$ 6 0.2 2 $3.455\times 10^{6}$ 6 0.2 2.5 $4.591\times 10^{6}$ 6 0.2 3 $4.647\times 10^{6}$ 43 0.6 0.01 $1.737\times 10^{5}$ 21 0.6 0.5 $3.126\times 10^{5}$ 28 0.6 1 $5.174\times 10^{5}$ 34 0.6 1.5 $8.151\times 10^{5}$ 39 0.6 2 $1.140\times 10^{6}$ 53 0.6 3 $1.530\times 10^{6}$ 72 : Critical Rayleigh numbers and corresponding azimuthal wave numbers for the simulations considered in this study (for all of them $\text{E}=10^{-4}$ and $\text{Pr}=1$). \[tab:rayleigh\] In all the simulations presented in this study, the Ekman number is kept fixed to a moderate value of $\text{E}=10^{-4}$, which allows to study a large number of cases. The Prandtl number is set to 1 and the magnetic Prandtl number to 2. Following [@Jones09] and our previous hydrodynamical models [@Gastine12], we adopt a polytropic index $m=2$ for the reference state. We consider two different aspect ratios: thick shells with $\eta=0.2$ and thin shells with $\eta=0.6$. Dynamo models in thin shells rely on the same setup as our previous hydrodynamical study [@Gastine12]. Gravity is adapted to better reflect the different shells, i.e. we use linear gravity in the thick shells ($g_1=1,\,g_2=0$) and $g\propto 1/r^2$ in the thin shells ($g_1=0,\,g_2=1$). Both cases also differ in the flow boundary conditions: thin shell cases assume stress-free conditions at both boundaries, thick shell cases assume a rigid inner boundary. Being mostly interested in the effects of the density stratification, we have performed numerical simulations with different density contrasts spanning the range from $N_\rho=0$ (i.e. Boussinesq) to $N_\rho=3$ (i.e. $\rho_\text{bot}/\rho_\text{top}\simeq 20$). For each density stratification, we vary the Rayleigh number from simulations close to onset of dynamo action to 10-50 times the critical Rayleigh number $\text{Ra}_c$ for the onset of convection. $\text{Ra}_c$ varies with $N_\rho$ and we list respective values for the cases explored here in Tab. \[tab:rayleigh\]. These have been determined with the anelastic linear stability analysis code developed by [@Jones09a]. As will be shown below, we find several cases of bistability where a strong dipole-dominated and a weak multipolar solution coexist at identical parameters. The starting solution for our time integration then determines which of the two solutions the simulation will assume. Most of the numerical simulations therefore have been initiated with both a strong dipolar magnetic field (with $\Lambda \sim 1$) and a weak multipolar magnetic field (model names end with a ’d’ or a ’m’ accordingly in Tab. \[tab:results\]). Altogether, more than 100 simulations have been computed, each running several magnetic diffusion time to ensure that a stable magnetic configuration has been reached. Table \[tab:results\] lists the input parameters of all cases along with key solution properties. Results {#sec:results} ======= Dynamo regimes -------------- ![Relative dipole strength plotted against the local Rossby number $\text{Ro}_\ell$ for two different aspect ratios: $\eta=0.2$ (upper panel) and $\eta=0.6$ (lower panel). Each type of symbol is associated to a given density stratification (from Boussinesq, i.e. $N_\rho=0$ to stratified simulations corresponding to $N_\rho=3$). The size of the symbol has been chosen according to the value of the modified Elsasser number (Eq. \[eq:elsloc\]). Vertical lines are tentative transitions between dipolar and multipolar dynamos.[]{data-label="fig:dipol"}](fig/dipEta02 "fig:"){width="9cm"} ![Relative dipole strength plotted against the local Rossby number $\text{Ro}_\ell$ for two different aspect ratios: $\eta=0.2$ (upper panel) and $\eta=0.6$ (lower panel). Each type of symbol is associated to a given density stratification (from Boussinesq, i.e. $N_\rho=0$ to stratified simulations corresponding to $N_\rho=3$). The size of the symbol has been chosen according to the value of the modified Elsasser number (Eq. \[eq:elsloc\]). Vertical lines are tentative transitions between dipolar and multipolar dynamos.[]{data-label="fig:dipol"}](fig/dipEta06 "fig:"){width="9cm"} Figure \[fig:dipol\] shows how the dipolarity $f_\text{dip}$ depends on the local Rossby number in the thick shell (top) and the thin shell (bottom) simulations. In both geometries we find two distinct branches: The upper branch corresponds to the dipolar regime at $f_\text{dip} > 0.5$ with strong magnetic fields at modified Elsasser number between $\Lambda_\ell= 0.1$ and $\Lambda_\ell =1$. This indicates a strong Lorentz force so that these dynamos likely operate in the so-called magnetostrophic regime where Coriolis forces, pressure gradients, buoyancy and Lorentz forces contribute to the first order Navier-Stokes equation. In contrast, the models belonging to the lower branch have a multipolar field geometry at $f_\text{dip} < 0.2$ and much weaker magnetic field with $\Lambda_\ell \lesssim 5\times 10^{-2}$. Since the Lorentz force may thus not enter the first order Navier-Stokes equation these cases are likely geostrophic. ![Regime diagram for dynamo simulations as a function of supercriticality and density stratification for two different aspect ratios: $\eta=0.2$ (upper panel) and $\eta=0.6$ (lower panel). Red circles correspond to dipolar dynamos, blue squares to multipolar dynamos and black crosses are decaying dynamos. The size of the symbols depend on the value of the modified Elsasser number (Eq. \[eq:elsloc\]). The grey-shaded area highlights the dipolar region, while the dashed lines correspond to the critical $\text{Ro}_{\ell c}$ values that mark the limit between dipolar and multipolar dynamos (see the vertical lines in Fig. \[fig:dipol\]).[]{data-label="fig:crit"}](fig/eta02 "fig:"){width="9cm"} ![Regime diagram for dynamo simulations as a function of supercriticality and density stratification for two different aspect ratios: $\eta=0.2$ (upper panel) and $\eta=0.6$ (lower panel). Red circles correspond to dipolar dynamos, blue squares to multipolar dynamos and black crosses are decaying dynamos. The size of the symbols depend on the value of the modified Elsasser number (Eq. \[eq:elsloc\]). The grey-shaded area highlights the dipolar region, while the dashed lines correspond to the critical $\text{Ro}_{\ell c}$ values that mark the limit between dipolar and multipolar dynamos (see the vertical lines in Fig. \[fig:dipol\]).[]{data-label="fig:crit"}](fig/eta06 "fig:"){width="9cm"} For both aspect ratios, the dipolar branch is bounded by a critical local Rossby number $\text{Ro}_{\ell c}$ beyond which all cases are multipolar. In the thin shell with $\eta=0.6$, the value $\text{Ro}_{\ell c}\simeq 0.15$ is slightly larger than the value of $\text{Ro}_{\ell c}\simeq 0.12$ predicted by [@Christensen06] in their geodynamo models with $\eta=0.35$. In the thick shell with $\eta=0.2$, it is smaller at $\text{Ro}_{\ell c}\simeq 0.08$. The critical local Rossby number thus seems to increase with the inner core size. A similar dependence on the aspect ratio has been already found by [@Aubert09]. In contrast to the previous Boussinesq studies that used rigid boundary conditions [e.g. @Christensen06; @Aubert09], the multipolar branch also extends here below the critical $\text{Ro}_{\ell c}$ where the dipolar and the multipolar branch now coexist. This multipolar dynamo branch for $\text{Ro}_{\ell} < 0.1$ has also been found in the Boussinesq models of [@Schrinner12] with stress-free boundary conditions. We further discuss the bistability phenomenon in section \[sec:bi\]. Figure \[fig:crit\] illustrates how the field morphology depends on the stratification and on the supercriticality $\text{Ra}/\text{Ra}_c$. The value $N_\rho\simeq 1.8$ marks an important regime boundary: below this value most cases are bistable (nested circles and squares), while above this value only multipolar cases remain (squares only). The reason for this boundary will be considered in section \[sec:strati\]. The dashed black lines in Fig. \[fig:crit\] mark the $\text{Ro}_{\ell c}$ boundary which moves to lower supercriticalities for increasing density stratifications. Larger stratifications promote larger flow amplitudes as well as smaller lengthscales since convection is progressively concentrated in a narrowing region close to the outer boundary [see Table \[tab:rayleigh\], Fig. \[fig:vortz\] and @Jones09; @Gastine12]. Both effects lead to larger $\text{Ro}_\ell$ for a given supercriticality. For $N_\rho< 1.8$ the boundary for onset of dynamo action also seems to move to smaller supercriticalities when $N_\rho$ increases. The reason is a mixture of a decreasing critical magnetic Reynolds number and an increasing flow amplitude for a given supercriticality. Beyond $N_\rho=1.8$, the critical magnetic Reynolds number once more increases and the onset of dynamo action moves to larger supercriticalities. The onset of dynamo action and the $\text{Ro}_{\ell c}$ boundary are less affected by the stratification for the thinner shell. One reason may be that the flow lengthscale is already quite small for thinner shells even in the Boussinesq case [see Tab. \[tab:rayleigh\] and @Shamali04]. The differences in the gravity profile and the flow boundary condition may also play a role here. The dipolar branch reaches down to lower local Rossby numbers than the multipolar branch (see Fig. \[fig:dipol\]). This is where we find examples of subcritical dynamo action (models 1d, 6d and 12d in Tab. \[tab:results\]), where the dynamo is only successful when started with a sizable magnetic field strength [@Morin09]. Simulations started with a weak field do not have the option to run to the multipolar branch and the field simply decays away. Bistability {#sec:bi} ----------- ![Left panels: iso-contours of axisymmetric azimuthal field and poloidal field lines. Right panels: surface radial field $B_r(r=r_o)$. From a model with $\eta=0.2$, $N_\rho=1.7$, $\text{Ra}=9\times 10^{6}$ initiated with a dipolar magnetic field (upper panels) and a multipolar one (lower panels). Red (blue) correspond to positive (negative) values.[]{data-label="fig:bistab"}](fig/magnSF "fig:"){width="9cm"} ![Left panels: iso-contours of axisymmetric azimuthal field and poloidal field lines. Right panels: surface radial field $B_r(r=r_o)$. From a model with $\eta=0.2$, $N_\rho=1.7$, $\text{Ra}=9\times 10^{6}$ initiated with a dipolar magnetic field (upper panels) and a multipolar one (lower panels). Red (blue) correspond to positive (negative) values.[]{data-label="fig:bistab"}](fig/magnWF "fig:"){width="9cm"} ![Time-averaged kinetic and magnetic energy spectra plotted against the spherical harmonic degree $\ell$. From a model with $\eta=0.2$, $N_\rho=1.7$, $\text{Ra}=9\times 10^{6}$ initiated with a dipolar magnetic field (black lines) and a multipolar one (blue lines), corresponding to Fig. \[fig:bistab\].[]{data-label="fig:spec"}](fig/spec){width="9cm"} As illustrated on Fig. \[fig:crit\], most of our models at $N_\rho<1.8$ show bistability. Exceptions are the subcritical thick shell cases and models 13d/m and 14d/m (see Tab. \[tab:results\]), where even small initial magnetic fields developed to strong dipoles. The top panel of Fig. \[fig:bistab\] shows the dipole-dominated solution with, however, stronger dynamo action in the southern hemisphere. The lower panel of Fig. \[fig:bistab\] illustrates the weak multipolar solution, which has pronounced equatorially symmetric and non-axisymmetric components (this is a feature also visible in the more stratified cases shown in the following Figs. \[fig:brsurfEta02\]-\[fig:brsurfEta06\]). The dipolar solution has a three times larger modified Elsasser number. Generally, $\Lambda_\ell$ is twice to ten times larger in the dipole-dominated than in the multipolar counterpart of bistable cases (see Fig. \[fig:dipol\] and Tab. \[tab:results\] for further details). This suggests that the two dynamo branches are also characterised by different force balances, at least in the bistability region (i.e. $\text{Ro}_\ell < 0.1$): the dipolar branch is magnetostrophic, while the multipolar one is more likely to be geostrophic. Beyond $\text{Ro}_\ell > 0.1$, however, the multipolar dynamos can be strong enough to yield a magnetostrophic force balance due to the larger Rayleigh numbers [@Brun05]. Figure \[fig:spec\] shows the corresponding time-averaged kinetic and magnetic spectra for the two bistable solutions displayed in Fig. \[fig:bistab\]. The two kinetic energy spectra nearly coincide, emphasising the similarities of the flow in the two solutions. We observe a broad plateau for degrees $4 \leq \ell \leq 30$ that reflects the convective driving (for this model, the critical azimuthal wavenumber at onset of convection is $m=6$, see Tab. \[tab:rayleigh\]). The magnetic spectra show the differences already visible in the surface magnetic fields displayed in Fig. \[fig:bistab\]. The amplitude of the dipolar component is more than one order of magnitude lower, higher-order components drop by roughly $50\%$. A mild maximum for $\ell=4$ corresponds to the large-scale contribution visible in Fig. \[fig:bistab\]. For spherical harmonic degrees $\ell > 30$, the spectra follow some clear power-law behaviour. Both magnetic and kinetic energy spectra have a steep decrease (between $\ell^{-4}$ and $\ell^{-5}$), very similar to previous stellar dynamo models [e.g. @Brun05; @Browning08] or quasi-geostrophic kinematic dynamos [e.g. @Schaeffer06]. The coexistence of a dipolar and a multipolar branch has already been reported for Boussinesq models with stress-free [@Busse06; @Schrinner12] or mixed mechanical boundary conditions [@Sasaki11]. For rigid boundary conditions only one such case has been found by [@Christensen06]. Stress-free boundaries allow stronger zonal winds to develop which seem to play a key role here. These geostrophic flows (constant on coaxial cylinders) are maintained by Reynolds stresses, i.e. by the correlation between the zonal $u_\phi$ and the cylindrically radial $u_s$ velocity components [e.g. @Gastine12]. For stress-free boundary conditions their amplitude is limited by the weak bulk viscosity. When rigid boundary conditions are employed, the much stronger boundary friction severely brakes these zonal winds. ![Relative dipole strength plotted against the ratio of axisymmetric toroidal kinetic energy divided by toroidal energy. Red (grey) symbols correspond to simulations in thick (thin) shells ($\eta=0.2$ and $\eta=0.6$, respectively). Each type of symbol is associated with a given density stratification (from Boussinesq, i.e. $N_\rho=0$ to $N_\rho=3$). The sizes of the symbols have been chosen according to the values of the modified Elsasser number (Eq. \[eq:elsloc\]).[]{data-label="fig:shear"}](fig/shear){width="9cm"} Figure \[fig:shear\] displays the dipolarity of the surface field against the ratio of axisymmetric toroidal to total toroidal kinetic energy, which is a good proxy of the relative amplitude of zonal winds. Strongly dipolar solutions cluster in the upper left corner where zonal winds are weak. For strong zonal winds only weak multipolar solutions can be found. For bistable cases, one solution belongs to the first type while the other one belongs to the second category. There clearly is a competition between strong zonal winds and strong dipolar magnetic fields. The third type of multipolar and sometimes strong solutions with mostly weak zonal winds are mainly strongly stratified cases that we will discuss in section \[sec:strati\]. The strong multipolar fields can have a strong impact on the zonal flows, in a similar way to the stellar models of [@Brun05]. ![Role of the $\Omega$-effect in the production of the toroidal field for two selected simulations in thick shells (upper panels) and in thin shells (lower panel). For each simulation, the $\Omega$-effect (color levels) and the axisymmetric azimuthal field (solid and dashed lines) are displayed on the left part, while the axisymmetric zonal flow is given on the right part. Red (blue) correspond to positive (negative) values.[]{data-label="fig:omeffect"}](fig/omegaNrho0 "fig:"){width="4.4cm"} ![Role of the $\Omega$-effect in the production of the toroidal field for two selected simulations in thick shells (upper panels) and in thin shells (lower panel). For each simulation, the $\Omega$-effect (color levels) and the axisymmetric azimuthal field (solid and dashed lines) are displayed on the left part, while the axisymmetric zonal flow is given on the right part. Red (blue) correspond to positive (negative) values.[]{data-label="fig:omeffect"}](fig/omegaNrho3 "fig:"){width="4.4cm"} ![Role of the $\Omega$-effect in the production of the toroidal field for two selected simulations in thick shells (upper panels) and in thin shells (lower panel). For each simulation, the $\Omega$-effect (color levels) and the axisymmetric azimuthal field (solid and dashed lines) are displayed on the left part, while the axisymmetric zonal flow is given on the right part. Red (blue) correspond to positive (negative) values.[]{data-label="fig:omeffect"}](fig/omegaNrho1Eta06 "fig:"){width="4.4cm"} ![Role of the $\Omega$-effect in the production of the toroidal field for two selected simulations in thick shells (upper panels) and in thin shells (lower panel). For each simulation, the $\Omega$-effect (color levels) and the axisymmetric azimuthal field (solid and dashed lines) are displayed on the left part, while the axisymmetric zonal flow is given on the right part. Red (blue) correspond to positive (negative) values.[]{data-label="fig:omeffect"}](fig/omegaNrho3Eta06 "fig:"){width="4.4cm"} The stronger zonal winds in the multipolar cases go along with a change in dynamo mechanism which is illustrated in Fig. \[fig:omeffect\]. In the dipolar cases (left panels), zonal flows are weak and mainly driven by thermal wind effects. The stronger Lorentz force can balance the Coriolis force and allows for these strongly non-geostrophic motions [@Aubert05]. The $\Omega$-effect, i.e. the production of axisymmetric toroidal magnetic field by zonal wind shear, plays only a secondary role (left half of left panels in Fig. \[fig:omeffect\]) so that the dynamo is of the $\alpha^2$ type in the mean field nomenclature [@Olson99], similar to the planetary dynamo models by [@Christensen06] or mean-field models of fully convective stars [@Chabrier06]. In the multipolar cases (right panels in Fig. \[fig:omeffect\]), significant zonal winds driven by Reynolds stresses develop close to the outer boundary and the associated $\Omega$-effect plays an important role for toroidal magnetic field production. Dynamos on the multipolar branch (for $N_\rho<2$) are thus of the $\alpha\Omega$ or $\alpha^2\Omega$ type. Many of these solutions show distinct oscillations connected to Parker dynamo waves that we discuss further in section \[sec:parker\]. Influence of stratification {#sec:strati} --------------------------- {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} For $N_\rho>1.8$ only the multipolar branch remains. Figures \[fig:brsurfEta02\] and \[fig:brsurfEta06\] illustrate how the convective flow and the surface magnetic field evolve on increasing the density stratification for simulations with very similar local Rossby numbers. All these models have been initiated with a strong dipolar field. For both aspect ratios, the weakly stratified cases (i.e. Boussinesq and $N_\rho=1$) have a dipole-dominated magnetic field, while the stronger stratified cases (i.e. $N_\rho=2$ and $N_\rho=3$) have multipolar fields with significant non-axisymmetric contributions. As a consequence of the background density contrast, the typical lengthscale of convection gradually decreases with $N_\rho$ [see also Tab. \[tab:rayleigh\] and @Gastine12]. While the imprints of these smaller convective scale are visible in the surface magnetic fields, the dynamos are always dominated by larger scale features even in the strongly stratified cases. For example, for the two models at $N_\rho=3$, a clear wave number $m=2$ signature emerges similar to the patterns already observed in the Boussinesq models of [@Goudard08]. This magnetic mode can only be clearly identified for moderate magnetic Reynolds numbers $\text{Rm} < 200$ close to the onset of dynamo action. For larger Rm, smaller scale more complex features once more take over. Some models also show a stronger concentration of the magnetic field in one hemisphere. Similar effects can be found for some multipolar simulations at low stratifications (see section \[sec:parker\]). [@Grote00] and [@Busse06] report that hemispherical dynamo action is typical for simulations with stress-free boundaries and Prandtl numbers around unity. The collapse of the dipolar branch for $N_\rho\gtrsim 2$ is caused by the strong concentration of convective features close to the outer boundary where density decreases most drastically. Figure \[fig:vortz\] illustrates the evolution of the convective columns when $N_\rho$ increases while the local Rossby numbers remains similar (see Figs. \[fig:brsurfEta02\]-\[fig:brsurfEta06\] for the corresponding magnetic fields and radial flow structures). As it has already been demonstrated by non-magnetic simulations [@Jones09a; @Gastine12], convection first develops close to the inner boundary for Boussinesq and weakly stratified models ($N_\rho \leq 1$). When increasing $N_\rho$, however, the convective columns gradually move outward and become confined to a thin region close to the equator ($N_\rho \geq 2$). This goes along with a decrease of the typical lengthscale (see also Tab. \[tab:rayleigh\] for the critical azimuthal wave numbers at onset). According to [@Gastine12], this is a consequence of the fact that buoyancy and thus the effective local Rayleigh number becomes much larger at the outer boundary than in the interior for strongly stratified models. ![Azimuthal average of kinetic helicity $\overline{\vec{u}'\cdot\vec{\nabla}\times\vec{u}'}$ for the models displayed in Fig. \[fig:vortz\]. Positive (negative) values are rendered in red (blue).[]{data-label="fig:helicity"}](fig/helicityEta02 "fig:"){width="9cm"} ![Azimuthal average of kinetic helicity $\overline{\vec{u}'\cdot\vec{\nabla}\times\vec{u}'}$ for the models displayed in Fig. \[fig:vortz\]. Positive (negative) values are rendered in red (blue).[]{data-label="fig:helicity"}](fig/helicityEta06 "fig:"){width="9cm"} The kinetic helicity ${\cal H} = \overline{\vec{u'}\cdot\vec{\nabla}\times \vec{u'}}$ is a key ingredient in the induction process via the $\alpha$-effect [e.g. @Moffatt78]. Figure \[fig:helicity\] illustrates how the helicity increasingly concentrates closer to the outer boundary equator when the stratification intensifies simply because the convective columns are the main carriers of helicity. Mean field dynamo models have demonstrated that a similar concentration of the $\alpha$-effect can promote non-axisymmetric dynamo modes of low spherical harmonic order [typically $m \sim 1-2$, see @Rudiger03; @Jiang06]. Non-axisymmetric $\alpha^2\Omega$ mean-field models also suggest a lower critical dynamo number for non-axisymmetric dynamo modes [e.g. @Ruzmaikin88; @Bassom05]. These findings may explain the onset of the new larger wave number magnetic mode in our strongly stratified cases. Parker waves {#sec:parker} ------------ ![Time evolution of axisymmetric toroidal magnetic field $\overline{B_\phi}$ for a simulation in a thick shell ($\eta=0.2$, $N_\rho=3$, $\text{Ra}=1.1\times 10^{7}$, upper panel) and a simulation in a thin shell ($\eta=0.6$, $N_\rho=1$, $\text{Ra}=10^{6}$, lower panel). Times are given here in magnetic diffusion units. Red (blue) correspond to positive (negative) values.[]{data-label="fig:parker"}](fig/parkerwaveEta02 "fig:"){width="9cm"} ![Time evolution of axisymmetric toroidal magnetic field $\overline{B_\phi}$ for a simulation in a thick shell ($\eta=0.2$, $N_\rho=3$, $\text{Ra}=1.1\times 10^{7}$, upper panel) and a simulation in a thin shell ($\eta=0.6$, $N_\rho=1$, $\text{Ra}=10^{6}$, lower panel). Times are given here in magnetic diffusion units. Red (blue) correspond to positive (negative) values.[]{data-label="fig:parker"}](fig/parkerwaveEta06 "fig:"){width="9cm"} ![Butterfly diagrams: time evolution of axisymmetric toroidal magnetic field $\overline{B_\phi}$ at $r=0.9 r_o$ for the two models of Fig. \[fig:parker\]. Time is given here in magnetic diffusion units. Red (blue) correspond to positive (negative) values.[]{data-label="fig:butterfly"}](fig/butterflyeta02 "fig:"){width="8cm"} ![Butterfly diagrams: time evolution of axisymmetric toroidal magnetic field $\overline{B_\phi}$ at $r=0.9 r_o$ for the two models of Fig. \[fig:parker\]. Time is given here in magnetic diffusion units. Red (blue) correspond to positive (negative) values.[]{data-label="fig:butterfly"}](fig/butterflyeta06 "fig:"){width="8cm"} The $\alpha\Omega$ or $\alpha^2\Omega$ dynamos on the multipolar branch also show a distinct and interesting oscillatory time dependence. Figures \[fig:parker\]-\[fig:butterfly\] illustrate two examples. The magnetic field is weak, multipolar and dominated by non-axisymmetric components in both cases (see for example Figs. \[fig:brsurfEta02\]-\[fig:brsurfEta06\]). For the thin shell (lower panels), a dipolar solution coexists at identical parameters. Toroidal magnetic field first emerges at low latitudes, then travels towards the poles and finally vanishes roughly where the tangent cylinder seems to prevent a further movement towards even higher latitudes [@Schrinner11]. When the patches have travelled half way, the cycle starts over with field of opposite polarity appearing close to the equatorial plane. The thick shell solution (top panels) demonstrates that during some periods one hemisphere clearly dominates. The characteristic period is roughly $0.1\tau_\lambda - 0.2\tau_\lambda$ where $\tau_\lambda=d^2/\lambda$ is the magnetic diffusion time. These are Parker dynamo waves, as we will demonstrate below, very similar to those that have been previously identified in some Boussinesq simulations [@Goudard08; @Schrinner11; @Simitev12]. These coherent oscillations lead to the periodic patterns observed in the “butterfly-diagram” style illustrations shown in Fig. \[fig:butterfly\]. The diagrams highlight the direction of travel which in Parker waves is controlled by the gradient of zonal flows: they travel polewards when like in our simulations the differential rotation decreases with depth [e.g. @Yoshimura75]. The mean-field formalism allows to derive a dispersion relation for these dynamo waves [@Busse06; @Schrinner11]. In this formalism, magnetic field and flow are decomposed into the axisymmetric or mean contributions denoted by overbars $$\overline{\vec{B}} = \vec{B}_p+B\,\vec{e_\phi}=\vec{\nabla}\times (A\ \vec{e_\phi})+B\,\vec{e_\phi} \quad \text{;} \quad \overline{\vec{u}} = u\, \vec{e_\phi},$$ and the primed non-axisymmetric or fluctuating contributions $\vec{B'}$ and $\vec{u'}$. This allows to reformulate the axisymmetric part of the induction equation (\[eq:induction\]) to $$\left\lbrace \begin{aligned} \dfrac{\partial A}{\partial t} & = \alpha B + \dfrac{1}{\text{Pm}}\nabla^2 A, \\ \dfrac{\partial B}{\partial t} & = \vec{B}_p\cdot \vec{\nabla}\,u + (\vec{\nabla}\times \alpha\vec{B_p}) \cdot\vec{e_\phi} + \dfrac{1}{\text{Pm}}\nabla^2 B, \end{aligned} \right. \label{eq:meanfields}$$ where the fluctuating mean electromotive force is approximated by an $\alpha$-effect, i.e. by $\overline{\vec{u'}\times\vec{B'}}=\alpha\, \vec{\overline{B}}$. To simplify this system we assume a homogeneous $\alpha$ and also adopt the plane layer formalism introduced by [@Parker55], where the cartesian coordinates $(x,y,z)$ correspond to $(\phi,\theta,r)$. Further assuming that $\overline{u}$ depends only on $z$ (radius) leads to $$\left\lbrace \begin{aligned} \dfrac{\partial A}{\partial t} & = \alpha B + \dfrac{1}{\text{Pm}}\nabla^2 A, \\ \dfrac{\partial B}{\partial t} & = -\alpha \nabla^2 A + \dfrac{du}{dz} \dfrac{\partial A}{\partial y} + \dfrac{1}{\text{Pm}}\nabla^2 B. \end{aligned} \right.$$ The ansatz $A, B \propto \exp(i(k_y y + k_z z)+\lambda t)$ with $\lambda = \tau +i\omega$ allows to derive the following dispersion relation $$\left(\lambda + \dfrac{|\vec{k}|^2}{\text{Pm}}\right)^2 =\alpha^2\,|\vec{k}|^2 + i k_y\,\alpha\,\dfrac{du}{dz}.$$ For $\alpha\Omega$ dynamos we concentrated on in the following, the imaginary part of $\lambda$ provides $$\omega = {\left(}\dfrac{1}{2}\,\alpha\,\dfrac{du}{dz}\,k_y {\right)}^{1/2}.$$ To further simplify this dispersion relation we assume that $z$ variations are of the order of the shell radius, i.e. $k_y\sim 1/r_o$ and that the shear can be approximated by $du/dz \sim \overline{u_\phi}/d \sim \text{Re}_{\text{zon}}$. The $\alpha$ is particularly difficult to estimate. A full derivation of the $\alpha$ tensor would, for example, require to employ the test-field method [e.g. @Schrinner07]. For homogeneous and isotropic MHD turbulence, however, it may at least crudely be approximated via the fluctuating kinetic helicity: $$\alpha = -\dfrac{\tau_c}{3}\,\overline{ \vec{u}'\cdot \nabla\times\vec{u}'}, \label{eq:helicity}$$ where $\tau_c$ is the typical lifetime of a convective feature [e.g. @Brandenburg05]. Following [@Brown10], we use $\tau_c=H_\rho/u'$ when the local density scale height $H_\rho=(d\ln{\bar{\rho}}/dr)^{-1}$ is smaller than $d$ and $\tau_c=d/u'$ otherwise. This finally leads to the simplified dispersion relation for Parker waves $$\omega \sim {\left(}\dfrac{\alpha\,\text{Re}_{\text{zon}}}{2\,r_o} {\right)}^{1/2}. \label{eq:freq}$$ ![Frequencies of oscillatory dynamos plotted against the dispersion relation of Parker waves given in Eq. (\[eq:freq\]). Red (grey) symbols correspond to simulations in thick (thin) shells ($\eta=0.2$ and $\eta=0.6$, respectively). Each type of symbol is associated to a given density stratification.[]{data-label="fig:freq"}](fig/frequency){width="9cm"} Figure \[fig:freq\] shows the comparison between the frequencies predicted by this simplified dispersion relation and those found in the numerical models. The later frequencies are obtained using a Fourier transform of the butterfly diagrams displayed in Fig \[fig:butterfly\]. The resulting $\widehat{B_\phi}(\theta, \omega)$ is then integrated over the colatitude $\theta$ to derive the power spectrum. Even though individual predictions fail by up to a factor two, the general agreement is quite convincing given the numerous approximation involved and is similar to that reached in previous Boussinesq studies [@Busse06; @Schrinner11]. This confirms that the oscillations in our multipolar cases are indeed Parker waves. When the magnetic Reynolds number becomes too large ($\text{Rm} \gtrsim 150-200$), the coherence of these oscillations is gradually lost and a definite frequency is increasingly difficult to determine. This limits our analysis to lower $\text{Re}_{\text{zon}}$ and $\alpha$ values and to frequencies $\omega<300$. Discussion and conclusions {#sec:discussion} ========================== We have investigated the influence of background density stratification on convection-driven dynamos in a rotating spherical shell. The use of the anelastic approximation allowed us to exclude sound waves and the related short time steps [@Glatz1; @Clune99; @Jones09]. Previous Boussinesq results have shown that inertial effects play a decisive role for determining the magnetic field geometry. When inertia becomes influential, only multipolar solutions with weak magnetic fields are possible. When inertia is weak, two solutions can coexist: dipole-dominated solutions with strong magnetic fields are found for stress-free as well as rigid boundary conditions. For stress-free and mixed boundary conditions, a multipolar branch is found at identical parameters [e.g. @Simitev09; @Schrinner12]. Alternatively, the recent study by [@King12] suggests that the transition between dipolar and multipolar dynamos may occur when inertia becomes larger than viscous forces. This transition is accompanied by an abrupt decrease of the kinetic helicity. These results however seem to be in contradiction with previous studies where inertia is always larger than viscosity [e.g. @Wicht10] and further investigations are required to clarify this contradiction. Our anelastic simulations confirm this scenario for mild stratifications corresponding to $N_\rho<1.8$. The reason for the bistability is a competition between zonal winds and dipolar magnetic fields. Strong dipolar magnetic fields prevent significant zonal winds to develop. Strong zonal winds, on the other hand prevent the production of significant dipolar fields. The two branches also differ in the induction mechanism. Strong zonal winds promote an $\Omega$-effect which leads to a $\alpha\Omega$ or $\alpha^2\Omega$ type of dynamo, while dipole-dominated magnetic fields are typically generated in an $\alpha^2$ process. The sizable axisymmetric toroidal fields produced by an $\Omega$-effect typically lead to a coherent cyclic time evolution of the magnetic field for moderate magnetic Reynolds numbers ($\text{Rm} < 200$). This is consistent with previous Boussinesq studies [e.g. @Schrinner07] and numerical models of young solar-type stars [@Brown11] and has been identified as Parker waves. Contrary to what is observed in the solar cycle, these waves start at the equator and travel towards the poles because of the opposite sign in the zonal shear. For stronger stratification with $N_\rho>1.8$, the dipolar branch is lost and close to onset of dynamo action a new magnetic mode characterised by a large wave number $m=2$ appears. This is likely due to the concentration of the $\alpha$-effect into a narrow region close the equator. According to mean-field models, this would preferentially promote non-axisymmetric dynamos [see also @Chabrier06] with large wave numbers. The collapse of the dipolar branch may explain the differences between the weakly stratified and significantly dipolar simulations by [@Dobler06] ($\rho_{\text{bot}}/\rho_{\text{top}} \simeq 5$, i.e. $N_\rho=1.6$) and the strongly stratified anelastic and multipolar models by [@Browning08] ($\rho_{\text{bot}}/\rho_{\text{top}} \simeq 100$, i.e. $N_\rho=4.6$). Numerical limitations force us to use excessively large diffusivities in our simulations. Ekman numbers are thus orders of magnitude too large and Reynolds number orders of magnitude too low. When extrapolated to fast rotating planets and stars this type of models nevertheless provides, for example, realistic magnetic field strengths [@Olson06; @Christensen09]. This gives us confidence to compare the observed field geometries of planets and stars with predictions based on our simulation results. Our result are compatible with the dipole-dominated magnetic fields on Jupiter and Saturn that are generated in their deeper metallic envelopes where the density stratification is only mild [roughly $N_\rho\sim 1.5-2$, see @Heimpel11; @Nettelmann12]. Since the local Rossby number is very small for these planets, however, bistability seems to be an option. This is also the case for Uranus and Neptune, where $\text{Ro}_\ell$ and the density contrast within the dynamo region are small [@Hubbard91; @Olson06]. Their multipolar magnetic field would then suggest that these planets would then occupy the alternative branch offered by the bistability phenomenon. Concerning rapidly rotating low-mass stars that may also fall into the low $\text{Ro}_\ell$ regime, the spectropolarimetric observations of [@lateM] suggest that late M stars with very similar parameters (mass and rotation rate) come in two categories: some stars present a strong dipole-dominated magnetic field while others show weaker and multipolar magnetic structures. These two geometries may represent the two coexisting dynamo branches at smaller local Rossby numbers. For the bistability to be a viable explanation, however, our simulations suggest that the dynamos must operate in a region with moderate density stratification ($N_\rho \sim 1-2$) and supercriticality ($\text{Ra}/\text{Ra}_c \sim 10-30$). Since the stellar dynamos operate presumably far from onset of convection, only multipolar fields would then be possible. However, when decreasing the Ekman number towards more realistic values, the simulations by [@Christensen06] suggest that the dipolar window may persist at higher supercriticalities. In addition, since these stars have huge density contrasts, our investigation would then generally predict multipolar fields. A further exploration of the parameter space seems here required to clarify this point. [@Stanley09], for example, suggest that lower Prandtl number may help in creating stronger dipole fields. Considering radial-dependent properties (e.g. viscosity, thermal diffusivity and electrical diffusivity) is also known to have a strong impact on the location of the convective columns that could possibly help to avoid the concentration of helicity close to the equator. All the computations have been carried out on the GWDG computer facilities in Göttingen. This work was supported by the Special Priority Program 1488 (PlanetMag, <http://www.planetmag.de>) of the German Science Foundation. It is a pleasure to thank C. A. Jones for providing us his linear stability code. [57]{} natexlab\#1[\#1]{} , F. M., [Heimpel]{}, M. H., & [Aurnou]{}, J. M. 2004, Geophysical and Astrophysical Fluid Dynamics, 98, 153 , J. 2005, Journal of Fluid Mechanics, 542, 53 , J., [Labrosse]{}, S., & [Poitou]{}, C. 2009, Geophysical Journal International, 179, 1414 , A. P., [Kuzanyan]{}, K. M., [Sokoloff]{}, D., & [Soward]{}, A. M. 2005, Geophysical and Astrophysical Fluid Dynamics, 99, 309 , N. & [Brun]{}, A. S. 2011, Astronomische Nachrichten, 332, 1045 , S. I. & [Roberts]{}, P. H. 1995, Geophysical and Astrophysical Fluid Dynamics, 79, 1 , A. & [Subramanian]{}, K. 2005, , 417, 1 , B. P., [Browning]{}, M. K., [Brun]{}, A. S., [Miesch]{}, M. S., & [Toomre]{}, J. 2010, , 711, 424 , B. P., [Miesch]{}, M. S., [Browning]{}, M. K., [Brun]{}, A. S., & [Toomre]{}, J. 2011, , 731, 69 , M. K. 2008, , 676, 1262 , A. S., [Browning]{}, M. K., & [Toomre]{}, J. 2005, , 629, 461 , F. H. & [Simitev]{}, R. D. 2006, Geophysical and Astrophysical Fluid Dynamics, 100, 341 , G. & [K[ü]{}ker]{}, M. 2006, , 446, 1027 , U. R. 2006, , 444, 1056 , U. R. 2010, , 152, 565 , U. R. & [Aubert]{}, J. 2006, Geophysical Journal International, 166, 97 , U. R., [Holzwarth]{}, V., & [Reiners]{}, A. 2009, , 457, 167 , T. C., [Elliott]{}, J. R., [Miesch]{}, M. S., [Toomre]{}, J., & [Glatzmaier]{}, G. A. 1999, Parallel Computing, 25, 361 , W., [Stix]{}, M., & [Brandenburg]{}, A. 2006, , 638, 336 , J., [Forveille]{}, T., [Cameron]{}, A. C., [et al.]{} 2006, Science, 311, 633 , J., [Morin]{}, J., [Petit]{}, P., [et al.]{} 2008, , 390, 545 , T. & [Wicht]{}, J. 2012, , 219, 428 , P. A. & [Glatzmaier]{}, G. A. 1981, , 45, 335 , G. A. 1984, Journal of Computational Physics, 55, 461 , L. & [Dormy]{}, E. 2008, Europhysics Letters, 83, 59001 , E. & [Busse]{}, F. H. 2000, , 62, 4457 , M. & [G[ó]{}mez P[é]{}rez]{}, N. 2011, , 38, L14201 , W. B., [Nellis]{}, W. J., [Mitchell]{}, A. C., [et al.]{} 1991, Science, 253, 648 , J. & [Wang]{}, J.-X. 2006, , 6, 227 , C. A., [Boronski]{}, P., [Brun]{}, A. S., [et al.]{} 2011, , 216, 120 , C. A. & [Kuzanyan]{}, K. M. 2009, , 204, 227 , C. A., [Kuzanyan]{}, K. M., & [Mitchell]{}, R. H. 2009, Journal of Fluid Mechanics, 634, 291 , C. & [Christensen]{}, U. R. 2002, Physics of the Earth and Planetary Interiors, 131, 29 , S. R. & [Fan]{}, Y. 1999, , 121, 247 , H. K. 1978, [Magnetic field generation in electrically conducting fluids]{} (Cambridge, England, Cambridge University Press, 1978. 353 p.) , J., [Donati]{}, J., [Petit]{}, P., [et al.]{} 2008, , 390, 567 , J., [Donati]{}, J.-F., [Petit]{}, P., [et al.]{} 2010, , 407, 2269 , V. & [Dormy]{}, E. 2009, International Journal of Modern Physics B, 23, 5467 , N., [Becker]{}, A., [Holst]{}, B., & [Redmer]{}, R. 2012, , 750, 52 , P., [Christensen]{}, U., & [Glatzmaier]{}, G. A. 1999, , 104, 10383 , P. & [Christensen]{}, U. R. 2006, Earth and Planetary Science Letters, 250, 561 , E. N. 1955, , 122, 293 , G., [Elstner]{}, D., & [Ossendrijver]{}, M. 2003, , 406, 15 , A. A., [Sokolov]{}, D. D., & [Starchenko]{}, S. V. 1988, , 115, 5 , Y., [Takehiro]{}, S.-I., [Kuramoto]{}, K., & [Hayashi]{}, Y.-Y. 2011, Physics of the Earth and Planetary Interiors, 188, 203 , N. & [Cardin]{}, P. 2006, Earth and Planetary Science Letters, 245, 595 , M., [Petitdemange]{}, L., & [Dormy]{}, E. 2011, , 530, A140 , M., [Petitdemange]{}, L., & [Dormy]{}, E. 2012, , 752, 121 , M., [R[ä]{}dler]{}, K.-H., [Schmitt]{}, D., [Rheinhardt]{}, M., & [Christensen]{}, U. R. 2007, Geophysical and Astrophysical Fluid Dynamics, 101, 81 , R. D. & [Busse]{}, F. H. 2009, Europhysics Letters, 85, 19001 , R. D. & [Busse]{}, F. H. 2012, , 749, 9 , K. M., [King]{}, E. M., & [Aurnou]{}, J. M. 2012, Earth and Planetary Science Letters, 333, 9 , B. & [Jones]{}, C. A. 2006, Geophysical Journal International, 164, 467 , S. & [Glatzmaier]{}, G. A. 2010, , 152, 617 , J. 2002, Physics of the Earth and Planetary Interiors, 132, 281 , J. & [Christensen]{}, U. R. 2010, Geophysical Journal International, 181, 1367 , H. 1975, , 201, 740 [ ]{}

ArXiv

--- abstract: 'We define and study the space of *mixed modular symbols* for a given finite index subgroup $\Gamma$ of $\operatorname{SL}_2(\mathbf{Z})$. This is an extension of the usual space of modular symbols, which in some cases carries more information about Eisenstein series. We make use of mixed modular symbols to construct some $1$-motives related to the generalized Jacobian of modular curves. In the case $\Gamma = \Gamma_0(p)$ for some prime $p$, we relate our construction to the work of Ehud de Shalit on $p$-adic periods of $X_0(p)$.' author: - Emmanuel Lecouturier bibliography: - 'biblio.bib' title: 'Mixed modular symbols and the generalized cuspidal $1$-motive' --- Introduction {#section_introduction} ============ Overview and motivations {#intro_overview .unnumbered} ------------------------ Consider the torsion free abelian group $\mathcal{M}$ generated by the symbols $\{\alpha, \beta\}$ where $(\alpha, \beta) \in \mathbf{P}^1(\mathbf{Q})^2$ with the following relations $$\{\alpha, \beta\}+\{\beta, \gamma\}+\{\gamma, \alpha\}=0$$ for $(\alpha, \beta, \gamma) \in \mathbf{P}^1(\mathbf{Q})^3$. We denote by $\operatorname{GL}_2^+(\mathbf{Q})$ the subgroup of $\operatorname{GL}_2(\mathbf{Q})$ consisting of positive determinant matrices. We let $\operatorname{GL}_2^+(\mathbf{Q})$ act on the left on $\mathbf{P}^1(\mathbf{Q})$. If $\Gamma$ is a finite index subgroup of $\operatorname{SL}_2(\mathbf{Z})$, we denote by $\mathcal{M}_{\Gamma}$ the largest torsion-free quotient of $\mathcal{M}$ on which $\Gamma$ acts trivially. This group is usually called the group of modular symbols of level $\Gamma$ (and weight $2$). The group $\mathcal{M}_{\Gamma}$ is canonically isomorphic to the relative homology group $H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z})$, where $X_{\Gamma}$ is the compact modular curve associated to $\Gamma$ and $\mathcal{C}_{\Gamma}$ is the set of cusps of $X_{\Gamma}$. By intersection duality, we have a canonical isomorphism $\mathcal{M}_{\Gamma} \xrightarrow{\sim} \operatorname{Hom}_{\mathbf{Z}}(H_1(Y_{\Gamma}, \mathbf{Z}), \mathbf{Z})$ where $Y_{\Gamma} = X_{\Gamma} - \mathcal{C}_{\Gamma}$ is the open modular curve. If $\Gamma$ is a congruence subgroup, there is a Hecke-equivariant injective map of $\mathbf{R}$-vector spaces $$p_{\Gamma} \otimes \mathbf{R}: H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{R} \hookrightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$$ given by $c \otimes 1 \mapsto \left(f \mapsto \int_{c} 2i\pi f(z)dz\right)$. Here, $M_2(\Gamma)$ is the complex vector space of modular forms of weight $2$ and level $\Gamma$ (this includes Eisenstein series). The map $p_{\Gamma}$ is not surjective in general, since the dimension of the left hand side is $2g(\Gamma)+c(\Gamma)-1$ whereas the dimension of the right hand side is $2g(\Gamma)+2\cdot(c(\Gamma)-1)$, where $g(\Gamma)$ (resp. $c(\Gamma)$) is the genus (resp. the number of cusps) of $X_{\Gamma}$. Dually one copy of Eisenstein series is missing to the space $\mathcal{M}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{C}$, which is a well-known Hodge theoretic phenomenon for non-projective smooth curves. In this note, we define an extension $\tilde{\mathcal{M}}_{\Gamma}$ of $\mathcal{M}_{\Gamma}$, called the space of *mixed modular symbols* of level $\Gamma$, with rank equal to $2\dim_{\mathbf{C}} M_2(\Gamma)$. The definition does not require any assumption on $\Gamma$ (in particular, it could be a non-congruence subgroup of $\operatorname{SL}_2(\mathbf{Z})$). The map $p_{\Gamma}$ extends to a Hecke-equivariant map $\tilde{p}_{\Gamma} \otimes \mathbf{R} : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$. The construction makes use of a cocycle construction by Glenn Stevens [@Stevens_book], and involves special values of $L$-functions. We define and study various objects related to the space of mixed modular symbols, namely Hecke operators, the complex conjugation, Manin symbols, intersection duality and its relation with the extended Petersson pairing of Don Zagier [@Zagier_Rankin] and Vinsentiu Pasol and Alexandru A. Popa [@Popa_Haberland], relation with generalized Jacobians and the associated $\ell$-adic Galois representations. We also study in more details the particular case $\Gamma = \Gamma_0(p)$ were $p$ is prime. In this case, we relate our construction to the one of Ehud de Shalit on generalized $p$-adic periods [@deShalit]. This will be used in our forthcoming work on the Mazur-Tate conjecture in conductor $p$ [@Lecouturier_MT]. We now describe in more details our main results. Definition of mixed modular symbols {#intro_def_mixed_modsymb} ----------------------------------- Consider the torsion-free abelian group $\tilde{\mathcal{M}}$ generated by the symbols $\{g,g'\}$ where $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$ with the following relations: 1. $\{g,g'\}+\{g',g''\}+\{g'',g\}=0$ for all $(g,g',g'') \in \operatorname{SL}_2(\mathbf{Z})^3$; 2. $\{g,g'\}-\{\epsilon_1g, \epsilon_2 g'\}=0$ for all $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$ and $(\epsilon_1,\epsilon_2)\in \{1,-1\}^2$; 3. $\{g,gT^n\}-n\cdot \{g,gT\}=0$ for all $g\in \operatorname{SL}_2(\mathbf{Z})$, $n\in \mathbf{Z}$. In this note, we let $T = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}$. There is a left action of $\operatorname{SL}_2(\mathbf{Z})$ on $\tilde{\mathcal{M}}$, given by $g \cdot \{g',g''\} = \{gg',gg''\}$ for all $(g,g',g'') \in \operatorname{SL}_2(\mathbf{Z})^3$. We let $\tilde{\mathcal{M}}_{\Gamma}$ be the largest torsion-free quotient of $\tilde{\mathcal{M}}$ on which $\Gamma$ acts trivially. There is a surjective group homomorphism $$\pi_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \mathcal{M}_{\Gamma}$$ given by $\pi_{\Gamma}(\{g,g'\}) = \{g\infty, g'\infty\}$. We show in Proposition \[rank\_computation\] that $\operatorname{Ker}(\pi_{\Gamma})$ is isomorphic to the cokernel of the group homomorphism $\mathbf{Z}\rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]$ given by $1 \mapsto \frac{1}{d_{\Gamma}} \sum_{c \in \mathcal{C}_{\Gamma}} e_c\cdot [c]$, where $e_c$ is the width of the cusp $c$ and $d_{\Gamma} = \gcd(e_c)_{c \in \mathcal{C}_{\Gamma}}$. In particular, $\tilde{\mathcal{M}}_{\Gamma}$ has the required rank. There is a boundary map $\partial_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$ (the upper $0$ meaning the augmentation subgroup), whose kernel contains $H_1(Y_{\Gamma}, \mathbf{Z}) \hookrightarrow \tilde{\mathcal{M}}_{\Gamma}$ with finite index equal to $\frac{1}{d_{\Gamma}}\prod_{c \in \mathcal{C}_{\Gamma}} e_c$. The generalized period map $\tilde{p}_{\Gamma}$ {#intro_generalized_periods} ----------------------------------------------- In §\[section\_period\_iso\], we extend the period map $p_{\Gamma}$ to a group homomorphism $$\tilde{p}_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \text{ .}$$ One of the main properties of $\tilde{p}_{\Gamma}$ is that it encodes the special values of $L$-functions. More precisely, for any $g \in \operatorname{SL}_2(\mathbf{Z})$ and any $f \in M_2(\Gamma)$, we have $\tilde{p}_{\Gamma}(\{g,gS\}) = L(f \mid g,1)$ where $S = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ and as usual $(f \mid g)(z) = (cz+d)^{-2}\cdot f(\frac{az+b}{cz+d})$. The map $p_{\Gamma} \otimes \mathbf{R}$ is not always an isomorphism (equivalently injective), for instance if $\Gamma$ is a principal congruence subgroup of level divisible by $6$. Nevertheless, we prove the following result. \[intro\_thm\_period\] The map $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism if $\Gamma$ is a congruence subgroup of level $p^n$ for some prime $p$ and integer $n \geq 1$. If $\Gamma=\Gamma_0(N)$ or $\Gamma = \Gamma_1(N)$ for some integer $N \geq 1$, let $\mathbb{T}$ be the Hecke algebra over $\mathbf{Z}$ acting faithfully on $M_2(\Gamma)$. We define in §\[paragraph\_Hecke\] an action of $\mathbb{T}$ on $\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}]$. More precisely, if $n \geq 1$ is an integer prime to $2N$, then the Hecke operator $T_n$ stabilises $\tilde{\mathcal{M}}_{\Gamma}$. However, it turns out that if $p$ is a prime dividing $N$ (resp. $2$) then $U_p$ (resp. $T_2$ or $U_2$) sends $\tilde{\mathcal{M}}_{\Gamma}$ into $\frac{1}{p}\cdot \tilde{\mathcal{M}}_{\Gamma}$, so it may not stabilises $\tilde{\mathcal{M}}_{\Gamma}$ in general. The reason for this non-integrality phenomenon is that there is a priori no direct way to define a general double coset action on $\tilde{\mathcal{M}}_{\Gamma}$: while $\operatorname{GL}_2^+(\mathbf{Q})$ acts naturally on $\mathbf{P}^1(\mathbf{Q})$, it does not acts naturally on $\operatorname{SL}_2(\mathbf{Z})$. To resolve this issue, we embed $\tilde{\mathcal{M}}_{\Gamma}$ as a lattice inside a $\mathbf{Q}$-vector space on which there is a natural action of the double coset operators (this vector space is defined in a similar way as $\tilde{\mathcal{M}}_{\Gamma}$, replacing $\operatorname{SL}_2(\mathbf{Z})$ by $\operatorname{GL}_2^+(\mathbf{Q})$). We also define similarly an action of the Atkin–Lehner involution $W_N$ on $\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{N}]$. In §\[paragraph\_complex\_conjugation\], we define an action of the complex conjugation on $\tilde{\mathcal{M}}_{\Gamma}$. The action of Hecke operators and the complex conjugation are compatible with the projection $\pi_{\Gamma}$ and with the embedding $H_1(Y_{\Gamma}, \mathbf{Z}) \hookrightarrow \tilde{\mathcal{M}}_{\Gamma}$. Manin symbols {#intro_Manin_symbols} ------------- In §\[paragraph\_Manin\], we define Manin symbols in $\tilde{\mathcal{M}}_{\Gamma}$. Let $$\tilde{\xi}_{\Gamma} : \mathbf{Z}[\Gamma\backslash \operatorname{PSL}_2(\mathbf{Z})]\rightarrow \tilde{\mathcal{M}}_{\Gamma}$$ be the map defined by $\tilde{\xi}_{\Gamma}(\Gamma g) = \{g,gS\}$ for all $g\in \operatorname{PSL}_2(\mathbf{Z})$. This is the Manin map, and the element $\tilde{\xi}_{\Gamma}(\Gamma g)$ is the Manin symbol in $\tilde{\mathcal{M}}_{\Gamma}$ associated to $\Gamma g$. We give a simple and concrete description of the image of $\tilde{\xi}_{\Gamma}$. In particular, we prove the following. \[intro\_thm\_Manin\] Let $p\geq 5$ be a prime and $n \in \mathbf{N}$. If $\Gamma=\Gamma_1(p^n)$ or $\Gamma = \Gamma_0(p^n)$, then the image of $\tilde{\xi}_{\Gamma}$ has index dividing $3$, the divisibility being strict of and only if $p \equiv 1 \text{ (modulo 3}\text{)}$ and $\Gamma = \Gamma_0(p^n)$. If $\Gamma = \Gamma_1(N)$ or $\Gamma = \Gamma_0(N)$ for some odd $N \geq 1$, we were not able to determine when the image of $\tilde{\xi}_{\Gamma}$ has finite index in $\tilde{\mathcal{M}}_{\Gamma}$ (we know it is not of finite index when $N$ is even). The question seems to be related to additive number theory in $(\mathbf{Z}/N\mathbf{Z})^{\times}$. See Remark \[Manin\_rmk\] for more details. We also give a description of the Manin relations, [*i.e.* ]{}of the group $\operatorname{Ker}(\tilde{\xi}_{\Gamma})$. It turns out that the $2$-terms Manin relations $\tilde{\xi}(\Gamma g)+\tilde{\xi}(\Gamma gS)=0$ are satisfied, but the $3$-terms Manin relations $\tilde{\xi}(\Gamma g)+\tilde{\xi}(\Gamma gU)+\tilde{\xi}(\Gamma gU^2)=0$, where $U = \begin{pmatrix} 1 & -1 \\ 1 & 0 \end{pmatrix}$, are not satisfied in general. Instead, we have to replace them with a subgroup of relations which is described in Theorem \[generalized\_Manin\_thm\]. Duality theory and the generalized Petersson product {#intro_duality} ---------------------------------------------------- In §\[paragraph\_Petersson\], we study an anti-symmetric bilinear pairing $\langle \cdot , \cdot \rangle : \tilde{\mathcal{M}}_{\Gamma}^* \times \tilde{\mathcal{M}}_{\Gamma}^* \rightarrow \frac{1}{6}\cdot \mathbf{Z}$ defined by the formula $$\begin{aligned} \langle \varphi_1, \varphi_2 \rangle = \frac{1}{6}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} & \varphi_1(\{gS,g\}) \cdot \varphi_2(\{gTS,gT\}) - \varphi_1(\{gTS,gT\})\cdot \varphi_2(\{gS,g\}) \\&-4\cdot \varphi_1(\{g,gT\})\cdot\varphi_2(\{g,gS\})+4\cdot\varphi_1(\{g,gS\})\cdot \varphi_2(\{g,gT\}) \text{ ,}\end{aligned}$$ where $\tilde{\mathcal{M}}_{\Gamma}^* = \operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{Z})$. By tensoring with $\mathbf{C}$, $\langle \cdot , \cdot \rangle$ extends to an anti-symmetric bilinear pairing on $\operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{C})$. The map $\tilde{p}_{\Gamma}$ induces a map $ \tilde{p}_{\Gamma}^* : M_2(\Gamma) \rightarrow \operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{C})$. Following Zagier, Pasol and Popa [@Popa_Haberland] extended the Pertersson pairing on $M_2(\Gamma)$, using a procedure of renormalization of divergent integrals. They generalize a formula of Klaus Haberland and Loïc Merel as follows: for any $f_1$, $f_2$ $\in M_2(\Gamma)$, we have $$-8 i \pi^2 (f_1, f_2) = \frac{1}{[\operatorname{SL}_2(\mathbf{Z}):\Gamma]}\cdot \langle \tilde{p}_{\Gamma}^*(f_1) , \overline{\tilde{p}_{\Gamma}^*(f_2)} \rangle \text{ .}$$ We expect that $\langle \cdot , \cdot \rangle $ is $\mathbf{Z}$-valued and non-degenerate. More precisely, we expect that the determinant of the pairing $\langle \cdot , \cdot \rangle $ is $\frac{1}{d_{\Gamma}}\cdot \prod_{c \in \mathcal{C}_{\Gamma}} e_c$. In this direction, we prove the following result. \[intro\_thm\_Petersson\] Assume that the cusps of $X_{\Gamma}$ are fixed by the complex conjugation (if $\Gamma = \Gamma_1(N)$, this is the case if and only if $N$ divides $2p$ for some prime $p$). Then $\langle \cdot , \cdot \rangle $ is perfect after inverting $2$ and the $\operatorname{lcm}$ of the widths of the cusps of $X_{\Gamma}$. See also Remark \[rem\_generality\_conjecture\_G\] and Proposition \[expression\_G\^+\_rho\_i\] for a partial result in the general case where the cusps are not assumed to be real. Relation with the generalized Jacobian {#intro_generalized_cuspidal_motive} -------------------------------------- Assume in this paragraph that $\Gamma$ is a congruence subgroup of $\operatorname{SL}_2(\mathbf{Z})$, whose level is denoted by $N$. Recall that for each cusp $c \in \mathcal{C}_{\Gamma}$, we have denoted by $e_c$ the width of $c$. Fix an algebraic closure $\overline{\mathbf{Q}}$ of $\mathbf{Q}$ and an embedding $\overline{\mathbf{Q}} \hookrightarrow \mathbf{C}$. We let $\zeta_N = e^{\frac{2i\pi}{N}} \in \overline{\mathbf{Q}}$. Let $k$ be the number field of definition of $X_{\Gamma}$ in $\overline{\mathbf{Q}}$; this is a subfield of $\mathbf{Q}(\zeta_N)$. Let $J_{\Gamma}$ be the Jacobian variety of $X_{\Gamma}$ over $k$. Let $J_{\Gamma}^{\#}$ be the generalized Jacobian variety of $X_{\Gamma}$ over $k$ with respect to the cuspidal divisor, [*i.e.* ]{}the sum of the closed points of $X_{\Gamma}\backslash Y_{\Gamma}$ over $k$. By definition, $J_{\Gamma}^{\#}$ parametrizes degree zero divisors supported on $Y_{\Gamma}$ modulo the divisors of functions which are constant ($\neq 0, \infty$) on the cusps. We have a canonical exact sequence of group schemes over $k$: $$1 \rightarrow \mathbf{G}_m \rightarrow \prod_{c \in X_{\Gamma}\backslash Y_{\Gamma}} \text{Res}_{k(c)/k}(\mathbf{G}_m)\rightarrow J_{\Gamma}^{\#} \rightarrow J_{\Gamma} \rightarrow 1 \text{ ,}$$ where $k(c)\subset \mathbf{C}$ is the field of definition of the cusp $c$ and $\text{Res}_{k(c)/k}$ denotes the Weil restriction. Let $C_{\Gamma} \subset J_{\Gamma}$ be the cuspidal subgroup, [*i.e.* ]{}the subgroup generated by the image of the difference of the cusps of $X_{\Gamma}$. This is a finite group by the Manin–Drinfeld theorem [@Manin_Drinfeld]. Our goal is to try to define a reasonable “lift” $C^{\#}_{\Gamma}$ of $C_{\Gamma}$ in $J_{\Gamma}^{\#}$. For technical (but seemingly necessary) reasons, we only define a lift $C_{\Gamma}^{\natural}$ of $C_{\Gamma}$ in $J_{\Gamma}^{\natural} := J_{\Gamma}^{\#}/\prod_{c \in X_{\Gamma}\backslash Y_{\Gamma}} \text{Res}_{k(c)/k}(\mu_{N})$, where $\mu_n \subset \mathbf{G}_m$ is the group of $n$th roots of unity. Note that there is a canonical projection $$\beta_{\Gamma}^{\natural} : J_{\Gamma}^{\natural} \rightarrow J_{\Gamma}$$ induced by the projection $$\beta_{\Gamma}^{\#} : J_{\Gamma}^{\#} \rightarrow J_{\Gamma} \text{ .}$$ Let $\delta_{\Gamma} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}$ be the map sending a divisor to its class in $J_{\Gamma}$. We define two maps $\delta_{\Gamma}^{\natural, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}(\mathbf{C})$ and $\delta_{\Gamma}^{\natural, \operatorname{an}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}(\mathbf{C})$ such that $$\beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{alg}} = \beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{an}} = \delta_{\Gamma} \text{ .}$$ The map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ is defined in an algebraic way using uniformizers at cusps, whereas the map $\delta_{\Gamma}^{\natural, \operatorname{an}}$ is defined in an analytic way (via an Abel-Jacobi map) using mixed modular symbols and the generalized period map $\tilde{p}_{\Gamma}$. The fact that we were only able to define $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ as a map valued in $J_{\Gamma}^{\natural}(\mathbf{C})$ and not in $J_{\Gamma}^{\#}(\mathbf{C})$ is explained by the fact that our choice of a uniformizer at a cusp $c$ is canonical only up to a $e_c$-th root of unity. Similarly, the fact that we were only able to define $\delta_{\Gamma}^{\natural, \operatorname{an}}$ as a map valued in $J_{\Gamma}^{\natural}(\mathbf{C})$ and not in $J_{\Gamma}^{\#}(\mathbf{C})$ is explained by the fact that the index of $H_1(Y_{\Gamma}, \mathbf{Z})$ in $\tilde{\mathcal{M}}_{\Gamma}$ is $\frac{1}{d_{\Gamma}}\cdot \prod_{c \in \mathcal{C}_{\Gamma}} e_c$. For convenience, we refer to section \[generalized\_cuspidal\] for the precise definitions of $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ and $\delta_{\Gamma}^{\natural, \operatorname{an}}$. While the map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ is easily seen to take values in $J_{\Gamma}^{\natural}(\mathbf{Q}(\zeta_N))$, it is unclear *a priori* whether the map $\delta_{\Gamma}^{\natural, \operatorname{an}}$ takes values in $J_{\Gamma}^{\natural}(\overline{\mathbf{Q}})$. This is in fact the case, as we show in the following comparison result between $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ and $\delta_{\Gamma}^{\natural, \operatorname{an}}$. \[intro\_main\_thm\_comparison\] 1. \[comparison\_i\] Let $n$ be the order of the cuspidal subgroup $C_{\Gamma}$ of $J_{\Gamma}$. Then we have $$n^2\cdot \delta_{\Gamma}^{\natural, \operatorname{alg}} = n^2\cdot \delta_{\Gamma}^{\natural, \operatorname{an}} \text{ .}$$ In particular, $\delta_{\Gamma}^{\natural, \operatorname{an}}$ takes values in $J_{\Gamma}^{\natural}(\mathbf{Q}(\zeta_N, \zeta_{n^2}))$. 2. \[comparison\_ii\] Assume that $N$ is odd, that $\begin{pmatrix} -1 & 0 \\ 0 & 1 \end{pmatrix}$ normalizes $\Gamma$, and that all the cusps in $\mathcal{C}_{\Gamma}$ are fixed by the complex conjugation, [*i.e.* ]{}for all $\frac{p}{q} \in \mathbf{P}^1(\mathbf{Q})$ we have $\Gamma \cdot (-\frac{p}{q}) = \Gamma \cdot \frac{p}{q}$. Then we have $\delta_{\Gamma}^{\natural, \operatorname{alg}} \equiv \delta_{\Gamma}^{\natural, \operatorname{an}}$ up to some element in the image of $\prod_{c \in \mathcal{C}_{\Gamma}} \{\pm 1\}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$. Additional results in the case $\Gamma = \Gamma_0(p)$ {#intro_Gamma_0(p)} ----------------------------------------------------- Assume in this paragraph that $\Gamma = \Gamma_0(p)$ for some prime $p$. The modular curve $X_{\Gamma_0(p)}$ has two cusps, namely $\Gamma_0(p) \infty$ and $\Gamma_0(p) 0$. Let $n$ be the order of the cuspidal subgroup of $J_{\Gamma_0(p)}$; Barry Mazur proved that $n = \frac{p-1}{d}$ where $d = \gcd(p-1,12)$. Let $j : X_{\Gamma_0(p)} \rightarrow \mathbf{P}^1$ be the usual $j$-invariant map. Let $\mathbb{T}$ be the Hecke algebra over $\mathbf{Z}$ acting faithfully on $M_2(\Gamma_0(p))$. ### The generalized cuspidal $1$-motive {#intro_refined_Gamma_0(p)} We define maps $$\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$$ and $$\delta_{\Gamma_0(p)}^{\#, \operatorname{an}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{C})$$ lifting our previous maps $\delta_{\Gamma_0(p)}^{\natural, \operatorname{alg}}$ and $\delta_{\Gamma_0(p)}^{\natural, \operatorname{an}}$. The reason we were able to construct $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$ is that there is a canonical uniformizer at the cusp $\Gamma_0(p) \infty$ (resp. $\Gamma_0(p) 0$), namely $j^{-1}$ (resp. $(j\circ w_p)^{-1}$ where $w_p$ is the Atkin–Lehner involution). The reason we were able to construct $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$ is that there is a canonical map $M_2(\Gamma_0(p)) \rightarrow \mathbf{C}$ given by $f \mapsto -L(f,1)$ where $L(f,s)$ is the complex $L$-function attached to $f$. This map, together with the (generalized) Abel–Jacobi isomorphism, gives a point in $\delta_{\Gamma_0(p)}^{\natural, \operatorname{an}}$ corresponding to $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}((\Gamma_0(p) \infty) - (\Gamma_0(p) 0))$. For convenience, we refer to §\[paragraph\_applications\_Gamma\_0(p)\] for the precise definitions of $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$ and $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$. We prove the following result, which is a refinement of Theorem \[intro\_main\_thm\_comparison\]. \[intro\_main\_comparison\_Gamma\_0\_case\] 1. \[main\_comparison\_Gamma\_0\_case\_i\] The maps $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$ and $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$ are $\mathbb{T}$-equivariant. 2. \[main\_comparison\_Gamma\_0\_case\_ii\] The element $n \cdot \delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}((\Gamma_0(p) \infty)-(\Gamma_0(p) 0))$ of $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$ is the image of $(1,p^{\frac{12}{d}}) \in \mathbf{Q}^{\times} \times \mathbf{Q}^{\times}$. 3. \[main\_comparison\_Gamma\_0\_case\_iii\] We have $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}} \equiv \delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$ modulo the image of $\mu_{\gcd(2,n)} \times \mu_{\gcd(2,n)}$ in $J_{\Gamma_0(p)}^{\#}(\mathbf{C})$. In particular, $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$ takes values in $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$. The map $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$ can be considered as a $1$-motive $\mathbf{Z} \rightarrow J_{\Gamma_0(p)}^{\#}$ over $\mathbf{Q}$, which we call the *generalized cuspidal $1$-motive*. Theorem \[intro\_main\_comparison\_Gamma\_0\_case\] (\[main\_comparison\_Gamma\_0\_case\_iii\]) describes the Betti realization of this $1$-motive (up to a sign ambiguity). ### Relation with the generalized $p$-adic period pairing of de Shalit {#intro_deShalit} There is a $p$-adic analogue of the constructions of §\[intro\_generalized\_cuspidal\_motive\], coming from the work of de Shalit [@deShalit] which we briefly recall below. Fix an algebraic closure $\overline{\mathbf{Q}}_p$ of $\mathbf{Q}_p$, and let $K \subset \overline{\mathbf{Q}}_p$ be the quadratic unramified extension of $\mathbf{Q}_p$. We also fix an algebraic closure $\mathbf{C}_p$ of the $p$-adic completion of $\overline{\mathbf{Q}}_p$. Let $S$ be the set of isomorphisms classes of supersingular elliptic curves over $\overline{\mathbf{F}}_p$. This is a finite set since the $j$-invariant of an element of $S$ is known to lie in $\mathbf{F}_{p^2}$. More precisely, we have $\operatorname{Card}(S)=g+1$ where $g$ is the genus of $X_0(p)$; we write $S = \{e_0, ..., e_g\}$. We denote by $\mathbf{Z}[S]$ the free $\mathbf{Z}$-module with basis the elements of $S$ (this is usually called the *supersingular module*) and by $\mathbf{Z}[S]^0$ its augmentation subgroup (the degree zero elements). There is a canonical bilinear pairing called the *$p$-adic period pairing* $$Q^0 : \mathbf{Z}[S]^0 \times \mathbf{Z}[S]^0 \rightarrow K^{\times}$$ inducing a map $q^0 : \mathbf{Z}[S]^0 \rightarrow \operatorname{Hom}(\mathbf{Z}[S]^0, K^{\times})$ via the formula $q^0(x)(y) = Q(x,y)$. The theory of $p$-adic uniformization gives a canonical $\operatorname{Gal}(\mathbf{C}_p/K)$-equivariant isomorphism $$J_{\Gamma_0(p)}(\mathbf{C}_p) \xrightarrow{\sim} \operatorname{Hom}(\mathbf{Z}[S]^0, \mathbf{C}_p^{\times})/q^0(\mathbf{Z}[S]^0) \text{ .}$$ de Shalit constructed in [@deShalit] a bilinear pairing $$Q : \mathbf{Z}[S] \times \mathbf{Z}[S] \rightarrow K^{\times}$$ extending $Q^0$ and with the property that there is a canonical $\operatorname{Gal}(\mathbf{C}_p/K)$-equivariant isomorphism $$\label{p-adic_uniformization_generalized} J_{\Gamma_0(p)}^{\#}(\mathbf{C}_p) \xrightarrow{\sim} \operatorname{Hom}(\mathbf{Z}[S], \mathbf{C}_p^{\times})/q(\mathbf{Z}[S]^0) \text{ ,}$$ where $q : \mathbf{Z}[S] \rightarrow \operatorname{Hom}(\mathbf{Z}[S], K^{\times})$ is such that $q(x)(y) = Q(x,y)$. We apologize to the reader for not recalling the precise construction of $Q$, as it is quite involved and is beautifully done in [@deShalit]. As de Shalit notes, the pairing $Q^0$ is canonical, but the choice of $Q$ depends on a choice of a tangent vector at the cusp $\Gamma_0(p)\infty$ [@deShalit §1.1]. This corresponds to the choice of the uniformizer $j^{-1}$ at $\Gamma_0(p) \infty$. de Shalit also proved that $Q$ is non-degenerate (this is even true after composing with the $p$-adic valuation $K^{\times} \rightarrow \mathbf{Z}$ since we essentially get the Kronecker pairing, [*cf.* ]{}[@deShalit §1.6 Main Theorem]). This means that $q$ is injective. There is a group homomorphism $$\delta^{\#, p\text{-adic}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{C}_p)$$ defined by $$\delta^{\#, p\text{-adic}}(\Gamma_0(p) 0 - \Gamma_0(p) \infty) = \text{image of }q(x) \text{ modulo }q(\mathbf{Z}[S]^0)$$ via (\[p-adic\_uniformization\_generalized\]) for any $x \in \mathbf{Z}[S]$ of degree $1$ (this does not depend on $x$). We prove the following comparison result in §\[section\_deShalit\]. \[intro\_main\_thm\_comparison\_p\_adic\] We have $\delta^{\#, p\text{-adic}} = \delta^{\#, \operatorname{alg}}$. ### The $\ell$-adic realization {#intro_Galois_representations} As an application of the above results of §\[intro\_Gamma\_0(p)\], we construct certain modular Galois representations in a “geometric” way. Let $\ell \geq 2$ be a prime and $\mathfrak{m}$ be a maximal ideal of $\mathbb{T}$ of residue characteristic $\ell$. The $\ell$-adic Tate module of the generalized cuspidal $1$-motive $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$ is denoted by $\mathcal{V}_{\ell}$, and its $\mathfrak{m}$-adic completion is denoted by $\mathcal{V}_{\mathfrak{m}}$([*cf.* ]{}§\[application\_Galois\_representations\] for the precise definition). We prove in Proposition \[multiplicity\_one\_V\] that $\mathcal{V}_{\mathfrak{m}}$ is a free $\mathbb{T}_{\mathfrak{m}}$-module of rank $2$, except possibly if $\ell=2$ and $\mathfrak{m}$ is supersingular (here, $\mathbb{T}_{\mathfrak{m}}$ is the $\mathfrak{m}$-adic completion of $\mathbb{T}$). If $\mathfrak{m}$ is non-Eisenstein (meaning it does not contain the Eisenstein ideal defined by Mazur), then $\mathcal{V}_{\mathfrak{m}}$ is canonically isomorphic to the usual Galois representation constructed from the $\mathfrak{m}$-adic Tate module of $J_{\Gamma_0(p)}$. We prove in §\[application\_Galois\_representations\] the following result if $\mathfrak{m}$ is Eisenstein. \[intro\_thm\_galois\_rep\] Assume that $\mathfrak{m}$ is Eisenstein (in particular, $\ell$ divides the numerator of $\frac{p-1}{12}$). We can choose a basis of $\mathcal{V}_{\mathfrak{m}}$ as a $\mathbb{T}_{\mathfrak{m}}$-module such that the following properties hold for the associated representation $\rho : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(\mathbb{T}_{\mathfrak{m}})$. 1. \[intro\_galois\_prop1\] The reduction of $\rho$ modulo $\mathfrak{m}$ is the residual representation $\overline{\rho} : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(\mathbf{F}_{\ell})$ given by $ \overline{\rho} = \begin{pmatrix} \overline{\chi}_{\ell} & \overline{b} \\ 0 & 1 \end{pmatrix}$, where $\chi_{\ell} : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \mathbf{Z}_{\ell}^{\times}$ is the $\ell$-adic cyclotomic character, $\overline{\chi}_{\ell}$ is the reduction of $\chi_{\ell}$ modulo $\ell$ and $\overline{b} : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \mathbf{F}_{\ell}$ is a Kummer cocycle in $Z^1( \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}), \overline{\chi}_{\ell})$ whose kernel cut out a number field isomorphic to $\mathbf{Q}(p^{\frac{1}{\ell}})$. 2. \[intro\_galois\_prop2\] The representation $\rho$ is unramified outside $p$ and $\ell$, has determinant $\chi_{\ell}$ and is finite flat at $\ell$. 3. \[intro\_galois\_prop3\] There is a free $\mathbb{T}_{\mathfrak{m}}$-submodule of rank one in $\mathcal{V}_{\mathfrak{m}}$ which is pointwise fixed by the inertia subgroup at $p$. 4. \[intro\_galois\_prop4\] For all primes $q \neq \ell, p$, the trace of $\rho(\operatorname{Frob}_q)$ is the Hecke operator $T_q$, where $\operatorname{Frob}_q$ is any (arithmetic) Frobenius element at $q$. 5. \[intro\_galois\_prop5\] If $\ell \geq 5$ then the representation $\rho$ is universal for the above properties, so we have an isomorphism $R \xrightarrow{\sim} \mathbb{T}_{\mathfrak{m}}$ where $R$ is the universal deformation rings with the prescribed above properties. Theorem \[intro\_thm\_galois\_rep\] is similar to a result of Frank Calegari and Matthew Emerton [@Calegari_Emerton Theorem 1.5], although the residual Galois representation they consider is $ \begin{pmatrix} \overline{\chi}_{\ell} & 0 \\ 0 & 1 \end{pmatrix}$. Property (\[intro\_galois\_prop5\]) is an immediate consequence of a result of Preston Wake and Carl Wang-Erickson [@WWE Corollary 7.1.3] (the restriction $\ell \geq 5$ comes from there, but we expect that the result still holds for $\ell \in \{2,3\}$). Our contribution here is really to the construction of the Galois representation $\rho$ satisfying the above properties. While we could maybe prove the existence of $\rho$ using a gluing argument using Ribet’s Lemma, the construction here is more geometric in nature since it is related to the generalized Jacobian of $X_0(p)$. \[intro\_rem\_galois\_rep\_duality\] There exists a perfect pairing of Galois modules $\mathcal{V}_{\ell} \times \mathcal{V}_{\ell} \rightarrow \mathbf{Z}_{\ell}(1)$, where as usual $\mathbf{Z}_{\ell}(1)$ is $\mathbf{Z}_{\ell}$ with the Galois action given by $\chi_{\ell}$ ([*cf.* ]{}Proposition \[galois\_rep\_duality\]). It seems therefore reasonable to expect that the $1$-motive $\mathbf{Z} \rightarrow J_{\Gamma_0(p)}^{\#}$ itself is self-dual. If $\Gamma$ is a congruence subgroup of level $N$ and $\ell$ is a prime not dividing $N$, then we can define similarly a canonical Hecke and Galois module by considering the $\ell$-adic realization of the $1$-motive $\delta_{\Gamma}^{\natural, \operatorname{alg}}$. We do not know whether this module is self-dual up to twist. An interesting case would be $\Gamma = \Gamma_1(p)$, since it would simplify some of our arguments in [@Lecouturier_MT]. Acknowledgements ---------------- I would like to thank my former Phd advisor Loïc Merel for his helpful suggestions and continuous support. I would also like to thank Takao Yamazaki for answering some questions about generalized Jacobians, and pointing out the usefulness of the Weil reciprocity law in §\[subsection\_comparison\]. This research was funded by Tsinghua University and the Yau Mathematical Sciences Center. Mixed modular symbols ===================== We keep the notation of §\[section\_introduction\]. First general properties {#section_first_properties} ------------------------ There is a group homomorphism $i_{\Gamma} : \Gamma \rightarrow \tilde{\mathcal{M}}_{\Gamma}$ given by $\gamma \mapsto \{1, \gamma\}$. We have a surjective group homomorphism $$\Pi_{\Gamma} : \Gamma \rightarrow H_1(Y_{\Gamma}, \mathbf{Z})$$ sending $\gamma \in \Gamma$ to the image in $Y_{\Gamma}$ of the geodesic path in the upper-half plane between $z_0$ and $\gamma(z_0)$ (for any $z_0$ in the upper-half plane). The kernel of $\Pi_{\Gamma}$ is generated by commutators and elliptic elements. Thus, the map $i_{\Gamma}$ factors through $\Pi_{\Gamma}$, so induces a group homomorphism $$\iota_{\Gamma} : H_1(Y_{\Gamma}, \mathbf{Z}) \rightarrow \tilde{\mathcal{M}}_{\Gamma} \text{ .}$$ Recall the following construction due to Stevens [@Stevens_book §2.3]. Let $M_2$ be the $\mathbf{C}$-vector space of weight $2$ modular forms with arbitrary level (in particular, this includes modular forms for non-congruence subgroups). Recall that there is a right action of $\operatorname{GL}_2^+(\mathbf{Q})$ on $M_2$, given by $(f \mid g)(z) = \det(g)\cdot (cz+d)^{-2}\cdot f(\frac{az+b}{cz+d})$ where $g = \begin{pmatrix} a & b \\ c & d \end{pmatrix}$ and $f \in M_2$. For any $f \in M_2$, we denote by $a_0(f)$ the Fourier coefficient at infinity of $f$. Fix $z_0$ in the upper-half plane. Following Stevens [@Stevens_book Definition 2.3.1 p. 51], we define a map $$\mathcal{S} : \operatorname{GL}_2^+(\mathbf{Q}) \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2, \mathbf{C})$$ by $$\mathcal{S}(g)(f) = 2i\pi \cdot \int_{z_0}^{g(z_0)} f(z)dz - 2i\pi\cdot z_0\cdot (a_0(f\mid g)-a_0(f)) + 2i\pi\cdot \int_{z_0}^{i\infty} \left((f\mid g)(z)-a_0(f\mid g)\right) - \left(f(z) - a_0(f)\right) dz \text{ .}$$ \[Stevens\_trivial\_lemma\] 1. \[Stevens\_trivial\_lemma\_1\] The map $\mathcal{S}$ does not depend on the choice of $z_0$. 2. \[Stevens\_trivial\_lemma\_2\] If $g \in \Gamma$ and $f$ is modular of level $\Gamma$, then $$\mathcal{S}(g)(f) = 2i\pi\int_{z_0}^{g(z_0)} f(z)dz \text{ .}$$ 3. The map $\mathcal{S}$ is a cocycle, [*i.e.* ]{}for all $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})^2$ and $f \in M_2$, we have $$\mathcal{S}(gg')(f) = \mathcal{S}(g)(f) + \mathcal{S}(g')(f\mid g) \text{ .}$$ 4. \[Stevens\_trivial\_lemma\_3\] For all $f \in M_2$, we have $\mathcal{S}\begin{pmatrix} a & b \\ 0 & d \end{pmatrix}(f) = \frac{b}{d}\cdot 2i\pi\cdot a_0(f)$. 5. \[Stevens\_trivial\_lemma\_4\] For all $f \in M_2$, we have $\mathcal{S}\begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}(f) = L(f,1)$. This follows from [@Stevens_book p. 51-52]. Using Proposition \[Stevens\_trivial\_lemma\], we check that there exists a unique group homomorphism $$\tilde{p}_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \text{ ,}$$ such that $\tilde{p}_{\Gamma}(\{g,g'\})(f)= \mathcal{S}(g^{-1}g')(f\mid g) = \mathcal{S}(g')(f)-\mathcal{S}(g)(f)$ for all $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$ and $f \in M_2(\Gamma)$. If $c \in \mathcal{C}_{\Gamma}$ is a cusp, let $e_c$ be the ramification index at the cusp $c$ of the map $X_{\Gamma}\rightarrow X_{\operatorname{SL}_2(\mathbf{Z})}$. Let $N$ be the l.c.m of the indices $e_c$; following [@general_level] we call $N$ the *general level of $\Gamma$*. If $\Gamma$ is a congruence subgroup of general level $N$, then $\Gamma$ has (usual) level $N$ or $2N$ [@general_level Proposition 3]. \[rank\_computation\] 1. \[rank\_computation\_1\] The map $\iota_{\Gamma} : H_1(Y_{\Gamma}, \mathbf{Z}) \rightarrow \tilde{\mathcal{M}}_{\Gamma}$ is injective. Thus, we can (and do) indentify $H_1(Y_{\Gamma}, \mathbf{Z})$ with a subgroup of $\tilde{\mathcal{M}}_{\Gamma}$. 2. \[rank\_computation\_2\] The element $\{g,gT\}$ of $\tilde{\mathcal{M}}_{\Gamma}$ only depends of the cusp $c=\Gamma g\infty$ of $X_{\Gamma}$, and $e_c \cdot \{g,gT\}$ is the image by $\iota_{\Gamma}$ of a small oriented circle around the cusp $c$. In particular, we have in $\tilde{\mathcal{M}}_{\Gamma}$: $$\sum_{c=\Gamma g\infty \in \mathcal{C}_{\Gamma}} e_c \cdot \{g, gT\}=0 \text{ .}$$ 3. \[rank\_computation\_3\] Consider the boundary map $\partial_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$ given by $\partial_{\Gamma}(\{g,g'\}) = [\Gamma g'\infty]-[\Gamma g\infty]$. The kernel of $\partial_{\Gamma}$ is spanned by $H_1(Y_{\Gamma}, \mathbf{Z})$ and the elements $\{g,gT\}$ for $g \in \operatorname{SL}_2(\mathbf{Z})$. In particular, the torsion subgroup of the coimage of $\iota_{\Gamma}$ has exponent $N$, and we have an exact sequence $$\label{exact_sequence_1} 0 \rightarrow H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{N}] \xrightarrow{\iota_{\Gamma} \otimes 1} \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{N}] \xrightarrow{\partial_{\Gamma} \otimes 1} \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{N}]\rightarrow 0$$ 4. \[rank\_computation\_4\] The kernel of $\pi_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \mathcal{M}_{\Gamma}$ is spanned by the elements $\{g,gT\}$ for $g \in \operatorname{SL}_2(\mathbf{Z})$. Thus, we have an exact sequence $$\label{exact_sequence_2} 0 \rightarrow \mathbf{Z} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]\rightarrow \tilde{\mathcal{M}}_{\Gamma} \xrightarrow{\pi_{\Gamma}} H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}) \rightarrow 0 \text{ .}$$ Here, the map $\mathbf{Z} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]$ is given by $1 \mapsto \frac{1}{d_{\Gamma}}\sum_{c \in \mathcal{C}_{\Gamma}} e_c\cdot [c]$, and the map $\mathbf{Z}[\mathcal{C}_{\Gamma}] \rightarrow \tilde{\mathcal{M}}_{\Gamma}$ is given by $[c] \mapsto \{g,gT\}$ where $g\in SL_2(\mathbf{Z})$ is such that $c=\Gamma g\infty$. Proof of (\[rank\_computation\_1\]). It suffices to show that the map $\tilde{p}_{\Gamma} \circ \iota_{\Gamma} : H_1(Y_{\Gamma}, \mathbf{Z}) \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2, \mathbf{C})$ is injective. This follows from Lemma \[Stevens\_trivial\_lemma\] (\[Stevens\_trivial\_lemma\_2\]) and the fact that the period map is injective. Proof of (\[rank\_computation\_2\]). Let $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$ such that $\Gamma g\infty = \Gamma g'\infty$. There exists $\gamma\in \Gamma$ and $n \in \mathbf{Z}$ such that $g' = \pm \gamma gT^n$. We have, in $\tilde{\mathcal{M}}_{\Gamma}$: $$\begin{aligned} \{g',g'T\}&=\{\gamma gT^n, \gamma g T^{n+1}\} \\&= \{gT^n, gT^{n+1}\} \\& = \{gT^n, g\}+\{g, gT^{n+1}\} \\&=\{g, gT^{n+1}\}-\{g,gT^n\} \\& = (n+1)\cdot \{g,gT\}-n\cdot \{g,gT\} \\&= \{g,gT\} \text{ .} \end{aligned}$$ This proves the first claim. For the second claim, note that the image by $\iota_{\Gamma}$ of a small oriented circle around $c$ is $i_{\Gamma}(gT^{e_c}g^{-1}) = \{1, gT^{e_c}g^{-1}\}$. Furthermore, we have in $\tilde{\mathcal{M}}_{\Gamma}$: $$\begin{aligned} e_c\cdot \{g,gT\} &= \{g,gT^{e_c}\} \\& = \{g,\gamma g\} \\& = \{g,1\}+\{1,\gamma\}+ \{\gamma, \gamma g\} \\& =\{g,1\}+\{1,\gamma\}+\{1,g\} \\&=\{1,\gamma\} \\&= \{1, gT^{e_c}g^{-1}\}\end{aligned}$$ where $\gamma=gT^{e_c}g^{-1} \in \Gamma$. This proves the second claim. The equality $$\sum_{c=\Gamma g\infty \in \mathcal{C}_{\Gamma}} e_c \cdot \{g, gT\}=0$$ is known to be true in $H_1(Y_{\Gamma}, \mathbf{Z})$, and hence in $\tilde{\mathcal{M}}_{\Gamma}$. Proof of (\[rank\_computation\_3\]). It is clear that $\operatorname{Ker}(\partial_{\Gamma})$ contains $H_1(Y_{\Gamma}, \mathbf{Z})$ and $\{g,gT\}$ for all $g \in \operatorname{SL}_2(\mathbf{Z})$. Note that $\tilde{\mathcal{M}}_{\Gamma}$ is spanned by elements of the form $\{1,g\}$ for $g \in \operatorname{SL}_2(\mathbf{Z})$. Let $\sum_{g \in \operatorname{SL}_2(\mathbf{Z})} \lambda_g \cdot \{1,g\}$ be an element of $\operatorname{Ker}(\partial_{\Gamma})$. For each cusp $c \in \mathcal{C}_{\Gamma}$, fix a $g_c \in \operatorname{SL}_2(\mathbf{Z})$ such that $c = \Gamma g_c \infty$. For each $c \neq \Gamma \infty$, we have $\sum_{g \in \operatorname{SL}_2(\mathbf{Z}), c=\Gamma g\infty} \lambda_g=0$. If $c = \Gamma g\infty$, then we have $g = \pm \gamma g_c T^n$ for some $\gamma \in \Gamma$ and $n \in \mathbf{Z}$. Thus $\{1,g\}=\{1,\gamma\}+\{\gamma, \gamma g_c T^n\}=\{1,\gamma\}+\{1,g_c\}+n\cdot \{g_c, g_c T\}$. Since $\{1,\gamma\} \in H_1(Y_{\Gamma}, \mathbf{Z})$ and $\sum_{g \in \operatorname{SL}_2(\mathbf{Z}), c=\Gamma g\infty} \lambda_g=0$, the element $\sum_{g \in \operatorname{SL}_2(\mathbf{Z}), c=\Gamma g\infty} \lambda_g \cdot \{1,g\}$ is in the span of $H_1(Y_{\Gamma}, \mathbf{Z})$ and the elements $\{g,gT\}$. Proof of (\[rank\_computation\_4\]). By (\[rank\_computation\_2\]) and (\[rank\_computation\_3\]), the kernel of $\pi_{\Gamma}$ is equal to the image of $\mathbf{Z}[\mathcal{C}_{\Gamma}]$ in $\tilde{\mathcal{M}}_{\Gamma}$. Furthermore, the exact sequence (\[exact\_sequence\_1\]) shows that the kernel of $\mathbf{Z}[\mathcal{C}_{\Gamma}]\rightarrow \tilde{\mathcal{M}}_{\Gamma}$ must be free of rank one over $\mathbf{Z}$. Since the kernel of $\mathbf{Z}[\mathcal{C}_{\Gamma}]\rightarrow \tilde{\mathcal{M}}_{\Gamma}$ contains $\sum_{c \in \mathcal{C}_{\Gamma}} e_c\cdot [c]$ by (\[rank\_computation\_2\]) and $\tilde{\mathcal{M}}_{\Gamma}$ is torsion-free, this kernel must be spanned by $\frac{1}{d_{\Gamma}}\sum_{c \in \mathcal{C}_{\Gamma}} e_c\cdot [c]$. Image of the period map $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ {#section_period_iso} --------------------------------------------------------------- By Proposition \[rank\_computation\], $\tilde{\mathcal{M}}_{\Gamma}$ is a free $\mathbf{Z}$-module of rank $\dim_{\mathbf{R}} M_2(\Gamma) = 2\cdot g(\Gamma) + 2 \cdot (c(\Gamma)-1)$. It thus makes sense to ask whether the map $\tilde{p}_{\Gamma}\otimes \mathbf{R} : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$ is an isomorphism (or equivalently a surjective map) of $\mathbf{R}$-vector spaces. Note that if $\Gamma'$ is a subgroup of $\Gamma$, then we have a commutative diagram of $\mathbf{R}$-vector spaces $$\begin{tikzcd} \tilde{\mathcal{M}}_{\Gamma'} \otimes_{\mathbf{Z}} \mathbf{R} \arrow{r}{\tilde{p}_{\Gamma'} \otimes \mathbf{R}} \arrow[swap]{d}{} & \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma'), \mathbf{C}) \arrow{d}{} \\ \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \arrow{r}{\tilde{p}_{\Gamma} \otimes \mathbf{R}} & \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \end{tikzcd}$$ where the two vertical maps are the canonical ones, and are surjective. In particular, if $\tilde{p}_{\Gamma'} \otimes \mathbf{R}$ is surjective, then $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is also surjective. Thus, if $\Gamma$ is a congruence subgroup containing the principal congruence subgroup $\Gamma(N)$, the map $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism if the map $\tilde{p}_{\Gamma(N)} \otimes \mathbf{R}$ is surjective. During the rest of this paragraph, we assume that $\Gamma$ is a congruence subgroup. We have a decomposition of $\mathbf{C}$-vector spaces $M_2(\Gamma) = S_2(\Gamma) \oplus \mathcal{E}_2(\Gamma)$, where $S_2(\Gamma)$ is the subspace of cuspidal modular forms and $\mathcal{E}_2(\Gamma)$ is the subspace of Eisenstein series. Let $\rho_{\operatorname{Cusp}, \Gamma} : \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \rightarrow \operatorname{Hom}_{\mathbf{C}}(S_2(\Gamma), \mathbf{C})$ and $\rho_{\operatorname{Eis}, \Gamma} : \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \rightarrow \operatorname{Hom}_{\mathbf{C}}(\mathcal{E}_2(\Gamma), \mathbf{C})$ be the maps induced by the inclusions $S_2(\Gamma) \hookrightarrow M_2(\Gamma)$ and $\mathcal{E}_2(\Gamma) \hookrightarrow M_2(\Gamma)$ respectively. \[restriction\_Eisenstein\_surjective\] The map $\rho_{\operatorname{Cusp}, \Gamma}\circ (\tilde{p}_{\Gamma}\otimes \mathbf{R})$ is surjective. Thus, $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism if and only if $\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma}\otimes \mathbf{R})$ is surjective. By Proposition \[rank\_computation\] (\[rank\_computation\_4\]), $\rho_{\operatorname{Cusp}, \Gamma}\circ (\tilde{p}_{\Gamma}\otimes \mathbf{R})$ factors through $\pi_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} : \tilde{M}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{R}$. The map $H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{C}}(S_2(\Gamma), \mathbf{C})$ is surjective, since its restriction to $H_1(X_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{R}$ is known to be an isomorphism. We thus need to understand the image of $\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma}\otimes \mathbf{R})$. We first recall some facts about Eisenstein series. We refer the reader to [@Stevens_TAMS §1] for details. The $\mathbf{C}$-vector space $\mathcal{E}_2$ is spanned by those $f \in M_2(\Gamma)$ such that $f(z)dz$ induces a meromorphic differential form on $X_{\Gamma}$ with integer residues at the cusps – such a differential is called a *differential of the third kind*. An example of differential of the third kind is the logarithmic derivative $d\log(u)$ of a *modular unit* $u$, [*i.e.* ]{}a meromorphic function on $X_{\Gamma}$ whose divisor is supported on $\mathcal{C}_{\Gamma}$. The set of $f \in \mathcal{E}_2(\Gamma)$ such that $f(z)dz = d\log(u)$ for some modular unit $u$ is denoted by $\mathcal{E}_2(\Gamma, \mathbf{Z})$; this is a $\mathbf{Z}$-module of rank $\dim_{\mathbf{C}} \mathcal{E}_2(\Gamma)$ by the Manin-Drinfeld theorem. If $M$ is a subgroup of $\mathbf{C}$, we let $\mathcal{E}_2(\Gamma, M)$ be the subgroup of $\mathcal{E}_2(\Gamma)$ generated by the elements $\lambda\cdot E$ where $\lambda \in M$ and $E \in \mathcal{E}_2(\Gamma, \mathbf{Z})$. We have $\mathcal{E}_2(\Gamma) = \mathcal{E}_2(\Gamma, \mathbf{C})$. If $\Gamma = \Gamma(N)$ for some $N \geq 1$, Stevens gave a spanning family for the $\mathbf{Q}$-vector space $\mathcal{E}_2(\Gamma, \mathbf{Q})$, together with the set of all possible relations. We refer to [@Stevens_book §2.4] for details. If $(x,y) \in (\mathbf{Q}/\mathbf{Z})^2 \backslash \{(0,0)\}$ has order dividing $N$, then Stevens defined an Eisenstein series $\phi_{(x,y)} \in \mathcal{E}_2(\Gamma(N))$, whose $q$-expansion at the cusp $\Gamma(N)\cdot \infty$ is $$\phi_{(x,y)}(z)=\frac{1}{2}\cdot\operatorname{\overline{B}}_2(x) - \sum_{ k \equiv x \text{(modulo }1\text{)} \atop k \in \mathbf{Q}_{>0} } k \cdot \sum_{m=1}^{\infty} q(m(kz+y)) - \sum_{ k \equiv -x \text{(modulo }1\text{)} \atop k \in \mathbf{Q}_{>0}} k \cdot \sum_{m=1}^{\infty} q(m(kz-y))$$ where $q(z) = e^{2i\pi z}$ and $\operatorname{\overline{B}}_2(x) = (x-E(x))^2-(x-E(x))+\frac{1}{6}$ is the second periodic Bernoulli polynomial function. Note that $\phi_{(-x,-y)} = \phi_{(x,y)}$. We have $2i\pi \phi_{(x,y)}(z)dz = d\log( g_{(x,y)})$ where $g_{(x,y)}$ is a Siegel unit defined by Kubert-Lang in [@Kubert_Lang Chapter 2 §1 Formula K$4$]. By [@Kubert_Lang Chapter 2, Theorem 1.2], the function $g_{(x,y)}^{12N}$ is modular of level $\Gamma(N)$. Thus we have $2i\pi \phi_{(x,y)} \in \frac{1}{12N} \cdot \mathcal{E}_2(\Gamma, \mathbf{Z})$, and in particular $2i\pi \phi_{(x,y)} \in \mathcal{E}_2(\Gamma, \mathbf{Q})$. By [@Stevens_book Chapter 2, §2.4], the Eisenstein series $2i\pi \phi_{(x,y)}$ span $\mathcal{E}_2(\Gamma, \mathbf{Q})$, and the linear relations between them are the so called *distribution relations* [@Stevens_book Chapter 2, Remark 2.4.4]. \[thm\_iso\_tilde\_p\] If $\Gamma(p^n) \subset \Gamma$ for some prime $p \geq 2$ and some integer $n \geq 1$, then $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism. Let $\mathcal{E} = \mathcal{E}_2(\Gamma, \mathbf{R}) \subset \mathcal{E}_2(\Gamma)$ and $\mathcal{E}' = \mathcal{E}_2(\Gamma, i\cdot \mathbf{R}) = i \cdot \mathcal{E}\subset \mathcal{E}_2(\Gamma)$. The $\mathbf{R}$-vector space $\mathcal{E}_2(\Gamma)$ is the direct sums its two $\mathbf{R}$-vector subspaces $\mathcal{E}$ and $\mathcal{E}'$. We get a canonical isomorphism of $\mathbf{R}$-vector spaces $$\operatorname{Hom}_{\mathbf{C}}(\mathcal{E}_2(\Gamma(p^n)), \mathbf{C}) \xrightarrow{\sim} \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}, \mathbf{R}) \times \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}', \mathbf{R})$$ given by taking restrictions and real parts. We denote by $p : \operatorname{Hom}_{\mathbf{C}}(\mathcal{E}_2(\Gamma(p^n)), \mathbf{C}) \rightarrow \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}, \mathbf{R}) $ and $p' : \operatorname{Hom}_{\mathbf{C}}(\mathcal{E}_2(\Gamma(p^n)), \mathbf{C}) \rightarrow \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}', \mathbf{R})$ the associated projections. By Lemma \[restriction\_Eisenstein\_surjective\], $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism if and only if $p\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R})) : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}, \mathbf{R}) $ and $p'\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R})): \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}', \mathbf{R}) $ are surjective $\mathbf{R}$-linear maps. To conclude the proof of Theorem \[thm\_iso\_tilde\_p\], it thus suffices to prove the following two lemmas. \[surjectivity\_part1\] The map $p'\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R}))$ is surjective (we need not assume anything on $\Gamma$, except that it is a congruence subgroup). We have a $\mathbf{R}$-linear isomorphism $\mathcal{E}' \xrightarrow{\sim} \operatorname{Div}^0(\mathcal{C}_{\Gamma} , \mathbf{R})$ given by $E \mapsto \sum_{c \in \mathcal{C}_{\Gamma}} 2i\pi \cdot \operatorname{Res}_c(E)\cdot (c)$, where $\operatorname{Res}_c(E)$ is the residue at $c$ of the differential form on $X_{\Gamma}$ induced by $E(z)dz$. Here, we have denoted by $\operatorname{Div}^0(\mathcal{C}_{\Gamma} , \mathbf{R})$ the group of degree zero divisors with coefficients in $\mathbf{R}$ supported on the set $\mathcal{C}_{\Gamma}$. We thus get a $\mathbf{R}$-linear isomorphism $f : \operatorname{Hom}_{\mathbf{R}}(\mathcal{E}', \mathbf{R}) \xrightarrow{\sim} \operatorname{Hom}_{\mathbf{R}}(\operatorname{Div}^0(\mathcal{C}_{\Gamma} , \mathbf{R}), \mathbf{R})$. We need to prove that the map $f\circ p'\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R})) : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{R} \rightarrow \operatorname{Hom}_{\mathbf{R}}(\operatorname{Div}^0(\mathcal{C}_{\Gamma} , \mathbf{R}), \mathbf{R})$ is surjective. The image of a little oriented circle around the cusp $c$ by the latter map is the restriction to $\operatorname{Div}^0(\mathcal{C}_{\Gamma})$ of the element of $\operatorname{Hom}_{\mathbf{R}}(\operatorname{Div}(\mathcal{C}_{\Gamma} , \mathbf{R}), \mathbf{R})$ sending $[c']$ to $0$ if $c'\neq c$ and $[c]$ to $1$. This concludes the proof of Lemma \[surjectivity\_part1\]. \[surjectivity\_part2\] The map $p\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R}))$ is surjective. For notational simplicity, denote by $\varphi$ the map $p\circ (\rho_{\operatorname{Eis}, \Gamma} \circ (\tilde{p}_{\Gamma} \otimes \mathbf{R}))$. By the discussion at the beginning of §\[section\_period\_iso\], we can assume without loss of generality that $\Gamma = \Gamma(p^n)$. The distribution relations show that a spanning family of the $\mathbf{R}$-vector space $\mathcal{E}$ is given by the Eisenstein series $2i\pi \phi_{(\frac{a}{p^n},\frac{b}{p^n})}$ where $(a,b) \in (\mathbf{Z}/p^n\mathbf{Z})^2/\pm 1$ is such that $\gcd(a,b,p)=1$. For simplicity, we write $2i\pi \phi_{(a,b)}$ for $2i\pi \phi_{(\frac{a}{p^n},\frac{b}{p^n})}$. The set of such $(a,b)$ is denoted by $S$. The only linear relation between these Eisenstein series is $$\sum_{(a,b) \in S} 2i\pi \phi_{(a,b)}=0 \text{ .}$$ There is a right action of $\operatorname{SL}_2(\mathbf{Z}/p^n\mathbf{Z})$ on $S$, given by $g \mapsto (a,b)\cdot g$. Note that $\operatorname{SL}_2(\mathbf{Z}/p^n\mathbf{Z}) \simeq \Gamma(p^n) \backslash \operatorname{SL}_2(\mathbf{Z})$ also acts on $\mathcal{E}$ via the slash operation, and we have [@Stevens_book Chapter 2, Remark 2.4.4]: $$(2i\pi \phi_{(a,b)})\mid g = 2i\pi \phi_{(a,b)\cdot g} \text{ .}$$ By [@Stevens_book Chapter 2, Proposition 2.5.4 (b)], for all $(g,g') \in \operatorname{SL}_2(\mathbf{Z})$ and $(a,b)\in S$, we have: $$\label{Stevens_formula_evalution_cusps} \varphi(\{g,g'\} \otimes 1)(2i\pi \phi_{(a,b)}) = F((a,b)\cdot g') - F((a,b)\cdot g)$$ where $F : S \rightarrow \mathbf{R}$ is given by $$F(a,b) = -\delta_{a}\cdot \log \mid 1-e^{\frac{2i\pi b}{p^n}} \mid \text{ .}$$ Here, $\delta_{a}=0$ if $a \neq 0$ and $\delta_{a}=1$ otherwise. Consider the matrix $M$ whose rows are indexed by $S$ and columns are indexed by $\mathcal{C}_{\Gamma}$, and such that the coefficient of $M$ at position $((a,b), \Gamma g\infty)$ is $F((a,b)\cdot g)$. Note that $M$ is a square matrix. We fix an ordering of $S$ and $\mathcal{C}_{\Gamma}$ as follows. Let $B$ be the subgroup of $\operatorname{SL}_2(\mathbf{Z}/p^n\mathbf{Z})$ consisting of upper-triangular matrices. Write $\operatorname{SL}_2(\mathbf{Z}/p^n\mathbf{Z})/B = \bigcup_{i=1}^k g_i B$ for some fixed elements $g_1$, ..., $g_k$. If $x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}$, let $\gamma_x = \begin{pmatrix} x & 0 \\ 0 & x^{-1} \end{pmatrix} \in B$. We then write $$\mathcal{C}_{\Gamma} = \bigsqcup_{i=1}^{k} \{g_i\gamma_x\infty, x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1 \} \text{ .}$$ We also write $$S = \bigsqcup_{i=1}^{k} \{(0,1)\gamma_x^{-1}g_i^{-1}, x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1 \} \text{ .}$$ This gives a decomposition of $M$ as a block matrix with $k^2$ blocks of size $\operatorname{Card}((\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm 1)$. We easily check that the non-diagonal blocks are zero, and the diagonal blocks are all equal to the matrix $M':=(-\log \mid 1 - e^{\frac{2i\pi x^{-1}y}{p^n}} \mid)_{(x,y)\in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1}$. To conclude the proof of Lemma \[surjectivity\_part2\], it suffices to prove that the matrices $M'$ and $$M'':=(-\log \mid 1 - e^{\frac{2i\pi x^{-1}y}{p^n}} \mid + \log \mid 1 - e^{\frac{2i\pi x^{-1}}{p^n}} \mid )_{(x,y)\in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1\atop x,y \neq \pm1}$$ are invertible. By [@Washington Lemma 5.26 (a), (b)], we have $$\det(M') = \prod_{\chi} \sum_{x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1} -\chi(x)\cdot \log \mid 1 - e^{\frac{2i\pi {x^{-1}}}{p^n}}\mid$$ $$(\text{resp. } \det(M'') = \prod_{\chi \neq 1} \sum_{x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1} -\chi(x)\cdot \log \mid 1 - e^{\frac{2i\pi {x^{-1}}}{p^n}}\mid \text{ )}$$ where $\chi$ goes through the (resp. non-trivial) even Dirichlet characters of level $p^n$. By the well-known formula for $L$ functions of primitive even Dirichlet characters [@Washington Theorem 4.9], we get: $$\det(M') = \frac{p}{2} \cdot \prod_{\chi \neq 1} \frac{f_{\chi}}{2\cdot \tau(\chi)} \cdot L(\chi,1)$$ and $$\det(M'') = \prod_{\chi \neq 1} \frac{f_{\chi}}{2\cdot \tau(\chi)}\cdot L(\chi,1)$$ where $f_{\chi}$ is the conductor of $\chi$ and $\tau(\chi)$ is the Gauss sum attached to the primitive Dirichlet character associated to $\chi$. These two quantities are known to be non-zero [@Washington Corollary 4.4]. \[rem\_period\_non\_iso\] In general, $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is not an isomorphism, as one can check numerically for instance if $\Gamma = \Gamma(6)$ using the method of the proof of Theorem \[thm\_iso\_tilde\_p\]. Consequently, for any integer $N \geq 1$, $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is not an isomorphism if $\Gamma = \Gamma(6N)$. Hecke operators and the complex conjugation ------------------------------------------- *In the rest of this paragraph, we assume that $\Gamma = \Gamma_1(N)$ or $\Gamma=\Gamma_0(N)$ for some integer $N \geq 1$.* Let $\mathbb{T}$ be the Hecke algebra acting faithfully on $M_2(\Gamma)$, generated by the Hecke operator $T_n$ for $n \geq 1$ and by the diamond operators. The abelian groups $\mathcal{M}_{\Gamma} \simeq H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z})$ and $H_1(Y_{\Gamma}, \mathbf{Z})$ both carry a faithfull action of $\mathbb{T}$ and of the complex conjugation ([*cf.* ]{}for instance [@Merel_Universal]). The goal of this paragraph is to define a natural action of $\mathbb{T}$ and of the complex conjugation on $\tilde{\mathcal{M}}_{\Gamma}$. ### The complex conjugation {#paragraph_complex_conjugation} If $M$ is an abelian group equipped with an action of an involution $m \mapsto \overline{m}$, we denote by $M(1)$ the abelian group $M$ equipped with the involution $m \mapsto -\overline{m}$. We also let $M^+ = \{m\in M, \overline{m}=m\}$ and $M^-=\{m\in M, \overline{m}=-m\}$. The action of the complex conjugation (denoted by a bar) on $\mathcal{M}_{\Gamma}$ and $H_1(Y_{\Gamma}, \mathbf{Z})$ is induced by the map $z \mapsto -\bar{z}$ in the upper-half plane. Thus, we have $\overline{\{\alpha, \beta\}} = \{-\alpha, -\beta\}$ in $\mathcal{M}_{\Gamma}$ for all $(\alpha, \beta) \in \mathbf{P}^1(\mathbf{Q})^2$. If $g=\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \operatorname{GL}_2^+(\mathbf{Q})$, we let $\overline{g} = \begin{pmatrix} a & -b \\ -c & d \end{pmatrix} \in \operatorname{GL}_2^+(\mathbf{Q})$. Note that we have $\overline{g\cdot g'} = \overline{g}\cdot \overline{g'}$ if $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})$. The complex conjugation acts on $\Gamma$ as via $g \mapsto \overline{g}$, and this induces an action on $H_1(Y_{\Gamma}, \mathbf{Z})$ via $\Pi_{\Gamma} : \Gamma \rightarrow H_1(Y_{\Gamma}, \mathbf{Z})$. There are canonical exact sequences of abelian groups, which are equivariant for the action of complex conjugation: $$\label{boundary_exact_seq_1} 0 \rightarrow H_1(X_{\Gamma}, \mathbf{Z}) \rightarrow \mathcal{M}_{\Gamma} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow 0$$ and $$\label{boundary_exact_seq_2} 0 \rightarrow \mathbf{Z}(1) \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}](1) \rightarrow H_1(Y_{\Gamma}, \mathbf{Z}) \rightarrow H_1(X_{\Gamma}, \mathbf{Z}) \rightarrow 0 \text{ .}$$ Here, the action of the complex conjugation on $\mathbf{Z}[\mathcal{C}_{\Gamma}]$ and $\mathbf{Z}[\mathcal{C}_{\Gamma}]^0$ is induced by the natural action on $\mathcal{C}_{\Gamma}$, and the action on $\mathbf{Z}$ is trivial. The action of the complex conjugation on $\tilde{\mathcal{M}}_{\Gamma}$ is defined by $\overline{\{g,g'\}}:=\{\overline{g}, \overline{g}'\}$. One easily checks that this is well-defined, and that the exact sequences (\[exact\_sequence\_1\]) and (\[exact\_sequence\_2\]) are equivariant with respect to the complex conjugation. There is a natural action of the complex conjugation on $\operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$, given by $\varphi \mapsto \overline{\varphi}$, where $\overline{\varphi} : f \mapsto \overline{\varphi(\overline{f})}$. Here, $\overline{f} \in M_2(\Gamma)$ is defined by $\overline{f}(z) = \overline{f(-\overline{z})}$. Equivalently, if the $q$-expansion of $f$ at $\Gamma\infty$ is $\sum_{n \geq 0} a_nq^n$, then the $q$-expansion of $\overline{f}$ at $\Gamma \infty$ is $\sum_{n\geq 0} \overline{a}_nq^n$. \[equivariance\_complex\_conjug\_period\] The map $\tilde{p}_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$ commutes with the action of the complex conjugation. Recall that $\tilde{p}_{\Gamma}(\{g,g'\})(f)=\mathcal{S}(g')(f) - \mathcal{S}(g)(f)$ where $$\mathcal{S}(g)(f) = 2i\pi \cdot \int_{z_0}^{g(z_0)} f(z)dz - 2i\pi\cdot z_0\cdot (a_0(f\mid g)-a_0(f)) + 2i\pi\cdot \int_{z_0}^{i\infty} \left((f\mid g)(z)-a_0(f\mid g)\right) - \left(f(z) - a_0(f)\right) dz$$ for any $z_0 \in \mathfrak{h}$. It thus suffices to prove that for any $f \in M_2(\Gamma)$ and $g \in \operatorname{SL}_2(\mathbf{Z})$, we have $\overline{\mathcal{S}(g)(\overline{f})} = \mathcal{S}(\tilde{g})(f)$. We shall make use the following straightforward result. \[trivial\_lemma\_cplx\_conjug\] For any $f \in M_2$, $g \in \operatorname{SL}_2(\mathbf{Z})$ and $z \in \mathfrak{h}$, we have $\overline{f \mid g} = \overline{f} \mid \tilde{g}$ and $-\overline{g(z)} = \tilde{g}(-\overline{z})$. We analyse separately the three terms in the definition of $\overline{\mathcal{S}(g)(\overline{f})}$. We have: $$\begin{aligned} \overline{2i\pi \cdot \int_{z_0}^{g(z_0)} \overline{f}(z)dz} &= -2i\pi \int_{z_0}^{g(z_0)} \overline{\overline{f(-\overline{z})}dz} \\&=2i\pi \int_{-\overline{z_0}}^{-\overline{g(z_0)}} f(z)dz \\&=2i\pi \int_{-\overline{z_0}}^{\tilde{g}(-\overline{z_0})} f(z)dz \text{ .}\end{aligned}$$ We have: $$\begin{aligned} \overline{2i\pi\cdot z_0\cdot (a_0(\overline{f}\mid g)-a_0(\overline{f}))} &= -2i\pi \cdot \overline{z_0}\cdot (a_0(\overline{\overline{f}\mid g})-a_0(f)) \\& = 2i\pi \cdot (-\overline{z}_0)\cdot (a_0(f\mid \tilde{g})-a_0(f)) \text{ .}\end{aligned}$$ Finally, we have: $$\begin{aligned} \overline{2i\pi\cdot \int_{z_0}^{i\infty} \left((\overline{f}\mid g)(z)-a_0(\overline{f}\mid g)\right) - \left(\overline{f}(z) - a_0(\overline{f})\right) dz } &= -2i\pi \cdot \int_{z_0}^{i\infty} \left((f\mid \tilde{g})(-\overline{z})-a_0(f\mid \tilde{g})\right) - \left(f(-\overline{z}) - a_0(f)\right) d\overline{z} \\&= 2i\pi \cdot \int_{-\overline{z_0}}^{i\infty} \left((f\mid \tilde{g})(z)-a_0(f\mid \tilde{g})\right) - \left(f(z) - a_0(f)\right) dz \text{ .}\end{aligned}$$ We have thus proved that $\overline{\mathcal{S}(g)(\overline{f})} = \mathcal{S}(\tilde{g})(f)$. This concludes the proof of Proposition \[equivariance\_complex\_conjug\_period\]. ### Hecke operators {#paragraph_Hecke} We are going to define a natural action of $\mathbb{T}$ on $\tilde{\mathcal{M}}_{\Gamma}$. We first briefly recall how $\mathbb{T}$ acts on $\mathcal{M}_{\Gamma}$ and $H_1(Y_{\Gamma}, \mathbf{Z})$. Let $g \in \operatorname{GL}_2^+(\mathbf{Q})$ and consider the double-coset $\Gamma g\Gamma = \bigsqcup_{i \in I} \Gamma g_i$ for some finite set $I$ and $g_i \in \operatorname{GL}_2^+(\mathbf{Q})$. The Hecke operator $T_g$ acts $X_{\Gamma}$ via the correspondance $(\Gamma z) \mapsto \sum_{i \in I}(\Gamma g_iz)$, where $z \in \mathfrak{H} \cup \mathbf{P}^1(\mathbf{Q})$ and $\mathfrak{H}$ is the upper-half plane. This does not depend on the choice of the elements $g_i$. If $g = \begin{pmatrix} 1&0\\ 0& p \end{pmatrix}$ for some prime $p$, the Hecke operator $T_{g}$ is denoted by $T_p$ if $p$ does not divide $N$ and $U_p$ otherwise. If $g=\begin{pmatrix} a & b \\ c& d\end{pmatrix} \in \Gamma_0(N)$, the Hecke operator $T_g$ is denoted by $\langle d \rangle$ (a so-called diamond operator). This induces an action of $T_g$ on $H_1(Y_{\Gamma}, \mathbf{Z})$ and $\mathcal{M}_{\Gamma}$, which is explicitly given as follows. For any $\{\alpha, \beta\} \in \mathcal{M}_{\Gamma}$, we have $T_g\{\alpha,\beta\}=\sum_{i \in I} \{g_i\alpha, g_i\beta\}$. For $H_1(Y_{\Gamma}, \mathbf{Z})$, the action of $T_g$ is similarly given in terms of geodesic paths in $\mathfrak{H}$, but it will be more convenient to describe this action using the map $\Pi_{\Gamma}$. For all $i \in I$ and $\gamma \in \Gamma$, we have $$g_i\gamma = t_{i, \Gamma}(\gamma)g_{\sigma_{\gamma}(i)}$$ for unique $t_{i, \Gamma}(\gamma) \in \Gamma$ and $\sigma_{\gamma}(i) \in I$. Note that $t_{i, \Gamma} : \Gamma \rightarrow \Gamma$ and $\sigma_{\gamma} : I \rightarrow I$ are bijective maps. The following result is well-known. \[Hecke\_group\_theoretic\] For all $\gamma \in \Gamma$, we have $$T_g \Pi_{\Gamma}(\gamma) = \sum_{i \in I} \Pi_{\Gamma}(t_{i, \Gamma}(\gamma)) \text{ .}$$ Fix $z_0 \in \mathfrak{H}$. Let $\alpha : \mathfrak{H} \rightarrow Y_{\Gamma}$ be the quotient map. Let $S = \alpha(\operatorname{GL}_2^+(\mathbf{Q})\cdot z_0)$. We have an inclusion $H_1(Y_{\Gamma}, \mathbf{Z}) \subset H_1(Y_{\Gamma}, S, \mathbf{Z})$ where $H_1(Y_{\Gamma}, S, \mathbf{Z})$ is the relative homology group with respect to the pair $(Y_{\Gamma}, S)$. We will do our computations inside $H_1(Y_{\Gamma}, Z, \mathbf{Z})$. If $z_1,z_2 \in \operatorname{GL}_2^+(\mathbf{Q})\cdot z_0$, let $\{z_1,z_2\}$ be the image in $H_1(Y_{\Gamma}, S, \mathbf{Z})$ of the geodesic path between $z_1$ and $z_2$ in $\mathfrak{H}$. For all $\gamma \in \Gamma$ we have in $H_1(Y_{\Gamma}, S, \mathbf{Z})$: $$\begin{aligned} T_g \Pi_{\Gamma}(\gamma) &= \sum_{i\in I} \{g_iz_0, g_i\gamma z_0\} \\&= \sum_{i\in I} \{g_iz_0, t_{i, \Gamma}(\gamma)g_{\sigma_{\gamma}(i)}z_0\} \\& = \sum_{i\in I} \{g_iz_0, g_{\sigma_{\gamma}(i)}z_0\} + \{g_{\sigma_{\gamma}(i)}z_0, t_{i, \Gamma}(\gamma)g_{\sigma_{\gamma}(i)}z_0\} \\&=\sum_{i\in I} \Pi_{\Gamma}(t_{i, \Gamma}(\gamma))+ \sum_{i\in I} \{g_iz_0, g_{\sigma_{\gamma}(i)}z_0\} \text{,}\end{aligned}$$ where in the last equality we have used that $$\{g_{\sigma_{\gamma}(i)}z_0, t_{i, \Gamma}(\gamma)g_{\sigma_{\gamma}(i)}z_0\} = \{z_0, t_{i, \Gamma}(\gamma)z_0\}=\Pi_{\Gamma}(t_{i, \Gamma}(\gamma)) \text{ .}$$ To conclude the proof of Proposition \[Hecke\_group\_theoretic\], we have to prove that $\sum_{i\in I} \{g_iz_0, g_{\sigma_{\gamma}(i)}z_0\}=0$. This follows formally from the fact that $\sigma_{\gamma}$ is a permutation of $I$ and the Chasles relations on symbols $\{., .\}$. One difficulty to define an action of $\mathbb{T}$ on $\tilde{\mathcal{M}}_{\Gamma}$ comes from the fact that while $\operatorname{GL}_2^+(\mathbf{Q})$ acts on $\mathfrak{H}\cup \mathbf{P}^1(\mathbf{Q})$, it does not act on $\operatorname{SL}_2(\mathbf{Z})$ so it does not make sense to talk about the symbols $\{g_ig,g_ig'\}$ in $\tilde{\mathcal{M}}_{\Gamma}$ for $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$. We are going to solve this issue by introducing a $\mathbf{Q}$-vector space $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ containing $\tilde{\mathcal{M}}_{\Gamma}$ as a lattice and for which there is a natural action of double cosets Hecke operators. Consider the $\mathbf{Q}$-vector space $\tilde{\mathcal{M}}_{\mathbf{Q}}$ generated by the symbols $\{g,g'\}_{\mathbf{Q}}$ where $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})^2$ with the following relations: 1. $\{g,g'\}_{\mathbf{Q}}+\{g',g''\}_{\mathbf{Q}}+\{g'',g\}_{\mathbf{Q}}=0$ for all $(g,g',g'') \in \operatorname{GL}_2^+(\mathbf{Q})^3$; 2. $\{g,g'\}_{\mathbf{Q}}-\{\lambda_1g, \lambda_2 g'\}_{\mathbf{Q}}=0$ for all $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})^2$ and $(\lambda_1,\lambda_2)\in (\mathbf{Q}^{\times})^2$; 3. $\{g,g\begin{pmatrix}a&b \\ 0 & d\end{pmatrix}\}_{\mathbf{Q}}-\frac{b}{d}\cdot \{g,gT\}_{\mathbf{Q}}=0$ for all $g\in \operatorname{GL}_2^+(\mathbf{Q})$ and $\begin{pmatrix}a&b \\ 0 & d\end{pmatrix} \in \operatorname{GL}_2^+(\mathbf{Q})$. There is a left action of $\operatorname{GL}_2^+(\mathbf{Q})$ on $\tilde{\mathcal{M}}_{\mathbf{Q}}$, given by $g \cdot \{g',g''\} = \{gg',gg''\}$ for all $(g,g',g'') \in \operatorname{GL}_2^+(\mathbf{Q})^3$. We let $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ be the largest quotient of $\tilde{\mathcal{M}}_{\mathbf{Q}}$ on which $\Gamma$ acts trivially. There is a canonical map $$\psi : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$$ given by $\{g,g'\} \mapsto \{g,g'\}_{\mathbf{Q}}$ for $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$. By Proposition \[Stevens\_trivial\_lemma\], there is a unique well-defined $\mathbf{Q}$-linear map $$\tilde{p}_{\mathbf{Q}, \Gamma} : \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$$ such that $\tilde{p}_{\mathbf{Q}, \Gamma}(\{g,g'\}_{\mathbf{Q}}) = \mathcal{S}(g')(f) - \mathcal{S}(g)(f)$ for all $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})^2$. We have $\tilde{p}_{\Gamma} = \tilde{p}_{\mathbf{Q}, \Gamma} \circ \psi$. \[lattice\_Q\] The map $\psi$ induces an isomorphism of $\mathbf{Q}$-vector spaces $\psi \otimes \mathbf{Q} : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Q} \xrightarrow{\sim} \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$. Thus, we can consider $\tilde{\mathcal{M}}_{\Gamma}$ as a lattice inside $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$. We first show that $\psi \otimes \mathbf{Q}$ is surjective. Let $(g,g') \in (\operatorname{GL}_2(\mathbf{Q})^+)^2$. Write $g = \alpha \begin{pmatrix} a & b \\ 0 & d \end{pmatrix}$ and $g' = \alpha' \begin{pmatrix} a' & b' \\ 0 & d' \end{pmatrix}$ for some $(\alpha, \alpha') \in \operatorname{SL}_2(\mathbf{Z})^2$ and $\begin{pmatrix} a & b \\ 0 & d \end{pmatrix}, \begin{pmatrix} a' & b' \\ 0 & d' \end{pmatrix} \in \operatorname{GL}_2^+(\mathbf{Q})$. We then have, in $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$: $$\begin{aligned} \{g,g'\}_{\mathbf{Q}} &= \{g, \alpha\}_{\mathbf{Q}}+\{\alpha, \alpha'\}_{\mathbf{Q}}+\{\alpha', g'\}_{\mathbf{Q}} \\& = \frac{b}{d}\cdot \{\alpha T, \alpha\}_{\mathbf{Q}}+ \{\alpha, \alpha'\}_{\mathbf{Q}} + \frac{b'}{d'}\cdot \{\alpha', \alpha'T\} _{\mathbf{Q}} \\& = \frac{b}{d}\cdot (\psi \otimes \mathbf{Q})(\{\alpha T, \alpha\}) + (\psi \otimes \mathbf{Q})(\{\alpha, \alpha'\}) + \frac{b'}{d'}\cdot (\psi \otimes \mathbf{Q})(\{\alpha', \alpha' T\}) \text{ .}\end{aligned}$$ This proves that $\psi \otimes \mathbf{Q}$ is surjective. To prove that $\psi \otimes \mathbf{Q}$ is an isomorphism, it suffices to prove that $\dim_{\mathbf{Q}} \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma} \geq \dim_{\mathbf{Q}}(\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Q}) = 2 g(\Gamma) +2\cdot (c(\Gamma)-1)$. There is a surjective $\mathbf{Q}$-linear map $$\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma} \rightarrow \mathcal{M}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Q}\text{ ,}$$ given by $\{g,g'\}_{\mathbf{Q}} \mapsto \{g\infty, g'\infty\}$. Its kernel contains the elements $\{g,gT\}_{\mathbf{Q}}$ for all $g \in \operatorname{SL}_2(\mathbf{Z})$, so it suffices to show that the $\mathbf{Q}$-vector space $C$ spanned by these elements has dimension $\geq c(\Gamma)-1$. The map $(\psi \otimes \mathbf{Q}) \circ (\iota_{\Gamma} \otimes \mathbf{Q}) : H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q} \rightarrow \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ is injective, since $$\tilde{p}_{\mathbf{Q}, \Gamma} \circ (\psi \otimes \mathbf{Q}) \circ (\iota_{\Gamma} \otimes \mathbf{Q}) : H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$$ is injective. Thus, we can identify $H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q}$ with a $\mathbf{Q}$-vector subspace of $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$. We know that $C$ is the subspace of $H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q}$ of dimension $c(\Gamma)-1$ spanned by the little circles around the cusps. This concludes the proof of Proposition \[lattice\_Q\]. Let $g \in \operatorname{GL}_2^+(\mathbf{Q})$, and write as before $\Gamma g \Gamma = \bigsqcup_{i=1}^n \Gamma g_i$. There is a well-defined double coset operator $$T_{g, \mathbf{Q}} : \tilde{\mathcal{M}}_{\Gamma, \mathbf{Q}} \rightarrow \tilde{\mathcal{M}}_{\Gamma, \mathbf{Q}}$$ given by $\{h,h'\}_{\mathbf{Q}} \mapsto \sum_{i=1}^n \{g_ih, g_ih'\}_{\mathbf{Q}}$. If $p$ is a prime and $g = \begin{pmatrix} 1 & 0 \\ 0 & p \end{pmatrix}$, we denote $T_{g, \mathbf{Q}}$ by $T_{p, \mathbf{Q}}$ if $p \nmid N$ and by $U_{p, \mathbf{Q}}$ otherwise. If $d$ is an integer coprime to $N$ and $g \in \Gamma_0(N)$ is such that its lower-right coefficient is congruent to $d$ modulo $N$, we denote $T_{g, \mathbf{Q}}$ by $\langle d \rangle_{\mathbf{Q}}$. Finally, if $g = \begin{pmatrix} 0 & -1 \\ N & 0 \end{pmatrix}$, then we denote $T_{g, \mathbf{Q}}$ by $W_{N, \mathbf{Q}}$. \[Hecke\_Q\] There is a unique action of $\mathbb{T}$ on $\tilde{\mathcal{M}}_{\Gamma, \mathbf{Q}}$ such that the following hold. 1. \[Hecke\_Q\_compatible\] If $p$ is a prime dividing (resp. not dividing) $N$, then $U_p$ (resp. $T_p$) acts as $U_{p,\mathbf{Q}}$ (resp. $T_{p,\mathbf{Q}}$). If $d$ is an integer coprime to $N$, then $\langle d \rangle$ acts as $\langle d \rangle_{\mathbf{Q}}$. 2. \[Hecke\_Q\_periods\] The map $\tilde{p}_{\mathbf{Q}, \Gamma} : \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}\rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$ is $\mathbb{T}$ and $W_{N, \mathbf{Q}}$-equivariant. 3. \[Hecke\_Q\_exact\_seq\] The exact sequences (\[exact\_sequence\_1\]) and (\[exact\_sequence\_2\]) tensorized by $\mathbf{Q}$ are $\mathbb{T}$ and $W_{N, \mathbf{Q}}$-equivariant after identifying $\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Q}$ with $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ via $\psi \otimes \mathbf{Q}$. 4. \[Hecke\_Q\_lattice\_odd\] For all integer $d$ coprime to $N$, the diamond operator $\langle d \rangle$ stabilizes the lattice $\tilde{\mathcal{M}}_{\Gamma}$. For all prime $p$ not dividing $2N$, $T_p$ stabilizes $\tilde{\mathcal{M}}_{\Gamma}$. More precisely, write $$\operatorname{SL}_2(\mathbf{Z}) \begin{pmatrix} 1 & 0 \\ 0 & p \end{pmatrix} \operatorname{SL}_2(\mathbf{Z}) = \operatorname{SL}_2(\mathbf{Z})g_{\infty} \bigcup_{i=-{\frac{p-1}{2}}}^{\frac{p-1}{2}} \operatorname{SL}_2(\mathbf{Z})g_i$$ where $g_{\infty} = \begin{pmatrix} p & 0 \\ 0 & 1 \end{pmatrix}$ and $g_i = \begin{pmatrix} 1 & i \\ 0 & p \end{pmatrix}$. Then we have in $\tilde{\mathcal{M}}_{\Gamma}$, for all $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$: $$T_p \{g,g'\} = \langle p \rangle \{t_{\infty, SL_2(\mathbf{Z})}(g), t_{\infty, SL_2(\mathbf{Z})}(g')\}+\sum_{i=-\frac{p-1}{2}}^{\frac{p-1}{2}} \{t_{i, SL_2(\mathbf{Z})}(g), t_{i, SL_2(\mathbf{Z})}(g')\} \text{ .}$$ 5. \[Hecke\_Q\_T\_2\] If $p$ divides $2N$, the operator $T_p$ (or $U_p$ if $p \mid N$) sends $\tilde{\mathcal{M}}_{\Gamma}$ into $\tilde{\mathcal{M}}_{\Gamma}+\frac{1}{p}\cdot \sum_{c=\Gamma g\infty \in \mathcal{C}_{\Gamma}} \{g,gT\} \subset \frac{1}{p}\cdot \tilde{\mathcal{M}}_{\Gamma}$. 6. \[Atkin\_Lehner\_M\] The operator $W_{N, \mathbf{Q}}$ induces an endomorphism of $\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{N}]$, which we denote by $W_N$ and call the Atkin–Lehner involution. 7. \[hecke\_conjugation\_M\] Assume that all the cusps of $X_{\Gamma}$ are fixed by the complex conjugation. Then the exact sequences (\[exact\_sequence\_1\]) and (\[exact\_sequence\_2\]) induce a canonical $\mathbb{T}$ and $W_N$-equivariant isomorphism $$\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}] \xrightarrow{\sim} H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^+ \bigoplus H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^- \text{ .}$$ If $\Gamma = \Gamma_1(N)$, the cusps of $X_{\Gamma}$ are fixed by the complex conjugation if and only if $N$ divides $2p$ for some prime number $p \geq 2$. If $\Gamma = \Gamma_0(N)$, the cusps of $X_{\Gamma}$ are fixed by the complex conjugation if and only if $N$ is squarefree or four times a squarefree integer. \[Hecke\_equivariance\_+\_-\] 1. \[Hecke\_equivariance\_pi\]The map $\pi_{\Gamma} \otimes \mathbf{Q} : \tilde{\mathcal{M}}_{\mathbf{Q},\Gamma} \rightarrow H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Q})$ is Hecke equivariant. This means the following. If $p$ is a prime dividing (resp. not dividing) $N$, then $(\pi_{\Gamma} \otimes \mathbf{Q}) \circ U_{p, \mathbf{Q}} = U_p \circ (\pi_{\Gamma} \otimes \mathbf{Q})$ (resp. $(\pi_{\Gamma} \otimes \mathbf{Q}) \circ T_{p, \mathbf{Q}} = T_p \circ (\pi_{\Gamma} \otimes \mathbf{Q})$). Similarly for the diamond operators and the Atkin-Lehner involution. 2. \[Hecke\_equivariance\_iota\] The map $\iota_{\Gamma} \otimes \mathbf{Q} : H_1(Y_{\Gamma}, \mathbf{Q}) \rightarrow \tilde{\mathcal{M}}_{\mathbf{Q},\Gamma} $ is Hecke equivariant. 3. \[Hecke\_equivariance\_periods\] The map $\tilde{p}_{\mathbf{Q}, \Gamma} :\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$ is Hecke equivariant. 4. \[Hecke\_equivariance\_produit\_injective\] The map $$\tilde{p}_{\mathbf{Q}, \Gamma} \times (\tilde{\pi}_{\Gamma} \otimes \mathbf{Q}) : \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \times (\mathcal{M}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Q})$$ is injective and Hecke equivariant. Point (\[Hecke\_equivariance\_pi\]) follows by definition of Hecke operators. We now prove (\[Hecke\_equivariance\_iota\]). We first check the compatibility of diamond operators. Let $d$ be an integer coprime to $N$. Let $g$ be a matrix whose lower right corner is congruent to $d$ modulo $N$. For all $\gamma \in \Gamma$, we have in $ \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$: $$\begin{aligned} \iota_{\Gamma}(\langle d \rangle\Pi_{\Gamma}(\gamma)) &= \iota_{\Gamma}(\Pi_{\Gamma}(g\gamma g^{-1})) \\& = \{1, g \gamma g^{-1}\}_{\mathbf{Q}} \\& = \{g,g\gamma\}_{\mathbf{Q}} \text{ .}\end{aligned}$$ The first equality follows from Proposition \[Hecke\_group\_theoretic\]. The second equality is justified as follows: $$\begin{aligned} \{g, g\gamma\} &= \{g, g\gamma g^{-1}g\} \\&= \{g,1\}+\{1,g\gamma g^{-1}\}+\{g\gamma g^{-1}, g\gamma g^{-1}g\} \\& = \{g,1\}+\{1,g\gamma g^{-1}\}+\{1,g\} \\&= \{1,g\gamma g^{-1}\} \text{ ,}\end{aligned}$$ where we have used the fact that $g\gamma g^{-1}\in \Gamma$. This proves the compatibility of $\iota_{\Gamma} \otimes \mathbf{Q}$ to diamond operators. We now consider the Hecke operator $T_p$ (or $U_p$) for a prime $p$. Write $\Gamma \begin{pmatrix} 1 & 0 \\ 0 & p \end{pmatrix} \Gamma = \bigcup_{i \in I} \Gamma g_i \Gamma$. By Proposition \[Hecke\_group\_theoretic\], for all $\gamma \in \Gamma$, we have in $ \tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$: $$\begin{aligned} (\iota_{\Gamma} \otimes \mathbf{Q})(T_p \Pi_{\Gamma}(\gamma)) &= \sum_{i \in I} \Pi_{\Gamma}(t_{i, \Gamma}(\gamma)) \\&= \sum_{i \in I} \{1, t_{i, \Gamma}(\gamma)\}_{\mathbf{Q}} \text{ .}\end{aligned}$$ On the other hand, we have: $$\begin{aligned} T_{p, \mathbf{Q}}(\iota_{\Gamma} \otimes \mathbf{Q})(\gamma)&=\sum_{i \in I} \{g_i, g_i\gamma\}_{\mathbf{Q}} \\& = \sum_{i \in I} \{g_i, t_{i,\Gamma}(\gamma)g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}} \\& = \sum_{i\in I} \{g_i, g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}}+\{g_{\sigma_{\gamma}(i)}, t_{i,\Gamma}(\gamma)g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}} \\& = \sum_{i\in I} \{g_i, g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}} + \sum_{i\in I} \{1, t_{i,\Gamma}(\gamma)\}_{\mathbf{Q}}\end{aligned}$$ where in the last equality we have used the fact that $t_{i,\Gamma}(\gamma) \in \Gamma$, so $\{g_{\sigma_{\gamma}(i)}, t_{i,\Gamma}(\gamma)g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}} = \{1, t_{i,\Gamma}(\gamma)\}_{\mathbf{Q}}$. Since $\sigma_{\gamma} : I \rightarrow I$ is a bijection, we have $\sum_{i\in I} \{g_i, g_{\sigma_{\gamma}(i)}\}_{\mathbf{Q}}=0$. This proves (\[Hecke\_equivariance\_iota\]). Lemma \[Hecke\_equivariance\_+\_-\] (\[Hecke\_equivariance\_periods\]) is an immediate consequence of the definition of Hecke operators as double coset operators, and of the formula $\tilde{p}_{\mathbf{Q}, \Gamma}(\{g,g'\}_{\mathbf{Q}})(f) = \mathcal{S}(g^{-1}g')(f\mid g)$ for all $(g,g') \in \operatorname{GL}_2^+(\mathbf{Q})^2$ and $f \in M_2(\Gamma)$. We finally prove Lemma \[Hecke\_equivariance\_+\_-\] (\[Hecke\_equivariance\_produit\_injective\]). The map $\tilde{p}_{\mathbf{Q}, \Gamma} \times (\tilde{\pi}_{\Gamma} \otimes \mathbf{Q})$ is Hecke equivariant by (\[Hecke\_equivariance\_pi\]) and (\[Hecke\_equivariance\_periods\]), so we only need to prove that it is injective. This follows from the facts that the map $\tilde{p}_{\Gamma} \otimes \mathbf{Q} : H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma),\mathbf{C})$ is injective and that the kernel of $\tilde{\pi}_{\Gamma} \otimes \mathbf{Q}$ is contained in $H_1(Y_{\Gamma}, \mathbf{Z}) \otimes_{\mathbf{Z}} \mathbf{Q}$ (considered as a subspace of $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ via $\iota_{\Gamma} \otimes \mathbf{Q}$). This concludes the proof of Lemma \[Hecke\_equivariance\_+\_-\]. By Lemma \[Hecke\_equivariance\_+\_-\] (\[Hecke\_equivariance\_produit\_injective\]), there is an action of $\mathbb{T}$ on $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$ satisfying Theorem \[Hecke\_Q\] (\[Hecke\_Q\_compatible\]). This action is obviously unique. Theorem \[Hecke\_Q\] (\[Hecke\_Q\_periods\]) and (\[Hecke\_Q\_exact\_seq\]) also follow from Lemma \[Hecke\_equivariance\_+\_-\]. We now prove Theorem \[Hecke\_Q\] (\[Hecke\_Q\_lattice\_odd\]) and (\[Hecke\_Q\_T\_2\]). The assertion about diamond operators is straightforward. Let $p$ be a prime. If $p \mid N$, then by convention the diamond operator $\langle p \rangle$ is zero. We have: $$\operatorname{SL}_2(\mathbf{Z})\begin{pmatrix} 1 & 0 \\ 0 & p \end{pmatrix} \operatorname{SL}_2(\mathbf{Z}) = \operatorname{SL}_2(\mathbf{Z})g_{\infty} \bigcup_{i \in I} \operatorname{SL}_2(\mathbf{Z}) g_i \text{ ,}$$ where $I = \{-\frac{p-1}{2}, ..., \frac{p-1}{2}\}$, $g_{\infty} = \begin{pmatrix} p & 0 \\ 0 & 1 \end{pmatrix}$ and $g_i = \begin{pmatrix} 1 & i \\ 0 & p \end{pmatrix}$. Recall that for all $g \in \operatorname{SL}_2(\mathbf{Z})$, we have $g_ig = t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}$ for some $t_{i, \operatorname{SL}_2(\mathbf{Z})} : \operatorname{SL}_2(\mathbf{Z}) \rightarrow \operatorname{SL}_2(\mathbf{Z})$ and $\sigma_{g} : I \cup \{\infty\}\xrightarrow{\sim} I \cup \{\infty\}$. For all $(g,g') \in \operatorname{SL}_2(\mathbf{Z})^2$, we have in $\tilde{\mathcal{M}}_{\mathbf{Q}, \Gamma}$: $$\begin{aligned} T_p\{g,g'\}_{\mathbf{Q}} &= \langle p \rangle \{g_{\infty}g, g_{\infty}g'\}_{\mathbf{Q}}+\sum_{i \in I} \{g_ig, g_ig'\}_{\mathbf{Q}} \\& = \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_{g}(\infty)}, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} + \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g), t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g')\}_{\mathbf{Q}} \\& + \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g'), t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g')g_{\sigma_{g'}(\infty)}\}_{\mathbf{Q}} \\&+ \sum_{i \in I}\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}}+\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g), t_{i, \operatorname{SL}_2(\mathbf{Z})}(g')\}_{\mathbf{Q}}\\&+\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g'), t_{i, \operatorname{SL}_2(\mathbf{Z})}(g')g_{\sigma_{g'}(i)}\}_{\mathbf{Q}} \text{ .}\end{aligned}$$ For all $i \in I \cup \{\infty\}$, we have $\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} \in \frac{1}{p}\mathbf{Z}\cdot \{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g), t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)T\}_{\mathbf{Q}}$ and $\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g'), t_{i, \operatorname{SL}_2(\mathbf{Z})}(g')g_{\sigma_{g'}(i)}\}_{\mathbf{Q}} \in \frac{1}{p}\mathbf{Z}\cdot \{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g'), t_{i, \operatorname{SL}_2(\mathbf{Z})}(g')T\}_{\mathbf{Q}}$. This proves Theorem \[Hecke\_Q\] (\[Hecke\_Q\_T\_2\]). Assume now that $p \nmid 2N$. To conclude the proof of Theorem \[Hecke\_Q\] (\[Hecke\_Q\_lattice\_odd\]), it suffices to show that for all $g \in \operatorname{SL}_2(\mathbf{Z})$, we have $$\label{lemma_integrality} \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_{g}(\infty)}, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} + \sum_{i \in I}\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} =0 \text{ .}$$ Write $g = \begin{pmatrix} a & b \\ c & d \end{pmatrix}$. We consider two cases. First, assume that $p \mid c$. Then $\sigma_g(\infty)=\infty$ and $\sigma_g$ induces a permutation of $I$. We have: $$\begin{aligned} \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_{g}(\infty)}, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} &+ \sum_{i \in I}\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} \\& = \frac{1}{p}\cdot \sum_{i \in I} \sigma_g(i) \cdot \{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}}\end{aligned}$$ Since $p\neq 2$, we have $ \sum_{i \in I} \sigma_g(i) = \sum_{i \in I} i = 0$, so to prove (\[lemma\_integrality\]) it suffices to prove that $\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}}$ is independant of $i$, [*i.e.* ]{}that the cusp $\Gamma t_{i, \operatorname{SL}_2(\mathbf{Z})}(g) \infty$ is independant of $i$. We have $t_{i, \operatorname{SL}_2(\mathbf{Z})}(g) \infty = g_i g \infty = \frac{a+ic}{pc}$. Since $p \mid c$ and $\gcd(a,c)=1$, the fraction $\frac{a+ic}{pc}$ is irreducible. Since $\Gamma = \Gamma_1(N)$ or $\Gamma=\Gamma_0(N)$ and $p \nmid N$, we see that $\Gamma \frac{a+ic}{pc} = \Gamma \frac{a}{pc}$ is independant of $i$. Assume now that $p \nmid c$. Let $j \in I$ such that $p \mid a+jc$. Then $\sigma_g(j) = \infty$ and $\sigma_g(\infty)=j'$ where $j' \in I$ is such that $j' \equiv dc^{-1}\text{ (modulo }p\text{)}$. We have: $$\begin{aligned} \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_{g}(\infty)}, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} &+ \sum_{i \in I}\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)g_{\sigma_g(i)}, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} \\& = \frac{\sigma_g(\infty)}{p}\cdot \langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} \\&+\sum_{i\in I \atop i \neq j} \frac{\sigma_g(i)}{p}\cdot \{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}} \text{ .}\end{aligned}$$ Since $p \neq 2$, we have $$\sigma_{g}(\infty) + \sum_{i \in I \atop i \neq j} \sigma_g(i) = \sum_{i \in I} i = 0 \text{ .}$$ Thus, to prove (\[lemma\_integrality\]) it suffices to prove that the elements $\langle p \rangle \{t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}}$ and $\{t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)T, t_{i, \operatorname{SL}_2(\mathbf{Z})}(g)\}_{\mathbf{Q}}$ for all $i \in I \backslash \{j\}$ coincide. Equivalently, it suffices to prove that the cusps $\langle p \rangle \Gamma t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g) \infty$ and $\Gamma t_{i, \operatorname{SL}_2(\mathbf{Z})}(g) \infty$ are the same, for all $i \in I \backslash \{j\}$. We have $\langle p \rangle \Gamma t_{\infty, \operatorname{SL}_2(\mathbf{Z})}(g) \infty = \langle p \rangle \Gamma \frac{pa}{c}$ and $\Gamma t_{i, \operatorname{SL}_2(\mathbf{Z})}(g) \infty = \Gamma \frac{a+ic}{pc}$. The facts that $p \nmid a+ic$ if $i \neq j$, $\Gamma = \Gamma_1(N)$ or $\Gamma=\Gamma_0(N)$ and $p \nmid N$ imply that these cusps are all the same. This concludes the proof of (\[lemma\_integrality\]). Theorem \[Hecke\_Q\] (\[Atkin\_Lehner\_M\]) follows from the fact that for any $g \in \operatorname{SL}_2(\mathbf{Z})$, we can write $\begin{pmatrix} 0 & -1 \\ N & 0 \end{pmatrix} g = g' \cdot \begin{pmatrix} a & b \\ 0 & d \end{pmatrix}$ for some $g' \in \operatorname{SL}_2(\mathbf{Z})$ and $a,b,d \in \mathbf{Z}$ with $ad=N$. Theorem \[Hecke\_Q\] (\[hecke\_conjugation\_M\]) follows from (\[Hecke\_Q\_exact\_seq\]). This concludes the proof of Theorem \[Hecke\_Q\] . Manin symbols {#paragraph_Manin} ------------- Following Manin [@Manin_parabolic], consider the map $$\xi_{\Gamma} : \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})] \rightarrow \mathcal{M}_{\Gamma}$$ defined by $\Gamma g \mapsto \{g0, g\infty\}$. Let $S = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ and $U = \begin{pmatrix} 1 & -1 \\ 1 & 0 \end{pmatrix}$. We have $S^2=U^3=\begin{pmatrix} -1 & 0 \\ 0 & -1 \end{pmatrix}$. \[Manin\_thm\] 1. \[Manin\_surjectivity\] The map $\xi_{\Gamma}$ is surjective. 2. \[Manin\_kernel\] We have $$\operatorname{Ker}(\xi_{\Gamma}) = \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^U + \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^S$$ where for any $g \in \operatorname{SL}_2(\mathbf{Z})$, $\mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^g$ denotes the subgroup of elements fixed by the right multiplication by $g$. Recall that in §\[intro\_Manin\_symbols\] we defined a map $\tilde{\xi}_{\Gamma} : \mathbf{Z}[\Gamma\backslash \operatorname{PSL}_2(\mathbf{Z})]\rightarrow \tilde{\mathcal{M}}_{\Gamma}$ by $\tilde{\xi}_{\Gamma}(\Gamma g) = \{g,gS\}$ for all $g\in \operatorname{PSL}_2(\mathbf{Z})$. We have $\xi_{\Gamma} = \pi_{\Gamma} \circ \tilde{\xi}_{\Gamma}$. Let $\mathbf{Z}[\mathcal{C}_{\Gamma}]_0$ be the quotient of $\mathbf{Z}[\mathcal{C}_{\Gamma}]$ by the element $\frac{1}{d_{\Gamma}}\sum_{c \in \mathcal{C}_{\Gamma}} e_c\cdot [c]$ where $e_c$ is the width of $c$. If $c \in \mathcal{C}_{\Gamma}$, we denote by $(c)$ the image of $[c]$ in $\mathbf{Z}[\mathcal{C}_{\Gamma}]_0$. We have a canonical injective group homomorphism $\mathbf{Z}[\mathcal{C}_{\Gamma}]_0 \hookrightarrow \tilde{\mathcal{M}}_{\Gamma}$ coming from the exact sequence (\[exact\_sequence\_2\]). Theorem \[intro\_thm\_Manin\] is a consequence of the following result. \[generalized\_Manin\_thm\] Let $\varphi_{\Gamma} : \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})] \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]_0$ be the group homomorphism given by $\varphi_{\Gamma}([\Gamma g]) = (\Gamma g \infty)$. 1. \[surjectivity\_generalized\_Manin\_general\] The coimage of $\tilde{\xi}_{\Gamma}$ is canonically isomorphic to the coimage of the restriction of $\varphi_{\Gamma}$ to $ \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^U$. 2. \[surjectivity\_generalized\_Manin\_special\] Let $p\geq 5$ be a prime and $n \in \mathbf{N}$. If $\Gamma=\Gamma_1(p^n)$ or $\Gamma = \Gamma_0(p^n)$, then the image of $\tilde{\xi}_{\Gamma}$ has index dividing $3$, the divisibility being strict of and only if $p \equiv 1 \text{ (modulo 3}\text{)}$ and $\Gamma = \Gamma_0(p^n)$. 3. \[kernel\_generalized\_Manin\] The kernel of $\tilde{\xi}_{\Gamma}$ is spanned by the following two subgroups of $\mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]$: - The subgroup $\mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^S$. - The kernel of the restriction of $\varphi_{\Gamma}$ to $ \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^U$. We prove (\[surjectivity\_generalized\_Manin\_general\]). By Theorem \[Manin\_thm\], $\pi_{\Gamma}\circ \tilde{\xi}_{\Gamma}$ is surjective and the coimage of $\tilde{\xi}_{\Gamma}$ in $\tilde{\mathcal{M}}_{\Gamma}$ equals the coimage in $\operatorname{Ker}(\pi_{\Gamma})$ of the restriction of $\tilde{\xi}_{\Gamma}$ to $\mathbf{Z}[\Gamma \backslash \operatorname{SL}_2(\mathbf{Z})]^U + \mathbf{Z}[\Gamma \backslash \operatorname{SL}_2(\mathbf{Z})]^S$. Note that the restriction of $\tilde{\xi}_{\Gamma}$ to $\mathbf{Z}[\Gamma \backslash \operatorname{SL}_2(\mathbf{Z})]^S$ is zero. \[lemma\_Manin\_relations\_cusps\] For all $g \in \operatorname{SL}_2(\mathbf{Z})$, we have in $\tilde{\mathcal{M}}_{\Gamma}$: $$\tilde{\xi}_{\Gamma}([\Gamma g]+ [\Gamma gU]+[\Gamma gU^2]) = \{g,gT\}+\{gU,gUT\}+\{gU^2,gU^2T\} \text{ .}$$ Using the following equalities: $US=-T$, $U^2S=-UT$ and $U^2T=S$, we have: $$\begin{aligned} \tilde{\xi}_{\Gamma}([\Gamma g]+ [\Gamma gU]+[\Gamma gU^2]) &= \{g,gS\}+\{gU,gUS\}+\{gU^2, gU^2S\} \\& = \{g,gS\}+\{gU,gT\}+\{gU^2, gUT\} \\& = \{g,gT\}+\{gT,gS\}+\{gU,gUT\}+\{gUT,gT\}+\{gU^2, gU^2T\}+\{gU^2T, gUT\} \\& =\{g,gT\}+\{gU,gUT\}+\{gU^2,gU^2T\}+\{gT,gS\}+\{gUT,gT\}+\{gS,gUT\} \\& = \{g,gT\}+\{gU,gUT\}+\{gU^2,gU^2T\} \text{ .}\end{aligned}$$ By Lemma \[lemma\_Manin\_relations\_cusps\], the restrictions of $\tilde{\xi}_{\Gamma}$ and $\varphi_{\Gamma}$ to $\mathbf{Z}[\Gamma \backslash \operatorname{SL}_2(\mathbf{Z})]^U$ are equal. This proves (\[surjectivity\_generalized\_Manin\_general\]). We now prove (\[surjectivity\_generalized\_Manin\_special\]). Let $p \geq 5$ be a prime and $n \in \mathbf{N}$. \[lemma\_g0\_g1\_goo\] Assume $\Gamma = \Gamma_1(p^n)$ or $\Gamma = \Gamma_0(p^n)$. Consider the map $f_{\Gamma} : \mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})] \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]$ sending $[\Gamma g]$ to $[\Gamma g 0]+[\Gamma g 1]+[\Gamma g \infty]$. The image of $f_{\Gamma}$ consists of the elements of degree divisible by $3$ in $\mathbf{Z}[\mathcal{C}_{\Gamma}]$. Since $f_{\Gamma}$ is functorial in $\Gamma$, it suffices to prove Lemma \[lemma\_g0\_g1\_goo\] when $\Gamma = \Gamma_1(p^n)$. Recall that if $(a,b,a',b') \in \mathbf{Z}^4$ and $\gcd(a,b)=\gcd(a',b')=1$, then we have $\Gamma \cdot \frac{a}{b} = \Gamma \cdot \frac{a'}{b'}$ if and only if $\begin{pmatrix} a' \\ b'\end{pmatrix} \equiv \pm \begin{pmatrix} a+kb \\ b\end{pmatrix} \text{ (modulo }p^n\text{)}$ for some $k \in \mathbf{Z}$. The class of the vector $\begin{pmatrix} a \\ b\end{pmatrix}$ modulo the previous equivalence relation is denoted by $\begin{Bmatrix} a \\ b \end{Bmatrix}$, and such a symbol is uniquely identified with an element of $\mathcal{C}_{\Gamma}$. If $r\geq 0$ is an integer, we denote by $\mathcal{C}_{\Gamma}^{(r)} \subset \mathcal{C}_{\Gamma}$ the set of $\begin{Bmatrix} a \\ b \end{Bmatrix}$ with $\gcd(b,p^n)=p^r$. We have $\mathcal{C}_{\Gamma} = \bigsqcup_{r=0}^n \mathcal{C}_{\Gamma}^{(r)}$. We first show that the elements of $\mathbf{Z}[\mathcal{C}_{\Gamma}^{(0)}]$ of degree divisible by $3$ are contained in $\text{Im}(f_{\Gamma})$. First note that there is a bijection $(\mathbf{Z}/p^n\mathbf{Z})^{\times}/\pm1 \xrightarrow{\sim} \mathcal{C}_{\Gamma}^{(0)}$ given by $x \mapsto \begin{Bmatrix} 0 \\ x \end{Bmatrix}$. For any $x,y \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}$ such that $x+y$ is prime to $p$, we have $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}]+[\begin{Bmatrix} 0 \\ y \end{Bmatrix}]+[\begin{Bmatrix} 0 \\ x+y \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. Note in particular that $2\cdot [\begin{Bmatrix} 0 \\ x \end{Bmatrix}]+[\begin{Bmatrix} 0 \\ 2x \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. If $x-y$ is prime to $p$, we also have $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}]+[\begin{Bmatrix} 0 \\ y \end{Bmatrix}]+[\begin{Bmatrix} 0 \\ x-y \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. Thus, for any $x,y \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}$ with $x \pm y$ prime to $p$, we have $[\begin{Bmatrix} 0 \\ x+y \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ x-y \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. Since $p$ is odd, for any $x,y \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}$ with $x \pm y$ prime to $p$ we have $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ y \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. Since $p>3$, by letting $x=2$ and $y=1$ we get $[\begin{Bmatrix} 0 \\ 2 \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ 1 \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. We also have $2\cdot [\begin{Bmatrix} 0 \\ 1\end{Bmatrix}]+ [\begin{Bmatrix} 0 \\ 2 \end{Bmatrix}] = f_{\Gamma}(\Gamma \begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix}) \in \text{Im}(f_{\Gamma})$. We thus get $3\cdot [\begin{Bmatrix} 0 \\ 1 \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. It remains to show that for all $x \in (\mathbf{Z}/p^n\mathbf{Z})^{\times}$, we have $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ 1 \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. We already know this if $x \not\equiv \pm 1 \text{ (modulo }p\text{)}$. If $x \equiv \pm 1 \text{ (modulo }p\text{)}$ then since $p>3$ we have $\frac{x}{2} \equiv \pm \frac{1}{2} \not\equiv \pm 1 \text{ (modulo }p\text{)}$, so we get $[\begin{Bmatrix} 0 \\ \frac{x}{2} \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ 1 \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. Since we know that $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}] + 2\cdot [\begin{Bmatrix} 0 \\ \frac{x}{2}\end{Bmatrix}] \in \text{Im}(f_{\Gamma})$, we get $[\begin{Bmatrix} 0 \\ x \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ 1 \end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. We now conclude the proof of Lemma \[lemma\_g0\_g1\_goo\]. Let $r$ be an integer such that $1 \leq r \leq n$. Let $\begin{Bmatrix} v \\ up^r \end{Bmatrix} \in \mathcal{C}_{\Gamma}^{(r)}$, where $(v,u) \in \mathbf{Z}^2$ is such that $\gcd(v,up^r)=1$ and $u \not\equiv 0 \text{ (modulo }p\text{)}$. Let $g = \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \operatorname{SL}_2(\mathbf{Z})$ be such that $\gcd(c,p)=\gcd(d,p)=1$, $c+d \equiv up^r\text{ (modulo }p^n\text{)}$ and $d \equiv v^{-1}\text{ (modulo }p^r\text{)}$. Then we have $f_{\Gamma}([\Gamma g]) = [\begin{Bmatrix} 0 \\ c\end{Bmatrix}]+[\begin{Bmatrix} 0 \\ d\end{Bmatrix}]+[\begin{Bmatrix} v \\ up^r\end{Bmatrix}]$. Thus, we have $[\begin{Bmatrix} v \\ up^r \end{Bmatrix}] - [\begin{Bmatrix} 0 \\ 1\end{Bmatrix}] \in \text{Im}(f_{\Gamma})$. By Theorem \[generalized\_Manin\_thm\] (\[surjectivity\_generalized\_Manin\_general\]) and Lemma \[lemma\_g0\_g1\_goo\], the index of the image of $\tilde{\xi}_{\Gamma}$ in $\mathbf{Z}[\mathcal{C}_{\Gamma}]_0$ is divisible by $3$. Moreover, this index is one if and only if either there exists an element of $\mathbf{Z}[\Gamma \backslash \operatorname{PSL}_2(\mathbf{Z})]^U$ of degree $1$ or $\sum_{c \in \mathcal{C}_{\Gamma}} e_c$ is relatively prime to $3$. If $\Gamma = \Gamma_1(p^n)$, we check that these conditions never hold. If $\Gamma = \Gamma_0(p^n)$, we check that these two conditions hold if $p\equiv 1 \text{ (modulo }3\text{)}$ and none of them hold otherwise. This proves Theorem \[generalized\_Manin\_thm\] (\[surjectivity\_generalized\_Manin\_special\]) Theorem \[generalized\_Manin\_thm\] (\[Manin\_kernel\]) follows from Theorem \[Manin\_thm\] (\[Manin\_kernel\]) and Lemma \[lemma\_Manin\_relations\_cusps\]. This concludes the proof of Theorem \[generalized\_Manin\_thm\]. \[Manin\_rmk\] It would be interesting to compute the index of $\tilde{\xi}_{\Gamma}$ if $\Gamma = \Gamma_0(N)$ or $\Gamma = \Gamma_1(N)$ for any $N\geq 1$. If $N$ is even, then this index is infinite (it suffices to check that for $N=2$). If $N$ is odd, we do not know if this index finite (the numerical computations seem to indicate that this index divides $3$). The proof would rely on a generalization of Lemma \[lemma\_g0\_g1\_goo\], which seems to be hard if $N$ is highly composite. Self-duality and Zagier’s Petersson inner product {#paragraph_Petersson} ------------------------------------------------- In this paragraph, we do not assume anything about $\Gamma$, which could in particular be a non-congruence subgroup. We denote by $N$ the general level of $\Gamma$. By taking duals, the map $\tilde{p}_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$ induces a map $\tilde{p}_{\Gamma}^* : M_2(\Gamma) \rightarrow \operatorname{Hom}_{\mathbf{Z}}(\tilde{M}_{\Gamma} , \mathbf{C})$. If $f \in M_2(\Gamma)$ and $g \in \operatorname{SL}_2(\mathbf{Z})$, recall that by construction we have $\tilde{p}_{\Gamma}^*(f)(\{g,gS\}) = L(f\mid g, 1)$ and $\tilde{p}_{\Gamma}^*(f)(\{g,gT\}) = 2i\pi \cdot a_0(f \mid g)$, where $S = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ and $T = \begin{pmatrix}1 & 1 \\ 0 & 1 \end{pmatrix}$. Following Zagier [@Zagier_Rankin] we can extend the usual Petersson Hermitian pairing on cuspidal forms to the whole space of modular forms $M_2(\Gamma)$. We now recall the definition. For any real number $T>1$, define the truncated fundamental domain $\mathcal{F}_T = \{z \in \mathfrak{h} \text{ such that } \abs{z} \geq 1, \abs{ \Re(z) }\leq \frac{1}{2} \text{ and } \Im(z)<T\}$, where $\mathfrak{h}$ is the upper-half plane. Then according to Zagier, if $f_1,f_2 \in M_2(\Gamma)$, we define the Petersson pairing of $f_1$ and $f_2$ to be $$(f_1, f_2) := \frac{1}{[\operatorname{SL}_2(\mathbf{Z}) : \Gamma]} \cdot \lim_{T \rightarrow +\infty} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \int_{\mathcal{F}_T} (f_1\mid g)(z) \cdot \overline{(f_2\mid g)(z)}dxdy - T\cdot a_0(f_1\mid g)\cdot \overline{a_0(f_2 \mid g)} \text{ .}$$ We get a Hermitian bilinear pairing on $M_2(\Gamma)$. Although the restriction of the Petersson pairing to cuspidal modular forms is non-degenerate, it can be degenerate on $M_2(\Gamma)$. Nevertheless, Pasol and Popa proved the following result [@Popa_Petersson Theorems 4.4 and 4.5], which seems related to Remark \[rem\_period\_non\_iso\]. \[non\_degeneracy\_petersson\] 1. Assume $\Gamma = \Gamma_0(N)$ or $\Gamma = \Gamma_1(N)$. The Petersson pairing $M_2(\Gamma) \times M_2(\Gamma) \rightarrow \mathbf{C}$ is non degenerate if $N$ is a prime power. 2. The Petersson pairing is degenerate if $\Gamma=\Gamma_1(N)$ and $N$ is divisible by $pq$ with $p\neq q$ primes such that $q$ is not a primitive residue modulo $p$ ([*e.g.* ]{}if $N$ is divisible by $6$). It is also degenerate if $\Gamma=\Gamma_1(N)$ and $N$ is divisible by $p^2q$ with $p\neq q$ primes. 3. The Petersson pairing is degenerate if $\Gamma=\Gamma_0(N)$ and $N$ is square-free and not prime. It is natural to ask whether the Petersson pairing can be expressed in terms of a pairing on $\operatorname{Hom}_{\mathbf{Z}}(\tilde{M}_{\Gamma}, \mathbf{C})$ *via* the map $\tilde{p}_{\Gamma}^*$. This was done by Merel when restricting to cuspidal modular forms [@Merel_Manin_L Théorème 2]. In the general case, this was done by Pasol and Popa. More precisely, the following result is a reformulation of [@Popa_Haberland Theorem 8.6 a)]. \[Haberland\] For any $f_1, f_2 \in M_2(\Gamma)$, we have: $$\begin{aligned} 48 i \pi^2\cdot (f_1, f_2) = \frac{1}{[\operatorname{SL}_2(\mathbf{Z}) : \Gamma]} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} &\tilde{p}_{\Gamma}^*(f_1)(\{gTS,gT\})\cdot \overline{\tilde{p}_{\Gamma}^*(f_2)(\{gS,g\})} \\& - \tilde{p}_{\Gamma}^*(f_1)(\{gS,g\}) \cdot \overline{\tilde{p}_{\Gamma}^*(f_2)(\{gTS,gT\})} \\& + 4\cdot \tilde{p}_{\Gamma}^*(f_1)(\{g,gT\})\cdot \overline{ \tilde{p}_{\Gamma}^*(f_2)(\{g,gS\})} \\& - 4\cdot \tilde{p}_{\Gamma}^*(f_1)(\{g,gS\})\cdot \overline{ \tilde{p}_{\Gamma}^*(f_2)(\{g,gT\})} \text{ .}\end{aligned}$$ Recall that in §\[intro\_duality\], we have defined an anti-symmetric pairing $\langle \cdot, \cdot \rangle$ on the $\mathbf{Z}$-dual $\tilde{\mathcal{M}}_{\Gamma}^*$ of $\tilde{\mathcal{M}}_{\Gamma}$. We easily check that $\langle \cdot, \cdot \rangle$ is anti-invariant for the complex conjugation $c$ acting on $\tilde{\mathcal{M}}_{\Gamma}$, namely for all $\varphi_1, \varphi_2 \in \tilde{\mathcal{M}}_{\Gamma}$, we have $$\label{complex_conjug_compatibility_pairing} \langle \varphi_1 \circ c , \varphi_2 \rangle = -\langle \varphi_1 , \varphi_2 \circ c \rangle \text{ .}$$ By tensoring with $\mathbf{C}$, $\langle \cdot , \cdot \rangle$ extends to an anti-symmetric bilinear pairing on $\operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{C})$. If $\varphi \in \operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{C})$, we define $\overline{\varphi} \in \operatorname{Hom}_{\mathbf{Z}}(\tilde{\mathcal{M}}_{\Gamma}, \mathbf{C})$ by the formula $\overline{\varphi}(m) = \overline{\varphi(m)}$ for all $m \in \tilde{\mathcal{M}}_{\Gamma}$. We can restate Theorem \[Haberland\] as $$\label{reformulation_Haberland} -8 i \pi^2 (f_1, f_2) = \frac{1}{[\operatorname{SL}_2(\mathbf{Z}):\Gamma]}\cdot \langle \tilde{p}_{\Gamma}^*(f_1) , \overline{\tilde{p}_{\Gamma}^*(f_2)} \rangle \text{ .}$$ We expect that two properties hold: 1. \[property\_degenerate\_pairing\] The pairing $\langle \cdot , \cdot \rangle$ takes values in $\mathbf{Z}$ and has determinant $\frac{1}{d_{\Gamma}} \prod_{c \in \mathcal{C}_{\Gamma}} e_c$. 2. \[property\_Hecke\_pairing\] The pairing $\langle \cdot , \cdot \rangle$ exchanges the Hecke operators with their duals, [*i.e.* ]{}for all $\varphi_1, \varphi_2 \in \tilde{\mathcal{M}}_{\Gamma}^*$ and $T \in \mathbb{T}$, we have $$\langle T\varphi_1, \varphi_2 \rangle = \langle \varphi_1, W_N T W_N^{-1} \varphi_2 \rangle$$ after inverting $2N$. If $N$ is a prime power, then by Theorem \[thm\_iso\_tilde\_p\] and Theorem \[non\_degeneracy\_petersson\] the pairing $\langle \cdot, \cdot \rangle$ is non-degenerate and (\[property\_Hecke\_pairing\]) is satisfied. Even if in general the Petersson pairing $(\cdot , \cdot)$ can be degenerate, this does not contradicts the conjecture that $\langle \cdot, \cdot \rangle$ should be non-degenerate. Indeed, we have seen that the map $\tilde{p}_{\Gamma}^*$ can have a kernel (when $N$ is not a prime power), and by Theorem \[Haberland\], the kernel of the Petersson pairing $(\cdot, \cdot)$ contains $\operatorname{Ker}(\tilde{p}_{\Gamma}^*)$. We do not know whether this inclusion is always an equality. We now prove Theorem \[intro\_thm\_Petersson\] Recall that by Theorem \[Hecke\_Q\] (\[hecke\_conjugation\_M\]), we have a canonical $\mathbb{T}$-equivariant isomorphism $$\label{complex_conjug_dec_duality} \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}] \xrightarrow{\sim} H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^+ \bigoplus H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^- \text{ .}$$ There is a canonical perfect bilinear pairing, namely the *intersection pairing* $$H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2}])^+ \times H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{2}])^- \rightarrow \mathbf{Z}[\frac{1}{2}] \text{ .}$$ This pairing exchanges the Hecke operators with their duals. Using (\[complex\_conjug\_dec\_duality\]), we get a canonical perfect bilinear pairing $$\cdot \bullet \cdot : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}] \times \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}] \rightarrow \mathbf{Z}[\frac{1}{2N}]$$ exchanging the Hecke operators with their duals. This gives us a linear isomorphism $$F : \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{6N}] \rightarrow \tilde{\mathcal{M}}_{\Gamma}^* \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{6N}] \text{ .}$$ The pairing $\langle \cdot, \cdot \rangle$ induces a linear map $$G : \tilde{\mathcal{M}}_{\Gamma}^* \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}] \rightarrow \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{6N}] \text{ .}$$ Theorem \[intro\_thm\_Petersson\] thus follows from the following result. \[prop\_intersection\_inverse\_pairing\] We have $F \circ G = \operatorname{Id}_{ \tilde{\mathcal{M}}_{\Gamma}^* \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{2N}]}$. By (\[complex\_conjug\_dec\_duality\]), $G$ is the direct sum of two maps $$G^+ : \operatorname{Hom}(H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^+, \mathbf{Z}[\frac{1}{2N}]) \rightarrow H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{6N}])^-$$ and $$G^- : \operatorname{Hom}(H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^-, \mathbf{Z}[\frac{1}{2N}]) \rightarrow H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{6N}])^+ \text{ .}$$ Similarly $F$ is the direct sum of $$F^+ : H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{6N}])^+ \rightarrow \operatorname{Hom}(H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{6N}])^-, \mathbf{Z}[\frac{1}{6N}])$$ and $$F^- : H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{6N}])^-\rightarrow \operatorname{Hom}(H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^+, \mathbf{Z}[\frac{1}{6N}]) \text{ .}$$ To prove Proposition \[prop\_intersection\_inverse\_pairing\], it suffices to prove that $$\label{eq1_prop_intersection_inverse_pairing} F^- \circ G^+ = \operatorname{Id}_{\operatorname{Hom}(H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^+, \mathbf{Z}[\frac{1}{2N}])}$$ and $$\label{eq2_prop_intersection_inverse_pairing} F^+ \circ G^- = \operatorname{Id}_{\operatorname{Hom}(H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{2N}])^-,\mathbf{Z}[\frac{1}{2N}])} \text{ .}$$ Since $G^-$ (resp. $F^+$) is the dual of $G^+$ (resp. $F^-$), it suffices to prove (\[eq2\_prop\_intersection\_inverse\_pairing\]). We use an explicit formula to compute the intersection product, due to Merel [@Merel_HK]. Let $\rho=e^{\frac{2i\pi}{3}} \in \mathfrak{h}$ and $i = e^{\frac{i\pi}{2}} \in \mathfrak{h}$. Let $R$ (resp. $I$) be the image in $Y_{\Gamma}$ of $SL_2(\mathbf{Z})\cdot \rho$ (resp. $\operatorname{SL}_2(\mathbf{Z})\cdot i$). If $g \in \operatorname{SL}_2(\mathbf{Z})$, we denote by $\{gi, g\rho\}$ the image in $H_1(Y_{\Gamma}, R\cup I, \mathbf{Z})$ of the geodesic path in $\mathfrak{h}$ between $gi$ and $g\rho$. Merel proved [@Merel_HK Théorèmes 2 and 3] that $H_1(Y_{\Gamma}, R\cup I, \mathbf{Z})$ is generated by the element $\{gi, g\rho\}$ and that $$H_1(Y_{\Gamma}, \mathbf{Z}) = \big\{\sum_{\Gamma g \in \operatorname{SL}_2(\mathbf{Z})} \lambda_g \cdot \{gi, g\rho\} \text{ such that } \lambda_g \in \mathbf{Z} \text{ and } \lambda_{g}+\lambda_{gS}=\lambda_g+\lambda_{g\tau}+\lambda_{g\tau^2}=0 \text{ for all }g\in \operatorname{SL}_2(\mathbf{Z})\big\}$$ where $\tau = \begin{pmatrix} 0 & -1 \\ 1 & 1 \end{pmatrix} = ST$ stabilizes $\rho$. Furthermore, Merel proved the following formula for the intersection product $\bullet : H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z})\times H_1(Y_{\Gamma}, \mathbf{Z}) \rightarrow \mathbf{Z}$. If $x = \sum_{g \in \operatorname{SL}_2(\mathbf{Z})} \mu_g\cdot \{g0, g\infty\} \in H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z})$ and $y = \sum_{g \in \operatorname{SL}_2(\mathbf{Z})} \lambda_g\cdot \{gi, g\rho\} \in H_1(Y_{\Gamma}, \mathbf{Z})$ then [@Merel_HK Corollaire 3] we have $$\label{Merel_formula_intersection} x\bullet y = \sum_{g \in \Gamma\backslash \operatorname{SL}_2(\mathbf{Z})} \lambda_g\cdot \mu_g \text{ .}$$ To prove (\[eq1\_prop\_intersection\_inverse\_pairing\]), it thus suffices to prove the following result, whose statement and proof are generalization of the first formula in [@Merel_Manin_L Théorème 2]. \[expression\_G\^+\_rho\_i\] We do not make any assumption on $\Gamma$ (which can thus be a non-congruence subgroup). For all $\varphi \in \operatorname{Hom}(H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}), \mathbf{Z}) \subset \tilde{\mathcal{M}}_{\Gamma}^*$, we have in $H_1(Y_{\Gamma}, \mathbf{Z}) \subset \tilde{\mathcal{M}}_{\Gamma}$: $$G(\varphi) = \sum_{g \in \Gamma\backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\rho\} \text{ .}$$ In other words, the pairing $\langle \cdot, \cdot \rangle$ induces by restriction a map $\operatorname{Hom}(H_1(X_{\Gamma}, \mathcal{C}_{\Gamma}, \mathbf{Z}), \mathbf{Z}) \rightarrow H_1(Y_{\Gamma}, \mathbf{Z})$ which is inverse to the intersection pairing. Let $N$ be the general level of $\Gamma$. We let $\tilde{\mathcal{M}}_{\Gamma, R\cup I}$ be the pushout of the two (injective) morphisms $H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{6N}]) \rightarrow \tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{6N}]$ and $H_1(Y_{\Gamma}, \mathbf{Z}[\frac{1}{6N}]) \rightarrow H_1(Y_{\Gamma}, R\cup I, \mathbf{Z}[\frac{1}{6N}])$. Let $\tilde{\mathcal{M}}_{\Gamma, R\cup I}' = \tilde{\mathcal{M}}_{\Gamma, R\cup I} \oplus \mathbf{Z}[\frac{1}{6N}]\cdot \{1, i\}$ (where $\{1, i\}$ is a formal symbol). For any $(g,g') \in \operatorname{SL}_2(\mathbf{Z})$, we define the symbol $\{g,g'i\}$ in $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$ to be $\{g,1\}+\{1,i\}+\{i,g'i\} \in\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$. Similarly, we define the symbol $\{g,g'\rho\}$ in $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$ to be $\{g,1\}+\{1,i\}+\{i,g'\rho\} \in\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$. Finally, we define $\{g'i,g\}$ to be $-\{g,g'i\}$ and $\{g'\rho,g\}$ to be $-\{g,g'\rho\}$. We can thus talk about symbols $\{\alpha, \beta\}$ in $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$, where $\alpha, \beta \in \{\operatorname{SL}_2(\mathbf{Z}), \operatorname{SL}_2(\mathbf{Z})\cdot i,\operatorname{SL}_2(\mathbf{Z})\cdot \rho\}$. We easily check that the Chasles relation are satisfied and that for any $\gamma \in \Gamma$, we have $\{\gamma \alpha, \gamma \beta\} = \{\alpha, \beta\}$. Since $\tilde{\mathcal{M}}_{\Gamma} \otimes_{\mathbf{Z}} \mathbf{Z}[\frac{1}{6N}]$ embeds into $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$, it is legitimate to do all our computations in $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$. On the one hand, by definition of $G$, we have in $\tilde{\mathcal{M}}_{\Gamma}$: $$\begin{aligned} 6\cdot G(\varphi) &= \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{gS\infty,g\infty\})\cdot \{gTS, gT\}- \varphi(\{gTS\infty, gT\infty\}) \cdot \{gS,g\} \\& + 4\cdot \varphi(\{g\infty,gS\infty\}) \cdot \{g,gT\} \\&= \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\infty,g0\})\cdot \{gSTS, gST\}-\varphi(\{gSTS\infty, gST\infty\}) \cdot \{g,gS\} \\&- 4\cdot \varphi(\{g0,g\infty\}) \cdot \{g,gT\} \\&=\sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\infty,g0\})\cdot \{g\tau S, g\tau\} - \varphi(\{g\tau S\infty, g\tau\infty\}) \cdot \{g,gS\}\\& - 4\cdot \varphi(\{g0,g\infty\}) \cdot \{g,gT\} \\& =\sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\tau0,g\tau\infty\})\cdot \{g\tau^2, g\tau^2S\}-\varphi(\{g\tau 0, g\tau\infty\}) \cdot \{g,gS\} \\& - 4\cdot \varphi(\{g0,g\infty\}) \cdot \{g,gT\} \text{ .}\end{aligned}$$ On the other hand, we have in $\tilde{\mathcal{M}}_{\Gamma, R\cup I}'$: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\rho\}&= \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \left(\{gi, g\}+\{g, g\rho\}\right) \text{ .}\end{aligned}$$ We have, using the fact that $Si=i$: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\}&= \frac{1}{2}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\} + \varphi(\{gS0, gS\infty\})\cdot \{gSi, gS\} \\&= \frac{1}{2}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gS,g\} \text{ .}\end{aligned}$$ We have: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{g, g\rho\}&= \frac{1}{3}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{g, g\rho\} \\& + \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau, g\tau\rho\} +\varphi(\{g\tau^20, g\tau^2\infty\})\cdot \{g\tau^2, g\tau^2\rho\} \text{ .}\end{aligned}$$ Since $\tau \rho = \rho$ and $ \varphi(\{g0, g\infty\}) = - \varphi(\{g\tau0, g\tau\infty\})-\varphi(\{g\tau^20, g\tau^2\infty\})$, we get $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{g, g\rho\}&= \frac{1}{3}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau, g\} \\& + \varphi(\{g\tau^20, g\tau^2\infty\})\cdot \{g\tau^2, g\} \text{ .}\end{aligned}$$ Note that $$\begin{aligned} \{g\tau^2, g\} = \{g\tau^2, g\tau^2\tau\}= \{g\tau^2, g\tau^2ST\} = \{g\tau^2, g\tau^2S\} + \{g\tau^2S, g\tau^2ST\} \text{ .}\end{aligned}$$ Thus, we have: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{g, g\rho\}&= \frac{1}{3}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau, g\} \\& + \varphi(\{g0, g\infty\})\cdot \{g, gS\}+ \varphi(\{g0, g\infty\})\cdot \{gS,gST\} \\&= \frac{1}{3}\cdot \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau, g\} \\& + \varphi(\{g0, g\infty\})\cdot \{g, gS\}- \varphi(\{g0, g\infty\})\cdot \{g,gT\} \text{ .}\end{aligned}$$ Combining the two computations above, we get: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\rho\}&=\sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \frac{1}{6}\cdot \varphi(\{g0, g\infty\})\cdot \{gS,g\}+\frac{1}{3}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau, g\} \\& - \frac{1}{3}\cdot \varphi(\{g0, g\infty\})\cdot \{g,gT\} \\&= \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \frac{1}{6}\cdot \varphi(\{g0, g\infty\})\cdot \{gS,g\}\\&+\frac{1}{3}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot (\{gST, gS\} +\{gS,g\})\\& - \frac{1}{3}\cdot \varphi(\{g0, g\infty\})\cdot \{g,gT\} \\&= \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \frac{1}{6}\cdot \varphi(\{g0, g\infty\})\cdot \{gS,g\}\ +\frac{1}{3}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot\{gS,g\} \\& - \frac{2}{3}\cdot \varphi(\{g0, g\infty\})\cdot \{g,gT\}\end{aligned}$$ where in the last equality, we have used the fact that $\{gST,gS\} = \{g\tau, g\tau T^{-1}\} = -\{g\tau, g\tau T\}$. Using again the fact that $ \varphi(\{g0, g\infty\}) + \varphi(\{g\tau0, g\tau\infty\})+\varphi(\{g\tau^20, g\tau^2\infty\})=0$, we have: $$\begin{aligned} \sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \varphi(\{g0, g\infty\})\cdot \{gi, g\rho\}&=\sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \frac{1}{6}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot \{gS,g\}-\frac{1}{6}\cdot \varphi(\{g\tau^20, g\tau^2\infty\})\cdot \{gS, g\} \\& - \frac{2}{3}\cdot \varphi(\{g0, g\infty\})\cdot \{g,gT\} \\&=\sum_{g \in \Gamma \backslash \operatorname{SL}_2(\mathbf{Z})} \frac{1}{6}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot \{gS,g\}-\frac{1}{6}\cdot \varphi(\{g\tau0, g\tau\infty\})\cdot \{g\tau^2S, g\tau^2\} \\& - \frac{2}{3}\cdot \varphi(\{g0, g\infty\})\cdot \{g,gT\} \\&=G(\varphi) \text{ .}\end{aligned}$$ This concludes the proof of Proposition \[expression\_G\^+\_rho\_i\], and thus the proof of Proposition \[prop\_intersection\_inverse\_pairing\] and Theorem \[intro\_thm\_Petersson\]. \[rem\_generality\_conjecture\_G\] Proposition \[expression\_G\^+\_rho\_i\] is the strongest general result we were able to prove toward Properties (\[property\_degenerate\_pairing\]) and (\[property\_Hecke\_pairing\]). Note that the pairing $\langle \cdot, \cdot \rangle$ must have determinant divisible by the index of $H_1(Y_{\Gamma}, \mathbf{Z})$ in $\tilde{\mathcal{M}}_{\Gamma}$, [*i.e.* ]{}$\frac{1}{d_{\Gamma}}\prod_{c \in \mathcal{C}_{\Gamma}} e_c$. Relation with the generalized Jacobian {#generalized_cuspidal} ====================================== We remind the reader that we keep the notation of §\[intro\_generalized\_cuspidal\_motive\]. We first define the two maps $\delta_{\Gamma}^{\natural, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}$ and $\delta_{\Gamma}^{\natural, \operatorname{an}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}$ such that $$\beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{alg}} = \beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{an}} = \delta_{\Gamma} \text{ .}$$ Definitions and comparison result --------------------------------- ### Algebraic definition {#subsection_alg_def} We first recall an alternative algebraic description of $J_{\Gamma}^{\#}$, which the author has learned in [@Yamazaki §2.2]. Let $F$ be a field extension of $k$ containing the compositum of the fields $k(c)$ for $c \in X_{\Gamma}\backslash Y_{\Gamma}$. Let $K$ be the function field of $X_{\Gamma} \times_k F$. If $P$ is a closed point in $X_{\Gamma} \times_k F$, let $K_P$ be the completion of $K$ at $P$ and $U_P \subset K_P^{\times}$ be the group of principal unit. Let $$\operatorname{Div}(X_{\Gamma}, \mathcal{C}_{\Gamma})(F):=\operatorname{Div}(Y_{\Gamma})(F) \oplus \bigoplus_{c \in \mathcal{C}_{\Gamma}} K_{c}^{\times}/U_c^{\times}\text{ ,}$$ where $\operatorname{Div}(Y_{\Gamma})(F)$ is the group of divisors of $Y_{\Gamma}$ defined over $F$. We denote by $\operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)$ the kernel of the degree map $\operatorname{Div}(X_{\Gamma}, \mathcal{C}_{\Gamma})(F) \rightarrow \mathbf{Z}$ given by $$D \oplus (f_c \text{ modulo }U_c)_{c \in \mathcal{C}_{\Gamma}} \mapsto \deg(D) + \sum_{c \in \mathcal{C}_{\Gamma}} \operatorname{ord}_c(f_c) \text{ .}$$ There is a canonical map $K^{\times} \rightarrow \operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)$, given by $$f \mapsto \text{div}_{Y_{\Gamma}}(f) \oplus (f \text{ modulo }U_{c})_{c \in \mathcal{C}_{\Gamma}}$$ where $\text{div}_{Y_{\Gamma}}(f)$ is the divisor of the restriction of $f$ to $Y_{\Gamma}$. Then there is a canonical $\operatorname{Gal}(F/k)$-equivariant group isomorphism $$\label{iso_gen_jac_alg} J_{\Gamma}^{\#}(F) \xrightarrow{\sim} \operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)/K^{\times}$$ sending the class of a divisor $D$ supported on $Y_{\Gamma}$ to the image of $(D \oplus 0)$ in $\operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)/K^{\times}$. Under this identification, the map $J_{\Gamma}^{\#}(F) \rightarrow J_{\Gamma}(F)$ corresponds to the map $\operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)/K^{\times} \rightarrow J_{\Gamma}(F)$ given by $$D \oplus (f_c \text{ modulo }U_c)_{c \in \mathcal{C}_{\Gamma}} \mapsto \text{class of the divisor }D + \sum_{c \in \mathcal{C}_{\Gamma}} \operatorname{ord}_c(f_c) \cdot (c) \text{ .}$$ We now define the map $\delta_{\Gamma}^{\natural, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}(\mathbf{C})$. We apply the discussion above to $F = \mathbf{C}$. For any $c \in \mathcal{C}_{\Gamma}$, let $\frac{p_c}{q_c} \in \mathbf{Q}$ such that $c = \Gamma \cdot \frac{p_c}{q_c}$ and let $g_c \in \operatorname{SL}_2(\mathbf{Z})$ such that $g_c( \frac{p_c}{q_c}) = \infty$. Then the function $t_c : z\mapsto j(\frac{g_c z}{e_c})^{-1}$ of the upper-half plane induces a uniformizer at $c$, still denoted by $t_c \in K_c^{\times}$ (we recall that $e_c$ is the width of $c$). The map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ sends a degree zero divisor $\sum_{c \in \mathcal{C}_{\Gamma}} n_c \cdot (c)$ to the image of $0 \oplus (t_c^{n_c} \text{ modulo }U_c)_{c \in \mathcal{C}_{\Gamma}}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$ via the identification (\[iso\_gen\_jac\_alg\]). The following result motivates the definition of $J_{\Gamma}^{\natural}$. \[beta\_alg\_lemma\] 1. \[i\_commutative\_beta\]We have $\beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{alg}} = \delta_{\Gamma}$. 2. \[ii\_well\_defined\_beta\]The map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ is independant of the choice of the elements $\frac{p_c}{q_c}$ and of $g_c$. 3. \[iii\_rationality\_alg\_beta\] The map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ takes values in $J_{\Gamma}^{\natural}(\mathbf{Q}(\zeta_N))$. 4. \[iv\_galois\_equivariance\] Assume that $\Gamma = \Gamma_0(N)$ or $\Gamma = \Gamma_1(N)$ (in which case we have $k=\mathbf{Q}$). Then the map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ is $\operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$-equivariant. Point (\[i\_commutative\_beta\]) is straightforward. We prove point (\[ii\_well\_defined\_beta\]). First, fix the choice of $\frac{p_c}{q_c}$. Recall that locally at $\infty$, we have $j(z)^{-1} \sim q(z) = e^{2i\pi z}$ modulo principal units. Thus, if we replace $g_c$ by $T^ng_c$ for some $n \in \mathbf{Z}$, then $t_c$ is replaced by $e^{\frac{2i \pi n}{e_c}}\cdot t_c$. By construction, the image of $\prod_{c \in \mathcal{C}_{\Gamma}} \mu_{e_c}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$ is trivial, so our construction does not depend on the choice of the elements $g_c$. The fact that $t_c$ does not depend on the choice of $\frac{p_c}{q_c}$ is obvious. Point (\[iii\_rationality\_alg\_beta\]) follows from the following result. For all $c \in \mathcal{C}_{\Gamma}$, $t_c$ is defined over $\mathbf{Q}(\zeta_N)$. Let $c \in \mathcal{C}_{\Gamma}$ and $\Gamma' = \Gamma \cap \Gamma(e_c)$, where $\Gamma(e_c)$ is the principal congruence subgroup of level $e_c$. Note that $\Gamma(N)$ is a subgroup of $\Gamma'$ since $e_c$ divides $N$ and $\Gamma(N) \subset \Gamma$. Let $c' = \Gamma' \cdot \frac{p_c}{q_c}$; it is a cusp of $X_{\Gamma'}$. The covering $X_{\Gamma'} \rightarrow X_{\Gamma}$ is unramified at $c'$ (where $X_{\Gamma'}$ and $X_{\Gamma}$ are considered over $\mathbf{Q}(\zeta_N)$), since the ramification index of both $X_{\Gamma'} \rightarrow X_{\operatorname{SL}_2(\mathbf{Z})}$ and $X_{\Gamma} \rightarrow X_{\operatorname{SL}_2(\mathbf{Z})}$ is $e_c$ at $c'$ and $c$ respectively. The function $j(\frac{g_c z}{e_c})$ is modular for the group $\Gamma'$, and so it suffices to prove that it is defined over $\mathbf{Q}(\zeta_N)$. This is true at the level of the modular curve $X_{\Gamma(N)}$, so it is true for $X_{\Gamma'}$ since the covering $X_{\Gamma(N)} \rightarrow X_{\Gamma'}$ is defined over $\mathbf{Q}(\zeta_N)$. We now prove point (\[iv\_galois\_equivariance\]). The action of $\operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$ on $\mathbf{Z}[\mathcal{C}_{\Gamma}]^0$ comes from its action on $\mathcal{C}_{\Gamma}$. If $c \in \mathcal{C}_{\Gamma}$ and $g \in\operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$, we denote by $g(c) \in \mathcal{C}_{\Gamma}$ the cusp corresponding to the action of $g$ on $c$. Under (\[iso\_gen\_jac\_alg\]), the action of $\operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$ on $\operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(F)/K^{\times}$ is given as follows. The action of $\operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$ on $K^{\times}$ and $\operatorname{Div}(Y_{\Gamma})(F)$ is the obvious one. If $g \in \operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$, then $g$ acts on $\bigoplus_{c \in \mathcal{C}_{\Gamma}} K_{c}^{\times}/U_c^{\times}$ by $g \cdot (f_c \text{ modulo }U_c)_{c \in \mathcal{C}_{\Gamma}} = (g\cdot f_c)_{g(c), c \in \mathcal{C}_{\Gamma}}$. To prove (\[iv\_galois\_equivariance\]), it suffices to prove that for all $c \in \mathcal{C}_{\Gamma}$ and $g \in \operatorname{Gal}(\mathbf{Q}(\zeta_N)/\mathbf{Q})$, we have $(g\cdot f_c)^{e_{g(c)}} \equiv f_{g(c)}^{e_{g(c)}} \text{ (modulo }U_{g(c)}\text{)}$. We have $f_{g(c)}^{e_{g(c)}} \equiv j(z) \text{ (modulo }U_{g(c)}\text{)}$. Since $e_{g(c)} = e_c$, we have $(g\cdot f_c)^{e_{g(c)}} = g\cdot (f_c^{e_c}) \equiv g\cdot j(z) \equiv j(z) \text{ (modulo }U_{g(c)}\text{)}$. We have used the fact that $j : X_{\Gamma} \rightarrow X_{\operatorname{SL}_2(\mathbf{Z})}$ is defined over $\mathbf{Q}$. ### Analytic definition There is a well-known analytic uniformization of the generalized Jacobian $J_{\Gamma}^{\#}(\mathbf{C})$. Namely, there is an isomorphism of complex abelian varieties (called the generalized Abel-Jacobi map) $$\operatorname{AJ}_{\Gamma} : J_{\Gamma}^{\#}(\mathbf{C}) \xrightarrow{\sim} \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})/H_1(Y_{\Gamma}, \mathbf{Z}) \text{ ,}$$ sending the class of a divisor $D$ to the class of the morphism $\varphi : M_2(\Gamma) \rightarrow \mathbf{C}$ given by $f \mapsto \int_{\gamma} 2i\pi f(z)dz$ where $\gamma$ is a $1$-chain in $Y_{\Gamma}$ whose boundary is $D$. Recall that we have a group homomorphism $\tilde{p}_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma),\mathbf{C})$. By Proposition \[rank\_computation\] (\[rank\_computation\_2\]) and (\[rank\_computation\_3\]), the compositum $$\tilde{\mathcal{M}}_{\Gamma} \xrightarrow{\tilde{p}_{\Gamma}} \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C}) \rightarrow \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})/H_1(Y_{\Gamma}, \mathbf{Z}) \xrightarrow{\operatorname{AJ}_{\Gamma}^{-1}} J_{\Gamma}^{\#}(\mathbf{C}) \rightarrow J_{\Gamma}^{\natural}(\mathbf{C})$$ factors through $\partial_{\Gamma} : \tilde{\mathcal{M}}_{\Gamma} \rightarrow \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$, so we have defined a map $$\delta_{\Gamma}^{\natural, \operatorname{an}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}(\mathbf{C}) \text{ .}$$ By Theorem \[Hecke\_Q\], the map $\delta_{\Gamma}^{\natural, \operatorname{an}}$ is $\mathbb{T}$-equivariant. \[beta\_an\_lemma\] We have $\beta_{\Gamma}^{\natural} \circ \delta_{\Gamma}^{\natural, \operatorname{an}} = \delta_{\Gamma}$. This follows from the fact that for any cuspidal modular form $f$ and any $g \in \operatorname{SL}_2(\mathbf{Z})$, we have $\mathcal{S}(g)(f) = 2i\pi\int_{\infty}^{g(\infty)} f(z)dz$. ### Comparison between the algebraic and analytic definitions {#subsection_comparison} We now prove Theorem \[intro\_main\_thm\_comparison\]. We first prove point (\[comparison\_i\]). By Proposition \[beta\_alg\_lemma\] (\[i\_commutative\_beta\]) and Lemma \[beta\_an\_lemma\], the map $\delta_{\Gamma}^{\natural, \operatorname{alg}} - \delta_{\Gamma}^{\natural, \operatorname{an}}$ takes values in the image of $\prod_{c \in \mathcal{C}_{\Gamma}} \mathbf{C}^{\times}/\mu_{N}$ inside $J_{\Gamma}^{\natural}(\mathbf{C})$. If $(\lambda_c)_{c \in \mathcal{C}_{\Gamma}} \in \mathbf{C}^{\times}$, we abusively use the same notation for its image in $J_{\Gamma}^{\natural}(\mathbf{C})$. Recall ([*cf.* ]{}paragraph \[subsection\_alg\_def\]) that for any $c \in \mathcal{C}_{\Gamma}$, we write $c = \Gamma \cdot \frac{p_c}{q_c}$ and we fix some matrix $g_c \in \operatorname{SL}_2(\mathbf{Z})$ such that $g_c(\frac{p_c}{q_c}) = \infty$. We then let $t_c(z) = j(\frac{g_c(z)}{e_c})^{-1}$ and $u_c(z) = e^{\frac{2i\pi g_c(z)}{e_c}}$. Note that $u_c(z) \sim t_c(z)$ near $\frac{p_c}{q_c}$. \[main\_thm\_comparison\_lemma\_1\] Let $D = \sum_{c \in \mathcal{C}_{\Gamma}} n_c \cdot (c) \in \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$, $m_D$ be the order of the image of $D$ in $J_{\Gamma}$ and $u$ be a modular unit with divisor $m_D\cdot D$. We have: $$m_D \cdot \delta_{\Gamma}^{\natural, \operatorname{alg}}(D) = ((\frac{t_c^{m_D\cdot n_c}}{u})(c))_{c \in \mathcal{C}_{\Gamma}} \text{ .}$$ By definition, $m_D \cdot \delta_{\Gamma}^{\natural, \operatorname{alg}}(D)$ is the image of $0 \oplus (t_c^{m_D\cdot n_c})_{c\in \mathcal{C}_{\Gamma}}\in \operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(\mathbf{C})$ in $J_{\Gamma}^{\natural}(\mathbf{C})$. Thus, $m_D \cdot \delta_{\Gamma}^{\natural, \operatorname{alg}}(D)$ is also the image of $0 \oplus ((\frac{t_c^{m_D\cdot n_c}}{u})(c))_{c \in \mathcal{C}_{\Gamma}} \in \operatorname{Div}^0(X_{\Gamma}, \mathcal{C}_{\Gamma})(\mathbf{C})$ in $J_{\Gamma}^{\natural}(\mathbf{C})$. This concludes the proof of Lemma \[main\_thm\_comparison\_lemma\_1\]. \[main\_thm\_comparison\_lemma\_2\] Let $D = \sum_{c \in \mathcal{C}_{\Gamma}} n_c \cdot (c) \in \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$. Lift $\delta_{\Gamma}^{\natural, \operatorname{an}}(D)$ to an element $ \varphi_D \in \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$. Let $v$ be a modular unit of level $\Gamma$, and $d\log(v) \in M_2(\Gamma)$ denote the Eisenstein series $\frac{v'}{v}$. The quantity $\varphi_D(\frac{1}{2i\pi}d\log(v))$ is canonical in $\mathbf{C}/\frac{2i \pi}{N} \mathbf{Z}$. We have the following equality in $\mathbf{C}^{\times}/\mu_{N}$: $$\exp(\varphi_D(\frac{1}{2i\pi}d\log(v))) = \prod_{c \in \mathcal{C}_{\Gamma}} (\frac{v}{t_c^{\operatorname{ord}_c(v)}})^{n_c}(c)\text{ .}$$ By construction of $\delta_{\Gamma}^{\natural, \operatorname{an}}$, we can choose $\varphi_D$ so that for any $f \in M_2(\Gamma)$, we have $$\varphi_D(f) = \sum_{c \in \mathcal{C}_{\Gamma}} n_c \cdot \mathcal{S}(g_c^{-1})(f) \text{ ,}$$ where we recall that $$\mathcal{S}(g)(f) = 2i\pi \cdot \int_{z_0}^{g(z_0)} f(z)dz - 2i\pi\cdot z_0\cdot (a_0(f\mid g)-a_0(f)) + 2i\pi\cdot \int_{z_0}^{i\infty} \left((f\mid g)(z)-a_0(f\mid g)\right) - \left(f(z) - a_0(f)\right) dz$$ for any $g \in \operatorname{SL}_2(\mathbf{Z})$ and $f \in M_2(\Gamma)$. We can simplify this formula when $f = \frac{1}{2i\pi}\frac{v'}{v}$, where we view $v$ as a function on the upper-half plane $\mathfrak{h}$. For any $c \in \mathcal{C}_{\Gamma}$, let $v_c = \frac{v}{u_c^{\operatorname{ord}_c(v)}} \circ g_c^{-1}$. Fix a logarithm of $v_c$ on $\mathfrak{h}$, denoted by $\log(v_c)$ (so by definition we have $\exp(\log(v_c))=v_c$ on $\mathfrak{h}$). We also fix a logarithm $\log(v)$ of $v$ on $\mathfrak{h}$; note that $\log(v)' = 2i\pi f$. For all $c \in \mathcal{C}_{\Gamma}$, we easily see that $a_0(f\mid g_c^{-1}) = \frac{1}{e_{c}}\cdot \operatorname{ord}_{c}(v)$ and $2i \pi \cdot (f\mid g_c^{-1})-2i\pi \cdot a_0(f\mid g_c^{-1}) = \log(v_c)'$. Thus, we have: $$\begin{aligned} \mathcal{S}(g_c^{-1})(f) &= \log(v)(g_c^{-1}(z_0)) - \log(v)(z_0) + 2i\pi \cdot z_0\cdot (\frac{1}{e_c}\cdot \operatorname{ord}_c(v)-\frac{1}{e_{\Gamma \infty}}\cdot \operatorname{ord}_{\Gamma \infty}(v)) \\& + \log(v_c)(i\infty)-\log(v_c)(z_0)-\log(v)(i\infty) + \log(v)(z_0) \text{ .}\end{aligned}$$ Since $\sum_{c \in \mathcal{C}_{\Gamma}} n_c = 0$, we have $$\begin{aligned} \sum_{c \in \mathcal{C}_{\Gamma}} n_c\cdot \mathcal{S}(g_c^{-1})(f) = \sum_{c \in \mathcal{C}_{\Gamma}} n_c\cdot \left( \log(v)(g_c^{-1}(z_0)) + 2i\pi \cdot z_0\cdot \frac{1}{e_c}\cdot \operatorname{ord}_c(v) + \log(v_c)(i\infty)-\log(v_c)(z_0) \right) \text{ .}\end{aligned}$$ Thus, we have in $\mathbf{C}^{\times}/\mu_N$: $$\begin{aligned} \exp(\varphi_D(\frac{1}{2i\pi}d\log(v))) &= \prod_{c \in \mathcal{C}_{\Gamma}} \left(v(g_{c}^{-1}(z_0))\cdot \exp(2i\pi \cdot z_0\cdot \frac{1}{e_c}\cdot \operatorname{ord}_c(v))\cdot v_c(i\infty)\cdot v_c(z_0)^{-1}\right)^{n_c} \\& = \prod_{c \in \mathcal{C}_{\Gamma}} v_c(i\infty)^{n_c} \\&= \prod_{c \in \mathcal{C}_{\Gamma}} (\frac{v}{t_c^{\operatorname{ord}_c(v)}})^{n_c}(c) \text{ .}\end{aligned}$$ This concludes the proof of Lemma \[main\_thm\_comparison\_lemma\_2\]. \[main\_thm\_comparison\_lemma\_3\] Let $(\lambda_c)_{c \in \mathcal{C}_{\Gamma}} \in \prod_{c \in \mathcal{C}_{\Gamma}} \mathbf{C}^{\times}$. Let $\varphi \in \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma),\mathbf{C})$ whose class in $J_{\Gamma}^{\#}(\mathbf{C})$ corresponds to $(\lambda_c)_{c \in \mathcal{C}_{\Gamma}}$ via the generalized Abel-Jacobi isomorphism $AJ_{\Gamma}$. For any modular unit $v$ of $X_{\Gamma}$, we have $$\exp(\varphi(\frac{1}{2i\pi}d\log(v))) = \prod_{c \in \mathcal{C}_{\Gamma}} \lambda_c^{-\operatorname{ord}_c(v)} \text{ .}$$ Let $f$ be a meromorphic function on $X_{\Gamma}$ whose divisor $\text{div}(f)$ is supported on $Y_{\Gamma}$, and such that for all $c \in \mathcal{C}_{\Gamma}$ we have $f(c) = \lambda_c$. We write $\text{div}(f) = \sum_{i=1}^{m} m_i(Q_i)$ for some $Q_i\in Y_{\Gamma}$ and $m_i \in \mathbf{Z}$. By the construction of $J_{\Gamma}^{\#}$ in paragraph \[subsection\_alg\_def\], we have in $\mathbf{C}/2i\pi \mathbf{Z}$: $$\varphi(\frac{1}{2i\pi}d\log(v)) = -\int_{\text{div}(f)} d\log(v)(z)dz \text{ ,}$$ where the integral is over any $1$-chain with boundary $\text{div}(f)$. The following computation was given to us by the mathoverflow user abx in his answer [@MO_abel-jacobi]. Such a $1$-chain can be written as a linear combination of paths $\gamma_j$ from $Q_{j_1}$ to $Q_{j_2}$ for some indexes $j_1$ and $j_2$. By choosing a determination of $\log(v)$ along the path $\gamma_j$, we get: $$\exp(\int_{\gamma_j} d\log(v)) = \frac{v(Q_{j_1})}{v(Q_{j_2})} \text{ .}$$ Consequently, we have: $$\exp(\int_{\text{div}(f)} d\log(v)(z)dz) = \prod_{i=1}^m v(Q_i)^{m_i} \text{ .}$$ By Weil’s reciprocity law, we have: $$\prod_{i=1}^m v(Q_i)^{m_i} = \prod_{c \in \mathcal{C}_{\Gamma}} f(c)^{\operatorname{ord}_c(v)} = \prod_{c \in \mathcal{C}_{\Gamma}} \lambda_c^{\operatorname{ord}_c(v)} \text{ .}$$ This concludes the proof of Lemma \[main\_thm\_comparison\_lemma\_3\]. Let $u$ be a modular unit of $Y_{\Gamma}$. Using the generalized Abel-Jacobi isomorphism, lift $\delta_{\Gamma}^{\natural, \operatorname{alg}}(\text{div}(u))$ (resp. $\delta_{\Gamma}^{\natural, \operatorname{an}}(\text{div}(u))$) to an element $\varphi_u^{\operatorname{alg}}$ (resp. $\varphi_u^{\operatorname{an}}$) of $\operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma), \mathbf{C})$. By Lemmas \[main\_thm\_comparison\_lemma\_1\] and \[main\_thm\_comparison\_lemma\_3\], for any modular unit $v$ we have in $\mathbf{C}^{\times}/\mu_{N}$: $$\exp(\varphi_u^{\operatorname{alg}}(\frac{1}{2i\pi}d\log(v))) = \prod_{c \in \mathcal{C}_{\Gamma}} (\frac{u}{t_c^{\operatorname{ord}_c(u)}}(c))^{\operatorname{ord}_c(v)} \text{ .}$$ On the other hand, by Lemma \[main\_thm\_comparison\_lemma\_2\] we have in $\mathbf{C}^{\times}/\mu_{N}$: $$\exp(\varphi_u^{\operatorname{an}}(\frac{1}{2i\pi}d\log(v))) = \prod_{c \in \mathcal{C}_{\Gamma}} (\frac{v}{t_c^{\operatorname{ord}_c(v)}}(c))^{\operatorname{ord}_c(u)} \text{ .}$$ By Weil reciprocity law, we have in $\mathbf{C}^{\times}/\mu_{N}$: $$\exp(\varphi_u^{\operatorname{alg}}(\frac{1}{2i\pi}d\log(v)))=\exp(\varphi_u^{\operatorname{an}}(\frac{1}{2i\pi}d\log(v))) \text{ .}$$ By Lemma \[main\_thm\_comparison\_lemma\_3\], the image of $\varphi_u^{\operatorname{alg}}-\varphi_u^{\operatorname{an}}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$ is equal to the image of $(\lambda_c)_{c \in \mathcal{C}_{\Gamma}}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$ for some $\lambda_c \in \mathbf{C}^{\times}/\mu_{N}$ such that for all modular unit $v$, we have in $\mathbf{C}^{\times}/\mu_{N}$: $$\prod_{c \in \mathcal{C}_{\Gamma}} \lambda_c^{\operatorname{ord}_c(v)}=1 \text{ .}$$ In particular, there exists $\lambda \in \mathbf{C}^{\times}/\mu_{N}$ such that for all $c \in \mathcal{C}_{\Gamma}$ we have $\lambda_c^n=\lambda$. This proves that for all modular unit $u$, we have $$n\cdot \delta_{\Gamma}^{\natural, \operatorname{alg}}(\text{div}(u)) = n\cdot \delta_{\Gamma}^{\natural, \operatorname{an}}(\text{div}(u)) \text{ .}$$ In particular, for all $D \in \mathbf{Z}[\mathcal{C}_{\Gamma}]^0$, we have $$n^2\cdot \delta_{\Gamma}^{\natural, \operatorname{alg}}(D) = n^2\cdot \delta_{\Gamma}^{\natural, \operatorname{an}}(D) \text{ .}$$ This concludes the proof of point (\[comparison\_i\]). We now prove point (\[comparison\_ii\]). The map $\delta_{\Gamma}^{\natural, \operatorname{alg}} - \delta_{\Gamma}^{\natural, \operatorname{an}}$ takes values in the image of $\prod_{c \in \mathcal{C}_{\Gamma}} \mathbf{C}^{\times}/\mu_{N}$ in $J_{\Gamma}^{\natural}(\mathbf{C})$, and has finite order by (\[comparison\_i\]). The map $\delta_{\Gamma}^{\natural, \operatorname{alg}} - \delta_{\Gamma}^{\natural, \operatorname{an}}$ is invariant by the action of the complex conjugation by assumption, by Proposition \[beta\_alg\_lemma\] (\[iv\_galois\_equivariance\]) and by Proposition \[equivariance\_complex\_conjug\_period\]. Since $N$ is assumed to be odd, an element of finite order in $\mathbf{C}^{\times}/\mu_{N}$ fixed by the complex conjugation is equal to $\pm1$. This concludes the proof of Theorem \[intro\_main\_thm\_comparison\]. \[remarks\_after\_comparison\] 1. \[remarks\_after\_comparison\_i\] By Theorem \[intro\_main\_thm\_comparison\] (\[comparison\_i\]), the map $\delta_{\Gamma}^{\natural, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J_{\Gamma}^{\natural}(\mathbf{C})$ is injective if and only if $\delta_{\Gamma}^{\natural, \operatorname{an}}$ is injective. We easily see that $\delta_{\Gamma}^{\natural, \operatorname{an}}$ is injective if $\tilde{p}_{\Gamma} \otimes \mathbf{R}$ is an isomorphism. By Theorem \[thm\_iso\_tilde\_p\], this is the case if $\Gamma$ has prime power level. 2. It would be interesting to remove the hypotheses of Theorem \[intro\_main\_thm\_comparison\] (\[comparison\_ii\]). It is not clear to us whether they are necessary. 3. \[remarks\_after\_comparison\_ii\] Assume that $\Gamma$ satisfies the assumptions of Theorem \[intro\_main\_thm\_comparison\] (\[comparison\_ii\]). Let $J_{\Gamma}^{\flat}$ be the semi-abelian variety over $k$ defined by $J_{\Gamma}^{\flat} = J_{\Gamma}^{\natural}/\prod_{c \in \mathcal{C}_{\Gamma}} \operatorname{Res}_{k(c)/k}(\mu_2)$. The map $\delta_{\Gamma}^{\natural, \operatorname{alg}}$ (or equivalently $\delta_{\Gamma}^{\natural, \operatorname{an}}$) gives a natural group homomorphism $\delta_{\Gamma}^{\flat} : \mathbf{Z}[\mathcal{C}_{\Gamma}]^0 \rightarrow J^{\flat}_{\Gamma}(\mathbf{Q}(\zeta_N))$. If $\Gamma = \Gamma_1(N)$ or $\Gamma = \Gamma_0(N)$ (with $N$ necessarily an odd prime given our assumptions), then by Theorem \[intro\_main\_thm\_comparison\] (\[comparison\_ii\]) and Theorem \[Hecke\_Q\], the map $\delta_{\Gamma}^{\flat}$ is $\mathbb{T}$-equivariant. Applications to the modular curve $X_0(p)$ {#paragraph_applications_Gamma_0(p)} ------------------------------------------ Let $p$ be an odd prime and $\Gamma = \Gamma_0(p)$. Let $w_p$ be the Atkin-Lehner involution acting on $X_0(p) = X_{\Gamma_0(p)}$. Note that we have $\mathcal{C}_{\Gamma_0(p)} = \{\Gamma_0(p) 0, \Gamma_0(p) \infty\}$. These two cusps are defined over $\mathbf{Q}$ and there is a canonical uniformizer at $\Gamma_0(p) \infty$ (resp. $\Gamma_0(p) 0$) given by $j^{-1}$ (resp. $(j \circ w_p)^{-1}$). The order of $(\Gamma_0(p) \infty) - (\Gamma_0(p) 0)$ in $J_{\Gamma_0(p)}$ is $n:=\frac{p-1}{d}$ where $d = \gcd(p-1,12)$. More precisely, $n\cdot((\Gamma_0(p) \infty) - (\Gamma_0(p) 0))$ is the divisor of the modular unit $$u(z) := \left(\frac{\Delta(pz)}{\Delta(z)}\right)^{\frac{1}{d}}$$ where $\Delta \in S_{12}(\operatorname{SL}_2(\mathbf{Z}))$ is Ramanujan’s Delta function and $z \in \mathfrak{H}$. We have an exact sequence of semi-abelian schemes over $\mathbf{Q}$: $$1 \rightarrow \mathbf{G}_m \rightarrow \mathbf{G}_m \times \mathbf{G}_m \rightarrow J_{\Gamma_0(p)}^{\#} \rightarrow J_{\Gamma_0(p)} \rightarrow 1$$ which induces (by Hilbert 90) an exact sequence of abelian groups $$1 \rightarrow \mathbf{Q}^{\times} \rightarrow \mathbf{Q}^{\times}\times \mathbf{Q}^{\times} \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{Q}) \rightarrow J_{\Gamma_0(p)}(\mathbf{Q}) \rightarrow 1 \text{ .}$$ By convention, the first copy of $\mathbf{G}_m$ in $\mathbf{G}_m \times \mathbf{G}_m$ corresponds to the cusp $\Gamma_0(p) \infty$, while the second copy corresponds to $\Gamma_0(p) 0$. Although we shall not use it, in contrast with the case of $J_{\Gamma_0(p)}(\mathbf{Q})$, the torsion subgroup of $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$ is the image of $\{\pm 1\} \times \{\pm 1\}$ [@Yamazaki Theorem 1.1.3]. ### The generalized cuspidal $1$-motive {#the-generalized-cuspidal-1-motive} In this case, we can be a little more precise both on the algebraic and analytic sides. Let $$\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$$ be the map sending $(\Gamma_0(p) \infty)-(\Gamma_0(p) 0)$ to the image of $0 \oplus (j^{-1} \text{ modulo } U_{\Gamma_0(p) \infty}, j \circ w_p \text{ modulo } U_{\Gamma_0(p) 0})$ in $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$ via (\[iso\_gen\_jac\_alg\]). Let $$\delta_{\Gamma_0(p)}^{\#, \operatorname{an}} : \mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]^0 \rightarrow J_{\Gamma_0(p)}^{\#}(\mathbf{C})$$ be the map sending $(\Gamma_0(p) \infty)-(\Gamma_0(p) 0)$ to the image of $(f \mapsto -L(f,1)) \in \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma_0(p)), \mathbf{C})$ via the Abel-Jacobi map $AJ_{\Gamma_0(p)}$ (where $L(f,1)$ is the special value at $s=1$ of the L-function of $f$). We now prove Theorem \[intro\_main\_comparison\_Gamma\_0\_case\]. Proof of (\[main\_comparison\_Gamma\_0\_case\_i\]). Note that $\mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]$ is annihilated by Mazur’s *Eisenstein ideal* $I$, generated by the Hecke operators $T_{\ell}-\ell-1$ for primes $\ell \neq p$ and by $U_p-1$. Thus, we only need to show that $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$ and $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$ are annihilated by $I$. We first consider the map $\delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}$. Let $\ell \neq p$ be a prime number, and consider the usual double coset $\Gamma_0(p) \begin{pmatrix}1 & 0\\ 0 & p \end{pmatrix} \Gamma_0(p) =\Gamma_0(p)g_{\infty} \bigcup_{i=0}^{\ell-1} \Gamma_0(p)\cdot g_i$ where $g_i = \begin{pmatrix} 1 & i \\ 0 & \ell \end{pmatrix}$ and $g_{\infty} = \begin{pmatrix} \ell & 0 \\ 0 & 1 \end{pmatrix}$. Since in our case the Hecke operators are self-dual and fix the two cusps, the action of $T_{\ell}$ on the image of a point $0 \oplus (f_c)_{c \in \mathcal{C}_{\Gamma_0(p)}} \in \operatorname{Div}^0(X_0(p),\mathcal{C}_{\Gamma_0(p)})$ in $J_{\Gamma_0(p)}^{\#}$ is the image of $$0 \oplus (\prod_{i \in \{0, ..., \ell-1, \infty\}}f_c\circ g_i)_{c \in \mathcal{C}_{\Gamma_0(p)}} \text{ .}$$ If $\ell > 2$, then we are done since $$\prod_{i \in \{0, ..., \ell-1, \infty\}}(j \circ g_i)(z) \sim e^{2i\pi \ell z}\cdot \prod_{i=0}^{\ell-1} e^{2i \pi \frac{z+i}{\ell}} = (e^{2i \pi z})^{\ell+1} \sim j(z)^{\ell+1}$$ and similarly for $j\circ w_p$. If $\ell=2$, we find $$\prod_{i \in \{0, ..., \ell-1, \infty\}}(j \circ g_i)(z) \sim -j(z)^{\ell+1}$$ and $$\prod_{i \in \{0, ..., \ell-1, \infty\}}(j\circ w_p \circ g_i)(z) \sim -(j\circ w_p)(z)^{\ell+1} \text{ .}$$ But the image of $0 \oplus (f_c)_{c \in \mathcal{C}_{\Gamma_0(p)}}$ in $J_{\Gamma_0(p)}^{\#}$ is the same as the image of $0 \oplus (\lambda \cdot f_c)_{c \in \mathcal{C}_{\Gamma_0(p)}}$ for any scalar $\lambda$. Thus, we have proved that $T_{\ell}-\ell-1$ annihilates $\delta_{\Gamma_0(p)}^{\#}$. The case $\ell=p$ is similar. We now consider the map $\delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$. Let $T$ be a Hecke operator in the Eisenstein ideal $I$. We need to show that the map $f \mapsto -L(T f, 1) \in \operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma_0(p)), \mathbf{C})$ comes from the integration of an element in $H_1(Y_0(p), \mathbf{Z})$. If $f=E$ is the unique Eisenstein series of $M_2(\Gamma_0(p))$, then $T f =0$ by definition. If $f$ is a cusp form, then there is a unique element $e$ in $H_1(Y_0(p), \mathbf{Q})^+$ (the so-called *winding element* of Mazur) such that $-L(f,1) = \int_{e} 2i\pi f(z)dz$. We know that $Te \in H_1(Y_0(p), \mathbf{Z})^+$. In particular, we have $\int_{Te} 2i\pi E(z)dz = 0$ (this is true for any cycle in $H_1(Y_0(p), \mathbf{Z})^+$ since $2i\pi E(z)dz$ is the logarithmic derivative of the modular unit $u(z)$). Thus, the map $f \mapsto -L(T f, 1)$ of $\operatorname{Hom}_{\mathbf{C}}(M_2(\Gamma_0(p)), \mathbf{C})$ coincides with the map $f \mapsto \int_{Te} 2i\pi f(z)dz$, which concludes the proof of (\[main\_comparison\_Gamma\_0\_case\_i\]) since $Te \in H_1(Y_0(p), \mathbf{Z})$. The proof of (\[main\_comparison\_Gamma\_0\_case\_ii\]) and (\[main\_comparison\_Gamma\_0\_case\_iii\]) is essentially a particular case of the proof of Theorem \[intro\_main\_thm\_comparison\], but we give the details for the convenience of the reader. Proof of (\[main\_comparison\_Gamma\_0\_case\_ii\]). By definition, $n \cdot \delta_{\Gamma_0(p)}^{\#, \operatorname{alg}}((\Gamma_0(p) \infty)-(\Gamma_0(p) 0))$ is the image of $((\frac{j^{-n}}{u})(\Gamma_0(p) \infty), (\frac{(j\circ w_p)^{n}}{u})(\Gamma_0(p) 0))$ in $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$. By the $q$-expansion product formula for $u$, we see that $(\frac{j^{-n}}{u})(\Gamma_0(p) \infty) =1$. On the other hand, it is well-known that $u \circ w_p = p^{-\frac{12}{d}} \cdot u^{-1}$ (this is noted for instance in [@deShalit p. 471]). Thus, $(\frac{(j\circ w_p)^{n}}{u})(\Gamma_0(p) 0) = (\frac{j^{n}}{u\circ w_p})(\Gamma_0(p) \infty) = p^{\frac{12}{d}}\cdot (\frac{j^{-n}}{u})^{-1}(\Gamma_0(p) \infty) = p^{\frac{12}{d}}$. Proof of (\[main\_comparison\_Gamma\_0\_case\_iii\]). It suffices to prove that $n \cdot \delta_{\Gamma_0(p)}^{\#, \operatorname{alg}} = n \cdot \delta_{\Gamma_0(p)}^{\#, \operatorname{an}}$. The element $n \cdot \delta_{\Gamma_0(p)}^{\#, \operatorname{an}}((\Gamma_0(p) \infty)-(\Gamma_0(p) 0))$ of $J_{\Gamma_0(p)}^{\#}(\mathbf{C})$ is the image of $(\lambda_{\infty}, \lambda_0) \in \mathbf{R}^{\times} \times \mathbf{R}^{\times}$. By Lemma \[main\_thm\_comparison\_lemma\_3\], we have $$\lambda_{\infty}^{n} \cdot \lambda_0^{-n} = \exp(n\cdot L(d\log(u), 1)) \text{ .}$$ We know that $d\log(u)$ is the Eisenstein series $$E = \frac{p-1}{d} + \frac{24}{d}\cdot \sum_{k=0}^{\infty} \left(\sum_{m \mid k\atop \gcd(m,p)=1} m \right)q^k \in M_2(\Gamma_0(p)) \text{ .}$$ Thus, we have $L(d\log(u), s) = \frac{24}{d}\cdot (1-p^{1-s})\zeta(s-1)\zeta(s)$, so $L(d\log(u), 1) = -\frac{12}{d}\cdot \log(p)$ and $\lambda_{\infty}^{n} \cdot \lambda_0^{-n} = p^{-n\cdot \frac{12}{d}}$. Thus, we have $\lambda_{\infty} \cdot \lambda_0^{-1} = \epsilon \cdot p^{-\frac{12}{d}}$ where $\epsilon \in \{-1,1\}$ and $\epsilon=1$ if $n$ is odd. This concludes the proof of (\[main\_comparison\_Gamma\_0\_case\_ii\]). It would be interested to remove the sign ambiguity in Theorem \[intro\_main\_comparison\_Gamma\_0\_case\] (\[main\_comparison\_Gamma\_0\_case\_iii\]). ### Comparison with de Shalit’s extended $p$-adic period pairing {#section_deShalit} We now prove Theorem \[intro\_main\_thm\_comparison\_p\_adic\]. The proof makes use of de Shalit’s explicit construction of $Q$, so we will use the results and notation of [@deShalit] (especially §1.5) without recalling them. Let $f$ be a meromorphic function on $X_0(p)$, defined over $\mathbf{Q}$, and such that $f \sim j^{-1}$ at the cusp $\Gamma_0(p) \infty$ and $f \sim j\circ w_p$ at the cusp $\Gamma_0(p) 0$. Write $$\text{div}(f) = D + (\Gamma_0(p) \infty) - (\Gamma_0(p) 0)$$ where $D = \sum_{i=1}^m n_i (P_i)$ is a divisor supported on $Y_0(p)$. Then, by definition, $\delta^{\#, \operatorname{alg}}( (\Gamma_0(p) \infty) - (\Gamma_0(p) 0))$ is the class of $-D$ in $J_{\Gamma_0(p)}^{\#}(\mathbf{Q})$. Recall that de Shalit denotes by $\Gamma$ a $p$-adic Schottky group uniformizing $X_0(p)$, $\mathfrak{H}_{\Gamma}$ the associated $p$-adic upper-half plane and $\tau : \mathfrak{H}_{\Gamma} \rightarrow X_0(p)(\mathbf{C}_p)$ the $p$-adic uniformization map. Thus, as a $\Gamma$-invariant function on $\mathfrak{H}_{\Gamma}$, we have $$f(z)= \lambda \cdot \Theta(\textbf{D}; z) \cdot \Theta(z_{\infty}^{(0)}, z_0^{(0)}; z)$$ where $\lambda \in \mathbf{C}_p^{\times}$, $\textbf{D}$ is some lift of $D$ to $\mathfrak{H}_{\Gamma}$ and $\Theta(\textbf{D}; z)$, $\Theta(z_{\infty}^{(0)}, z_0^{(0)}; z)$ are the theta functions defined in [@deShalit §0.2]. We can assume that $\textbf{D}$ is disjoint from $\tau^{-1}(\mathcal{C}_{\Gamma_0(p)})$. By [@deShalit §1.4], under (\[p-adic\_uniformization\_generalized\]), the class of $-D$ corresponds to the group homomorphism $\psi_D : \mathbf{Z}[S]\rightarrow K^{\times}$ given by $$\psi_D(e_i)=\frac{\Theta(\textbf{D}; z_0^{(i)})}{\Theta(\textbf{D}; z_{\infty}^{(i)})} \text{ .}$$ On the other hand, by definition [@deShalit §1.5] for all $i,j \in \{0, ..., g\}$ we have $$Q(e_j, e_i) = \lim_{z \rightarrow z_0^{(j)} \atop z' \rightarrow z_{\infty}^{(j)}} (j\circ w_p(z))^2\cdot \frac{\Theta(z',z;z_{\infty}^{(i)})}{\Theta(z',z;z_0^{(i)})}$$ where $z$ and $z'$ satisfy the constraint $\tau(z') = w_p(\tau(z))$. By [@deShalit §0.2 Properties (e)], we also have $$\begin{aligned} Q(e_j, e_i) &= \lim_{z \rightarrow z_0^{(j)} \atop z' \rightarrow z_{\infty}^{(j)}} (j\circ w_p(z))^2\cdot \frac{\Theta(z_{\infty}^{(i)},z_0^{(i)};z')}{\Theta(z_{\infty}^{(i)},z_0^{(i)};z)} \\& = \lim_{z \rightarrow z_0^{(j)} \atop z' \rightarrow z_{\infty}^{(j)}} \frac{j(z')\cdot \Theta(z_{\infty}^{(i)},z_0^{(i)};z')}{(j \circ w_p)^{-1}(z)\cdot \Theta(z_{\infty}^{(i)},z_0^{(i)};z)} \text{ .}\end{aligned}$$ On the other hand, we have by assumption: $$\begin{aligned} 1 &= \lim_{z \rightarrow z_0^{(i)} \atop z' \rightarrow z_{\infty}^{(i)}} \frac{j(z') \cdot f(z')}{(j\circ w_p)^{-1}(z)\cdot f(z)} \\& = \psi_D(e_i)^{-1}\cdot Q(e_i,e_0) \text{ .}\end{aligned}$$ Thus, we have $\psi_D(e_i) = Q(e_i,e_0)$. To conclude the proof of Theorem \[intro\_main\_thm\_comparison\_p\_adic\], it suffices to prove the following result (expected in [@deShalit §1.5], altough no proof is given). \[symmetry\_Q\] The pairing $Q$ is symmetric. Recall ([*cf.* ]{}[@deShalit §1.2 and 1.3]) that $\Gamma$ is a free group with generators denoted by $\{\alpha_1, ..., \alpha_g\}$, and that for all $i \in \{0, ..., g\}$ we have the relation $z_{\infty}^{(i)} = \alpha_i^{-1}(z_{\infty}^{(0)})$ (with the convention $\alpha_0=1$). To keep track of indices, let $z_i$ (resp. $z_i'$) be what we called $z$ (resp. $z'$) above (so $z_i$ goes to $z_0^{(i)}$ and $z_i'$ goes to $z_{\infty}^{(i)}$). As de Shalit did, we can and do assume that $z_0^{(i)}$ and $z_i$ do not depend on $i$. By the argument of [@deShalit §3.1], we have $z_i' = \alpha_i^{-1}(z_0')$. Thus, we have $z_i' = \alpha_i^{-1}\alpha_j(z_j')$ and $z_{\infty}^{(i)} = \alpha_i^{-1}\alpha_j(z_{\infty}^{(j)})$ for all $(i,j) \in \{0, ..., g\}^2$. We have: $$\begin{aligned} Q(e_i, e_j) &= \lim_{z_i \rightarrow z_{0}^{(i)} \atop z_i' \rightarrow z_{\infty}^{(i)}} ((j\circ w_p)(z_i))^2 \cdot \frac{\Theta(z_i', z_i, z_{\infty}^{(j)})}{\Theta(z_i', z_i, z_{0}^{(j)})} \\&= \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} ((j\circ w_p)(z_j))^2 \cdot \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j, \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j, z_{0}^{(i)})} \\&= \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} ((j\circ w_p)(z_j))^2 \cdot \frac{\Theta(z_j', z_j, \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(z_j', z_j, z_{0}^{(i)})} \cdot \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})} \\& = \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} ((j\circ w_p)(z_j))^2 \cdot \frac{\Theta(z_j', z_j, z_{\infty}^{(i)})}{\Theta(z_j', z_j, z_{0}^{(i)})} \cdot \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})} \cdot \frac{\Theta(z_j', z_j, \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(z_j', z_j, z_{\infty}^{(i)})} \\& = Q(e_j, e_i) \cdot \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})} \cdot \frac{\Theta(z_j', z_j, \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(z_j', z_j, z_{\infty}^{(i)})} \text{ .}\end{aligned}$$ By [@deShalit §0.2 Properties], for any $a$, $b$ $\in \mathfrak{H}_{\Gamma}$ and any $\alpha \in \Gamma$, there exists a constant $c(a,b;\alpha)$ such that for all $z \in \mathfrak{H}_{\Gamma}$ not in $\Gamma \cdot a \cup \Gamma \cdot b$, we have $\Theta(a,b;z) = c(a,b;\alpha) \cdot \Theta(a,b;\alpha z)$. Thus, we have: $$\begin{aligned} \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})} &= c(\alpha_i^{-1}\alpha_j(z_j'),z_j'; \alpha_j^{-1}\alpha_i)^{-1} \cdot \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{\infty}^{(i)})}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})}\end{aligned}$$ and $$\frac{\Theta(z_j', z_j, \alpha_j^{-1}\alpha_i(z_{\infty}^{(i)}))}{\Theta(z_j', z_j, z_{\infty}^{(i)})}=c(z_j',z_j, \alpha_j^{-1}\alpha_i)^{-1} \text{ .}$$ By construction, we have $$Q(e_i-e_j, e_i) = \frac{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{\infty}^{(i)})}{\Theta(\alpha_i^{-1}\alpha_j(z_j'), z_j', z_{0}^{(i)})} \text{ .}$$ Thus we have: $$\begin{aligned} Q(e_i,e_j) &= Q(e_j,e_i) \cdot Q(e_i-e_j,e_i)\cdot \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} c(\alpha_i^{-1}\alpha_j(z_j'),z_j'; \alpha_j^{-1}\alpha_i)^{-1} \cdot c(z_j',z_j, \alpha_j^{-1}\alpha_i)^{-1} \\& = Q(e_j,e_i) \cdot Q(e_i-e_j,e_i) \cdot \lim_{z_j \rightarrow z_{0}^{(j)} \atop z_j' \rightarrow z_{\infty}^{(j)}} c(\alpha_i^{-1}\alpha_j(z_j'), z_j, \alpha_j^{-1}\alpha_i)^{-1} \\& = Q(e_j,e_i) \cdot Q(e_i-e_j,e_i) \cdot c(z_{\infty}^{(i)}, z_0^{(i)}, \alpha_j^{-1}\alpha_i)^{-1} \text{ .}\end{aligned}$$ But by construction we have $c(z_{\infty}^{(i)}, z_0^{(i)}, \alpha_j^{-1}\alpha_i) = Q(e_i-e_j,e_i)$. Thus, we have proved that $Q(e_i,e_j)=Q(e_j,e_i)$. We conclude this paragraph by two important properties of $Q$, which follow easily from the work of de Shalit [@deShalit]. \[Q\_T\_equiv\] 1. \[Q\_T\_equiv\_i\] The pairing $Q : \mathbf{Z}[S] \times \mathbf{Z}[S] \rightarrow K^{\times}$ is $\mathbb{T}$-equivariant, [*i.e.* ]{}for all $T \in \mathbb{T}$ and $(x,y) \in \mathbf{Z}[S] \times \mathbf{Z}[S]$, we have $Q(Tx, y) = Q(x,Ty)$. 2. \[Q\_T\_equiv\_ii\] Modulo the principal units of $K^{\times}$, the pairing $Q$ takes values in $\mathbf{Q}_p^{\times}$ and is $\operatorname{Gal}(\mathbf{F}_{p^2}/\mathbf{F}_p)$-equivariant, [*i.e.* ]{}for all $h \in \operatorname{Gal}(\mathbf{F}_{p^2}/\mathbf{F}_p)$ and $(x,y) \in \mathbf{Z}[S] \times \mathbf{Z}[S]$, we have $Q(hx, hy) = h(Q(x,y))$ (modulo principal units). Proof of (\[Q\_T\_equiv\_i\]). The restriction of $Q \otimes_{\mathbf{Z}} \mathbf{Z}_{\ell}$ to $\mathbf{Z}[S]^0 \times \mathbf{Z}[S]$ (and thus to $\mathbf{Z}[S] \times \mathbf{Z}[S]^0$ by symmetry of $Q$) is known to be $\mathbb{T}$-equivariant, since it has an interpretation in termes of the generalized Jacobian $J^{\sharp}_{\Gamma_0(p)}$ ([*cf.* ]{}[@deShalit_X_1 §2.3]). It follows that for any $T \in \mathbb{T}$, the quantity $\lambda_{i,j}:=\frac{Q(Te_i, e_j)}{Q(e_i, Te_j)}$ does not depend on $(i,j) \in \{0,.., g\}^2$ (recall that we have denoted $S = \{e_0, ..., e_g\}$). We have $\lambda_{i,i}=1$ by symmetry of $Q$, so for all $(i,j) \in \{0,.., g\}^2$ we have $Q(Te_i, e_j) = Q(e_i, Te_j)$. By bilinearity of $Q$, this proves (\[Q\_T\_equiv\_i\]). Proof of (\[Q\_T\_equiv\_ii\]). The fact that $Q$ takes values in $\mathbf{Q}_p^{\times}$ modulo principal units is [@deShalit Lemma 1.7]. Let $h$ be the non-trivial element of $\operatorname{Gal}(\mathbf{F}_{p^2}/\mathbf{F}_p)$. For any $i \in \{0, ... g\}$, we have $h(e_i) = U_p(e_i)$ where $U_p$ is the Hecke operator of index $p$. Thus, $$Q(h(e_i), h(e_j)) = Q(U_p(e_i), U_p(e_j)) = Q(e_i, U_p^2(e_j)) = Q(e_i, e_j)$$ where we have used (\[Q\_T\_equiv\_i\]) and the fact that $U_p^2=1$. Since $Q$ takes values in $\mathbf{Q}_p^{\times}$ modulo principal units, we have $h(Q(e_i,e_j)) = Q(e_i,e_j)$ modulo principal units. 1. The pairing $Q$ itself should be $\mathbf{Q}_p^{\times}$-valued (and thus automatically $\operatorname{Gal}(\mathbf{F}_{p^2}/\mathbf{F}_p)$-equivariant). 2. The Hecke-equivariance property is specific to the level $\Gamma_0(p)$. Indeed, an analogue of Oesterlé’s conjecture at level $\Gamma(2) \cap \Gamma_0(p)$ (basically replacing the $j$-invariant by Legendre $\lambda$ invariant) was proved in [@Lecouturier_Betina]. It appears that $Q$ does not commute with the Hecke operator $U_2$ (although it commutes with $T_{\ell}$ if $\ell \neq 2,p$ and with $U_p$). ### Application to Galois representations {#application_Galois_representations} To conclude this paper, we give an application of our results to the construction of Galois representations. Recall that $\mathbb{T}$ is the Hecke algebra acting on $M_2(\Gamma_0(p))$, generated by the Hecke operators $T_{q}$ for primes $q \neq p$ and by $U_p$. Let $\mathbb{T}^0$ be the cuspidal Hecke algebra, acting faithfully on the cuspidal modular forms $S_2(\Gamma_0(p))$. Let $I \subset \mathbb{T}$ be the Eisenstein ideal, generated by the operators $T_{q} - q-1$ ($q\neq p$ prime) and $U_p-1$. A maximal ideal $\mathfrak{m}$ of $\mathbb{T}$ is said to be *Eisenstein* if $I \subset \mathfrak{m}$. We know that $\mathbb{T}/I = \mathbf{Z}$ and that $\mathbb{T}^0/I = \mathbf{Z}/n\mathbf{Z}$ [@Mazur_Eisenstein Proposition II.9.7] (recall that $n$ is the numerator of $\frac{p-1}{12}$). In particular, the residue characteristic $\ell$ of a maximal Eisenstein ideal divides $\frac{p-1}{12}$, and such a maximal ideal is unique. Let $\mathcal{G} = J_{\Gamma_0(p)}^{\#}(\overline{\mathbf{Q}})/\text{Im}(\delta^{\#, \operatorname{alg}})$, where $\text{Im}(\delta^{\#, \operatorname{alg}})$ is the image of $\delta^{\#, \operatorname{alg}}$. If $\ell$ is a prime, let $\mathcal{V}_{\ell}$ be the $\ell$-adic Tate module of $\mathcal{G}$, [*i.e.* ]{}$\mathcal{V}_{\ell} := \varprojlim_{n} \mathcal{G}[\ell^n]$. This is a $\mathbb{T}[\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})]$-module. Note that $\mathcal{V}_{\ell}$ is also the $\ell$-adic Tate module of $J_{\Gamma_0(p)}^{\#}(\mathbf{C})/\text{Im}(\delta^{\#, \operatorname{an}})$ if $\ell \neq 2$ and of $J_{\Gamma_0(p)}^{\#}(\overline{\mathbf{Q}}_p)/\text{Im}(\delta^{\#, p-\text{adic}})$ (for all prime $\ell$) by Theorems \[intro\_main\_comparison\_Gamma\_0\_case\] and \[intro\_main\_thm\_comparison\_p\_adic\] respectively. If $\mathfrak{m}$ is a maximal ideal of residue characteristic $\ell$ of $\mathbb{T}$, let $\mathbb{T}_{\mathfrak{m}}$ and $\mathcal{V}_{\mathfrak{m}}:=\mathcal{V}_{\ell} \otimes_{\mathbb{T} \otimes_{\mathbf{Z}} \mathbf{Z}_{\ell}} \mathbb{T}_{\mathfrak{m}}$ be the $\mathfrak{m}$-adic completion of $\mathbb{T}$ and $\mathcal{V}_{\ell}$ respectively. \[multiplicity\_one\_V\] The $\mathbb{T}_{\mathfrak{m}}$-module $\mathcal{V}_{\mathfrak{m}}$ is free of rank $2$ if and only if $J_0(p)[\mathfrak{m}]$ is free of rank $2$ over $\mathbb{T}^0/\mathfrak{m}$ (where we view abusively $\mathfrak{m}$ as an ideal of $\mathbb{T}^0$). By Mazur [@Mazur_Eisenstein Lemma II.15.1 and Corollary II.15.2], the latter assertion is always true, except possibly if $\mathfrak{m}$ is a non-Eisenstein maximal ideal of characteristic $2$ and $\mathfrak{m}$ is ordinary ([*i.e.* ]{}the image of $U_p$ in $\mathbb{T}^0/\mathfrak{m}$ is non-zero). If $\mathfrak{m}$ is not Eisenstein, then this follows from Theorem \[intro\_main\_comparison\_Gamma\_0\_case\] (\[main\_comparison\_Gamma\_0\_case\_i\]) and the fact that the $\mathbf{Z}[\mathcal{C}_{\Gamma_0(p)}]$ is annihilated by $I$. Assume that $\mathfrak{m}$ is Eisenstein, of residue characteristic $\ell$. By Theorem \[intro\_main\_thm\_comparison\_p\_adic\], $\mathcal{V}_{\ell}$ is the $\ell$-adic Tate module of the $\mathbb{T}$-module $\operatorname{Hom}(\mathbf{Z}[S], \overline{\mathbf{Q}}_p^{\times})/q(\mathbf{Z}[S])$. Since we know that $q$ is injective, we get an exact sequence of $\mathbb{T} \otimes_{\mathbf{Z}} \mathbf{Z}_{\ell}$-modules $$0 \rightarrow \operatorname{Hom}(\mathbf{Z}[S], \mathbf{Z}_{\ell}) \rightarrow \mathcal{V}_{\ell} \rightarrow \mathbf{Z}_{\ell}[S] \rightarrow 0 \text{ .}$$ We get an exact sequence of $\mathbb{T}_{\mathfrak{m}}$-modules $$\label{ordinary_exact_sequence_supersingular} 0 \rightarrow \operatorname{Hom}(\mathbf{Z}[S], \mathbf{Z}) \otimes_{\mathbb{T}} \mathbb{T}_{\mathfrak{m}} \rightarrow \mathcal{V}_{\mathfrak{m}} \rightarrow \mathbf{Z}[S] \otimes_{\mathbb{T}} \mathbb{T}_{\mathfrak{m}} \rightarrow 0 \text{ .}$$ By [@Emerton_supersingular Theorem 0.5] and [@Mazur_Eisenstein Corollary II.16.3], the $\mathbb{T}_{\mathfrak{m}}$-modules $\mathbf{Z}[S] \otimes_{\mathbb{T}} \mathbb{T}_{\mathfrak{m}}$ and $ \operatorname{Hom}(\mathbf{Z}[S], \mathbf{Z}) \otimes_{\mathbb{T}} \mathbb{T}_{\mathfrak{m}} $ are free of rank $1$. Thus, $\mathcal{V}_{\mathfrak{m}}$ is free free of rank $2$ over $\mathbb{T}_{\mathfrak{m}}$. This concludes the proof of Proposition \[multiplicity\_one\_V\] (another approach if $\ell \neq 2$ would have been to use modular symbols via Theorem \[intro\_main\_comparison\_Gamma\_0\_case\] instead of the supersingular module). \[galois\_rep\_duality\] There is a Hecke and Galois equivariant perfect $\mathbf{Z}_{\ell}$-bilinear pairing $\langle \cdot, \cdot \rangle : \mathcal{V}_{\ell} \times \mathcal{V}_{\ell} \rightarrow \mathbf{Z}_{\ell}(1)$ (the equivariance means that for any $(x,y) \in \mathcal{V}_{\ell} \times \mathcal{V}_{\ell}$, $T \in \mathbb{T}$ and $g \in \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$, we have $\langle gx, gy \rangle = \chi_{\ell}(g)\cdot \langle x, y \rangle$ and $\langle T x, y \rangle = \langle x, Ty \rangle$). It suffices indeed to prove the analogous statement for $\mathcal{V}_{\mathfrak{m}}$ for all maximal ideal $\mathfrak{m}$ of $\mathbb{T}$ containing $\ell$. If $\mathfrak{m}$ is not Eisenstein, the pairing $\mathcal{V}_{\mathfrak{m}} \times \mathcal{V}_{\mathfrak{m}} \rightarrow \mathbf{Z}_{\ell}(1)$ comes from the Weil pairing on $J_0(p)$. Assume now that $\mathfrak{m}$ is Eisenstein. Let $\mathcal{V}_{\mathfrak{m}}^* = \operatorname{Hom}_{\mathbf{Z}_{\ell}}(\mathcal{V}_{\mathfrak{m}}, \mathbf{Z}_{\ell}(1))$, with the action of $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$ given by $(g\cdot \varphi)(x) = \chi_{\ell}(g)\cdot \varphi(g^{-1}\cdot x)$ for all $\varphi \in \mathcal{V}_{\mathfrak{m}}^*$ and $x \in \mathcal{V}_{\mathfrak{m}}$. By Proposition \[multiplicity\_one\_V\], we can choose a basis $(e_1,e_2)$ of the $\mathbb{T}_{\mathfrak{m}}$-module $\mathcal{V}_{\mathfrak{m}}$. Let $(e_1^*, e_2^*)$ be the dual basis in $\mathcal{V}_{\mathfrak{m}}^*$, where $e_i^* \in \operatorname{Hom}_{\mathbf{Z}_{\ell}}(\mathbb{T}_{\mathfrak{m}}, \mathbf{Z}_{\ell})$. By [@Mazur_Eisenstein Corollary II.15.2], the $\mathbb{T}_{\mathfrak{m}}$-module $\operatorname{Hom}_{\mathbf{Z}_{\ell}}(\mathbb{T}_{\mathfrak{m}}, \mathbf{Z}_{\ell})$ is free of rank one, so that $(-e_2^*, e_1^*)$ is a basis of the $\mathbb{T}_{\mathfrak{m}}$-module $\mathcal{V}_{\mathfrak{m}}^*$. The Galois representation $ \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(\mathbb{T}_{\mathfrak{m}})$ in that basis is equal to $\rho$, so we get an isomorphism of $\mathbb{T}_{\mathfrak{m}}[\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})]$-modules $\mathcal{V}_{\mathfrak{m}}^* \simeq \mathcal{V}_{\mathfrak{m}}$. Assume from now on, and until the end of the paper, that $\ell$ is a prime dividing the numerator of $\frac{p-1}{12}$. Fix an embedding $\overline{\mathbf{Q}} \hookrightarrow \overline{\mathbf{Q}}_p$. This fixes the choice of a decomposition group $G_p$ at $p$ in $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$. We denote by $I_p \subset G_p$ the inertia subgroup. Fix a cocycle $b \in Z^1(\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}), \mathbf{Z}_\ell(1))$ whose class in $H^1(\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}), \mathbf{Z}_\ell(1))$ corresponds to the class on $p^{\frac{12}{d}}$ in $\mathbf{Q}^{\times}/(\mathbf{Q}^{\times})^{\ell}$ via Kummer theory. Let $\overline{b} : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \mathbf{F}_\ell(1)$ be the reduction of $b$ modulo $\ell$. Let $\overline{\rho} : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(\mathbf{F}_\ell)$ be given by $$\overline{\rho} = \begin{pmatrix} \overline{\chi}_{\ell} & \overline{b} \\ 0 & 1 \end{pmatrix} \text{ .}$$ Let $\overline{L}$ be the line in $\mathbf{F}_{\ell}^{2}$ spanned by the vector $(1,0)$. Note that $\overline{L}$ is the unique line fixed (pointwise) by $\overline{\rho}(I_p)$ We consider the following classical deformation problem. Let $\mathcal{C}$ be the category of local Artinian rings with residue field $\mathbf{F}_{\ell}$. Let $\operatorname{Def}: \mathcal{C} \rightarrow \underline{\text{Set}}$ be the functor such that if $A \in \mathcal{C}$, then $\operatorname{Def}(A)$ is the set of strict-equivalence classes of morphisms $\rho : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(A)$ such that the following conditions hold: 1. \[def\_condition\_reduction\] The reduction of $\rho$ modulo the maximal ideal of $A$ is $\overline{\rho}$. 2. \[def\_condition\_ramification\] The representation $\rho$ is unramified outside $p$ and $\ell$. 3. \[def\_condition\_determinant\] The determinant of $\rho$ if $\chi_p$ (where we abusively view $\chi_p$ as $A^{\times}$ valued via the ring homomorphism $\mathbf{Z}_p \rightarrow A$). 4. \[def\_condition\_finite\_flat\] The representation $\rho$ is finite flat at $\ell$ (meaning that the restriction of $\rho$ to a decomposition group $\operatorname{Gal}(\overline{\mathbf{Q}}_{\ell}/\mathbf{Q}_{\ell})$ arises from the $\overline{\mathbf{Q}}_{\ell}$-points of a finite flat group scheme over $\mathbf{Z}_{\ell}$). 5. \[def\_condition\_semistable\] There is a line $L$ in $A^2$ stable by $\rho(I_p)$. Note that $\overline{\rho}$ gives an element of $\operatorname{Def}(\mathbf{F}_{\ell})$ so our deformation problem makes sense. Since the endomorphisms of $\mathbf{F}_{\ell}^2$ commuting with $\overline{\rho}$ are the scalars, we know that $\operatorname{Def}$ is pro-representable by a local Noetherian $\mathbf{Z}_{\ell}$-algebra $R$. We now prove Theorem \[intro\_thm\_galois\_rep\]. Let $\overline{\mathcal{V}} := \mathcal{V}_{\mathfrak{m}}/\mathfrak{m}\cdot \mathcal{V}_{\mathfrak{m}}$. This is a $\mathbf{F}_{\ell}$-vector space of rank $2$, with an action of $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$. The kernel of the projection $\mathbb{T}\rightarrow \mathbb{T}^0$ is $\mathbf{Z}\cdot T_0$ for some $T_0 \in \mathbb{T}$. One can choose $T_0$ so that $T_0-n\in I$ [@Emerton_supersingular Proposition 1.8]. In particular, we have $T_0 \in \mathfrak{m}$. Let $\mathcal{V}_{\mathfrak{m}}^0 := \mathcal{V}_{\mathfrak{m}}/T_0\cdot \mathcal{V}_{\mathfrak{m}}$ and $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} := \mathcal{V}_{\mathfrak{m}}/I\cdot \mathcal{V}_{\mathfrak{m}}$; these are naturally $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$-modules. Note that $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}}$ is free of rank $2$ over $\mathbf{Z}_{\ell}$. \[lemma\_galois\_rep\] The projection map $f : \mathcal{V}_{\mathfrak{m}} \rightarrow \mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} \times \mathcal{V}_{\mathfrak{m}}^{0}$ is injective, $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$-equivariant, and its image has finite index in the fiber product $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} \times_{\overline{\mathcal{V}}} \mathcal{V}_{\mathfrak{m}}^{0}$. Furthermore, there is a $\mathbf{Z}_{\ell}$-basis of $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}}$ such that the action of $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$ is given by $\begin{pmatrix} \chi_{\ell} & b \\ 0 & 1 \end{pmatrix}$. In particular, the $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$-module $\overline{\mathcal{V}}$ is isomorphic to $\overline{\rho}$. We have $I \cdot T_0 = 0$ in $\mathbb{T}$, so $I\cdot \mathcal{V}_{\mathfrak{m}} \cap T_0\cdot \mathcal{V}_{\mathfrak{m}}$ is annihilated by $I$ and by $T_0$, and hence by $n$ since $T_0-n \in I$. Since $\mathcal{V}_{\mathfrak{m}}$ is a free $\mathbf{Z}_{\ell}$-module, we get $I\cdot \mathcal{V}_{\mathfrak{m}} \cap T_0\cdot \mathcal{V}_{\mathfrak{m}} = 0$. Thus $f$ is injective and takes values in $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} \times_{\overline{\mathcal{V}}} \mathcal{V}_{\mathfrak{m}}^{0}$ since $I+(T_0) \subset \mathfrak{m}$. Furthermore, the image of $f$ into $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} \times_{\overline{\mathcal{V}}} \mathcal{V}_{\mathfrak{m}}^{0}$ has finite index since $\operatorname{rk}_{\mathbf{Z}_{\ell}}(\mathcal{V}_{\mathfrak{m}}) = 2\cdot \operatorname{rk}_{\mathbf{Z}_{\ell}}(\mathbb{T}_{\mathfrak{m}}) = \operatorname{rk}_{\mathbf{Z}_{\ell}}( \mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}} \times_{\overline{\mathcal{V}}} \mathcal{V}_{\mathfrak{m}}^{0})$. There is a $\mathbf{Z}_{\ell}$ linear isomorphism $T_0 \cdot \mathbb{T}_{\mathfrak{m}} \xrightarrow{\sim} \mathbb{T}_{\mathfrak{m}}/I\cdot \mathbb{T}_{\mathfrak{m}}$ sending $T_0\cdot T$ to the class of $T$ modulo $I$. This induces a $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$-equivariant isomorphism $T_0\cdot \mathcal{V}_{\mathfrak{m}} \xrightarrow{\sim} \mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}}$. The facts that $T_0 \cdot J_{\Gamma_0(p)}^{\#}$ is the image of $\mathbf{G}_m\times \mathbf{G}_m$ in $J_{\Gamma_0(p)}^{\#}$, $T_0 - n \in I$ and $n\cdot \delta^{\#, \operatorname{alg}}((\infty)-(0)) = (1, p^{\frac{12}{d}}) \in \mathbf{Q}^{\times} \times \mathbf{Q}^{\times}$ imply that $T_0 \cdot \mathcal{V}_{\mathfrak{m}}$ is isomorphic to the $\ell$-adic Tate module of the $1$-motive $\mathbf{Z} \rightarrow \mathbf{G}_m$ sending $1$ to the class of $p^{\frac{12}{d}}$, which is given by $\begin{pmatrix} \chi_{\ell} & b \\ 0 & 1 \end{pmatrix}$. This concludes the proof of Lemma \[lemma\_galois\_rep\]. We now prove that $\mathcal{V}_{\mathfrak{m}}$ gives rise to a continuous homomorphism $R \rightarrow \mathbb{T}_{\mathfrak{m}}$ of local $\mathbf{Z}_{\ell}$-algebras. By Lemma \[lemma\_galois\_rep\] , there is a $\mathbb{T}_{\mathfrak{m}}$ basis of $\mathcal{V}_{\mathfrak{m}}$ giving rise to a representation $\rho : \operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q}) \rightarrow \operatorname{GL}_2(\mathbb{T}_{\mathfrak{m}})$ such that the reduction of $\rho$ modulo $\mathfrak{m}$ is $\overline{\rho}$, which is the deformation condition \[def\_condition\_reduction\]. The conditions \[def\_condition\_ramification\], \[def\_condition\_determinant\] and \[def\_condition\_finite\_flat\] follow from the analogous statement for the $\operatorname{Gal}(\overline{\mathbf{Q}}/\mathbf{Q})$-modules $\mathcal{V}_{\mathfrak{m}}^{\operatorname{Eis}}$ (by Lemma \[lemma\_galois\_rep\]) and $\mathcal{V}_{\mathfrak{m}}^{0}$ (which is well-known). Condition \[def\_condition\_semistable\] follows from (\[ordinary\_exact\_sequence\_supersingular\]). Thus, we get a canonical continuous homomorphism $u : R \rightarrow \mathbb{T}_{\mathfrak{m}}$ of local $\mathbf{Z}_{\ell}$-algebras. It remains to show that $u$ is an isomorphism if $\ell \geq 5$. This follows from [@WWE Corollary 7.1.3]: the authors construct a universal pseudo-deformation ring $R'$ with an isomorphism $v:R' \xrightarrow{\sim} \mathbb{T}_{\mathfrak{m}}$. Obviously, $v$ is the composition of a map $R' \rightarrow R$ and $u$. Since $v$ is an isomorphism, so is $u$.

ArXiv

--- abstract: 'We analytically evaluate the large deviation function in a simple model of classical particle transfer between two reservoirs. We illustrate how the asymptotic large time regime is reached starting from a special propagating initial condition. We show that the steady state fluctuation theorem holds provided that the distribution of the particle number decays faster than an exponential, implying analyticity of the generating function and a discrete spectrum for its evolution operator.' author: - Upendra Harbola - Christian Van den Broeck - Katja Lindenberg title: Large deviation function and fluctuation theorem for classical particle transport --- Introduction ============ The second law of thermodynamics has two major ingredients: the existence, in an equilibrium state, of a state function called the entropy, and the increase of the total entropy in a spontaneous transition between two equilibrium states. The discovery of the fluctuation theorem entails a double departure from this standard formulation of the second law. Entropy is defined in nonequilibrium states, and it is even defined for a single realization in a non-extensive system, where it is a stochastic quantity which can increase as well as decrease with time. Furthermore, the fluctuation theorem states (in its simplest formulation) that the fluctuating total entropy production $\Delta S_{tot}$ obeys a symmetry property: $P(\Delta S_{tot})/P(-\Delta S_{tot})=\exp{\Delta S_{tot}/k_B}$ [@ft]. This equality implies the second law inequality $\langle{\Delta S}_{tot}\rangle\geq 0$. The fluctuation theorem was first discovered in an “asymptotic version," preceding the definition of a proper stochastic entropy [@udo]. In this version, one only focuses on the entropy production in idealized reservoirs. For a heat reservoir $i$ at temperature $T^{(i)} $, the entropy change is given by $\Delta S^{(i)} =Q^{(i)} /T^{(i)} $, where $Q^{(i)} $ is the net amount of heat transferred from the system to reservoir $i$. Note that $Q^{(i)} $ and hence $\Delta S^{(i)} =Q^{(i)} /T^{(i)} $ are well defined even though the system need not be at equilibrium. Furthermore, in the case of a small system, the amount of heat $Q^{(i)} $ will differ from one realization to another, so that this is a genuine stochastic quantity. Coming back to the fluctuation theorem, one notes that the total entropy change is the sum of all the contributions: $\Delta S_{tot}=\sum_i \Delta S^{(i)} +\Delta S$, which includes the entropy change $\Delta S$ of the system. Since the entropy productions in the reservoirs are easier to monitor than the nonequilibrium stochastic system entropy $\Delta S$ (where, for simplicity, we also avoid a discussion of the possible contribution to the entropy coming from interaction terms between system and reservoirs), it is of interest to identify situations in which the latter contribution is negligible. A good candidate is the asymptotic time regime for systems operating under steady state nonequilibrium conditions. Indeed one intuitively expects that the heat (and/or particle) flows will be proportional to time, while the (stochastic) entropy of the system, being in a steady state, should remain more or less constant. Hence the latter contribution becomes negligible in the long-time limit. This was indeed proven to be the case for a class of systems with bounded energy. One can then write $$\frac{P(\sum_i\Delta S^{(i)} )}{P(-\sum_i\Delta S^{(i)} )}\sim\mbox{e}^{\sum_i\Delta S^{(i)} /k_B},$$ known as the steady-state fluctuation theorem [@RMP]. This asymptotic version of the fluctuation theorem can, however, break down when the energy of the system is unbounded [@farago]. A well documented example is that of Brownian particle in contact with one or several heat reservoirs [@vanzon; @visco; @baiesi; @puglisi; @fogedby]. The large deviation function probes exponentially unlikely events for the heat evacuated to the reservoirs. Such an event can however be the result of an exponentially unlikely initial energy of the system. In this case the entropy contribution of the system is no longer negligible and a fluctuation theorem in terms of reservoir entropies alone breaks down. Here, we study the less-documented case of particle transport. More precisely, we analyze particle transport through a system in contact with particle reservoirs [@vdb]. We evaluate analytically the large deviation function for the particle flux. We find that the asymptotic fluctuation theorem is satisfied under “natural" conditions, which are typical for the distribution of the number of particles. More precisely, whereas the Boltzmann factor allows for large energies, albeit in an exponentially unlikely way, the distribution $P(n)$ for having $n$ particles in a system typically decays faster than exponential, due to the indistinguishability property giving rise to a $n!$ contribution in the denominator. As a result the generating function $F(s)=\Sigma_n s^n P(n)$ is analytic in $s$ in the entire complex plane, and the stochastic operator describing the particle exchange has a discrete spectrum. Under this condition, we are able to prove the validity of the steady state fluctuation theorem. When the system is coupled to particle reservoirs, a cumulated particle flux develops between the system and the reservoirs. The probability distribution of this flux typically depends on time in a complicated way. One may wonder whether there exist special initial conditions for which the functional form of this distribution does not change with time. We will identify such an initial distribution for the model under consideration. Note that, from the point of view of a detector (observer), the natural initial condition is to start to count the particle flux at time zero starting from a zero value. However, as we shall see, this is an unnatural initial condition from the “system’s point of view," as correlations then develop between the observed flux and the state of the system. The paper is organized as follows. In the next section we present the particle transfer model and its basic mathematical framework. In Sec. \[GF\] we introduce the generating function method to compute statistics of the particle fluxes between system and reservoirs. In Sec. \[time\] we identify the joint distribution of system state and particle fluxes that propagates in time. In the long-time limit, it approaches an asymptotic regime characterized by a large deviation function. In Sec. \[discussion\] we verify the steady-state fluctuation relation and discuss its validity in the specific context of particle transfer. Master equation =============== We consider a system which can exchange matter (particles) with two particle baths, with chemical potentials $\mu^{(1)}$ and $\mu^{(2)}$ and temperatures $T^{(1)}$ and $T^{(2)}$, respectively. The number of particles in the system will be denoted by $n$, $ n\in \mathbb{N} $. For simplicity, we will operate in the limit of non-interacting, classical but indistinguishable particles, each of energy $\epsilon$. We furthermore assume that the exchange of particles between system and baths can be described by a Markovian jump process. Consider an instant at which the system contains $n$ particles. Let $k_{+}^{(i)}$ be the rate for a particle to jump into the system, while $n k_{-}^{(i)}$ is the rate for a particle to jump out of the system, to and from thermal bath $i$, with $i=1,2$. The corresponding master equation for the probability to have $n$-particles in the system at time $t$ then reads $$\begin{aligned} \label{eq-00} \frac{\partial}{\partial t}P(n;t) &=& k_+ P(n-1;t) + (n+1)k_- P(n+1;t) \nonumber\\ &-& (k_+ + nk_-)P(n;t),\end{aligned}$$ with $$k_{\pm}=k_\pm^{(1)}+k_\pm^{(2)}.$$ We start with a number of remarks. First, when in contact with a single bath $i$, the steady state probability solution of (\[eq-00\]) reduces to the Poissonian equilibrium distribution $P_{eq}^{(i)}(n)$, $$\begin{aligned} \label{eq-0} P_{eq}^{(i)}(n) = \frac{\rho_i ^n}{n!} \mbox{e}^{-\rho_i},\end{aligned}$$ with a corresponding average number of particles $\rho_i$ given by $$\rho_i =\frac{k_+^{(i)}}{k_-^{(i)}}.$$ We mention for further use the relation between this equilibrium density and the reservoir properties ($\beta^{(i)}=1/k_B T^{(i)}$), $$\label{0} \rho_i=g e^{-\beta^{(i)}(\epsilon-\mu^{(i)})},$$ where $g$ is the number of single particle states with energy $\epsilon$ in the system and $\mu_i$ is the chemical potential of the $i$-th bath. Since we are dealing with the classical limit, $g$ should be much larger than $n$ and the exponential in (\[0\]) is much smaller than unity. Equation (\[0\]) is the detailed balance condition for the transport between the system and the $i$-th reservoir. Second, due to the combinatorial factor $n!$, the probability distribution $P_{eq}^{(i)}(n)$ decays more quickly than exponentially for large $n$. Since there is a compelling physical reason for this factor, namely, the indistinguishability of particles, we expect this to be a genuine feature of a particle distribution function, and we will assume below a faster than exponential decay for the probability distribution even when operating under nonequilibrium conditions. Third, we mention the following exact time-dependent solution of (\[eq-00\]), namely, a propagating Poisson distribution: $$\begin{aligned} \label{eq-4} P(n;t) = \frac{(\rho(t))^n \mbox{e}^{-\rho(t)}}{n!}\end{aligned}$$ with the mean number of particles in the system at time $t$ given by $$\begin{aligned} \label{eq-5} \rho(t) = \frac{k_+}{k_-}+\left(\rho(0)-\frac{k_+}{k_-}\right)\mbox{e}^{-k_-t}.\end{aligned}$$ At steady state, $\rho_{ss}=k_+/k_-$. Hence, the particle distribution maintains a Poissonian equilibrium-like shape, even though it is in a nonequilibrium state. We next turn to our main quest, namely, the study of the fluctuation theorem. We first identify the entropy change $\Delta S_r^{(i)}$ in bath $(i)$, for a given total elapsed time $t$: $$\begin{aligned} \label{ep} \Delta S^{(i)}&=&\frac{ Q^{(i)}}{T^{(i)}}\nonumber\\ &=&-\frac{(\epsilon-\mu^{(i)}) N_i }{T^{(i)}}\nonumber\\ &=&k_B N_i \ln \frac{\rho_i}{g} \end{aligned}$$ Here $N_i$ is a register that adds (subtracts) $1$ whenever a particles crosses from heat bath $i$ to the system (from the system to heat bath $i$). Thus, $N_i$ is the [*n*et]{} number crossing from bath $i$ to the system between time $0$ and $t$ [*p*lus]{} the number that was on the register at time $t=0$ (see below). All the particles have the same energy $\epsilon$. In going from the first to the second line, we have used the conservation of energy, with the change in bath energy $-\epsilon N_i$ being equal to heat plus chemical energy, $Q^{(i)} -\mu^{(i)} N_i$. Transition to the third line is based on (\[0\]). The evaluation of the stochastic bath entropies is thus reduced to that of the number of particles $N_{i}$ ($ N_i \in \mathbb{Z}, i=1,2$) transferred from baths to the system in time $t$. Since these numbers are deterministic functions of the system dynamics, the enlarged set of variables $n,N_1,N_2$ again defines a Markov jump process, and the joint probability $P(n,N_1,N_2;t)$, to find $n$ particles in the system while having a cumulative transfer of $N_1,N_2$ particles in time $t$, evolves according to the following master equation: $$\begin{aligned} \label{eq-2} \frac{\partial}{\partial t}P(n,N_1,N_2;t) &=& k_+^{(1)} P(n-1,N_1-1,N_2;t)\nonumber\\ &+& k_+^{(2)} P(n-1,N_1,N_2-1;t) \nonumber\\ &+& (n+1)k_-^{(1)} P(n+1,N_1+1,N_2;t) \nonumber\\ &+& (n+1) k_-^{(2)} P(n+1,N_1,N_2+1;t) \nonumber\\ &-& (k_+ + nk_-)P(n,N_1,N_2;t).\end{aligned}$$ As we proceed to show, it is possible to find an exact time-propagating solution of this equation, which allows us to find the asymptotic large time properties, in particular those of the stochastic entropy. This solution furthermore illustrates how this asymptotic regime is reached in the course of time. We finally note that due to particle conservation, the following identity holds at all times: $$\begin{aligned} \label{par-cons} n(t)= N_1(t)-N_1(0)+N_2(t)-N_2(0) + n(0),\end{aligned}$$ where $n(0)$, $N_1(0)$ and $N_2(0)$ are, respectively, the number of particles in the system and the number of particles on registers $1$ and $2$ at time $t=0$. We make two observations whose significance will reveal themselves when discussing the time-propagating solution of (\[eq-2\]). First, while it is quite natural and tempting, from the observer’s point of view, to choose $N_1(0)=N_2(0)=0$, the choice of $N_1(0)$ and $N_2(0)$ is in principle free, and could even be stochastic. Second, it follows from (\[par-cons\]) that the condition $n(0)=N_1(0)+N_2(0)$ propagates in time, i.e., it implies that $n(t)=N_1(t)+N_2(t)$ holds at all times. Generating function {#GF} =================== The solution of the master equation is facilitated by switching to the following generating function: $$\begin{aligned} \label{eq-6} F_{\boldsymbol \lambda}(s;t) = \sum_{N_1,N_2=-\infty}^\infty \;\sum_{n=0}^\infty \; \mbox{e}^{\lambda_1 N_1+\lambda_2N_2}\;s^n\; P(n,N_1,N_2;t).\nonumber\\\end{aligned}$$ The parameters ${\boldsymbol \lambda}=\{\lambda_1,\lambda_2\}$ are so-called “counting parameters" that keepi track of the net number of particles transferred between the system and the corresponding reservoirs. We shall use $\boldsymbol \lambda$ to denote a dependence on $\lambda_1$ and $\lambda_2$. By combination with (\[eq-2\]) we find $$\begin{aligned} \label{eq-7} \frac{\partial}{\partial t} F_{\boldsymbol \lambda}(s;t) &=& {\cal L}_{\boldsymbol \lambda}(s)F_{\boldsymbol \lambda}(s,0),\end{aligned}$$ where the operator ${\cal L}$ is defined as, $$\begin{aligned} \label{eq-8} {\cal L}_{\boldsymbol \lambda}(s) = k_- \alpha_{\boldsymbol \lambda} s + k_- \beta_{\boldsymbol \lambda} \frac{\partial}{\partial s} - k_- s\frac{\partial}{\partial s} -k_+\end{aligned}$$ with $$\begin{aligned} \label{a-b} \alpha_{\boldsymbol \lambda} &=& \frac{k_+^{(1)}\mbox{e}^{\lambda_1}+k_+^{(2)}\mbox{e}^{\lambda_2}}{k_-},\\ \beta_{\boldsymbol \lambda} &=& \frac{k_-^{(1)}\mbox{e}^{-\lambda_1}+k_-^{(2)}\mbox{e}^{-\lambda_2}}{k_-}. \end{aligned}$$ Since the dependence on the counting parameters in Eq. (\[eq-7\]) is parametric, it suffices to evaluate the eigenvectors and eigenvalues of ${\cal L}$ as an operator with respect to the variable $s$. Let $\Psi_{\boldsymbol \lambda}(s)$ be an eigenvector of ${\cal L}_{\boldsymbol \lambda}(s)$ with eigenvalue $\zeta_{\boldsymbol \lambda}$, $$\begin{aligned} \label{eq-9} {\cal L}_{\boldsymbol \lambda}(s) \Psi_{\boldsymbol \lambda}(s) = \zeta_{\boldsymbol \lambda} \Psi_{\boldsymbol \lambda}(s).\end{aligned}$$ With the expression (\[eq-8\]) for the operator, the eigenfunctions are found by straightforward integration, $$\begin{aligned} \label{eq-10} \Psi_{\boldsymbol \lambda}(s) = \left(s-\beta_{\boldsymbol \lambda}\right)^{g_{\boldsymbol \lambda}} \mbox{exp} \left\{\alpha_{\boldsymbol \lambda}\left(s-\beta_{\boldsymbol \lambda}\right)\right\},\end{aligned}$$ where $$\begin{aligned} \label{eq-10a} g_{\boldsymbol \lambda} = \alpha_{\boldsymbol \lambda}\beta_{\boldsymbol \lambda}-\frac{k_+ + \zeta_{\boldsymbol \lambda}}{k_-}.\end{aligned}$$ We now make the following crucial assumption, already mentioned in the introduction: We request that the eigenfunctions be analytic in the variable $s$ in the entire complex plane. Analyticity imposes two restrictions on the exponent $g_{\boldsymbol \lambda}$: it must be greater than or equal to zero, $g_{\boldsymbol \lambda} \geq 0$ and it must be an integer. Hence, setting $g_{\boldsymbol \lambda} =l$ with $l\in \mathbb{N}$ we find from (\[eq-10a\]) for the eigenspectrum of the operator ${\cal L}_{\boldsymbol \lambda}$ $$\begin{aligned} \label{eq-11} \zeta^{(l)} _{\boldsymbol \lambda} = k_-\alpha_{\boldsymbol \lambda}\beta_{\boldsymbol \lambda}-k_+ - lk_-,\;\;\; l\in \mathbb{N}.\end{aligned}$$ The corresponding eigenfunction can now be written in the following compact way: $$\begin{aligned} \label{eq-11a} \Psi_{\boldsymbol \lambda}^{(l)}(s) = \frac{\partial^l} {\partial \alpha^l_{\boldsymbol \lambda}} \mbox{exp} \left\{\alpha_{\boldsymbol \lambda}\left(s-\beta_{\boldsymbol \lambda}\right)\right\}.\end{aligned}$$ At this point, we make two observations. First, the eigenvalue $\zeta^{(l)}_{\boldsymbol \lambda} $ depends on ${\boldsymbol \lambda}$ only via $\lambda=\lambda_1-\lambda_2$ and is invariant under the following interchanges: $$\begin{aligned} \label{eq-12} \lambda \leftrightarrow{-\lambda-\ln \frac{\rho_1}{\rho_2}}\\ \lambda_i\leftrightarrow-\lambda_i-\ln (\rho_i)\;\;\;\mbox{for}\;i=1 \;\mbox{and}\; 2.\end{aligned}$$ The eigenfunctions themselves, however, do not obey this symmetry. This property will be crucial to verify the steady state fluctuation theorem. Second, the above set of eigenfunctions is complete in the sense that any analytic function of $s$ can be expanded in terms of this basis. Furthermore, the expansion is unique as it corresponds to a Taylor expansion around the point $\beta_{\bf \lambda}$. Hence we obtain the following explicit expression for the generating function (assumed to be analytic in $s$) obeying (\[eq-7\]): $$\begin{aligned} \label{eq-17} F_{\boldsymbol \lambda}(s;t) = \sum_{l=0}^\infty a^{(l)}_{\boldsymbol\lambda}\; \Psi_{\boldsymbol \lambda}^{(l)}(s)\; \mbox{e}^{t\zeta^{(l)}_{\boldsymbol \lambda}},\end{aligned}$$ where $a^{(l)}_{\boldsymbol\lambda}$ is expansion coefficient of the eigenfunction $\Psi_{\boldsymbol \lambda}^{(l)}(s)$ for the initial function $F_{\boldsymbol \lambda}(s,0)$. In the sequel, we will focus on a particular simple initial condition, corresponding to $a^{(l)}_{\boldsymbol\lambda}=\delta^K_{l,0} a^{(0)}_{\boldsymbol\lambda}$, where $\delta^{K}$ is the Kronecker delta. One reason is obvious: the corresponding eigenfunction is dominating the long-time limit, since it has the lowest eigenvalue, cf. (\[eq-11\]). The other reason is that such an initial condition corresponds, for an appropriate choice of the coefficient $a^{(0)}_{\boldsymbol\lambda}$, to a genuine probability distribution. The explicit expression for $\Psi_{\boldsymbol \lambda}^{(0)}(s)$ suggests the following choice for $a^{(0)}_{\boldsymbol\lambda}$: $$\label{a0} a^{(0)}_{\boldsymbol\lambda}=\mbox{e}^{{\alpha}_{\boldsymbol\lambda}\beta_{\boldsymbol\lambda}-k_+/k_-},$$ where the subtraction of $k_+/k_-$ guarantees normalization. Referring to the appendix for the calculation of the inverse, we find that it leads to the following initial probability distribution: $$\begin{aligned} \label{distribution} P(n,N_1,N_2;t=0)&=& \frac{\mbox{e}^{-\frac{k_+}{k_-}}}{N_1!N_2!} \left(\frac{k_+^{(1)}}{k_-}\right)^{N_1} \left(\frac{k_+^{(2)}}{k_-}\right)^{N_2}\nonumber\\ &\times&\delta^K_{n,N_1+N_2}\Theta(N_1)\Theta(N_2),\end{aligned}$$ where $\Theta(x)$ is a Heaviside theta-function. The Kronecker delta in (\[distribution\]) imposes the condition that initially $n(0)=N_1(0)+N_2(0)$. This condition was to be expected since, as mentioned earlier, it propagates in time with $n=N_1+N_2$ at all times. The corresponding reduced distribution for the number of particles in the system, $P^{st}(n)$, is, as expected, the steady state distribution, cf. (\[eq-4\]), $$\begin{aligned} \label{eq-4eq} P^{st}(n) \equiv P(n;t=0) = \frac{\rho^n \mbox{e}^{-\rho}}{n!},\end{aligned}$$ with $\rho=k_+/k_-$. The reduced distribution for the initial cumulated particle transfer $P(N_1,N_2;t=0)$, see also Fig. \[fig-1\], is obtained by summation of (\[distribution\]) over $n$. The summation only affects the Kronecker delta; hence $P(N_1,N_2;t=0)$ is obtained from Eq. (\[distribution\]) by the replacement of the Kronecker delta $\delta^K_{n,N_1+N_2}$ with $\Theta(N_1+N_2)$, but this factor is superfluous due to the presence of $\Theta(N_1)\Theta(N_2)$. We conclude that the reduced initial distribution $P(N_1,N_2;t=0)$ is a product of two independent Poissonian distributions. One can verify that the “natural initial condition“ $P(N_1,N_2;t=0)= \delta^{K}_{N_1,0} \delta^{K}_{N_2,0}$ does not lead to a time propagating solution, that is, it cannot be expressed solely in terms of the eigenvector $\Psi_{\boldsymbol \lambda}^{(0)}(s)$. This could have been anticipated from the fact that the propagating condition $n(0)=N_1(0)+N_2(0)$ would then imply $n(0)=0$, which is incompatible with the ”propagating" steady state statistics $P^{st}(n)$ for $n$. Propagating solution and large deviation function {#time} ================================================= Our main focus is the evaluation of the joint reduced distribution, $P(N_1,N_2;t)$. Its generating function $F_{\boldsymbol \lambda}(t)$ is found by setting $s=1$ in (\[eq-17\]): $$\begin{aligned} \label{eq-17a} F_{\boldsymbol \lambda}(t) &=& \sum_{N_1,N_2=-\infty}^\infty \mbox{e}^{\lambda_1 N_1+\lambda_2N_2} P(N_1,N_2;t)\\ &=& \sum_{l=0}^{\infty} a^{(l)}_{\boldsymbol\lambda}\; \Psi_{\boldsymbol \lambda}^{(l)}(s=1)\;\mbox{e}^{t\zeta^{(l)}_{\boldsymbol \lambda}}.\label{eq-17b}\end{aligned}$$ The $l=0$ term dominates the series in Eq. (\[eq-17\]) for asymptotically long times $t \rightarrow \infty$. Alternatively, this term corresponds to the full solution at all times for the initial condition identified in the previous section. We henceforth consider this case and can thus write $$\begin{aligned} \label{eq-17b} F_{\boldsymbol \lambda}(t)= a^{(0)}_{\boldsymbol\lambda} \mbox{e}^{t\zeta^{(0)}_{{\boldsymbol \lambda}}} \Psi^{(0)}_{\boldsymbol\lambda}(s=1),\end{aligned}$$ with $a^{(0)}_{\boldsymbol\lambda} $ given by Eq. (\[a0\]). The corresponding joint probability $P(N_1,N_2;t)$ is computed by taking the inverse transform of (\[eq-17b\]): $$\begin{aligned} \label{eq-17c} P(N_1,N_2;t) &=& \oint \frac{d\lambda_1}{2\pi i} \oint \frac{d\lambda_2}{2\pi i} \mbox{e}^{-(\lambda_1 N_1+\lambda_2 N_2-t\zeta^{(0)}_{{\boldsymbol \lambda}})} \nonumber\\ &\times& a^{(0)}_{\boldsymbol\lambda} \Psi_{\boldsymbol\lambda}^{(0)}(s=1).\end{aligned}$$ Switching to the integration variable $\lambda=\lambda_1-\lambda_2$, and expanding the exponential of $\alpha_{\boldsymbol\lambda}/k_-$ which appears inside $\Psi_{\boldsymbol \lambda}^{(0)}(s=1)$, cf. Eq. (\[eq-11a\]), Eq. (\[eq-17c\]) can be rewritten as $$\begin{aligned} \label{eq-17x1} &&P(N_1,N_2;t) = \sum_{m=0}^\infty \frac{1}{m!}\frac{1}{k_-^m} \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{\lambda_1(m-N_1-N_2))}\nonumber\\ &&\times \oint \frac{d\lambda}{2\pi i} (k_+^{(1)}+k_+^{(2)}\mbox{e}^{-\lambda})^m \mbox{e}^{\lambda (N_2+t\zeta^{(0)}_\lambda)}\mbox{e}^{-k_+/k_-} .\end{aligned}$$ Since $\zeta^{(0)}_{\lambda}$ and $\alpha_\lambda \beta_\lambda$ are functions of only $\lambda$, the integral over $\lambda_1$ reduces to the Kronecker delta $\delta^K_{m,N_1+N_2}$. Expanding the exponential in $\alpha_\lambda \beta_\lambda$, the remaining integral over $\lambda$ can be performed. Following the same steps which led to (\[distribution\]), we obtain the following propagating solution for the joint distribution function: $$\begin{aligned} \label{eq-17x2} P(N_1,N_2;t) &=& \frac{\mbox{e}^{-t \frac{k_+^{(1)}k_-^{(2)}+k_+^{(2)}k_-^{(1)}}{k_-}}} {(N_1+N_2)!} \mbox{e}^{-k_+/k_-}\nonumber\\ &\times& \left(\frac{x}{2k_-^{(2)}}\right)^{N_2} \left(\frac{k_+^{(1)}}{k_-}\right)^{N_1} \nonumber\\ &\times&\sum_{m=0}^{N_1+N_2} \binom{N_1+N_2}{m} \left(\frac{k_+^{(2)}k_-^{(2)}}{k_+^{(1)}k_-^{(1)}}\right)^{m/2} \nonumber\\ &\times& I_{m-N_2}\left(xt\right)\Theta(N_1+N_2),\end{aligned}$$ where $$\begin{aligned} \label{x-eq} x=\frac{2}{k_-}\sqrt{k_+^{(1)}k_+^{(2)}k_-^{(1)}k_-^{(2)}}\end{aligned}$$ and $I_n(y)$ is the modified bessel function of the first kind of order $n$. It is straightforward to check that for $t=0$ this reduces to the product of Poissonians, cf. (\[distribution\]). The distribution function (\[eq-17x2\]) is shown in Fig. \[fig-1a\] for different values of $t$. We next focus on the large $t$ limit. Since the particle fluxes $N_1$ and $N_2$ diverge for $t \rightarrow \infty$, we introduce the fluxes per unit time $j_i=N_i/t$. In this limit, an additional simplification takes place as one finds asymptotically that the stochastic quantities $j_1$ and $j_2$ become identical (see Fig. \[fig-1a\]). Hence, the statistics in the long time limit are expressed in terms of a single flux $j_1=-j_2=j$. The corresponding probability distribution function for $j$ has the typical shape from large deviation theory, namely, $$\label{pdf-t} P(j;t) \sim \mbox{e}^{-t{\cal L}(j)}$$ with the large deviation function $$\begin{aligned} \label{eq-17f} {\cal L}(j) = -\lim_{t\to \infty}\frac{1}{t}\ln P(j;t)\end{aligned}$$ given by the following convex non-negative function: $$\begin{aligned} \label{eq-17g} {\cal L}(j) &=& \frac{k_+^{(1)}k_-^{(2)}+k_+^{(2)}k_-^{(1)}}{k_-} -\sqrt{x^2+j^2}\nonumber\\ &+&\frac{j}{2}\ln \left(\frac{\rho_2}{\rho_1}\frac{\sqrt{x^2+j^2}+j}{\sqrt{x^2+j^2}-j}\right).\end{aligned}$$ From Eq. (\[eq-17x2\]), the probability $P(N,t)$ to have $N_1=-N_2=N$ at time $t$ takes a simple form, $$\begin{aligned} \label{pnt} P(N;t) &=& \mbox{e}^{-t \frac{k_+^{(1)}k_-^{(2)}+k_+^{(2)}k_-^{(1)}}{k_-}} \left(\frac{k_+^{(1)}k_-^{(2)}}{k_+^{(2)}k_-^{(1)}}\right)^{N/2}\nonumber\\ &\times&\mbox{e}^{-k_+/k_-} I_{N}\left(xt\right).\end{aligned}$$ In Fig. \[fig-2\] we plot $-(1/t)\ln P(N,t)$ to illustrate how the asymptotic form of the large deviation function is approached in the course of time. The dots represent the exact analytical result, Eq. (\[eq-17g\]). An alternative procedure to obtain the large deviation function is to work via the scaled cumulant generating function, $$\label{cumulant-GF} G_{\boldsymbol\lambda}=\frac{1}{t}\ln F_{\boldsymbol\lambda}(t).$$ For large $t$, the generating function $F_{\boldsymbol\lambda}(t)$ in Eq. (\[eq-17b\]) can be approximated by $$\begin{aligned} \label{approx-gf} F_{\lambda}(t) = \mbox{e}^{t\frac{k_+^{(1)}k_-^{(2)}}{k_-}(\mbox{e}^\lambda-1) +t\frac{k_+^{(2)}k_-^{(1)}}{k_-}(\mbox{e}^{-\lambda}-1)},\end{aligned}$$ which on substituting in (\[cumulant-GF\]) gives $$\begin{aligned} \label{approx-cgf} G_{\lambda} = \frac{k_+^{(1)}k_-^{(2)}}{k_-}(\mbox{e}^\lambda-1) +\frac{k_+^{(2)}k_-^{(1)}}{k_-}(\mbox{e}^{-\lambda}-1).\end{aligned}$$ $G_\lambda$ is related to ${\cal L}(j)$ by a Legendre transformation, $$\begin{aligned} \label{LDF-new} {\cal L}(j) = ext_{\lambda}\left(\lambda j - G_{\lambda}\right).\end{aligned}$$ The extremum is found at $$\begin{aligned} \label{eq-18} \lambda(j) = \frac{1}{2}\ln\left(\frac{\rho_2}{\rho_1} \frac{\sqrt{j^2+x^2}+j}{\sqrt{j^2+x^2}-j}\right).\end{aligned}$$ Substituting this result in (\[LDF-new\]), we recover Eq. (\[eq-17g\]). The cumulant generating function $G_\lambda$ allows for a swift calculation of the cumulants of the current. For large times, the $n$th cumulant $\kappa_n(N_1)$ of $N_1$, the net number of particles transferred between the system and the reservoir by time $t$, is obtained from $G_\lambda$ as $$\begin{aligned} \label{cumulant-def} \kappa_n(N_1) = t \left.\frac{d^n G_{\lambda}}{d\lambda^n}\right|_{\lambda=0}.\end{aligned}$$ Thus according to (\[cumulant-def\]), all cumulants $\kappa_n (N_1)$ vary linearly with time. For particle current $j$, the average and the variance are given by: $$\begin{aligned} \label{eq-19} \kappa_1(j) &=& \frac{1}{k_-}\left(k_+^{(1)}k_-^{(2)}-k_+^{(2)}k_-^{(1)}\right),\\ \kappa_2(j) &=& \frac{1}{tk_-}\left(k_+^{(1)}k_-^{(2)}+k_+^{(2)}k_-^{(1)}\right).\end{aligned}$$ . \[fig-3\] Furthermore, all odd and even cumulants of $j$ are proportional to the first and second cumulants, respectively: $$\begin{aligned} \label{eq-21} \kappa_n(j) &=& \frac{\kappa_2(j)}{t^{n-2}},~~~ n~~~ \mbox{even}\nonumber\\ \kappa_n(j) &=& \frac{\kappa_1(j)}{t^{n-1}},~~~ n~~~ \mbox{odd}.\end{aligned}$$ Fluctuation Theorem {#discussion} =================== From (\[ep\]), we find that the total entropy production in the reservoirs (divided by $k_B$) is asymptotically given by $$\begin{aligned} \label{eq-22} \frac{\Sigma_i \Delta S^{(i)}}{k_B} &=& \Sigma_i N_i \ln \frac{\rho_i}{g} \\ &\sim & t j \ln\frac{\rho_1}{\rho_2},\end{aligned}$$ where we used the fact that asymptotically $j=j_1=-j_2$ ($N_i=t j_i$). The steady state fluctuation theorem then requires that $$\frac{P(j)}{P(-j)}\sim \mbox{e}^{t j \ln\frac{\rho_1}{\rho_2}},$$ or, more precisely, that the large deviation function, Eq. (\[eq-17g\]), obeys the symmetry relation, $$\label{eq-23} {\cal L}(j)-{\cal L}(-j)= j \ln(\rho_2/\rho_1),$$ which is easily verified by Eq. (\[eq-17g\]). This symmetry of the large deviation function implies an analogous symmetry for the cumulant generating function, $$\label{eq-24} G_\lambda =G_{-\lambda-\ln(\rho_1/\rho_2)}.$$ We recall that the analyticity of the generating function for the particle number $n$ is an essential assumption in the above derivation of the steady state fluctuation theorem. This requires that the corresponding probability $P(n)$ decay faster than an exponential in $n$. This property is verified by the steady state Poisson distribution, Eq. (\[eq-4\]), which decays (for large $n$) logarithmically faster than the exponential: $ P(n) \sim \mbox{e}^{-n\ln(n)},~~ \mbox{for}~ n\gg1 $. It is therefore quite natural to assume that the initial condition satisfies the same property, that is, that it ask decays faster than an exponential. Without this assumption, exponentially rare fluctuations in the initial particle distribution will lead to a breakdown of the fluctuation theorem. We finally mention that the large deviation function Eq. (\[eq-17g\]) is identical to that for a random walker on a line with jump rates ${k_+^{(1)}k_-^{(2)}}/{k_-}$ to the right and ${k_+^{(2)}k_-^{(1)}}/{k_-}$ to the left. The physical interpretation is clear. Since the probability distribution of the number of particles contained in the system decays faster than an exponential, the large deviation statistics is essentially described by the transfer statistics between the reservoirs only, cf. the asymptotic identity of $j_1$ and $-j_2$. It is intuitively clear that this long-time process will be identical to the asymptotic properties of a random walk. An identical result is obtained for the effusion of particles between two reservoirs connected through a small opening [@effusion]. Acknowledgments {#acknowledgments .unnumbered} =============== UH acknowledges financial support from the Indian Institute of Science, Bangalore, India. CVdB thanks the European Science Foundation through the network “Exploring the Physics of Small Devices." KL gratefully acknowledges support of the Office of Naval Research through Grant No. N00014-13-1-0205. ..... Appendix {#appendix .unnumbered} ======== Derivation of Eq. (\[distribution\]) {#derivation-of-eq.-distribution .unnumbered} ===================================== In order to compute the initial probability distribution for which only the term with $l=0$ in the series (\[eq-17\]) survives, we need to inverse transform $a^{(0)}_{\boldsymbol\lambda}\Psi_\lambda^{(0)}(s)$, where we choose $a^{(0)}_{\boldsymbol\lambda}= \mbox{e}^{{\alpha}_{\boldsymbol\lambda}\beta_{\boldsymbol\lambda}-k_+/k_-}$. Note that for $\boldsymbol\lambda=0$, $a^{(0)}_{\boldsymbol\lambda}=1$ to preserve the normalization of the probability distribution. This is not the only possible choice for $a^{(0)}_{\boldsymbol\lambda}$, however this choice leads to a simple natural initial Poissoinian distribution, cf. Eq. (\[distribution\]), which propagates in time. Thus we have, $$\begin{aligned} \label{app-1} P(n,N_1,N_2;t=0) &=& \oint \frac{ds}{2\pi i} s^{n-1} \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{-\lambda_1N_1} \nonumber\\ && \oint \frac{d\lambda_2}{2\pi i} \mbox{e}^{-\lambda_2N_2} a^{(0)}_{\boldsymbol\lambda} \Psi_{\boldsymbol \lambda}^{(0)}(s).\end{aligned}$$ Substituting for $\Psi_{\boldsymbol\lambda}^{(0)}(s)$ from Eq. (\[eq-11a\]), we get $$\begin{aligned} \label{app-2} P(n,N_1,N_2;t=0) &=& \oint \frac{ds}{2\pi i}\frac{1}{s^{n+1}} \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{-\lambda_1N_1} \nonumber\\ &&\oint \frac{d\lambda_2}{2\pi i} \mbox{e}^{-\lambda_2N_2} \mbox{e}^{s\alpha_{\boldsymbol\lambda}} \mbox{e}^{-k_+/k_-}\end{aligned}$$ Next we expand the exponential which contains the variable $s$. This allows us to perform the $s$-integral and gives the Kronecker delta function $\delta^K_{n,m}$. We find $$\begin{aligned} \label{app-3} P(n,N_1,N_2;t=0) &=& \frac{1}{n!} \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{-\lambda_1N_1} \oint \frac{d\lambda_1}{2\pi i} \alpha_\lambda^n\nonumber\\ && \mbox{e}^{-\lambda_2N_2} \mbox{e}^{-k_+/k_-}.\end{aligned}$$ Using $\alpha_{\boldsymbol\lambda}$ from (\[a-b\]), and expanding $\alpha_{\boldsymbol\lambda}^n$ using binomial expansion, we obtain $$\begin{aligned} \label{app-3} P(n,N_1,N_2;t=0) &=& \frac{\mbox{e}^{-\frac{k_+}{k_-}}}{n!} \left(\frac{k_+^{(1)}}{k_-}\right)^n \sum_{l=0}^n \binom{n}{l} \left(\frac{k_+^{(2)}}{k_+^{(1)}}\right)^l \nonumber\\ &\times& \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{\lambda_1(n-l-N_1)} \oint \frac{d\lambda_1}{2\pi i} \mbox{e}^{\lambda_2(l-N_2)}.\nonumber\\\end{aligned}$$ The integrals over $\lambda_1$ and $\lambda_2$ give Kronecker deltas, $\delta^{K}_{n,N_1+N_2}$ and $\delta^{K}_{l,N_2}$, respectively. Using these in (\[app-3\]), we get (for $n=N_1+N_2$) Eq. (\[distribution\]). An alternative method to obtain the probability distribution is to expand the generating function $F_{\boldsymbol\lambda}$ and compare it term-by-term with the definition (\[eq-6\]). Here we present this method to recover Eq. (\[distribution\]). At $t=0$, the generating function $F_{\boldsymbol\lambda}$ is (keeping only the $l=0$ term) $$\begin{aligned} \label{gf-n1} F_{\boldsymbol\lambda}(s;0) &=& a^{(0)}_{\boldsymbol\lambda} \Psi_{\boldsymbol \lambda}^{(0)}(s).\end{aligned}$$ Substituting for $a^{(0)}_{\boldsymbol\lambda}$ and $\Psi_{\boldsymbol \lambda}^{(0)}(s)$, we can re-express it as $$\begin{aligned} \label{gf-n2} F_{\boldsymbol\lambda}(s;0) &=& \mbox{e}^{-\frac{k_+}{k_-}} \mbox{e}^{s\alpha_{\boldsymbol\lambda}}\nonumber\\ &=& \mbox{e}^{-\frac{k_+}{k_-}} \sum_{n=0}^\infty \frac{s^n}{n!} \alpha_{\boldsymbol\lambda}^n\nonumber\\ &=& \mbox{e}^{-\frac{k_+}{k_-}} \sum_{n=0}^\infty \frac{s^n}{n!} \left(\frac{k_+^{(1)}}{k_-}\right)^n \nonumber\\ &\times&\sum_{m=0}^n \binom{n}{m} \left(\frac{k_+^{(2)}}{k_+^{(1)}}\right)^m \mbox{e}^{(n-m)\lambda_1} \mbox{e}^{m\lambda_2}.\end{aligned}$$ Since $m,n$ are dummy variables, we can rewrite the last line as $$\begin{aligned} \label{gf-n3} F_{\boldsymbol\lambda}(s;0) &=& \mbox{e}^{-\frac{k_+}{k_-}} \sum_{n=0}^\infty \sum_{N_2=0}^n \frac{s^n}{n!} \left(\frac{k_+^{(1)}}{k_-}\right)^n \nonumber\\ &\times&\binom{n}{N_2} \left(\frac{k_+^{(2)}}{k_+^{(1)}}\right)^{N_2} \mbox{e}^{(n-N_2)\lambda_1} \mbox{e}^{N_2\lambda_2}.\end{aligned}$$ In order to put it in a convenient form which will allow an easy comparison with (\[eq-6\]), we introduce a Kronecker delta $\delta^K_{n-N_2,N_1}$. This allows us to rewrite Eq. (\[gf-n3\]) as, $$\begin{aligned} \label{gf-n4} F_{\boldsymbol\lambda}(s;0) &=& \mbox{e}^{-\frac{k_+}{k_-}} \sum_{n=0}^\infty \sum_{N_2=0}^n \sum_{N_1=-\infty}^{\infty} \frac{s^n}{n!} \left(\frac{k_+^{(1)}}{k_-}\right)^n \nonumber\\ &\times&\binom{n}{N_2} \left(\frac{k_+^{(2)}}{k_+^{(1)}}\right)^{N_2} \mbox{e}^{N_1\lambda_1} \mbox{e}^{N_2\lambda_2} \delta^K_{N_1,n-N_2}.\end{aligned}$$ Finally, rearranging the Kronecker delta and using the fact that due to the binomial coefficients all terms for $N_2<0$ and $n<N_2$ vanish, we can recast this expression as $$\begin{aligned} \label{gf-n4} F_{\boldsymbol\lambda}(s;0) &=& \mbox{e}^{-\frac{k_+}{k_-}} \sum_{n=0}^\infty \sum_{N_2=-\infty}^\infty \sum_{N_1=-\infty}^{\infty} s^n \mbox{e}^{\lambda_1N_1} \mbox{e}^{\lambda_2N_2} \nonumber\\ &\times& \frac{1}{N_1!N_2!} \left(\frac{k_+^{(1)}}{k_-}\right)^n \left(\frac{k_+^{(2)}}{k_+^{(1)}}\right)^{N_2} \delta^K_{n,N_1+N_2}.\end{aligned}$$ Comparing this with (\[eq-6\]), we recover (\[distribution\]). Similar steps can be followed to obtain the time dependent joint distribution function given in Eq. (\[eq-17x2\]). G. N. Bochkov and Y. E. Kuzovlev, Physica A [**106**]{}, 443 (1981); ibid [**480**]{} (1981); D. Evans, E. G. D. Cohen and G. P. Morris, Phys. Rev. Lett. [**71**]{}, 2401 (1993); D. J. Evans and D. J. Searles, Phys. Rev. E [**50**]{}, 1645 (1994); G. Gallavotti and E. G. D. Cohen, Phys. Rev. Lett. [**74**]{}, 2694 (1995); J. Stat. Phys. [**80**]{}, 931 (1995); C. Jarzynski, Phys. Rev. Lett. [**78**]{}, 2690 (1997); Phys. Rev. E [**56**]{}, 5018 (1997); G. E. Crooks, J. Stat. Phys. [**90**]{}, 1481 1(998); L. Lebowitz and H. Spohn, J. Stat. Phys. [**95**]{}, 333 (1999); T. Hatano and S. I. Sasa, Phys. Rev. Lett. [**86**]{}, 3463 (2001); T. Speck and U. Seifert, Europhys. Lett. [**74**]{}, 391 (2006). R. J. Harris and G. M. Schutz, J. Stat. Mech. P07020 (2007); M. Esposito and C. Van den Broeck, Phys. Rev. Lett.[**104**]{}, 090601 (2010); M. Esposito, U. Harbola, and S. Mukamel, Phys. Rev. E [**76**]{}, 031132 (2007). U. Seifert, Phys. Rev. Lett. [**95**]{}, 040602 (2005). M. Esposito, U. Harbola and S. Mukamel, Rev. Mod. Phys. [**81**]{}, 1665 (2009). J. Farago, J. Stat. Phys. [**107**]{} 781 (2002); ibid, Physica [**331**]{}, 69 (2003). R. van Zon and E. G. D. Cohen, Phys. Rev. Lett. [**91**]{}, 110601 (2003). P. Visco, J. Stat. Mech. P06006 (2006). M. Baiesi, T. Jacobs, C. Maes, and N. S. Skantzos Phys. Rev. E [**74**]{}, 021111 (2006). A. Puglisi, P. Visco, E. Trizac and F. van Wijland, Phys. Rev. E [**73**]{}, 021301 (2006). H. C. Fogedby, A. Imparato, J. Stat. Mech., P05015 (2011). C. Van den Broeck and K. Lindenberg, Phys. Rev. E [**86**]{}, 041144 (2012). B. Cleuren, C. Van den Broeck and R. Kawai, Phys. Rev. E [**74**]{}, 021117 (2006).

ArXiv

--- abstract: 'The behaviors of coupled oscillators, each of which has periodic motion with random natural frequency in the absence of coupling, are investigated. Some novel collective phenomena are revealed. At the onset of instability of the phase-locking state, simultaneous phase slips of all oscillators and quantized phase shifts in these phase slips are observed. By increasing the coupling, a bifurcation tree from high-dimensional quasiperiodicity to chaos to quasiperiodicity and periodicity is found. Different orders of phase synchronizations of chaotic oscillators and chaotic clusters play the key role for constructing this tree structure.' address: | $^a$Department of Physics and Center for Nonlinear Studies, Hong Kong Baptist University, Hong Kong\ $^b$Department of Physics, Beijing Normal University, Beijing 100875, China\ $^c$Department of Physics, University of Houston, Houston TX77204, USA author: - 'Zhigang Zheng$^{a,b}$ Gang Hu$^{a,b}$ and Bambi Hu$^{a,c}$' title: Phase slips and phase synchronization of coupled oscillators --- [2]{} The investigation of coupled oscillators has attracted constant interest for many decades \[1\]. The rich collective behaviors of these systems, such as mutual entrainment, self-synchronization, and so on, are observed in many fields, e.g., coupled laser systems, Josephson junction arrays, biological and chemical oscillators etc. \[2-10\]. In early studies, interest was focused on coupled oscillators, of which each is periodic without coupling. Recently, the investigation has been extended to coupled chaotic systems (i.e., individual systems are chaotic without coupling). Significant phenomena were found, such as phase synchronization of two mutually coupled chaotic oscillators \[11, 12\], and clustering and cluster-cluster synchronization of multiple coupled chaotic units for local \[13\] and global \[14\] couplings. In this letter, we study the following $N$ coupled oscillators with the nearest coupling, $$\label{1}\stackrel{.}{\theta }_i=\omega _i+\frac K3[\sin (\theta _{i+1}-\theta _i)+\sin (\theta _{i-1}-\theta _i)],$$ $i=1,2,...,N$, where $K$, $\theta _i$ and $\omega _i$ are the coupling strength, the angle of modulo $2\pi $ and the natural frequency of the $i$-th oscillator, respectively. Model (1) has been extensively investigated in the past several decades. Here we concentrate on the dynamical behavior of the system. In particular, we are interested in the characteristic features of the motions of individual oscillators, i.e., the microscopic motions, in the regime of desynchronization of the phase-locking state, which have not yet been well investigated by the previous works. Several novel features of this system are found. First, we find simultaneous phase slips of all oscillators at the onset of desynchronization of the phase locking state, and quantized phase shifts in these slips are observed. Moreover, we find the interesting cascade behavior of coupling-induced chaos and a nice tree structure of transitions from qausiperiodicity to chaos to qausiperiodicity and periodicity. Then rich behaviors of synchronizations between chaotic oscillators and chaotic clusters, which have attracted much attention recently for the coupled Rossler and Lorenz oscillators, can be also identified for the rather old as well as popular system (1), of which the individual oscillators have simple periodic motions without coupling. These findings greatly enlarge the application perspectives of chaos synchronization. These features of phase dynamics are expected to be observable in practical systems by experiments, such as coupled laser arrays, Josephson junction chains and coupled electrical circuits. In Eq. (1) the periodic boundary condition $\theta _{i+N}(t)=\theta _i(t)$ is applied. Without losing generality we scale $\omega _i$ such that $$\label{2}\sum\limits_{i=1}^{N} \omega _i=0.$$ It is well known that for a given $N$ and $\{\omega _i,\ i=1,...N\}$, there is a critical coupling $K=K_c$. For $K>K_c$, phase-locking can be observed, and then we have $\{\stackrel{.}{\theta }_i=0,\ i=1,...,N\}$ , and each $\theta _i$ is locked to a fixed value. For $K<K_c$, no phase-locking exists, and $\stackrel{.}{\theta }_i(t)$ are no longer zero. In \[8-10\], it is found that if we define an average frequency as $$\label{3}\stackrel{\_}{\omega }_i=\lim_{T \rightarrow \infty }% \frac 1T\int_0^T\stackrel{.}{\theta }_i(t)dt,$$ synchronization between different oscillators, in the sense of $\stackrel{\_% }{\omega }_i=\stackrel{\_}{\omega }_j$, $i\neq j$, can be observed in the region where strict phase-locking of $\stackrel{.}{\theta }_i=0$ is broken. It is interesting to investigate how the various oscillators are led to complete synchronization (phase-locking for $K>K_c$) via a sequence of bifurcations by increasing the coupling $K$ from $K=0$. In order to get a general idea about the global behavior of the system, we first measure the following two positive macroscopic quantities $R$ and $\Omega $: $$\label{4} R= \frac 1N\left| \sum \limits_{j=1}^{N} e^{i\theta_j}\right| , \quad \Omega =\lim_{T\rightarrow \infty }\frac 1T \int_0^T {\sum\limits_{i=1}^{N} }\left| \stackrel{.}{\theta }_i(t)\right| dt.$$ It is clear that $R$ is time-dependent beyond the phase-locking region. Then we further measure its time average as an order parameter: $$\label{5} \left< R \right> =\lim_{T\rightarrow \infty}\frac 1T \int_0^T R(t)dt,$$ In Figs. 1(a) and (c), we plot $\Omega $ vs. $K$ for $N=5$ and $15$, respectively. In both cases, natural frequencies are randomly chosen from a normal Gaussian distribution. The actual frequencies can be seen in Fig. 3(a) for $N=5$ and Fig. 3(b) for $N=15$ at $K=0$, respectively. These natural frequencies are used for all figures in this letter. We found, for the given natural frequencies, $K_c=5.08$ for $N=5$ and $K_c=6.22$ for $N=15$. When $K>K_c$ we have identically $\Omega =0$, and then complete synchronization (phase-locking) is justified. In Figs.1(b) and (d) we do the same as (a) and (c), respectively, with the measured quantity replaced by $<R>$. In computing Figs.1, initial conditions of $\theta_j(0)$ are randomly chosen, and then in the shaded region of Fig. 1(d) the coexistence of multiple attractors of phase-locking states is justified for $K>K_c$. In Fig.1, the quantity $T$ in Eqs. (4) and (5) is taken sufficiently long in our simulations so that the fluctuations due to finite $T$ are invisible. It is interesting to observe that several discontinuities appear in the region $K<K_c$, indicating that, apart from the apparent phase-locking transition, some additional transitions exist even before $K_c$. For the microscopic quantities, it is natural to study the velocities of various oscillators. In Figs. 2(a) and (b), we present the motions of $\stackrel{.}{\theta }_i$ vs. $t$ for $N=5$ and different $K$’s. For $K=0$, $\stackrel{.}{\theta }_i$ must be equal to the constant $\omega _i$, and for small $K$ \[see (a)\] $\stackrel{.}{\theta }_i$ varies oscillatorily around its natural frequency. As $K$ increases, the oscillations of $\stackrel{.}{\theta }_i$ become large, and the oscillation centers of all oscillators shift close to each other. An interesting feature is that near the onset of synchronization \[see (b)\], we can find simultaneous phase slips, i.e., all oscillators keep in a the phase-locking condition (OFF-state) for a long time, and then simultaneous bursts of all oscillators (ON-state) break the locking state. After a short firings all oscillators calm down again simultaneously to another phase-locking state, and then repeat the same process periodically. As $K$ gets closer to $K_c$, the length of the OFF state $\tau $ becomes longer. We find a clear scaling between $\tau $ and $K_c-K$: $$\label{6}\tau \propto (K_c-K)^{-0.5}.$$ The above features of synchronized actions of phase slips can be well understood by an intuitive explanation. Suppose various oscillators can be locked to a set of phases $\stackrel{\_}{\theta }_i(K)$ for $K\geq K_c$. Then all solutions satisfying $$\label{7}\stackrel{\_}{\theta }_{i+1}(K,{\bf m})-\stackrel{\_}{\theta }_i(K,% {\bf m})=\stackrel{\_}{\theta }_{i+1}(K)-\stackrel{\_}{\theta }_i(K)+2\pi m_i$$ with ${\bf m}=(m_1,...,m_j,...m_N)$ and $m_i$ being any integer, must also be phase-locking solutions of (1). By reducing $K$ lower than $K_c$, all the above solutions lose their stability via the saddle-node bifurcation. At $K=K_c$, there exists a heteroclinic path linking some of the above solutions, which has the lowest potential, and is attracting (the existence of such a heteroclinic path can be rigorously proven for $N=2$). For $K<K_c$, and $\left| K-K_c\right| \ll 1$, the system takes such a periodic path, which stays in the vicinity of one of the above stationary solution for a long time, and escapes away from this solution (simultaneous phase slips for all oscillators), and then approaches to the vicinity of the next stationary solution alone the heteroclinic path of $K=K_c$; this produces the periodic pulses of Fig. 2(b). The scaling property of time length of the OFF-state $\tau $ can be also computed since for the saddle-node bifurcation we have a universal form $\stackrel{.}{x}=(K_c-K)+x^2.$ The time $\tau $ for $x$ to move from $x=0$ to $x\rightarrow \infty $ reads $\tau \propto \int_{0}^{\infty} \frac{dx}{(K_c-K)+x^2}=\frac \pi {2\sqrt{K_c-K}}$; this explains our observation of the scaling law of Eq. (6). It is interesting to compute the phase shifts for various oscillators in each phase slip pulse in Figs. 2(b). Let $\Delta \theta _i$ represent the phase shift of $\theta _i$ during each pulse. From (1) and (2) we have $$\label{8}{\sum\limits_{i=1}^{N} \Delta \theta _i=0.}$$ We argue that any two adjacent fixed points in a heteroclinic path at $K=K_c$ take $m_i=0$ or $\pm 1$ in Eqs. (7), i.e., $$\label{9}\Delta \theta _{i+1}-\Delta \theta _i=0\text{ or }\pm 2\pi$$ for $i=1,...,N-1$. Considering both conditions (8) and (9), $\Delta \theta _i$ can take only quantized values $$\label{10}\Delta \theta _i=0,\pm \frac{2\pi }N,\pm \frac{4\pi }N,...,\pm \frac{ 2(N-1)\pi }N,\pm 2\pi .$$ The concrete value for each $\Delta \theta _i$ depends on the particular distribution of $\omega _i$. In Figs.2(c) and (d) the phase shifts observed in all pulses fully agree with our heuristic argument. Further considering Fig.3, the phase shifts in Fig.2 can be exactly predicted (see \[15\]). One more conclusion from the above analysis is that at the onset of instability of the phase-locking state, all velocities $\stackrel{\_}{\omega }_j$ have the scaling $\stackrel{\_}{\omega }_j\propto \sqrt{K_c-K}$, and the ratios between different $\stackrel{\_}{\omega }_j$ must be quantized to discrete rational numbers. As we further reduce $K$ to values considerably smaller than $K_c$, no more phase-locking exists, and no apparent synchronization can be observed directly for $\stackrel{.}{\theta }_j(t)$. However, some other implicit synchronization — phase synchronization — can be still found. To get a general idea, we plot $\stackrel{\_}{\omega }_i$ defined in (3) vs. $K$ for $N=5$ and $15$ in Fig. 3(a) and (b), respectively, by varying $K$ from $K=0$ to $K>K_c$. Interesting behavior of transition tree for phase synchronization is clearly shown. Though some synchronizations can be expected from the observations of frequency plateaus reported in previous papers \[8-10\], a number of characteristic features revealed in these trees are novel and interesting. Three kinds of transitions can be observed in these trees. First, if two [*adjacent*]{} oscillators (or adjacent clusters of oscillators) have close frequencies, they can be easily synchronized by increasing $K$. In this case, one always finds two branches merging to a single one (indicated by [**A**]{}). Second, if two [*non-adjacent*]{} oscillators (or two non-adjacent clusters) have close frequencies while the oscillators between them have considerably different frequencies, one can find the non-adjacent oscillators can be also synchronized to each other, i.e., non-local clusters can be formed, and these non-local clusters can quickly bring the oscillators between them to the synchronized status, and form a solidly larger synchronized cluster. In this case, the transition may be from two to one or from multiple branches to one (indicated by [**B**]{}). An oscillator, which is synchronized to a cluster for certain $K$, can be [*desynchronized*]{} from the original cluster by increasing $K$. This desynchronization always happens at an edge oscillator of a cluster, due to the competition between two neighbor clusters (indicated by [**C**]{}, see 2nd and 3rd oscillators of Fig. 3(b)). It is obvious that [**C**]{} is the inverse of [**A**]{}. The transitions of types [**B**]{} and [**C**]{} have never been realized before. The most interesting and novel finding is the nature of these synchronizations. The motions in these synchronization trees may be very different. They can be periodic, quasiperiodic and chaotic. In Fig. 4(a), we take $N=15$ and plot the largest Lyapunov exponent of the system vs. $K$. In a large interval of $K$, we find the positive Lyapunov exponent, indicating chaos. Thus, in this region phase synchronizations of chaotic oscillators are identified. Recently, the phase synchronization of coupled chaotic systems has attracted great attention \[11\]. Here the major difference between our system and the previous chaos synchronization is that our oscillators are periodic without coupling, and chaos is induced by the coupling, while in the latter case the individual systems are chaotic without coupling. Moreover, we find that these coupling-induced chaotic motions of different oscillators may have different levels of synchronizations by varying the coupling, and the cascade of synchronization forms a tree-like structure of Fig. 3. In Figs. 4(b) to (f) we plot the maps of $\stackrel{.}{\theta }_1(n)$ to $\stackrel{.}{\theta }_1(n+1)$, where $\stackrel{.}{\theta }_1(n)$ is the $\stackrel{.}{\theta }_1$ value at the time $t$ when $\theta _1(t)$ crosses the angles $2n\pi $ with $n$ being an integer. For $K>K_c$, we have fixed point solution, and the map is fixed at $\stackrel{.}{\theta }_1(n)=\stackrel{.}{\theta }_1(n+1)=0$. For $K$ slightly smaller than $K_c$ we have periodic solution represented by the finite number of dots in Fig. 4(b). The period $8$ can be easily understood from Fig. 2(d). The period of the total system is $15\tau $, where $\tau $ is the time interval between the two adjacent slips, and the change of of $\theta _1$ in $15\tau $ is $-16\pi $. This leads to the period-$8$ solution of Fig. 4(b). Two-frequency torus can be identified in the three-cluster regime \[see Fig. 4(c)\]. For very small $K$, we can find high-dimensional quasiperiodicity \[e.g., Fig. 4(f)\]. Chaos is prevailing in the $K$ region between Fig. 4(c) and 4(f) \[see Fig. 4(d) and (e), and Fig. 4(a)\]. Many periodic windows are found in the quasiperiodic and chaotic regions, which will be investigated in detail in our forthcoming extended paper. The entire variation from high-dimensional quasiperiodicity (for very small $K$) to periodic motion ($K<K_c$, $\left| K-K_c\right| \ll 1$) through various orders of chaos synchronizations can be vitally seen in Figs. 3 and 4. Starting from the high-dimensional quasiperiodicity for $K\ll 1$, by increasing $K$ various neighboring oscillators with close frequencies start to form clusters via phase synchronization, and chaos is induced near the first synchronization. Then in each cluster, different oscillators perform different chaotic motions, while having identical winding number. The winding numbers for different chaotic clusters are different. When further increasing $K$, adjacent chaotic clusters can be synchronized to form larger clusters, until two large clusters are formed when the motion becomes periodic. This tree picture of the transitions is expected to be the same for general large number of coupled non-identical oscillators, which are periodic in bare case. Since system (1) qualitatively describes practical situations in wide fields, ranging from physics, chemistry to biology, the findings in this letter are expected to be of general significance, and they can be used for understanding the mechanisms of rich collective behaviors of coupled systems. Laser arrays, Josephson junction chains and coupled electrical circuits may be ideal candidates for experimentally revealing the features explored in this letter. This work is supported in part by the Research Grant Council RGC and the Hong Kong Baptist University Faculty Research Grant FRG and in part by the National Natural Science Foundation of China.

ArXiv

--- abstract: 'We give a characterization of irreducible symplectic fourfolds which are given as Hilbert scheme of points on a K3 surface.' address: 'Yasunari Nagai, Graduate School of Mathematical Sciences, University of Tokyo, 3-8-1 Komaba, Meguro, Tokyo 153-8914, Japan' author: - Yasunari Nagai date: 'Revised 21/09/2002' title: 'A characterization of certain irreducible symplectic 4-folds' --- [^1] Introduction {#intro} ============ In the theory of the moduli problem of K3 surfaces, Kummer surfaces played a very important role. It is easy to characterize Kummer surfaces. \[kum\] Let $S$ be a K3 surface. If $S$ contains 16 disjoint $\mathbb P^1$’s $C_1,\dots ,C_{16}$ and $D=\sum C_i$ is $2$-divisible in $\operatorname{Pic}(S)$ then $S$ is isomorphic to a Kummer surface. The density of Kummer surfaces in the moduli space and this characterization enable us to derive the Global Torelli Theorem for arbitrary K3 surfaces from that for Kummer surfaces. A higher dimensional analogue of a K3 surface is an irreducible symplectic manifold. A compact Kähler manifold $X$ of dimension $2n$ is said to be *irreducible symplectic* if the following conditions are satisfied. 1. $X$ admits a symplectic form, *i.e.* there exists a $d$-closed holomorphic 2-form $\sigma$ such that $\sigma ^{\wedge n}$ is nowhere vanishing. 2. $h^0(X,\Omega ^2 _X)=1$, *i.e.* any non-zero holomorphic 2-form is the symplectic form up to constant. 3. $X$ is simply connected. An irreducible symplectic manifold is also called *hyper-Kähler* in the literature (see [@Be; @H]). It seems that the moduli behaviour of irreducible symplectic manifolds is similar to that of K3 surfaces. Although Namikawa recently found a counterexample to the Global Torelli Problem in higher dimensions [@Na], one still believes that some kind of Global Torelli Theorem should hold, but even a convincing conjectural version of it is missing for the time being. With a view towards the Global Torelli Problem for irreducible symplectic manifolds, it is important to ask for some “typical” objects in the moduli spaces in question and to give their characterization. This question in general seems to be quite hard. It is natural to restrict ourselves to a special case as our first step. The Hilbert scheme of points on a K3 surface is an example of irreducible symplectic manifold which is important and seems to be rather easy to handle, for it has a very explicit description, in particular in dimension four. \[hilb\] Let $S$ be a smooth surface, $\operatorname{Hilb}^n (S)$ the Hilbert scheme of 0-dimensional sub-schemes of length $n$ and $\operatorname{Sym}^n (S)=S^n/\mathfrak S_n$ the $n$-th symmetric product of S. Beauville [@Be] showed that the natural morphism (*Hilbert-Chow morphism*) $$F : \operatorname{Hilb}^n (S)\to \operatorname{Sym}^n (S)$$ is a crepant birational morphism and that if $S$ is a K3 surface, the Hilbert scheme $\operatorname{Hilb}^n(S)$ is an irreducible symplectic manifold of dimension $2n$. In the case $n=2$, the description of $F$ is quite easy. The singular locus $\Sigma$ of $\operatorname{Sym}^2(S)$ is isomorphic to $S$ and $\operatorname{Sym}^2(S)$ is locally of the form $\mathbb C^2\times (A_1\mbox{ surface singularity})$ along $\Sigma$. It is easy to show that $F$ is simply the blowing-up of $\operatorname{Sym}^2(S)$ along $\Sigma$. Considering the action of $\mathfrak S_2$, we have the following diagram $$\label{hilbdiagram} \vcenter{\xymatrix{ Bl_{\Delta}(S\times S) \ar[r]^{\quad \widetilde F} \ar[d] & S\times S \ar[d]\\ \operatorname{Hilb}^2(S) \ar[r]^F & \operatorname{Sym}^2(S) }}$$ where $\Delta $ is the diagonal of $S\times S$. We give the following result as an analogy of Proposition \[kum\]. \[main\] Let $X$ be a projective irreducible symplectic fourfold. Assume that there exists a birational morphism $f:X\to Y$ which contracts an irreducible divisor $E$ to a surface $S\subset Y$ such that 1. The general fibre of $f_{|E}:E\to S$ is isomorphic to $\mathbb P^1$, 2. $E$ is $2$-divisible in $\operatorname{Pic}(X)$, 3. $E^4=192$. Then, $S$ is a K3 surface and $X$ is isomorphic to the Hilbert scheme $\operatorname{Hilb}^2(S)$ of $S$. If $X$ is deformation equivalent to some $\operatorname{Hilb}^2(T)$ for a K3 surface $T$, the condition (iii) can be replaced by [(iii’)]{} $q_X(E)=-8$. where $q_X$ is the Beauville-Bogomolov form on $H^2(X,\mathbb{Z})$ (see [@Be; @H], see also Remark \[conv\]). It would be natural to pose the following question: *If $E$ is an irreducible divisor on $X$ with $q_X(E)<0$, then there exist an irreducible symplectic fourfold $X'$ birational to $X$ and a birational morphism $f:X'\to Y'$ which contracts the strict transform of $E$ on $X'$ ?* Clearly, the answer will be affirmative if every flop of symplectic 4-fold is a Mukai flop as conjectured, for the termination of flops for terminal fourfolds is already known. The next natural problem to consider is the density of the *birational (bimeromorphic) models* of Hilbert schemes made from Kummer surfaces in the connected component of the moduli space containing an irreducible symplectic fourfold which is birational to the Hilbert scheme of some K3 surface. But even this seems to be rather hard question. The rough idea to prove the theorem is to trace backward Beauville’s proof of Example \[hilb\]. It uses more or less elementary and standard techniques. It contains several ingredients. One is a numerical computation using Holomorphic Lefschetz theorem of Atiyah-Singer. Another is the decomposition theorem of Kähler manifolds with trivial first Chern class. The result of Wierzba [@W] on divisorial contractions of symplectic manifolds is also used in an essential way. The remaining part consists of geometric arguments based on the geometry of K3 surfaces. #### Notation. Through this paper we work with the following notation. Let $X$, $Y$, $f$, $E$ and $S$ be as in the theorem above. Theorems 1.4(ii) and 1.5 in [@W] imply that $E$ is a $\mathbb P^1$ bundle and $S$, which is the singular locus of $Y$, is a smooth surface with $K_S\sim 0$. Furthermore they infer that $Y$ is analytically locally isomorphic to $\mathbb C^2\times (\mbox{$A_1$ surface singularity})$ at each point of $S$. Put $D=\frac 12 E$ and take a double covering $p:\widetilde{X}\to X$ defined by $\mathscr{O}(D)$. Then $p$ is ramified at $\widetilde E\subset \widetilde X$ and $p(\widetilde E)=E$. $\widetilde X$ is smooth since $E$ is smooth and we have the following diagram $$\label{modeldiagram} \vcenter{\xymatrix{ \widetilde X \ar[r]^{\tilde f} \ar[d]_p & \widetilde Y \ar[d]^q\\ X \ar[r]^f & Y }}$$ where $\tilde f$ and $q$ is the Stein factorization of $f\circ p$. In fact $\widetilde Y$ is a smooth fourfold, $\tilde f$ is the blowing-up of $\widetilde Y$ along a smooth centre $\widetilde S$ with the exceptional divisor $\widetilde E$, and $q_{|\widetilde S}:\widetilde S\overset{\sim}{\to} S$ is an isomorphism. Note that $K_{\widetilde Y}\sim 0$, for $f$ is crepant and $q$ has no ramification divisor. The projectivity assumption in Theorem \[main\] is made to apply Wierzba’s result in our argument. If Wierzba’s description on divisorial contraction of symplectic manifolds is valid for non-projective ones, the projectivity assumption would not be necessary. Geometry of $\widetilde Y$ {#sec:2} ========================== In this section, we prove the following proposition. \[P1\] Under the assumption of Theorem \[main\] and the notation above, $\widetilde Y$ is isomorphic to a product of two K3 surfaces. Our strategy is to apply the following famous decomposition theorem to $\widetilde Y$. \[Bogomolov\] Let $Z$ be a compact Kähler manifold with $K_Z\sim 0$. Then there exists a finite étale covering $\widetilde Z\to Z$ such that $\widetilde Z$ is isomorphic to a product of varieties of following types 1. complex torus, 2. Calabi-Yau manifold *i.e.* compact Kähler manifold $W$ for which $K_W\sim 0$, $h^i(W,\mathscr O_W)=0$ for $0<i<\dim W$, and $\pi _1(W) =\{e\}$, 3. irreducible symplectic manifold. Thanks to this powerful theorem, Proposition \[P1\] is reduced to the following \[quant\] [(i)]{} $\pi _1(\widetilde Y)=\{e\}$. $h^0(\widetilde Y, \Omega ^2 _{\widetilde Y})=2$. Applying Theorem \[Bogomolov\] under the condition (i) of Proposition \[quant\], $\widetilde Y$ itself decomposes into a product of Calabi-Yau manifolds and irreducible symplectic manifolds. Since $\widetilde Y$ is of dimension $4$, a product of two K3 surfaces, a Calabi-Yau fourfold or a compact irreducible symplectic fourfold is possible. But (ii) of Proposition \[quant\] asserts that the last two cases do not happen. To compute these quantities from the condition (iii) of Theorem \[main\], we use Holomorphic Lefschetz formula of Atiyah-Singer [@A-S]. Let $M$ be a compact complex manifold, and $g$ an automorphism of finite order of $M$. The *holomorphic Lefschetz number* $L_{hol}(g)$ is defined by $$L_{hol}(g)=\sum (-1)^p \operatorname{trace}(g^*:H^p(M,\mathscr O_M)) .$$ \[hollef\] As in the notation of the definition above. Assume further that the fixed point set $M^g=\{x\in M\mid g(x)=x\}$ is smooth. Then the formula $$L_{hol}(g) =\int _{M^g} \frac{\prod _{\theta} \mathcal U^{\theta} (N_{M^g/M}(\theta)) \cdot \operatorname{td}(M^g)} {\det (1-(g_{N_{M^g/M}})^*)}$$ holds, where $\operatorname{td}(M^g)$ denotes the Todd class of $M^g$, $N_{M^g/M}$ the normal bundle, $N_{M^g/M}(\theta)\subset N_{M^g/M}$ the eigen-sub-bundle of $(g_{N_{M^g/M}})^*$ with the eigenvalue $e^{i\theta}$, and $$\mathcal U^{\theta} (x_1,x_2,\dots) =\left\{ \prod _j \left( \frac{1-e^{-x_j-i\theta}}{1-e^{-i\theta}} \right) \right\}^{-1} .$$ The general formula itself is very complicated, but in our case the formula becomes easy to handle. \[appdhollef\] Notation as in [§1]{}. Let $g$ be an involution of $\widetilde Y$ induced by $q:\widetilde Y\to Y$. Then $L_{hol}(g)\in \mathbb Z$ and we have $$L_{hol}(g)=\frac{1}{48}(c_2(T_{\widetilde S}) +3 c_2(N_{\widetilde S/\widetilde Y})).$$ In particular if $\widetilde S$ is a K3 surface, $$L_{hol}(g)=\frac 12 +\frac 1{16} c_2(N_{\widetilde S/\widetilde Y})$$ and if $S$ is an abelian surface, $$L_{hol}(g)=\frac 1{16} c_2(N_{\widetilde S/\widetilde Y}).$$ Since $g$ is an involution, eigenvalues of $g^*$ on each cohomology group must be $\pm 1$, therefore $\operatorname{trace}g^*\in \mathbb Z$, in particular $L_{hol}(g)\in \mathbb Z$. We apply Theorem \[hollef\] under $M=\widetilde Y$ and $M^g=\widetilde S$. Since $g$ produces the two dimensional locus of $A_1$ singularities $S$, we have $$(g_{N_{\widetilde S/\widetilde Y}})^* = \begin{pmatrix} -1 & 0\\ 0 & -1 \end{pmatrix}.$$ Therefore, $\det (1-(g_{N_{\widetilde S/\widetilde Y}})^*)=4$ and the only possible $\theta$ is $\theta =\pi$, *i.e.* $N_{\widetilde S/\widetilde Y}(\pi)=N_{\widetilde S/\widetilde Y}$. Since the rank of this bundle is 2, we have only to consider $\mathcal U^{\pi}$ in $2$ variables. By definition $$\begin{aligned} \mathcal U^{\pi}(x_1,x_2) &= \frac 2{1+e^{-x_1}} \cdot \frac 2{1+e^{-x_2}}\\ &=1-\frac 12(x_1+x_2)+\frac 14 x_1x_2 +O(x_1,x_2)^3 . \end{aligned}$$ This implies $$\mathcal U^{\pi}(N_{\widetilde S/\widetilde Y}(\pi)) =1-\frac 12 c_1(N_{\widetilde S/\widetilde Y}) +\frac 14 c_2(N_{\widetilde S/\widetilde Y}) .$$ Note that $\operatorname{td}(\widetilde S)=1+ \frac 1{12} c_2 (T_{\widetilde S})$ because $K_{\widetilde S} \sim 0$. Combining these, we get the formula. For the last assertion, note that $\widetilde S$ is either a K3 surface or a complex $2$-torus by the Enriques-Kodaira’s classification. For the computation of the second Chern class $c_2(N_{\widetilde S/\widetilde Y})$, we prepare an easy lemma. \[easy\] $\widetilde E^4=c_2(N_{\widetilde S/\widetilde Y})$. Note that $\tilde f$ is a blowing-up of the smooth variety $\widetilde Y$ along the smooth centre $\widetilde S$. Therefore, we can regard $\tilde h=\tilde f _{|\widetilde E}$ as $\tilde h: \widetilde E=\mathbb P(N_{\widetilde S/\widetilde X}) \to S$. We have a line bundle $L$ on $\widetilde E$ such that there is an exact sequence of bundle maps $$0\lto L\lto \tilde h^* N_{\widetilde S/\widetilde X} \lto \mathscr{O}_{\widetilde E}(1)\lto 0 .$$ By naturality, we have $c_2(\tilde h^* N_{\widetilde S/\widetilde X}) =c_2(N_{\widetilde S/\widetilde X})\cdot [F] \in H^4(\widetilde E,\mathbb{Z})$, where $F$ is a fibre of $\tilde h$. This implies $c_2(N_{\widetilde S/\widetilde X}) =c_2(\tilde h^*N_{\widetilde S/\widetilde X}) \cdot \mathscr{O}_{\widetilde E}(1) =L\cdot \mathscr{O}_{\widetilde E}(1)^2$. On the other hand, one has trivially $\tilde h^* c_1(N_{\widetilde S/\widetilde X}) =c_1(\tilde h^* N_{\widetilde S/\widetilde X}) = c_1(\mathscr O_{\widetilde E}(1))+c_1(L)$. Combining these we get $c_2(N_{\widetilde S/\widetilde X}) = \tilde h^* c_1(N_{\widetilde S/\widetilde X})\cdot c_1(\mathscr O_{\widetilde E}(1))^2 + \widetilde E^4$, for $\mathscr O_{\widetilde E}(1)=(-\widetilde E)_{|\widetilde E}$. But we have $c_1(N_{\widetilde S/\widetilde X}) =c_1(T_{\widetilde Y|\widetilde S}) - c_1(T_{\widetilde S})=0$ and therefore $c_2(N_{\widetilde S/\widetilde X})=\widetilde E^4$. \[conv\] We show the converse of Theorem \[main\] using this lemma. Let $X=\operatorname{Hilb}^2 (S)$ for some K3 surface $S$. Noting that fits into the diagram , (i,ii) of the Theorem \[main\] are evident. In the notation of we have $N_{\Delta/S\times S} \cong T_S\cong \Omega^1 _S$, for $S$ is K3. Therefore, for the exceptional divisor $\widetilde E$ of $F$, we get $\widetilde E^4=c_2(N_{\Delta/S\times S}) =c_2(\Omega ^1_S)=24$. By the projection formula and ramification, we have $$\label{EandEtild} \widetilde E^4=\left(\frac 12 p^* E\right)^4 =\frac 1{16}(p^*E)^4=\frac 18E^4 ,$$ so that $E^4=8\cdot 24=192$. \[K3\] Under the assumption and notation as in [§1]{}, $\widetilde S$ is a K3 surface. Assume the contrary, *i.e.* assume $\widetilde S$ be an abelian surface. Note that $\widetilde E^4=24$ by . Lemmas \[appdhollef\] and \[easy\] imply $L_{hol}(g) =\displaystyle \frac{24}{16}\not\in \mathbb Z$, which is a contradiction. Now is the time to prove Proposition \[quant\]. \(i) Let $\widehat E$, $\widehat{\widetilde E}$ be tubular neighbourhoods of $E$, $\widetilde E$ and set $X^{\circ}=X\backslash E$, $\widetilde X^{\circ}=\widetilde X \backslash \widetilde E$. Then we have $$\begin{aligned} X=X^{\circ}\cup \widehat E, \quad & X^{\circ}\cap \widehat E=\widehat E\backslash E,\\ \widetilde X=\widetilde X^{\circ}\cup \widehat{\widetilde E}, \quad &\widetilde X^{\circ}\cap \widehat{\widetilde E} =\widehat{\widetilde E}\backslash \widetilde E.\end{aligned}$$ Since the K3 surface $S\cong \widetilde S$ is simply connected and $E\to S$ and $\widetilde E\to \widetilde S$ are $\mathbb P^1$-bundles, the homotopy exact sequence infers $$\pi _1(E)=\pi _1\left(\widehat E\right)=\{e\},\quad \pi _1\left(\widetilde E\right) =\pi _1\left(\widehat{\widetilde E}\right)= \{e\}.$$ Note that $\pi _1(X)=\{e\}$ because $X$ is irreducible symplectic. By Van Kampen’s theorem $$\xymatrix @ur { \pi _1 (X^{\circ}) \ar[r] &\pi _1(X)\cong \{e\} \\ \pi _1\left(\widehat E\backslash E\right)\cong \mathbb Z \ar[u]_{\varphi} \ar[r] & \pi _1\left(\widehat E\right)\cong \{e\} \ar[u] }\lower38pt\hbox{,}$$ we know $\varphi$ is surjective. Consider the following commutative diagram $$\xymatrix{ \pi _1\left(\widetilde X^{\circ}\right)\ar[d]^{p_*} & \ar[l]^{\widetilde \varphi \quad} \pi _1\left(\widehat{\widetilde E}\backslash \widetilde E\right) \cong \mathbb Z \ar[d]_{p_*}\\ \pi _1(X^{\circ}) &\ar[l]^{\varphi\quad} \pi _1\left(\widehat E\backslash E\right)\cong \mathbb Z }\lower50pt\hbox{.}$$ Since $\widetilde X^{\circ}\to X^{\circ}$ is étale of degree 2, $\pi _1(\widetilde X^{\circ})\overset{p_*}{\to} \pi _1(X^{\circ})$ is injective but not surjective. It follows that $\widetilde \varphi$ is also surjective, because $\pi _1\left(\widehat{\widetilde E}\backslash \widetilde E\right) \to \pi _1\left(\widehat E\backslash E\right)$ is of index 2. Again using Van Kampen’s theorem, we get $\pi _1\left(\widetilde X\right)=\{e\}$, therefore $\pi _1 \left(\widetilde Y\right)=\{e\}$, for a birational map $\tilde f$ does not change the fundamental group. \(ii) The condition $h^0(X,\Omega ^2_X)=1$ implies $h^0(Y,\Omega ^2_Y)=1$. Therefore, $$L_{hol}(g)=1+\{1-(h^0(\widetilde Y,\Omega ^2_{\widetilde Y})-1)\} +1=4-h^0(\widetilde Y,\Omega ^2_{\widetilde Y}).$$ On the other hand, by Lemmas \[appdhollef\], \[easy\], and Corollary \[K3\], we see $$L_{hol}(g)=\frac 12+\frac{24}{16}=2 .$$ Combining these, we get $h^0(\widetilde Y,\Omega ^2_{\widetilde Y}) =2$. Conclusion of the proof of Theorem \[main\] {#sec:3} =========================================== In this section, we complete the proof of Theorem \[main\]. In the last section, we have shown that $\widetilde Y \cong T_1\times T_2$, where, $T_1$, $T_2$ are K3 surfaces. To prove Theorem \[main\], we investigate the action of $g$ on $\widetilde Y$. \[P2\] $T_1\cong T_2\cong \widetilde S$. We may assume $\widetilde S$ is the diagonal of $\widetilde Y\cong \widetilde S\times \widetilde S$. Let $p_i : \widetilde Y\to T_i\; (i=1,2)$ be the projections. To prove the proposition, it is enough to show $$\label{cl1} \dim p_1(\widetilde S)=\dim p_2(\widetilde S)=2 ,$$ for these imply $p_{i\; |\widetilde S}=:\varphi _i : \widetilde S\to T_i\; (i=1,2)$ is generically finite, but $K_{\widetilde S}\sim 0$ and $K_{T_i}\sim 0$ imply $\varphi _i$ has neither exceptional divisor nor ramification divisor, *i.e.* $\varphi _i$ is isomorphism. In the following, we show via case by case consideration. If $\dim p_1(\widetilde S)=0$, *i.e.* $p_1(\widetilde S) =\{t\}\subset T_1$, we get $\widetilde S=p_1^{-1}(t)$, therefore $N_{\widetilde S/\widetilde Y} =\mathscr{O}_{\widetilde S}^{\oplus 2}$ so that $c_2(N_{\widetilde S/\widetilde Y})=0$. This contradicts to Lemma \[easy\]. By the same argument for $p_2$ we have $\dim p_i (\widetilde S)\geqslant 1\; (i=1,2)$. If $\dim p_1(\widetilde S)=1$, then $\dim p_2 (\widetilde S)=2$. Moreover for $Z=p_1^{-1}(p_1(\widetilde S))$, we have $g(Z)=Z$. Assume $\dim p_1(\widetilde S)=\dim p_2(\widetilde S)=1$. Let $C_i=p_i(\widetilde S)\; (i=1,2)$. Then we have $$\widetilde S\subset p_1^{-1} (C_1)\cap p_2^{-1}(C_2) =C_1\times C_2.$$ Since $\widetilde S$ is irreducible, so are $C_1$ and $C_2$. But this implies that $\widetilde S$ is actually the product of these two curves, which is absurd. Therefore we may assume $\dim p_2(\widetilde S)=2$ and $\varphi _2 : \widetilde S\overset{\sim}{\to} T_2$ is an isomorphism. The second assertion of the claim is equivalent to $p_1(g(Z))=C_1$. Assume the contrary, *i.e.* $p_1(g(Z))=T_1$. Then, we have the following diagram $$\xymatrix{ Z=C_1\times T_2 \ar[r]^{\quad g} \ar@{}[d]|{\bigcup} \ar `l[d] `[dd]_{p_{1\; |Z}} [dd] & g(Z)\ar@{}[d]|{\bigcup} \ar@/^1pc/[ddr]^{\psi _1=p_{1\; |g(Z)}}\\ \widetilde S \ar[d]^{\varphi _1} \ar@{=}[r] & \widetilde S \ar[d]_{\varphi _1}\\ C_1 \ar@{=}[r] & C_1 \ar[r]^{\subset} & T_1 }\lower79pt\hbox{.}$$ Clearly $\kappa (Z)=-\infty$, where $\kappa(Z)$ the Kodaira dimension of $Z$. The sub-additivity property of Kodaira dimension (cf. [@K]) implies that irreducible components of any fibre of $\psi _1$ are rational curves. On the other hand, connected components of the general fibre of $\varphi _1 : \widetilde S\to C$ are elliptic curves so that $\psi _1$ contains an elliptic curve as its fibre, a contradiction. (End of the proof of the claim) To prove the proposition, what we have to do is to get a contradiction assuming $\dim p_1(\widetilde S)=1$ and $g(Z)=Z$, thanks to the claim. Consider the diagram $$\xymatrix@R=10pt{ & Z=C_1\times T_2 \ar[dl]_{p_1}\ar@{}[d]|{\bigcup}\ar[dr]^{p_2}\\ C_1 & \widetilde S\ar[l]_{\varphi _1} \ar[r]^{\varphi _2} & T_2 }\lower26pt\hbox{.}$$ Let $U_t= p_1^{-1}(t)\; (t\in C_1)$ and consider the normalizations $$\xymatrix{ \overline{Z}=\mathbb P^1\times T_2\ar[d]^{\bar p_1} \ar[r]^{norm.} & Z=C_1\times T_2 \ar[d]^{p_1}\\ \mathbb P^1\ar[r]^{norm.}& C_1 }\lower39pt\hbox{.}$$ The involution $g$ on $Z$ ascends to $\overline Z$. Since $\bar p_1$ is the anti-canonical map of $\overline Z$, $g$ descends to $\mathbb P^1$. This implies $\dim p_1(g(U_t))=0$ therefore $p_1(g(U_t))=\{t\}$, since $U_t\cap \widetilde S=g(U_t)\cap \widetilde S\neq \emptyset$. Then we have a family of automorphisms $\{\psi _t : U_t\to U_t\}$. There is a commutative diagram of isomorphisms $$\xymatrix{ U_t \ar[d]_{\psi _t} \ar[r]^{p_{2\; |U_t}} & T_2 \ar[d] & \widetilde S \ar[l]_{p_{2\; |\widetilde S}} \ar[d]^{\rho _t}\\ U_t \ar[r]^{p_{2\; |U_t}} & T_2 & \widetilde S \ar[l]_{p_{2\; |\widetilde S}} }\lower42pt\hbox{.}$$ Since a K3 surface has no infinitesimal automorphism, $\rho _t$ is independent of $t$. But $\psi _t$ fixes the points on $D_t=U_t\cap \widetilde S$ and $\widetilde S=\bigcup _{t\in C_1} D_t$, $\rho _t$ induces the identity on $\widetilde S$ and also on $U_t$. This contradicts $g\neq \operatorname{id}$. Finally the following proposition completes our proof of the Main Theorem, in view of Example \[hilb\]. Notation as above. The action of $g$ on $\widetilde Y=\widetilde S\times \widetilde S$ is the permutation of two components. Consider the diagram $$\xymatrix{ \widetilde S \ar[r]^{\mbox{\footnotesize diag.}\qquad } & \widetilde Y=\widetilde S \times \widetilde S \ar[d]^{p_1} \ar[r]^{\qquad p_2} & \widetilde S\\ & \widetilde S & }\lower44pt\hbox{.}$$ Let $T_t=p_1^{-1} (t)\; (t\in \widetilde S)$ and $\Gamma _i=\{t\in \widetilde S\mid \dim p_i\circ g(T_t)=1\}$. Note that $\Gamma _i$ is a locally closed set in Zariski topology. $\dim \Gamma _i\leqslant 1\quad (i=1,2)$. Assume the contrary, *i.e.* assume that $\Gamma _1\subset \widetilde S$ contains an Zariski open set. For any $t\in \widetilde S$, $$g(T_t)\overset{p_1}{\lto} C_t\subset \widetilde S$$ with $C_t$ is a rational curve containing $t$. Since $t$ sweeps over an open set of $\widetilde S$, $\{C_t\}$ is a covering family of rational curves on $\widetilde S$. This is impossible because $\widetilde S$ is a K3 surface. (End of the proof of the claim) Therefore, there exists a Zariski open set $V\subset \widetilde S$ such that $$(\dim p_1\circ g(T_t),\dim p_2\circ (T_t)) =(0,2)\mbox{, }(2,2)\mbox{ or }(2,0)$$ for any $t\in V$. In the first case, we have $g(T_t)=T_t$ and $g$ induces $j_t\in \operatorname{Aut}(T_t)\cong \operatorname{Aut}(\widetilde S)$. Since $\widetilde S$ has no infinitesimal automorphism, $j_t$ is constant with respect to $t\in \widetilde S$. But $j_t(t)=t$ for any $t\in V$ implies $j_t=\operatorname{id}$, contradiction. In the second case, $g(T_t)$ is the graph of an automorphism $f_t\in \operatorname{Aut}(\widetilde S)$. Since $g(T_t)\cap g(T_{t'})=\emptyset$ for $t\neq t'$, we have $f_t\neq f_{t'}$. This contradicts the discreteness of $\operatorname{Aut}(\widetilde S)$. Therefore only the last case can happen. This implies $g(p_1^{-1}(t))=p_2^{-1}(t)$, for $T_t\cap \widetilde S\neq \emptyset$. Exchanging the roles of $p_1$ and $p_2$, we also have $g(p_2^{-1}(t))=p_1^{-1}(t)$. Finally we get $$\xymatrix{ \{(s,t)\} \ar@{=}[r] & p_1^{-1}(s)\cap p_2^{-1}(t) \ar[d]^g\\ &p_2^{-1}(s)\cap p_1^{-1}(t) \ar@{=}[r] & \{(t,s)\} }$$ for $s,t\in V$. This shows $g$ is a permutation of $p_1$ and $p_2$. #### [**Acknowledgement**]{} The author would like to express his profound gratitude to Prof. Yujiro Kawamata, his supervisor, for comments and continuous encouragement. He also thanks Tetsushi Ito and Shunsuke Takagi for stimulating discussions, Prof. Keiji Oguiso for his comments and indicating an improvement of the proof of Proposition \[P2\]. He is also indebted to the referee for his careful reading of the manuscript and many valuable comments and suggestions. Atiyah, M.F., Singer, I., *The index of elliptic operators III*, Ann. Math. (2) [**87**]{}, 546-604 (1968) Barth, W., Peters, C., Van de Ven, A., *Compact Complex Surfaces*, Springer-Verlag (1984) Beauville, A., *Variétés Kähleriennes dont la première classe de Chern est nulle*, J. Diff. Geom. [**18**]{}, 775-782 (1983) Kawamata, Y., *Minimal models and the Kodaira dimension of algebraic fiber spaces*, J. Reine Angew. Math. [**363**]{}, 1-46 (1985) Namikawa, Y., *Counter-example to global Torelli problem for irreducible symplectic manifolds*, preprint, [e-print arXiv:math.AG/0110114]{}, to appear in Math. Ann. Huybrechts, D., *Compact hyperkähler manifolds: basic results*, Invent. math. [**135**]{}, No.1, 63-113 (1999) Wierzba, J., *Contractions of Symplectic Varieties*, preprint, [e-print arXiv:math.AG/9910130]{}, to appear in J. Alg. Geom. [^1]: *Mathematics Subject Classification* (2000): Primary 14J32, Secondary 32Q20

ArXiv

--- abstract: 'This paper presents two sufficient conditions to ensure a faithful evaluation of polynomial in IEEE-754 floating point arithmetic. Faithfulness means that the computed value is one of the two floating point neighbours of the exact result; it can be satisfied using a more accurate algorithm than the classic Horner scheme. One condition here provided is an [[*a priori*]{} ]{}bound of the polynomial condition number derived from the error analysis of the compensated Horner algorithm. The second condition is both dynamic and validated to check at the running time the faithfulness of a given evaluation. Numerical experiments illustrate the behavior of these two conditions and that associated running time over-cost is really interesting.' author: - | Philippe Langlois, Nicolas Louvet\ [Université de Perpignan Via Domitia [^1]]{}\ bibliography: - 'LaLo06.bib' title: | Faithful Polynomial Evaluation\ with Compensated Horner Algorithm --- **Keywords:** Polynomial evaluation, faithful rounding, Horner algorithm, compensated Horner algorithm, floating point arithmetic, IEEE-754 standard. Introduction {#sec:Introduction} ============ Motivation ---------- Horner’s rule is the classic algorithm when evaluating a polynomial $p(x)$. When performed in floating point arithmetic this algorithm may suffer from (catastrophic) cancellations and so yields a computed value with less exact digits than expected. The relative accuracy of the computed value ${{\,\widehat{p}}}(x)$ verifies the well known following inequality, $$\label{rel:HornerBound} \frac{|p(x) - {{\,\widehat{p}}}(x)|}{|p(x)|} \leq \alpha(n) {\operatorname{cond}}(p, x) \ {\ensuremath{{\mathbf{u}}}}.$$ In the right-hand side of this accuracy bound, ${\ensuremath{{\mathbf{u}}}}$ is the computing precision and $\alpha(n) \approx 2n$ for a polynomial of degree $n$. The condition number ${\operatorname{cond}}(p,x)$ that only depends on $x$ and on $p$ coefficients will be explicited further. The product $\alpha(n) {\operatorname{cond}}(p, x)$ may be arbitrarily larger than $1/{\ensuremath{{\mathbf{u}}}}$ when cancellations appear, [[*i.e.*]{}, ]{}when evaluating the polynomial $p$ at the $x$ entry is ill-conditioned.\ When the computing precision ${\ensuremath{{\mathbf{u}}}}$ is not sufficient to guarantee a desired accuracy, several solutions simulating a computation with more bits exist. Priest-like “double-double” algorithms are well-known and well-used solutions to simulate twice the IEEE-754 double precision [@Prie:91; @XBLAS:02]. The compensated Horner algorithm is a fast alternative to “double-double” introduced in [@grll05] — fast means that the compensated algorithm should run at least twice as fast as the “double-double” counterpart with the same output accuracy. In both cases this accuracy is improved and now verifies $$\label{rel:CompHornerBound} \frac{|p(x)-{{\,\widehat{p}}}(x)|}{|p(x)|} \leq {\ensuremath{{\mathbf{u}}}}+ \beta(n) {\operatorname{cond}}(p, x) \ {\ensuremath{{\mathbf{u}}}}^2,$$ with $\beta(n) \approx 4n^2$. This relation means that the computed value is as accurate as the result of the Horner algorithm performed in twice the working precision and then rounded to this working precision.\ This bound also tells us that such algorithms may yield a full precision accuracy for not too ill-conditioned polynomials, [[*e.g.*]{}, ]{}when $\beta(n) {\operatorname{cond}}(p, x) {\ensuremath{{\mathbf{u}}}}< 1$.\ This remark motivates this paper where we consider [*faithful polynomial evaluation*]{}. By faithful (rounding) we mean that the computed result ${{\,\widehat{p}}}(x)$ is one of the two floating point neighbours of the exact result $p(x)$. Faithful rounding is known to be an interesting property since for example it guarantees the correct sign determination of arithmetic expressions, [[*e.g.*]{}, ]{}for geometric predicates. We first provide an [[*a priori*]{} ]{}sufficient criterion on the condition number of the polynomial evaluation to ensure that the compensated Horner algorithm provides a faithful rounding of the exact evaluation ([Theorem \[prop:FaithfulRoundingCompHorner\]]{} in [Section \[sec:AprioriBound\]]{}). We also propose a validated and dynamic bound to prove at the running time that the computed evaluation is actually faithful ([Theorem \[prop:DynErrorBounds\]]{} in [Section \[sec:DynBound\]]{}). We present numerical experiments to show that the dynamic bound is sharper than the [[*a priori*]{} ]{}condition and we measure that the corresponding over-cost is reasonable ([Section \[sec:ExperimentalResults\]]{}). Notations --------- Throughout the paper, we assume a floating point arithmetic adhering to the IEEE-754 floating point standard [@IEEE:85]. We constraint all the computations to be performed in one working precision, with the “round to the nearest” rounding mode. We also assume that no overflow nor underflow occurs during the computations. Next notations are standard (see [@ASNA:02 chap. 2] for example). ${\mathbb{F}}$ is the set of all normalized floating point numbers and ${\ensuremath{{\mathbf{u}}}}$ denotes the unit roundoff, that is half the spacing between $1$ and the next representable floating point value. For IEEE-754 double precision with rounding to the nearest, we have ${\ensuremath{{\mathbf{u}}}}= 2^{-53} \approx 1.11 \cdot 10^{-16}$. We define the floating point predecessor and successor of a real number $r$ as follows, $${\operatorname{pred}}(r) = \max\{f\in{\mathbb{F}}/ f<r\} \quad \mbox{and} \quad {\operatorname{succ}}(r) = \min\{f\in{\mathbb{F}}/ r<f\}.$$ A floating point number $f$ is defined to be a faithful rounding of a real number $r$ if $${\operatorname{pred}}(f) < r < {\operatorname{succ}}(f).$$ The symbols $\oplus$, $\ominus$, $\otimes$ and $\oslash$ represent respectively the floating point addition, subtraction, multiplication and division. For more complex arithmetic expressions, ${\operatorname{fl}}(\cdot)$ denotes the result of a floating point computation where every operation inside the parenthesis is performed in the working precision. So we have for example, $a \oplus b = {\operatorname{fl}}(a+b)$. When no underflow nor overflow occurs, the following standard model describes the accuracy of every considered floating point computation. For two floating point numbers $a$ and $b$ and for $\circ$ in $\{+, -, \times, /\}$, the floating point evaluation ${\operatorname{fl}}(a \circ b)$ of $a \circ b$ is such that $$\label{rel:StdModel} {\operatorname{fl}}(a \circ b) = (a \circ b)(1+{\varepsilon}_1) = (a \circ b)/(1+{\varepsilon}_2), \mbox{with} \quad |{\varepsilon}_1|,|{\varepsilon}_2| \leq {\ensuremath{{\mathbf{u}}}}.$$ To keep track of the $(1+{\varepsilon})$ factors in next error analysis, we use the classic $(1+\theta_k)$ and $\gamma_k$ notations [@ASNA:02 chap. 3]. For any positive integer $k$, $\theta_k$ denotes a quantity bounded according to $$|\theta_k| \leq \gamma_k = \frac{k{\ensuremath{{\mathbf{u}}}}}{1-k{\ensuremath{{\mathbf{u}}}}}.$$ When using these notations, we always implicitly assume $k{\ensuremath{{\mathbf{u}}}}<1$. In further error analysis, we essentially use the following relations, $$(1+\theta_k)(1+\theta_j) \leq (1+\theta_{k+j}), \quad k{\ensuremath{{\mathbf{u}}}}\leq \gamma_k, \quad \gamma_k \leq \gamma_{k+1}.$$ Next bounds are computable floating point values that will be useful to derive dynamic validation in [Section \[sec:DynBound\]]{}. We denotes ${\operatorname{fl}}(\gamma_k) = (k {\ensuremath{{\mathbf{u}}}}) \oslash (1 \ominus k {\ensuremath{{\mathbf{u}}}})$ by ${{\,\widehat{\gamma}}}_k$. We know that ${\operatorname{fl}}(k{\ensuremath{{\mathbf{u}}}}) = k{\ensuremath{{\mathbf{u}}}}\in {\mathbb{F}}$, and $k{\ensuremath{{\mathbf{u}}}}<1$ implies ${\operatorname{fl}}(1-k{\ensuremath{{\mathbf{u}}}}) = 1-k{\ensuremath{{\mathbf{u}}}}\in {\mathbb{F}}$. So ${{\,\widehat{\gamma}}}_k$ only suffers from a rounding error in the division and $$\label{rel:GammaK} \gamma_k \leq (1+{\ensuremath{{\mathbf{u}}}}) {{\,\widehat{\gamma}}}_k.$$ The next bound comes from the direct application of Relation (\[rel:StdModel\]). For $x \in {\mathbb{F}}$ and $n \in {\mathbf{N}}$, $$\label{rel:CompBound} (1+{\ensuremath{{\mathbf{u}}}})^n|x| \leq {\operatorname{fl}}\left( \frac{|x|}{1-(n+1){\ensuremath{{\mathbf{u}}}}} \right).$$ From Horner to compensated Horner algorithm {#sec:CompHorner} =========================================== The compensated Horner algorithm improves the classic Horner iteration computing a correcting term to compensate the rounding errors the classic Horner iteration generates in floating point arithmetic. Main results about compensated Horner algorithm are summarized in this section; see [@grll05] for a complete description. Polynomial evaluation and Horner algorithm ------------------------------------------ The classic condition number of the evaluation of $p(x) = \sum_{i=0}^n a_i x^i$ at a given data $x$ is $$\label{rel:CondPoly} {\operatorname{cond}}(p,x) = \frac{\sum_{i=0}^n |a_i| |x|^i}{|\sum_{i=0}^n a_i x^i|}= \frac{{\ensuremath{\,\widetilde{p}}}(x)}{|p(x)|}.$$ For any floating point value $x$ we denote by ${\mbox{{{\sf Horner}}}\left(p, x\right)}$ the result of the floating point evaluation of the polynomial $p$ at $x$ using next classic Horner algorithm. \[algo:Horner\] Horner algorithm =function $r_0 = {\mbox{{{\sf Horner}}}\left(p,x\right)}$\ $r_n = a_n$\ for $i=n-1:-1:0$\ $r_i = r_{i+1} \otimes x \oplus a_i$\ end The accuracy of the result of [Algorithm \[algo:Horner\]]{} verifies introductory [inequality (\[rel:HornerBound\])]{} with $\alpha_n {\ensuremath{{\mathbf{u}}}}= \gamma_{2n}$ and previous condition number (\[rel:CondPoly\]). Clearly, the condition number ${\operatorname{cond}}(p, x)$ can be arbitrarily large. In particular, when ${\operatorname{cond}}(p, x) > 1/ \gamma_{2n}$, we cannot guarantee that the computed result ${\mbox{{{\sf Horner}}}\left(p,x\right)}$ contains any correct digit. We further prove that the error generated by the Horner algorithm is exactly the sum of two polynomials with floating point coefficients. The next lemma gives bounds of the generated error when evaluating this sum of polynomials applying the Horner algorithm. \[lemma:HornerSum\] Let $p$ and $q$ be two polynomials with floating point coefficients, such that $p(x) = \sum_{i=0}^n a_i x^i$ and $q(x) = \sum_{i=0}^n b_i x^i$. We consider the floating point evaluation of $(p+q)(x)$ computed with ${\mbox{{{\sf Horner}}}\left(p \oplus q, x\right)}$. Then, in case no underflow occurs, the computed result satisfies the following forward error bound, $$\label{rel:ErrorBoundHornerSum} |(p+q)(x) - {\mbox{{{\sf Horner}}}\left(p \oplus q, x\right)}| \leq \gamma_{2n+1} ({\ensuremath{\,\widetilde{p+q}}})(x).$$ Moreover, if we assume that $x$ and the coefficients of $p$ and $q$ are non-negative floating point numbers then $$\label{rel:ComputableBoundHornerSum} (p+q)(x) \leq (1+{\ensuremath{{\mathbf{u}}}})^{2n+1} {\mbox{{{\sf Horner}}}\left(p \oplus q, x\right)}.$$ The proof of the error bound (\[rel:ErrorBoundHornerSum\]) is easily adapted from the one of the Horner algorithm (see [@ASNA:02 p.95] for example). To prove (\[rel:ComputableBoundHornerSum\]) we consider [Algorithm \[algo:Horner\]]{}, where $$r_{n} = a_n \oplus b_n \quad \mbox{and} \quad r_{i} = r_{i+1} \otimes x \oplus (a_i \oplus b_i) \quad \mbox{for} \quad i=n-1,\ldots,0.$$ Next, using the standard model (\[rel:StdModel\]) it is easily proved by induction that, for $i=0,\ldots,n$, $$\label{rel:ComputableBoundHornerSum:rel1} \sum_{j=0}^{i}(a_{n-i+j}+b_{n-i+j})x^j \leq (1+{\ensuremath{{\mathbf{u}}}})^{2i+1} r_{n-i},$$ which in turn proves (\[rel:ComputableBoundHornerSum\]) for $i=n$. EFT for the elementary operations --------------------------------- Now we review well known results concerning error free transformation (EFT) of the elementary floating point operations $+$, $-$ and $\times$.\ Let $\circ$ be an operator in $\{ +, -, \times\}$, $a$ and $b$ be two floating point numbers, and ${{\,\widehat{x}}} = {\operatorname{fl}}(a \circ b)$. Then their exist a floating point value $y$ such that $$\label{rel:DefEFT} a \circ b = {{\,\widehat{x}}} + y.$$ The difference $y$ between the exact result and the computed result is the rounding error generated by the computation of ${{\,\widehat{x}}}$. Let us emphasize that relation (\[rel:DefEFT\]) between four floating point values relies on real operators and exact equality, [[*i.e.*]{}, ]{}not on approximate floating point counterparts. Ogita [*et al.*]{} [@OgRO:05] name such a transformation an error free transformation (EFT). The practical interest of the EFT comes from next Algorithms \[algo:TwoSum\] and \[algo:TwoProd\] that compute the exact error term $y$ for $\circ = +$ and $\circ = \times$.\ For the EFT of the addition we use [Algorithm \[algo:TwoSum\]]{}, the well known [[TwoSum]{}]{} algorithm by Knuth [@Knut:98] that requires 6 flop (floating point operations). For the EFT of the product, we first need to split the input arguments into two parts. It is done using [Algorithm \[algo:Split\]]{} of Dekker [@Dekk:71] where $r=27$ for IEEE-754 double precision. Next, [Algorithm \[algo:TwoProd\]]{} by Veltkamp (see [@Dekk:71]) can be used for the EFT of the product. This algorithm is commonly called [[TwoProd]{}]{} and requires 17 flop. EFT of the sum of two floating point numbers. \[algo:TwoSum\] ==function $[x, y] = {\mbox{{{\sf TwoSum}}}\left(a, b\right)}$\ $x = a \oplus b$\ $z = x \ominus a$\ $y = (a \ominus (x \ominus z)) \oplus (b \ominus z)$ Splitting of a floating point number into two parts. \[algo:Split\] ==function $[x, y] = {\mbox{{{\sf Split}}}\left(a\right)}$\ $z = a \otimes (2^r + 1)$\ $x = z \ominus (z \ominus a)$\ $y = a \ominus x$ EFT of the product of two floating point numbers. \[algo:TwoProd\] ==function $[x, y] = {\mbox{{{\sf TwoProd}}}\left(a, b\right)}$\ $x = a \otimes b$\ $[a_h, a_l] = {\mbox{{{\sf Split}}}\left(a\right)}$\ $[b_h, b_l] = {\mbox{{{\sf Split}}}\left(b\right)}$\ $y = a_l \otimes b_l \ominus (((x \ominus a_h \otimes b_h) \ominus a_l \otimes b_h) \ominus a_h \otimes b_l )$ The next theorem exhibits the previously announced properties of [[TwoSum]{}]{} and [[TwoProd]{}]{}. \[theorem:EFT\] Let $a, b$ in ${\mathbb{F}}$ and $x, y \in {\mathbb{F}}$ such that $[x, y] = {{\sf TwoSum}}(a,b)$ (Algorithm \[algo:TwoSum\]). Then, ever in the presence of underflow, $$a + b = x + y, \quad x = a \oplus b, \quad |y| \leq {\ensuremath{{\mathbf{u}}}}|x|, \quad |y| \leq {\ensuremath{{\mathbf{u}}}}|a+b|.$$ Let $a, b \in {\mathbb{F}}$ and $x, y \in {\mathbb{F}}$ such that $[x, y] = {{\sf TwoProd}}(a,b)$ (Algorithm \[algo:TwoProd\]). Then, if no underflow occurs, $$a \times b = x + y, \quad x = a \otimes b, \quad |y| \leq {\ensuremath{{\mathbf{u}}}}|x|, \quad |y| \leq {\ensuremath{{\mathbf{u}}}}|a \times b|.$$ We notice that algorithms [[TwoSum]{}]{} and [[TwoProd]{}]{} only require well optimizable floating point operations. They do not use branches, nor access to the mantissa that can be time-consuming. We just mention that significant improvements of these algorithms are defined when a Fused-Multiply-and-Add operator is available [@grll05]. An EFT for the Horner algorithm ------------------------------- As previously mentioned, next EFT for the polynomial evaluation with the Horner algorithm exhibits the exact rounding error generated by the Horner algorithm together with an algorithm to compute it. \[algo:EFTHorner\] EFT for the Horner algorithm =function $[s_0, p_{\pi}, p_{\sigma}] = {{\sf EFTHorner}}(p,x)$\ $s_n = a_n$\ for $i=n-1:-1:0$\ $[p_i , \pi_i ]$ = `TwoProd`$(s_{i+1},x)$\ $[s_i , \sigma_i ]$ = `TwoSum`$(p_i,a_i)$\ Let $\pi_i$ be the coefficient of degree $i$ in $p_{\pi}$\ Let $\sigma_i$ be the coefficient of degree $i$ in $p_{\sigma}$\ end \[theorem:EFTHorner\] Let $p(x) = \sum_{i=0}^{n} a_i x^i$ be a polynomial of degree $n$ with floating point coefficients, and let $x$ be a floating point value. Then [Algorithm \[algo:EFTHorner\]]{} computes both i) the floating point evaluation ${\mbox{{{\sf Horner}}}\left(p, x\right)}$ and ii) two polynomials $p_{\pi}$ and $p_{\sigma}$ of degree $n-1$ with floating point coefficients, such that $$[{\mbox{{{\sf Horner}}}\left(p, x\right)}, p_{\pi}, p_{\sigma}] = {\mbox{{{\sf EFTHorner}}}\left(p, x\right)}.$$ If no underflow occurs, $$\label{rel:EFTHorner} p(x) = {\mbox{{{\sf Horner}}}\left(p, x\right)} + (p_{\pi}+p_{\sigma})(x).$$ Moreover, $$\label{rel:BoundErrorTerm} ({\ensuremath{\,\widetilde{p_{\pi}+p_{\sigma}}}})(x) \leq \gamma_{2n} {\ensuremath{\,\widetilde{p}}}(x).$$ [Relation (\[rel:EFTHorner\])]{} means that algorithm [[EFTHorner]{}]{} is an EFT for polynomial evaluation with the Horner algorithm. Since [[TwoProd]{}]{} and [[TwoSum]{}]{} are EFT from [Theorem \[theorem:EFT\]]{} it follows that $s_{i+1} x = p_i + \pi_i$ and $p_i + a_i = s_i + \sigma_i$. Thus we have $s_i = s_{i+1} x + a_i - \pi_i - \sigma_i$, for $i=0,\ldots,n-1$. Since $s_n = a_n$, at the end of the loop we have $$s_0 = \sum_{i=0}^{n}a_i x^i-\sum_{i=0}^{n-1}\pi_i x^i-\sum_{i=0}^{n-1}\sigma_i x^i,$$ which proves (\[rel:EFTHorner\]). Now we prove relation (\[rel:BoundErrorTerm\]) According to the error analysis of the Horner algorithm (see [@ASNA:02 p.95]), we can write $${\mbox{{{\sf Horner}}}\left(p, x\right)} = (1+\theta_{2n})a_n x^n + \sum_{i=0}^{n-1} (1+\theta_{2i+1}) a_i x^i,$$ where every $\theta_{k}$ satisfies $|\theta_k| \leq \gamma_k$. Then using (\[rel:EFTHorner\]) we have $$(p_{\pi} + p_{\sigma})(x) = p(x) - {\mbox{{{\sf Horner}}}\left(p, x\right)} = -\theta_{2n} a_n x^n - \sum_{i=0}^{n-1} \theta_{2i+1} a_i x^i.$$ Therefore it yields next expected inequalities between the absolute values, $$({\ensuremath{\,\widetilde{p_{\pi} + p_{\sigma}}}})(x) \leq \gamma_{2n} |a_n| |x|^n + \sum_{i=0}^{n-1} \gamma_{2i+1} |a_i| |x|^i \leq \gamma_{2n}{\ensuremath{\,\widetilde{p}}}(x).$$ Compensated Horner algorithm ---------------------------- From [Theorem \[theorem:EFTHorner\]]{} the final forward error of the floating point evaluation of $p$ at $x$ according to the Horner algorithm is $$c = p(x) - {\mbox{{{\sf Horner}}}\left(p, x\right)} = (p_{\pi} + p_{\sigma})(x),$$ where the two polynomials $p_{\pi}$ and $p_{\sigma}$ are exactly identified by [[EFTHorner]{}]{} ([Algorithm \[algo:EFTHorner\]]{}) —this latter also computes ${\mbox{{{\sf Horner}}}\left(p, x\right)}$. Therefore, the key of the compensated algorithm is to compute, in the working precision, first an approximate ${{\,\widehat{c}}}$ of the final error $c$ and then a corrected result $${\ensuremath{\,\overline{r}}} = {\mbox{{{\sf Horner}}}\left(p, x\right)} \oplus {{\,\widehat{c}}}.$$ These two computations leads to next compensated Horner algorithm [[CompHorner]{}]{} ([Algorithm \[algo:CompHorner\]]{}). \[algo:CompHorner\] Compensated Horner algorithm =function ${\ensuremath{\,\overline{r}}} = {\mbox{{{\sf CompHorner}}}\left(p, x\right)}$\ $\left[{{\,\widehat{r}}}, p_{\pi}, p_{\sigma}\right] = {\mbox{{{\sf EFTHorner}}}\left(p, x\right)}$\ ${{\,\widehat{c}}} = {\mbox{{{\sf Horner}}}\left(p_{\pi} \oplus p_{\sigma}, x\right)}$\ ${\ensuremath{\,\overline{r}}} = {{\,\widehat{r}}} \oplus {{\,\widehat{c}}}$ We say that ${{\,\widehat{c}}}$ is a correcting term for ${\mbox{{{\sf Horner}}}\left(p, x\right)}$. The corrected result $\bar{r}$ is expected to be more accurate than the first result ${\mbox{{{\sf Horner}}}\left(p, x\right)}$ as proved in next section. An [*a priori*]{} condition for faithful rounding {#sec:AprioriBound} ================================================= We start proving the accuracy behavior of the compensated Horner algorithm we previously mentioned with introductory inequality (\[rel:CompHornerBound\]) and that motivates the search for a faithful polynomial evaluation. This bound (and its proof) is the first step towards the proposed [[*a priori*]{} ]{}sufficient condition for a faithful rounding with compensated Horner algorithm. Accuracy of the compensated Horner algorithm -------------------------------------------- Next result proves that the result of a polynomial evaluation computed with the compensated Horner algorithm ([Algorithm \[algo:CompHorner\]]{}) is as accurate as if computed by the classic Horner algorithm using twice the working precision and then rounded to the working precision. \[theorem:CompHorner\] Consider a polynomial $p$ of degree $n$ with floating point coefficients, and $x$ a floating point value. If no underflow occurs, $$\label{rel:EBCompHorner} |{\mbox{{{\sf CompHorner}}}\left(p, x\right)} - p(x)| \leq {\ensuremath{{\mathbf{u}}}}|p(x)| + \gamma_{2n}^{2} {\ensuremath{\,\widetilde{p}}}(x).$$ The absolute forward error generated by [Algorithm \[algo:CompHorner\]]{} is $$| {\ensuremath{\,\overline{r}}} - p(x) | = \left| ({{\,\widehat{r}}} \oplus {{\,\widehat{c}}}) - p(x) \right| = \left| (1 +{\varepsilon})({{\,\widehat{r}}} + {{\,\widehat{c}}}) - p(x) \right| \quad \mbox{with} \quad |{\varepsilon}| \leq {\ensuremath{{\mathbf{u}}}}.$$ Let $c = (p_{\pi}+p_{\sigma})(x)$. From [Theorem \[theorem:EFTHorner\]]{} we have ${{\,\widehat{r}}} ={\mbox{{{\sf Horner}}}\left(p, x\right)} = p(x) - c$, thus $$| {\ensuremath{\,\overline{r}}} - p(x) | =\left| (1 +{\varepsilon}) \left( p(x) - c + {{\,\widehat{c}}}\right) - p(x) \right| \leq {\ensuremath{{\mathbf{u}}}}|p(x)| + (1 + {\ensuremath{{\mathbf{u}}}}) | {{\,\widehat{c}}} - c |.$$ Since ${{\,\widehat{c}}} = {\mbox{{{\sf Horner}}}\left(p_{\pi} \oplus p_{\sigma}, x\right)}$ with $p_{\pi}$ and $p_{\sigma}$ two polynomials of degree $n-1$, [Lemma \[lemma:HornerSum\]]{} yields $|{{\,\widehat{c}}} - c| \leq \gamma_{2n-1} ({\ensuremath{\,\widetilde{p_{\pi}+ p_{\sigma}}}})(x)$. Then using (\[rel:BoundErrorTerm\]) we have $|{{\,\widehat{c}}} - c| \leq \gamma_{2n-1} \gamma_{2n} {\ensuremath{\,\widetilde{p}}}(x)$. Since $(1+{\ensuremath{{\mathbf{u}}}})\gamma_{2n-1} \leq \gamma_{2n}$, we finally write the expected error bound (\[rel:EBCompHorner\]). For later use, we notice that $|{{\,\widehat{c}}} - c| \leq \gamma_{2n-1} \gamma_{2n} {\ensuremath{\,\widetilde{p}}}(x)$ implies $$\label{rel:EBCorrectingTerm} |{{\,\widehat{c}}} - c| \leq \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x).$$ It is interesting to interpret the previous theorem in terms of the condition number of the polynomial evaluation of $p$ at $x$. Combining the error bound (\[rel:EBCompHorner\]) with the condition number (\[rel:CondPoly\]) of polynomial evaluation gives the precise writing of our introductory [inequality (\[rel:CompHornerBound\])]{}, $$\label{rel:REBCompHorner} \frac{| {\mbox{{{\sf CompHorner}}}\left(p, x\right)} - p(x) |}{|p(x)|} \leq {\ensuremath{{\mathbf{u}}}}+ \gamma_{2n}^{2} {\operatorname{cond}}(p, x).$$ In other words, the bound for the relative error of the computed result is essentially $\gamma_{2n}^2$ times the condition number of the polynomial evaluation, plus the inevitable summand ${\ensuremath{{\mathbf{u}}}}$ for rounding the result to the working precision. In particular, if ${\operatorname{cond}}(p, x) < {\ensuremath{{\mathbf{u}}}}/ \gamma_{2n}^2$, then the relative accuracy of the result is bounded by a constant of the order ${\ensuremath{{\mathbf{u}}}}$. This means that the compensated Horner algorithm computes an evaluation accurate to the last few bits as long as the condition number is smaller than ${\ensuremath{{\mathbf{u}}}}/ \gamma_{2n}^2 \approx 1/4n^2{\ensuremath{{\mathbf{u}}}}$. Besides that, relation (\[rel:REBCompHorner\]) tells us that the computed result is as accurate as if computed by the classic Horner algorithm with twice the working precision, and then rounded to the working precision. An [[*a priori*]{} ]{}condition for faithful rounding {#an-a-priori-condition-for-faithful-rounding} ----------------------------------------------------- Now we propose a sufficient condition on ${\operatorname{cond}}(p, x)$ to ensure that the corrected result ${\ensuremath{\,\overline{r}}}$ computed with the compensated Horner algorithm is a faithful rounding of the exact result $p(x)$. For this purpose, we use the following lemma from [@RuOO06]. \[lemma:FaithfulRounding\] Let $r, \delta$ be two real numbers and ${\ensuremath{\,\overline{r}}} = {\operatorname{fl}}(r)$. We assume here that ${\ensuremath{\,\overline{r}}}$ is a normalized floating point number. If $|\delta| < \frac{{\ensuremath{{\mathbf{u}}}}}{2} |{{\,\widehat{r}}}|$ then ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of $r +\delta$. From [Lemma \[lemma:FaithfulRounding\]]{}, we derive a useful criterion to ensure that the compensated result provided by [[CompHorner]{}]{} is faithfully rounded to the working precision. \[lemma:FaithfulRoundingCompHorner\] Let $p$ be a polynomial of degree $n$ with floating point coefficients, and $x$ be a floating point value. We consider the approximate ${\ensuremath{\,\overline{r}}}$ of $p(x)$ computed with ${\mbox{{{\sf CompHorner}}}\left(p, x\right)}$, and we assume that no underflow occurs during the computation. Let $c$ denotes $c = (p_{\pi} + p_{\sigma})(x)$. If $|{{\,\widehat{c}}} - c| < \frac{{\ensuremath{{\mathbf{u}}}}}{2}|{\ensuremath{\,\overline{r}}}|$, then ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of $p(x)$. We assume that $|{{\,\widehat{c}}} - c| < \frac{{\ensuremath{{\mathbf{u}}}}}{2}|{\ensuremath{\,\overline{r}}}|$. From the notations of [Algorithm \[algo:CompHorner\]]{}, we recall that ${\operatorname{fl}}({{\,\widehat{r}}} + {{\,\widehat{c}}}) = {\ensuremath{\,\overline{r}}}$. Then from [Lemma \[lemma:FaithfulRounding\]]{} it follows that ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of ${{\,\widehat{r}}}+{{\,\widehat{c}}}+c-{{\,\widehat{c}}} = {{\,\widehat{r}}}+c$. Since $[{{\,\widehat{r}}}, p_{\pi}, p_{\sigma}] = {\mbox{{{\sf EFTHorner}}}\left(p, x\right)}$, [Theorem \[theorem:EFTHorner\]]{} yields $p(x) = {{\,\widehat{r}}}+c$. Therefore ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of $p(x)$. The criterion proposed in [Lemma \[lemma:FaithfulRoundingCompHorner\]]{} concerns the accuracy of the correcting term ${{\,\widehat{c}}}$. Nevertheless [Relation (\[rel:EBCorrectingTerm\])]{} pointed after the proof of [Theorem \[theorem:CompHorner\]]{} says that the absolute error $|{{\,\widehat{c}}} - c|$ is bounded by $\gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x)$. This provides us a more useful criterion, since it relies on the condition number ${\operatorname{cond}}(p, x)$, to ensure that [[CompHorner]{}]{} computes a faithfully rounded result. \[prop:FaithfulRoundingCompHorner\] Let $p$ be a polynomial of degree $n$ with floating point coefficients, and $x$ a floating point value. If $$\label{rel:APrioriCriterion} {\operatorname{cond}}(p, x) < \frac{1-{\ensuremath{{\mathbf{u}}}}}{2+{\ensuremath{{\mathbf{u}}}}} {\ensuremath{{\mathbf{u}}}}{\gamma_{2n}}^{-2},$$ then ${\mbox{{{\sf CompHorner}}}\left(p, x\right)}$ computes a faithful rounding of the exact $p(x)$. We assume that (\[rel:APrioriCriterion\]) is satisfied and we use the same notations as in [Lemma \[lemma:FaithfulRoundingCompHorner\]]{}. First we notice that ${\ensuremath{\,\overline{r}}}$ and $p(x)$ are of the same sign. Indeed, from (\[rel:EBCompHorner\]) it follows that $\left|{{\ensuremath{\,\overline{r}}}}/{p(x)} - 1\right| \leq {\ensuremath{{\mathbf{u}}}}+ \gamma_{2n}^2 {\operatorname{cond}}(p, x)$, and therefore ${{\ensuremath{\,\overline{r}}}}/{p(x)} \geq 1 - {\ensuremath{{\mathbf{u}}}}-\gamma_{2n}^2 {\operatorname{cond}}(p, x)$. But (\[rel:APrioriCriterion\]) implies that $1 - {\ensuremath{{\mathbf{u}}}}-\gamma_{2n}^2 {\operatorname{cond}}(p, x) > 1 - 3{\ensuremath{{\mathbf{u}}}}/(2+{\ensuremath{{\mathbf{u}}}}) > 0$, hence ${{\ensuremath{\,\overline{r}}}}/{p(x)} > 0$. Since ${\ensuremath{\,\overline{r}}}$ and $p(x)$ have the same sign, it is easy to see that $$\label{proof:FaithfulRoundingCompHorner:rel1} (1-{\ensuremath{{\mathbf{u}}}})|p(x)| - \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) \leq |{\ensuremath{\,\overline{r}}}|.$$ Indeed, if $p(x)>0$ then (\[rel:EBCompHorner\]) implies $p(x) - {\ensuremath{{\mathbf{u}}}}|p(x)| - \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) \leq {\ensuremath{\,\overline{r}}} = |{\ensuremath{\,\overline{r}}}|$. If $p(x)<0$, from (\[rel:EBCompHorner\]) it follows that ${\ensuremath{\,\overline{r}}} \leq p(x) + {\ensuremath{{\mathbf{u}}}}|p(x)| + \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x)$, hence $- p(x) - {\ensuremath{{\mathbf{u}}}}|p(x)| - \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) \leq - {\ensuremath{\,\overline{r}}} = |{\ensuremath{\,\overline{r}}}|$. Next, a small computation proves that $${\operatorname{cond}}(p, x) < \frac{1-{\ensuremath{{\mathbf{u}}}}}{2+{\ensuremath{{\mathbf{u}}}}} {\ensuremath{{\mathbf{u}}}}{\gamma_{2n}}^{-2} \quad \mbox{if and only if} \quad \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) < \frac{{\ensuremath{{\mathbf{u}}}}}{2} \left[(1-{\ensuremath{{\mathbf{u}}}})|p(x)| - \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x)\right].$$ Finally, from (\[rel:EBCorrectingTerm\]) and (\[proof:FaithfulRoundingCompHorner:rel1\]) it follows $$|{{\,\widehat{c}}} - c| \leq \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) < \frac{{\ensuremath{{\mathbf{u}}}}}{2} \left[ (1-{\ensuremath{{\mathbf{u}}}})|p(x)| - \gamma_{2n}^2 {\ensuremath{\,\widetilde{p}}}(x) \right] \leq \frac{{\ensuremath{{\mathbf{u}}}}}{2} |{\ensuremath{\,\overline{r}}}|.$$ From [Lemma \[lemma:FaithfulRoundingCompHorner\]]{} we deduce that ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of $p(x)$. Numerical values of condition numbers for a faithful polynomial evaluation in IEEE-754 double precision are presented in [Table \[tab:BoundsCond\]]{} for degrees varying from 10 to 500. n 10 100 200 300 400 500 --------------------------------------------------------------------------------------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- --------------------- --------------------- -- $\frac{1-{\ensuremath{{\mathbf{u}}}}}{2-{\ensuremath{{\mathbf{u}}}}} {\ensuremath{{\mathbf{u}}}}{\gamma_{2n}}^{-2}$ $1.13 \cdot 10^{13}$ $1.13 \cdot 10^{11}$ $2.82 \cdot 10^{10}$ $1.13 \cdot 10^{10}$ $7.04 \cdot 10^{9}$ $4.51 \cdot 10^{9}$ : [*A priori*]{} bounds on the condition number to ensure faithful rounding in IEEE-754 double precision for polynomials of degree 10 to 500[]{data-label="tab:BoundsCond"} Dynamic and validated error bounds for faithful rounding and accuracy {#sec:DynBound} ===================================================================== The results presented in [Section \[sec:AprioriBound\]]{} are perfectly suited for theoretical purpose, for instance when we can [[*a priori*]{} ]{}bound the condition number of the evaluation. However, neither the error bound in [Theorem \[theorem:CompHorner\]]{}, nor the criterion proposed in [Theorem \[prop:FaithfulRoundingCompHorner\]]{} can be easily checked using only floating point arithmetic. Here we provide dynamic counterparts of [Theorem \[theorem:CompHorner\]]{} and [Proposition \[prop:FaithfulRoundingCompHorner\]]{}, that can be evaluated using floating point arithmetic in the “round to the nearest” rounding mode. \[lemma:DynBoundErrorTerm\] Consider a polynomial $p$ of degree $n$ with floating point coefficients, and $x$ a floating point value. We use the notations of [Algorithm \[algo:CompHorner\]]{}, and we denote $(p_{\pi}+p_{\sigma})(x)$ by $c$. Then $$\label{rel:DynBoundErrorTerm} |c-{{\,\widehat{c}}}| \leq {\operatorname{fl}}\left( \frac{ {{\,\widehat{\gamma}}}_{2n-1} {\mbox{{{\sf Horner}}}\left(|p_{\pi}|\oplus|p_{\sigma}|,|x|\right)} }{ 1-2(n+1){\ensuremath{{\mathbf{u}}}}} \right) := {{\,\widehat{\alpha}}}.$$ Let us denote ${\mbox{{{\sf Horner}}}\left(|p_{\pi}|\oplus|p_{\sigma}|,|x|\right)}$ by ${{\,\widehat{b}}}$. Since $c = (p_{\pi}+p_{\sigma})(x)$ and ${{\,\widehat{c}}} = {\mbox{{{\sf Horner}}}\left(p_{\pi} \oplus p_{\sigma}, x\right)}$ where $p_{\pi}$ and $p_{\sigma}$ are two polynomials of degree $n-1$, [Lemma \[lemma:HornerSum\]]{} yields $$|c-{{\,\widehat{c}}}| \leq \gamma_{2n-1} ({\ensuremath{\,\widetilde{p_{\pi}}}}+{\ensuremath{\,\widetilde{p_{\sigma}}}})(x) \leq (1+{\ensuremath{{\mathbf{u}}}})^{2n-1} \gamma_{2n-1} {{\,\widehat{b}}}.$$ From (\[rel:GammaK\]) and (\[rel:StdModel\]) it follows that $$|c-{{\,\widehat{c}}}| \leq (1+{\ensuremath{{\mathbf{u}}}})^{2n} {{\,\widehat{\gamma}}}_{2n-1} {{\,\widehat{b}}} \leq (1+{\ensuremath{{\mathbf{u}}}})^{2n+1} {\operatorname{fl}}({{\,\widehat{\gamma}}}_{2n-1}{{\,\widehat{b}}}).$$ Finally we use relation (\[rel:CompBound\]) to obtain the error bound. [Lemma \[lemma:DynBoundErrorTerm\]]{} allows us to compute a validated error bound for the computed correcting term ${{\,\widehat{c}}}$. We apply this result twice to derive next [Theorem \[prop:DynErrorBounds\]]{}. First with [Lemma \[lemma:FaithfulRoundingCompHorner\]]{} it yields the expected dynamic condition for faithful rounding. Then from the EFT for the Horner algorithm ([Theorem \[theorem:EFTHorner\]]{}) we know that $p(x) = {{\,\widehat{r}}}+c$. Since ${\ensuremath{\,\overline{r}}} = {{\,\widehat{r}}}\oplus{{\,\widehat{c}}}$, we deduce $|{\ensuremath{\,\overline{r}}} - p(x)| = |({{\,\widehat{r}}}\oplus{{\,\widehat{c}}}) - ({{\,\widehat{r}}}+{{\,\widehat{c}}})+({{\,\widehat{c}}}-c)|$. Hence we have $$|{\ensuremath{\,\overline{r}}} - p(x)| \leq |({{\,\widehat{r}}}\oplus{{\,\widehat{c}}}) - ({{\,\widehat{r}}}+{{\,\widehat{c}}})| + |({{\,\widehat{c}}}-c)|.$$ The first term $|({{\,\widehat{r}}}\oplus{{\,\widehat{c}}}) - ({{\,\widehat{r}}}+{{\,\widehat{c}}})|$ in the previous inequality is basically the absolute rounding error that occurs when computing ${\ensuremath{\,\overline{r}}}={{\,\widehat{r}}}\oplus{{\,\widehat{c}}}$. Using only the bound (\[rel:StdModel\]) of the standard model of floating point arithmetic, it could be bounded by ${\ensuremath{{\mathbf{u}}}}|{\ensuremath{\,\overline{r}}}|$. But here we benefit again from error free transformations using algorithm ${{\sf TwoSum}}$ to compute the actual rounding error exactly, which leads to a sharper error bound. Next [Relation (\[rel:DynErrorBound\])]{} improves the dynamic bound presented in [@grll05]. \[prop:DynErrorBounds\] Consider a polynomial $p$ of degree $n$ with floating point coefficients, and $x$ a floating point value. Let ${\ensuremath{\,\overline{r}}}$ be the computed value, ${\ensuremath{\,\overline{r}}}={\mbox{{{\sf CompHorner}}}\left(p,x\right)}$ ([Algorithm \[algo:CompHorner\]]{}) and let ${{\,\widehat{\alpha}}}$ be the error bound defined by [Relation (\[rel:DynBoundErrorTerm\])]{}. i) If ${{\,\widehat{\alpha}}} < \frac{{\ensuremath{{\mathbf{u}}}}}{2}|{\ensuremath{\,\overline{r}}}|$, then ${\ensuremath{\,\overline{r}}}$ is a faithful rounding of $p(x)$ \[rel:DynFaithfulBound\]. ii) Let $e$ be the floating point value such that ${\ensuremath{\,\overline{r}}}+e = {{\,\widehat{r}}}+{{\,\widehat{c}}}$, [[*i.e.*]{}, ]{}$[{\ensuremath{\,\overline{r}}}, e] = {\mbox{{{\sf TwoSum}}}\left({{\,\widehat{r}}},{{\,\widehat{c}}}\right)}$, where ${{\,\widehat{r}}}$ and ${{\,\widehat{c}}}$ are defined by [Algorithm \[algo:CompHorner\]]{}. The absolute error of the computed result ${\ensuremath{\,\overline{r}}} = {\mbox{{{\sf CompHorner}}}\left(p,x\right)}$ is bounded as follows, $$\label{rel:DynErrorBound} |{\ensuremath{\,\overline{r}}}-p(x)| \leq {\operatorname{fl}}\left(\frac{{{\,\widehat{\alpha}}}+|e|}{1-2u}\right) := {{\,\widehat{\beta}}}.$$ The first proposition follows directly from [Lemma \[lemma:FaithfulRoundingCompHorner\]]{}. By hypothesis ${\ensuremath{\,\overline{r}}}= {{\,\widehat{r}}}+{{\,\widehat{c}}}-e$, and from [Theorem \[theorem:EFTHorner\]]{} we have $p(x) = {{\,\widehat{r}}}+c$, thus $$|{\ensuremath{\,\overline{r}}}-p(x)| = |{{\,\widehat{c}}}-c-e| \leq |{{\,\widehat{c}}}-c| + |e| \leq {{\,\widehat{\alpha}}} + |e|.$$ From (\[rel:StdModel\]) and (\[rel:CompBound\]) it follows that $$|{\ensuremath{\,\overline{r}}}-p(x)| \leq (1+{\ensuremath{{\mathbf{u}}}}) {\operatorname{fl}}({{\,\widehat{\alpha}}} + |e|) \leq {\operatorname{fl}}\left(\frac{{{\,\widehat{\alpha}}}+|e|}{1-2u}\right);$$ which proves the second proposition. From [Theorem \[prop:DynErrorBounds\]]{} we deduce the following algorithm. It computes the compensated result ${\ensuremath{\,\overline{r}}}$ together with the validated error bound ${{\,\widehat{\beta}}}$. Moreover, the boolean value $\textsf{isfaithful}$ is set to true if and only if the result is proved to be faithfully rounded. \[algo:FaithfulRoundingCompHorner\] Compensated Horner algorithm with check of the faithful rounding =function $[{\ensuremath{\,\overline{r}}}, {{\,\widehat{\beta}}}, \mbox{\textsf{isfaithful}}] = {\mbox{{{\sf CompHornerIsFaithul}}}\left(p, x\right)}$\ $\left[{{\,\widehat{r}}}, p_{\pi}, p_{\sigma}\right] = {\mbox{{{\sf EFTHorner}}}\left(p, x\right)}$\ ${{\,\widehat{c}}} = {\mbox{{{\sf Horner}}}\left(p_{\pi} \oplus p_{\sigma}, x\right)}$\ ${{\,\widehat{b}}} = {\mbox{{{\sf Horner}}}\left(|p_{\pi}| \oplus |p_{\sigma}|, |x|\right)}$\ $[{\ensuremath{\,\overline{r}}}, e] = {\mbox{{{\sf TwoSum}}}\left({{\,\widehat{r}}}, {{\,\widehat{c}}}\right)}$\ ${{\,\widehat{\alpha}}} = ({{\,\widehat{\gamma}}}_{2n-1} \otimes {{\,\widehat{b}}}) \oslash (1 \ominus 2(n+1) \otimes {\ensuremath{{\mathbf{u}}}})$\ ${{\,\widehat{\beta}}} = ({{\,\widehat{\alpha}}} \oplus |e|) \oslash (1-2 \otimes {\ensuremath{{\mathbf{u}}}})$\ $\mbox{\textsf{isfaithful}} = ({{\,\widehat{\alpha}}} < \frac{{\ensuremath{{\mathbf{u}}}}}{2}|{\ensuremath{\,\overline{r}}}|)$ Experimental results {#sec:ExperimentalResults} ==================== We consider polynomials $p$ with floating point coefficients and floating point entries $x$. For presented accuracy tests we use Matlab codes for [[CompHorner]{}]{} ([Algorithm \[algo:CompHorner\]]{}) and [[CompHornerIsFaithul]{}]{} ([Algorithm \[algo:FaithfulRoundingCompHorner\]]{}). These Matlab programs are presented in Appendix \[sec:MatlabCodes\]. From these Matlab codes, we see that [[CompHorner]{}]{} requires $O(21n)$ flop and that [[CompHornerIsFaithul]{}]{} requires $O(26n)$ flop. For time performance tests previous algorithms are coded in C language and several test platforms are described in next [Table \[tab:MeasTimePerf\]]{}. Accuracy tests -------------- We start testing the efficiency of faithful rounding with compensated Horner algorithm and the dynamic control of faithfulness. Then we focus more on both the [[*a priori*]{} ]{}and dynamic bounds with two other test sets. Three cases may occur when the dynamic test for faithful rounding in [Algorithm \[algo:FaithfulRoundingCompHorner\]]{} is performed. 1. The computed result is faithfully rounded and this is ensured by the dynamic test. Corresponding plots are green in next figures. 2. The computed result is actually faithfully rounded but the dynamic test fails to ensure this property. Corresponding plots are blue. 3. The computed result is not faithfully rounded and plotted in red in this case. Next figures should be observed in color. ### Faithful rounding with compensated Horner In the first experiment set, we evaluate the expanded form of polynomials $p_n(x) = (1-x)^n$, for degree $n=6,8,10,12$, at $2048$ equally spaced floating point entries being near the multiple root $x=1$. These evaluations are extremely ill-conditioned since $${\operatorname{cond}}(p_n, x) = \left| \frac{1+|x|}{1-x} \right|^n.$$ These condition numbers are plotted in the lower frame of [Figure \[fig:CHFaithBinom\]]{} while $x$ varies around the root. These huge values have a sense since polynomials $p$ are exact in IEEE-754 double precision. Results are reported on [Figure \[fig:CHFaithBinom\]]{}. The well known relation between the lost of accuracy and the nearness and the multiplicity of the root, [[*i.e.*]{}, ]{}the increasing of the condition number, is clearly illustrated. These results also illustrate that the dynamic bound becomes more pessimistic as the condition number increases. In next figures the horizontal axis does not represent the $x$ entry range anymore but the condition number which governs the whole behavior.\ For the next experiment set, we first designed a generator of arbitrary ill-conditioned polynomial evaluations. It relies on the condition number definition (\[rel:CondPoly\]). Given a degree $n$, a floating point argument $x$ and a targeted condition number $C$, it generates a polynomial $p$ with floating point coefficients such that ${\operatorname{cond}}(p, x)$ has the same order of magnitude as $C$. The principle of the generator is the following. 1. $\lfloor n/2 \rfloor$ coefficients are randomly selected and generated such that ${\ensuremath{\,\widetilde{p}}}(x) = \sum |a_i||x|^i \approx C$, 2. the remaining coefficients are generated ensuring $|p(x)| \approx 1$ thanks to high accuracy computation. Therefore we obtain polynomials $p$ such that ${\operatorname{cond}}(p, x) = {\ensuremath{\,\widetilde{p}}}(x) / |p(x)| \approx C$, for arbitrary values of $C$. In this test set we consider generated polynomials of degree $50$ whose condition numbers vary from about $10^2$ to $10^{35}$. These huge condition numbers again have a sense here since the coefficients and the argument of every polynomial are floating point numbers. The results of the tests performed with [[CompHornerIsFaithul]{}]{} ([Algorithm \[algo:FaithfulRoundingCompHorner\]]{}) are reported on [Figure \[fig:CHFaithGen\]]{}. As expected every polynomial with a condition number smaller than the [[*a priori*]{} ]{}bound (\[rel:APrioriCriterion\]) is faithfully evaluated with [Algorithm \[algo:FaithfulRoundingCompHorner\]]{} —green plots at the left of the leftmost vertical line. On [Figure \[fig:CHFaithGen\]]{} we also see that evaluations with faithful rounding appear for condition numbers larger than the [[*a priori*]{} ]{}bound (\[rel:APrioriCriterion\]) — green and blue plots at the right of the leftmost vertical line. As expected a large part of these cases are detected by the dynamic test introduced in [Theorem \[prop:DynErrorBounds\]]{} —the green ones. Next experiment set comes back to this point. We also notice that the compensated Horner algorithm produces accurate evaluations for condition numbers up to about $1/{\ensuremath{{\mathbf{u}}}}$ —green and blue plots. ### Significance of the dynamic error bound We illustrate the significance of the dynamic error bound (\[rel:DynErrorBound\]), compared to the [[*a priori*]{} ]{}error bound (\[rel:EBCompHorner\]) and to the actual forward error. We evaluate the expanded form of $p(x) = (1-x)^5$ for $400$ points near $x=1$. For each value of the argument $x$, we compute ${\mbox{{{\sf CompHorner}}}\left(p, x\right)}$ ([Algorithm \[algo:CompHorner\]]{}), the associated dynamic error bound (\[rel:DynErrorBound\]) and the actual forward error. The results are reported on [Figure \[fig:CHBounds\]]{}. As already noticed, the closer the argument is to the root $1$ ([[*i.e.*]{}, ]{}, the more the condition number increases), the more pessimistic becomes the [[*a priori*]{} ]{}error bound. Nevertheless our dynamic error bound is more significant than the [[*a priori*]{} ]{}error bound as it takes into account the rounding errors that occur during the computation. Time performances ----------------- [|ll|c|c|c|]{} & $\frac{{{\sf CompHorner}}}{{{\sf Horner}}}$ & $\frac{{{\sf CompHornerIsFaith}}}{{{\sf Horner}}}$ & $\frac{{{\sf DDHorner}}}{{{\sf Horner}}}$\ & & 3.77 & 5.52 & 10.00\ & & 3.06 & 5.31 & 8.88\ & & 3.89 & 4.43 & 10.48\ & & 3.64 & 4.59 & 5.50\ & & 1.87 & 2.30 & 8.78\ & $\sim 2 - 4$ & $\sim 4 - 6$ & $\sim 5 - 10$\ All experiments are performed using IEEE-754 double precision. Since the double-doubles [@LiHB:01; @XBLAS:02] are usually considered as the most efficient portable library to double the IEEE-754 double precision, we consider it as a reference in the following comparisons. For our purpose, it suffices to know that a double-double number $a$ is the pair $(a_h, a_l)$ of IEEE-754 floating point numbers with $a = a_h + a_l$ and $|a_l| \leq {\ensuremath{{\mathbf{u}}}}|a_h|$. This property implies a renormalisation step after every arithmetic operation with double-double values. We denote by [[DDHorner]{}]{} our implementation of the Horner algorithm with the double-double format, derived from the implementation proposed in [@XBLAS:02]. We implement the three algorithms [[CompHorner]{}]{}, [[CompHornerIsFaith]{}]{} and [[DDHorner]{}]{} in a C code to measure their overhead compared to the [[Horner]{}]{} algorithm. We program these tests straightforwardly with no other optimization than the ones performed by the compiler. All timings are done with the cache warmed to minimize the memory traffic over-cost. We test the running times of these algorithms for different architectures with different compilers as described in [Table \[tab:MeasTimePerf\]]{}. Our measures are performed with polynomials whose degree vary from 5 to 200 by step of 5. For each algorithm, we measure the ratio of its computing time over the computing time of the classic Horner algorithm; we display the average time ratio over all test cases in [Table \[tab:MeasTimePerf\]]{}.\ The results presented in [Table \[tab:MeasTimePerf\]]{} show that the slowdown factor introduced by [[CompHorner]{}]{} compared to the classic [[Horner]{}]{} roughly varies between 2 and 4. The same slowdown factor varies between 4 and 6 for [[CompHornerIsFaithul]{}]{} and between 5 and 10 for [[DDHorner]{}]{}. We can see that [[CompHornerIsFaithul]{}]{} runs a most 2 times slower than [[CompHorner]{}]{}: the over-cost due to the dynamic test for faithful rounding is therefore quite reasonable. Anyway [[CompHorner]{}]{} and [[CompHornerIsFaithul]{}]{} run both significantly faster than [[DDHorner]{}]{}. We provide time ratios for IA’64 architecture (Itanium 2). Tested algorithms take benefit from IA’64 instructions, [[*e.g.*]{}, ]{}[[fma]{}]{}, but are not described in this paper. Conclusion {#sec:Conclusion} ========== Compensated Horner algorithm yields more accurate polynomial evaluation than the classic Horner iteration. Its accuracy behavior is similar to an Horner iteration performed in a doubled working precision. Hence compensated Horner may perform a faithful polynomial evaluation with IEEE-754 floating point arithmetic in the “round to the nearest” rounding mode. An [[*a priori*]{} ]{}sufficient condition with respect on the condition number that ensures such faithfulness has been defined thanks to the error free transformations. These error free transformations also allow us to derive a dynamic sufficient condition that is more significant to check for faithful rounding with compensated Horner algorithm. It is interesting to remark here that the significance of this dynamic bound can be improved easily —how to transform blue plots in green ones? Whereas bounding the error in the computation of the (polynomial) correcting term in [Relation (\[rel:DynBoundErrorTerm\])]{}, a good approximate of the actual error could be computed (applying again [[CompHorner]{}]{} to the correcting term). Of course such extra computation will introduce more running time overhead not necessary useful —green plots are here! So it suffices to run such extra (but costly) checking only if the previous dynamic one fails (a similar strategy as in dynamic filters for geometric algorithms). Compared to the classic Horner algorithm, experimental results exhibit reasonable over-costs for accurate polynomial evaluation (between 2 and 4) and even for this computation with a dynamic checking for faithfulness (between 4 and 6). Let us finally remark than such computation that provides as accuracy as if the working precision is doubled and a faithfulness checking is no more costly in term of running time than the “double-double” counterpart without any check. Future work will be to consider subnormals results and also an adaptative algorithm that ensure faithful rounding for polynomials with an arbitrary condition number. Appendix {#sec:MatlabCodes} ======== Accuracy tests use next Matlab codes for algorithms [Algorithm \[algo:CompHorner\]]{} ([[CompHorner]{}]{}) and [Algorithm \[algo:FaithfulRoundingCompHorner\]]{} ([[CompHornerIsFaithul]{}]{}). Following Matlab convention, $p$ is represented as a vector `p` such that $p(x) = \sum_{i=0}^{n} \mathtt{p}(n-i+1) x^i$. We also recall that Matlab `eps` denotes the machine epsilon, which is the spacing between $1$ and the next larger floating point number, hence ${\ensuremath{{\mathbf{u}}}}= \mathtt{eps}/2$. \[algo:CodeCompHorner\] Code for [Algorithm \[algo:CompHorner\]]{}. function r = CompHorner(p, x) n = length(p)-1; % degree of p [xh, xl] = Split(x); r = p(1); c = 0.0; for i=2:n+1 %[r, pi] = TwoProd(r, x) p = r*x; [rh, rl] = Split(r); pi = rl*xl-(((p-rh*xl)-rl*xh)-rh*xl); %[r, sigma] = TwoSum(r, p(i)) r = p+p(i); t = r-p; sigma = (p-(r-t))+(p(i)-t); % Computation of the correcting term c = c*x+(pi+sig); end % Final correction of the result r = r+c; \[algo:CodeFaithfulRoundingCompHorner\] Code for [Algorithm \[algo:FaithfulRoundingCompHorner\]]{}. function [r, beta, isfaith] = CompHornerIsFaithul(p, x) n = length(p)-1; % degree of p [xh, xl] = Split(x); absx = abs(x); r = p(1); c = 0.0; beta = 0.0; for i=2:n+1 % [r, pi] = TwoProd(r, x) p = r*x; % [rh, rl] = Split(r); pi = rl*xl-(((p-rh*xl)-rl*xh)-rh*xl); % [r, sigma] = TwoSum(r, p(i)) r = p+p(i); t = r-p; sigma = (p-(r-t))+(p(i)-t); % Computation of the correcting term c = c*x+(pi+sig); b = b*absx+(abs(pi)+ abs(sig)); end % Final correction of the result [r, e] = TwoSum(r,c); % Check for faithful rounding alpha = gam(2*n-1)*b / (1-(n+1)*eps); isfaith = alpha < 0.25*eps*abs(r); % Absolute error bound beta = (alpha + abs(e))/(1-2*u); [^1]: DALI Research Team. Laboratory LP2A. 52, avenue Paul Alduy. F-66860 Perpignan, France.

ArXiv

--- abstract: 'We give a new construction of conserved charges in asymptotically anti-de Sitter spacetimes in Einstein’s gravity. The new formula is explicitly gauge-invariant and makes direct use of the linearized curvature tensor instead of the metric perturbation. As an example, we compute the mass and angular momentum of the Kerr-AdS black holes.' author: - Emel Altas - Bayram Tekin title: 'Conserved charges in $AdS$: A new formula' --- Introduction ============ In Einstein’s gravity, outside a source, in a vacuum, all the effects of gravity are encoded in the Riemann tensor (or the Weyl tensor when there is no cosmological constant). This should also be the case for conserved charges, such as mass-energy and angular momentum. Here we show that such a construction of conserved charges exists in asymptotically anti de Sitter $(AdS)$ spacetimes. Namely the total mass-energy or angular momentum of an asymptotically $AdS$ spacetime can be directly computed from an integral that is written in terms of the linearized part of the Riemann tensor. Just like in any other theory in a flat spacetime, conserved charges in gravity play a major role in understanding the integration parameters that appear in the classical solutions such as the black holes and their thermodynamics. But in contrast to the flat spacetime, a generic curved spacetime does not have any symmetries and hence one should not expect any conserved quantities. Fortunately, for some spacetimes which are important in black hole physics and cosmology, one can define total mass (energy) and angular momentum given that the spacetime is asymptotically flat or (anti)-de-Sitter. For an asymptotically flat spacetime, we have the celebrated Arnowitt-Deser-Misner (ADM) mass [@adm] which is also a geometric invariant for the spacelike hypersurface of the four dimensional spacetime as long as certain asymptotic conditions on the decay of the metric tensor and the extrinsic curvature are satisfied. One can also give a similar formula for the total angular momentum of asymptotically flat spacetimes. A generalization to asymptotically $(A)dS$ spacetimes was carried out by Abbott and Deser (AD) [@Abbott_Deser]. In the usual formulation of conserved charges [@Deser_Tekin], given a background Killing vector $\bar{\xi}^{\mu}$ a partially conserved current in Einstein’s theory can be found as $$J^{\mu}:=\sqrt{-g}\bar{\text{\ensuremath{\xi}}}_{\nu}\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}, \hskip 0.5 cm \partial_{\mu}J^{\mu}=0,$$ where $\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$ is the linearized cosmological Einstein tensor and the linearization of the field equations read $\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}=\kappa\tau^{\mu\nu}+{\cal {O}}(h^{2},h^{3},...) =: \kappa T^{\mu\nu}$. So the conserved charge is $$Q(\ensuremath{\xi}):=\intop_{\bar \Sigma}d^{n-1}y \thinspace\sqrt{-g}\thinspace\bar{\text{\ensuremath{\xi}}}_{\nu}\left(\text{\ensuremath{{\cal {G}}}}^{\nu0}\right)^{\left(1\right)}, \label{ilk}$$ where $\bar \Sigma$ is a spatial hypersurface. Note that as $T^{\mu \nu}$ includes all the localized matter and higher order gravitational corrections, despite appearance, (\[ilk\]) captures all the non-linear terms. See the recent review articles [@Adami; @Compere] for more details. To proceed further one needs to write $\bar{\text{\ensuremath{\xi}}}_{\nu}\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$ to be the divergence of a tensor. This requires writing $\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$ explicitly in terms of the metric perturbation $h_{\mu\nu}$ which yields [@Abbott_Deser] $$\bar{\text{\ensuremath{\xi}}}_{\nu}\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}=\nabla_{\alpha}\left(\bar{\text{\ensuremath{\xi}}}_{\nu}\nabla_{\beta}K^{\mu\alpha\nu\beta}-K^{\mu\beta\nu\alpha}\nabla_{\beta}\bar{\text{\ensuremath{\xi}}}_{\nu}\right), \label{iki}$$ with the superpotential given as $$K^{\mu\alpha\nu\beta}:=\frac{1}{2}\left(\bar{g}^{\alpha\nu}\tilde{h}^{\mu\beta}+\bar{g}^{\mu\beta}\tilde{h}^{\alpha\nu}-\bar{g}^{\alpha\beta}\tilde{h}^{\mu\nu}-\bar{g}^{\mu\nu}\tilde{h}^{\alpha\beta}\right),$$ and $\tilde{h}^{\mu\nu}:=h^{\mu\nu}-\frac{1}{2}\bar{g}^{\mu\nu}h.$ The crux of the above construction is that one must use the explicit form the linearized Einstein tensor in terms of the metric perturbation (or deviation from the $(A)dS$ background). This yields (\[iki\]) which is invariant under gauge transformations of the form $\delta h_{\mu \nu} = \bar \nabla_\mu \zeta_\nu +\bar \nabla_\nu \zeta_\mu$, but neither $K^{\mu\alpha\nu\beta}$ , nor the two-from current $\left(\bar{\text{\ensuremath{\xi}}}_{\nu}\nabla_{\beta}K^{\mu\alpha\nu\beta}-K^{\mu\beta\nu\alpha}\nabla_{\beta}\bar{\text{\ensuremath{\xi}}}_{\nu}\right)$ that appears in the right-hand side of (\[iki\]) are gauge invariant: under these transformations a boundary terms appears [@Emel_PRD_uzun]. So, even though the charge $Q(\ensuremath{\xi})$ is gauge invariant, the integrand defining the charge is not. The question is if one can find a way to make all this construction explicitly gauge invariant. The answer is not obvious because not every gauge-invariant physical quantity can be written explicitly gauge-invariant in local way. To achieve our goal of finding a fully gauge-invariant expression, here we shall provide another method of expressing $\bar{\text{\ensuremath{\xi}}}_{\nu}\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$ in such a way that one does not explicitly use the expression of $\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$, instead purely geometric considerations will be used such that the charges are expressed in terms of the linearized Riemann tensor. The formula, whose derivation will be given below, reads as $$Q\left(\bar{\xi}\right)=k\int_{\partial\bar{\Sigma}}d^{n-2}x\,\sqrt{\bar{\gamma}}\,\bar{\epsilon}_{\mu\nu}\left(R^{\nu\mu}\thinspace_{\beta\sigma}\right)^{\left(1\right)}\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}, \label{newcharge}$$ with the constant coefficient $$k=\frac{(n-1)(n-2)}{8(n-3)\Lambda G\Omega_{n-2}} \label{katsayi}$$ and $(R^{\nu\mu}\thinspace_{\beta\sigma})^{\left(1\right)}$ is the linearized part of the Riemann tensor about the $AdS$ background. All the barred quantities refer to the background spacetime $\bar{{\cal {M}}}$ whose boundary is $\partial\bar{{\cal {M}}}$. The Killing vector is $\bar{\xi}^{\sigma}$ and the antisymmetric tensor is $\bar{{\cal {F}}}^{\beta\sigma}:=\bar{\nabla}^{\beta}\bar{\xi}^{\sigma}$. $\bar{\Sigma}$ is a spatial hypersurface which is not equal to $\partial\bar{{\cal {M}}}$, hence $\bar{\Sigma}$ can have a boundary of its own which is $\partial\bar{\Sigma}$. Here $$\bar{\epsilon}_{\mu\nu}:=\frac{1}{2}\left(\bar{n}_{\mu}\bar{\sigma}_{\nu}-\bar{n}_{\nu}\bar{\sigma}_{\mu}\right),$$ where $\bar{n}_{\mu}$ is a normal one form on $\partial\bar{{\cal {M}}}$ and $\bar{\sigma}_{\nu}$ is the unit normal one form on $\partial\bar{\Sigma}$ and $\bar{\gamma}$ is the induced metric on the boundary. Derivation of the new formula ============================= Let us now provide the derivation of (\[newcharge\]): we start with the second Bianchi identity $$\nabla_{\nu}R_{\sigma\beta\mu\rho}+\nabla_{\sigma}R_{\beta\nu\mu\rho}+\nabla_{\beta}R_{\nu\sigma\mu\rho}=0,$$ multiplying with $g^{\nu\rho}$ and making use of the definition of the cosmological Einstein tensor ${\cal {G}}_{\beta}^{\nu}:=R_{\beta}^{\nu}-\frac{1}{2}R\delta_{\nu}^{\beta}+\Lambda\delta_{\nu}^{\beta}$, one arrives at $$\nabla_{\nu}{\cal{P}}^{\nu}\thinspace_{\mu\beta\sigma}=0,\label{divK}$$ where the ${\cal{P}}$ tensor reads $$\begin{gathered} \text{\ensuremath{{\cal{P}}}}^{\nu}\thinspace_{\mu\beta\sigma}:=R^{\nu}\thinspace_{\mu\beta\sigma}+\delta_{\sigma}^{\nu}\text{\ensuremath{{\cal {G}}}}_{\beta\mu}-\delta_{\beta}^{\nu}\text{\ensuremath{{\cal {G}}}}_{\sigma\mu}+\text{\ensuremath{{\cal {G}}}}_{\sigma}^{\nu}g_{\beta\mu}-\text{\ensuremath{{\cal {G}}}}_{\beta}^{\nu}g_{\sigma\mu}\\ +(\frac{R}{2}-\frac{\Lambda(n+1)}{n-1})(\delta_{\sigma}^{\nu}g_{\beta\mu}-\delta_{\beta}^{\nu}g_{\sigma\mu}).\end{gathered}$$ In the construction of this tensor we have used $\nabla_{\mu}{\cal {G}}^{\mu\nu}=0$ and $\nabla_{\mu}g^{\mu\nu}=0$ and defined it in such a way that its $AdS$ value vanishes. For any smooth metric, (\[divK\]) is valid [*identically*]{} without the use of the field equations. We can also express the ${\cal{P}}$ tensor in terms of the Weyl tensor as $$\begin{gathered} \text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma}=C^{\nu\mu\beta\sigma}-\frac{2(n-3)}{n-2}(\text{\ensuremath{{\cal {G}}}}^{\nu[\beta}g^{\sigma]\mu}+\text{\ensuremath{{\cal {G}}}}^{\mu[\sigma}g^{\beta]\nu})\\ +\frac{n-3}{n-1}(\frac{\Lambda n}{n-2}-\frac{R}{2})(g^{\nu\beta}g^{\mu\sigma}-g^{\mu\beta}g^{\nu\sigma}).\end{gathered}$$ Let $\text{\ensuremath{{\cal {F}}}}^{\beta\sigma}$ be a generic antisymmetric tensor. Then, contracting (\[divK\]) with $\text{\ensuremath{{\cal {F}}}}_{\beta\sigma}$ yields $$\nabla_{\nu}(\text{\ensuremath{{\cal {F}}}}_{\beta\sigma}\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})-\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma}\nabla_{\nu}\text{\ensuremath{{\cal {F}}}}_{\beta\sigma}=0,\label{eq:ddimensionalmainequation}$$ which is an exact equation. Let us now consider the metric perturbation which defines asymptotically $AdS$ spacetimes $$g_{\mu\nu}=\overline{g}_{\mu\nu}+ h_{\mu\nu},$$ where the background metric is $AdS$ and satisfies $$\bar{R}_{\alpha\beta\gamma\delta}=\frac{2\Lambda}{(n-1)(n-2)}\left(\overline{g}_{\alpha\gamma}\overline{g}_{\beta\delta}-\overline{g}_{\alpha\delta}\overline{g}_{\beta\gamma}\right),$$ together with Ricci tensor $\bar{R}_{\alpha\beta}=\frac{2\Lambda}{n-2}\overline{g}_{\alpha\beta}$ and the scalar curvature $\bar{R}=\frac{2\Lambda n}{n-2}$. For the AdS background we have $\bar{\text{\ensuremath{{\cal {G}}}}}^{\mu\nu}$= 0 and $\bar{\text{\ensuremath{{\cal{P}}}}}^{\nu\mu\beta\sigma}=0$ as already noted. Let us now consider the following particular anti-symmetric tensor $$\text{\ensuremath{{\cal {F}}}}_{\alpha\beta}:=\frac{1}{2}\left(\nabla_{\alpha}\xi_{\beta}-\nabla_{\beta}\text{\ensuremath{\xi}}_{\alpha}\right).$$ When $\text{\ensuremath{\xi}}$ is a background Killing vector one has $\text{\ensuremath{{\cal {F}}}}_{\alpha\beta}=\bar{\text{\ensuremath{{\cal {F}}}}}_{\alpha\beta}$. The linear order expansion of (\[eq:ddimensionalmainequation\]) reads $$\bar{\nabla_{\nu}}((\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})^{(1)}\bar{\text{\ensuremath{{\cal {F}}}}}_{\beta\sigma})-(\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})^{(1)}\bar{\nabla}_{\nu}\bar{\text{\ensuremath{{\cal {F}}}}}_{\beta\sigma}=0.\label{eq:ddimensionalmainequationlinear}$$ We now need to calculate the first order linearization of the ${\cal{P}}$ tensor which reads $$\begin{gathered} ({\cal{P}}^{\nu\mu\beta\sigma})^{\left(1\right)}=(R^{\nu\mu\beta\sigma})^{1}+2(\text{\ensuremath{{\cal {G}}}}^{\mu[\beta})^{(1)}\overline{g}^{\sigma]\nu}+2 (\text{\ensuremath{{\cal {G}}}}^{\nu[\sigma})^{(1)}\overline{g}^{\beta]\mu}\\ +(R)^{\left(1\right)}\overline{g}^{\mu[\beta}\overline{g}^{\sigma]\nu} +\frac{4\Lambda}{(n-1)(n-2)}(h^{\mu[\sigma}\overline{g}^{\beta]\nu}+ \overline{g}^{\mu[\sigma}{h}^{\beta]\nu}).\label{ktensorlinear}\end{gathered}$$ After some manipulations and using the identity $\bar{\nabla}_{\mu}\bar{\nabla}_{\nu}\text{\ensuremath{\bar{\xi}}}_{\rho}=\bar{R}^{\sigma}\thinspace_{\mu\nu\rho}\,\text{\ensuremath{\bar{\xi}}}_{\sigma}$, one arrives at $$(\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})^{(1)}\bar{\nabla}_{\nu}\bar{\text{\ensuremath{{\cal {F}}}}}_{\beta\sigma}=\frac{4\Lambda (n-3)}{(n-1)(n-2)}\bar{\text{\ensuremath{\xi}}}_{\lambda}(\text{\ensuremath{{\cal {G}}}}^{\lambda\mu})^{(1)},$$ then from (\[eq:ddimensionalmainequationlinear\]) we obtain the main expression $$\bar{\text{\ensuremath{\xi}}}_{\lambda}(\text{\ensuremath{{\cal {G}}}}^{\lambda\mu})^{(1)}=\frac{(n-1)(n-2)}{4\Lambda (n-3)}\bar{\nabla_{\nu}}\Big ((\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})^{(1)}\bar{\text{\ensuremath{{\cal {F}}}}}_{\beta\sigma}\Big).\label{eq:finallinearequation}$$ This proves our desired formula: without writing the linearized Einstein tensor explicitly in terms of the metric perturbation, we were able to express the conserved current as a boundary term involving the linearization of the Riemann and Einstein tensors as well as the Ricci scalar. To arrive at the total charge expression, we use the Stokes’ theorem and the resulting integral must be evaluated at spatial infinity. This simplifies the expression further: $\left(\text{\ensuremath{{\cal {G}}}}^{\nu\mu}\right)^{\left(1\right)}$ and the linearized scalar curvature vanishes at infinity. Moreover lowering the last two indices of the $(\text{\ensuremath{{\cal{P}}}}^{\nu\mu\beta\sigma})^{(1)}$ tensor one arrives at the charge expression (\[newcharge\]). Application to Kerr-AdS black holes {#application-to-kerr-ads-black-holes .unnumbered} ----------------------------------- As an application of our formula, let us consider the Kerr-$AdS$ black hole in four dimensions. One can take the solution to be in the Kerr-Schild form which reads $$ds^{2}=d\bar{s}^{2}+\frac{2M r }{\rho^2}\left(k_{\mu}dx^{\mu}\right)^{2},$$ where $ \rho^{2}=r^{2}+a^{2}\cos^{2}\theta$ and with the $AdS$ seed metric given as $$\begin{aligned} d\bar{s}^{2} & = & -\frac{\left(1-\frac{\Lambda r^{2}}{3}\right)\Delta_{\theta}dt^{2}}{\left(1+\frac{\Lambda a^{2}}{3}\right)}+\frac{\rho^{2}dr^{2}}{\left(1-\frac{\Lambda r^{2}}{3}\right)\left(r^{2}+a^{2}\right)}\nonumber \\ & + & \frac{\rho^{2}d\theta^{2}}{\Delta_{\theta}}+\frac{\left(r^{2}+a^{2}\right)\sin^{2}\theta d\phi^{2}}{\left(1+\frac{\Lambda a^{2}}{3}\right)},\end{aligned}$$ where $\Delta_{\theta} = 1+ \frac{\Lambda}{3} \cos^2 \theta$. The null vector $k_{\mu}$ is given by $$k_{\mu}dx^{\mu}=\frac{\Delta_{\theta}dt}{\left(1+\frac{\Lambda a^{2}}{3}\right)}+\frac{\rho^{2}dr}{\left(1-\frac{\Lambda r^{2}}{3}\right)\left(r^{2}+a^{2}\right)}-\frac{a \sin^{2}\theta d\phi}{\left(1+\frac{\Lambda a^{2}}{3}\right)}. \nonumber$$ Taking the Killing vector to be $ \bar \xi = ( -1,0,0,0)$, and $G=1$, the charge expression (\[newcharge\]) becomes $$E = \frac{3}{ 16 \pi \Lambda} \int_{S^2_\infty} d\Omega (R^{r t}\thinspace_{\beta\sigma})^{\left(1\right)}\bar{\nabla}^\beta\bar{\xi}^\sigma,$$ with $\sqrt{\bar \gamma} = \frac{r^2+ a^2 \cos^2\theta}{1+ \frac{\Lambda}{3} a^2}$. The integral is over a sphere at $r \rightarrow \infty$ which yields the answer $$E=\frac{M}{\left(1+\frac{\Lambda a^{2}}{3}\right)^{2}}.$$ Similarly for the Killing vector $\bar \xi = ( 0,0,0,1)$ one finds the angular momentum of the black hole as $$J=\frac{aM}{\left(1+\frac{\Lambda a^{2}}{3}\right)^{2}}.$$ These relations satisfy $E=J/a$ and they match the ones computed in [@Kanik]. Relation of the new formula with the Abbott-Deser formula ========================================================= Let us derive the explicit connection between the AD expression (\[iki\]) and the one we have given here (\[eq:finallinearequation\]). Going from the former to the latter is extremely difficult, one needs judicious additions of terms that vanish, so we shall start from our expression and expand it to find out the relation. For this purpose, let us start from the linearized form of the $(2,2)$ background tensor $$\begin{gathered} \text{\ensuremath{{\cal {P}}}}^{\nu\mu}\thinspace_{\beta\sigma}:=R^{\nu\mu}\thinspace_{\beta\sigma}+\delta_{\sigma}^{\nu}(R_{\beta}^{\mu})^{\left(1\right)}-\delta_{\beta}^{\nu}(R_{\sigma}^{\mu})^{\left(1\right)}+\delta_{\beta}^{\mu}(R_{\sigma}^{\nu})^{\left(1\right)}\\ -\delta_{\sigma}^{\mu}(R_{\beta}^{\nu})^{\left(1\right)}-\frac{1}{2}(R)^{\left(1\right)}(\delta_{\sigma}^{\nu}\delta_{\beta}^{\mu}-\delta_{\beta}^{\nu}\delta_{\sigma}^{\mu}),\label{Ptensorlinear}\end{gathered}$$ which, due to the symmetries, yields $$\begin{gathered} (\text{\ensuremath{{\cal {P}}}}^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}=\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}(R^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}\\ +2\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}(R_{\sigma}^{\mu})^{(1)}-2\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\mu}(R_{\sigma}^{\nu})^{(1)}-\bar{\text{\ensuremath{{\cal {F}}}}}^{\mu\nu}(R)^{(1)}.\label{eq:newformulationtoad}\end{gathered}$$ Let us compute the right-hand side of the last expression term by term. The first term can be written as $$\begin{gathered} \bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}(R^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}=\frac{1}{2}\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}\left (-\bar{R}^{\nu}\thinspace_{\lambda\beta\sigma}h^{\lambda\mu}+\bar{R}^{\mu}\thinspace_{\lambda\beta\sigma}h^{\lambda\nu}\right.\\ \left.+\bar{g}^{\lambda\mu}(R^{\nu}\thinspace_{\lambda\beta\sigma})^{(1)}-\bar{g}^{\lambda\nu}(R^{\mu}\thinspace_{\lambda\beta\sigma})^{(1)}\right).\end{gathered}$$ Using the first order linearized Riemann tensor $$(R^{\nu}\thinspace_{\lambda\beta\sigma})^{(1)}=\bar{\nabla}_{\beta}(\Gamma_{\lambda\sigma}^{\nu})^{\left(1\right)}-\bar{\nabla}_{\sigma}(\Gamma_{\lambda\beta}^{\nu})^{\left(1\right)},$$one finds $$\begin{gathered} \bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}(R^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}=\frac{2\Lambda}{(n-1)(n-2)}(\bar{\text{\ensuremath{{\cal {F}}}}}^{\mu\sigma}h_{\sigma}^{\nu}-\bar{\text{\ensuremath{{\cal {F}}}}}^{\nu\sigma}h_{\sigma}^{\mu})\\ +\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}\bar{\nabla}_{\beta}(\bar{\nabla}^{\mu}h_{\sigma}^{\nu}-\bar{\nabla}^{\nu}h_{\sigma}^{\mu}).\end{gathered}$$ We can rewrite this as follows: $$\begin{gathered} \bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}(R^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}=\frac{2\Lambda}{(n-1)(n-2)}\left(h_{\sigma}^{\nu}\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\sigma}\right.\\ \left.-h_{\sigma}^{\mu}\bar{\nabla}^{\nu}\bar{\text{\ensuremath{\xi}}}^{\sigma}+(n-1)\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{\mu}h_{\sigma}^{\nu}-(n-1)\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{\nu}h_{\sigma}^{\mu}\right)\\ +\bar{\nabla}_{\beta}\left(\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}(\bar{\nabla}^{\mu}h_{\sigma}^{\nu}-\bar{\nabla}^{\nu}h_{\sigma}^{\mu})\right)\end{gathered}$$ Now, we can compute the second term in (\[eq:newformulationtoad\]) as $$2\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}(R_{\sigma}^{\mu})^{(1)}=2\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}\Big (\bar{g}^{\lambda\mu}(R_{\lambda\sigma})^{\left(1\right)}-h^{\lambda\mu}\bar{R}_{\lambda\sigma})\Big)$$ where $$(R_{\lambda\sigma})^{\left(1\right)}=\frac{1}{2}(\bar{\nabla}_{\rho}\bar{\nabla}_{\lambda}h_{\sigma}^{\rho}+\bar{\nabla}_{\rho}\bar{\nabla}_{\sigma}h_{\lambda}^{\rho}-\bar{\nabla}_{\rho}\bar{\nabla}^{\rho}h_{\lambda\sigma}-\bar{\nabla}_{\lambda}\bar{\nabla}_{\sigma}h)$$ Then we have $$\begin{gathered} 2\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}(R_{\sigma}^{\mu})^{(1)}=\frac{2\Lambda}{(n-1)(n-2)}\Bigl(\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{\nu}h_{\sigma}^{\mu}-h\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\nu}\\ -\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{\mu}h_{\sigma}^{\nu}+(n-2)(h_{\sigma}^{\mu}\bar{\nabla}^{\nu}\bar{\text{\ensuremath{\xi}}}^{\sigma}+\bar{\text{\ensuremath{\xi}}}^{\nu}\bar{\nabla}_{\sigma}h^{\sigma\mu}-\bar{\text{\ensuremath{\xi}}}^{\nu}\bar{\nabla}^{\mu}h)\Bigr) \\ +\bar{\nabla}_{\rho}\left(\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}(\bar{\nabla}^{\mu}h_{\sigma}^{\rho}+\delta_{\sigma}^{\rho}\bar{\nabla}_{\beta}h^{\beta\mu}-\bar{\nabla}^{\rho}h_{\sigma}^{\mu}-\delta_{\sigma}^{\rho}\bar{\nabla}^{\mu}h)\right).\end{gathered}$$ Finally we can compute the last term in (\[eq:newformulationtoad\]) as $$\begin{gathered} \bar{\text{\ensuremath{{\cal {F}}}}}^{\mu\nu}(R)^{(1)}=\frac{2\Lambda}{(n-1)(n-2)}\Bigl(-\bar{\text{\ensuremath{\xi}}}^{\mu}\bar{\nabla}_{\sigma}h^{\sigma\nu}+\bar{\text{\ensuremath{\xi}}}^{\mu}\bar{\nabla}^{\nu}h\\ +\bar{\text{\ensuremath{\xi}}}^{\nu}\bar{\nabla}_{\sigma}h^{\sigma\mu}-\bar{\text{\ensuremath{\xi}}}^{\nu}\bar{\nabla}^{\mu}h-(n-1)h\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\nu}\Bigr) \\ \bar{\nabla}_{\rho}\left(\bar{\text{\ensuremath{{\cal {F}}}}}^{\mu\nu}(\bar{\nabla}_{\sigma}h^{\rho\sigma}-\bar{\nabla}^{\rho}h)\right)\\\end{gathered}$$ Collecting all the pieces together, we have the following expression $$\begin{gathered} (\text{\ensuremath{{\cal {P}}}}^{\nu\mu}\thinspace_{\beta\sigma})^{(1)}\bar{\text{\ensuremath{{\cal {F}}}}}^{\beta\sigma}=\frac{4\Lambda(n-3)}{(n-1)(n-2)}\Bigl(h_{\sigma}^{[\mu}\bar{\nabla}^{\nu]}\bar{\text{\ensuremath{\xi}}}^{\sigma}+\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{[\mu}h_{\sigma}^{\nu]}\\ +\bar{\text{\ensuremath{\xi}}}^{[\nu}\bar{\nabla}_{\sigma}h^{\mu] \sigma}+\bar{\text{\ensuremath{\xi}}}^{[\mu}\bar{\nabla}^{\nu]}h+\frac{1}{2}h\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\nu}\Bigr) \\ +\bar{\nabla}_{\rho}\left(\bar{\text{\ensuremath{{\cal {F}}}}}^{\sigma\nu}(\bar{\nabla}^{\mu}h_{\sigma}^{\rho}+\delta_{\sigma}^{\rho}\bar{\nabla}_{\beta}h^{\beta\mu}-\bar{\nabla}^{\rho}h_{\sigma}^{\mu}-\delta_{\sigma}^{\rho}\bar{\nabla}^{\mu}h)\right.\\ \left. -\frac{1}{2}\bar{\text{\ensuremath{{\cal {F}}}}}^{\mu\nu}(\bar{\nabla}_{\sigma}h^{\rho\sigma}-\bar{\nabla}^{\rho}h)+ \bar{\text{\ensuremath{{\cal {F}}}}}^{\rho\sigma}\bar{\nabla}^{\mu}h_{\sigma}^{\nu} - (\mu \leftrightarrow \nu) \right).\end{gathered}$$ from (\[eq:finallinearequation\]), we can write $$\begin{gathered} \bar{\text{\ensuremath{\xi}}}_{\lambda}(\text{\ensuremath{{\cal {G}}}}^{\lambda\mu})^{(1)}=\bar{\nabla_{\nu}}\Bigl(h_{\sigma}^{[\mu}\bar{\nabla}^{\nu]}\bar{\text{\ensuremath{\xi}}}^{\sigma}+\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{[\mu}h_{\sigma}^{\nu]}\\ +\bar{\text{\ensuremath{\xi}}}^{[\nu}\bar{\nabla}_{\sigma}h^{\mu] \sigma}+\bar{\text{\ensuremath{\xi}}}^{[\mu}\bar{\nabla}^{\nu]}h+\frac{1}{2}h\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\nu}\Bigr) \\ +\frac{(n-1)(n-2)}{4\Lambda (n-3)}\bar{\nabla_{\nu}}\bar{\nabla}_{\rho}\Bigl( -\frac{1}{2}\bar{\nabla}^{\mu}\bar{\text{\ensuremath{\xi}}}^{\nu}(\bar{\nabla}_{\sigma}h^{\rho\sigma}-\bar{\nabla}^{\rho}h)\\ +\bar{\nabla}^{\sigma}\bar{\text{\ensuremath{\xi}}}^{\nu} (\bar{\nabla}^{\mu}h_{\sigma}^{\rho}+\delta_{\sigma}^{\rho}\bar{\nabla}_{\beta}h^{\beta\mu}-\bar{\nabla}^{\rho}h_{\sigma}^{\mu} -\delta_{\sigma}^{\rho}\bar{\nabla}^{\mu}h)\\ +\bar{\nabla}^{\rho}\bar{\text{\ensuremath{\xi}}}^{\sigma}\bar{\nabla}^{\mu}h_{\sigma}^{\nu} - (\mu \leftrightarrow \nu)\Bigr). \label{lasteqn}\end{gathered}$$ The first two lines yield the AD expression as given in [@Deser_Tekin] while the remaining part is of the form $\bar{\nabla}_{\nu}\bar{\nabla}_{\rho} Q^{\rho \mu \nu}[h]$. Integrating the above expression on a spatial hypersurface, after making use of the Stokes’ theorem, the first two lines give the AD charge, while the other part having two derivatives remain a [*total divergence*]{} on the boundary of the hypersurface, vanishes since the boundary of the boundary is nil. Note that this equivalence does not work in 3 spacetime dimensions and for the asymptotically flat spacetimes. It is important to recognize the following: under gauge transformations, the left-hand side of (\[lasteqn\]) is gauge invariant and so is the right-hand side. But, it is easy to see that the first two lines are gauge-invariant only up to a boundary term. Full gauge invariance is recovered with the additional parts. The details of this discussion were given in [@Emel_PRD_uzun]. Conclusions =========== We have given a conserved charge expression in Einstein’s theory for asymptotically $(A)dS$ spacetimes which is directly written in terms of the linearized Riemann tensor and an anti-symmetric tensor that appears as the potential of the Killing vector on the boundary of the spatial hypersurface. The expression is explicitly gauge-invariant as the up-up-down-down linearized Riemann tensor is gauge invariant under small variations $\delta h_{\mu \nu} = \bar \nabla_\mu \zeta_\nu +\bar \nabla_\nu \zeta_\mu$. Our construction started from the second Bianchi Identity on the Riemann tensor and as such, the final expression of conserved charges is valid for $n >3$ and not valid for the case of three dimensions. A naive extension of this construction to generic gravity theories as discussed in [@Deser_Tekin] is not that obvious and was carried out [@Emel_PRD_uzun] after this work appeared. Once a higher order theory’s field equations are given one can work out a similar computation for these theories and the coefficient $k$ in (\[katsayi\]) receives corrections from the higher curvature terms. It would be interesting to relate our construction to the one given in [@kim]. [1]{} R. Arnowitt, S. Deser and C.W. Misner, Canonical variables for general relativity, Phys. Rev. **117**, 1595 (1960); The dynamics of general relativity, Gen. Rel. Grav. **40**, 1997 (2008). L.F. Abbott and S. Deser, Stability of gravity with a cosmological constant, Nucl. Phys. B **195**, 76 (1982). S. Deser and B. Tekin, Energy in generic higher curvature gravity theories, Phys. Rev. D **67**, 084009 (2003); Gravitational energy in quadratic curvature gravities, Phys. Rev. Lett. **89**, 101101 (2002). H. Adami, M. R. Setare, T. C. Sisman, and B. Tekin, Conserved charges in extended theories of gravity, arXiv:1710.07252 [\[]{}hep-th[\]]{}. G. Compere and A. Fiorucci, Advanced Lectures in General Relativity, arXiv:1801.07064 \[hep-th\]. E. Altas and B. Tekin, A new approach to conserved charges of generic gravity in AdS,” arXiv:1811.11525 \[hep-th\]. S. Deser, I. Kanik and B. Tekin, Conserved charges of higher D Kerr-AdS spacetimes, Class. Quant. Grav.  [**22**]{}, 3383 (2005). W. Kim, S. Kulkarni and S. H. Yi, Quasilocal Conserved Charges in a Covariant Theory of Gravity, Phys. Rev. Lett.  [**111**]{}, no. 8, 081101 (2013). Erratum: \[Phys. Rev. Lett.  [**112**]{}, no. 7, 079902 (2014)\].

ArXiv

--- abstract: 'A link of an isolated singularity of a two-dimensional semialgebraic surface in ${\mathbb R}^4$ is a knot (or a link) in $S^3$. Thus the ambient Lipschitz classification of surface singularities in ${\mathbb R}^4$ can be interpreted as a bi-Lipschitz refinement of the topological classification of knots (or links) in $S^3$. We show that, given a knot $K$ in $S^3$, there are infinitely many distinct ambient Lipschitz equivalence classes of outer metric Lipschitz equivalent singularities in ${\mathbb R}^4$ with the links topologically equivalent to $K$.' address: - 'Dept Matemática, Universidade Federal do Ceará (UFC), Campus do Picici, Bloco 914, Cep. 60455-760. Fortaleza-Ce, Brasil' - 'Dept Mathematics, Purdue University, 150 N. University Street, West Lafayette, IN 47907-2067, USA ' author: - 'Lev Birbrair\*' - 'Andrei Gabrielov\*\*' title: 'Surface singularities in ${\mathbb R}^4$: first steps towards Lipschitz knot theory' --- [^1] [^2] Introduction ============ There are three kinds of equivalence relations in Lipschitz Geometry of Singularities. One equivalence relation is bi-Lipschitz homeomorphism (of the germs at the origin) of singular sets with respect to the inner metric, where the distance between two points of a set $X$ is defined as infimum of the lengths of paths inside $X$ connecting the two points. The second equivalence relation, outer Lipschitz equivalence, is bi-Lipschitz homeomorphism with respect to the outer metric, where the distance is defined as the distance between the points in the ambient space. A set $X$ is called normally embedded if its inner and outer metrics are equivalent. In [@BG], we considered the third equivalence relation, Lipschitz ambient equivalence. Two germs $X$ and $Y$ of semialgebraic sets at the origin of ${\mathbb R}^n$ are called Lipschitz ambient equivalent if there exists a germ of a bi-Lipschitz homeomorphism $h$ of $({\mathbb R}^n,0)$ such that $Y=h(X)$. In particular, such sets $X$ and $Y$ are outer Lipschitz equivalent. Two outer Lipschitz equivalent sets are always inner Lipschitz equivalent, but can be ambient topologically non-equivalent (see Neumann-Pichon [@NP]). Let $X$ and $Y$ be two semialgebraic surface singularities (two-dimensional germs at the origin) in ${\mathbb R}^n$ which are outer Lipschitz equivalent. Suppose also that $X$ and $Y$ are topologically ambient equivalent. Does it imply that the sets $X$ and $Y$ are Lipschitz ambient equivalent? It seems plausible that the answer is “yes” when $n\ge 5$, or when $X$ and $Y$ are normally embedded. However, examples in [@BG] show that the answer may be “no” when $n=3$ or $n=4$. One class of examples in ${\mathbb R}^3$ and ${\mathbb R}^4$ is based on the theorem of Sampaio [@S]: Lipschitz ambient equivalence of two sets implies Lipschitz ambient equivalence of their tangent cones. Thus any two sets with topologically ambient non-equivalent tangent cones cannot be Lipschitz ambient equivalent. The case $n=4$ is especially interesting, as in that case the link of a two-dimensional germ $X$ in ${\mathbb R}^4$ is a knot (or a link) in $S^4$, and the arguments are based on the knot theory. For a given surface $X\subset {\mathbb R}^3$ there are finitely many distinct Lipschitz ambient equivalence classes of the surfaces which are topologically ambient equivalent and outer Lipschitz equivalent to $X$. However, there may be infinitely many such Lipschitz ambient equivalence classes for a surface in ${\mathbb R}^4$. Moreover, a more delicate argument, based on the “bridge construction” below, provides infinitely many distinct Lipschitz ambient equivalence classes of surfaces which are topologically ambient equivalent to a given surface $X\subset{\mathbb R}^4$ and belong to the same outer metric Lipschitz equivalency class, even when each of these surfaces has a tangent cone consisting of a single ray. ![The links of the surfaces $X_1$ and $X_2$ in Example 1, and of their tangent cones.[]{data-label="fig:example2b"}](example2b.eps){width="4in"} ![The links of the surfaces $X_1$ and $X_2$ in Example 2, and of their tangent cones.[]{data-label="fig:knot1bt"}](knot1bt.eps){width="4in"} Examples in ${\mathbb R}^3$ and ${\mathbb R}^4$ based on Sampaio’s theorem ========================================================================== [**Example 1**]{} (see [@BG]). Let $X_1$ and $X_2$ be two surfaces in ${\mathbb R}^3$ with the links at the origin shown in Fig. \[fig:example2b\]$a$ and Fig. \[fig:example2b\]$b$, and the links of their tangent cones at the origin shown in Fig. \[fig:example2b\]$a'$ and Fig. \[fig:example2b\]$b'$. The tangency exponent of the arcs $\gamma_+$ and $\gamma_-$ in Fig. \[fig:example2b\]$a$ and Fig. \[fig:example2b\]$b$ is $\alpha>1$, and the tangency exponent of the arcs $\zeta_+$ and $\zeta_-$ is $\beta>1$. Thus the arcs $\gamma_+$ and $\gamma_-$ correspond to a single ray $\gamma$ of the tangent cone, and the arcs $\zeta_+$ and $\zeta_-$ correspond to a single ray $\zeta$. One can define $X_1$ and $X_2$ by explicit semialgebraic formulas. Both surfaces $X_1$ and $X_2$ are ambient topologically equivalent to a cone over a circle and bi-Lipschitz equivalent with respect to the outer metric, but not ambient Lipschitz equivalent by Sampaio’s theorem, since their tangent cones are not ambient topologically equivalent. [**Example 2**]{} (see [@BG]). Let $X_1$ and $X_2$ be two surfaces in ${\mathbb R}^4$ with the links at the origin shown in Fig. \[fig:knot1bt\]$a$ and Fig. \[fig:knot1bt\]$b$, and the links of their tangent cones at the origin shown in Fig. \[fig:knot1bt\]$a'$ and Fig. \[fig:knot1bt\]$b'$. The tangency exponent of the arcs $\gamma_+$ and $\gamma_-$ in Fig. \[fig:knot1bt\]$a$ and Fig. \[fig:knot1bt\]$b$ is $\alpha>1$, thus the arcs $\gamma_+$ and $\gamma_-$ correspond to a single ray $\gamma$ of the tangent cone. One can define $X_1$ and $X_2$ by explicit semialgebraic formulas. Both surfaces $X_1$ and $X_2$ are ambient topologically equivalent to a cone over a circle and bi-Lipschitz equivalent with respect to the outer metric, but not ambient Lipschitz equivalent by Sampaio’s theorem, since their tangent cones are not ambient topologically equivalent. Bridge construction =================== A $(q,\beta)$-*bridge* is the set $A_{q,\beta}=T_+\cup T_-\subset {\mathbb R}^4$ where $1<\beta<q$ and $$T_\pm=\left\{ 0\le t\le 1,\; -t^\beta\le x\le t^\beta,\; y=\pm t^q,\; z=0\right\}.$$ Its link is shown in Fig. \[fig:bridge\] (left). A *broken $(q,\beta)$-bridge* $B_{q,\beta}$ is obtained from $A_{q,\beta}$ by the *saddle operation*, removing from $T_\pm$ two $p$-Hölder triangles $\left\{t\ge 0,\; |x|\le t^p,\; y=\pm t^q,\; z=0\right\}$ where $p>q$, and replacing them by two $q$-Hölder triangles $\left\{ 0\le t\le 1,\; x=\pm t^p,\; |y|\le t^q,\; z=0\right\}.$ Its link is shown in Fig. \[fig:bridge\] (right). We call $(q,\beta)$-bridge any surface outer Lipschitz equivalent to $A_{q,\beta}$. It was shown in [@BG] that ambient Lipschitz equivalence $h:X\to Y$ of two surfaces in ${\mathbb R}^4$ maps a $(q,\beta)$-bridge in $X$ to a $(q,beta)$-bridge in $Y$, and that the two surfaces remain ambient Lipschitz equivalent when their $(q,\beta)$-bridges are replaced by the broken $(q,\beta)$-bridges. \[broken\]*Our definition of a broken bridge is slightly different from the definition in Example 4 of [@BG], where it was defined with $p<q$. Condition $p>q$ makes the “broken bridge” operation invertible: *two surface germs with the same $(q,\beta)$-bridge are ambient Lipschitz equivalent if and only if they are ambient Lipschitz equivalent after the bridge is broken (with the same $p>q$)*. Note that this invertibility is never used here or in [@BG].* ![The links of a $(q,\beta)$-bridge $A_{q,\beta}$ and a broken $(q,\beta)$-bridge $B_{q,\beta}$.[]{data-label="fig:bridge"}](bridge3.eps){width="4.5in"} ![The link of the surface $G$ in Example 3.[]{data-label="fig:G"}](G.eps){width="4.5in"} ![The link of the surface $H$ in Example 3.[]{data-label="fig:H"}](H.eps){width="4.5in"} [**Example 3**]{} (see [@BG]). The common boundary of $A_{q,\beta}$ and $B_{q,\beta}$ consists of the four arcs $\left\{0\le t\le 1,\; x=\pm t^\beta,\; y=\pm t^q,\; z=0\right\}$ shown as $m,\,n,\,m',\,n'$ in Fig. \[fig:bridge\]. Let $G \subset {\mathbb R}^4$ be a semialgebraic surface containing $A_{q,\beta}$ and bounded by the four straight line segments $\{0\le t\le 1,\; \pm x =\pm y = t,\;z=0\}$ (see Fig. \[fig:G\] where the boundary arcs of $G$ are shown as $M,\,N,\,M',\,N'$). Let $H$ be the surface obtained from $G$ by replacing the bridge $A_{\beta,q}$ by the broken bridge $B_{\beta,q}$ (see Fig. \[fig:H\]). Consider two topologically trivial knots $K$ and $L$ in the hyperplane $\{t=1\}\subset {\mathbb R}^4_{x,y,z,t}$ as shown in Fig. \[fig:knot2\]a and Fig. \[fig:knot2\]b. Each of these two knots contains the curve $g=G\cap\{t=1\}$, We define the surface $X$ as the union of $G$ and a straight cone over $K\setminus g$, and the surface $Y$ as the union of $G$ and a straight cone over $L\setminus g$. ![The links of the surfaces $X$ and $Y$ in Example 3.[]{data-label="fig:knot2"}](knot2aGH.eps){width="4.5in"} ![The links of the surfaces $X'$ and $Y'$ in Example 3.[]{data-label="fig:knot2b"}](knot2bGH.eps){width="4.5in"} ![The links of the surfaces $X$ and $Y$ with an extra knot attached.[]{data-label="fig:knot3"}](knot3MN.eps){width="4.5in"} ![The link of the surface $X$ in Theorem \[CC\].[]{data-label="doubleknot"}](doubleknot.eps){width="4.5in"} \[thm:ex3\] (see [@BG] Theorem 3.2.) The germs of the surfaces $X$ and $Y$ at the origin are outer Lipschitz equivalent, topologically ambient equivalent, but not Lipschitz ambient equivalent. This is proved by replacing the $(q,\beta)$-bridges in $X$ and $Y$ by the broken $(q,\beta)$-bridges, resulting in the new surfaces $X'$ and $Y'$, shown in Fig. \[fig:knot2b\]. The link of $X'$ consists of two unlinked circles while the two circles in the link of $Y'$ are linked. Thus $X'$ and $Y'$ are not topologically ambient equivalent, which implies that $X$ and $Y$ are not Lipschitz ambient equivalent. *Notice that the tangent cones of both surfaces $X$ and $Y$ in Example 3 are ambient topologically equivalent to a cone over two unknotted circles, pinched at one point. Thus Sampaio’s theorem does not apply, and we need the bridge construction in this example. Notice also that the bridge construction employed in this example allows one to construct examples of outer bi-Lipschitz equivalent, topologically ambient equivalent but Lipschitz ambient non-equivalent surface singularities in $\mathbb{R}^4$ with the tangent cones as small as a single ray.* The surfaces $X$ and $Y$ in Example 3 differ by a “twist” of the $(q,\beta)$-bridge, which can be extended to a homeomorphism of the ambient space, but not to a bi-Lipschitz homeomorphism. One can iterate such a twist to obtain infinitely many Lipschitz ambient non-equivalent surfaces. On can also attach an additional knot to the links of both surfaces $X$ and $Y$ (see Fig. \[fig:knot3\]). This yields the following “universality” result (see [@BG] Theorem 4.1). \[thm:main\] For any semialgebraic surface germ $S \subset {\mathbb R}^4$ there exist infinitely many semialgebraic surface germs $X_i \subset {\mathbb R}^4$ such that $1) \ $ For all $i$, the germs $X_i$ are topologically ambient equivalent to $S$; $2) \ $ All germs $X_i$ are outer Lipschitz equivalent; $3) \ $ The tangent cones of all germs $X_i$ at the origin are ambient topologically equivalent; $4) \ $ For $ i \neq j$ the germs $X_i$ and $X_j$ are not Lipschitz ambient equivalent. Other versions of universality can be formulated. In particular, the following theorem is proved in [@BB]. \[CC\] For each knot $K$ there exists a semialgebraic surface $X \subset {\mathbb R}^4$ with a singular point at the origin, such that $1) \ $ The germ of $X$ at the origin has one $(q,\beta)$-bridge. $2) $ The link of $X$ at the origin is a trivial knot. $3) $ The link of $S(X)$ at zero is a two component link, such that each component of this link is the knot $K$. This result is illustrated in Fig. \[doubleknot\]. Notice that the tangent cones of the surface $X$ in Fig. \[doubleknot\] are not equivalent for non-equivalent knots $K$, so this result could be also obtained by Sampaio’s theorem. [99]{} L. Birbrair and Brandenbursky. Lipschitz geometry of singularities (in preparation), 2019. L. Birbrair and A. Gabrielov. Ambient Lipschitz equivalence of real surface singularities. IMRN, https://doi.org/10.1093/imrn/rnx328, 2018. A. Fernandes. Topological equivalence of complex curves and bi-Lipschitz maps. The Michigan Mathematical Journal, Michigan, v. 51, n.3, p. 593-606, 2003. W. Neumann and A. Pichon, Lipschitz geometry does not determine embedded topological type, Journal of Singularities, v 10 , 2014 , p. 225-234. P. Pham and B. Teissier, Saturation Lipschitzienne d’une algèbre analytique complexe et saturation de Zariski. Prépublication Ecole Polytechnique 1969. J. E. Sampaio. Bi-Lipschitz homeomorphic subanalytic sets have bi-Lipschitz homeomorphic tangent cones. Selecta Mathematica-New Series, v. 22, p. 553-559, [^1]: \*Research supported under CNPq 302655/2014-0 grant and by Capes-Cofecub [^2]: \*\*Research supported by the NSF grant DMS-1665115

ArXiv

--- abstract: 'In the scope of the Physics Beyond Colliders studies, the Gamma-Factory initiative proposes the use of partially stripped ions as a driver of a new type of high-intensity photon source in CERN Large Hadron Collider (LHC). In 2018, the LHC accelerated and stored partially stripped $^{208}\text{Pb}^{81+}$ ions for the first time. The collimation system efficiency recorded during this test was found to be prohibitively low, so that only a very low-intensity beam could be stored without the risk of triggering a beam dump when regular, minor beam losses occur. The worst losses were localised in the dispersion suppressor of the betatron-cleaning insertion. This article presents an analysis to understand in detail the source of these losses. Based on this understanding, possible mitigation measures that could significantly improve the cleaning efficiency and enable regular operation with partially-stripped ions in the future are developed.' author: - 'A. Gorzawski[^1,2^]{}' - 'A. Abramov[^1,3^]{}' - 'R. Bruce[^1^]{}' - 'N. Fuster-Martínez[^1^]{}' - 'M. Krasny[^1,4^]{}' - 'J. Molson[^1^]{}' - 'S. Redaelli[^1^]{}' - 'M. Schaumann[^1^]{}' bibliography: - 'references.bib' title: Collimation of partially stripped ions in the CERN Large Hadron Collider --- [^1] [^2] Introduction {#sec:intro} ============ The CERN Large Hadron Collider (LHC) is designed to provide proton collisions at an energy of 7 TeV per beam and heavy-ion collisions at the equivalent magnetic rigidity [@lhcdesignV1]. It consists of eight arcs and eight straight Insertion Regions (IRs). The accelerated beams collide in four out of the eight IRs, where the particle-physics experiments are located. The baseline design of the LHC includes operation with protons and lead ion beams ([$^{208}$Pb$^{82+}$ ]{}), but other beam particle species have also been considered [@YR_WG5_2018; @Schaumann:IPAC2018-MOPMF039]. The Physics Beyond Colliders initiative [@alemany2019summary] is a dedicated research campaign steered by CERN, focused on exploring alternative options to colliders for future particle-physics experiments. One of the projects under consideration is the Gamma Factory (GF), a design study for a novel type of light source [@krasny2015gamma]. The goal of the study group is to explore the possibilities to use the LHC beams for creating high-intensity, high-energy photon beams that can serve many applications, spanning from atomic to high-energy physics. The concept relies on using partially stripped ions (PSI) beams in the LHC as a driver. PSI retain one or more bound electrons. To produce photons, the remaining electrons in PSI are exited using a laser. The energy of the photons emitted during the spontaneous de-excitation of the excited atomic states is proportional to the square of the Lorentz factor of the ion beam, which allows photon energies of up to in the LHC. While the proof-of-principle experiment for the GF is proposed to operate at the Super Proton Synchrotron (SPS) at a beam energy of  [@Krasny:2690736], the ultimate implementation is intended to operate at the LHC’s top energy of . The GF depends on the acceleration and storage of PSI beams in CERN’s accelerator complex. PSI at varying states of ionisation are routinely used at CERN during the different stages of acceleration of the typical lead or argon beams in the injectors [@lhcdesignV3; @kuchler14; @Arduini:308372]. However, the LHC was never used to accelerate PSI beams. In 2018, the first operational tests with PSI beams in the LHC were performed with the goal of studying the beam lifetime and characterising the beam losses for such a beams [@Schaumann_2019_ipac_part_strip_ions; @Schaumann:2670544]. During one dedicated machine development (MD) session, PSI beams were successfully injected, accelerated, and stored. At the same time, it was found that the current LHC configuration poses a critical limitation on PSI operation. The collimation-system efficiency recorded during this test was found to be orders of magnitude worse than in standard operation and hence prohibitively low, effectively imposing an intensity limit. The worst losses were localised in the dispersion suppressor (DS) of the betatron-cleaning insertion. These findings clearly put in question the overall feasibility to operate the LHC with PSI beams of sufficient intensities for a future GF facility. In this article we study the underlying physical processes responsible for the worsening in collimation efficiency with PSI beams and study possible mitigation measures. In Section \[sec:lhc-recap\] we provide a brief recap of the LHC operation, collimation, and beam loss limitations. Section \[sec:experiments\] describes the experiments performed in the LHC using PSI beams, as well as the encountered limitations. Section \[sec:simulations\] describes the interpretation of the observed losses with PSI beams and then describes the available simulation tools for the case of partially stripped ions. In the same section, the measured loss maps are compared against results from simulations. Finally, in Section \[sec:mitigations\], different mitigation strategies for the found limitations are outlined and investigated, including a new DS collimator, crystal collimation, or an orbit bump. LHC collimation and beam losses {#sec:lhc-recap} =============================== LHC collimation system ---------------------- In the LHC, the proton-beam stored energy in Run 2 (2015-2018) at exceeded , approaching the design value of planned for the operation at . Uncontrolled beam losses of even a tiny fraction of the full beam could cause a superconducting magnet to quench or even cause material damage of exposed accelerator components. Therefore, a multi-stage collimation system is installed. It consists of more than 100 movable devices, installed to provide beam cleaning and passive protection against beam losses during regular operation and accidents [@lhcdesignV1; @assmann05chamonix; @assmann06; @bruce14_PRSTAB_sixtr; @bruce15_PRSTAB_betaStar; @valentino17_PRSTAB]. The betatron-halo cleaning is done by a three-stage collimator hierarchy in IR7 while off-momentum cleaning is done in IR3 by a similar system. Collimators for local triplet-magnet protection are located in all experimental IRs (IR1, IR2, IR5 and IR8) and, in addition, collimators for the physics-debris cleaning are installed in the high-luminosity experiments in IR1 and IR5. In the three-stage collimation system of the LHC, the primary collimators (TCP) are the devices closest to the beam of the whole machine and their purpose is to intercept any halo particles drifting out to large amplitudes [@redaelli_coll]. Particles that are not absorbed by the TCPs, but scattered to larger amplitudes should be caught by the secondary collimators (TCS), which are designed to intercept the secondary beam halo. As a third stage that uses shower absorber collimators (TCLAs) is in place to intercept the tertiary halo and the products of hadronic showers, leaking from the TCSs. Additional tertiary collimators (TCT) are placed around the experimental insertions. Figure \[fig:collimators-around-LHC\] illustrates the collimator locations around the LHC. It is important to note that beam-halo particles interacting with the TCPs are not always deflected onto the TCSs; in some cases they can escape the collimation insertion and complete further revolutions around the machine before being disposed of at collimators. Beam particles with momentum offsets escaping the collimation section can be lost on the cold aperture in the dispersion suppressor immediately downstream, where the rising dispersion affects their trajectories [@bruce14ipac_DS_coll]. The DS in IR7 is thus the main bottleneck for beam-halo losses in the LHC and the amount of local losses in the DS may impose limitations on the total achievable intensity, in particular for heavy-ion operation [@epac2004; @hermes16_ion_quench_test; @hermes16_nim] ![Conceptual sketch of the layout of LHC collimators around the ring. Collimators are located mainly in IR3 and IR7, but also protect the experiments, the beam dump, and the injection regions.[]{data-label="fig:collimators-around-LHC"}](collimation_system_lhc.pdf){width="\columnwidth"} Beam loss measurements and loss maps ------------------------------------ In order to prevent quenches and damage of sensitive equipment, beam loss monitors (BLMs) are installed around the LHC ring [@blmSystem; @holzer05; @holzer08a]. They trigger a beam dump if the local losses exceed a pre-defined threshold. The LHC BLM system uses about 4000 ionization chambers, placed outside the cryostat on cold magnets and on other key equipment such as collimators. The BLM system provides a measurement of electromagnetic and hadronic showers resulting from nearby impacts of beam losses. They provide data at a time resolution down to , i.e. around half the beam revolution time. In order to validate any operational configuration in the LHC, including the optical configuration and the collimator positions, validation tests called loss maps (LMs) are performed. During these tests a safe, low-intensity beam is artificially excited to create losses while the BLMs provide a continuous measurement of the loss distribution around the ring. An example of a loss map from standard operation with [$^{208}$Pb$^{82+}$ ]{}ions can be seen in the top graph of Fig.\[fig:example-loss-map\]. Loss maps are carried out at all stages of the LHC operational cycle. Firstly, the beams are injected in the machine at , at the so-called injection plateau. After the injection of bunches is finished, the energy of the two beams is increased for about 20 minutes during the energy ramp during which the first optics change occur. Once at top energy, the optics is adjusted in order to achieve a lower $\beta$-function at the collision points, which is called the squeeze. Then, the beams are put in collision, and this part of operation is called physics. {width="\textwidth"} PSI beam tests in the LHC {#sec:experiments} ========================= Experimental setup ------------------ The collimation system in the LHC has been designed and optimised for proton operation and it is important to evaluate its performance for any other species considered for operation. The first run with PSI beams in the LHC was performed during MD studies in July 2018 [@Schaumann_2019_ipac_part_strip_ions; @Schaumann:2670544], where [$^{208}$Pb$^{81+}$ ]{}ions with one electron left were injected and stored in the LHC. While this experiment had as a main objective the demonstration of the possibility to store in the LHC PSI beams with good lifetime, it also gave the opportunity for the first tests of the PSI collimation process. The beam provided by the injectors for this experiment consisted of two [$^{208}$Pb$^{81+}$ ]{}bunches spaced by per SPS injection in the LHC. Each bunch featured an intensity of up to about $1.1\times10^{10}$ charges, or $1.3\times10^{8}$ ions. Because of machine protection requirements, the total circulating beam intensity had to stay below $3\times10^{11}$ charges during the experiment. Table \[tab:paramaters\] lists all the beam and machine parameters used during the test. In Table \[tab:summary\] we list all detailed settings of the collimators used for the experiment and later for the simulation. Parameter LHC MD HL–LHC --------------------------------------------- ------------------------- ------------------------- Ion $^{208}_{82}$Pb$^{81+}$ $^{208}_{82}$Pb$^{81+}$ Equiv. proton beam energy \[TeV\] 6.5 7 PSI beam energy \[TeV\] 526.5 [^3] 567 Proton beam emmitance \[$\mu$m\] 3.5 2.5 Ion beam emmitance \[$\mu$m\] 1.39 1.00 Bunch population \[Pb ions $\times10^{8}$\] 0.9 [^4] 1.8 Nb of bunches 6 1240 $\beta^*$ in IP1/2/5/8 (Top energy) \[m\] 1/1/1/1.5 – : Collection of beam and machine parameters used in experiments with [$^{208}$Pb$^{81+}$ ]{}and for the simulated beam in the SixTrack-FLUKA setup. The HL–LHC beam parameters are taken from Ref. [@bruce20_HL_ion_report]. \[tab:paramaters\] The cleaning performance for [$^{208}$Pb$^{81+}$ ]{}was tested through dedicated loss maps at injection and at top energy. The latter case can be seen in Fig. \[fig:example-loss-map\] (bottom graph) together with the loss pattern obtained for [$^{208}$Pb$^{82+}$ ]{}beams (top graph). Losses are normalized by the signal recorded at the primary collimator with highest losses. Both at injection and at top energy, severe losses are observed in the DS of IR7 with PSI beams. These losses turned out to be a real operational limitation, when a beam dump was triggered, two minutes after reaching top energy in the first fill. At the time, only 24 low-intensity bunches were stored, and the beam was dumped because of too high losses around $s=$, i.e. in the DS of IR7. In the second and last fill of the experiment, the number of bunches was reduced to six and the intensity per bunch was reduced to about $0.75\times10^{10}$ charges. This beam could successfully be accelerated to and stored for about two hours [@Schaumann_2019_ipac_part_strip_ions; @Schaumann:2670544]. Still, the losses reached around 60% of the dump threshold level. [ l r r ]{} Collimator name & LHC & HL–LHC\ & \[$\sigma_N$\] & \[$\sigma_N$\]\ \ TCP & 5.0 & 5.0\ TCSG & 6.5 & 6.5\ TCLA & 10.0& 10.0\ TCLD &n/a &14.0\ \ TCP & 15.0 & 15.0\ TCSG & 18.0 & 18.0\ TCLA & 20.0& 20.0\ \ TCT IR1/2/5 & 15 & 10\ TCT IR8 & 15 & 15\ Measured cleaning performance of PSI beams ------------------------------------------ In the second fill, the collimation performance at was tested through loss maps. The worst losses were observed for Beam 1 in the horizontal plane (B1H)and the measured loss map for this case can be seen in the bottom graph of Fig. \[fig:example-loss-map\], showing the full-ring loss pattern, and in the bottom graph of Fig. \[fig:psi-lossmaps-b1h\] (IR7 zoom). The peak losses were recorded by the BLMs located around the quadrupole magnet in cell 11, at position $s\approx$ from the center of IR1 (see for the local dispersion plotted in Fig. \[fig:psi-lossmaps-b1h\]). Please note, that Beam 2 did not reach the top energy, therefore there is no data for it [@Schaumann:2670544]. The performance of the collimation system with PSI beams is significantly worse than for protons or fully stripped ion beams [@bruce14_PRSTAB_sixtr; @hermes16_nim]. The recorded magnitude of the highest losses on the cold aperture of the DS, normalized to the intensity impacting on the TCP, is about 4 orders of magnitude larger than for standard proton operation, and about two orders of magnitude larger than for standard [$^{208}$Pb$^{82+}$ ]{}operation, as can be seen in Fig. \[fig:psi-lossmaps-b1h\]. The BLM signal is also about 4 times larger at the peak in the DS than on the collimators, however, this does not mean that primary losses occur in the DS. Instead, this is due to the fact that the BLM response per locally lost particle is different at the two locations, because of the local geometry and materials. ![Loss map for Beam 1 in the horizontal plane, recorded during the standard [$^{208}$Pb$^{82+}$ ]{}fill (top) and during the experiments with [$^{208}$Pb$^{81+}$ ]{}beams that were carried out in LHC in 2018 at top energy (bottom). The PSI case shows a visible excess of the BLM signal around $s=$. The machine lattice and the horizontal locally generated dispersion from the TCP are shown on the very top of the figure.[]{data-label="fig:psi-lossmaps-b1h"}](two_lms_zoom_dispersion.pdf){width="\columnwidth"} Simulations of PSI beam losses in the LHC {#sec:simulations} ========================================= Process of DS losses through stripping action --------------------------------------------- We hypothesize that the loss pattern in Fig. \[fig:psi-lossmaps-b1h\] can be explained by the stripping action of the collimators in combination with the increasing value of the dispersion in the DS (see the top plot of the Fig. \[fig:psi-lossmaps-b1h\]). When passing through the TCP, partially stripped [$^{208}$Pb$^{81+}$ ]{}ions from the beam halo may lose their electron. In this stripping process, they do not experience an angular deflection sufficient to be intercepted by the TCSs. The resulting fully-stripped [$^{208}$Pb$^{82+}$ ]{}ions have energies very close to nominal, i.e. , but have an altered charge-to-mass ratio and thus a magnetic rigidity that differs from that of the circulating beam by about 1.2 %. If they escape the betatron-cleaning insertion, the dispersion in the DS can push their trajectories onto the cold aperture. The simulation tools used for this study do not support the tracking of the partially stripped ions. Therefore, a simplified simulation setup was used, assuming a 100% stripping efficiency, i.e. PSI immediately losing their electron when they impact on a primary collimator such that any ions escaping the collimators are fully stripped. This assumption comes from the analysis of the mean free path (MFP) for the stripping process. The MFP was calculated to be for [$^{208}$Pb$^{81+}$ ]{}ions in the Carbon-Fiber-Carbon material, of which the 0.6 m-long primary collimators are made of, using the methods in Ref. [@Tolstikhina_2018]. The MFP can be compared to the distance traveled in the material for the different impact parameters, i.e. the distance between the collimator edge and the impact, where we assume that any impact takes place at the phase with maximum amplitude in phase space. This determines the impact angle. In this study we have considered impact parameter values , and (as in Ref. [@bruce14_PRSTAB_sixtr]) and we obtained traverse distances of: , , and , respectively. For each impact parameter value in the given range, the distance traveled inside the collimator material is at least two orders of magnitude larger than the MFPs reported earlier. Therefore, it is a very good approximation to assume a full stripping of any PSI that approaches the TCPs. Trajectories of fully stripped ions ----------------------------------- To test the theory of the stripping action of the collimators as explanation for the observed losses, we performed tracking simulations in MAD-X [@herr04; @madx] of fully stripped ions emerging from the horizontal TCP. The trajectories of the fully stripped [$^{208}$Pb$^{82+}$ ]{}ions were calculated with an effective $\Delta p/p=-1/82$ originating at one of the TCP jaws. Those trajectories were tracked through the betatron-cleaning insertion and the downstream DS, where the point at which they are intercepted by the aperture was calculated. In this simplified study, we do not assume any angular kicks at the TCP, but instead that the particles are at the phase of maximum horizontal excursion in phase space, as would be the case when they first hit the TCP after a slow diffusion process. A selected range of trajectories with the physical aperture overlaid is shown in Fig. \[fig:psi-lhc-experiment-fluka\]. As seen in the figure, the MAD-X trajectories of [$^{208}$Pb$^{82+}$ ]{}escaping the TCP jaws bypass all downstream collimators and travel directly to the DS. We observe a calculated loss position very similar to the one measured in the machine (see Fig. \[fig:psi-lossmaps-b1h\]). This result strengthens the hypothesis on the origin of the large losses in the DS. Furthermore, it shows that since the loss mechanism involves dispersion, the loss location is relatively constant regardless of which TCP and jaw caused the stripping. ![The top plot shows trajectories of fully stripped off-rigidity [$^{208}$Pb$^{82+}$ ]{}ions escaping the B1 TCPs, as calculated with MAD-X. The 4 trajectories depicted correspond to starting positions at both jaws of the horizontal and vertical TCPs. The middle plot shows the loss pattern simulated for PSI beams using the Sixtrack–FLUKA simulations. The bottom plot shows the measured B1H loss map with [$^{208}$Pb$^{81+}$ ]{}beams at top energy. []{data-label="fig:psi-lhc-experiment-fluka"}](comparison_three.pdf){width="\linewidth"} Simulations of the LHC experiment --------------------------------- The next step to further investigate the stripping effect of the TCPs was to perform an integrated beam tracking and collimator interaction simulation. Using the coupling between the SixTrack [@schmidt94; @sixtrack-web; @bruce14_PRSTAB_sixtr] and FLUKA [@fluka14; @Battistoni:2015epi] codes, described further in Refs. [@mereghetti13_ipac; @skordis18_tracking_workshop; @pascal-thesis], the aim was first to reproduce the measurement results and second to validate the proposed mitigation solutions (see Section \[sec:mitigations\]). These simulations track beam halo particles through the magnetic lattice using SixTrack, accounting for a detailed aperture model to infer beam losses. When a particle hits a collimator, the particle-matter interaction is simulated using FLUKA, and any surviving particles that return to the beam vacuum are sent back to SixTrack for further magnetic tracking. Similar simulations have been extensively compared to measurements in previous publications for protons [@bruce14_PRSTAB_sixtr; @auchmann15_PRSTAB; @bruce17_NIM_beta40cm; @bruce19_PRAB_beam-halo_backgrounds_ATLAS] and Pb ions [@hermes16_nim; @hermes16_ipac_coupling]. The simulation was started at the upstream edge of the horizontal TCP, tracking [$^{208}$Pb$^{82+}$ ]{}ions with the electron already stripped, but in a machine configuration with the magnetic rigidity adjusted for [$^{208}$Pb$^{81+}$ ]{}, i.e. . To simulate the experiment we used $5\times10^5$ macro particles hitting a collimator with an impact parameter of (to remain conform to the regular [$^{208}$Pb$^{82+}$ ]{}simulations). The cleaning hierarchy setup was reproduced as in the experiment, and it is listed in detail in Table \[tab:summary\]. Figure \[fig:psi-lhc-experiment-fluka\] shows the simulated loss distribution for the LHC configuration used during the PSI machine test, including the same optics and collimator settings. One can see the very good qualitative agreement between the measured loss map and the simulated one, with the highest cold loss peak at the place of the aperture impact predicted with MAD-X [@madx], used to estimate the single-pass trajectory for off-momentum particles. While the agreement on the peak losses (see Fig. \[fig:psi-lhc-experiment-fluka\]) is visible in the DS region around $s=\unit[20400]{m}$, some differences may be noticed in the TCP/TCS region ($s=\unit[19800-19900]{m}$). This apparent discrepancy could be explained by the fact that a full quantitative comparison in Fig. \[fig:psi-lhc-experiment-fluka\] cannot be made, since the BLM measurement is sensitive to the secondary shower particles that emerge outside of the impacted elements, while our simulations show the number of primary nuclei impacting on the aperture or disintegrating on the collimators. The measured loss pattern is also affected by a cross-talk between nearby BLMs, where any BLM intercepts the shower not only from the element at which it is placed, but also from losses on nearby upstream elements. Previous studies for protons have shown that the agreement improves dramatically when a further simulation of the shower development and the energy deposition in the BLM is performed [@bruce14_PRSTAB_sixtr; @auchmann15_PRSTAB]. Although the experiment was only performed for Beam 1, a simulation for Beam 2 was also carried out. A similar loss peak as for Beam 1 was noticed in the simulation results. However, due to small asymmetries of the optics functions with respect to Beam 1, the Beam 2 loss peak is found more downstream. Figure \[fig:psi-fluka-lhc-beam2\] shows the result of the simulation for Beam 2. The main loss position of fully stripped Pb ions given by the trajectories and tracking simulations in Fig. \[fig:psi-lhc-experiment-fluka\] is a strong indirect demonstration that the stripping process is the main source of the observed worsening in cleaning performance. ![SixTrack–FLUKA simulation results for Beam 2. The presence of an excess losses in the first cell after the dispersion suppressor (around $s=$) is visible, as for the case of Beam 1 (see Fig.\[fig:psi-lhc-experiment-fluka\]).[]{data-label="fig:psi-fluka-lhc-beam2"}](lhc_beam2_simulation.pdf){width="\columnwidth"} Mitigation techniques {#sec:mitigations} ===================== Dispersion suppressor collimators --------------------------------- The High Luminosity LHC (HL–LHC) [@hl-lhc-tech-design] will start operating in 2027 and will push forward the luminosity frontier, increasing by about a factor of 10 the LHC integrated luminosity. Several hardware upgrades will be implemented for the collimation system of the HL–LHC. Among other upgrades not relevant for this work, new dispersion suppressor collimators – called TCLD – will be installed to protect the downstream DS region [@redaelli14_chamonix]. During the current long shutdown period (LS2, in 2019-2021) it is planned to install one TCLD collimator per beam in the DS regions around IR7. The primary purpose of those collimators is to intercept dispersive losses coming from the betatron-collimation insertion and studies show that their presence can reduce the losses in the DS for both proton and ion beams [@hl-lhc-tech-design; @bruce14ipac_DS_coll; @hermes15_ipac]. To make space for the TCLD, a standard 8.33$\mathrm{T}$, 15 m-long main dipole will be replaced by two shorter, 5 m-long 11$\mathrm{T}$ dipoles, based on the Nb$_3$Sn superconducting alloy, with the collimator assembly in the middle [@Zlobin:2019ven]. The planned location for the TCLD installation at the longitudinal position $s$=, see Fig.6. Since the purpose of the TCLD is to catch dispersive losses, we investigate in detail whether it could potentially be used to intercept the fully stripped [$^{208}$Pb$^{82+}$ ]{}ions during PSI operation. A plot of the [$^{208}$Pb$^{82+}$ ]{}trajectories after electron stripping at the TCPs is shown in the top graph of Fig. \[fig:hllhc-tcld-in-trajectories-and-losses\]. The longitudinal position of the TCLD is indicated by the black line, showing also its expected operational opening of 14$\sigma$. The aperture of the fixed layout elements is also shown. It is shown that the TCLD can indeed intercept the fully-stripped [$^{208}$Pb$^{82+}$ ]{}ions before they reach the cold aperture of DS magnets. Complete loss maps simulations were performed to confirm more quantitatively this finding. The collimator settings of Table \[tab:summary\] were considered. The simulated loss distribution is shown in the middle and bottom graphs of Fig. \[fig:hllhc-tcld-in-trajectories-and-losses\], for the cases with and without TCLD, respectively. The TCLD efficiently catches the fully stripped ions, reducing the downstream cold losses by about four orders of magnitude. It is also noticeable that the TCLD collimator becomes (in case of operation with [$^{208}$Pb$^{81+}$ ]{}) the collimator with the highest fraction of the energy absorbed. ![\[fig:hllhc-tcld-in-trajectories-and-losses\]The top figure shows trajectories calculated with MAD-X of fully stripped off-rigidity [$^{208}$Pb$^{82+}$ ]{}ions escaping the TCPs together with the aperture model and the TCLD with of 14$\sigma$ opening. The middle plot shows the SixTrack-FLUKA coupling simulation of the loss map for HL–LHC version 1.2, including the TCLD at 14 $\sigma$ in cell 9. The bottom plot shows the simulated loss map for the same optics but without the TCLD, resulting in the similar loss pattern as observed in the LHC.](hllhc_tcld_3.pdf){width="\linewidth"} Another option including the dispersion suppressor collimator, would be to install one additional unit in the empty cryostat. The added values oh that solution would be lack of need for additional magnets, and absorber’s closer position to the peak losses. Moreover orchestrating the setting in a way to have the losses split between the two collimators is potentially possible. But detailed look in this aspect goes beyond the scope of this paper. Power load on coil and TCLD collimator -------------------------------------- Even though the TCLD effectively intercepts the losses of fully stripped ions, the risk of quenching the downstream magnets must be assessed by taking into account the energy leaking out of the TCLD. To estimate this risk, we assume a quench limit of for the dipole [@bottura18], and a minimum beam lifetime, according to the design specification of the HL–LHC collimation system. We also assume the maximum intensity for the [$^{208}$Pb$^{81+}$ ]{}beams as considered for the [$^{208}$Pb$^{82+}$ ]{}for HL–LHC, namely $2.2\times10^{11}$ [@bruce20_HL_ion_report; @hl-lhc-tech-design]. These assumptions correspond to a loss rate of about $3.1\times10^8$ ions per second on the TCP, carrying a total power of about 28 kW [@hl-lhc-tech-design]. Furthermore, we assume that each [$^{208}$Pb$^{82+}$ ]{}ion impacting on the TCLD causes an energy deposition of in the coils of the downstream magnet [@anton-et-all-on-TCLD-losses]. This number is extracted from an energy deposition study with FLUKA of betatron losses during standard [$^{208}$Pb$^{82+}$ ]{}operation. However, it should be noted that this is a very approximate number, since the impact distribution on the TCLD will be different for PSI losses. Therefore, for a more detailed assessment, specific energy deposition simulations should be repeated for this case. The total maximum [$^{208}$Pb$^{81+}$ ]{}beam intensity that is acceptable without quenching can then be calculated to $3\times 10^{11}$ Pb ions, which is beyond the baseline Pb total intensity for the ion runs in HL–LHC. Therefore, it is not expected that the total PSI intensity will be severely limited by the shower on the downstream 11 T magnet. To refine the intensity estimate, energy deposition simulations should be performed using the realistic PSI impact distribution on the TCLD. From the Fig. \[fig:hllhc-tcld-in-trajectories-and-losses\] the power load on the TCLD during a 0.2 h beam lifetime drop is estimated to about 30 % of the total beam losses, i.e. about . A simple scaling of previous studies on other tungsten collimators [@TCT_fede] shows that the TCLD risks to have a peak temperature of 150 $^\circ$C or more, and that it could temporarily deform by a few hundred microns but without any permanent damage to the collimator. Therefore, we do not expect severe limitations on the total beam intensity below the assumed baseline. However, a more detailed assessment, including further simulations of the energy deposition and thermo-mechanical response of the TCLD, is needed to draw a firm conclusion. In addition to the 0.2 h beam lifetime scenario, the impact of steady-state losses with a 1 h beam lifetime could also be studied as done for the standard HL–LHC running scenario [@hl-lhc-tech-design]. Since the TCLDs are planned to be installed in LS2, the LHC run 3 starting in 2021 will provide a unique opportunity to perform dedicated tests of cleaning efficiency with [$^{208}$Pb$^{81+}$ ]{}ion beams with the final DS layouts. Alternative mitigation measures ------------------------------- In case of problems with the TCLD collimators, or if a further mitigation would be needed, other alleviation techniques could be envisaged. We discuss these concepts here, although without detailed simulations. Crystal collimation is a novel technique being investigated for the LHC and HL–LHC [@daniele-thesis; @scandale16; @Mirarchi2017_crystals]. A bent silicon crystal is used instead of the amorphous carbon TCP in the standard collimation setup. Beam halo particles incident on the crystal can enter a channeling regime, in which their trajectories are guided by the potential between crystalline planes. The angular deflection achieved by channeling in the crystal is much larger than the deflection achieved by scattering in an amorphous material and beam halo particles can be directed onto a single massive absorber with a large impact parameter, reducing the need for additional secondary collimators and additional absorbers. While the interaction of the PSI with the crystal is currently not well characterised, crystal collimation has shown promise for improving the cleaning efficiency with heavy ion beams [@D'Andrea:2678781] and it is proposed to study if it can also alleviate the losses for PSI beams. If the impacting ions could be channelled by the crystal even if their electron has been stripped, they could could be steered onto the absorber as in the standard crystal collimation scheme. This concept requires thorough experimental feasibility studies before being relied upon. A crystal test stand installation is available in the LHC for collimation studies [@Mirarchi2017_crystals], and it is planned to test this with PSI beams when available again in the LHC. As the stripping action of collimators produces a secondary beam of similar particles, another mitigation strategy may involve an orbit bump to shift the losses to a less sensitive location. Orbit bumps have been successfully employed for the case of secondary beams created in the collisions between [$^{208}$Pb$^{82+}$ ]{}beams, formed by the process of Bound-Free Pair Production (BFPP) [@klein01; @prl07; @prstabBFPP09; @schaumann16_md_BFPPquench; @jowett16_ipac_bfpp]. The bump is used to shift the impact location of the secondary beam on the mechanical aperture out of the main dipole and into an empty connection cryostat so that no impacts occur directly on superconducting magnets. A similar strategy can be considered for PSI secondary beams. A local closed orbit bump could be used to optimize the loss location of the fully stripped [$^{208}$Pb$^{82+}$ ]{}beam, presently around $s$=, to the nearby empty connection cryostat. However, a rather large bump amplitude of about 6.5 mm would be needed to move the losses away from the quadrupole, and it needs to be studied if such a bump is operationally feasible. It should also be noted that the total peak power of the local losses on the connection cryostat during PSI operation will be much higher than the power of the BFPP losses that were successfully handled in the LHC. The peak collimation losses are assumed to be transient and last only for a few seconds, but steady collimation losses during standard operation might also cause limitations. Therefore, to conclude on the feasibility of this mitigation strategy, the power deposition in the connection cryostat and downstream magnets needs to be studied in further detail for different scenarios of losses on the TCPs. Conclusions and outlook ======================= Partially stripped [$^{208}$Pb$^{81+}$ ]{}ions with one electron left were injected, accelerated, and stored in the LHC for the first time. The results of the first measurement of the collimation performance with those beams at the LHC show that the cleaning efficiency is prohibitively low for high-intensity operation, due to very high localized losses in the dispersion suppressor. We have presented studies showing that the likely reason for the poor collimation performance is the stripping action of the primary collimators, which causes nearly every ion that touches the collimator to lose its electron. If an ion does not fragment in the collimator, it re-enters the beam with a higher charge, causing it to be lost in the dispersion suppressor where the dispersion rises. For a detailed investigation of the observed losses, tracking simulations were performed. The simulated impact location of the fully stripped [$^{208}$Pb$^{82+}$ ]{}ions, as well as the finding that this loss location is by far dominating, are in excellent agreement with the measurements. This demonstrates that the stripping action of the material in the primary collimators is indeed very likely responsible for the observed loss peaks, which have never been observed before at this magnitude with other particle species. The simulations were extended for the machine layout including upgrades foreseen for HL–LHC. Several mitigation strategies are under consideration for reducing the losses on the dispersion suppressor and thus increasing the intensity reach of partially stripped ions operation. The use of a new TCLD collimator, which is scheduled to be installed before the next LHC run, was identified as the most promising option, as it could efficiently intercept the fully stripped [$^{208}$Pb$^{82+}$ ]{}ions before reaching the cold magnets. A preliminary estimate extrapolated from energy deposition studies for [$^{208}$Pb$^{81+}$ ]{}indicates that the magnet downstream of the TCLD is not likely to quench with the assumption of beam lifetime and total beam intensity as for the [$^{208}$Pb$^{82+}$ ]{}beam, even with a potential full HL–LHC PSI beam. On the other hand, there is a risk that the TCLD itself could suffer significant deformations, potentially limiting the maximum intensity to a factor of a few below the nominal HL–LHC design intensity for [$^{208}$Pb$^{82+}$ ]{}. However, additional comprehensive energy deposition studies should be carried out to better quantify the current beam intensity limit of safe operation for both the magnets and the TCLD. The non-standard beam particle type and the additional physical interactions the PSI can undergo are not supported by the available simulation tools. An active effort is directed to extending the existing simulation frameworks and developing new ones that would enable future studies of various PSI collimation aspects. Two other mitigation techniques were discussed: crystal collimation and dedicated orbit bumps. Both options, alone or in combination with other mitigation techniques, may be considered as useful in case the TCLD would turn out not being sufficient regarding the quench limits of the nearby magnets, or if the load on the TCLD itself would be too high. The authors would like to thank the LHC operations crew for their help during the measurements. Additionally, the authors thank F. Carra for the detailed and instructive comments on the tungsten collimator power-deposition loads. [^1]: arek.gorzawski@cern.ch [^2]: andrey.abramov@cern.ch [^3]: it is $\times Q$ where $Q=Z-1$ in this case is the charge number of the ion. [^4]: first fill during the MD featured $1.3\times 10^8$ ions.

ArXiv

--- abstract: | We report the first results of AS2UDS: an 870$\mu$m continuum survey with the Atacama Large Millimeter/Submillimeter Array (ALMA) of a total area of $\sim$50arcmin$^2$ comprising a complete sample of 716 submillimeter sources drawn from the SCUBA-2 Cosmology Legacy Survey (S2CLS) map of the UKIDSS/UDS field. The S2CLS parent sample covers a 0.96degree$^2$ field at $\sigma_{850}=0.90\pm0.05$mJybeam$^{-1}$. Our deep, high-resolution ALMA observations with $\sigma_{\rm 870}\sim$0.25mJy and a 0$\farcs$15–0$\farcs$30 FWHM synthesized beam, provide precise locations for 695 submillimeter galaxies (SMGs) responsible for the submillimeter emission corresponding to 606 sources in the low resolution, single-dish map. We measure the number counts of SMGs brighter than $S_{\rm 870}\geq$4mJy, free from the effects of blending and show that the normalisation of the counts falls by 28$\pm$2% in comparison to the SCUBA-2 parent sample, but that the shape remains unchanged. We determine that 44$^{+16}_{-14}$% of the brighter single-dish sources with $S_{\rm 850}\geq$9mJy consist of a blend of two or more ALMA-detectable SMGs brighter than $S_{\rm 870}\sim$1mJy (corresponding to a galaxy with a total-infrared luminosity of $L_{\rm IR}{\mathrel{\raise0.35ex\hbox{$\scriptstyle >$}\kern-0.6em \lower0.40ex\hbox{{$\scriptstyle \sim$}}}}$10$^{12}$L$_\odot$), in comparison to 28$\pm$2% for the single-dish sources at $S_{\rm 850}\geq$5mJy. Using the 46 single-dish submillimeter sources that contain two or more ALMA-detected SMGs with photometric redshifts, we show that there is a significant statistical excess of pairs of SMGs with similar redshifts ($<$1% probability of occurring by chance), suggesting that *at least* 30% of these blends arise from physically associated pairs of SMGs. author: - 'Stuart M. Stach' - Ian Smail - 'A.M. Swinbank' - 'J.M. Simpson' - 'J.E. Geach' - Fang Xia An - Omar Almaini - Vinodiran Arumugam - 'A.W. Blain' - 'S.C. Chapman' - 'Chian-Chou Chen' - 'C.J. Conselice' - 'E.A. Cooke' - 'K.E.K. Coppin' - 'J.S. Dunlop' - Duncan Farrah - 'B. Gullberg' - 'W. Hartley' - 'R.J. Ivison' - 'D.T. Maltby' - 'M.J. Michałowski' - Douglas Scott - Chris Simpson - 'A.P. Thomson' - 'J.L. Wardlow' - 'P. van der Werf' bibliography: - 'Countsref.bib' title: 'An ALMA survey of the SCUBA-2 Cosmology Legacy Survey UKIDSS/UDS field: Number counts of submillimeter galaxies' --- Introduction {#sec:intro} ============ It has been two decades since the Submillimeter Common User Bolometer Array (SCUBA) instrument on the James Clerk Maxwell Telescope (JCMT) enabled deep observations of high-redshift submillimeter sources which expanded the number of known high-redshift submillimeter luminous infrared sources up to hundreds [e.g. @smail1997deep; @hughes1998high; @barger1998submillimetre]. These submillimeter galaxies (SMGs) constitute a population of the most intensely star-forming galaxies, with star-formation rates (SFRs) in the 100s–1000s of M$_{\odot}$yr$^{-1}$ [@blain2002submillimeter; @magnelli2012herschel; @swinbank2013alma; @casey2013characterization] at typical redshifts $z\sim $2–3 [@chapman2005redshift; @wardlow2011laboca; @simpson2014alma; @chen2016scuba]. This level of star formation means that in a single starburst event, an SMG would need just a few hundred million years to form the stellar mass of a massive galaxy ($M_{\ast} \gtrsim $10$^{11}$M$_{\odot}$). This has led to the suggestion that SMGs have many of the properties expected for the progenitors of the luminous massive elliptical and spheroid galaxies in the local Universe [@lilly1999canada; @fu2013rapid; @simpson2014alma] with speculation that they could represent a phase in a single evolutionary path linking SMGs to luminous quasi-stellar objects (QSOs) at $z\sim$2 and massive, passive galaxies found at $z\sim$1–2 [@coppin2008testing; @cimatti2008gmass; @whitaker2012large; @toft2014submillimeter]. Further evidence for this evolutionary path comes from clustering studies from single-dish detections, suggesting they reside in halos of mass $\sim $10$^{13}$M$_{\odot}$, consistent with that of $z\sim$2 QSOs and with their subsequent evolution into local ellipticals [@farrah2006spatial; @hickox2012laboca; @wilkinson2016scuba]. However, whilst SMGs may play a significant role in the stellar mass growth of massive galaxies, measuring their basic properties have been hampered by the coarse angular resolution of the single-dish telescopes, with beams of $\sim$15$\arcsec$FWHM. One of the questions raised is whether the (coarse resolution) single-dish detections arises from a single SMG or are blends of multiple SMGs within the single-dish beam. To measure the blending and to accurately identify SMG counterparts at other wavelengths requires high-resolution interferometric studies, which were initially performed via radio counterpart identification [e.g. @chapman2005redshift; @ivison2007scuba], but more recently with submillimeter interferometers. [@wang2010sma] use deep 850$\mu$m integrations of two bright submillimeter sources in the GOODS-N field to suggest that both sources break into multiple components and suggested that around 30% of 850-$\mu$m sources with flux densities ($S_{\rm 850}$) $S_{\rm 850}\geq$5mJy could be composed of blends of more than one SMG. ALMA observations of much larger samples suggested that this rises to $>$90% for $S_{\rm 850}\sim$8mJy sources selected in single-dish surveys [e.g. @simpson2015scuba2]. More recently, [@hill2018high] used the Submillimeter Array (SMA) to observe 75 of the brightest S2CLS sources (S$_{\rm 850}\gtrsim$8mJy) at 870$\mu$m with a resolution of $\sim$2$\farcs$4. Combining their SMA data with archival observations they determine a lower multiplicity rate of $\sim$15%, which is consistent with previous work with the SMA [@chen2013resolving]. However these SMA observations are limited by the sensitivity, with [@hill2018high] using maps with an average rms depth of $\sim$1.5mJy. This meant that multiples can only be identified in a bright single-dish source if both components have near equal flux density, which is unlikely to be a frequent occurrence. Therefore, care needs to be taken when comparing such multiplicity studies since they can use different criteria for the brightness ratio of detected sources. To make definitive progress in understanding the properties of SMGs area requires the improvements in sensitivity and resolution provided by the Atacama Large Millimeter/Submillimeter Array (ALMA). The first such study, comprising Cycle 0 observations of the 122 submillimeter sources detected in the LABOCA survey of the Extended *Chandra* Deep Field South [LESS: @weiss2009large] found that 30% of LABOCA sources resolved into multiple components with $S_{\rm 850}\gtrsim$1.5mJy when observed at 1$\farcs$5 resolution [@karim2013alma; @hodge2013alma]. Following this result, in ALMA Cycle 1, 30 of the brightest submillimeter sources (median single-dish flux density of $S_{\rm 850}{\mathrel{\raise0.35ex\hbox{$\scriptstyle >$}\kern-0.6em \lower0.40ex\hbox{{$\scriptstyle \sim$}}}}$9mJy) from the SCUBA-2 Cosmology Legacy Survey [S2CLS: @geach2017scuba] map of the UKIDSS Ultra Deep Survey (UDS, @lawrence2007ukirt) field were observed with ALMA by @simpson2015scuba2. This confirmed that the majority (61$^{+19}_{-15}$%) of bright, single-dish submillimeter sources are comprised of blends of multiple SMGs brighter than $S_{\rm 850}\sim$1.5mJy [@simpson2015scuba2; @simpson2015scuba]. Each of these bright single-dish sources consists of 2–4 SMGs, which themselves are ultraluminous infrared galaxies (ULIRGs; $L_{\rm IR}\geq10^{12}L_{\odot}$), seen within a projected diameter of $\sim$150kpc. [@simpson2015scuba2] suggest that such a high over-density of SMGs requires that the majority of such detections result from physical association, as opposed to chance projections along the line of sight. Several studies have used spectroscopic observations of molecular gas emission to test the origin of blends of SMGs. For example, [@zavala2015early] used spectroscopic detections for the components in one blended submillimeter-bright lensed galaxy to show that it split into three distinct galaxies, each at significantly different redshifts. More recently, [@wardlow2018] used ALMA observations to search for CO emission in the fields of six submillimeter sources, which include a total of 14 SMGs, to determine that $\gtrsim$75% of blends of multiple SMGs are not physically associated. Similarly, [@hayward2018observational] report optical and near-infrared spectroscopy of a sample of seven single-dish sources, where three showed a blending of physically associated SMGs, whilst four contained at least one pair of components that was physically unassociated. This mix of physically associated and unassociated components in the blended single-dish submillimeter sources is consistent with semi-analytic modelling, for example [@cowley2014simulated] have suggested that most blends of SMGs in single-dish sources arise from projections of unrelated galaxies seen along the line of sight. The presence of multiple SMG counterparts to individual single-dish submillimeter sources indicates that the number counts derived from low-resolution single-dish surveys do not represent the true number counts of SMGs. Even a small change in the expected form of the counts of SMGs has a potentially significant impact on models that use them as a constraint on the evolution of high-redshift, dust obscured starbursts [e.g. @cowley2014simulated; @lacey2016unified]. In this paper we present the first results of the recently completed ALMA survey of the full S2CLS UDS sample, which comprises 870$\mu$m maps of the 716 $>$4$\sigma$ single-dish sources with observed $S_{\rm 850}\geq$3.4mJy in this 0.96degree$^2$ field. Our deep, high-resolution ALMA survey, with rms depths of $\sigma_{\rm 870}\sim$0.25mJybeam$^{-1}$ at 0$\farcs$15–0$\farcs$30 resolution, provides the statistical sample necessary to study the SMG population in detail and supplies us with the largest sample of ALMA-detected SMGs currently available. From this we construct resolved 870-$\mu$m SMG number counts and investigate the multiplicity in single-dish surveys. In §\[sec:data\] we describe the sample selection, observations, data reduction and source extraction. §\[sec:results\] covers our results and discussions and §\[sec:conclusion\] gives our conclusions. Observations and Data Reduction {#sec:data} =============================== Sample Selection ---------------- Our survey (the ALMA-SCUBA-2 Ultra Deep Survey field survey, hereafter AS2UDS) is based on a complete sample of 850-$\mu$m sources selected from the S2CLS map of the UDS field [@geach2017scuba]. The S2CLS UDS map covers an area of 0.96deg$^{2}$, with noise levels below 1.3mJy and a median depth of $\sigma_{\rm 850}=$0.88mJybeam$^{-1}$ with 80% of sources having $\sigma_{\rm 850}=$0.86–1.02mJybeam$^{-1}$. Between Cycles 1, 3 and 4 we observed all 716 $>4\sigma$ sources from the SCUBA-2 map, giving an observed flux density limit of $S_{\rm 850}\geq $3.4mJy, or a deboosted flux density of $S^{\rm deb}_{\rm 850}\geq$2.5mJy [@geach2017scuba]. As a pilot project in Cycle 1 (Project ID: 2012.1.00090.S), 30 of the brightest sources from an early version of the SCUBA-2 map (data taken before 2013 February) were observed in Band 7 [@simpson2015scuba; @simpson2015scuba2; @simpson2017scuba]. This early version of the map had a depth of $\sigma_{\rm 850}\sim$2.0mJy$^{-1}$ and subsequent integration time scattered three of these sources below our final sample selection criteria, leaving 27 of these original single-dish detected sources in our final sample. In Cycles 3 and 4 (Project ID: 2015.1.01528.S and 2016.1.00434.S, respectively) we observed the remaining 689 single-dish sources in the final S2CLS catalog. To cross calibrate the data, a fraction of these sources were observed twice in Cycles 3 and 4 or twice in Cycle 4. Data Reduction and Source Detection ----------------------------------- Full details of the data reduction and source detection will be presented in Stach et al. (in prep.) but here we provide a brief overview. Our ALMA targets were observed in Band 7 (344GHz$\sim$870$\mu$m), where the frequency closely matches the central frequency of the SCUBA-2 filter transmission and the FWHM of the ALMA primary beam at this frequency (17$\farcs$3) comfortably covers the whole of the SCUBA-2 beam (14$\farcs$7 FWHM). Cycle 1 observations were carried out on 2013 November 1, Cycle 3 between 2016 July 23 and August 11 and Cycle 4 between 2016 November 9 and 17 and 2017 May 6. The phase center for each pointing was set to the SCUBA-2 positions from the S2CLS DR1 submillimeter source catalog [@geach2017scuba], with observations taken with 7.5GHz bandwidth centred at 344GHz using a single continuum correlator set-up with four basebands. Observations of 40seconds were employed with the aim to yield 0$\farcs$3 resolution maps with a depth of $\sigma_{\rm 870}=$0.25mJybeam$^{-1}$. However, the Cycle 3 observations were taken in a more extended ALMA configuration, yielding a median synthesised beam of 0$\farcs$19 FWHM. Calibration and imaging were carried out with the <span style="font-variant:small-caps;">Common Astronomy Software Application</span> [<span style="font-variant:small-caps;">casa</span> v4.6.0; @mcmullin2007casa]. For source detection we created “detection” maps by applying a 0$\farcs$5 FWHM Gaussian taper in the *uv*-plane, to ensure sensitivity to extended flux from our SMGs that might fall below our detection threshold, as well as improving efficiency for selecting extended sources. This down-weighting of the long baseline information results in final “detection” maps with a mean synthesized beam size of 0$\farcs$73$\times$0$\farcs$59 for Cycle 1, 0$\farcs$56$\times$0$\farcs$50 for Cycle 3 and 0$\farcs$58$\times$0$\farcs$55 for Cycle 4. The <span style="font-variant:small-caps;">clean</span> algorithm was used to create the continuum maps using multi-frequency synthesis mode with a natural weighting to maximise sensitivity. We initially created a dirty image from the combined spectral windows (SPWs) for each field and calculated the rms noise values. The fields were then initially cleaned to 3$\sigma$ and then masking ellipses are placed on sources above 4$\sigma$ and the sources are then cleaned to 1.5’,$\sigma$. The final cleaned, *uv*-tapered detection maps have mean depths of $\sigma_{\rm 870}=$0.25mJybeam$^{-1}$ for Cycle 1, $\sigma_{\rm 870}=$0.34mJybeam$^{-1}$ for Cycle 3 and $\sigma_{\rm 870}=$0.23mJybeam$^{-1}$ in Cycle 4, the differences here largely being due to the varying resolutions of the observations in each ALMA cycle. For source detection, <span style="font-variant:small-caps;">sextractor</span> was initially used to find $>$2$\sigma$ peaks within the “detection” maps. Noise estimates were then calculated from the standard deviation in the integrated fluxes in 100 randomly placed 0$\farcs$5 diameter apertures in each map. These were then used, along with the 0$\farcs$5 diameter flux measured for each detection, to determine the signal-to-noise ratio (SNR) of the sources. As we used an aperture smaller than the beam size the mean 0$\farcs$5 aperture depths in the detection maps are approximately a factor of two deeper than the noise per beams quoted above (with the caveat of a corresponding aperture correction). The choice of the size of the detection aperture and the SNR cut for the sample selection were made based on a trade-off between purity and depth of the catalog. The final catalog consists of the 695[^1] sources that have a 0$\farcs$5 aperture SNR $\geq$4.3 and fall within the primary beam of the ALMA maps. This threshold and aperture size was chosen to give us a 98% purity rate, $P_{\rm r}$ (2% contamination), calculated as follows: $$P_{\rm r}=\frac{N_{p}-N_{n}}{N_{p}}, \label{eq:doublepower}$$ where $N_{p}$ is the number of positive sources detected above the chosen SNR limit (i.e. 695) and $N_{n}$ is the number of sources detected above the same limit in the inverted detection maps (made by multiplying the detection maps by $-1$, Figure \[fig:purity\]). We confirm the behaviour of the noise in our maps by comparing our number of “negative” sources from the inverted maps at our selected SNR threshold against that expected from a simple Gaussian distribution of independent synthesised beams [@dunlop2016deep]. In AS2UDS, for our average restored beam size, there are roughly $\sim$450,000 independent beams across the 716 ALMA pointings. For Gaussian statistics we would then expect $\sim$8 “negative” sources at 4.3$\sigma$. However, as noted by [@dunlop2016deep], based on [@condon1997errors; @condon1998nrao], there are effectively twice as many statistically independent noise samples as one would expect from a naive Gaussian approach due to the non-independence of pixel values in synthesised imaging. This would result in an expected $\sim$16 “negative” sources or 2.3$\pm$0.5%, which is consistent with the number we detect. For each of the detected sources we then derived a 1$\farcs$0 diameter aperture flux density from the primary beam corrected maps, these flux densities are aperture corrected and flux deboosted using the same methodology as [@simpson2015scuba2], as briefly described below. Completeness & Flux Deboosting ------------------------------ To calculate the completeness and flux deboosting factors for our ALMA catalog we inserted model sources into simulated ALMA maps and determined the properties of those which were recovered. We start with simulated noise maps, to make these as realistic as possible we used ten residual maps output from [casa]{} (i.e. an observed ALMA map where the source flux from any detected sources has been removed). The maps were selected to match the distribution in observed $\sigma_{\rm 870}$ for all 716 AS2UDS pointings. Model sources with flux densities drawn from a steeply declining power-law distribution with an index of $-2$, consistent with [@karim2013alma; @simpson2015scuba2], and intrinsic FWHM sizes drawn uniformly from a range 0–0$\farcs$9, were convolved with ALMA synthesized beams and inserted into 60,000 simulated noise maps. Then we applied our source detection algorithm and measured recovered fluxes as detailed above, with a successful recovery claimed for detections within the size of a synthesized beam, i.e. 0$\farcs$6, from the injected model source position. The result of these simulations is that we estimate our catalog is 98$\pm$1% complete for all our simulated sources at $S_{870}\geq$4mJy, with the incompleteness exclusively arising from the most extended simulated sources (intrinsic FWHM $>$0$\farcs$6). As found in [@franco2018goods] our simulated maps show the intrinsic sizes of the submillimeter galaxies strongly effects the completeness fractions at low signal-to-noise. But, at our 4mJy threshold we are only miss a small number of the most extended galaxies. We note that our simulated sources had sizes which were uniformly distributed up to 0$\farcs$9, whereas previous studies suggest median submillimeter sizes of $\sim0\farcs3$ [@tacconi2006high; @simpson2015scuba] therefore the 98$\pm$1% completeness is probably conservative. We estimate the flux boosting, the effect of noise fluctuations in the overestimation of a source’s flux density, by calculating the ratio of the flux density for each recovered simulated source to the original input flux density. The fact that noise in the maps is approximately Gaussian, combined with the steep counts of faint sources, means that we find that fluxes are typically overestimated in the lower flux bins. However, again brighter than $S_{870}\geq$4mJy the flux deboosting becomes a minor correction with a median correction factor of 0.98$\pm$0.04 for the SMGs considered in this paper. The complete catalog of SMGs from AS2UDS, with full descriptions of the source extraction, flux density measurements and flux deboosting will be presented in Stach et al. (in prep.). Analysis, Results and Discussion {#sec:results} ================================ The AS2UDS catalog contains 695 SMGs (detected in 606 ALMA maps), with $S_{\rm 870}\geq$0.9mJy (4.3$\sigma$), across 716 ALMA fields centred on $>4\sigma$ single-dish submillimeter sources from S2CLS [@geach2017scuba]. The total area of the primary-beam coverage in our ALMA survey is equivalent to 47.3arcmin$^2$. The AS2UDS SMG sample is roughly seven times larger than the previous largest sub/millimeter interferometric survey of single-dish submillimeter sources [ALESS: @hodge2013alma; @karim2013alma] and drawn from a field which is four times larger in terms of contiguous area. As was also found in ALESS, a fraction of our ALMA maps do not contain any detected SMGs (above 4.3$\sigma$ significance) – there are 108 of these “blank” maps (15$\pm$2% of the survey). In addition, we have 79 maps (11$\pm$1%) where the single-dish SCUBA-2 source breaks up into multiple SMGs at ALMA resolution. In §3.2 we show that the blank maps may in part be a result of similar “multiplicity” effects, as opposed to false positive detections in the original SCUBA-2 catalog. With this nearly order-of-magnitude increase in the sample of SMGs, in this paper we present number counts of SMGs brighter than $S_{\rm 870}\sim$4mJy, above the original 4-$\sigma$ limit of the single-dish SCUBA-2 survey. We also utilise the available multi-wavelength data for the UKIDSS/UDS field to employ photometric redshifts for our SMGs to quantify what fraction of the SCUBA-2 sources corresponding to multiple ALMA SMGs are due to chance projections, rather than physical associations. Flux Recovery ------------- We start by determining the fraction of the original SCUBA-2 sources fluxes which are recovered in the sources we detect in the corresponding maps from ALMA. In the flux regime that we are interested in this paper, $S_{\rm 870}\geq$4mJy, we find that we recover a median fraction of 97$^{+1}_{-2}$% of the original SCUBA-2 flux from SMGs detected within the ALMA primary beam pointing of the corresponding SCUBA-2 parent source. In respect of the “blank” maps: both the noise properties of the SCUBA-2 sources which resulted in“blank” maps and the noise properties of the ALMA observations of these maps are indistinguishable from those where ALMA detected an SMG. This suggests that these “blank” maps are not simply due to variations in the quality of the input catalog or follow-up observations. Similarly, it could be that many of the “blank” map sources are due to spurious false positives in the S2CLS parent sample. We test this by stacking *Herschel*/SPIRE maps at the locations of the 108 “blank” map sources, ranked in five bins of their SCUBA-2 flux. We recover emission in all the SPIRE bands (250, 350 and 500 $\mu$m) with flux densities between 7–20mJy for all five flux bins. Even for the faintest 10% of SCUBA-2 sources with corresponding “blank” ALMA maps, we still recover SPIRE detections at 250 and 350$\mu$m. Hence we are confident that the majority of the “blank” maps are a result of genuine non-detections in ALMA and not false positive sources in the S2CLS map. However, these “blank” maps do typically correspond to fainter single-dish sources: the median flux of the “blank” maps is $S_{\rm 850}=$4.0$\pm$0.1mJy, compared to $S_{\rm 850}=$4.5$\pm$0.1mJy for the whole sample. Thus it is possible that a strong increase in flux boosting in the original S2CLS catalog at SNR of $\lesssim$4–4.5$\sigma$ ($S_{\rm 870}\sim$3.6–4.0mJy) may play a part in explaining why ALMA detects no SMGs in these maps. To remove this concern, in our analysis we only consider the number counts brighter than $S_{\rm 870}\geq$4mJy. We conclude that with the sensitivities of our ALMA maps we can detect $S_{\rm 870}=$4mJy SMGs in even the shallowest AS2UDS maps across the entirety of the primary beam. In addition, based on our simulated ALMA maps described above, we have shown we have with reliable measured flux densities for the complete sample of 299 $S_{\rm 870}\geq$4mJy SMGs in the AS2UDS catalog presented here. Number Counts ------------- In Figure \[fig:Counts\], we show the cumulative and differential number counts of the 299 870$\mu$m-selected SMGs from AS2UDS to a flux limit of $S_{\rm 870}=4$mJy. Both the cumulative and differential number counts are normalized by the area of the S2CLS UDS map from which the original targets were selected: 0.96degree$^{2}$. Whilst the ALMA completeness factors are minimal for AS2UDS the number counts do have to be adjusted for the incompleteness of the parent S2CLS survey. We correct our counts by factoring in the estimated incompleteness of the catalog of the S2CLS UDS map from [@geach2017scuba] who reported that the parent sample is effectively complete at $\geq$5mJy, dropping to $\sim$88% at $\geq$4.5mJy and $\sim$83% at $\geq$4mJy. As in [@karim2013alma] the errors are calculated from both the Poissonian error and the individual flux uncertainties added in quadrature, where the flux uncertainty error is the standard deviation of the mean of the counts for each bin based on 1,000 re-samples of the catalog, assigning random flux densities to each source within their individual error margins, Table 1. We also compare these counts to those from the parent single-dish catalog of the S2CLS UDS field [@geach2017scuba], and the earlier ALESS survey [@karim2013alma]. To convert the S2CLS 850-$\mu$m counts to a common $S_{\rm 870}$ we use a factor of $S_{\rm 870}/S_{\rm 850}=$0.95 derived from a redshifted ($z=$2.5), composite spectral energy distribution (SED) for SMGs from the ALESS survey [@swinbank2013alma], although we note that this correction is smaller than the estimated absolute calibration precision from S2CLS of 15% [@geach2017scuba]. Compared to a single power-law fit, the number counts of SMGs show a steepening decline at brighter fluxes. As a result the best fit to the differential number counts is with a double power-law function with the form: $$\frac{dN}{dS} = \frac{N_{0}}{S_{0}}\Big[\Big(\frac{S}{S_{0}}\Big)^{\alpha}+\Big(\frac{S}{S_{0}}\Big)^{\beta}\Big]^{-1}, \label{eq:doublepower}$$ where $N_{0}$ describes the normalisation, $S_{0}$ the break flux density, $\alpha$ and $\beta$ the two power-law slopes. For our AS2UDS data the best-fit parameters found are $N_{0}=1200^{+200}_{-300}$deg$^{-2}$, $S_{0}=5.1\pm0.7$mJy, $\alpha=5.9^{+1.3}_{-0.9}$ and $\beta=0.4\pm0.1$. [ccc]{}\[h\] 4.5 & 385.3$^{+21.1}_{-7.7}$& 168.5$^{+14.8}_{-7.9}$\ 5.5 & 216.7$^{+17.3}_{-6.6}$ & 110.5$^{+12.1}_{-4.1}$\ 6.5 & 106.2$^{+11.4}_{-3.5}$ & 52.6$^{+8.3}_{-2.6}$\ 7.5 & 53.6$^{+8.4}_{-2.5}$ & 24.1$^{+6.0}_{-1.9}$\ 8.5 & 29.6$^{+6.5}_{-1.9}$ & 9.5$^{+4.2}_{-1.1}$\ 9.5 & 20.0$^{+5.7}_{-1.8}$ & 9.4$^{+4.2}_{-1.1}$\ 10.5 & 10.5$^{+4.4}_{-1.2}$ & 5.2$^{+3.5}_{-0.9}$\ 11.5 & 5.3$^{+3.5}_{-0.9}$ & 3.1$^{+3.0}_{-0.7}$\ 12.5 & 2.1$^{+2.8}_{-0.6}$ & –\ 13.5 & 2.1$^{+2.8}_{-0.6}$ & 1.0$^{+2.4}_{-0.5}$\ 14.5 & 1.0$^{+2.4}_{-0.5}$ & –\ At $S_{\rm 870}\geq$4mJy we derive a surface density of 390$^{+70}_{-80}$ deg$^{-2}$, corresponding to one SMG per $\sim$arcmin$^{2}$ or one source per $\sim$130 ALMA primary beams at this frequency. Figure \[fig:Counts\] shows a systematic reduction in the surface density of SMGs compared to the single-dish estimate at all fluxes. This reduction from the SCUBA-2 counts to AS2UDS is statistically significant for sources fainter than $S_{\rm 870}=$8mJy, with a reduction of a factor of 28$\pm$2% at $S_{\rm 870}\geq$4mJy and 41$\pm$8% at $S_{\rm 870}\geq$7mJy. At the very bright end ($S_{\rm 870}\geq$12mJy) the number of SMGs is so low (just two in our $\sim$1deg$^2$ field) that the reduction in the relative number counts is poorly constrained, 30$\pm$20%. Our bright-end reduction does agree with that seen in [@hill2018high] where they found a 24$\pm$6% reduction between 11–15mJy in their SMA follow-up counts compared to the original SCUBA-2 parent sample. This agreement is unsurprising as a large number of their sources are drawn from our ALMA survey of the UDS field. We also note that, as with our earlier pilot study of UDS in @simpson2015scuba2, that we do not see an extreme drop-off of the counts above $S_{\rm 870}\sim$9mJy as was suggested from the smaller-area ALESS survey [@karim2013alma]. As we discuss below, the main factor which appears to be driving the the systematically lower counts of SMGs from interferometric studies, compared to the single-dish surveys, is that a fraction of the brighter single-dish sources break up into multiple fainter sources (with flux densities of $S_{\rm 870}\lesssim$1–4mJy) in the interferometer maps and thus fall below the single-dish limit adopted for our counts. This effect has been termed “multiplicity” [@karim2013alma; @simpson2015scuba2]. An additional factor is the twelve ALMA “blank” maps of S2CLS sources brighter than $S^{\rm deb}_{\rm 870}\geq$4mJy, which also contribute to lowering the normalization of the number counts. These S2CLS sources, have a mean SNR of 5.8$\pm$0.8, and are therefore unlikely to be spurious SCUBA-2 detections and our *Herschel*/SPIRE stacking confirms this; instead the most likely explanation for their ALMA non-detection is “extreme” multiplicity, where the single-dish source breaks up into several faint SMGs below the detection limit of our ALMA maps. For these brighter SCUBA-2 sources with “blank” ALMA maps this would require that the single-dish source breaks up into $\geq$4 sources to result in a non-detection. Multiplicity ------------ There are differing claims in the literature regarding the influence of multiplicity of SMGs on single-dish submillimeter surveys. This is a result of both the differing depths of the interferometric studies used to investigate this issue and the different definitions of “multiplicity” adopted in these works. Our survey has a relatively uniform sensitivity of $\sigma_{\rm 870}\sim$0.25mJybeam$^{-1}$, and therefore we adopt a fixed $S_{\rm 870}$ limit to identify multiple SMGs. We follow [@simpson2015scuba2] and define a multiple map as any field with more than one $S_{\rm 870}\geq$1mJy SMG within our ALMA Band 7 primary beam (i.e. within $\sim$9$\arcsec$ of the original SCUBA-2 detection locations). At the redshift of SMGs this corresponds to borderline U/LIRG systems, $L_{\rm IR}\geq$10$^{12}$$L_{\odot}$ which have SFRs of the order of 10$^{2}$M$_{\odot}$yr$^{-1}$ [@swinbank2013alma]. We also believe this is a more physical choice than, for example, using the [*relative*]{} submillimeter brightness of the two sources to decide if they constitute a “multiple”, as the relative fluxes may have little relevance to their other physical properties (e.g., mass or redshift) which are essential to understand their significance. In our full sample we have maps with more than one $S_{\rm 870}\geq1$mJy SMG in 79 of the 716 observations (11$\pm$1%). We note that at 1mJy our ALMA observations are not complete, therefore this sets the multiplicity as a lower limit. The surface density of $S_{\rm 870}\sim$1mJy SMGs is $\sim$1arcmin$^{-2}$, as estimated from unbiased ALMA surveys [@aravena2016alma; @dunlop2016deep]. Hence we expect to find one $S_{\rm 870}\sim$1mJy SMG per $\sim$19 ALMA primary beams or in $\sim$5% of the maps, compared to the observed rate of $\sim$11% (one per nine ALMA maps). We note, however, that the presence of a secondary source in these maps may act to increase the likelihood of the inclusion of that map into our sample by boosting the apparent SCUBA-2 flux into the S2CLS catalog. To address this potential bias we estimate the multiplicity rate for the 179 brighter single-dish sources with deboosted SCUBA-2 flux densities of $S^{\rm deb}_{\rm 850}\geq$5mJy. The rate of multiples in these brighter SCUBA-2 sources is much higher 51/179 (28$\pm$2%), suggesting that the presence of a [*detected*]{} secondary SMG in faint single-dish sources does not strongly influence the inclusion of that single-dish source into our parent catalog. Instead, the influence of multiplicity in faint single-dish sources is more likely to be seen through the presence of “blank” maps. Hence we also place an upper limit on the multiplicity in our full survey by assuming that [*all*]{} the blank fields are a result of the blending of multiple faint SMGs, giving 187/716 (26$\pm$2%) multiples. As implied above, the multiplicity appears to depend on the single-dish flux: as expected as the inclusion of emission from other SMGs within the beam can only act to increase the apparent flux of the (blended) single-dish source. As described in §1, early observations suggested that roughly a third of $S_{\rm 850}>$5mJy single-dish sources could be blends of multiple SMGs, with this rate increasing to 90% for $S_{\rm 870}>$9mJy [e.g. @karim2013alma]. As shown in Figure\[fig:Multiplicity\], for AS2UDS we find a frequency of multiplicity (ignoring “blank” maps) of 28$\pm$2% for $S^{\rm deb}_{\rm 850}\geq$5mJy rising to 44$^{+16}_{-14}$% at $S^{\rm deb}_{\rm 850}\geq$9mJy. In Figure\[fig:Multiplicity\] we also plot the fractional contribution of each secondary and tertiary ALMA SMG (ranked by flux density) to the total recovered ALMA flux density of all the SMGs for each field with multiple SMGs. The mean fraction of the total flux contributed by the secondary component is 34$\pm$2% with no significant variation of this fraction as a function of the original deboosted SCUBA-2 source flux. The 64$\pm$2% contribution from the primary components in maps with multiple SMGs is broadly consistent with the semi-analytic model of [@cowley2014simulated] which suggested that $\sim$70% of the flux density in blended sources would arise from the brightest component. ### Physical association of the multiple SMGs Based on our Cycle 1 pilot study, [@simpson2015scuba2] showed that the number density of secondary SMGs in the maps of their 30 bright SCUBA-2 sources was 80$\pm$30 times that expected from blank-field number counts, suggesting that at least a fraction of these SMGs must be physically associated. Using our large sample we now seek to test this further. The most reliable route to test for physical association between SMGs in the same ALMA map would be to use spectroscopic redshifts for the SMGs. However, as the current spectroscopic coverage of SMGs in AS2UDS is sparse, we instead exploit photometric redshifts to undertake this test. We use the photometric redshift catalog constructed from the UKIDSS DR11 release (Hartley et al. in prep.), where a full description of the DR11 observations will be given in Almaini et al. (in prep.). These photometric redshifts are derived from twelve photometric bands ($U,B,V,R,I,z,Y,J,H,K,[3.6],[4.5]$) and applied to 296,007 $K$-band-detected sources using [eazy]{} [@brammer2008eazy]; details of the methodology can be found in [@simpson2013prevalence]. The accuracy of these photometric redshifts is investigated in Hartley et al. (in prep.) from comparison with the $\sim$6,500 sources in the UKIDSS DR11 catalog which have spectroscopic redshifts, finding $|z_{\rm spec}-z_{\rm phot}|/(1+z_{\rm spec})=0.019\pm0.001$ with a median precision of $\sim$9%. Around 85% of the ALMA maps fall in regions of the UDS with high-quality photometric redshifts and these are considered in the following analysis. In Figure\[fig:Deltaz\] we plot the distribution of the differences in photometric redshifts ($\Delta z_{\rm phot}$) for pairs of SMGs in those single-dish maps with multiple ALMA-detected SMGs. We limit our analysis to SMGs that fall within the region with high-quality photometric redshifts and which have $K$-band detections within 0$\farcs$6 radius from the ALMA positions (497 of the 695 SMGs) for both sources in the map. This yields 46 pairs of SMGs (92 SMGs in total) from the 164 SMGs in the 79 maps with multiple SMGs. We find that 52% of these pairs (24/46) have $\Delta z_{\rm phot} <$0.25. We note that 2$\arcsec$ diameter apertures were employed for the photometry in the DR11 catalog, therefore the $\Delta z_{\rm phot}$ was additionally calculated for only pairs that are separated by greater than 2$\arcsec$, thus removing the possibility of neighbours contaminating photometry and thus photometric redshifts. This *still* results in 53% of pairs having $\Delta z_{\rm phot} <$0.25 (23/43). To assess the significance of this result we next quantify whether the 24 pairs of blended SMGs with $\Delta z_{\rm phot}<0.25$ is statistically in excess of expectations for 46 random SMG pairs. To do this we determine the expected distribution of $\Delta z_{\rm phot}$ for pairs of SMGs randomly selected from the 497 SMGs with high-quality photometric redshifts across the full field, and plot this in Figure\[fig:Deltaz\]. To perform this test we sample the random distribution of our unassociated SMGs 10,000 times, each time drawing 46 pairs, and testing how frequently $>$52% of these are found to have $\Delta z_{\rm phot} <$0.25. This analysis shows that the median fraction of random pairs with $\Delta z_{\rm phot} <$0.25 is 20$\pm$2% compared to the 52% for the actual pairs of SMGs. This strongly suggests that a significant fraction of the single-dish sources that resolve into multiple optically-bright (e.g. those with photometric redshifts) SMGs are in fact physically associated galaxies on projected angular scales of $\sim$10–100kpc scales. If we assume that all pairs without photometric redshifts for *both* SMGs are physically unassociated, a conservative estimate, then comparing to the total number of ALMA fields with multiple SMGs, we can place a lower limit of at least 30% (24 pairs out of 79) on the fraction of all multiple-SMG fields arising from closely associated galaxies. This is consistent with previous spectroscopic studies of SMG multiples e.g. $\sim$40% of SMG pairs physically associated combining the estimates from [@wardlow2018] and [@hayward2018observational]. Of course, to truly test this requires a spectroscopic redshift survey of a much large sample of these multiple-SMG systems. Conclusions {#sec:conclusion} =========== We have presented the first results from a large ALMA 870-$\mu$m continuum survey of 716 single-dish submillimeter sources drawn from the SCUBA-2 Cosmology Legacy Survey map of the UKIDSS UDS field. These sensitive, high-resolution ALMA observations provide the largest sample of interferometrically detected submillimeter galaxies constructed to date, with 695 SMGs above 4.3$\sigma$ (corresponding to a false detection rate of 2%). This sample is seven times larger in terms number of SMGs and drawn from a single-dish survey which has four times the area of the previous largest interferometric SMG survey. The main conclusions of this work are as follows: - We construct resolved 870$\mu$m differential and cumulative number counts brighter than S$_{\rm 870}\geq$4mJy (a conservative choice based on the flux limit of the parent single-dish S2CLS survey), which show a similar shape to the number counts from S2CLS, but with a systematically lower normalization at fixed flux density, by a factor of 1.28$\pm$0.02. Much of this reduction in the SMG counts, is due to the influence of multiplicity, i.e. single-dish sources splitting into two or more SMGs detected by ALMA. We fit a double power-law function to our differential number counts to easily facilitate future comparison with observations in other fields and simulations. - In 11$\pm$1% of our 716 ALMA maps we detect more than one SMG with S$_{\rm 870}\geq$1mJy corresponding to a $L_{\rm IR} \geq $10$^{12}$ L$_{\odot}$ galaxy in a region with a projected diameter of $\sim$100kpc at $z=$2. This multiplicity fraction varies from 26$\pm$2% for all single-dish sources with $S^{\rm deb}_{\rm 850}\geq$5mJy, to 44$^{+16}_{-14}$% at $S^{\rm deb}_{\rm 850}\geq$9mJy. The brightest of these multiple-SMG components typically contributes 64$\pm$2% of the total flux of the SCUBA-2 source, with no detectable variation in this fraction with with single-dish source flux, consistent with results from semi-analytic models of blending in single-dish surveys. - By comparing the photometric redshift differences between pairs of SMGs in ALMA maps with multiple components, we show evidence that a significant fraction of these pairs are likely to be physically associated, with $\gtrsim$30% of all multiple-SMG maps arising from physically associated galaxies. SMS acknowledges the support of STFC studentship (ST/N50404X/1). AMS and IS acknowledge financial support from an STFC grant (ST/P000541/1). IS EAC and BG also acknowledge support from the ERC Advanced Investigator program DUSTYGAL 321334, and a Royal Society/Wolfson Merit Award. JEG acknowledges support from a Royal Society University Research Fellowship. JLW acknowledges the support of an STFC Ernest Rutherford Fellowship. MJM acknowledges the support of the National Science Centre, Poland through the POLONEZ grant 2015/19/P/ST9/04010; this project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sk[ł]{}odowska-Curie grant agreement No. 665778. The ALMA data used in this paper were obtained under programs ADS/JAO.ALMA\#2012.1.00090.S, \#2015.1.01528.S and \#2016.1.00434.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO, and NAOJ. This paper used data from project MJLSC02 on the James Clerk Maxwell Telescope, which is operated by the East Asian Observatory on behalf of The National Astronomical Observatory of Japan, Academia Sinica Institute of Astronomy and Astrophysics, the Korea Astronomy and Space Science Institute, the National Astronomical Observatories of China and the Chinese Academy of Sciences (Grant No. XDB09000000), with additional funding support from the Science and Technology Facilities Council of the United Kingdom and participating universities in the United Kingdom and Canada. UKIDSS-DR11 photometry made use of UKIRT. UKIRT is owned by the University of Hawaii (UH) and operated by the UH Institute for Astronomy; operations are enabled through the cooperation of the East Asian Observatory. When (some of) the data reported here were acquired, UKIRT was supported by NASA and operated under an agreement among the University of Hawaii, the University of Arizona, and Lockheed Martin Advanced Technology Center; operations were enabled through the cooperation of the East Asian Observatory. When (some of) the data reported here were acquired, UKIRT was operated by the Joint Astronomy Centre on behalf of the Science and Technology Facilities Council of the U.K. [^1]: We detect the strongly lensed SMG ’Orochi’ [@ikarashi2011detection] but remove this from our analysis.

ArXiv

--- abstract: 'We study the scattering of massless scalar waves by a Kerr black hole by letting plane monochromatic waves impinge on the black hole. We calculate the relevant scattering phase-shifts using the Prüfer phase-function method, which is computationally efficient and reliable also for high frequencies and/or large values of the angular multipole indices ($l$,$m$). We use the obtained phase-shifts and the partial-wave approach to determine differential cross sections and deflection functions. Results for off-axis scattering (waves incident along directions misaligned with the black hole’s rotation axis) are obtained for the first time. Inspection of the off-axis deflection functions reveals the same scattering phenomena as in Schwarzschild scattering. In particular, the cross sections are dominated by the glory effect and the forward (Coulomb) divergence due to the long-range nature of the gravitational field. In the rotating case the overall diffraction pattern is “frame-dragged” and as a result the glory maximum is not observed in the exact backward direction. We discuss the physical reason for this behaviour, and explain it in terms of the distinction between prograde and retrograde motion in the Kerr gravitational field. Finally, we also discuss the possible influence of the so-called superradiance effect on the scattered waves.' address: - '$^1$ Department of Physics and Astronomy, Cardiff University, Cardiff CF2 3YB, United Kingdom' - '$^2$ Department of Mathematics, University of Southampton, Southampton SO17 1BJ, United Kingdom' author: - 'Kostas Glampedakis$^{1}$ and Nils Andersson$^{2}$' title: Scattering of scalar waves by rotating black holes --- INTRODUCTION ============ Diffraction of scattered waves provides the explanation for many of Nature’s most beautiful phenomena, such as rainbows and glories. It has long been recognized that these optical phenomena have analogies in many other branches of physics. They are of particular relevance to quantum physics, where plane wave “beams” are routinely used to probe the details of atoms, nuclei or molecules. Such experiments provide a deep understanding of the scatterer’s physics and can be used as a powerful test of various theoretical models. The analogy can be extended also to gravitational physics and extreme astrophysical objects like black holes. In fact, black hole scattering has been the subject of a considerable amount of work carried out over the last 30 years (see [@book] for an extensive review). In the case of astrophysical black holes it is unlikely that the various diffraction effects will ever be observed (although it is not entirely implausible that advances of current technology will eventually enable us to study interference effects in gravitationally lensed waves). However, it is nevertheless useful to have a detailed theoretical understanding of the scattering of waves from black holes. After all, a study of these problems provides a deeper insight into the physics of black holes as well as wave-propagation in curved spacetimes. The benchmark problem for black-hole scattering is massless scalar waves impinging on a Schwarzshild black hole. This problem is well understood [@Ryan; @Sanchez1; @Sanchez2; @Sanchez3; @Sanchez4; @path_int1; @Nils], and it is known that it provides a beautiful example of the glory effect. Handler and Matzner [@Handler] have shown that the situation remains almost unchanged if, instead of scalar waves, one decides to “shoot” plane electromagnetic or gravitational waves towards the black hole. These authors have also considered on-axis scattering of gravitational waves in the case when the black hole is rotating [@Handler]. Their results suggest that the scattering cross sections consist of essentially the same features as in the non-rotating case (there is a forward divergence due to the long-range nature of the gravitational field and a backward glory). In addition, they find some peculiar features that are, at the present time, not well understood. An explanation of these effects is complicated by the fact that they could be caused by several effects, the most important being the coupling between the black hole’s spin and the spin/polarisation of the incident wave. Given that the available investigations have not been able to distinguish between these various effects, we feel that our current understanding is somewhat unsatisfactory. This feeling is enhanced by the fact that no results for the most realistic case, corresponding to off-axis incidence, have yet been obtained. This paper provides an attempt to further our understanding of the scattering from rotating black holes. Our aim is to isolate those scattering effects that are due to the spin of the black hole. In order to do this, we focus our attention on the scattering of massless scalar waves. For this case, the infalling waves have neither spin nor polarisation and therefore one would expect the scattered wave to have a simpler character than in the physically more relevant case of gravitational waves. However, one can be quite certain that the features discussed in this paper will be present also in the case of gravitational waves. It is, after all, well known that the propagation of various fields in a given black hole geometry is described by very similar wave equations. Although we will re-examine the case of axially incident waves, our main attention will be on the more interesting off-axis scattering cross sections. These cross sections turn out to be quite different from the ones available in the literature. Obviously, they have two degrees of freedom (corresponding to the two angles $\theta$ and $\varphi$ in Boyer-Lindquist coordinates). In addition we will show that the cross sections are asymmetric with respect to the incidence direction. In particular, the glory moves away from the backward direction as a result of rotational frame dragging that provides a distinction between prograde and retrograde motion in the Kerr geometry. We construct our differential cross sections using the well-known partial wave decomposition (for an introduction see [@Newton]) — the standard approach in quantum scattering theory. That this method is equally useful in black-hole scattering is well established [@book]. We should point out, however, that alternatives (such as the complex-angular momentum approach [@CAM1; @CAM2] and path-integral methods [@path_int1; @path_int2; @path_int3]) have also been succesfully applied to the black-hole case. In the partial wave picture all scattering information is contained in the radial wavefunction’s phase shifts. The calculation of these phase-shifts must, apart from in exceptional cases like Coulomb scattering, be performed numerically. Various techniques have been developed for this task. Basically, one must be able to determine the phase-shifts accurately up to sufficiently large $l$ partial waves that no interference effects are lost. This boils down to a need for many more multipoles to be studied as the frequency of the infalling wave is increased. In black-hole scattering several methods have been employed for the phase-shift calculation: Matzner and Ryan [@Ryan] numerically integrated the relevant radial wave equation (Teukolsky’s equation). Since the desired solution is an oscillating function, this calculation becomes increasingly difficult (and time consuming) as the frequency is increased. Consequently, Matzner and Ryan restricted their study of electromagnetic and gravitational wave scattering to $\omega M \leq 0.75 $ and $ l \leq 10 $. In order to avoid this difficulty, Handler and Matzner [@Handler] combined a numerical solution in the region where the gravitational curvature potential varies rapidly, with an approximate WKB solution for relatively large values of the radial coordinate. This trick allowed them to perform calculations for $l \leq 20 $ and $\omega M \leq 2.5 $. Some years ago one of us used the phase-integral method [@Froman1; @Froman2] to derive an approximate formula for the phase-shifts in the context of Schwarzschild scattering [@Nils]. This formula was shown to be reliable and efficient even for high frequencies and/or large $l$ values (in [@Nils] results for $\omega M= 10 $ and $l \leq 200 $ were presented). This means that the differential cross sections determined from the phase-integral phase-shifts were reliable also for rather high frequencies. Even though the phase-integral formula could be generalised to scattering by a Kerr black hole and therefore used for the purposes of the present study, we have chosen a different approach here. Our phase-shift determination is based on the so-called Prüfer method (well-known in quantum scattering theory [@Pajunen1; @Pajunen2] and, in general, in numerical treatments of Sturm-Liouville problems [@Pryce] ) which, in a nutshell, involves transforming the original radial wavefunction to specific phase-functions and numerical integration of the resulting equations. In essence, this method is a close relative of the phase-amplitude method that was devised by one of us to study black-hole resonances [@pam]. The remainder of the paper is organised as follows. In Sections IIA and IIB the problem of scattering by a Kerr black hole is rigorously formulated. In Section IIC the important notion of the deflection function is discussed. Section III is devoted to our numerical results. First, in Section IIIA our numerical method for calculating phase-shifts is presented. In Section IIIB familiar Schwarzschild results are reproduced as a code validation. Sections IIIC and IIID contain entirely new information: Differential cross sections and deflection functions for on and off-axis scattering respectively. These are the main results of the paper. Furthermore, in Section IIIE we present numerical results concerning forward glories. The role of superradiance for scattering of monochromatic waves is discussed in Section IIIF. Our conclusions are briefly summarised in Section IV. Three appendices are devoted to technical details, which are included for completeness. In Appendix A we discuss the notion of “plane waves” in the presence of a gravitational field. In Appendix B the partial-wave decomposition of a plane wave in the Kerr background is determined, and finally in Appendix C we briefly describe the method we have used to calculate the spin-$0$ spheroidal harmonics and their eigenvalues. Throughout the paper we adopt geometrised units ($c=G=1$). Scattering from black holes =========================== Formulation of the problem -------------------------- We consider a massless scalar field in the Kerr black-hole geometry. Then, first-order black-hole perturbation theory, basically the Teukolsky equation [@Teukolsky], applies. The scalar field satisfies the curved spacetime wave equation $ \Box \Phi =0$. Adopting standard Boyer-Lindquist coordinates we can always decompose the field as (since the spacetime is axially symmetric) $$\Phi(r,\theta,\varphi,t)= \frac{1}{\sqrt{r^2 +a^2}}\sum_{m=-\infty}^{+\infty} \phi_{m}(r, \theta,t) e^{im\varphi} \label{eq1}$$ In scattering problems it is customary to consider monocromatic waves with given frequency $\omega$. Therefore we can further write $$\phi_{m}(r,\theta,t)= \sum_{l=|m|}^{+\infty} c_{lm} u_{lm}(r,\omega) S_{lm}^{a\omega} (\theta) e^{-i\omega t} \label{eq2}$$ where $c_{lm}$ is some expansion coefficient and $S_{lm}^{a\omega}(\theta)$ are the usual spin-0 spheroidal harmonics. These are normalised as $$\int_{0}^{\pi} d\theta \sin\theta |S_{lm}^{a\omega}(\theta)|^2 = \frac{1}{2\pi}$$ Finally, the function $u_{lm}(r,\omega) $ is a solution of the radial Teukolsky equation: $${d^2 u_{lm} \over dr_\ast^2} + \left[ {K^2 + (2am\omega -a^2\omega^2 - E_{lm})\Delta \over (r^2+a^2)^2} -{dG \over dr_\ast} - G^2 \right] u_{lm} = 0 \label{Teuk}$$ where $K=(r^2 + a^2)\omega - am$ and $G=r\Delta/(r^2+a^2)^2$. Furthermore, $E_{lm}$ denotes the angular eigenvalue, cf. Appendix C. As usual, $\Delta= r^2 -2Mr + a^2$ and the “tortoise” radial coordinate $ r_{\ast}$ is defined as (with $r_{\pm}$, the two solutions to $\Delta =0$, denoting the event horizon and the inner Cauchy horizon of the black hole) $$r_{\ast} = r + \frac{2Mr_{+}}{r_{+}-r_{-}}\ln \left ( \frac{r}{r_{+}} -1 \right ) -\frac{2Mr_{-}}{r_{+}-r_{-}}\ln \left ( \frac{r}{r_{-}} -1 \right ) + c \label{rstar}$$ Usually, the arbitrary integration constant $c$ is disregarded in this relation. However, in scattering problems it turns out to be useful to keep it, as we shall see later. We are interested in a causal solution to (\[Teuk\]) which describes waves that are purely “ingoing” at the black hole’s horizon. This solution can be written $$u_{ lm}^{\rm in} \sim \left\{ \begin{array} {ll} e^{-ikr_\ast} \quad \mbox{as } r\to r_+ \ , \\ A^{\rm out}_{lm} e^{i\omega r_\ast} + A^{\rm in}_{lm} e^{-i\omega r_\ast} \quad \mbox{as } r\to +\infty \ . \end{array} \right. \label{uin}$$ where $k= \omega -ma/2Mr_{+}= \omega - m\omega_+$. In addition, we want to impose an “asymptotic scattering boundary condition”. We want the total field at spatial infinity to be the sum of a plane wave plus an outgoing scattered wave. In other words, we should have $$\Phi(r,\theta,\varphi) \sim \Phi_{\rm plane} + \frac{1}{r} f(\theta,\varphi) e^{i\omega r_{\ast}} \quad \mbox{as} \quad r \to +\infty \label{condition}$$ where we have omitted the trivial time-dependence. All information regarding scattering is contained in the (complex-valued) scattering amplitude $f(\theta,\varphi)$. Note that, unlike in axially symmetric scattering the scattering amplitude will depend on both angles: $\theta$ and $\varphi$. Up to this point, we have used the term “plane wave” quite loosely. In the presence of a long-range field such as the Kerr gravitational field (which falls off as $\sim 1/r $ at infinity) we cannot write a plane wave in the familiar flat space form. This problem has been discussed in several papers, see [@Matzner; @Chrzanowski]. Remarkably, it turns out that in a black hole background the long-range character of the field is accounted for by a logarithmic phase-modification of the flat space plane-wave expression. In practice, the substitution $ r \to r_{\ast} $ is made in the various exponentials. In order to make this paper as self-contained as possible, we discuss this point in some detail in Appendix A. The asymptotic expression for a plane wave travelling along a direction making an angle $\gamma $ with the black hole’s spin axis, see Figure \[draw1\], is $$\Phi_{\rm plane}= e^{i\omega r_{\ast}(\sin\gamma \sin\theta \sin\varphi + \cos\gamma \cos\theta)} \label{ofplane}$$ where without any loss of generality we have assumed an amplitude of unity. We can decompose this plane wave in a way similar to (\[eq1\]) and (\[eq2\]); $$\Phi_{\rm plane} \approx \frac{1}{r}\sum_{m=-\infty}^{+\infty} \sum_{l=|m|}^{+\infty} c_{lm}^{(0)} u_{lm}^{(0)}(r,\omega) S_{lm}^{a\omega}(\theta)e^{im\varphi}$$ where $u_{lm}^{(0)}$ are asymptotic solutions of (\[Teuk\]). For $r \to \infty$ we have (see Appendix B), $$c_{lm}^{(0)}u_{lm}^{(0)}(r,\omega) \approx 2\pi S_{lm}^{a\omega}(\gamma) \left \{ (-i)^{m+1} e^{i\omega r_{\ast}} + i^{m+1} (-1)^{l+m} e^{-i\omega r_{\ast}} \right \} \label{plane_asym}$$ Similarly, the full field at infinity can be approximated as : $$\Phi \approx \frac{1}{r}\sum_{m=-\infty}^{+\infty}\sum_{l=|m|}^{+\infty} c_{lm}(A^{\rm in}_{lm} e^{-i\omega r_{\ast}}+ A^{\rm out}_{lm}e^{i\omega r_{\ast}})S_{lm}^{a\omega}(\theta) e^{im\varphi}$$ By imposing the scattering condition (\[condition\]), we can fix $c_{lm}$ by demanding that the ingoing wave piece of $\Phi -\Phi_{\rm plane}$ vanishes. After some straightforward manipulations we get for the scattering amplitude $$f(\theta,\varphi)= \frac{2\pi}{i\omega}\sum_{m=-\infty}^{+\infty} \sum_{l=|m|}^{+\infty}(-i)^{m}S_{lm}^{a\omega}(\theta)S_{lm}^{a\omega}(\gamma) e^{im\varphi}\left [ (-1)^{l+1} \frac{A^{\rm out}_{lm}}{A^{\rm in}_{lm}} -1 \right ] \label{amp_off}$$ By defining the “scattering matrix element” ${\cal S}_{lm}=(-1)^{l+1} A^{\rm out}_{lm}/A^{\rm in}_{lm}$ we can equivalently write $${\cal S}_{lm}= e^{2i\delta_{lm}}$$ where we have introduced the phase-shift $\delta_{lm}$. Thus we see that the phase-shifts $\delta_{lm}$ contain all relevant information regarding the scattered wave. It is worth emphasising that for non-axisymmetric scattering the phase-shifts will depend on both $l$ and $m$. Also, the $\delta_{lm}$ are in general complex valued in order to account for absorption by the black hole. For later convenience, we also point out that the full field at infinity can be written $$\Phi \sim \sin(\omega r_{\ast} + \delta_{lm} -\frac{l\pi}{2}) \quad \mbox{ as } r_\ast \to \infty$$ In the case of on-axis incidence ($\gamma=0$) the scattering amplitude simplifies considerably, and we get $$f(\theta)= \frac{2\pi}{i\omega}\sum_{l=0}^{+\infty} S_{l0}^{a\omega}(\theta)S_{l0}^{a\omega}(0) \left [ (-1)^{l+1} \frac{A^{\rm out}_{l}}{A^{\rm in}_{l}} -1 \right ] \label{amp_on}$$ Here we see that the outcome is no longer dependent on $m$, which is natural given the axial symmetry of the problem. Furthermore, it is easy to see that we recover the familiar Schwarzschild expression [@Nils] by setting $a=0$. The differential cross section (often simply called the cross section in this paper) is the most important “observable” in a scattering problem. It provides a measure of the extent to which the scattering target is “visible” from a certain viewing angle. As demonstrated in standard textbooks [@Newton], the differential cross section follows immediately from the scattering amplitude $$\frac{d\sigma}{d\Omega}= |f(\theta,\varphi)|^2 \ . \label{dcs}$$ This cross section corresponds to “elastic” scattering only, that is, it describes the angular distribution of the waves escaping to infinity. We can similarly define an “absorption cross section” but we shall not be concerned with this issue here. Nevertheless, as we have already pointed out, black hole absorption has an effect on the phase-shifts that are used to compute the cross section (\[dcs\]). The strategy then for a cross section calculation (for given black hole parameters and wave frequency) involves three steps: i) calculation of the phase-shifts $\delta_{lm}$, (or, equivalently, of the asymptotic amplitudes $A_{lm}^{\rm out/in}$), ii) calculation of the spheroidal harmonics $S_{lm}^{a\omega}(\theta)$, and finally iii) evaluation of the sums in (\[amp\_off\]) and/or (\[amp\_on\]) including a sufficiently large number of terms. Approximating the scattering amplitude -------------------------------------- In practice, the partial-wave sum calculation is problematic as it converges slowly. In fact, the sum is divergent for some angles. This is just an artifact due to the long-range nature of the gravitational field. No matter how far from the black hole a partial wave may travel, it will always “feel” the presence of the gravitational potential (that falls off as $1/r$). A similar behaviour is known to exist in Coulomb scattering. The divergence always occurs at the angle that specifies the incident wave’s propagation direction. That this will be the case is easily seen from the identity $$\sum_{l,m} S_{lm}^{a\omega}(\theta)e^{im\varphi}S_{lm}^{a\omega}(\gamma) e^{-im\pi/2}= \delta(\cos\theta -\cos\gamma)\delta(\varphi-\pi/2) \ , \label{orth}$$ which follows directly from the fact that the functions $ S_{lm}^{a\omega}(\theta) e^{im\varphi}$ form an orthonormal set. The corresponding identity for on axis incidence is $$\sum_{l} S_{l0}^{a\omega}(\theta)S_{l0}^{a\omega}(0) = \frac{1}{2\pi}\delta(\cos\theta -1) \ .$$ From (\[orth\]) we can deduce a peculiar feature: Although the scattering problem is physically insensitive to the actual $\varphi$ of the incidence direction, the specific value $\varphi=\pi/2$ is imposed by the above relations. Of course, this has no physical relevance since the problem at hand is axially symmetric and we can, without any loss of generality, assume an incoming wave travelling along the direction $(\theta,\varphi)=(\gamma,\pi/2)$. The fact that the Kerr gravitational field behaves asymptotically as a Newtonian one considerably simplifies the scattering amplitude calculation. We would expect that large $l$ partial waves (strictly speaking when $l/\omega M \gg 1 $) to essentially feel only the far-zone Newtonian field. In terms of the phase-shifts, we expect them to approach their Newtonian counterparts $ \delta_{lm} \to \delta_{l}^{N}$ asymptotically. In order to secure this matching we add to our phase-shifts an “integration constant” $-2\omega M \ln(4\omega M) + \omega M $. In this way, we also get $ r_{\ast} \to r_{\rm c} $, where $r_{\rm c}= r + 2M\ln(2\omega r)$ is the respective tortoise coordinate of the Coulomb/Newtonian problem. Such a manipulation is admissible given the arbitrariness in the choice of the constant $c$ in (\[rstar\]). In calculating the partial-wave sum for the scattering amplitude, it is convenient to split it into two terms: $$f(\theta,\varphi)= f_D(\theta,\varphi) + f_N(\theta,\varphi)$$ Here, $f_D(\theta,\varphi)$ represents the part of the scattering amplitude that carries the information of the main diffraction effects, while $f_{N}(\theta,\varphi)$ denotes the Newtonian (Coulomb) amplitude. Explicitly we have $$f_N(\theta,\varphi)= \frac{2\pi}{i\omega} \sum_{l,m} Y_{lm}(\theta) e^{im\varphi} Y_{lm}(\gamma) (-i)^m \left [ e^{2i\delta_l^N} -1 \right ]$$ where we have deliberately “forgotten” the spherical symmetry of the Newtonian potential (which would had allowed us to write $f_{N}$ as a function of $\theta$ only, and thus in terms of a sum over $l$). However, the Newtonian phase-shifts $\delta_l^{N}$ are still given by the well-known expression [@book], $$e^{2i\delta_l^N}= \frac{\Gamma(l+1-2i\omega M)} {\Gamma(l+1+2i\omega M)}$$ After simple manipulations we get $$f_N(\xi)= \frac{1}{2i\omega}\sum_{l=0}^{+\infty} (2l+1)P_l(\cos\xi) (e^{2i\delta_l^N} -1) \label{fN2}$$ where $\cos\xi= \cos\theta\cos\gamma + \sin\theta \sin\gamma \sin\varphi $. The sum in (\[fN2\]) is known in closed form [@book]; $$f_N(\xi)= M\frac{\Gamma(1-2i\omega M)}{\Gamma(1+2i\omega M)} \left [\sin\frac{\xi}{2} \right ]^{-2+4i\omega M} \label{fN}$$ From this we can see that $f_N(\theta,\varphi)$ diverges in the $\xi=0$ direction. Let us now focus on the “diffraction” amplitude $f_{D}(\theta,\varphi)$. It has the form $$f_D(\theta,\varphi)= \frac{2\pi}{i\omega}\sum_{m=-\infty}^{+\infty} \sum_{l=|m|}^{+\infty}(-i)^{m}e^{im\varphi} \left \{ S_{lm}^{a\omega}(\theta)S_{lm}^{a\omega}(\gamma)(e^{2i\delta_{lm}} -1) - Y_{lm}(\theta) Y_{lm}(\gamma) (e^{2i\delta_l^N} -1) \right \} \label{fDoff}$$ The corresponding on-axis expression is, $$f_D(\theta)= \frac{1}{2i\omega}\sum_{l=0}^{+\infty} \left \{ 4 \pi S_{l0}^{a\omega}(\theta) S_{l0}^{a\omega}(0) ( e^{2i\delta_l} -1) -(2l+1)P_l(\cos\theta) (e^{2i\delta_l^N} -1) \right \} \label{fDon}$$ One would expect the sums in (\[fDoff\]) and (\[fDon\]) to converge. This follows from the fact that for $l/\omega M \to \infty $ we have $\delta_{lm}\to \delta_{l}^{N}$ and $ S_{lm}^{a\omega}(\theta) e^{im\varphi} \to Y_{lm}(\theta,\varphi)$ [@Handler]. We introduce a negligible error by truncating the sums at a large value $l_{\rm max}$ (say). In practice, $l_{\rm max}$ need not be very large. We find that a value $\sim 30-50$ for $\omega M \lesssim 2 $ typically suffices. Since each partial wave can be labelled by an impact parameter $b(l)$ (see Section IID), the criterion for $l_{\max}$ to be a “good” choice, is that $ b(l_{\rm max}) \gg b_{\rm c}$, where $ b_{\rm c}$ is the (largest) critical impact parameter associated with an unstable photon orbit in the Kerr geometry. The truncation of the partial-wave sums will introduce interference oscillations in the final cross sections (roughly with a wavelength $2\pi/l_{\rm max}$ [@Handler]). These unphysical oscillations can be eliminated by following the approach of Handler and Matzner [@Handler]. For a chosen $\l_{\rm max}$ we add a constant $\beta$ to all the phase-shifts in (\[fDoff\]) and (\[fDon\]). This constant is chosen such that $\delta_{l_{\rm max},m} + \beta = \delta_{l}^{N} $. This means that the resulting cross section is effectively smoothed. Deflection functions -------------------- It is well-know that the so-called deflection function is of prime importance in scattering problems. It arises in the semiclassical description of scattering, as discussed in the pioneering work of Ford and Wheeler [@Ford]. Although these authors considered scattering in the context of quantum theory, their formalism is readily extended to the black-hole case. In the semiclassical paradigm, the phase-shifts are approximated by a one-turning point WKB formula (typically useful for $l$ much larger than unity). In a problem which has only one classical turning point, the deflection function is defined as $$\Theta(l)= 2\frac{d\delta^{\rm WKB}}{dl}$$ where $l$ is assumed to take on continous real values. As a convention, the deflection function is negative for attractive potentials. The right-hand side of this equation resembles the expression for the deflection angle of classical motion in the given potential, provided that we define the following effective impact parameter $b$ for the wave motion [@Newton] $$b= \frac{l+ 1/2}{\omega} \label{impact}$$ The black-hole effective potential has two turning points, but for $l/\omega M \to \infty $ the scattering is mainly due to the outer turning point and one can derive a one turning point WKB approximation for the phase shifts. For a Schwarzschild black hole this expression is [@Nils] $$\delta_{l}^{\rm WKB}= \int_{t}^{+\infty} \left [ Q_{\rm s} - \left (1 -\frac{2M}{r} \right )^{-1} \omega \right ] dr - \omega t_{\ast} + (2l +1)\frac{\pi}{4} \label{d_wkb1}$$ where $$Q_{\rm s}^2= \left ( 1 -\frac{2M}{r} \right )^{-2} \left [ \omega^2 - \left ( 1-\frac{2M}{r} \right ) \frac{l(l+1)}{r^2} + \frac{M^2}{r^4} \right ]$$ Here $t_{\ast}$ denote the value of the tortoise coordinate corresponding to the (outer) turning point $t$. We now define the deflection function as $$\Theta(l)= 2 Re \left [ \frac{d \delta_{l}}{dl} \right ] \label{def_sch}$$ where only the real part of the phase shift is considered, as the whole discussion is relevant for elastic scattering only. Using (\[d\_wkb1\]), we find that the real scattering angle is $$\Theta(l)= \pi -2\left ( \frac{l+1/2}{\omega} \right ) \int_{t}^{+\infty} \frac{dr}{r^2} \left [ 1- \left ( 1-\frac{2M}{r} \right ) \left (\frac{l+1/2}{\omega r} \right )^2 + {\cal O }(M^2/r^2) \right ]^{-1/2} \label{def_sch2}$$ This WKB result should be compared to the deflection angle for a null geodesic in the Schwarzschild geometry, which is given by $$\Theta_c(b)= \pi - 2 b\int_{t}^{+\infty} \frac{dr}{r^2} \left [ 1- \left ( 1-\frac{2M}{r} \right )\frac{b^2}{r^2} \right ]^{-1/2} \label{def_cl}$$ where $ b= L_{z}/E $ is the orbit’s impact parameter ($L_{z}$ and $E$ denote, respectively, the orbital angular momentum component along the black hole’s spin axis and the orbital energy) and $t$ is the (classical) turning point. In writing down these expressions we have chosen the signs in such a way that the deflection angle is negative for attractive potentials. Clearly, it is possible to “match” the deflection function (\[def\_sch\]) with the classical deflection angle, albeit only at large distances. Since the effective impact parameter will be given by (\[impact\]), it is clear that in (\[def\_sch2\]) the integral will be over large $r$ only. Owing to its clear geometrical meaning the deflection function is an exceptionally useful tool in scattering theory. It can be used to define diffraction phenomena like glories, rainbows etc. [@Ford]. For example, in axisymmetric scattering backward glories are present if the deflection function takes on any of the values $\Theta=-n\pi$, where $n$ a positive odd integer. It seems natural to try and define deflection functions for Kerr scattering as well. In general, we anticipate the need for two deflection functions $\Theta(l,m)$ and $\Phi(l,m)$ (with only the first being relevant for the special case of on-axis scattering). The WKB phase-shift formula becomes in the Kerr case: $$\delta_{lm}^{\rm WKB}= \int_{t}^{+\infty} \left [ Q_{\rm k} - \left ( \frac{r^2 + a^2}{\Delta} \right ) \omega \right ] dr - \omega t_{\ast} + (2l +1)\frac{\pi}{4} \label{d_wkb2}$$ where $$Q_{\rm k}^2= \frac{1}{\Delta^2} \left [ K^2 - \lambda\Delta + M^2 -a^2 \right ]$$ where $\lambda = E_{lm} + a^2 \omega^2 - 2am\omega$. The next step is to derive the deflection angles for null geodesics approaching a Kerr black hole from infinity. Such orbits are studied in detail in [@Chandra]. From the results in [@Chandra] it is clear that it is not easy to write down a general expression for the deflection angle in the Kerr case. But we can obtain useful results in two particular cases. We begin by considering a null ray with $L_z=0$ (which would correspond to an axially incident partial wave). For such an orbit we find that the deflection angle $\Theta_{\rm c}$ obeys the following relation $$\int_{\pi}^{\Theta_{\rm c}(\eta)} d\theta \left [ 1 + \frac{a^2}{\eta^2} \cos^2\theta \right ]^{-1/2} = -2\eta\int_{t}^{+\infty} \frac{dr}{r^2} \left [ 1 -\frac{\eta^2}{r^2} \left (1 -\frac{2M}{r} + \frac{a^2}{r^2} \right ) + \frac{a^2}{r^2} + \frac{2a^2 M}{r^3} \right ]^{-1/2} \label{latdefc}$$ where $\eta= C^{1/2}/E$, with $C$ denoting the orbit’s Carter constant. This expression is valid provided the ray’s $\theta$-coordinate varies monotonically during scattering. This should be true in the cases we are interested in, at least for large impact parameters such that $ \eta \gg M$. The ray will also be deflected in the $\varphi$-direction but this deflection carries no information regarding plane-wave scattering due to the axisymmetry of the problem. We next consider a null ray travelling in the black hole’s equatorial plane. This situation will be particularly relevant for a plane wave incident along $\gamma=\pi/2$. The net azimuthal deflection $\Phi_{\rm c}(b)$ for an impact parameter $b= L_{\rm z}/E$ is $$\Phi_{\rm c}(b)= \pi - 2 b \int_{t}^{+\infty} \frac{dr}{r} \left [ 1 - \frac{a^2}{\Delta} + \frac{2aMr}{\Delta b} \right ] \left [ r^2 + a^2 + \frac{2a^2 M}{r} - b^2 \left ( 1 -\frac{2M}{r} \right) -\frac{4aMb}{r} \right ]^{-1/2} \label{azimdefc}$$ Working to the same accuracy in terms of $M/r$ as in the Schwarzschild case, we can match (\[latdefc\]) and (\[azimdefc\]) to $\partial\delta_{lm}/\partial l$. This matching becomes possible if we use the following approximate expression for the eigenvalue $E_{lm}$ [@Handler] $$E_{lm}\approx l(l+1) - \frac{1}{2}a^2\omega^2 + {\cal O}(\frac{a^3\omega^3}{l})$$ As in the Schwarzschild case we assume that the effective impact parameter is given by (\[impact\]). Although there is no occurrence of the multipole $m$ in the above expressions, one can argue (from the symmetry of the various spheroidal harmonics, which is similar to that of the spherical harmonic of the same ($l,m)$) that the classical angles (\[latdefc\]) and (\[azimdefc\]) are related to partial waves with $m=0$ and incidence $\gamma=0$ and partial waves with $m= \pm l$ and incidence $\gamma=\pi/2$, respectively. Hence, we define the latitudinal deflection function $$\Theta(l) = 2Re \left [ \frac{\partial\delta_{lm}}{\partial l}(m=0) \right ] \label{latdef}$$ and the azimuthal (“equatorial”) deflection function $$\Phi(l)= 2Re \left [ \frac{\partial\delta_{lm}}{\partial l}(m=\pm l) \right ] \label{azimdef}$$ Numerical results ================= Phase-shifts calculation via the Prüfer transformation ------------------------------------------------------ In order to determine the required scattering phase-shifts we have used a slightly modified version of the simple Prüfer transformation, well-known from the numerical analysis of Sturm-Liouville problems [@Pryce]. The method is best illustrated by a standard second order ordinary differential equation: $$\frac{d^2 \psi}{dx^2} + U(x) \psi=0 \label{sle}$$ where we can think of $x$ as being a radial coordinate, spanning the entire real axis, and $U$ an effective potential (in our problem corresponding to a single potential barrier) with asymptotic behaviour $$U(x) \sim \left\{ \begin{array} {ll} k^2 \quad \mbox{as } x \to -\infty \ , \\ \omega^2 \quad \mbox{as } x \to +\infty \ . \end{array} \right.$$ with $\omega$ and $k$ real constants. (The black hole problem we are interested in does, of course, have exactly this nature.) The solution of (\[sle\]) will take the form of oscillating exponentials for $x \to \pm \infty$. Let us assume that we are looking for a solution to (\[sle\]) with purely “ingoing” behaviour at the “left” boundary ($x \to - \infty$) and mixed ingoing/outgoing behaviour as $x \to + \infty$: $$\psi \sim \left\{ \begin{array}{ll} e^{-ikx} \quad \mbox{as } x \to -\infty \ , \\ B \sin[\omega x + \zeta ] \quad\mbox{as } x \to +\infty \ . \end{array} \right.$$ where $\zeta$ and $B$ are complex constants. We can then write the exact solution of (\[sle\]) in the form $$\psi(x)= e^{\int P(x) dx}$$ The function $P(x)$ is the logarithmic derivative of $\psi(x)$ (a prime denotes derivative with respect to $x$) $$\frac{\psi^{\prime}}{\psi}= P$$ which obeys the boundary condition $P(x) \to -ik$ for $x \to -\infty$. Similarly, we can express the function $\psi$ and its derivative via a Prüfer transformation; $$\begin{aligned} \psi(x) &=& B\sin[\omega x +\tilde{P}(x)] \\ \psi^{\prime}(x)&=&B \omega \cos[\omega x + \tilde{P}(x) ]\end{aligned}$$ with $\tilde{P}(x)$ a Prüfer phase function which has $\zeta$ as its limiting value for $x \to + \infty$. Direct substitution in (\[sle\]) yields the equations $$\frac{dP}{dx} + P^2 + U(x) = 0 \label{P}$$ $$\frac{d\tilde{P}}{dx} + \left [ \omega -\frac{U(x)}{\omega} \right ] \sin^2(\tilde{P} + \omega x)= 0 \label{P2}$$ The idea is to numerically integrate (\[P\]) and (\[P2\]) instead of the original equation (\[sle\]). The motivation for this is that, while the original solution may be rapidly oscillating, the phase-functions $P$ and $\tilde{P}$ are expected to be slowly varying functions of $x$. We expect this integration scheme to be considerably more stable, especially for high frequencies, than any direct approach to (\[sle\]). Moreover, eqs. (\[P\]) and (\[P2\]) are well behaved also at the classical turning points and are well suited for barrier penetration problems. However, if we want to ensure that the phase-functions are smooth and non-oscillatory we must account for the so-called Stokes phenomenon — the switching on of small exponentials in the solution to an equation of form (\[sle\]). To do this we simply shift from studying $P(x)$ (which is calculated from $ x=-\infty $ up to the relevant matching point $x_{\rm m}$) to $\tilde{P}(x)$ (which is calculated outwards to $x=+\infty$). In practice, the calculation is stable and reliable if the switch is done in the vicinity of the maximum of the black-hole potential barrier (the essential key is to not use one single representation of the solution through the entire potential barrier). The two phase functions are easily connected by $$\tilde{P}(x)= -\omega x + \frac{1}{2i} \ln\left[ \frac{iP -\omega} {iP +\omega} \right ]$$ Finally, the desired phase-shift can be easily extracted as $\delta_{lm}= \zeta + l\pi/2$. A major advantage of the adopted method is that it permits direct calculation of the partial derivatives $\partial \delta_{lm}/\partial l~$ and $\partial\delta_{lm}/\partial m$, which are required for the evaluation of the deflection functions from Section IID. The equations for the $l,m$-derivatives of $P$ and $\tilde{P}$ are simply found by differentiation of (\[P\]) and (\[P2\]). For the case of scattering by a Kerr black hole, calculation of $\delta_{ lm}$ and its derivatives with respect to $l,m$ requires knowledge of the angular eigenvalue $E_{lm}$ (see Appendix C) and its $l,m$-derivatives. We have used an approximate formula which is a polynomial expansion in $a\omega$ (formula $21.7.5$ of [@Stegun]). This expression (and its derivatives) is well behaved for all integer values of $l$ and $m$. However, it is divergent for the half-integer values $ l= 1/2, 3/2, 5/2 $. Hence, the numerical calculation of the deflection function will fail at these points, and will be generally ill-behaved in their neighbourhood. For the full cross section calculation, we additionally need to calculate the spin-0 spheroidal harmonics $S_{lm}^{a\omega}(\theta)$. This calculation is discussed in detail in Appendix C. Schwarzschild results --------------------- In this section we reproduce phase-shifts and cross sections for Schwarzschild scattering. The purpose of this exercise is to validate, and demonstrate the reliability of, our numerical methods. We compare our numerical integration results to ones obtained using the phase-integral method [@Nils]. As a first crucial test we compare, in Fig. \[fig1\], the first 100 phase-shifts for $\omega M=1$. Because of the multi-valued nature of the phase-shifts we always plot the quantity ${\cal S}_l=e^{2i\delta_l}$. As is evident from Fig. \[fig1\], the agreement between our numerical phase-shifts and the phase-integral ones is excellent. This is equally true for a all frequencies examined (up to $\omega M=10$). It should be noted that ${\cal S}_l$ is essentially zero ($\delta_l$ has a positive imaginary part) for those values of $l$ for which absorption by the black hole is important. That this is the case for the lowest multipoles is clear from Fig. \[fig1\]. As $l$ increases $\delta_l$ becomes almost real and as a consequence ${\cal S}_l$ is almost purely oscillating. In Fig. \[fig2\] we present an $\omega M=10$ cross section generated from our numerical phase-shifts. For this particular calculation we have used $l_{\rm max}= 200$. The resulting cross section matches the one constructed using approximate phase-integral phase shifts perfectly. This demonstrates the efficiency of our approach in the high frequency regime, and it is clear that our study of Kerr scattering will not be limited by the lack of reliable phase shifts. However, the Kerr study is nevertheless limited in the sense that $\omega$ cannot be taken to be arbitrarily large. This restriction is imposed by the calculation of the spheroidal harmonics (Appendix C). However, it is important to emphasize that the most interesting frequency range, as far as diffraction phenomena is concerned, is $\omega M \sim 2$ [@Nils]. Thus, we expect that our investigation should be able to reliable unveil all relevant rotational effects in the scattering problem. We also find that the numerical values for $d\delta_l/dl$ are in good agreement with the respective phase-integral results. As a final remark we should emphasize that the numerical approach adopted in this work is very efficient from a computational point of view. On-axis Kerr scattering ----------------------- Having confirmed the reliability of our numerical results we now turn to the study of scattering of axially incident scalar waves by a Kerr black hole, cf. Fig. \[draw2\]. In principle, we would expect the corresponding cross sections to be qualitatively similar to the Schwarzschild ones. The main reason for this is the inability of axially impinging partial waves to distinguish between prograde and retrograde orbits. However, examination of the orbital equations [@Chandra] reveals that the critical impact parameter (associated with the unstable photon orbit) decrease sligthly from the value $3\sqrt{3}M $ as the black hole spins up. For example, for a=0.99M we have $b_{c}= 4.74M$. Indeed, Fig. \[figoncs\] confirms our expectations. The data in the figure corresponds to a black hole with spin $a=0.99M$ and a wave frequency of $\omega M=2$. For purposes of comparison, we also show the corresponding Schwarzschild cross section. The two cross sections are very similar. In particular, they are both dominated by the backward glory. It is well-known [@book; @path_int1] that for Schwarzschild scattering the glory effect can be described in terms of Bessel functions. It has been shown that for $\theta \approx \pi$ the glory cross section can be approximated by $$\frac{d\sigma}{d\Omega}(\theta)|_{\rm glory} \propto J^{2}_{0}[\omega b_{\rm c} \sin\theta] \label{glory}$$ A similar result holds for the backward glory in the case of on-axis Kerr scattering. Since $b_c$ gets smaller, one would expect the zeros of the Bessel function (the diffraction minima) to move further away from $\theta = \pi$ as $a$ increases. This effect is indicated by the data in Fig. \[figoncs\]. The above conclusions are further supported by the results for the deflection function (\[latdef\]), as shown in Fig. \[figThet\]. Because of the inaccuracies inherent in our method of calculating the deflection function for the lowest $l$-multipoles, see the discussion in Section IIIA, we do not show results for this regime. This is, however, irrelevant as the corresponding partial waves are expected to be more or less completely absorbed by the black hole. Off-axis Kerr scattering ------------------------ The conclusions of our study of on-axis scattering are perhaps not very exciting. Once the Schwarzschild case is understood, the on-axis results for Kerr come as no surprise. This is, however, not the case for off-axis scattering, cf. Fig. \[draw3\], where several new features appear. Our study of the off-axis case provides the first results for non-axisymmetric wave scattering in black hole physics. Since this problem has not been discussed in great detail previously, it is worthwhile asking whether we can make any predictions before turning to the numerical calculations. Two effects ought to be relevant: First of all, the partial waves now have orbital angular momentum which couples to the black hole’s spin. As a result the partial waves can be divided into prograde ($m >0$) and retrograde ($m<0$) ones. We expect prograde waves to be able to approach closer to the horizon than retrograde ones. In the geometric optics limit, prograde and retrograde rays tend to have increasingly different critical impact parameters as $a \to M$. As a second feature, we expect to find that large $l$ partial waves will effectively feel only the spherically symmetric (Newtonian) gravitational potential. In other words, partial waves with the same (large) $l$ and different values of $m$ will approximately acquire the same phase-shift. Our numerical results essentially confirm these expectations, as is clear from the phase-shifts (calculated for $a=0.9M$ and $\omega M=1$) shown in Fig. \[figofps\]. As above, we have graphed the single-valued quantity ${\cal S}_{lm}=e^{2i\delta_{lm}}$ as a function of $l$. For each value of $l$ we have included all the phase-shifts for $ -l \leq m \leq +l$. The solid (dashed) line corresponds to $m=+l$ ($m=-l$) and the intermediate values of $m$ lead to results in between these two extremes. For $l \gg 1$, partial waves with different values of $m$ have almost the same phase-shift. This is easy to deduce from the fact that the two curves approach each other as $l$ increases. On the other hand, for the first ten or so partial waves we get very different results for the various values of $m$. In particular, we see that phase-shifts with $m>0$ become almost real (that is, $| {\cal S}_{lm}|$ becomes non-zero) for a smaller $l$-value as compared to the $m<0$ ones. As anticipated, this is due to the different critical impact parameters associated with prograde/retrograde motion, and the fact that a larger number of prograde partial waves are absorbed by the black hole. We now turn to the cross section results for the off-axis case. We have considered a plane wave incident along the direction $\gamma=\varphi=\pi/2$. Even though our formalism allows incidence from any direction we have focussed on this case, which is illustrated in Fig. \[draw3\]. The motivation for this is that there will then be partial waves (specifically the ones with $m=\pm l$) that are mainly travelling in the black hole’s equatorial plane. These partial waves are important because one would expect them to experience the strongest rotational effects. Besides, we can obtain an understanding of these waves by studying equatorial null geodesics in the geometric optics limit. Equatorial null rays are much easier to describe than nonequatorial ones. This proves valuable in attempts to “decipher” the off-axis cross sections, and the obtained conclusions provide an understanding also of the general case. In Fig. \[fig6\] we present a series of cross sections as functions of $\varphi$ for the specific values $\theta= \pi/8, \pi/4, 3\pi/8, \pi/2 $. These results correspond to viewing the scattered wave on the circumference of cones (like that shown in Fig. \[draw3\]) with increasing opening angles. Two different frequencies $\omega M=1 $ and $\omega M=2$ have been considered for a black hole with spin $a=0.9M$. A first general remark concerns the asymmetry of the cross sections with respect to the incidence direction (note, however, that as a consequence of our particular choice of incidence direction there is still a reflection symmetry with respect to the equator). We can also easily distinguish the Coulomb forward divergence in the direction $\theta=\varphi=\pi/2$. Another obvious feature in Fig. \[fig6\] is the markedly different appearance of the cross section for different values of $\theta$. As we move away from the equatorial plane the cross sections becomes increasingly featureless. This behaviour is artificial in the sense that as $\theta$ decreases, we effectively observe along a smaller circumference. At $\theta=0$ this circumference degenerates into a point, cf. Fig. \[draw3\]. In order to understand the features seen in Fig. \[fig6\] further, we focus on the $\theta=\pi/2$ cross section. In Fig. \[var\_a\] we show these “equatorial” cross sections for a sequence of spin rates $a/M= 0.2, 0.5, 0.7, 0.9 $. As before, we have considered two different wave frequencies, $\omega M=1$ and $\omega M=2$. From the results shown in Fig. \[var\_a\] it is clear that that the glory maximum is typically not observed in the backward ($\varphi= -\pi/2$) direction. In fact, it is clear that the maximum of the glory oscillations move away from the backward direction as the spin of the black hole is increased. A similar shift is seen in all interference oscillations. This behaviour is easy to explain in terms of the anticipated rotational frame-dragging. In order to illustrate this argument, we consider the geometric optics limit where partial waves are represented by null rays. Recall that in axisymmetric scattering the backward glory is associated with the divergence of the classical cross section at $\Theta=\pi$ in such a way that $$\left (\frac{d\sigma}{d\Omega}\right )_{\rm cl}= \frac{b}{\sin\Theta} \left (\frac{d\Theta}{db} \right )^{-1}$$ The divergence is a result of the intersection of an infinite number of rays. For simplicity, let us consider rays travelling in a specified plane. Scattering near the backward direction by a Schwarzschild black hole is illustrated in Fig. \[fignulls\] (left panel). Two rays with different impact parameters emerge at any given angle. These two waves make the main contribution to cross section at that particular angle. It is clear that for scattering at $\theta=\pi$ the two rays in Fig. \[fignulls\] will follow symmetric trajectories. This means that when observed at infinity, after being scattered, the two waves will have equal phases (provided their initial phases were equal). In effect, these two rays will then constructively interfere in the exact backward direction. As we move away from the backward direction, we should observe a series of interference maxima and minima — the two rays will now have an overall phase difference since they follow different orbits (see Fig. \[fignulls\]). A very crude estimate of the location of the successive maxima would be $\theta_{\rm n} \sim n\pi/3\omega M$ where $n=0,1,2, ...$, in reasonable agreement with the exact results. Similar arguments apply in the case of a Kerr black hole. We shall consider only equatorial rays, cf. Fig \[fignulls\] (right panel). As a result of the discrimination between prograde and retrograde orbits, the two rays contributing to the cross section in the exact backward direction will no longer follow symmetric paths. In fact, the ray symmetric to the prograde ray shown in Fig \[fignulls\] will follow a plunging orbit. Therefore, we should not expect the interference maximum to be located in the exact backward direction. An estimate (based on a crude calculation of the phase difference between the two null rays) of the location of the main backward glory maximum yields $$\varphi_{\rm max} \sim \pi \left ( \frac{r_{\rm ph+} -r_{\rm ph-} } {r_{\rm ph+} +r_{\rm ph-}} \right )$$ where $ r_{\rm ph+}$ and $ r_{\rm ph-}$ denote, respectively, the location of the prograde and retrograde unstable photon orbits (in Boyer-Lindquist coordinates). This angle is measured from the backward direction in the direction of the black hole’s rotation. This simple prediction agrees reasonably well with the results inferred from our numerical cross sections. In a similar way, all other maxima and minima will be frame-dragged in the black hole’s rotational direction. To complete this discussion we consider the deflection function $\Phi(l,m)$ for “equatorial” partial waves ($m=\pm l$), an $a=0.9M$ black hole and $\omega M=1$. The corresponding data is shown in Fig. \[figPhi\]. There are two distinct logarithmic divergences which are associated with the existence of separate unstable circular photon orbits for prograde and retrograde motion. Note that for $m>0$ the deflection function diverges steeper than it does for $m<0$. The origin of this effect is the fact that prograde partial waves with $ b \sim b_{\rm c}$ perform a greater number of revolutions (before escaping to infinity) than retrograde ones. Finally, for $ |m|\gg 1$ we recover, as expected, the Einstein deflection angle $\Phi \approx -4M/b $ (not explicitly shown in the figure). Digression: forward glories --------------------------- The results presented in the preceeding sections clearly show that, in general, black hole cross sections are dominated by a “Coulomb divergence” in the forward direction and (frame-dragged) glory oscillations near the backward direction. However, according to the predictions of geometrical optics [@book], one would expect to find glory oscillations also in the forward direction (see comments in [@Nils]). For the case of Schwarzschild scattering, this effect would be associated with partial waves scattered at angles $\Theta= 0, -2\pi, -4\pi, ...$. Inspection of the relevant deflection function (Fig. \[figThet\]) indicates that a partial wave which has $\Theta=0$ will also be strongly absorbed, since it has an impact parameter $b < b_{\rm c}$. Hence, we would expect its contribution to the forward glory to be severely supressed. It thus follows that, as far as the possible forward glory is concerned, the most important partial waves are those with $\Theta= -2\pi$. These partial waves “whirl” around the black hole as they have $b \approx b_{\rm c}$. Ford and Wheeler’s semiclassical approach [@Ford] shows that the forward glory is well approximated (for $\theta \approx 0$) by (\[glory\]), although with a slightly different proportionality factor. However, we should obviously not expect to see a pronounced forward glory in the cross section, as it will drown in the forward Coulomb divergence. Still, as an experiment aimed at supporting our intuition, we can try to “dig out” the forward glory pattern. This has to be done in a somewhat artificial manner, but since the forward glory is due to scattering and interference of partial waves with $b \approx b_{\rm c} $ we can isolate their contribution by truncating the sum in $f_D$ at some $l_{\rm max} \sim \omega b_{\rm c} $ and at the same time neglecting the Newtonian part $f_N$ entirely. It is, of course, important to realize that this “truncated” cross section is not a physical (observable) quantity. In Fig. \[Sfglor\] we show the result of this “truncated cross section” calculation for the case of a Schwarzschild black hole and $\omega M=2 $. We compare results for two levels of truncation, $l_{\rm max}= 10 $ and $15 $. In the first case, a clear Bessel- function like behaviour arises (it is straightforward to fit a $ J_{0}^2(\omega b_{\rm c} \sin\theta) $ function to the solid curve in Fig. \[Sfglor\]). This confirms our expectation that there is, indeed, a forward glory present in the data. As more partial waves are included the cross section begins to deviate from the glory behaviour, and if we increase $l_{\rm max}$ further the forward glory is swamped by terms that contribute to the Coulomb divergence. The role of superradiance in the scattering of monochromatic waves. ------------------------------------------------------------------- Superradiance is an interesting effect known to be relevant for rotating black holes. It is easily understood from the asymptotic behaviour (\[uin\]) of the causal solution to the scalar-field Teukolsky equation (\[Teuk\]). If we use this solution and its complex conjugate, and the fact that two linearly independent solutions to (\[Teuk\]) must lead to a constant Wronskian, it is not difficult to show that $$( 1 - m \omega_+/\omega ) |{\cal T}_{lm}|^2 = 1 - |{\cal S}_{lm}|^2 \ .$$ where we have defined $$|{\cal T}_{lm}|^2 = \left| {1 \over A^{\rm in}_{lm} } \right|^2 \ .$$ From the above result it is evident that the scattered waves are amplified ($|{\cal S}_{lm}|^2>1$) if $\omega < m\,\omega_+$. This amplification is known as superradiance. In principle, one would expect superradiance to play an important role in the scattering problem for rapidly spinning black holes. For example, one could imagine that some partial waves which would otherwise be absorbed, could escape back to infinity. These waves might then possibly make a noticeable contribution to the diffraction cross section, provided that there were a sufficient number of them (as compared to the total number $l_{\rm max}$ of partial waves contributing to the diffraction scattering amplitude). In order to investigate this possibility, we have performed a number of off-axis cross section calculations for a variety of wave frequencies $\omega M= 0.5 - 10$ and for $a \approx M$, i.e. black holes spinning near the extreme Kerr limit. We have found no qualitive difference whatsoever between those cross-sections and the ones for a somewhat smaller spin value, $a=0.9M$ (say). In essence, we were unable to find any effects in the cross section that could be attributed to superradiance. Consequently, we are led to suspect that our intuition regarding the importance of superradiance for the scattering problem may be wrong. This suspicion is confirmed by the following simple argument. In order for a partial wave to be superradiant we should have $ 0 < \omega < m\omega_{+}$. Considering an extreme Kerr black hole (which provides the best case for superradiant scattering) and the fact that $ m \leq l$, we have the condition $$0 < 2\omega M < l \label{sr1}$$ As already mentioned, the partial waves for which superradiance will be important are the ones with impact parameters $b \lesssim b_{\rm c}$, i.e. those that would be absorbed under different circumstances. This then requires that $$l \lesssim \omega b_{\rm c} -1/2 \label{sr2}$$ Combining (\[sr1\]) and (\[sr2\]) we arrive at the inequality $$0 < 2M \lesssim b_{\rm c} - 1/2\omega \label{sr3}$$ The critical impact parameter (for prograde motion) for an $a=M$ black hole is $b_{\rm c} \approx 2M$. Hence the condition (\[sr3\]) will not be satisfied, and it is unlikely that we would get a significant number of (if any) superradiant partial waves that could affect the cross section. This conclusion may seem surprising given results present in the literature [@book; @Handler]. In particular, Handler and Matzner have briefly discussed the effect of superradiance on axially incident gravitational waves. They argue that (see figure 14 in [@Handler]) “superradiance has the effect of imposing a large background over the pattern, filling in the interference minima”. Given our current level of understanding (or lack thereof) we cannot at this point say whether superradiance can be the explanation for the effects observed by Handler and Matzner. After all, one should remember that superradiant scattering strongly depends on the spin of the field that is being scattered. It is well known [@Press] that gravitational perturbations can be amplified up to $138\%$ compared to a tiny $0.04\%$ amplification for scalar fields (which is the case considered in this paper). This means that superradiance may significantly affect also partial waves with $b > b_{\rm c}$ in the gravitational wave case, which could lead to our simple argument not being valid. This issue should be addressed by a detailed study of the scattering of gravitational waves from rotating black holes. Concluding Discussion ===================== We have presented an investigation of scattering of massless scalar waves by a Kerr black hole. Our numerical work is based on phase-shifts obtained via integration of the relevant radial wavefunction with the help of the Prüfer phase-function method. This method has been shown to be computationally efficient and to provide accurate results. Using the obtained phase-shifts we have constructed differential cross sections for several different cases. First we have discussed the case of waves incident along the black hole’s rotation axis, for which we showed that the resulting cross sections are similar to ones obtained in the (non-rotating) Schwarzschild case. We then turned to the case of off-axis incidence, where the situation was shown to change considerably. In that case the cross sections are generically asymmetric with respect to the incidence direction. The overall diffraction pattern is “frame dragged”, and as a result the backward glory maximum is shifted along with of the black hole’s rotation. Moreover, we have concluded that (at least for scalar waves) the so-called superradiance effect is unimportant for monochromatic scalar wave scattering. To summarize, our study provides a complete understanding of the purely rotational effects involved in black-hole scattering. Given this we are now well equipped to proceed to problems of greater astrophysical interest, particularly ones concerning gravitational waves. In these problems one would expect further features to arise as the spin and polarisation of the impinging waves interact with the spin of the black hole. For incidence along the hole’s spin axis, one can have circularly polarised waves which are either co- or counter-rotating. The two cases can lead to quite different results. Although the general features of the corresponding cross sections are similar, they show different structure in the backward direction [@book]. This is possibly due to interference between the two polarisation states of gravitational waves, an effect that has not yet been explored in detail. Some initial work on gravitational-wave scattering has been done, see [@Handler], but we believe that the results of the present paper sheds new light on previous results, and could help interpret the rather complex cross sections that have been calculated in the gravitational-wave case. In this context, it should be stressed that the choice of studying scalar waves was made solely on grounds of clarity and simplicity. Our approach can readily be extended to other cases. Moreover, it is relevant to point out that a full off-axis gravitational wave scattering cross section calculation is still missing. We would expect such cross sections to be rather complicated, combining the frame-dragging effects discussed in this paper with various spin-induced features. We hope to be able to study this interesting problem in the near future. K.G. thanks the State Scholarships Foundation of Greece for financial support. N.A. is a Philip Leverhulme Prize Fellow, and also acknowledges support from PPARC via grant number PPA/G/1998/00606 and the European Union via the network “Sources for Gravitational Waves”. Plane waves in the Kerr geometry ================================ The long-range character of the gravitational field modifies the form of “plane waves”. This non-trivial issue has been discussed in the context of black hole scattering by Matzner [@Matzner] and Chrzanowski [*et al.*]{} [@Chrzanowski]. For completeness, we provide a brief discussion here. In a field-free region a monochromatic plane wave is, of course, given by the familiar expression $$\Phi_{\rm plane}= e^{i\omega r \cos\theta -i\omega t} \label{fplane}$$ when a spherical coordinate frame is employed. The plane wave is taken to travel along the $z$-axis. The field in (\[fplane\]) solves the the wave equation $\Box\Phi_{\rm plane}=0 $. Moreover, we can assume a decomposition of the form $$\Phi_{\rm plane}= \frac{e^{-i\omega t}}{r} \sum_{l} c_{l}^{(0)} u_{l}^{(0)} (r) P_{l}(\cos\theta) e^{-i\omega t}$$ where the radial wavefunction satisfies $$\frac{d^2 u_{l}^{(0)}}{dr^2} + \left [ \omega^2 -\frac{l(l+1)}{r^2} \right ] u_{l}^{(0)}= 0$$ Let us now consider a “plane wave” in the Schwarzschild geometry. First of all, we expect such a field to be only an asymptotic solution (as $r \to \infty $) of the full wave equation $\Box \Phi=0$ (where $\Box$ represents the covariant d’Alembert operator). The plane wave field can then be represented at infinity as $$\Phi_{\rm plane} \approx \frac{e^{-i\omega t}}{r} \sum_{l} c_{l}^{(0)} u_{l}^{(0)} (r) P_{l} (\cos\theta) e^{-i\omega t}$$ and the radial wavefunction will be a solution of $$\frac{d^2 u_{l}^{(0)}}{dr_{\ast}^2} + \left [ \omega^2 - \frac{l(l+1)}{r_{\ast}^2} + {\cal O} \left ( \frac{\ln r_{\ast}}{r_{\ast}^3} \right ) \right ] u_{l}^{(0)}= 0$$ This equation is similar to the corresponding flat space equation. The only difference is the appearance of the tortoise coordinate $r_{\ast}$ instead of $r$. Hence, we are inspired to write the plane wave field as $$\Phi_{\rm plane}= e^{i\omega r_{\ast} \cos\theta -i\omega t} \label{splane}$$ From this discussion, it should be clear that this form is valid only for $r \to \infty$. It is straightforward to see that in the same regime (\[splane\]) solves $\Box\Phi=0$. Expression (\[splane\]) is the closest we can get to the usual plane wave form (\[fplane\]). The long-range gravitational field is simply taken into account by an appropriate phase modification. Next, we consider a plane wave in Kerr geometry. For simplicity we take the $z$-axis to coincide with the black hole’s spin axis. We can then write $$\Phi_{\rm plane} \approx \frac{e^{-i\omega t}}{r} \sum_{l} c_{l}^{(0)} u_{l}^{(0)}(r) S_{l0}^{a\omega}(\theta) e^{im\varphi} \label{kplane}$$ The radial wavefunction is solution of $$\frac{d^2 u_{l}^{(0)}}{dr_{\ast}^2} + \left [ \omega^2 - \frac{\lambda + 2am\omega}{r_{\ast}^2} + {\cal O} \left (\frac{\ln r_{\ast}} {r_{\ast}^3} \right ) \right ] u_{l}^{(0)}=0 \label{kplane2}$$ It is obvious that both (\[kplane\]) and (\[kplane2\]) are different from the corresponding flat space expressions. That is, unlike in the Schwarzschild case, we are not able to derive an explicit form for a plane wave. Thus, we the following asymptotic expression for a plane wave travelling along the $z$-axis $$\Phi_{\rm plane}= e^{i\omega r_{\ast}\cos\theta -i\omega t} \label{kplane3}$$ where $r_{\ast}$ is the appropriate tortoise coordinate (\[rstar\]). The field given by (\[kplane3\]) is a solution of $\Box \Phi=0$ for $r \to \infty$. For the general case of a plane wave travelling along a direction that makes an angle $\gamma$ with the $z$-axis the appropriate expression is given by (\[ofplane\]). Asymptotic expansion of plane waves =================================== In this Appendix the asymptotic expansion of a plane wave in Kerr background is worked out. The calculation presented here is identical to the one found in the Appendix A1 of [@book], but here it is specialised to $s=0$. We have seen that for $r \to \infty $ the plane wave decomposition becomes $$e^{i\omega r_{\ast} [\sin\gamma\sin\theta\sin\varphi +\cos\gamma\cos\theta]} \approx \frac{1}{\omega r} \sum_{l,m} c_{lm}^{(0)} u_{lm}^{(0)}(r \to \infty) S_{lm}^{a\omega}(\theta) e^{im\varphi}$$ where $\gamma$ is the angle between the wave’s propagation direction and the positive $z$-axis. We multiply this expression by $S_{l^{\prime}m^{\prime}}^{a\omega}(\theta^{\prime})e^{-im^{\prime}\varphi^ {\prime}}$ and integrate over the angles to get (after a trivial change $l^{\prime} \to l$, $ m^{\prime} \to m$ at the end) $$c_{lm}^{(0)} u_{lm}^{(0)}(r \to \infty) \approx \omega r \int_{0}^{\pi} d\theta \sin\theta ~ S_{lm}^{a\omega}(\theta) e^{i\omega r_{\ast} \cos\gamma \cos\theta } \int_{0}^{2\pi} d\varphi e^{i\omega r_{\ast} \sin\gamma \sin\theta \sin\varphi -im\varphi } \label{c1}$$ The integration over $\varphi$ can be performed with a little help from [@Gradshteyn], and the result is $$\int_{0}^{2\pi} d\varphi e^{i\omega r_{\ast} \sin\gamma \sin\theta \sin\varphi -im\varphi }= 2\pi J_{m}(\omega r_{\ast} \sin\gamma \sin\theta )$$ Since the Bessel function has a large argument it can be approximated as [@Stegun], $$J_{\nu}(z) \approx \frac{1}{\sqrt{2\pi z}} \left ( e^{i(z- \nu\pi/2 -\pi/4)} + e^{-i(z -\nu\pi/2 -\pi/4)} \right )$$ Note that this approximation is legal as long as $\gamma \neq 0 $. The on-axis case $\gamma=0$ can be treated seperately, in a way similar to the one sketched here. Using this approximation in (\[c1\]) we get $$\begin{aligned} & & c_{lm}^{(0)}u_{lm}^{(0)}(r \to \infty) \approx \sqrt{\frac{2\pi\omega r} {\sin\gamma} } \left ( e^{-\frac{i}{2}(m\pi + \pi/2)} {\cal I}_{-} + e^{\frac{i}{2}(m\pi + \pi/2)} {\cal I}_{+} \right ) \\ && {\cal I}_{\pm} = \int_{0}^{\pi} d\theta \sqrt{\sin\theta} S_{lm}^{a\omega} (\theta) e^{i\omega r_{\ast} \cos(\theta \pm \gamma) }\end{aligned}$$ The ${\cal I}_{\pm} $ integrals can be evaluated using the stationary phase approximation. We obtain $$c_{lm}^{(0)} u_{lm}^{(0)} (r \to \infty) \approx 2\pi \left[ (-i)^{m+1} e^{i\omega r_{\ast}} S_{lm}^{a\omega}(\gamma) + i^{m+1} e^{-i\omega r_{\ast} } S_{lm}^{a\omega}(\pi -\gamma) \right]$$ We finally get (\[plane\_asym\]) by using the symmetry relation $$S_{lm}^{a\omega}(\pi -\theta)= (-1)^{l+m} S_{lm}^{a\omega}(\theta)$$ Calculation of spheroidal harmonics and their eigenvalues ========================================================= For the numerical calculation of the spheroidal harmonics we have adopted a “spectral decomposition” method, first developed by Hughes [@Scott] in the context of gravitational wave emission and radiation backreaction on particles orbiting rotating black holes. In the present work we have specialised this technique for the spin-$0$ spheroidal harmonics. The angular equation satisfied by $S^{a\omega}_{lm}(\theta)$ is, $$\frac{1}{\sin\theta} \frac{d}{d\theta} \left ( \sin\theta \frac{d S_{lm}^ {a\omega}}{d\theta} \right ) + [ (a\omega)^2 \cos^{2}\theta -\frac{m^2} {\sin^{2}\theta} + E_{lm} ]S^{a\omega}_{lm}=0 \label{spheroid}$$ where $E_{lm}$ denotes the corresponding eigenvalue. For the special case $a\omega=0$ we have $E_{lm}=l(l+1)$ and the solution of (\[spheroid\]) is the familiar spherical harmonic (here we are suppressing the dependence on $\varphi$) $$Y_{lm}(\theta)= \left [ \frac{2l+1}{4\pi} \frac{(l-m)!}{(l+m)!} \right ]^{1/2} P_{lm}(\cos\theta)$$ where $P_{lm}$ is the associated Legendre polynomial. It’s numerical calculation is quite straigthforward [@Recipes] based on the recurrence relation $$P_{lm}(x)= \frac{1}{l-m} [x(2l-1)P_{l-1,m} -(l+m-1)P_{l-2,m} ]$$ with “initial conditions” $$\begin{aligned} P_{mm}(x)&=&(-1)^m (2m-1)!!(1-x^2)^{m/2} \\ P_{m+1,m}(x) &=& (2m+1)xP_{mm}(x) \end{aligned}$$ We can always expand the spheroidal harmonic in terms of spherical harmonics, $$S_{lm}^{a\omega}(\theta)= \sum_{j=|m|}^{\infty} b_{j}^{a\omega} Y_{jm}(\theta) \label{sp_dec}$$ Substituting this spectral decomposition in (\[spheroid\]), multiplying with $Y_{lm}(\theta) $ and integrating over $\theta$ we get $$(a\omega)^2 \sum_{j=|m|}^{\infty} b_{j}^{a\omega} c_{jl2}^{m} -b_{l}^{a\omega} l(l+1) = -E_{lm} b_{l}^{a\omega} \label{step1}$$ where $$c_{jl}^{m}= 2\pi \int_{0}^{\pi} d\theta \sin\theta \cos^2\theta ~ Y_{lm}(\theta)Y_{jm}(\theta)$$ This integral can be evaluated in terms of Clebsch-Gordan coefficients [@Stegun] $$c_{jl}^{m}= \frac{1}{3} \delta_{lj} + \frac{2}{3} \sqrt{\frac{2j+1}{2l+1}} <j2m0|lm> <j200|l0>$$ It follows that $c_{jl}^{m} \neq 0$ only for $ j= l-1,l,l+1$. Then (\[step1\]) gives $$[(a\omega)^2 c_{l-2,l}^{m} ]b_{l-2}^{a\omega} + [(a\omega)^2 c_{l,l}^{m} -l(l+1) ] b_{l}^{a\omega} + [(a\omega)^2 c_{l+2,l}^{m} ] b_{l+2}^{a\omega}= -E_{lm} b_{l}^{a\omega} \label{step2}$$ We can rewrite (\[step2\]) as an eigenvalue problem for the matrix $M_{ij}= (a\omega)^2 c_{ji}^{m}$ with eigenvector $b^{i}= b^{a\omega}_{i}$ and eigenvalue $E_{lm}$. Clearly, ${\bf M}$ is a real band-diagonal matrix. Standard routines from [@Recipes] can be employed to find the eigenvectors and eigenvalues of such a matrix. Then, the spheroidal harmonic is directly obtained from (\[sp\_dec\]) (even though it involves an infinite sum, in reality only few coefficients $b_{j}^{a\omega}$ are significant). The described spectral decomposition method is reliable, unless $a\omega$ becomes large compared to unity (under such conditions the matrix ${\bf M}$ is no longer diagonally dominant, and the convergence of the method is very slow). In effect, very high frequency cross sections for Kerr scattering will be inaccurate (especially when the black hole is rapidly rotating). J.A.H. Futterman, F.A. Handler and R.A. Matzner, *Scattering from Black Holes* (Cambridge University Press, Cambridge, England, 1988). R.A. Matzner and M.P. Ryan, Jr., Astrophys. J. Suppl. [**36**]{}, 451 (1978). N. Sanchez, J. Math. Phys. [**17**]{}, 688 (1976) N. Sanchez, Phys. Rev. D [**16**]{}, 937 (1977) N. Sanchez, Phys. Rev. D [**18**]{}, 1030 (1978) N. Sanchez, Phys. Rev. D [**18**]{}, 1798 (1978) P. Anninos, C. DeWitt-Morette, R.A. Matzner, P. Yioutas, and T-R Zhang, Phys. Rev. D [**46**]{}, 4477 (1992) N. Andersson, Phys. Rev. D [**52**]{}, 1808 (1995) F.A. Handler and R.A. Matzner, Phys. Rev. D [**22**]{}, 2331 (1980) R.G. Newton, *Scattering Theory of Waves and Particles* (McGraw-Hill, New York, 1966). N. Andersson and K-E. Thylwe, Class. Quantum Grav. [**11**]{}, 2991 (1994) N.Andersson, Class. Quantum Grav. [**11**]{}, 3003 (1994) C. DeWitt-Morette and B.L. Nelson, Phys. Rev. D [**29**]{}, 1663 (1984) T-R Zhang and C. DeWitt-Morette, Phys. Rev. Lett. [**52**]{}, 2313 (1984) N. Fröman, P.O. Fröman, and B. Lundborg, Math. Proc. Cambridge Philos. Soc. [**104**]{}, 153 (1988) N. Fröman, P.O. Fröman, in *Forty More Years of Ramifications: Spectral Asymptotics and its Applications*, edited by S.A. Fulling and F.J.Narcowich, Discourses in Mathematics and its Applications, No. 1(Texas A& M University, Department of Mathematics, 1992), pp 121-159 P. Pajunen, J. Chem. Phys. [**88**]{}, 4268, (1988) P. Pajunen, J. Comp. Phys. [**82**]{}, 16, (1989) J.D. Pryce, *Numerical Solution of Sturm-Liouville Problems* (Oxford Science Publications, Oxford, England, 1993) N. Andersson, Proc. R. Soc. Lon. A [**439**]{}, 47, (1992) S.A. Teukolsky, Astrophys. J. [**185**]{}, 635 (1973) S.A. Teukolsky and W.H. Press, Astrophys. J. [**193**]{}, 443 (1974) R.A. Matzner, J.Math.Phys. [**9**]{}, 163 (1968) P.L. Chrzanowski, R.A. Matzner, V.D. Sandberg, and M.P. Ryan, Jr. , Phys. Rev. D [**14**]{}, 317 (1976) K.W. Ford and J.A. Wheeler, Ann. Phys. (N.Y.) [**7**]{}, 287 (1959) S. Chandrasekhar, *The Mathematical Theory of Black Holes* (Oxford University Press, New York, 1983) W.H. Press, S.A. Teukolsky, W.T. Vetterling and B.P. Flannery, *Numerical Recipes* (Cambridge University Press, Cambridge, England, 1992). M. Abramowitz and I.A. Stegun *Handbook of Mathematical Functions* (Dover Publications INC., New York, 1985) I.S. Gradshteyn and I.M. Ryzhik, *Tables of Integrals, Series and products* (Academic Press INC., London 1980) S.A. Hughes, Phys. Rev. D [**61**]{}, 084004 (2000)

ArXiv