Datasets:

timaeus

/

pile-arxiv

like 0

Follow

Timaeus 9

--- author: - 'F. Hinterberger' - 'A. Sibirtsev' date: 'Received: date / Revised version: date' title: | Analysis of the $\Lambda p$ Final State Interaction\ in the Reaction $p+p\to K^+(\Lambda p)$. --- [leer.eps]{} gsave 72 31 moveto 72 342 lineto 601 342 lineto 601 31 lineto 72 31 lineto showpage grestore Introduction. ============= Experimental information on the $\Lambda{p}$ interaction has been derived from the analysis of hypernuclei, $\Lambda{p}$ scattering experiments and studies of the $\Lambda{p}$ final state interaction (FSI) in strangeness transfer reactions. The binding energy of light hypernuclei shows that the low-energy $\Lambda{p}$ interaction is attractive. In addition, the $\Lambda{p}$ interaction is spin-dependent and the singlet interaction is stronger than the triplet one [@fhint:dal65; @fhint:dal81; @fhint:dov84]. The free $\Lambda{p}$ scattering was studied in bubble chamber measurements [@fhint:ale68; @fhint:sec68; @fhint:ale69]. In the low momentum region the elastic cross sections were analyzed in terms of the $S$-wave singlet and triplet scattering lengths and effective ranges. However these determinations are characterized by large variances and covariances since the data only support the determination of a spin-averaged scattering length and effective range [@fhint:ale69]. The strong effect due to the $\Lambda{p}$ FSI was observed in strangeness transfer reactions [@fhint:tai69; @fhint:pig77; @fhint:pie88] and and in associated strangeness production reactions [@fhint:mel65; @fhint:ree68; @fhint:hog68; @fhint:sie94; @fhint:bal98; @fhint:bil98; @fhint:mag01]. The $\Lambda{p}$ production in the $K^-d{\to}\pi^-\Lambda{p}$, $\pi^+d{\to}K^+\Lambda{p}$, $\gamma{d}{\to}K^0\Lambda{p}$ and $pp{\to}K^+\Lambda{p}$ reactions can provide substantial improvement in an evaluation of the low energy $\Lambda{p}$ scattering. Intensive theoretical studies of the hyperon-nucleon interaction with the Nijmegen [@fhint:nag77; @fhint:nag79; @fhint:mae89; @fhint:sto99] and Jülich [@fhint:hol89; @fhint:reu94; @Haidenbauer] potential models predict the hyperon-nucleon potentials and phase shift parameters. These models also predict the singlet and triplet scattering length and effective range parameters of the free $S$-wave $\Lambda{p}$ interaction. The present paper refers to the associated strangeness reaction $pp{\to}K^+(\Lambda{p})$ which is characterized by a strong FSI near the $\Lambda{p}$ production threshold. In most experiments the $\Lambda{p}$ system was measured inclusively. Exclusive measurements of the $pp{\to}K^+\Lambda{p}$ reaction were performed at COSY [@fhint:bal98; @fhint:bil98] and SATURNE [@fhint:mag01]. Theoretical analyses of the reactions were done [@fhint:del89; @fhint:lag91; @Sibirtsev1; @fhint:fae97; @fhint:sib98; @Sibirtsev2; @Tsushima1; @Sibirtsev3; @Shyam1; @Gasparian1; @fhint:kel00; @Shyam2; @Gasparian2; @Gasparian3] by applying the meson exchange model and including the $\Lambda{p}$ FSI generally modelled by Nijmegen or Jülich potentials. The aim of the present paper is to perform an analysis of the $\Lambda{p}$ FSI in the reaction $pp{\to}K^+(\Lambda{p})$ and an evaluation of the $\Lambda{p}$ low energy interaction parameters. Using the Watson-Migdal approximation [@wat52; @mig55; @gol64; @Gillespie] the reaction amplitude is factorized in terms of a production matrix element and a FSI enhancement factor, which can be represented by the inverse Jost function [@jos47; @jos52; @bar49a; @bar49b]. We analyze experimental results from SATURNE collected by Siebert et al. [@fhint:sie94], which are characterized by a high statistical accuracy and a high invariant mass resolution. Furthermore, in the fitting procedure the missing mass resolution is taken into account by folding the theoretical expressions with the experimental resolution function. The analysis shows that the shape of the sharply rising invariant mass spectrum depends strongly on the singlet and triplet scattering length and effective range parameters. But only two parameters, the spin-averaged scattering length and effective range parameters, can be deduced within an acceptable confidence level by fitting the $\Lambda{p}$ missing mass spectrum. Additional information can be obtained by taking the total cross section data for the free $\Lambda{p}$ scattering into account and including these data in an overall fit. At low energies the total cross section can be described in a model-independent way using the effective range approximation [@fhint:sch47; @fhint:bet49; @fhint:bla54]. Thus, by fitting simultaneously the $\Lambda{p}$ invariant mass spectrum and the available total cross section data of the free $\Lambda{p}$ scattering severe constraints on the singlet and triplet scattering length and effective range parameters can be deduced. This method allows also to test theoretical model predictions. The formalism. ============== Phase space distribution. ------------------------- The $pp{\to}K^+\Lambda{p}$ double differential cross section is given as $$\frac{d^2\sigma}{d\Omega_{K}dM_{\Lambda{p}}}= |{\tilde{\cal M}|^2} \, \Phi_3 ,$$ where ${\tilde {\cal M}}$ is the Lorentz-invariant reaction amplitude and the three-body phase space distribution function is $$\Phi_3{=} \frac{\pi}{16(2\pi)^5} \frac{p_{K}^2 q}{p_p m_p [(E_{p}+m_p)p_{K}-E_{K}p_{p}\cos\theta_{K}]}, \label{phase}$$ where $q$ is the momentum of $\Lambda$ in the Gottfried-Jackson rest-frame of the produced two-particle subsystem $X{=}\Lambda{+}p$, $M_{\Lambda{p}}$ is the corresponding invariant mass and $p_p$, $E_p$, $p_K$, $E_K$, $\theta_K$, $\Omega_K$ are defined in the laboratory system. Obviously, in inclusive measurements the invariant mass $M_{\Lambda{p}}$ is equal to the missing mass $M_X$ below the $\Sigma$-hyperon production threshold. Eq.(\[phase\]) is consistent with the kinematical definitions of Refs. [@fhint:byc73; @fhint:pdg98]. Final state interaction. ------------------------ In the Watson-Migdal approximation [@wat52; @mig55; @gol64] the FSI is taken into account by introducing a FSI enhancement factor $|C_{FSI}|^2$, $$\frac{d^2\sigma}{d\Omega_{K} dM_{\Lambda{p}}}= |{\cal M}|^2 \, |C_{FSI}|^2 \, \Phi_3, \label{watson}$$ where now ${\cal M}$ is pure production matrix element and the FSI amplitude $C_{FSI}$ depends on the internal momentum $q$ of the $\Lambda{p}$ subsystem. It converges to 1 for $q{\to}\infty$ where the $S$-wave FSI enhancement vanishes. Applying the factorization we assume that the production operator ${\cal M}$ is constant, i.e. does not depend on the internal kinetic energy of the $\Lambda{p}$ subsystem. In case of the $pp{\to}K^+\Lambda{p}$ reaction this assumption is supported by the kinematics which provides a focus onto the $\Lambda{p}$ FSI. The internal kinetic energy of the $\Lambda{p}$ subsystem is almost zero near the $\Lambda{p}$ threshold whereas the $K^+ \Lambda$- and $K^+{p}$-subsystems have large internal kinetic energies. Even if the $pp{\to}K^+\Lambda{p}$ reaction is dominated [@Sibirtsev2; @Tsushima1; @Sibirtsev3; @Shyam1] by intermediate baryonic resonances coupled to the $K^+\Lambda$ system a small variation of the invariant $\Lambda{p}$ mass does practically not affect the production amplitude. The methods for studying the FSI between the particles have been developed in different areas of physics, ranging from atomic physics to high energy particle physics [@Gillespie]. Taking the inverse Jost function [@jos47; @jos52] the correction due to the FSI is given as $$C_{FSI}=\frac{q-i\beta}{q+i\alpha},\; \; \; |C_{FSI}|^2=\frac{q^2+\beta^2}{q^2+\alpha^2}.$$ The potential parameters $\alpha$ and $\beta$ can be used to establish phase-equivalent Bargmann potentials [@bar49a; @bar49b]. They are related to the scattering lengths $a$, and effective ranges $r$ of the low-energy $S$-wave scattering $$\alpha=\frac{1}{r}\left(1-\sqrt{1-2\frac{r}{a}}\right),\; \; \; \beta =\frac{1}{r}\left(1+\sqrt{1-2\frac{r}{a}}\right).$$ The $\Lambda{p}$ system can couple to singlet $^1S_0$ and triplet $^3S_1$ states. Near production threshold the singlet-triplet transitions due to the final state interaction cannot occur. Therefore, the contributions of the spin-singlet and spin-triplet final states can be added incoherently. Taking the spin-statistical weights into account the unpolarized double differential cross section may be written as $$\begin{aligned} \frac{d^2\sigma}{d\Omega_{K}dM_{\Lambda{p}}}= \Phi_3 \left[\, 0.25 \, |{\cal M}_s|^2 \, \frac{q^2+\beta_s^2}{q^2+\alpha_s^2} \right. \nonumber \\ {+}\left.\, 0.75 \, |{\cal M}_t|^2 \, \frac{q^2+\beta_t^2}{q^2+\alpha_t^2}\right]. \label{miss}\end{aligned}$$ This equation leaves six free parameters, the singlet and triplet potential parameters $\alpha_s$, $\beta_s$, $\alpha_t$, $\beta_t$ and the production matrix elements $|{\cal M}_s|$ and $|{\cal M}_t|$. Instead of the parameters $\alpha_s$, $\beta_s$, $\alpha_t$ and $\beta_t$ one can equally well use the singlet and triplet scattering length and effective range parameters $a_s$, $r_s$, $a_t$ and $r_t$. The functional dependence on the invariant mass $M_{\Lambda{p}}$ can be evaluated by inserting the corresponding expression for the internal momentum $q$ of the $\Lambda{p}$ system, $$q=\frac{\sqrt{M_{\Lambda{p}}^2-(m_\Lambda+m_p)^2} \sqrt{M_{\Lambda{p}}^2-(m_\Lambda-m_p)^2}} {2M_{\Lambda{p}}}.$$ Missing mass resolution. ------------------------ The theoretical missing mass spectrum of Eq.(\[miss\]) has to be folded with the missing mass resolution function before comparing with the data. Introducing the missing mass resolution function $g({\tilde M}_{\Lambda{p}}{-}M_{\Lambda{p}})$ and denoting the r.h.s. of Eq.(\[miss\]) by $f_0(M_{\Lambda{p}})$ yields the distribution function $f(M_{\Lambda{p}})$ which has to be compared with the data, $$f(M_{\Lambda{p}})=\int \!\!f_0({\tilde M}_{\Lambda{p}}) \, g({\tilde M}_{\Lambda{p}}{-}M_{\Lambda{p}}) \, d{\tilde M}_{\Lambda{p}} \label{res}$$ where the resolution function is given as $$g({\tilde M}_{\Lambda{p}}{-}M_{\Lambda{p}}){=} \frac{1}{\sqrt {2\pi}}\frac{1}{\sigma_M} \exp-\frac{({\tilde M}_{\Lambda{p}}{-}M_{\Lambda{p}}{+} \Delta M_{\Lambda{p}})^2}{2\sigma_{M}^2}.$$ Here, $\sigma_{M}$ denotes the one standard deviation width and the shift $\Delta M_{\Lambda{p}}$ takes a systematic calibration error of the invariant mass scale into account. For the comparison with the SATURNE data the values $\Delta M_{\Lambda{p}}{=}1.7$ MeV and $\sigma_M{=}2.0$ MeV [@fhint:sie94] are used. The $\Lambda{p}$ cross section. ------------------------------- In the effective range approximation [@fhint:sch47; @fhint:bet49; @fhint:bla54] the total cross section of the free $\Lambda{p}$ elastic scattering can be expressed in terms of the singlet and triplet scattering length and effective range parameters $a_s$, $a_t$, $r_s$ and $r_t$, $$\sigma_{\Lambda{p}}{=} \frac{\pi}{q^2{+}\left(-\frac{1}{a_s}{+}\frac{r_s q^2}{2}\right)^2} {+} \frac{3\pi}{q^2{+}\left(-\frac{1}{a_t}{+}\frac{r_t q^2}{2}\right)^2}. \label{cross}$$ Here, $q$ is the cm-momentum of the $\Lambda{p}$ scattering. This equation can be applied at low energies where $S$-wave contributions dominate. The results of the fits. ======================== The program Minuit of the CERN program library [@fhint:cer94] was used in order to perform nonlinear least-square fits. We fit the experimental results  [@fhint:sie94] for missing mass spectrum from $pp{\to}K^+X$ reaction measured at proton beam energy $T_p{=}2.3$ GeV and kaon emission angle of $\theta_K{=}10.3^\circ$ using Eqs.(\[miss\]) and (\[res\]), respectively. We fit the strong FSI enhancement near the $\Lambda{p}$ threshold assuming that the $S$-wave FSI is dominant. Therefore, we take only data corresponding to very low c.m. momenta $q$ in the $\Lambda{p}$ system. The fits include 29 experimental points of the missing mass spectrum from the $\Lambda{p}$ threshold at 2054 MeV up to the invariant mass of the $\Lambda{p}$ system of 2095 MeV. This $M_{\Lambda{p}}$ range corresponds to c.m. momenta $q$ in the $\Lambda{p}$ system between 0 and 200 MeV/c. Low energy total $\Lambda{p}$ cross section data from bubble chamber measurements [@fhint:ale68; @fhint:sec68; @fhint:ale69] are fitted simultaneously using Eq.(\[cross\]). The fits include 12 total $\Lambda{p}$ cross section data points covering c.m. momentum range 60${\le}q{\le}$135 MeV/c. It is worthwhile to mention that the $pp{\to}K^+X$ reaction provides high quality data even at very small $q$ near the $\Lambda p$ threshold whereas $\Lambda{p}{\to}\Lambda{p}$ data are available only for $q{\geq}60$ MeV/c. Also the number of data and the accuracy of the $pp{\to}K^+X$ measurement is substantially higher. Therefore it is very attractive to include the $pp{\to}K^+X$ data in the evaluation of the low-energy phase-equivalent potential parameters as far as the evaluation remains model independent. The quality of the least-square fits is given by total $\chi^2$ and reduced $\chi^2/ndf$ where $ndf$ is the number of degrees of freedom given by the number of data minus the number of the fit parameter. In calculating the parameter errors nonlinearities and parameter correlations were taken into account. The error for a given parameter is defined as the change of that parameter which causes $\chi^2$ to increase by one while re-fitting all other free parameters. Due to the nonlinearities the resulting error intervals are in general asymmetric. Three-parameter fit. -------------------- In a three-parameter fit we determine spin-averaged parameters by applying the constraints $$|{\cal M}_s|^2{=}|{\cal M}_t|^2{=}|\bar{{\cal M}}|^2,\; \; a_s{=}a_t{=}\bar{a},\; \; r_s{=}r_t{=}\bar{r}.$$ The three-parameter fit yields the production matrix element squared $|\bar{\cal M}|^2$ and the spin-averaged scattering length $\bar{a}$ and effective range $\bar{r}$. In a first step we fit only the missing mass spectrum without taking the $\Lambda p$ total cross section data into account. The three-parameter fit yields an excellent description of the missing mass spectrum but fails completely to reproduce the total cross section data ($\chi^2/ndf{=}18.0$, see Fig. \[mm3cross0\]). The fit parameters of the missing mass spectrum are $$\begin{aligned} |\bar{\cal M}|^2&=&15.4_{-1.6}^{+1.5}\;{\rm b/sr},\; \chi^2/ndf=0.98 \nonumber \\ \bar{a}&=&-2.57_{-0.23}^{+0.20}\; {\rm fm},\; \bar{r}=2.47_{-0.24}^{+0.23}\; {\rm fm}. \label{fitmm3cross0}\end{aligned}$$ The dotted and dashed-dotted lines in Fig. \[mm3cross0\] and in the following Figs. \[mm1cross2\] - \[mm5cross2\] show the corresponding singlet and triplet contributions. The dashed lines show the phase space distributions without the FSI enhancement factor, i.e. $(0.25|{\cal M}_s|^2 + 0.75|{\cal M}_t|^2)\Phi_3$. Vice versa, we take only the total cross section data into account and determine $\bar{a}$ and $\bar{r}$ in a two-parameter fit. The resulting fit parameters are $$\bar{a}=-1.81_{-0.21}^{+0.18}\; {\rm fm},\; \bar{r}=3.24_{-0.48}^{+0.48}\; {\rm fm},\; \chi^2/ndf=0.39. \label{cross2}$$ Taking those parameters fixed and fitting only $|\bar{\cal M}|^2$ yields for the missing mass spectrum $$|\bar{\cal M}|^2=19.5_{-0.2}^{+0.2}\;{\rm b/sr},\; \chi^2/ndf=3.7. \label{fitmm1cross2}$$ This procedure yields an excellent fit of the total cross section data but fails completely to describe the missing mass spectrum (see Fig. \[mm1cross2\]). In a next step, we determine spin-averaged parameters in a combined fit, i.e. by fitting simultaneously the missing mass spectrum and the total cross section data. The resulting parameters are $$\begin{aligned} |\bar{\cal M}|^2&=&16.9_{-1.2}^{+1.2}\;{\rm b/sr},\; \chi^2/ndf=2.2, \nonumber \\ \bar{a}&=&-1.91_{-0.11}^{+0.10}\; {\rm fm},\; \bar{r}=2.74_{-0.20}^{+0.20}\; {\rm fm}. \label{fitmm3cross2}\end{aligned}$$ This procedure fails to describe both, the missing mass spectrum and the total cross section data (see Fig. \[mm3cross2\]). This failure is a direct indication that the spin dependence of the $\Lambda{p}$ interaction must be taken into account. Five-parameter fit. ------------------- Now the data on total $\Lambda{p}$ cross section and $pp{\to}K^+X$ missing mass spectrum are fitted in a combined fit with the singlet and triplet scattering lengths and effective ranges $a_s$, $r_s$, $a_t$, $r_t$ as separate free parameters. Taking the unknown quantities $|{\cal M}_s|^2$ and $|{\cal M}_t|^2$ into account a six-parameter fit should be performed. However, it turned out that the $\chi^2$-criterion cannot be used to determine simultaneously $|{\cal M}_s|^2$ and $|{\cal M}_t|^2$. This is due to the fact that the resulting $\chi^2$ depends only weakly on the ratio $|{\cal M}_t|^2{/}|{\cal M}_s|^2$ as is indicated in Table \[tab1\]. Therefore, five-parameter fits were performed taking $|{\cal M}_s|^2$ as free parameter and the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ as fixed parameter. By this method valuable constraints on the singlet and triplet scattering lengths and effective ranges can be deduced from the data. It should be mentioned that the spin-statistical weights 0.25 and 0.75 of the singlet and triplet contributions have already been taken into account in the theoretical ansatz (\[miss\]). Therefore the five-parameter search was started with the constraint $|{\cal M}_t|^2/|{\cal M}_s|^2{=}1$ yielding a solution with $\chi^2/ndf{=}0.82$. The resulting parameters are $$\begin{aligned} |{\cal M}_s|^2&=&15.0_{-1.6}^{+1.4}\;{\rm b/sr},\; \chi^2/ndf=0.82, \nonumber \\ a_s&=&-3.2_{-0.6}^{+0.4}\; {\rm fm},\; r_s=1.25_{-0.15}^{+0.13}\; {\rm fm}, \nonumber \\ a_t&=&-1.3_{-0.5}^{+0.4}\; {\rm fm},\; r_t=5.4_{-1.6}^{+1.6}\; {\rm fm}. \label{fitmm5cross2}\end{aligned}$$ Both, the missing mass spectrum and the total cross section data (see Fig.\[mm5cross2\]) are perfectly reproduced. Then, the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ was varied over a wide range of values and best fit solutions of similar quality with $\chi^2/ndf$ varying between 0.80 and 0.83 were found. The resulting fit parameters are listed in Table \[tab1\] for $0{\leq} |{\cal M}_t|^2/|{\cal M}_s|^2{\leq} 8$. A characteristic feature of the best fit solutions is the fact that the singlet FSI enhancement factor at $q{=}0$ is much larger than the triplet one, e.g. $\beta_s^2/\alpha_s^2{=}49.1$ and $\beta_t^2/\alpha_t^2{=}3.9$ for $|{\cal M}_t|^2/|{\cal M}_s|^2{=}1$. In spite of the statistical weight 0.25 the singlet contribution dominates the ${\rm p}{\rm p}{\to}{\rm K}^+ \Lambda {\rm p}$ cross section at $q{=}0$, e.g. $0.25|{\cal M}_s|^2\beta_s^2/\alpha_s^2{=}183$ b/sr and $0.75|{\cal M}_t|^2\beta_t^2/\alpha_t^2{=}43.5$ b/sr for $|{\cal M}_t|^2/|{\cal M}_s|^2{=}1$. Solutions where the triplet contribution is larger than the singlet contribution do not fit the FSI enhancement of the missing mass spectrum near $q=0$ and can be excluded on the basis of the $\chi^2$ criterion. This holds true even if one varies the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ in a wide range. The resulting best fit parameters are shown in Fig. \[correl3\] as a function of the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$. The variation of the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ causes rather small [cccccccc]{} $|{\cal M}_t|^2/|{\cal M}_s|^2$&$ |{\cal M}_s|^2$ (b/sr) & $a_s$ (fm) & $r_s$ (fm) & $a_t$ (fm) & $r_t$ (fm) & $\chi^2$ & $\chi^2/ndf$\ 0.00 & $61.4_{-6.3}^{+5.9}$ & $-2.6_{-0.2}^{+0.2}$ & $2.47_{-0.24}^{+0.23}$ & $-1.5_{-0.3}^{+0.2}$ & $3.6_{-0.9}^{+0.9}$ & 29.0 & 0.805\ 0.10 & $46.7_{-4.5}^{+4.2}$ & $-2.7_{-0.4}^{+0.3}$ & $2.25_{-0.24}^{+0.23}$ & $-1.5_{-0.3}^{+0.3}$ & $3.8_{-1.0}^{+1.0}$ & 29.0 & 0.807\ 0.25 & $34.4_{-3.2}^{+3.0}$ & $-2.9_{-0.5}^{+0.3}$ & $1.97_{-0.23}^{+0.22}$ & $-1.4_{-0.4}^{+0.3}$ & $4.0_{-1.2}^{+1.1}$ & 29.1 & 0.809\ 0.50 & $24.0_{-2.3}^{+2.1}$ & $-3.1_{-0.6}^{+0.4}$ & $1.63_{-0.19}^{+0.18}$ & $-1.3_{-0.4}^{+0.4}$ & $4.5_{-1.3}^{+1.3}$ & 29.2 & 0.812\ 1.00 & $15.0_{-1.6}^{+1.4}$ & $-3.2_{-0.6}^{+0.4} $ & $1.25_{-0.15}^{+0.13}$ & $-1.3_{-0.5}^{+0.4}$ & $5.4_{-1.}^{+1.6}$ & 29.4 & 0.816\ 2.00 & $8.7_{-1.0}^{+0.8}$ & $-3.3_{-0.6}^{+0.4}$ & $0.90_{-0.10}^{+0.09}$ & $-1.2_{-0.6}^{+0.5}$ & $6.5_{-2.1}^{+2.0}$ & 29.6 & 0.821\ 4.00 & $4.7_{-0.5}^{+0.4}$ & $-3.3_{-0.6}^{+0.4}$ & $0.63_{-0.07}^{+0.06}$ & $-1.2_{-0.8}^{+0.5}$ & $7.8_{-2.7}^{+2.7}$ & 29.8 & 0.827\ 8.00 & $2.4_{-0.2}^{+0.2}$ & $-3.4_{-0.6}^{+0.4}$ & $0.44_{-0.05}^{+0.04}$ & $-1.2_{-1.0}^{+0.6}$ & $9.2_{-3.5}^{+3.5}$ & 29.9 & 0.832\ variations of the parameters $a_s$, $a_t$ and rather large variations of the parameters $r_s$, $r_t$, respectively. Therefore one can deduce important constraints on the singlet and triplet scattering lengths and effective ranges of the low energy $\Lambda{p}$ interaction: $$\begin{aligned} -4.1\; {\rm fm}&<&a_s < -2.3\; {\rm fm},\; \; \; r_s < 2.7\; {\rm fm}, \nonumber \\ -1.8\; {\rm fm} &<&a_t < -0.6\; {\rm fm},\; \; \; r_t > 2.7\; {\rm fm}. \label{constraint}\end{aligned}$$ Tests of potential model results. ================================= Available meson-exchange potential model predictions of the S-wave $\Lambda{p}$ singlet and triplet scattering length and effective range parameters are listed in Table \[tab2\]. Here the results from Nijmegen model are denoted as Nijm D [@fhint:nag77], Nijm F [@fhint:nag79] and Nijm a-e [@fhint:sto99]. The NSC parameters are taken from Ref. [@fhint:mae89]. The parameters given by Jülich model are denoted as Jül A,B [@fhint:hol89] and Jül $\tilde{\rm A}$,$\tilde{\rm B}$ [@fhint:reu94]. The most recent Jülich parameters are indicated as Jül 03 [@Haidenbauer]. In Fig. \[correl2\] the model predictions of $a_s$, $r_s$, $a_t$ and $r_t$ are compared with the five-parameter fit results of Sect. 3. The experimentally allowed regions deduced from five-parameter fits are indicated by hatched rectangles. All model predictions of ($a_s$, $r_s$) and most predictions of ($a_t$, $r_t$) lie outside the experimentally allowed regions. In most cases the singlet scattering lengths $a_s$ are too small and the triplet scattering lengths $a_t$ are too large. [lccccccccc]{} Model & $ a_s$ (fm) & $r_s$ (fm) & $a_t$ (fm) & $r_t$ (fm) & $|{\cal M}_s|^2$ (b/sr) & $|{\cal M}_t|^2$ (b/sr) & $\chi^2$ & $\chi^2/ndf$\ Nijm a & -0.71 & 5.86 & -2.18 & 2.76&$0.\pm 0.1$&$22.8\pm 0.2$ & 47.6 & 1.22\ Nijm b & -0.90 & 4.92 & -2.13 & 2.84&$0.\pm 0.1$ &$23.4\pm 0.2$ & 53.9 & 1.38\ Nijm c & -1.20 & 4.11 & -2.08 & 2.92 &$0.\pm 0.1$&$23.9\pm 0.2$ & 62.2 & 1.60\ Nijm d & -1.71 & 3.46 & -1.95 & 3.08 &$0.\pm 0.1$ &$25.0\pm 0.2$ & 83.3 & 2.14\ Nijm e & -2.10 & 3.19 & -1.86 & 3.19 &$77.5\pm 0.7$ &$0.\pm 0.1$& 76.6 & 1.97\ Nijm f & -2.51 & 3.03 & -1.75 & 3.32 &$74.7\pm 0.7$&$0.\pm 0.1$ & 44.8 & 1.15\ Jül $\tilde{\rm A}$ & -2.04 & 0.64 & -1.33 & 3.91 & $7.3\pm 0.5$& $7.8\pm 1.6$ & 55.9 & 1.43\ Jül $\tilde{\rm B}$ & -0.40 & 12.28 & -2.12 & 2.57 & $ 0. \pm 0.1$& $ 21.3\pm 0.2$ & 43.6 & 1.12\ Jül A & -1.56 & 1.43 & -1.59 &3.16 &$33.8\pm 0.3$ &$0.\pm 0.1$ & 80.2&2.06\ Jül B & -0.56 & 7.77 & -1.91 & 2.43 & $0.\pm 0.1$ & $20.1\pm 0.2$ & 55.5 &1.42\ NSC & -2.78 & 2.88 & -1.41 & 3.11 & $71.6\pm 0.7$ & $0.\pm 0.1$ & 32.8 &0.84\ Nijm D & -1.90 & 3.72 & -1.96 & 3.24 & $ 0.\pm 0.2$ & $ 26.1\pm 0.2$ & 91.6&2.35\ Nijm F & -2.29 & 3.17 & -1.88 & 3.36 & $77.4\pm 0.7$ & $0.\pm 0.1$ & 62.6&1.61\ Jül 03 & -1.02 & 4.49 & -1.89 & 2.57 & $0.\pm 0.1$ & $21.3\pm 0.2$ & 62.4&1.60\ A more direct test of the model predictions can be performed using Eqs.(\[miss\]) and (\[cross\]). Keeping the predicted parameters fixed and fitting only the two free parameters $|M_s|^2$ and $|M_t|^2$ the calculated cross sections are compared with the ${\rm p}{\rm p}{\to}{\rm K}^+{\rm X}$ missing mass spectrum as well as the total $\Lambda{p}$ cross section data and the $\chi^2$ are deduced (see Table \[tab2\]). Excepting the model Jül ${\tilde{\rm A}}$ this procedure yields either $|M_s|^2{=}0$ or $|M_t|^2{=}0$. This is due to the fact that the models predict either a dominating singlet or triplet FSI enhancement. The resulting values of $\chi^2$ should be compared with the best-fit value $\chi^2{=}29.0$ obtained in direct five-parameter fits (see Table \[tab1\]). Excepting the result of the NSC model all $\chi^2$ values are essentially larger than 29.0. They indicate that those model predictions fail to reproduce simultaneously the missing mass spectrum and the total cross section data. Summary and conclusions. ======================== We analyzed the high resolution $pp{\to}K^+X$ data of Siebert et al. [@fhint:sie94] with respect to the strong FSI near the $\Lambda{p}$ production threshold. The observed missing mass spectrum was described by factorizing the reaction amplitude in terms of a production amplitude and FSI amplitude which was parametrized in terms of the inverse Jost function. It was found that a three-parameter fit with spin-averaged scattering length and effective range parameters $\bar{a}$ and $\bar{r}$ can reproduce the missing mass spectrum but fails to describe simultaneously the total $\Lambda{p}$ cross section data. Vice versa deducing $\bar{a}$ and $\bar{r}$ from a fit to the total cross section data fails to describe the missing mass spectrum. Also a combined three-parameter fit of the missing mass spectrum and the total $\Lambda{p}$ cross section data fails to reproduce simultaneously both data sets with spin-averaged parameters $\bar{a}$ and $\bar{r}$. Therefore the singlet and triplet scattering lengths and effective ranges $a_s$, $r_s$, $a_t$, $r_t$ are fitted as separate free parameters. Taking $|{\cal M}_s|^2$ as free parameter and the ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ as fixed parameter both the missing mass spectrum and the total cross scetion data are perfectly reproduced in five-parameter fits. The ratio $|{\cal M}_t|^2/|{\cal M}_s|^2$ was varied over a wide range of values and important constraints on the parameters $a_s$, $r_s$, $a_t$, $r_t$ were deduced. These are indicated in Fig. \[correl2\] as experimentally allowed regions. A characteristic feature of the best-fit solutions is the fact that the $\Lambda{p}$ FSI in the reaction $pp{\to}K^+\Lambda{p}$ is dominated by the singlet contribution. Another important result follows from a comparison of the singlet and triplet scattering lengths. The fact that $-a_s > -a_t$ means that the $\Lambda{p}$ interaction is more attractive in the singlet state than in the triplet state. This result is in accordance with the expectation deduced from an analysis of the binding energies of light hypernuclei [@fhint:dal65]. Though the $\chi^2$ values are only weakly dependent on $|{\cal M}_t|^2/|{\cal M}_s|^2$ ratios with $|{\cal M}_t|^2/|{\cal M}_s|^2{>}8$ can be excluded on the basis of the $\chi^2$ criterion. In this context it is interesting to note that the fit yields for $|{\cal M}_t|^2/|{\cal M}_s|^2{>}1$ abnormally small singlet and large triplet effective ranges, $r_s{<}1$ fm and $r_t{>}6$ fm. Taking the definition of the effective range as interaction range such values are questionable. However, this problem can only be decided by experiments which allow to determine separately the absolute values of the singlet and triplet production matrix elements. Most previous experiments provide only informations on the spin-averaged parameters $\bar{a}$ and $\bar{r}$ and the deduced values are given without errors estimates. But since the concept of spin-averaged parameters fails to describe simultaneously the $pp{\to}K^+ X$ FSI enhancement and the $\Lambda {p}$ total cross sections we do not compare our spin-averaged parameters with previous determinations of $\bar{a}$ and $\bar{r}$. In this context we mention the general problem of defining spin-averaged parameters if singlet and triplet parameters are different. Inspecting the world data set there is one experiment which provides direct information on the triplet parameters $a_t$ and $r_t$. The data on the $\Lambda{p}$ production in the reaction $K^-d{\to}\pi^-\Lambda{p}$ [@fhint:tai69] show a marked FSI enhancement near the $\Lambda{p}$ threshold. Tai Ho Tan deduced $a_t$ and $r_t$ assuming that the $K^- d$ capture at rest occurs mainly from S-wave orbital admixtures. Spin and parity considerations imply that the final $\Lambda{p}$ state has spin one as the deuteron. Assuming that the FSI amplitude near threshold is dominated by the S-wave contribution, i.e. the $^3{\rm S}_1$-state, the triplet scattering length and effective range parameter $a_t$ and $r_t$ can directly be extracted from those data. The deduced values $a_t{=}-2.0\pm 0.5$ fm and $r_t{=}3.0\pm 1.0$ fm [@fhint:tai69] agree within one standard deviation with the experimentally allowed region of ($a_t, r_t$). They favor our five-parameter solutions for small ratios $|{\cal M}_t|^2/|{\cal M}_s|^2$. It should be mentioned that a reanalysis of the $K^-d{\to}\pi^-\Lambda{p}$ data is highly wanted in view of the importance and the high quality of those data. The meson theoretical model predictions of the scattering lengths and effective ranges were compared with the experimentally allowed regions of ($a_s$, $r_s$) and ($a_t$, $r_t$) deduced from five-parameter fits of the data. In most cases the predicted singlet scattering lengths are too small and the triplet scattering lengths are too large. Only one model prediction is in agreement with the constraints deduced from the data. This finding was confirmed by a more direct test of the model predictions. The test was performed by keeping the predicted parameters in the fit routine fixed and fitting only $|M_s|^2$ and $|M_t|^2$. With one exception the model predictions fail to reproduce simultaneously the missing mass spectrum near the $\Lambda{p}$ threshold and the total $\Lambda{p}$ cross section data. It should be emphasized that the shape of the FSI enhancement near the $\Lambda{p}$ threshold is the essential experimental information in order to evaluate the FSI parameters from the reaction $pp{\to}K^+ \Lambda{p}$. Since the most important part of the FSI enhancement is located in the sharply rising part of the missing mass spectrum a high missing mass resolution and sufficient statistical accuracy are needed. The effective missing mass resolution of the SATURNE experiment [@fhint:sie94] corresponded to a $1\sigma$ width of 2 MeV. Improving this value by an order of magnitude would allow to study the $\Lambda{p}$ FSI near $q{=}0$ with higher accuracy. But not only the improvement of precision but also a systematic check of the method are important. An essential assumption of the factorization ansatz (Eq. \[watson\]) is that the production matrix element is constant, i.e. does not depend on the internal energy of the $\Lambda{p}$ final state. A direct check is to measure the reaction $pp{\to}K^+ \Lambda{p}$ at different energies and angles. Ultimately, measurements with polarized beams and polarized targets are needed in order to disentangle the spin singlet and triplet contributions in the $pp{\to}K^+\Lambda{p}$ reaction. Such measurements would allow to determine separately the singlet and triplet scattering lengths and effective ranges. The feasibility depends not only on the polarization of beam and target but also on the effective missing mass resolution and luminosity. We appreciate discussions with J. Haidenbauer, C. Hanhart and H. Rohdjeß. During the preparation of the paper we were informed about a theoretical study [@Gasparian3] of the $pp{\to}K^+\Lambda{p}$ reaction using dispersion relations. [byc73]{} R.H. Dalitz, [*Nuclear Interactions of the Hyperons*]{} (Oxford University Press, Oxford, 1965). R.H. Dalitz, Nucl. Phys. A [**354**]{}, 101c (1981). C.B. Dover and A. Gal, Prog. Part. Nucl. Phys. [**12**]{}, 171 (1984). G. Alexander et al., Phys. Rev. [**173**]{}, 1452 (1968). B. Sechi-Zorn et al., Phys. Rev. [**175**]{}, 1735 (1968). G. Alexander, [*Proc. Int. Conf. on Hypernuclei*]{} (Argonne National Laboratory, 1969). Tai Ho Tan, Phys. Rev. Lett. [**23**]{}, 395 (1969). C. Pigot et al., Nucl. Phys. B [**249**]{}, 172 (1985). H. Piekarz, Nucl. Phys. A [**479**]{}, 263c (1988). A.C. Melissinos et al., Phys. Rev. Lett. [**14**]{}, 604 (1965). J.T. Reed et al., Phys. Rev. [**168**]{}, 1495 (1968). W.J. Hogan, P.A. Piroue and A.J.S. Smith, Phys. Rev. [**166**]{}, 1472 (1968). R. Siebert et al., Nucl. Phys. A [**567**]{}, 819 (1994). J.T. Balewski et al., Eur. Phys. J. A [**2**]{}, 99 (1998) 99. B. Bilger et al., Phys. Lett. B [**420**]{}, 217 (1998). M. Maggiora et al., Nucl. Phys. A [**691**]{}, 329c (2001). R. Karplus and L. Rodberg, Phys. Rev. [**115**]{}, 1058 (1959). T. Kotani and M. Ross, Nuovo Cimento [**14**]{}, 1282 (1959). M.M. Nagels, T. A. Rijken and J.J. de Swart, Phys. Rev. D [**15**]{}, 2547 (1977). M.M. Nagels,T. A. Rijken and J.J. de Swart, Phys. Rev. D [**20**]{}, 1633 (1979). P.M.M. Maessen, T. A. Rijken and J.J. de Swart, Phys. Rev. C [**40**]{}, 226 (1989). V.C.J. Stoks and Th.A. Riken, Phys. Rev. C [**59**]{}, 3009 (1999). B. Holzenkamp, K. Holinde and J. Speth, Nucl. Phys. A [**500**]{}, 485 (1989). A. Reuber, K. Holinde and J. Speth, Nucl. Phys. A [**570**]{}, 543 (1994). J. Haidenbauer, W. Melnitchouk and J. Speth, nucl-th 0108062 (2001) and AIP Conf. Proc. [**603**]{}, 421 (2001). A. Deloff, Nucl. Phys. A [**505**]{}, 583 (1989). J.M. Laget, Phys. Lett. B [**259**]{}, 24 (1991). A. Sibirtsev, Phys. Lett. B [**359**]{}, 29 (1995). G. Fäldt and C. Wilkin, Z. Phys. A [**357**]{}, 241 (1997). A. Sibirtsev and W. Cassing, nucl-th/9802025 (1998). A. Sibirtsev, K. Tsushima and A.W. Thomas, Phys. Lett. B [**421**]{}, 59 (1998) K. Tsushima, A. Sibirtsev and A.W. Thomas, Phys. Rev. C [**59**]{}, 369 (1999). A. Sibirtsev, K. Tsushima, W. Cassing, A.W. Thomas, Nucl. Phys. A [**646**]{}, 427 (1999) R. Shyam, Phys. Rev. C [**60**]{}, 055213 (1999). A. Gasparian, J. Haidenbauer, C. Hanhart, L. Kondratyuk and J. Speth, Phys .Lett. B [**480**]{}, 273 (2000) N.G. Kelkar and B.K. Jain, Int J. Mod. Phys. E [**9**]{}, 431 (2000). R. Shyam, G. Penner and U. Mosel, Phys. Rev. C [**63**]{}, 022202 (2001). A. Gasparian, J. Haidenbauer, C. Hanhart, L. Kondratyuk and J. Speth, Nucl. Phys. A [**684**]{}, 397 (2001). A. Gasparyan, J. Haidenbauer, C. Hanhart and J. Speth, hep-ph/0311116 (2003). K.M. Watson, Phys. Rev. [**88**]{}, 1163 (1952). A.B. Migdal, Sov. Phys. JETP [**1**]{}, 2 (1955). M.L. Goldberger and K.M. Watson, [*Collision Theory*]{} (J. Wiley, New York, 1964) p. 549. J. Gillespie, [*Final-State Interactions*]{}, ed. K.M. Watson, (Holden-Day Adv. Phys. Monographs, Holden-Day, San Francisco, London, Amsterdam, 1964) R. Jost, Helv. Phys. Acta [**20**]{}, 356 (1947). R. Jost and W. Kohn, Phys. Rev. [**87**]{}, 977 (1952). V. Bargmann, Phys. Rev. [**75**]{}, 301 (1949). V. Bargmann, Rev. Mod. Phys. [**21**]{}, 488 (1949). J.S. Schwinger, Phys. Rev. [**72**]{}, 742 (1947). H.A. Bethe, Phys. Rev. [**76**]{}, 38 (1949). J.M. Blatt and V. Weisskopf, [*Theoretical Nuclear Physics*]{}, (J. Wiley, New York, London, 1954). E. Byckling and K. Kajantie, [*Particle Kinematics*]{}, (J. Wiley, New York, London, 1973). Particle Data Group, Eur. Phys. Jour. C [**3**]{}, 1 (1998). F. James, M. Roos, Comput. Phys. Commun. [**10**]{}, 343 (1975).

--- abstract: 'Measurement of an observable on a quantum system involves a probabilistic collapse of the quantum state and a corresponding measurement outcome. Lüders and von Neumann state update rules attempt to describe the above phenomenological observations. These rules are identical for a nondegenerate observable, but differ for a degenerate observable. While Lüders rule preserves superpositions within a degenerate subspace under a measurement of the corresponding degenerate observable, the von Neumann rule does not. Recently Hegerfeldt and Mayato \[Phys. Rev. A, 85, 032116 (2012)\] had formulated a protocol to discriminate between the two types of measuring devices. Here we have reformulated this protocol for quantum registers comprising of system and ancilla qubits. We then experimentally investigated this protocol using nulear spin systems with the help of NMR techniques, and found that Lüders rule is favoured.' author: - 'C. S. Sudheer Kumar, Abhishek Shukla, and T. S. Mahesh' bibliography: - 'bib\_ch.bib' title: | Discriminating between Lüders and von Neumann measuring devices:\ An NMR investigation --- Introduction ============ Quantum measurement paradox lies at the heart of foundations of quantum mechanics[@D_Home_book]. It’s an experimental fact that, upon measurement, a quantum state collapses into an eigenstate of the observable being measured. However there is no collapse in the unitary evolution described by Schrödinger equation, and therefore, the collapse has to be imposed from outside the formalism. Let us assume an observable $A_N$ with discrete and nondegenerate eigenspectrum. In that case, the measurement leads to a collapse of the state to one of the eigenstates of $A_N$ (see Fig. \[luder\]). On the other hand, if we consider an observable $A$ with a degenerate eigenspectrum, there are two extreme rules to update the state after the measurement. The most commonly used rule was postulated by Gerhart Lüders in 1951 [@luders1951zustandsanderung; @Luders_originl_paper]. According to it, a system existing in a superposition of degenerate eigenstates is unaffected by the measurement such that the superposition is preserved. However, an earlier postulate by von Neumann, proposed in 1932 [@von_math_foundof_QM], does not preserve such a superposition. In the latter postulate, the measuring device refines the observable $A$ into another commuting observable $A'$ (actual system observable) having a nondegenerate spectrum. The resulting measurement collapses the state to an eigenstate of $A'$, and the original superposition is not preserved under the measurement as if the degeneracy has been lifted [@Discriminate_von_Luder_protocol]. Although, one generally assumes Lüders state update rule implicitly in quantum physics, occassionally one encounters applications of the von Neumann state update rule. One example is in the context of Leggett-Garg inequality in multilevel quantum systems [@multi_level_LGI]. In principle, measurements which are intermediate between Lüders and von Neumann can also be conceived [@Discriminate_von_Luder_protocol; @multi_level_LGI]. Recently, Hegerfeldt and Mayato have proposed a general protocol (HM protocol) to discriminate between Lüders and von Neumann kind of measuring devices [@Discriminate_von_Luder_protocol]. To explain this protocol we consider an observable $A$, having two-fold degenerate eigenvalues, say $+1$ and $-1$ (see Fig. \[HMprotocolfig\]). The HM protocol involves the following steps: (i) prepare an eigenstate ${\vert{\xi_\mathrm{in}}\rangle}$ of $A$, (ii) let the device measure $A$, and (iii) characterize the output state. In step (ii) a Lüders measurement will preserve the state, while a von Neumann measurement may not. The last step is simply to determine if the step (ii) has changed the state or not. If the state has changed, we conclude that the device is von Neumann. Else, either the device is of Lüders type, or the chosen initial state ${\vert{\xi_\mathrm{in}}\rangle}$ happens to be a nondegenerate eigenstate of the actual system observable $A'$. To rule out the latter possibility, one may change the initial state and repeat the above steps (Fig. \[HMprotocolfig\]). This way one can attempt to discriminate between the Lüders and von Neumann measurement devices. In this work, we reformulate the HM protocol for a quantum register and try to investigate it using experiments. Nuclear spin ensembles in liquid, liquid-crystalline, or solid-state systems have often been chosen as convenient testbeds for studying foundations of quantum physics [@suter1988study; @Moussa; @LGI_Soumya; @Context_cssk]. Their main advantages are long coherence times and excellent control over quantum dynamics via highly developed nuclear magnetic resonance (NMR) techniques. In section II, we briefly explain the HM protocol as adapted to an NMR setup. The experimental details to discriminate between the Lüders and von Neumann measuring devices is described in section III. Finally we conclude in section IV. Theory ====== For the sake of clarity, and also to match the experimental details described in the next section, we consider a system of two qubits. Since the system is to be measured projectively, dimension of the pointer basis should be greater than or equal to that of the system, and hence we need at least two ancillary qubits. We refer to the ancillary qubits as (1,2) and system qubits as (3,4). We use Zeeman product basis as our computational basis and denote eigenkets of $\sigma_z$, the Pauli $z$-operator, by ${\vert{0}\rangle}$ and ${\vert{1}\rangle}$. We denote the basis vectors of system qubits as $$\begin{aligned} {\vert{\phi_{0}}\rangle} = {\vert{00}\rangle}, {\vert{\phi_{1}}\rangle} = {\vert{01}\rangle}, {\vert{\phi_{2}}\rangle} = {\vert{10}\rangle}, {\vert{\phi_{3}}\rangle} = {\vert{11}\rangle}. \label{phis}\end{aligned}$$ Let us assume a two-fold degenerate system-observable with spectral decomposition $$\begin{aligned} A = (\Pi_0+\Pi_1)-(\Pi_2+\Pi_3), ~~~~ \label{A}\end{aligned}$$ where the projectors are defined as $\Pi_j={\vert{\chi_j}\rangle\langle{\chi_j}\vert},{\vert{\chi_0}\rangle}=\alpha_0{\vert{\phi_0}\rangle}+\beta_0{\vert{\phi_1}\rangle}$, ${\vert{\chi_1}\rangle}=\alpha_1{\vert{\phi_0}\rangle}+\beta_1{\vert{\phi_1}\rangle}$, ${\vert{\chi_2}\rangle}=\alpha_2{\vert{\phi_2}\rangle}+\beta_2{\vert{\phi_3}\rangle}$, ${\vert{\chi_3}\rangle}=\alpha_3{\vert{\phi_2}\rangle}+\beta_3{\vert{\phi_3}\rangle}$ are eigenvectors of $A$. The projectors have the property $\Pi_k\Pi_l=\delta_{kl}\Pi_k$. We note that $A$ has no unique spectral decomposition due to the degeneracy. We consider a measurement model, wherein a quantum system being measured undergoes a joint evolution with the measuring device, ultimately forming an entangled state. When the measuring device collapses to a particular pointer state, the system also collapses to the corresponding eigenstate. Let $Q$ be the observable corresponding to the ancilla (measuring device) and $g$ be the system-ancilla interaction strength. The joint evolution is then of the form $$\begin{aligned} U_\mathrm{int} = \exp(-i{\cal H}_\mathrm{int}\tau),\end{aligned}$$ where ${\cal H}_\mathrm{int} = g ~Q \otimes A$ is the interaction Hamiltonian in units of angular frequency. To fix the basis inside a degenerate subspace, we should choose a nondegenerate observable $A'$ which commutes with $A$, so that they are simultaneously diagonalizable and hence we can find a common eigenbasis. For simplicity we choose the computional basis $\{{\vert{\phi_j}\rangle}\}$ as the common eigenbasis. Then the observable $A'$ must have the following spectral decomposition $$\begin{aligned} A'= \sum_{j=0}^3 a'_j P_j, \label{A'_defined}\end{aligned}$$ where $P_j = {\vert{\phi_j}\rangle\langle{\phi_j}\vert}$ and the nondegenerate eigenvalues $a'_j$ are yet to be determined. Let us assume the device to be von Neumann which refines the degenerate observable $A$ that is being measured, into a nondegenerate observable $A'$, via a mapping $f(A')=A$. As the refined observable $A'$ has nondegenerate eigenvalues and commutes with $A$, it fixes the basis inside the degenerate subspace. However, the choice of $A'$ is not unique, i.e., any orthonormal basis inside the degenerate subspace can be nondegenerate eigenkets of $A'$, and the von Neumann device has the freedom to choose among them [@von_math_foundof_QM]. The measurement outcome is passed via the refining function $f$, such that $f(a'_0)=f(a'_1)=+1$ and $f(a'_2)=f(a'_3)=-1$. Hence the outcome is same as if $A$ is being measured. To projectively measure the observable $A'$, the measuring device has to jointly evolve with the system under the interaction Hamiltonian, $$\begin{aligned} {\cal H}'_\mathrm{int}= g ~Q\otimes A'.\end{aligned}$$ For instance, we choose $Q=q_1\sigma_{1z}+q_2\sigma_{2z}$, where $\sigma_{1z} = \sigma_z \otimes\mathbbm{1}_2$, $\sigma_{2z} = \mathbbm{1}_2\otimes\sigma_z$ and $\mathbbm{1}_2$ is $2\times 2$ identity operator. The joint evolution between the measuring device (ancillary qubits) and the system is described by the unitary operator $$\begin{aligned} U'_\mathrm{int} = \exp(-i\cal{H}_\mathrm{int}' \tau),\end{aligned}$$ where $\tau$ is duration of the evolution. If each of the quantum register is initially prepared in ${\vert{\Phi_0}\rangle} = {\vert{++++}\rangle}$, with ${\vert{+}\rangle} = ({\vert{0}\rangle}+{\vert{1}\rangle})/\sqrt{2}$, the state after the joint evolution is given by $$\begin{aligned} U'_\mathrm{int}{\vert{\Phi_0}\rangle} &=& \frac{1}{2} \bigg( e^{-iga'_0Q \tau}{\vert{++}\rangle}{\vert{\phi_0}\rangle}+ e^{-iga'_1Q \tau}{\vert{++}\rangle}{\vert{\phi_1}\rangle}+ \nonumber \\ && e^{-iga'_2Q \tau}{\vert{++}\rangle}{\vert{\phi_2}\rangle}+ e^{-iga'_3Q \tau}{\vert{++}\rangle}{\vert{\phi_3}\rangle} \bigg) \nonumber \\ &=& \frac{1}{2} \sum_{j=0}^3 {\vert{\psi_j}\rangle}{\vert{\phi_j}\rangle}, \label{von_evolutn}\end{aligned}$$ where ${\vert{\phi_j}\rangle}$ are as defined in Eqs. \[phis\] and ${\vert{\psi_j}\rangle} = \exp(-iga'_jQ \tau){\vert{++}\rangle}$ represent states of the ancillary qubits. To realize the projective measurement, the pointer basis $\{{\vert{\psi_j}\rangle}\}$ must be orthonormal. Imposing the mutual orthogonality condition results in trigonometric constraint equations leading to a set of possible solutions. One such possible solution is $$\begin{aligned} \begin{array}{l|l} a'_0 = -a'_2 = -3 & q_1 = \pi/(4g\tau) \\ a'_1 = -a'_3 = 1 & q_2 = -q_1/2.\\ \end{array}\end{aligned}$$ Again, the von Neumann measuring device has the freedom to choose a particular pointer basis among several possible ones. Substituting the above values in Eq. \[A’\_defined\], we obtain, $$\begin{aligned} A'=-3 P_0+ P_1 +3 P_2 -P_3, \end{aligned}$$ which is obviously nondegenerate in the computational basis. The refining function $f$ can now be setup by interpolating the eigenvalue distribution (see Fig. \[fofx\]). For the above example, we find a possible map to be $f(A') = (-A'^3+7A')/6 = A$. The quantum circuit for discriminating Lüders and von Neumann devices, illustrated in Fig. \[fig\_discrimnt\_von\_lud\], involves four qubits each of which is initialized in state ${\vert{+}\rangle}$. If the device is Lüders, the system undergoes a joint evolution $U_\mathrm{int}$ with the ancilla resulting in the state $$\begin{aligned} U_\mathrm{int} {\vert{\Phi_0}\rangle} &=& \frac{1}{\sqrt{2}} \bigg( e^{-igQ \tau}{\vert{++}\rangle} \frac{d_0{\vert{\chi_0}\rangle}+d_1{\vert{\chi_1}\rangle}}{\sqrt{2}} + \nonumber \\ && e^{igQ \tau}{\vert{++}\rangle} \frac{d_2{\vert{\chi_2}\rangle}+d_3{\vert{\chi_3}\rangle}}{\sqrt{2}} \bigg) \nonumber \\ & = & \frac{1}{\sqrt{2}} \bigg( {\vert{\psi_1}\rangle} \frac{{\vert{\phi_0}\rangle}+{\vert{\phi_1}\rangle}}{\sqrt{2}} + {\vert{\psi_3}\rangle} \frac{{\vert{\phi_2}\rangle}+{\vert{\phi_3}\rangle}}{\sqrt{2}} \bigg),~~~~ \label{Luder_evolution}\end{aligned}$$ where the coefficients $d_j$ depend on the choice of ${\vert{\chi_j}\rangle}$ (defined after Eq. \[A\]) [@ludernote1]. Note that if the measuring device is of von Neumann type, it will instead measure $A'$, and pass the measurement outcome via the function $f$, as explained before. After the joint evolution of system and ancilla, a selective measurement of ancilla qubits is carried out. Generally in a quantum measurement the measuring device collapses to its pointer basis. In our scheme, we perform the projective measurement in the computational basis after transforming the ancilla qubits onto the computational basis using a similarity transformation $U_{a}^\dagger$, such that $$\begin{aligned} U_{a}{\vert{00}\rangle}={\vert{\psi_0}\rangle},U_{a}{\vert{01}\rangle}={\vert{\psi_1}\rangle},\nonumber\\ U_{a}{\vert{10}\rangle}={\vert{\psi_2}\rangle},U_{a}{\vert{11}\rangle}={\vert{\psi_3}\rangle}. \label{U_ra_constraints}\end{aligned}$$ By substituting the explicit forms of ${\vert{\psi_j}\rangle}$, we obtain $$\begin{aligned} U_{a}=\frac{1}{2} \left[ \begin{array}{cccc} z^3 & z^{-1} & z^{-3} & z \\ z^{9} & z^{-3} & z^{-9} & z^{3} \\ z^{-9} & z^{3} & z^{9} & z^{-3} \\ z^{-3} & z & z^{3} & z^{-1} \end{array} \right], \label{U_ra}\end{aligned}$$ where $z=\exp(i\pi/8)$. Finally, the ancilla is traced-out and the state of system qubits is characterized with the help of quantum state tomography. According to the Lüders state update rule, if a degenerate observable $A$ (as in Eq. \[A\]) is measured on a system in state $\rho_0$, then the postmeasurement state of the ensemble is described by $$\begin{aligned} \rho_{L}=\sum_{l=\pm 1}{\mathbb P}_l\rho_{0}{\mathbb P}_l, \label{rhoLgen}\end{aligned}$$ where ${\mathbb P}_{+1} = \Pi_{0}+\Pi_{1}$, ${\mathbb P}_{+1} = \Pi_{2}+\Pi_{3}$. For the initial state $\rho_0 = {\vert{\Phi_0}\rangle\langle{\Phi_0}\vert}$, we obtain $$\begin{aligned} \rho_L = (\mathbbm{1}_4 + \mathbbm{1}_2 \otimes \sigma_x)/4. \label{rhoL}\end{aligned}$$ However according to von Neumann’s degeneracy breaking state update rule, the postmeasurement state of the ensemble is given by $$\begin{aligned} \rho_N = \sum_{j=0}^3 \Pi_j \rho_0 \Pi_j, \label{rhoNgen}\end{aligned}$$ where, $\Pi_j$’s are fixed by the refining observable $A'$. Therefore, for the initial state $\rho_0 = {\vert{\Phi_0}\rangle\langle{\Phi_0}\vert}$ and the observable $A'$ (Eq. \[A’\_defined\]), the postmeasurement state collapses to a maximally mixed state, i.e., $$\begin{aligned} \rho_N = \mathbbm{1}_4/4. \label{rhoN}\end{aligned}$$ In both the cases, the probabilities of obtaining the eigenvalues $\pm 1$ are identical, i.e., $$\begin{aligned} p_{+1} &=& \mathrm{Tr}({\mathbb P}_{+1}\rho_{0}{\mathbb P}_{+1})=\sum_{j=0,1}\mathrm{Tr}(\Pi_j \rho_0 \Pi_j) ~~\mathrm{and,}\nonumber \\ p_{-1} &=& \mathrm{Tr}({\mathbb P}_{-1}\rho_{0}{\mathbb P}_{-1})=\sum_{j=2,3}\mathrm{Tr}(\Pi_j \rho_0\Pi_j).\end{aligned}$$ Thus although, the measurement outcomes (eigenvalues) and their probabilities are identical, the postmeasurement states $\rho_L$ and $\rho_N$ are different [@Luders_originl_paper; @von_math_foundof_QM; @Discriminate_von_Luder_protocol; @multi_level_LGI]. In fact, the Uhlmann fidelity between $\rho_L$ and $\rho_N$ turns out to be $F(\rho_L,\rho_N)=\mbox{Tr}\sqrt{\sqrt{\rho_N}\rho_L\sqrt{\rho_N}}=1/\sqrt{2}$ [@quant_info_neilson_chuang]. Therefore, it is possible to discriminate between the Luders and von Neumann devices by simply characterizing the final state of the system as shown by the circuit in Fig. \[fig\_discrimnt\_von\_lud\]. Experiment ========== We utilize the four spin-1/2 nuclei of 1,2-dibromo-3,5-difluorobenzene (DBDF) as our quantum register. About 12 mg of DBDF was partially oriented in 600 $\upmu$l of liquid crystal MBBA. The molecular structure of DBDF and its NMR Hamiltonian parameters are shown in Fig. \[mol\_H\]. The experiments were performed at 300 K on a 500 MHz Bruker UltraShield NMR spectrometer. The secular part of the spin-Hamiltonian is of the form [@cavanagh], $$\begin{aligned} {\cal H}_0 &=& -\sum_{j=1}^4 \omega_j I_{jz} + 2\pi\sum_{j,k>j} (J_{jk}+2D_{jk}) I_{jz}I_{kz} \nonumber \\ &&+ 2\pi (J_{12}-D_{12})( I_{1x}I_{2x}+ I_{1y}I_{2y}),\end{aligned}$$ where $\omega_j$, $J_{ij}$, and $D_{ij}$ are the resonance off-sets, indirect scalar coupling constants, and direct dipole-dipole coupling constants (Fig. \[mol\_H\]). The strong-coupling term (i.e., the last term) is relevant only for (H$_1$, H$_2$) spins since $\vert\omega_1-\omega_2\vert < 2\pi \vert D_{12} \vert$. We choose H$_1$, H$_2$ as ancilla (qubits 1, 2) and F$_3$, F$_4$ as the system (qubits 3, 4). The NMR pulse diagram to implement the quantum circuit in Fig. \[fig\_discrimnt\_von\_lud\](a) is shown in Fig. \[fig\_discrimnt\_von\_lud\](b). It begins with the initial state preparation. The thermal equilibrium state of the NMR system in the Zeeman eigenbasis under high-field, high-temperature, and secular approximation is given by [@Levitt_Spindynabook; @cavanagh], $$\rho_{\mathrm{eq}} = \mathbbm{1}_{16}/16 + \sum_{j=1}^4 \epsilon_j I_{zj}, \label{rho_eq}$$ where $\epsilon_j \sim 10^{-5}$ are the purity factors and the second term in the right hand side corresponds to the traceless deviation density matrix. The identity part is invariant under the unitary transformations and does not give rise to observable signal. Therefore only the deviation part is generally considered for both state preparation and characterization [@cory]. The initial state of the quantum register assumed in the theory section, i.e., ${\vert{\Phi_0}\rangle}$ can be prepared by applying an Hadamard operator on each of the four qubits in a pure ${\vert{0}\rangle}$ state. However, in NMR, the preparation of such pure states is difficult and instead a pseudopure state is used [@cory]. In our work, we utilize a technique based on preparing a pair of pseudopure states (POPS) [@POPS_Fung]. It involves inverting a single transition and subtracting the resulting spectrum from that of the thermal equilibrium. By inverting the transition ${\vert{0000}\rangle}$ to ${\vert{0001}\rangle}$ transition using a transition selective $\pi$ pulse, followed by Hadamard gates ([H]{}) on all the spins we obtain the POPS deviation density matrix: $$\begin{aligned} \rho_\mathrm{POPS} = \big({\vert{++++}\rangle\langle{++++}\vert}-{\vert{+++-}\rangle\langle{+++-}\vert}\big) \nonumber.\end{aligned}$$ We then implemented the quantum circuit shown in Fig. \[fig\_discrimnt\_von\_lud\] (a) using the pulse sequence in Fig. \[fig\_discrimnt\_von\_lud\] (b). As evident from circuit in Fig. \[fig\_discrimnt\_von\_lud\], controls are designed to implement $U_{\mathrm{int}}$ (Eq. \[Luder\_evolution\]) since we intend to measure $A$. Whether to map it to $A'$ or not is left to the device. The unitary operators $U_\mathrm{int}$ and $U_{a}^\dagger$ were realized by bang-bang optimal control [@Bangbang_TSM]. Hadamard and tomography operations were only few hundred micro seconds long and had a simulated fidelity of about 0.99, when averaged over $\pm 10\%$ inhomogeneous RF fields. The combined operation of $U_\mathrm{int}$ and $U_a^\dagger$ was about 17 ms in duration and had an average fidelity over $0.933$. The intermediate measurement on ancilla was realized by applying strong pulse-field-gradients (PFG). By applying a $\pi_x$ pulse on the system spins in between two symmetrically spaced PFG pulses, we realize the selective dephasing of the ancilla spins (Fig. \[fig\_discrimnt\_von\_lud\] (b)). The central $\pi_x$ also refocuses all the system-ancilla coherent evolutions during the ancilla measurement. When averaged over the sample volume this process retains only the diagonal terms in the density matrix of the ancialla spins and thus simulates a projective measurement of ancilla. Setting the total duration of this process to $1/(J_{34}+2D_{34})$ also ensures refocusing of (F$_3$, F$_4$) interactions. Finally, the density matrix of the system qubits was characterized using quantum state tomography. It involved nine independent measurements with different tomography pulses ([T]{}) (Fig. \[fig\_discrimnt\_von\_lud\] (b)) [@Tomography_Chuang447; @Tomo_Soumya]. The results of the quantum circuit (Fig. \[fig\_discrimnt\_von\_lud\]) on ${\vert{++++}\rangle\langle{++++}\vert}$ state by Lüders and von Neumann devices are described in Eqs. \[rhoL\] and \[rhoN\] respectively. For Lüders measurement with the POPS input state ${\vert{++++}\rangle\langle{++++}\vert}-{\vert{+++-}\rangle\langle{+++-}\vert}$, the final deviation density matrix (in circuit \[fig\_discrimnt\_von\_lud\]) is expected to be $$\begin{aligned} \rho_L' = \mathbbm{1}_2 \otimes \sigma_x/2. \label{Lud_finalstate}\end{aligned}$$ On the other hand, for von Neumann measurement, the POPS input state leads to a maximally mixed final state with a null deviation density matrix $(\rho_N')$. Fig. \[fig\_tomo\] compares the experimental results with the theoretically expected deviation density matrices. The correlation [@Correltn_fidlty_cory] $$\begin{aligned} C = \frac{\mathrm{Tr}[\rho'_{L}\rho'_\mathrm{exp}]}{\sqrt{\mathrm{Tr}[\rho_{L}^{\prime 2}]\mathrm{Tr}[\rho_\mathrm{exp}^{\prime 2}]}} \label{C_corelatn}\end{aligned}$$ between the theoretical ($\rho'_{L}$, Eq. \[Lud\_finalstate\]) and the experimental ($\rho'_\mathrm{exp}$) deviation density matrices was 0.923. The reduction in the correlation is mainly due to coherent errors caused by imperfect unitary operators, fluctuations in the dipolar coupling constants due to temperature gradients over the sample volume, inhomogeneous RF fields, as well as due to decoherence. The correlation expression in Eq. \[C\_corelatn\] is not directly applicable for the null-matrix $\rho'_N$. Therefore, we replace $\rho'_N$ with random traceless diagonal matrices, and obtained 0.28 as the upper bound for the correlation of $\rho'_\mathrm{exp}$ with $\rho'_N$. Therefore we conclude that the experimental deviation density matrix is much closer to $\rho'_L$ (Eq. \[Lud\_finalstate\]), and strongly favors the Lüders update rule. Conclusions =========== Quantum measurements, involving probabilistic state collapse and corresponding measurement outcomes, has always been mysterious. There have been attempts to deduce rules based on phenomenological observations. According to one of the earliest reduction rules, given by von Neumann, superposition in a degenerate subspace is destroyed by the measurement of the respective degenerate observable. This rule was later substantially modified by Gerhart Lüders. The modified rule, which is most commonly used, implies that superpositions within the degenerate subspaces are preserved under such a measurement. A protocol to determine whether a given measuring device is Lüders or von Neumann was recently formulated by Hegerfeldt and Mayato [@Discriminate_von_Luder_protocol]. In this work, we have adapted this protocol for quantum information systems, and utilize ancilla qubits for performing a desired measurement on system qubits. Moreover, we describe an NMR experiment, with two system qubits and two ancilla qubits, to discriminate between Lüders and von Neumann devices. Within the limitations of experimental NMR techniques, we found that the measurements are of Lüders type. There is a possibility that the above measurement is still of von Neumann type, if the chosen initial state happens to be a nondegenerate eigenstate of the actual system observable ($A'$). One way to rule out this possibility is by changing the initial state (Fig. \[HMprotocolfig\]). However, it is also possible that the actual system observable is dynamic, in which case it is even more difficult to discriminate between Lüders and von Neumann measurements. In this work we have not excluded these possibilities. Nevertheless, the present work opens many interesting questions. For example, how can we build a von Neumann measuring device, or even an intermediate measuring device that partly breaks the degeneracy? More importantly, further research in this direction may throw some light on fundamental aspects of quantum measurement itself. Acknowledgement {#acknowledgement .unnumbered} =============== Authors acknowledge useful discussions with Anjusha V. S. Authors are also grateful to Prof. A. K. Rajagopal and Prof. A. R. Usha Devi for numerous comments and suggestions. This work was supported by DST/SJF/PSA-03/2012-13 and CSIR-03(1345)/16/EMR-II.

--- abstract: | Let $\mu$ be a finite positive measure on the real line. For $a>0$ denote by $\EE_a$ the family of exponential functions $$\EE_a=\{e^{ist}| \ s\in[0,a]\}.$$ The exponential type of $\mu$ is the infimum of all numbers $a$ such that the finite linear combinations of the exponentials from $\EE_a$ are dense in $L^2(\mu)$. If the set of such $a$ is empty, the exponential type of $\mu$ is defined as infinity. The well-known type problem asks to find the exponential type of $\mu$ in terms of $\mu$. In this note we present a solution to the type problem and discuss its relations with known results. address: | Texas A&M University\ Department of Mathematics\ College Station, TX 77843, USA author: - 'A. Poltoratski' title: A problem on completeness of exponentials --- **Introduction** ================ Completeness of exponentials. ----------------------------- Let $\mu$ be a finite positive Borel measure on $\R$. Let us consider the family $\EE_\L$ of exponential functions $\exp(i\l t)$ on $\R$ whose frequencies $\l$ belong to a certain set $\L\subset \C$: $$\EE_\L=\{\exp(i\l t)|\ \l\in \L\}.$$ One of the classical problems of Harmonic analysis is to find conditions on $\mu$ and $\L$ that ensure completeness, i.e. density of finite linear combinations, of functions from $\EE_\L$ in $L^2(\mu)$. Versions of this problem were considered by many prominent analysts. The case when $\L$ is a sequence and $\mu$ is Lebesgue measure on an interval was solved by Beurling and Malliavin in the early sixties [@BM1; @BM2]. The so-called Beurling-Malliavin theory, created to treat that problem, is considered to be one of the deepest parts of the 20th century Harmonic Analysis. Other cases of the problem and its multiple reformulations were studied by Wiener, Levinson, Kolmogorov, Krein and many others. Such an extensive interest is largely due to the fact that it is naturally related to other fields of classical analysis, such as stationary Gaussian processes and prediction theory, spectral problems for differential operators, approximation theory, signal processing, etc. Despite considerable efforts by the analytic community many important cases of the problem remain open. The type problem ---------------- Perhaps the most studied among such open cases is the so-called type problem. Consider a family $\EE_a=\EE_{[0,a]}$ of exponential functions whose frequencies belong to the interval from 0 to $a$. If $\mu$ is a finite positive measure on $\R$ we denote by $\GG^2_\mu$ its exponential type that is defined as $$\GG^2_\mu=\inf\{\ a>0\ |\ \EE_a \textrm{ is complete in } L^2(\mu)\ \}\label{type}$$ if the set of such $a$ is non-empty and as infinity otherwise. The type problem asks to calculate $\GG^2_\mu$ in terms of $\mu$. This question first appears in the work of Wiener, Kolmogorov and Krein in the context of stationary Gaussian processes (see [@Krein1; @Krein2] or the book by Dym and McKean [@DM]). If $\mu$ is a spectral measure of a stationary Gaussian process, completeness of $\EE_a$ in $L^2(\mu)$ is equivalent to the property that the process at any time is determined by the data for the time period from 0 to $a$. Hence the type of the measure is the minimal length of the period of observation necessary to predict the rest of the process. Since any even measure is a spectral measure of a stationary Gaussian process, and vice versa, this reformulation is practically equivalent. The type problem can also be restated in terms of the Bernstein weighted approximation, see for instance the book by Koosis [@KoosisLog]. Important connections with spectral theory of second order differential operators were studied by Gelfand and Levitan [@GL] and Krein [@Krein2; @Krein3]. Closely related to spectral problems for differential operators is Krein – de Branges’ theory of Hilbert spaces of entire functions, see [@dBr]. One of the deep results of the theory says that for any positive finite (or more generally Poisson-finite) measure $\mu$ on $\R$ there is a unique nested regular chain of de Branges’ spaces of entire functions isometrically embedded in $L^2(\mu)$. An important characteristic of such a chain is the supremum $S_\mu$ of the exponential type taken over all entire functions contained in the embedded spaces. For instance, if such a chain corresponds to a regular Schrödinger operator on an interval, i.e. if $\mu$ is the spectral measure of such an operator, then $S_\mu$ is equal to the length of the interval and all spaces of the chain can be parametrized by their exponential type. It is well-known, and not difficult to show, that the problem of finding the value of $S_\mu$ is equivalent to the type problem, i.e. $S_\mu=G^2_\mu$. For more on the history and connections of the type problem see, for instance, a note by Dym [@Dym] or a recent paper by Borichev and Sodin [@BS]. General case $p\neq 2$ ---------------------- The family $\EE_a$ is incomplete in $L^2(\mu)$ if and only if there exists a function $f\in L^2(\mu)$ orthogonal to all elements of $\EE_a$. Expanding to other $1\leqslant p\leqslant \infty$ we define $$\GG^p_\mu=\sup\{\ a\ |\ \exists \ f\in L^p(\mu),\int f(x)e^{i\l x}d\mu(x)=0, \forall \ \l\in[0,a]\ \}. \label{typep}$$ We put $\GG^p_\mu=0$ if the set in is empty. By duality, for $1<p\leqslant\infty$, $\GG^p_\mu$ can still be defined as the infimum of $a$ such that $\EE_a$ is complete in $L^q(\mu),\ \frac 1p +\frac 1q =1$. Cases $p\neq 2$ were considered in several papers, see for instance articles by Koosis [@Koosis2] or Levin [@Levin2] for the case $p=\infty$ or [@GAP] for $p=1$. Since $\mu$ is a finite measure we have $$\GG^p_\mu\leqslant\GG^q_\mu\textrm{ for }p\geqslant q.\label{pq}$$ Apart from this obvious observation, the problems of finding $\GG^p_\mu$ for different $p$ were generally considered non-equivalent. One of the consequences of theorem \[main\], section \[secMain\], is that, in some sense, there are only two significantly different cases, $p=1$ (the gap problem) and $1<p\leqslant\infty$ (the general type problem). In this paper we restrict our attention to the class of finite measures. The formal reason for that is the fact that $\mu$ has to be finite for exponentials to belong to $L^2(\mu)$. This obstacle can be easily overcome if instead of $\EE_a$ one considers $E_a$, the set of Fourier transforms of smooth functions supported on $[0,a]$. All elements of $E_a$ decay fast at infinity and one one can ask about the density of $E_a$ in $L^p(\mu)$ for wider sets of $\mu$, see for instance [@BS]. One of such traditional sets is the class of Poisson-finite measures satisfying $$\int\frac {d|\mu|(x)}{1+x^2}<\infty.$$ However, due to the reasons similar to lemma \[polynomial\] below (note that if $\mu$ is Poisson-finite then $\mu/(1+x^2)$ is finite and vice versa), considering such a wider set of measures will not change the problem and all of the statements will remain the same or analogous. The gap problem {#introGAP} --------------- One of the important cases is the so-called gap problem, $p=1$. Here one can reformulate the question as follows. Let $X$ be a closed subset of the real line. Denote $$\GG_X=\sup\{\ a\ |\ \exists\ \mu\neq 0, \ \supp\mu\subset X, \hat\mu=0 \textrm{ on }[0,a]\ \}.$$ Here and in the rest of the paper $\hat\mu$ denotes the (inverse) Fourier transform of a finite measure $\mu$ on $\R$: $$\hat\mu(z)=\int_\R e^{izt}d\mu(t).$$ As was shown in [@GAP], for any finite measure $\mu$ on $\R$, $\GG^1_\mu$, as defined in the previous section, depends only on its support: $$\GG^1_\mu=\GG_X,\ X=\supp\mu.$$ This property separates the gap problem from all the cases $p>1$. For a long time both the gap problem and the type problem were considered by experts to be “transcendental,” i.e. not having a closed form solution. Following an approach developed in [@MIF1] and [@MIF2], a solution to the gap problem was recently suggested in [@GAP], see section \[secGAP\]. Some of definitions and results from [@GAP] are used in the present paper. Known examples -------------- We say that a function $f$ on $\R$ is Poisson-summable if it is summable with respect to the Poisson measure $\Pi$, $$d\Pi=dx/(1+x^2).$$ We say that a sequence of real numbers $A=\{a_n\}$ is discrete if it does not have finite accumulation points. We always assume that a discrete sequence is enumerated in the natural increasing order: $a_n\geqslant a_{n-1}$. Since the sequences considered here have $\pm\infty$ as their density points, the indices run over $\Z$. In most of our statements and definitions, the sequences do not have multiple points. We call a discrete sequence $\{a_n\}\subset \R$ separated if $|a_n-a_k|>c$ for some $c>0$ and any $n\neq k$. A classical result by Krein [@Krein1] says that if $d\mu=w(x)dx$ and $\log w$ is Poisson-summable then $\GG^p_\mu=\infty$ for all $p, \ 1\leqslant p\leqslant\infty$. A partial inverse, proved by Levinson and McKean, holds for even monotone $w$, see section \[secKLM\]. A theorem by Duffin and Schaeffer [@DS] implies that if $\mu$ is a measure such that for any $x\in \R$ $$\mu([x-L,x+L])>d$$ for some $L,d>0$ then $\GG^2_\mu\geqslant 2\pi/L$, see section \[secDS\]. For discrete measures, in the case $\supp\mu=\Z$, a deep result by Koosis shows an analogue of Krein’s result: if $\mu=\sum w(n)\delta_n$, where $$\sum\frac{\log w(n)}{1+n^2}>-\infty,$$ then $\GG^p_\mu=2\pi$ for all $p, \ 1\leqslant p\leqslant\infty$ [@Koosis2]. Not much was known about supports other than $\Z$ besides a recent result from [@Polya], which implies that if $$\mu=\sum \frac{\delta_{a_n}}{1+a_n^2}$$ for a separated sequence $A=\{a_n\}\subset\R$ then $\GG^p_\mu=2\pi D_*(A)$, where $D_*$ is the interior Beurling-Malliavin density of $A$, see section \[secGAP\] for the definition. We generalize these results in section \[secDisc\]. In addition to these few examples, classical theorems by Levinson-McKean, Beurling and de Branges show that if a measure has long gaps in its support or decays too fast, then $\GG^p_\mu=0$, see section \[classical\]. Examples of measures of positive type can be constructed using the results by Benedicks [@Benedicks], see section \[secBen\]. The most significant recent development, that allows one to modify existing examples, is the result by Borichev and Sodin [@BS], which says that “exponentially small” changes in weight or support do not change the type of a measure, see section \[secBS\]. Approach and goals of the paper ------------------------------- The problems discussed above belong to the area often called the Uncertainty Principle in Harmonic Analysis [@HJ]. A new approach developed by N. Makarov and the author in [@MIF1; @MIF2] allows one to study this area with modern tools of analytic function theory and singular integrals. Together with traditional methods, such as de Branges’ theory of Hilbert spaces of entire functions or the Beurling-Malliavin theorems, these techniques have produced some new ideas and developments. Among them is an extension of the Beurling-Malliavin theory [@MIF2], a solution to the Pólya-Levinson problem on sampling sets for entire functions of zero type [@Polya] and a solution to the gap problem [@GAP]. In the present paper we continue to apply the same approach. We focus on the type problem, the problem of finding $\GG^2_\mu$ in terms of $\mu$. Our main results are theorem \[main\] and its corollaries contained in section \[mainresults\]. In most of our statements, treating $p>1,\ p\neq 2$ did not require any additional efforts, and hence they were formulated for general $p>1$. The case $p=1$, studied in [@GAP], provided us with some useful definitions and statements, see section \[secGAP\]. **Acknowledgements.** I am grateful to Nikolai Makarov whose deep mathematical insight and intuition led to the development of the methods used in this paper. I would also like to thank Misha Sodin for getting me interested in the gap and type problems and for numerous invaluable discussions. Contents -------- The paper is organized as follows: - Section \[prelim\] contains preliminary material, including the basics of the so-called Clark theory, definitions of Beurling-Malliavin densities and a short discussion of the gap problem. - In section \[mainresults\] we state the main results of the paper. - Section \[classical\] discusses connections of our results with classical theorems by Beurling, de Branges, Duffin and Schaeffer, Krein, Levinson and McKean as well as more recent results by Benedicks, Borichev and Sodin. - Section \[lemmas\] contains several lemmas needed for the main proofs. - In section \[mainproofs\] we give the proofs of the main results. Preliminaries {#prelim} ============= Clark theory {#Clark} ------------ By $H^2$ we denote the Hardy space in the upper half-plane $\C_+$. We say that an inner function $\theta(z)$ in $\C_+$ is meromorphic if it allows a meromorphic extension to the whole complex plane. The meromorphic extension to the lower half-plane $\C_{-}$ is given by $$\theta(z)=\frac{1}{\theta^{\#}(z)}$$ where $\theta^{\#}(z)=\bar\theta(\bar z)$. Each inner function $\theta(z)$ determines a model subspace $$K_\theta=H^2\ominus \theta H^2$$ of the Hardy space $H^2(\C_+)$. These subspaces play an important role in complex and harmonic analysis, as well as in operator theory, see [@Ni2]. For each inner function $\theta(z)$ one can consider a positive harmonic function $$\Re \frac{1+\theta(z)}{1-\theta(z)}$$ and, by the Herglotz representation, a positive measure $\mu$ such that $$\label{for1} \Re \frac{1+\theta(z)}{1-\theta(z)}=py+\frac{1}{\pi}\int{\frac{yd\mu (t)}{(x-t)^2+y^2}}, \hspace{1cm} z=x+iy,$$ for some $p \geqslant 0$. The number $p$ can be viewed as a point mass at infinity. The measure $\mu$ is Poisson-finite, singular and supported on the set where non-tangential limits of $\theta$ are equal to $1$. The measure $\mu +p\delta_\infty$ on $\hat\R$ is called the Clark measure for $\theta(z)$. Following standard notations, we will sometimes denote the Clark measure defined in by $\mu_1$. More generally, if $\alpha\in \C, |\alpha|=1$ then $\mu_\alpha$ is the measure defined by with $\theta$ replaced by $\bar\alpha\theta$. Conversely, for every positive singular Poisson-finite measure $\mu$ and a number $p \geqslant 0$, there exists an inner function $\theta(z)$ satisfying . Every function $f \in K_\theta$ can be represented by the formula $$\label{for2} f(z)=\frac{p}{2\pi i}(1-\theta(z))\int{f(t)\overline{(1-\theta(t))}dt}+\frac{1-\theta(z)}{2\pi i}\int{\frac{f(t)}{t-z} d\mu (t)}.$$ If the Clark measure does not have a point mass at infinity, the formula is simplified to $$f(z)=\frac1{2\pi i}(1-\theta(z))Kf\mu$$ where $Kf\mu$ stands for the Cauchy integral $$Kf\mu(z)=\int\frac{f(t)}{t-z} d\mu (t).$$ This gives an isometry of $L^2(\mu)$ onto $K_\theta$. Similar formulas can be written for any $\mu_\alpha$ corresponding to $\theta$. In the case of meromorphic $\theta(z)$, every function $f \in K_\theta$ also has a meromorphic extension in $\C$, and it is given by the formula . The corresponding Clark measure is discrete with atoms at the points of $\{\theta=1\}$ given by $$\mu(\{x\})=\frac{2\pi }{|\theta'(x)|}.$$ If $\L\subset \R$ is a given discrete sequence, one can easily construct a meromorphic inner function $\theta$ satisfying $\{\theta=1\}=\L$ by considering a positive Poisson-finite measure concentrated on $\L$ and then choosing $\theta$ to satisfy . One can prescribe the derivatives of $\theta$ at $\L$ with a proper choice of pointmasses. For more details on Clark measures and further references the reader may consult [@PS]. Interior and exterior densities ------------------------------- A sequence of disjoint intervals $\{I_n\}$ on the real line is called *long* (in the sense of Beurling and Malliavin) if $$\sum_n\frac{|I_n|^2}{1+\dist^2(0,I_n)}=\infty,\label{long}$$ where $|I_n|$ stands for the length of $I_n$. If the sum is finite, we call $\{I_n\}$ *short*. One of the obvious properties of short sequences is that $|I_n|=o(\dist(0,I_n))$ as $n\to\infty$. In particular, $\dist(0,I_n)$ can be replaced with any $x_n\in I_n$ in . Following [@BM2] we say that a discrete sequence $\L\subset \R$ is $a$-*regular* if for every $\epsilon>0$ any sequence of disjoint intervals $\{I_n \}$ that satisfies $$\left|\frac{\#(\L\cap I_n)}{|I_n|}-a\right|\geqslant \epsilon$$ for all $n$, is short. A slightly different $a$-regularity can be defined in the following way, that is more convenient in some settings. For a discrete sequence $\L\subset \R$ we denote by $n_\L (x)$ its counting function, i.e. the step function on $\R$, that is constant between any two points of $\L$, jumps up by $1$ at each point of $\L$ and is equal to $0$ at $0$. We say that $\L$ is *strongly* $a$-*regular* if $$\int\frac{|n_{\L}(x)-ax|}{1+x^2}<\infty.$$ Conditions like this can be found in many related results, see for instance [@dBr] or [@KoosisLog]. Even though $a$-regularity is not equivalent to strong $a$-regularity, in the following definitions of densities changing “$a$-regular” to “strongly $a$-regular” will lead to equivalent definitions. The interior BM (Beurling-Malliavin) density of a sequence $\L$ is defined as $$\label{id}D_*(\L):=\sup \{\ a\ |\ \exists \ \textrm{$a$-regular subsequence}\ \ \L'\subset \L \ \}.$$ If the set is empty we put $D_*(\L)=0$. Similarly, the exterior BM density is defined as $$\label{ed}D^*(\L):=\inf \{\ a\ |\ \exists \ \textrm{$a$-regular supsequence}\ \ \L'\supset \L\ \}.$$ If no such sequence exists, $D^*(\L)=\infty$. It is interesting to observe that after the two densities were simultaneously introduced over fifty years ago, the exterior density immediately became one of the staples of harmonic analysis and spectral theory, mostly due to its appearance in the celebrated Beurling-Malliavin theorem, see [@BM2], [@HJ] or [@KoosisLog]. Meanwhile, the interior density remained largely forgotten until its recent comeback in [@Polya] and [@GAP]. It will continue to play an important role in our discussions below. The gap problem and $d$-uniform sequences {#secGAP} ----------------------------------------- Let $\Lambda=\{\lambda_1,...,\lambda_n\}$ be a finite set of points on $\R$. Define $$E(\Lambda)=\sum_{\lambda_k,\lambda_j\in\L} \log|\lambda_k-\lambda_l|.\label{electrons}$$ According to the 2D Coulomb law, the quantity $E(\L)$ can be interpreted as potential energy of the system of “flat electrons” placed at $\L$, see [@GAP]. That observation motivates the term we use for the condition below. The following example is included to illustrate our next definition. **Key example:** *Let $I\subset \R$ be an interval and let $\L=d^{-1}\Z\cap I$ for some $d>0$. Then $$\Delta=\#\L= d|I|+O(1)$$ and $$E=E(\Lambda)=\sum_{1\leqslant m\leqslant \Delta}\log\left[d^{-\Delta+1}(m-1)!(\Delta-m)!\right]= \Delta^2\log |I| + O(|I|^2)\label{eqkey}$$ as follows from Stirling’s formula. Here the notation $O(\cdot)$ corresponds to the direction $|I|\to\infty$.* The uniform distribution of points on the interval does not maximize the energy $E(\Lambda)$ but comes within $O(|I|^2)$ from the maximum, which is negligible for our purposes, see the main definition and its discussion below. It is interesting to observe that the maximal energy for $k$ points is achieved when the points are placed at the endpoints of $I$ and the zeros of the Jacobi $(1,1)$-polynomial of degree $k-2$, see for example [@Kerov]. Let $$...<a_{-2}<a_{-1}<a_0=0<a_1<a_2<...$$ be a discrete sequence of real points. We say that the intervals $I_n=(a_n,a_{n+1}]$ form a short partition of $\R$ if $|I_n|\to\infty$ as $ n\to \pm\infty$ and the sequence $\{I_n\}$ is short. **Main Definition:** Let $\L=\{\lan\}$ be a discrete sequence of real points. We say that $\L$ is $d$-uniform if there exists a short partition $\{I_n\}$ such that $$\Delta_n= d|I_n|+o(|I_n|)\ \ \ \text{for all} \ \ \ n\ \ \textrm{(density condition)}\label{density}$$ as $n\to\pm\infty$ and $$\sum_n \frac{\Delta_n^2\log|I_n|-E_n}{1+\dist^2(0,I_n)}<\infty\ \ \textrm{(energy condition)}\label{energy}$$ where $\Delta_n$ and $E_n$ are defined as $$\Delta_n=\#(\L\cap I_n)\ \ \text{ and }\ \ E_n=E(\Lambda\cap I_n)=\sum_{\lambda_k,\lambda_l\in I_n,\ \lambda_k\neq\lambda_l}\log|\lambda_k-\lambda_l|.$$ Note that the series in the energy condition is positive: every term in the sum defining $E_n$ is at most $\log|I_n|$ and there are no more than $\Delta_n^2$ terms. As follows from the example above, the first term in the numerator of is approximately equal to the energy of $\Delta_n$ electrons spread uniformly over $I_n$. The second term is the energy of electrons placed at $\L\cap I_n$. Thus the energy condition is a requirement that the placement of the points of $\Lambda$ is close to uniform, in the sense that the work needed to spread the points of $\L$ uniformly on each interval is summable with respect to the Poisson weight. For a more detailed discussion of this definition see [@GAP] In [@GAP], $d$-uniform sequences were used to solve the gap problem mentioned in the introduction. Recall that with any closed $X\subset\R$ one can associate its (spectral) gap characteristic $\GG_X$ defined as in section \[introGAP\]. The main result of [@GAP] is the following statement: \[mainGAP\][@GAP] Let $X$ be a closed set on $\R$. Then $$\GG_X=\sup\{\ d\ |\ X \textrm{ contains a } d-\textrm{uniform sequence }\}.$$ Recall that, as was proved in [@GAP], $\GG_X=\GG^1_\mu$ for any $\mu$ such that $\supp\mu=X$. The following simple observations will also be useful to us in the future: \[rem1\] $$$$ - If $\L$ is a $d$-uniform sequence then $D_*(\L)=d$, as follows easily from the density condition . - Among other things, the energy condition ensures that the points of $\L$ are not too close to each other. In particular, if $\L$ is $d$-uniform for some $d>0$ and $\L'=\{\lambda_{n_k}\}$ is a subsequence such that for all $k$, $$\lambda_{n_k+1}-\lambda_{n_k}\leqslant e^{-c|\lambda_{n_k}|}$$ for some $c>0$, then $D_*(\L')=0$. - An exponentially small perturbation of a $d$-uniform sequence contains a $d$-uniform subsequence. More precisely, if $c>0$ and $\L$ is a $d$-uniform sequence then any sequence $A=\{\alpha_n\}$ such that $|\lan-\alpha_n|\leqslant e^{-c|\lan|}$ contains a $d$-uniform subsequence $A'$ consisting of all $\alpha_{n_k}$ such that $$\lambda_{n_k+1}-\lambda_{n_k}\geqslant e^{-(c-\e)|\lambda_{n_k}|}.$$ - As discussed in [@GAP], the energy condition always holds for separated sequences. If $\L$ is separated then it is $d$-uniform if and only if $D_*(\L)=d$. Polynomial decay {#poly} ---------------- In this section we prove a version of the well-known property that adding or removing polynomial decay cannot change the type of a measure. \[polynomial\] Let $\mu$ be a finite positive measure on $\R$ and let $\alpha>0$. Consider the measure $\nu$ satisfying $$d\nu(x)=\frac{d\mu(x)}{1+|x|^\alpha}.$$ Then for any $1\leqslant p\leqslant \infty$ $$\GG^p_\mu=\GG^p_\nu.$$ Since $d\nu/d\mu\leqslant 1$, one only needs to show that $\GG^p_\mu\leqslant \GG^p_\nu.$ Suppose that $f\in L^p(\mu)$ is such that $\bar f\mu$ annihilates all $e^{iaz}, a\in (0,d)$. This is equivalent to the property that the Cauchy integral $Kf\mu$ is divisible by $e^{idz}$ in $\C_+$, i.e. it decays like $e^{idz}$ along the positive imaginary axis $i\R_+$, see for instance lemma 2 in [@Polya]. Let $N\geqslant\alpha$ be an integer. It is enough to prove the statement for $N=1$: the general case will follow by induction. First let us assume that $Kf\mu$ has at least one zero $a$ in $\C\setminus\R$. It is well-known, and not difficult to verify, that then the measure $\frac f{x-a}\mu$ satisfies $$K\left(\frac f{x-a}\mu\right)=\frac {Kf\mu}{z-a}.$$ Hence the Cauchy integral in the left-hand side still decays like $e^{idz}$ along $i\R_+$ and therefore the measure still annihilates $e^{iaz}, a\in (0,d)$. It is left to notice that $$f(x)\frac{1+|x|^\alpha}{x-a}\in L^p(\nu).$$ If $Kf\mu$ does not have any zeros outside of $\R$, note that the Cauchy integral of the measure $\eta=e^{-i\e x}f\mu$ satisfies $$K\eta=e^{-i\e z}Kf\mu$$ (see for instance theorems 3.3 and 3.4 in [@Clark]) and therefore $$K(f\mu-c\eta)=K(1-ce^{-i\e x})f\mu$$ has infinitely many zeros in $\C\setminus\R$ for any $c, |c|\neq 1$, while still decaying like $e^{i(d-\e)z}$ along $i\R_+$. Let $\mu$ be a finite measure on a separated sequence $X$, with point masses decaying polynomially. Lemma \[polynomial\] together with elementary estimates imply that then $\GG^2_\mu=\GG^1_\mu$. Hence in this case theorem \[main\] below becomes the statement from [@Polya] mentioned above: $$\GG^2_\mu=2\pi D_*(X).$$ Main Results {#mainresults} ============ Main Theorem {#secMain} ------------ Let $\tau$ be a finite positive measure on the real line. We say that a function $W\geqslant 1$ on $\R$ is a $\tau$-*weight* if $W$ is lower semi-continuous, tends to $\infty$ at $\pm\infty$ and $W\in L^1(\tau)$. \[main\] Let $\mu$ be a finite positive measure on the line. Let $1<p\leqslant\infty$ and $a>0$ be constants. Then $\GG^p_\mu\geqslant a$ if and only if for any $\mu$-weight $W$ and any $0<d<a$ there exists a $d$-uniform sequence $\L=\{\lan\}\subset \supp \mu$ such that $$\sum\frac{\log W(\lan)}{1+\lan^2}<\infty.\label{ur4}$$ We postpone the proof until section \[mainproofs\]. One of the immediate corollaries of the above statement is that the $p$-type of a measure, $\GG^p_\mu$ for $1<p\leqslant\infty$, does not depend on $p$, which may come as a surprise to some of the experts. Further corollaries of theorem \[main\] and its connections with classical results are discussed in the following sections. Discrete case {#secDisc} ------------- The conditions of theorem \[main\] are simplified for many specific classes of measures. In particular, if the measure is discrete, or absolutely continuous with regular enough density, the weight $W$ may be eliminated from the statement. Here we treat the discrete case that is important in spectral theory of differential operators and other adjacent areas. Our results in this section may be viewed as extensions of the result by Koosis mentioned in the introduction. The following statement gives a simplified formula for the type of a measure supported on a discrete sequence, excluding pathological cases when the counting function of the sequence grows exponentially. \[mainDiscrete\] Let $B=\{b_n\}$ be a discrete sequence of real points. Let $$\mu=\sum w(n)\delta_{b_n}$$ be a finite positive measure supported on $B$. Define $$D = \sup\{\ d\ |\ \exists\textrm{ d-uniform }B'\subset B,\ \sum_{\lan\in B'}\frac {\log w(n)}{1+{n}^2}>-\infty\ \}.$$ Then for any $1<p\leqslant\infty$, $$\GG^p_\mu\geqslant 2\pi D.$$ If the counting function of $B$ satisfies $\log (|n_B|+1)\in L^1_\Pi$ then $$\GG^p_\mu=2\pi D.$$ The proof is given in section \[mainproofs\]. The condition $\log (|n_B|+1)\in L^1_\Pi$ in the second part of the statement is sharp. The corresponding examples can be easily constructed using theorem \[main\] or the result by Borichev and Sodin [@BS], see theorem \[BSthm\] below. In the case when the sequence is separated, the condition can be simplified even further. Note that for $p=1$, $\GG^1_\mu=D_*(\L)$ for any separated sequence $\L$ and any measure $\mu,\ \supp\mu=\L$, by theorem \[mainGAP\]. For $p>1$ we have \[mainSeparated\] Let $\L=\{\lambda_n\}$ be a separated sequence and let $$\mu=\sum w(n)\delta_{\lambda_n}$$ be a finite positive measure supported on $\L$. Define $$D=\sup D_*(\L'),$$ where the supremum is taken over all subsequences $\L'\subset \L$ satisfying $$\sum_{\lambda_n\in \L'}\frac{\log w(n)}{1+n^2}>-\infty.\label{ur10}$$ Then $$\GG^p_\mu=2\pi D$$ for all $1<p\leqslant \infty$. Suppose that $G^p_\mu>2\pi D$ for some $D>0, p>1$. Define the $\mu$-weight $W$ as $W(\lambda_n)=(\mu(\{\lambda_n\})(1+\lambda^2_n))^{-1}$. Then by theorem \[mainDiscrete\] there exists a subsequence $\L'\subset \L$ such that $D_*(B')>D$ and is satisfied. In the opposite direction the statement follows directly from theorem \[mainDiscrete\] and remark \[rem1\]. A general sufficient condition ------------------------------ As a corollary of theorem \[mainDiscrete\] we obtain the following sufficient condition for general measures. The condition seems to be reasonably sharp, as it is satisfied by all examples of measures with positive type existing in the literature. \[SuffGen\] Let $\mu$ be a finite positive measure on $\R$. Let $A=\{a_n\}$ be a $d$-uniform sequence of real numbers such that $$\sum\frac{\log\mu((a_n-\e_n,a_n+\e_{n}))}{1+n^2}>-\infty,\label{log1}$$ where $$\e_n=\frac 13\min\left((a_{n+1}-a_n),(a_n-a_{n-1})\right).$$ Then $\GG^\infty_\mu\geqslant 2\pi d.$ For each $\tau\in [0,1]$ let us define a discrete measure $\nu_\tau$ as follows. The measure $\nu_\tau$ has exactly one pointmass of the size $$\mu((a_n-\e_n,a_n+\e_{n}))$$ in each interval $$(a_n-\e_n,a_n+\e_{n})$$ at the point $x^\tau_n$ chosen as $$x^\tau_n=\inf\{\ a\ |\ \mu((a_n-\e_n,a))\geqslant\tau \mu((a_n-\e_n,a_n+\e_{n}))\ \}.$$ Notice that $\{x^\tau_n\}$ is a $d$-uniform sequence. In view of and theorem \[mainDiscrete\], $\nu_\tau$ satisfies $$\GG^\infty_{\nu_\tau}\geqslant 2\pi d.$$ Then $$\nu=\int_0^1\nu_\tau d\tau$$ satisfies $d\nu/d\mu\leqslant 1$ and therefore $$\GG^\infty_\mu\geqslant\GG^\infty_\nu\geqslant 2\pi d.$$ Classical results and further corollaries {#classical} ========================================= The goal of this section is to give examples of applications of theorem \[main\] and discuss its connections with classical results on the type problem. Due to this reason, we prefer to deduce each statement directly from the results of the last section, rather than obtaining them from each other, even when the latter approach may slightly shorten the proof. In our estimates we write $a(n)\lesssim b(n)$ if $a(n)< C b(n)$ for some positive constant $C$, not depending on $n$, and large enough $|n|$. Similarly, we write $a(n)\asymp b(n)$ if $c a(n)< b(n)<C a(n)$ for some $C\geqslant c>0$. Some formulas will have other parameters in place of $n$ or no parameters at all. Beurling’s Gap Theorem ---------------------- If $\mu$ is a finite measure supported on a set with long gaps and the Fourier transform of $\mu$ vanishes on an interval, then $\mu\equiv 0$. If $\supp\mu$ has long gaps than for every short partition of $\R$ infinitely many intervals of the partition must be contained in the gaps of $\supp\mu$. Therefore $\supp\mu$ does not contain a sequence satisfying the density condition , i.e. it does not contain a $d$-uniform sequence for any $d>0$. Levinson’s Gap Theorem ---------------------- \[Lev\] Let $\mu$ be a finite measure on $\R$ whose Fourier transform vanishes on an interval. Denote $$M(x)=|\mu|((x,\infty)).$$ If $\log M$ is not Poisson-summable on $\R_+$ then $\mu\equiv 0$. Suppose that $\log M$ is not Poisson-summable on $\R_+$. Without loss of generality, $M(0)=1$. Let $0=a_0\leqslant a_1 \leqslant a_2\leqslant...$ be the sequence of points such that $$a_n=\inf\{\ a\ |\ M(a)\leqslant 3^{-n}\ \}.$$ Define a $|\mu|$-weight $W$ as $2^n$ on each $(a_{n-1},a_n],\ a_{n-1}<a_n$. Since $\hat\mu$ vanishes on an interval, by theorem \[main\] there exists a sequence $\L\subset\supp\mu$ satisfying the density condition with some $a>0$ on a short partition $I_n=(b_n,b_{n+1}]$, such that holds. WLOG $b_0=0$. Notice that $\log W$ is an increasing step function on $\R_+$ satisfying $\log W\gtrsim -\log M$. Also, since $\{I_n\}$ is short, $cb_{n+1}\leqslant b_n$ for some $0<c<1$ and all $n>0$. Hence, $$\sum_n\frac{\log W(\lan)}{1+\lan^2}\gtrsim\sum_{n=1}^\infty\frac{\log W(b_n)|I_n|}{1+b_n^2}$$$$\gtrsim \sum_{n=1}^\infty\frac{\log W(cb_{n+1})|I_n|}{1+b_n^2}\gtrsim \int_0^\infty \frac{-\log M(cx) dx}{1+x^2}=\infty.$$ Levinson’s result above was later improved by Beurling [@Beurling-Stanford] who showed that instead of vanishing on an interval $\hat\mu$ may vanish on a set of positive Lebesgue measure with the same conclusion. Note that an analogous improvement cannot be made in Beurling’s own gap theorem above, as illustrated by Kargaev’s counterexample, see [@KoosisLog vol. 1, p. 305]. A hybrid theorem ---------------- Beurling’s and Levinson’s Gap Theorems compliment each other by treating measures with sparse supports and fast decay correspondingly. In this section we suggest a hybrid theorem that combines the features of both statements. In comparison with Beurling’s result it shows that the measure does not have to be zero on a long sequence of intervals, it just has to be small on it. In regard to Levinson’s theorem, our statement says that the measure does not have to decay fast along the whole axis, just along a large enough set. One can show that the statement is sharp in both scales. \[BerLev\] Let $\mu$ be a finite measure on $\R$ whose Fourier transform vanishes on an interval. Suppose that there exists a sequence of disjoint intervals $\{I_n\}$ such that $$\sum \frac{|I_n|\min\left(|I_n|,\log\frac 1{|\mu|(I_n)}\right)}{1+\dist^2(I_n,0)}=\infty. \label{bleq}$$ Then $\mu\equiv 0$. We can assume that $|I_n|\to\infty$ because any subsequence of intervals with uniformly bounded lengths can be deleted from $\{I_n\}$ without affecting . Define the $|\mu|$-weight $W$ as $$W=\left[|\mu|(I_n)(1+\dist^2(I_n,0))\right]^{-1}$$ on each $I_n$. If $\hat\mu$ vanishes on an interval then for some $d>0$ there exists a $d$-uniform sequence $\L\subset\supp\mu$ satisfying . Let $$N=\{\ n\ |\ \#(\L \cap I_n)>\frac d2|I_n|\}.$$ Note that the sequence $\{I_n\}_{n\not\in N}$ cannot be long because otherwise $\L$ will not satisfy the density condition on any short partition. Therefore the part of the sum in corresponding to $n\not\in N$ is finite and $$\sum_{n\in \Z}\frac{\log W(\lan)}{1+\lan^2}\geqslant\sum_{\lan\in\cup_{k\in N} I_k}\frac{\log W(\lan)}{1+\lan^2}\gtrsim\sum_{n\in N}\frac{|I_n|\log \frac 1{|\mu|(I_n)}}{1+\lan^2}=\infty.$$ De Branges’ Gap Theorem ----------------------- \[Branges\] Let $K(x)$ be a continuous function on $\R$ such that $K(x)\geqslant 1$, $\log K$ is uniformly continuous and Poisson-unsummable. Then there is no nonzero finite measure $\mu$ on $\R$ such that $$\int^{\infty}_{-\infty}Kd|\mu|<\infty \label{messum}$$ and $\hat\mu$ vanishes on an interval. Suppose that $\mu$ satisfies and its Fourier spectrum has a gap. Since $K$ is a $\mu$-weight there must exist a $d>0$ and a $d$-uniform sequence $\L\subset\supp \mu$ satisfying with $K$ in place of $W$. Since $\L$ has positive interior density and $\log K$ is uniformly continuous, implies that $\log K$ is Poisson-summable. A theorem by Krein, Levinson and McKean {#secKLM} --------------------------------------- Our next statement combines results by Krein (part I in the statement below, case $p=2$) and by Levinson and McKean (part II, $p=2$). \[KLM\] Let $\mu$ be a finite measure on $\R$, $\mu=w(x) dx$, where $w(x)\geqslant 0$. Then I\) If $\log w$ is Poisson-summable then for any $1\leqslant p\leqslant \infty$, $\GG^p_\mu=\infty$. II\) If $\log w$ is monotone and Poisson-unsummable on a half-axis $(-\infty,x)$ or $(x,\infty)$ for some $x\in\R$ then for any $1< p\leqslant \infty$, $\GG^p_\mu=0$. If $\log w$ is Poisson-summable, denote by $H(z)$ the outer function in $\C_+$ satisfying $|H|=w$ on $\R$. Then for any $a>0$ the measure $\eta=e^{-iax}\bar H(x)dx$ annihilates all exponentials with frequencies from $[0,a)$. (Here we use the fact that the integral over $\R$ for any function from $H^1(\C_+)$ is 0.) Since $|\eta|=\mu$, it follows that $\GG^p_\mu=\infty$ for any $1\leqslant p\leqslant \infty$. In the opposite direction, suppose that $\log w$ is Poisson-unsummable and monotone on $\R_+$. Consider a $\mu$-weight $W(x)=(w(x)(1+x^2))^{-1}$. If $\GG^p_\mu>2\pi d>0$, there exists a $d$-uniform sequence $\L$ satisfying . Suppose that $\L$ satisfies on a short partition $I_n=(b_n,b_{n+1}], \ b_0=0$. Then, similarly to the proof of theorem \[Lev\], for some $0<c<1$, $cb_{n+1}<b_n$. Together with monotonicity of $\log w=-\log W-\log(1+x^2)$, we obtain $$\sum\frac{\log W(\lan)}{1+\lan^2}\gtrsim\sum_{n=0}^\infty\frac{\log W(b_n)|I_n|}{1+b_n^2}+\const$$$$\gtrsim \sum_{n=0}^\infty\frac{\log W(cb_{n+1})|I_n|}{1+b_n^2}+\const\gtrsim \int_0^\infty \frac{-\log w(cx) dx}{1+x^2}+\const=\infty.$$ A result by Borichev and Sodin on stability of type {#secBS} --------------------------------------------------- If $I\subset \R$ is an interval and $D>0$ is a constant we denote by $DI$ the interval concentric with $I$ of length $D|I|$. Following [@BS], for $\delta > 0$ and $x\in\R$, we denote $$I_{x,\delta} = [x - e^{-\delta |x|}, x + e^{-\delta |x|}].$$ If $\mu$ and $\nu$ are two finite positive measures on $\R$ we write $\mu\preccurlyeq\nu$ if there exist constants $\delta > 0,\ C > 0$, and $l > 0$, such that, for all $x\in R$, $$\mu(I_{x,\delta} )\leqslant C(1 + |x|)^l \left(\nu(2I_{x,\delta}) + e^{-2\d |x|}\right) .$$ Instead of finite measures [@BS] deals with a wider class of polynomially growing measures and uses the corresponding definition of type. As was mentioned in the introduction, in view of the statements like lemma \[polynomial\] above, such differences are not essential for the type problem and the corresponding results are equivalent. [@BS]\[BSthm\] If $\mu\preccurlyeq\nu$ then $\GG^2_\mu\leqslant\GG^2_\nu$. Let $\{a_n\}_{n\in\Z}$ be a strictly increasing discrete sequence of real points satisfying $a_{-n}=-a_n$ and $$a_{n+1}-a_n=2e^{-\d b_n},\ b_n=\frac{a_{n+1}+a_n}2\textrm{ for all }n\in\N,$$ where $\d>0$ is the constant from the definition of the relation $\mu\preccurlyeq\nu$. Denote $I_n=(a_n,a_{n+1}]$. Let $W$ be a $\nu$-weight. Then the step-function $W^*$ defined as $$W^*(x)= 1+(1 + |b_n|)^{-l}\left[\frac 1{\nu(2I_n)+e^{-2\d b_n}}\int_{2I_n} Wd\nu\right]\textrm{ on each }I_n$$ is a $\mu$-weight, as follows from the condition $\mu\preccurlyeq\nu$. Assume that $\GG^2_\mu=2\pi d>0$. Then there exists an $(d-\e)$-uniform sequence $\L\in\supp \mu$, that satisfies with $W^*$. Our goal is to modify $\L$ into an $(d-\e)$-uniform sequence in $\supp \nu$ satisfying with $W$. Notice that WLOG we can assume that each interval $2I_n$ contains at most one point of $\L$, see remark \[rem1\]. Choose $k_n$ so that $\lan\in 2I_{k_n}$. Now for each $\lan\in 2I_{k_n}$ choose a point $\alpha_n\in 2I_{k_n}\cap \supp \nu$ such that $$W(\alpha_n)\leqslant \frac 1{\nu(2I_n)}\int_{2I_n} Wd\nu.$$ WLOG $$\int_{2I_n} Wd\nu\geqslant e^{-\frac32\d b_n}$$ for all $n$: otherwise we can increase the weight $W$ to satisfy this condition and it will still remain $\nu$-summable. If $W$ is such a weight, then the interior density of the subsequence of $\L$ that falls in the intervals $I_n$ satisfying $\nu(2I_n)\leqslant e^{-2\d b_n}$ must be zero: otherwise the sum for $\L$ and $W^*$ would diverge. We can assume that $\L$ does not have such points. Then $$\log W(\alpha_n)\leqslant \log W^*(\lan)+2l\log(1+|\lan|)$$ and therefore $A=\{\alpha_n\}$ satisfies with $W$. By remark \[rem1\], $A$ has a $(d-\e)$-uniform subsequence. Hence $\GG^2_\nu\geqslant\GG^2_\mu-2\pi\e$. Notice that our proof is $p$-independent, i.e. $\GG^2$ can be replaced with $\GG^p$ for any $1<p\leqslant\infty$ in the Borichev-Sodin result. A sufficient condition by Duffin and Schaeffer {#secDS} ---------------------------------------------- Our next statement is formulated in [@DS; @BS] for Poisson-finite measures. Here we present an equivalent finite version. \[DSthm\] Let $\mu$ be a finite positive measure on $\R$ such that for any $x\in \R$ $$\mu([x-L,x+L])>c(1+x^2)^{-1}$$ for some $L,c>0$. Then $\GG^2_\mu\geqslant \pi/L$. If $\e>0$ consider $a_n=n(2L+\e)$. Then in every interval $(a_n-L,a_n+L)$ there exists a subinterval $I_n$ of the length $\e$ satisfying $$\mu(I_n)\geqslant \frac{d\e}{L(1+a_n^2)}.$$ It is left to apply theorem \[SuffGen\] to the sequence of centers of $I_n$. Benedicks’ result on unions of intervals {#secBen} ---------------------------------------- The following reslut contained in [@Benedicks] provides non-trivial examples of measures with positive type. Until now, only a few examples of this kind existed in the literature. \[ben\][@Benedicks] Let $...<a_{-1}<a_0=0<a_1<a_2<...$ be a discrete sequence of points and let $I_n=(a_n,a_{n+1}]$ be the corresponding partition of $\R$. Suppose that there exist positive constants $C_1,C_2,C_3$ such that 1\) if $$C_1^{-1}a_{2n+1}<a_{2k+1}<C_1a_{2n+1},$$ for some $n,k$, then $$C_2^{-1}|I_{2n+1}|<|I_{2k+1}|<C_2|I_{2n+1}|;$$ 2\) for all $n$ $$C^{-1}_1|a_{2n+1}|<|a_{2n-1}|<C_1|a_{2n+1}| ;$$ 3\) for all $n$ $$|I_{2n+1}|>C_3\max (|I_{2n}|,1);$$ 4) $$\sum\frac{|I_{2n+1}|^2}{1+a^2_{2n+1}}\left[\log_+\frac{|I_{2n+1}|}{|I_{2n}|}+1 \right]<\infty.\label{ur6}$$ Then for any real number $A>0$ and $1\leqslant p<\infty$ there exists a nonzero function $$f\in L^1(\R)\cap L^p(\R)\cap C^\infty(\R),\ \ \ \supp f\subset\cup I_{2n},$$ such that $\hat f=0$ on $[0,A]$. Here we will not concern ourselves with the condition $f\in C^\infty$. The rest of the statement, i.e. the existence of $f\in L^1(\R)\cap L^p(\R)$, follows from theorem \[main\]. Moreover, conditions 1 and 2 prove to be redundant. Let $\{b_n\}_{n\in\Z}$ be a sequence of positive integers, monotonically increasing to $\infty$ as $n\to\infty$ and as $n\to-\infty$, such that if one replaces $|I_{2n+1}|$ in with $b_{n}|I_{2n+1}|$ the series still converges. Consider the sup-partition of $\{I_n\}$ defined in the following way. Let $$n_0=0, n_{k+1}-n_k=b_{n_k}$$ for $k> 0$ and $$n_{k+1}-n_k=b_{n_{k+1}}$$ for $n<0$. Define $J_k=(a_{2n_k},a_{2n_{k+1}}]$. By 3, the new partition satisfies the property $|J_n|\to\infty$ and, because of monotonicity of $b_n$, $$\sum\frac{|J_{n}|^2}{1+\dist^2(0,J_n)}\left[\log_+\frac{|J_n|}{|J_n\cap\left(\cup I_{2k}\right)|}+1 \right]<\infty.\label{ur7}$$ In particular $\{J_n\}$ is short. Let $C$ be a large positive number. By $[\cdot]$ we will denote the integer part of a real number. Define a sequence $\L$ as follows. On each $J_k=(a_{2n_k},a_{2n_{k+1}}]$ place $N=[C|J_k|]$ points of $\L$ inside $J_k\cap\left(\cup I_{2n}\right)$ so that $$\lambda_{m_k}<\lambda_{m_{k+1}}<...<\lambda_{m_k+N}$$ and $$|\left(\cup I_{2n}\right)\cap (a_{2n_k},\lambda_{m_k}]|= |\left(\cup I_{2n}\right)\cap (\lambda_{m_k+N},a_{2n_{k+1}}]|=|\left(\cup I_{2n}\right)\cap (\lambda_{l},\lambda_{{l+1}}]|,$$ for all $l, m_k\leqslant l<m_k+N-1$. Then conditions 3 and 4 of the theorem imply that $\L$ satisfies the energy condition on $J_n$ and that $D_*(\L)=C$. Also the measure $$\nu={\raisebox{\depth}{\(\chi\)}}_{\cup I_{2n}}\Pi$$ and $\L$ satisfy conditions of theorem \[SuffGen\]. Therefore $\GG^p_\nu\geqslant 2\pi C$ for any $1\leqslant p\leqslant\infty$ which implies the existence of the desired function $f$ satisfying $\hat f =0$ on $(0,2\pi C)$. Notice that our proof actually produces $f\in L^\infty$. If, in addition to the conditions of the theorem, $|I_{2n}|>\const>0$, then the remaining property $f\in C^\infty$ can be added with little effort. One would need to construct $f$ supported on $\cup \frac 12 I_{2n}$ and then consider a convolution $f*\phi$ with a $C^\infty$-function $\phi$ with small support. In the general case $f$ can be “smoothed out” using functions with exponentially decreasing size of support and involving arguments like theorem \[BSthm\]. Proofs: Auxiliary Statements {#lemmas} ============================ This section contains the results that will be needed to prove theorems \[main\] and \[mainDiscrete\]. A measure with positive type ---------------------------- The following lemma is essentially proved, but not explicitly stated in [@GAP]. \[disclemma\] Let $A=\{a_n\}$ be a discrete sequence of distinct real numbers that has bounded gaps, i.e. $a_{n+1}-a_{n}<C$ for some $0<C<\infty$. Denote by $b_n$ the middle of the interval $(a_n,a_{n+1})$, $b_n=(a_n+a_{n+1})/2$. Suppose that the sequence $A$ is $d$-uniform for some $d>0$. Then there exists a finite positive measure supported on $B=\{b_n\}$, $$\mu=\sum\beta_n\delta_{b_{n}},$$ satisfying $$0<\beta_{n}\leqslant \frac{\sqrt{a_{n+1}-a_{n}}}{1+a_n^2},\label{ur100}$$ such that $\GG^\infty_\mu\geqslant 2\pi d$. Let $\theta$ be the meromorphic inner function constructed for the sequence $A$ as in lemma 5 from [@GAP]. By construction, the Clark measure $\nu=\mu_{-1}$ corresponding to $\theta$ is supported on $B$ and satisfies $$\nu(\{b_n\})\lesssim a_{n+1}-a_{n},\label{ur101}$$ see the estimate (7.3) in [@GAP]. Let $c=d-\e$. As was proved in [@GAP], if $\theta$ satisfies the conditions of lemma 5, [@GAP], and $A$ is $d$-uniform, then there exists $f\in K_\theta$ that is divisible by $e^{icz}$ in $\C_+$. (This is one of the main steps in the proof of theorem 2, [@GAP]. See the part from the fourth line before claim 1 to the end of part I of the proof.) Then, by the Clark representation, $2\pi i f=(1+\theta)Kf\nu$. Since $1+\theta$ is outer, $Kf\nu$ is divisible by $e^{icz}$ in $C_+$. Because $\e$ is arbitrary, by lemma \[polynomial\], the measure $\mu=|f|\nu/(1+x^2)$ satisfies $\GG^\infty_\mu\geqslant 2\pi d$. Since $f\in L^2(\nu)$ and $\nu$ satisfies , considering a constant multiple of $\mu$ if necessary, we obtain . Construction of an auxiliary sequence ------------------------------------- To apply our previous lemma in the main proofs we will need the following \[sequences\] Let $B=\{b_n\}$ be a $d$-uniform sequence satisfying and on a short partition $\{I_n\}$. Let $w(n)$ be a positive bounded function on $\Z$ such that $$\sum\frac {\log w(n)}{1+n^2}>-\infty.\label{w}$$ Then for any $\e>0$ there exists a discrete sequence $A=\{a_n\}$ satisfying: 1\) $a_{n+1}-a_{n}<1/\e$. 2\) Define the sequence $C=\{c_k\}$ as $c_k=\frac{a_{k+1}+a_k}2$. Then the sequence $B'=B\cap C$ satisfies $$\#\left(B'\cap I_n\right)\geqslant (d-e)|I_n|$$ for large enough $|n|$. 3\) If $b_n=c_k$, i.e. $b_n$ is the middle of $(a_k, a_{k+1})$, then $a_{k+1}-a_k\leqslant w(n)$. 4\) $A$ is $2d$-uniform 5) $D^*(C\setminus B)\leqslant d+\e$. Denote $$l_n=\min (b_{n+1}-b_n, b_n-b_{n-1}, w(n)).$$ Consider the sequence $P=\{p_n\}$ defined as $$p_{2n}=b_n-\frac 13 l_n, \ \ p_{2n+1}=b_n+\frac 13 l_n.$$ Choose a large $L>>1/\e$. Define the sequence $Q$ as follows: if $p_{2n+2}-p_{2n+1}>L$, insert $M=[(p_{2n+2}-p_{2n+1})/L]$ points of $Q$ into the interval $(p_{2n+1},p_{2n+2})$ uniformly, i.e. at the points $$p_{2n+1}+k\frac{p_{2n+1}-p_{2n+2}}{M+1},\ k=1,2,...M.$$ Now put $A=P\cup Q$. By our construction the sequence $A$ satisfies $$2\#(B\cap I_n)-2\leqslant \#(A\cap I_n)\leqslant 2\#(B\cap I_n)+\e|I_n|.$$ To make $A$ satisfy the more precise density condition with $2d$ we may need to delete some points of $B$ on each interval $I_n$ and consider a smaller sequence $B'$ in place of $B$ in the above construction. Note that we would have to delete at most $\e|I_n|$ points from $B$ on each $I_n$ and that $B'$ will satisfy the energy condition as a subsequence of $B$. After such an adjustment, $A$ will satisfy 1), 2), 3) and the density condition with $2d$. Note that $A$ satisfies the energy condition on $\{I_n\}$. Indeed, let us denote $\Delta_n=\#(P\cap I_n)$ and $\Gamma_n=\#(Q\cap I_n)$. Then $$\#(A\cap I_n)^2\log |I_n|-\sum_{a_n,a_k\in A\cap I_n}\log |a_n-a_k|=$$$$\left(\Delta_n^2\log |I_n|-\sum_{a_n,a_k\in P\cap I_n}\log |a_n-a_k|\right)+$$$$\left(\Gamma_n^2\log |I_n|-\sum_{a_n,a_k\in Q\cap I_n}\log |a_n-a_k|\right)+$$$$2\left(\Delta_n\Gamma_n\log |I_n|-\sum_{a_n\in P\cap I_n,a_k\in Q\cap I_n}\log |a_n-a_k|\right)=$$$$I+II+III.$$ To estimate $I$ notice that for any $p_{2k}\in P\cap I_n$, $$-\log (p_{2k+1}-p_{2k})\leqslant -\log w(k),$$ by our choice of points $p_{2k},p_{2k+1}$. The rest of the terms in $I$ can be estimated by the similar terms for $B'$, i.e. $$I\lesssim \left(\#(B'\cap I_n)^2\log |I_n|-\sum_{b_n,b_k\in B'\cap I_n}\log |b_n-b_k|\right)$$ $$-\sum_{p_{2k}\in P\cap I_n}\log w(k)+O(|I_n|^2).$$ Since $B'$ satisfies the energy condition and because of and shortness of the partition, $I$ will give finite contribution to the energy sum in . To estimate $II$ notice that points in $Q$ are at a distance at least $L/2$ from each other. Therefore $$II\lesssim \left(\Gamma_n^2\log |I_n|-\sum_{0\leqslant n,k\leqslant \Gamma_n}\log |n-k|\right) + O(\Gamma_n^2)=$$$$\Gamma_n^2\log \frac{|I_n|}{\Gamma_n} + O(\Gamma_n^2)$$ after estimating the sum via Stirling’s formula. Notice that since $\Gamma_n<|I_n|$ and $$\log \frac{|I_n|}{\Gamma_n} <\frac{|I_n|}{\Gamma_n},$$ the last quantity will also give finite contribution to . Finally, $III$ can be estimated similarly to $II$. Just notice that any point $a_j$ in $P$ is at a distance at least $L/2$ from $Q$ and therefore $$\Gamma_n\log |I_n|-\sum_{a_k\in Q\cap I_n}\log |a_j-a_k|\lesssim \Gamma_n \log \frac{|I_n|}{\Gamma_n}+O(|I_n|^2).$$ Summing over all $a_j\in P\cap I_n$ and recalling that $\#(P\cap I_n)=\Delta_n\lesssim |I_n|$ we again get a finite quantity in . To prove 5), let us split $C$ into two subsequences: $$C_1=\left\{(a_n+a_{n+1})/2\ |\ a_n, a_{n+1}\in P\right\}\textrm{ and }C_2=C\setminus C_1.$$ Notice that $C_1\setminus B'$ has at most one point between each two points of $B'$. Therefore, $$D^*(C_1\setminus B)\leqslant D^*(B)\leqslant d+\e.$$ Also, if $2/L<<\e$ then $D^*(C_2)<\e$, because any two points of $C_2$ are at a distance at least $L/2$ from each other. Existence of extremal measure with a spectral gap ------------------------------------------------- The lemma in this section can be viewed as a version of de Branges’ theorem 66 from [@dBr]. The last section of [@GAP] contains a discussion of that theorem and its equivalent reformulations. Here and throughout the rest of the paper we will use the standard notation $S(z)=e^{iz}$ for the exponential inner function in the upper half-plane. In general, $S^a(z)=e^{iaz}$ is inner in $\C_+$ if $a>0$ and inner in $\C_-$ if $a<0$. \[t66\] Let $\mu$ be a finite complex measure such that $\hat\mu\equiv 0 $ on $[0,a]$. Let $W$ be a $|\mu|$-weight. Then there exists a finite measure $\nu=\sum\alpha_n\delta_{\lan}$ concentrated on a discrete sequence $\L=\{\lan\}$ such that 1\) $\L\subset\supp\mu$; 2\) $W$ is a $|\nu|$-weight; 3\) $\hat\nu\equiv 0$ on $[0,a]$; 4\) The Cauchy integral $K\nu$ has no zeros in $\C$, $K\nu/S^a$ is outer in $\C_+$ and $K\nu$ is outer in $\C_-$. It will be more convenient for us to assume that $\hat\mu\equiv 0$ on a symmetric interval $[-a,a]$. Then $\bar\mu$ has the same property. Hence we can assume that the measure is real (otherwise consider $\mu\pm\hat\mu$). Consider the following set of measures on $\supp\mu$: $$M_W=\{\ \nu\ |\ \int Wd|\nu|\leqslant 1,\ \hat\nu=0\textrm{ on }[-a,a], \ \supp\nu\subset\supp\mu,\ \nu=\bar\nu\}.$$ Notice that the set is non-empty, because $\mu\in M_W$, and convex. It is also $*$-weakly closed in the space of all finite measures on $\supp\mu$. Therefore by the Krein-Milman theorem it has an extreme point. Let $\nu$ be such a point. We claim that it is the desired measure. First, let us note that $\hat\nu\equiv 0$ on $[-a,a]$. It is well-known that this property is equivalent to the property that $\nu$ annihilates the Payley-Wiener class $PW_a$, i.e. that for any bounded $f\in PW_a$, $$\int fd\nu=0,$$ see for instance the last section of [@GAP]. Next, let us show that the set of real $L^\infty(|\nu|)$-functions $h$, such that $ \widehat{h\nu}\equiv 0$ on $[-a,a]$, is one-dimensional and therefore $h=c\in \R$. (This is equivalent to the statement that the closure of $PW_a$ in $L^1(|\nu|)$ has deficiency 1, i.e. the space of its annihilators is one dimensional.) Let there be a bounded real $h$ such that $ \widehat{h\nu}\equiv 0$ on $[-a,a]$. WLOG $h\geqslant 0$, since one can add constants, and $\int W|h|d|\nu|=1$. Choose $0<\alpha<1$ so that $0\leqslant\alpha h<1$. Consider the measures $\nu_1=h\nu$ and $\nu_2=(1-\alpha)^{-1}(\nu-\alpha\nu_1)$. Then both of them belong to $M_W$ and $\nu=\alpha \nu_1+(1-\alpha)\nu_2$ which contradicts the extremality of $\nu$. Now let us show that $\nu$ is discrete. Let $g$ be a continuous compactly supported real function on $\R$ such that $\int gd|\nu|=0$. By the previous part, there exists a sequence $f_n\in PW_a$, $f_n\to g$ in $L^1(|\nu|)$. Indeed, otherwise there would exist a function $h\in L^\infty(|\nu|)$ annihilating all $f\in PW_a\cap L^1(|\nu|)$ and such that $\int hgd|\nu|=1$. Since $\int gd|\nu|=0$, $h\neq const$ and we would obtain a contradiction with the property that the space of annihilators is one-dimensional. Since $\nu$ annihilates $PW_a$ and $(f_n(z)-f_n(w))/(z-w)\in PW_a$ for every fixed $w\in\C\sm\R$, $$0=\int \frac{f_n(z)-f_n(w)}{z-w}d\nu(z)=Kf_n\bar\nu(w)- f_n(w)K\nu(w)$$ and therefore $$f_n(w)=\frac{Kf_n\nu}{K\nu}(w).$$ Taking the limit, $$f=\lim f_n=\lim \frac{Kf_n\nu}{K\nu}=\frac{K g\nu}{K\nu}.$$ Since all of $f_n$ are entire, one can show that the limit function $f$ is also entire. Indeed, first notice that there exists a positive function $V\in L^1(|\nu|)$ such that $f_{n_k}/V \to g/V$ in $L^\infty(|\nu|)$, for some subsequence $\{f_{n_k}\}$. To find such a $V$ first choose $f_{n_k}$ so that $||f_{n_k}-g||_{L^1(|\nu|)}<3^{-k}$ and then put $$V=1+\sum 2^k |f_{n_k}-g|.$$ Denote $F_k=f_{n_k}/V$ and $\eta=V|\nu|$. Then $F_k$ converge in $L^2(\eta)$ and by the Clark theorem $(1-I)KF_k\eta$ converge in $H^2(\C_+)$, where $I$ is the inner function whose Clark measure is $\eta$. Notice that $$f_{n_k}=\frac{Kf_{n_k}\nu}{K\nu}=\frac{KF_k\eta}{K\nu}=\frac{(1-I)KF_k\eta}{(1-I)K\nu}.$$ Now let $T$ be a large circle in $\C$ such that $|(1-I)K\nu|>\const>0$ on $T$. Denote $T_\pm=T\cap \C_\pm$ and let $m_T$ be the Lebesgue measure on $T$. Since $(1-I)KF_k\eta$ converge in $H^2(\C_+)$, $f_{n_k}$ converge in $L^1(T_+, m_T)$. Similarly, $f_{n_k}$ converge in $L^1(T_-, m_T)$. By the Cauchy formula it follows that $f_{n_k}$ converge normally inside $T$ and therefore $f$ is analytic inside $T$. Since such a circle $T$ can be chosen to surround any bounded subset of $\C$, $f$ is entire. Since the numerator in the representation $$f=\frac{K g\bar\nu}{K\bar\nu}$$ is analytic outside the compact support of $g$, the measure in the denominator must be singular outside of that support: Cauchy integrals of non-singular measures have jumps at the real line on the support of the a.c. part, which would contradict the property that $f$ is entire. Choosing two different functions $g$ with disjoint supports we conclude that $\nu$ is singular. Moreover, since $f$ is entire, the zero set of $f$ has to be discrete. Since $\nu$ is singular, $K\nu$ tends to $\infty$ nontangentially in $\C_+$ at $\nu$-a.e. point and $f=0$ at $\nu$-a.e. point outside of the support of $g$. Again, by choosing two different $g$ with disjoint supports, we can see that $\nu$ is concentrated on a discrete set. It remains to verify 4). Since we chose to deal with the symmetric interval $[-a,a]$, we need to show that $K\nu/S^{\pm a}$ are outer in $\C_\pm$ correspondingly. Let $J$ be the inner function corresponding to $|\nu|$ ($|\nu|$ is the Clark measure for $J$). Denote $$G=\frac 1{2\pi i}(1-J)K\nu\in K_J.$$ As was mentioned in section \[Clark\], $G$ has non-tangential boundary values $|\nu|$-a.e. and $$\nu=G|\nu|.$$ Since $K\nu$ is divisible by $S^a$ in $\C_+$, $G$ is divisible by $S^a$ in $\C_+$. Suppose that $G=S^a UH$ for some inner $U$. Since the measure $\nu$ is real, $\bar G=G$, $|\nu|$-a.e. Let $F\in K_J$ be the function such that $\bar J G=\bar F$. Since $J=1$, $|\nu|$-a.e., $F=\bar G= G$, $|\nu|$-a.e. Since functions in $K_J$ are uniquely determined by their traces on the support of the Clark measure $|\nu|$, $F=G=S^a UH$. Notice that the function $h=S^a (1+U)^2H$ also belongs to $K_J$: $$\bar J h=\bar J S^a (1+U)^2H=(\bar J G)\bar U(1+U)^2=\bar G\bar U(1+U)^2$$$$=\overline{S^a (1+U)^2H}= \overline{h}\in \overline{H^2}(\C_+),$$ because $\bar U(1+U)^2$ is real a.e. on $\R$. Denote by $\gamma$ the measure from the Clark representation of $h$, i.e. $$\gamma=h|\nu|, \ \ h=\frac 1{2\pi i}(1-J)K\gamma.$$ Then $$\gamma=h|\nu|=\bar U(1+U)^2G|\nu|=\bar U(1+U)^2\nu.$$ The Cauchy integral of $\gamma$ is divisible by $S^a$ in $\C_+$ because $h$ is divisible by $S^a$ in $\C_+$. Since $\bar U(1+U)^2$ is real, a constant multiple of $\gamma$ belongs to $M_W$. Since $U$ is non-constant and $|\nu|$ is the Clark measure for $J$, $\gamma$ is not a constant multiple of $\nu$. Again we obtain a contradiction with the property that the space of annihilators is one-dimensional. Thus $G/S^a\in K_J$ is outer in $C_+$. Since $ J \bar G=\bar G$, the pseudocontinuation of $G$ does not have an inner factor except $S^{-a}$ in $\C_-$ as well. Hence $K\nu/S^{\pm a}$ is outer in $\C_\pm$. If $G$ has a zero at $x=a\in \R$ outside of $\supp \nu$ then $$\frac G{x-a}\in K_J$$ and the measure $$\gamma= \frac G{x-a}|\nu|$$ leads to a similar contradiction with the property that the space of annihilators is one-dimensional, since $(x-a)^{-1}$ is bounded and real on the support of $\nu$. Since $G=\frac 1{2\pi i}(1-J)K\nu,$ $K\nu$ does not have any zeros on $\R$. A statement similar to lemma 9 from [@GAP], where $S^a$ was replaced with an arbitrary inner function, can also be formulated in the case of lemma \[t66\]. Estimates of $\log|\theta|$ for a meromorphic inner function ------------------------------------------------------------ \[thetasum\] Consider a short partition $\{I_n\}$ of $\R$. Consider the set of circles $T_n=\{z\ |\ |z-\xi_n|=2|I_n|\}$ where $\xi_n\in I_n$. Let $\theta$ be a meromorphic inner function such that $$\sum_n\frac{\#\left(\{\theta=1\}\cap 10I_n\right)|I_n|}{1+\dist^2(0,I_n)}<\infty. \label{ur30}$$ Then the integrals $$p_n=\int_{T_n}\left|\log |\theta(z)|\right|d|z|$$ satisfy $$\sum_n\frac {p_n}{1+\dist^2(0,I_n)}<\infty.\label{ur1}$$ Suppose that $\theta=S^aB$ for some $a>0$ and some Blaschke product $B, \{B=0\}=\{a_n\}\subset \C_+$. Then $$\log |\theta|=\log |S^a| +\log |B|.$$ The integrals of $|\log |S^a||$ are summable because $$|\log |S^a||\lesssim |I_n|$$ on $T_n$ and the sequence $\{I_n\}$ is short. To estimate the integral of $|\log |B||$ notice that $$|\log |B(z)||=\sum_{a_k\in D_n}\left|\log\frac{|z-a_k|}{|z-\bar a_k|}\right|+\sum_{a_k\not\in D_n}\left|\log\frac{|z-a_k|}{|z-\bar a_k|}\right|,$$ where $D_n$ is the disk, $D_n=\{z\ |\ |z-\xi_n|\leqslant 3|I_n|\}$. Elementary estimates show that for any $a_n\in D_n$ $$\int_{T_n}\left|\log\frac{|z-a_k|}{|z-\bar a_k|}\right|d|z|\lesssim |I_n|.$$ Also, since for each $a_k\in D_n$ the argument of $\frac{z-a_k}{z-\bar a_k}$ increases by at least $\pi$ on the diameter of $D_n$, that is contained in $10I_n$, the number of points $a_k\in D_n$ is $\lesssim \#\left(\{\theta=1\}\cap 10I_n\right)$. Hence, because of , such integrals will give a finite contribution to the sum in . For each $a_k\not\in D_n$ one can show that $$\int_{T_n}\left|\log\frac{|z-a_k|}{|z-\bar a_k|}\right|d|z|\lesssim\int_{I_n} \frac {|I_n|\Im a_kdx}{(\Re a_k -x)^2+(\Im a_k)^2} .$$ Notice that $$\sum_k\int_{I_n} \frac {\Im a_kdx}{(\Re a_k -x)^2+(\Im a_k)^2}=\int_{I_n} (\arg B)'\leqslant 2\pi\cdot\#\left(\{\theta=1\}\cap I_n\}\right)+\const.$$ Again, because of , the integrals for $a_k\not\in D_n$ will give a finite contribution in A version of the first BM theorem --------------------------------- The following lemma is essentially a version of the so-called first Beurling-Malliavin theorem, see also [@MIF2]. \[BM1\] Let $\{I_n\}$ be a long sequence of intervals and let $c$ be a positive constant. Denote by $I_n'$ and $I_n''$ the intervals of the length $c|I_n|$ adjacent to $I_n$ from the left and from the right correspondingly. Let $u$ be a real function on $\R$ such that $$\Delta_n=\sup_{I_n''}u-\inf_{I_n'}u\geqslant d|I_n|$$ for all $n$ and for some $d>0$. Then $u$ is not a harmonic conjugate of a Poisson-summable function. Note that if $\ti u\in L^1_\Pi$ then $f=e^{-iu+\ti u}$ is an outer function in the Smirnov class in $\C_+$. Moreover, $f$ belongs to the kernel $N^+[e^{iu}]$ of the Toeplitz operator with the symbol $e^{iu}$ in the Smirnov class. This contradicts a Toeplitz version of the first BM theorem, see section 4.4 of [@MIF2]. An estimate for an extremal discrete measure of positive type ------------------------------------------------------------- In this section we show that that a discrete measure of positive type, like in the statement of lemma \[t66\], must have log-summable pointmasses. We start with the following elementary statement that can be easily verified. \[short2I\] Let $\{I_n\}$ be a short sequence of intervals and let $C>1$. Denote $$l_n=\sum_{I_m\cap CI_k\neq \emptyset}|I_m|.$$ Then $$\sum\frac{l_n|I_n|}{1+\dist^2(0,I_n)}<\infty.$$ Our main statement in this section is \[denlog\] Let $\nu$ be a finite measure $$\nu=\sum\alpha_n\delta_{\lan}$$ on a discrete sequence $\L=\{\lan\}$, such that $\alpha_n\neq 0$, $\hat\nu\equiv 0$ on $[0,2\pi d]$, $K\nu$ does not have any zeros in $\C$, $K\nu/S^d$ is outer in $\C_+$ and $K\nu$ is outer in $\C_-$. Then for any $\e>0$, $\L$ contains a $(d-\e)$-uniform subsequence and $$\sum\frac{\log |\alpha_n|}{1+n^2}>-\infty.\label{ur2}$$ The statement that $\L$ contains a $(d-\e)$-uniform subsequence follows from the property that $\GG_\L\geqslant d$ and theorem \[mainGAP\]. To establish , let us first show that there exists a short partition $\{I_k\}$ of $\R $ such that $\L$ satisfies with $d$ on that partition. Let $J$ be the inner function whose Clark measure is $|\nu|$. Then by the Clark theorem the function $$Q(z)=(1-J)K\nu$$ belongs to $K_J$. It follows from the properties of $K\nu$ that $Q=S^dO$ in $\C_+$ for some outer $O$ and $\bar JQ=\bar O$. Therefore the argument of $O$ satisfies $u=2\arg O=\arg J- dx$. Notice that $\arg J$ is a growing function that is equal, up to a bounded term, to the counting function of $\L$. Also, since $O\in H^2$, $\tilde u\in L^1_\Pi$. If the desired short partition $\{I_k\}$, where $\L$ satisfies , does not exist then there exists a long sequence of intervals $\{J_k\}$ such that $$\left|\#(\L\cup J_k) - d|J_k|\right|\geqslant c_1 |J_k|\label{ur20}$$ for each $k$ and for some $c_1>0$. First, let us assume that the difference in the left-hand side is positive for a long subsequence of $\{J_k\}$. Let $J_k', J_k''$ denote the intervals of the length $c_2|J_k|, 0< c_2<<c_1$, adjacent to $J_k$ from the left and from the right correspondingly. Since $u'$ is bounded from below we get that $$\Delta_k=\inf_{J_k''}u-\sup_{J_k'}u>c_3 |J_k|$$ for some $c_3>0$ on a long subsequence of $\{J_k\}$, if $c_2$ is small enough. By lemma \[BM1\], this contradicts the property that $\tilde u\in L^1_\Pi$. If the difference in is negative for a long subsequence of $\{J_k\}$ then lemma \[BM1\] can be applied to $-u$ and the intervals $J_k', J_k''$ chosen so that $J_k', J_k''\subset J_k,\ |J_k'|=|J_k''|=c_2|J_k|$, $J_k'$ shares its left endpoint with $J_k$ and $J_k''$ shares its right endpoint with $J_k$, to arrive at the same contradiction. Hence a short partition where $\L$ satisfies with $d$ does exist. Let $\{I_k\}$ be such a partition. Let $\lambda_{n_k}\in I_k$ be such that $$\log_-\alpha_{n_k}=\max_{\lan\in I_k} \log_-\alpha_{n}.$$ Suppose that is not satisfied. Then $$\sum_k \#(\L\cap I_k)\frac{\log_-\alpha_{n_k}}{1+n_k^2}\asymp \sum_k |I_k|\frac{\log_-\alpha_{n_k}}{1+n_k^2}=\infty. \label{ur32}$$ Consider $\mu=\sum_n \delta_{\lan}$ the counting measure of $\L$. Since $\L$ satisfies , $\mu$ is Poisson-finite. Let $\theta$ be the inner function such that $\mu$ is its Clark measure. Since $\nu$ is finite, it can be represented as $\nu=f\mu$ with $f\in L^2(\mu)$. Hence, by Clark theory, $$F=\frac 1{2\pi i}(1-\theta)K\nu\in K_\theta$$ with $F(\lambda_n)=f(\lan)=\alpha_n$. For each $k$ consider the disk $$D_k=\{z \ |\ |z-\lambda_{n_k}|< 2|I_k|\}$$ and its boundary circle $T_k=\partial D_k$. Notice that for each $k$, $F$ does not have zeros in $D_k$. It does have poles at the points $\bar a_n\in \C_-$, where $A=\{a_n\}$ are the zeros of $\theta$ in $\C_+$. Hence in $D_k$ the function $F$ admits factorization $F= H_k/B_k$, where $B_k$ is the finite Blaschke product in $D_k$ ($|B_k|=1$ on $T_k$) with zeros at $\bar A\cap D_k$, and $H_k$ is analytic without zeros in $D_k$. Notice that $$-\int_{T_k}\log_- |F(z)|d|z|\leqslant\int_{T_k}\log |F(z)|d|z|=\int_{T_k}\log |H_k(z)|d|z|\lesssim |I_k|\log\alpha_{n_k} \label{ur2a}$$ by Jensen’s inequality, because $F$ has only poles and no zeros in $D_k$. At the same time, since $F\in K_\theta$, it belongs to $H^2(\C_+)$ and is equal to $\theta \bar G, G\in H^2(\C_+)$ in $\C_-$. Denote by $T_k^\pm$ the upper and lower halves of $T_k$. Since the absolute value of an $H^2$ function is bounded by $$\const+\const\ |y|^{-1/2}$$ inside the half-plane, we have $$\int_{T_k}\log_+ |F(z)|d|z|\leqslant\int_{T^+_k}\log_+ |F(z)|d|z|+ \int_{T_k^-}\log_+| G(z)|d|z|$$$$+\int_{T_k^-}\log_+|\theta (z)|d|z|\lesssim |I_k| + v_k,\label{ur3}$$ where $\sum v_k/(1+a_{n_k}^2)<\infty$ by lemma \[thetasum\], because $$\#(\{\theta=1\}\cap 10I_k)\lesssim \sum_{I_m\cap 10I_k\neq\emptyset}|I_m|$$ and is satisfied by lemma \[short2I\]. Since $H_k\neq 0$ in $D_k$, $\log |H|$ is harmonic in $D_k$. Hence its values on $I_k$ can be recovered from the values of $\log |H_k|=\log |F|$ on $T_k$ via the Poisson formula. By , the Poisson integral of $\log_+|F|$ will deliver a small contribution, i.e. on each $I_k$ it will be equal to a function $h^+_k$ such that $$\sum \int_{I_k}h^+_k(x)d\Pi(x)<\infty.$$ On the other hand, the Poisson integral of $\log_-|F|$ in $D_k$, restricted to $I_k$, will be equal to $h^-_k$, where $h^-_k(x)\asymp \log\alpha_{n_k}$ for all $x\in I_k$ by . Hence by $$\sum \frac{\int_{I_k}\log|H_k|dx}{1+\dist^2(0,I_k)}=-\infty.$$ Furthermore, similarly to the proof of lemma \[thetasum\], $$\deg B_k\lesssim \#(\{\theta=1\}\cap 5I_k)\lesssim \sum_{I_m\cap 5I_k\neq\emptyset} |I_m|.$$ Therefore by lemma \[short2I\] $$\sum_k \frac{|I_k|\deg B_k}{1+\dist^2(0,I_k)}<\infty.$$ Thus $$\int_\R\log |F(x)|d\Pi\asymp \sum_k \frac{\int_{I_k}\left(\log |B_k(x)|+\log |H_k(x)|\right)dx}{1+\dist^2(0,I_k)}\lesssim$$$$\sum_k \frac{|I_k|\deg B_k+\int_{I_k}\log |H_k(x)|dx}{1+\dist^2(0,I_k)}=-\infty$$ and we obtain a contradiction. Using results of [@GAP] one can prove a slightly stronger statement that $\L$ itself is $d$-uniform. Equivalence of completeness in $L^p$ and $C_W$ ---------------------------------------------- The theorem we discuss in this section relates the type problem to Bernstein’s study of weighted uniform approximation, see [@KoosisLog] or [@BS]. Consider a weight $W$, i.e. a lower semicontinuous function $W:\R\to [1,\infty)$ that tends to $\infty$ as $x\to\pm\infty$. We define $C_W$ to be the space of all continuous functions on $\R$ satisfying $$\lim_{x\to\pm\infty} \frac{f(x)}{W(x)}=0.$$ We define the norm in $C_W$ as $$||f||=|| fW^{-1}||_\infty.$$ The following is a well-known result by A. Bakan. For reader’s convenience we supply a short proof. \[Bakan\] Let $\mu$ be a finite positive measure on $\R$. Then the system of exponentials $\EE_d$ is complete in $L^p(\mu)$ for some $1\leqslant p\leqslant \infty$ if and only if there exists a $\mu$-weight $W\in L^p(\mu)$ such that $\EE_d$ is complete in $C_W$. If $\EE_d$ is complete in $C_W$ for some $\mu$-weight $W\in L^p(\mu)$ then for any bounded continuous function $f$ there exists a sequence $\{S_n\}$ of finite linear combinations of exponentials from $\EE_d$ such that $S_n/W$ converges to $f/W$ uniformly. Then $S_n$ converges to $f$ in $L^p(\mu)$. Hence $\EE_d$ is complete in $L^p(\mu)$. Suppose that $\EE_d$ is complete in $L^p(\mu)$. Let $\{f_n\}_{n\in\N}$ be a set of bounded continuous functions on $\R$, that is dense in $C_W$. Let $\{S_{n,k}\}_{n,k\in\N}$ be a family of finite linear combinations of exponentials from $\EE_d$ such that $$||f_n-S_{n, k}||_{L^p(\mu)}<4^{-(n+k)}.$$ Denote $$W=1+\sum_{n,k\in \N}2^{n+k}|f_n-S_{n, k}|.$$ Notice that then $W\in L^p(\mu)$ and $S_{n, k}/W\to f_n/W$ uniformly as $k\to\infty$. Since $\{f_n\}$ is dense in $C_W$, $\EE_d$ is complete in $C_W$. Proofs of main results {#mainproofs} ====================== Proof of theorem \[mainDiscrete\] --------------------------------- $ $ To prove that $\GG^p_\mu\geqslant 2\pi D$, WLOG we can assume that $B$ itself is a $d$-uniform sequence for some $d>0$ and that $w$ satisfies . Fix a small $\e>0$. Let $C=\{c_n\}$ be the sequence provided by lemma \[sequences\]. Then by lemma \[disclemma\] (applied to $C$ and $w^2$) there exists a finite positive measure $\nu=\sum \sigma_n\delta_{c_n}$ concentrated on $C$, satisfying $$0<\sigma_n<w(k)\textrm{ for } c_n=b_k\textrm{ and }\GG^\infty_\nu\geqslant 2\pi (2d).$$ Let $\theta$ be the Clark inner function corresponding to $\nu$. Then there exists a function in $K_\theta$ divisible by $S^{2\pi (2d-\e)}$ in the upper half-plane, i.e. $S^{2\pi (2d-\e)}h\in K_\theta$ for some $h\in H^2$: if $\widehat{\phi\nu}=0$ on $[0,2\pi (2d-\e)]$ for some $\phi\in L^\infty(\nu)$, put $$h=\frac 1{2\pi i}(1-\theta)K\phi\nu.$$ By lemma \[sequences\], $D^*(C\setminus B)< d+\e$. Let $J$ be an inner function such that $\{J=1\}=C\setminus B$. By a version of the Beurling-Malliavin theorem, see [@MIF1] section 4.6, the kernel of the Toeplitz operator with the symbol $S^{2\pi (-d-\e)}J$ in $H^\infty$ is non-empty, i.e. there exists a function $g\in H^\infty(\C_+)$ such that $$S^{2\pi (-d-\e)}Jg\in \bar H^\infty.$$ Since $$\bar\theta S^{2\pi (-d-\e)}Jg S^{2\pi (2d-\e)}h=\bar\theta S^{2\pi (d-2\e)}Jgh\in \bar H^2,$$ we have $$S^{2\pi (d-2\e)}Jgh\in K_\theta.$$ Since $K_\theta$ is closed under division by inner components, $S^{2\pi (d-2\e)}gh\in K_\theta$ and therefore $$p=S^{2\pi (d-2\e)}Jgh-S^{2\pi (d-2\e)}gh=S^{2\pi (d-2\e)}(J-1)gh\in K_\theta.$$ By the Clark representation formula, $p=\frac1{2\pi i}(1-\theta)Kp\nu$, and since $1-\theta$ is outer, $Kp\nu$ is divisible by $S^{2\pi (d-2\e)}$ in $\C_+$. Notice that $p=(1-J)gh=0$ on $C\setminus B=\{J=1\}$ and $p\in L^\infty(\nu)$ on $B\cap C$. Therefore, if $\eta$ is the restriction of $\nu$ on $B\cap C$, the existence of such $p$ implies $$\GG^\infty_\eta\geqslant 2\pi (d-2\e).$$ For any $\e>0$, the measure $\eta$ constructed as above will have a bounded density with respect to $\mu$. Hence $\GG^\infty_\mu\geqslant 2\pi d.$ To prove the second part of the statement suppose that $\log(|n_B|+1)\in L^1_\Pi$ but $\GG^p_\mu>2\pi A> 2\pi D$ for some $A>D$. WLOG assume that the counting function $n_B$ is non-zero outside of $[-1,1]$ and define $$W(b_k)=\begin{cases}\frac1{2^{n}n_B(2^n)w(k)}+1\textrm{ if }b_k\in (2^{n-1},2^n]\textrm{ for some }n\in\N\\ -\frac1{2^{n}n_B(-2^n)w(k)}+1\textrm{ if }b_k\in (-2^{n},-2^{n-1}]\textrm{ for some }n\in\N\\ 1\textrm{ if }b_k\in (-1,1]\end{cases}.$$ Then $W$ is a $\mu$-weight and by lemmas \[t66\] and \[denlog\] there exists a measure $\nu=\sum \alpha_k\delta_{b_{n_k}}$ supported on $B'=\{b_{n_k}\}\subset B$ such that $W$ is a $\nu$-weight, $B'$ is an $A$-uniform sequence and $\alpha_k$ satisfy . Since $W$ is a $\nu$-weight, $|\alpha_k|\leq C/W(b_{n_k})$. Since $\alpha_k$ satisfy , the definition of $W$ implies that $$\sum_k\frac {\log w(n_k)}{1+{n_k}^2}>-\infty.$$ Hence $D\geqslant A$ and we obtain a contradiction. $\square$ Proof of theorem \[main\] ------------------------- $ $ I\) First, suppose that $\GG^p_\mu\geqslant a$ for some $1<p\leqslant \infty$. Then for any $d>0,\ 2\pi d<a$, there exists $f\in L^p(\mu)$ such that $\widehat{f\mu}=0$ on $[0,2\pi d]$. Let $W$ be a $\mu$-weight. Denote $V=W^{1/q}$ where $\frac1p+\frac1q=1$. Then $$\int V|f|d\mu<\infty.$$ Therefore by lemma \[t66\] there exists a discrete measure $\nu=\sum \alpha_n\delta_{\lan},\ \L=\{\lan\}\subset\supp\mu$ such that $\hat\nu=0$ on $[0,2\pi d]$, $V$ is a $|\nu|$-weight and $\nu$ satisfies the rest of the conditions of lemma \[denlog\]. Then by lemma \[denlog\], $\L$ contains a $d$-uniform subsequence $\L'$ and $\alpha_n$ satisfy . Since $V$ is a $\nu$-weight, $V(\lan)<C/|\alpha_n|$ for all $n$. It is left to notice that $\log W(\lan)=q\log V(\lan)$ and therefore $$\sum_{\lan\in \L'}\frac{\log W(\lan)}{1+\lan^2}<\infty.$$ II) Now suppose that $\GG^p_\mu< d<a$ for some $1<p\leqslant \infty$. Since $\GG^p_\mu< d$, by theorem \[Bakan\] there exists a $\mu$-weight $W$ such that finite linear combinations of exponentials from $\EE_{d-\e}$ are dense in $C_W$ for some $\e>0$. Suppose that there exists a $d$-uniform sequence $\L=\{\lan\}\subset \supp \mu,$ satisfying . Then by theorem \[mainDiscrete\] there exists a measure $\nu=\sum\alpha_n \delta_{\lan}$ such that $|\alpha_n|\leqslant W^{-1}(\lan)/(1+\lan^2)$ and $\hat\nu=0$ on $[0,d-\e]$. Then the finite measure $W\nu$ annihilates all functions $e^{ict}/W,\ c\in [0,d-\e]$. This contradicts completeness of $\EE_{d-\e}$ in $C_W$. $\square$ [24]{} Proc. Amer. Math. Soc., 136 (2008), no. 10, 3579�3589 Math. Scand., 55 (1984), 285–309 Mimeographed lecture notes, Summer institute, Stanford University (1961) Acta Math. 107 (1962), 291–302 Acta Math. 118 (1967), 79–93 , preprint, arXiv:1004.1795v1 Prentice-Hall, Englewood Cliffs, NJ, 1968 Trans. Amer. Math. Soc. 99 (1961), 118-152 American Journal of Mathematics, 67 (1945), 141�154. Linear and Complex Analysis Problem Book 3, Part II, Lecture Notes in Math., Springer, 1994, 87 – 88 Academic Press, New York, 1976 (Russian), Izvestiya Akad. Nauk SSSR, Ser. Mat., 15 (1951), 309�360; English translation in Amer. Math. Soc. Translation, Ser. 2, 1 (1955), 253�304. Springer-Verlag, Berlin, 1994. Algebra i Analiz, 12:6 (2000), 224�237 Cambridge Univ. Press, Cambridge, 1988 , Algebra i Analiz 10 (1998), 45�64; English translation in St. Petersburg Math. J. 10 (1999), no. 3, 441�455. Dokl. Akad. Nauk SSSR 46 (1945), 306–309 (Russian). Doklady Akad. Nauk SSSR (N.S.) 94, (1954), 13–16 (Russian). (Russian), Doklady Akad. Nauk SSSR (N.S.) 88 (1953), 405�408. AMS, Providence, RI, 1996 (Russian), Issled. Linein. Oper. Teorii Funktsii, 17, Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov (LOMI) 170 (1989), 102�156; English translation in J. Soviet Math., 63 (1993), no. 2, 171�201. AMS Colloquium Publications, 26 (1940) in [*Perspectives in Analysis*]{}, Springer Verlag, Berlin, 2005, 185–252 Invent. Math., Vol. 180, Issue 3 (2010), 443-480 Advances in Math., 224 (2010), pp. 1057-1070 , Dokl. Acad. Nauk SSSR 252, (1980), 1316–1320 , Springer-Verlaag, Berlin (1986) , Algebra i Analiz 5 (1993), no. 2, 189–210; translation in St. Petersburg Math.J. 5 (1994), no. 4, 389–408 , to appear in Acta Math., arXiv:0908.2079. Recent advances in operator-related function theory, 1–14, Contemp. Math., 393, Amer. Math. Soc., Providence, RI, 2006

--- abstract: 'Rationally null-homologous links in Seifert fibered spaces may be represented combinatorially via labeled diagrams. We introduce an additional condition on a labeled link diagram and prove that it is equivalent to the existence of a rational Seifert surface for the link. In the case when this condition is satisfied, we generalize Seifert’s algorithm to explicitly construct a rational Seifert surface for any rationally null-homologous knot. As an application of the techniques developed in the paper, we derive closed formulae for the rational Thurston-Bennequin and rotation numbers of a Legendrian knot in a contact Seifert fibered space.' address: - 'Institute for Advanced Study, Princeton, NJ 08540 & Australian National University' - 'Haverford College, Haverford, PA 19041' author: - 'Joan E. Licata' - 'Joshua M. Sabloff' bibliography: - 'main.bib' title: Rational Seifert Surfaces in Seifert Fibered Spaces --- Introduction ============ This paper studies rationally null-homologous links in Seifert fibered spaces, with the goal of extending techniques from classical knot theory to a more general setting. Previous work in this vein includes Gilmer’s signatures for rationally null-homologous links [@gilmer:ratl-link-cob] and Calegari and Gordon’s classification of knots with small rational genus [@cg:ratl-genus]. More generally, recent work on the Berge Conjecture has shown that the study of rationally null-homologous links is important for understanding Dehn surgery questions; see, for example, [@ras:berge]. Rationally null-homologous knots are also interesting in a contact geometric setting. For example, Baker and Etnyre generalized the definition of classical invariants for Legendrian knots to the case of rational homology three-spheres and classified rational Legendrian unknots [@be:rational-tb], and Cornwell has studied Bennequin-type inequalities in lens spaces [@cornwell:lens-invts]. Our interest in this topic was also prompted by contact geometry [@joan-josh:lch4sfs], but we hope the techniques developed in this paper will find applications within the wider context of the link theory in rational homology three-spheres. Just as a knot in $\mathbb{R}^3$ is often studied via the combinatorics of its planar projection, we consider the projection of a knot in a Seifert fibered space to its two-dimensional orbifold base. As we show in Section \[sect:sfs\], labeling this projection with some ancillary data permits the topological type of the knot to be recovered. Turaev initiated this “shadow” approach in the case of knots in an $S^1$ bundle over a surface, and the extension to $S^1$ bundles over orbifolds answers a question he posed in [@turaev:shadow]. After discussing labeled knot diagrams, we will introduce two further combinatorial objects: a is an assignment of an integer to each complementary components of the labeled knot projection, while a compatible is an assignment of integers to each quadrant around each double point of a labeled diagram. The precise definitions are given in Section \[sec:frss\] and allow us to state the following theorems: \[thm:label\] If $K$ is rationally null-homologous in a Seifert fibered space, then any labeled diagram for $K$ admits a formal rational Seifert surface with a compatible fiber distribution. \[thm:bounds\] If a labeled diagram for $K$ admits a formal rational Seifert surface with a compatible fiber distribution, then $K$ bounds a rational Seifert surface in $M$. It is clear that $K$ bounding a rational Seifert surface implies that $K$ is rationally null-homologous; thus, these two theorems also show that the existence of a formal rational Seifert surface with a compatible fiber distribution is equivalent to the geometric condition that $K$ is rationally null-homologous. A key construction in this paper is a generalization of Seifert’s algorithm for knots in $\rr^3$ to rationally null-homologous knots in Seifert fibered spaces. This algorithm, which provides the proof of Theorem \[thm:bounds\], explicitly constructs a rational Seifert surface in $M$ from the given combinatorial data. The algorithm is described in Section \[sec:seifert-algorithm\]. In the final section, we turn our attention to the special case of a Legendrian knot in a Seifert fibered space equipped with a transverse, $S^1$-invariant contact structure. (This setting was studied in more detail in [@joan-josh:lch4sfs].) As an application of the algorithm defined in Section \[sec:seifert-algorithm\], we compute the rational classical invariants of a Legendrian knot from its labeled diagram; this result generalizes the familiar formulae for classical invariants in the standard contact $\mathbb{R}^3$. \[prop:initial-leg\] Let $K$ be a rationally null-homologous Legendrian knot in a contact Seifert fibered space. The rational rotation number of $K$ may be computed directly from a formal rational Seifert surface, and the rational Thurston-Bennequin number of $K$ may be computed directly from a compatible fiber distribution. See Proposition \[prop:legendrian\] for a more precise statement. Labeled diagrams ================ Background {#sect:sfs} ---------- We view Seifert fibered spaces as $S^1$ bundles over two-dimensional orbifolds, following the notational conventions of [@lisca-matic:transverse; @massot]. Let $\Sigma'$ be an oriented surface, possibly with boundary, with $r+1$ discs removed from its interior. Orient the new components of $\partial \Sigma'$ as the boundary of the missing disc, and let $M' =\Sigma' \times S^1$. The first homology groups of the boundary tori of $M'$ are generated by classes $\langle m_i, \ell_i \rangle$, with $\cup_i m_i=[\partial \Sigma'\times \{\text{pt}\}]$ and $\ell_i = [\{\text{pt}\} \times S^1]$, oriented so that $m_i \cdot \ell_i = 1$. Note that this orients all the fibers in $M$. For $1\leq i\leq r$, let $\alpha_i$ and $\beta_i$ be relatively prime integers satisfying $0 < \beta_i <\alpha_i$. Glue a solid torus $W_i$ to the $i^{th}$ boundary component of $M'$ so that the image of a meridian represents the homology class $\alpha_i m_i+ \beta_i \ell_i$. To the remaining boundary component, glue a solid torus so that the meridian is sent to a curve representing the class of $m_0 + b \ell_0$. The fiber structure on the boundary of $M'$ extends uniquely to a fiber structure on the interior of the surgery solid tori, and the resulting identification space $M$ is said to have $(g,b; (\alpha_1, \beta_1), \ldots, (\alpha_r, \beta_r))$. Note if $\Sigma'$ has boundary, then $M$ has an $S^1$-fibered boundary. Every orientable Seifert fibered space with an orientable fiber space can be realized via this construction; given two Seifert invariants, it is straightforward to determine whether they correspond to the same Seifert fibered manifold [@orlik]. The of a Seifert fibered space with Seifert invariants $(g,b; (\alpha_1, \beta_1), \ldots, (\alpha_r, \beta_r))$ is the rational number $$e(M)=-b-\sum_{i=1}^r \frac{\beta_i}{\alpha_i} .$$ Labeled Diagrams {#sec:labeled-diagrams} ---------------- Let $L$ be an oriented link in a Seifert fibered space $M$ with Seifert invariants $(g,b; (\alpha_1, \beta_1), \ldots, (\alpha_r, \beta_r))$. We suppose throughout that $L$ is everywhere transverse to the fibers, and we let $(\Sigma, \Gamma_L)$ denote the image of $(M, L)$ under the quotient map $\pi$ which sends each fiber to a point. In order to recover the isotopy class of $L$ from this projection, we will use a ; this notion was introduced in [@s1bundles] and is similar to Turaev’s notion of a shadow for a link in a circle bundle [@turaev:shadow]. The fiber over a double point of $\Gamma_L$ is separated by its intersections with $L$ into two oriented chords, and we systematically select a preferred chord at each crossing. Near a crossing, there is a unique quadrant which is coherently and positively oriented by $L$. Declare this quadrant and the opposite quadrant to be , and declare the adjacent quadrants to be . When the oriented boundary of a positive (respectively, negative) quadrant is lifted to segments of $K$ connected by a chord, the preferred chord is the one traversed positively (negatively). Given a region $R$ in $\Sigma\setminus \Gamma_L $, define $M_R$ to be the restriction of the orbibundle $M \to \Sigma$ to $R$. Let $A_R$ be the least common multiple of the orders of the orbifold points in $R$; if $R$ contains no orbifold points, set $A_R = 1$. The subcurves of $L$ which project to $\partial R$ may be concatenated with the preferred chords over the corners of $R$ to yield a closed curve $L_R$ in $\partial M_R$; orient $L_R$ so that the orientation induced by its projection to $\Sigma$ agrees with that of $\partial R$. Let $\widetilde{R}$ be the the $A_R$-fold branched covering $\widetilde{R}$ of $R$. Use the covering map to pull back the bundle $M_R$ to $\widetilde{R}$. This lifts $L_R$ to a closed $1$-manifold $\widetilde{L}_R$ in an honest $S^1$ bundle over $\widetilde{R}$. Let $$\gamma_1 \times \cdots \times \gamma_{k_R}: S^1 \sqcup \cdots \sqcup S^1 \rightarrow S^1\times \partial \widetilde{R}$$ denote the map whose image is $\widetilde{L}_R$. Choose a trivialization of $S^1\times \widetilde{R}$ and let $\iota:S^1\times \partial \widetilde{R} \hookrightarrow S^1\times \widetilde{R}$ denote the inclusion. Finally, let $p: S^1\times \widetilde{R}\rightarrow S^1$ be projection to the first factor. Given a region $R$, the $n(R)$ of the region is $$n(R)=\frac{1}{A_R} \sum_{i=1}^{k_R} \operatorname{deg}(p \circ \iota \circ \gamma_i).$$ It is immediate from the definition that $n(R)=0$ if and only if the (multi)curve $\widetilde{L}_R$ bounds a section of the $S^1$ bundle. In fact, this implies that the defect is independent of the chosen trivialization. It follows from this definition that the defect is additive on regions. When $R$ contains no orbifold points, then the defect $n(R)$ is an integer; in general, the defect contains information about the Euler number of $M_R$. \[lem:orbi-defect\] The difference between the defect $n(R)$ and the Euler number $e(M_R)$ is an integer, i.e. $n(R) - e(M_R) \in \zz$. Recall that each exceptional fiber $F'$ can be viewed as the core of a solid torus where Dehn surgery was performed on some regular fiber $F$. Let $K$ be a loop bounding a meridional disc in a regular neighborhood of $F$. After performing $(\alpha, \beta)$ surgery, $K$ intersects a meridian of the surgered torus $-\beta$ times, so the defect of the region bounded by $K$ is $\frac{-\beta}{\alpha}$. Now let $K_1, \ldots, K_l$ be small loops in $R$ around the $l$ exceptional fibers in $M_R$. The multicurve $\partial R \bigcup (\cup K_i)$ bounds a region with no orbifold points, and hence has an integral defect $d$. Since the defect is additive, we see that $$n(R) = d - \sum_{i=1}^l \frac{\beta_i}{\alpha_i} = d' + e(M_R)$$ for some $d' \in \zz$. We say that a diagram $(\Sigma, \Gamma_L)$ is when it is decorated with a defect in each region and with the fiber invariants associated to each orbifold point. Abusing notation, we will refer to both the projection and the labeled diagram by $\Gamma_L$. Isotopy of the link changes the labeled diagram in one of several ways. Figure \[fig:moves\] shows labeled Reidemeister moves for links in a Seifert fibered space; these correspond to isotopies of $L$ in the complement of the exceptional fibers. When a strand of $L$ passes through an exceptional fiber of type $(\alpha, \beta)$ the labeled diagram changes by a move which wraps $\Gamma$ around the orbifold point $\alpha$ times. See Figure \[fig:tearlabels\]. In order to label the new regions created by a teardrop, we assume that the isotopy occurs in an arbitrarily small neighborhood of the exceptional fiber. The defect is therefore completely determined by the preferred chords at the new crossings. We may choose a local metric on the solid torus over the neighborhood of an orbifold point so that each regular fiber has length $1$ and the exceptional fiber has length $\frac{1}{\alpha}$. With such a choice, the chords created by the teardrop have lengths in the set $\{ \frac{1}{\alpha}, \frac{2}{\alpha}, \hdots \frac{\alpha-1}{\alpha}\}$. The defect of a region is the signed sum of the lengths of the chords assigned to its corners, where the sign is positive at coherent corners and negative otherwise. Since the innermost region of the teardrop has a coherent corner, it follows from Lemma \[lem:orbi-defect\] that the defect of this region is $\frac{\alpha-\beta}{\alpha}$. The defects of the other regions are determined by the signs of the corners and the requirement that the defect of any region not containing an orbifold point is integral. We say that two labeled diagrams are if they differ only by sequence of surface isotopies in $\Sigma$, labeled Reidemeister moves, or labeled teardrop moves. The discussion above, together with the classical Reidemeister theorem, establishes the following lemma: \[prop:diagram-isotopy\] If two generic links in $M$ are isotopic, then their labeled diagrams are equivalent. Although it is possible to define an inverse for the teardrop move, we present it as unidirectional; passing the innermost strand of the teardrop back across the fiber introduces a second teardrop, and a sequence of Reidemister II moves returns a projection isotopic to the original one. See Figure \[fig:octo\] for an example. Next, we define a diagram move that preserves $\Gamma$ but alters the defects of a pair of adjacent regions. Following Turaev, we say is the operation that replaces an oriented segment of $L$ with a segment that has the same projection but travels once around the fiber. We define an action of $H_1(\Sigma)$ on the set of oriented links $\mathcal{L}(M)$ as follows: let $\gamma$ be a generic simple closed curve on $\Sigma$ that represents a class $[\gamma] \in H_1(\Sigma)$; in particular, we assume that $\gamma$ intersects $\pi(L)$ transversely in finitely many points and misses the double points of $\pi(L)$ and the orbifold points of $\Sigma$. Construct the link $\gamma \cdot L$ by performing fiber fusion on $L$ in a neighborhood of each point of $\gamma \cap \pi(L)$, where the sign of intersection dictates the sign of the fusion. \[lem:h1-action\] The isotopy type of the link $\gamma \cdot L$ depends only on the homology class $[\gamma]$. The proof is the same as that in [@turaev:shadow]. We note that the labeled diagrams associated to $L$ and to $\gamma\cdot L$ have the same defects; this follows from the fact that for each region $R$, the closed loop $\gamma$ intersects $\partial R$ zero times algebraically. Consequently, a labeled diagram of genus greater than zero cannot determine an isotopy class of link. We show next that each labeled diagram corresponds to an equivalence class of links related by this $H_1(\Sigma)$ action. Let $\bar{\alpha}=(\alpha_1, \alpha_2, \hdots \alpha_k)$ be a list of the orders of the orbifold points on $\Sigma$. Pick a list $\bar{\beta}=(\beta_1, \beta_2, \hdots \beta_k)$ such that $(\alpha_i, \beta_i)$ are relatively prime and $1\leq \beta_i<\alpha_i$. Let $\mathcal{D}(\Sigma,q,\bar{\alpha}, \bar{\beta})$ denote the set of labeled diagrams whose defects sum to $q$ and satisfy Lemma \[lem:orbi-defect\] in each region. \[thm:realization\] Let $M$ be a Seifert fibered space with exceptional fiber invariants $\{ (\alpha_i, \beta_i)\}$. There is a bijective correspondence between the set $\mathcal{D}(\Sigma, e(M), \mathbf{\alpha}, \mathbf{\beta})$, up to equivalence, and the set $\mathcal{L}(M)$, up to isotopy and the action of $H_1(\Sigma)$. In the absence of exceptional fibers, we note that this result follows from a theorem of Turaev which establishes a bijection between his “shadow links” and isotopy classes of links in $M$, up to the action of $H_1(\Sigma)$. To see the theorem in this special case, we describe a bijection between labeled diagrams and shadow links. Let $R$ be a region of $\Sigma \setminus \Gamma$ with $p(R)$ positive corners and $q(R)$ negative corners with respect to the preferred chords. In the notation of [@turaev:shadow], $\alpha=2n(R)-p(R)+q(R)$ and $\beta=p(R)+q(R)$. It follows that the “gleam” of $R$ is $p(R)-n(R)$. As a first step, we show for a given labeled diagram in $\mathcal{D}(\Sigma, e(M), \mathbf{\alpha}, \mathbf{\beta})$, one may always find a link $L$ realizing this diagram. Fix a labeled diagram $(\Sigma, \Gamma) \in \mathcal{D}(\Sigma, e(M), \mathbf{\alpha}, \mathbf{\beta})$. One may easily find a link $L$ in $M$ which projects to $\Gamma$, and by Lemma \[lem:orbi-defect\], the defect of any region will differ from the Euler number of the bundle over that region by an integer. We induct on the number of crossings to show that $L$ may be modified so that its defect in each region agrees with the given label. For the base case, consider a diagram consisting of a collection of disjoint embedded circles. Selecting an arbitrary component of $\Sigma \setminus \Gamma$ to be “outermost” gives a partial order on the components of $\Gamma$. Perform fiber fusions on the curves of $L$ which project to the boundary of any innermost region in order to adjust its defect to the given label. Proceed outward, region by region. Upon reaching the outermost region, there will be no free edges available for fiber fusion, but since each fusion operation preserves the sum of the labels, the defect of the outermost region will automatically agree with the given label. Now suppose that for any labeled diagram with fewer than $n$ crossings, we can find a knot $L\subset M$ whose defects agree with the labels. Let $(\Sigma, \Gamma)\in \mathcal{D}(\Sigma, e(M), \mathbf{\alpha}, \mathbf{\beta})$ have $n$ crossings. Resolve one crossing so as to preserve the orientation of $\Gamma$ and apply the inductive hypothesis to construct a link $L'$ whose defects agree with the labels. Replacing the crossing splits one region into two pieces, and Figure \[fig:adjust\] indicates how to perform fiber fusions to construct the desired $L$. As in [@turaev:shadow], the remainder of the proof of Theorem \[thm:realization\] follows from two further steps. The first step is showing that any two isotopy classes of links which correspond to the same labeled diagram are related by the action of $H_1(\Sigma)$. The second step establishes that two generic links corresponding to equivalent labeled diagrams are related by a sequence of fiber fusions and isotopies. Turaev’s arguments apply with little modification to both cases; in the second case, we additionally note that any teardrop move on labeled diagrams can be realized by a local isotopy of the link across an exceptional fiber. Combinatorics for Rational Seifert Surfaces {#sec:frss} =========================================== In this section, we develop a combinatorial description of a rational Seifert surface for a rationally null-homologous knot $K$. The description has the form of two decorations of the labeled diagram $\Gamma_K$ of $K$: a “formal rational Seifert surface” and a compatible “fiber distribution”. The two decorations will be used in the next section to describe a generalization of the Seifert algorithm. Two Decorations of Labeled Diagrams ----------------------------------- A surface in a Seifert fibered space is said to be horizontal if it is everywhere transverse to the fibers; we relax this condition slightly and consider rational Seifert surfaces which are transverse except near fibers over double points of $\Gamma$. The idea of the first decoration is that any such surface assigns a multiplicity to each region $R$. Conversely, we may characterize the sets of multiplicities on $\Sigma$ which are induced by such a surface using the following combinatorial object: \[def:rfss\] A is an assignment of an integral multiplicity $m(R_j)$ to each region $R_j$ of $\Sigma\setminus \Gamma$ which satisfies the following conditions: 1. \[except\] The least common multiple $A_{R_j}$ of the orders of the orbifold points in $R_j$ divides $m(R_j)$; 2. \[order\] if $R_k$ and $R_l$ share an edge oriented as $\partial R_k$, then $$m(R_k)-m(R_l)=r;$$ 3. \[total\] summing over all regions, $$\sum_j m(R_j) n(R_j)=0.$$ A formal rational Seifert surface may be viewed as a secondary labeling on a knot diagram, and we introduce a tertiary labeling as well. Let $x_j^i$ denote a corner of the region $R_j$ at the $i^{th}$ crossing. (It is possible for a single region to fill more than one corner at a given crossing, but for notational convenience, we avoid introducing a third index to distinguish them.) \[defn:fiber-dist-g\] Given a formal rational Seifert surface $\mathbf{m}$ for a labeled diagram $\Gamma$, a compatible with $\mathbf{m}$ is an assignment $\mathbf{f}$ of integers $f(x_j^i)$ to the corners of regions of $\Sigma \setminus \Gamma$ which satisfies the following properties: 1. \[flatdiscg\] for each region $R_j$ with corners $x^i_j$ for $i\in \mathcal{C}_R = \{i_1, \ldots, i_{k_R}\}$, $$m(R_j) n(R_j)+\sum_{i \in \mathcal{C}_R} f(x_j^i)=0;$$ 2. \[welllabg\] for each crossing labeled $i$ with incident regions $R_{j_1}, \ldots, R_{j_4}$, $$\sum_{k=1}^4 f(x_{j_k}^i) =0.$$ Rational formal Seifert sufaces and their fiber distributions are best understood in terms of a special cell decomposition of $M$, which is constructed in Section \[sec:cell\]. As motivation, however, one may view the rational formal Seifert surface as describing how a surface interacts with the base orbifold $\Sigma$, whereas a fiber distribution captures its interaction with the bundle structure of $M$. \[ex:1\] The figure shows a labeled diagram for a knot in $L(5,2)$, together with a rational formal Seifert surface and fiber distribution. We will use this example to illustrate the generalized Seifert algorithm in Section \[sec:seifert-algorithm\]. A cell decomposition for $M$ {#sec:cell} ---------------------------- In this section, we construct a cell decomposition of $M$ using data from the knot $K$. We begin by enlarging the graph $\Gamma_K$ so that each complementary region is homeomorphic to a disc and contains at most one orbifold point. If a region has nontrivial topology or contains more than one orbifold point, subdivide it using a collection of arcs $\Gamma_0 \subset \Sigma$ whose endpoints lie on $\Gamma_K$; let $\bar{\Gamma}$ denote the graph $\Gamma_K \cup \Gamma_0$. Lift the arcs of $\Gamma_0$ to curves $K_0$ in $M$ whose endpoints lie on $K$. The knot $K$, the arcs $K_0$, and the fibers over each vertex of $\bar{\Gamma}$ form a $1$-complex in $M$. The $2$-skeleton of $M$ consists of two types of cells. First, for each edge $e$ of $\bar{\Gamma}$, let $D_e$ be the preimage of $e$ in $M$, thought of as a disc whose boundary consists of the fibers over the ends of the edge, together with two oppositely-oriented copies of the corresponding segment of $K \cup K_0$. Refer to this type of cell as . Second, for each region $R$ of $\Sigma \setminus \bar{\Gamma} $, we construct the cell $D_R$ as follows. Denote the fibers over double points in $\partial R$ by $\{F_i\}$. The lifted curve $K_{R}$ satisfies $[A_R K_{R}-\sum b_i F_i]=0\in H_1(M)$ for any $b_i$ such that $\sum b_i=A_R n(R)$. The $1$-chain $A_R K_{R}-\sum b_i F_i$ bounds a disc in $M_R$, and we include this as the $2$-cell $D_R$. The remainder of $M$ consists of $3$-balls that come from removing a meridian disc from the solid tori over each region of $\bar{\Gamma}$; these balls make up the $3$-skeleton. Proof of Theorem \[thm:label\] ------------------------------ Recall the statement of Theorem \[thm:label\] from the introduction: If $K$ is rationally null-homologous in a Seifert fibered space, then any labeled diagram for $K$ admits a formal rational Seifert surface with a compatible fiber distribution. Suppose that $K$ is rationally null-homologous with order $r$. The knot $K$ has an obvious representative (which we shall also call $K$) as a $1$-chain in the cell decomposition described above. Hence, there exists a $2$-chain $S$ such that $\partial S = rK$. For each region $R_j\in \Sigma\setminus (\bar{\Gamma} )$, let $c_j$ denote the coefficient of $D_j$ in $S$. Assign the multiplicity $m(R_j)$ to be $c_j\alpha_{R_j}$. We begin by verifying Condition \[except\] of Definition \[def:rfss\]. It is clearly satisfied on disc components of $\Sigma \setminus \bar{\Gamma}$. Now suppose that $R_1$ and $R_2$ in $\Sigma\setminus \bar{\Gamma})$ are separated by the edge $e_0\in \Gamma_0$. The assumption that $\partial S=rK$ implies that this edge has multiplicity $0$ in $\partial S$, so $m(R_1)=m(R_2)$. This shows that the multiplicities are well-defined on components of $\Sigma \setminus \Gamma_K$. Since $\alpha_j \ | \ m(R_j)$ for $j=1,2$ and $m(R_1)=m(R_2)$, Condition \[except\] is satisfied on $R_1\cup R_2$, and an inductive argument shows that it holds for all components of $\Sigma \setminus \Gamma_K$. Adding a vertical $2$-cell to a chain does not change the coefficient of any edge of $K$ in the boundary $1$-chain. Each edge of $K$ appears $r$ times in $\partial S$, so the difference in multiplicities between the two adjoining regional cells is $r$, establishing Condition \[order\]. Finally, we show that Condition \[total\] of Definition \[def:rfss\] holds. By construction, the boundary of each regional cell consists of $A_{R_j}K_{R_j}$ and $-A_{R_j} n(R_j)$ copies of the fiber. Thus the total number of copies of the fiber coming from regional $2$-cells is $\sum_j -m(R_j)n(R_j)$. The addition of any vertical $2$-cell preserves this sum, and the assumption that $\partial S=rK$ implies that the copies of the fiber must cancel algebraically: $\sum_j m(R_j)n(R_j)=0$. To construct a compatible fiber distribution $\mathbf{f}$, consider a quadrant $x^i_j$ of a crossing $i$ lying in the region $R_j$. Suppose that this quadrant lies to the right of the oriented edges $\mathcal{E}(x^i_j)$ of $\bar{\Gamma}$; note that this set may be empty and has at most two elements. We then define $f(x^i_j)$ to be $$f(x^i_j) = -c_j b_i + \sum_{e \in \mathcal{E}(x^i_j)} \epsilon_e f_e,$$ where the integer $b_i$ comes from the construction of the regional cell $D_{R_j}$, $f_e$ is the coefficient of the vertical cell $D_e$ in $S$, and $\epsilon_e$ is positive if and only if the head of $e$ is incident to the double point $i$. Condition \[flatdiscg\] of Definition \[defn:fiber-dist-g\] now follows from two facts. First, observe that $c_j \sum b_i = c_j A_{R_j} n(R_j) = m(R_j) n(R_j)$. Second, note that each edge with $R_j$ on its right contributes $f_e$ to the sum associated to the quadrant at its head and and $-f_e$ to the sum associated to the quadrant at its foot; thus, the contributions coming from the vertical $2$-cells cancel around any given region. Condition \[welllabg\] holds because $\sum_{k=1}^4 f(x_{j_k}^i)$ is the coefficient of the fiber over the double point $i$ in $\partial S$, but we know that $\partial S = rK$, and hence this coefficient must vanish. One may show that every formal rational Seifert surface admits a compatible fiber distribution, a fact which permits a stronger formulation of Theorem \[thm:bounds\]. The proof is by induction on the number of double points of $\Gamma$, and we leave the details to the reader. Seifert algorithm for knots in $S^1$ orbifold bundles {#sec:seifert-algorithm} ===================================================== Given a formal rational Seifert surface $\mathbf{m}$ and a fiber distribution $\mathbf{f}$ for a rationally null-homologous knot of order $r$, we construct a rational Seifert surface of the same order. The classical Seifert algorithm for knots in $\rr^3$ proceeds in three steps: first, one resolves the crossings in a projection of the knot to obtain a collection of Seifert circles in the plane. Second, one views the Seifert circles as bounding disjoint embedded disks. Finally, the Seifert disks are connected by twisted bands at the crossings. The generalized algorithm for a knot in a Seifert fibered space parallels the classical algorithm. As a first step, we let $D_i$ denote a neighborhood of the $i^{th}$ double point of $\Gamma$ and let $U_i = \pi^{-1}(D_i)$. We use **m** and **f** to resolve the knot into circles in $M \setminus \bigcup U_i$ (Section \[ssec:seifert-circle\]). Next, we view these resolved circles as bounding embedded surfaces in $M \setminus U_i$ (Section \[ssec:seifert-surfaces\]). Finally, we extend these surfaces across the solid tori $U_i$ (Section \[ssec:crossing-construction\]). We begin by establishing notation which will be useful throughout the algorithm. For each double point of $\Gamma$, parameterize the neighborhood $D_i$ as a unit disc and let $C^i_t$ denote the $S^1$ bundle over the circle of radius $t$. Dropping the superscript when the crossing is obvious, we split the torus $C_1$ into annuli denoted $A_{I}$, $A_{II}$, $A_{III}$, and $A_{IV}$ according to the corresponding quadrants of $\Sigma$; see Figure \[fig:notation\]. Let $K_0$ be the curves constructed in Section \[sec:cell\]. Near $K \cup K_0$ but away from the double points of $\bar\Gamma$, the local behavior of any rational Seifert surface is dictated by the multiplicities of the adjacent regions; note that the multiplicities on regions of $\Gamma$ induce multiplicities on the regions of $\bar\Gamma$. Let $N$ be a regular neighborhood of $K \cup K_0$, and suppose that the projection of a segment of $K \cup K_0$ separates regions with multiplicities $m_1$ and $m_2$. In this case, the rational Seifert surface $S$ intersects $\partial N$ $m_i$ times on each side. Correspondingly, to each side of a cross section of $N$ we draw $m_i$ parallel, transversely-oriented lines. The endpoints of these lines trace out $m_i$ parallel curve segments on $\partial N$ as the cross-section varies; see Figure \[fig:notation\]. Resolution into Seifert Circles {#ssec:seifert-circle} ------------------------------- The first step of the construction replaces $K$ with a collection of circles. Remove the interior of $N$ and the fibered solid tori $U_i$ from $M$. As described above, the portions of $\partial N$ away from the $U_i$ and neighborhoods of the intersection points $K \cap K_0$ are decorated with collections of parallel curves. Near the intersection points $K \cap K_0$, we simply join the endpoints of corresponding parallel curves. Near the solid tori $U_i$, we will use **m** and **f** to construct a pattern of curves on $C^i_1$ which connect the endpoints of the parallel curves. Fix a crossing, and for convenience, cut the corresponding solid torus along a meridional disc so that $C_1$ becomes a cylinder composed of four rectangles still labeled by I, II, III, and IV. Orienting each rectangle as if viewed from $t>1$, decorate it with a pattern of multicurves as shown in Figure \[fig:basicpattern\]. Each curve is decorated with an arrow indicating its transverse orientation and by an integer weight indicating its multiplicity. Reversing the arrrow changes the sign of this weight. By construction, the endpoints of these curves can be glued to the endpoints of the curves on $\partial N$. The resulting pattern of curves on $\partial \big(N \cup (\cup_iU_i)\big)$ will serve as our Seifert circles. Before proceeding, we note the following: \[lem:cancel\] The sum of the algebraic intersection numbers of the pattern curves with the meridian of $C_1$ is zero around each double point. This follows from Condition \[welllabg\] of Definition \[defn:fiber-dist-g\]. Surfaces Bounded by Seifert Circles {#ssec:seifert-surfaces} ----------------------------------- We begin the second step by constructing surfaces in $M\setminus \big( N \cup (\cup_iU_i)\big)$ bounded by the Seifert circles. Condition \[flatdiscg\] of Definition \[defn:fiber-dist-g\] implies that all $\frac{m(R)}{A_R}$ Seifert circles over the boundary of a given region $R$ are null-homologous and hence bound horizontal embedded discs in $M_R$. By construction, the signed intersection number of each curve pattern with $\partial N \cap C_1$ is $r$; see Figure \[fig:notation\]. To complete this step, we extend this surface over the cylinders $N \cap \big(M \setminus \cup U_i\big)$. There are two cases to consider for the extension over such a cylinder. If the multiplicities of the adjoining regions have the same sign, then we extend the embeddings of the surfaces as in Figure \[fig:nbhd-extend\](a) for an appropriate choice of $k,l \geq 0$. In particular, if the regions in question are separated by an edge of $\pi(K_0)$, then the multiplicities of the adjacent regions are the same and we use $k+l = m_-$ in the figure. If, on the other hand, the multiplicities of the adjoining regions have opposite signs, then the extension is as in Figure \[fig:nbhd-extend\](b); in this case, there is no choice to make. Extending across solid tori over crossings {#ssec:crossing-construction} ------------------------------------------ We have now constructed a surface in the complement of the crossing tori $U_i$. In this section, we extend the surface across each $U_i$ by describing how it intersects a collection of concentric cylinders $C_t$ of decreasing radius. Modifications to the intersection pattern describe changes in the surface. In addition to surface isotopy of the curves, we allow the following three primitive moves: Finger Moves : We may replace a curve segment adjacent to $C_t\cap N$ with a pair of arcs ending on $C_t\cap\partial N$; these intersections will have opposite signs. This move preserves the topology of the surface, but pushes it locally into the neighborhood of $K$. See Figure \[fig:extension-choice\]. Capping a Circle : Any embedded circle may be removed from the intersection pattern. This corresponds to capping off the corresponding component of $S\cap C_{t_0}$ with a disc embedded in the solid torus defined by $t<t_0$. Saddle Moves : We may perform a saddle resolution between two curves with opposite transverse oreintations. This corresponds to reducing the Euler characteristic of the surface by $1$. We will also make use of two consequences of these three moves. Cancellation of Parallel Strands : Two oppositely-oriented adjacent parallel strands between components of $N \cap C_t$ may be removed. See Figure \[fig:derived\]. Reconfiguration in $N$ : Any two configurations that appear in Figure \[fig:nbhd-extend\] are related by a sequence of saddle moves. See Figure \[fig:derived\]. We now begin to extend the surface $S$ across the solid torus $U_i$. Isotope all the intersections of the Seifert circles to the annulus $A_{II}$. Fixing these intersections, standardize the pattern of curves on $C_t$ via isotopy, finger moves, and cancellations of oppositely-oriented parallel strands. Note that after cancellation, the configurations inside $N \cap C_t$ are again of the form in Figure \[fig:nbhd-extend\]. Lemma \[lem:cancel\] states that the algebraic intersection number of these curves with the meridian is zero, and saddle resolutions between oppositely-oriented curves reduce the geometric intersection number to zero as well. The resulting pattern may contain curves with both endpoints on the same component of $N \cap C_t$; these may be again be removed using sequences of the moves above, especially capping circles. As $t\rightarrow 0$, the strands of $K$ cross; this rotates a region containing two components of $N \cap C_t$ by $\pi$. Further finger moves, cancellations, and isotopies yield a standard pattern consisting solely of horizontal curves. It is clear that these bound a collection of discs, completing $S$. Note that reconfigurations inside $N$ allow us to match those configurations coming from opposite sides of the intersection of one component of $N \cap U_i$. See Figure \[fig:torusexample2\] for an example. Legendrian invariants {#sec:legendrian} ===================== In this section we use the use the generalized Seifert algorithm to compute the rational classical invariants for a Legendrian knot from a formal rational Seifert surface and fiber distribution. Contact Seifert fibered spaces {#sec:legendrian:intro} ------------------------------ We will use the phrase to denote an orientable Seifert fibered space over an orientable base, equipped with a contact structure $\xi$ transverse to the Seifert fibers. Such a contact structure exists whenever the rational Euler number of a Seifert fibered space is negative [@kt:cr-seifert; @lisca-matic:transverse]. If we further specify a contact form $\alpha$ for $\xi$ with the property that its Reeb field points along the fibers (see [@joan-josh:lch4sfs]), then the defect defined in Section \[sec:labeled-diagrams\] can be interpreted as an integral of the curvature form associated to $\alpha$ on the Reeb orbit space. We note that the Legendrian condition precludes the Reidemeister I move of Section \[sec:labeled-diagrams\]. A formal rational Seifert surface $\mathbf{m}$ and a compatible fiber distribution $\mathbf{f}$ may be used to compute may be used to compute the rational classical invariants of a Legendrian knot in a contact Seifert fibered space. We prove this using the rational Seifert surfaces constructed in Section \[sec:seifert-algorithm\]. For each region $R_j\in \Sigma\setminus \Gamma$, let $\chi_{orb}(R_j)$ denote the orbifold Euler characteristic of $R_j$ as a sub-orbifold of $\Sigma$; recall that this quantity is defined to be: $$\label{eq:orb-euler-char} \chi_{orb}(R) = \chi(R) + \sum_{j=1}^r \left( \frac{1}{\alpha_j} - 1 \right).$$ Let $k_j$ and $l_j$ denote the number of double points of $\Gamma$ where $R_j$ fills one or three quadrants, respectively. We restate Proposition \[prop:initial-leg\] as follows: \[prop:legendrian\] The rational classical invariants of a null-homologous Legendrian knot $K$ maybe be computed from a formal rational Seifert surface $\mathbf{m}$ and a compatible fiber distribution $\mathbf{f}$ using the following formulae: $$\begin{aligned} \label{eq:r} \text{rot}_{\mathbb{Q}}(K)&=\frac{1}{r}\sum_{\text{regions }R_j} m(R_j) \big[\chi_{orb}(R_j)+\frac{1}{4}(l_j-k_j)\big], \\ \label{eq:tb} \text{tb}_{\mathbb{Q}}(K)&=\frac{1}{r}\sum_{\text{dble pts }i} (-r-f^i_{II}+f^i_{IV}).\end{aligned}$$ The subsequent sections discuss these invariants and develop proofs of these propositions. The knot in Example \[ex:1\] can be realized as a Legendrian knot whose Lagrangian projection is shown in Figure \[fig:lensex\]. To see this, begin with the unknot with maximal Thurston-Bennequin number in the standard contact $S^3$. Performing $1$ and $\frac{1}{2}$ surgery on a pair of regular fibers yields the labeled diagram of Figure \[fig:lensex\], and the contact form may be extended across the surgery tori so that the induced Reeb orbits are the Seifert fibers. The results above show that this knot has rational rotation number $$rot_\qq(K)= \frac{1}{5}\left[6(\frac{1}{4})+1(\frac{1}{2})-4(\frac{3}{4})\right] =-\frac{1}{5}$$ and rational Thurston-Bennequin number $$tb_\qq (K)=\frac{1}{5}(-5-3-4)=\frac{-12}{5}.$$ The rational rotation number ---------------------------- In [@be:rational-tb], Baker and Etnyre define the rational rotation number of a rationally null-homologous knot by analogy with the classical rotation number for a null-homologous knot. Let $j:S\hookrightarrow M$ be a rational Seifert surface for $K$. Trivialize the pulled back contact bundle $j^*\xi$ over $S$ using a nonvanishing vector field $v$; since $K$ is Legendrian, $TK$ lies in the restriction of $\xi$ to $\partial S$. One may therefore define the winding number of $j^*TK$: $$\text{rot}_{\mathbb{Q}}(K)=\frac{1}{r} \text{wind}_V(j^*TK).$$ To better understand a trivialization of $j^*\xi$, we will cut $S$ along its intersection with the vertical tori $\partial U_i$. This creates a collection of disjoint surfaces with boundary, denoted collectively by $\hat{S}$; we compute the rotation of each component individually and sum them to compute the rational rotation number of $K$. Note that cutting introduces new segments to the boundary curves; although these could be isotoped to be Legendrian, their contributions to the rotation will cancel under gluing. We may therefore ignore these segments and compute only the contributions to the rotation number of $T(\partial \hat{S})$ by $TK$. We begin by showing that the contribution of a component $X$ of $\hat{S}$ lying in the solid torus $U_i$ to $\text{rot}_{\mathbb{Q}}(K)$ is zero. We may assume that the complex structure on $\Sigma$ is chosen so that the arcs of $\Gamma$ intersect the boundary of the neighborhood of the double point orthogonally. Choosing the neighborhood of a fixed double point small enough, we may trivialize $T\Sigma$ over the disc $D_i$ with vector fields $\{v, iv\}$ so that $TK$ never coincides with the lines spanned by $v$ and $iv$. Pull back this trivialization to $\xi|_{U_i}$, and then again to $j^*\xi|_{X}$. With respect to this trivialization, it is obvious that $K \cap U_i$ contributes zero to the rotation number. We now turn to the portions of $S$ constructed from Seifert circles in Section \[ssec:seifert-surfaces\], i.e., the components of $j(S) \cap (M_R \setminus \bigcup U_i)$. Recall that these components of $\hat{S}$ are horizontal, and hence that we may identify $TS$ and $j^*\xi$ on these portions. The next lemma extends the existing trivialization of $j^*\xi$ from $j(S) \cap \partial U_i$ and describes the contribution to $\text{rot}_{\mathbb{Q}}(K)$ coming from a single region $R$. Suppose that the region $R$ has multiplicity $m(R)$ in a formal rational Seifert surface for $K$, and that $k$ and $l$ donote the number of double points in $\partial R$ where $R$ fills one and three quadrants, respectively. The contribution of $j(S) \cap (M_R \setminus \bigcup U_i)$ to $\text{rot}_{\mathbb{Q}}(K)$ is $\frac{1}{r}m(R)\big[\chi_{orb}(R)+\frac{1}{4}(l-k)\big]$. Note that, together with the discussion above, this lemma finishes the proof of the first part of Proposition \[prop:legendrian\]. As a consequence of trivializing $\xi$ over the solid tori $U_i$, each truncated region may be replaced by the original region without affecting its contribution to the rotation. Let $S_R$ be a component of $j(S) \cap M_R$. Note that $S_R$ an $A_R$-fold branched cover of $R$, branched over the orbifold points of $R$. We represent a trivialization of $\xi|_{S_R}$ by a non-vanishing vector field in $T S_R$, and we use the Poincaré-Hopf Theorem to compute the winding number of $T \partial S_R$ with respect to this framing on the boundary. Embed $S_R$ as a subsurface of a closed surface $\bar{S}_R$ satisfying $\chi(\bar{S}_R)=\chi(S_R)+1$. Choose a vector field $v$ on $\bar{S}_R$ that extends the trivialization of $\xi$ in the tori $U_i$ and which has the property that its unique critical point $c$ lies in $\bar{S}_R \setminus S_R$. Because $S_R$ is a branched cover of $R$, we may use the Riemann-Hurwitz Theorem to compute the Euler characteristic of $S_R$: $$\chi(S_R)= A_R\big[\chi(R)+\sum_{i=1}^r(\frac{1}{\alpha_i}-1)\big].$$ The Poincaré-Hopf Theorem implies that the index of $v$ at the unique critical point $c$ is $$\label{eq:gr1} \text{ind}_c v= 1+ A_R\big[ \chi(R)+\sum_{i=1}^r(\frac{1}{\alpha_i}-1)\big].$$ We now compute the winding number of $\partial S_R$ as an embedded curve with corners which encircles the singular point of the vector field. For simplicity, consider the curve $-\partial S_R$ (which bounds a neighborhood of the critical point positively). Identifying this neighborhood with a neighborhood of the origin in $\mathbb{C}$, and compute the winding number of the tangent to $-\partial S_R$ with respect to the translation-invariant page framing: $$\label{eq:gr2} \text{wind}_{page}(-\partial S_R)-A_R(\frac{k}{4}) +A_R(\frac{l}{4})=1.$$ To convert the winding number with respect to the page framing to the winding number with respect to $v$, subtract the index of $c$: $$\text{wind}_v(-\partial S_R)=\text{wind}_{page}(-\partial S_R)-\text{ind}_c v.$$ The Seifert surface is constructed locally using $\frac{m(R)}{A_R}$ copies of $S_R$, so the result follows from Equations (\[eq:orb-euler-char\]), (\[eq:gr1\]), and (\[eq:gr2\]). The rational Thurston-Bennequin number {#sect:tb} -------------------------------------- In this final section, we use a rational formal Seifert surface and a fiber distribution to compute the rational Thurston-Bennequin invariant. Recall from [@be:rational-tb] that the rational Thurston-Bennequin number of a Legendrian knot $K$ is defined to be the rational linking number of $K$ with a transverse push-off $K'$ with respect to some rational Seifert surface for $K$. Since the fibers are transverse to the contact planes, we may take $K'$ to be the Legendrian push-off along the Reeb direction; we may think of $K'$ as lying at the bottom of $\partial N$. Away from the double points of $\Gamma$, the conventions for how a rational Seifert surface $S$ interacts with $N$ in Figure \[fig:notation\] imply that there will be no intersection points. Thus, computing $\text{tb}_{\mathbb{Q}}(K)$ reduces to counting intersections between $S$ and $K'$ in the solid tori over the double points of $\Gamma$. As discussed above, it suffices to examine how the generalized Seifert algorithm extends the Seifert surface $S$ across a fibered neighborhood of a double point of $\Gamma$. The only interactions of $S$ and $K'$ will be when the generalized Seifert algorithm uses finger moves to push $S$ across the bottom of $N$. The sign of these intersections may be computed combinatorially as in Figure \[fig:intersect\]. We need to count (with sign) finger moves of $S$ across the bottom of $\partial N$. The first step in extending $S$ requires sliding each intersection between the fiber and the top edge of $C_t$ into $A_{II}$ and then standardizing the resulting pattern. Isotope the intersections from $A_{III}$ and $A_{IV}$ to the left across discs where $K$ is oriented to point into the page, and isotope the intersections from $A_I$ to the right across a disc where $K$ is oriented to point out of the page. Figure \[fig:standardize\] shows that moving all the intersections and standardizing the resulting pattern contributes $$2f_{IV}+f_{III}+f_I$$ to the signed intersection number. Performing saddle moves to eliminate all the longitudinal curves in the pattern does not change the intersection number. Furthermore, observe that the weight of the curves intersecting each side of the $N$ disc is preserved by the standardization process. When the strands of $K$ cross, the two $N$ discs on the edges of $A_{II}$ exchange places. Standardizing the resulting pattern introduces an additional $-m_{II}+m_{III}=-r$ intersections between $S$ and $K'$. Summing these with the previous intersections and repeating the process at every solid torus yields the following formula for the rational Thurston Bennequin number: $$tb_{\mathbb{Q}}(K)=\frac{1}{r}\sum_i (-r+f^i_{I}+f^i_{III}+2f^i_{IV}).$$ To make the formula more elegant, we repeat the same computation, but this time isotope all the intersections to $A_{IV}$ instead. Counting intersections yields: $$tb_{\mathbb{Q}}(K)=\frac{1}{r}\sum_i (-r-f^i_{III}-f^i_{I}-2f^i_{II}).$$ We sum the two formulae for $tb_{\mathbb{Q}}(K)$ and divide by $2$, which yields the desired formula: $$\text{tb}_{\mathbb{Q}}(K)=\frac{1}{r}\sum_i (-r-f^i_{II}+f^i_{IV}).$$

--- abstract: 'We study a one-parameter family of vector-valued polynomials associated to each simple Lie algebra. When this parameter $q$ equals $-1$ one recovers Joseph polynomials, whereas at $q$ cubic root of unity one obtains ground state eigenvectors of some integrable models with boundary conditions depending on the Lie algebra; in particular, we find that the sum of its entries is related to numbers of Alternating Sign Matrices and/or Plane Partitions in various symmetry classes.' address: - 'P. Di Francesco, Service de Physique Théorique de Saclay, CEA/DSM/SPhT, URA 2306 du CNRS, C.E.A.-Saclay, F-91191 Gif sur Yvette Cedex, France' - 'P. Zinn-Justin, Laboratoire de Physique Théorique et Modèles Statistiques, UMR 8626 du CNRS, Université Paris-Sud, Bâtiment 100, F-91405 Orsay Cedex, France' author: - 'P. Di Francesco' - 'P. Zinn-Justin' title: From Orbital Varieties to Alternating Sign Matrices --- Introduction ============ Recently, a remarkable connection between integrable models and combinatorics has emerged. It first appeared in a series of papers concerning the XXZ spin chain and the Temperley–Lieb (TL) loop model [@BdGN; @RSa] and which culminated with the so-called Razumov–Stroganov (RS) conjecture [@RS]. One of the main observations of [@BdGN], a weak corollary of the RS conjecture, is that the sum of entries of the properly normalized ground state vector of the TL(1) loop model is (unexpectedly!) equal to the number of Alternating Sign Matrices. This result was eventually proved in [@DFZJ] by using the [*integrability*]{} of the TL loop model in the following way: the model is generalized by introducing $N$ complex numbers (spectral parameters, or inhomogeneities) in the problem, where $N$ is the size of the system. The ground state entries become polynomials in these variables, and integrability provides many new tools for analyzing them, leading eventually to the exact computation of their sum, identified as the so-called Izergin–Korepin (IK) determinant, known to specialize to the number of Alternating Sign Matrices in the homogeneous limit [@Kup]. Note that in this work, the meaning of the spectral parameters is not very transparent; in particular, it is unclear how to generalize the full RS conjecture in their presence. Next, it was observed in [@Pas] that the polynomials obtained above really belong to a one-parameter family of solutions of a certain set of linear equations, in which the parameter $q$ has been set equal to a cubic root of unity. This observation is not obvious because the equations for generic $q$ are not a simple eigenvector equation; in fact, as explained in [@DFZJc], they are precisely the quantum Knizhnik–Zamolodchikov ($q$KZ) equations at level 1 for the algebra $U_q(\widehat{\mathfrak{sl}(2)})$. Furthermore, in the “rational” limit $q\to -1$, these polynomials have a remarkable geometric interpretation: they are equivariant Hilbert polynomials (or “multidegrees”) of $A_{N-1}$ orbital varieties $M^2=0$ ([@DFZJc], see also [@KZJ]), which are extensions of the Joseph polynomials [@Jo]. Note that here, the spectral parameters quite naturally appear as the basis of weights of $\mathfrak{gl}(N)$. In [@DFZJc], these ideas were generalized to higher algebras $U_q(\widehat{\mathfrak{sl}(k)})$, which correspond to the orbital varieties $M^k=0$. Here, we pursue a [*different*]{} type of generalization: we investigate orbital varieties corresponding to the other infinite series of simple Lie algebras: $B_r$, $C_r$, $D_r$; but we stick to the $U_q(\widehat{\mathfrak{sl}(2)})$ case by choosing the orbital varieties $M^2=0$, $M$ a complex matrix in the fundamental representation. Indeed, we show below that such orbital varieties are related to the same loop model, but with different [*boundary conditions*]{} (corresponding to variants of the Temperley–Lieb algebra). Furthermore, one can now $q$-deform the resulting polynomials to produce solutions of $q$KZ equations of type $B$, $C$, $D$ and set $q$ to be a cubic root of unity. Taking the homogeneous limit, the entries become integer numbers, which we conjecture to be related to [*symmetry classes*]{} of Alternating Sign Matrices and/or Plane Partitions; in particular we identify the sums of entries. In what follows we state most results without proofs; some will appear in a joint paper with A. Knutson [@DFKZJ] on a closely related subject. General setup ============= Orbital varieties ----------------- Let $\mathfrak{g}$ be a simple complex Lie algebra of rank $r$, $\mathfrak{b}$ a Borel subalgebra. $\mathfrak{b}=\mathfrak{t}\oplus \mathfrak{n}$ where $\mathfrak{t}$ is the corresponding Cartan subalgebra and $\mathfrak{n}$ is the space of nilpotent elements of $\mathfrak{b}$. $B$ and $T$ are Borel and Cartan subgroups. Let $W$ denote the Weyl group of $\mathfrak{g}$, and $s_\alpha$ its standard generators, where $\alpha$ runs over the set of simple roots of $\mathfrak{g}$. Fixing an orbit $G\cdot x$, with $x \in \mathfrak{n}$ and $G$ acting by conjugation, one can consider the irreducible components of $\overline{{\mathfrak b}\cap (G\cdot x)}$, which are called orbital varieties. Even though much of what follows can be done for any orbital varieties, we focus below on the following special case: we fix an irreducible representation $\rho$ (of dimension $N$) and consider the scheme $E=\{ x\in \mathfrak{b} \mid \rho(x)^2=0 \}$. The underlying set is precisely a $\overline{{\mathfrak b}\cap (G\cdot x)}$, where $x$ is any element of $E$ such that $\rho(x)$ is of maximal rank. In some sense, its components are the “simplest possible” orbital varieties. Hotta construction ------------------ It is known that there exists a representation of the Weyl group $W$ on the vector space $V$ of formal linear combinations of orbital varieties (Springer/Joseph representation); for each $G$-orbit, it is an irreducible representation. We use the following explicit form of the representation: note that orbital varieties are invariant under $T\times {\mathbb{C}}^\times$, where $T$ acts by conjugation and ${\mathbb{C}}^\times$ acts by overall scaling. We can therefore consider equivariant cohomology $H^*_{T\times {\mathbb{C}}^\times}(\cdot)$ and in particular via the inclusion map from each orbital variety $\pi$ to the space $\mathfrak{n}$, the unit of $H^*_{T\times{\mathbb{C}}^\times}(\pi)$ is pushed forward to some cohomology class $\Psi_\pi$ in $H^*_{T\times{\mathbb{C}}^\times}(\mathfrak{n})={\mathbb{C}}[\mathfrak{t},A]$, that is a polynomial in $r+1$ variables $\alpha_1$, $\ldots$, $\alpha_r$, $A$ (the $r$ simple roots plus the ${\mathbb{C}}^\times$ weight), sometimes called [*multidegree*]{} of $\pi$. Suppressing the ${\mathbb{C}}^\times$ action, that is setting $A=0$, one recovers the Joseph polynomials [@Jo]. The way that $W$ acts on these polynomials can be described explicitly, by extending slightly the results of Hotta [@Ho] to include the additional ${\mathbb{C}}^\times$ action. One starts by associating to each simple root $\alpha$ a certain geometric construction, which we briefly recall. For $x\in \mathfrak{b}$ write $x=\sum_\alpha x_\alpha e_\alpha$ where $\alpha$ runs over positive roots, $e_\alpha\in\mathfrak{g}$ being a vector of weight $\alpha$. Define $\mathfrak{b}_\alpha=\{ x\in \mathfrak{b}\mid x_\alpha=0 \}$, and $L_\alpha$ to be Lévy subgroup whose Lie algebra is $\mathfrak{b}\oplus {\mathbb{C}}e_{-\alpha}$. Starting from an orbital variety $\pi$, we distinguish two cases: - $\pi\subset \mathfrak{b}_\alpha$. Then set $s_\alpha \pi=\pi$. - $\pi \not\subset \mathfrak{b}_\alpha$. Then let $L_\alpha$ acts by conjugation: the top-dimensional components of $L_\alpha\cdot (\pi\cap \mathfrak{b}_\alpha)$ are again orbital varieties; set $s_\alpha \pi = - \pi - \sum_{\pi'} \mu_{\alpha}{}_\pi^{\pi'} \pi'$ where $\mu_{\alpha}{}^{\pi'}_\pi$ is the multiplicity of $\pi'$ in $L_\alpha\cdot (\pi\cap \mathfrak{b}_\alpha)$. These elementary operations have a counterpart when acting on multidegrees, and a simple calculation shows that both cases are covered by a single formula: $$\label{funda} s_\alpha \Psi_\pi = (-\tau_\alpha + A {\partial}_\alpha) \Psi_\pi$$ where $\tau_\alpha$ is the reflection orthogonal to the root $\alpha$ in ${\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$, and ${\partial}_\alpha={1\over\alpha}(\tau_\alpha-1)$ is the associated [*divided difference operator*]{}, whereas on the left hand side $s_\alpha$ implements right action on the $\Psi_\pi$, namely $s_\alpha\Psi_\pi:=-\Psi_\pi-\sum_{\pi'} \mu_{\alpha}{}^{\pi'}_\pi \Psi_{\pi'}$. One can check that $s_\alpha\mapsto -\tau_\alpha+A {\partial}_\alpha$ is a representation of the Weyl group $W$ on polynomials. Note that at $A=0$, we recover the natural action of $W$ (up to a sign, with our conventions). Yang–Baxter equation and integrable models ------------------------------------------ Let us define the operator $$R_\alpha(u):={A-u s_\alpha\over A+u}$$ which acts in the space $V\otimes {\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$, $u$ being a formal parameter. Rewriting slightly the relation above we find that $\tau_\alpha$ acts as $R_\alpha(\alpha)$. Using the fact that $\tau_\alpha$, just like the $s_\alpha$, satisfy the Weyl group relations, we find that the operators $\tau_\alpha R_\alpha(\alpha)$ also satisfy those. In the case of non-exceptional Lie algebras, there are only 2 types of edges in the Dynkin diagram, and therefore we have Coxeter relations of the form $(s_\alpha s_\beta)^{m_{\alpha\beta}}=1$, where $m_{\alpha\beta}=1,2,3,4$ depending on whether $\alpha=\beta$, there is no edge, a single or a double edge between $\alpha$ and $\beta$. Writing these relations for $\tau_\alpha R_\alpha$ and eliminating the $\tau_\alpha$, we find that relations with $m_{\alpha\beta}=1,3,4$ correspond respectively to the [*unitarity*]{} equation: $$R_\alpha(\alpha)R_\alpha(-\alpha)=1\ ,$$ the [*Yang–Baxter*]{} equation: $$R_\alpha(\alpha)R_\beta(\alpha+\beta)R_\alpha(\beta)= R_\beta(\beta)R_\alpha(\alpha+\beta)R_\beta(\alpha) \qquad \alpha\ \epsfig{file=singleedge.eps, width=1cm}\ \beta$$ and the [*boundary Yang–Baxter*]{} (or reflection) equation: $$R_\alpha(\alpha)R_\beta(\beta+\alpha)R_\alpha(\alpha+2\beta)R_\beta(\beta) = R_\beta(\beta)R_\alpha(\alpha+2\beta)R_\beta(\beta+\alpha)R_\alpha(\alpha) \qquad \alpha\ \epsfig{file=doubleedge.eps, width=1cm}\ \beta$$ whereas the case $m_{\alpha\beta}=2$ expresses a simple commutation relation for distant vertices. Indeed one recognizes in $R_\alpha(u)$ a standard form of the rational solution of the Yang–Baxter equation, the parameter $u$ playing the role of difference of spectral parameters. Thus the multidegrees $\Psi_\alpha$ are closely connected to integrable models with rational dependence on spectral parameters, as will be discussed now. Before doing so, let us remark that in the special case investigated here of orbital varieties associated to $M^2=0$, the $s_\alpha$ obey more than just the Coxeter relations. In the $A_r$ case they actually generate a quotient of the symmetric group algebra $S_{r+1}$ known as the Temperley–Lieb algebra $TL_{r+1}(2)$ (here 2 is the value of the parameter in the definition of the algebra, as will be explained below). The same type of phenomena will be described for other simple Lie algebras, and will lead to variants of the Temperley–Lieb algebra; in particular, the “bulk” (i.e. everything but a finite number of edges at the boundary) of the Dynkin diagrams being sequences of simple edges, these variants will only differ at the level of “boundary conditions” of the model. Affinization and rational $q$KZ equation ---------------------------------------- Let us now discuss the meaning of the equation $$R_\alpha(\alpha) \Psi = \tau_\alpha \Psi$$ where $\tau_\alpha$ is the reflection associated to the root $\alpha$ acting on the “spectral parameters” $\alpha_1$, $\ldots$, $\alpha_r$, $R_\alpha(\alpha)$ is a certain linear operator defined above acting in the space $V\otimes {\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$ and $\Psi=\sum_\pi \pi\otimes \Psi_\pi$ is a vector in that space. When $R_\alpha(u)$ is the $R$-matrix (or boundary $R$-matrix) of some integrable model, such equations are satisfied by eigenvectors of the corresponding integrable transfer matrix. More generally, these equations appear in the context of the quantum Knizhnik–Zamolodchikov ($q$KZ) equation, in connection with the representation theory of affine quantum groups [@FR]. In either case, it is known that we need an additional equation to fix the $\Psi_\pi$ entirely. Define $\hat{W}$ to be the semi-direct product of $W$ and of the weight lattice of $\mathfrak{g}$. It contains as a finite index subgroup the usual affine Weyl group defined as the Coxeter group of the affinized Dynkin diagram. Just like the affine Weyl group, it has a natural action on $\mathfrak{t}$ and therefore on ${\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$ which extends the action of $W$ generated by the reflections $\tau_i$; by definition, in this representation, an element of the weight lattice acts as translation in $\mathfrak{t}$ of the weight multiplied by $3A$ ($3=l+\check h$ where $l=1$ is the level of the $q$KZ equation and $\check h=2$ is the dual Coxeter number of $\mathfrak{sl}(2)$). Then we claim that one can extend the representation of $W$ on $V\otimes {\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$ (the operators $\tau_\alpha R_\alpha(\alpha)$) into a representation of $\hat{W}$, in such a way that each element of $\hat W$ is the product of its natural action on ${\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$ and of a ${\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$-linear operator. Describing here the geometric procedure that leads to this action is beyond the scope of this paper. The action will however be described explicitly in each of the cases below. An important property is that if one sets $A=0$ the representation of $\hat{W}$ factors through the projection $\hat W \to W$. So the ${\mathbb{C}}^\times$ action actually produces the affinization. Imposing that $\Psi$ be invariant under the action of the whole group $\hat W$ leads to a full set of equations, which are precisely equivalent to the so-called rational $q$KZ equation (or more precisely, a generalization of it for arbitrary Dynkin diagram, the original $q$KZ equation corresponding to the case $A_r$) at level $1$; and it turns out that they have a unique polynomial solution of the prescribed degree (up to multiplication by a scalar). $q$-deformation and Razumov–Stroganov point ------------------------------------------- The integrability suggests how to $q$-deform the above construction. Indeed, we have considered thus far $R$-matrices that form so-called [*rational*]{} solutions of the Yang–Baxter Equation, and $\Psi$’s that are solutions of the rational $q$KZ equation. It is known however that the trigonometric $R$-matrices are a special degeneration of a one-parameter family of [*trigonometric*]{} solutions of the Yang–Baxter Equation, depending on a parameter $q$. Setting $q=-e^{-\hbar A/2}$, one customarily uses exponentiated “multiplicative" spectral parameters of the form $e^{-\hbar \alpha_i}$. We then look for polynomial solutions $\Psi$ of these parameters, to the corresponding trigonometric $q$KZ equations. The rational solutions are then recovered from the trigonometric ones via the limit $\hbar\to0$, at the first non-trivial order in $\hbar$. The details of the bulk and boundary $R$-matrices will be given below for the cases $A_r$, $B_r$, $C_r$ and $D_r$. We thus obtain, for any $q$, a representation of the group $\hat W$, the $W$ relations satisfied by the $\tau_\alpha R_\alpha(\alpha)$ and more generally the $\hat W$ relations being undeformed. In terms of the new variables $e^{-\hbar \alpha_i}$ living in $T$, the natural action of an element of the weight lattice $\omega$ (as the abelian subgroup of $\hat W$) is the multiplication by $q^{6\omega}$. Since for all simple Lie algebras, $\omega$ has half-integer coordinates, we reach the important conclusion that when $q^3=1$, this action becomes trivial. Therefore, all operators associated to the weight lattice by the procedure outlined in the previous section become ${\mathbb{C}}[\alpha_1,\ldots,\alpha_r,A]$-linear (i.e. correspond to finite-dimensional operators on $V$ after evaluation of the parameters $\alpha_1$, $\ldots$, $\alpha_r$, $A$). In this case they are simply the scattering matrices of [@Pasb], and they commute with the usual (inhomogeneous) integrable transfer matrix of the model. This implies that $\Psi$ is an eigenvector of the latter; in fact, we can call it “ground state eigenvector” because in the physical situation where the transfer matrix elements are positive, the Perron–Frobenius theorem applies and the eigenvalue $1$ of $\Psi$ is the largest eigenvalue in modulus. The value $q=e^{2i\pi/3}$ (also called “Razumov–Stroganov point") is henceforth quite special and deserves a particular study. In particular, in the homogeneous limit where the spectral parameters $\alpha_i$ are specialized to zero, $\Psi$ can be normalized so that its entries are all [*non-negative integers*]{}, and we are interested in their combinatorial significance, in relation to the counting of Alternating Sign Matrices and/or Plane Partitions. We do not claim to have a full understanding of the general correspondence principle between simple Lie algebras and these combinatorial problems, but we will perform a case-by-case study for $A_r$, $B_r$, $C_r$ and $D_r$. A last remark is in order. As we shall see, it is simple to see that the solutions $\Psi$ to the $A$, $B$, $C$, $D$ $q$KZ equations obey recursion relations, that allow to obtain the rank $r$ case from rank $r+1$, hence we will content ourselves with the detailed description for $r$ with a given parity, namely $A_{2n-1}$, $B_{2n}$, $C_{2n+1}$, $D_{2n+1}$. $A_r$ case ========== We review the $A_r$ case, already explored in [@DFZJc]. We set $\alpha_i=z_i-z_{i+1}$, $i=1,\ldots,r$. The fact that there are $r+1\equiv N$ of these new variables $z_i$, the spectral parameters, as opposed to the $r$ simple roots, is a reflection of the usual embedding $\mathfrak{sl}(N)\subset \mathfrak{gl}(N)$. $\mathfrak{b}$ (resp. $\mathfrak{n}$) is simply the space of upper triangular (resp. strictly upper triangular) matrices of size $N$, and the orbital varieties under consideration are the irreducible components of the scheme $\{ M\in \mathfrak{n} \mid M^2=0\}$. We also restrict ourselves to the case of $N=2n$ even, which is technically simpler. Orbital varieties and Temperley–Lieb algebra -------------------------------------------- In general, $\mathfrak{sl}(N)$ nilpotent orbits are classified by their Jordan decomposition type, which can be expressed as a Young diagram; the orbital varieties are then indexed by Standard Young Tableaux (SYT). The condition $M^2=0$ ensures that only Young diagrams with at most 2 rows can appear (blocks in the Jordan decomposition are of size at most 2), and it is easy to check that all orbits are in the closure of the largest orbit, whose Young diagram is of the form $(n,n)$. It is convenient to describe the corresponding SYT by “link patterns”, that is $N$ points on a line connected in the upper-half plane via $n$ non-intersecting arches, see fig. \[linkpatt\]. The numbers in the first (resp. second) row of the SYT are the labels of the openings (resp. closings) of the arches. There are ${(2n)!\over n!(n+1)!}$ such configurations. 0.5cm In this language, one has a rather convenient description of orbital varieties [@Roth; @Mel], which we mention for the sake of completeness. Indeed, to each orbital variety $\pi$ we associate the upper triangular matrix $\pi^<$ with $\pi^<_{ij}=1$ if points labelled $i$ and $j$ are connected by an arch, $i<j$, $0$ otherwise. Then $\pi=\overline{B\cdot\pi^<}$, $B$ acting by conjugation. Equivalently, $\pi$ is given by the following set of equations: (i) $M^2=0$ and (ii) $r_{ij}(M)\le r_{ij}(\pi^<)$, $i,j=1,\ldots,N$, where $r_{ij}$ is the rank of the $i\times j$ lower-left rectangle. $$e_4\ \raise-2.6mm\hbox{\epsfig{file=arch3.eps,width=4cm}} = \raise-9.2mm\hbox{\epsfig{file=ei1.eps,width=4cm}} = \raise-2.6mm\hbox{\epsfig{file=arch1.eps,width=4cm}}$$ $$e_2\ \raise-2.6mm\hbox{\epsfig{file=arch3.eps,width=4cm}} = \raise-9.2mm\hbox{\epsfig{file=ei2.eps,width=4cm}} = \beta\ \raise-2.6mm\hbox{\epsfig{file=arch3.eps,width=4cm}}$$ It is equally simple to describe the action of the Weyl group, namely the symmetric group $S_N$. Rather than the generators corresponding to the simple roots: $s_i\equiv s_{\alpha_i}$, $i=1,\ldots,r$ used so far, it proves simpler to consider the action of the projectors $e_i=1-s_i$ in the symmetric group algebra. The operator $e_i$ acts on link patterns $\pi$ by connecting the arches ending at $i$ and $i+1$ and creates a new little arch between these 2 points; this action is described on Fig. \[tla\]. When a closed loop is formed, it is erased but contributes a weight $\beta=2$. The $q$-deformed version of this is obtained by attaching a weight $\beta=-(q+q^{-1})$ to each erased loop, thus leading to the following (pictorially clear) relations: $$\label{tlarel} e_i^2=\beta e_i\qquad e_i=e_i e_{i\pm 1}e_i\qquad [e_i,e_j]=0\quad |i-j|>1$$ all indices taking values in $1,\ldots,r$. These are the defining relations of the Temperley–Lieb algebra $TL_{r+1}(\beta)$. When $q=-1$, i.e. $\beta=2$, it is simply a quotient of the symmetric group algebra. Alternatively, the deformed generators $s_i=-q^{-1}-e_i$ satisfy the usual relations of the Hecke algebra (of which the Temperley–Lieb algebra is a quotient). In what follows, one special element of $TL_N(\beta)$ will be needed: it is the cyclic rotation $S$. Its effect is to rotate the endpoints of the link patterns: $1\to2\to\cdots\to N\to1$ without changing their connectivity. It can also be expressed as: $S=q^{n-2} s_1 \cdots s_{N-1}$. $q$KZ equation -------------- For each simple root $\alpha_i$, we have the trigonometric $R$-matrix: $$\label{qRmat} R_i(w)\equiv R_{\alpha_i}(w)={(qw-q^{-1})+(w-1)e_i\over q-q^{-1}w},$$ where the $e_i=-q^{-1}-s_i$ generate $TL_N(\beta)$ and act in the space of link patterns as explained above. We first write the system of equations: $$\label{qkza} R_i(w_{i+1}/w_{i})\Psi=\tau_i \Psi \qquad i=1,\ldots,N-1$$ where $\tau_i\equiv \tau_{\alpha_i}$ acts by interchanging multiplicative spectral parameters $w_i:=e^{-\hbar z_i}$ and $w_{i+1}$ in the polynomial $\Psi$ of the $w$’s, homogeneous of degree $n(n-1)$. These equations are supplemented by the “affinized” equation satisfied by $\Psi$. Since the affine Dynkin diagram $A_r^{(1)}$ is a circular chain, this equation quite naturally involves the cyclic rotation $S$. Define the operator $\rho$ on ${\mathbb{C}}[w_1,\ldots,w_N]$ which shifts the variables $w_i$ according to the rule: $w_i\to w_{i+1}$, $i=1,\ldots,N-1$ and $w_N\to q^6 w_1$. Then the additional equation is $$\label{qkzaa} q^{3(n-1)}S^{-1} \Psi=\rho\Psi$$ Together with this equation, the above system forms the so-called level one $q$KZ equation. We claim that the ${\bf R}_i:=\tau_i R_i(w_{i+1}/w_i)$ and ${\bf S}:=q^{3(1-n)}\rho S$ generate together $\hat W$. In order to see that, it is sufficient to build the $N$ generators ${\bf T}_i$ of the abelian subgroup (the lattice of weights). They are given by ${\bf T}_i={\bf R}_{i-1}{\bf R}_{i-2}\cdots {\bf R}_1 {\bf S} {\bf R}_{N-1}\cdots {\bf R}_{i+1} {\bf R}_i$, $i=1,\ldots,N$. The original definition of the $q$KZ equation is in fact the eigenvector equation for these “scattering” matrices; with reasonable assumptions it is equivalent to the above system. Also, note that if one defines ${\bf R}_N:={\bf S}^{-1} {\bf R}_1 {\bf S}$, then the ${\bf R}_i$, $i=1,\ldots,N$ generate the usual affine Weyl group (a subgroup of order $N$ of $\hat W$). The minimal degree polynomial solution of the level one $q$KZ equation was obtained in [@Pas; @DFZJc], and is characterized by its “base” entry $\Psi_{\pi_0}$ corresponding to the link pattern $\pi_0$ that connects points $i\leftrightarrow 2n+1-i$, with the value $$\Psi_{\pi_0}=\prod_{1\leq i<j\leq n} (qw_i-q^{-1}w_j)\prod_{n+1\leq i<j\leq 2n}(qw_i-q^{-1}w_j)$$ in which all factors are a direct consequence of the $\tau_i\Psi=R_i\Psi$ equations. It is then easy to prove that all the other entries of $\Psi$ may be obtained from $\Psi_{\pi_0}$ in a triangular way. [*Example:*]{} at $N=6$, there are 5 link patterns. The minimal degree polynomial solution of the level one $q$KZ equation reads: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0.eps,width=1.8cm}} =(qw_1-q^{-1}w_2)(qw_2-q^{-1}w_3)(qw_1-q^{-1}w_3) (qw_4-q^{-1}w_5)(qw_5-q^{-1}w_6)(qw_4-q^{-1}w_6)\\ &&\Psi_{\epsfig{file=arch1.eps,width=1.8cm}} =(qw_1-q^{-1}w_2)(qw_3-q^{-1}w_4)(qw_5-q^{-1}w_6)\\ &&\times \Big((w_1+w_2)(q^2w_3w_4-q^{-2}w_5w_6)-(w_3+w_4)(q^4w_1w_2-q^{-4}w_5w_6) +(w_5+w_6)(q^2w_1w_2-q^{-2}w_3w_4)\Big)\\ &&\Psi_{\epsfig{file=arch2.eps,width=1.8cm}} =(qw_2-q^{-1}w_3)(qw_2-q^{-1}w_4)(qw_3-q^{-1}w_4) (qw_5-q^{-1}w_6)(q^{-2}w_6-q^{2}w_1)(q^{-2}w_5-q^{2}w_1)\\ &&\Psi_{\epsfig{file=arch3.eps,width=1.8cm}} =(qw_1-q^{-1}w_2)(qw_3-q^{-1}w_4)(qw_4-q^{-1}w_5) (qw_3-q^{-1}w_5)(q^{-2}w_6-q^{2}w_1)(q^{-2}w_6-q^{2}w_2)\\ &&\Psi_{\epsfig{file=arch4.eps,width=1.8cm}} =(qw_2-q^{-1}w_3)(qw_4-q^{-1}w_5)(q^{-2}w_6-q^{2}w_1)\\ &&\times \Big((q^3w_1+q^{-3}w_6)(q^2w_2w_3-q^{-2}w_4w_5)-(w_2+w_3)(qw_1w_6-q^{-1}w_4w_5) -(w_4+w_5)(qw_2w_3-q^{-1}w_1w_6)\Big)\\\end{aligned}$$ Performing the rational limit $\hbar\to 0$, $z_i=e^{-\hbar w_i}$, $q=-e^{-\hbar A/2}$ yields the following multidegrees: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0.eps,width=1.8cm}} =(A+z_1-z_2)(A+z_2-z_3)(A+z_1-z_3) (A+z_4-z_5)(A+z_5-z_6)(A+z_4-z_6)\\ &&\Psi_{\epsfig{file=arch1.eps,width=1.8cm}} =(A+z_1-z_2)(A+z_3-z_4)(A+z_5-z_6)\Big(4A^3 + 3A^2(z_1+z_2-z_5-z_6) +\\ &&+ A(2(z_1z_2-2z_3z_4-z_1z_5-z_2z_5-z_1z_6-z_2z_6+z_5z_6)+ (z_3+z_4)(z_1+z_2+z_5+z_6))\\ &&+ (z_1+z_2)(z_5z_6-z_3z_4)+(z_3+z_4)(z_1z_2-z_5z_6)+(z_5+z_6)(z_3z_4-z_1z_2)\Big)\\ &&\Psi_{\epsfig{file=arch2.eps,width=1.8cm}} =(A+z_2-z_3)(A+z_2-z_4)(A+z_3-z_4) (A+z_5-z_6)(2A+z_1-z_6)(2A+z_1-z_5)\\ &&\Psi_{\epsfig{file=arch3.eps,width=1.8cm}} =(A+z_1-z_2)(A+z_3-z_4)(A+z_4-z_5) (A+z_3-z_5)(2A+z_1-z_6)(2A+z_2-z_6)\\ &&\Psi_{\epsfig{file=arch4.eps,width=1.8cm}} =(A+z_2-z_3)(A+z_4-z_5)(2A+z_1-z_6)\Big(5A^3 + 3A^2(z_1+z_2+z_3-z_4-z_5-z_6) +\\ &&+A(2z_1(z_2+z_3-z_6)+z_2z_3+z_4z_5-(z_2+z_3)z_6+(z_4+z_5)(2z_6-z_1-z_2-z_3))\\ &&+(z_1+z_6)(z_2z_3-z_4z_5)+(z_2+z_3)(z_4z_5-z_1z_6)+(z_4+z_5)(z_1z_6-z_2z_3)\Big)\\\end{aligned}$$ and in particular the degrees $1,4,4,4,10$ respectively, upon taking $z_i=0$ and $A=1$. Razumov–Stroganov point and ASM ------------------------------- At $q=e^{2i\pi/3}$, $\Psi$ becomes the ground state eigenvector of the integrable transfer matrix with periodic boundary conditions and inhomogeneities $w_1$, $\ldots$, $w_N$, or equivalently of the scattering matrices ${\bf T}_i=R_{i-1}(w_{i-1}/w_i)\cdots R_1(w_1/w_i) S R_{N-1}(w_{N-1}/w_i)\cdots R_i(w_{i+1}/w_i)$. Consider now the particular case $w_1=\cdots=w_N=1$, when $\Psi$ is the Perron–Frobenius eigenvector of the Hamiltonian $H=e_1+\cdots+e_N$ where $e_N=S^{-1} e_1 S$. Note that the periodic boundary conditions mean that $H$ is cyclic-invariant: $SH=HS$. Normalizing $\Psi$ so that its smallest entry $\Psi_{\pi_0}$ is $1$, we have the following [@DFZJ] The sum of entries $\sum_\pi \Psi_\pi$ is equal to the number of Alternating Sign Matrices, $A(n)$. The result of [@DFZJ] is actually much more general, as the sum $\sum_\pi \Psi_\pi$ was evaluated in the presence of all the spectral parameters $w_i$, and identified with proper normalization to the so-called Izergin–Korepin determinant [@Iz; @Kor], also equal to a particular Schur function [@Oka]. Still unproven, however, is the [@BdGN] The largest entry of $\Psi$, with arches connecting consecutive points, is $A(n-1)$. For instance, plugging $w_i=1$ and $q=e^{2i\pi/3}$ into the above example, we get for $N=6$, $\Psi=(1,2,1,1,2)$ and $\sum_\pi \Psi_\pi=7=A(3)$, the total number of $3\times 3$ ASMs. $B_r$ case ========== We now develop the $B_r$ case, which allows us to recover and interpret geometrically the results of [@DF]. We concentrate on the even case $r=2n$. We parametrize as usual the roots $\alpha_i=z_i-z_{i+1}$ for $i=1,2,\ldots,r-1$ and $\alpha_r=z_r$. We consider matrices that square to zero in the fundamental representation of dimension $N=2r+1$: a possible choice is to select upper triangular matrices satisfying $M^TJ+JM=0$, $J$ antidiagonal matrix with 1’s on the second diagonal. It turns out that the orbital varieties are indexed by the same link patterns as before, of size $r$; and that the Weyl group representation is actually a representation of the same quotient, the Temperley–Lieb algebra $TL_r(\beta)$, the additional reflection $s_r$ being represented by a multiple of the identity. $B$-type $q$KZ equation ----------------------- According to the dicusssion above, the B $q$KZ system reads: $$\begin{aligned} R_i(w_{i+1}/w_i)\Psi&=&\tau_i \Psi, \quad i=1,2,...,r-1\label{firBx} \\ w_r^{-m_r}{q^{-1}w_r-q\over q^{-1}-q w_r} \Psi&=&\tau_r \Psi\label{firB}\end{aligned}$$ where $\tau_r$ stands for the inversion of the last spectral parameter, namely $\tau_r\Psi(w_1,...w_{r-1},w_r)= \Psi(w_1,...,w_{r-1},1/w_r)$ and $m_r$ is the degree of $\Psi$ in $w_r$. Finally, these equations are to be supplemented by the affinization relation. The latter is expressed by considering the reflection with respect to the extra root $z_1$. One finds that $$\label{secB} (q^3w_1)^{-m_1} {q^{-2}-q^2 w_1\over q w_1-q^{-1}} \Psi(w_1,w_2,...,w_r)= \Psi\Big({1\over q^6w_1},w_2,...,w_{r}\Big)$$ where $m_1$ is the degree of $\Psi$ in $z_1$. Introducing the boundary operators ${\bf K}_1$ and ${\bf K}_r$ so that Eqs. (\[firB\]–\[secB\]) reduce to ${\bf K}_1 \Psi={\bf K}_2 \Psi=\Psi$, as well as the usual ${\bf R}_i=\tau_i R_i(w_{i+1}/w_i)$, the generators of the weight lattice (as abelian subgroup of $\hat W$) are: (i) ${\bf T}_i={\bf R}_i{\bf R}_{i+1}\cdots {\bf R}_{r-1}{\bf K}_r {\bf R}_{r-1}\cdots$ ${\bf R}_1 {\bf K}_1 {\bf R}_1\cdots{\bf R}_{i-1}$ that implements $w_i\to q^6 w_i$ and (ii) one additional generator implementing $w_i\to q^3 w_i$ simultaneously for all $i$. The latter is a combination of [**R**]{} and [**K**]{} matrices as well as an additional operator implementing the reflection $w_i\leftrightarrow q^{-3}/w_{r+1-i}$ for all $i$. The minimal polynomial solution to the system (\[firBx\]–\[secB\]) has degree $m_1=m_r=r-1=2n-1$ in each spectral parameter and total degree $n(3n-1)$. As before it has a simple factorized base entry $$\Psi_{\pi_0}=C\prod_{1\leq i<j\leq n}(qw_i-q^{-1}w_j)(q^{-2}-q^{2}w_iw_j) \prod_{n+1\leq i<j\leq 2n} (qw_i-q^{-1}w_j)(qw_iw_j-q^{-1})$$ where $C=2^n\prod_{i=1}^r (q w_i-q^{-1})$ is a common (symmetric) factor to all entries of $\Psi$. All other entries may be obtained from this one in a triangular manner. [*Example:*]{} For $B_4$, there are 2 link patterns as for the case $A_3$. The minimal degree polynomial solution of the level one $B_4$ $q$KZ equation reads: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0b.eps,width=1.2cm}} =C(qw_1-q^{-1}w_2)(q^{-2}-q^{2}w_1w_2)(qw_3-q^{-1}w_4)(qw_3w_4-q^{-1})\\ &&\Psi_{\epsfig{file=arch1b.eps,width=1.2cm}} =C(qw_2-q^{-1}w_3)(q^{-1}w_1-qw_1w_2w_3-q^{-5}w_4-qw_1^2w_4+(q^{-1}-q)w_1(w_2+w_3)w_4\\ &&\ \ \ +q^{-1}w_2w_3w_4+q^5 w_1^2 w_2w_3w_4+q^{-1}w_1w_4^2-qw_1w_2w_3w_4^2)\\\end{aligned}$$ As before, we get the corresponding multidegrees upon taking the rational limit, with the result: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0b.eps,width=1.2cm}} =C'(A+z_1-z_2)(2 A + z_1+ z_2) (A + z_3 - z_4) (A + z_3 + z_4)\\ &&\Psi_{\epsfig{file=arch1b.eps,width=1.2cm}} =C'(A+z_2-z_3)\Big(5A^3+3A^2 (2z_1 + z_2 + z_3)+A(2z_1^2 + 3z_1(z_2 +z_3) +z_2z_3 -z_4^2)\\ &&+ (z_2+z_3)(z_1^2-z_4^2)\Big)\\\end{aligned}$$ with $C'=4(A+z_1)(A+z_2)(A+z_3)(A+z_4)$; hence the degrees $4\times2,4\times5$ for $A=1$ and $z_i=0$. RS point, VSASM and CSTCPP -------------------------- As explained in Sect. 2, the case $q=e^{2i\pi/3}$ is special in that the problem admits a transfer matrix, and its solution $\Psi$ in the homogeneous limit where all $w_i=1$ is the groundstate of a Hamiltonian $$H_B=e_1+e_2+...+e_{N-1}$$ which is the open boundary version of the $A_r$ Hamiltonian $H$. As shown in [@DFb], at the RS point $q=e^{2i\pi/3}$, and in the homogeneous limit where $w_i=1$ for all $i$, and in which $\Psi$ is normalized so that its smallest entry is $\Psi_{\pi_0}=1$, we have the following [@DF] The sum of entries $\sum_\pi \Psi_\pi$ is equal to the number of Vertically Symmetric Alternating Sign Matrices (VSASM), $A_V(2n+1)$. This was actually proved in the same spirit as for the $A_r$ case, by identifying the sum of components including all spectral parameters $w_i$ as yet another determinant, which takes the form of a particular symplectic Schur function. A similar result holds for the case of odd $r=2n-1$, namely once properly normalized, the sum of entries $\sum_\pi \Psi_\pi$ is equal to an integer we call $A_V(2n)$ by analogy. It turns out that $A_V(2n)=N_8(2n)$ is the number of Cyclically Symmetric Transpose Complement Plane Partitions (CSTCPP) in an hexagon of size $2n\times 2n\times 2n$ [@Bre]. The numbers $A_V(i)$ both have determinant formulae, namely $A_V(2n)=\det{i+j\choose 2i-j}_{0\leq i,j\leq n-1}$, and $A_V(2n+1)=\det{i+j+1\choose 2i-j}_{0\leq i,j\leq n-1}$. As in the $A$ case, we have the [@BdGN] The largest entry of $\Psi$, with arches connecting consecutive points, is $A_V(r)$. Example: for $r=2n=4$, taking $w_i\to 1$ and $q=e^{2i\pi/3}$ in the above expressions, we get the components $\Psi=(1,2)$, which sum to $3=A_V(5)$, the number of $5\times 5$ VSASMs, and the maximal entry of $\Psi$ is $2=N_8(4)$. $C_r$ case ========== The simple roots of $C_r$ are $\alpha_i=z_i-z_{i+1}$, $i=1,2,\ldots,r-1$ and $\alpha_r=2z_r$. We concentrate on the odd case $r=2n+1$, and consider the fundamental representation of dimension $N=2r$. One choice is to select upper triangular matrices satisfying $M^T J+JM=0$, $J$ antidiagonal matrix with $1$’s (resp. $-1$’s) in the upper (resp. lower) triangle. Orbital varieties and $C$-type Temperley–Lieb algebra ----------------------------------------------------- There are $r\choose\lfloor{r+1\over2}\rfloor$ orbital varieties, which are now indexed by [*open link patterns*]{}, that is configurations of $r$ points on a line connected in the upper-half plane either in pairs via (closed) arches or to infinity via half-lines (open arches). The representation of the Weyl group on these open link patterns takes the form of a modified Temperley–Lieb algebra. We describe now its $q$-deformed version, $CTL(\beta)$ (see also [@Gr] for other variants of Temperley–Lieb algebra). The generators $e_1,e_2,\ldots,e_{r-1}$ obey the standard $TL(\beta)$ relations and the additional “boundary" generator $e_r$ satisfies: $e_r^2=\beta e_r$, $e_{r-1}e_re_{r-1}=2 e_{r-1}$. These generators act on open link patterns as follows. Open link patterns are represented with their open arches connected to a vertical line on the right. The $e_i$, $i=1,2,...,r-1$ act as usual, and $e_r$ like the left half of an $e$, connecting the point $2n+1$ to the vertical line (first line of Fig. \[crules\]). The rule is that any loop may be erased and replaced by a factor $\beta$. Moreover, whenever a connection between points on the vertical line (consecutive open arches) is created, they may also be erased and replaced by a factor $\beta$ (resp. $2$) if this is created by the action of some $e_{2i-1}$ (resp. $e_{2i}$). As $r$ is odd, the loop created by $e_r^2$ yields a weight $\beta$, while that created by $e_{r-1}e_re_{r-1}$ yields a weight $2$, hence the result $2e_{n-1}$ (second line of Fig. \[crules\]). We shall also need an additional operator $e_1'$ satisfying the relations: $(e_1')^2=\beta e_1'$ and $e_1e_1'=e_1'e_1=e_1'e_2e_1'-e_1'=e_2e_1'e_2-e_2=0$. It is defined as $e'_1=s e_1 s$, where $s$ is the involution acting on link patterns as follows: (i) $s\pi=\pi$ if the arch connected to point $1$ is open, and (ii) $s\pi=-\pi+\pi'$ otherwise, where $\pi'$ is the link pattern in which the closed arch connected to $1$ is cut into two open arches. $C$-type $q$KZ equation ----------------------- To each simple root we attach respectively the standard trigonometric $R$-matrices $R_i(w_{i+1}/w_i)$, $i=1,2,\ldots,r-1$ of Eq. , and the boundary $R$-matrix $R_r(1/w_r^2)\equiv R_{\alpha_r}$, with the same expression. The level one $C$ $q$KZ equation consists of the following system $$\begin{aligned} R_i(w_{i+1}/w_i)\Psi&&=\tau_i\Psi \label{CqKZa}\\ w_r^{-m_r}R_r(1/w_r^2)\Psi&&=\tau_r\Psi\label{CqKZb}\end{aligned}$$ where as usual $\tau_i$ acts by interchanging the spectral parameters $w_i$ and $w_{i+1}$, $i=1,2,...,r-1$ and $\tau_r$ acts on $\Psi$ by letting $w_r\to 1/w_r$, and $m_r$ is the degree of $\Psi$ in $w_r$. These are finally supplemented by the affinization relation, obtained by considering an extra root, say $\alpha_1'=-z_1-z_2$, and the associated boundary operator $R_1'(q^6w_1w_2)$: $$R_1'(q^6w_1w_2)\Psi=\tau_1'\Psi\label{CqKZc}$$ where $\tau_1'$ interchanges $w_2$ and $1/(q^6 w_1)$, and $R'_1$ is of the form of Eq.  with $e'_1$ in place of $e_i$. Using $R'_1(w)=s R_1(w) s$, the relation can also be recast into $$(q^3 z_1)^{-m_1} s\Psi(w_1,\ldots,w_r)=\Psi\Big({1\over q^6 w_1},w_2,\ldots,w_r)\label{CqKZd}$$ The generators of the weight lattice (as abelian subgroup of $\hat W$) are very similar to the generators (i) of the case $B_r$: the only change concerns the boundary operators ${\bf K}_1$ and ${\bf K}_r$ now implementing Eqs.  and . The polynomial solution $\Psi$ to the level one $C_r$ $q$KZ system has degree $m_1=m_r=2n$ in each variable, total degree $n(2n+1)$ and base entry $$\Psi_{\pi_0}=\prod_{1\leq i<j\leq 2n+1} (qz_i-q^{-1}z_j)$$ and all the other entries of $\Psi$ may be obtained in a triangular way from this one. Example: for $r=3$, we have the following minimal polynomial solution to the level one $C_3$ $q$KZ system: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0c.eps,width=0.9cm}} =(qw_1-q^{-1}w_2)(qw_1-q^{-1}w_3)(qw_2-q^{-1}w_3)\\ &&\Psi_{\epsfig{file=arch1c.eps,width=0.9cm}} =(qw_1-q^{-1}w_2)(q^2 w_1w_2-q^{-2})(q^{-1}-q w_3^2)\\ &&\Psi_{\epsfig{file=arch2c.eps,width=0.9cm}} =(q^3w_1^2-q^{-3})(qw_2-q^{-1}w_3)(qw_2w_3-q^{-1}) \\\end{aligned}$$ which, upon taking the rational limit yields the multidegrees: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0c.eps,width=0.9cm}} =(A+z_1-z_2)(A+z_1-z_3)(A+z_2-z_3)\\ &&\Psi_{\epsfig{file=arch1c.eps,width=0.9cm}} =(A+z_1-z_2)(2A+z_1+z_2)(A+2z_3)\\ &&\Psi_{\epsfig{file=arch2c.eps,width=0.9cm}} =(3A+2z_1)(A+z_2-z_3)(A+z_2+z_3) \\\end{aligned}$$ and the degrees $\Psi=(1,2,3)$ for $A=1$ and $z_i=0$. RS point and CSSCPP ------------------- At the point $q=e^{2i\pi/3}$, $\Psi$ may be viewed as the ground state eigenvector of a transfer matrix, corresponding in the homogeneous limit to the Hamiltonian $$H_C={e_1+e_1'\over 2}+\sum_{i=2}^{r-1}e_i +e_r$$ Normalizing $\Psi$ so that its smallest entry $\Psi_{\pi_0}=1$, we have been able to compute the sum of entries to be $A(n)A(n+1)$. In the case of even $r=2n$, the above may be repeated almost identically: in the presence of spectral parameters, the even case may be recovered from the odd one by taking $w_{2n+1}\to -q^{-1}$, and dividing out the result by $\prod_{1\leq i\leq 2n} (1+q^3w_i)$. Indeed, this specialization leaves us with only non-vanishing components whith an open arch at the rightmost point, in bijection with open link patterns with that point erased, hence the projection onto the case of size one less. This leads us to the $$\sum_\pi \Psi_\pi =A(\lfloor r/2\rfloor)A(\lceil r/2 \rceil)$$ Note that the sum in the even case, $A(n)^2$, also counts the Cyclically Symmetric Self-Complementary Plane Partitions (CSSCPP) in an hexagon of size $2n\times 2n\times 2n$ [@Bre]. Also note the determinant formulae $A(n)^2=\det\Big({i+j\choose 2i-j-1}+{i+j+1\choose 2i-j}\Big)_{0\leq i,j\leq n-1}$ and $A(n)A(n+1)=\det\Big({i+j+1\choose 2i-j}+{i+j+2\choose 2i-j}\Big)_{0\leq i,j\leq n-1}$. Furthermore, consider the left eigenvector $v$ of $H_C$ with the same eigenvalue ($r$ for $r$ odd, $r+1/2$ for $r$ even). Normalize $v$ so that its entries are coprime positive integers. We have found empirically the following $$\sum_\pi v_\pi \Psi_\pi = A(r)\ .$$ Finally, we formulate the The largest entry of $\Psi$ for $C_r$ is the sum of entries for $C_{r-1}$. Example: at $r=5$, $\Psi=(1,2,3,3,0,1,4,0,0,0)$, $v=(48,36,28,34,24,23,25,18,17,14)$, $\sum_\pi \Psi_\pi=14=2\times 7=A(2)A(3)$, $\sum_\pi v_\pi \Psi_\pi=429=A(5)$, and the maximal entry of $\Psi$ is $4=A(2)^2$. $D_r$ case ========== The simple roots of $D_r$ are $\alpha_i=z_i-z_{i+1}$ for $i=1,2,\ldots,n-1$ and $\alpha_r=z_{r-1}+z_r$. We concentrate on the odd case $r=2n+1$, and consider again the fundamental representation of dimension $N=2r$. Just like in the $B_r$ case, one choice is to select upper triangular matrices satisfying $M^T J+JM=0$, $J$ antidiagonal matrix with $1$’s on the second diagonal. Orbital varieties and $D$-type Temperley–Lieb algebra ----------------------------------------------------- Just as in the case $C$, there are $r\choose\lfloor{r+1\over2}\rfloor$ orbital varieties, indexed by open link patterns. We now deal with $D$-type Temperley–Lieb algebras, denoted $DTL(\beta)$, with generators $e_i$, $i=1,2,...,r-1$ obeying the $TL(\beta)$ relations and an extra generator $e_{r-1}'$, satisfying the relations: $$(e_{r-1}')^2=\beta e_{r-1}, \qquad e_{r-1}e_{r-1}'=e_{r-1}'e_{r-1}=e_{r-2}e_{r-1}'e_{r-2}-e_{r-2}=e_{r-1}'e_{r-2}e_{r-1}'-e_{r-1}'=0$$ These operators act on open link patterns as follows. The $e_i$, $i=1,2,\ldots,r-1$ act in the usual way, by creating a little arch between points $i$ and $i+1$ and by gluing the two former points. To describe the action of $e_{r-1}'$, let us first connect the open arches of the open link patterns by pairs of consecutive open arches from the left to the right, and represent the newly formed arches in a different color (dashed lines, cf Fig. \[linkpaD\] for the $D_5$ example). We then define an involution $s$ on open link patterns that simply switches the color (solid $\leftrightarrow$ dashed) of the rightmost arch if it is closed, and leaves it invariant if it is open. Then $e_{r-1}'=s e_{r-1}s$. Finally, we introduce an extra boundary operator $e_0$, which is the right half of an $e$ (like a reflected $e_r$ of $C_r$), with its open end connected to the vertical line, and acts as such, with the same rules as for $C_r$, but upon reflection of indices $i\leftrightarrow r-i$. It satisfies the relations: $e_0^2=\beta e_0$ and $e_1e_0e_1=2 e_1$. $D$-type $q$KZ equation ----------------------- We associate to the roots the $R$-matrices $R_i(w_{i+1}/w_i)$ of Eq. , and $R_r(1/(w_r w_{r-1}))$ defined by the same equation in which $e_i$ is replaced with $e'_{r-1}$, so that $R_r(w)=s R_{r-1}(w) s$. The level one $D$ $q$KZ equation consists of the following system $$\begin{aligned} R_i(w_{i+1}/w_i)\Psi&&=\tau_i\Psi, \qquad i=1,2,...,r-1\\ R_r(1/(w_rw_{r-1}))\Psi&&=\tau_{r-1}'\Psi\end{aligned}$$ where as usual $\tau_i$ acts by interchanging the spectral parameters $w_i$ and $w_{i+1}$, $i=1,2,...,r-1$ and $\tau_r'$ acts on $\Psi$ by interchanging $w_{r-1}$ and $1/w_r$. Upon using the above relation $e_{r-1}'=s e_{r-1}s$, the latter equation may be equivalently replaced by $$z_r^{-m_r} s\Psi(z_1,\ldots,z_r)=\Psi\Big(z_1,\ldots,z_{r-1},{1\over z_r}\Big)$$ These are finally supplemented by the affinization relation, obtained by considering the extra root $\alpha_0=-2z_1$, and the associated boundary operator $R_0(q^6w_1^2)$ involving the extra operator $e_0$: $$w_1^{-m_1}R_0(q^6 w_1^2)\Psi=\tau_0\Psi$$ where $\tau_0 f(w_1)=f( 1/(q^6 w_1))$ and $m_1$ the degree of $\Psi$ in $w_1$. The construction of the abelian subgroup of $\hat W$ is similar to the cases $B$ and $C$, and is skipped for the sake of brevity. The minimal degree polynomial solution to the level one $D_r$ $q$KZ system has total degree $r(r-1)/2$ and partial degree $m_1=m_r=r-1$ in all variables. Its base entry, corresponding to the open link pattern $\pi_0$ with only open arches reads $$\Psi_{\pi_0}=\prod_{1\leq i<j\leq 2n+1} (qz_i-q^{-1}z_j)$$ and all the other entries of $\Psi$ may be obtained in a triangular way from this one. Example: for $r=3$, we have the following minimal polynomial solution to the level one $D_3$ $q$KZ system: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0c.eps,width=0.9cm}} =(qw_1-q^{-1}w_2)(qw_1-q^{-1}w_3)(qw_2-q^{-1}w_3)\\ &&\Psi_{\epsfig{file=arch1c.eps,width=0.9cm}} =(qw_1-q^{-1}w_2)(q w_1w_3-q^{-1})(q w_2w_3-q^{-1})\\ &&\Psi_{\epsfig{file=arch2c.eps,width=0.9cm}} =(q^{-2}-q^{2}w_1^2)(qw_2-q^{-1}w_3)(qw_2w_3-q^{-1}) \\\end{aligned}$$ which, upon taking the rational limit gives the multidegrees: $$\begin{aligned} &&\Psi_{\epsfig{file=arch0c.eps,width=0.9cm}} =(A+z_1-z_2)(A+z_1-z_3)(A+z_2-z_3)\\ &&\Psi_{\epsfig{file=arch1c.eps,width=0.9cm}} =(A+z_1-z_2)(A+z_1+z_3)(A+z_2+z_3)\\ &&\Psi_{\epsfig{file=arch2c.eps,width=0.9cm}} =2(A+z_1)(A+z_2-z_3)(A+z_2+z_3) \\\end{aligned}$$ and the degrees $\Psi=(1,1,2)$ for $A=1$ and $z_i=0$. RS point and HTASM ------------------ At the point $q=e^{2i\pi/3}$, $\Psi$ may be viewed as the Perron–Frobenius eigenvector of a transfer matrix, corresponding in the homogeneous limit to the Hamiltonian $$H_D=e_0+\sum_{i=1}^{r-2}e_i +{e_{r-1}+e_{r-1}'\over 2}$$ Note that upon the reflection $e_i\to e_{r-i}$, this Hamiltonian is mapped onto $H_C$: we are dealing with the same algebra, but in different representations. Going to the RS point $q=e^{2i\pi/3}$ and taking the homogeneous limit $w_i=1$ for all $i$, and normalizing $\Psi$ so that its smallest entry is $\Psi_{\pi_0}=1$, we have found the The sum of entries $\sum_\pi \Psi_\pi$ is the number of Half-Turn Symmetric Alternating Sign Matrices of size $r$, $A_{HT}(r)$. This conjecture also works in the even case $r=2n$, which may be obtained from the odd one by taking $z_1=-q^{-2}$, shifting all remaining spectral parameters $w_i\to w_{i-1}$, $i=2,3,...,2n+1$, and dividing out by $\prod_{1\leq i\leq 2n} (1+z_i)$. Note the formulae $A_{HT}(2n)=\det\Big({i+j\choose 2i-j}+{i+j+1\choose 2i-j}\Big)_{0\leq i,j\leq n-1}$ and $A_{HT}(2n+1)=\det\Big({i+j+1\choose 2i-j}+{i+j+2\choose 2i-j+1}\Big)_{0\leq i,j\leq n-1}$. Introduce as before the left Perron–Frobenius eigenvector $v$ of $H_D$ with coprime positive integer entries. $$\sum_\pi v_\pi \Psi_\pi = A(r)\ .$$ Finally, we also find the The largest entry of $\Psi$ for $D_r$ is the sum of entries for $C_{r-1}$. Example: at $r=5$, $\Psi=(1,1,3,4,2,3,1,4,2,4)$, $v=(10,10,17,14,18,17,23,14,18,25)$, $\sum_\pi \Psi_\pi=25=A_{HT}(5)$, $\sum_\pi v_\pi \Psi_\pi=429=A(5)$, and the maximal entry of $\Psi$ is $4=A(2)^2$, the sum of the components of the $C_4$ solution. [A]{} M.T. Batchelor, J. de Gier and B. Nienhuis, The quantum symmetric XXZ chain at $\Delta=-1/2$, alternating sign matrices and plane partitions, [*J. Phys.*]{} A34 (2001) L265–L270, cond-mat/0101385. A.V. Razumov and Yu.G. Stroganov, Spin chains and combinatorics, [*J. Phys*]{} A34 (2001), 3185, cond-mat/0012141; Spin chains and combinatorics: twisted boundary conditions, [*J. Phys*]{} A34 (2001), 5335, cond-mat/0012247. A.V. Razumov and Yu.G. Stroganov, Combinatorial nature of ground state vector of $O(1)$ loop model, [*Theor. Math. Phys.*]{} [**138**]{} (2004) 333–337; [*Teor. Mat. Fiz.*]{} 138 (2004) 395–400, math.CO/0104216. P. Di Francesco and P. Zinn-Justin, Around the Razumov–Stroganov conjecture: proof of a multi-parameter sum rule, [*E. J. Combi.*]{} 12 (1) (2005), R6, math-ph/0410061. G. Kuperberg, Another proof of the alternating sign matrix conjecture, [*Int. Math. Research Notes*]{} (1996) 139–150, math.CO/9712207. V. Pasquier, Quantum incompressibility and Razumov Stroganov type conjectures, cond-mat/0506075. P. Di Francesco and P. Zinn-Justin, Quantum Knizhnik–Zamolodchikov equation, generalized Razumov–Stroganov sum rules and extended Joseph polynomials, to appear in [*J. Phys.*]{} A, math-ph/0508059. A. Knutson and P. Zinn-Justin, A scheme related to the Brauer loop model, math.AG/0503224. P. Di Francesco and P. Zinn-Justin, Inhomogeneous model of crossing loops and multidegrees of some algebraic varieties, to appear in [*Commun. Math. Phys.*]{} (2005), math-ph/0412031. G. Kuperberg, Symmetry classes of alternating-sign matrices under one roof, [*Ann. of Math.*]{} (2) 156 (2002), no. 3, 835–866, math.CO/0008184. A. Joseph, On the variety of a highest weight module, [*J. Algebra*]{} 88 (1) (1984), 238–278. R. Hotta, On Joseph’s construction of Weyl group representations, Tohoku Math. J. Vol. 36 (1984), 49–74. I.B. Frenkel and N. Reshetikhin, Quantum affine Algebras and Holonomic Difference Equations, [*Commun. Math. Phys.*]{} 146 (1992), 1–60. M. Jimbo and T. Miwa, Algebraic analysis of Solvable Lattice Models, CBMS Regional Conference Series in Mathematics vol. 85, American Mathematical Society, Providence, 1995. P. Di Francesco, A. Knutson and P. Zinn-Justin, Extended Orbital Varieties and the Yang–Baxter equation, work in progress. P. Di Francesco, Boundary $q$KZ equation and generalized Razumov–Stroganov sum rules for open IRF models, math-ph/0509011. P. Di Francesco, Inhomogenous loop models with open boundaries, [*J. Phys.*]{} A 38 (2005), 6091–6120, math-ph/0504032. N. Chriss and V. Ginzburg, Representation Theory and Complex Geometry, Birkhauser 1997. V. Pasquier, Scattering matrices and Affine Hecke Algebras, q-alg/9508002. A. Izergin, Partition function of the six-vertex model in a finite volume, [*Sov. Phys. Dokl.*]{} [**32**]{} (1987) 878-879. V. Korepin, Calculation of norms of Bethe wave functions, [*Comm. Math. Phys.*]{} [**86**]{} (1982) 391-418. S. Okada, Enumeration of Symmetry Classes of Alternating Sign Matrices and Characters of Classical Groups, math.CO/0408234. R.M. Green, Generalized Temperley–Lieb algebras and decorated tangles, [*Journal of Knot Theory and its Ramifications*]{} 7 (1998), 155–171. D. Bressoud, Proofs and confirmations. The story of the alternating sign matrix conjecture, Cambridge University Press (1999). B. Rothbach, unpublished. A. Melnikova, Description of B-orbit closures of order 2 in upper-triangular matrices, math.RT/0312290, to appear in [*Transformation Groups*]{}.

--- abstract: 'Protein aggregation in cell membrane is vital for the majority of biological functions. Recent experimental results suggest that transmembrane domains of proteins such as $\alpha$-helices and $\beta$-sheets have different structural rigidities. We use molecular dynamics simulation of a coarse-grained model of protein-embedded lipid membranes to investigate the mechanisms of protein clustering. For a variety of protein concentrations, our simulations under thermal equilibrium conditions reveal that the structural rigidity of transmembrane domains dramatically affects interactions and changes the shape of the cluster. We have observed stable large aggregates even in the absence of hydrophobic mismatch, which has been previously proposed as the mechanism of protein aggregation. According to our results, semi-flexible proteins aggregate to form two-dimensional clusters, while rigid proteins, by contrast, form one-dimensional string-like structures. By assuming two probable scenarios for the formation of a two-dimensional triangular structure, we calculate the lipid density around protein clusters and find that the difference in lipid distribution around rigid and semiflexible proteins determines the one- or two-dimensional nature of aggregates. It is found that lipids move faster around semiflexible proteins than rigid ones. The aggregation mechanism suggested in this paper can be tested by current state-of-the-art experimental facilities.' author: - 'Hamidreza Jafarinia$^{1}$' - Atefeh Khoshnood$^2$ - 'Mir Abbas Jalali$^{3}$' bibliography: - 'ref.bib' title: Rigidity of transmembrane proteins determines their cluster shape --- Introduction {#sec:intro} ============ Transmembrane (TM) proteins are regulators of several cellular processes. In order to perform their functions, they often aggregate and distribute non-uniformly in the cell membrane [@sieber2007anatomy]. The aggregation process of proteins sometimes becomes abnormal, and causes amyloid diseases [@chiti2006protein]. In recent years, there has been increasing interest in identifying various mechanisms that affect protein–membrane interactions [@mouritsen1984mattress; @west2009membrane] and, consequently, the aggregation behavior of different proteins. The specific structure of TM proteins defines their physical properties and enables them to perform their biological functions [@buehler2006nature]. TM proteins differ in size and physical properties [@becker2009biogenesis], and may have single or multiple $\alpha$-helical [@bocharov2010structure; @westfield2011structural] or $\beta$-structure [@tang2013desalination] domains. Recent experimental studies show that $\alpha$-helical structures have softer domains than $\beta$-structures in both dry and hydrated states, and $\beta$-barrels and $\beta$-sheets are more rigid structural units than $\alpha$-helices [@perticaroli2014rigidity]. The higher number of hydrogen bonds per residue is also considered to be the probable cause of the more rigid structure of $\beta$-sheets compared to $\alpha$-helices [@perticaroli2013secondary]. It has also been shown that the secondary structure of proteins affects the rigidity and dynamics of the protein. Proteins containing $\beta$-structures have a higher Young’s modulus and higher frequency of the collective vibration [@perticaroli2013secondary]. Consequently, the effect of the class and structural rigidity of proteins on their aggregation behavior and biological function cannot be ignored. Although lipid raft formation [@mcintosh2006roles] and direct linking [@feng1998dual] of proteins cause the aggregation of membrane proteins, membrane curvature [@bahrami2011vesicle] and membrane-mediated interactions play important roles on the formation and fragmentation of protein clusters. Among mechanisms that generate lipid-mediated protein interactions, the hydrophobic mismatch interaction, which is due to the difference between the hydrophobic lengths of the integral proteins and the hydrophobic thickness of their host membrane, has been widely studied by several groups [@venturoli2005simulation; @schmidt2008cluster; @de2008molecular; @west2009membrane]. However, we know little about the effect of the structural properties of proteins on the cluster formation. It is known that the shapes and sizes of proteins determine the distribution of their surrounding lipid molecules [@de2008molecular; @schmidt2010hydrophobic; @morozova2012shape; @yoo2013membrane], which in turn affect the stability and function of proteins [@jensen2004lipids; @lee2004lipids]. What is poorly understood is how the structural rigidity of proteins integrates with other factors to shape the patterns of aggregates. In this paper, we use coarse-grained molecular dynamics simulations to systematically investigate the effect of structural rigidity of TM proteins on the formation of clusters. We design specific model proteins to exclude the effect of hydrophobic mismatch and isolate the role of structural rigidity. We use two sets of proteins: semiflexible and rigid. In §\[sec:model\], we present our model and simulation method and setup. In §\[sec:results\], the results of molecular dynamics simulations are presented. We discuss our findings in §\[sec:discuss\] and show how they are comparable with experimental observations and previous theoretical modelings. Our simulations show that proteins form clusters even in the absence of hydrophobic mismatch and the final shapes of protein aggregates in lipid bilayers depend strongly on the rigidity of proteins. Model and Methods {#sec:model} ================= The model of lipid molecules is composed of one hydrophilic head particle and a hydrophobic tail chain [@goetz1998computer; @reynwar2007aggregation], which contains four particles. Our TM proteins are modeled as hexagonal prisms with middle hydrophobic particles and hydrophilic groups at both ends [@schmidt2008cluster; @schmidt2010hydrophobic]. The hydrophobic mismatch is tuned by changing the length $\Delta r=r_p-r_l$ of the hydrophobic part of proteins, where $r_{p}$ is the length of the hydrophobic part of proteins, and $r_{l}$ is the average thickness of the hydrophobic part of the bilayer. Models of lipid and protein molecules and their corresponding bonds are displayed in Fig. \[fig1\]. Hydrophilic and hydrophobic particles are labeled H and T, respectively. In this study, our length scale is $\sigma=1/3$ nm and the energy unit is $N_{\rm A}\epsilon=2$ kJ/mol, with $N_{\rm A}$ being the Avogadro number. In each lipid or protein molecule, the adjacent $i$th and $(i+1)$th particles interact through the harmonic bond potential $$\label{six} U_b(r_{i,i+1})=k_b(r_{i,i+1}-\sigma_{eq})^2,$$ where $r_{i,i+1}$ is the distance between particles and $\sigma_{eq}$ is the equilibrium length of the bonds. Three kinds bonds— planar, vertical, and oblique—are used to build two types of protein molecules—rigid and semiflexible. The planar and vertical bonds have identical spring constants of $k_{b1}=5000 \, \epsilon/\sigma^{2}$, which is fixed in all simulations. For the oblique bonds of rigid proteins (RPs) and semiflexible proteins (SFPs) we set $k_{b1}=5000 \, \epsilon/\sigma^{2}$ and $k_{b2}=35 \, \epsilon/\sigma^{2}$, respectively. Both SFPs and RPs are more rigid than lipid molecules. We set $\sigma_{eq}=\sigma$ for lipid bonds and planar and vertical bonds of proteins. For oblique protein bonds, we use $\sigma_{eq}=\sqrt{2} \, \sigma$. SFPs are stiffer than lipid molecules and mostly maintain their hexagonal cross section when they are bent due to interactions with other proteins and lipids. The angle between consecutive bonds in a lipid molecule is controlled by $$U=k_{a}(\cos\theta-cos\theta_{eq})^2,$$ where $k_{a}=1.85\,\epsilon$ and $\theta_{eq}=\pi$. Interactions between the particles of different molecules are governed by soft-core and Lennard-Jones potentials, defined as $$\begin{aligned} U_{\rm sc}(r_{ij}) &=& 4\epsilon \left (\frac{\sigma_{sc}}{r_{ij}} \right )^9, \label{eq1} \\ U_{\rm LJ}(r_{ij}) &=& 4\epsilon \left [\left (\frac{\sigma}{r_{ij}} \right )^{12}-\left ( \frac{\sigma}{r_{ij}} \right )^6 \right ], \label{eq2}\end{aligned}$$ where $r_{ij}=\vert {\textit{\textbf{r}} }_i-{\textit{\textbf{r}} }_j \vert$ and $\sigma_{\rm sc}=1.05 \, \sigma$. In these equations, ${\textit{\textbf{r}} }_i$ is the global position vector of the $i$th particle. The repulsive soft-core potential is used to model the interaction between hydrophobic and hydrophilic particles, and between solvent and hydrophobic particles. All other interactions are modeled by the Lennard-Jones potential. A cutoff radius of $r_c=2.5 \, \sigma$ is applied to $U_{\rm sc}$ and $U_{\rm LJ}$, which are then shifted in order to vanish at $r_{ij}=2.5 \, \sigma$ [@goetz1998computer]. This guarantees the continuity of both potential fields. We carry out molecular dynamics simulations of $NVT$ ensembles using the package. Periodic boundary conditions are imposed and the temperature is kept constant at $T_0=310$ K (with $k_B T=1.29 \, \epsilon$) utilizing the Nose-Hoover thermostat. Here $k_B$ is the Boltzmann constant and the lipid bilayer is in the liquid phase. All particles have the same mass, $N_{\rm A}m=36$ gr/mol [@goetz1998computer]. The integration time step is set to $\Delta t=0.005 \, \tau$, where $\tau=\sqrt {m\sigma^2/\epsilon}$ is the intrinsic time scale. The actual value of $\Delta t$ is $7.07$ fs. In all simulations, the number density of particles is $n=0.66/\sigma^{3}$ and the area per lipid (for a bilayer without proteins) is $A_s=2.09 \, \sigma^{2}$. $A_s$ is the area of lipid bilayer divided by the number of lipids. We select the number of lipids such that bilayers with minimum surface tension and without permanent curvatures are obtained. This helps us to eradicate the effect of curvature-mediated interactions. Under this condition the surface tension of the bilayer is positive and approximately equals $0.24\, \epsilon\sigma^{-2}$, corresponding to 7 mN/m. This has led to stable bilayers in all of our simulations. It is remarked that the rupture surface tension of biological membranes varies from 1 to 30 mN/m, and it depends on the lipid composition of the bilayer [@evans2003dynamic; @needham1990elastic]. Simulations are performed for various concentrations of proteins embedded in the bilayer. We denote the concentration of proteins $c_p=N_{p}/(N_{p}+N_{l})$, where $N_{p}$ and $N_{l}$ are the numbers of protein and lipid particles, respectively. At the beginning of simulations, each protein molecule is placed in the bilayer by removing nine lipid molecules. Results {#sec:results} ======= We investigate the clustering process for low and high protein concentrations and for two protein types: RPs and SFPs. To quantify the flexibilities of proteins, we have compared the longitudinal flexibilities of our protein models with each other and, also, with a lipid patch that occupies the same area that proteins do. The torsional flexibilities of protein models have also been computed. We have carried out simulations for a single RP and SFP in vacuum as well as in a lipid bilayer and calculated the standard deviation of the length of the hydrophobic part for each protein. A bilayer consisting of nine lipid molecules in a box of size $3.06 \sigma \times 3.06 \sigma \times 10\sigma$ was simulated. The area of this bilayer patch approximately equals the area of proteins, and it has the same area per lipid ($A_s=2.09 \sigma^2$) as other bilayers in our simulations. Let us define the lengths of the hydrophobic parts of the RP and SFP by $ r_{\rm RP}$ and $r_{\rm SFP}$, respectively. Denoting the standard deviation $SD(\cdot)$, we find $$\begin{aligned} SD(r_{\rm RP})= 0.025\sigma,~~SD(r_{\rm SFP})=0.067\sigma, \end{aligned}$$ for proteins in vacuum, $$\begin{aligned} SD(r_{\rm RP})= 0.025\sigma,~~SD(r_{\rm SFP})=0.055 \sigma,\end{aligned}$$ for proteins in the bilayer, and $SD(r_l)=0.32 \sigma$ for the lipid bilayer. It is seen that transmembrane RPs are stiffer than SFPs, and SFPs are stiffer than a bilayer that has the same surface area of proteins. Due to the lack of interactions with lipid and solvent molecules, SFPs are more flexible in vacuum. The weaker oblique bonds result in more torsional flexibility for SFPs. The torsional rigidity of proteins can be measured by the relative twist angle of their two hydrophilic ends. Defining $\theta_{\rm RP}$ and $\theta_{\rm SFP}$ as the relative twist angles of the upper and lower hydrophobic parts, we obtain $$\begin{aligned} SD(\theta_{\rm RP})= 0.11^{\circ},~~SD(\theta_{\rm SFP})=0.28^{\circ},\end{aligned}$$ for proteins in vacuum and $$\begin{aligned} SD(\theta_{\rm RP})= 0.11^{\circ},~~SD(\theta_{\rm SFP})=0.22^{\circ}, \end{aligned}$$ for proteins in the bilayer. These results show that RPs are torsionally stiffer than SFPs. Our findings are consistent with the higher conformational displacements of $\alpha$-helices compared to $\beta$-sheets [@gaspar2008dynamics; @perticaroli2013secondary]. In all of our simulations, the effect of hydrophobic mismatch is neutralized, $\Delta r=0.01$, by using the $HT_{5}H$ model proteins. Figure \[fig2\] shows snapshots of protein-embedded membranes with SFPs and RPs. We have used $N=3$, 4, 6, and 8 protein molecules for both protein types. The size of the lipid bilayer is $16\,\sigma \times 16\,\sigma$ for $N=3$, $23\,\sigma \times 23\,\sigma$ for $N=4$ and 6, and $35\,\sigma \times 35\,\sigma$ for $N=8$. The snapshots have been taken at $t=5 \times 10^{6}\Delta t$. The difference between SFP and RP clusters is prominent: RPs form string-like one-dimensional structures, while SFPs form two-dimensional clusters. These results differ significantly from previous work [@venturoli2005simulation], which suggests that there is only weak attraction between inclusions in the absence of mismatch. Our results, clearly, show that protein inclusions form stable clusters in the absence of hydrophobic mismatch. One of the distinct features of SFPs is their clustering in triangular structures. For the model with six and eight SFPs, two triangular structures can be seen in Fig. \[fig2\]. RPs do not share this property. To quantify the discrepancies between one-dimensional and two-dimensional clusters, we measure the radius of gyration of protein structures as $$R_g^2=\frac{1}{M}\sum_{i=1}^{N_{p}} m_i \left | {\textit{\textbf{r}} }_i-{\textit{\textbf{r}} }_{\rm c} \right | ^2, \label{Rg}$$ where $N_{p}$ and $M$ are the total number and the total mass of the head particles of proteins, respectively, and $r_{\rm c}$ and ${\textit{\textbf{r}} }_i$ define the position vectors of the $i$th particle and the center of mass of protein heads, respectively. The mass of each particle is denoted $m_{i}$. To compute $R_g$, we have used protein heads lying in one monolayer. The variation of $R_g$ over time has been plotted in Fig. \[fig3\] for a system with four and seven proteins. We have also plotted the time-averaged radius of gyration, $\langle R_g \rangle =\frac 1{t_2-t_1} \int_{t_1}^{t_2} R_g(t) \, dt$, for several number of proteins in a cluster. We have used $(t_1,t_2)=(30,35\,{\rm ns})$ for both RPs and SFPs. It is shown that for the same number of proteins, two-dimensional structures formed by SFPs always have a lower radius of gyration; the radius of gyration increases proportionally to the number of proteins in the cluster. Since the linear aggregates have more contact and interaction with surrounding lipids, these structures are more likely to change their shape slightly, for example, from a completely straight linear structure to a curved line, which, in turn, alters the radius of gyration. As Fig. \[fig2\] shows, string-like clusters have a variety of configurations with large variations in their gyration radii (see Fig. \[fig3\]). In simulations of SFPs we have observed deformed cross sections of proteins (deviations from hexagonal shapes), especially when they do not belong to a cluster. Deformations of proteins can occasionally stabilize them in the membrane, so that they do not participate in cluster forming processes. So SFPs are stabilized by either contributing to a cluster or undergoing deformation while they are singly dispersed in the membrane. It is remarked that we have repeated our simulations starting with different initial conditions and observed similar clustering trends for each protein class. Moreover, we continued our simulations over longer time scales, up to 200 ns, and obtained the same results as in our 35-ns-long simulations for both SFP and RP clusters. Figure \[fig4\] shows the temporal evolution of cluster formation from an initially random distribution of proteins. Interestingly, RPs immediately aggregate to one-dimensional clusters, whereas SFPs first make a cluster with several branches separated by lipid molecules, then evolve to a compact cluster as trapped lipids are released. To eliminate possible boundary effects in simulation boxes with periodic boundary conditions, and also understand how the concentration of proteins affect the aggregation process, we have carried out simulations in a larger membrane, of size $90 \sigma \times 90 \sigma$ with higher protein concentrations, $c_p=0.2$, 0.27, 0.35, and 0.48. We have studied several snapshots of these simulations and our previous conclusions for smaller membranes are unaltered. A few more results for high-$c_p$ simulations are as follows: (i) SFPs rarely form one-dimensional clusters, though the lengths of such clusters are much shorter than the aggregates of RPs; (ii) one-dimensional structures of RPs may connect to each other to form longer or branched web-like structures; and (iii) on rare occasions RPs cluster as two-dimensional domains. Observations i and iii might be due to the longer relaxation time scales of large membranes with a high $c_p$. The structure formation process by RPs and SFPs is better understood by computing the average number of neighboring proteins, $N_{\rm av}$. For a given (subject) protein $i$, we find the number of neighboring proteins $N_{i}$ within a distance of $6.1\, \sigma$, measured from the center of mass of the protein. We then calculate the average number $N_{\rm av}=\frac 1N \sum_{i=1}^{N} N_i$. Figure \[fig5\] shows the variation of $N_{\rm av}$ over time for two protein types and several protein concentrations. It is shown that $N_{\rm av,SFP}$ is consistently larger than $N_{\rm av,RP}$, and the difference $\Delta N_{\rm av}=N_{\rm av,SFP}-N_{\rm av,RP}$ increases versus time as the aggregate size and structure reach steady state. We note that the time for $N_{\rm av}$ to reach the steady-state value is shorter in systems with RPs and at higher protein concentrations. We explain the physical origin of different cluster-forming pathways of SFPs and RPs by investigating the distribution of lipids around proteins [@morozova2012shape; @de2008molecular]. Our numerical simulations show that the lipid head and tail densities around a subject protein depend on the rigidity of the protein: RPs induce order and structure in surrounding lipids, which have formed ring-like structures around proteins (Fig. \[fig6\]). The lipid density around rigid inclusions is generally lower around SFPs because the thermally vibrating flexible structure of relatively massive proteins continuously kicks and scatters lighter lipid molecules. For a complex of two proteins, the ordered structure of lipids takes an oval shape and constrains the formation of a complex with three proteins. To investigate how the process works, we designed two experimental scenarios. In the first scenario, two proteins are kept close to each other almost at the center of the lipid membrane by means of a spring of constant $500\,\epsilon/\sigma^2$. We then fix the position of the third protein on the double-complex vertical symmetry line at $d=4\,\sigma$ \[Fig. \[fig6\](a)\] with the same spring constant. The next step is to obtain the lipid distribution around the three-protein complex. When the triple-complex is composed of RPs, the high density of trapped lipids between the third protein and the double ones prohibits the formation of a triple, triangular-shaped complex as Fig. \[fig6\](b) shows. For SFPs, however, lipids are weakly bound to the double-complex, easily diffuse out of the triple-contact area, and thus facilitate the formation of bigger two-dimensional clusters if the constraint is released \[Fig. \[fig6\](c)\]. In the second scenario, the third protein is placed close to the double-complex and along an oblique line with an angle of $\theta=2\pi/3$ \[Fig. \[fig6\](d)\], forming a string-like structure. In the case of three RPs, lipids are trapped between two proteins and have completely filled the region between the two adjacent proteins. Therefore, these lipids resist the reduction of $\theta$ if the constraint is released. This is how RPs maintain their one-dimensional structure \[Fig. \[fig6\](e)\]. Note that in Figs. \[fig6\](a) and 6(d), the arrows show the most probable path that the third protein chooses to follow in order to form a triangular structure if the constraint is released. The distribution of lipids around SFPs is more homogeneous than that around RPs; compare Figs. \[fig6\](e) and 6(f). Our simulations with three proteins show that lipid molecules more easily diffuse in the space between SEPs and result in different arrangements of SEPs compared to RPs. More diffusive lipids around SFPs can also be identified by analyzing the mean squared displacement of lipids. For $c_p=0.48$, we have calculated the mean squared displacement for lipid molecules using the method in [@khoshnood2013anomalous]. Let us define ${\textit{\textbf{r}} }_{\parallel}(t)$ as the component of the position vector ${\textit{\textbf{r}} }(t)$ of lipid particles parallel to the bilayer surface. The two-dimensional diffusion coefficients $D=\lim_{t\rightarrow \infty}\langle \vert {\textit{\textbf{r}} }_{\parallel}(t)-{\textit{\textbf{r}} }_{\parallel}(0) \vert ^2\rangle /(4 t)$ that we find are approximately $43.1 \times 10^{-7}$ and $56.2\times 10^{-7} \, {\rm cm}^{2}/s$ for the ensemble of lipid molecules around RPs and SFPs, respectively. The diffusion coefficients are different from experimental data because the model is coarse grained. Here the relative change in diffusion coefficients is important. Our results show a $30\%$ reduction in diffusion coefficient for a membrane with RPs in comparison with one hosting SEPs. This is consistent with our observation of more restricted lipids surrounding RPs. We note that diffusion coefficients are measured from the part of the mean squared displacement profile that is linear in time. The initial anomalous region is not included in the calculations [@khoshnood2013anomalous]. Discussion {#sec:discuss} ========== Using coarse-grained molecular dynamic simulations, we showed that the rigidity of proteins has a profound effect on the cluster shape of TM proteins in lipid membranes. For proteins with zero hydrophobic mismatch, regardless of the protein concentration, RPs aggregate to form one-dimensional clusters while SFPs form two-dimensional clusters. In contrast to previous studies [@venturoli2005simulation; @schmidt2008cluster; @de2008molecular; @west2009membrane] where hydrophobic mismatch is considered to be essential for clustering, we showed that proteins form stable clusters even in the absence of mismatch. Lipid-induced depletion interactions have been suggested as one of the main contributing factors to the interactions of cylindrical inclusions in bilayers [@west2009membrane; @lague2001lipid; @sintes1997protein]. This type of attraction occurs at distances smaller than the diameter of one lipid molecule. It has also been suggested that the interaction between membrane proteins largely depends on indirect lipid-mediated interactions [@west2009membrane]. Therefore, in this study, the interactions between proteins at short distances are most likely due to a strong depletion force of entropic origin explained by the Asakura-Oosawa model. This is a consequence of the high flexibility of the lipid chains. Direct protein–protein interactions can also play a role in the aggregation of proteins when protein particles are in the range of the cut-off distance. Khoshnood et al. [@khoshnood2010lipid] report on the same effect of depletion force for aggregation of rigid inclusions compared with completely flexible ones while hydrophobic mismatch exists. A key factor in association of proteins is the distribution of lipid molecules in close proximity to the inclusions [@yoo2013membrane]. These lipid molecules lose their entropy due to interaction with the lateral surface of proteins and form patterns around the inclusion. These patterns are more structured around RPs than SFPs; compare Figs. \[fig6\](b) and \[fig6\](e) with Figs. \[fig6\](c) and \[fig6\](f), respectively. The strong induced lipid structures around RPs lead to a low mobility of lipids, and consequently they have lower diffusion coefficients compared to lipids in a system with SFPs. As a result, the formation of a two-dimensional cluster, is obstructed by the induced lipid structure surrounding RPs, and in such systems one-dimensional aggregates are dominant. We have observed fluctuation-induced attraction between two RPs and SFPs by applying the same constraint described in [@sintes1997protein]. This type of long-range interaction is caused by inclusions that affect the elastic properties of membranes and hence the fluctuation of lipid molecules. Long-range interactions play important roles in the clustering of both RPs and SFPs. However, the shape of the clusters is determined by a repulsive zone when proteins are close to each other and form groups of three or more proteins. The repulsive zone occurs when the distances between proteins are slightly larger than the size of one lipid molecule (beyond the depletion zone) [@lague2001lipid; @sintes1997protein]. When a dimer forms, the fluctuation-induced attraction between proteins overcomes the effect of the repulsive zone and RPs and SFPs are able to create string-like aggregates at the early stages of simulations. The effect of the repulsive zone becomes more prominent when RPs want to create two-dimensional aggregates and can not be overcome by the fluctuation-induced attraction between inclusions. In the case of SFPs, the fluctuation-induced attraction overcomes the effect of the repulsive zone and pushes the third protein into the depletion zone. Both the fluctuation-induced attraction and the repulsive zone are classified as lipid-mediated interactions. Cell adhesion to extracellular matrices is regulated by the size and position of focal adhesions and, most importantly, how they are distributed [@elineni2011regulation]. The latter is controlled by integrin protein association. Integrin has both $\beta$ and $\alpha$ subunits and the number of $\beta$ and $\alpha$ subunits varies depending on the type of integrin, which means a variety of structural rigidity. According to our results variation in integrin stiffness affects their aggregation patterns and consequently may have an impact on the quality of cell adhesion. Cell receptor activation depends on receptors’ clustering and their conformational changes. The latter requires rearrangements of proteins in the cluster [@minguet2007full]. It means that the pattern in which receptors are attached to each other in a cluster plays a role in the activation process. Receptors may have subunits with different structural rigidities, and based on our findings this feature can affect their final aggregation pattern. Remodeling of the biomembrane is achievable by proteins [@simunovic2013linear] and nanoparticles [@vsaric2012fluid], and it is a vital step in endocytosis and vesiculation [@reynwar2007aggregation]. Interestingly, in a system of spherical nano-particles on lipid membranes [@vsaric2012fluid] different aggregation patterns have been induced by varying the membrane rigidity. The majority of living cell membranes are made up of phospholipids and the rigidity of the bilayer is known. Our results suggest that flexibility of the inclusion may work as a controlling parameter for membrane remodeling when we can not alter the membrane rigidity. The new controlling parameter can also help in the design of therapeutic peptides to tackle protein-aggregation diseases. This is obtainable by protein engineering methods without perturbing significantly the overall stability or activity of the protein [@villegas2000protein]. Different mixtures of amino acids have different amounts of alpha-helical and beta-structural units and consequently have different rigidity [@perticaroli2014rigidity]. Therefore, it is possible to design proteins with a specifically defined rigidity and certain aggregation behavior, which lead them to perform a specific biological function. The protein model in our study is a toy model with SFPs and RPs resembling $\alpha$-helices and $\beta$-sheets, respectively, as their structural rigidities differ substantially [@perticaroli2014rigidity]. The simulations by Parton et al. [@parton2011aggregation] with $\alpha$-helical and $\beta$-barrel proteins on vesicles and flat membranes show that while $\alpha$-helical proteins form two-dimensional clusters, $\beta$-barrel proteins constitute linear aggregates. In their experiments, a combination of several factors such as hydrophobic mismatch, membrane curvature, and the shape and class of proteins can be held responsible for the observed discrepancies in the aggregation patterns of proteins. Our results demonstrate that structural rigidity as a sole factor determines the shape and pattern of protein aggregates.

--- abstract: 'We derive covariant baryon wave functions for arbitrary Lorentz boosts. Modeling baryons as quark-diquark systems, we reduce their manifestly covariant Bethe-Salpeter equation to a covariant 3-dimensional form by projecting on the relative quark-diquark energy. Guided by a phenomenological multigluon exchange representation of a covariant confining kernel, we derive for practical applications explicit solutions for harmonic confinement and for the MIT Bag Model. We briefly comment on the interplay of boosts and center-of-mass corrections in relativistic quark models.' author: - | M. Dillig\ Institute for Theoretical Physics, University of Erlangen-Nuernberg\ D-91058 Erlangen, Germany title: 'Baryon Wave Functions in Covariant Relativistic Quark Models [^1]' --- 1.0cm PACS: 03.65Pm,11.30Cp,12.39Ki\ Key Words: Covariance, Quark Models, Diquarks One of the major goals of modern electron and hadron accelerators is the investigation of the internal structure of hadrons, in particular of baryons: detailed information is extracted from scattering experiments at large momentum transfers of typically 1 GeV/c and beyond. The corresponding form factors map out the various internal (generalized) charge distributions and provide stringent information on the underlying quark and gluon degrees of freedom. Presently various experiments are ongoing with electron or photon beams at MAMI,ELSA, MIT, JLAB und DESY (1) and with proton beams at COSY and CELSIUS (2) and other labs. 0.2cm Practical calculations of form factors suffer in general from the pertinent problem of center-of-mass (CM) corrections for the many-body problem and from drastic effects from Lorentz contraction at increasing momentum transfers. While for the CM corrections various recipes have been developed and applied in practical calculations (3-6), less progress has been achieved in the formulation of covariant baryon wave functions suitable for practical calculations (7-24). A possible alternative, the evaluation of formfactors on the light cone, where Lorentz boosts are completely kinematical, has so far entered only selectively in practical applications at low scattering energies, beyond that such an approach suffers from other decreases, such as the loss of strict rotational invariance (25). As in general the construction of boosted, Lorentz contracted wavefunctions is nearly as complicated as the solution of the full problem, in most practical applications ad hoc and purely kinematical prescriptions for the rescaling of the coordinate along the direction of the momentum transfer are applied (examples are given ref. (26-27)). Thus, specific questions, as the dependence of Lorentz corrections on the confining kernel in quark models, are not addressed. In addition, to hopefully minimize the influence of Lorentz contractions formfactors are in general evaluated in the Breit frame, though experimentally they are measured in the lab system. 0.2cm In this note we formulate an economical model for covariant baryon wave functions, which leads to results suitable for practical applications. As it our main goal to end up with analytical formulae, we model the baryon - in the following we use the word proton, though our approach is fairly general - as a quark-diquark system and restict ourselves, without any loss of generality, to spin-isospin scalar diquarks (28). 0.2cm Our starting point is the manifestly covariant 4-dimensional Bethe-Salpeter equation (29) $$\Gamma = K \, G\, \Gamma \quad \mbox{and} \quad \Psi = G \, \Gamma$$ with the vertex function and the Bethe-Salpeter amplitude $ \Gamma $ and $\Psi $, respectively, and the interaction kernel K. In the two-body Greens function for the quark with mass m and the diquark with mass m\* we fix the relative energy dependence from the covariant projection on the diquark (30) $$G (P, q) = \frac{q\!\!\!/ + m}{q^{2} - m ^{2} - i \epsilon} \; i \pi \delta _{+} \left ( ( P - q)^{2} - m^{*2} \right )$$ which results up to $0 \left ( \frac{q^{2}}{2M} \right ) $ in the single particle Dirac equation for the quark for systems with arbitrary overall 4-momenta $ Q = ( E(P) = \sqrt{P^{2} + M^{2}}\, , 0, 0, P ) $ $$\begin{aligned} \left ( \frac{M}{E(P)} \right. \left. \left ( \epsilon + \frac{P}{M} \, q_{z} - \frac{{\bf q}^{2}}{2 M} \; \right ) - (\mbox{\boldmath$\alpha$} {\bf q} + \beta m) \right ) \; \varphi (Q, {\bf q}) & = & \nonumber\\ && \hskip -4cm \frac{1}{E(P)} \int K (Q, {\bf q},{\bf k}) \varphi (Q, {\bf k}) d {\bf k}\end{aligned}$$ 0.2cm Without any details we add a brief comment on the CM corrections in our model: evidently there is a direct coupling between the internal and external momenta [**q**]{} and [**P**]{}, or equivalently, between boosts and the CM motion. In the rest system the leading center-of-mass corrections are absorbed for $ \epsilon = m + \epsilon _{b}$, where $ \epsilon _{b} $ is the binding energy of the quark in $$\left ( \frac{q^{2}}{2 \mu} + \epsilon _{b} - V_{n} (r) \right ) \varphi ({\bf r}) = 0$$ with the reduced mass $1 /\mu = 1 /m + 1 /(m + m^{*}) $ for an arbitrary quark potential $ V_{n} (r)$ (a detailed discussion of CM corrections are presented in a separate paper). 0.2cm The decisive step for a practical model is the formulation of a covariant interaction kernel in eq. (4). As the dynamics of the quark - quark interactions, particularly the microscopic nature of the confinement, lacks an understanding on the fundamental level of QCD, all current models in practical calculations rely on phenomenological formulations of the interaction kernel. Being unable to do better, we proceed here along similar lines: we assume that the interaction kernel can be presented as a superposition of appropriately weigthed gluon exchange contributions; quantitative parameters can be extracted in comparison with studies to baryon spectroscopy, decay rates or form factors (31). Thus we start from the general kernel $$K(P, q,k) = \sum _{n} \; \frac{k_{n} (P)}{((q - k)^{2} - m ^{2} + i \epsilon )^{n+1}}$$ for arbitrary powers of n (which reflect different parametrizations of the confining kernel) (eq. (5) contains the linear confinement in the Cornell potential (32)). Upon projecting out the relative energy dependence this yields the covariant, 3-dimensional kernel $$K_{n} (P, q, k) \propto \;_{ {\lim \atop \mu \to 0}} \; \left(\frac{d}{d \mu^{2}} \right )^n \; \frac{1}{\lambda ^{2} (P) q _{z} ^{2} + {\bf q}^{2}_{\bot} + \mu ^{2} - i \epsilon}$$ with the “quenching parameter” $$\lambda (P) = M / \sqrt{M^{2} + P^{2} } \, = M / E(P)$$ where we introduced the mass scale $\mu $ (to regularize the Fourier transform to coordinate space). Already simple power counting signals, that a kernel with the power n leads to confinement with $\sim r^{2n}$. Upon performing the corresponding Fourier transform to coordinate space and performing the limit $ \mu \to 0 $ we find $$K_{n} (P,q) \to (1+\beta)/2\, V_{n} (\sqrt{(z/\lambda(P))^{2} + \mbox{\boldmath$\rho$} ^{2})}$$ where we introduced for convenience the particular Dirac structure of the kernel to facilitate the evaluation of the resulting Dirac equation. Eliminating the small component in eq. (3) with the kernel from eq. (6) and upon dropping CM corrections and $ \epsilon ^{2}_{b} $ terms for compactness, we end up with the Schroedinger type equation for the large component of the Dirac equation $$\begin{aligned} & & \ (2m\epsilon_{b} -\lambda^{2}( q_{z}- \frac{P}{M} m)^2 ) - {\bf{q} ^{2}_{\bot}} - \nonumber\\ & & \qquad \qquad -\, V_{n} \left ( \sqrt{(z /\lambda (P))^{2} + \mbox{\boldmath$\rho$} ^{2} ) } / R \right ) \; u (z, \bf{\rho} ) = 0\end{aligned}$$ (with the typical length scale R; in the rest system the equation above reduces to the standard spherical Schrödinger type equation for a particle with mass m). The final equation defines with its connection to the small component by a simple differentation the full relativistic covariant quark - diquark wave function for arbitrary Lorentz systems. In the equation above we see the shortcoming from the phenomenological nature of the interaction kernel: we absorb the explicit $\epsilon$ and P dependence of the kernel in the definition of the energy scale $V_{n} $ for the confining force; including an explicit P dependence in $ V_{n}$ would require a detailed knowledge of its microscopic origin. 0.2cm Approximate or numerical solutions for eq. (9) can be obtained for different confining szenarios (a more detailed investigation, such as also of the popular linear (heavy quark) confinement (33), is presented elsewhere). Here we enter only briefly into two szenarios, which allow a rigorous analytic solution for arbitrary systems: i. e. harmonic confinement and bag models in the limit $ n \to \infty $ in eq(6). 0.2cm - Harmonic confinement:\ With the harmonic kernel defined as (34) $$\begin{aligned} K(P,q,k) & = & - 12/\pi _{{\lim \atop (\mu \to 0)}} \left ( ( d/d \mu^{2})^{2} (\mu / 2 + (d/d \mu ^{2}) \mu^{3} /3 \right ) \nonumber\\ & \to & - \left ( (1/\lambda (P))^{2} (d/d q_{z})^{2} + (d/d{\bf q}^{2}_{\bot}) \right ) \delta (q_{z}-k_{z}) \delta ({\bf q}_{\bot} - {\bf k}_{\bot}) \, , \end{aligned}$$ the solutions for arbitrary excitations of the baryon are easily obtained in momentum space. After a redefinition of the longitudinal momentum and upon separating the longitudinal and the perpendicular component, the general solution is given by a product of confluent hypergeometric functions (35). Here we focus only on the nucleon as the quark - diquark ground state and obtain explicitly $$u(q_{z}, {\bf q}_{\bot}) = N \; e^{- \frac{a^{2}}{2} \;\left ( \lambda ^{2} (q_{z} - \frac{m}{M} P \right ) ^{2} + {\bf q}_{\bot} )^{2}}$$ with the oscillator parameter $ a^{2} = \frac{2}{\sqrt{V_{c}}} $, with $ V_{c}\propto 1/R^{4} $ being the confinement strength and with the ground state energy $$\epsilon_{b} (P) \cong (1+ \frac{ P^{2}}{M^{2}}) \cdot \frac{\sqrt{V_{c}}}{m}$$ As expected the standard solution for the spherical harmonic oscillator is recovered in the rest system, i.e. for P=0 and $\lambda $(0)=1.As the characteristic result we find a quenching of the effective P-dependent with size parameter $$a ^{2}(P) ^{2} = (\lambda (P)a )^{2} = \frac{M^{2}}{P^{2} + M^{2}} \; a^{2}$$ which leads to Lorentz quenching in coordinate space along the z - axis and thus to a significant increase of the longitudinal high momentum components with increasing P (Fig.1(a,b)); 0.2cm - Bag Model:\ As mentioned above we generate the Bag from the transition $ n \to \infty $ in the power of gluon-exchange kernel. As we are unable to present an analytical solution for arbitrary n (a closed solution for the z-component exists only in the limit of vanishing binding $ \epsilon _{b} \to 0 $ (35)) we first perform the limit $ n \to \infty $ and then solve the equation $$\left (\epsilon _{b} + \frac{P}{M} \; q _{z} + m - ( \alpha \, {\bf q} + \beta m) \right ) \quad u(z, \mbox{\boldmath$\rho$}) = 0$$ with the standard MIT boundary condition for the large and small components at $ z = \lambda (P) R$ for the bag radius R. For the large component the ground state solution can be represented as $$u (z, \mbox{\boldmath$\rho$}) = N \cos (\frac{k_{z}}{\lambda}\, z) \,\, J_{0} (k_{\bot} \mbox {\boldmath$\rho$})$$ where the $\lambda$ dependence of the z-component again reflects the quenching of the bag (The extension to excited baryon states again is straightforward). The quenching fo the bag along the boost momentum is also reflected in the boundary condition $z=\lambda R$ for $ \mbox{\boldmath$\rho$} =0$, which for the deformed bag can be solved only numerically (36). Characteristic results for 3 different boost momenta are presented for the large and small component of the bag ground state solution in Fig. 2. 0.2cm Comparing our findings with current more phenomenological recipes we find that a general and simple extension of the parametrization of the spherical wave functions and momentum distributions in the rest system to a boosted system, by rescaling the size parameter of the system, but keeping otherwise the spherical character of the solutions, is certainly very unsatisfactory and breaks down completely for boost momenta of typically $P/M \ge 1$. Only for very small boost momenta P simple approximations, such as $$u (z, \mbox{\boldmath$\rho$} , a) \cong u \left ( r, a/ \left ( \sqrt{3} \; \lambda (P) \right) \right) \quad \mbox{and}$$ $$u( z, \mbox{\boldmath$\rho$} , R) = \exp ( - (r / (\sqrt{3} \; \lambda (P) R)) ^{2} \, u (r, R)$$ simulate very qualitatively Lorentz quenching of slowly moving systems. 0.2cm With increasing boost momenta the breaking of the spherical symmetry for the quenched bay leads for the ground (and all excited) state to the admixture of additional angular momenta, which drastically enhance the momentum spectrum of the ground state with increasing q. A characteristic ressult is shown in Fig. 3 for the d-wave admixture for different boost momenta. 0.2cm Summarizing our main findings in this note, we have formulated covariant wave functions and their transformation properties in an analytical quark - diquark model for the baryon and we find characteristic modifications from the baryon rest system to moving Lorentz-systems for different confining kernels. 0.2cm Our findings suggest possible extensions and basic shortcomings of the model. We feel that an extension of the model to mesons as $q\overline{q}$ systems, towards a more realistic quark-diquark description of baryons or to genuine 3-quark systems (together with a systematic inclusion of CM corrections) imposes only technical problems and is certainly feasible. Here we only mention that the quenching factor from eq.(7) is recovered in leading order for all current projections of the BS equation: as an example the Blankenbecler-Sugar reduction (30) for quarks with equal masses yields immediately $$G_{BBS} \, (P, q) \sim \delta (q_{0} - P/M q_{z} )$$ A more serious problem for confining kernels with a finite power in the interquark distance r is the precise formulation of Lorentz quenching for the kernels itself (in Bag models the dependence is absorbed in the boundary condition). Here the unsurmountable problem is our current lack in understanding the confining mechanism: it is not clear, how the full P dependence enters into the kernel (for an example compare ref.(37); however, different szenarios may lead to quantitatively very different results for large P; for an example compare ref. (37)). We feel that a more realistic extension of present phenomenological quark models undoubtedly requires a much deeper analytical understanding of confinement. Here significant progress in various directions has been achieved recently, to mention only the modelling of confinement of QCD in the Coulomb gauge (38) or the extension of conecpt of instantons to merons as solutions of the classical QCD equations (39). [99]{} T. Walcher: Nucl. Phys. A 680 (2000)25; U. Thoma (for CB-ELSA Coll.): “Hirschegg 2001: Structure of Hadrons” (2001) 193; BATES-ANN-REPT-1999 (1999); V. D. Burkert: “ICTP 3. Int. Conf. on Persp. in Hadr. Phys., Trieste 2001”; nucl-ex/0108023; N.C.R. Markins (for Hermes Coll.): Nucl. Phys. A684 (2001) 279 D. Grzonka and K. Kilian: “Hirschegg 2001: Structure of Hadrons” (2001)330; B. Jakobsson: Nucl. Phys. News 9 (1999) 22; K. Shimizu et al.: Phys. Rev. C63 (2001) 025212; R. E. Peierls and J. Yoccoz: Proc. Phys. Soc. London A 70 (1957) 381; R. E. Peierls and D. J. Thouless: Nucl. Phys. 38 (1962) 1154 R. Tegen, R. Brockmann and W. Weise: Z. Phys. A 307 (1982) 329; D. H. Lu, A. W. Thomas and A. G. Williams: Phys. Rev. C 55 (1997) 3108; R. Bakamjian and L. H. Thomas: Phys. Rev. 92 (1953) 1300; Y. S. Kim and R. Zaoui: Phys. Rev. D 4 (1971) 1764; A. L. Licht and A. Pagnamenta: Phys. Rev. D 4 (1971) 2810; F. Coester and W. W. Polyzou: Phys. Rev. 26 (1982) 1348; M. Betz and R. Goldflam: Phys. Rev. D 28 (1983) 2848; B. D. Keister: Nucl. Phys. A 402 (1983) 445; W. Gloeckle and Y. Nogmai: Phys. Rev. D 35 (1987) 3840; D 35 (1987) 584; P. L. Chung and F. Coester: Phys. Rev. D 44 (1991) 229; B. D. Keister and W. W. Plyzou: Adv. Nucl. Phys. 20 (1991) 225; D. Tadic and S. Zganec: Phys. Rev. D 50 (1994)5853; Phys. Rev. D 52 (1995) 6466; S. Giller et al.: Mod. Phys. Lett. A 8 (1993) 3785; Y. B. Dai, C. S. Huang and H. Y. Jin: Phys. Rev. D 51 (1995) 2347; V. Morenas et al.: Phys. Lett. B 386 (1996) 315; F. Coester and D. O. Riska: Few Body Syst. 25 (1998) 29; L. Ya. Glozman et al.: Phys. Rev. D 58 (1998) 094030; hep-ph/9706507 R. Oda et al.: Progr. Theor. Phys. 102 (1999) 297; L. Micu: hep-th/0107210; S. Boffi et al.: hep-ph/01088271; hep-ph/0104223; L. Ya. Glozman et al.; nucl-th/0105028; S. J. Brodsky: Nucl. Phys. Roc. Suppl 90 (2000) 3; hep-ph/0009229; S. J. Brodsky, H.-C. Pauli and S. S. Pinsky: Phys. Rept. 301 (1998) 299; hep-ph/9705477; B. Burkardt: Adv. Nucl. Phys. 23 (1996) 1; A. L. Lewis and A. Pagnamenta: Phys. Rev. D 2 (1970) 1150; Phys. Rev. D 2 (1970)1156; D. A. Lu, A. W. Thomas and A. G. Williams: Phys. Rev. C 55 (1997) 3108; Phys. Rev. C 57 (1998) 2638; M. Anselmino and E. Predazzi: Proc. “Diquarks 3” (1998) (Ed. World Scientific, Singapore); C. F. Berger, B. Lechner and W. Schweiger: Fizika B 8(1999) 371; M. Oettel, R. Alkofer and L. von Smekal: Eur. Phys. J A 8 (2000) 553; Phys. Rept. 353 (1001) 281; M. Oettel: “Hirschegg 2001: Structure of Hadrons” (2001) 140; hep-ph/0102054; S. M. Gerasyuta and D. V. Ivanov: hep-th/0102024; E. E. Salpeter and H. A. Bethe: Phys. Rev. 84 (1951) 1232; C. Itzykson and J. B. Zuber: Quantum Field Theory (1980) (McGraw - Hill, New York); R. Blankenbecler and R. Sugar: Phys. Rev. 142 (1961) 1051; R. J. Yaes: Phys. Rev. D 3 (1971)3086; F. Gross: Phys. Rev. 186 (1969) 1448; Relativ. Quant. Mech. and Quantum Field Theory (1993); (J. Wiley & Sons, Inc. New York); F. Coester, K. Dannborn and D. O. Riska: Nucl. Phys. A 634 (1998) 335; F. Cardarelli et al.: Few. Body. Syst. Suppl. 11 (1999)66; S. Capstick and W. Roberts: nucl-th/0008028; U. Löring et al.: Eur. Phys. J. A 10 (2001) 395; hep-ph/0103289; J. L. Richardson: Phys. Lett. 382 (1979) 272 H. Q. Ding: Int. Journ. Mod. Phys. C 2 (1991) 637; V. V. Kiselev, A. E. Kovalsky and A. I. Onishchenko: Phys. Rev. D 64 (2001) 054009; hep-ph/0005020; N. N. Singh, Y. K. Mathur and A. N. Mitra: Few Body Syst. 1 (1987) 47; M. Schmitt: PhD thesis, Univ. Erlangen (1996); M. Abramowitz and I. A. Stegun: Handbook of Mathematical Functions (1964) (US Nat. Bureau of Standards. Wash. DC.); D. Vasak et al.: J. Phys. G 9 (1983) 511; K. Hahn, R. Goldflam and L. Wilets: Phys. Rev. D 27 (1983) 635; S. J. Wallace and V. B. Mandelzweig: Nucl. Phys. A 503 (1989) 673; A. Cucchieri and D. Zwanziger: hep-lat/0110189; A. Cucchieri, T. Mendes and D. Zwanziger: hep-lat/0110188; A. P.. Szczepaniak and E. S. Swanson: Phys. Rev. D 65 (2002) 025012; C. G. Callan et al.: Phys. Lett. B 66 (1977) 375; O. Jahn et al.: Nucl. Phys. Proc. Suppl. 83 (2000) 524; V. Steel: “Contin. Advances in QCD” (Minneapolis, 2000) hep-lat/0007030 1. Z-dependence of the large component of the harmonic oscillator ground state in coordinate (a) and momentum space (b) for different boost momenta P = 0 (thin line), M (middle line) and 2M (thick line). 2. \(a) Quenching and boundary conditions for the bag along the boost momenta P=0, M, 2M. The functions f(z) and g(z) denote the large and small components of the ground state wave function (for a bag radius R = 1 fm). 0.2cm 3. D-state admixture to the harmonic oscillator ground state. Compared are the s-wave distribution for P=0 with the d-state component for P=M and 2M (middle and thick line, respectively). [^1]: Supported in part by the Kernforschungszentrum KFZ Juelich, Germany, mdillig@theorie3.physik.uni-erlangen.de

--- abstract: 'Chiral molecules may exist in superpositions of left- and right-handed states. We show how the amplitudes of such superpositions may be teleported to the polarization degrees of freedom of a photon and thus measured. Two experimental schemes are proposed, one leading to perfect, the other to state-dependent teleportation. Both methods yield complete information about the amplitudes.' address: | Department of Chemistry, The University of California, Berkeley,\ CA 94720 author: - 'Christopher S. Maierle, Daniel A. Lidar and Robert A. Harris' title: How to Teleport Superpositions of Chiral Amplitudes --- [2]{} Introduction ============ “Quantum teleportation,” proposed in 1993 by Bennett et al. [@Bennett:93], has become a reality [@Bouwmeester:97; @Boschi:98]. However, to date all of the teleportation work deals with [ *intra*]{}-species teleportation (e.g., atom-atom or photon-photon). In this Letter we propose an inter-species teleportation scheme. Specifically, we outline experiments in which the information contained in a superposition of chiral amplitudes: $|\phi _{M}\rangle =a|L\rangle +b|R\rangle$, may be teleported to a photon. Here, $|L\rangle $ and $|R\rangle $ are the left and right handed states of a chiral molecule. In special cases, the teleportation scheme presented here can also be used to teleport the state of more general molecular superpositions such as superpositions of cis and trans isomers. While methods for creating and detecting a molecular state of the form $|\phi _{M}\rangle$ have already been discussed [@Quack:86;Harris:94b;Cina:94;Cina:95; @Harris:94], the corresponding experiments have not been performed. Indeed, no chiral superposition has ever been measured. By teleporting the information contained in the amplitudes $a$ and $b$ to the polarization vector of a photon, the superposition becomes easy to detect via standard photon polarization measurements. Then by performing another teleportation to a spin 1/2 nucleus or a trapped ion (as envisioned in [@Bouwmeester:97]) one can imprint the chirality information onto a much more stable state, suitable for further manipulations: the nucleus or ion acts as a [*quantum memory*]{} device. To teleport the chiral superposition state $|\phi _{M}\rangle$ we take our entangled pair to be two photons in the state: $|\Psi _{12}^{-}\rangle =\frac{1}{\sqrt{2}}(|l_{1}\rangle |r_{2}\rangle -|r_{1}\rangle |l_{2}\rangle )$. Here, $|l\rangle $ and $|r\rangle $ denote left and right circularly polarized photons. Photon 1 will become entangled with the molecule and photon 2 will be the one whose polarization state will receive the amplitudes $a$ and $b$. Initially, the total molecule-photons state is unentangled: $|\psi \rangle =|\phi \rangle |\Psi _{12}^{-}\rangle$. As in [@Bennett:93], this state can be rewritten as: $$|\psi \rangle =\frac{1}{2} \left( |\Psi _{M1}^{-}\rangle |1\rangle +|\Psi _{M1}^{+}\rangle |2\rangle +|\Phi _{M1}^{-}\rangle |3\rangle +|\Phi _{M1}^{+}\rangle |4\rangle \right). \label{eq:fullstatenum}$$ Here the four maximally entangled “Bell states” are: $|\Psi _{M1}^{\pm }\rangle = \frac{1}{\sqrt{2}}(|L\rangle |r_{1}\rangle \pm |R\rangle |l_{1}\rangle )$ , $|\Phi _{M1}^{\pm }\rangle = \frac{1}{\sqrt{2}}(|L\rangle |l_{1}\rangle \pm |R\rangle |r_{1}\rangle )$ , and the four states involving photon 2 are: $$\begin{aligned} |1\rangle & = & -{a \choose b}=-a|\,l_{2}\rangle -b|\,r_{2}\rangle \\ |2\rangle & = & -\sigma _{z}{a \choose b} =-a|\,l_{2}\rangle +b|\,r_{2}\rangle \label{eq:1+2} \\ \,|3\rangle & = & \sigma _{x} {a \choose b}=b|\,l_{2}\rangle +a|\,r_{2}\rangle \\ |4\rangle & = & -i\,\sigma _{y} {a \choose b} =-b|\,l_{2}\rangle +a|\,r_{2}\rangle . \label{eq:3+4}\end{aligned}$$ The $\sigma $’s are the Pauli matrices and $|\,l_{2}\rangle = {1 \choose 0} $, $|\,r_{2}\rangle = {0 \choose 1}$. As emphasized in [@Bennett:93], the above form implies that the “teleportee” photon (2) can be transformed into the state ${a \choose b}$ by one of four simple unitary operations. Which of the four operations needs to be applied depends [*only*]{} on the measurement outcome of the projection onto the Bell states. The scheme thus requires two bits of classical communication to transfer the information regarding the [*continuum*]{} of quantum states ${a \choose b}$, at the price of establishing prior entanglement. However, this is not to say that a unitary transformation of the original teleported photon state is the only way to obtain complete information. Below we give an example of “state-dependent” teleportation which nonetheless does yield complete information. Note further that the circular polarized basis plays no special role in the above discussion. We now show how to extend the previous work by proposing an apparatus which can be used to teleport the superposition of chiral amplitudes. [*Teleportation of Chirality*]{} —. Consider the apparatus shown in Figure 1. Before photon 1 reaches the interferometer, the system is in the direct product state $|\phi \rangle |\Psi _{12}^{-}\rangle$. At some later time, photon 1 will have reached the beam-splitter and thus will have some amplitude to be found in the top arm and some amplitude to be found in the bottom arm where it will interact with the molecule. Now it is well known that left and right circularly polarized photons acquire different phase shifts when scattering through a chiral molecule [@Barron:82]. Ordinarily the phase shifts due to a single molecule are undetectably small. However, it has been shown that by utilizing a high-finesse optical resonator (cavity), phase shifts due to the coupling of a single photon to a single cesium atom may be as large as $16^\circ$ [@Turchette:95]. The interaction which gives rise to this phase shift is an electric dipole-dipole scattering process, whereas optical activity is mediated by an electric dipole-magnetic dipole interaction. Since molecular magnetic dipole moments are about $10^{-2}$ smaller than electric dipole moments, we expect that with current technology, the phase shift due to optical activity in a high-finesse cavity would be on the order of a tenth of a degree, perhaps too small to be useful in our scheme. On the other hand, the last decade has seen tremendous progress in the fabrication of high-finesse optical cavities[@Kimble:94] and we expect that our proposed experiment will be feasible in the future. Furthermore, natural optical activity cannot be enhanced by a standing wave cavity because it is erased upon reflection back through the optically active medium [@Landau+Lifshitz]. Thus, natural optical activity is enhanced only by use of a [*ring*]{}-cavity. Another possibility is to utilize the effect of E-field optical activity [@Harris:94]. This type of optical activity does not vanish upon reflection through the medium so a standing wave cavity may be used. The formalism which we present here is applicable to natural optical activity. For the case of E-field optical activity, the superposition of chiral states should be written in the ‘false chirality’ basis made from the states $|L \rangle \pm i|R \rangle$ [@Barron:86]. These states are converted into one another by inversion and they are to E-field optical activity what the states $|L \rangle$ and $|R \rangle$ are to natural optical activity [@Harris:94]. The photon-molecule scattering implements a [*conditional phase-shift*]{} mechanism. Specifically: $$\begin{aligned} |l\rangle |L\rangle & \rightarrow & e^{i(kz-\varphi )}|l\rangle |L\rangle \text{ ; \ \ \ \ }|r\rangle |R\rangle \rightarrow e^{i(kz-\varphi )}|r\rangle |R\rangle \label{eq:PS0} \\ |l\rangle |R\rangle & \rightarrow & e^{i(kz+\varphi )}|l\rangle |R\rangle \text{ ; \ \ \ \ }|r\rangle |L\rangle \rightarrow e^{i(kz+\varphi )}|r\rangle |L\rangle . \label{eq:PS}\end{aligned}$$ Here $e^{\pm i\varphi }$ is the phase shift due to optical activity and $z$ is the free-space optical path length. For arbitrary molecular superpositions, the Faraday effect[@Barron:82] gives rise to a table similar to the one above. However, with the Faraday effect, the two different molecular states give rise to different rotations and spin-independent indices of refraction (related to the quantities $\phi$ and $kz$ in the above equations). For the special case where the two states give equal and opposite rotations of the polarization vector and also give rise to the same spin-independent phase shift, the Faraday effect gives rise to a set of equations identical to the ones above but where $|L \rangle$ and $|R \rangle$ are replaced by kets representing the two molecular states which are superposed. For this special situation, the Faraday effect can be used in place of optical activity and the teleportation scheme can be used to teleport more general types of molecular superpositions. As an added bonus, the Faraday effect does not vanish upon reflecting the beam back through the medium and can therefore be enhanced in a standing wave cavity. In any case, using the above equations, it is easy to show that after the interaction with the chiral superposition, the amplitude on the bottom arm of the interferometer (Figure 1) is proportional to: $$\begin{aligned} |\psi _{\text{bot}}\rangle &\propto& \frac{e^{i\varphi }}{2}\left[ |\Psi _{M1}^{-}\rangle |1\rangle +|\Psi _{M1}^{+}\rangle |2\rangle \right] \nonumber \\ &+& \frac{ e^{-i\varphi }}{2}\left[ |\Phi _{M1}^{-}\rangle |3\rangle +|\Phi _{M1}^{+}\rangle |4\rangle \right] \label{eq:psi-bot}\end{aligned}$$ The amplitude for going through the upper arm is still described by Eq. (\[eq:fullstatenum\]) and the full state is a superposition of the amplitude on the top and bottom arms. Now, by adjusting the path-length and thus the phase of the amplitude in the top arm of the interferometer, we can arrange so that only one pair of the states $|\Psi _{M1}^{\pm }\rangle $ and $|\Phi _{M1}^{\pm }\rangle $ has non-zero amplitude to reach the detector 2. Suppose we adjust the top arm so that only the $|\Psi _{M1}^{\pm }\rangle $ reach detector 2. Then after the photon has left the interferometer the state can be written (up to an overall phase): $|\psi \rangle =\frac{\sin (\varphi )}{\sqrt{2}}\left[ |\Psi _{M1}^{-}\rangle _{ \text{D}_{2}}|1\rangle +|\Psi _{M1}^{+}\rangle _{\text{D}_{2}}|2\rangle \right] +|\psi' \rangle$. Here the subscript D$_{2}$ indicates that the photon in that state is heading for detector 2. The state $| \psi' \rangle $ is the amplitude which in Figure 1 is now traveling vertically away from this detector (denoted “other”). We are not concerned with the precise form of $| \psi' \rangle $; for our purposes it suffices to know that this state involves photon amplitude which will never intersect D$_{2}$. Using the basis of parity eigenstates $ |\pm \rangle $ and linear polarizations $|x\rangle $, $|y\rangle $: $$\begin{aligned} |\pm \rangle &=&\frac{1}{\sqrt{2}}(|L\rangle \pm |R\rangle ) \\ |x\rangle &=&\frac{1}{\sqrt{2}}(|l\rangle +|r\rangle ) \:;\:\: |y\rangle =\frac{i}{\sqrt{2}}(|l\rangle -|r\rangle )\text{ ,}\end{aligned}$$ we can rewrite the post-interferometer state as: $$\begin{aligned} |\psi \rangle & = & \frac{\sin (\varphi )}{\sqrt{2}}(\frac{1}{\sqrt{2}}|+\rangle |x\rangle _{\text{D}_{2}}-\frac{i}{\sqrt{2}}|-\rangle |y\rangle _{\text{D} _{2}})|1\rangle \nonumber \\ & + &(\frac{1}{\sqrt{2}}|-\rangle |x\rangle _{\text{D}_{2}}-\frac{i}{\sqrt{2}} |+\rangle |y\rangle _{\text{D}_{2}})|2\rangle +|\psi' \rangle \text{ .}\end{aligned}$$ The laser in Figure 1 is used to produce a $\pi$ pulse, tuned to a transition between the ground state of the molecule and an excited state of definite parity which fluoresces. Suppose that the excited state is of odd parity; in the electric dipole approximation the parity must change upon electronic excitation, so only the state $|+\rangle $ will be excited. Therefore, after excitation and fluorescence, we arrive at the state: $$\begin{aligned} |\psi \rangle & = & \frac{\sin (\varphi )}{\sqrt{2}} \left[ \left( \frac{1}{\sqrt{2}}|+\rangle |\nu _{1}\rangle |x\rangle _{\text{D}_{2}}-\frac{i}{\sqrt{2}}|-\rangle |\nu _{0}\rangle |y\rangle _{\text{D}_{2}} \right) |1\rangle \right. \nonumber \\ & + & \left. \left( \frac{1}{\sqrt{2}}|-\rangle |\nu _{0}\rangle |x\rangle _{\text{D}_{2}}- \frac{i}{\sqrt{2}}|+\rangle |\nu _{1}\rangle |y\rangle _{\text{D} _{2}} \right) |2\rangle \right] +|\psi' \rangle . \nonumber \\\end{aligned}$$ The $|+\rangle $ molecular state has become coupled to a spontaneously emitted photon ($|\nu_{1}\rangle $), whereas the $|-\rangle $ state has not (the vacuum state $|\nu _{0}\rangle $). Teleportation can now be performed by projecting onto an [*unentangled*]{} state [@Cirac:94]: one places a polaroid oriented in the $x$ direction in front of the detector 2 and looks for coincidences with detector 1 (which detects the spontaneous emission $|\nu _{1}\rangle$). A coincidence measurement then constitutes a projection onto the state $|\nu _{1}\rangle |x\rangle _{\text{D}_{2}}$. This implies that the teleportee photon is in the state $|1\rangle =-a|l_{2}\rangle -b|r_{2}\rangle $. Teleportation has been achieved. By altering the length of the top arm of the interferometer and/or exciting the molecule to an even parity state, we can teleport to any of the four photon states in Eq. (\[eq:fullstatenum\]). [*State-Dependent Teleportation*]{} —. The latter scheme accomplishes perfect teleportation: an [*unknown*]{} amplitude of the chiral superposition appears in the polarization vector of photon 2. We will next consider a simplified experiment which avoids the use of interferometry, and accomplishes [*state-dependent*]{}, imperfect teleportation. Suppose we remove the upper arm of the interferometer. Then the entire apparatus is represented by $|\psi _{\text{bot}}\rangle $ as in Eq. (\[eq:psi-bot\]). We again rewrite the amplitude in the $|+\rangle $, $ |-\rangle $, $|x\rangle $ and $|y\rangle $ representation, yielding: $|\psi _{\text{bot}}^{\prime }\rangle =\left[ |-\rangle |y\rangle |1^{\prime }\rangle +|+\rangle |y\rangle |2^{\prime }\rangle +|-\rangle |x\rangle |3^{\prime }\rangle +|+\rangle |x\rangle |4^{\prime }\rangle \right]$ where the four unnormalized teleportee photon states are: $$\begin{aligned} |1^{\prime }\rangle &\equiv &\frac{i}{2\sqrt{2}}\left( \,e^{-i\varphi }|2\rangle -\,e^{i\varphi }|4\rangle \right) \label{eq:1'} \\ |2^{\prime }\rangle &\equiv &\frac{i}{2\sqrt{2}}\left( \,e^{-i\varphi }|1\rangle -\,e^{i\varphi }|3\rangle \right) \label{eq:2'} \\ |3^{\prime }\rangle &\equiv &\frac{1}{2\sqrt{2}}\left( \,e^{-i\varphi }|1\rangle +\,e^{i\varphi }|3\rangle \right) \label{eq:3'} \\ |4^{\prime }\rangle &\equiv &\frac{1}{2\sqrt{2}}\left( \,e^{-i\varphi }|2\rangle +\,e^{i\varphi }|4\rangle \right) \label{eq:4'}\end{aligned}$$ Denoting $a=\alpha e^{i\theta _{a}}$ and $b=\beta e^{i\theta _{b}}$, the norms are: $$\begin{aligned} \text{Pr}(1^{\prime }) &=&|\langle \psi _{\text{bot}}^{\prime }|-,y\rangle |^{2}=\text{Pr}(3^{\prime })=|\langle \psi _{\text{bot}}^{\prime }|-,x\rangle |^{2} \nonumber \\ &=& \frac{1}{4}[1-2\alpha \beta \cos \left( \theta _{a}-\theta _{b})\cos (2\varphi \right) ] \label{eq:norm1} \\ \text{Pr}(2^{\prime }) &=&|\langle \psi _{\text{bot}}^{\prime }|+,y\rangle |^{2}=\text{Pr}(4^{\prime })=|\langle \psi _{\text{bot}}^{\prime }|+,x\rangle |^{2} \nonumber \\ &=& \frac{1}{4}[1+2\alpha \beta \cos \left( \theta _{a}-\theta _{b})\cos (2\varphi \right) ]\text{.} \label{eq:norm2} \end{aligned}$$ We notice that the transformation matrices that send ${a \choose b}$ into $|1'\rangle$ through $|4'\rangle$ are not unitary. For example, $$|1^{\prime }\rangle = \frac{1}{i 2\sqrt{2}}\left( \begin{array}{cc} e^{-i\varphi } & -e^{i\varphi } \\ e^{i\varphi } & -e^{-i\varphi } \end{array} \right) \left( \begin{array}{c} a \\ b \end{array} \right).$$ The resulting states are, however, [*pure*]{}. Appropriately renormalized, they may be represented by polarization vectors on the unit Bloch sphere. Hence there is a unitary rotation which carries each teleported vector into the original state, ${a \choose b}$. However, as seen from Eqs. (\[eq:norm1\]) and (\[eq:norm2\]), the transformation depends upon the values of $a,b$, and the phase-shift angle $\varphi$. In this sense the present scheme constitutes a state-dependent, imperfect teleportation, since it cannot be used to teleport an unknown quantum state [@Mor:96]. Nevertheless, it is possible to obtain [*full*]{} information about $a$ and $b$ by standard optical methods. One can measure the relative phase and relative magnitude of the polarization components of photon 2. The relative phases and separately, the relative magnitudes, are equal in pairs. Thus by transmission of a single bit of classical information, it is possible to tell cases $1^{\prime },3^{\prime }$ from $2^{\prime },4^{\prime }$ and to obtain complete information on the chiral superposition. Of course, each measurement which contributes to this process destroys the superposition of polarizations of photon 2. This loss cannot be prevented in the perfect teleportation scheme either if the actual values of $a$ and $b$ are needed. From this perspective there is no real advantage to the perfect scheme. Indeed, the perfect scheme is better only if photon 2 is put to use in a later quantum information processing stage, such as an input to a quantum computer. [*Discussion*]{} —. This work has, for the first time, considered in detail the possibility of inter-species teleportation. This led us to propose a concrete scheme by which the quantum chiral state of a single molecule could be measured, by means of transferring the chiral information to an easily measurable photon polarization state. We have outlined two experiments, one leading to perfect, or unitary, teleportation of the amplitudes of a chiral superposition, the other to state-dependent teleportation. However we have shown that the latter is able to transmit, in general, full amplitude information as well. The key to the schemes proposed here is the use of the parity-conserving [*symmetry*]{} governing the interaction between light and a chiral molecule, Eqs. (\[eq:PS0\]) and (\[eq:PS\]). This symmetry leads to the possibility of implementing a conditional phase-shift, without which teleportation cannot take place. We conjecture that it is possible to exploit other symmetries in order to affect inter-species teleportation in other cases. Indeed, we have discussed how, using the Faraday effect, the state of a more general molecular superposition can be teleported. In the same vein, any other spin $1/2$ particle can be used to replace the photons in our scheme, but with a different interaction, such as spin-orbit coupling [@Maierle:97]. A virtue of the method of chiral teleportation is that it provides a genuine new way of measuring chiral superpositions of chiral amplitudes. Indeed, even if the original molecular state is $|L \rangle$, successful teleportation is a manifestation of at least one pair of superpositions, $|+\rangle$ and $|-\rangle$. Model calculations show that chiral superpositions in media are extremely short-lived: they decohere on a time-scale of pico- to femtoseconds [@Cina:94a]. Hence one might wonder whether the chiral superposition will not decohere over the time scale needed for the photon to interact with it. Collisions with the walls and asymmetries of the cavity are the chief decoherence agents as they lead to fluctuating chiral environments [@Harris:81]. In the case of a high-finesse cavity we estimate that this should lead to decoherence times of the order of 1 second. Decoherence thus does not present a significant obstacle in accomplishing the proposed experiment. Another issue brought up by this work is the possibility of probing quantum properties in “large” objects, and thus the transition (be it continuous or sharp) as a function of object size to classical behavior. The emergence of the latter is one of the most fascinating unsolved problems of present-day physics. At which point is the object “too large” to enable teleportation of its chiral superposition? The number of degrees of freedom of the object will set the decoherence time-scale, since it determines the coupling to the bath degrees of freedom. It thus controls the extent of “environmental-monitoring,” leading to classical behavior, i.e., the absence of quantum interference in large objects [@Zurek]. In principle, as long as the object is chiral and there exist excited non-degenerate states of definite parity, our schemes apply. [*Ceteris paribus*]{}, failure to teleport may thus be taken as an indication of classicality of the chiral object. Finally, the extension of quantum teleportation to superpositions of molecular states has yielded an entirely new way of measuring given superpositions of chiral amplitudes. It is therefore tempting to consider [*reverse*]{} inter-species teleportation. Consider, e.g., diasterioisomers: these are two molecules connected by a chemical bond or a Van der Waals complex. The molecules are are mirror images of one another. That is, one is $|L\rangle$ and the other is $|R \rangle$. For example, a left-handed oligomer of diphenyl alanine (DA) connected by a disulphide bridge (DB) to a right-handed oligomer of DA. The DB bridge is easily cleaved. Let us imagine that the dimer is cooled down to a $J=0$ state of total angular momentum. When cleaved the dimer state breaks up into monomer states of equal total angular momentum and equal and opposite $z$ component of angular momentum, $M$. The fragments also move in opposite directions. Clearly then there is a sum of entangled states in $M$. Using the same scheme as described above for a simultaneous measurement on a chiral molecule and a photon, we can now teleport a superposition of [*photon*]{} polarization states to create a superposition of handed states in one of the molecules. Thus we can create arbitrary superpositions of chiral amplitudes through inter-species teleportation. Acknowledgements {#acknowledgements .unnumbered} ================ We would like to acknowledge interesting conversations with T. Lynn, and Profs. I. Tinoco, K.B. Whaley and W.H. Miller. This work was supported by a grant to R.A.H. from the N.S.F. [10]{} , [*Phys. Rev. Lett.*]{} [**70**]{}, 1895 (1993). , [*Nature*]{} [**390**]{}, 575 (1997). , [*Phys. Rev. Lett.*]{} [**80**]{}, 1121 (1997). , [*Chem. Phys. Lett.*]{} [**132**]{}, 147 (1986); [R.A. Harris, Y. Shi and J.A. Cina]{}, [*J. Chem. Phys.*]{} [**101**]{}, 3459 (1994); [J.A. Cina and R.A. Harris]{}, [*J. Chem. Phys.*]{} [**100**]{}, 2531 (1994); [J.A. Cina and R.A. Harris]{}, [*Science*]{} [**267**]{}, 832 (1995). , [*Chem. Phys. Lett.*]{} [**223**]{}, 250 (1994). , [*[Molecular Light Scattering and Optical Activity]{}*]{} ([Cambridge University Press]{}, [Cambridge, England]{}, 1982). , [*Phys. Rev. Lett.*]{} [**75**]{}, 4710 (1995). H. J. Kimble in [*Cavity Quantum Electrodynamics,*]{} ed. by P. R. Berman (Academic, San Diego, CA 1994) p 203. , [*Electrodynamics of Continuous Media*]{} ([Addison-Wesley]{}, [Reading, Mass.]{}, 1960). , [*Chem. Phys. Lett.*]{} [**123**]{}, 423 (1986). Application of additional degrees of freedom (here $|\nu_{0,1}\rangle$) in order to avoid the measurement in a Bell basis has been considered before by J.I. Cirac and A.S. Parkins, [*Phys. Rev. A*]{} [**50**]{}, R4441 (1994). , [“TelePOVM” – More Faces of Teleportation]{}, LANL Report No. quant-ph/9608005; [M.A. Nielsen and C.M. Caves]{}, [*Phys. Rev. A*]{} [**55**]{}, 2547 (1997). , [*Chem. Phys. Lett.*]{} [**267**]{}, 199 (1997). , in [*[Ultrafast Phenomena]{}*]{}, edited by [P.F. Barbara, W.H. Knox, G.A. Mourou and A.H. Zewail]{} ([Springer-Verlag]{}, Berlin, 1994), Vol. [IX]{}, p. 486. , [*J. Chem. Phys.*]{} [**74**]{}, 2145 (1981). , [*Physics Today*]{} [**44**]{}, 36 (1991); [W.H. Zurek]{}, [*Prog. Theor. Phys.*]{} [**89**]{}, 281 (1993); [W.H. Zurek, S. Habib and J.P. Paz]{}, [*Phys. Rev. Lett.*]{} [**70**]{}, 1187 (1993). =5.5cm \[fig:expt\] Fig. 1: [A source produces an entangled pair of photons, 1 and 2. Photon 1 enters an interferometer and interacts with the molecule, labeled M. A $\pi$ pulse, produced by the box labeled laser, excites the molecule to a state of definite parity and the resulting fluorescence is detected by detector 1. Detector 2 is set to detect photons of a definite polarization. A coincidence measurement then teleports the state of the molecular superposition to the polarization state of photon 2.]{}

--- author: - | C. Aberle$^a$, A. Elagin$^b$, H.J. Frisch$^b$, M. Wetstein$^b$, and L. Winslow$^a$[^1]\ University of California, Los Angeles, Los Angeles, CA 90095, USA\ University of Chicago, Chicago, IL 60637, USA\ E-mail: bibliography: - 'DirectionBibliography\_v1.bib' title: 'Measuring directionality in double-beta decay and neutrino interactions with kiloton-scale scintillation detectors' --- =1 Introduction ============ Liquid scintillator-based detectors are responsible for several of the critical measurements that have determined our present understanding of neutrino masses and mixings. These measurements include KamLAND’s measurement of reactor anti-neutrino oscillation at a distance of $\sim$200 km[@kam2013], Borexino’s measurement of $^{7}$Be solar neutrino oscillation[@borexino], and most recently the short baseline reactor anti-neutrino experiments that measured oscillations due to $\theta_{13}$ at a distance of 1 km: Daya Bay[@dbtwo], Double Chooz[@dctwo; @dchydrogen], and RENO[@reno]. Scintillator-based neutrino detectors will continue to be important for the next set of neutrino measurements, from the determination of the neutrino mass hierarchy[@juno; @reno50] to elastic scattering measurements[@isodarscatt] and sterile neutrino searches[@isodar; @nist], and for non-proliferation applications[@nucifer; @songs]. The scalability of these detectors to large volumes also makes them highly competitive for neutrino-less double-beta ($0\nu\beta\beta$) decay searches in which the final state consists of a pair of electrons with energies in the $\sim$1-2 MeV range. The observation of this rare decay would prove that the neutrino is a Majorana particle, which would have profound consequences to our understanding of the generation of mass and may provide a possible explanation of the matter-antimatter asymmetry in the universe[@leptogenesis]. Currently one of the best limits for the $0\nu\beta\beta$ half-life comes from the scintillating detector KamLAND-Zen[@KZ0nu]. The advantage of liquid scintillators for measurements in the $\sim$1 MeV range is their scalability from 1 ton to 1 kiloton while providing energy resolutions of $\sigma(E)=\sim$5 % $/\sqrt{E(MeV)}$[@kam2013; @borexino]. This is roughly a factor of two better than for water Cherenkov detectors, the other developed technology for neutrino detectors that can be economically scaled to these large masses. However, this energy resolution is much poorer than other technologies being used for $0\nu\beta\beta$ searches: Ge detectors[@gerda2013], Te bolometers[@Alessandria:2011rc], tracking detectors[@SuperNEMO], liquid Xe time projection chambers (TPCs)[@EXO2012] and high pressure gaseous Xe TPCs[@NEXTsipm]. Scintillation light is isotropic. At these low energies, it does not contain sufficient information to reconstruct the track of the outgoing particles, although at higher energies it may[@john]. Cherenkov light is also produced for electrons above threshold. Most is absorbed and re-emitted as part of the scintillation processes; however some fraction retains its directional information. If this directional Cherenkov light can be isolated from the copious isotropic scintillation light, it may be possible to reconstruct the direction of the primary particle. The addition of directionality is a powerful tool for background rejection, especially for $0\nu\beta\beta$ searches. In high pressure TPCs, reduction factors on the order of $\sim10^4$ have been achieved[@Gotthard]. It is also possible to look for new physics in the angular correlation of the emerging electrons[@newphysics0nuBB], as has been proposed for the tracking-based detectors[@SuperNEMO; @NEXTsipm]. The addition of a directional signal would make large-scale liquid scintillator detectors more competitive for the next generation of $0\nu\beta\beta$ searches. This is the first in a series of papers exploring directionality in large-scale liquid scintillator detectors. In this paper, we develop a technique for separating the Cherenkov and scintillation light using the photon arrival times and evaluate several detector advances in timing, photodetector spectral response, and scintillator emission spectra that would allow the realization of direction reconstruction in kilo-ton scale scintillating neutrino detectors. This is different from the direction reconstruction described for high-energy neutrino interactions[@john] or that for neutrons from inverse beta decay[@chooz; @dcDirection]. We then use these results as input into a traditional direction reconstruction developed for water Cherenkov detectors. Since the reconstruction of the direction of $\sim$1 MeV particles has not been achieved before, we start these studies with the simple case of a single particle at the center of the detector. We also start with an easier test case of a 5 MeV electron like that from a $^{8}$B solar neutrino interaction. With the higher photon statistics at this energy, we verify the technique and the different detector parameters that affect it. We then study the technique for two lower energies, 1.4 and 2.1 MeV, that are more relevant to $0\nu\beta\beta$. Liquid scintillator detectors ============================= Liquid scintillators are ‘cocktails’ of aromatic hydrocarbons. When charged particles move through a scintillator, the molecules are excited, predominantly via the non-localized electrons in the $\pi$-bonds of the phenyl groups[@birks_book]. Vibrational and rotational modes of the molecules are turned into heat within picoseconds through collisions with other molecules. Within $\sim$10 picoseconds, the $\pi$-electrons de-excite to the first excited state from higher levels through radiationless transitions. The first excited state can de-excite through photon emission. There are two characteristic times for this de-excitation, depending if the singlet state or the triplet state was excited. The singlet state will de-excite within nanoseconds while the triplet state de-excites on the order of 10’s or 100’s of nanoseconds. These two processes are fluorescence and phosphorescence respectively. The exact time constants for these processes are determined by the composition of the scintillator. The absorption and emission spectra overlap at some level in all molecules. Consequently, if there is only one type of molecule in the scintillator cocktail the light output is reduced due to inefficiencies in the energy transfer through multiple absorption and re-emission processes. Aromatic solutes or fluorophores are added to the primary solvent to shift the wavelengths of the photons to higher values where the scintillator is more transparent. This wavelength-shifting is also used to match the quantum efficiency as a function of wavelength for the photodetectors being used. One typical scintillator mixture uses pseudocumene(1,2,4-trimethylbenzene) as the solvent with 1-5 g/l of PPO (2,5-diphenyloxazole) as the fluorophore. This mixture has a peak emission at about 400 nm where bialkali photomultiplier tubes (PMTs) are most sensitive and the pseudocumene is relatively transparent. A good liquid scintillator will produce $\sim$10,000 photons isotropically per MeV of deposited energy. Although less abundant, Cherenkov light will be produced as well if a particle is moving faster than the speed of light in the medium. This light is emitted in a cone centered on the direction of the particle trajectory, and with a continuous spectrum weighted toward shorter wavelengths but extending well into the red. The spectrum is described by[@Cherenkov34]: $$\label{eqCherenkov} \frac{d^2N}{d\lambda dx} = \frac{2 \pi \alpha Z^2}{\lambda^2} \left [ 1 - \frac{1}{\beta^2 n(\lambda)^2} \right ]$$ where $N$ is the number of Cherenkov photons, $\lambda$ is the wavelength of the photon, x is the distance travelled by the particle, $Z$ is the charge of the particle, $\alpha$ is the fine structure constant, $n(\lambda)$ is the wavelength-dependent index of refraction and $\beta$ is the velocity of the particle. The Cherenkov light produced at wavelengths shorter than the absorption cutoff of the scintillator will be absorbed and re-emitted as isotropic light, but wavelengths longer than this cutoff will propagate across the detector, retaining their directional information. The absorption cutoff for the example scintillator above is 370 nm. The index of refraction at $\lambda$=370 nm is n=1.466, which translates into an effective Cherenkov threshold for electrons at 0.188 MeV. For a 5 MeV electron, this yields 685 Cherenkov photons per event when integrating from the cutoff wavelength at 370 nm to 550 nm, the wavelength above which a small number of photons is detected due to the quantum efficiency of a standard bialkali photocathode. Lowering the energy to 1 MeV yields 82 Cherenkov photons between 370-550 nm. All photons including these undisturbed Cherenkov photons will have timing determined by the group velocity[@group_velocity_article; @pdg_review_2012; @tamm1939] in the liquid, $$\label{eqGroup} v_{g}(\lambda) = \frac{c_{vacuum}}{n(\lambda) - dn(\lambda)/d\textnormal{log}(\lambda)}.$$ Photons at the scintillation cutoff of 370 nm have a velocity of 0.191 m/ns, while photons with wavelengths of 600 nm are appreciably faster, with a velocity of 0.203 m/ns. Since on average the undisturbed Cherenkov photons have longer wavelengths, they will arrive before the scintillation light, which is slowed by both the scintillation processes and the shorter wavelengths involved. Thus, with sufficient timing resolution and sensitivity to longer wavelengths it should be possible to separate the directional Cherenkov light and the isotropic scintillation light, and then to reconstruct the direction of the initial particle. In neutrino-electron scattering events, a single electron emerges with a distribution of energies with a maximum energy related to the incoming neutrino’s energy. In comparison, $0\nu\beta\beta$ events are more complicated, with two electrons emerging with a combined energy equal to the Q-value of the particular isotope. The individual electrons follow distributions of energies and angular correlations which depend on the underlying $0\nu\beta\beta$ decay mechanism[@SuperNEMO; @newphysics0nuBB; @bandv]. Figure 1 shows a simulated $^{116}$Cd $0\nu\beta\beta$ event in a model with light Majorana neutrino exchange, for which a probable case is the emission of two electrons with comparable energies at a large angle relative to each other. High Q-value candidates are preferred for $0\nu\beta\beta$ for two reasons. First, as the measured half-life is inversely proportional to the phase-space factor, the measured rate is expected to be higher from isotopes with higher Q-values. Second, the main background for these experiments come from the daughters of the $^{238}$U and $^{232}$Th decay chains. The Compton shoulder from gamma-rays is particularly problematic. The 2.6 MeV gamma-ray from $^{208}$Tl decay is the highest energy gamma from these decay chains and so isotopes above 2.6 MeV are preferred. Experiments using isotopes with Q-values below this energy must compensate with improved background rejection techniques. Most of the high Q-value candidates[@tabledbb] have been considered as a dopant for a liquid scintillator: $^{150}$Nd (Q=3.367 MeV)[@minfang; @nd1], $^{96}$Zr (Q=3.350 MeV)[@zr1], $^{100}$Mo (Q=3.034 MeV)[@mo1], $^{82}$Se (Q=2.995 MeV)[@qdot], $^{116}$Cd (Q=2.81 MeV)[@qdot; @cd1], $^{130}$Te (Q=2.533 MeV)[@qdot; @biller], $^{136}$Xe (Q=2.479 MeV)[@KZ0nu] and $^{124}$Sn (Q=2.29 MeV)[@sn1]. Xenon gas readily dissolves into liquid scintillator. For the other isotopes, a suitable organometallic compound needs to be found that produces a stable scintillator with a long attenuation length in the wavelength region of interest. Recently nanocrystals formed by candidate isotopes have been explored as an alternative to doping by single atoms[@qdot; @qdot2]. ![The detector geometry and coordinate system. The radial rays (green lines) are photons emitted by two back-to-back electrons with 1.4 MeV each (equally divided energy of $^{116}$Cd $0\nu\beta\beta$ decay). The electrons originate at the center of the sphere with initial directions along the x and -x-axis. Only Cherenkov photons are drawn to illustrate the directionality of the event. \[detector\_view\]](geometry_plot_labels.pdf) Geant4 simulation {#sim_section} ================= In order to study the effects relevant to directional reconstruction in liquid scintillators, a Geant4[@geant4one; @geant4two] simulation has been constructed. The simulation uses Geant4 version 4.9.6 with the default liquid scintillator optical model, in which optical photons are assigned the group velocity in the wavelength region of normal dispersion. The detector geometry is a sphere of 6.5 m radius filled with scintillator. Figure \[detector\_view\] shows the geometry and the Cherenkov light from an example $^{116}$Cd $0\nu\beta\beta$ event. The default scintillator properties have been chosen to match a KamLAND-like scintillator[@kamland2003]: 80% n-dodecane, 20% pseudocumene and 1.52 g/l PPO. The scintillator properties implemented in the simulation include the atomic composition and density ($\rho$ = 0.78 g/ml), the wavelength-dependent attenuation length[@tajimaMaster] and refractive index[@OlegThesis], the scintillation emission spectrum[@tajimaMaster], emission rise time ($\tau_r$ = 1.0 ns) and emission decay time constants ($\tau_{d1}$ = 6.9 ns and $\tau_{d2}$ = 8.8 ns with relative weights of 0.87 and 0.13)[@tajimaThesis], scintillator light yield (9030 photons/MeV), and the Birks constant ($kB$ $\approx$ 0.1 mm/MeV)[@ChrisThesis]. This is a standard scintillator. The attenuation length at 400 nm, the position of the peak standard bialkali photocathode efficiency, is 25 m. The attenuation length drops precipitously between 370 nm and 360 nm from 6.5 m to 0.65 m. We use this drop to define the cutoff wavelength at 370 nm. Variations from the baseline KamLAND case are discussed below. Re-emission of absorbed photons in the scintillator bulk volume and optical scattering, specifically Rayleigh scattering, have not yet been included by default. A test simulation shows that the effect of optical scattering is negligible. As shown in figure \[scattplot\], scattering causes a small tail at longer times. The reason is that the cutoff is very steep below 360 nm and almost no photons reach the sphere, so optical scattering makes no difference for short wavelengths. Above about 395 nm, the attenuation length is greater than 20 m so both scattering and absorption are not very likely and scattering is negligible. The intermediate region is rather small. A similar argument holds for re-emission. Scattering length measurements and discussions can be found in Ref. [@Wurm:2010ad]. ![Cherenkov photoelectron (PE) arrival times after application of the transit-time spread (TTS) for the simulation of 1000 electrons (5 MeV). The default simulation is shown with optical scattering turned on: all photons (black solid), un-scattered photons (red dotted), and scattered (blue dashed). Scattering causes a very small tail at longer times as expected. \[scattplot\]](PEtime_Cerenkov_scattered_vs_unscattered.pdf) The inner sphere surface is used as the photodetector. It is treated as fully absorbing (no reflections), with a photodetector coverage of 100%. As in the case of optical scattering, reflections at the sphere are a small effect that would create a small tail at longer times. Two important photodetector properties have been varied: 1) the transit-time spread (TTS, default $\sigma$ = 0.1 ns) and 2) the wavelength-dependent quantum efficiency (QE) for photoelectron production. The default is the QE of a bialkali photocathode (Hamamatsu R7081 PMT)[@Hamamatsu_R7081]. The QE values as a function of wavelength come from the Double Chooz[@dctwo] Monte Carlo simulation. We note that the KamLAND 17-inch PMTs use the same photocathode type with similar quantum efficiency. We are neglecting any threshold effects in the photodetector readout electronics. Four effects primarily contribute to the timing of the scintillator detector system: the travel time of the particle, the time constants of the scintillation process, chromatic dispersion, and the timing of the photodetector. First, the simulated travel time of a 5 MeV electron is 0.108$\pm$0.015 ns. This corresponds to an average path length of 3.1 cm and a final distance from the origin of 2.6 cm. The time until the electron drops below Cherenkov threshold is 0.106$\pm$0.015 ns. We note that due to scattering the final direction of the electron before it stops does not correspond to the initial direction; however the scattering angle is small while the majority of Cherenkov light is produced. The Cherenkov light thus encodes the direction of the primary electron. The scattering physics is handled by Geant4’s “Multiple Scattering" process which is valid down to 1 keV, where atomic shell structure becomes important[@geant4scatt]. In the energy range important for $0\nu\beta\beta$, a 1.4 MeV electron travels a total path length of 0.8 cm, has a distance from the origin of 0.6 cm in 0.030$\pm$0.004 ns and takes 0.028$\pm$0.004 ns to drop below Cherenkov threshold. The scattering follows the same pattern. The scintillator-specific rise and decay times are the second effect that determines the timing in a scintillator detector. The first step in the scintillation process is the transfer of energy from the solvent to the solute. The time constant of this energy transfer accounts for a rise time in scintillation light emission. Past neutrino experiments were not highly sensitive to the effect of the scintillation rise time, which is the reason why there is a lack of accurate numbers. We assume a rise time of 1.0 ns; more detailed studies are needed in the future. The two time constants used to describe the falling edge of the scintillator emission time distribution (quoted above) are values specific to the KamLAND scintillator. Chromatic dispersion is the third effect that determines the timing in a scintillator detector. Due to the wavelength-dependence of the refractive index, the speed of light in the scintillator (see Equation (\[eqGroup\])) increases with increasing photon wavelengths for normal dispersion, with red light traveling faster than blue light. In order to study the time differences due to this chromatic dispersion, we used a simplified simulation of 5 MeV electrons at the center of the sphere where we used instantaneous scintillation emission with the quantum-efficiency applied, but not including a transit-time spread. The higher energy 5 MeV electron provides larger photon statistics with which to evaluate the method, debug the simulation, and is of interest to neutrino-electron scattering experiments. The true hit time distributions of photoelectrons were analyzed for scintillation light and Cherenkov light separately. Photoelectrons coming from Cherenkov light are on average created about 0.5 ns earlier than PEs from scintillation light. The RMS values from PE time distributions for Cherenkov and scintillation light are both about 0.5 ns. Note that these numbers include the effect of the finite electron travel time. The fourth effect determining the timing in a scintillator detector is the timing of the photodetectors. The measurement of the arrival times of single photoelectrons is affected by the transit-time spread (TTS) of the photodetectors, a number which can be different by orders of magnitude depending on the detector type. The default TTS of 0.1 ns ($\sigma$) can be achieved with large area picosecond photodetectors (LAPPDs)[@Adams:2013nva; @RSI_paper; @PSEC4_paper; @anode_paper] and possibly hybrid photodetectors (HPDs)[@hpdThesis]; even significantly lower TTS numbers are realistic with the LAPPD[@RSI_paper; @PSEC4_paper; @anode_paper]. In sections \[detector\_timing\_sec\] to \[scintillator\_emission\_sec\], we study the photoelectron timing for different detector configurations at 5 MeV. We focus on the idea of increasing the discrimination between Cherenkov and scintillation light by using improved detector timing. The primary quantities provided by the Geant4 simulation are the photoelectron hit positions and the detection times after the TTS resolution has been applied. In section \[reconstruction\_sec\] these quantities are then used for event reconstruction. With the successful reconstruction at 5 MeV, we then lower the energy of the simulated electrons and show that it is possible to reconstruct electrons in the range interesting for $0\nu\beta\beta$. Detector timing {#detector_timing_sec} =============== We first discuss results for the default simulation settings described in the previous section. Figure \[time\_plots\_comparison\] (a) shows the TTS-smeared photoelectron (PE) detection times for 1000 simulated electrons with 5 MeV energy in the center of the detector, with initial momentum directions coinciding with the x-axis. The photoelectrons induced by Cherenkov light arrive earlier, as expected due to the instantaneous emission and the higher average photon speed compared to scintillation light. There is, however, significant overlap of the two arrival time distributions. In order to compare simulations with different parameters, a fixed time cut of $t\leq$ 34.0 ns is applied using the truth information to isolate the Cherenkov light in this early time window, as shown in figure \[time\_plots\_comparison\] (a). For the default simulation case, the average number of PEs per event coming from Cherenkov light in the early time window (108) is 98% of the total average number of PEs from Cherenkov light (110). For scintillation light, the average number of PEs (171) is only 3.1% of the average total scintillation-induced PEs (5445). This demonstrates the effectiveness of a time cut to separate Cherenkov light from scintillation light. The ratio of Cherenkov-induced to scintillation-induced photoelectrons in the early time window ($R_{C/S}$) is a useful figure-of-merit when comparing different simulation settings, since a higher ratio means more directional information per PE. For the default simulation settings $R_{C/S}=0.63$. ![The angular distribution of photoelectron hits relative to the original electron direction, $\cos(\Psi) = x_{hit}/|\vec{r}_{hit}|$. Three energies are shown: 5 MeV (Black - Top), 2.1 MeV (Red - Middle), 1.4 MeV (Blue - Bottom). Each sample consists of 1000 events produced at the detector center. Default simulation settings are used and both Cherenkov and scintillation light are included. The $t\leq$ 34.0 ns cut is applied.[]{data-label="Cherenkov_cone"}](combination_cos_psi_34_h.pdf) Figure \[Cherenkov\_cone\] displays the angular distribution of PE hits after the time cut. Although this time cut is a simplification of actual time reconstruction effects, we can use it to indicate the spatial distribution of hits in the early time window. The Cherenkov ring structure can be clearly seen in the peak near 46, demonstrating that the directional signal conveyed by the Cherenkov photons is not erased by scattering of the initial electrons even at 1.4 MeV. When the 17-inch KamLAND PMTs[@tajimaMaster; @kume_1983] (TTS = 1.28 ns) are used in the simulation, the broadening of the time distributions leads to a strongly decreased ratio of Cherenkov over scintillation light ($R_{C/S}=0.25$) for $t<34.0$ ns (see figure \[time\_plots\_comparison\] (b)). This shows that a photodetector with a low TTS is critical for directionality reconstruction and motivates the use of novel photodetector types. Detector wavelength response {#detector_wavelength_response_sec} ============================ Since Cherenkov photons that pass through meters of scintillator have on average longer wavelengths than scintillation photons, a photodetector that is more sensitive at long wavelengths increases not only the absolute number of PEs but also the ratio between Cherenkov- and scintillation-induced PEs. We have run the simulation with the QE of an extended red-sensitive GaAsP photocathode (Hamamatsu R3809U-63)[@Hamamatsu_R3899U], but with the default TTS of 0.1 ns. Figure \[time\_plots\_comparison\](c) shows the results for the modified simulation with high QE in the red spectral region. The higher absolute number of photoelectrons coming from Cherenkov light (factor of $\approx$ 2) and the increased Cherenkov/scintillation ratio ($R_{C/S}=1.01$) in the early time window would significantly improve the directionality reconstruction. Scintillator emission spectrum {#scintillator_emission_sec} ============================== An alternative route towards increasing the separation in time between Cherenkov and scintillation photon hits is the tuning of the scintillator emission spectrum. Recently, the use of quantum dots (QDs) in liquid scintillators has been studied as a possibility to improve future large scale neutrino experiments[@qdot; @qdot2]. One major motivation for quantum-dot-doped scintillator is control of the emission spectrum by tuning the size or composition of the quantum dots. Quantum dots can also provide a mechanism for introducing an isotope for studying double-beta decay. The emission spectrum of commercial alloyed core/shell CdS$_x$Se$_{1-x}$/ZnS quantum dots was measured in Ref.[@qdot2]. This spectrum shows a symmetric peak centered around 461 nm with FWHM = 29 nm. In order to isolate the effect of the different emission spectrum, the other simulation settings, including the KamLAND absorption spectrum, were kept unchanged; we find $R_{C/S}=0.17$ for the default 34.0 ns timing cut. Compared to the default case shown in figure \[time\_plots\_comparison\](a), the separation is worse (as expected) because the scintillation light wavelengths are longer than in the KamLAND emission spectrum. However, advances in the production of commercial quantum dot samples could yield quantum dots that have similar, single peak emission shapes at shorter wavelengths. This case has been simulated using the same spectral shape of the measured core-shell quantum dot emission but shifted to shorter wavelengths such that the emission peak is centered at 384 nm. This peak emission value has been measured for other types of QDs, however with a much more pronounced tail[@qdot2]. The resulting PE time distribution shows improved separation of Cherenkov and scintillation light compared to the default simulation. After the 34.0 ns cut on the TTS-smeared PE time we obtain a Cherenkov/scintillation ratio of $R_{C/S}=0.86$ (107 PE from Cherenkov light and 124 PE from scintillation). The number of Cherenkov-induced PEs after the time cut is unchanged while the number of PEs coming from scintillation light is decreased due to the higher average photon travel times. Reconstruction {#reconstruction_sec} ============== The timing studies show that in the early time window, $t\leq$ 34.0 ns, the ratio $R_{C/S}$ is high, improving the photoelectron hit selection. In this section, we apply reconstruction tools for a water Cherenkov detector, WCSimAnalysis, to the problem of reconstructing the position and direction of 5 MeV electrons from this early light. WCSimAnalysis is a water Cherenkov reconstruction package developed by the Long Baseline Neutrino Experiment (LBNE) Collaboration[@Blake]. It provides a framework for generic event cleaning, track reconstruction, and particle identification, and comes equipped with variety of pre-built algorithms. It is continuing to be expanded using new track-fitting techniques for water Cherenkov detectors[@Sanchez2012525] based on advanced photosensors with sub-cm imaging capabilities and timing resolutions below 100 picoseconds. To start, we are neglecting the effects of position dependence. In future work the arrival times can be corrected by the time of flight from the reconstructed vertex and the position and direction fitted simultaneously. For isotropic light, the vertex reconstruction uncertainty leads to an additional smearing of the arrival time distribution of 0.15 ns for every 3 cm of position reconstruction uncertainty. For directional light, the vertex reconstruction uncertainty leads to an effective shift of the arrival time distribution. This would change the ideal time cut by 0.15 ns for every 3 cm of position reconstruction uncertainty. We have studied other time cuts from 33 ns to 34.5 ns and the reconstruction remains reasonable. The results presented in this paper rely on a simple vertex reconstruction algorithm, commonly known as a “point fit”[@SuperKalgo]. It assumes that all of the scintillation and Cherenkov light is emitted from a single point in space-time $(x_0,y_0,z_0,t_0)$. In actuality, the light is emitted along a multi-scattered electron track. However, at the energies discussed in this paper, the extent of this track is small (a few cm) compared to the scale of the detector (R=6.5 meters) that sets the typical photon transit distances. The first step of the reconstruction process relies on exact numerical calculations of vertex candidates from quadruplets of hits. Given a single point source, we need four constraints to solve for the four unknowns of the vertex $(x,y,z,t_0)$[@Smy]. This approach would provide an exact solution in the case of four prompt, un-scattered photons originating from a common point. However, many of these randomly chosen quadruplets will produce anomalous solutions due to ‘real world’ effects such as delayed emission and deviations from the point-like geometry. Nonetheless, we found that any chosen subset of 400 quadruplets was a sufficiently large ensemble to assure that some solutions will be close to the true vertex. Once a set of vertex candidates has been found, we test the goodness of each vertex and select the one that best fits the full ensemble of photon hits. The goodness of fit is determined based on the distribution of an observable known as the “point time residual”[@SuperKalgo]. The point time residual is calculated by first choosing a hypothesis for the vertex position and $t_0$ for the event. The goodness of fit for these values is then calculated by taking the difference between the measured and predicted times of each photon hit, using a single effective speed of light in the scintillator. The width of the time residual distribution over all hits is minimized when the hypothesized vertex is near the true vertex. Based on this figure of merit, we select the vertex with the narrowest time residual distribution from among the 400 candidates. The direction of the electron track is then determined by taking the centroid of all vectors pointing from the fitted vertex to the hits on the detector. Since the Cherenkov light is highly directional, and since the timing cut enhances the purity of the Cherenkov light in the sample, this calculation provides a good measure of the track direction. For the purpose of testing the reconstruction algorithm we use 1000 simulated electrons with an energy of 5 MeV; lower energies are studied in the next section. The electrons are simulated at the center of the detector, $\vec{r}$ = (0,0,0), along the x-axis, $\vec{p}/|\vec{p}|$ = (1,0,0). Figure \[fig:reco\] (Top) shows the vertex reconstruction. The vertex is reasonably well-reconstructed around the center of the detector, $\vec{r}$ = (0,0,0), except along the x-axis. The RMS values of the distributions for all three reconstructed coordinates are smaller than 3.5 cm. The shift along the x-axis is due to two effects for which the reconstruction has to use average values rather than the unknown true value for each hit: the wavelength and hence the speed of the light in the medium, and the point of emission for each of the photons, reconstructed as coming from a common point. The reconstruction of the direction also is shown in figure \[fig:reco\] (Top). It shows that for the majority of the events the initial electron direction is reconstructed well. This is a promising result given the simplicity of the algorithms. Energy dependence {#edep_section} ================= In the previous sections we presented results on single 5 MeV electrons such as might be observed in neutrino-electron scattering. In this section we study two lower energies, 1.4 MeV and 2.1 MeV, appropriate for searches for $0\nu\beta\beta$. These energies correspond to Q/2 for the double-beta decay of $^{116}$Cd and $^{48}$Ca, respectively[@cd1; @biller]. The isotope $^{116}$Cd was chosen because of its potential use in quantum-dot-doped scintillators[@qdot; @qdot2] and $^{48}$Ca was chosen to cover the Q-value range of $0\nu\beta\beta$ candidate isotopes. ![The energy dependence of the mean number of PEs after the 34.0 ns time cut is shown for Cherenkov-induced PEs (black open circles, dotted line) and scintillation-induced PEs (black open squares, dashed line). The ratio between the mean number of Cherenkov-induced and scintillation-induced PEs is shown as blue filled triangles, values are given on the right y-axis. The statistical errors are too small to be seen. \[Edep\_NPE\]](PEandR_vs_energy_final.pdf) The two additional simulation sets with 1.4 MeV and 2.1 MeV electrons were generated using the default simulation configuration described in section \[sim\_section\]. The PE time distribution for the default settings is shown in figure \[time\_plots\_comparison\] (a) for 5 MeV electrons. The shapes of the scintillation and Cherenkov spectra are similar for the lower energies (not shown here). In figure \[Edep\_NPE\], the energy-dependent mean number of PEs per event after the 34.0 ns time cut is shown for Cherenkov-induced and scintillation-induced PEs, as well as their ratio $R_{C/S}$. The mean number of PEs from Cherenkov (scintillation) light is 21.8 (52.1), 38.4 (76.8) and 108 (171) for electron energies of 1.4 MeV, 2.1 MeV and 5 MeV, respectively. This gives the ratios $R_{C/S}$ = 0.42, 0.50 and 0.63: The decrease in Cherenkov-induced PEs is stronger than the decrease in scintillation-induced PEs as the energy is lowered. ![Mean cosine of the angle between the true initial electron direction and the reconstructed direction, as a function of the electron energy. For each energy 1000 events have been simulated. Statistical errors are shown. \[Edep\_angle\]](hCos_vs_E_final.pdf) The reconstruction algorithms outlined in section \[reconstruction\_sec\] have also been applied to the simulations at lower energies. Figure \[fig:reco\] (Middle) shows the results for 2.1 MeV and figure \[fig:reco\] (Bottom) shows the results for 1.4 MeV. Most events are still reconstructed well, despite the lower number of PEs and the decreased $R_{C/S}$. For 1.4 MeV electrons, the RMS values of the distributions for all three reconstructed coordinates are smaller than 4.5 cm. The mean cosine of the angle between the true direction and the reconstructed direction for different energies of the initial electron is shown in figure \[Edep\_angle\]. The direction reconstruction performance is still promising for energies as low as 1.4 MeV. Conclusions =========== We have developed a technique to separate scintillation and Cherenkov light to reconstruct direction of electrons with energies 5 MeV, 2.1 MeV and 1.4 MeV in a liquid scintillator detector. These energies have been chosen to represent typical neutrino-electron elastic scattering energies and $0\nu\beta\beta$ energies. The Cherenkov threshold for an electron in a typical liquid scintillator is $\sim$0.2 MeV. While scintillation light is isotropic, Cherenkov light with wavelengths above the absorption cutoff $\sim$370 nm carries the information about the direction of the electrons. All light shorter than this gets absorbed and re-emitted isotropically as a part of the scintillation process. On average scintillation light is delayed with respect to the direct Cherenkov light due to chromatic dispersion and the finite time of the scintillation processes; the early light thus contains directional information. Using a Geant4 simulation of a spherical detector with radius of 6.5 m and photodetectors with TTS of 0.1 ns, we have shown that for electrons originated at the center a time cut on the early light is effective at isolating the directional light, improving the ratio of Cherenkov light to scintillation light from 0.02 to 0.63. This ratio is degraded by a factor of 2.5 if current photodetectors with $\sim$1 ns resolution were used. This can be improved by factors of 1.6 or 1.4 if more red-sensitive photodetectors or scintillators with narrower emission spectra are used. Reconstruction algorithms developed for water Cherenkov detectors have been applied to this early light. The algorithms are able to converge on reasonable reconstructed vertices and directions for all simulated energies: 5 MeV, 2.1 MeV, 1.4 MeV. As expected, the reduction in photon statistics with lower energies leads to a broadening of the reconstructed vertex by $\sim$1 cm. Similarly, the mean $\cos(\theta)$ between the reconstructed and initial directions drops from 0.88 to 0.74. This technique is promising, and we plan to continue work on the topic. The ability to reconstruct direction in kiloton-scale scintillation detectors would expand capabilities of neutrino experiments, especially those also searching for neutrino-less double-beta decay. More generally, this technique could be applied wherever scintillation-based detectors are used. The authors thank Andrew Blake at University of Cambridge for his work authoring the WCSimAnalysis code. The authors thank the neutrino reconstruction group at Iowa State, particularly Mayly Sanchez, Ioana Anghel, and Tian Xin, for their continued work in developing the WCSimAnalysis algorithms and for their insights and expertise regarding issues related to Cherenkov reconstruction with fast-timing. The authors also thank Michael Smy for his development of the quadruplet-based vertex-finding method. L. Winslow would like to thank Janet Conrad for many useful discussions on the topic, and Katsushi Arisaka for discussions on the possible reach of traditional PMTs and the characteristics of HPDs. C. Aberle and L. Winslow are supported by funds from University of California Los Angeles. The work at the University of Chicago is partially supported by DOE contract DE-SC0008172 and NSF grant PHY-1066014. Matthew Wetstein gratefully acknowledges support by the Grainger Foundation. [^1]: corresponding author

--- author: - | R. Casana, C. A. M. de Melo and B. M. Pimentel\ [Instituto de Física Teórica, Universidade Estadual Paulista]{}\ [Rua Pamplona 145, CEP 01405-900, São Paulo, SP, Brazil]{}\ [**Abstract:** We discuss the coupling of the electromagnetic field with a curved]{}\ [and torsioned Lyra manifold using the Duffin-Kemmer-Petiau theory. We will show]{}\ [how to obtain the equations of motion and energy-momentum and spin density tensors]{}\ [by means of the Schwinger Variational Principle.]{} title: 'Electromagnetic Field in Lyra Manifold: A First Order Approach' --- Introduction ============ First order Lagrangians are one of the most profitable tools in Field Theory. By means of first order approach, Hamiltonian dynamics becomes more transparent, constrained systems can be dealt with a wide range of methods [@FirstOrderConstrained], and CPT and spin-statistics theorems can be proved by variational statements [@SchwSpinStat]. Otherwise, the coupling between electromagnetism and the torsion content of spacetime has been  an intringuing puzzle for many years. Minimal coupling of the Einstein-Cartan gravity with eletromagnetism breaks local gauge covariance by the presence of the torsion interaction [@sabbata; @RiemCart; @MassiveRCartan]. Here, we want to add another piece to the puzzle, showing that the torsion coupling problem is related to scale invariance which we will model together with the gravitational field by means of the Lyra geometry. Electromagnetic field will be described by the first order approach of Duffin-Kemmer-Petiau (DKP). The Lyra Geometry\[LyraGeom\] ============================= The Lyra manifold [@Lyra] is defined giving a tensor metric $g_{\mu \nu } $ and a positive definite scalar function $\phi $ which we call the scale function. In Lyra geometry, one can change scale and coordinate system in an independent way, to compose what is called a *reference system* transformation: let $M\subseteq \mathbb{R}^{N}$ and $U$ an open ball in $% \mathbb{R}^{n}$, ($N\geq n$) and let $\chi :U\curvearrowright M$. The pair $% \left( \chi ,U\right) $ defines a *coordinate system*. Now, we define a reference system by $\left( \chi ,U,\phi \right) $ where $\phi $ transforms like $$\bar{\phi}\left( \bar{x}\right) =\bar{\phi}\left( x\left( \bar{x}\right) ;\phi \left( x\left( \bar{x}\right) \right) \right) \quad ,\quad \frac{% \partial \bar{\phi}}{\partial \phi }\not=0$$under a reference system transformation. In the Lyra’s manifold, vectors transform as $$\bar{A}^{\nu }=\frac{\bar{\phi}}{\phi }\frac{\partial \bar{x}^{\nu }}{% \partial x^{\mu }}A^{\mu }$$In this geometry, the affine connection is $$\tilde{\Gamma}_{\;\,\mu \nu }^{\rho }=\frac{1}{\phi }\mathring{\Gamma}^{\rho }{}_{\mu \nu }+\frac{1}{\phi }\left[ \delta _{\,\mu }^{\rho }\partial _{\nu }\ln \left( \frac{\phi }{\bar{\phi}}\right) -g_{\mu \nu }g^{\rho \sigma }\partial _{\sigma }\ln \left( \frac{\phi }{\bar{\phi}}\right) \right]$$whose transformation law is given by $$\begin{aligned} \tilde{\Gamma}_{\;\,\mu \nu }^{\rho } &=&\frac{\bar{\phi}}{\phi }\bar{\Gamma}% _{\;\,\lambda \varepsilon }^{\sigma }\frac{\partial x^{\rho }}{\partial \bar{% x}^{\sigma }}\frac{\partial \bar{x}^{\lambda }}{\partial x^{\mu }}\frac{% \partial \bar{x}^{\varepsilon }}{\partial x^{\nu }}+\frac{1}{\phi }\frac{% \partial x^{\rho }}{\partial \bar{x}^{\sigma }}\frac{\partial ^{2}\bar{x}% ^{\sigma }}{\partial x^{\mu }\partial x^{\nu }}+ \\ &&+\frac{1}{\phi }\delta _{\,\nu }^{\rho }\frac{\partial }{\partial x^{\mu }}% \ln \left( \frac{\bar{\phi}}{\phi }\right) \,.\end{aligned}$$ One can define the covariant derivative for a vector field as $$\nabla _{\mu }A^{\nu }={\frac{1}{\phi }}\partial _{\mu }A^{\nu }+\tilde{% \Gamma}_{\;\,\mu \alpha }^{\nu }A^{\alpha }\,,\quad \nabla _{\mu }A_{\nu }={% \frac{1}{\phi }}\,\partial _{\mu }A_{\nu }-\tilde{\Gamma}_{\;\,\mu \nu }^{\alpha }A_{\alpha }\,.$$We use the notation $\Gamma _{\;\,[\alpha \mu ]}^{\nu }=\frac{1}{2}\left( \Gamma _{\;\alpha \mu }^{\nu }-\Gamma _{\;\mu \alpha }^{\nu }\right) $ for the antisymmetric part of the connection and $\mathring{\Gamma}_{\,\mu \nu }^{\rho }\equiv \frac{1}{2}g^{\rho \sigma }\left( \partial _{\mu }g_{\nu \sigma }+\partial _{\nu }g_{\sigma \mu }-\partial _{\sigma }g_{\mu \nu }\right) $ for the analogous of the Levi-Civita connection. The richness of the Lyra’s geometry is demonstrated by the *curvature* [@Sen] $$\begin{aligned} \tilde{R}_{\,\beta \alpha \sigma }^{\rho } &\equiv &\frac{1}{\phi ^{2}}\left[ \partial _{\beta }\left( \phi \tilde{\Gamma}_{\,\alpha \sigma }^{\rho }\right) -\partial _{\alpha }\left( \phi \tilde{\Gamma}_{\,\beta \sigma }^{\rho }\right) \right] + \\ &&+\frac{1}{\phi ^{2}}\left[ \phi \tilde{\Gamma}_{\,\beta \lambda }^{\rho }\phi \tilde{\Gamma}_{\,\alpha \sigma }^{\lambda }-\phi \tilde{\Gamma}% _{\,\alpha \lambda }^{\rho }\phi \tilde{\Gamma}_{\,\beta \sigma }^{\lambda }% \right]\end{aligned}$$and the *torsion* $$\tilde{\tau}_{\mu \nu }^{\,\quad \rho }=-\frac{2}{\phi }\delta _{\,[\mu }^{\rho }\partial _{\nu ]}\ln \bar{\phi} \label{torsion1}$$which has intrinsic link with the scale functions and whose trace is given by $$\tilde{\tau}_{\mu \rho }^{\,\quad \rho }\equiv \tilde{\tau}_{\mu }=\ \frac{3% }{\phi }\partial _{\mu }\ln \bar{\phi}\,. \label{torsion-trace}$$ In the next section we introduce the behavior of massless DKP field in the Lyra geometry. The massless DKP field in Lyra manifold ======================================= DKP theory describes in a unified way the spin $0$ and $1$ fields [DKPoriginais, KrajcikNieto, Lunardi1]{}. The massless DKP theory can not be obtained as a zero mass limit of the massive DKP case, so we consider the Harish-Chandra Lagrangian density for the massless DKP theory in the Minkowski space-time $\mathcal{M}^{4}$, given by [@HarishC] $$\mathcal{L_{M}}=i\bar{\psi}\gamma \beta ^{a}\partial _{a}\psi -i\partial _{a}% \bar{\psi}\beta ^{a}\gamma \psi -\bar{\psi}\gamma \psi \;, \label{minkowski}$$where the $\beta ^{a}$ matrices satisfy the usual DKP algebra $$\beta ^{a}\beta ^{b}\beta ^{c}+\beta ^{c}\beta ^{b}\beta ^{a}=\beta ^{a}\eta ^{bc}+\beta ^{c}\eta ^{ba}$$and $\gamma $ is a *singular* matrix satisfying[^1] $$\beta ^{a}\gamma +\gamma \beta ^{a}=\beta ^{a}\qquad ,\qquad \gamma ^{2}=\gamma \,.$$ From the above lagrangian it follows the massless DKP wave equation $$i\beta ^{a}\partial _{a}\psi -\gamma \psi =0\;.$$ As it was known, the Minkowskian Lagrangian density (\[minkowski\]) in its massless spin 1 sector reproduces the electromagnetic or Maxwell theory with its respective $U(1)$ local gauge symmetry. To construct the covariant derivative of massless DKP field in Lyra geometry, we follow the standard procedure of analyzing the behavior of the field under local Lorentz transformations, $$\psi \left( x\right) \rightarrow \psi \prime \left( x\right) =U\left( x\right) \psi \left( x\right) \label{LocalLorentz}$$where $U$ is a spin representation of Lorentz group characterizing the DKPfield. Now we define a *spin connection* $S_{\mu }$ in a such way that the object $$\nabla _{\mu }\psi \equiv \frac{1}{\phi }\partial _{\mu }\psi +S_{\mu }{}\psi \label{cov-fer}$$transforms like a DKP field in (\[LocalLorentz\]), thus, we set$$\nabla _{\mu }\psi \rightarrow \left( \nabla _{\mu }\psi \right) ^{\prime }=U\left( x\right) \nabla _{\mu }\psi$$and therefore $S$ transforms like $$S_{\mu }^{\prime }=U\left( x\right) S_{\mu }U^{-1}\left( x\right) -\frac{1}{% \phi }\left( \partial _{\mu }U\right) U^{-1}\left( x\right) \label{TransfSpin}$$ From the covariant derivative of the DKP field (\[cov-fer\]) and remembering that $\bar{\psi}\psi $ must be a scalar under the transformation (\[LocalLorentz\]), it follows that $\nabla _{\mu }\bar{\psi}=\frac{1}{% \phi }\partial _{\mu }\bar{\psi}-\bar{\psi}S_{\mu }$. Using the covariant derivative of the DKP current $$\begin{gathered} \nabla _{\mu }\left( \bar{\psi}\beta ^{\nu }\psi \right) =\frac{1}{\phi }% \partial _{\mu }\left( \bar{\psi}\beta ^{\nu }\psi \right) +\Gamma ^{\nu }{}_{\mu \lambda }\left( \bar{\psi}\beta ^{\lambda }\psi \right) = \\ =\left( \nabla _{\mu }\bar{\psi}\right) \beta ^{\nu }\psi +\bar{\psi}\left( \nabla _{\mu }\beta ^{\nu }\right) \psi +\bar{\psi}\beta ^{\nu }\left( \nabla _{\mu }\psi \right)\end{gathered}$$one gets the following expression for the covariant derivative of $\beta ^{\nu }$$$\nabla _{\mu }\beta ^{\nu }=\frac{1}{\phi }\partial _{\mu }\beta ^{\nu }+\Gamma ^{\nu }{}_{\mu \lambda }\beta ^{\lambda }+S_{\mu }\beta ^{\nu }-\beta ^{\nu }S_{\mu }$$ A particular solution to this equation is given by $$S_{\mu }=\frac{1}{2}\omega _{\mu ab}S^{ab}~,\;S^{ab}=\left[ \beta ^{a},\beta ^{b}\right] .$$ With a covariant derivative of the DKP field well-defined we can consider the Lagrangian density (\[minkowski\]) of the massless DKP field minimally coupled [@misner; @hehl; @sabbata] to the Lyra manifold, introducing the tetrad field,$$g^{\mu \nu }(x)=\eta ^{ab}\,e^{\mu }{}_{a}(x)e^{\nu }{}_{b}(x)\,,\;g_{\mu \nu }(x)=\eta _{ab}e_{\mu }{}^{a}(x)e_{\nu }{}^{b}(x)~.$$$$S=\int_{\Omega }\!\!d^{4}x\;\phi ^{4}e\;\left( \!i\,\bar{\psi}\gamma e_{\;\,a}^{\mu }\beta ^{a}\nabla _{\mu }\psi -i\nabla _{\mu }\bar{\psi}\beta ^{a}e_{\;\,a}^{\mu }\gamma \psi -\bar{\psi}\gamma \psi \right) \;. \label{lag-cv}$$where $\nabla _{\mu }$ is the Lyra covariant derivative of DKP field defined above. Equations of Motion and the Description of Matter Content ========================================================= In following we use a classical version of the Schwinger Action Principle such as it was treated in the context of Classical Mechanics by Sudarshan and Mukunda [@Sudarshan]. The Schwinger Action Principle is the most general version of the usual variational principles. It was proposed originally at the scope of the Quantum Field Theory [@SchwSpinStat], but its application goes beyond this area. Here, we will apply the Action Principle to derive equations of motion of the Dirac field in an external Lyra background and expression for the energy-momentum and spin density tensors. Thus, making the variation of the action integral (\[lag-cv\]) we get$$\delta S=\int_{\Omega }dx~e\phi ^{4}\left[ 4\mathcal{L}-\frac{i}{\phi }\bar{% \psi}\gamma \beta ^{\mu }\partial _{\mu }\psi +\frac{i}{\phi }\partial _{\mu }\bar{\psi}\beta ^{\mu }\gamma \psi \right] \left( \frac{\delta \phi }{\phi }% \right) +$$$$\begin{gathered} +\int_{\Omega }dx~\phi ^{4}e~\left( \frac{\delta e}{e}\right) \mathcal{L~}+ \label{action-variation} \\ +\int_{\Omega }dx~e\phi ^{4}\left[ \frac{{}}{{}}i\bar{\psi}\gamma \left( \delta \beta ^{\mu }\right) \nabla _{\mu }\psi -i\nabla _{\mu }\bar{\psi}% \left( \delta \beta ^{\mu }\right) \gamma \psi \right] + \notag \\ +\int_{\Omega }dx~e\phi ^{4}\left[ \frac{{}}{{}}i\bar{\psi}\gamma \beta ^{\mu }\left( \delta S_{\mu }\right) \psi +i\bar{\psi}\left( \delta S_{\mu }\right) \beta ^{\mu }\gamma \psi \right] + \notag \\ +\int_{\Omega }dx~e\phi ^{4}~\delta \bar{\psi}\left( i\gamma \beta ^{\mu }\nabla _{\mu }\psi -\gamma \psi +iS_{\mu }\beta ^{\mu }\gamma \psi \right) + \notag \\ +\int_{\Omega }dx~e\phi ^{4}\left[ \frac{i}{\phi }\bar{\psi}\gamma \beta ^{\mu }\left( \delta \partial _{\mu }\psi \right) -\frac{i}{\phi }\left( \delta \partial _{\mu }\bar{\psi}\right) \beta ^{\mu }\gamma \psi \right] \notag \\ -\int_{\Omega }dx~e\phi ^{4}\left( i\nabla _{\mu }\bar{\psi}\beta ^{\mu }\gamma +\bar{\psi}\gamma -i\bar{\psi}\gamma \beta ^{\mu }S_{\mu }\right) \delta \psi \notag\end{gathered}$$Choosing different specializations of the variations, one can easily obtain the equations of motion and the energy-momentum and spin density tensor. Equations of Motion ------------------- We choose to make functional variations only in the massless DKP field thus we set $\delta \phi =\delta e^{\mu }{}_{b}=\delta \omega _{\mu ab}=0$ and considering $\left[ \delta ,\partial _{\mu }\right] =0$, from ( [action-variation]{}) we get$$\begin{gathered} \delta S=\int_{\partial \Omega }d\sigma _{\mu }~e\phi ^{3}~i\left[ \frac{{}}{% {}}\bar{\psi}\gamma \beta ^{\mu }\left( \delta \psi \right) -\left( \delta \bar{\psi}\right) \beta ^{\mu }\gamma \psi \right] + \\ +i\int_{\Omega }dx~e\phi ^{4}\left( \delta \bar{\psi}\right) \left[ i\beta ^{\mu }\nabla _{\mu }\psi +i\tilde{\tau}_{\mu }\beta ^{\mu }\gamma \psi -\gamma \psi \frac{{}}{{}}\right] + \\ -\int_{\Omega }dx~e\phi ^{4}\left[ i\nabla _{\mu }\bar{\psi}\beta ^{\mu }+i% \tilde{\tau}_{\mu }\bar{\psi}\gamma \beta ^{\mu }+\bar{\psi}\gamma \frac{{}}{% {}}\right] \delta \psi\end{gathered}$$ Following the action principle we get the generator of the variations of the massless DKP field$$G_{\delta \psi }=\int_{\partial \Omega }d\sigma _{\mu }~e\phi ^{3}~i\left[ \frac{{}}{{}}\bar{\psi}\gamma \beta ^{\mu }\left( \delta \psi \right) -\left( \delta \bar{\psi}\right) \beta ^{\mu }\gamma \psi \right]$$and its equations of motion in the Lyra’s manifold are$$\left\{ \begin{array}{c} i\beta ^{\mu }\left( \nabla _{\mu }+\tilde{\tau}_{\mu }\gamma \right) \psi -\gamma \psi =0 \\ i\nabla _{\mu }\bar{\psi}\beta ^{\mu }+i\tilde{\tau}_{\mu }\bar{\psi}\gamma \beta ^{\mu }+\bar{\psi}\gamma =0% \end{array}% \right.$$ The spin 1 projectors $R^{\mu }$($=e^{\mu }{}_{a}R^{a}$ ) and $R^{\mu \nu }$($=e^{\mu }{}_{a}e^{\nu }{}_{b}R^{ab}$) [@Umezawa; @Lunardi1] are such that $R^{\mu }\psi $ and $R^{\mu \nu }\psi $ transform respectively as a vector and a second rank tensor under general coordinate transformation. Thus, using the projectors we have $$R^{\mu }\,\ \ \rightarrow \,\ \ i\nabla _{\nu }\left( R^{\mu \nu }\psi \right) +i\tilde{\tau}_{\nu }\left( R^{\mu \nu }\gamma \psi \right) -R^{\mu }\gamma \psi =0$$multiplying by $\left( 1-\gamma \right) $ we get $$i\left( \nabla _{\nu }+\tilde{\tau}_{\nu }\right) \left( R^{\mu \nu }\gamma \psi \right) =0$$and $$R^{\mu \nu }\,\ \ \rightarrow \,\ \ i\nabla _{\rho }\left( R^{\mu \nu }\beta ^{\rho }\psi \right) +i\tilde{\tau}_{\rho }\left( R^{\mu \nu }\beta ^{\rho }\gamma \psi \right) -R^{\mu \nu }\gamma \psi =0$$ $$R^{\mu \nu }\gamma \psi =i\left( \nabla _{\rho }+\tilde{\tau}_{\rho }\gamma \right) \left[ g^{\rho \nu }\left( R^{\mu }\psi \right) -g^{\rho \mu }\left( R^{\nu }\psi \right) \right]$$ from the above equations we get the equation of motion for the massless vector field $R^{\mu }\psi $ $$(\nabla _{\nu }+\tilde{\tau}_{\nu })(\nabla _{\rho }+\tilde{\tau}_{\rho }\gamma )\left[ g^{\rho \nu }\left( R^{\mu }\psi \right) -g^{\rho \mu }\left( R^{\nu }\psi \right) \right] =0\,,$$ We use a specific representation of the DKP algebra in which the singular $% \gamma $ matrix is $$\gamma =\mbox{diag}(0,0,0,0,1,1,1,1,1,1)~.$$Then in this representation the DKP field $\psi $ is now a 10-component column vector $$\psi =\left( \begin{array}{c} \psi ^{0},\psi ^{1},\psi ^{2},\psi ^{3},\psi ^{23},\psi ^{31},\psi ^{12},\psi ^{10},\psi ^{20},\psi ^{30}% \end{array}% \right) ^{T}\;,$$where $\psi ^{a}$ ($a=0,1,2,3$) and $\psi ^{ab}$ behave, respectively, as a 4-vector and an antisymmetric tensor under *Lorentz* transformations on the Minkowski tangent space. And we also get $$\begin{gathered} \gamma \psi =\left( \begin{array}{c} 0,0,0,0,\psi ^{23},\psi ^{31},\psi ^{12},\psi ^{10},\psi ^{20},\psi ^{30}% \end{array}% \right) ^{T} \\ R^{\mu }\psi =\left( \begin{array}{c} \psi ^{\mu },0,0,0,0,0,0,0,0,0% \end{array}% \right) ^{T} \\ R^{\mu \nu }\psi =\left( \begin{array}{c} \psi ^{\mu \nu },0,0,0,0,0,0,0,0,0% \end{array}% \right) ^{T}\end{gathered}$$ due to $R^{\mu }\gamma =\gamma R^{\mu }$ and $R^{\mu \nu }\gamma =(1-\gamma )R^{\mu \nu }$. Then, we get the following relations among $\psi $ components $$i\,\psi _{\mu \nu }=\nabla _{\mu }\psi _{\nu }-\nabla _{\nu }\psi _{\mu }$$which leads to the equation of motion for the spin $1$ sector of the massless DKP field in Lyra space-time $$(\nabla _{\mu }+\tilde{\tau}_{\mu })\left( {\nabla }^{\mu }\psi ^{\nu }-{% \nabla }^{\nu }\psi ^{\mu }\right) =0\,.$$ Energy-momentum tensor and spin tensor density ---------------------------------------------- Now, we only vary the background manifold and we assume that $\delta \omega _{\mu ab}$ and $\delta e^{\mu }{}_{a}$ are independent variations, $$\begin{gathered} \delta S=\int_{\Omega }dxe\phi ^{4}\left[ i\left( \bar{\psi}\gamma \beta ^{a}\nabla _{\mu }\psi -\nabla _{\mu }\bar{\psi}\beta ^{a}\gamma \psi \right) \delta e^{\mu }{}_{a}+\right. \\ \left. +\left( \frac{1}{e}\delta e\right) \mathcal{L}+i\left( \bar{\psi}% \gamma \beta ^{\mu }S^{ab}\psi +\bar{\psi}S^{ab}\beta ^{\mu }\gamma \psi \right) \frac{1}{2}\delta \omega _{\mu ab}\right] .\end{gathered}$$ First, holding only the variations in the tetrad field, $\delta \omega _{\mu ab}=0$, we found for the variation of the action $$\delta S=\int_{\Omega }dx~e\phi ^{4}~\left[ \frac{{}}{{}}i\left( \bar{\psi}% \gamma \beta ^{a}\nabla _{\mu }\psi -\nabla _{\mu }\bar{\psi}\beta ^{a}\gamma \psi \right) -e_{\mu }{}^{a}\mathcal{L}\right] \delta e^{\mu }{}_{a}$$ Defining the energy-momentum density tensor as $$T_{\mu }{}^{a}\equiv \frac{1}{\phi ^{4}e}\frac{\delta S}{\delta e^{\mu }{}_{a}}=i\bar{\psi}\gamma \beta ^{a}\nabla _{\mu }\psi -i\nabla _{\mu }\bar{% \psi}\beta ^{a}\gamma \psi -e_{\mu }{}^{a}\mathcal{L}$$which can be written in coordinates as $$T_{\mu }{}^{\nu }\equiv e^{\nu }{}_{a}T_{\mu }{}^{a}=i\bar{\psi}\gamma \beta ^{\nu }\nabla _{\mu }\psi -i\nabla _{\mu }\bar{\psi}\beta ^{\nu }\gamma \psi -\delta _{\mu }{}^{\nu }\mathcal{L}$$On the mass shell, $$T_{\mu }{}^{\nu }=i\bar{\psi}\gamma \beta ^{\nu }\nabla _{\mu }\psi -i\nabla _{\mu }\bar{\psi}\beta ^{\nu }\gamma \psi -\delta _{\mu }{}^{\nu }\bar{\psi}% \gamma \psi$$ Now, making functional variations only in the components of the spin connection, $\delta e^{\mu }{}_{a}=0$, we found for the action variation $$\delta S=\int_{\Omega }dx~e\phi ^{4}~\frac{1}{2}\left( \delta \omega _{\mu ab}\right) i\bar{\psi}\left( \gamma \beta ^{\mu }S^{ab}+S^{ab}\beta ^{\mu }\gamma \right) \psi ,$$we define the spin tensor density as being $$S^{\mu ab}\equiv \frac{2}{\phi ^{4}e}\frac{\delta S}{\delta \omega _{\mu ab}}% =i\bar{\psi}\left( \gamma \beta ^{\mu }S^{ab}+S^{ab}\beta ^{\mu }\gamma \right) \psi$$ The spin 1 component of DKP energy momentum tensor is$$\begin{aligned} T_{\mu }{}^{\nu } &=&\frac{i}{2}\,\psi ^{\ast \nu \alpha }\left( \nabla _{\mu }\,\psi _{\alpha }-\nabla _{\alpha }\,\psi _{\mu }\right) + \\ &&-\frac{i}{2}\,\psi ^{\nu \beta }\left( \nabla _{\mu }\,\psi _{\beta }^{\ast }-\nabla _{\beta }\,\psi _{\mu }^{\ast }\right) + \\ &&-\delta _{\mu }{}^{\nu }\left( \psi ^{\ast \alpha \beta }\psi _{\alpha \beta }\right)\end{aligned}$$which coincides with the first order energy momentum tensor of the electromagnetic field in the real case. Final Remarks ============= The coupling between torsion and massless vectorial field was showed to be related to scale transformations in Lyra background. Since this scale transformations are governed by an arbitrary function $\phi $, it seems plausible that the problem of breaking the local gauge invariance associated with this coupling could be removed from the theory if we had chosen an gauge transformations to be linked to scale invariance in Lyra manifold. A deeper study of this line is under construction. **Acknowledgments** This work is supported by FAPESP grants 01/12611-7 (RC), 01/12584-0 (CAMM) and 02/00222-9 (BMP). BMP also thanks CNPq for partial support. [99]{} M. C. Bertin, B. M. Pimentel and P. J. Pompeia, *First Order Actions: a New View* available as \[hep-th/0503064\]; L. Faddeev and R. Jackiw, Phys. Rev. Lett. **60**, 1692 (1988); D. M. Gitman and I. V. Tyutin, *Quantization of Fields with Constraints*, Springer-Verlag (1990); K. Sundermeyer, *Constrained Dynamics*, Lecture Notes in Physics Vol. 169, Springer-Verlag (1982). J. S. Schwinger, Phys. Rev.* ***82**, 914 (1951); Proc. Natl. Acad. Sci. U.S.A.* ***44**, 223 (1958). V. De Sabbata and M. Gasperini, *Introduction to Gravitation*, World Scientific (1985). R. Casana, V.Ya. Fainberg, J.T. Lunardi, B.M. Pimentel and R.G. Teixeira, Class. Quant. Grav. **20**, 2457 (2003). R. Casana, B. M. Pimentel, J. T. Lunardi and R. G. Teixeira, Gen. Rel. Grav. **34**, 1941 (2002). G. Lyra, *Math. Z.* **54**, 52 (1951); E. Scheibe, Math. Z. **57**, 65 (1952). D. K. Sen and J.R. Vanstone, J. Math. Phys. **13**, 990 (1972). G. Petiau, Acad. R. Soc. Belg. Cl. Sci. Mém. Collect. 8 16, No 2 (1936); R. J. Duffin, Phys. Rev. 54, 1114 (1938); N. Kemmer, Proc. R. Soc. **A 173**, 91 (1939). R.A. Krajcik and M.M. Nieto, Am. J. Phys. **45**, 818 (1977). J.T. Lunardi, B.M. Pimentel and R.G. Teixeira, in *Geometrical Aspects of Quantum Fields*, Proceedings of the 2000 Londrina Workshop, Londrina, Brazil; edited by A. A. Bytsenko, A. E. Gonçalves and B. M. Pimentel; World Scientific, Singapore (2001), p. 111. Also available as \[gr-qc/9909033\]. Harish-Chandra, Proc. Roy. Soc. Lond. **A** **186**, 502 (1946). C. W. Misner, K. S. Thorne and J. A. Wheeler, *Gravitation*, Freeman, San Francisco (1973). F. W. Hehl, P. von der Heyde and G. D. Kerlick, Rev. Mod. Phys. **48**, 393 (1976). E. C. G. Sudarshan and N. Mukunda, *Classical Dynamics: A Modern Perspective* (Wiley, 1974). H. Umezawa, *Quantum Field Theory*, North-Holland (1956). [^1]: We choose a representation in which ${\beta ^{0}}^{\dag }={\beta ^{0}}$, ${% \beta ^{i}}^{\dag }=-{\beta ^{i}}$ and $\gamma ^{\dag }=\gamma $ .

--- bibliography: - 'tenpy.bib' --- [ **Efficient numerical simulations with Tensor Networks: Tensor Network Python ([TeNPy]{})** ]{} Johannes Hauschild^1\*^, Frank Pollmann^1^ [**1**]{} Department of Physics, TFK, Technische Universit[ä]{}t M[ü]{}nchen, James-Franck-Stra[ß]{}e 1, D-85748 Garching, Germany\ \* johannes.hauschild@tum.de Abstract {#abstract .unnumbered} ======== [**Tensor product state (TPS) based methods are powerful tools to efficiently simulate quantum many-body systems in and out of equilibrium. In particular, the one-dimensional matrix-product (MPS) formalism is by now an established tool in condensed matter theory and quantum chemistry. In these lecture notes, we combine a compact review of basic TPS concepts with the introduction of a versatile tensor library for Python ([TeNPy]{}) [@tenpy]. As concrete examples, we consider the MPS based time-evolving block decimation and the density matrix renormalization group algorithm. Moreover, we provide a practical guide on how to implement abelian symmetries (e.g., a particle number conservation) to accelerate tensor operations.** ]{} ------------------------------------------------------------------------ ------------------------------------------------------------------------

--- abstract: 'We develop a new and further generalized form of the fractional kinetic equation involving generalized k-Bessel function. The manifold generality of the generalized k-Bessel function is discussed in terms of the solution of the fractional kinetic equation in the present paper. The results obtained here are quite general in nature and capable of yielding a very large number of known and (presumably) new results.' address: - 'P. Agarwal: Department of Mathematics, Anand International College of Engineering, Jaipur303012, India' - 'S. Jain: Department of Mathematics, Poornima College of Engineering, Jaipur-303029, India' - 'A. Atangana: Faculty / Fakulteit: Natural and Agricultural Sciences / Natuur-en Landbouwetenskappe PO Box / Posbus 339, Bloemfontein 9300, Republic of South Africa / Republiek van Suid-Afrika' - 'M. Chand: Department of Mathematics, Fateh College for Women, Bathinda 151103, India' - 'G. Singh:Department of Mathematics, Mata Sahib Kaur Girls College, Talwandi Sabo, Bathinda-151103 (India)Research Scholar, Department of Mathematics, Singhania University, Pacheri Bari, Jhunjhunu-(India)' author: - 'Praveen Agarwal, Shilpi Jain, Abdon Atangana, Mehar Chand and Gurmej Singh' title: 'Certain Fractional Kinetic Equations Involving Generalized k-Bessel Function' --- [^1] Introduction and Preliminaries ============================== In recent years, unified integrals involving Special functions attract the attention of the many researchers due to various application point of view(see, [@11; @12]). In the sequel, Diaz and Pariguan [@111] introduced the $k$-Pochhemmer symbol and $k$-gamma function defined as follows: $$\label{Poch} \aligned & (\lambda)_{n,k}: =\left\{\aligned & \frac{\Gamma_k(\gamma+nk)}{\Gamma_k(\gamma)} \hskip 41 mm (k\in\mathbb{R};\gamma \in \mathbb{C}\setminus \{0\}) \\ & \gamma(\gamma+k)...(\gamma+(n-1)k) \hskip20mm (n\in\mathbb{N};\gamma \in \mathbb{C}), \endaligned \right. \endaligned$$ They gave the relation with the classical Euler’s gamma function(see[@18; @10]) as: $$\label{Poch-1} \aligned \Gamma_k(\gamma)=k^{\frac{\gamma}{k}-1}\Gamma\left(\frac{\gamma}{k}\right) \endaligned$$ Clearly, for $ k=1 $, reduces to the classical Pochhemmer symbol and Euler’s gamma function, respectively (see[@116]) Recently ,Romero et. al.[@10] (see, also[@118]) introduced the k-Bessel function of the first kind for $\alpha,\lambda,\gamma,\nu\in \mathbb{C} $ and $ \Re(\lambda)>0, \Re(\nu)>0 $ as follows: $$\label{k-bessel} \aligned J^{(\gamma),(\lambda)}_{k,\nu}(z)=\sum^\infty_{n=0}\frac{(\lambda)_{n,k}}{\Gamma_k(\lambda n+\nu+1)}\frac{(-1)^n\left(\frac{z}{2}\right)^n}{(n!)^2} \endaligned$$ The Fox-Wright function ${}_p\psi_q(z)$ with $p$ numerator and $q$ denominators, such that $a_i,b_j\in\mathbb{C} (i=1,...,p;j=1,...,q)$ is defined by (see, for detail[@112]): $$\begin{aligned} \label{wright-function}{}_p\psi_q(z)={}_p\psi_q\left[ \left. \begin{array}{cc} (a_i,\alpha_i)_{1,p} \\(b_j,\beta_j)_{1,q} \end{array}\right|z \right]=\sum^\infty_{n=0}\frac{\prod^p_{i=1}\Gamma(a_i+\alpha_in)}{\prod^q_{j=1}\Gamma(b_j+\beta_Jn)}\frac{z^n}{n!} \end{aligned}$$ under the condition $$\begin{aligned} \sum^{q}_{j=1}\beta_j-\sum^p_{i=1}\alpha_i>-1 \end{aligned}$$ In particular, when $a_i=b_j=1(i=1,...,p;j=1,...,q),$ immediate reduces to the generalized hypergeometric function ${}_pF_q (p,q \in \mathbb{N}_0) $ (see, for details[@15]): $$\begin{aligned} \label{wright-function-case-1}{}_p\psi_q(z)={}_p\psi_q\left[ \left. \begin{array}{cc} (a_i,1)_{1,p} \\(b_j,1)_{1,q} \end{array}\right|z \right]=\frac{\prod^p_{i=1}\Gamma(a_i)}{\prod^q_{j=1}\Gamma(b_j)}{}_pF_q\left[\begin{array}{cc} a_1,...,a_p; \\b_1,...,b_q; \end{array} z\right] \end{aligned}$$ In terms of the $k$-Pochhamer symbol $(\gamma)_{n,k}$ defined by , we introduce more generalized form of k-Bessel function $\omega^{\gamma,\lambda}_{k,\nu,b,c}(z)$ as follows: $$\begin{aligned} \label{k-bessel-function}\omega^{\gamma,\lambda}_{k,\nu,b,c}(z)=\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\nu+\lambda n+\frac{b+1}{2})}\frac{\left(\frac{z}{2}\right)^{\nu+2n}}{(n!)^2} \end{aligned}$$ where $\alpha,\lambda,\gamma,\nu,c,b\in\mathbb{C}$ and $\Re(\lambda)>0,\Re(\nu)>0$. The importance of fractional differential equations in the field of applied science has gained more attention not only in mathematics but also in physics, dynamical systems, control systems and engineering, to create the mathematical model of many physical phenomena. Especially, the kinetic equations describe the continuity of motion of substance. The extension and generalization of fractional kinetic equations involving many fractional operators were found [@5; @6; @7; @8; @9; @10; @11; @12; @13; @14; @15; @16; @17; @18]. In view of the effectiveness and a great importance of the kinetic equation in certain astrophysical problems the authors develop a further generalized form of the fractional kinetic equation involving generalized k-Bessel function. The fractional differential equation between rate of change of the reaction was established by Haubold and Mathai[@7], the destruction rate and the production rate are given as follows: $$\begin{aligned} \label{D-1} \frac{dN}{dt}=-d(N_{t})+p(N_{t}), \end{aligned}$$ where $ N=N(t) $ the rate of reaction, $ d=d(N) $ the rate of destruction, $ p=p(N) $ the rate of production and $N_{t} $ denotes the function defined by $N_{t}(t^{*})= N(t-t^{*}), t^{*}>0 $ The special case of for spatial fluctuations and inhomogeneities in $ N(t) $ the quantities are neglected , that is the equation $$\begin{aligned} \aligned &\label{D-2} \frac{dN}{dt}=-c_{i}N_{i}(t),\endaligned \end{aligned}$$ with the initial condition that $ N_{i}(t=0)=N_{0} $ is the number density of the species $ i $ at time $ t=0 $ and $ c_{i}>0 $. If we remove the index $ i $ and integrate the standard kinetic equation , we have $$\begin{aligned} \aligned &\label{D-3} N(t)-N_{0}=-c{}_{0}D^{-1}_{t}N(t)\endaligned \end{aligned}$$ where $ {}_{0}D^{-1}_{t} $ is the special case of the Riemann-Liouville integral operator $ {}_{0}D^{-\nu}_{t} $ defined as $$\begin{aligned} \aligned &\label{D-4} {}_{0}D^{-\nu}_{t}f(t)=\frac{1}{\Gamma(\nu)}\int_{0}^{t}\left(t-s\right)^{\nu-1}f(s)ds,\hskip 6mm (t>0,R(\nu)>0)\endaligned \end{aligned}$$ The fractional generalization of the standard kinetic equation is given by Haubold and Mathai[@7] as follows: $$\begin{aligned} \label{D-5} N(t)-N_{0}=-c^{\nu}{}_{0}D^{-1}_{t}N(t) \end{aligned}$$ and obtained the solution of as follows: $$\begin{aligned} \aligned &\label{D-6} N(t)=N_{0}\sum _{k=0}^{\infty }\frac{(-1)^{k}}{\Gamma\left(\nu k+1\right)}\left(ct\right)^{\nu k}\endaligned \end{aligned}$$ Further, [@11] considered the the following fractional kinetic equation: $$\begin{aligned} \aligned &\label{D-7} N(t)-N_{0}f(t)=-c^{\nu}{}_{0}D^{-\nu}_{t}N(t),\hskip 6mm (\Re(v)>0),\endaligned\end{aligned}$$ where $ N(t) $ denotes the number density of a given species at time $ t $, $ N_{0}=N(0) $ is the number density of that species at time $t=0 $, $ c $ is a constant and $ f \in \mathcal{L}(0,\infty) $. By applying the Laplace transform to (see[@17]), $$\begin{aligned} \aligned &\label{D-8} L\left\{ N(t);p\right\}=N_{0}\frac{F(p)}{1+c^{\nu}p^{-\nu}}=N_{0}\left(\sum _{n=0}^{\infty }(-c^{\nu})^{n}p^{-\nu n}\right)F(p),\\&\hskip 6mm \left(n\in N_{0},\left|\frac{c}{p}\right|<1\right)\endaligned \end{aligned}$$ where the Laplace transform [@19] is given by $$\begin{aligned} \aligned &\label{D-9} F(p)=L\left\{ N(t);p\right\} =\int_{0}^{\infty} e^{-pt}f(t)dt,\hskip 6mm (\mathcal{R}(p)>0).\endaligned \end{aligned}$$ The objective of this paper is to derive the solution of the fractional kinetic equation involving generalized k-Bessel function. The results obtained in terms of Mittag-Leffler function are rather general in nature and can easily construct various known and new fractional kinetic equations. Solution of generalized fractional kinetic equations ==================================================== In this section, we will investigate the solution of the generalized fractional kinetic equations by considering generalized k-Bessel function. The results are as follows. \[Th-1\] If $ d>0, \nu>0, \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned &\label{Th-1-1} N(t)-N_{0}\omega^{\gamma,\lambda}_{k,\mu,b,c}(t)=-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned &\label{Th-1-2} N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n} E_{\nu,\hskip .5mm\mu+2n+1}(-d^{\nu}t^{\nu}),\endaligned \end{aligned}$$ where the generalized Mittag-Leffler function $ E_{\alpha,\beta}(x) $ is given by [@20] $$\begin{aligned} \label{2.3}\aligned &\label{ML} E_{\alpha,\beta}(x)=\sum _{n=0}^{\infty }\frac{(x)^{n}}{\Gamma\left(\alpha n+\beta\right)}.\endaligned \end{aligned}$$ Proof: the Laplace transform of Riemann-Liouville fractional integral operator is given by [@21] [@22] $$\begin{aligned} \aligned &\label{Th-1-3 } L\left\{{}_{0}D^{-\nu}_{t}f(t);p\right\}=p^{-\nu}F(p)\endaligned \end{aligned}$$ where $ F(p) $ is defined in .Now ,applying the Laplace transform to the both sides of gives $$\begin{aligned} \aligned &\label{Th-1-4}L\left\{ N(t);p\right\}=N_{0}L\left\{ \omega^{\gamma,\lambda}_{k,\mu,b,c}(t);p\right\}-d^{\nu}L\left\{{}_{0}D^{-\nu}_{t}N(t);p\right\}\endaligned \end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-5} N(p)=N_{0}\left(\int^{\infty}_{0}e^{-pt}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{1}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n}dt\right)\\&\hskip 6mm-d^{\nu}p^{-\nu}N(p)\endaligned\end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-6} N(p)+d^{\nu}p^{-\nu}N(p)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{1}{(n!)^2}\left(\frac{1}{2}\right)^{\mu+2n}\\&\hskip 6mm\times\int^{\infty}_{0}e^{-pt}t^{\mu+2n}dt\endaligned \end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-7} =N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{1}{(n!)^2}\left(\frac{1}{2}\right)^{\mu+2n}\\&\hskip 6mm\times\frac{\Gamma(\mu+2n+1)}{p^{\mu+2n+1}}\endaligned\end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-8} N(p)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{1}{2}\right)^{\mu+2n}\\&\hskip 6mm\times\left\{p^{-(\mu+2n+1)}\sum _{r=0}^{\infty }\left[-\left(\frac{p}{d}\right)^{-\nu}\right]^{r}\right\}\endaligned \end{aligned}$$ Taking Laplace inverse of ,and by using $$\begin{aligned} \aligned &\label{Th-1-9} L^{-1}\left\{p^{-\nu};t\right\}=\frac{t^{\nu-1}}{\Gamma(\nu)},(R(\nu)>0)\endaligned\end{aligned}$$ we have $$\begin{aligned} \aligned &\label{Th-1-10} L^{-1}\left\{N(p)\right\}=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{1}{2}\right)^{\mu+2n}\\&\hskip 6mm\times L^{-1}\left\{\sum _{r=0}^{\infty }(-1)^{r}d^{\nu r}p^{-(\mu+2n+1+\nu r)}\right\}\endaligned \end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-11} N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{1}{2}\right)^{\mu+2n}\\&\hskip 6mm\times\left\{\sum _{r=0}^{\infty }(-1)^{r}d^{\nu r}\frac{t^{(\mu+2n+\nu r)}}{\Gamma\left(\mu+2n+\nu r+1\right)}\right\}\endaligned\end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-12} =N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n}\\&\hskip 6mm\times\left\{\sum _{r=0}^{\infty }(-1)^{r}d^{\nu r}\frac{t^{(\nu r)}}{\Gamma\left(\mu+2n+\nu r+1\right)}\right\}\endaligned \end{aligned}$$ $$\begin{aligned} \aligned &\label{Th-1-13} N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n} E_{\nu,\mu+2n+1}(-d^{\nu}t^{\nu})\endaligned \end{aligned}$$ \[Th-2\] If $ d>0, \nu>0, \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned &\label{Th-2-1} N(t)=N_{0}\omega^{\gamma,\lambda}_{k,\mu,b,c}(d^{\nu}t^{\nu})-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned &\label{Th-2-2} N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-d^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . \[Th-3\] If $ a>0, d>0, \nu>0; a\neq d; \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned &\label{Th-3-1} N(t)=N_{0}\omega^{\gamma,\lambda}_{k,\mu,b,c}(d^{\nu}t^{\nu})-a^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned &\label{Th-3-2} N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^nc^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+\frac{b+1}{2})}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-a^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . Proof: The proof of theorem 2 and 3 would run parallel to those of theorem 1. Special Cases ============= If we choose $b=c=1$ then generalized k-Bessel function reduced to the following form: $$\begin{aligned} \omega^{\gamma,\lambda}_{k,\mu,1,1}(z)=\left(\frac{z}{2}\right)^{\mu}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\lambda n+\mu+1)}\frac{\left(\frac{z^2}{4}\right)^{n}}{(n!)^2}=\left(\frac{z}{2}\right)^{\mu}J^{(\gamma),(\lambda)}_{k,\mu}\left(\frac{z^2}{2}\right) ,\end{aligned}$$ where $\lambda,\gamma,\mu,\in\mathbb{C}$ and $\Re(\lambda)>0,\Re(\mu)>0$. Then the Theorems 1, 2 and 3 reduced to the following the form: If $ d>0, \nu>0, \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)-N_{0}\left(\frac{t}{2}\right)^{\mu}J^{(\gamma),(\lambda)}_{k,\mu}\left(\frac{t^2}{2}\right)=-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n} E_{\nu,\hskip .5mm\mu+2n+1}(-d^{\nu}t^{\nu}),\endaligned \end{aligned}$$ where $ E_{\nu,\mu+2n+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ d>0, \nu>0, \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu}J^{(\gamma),(\lambda)}_{k,\mu}\left(\frac{(d^{\nu}t^{\nu})^2}{2}\right)-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-d^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ a>0, d>0, \nu>0;a\neq d; \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu}J^{(\gamma),(\lambda)}_{k,\mu}\left(\frac{(d^{\nu}t^{\nu})^2}{2}\right)-a^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-a^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If we choose $b=-1,c=1$ then generalized k-Bessel function reduced to the k-Wright function [@23] associated with the following relation: $$\begin{aligned} \omega^{\gamma,\lambda}_{k,\mu,-1,1}(z)=\left(\frac{z}{2}\right)^{\mu}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\lambda n+\mu)}\frac{\left(\frac{z^2}{4}\right)^{n}}{(n!)^2}=\left(\frac{z}{2}\right)^{\mu}W^{\gamma}_{k,\lambda,\mu}\left(\frac{-z^2}{2}\right) \end{aligned}$$ where $\lambda,\gamma,\mu,\in\mathbb{C}$ and $\Re(\lambda)>0,\Re(\mu)>0$. Then the Theorems 1, 2 and 3 reduced to the following the form: If $ d>0, \nu>0, \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)-N_{0}\left(\frac{t}{2}\right)^{\mu}W^{\gamma}_{k,\lambda,\mu}\left(\frac{-t^2}{2}\right)=-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\left(\frac{t}{2}\right)^{\mu+2n} E_{\nu,\hskip .5mm\mu+2n+1}(-d^{\nu}t^{\nu}),\endaligned \end{aligned}$$ where $ E_{\nu,\mu+2n+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ d>0, \nu>0, \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu}W^{\gamma}_{k,\lambda,\mu}\left(\frac{-(d^{\nu}t^{\nu})^2}{2}\right)-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-d^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ a>0, d>0, \nu>0 ;a\neq d; \lambda,\gamma,\mu\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu}W^{\gamma}_{k,\lambda,\mu}\left(-\frac{(d^{\nu}t^{\nu})^2}{2}\right)-a^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\sum^\infty_{n=0}\frac{(-1)^n(\gamma)_{n,k}}{\Gamma_k(\mu+\lambda n+1)}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-a^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . By applying the results in equations and , after little simplification the Theorems 1, 2 and 3 reduced to the following form: If $ d>0, \nu>0, \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)-N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\,{}_1\psi_2\left[ \begin{array}{cc} (\gamma/k,1);\\(\mu/k+(b+1)/k,\lambda/k),(1,1);\end{array} t\right]\\&=-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\sum^\infty_{n=0}\frac{(-ck^{\lambda/k-1})^n}{\Gamma(\mu/k+\lambda n/k+\frac{b+1}{2k})}\frac{\Gamma(\mu+2n+1)}{(n!)^2}\\&\times\left(\frac{t}{2}\right)^{\mu+2n} E_{\nu,\hskip .5mm\mu+2n+1}(-d^{\nu}t^{\nu}),\endaligned \end{aligned}$$ where $ E_{\nu,\mu+2n+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ d>0, \nu>0, \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\,{}_1\psi_2\left[ \begin{array}{cc} (\gamma/k,1);\\(\mu/k+(b+1)/k,\lambda/k),(1,1);\end{array} d^{\nu}t^{\nu}\right]\\&-d^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\sum^\infty_{n=0}\frac{(-ck^{\lambda/k-1})^n}{\Gamma(\mu/k+\lambda n/k+\frac{b+1}{2k})}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}\\&\hskip 10mm\times E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-d^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . If $ a>0, d>0, \nu>0 ;a\neq d; \lambda,\gamma,\mu,c,b\in\mathbb{C}, k\in\mathbb{N}$ and $\Re(\lambda)>0,\Re(\mu)>0$ then the solution of the equation $$\begin{aligned} \aligned & N(t)=N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\,{}_1\psi_2\left[ \begin{array}{cc} (\gamma/k,1);\\(\mu/k+(b+1)/k,\lambda/k),(1,1);\end{array} d^{\nu}t^{\nu}\right]\\&-a^{\nu}{}_{0}D^{-\nu}_{t}N(t)\endaligned\end{aligned}$$ is given by the following formula $$\begin{aligned} \aligned & N(t)=N_{0}\frac{k^{1-\mu/k-(b+1)/2k}}{\Gamma(\gamma/k)}\sum^\infty_{n=0}\frac{(-ck^{\lambda/k-1})^n}{\Gamma(\mu/k+\lambda n/k+\frac{b+1}{2k})}\frac{\Gamma(\nu(\mu+2n)+1)}{(n!)^2}\\&\hskip 10mm\times \left(\frac{d^{\nu}t^{\nu}}{2}\right)^{\mu+2n}E_{\nu,\hskip .5mm\nu(\mu+2n)+1}(-a^{\nu}t^{\nu}),\endaligned\end{aligned}$$ where $ E_{\nu,\nu(\mu+2n)+1}(.) $ is the generalized Mittag-Leffler function defined in equation . Graphical Interpretation ======================== In this section we plot the graphs of main results established in equation , and . Graphs of the solution of the equation are depicted below for some parameter values i.e. $N_0=c=k=2;b=d=3;\mu=\nu=\gamma=1;\lambda=1,1.25,1.5,1.75,2$ in Fig. 1, Fig. 2 and Fig. 3 for time interval $t=0:1$, $t=0:2$ and $t=0:3$ respectively; graphs of the solution of the equation are depicted below for some parameter values i.e. $N_0=c=k=2;b=d=3;\mu=\nu=\gamma=1;\lambda=1,1.25,1.5,1.75,2$ in Fig. 4 and Fig. 5 for time interval $t=0:.05$ and $t=0:.06$ respectively. graphs of of the solution of the equation are depicted below for some parameter values i.e. $N_0=c=k=2;b=d=3;a=\mu=\nu=\gamma=1;\lambda=1,1.25,1.5,1.75,2$ in Fig. 6 and Fig. 7 for time interval $t=0:.05$ and $t=0:.06$ respectively. It is clear from these figures that $N(t)>0$ and the behavior of the solutions for different parameters and time interval can be studied and observed very easily. It is also observed that if we select $a=d$ in equations and give the identical solutions as we select in figures 4, 5, 6 and 7. Figures 4 and 6; 5 and 7 represents the identical solutions. Conclusion ========== In this work we give a new fractional generalization of the standard kinetic equation and derived solution for the same. From the close relationship of the k-Bessel function with many special functions, we can easily construct various known and new fractional kinetic equations. [99]{} Cerutti, R. A., 2012. On the $k$-Bessel functions, *International Mathematical forum*, 7(38),1851-1857. Choi, J., Kumar, D., 2015. Solutions of generalized fractional kinetic equations involving Aleph functions. Math. Commun. 20, 113-123. Choi, J., Agarwal, P., 2013. Certain unified integrals associated with Bessel functions,*Boundary Value Problems*, 1(2013), 1-9. Choi, J., Agarwal, P., 2013. Certain unified integrals involving a product of Bessel functions of the first kind,*Honam Mathematical Journal*, 35(4), 667-677. Chouhan, A., Sarswat, S., 2012. On solution of generalized kinetic equation of fractional order. Int. J. Math. Sci. Appl. 2(2),813-818. Chouhan, A., Purohit, S.D., Saraswat, S., 2013. An alternative method for solving generalized differential equations of fractional order. Kragujevac J. Math. 37(2), 299-306. Chaurasia, V.B.L., Pandey, S.C., 2008. On the new computable solution of the generalized fractional kinetic equations involving the generalized function for the fractional calculus and related functions. Astrophys. Space Sci. 317, 213-219. Diaz, R. and Pariguan, E., 2007. On hypergeometric functions and $k$-Pochhammer symbol, *Divulgaciones Mathematicas*, 15(2), 179-192. Gupta, V.G., Sharma, B., 2011. On the solutions of generalized fractional kinetic equations. Appl. Math. Sci. 5(19), 899-910. Erdelyi, A., Magnus, W., Oberhettinger, F., Tricomi, F.G., 1954. In: Tables of Integral Transforms, vol. 1. McGraw-Hill, New York-Toronto-London. Fox, C., 1928. The asymptotic expansion of generalized hypergeometric functions, *Proc. London. Math. Soc.* 27(4), 389-400. Gupta, A., Parihar, C.L., 2014. On solutions of generalized kinetic equations of fractional order. Bol. Soc. Paran. Mat. 32 (1), 181-189. Haubold, H.J., Mathai, A.M., 2000. The fractional kinetic equation and thermonuclear functions. Astrophys. Space Sci. 327, 53–63. function of the first kind. Math. Probl. Eng. 2015, 7 Article ID 289387. Kumar, D., Purohit, S.D., Secer, A., Atangana, A., 2015. On generalized fractional kinetic equations involving generalized Bessel, Mathematical Problems in Engineering, 2015, Article ID 289387, 7 pages. http://dx.doi.org/10.1155/2015/289387 Spiegel, M.R., 1965. Theory and Problems of Laplace Transforms, Schaums Outline Series. McGraw-Hill, New York. Mittag-Leffler, G.M., 1905. Sur la representation analytiqie d’une fonction monogene cinquieme note. Acta Math. 29, 101-181. Rainville,E. D., 1960. Special functions, Macmillan, New York. Romero, L; Cerutti, R. Fractional calculus of a $k$-Wright type function. To appear. Romero, L. G., Dorrego,G. A. and Cerutti, R. A., 2014. The $k$-Bessel function of first kind, *International Mathematical forum*,38(7), 1859-1854. Saichev, A., Zaslavsky, M., 1997. Fractional kinetic equations: solutions and applications. Chaos 7, 753-764. Saxena, R.K., Mathai, A.M., Haubold, H.J., 2002. On fractional kinetic equations. Astrophys. Space Sci. 282, 281-287. Saxena, R.K., Mathai, A.M., Haubold, H.J., 2004. On generalized fractional kinetic equations. Physica A 344, 657-664. Saxena, R.K., Mathai, A.M., Haubold, H.J., 2006. Solution of generalized fractional reaction-diffusion equations. Astrophys. Space Sci. 305, 305-313. Saxena, R.K., Kalla, S.L., 2008. On the solutions of certain fractional kinetic equations. Appl. Math. Comput. 199, 504-511. Srivastava, H.M., Saxena, R.K., 2001. Operators of fractional integration and their applications. Appl. Math. Comput. 118, 1-52. Zaslavsky, G.M., 1994. Fractional kinetic equation for Hamiltonian chaos. Physica D 76, 110-122. [^1]: $^{*}$ corresponding author

--- abstract: 'We present a study of planar physical solutions to the Lorentz-Dirac equation in a constant electromagnetic field. In this case, we reduced the Lorentz-Dirac equation to the one second order differential equation. We obtained the asymptotics of physical solutions to this equation at large proper times. It turns out that, in the crossed constant uniform electromagnetic field with vanishing invariants, a charged particle goes to a universal regime at large times. We found the ratio of momentum components which tends to a constant determined only by the external field. This effect is essentially due to a radiation reaction. There is not such an effect for the Lorentz equation in this field.' author: - 'P.O. Kazinski' - 'M.A. Shipulya' title: 'Asymptotics of physical solutions to the Lorentz-Dirac equation for a planar motion in constant electromagnetic fields' --- Introduction ============ The Lorentz-Dirac (LD) equation has an ill fame to suffer from the various type inconsistencies. The latter come from the higher derivative Schott term entering the LD equation and appear as the blowing up (runaway) and acausal solutions. However, despite of its undesirable features, we have to accept the LD equation as a correct one in its range of applicability by the following reasons. First, as it was shown in the seminal paper by Dirac [@Dir] and then more elaborately in [@Teit], the LD equation stems from the energy-momentum conservation law provided a charged particle is sufficiently small and possesses neglible higher multipoles of a charge distribution. Second, under these assumptions the LD equation is a minimal evolutionary equation describing a radiation reaction which complies with all the symmetries of the model: the Poincare and reparameterization invariance. Furthermore, having made certain approximations, the LD equation was derived in the context of quantum electrodynamics (see, e.g., [@JoHu]), where the LD equation can be considered as a leading quasiclassical asymptotics to the Schwinger-Dyson equations for an electron. Therefore, the LD equation makes a physical sense and, under certain conditions, its solutions should give rise to predictions which can be observed in experiments. It is clear that the LD equation is valid in the range of energies and field strengths where the quantum corrections are neglible in comparison with the classical contribution. Rough general estimates of this range can be found, e.g., in [@Shen; @Klepik], and the more accurate analysis for the particular case of a constant homogeneous magnetic field is presented in [@Bagrov]. Various generalizations of the LD equation to include spin and higher multipoles [@Frenk], or an interaction with non-Abelian gauge fields [@Wong] and gravity [@DeWBr], to higher dimensions [@Kos], to dyonic [@Rohrl] and massless charged particles [@rrmlp] are also known. All of them have the higher derivative terms and, hence, possess the same unwanted properties as the LD equation. There is a coherent approach [@Bhabha; @Plass; @GuptaEl; @Barut; @Klepik; @RohrlBook] how to extract a physical information from the LD equation and its analogues. It is based on the notion of a physical solution. In a general setting, it looks as follows. Given a system of interacting fields $\phi^{\gamma}_a$ with the action functional $$\label{action0} S[\phi^{\gamma}_1,\ldots,\phi^{\gamma}_N]=\sum_{a=1}^N{\lambda}^{-1}_a S^0_a[\phi_a^{\gamma}]+S_{int}[\phi^{\gamma}_1,\ldots,\phi^{\gamma}_N],$$ where ${\gamma}$ is a condensed index representing group and spacetime indices and spacetime points, $a$ numerates the fields, and ${\lambda}_a$ are some constants. Then the solution $\phi^{\gamma}_a({\lambda})$ of the coupled system of equations of motion corresponding to the action at given initial and boundary conditions is called physical if there exist finite limits $$\lim_{{\lambda}_b\rightarrow0}\phi^{\gamma}_a({\lambda}),\qquad a=\overline{1,N},$$ other lambdas being fixed. This regularity condition completely rules out the runaway solutions to the LD equation. Besides, its physical solutions are unambiguously specified by the six initial data – three position coordinates and three momentum components – as it should be in the realm of Newtonian mechanics. Since the LD equation is nonlinear, it is hard to solve it even in simple external field configurations. Almost all the exact solutions to the LD equation can be found in [@Plass; @RohrlBook; @SokTer; @DeRaStTr; @GuptaEl]. In this paper, we address the problem of a finding and description of exact physical solutions to the LD equation in a constant homogeneous electromagnetic field. Moreover, we restrict ourself by a planar motion only. Even in this rather simple situation we did not succeed in finding the exact essentially planar (i.e. non linear) solutions to the LD equation. However, we reduce the LD equation to the one second order differential equation and investigate the asymptotics of its physical solutions at large times. For the constant homogeneous magnetic field this asymptotics was found in [@Plass]. As far as the constant electric and crossed fields are concerned these asymptotics, to our knowledge, are obtained for the first time. It turns out that, in the crossed field configuration, the LD equation possesses an attractor and the system passes into a universal regime at large times. After the lapse of time, the identical charged particles moving on the plane in such an electromagnetic field “forget” their initial data. Their trajectories become parallel and a certain ratio of momenta components tends to a constant which is independent of the initial conditions and determined by the external field only. This effect is essentially due to the radiation reaction. It is absent for the solution to the Lorentz equation in this field and can serve as an explicit manifestation of a validity of the classical radiation reaction theory in the domain of its applicability. The paper is organized as follows. In Sec. \[general\], we present general formulas regarding a radiation reaction and define the physical solutions to the LD equation. Here we also give an integro-differential equation for the physical solutions. Sec. \[linplan\] is a main part of the article. In Sec. \[linmot\], we briefly describe the linear motion and exact solutions to the LD equation in this case. In Sec. \[planmotgen\], we investigate the symmetries of the LD equation and provide the necessary and sufficient condition for the motion of a charged particle to be planar. In Sec. \[planmotsecor\], we derive the second order differential equation describing planar solutions to the LD equation. Sec. \[planmotasym\] is devoted to the asymptotics of the physical solutions to the LD equation at large times. In Sec. \[planmotLL\], we consider the same problem in the framework of the so-called Landau-Lifshitz equation [@LandLifsh]. In conclusion, we summarize the main results of the paper and discuss the prospects for a further research. General formulas {#general} ================ Consider a particle with the charge $e$ and mass $m$ interacting with the electromagnetic field $A_\mu$ on Minkowski background $\mathbb{R}^{1,3}$ with the metric $\eta_{\mu\nu}=diag(1,-1,-1,-1)$ and coordinates $x^\mu$, $\mu=\overline{0,3}$. The action functional for such a system has the form $$\label{action particl} S[x(\tau),A(x)]=-m\int{d\tau\sqrt{\dot{x}^2}}-e\int{d\tau A_\mu \dot{x}^\mu}-\frac1{16\pi}\int{d^4xF_{\mu\nu}F^{\mu\nu}},$$ where $x^\mu(\tau)$ defines the particle worldline, $F_{\mu\nu}:=\partial_{[\mu}A_{\nu]}$ is the strength tensor of the electromagnetic field (the square brackets denote an antisymmetrization without $1/2$) $$\label{fmunu} F_{\mu\nu}=\begin{bmatrix} 0 & E_x & E_y & E_z \\ -E_x & 0 & -H_z & H_y \\ -E_y & H_z & 0 & -H_x \\ -E_z & -H_y & H_x & 0 \\ \end{bmatrix},$$ and we take the system of units in which the speed of light $c=1$. In the proper time parameterization $\dot{x}^2=1$, the LD equation [@Lor; @Dir] reads as $$\label{lde_ini} m\ddot{x}_\mu=eF_{\mu\nu}\dot{x}^\nu+\frac23e^2(\dddot{x}_\mu+\ddot{x}^2\dot{x}_\mu),$$ where $F_{\mu\nu}$ is the strength tensor of the external electromagnetic field. Introducing the dimensionless quantities $$x^\mu\rightarrow m^{-1}x^\mu,\qquad\tau\rightarrow m^{-1}\tau,\qquad F_{\mu\nu}\rightarrow m^2e^{-1}F_{\mu\nu},$$ we rewrite it in the form $$\label{lde} \dot{{\upsilon}}_\mu=f_\mu+{\lambda}(\ddot{{\upsilon}}_\mu+\dot{{\upsilon}}^2{\upsilon}_\mu),\qquad f_\mu:=F_{\mu\nu}{\upsilon}^\nu,$$ where ${\lambda}:=2e^2/3$ and $m{\upsilon}^\mu:=m\dot{x}^\mu$ is the $4$-momentum of the particle. The LD equation possesses unphysical solutions. Following [@Bhabha; @GuptaEl], we shall call the solution $x^\mu({\lambda},\tau)$ physical if it tends to the solution $x^\mu(0,\tau)$ of the corresponding Lorentz equation as ${\lambda}$ goes to zero. It is a realization of the general definition given in Introduction in the case of classical electrodynamics. According to this definition, the physical solution should be regular at the large mass $m$ and small $e^2$. From Eq. we see that this requirement leads to a regularity of the physical solution with respect to the external field and the parameter ${\lambda}$ entering . The former simply follows from the general theorems regarding dependence of solutions to ordinary differential equations on a parameter (see, e.g., [@Golubev]), while the latter condition is not trivial. Also notice that all the known physically reasonable solutions to the LD equation are physical in the sense adopted by us. Some extra arguments in favor of this definition of physical solutions are given in Appendix. We can find these solutions perturbatively as a (formal) series in ${\lambda}$. This perturbative scheme reduces the order of the LD equation and provides a unique solution to it at some fixed initial position and velocity of the particle. The first iteration of this perturbative procedure yields the Landau-Lifshitz equation [@LandLifsh]. It is not difficult to write an integro-differential equation which describes the physical solutions to the LD equation [@Barut; @Plass; @Klepik; @RohrlBook]. In the proper time parameterization, it looks like $$\label{lde_integral} \dot{{\upsilon}}_\mu(\tau)=\operatorname{pr}_\mu^\nu(\tau)\int_0^\infty dte^{-t}\mathcal{P}_\nu(\tau+{\lambda}t),\qquad\operatorname{pr}_\mu^\nu:={\delta}^\nu_\mu-{\upsilon}_\mu{\upsilon}^\nu,$$ where $\mathcal{P}_\mu=f_\mu+{\lambda}\dot{{\upsilon}}^2{\upsilon}_\mu$. Solutions to Eq. are solutions to the LD equation with ${\upsilon}_\mu\dot{{\upsilon}}^\mu=0$. It is the latter requirement that gives rise to the projector entering Eq. . The solutions of this equation tend to the solutions of the Lorentz equation at ${\lambda}\rightarrow0$. Expanding Eq. in a series in ${\lambda}$, we see that solutions to Eq. are those solutions to the LD equation which are obtained from it by the aforementioned perturbative scheme. If we knew all the terms of the perturbation series for the acceleration $\dot{{\upsilon}}(\tau,{\lambda})$ then formula would tell us that this series in ${\lambda}$ must be summed by the Borel method [@Hardy]. So, if the following conditions are satisfied at some fixed initial position and velocity 1. There exists a unique solution to the corresponding Lorentz equation, which is defined at any $\tau>\tau_0$ and tends to infinity not faster than $Me^{a\tau}$ at $\tau\rightarrow\infty$. Here, $\tau_0$, $M$ and $a>0$ are some constants; 2. The perturbative series in ${\lambda}$ converges absolutely in a vicinity of the point ${\lambda}=0$ at sufficiently small ${\lambda}$; then a unique solution to Eq. exists at sufficiently small ${\lambda}={\lambda}_0>0$. If the value of ${\lambda}_0$ is smaller than the physical value of ${\lambda}$ then the physical solution to the LD equation at the physical value of ${\lambda}$ is obtained by an analytical continuation in ${\lambda}$. A concrete prescription how to construct this analytical continuation depends on analytical properties of the solution to the Lorentz equation. For example, if this solution satisfies the first condition above and has a finite number of singularities in the part of the complex $\tau$-plane where $\operatorname{Re}\tau>\tau_0$ and $\operatorname{Im}\tau>0$, or $\operatorname{Re}\tau>\tau_0$ and $\operatorname{Im}\tau<0$ for some $\tau_0$, then $x^\mu({\lambda},\tau)$ has the same properties at sufficiently small ${\lambda}$. In that case, we can rotate the ray along which the integration contour tends to infinity so as to make the integral convergent for any $a$. In particular, this procedure makes the right hand side of Eq. convergent when we perturbatively solve Eq. by Picard iterations starting with the solution to the Lorentz equation $x^\mu(0,\tau)$ satisfying the first condition above, while $3a{\lambda}\geq1$. Of course, the latter situation is rather unphysical since $$a{\lambda}\sim{\alpha}\frac{H}{H_0},\qquad H_0=\frac{m^2}{|e|\hbar},$$ where $H$ is a characteristic value of the field strength, ${\alpha}$ is the fine structure constant and $H_0$ is the Schwinger field. However, we can define the physical solution even in this case. Another, possibly more convenient, form of Eq. can be derived if we write [@Plass] $$\dot{{\upsilon}}^2=\int_0^\infty dte^{-t}(\dot{{\upsilon}}f)(\tau+{\lambda}t/2),$$ for physical solutions. Then $$\label{lde_integral1} \dot{{\upsilon}}_\mu(\tau)=\operatorname{pr}_\mu^\nu(\tau)\int_0^\infty dte^{-t}\left[f_\nu(\tau+{\lambda}t)+{\lambda}\int_0^tds(\dot{{\upsilon}}f)(\tau+{\lambda}s/2){\upsilon}_\nu\left(\tau+{\lambda}(t-s)\right)\right].$$ In this form, it is obvious that, in the absence of external fields, physical solutions to the LD equation are straight lines in the spacetime. The integro-differential equations for physical solutions to the LD equation, which we have presented in this section, are not very useful in finding analytical solutions. But they are pertinent to numerical simulations, for example, by the Picard iterations, rather than the LD equation itself. Even if we set the initial conditions to their physical values and would solve numerically the LD equation then we shall obtain an unphysical runaway solution owing to machine approximation errors. Also, Eqs. and are a good starting point for an investigation of the stochastic LD equation (see, e.g., [@JoHu; @sd]). As in the case of numerical simulations, the unphysical solutions should be explicitly excluded to get rid of stochastically induced runaways. The study of this problem will be given elsewhere. Physical solutions for a planar motion {#linplan} ====================================== In this section, we study the planar motion of a charged particle obeying the LD equation. We call the motion of a particle planar (linear) provided that the trajectory of this particle in the space can be made planar (linear) by an appropriate Lorentz transform. For the linear motion, the worldline of a particle lies in a two-dimensional plane of the spacetime. For the planar motion, it lies in a three-dimensional hyperplane. Throughout of this section, we mostly assume that the particle moves in the constant homogeneous external electromagnetic field, but some results can be generalized to a non-constant electromagnetic field of the special configuration. It will be mentioned in its place. Linear motion {#linmot} ------------- Let us consider, at first, the hyperbolic motion, which is the motion with vanishing the LD force, $$\begin{gathered} \label{hyperbol_mot} \ddot{{\upsilon}}_\mu+\dot{{\upsilon}}^2{\upsilon}_\mu=0\;\Rightarrow\;\dot{{\upsilon}}^2=-\omega^2=const\;\Rightarrow\\ {\upsilon}_\mu(\tau)={\alpha}_\mu\cosh(\omega\tau)+{\beta}_\mu\sinh(\omega\tau),\quad{\alpha}^2=-{\beta}^2=1,\quad{\alpha}_\mu{\beta}^\mu=0.\end{gathered}$$ The hyperbolic motion is the solution to the LD equation with a constant and homogeneous external electromagnetic field if, and only if, $$\ddot{{\upsilon}}^\mu=F^\mu_{\ \rho}F^\rho_{\ \nu}{\upsilon}^\nu=-\omega^2{\upsilon}^\mu.$$ By the use of canonical forms (see [@DubNovFom] and also below) of the tensor $F_{\mu\nu}$, it is not difficult to show that the last equality (the equation on eigenvectors) is fulfilled if, and only if, there exists a Lorentz frame in which the particle moves along $3$-vectors of the electric and magnetic field strengths. In this system of coordinates, the problem reduces to a description of the linear motion. For a linear motion, the LD equation is equivalent to (see, e.g., [@DeRaStTr; @Plass; @Dir] ) $$\label{solution1d} \frac{\dot{{\upsilon}}}{\sqrt{1+{\upsilon}^2}}={\lambda}\frac{d}{d\tau}\frac{\dot{{\upsilon}}}{\sqrt{1+{\upsilon}^2}}+E\;\Rightarrow\;{\upsilon}(\tau)=\sinh(c_2+E\tau+c_1e^{{\lambda}^{-1}\tau}).$$ where ${\upsilon}:=\dot{x}(\tau)$. A generalization of this solution to the case of the electric field depending on $\tau$ is trivial. The solution with $E=0$ is also a general solution to the free LD equation written in the Lorentz frame, where the initial $3$-velocity and $3$-acceleration are parallel. The solution becomes physical if we take $c_1=0$. It is easy to verify that the solution satisfies the integro-differential equation only at vanishing $c_1$. In order to make the integral convergent at ${\lambda}E\geq1$, we have to rotate the integration contour in Eq. as it was described in the previous section. Planar motion ------------- ### General consideration {#planmotgen} Let us turn to the planar motion. When one of the Poincare-invariants of the electromagnetic field is not zero, the strength tensor can be represented as $$\label{strength_nondeg} F^{\mu\nu}=\omega_1e_0^{[\mu} e_1^{\nu]}+\omega_2e_2^{[\mu} e_3^{\nu]},\qquad (e_{\alpha}e_{\beta})=\eta_{{\alpha}{\beta}},$$ where $e_{\alpha}^\mu$, ${\alpha}=\overline{0,3}$, is a tetrad of eigenvectors of the tensor $(F^2)^\mu_\nu$. The eigenvalue $\omega^2_1$ of the tensor $(F^2)^\mu_\nu$ corresponds to the vectors $e^\mu_{0,1}$, and the eigenvalue $-\omega^2_2$ corresponds to the vectors $e^\mu_{2,3}$. In terms of the Poincare-invariants of the electromagnetic field $I_1=\mathbf{E}^2-\mathbf{H}^2$ and $I_2=2(\mathbf{EH})$, these omegas read as $$\omega^2_1=(\sqrt{I_1^2+I_2^2}+I_1)/2,\qquad\omega^2_2=(\sqrt{I_1^2+I_2^2}-I_1)/2.$$ The Lorentz transforms, which do not change the strength tensor , constitute the group $SO(1,1)\times SO(2)$. Its matrix representation is obvious. In the degenerate case, when $I_1=I_2=0$, the strength tensor is given by $$\label{strength_deg} F^{\mu\nu}=\omega e_-^{[\mu}e_1^{\nu]},\qquad (e_ae_b)=\begin{bmatrix} 0 & 0 \\ 0 & -1 \\ \end{bmatrix},$$ where $e_a^\mu$, $a=\{-,1\}$, are eigenvectors of the tensor $(F^2)^\mu_\nu$ corresponding to zero eigenvalue. The normalized eigenvector $e_1^\mu$ is orthogonal to the vector $e_3^\mu$, which is a normalized eigenvector of $F^\mu_{\ \nu}$ and $(F^2)^\mu_\nu$ corresponding to zero eigenvalue. These conditions determine the vectors $e_1$ and $e_3$ uniquely up to adding the isotropic vector $e_-$ and inversion. The factor $\omega$ can be included to $e_-$, but we leave it in the expression so as to control the external field. The strength tensor is invariant with respect to the two-dimensional Abelian subgroup of the Lorentz group generated by the two elements $${\Lambda}_1^{\mu\nu}=\eta^{\mu\nu}+r_1e_3^{[\mu}e_-^{\nu]}+\frac{r_1^2}2e_-^\mu e_-^\nu,\qquad {\Lambda}_2^{\mu\nu}=\eta^{\mu\nu}+r_2e_1^{[\mu}e_-^{\nu]}+\frac{r_2^2}2e_-^\mu e_-^\nu,$$ where $r_1$ and $r_2$ are the group parameters. Thus, in both degenerate and non-degenerate cases, we anticipate two integrals of motion of the LD equation provided the eigenvectors of the tensors $F^\mu_{\ \nu}$ and $(F^2)^\mu_\nu$ do not depend on a point of the spacetime. In order to obtain these integrals of motion, it is useful to introduce new variables adjusted to the action of the symmetry group. In the non-degenerate case, they are $$\label{subs_eh} \mathrm{v}_0=:\sqrt{u_e}\cosh\psi,\qquad \mathrm{v}_1=:\sqrt{u_e}\sinh\psi,\qquad \mathrm{v}_2=:\sqrt{u_h}\sin{\varphi},\qquad \mathrm{v}_3=:\sqrt{u_h}\cos{\varphi},$$ where $\mathrm{v}_{\alpha}:=(e_{\alpha}{\upsilon})$, $u_e-u_h=1$, and ${\varphi}$ and $\psi$ are the symmetry group parameters. Then, convolving the LD equation with the eigenvectors, we have $$\label{integrals_mot} \begin{split} u_e(\tau)\dot{\psi}(\tau)&=c_1e^{{\lambda}^{-1}\tau}+\int_\tau^\infty\frac{dt}{{\lambda}}e^{-{\lambda}^{-1}(t-\tau)}\omega_1 u_e(t),\\ u_h(\tau)\dot{{\varphi}}(\tau)&=c_2e^{{\lambda}^{-1}\tau}-\int_\tau^\infty\frac{dt}{{\lambda}}e^{-{\lambda}^{-1}(t-\tau)}\omega_2 u_h(t). \end{split}$$ In the degenerate case, the analogous variables read as $$\label{subs_iso} \mathrm{v}_1=:\mathrm{v}_-r_2,\qquad \mathrm{v}_3=:\mathrm{v}_-r_1,\qquad \mathrm{v}_+=:\mathrm{v}_-^{-1}+\mathrm{v}_-(r_1^2+r_2^2),$$ where $e_+^\mu$ is an isotropic vector orthogonal to $e_1^\mu$ and $e_3^\mu$, and such that $(e_+e_-)=2$. The respective integrals of motion become $$\label{integrals_mot1} \begin{split} u(\tau)\dot{r}_1(\tau)&=c_1e^{{\lambda}^{-1}\tau},\\ u(\tau)\dot{r}_2(\tau)&=c_2e^{{\lambda}^{-1}\tau}+\int_\tau^\infty\frac{dt}{{\lambda}}e^{-{\lambda}^{-1}(t-\tau)}\omega u(t), \end{split}$$ where $u(\tau):=\mathrm{v}_-^2(\tau)$. The non-vanishing constants $c_1$ and $c_2$ in Eqs. and correspond to unphysical solutions. So, they should be set to zero and we do not take them into account henceforth. Consider the particular case $I_2=0$ (see [@GuptaEl]). If $I_1>0$ then $\omega_2$ is zero and we get $$\mathrm{v}_2=\mathrm{v}_3=0\quad\text{or}\quad\dot{{\varphi}}=0.$$ The first case is a linear motion, the second case is a planar one. When $I_1<0$, we have $\omega_1=0$ and $$\dot{\psi}=0.$$ In the degenerate case $I_1=I_2=0$, $$\dot{r}_1=0.$$ By definition of the variables $\psi$ and $r_1$ as the group parameters, the particle can be confined to a plane by an appropriate Lorentz transform in these cases. Thus we have proved the following statement. If $I_2=0$ and the field strength tensor admits the representation or with the constant eigenvectors $e_{\alpha}^\mu$ then a charged particle obeying the LD equation executes a planar motion. In a constant homogeneous electromagnetic field, the converse statement is also true. Namely, if a charged particle obeying the LD equation executes an essentially planar (i.e. non linear) motion then $I_2=0$. ### Second order equation {#planmotsecor} Now we investigate the planar motion in detail. Let us characterize this motion by a tetrad $$e^\mu_{\alpha}\eta_{\mu\nu}e^\nu_{\beta}=\eta_{{\alpha}{\beta}},\qquad e^\mu_3 e^{\alpha}_\mu=0,\qquad (e_3)^2=-1,$$ where the indices ${\alpha}$ and ${\beta}$ run the values $0,1,2$. They are risen and lowered by the metric $\eta_{{\alpha}{\beta}}=diag(1,-1,-1)$. The worldline of the particle and the external electromagnetic field admit a representation $$\begin{gathered} {\upsilon}^\mu(\tau)={\upsilon}^{\alpha}(\tau)e_{\alpha}^\mu,\qquad e^\mu_3{\upsilon}_\mu(\tau)=0,\\ F_{\mu\nu}=f_{{\alpha}{\beta}}e^{\alpha}_\mu e^{\beta}_\nu,\qquad f_{{\alpha}{\beta}}=\omega{\varepsilon}_{{\alpha}{\beta}{\gamma}}\xi^{\gamma}, \end{gathered}$$ where ${\varepsilon}_{012}=1$ and $\xi^2=\{\pm1,0\}$. The LD equation is rewritten as (for the Lorentz equation see, e.g., [@Plyushch]) $$\label{lde tetr} \dot{{\upsilon}}_{\alpha}=\omega{\varepsilon}_{{\alpha}{\beta}{\gamma}}{\upsilon}^{\beta}\xi^{\gamma}+{\lambda}(\ddot{{\upsilon}}_{\alpha}+\dot{{\upsilon}}^2{\upsilon}_{\alpha}),\qquad{\upsilon}^2=1.$$ We see that the LD equation of a charged particle confined to a plane possesses a symmetry. This is a residue of the symmetry discussed above after the reduction to the plane. The residue symmetry group is constituted by the Lorentz transforms leaving the vector $\xi^{\alpha}$ intact. So, if $\xi^2\leq0$ this symmetry group is isomorphic to $SO(1,1)$, and if $\xi^2>0$ it is isomorphic to $SO(2)$. This symmetry allows us to reduce the problem of integration of the system of equations to an integration of the autonomous system of three first order equations or the one second order equation. To this end, we introduce new more convenient variables $$m^{\alpha}:={\varepsilon}^{{\alpha}{\beta}{\gamma}}\dot{{\upsilon}}_{\beta}{\upsilon}_{\gamma}.$$ In these variables, the LD equation turns into a system of the first order equations $$\label{lde in m} {\lambda}\dot{m}_{\alpha}=m_{\alpha}+\omega(\xi_{\alpha}-p{\upsilon}_{\alpha}),\qquad\dot{{\upsilon}}_{\alpha}=-{\varepsilon}_{{\alpha}{\beta}{\gamma}}m^{\beta}{\upsilon}^{\gamma},\qquad m_{\alpha}{\upsilon}^{\alpha}=0,\qquad{\upsilon}^2=1,$$ where $p:=\xi_{\alpha}{\upsilon}^{\alpha}$. Then we introduce the invariants of the symmetry group action $$\label{invar} \begin{aligned} a&=\xi_{\alpha}m^{\alpha},&\qquad b&=p^{-1}{\varepsilon}^{{\alpha}{\beta}{\gamma}}\xi_{\alpha}m_{\beta}{\upsilon}_{\gamma}=-p^{-1}\dot{p},\\ s&=p^{-2}(m^2-\xi^2a^2),&\qquad u&=p^2-\xi^2. \end{aligned}$$ These invariants are dependent. From their definition, it is not difficult to obtain the identity $$\label{ident} a^2\frac{1+\xi^2u}{\xi^2+u}+b^2=-su.$$ The invariants evolve according to the equations $$\label{invar_evol0} \begin{aligned} {\lambda}\dot{a}&=a-\omega u,&\qquad\dot{u}&=-2b(\xi^2+u),\\ {\lambda}\dot{s}&=2s(1+{\lambda}b)+2\omega a\frac{1+\xi^2u}{\xi^2+u},&\qquad {\lambda}\dot{b}&=b+{\lambda}\xi^2s+{\lambda}a^2\frac{|\xi^2|-1}{u}=b-{\lambda}\frac{a^2+\xi^2b^2}{u}. \end{aligned}$$ The first equation in this system is the one of the equations of the system or written in a differential form. The physical solutions are described by the one integro-differential equation on $u(\tau)$. It is obtained from the above equations if we write $$a(\tau)=\int_\tau^\infty\frac{dt}{{\lambda}}e^{-{\lambda}^{-1}(t-\tau)}\omega u(t),\qquad p^2(\tau)s(\tau)=\int_\tau^\infty\frac{2dt}{{\lambda}}e^{-2{\lambda}^{-1}(t-\tau)}\omega p^2(t)a(t),$$ and substitute these expressions to the identity with the function $b(\tau)$ taken from the second equation of the system . The solution of this integro-differential equation is specified by the only one arbitrary constant. If this solution or some unphysical solution to the system are known, we can integrate the LD equation. Indeed, in a general position the vector $\dot{{\upsilon}}_{\alpha}$ can be expressed as a linear combination of the vectors $\xi_{\alpha}$, ${\upsilon}_{\alpha}$ and ${\varepsilon}_{{\alpha}{\beta}{\gamma}}\xi^{\beta}{\upsilon}^{\gamma}$. The coefficients of this decomposition are certain functions of the invariants which are already known. Therefore, we need to integrate a system of linear equations with variable coefficients. So, $$\label{accelera} \dot{{\upsilon}}_{\alpha}=\frac{a}{u}{\varepsilon}_{{\alpha}{\beta}{\gamma}}\xi^{\beta}{\upsilon}^{\gamma}+\frac{bp}{u}(\xi_{\alpha}-p{\upsilon}_{\alpha}).$$ Because of the orthogonality condition, only two equations are independent. Now we make a substitution to Eq. of the form , . In the case $\xi^{\alpha}=(1,0,0)$, we have $$\label{subs h} {\upsilon}_0=p,\qquad{\upsilon}_1=\sqrt{u}\cos{\varphi},\qquad{\upsilon}_2=\sqrt{u}\sin{\varphi},$$ and $$\label{freq_H} u\dot{{\varphi}}=-a,$$ whence the momenta ${\upsilon}_{\alpha}$ are found in quadratures. If $\xi^{\alpha}=(0,0,1)$ then we substitute $$\label{subs e} {\upsilon}_0=\sqrt{u}\cosh\psi,\qquad{\upsilon}_1=\sqrt{u}\sinh\psi,\qquad{\upsilon}_2=p.$$ This results in $$\label{freq_E} u\dot{\psi}=a.$$ In the third case $\xi^{\alpha}=(1,0,1)$, we do $$\label{isotropic} {\upsilon}_1=rp,\qquad{\upsilon}_-=p,\qquad p{\upsilon}_+=1+ur^2,$$ and arrive at $$\label{freq_HE} u\dot{r}=a.$$ The case of an arbitrary vector $\xi^{\alpha}$ is reduced to the considered ones by a proper Lorentz transform of the tetrad indices. Notice that the equations of motion , , and are valid for a non-constant external field parameter $\omega$. In accordance with our general considerations, physical solutions are specified by four constants – two constants specify the initial position on the plane, one determines $u(\tau)$ and another one is needed to pick out the unique solution from Eqs. , or . Thus, we have to find the evolution of invariants described by Eqs. . In case of a constant $\omega$, the autonomous system is equivalent to the one second order differential equation on the function $a(u)$ or its inverse $u(a)$ $$\begin{split} a''&=-\frac{[2au+\xi^2(a+\omega u)]a'}{2u(a-\omega u)(\xi^2+u)}-\frac{2{\lambda}^2 a^2(\xi^2+u)a'^3}{u(a-\omega u)^2},\\ \ddot{u}&=\frac{[2au+\xi^2(a+\omega u)]\dot{u}^2}{2u(a-\omega u)(\xi^2+u)}+\frac{2{\lambda}^2 a^2(\xi^2+u)}{u(a-\omega u)^2}. \end{split}$$ Even in the simplest case $\xi^2=0$, when these equations can be cast into the form $$a''=-\frac{aa'}{u(a-u)}-\frac{2a^2a'^3}{(a-u)^2},\qquad\ddot{u}=\frac{a\dot{u}^2}{u(a-u)}+\frac{2a^2}{(a-u)^2},$$ we have not succeeded in finding a general solution. ### Asymptotics {#planmotasym} However, we can investigate the asymptotics of exact physical solutions to the LD equation at large times. It is easily done in the coordinates where the vector field of the system has no singularities. Making a change of variables $$a=u\bar{a},\qquad b=u\bar{b},$$ we come to $$\label{invar evol 1} \lambda\dot{\bar{a}}=\bar{a}[1+2{\lambda}\bar{b}(\xi^2+u)]-\omega,\qquad\lambda\dot{\bar{b}}=\bar{b}-{\lambda}[\bar{a}^2-\bar{b}^2(\xi^2+2u)],\qquad\dot{u}=-2\bar{b}u(\xi^2+u).$$ It is not difficult to find the stationary points of this system. We are interested only in physical solutions and physical stationary points. A physical stationary point as a particular case of a physical solution should be regular in ${\lambda}$. Again we have three cases. The case $\xi^2=1$ is a planar motion in a constant homogeneous magnetic field [@Plass; @Endres]. The system has two stationary points, one of them being physical $$\label{stationary_h} \bar{a}=\frac{\omega}{g},\qquad\bar{b}=\frac{g-1}{2{\lambda}},\qquad u=0,\qquad g:=2^{-1/2}\left(1+\sqrt{1+16{\lambda}^2\omega^2}\right)^{1/2}.$$ Linearizing the system in a vicinity of this point, we obtain the asymptotics of the exact solution to the LD equation $$\label{solution_H} \begin{split} {\delta}\bar{a}&=-u(0)Ae^{{\lambda}^{-1}(1-g)\tau}+e^{{\lambda}^{-1}g\tau}\left[(c_1+u(0)A)\cos\frac{2\omega\tau}{g}+(c_2+u(0)B)\sin\frac{2\omega\tau}{g}\right],\\ {\delta}\bar{b}&=-u(0)Be^{{\lambda}^{-1}(1-g)\tau}+e^{{\lambda}^{-1}g\tau}\left[(c_2+u(0)B)\cos\frac{2\omega\tau}{g}-(c_1+u(0)A)\sin\frac{2\omega\tau}{g}\right],\\ {\delta}u&=u(0)e^{{\lambda}^{-1}(1-g)\tau},\qquad A:=\frac{\omega(g-1)}{g(5g-4)},\qquad B:=\frac{3(g-1)^2}{2{\lambda}(5g-4)}. \end{split}$$ The terms in the brackets describe runaway solutions. They are unphysical and have to be set to zero by a proper choice of the initial conditions. Then the physical solution to Eqs. is solely specified by the initial condition on $u$. The first term in the first line in Eqs. describes a correction to the rotational speed of a charged particle due to the radiation reaction. Inasmuch as $u$ is non-negative, this correction has an opposite sign with respect to the main contribution, i.e., the rotational speed increases with time and tends exponentially to its limiting value . The limiting value is, of course, lesser than the cyclotron frequency. In case at hand, $u$ is related to the kinetic energy of the particle and so the expression for ${\delta}u $ in describes its decreasing. The case $\xi^2=-1$ corresponds to a planar motion in a constant homogeneous electric field. The system possesses two stationary points. Only one of these points is physical $$\label{stationary_e} \bar{a}=\omega,\qquad\bar{b}=\frac{g-1}{2{\lambda}},\qquad u=1,\qquad g:=\sqrt{1+4{\lambda}^2\omega^2}.$$ The linearized in a neighbourhood of this point LD equation has the solution $$\label{solution_E} \begin{split} {\delta}\bar{a}&=-\omega u(0)(1-g^{-1})e^{{\lambda}^{-1}(1-g)\tau}+e^{{\lambda}^{-1}\tau}\left[c_1+\omega u(0)(1-g^{-1})\right],\\ {\delta}\bar{b}&=-u(0)Be^{{\lambda}^{-1}(1-g)\tau}+\frac{2{\lambda}\omega}{g-1}e^{{\lambda}^{-1}\tau}\left[c_1+\omega u(0)(1-g^{-1})\right]+e^{{\lambda}^{-1}g\tau}\left[c_2-\frac{2c_1{\lambda}\omega}{g-1}+\frac{u(0)(g-1)}{{\lambda}(1-2g)}\right],\\ {\delta}u&=u(0)e^{{\lambda}^{-1}(1-g)\tau},\qquad B:=\frac{(g-1)^2(2g+1)}{2{\lambda}g(2g-1)}. \end{split}$$ Unphysical solutions are the terms in the square brackets. Demanding their vanishing, we uniquely determine the integration constants $c_1$ and $c_2$ through $u(0)$. The correction to the “frequency” $\dot{\psi}$ increases with time and tends exponentially to the limiting value . The expression for ${\delta}u(0)$ in describes an evolution of the square of the momentum component normal to the electric field. As expected, this component exponentially tends to zero and the solution passes into the hyperbolic motion . The case $\xi^2=0$ is more involved. To shorten formulas, we redefine the variables entering $$\label{replacement} \bar{a}\rightarrow \omega\bar{a},\qquad\bar{b}\rightarrow{\lambda}\omega^{2}\bar{b},\qquad u\rightarrow({\lambda}\omega)^{-2}u,\qquad\tau\rightarrow{\lambda}\tau,$$ and shall restore the original notation where it will be necessary to make estimations. After this redefinition, a regularity in ${\lambda}$, which distinguishes physical solutions, means a regularity of the solution in $\tau^{-1}$. Then the system has a single stationary point $$\label{crit_point} \bar{a}=1,\qquad\bar{b}=1,\qquad u=0.$$ This point is degenerate and, therefore, the solutions to the linearized system improperly describe a behaviour of solutions to the LD equation in a vicinity of this point. To obtain a correct asymptotics, we integrate the last equation in $$u=u(0)\left[1+2u(0)\tau+2u(0)\int_0^\tau dt{\delta}\bar{b}(t)\right]^{-1}.$$ The integrand of the third term in the square brackets tends to zero. Consequently, the second term in the square brackets will dominate at large times $$\label{assumptions} \tau\gg{\lambda},\qquad\tau\gg(2{\lambda}\omega^2u(0))^{-1},$$ and we can take $$u\approx\tau^{-1}/2.$$ Then the equations for the leading asymptotics read as $$\label{linearized_he} {\delta}\dot{\bar{a}}={\delta}\bar{a}+\tau^{-1}{\delta}\bar{b}+\tau^{-1},\qquad{\delta}\dot{\bar{b}}=-2{\delta}\bar{a}+{\delta}\bar{b}+\tau^{-1},$$ where we keep only the leading terms. The system possesses runaway solutions, which are nonregular in ${\lambda}$ ($\tau^{-1}$) and, consequently, unphysical. They can be removed by an appropriate choice of the initial data. As for the physical solutions, their leading asymptotics takes the form $$\label{asympt} \begin{split} \bar{a}&=1-\tau^{-1}-\tau^{-2}(3\ln\tau+2u(1)-6)+o(\tau^{-2})=1-\frac1{\tau}\left(\frac{e^{-5/3}\tau}{8u^3(1)}\right)^{3/\tau}+o(\tau^{-2}),\\ \bar{b}&=1-3\tau^{-1}-\tau^{-2}(9\ln\tau+6u(1)-23)+o(\tau^{-2})=1-\frac3{\tau}\left(\frac{e^{-20/9}\tau}{8u^3(1)}\right)^{3/\tau}+o(\tau^{-2}),\\ u&=\frac12\left[\tau^{-1}+\tau^{-2}(3\ln\tau+2u(1)-1)\right]+o(\tau^{-2})=\frac1{2\tau}\left(\frac{\tau}{8u^3(1)}\right)^{3/\tau}+o(\tau^{-2}). \end{split}$$ Here we also add a next to leading correction to the asymptotics, which can be derived from the initial non-linear system . The last equalities in show how the power of the proper time entering the asymptotics tends to its limiting value. The initial value $u(1)$ in these last formulas differs from $u(1)$ appearing in the second equalities in . These initial values are related in an evident manner. Substituting the asymptotics into Eqs. , and bearing in mind the replacement , we find $$\begin{split} {\upsilon}_1&=\left(\frac{\tau}{2{\lambda}}\right)^{1/2}\left(2{\upsilon}^2_1({\lambda})\frac{\tau}{{\lambda}}\right)^{{\lambda}/2\tau}+o\left(({\lambda}/\tau)^{1/2}\right),\\ {\upsilon}_0&={\lambda}\omega\left(\frac{\tau}{2{\lambda}}\right)^{3/2}\left(\frac{{\lambda}\omega^2\tau}{8{\upsilon}_0^2({\lambda})}\right)^{-{\lambda}/2\tau}+o\left((\tau/{\lambda})^{1/2}\right)=-{\upsilon}_2. \end{split}$$ As we see, the system goes to the universal regime. For example, the quantity $$\label{ratio} \frac{{\upsilon}_0}{{\upsilon}_1^3}\approx-\frac{{\upsilon}_2}{{\upsilon}_1^3}\approx{\lambda}\omega\left(\frac{\omega{\upsilon}_1^3({\lambda})\tau^2}{{\lambda}{\upsilon}_0({\lambda})}\right)^{-{\lambda}/\tau}$$ ceases to depend on the initial data and tends to ${\lambda}\omega$. This occurs on the proper time scales $$\tau\gg{\lambda}\left|\ln\frac{{\upsilon}_0({\lambda})}{{\lambda}\omega{\upsilon}_1^3({\lambda})}\right|.$$ This effect is essentially due to the radiation reaction. After the lapse of a certain time, the charged particles moving on the plane in the electromagnetic field with the invariants $I_1=I_2=0$ will have the same ratio of momenta of the form irrespective of their initial momenta. So, if we measure the momenta components of these charged particles, the measured data will lie on the cubic parabola determined by Eq. . For comparison, we give here the well-known solution to the Lorentz equation in this electromagnetic field $$\begin{gathered} {\upsilon}_1={\upsilon}_1(0)+\sqrt{u(0)}\omega\tau,\qquad{\upsilon}_0={\upsilon}_0(0)+{\upsilon}_1(0)\omega\tau+\sqrt{u(0)}\frac{\omega^2\tau^2}2,\\ {\upsilon}_2={\upsilon}_2(0)-{\upsilon}_1(0)\omega\tau-\sqrt{u(0)}\frac{\omega^2\tau^2}2. \end{gathered}$$ In this case, the quantity analogous to , which tends to a constant at large times, is the ratio $$\frac{{\upsilon}_0}{{\upsilon}_1^2}\approx-\frac{{\upsilon}_2}{{\upsilon}_1^2}\approx\frac1{2\sqrt{u(0)}}.$$ Its limiting value depends on the initial conditions. The ratio goes to zero as $\tau^{-1}$ with asymptotics depending on the initial data. ### Landau-Lifshitz equation {#planmotLL} Now we investigate the planar motion of a charged particle in a constant homogeneous electromagnetic field in the framework of the so-called Landau-Lifshitz equation [@LandLifsh]. This is an approximate equation describing the physical solutions to the LD equation. It is obtained from the LD equation by the reduction of order procedure, while ${\lambda}$ being assumed to be a small parameter. Thus, let us seek for solutions to the LD equation in a class of functions which $$\left|\frac{d^km_{\alpha}}{d\tau^k}\right|=O(1),\qquad\left|\frac{d^k{\upsilon}_{\alpha}}{d\tau^k}\right|=O(1),\qquad k=\overline{0,\infty},$$ with respect to the small parameter ${\lambda}$. Also, we restrict ourself by the first correction in ${\lambda}$ to the Lorentz equation. Differentiating the first equation in with respect to $\tau$, we find $\dot{m}_{\alpha}$. Then we substitute it to the initial equation and come to $$m_{\alpha}=-\omega(\xi_{\alpha}-p{\upsilon}_{\alpha}-{\lambda}\dot{p}{\upsilon}_{\alpha}-{\lambda}p\dot{{\upsilon}}_{\alpha}).$$ By the use of this relation, the second equation in describing the evolution of the momentum ${\upsilon}_{\alpha}$ has the form $$\dot{{\upsilon}}_{\alpha}=\omega{\varepsilon}_{{\alpha}{\beta}{\gamma}}\xi^{\beta}{\upsilon}^{\gamma}+{\lambda}\omega^2p(\xi_{\alpha}-p{\upsilon}_{\alpha}),$$ where we neglect the higher orders in ${\lambda}$. Convolving this equation with $\xi_{\alpha}$, we arrive at $$\dot{u}=-2{\lambda}\omega^2u(\xi^2+u),$$ whence $$u=\frac{\xi^2u(0)}{(\xi^2+u(0))e^{2{\lambda}\omega^2\xi^2\tau}-u(0)},\quad\xi^2=\pm1;\qquad u=\frac{u(0)}{1+2{\lambda}\omega^2u(0)\tau},\quad\xi^2=0.$$ If $\xi^{\alpha}=(1,0,0)$, the substitution gives $$\dot{{\varphi}}=-\omega.$$ If $\xi^{\alpha}=(0,0,1)$, the substitution results in $$\dot{\psi}=\omega.$$ These formulas are in agreement with Eqs. and up to the first order in ${\lambda}$. In the case $\xi^{\alpha}=(1,0,1)$, we have $$\begin{gathered} {\upsilon}_1=\frac{{\upsilon}_1(0)+\sqrt{u(0)}\omega\tau}{\sqrt{1+2{\lambda}\omega^2u(0)\tau}},\qquad{\upsilon}_0+{\upsilon}_2=\frac{\sqrt{u(0)}}{\sqrt{1+2{\lambda}\omega^2u(0)\tau}},\\ {\upsilon}_0-{\upsilon}_2=\frac{{\upsilon}_0(0)-{\upsilon}_2(0)+2\omega({\upsilon}_1(0)+{\lambda}\omega\sqrt{u(0)})\tau+\sqrt{u(0)}\omega^2\tau^2}{\sqrt{1+2{\lambda}\omega^2u(0)\tau}}. \end{gathered}$$ We see that the system passes to the universal regime at sufficiently large proper times $${\upsilon}_1\approx\frac{\tau^{1/2}}{\sqrt{2{\lambda}}},\qquad{\upsilon}_0+{\upsilon}_2\approx \omega^{-1}\frac{\tau^{-1/2}}{\sqrt{2{\lambda}}},\qquad{\upsilon}_0-{\upsilon}_2\approx \omega\frac{\tau^{3/2}}{\sqrt{2{\lambda}}}.$$ The limiting value of the ratio is the same for these solutions as for the physical solutions to the exact LD equation. Thus, the Landau-Lifshitz equation correctly reproduces the asymptotics of the physical solution at large $\tau$ in spite of the fact that this asymptotics is not regular in ${\lambda}$. Discussion {#discuss} ========== In this paper, we have investigated a planar motion of charged particles obeying the LD equation. We gave a detailed study of asymptotics of the planar physical solutions to the LD equation in a constant homogeneous electromagnetic field. One of the main results of our study is an interesting asymptotics of the physical solution to the LD equation in this electromagnetic field with vanishing invariants $I_1=I_2=0$. According to the classical radiation reaction theory, the charged particles moving on the plane in such the electromagnetic field must have the same ratio of the momenta components at sufficiently large times. The existence of this asymptotics can be verified in a purely quantum electrodynamical context. Namely, the Dirac equation in that electromagnetic field can be exactly solved [@BagGit]. Then, by the use of this complete set of solutions, we construct quantum electrodynamics on the given background [@GFSh] and define the $S$-matrix. The operators of covariant momenta $\mathcal{P}_0=m{\upsilon}_0$ and $\mathcal{P}_1=m{\upsilon}_1$ commute. Therefore, we can construct a complete set of their eigenfunctions and evaluate the transition amplitudes to these states. For the one-photon radiation amplitude, the transition probability reads as $$\label{trans_prob} \sum_{k,{\lambda}} {\langle}{\beta}|\hat{U}^\dag|\mathcal{P}_0,\mathcal{P}_1;k,{\lambda}{\rangle}{\langle}\mathcal{P}_0,\mathcal{P}_1;k,{\lambda}|\hat{U}|{\beta}{\rangle},$$ where $\hat{U}$ is the evolution operator over an infinite time, ${\beta}$ are the quantum numbers characterizing the initial state of the electron, and $k$ and ${\lambda}$ are the momentum and polarization of the radiated photon. Provided the classical radiation reaction theory is viable, there must exist a range of the quantum numbers ${\beta}$ and strengths of the electromagnetic field such that the transition probability is mostly concentrated on the cubic parabola $$\mathcal{P}_0=\frac{2e^2 E}{3m^4}\mathcal{P}_1^3.$$ Of course, using formula , we disregard the multiple photon production and the production of the electron-position pairs, but these contributions are proportional to the fine structure constant and neglible for reasonable strengths of the electromagnetic field. We postpone a detail study of this effect in quantum electrodynamics for a future research. Also note that, as it was pointed out in [@Bhabha], the notion of a physical solution seems to be quite general and can be exploited to give a proper interpretation of the higher-order derivative corrections to effective actions in quantum field theory (see, e.g., [@JaLloMol; @Woodart; @Moroz] and also their quantization in [@LyakhovichPhD; @Plyushch1]). Usually, these terms arise from the heat kernel expansion over the regularization parameter ${\Lambda}$ or large mass $m$ of the one-loop correction [@VasilHeatKer] whereas the higher terms of this expansion are casted out. By analogy with the considerations presented in Appendix, we should demand a regularity of solutions to the effective equations of motion in the small expansion parameter – the inverse regularization parameter or inverse large mass. Then a neglect of the nonlocal reminder of the effective action is justified. In many instances, this regularity can be related to the regularity with respect to the coupling constants. Though, it does not mean that we assume a smallness of the couplings or these higher-order derivative terms. As far as the heat kernel expansion is concerned, one can distinguish the corrections of three types: i) the divergent higher derivative terms like, for example, in $R^2$ gravity; ii) the higher derivative terms disappearing in the regularization removal limit; iii) the higher derivative terms resulting from the large mass expansion of a finite part of the effective action. In these cases, the coefficients at the higher derivative terms have a form $$i)\;\bar{{\lambda}}_a^{-1}+e_a(\bar{{\lambda}})/f_a(m/{\Lambda}),\qquad ii)\;e_b(\bar{{\lambda}})f_b(m/{\Lambda}),\qquad iii)\;e_c(\bar{{\lambda}})f_c(R/m^2),$$ where $e$ and $f$ are some functions going to zero at the origin, $R$ schematically denotes the field strengths or their derivatives. The coefficients in the first case are the renormalized inverse couplings ${\lambda}_a^{-1}$. Now it is easy to see that a regularity of the expression in the coupling constants ${\lambda}_a$ implies its regularity in ${\Lambda}^{-1}$. As for the large mass expansion, we additionally have to require a regularity of solutions to the effective equations of motion in $m^{-1}$. Just to demonstrate what we mean, consider the effective action with a higher derivative correction coming from the (self)interaction $$S[\phi]=\frac12\int dx\phi(-\Box-m^2+{\lambda}\Box^2)\phi,$$ where ${\lambda}$ is proportional to the coupling constants or, possibly, to the inverse large mass. The physical sector of this model is equivalent to $$S_{phys}[\phi]=\frac12\int dx\phi\left[-\Box-(\sqrt{1+4{\lambda}m^2}-1)/2{\lambda}\right]\phi.$$ As long as as the coupling constants are scalars with respect to a symmetry group of the model, its physical sector possesses this symmetry as well. An elimination of the unphysical sector in free models is a simple task, but, for full interacting models, such an explicit elimination becomes complicated and results in an appearance of the nonlocal projectors to the physical states in the effective action. Regularity condition in Maxwell electrodynamics =============================================== In this appendix, we show that the solutions to the coupled system of Maxwell-Lorentz equations are regular in the coupling constants before taking a regularization removal limit. Consider a model with the action functional . The effective equations of motion of a charged particle with a bare mass $\bar{m}$ look like $$\label{lde_int_gen} \bar{m}\ddot{x}_\mu(\tau)=eF_{\mu\nu}(x(\tau))\dot{x}^\nu(\tau)+4e^2\int d\tau'\theta(X^0(\tau,\tau')){\delta}'(X^2(\tau,\tau')-{\varepsilon})X_{[\mu}(\tau,\tau')\dot{x}_{\nu]}(\tau')\dot{x}^\nu(\tau),$$ where $X_\mu(\tau,\tau'):=x_\mu(\tau)-x_{\mu}(\tau')$, and we have used the regularization of the retarded Green function of the form $$\label{green_func_reg_0} G^{-}(x)=\frac{\theta(x^0)}{2\pi}{\delta}(x^2)\;\rightarrow\;G^{-}_{\varepsilon}(x)=\frac{\theta(x^0)}{2\pi}{\delta}(x^2-{\varepsilon}),$$ where ${\varepsilon}$ is a regularization parameter. We shall assume that $F_{\mu\nu}(x)$ is smooth and bounded on the spacetime, $F_{\mu\nu}(x(\tau))$ vanishes at $\tau<\tau_0$ for some $\tau_0$, and the solution to Eq. has the following asymptotics in the past $$\label{asympt_past} x_\mu(\tau)=(\tau-\tau_0){\upsilon}_\mu+\bar{x}_\mu(\tau),\qquad{\upsilon}^2=1,$$ where $\bar{x}_\mu(\tau)$ is zero at $\tau<\tau_0$, and ${\upsilon}_\mu$ is a constant $4$-vector. The unphysical solutions may appear on the scales of a classical electron radius. Therefore, we are looking for the solution of the form $$\label{subs_irregul} x_\mu(\tau)=\bar{r}_e y_\mu(\tau/\bar{r}_e),\qquad \bar{r}_e:=e^2/\bar{m}.$$ The function $y_\mu(s)$ also depends on other dimensionless combinations of parameters entering Eq. , but we do not write them explicitly. Upon substitution , we arrive at $$\label{lde_int_gen_y} \ddot{y}_\mu(s)=\frac{e}{\bar{m}}\bar{r}_eF_{\mu\nu}(y(s))\dot{y}^\nu(s)+4\int ds'\theta(Y^0(s,s')){\delta}'(Y^2(s,s')-\bar{{\varepsilon}})Y_{[\mu}(s,s')\dot{y}_{\nu]}(s')\dot{y}^\nu(s),$$ where $\bar{{\varepsilon}}:={\varepsilon}/\bar{r}_e^2$ and other evident redefinitions have been done. The integral in this equation can be taken $$\label{I} I=\left[\frac{Y_{[\mu}(s,s')\ddot{y}_{\nu]}(s,s')}{(\dot{y}_\rho(s')Y^\rho(s,s'))^2}+\frac{Y_{[\mu}(s,s')\dot{y}_{\nu]}(s')}{(\dot{y}_\rho(s')Y^\rho(s,s'))^3}(1-\ddot{y}_\rho(s')Y^\rho(s,s'))\right]\dot{y}^\nu(s),$$ where $s'$ is determined by the conditions $Y^2(s,s')=\bar{{\varepsilon}}$, $Y^0(s,s')<0$. The LD equation is obtained by expanding Eqs. and in the asymptotic series in ${\varepsilon}$ around zero and discarding the terms vanishing at ${\varepsilon}\rightarrow0$. In view of the asymptotic behaviour of the solution , the asymptotics of the integral $I$ at small $\bar{r}_e$ readily follows $$\label{I1} I\underset{\bar{r}_e\rightarrow0}{\rightarrow}{\varepsilon}^{-3/2}\bar{r}_e^3\bar{y}_{[\mu}(s){\upsilon}_{\nu]}\dot{y}^\nu(s).$$ Hence, in this limit, the integro-differential equation reduces the Lorentz-type differential equation with the effective strength tensor: $F_{\mu\nu}$ plus the correction . According to the general theorems of ordinary differential equations theory, the solutions to this equation are regular in $\bar{r}_e$. The effective strength tensor tends to zero when $\bar{r}_e\rightarrow0$, and, in this limit, the particle moves along a straight line $$\ddot{y}_\mu(s)=0.$$ Restituting the notation , we see that the limiting trajectory of the particle is indeed regular in $\bar{r}_e$. If we compel the Lorentz force to be constant at $\bar{r}_e\rightarrow0$ increasing the strength of the electromagnetic field, the equation of motion turns into the ordinary Lorentz equation and its solutions cease to depend on $\bar{r}_e$. Another possibility is not to scale $x_\mu$ with $\bar{r}_e$ as in , but consider Eq. with the solutions of the form $x^\mu(\tau/\bar{r}_e)$. Drawing the same reasonings, it is easy to see that these solutions are regular in $\bar{r}_e$ as well. As long as $$r_e=\frac{\bar{r}_e}{1+b\bar{r}_e/{\varepsilon}^{1/2}}=\frac{b^{-1}{\varepsilon}^{1/2}}{1+b^{-1}{\varepsilon}^{1/2}/\bar{r}_e},$$ where $b$ is some constant, the regularity of a solution in $\bar{r}_e$ implies its regularity in the renormalized classical electron radius $r_e$ and the regularization parameter ${\varepsilon}^{1/2}$. If we used another regularization of the Green function $$\label{green_func_reg} G^{-}_{\varepsilon}(x)=\frac{\theta(x^0)}{2\pi{\varepsilon}}\theta(x^2)g(x^2/{\varepsilon}),\qquad\int_0^\infty dxg(x)=1,$$ then the asymptotics of the integral $I$ would become $$I\underset{\bar{r}_e\rightarrow0}{\rightarrow}4{\varepsilon}^{-3/2}\bar{r}_e^2\bar{y}_{[\mu}(s){\upsilon}_{\nu]}\dot{y}^\nu(s)\int_0^\infty dtg'(t^2).$$ The above arguments applied to this case reveal that there exists a regular limit $r_e\rightarrow0$ of solutions to the effective equations of motion of a charged particle with the regularization as well. A regularity of the electromagnetic field generated by this charged particle is also obvious. The regularization of the Green function is equivalent to the regularization of the current $$\label{current_reg} j^\mu(x)\;\rightarrow\;j^\mu_{\varepsilon}(x)=\Box_x\int dyG^+_{\varepsilon}(x-y)j^\mu(y),$$ where $j_\mu(x)$ is the current of a point charge. The regularization $G^+_{\varepsilon}(x)$ of the advanced Green function is analogous to the retarded one . The effective equations of motion of a charge with the regularized current have the form , but with the “effective” Green function (see for details, e.g., [@siss]) $$G^{eff}_{\varepsilon}(x-y):=\int dzdz'\Box_x G^-_{\varepsilon}(x-z)G^-(z-z')\Box_{z'}G^+_{\varepsilon}(z'-y).$$ It is not difficult to show that the effective regularized Green function is zero in the past light cone. A Poincare-invariance of this Green function implies $$\label{green_func_eff} G^{eff}_{\varepsilon}(x)=\frac{\theta(x^0)}{2\pi{\varepsilon}}\theta(x^2)\tilde{g}(x^2/{\varepsilon})+{\varepsilon}^{-1}\theta(-x^2)h(-x^2/{\varepsilon}),$$ where $\tilde{g}(x)$ is a generalized function satisfying the condition , while $$\int_0^\infty dxh(x)=0.$$ The part of the effective Green function with the support lying outside of the light cone does not contribute to the integral in Eq. . Thus, we revert to the case of the Green function regularization considered above. Notice also that we can use the retarded Green function of the form in the regularized current instead of the advanced Green function. It can be proven that the effective Green function takes the form in this case too. When we pass from Eq. to the LD equation , we turn to a “truncated” description of a charge. Because of the truncation, spurious solutions arise which are non-regular in the coupling constants. They should be excluded, since only the regular solutions, which we call physical, to the LD equation are close to the solutions of at small the regularization parameter ${\varepsilon}$. Indeed, the LD equation is derived from breaking off the series in ${\varepsilon}^{1/2}$. A regular in $r_e$ solution is regular in ${\varepsilon}^{1/2}$. If we substitute such a solution of the LD equation to Eq. expanded in the series in ${\varepsilon}^{1/2}$, we shall ascertain that the obtained expression tends to zero with ${\varepsilon}\rightarrow0$ for any $\tau$. On the other hand, if we substitute a non-regular solution of the LD equation to Eq. then the limit ${\varepsilon}\rightarrow0$ of the obtained expression may not exist. Making a general non-regular in $r_e$ ansatz to the LD equation , it is not difficult to see that the non-regular solutions to LD equation have to possess an essentially singular point at $r_e=0$ provided some miraculous cancelations do not occur. Since it is the LD force which is responsible for the singularity, it should dominate over the Lorentz force at small $r_e$ and certain proper times $\tau$. Therefore, the non-regular solution to the LD equation tends to the non-regular solution to the free LD equation at $r_e\rightarrow0$. It can be directly verified from stretching the proper time $\tau\rightarrow r_e\tau$. The solution does own the essentially singular point at $r_e=0$. But if we substitute this solution to Eq. , it blows up at ${\varepsilon}\rightarrow0$ (the regularization parameter also enters the renormalized mass). So, the non-regular solutions cannot be regarded as approximate solutions to at small the regularization parameter ${\varepsilon}$. We appreciate Prof. V.G. Bagrov for stimulating discussions on the subject. The work is supported by the Russian Ministry of Education and Science, contract No 02.740.11.0238, the FTP “Research and Pedagogical Cadre for Innovative Russia”, contracts No P1337, P2596, and the RFBR grant 09-02-00723-a. [999]{} P.A.M. Dirac, *Classical theory of radiating electrons*, Proc. Roy. Soc. London A **167**, 148 (1938). C. Teitelboim, *Splitting of the Maxwell tensor: Radiation reaction without advanced fields*, Phys. Rev. D **1**, 1572 (1970); *Splitting of the Maxwell tensor. II. Sources*, Phys. Rev. D **3**, 297 (1971); *Radiation reaction as a retarded self-interaction*, Phys. Rev. D **4**, 345 (1971); P.E.G. Rowe, *Structure of the energy tensor in the classical electrodynamics of point particles*, Phys. Rev. D **18**, 3639 (1978); K. Lechner, P.A. Marchetti, *Variational principle and energy-momentum tensor for relativistic electrodynamics of point charges*, Ann. Phys. (NY) **322**, 1162 (2007), hep-th/0602224; K. Lechner, *Radiation reaction and 4-momentum conservation for point-like dyons*, J. Phys. A **39**, 11647 (2006), hep-th/0606097. V. Krivitskii, V. Tsytovich, *Average radiation-reaction force in quantum electrodynamics*, Sov. Phys. Usp. **34**, 250 (1991); Ph.R. Johnson, B.L. Hu, *Stochastic theory of relativistic particles moving in a quantum field: I. Influence functional and Langevin equation*, quant-ph/0012137; *Stochastic theory of relativistic particles moving in a quantum field: II. Scalar Abraham-Lorentz-Dirac-Langevin equation, radiation reaction and vacuum fluctuations*, Phys.Rev. D **65**, 065015 (2002), quant-ph/0101001; *Worldline influence functional: Abraham-Lorentz-Dirac-Langevin equation from QED*, quant-ph/0012135. N.P. Klepikov, *Radiation damping forces and radiation from charged particles*, Sov. Phys. Usp. **28**, 506 (1985). C.S. Shen, *Radiation and acceleration of a relativistic charged particle in an electromagnetic field*, Phys. Rev. D **17**, 434 (1978). V.G. Bagrov, G.S. Bisnovatyi-Kogan, V.A. Bordovitsyn, A.V. Borisov, O.F. Dorofeev, V.Ya. Epp, V.S. Gushchina, and V.Ch. Zhukovsky, *Synchrotron Radiation Theory and Its Development* (World Scientific, Singapore, 1999). J. Frenkel, *Die Elektrodynamik des rotierenden Elektrons*, Z. Physik **37**, 243 (1926); H.J. Bhabha, H.C. Corben, *General classical theory of spinning particles in a Maxwell field*, Proc. Roy. Soc. London A **178**, 273 (1941); A.O. Barut, N. Unal, *Generalization of the Lorentz-Dirac equation to include spin*, Phys. Rev. A **40**, 5404 (1989); P.O. Kazinski, *Radition reaction of multipole moments*, Zh. Eksp. Teor. Fiz. **132**, 370 (2007) \[J. Exp. Theor. Phys. **105**, 327 (2007)\], hep-th/0604168. S.K. Wong, *Field and particle equations for the classical Yang-Mills field and particles with isotopic spin*, Nuovo Cimento **65**, 689 (1970); B.P. Kosyakov, *Radiation in electrodynamics and in Yang-Mills theory*, Sov. Phys. Usp. **35**, 135 (1992). B.S. DeWitt, R.W. Brehme, *Radiation damping in a gravitational field*, Ann. Phys. **9**, 220 (1960); J.M. Hobbs, *A vierbein formalism of radiation damping*, Ann. Phys. **47**, 141 (1968); D.V. Gal’tsov, P. Spirin, *Radiation reaction in curved even-dimensional spacetime*, Grav. Cosmol. **13**, 241 (2007), arXiv:1012.3085. B.P. Kosyakov, *Exact solutions of classical electrodynamics and the Yang-Mills-Wong theory in even-dimensional spacetime*, Teor. Mat. Fiz. **119**, 119 (1999) \[Theor. Math. Phys. **199**, 493 (1999)\], hep-th/0207217; P.O. Kazinski, S.L. Lyakhovich, and A.A. Sharapov, *Radiation reaction and renormalization in classical electrodynamics of a point particle in any dimension*, Phys. Rev. D **66**, 025017 (2002), hep-th/0201046. F. Rohrlich, *Classical theory of magnetic monopoles*, Phys. Rev. **150**, 1104 (1966); P. Chen, F.V. Hartemann, J.R. van Meter, and A.K. Kerman, *Radiative corrections in symmetrized classical electrodynamics*, Phys. Rev. E **62**, 8640 (2000). P.O. Kazinski, A.A. Sharapov, *Radiation reaction for a massless charged particle*, Class. Quant. Grav. **20**, 2715 (2003), hep-th/0212286. H.J. Bhabha, *On the expansibility of solutions in powers of interaction constant*, Phys. Rev. **70**, 759 (1946). N.D. Sen Gupta, *On the motion of a charged particle in a uniform electric field with radiation reaction*, Int. J. Theor. Phys. **4**, 179 (1971); *Synchrotron motion with radiation reaction*, Int. J. Theor. Phys. **4**, 389 (1971); *Comments on relativistic motion with radiation reaction*, Phys. Rev. D **5**, 1546 (1972); *A charged particle with radiation reaction*, Int. J. Theor. Phys. **8**, 301 (1973). G.N. Plass, *Classical electrodynamic equations of motion with radiative reaction*, Rev. Mod. Phys. **33**, 37 (1961). A.O. Barut, *Electrodynamics and Classical Theory of Fields and Particles* (Dover, New York, 1964). F. Rohrlich, *Classical Charged Particles* (Addison-Wesley, Reading, MA, 1965). A.A. Sokolov, I.M. Ternov, *Radiation from Relativistic Electrons* (American Institute of Physics, New York, 1986). F. Denef, J. Raeymaekers, U.M. Studer, and W. Troost, *Classical tunneling as a consequence of radiation reaction forces*, Phys. Rev. E **56**, 3624 (1997), hep-th/9906080. L.D. Landau, E.M. Lifshitz, *The Classical Theory of Fields* (Pergamon, Oxford, 1962). H.A. Lorentz, *Theory of Electrons* (B.G. Teubner, Leipzig, 1909). V.V. Golubev, *Lectures on the Analytic Theory of Differential Equations* (Gostekhizdat, Moscow-Leningrad, 1950) \[in Russian\]. G.H. Hardy, *Divergent Series* (Clarendon Press, Oxford, 1949). P.O. Kazinski, *Fluctuations as stochastic deformation*, Phys. Rev. E **77**, 041119 (2008), arXiv:0711.3644. A.T. Fomenko, B.A. Dubrovin, and S.P. Novikov, *Modern Geometry: Methods and Applications. Vol. 1: The Geometry of Surfaces, Transformation Groups, and Fields* (Springer, New York, 1984) M.S. Plyushchay, *Relativistic particle with torsion and charged particle in a constant electromagnetic field: Identity of evolution*, Mod. Phys. Lett. A **10**, 1463 (1995), hep-th/9309147. D.J. Endres, *The physical solution to the Lorentz-Dirac equation for planar motion in a constant magnetic field*, Nonlinearity **6**, 953 (1993). V.G. Bagrov, D.M. Gitman, *Exact Solutions of Relativistic Wave Equations* (Kluwer Acad. Pub., Dordrecht, 1990). E.S. Fradkin, D.M. Gitman, and Sh.M. Shvartsman, *Quantum Electrodynamics with Unstable Vacuum* (Springer, Berlin, 1991). X. Jaen, J. Llosa, and A. Molina, *A reduction of order two for infinite-order Lagrangians*, Phys Rev. D **34**, 2302 (1986). R.P. Woodard, *Avoiding dark energy with $1/R$ modifications of gravity*, Lect. Notes Phys. **720**, 403 (2007), astro-ph/0601672. A. Morozov, *Hamiltonian formalism in the presence of higher derivatives*, Theor. Math. Phys. **157**, 1542 (2008), arXiv:0712.0946. D.M. Gitman, S.L. Lyakhovich, and I.V. Tyutin, *Hamilton formulation of a theory with high derivatives*, Russ. Phys. J. **26**, 730 (1983); S.L. Lyakhovich, *The method of canonical quantization of theories with higher derivatives and its application in the $R^2$-gravity*, Ph.D. thesis, Tomsk State University, 1985 \[in Russian\]; I.L. Buchbinder, S.L. Lyahovich, *Canonical quantisation and local measure of $R^2$ gravity*, Class. Quantum Grav. **4**, 1487 (1987). M.S. Plyushchay, *Massive relativistic point particle with rigidity*, Int. J. Mod. Phys. A **4**, 3851 (1989); *Majorana equation and exotics: higher derivative models, anyons and noncommutative geometry*, Electron. J. Theor. Phys. **3**, 17 (2006), math-ph/0604022. D.V. Vassilevich, *Heat kernel expansion: user’s manual*, Phys. Rep. **388**, 279 (2003), hep-th/0306138. P.O. Kazinski, A.A. Sharapov, *Radiation back-reaction and renormalization in classical field theory with singular sources*, Teor. Mat. Fiz. **143**, 375 (2005) \[Theor. Math. Phys. **143**, 798 (2005)\].

--- abstract: 'Most of the research effort on image-based place recognition is designed for urban environments. In bucolic environments such as natural scenes with low texture and little semantic content, the main challenge is to handle the variations in visual appearance across time such as illumination, weather, vegetation state or viewpoints. The nature of the variations is different and this leads to a different approach to describing a bucolic scene. We introduce a global image descriptor computed from its semantic and topological information. It is built from the wavelet transforms of the image semantic edges. Matching two images is then equivalent to matching their semantic edge descriptors. We show that this method reaches state-of-the-art image retrieval performance on two multi-season environment-monitoring datasets: the CMU-Seasons and the Symphony Lake dataset. It also generalises to urban scenes on which it is on par with the current baselines NetVLAD and DELF.' author: - Assia Benbihi - Matthieu Geist - 'C[é]{}dric Pradalier' bibliography: - 'root.bib' title: ' Image-Based Place Recognition on Bucolic Environment Across Seasons From Semantic Edge Description' --- INTRODUCTION ============ Place recognition is the process by which a place that has been observed before can be identified when revisited. Image-based place recognition achieves this task using images taken with similar viewpoints at different times. This is particularly challenging for images captured in natural environments over multiple seasons (e.g. [@griffith2017symphony] or [@sattler2018benchmarking]) because their appearance is modified as a result of weather, sun position, vegetation state in addition to view-point and lighting, as usual in indoor or urban environments. In robotics, place recognition is used for the loop-closure stage of most large scale SLAM systems where its reliability is critical [@cummins2011appearance]. It is also an important part of any long-term monitoring system operating outdoor over many seasons [@griffith2017symphony; @churchill2013experience]. In practice, place recognition is usually cast as an image retrieval task where a query image is matched to the most similar image available in a database. The search is computed on a projection of the image content on a much lower-dimensional space. The challenge is then to compute a compact image encoding such that images of the same location are near to each other despite their change of appearance due to environmental changes. ![WASABI computes a global image descriptor for place recognition over bucolic environments across seasons. It builds upon the image semantics and its edge geometry that are robust to strong appearance variations caused by illumination and season changes. While existing methods are tailored for urban-like scenes, our approach generalises to bucolic scenes that offer distinct challenges.[]{data-label="fig:method"}](fig0.png){width="0.9\linewidth"} Most of the existing methods start with detecting and describing local features over the image before aggregating them into a low-dimensional vector. The methods differ on the local feature detection, description and aggregation. Most of the research efforts have focused on environments with rich semantics such as cities or touristic landmarks [@arandjelovic2016netvlad; @noh2017large]. Early methods relied on hand-crafted feature description (e.g. SIFT [@lowe2004distinctive]) and simple aggregation based on histograms constructed on a clustering of the feature space [@sivic2003video]. Recent breakthroughs use deep-learning to learn retrieval-specific detection [@noh2017large], description [@babenko2014neural] and aggregation [@arandjelovic2016netvlad]. Another line of work relies on the geometric distribution of the image semantic elements to characterise it [@gawel2018x]. However, all of these approaches assume that the images have rich semantics or strong textures and focus on urban environments. To the contrary, we are interested in scenes described by images depicting nature or structures with few semantic or textured elements. In the following, such environments, including lakeshores and parks, will be qualified as ‘bucolic’. In this paper, we show that an image descriptor based on the geometry of semantic edges is discriminative enough to reach image-retrieval performance on bucolic environments. The detection step consists in extracting semantic edges and sorting them by their label. Continuous edges are then described with the wavelet transform [@chuang1996wavelet] over a fixed-sized subsampling of the edge. This constitutes the local description step. The aggregation is a simple concatenation of the edge descriptors and their labels which, together, make the global image descriptor. Figure \[fig:method\] illustrates the image retrieval pipeline with our novel descriptor dubbed WASABI[^1]: A collection of images is recorded along a road during the Spring. Global image descriptors are computed and stored in a database. Later in the year, in Autumn, while we traverse the same road, we describe the image at the current location. Place recognition consists in retrieving the database image which descriptor is the nearest to the current one. To compute the image distance, we associate each edge from one image to the nearest edge with the same semantic label in the other image. The distance between two edges is the Euclidean distance between descriptors. The image distance is the sum of the distances between edge descriptors of associated edges. WASABI is compared to existing image retrieval methods on two outdoor bucolic datasets: the park slices of the CMU-Seasons[@sattler2018benchmarking] and Symphony[@griffith2017symphony], recorded over a period of 1 year and 3 years respectively. Experiments show that it outperforms existing methods even when the latter are finetuned for these datasets. It is also on par with NetVLAD, the current on urban scenes, which is specifically optimised for city environments. This shows that WASABI can generalise across environments. The contribution of this paper is a novel global image descriptor based on semantics edge geometry for image-based place recognition in bucolic environment. Experiments show that it is also suitable for urban settings. The descriptor as well as the evaluation code are available at <https://github.com/abenbihi/wasabi.git>. RELATED WORK ============ This section reviews the current state-of-the-art on place recognition. A common approach to place recognition is to perform global image descriptor matching. The main challenge is defining a compact yet discriminative representation of the image that also has to be robust to illumination, viewpoint and appearance variations. Early methods built such a global descriptor by aggregating locally invariant features such as SIFT [@lowe2004distinctive]. A first step consists in generating a codebook of visual words by clustering local feature descriptors over a training dataset. This dataset must be different from the place recognition one to generalise well. An image is then described with the statistics of its local features with respect to this codebook. In [@sivic2003video], the local features of the image are assigned to the codebook clusters and the descriptor is simply the clustering histogram. [@perronnin2010large] improves over the previous clustering by fitting a mixture of Gaussians over the training dataset local features. Then, for each local feature of the image of interest, they concatenate the gradient of the probability of this feature to belong to one of the gaussian. This high-dimensional vector is then reduced with Principal Component Analysis (PCA). This approach is simplified in [@JDSP10] by computing clusters as in [@sivic2003video] even though it does not use cluster-histogram for aggregation. Instead, they concatenate the distance vector between each local feature and its nearest cluster which is a specific case of the derivation in [@perronnin2010large]. All these methods rely on features based on pixel distribution that assumes that images have strong textures, which is not the case for bucolic image. They are also sensitive to variations in the image appearance such as seasonal changes. In contrast, we rely on the geometry of semantic elements and that proves to be robust to strong appearance changes. Recent works also aim at disentangling local features and pixel intensity through learned feature descriptions. [@babenko2014neural] uses pre-trained Convolutional Neural Network (CNN) feature maps as local descriptor and aggregates them using the previous methods. The current in place recognition, NetVLAD[@arandjelovic2016netvlad] also takes advantage of the rich representation space of CNN but specifically trains a CNN to generate local feature descriptors relevant for image retrieval. They transformed the hand-crafted aggregation from VLAD[@jegou2010aggregating] into an end-to-end learning pipeline and reached top performances on urban scenes such as the Pittsburg or the Tokyo time machine datasets [@torii2013visual; @torii201524]. DELF[@noh2017large] tackles the selection of local features and trains a network to sample only features relevant to the image retrieval through an attention mechanism on a landmark dataset. We also rely on CNNs to segment images but not to describe them. Instead, we use this high-level information to hand-craft a novel descriptor that relies on a combination of segmentation and image geometry. Segmentation is indeed robust to appearance changes but bucolic environments are typically not rich enough for the segmentation to be sufficient for place recognition. By combining it with information from the scene geometry we succeed in augmenting the discriminative power of the representation. Another recent approach leverages the image geometry. [@gawel2018x] converts images into a semantic graph, uses temporal information to fuse the graphs over time and generates a global database graph. Then, given a new image expressed as a semantic graph, image retrieval is reduced to a graph matching problem. However, this approach assumes again that the environment is rich in semantic elements to avoid ambiguous graphs. This is not the case in bucolic environments which leads us to leverage edges as another robust and discriminative image element. Edge-based image retrieval is not novel [@zhou2001edge] and the literature offers a wide range of edge descriptors [@merhy2017reconnaissance]. But these local descriptors are usually less robust to illumination and viewpoint variations than their pixel-based counterparts. In this work, we fuse edge description with semantic information to reach performance on bucolic image retrieval across seasons. We rely on the wavelet descriptor for its compact representation while offering uniqueness and invariance properties [@chuang1996wavelet]. METHOD ====== This section details the three steps of image retrieval: the detection and description of local features, their aggregation and the image distance computation. In this paper, a local feature is constructed as a vector that embeds the geometry of semantic edges. Local feature detection and extraction. --------------------------------------- The local feature detection stage takes a color image as input, and outputs a list of continuous edges together with their semantic labels. Two equivalent approaches can be considered. The first is to extract edges from the semantic segmentation of the image, i.e. its pixel-wise classification. The relies on CNN trained on labeled data [@larsson2019cross; @zhao2017pyramid]. The second approach is also based on CNN but directly outputs the edges together with their labels [@yu2017casenet; @acuna2019devil]. The first approach is favored as there are many more public segmentation models than semantic edges ones. Hence, starting from the semantic segmentation, a post-processing stage is necessary to reduce the labelling noise. Most of this noise consists in labeling errors around edges or small holes inside bigger semantic units. To reduce the influence of these errors, semantic blobs smaller than `min_blob_size` are merged with their nearest neighbours. Furthermore, to make semantic edges robust over long periods of time, it is necessary to ignore classes corresponding to dynamic objects such as cars or pedestrians. Otherwise, they would alter the semantic edges and modify the global image descriptor. These classes are removed from the segmentation maps and the resulting hole is filled with the nearest semantic labels. Taking the cleaned-up semantic segmentation as input, a simple Canny-based edge detection is performed and edges smaller than `min_edge_size` pixels are filtered out. Segmentation noise may also break continuous edges. So the remaining edges are processed so as to re-connect edges actually belonging with each other. For each class, if two edge extremities are below a pixel distance `min_neighbour_gap`, the corresponding edges are grouped together into a unique edge. The parameters are chosen empirically based on the segmentation noise of the images. We use the segmentation model from [@larsson2019cross]. It features a PSP-Net [@zhao2017pyramid] network trained on Cityscapes [@Cordts2016Cityscapes]. As a result, we fine-tuned it on CMU-Seasons. In this case, the relevant detection parameters were `min_blob_size=50`, `min_edge_size=50` and `min_neighbour_gap=5`. Local feature description ------------------------- Among the many existing edge descriptor, we favor the wavelet descriptor [@chuang1996wavelet] for its properties relevant to image retrieval. It consits in projecting a signal over a basis of known function and is often used to generate compact yet unique representation of a signal. Wavelet description is not the only transform to generates a unique representation for a signal. The Fourier descriptors [@zahn1972fourier; @granlund1972fourier] also provides such a unique embedding. However, the wavelet description is more compact than the Fourier one due to its multiple-scale decomposition. Empirically, we confirmed that the former was more discriminative than the latter for the same number of coefficients. ![Symphony. Semantic edge association across strong seasonal and weather variations.[]{data-label="fig:asso"}](./asso.PNG){width="0.9\linewidth"} Given a 2D contour extracted at the previous step, we subsample the edge at regular steps and collect $N$ pixels. Their $(x,y)$ locations in the image are concatenated into a 2D vector. We compute the discrete Haar-wavelet decomposition over each axis separately and concatenate the output that we L2 normalise. In the experiments, we set $N=64$ and keep only the even coefficients of the wavelet transforms. This does not destroy information as the coefficients are redundant. The final edge descriptor is a 128-dimension vector. Aggregation and Image distance ------------------------------ Aggregation is a simple accumulation of the edge descriptors together with their label. Given two images and using the aggregated edge descriptors, the image distance is the average distance between matching edges. More precisely, edges belonging to the same semantic class are associated between the images solving an assignment problem (see Fig. \[fig:asso\]). The distance used is the Euclidean distance between edge descriptors and the image distance is the average of the associated descriptor distances. In a retrieval setting, we compute such a distance between the query image and every image in the database and return the database entry with the lowest distance. ![Extended CMU-Seasons. Top: images. Down: segmentation instead of the semantic edge for better visualisation. Each column depicts one location from a `i` and a camera `j` that we note `i_cj`. Each lines depitcs the same location over several traversals noted `T`.[]{data-label="fig:cmu"}](./mosaic_cmu_img.png "fig:"){width="0.9\linewidth"} ![Extended CMU-Seasons. Top: images. Down: segmentation instead of the semantic edge for better visualisation. Each column depicts one location from a `i` and a camera `j` that we note `i_cj`. Each lines depitcs the same location over several traversals noted `T`.[]{data-label="fig:cmu"}](./mosaic_cmu_seg.png "fig:"){width="0.9\linewidth"} EXPERIMENTS =========== This section details the experimental setup and presents results for our approach against methods for which public code is available: BoW[@sivic2003video], VLAD[@jegou2011aggregating], NetVLAD[@arandjelovic2016netvlad], DELF[@noh2017large]. We demonstrate the retrieval performance on two outdoor bucolic datasets: CMU-Seasons[@sattler2018benchmarking] and Symphony[@griffith2017symphony], recorded over a period of 1 year and 3 years respectively. Although existing methods reach SoA performance on urban environment, our approach proves to outperform them on bucolic scenes, and so, even when they are finetuned. It also shows better generalisation as it achieves near SoA performance of the urban slices on the CMU-Seasons dataset. Datasets -------- #### Extended CMU-Seasons The Extended CMU-Seasons dataset (Fig. \[fig:cmu\]) is an extended version of the CMU-Seasons [@Badino2011] dataset. It depicts urban, suburban, and park scenes in the area of Pittsburgh, USA. Two front-facing cameras are mounted on a car pointing to the left/right of the vehicle at approximately 45 degrees. Eleven traversals are recorded over a period of 1 year and the images from the two cameras do not overlap. The traversals are divided into 24 spatially disjoint slices, with slices \[2-8\] for urban scenes, \[9-17\] for suburban and \[18-25\] for park scenes respectively. All retrieval methods are evaluated on the park scenes for which ground-truth poses are available \[22-25\]. The other park scenes \[18-21\] can be used to train learned approaches. For each slice in \[22-25\], one traversal is used as the image database and the 10 other traversals are the queries. In total, there are 78 image sets of roughly 200 images with ground-truth camera poses. Figure \[fig:cmu\] shows examples of matching images over multiple seasons with significant variations. #### Lake The Symphony [@griffith2017symphony] dataset consists of 121 visual traversals of the shore of Symphony Lake in Metz, France. The 1.3 km shore is surveyed using a pan-tilt-zoom (PTZ) camera and a 2D LiDAR mounted on an unmanned surface vehicle. The camera faces starboard as the boat moves along the shore while maintaining a constant distance. The boat is deployed on average every 10 days from Jan 6, 2014 to April 3, 2017. In comparison to the roadway datasets, it holds a wider range of illumination and seasonal variations and much less texture and semantic features, which challenges existing place recognition methods. ![Symphony dataset. Top-Down: images and their segmentation. First line: reference traversal at several locations. Each column `k` depicts one location `Pos.k`. Each line depicts `Pos.k` over random traversals noted `T`. Note that contrary to CMU-Seasons, we generate mixed-conditions evaluation traversals from the actual lake traversals. So there is no constant illumination or seasonal condition over one query traversal `T`.[]{data-label="fig:lake"}](./mosaic_lake_img.png "fig:"){width="0.9\linewidth"} ![Symphony dataset. Top-Down: images and their segmentation. First line: reference traversal at several locations. Each column `k` depicts one location `Pos.k`. Each line depicts `Pos.k` over random traversals noted `T`. Note that contrary to CMU-Seasons, we generate mixed-conditions evaluation traversals from the actual lake traversals. So there is no constant illumination or seasonal condition over one query traversal `T`.[]{data-label="fig:lake"}](./mosaic_lake_seg.png "fig:"){width="0.9\linewidth"} We generate 10 traversals over one side of the lake from the ground-truth poses computed with the recorded 2D laser scans [@pradalier2018multi]. The other side of the lake can be used for training. One of the 121 recorded traversal is used as the reference from which we sample images at regular locations to generate the database. For each database image, the matching images are sampled from 10 random traversals out of the 120 left. Note that contrary to the CMU-Seasons dataset, this means that there is no light and appearance continuity over one traversal (Fig. \[fig:lake\]). Experimental setup ------------------ This section describes the rationale behind the evaluation. On CMU-Seasons we evaluate place recognition methods the semantic elements of a traversal on one hand, and the lighting and seasonal conditions on the other hand. On the Symphony dataset, we assess their robustness to low texture images with few semantic elements and even harsher lighting and seasonal variations. The first CMU-Seasons evaluation, the semantic elements, consists in running independent place recognition over each slice and average the performance over the traversals. Since the slices are spatially disjoint, they hold different semantic elements that challenge the image retrieval in various ways. For example, slice 23 seen from camera 0 holds mostly repetitive patterns of trees that are harder to differentiate than the building skyline seen from camera 1 on slice 25. Averaging over the traversals is a way to put aside the influence of the lighting and season for each traversal. The second evaluation, the lighting and season, starts the same way with independent place recognition over each slice. But the scores are averaged over the slices for each traversal. This way, the semantic content is the same for all the traversals and only the lighting and season change. On Symphony, only the lighting and seasonal robustness are assessed as the semantic content is constant over the lake. As mentioned previously, our approach is evaluated against BoW, VLAD, NetVLAD and DELF. In their version available online, these methods are mostly tailored for rich semantic environments: the codebook for BoW and VLAD is trained on Flickr60k [@jegou2008hamming], NetVLAD is trained on the Pittsburg dataset [@torii2013visual] and DELF on the Google landmark one [@noh2017large]. For fair comparison, we finetune them on CMU-Seasons and Symphony when possible, and report both original scores and the finetuned ones noted with (\*). A new codebook generated for BOW an VLAD, using the CMU park slices 18-21. The NetVLAD training requires images with ground-truth poses, which is not the cases for these slices. So we train it on three slices from 22-25 and evaluate it on the remaining one. On Symphony, images together with their ground-truth poses are sampled from the west side of the lake that is spatially disjoint from the evaluation traversals. The DELF learned local features are not finetuned as the training code is not available even though the model is. Finally, our approach is tested against the original available methods on the three urban CMU-Seasons slices for which ground-truth poses are available \[6-8\]. This assesses whether our approach is also relevant for urban settings and hence better generalise across environments than methods tailored specifically for urban scenes. Metrics ------- The place recognition metrics are the [*recall$@\!N$*]{} and the [@philbin2007object]. Both depend on a distance threshold $\epsilon$: a retrieved database image matches the query if the distance between their camera center is below $\epsilon$. The [*recall$@\!N$*]{} is the percentage of queries for which there is at least one matching database image in the first $N$ retrieved images. We set $N \in \{1,5,10,20\}$, and $\epsilon$ to $5m$ and $2m$ for the CMU-Seasons and the Symphony datasets respectively. Both metrics are available in the code. Results ------- WASABI shows better performance on bucolic scenes than existing methods while only slightly underperforming the NetVLAD and DELF on urban environments. This is expected as the is optimised for such settings. Still, this shows that our method generalises to both types of environments. Finetuning existing methods to the bucolic scenes proves to be useful for VLAD but does not improve the overall performance for BoW and NetVLAD. A plausible explanation is that these methods require more data than the one available. Investigating the finetuning of these methods is out of the scope of this paper. The rest of this section details the results. #### Performance *w.r.t* semantic elements on CMU-Seasons ![CMU-Seasons. [*Recall$@\!N$*]{} for the 8 park slices with different semantic appearance. Our approach tends to outperform methods unless they are not enough edges (`23_c1`) or the slice holds many urban elements (`25_c1`).[]{data-label="fig:recN_avg_for_each_slice"}](./rec_avg_over_survey_per_slice.png){width="\linewidth"} Fig \[fig:recN\_avg\_for\_each\_slice\] plots the recognition [*recall$@\!N$*]{} averaged over several seasonal conditions for 8 locations. All methods show the same sensitivity to the retrieval error tolerance: the [*recall$@\!N$*]{} curves increase similarly as a function of $N$. On half the locations (`22_c0, 22_c1, 24_c0, 25_c0`), WASABI doubles the NetVLAD score. Elsewhere, it reaches similar scores except for `25_c1` for which it is slightly outperformed. This is expected as this slice holds urban elements for which DELF and NetVLAD are specifically trained. All recalls drop on slice `23_c1` that holds dense trees along the road. The images not only have few features usually leveraged by other methods, but also few semantic edges on which WASABI relies. This limit suggests that we exploit multiple levels of edges in future work and not only semantic edges. While finetuning NetVLAD shows no advantage over the original one on other slices, here it reaches the retrieval score. This suggests that retrieval could be learned on challenging bucolic environments. Still, the performance on the remaining slices shows that a simple finetuning may not be enough to transfer this method and additional research is necessary. These observations are confirmed with the results, not plotted here for the sake of page limits. #### Performance *w.r.t* light/season on CMU-Seasons ![CMU-Seasons. [*Recall$@\!N$*]{} for the 10 traversals with different light/season conditions averaged over various semantic scenes. Only 9 plots are displayed for the sake of visualisation.[]{data-label="fig:recN_avg_for_each_survey"}](./rec_avg_over_slice.png){width="\linewidth"} Fig \[fig:recN\_avg\_for\_each\_survey\] assesses the influence of various light/season conditions averaged over multiple semantic scenes. Overall, WASABI is as robust as SoA methods over light/seasons variations or slightly more robust. All approaches are relatively constant whereas our method displays a slight drop on survey 6 and 8. This can be explained by the presence of a strong sunglare (Fig. \[fig:sunglare\]) that makes the segmentation noisy. The resulting semantic edges are then less reliable even though they lead to similar retrieval performance as SoA. #### Global performance on Symphony Figure \[fig:lake\_recN\_global\]-left plots the *Recall@5* with respect to the image query averaged over all traversals. This shows whether retrieval is easier on a specific location of the shore. As expected, the score shows no bias on a location, which backs our claim that a semantic analysis on Symphony is pointless. [*Recall$@\!N$*]{} for other values of N and other methods support this claim. The right plot shows the [*Recall$@\!N$*]{} averaged over all the query images. WASABI presents a slight advantage over NetVLAD and DELF although one could have expected higher performance based on the previous solid results on CMU-Seasons. One explanation is the segmentation noise induced by the image noise in one hand (*e.g.* sunglare) and the lack of domain adaptation on the other hand (Fig. \[fig:sunglare\]). As there is no ground-truth segmentation for the Symphony dataset, finetuning the segmentation is currently not possible. However, the satisfying results on CMU-Seasons motivate future work to improve the Symphony segmentation as well as the robustness of the descriptor to failures of the segmentation stage. ![Segmentation failures. Left column: CMU-Seasons. A strong sunglare is present along survey 6 (sunny spring) or 8(snowy winter). Other columns: Symphony. The segmentation is not finetuned on the lake and produces a noisier output. It is also sensitive to sunglare.[]{data-label="fig:sunglare"}](./seg_fails.png){width="0.6\linewidth"} #### Generalisation to urban setting Figure \[fig:urban\_recN\_avg\_for\_each\_slice\] shows that WASABI compares with NetVLAD scores even though it is not specifically tailored for urban environments. It is interesting to note that on slice 7, WASABI slightly outperforms NetVLAD on camera 0 whereas we observe the opposite on camera 1. The former mostly shows grass and trees along a parking lot whereas the latter images a building. This observation supports the bias that existing methods have toward urban environments. ![CMU-Seasons. Urban slices \[6-8\]. [*Recall$@\!N$*]{} averaged over traversals for each slice for semantic analysis. WASABI exhibits strong generalisation as it compares to NetVLAD.[]{data-label="fig:urban_recN_avg_for_each_slice"}](./urban_rec_avg_over_survey_per_slice.png){width="\linewidth"} Figure \[fig:urabn\_recN\_avg\_for\_each\_survey\] reports the evolution of the retrieval performance over several traversal with various light/season conditions in urban scenes. The scores are consistent over the traversals which shows that all methods are robust to such variations. Their scores are even more consistent on these slices than on the park ones. This is expected as seasonal variations induce fewer alterations on urban environments than on bucolic ones. Additionally, there is more texture and distinctive structures to exploit for place recognition. Nevertheless, it is interesting to note that our method, although designed for bucolic environments, can adapt to semantically rich urban settings. ![CMU-Seasons. Urban slices \[6-8\]. [*Recall$@\!N$*]{} for the 10 traversals averaged over the various semantic locations. All methods are relatively constant across the traversals which exhibits their robustness to light/season variations. Only 9 plots are displayed for the sake of visualisation.[]{data-label="fig:urabn_recN_avg_for_each_survey"}](./urban_rec_avg_over_slice.png){width="\linewidth"} CONCLUSIONS =========== In this paper, we presented WASABI, a novel image descriptor for place recognition across seasons in bucolic environments. It represents the image content through the geometry of its semantic edges, taking advantage of their invariance seasons, weather and illumination. Experiments show that it outperforms existing image-retrieval approaches better suited for urban environments. Tuning these methods for bucolic datasets proves to be insufficient to reach the same performances as our approach. Conversely, WASABI generalises well to other settings and reach scores on par with on urban scenes. Current research now focuses on improving the segmentation and disentangling the image description from the noise segmentation. [^1]: WAvelet SemAntic edge descriptor for BucolIc environment

--- abstract: 'The Casimir force between two parallel thick plates, one perfectly dielectric, the other purely magnetic, has been calculated long ago by Boyer \[T. H. Boyer, Phys. Rev. A [**9**]{}, 2078 (1974)\]. Its most characteristic property is that it is repulsive. The problem is actually delicate and counterintuitive. In the present paper we analyze the problem by first considering the simple harmonic oscillator model introduced by us earlier \[J. S. H[ø]{}ye [*et al.*]{}, Phys. Rev. E [**67**]{}, 056116 (2003); Phys. Rev. A [**94**]{}, 032113 (2016)\]. Extension of this model shows how the repulsive behavior can be understood on a microscopic basis, due to the duality between canonical and mechanical momenta in presence of the electromagnetic vector potential. This duality corresponds to the TM and TE modes in electrodynamics. We analyze the generalized Boyer case where the permittivities and permeabilities of the parallel plates are arbitrary. In this respect we first find the induced interaction between a pair of particles with given electric and magnetic polarizabilities and then find it for a pair of parallel plates. The method used for our evaluations is the statistical mechanical one that we have introduced and applied earlier. Whether the pair of particles or plates attract or repel each other depends on their polarizabilities or permittivities and permeabilities respectively. For equal particles or equal plates there is always attraction.' author: - 'Johan S. Høye$^1$ and Iver Brevik$^2$' title: Repulsive Casimir force --- Introduction {#secintro} ============ The Casimir force between two equal parallel isotropic nonmagnetic materials separated by a vacuum (air) gap $a$ is under usual circumstances known to be attractive (for reviews on the Casimir effect. cf., for instance, Ref. [@bordag09; @milton01; @parsegian06]). It turns out to be possible, however, in some cases to make the Casimir force switch sign so that it becomes repulsive. Typically, this occurs if a special liquid like bromobenzene is immersed between gold and silica surfaces as shown by Munday [*et al.*]{} [@munday09]. Then Casimir repulsion takes place at distances above approximately 5 nm since the dielectric constant of bromobenzene lies between those of gold and silica for imaginary frequencies below a certain value [@bostrom12; @wennerstrom99; @ninham98]. The analog case of two equal purely magnetic plates behaves in the same way; the Casimir force is under usual circumstances attractive. The case of two unequal plates - one purely dielectric and the other purely magnetic - is however different in the sense that the Casimir force becomes [*repulsive*]{}. This peculiar effect has been known for several years. A well known reference in this area is the paper of Boyer [@boyer74] (the present problem is sometimes called the Boyer problem). Based upon earlier investigations of Feinberg and Sucher [@feinberg68] on the van der Waals force between two electrically and magnetically particles, he calculated the repulsive force between a conducting and a permeable plate in the limit $\varepsilon \rightarrow \infty, \mu \rightarrow \infty$. Boyer’s formalism was based upon so-called random electrodynamics. We ought to point here that this change of force direction is actually quite nontrivial. Consider for definiteness the left plate with boundary $z=0$ to be purely dielectric, with finite permittivity $\varepsilon >1$, while the right plate with boundary $z=a$ is purely magnetic with permeability $\mu > 1$. In the boundary region of the dielectric plate around $z=0$ the only electromagnetic volume force density is $-\frac{1}{2}\varepsilon_0 E^2\nabla \varepsilon$. Here the $E$ is the magnitude of the varying electric field from the outside to the inside of the dielectric boundary. This should be expected to give a surface force density acting in the positive $z$ direction, i.e. toward the vacuum region, in accordance with the general property of classical electrodynamics saying that the surface force acts in the direction of the optically thinner medium. In our case the force direction is however reversed. Taken literally, this has to be interpreted to imply that the square $E^2$ becomes negative, thus a counterintuitive result whose physical significance is not obvious. We mention that the problem has been considered in the literature from a more wide viewpoint also, not restricted to the idealized case of Boyer where one plate was electric and the other magnetic. Some papers dealing with these topics are Refs. [@richmond78; @kenneth02; @inui11; @pappakrishnan14; @sinha18]. Most of them, such as Refs. [@kenneth02; @inui11; @pappakrishnan14], analyze Casimir repulsive forces in magnetodielectric configurations, while the very recent Ref. [@sinha18] analyzes the repulsive force on a magnetic point particle near a surface. In general, whether the force is found to be repulsive or attractive, depends on the relative strengths of $\varepsilon_i$ and $\mu_1$ ($i=1,2$). The delicate point mentioned above - that a Casimir calculation leads to an effective square $E^2 <0$ while always $E^2 \ge 0$ for $E$ real - will not be investigated further in this work. In the present paper our purpose is instead to analyze the Boyer problem in a generalized sense, i.e. take the polarizabilities of the two particles or the permittivities and permeabilities to be general. We start from a microscopic harmonic oscillator model where two oscillators interact via a third one. We have employed this model before [@hoye03]. The model was later extended to cover the case where the third oscillator describes a set of oscillators [@hoye16]. The two oscillators of the model correspond to two polarizable particles that interact via the harmonic oscillator modes of the electromagnetic field. For the latter there are two types of waves that are clearly noticed when a pair of parallel polarizable plates are considered. They are the TM (transverse magnetic) and TE (transverse electric) waves or modes. The former is the one most easily understood in view of the induced attractive interaction. Also the induced interaction increases in magnitude with temperature to reach the classical limit. For the TE mode, however, the situation is less obvious as the force decreases to zero in the classical limit since its Matsubara frequency $\zeta=0$ does not contribute. To cope with this situation interaction terms similar to the interaction of a particle with the electromagnetic vector potential was used. To obtain the situation corresponding to a dielectric plate interacting with a magnetic one, we find that our model can be extended in a straightforward way to encompass such a situation. During computations two quantities $D_1$ and $D_2$ appear (defined in Eq. (\[18\]) below), where their product $D_1 D_2>0$ determines an induced attractive force. Now one notes that for the analog of the TM case both quantities are positive while for the TE case both are negative by which $D_1 D_2>0$ remains. We will show how the model can be extended such that $D_1$ and $D_2$ have opposite signs. This will be the analog of a pair of dielectric and magnetic plates that repel each other. In Sec. \[sec2\] we consider the harmonic oscillator model, showing how the transition from attractive to repulsive force between dipoles can be traced back to the fundamental distinction between mechanical momentum and canonical momentum in classical electrodynamics. In Sec. \[sec3\] we derive the induced interaction between a pair of particles where there are couplings between both electric and magnetic dipole moments. In Sec. \[sec4\] we generalize the statistical mechanical derivations of Ref. [@hoye98] to two dielectric half-planes with different dielectric constants. From this the well-known Lifshitz result is recovered. Finally, in Sec. \[sec5\] this is generalized further to obtain the Casimir force with magnetic properties included. Section VI summarizes the obtained results. Harmonic Oscillator model for repulsive Casimir force {#sec2} ===================================================== Now consider the harmonic oscillator model treated in Refs. [@hoye03] and [@hoye16]. We will here only sketch the previous derivations while more details can be found in the mentioned references. Let the three oscillators have eigenfrequencies $\omega_i$ ($i=1,2,3$). Their classical partition function is then proportional to the inverse of $\sqrt{Q}$ where $$Q=a_1 a_2 a_3, \quad a_i=\omega_i^2 \quad (i=1,2,3). \label{10}$$ For simplicity all three oscillators of the model are one-dimensional. By quantization using the path integral method [@hoye81; @brevik88] the classical system turns out to be split into a set of classical harmonic oscillator systems where the Matsubara frequencies are added to the eigenfrequencies. This replaces expression (\[10\]) by $$Q=A_1 A_2 A_3, \quad A_i=a_i+\zeta^2 = \omega_i^2+\zeta^2 \label{11}$$ where $\zeta=i\omega$ (the convention $\zeta=-i\omega$ may also be used). When the oscillators interact (via bilinear terms) the $Q$ can be expressed through a determinant. With interaction parameter $c$ for the interactions via oscillator 3, one obtains the result (9) of Ref. [@hoye16] (or (4.3a) of [@hoye03]) $$\begin{aligned} Q&=&\left| \begin{array}{ccc} A_1 & 0 & c\\ 0 & A_2 & c\\ c & c & A_3\\ \end{array} \right| =A_1 A_2 A_3 -c^2(A_1+A_2)=A_1 A_2 A_3(1-D_1-D_2) \nonumber\\ &=&A_1 A_2 A_3(1-D_1)(1-D_2)\left(1-\frac{D_1 D_2}{(1-D_1)(1-D_2)}\right), \label{12}\end{aligned}$$ $$D_j=\frac{c^2}{A_j A_3}\quad (j=1,2). \label{13}$$ Here the $A_j(1-D_j)$ terms represent each of the two oscillators and their radiation reaction with the third oscillator while the last factor corresponds to the Casimir free energy. With $D_j$ given by (\[13\]) this factor is larger than zero which means that the induced force between the two oscillators is attractive. This situation is the analog of the TM mode. The second situation is the analog of the TE mode. To analyze this we may start from the situation where oscillator 3 interacts with the momenta of oscillators 1 and 2. Then the interaction with the third oscillator is like the interaction via the electromagnetic vector field. With this field the Hamiltonian for a particle $j$ is $({\bf p}_j-(e/c){\bf A})^2/2m_j$, where here $c$ is the velocity of light, ${\bf p}_j$ the canonical momentum, and ${\bf A}={\bf A}({\bf r}_j)$ the vector potential (in Gaussian units). The mechanical momentum is ${\bf p}_{Mj}={\bf p}_j-(e/c){\bf A}$. In our model the analogous interaction can be written as $a_j(p_j-(c/a_j)x_3)^2$ ($j=1,2$), where again $c$ is the coupling parameter and the coordinate $x_3$ corresponds formally to the vector potential ${\bf A}$. Now for harmonic oscillators the roles of the momenta and coordinates can be exchanged to obtain the energy term $$a_j\left(x_j-\frac{c}{a_j}x_3\right)^2. \label{14}$$ With this the $A_3$ term in the determinant (\[12\]) is changed into $$A_3\rightarrow A_3+\frac{c^2}{a_1}+\frac{c^2}{a_2}. \label{15}$$ The rows $j=1,2$ of the determinant can be multiplied with $c/a_1$ and $c/a_2$ respectively to be subtracted from row 3. Then rows 1 and 2 can be multiplied with $\zeta/\sqrt{a_j}$, and columns 1 and 2 can be divided with the same factors. These operations do not change the determinant. We obtain $$\begin{aligned} Q&=&\left| \begin{array}{ccc} A_1 & 0 & c\\ 0 & A_2 & c\\ cq_1 & cq_2 & A_3\\ \end{array} \right| =\left| \begin{array}{ccc} A_1 & 0 & \zeta c/\sqrt{a_1}\\ 0 & A_2 & \zeta c/\sqrt{a_2}\\ -\zeta c/\sqrt{a_1} &-\zeta c/\sqrt{a_2} & A_3\\ \end{array} \right|, \label{16}\end{aligned}$$ $$q_j=1-\frac{A_j}{a_j}=-\frac{\zeta^2}{a_j}, \quad(j=1,2). \label{17}$$ With determinant (\[16\]) the $Q$ when evaluated can still be written as result (\[12\]) but now with $$D_j=-\frac{\zeta^2 c^2}{a_j A_j A_3}<0. \label{18}$$ Again the induced force will be attractive since $D_1 D_2>0$, but this time both factors are negative. However, in the classical limit where only $\zeta=0$ contributes this force will vanish as also is the case for the TE mode. From the above it is now easily seen that the induced force can be repulsive too. This will be the case if only one of the energies has form (\[14\]), say for $j=2$ but not for $j=1$. Then instead of (\[15\]) the change of $A_3$ will be $$A_3\rightarrow A_3+\frac{c^2}{a_2}. \label{19}$$ The determinant (\[16\]) will be modified to $$\begin{aligned} Q=\left| \begin{array}{ccc} A_1 & 0 & c\\ 0 & A_2 & \zeta c/\sqrt{a_2}\\ c &-\zeta c/\sqrt{a_2} & A_3\\ \end{array} \right|, \label{17}\end{aligned}$$ by which relation (\[18\]) changes into $$D_1=\frac{c^2}{A_1 A_3}>0 \quad \mbox{and} \quad D_2=-\frac{\zeta^2 c^2}{a_j A_j A_3}<0. \label{21}$$ With this $D_1 D_2<0$ by which the induced force will be repulsive. This force will be the analog of the repulsive force between a dielectric medium and a magnetic one. Induced interaction between a pair of particles with both dielectric and magnetic polarization {#sec3} ============================================================================================== In the static case electric and magnetic properties separate. However, for non-zero frequencies there is a coupling between electric and magnetic dipole moments. The fields are determined by Maxwell’s equations which in standard notation can be written (Gaussian units) $$\begin{aligned} \nonumber \nabla {\bf D}&=&0,\quad \nabla \times{\bf E}=-\frac{\partial{\bf B}}{c\,\partial t},\quad {\bf D}={\bf E}+4\pi{\bf P}=\varepsilon {\bf E},\\ \nabla {\bf B}&=&0,\quad \nabla \times{\bf H}=\frac{\partial{\bf D}}{c\,\partial t},\quad {\bf B}={\bf H}+4\pi{\bf M}=\mu {\bf H}. \label{30}\end{aligned}$$ The equations can be Fourier transformed by which $$\nabla\rightarrow-i{\bf k}\quad\mbox{and}\quad\frac{\partial}{c\,\partial t}\rightarrow i\frac{\omega}{c}=\zeta. \label{31}$$ (The signs of Fourier transforms depend upon the convention chosen.) Note, here and below definition (\[31\]) is used for $\zeta$. For a given polarization one finds the solution [@hoye82] $$\begin{aligned} \nonumber {\bf E}&=&-\frac{4\pi}{3}\left\{\frac{k^2}{k^2+\zeta^2}[3(\hat{\bf k}\cdot{\bf P})\hat{\bf k} -{\bf P}]+\left[\frac{2\zeta^2}{k^2+\zeta^2}+1\right]{\bf P}\right\}\\ &=&-\frac{4\pi}{3}\left\{\frac{k^2}{k^2+\zeta^2}3(\hat{\bf k}\cdot{\bf P})\hat{\bf k} +\frac{3\zeta^2}{k^2+\zeta^2}{\bf P}\right\} \label{32}\end{aligned}$$ where the hat is used for unit vectors. For simplicity the same vector notation is used both in $r$-space and $k$-space as the argument used will specify. In $r$-space (\[32\]) becomes [@hoye82] $${\bf E}={\bf E_1}=\frac{e^{-x}}{r^3}\left\{\left(1+x+\frac{1}{3}x^2\right)[3(\hat{\bf r}\cdot{\bf s}_1)\hat{\bf r}-{\bf s}_1]-\frac{2}{3}x^2{\bf s}_1\right\}-\frac{4\pi}{3}\delta({\bf r}){\bf s}_1 \label{33}$$ where $x=\zeta r$. Here the polarization used is a point dipole $${\bf P}={\bf P}({\bf r})={\bf s}_1\delta({\bf r}), \quad ({\bf P}={\bf P}({\bf k})={\bf s}_1). \label{34}$$ The interaction with a dipole moment ${\bf s}_2$ at position ${\bf r}$ is then $-{\bf E}_1{\bf s}_2$ which also is proportional to the correlation function for the two dipoles ${\bf s}_1$ and ${\bf s}_2$. Thus for (large) $r$ the induced free energy is proportional to $\langle({\bf E}_1{\bf s}_2)^2\rangle$ where the thermal average is over the fluctuating dipole moments for each value of $\zeta$. With $\langle({\bf a}\cdot\hat{\bf s})({\bf b}\cdot\hat{\bf s})\rangle={\bf a}\cdot{\bf b}/3$ one obtains $$\langle({\bf E}_1\cdot{\bf s}_2)^2\rangle=\frac{1}{3}\langle s_1^2\rangle\langle s_2^2\rangle\frac{1}{r^6}L_{EE}(x), \label{36}$$ $$L_{EE}(x)=e^{-2x}\left[2\left(1+x+\frac{1}{3}x^2\right)^2+\left(\frac{2}{3}x^2\right)^2\right]. \label{37}$$ Now the electric polarizability is given by ($i=1,2$) $$\alpha_{iE}=\alpha_{iE}(\zeta))\propto \langle s_i^2\rangle. \label{38}$$ With other details that follow from the derivations in Ref. [@brevik88], the induced free energy becomes Eq. (5.15) of the reference $$F_{EE}=-\frac{3}{2\beta r^6}\sum_{n=-\infty}^\infty L_{EE}(x)\alpha_{1E}\alpha_{2E} \label{39}$$ with $\beta=1/(k_B T)$ where $k_B$ is Boltzmann’s constant. (The factor 2 in the denominator is missing in the reference.) The sum is over the Matsubara frequencies $$x=\zeta r=\frac{Kr}{\hbar c},\quad K=\frac{2\pi n}{\beta}\quad\mbox{with $n$ integer}. \label{40}$$ At $T=0$ the sum can be replaced by an integral with $$dx=r\,d\zeta=\frac{2\pi r}{\hbar c\beta}\,dn\quad\mbox{or}\quad \sum_{n=-\infty}^\infty\rightarrow2\frac{\hbar c\beta}{2\pi r}\int_0^\infty dx. \label{41}$$ With constant (i.e. independent of $\zeta$) polarizabilities this further gives $$F_{EE}=-\frac{3\hbar c \alpha_{1E}\alpha_{2E}}{2\pi r^7}I_{EE} \label{42}$$ $$I_{EE}=\int\limits_0^\infty L_{EE}(x)\,dx=\frac{23}{6} \label{43}$$ which is the well established result (5.16) of Ref. [@brevik88]. Another situation is the induced interaction between the magnetic moments of the two particles. But due to the symmetry of Maxwell’s equations (\[30\]) this will be precisely like results (\[39\]) and (\[42\]) with the subscript $_E$ replaced by $_H$. What remains is the induced interaction between electric and magnetic moments. To do so we need to obtain the $L_{EH}$ function to replace $L_{EE}$ and replace an electric polarizability with a magnetic one. So consider the electric field created by a polarization. For non-zero frequencies it also creates a magnetic field. From Eqs. (\[30\]) and (\[31\]) one has $$-i({\bf k}\times{\bf E})=-i\frac{\omega}{c}{\bf B}=-\zeta {\bf H},\quad {\bf B}={\bf H}\,\,({\bf M}=0). \label{44}$$ With the last equality of Eq. (\[32\]) this simplifies to $${\bf H}=\frac{i}{\zeta}\left({\bf k}\times\left(\frac{-4\pi\zeta^2}{k^2+\zeta^2}{\bf P}\right)\right) \label{45}$$ (since $\hat{\bf k}\times\hat{\bf k}=0$). From this follows ($\nabla\times(\psi{\bf P})=(\nabla\psi)\times{\bf P}+\psi\nabla\times{\bf P}$) $${\bf H}=-\frac{1}{\zeta}\nabla\times\left(\frac{-\zeta^2e^{-\zeta r}}{r}{\bf P}\right)={\textcolor}{red}{-}\zeta(1+\zeta r)\frac{e^{-\zeta r}}{r^2}(\hat{\bf r}\times{\bf P}). \label{46}$$ So the interaction of an electric dipole moment ${\bf s}_1$ with a magnetic one ${\bf s}_2$ is $$-{\bf H}_1{\bf s}_2=-\zeta(1+\zeta r)\frac{e^{-\zeta r}}{r^2}(\hat{\bf r}\times{\bf s}_1)\cdot{\bf s}_2. \label{47}$$ The average of interest is the square of (\[47\]), but now with minus sign when compared with expression . This change of sign is due to a property of Fourier transforms. For two functions $f(x)$ and $g(x)$ and their Fourier transforms $\tilde f(k)$ and $\tilde g(k)$ one har the relation $$\int f(x) g(x)\,dx = \frac{1}{2\pi}\int \tilde f(k) \tilde g(-k)\,dk. \label{48a}$$ This relation is similar for situations when $x$ and/or $k$ have discrete values too. In the present case the $x$ corresponds to imaginary time with $\zeta$ the Fourier variable. In the free energy expression the sum over $\zeta$corresponds to the $k$ integration in relation . The transform changes sign when $\zeta\rightarrow-\zeta$ (as its $\zeta r=|\zeta|r$). In standard notation one can write $({\bf r}\times{\bf s}_1){\bf s}_2=\varepsilon_{ijk}x_i s_{1j}s_{2k}$ to obtain the average $$\langle[({\bf r}\times{\bf s}_1){\bf s}_2]^2\rangle=\varepsilon_{ijk}\varepsilon_{npq}x_i x_n\langle s_{1j}s_{1p}\rangle\langle s_{2k}s_{2q}\rangle=\frac{2}{9}r^2\langle s_1^2\rangle\langle s_2^2\rangle. \label{49}$$ Thus $$\langle({\bf H}_1\cdot{\bf s}_2)^2\rangle=\frac{1}{3}\langle s_1^2\rangle\langle s_2^2\rangle\frac{1}{r^6}L_{EH}(x) \label{50}$$ $$L_{EH}(x)=\frac{2}{3}e^{-2x}[x(1+x)]^2. \label{51}$$ Like Eq. (\[39\]), but now with opposite sign, the induced free energy expression becomes $$F_{EH}=\frac{3}{2\beta r^6}\sum_{n=-\infty}^\infty L_{EH}(x)\alpha_{1E}\alpha_{2H} \label{52}$$ plus the contribution $F_{HE}$ which will be the same except the product of polarizabilities that is modified into $\alpha_{1H}\alpha_{2E}$. With constant polarizabilities this similar to (\[42\]) gives $$F_{EH}=\frac{3\hbar c \alpha_{1E}\alpha_{2H}}{2\pi r^7}I_{EH} \label{53}$$ $$I_{EH}=\int\limits_0^\infty L_{EH}(x)\,dx=\frac{7}{6} \label{54}$$ This part of the free energy gives a repulsive force in agreement with earlier derivations [@boyer74]. Induced interaction between a pair of dielectric half-planes {#sec4} ============================================================ For a pair of half-planes one can generalize the statistical mechanical derivations of H[ø]{}ye and Brevik where two dielectric half-planes with the same dielectric constant were considered [@hoye98]. First we will reconsider this situation and modify it somewhat to include extension to half-planes with different dielectric constants. With this modification the method becomes more suitable for the extension to include different magnetic permeabilities too. Since equations of Ref. [@hoye98] will be used repeatedly, we will designate them with the numeral I when referred to. The half-planes with separation $a$ are parallel to the $xy$ plane. With this the dipole-dipole interaction can be Fourier transformed along the $x$ and $y$ directions to obtain Eq. (I6.4) for the radiating interaction in vacuum between electric dipoles ${\bf s}_1$ and ${\bf s}_2$ ($z\neq0$) $$\hat\phi(12)=-2\pi s_1 s_2\frac{e^{-q|z|}}{q} [({\bf h}\cdot\hat{\bf s}_1)({\bf h}\cdot\hat{\bf s}_2)-\zeta^2\hat{\bf s}_1\cdot\hat{\bf s}_2] \label{60}$$ $${\bf h}=\{ik_x, ik_y,\pm q\}, \quad q^2=k_\perp^2+\zeta^2, \quad k_\perp^2=k_x^2+k_y^2, \label{61}$$ Vectors with hats mean unit vectors. Without loss of generality the ${\bf k}$ can be directed along the $x$ axis by which ${\bf h}=\{ik_\perp, 0,\pm q\}$. With Eq. (I6.5) the interaction can be split in two parts $$\hat\phi(12)=-2\pi s_1 s_2\frac{e^{-q|z|}}{q}(H_1+H_2) \label{63}$$ $$H_1=({\bf h}\cdot\hat{\bf s}_1)({\bf h}\cdot\hat{\bf s}_2)-\zeta^2\hat{\bf s}_1^{||}\cdot\hat{\bf s}_2^{||}, \quad H_2=-\zeta^2\hat{\bf s}_1^\perp\cdot\hat{\bf s}_2^\perp \label{64}$$ where ${\bf s}^{||}$ is the component of ${\bf s}$ in the ${\bf h}$ plane while ${\bf s}^\perp$ is transverse to it. These two cases are the separation into the TM (transverse magnetic) and TE (transverse electric) modes respectively. For a medium with dielectric constant ${\varepsilon}$ and magnetic permeability $\mu$ the interaction (\[63\]) is modified into $$\hat\phi_{\varepsilon}(12)=-2\pi s_1 s_2\frac{e^{-q_{\varepsilon}|z|}}{q_{\varepsilon}}\left(\frac{1}{{\varepsilon}}H_{{\varepsilon}1}+\mu H_2\right) \label{66}$$ $$H_{\varepsilon1}=({\bf h}_\varepsilon\cdot\hat{\bf s}_1)({\bf h}_\varepsilon\cdot\hat{\bf s}_2)-\varepsilon\mu\zeta^2\hat{\bf s}_1^{||}\cdot\hat{\bf s}_2^{||} \label{65}$$ $${\bf h}_\varepsilon=\{ik_\perp, 0, \pm q_\varepsilon\}, \quad q_\varepsilon^2=k_\perp^2+\varepsilon\mu\zeta^2. \label{62}$$ while $H_2$ is unchanged. The interaction is due to the electric field from ${\bf s}_1$ that acts upon ${\bf s}_2$, or vice versa. Therefore it should be symmetric in the two dipole moments also when the dielectric constants are different. Thus $H_{{\varepsilon}1}$ must be modified. With use of (\[62\]) one finds Eq. (I6.10) which can be written as $$H_{{\varepsilon}1}={\bf G}_{{\varepsilon}1}\cdot\hat{\bf s}_2=g_{{\varepsilon}1} g_{{\varepsilon}2}, \quad {\bf G}_{{\varepsilon}1}= g_{{\varepsilon}1}\frac{q_{\varepsilon}}{k_\perp}{\bf u}_{{\varepsilon}\pm}, \quad {\bf u}_{{\varepsilon}\pm}=\left\{ik_\perp, 0, \pm\frac{k_\perp^2}{q_{\varepsilon}}\right\} \label{67a}$$ $$g_{{\varepsilon}i}=|{\bf h}_{\varepsilon}\times\hat{\bf s}_i|=\frac{q_{\varepsilon}}{k_\perp}\left(ik_\perp\hat s_{\perp i}\pm\frac{k_\perp^2}{q_{\varepsilon}}\hat s_{||i}\right)=\left(iq_{\varepsilon}\hat s_{\perp i}\pm k_\perp\hat s_{||i}\right), \quad (i=1,2) \label{67b}$$ Here the $\hat s_\perp$ and $\hat s_{||}$ are the components of $\hat {\bf s}^{||}$ along the direction of ${\bf k}_\perp$ (in the $xy$ plane) and the $z$ axis respectively. The advantage of this form is that it can be extended to half-planes with different dielectric constants with energy symmetric in ${\hat s}_1$ and ${\hat s}_2$. Thus Eq. (\[67a\]) can be modified to $$H_{{\varepsilon}1}={\bf G}_{{\varepsilon}1}\cdot\hat{\bf s}_2=g_{{{\varepsilon}_1}1} g_{{{\varepsilon}_2}2}, \quad {\bf G}_{{\varepsilon}1}= g_{{{\varepsilon}_1}1}\frac{q_{{\varepsilon}_2}}{k_\perp}{\bf u}_{{{\varepsilon}_2}\pm} \label{68}$$ For the interaction in vacuum (${\varepsilon}, \mu\rightarrow1$ one has similar expressions where $q_{\varepsilon}\rightarrow q$. (Note that with vector ${\bf u}_{{{\varepsilon}_2}\pm}$ the $\nabla{\bf D}=0$ of Eq. (\[30\]) is fulfilled.) The resulting Green function (electric interaction) for the two half-planes can now still be written in the form of Eq. (I6.20) $$\hat\phi_g(12)=J_{\varepsilon}\left(D_E^{||}H_{{\varepsilon}1}+D_E^\perp H_2\right), \quad J_{\varepsilon}=-2\pi s_1 s_2\frac{e^{-q_{{\varepsilon}_1}u_1-q_{{\varepsilon}_2}(u_2+a)}}{q_{{\varepsilon}_1}} \label{69}$$ where $u_1=-z_1>0$ and $u_2=z_2-a>0$. The sought coefficients $D_E^{||}$ and $D_E^\perp$ will make the resulting expression fully symmetric. (Within one medium this becomes interaction (\[66\]).) With several parallel layers the ${\bf u}_{{\varepsilon}\pm}$ and some other factors will change along with the change of the dielectric constant for various layers. But the dielectric constant for the half-plane $z<0$ stays fixed at $\varepsilon_1$. Thus for two half-planes the factors not containing $\varepsilon_2$ or $\mu_2$ can be separated out to be put into a quantity $L$ when determining the electric field through the layers. With this the electric field for the TM mode ($\sim{\bf G}_{{\varepsilon}1}\sim e^{\mp q_{\varepsilon_1}z}q_{\varepsilon_1}{\bf u}_{\varepsilon_1\pm}$) will be a slight modification of Eq. (I6.14) $${\bf E}/L=\left\{ \begin{array}{ll} q_{\varepsilon_1}\left(\frac{1}{{\varepsilon}_1}e^{-q_{\varepsilon_1}z}q_{\varepsilon_1}{\bf u}_{\varepsilon_1+}+Be^{q_{\varepsilon_1}z}{\bf u}_{\varepsilon_1-}\right),\quad & z_0<z<0\\ q(Ce^{-qz}{\bf u}_+ +C_1 e^{qz}{\bf u}_-), \quad & 0<z<a\\ q_{\varepsilon_2}De^{-q_{\varepsilon_2} z}{\bf u}_{\varepsilon_2+},\quad & a<z. \end{array} \right. \label{70}$$ where $B$, $C$, $C_1$, and $D$ are coefficients to be determined (relative to the $1/{\varepsilon}_1$ term of ). One condition to determine the coefficients is that the component of $\bf E$ along the $xy$ plane is continuous. However, for the other condition we here find it more convenient to turn to the magnetic field (instead of continuous ${\bf D}$ field normal to the interfaces) due to later extension to magnetic media. Then from Eq. (\[30\]) we will find Eq. (I6.17) $${\bf h}_\varepsilon\times{\bf E}=\zeta{\bf B}=\zeta\mu {\bf H}. \label{71}$$ With ${\bf E}/L=e^{\mp q_{\varepsilon}z} q_{\varepsilon}{\bf u}_{\varepsilon\pm}$ one finds (for $\varepsilon=1$ and $\varepsilon=\varepsilon_i$, $i=1,2$) $$\zeta{\bf B}/L=\{ik_\perp, 0, \pm q_\varepsilon\}\times\{ik_\perp, 0, \pm\frac{k_\perp^2}{q_\varepsilon}\}q_\varepsilon e^{\mp q_{\varepsilon}z}=\pm ik_\perp(q_\varepsilon^2-k_\perp^2)e^{\mp q_{\varepsilon}z}{\textcolor}{red}{\hat{\bf e}_y}=\pm ik_\perp\varepsilon\mu \zeta^2e^{\mp q_{\varepsilon}z}\hat{\bf e}_y \label{72}$$ where $\hat{\bf e}_y$ is the unit vector in the $y$ direction. The components of ${\bf E}$ and ${\bf H}$ parallel to the the interfaces are both continuous by which one gets the equations $$\begin{aligned} \nonumber q_{\varepsilon}\left(\frac{1}{{\varepsilon}_1}+B\right)&=&q\left( C+C_1\right)\\ \nonumber {\varepsilon}_1\left(\frac{1}{{\varepsilon}_1}-B\right)&=& C-C_1\\ \label{73} q(Ce^{-qa}+C_1e^{qa})&=&q_{{\varepsilon}_2} De^{-q_{{\varepsilon}_2}a}\\ \nonumber Ce^{-qa}+C_1e^{qa}&=&{\varepsilon}_2 De^{-q_{{\varepsilon}_2}a}\end{aligned}$$ The two last equations can be solved for $C$ and $C_1$ to be inserted in the equation that results from the elimination of $B$. This extends solution (I6.15) for $D$ to be $$D=D_E^{||}=\frac{4\kappa_1 e^{(q_{{\varepsilon}_2}-q)a}}{({\varepsilon}_1+\kappa_1)({\varepsilon}_2+\kappa_2)[1-A_n e^{-2qa}]}, \label{74a}$$ $$A_n=\frac{({\varepsilon}_1-\kappa_1)({\varepsilon}_2-\kappa_2)}{({\varepsilon}_1+\kappa_1)({\varepsilon}_2+\kappa_2)}, \quad \kappa_i=\frac{q_{{\varepsilon}_i}}{q}\quad (i=1,2). \label{74b}$$ The subscript $n=1,2,3,\cdots$ indicate the Matsubara frequencies (\[40\]) $\zeta=\zeta_n=K/(\hbar c)$, $K=2\pi n/\beta$. For the TE mode the electric field lies along the $y$ axis transverse to the ${\bf h}$ plane. Thus this component of the field can be written like Eq. (I6.18) $$E/L=\left\{ \begin{array}{ll} \mu_1 e^{-q_{\varepsilon_1}z}+Be^{q_{\varepsilon_1}z},\quad & z_0<z<0\\ Ce^{-qz} +C_1 e^{qz}, \quad & 0<z<a\\ De^{-q_{\varepsilon_2}z},\quad & a<z. \end{array} \right. \label{76}$$ where again $L$ is a quantity independent of ${\varepsilon}_2$ and $\mu_2$. The corresponding magnetic field will have components along both the $x$ and $z$ directions $$\zeta\mu{\bf H}/L=\zeta{\bf B}/L={\bf h}_\varepsilon\times{\bf E}/L= \{ik_\perp, 0, \pm q_\varepsilon\}\times\{0, 1, 0\}e^{\mp q_{\varepsilon}z}=(\mp q_{\varepsilon}\hat{\bf e}_x+ik_\perp\hat{\bf e}_z)e^{\mp q_{\varepsilon}z}. \label{77}$$ From this the boundary conditions give the equations (keeping $\mu$ although $\mu=1$ is considered so far) $$\begin{aligned} \nonumber \mu_1+B&=&C+C_1\\ \nonumber \frac{q_{{\varepsilon}_1}}{\mu_1}\left(\mu_1-B\right)&=&q( C-C_1)\\ \label{78} Ce^{-qa}+C_1e^{qa}&=&De^{-q_{{\varepsilon}_2}a}\\ \nonumber q(Ce^{-qa}+C_1e^{qa})&=&\frac{q_{{\varepsilon}_2}}{\mu_2} De^{-q_{{\varepsilon}_2}a}\end{aligned}$$ (The coefficients are relative to the $\mu_1$ term from .) The solution of these equations is $$D=D_E^{\perp}=\frac{4\mu_1\mu_2\kappa_1 e^{(q_{{\varepsilon}2}-q)a}}{(\kappa_1+\mu_1)(\kappa_2+\mu_2)[1-B_n e^{-2qa}]}, \label{79a}$$ $$B_n=\frac{(\kappa_1-\mu_1)(\kappa_2-\mu_2)}{(\kappa_1+\mu_1)(\kappa_2+\mu_2)}. \label{79b}$$ The Casimir force per unit area between the plates is given by Eq. (I6.32). This follows from Eq. (I4.4) where the product of the interaction (\[60\]) and the Green function (\[69\]) multiplied by $-q$ are integrated. With extension to plates with different dielectric constants this becomes $$f_{sur}=-\frac{1}{\pi\beta}\sum\limits_{n=0}^\infty{^\prime}\int\limits_{\zeta_n}^\infty q^2\,dq \, IQS. \label{80}$$ $$I=\int\limits_0^\infty \int\limits_0^\infty \frac{e^{-(q_{{\varepsilon}_2}+q)(u_2+a)} e^{-(q_{{\varepsilon}_1}+q)u_1}}{q_{{\varepsilon}_1}q}\,du_1du_2= \frac{e^{-(q_{{\varepsilon}_2}+q)a}}{q_{{\varepsilon}_1}q(q_{{\varepsilon}_1}+q)(q_{{\varepsilon}_2}+q)}. \label{81}$$ (with $z=z_2-z_1=u_1+u_2+a$). The prime means that the $n=0$ term is to be taken with half weight, and $\zeta_n=2\pi n/(\beta\hbar c)$ as follows from (\[40\]). The $Q$ is the $(9y/2)^2A$ term of (I6.32) which with (I5.5) is $({\varepsilon}-1)^2/4$. For different media this becomes $$Q=({\varepsilon}_1-1)({\varepsilon}_2-1)/4. \label{82}$$ Further from (I5.5) the $3y=(4\pi/3)\rho\beta\langle s^2\rangle$ which is the classical case. Now the polarizability $\alpha=\beta\langle s^2\rangle/3$ by which $3y=4\pi\rho\alpha$. This also holds for the quantum case where non-zero frequencies contribute, $\alpha=\alpha(\omega)$. This follows from Sec. 5 of Ref. [@brevik88].) Finally the last factor is $$S=S_{EE}=S^{||}+S^\perp, \quad S^{||}=D_E^{||}\langle H_{{\varepsilon}1}H_1^*\rangle, \quad S^\perp=D_E^{\perp}\langle H_2^2\rangle \label{83}$$ where $H_1=g_1 g_2$, $g_i=iq\hat s_{\perp i}\pm k_\perp \hat s_{||i}$, i.e. $H_1=H_{{\varepsilon}1}$ for ${\varepsilon}=1$. And from the Fourier transform relation the $H_1(-k_\perp)$ is the quantity needed. For the complex conjugate we have $H_1^*(k_\perp)=H_1(-k_\perp)$). With extension of Eqs. (I6.27) - (I6.30) to different, but non-magnetic, media we now get by use of Eqs.  and and then Eqs.  and ($\langle\hat s_{\perp i}^2\rangle=1/3$ etc.) $$\langle H_{{\varepsilon}1}H_1^*\rangle=\frac{1}{9}(k_\perp^2+q_{{\varepsilon}_1}q)(k_\perp^2+q_{{\varepsilon}_2}q), \label{84}$$ $$k_\perp^2+q_{{\varepsilon}_i}q=\frac{1}{{\varepsilon}_i-1}[{\varepsilon}_i q^2-q_{{\varepsilon}_i}+({\varepsilon}_i-1)q_{{\varepsilon}_i}q]=\frac{q}{{\varepsilon}_ i-1}({\varepsilon}_i-\kappa_i)(q+q_{{\varepsilon}_i}),\quad (i=1,2), \label{85}$$ $$\langle H_2^2\rangle=\frac{1}{9}\zeta^4; \quad \zeta^2=\frac{1}{{\varepsilon}_ i-1}(q_{{\varepsilon}_i}^2-q^2)=\frac{q}{{\varepsilon}_ i-1}(\kappa_i-1)(q_{{\varepsilon}_i}+q). \label{86}$$ Note these expressions for $k_\perp^2+q_{{\varepsilon}_i}q$ and $\zeta^2$ are valid only for $\mu_1=\mu_2=1$. Altogether for this case with $I$, $Q$, and $S$ inserted in (\[80\]) the known Lifshitz result (I2.9) for the force is recovered [@lifshitz55] $$f_{sur}=-\frac{1}{\pi\beta}\sum\limits_{n=0}^\infty{^\prime}\int\limits_{\zeta_n}^\infty q^2\,dq \,\left[\frac{A_n e^{-2qa}}{1-A_ne^{-2qa}}+\frac{B_n e^{-2qa}}{1-B_ne^{-2qa}}\right] \label{87}$$ with $A_n$ and $B_n$ given by Eqs. (\[74b\]) and (\[79b\]) respectively and with $q_{{\varepsilon}_i}$ ($i=1,2$) given by Eq. (\[62\]). Induced interaction with magnetic properties included {#sec5} ===================================================== In this section we want to show that expression for the Casimir force is valid also when magnetic interactions are present, i.e. $\mu_1$ and/or $\mu_2$ are different from one. In the previous section Sec. \[sec4\] the electric field from dipoles and their interactions for a pair of half-planes were found. This also included the induced magnetic fields and influence from magnetic permeabilities. Thus the coefficients $D_E^{||}$ and $D_E^\perp$ will stay unchanged. Also the main structure of the surface force (\[80\]) and integral (\[81\]) will be the same. However, expression (\[83\]) will have additional terms that involve magnetic moments too. For these additional terms the $Q$ given by Eq. (\[82\]) is modified by replacing ${\varepsilon}_i$ ($i=1,2$) with $\mu_i$ in the ways possible. Further the symmetry of electric and magnetic fields in Maxwell’s equations (\[30\]) is utilized. The fully new term to obtain is the induced interaction between electric and magnetic dipole moments. With $\mu\neq1$ expressions (\[83\]) and (\[84\]) for induced interaction between electric dipole moments will remain unchanged, but results (\[85\]) and (\[86\]) will not hold any longer. We find the (\[85\]) can be modified in two useful ways ($q_{\varepsilon}^2=k_\perp^2+{\varepsilon}\mu\zeta^2$, $\kappa=q_{\varepsilon}/q$) $$\begin{aligned} \nonumber k_\perp^2+q_{\varepsilon}q&=&\frac{1}{{\varepsilon}-1}[{\varepsilon}k_\perp^2-k_\perp^2+({\varepsilon}-1)q_{\varepsilon}q]\\ &=&\frac{1}{{\varepsilon}-1}[({\varepsilon}q^2-{\varepsilon}\zeta^2)-(q_{\varepsilon}^2-{\varepsilon}\mu\zeta^2)+({\varepsilon}-1)q_{\varepsilon}q] \label{90}\\ \nonumber &=&\frac{q}{{\varepsilon}-1}[{\varepsilon}-\kappa)(q+q_{\varepsilon})+(\mu-1){\varepsilon}\zeta^2],\end{aligned}$$ and likewise from symmetry with respect to ${\varepsilon}$ and $\mu$ $$k_\perp^2+q_{\varepsilon}q=\frac{q}{\mu-1} [(\mu-\kappa)(q+q_{\varepsilon})+({\varepsilon}-1)\mu\zeta^2]. \label{91}$$ Due to symmetry the contributions from the induced interactions between magnetic dipole moments will be like the ones of Eq. (\[83\]) since with $q_{\varepsilon}^2=k_\perp^2+{\varepsilon}\mu\zeta^2$ or with and the $k_\perp^2+q_{\varepsilon}q$ stays unchanged by interchange of $\mu$ and ${\varepsilon}$ $$S_{HH}=D_H^{||}\langle H_{{\varepsilon}1}H_1^*\rangle+D_H^{\perp}\langle H_2^2\rangle. \label{92}$$\ tc[red]{}[Further comparing]{} with expressions (\[74a\]) and (\[79a\]) with interchange of $\mu$ and ${\varepsilon}$ one finds $$D_E^\perp=\mu_1\mu_2 D_H^{||}, \quad D_H^\perp={\varepsilon}_1 {\varepsilon}_2 D_E^{||}. \label{93}$$ The remaining terms are those from induced interactions between electric and magnetic dipoles. Thus we need the magnetic fields due to the electric ones and vice versa that follows from symmetry. With (\[66\]), (\[68\]), and (\[69\]) the TM electric field for $z_2>a$ can be written in the form ${\bf E}=(J_{\varepsilon}/s_2) D_E^{||}g_{{\varepsilon}_1 1}(q_{{\varepsilon}_2}/k_\perp){\bf u}_{ {{\varepsilon}_2}\pm}$. Thus with Eq. (\[72\]) the corresponding magnetic field will be $$\zeta\mu_2 {\bf H}=\zeta{\bf B}=(J_{\varepsilon}/s_2) D_E^{||} g_{{\varepsilon}_1 1}(\pm ik_\perp{\varepsilon}_2 \mu_2 \zeta^2 \hat {\bf e}_y)/k_\perp \label{94}$$ with $g_{{\varepsilon}1}$ given by Eq. (\[67b\]). This is to be multiplied with the magnetic dipole moment ${\bf s}_2$ (instead of electric dipole moment) whose unit vector has the transverse component $\hat{\bf s}_2^\perp$ ( in the $y$ direction) to obtain the equivalent of $H_{{\varepsilon}1}$ in Eq.  in view of the common prefactor $J_{\varepsilon}D_E^{||}$ $$H_{{\varepsilon}EH}^{||}=\frac{1}{\zeta\mu_2 J_{\varepsilon}D_E^{||}}{\bf H}\cdot{\bf s}_2=\pm \zeta g_{{\varepsilon}_1 1} {\varepsilon}_2 \hat s_2^\perp. \label{95}$$ This multiplied with the corresponding quantity in vacuum is averaged over the orientations of the dipole moments to give $$\langle H_{{\varepsilon}EH}^{||} H_{EH}^{||}\rangle_1=-\frac{1}{9}\zeta^2(k_\perp^2+q_{{\varepsilon}_1} q){\varepsilon}_2. \label{96}$$ Note in the magnitude of the electric dipole moment $s_2$ is replaced by the magnetic one. This is taken into account by replacing ${\varepsilon}_2-1$ with $\mu_2-1$ in Eq.  for the factor $Q$ to be used in Eq.  below. The $H_{EH}^{||}=H_{{\varepsilon}EH}^{||}$ for ${\varepsilon}_i=\mu_i=1$ ($i=1,2$) with $\zeta\rightarrow -\zeta$. The $\zeta\rightarrow -\zeta$ (and $k_\perp\rightarrow -k_\perp)$ again follows from the Fourier transform relation . Likewise one has the interaction between an electric dipole in halfplane 2 with a magnetic one in halfplane 1. Due to symmetry the corresponding expression must be $$\langle H_{{\varepsilon}EH}^{||} H_{EH}^{||}\rangle_2=-\frac{1}{9}\zeta^2(k_\perp^2+q_{{\varepsilon}_2} q){\varepsilon}_1 \label{97}$$ Finally for the TM mode one has the interaction between the magnetic moments in the two halfplanes. But again due to symmetry of Maxwell’s equations (\[30\]) this for the magnetic field (in the $y$ direction) is equivalent to the TE field. Thus the corresponding magnetic interactions will be the same and follow from the general form of (\[86\]) as $\langle H_2^2\rangle=\zeta^2/9$. The resulting surface force is obtained by adding the various contributions multiplied with factors $Q$ given by (\[82\]) with one or both ${\varepsilon}$ replaced by $\mu$. Adding together for the TM contribution the product $QS^{||}$ from Eqs. (\[82\]) and is replaced by $$\begin{aligned} \nonumber S_Q^{||}&=&\frac{1}{4}[({\varepsilon}_1-1)({\varepsilon}_2-1)D_E^{||}\langle H_{{\varepsilon}1}H_1^*\rangle+({\varepsilon}_1-1)(\mu_2-1)D_E^{||}\langle H_{{\varepsilon}EH}^{||}H_{EH}^{||}\rangle_1\\ &+&({\varepsilon}_2-1)(\mu_1-1)D_E^{||}\langle H_{{\varepsilon}EH}^{||}H_{EH}^{||}\rangle_2+(\mu_1-1)(\mu_2-1)D_H^{\perp}\langle H_2^2\rangle]. \label{100}\end{aligned}$$ Inserting from Eqs. , , , and one finds $$\begin{aligned} \nonumber S_Q^{||}&=&\frac{D_E^{||}}{36}[({\varepsilon}_1-1)({\varepsilon}_2-1)(k_\perp^2+q_{{\varepsilon}_1}q)(k_\perp^2+q_{{\varepsilon}_2}q)-({\varepsilon}_1-1)(\mu_2-1){\varepsilon}_2\zeta^2(k_\perp^2+q_{{\varepsilon}_1}q)\\ &-&({\varepsilon}_2-1)(\mu_1-1){\varepsilon}_1\zeta^2(k_\perp^2+q_{{\varepsilon}_2}q)+(\mu_1-1)(\mu_2-1){\varepsilon}_1 {\varepsilon}_2\zeta^4] \label{101}\\ &=&\frac{D_E^{||}}{36}[({\varepsilon}_1-1)(k_\perp^2+q_{{\varepsilon}_1}q)-(\mu_1-1){\varepsilon}_1\zeta^2][({\varepsilon}_2-1)(k_\perp^2+q_{{\varepsilon}_2}q)-(\mu_2-1){\varepsilon}_2\zeta^2]. \nonumber\end{aligned}$$ Finally with use of Eq.  the result is $$S_Q^{||}=\frac{D_E^{||}}{36}q^2({\varepsilon}_1-\kappa_1)(q+q_{{\varepsilon}_1})({\varepsilon}_2-\kappa_2)(q+q_{{\varepsilon}_2}). \label{102}$$ This result is precisely the result for $S_Q^{||}=QS^{||}$ wth Eqs.  and inserted. However, the former result was only valid for $\mu=1$ while result is valid for general $\mu$. To obtain the resulting force the contribution from the TE mode with magnetic interaction included (i.e. $\mu\neq 1$) is also needed. Again one may find the magnetic interactions via the corresponding magnetic field . However, we simplify by utilizing the symmetry between the electric and magnetic fields in Maxwells equations. Thus for the magnetic field the TE mode will be similar to the TM mode with exchange of ${\varepsilon}$ and $\mu$. The various contributions will be similar to those of Eq.  including the $S^\perp$ of . Adding together Eq.  will be replaced by $$S_Q^\perp=\frac{D_E^\perp}{36\mu_1 \mu_2}[(\mu_1-1)(k_\perp^2+q_{{\varepsilon}_1}q)-({\varepsilon}_1-1)\mu_1\zeta^2][(\mu_2-1)(k_\perp^2+q_{{\varepsilon}_2}q)-({\varepsilon}_2-1)\mu_2\zeta^2]. \label{103}$$ With Eqs.  and this becomes $$S_Q^\perp=\frac{D_E^\perp}{36\mu_1\mu_2}q^2(\mu_1-\kappa_1)(q+q_{{\varepsilon}_1})(\mu_2-\kappa_2)(q+q_{{\varepsilon}_2}). \label{104}$$ With $\mu_1=1$ and $\mu_2=1$ this result is the same as the result $S_Q^\perp=QS^\perp$ with Eq.  inserted in . Result generalizes expression for the Casimir force to arbitrary $\mu_1$ and $\mu_2$. Summary {#sec6} ======= We have studied the induced Casimir force between media that can have both dielectric and magnetic properties. Usually this force is attractive, but with both properties present it turns out that it can be repulsive. It is not obvious how this can be understood on physical grounds. In Sec. \[sec2\] we established a simple oscillator model that shows precisely this induced repulsive behavior. The model is a simple extension of the model studied earlier [@hoye03; @hoye16], where we showed the analog of the attractive forces produced by the TM and TE modes of dielectric media. Then by use of the statistical mechanical method introduced in Ref. [@brevik88] we evaluated the induced force between a pair of particles possessing both dielectric and magnetic properties. Further in Sec. \[sec4\] the statistical mechanical theory used in Ref. [@hoye98] was extended to a pair of half-planes with different dielectric constants. Finally, in Sec. \[sec5\] we extended the theory to a pair of half-planes that have both the mentioned properties and are separated by a distance $a$. The result of the latter, as might be expected, is a generalization of the well known Lifshitz formula as given by Eq. . Again the Casimir force is repulsive if of the media is mainly dielectric while the other is mainly magnetic. With equal media in the two half-planes, the Casimir force is always attractive. Acknowledgment {#acknowledgment .unnumbered} ============== We acknowledge financial support from the Research Council of Norway, Project 250346. [99]{} M. Bordag, G. L. Klimchitskaya, U. Mohideen,, and V. M. Mostepanenko, [*Advances in the Casimir Effect*]{} (Oxford University Press, Oxford, UK, 2009). K. A. MIlton, [*The Casimir Effect: Physical Manifestations of Zero-Point Energy*]{} (World Scientific, Singapore, 2001). V. A. Parsegian, [*Van der Waals Forces: A Handbook for Biologists, Chemists, Engineers, and Physicists*]{} (World Scientific, Singapore, 2001). J. N. Munday, F. Capasso, and V. A. Parsegian, Nature (London) [**457**]{}, 07610 (2009). M. Bostr[ö]{}m, Bo E. Sernelius, I. Brevik, and B. W. Ninham, Phys. Rev. A [**85**]{}, 010701(R) (2012). H. Wennerstr[ö]{}m, J. Daicic, and B. W. Ninham, Phys. Rev. A [**60**]{}, 2581 (1999). B. W. Ninham and J. Daicic, Phys. Rev. A [**57**]{}, 1870 (1998). T. H. Boyer, Phys. Rev. A [**9**]{}, 2078 (1974). G. Feinberg and J. Sucher, J. Chem. Phys. [**48**]{}, 3333 (1968); Phys. Rev. A [**2**]{}, 2395 (1970). P. Richmond and B. W. Ninham, J. Phys. C: Solid St. Phys. [**4**]{}, 1988 (1971). O. Kenneth, I. Klich, A. Mann, and M. Revzen, Phys. Rev. Lett. [**89**]{}, 033001 (2002). N. Inui, Phys. Rev. A [**83**]{}, 032513 (2011); [**84**]{}, 052505 (2011). V. K. Pappakrishnan, P. C. Mundru, and D. A. Genov, Phys. Rev. B [**89**]{}, 045430 (2014). K. Sinha, Phys. Rev. A [**97**]{}, 032513 (2018). J.S. H[ø]{}ye, I. Brevik, J.B. Aarseth, and K.A. Milton, Phys. Rev. E [**67**]{}, 056116 (2003). J.S. H[ø]{}ye, I. Brevik, and K.A. Milton, Phys. Rev. A [**94**]{}, 032113 (2016). I. Brevik and J.S. H[ø]{}ye, Physica A [**153**]{}, 420 (1988). J.S. H[ø]{}ye and I. Brevik, Physica A [**259**]{}, 165 (1998). J.S. H[ø]{}ye and G. Stell, J. Chem. Phys. [**75**]{}, 5133 (1981); M.J. Thomson, K. Schweizer, and D. Chandler, J. Chem. Phys. [**76**]{}, 1128 (1982). J.S. H[ø]{}ye and G. Stell, J. Chem. Phys. [**77**]{}, 5173 (1982). E.M. Lifshitz, Zh. Exp. Teor. Fiz. [**29**]{}, 94 (1955); E.M. Lifshitz and L.P. Pitaevskii, [*[Statistical Physics]{}*]{}, part 2, in L.D. Landau and E. M. Lifshitz, [*Course of Theoretical Physics*]{}, vol. 9 (Pergamon, Oxford, 1980).

--- abstract: 'We study two different initial conditions for fermions for the problem of pair production of fermions coupled to a classical electromagnetic field with backreaction in [(1+1)]{} boost-invariant coordinates. Both of these conditions are consistent with fermions initially in a vacuum state. We present results for the proper time evolution of the electric field $E$, the current $J$, the matter energy density $\varepsilon$, and the pressure $p$ as a function of the proper time for these two cases. We also determine the interpolating number density as a function of the proper time. We find that when we use a “first order adiabatic” vacuum initial condition or a “free field” initial condition for the fermion field, we obtain essentially similar behavior for physically measurable quantities. The second method is computationally simpler, it is twice as fast and involves half the storage required by the first method.' author: - Bogdan Mihaila - 'John F. Dawson' - Fred Cooper bibliography: - 'johns.bib' title: | Fermion pair production in QED and the backreaction problem\ in (1+1)-dimensional boost-invariant coordinates revisited --- Introduction {#s:intro} ============ Particle production from strong fields has a long history starting with Schwinger’s classic paper [@r:Schwinger:1951fk]. A detailed history of this subject can be found in two recent reviews [@ref:Dunne06; @CooperDawsonMihaila0806.1249]. One of the many applications of pair production has been as a model for particle production in the central rapidity region following a relativistic heavy ion collision. Following such a collision, there is experimental evidence that the production of particles is “boost-invariant” [@r:CFSprd75; @r:Bjorken:1983zr] which leads to measurable quantities such as energy densities being functions of the fluid proper time alone. The initial conditions we want to study for this problem are that the number of pairs starts out zero and that the initial induced current in the Maxwell (backreaction) equation for the electric field is also zero. We then want to study the proper time evolution of the expectation value of the energy density, pressure, and current of the produced particles and the evolution of the electric field. In an earlier paper on this topic [@r:Cooper:1993uq] one particular set of initial conditions consistent with having no pairs of particles produced before the collision at initial proper time, $\tau=\tau_0$, led to the need for doubling the number of fermion solutions in order to start with zero induced current in the Maxwell equation for the electric field. In the paper by Cooper *et al.* [@r:Cooper:1993uq], two sets of solutions for the second-order squared Dirac equation were used in order to satisfy the desired initial conditions of having zero initial current. Similar results were presented in a 1992 paper by Kluger *et al.* [@r:Kluger:1992fk] for the Cartesian case. In those papers, the initial conditions were taken to correspond to a first-order adiabatic approximation to the second-order Dirac equation, which forced them to average over two different solutions of the Dirac equation so that the current vanished at $\tau = \tau_0$. In the present work, we consider a slightly different initial state, namely approximate free fields for the fermions, which automatically leads to a zero current at $\tau=\tau_0$. This initial condition was used earlier by Cooper and Savage [@r:CS02] in their study of the dynamics of the chiral phase transition in the [(2+1)]{} dimensional Gross-Neveu model. The free field initial condition does not require doubling the number of solutions as did the adiabatic choice. We compare the evolution of the problem for both initial conditions and show that at short to moderate times they are equivalent and are slightly different at very late times. This semi-classical approximation to the initial value QED problem describes the fermions as a quantum field but treats the electric field classically. The current used in Maxwell’s equation is calculated using the vacuum expectation value of the quantum Dirac current. As discussed in previous papers [@r:CHKMPAprd94], this approximation is equivalent to the first term in a large-N approximation to N-QED where there are N flavors of fermions present. The method we use for numerically solving this problem is a shooting method to numerically step out solutions of the equations from initial conditions. An adiabatic analysis of the form of the solutions is used to determine the behavior of the solutions at large momentum and to isolate divergences and perform renormalization as well as to choose appropriate initial states for fermions that are appropriate vacuum states. We study here the problem in [(1+1)]{} boost-invariant coordinates. This kinematic situation is related to the kinematics of the early phase of plasma evolution following a relativistic heavy ion collision with the electric field a simplification for the semiclassical chromoelectric field expected to be produced in that situation. We study [(1+1)]{} dimensions for simplicity here, where charge renormalization is finite. However, the same methods of solution used here can be applied to the case of [(3+1)]{} dimensions in both Cartesian and boost-invariant coordinates. We will present results for [(3+1)]{} dimensions for QED and QCD elsewhere. This paper is organized as follows: In Sec. \[s:notations\] we review briefly the equations we will need to solve for QED in [(1+1)]{} dimensional boost-invariant coordinates (for a detailed derivation, see e.g. Refs.  and ). In Secs. \[s:Maxwell\] and \[s:adiabaticexp\] we discuss the backreaction equation and the adiabatic expansion and charge renormalization, whereas in Sec. \[s:energymomentum\] we review the calculation of the energy-momentum tensor. The two types of initial conditions are introduced in Sec. \[s:initialconditions\]. We present results of our numerical simulations in Sec. \[s:results\] and conclude in Sec. \[s:conclusions\]. Notation and equations {#s:notations} ====================== In our simplified kinematics, in [(1+1)]{} dimensions, we choose the longitudinal axis of the collision to be the $z$-axis. Then, in the Cartesian frame, we want to solve the set of equations: $$\label{bi.e:diracI} \bigl \{ \, \gamma^{a} \, [ \, i \partial_a - e \, A_{a}(\xi) \, ] - m \, \bigr \} \, \hat{\psi}(\xi) = 0 \>,$$ where $\xi$ is shorthand for the Cartesian pair $(t,z)$, $\hat{\psi}(\xi)$ is a fermi field satisfying the anti-commutation relation: $$\label{bi.e:fermianticomm} {\ensuremath{\{ \, \hat{\psi}_{\alpha}^{\phantom\dagger}(z,t), \hat{\psi}_{\alpha}^{\dagger}(z',t) \, \}}} = \delta_{\alpha,\alpha'} \, \delta(z - z') \>,$$ and $A_a(\xi)$ is a classical field satisfying Maxwell’s equations: $$\label{bi.e:maxwellI} \partial_a \, F^{ab}(\xi) = J^b(\xi) \>, \quad F^{ab}(\xi) = \partial^a A^b(\xi) - \partial^b A^a(\xi) \>.$$ The current is given by: $$\label{bi.e:JdefI} J^b(\xi) = \frac{e}{2} \, {\ensuremath{\langle \, {\ensuremath{[ \, \hat{\bar{\psi}}(\xi) , \gamma^a \, \hat{\psi}(\xi) \, ]}} \, \rangle}} \>.$$ The $\gamma$-matrices satisfy ${\ensuremath{\{ \, \gamma^a , \gamma^b \, \}}}=2\, \eta^{a,b}$ and are given by: $$\gamma^0 = \begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix} \>, \>\> \gamma^3 = \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix} \>, \>\> \gamma^5 = \gamma^0 \gamma^3 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} \>.$$ Boost-invariant coordinates $x^\mu = ( \, \tau, \eta \, )$ are defined by: $$\label{bi.e:tauetadef} t = \tau \, \cosh{\eta} \>, \qquad z = \tau \, \sinh{\eta} \>.$$ The connection between the Cartesian frame ($d\xi^a$), and the boost-invariant frame ($dx^{\mu}$) is described by a vierbein matrix $V^{a}{}_{\mu}(x)$, given by: $$\begin{gathered} {\mathrm{d}}\xi^{a} = V^{a}{}_{\mu}(x) \, {\mathrm{d}}x^{\mu} \>, \qquad \partial_{\mu} = V^{a}{}_{\mu}(x) \, \partial_{a} \>, \label{bi.e:Vdef} \\ V^{a}{}_{\mu}(x) {\equiv}\frac{\partial \xi^{a}}{\partial x^{\mu}} = \begin{pmatrix} \cosh \eta, & \tau \sinh \eta \\ \sinh \eta, & \tau \cosh \eta \end{pmatrix} \>, \notag\end{gathered}$$ and its inverse: $$\begin{gathered} {\mathrm{d}}x^{\mu} = V^{\mu}{}_{a}(x) \, {\mathrm{d}}\xi^{a} \>, \quad \partial_{a} = V^{\mu}{}_{a}(x) \, \partial_{\mu} \>, \label{bi.e:Vinvdef} \\ V^{\mu}{}_{a}(x) {\equiv}\frac{\partial x^{\mu}}{\partial \xi^{a}} = \begin{pmatrix} \cosh \eta, & - \sinh \eta \\ - \sinh \eta / \tau, & \cosh \eta / \tau \end{pmatrix} \>. \notag\end{gathered}$$ The $\gamma$-matrices in this frame are denoted by a tilde: $\tilde{\gamma}^{\mu}(x) {\equiv}V^{\mu}{}_{a}(x) \, \gamma^{a}$, and satisfy: $$\label{bi.ae.e:gammamugammanu} {\ensuremath{\{ \, \tilde{\gamma}^{\mu}(x) , \tilde{\gamma}^{\nu}(x) \, \}}} = 2 \, g^{\mu\nu}(x) \>, \quad g^{\mu\nu}(x) = {\mathrm{diag}( \, 1, -1/\tau^2 \, )} \>.$$ Dirac’s equation becomes in this frame: $$\label{bi.e:diracII} \bigl [ \, \tilde{\gamma}^{\mu}(x) \, ( \, i \, \partial_{\mu} - e \, A_{\mu}(x) \, ) - m \, \bigr ] \, \hat{\psi}(x) = 0 \>,$$ where the fermi field obeys the anti-commutation relation: $$\label{bi.e:psifieldanticomm} {\ensuremath{\{ \, \hat{\psi}_{\alpha}^{\phantom(}(\tau,\eta) , \hat{\bar{\psi}}_{\alpha'}^{\phantom(}(\tau,\eta') \, \}}} = \tilde{\gamma}^{\tau}_{\alpha,\alpha'}(\eta) \, \delta( \eta - \eta' ) / \tau \>.$$ However, it is much simpler to make a similarity transformation to a system of coordinates where the vierbein becomes diagonal [@ref:CooperDawsonMihaila06]. In this rotated system, the $\gamma$-matrices are denoted by a bar: $$\label{bi.e:gammabar} S^{-1}(\eta) \, \tilde{\gamma}^{\mu}(x) \, S(\eta) = \bar{\gamma}^{\mu}(\tau) \>,$$ where the matrix $S(\eta)$ is given by: $$\label{bi.e:Sdef} S(\eta) = {\exp [ \, \eta \, \gamma^5 / 2 \, ]} = \cosh( \eta/2 ) + \gamma^5 \, \sinh( \eta /2 ) \>.$$ The $\bar{\gamma}^{\mu}(\tau)$ matrices are given explicitly by: $$\label{bi.e:Vbardefs} \bar{\gamma}^{\tau} = \gamma^{0} \>, \qquad \bar{\gamma}^{\eta}(\tau) = \gamma^3 / \tau \>,$$ So if we define a new fermi field $\hat{\phi}(x)$ by: $$\label{bi.e:psitophi} \hat{\psi}(x) = S(\eta) \, \hat{\phi}(x) / \sqrt{\tau} \>,$$ Dirac’s equation becomes: $$\label{bi.e:diracIII} \bigl [ \, i \, \bar{\gamma}^{\mu}(\tau) \, \nabla_{\mu} - m \, \bigr ] \, \hat{\phi}(\tau,\eta) / \sqrt{\tau} = 0 \>,$$ where $\nabla_{\mu} = \partial_{\mu} + \Pi_{\mu}(x) + i e \, A_{\mu}(x)$, with $\Pi_{\mu}(x) = S^{-1}(\eta) \, ( \, \partial_{\mu} S(\eta) \, )$, is the covariant derivative. Here $\hat{\phi}(x)$ obeys the simpler anti-commutation relation: $$\label{bi.e:phifieldanticomm} {\ensuremath{\{ \, \hat{\phi}_{\alpha}^{\phantom\dagger}(\tau,\eta) , \hat{\phi}_{\alpha'}^{\dagger}(\tau,\eta') \, \}}} = \delta_{\alpha,\alpha'}^{\phantom\dagger} \, \delta( \eta - \eta' ) \>.$$ For our case, the only non-vanishing $\Pi_{\mu}(x)$ is for $\Pi_{\eta} = \gamma^5 / 2$. In the boost-invariant frame, we work in the temporal gauge and choose $A_{\mu}(x) = (\, 0, -A(\tau) \,)$. That is $A(\tau)$ as the negative of the covariant component in the boost-invariant frame. So simplifies to: $$\label{bi.e:diracIV} \bigl \{ \, i \, \gamma^0 \, \partial_\tau + \gamma^3 \, \bigl [ \, i \, \partial_{\eta} + e \, A(\tau) \, \bigr ] / \tau - m \, \bigr \} \, \hat{\phi}(\tau,\eta) = 0 \>.$$ We now expand the field $\hat{\phi}(\tau,\eta)$ in a fourier series given by: $$\label{bi.e:phiexpand} \hat{\phi}(\tau,\eta) = {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda=\pm} \hat{A}_{k}^{(\lambda)} \, e^{i k \eta} \, \phi_{k}^{(\lambda)}(\tau) \>,$$ where we have introduced the notation $[\mathrm{d}k] = \mathrm{d}k/(2\pi)]$. Here, $\hat{A}_{k}^{(\lambda)}$ are mode operators and $\phi_{k}^{(\lambda)}(\tau)$ are two independent mode functions satisfying the equation: $$\label{bi.e:diracV} \bigl [ \, i \, \gamma^0 \, \partial_\tau - \gamma^3 \, {\pi_k}(\tau) - m \, \bigr ] \, \phi_k^{(\lambda)}(\tau) = 0 \>,$$ where $${\pi_k}(\tau) = \frac{1}{\tau} \ [ \, k - e \, A(\tau) \, ] \>.$$ It is now useful to add and subtract the upper and lower components of the spinor $\phi_k^{(\lambda)}$ by writing: $$\label{bi.e:rot} \phi_k^{(\lambda)}(\tau) = U \, F_k^{(\lambda)}(\tau) \>, \>\>\text{with}\>\> U = \frac{1}{\sqrt{2}} \begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix} \>.$$ Here $U^{\dagger} = U^{-1} = U^{T}$. Then $F_k^{(\lambda)}(\tau)$ satisfies an equation of Hamiltonian form: $$\label{bi.e:Fequ} i \, \partial_\tau \, F_k^{(\lambda)}(\tau) = H(\tau) \, F_k^{(\lambda)}(\tau) \>,$$ with $$\label{bi.de.e:Hdef} H(\tau) = \begin{pmatrix} {\pi_k}(\tau) & m \\ m & - {\pi_k}(\tau) \end{pmatrix} = {\mathbf{K}}_k(\tau) \cdot {\boldsymbol{\sigma}}\>.$$ Here ${\mathbf{K}}_k(\tau)$ is a vector defined in an abstract space $\mathcal{R}$ with unit vectors $({\hat{\mathbf{e}}}_1,{\hat{\mathbf{e}}}_2,{\hat{\mathbf{e}}}_3)$ and given by: $$\label{bi.e:kvecdef} {\mathbf{K}}_k(\tau) = m \, {\hat{\mathbf{e}}}_1 + {\pi_k}(\tau) \, {\hat{\mathbf{e}}}_3 \>.$$ We can introduce the $2 \times 2$ dimensional density matrix $\rho_k(\tau)$ and a “polarization” vector ${\mathbf{P}}_k(\tau)$ in $\mathcal{R}$ with the definitions: $$\label{bi.e:rhobPdefs} \rho_k^{(\lambda)}(\tau) = F_k^{(\lambda)}(\tau) \, F_k^{(\lambda)\,\dagger}(\tau) = \frac{1}{2} \, ( \, 1 + {\mathbf{P}}_k^{(\lambda)}(\tau) \cdot {\boldsymbol{\sigma}}\, ) \>.$$ Then from , the polarization vector ${\mathbf{P}}_k^{(\lambda)}(\tau)$ obeys the vector equation of motion: $$\label{bi.e:Peom} \partial_\tau \, {\mathbf{P}}_k^{(\lambda)}(\tau) = 2 \, {\mathbf{K}}_k(\tau) \times {\mathbf{P}}_k^{(\lambda)}(\tau) \>.$$ Since $H(\tau)$ in Eq.  is hermitian, $F_k^{(\lambda)}(\tau)$ satisfies a conservation equation: $$\label{bi.e:conseqs} \partial_\tau \, [ \, F_k^{(\lambda)\,\dagger}(\tau) \, F_k^{(\lambda')}(\tau) \, ] = 0 \>.$$ So if we choose the two spinors to be orthonormal at $\tau = \tau_0$, they remain orthonormal for all $\tau$. In Sec. \[s:initialconditions\] we show how to do this. So we can assume that these spinors are orthonormal and complete for all $\tau$: \[bi.e:orthocomp\] $$\begin{aligned} F_{k}^{(\lambda)\,\dagger}(\tau) \, F_{k}^{(\lambda')}(\tau) &= \delta_{\lambda,\lambda'} \>, \label{bi.de.e:ortho} \\ \sum_{\lambda=\pm} F_{k}^{(\lambda)}(\tau) \, F_{k}^{(\lambda)\,\dagger}(\tau) &= 1 \>. \label{bi.de.e:complete}\end{aligned}$$ Probability conservation also requires that the polarization vector ${\mathbf{P}}_{k}^{(\lambda)}(\tau)$ for both of these solutions to remain on the unit sphere for all time $\tau$. So to summarize, the fermi field can be written as: $$\label{bi.e:psitoF} \hat{\psi}(\tau,\eta) = S(\eta) \, U \, \hat{F}(\tau,\eta) / \sqrt{\tau} \>,$$ where the field $\hat{F}(\tau,\eta)$ obeys the anti-commutation relation: $$\label{bi.e:Ffieldanticomm} {\ensuremath{\{ \, \hat{F}_{\alpha}^{\phantom\dagger}(\tau,\eta) , \hat{F}_{\alpha'}^{\dagger}(\tau,\eta') \, \}}} = \delta_{\alpha,\alpha'}^{\phantom\dagger} \, \delta( \eta - \eta' ) \>.$$ and is expanded in terms of the spinors $F_{k}^{(\lambda)}(\tau)$ which satisfy : $$\label{bi.e:Fexpansion} \hat{F}(\tau,\eta) = {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda=\pm} \hat{A}_{k}^{(\lambda)} \, e^{i k \eta} \, F_{k}^{(\lambda)}(\tau) \>.$$ We can use this orthogonality to invert to get: $$\label{bi.e:Aexp} \hat{A}_{k}^{(\lambda)} = \int_{-\infty}^{+\infty} {\mathrm{d}}\eta \, e^{-i k \eta} \, F_{k}^{(\lambda)\,\dagger}(\tau) \, \hat{F}(\tau,\eta) \>,$$ for any time $\tau$. Using , we then find that the mode operators $\hat{A}_{k,s}^{(\lambda)}$ obey the anti-commutation relation: $$\label{bi.e:Aanticomm} {\ensuremath{\{ \, \hat{A}_{k}^{(\lambda)} , \hat{A}_{k'}^{(\lambda')\,\dagger} \, \}}} = ( 2\pi ) \, \delta_{\lambda,\lambda'}^{\phantom(} \, \delta( k - k' ) \>.$$ It is traditional to define separate positive and negative energy operators by setting: $$\label{bi.e:tradition} \hat{A}_{k}^{(+)} = \hat{a}_{k}^{\phantom(} \>, \qquad\text{and}\qquad \hat{A}_{k}^{(-)} = \hat{b}_{-k}^{\dagger} \>.$$ We choose our initial state to be the vacuum with no particle or anti-particle present. Then: $$\label{bi.e:abvac} \hat{a}_{k}^{\phantom(} \, { | \, 0 \, \rangle } = 0 \>, \qquad\text{and}\qquad \hat{b}_{k}^{\phantom(} \, { | \, 0 \, \rangle } = 0 \>.$$ This means that: $$\label{bi.e:expectcomm} {\ensuremath{\langle \, {\ensuremath{[ \, \hat{A}_{k}^{(\lambda)\,\dagger} , \hat{A}_{k'}^{(\lambda')} \, ]}} \, \rangle}} = - \, ( 2\pi ) \, \lambda \, \delta_{\lambda,\lambda'}^{\phantom(} \, \delta( k - k' ) \>,$$ a result we will use in the next section. Maxwell’s equation {#s:Maxwell} ================== Maxwell’s equation is given in Cartesian coordinates in Eq.  with the current given in Eq. . For our boost-invariant coordinates, Maxwell’s equation reads: $$\label{bi.me.e:maxwellcov} \frac{1}{\sqrt{-g}} \, \partial_{\mu} \bigl [ \, \sqrt{-g} \, F^{\mu\nu}(x) \, \bigr ] = J^{\nu}(x) \>,$$ where $\sqrt{-g} = \tau$. Now $A_{\mu} = ( \, 0, - A(\tau) \, )$, so the only non-vanishing elements of the field tensor are: $$\label{bi.me.e:Edef} F_{\tau,\eta}(x) = - F_{\eta,\tau}(x) = - \partial_\tau A(\tau) {\equiv}\tau \, E(\tau) \,$$ This last equation defines what we call the electric field $E(\tau) {\equiv}( \, \partial_\tau A(\tau) \, ) / \tau$. Then using the metric $g^{\mu\nu}(x) = {\mathrm{diag}( \, 1, -1/\tau^2 \, )}$, we get: $$\label{bi.me.e:Fupper} F^{\tau,\eta}(\tau) = - F^{\eta,\tau}(\tau) = - E(\tau) / \tau \>,$$ and Maxwell’s equation becomes: $$\label{bi.me.e:maxwellcovII} \partial_\tau E(\tau) = - J(\tau) \>.$$ Here we have defined a “reduced” current $J(\tau)$ by: $$\label{bi.me.e:jsdef} \begin{split} J(\tau) &= \frac{e \, \tau}{2} \, {\ensuremath{\langle \, {\ensuremath{[ \, \hat{\bar{\psi}}(\eta,\tau) , \tilde{\gamma}^{\eta}(\tau) \, \hat{\psi}(\eta,\tau) \, ]}} \, \rangle}} \\ &= \frac{e}{2 \, \tau} \, {\ensuremath{\langle \, {\ensuremath{[ \, \hat{\phi}^{\dagger}(\eta,\tau) , \gamma^5 \, \hat{\phi}(\eta,\tau) \, ]}} \, \rangle}} \>, \end{split}$$ Using the field expansion and the expectation value of the mode operators, we find for the reduced current: $$\begin{aligned} J(\tau) &= \frac{e}{2 \, \tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda=\pm 1} {\int_{-\infty}^{+\infty} [\mathrm{d} k'] \, } \sum_{\lambda'=\pm 1} \notag \\ &\times e^{i (k - k') \eta} \, \bigl [ \, \phi_{k}^{(\lambda)\,\dagger}(\tau) \, \gamma^5 \, \phi_{k'}^{(\lambda')\phantom\dagger}(\tau) \, \bigr ] \, {\ensuremath{\langle \, {\ensuremath{[ \, \hat{A}_{k}^{(\lambda)\,\dagger} , \hat{A}_{k'}^{(\lambda')} \, ]}} \, \rangle}} \notag \\ &= - \frac{e}{2 \, \tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda=\pm 1} \lambda \, \bigl [ \, F_{k}^{(\lambda)\,\dagger}(\tau) \, \sigma_3 \, F_{k}^{(\lambda)\phantom\dagger}(\tau) \, \bigr ] \notag \\ &= - e {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> P_3^{(+)}({\pi_k},\tau) \>. \label{bi.me.e:js}\end{aligned}$$ Here we have used the completeness statement to write the current in terms of positive energy solutions only. In the last line, we changed integration variables from $k$ to ${\pi_k}(\tau)$, using ${\mathrm{d}}{\pi_k}= {\mathrm{d}}k / \tau$, and defined ${\mathbf{P}}({\pi_k},\tau) {\equiv}{\mathbf{P}}_k(\tau)$. Maxwell’s equation becomes: $$\label{bi.me.e:maxwellfinal} \partial_\tau E(\tau) = e {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> P_3^{(+)}({\pi_k},\tau) \>.$$ Recall that $P_3$ is the third component of the polarization vector in the space $\mathcal{R}$. Adiabatic expansion {#s:adiabaticexp} =================== The large momentum behavior of the solutions of the Dirac equation can be obtained by looking at the adiabatic expansion of these solutions. Perhaps the simplest way to do this is from the polarization equation . In order to count powers of time derivatives, we put: $\partial_\tau \mapsto \epsilon \, \partial_\tau$, and set: $$\label{bi.ae.e:PexpandI} {\mathbf{P}}_k^{\phantom)}(\tau) = {\mathbf{P}}_k^{(0)}(\tau) + \epsilon \, {\mathbf{P}}_k^{(1)}(\tau) + \epsilon^2 \, {\mathbf{P}}_k^{(2)}(\tau) + \dotsb$$ Substitution of this into Eq.  and equating powers of $\epsilon$ give the results: \[bi.ae.e:PexpandII\] $$\begin{aligned} {\mathbf{P}}_k^{(0)} &= \frac{{\mathbf{K}}_k}{{\omega}} \>, \label{bi.ae.e:PexpA} \\ {\mathbf{P}}_k^{(1)} &= \frac{ \dot{{\mathbf{K}}}_k \times {\mathbf{K}}_k }{ 2 \, {\omega}^3 } \>, \label{bi.ae.e:PexpB} \\ {\mathbf{P}}_k^{(2)} &= \frac{ 3 \, ( \dot{{\mathbf{K}}}_k \cdot {\mathbf{K}}_k ) \, \dot{{\mathbf{K}}}_k - {\omega}^2 \, \ddot{{\mathbf{K}}}_k } { 4 \, {\omega}^5 } + {\mathcal{N}}_k \, {\mathbf{K}}\, \>, \label{bi.ae.e:PexpC}\end{aligned}$$ where ${\omega}= \sqrt{ {\pi_k}^2 + m^2 }$ and $$\label{bi.ae.e:canNdef} {\mathcal{N}}_k = - \frac{1}{8} \, \frac{\dot{{\pi_k}}^2}{{\omega}^5} + \frac{1}{4} \, \frac{{\pi_k}\, \ddot{{\pi_k}}}{{\omega}^5} - \frac{5}{8} \, \frac{{\pi_k}^2 \, \dot{{\pi_k}}^2}{{\omega}^7} \>.$$ We have suppressed the $\tau$ dependence here of these quantities. The dot denotes a partial derivative with respect to $\tau$. Explicitly, we find: $$\begin{aligned} P_{1} &= \frac{m}{{\omega}} + \epsilon^2 \, m \, \Bigl ( \, - \frac{1}{8} \, \frac{{\dot \pi_k}^2}{{\omega}^5}+ \frac{1}{4} \, \frac{ {\pi_k}\, {\ddot \pi_k}}{{\omega}^5} - \frac{5}{8} \, \frac{{\pi_k}^2 \, {\dot \pi_k}^2}{{\omega}^7} \, \Bigr ) + \dotsb \notag \\ P_{2} &= \epsilon \, m \, \frac{{\dot \pi_k}}{2 \, {\omega}^3} + \dotsb \notag \\ P_3 &= \frac{{\pi_k}}{{\omega}} - \epsilon^2 \, m^2 \, \Bigl ( \, \frac{1}{4} \, \frac{ {\ddot \pi_k}}{ {\omega}^5 } - \frac{5}{8} \, \frac{ {\pi_k}\, {\dot \pi_k}^2 }{ {\omega}^7 } \, \Bigr ) + \dotsb \label{bi.ae.e:PxPyPz}\end{aligned}$$ So setting $\epsilon \rightarrow 1$, Maxwell’s equation becomes: $$\label{bi.ae.e:Maxwellad} \dot{E}(\tau) = e {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \biggl [ \frac{{\pi_k}}{{\omega}} - m^2 \, \Bigl ( \frac{1}{4} \, \frac{ {\ddot \pi_k}}{ {\omega}^5 } - \frac{5}{8} \, \frac{ {\pi_k}\, {\dot \pi_k}^2 }{ {\omega}^7 } \Bigr ) \biggr ] + \dotsb$$ All terms odd in ${\pi_k}$ vanish by symmetric integration. After integration, Eq.  becomes: $$\label{bi.ae.e:MaxwelladII} \dot{E}(\tau) = - \frac{ e^2 }{ 6 \pi \, m^2} \, \dot{E}(\tau) + J^\mathrm{sub}(\tau) \>.$$ Here, the first term corresponds to finite charge renormalization in [(1+1)]{} dimensions and can be brought over to the left hand side of the equation. The current $J^\mathrm{sub}(\tau)$ is explicitly finite by power counting and is initially zero. An adiabatic expansion of the Dirac equation can also be carried out from solutions of the second-order form of the Dirac equation. In Section \[ss:twofield\] below, we show that this gives the same result as in Eqs. . Energy-momentum tensor {#s:energymomentum} ====================== In the boost-invariant coordinate system, the average value of the total energy-momentum tensor is given by Eqs. (4.1) and (4.2) of Ref. , and is the sum of two terms: $$\label{bi.em.e:Tmunu} T_{\mu\nu} = T_{\mu\nu}^{\text{matter}} + T_{\mu\nu}^{\text{field}} = {\mathrm{diag}( \, \mathcal{E}, \tau^2 \, \mathcal{P} \, )} \>,$$ where \[bi.em.e:Tmf\] $$\begin{aligned} T_{\mu\nu}^{\text{matter}} &= \frac{1}{4} \bigl \langle {\ensuremath{[ \, \hat{\bar{\psi}}(x) , \tilde{\gamma}_{(\mu}^{\phantom\ast}(x) \, ( i \, D_{\nu)}^{\phantom\ast} \, \hat{\psi}(x) ) \, ]}} + \text{h.c.} \, \bigr \rangle \label{em.e:Tmatter} \\ T_{\mu\nu}^{\text{field}} &= g_{\mu\nu} \, \frac{1}{4} \, F^{\alpha\beta} F_{\alpha\beta} + F_{\mu\alpha} \, g^{\alpha\beta} \,F_{\beta\nu} \>. \label{em.e:Tfield}\end{aligned}$$ Here $D_{\mu} = \partial_{\mu} + i e \, A_{\mu}(x)$ and the subscript notation $(\mu,\nu)$ means to symmetrize the term. From our definitions in Section \[s:Maxwell\] and Eq. , the field part of the energy-momentum tensor is given by: $$\label{bi.em.e:fieldEMtensor} T_{\mu\nu}^{\text{field}} = {\mathrm{diag}( \, E^2/2, - \tau^2 \, E^2/2 \, )} \>.$$ We denote the matter part of the energy-momentum tensor as: $$\label{matter:Tmunu} T_{\mu\nu}^{\text{matter}} = {\mathrm{diag}( \, \varepsilon, \tau^2 \, p \, )} \>.$$ For the matter field, we first note that $D_{\nu} \, \hat{\psi}(x) = S(x) \, \nabla_{\nu} \, \hat{\phi}(x) / \sqrt{\tau}$, where $\nabla_{\nu}$ is the covariant derivative defined below Eq. . For the $T_{\tau\tau}=\varepsilon(\tau)$ component, $\nabla_{0} = \partial_\tau$, and using the field expansion , Eqs. ,, and , we find: $$\begin{aligned} \varepsilon(\tau) &= - \frac{1}{2 \tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda} \lambda \, {\mathrm{Tr} [ \, \rho_k^{(\lambda)}(\tau) \, H(\tau) \, ]} \notag \\ &= - \frac{1}{\tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \, {\mathrm{Tr} [ \, \rho_k^{(+)}(\tau) \, H(\tau) \, ]} \notag \\ &= - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> {\mathbf{K}}({\pi_k}) \cdot {\mathbf{P}}^{(+)}({\pi_k},\tau) \>. \label{bi.em.e:Ttt}\end{aligned}$$ For the $T_{\eta\eta}=\tau^2 p(\tau)$ component, $\nabla_{\eta} = \partial_{\eta} - i e A(\tau) + \gamma^5/2$. Following similar steps to the preceding calculation, we find: $$\begin{aligned} p(\tau) &= - \frac{1}{2\tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda} \lambda \, {\pi_k}(\tau) \, {\mathrm{Tr} [ \, \rho_k^{(\lambda)}(\tau) \, \sigma_3 \, ]} \notag \\ &= - \frac{1}{\tau} {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \, {\pi_k}(\tau) \, {\mathrm{Tr} [ \, \rho_k^{(+)}(\tau) \, \sigma_3 \, ]} \notag \\ &= - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> {\pi_k}\, P_3^{(+)}({\pi_k},\tau) \>. \label{bi.em.e:Tee}\end{aligned}$$ So from , \[bi.em.e:EPvalues\] $$\begin{aligned} \mathcal{E} &= - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> {\mathbf{K}}({\pi_k}) \cdot {\mathbf{P}}^{(+)}({\pi_k},\tau) + \frac{E^2}{2} \>, \label{bi.em.e:Evalue} \\ \mathcal{P} &= - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \> {\pi_k}\, P_3^{(+)}({\pi_k},\tau) - \frac{E^2}{2} \>. \label{bi.em.e:Pvalue}\end{aligned}$$ The covariant derivative of the energy-momentum tensor in boost-invariant coordinates is conserved: $$\label{bi.em.e:consI} T^{\mu\nu}{}_{;\mu} = \partial_{\mu} T^{\mu\nu} + \Gamma^{\mu}_{\mu\sigma} T^{\sigma\nu} + \Gamma^{\nu}_{\mu\sigma} T^{\mu\sigma} = 0 \>.$$ The Christoffel symbols are defined by: $\Gamma^{\lambda}_{\mu\nu}(x) = V^{\lambda}{}_{a}(x) \, ( \partial_{\mu} V^{a}{}_{\nu}(x) )$. In our case, the non-vanishing symbols are given by: $$\label{bi.em.e:Christoffel} \Gamma^{\tau}_{\eta\eta} = \tau \>, \qquad \Gamma^{\eta}_{\tau\eta} = \Gamma^{\eta}_{\eta\tau} = 1 / \tau \>.$$ So we find that $$\label{bi.em.e:ConsI} \partial_\tau \, T^{\tau\tau} + T^{\tau\tau} / \tau + \tau \, T^{\eta\eta} = 0 \>,$$ or $$\label{bi.em.e:ConsII} \partial_\tau \, ( \tau \mathcal{E} ) + \mathcal{P} = 0 \>.$$ Using the equation of motion and Maxwell’s equation , one can show that Eq.  is automatically satisfied. Using Eqs. , the adiabatic expansion for the energy is given by: $$\label{bi.em.e:Eadiab} \mathcal{E} = \biggl ( 1 + \frac{e^2}{6\pi \, m^2} \biggr ) \, \frac{E^2}{2} + \frac{1}{24 \pi \, \tau^2} - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } 2 \, {\omega}+ \dotsb$$ We recognize the first term as a finite renormalization of the charge, the second term as a renormalization of the cosmological constant, and the third term as a sum of the zero point energies of pairs of particles and anti-particles with energy ${\omega}({\pi_k})$. We subtract these terms from the calculation of the energy and arrive at a finite energy $\mathcal{E}^{\text{sub}}$ given by: $$\label{bi.em.e:Esub} \mathcal{E}^{\text{sub}} = \frac{E^2}{2} + {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \, \Bigl [ \, - {\mathbf{K}}({\pi_k}) \cdot {\mathbf{P}}({\pi_k},\tau) + {\omega}- \frac{{\dot \pi_k}^2}{{\omega}^5} \, \Bigr ] \>.$$ For the pressure, the adiabatic expansion gives: $$\label{bi.em.e:Padiab} \mathcal{P} = - \Bigl ( 1 + \frac{e^2}{6\pi \, m^2} \Bigr ) \, \frac{E^2}{2} - \frac{1}{8 \pi \, \tau^2} - {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \frac{2\, {\pi_k}^2 }{{\omega}} + \dotsb$$ Again, the first term renormalizes the charge, the second term in canceled by the cosmological constant term and the third is the usual pressure. We subtract these terms from the pressure to get: $$\begin{aligned} \mathcal{P}^{\text{sub}} = - \frac{E^2}{2} + & {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \, \Bigl [ \, - {\pi_k}\, P_3({\pi_k},\tau) \label{bi.em.e:Psub} \\ \notag & \quad + \frac{{\pi_k}^2}{{\omega}} - m^2 \, \Bigl ( \, \frac{1}{4} \frac{ {\pi_k}\, {\dot \pi_k}^2}{{\omega}^5} - \frac{5}{8} \frac{{\pi_k}^2 \, {\dot \pi_k}^2}{{\omega}^7} \, \Bigr ) \, \Bigr ] \>.\end{aligned}$$ Eqs.  and are now finite.  Proper-time evolution of the electromagnetic fields and current for the one-field and two-field methods described in text. Here we choose $m=1$, $A(\tau_0)=0$ and $E(\tau_0)=4$.](Fig1){width="0.9\columnwidth"} Initial conditions {#s:initialconditions} ================== The simplest choice of initial conditions is to find approximate free-field solutions of Eq.  near $\tau = \tau_0$. This strategy was used in Ref. , and automatically provides a zero current at $\tau=\tau_0$. We call this the “one-field” method, and is discussed in Section \[ss:onefield\] below. In previous studies of the backreaction problem by Cooper *et al.* [@r:Cooper:1993uq] adiabatic initial conditions were used which required averaging over two different solutions to the Dirac equation to obtain an zero current at initial proper time $\tau_0$. We call this the “two-field” method. We discuss this method in Section \[ss:twofield\]. One-field method {#ss:onefield} ---------------- At $\tau = \tau_0 \equiv 1/m$, $A(\tau_0) = 0$ and $H(\tau_0)$ is given by: $$\label{bi.ic.e:Hzero} H(\tau_0) = m \begin{pmatrix} k & 1 \\ 1 & - k \end{pmatrix} \>.$$ So at $\tau \approx \tau_0$, $F_k(\tau)$ obeys the approximate equation of motion: $$\label{bi.ic.e:F0eom} i \, \partial_\tau \, F_{0;k}(\tau) = H(\tau_0) \, F_{0;k}(\tau) \>.$$ Writing $$\label{bi.ic.e:F0e} F_{0;k}(\tau) = \tilde{F}_{0;k} \, e^{-i {\omega}( \tau - \tau_0 )} \>,$$ We find that ${\omega}(\tau_0) = \pm \omega_0$, where $\omega_0 = m \sqrt{k^2 + 1}$. Positive frequency solutions given by: $$\label{bi.ic.e:F0p} \tilde{F}_{0;k}^{(+)} = \sqrt{ \frac{ \omega_0 + m k }{ 2 \omega_0 } } \begin{pmatrix} 1 \\[2pt] \zeta \end{pmatrix} = \begin{pmatrix} \cos ( \theta_k / 2 ) \\[4pt] \sin ( \theta_k / 2 ) \end{pmatrix} \>,$$ and negative frequency solutions by: $$\label{bi.ic.e:F0n} \tilde{F}_{0;k}^{(-)} = \sqrt{ \frac{ \omega_0 + m k }{ 2 \omega_0 } } \begin{pmatrix} - \zeta \\[2pt] 1 \end{pmatrix} = \begin{pmatrix} - \sin ( \theta_k / 2 ) \\[4pt] \cos ( \theta_k / 2 ) \end{pmatrix} \>,$$ with $\zeta = m / (\omega_0 + m k)$. Here $\sin \theta_k = 1 / \sqrt{k^2 + 1}$ and $\cos \theta_k = k / \sqrt{k^2 + 1}$, with $0 \le \theta_k \le \pi$. Density matrices for these solutions are given by: \[bi.ic.e:rho0pm\] $$\begin{aligned} \rho_k^{(+)} &= F_{0;k}^{(+)}(\tau) \, F_{0;k}^{(+)\,\dagger}(\tau) \label{bi.ic.e:rho0p} \\ &= \begin{pmatrix} \cos^2 ( \theta_k / 2 ) & \sin ( \theta_k / 2 ) \, \cos ( \theta_k / 2 ) \\ \sin ( \theta_k / 2 ) \, \cos ( \theta_k / 2 ) & \sin^2 ( \theta_k / 2 ) \end{pmatrix} \>, \notag \\ \rho_k^{(-)} &= F_{0;k}^{(-)}(\tau) \, F_{0;k}^{(-)\,\dagger}(\tau) \label{bi.ic.e:rho0m} \\ &= \begin{pmatrix} \sin^2 ( \theta_k / 2 ) & - \sin ( \theta_k / 2 ) \, \cos ( \theta_k / 2 ) \\ - \sin ( \theta_k / 2 ) \, \cos ( \theta_k / 2 ) & \cos^2 ( \theta_k / 2 ) \end{pmatrix} \>, \notag\end{aligned}$$ and are independent of $\tau$. The corresponding polarization vectors are also independent of $\tau$ and are given by: $$\label{bi.ic.e:Pol0pm} {\mathbf{P}}_{0;k}^{(+)} = \sin \theta_k \, {\hat{\mathbf{e}}}_1 + \cos \theta_k \, {\hat{\mathbf{e}}}_3 = \frac{{\mathbf{K}}_k(\tau_0)}{\omega_0} = - {\mathbf{P}}_{0;k}^{(-)} \>,$$ The initial spinors are orthogonal and complete: \[bi.ic.e:orthogcomp\] $$\begin{aligned} F_{0;k}^{(\lambda)\,\dagger}(\tau) \, F_{0;k}^{(\lambda')}(\tau) &= \delta_{\lambda,\lambda'} \>, \\ \sum_{\lambda=\pm} F_{0;k}^{(\lambda)}(\tau) \, F_{0;k}^{(\lambda)\,\dagger}(\tau) &= 1 \>.\end{aligned}$$ So if we set $F_{k}^{(\lambda)}(\tau_0) = \tilde{F}_{0;k}^{(\lambda)}$ at $\tau = \tau_0$, then the exact solutions remain orthogonal and complete for all $\tau$ and is satisfied. As we have seen in Section \[s:Maxwell\], only the positive energy solutions are needed for the backreaction calculation.  Proper-time evolution of the matter components of the renormalized energy-momentum tensor for the one-field and two-field methods described in text.](Fig2){width="0.9\columnwidth"} The initial spinors can serve to define a particle number operator. Since these initial mode functions form a complete set, we can expand the quantum field in terms of them: $$\label{bi.ic.e:phi0field} \hat{F}_{\alpha}(\tau,\eta) = {\int_{-\infty}^{+\infty} [\mathrm{d} k] \, } \sum_{\lambda} \> \hat{A}_{0;k}^{(\lambda)}(\tau) \> e^{i k \eta } \> F_{0,\alpha;k}^{(\lambda)}(\tau) \>,$$ where $\hat{A}_{0;k}^{(\lambda)}(\tau)$ are mode operators for the $F_{0;\alpha;k}^{(\lambda)}(\tau)$ functions, which now depend on time. Inverting , we find: $$\label{bi.ic.e:A0invert} \hat{A}_{0;k}^{(\lambda)}(\tau) = \int_{-\infty}^{+\infty} {\mathrm{d}}x \, \sum_{\alpha} e^{- i k \eta } \> F_{0,\alpha;k}^{(\lambda)\,\ast}(\tau) \, \hat{F}_{\alpha}^{\phantom{(}}(\tau,\eta) \>,$$ from which we obtain the equal time anti-commutation relation: $$\label{ic.e:A0anticomms} {\ensuremath{\{ \, \hat{A}_{0;k}^{(\lambda)}(\tau) , \hat{A}_{0;k'}^{(\lambda')\,\dagger}(\tau) \, \}}} = ( 2\pi ) \, \delta_{\lambda,\lambda'} \, \delta( k - k' ) \>.$$ Inserting the expansion into the right-hand-side of Eq. , we can relate the $\hat{A}_{0;k}^{(\lambda)}(\tau)$ mode operators to the $\hat{A}_{k}^{(\lambda)}$ mode operators. We find: $$\label{ic.e:A0toA} \hat{A}_{0;k}^{(\lambda)}(\tau) = \sum_{\lambda'} C_{k}^{(\lambda,\lambda')}(\tau) \, \hat{A}_{k}^{(\lambda')} \>,$$ where $$\label{ic.e:Cdef} C_{k}^{(\lambda,\lambda')}(\tau) = F_{0;k}^{(\lambda)\,\dagger}(\tau) \, F_{k}^{(\lambda')}(\tau) \>.$$ Particles are defined in reference to these initial states where a clear distinction between particles and anti-particles can be made. We define an average phase space number density $n_{k}(\tau)$ by: $$\label{ic.e:ndef} n_{k}(\tau) = \frac{ {\mathrm{d}}^2 N(\tau) }{ {\mathrm{d}}k \, {\mathrm{d}}\eta } \>,$$ and is computed using the relation: $$\label{ic.e:ndefI} n_{k}(\tau) \, ( 2\pi ) \, \delta( k - k' ) = {\ensuremath{\langle \, \hat{A}_{0;k}^{(+)\,\dagger}(\tau) \, \hat{A}_{0;k'}^{(+)}(\tau) \, \rangle}} \>.$$ Inserting into , and using $$\label{ic.e:expectAA} {\ensuremath{\langle \, \hat{A}_{k}^{(\lambda)\,\dagger} \, \hat{A}_{k'}^{(\lambda')} \, \rangle}} = \delta_{\lambda,-} \, \delta_{\lambda',-} \, ( 2\pi ) \, \delta( k - k' ) \>,$$ we find: $$\label{ic.e:nresult} \begin{split} n_{k}(\tau) &= | \, C_{k}^{(+,-)}(\tau) \, |^2 = | \, F_{0;k}^{(+)\,\dagger}(\tau) \, F_{k}^{(-)}(\tau) \, |^2 \\ &= 1 - | \, F_{0;k}^{(+)\,\dagger}(\tau) \, F_{k}^{(+)}(\tau) \, |^2 \\ &= 1 - {\mathrm{Tr} [ \, \rho_{0;k}^{(+)} \, \rho_{k}^{(+)}(\tau) \, |^2 \, ]} \\ &= \frac{1}{2} \, \bigl [ \, 1 - {\mathbf{P}}_{0;k}^{(+)} \cdot {\mathbf{P}}_{k}^{(+)}(\tau) \, \bigr ] \>. \end{split}$$ We see immediately that $n_k(\tau_0) = 0$ at $\tau=\tau_0$. We note that in the one-field method the current is automatically zero at $\tau=\tau_0$: Eq.  with $$P_3^{(+)}(\tau_0) = \frac{K_3(\tau_0)}{\omega_0} = \frac{{\pi_k}}{\omega_0} \>,$$ leads to a zero current because the integrand in Eq.  is odd in ${\pi_k}$. Furthermore, one of the subtleties of the one-field method is that the zero-current point is an unstable equilibrium point. This is most easily seen from the equation of motion, Eq. , of the polarization vector. For $\tau = \tau_0$, we find that $$\label{bi.ic.e:dotPzero} \begin{split} \partial_\tau \, {\mathbf{P}}_k^{(+)}(\tau_0) &= 2 \, {\mathbf{K}}_k(\tau_0) \times {\mathbf{P}}_k^{(+)}(\tau_0) \\ &= 2 \, {\mathbf{K}}_k(\tau_0) \times {\mathbf{K}}_k(\tau_0) / \omega_0 = 0 \>. \end{split}$$ However the second derivative is not zero: $$\label{bi.ic.e:ddotPzero} \partial_\tau^{2} \, {\mathbf{P}}_k^{(+)}(\tau_0) = - \frac{ 2 m^2 }{ \sqrt{k^2 + 1} } \, \bigl ( \, k - e E_0 / m^2 \, \bigr ) \, {\hat{\mathbf{e}}}_y \>.$$  Proper-time evolution of the particle density, $\mathrm{d}N/\mathrm{d}\eta$ and proper time evolution of the ratio $\tau \varepsilon(\tau) / [ \mathrm{d}N/\mathrm{d}\eta]$ for the one-field and two-field methods.](Fig3){width="0.9\columnwidth"} {width="70.00000%"} Two field method {#ss:twofield} ---------------- Here we start from solutions of the second-order Dirac equation. Writing the spinor $F_k(\tau)$ in the form: $$\label{bi.tf.e:Ffg} F_k^{(+)}(\tau) = \begin{pmatrix} f_{k,+}^{(+)}(\tau) \\ f_{k,-}^{(+)}(\tau) \end{pmatrix} \>,$$ from Dirac’s Eq. , we can find a second-order equation for either the upper or lower component: $$\label{bi.tf.e:fdeII} \bigl \{ \, \partial_\tau^2 + {\omega}^2(\tau) - i \, s \, {\dot \pi_k}(\tau) \, \bigr \} \, f_{k,s}^{(+)}(\tau) = 0 \>,$$ where $s=\pm 1$ designates the upper or lower component. A parametrization of these mode functions of the form: $$\begin{gathered} \label{bi.tf.e:fparaI} f_{k,s}^{(+)}(\tau) = \frac{ \mathcal{A}_{k,s}^{(+)} } { \sqrt{2 \, \Omega_{k,s}^{(+)}(\tau)} } \, \\ \times \exp \Biggl \{ - i \int_{\tau_0}^{\tau} \Bigl [ \, \Omega_{k,s}^{(+)}(\tau') - s \, \frac{ i {\dot \pi_k}(\tau') } { 2 \, \Omega_{k,s}^{(+)}(\tau') } \, \Bigr ] \, {\mathrm{d}}\tau' \Biggr \} \>,\end{gathered}$$ leads to a second-order nonlinear equation for $\Omega_{k,s}^{(+)}(\tau)$ given by: $$\label{bi.tf.e:nonlinear} \frac{1}{2} \frac{ \ddot{\Omega}_{s} } { \Omega_{s} } - \frac{3}{4} \biggl [ \frac{ \dot{\Omega}_{s} } { \Omega_{s} } \biggr ]^2 \!\!\! + \frac{1}{2} \frac{ s \, {\ddot \pi_k}} { \Omega_{s} } - \frac{1}{4} \biggl [ \frac{ {\dot \pi_k}} { \Omega_{s} } \biggr ]^2 \!\!\! - \frac{ s \,{\dot \pi_k}\, \dot{\Omega}_{s} } { \Omega_{s}^{2} } + \Omega_{s}^{2} = {\omega}^2 \>.$$ Here, and in the following, we suppress the dependencies on $\tau$, $k$, and the positive energy superscript. Solutions of the nonlinear equation for $\Omega_{s}$, subject to initial conditions given below, completely determine $f_{s}$. Once we find $f_{s}$, we can get the other Dirac component from Dirac’s equation: $$\label{bi.tf.e:fms} f_{-s} = \frac{1}{m} \, \bigl ( \, i \, \partial_\tau + s \, {\pi_k}\, \bigr ) \, f_{s} = \frac{ Z_{s}}{ m } \, f_{s} \>,$$ where $Z_{k,s}^{(+)}(\tau)$ is given by: $$\label{bi.tf.e:Zdef} Z_{s} = X_{s} + i \, Y_{s} = \Omega_{s} + s \, {\pi_k}\, - i \ \frac{ \dot{\Omega}_{s} + s \, {\dot \pi_k}\, }{ 2 \, \Omega_{s} } \, \>,$$ The normalization requirement: $ \sum_s|f_{s}|^2 = 1$ means that: $$\label{bi.tf.e:magf2pm} | \, f_{s} \, |^2 = \frac{m^2}{ m^2 + | \, Z_{s} \, |^2 } \>, \quad | \, f_{-s} \, |^2 = \frac{| \, Z_{s} \, |^2} { m^2 + | \, Z_{s} \, |^2 } \>,$$ which fixes the normalization factor $\mathcal{A}_{s}$. It is an easy matter now to get *all* the terms of the density matrix $\rho_{s}$, and we find: \[bi.tf.e:pOmega\] $$\begin{aligned} P_{1;s} &= \frac{ 2 \, X_{s} } { m^2 + | \, Z_{s} \, |^2 } \>, \label{bi.tf.e:pxOmega} \\ P_{2;s} &= \frac{ 2 \, Y_{s} } { m^2 + | \, Z_{s} \, |^2 } \>, \label{bi.tf.e:pyOmega} \\ P_{3;s} &= \frac{ m^2 - | \, Z_{s} \, |^2 } { m^2 + | \, Z_{s} \, |^2 } \>. \label{bi.tf.e:pzOmega}\end{aligned}$$ We are now in a position to carry out an adiabatic expansion of the nonlinear equation . We again count derivatives with respect to $\tau$ by putting: $\partial_\tau \mapsto \epsilon \, \partial_\tau$, and expand $$\label{bi.tf.e:Omegaexp} \Omega_{s} = \Omega^{(0)}_{s} + \epsilon \, \Omega^{(1)}_{s} + \epsilon^2 \, \Omega^{(2)}_{s} + \dotsb$$ Inserting this into and inverting the equation gives $\Omega^{(0)}_{s} = {\omega}$ and $\Omega^{(1)}_{s} = 0$, from which we find: $$\label{bi.tf.e:Omega2} \Omega^{(2)}_{s} = \frac{{\omega}- s \, {\pi_k}}{ 2 \, {\omega}} \, \Bigl [ \, \frac{1}{2} \, \frac{ s \, {\ddot \pi_k}}{ {\omega}^2 } \, + \frac{ {\dot \pi_k}^2 }{ {\omega}^3 } \, - \frac{5}{4} \, \frac{ {\dot \pi_k}^2 }{ {\omega}^4 } \, ( \, {\omega}+ s \, {\pi_k}\, ) \, \Bigr ] \>.$$ From this we find that $$\begin{aligned} Z_{s} &= X_s + i \, Y_s \label{bi.tf.e:Zexp} \\ &= ( \, {\omega}- s \, {\pi_k}\, ) \, \Bigl [ \, 1 + i \epsilon \, \frac{s \, {\dot \pi_k}}{2 \, {\omega}^2 } + \epsilon^2 \, \frac{ \Omega_{s}^{(2)} } { ( \, {\omega}- s \, {\pi_k}\, ) } + \dotsb \Bigr ] \>. \notag\end{aligned}$$ So from our general expressions , it is easy to show that: $$\begin{aligned} P_{1;s} &= \frac{m}{{\omega}} + \epsilon^2 \, m \, \Bigl ( \, - \frac{1}{8} \, \frac{{\dot \pi_k}^2}{{\omega}^5}+ \frac{1}{4} \, \frac{ {\pi_k}\, {\ddot \pi_k}}{{\omega}^5} - \frac{5}{8} \, \frac{{\pi_k}^2 \, {\dot \pi_k}^2}{{\omega}^7} \, \Bigr ) + \dotsb \notag \\ P_{2;s} &= \epsilon \, m \, \frac{s \, {\dot \pi_k}}{2 \, {\omega}^3} + \dotsb \label{eq:Pzs} \\ P_{3;s} &= \frac{s \, {\pi_k}}{{\omega}} - \epsilon^2 \, s \, m^2 \, \Bigl ( \, \frac{1}{4} \, \frac{ {\ddot \pi_k}}{ {\omega}^5 } - \frac{5}{8} \, \frac{ {\pi_k}\, {\dot \pi_k}^2 }{ {\omega}^7 } \, \Bigr ) + \dotsb \notag\end{aligned}$$ For $s=1$, Eqs.  are in agreement with Eqs. . So to second adiabatic order $P_{1;s}$ is independent of $s$, but $P_{2;s}$ and $P_{3;s}$ change sign with $s$. To specify the initial conditions for second-order nonlinear Eq.  at $\tau = \tau_0 = 1/m$ one needs two initial conditions. Since the vacuum state is not unique when particles are being produced, one usually chooses some approximate adiabatic vacuum state of given order as discussed in Ref. . The authors in Ref.  chose the first-order adiabatic conditions as \[bi.tf.e:yuvalinitial\] $$\begin{aligned} \Omega_{k,s}^{(+)}(\tau_0) &= \omega_0 = m \, \sqrt{k^2 + 1} \>, \\ \dot{\Omega}_{k,s}^{(+)}(\tau_0) &= \dot \omega_0 = m^2 \, \frac{ k \, ( \, \tilde{E}_0 - k \, ) } { \sqrt{ k^2 + 1 } } \>,\end{aligned}$$ where $\tilde{E}_0 = e E_0 / m^2$. The initial conditions are independent of $s$. Using Eq. , we obtain that for each value of $s$ this choice of initial conditions at $\tau = \tau_0$ will lead to a non-vanishing current, $J_{s}(\tau_0)$. However, if we average over the two sets of solutions $s= \pm$ and chooses for the Maxwell equation: $$\label{bi.me.e:maxwellfinal2} \partial_\tau E(\tau) = \frac{e}{2} {\int_{-\infty}^{+\infty} [\mathrm{d} {\pi_k}] \, } \left[ P_{3,+}^{(+)}({\pi_k},\tau) + P_{3,-}^{(+)}({\pi_k},\tau )\> \right].$$ then the renormalized Maxwell equation will start with a zero value for the current. Numerical results {#s:results} ================= We have performed numerical calculations for both sets of initial conditions described above. We employed a fourth-order Runge-Kutta method to solve the coupled Dirac equation and backreaction problem. The $k$-momentum variable, which is dimensionless, was discretized on a nonuniform piece-wise momentum grid with a cutoff at $k = \Lambda_k$. We found that a value of $\Lambda_k \approx 200$ was necessary to obtain numerical results insensitive with respect to the cutoff. For the purpose of calculating the subtracted values of the current $J(\tau)$, matter energy $\varepsilon(\tau)$, matter pressure $p(\tau)$, and fermion particle density $\mathrm{d}N(\tau)/\mathrm{d}\eta$, we needed to compute the momentum integrals with respect to the variable, ${\pi_k}$ rather than $k$. The corresponding momentum cutoff in ${\pi_k}$-space was chosen to be 20% greater than $\tau_\mathrm{max} \Lambda_k$ to allow for possible very large values of $A(\tau)$, which is unknown at the beginning of the calculation. The momentum integrals in ${\pi_k}$-space were performed using a Chebyshev integration method with spectral convergence [@r:MM02]. Using the procedure outlined here, we found that approximately 8000 mode functions were necessary to obtain a converged numerical result. The conservation of the energy-momentum tensor, see Eq. , served as a numerical test: we found that the renormalized energy-momentum tensor was conserved within machine precision. For the purpose of this comparison, we took: $m=1$, $e=1$, $\tau_0 = 1/m = 1$, $A(\tau_0)=0$, and $E(\tau_0)=4$. These strong-field initial conditions have been shown to produce sufficient fermion pairs at $\tau = \tau_0$ for plasma oscillations to take place. In Fig. \[fig1\], we show the proper-time evolution of the electromagnetic field, $A(\tau)$, electric field, $E(\tau)$, and current, $J(\tau)$, for the one-field and two-field methods described in text. The components of the matter part of the energy-momentum tensor, $\varepsilon(\tau)$ and $p(\tau)$, for the two simulations are shown in Fig. \[fig2\]. Finally, the proper-time evolution of the particle density, $\mathrm{d}N/\mathrm{d}\eta$, defined in Eq. , is given in Fig. \[fig3\]. For both methods, the ratio $\tau \varepsilon(\tau) / [ \mathrm{d}N/\mathrm{d}\eta]$ is seen to oscillate around the numerical value of 1, consistent with the hydrodynamical picture, as explained in Ref. . We notice that the two sets of solutions are almost identical at short and intermediate times. The two solutions become out of phase at late times due to the slightly different initial conditions. However, in the real problem we expect that interactions between the fermions would eliminate these oscillations. The proper-time evolution of the momentum-dependent particle-density distribution, $n_{{\pi_k}}$, corresponding to the choice of initial conditions in the one-field method, is shown in Fig. \[fig4\]. We note that the centroid of the particle-density distribution oscillates between positive and negative values of ${\pi_k}$. The oscillation of the number density is a result of the current oscillating in sign, the current in momentum space being related to the number density times the velocity of light. This effect is also seen classically when two infinite oppositely charged parallel plates initially a finite distance apart are released and allowed to pass through one another. In that case both the current and electric field oscillate in an analytically derivable manner [@r:CB1989sf]. Results for the case of the two-field method (not shown) are very similar, as it is to be expected from the results depicted in Fig. \[fig3\] (see also Ref. ). Conclusions {#s:conclusions} =========== To conclude, in this paper we report an initial-conditions sensitivity study for the problem of pair production of fermions coupled to a “classical” electromagnetic field with backreaction in [(1+1)]{} boost-invariant coordinates. We discuss two methods of choosing the initial conditions which are consistent with having the fermions in a “vacuum state.” We conclude that the two methods of starting out the calculation produce essentially the same answer. Based on our numerical simulations, there seems to be little reason theoretically or otherwise to use the two-field method discussed previously in Ref. , as it doubles the storage requirements and computational time. This is important for our forthcoming studies of fermion particle production with backreaction in QED and QCD. We emphasize here that in the case of the squared Dirac equation (two-field method) there are two independent solutions of the second-order differential equation for the mode functions, each of which provides a basis for two different fermi fields. In order to make compatible the physical requirement that the initial current is zero with the initial choice that the fermions were initially chosen to be a first-order adiabatic vacuum state, the authors of Ref.  simply *averaged* these two solutions to produce a current which was zero at $\tau = \tau_0$. So doubling the number of fermi fields allows one to produces consistent initial conditions if we define the current by averaging over the two sets of solutions. By staying with the original first-order Dirac equation, in the one-field method we were able to satisfy the initial condition of zero current by choosing a slightly cruder initial state for the fermion fields. This choice, however, reduces by half the size and duration of the calculation. This work was performed in part under the auspices of the United States Department of Energy. The authors would like to thank the Santa Fe Institute for its hospitality during the completion of this work.

--- abstract: 'Given a domain $\Omega \subset \mathbb C^n$, the Lempert function is a functional on the space $Hol ({{\mathbb{D}}},\Omega)$ of analytic disks with values in $\Omega$, depending on a set of poles in $\Omega$. We generalize its definition to the case where poles have multiplicities given by local indicators (in the sense of Rashkovskii) to obtain a function which still dominates the corresponding Green function, behaves relatively well under limits, and is monotonic with respect to the local indicators. In particular, this is an improvement over the previous generalization used by the same authors to find an example of a set of poles in the bidisk so that the (usual) Green and Lempert functions differ.' author: - 'Pascal J. Thomas, Nguyen Van Trao' title: Convergence and multiplicities for the Lempert function --- Introduction {#intro} ============ We assume throughout that $\Omega $ is a bounded domain in ${{\mathbb{C}}}^n$. Let ${{\mathbb{D}}}$ stand for the unit disk in ${{\mathbb{C}}}$. The classical Lempert function with pole at $a \in \Omega $ [@Lempert] is defined by $$\ell_a (z):=\inf \big\{ \log|\zeta|:\exists \varphi\in Hol (\mathbb D,\Omega), \varphi(0)=z, \varphi(\zeta)=a\big\}.$$ Given a finite number of points $a_j \in \Omega$, $j=1,...,N$, Coman [@Coman] extended this to: $$\begin{gathered} \label{coman} \ell (z):=\ell_{\{a_1,\dots,a_N\}} (z):= \inf \big\{ \sum^N_{j=1} \log|\zeta_j|: \\ \exists \varphi\in Hol (\mathbb D,\Omega) : \varphi(0)=z, \varphi(\zeta_j)=a_j, j=1,...,N \big\} .\end{gathered}$$ The *Green function* for the same poles is $$\begin{gathered} g := \sup \left\lbrace u \in PSH(\Omega, \mathbb R_-) : u(z) \le \log |z-a_j|+C_j, \right. \\ \left. \mbox{ for } z \mbox{ in a neighborhood of } a_j, j=1,...,N \right\rbrace ,\end{gathered}$$ where $PSH(\Omega, \mathbb R_-)$ stands for the set of all negative plurisubharmonic functions in $\Omega$. The inequality $g(z)\le \ell(z)$ always holds, and it is known that it can be strict [@CarlWieg], [@TraoTh], [@NikoZwo]. If $\ell$ ever turns out to be plurisubharmonic itself, then $\ell$ must be equal to $g$ [@Coman]. There are natural extensions of the definition of the Green function. In one dimension, considering a finite number of poles in the same location $a$, say $m$ poles, has a natural interpretation in terms of multiplicities: the point mass in the Riesz measure of the Green function is multiplied by $m$. Locally, the Green function behaves like $\log |f|$, where $f$ is a holomorphic function vanishing at $a$ with multiplicity $m$. Lelong and Rashkovskii [@lelongrash], [@Rash] defined a generalized Green function. The function $\log |z|$ was replaced by “local indicators", i.e. circled plurisubharmonic functions $\Psi$ whose Monge Ampère measure $(dd^c \Psi)^n$ is concentrated at the origin, such that whenever $\log |w_j|=c\log |z_j|$ for all $j \in \{1,\dots,n\}$, then $\Psi(w)=c\Psi(z)$. This has the advantage of allowing the consideration of non-isotropic singularities such as $\max(2\log |z_1|, \log |z_2|)$, but the “circled" condition privileges certain coordinate axes, so that the class isn’t invariant under linear changes of variables. We will have to remove this restriction to obtain a class large enough to describe some natural limits. In several complex variables, we would like to know which notion of multiplicity can arise when we take limits of ordinary Green (or Lempert) functions with several poles tending to the same point. This idea was put to use in [@TraoTh] to exhibit an example where a Lempert function with four poles is different from the corresponding Green function. The definition of a generalized Lempert function chosen in [@TraoTh] had some drawbacks — essentially, it was not monotonic with respect to its system of poles (in an appropriate sense) [@TraoTh Proposition 4.3] and did not pass to the limit in some very simple situations [@TTppt Theorem 6.3]. We recall that monotonicity holds when no multiplicities are present, see [@WikstromAMS] and [@TraoTh Proposition 3.1] for the convex case, and the more recent [@NiPfl] for arbitrary domains and weighted Lempert functions, or more generally when a subset of the original set of poles is considered with the same generalized local indicators. In section \[definition\], we successively define a class of indicators, a subclass which is useful to produce “monomial" examples, a notion of multiplicity for values attained by an analytic disk, and a generalization of Coman’s Lempert function to systems of poles with generalized local indicators, different from [@TraoTh]. In section \[main\], we state our two main results: monotonicity, and convergence under certain restrictive (but, we hope, natural) conditions. Further sections are devoted to the proofs of those results. Finally, in Section \[compprev\] we summarize the differences between our new definition and that given in [@TraoTh]. The first named author would like to thank Nikolai Nikolov for stimulating discussions on this topic, and his colleague Anne Bauval from Toulouse for showing him a nice purely combinatorial proof of Lemma \[order\]. Special thanks are due to the referee for his very thorough reading of our paper. Definitions {#definition} =========== [@lelongrash] \[LRindic\] Let $\Psi \in PSH({{\mathbb{D}}}^n)$. We call $\Psi$ a *local indicator* and write $\Psi \in \mathcal I_0$ if 1. $\Psi$ is bounded from above on ${{\mathbb{D}}}^n$; 2. $\Psi$ is circled, i.e. $\Psi (z_1,\dots,z_n)$ depends only on $(|z_1|, \dots, |z_n|)$; 3. for any $c>0$, $\Psi (|z_1|^c, \dots, |z_n|^c) =c \Psi (|z_1|, \dots, |z_n|)$. As a consequence, $(dd^c \Psi)^n = \tau_\Psi \delta_0$ for some $\tau_\Psi \ge 0$. Notice that if $\Psi_1 \in PSH({{\mathbb{D}}}^n)$, $\Psi_2 \in PSH({{\mathbb{D}}}^m)$, and they are both local indicators, then $$\Psi (z,z'):= \max ( \Psi_1(z), \Psi_2(z'))$$ defines a local indicator on ${{\mathbb{D}}}^{n+m}$. We need to remove the restriction to a single coordinate system in Definition \[LRindic\]. \[genindic\] We call $\Psi$ a *generalized local indicator*, and we write $\Psi \in \mathcal I$ if there exists $U$ a neighborhood of $0$, $\Psi_0 \in \mathcal I_0$ and a one-to-one linear map $L$ of ${{\mathbb{C}}}^n$ to itself such that $L(U ) \subset {{\mathbb{D}}}^n$ and $\Psi = \Psi_0 \circ L$. We will concentrate on a class of simple examples. Given two vectors $z, w \in {{\mathbb{C}}}^n$, their standard hermitian product is denoted by $z\cdot \bar w := \sum_j z_j \bar w_j$. We also write $\| z \| := |z \cdot \bar z |^{1/2}.$ \[elemindic\] We say that $\Psi$ is an *elementary local indicator* if there exists a basis $\{v_1, \dots, v_n\}$ of vectors of ${{\mathbb{C}}}^n$ and scalars $m_j \in {{\mathbb{R}}}_+$, $1\le j \le n$, such that for $z\in {{\mathbb{D}}}^n$, $$\label{elindform} \Psi(z) = \max_ {1\le j \le n} m_j \log |z \cdot \bar v_j | .$$ One easily checks that any elementary local indicator is a generalized local indicator. The most interesting case is the one for which the basis is orthornormal. In fact, it is essentially the only case. \[ortho\] Given an elementary local indicator $\Psi$ as in Definition \[elemindic\] there exists an orthonormal basis $\{\tilde v_1, \dots, \tilde v_n\}$ of ${{\mathbb{C}}}^n$ such that the associated elementary local indicator $\tilde \Psi (z) := \max_ {1\le j \le n} m_j \log |z \cdot \overline {{\tilde v}_j} |$ verifies $\tilde \Psi - \Psi \in L^\infty({{\mathbb{D}}}^n)$. As a consequence, we could have restricted the map $L$ in Definition \[genindic\] to be unitary, and it would not have changed things in any essential way. The proof of Lemma \[ortho\] is given in Section \[easypf\] below. [@lelongrash example in Section 3], [@Rash] \[mass\] If $\Psi$ is an elementary local indicator, then $\tau_\Psi = m_1 \cdots m_n.$ We take the same definition of the generalized Green function as in [@lelongrash]. \[defgreen\] Let $\Omega $ be a bounded domain in ${{\mathbb{C}}}^n$. Given $$S := \{ (a_j,\Psi_j), 1\le j \le N\}, \mbox{ where } a_j \in \Omega, a_j \neq a_k \mbox{ for }j \neq k, \Psi_j \in \mathcal I,$$ its *Green function* is $$\begin{gathered} G_S := \sup \left\lbrace u \in PSH(\Omega, \mathbb R_-) : u(z) \le \Psi_j (z) + C_j, \right. \\ \left. \mbox{ for } z \mbox{ in a neighborhood of } a_j, j=1,...,N \right\rbrace .\end{gathered}$$ To generalize the Lempert function, the first step is to quantify the way in which an analytic disk, i.e. an element of $Hol ({{\mathbb{D}}},\Omega)$, meets a pole provided with a generalized local indicator. \[multphia\] Let $\alpha \in {{\mathbb{D}}}$, $a \in \Omega$, $\Psi \in \mathcal I$. Then the *multiplicity* of $\varphi \in Hol ({{\mathbb{D}}},\Omega)$ at $\alpha$, with respect to $a$, is given by $$\begin{aligned} \mbox{If } \varphi (\alpha) =a, & \mbox{ then } & m_{\varphi, a, \Psi} (\alpha) := \min \left( \tau_\Psi , \liminf_{\zeta \to 0} \frac{\Psi(\varphi(\alpha+\zeta)-a)}{\log |\zeta|} \right); \\ \mbox{if } \varphi(\alpha) \neq a , &\mbox{ then }& m_{\varphi, a, \Psi} (\alpha) :=0.\end{aligned}$$ Notice that if $\Psi_1-\Psi_2$ is locally bounded near the origin, then $m_{\varphi, a, \Psi_1} (\alpha) =m_{\varphi, a, \Psi_2} (\alpha)$. The quantity $\liminf_{\zeta \to 0} \frac{\Psi(\varphi(\alpha+\zeta)-a)}{\log |\zeta|}$ is exactly the Lelong number at $0$ of the subharmonic function $\Psi \circ \varphi$, compare with [@RashSig pp. 334–335]. Truncating at the level of the local Monge-Ampère mass $\tau_\Psi$ will turn out to be convenient in Definition \[defLempert\], and the proofs that use it. It is useful to see what this means in the case of elementary local indicators. [**Elementary examples.**]{} Suppose that $\alpha=0$, $a=0$, and that $\Psi(z) = \max_ {1\le j \le n} m_j \log |z_j | .$ We write $$\varphi(\zeta) = (\varphi_1(\zeta), \dots , \varphi_n(\zeta)),$$ and define the valuations $$\nu_j := \nu_j (0,\varphi) := \min \{ k : (\frac{d}{d\zeta})^k \varphi_j(0) \neq 0 \}.$$ Then we have $$\label{multex} m_{\varphi, 0, \Psi} (0) = \min \left( \min_ {1\le j \le n} m_j \nu_j , \prod_{j=1}^n m_j \right).$$ [**Example 1.**]{} If $m_j=1$ for all $j$, $m_{\varphi, 0, \Psi} (0)=1$ if $\varphi(0)=0$, $m_{\varphi, 0, \Psi} (0)=0$ otherwise. This is the basic case where one just records whether a point has been hit by the analytic disk or not. [**Example 2.**]{} In more general cases, the use of an elementary local indicator will impose higher-order differential conditions on the map $\varphi$. For instance, if $m_1=2$ and $m_j=1$, $2\le j \le n$, then $$\begin{aligned} m_{\varphi, 0, \Psi} (0)& =& 0 \mbox{ if } \varphi(0) \neq 0; \\ m_{\varphi, 0, \Psi} (0)& =& 1 \mbox{ if } \varphi(0) = 0 \mbox{ and } \varphi_j'(0) \neq 0 \mbox{ for some }j \in \{2,\dots, n\}; \\ m_{\varphi, 0, \Psi} (0)& =& 2 \mbox{ if } \varphi(0) = 0 \mbox{ and } \varphi_j'(0) = 0\mbox{ for any }j \in \{2,\dots, n\}.\end{aligned}$$ \[defLempert\] Given a system $S$ as in Definition \[defgreen\], we write $\tau_j := \tau_{\Psi_j}$. Let $\varphi \in Hol({{\mathbb{D}}},\Omega)$ and $A_j \subset {{\mathbb{D}}}$, $1\le j \le N$. We say that $(\varphi, (A_j)_{1\le j \le N})$ is *admissible (for $S$, $z$)* if $$\varphi (0)=z; \quad A_j \subset \varphi^{-1}(a_j) \mbox{ and } \sum_{\alpha \in A_j} m_{\varphi, a_j, \Psi_j} (\alpha) \le \tau_j , 1\le j \le N.$$ In this case, we write (with the convention that $0 \cdot \infty =0$) $$\mathcal S (\varphi, (A_j)_{1\le j \le N}):= \sum_{j=1}^N \sum_{\alpha \in A_j} m_{\varphi, a_j, \Psi_j} (\alpha) \log |\alpha| .$$ Then the generalized Lempert function is defined by $$\begin{gathered} \mathcal L^\Omega_S (z):= \mathcal L_S (z) \\ := \inf \left\lbrace \mathcal S (\varphi, (A_j)_{1\le j \le N}) : (\varphi, (A_j)_{1\le j \le N}) \mbox{ is admissible for }S, z \right\rbrace .\end{gathered}$$ Notice that we allow any of the $A_j$ to be the empty set (in which case the $j$-th term drops from the sum). Consider the *single poles case* where $$\label{sglp} \mbox{ for each }j, \quad \Psi_j (z) = \max_ {1\le l \le n} \log |z_l | ,\mbox{ or }\Psi_j (z) = \log \|z \|$$ – it is the same, since both functions differ by a bounded term near $0$; in fact, one could use any norm that is homogenous under complex scalar multiplication. In this case, $\tau_j=1$ for every $j$. With a slight abuse of notation, we write $S = \{ a_1, \dots , a_N\}.$ Then $\mathcal L_S (z)= \min_{S'\subset S} \ell_{S'}(z)$, where $\ell_S$ is defined in . And in fact $\min_{S'\subset S} \ell_{S'}(z) = \ell_{S}(z)$ [@NiPfl] (see also [@Wikstrom], [@WikstromAMS] for the case when the domain $\Omega$ is convex). The Lempert function is different from the functionals considered by Poletsky and others in that it is restricted to one pre-image per pole $a_j$ (thus the Lempert function can fail to be equal to the corresponding Green function). In our definition, the number of pre-images per pole is bounded above by the Monge-Ampère mass at that pole of its generalized local indicator. In [@TraoTh], each pole only could have one pre-image, but (essentially) $\varphi$ had to hit the pole with maximum multiplicity at that pre-image. Although Definition \[defLempert\] may seem contrived, it is required to obtain the reasonable convergence theorem \[conv\]. See the discussion in Section \[compprev\]. We remark right away that the usual relationship holds between this generalized Lempert function and the corresponding Green function. \[inegGL\] For $\Omega $ a bounded domain, for any system $S$ as in Definition \[defgreen\], for any $z \in \Omega$, $G_S (z) \le \mathcal L_S (z)$. If $\varphi \in Hol({{\mathbb{D}}},\Omega)$, and $u \in PSH_- (\Omega)$ is a member of the defining family for the Green function of $S$, then $u \circ \varphi$ is subharmonic and negative on ${{\mathbb{D}}}$. Furthermore, if $(\varphi, (A_j)_{1\le j \le N})$ is admissible (for $S$, $z$) and $\alpha \in A_j$, then given any ${\varepsilon}>0$, for $|\zeta|$ small enough, $$u \circ \varphi (\alpha +\zeta ) \le C_j + \Psi_j (\varphi (\alpha +\zeta )-a_j) \le C_j + (m_{\varphi, a_j, \Psi_j} (\alpha) - {\varepsilon}) \log |\zeta| .$$ So $u \circ \varphi$ is a member of the defining family for the Green function on ${{\mathbb{D}}}$ with poles $\alpha$ and weights $m_{\varphi, a_j, \Psi_j} (\alpha) - {\varepsilon}$ at $\alpha$. This implies that $$u \circ \varphi (\zeta) \le \sum_{j=1}^N \sum_{\alpha \in A_j} (m_{\varphi, a_j, \Psi_j} (\alpha) - {\varepsilon}) \log \left| \frac{\alpha -\zeta}{1-\zeta \bar \alpha} \right| .$$ Letting ${\varepsilon}$ tend to $0$ and setting $\zeta=0$, we get $u (z) \le \mathcal S (\varphi, (A_j)_{1\le j \le N})$. Passing to the supremum over $u$, then to the infimum over $(\varphi, (A_j)_{1\le j \le N})$, we get the Lemma. Main Results {#main} ============ We start with a remark. \[smallerS\] If $S$ is as in Definition \[defgreen\], $1\le N'\le N$, and $$S':= \{ (a_j,\Psi_j), 1\le j \le N'\},$$ then for any $z \in \Omega$, $\mathcal L_{S'} (z) \ge \mathcal L_{S} (z)$. If we take $A_j=\emptyset$ for $N'+1 \le j \le N$, any member of the defining family for $\mathcal L_{S'} (z)$ becomes a member of the defining family for $\mathcal L_{S} (z)$, and the sum remains the same. The above lemma goes in the direction of monotonicity of the Lempert function with respect to its system of poles. For the Green function, it is immediate that the more poles there are, the more negative the function must be. More generally the more negative the generalized local indicators are (removing a pole corresponds to replacing a local indicator by $0$), the more negative the function must be. This is not immediately apparent in Definition \[defLempert\], but it does hold for elementary local indicators. \[monotone\] Let $\Omega$ be a bounded domain in ${{\mathbb{C}}}^n$, $$S := \{ (a_j,\Psi_j), 1\le j \le N\}, S' := \{ (a_j,\Psi'_j), 1\le j \le N\}, \mbox{ where } a_j \in \Omega,$$ and $\Psi_j$, $\Psi'_j$, are elementary local indicators such that $\Psi_j \le \Psi'_j + C_j$ in a neighborhood of $0$, $C_j \in {{\mathbb{R}}}$, $1\le j \le N$. Then $\mathcal L_{S'} (z) \ge \mathcal L_{S} (z)$, for all $z \in \Omega$. The proof is given in Section \[pfmono\]. Now we turn to a result about the convergence of some families of (ordinary) Lempert functions with single poles, whose limits can be described naturally as generalized Lempert functions. Note that the proof of this next theorem doesn’t require the relatively difficult Theorem \[monotone\], only the easy Lemma \[smallerS\]. For $z\in {{\mathbb{C}}}^n \setminus \{0\}$, we denote by $[z]$ the equivalence class of $z$ in the complex projective space $\mathbb P^{n-1}$. \[conv\] Let $\Omega$ be a bounded and convex domain in ${{\mathbb{C}}}^n$. Let $0\le M \le N$ be integers. For ${\varepsilon}$ belonging to a neighborhood of $0$ in ${{\mathbb{C}}}$, using the simplified notation of the single pole case , let $$S({\varepsilon}) := \left\lbrace a_j({\varepsilon}), 1\le j\le M ; a'_j({\varepsilon}), a''_j({\varepsilon}), M+1\le j\le N \right\rbrace \subset \Omega.$$ Suppose that all the points of $S({\varepsilon})$ are distinct for any fixed ${\varepsilon}$, that $$\lim_{{\varepsilon}\to 0} a_j({\varepsilon}) = a_j \in \Omega, 1\le j\le M ;$$ $$\lim_{{\varepsilon}\to 0} a'_j({\varepsilon}) = \lim_{{\varepsilon}\to 0} a''_j({\varepsilon}) =a_j \in \Omega, M+1\le j\le N ;$$ and that $$\label{limproj} \lim_{{\varepsilon}\to 0} [a''_j({\varepsilon})-a'_j({\varepsilon})] = [v_j],$$ where the limit is with respect to the distance in $\mathbb P^{n-1}$ and the representative $v_j$ is chosen of unit norm. Let $\Psi_j (z) := \log \| z \|$, $1\le j\le M$. Denote by $\pi_j$ the orthogonal projection onto $\{v_j\}^\bot$, $M+1\le j\le N $, and by $\Psi_j$ the generalized local indicator $$\Psi_j (z) := \max (\log \| \pi_j(z)\|, 2 \log |z\cdot \bar v_j|), \quad M+1\le j\le N.$$ Set $S:=\{(a_j,\Psi_j), 1 \le j \le N\}$. Then $$\lim_{{\varepsilon}\to 0} \ell_{S({\varepsilon})} (z) = \lim_{{\varepsilon}\to 0} \mathcal L_{S({\varepsilon})} (z) = \mathcal L_{S} (z) \mbox{ for all } z \in \Omega.$$ Remarks : (a) as in the comments after , one could replace $\Psi_j$ by an elementary local indicator; (b) the convexity requirement is imposed by Lemma \[relax\], and we conjecture that it is not essential. Note that in the case where $a'_j({\varepsilon})=a_j$ does not depend on ${\varepsilon}$, the hypothesis means that the point $a''_j({\varepsilon})$ converges to a limit in the blow-up of ${{\mathbb{C}}}^n$ around the point $a_j$. It seems to us that this is the only reasonable convergence result that can be obtained for a family of ordinary Lempert functions. If is not satisfied, one can find two distinct limit points for our family of Lempert functions. Thus hypothesis is required. We are restricting ourselves to the case where no more than two points converge to the same point: examples where three points converge to the origin in the bidisk are explicitly studied in [@Th3p], and show that the situation leads to results that probably can’t be described in terms of our generalized local indicators. The proof is given in Section \[pfconv\]. Proof of Lemma \[ortho\] {#easypf} ======================== Multiplying one of the vectors $v_j$ by a scalar only modifies the function $\Psi$ by a bounded additive term, so it will be enough to exhibit an orthogonal basis of vectors complying with the conclusion of the Lemma. Renumber the vectors $v_j$ so that we have $0\le m_1 \le \cdots \le m_n$. Using the Gram-Schmidt orthogonalization process, we produce an orthogonal system of vectors $\tilde v_k$ such that $\mbox{Span} (\tilde v_1, \dots, \tilde v_k) = \mbox{Span} ( v_1, \dots, v_k)$ for any $k$, $1\le k \le n$. We proceed by induction on the dimension $n$. When $n=1$ the property is immediate. Assume that the result holds up to dimension $n-1$. Write $$\Psi_1(z) := \max_ {1\le j \le n-1} m_j \log |z \cdot \bar v_j | , \quad \tilde \Psi_1 (z) := \max_ {1\le j \le n-1} m_j \log |z \cdot \overline {{\tilde v}_j} |.$$ Denote $z_n:= z \cdot \overline {{\tilde v}_n} $. It is enough to obtain the estimates on a neighborhood $U$ of $0$. We choose it so that for $z \in U$, $|z_n|\le 0$, $ \Psi_1 (z), \tilde \Psi_1 (z) \le 0$. Since $v_n=\tilde v_n - w$, where $w \in \mbox{Span}( v_1, \dots, v_{n-1})$, we have $$\begin{gathered} \label{psiw} \Psi (z) = \max ( \Psi_1 (z'), m_n \log |z_n-z'\cdot \bar w |), \\ \tilde \Psi (z) = \max ( \tilde \Psi_1 (z'), m_n \log |z_n |),\end{gathered}$$ where $z'$ stands for the orthogonal projection of $z$ on $\mbox{Span}( v_1, \dots, v_{n-1})$ $=\mbox{Span}( \tilde v_1, \dots, \tilde v_{n-1})$. By the induction hypothesis, $ \Psi_1=\tilde \Psi_1 + O(1),$ so it is enough to prove that $$\Psi'(z):= \max ( \Psi_1 (z'), m_n \log |z_n |)$$ differs from $\Psi(z)$ by a bounded additive term. There is a constant $C_0>0$ such that $\Psi_1 (z') \ge m_{n-1} \log \| z' \| - \log C_0$, for $z'\in U$. Choose a constant $A>1$ large enough so that $ \|w\| (C_0/A)^{1/ m_{n-1}} < 1/2$. Then, since $ \Psi_1 (z) \le 0$ and $m_{n-1}\le m_{n}$, $$\begin{gathered} \label{zw} |z'\cdot \bar w | \le \|w\| C_0^{1/ m_{n-1}} \exp( \frac{\Psi_1 (z') }{m_{n-1}}) \\ \le \|w\| C_0^{1/ m_{n-1}} \exp( \frac{\Psi_1 (z') }{m_{n}}).\end{gathered}$$ [*Case 1.*]{} $\Psi_1 (z') \ge m_n \log |z_n| - \log A.$ By the inequality above, $\Psi'(z) \le \Psi_1 (z') + \log A \le \Psi (z) + \log A$. On the other hand, using , we get $$|z_n-z'\cdot \bar w |^{m_{n}} \le \left(A^{1/m_{n-1}} + \|w\| C_0^{1/m_{n-1}} \right)^{m_{n}} \exp( \Psi_1 (z') ),$$ so $\Psi(z) \le \Psi_1 (z') +O(1) \le \Psi' (z) +O(1)$. [*Case 2.*]{} $\Psi_1 (z') \le m_n \log |z_n| - \log A.$ Then and the choice of $A$ imply $$|z'\cdot \bar w | \le \|w\| C_0^{1/ m_{n-1}} \exp\left( \log |z_n| -\frac{ \log A}{m_{n-1}} \right) \le \frac12 |z_n|,$$ thus implies that $$\Psi'(z) +\log \frac12 \le \Psi(z) \le \Psi'(z) +\log \frac32.$$ Proof of Theorem \[monotone\] {#pfmono} ============================= Without loss of generality, we may assume that $\tau'_j>0$ for all $j$. We have $\Psi_j \le \Psi'_j + C_j$ in a neighborhood of $0$ and $$\mbox{supp}\, (dd^c \Psi_j)^n \subset \{ 0 \}, \quad \mbox{supp}\, (dd^c \Psi'_j)^n \subset \{ 0 \}.$$ Thus it follows from Bedford and Taylor’s comparison theorem [@BT], [@Kli p. 126, Theorem 3.7.1] that $\tau_j \ge \tau'_j>0$. For any $\alpha, a_j$, $$\label{mgtm} m_{\varphi, a_j, \Psi_j} (\alpha) \ge m_{\varphi, a_j, \Psi'_j} (\alpha).$$ Therefore $$\label{comp_sums} \sum_{j=1}^N \sum_{\alpha \in A_j} m_{\varphi, a_j, \Psi_j} (\alpha) \log |\alpha| \le \sum_{j=1}^N \sum_{\alpha \in A_j} m_{\varphi, a_j, \Psi'_j} (\alpha) \log |\alpha| .$$ To finish the proof, it suffices to show that the family over which we take the infimum is smaller for $\mathcal L_{S'} (z)$ than the one for $\mathcal L_{S} (z)$. This can be checked for each $j$ separately, hence we drop the index $j$. \[comp\_sets\] Let $\Omega$ be a bounded domain in ${{\mathbb{C}}}^n$. If $\Psi, \Psi'$ are elementary local indicators such that $\Psi \le \Psi' + C$ and $\tau':=\tau_{\Psi'}>0$, if $A\subset {{\mathbb{D}}}$, $a\in \Omega$ and $\varphi \in Hol({{\mathbb{D}}},\Omega)$ verify $$\sum_{\alpha \in A} m_{\varphi, a, \Psi'} (\alpha) \le \tau',$$ then $$\sum_{\alpha \in A} m_{\varphi, a, \Psi} (\alpha) \le \tau:=\tau_{\Psi}.$$ Since the point $a$ plays no role, we assume $a=0$ and write $m_{\varphi, 0, \Psi} (\alpha)= m_{\varphi, \Psi} (\alpha)$. By , we may assume that $ m_{\varphi, \Psi} (\alpha) >0$ and the sums in will not change. Using Lemma \[ortho\], we reduce ourselves to the case where the elementary local indicators are given by orthonormal systems of vectors. We use the same “valuations" as in the Elementary Example: $$\nu_j (\alpha) := \nu_j (\alpha ,\varphi) := \min \{ k : (\frac{d}{d\zeta})^k (\varphi (\zeta) \cdot \bar v_j ) (\alpha )\neq 0 \},$$ and $\nu'_j (\alpha)$ is defined analogously using the vectors $v'_j$. [**Case 1.**]{} There exists $\alpha_0$ such that $m_{\varphi, \Psi'} (\alpha_0) = \tau'$. Then the hypothesis of Lemma \[comp\_sets\] implies that for all $\alpha \in A\setminus \{\alpha_0\}$, $m_{\varphi, \Psi'} (\alpha) =0$, so that $\min_{1\le k \le n} m'_k \nu'_k(\alpha)=0$. Since $\tau'>0$, we have $m'_k>0$ for all $k$, so there must exist $k$ such that $\nu'_k(\alpha)=0$. Then $\varphi(\alpha)\neq 0$, which implies that $m_{\varphi, \Psi} (\alpha) =0$, and $$\sum_{\alpha \in A} m_{\varphi, \Psi} (\alpha) = m_{\varphi, \Psi} (\alpha_0) \le \tau,$$ by definition of the multiplicity. [**Case 2.**]{} For all $\alpha \in A$, $m_{\varphi, \Psi'} (\alpha) < \tau'$. Therefore $m_{\varphi, \Psi'} (\alpha) = \min_{1\le k \le n} m'_k \nu'_k(\alpha)$, and since we always have $m_{\varphi, \Psi} (\alpha) \le \min_{1\le k \le n} m_k \nu_k(\alpha)$, it becomes enough to work with those quantities in . By dividing by $\tau$ and $\tau'$ respectively, it will be enough to prove the following Lemma. \[dirmult\] Under the hypotheses of Lemma \[comp\_sets\] and Case 2 above, for each $\alpha \in A$, $$\frac{\min_{1\le k \le n} m'_k \nu'_k(\alpha)}{\prod_{k=1}^n m'_k} \ge \frac{\min_{1\le k \le n} m_k \nu_k(\alpha)}{\prod_{k=1}^n m_k} .$$ Since we are now dealing with a single $\alpha$, we also drop it from the notation. We introduce a binary relation on the index set $\{1,\dots,n\}$. Given $k, l \in \{1,\dots,n\}$, we say that $k \mathcal R l$ if and only if $v_k \cdot \bar v'_l \neq 0$. \[relmult\] If $\Psi' + C \ge \Psi$ and $k \mathcal R l$, then $m_k \ge m'_l$. For any nonzero $\lambda \in \mathbb C$, $$\Psi'(\lambda v'_l)= m'_l \log |\lambda| + m'_l \log \|v'_l\|^2,$$ while, for $|\lambda|$ small enough, $$\Psi(\lambda v'_l)= \max_{1\le j \le n} \left( m_j (\log |\lambda| + \log |v_j \cdot \bar v'_l |) \right) =(\min_{k: k \mathcal R l} m_k) \log |\lambda| + O(1),$$ therefore by letting $\lambda$ tend to $0$ we see that $\min_{k: k \mathcal R l} m_k \ge m'_l$. \[relval\] If $\Psi' + C \ge \Psi$, then 1. $\nu'_l \ge \min \{ \nu_k : k \mathcal R l \},$ 2. $\nu_k \ge \min \{ \nu'_l : k \mathcal R l \}.$ We will use and prove part (1) only. The other one has a similar proof. Since $v'_l$ is orthogonal to $v_k$ unless $k \mathcal R l$, we must have complex scalars $c_k$ such that $v'_l =\sum_{k : k \mathcal R l} c_k v_k$, thus for $\varphi $ as in Lemma \[comp\_sets\], $$\varphi(\zeta) \cdot \bar v'_l = \sum_{k : k \mathcal R l} \bar c_k \varphi(\zeta) \cdot \bar v_k .$$ Now take $m < \nu_k = \nu_k(\alpha, \varphi)$, for all $k$ such that $k \mathcal R l$. Then $$(\frac{d}{d\zeta})^m (\varphi \cdot \bar v'_l ) (\alpha ) = \sum_{k : k \mathcal R l} \bar c_k (\frac{d}{d\zeta})^m (\varphi (\zeta) \cdot \bar v_k ) (\alpha) = 0,$$ so we must have $\nu'_l > m$, which proves the result. Now renumber the vectors $v'_l$ so that $\min_k (m'_k \nu'_k) = m'_1 \nu'_1$. Pick an index $k_0$ such that $k_0 \mathcal R 1$ and $\nu_{k_0} = \min \{ \nu_k : k \mathcal R 1 \}$. By renumbering the vectors $v_k$, we may assume $k_0=1$. By Lemma \[relval\], we may assume $\nu'_1 \ge \nu_1$. The conclusion of Lemma \[dirmult\] thus reduces to: $$\label{ineqfin} \frac{ \nu'_1}{\prod_{k=2}^n m'_k} \ge \frac{\nu_1}{\prod_{k=2}^n m_k} .$$ This is a consequence of the next result. \[order\] There exists a bijection $\sigma$ from $\{2,\dots,n\}$ onto itself such that for any $l \in \{2,\dots,n\}$, $ \sigma(l)\mathcal R l $. This Lemma will be proved below. It implies $$\prod_{k=2}^n m_k = \prod_{l=2}^n m_{\sigma(l)} \ge \prod_{l=2}^n m'_l,$$ by Lemma \[relmult\], so holds and this concludes the proof of Lemma \[dirmult\]. [*Proof of Lemma \[order\]* ]{} Denote $A:=(a_{kl})_{2\le k,l \le n} := (v_k \cdot \overline{v'_l})_{2\le k,l \le n}$. First we prove that this matrix is non singular. Let $\pi$ be the orthogonal projection on $\{v'_1\}^\bot$. If $\mbox{rank}\, \{\pi(v_k), 2 \le k \le n\} < n-1$, there exists $w \in \{v'_1\}^\bot$, $w\neq 0$, such that $w \bot \pi(v_k)$, $2\le k \le n$. This implies $w \bot v_k$, $2\le k \le n$. Since we have orthogonal bases, $v_1= \lambda w$, for some $\lambda \in {{\mathbb{C}}}$. So $v_1 \cdot \overline{v'_1}=0$, which contradicts the fact that $1\mathcal R 1$. We construct the bijection $\sigma$ by induction on $n$. For $n=2$ it’s obvious. Suppose that the property holds for $n-1$. Then $$0\neq \mbox{\rm det} A = \sum_{k=2}^n (-1)^k a_{k2} \mbox{\rm det} A_k,$$ where $A_k$ stands for the minor matrix with the first column and the $k$-th row removed. There must be some $k$ for which $a_{k2} \mbox{\rm det} A_k \neq 0$. Let $\sigma(2)=k$; the induction hypothesis gives us a bijection $\sigma'$ from $\{3,\dots,n\}$ to $\{2,\dots,n\}\setminus \{k\}$ such that $a_{\sigma'(l)l}\neq 0$, and this finishes the proof. Proof of Theorem \[conv\] {#pfconv} ========================= First observe that we can relax the conditions used in Definition \[defLempert\]. \[relax\] Let $\Omega$ be a convex bounded domain in $\mathbb C^n$ containing the origin, and let $z\in \Omega$. \(i) Let $a_j \in \Omega$, $\Psi_j \in \mathcal I$ and, as in Definition \[defgreen\] $$S := \left\lbrace (a_j, \Psi_j), 1\le j\le N \right\rbrace .$$ Suppose that for any $\delta >0$, there exists a map $\varphi^\delta $ holomorphic from ${{\mathbb{D}}}$ to $(1+\delta)\Omega$ and sets $(A_j(\delta))_{1\le j \le N}$ such that $(\varphi^\delta, (A_j(\delta))_{1\le j \le N})$ is admissible for $ S,z$ with respect to $(1+\delta)\Omega$ and $$\mathcal S (\varphi^\delta, (A_j(\delta))_{1\le j \le N}) \le \ell + h(\delta),$$ where $ h(\delta)\ge 0$, $\lim_{\delta \to 0} h(\delta)= 0$. Then $\mathcal L^\Omega_S (z) \le \ell.$ \(ii) For ${\varepsilon}$ in a neighborhood $V$ of $0$ in ${{\mathbb{C}}}$, let $a_j({\varepsilon}) \in \Omega$, $1\le j\le N$. $$S({\varepsilon}) := \left\lbrace (a_j({\varepsilon}), \Psi_j), 1\le j\le N \right\rbrace.$$ Let $g: V \longrightarrow {{\mathbb{R}}}_+^*$ be such that $\lim_{{\varepsilon}\to0}g({\varepsilon})=0$. Then $$\limsup_{{\varepsilon}\to0} \mathcal L^\Omega_{S({\varepsilon})} (z) \le \limsup_{{\varepsilon}\to0} \mathcal L^{(1+g({\varepsilon}))\Omega}_{S({\varepsilon})} (z).$$ Without loss of generality, we may assume $z=0$. Let $$\Omega_r := \left\{ \varphi (\zeta) : \varphi \in Hol({{\mathbb{D}}}, \Omega), \varphi(0)=0, |\zeta|<r \right\}.$$ A bounded convex domain is Kobayashi complete hyperbolic [@Dineen Proposition 6.9 (b), p. 88], so $\Omega_r $ is relatively compact in $\Omega $. Let $\rho_\Omega$ stand for the Minkowski function of $\Omega$: $$\rho_\Omega(z):= \inf \{ r>0 : \frac{z}r \in \Omega\}.$$ We set $\gamma_\Omega (r) := \sup_{\Omega_r } \rho_\Omega.$ The function $\gamma$ is increasing and continuous from $(0,1)$ to itself. For any $\mu \in (0,1)$, $\phi \in Hol({{\mathbb{D}}}, {{\mathbb{C}}}^n)$, denote $\phi_\mu (\zeta) := \phi (\mu\zeta)$. Note that for any points and generalized local indicators, $m_{\phi_\mu, a, \Psi} (\alpha/\mu) = m_{\phi, a, \Psi} (\alpha)$. Take $\varphi^\delta$ as in Part (i) of the Lemma, in particular $\varphi^\delta(0)=0$, so by construction of $\gamma$, $$\frac1{(1+\delta)}\varphi_\mu^\delta ({{\mathbb{D}}}) \subset \gamma(\mu) \Omega .$$ Choose some $\mu(\delta)$ such that $ \gamma(\mu(\delta)) = (1+\delta)^{-1}$, and set $\tilde \varphi^\delta := \varphi_{\mu(\delta)}^\delta $, then $\tilde \varphi^\delta \in Hol({{\mathbb{D}}}, \Omega)$. Note that $\lim_{\delta\to 0} \mu(\delta)=1$, by the relative compactness of each $\Omega_r$. Let $$\tilde A_j (\delta) := \left\lbrace \frac{\alpha}{\mu(\delta)}: \alpha \in A_j (\delta), |\alpha|< \mu(\delta) \right\rbrace .$$ Then $$\begin{gathered} \label{contract} \left| \mathcal S (\tilde \varphi^\delta, (\tilde A_j(\delta))_{1\le j \le N}) - \mathcal S (\varphi^\delta, (A_j(\delta))_{1\le j \le N}) \right| \\ = \left| \sum_j \sum_{\alpha \in A_j, |\alpha|< \mu(\delta)} m_{\varphi^\delta, a_j, \Psi_j} (\alpha) | \log \mu(\delta)| - \sum_j \sum_{\alpha \in A_j, |\alpha|\ge \mu(\delta)} m_{\varphi^\delta, a_j, \Psi_j} (\alpha) \log |\alpha| \right| \\ \le 2 (\sum_j \tau_{\Psi_j}) |\log \mu(\delta)|,\end{gathered}$$ and this last quantity tends to $0$, which concludes the proof of (i). To prove (ii), take maps $\varphi^{\varepsilon}$ and systems of points $(A_j({\varepsilon}))$, admissible for $S({\varepsilon})$, such that $$\lim_{{\varepsilon}\to 0} \mathcal S ( \varphi^{\varepsilon}, ( A_j({\varepsilon}))_{1\le j \le N}) = \limsup_{{\varepsilon}\to0} \mathcal L^{(1+g({\varepsilon}))\Omega}_{S({\varepsilon})} (0).$$ Use the above proof with $\delta=g({\varepsilon})$ to construct maps $\tilde \varphi^{\varepsilon}$ into $\Omega$ and systems of points $(\tilde A_j({\varepsilon}))$, admissible for $S({\varepsilon})$, such that $$\left| \mathcal S (\tilde \varphi^{\varepsilon}, (\tilde A_j({\varepsilon}))_{1\le j \le N}) - \mathcal S (\varphi^{\varepsilon}, (A_j({\varepsilon}))_{1\le j \le N}) \right| \le 2 (\sum_j \tau_{\Psi_j}) |\log \mu(g({\varepsilon}))|,$$ and by definition $\mathcal S (\tilde \varphi^{\varepsilon}, (\tilde A_j({\varepsilon}))_{1\le j \le N}) \ge \mathcal L^\Omega_{S({\varepsilon})} (0) $. Consider as in Theorem \[conv\] a bounded convex domain $\Omega,$ and distinct points $a_j \in \Omega$, $1\le j \le N$. Let $z \in \Omega \setminus \{a_j, 1 \le j \le N\}$ (otherwise the property is trivially true). Again we may assume $z=0$. By Lemma \[relax\] applied to $S(\delta)=S$ for any $\delta$, to show that $$\label{liminfgr} \mathcal L_S (z) \le \liminf_{{\varepsilon}\to 0} \mathcal L_{S({\varepsilon})} (z)=: \ell ,$$ it will be enough to provide: some increasing function $g$ such that $g(0)=0$ and, for any $\delta>0$, $\varphi^\delta \in Hol({{\mathbb{D}}},(1+g(\delta))\Omega)$ and subsets $(A_j^\delta )_{1\le j \le N})$ of $\Omega$ such $(\varphi^\delta , (A_j^\delta )_{1\le j \le N})$ is admissible for $S, z$ and that $$\mathcal S (\varphi^\delta, (A_j^\delta)_{1\le j \le N}) = \ell.$$ The systems $S({\varepsilon})$ all have single poles, so the definition of $\ell$ means that there exist $\varphi_m \in Hol ({{\mathbb{D}}}, \Omega)$, ${\varepsilon}_m\to 0$, and points $\alpha_{j,m}, \alpha'_{j,m}, \alpha''_{j,m} \in {{\mathbb{D}}}$ such that $\varphi_m ( \alpha_{j,m})= a_j({\varepsilon}_m),$ $1\le j \le M$, and $\varphi_m ( \alpha'_{j,m})= a'_j({\varepsilon}_m),$ $\varphi_m ( \alpha''_{j,m})= a''_j({\varepsilon}_m),$ $M+1\le j \le N$; and they satisfy $$\sum_{j=1}^M \log | \alpha_{j,m} | + \sum_{j=M+1}^N \left( \log |\alpha'_{j,m} | + \log | \alpha''_{j,m} | \right) = \ell + \delta (m) ,$$ with $\lim_{m\to \infty} \delta (m) =0$. Passing to a subsequence, for which we keep the same notations, we may assume that $\alpha_{j,m}\to \alpha_{j} \in \overline {{\mathbb{D}}},$ $\alpha'_{j,m}\to \alpha'_{j} \in \overline {{\mathbb{D}}},$ $\alpha''_{j,m}\to \alpha''_{j} \in \overline {{\mathbb{D}}}$ as $m\to \infty$, and that $\varphi_m \to \tilde \varphi \in Hol ({{\mathbb{D}}}, \overline \Omega)$ uniformly on compact subsets of ${{\mathbb{D}}}$. Furthermore, by compactness of the unit circle, there exists $\tilde {v_j} \in [v_j] \cap S^{2n-1}$ such that, taking a further subsequence, $$\lim_{m\to \infty} \frac{a''_j({\varepsilon}_m)-a'_j({\varepsilon}_m)}{\| a''_j({\varepsilon}_m)-a'_j({\varepsilon}_m) \|} = \tilde {v_j}.$$ By renumbering the points and exchanging $a'_j$ and $a''_j$ as needed, we may assume that there are integers $M' \le M$, $M\le N_1 \le N_2 \le N_3 \le N$ such that $$\begin{aligned} \alpha_j \in {{\mathbb{D}}}&\mbox{ for }& 1 \le j \le M' \\ \alpha_j \in \partial {{\mathbb{D}}}&\mbox{ for }& M'+1 \le j \le M\\ \alpha'_j = \alpha''_j\in {{\mathbb{D}}}&\mbox{ for }& M+1 \le j \le N_1 \\ |\alpha'_j| < |\alpha''_j| <1 &\mbox{ for }& N_1+1 \le j \le N_2 \\ |\alpha'_j |<1 , |\alpha''_j| =1 &\mbox{ for }& N_2+1 \le j \le N_3 \\ |\alpha'_j|=|\alpha''_j|=1 &\mbox{ for }& N_3+1 \le j \le N.\end{aligned}$$ Then $$\begin{gathered} \ell = \lim_{m\to\infty} \left( \sum_{j=1}^M \log |\alpha_{j,m}| + \sum_{j=M+1}^N \left( \log |\alpha'_{j,m}| + \log |\alpha''_{j,m}|\right) \right) \\ = \sum_{j=1}^{M'} \log |\alpha_{j}| + \sum_{j=M+1}^{N_1} 2 \log |\alpha'_{j}| + \sum_{j=N_1+1}^{N_2} \left( \log |\alpha'_{j}| + \log |\alpha''_{j}| \right) + \sum_{j=N_2+1}^{N_3} \log |\alpha'_{j}| .\end{gathered}$$ Now we choose $$\begin{aligned} A_j = \{\alpha_j \} &\mbox{ for }& 1 \le j \le M' \\ A_j = \emptyset &\mbox{ for }& M'+1 \le j \le M\\ A_j = \{\alpha'_j \} &\mbox{ for }& M+1 \le j \le N_1 \\ A_j = \{\alpha'_j, \alpha''_j\} &\mbox{ for }& N_1+1 \le j \le N_2 \\ A_j = \{\alpha'_j \} &\mbox{ for }& N_2+1 \le j \le N_3 \\ A_j = \emptyset &\mbox{ for }& N_3+1 \le j \le N.\end{aligned}$$ Notice that $(\tilde \varphi, (A_j)_{1\le j \le N})$ hits the correct points but doesn’t necessarily produce an admissible choice, because for some $j$, $N_1+1 \le j \le N_2$, we could have $$m_{\tilde \varphi, a_j, \Psi_j} (\alpha'_j) + m_{\tilde \varphi, a_j, \Psi_j} (\alpha''_j) > 2=\tau_j.$$ So, in order to apply Lemma \[relax\] with $\delta\to 0$, we set $A_j^\delta =A_j$ for any $\delta>0$ and $$\tilde \varphi^\delta (\zeta) := \tilde \varphi (\zeta) + \delta \zeta \left[ \prod_1^{M'} (\zeta-\alpha_j) \prod_{M+1}^{N_1} (\zeta-\alpha'_j)^2 \prod_{N_1+1}^{N_2} (\zeta-\alpha'_j) (\zeta-\alpha''_j) \prod_{N_2+1}^{N_3} (\zeta-\alpha'_j) \right] \, v,$$ where $v \in {{\mathbb{C}}}^n$ is a unit vector chosen such that $\pi_j(v) \neq 0$, $N_1+1 \le j \le N_3$. For any $\alpha \in \cup_1^N A_j$, $\tilde \varphi^\delta (\alpha) = \tilde \varphi (\alpha)$. There is a constant $C>0$ such that $\tilde \varphi^\delta ({{\mathbb{D}}}) \subset \Omega + C \delta B(0,1)$. All the following considerations apply when $\delta$ is small enough. For $1\le j \le M'$, $m_{\tilde \varphi^\delta, a_j, \Psi_j} (\alpha_j)=1$, because $\tilde \varphi^\delta$ takes on the correct value, and the multiplicity cannot be more than $1=\tau_j$ in those cases. For $N_1+1\le j \le N_3$, we have $$\pi_j ((\tilde \varphi^\delta)'(\alpha'_j)) = \pi_j ((\tilde \varphi)'(\alpha'_j)) + \delta p_j \pi_j(v),$$ where $p_j$ is some complex scalar which doesn’t depend on $\delta$, so for $\delta >0$ and small enough, this projection doesn’t vanish and we have $m_{\tilde \varphi^\delta, a_j, \Psi_j} (\alpha'_j)=1$. An analogous reasoning shows that $m_{\tilde \varphi^\delta, a_j, \Psi_j} (\alpha''_j)=1$ for $N_1+1\le j \le N_2$. For $M+1 \le j \le N_1$, we have $$(\tilde \varphi^\delta)'(\alpha'_j) = (\tilde \varphi)'(\alpha'_j),$$ and by the uniform convergence on compact sets, $$(\tilde \varphi)'(\alpha'_j) = \lim_{m\to\infty} \frac{\varphi_m (\alpha'_{j,m})-\varphi_m (\alpha''_{j,m})}{\alpha'_{j,m}-\alpha''_{j,m}} = \lim_{m\to\infty} \frac{a'_j({\varepsilon}_m)-a''_j({\varepsilon}_m)}{\alpha'_{j,m}-\alpha''_{j,m}},$$ which must be colinear to $v_j$ by definition. Therefore $m_{\tilde \varphi^\delta, a_j, \Psi_j} (\alpha'_j)=2$ for $M+1 \le j \le N_1$. Thus $(\tilde \varphi^\delta, (A_j)_{1\le j \le N})$ is admissible for $S, 0$ and $\mathcal S(\tilde \varphi^\delta, (A_j)_{1\le j \le N})=\ell$, which proves . Now we need to show that $$\label{limsupsm} \mathcal L_S(z) \ge \limsup_{{\varepsilon}\to0} \mathcal L_{S({\varepsilon})}(z).$$ We use Lemma \[relax\](ii). For any $\delta >0$, we need to construct a positive function $g$ such that $\lim_{{\varepsilon}\to0}g({\varepsilon})=0$ and, for ${\varepsilon}$ small enough, $\varphi^{\varepsilon}\in Hol({{\mathbb{D}}}(1+g({\varepsilon}))\Omega)$ and sets $(A_j({\varepsilon}))_{1\le j \le N}$ such that $ (\varphi^{\varepsilon}, (A_j({\varepsilon}))_{1\le j \le N})$ is admissible for $S({\varepsilon}),0$ and $$\mathcal S (\varphi^{\varepsilon}, (A_j({\varepsilon}))_{1\le j \le N}) \le \mathcal L_S(z) +\delta.$$ We start with an admissible choice $(\varphi, (A_j)_{1\le j \le N})$ for $S$, such that $$\mathcal S (\varphi, (A_j)_{1\le j \le N}) \le \mathcal L_S (z) + \delta/2.$$ To fix notations, suppose that, after renumbering and exchanging the points as needed, there exist integers $M'\le M$, $N_1, N_2, N_3 \in \{ M, \dots N \}$ such that $$\begin{aligned} A_j = \{\alpha_j \} &\mbox{ for }& 1 \le j \le M' ,\\ A_j = \emptyset &\mbox{ for }& M'+1 \le j \le M,\\ A_j = \{\alpha'_j \}, m_{\varphi, a_j, \Psi_j} (\alpha'_j)=2 &\mbox{ for }& M+1 \le j \le N_1 ,\\ A_j = \{\alpha'_j, \alpha''_j\}, \alpha'_j \neq \alpha''_j &\mbox{ for }& N_1+1 \le j \le N_2 ,\\ A_j = \{\alpha'_j \}, m_{\varphi, a_j, \Psi_j} (\alpha'_j)=1 &\mbox{ for }& N_2+1 \le j \le N_3 ,\\ A_j = \emptyset &\mbox{ for }& N_3+1 \le j \le N.\end{aligned}$$ The definition of $\Psi_j$ (see the computations performed in the Elementary example) implies that, for $M+1 \le j \le N_1$, $\varphi'(\alpha'_j) \cdot \bar w = 0$, for any $w \in v_j^\perp$. We perturb $\varphi$ to make sure that, on the other hand, $\varphi'(\alpha'_j) \cdot \bar v_j \neq 0$ in the same index range. For $\eta({\varepsilon}) \in \mathbb C$ to be chosen later, set $$\begin{gathered} \tilde \varphi (\zeta) := \varphi (\zeta) + \\ \eta({\varepsilon}) \, \left[ \zeta \, \prod_1^{M'} (\zeta - \alpha_j) \prod_{j=N_1+1}^{N_2} (\zeta - \alpha'_j)(\zeta - \alpha''_j) \prod_{j=N_2+1}^{N_3} (\zeta - \alpha'_j) \right] \times \\ \times \left\{ \sum_{j=M+1}^{N_1} \left[ (\zeta - \alpha'_j) \prod_{M+1\le k \le N_1, k\neq j} (\zeta - \alpha'_k)^2 \right] \, v_j \right\}.\end{gathered}$$ The map $\tilde \varphi$ depends on ${\varepsilon}$ and is admissible again. We have positive constants $C_1, C_2, C_3$ such that - $\tilde \varphi'(\alpha'_j)=\lambda_j v_j$, with $C_1^{-1} |\eta({\varepsilon})| \le |\lambda_j | \le C_1 |\eta({\varepsilon})| $, - $\|\tilde \varphi - \varphi\|_\infty \le C_2 |\eta({\varepsilon})|$, - $\tilde \varphi({{\mathbb{D}}}) \subset (1+C_3|\eta({\varepsilon})|) \Omega$; in particular $\tilde \varphi$ will be bounded by constants independent of ${\varepsilon}$, along with all its derivatives on any given compact subset of ${{\mathbb{D}}}$. For $M+1 \le j \le N$ and ${\varepsilon}$ in a neighborhood of $0$, $a''_j({\varepsilon}) - a'_j({\varepsilon}) = n_j({\varepsilon}) v_j({\varepsilon})$, where $\|v_j({\varepsilon})\|=1$, $\lim_{{\varepsilon}\to 0} v_j({\varepsilon}) = v_j$ and $n_j({\varepsilon}) \in {{\mathbb{C}}}$. For $|{\varepsilon}|$ small enough, we now may define $$\begin{gathered} A_j({\varepsilon}) := A_j, \mbox{ for } 1 \le j \le M, N_1+1 \le j \le N, \\ \mbox{ and } A_j({\varepsilon}) :=\{ \alpha'_j, \alpha'_j +\frac{n_j({\varepsilon})}{\lambda_j} \}, \mbox{ for } M+1 \le j \le N_1.\end{gathered}$$ We shall need to add to $\tilde \varphi$ a vector-valued correcting term obtained by Lagrange interpolation. To this end, we write $B({\varepsilon}) := \cup_j A_j({\varepsilon})$, and values to be interpolated, $w (\alpha)$, for $\alpha \in B({\varepsilon})$. Let $$\begin{aligned} w(\alpha_j ) := a_j({\varepsilon}) - a_j = a_j({\varepsilon}) - \tilde \varphi(\alpha_j ) &\mbox{ for }& 1 \le j \le M' ,\\ w(\alpha'_j ) := a'_j({\varepsilon}) - a_j = a'_j({\varepsilon}) - \tilde \varphi(\alpha'_j ) &\mbox{ for }& M+1 \le j \le N_1 ,\\ w(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) := a''_j({\varepsilon}) - \tilde \varphi(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) &\mbox{ for }& M+1 \le j \le N_1 ,\\ w(\alpha'_j ) := a'_j({\varepsilon}) - a_j = a'_j({\varepsilon}) - \tilde \varphi(\alpha'_j ) &\mbox{ for }& N_1+1 \le j \le N_2 ,\\ w(\alpha''_j ) := a''_j({\varepsilon}) - a_j = a''_j({\varepsilon}) - \tilde \varphi(\alpha''_j ) &\mbox{ for }& N_1+1 \le j \le N_2 ,\\ w(\alpha'_j ) := a'_j({\varepsilon}) - a_j = a'_j({\varepsilon}) - \tilde \varphi(\alpha'_j ) &\mbox{ for }& N_2+1 \le j \le N_3 .\end{aligned}$$ We denote by $P_{\varepsilon}$ the solution to the interpolation problem $$\left(P(\alpha)= w(\alpha): \alpha \in B({\varepsilon}) \right).$$ Let $\varphi^{\varepsilon}:= \tilde \varphi + P_{\varepsilon}\in Hol({{\mathbb{D}}}, \Omega^{\varepsilon})$. The domain $\Omega^{\varepsilon}$ will be specified below. By construction $ (\varphi^{\varepsilon}, (A_j({\varepsilon}))_{1\le j \le N})$ is admissible for $S({\varepsilon})$, and for $|{\varepsilon}|$ small enough, $$\mathcal S (\varphi^{\varepsilon}, (A_j({\varepsilon}))_{1\le j \le N}) \le \mathcal L_S(z) +\delta,$$ provided that, for $M+1 \le j \le N_1$, $$\label{lambdacond} \lim_{{\varepsilon}\to 0} \frac{n_j({\varepsilon})}{\lambda_j} = 0,$$ Now we need to show that the correction is small, more precisely that we can choose $\eta({\varepsilon})$ so that the above condition is satisfied and $\lim_{{\varepsilon}\to 0} \|P_{\varepsilon}\|_\infty = 0$. Then we can choose a function $g$ tending to $0$ such that $$\Omega^{\varepsilon}= (1+g({\varepsilon}) ) \Omega \supset (1+C_3|\eta({\varepsilon})|) \Omega + B(0,\|P_{\varepsilon}\|_\infty).$$ Write $\Pi_\alpha$ for the unique (scalar) polynomial of degree less or equal to $d:= \# B({\varepsilon}) - 1$ ($d$ does not depend on ${\varepsilon}$) such that $$\Pi_\alpha (\alpha) = 1, \Pi_\alpha (\beta) = 0 \mbox{ for any } \beta \in B({\varepsilon}) \setminus \{\alpha\}.$$ Then $$P_{\varepsilon}= \sum_{\alpha \in B({\varepsilon})} \Pi_\alpha w(\alpha).$$ For $\alpha \in \bigcup_{1 \le j \le M, N_1+1 \le j \le N} A_j$, $\|\Pi_\alpha\|_\infty$ is uniformly bounded, because $\mbox{dist} (\alpha, B({\varepsilon}) \setminus \{\alpha\}) \ge \gamma > 0$ with $\gamma$ independent of ${\varepsilon}$. It also follows from the hypotheses of the theorem and the choice of $w$ that $$\lim_{{\varepsilon}\to 0} \, \max \{ \|w(\alpha)\| , \alpha \in \bigcup_{1 \le j \le M, N_1+1 \le j \le N} A_j\} = 0.$$ For $M+1 \le j \le N_1$, we need an elementary lemma about Lagrange interpolation. \[Lagrange\] Let $x_0, \dots , x_d \in {{\mathbb{D}}}$, $w_0, w_1 \in {{\mathbb{C}}}^n$. Suppose that there exists $\gamma >0$ such that $|x_0-x_1|\le \gamma $ and $\mbox{dist} ([x_0,x_1], \{x_2, \dots, x_d\})\ge 2 \gamma$, where $[x_0,x_1]$ is the real line segment from $x_0$ to $x_1$. Let $P$ be the unique (${{\mathbb{C}}}^n$-valued) polynomial of degree less or equal to $d$ such that $$P(x_0)=w_0, P(x_1)=w_1, P(x_j)=0, 2 \le j \le d.$$ Then there exist constants $L_1, L_0$ depending only on $\gamma$ and $d$ such that $$\sup_{\zeta \in {{\mathbb{D}}}} \|P(\zeta)\| \le L_1 \left\| \frac{w_1-w_0}{x_1-x_0} \right\| + L_0 \|w_0\|.$$ We will prove this Lemma a little later. It yields, for $M+1 \le j \le N_1$, $$\begin{gathered} \sup_{\zeta \in {{\mathbb{D}}}} \left\| \Pi_{\alpha'_j}(\zeta) w(\alpha'_j ) + \Pi_{\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}} (\zeta) w(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) \right\| \\ \le L_1 \left|\frac{\lambda_j} {n_j({\varepsilon})}\right| \left\| a''_j({\varepsilon}) - \tilde \varphi(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) \right\| + L_0 \|a'_j({\varepsilon}) - a_j\|.\end{gathered}$$ We now estimate the first term in the last sum above. By the Taylor formula, $$a''_j({\varepsilon}) - \tilde \varphi(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) = a''_j({\varepsilon}) - a'_j({\varepsilon}) - n_j({\varepsilon}) v_j + R_2 ({\varepsilon}) = n_j({\varepsilon}) ( v_j({\varepsilon}) - v_j ) + R_2 ({\varepsilon}),$$ where $\| R_2 ({\varepsilon})\| \le C |n_j({\varepsilon})|^2|\lambda_j|^{-2}$ with $C$ a constant independent of ${\varepsilon}$ by the boundedness of the derivatives of $\tilde \varphi$. Finally $$\begin{gathered} \sup_{\zeta \in {{\mathbb{D}}}} \left\| \Pi_{\alpha'_j}(\zeta) w(\alpha'_j ) + \Pi_{\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}} (\zeta) w(\alpha'_j + \frac{n_j({\varepsilon})}{\lambda_j}) \right\| \\ \le C \left( \|v_j({\varepsilon}) - v_j\| |\eta({\varepsilon})| + |n_j({\varepsilon})||\eta({\varepsilon})|^{-1} + \|a'_j({\varepsilon}) - a_j\| \right).\end{gathered}$$ To to satisfy , we need to have $\lim_{{\varepsilon}\to 0} n_j({\varepsilon})/\eta({\varepsilon})=0$; to make sure, in addition, that the whole sum above tends to $0$ as ${\varepsilon}$ tends to $0$, it will be enough to choose $\eta({\varepsilon})$ going to zero, but more slowly that $|n_j({\varepsilon})| =\|a''_j({\varepsilon}) - a'_j\|$, for $M+1 \le j \le N_1$. [*Proof of Lemma \[Lagrange\]*]{} Let $$Q(X,Y):= \prod_2^d \frac{X-x_k}{Y-x_k}.$$ Then $Q$ and all of its derivatives are bounded for $X \in \overline {{\mathbb{D}}}$ and $Y\in [x_0,x_1]$. $$\begin{gathered} P(X) = \frac{X-x_0}{x_1-x_0} Q(X,x_1) w_1 + \frac{X-x_1}{x_0-x_1} Q(X,x_0) w_0 \\ = \frac{w_1 - w_0}{x_1-x_0} (X-x_0) Q(X, x_1) + \left( -Q(X,x_1) + (X-x_1) \frac{Q(X, x_1)-Q(X, x_0)}{x_0-x_1} \right) \, w_0.\end{gathered}$$ Then the conclusion follows from the boundedness of $Q$ and $Q'$ and the mean value theorem. Comparison with previous results {#compprev} ================================ In [@TraoTh], we had used a different definition for a Lempert function with multiplicities. We state it with the same notations as in Definition \[defLempert\]. \[prevdef\] Given a system $S$ as in Definition \[defgreen\], we write $\tau_j := \tau_{\Psi_j}$. Let $\varphi \in Hol({{\mathbb{D}}},\Omega)$ and $\alpha_j \in {{\mathbb{D}}}$, $1\le j \le N$. We say that $(\varphi, (\alpha_j)_{1\le j \le N})$ is *admissible (for $S$, $z$) in the old sense* if $$\begin{gathered} \varphi (0)=z, \mbox{ and there exists } U_j \mbox{\rm\ a neighborhood of }\zeta_j \\ \mbox{ s.t. } \Psi_j (\varphi(\zeta) -a_j) \le \tau_j \log|\zeta-\zeta_j| + C_j, \forall \zeta \in U_j, 1\le j \le N.\end{gathered}$$ In this case, we write (with the convention that $0 \cdot \infty =0$) $$\mathcal S (\varphi, (\alpha_j)_{1\le j \le N}):= \sum_{j=1}^N \tau_j \log |\alpha_j| .$$ Then the *old generalized Lempert function* is defined by $$\begin{gathered} L^\Omega_S (z):= L_S (z) \\ := \inf \left\lbrace \mathcal S (\varphi, (\alpha_j)_{1\le j \le N}) : (\varphi, (\alpha_j)_{1\le j \le N}) \mbox{ is admissible for }S, z \mbox{ in the old sense } \right\rbrace .\end{gathered}$$ Recall also that since the functional $L$ did not enjoy monotonicity properties, another definition was given in [@TraoTh]. \[modifLemp\] Let $S:= \{(a_j,\Psi_j): 1 \le j \le N\}$ and $S_1:= \{(a_j,\Psi^1_j): 1 \le j \le N\}$ where $a_j \in \Omega$ and $\Psi_j$, $\Psi^1_j$ are local indicators. We define $$\tilde L_S (z) := \inf \{ L_{S^1}(z) : \Psi^1_j \ge \Psi_j + C_j, 1 \le j \le N\} .$$ \[comparison\] If $S=\{(a_j, \Psi_j), 1\le j \le N\}$, where the $\Psi_j$ are elementary local indicators, then for any $z \in \Omega$, $\mathcal L_S (z) \le \tilde L_S (z)$. Since the functional $\mathcal L$ is monotonic by Theorem \[monotone\], it will be enough to show that $\mathcal L_S (z) \le L_S (z)$ for any system $S$. If we have a map $\varphi$ which is admissible in the sense of Definition \[prevdef\], we can take $A_j:=\{\alpha_j\}$, and $\Psi_j (\varphi(\zeta) -a_j) \le \tau_j \log|\zeta-\alpha_j| + C_j$ implies that $m_{\varphi, a_j, \Psi_j} (\alpha_j) \ge \tau_j$, which by Definition \[multphia\] means that $m_{\varphi, a_j, \Psi_j} (\alpha_j) = \tau_j$. So that any such $\varphi$ is admissible in the sense of Definition \[defLempert\], and $$\mathcal S (\varphi, (a_j)_{1\le j \le N}) = \mathcal S (\varphi, (A_j)_{1\le j \le N}),$$ and the desired inequality follows. We now return to the study of the example presented in [@TraoTh]. Let us recall the notations. For $z \in \mathbb D^2$, $$\Psi_0 (z) := \max(\log|z_1|,\log|z_2|), \quad \Psi_V (z) := \max(\log|z_1|,2\log|z_2|).$$ Here $V$ stands for “vertical”, for the obvious reasons : for $a \in \mathbb D^2$, $\Psi_j ( \varphi (\zeta) -a) \le \tau_j \log|\zeta -\zeta_0| + C$ translates to ($\tau_0=1$, $\tau_V=2$): $$\begin{aligned} \varphi (\zeta_0) = a, &\mbox{ when }& j =0, \\ \varphi (\zeta_0) = a, \varphi'_1(\zeta_0) =0 &\mbox{ when }& j =V .\end{aligned}$$ For $a$, $b \in \mathbb D$ and ${\varepsilon}\in \mathbb C$, let $$\begin{aligned} S_{\varepsilon}&:=& \{ ((a,0), \Psi_0);((b,0), \Psi_0);((b,{\varepsilon}), \Psi_0);((a,{\varepsilon}), \Psi_0) \} \\ S &:=& \{ ((a,0), \Psi_V);((b,0), \Psi_V) \}.\end{aligned}$$ Those are product set situations, and the Green functions are explicitly known. For $w \in \mathbb D$, denote by $\phi_w$ the unique involutive holomorphic automorphism of the disk which exchanges $0$ and $w$: $$\phi_w (\zeta) := \frac{w-\zeta}{1-\zeta \bar w}.$$ Then $$\begin{gathered} G_S (z_1,z_2) = \max \left( \log |\phi_a(z_1) \phi_b(z_2)|, 2 \log |z_2| \right), \\ G_{S_{\varepsilon}} (z_1,z_2) = \max \left( \log |\phi_a(z_1) \phi_b(z_2)|, \log |z_2 \phi_{\varepsilon}(z_2)| \right).\end{gathered}$$ The following is proved in [@TraoTh p. 397]. \[basicineq\] If $b=-a$ and $|a|^2<|\gamma|<|a|$, then $G_S (0,\gamma)<\tilde L_S (0,\gamma)$. It follows from our Theorem \[conv\] that for any $z \in \mathbb D^2$, $\lim_{{\varepsilon}\to 0} L_{S_{\varepsilon}} (z) = \mathcal L_S(z)$, and in particular, using Lemma \[comparison\], we find again the result laboriously obtained in [@TTppt Proposition 6.1]: $\limsup_{{\varepsilon}\to 0} L_{S_{\varepsilon}} (z) \le \tilde L_S(z)$. It is a consequence of [@TraoTh Theorem 5.1] (or equivalently [@TTppt Theorem 6.2]) that for $b=-a$ and $|a|^{3/2} < |\gamma| <|a|$, then $\mathcal L_S (0,\gamma) > G_S (0,\gamma) $; the motivation then was to obtain the counterexample $ L_{S_{\varepsilon}} (0,\gamma) > G_{S_{\varepsilon}} (0,\gamma)$ for $|{\varepsilon}|$ small enough. On the other hand, when $ |\gamma| < |a|^{3/2}$, the old generalized Lempert function doesn’t provide the correct limit of the single pole Lempert functions. \[distinct\] For $b=-a$ and $|a|^2 < |\gamma| < |a|^{3/2}$, $\mathcal L_S (0,\gamma) < \tilde L_S (0,\gamma) $. Since Proposition \[basicineq\] implies that $\tilde L_S (0,\gamma) > G_S (0,\gamma) = 2 \log |a|$, it will be enough to provide a mapping $\varphi$ and sets $A_1, A_2$ admissible in the sense of Definition \[defLempert\] such that $\mathcal S (\varphi; A_1, A_2) \le 2 \log |a|$. We restrict ourselves to $a>0$. We now choose $A_1:=\{\zeta_1, \zeta_4\}$, $A_2:=\{\zeta_2\}$, with $$\zeta_2:= \sqrt{a}, \quad \zeta_1 := \phi_{\zeta_2} \left( \sqrt{\frac{2a}{1+a^2}} \right), \quad \zeta_4 := \phi_{\zeta_2} \left( -\sqrt{\frac{2a}{1+a^2}} \right),$$ and $$\varphi_1 (\zeta) := \phi_{-a} \left( - \phi_{\zeta_2} (\zeta)^2 \right), \varphi_2 (\zeta) := \frac{\gamma}{\zeta_1\zeta_2\zeta_4} \phi_{\zeta_1} (\zeta) \phi_{\zeta_2} (\zeta) \phi_{\zeta_4} (\zeta).$$ From those definitions it is clear that $\varphi_1 (\mathbb D) \subset \mathbb D$ and that $$\varphi_1 (\zeta_2) = -a, \varphi_1' (\zeta_2) =0; \quad \varphi_2 (\zeta_j) = 0, \mbox{ for } j = 1, 2, 4.$$ Furthermore, using the involutivity of $\phi_{\zeta_2} $, $$\varphi_1 (\zeta_1) = \varphi_1 (\zeta_4) = \phi_{-a} \left(- \frac{2a}{1+a^2}\right) = \phi_{-a} \left( \phi_{-a} (a) \right) = a.$$ So the map $\varphi$ hits the poles, and $$m_{\varphi, (a,0), \Psi_V} (\zeta_1) \ge 1, \quad m_{\varphi, (a,0), \Psi_V} (\zeta_4) \ge 1, \quad m_{\varphi, (-a,0), \Psi_V} (\zeta_2) = 2.$$ To see that actually $m_{\varphi, (a,0), \Psi_V} (\zeta_j)=1$, for $j=1, 4$, notice that, since $\varphi_1$ only admits one critical point, $\zeta_2$, and since $\zeta_1\neq \zeta_2$ and $\zeta_4\neq \zeta_2$, we must have $ \varphi_1' (\zeta_j)\neq 0$, $j=1, 4$. Thus $\varphi$ is admissible in the sense of Definition \[defLempert\], and $$\mathcal S (\varphi; A_1, A_2) = \log |\zeta_1| + 2 \log |\zeta_2| +\log |\zeta_4| = \log |\zeta_1 \zeta_4 \zeta_2^2|.$$ We need to compute $$\begin{gathered} \zeta_1 \zeta_4= \phi_{\sqrt a} \left( \sqrt{\frac{2a}{1+a^2}} \right) \cdot \phi_{\sqrt a} \left( -\sqrt{\frac{2a}{1+a^2}} \right) \\ = \frac{\sqrt a - \sqrt{\frac{2a}{1+a^2}} }{1-\sqrt a \sqrt{\frac{2a}{1+a^2}}} \cdot \frac{\sqrt a + \sqrt{\frac{2a}{1+a^2}} }{1 + \sqrt a \sqrt{\frac{2a}{1+a^2}}} =\frac{ a - \frac{2a}{1+a^2} }{1- a \frac{2a}{1+a^2} } = \phi_a \left( \phi_a (-a) \right) = -a.\end{gathered}$$ From this we deduce $|\zeta_1 \zeta_2 \zeta_4| = a^{3/2} > |\gamma|$, and therefore $\varphi_2 (\mathbb D) \subset \mathbb D$; and $ |\zeta_1 \zeta_4 \zeta_2^2| = a^2$, therefore $$\mathcal S (\varphi; A_1, A_2) \le \log |\zeta_1 \zeta_4 \zeta_2^2| = 2 \log |a|,$$ q.e.d. [ABC]{} E. Bedford, B. A. Taylor, [*A new capacity for plurisubharmonic functions,*]{} Acta Math. [**149**]{} (1982), 1–40. M. Carlehed, J. Wiegerinck, [*Le cône des fonctions plurisousharmoniques négatives et une conjecture de Coman,*]{} Ann. Pol. Math. [**80**]{} (2003), 93–108. D. Coman, [*The pluricomplex Green function with two poles of the unit ball of $\mathbb C^n$,*]{} Pacific J. Math. [**194,**]{} no 2, 257–283 (2000). S. Dineen, [*The Schwarz Lemma,*]{} Clarendon Press, Oxford, 1989. A. Edigarian, [*Remarks on the pluricomplex Green function,*]{} Univ. Iagel. Acta Math. [**37**]{} (1999), 159–164. J. B. Garnett, [*Bounded Analytic Functions,*]{} Academic Press, Inc. New York-London, 1981. M. Klimek, [*Pluripotential Theory,*]{} Oxford Science Publications, Oxford-New York-Tokyo, 1991. P. Lelong and A. Rashkovskii, [*Local indicators for plurisubharmonic functions,*]{} J. Math. Pure Appl. [**78**]{}, 233–247 (1999). L. Lempert, [*La métrique de Kobayashi et la représentation des domains sur la boule,*]{} Bull. Soc. Soc. Math. France [**109**]{}, 427–474 (1981). N. Nikolov, P. Pflug, [*The multipole Lempert function is monotone under inclusion of pole sets,*]{} Mich. Math. J., [**54**]{}, no. 1,111–116 (2006). N. Nikolov, W. Zwonek, [*On the product property for the Lempert function,*]{} Complex Var. Theory Appl. [**50**]{}, no. 12, 939–952 (2005). A. Rashkovskii, [*Newton numbers and residual measures of plurisubharmonic functions,*]{} Ann. Polon. Math. [**75**]{} (2000), 213–231. A. Rashkovskii, R. Sigurdsson, [*Green functions with singularities along complex spaces,*]{} Internat. J. Math. [**16**]{} (2005), no. 4, 333–355. P. J. Thomas, [*An example of limit of Lempert Functions,*]{} Viet Nam J. Math. 35, no. 3, 1–14 (2007). P. J. Thomas, N. V. Trao, [*Pluricomplex Green and Lempert functions for equally weighted poles.*]{} Prépublication no. 240 du Laboratoire de Mathématiques Emile Picard, Université Paul Sabatier, Toulouse, France, march 2002 (25 pp.). Available as an e-print on math arXiv: http://xxx.lanl.gov/abs/math.CV/0206214. P. J. Thomas, N. V. Trao, [*Pluricomplex Green and Lempert functions for equally weighted poles.*]{} Ark. Mat. 41, no. 2, 381–400 (2003). F. Wikström, [*Jensen Measures, Duality and Plurisubharmonic Green Functions*]{}, Doctoral thesis No 18, Umeå University, 1999. F. Wikström, [*Non-linearity of the pluricomplex Green function,*]{} Proc. Amer. Math. Soc. 129, no. 4, 1051–1056 (2001). .2cm Pascal J. Thomas Institut de Mathématiques de Toulouse CNRS UMR 5219 UFR MIG Université Paul Sabatier F-31062 TOULOUSE CEDEX 9 France pthomas@math.univ-toulouse.fr .5cm Nguyen Van Trao Department of Mathematics Dai Hoc Su Pham 1 (Pedagogical Institute of Hanoi) Cau Giay, Tu Liem Ha Noi Viet Nam ngvtrao@yahoo.com

--- abstract: 'We formulate a statistical model for description of nuclear composition and equation of state of stellar matter at subnuclear densities and temperature up to 20 MeV, which are expected during the collapse and explosion of massive stars. The model includes nuclear, electromagnetic and weak interactions between all kinds of particles, under condition of statistical equilibrium. We emphasize importance of realistic description of the nuclear composition for understanding stellar dynamics and nucleosynthesis. It is demonstrated that the experience accumulated in studies of nuclear multifragmentation reactions can be used for better modelling properties of stellar medium.' --- [**Statistical approach for supernova matter.**]{} [A.S. Botvina$^{a,b}$ and I.N. Mishustin$^{b,c}$]{}\ [ *$^a$Institute for Nuclear Research, Russian Academy of Sciences, 117312 Moscow, Russia\ $^b$Frankfurt Institute for Advanced Studies, J.W. Goethe University, D-60438 Frankfurt am Main, Germany\ $^c$Kurchatov Institute, Russian Research Center, 123182 Moscow, Russia\ *]{} [PACS: 26.50.+x , 21.65.-f, 25.70.Pq , 26.30.-k, 97.60.Bw]{} [**1. Introduction**]{}\ In violent nuclear reactions strong interaction between many nucleons leads to a fast equilibration. This short-range interaction is responsible for a sharp freeze-out when the inter-particle distance becomes larger than the interaction range. For these reasons statistical models have proved to be very successful for interpretation of nuclear reactions at various energies. They are widely used for description of fragment production when one or several equilibrated sources can be identified. Originally this concept was proposed for compound nucleus decays, such as evaporation or fission of excited nuclei [@Bohr]. Recently, it was demonstrated that the concept of equilibrated source can even be effectively used for more violent multifragmentation reactions leading to production of many fragments [@SMM]. On other side, the statistical equilibrium is expected in many astrophysical processes, when the characteristic time for formation of nuclei is much shorter than the time-scale of these processes. For example, one of the most spectacular astrophysical events is a type II supernova explosion, with a huge energy release of about several tens of MeV per nucleon [@Brown; @Bethe]. When the core of a massive star collapses, it reaches densities several times larger than the normal nuclear density $\rho_0=0.15$ fm$^{-3}$. The repulsive nucleon-nucleon interaction gives rise to a bounce-off and creation of a shock wave propagating through the in-falling stellar material. This shock wave is responsible for the ejection of a star envelope that is observed as a supernova explosion. During the collapse and subsequent explosion the temperatures $T\approx (0.5\div 10)$ MeV and densities $\rho \approx (10^{-6}\div 2) \rho_0$ can be reached. It is widely believed that the nuclear statistical equilibrium should be reached under these conditions. As shown by many theoretical studies, a liquid-gas phase transition should take place in nuclear matter under such conditions. As discussed by several authors (see e.g. [@Janka; @Thielemann; @Sumi; @Burrows]), there are problems in producing successful explosions in hydrodynamical simulations of the core-collapse supernovae, even when neutrino heating and convection effects are included. The hope is that full 3-$d$ simulations will help to solve this problem [@3d]. On the other hand, it is known that nuclear composition is extremely important for understanding the physics of supernova explosions. In particular, the weak reaction rates and energy spectra of emitted neutrinos are very sensitive to the presence of heavy nuclei (see e.g. [@Ring; @Hix; @Langanke; @Horowitz]). This is also true for the equation of state (EOS) used in hydrodynamical simulations since the shock strength is diminished by dissociation of heavy nuclei. Nuclear reactions in a supernova environment are especially important because supernova explosions may be considered as breeders for creating chemical elements. According to present understanding, there are three main sources of chemical element production in the Universe. The lightest elements (up to He, and, partly, Li) are formed during the first moments of the Universe expansion, immediately after the Big Bang. Light elements up to $^{16}$O can be produced in thermonuclear reactions in ordinary stars like our Sun, while heavier elements up to Fe and Ni can be formed in heavy stars at the end of the nuclear burning epoch. It is most likely that heavy elements up to U were synthesized in the course of supernova explosions. Pronounced peaks in the element abundances can be explained by neutron capture reactions in s- and r-processes [@Cowan; @Qian]. It is believed that suitable conditions for the r-process were provided by free neutrons abundantly produced in supernova environments together with appropriate seed nuclei. The EOS of supernova matter is under investigation for more than 25 years. One of the first EOS, frequently used in supernova simulations, was obtained in refs. [@Lamb; @Lattimer] many years ago. It includes both light and heavy nuclei in statistical equilibrium. However, it does not include the whole ensemble of hot heavy nuclei, replacing them by a single “average” nucleus. The same assumption within a relativistic mean-field approach was used in the EOS of ref. [@Shen]. As was already pointed out by many authors [@Japan; @Botvina04; @Langanke_structure; @Janka2] this assumption is not sufficient for accurate treatment of the supernova processes. We think that this kind of approximation can distort the true statistical ensemble in many cases. There are other statistical calculations which consider the ensemble with different nuclear species, but in the partition sum they include only nuclei in long-lived states known from terrestrial experiments (see refs. [@Japan; @Langanke_structure; @Janka2]). Also, for description of unknown neutron-rich hot nuclei only properties (e.g., the symmetry energies) of cold and slightly excited isolated nuclei have been used up to now. These assumptions are not justified for supernova environments characterized by relatively high temperatures (up to 10 MeV) and densities of electrons and baryons in the range $\rho \approx 10^{-4}\div10^{-1} \rho_0$. We believe, in order to achieve a more realistic description of supernova matter, it is necessary to use rich experience accumulated in recent years by the nuclear community in studying highly excited equilibrated systems in nuclear reactions. In particular, multifragmentation reactions provide valuable information about hot nuclei in dense surrounding of nucleons and other nuclei, which in many aspects is similar to supernova interior [@Botvina05]. [**2. Nuclear reactions in supernova environments.**]{}\ [**2.1 Studying equilibrated nuclear matter in laboratory.** ]{}\ Properties of strongly-interacting nuclear matter are studied experimentally and theoretically for a long time. At present there exist consensus about phase diagram of nuclear and neutron matter (see, e.g., refs. [@Mosel76; @Lamb78; @Lamb]). It is shown schematically in Fig. 1 for symmetric nuclear matter, for range of densities and temperatures expected in the Supernova II explosions. It is commonly accepted that this diagram contains a liquid-gas phase transition. From this phase diagram one can conclude that nuclear matter at densities $\rho \approx 0.3-0.8\rho_0$ and temperatures $T < T_c$ should be in the mixed phase. This phase is strongly inhomogeneous with intermittent dense and dilute regions. In the case of electrically-neutral matter this mixed phase may have different topologies such as spherical droplets, cylindrical nuclei, slab-like configurations and others. These configurations are generally referred to as nuclear ’pasta’ phases [@Pethick], which were recently under intensive theoretical investigation [@Pethick2; @Horowitz2; @Watanabe]. However, in the coexistence region at lower densities, $\rho < 0.3\rho_0$, which are considered in this paper, the nuclear matter breaks up into compact nuclear droplets surrounded by nucleons. These relatively low densities dominate during the main stages of stellar collapse and explosion. Description of nuclear composition in this region requires theoretical extrapolation of nuclear properties to these extreme conditions. As became obvious after intensive experimental studies of nuclear multifragmentation reactions, they proceed through formation of thermalized nuclear systems characterized by subnuclear densities $\rho \sim 0.1 \rho_0$ and temperatures of 3–8 MeV. Thermodynamic conditions associated with these reactions are indicated by the shaded area in Fig. 1. This gives us a chance to extract properties of hot nuclei in the environment of other nuclear species directly from the experimental multifragmentation data and then use this information for more realistic calculations of nuclear composition of stellar matter. We have also shown isentropic trajectories with the entropy per baryon (S/B) of 1, 2, and 4 units (see section [**5.2**]{}) typical for supernova explosions. One can see, for example, that an adiabatic collapse (and expansion) of stellar matter with typical entropies of 1-2 units per baryon passes exactly through the multifragmentation domain. Nuclear multifragmentation, i.e. break-up of hot heavy nuclei into many fragments, was under intensive investigation during the last 20 years. It was solidly established by both theoretical and experimental studies that this channel dominates at high excitation energies, above 3–4 MeV per nucleon, replacing sequential evaporation and fission of the compound nucleus, which is conventional mechanism at low excitation energies. In this respect, multifragmentation is a universal process expected in all types of nuclear reactions, induced by hadrons, heavy ions, and electromagnetic collisions, where the nucleus receives a high excitation energy. The Statistical Multifragmentation Model (SMM) [@SMM] is one of the most successful models used for theoretical interpretation of these reactions. Some examples, how low-density equilibrated nuclear systems can be produced, and how well their decay can be described within this statistical approach, can be found in refs. [@SMM; @Botvina90; @ALADIN; @EOS; @MSU; @INDRA; @FASA; @Dag]. Recently, experimental evidences have appeared [@LeFevre; @Iglio; @Souliotis] that the symmetry energy of hot fragments produced in multifragmentation reactions is significantly reduced as compared with the value in cold nuclei. The surface and bulk energy of nuclei in this hot and dense environment can be modified too [@Botvina06; @bulk]. These conclusions, made after appropriate analyses of experimental data, suggest that the same modifications will occur in stellar matter at similar densities and temperatures. Actually, the “in-medium” modifications are quite expected, since the nuclei will interact with the surrounding matter, and, therefore, should change their properties. These modifications may have important consequences for nuclear composition, equation of state, and weak reaction rates on these nuclei. [**2.2 Nuclear and electro-weak reaction rates.**]{}\ In the supernova environment, as compared to the nuclear reactions, several new important ingredients should be taken into consideration. First, the matter at stellar scales must be electrically neutral and therefore electrons should be included to balance a positive nuclear charge. Second, energetic photons present in hot matter may change nuclear composition via photo-nuclear reactions. And third, the matter is irradiated by a strong neutrino wind from the protoneutron star. We consider macroscopic volumes of matter consisting of various nuclear species with mass number $A$ and charge $Z$, $(A,Z)$, nucleons $(n=(1,0)$ and $p=(1,1))$, electrons $(e^-)$ and positrons $(e^+)$ under condition of electric neutrality. There exist several reaction types responsible for the chemical composition in supernova matter. At low densities and temperatures around a few MeV the most important ones are: 1) neutron capture and photodisintegration of nuclei, which proceed via production of a hot compound nucleus $$\begin{aligned} (A,Z)+n\rightarrow (A+1,Z)^{*}\rightarrow (A+1,Z)+\gamma~,~...\nonumber\\ (A,Z)+\gamma\rightarrow (A,Z)^{*}\rightarrow (A-1, Z)+n~,~... $$ 2) neutron and light charged particle emission (evaporation) by the hot nuclei $$(A,Z)^{*}\rightarrow(A-1,Z)+n~,~ (A,Z)^{*}\rightarrow(A-1,Z-1)+p~,~...$$ and 3) weak processes induced by electrons/positrons and neutrinos/antineutrinos $$\label{enu} (A,Z)+e^-\leftrightarrow (A,Z-1)+\nu~, ~(A,Z)+e^+\leftrightarrow (A,Z+1)+\tilde{\nu}~,$$ which transfer protons to neutrons and vice versa. There are many other reactions not shown here, which are naturally taken into account within the assumption of statistical equilibrium. The characteristic reaction times for neutron capture, photodisintegration of nuclei and nucleon emission are defined as $$\begin{aligned} \label{rate} \tau_{\rm cap}= \left[\langle\sigma_{nA}v_{nA}\rangle \rho_n\right]^{-1}~,\nonumber\\ \tau_{\gamma A}= \left[\langle\sigma_{\gamma A}v_{\gamma A}\rangle \rho_\gamma\right]^{-1}~, \nonumber\\ \tau_{n,p}=\hbar/\Gamma_{n,p}~,$$ respectively. Here $\sigma_{nA}$ and $\sigma_{\gamma A}$ are the corresponding cross sections, $v_{nA}$ and $v_{\gamma A}$ are the relative (invariant) velocities, and $\Gamma_{n,p}$ is the neutron (proton) decay width. In our calculations for $\sigma_{nA}$ we use the geometrical neutron–nuclear cross sections, that is a good approximation for the considered range of temperatures ($T\approx (0.5\div 10)$ MeV). The photo–nucleus cross section $\sigma_{\gamma A}$ was taken phenomenologically under assumption that it is dominated by the giant dipole resonance. These parametrizations of neutron and photon cross-sections are in a good agreement with experimental data (see discussion in ref. [@SMM]). The evaporation decay widths were calculated according to the Weisskopf evaporation model as described in ref. [@Botvina87]. Our estimates show that at temperatures and densities of interest these reaction times vary within the range from 10 to 10$^6$ fm/c, that is indeed very short time scale compared to the characteristic hydrodynamic time of a supernova explosion, about 100 ms [@Janka]. The nuclear statistical equilibrium is a reasonable approximation under these conditions. We have calculated the reaction rates of Eq. (\[rate\]) for nuclei with $A=$60 and $Z=$24, which are typical for stellar nucleosynthesis in dense matter and at a typical electron fraction (i.e., the ratio of electron and baryon densities $\rho_e/\rho_B$) $Y_e \sim 0.4$. They are presented in Fig. 2 as function of neutron density for several temperatures. By analyzing this figure one should take into account that the neutron density $\rho_n$ is usually by 2–5 times smaller than $\rho$. One can see clearly that for densities $\rho_n > 10^{-5}\rho_0$ and for the expected temperatures of the environment, $T {\,\raisebox{-.5ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}5$ MeV, we obtain $\tau_{\gamma A} >> \tau_{\rm cap}, \tau_{n,p}$, i.e. the photodisintegration is more slow than other processes. There exists a range of densities and temperatures, for example, $\rho_n {\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}10^{-5}\rho_0$ at $T=1$ MeV, $\rho_n {\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}10^{-3}\rho_0$ at $T=3$ MeV, and $\rho_n {\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}10^{-2}\rho_0$ at $T=5$ MeV, where the neutron capture dominates, i.e. $\tau_{\rm cap} < \tau_{n,p}$. Under these conditions new channels for production and decay of nuclei will appear (e.g. a fast break-up with emission of $\alpha$-particles or heavier clusters) which restore the detailed balance. We expect that in this situation an ensemble of various nuclear species will be in chemical equilibrium like in a liquid-gas coexistence region, as also observed in the multifragmentation reactions. Here the nuclear system is fully characterized by the temperature $T$, density $\rho$ (which is nearly the same as baryon density $\rho_B$), and electron fraction $Y_e$. One may expect that modifications of nuclear properties come into force in this environment, because of intensive interaction between clusters. This is complementary to the well known effects in isolated nuclei: at high temperature the masses and level structure in hot nuclei can be different from those observed in cold nuclei (see, e.g., ref. [@Ignat]). The weak interaction reactions are much slower. The direct and inverse reactions in Eq. (\[enu\]) involve both free nucleons and all nuclei present in the matter. It is most likely that at early stages of a supernova explosion neutrinos/antineutrinos are trapped inside the neutrinosphere around a protoneutron star [@Prakash]. In this case we should impose the lepton number conservation condition by fixing the lepton fraction $Y_L$. Then one should take into account the continuous neutrino flux out of the surface of the neutrinosphere propagating through the hot bubble. Due to large uncertainties in the weak interaction rates, below we consider three physically distinctive situations: 1\) fixed lepton fraction $Y_L$ corresponding to a $\beta$-equilibrium with trapped neutrinos inside the neutrinosphere (early stage of the explosion); 2\) fixed electron fraction $Y_e$ but no $\beta$-equilibrium inside a hot bubble (early and intermediate times); 3\) full $\beta$-equilibrium without neutrino (the late times of the explosion, after neutrino escape). The second case corresponds to a non-equilibrium situation which may take place in the bubble, before the electron capture becomes efficient. Actually, this case is considered as basic for calculations of nuclear composition in the hot bubble behind the shock. Generally, one should keep in mind that weak reactions are often out of equilibrium. Our estimates show that their characteristic times range from 10 ms to 10 s depending on thermodynamical conditions and intensity of the neutrino wind. Therefore, one should specify what kind of statistical equilibrium is expected with respect to weak interaction. [**3. Formulation of the statistical model.**]{}\ Below we describe supernova matter as a mixture of nuclear species, electrons, photons, and perhaps neutrinos in thermal equilibrium. For the macroscopic scales one can safely apply the grand-canonical approximation. We call this model the Statistical Model for Supernova Matter (SMSM). It was first proposed in ref.[@Botvina04]. [**3.1 Equilibrium conditions.**]{}\ Within the SMSM each particle $i$ with baryon number $B_i$, charge $Q_i$ and lepton number $L_i$ is characterized by a chemical potential $\mu_i$, which can be represented as $$\mu_i=B_i\mu_B+Q_i\mu_Q+L_i\mu_L .$$ Here $\mu_B$, $\mu_Q$ and $\mu_L$ are three independent chemical potentials which are determined from the conservation of total baryon number $B=\sum_iB_i$ electric charge $Q=\sum_iQ_i$ and lepton number $L=\sum_iL_i$ of the system. Explicitly, the chemical potentials for nuclear species $(A,Z)$, electrons $(e^{-}, e^{+})$, and neutrinos $(\nu, \tilde{\nu})$ can be expressed as $$\label{chem} \begin{array}{ll} \mu_{AZ}=A\mu_B+Z\mu_Q~,~\\ \mu_{e^-}=-\mu_{e^+}=-\mu_Q+\mu_L~,~\\ \mu_\nu=-\mu_{\tilde{\nu}}=\mu_L~. \end{array}$$ These relations are valid also for nucleons, $\mu_n=\mu_B$ and $\mu_p=\mu_B+\mu_Q$. The corresponding conservation laws can be written as $$\begin{aligned} \rho_B=\frac{B}{V}=\sum_{AZ}A\rho_{AZ}~,~\nonumber\\ \rho_Q=\frac{Q}{V}=\sum_{AZ}Z\rho_{AZ}-\rho_e=0~,~\\ \rho_L=\rho_e +\rho_{\nu} -\rho_{\tilde{\nu}}=Y_L \rho_B ~.\nonumber $$ Here $\rho_e=\rho_{e^-}-\rho_{e^+}$ is the net electron density, $Y_L$ is the lepton fraction. The second equation requires that any macroscopic volume of the star is electrically neutral. The lepton number conservation is a valid concept only if $\nu$ and $\tilde{\nu}$ are trapped in the system within the neutrinosphere [@Prakash]. If they escape freely from the system, the lepton number conservation is irrelevant and $\mu_L=0$. In this case two remaining chemical potentials are determined from the conditions of baryon number conservation and electro-neutrality. Since the $\beta-$ equilibrium may not be achieved in a fast explosive process, we also often fix the electron fraction $Y_e=\rho_e / \rho_B$ in this case. [**3.2 Ensemble of nuclear species.**]{}\ Our treatment of nuclear reactions is based on the statistical multifragmentation model (SMM) [@SMM], which was very successfully applied for description of experimental data. For describing an ensemble of nuclear species under supernova conditions we use the Grand Canonical version of the SMM [@Botvina85]. After integrating out translational degrees of freedom one can represent pressure of nuclear species as $$\begin{aligned} \label{naz} P_{\rm nuc}=T\sum_{AZ}\rho_{AZ} \equiv T\sum_{AZ}g_{AZ}\frac{V_f}{V}\frac{A^{3/2}}{\lambda_T^3} {\rm exp}\left[-\frac{1}{T}\left(F_{AZ}-\mu_{AZ}\right)\right]~, $$ where $\rho_{AZ}$ is the density of nuclear species with mass $A$ and charge $Z$. Here $g_{AZ}$ is the ground-state degeneracy factor of species $(A,Z)$, $\lambda_T=\left(2\pi\hbar^2/m_NT\right)^{1/2}$ is the nucleon thermal wavelength, $m_N \approx 939$ MeV is the average nucleon mass. $V$ is the actual volume of the system, and $V_f$ is so called free volume, which accounts for the finite size of nuclear species. We assume that all nuclei have normal nuclear density $\rho_0$, so that the proper volume of a nucleus with mass $A$ is $A/\rho_0$. At low densities the finite-size effect may be included via the excluded volume approximation $V_f/V \approx \left(1-\rho_B/\rho_0\right)$. We emphasize that this last approximation is commonly accepted in statistical models. As we know from nuclear multifragmentation studies information about free volume can be extracted from analysis of experimental data [@EOS]. At densities $\rho_B > 0.1 \rho_0$ the extracted $V_f$ may slightly deviate from above approximation, however, they are in qualitative agreement. In the present work we remain in the framework of the conventional statistical approach, although allowing for modifications of nuclear properties. The internal excitations of nuclei play an important role in regulating their abundance, since they increase significantly their entropy. Some authors (see, e.g., ref. [@Japan]) limit the excitation spectrum by particle-stable levels known for low excited nuclei. Within the SMM we follow quite different philosophy. Namely, we calculate internal excitation of nuclei by assuming that they have the same internal temperature as the surrounding medium. In this case not only particle-stable states but also particle-unstable states will contribute to the excitation energy and entropy. This can be justified by the dynamical equilibrium of nuclei in hot environment, and supported by numerous comparisons with experiment (see part [**2.1**]{}). Moreover, in the supernova environment both the excited states and the binding energies of nuclei will be strongly affected by the surrounding matter. By this reason, we find it more appropriate to use an approach which can easily be generalized to include in-medium modifications. Namely, the internal free energy of species $(A,Z)$ with $A>4$ is parameterised in the spirit of the liquid drop model, which has been proved to be very successful in nuclear physics [@Bohr]: $$F_{AZ}(T,\rho)=F_{AZ}^B+F_{AZ}^S+F_{AZ}^{\rm sym}+F_{AZ}^C~~.$$ Here the right hand side contains, respectively, the bulk, the surface, the symmetry and the Coulomb terms. The first three terms are taken in the standard form [@SMM], $$\begin{aligned} F_{AZ}^B(T)=\left(-w_0-\frac{T^2}{\varepsilon_0}\right)A~~, \\ F_{AZ}^S(T)=\beta_0\left(\frac{T_c^2-T^2}{T_c^2+T^2}\right)^{5/4}A^{2/3}~~,\\ \label{fres} F_{AZ}^{\rm sym}=\gamma \frac{(A-2Z)^2}{A}~~,\end{aligned}$$ where $w_0=16$ MeV, $\varepsilon_0=16$ MeV, $\beta_0=18$ MeV, $T_c=18$ MeV and $\gamma=25$ MeV are the model parameters which are extracted from nuclear phenomenology and provide a good description of multifragmentation data [@SMM; @ALADIN; @EOS; @MSU; @INDRA; @FASA; @Dag]. However, these parameters, especially the symmetry coefficient $\gamma$, can be different in hot nuclei at multifragmentation conditions, and they should be determined from corresponding experimental observables (see discussion in refs. [@Botvina06; @bulk; @traut]). In the electrically-neutral environment the nuclear Coulomb term should be modified to include the screening effect of electrons. Within the Wigner-Seitz approximation with constant electron density it can be expressed as $$\begin{aligned} F_{AZ}^C(\rho)=\frac{3}{5}c(\rho)\frac{(eZ)^2}{r_0A^{1/3}}~~,\\ c(\rho)=\left[1-\frac{3}{2}\left(\frac{\rho_e}{\rho_{0p}}\right)^{1/3} +\frac{1}{2}\left(\frac{\rho_e}{\rho_{0p}}\right)\right]~,\nonumber $$ where $r_0=1.17$ fm and $\rho_{0p}=(Z/A)\rho_0$ is the proton density inside the nuclei. The screening function $c(\rho)$ is 1 at $\rho_e=0$ and 0 at $\rho_e=\rho_{0p}$. Here one can also use an approximation $\rho_e/\rho_{0p}=\rho_B/\rho_0$, as in ref. [@Lamb]. In our calculations, we have checked that these two choices lead to very similar results, especially at small densities. We want to stress that both the reduction of the surface energy due to the finite temperature and the reduction of the Coulomb energy due to the finite electron density favour the formation of heavy nuclei. Nucleons and light nuclei $(A \leq 4)$ are considered as structure-less particles characterized only by exact masses and proper volumes [@SMM]. Their Coulomb interaction is taken into account within the same Wigner-Seitz approximation. As follows from Eq. (\[naz\]), the fate of heavy nuclei depends strongly on the relationship between $F_{AZ}$ and $\mu_{AZ}$. In order to avoid an exponentially divergent contribution to the baryon density, at least in the thermodynamic limit ($A \rightarrow \infty$), inequality $F_{AZ}{\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}\mu_{AZ}$ must hold. The equality sign here corresponds to the situation when a big (infinite) nuclear fragment coexists with the gas of smaller clusters [@Bugaev]. When $F_{AZ}>\mu_{AZ}$, only small clusters with nearly exponential mass spectrum are present. However, there exist a region of thermodynamic quantities corresponding to $F_{AZ}\approx\mu_{AZ}$ when the mass distribution of nuclear species is a power-law $A^{-\tau}$ with $\tau \approx 2$. The advantage of our approach is that we consider all the fragments present in this transition region, contrary to the previous calculations [@Lamb; @Lattimer], which consider only one “average” nucleus characterizing the liquid phase. [**3.3 Electromagnetic and weak processes.**]{}\ At $T,\mu > m_e$ the pressure of the relativistic electron-positron gas can be written as $$\begin{aligned} P_e=\frac{g_e\mu_e^4}{24\pi^2}\left[1+2\left(\frac{\pi T}{\mu_e}\right)^2+ \frac{7}{15}\left(\frac{\pi T}{\mu_e}\right)^4-\frac{m_e^2}{\mu_e^2} \left(3+\left(\frac{\pi T}{\mu_e}\right)^2\right)\right] , $$ where first-order correction ($\sim m_e^2$) due to the finite electron mass is included, $g_e$=2 is the spin degeneracy factor. The net number density $\rho_e$ and entropy density $s_e$ can be obtained from standard thermodynamic relations, $\rho_e=\partial P_e/\partial \mu_e$, and $s_e=\partial P_e/\partial T$, which give $$\begin{aligned} \label{eden} \rho_e=\frac{g_e}{6\pi^2}\left[\mu_e^3+\mu_e\left(\pi^2T^2- \frac{3}{2}m_e^2\right)\right]~,\\ s_e= \frac{g_e T\mu_e^2}{6}\left[1+\frac{7}{15}\left(\frac{\pi T}{\mu_e}\right)^2- \frac{m_e^2}{2\mu_e^2}\right]~. $$ Electron neutrinos and antineutrinos are taken into account in the same way, but as massless fermions, and with the degeneracy factor twice smaller than for the electrons, i.e., $g_{\nu}$=1. The photons are always close to the thermal equilibrium, and they are treated as massless Bose gas with zero chemical potential. The corresponding density $\rho_{\gamma}$, energy density $e_{\gamma}$, pressure $P_{\gamma}$, and entropy density $s_{\gamma}$ of photons gas are given by standard formulae: $$\rho_{\gamma}=\frac{g_{\gamma}\xi(3) T^3}{\pi^2 \hbar^3 c^3}~,~ e_{\gamma}=\frac{g_{\gamma}\pi^2 T^4}{30 \hbar^3 c^3}~,~ P_{\gamma}=\frac{e_{\gamma}}{3}~,~ s_{\gamma}=\frac{4 e_{\gamma}}{3 T},$$ where $g_{\gamma}$=2. All kinds of particles (nuclei, baryons, electrons, neutrinos, photons) contribute to the free energy, pressure and other thermodynamical characteristics of the system, and we sum up all these contributions. Within the model we calculate densities of all particles self-consistently by taking into account the relations between their chemical potentials. [**3.4 Comparison with the Lattimer-Swesty model.**]{}\ We have performed calculations for sets of physical conditions expected during the collapse of massive stars and subsequent supernova explosions. We take baryon number $B=$1000 and perform calculations for all fragments with 1$\leq A \leq$1000 and 0$\leq Z \leq A$ in a box of fixed volume $V$. This volume is determined by the average baryon density $\rho_B=B/V$. This restriction on the size of nuclear fragments is fully justified in our case, since fragments with larger masses ($A>$1000) can be produced only at very high densities $\rho {\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}0.5\rho_0$ [@Bethe; @Lamb], which are appropriate for the regions deep inside the protoneutron star, and which are not considered here. In the beginning we compare our model with calculations within other models on the marked. It is necessary to mention that all models treat electrons and photons in the same way, therefore, differences appear entirely due to different description of nuclear species. One can expect that in the case of domination of radiation processes the results will be similar. Most supernova simulations are performed with the equation of state of Lattimer–Swesty (LS) [@Lamb; @Lattimer], where an ensemble of heavy nuclei is replaced by a single “average” nucleus. Therefore, one can compare only integral characteristics of the stellar matter. In Fig. 3 we compare our SMSM results with the LS model. We show entropies, pressure, mass fractions of alpha particles ($X_{alpha}$), and heavy fragments ($X_{heavy}$), at different temperatures, and at a fixed electron fraction, versus densities. In the SMSM $X_{heavy}$ includes all fragments with $A>$4, whereas in the LS model it is only the share of the “average” heavy nucleus. The LS calculations were taken from ref. [@Janka2]. One can see that the average thermodynamical characteristics (pressure, entropy) are very close in the two models. This remains also true if we extend comparison to other models, e.g., reported in [@Janka2; @Japan]. However, as seen from Fig. 3, mass fractions of nuclei are very different. As was mentioned in [@Janka2], a small yield of alpha clusters in the LS model may be caused by mistakes in calculations of the Coulomb corrections to their binding energies. We stress again that the details of the nuclear composition are very important for dynamics of the explosion, since it influences the total energy balance, and determines the weak reaction rates. [**4. Composition of matter.**]{}\ In this section we present the SMSM results concerning nuclear and lepton composition of stellar matter, which are important for determining general thermodynamical characteristics of the matter, such as energy density, pressure, and entropy. The energy deposition into the matter with photons and neutrinos produced by external sources is also considerably influenced by this composition. We pay special attention to the contribution of nuclear species whose properties may be modified in dense environment as follows from recent findings in nuclear multifragmentation reactions. [**4.1 Electron and neutron fractions of stellar matter.**]{}\ Within the SMSM we can calculate the electron fraction in the electrically-neutral matter under assumption of full $\beta$-equilibrium. The appropriate astrophysical sites, where this may happen, are the relatively slow collapse stage, and the very late stages of the explosion, after cooling down the matter and neutrino escape. In this case we calculate self-consistently densities of all species by using relations between their chemical potentials, Eq. (\[chem\]), with $\mu_L$=0. The net electron density, which is equal to the proton density, is explicitly given as a function of the chemical potential in Eq. (\[eden\]). Fig. 4 presents the fractions of free electrons and neutrons as functions of baryon density. On the left panels the results are shown for the $\beta$-equilibrium neutrino-less matter at $T=$1 and 3 MeV. One can clearly see two general trends: with increasing baryon density the electron fraction $Y_e$ gradually decreases and the neutron fraction $Y_n$ increases. At small densities, which correspond to low $\mu_e$, the electrically-neutral matter tends to be isospin symmetric, with a large amount of electrons. In the case of low temperatures ($T{\,\raisebox{-.5ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}1$ MeV) the protons are captured in most bound nuclei with $A\sim$ 50–60. As was realized long time ago [@Bethe], at large densities the electrons are absorbed by protons in the inverse $\beta$-decay process, that is driven by high electron chemical potential. When we increase temperature ($T {\,\raisebox{-.5ex}{$\stackrel{>}{\scriptstyle\sim}$}\,}3$ MeV) the nuclei dissociate into protons and neutrons, that helps to capture electrons at large densities also. At the same time, the number of free neutrons increases rapidly at higher baryon densities for both low and high temperatures. This is important for maintaining a high rate of nuclear reactions to generate equilibrium ensemble of nuclei. At low temperatures ($T \sim 1$ MeV) a noticeable change in the trend is seen below $\rho_B\approx 10^{-4}\rho_0$. At these densities most neutrons are bound in large nuclei, which are still present in the matter. For example, $Y_n \approx$ 0.2 means that 80% of neutrons are trapped in the nuclei. At even lower densities, when heavy nuclei disappear (see Fig. 3) the number of free neutrons increases. The same trend at higher densities is explained by the fact that more and more neutrons are dripping out of nuclei, since the matter contains less and less protons. For example, at $\rho_B>10^{-3}\rho_0$ more than half of the neutrons are free. This behavior correlates with decreasing the number of electrons and a relatively small share of heavy nuclei in the system. At higher densities, the structure of matter may change because of the neutrino/antineutrino and electron/positron capture reactions [@Bethe]. The right panel of Fig. 4 shows the results obtained under the condition of lepton number conservation, i.e., at fixed values of lepton fractions $Y_L$=0.1, 0.2, 0.3. These values are consistent with uncertainties concerning the neutrinosphere radius discussed in literature. One can see that in this case the number of free neutrons always drops with density reflecting formation of very big nuclei and transition to the liquid phase at $\rho_B \rightarrow \rho_0$. While the electron fraction stays nearly constant, exhausting around 80–90% of the total $Y_L$. These results show that weak reactions affect significantly the composition of supernova matter. [**4.2 Mass fractions of light and heavy nuclei.**]{}\ As well known, at low densities and temperatures the nuclear matter exists in the form of isolated nuclei and nucleons. At terrestrial conditions the nuclei capture electrons and become atoms. However, at supernova conditions the atoms are fully ionized, therefore, the nuclei are embedded in more or less uniform background of electrons and neutrons. This surrounding to a large extent determines the nuclear composition of stellar matter. Figures 5 and 6 demonstrate the mass fractions of nuclear matter contained in heavy nuclei (with mass numbers $A > 4$), $\alpha$-particles, neutrons and protons, for different electron fractions $Y_e$. One can see that at low temperatures ($T < 1$ MeV) the matter is mainly composed of heavy nuclei. If the share of electrons is small, the free neutrons are also present. With increasing temperatures the heavy nuclei gradually disintegrate into $\alpha$’s, neutrons and protons. At low densities this disintegration happens already at moderate temperatures $T \sim$ 1–2 MeV, while at subnuclear densities ($\rho \sim 0.1 \rho_0$) the heavy nuclei survive even at higher temperatures, though they become very excited. One should bear in mind, however, that in this case we are dealing with the dynamical equilibrium between decay of excited nuclei and absorption of surrounding nucleons, as regulated by the reaction rates presented in Fig. 2. An interesting observation is that heavy nuclei first break-up into light clusters (like $\alpha$) and then these clusters dissolve into nucleons with temperature. This is clear seen in the yields of $\alpha$–particles, which demonstrate a ’rise and fall’ behaviour both with increasing temperature and decreasing density. As seen from Figs. 7 and 8, small clusters with $A$=2 and 3 are also produced in the transition region. [**4.3 Nuclear mass distributions.**]{}\ The properties of heavy nuclei are very important for understanding processes taking place in stellar matter. In the bottom panel of Fig. 7 we show mass distributions of nuclear species at $\rho_B = 10^{-3} \rho_0$ and several temperatures. At low temperatures the distribution of heavy nuclei looks like a Gaussian with a well defined peak (see also ref. [@Botvina04]). In this case the average thermodynamic characteristics of the system may not be much different from the ones calculated under assumption of an “average” nucleus as in ref. [@Lattimer], see Fig. 3. However, even in this case, the width of the distributions may be important for calculations of weak reactions in matter. By increasing temperature we move into a coexistence region of the nuclear liquid-gas phase transition: The mass distributions become ’U-shaped’, and they contain all nuclei from light to heavy ones. At higher temperatures the mass distributions have exponential shape. These distributions can not be even approximately characterized by an “average” nucleus. This evolution of fragment mass distributions is well established in nuclear multifragmentation reactions [@SMM; @Dag]. The average charge-to-mass ratio for all nuclei is demonstrated in the top panel of Fig. 7. For this quantity both the symmetry and Coulomb energies of fragments are crucially important. The charges of heavy fragments show rather regular behaviour: The $Z/A$ ratio decreases slowly with mass number because of Coulomb interaction. It decreases less rapidly than in multifragmentation reactions due to the screening effect of electrons. From our calculations we came to the conclusion that the charge distribution of nuclei at given $A$ can be approximated by a Gaussian, with the width $\sigma_Z \approx \sqrt{AT/8\gamma}$ [@Botvina87; @Botvina85; @Botvina01], where $\gamma$ is the coefficient in the symmetry energy (see Eq. (\[fres\])). We have found that the phase transition from heavy nuclei to light fragments (nuclear gas) always proceeds through the same sequence of mass distributions: ’U-shape’, power-law, and exponential ones, both with increasing temperature and decreasing density. This opens the possibility to study the critical behaviour in stellar matter, in the same way as was previously done in multifragmentation reactions [@EOS; @Dag]. One can find examples of the temperature-driven transitions in refs. [@Botvina04; @Botvina05] and in Fig. 7 (see also Fig. 22). In Fig. 8 we demonstrate an example of the density-driven transition at fixed temperature of 1 MeV. As we can clearly see, the mass distributions evolve from ’U-shape’ at densities (in units $10^{-5}\rho_0$) 1.0 and 0.32 to power-law at 0.18 and 0.1, and finally to exponential at 0.03. At density 0.18 the distribution of large clusters is most flat, that may be considered as a critical point of the phase transition [@Bugaev]. [**4.4 Shell effects.**]{}\ Up to now we have used the liquid-drop description of nuclei in the stellar matter. However, it is well known that at low temperatures ($T{\,\raisebox{-.5ex}{$\stackrel{<}{\scriptstyle\sim}$}\,}1$ MeV) the shell corrections to nuclear masses becomes important, at least, at terrestrial densities of matter. It is likely that the shell effects may also play a role in stellar matter at subnuclear densities. In this section we demonstrate how shell effects may influence the mass and charge distributions of nuclei. In particular, we have analyzed possible existence of superheavy elements in stellar matter, assuming that they are sufficiently long-lived. It was assumed that there is a certain shell correction to the free energy $F_{AZ}$ of a specific nucleus, $\Delta F_{AZ}$, taken as follows: $$\Delta F_{AZ}=-\left[E_{sh}exp \left(-\frac{(N-N_{sh})^2}{2\sigma_N^2}\right)+ E_{sh}exp \left(-\frac{(Z-Z_{sh})^2}{2\sigma_Z^2}\right)\right].$$ We have assumed existence of the island of stability around $Z_{sh}$=120 and $N_{sh}$=180 [@Greiner]. The maximum energy of the shell correction is chosen to be $E_{sh}$=5 MeV, and the widths of the shell are $\sigma_{N}= \sigma_{Z}=$5 MeV. These values are consistent with the magnitudes of the shell effects discussed for superheavy elements. In Fig. 9 we compare the calculated charge and mass distributions without and with the shell corrections (i.e., without and with the term $\Delta F_{AZ}$). We consider typical conditions in supernova matter where the production of superheavy nuclei is still possible, i.e., $T=$1 MeV, $\rho_B \approx 0.05\rho_0$, $Y_e=0.2$. One can see that a pure liquid-drop description predict Gaussian-like mass and charge distributions of fragments which move to smaller values with decreasing density. When the shell constraint is imposed, the yields become essentially, by factor 2–5, larger in the vicinity of neutron ($N_{sh}$) and proton ($Z_{sh}$) ’magic’ numbers. Moreover, because of the increased binding, these magic nuclei are abundantly produced in a rather broad region of density. For example, the neutron shell of $N_{sh}$=180 dominates clearly at densities of both $0.03\rho_0$ and $0.05\rho_0$. This result makes possible to discuss nucleosynthesis of superheavy elements at supernova conditions. As has been already pointed out in ref. [@Botvina04], there is a chance that heavy nuclei could be produced at subnuclear densities, and then ejected into the space. At considered small temperature (1 MeV) the effect of secondary deexcitation will be minimal, only few nucleons will be lost. The fission channels for these neutron-rich nuclei may also be suppressed by the shell effects [@Nix] and electron screening [@Burven]. However, one can expect that these nuclei will fastly emit neutrons above the neutron drip-line, undergo abundant $\beta$ decay and, possibly, $\alpha$ emission. We are planning to analyze all these processes in the forthcoming publications [@s-heavy]. Here we mention only a possible scenario how such superheavy nuclei can be ejected into space. Since the synthesis of heavy and superheavy nuclei is only possible at rather high baryon density $\rho_B \sim 0.05\rho_0$ it is most likely that such nuclei will not be ejected in the course of the supernova explosion. Instead, they will be accumulated at the surface of a newly produced neutron star. If this star is in a binary system with another neutron star, white dwarf, or even a black hole, there is a chance of their collision at later stages of evolution. Then a part of the stellar material will be ejected in space, while the other part may collapse into a black hole. One can also speculate about asymmetric explosions of supernovas, acceleration of nuclei by the neutrino wind, and starquakes which may provide this ejection (see, e.g., refs. [@star-collide; @neutrino-wind; @star-quakes]). The search for new mechanisms of nucleosynthesis is motivated by the fact that the traditional $s$- and $r$-processes have serious problems to explain synthesis of fissioning nuclei larger than lead [@Qian]. [**5. Thermodynamical characteristics of stellar matter.**]{}\ In this section we present general thermodynamical characteristics of stellar matter, such as energy density, pressure and entropy as functions of temperature $T$, baryon density $\rho_B$, electron fraction $Y_e$, as well as the nuclear composition. In hydrodynamical simulations most important role is played by Equation of State (EOS), which connects pressure with the energy density. The dynamics of collapse and explosion depends essentially on the EOS [@Janka]. [**5.1 Caloric curve, pressure and entropy**]{}\ One of the main inputs for dynamical simulations of supernova explosions is the total thermal energy deposited in the matter. In Fig. 10 we show so-called caloric curves, i.e., the thermal energy per nucleon of the matter as a function of temperature. For convenience, the actual energy per nucleon is shifted by the value of 16 MeV, which corresponds to the bulk binding energy of nuclear matter at $\rho_B = \rho_0$. Our calculations show that nuclear contributions dominate at high densities and low temperatures, where heavy nuclei survive. Due to the internal excitation of these nuclei according to the compound nucleus law, $E^{*} \sim T^2$, the caloric curve has a parabolic shape at $\rho_B \sim 0.1-0.001 \rho_0$. However, at low densities and high temperatures electrons and photons dominate, and the caloric curve behaves according to the Stefan–Boltzmann law $E^{*} \sim T^4$, i.e., the energy per nucleon grows very rapidly with temperature. One can also see that the nuclear contributions become nearly independent on baryon density at high temperatures. Namely, they approach the Boltzmann limit, $E^{*} \sim 1.5T$, when nuclear matter disintegrate completely into nucleons. It is instructive to note that at high densities $\rho \sim 0.1\rho_0$ and proton fractions $Y_p=Y_e \sim 0.4$ which are typical for normal nuclei, the caloric curve is determined mainly by the nuclear species, and it reminds very much the caloric curves extracted from the multifragmentation reactions [@SMM]. In this case, the excitation energy increases rapidly at temperatures $T \approx 4-6$ MeV, which correspond to the maximum in the heat capacity. This is a characteristic feature of the liquid-gas phase transition [@Dag; @Bugaev]. Pressure is another important characteristic of the matter, which, in competition with the gravitational pressure, determines the structure and dynamics of the stellar system. In Figs. 11 and 12 we plot the pressure versus baryon density for different temperatures. There are several important features to be mentioned. First, the nuclear contributions are mainly important in the intermediate density region $(10^{-4}-10^{-2})\rho_0$, where, depending on temperature, more and more free nucleons are present in the nuclear matter. In the case of full disintegration of nuclear species into nucleons, the nucleons may contribute up to 50% to the total pressure. Second, at higher densities the pressure is dominated by the relativistic electrons, since their chemical potential becomes very high. Third, at very low densities and high temperatures the radiation pressure dominate, which is proportional to $T^4$ and does not depend on baryon density. One can see that the nuclear pressure is higher in the case of a low electron fraction ($Y_e$=0.2), because of a considerable abundance of free neutrons, even if large nuclei are present in the matter. In the modern hydrodynamical simulations of supernova explosions the shock stalls at densities around $10^{-6}-10^{-5}\rho_0$ [@Janka]. This happens partly because a significant fraction of shock energy is used for disintegration of infalling nuclei, from C to Fe. Therefore, survival of medium and heavy nuclei would contribute essentially to the revival of the shock. The entropy per baryon $S/B$, which is shown in Fig. 13 as a function of temperature, is an important characteristic of the exploding matter. One can notice that it correlates strongly with behavior of the caloric curve. At low temperatures and high densities the nuclear contribution to the entropy dominates. At high temperatures, the nuclei disintegrate into nucleons and the nuclear entropy depends only logarithmically on temperature and density according to the Boltzmann gas law. Usually, at entropy greater than 10 units per nucleon only nucleon gas without heavy nuclei is present in the system. The main contribution to the total entropy in this case is provided by the radiation and electron-positron pairs. This contribution does not depend on density and is proportional to $T^3$. The total entropy has a jump across the shock, which can be explained in part by disintegration of heavy nuclei. Therefore, even small differences in nuclear properties in medium, in comparison with isolated nuclei, may lead to significant effects in the shock dynamics. [**5.2 Adiabatic trajectories.**]{}\ Some important processes in stellar matter proceed with approximately constant entropy. For example, the collapse of a massive star before re-bounce is characterized by the entropy around one unit per nucleon. After propagation of the shock wave the entropy increases drastically. However, subsequent evolution of the matter is again close to isentropic, and this assumption is quite valid for nucleosynthesis. For this reason we consider adiabatic trajectories in the $T - \rho_B$ plane, and calculate fragment mass distributions and thermodynamical functions along these trajectories. In Fig. 14 we show adiabatic trajectories for several fixed S/B. This representation of the phase diagram on $T - \rho$ plane is very convenient for understanding thermodynamical properties of stellar matter (compare also with Fig. 1). It is also possible to make a rough estimation of nuclear composition of the matter: As well known from previous studies [@Bethe], at S/B=1 many heavy nuclei exist in the matter, while at the entropy as high as $S/B$=20 the baryonic matter consist mainly of free nucleons and hadron resonances. Our calculations confirm this expectation. It is instructive to compare mass fractions of different nuclear species along the adiabats for different thermodynamical conditions, in order to get an idea about evolution of nuclei during the whole collapse–explosion process. In Fig. 15 we show these fractions (similar to Fig. 6) for S/B=1, where the mass fraction of heavy nuclei is close to 1, and for S/B=8, where the nuclei undergoes deep disintegration. In the latter case the fraction of nuclei with $A>4$ is essential only at high densities, and there are no large fragments at low densities. At S/B=8 the fraction of $\alpha$ particles has a very interesting behavior: It has a minimum at intermediate densities ($\rho_B \sim 10^{-3}\rho_0$), but then it increases with decreasing density, and completely dominates at low densities. For this fixed entropy a large binding energy of $\alpha$ particles is very important for the thermodynamical balance in the system. Therefore, $\alpha$ particles can be preferable instead of individual nucleons at the low density conditions. The production of $\alpha$ clusters may help to revive the shock wave by maintaining a sufficiently high temperature behind the shock. The mass distributions of nuclei evolve strongly along the isentropic trajectories. In Fig. 16 we demonstrate this evolution in stellar matter with a large electron fraction $Y_e =0.4$ for two cases corresponding to a low entropy S/B=1, and to a higher entropy S/B=4, where the contribution of heavy fragments is still essential. It is important that in the first case (S/B=1) at high baryon density ($\rho_B = 0.1\rho_0$) and temperature ($T=$3.39 MeV), the distribution of heavy nuclei is centred around $A \sim$130 and has a large width $\sigma_A \sim 50$. Therefore, heavy nuclei with mass number higher than 200, as well as very small clusters, coexist in the system. With decreasing density the distribution of nuclei shifts to smaller masses and becomes more narrow. Finally, the distribution moves into the iron region, where the binding energy is maximal. At higher entropy (S/B=4) we have an opposite situation. The masses of produced nuclei become larger with decreasing density. And at very low densities the nuclei reach finally the iron region. The reason for this interesting behavior is that the temperature drops essentially (from 10 to 1 MeV, as seen from Fig. 14) along the adiabatic trajectory. All above discussed trends take also place in the case of isospin asymmetric matter with a small electron fraction, $Y_e=0.2$, see Fig. 17. Besides of the expected effect of increasing number of free neutrons, we obtain here that more heavy nuclei are produced at the same densities. In the same time these nuclei are very neutron-rich. For example, at S/B=1 and $\rho=0.1\rho_0$, for $Y_e=0.2$ we have an average charge of big nuclei $\langle Z \rangle /A \approx 0.28$, in comparison with $\langle Z \rangle /A \approx 0.41$ for $Y_e=0.4$. Because of formation of the unusual heavy nuclei, the temperature is lower for small electron fractions. One can conclude from analysis of Figs. 15, 16, and 17 that clusterization of nuclear matter at low densities should have important consequences for the explosion, at least via the energy balance. However, this effect will be even stronger if we include into consideration the modifications of the weak reaction rates caused by the clustering (see below). In Fig. 18 we demonstrate the total adiabatic pressure as function of baryon density. As expected, it increases with density, and with entropy. One can see a nearly linear relation between ln(P) and ln($\rho$). For a fixed entropy this relation is usually expressed as the politropic equation $$P \sim \rho^{\Gamma_{ad}},$$ where $\Gamma_{ad}$ is an effective adiabatic index. In Fig. 19 we show the behavior of $\Gamma_{ad}$ as a function of density for two values of entropy per baryon. At S/B=1, when heavy nuclei still exist, the adiabatic index is nearly constant and close to 4/3, which is expected for relativistic electron gas. At S/B=8 the $\Gamma_{ad}$ coefficient shows more interesting behavior: it changes considerably and goes through the maximum. The maximal value of 1.5 is reached around $10^{-4}-10^{-2}\rho_0$. This corresponds to a change of the pressure slope seen in Fig. 18. This effect is caused by a nearly complete disintegration of nuclei into nucleons, which takes place in this density region (see Fig. 15). The production of heavy nuclei at higher densities, and production of $\alpha$ particles at lower densities, lead to decreasing $\Gamma_{ad}$ at the both side of the maximum. Matter with $\Gamma_{ad} < 4/3$ can not resist the gravity and, therefore, unstable with respect to gravitational collapse. On the other hand, matter with $\Gamma_{ad} > 4/3$ can provide conditions for outward propagation of the shock wave during the supernova explosion. Finally, let us consider now the adiabatic sound velocity, $c_s^2=\partial P/ \partial \rho |_s$, which plays an important role in hydrodynamical simulations. For example, when the collective velocity of matter exceeds $c_s$ a shock wave is generated. In Fig. 20 we present the sound velocity along different adiabates. In this case it can be obtained as $c_s^2=\Gamma_{ad}\cdot P/\rho$. As expected, $c_s$ increases with density. However, in the case of S/A=8, there is a peak around $10^{-2}\rho_0$, which is caused by the same physical reasons as the maximum of $\Gamma_{ad}$, shown in Fig. 19. We note that a small sound velocity in comparison with the light velocity give a justification for using nonrelativistic formulas for baryonic and nuclear degree of freedom. [**6. Possible in-medium modification of nuclear properties.**]{}\ [**6.1 Reduction of symmetry energy.** ]{}\ As we have mentioned in section [**3**]{} multifragmentation reactions open a unique possibility to investigate clusterization of nuclear matter at subnuclear densities. Recently, the symmetry energy of hot nuclei was extracted from experimental data [@Botvina05; @LeFevre; @Iglio; @Souliotis], and it was demonstrated that the $\gamma$ coefficient, see Eq. (\[fres\]), is considerably reduced as compared with the values expected for cold isolated nuclei. This effect becomes stronger with increasing excitation energy. For example, in Fig. 21 we show the extracted values of the $\gamma$ coefficient, which go significantly down with decreasing impact parameter $b$, i.e., for more and more central collisions. The empirical value of the $\gamma$ coefficient, approximately $25$ MeV, was obtained for isolated nuclei from the liquid-drop description of their binding energies. As one can see, at high excitation energies it drops down to $\approx 15$ MeV, and may even lower, as follows from the analysis of ref. [@LeFevre]. As discussed in refs. [@traut; @LeFevre; @Iglio; @Souliotis; @Botvina06; @bulk] this change can be explained by the reduction of the fragment density, modification of the nuclear surface energy, and the influence of nuclear environment. Now we come back to the comparison of conditions which can be reached in multifragmentation reactions and in supernova explosions. Specifically, in Fig. 22 we demonstrate the similarity of fragment mass distributions for these two physical systems, calculated within the SMM and SMSM using the same description of hot fragments in dense environment. The density and temperature of stellar matter are chosen close to the ones typical for multifragmentation reactions. The electron fraction ($Y_e=$ 0.2) corresponds approximately to the deleptonization values obtained at this density. One can see that the evolution of mass distributions with excitation energy and temperature is qualitatively similar for both cases (see also section [**4.3**]{} and refs. [@SMM; @Botvina08]). The transition from the ’U-shaped’ mass distribution to the exponential one, a characteristic feature of the liquid-gas phase transition, is well pronounced in both cases too. However, in the supernova environments much heavier and neutron-rich nuclei can be produced because of screening effect of surrounding electrons. In this figure we also demonstrate how important is to extract reliable information about the symmetry energy of hot nuclei. As one can see from mass yields at 3 MeV per nucleon in top panel, changing $\gamma$ coefficient from 25 to 15 MeV has practically no effect on the mass distributions of fragments produced in nuclear reactions. As was shown in refs. [@Botvina06; @nihal], in nuclear multifragmentation reactions the most noticeable effect is that the isotope distributions become broader at smaller $\gamma$. However, the symmetry coefficient $\gamma$ has a dramatic influence on masses of nuclei produced in supernova environments. One can see from Fig. 22 that much more heavy (and more neutron-rich) nuclei can be formed in this case. This effect makes very likely production of heavy and superheavy nuclei in supernova environments. In the following these hot nuclei should undergo de-excitation, and their decay products may either survive on the surface of a neutron star or be ejected into inter-stellar space. Also they can serve as seeds for subsequent $r-$process, as discussed in section [**4.4**]{}. Therefore, studying the multifragmentation reactions in the laboratory is important for understanding how heavy elements were synthesized in the Universe. [**6.2 Electron capture rate.** ]{}\ Reduction of the symmetry energy of hot nuclei can be very important for weak processes. Here we consider a typical example related to deleptonization of matter, e.g., the electron capture by nuclei. One should bear in mind that the electron fraction is crucial for dynamics of stellar matter, since the electron pressure dominates at subnuclear densities. The calculation of the electron capture rate, $R_e$, was carried out with the method suggested in ref. [@langanke1]. It is based on an independent particle model and assumes dominance of Gamow-Teller transitions. The electron chemical potential $\mu_e$ and the reaction $Q$-value are the most important energy scales of the capture process. It is clear that the $Q$-value is directly related to the symmetry energy coefficient $\gamma$. A good approximation for the capture rate (per second) on an isolated nucleus is the expression [@langanke1]: $$R_e=\frac{0.693B_g}{t_g}\left(\frac{T}{m_{e}c^2}\right)^5 \left[F_4(\eta)-2\xi F_3(\eta)+\xi^2 F_2(\eta)\right],$$ where $t_g=$6146$s$, $B_g=$4.6 represents a typical (Gamow-Teller plus forbidden) matrix element, $\xi=(Q-\delta E)/T$, $\eta=(\mu_e+Q-\delta E)/T$, $\delta E=$2.5 MeV, and $F_k$ are the relativistic Fermi integrals of order $k$. It is instructive to normalize per nucleon this rate by taking into account the whole ensemble of heavy nuclei produced in stellar matter: $\langle R_e \rangle=\sum\rho_{AZ}R_e/\rho_B$. Figure 23 demonstrates that the electron capture rate in stellar matter depends very essentially on the symmetry energy of nuclei. One can see that at relatively high densities $\rho_B \sim 0.1\rho_0$ the electron capture rate changes only by 20-50%, if we adopt the reduced symmetry energy coefficient $\gamma \approx 15$ MeV. This is because a high electron chemical potential drives the reaction. However, at small densities (below $10^{-3}\rho_0$), when heavy nuclei with large charge still exist (at least at low temperatures), the effect of reduced $\gamma$ is dramatic, of two-three orders of magnitude. We note, that at these relatively low densities and temperatures the nuclear chemical equilibrium may already be problematic [@Botvina09], although some authors keep using it in network calculations [@network]. We believe that hot nuclei can interact with each other by neutron exchange in this case. This situation is similar to what we have at higher densities of nuclear matter in multifragmentation reactions. Therefore, the effect of reduction of the symmetry energy observed in multifragmentation may also take place in the supernova environments and be responsible for significant enhancement of weak reaction rates. [**Conclusions**]{}\ We have formulated a statistical approach (SMSM) designed to describe supernova matter at subnuclear densities. It may be applied for a broad variety of stellar processes, including the collapse of massive stars and supernova explosions, clusterization of nuclear matter in the crust of neutron stars, nuclear composition in merging binary stars, etc. The model includes the whole ensemble of nuclear species, as well as photons and leptons ($e^{-}, e^{+}, \nu, \tilde{\nu}$). The model fully accounts for the nuclear liquid-gas phase transition, which was previously under active investigation in nuclear reactions. In general, we emphasize a close connection of the processes in stellar matter with multifragmentation reactions studied in laboratories. We have calculated main thermodynamical characteristics of the stellar matter under different assumption on the lepton fractions. Nuclear degrees of freedom contribute essentially to the energy and the entropy at high densities. Whereas, at low densities and high temperatures the photons and leptons contributions dominate. The nuclear contribution to pressure becomes essential only when nuclei completely dissociate into nucleons. Accordingly, the adiabatic index increases considerably in this region. On the other hand, we have found that the $\alpha$ particle production at low densities and moderate entropies can be an important process, which should be correctly taken into account in the dynamical simulations of the explosion. The comparison with the Lattimer-Swesty model shows that thermodynamical quantities of the matter, e.g., pressure, are not very different in two models. The reason is that both models treat the leptons and photons in a similar way. Considerable differences appear in the yields of $\alpha$ particles and heavy nuclei. We believe that the SMSM provides more realistic mass and charge distributions of hot nuclei, without any additional constraint on their sizes. As a result of our calculations, we especially emphasize the evolutionary nature of the mass and charge distributions of produced heavy fragments. These distributions carry important information regarding the nuclear liquid-gas phase transition in the stellar matter. Also these nuclei participate in many processes which determine the energy deposition and dynamics of the collapse and explosion. Motivated by recent findings in nuclear multifragmentation reactions we have analyzed how possible in-medium modifications of nuclear properties, in particular, a reduction of the symmetry energy, can influence the fragment yields and weak processes. We have found that these effects can be very important, e.g., for the electron capture by nuclei, which is responsible for deleptonization of matter. At the same time, they can increase the yield of big neutron-rich nuclei. We have discussed new mechanisms of nucleosynthesis leading to the production of heavy and superheavy nuclei in supernova environments. In particular, the shell effects existing at relatively low temperatures may provide an additional enhancement factor for their formation. The authors thank W. Trautmann, J. Schaffner-Bielich, M. Hempel, K. Langanke, Th. Janka, J. Lattimer, M. Liebendoerfer, Th. Buervenich, H. St[ö]{}cker, and W. Greiner for many fruitful discussions. One of us (A.S.B.) acknowledges financial support from the Helmholz International Center for FAIR. This work was partly supported by DFG grant 436RUS 113/711/02 (Germany) and by grant NS-3004.2008.2 (Russia). [99]{} N. Bohr, Nature, 137 (1936) 344. J.P. Bondorf, A.S. Botvina, A.S. Iljinov, I.N. Mishustin, K. Sneppen, Phys. Rep. 257 (1995) 133. G.E. Brown, H.A. Bethe and G. Baym, Nucl. Phys. A 375 (1982) 481. H.A. Bethe, Rev. Mod. Phys. 62 (1990) 801. H.-T. Janka, R. Buras, K. Kifonidis, M. Rampp, T. Plewa, Review in “Core Collapse of Massive Stars”, Fryer, C.L. (ed.), astro-ph/0212314 (2001); H.-T. Janka and E. Mueller, Astron. Astrophys. 306 (1996) 167. M. Liebend[ö]{}rfer et al., Nucl. Phys. A 719 (2003) 144c; Astrophys. J. 620 (2005) 840. K. Sumiyoshi et al., Astrophys. J. 629 (2005) 922. A. Burrows et al., Astrophys. J. 640 (2006) 878. H.-Th. Janka, A. Marek, B. Mueller, L. Schec, arXiv:0712.3070 \[astro-ph\] (2007); AIP Conf. Proc. 983 (2008) 369. F.K. Sutaria, A. Ray, J.A. Sheikh, P. Ring, Astron. Astrophys. 349 (1999) 135. W.R. Hix et al., Phys. Rev. Lett. 91 (2003) 201102. K. Langanke, G. Martinez-Pinedo, Nucl. Phys. A 673 (2000) 481. C.J. Horowitz, Phys. Rev. D 55 (1997) 4577. J.J. Cowan, F.-K. Thieleman, J.W. Truran, Phys. Rep. 208 (1991) 267. Yong-Zhong Qian, Prog. Part. Nucl. Phys. 50 (2003) 153. D.Q. Lamb, J.M. Lattimer, C.J. Pethick, D.G. Ravenhall, Nucl. Phys. A 360 (1981) 459; J.M. Lattimer, C.J. Pethick, D.G. Ravenhall, D.Q. Lamb, Nucl. Phys. A 432 (1985) 646. J.M. Lattimer, F.D. Swesty, Nucl. Phys. A 535 (1991) 331. H. Shen, H. Toki, K. Oyamatsu, K. Sumiyoshi, Nucl. Phys. A 637 (1998) 435. C. Ishizuka, A. Ohnishi, K. Sumiyoshi, Nucl.Phys. A 723 (2003) 517. A.S. Botvina, I.N. Mishustin, Phys. Lett. B 584 (2004) 233. A. Aprahamian, K. Langanke, M. Wiescher, Prog. Part. Nucl. Phys. 54 (2005) 535. R.Buras et al., astro-ph/0507135 (2005). A.S. Botvina, I.N. Mishustin, Phys. Rev. C 72 (2005) 048801. G. Sauer, H. Chandra, U. Mosel, Nucl.Phys. A 264 (1976) 221. D.Q. Lamb, J.M. Lattimer, C.J. Pethick, D.G. Ravenhall, Phys. Rev. Lett. 41 (1978) 1623. D.G. Ravenhall, C.J. Pethick, J.R. Wilson, Phys. Rev. Lett. 50 (1983) 2066. C.J. Pethick, D.G. Ravenhall, Annu. Rev. Nucl. Part. Sci. 45 (1995) 429. C.J. Horowitz, M.A. Perez-Garcia, J. Piekarewicz, Phys. Rev. C 69 (2004) 045804. G. Watanabe, K. Sato, K. Yasuoka, T. Ebisuzaki, Phys. Rev. C 69 (2004) 055805. A.S. Botvina, A.S. Iljinov, I.N. Mishustin, Nucl. Phys. A 507 (1990) 649. A.S. Botvina et al., Nucl. Phys. A 584 (1995) 737. R.P. Scharenberg et al., Phys. Rev. C 64 (2001) 054602. M. D’Agostino et al., Phys. Lett. B 371 (1996) 175. N. Bellaize et al., Nucl. Phys. A 709 (2002) 367. S.P. Avdeyev et al., Nucl. Phys. A 709 (2002) 392. M. D’Agostino et al., Nucl. Phys. A 650 (1999) 329. A. Le Fevre et al., Phys. Rev. Lett. 94 (2005) 162701. J. Iglio et al., Phys. Rev. C 74 (2006) 024605. G. Souliotis et al., Phys. Rev. C 75 (2007) 011601(R). A.S. Botvina et al., Phys. Rev. C 74 (2006) 044609. N. Buyukcizmeci, A.S. Botvina I.N. Mishustin, R. Ogul, Phys. Rev. C 77 (2008) 034608. A.S. Botvina et al., Nucl. Phys. A 475 (1987) 663. A.V. Ignatiuk et al., Phys. Lett. B 76 (1978) 543. M. Prakash et al., Phys. Rep. 280 (1997) 1. A.S. Botvina, A.S. Iljinov, I.N. Mishustin, Sov. J. Nucl. Phys. 42 (1985) 712. A.S. Botvina, O.V. Lozhkin, W. Trautmann, Phys. Rev. C 65 (2002) 044610. K.A. Bugaev, M.I. Gorenstein, I.N. Mishustin, Phys. Lett. B 498 (2001) 144. A.S. Botvina, I.N. Mishustin, Phys. Rev. C 63 (2001) 061601(R). M. Bender, K. Rutz, P.-G. Reinhardt, J.A. Maruhn, W. Greiner, Phys. Rev. C 60 (1999) 034304. J.R. Nix, Ann. Rev. Nucl. Sci. 22 (1972) 65. T.J. Buervenich, I.N. Mishustin, W. Greiner, Phys. Rev. C 76 (2007) 034310. A.S. Botvina, I.N. Mishustin, W. Greiner, work in progress. G.J. Mathews, P. Marronetti, J.R. Wilson, Phys. Rev. D 58 (1998) 043003. T.A. Thompson, A. Burrows, B.S. Meyer, Astrophys. J. 562 (2001) 887. T. Ohnishi, Astroph. & Space Sci. 69 (1980) 155. A.S. Botvina, I.N. Mishustin, Phys. At. Nucl. 71 (2008) 1088. N. Buyukcizmeci, R. Ogul, A.S. Botvina, Eur. Phys. J. A 25 (2005) 57. K. Langanke et al., Phys. Rev. Lett. 90 (2003) 241102. A.S. Botvina, I.N. Mishustin, work in progress. C. Travaglio et al., Astron. Astrophys. 425 (2004) 1029. \[tbh\] {width="10cm"} \[tbh\] {width="8.6cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="10cm"} \[tbh\] {width="8.8cm"}

**Polyhedral Deformations of Cone Manifolds**\ A Aalam **Abstract** Two single parameter families of polyhedra $P(\psi)$ are constructed in three dimensional spaces of constant curvature $C(\psi)$. Identification of the faces of the polyhedra via isometries results in cone manifolds $M(\psi)$ which are topologically $S^1\times S^2$, $S^3$ or singular $S^2$ . The singular set of $M(\psi)$ can have self intersections for some values of $\psi$ and can also be the Whitehead link or form other configurations. Curvature varies continuously with $\psi$. At $\psi=0$ spontaneous surgery occurs and the topological type of $M(\psi)$ changes. This phenomenon is described. 0 Introduction {#introduction .unnumbered} ============== We study continuous families of cone manifolds $M_\psi$ parametrised by cone angle which begin at cone angle zero with the complement of the Whitehead link in $S^3$. We consider the case of equal cone angles on all singular link components. Increasing cone angles the families trace different paths in Dehn surgery space joined by what we call a [*Dehn surgery transition point*]{}. The cone structures for certain non-zero values of cone angles exist in projective models or in $S^3$. In one Dehn surgery direction the cone manifolds are for certain cone angles, obtained by surgery on the Whitehead link in $S^3$ resulting in a topologically distinct singular set in $S^2 \times S^1$. As cone angle is increased the topological type of the singular set changes and the hyperbolic cone manifold develops two cusps and becomes $S^3$ at cone angle $\frac{2}{3}$$\pi$. The topological type of the singular set and the structure of $M_\psi$ remain unaltered as cone angle increases beyond $\frac{2}{3}$$\pi$ until we reach a cone angle $\omega$ where $M_\psi$ becomes $\Bbb R^3$ with topologically the same type of singularity. Increasing cone angle past $\omega$ the singular set reverts back to its pre-$\frac{2}{3}$$\pi$ cone angle topological type and $M_\psi$ becomes spherical in $S^2 \times S^1$. At cone angle $\pi$ the underlying polyhedron becomes a lens in $S^3$ from which the cone manifold is obtained by suitable identifications. For cone angles in the interval \[$\pi, \zeta$\], $M_\psi$ is spherical and the topological type of its singular set is unchanged but it is now in $S^3$. At cone angle $\zeta$, $M_\psi$ becomes the suspension of a sphere with four cone points. It remains the well understood sphere with four singularities for cone angles larger than $\zeta$. Investigating the deformation on the other side of Dehn surgery we obtain the Whitehead link in $S^3$ for certain non-zero cone angles. A complete investigation will be carried out later. Working in hyperbolic or spherical three space of constant curvature it is usual for curvature to be normalised to plus or minus one. While there are good reasons for this convention, this does not allow us to envisage a continuous family of cone structures in which curvature changes from positive to negative with Euclidean space a point in a continuum.cf [@HLM]. Here a [*Cone Manifold*]{} is a PL manifold with a possibly empty codimension two locally flat submanifold called the [*singular set*]{}. In dimensions two and three the singular set consists of isolated points and curves but not graphs respectively. The geometric model is a spherical, Euclidean, or hyperbolic space of constant curvature where the constant is any real number. Points in the complement of the singular set have neighbourhoods homeomorphic to neighbourhoods in the model. Points on the singular set have neighbourhoods homeomorphic to neighbourhoods in the topological space obtained by identifying boundaries of the intersection of two half spaces referred to as a [*wedge*]{} in the model. The homeomorphism takes the singular set to the axis of rotation in the topological space and transition functions are isometries. [*Orbifolds*]{} are represented by discrete structures where cone angles are of the form $\frac{2\pi}{n} , n \in \Bbb N$. cf [@HLM] The topological space obtained by identifying in pairs faces of a polyhedron in a space of constant curvature via isometries is a cone manifold if : 1. No edge is identified with its inverse in the equivalence class induced by the identifications, and the identifications of wedges along faces are cyclic for each equivalence class of edges. 2. The cone angle at each edge, i.e. the sum of the dihedral angles about the edge is $\leq 2\pi$. 3. The neighbourhood of each vertex is a cone on a sphere. If a vertex neighbourhood is, for example, a cone on a torus the topological space is not a manifold. 4. There are either two or no edges emanating from a vertex where cone angle is not 2$\pi$. If two edges then the cone angles must be equal and the edges must be lined up. The vertex neighbourhood is obtained by identifying wedge half planes. cf [@HLM]. It is reasonable to expect that our deformation process holds for any hyperbolic link complement since the polyhedral description of a link complement utilised here is canonical and we can expect to be able to deform this construction by opening cusps in $M_0$ in the way we have described. Our methods can be viewed as part of an approach to describe all compact connected 3-manifolds through deformations by removing singularities of branched singular covers of $S^3$ along universal links thereby providing an approach to the Poincare Conjecture in dimension three. The important phenomenon of self intersecting singular set in a cone manifold occurs in our work. Thus our construction may prove useful in studying this phenomenon. Cone angle interval $[0,\omega]$ {#section:s1} ================================ $M_\psi$ is $S^2 \times S^1$ with two singular components and is hyperbolic for cone angles $\psi \in (0,\frac{2}{3} \pi)$. $M_{\frac{2}{3}\pi}$ is cusped hyperbolic and has a self intersecting singular set. There exists $\omega \in \Bbb R$ such that $M_\psi$ is hyperbolic with a self intersecting singular set when $\frac{2}{3}\pi < \psi < \omega$ and $M_\omega$ is $\Bbb R^3$. [**Proof**]{} $M_0$ the complement of the Whitehead link in $S^3$ is obtained from a two cusped octahedron $P_0$ with identifications described in the projective Klein model of $\Bbb H^3$ in figure \[fig:fig3\]. We refer the reader to [@Rat], [@Thu2] and [@Thu1] for details of this construction. ![The Whitehead link[]{data-label="fig:fig2"}](fig14){height="35mm" width="40mm"} ![$P_0$ with identifications[]{data-label="fig:fig3"}](fig15){height="70mm" width="70mm"} To find out more about $P_\psi$ we view it relative to a coordinate system and inside a [*reference box*]{} in the Klein model as in figure \[fig:fig5\]. The length, width and height of the reference box are denoted by $a, b$ and $c$ respectively. Opening cusps in $M_0$ we obtain $P_\psi$ for $\psi \in (0,\frac{2}{3}\pi)$ as in figure \[fig:fig4\]. ![opening cusps[]{data-label="fig:fig4"}](fig20){height="60mm"} ![opening cusps[]{data-label="fig:fig5"}](fig16c){height="80mm" width="80mm"} From figure \[fig:fig5\] we have the following identifications:\ Face identifications: $$\begin{gathered} A \longleftrightarrow A' \hspace{5mm} i.e. \hspace{5mm} \bigtriangleup(onp) \longleftrightarrow \bigtriangleup(jhi) \\ C \longleftrightarrow C' \hspace{5mm} i.e. \hspace{5mm} \bigtriangleup(sqr) \longleftrightarrow \bigtriangleup(klm)\\ D \longleftrightarrow D' \hspace{5mm} i.e. \hspace{5mm} hexagon (inpqsj) \longleftrightarrow hexagon (hmlrsj)\\ B \longleftrightarrow B'\hspace{5mm} i.e. \hspace{5mm} hexagon (nokmhi) \longleftrightarrow hexagon (poklrq)\end{gathered}$$\ Edge pairings: $$\begin{gathered} ok \longleftrightarrow ok\label{5}\\ sj \longleftrightarrow sj\label{6}\\ on \longleftrightarrow op \longleftrightarrow ji \longleftrightarrow jh\label{7}\\ sr\longleftrightarrow sq \longleftrightarrow km \longleftrightarrow kl\label{8}\\ np \longleftrightarrow ml \longleftrightarrow hi \longleftrightarrow rq\label{9}\\ in \longleftrightarrow hm \longleftrightarrow lr \longleftrightarrow pq \label{10}\end{gathered}$$ Edge pairings \[5\], \[6\] and \[10\] represent singular components of $M_\psi$. We observe: $$\begin{gathered} \notag\text{The dihedral angle between planes incident}\\ \text{in a member of (\ref{5}) or (\ref{6}) is the cone angle}\; \psi.\label{11}\end{gathered}$$\ Moreover (\[7\]), (\[8\]) and (\[9\]) are not part of the singular set. Therefore $$\begin{gathered} \notag\text{The dihedral angle between planes incident}\\ \text{in a member of (\ref{7}), (\ref{8}) or (\ref{9}) is}\; \frac{\pi}{2}.\label{12}\end{gathered}$$ Information in (\[10\]) represents segments $\beta_1, \beta_2, \beta_3$ and $\beta_4$ which are identified to give a singular component $\beta$ of the singular set of $M_\psi$. Considering (\[10\]) with equal cone angles on all singular components we deduce: $$\begin{gathered} \text{Planes incident in a member of (\ref{10}) intersect at dihedral angle}\; \frac{\psi}{4}.\label{13}\end{gathered}$$ We deduce from (\[10\]) that $a=b$. Box coordinates can therefore be normalised to give $a=1=b$ and $c$ where $c$ is measured in the $z-$direction. The reference box is therefore a cube with square base and height $2c$ as in figure \[fig:fig5\]. Topologically $M_\psi$ is $S^1 \times S^2$ with two unlinked singular components as in figure \[fig:OC\]. ![Topology of $M_\psi$[]{data-label="fig:OC"}](OC){height="70mm" width="70mm"} To find out about the geometry of $M_\psi$ we look at $P_\psi$ inside the Klein model of $\Bbb H^3$. Let $R$ be the radius of the Klein ball $B_R$ and let $c$ represent the height of the main box. We will show that $P_\psi$ lives inside $B_R$ and can be specified by $R$ and $c$. We also have $$\begin{gathered} \text{curvature of}\; \Bbb H^3 = -\frac{1}{R^2}\label{curv}\end{gathered}$$ The geometry of $M_\psi$ can therefore be described in terms of $\psi$. We now obtain $R$ and $c$ as functions of $\psi$. Figure \[fig:fig6\] depicts $P_\psi$ in the Klein model. ![$P_\psi$ in the Klein model[]{data-label="fig:fig6"}](fig16d){height="57mm" width="60mm"} In homogeneous coordinates the equations of planes of $P_\psi$ and their poles are: $$\label{planes} \begin{gathered} plane \\ A \\ B \\ C \\ D \end{gathered} \begin{gathered} equation \\ cx+cy+z=c \\ cx-cy-z=c \\ -cx-cy+z=c \\ -cx+cy-z=c \end{gathered}\qquad \begin{gathered} pole \\ (cR,cR,R,c) \\ (cR,-cR,-R,c) \\ (-cR,-cR,R,c) \\ (-cR,cR,-R,c) \end{gathered}$$ Let ${\bf v}=(v_1,v_2,v_3,v_4), {\bf w}=(w_1,w_2,w_3,w_4) \in \Bbb R^4$. Then the hyperbolic bilinear form is given by $$\label{HBF} {\langle{\bf v},{\bf w} \rangle}_\Bbb H = v_1 w_1 + v_2 w_2 + v_3 w_3 - v_4 w_4$$ Let $\theta$ be the dihedral angle of intersection between planes $P$ and $Q$ with poles ${\bf v}$ and ${\bf w}$ respectively. Then $$\label{angle} \cos \theta = -\dfrac{{\langle {\bf v},{\bf w} \rangle}_\Bbb H}{\sqrt {{\langle {\bf v}, {\bf v} \rangle}_\Bbb H} \sqrt {{\langle {\bf w},{\bf w} \rangle}_\Bbb H}}$$ Combining (\[10\]), (\[13\]), (\[planes\]) and (\[angle\]) we can derive expressions for $\cos \psi$ and $\cos \dfrac{\psi}{4}$ in terms of $R$ and $c$ to obtain : $$\begin{gathered} c^2 = \dfrac{1+\cos\psi}{2\cos\dfrac{\psi}{4}-\cos\psi+1}\label{c2}\\ \notag \\ R^2=\dfrac{1+\cos\psi}{2\cos\dfrac{\psi}{4}+\cos\psi-1}\label{R2}\end{gathered}$$ Substituting $(R, c)-$expressions of $\cos\dfrac{ \psi}{4}$ and $\cos\psi$ in (\[c2\]) and (\[R2\]) we can verify these identities. We note $\omega=4\cos^{-1}(\dfrac{2}{\sqrt3}\cos(\dfrac{1}{3}\cos^{-1}(\dfrac{-3\sqrt3}{8})))\approx 2.311984...$ is the smallest positive value of $\psi$ for which the denominator of (\[R2\]) is zero. Hence $\omega$ is the smallest positive value of $\psi$ for which $R^2$ is infinite. This combined with (\[curv\]) implies that curvature is zero and $M_\psi$ is therefore Euclidean when $\psi=\omega$. Combining (\[12\]), (\[planes\]), (\[HBF\]) and (\[angle\]) the equation of plane $N$ in figure \[fig:fig6\] is : $$\label{N} N : -x+y+cz=R^2$$ Let ${\bf p}=(-1,1,-c,R)$ denote the pole of $N$ with $\Bbb H^3$ embedded in $\Bbb RP^3$. We have $$\langle{\bf p},{\bf p}\rangle=2+c^2-R^2=\dfrac{3-4\cos^2\dfrac{\psi}{4}}{(4\cos^3\dfrac{\psi}{4}-4\cos\dfrac{\psi}{4}-1)(4\cos^3\dfrac{\psi}{4}-4\cos\dfrac{\psi}{4}+1)}$$ Therefore $$\label{RP3} \langle {\bf p},{\bf p} \rangle \begin{cases} >0 & \text{when \;$0 < \psi<\dfrac{2}{3} \pi$},\\ =0 & \text{when \;$\psi= \dfrac{2}{3} \pi$},\\ <0 & \text{when \;$\dfrac{2}{3}\pi< \psi < \omega$}. \end{cases}$$ Information contained in (\[RP3\]) corresponds to the configurations depicted in figure \[fig:RP2\] in the case of $\Bbb H^2$ embedded in $\Bbb RP^2$ and is true for $n$ hyperplanes in $\Bbb H^n$ embedded in $\Bbb RP^n$. ![(a) $\langle {\bf p},{\bf p}\rangle>0$ : lines meet outside $\Bbb H^2$. (b) $\langle {\bf p},{\bf p}\rangle=0$ : lines meet on $\partial\Bbb H^2$. (c) $\langle {\bf p},{\bf p}\rangle<0$ : lines meet inside $\Bbb H^2$.[]{data-label="fig:RP2"}](figRP2){height="70mm" width="90mm"} Referring to (\[N\]), figure \[fig:fig6\] and figure \[fig:RP2\], when $\langle {\bf p},{\bf p}\rangle>0$ there is a plane $N$ inside $\Bbb H^3$. When $\langle {\bf p},{\bf p}\rangle=0$ the plane $N$ is the point of intersection of the planes $A, C$ and $D$ on $\partial\Bbb H^3$. When $\langle {\bf p},{\bf p}\rangle<0$ the plane $N$ is the point of intersection of the planes $A, C$ and $D$ inside $\Bbb H^3$. We now verify that $P_\psi \subset \Bbb H^3 \cup \partial\Bbb H^3$ for $\psi\in[0,\frac{2}{3}\pi]$. Since $P_\psi$ is symmetric with respect to the origin it is sufficient to show that the vertices $(A \cap N \cap C)$ and $(A \cap N \cap D)$ of figure \[fig:fig6\] live inside $\Bbb H^3 \cup \partial\Bbb H^3$. Let $$\label{ANC} f(\psi)=R^2-(A \cap N \cap C)^2 = \frac{1}{2}(R^2-c^2)(2+c^2-R^2)$$ where $(A \cap N \cap C)^2$ denotes square of the distance of vertex $(A\cap N \cap C)$ from the origin. We have $f>0$ for $\psi \in (0,\frac{2}{3}\pi)$ and $f=0$ when $\psi=0$ or $\frac{2}{3}\pi$. Let $$\label{AND} g(\psi)=R^2-(A\cap N \cap D)^2= \frac{(R^2-1)(2+c^2-R^2)}{c^2+1}$$ We note $g>0$ when $\psi \in (0,\frac{2}{3}\pi)$ and $g=0$ if $\psi=0$ or $\frac{2}{3}\pi$. Therefore $P_\psi \subset {\Bbb H^3 \cup \partial \Bbb H^3}$ when $\psi \in [0,\frac{2}{3}\pi]$. Hence $M_\psi$ is hyperbolic when $\psi \in [0,\frac{2}{3}\pi]$. We note from (\[RP3\]) and figure \[fig:RP2\] that $M_{\frac{2}{3}\pi}$ is cusped. From figure \[fig:P23Pi\] we observe that $M_{\frac{2}{3}\pi}$ has two cusps, it is topologically $S^3$ with singular set as shown in figure \[fig:S23Pi\] and its cusp neighborhoods are Euclidean turnovers as in figure \[fig:Eturnover\]. ![$P_{\frac{2}{3}\pi}$ inside $\Bbb H^3 \cup \partial \Bbb H^3$[]{data-label="fig:P23Pi"}](P23Pi){height="50mm" width="50mm"} ![Singular set of $M_\psi$ when $\psi\in[\frac{2}{3}\pi,\omega)$ and its cusp neighbourhoods when $\psi=\frac{2}{3}\pi$.[]{data-label="fig:S23Pi"}](S23Pi){height="40mm" width="60mm"} ![A Euclidean turnover[]{data-label="fig:Eturnover"}](Eturnover){height="27mm" width="50mm"} From (\[RP3\]) and figure \[fig:RP2\] we observe that the bounding planes of $P_\psi$ meet inside $\Bbb H^3$ when $\psi\in(\frac{2}{3}\pi,\omega)$. Therefore, $P_\psi$ is in $\Bbb H^3$. Hence $M_\psi$ is hyperbolic when $\psi\in(\frac{2}{3}\pi,\omega)$. The singular set of $M_\psi$ for $\psi\in(\frac{2}{3}\pi,\omega)$ is shown in figure \[fig:S23Pi\]. We note from (\[R2\]) that $R(\omega)=\infty$ hence $M_\omega$ is Euclidean. $P_\omega$ is a tetrahedron with deleted vertices from which $M_\omega$ is obtained using identifications. We observe that $M_\omega$ is $R^3$. The singular set of $M_\omega$ is shown in figure \[fig:Somega\]. ![Singular set of $M_\omega$[]{data-label="fig:Somega"}](Somega){height="40mm" width="70mm"} Cone angles larger than $\omega$ ================================ $M_\psi$ is spherical when cone angle $\psi$ is larger than $\omega$. There exists $\zeta\in \Bbb R$ such that $M_\psi$ is topologically $S^3$ with a self intersecting singular set when $\omega < \psi <\zeta$. $M_\psi$ is the suspension of a sphere with four cone points when $\zeta \leq \psi$. [**Proof**]{} When $\omega<\psi<\pi$ we note $R^2$ the square radius of the Klein model becomes negative so that the model has imaginary radius. This leads us to use the spherical bilinear form for ${\bf v}=(v_1,v_2,v_3,v_4),{\bf w}=(w_1,w_2,w_3,w_4)\in \Bbb R^4$ : $$\label{SBF} {\langle {\bf v},{\bf w} \rangle}_\Bbb S = v_1w_1+v_2w_2+v_3w_3+v_4w_4$$ when $\psi\in(\omega,\pi)$. This is the usual scalar product on $\Bbb R^4$. Let ${\bf x}=(x_1,x_2,x_3,x_4)\in \Bbb R^4$. Since ${\Bbb H}^3_R=\{x \mid {\langle {\bf x},{\bf x} \rangle}_\Bbb H = -R^2\}$ we define ${\Bbb S}^3_R=\{x \mid {\langle {\bf x},{\bf x} \rangle}_\Bbb S = R^2\}$. Thus ${\Bbb S}^3_R$ is the sphere of radius $R$. The “Klein model” ${\Bbb K}^3_R$ for the sphere of radius $R$ is the hyperplane ${\Bbb K}^3_R = \{{\bf x} \mid t=R\}$. In contrast to the hyperbolic case we don’t need to verify that the polytope lies inside the sphere of radius $R$ in ${\Bbb K}^3_R$ since projection from the origin which is not conformal defines a 1-1 correspondence between ${\Bbb K}^3_R$ and the upper hemisphere of ${\Bbb S}^3_R$. The metric on ${\Bbb K}^3_R$ is then the pull back of the metric on ${\Bbb S}^3_R$. As in the hyperbolic case reflections in planes through the origin and rotation about axes through the origin are both Euclidean and spherical isometries. If a plane in ${\Bbb K}^3_R$ has equation $\alpha x+\beta y+\gamma z=\delta$ and $t=R$ then in homogeneous coordinates the equation is $\alpha x+\beta y+\gamma z-(\delta /R)t=0$. In the spherical case the pole has homogeneous coordinates $(\alpha,\beta,\gamma,-\delta /R)$ whereas in the hyperbolic case the pole has coordinates $(\alpha,\beta,\gamma,\delta /R)$. If a pair of planes with poles ${\bf v}$ and ${\bf w}$ in the spherical case intersect with dihedral angle $\theta$ then $$\label{Sangle} \pm\cos\theta=\dfrac{{\langle{\bf v},{\bf w}\rangle}_\Bbb S}{\sqrt{{\langle{\bf v},{\bf v}\rangle}_\Bbb S {\langle{\bf w},{\bf w}\rangle}_\Bbb S}}$$ Using (\[SBF\]), (\[Sangle\]) and the spherical version of (\[planes\]) we obtain $$0\leq R^2 = -\dfrac{1+\cos \psi}{2\cos \dfrac{\psi}{4}-1+\cos \psi}$$ when $\omega\leq\psi\leq \zeta$. When $\psi \in (\omega,\pi)$ we observe that $P_\psi$ is a tetrahedron with identification as in figure \[fig:tetra\]. ![$P_\psi$ inside the reference box[]{data-label="fig:tetra"}](tetra){height="59mm" width="63mm"} Therefore $M_\psi$ is $S^3$ with a self intersecting singular set as in figure \[fig:S23Pi\] with spherical turnover neighbourhoods. ![A spherical turnover[]{data-label="fig:turnover"}](turnover){height="35mm" width="40mm"} Since $R^2(\pi)=0=c^2(\pi)$ the three dimensional model has collapsed into a two dimensional disc at cone angle $\pi$. By projection onto $S_1^3=\{{\bf x} \mid x^2+y^2+z^2+t^2=1\}$ the sphere of radius 1 we observe that $P_\pi$ is a lens with angle $\psi/4$ as in figure \[fig:PPI\] from which $M_\pi$ is obtained by identifications. ![$P_\pi$ is a lens with angle $\dfrac{\pi}{4}$[]{data-label="fig:PPI"}](PPI){height="35mm" width="85mm"} We note $\zeta=4\cos^{-1}(\dfrac{2}{\sqrt3}\cos(\dfrac{1}{3}\cos^{-1}(\dfrac{-3\sqrt3}{8}+\dfrac{4}{3}\pi)))\approx 5.191298...$ is the second smallest value of $\psi$ for which $R$ is infinite. The equation of $c^2$ is the same as in the hyperbolic case. The model continues to be $S^3$ for $\pi<\psi<\zeta$ therefore $M_\psi$ is spherical with a self intersecting singular set as in figure \[fig:SgPi\] when $\psi \in (\pi,\zeta)$. ![Singular set of $M_\pi$[]{data-label="fig:SgPi"}](SgPi){height="40mm" width="70mm"} As cone angle $\psi$ approaches $\zeta$ the segments $h_1$ and $h_2$ decrease in length towards $0$ as shown in figure \[fig:Ppih\] so that we get the singular set shown in figure \[fig:Pzeta\] when $\psi=\zeta$. ![Segments $h_1$ and $h_2$ decrease in length as $\psi$ increases towards $\zeta$[]{data-label="fig:Ppih"}](Ppih){height="40mm" width="80mm"} We therefore obtain $M_\psi$ as the suspension of a sphere with four cone points an in figure \[fig:Pzeta\] when $\zeta<\psi$. ![$M_\zeta$ is the suspension of a sphere with four cone points[]{data-label="fig:Pzeta"}](Pzeta){height="40mm" width="80mm"} Spontaneous surgery =================== We now look at how we get spontaneous surgery and the Whitehead link as the singular set. $P_\psi$ and its identifications “after”spontaneous surgery are shown in figure \[fig:wlpoly\]. Face pairings are $A \longleftrightarrow A', B\longleftrightarrow B', C\longleftrightarrow C'$ and $D \longleftrightarrow D'$ and edges with the same label are identified. ![$P_\psi$ after spontaneous surgery gives $S^3$ with the Whitehead link as its singular set[]{data-label="fig:wlpoly"}](wlpoly){height="60mm" width="80mm"} We note that $M_\psi$ “after” spontaneous surgery is the result of performing $(0,1)$-Dehn surgery on one component of the Whitehead link in $S^3$. Dihedral angles at the $\beta$ edges of $P_\psi$ are now the cone angle $\psi$ and other incidence angles between the planes of $P_\psi$ are as they were prior to spontaneous surgery. Calculations to show the behaviour of $M_\psi$ after surgery remain to be done. Dr Aalam\ PO Box 18810\ London SW7 2ZR\ UK.\ email: aalam@mth.kcl.ac.uk

--- author: - 'N. Antonietti,$^{,\ast}$ M. Mondin, G. Catastini, G. Brida, M. Genovese' title: 'Numerical systematical study of atmospheric effects on Earth-Space QKD' --- Introduction ============ In the last ten years a new discipline called quantum information and devoted to codification, elaboration and transmission of information by exploiting the specific properties of quantum systems has been widely studied and tested.\ The information is said to be quantum when it is encoded in quantum system (for instance the spin of a particle, the polarization or phase of a photon). A secure cryptosystem can be achieved if one encodes information in a quantum system. To be more precise, by exploiting the properties of quantum systems, Alice and Bob can share secret keys which can then be used in standard secret-key protocols [@gis; @w; @5; @6; @7]. The most natural carrier of quantum information is the photon. In fact it travels with the speed of light, has a very limited interaction with the environment and allows to encode information in several degrees of freedom, such as polarization, phase or energy. All the experimental realizations of quantum cryptographic protocols (more properly Quantum Key Distribution protocols, QKD), since the first one at IBM in 1989 (published in 1992 [@5]) used single photons as quantum bits (qubits). Afterwards, many research teams have made quantum key distribution using different protocols and different physical implementations [@gis]. Following these experimental verifications, the research on QKD has been then addressed toward the realization of long distance communications [@lon], the implementation of protocols suited for commercial purposes [@com], the study of eavesdropping in realistic conditions [@eav], protocols in higher dimensions and/or many particles [@qud; @las2], ... Nowadays the frontier of QKD is the realization of a Earth-Space (Space-Earth) or a Space-Space quantum communication channel [@zei; @rar; @giap; @nos]. A communication channel, in this case the atmosphere, is said to be quantum when quantum information is transmitted through it and we call this process quantum communication. This realization would be of utmost relevance both for quantum key distribution [@gis], since it would allow intercontinental quantum transmissions, and for studies concerning foundations of quantum mechanics [@zei; @mg]. Thus, preliminary feasibility studies have been performed showing its practical realizability. In little more details, in ref. [@rar] a BB84 scheme was studied for Earth-Space communication. By considering gaussian optics, a 15 dB loss was attributed to diffraction, whilst aerosol loss was considered of secondary relevance (0.04-0.06 dB) for transmission with clear sky and from high elevation above sea. Altogether atmospheric losses were estimated to be about 2-5 dB. On the other hand, the security level for quantum transmission (the amount of losses which the communication is still safe with) was estimated to be 40 dB (from estimated background and detectors dark counts), lowering to 10 dB if the eavesdropper (conventionally dubbed Eve) had technologies for intercepting selectively a possible multi-photon component. In another study, ref. [@zei], a 6.5 dB loss was estimated by considering optics and finite quantum efficiency of detectors [@las1]. The limit for secure quantum transmission was estimated to be at 60 dB loss. Finally, here atmospheric losses were estimated to be around 1 dB.\ Effectively, daylight and open space transmission at a distance of 10 km [@10; @rz] was achieved and, very recently, an European collaboration has achieved preliminary results on a quantum channel (at Canary islands) at a distance up to 144 km [@tom]. All these theoretical and experimental studies guarantee the feasibility of a ground-space channel. Nevertheless, the analysis of atmospheric effects is rather incomplete and is far from considering various realistic atmospheric situations that could be met during a real transmission. Even in the very general review of ref. [@gil], only few results are presented, and with small detail. Thus, a detailed analysis of atmospheric effects in various realistic situations would be of the utmost relevance. In particular one should consider both static atmosphere effects (as absorbtion, scattering and emissions) and turbulent atmospheric effects (as wandering). Even if the latter represent the main contribution to losses, more than reflection, in perspective they can be coped with technological solutions (as adaptative optics), whilst absorption (even if smaller) will represent the final limit of this kind of communication. Errors due to electronic and optical imperfections must be taken into account separately. Purpose of this paper is to describe a work that addresses a precise characterization of static atmospheric effects on the quantum communication process. We want to determine the losses in the quantum information carriers (photons) and, in this work of ours, we consider lost any photon that has interacted with the atmosphere. Then we compare the estimated losses with the security loss upper bounds mentioned before. After estimating the dependence of a secure transmission in different meteorological situation, for example, it would then be possible to evaluate the average available time for a secure communication for a certain ground station by the statistical meteorological conditions of the station itself. To investigate this topic we have used a free source library for radiative transfer calculations named libRadtran[@libradtran]. This library can solve the radiative transfer equations, set some input parameters and exploit the HITRAN[@hitran] database, which is a high-resolution atmospheric parameters database (for example there are the atmospherical components cross sections for different wavelengths).\ In our simulations we can determine what part of the irradiance of a source at the top of the atmosphere can reach the ground without interactions with the atmosphere. The effects on photon polarization can be estimated to be small albeit not completely negligible (depolarization being of the order of $3.5 \%$ by including single forward Rayleigh scattering only [^1]): their precise evaluation is now in progress. Various parameters can influence the atmospheric effects on the photon transmission, as, for instance, aerosols, pressure, temperature, air density, precipitations, cloud composition, humidity, chemical components. As a first step in order to evaluate their relevance in various meteorological conditions, here we present some preliminary results obtained by varying some of them in realistic intervals. Numerical results {#sec:real_experiments} ================= In [@chand], a *pencil of radiation* is defined in terms of the *specific intensity* $u_{\nu}$, as the amount of energy $dE_{\nu}$, in a specified frequency interval $(\nu ,\nu + d\nu)$, which is transported across an element of area $d\sigma$, along directions confined to an element of solid angle $d\omega$, during a time $dt$, according to the following formula: $$dE_{\nu} = u_{\nu} \cos \theta d\nu d\sigma d\omega dt,$$ where $\theta$ is the angle between the direction of the incoming radiation and the outward direction normal to $d\sigma$.\ We can consider a photon traveling through the atmosphere, as a pencil of radiation whose radiant energy is discrete and ideally as a monochromatic radiation (they have actually a bandwidth, although very narrow).\ Thus, in this sense, the photon can be considered with a classical treatment, undergoing absorption and scattering.\ Let’s then consider an atmosphere modeled as stratified in parallel planes. In such an atmosphere, all the physical properties (temperature, pressure, relative humidity...) and the chemical component density are invariant over a plane. That’s a very common model for the atmosphere that is, indeed very often in nature, structured as plane parallel [@chand].\ When a pencil of radiation traverses a medium of density $\rho$ and thickness $ds$, it’s weakened according to the following equation, $du_{\nu} = -k_{\nu} \rho u_{\nu} ds$, where $k_{\nu}$ is called the mass absorption coefficient.\ The optical depth at the height $z$ from the Earth surface is $\tau = \int_z^{\infty}k \rho dz$. The transfer of monochromatic radiation through a parallel plane atmosphere is described by the radiative transfer equation [@chand]: $$\label{rte} \mu \frac{du_{\nu}(\tau_{\nu}, \mu, \phi)}{d\tau}=u_{\nu}(\tau_{\nu}, \mu, \phi) - S_{\nu}(\tau_{\nu}, \mu, \phi)$$ where $\mu=\cos \theta$, $u_{\nu}(\tau_{\nu}, \mu, \phi)$ is the specific intensity along the direction $(\mu ,\phi )$, at optical depth $\tau _{\nu}$.\ $S$ is the source function: $$\label{eq:soft_sourc_func} \begin{aligned} & S_{\nu}(\tau _{\nu},\mu ,\phi) = \frac{\omega _{\nu}(\tau _{\nu})}{4\pi} \int_0^{2\pi} d\phi ' \int_{-1}^1 d\mu 'P_{\nu}(\tau _{\nu},\mu ,\phi; \mu ',\phi ')\\ & \times u_{\nu}(\tau _{\nu},\mu' ,\phi') + Q_{\nu}(\tau _{\nu},\mu ,\phi), \end{aligned}$$ $\omega_{\nu}(\tau_{\nu})= \int p(\cos \Theta)\frac{d\omega '}{4\pi}$ is the single scattering albedo and $p(\cos \Theta )$ is the phase function, that gives the rate at which energy is being scattered into an element of solid angle $d\omega '$ and in a direction inclined at a direction $\Theta$ to the direction of incidence of a pencil of radiation on an element of mass $dm$. For our investigations, a Henyey-Greenstein phase function is assumed.\ $Q_{\nu}(\tau _{\nu},\mu ,\phi) = [1-\omega_{\nu} (\tau_{\nu})B_{\nu}(T(\tau_{\nu}))]$ is the amount of radiant energy emitted by the atmosphere as thermal emission and $B_{\nu}(T(\tau_{\nu}))$ is the Planck function at frequency $\nu$and temperature $T$.\ The discrete approximation to (\[rte\]) can be written as [@stamnes]: $$\label{rte_disc} \begin{aligned} & \mu_i \frac{du^m (\tau ,\mu_i )}{d\tau} = u^m (\tau ,\mu_i ) - \sum_{\substack{j=-N \\ j\neq 0}}^N w_jD^m (\tau ,\mu_i ,\mu_j ) \times \\ & u^m (\tau ,\mu_j) - Q^m (\tau ,\mu_i) \hspace{3mm} (i=\pm 1 ,\ldots \pm N ). \end{aligned}$$ The phase function is expanded in a series of Legendre polynomials ($D^m$) and the intensity in a Fourier cosine series whose coefficient are $u^m$. $w_i$ are the quadrature weights of the series of Legendre polynomials. Equation (\[rte\]) is replaced by 2N independent equations (whose unknown quantities are the coefficients $u^m$) and the procedure is repeated for each layer we have divided the atmosphere in. Equation (\[rte\_disc\]) is transformed in a system of $2N$-coupled ordinary differential equations with constants coefficients which can be solved numerically. The solution of (\[rte\_disc\]) is reported in [@stamnes]. The libRadtran library we are going to use solves the radiative transfer equation (\[rte\]) and gives, at the output, the amount of energy which doesn’t interact with the atmosphere, by means of the algorithm presented in [@stamnes]. Different distributions of aerosols {#subsec:aerosols} ----------------------------------- In libRadtran, a database of aerosols distributions and their optical properties can be found. It has been written according to ref. [@shettle:1989]. There, four aerosols distributions for four different environment conditions are described (rural, maritime, urban, tropospheric).\ For our first analysis, the atmospheric conditions (for a plane-parallel atmosphere model) are selected to be in summer season, at midlatitudes, according to ref. [@afgl:1986]; there, the atmosphere is described with its pressure, temperature, air density, relative humidity profiles, etc and the optical properties are derived. The source irradiance, posted at the top of the atmosphere, is chosen in accordance to ref. [@kato:1999]; both databases are present in libRadtran.\ For these conditions the direct downward irradiance (the amount of radiation which doesn’t experience any interaction with the atmosphere) at the Earth surface with the source at the zenith and at a zenith angle of either 50$^o$ and 80$^o$, outside the whole atmosphere has been evaluated. Calculations are performed invoking a correlated-k band parametrization by Kato et al. [@kato:1999] and this choice will be maintained for the next calculations, unless differently specified. We are mainly interested in the visible wavelengths (actually in between about 700 nm and 900 nm) because that range is not affected by strong absorption as UV and IR bands; anyway, for completeness we extended our investigations in the infrared band and actually the exact range is in between 256.3 nm and 2638.5 nm (being some good transmission windows present here as well); the extreme points in the wavelength range are determined by the choice of the database [@kato:1999]. In figure (\[transmittance\_shettle\_aerosols\]) we show the direct downward transmittance $T$ (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance). A first result of this analysis is that this quantity is largely independent of the aerosol type. Moreover the behavior of the transmittances for different aerosols conditions are largely independent of the source zenith angle too.\ Of course, the evaluation of $T$ next to the extreme source zenith angle of 90$^o$ is useless, because fairly no radiation reaches the ground (besides the parallel planes model of atmosphere used in the program cannot be extended beyond 80$^o$ from zenith). ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm for 0$^o$, 50$^o$, 80$^o$ source zenith angles and different aerosols conditions, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986] and the source irradiance is chosen in accordance to ref. [@kato:1999]. The aerosol distributions are the rural, maritime, urban and tropospheric as described in ref. [@shettle:1989]. Here and in the following large scale figures the wave length resolution is kept poor in order not to compromise the readability: the code allow a much more detailed resolution than can be exploited for the region of interest (and that will be used in some smaller scale figures). []{data-label="transmittance_shettle_aerosols"}](./relative_transmission_aerosols.eps "fig:"){width="8.5cm"}\ It can be observed that the best range for communications is roughly from 600 nm to almost 900 nm, but some window is present also in infrared region (as 1,564 nm or 2,214 nm). ![As in fig.1, but selecting three specific wave length windows with LOWTRAN resolution. []{data-label="aerdet"}](./relative_transmission_aerosols_zoom1200_LOWTRAN.eps "fig:"){width="8.5cm"} ![As in fig.1, but selecting three specific wave length windows with LOWTRAN resolution. []{data-label="aerdet"}](relative_transmission_aerosols_zoom1600_LOWTRAN.eps "fig:"){width="8.5cm"} ![As in fig.1, but selecting three specific wave length windows with LOWTRAN resolution. []{data-label="aerdet"}](relative_transmission_aerosols_zoom700_LOWTRAN.eps "fig:"){width="8.5cm"}\ A detail of absorptions in some more restricted wave-length windows is shown in fig. (\[aerdet\]), using the LOWTRAN atmospheric database. Incidentally, the use of strong resolution in wave length makes unreadable figures with a large scales and thus we only show it in lower scale figures as (\[aerdet\]); here one can appreciate the details and clearly distinguish specific absorption lines that should be avoided for communication. On the other hand large scale figures show the rough general dependence of absorption, pointing out the regions where a more detailed analysis can be interesting. In the range 600-900 nm, for a source zenith angle of 0$^o$, the fraction source light which gets across the atmosphere without any interaction with the atmosphere is, excluding some specific absorption lines (as, for instance, around 760 nm, that is due to the oxygen absorption line), about 70%.\ This means that a photon has a 70% probability to get across the model of atmosphere we have built without interacting with it.\ We can present, as usual, the losses in dB ($l_{dB}$), from the direct downward transmittance in percent ($T_{\%}$), $$l_{dB} = -10 \log_{10}(T_{\%})$$ Then, in the visible window from 700 nm to 900 nm, the losses due to atmospheric interaction are less than 4 dB if the source zenith angle is 0$^o$ and less than 20 dB if the source zenith angle is 80$^o$, but at any rate lower than requested in [@zei] for establishing a secure communication. Thus, one can infer that the transmission can be carried on for almost the whole visibility range of a satellite, when the atmosphere is in the conditions we have described before. In the following picture (\[losses\_aerosols\] )the losses versus the wavelength are depicted: ![Losses in dB vs wavelength in nm for 0$^o$, 50$^o$, 80$^o$ source zenith angles and different aerosols conditions, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986] and the source irradiance is chosen in accordance to ref. [@kato:1999]. The aerosol distributions are the rural, maritime, urban and tropospheric as described in ref. [@shettle:1989]. The lowest and most constant losses are in the range from 700 nm to 900 nm[]{data-label="losses_aerosols"}](./losses_aerosols.eps "fig:"){width="8.5cm"}\ When the zenith angle is 0$^o$, the losses are less than 60 dB for wavelengths from 295.1 nm; when zenith angle is 50°$^o$ losses are around 60 dB at wavelengths equal to 295.1 nm; finally, at zenith angle equal to 80°$^o$, losses are less than 60 dB only starting from 317.3 nm. Different temperature profile {#subsec:temp_prof} ----------------------------- In the considered atmosphere database, the temperature decreases fairly linearly from the ground level value $T_{0}$ up to 15 km, where it assumes a given value $T_{15}$. $T_{0}$ is actually a free parameter, and we have varied its value from -10$^o$C up to 30$^o$C, with steps of 5$^o$C. The air density is modified according to the perfect gas law. Above 15 km, the parameters have been left unchanged, and no aerosols presence has been considered. We evaluated the Transmittance for three different source zenith angles (0$^o$, 50$^o$, 80$^o$). It turns out that the ground level temperature doesn’t affect the transmittance at all. Moreover, as in the previous case, the behavior of the transmittances for different level ground temperatures are largely independent of the source zenith angle. This atmosphere and the following ones are aerosol-free\ So, we depict in figure (\[transmittance\_temperature\]) the results of the simulations for the source zenith angle equal to 0$^o$. ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm for 0$^o$ source zenith angle, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986], but with a further modification: the surface temperature has been made vary from -10$^o$C up to 30$^o$C, with steps of 5$^o$C and the temperature has been decreased linearly with the altitude (for the first fifteen kilometers), up to the value it assumed in the unmodified atmosphere [@afgl:1986]. The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free[]{data-label="transmittance_temperature"}](./relative_transmission_temperature.eps "fig:"){width="8.5cm"}\ Once again, it is possible to observe that the best range for the communications is roughly from 700 nm to 900 nm as well. In fact, in this case where no aerosol was included, at a solar zenith angle of 80$^o$ the losses are less than 14 dB. The detailed dependence, according to the LOWTRAN atmospheric database, can be observed in picture (\[transmittance\_temperature\_zoom\]). ![Detail of the figure \[transmittance\_temperature\], from 600 nm to 900 nm[]{data-label="transmittance_temperature_zoom"}](./relative_transmission_temperature_zoom_LOWTRAN.eps "fig:"){width="8.5cm"}\ As for the previous analysis, when the zenith angle is 0$^o$, the losses are less than 60 dB for wavelengths from 295.1 nm; when zenith angle is 50$^o$, losses are around 60 dB at wavelengths equal to 295.1 nm and then decrease as wavelengths increase; finally, at zenith angle equal to 80$^o$, losses are less than 60 dB only starting from 317.3 nm.\ In this case, the range from 700 nm to 900 nm is the actual minimum absorbtion range in the visible wavelengths and in the picture (\[losses\_temperature\]), it is seen in detail (hereafter, all the detailed figures are imlied to be evaluated with the LOWTRAN database). ![losses in dB vs wavelength in nm for 0$^o$, 50$^o$ and 80$^o$ source zenith angle, in the range between 700 nm and 900 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986], but with a further modification: the surface temperature has been made vary from -10$^o$C up to 30$^o$C, with steps of 5$^o$C and the temperature has been decreased linearly with the altitude (for the first fifteen kilometers), up to the value it assumed in the unmodified atmosphere [@afgl:1986]. The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free.[]{data-label="losses_temperature"}](./losses_temperature_zoom_LOWTRAN.eps "fig:"){width="8.5cm"}\ Different humidity profiles {#subsec:hum_prof} --------------------------- In order to observe the effect of humidity, a further modification has been added to the atmospheric conditions of ref. [@afgl:1986]. This time, the relative humidity has been set in different times as a constant value along the first 15 km of the atmosphere. The values are 5% and from 10% to 100% with steps of 10%. No aerosols have been considered in this configuration. In figure (\[transmittance\_humidity\]), the evaluated direct downward transmittance can be observed for source zenith angles of 0$^o$, 50$^o$ and 80$^o$. ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm for 0$^o$, 50$^o$, 80$^o$ source zenith angles, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986], but with a further modification: the relative humidity has been set constant along the first 15 kilometers of the atmosphere, with values of 5%, 40%, 70% and 100% values. The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free[]{data-label="transmittance_humidity"}](./relative_transmission_humidity.eps "fig:"){width="8.5cm"}\ Also in this scenario, the most advantageous range for communication is from 700 nm to 900 nm, where losses are less than 10 dB, but in this case we can observe some differences among different humidity conditions. As it can be observed from figure (\[transmittance\_humidity\]) and, in more details, in figure (\[transmittance\_humidity\_zoom\]), the absorption in the range between 800 nm and 1000 nm is water vapor dependent and is strongly affected by its presence. ![Detail of the figure \[transmittance\_humidity\], from 550 nm to 900 nm and a 0$^o$ zenith angle[]{data-label="transmittance_humidity_zoom"}](./relative_transmission_humidity_zoom_LOWTRAN.eps "fig:"){width="8.5cm"}\ The losses in dB are depicted in figure (\[losses\_humidity\]), ![Losses in dB vs wavelength in nm for 0$^o$ source zenith angle, in the range between 550 nm and 900 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986], but with a further modification: the relative humidity has been set constant along the first 15 kilometers of the atmosphere, with values of 5%, 40%, 70% and 100% values. The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free[]{data-label="losses_humidity"}](./losses_humidity_zoom_LOWTRAN.eps "fig:"){width="8.5cm"}\ When the zenith angle is 0$^o$, the losses are less than 60 dB for wavelengths from 295.1 nm; when zenith angle is 50°$^o$ losses are around 60 dB at wavelengths equal to 295.1 nm; finally, at zenith angle equal to 80°$^o$, losses are less than 60 dB only starting from 317.3 nm.\ There are minima outside the range from 700 nm to 900 nm but the former is the most stable. Presence of clouds {#subsec:cloud} ------------------ In order to study the possibility of establish a quantum communication channel, the presence of clouds has to be considered as well. In order to do this, we have added clouds to the atmosphere [@afgl:1986] without aerosols. We set at an altitude of 10 km, a 1 km deep layer of clouds, whose liquid water content is $0.06 gm^{-3}$ and the effective droplet radius is $50\mu m$. This configuration matches a cirrus and the estimation of direct downward transmittance is depicted in the figure (\[transmissivity\_clouds\]). ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm for 0$^o$ source zenith angle, in the range between 256.3 nm and 2638.59 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. A cirrus cloud at an altitude of 10 km has been added; it is 1 km deep, with a liquid water content of $0.06 gm^{-3}$ and an effective droplet radius of $50\mu m$ The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free[]{data-label="transmissivity_clouds"}](./relative_transmission_clouds.eps "fig:"){width="8.5cm"}\ Liquid water content and effective droplet radius are translated into optical properties in [@hu_st]. As it can be seen in the figure (\[losses\_clouds\]), the presence of this kind of cloud doesn’t disable the communication. For 0$^o$ and 50$^o$ zenith angles, the losses are smaller than 60 dB starting from the 295.1 nm wavelength, nevertheless they remain always around 15-20dB. At the zenith angle of 80$^o$, losses are dramatically close to 60 dB. Thus, these results suggest that even the presence of thin clouds as cirri makes the transmission substantially delicate. On the other hand the presence of any other kind of clouds with a higher water content (stratus $L=0.28 gm^{-3}$, cumulus $L=0.26 gm^{-3}$, cumulonimbus $L=1 gm^{-3}$, stratocumulus $L=0.44 gm^{-3}$, ...) makes the communication impossible. This information together with a description of passages of different clouds within different perturbations and statistical data on average meteorological evolution in a year for a given station allows an estimate of the available time for transmission from a specific place. ![Losses in dB vs wavelength in nm for 0$^o$, 50$^o$ and 80$^o$ source zenith angles, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. A cirrus cloud at an altitude of 10 km has been added; it is 1 km deep, with a liquid water content of $0.06 gm^{-3}$ and an effective droplet radius of $50\mu m$ The source irradiance is chosen in accordance to ref. [@kato:1999]. This atmosphere is considered aerosol-free[]{data-label="losses_clouds"}](./losses_clouds.eps "fig:"){width="8.5cm"}\ The program allows a similar analysis for fog with different degrees of optical depth as well. Comparison between two extremely different conditions {#subsec:diff_cond} ----------------------------------------------------- Then we want to get an idea of how is transmissivity for two extremely different conditions. On one side there is a city environment with relevant aerosols concentrations and 90% relative humidity, on the other side a dry desert without aerosols. The results for these two cases are depicted in figure (\[transmissivity\_citydesert\]) for source zenith angle of 0$^o$, 50$^o$ and 80$^o$. ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm for 0$^o$, 50$^o$ and 80$^o$ source zenith angles, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. The source irradiance is chosen in accordance to ref. [@kato:1999]. In the first case, there is a urban aerosol environment and a 90% relative humidity and on the other side, there is a dry desert without aerosols[]{data-label="transmissivity_citydesert"}](./relative_transmission_comparison.eps "fig:"){width="8.5cm"}\ As we could expect, a dry desert is a much better environment for quantum communication than a humid city. The losses in the desert are always at least 3 dB smaller than in the city\ Anyway, either a dry desert and a city are secure environment for quantum communications. For 0$^o$ and 50$^o$ zenith angles, starting with wavelength equal to 295.1 nm, losses are lower than 60 dB, for 80$^o$ the first secure wavelength is 317.3 nm.\ We want to point out once more that the losses calculated so far, are only expect from the atmosphere. The security limit of 60 dB is a total loss limit, including, for instance, quantum efficiency of detectors, optical losses in the devices,...\ For all the cases under consideration, the best range for telecommunications is from 700 nm to 900 nm. ![Losses in dB vs wavelength in nm for 0$^o$, 50$^o$ and 80$^o$ source zenith angles, in the range between 256.3 nm and 2638.5 nm. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. The source irradiance is chosen in accordance to ref. [@kato:1999]. In the first case, there is a urban aerosol environment and a 90% relative humidity and on the other side, there is a dry desert without aerosols[]{data-label="losses_citydesert"}](./losses_comparison.eps "fig:"){width="8.5cm"}\ Quantum Entanglement over the Danube {#sec:quant_ent_over_danube} ------------------------------------ Finally we would like to consider the realistic situation of some experiment. As a first example, we consider a quantum entanglement distribution experiment [@danube] performed over the Danube in Vienna, for a distance of 600m. The information we can infer from the paper some of the meteorological conditions of when such experiment was realized, e.g. the temperature was around 0$^o$C and wind had strength up to 50 km/h. The bottom (at the Danube level) of the atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Urban aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999]. As an application of our program to a realistic situation, here we report the atmospheric effects for this experiment as deduced from our analysis. Although the receivers were located at a distance of either 150 m and 500 m from the source of entangled photon, we discuss the atmospheric effects over 600 m, the distance between the two receivers. The direct downward transmittance is shown in figure (\[transmittance\_danube\]). ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm, in the range between 256.3 nm and 2638.5 nm. The bottom (at the Danube level) of the atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Urban aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999]. The path the light has to cross is 600 m long, the wind blows at 50 km/h and the surface temperature is 0$^o$C[]{data-label="transmittance_danube"}](./relative_transmission_danube_LOWTRAN.eps "fig:"){width="8.5cm"}\ For 810 nm (the wavelength in the experiment), the direct downward transmittance percentage is about 94%. In [@danube] it is reported that “The attenuation in each of the links was about 12dB”, but no more indications are given about the sources of attenuation. According to our simulation, the atmospheric losses at that wavelength are less than 0.3 dB (see fig.\[losses\_danube\]). We can thus suppose that the main attenuation factors were the optical losses and finite quantum efficiency of detectors. ![Losses in dB vs wavelength in nm, in the range between 256 nm and 2638 nm. The bottom (at the Danube level) of the atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Urban aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999]. The path the light has to cross is 600 m long, the wind blows at 50 km/h and the surface temperature is 0$^o$C. The best range is, as usual so far, from 700 nm to 900 nm[]{data-label="losses_danube"}](./losses_danube_LOWTRAN.eps "fig:"){width="8.5cm"}\ 144 km transmission {#subsec:120km} ------------------- As hinted in the Introduction, an ongoing experiment at Canary Islands is devoted to establish a 140 km quantum-link. Here we discuss photon transmission for a reasonable range of atmospheric conditions in a foreseen Canary Island scenario. Generic environmental conditions are taken into considerations: the atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Maritime aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999], a 2mm monthly averaged water precipitation (value that will scale the water vapor profile accordingly). The results of a transmission on a 144 km distance in this scenario are reported in figure (\[transmittance\_laspalmas\]). ![Direct downward transmittance (ratio between the direct downward irradiance at the Earth surface with respect to the source irradiance) vs wavelength in nm, in the range between 256.3 nm and 2638.5 nm. The scenario is at the Canary islands. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Maritime aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999]. The path the light has to cross is 144 km long, the average monthly precipitation is 2 mm[]{data-label="transmittance_laspalmas"}](./relative_transmission_laspalmas.eps "fig:"){width="8.5cm"}\ Results show that higher transmission percentages can be obtained for high wavelengths. Anyway, the losses are always less than 20 dB in the range from 700 nm to 900 nm.\ The secure communication (losses lower than 60 dB) starts from the wavelength equal to 345.1 nm (see figure (\[losses\_laspalmas\])). ![Losses in dB vs wavelength in nm, in the range between 256.3 nm and 2638.5 nm. The scenario is at the Canary islands. The atmosphere is in summer conditions and at midlatitudes, according to ref. [@afgl:1986]. Maritime aerosols are used. The source irradiance is chosen in accordance to ref. [@kato:1999]. The path the light has to cross is 120 km long, the average monthly precipitation is 2 mm[]{data-label="losses_laspalmas"}](./losses_laspalmas.eps "fig:"){width="8.5cm"}\ From a comparison of figures \[losses\_laspalmas\] and \[losses\_danube\] one can also appreciate as different atmospheric conditions (aerosols, humidity, etc.) affect specific absorption regions. Conclusions {#sec:conclusions} =========== In this paper we have presented some preliminary results on atmospheric interaction with photons, obtained by using the free source library libRadtran.\ Our results show that a secure communication can be established under many realistic meteorological conditions even up to only $10^o$ from horizon. Thus, a Earth-satellite quantum channel can be realized for a large fraction of visibility each orbit. Furthermore, our results can be used for a first estimate of the fraction of time per year when a secure communication quantum channel with a certain satellite can be achieved. A further deeper analysis of atmospheric effects based on this approach could effectively be a useful tool for predicting precisely the performances of a quantum communication channel in various realistic operative meteorological situations. Acknowledgements ================ This work has been supported by Regione Piemonte (E14), by MIUR FIRB RBAU01L5AZ-002 and by “San Paolo foundation”. [0]{} N. Gisin, Rev. Mod. Phys. 74 (02) 145. S. Wiesner, Sigact News 15 (1983) 78. C. H. Bennett, F. Bessette, G. Brassard, L. Salvail, J. Smolin, J. Cryptology 5, 3 (1992). C. H. Bennett and G. Brassard, Int. Conf. Computers, Systems and Signal processing, Bangalore, India, p. 175 (1984). A. K. Ekert, Phys. Rev. Lett. 67, 661 (1991). E. Diamanti et al., quant-ph 0608110; C. Peng et al., quant-ph 0607129; I. Marcikic et al., quant-ph 0606072; C. Kurtsiefer et al., Nature 419 (02) 450; I. Marcikic et al., quant-ph 0404124. F. Grosshans et al., Nature 421 (03) 238; D. Stucki et al., Appl. Phys. Lett. 87 (05) 194108. C. Kim et al., quant-ph 0603013; Q. Kai, Phys. Lett. A 351 (06) 23; F.A. Bovino et al., quant-ph 0308030; M. Genovese, Phys. Rev. A 63 (01) 044303. H. Bechmann-Pasquinucci and W. Tittel, (2000) 062308. H. Bechmann-Pasquinucci and A. Peres, (2000) 3313. M.Bourennane et al., (2001) 062303.N.J. Cerf et al., (2002) 127902. D. Bruss and C. Macchiavello, (2002) 127901. M. Genovese and C. Novero, Eur. Journ. of Phys. D. 21 (2002) 109. A.V. Burlakov et al., Phys. Rev. A 60 (1999) R4209-R4212. M.V. Chekhova et al., Phys. Rev. A 70 (2004) 053801. J.C. Howell, A. Lamas-Linares, and D. Bouwmeester, Phys. Rev. Lett. 88 (02) 030401. A. Vaziri et al., Phys. Rev. Lett. 89 (2002) 240401. R.T.Thew et al., 93 (2004) 010503. V. N. Gorbachev, A. I. Trubilko, Laser Physics Letters Volume 3, Issue 2, Date: February 2006, Pages: 59-70. R. Kaltenbaeck et al., “Quantum Comm. and Quantum Imag.”, ed R. Meyers and Y. Shih, Proc. of SPIE 5161, pag. 252. P. Villoresi et al., quant-ph 0408067. M. Aspelmeyer et al., IEEE Sel. Top. Quant. El. 9 (03) 1541. J.. Rarity et al., “Quantum Comm. and Quantum Imag.”, ed R. Meyers and Y. Shih, Proc. of SPIE 5161, pag. 240. P.J. Edwards et al., “Quantum Comm. and Quantum Imag.”, ed R. Meyers and Y. Shih, Proc. of SPIE 5161, pag. 152. G. Catastini, M. Rasetti, R. Ionicioiu, G. Brida, M. Genovese, Communication presented at ONERA workshop, Paris, April 2005. M. Genovese, Physics Reports 413/6 (2005) 319. G. Brida, M. Genovese, M. Gramegna, Laser Physics Letters Volume 3, Issue 3, Date: March 2006, Pages: 115-123. C. Peng et al., 94 (05) 150501; R.J. Hughes et al., La-UR-02-449; R. Alleaume et al., quant-ph 0402110. C. Kurtsiefer et al., Nature 419 (02) 450. T. Occhipinti, personal communication. R. Ursin et al., quant-ph 0607182. G. Gilbert and M. Hamrick, quant-ph 0009027. http://www.libradtran.org/ http://cfa-www.harvard.edu/hitran/. S. Chandrasekhar, Radiative transfer (Dover, New York, 1960) Stamnes, K., S.-C. Tsay, W. Wiscombe and K. Jayaweera, 1988, ‘Numerically stable algorithm for discrete-ordinate-method radiative transfer in multiple scattering and emitting layered media’, Applied Optics, 27, 2502. E. P. Shettle, Models of aerosols, clouds and precipitation for atmospheric propagation studies, “Atmospheric propagation in the UV, visible, ir and mm-region and related system aspects”, vol. 454, (1989) G. P. Anderson, S. A. Clough,F. X. Kneizys, J. H. Chetwynd and E. P. Shettle, AFDL atmospheric constituent profiles, “AFGL Tech. Rep., AFGL-TR-86-0110”, Air Force Geophysical Laboratories (1986). S. Kato and T. P. Ackerman and J. H. Mather and E. E. Clothiaux, The k-distribution method and correlated-k approximation for a shortwave radiative transfer model, “J. Quant. Spectrosc. Radiat. Transfer”, vol. 62 pp. 109-121, (1999). P. Ricchiazzi and S. Yang and C. Gautier and D. Sowle, SBDART: A research and teaching software tool for plane-parallel radiative transfer in the Earth’s atmosphere, “Bull. Am. Met. Soc.”, vol. 79, pp. 2101-2114, (1998). Y.X. Hu and K. Stamnes, An accurate parameterization of the radiative properties of water clouds suitable for use in climate models, “J. Climate”, vol. 6 pp. 728-742, (1993). Aspelmeyer A. et al., Long-Distance Free-Space Distribution of Quantum Entanglement, Science, vol. 301, no. 5633, (2003) [^1]: forward Mie scattering depolarization being anyway negligible.

--- abstract: | Let $\Omega$ be a smooth bounded domain in $\mathbb{R}^{N}$ and let $m$ be a possibly discontinuous and unbounded function that changes sign in $\Omega$. Let $f:\left[ 0,\infty\right) \rightarrow\left[ 0,\infty\right) $ be a nondecreasing continuous function such that $k_{1}\xi^{p}\leq f\left( \xi\right) \leq k_{2}\xi^{p}$ for all $\xi\geq0$ and some $k_{1},k_{2}>0$ and $p\in\left( 0,1\right) $. We study existence and nonexistence of strictly positive solutions for nonlinear elliptic problems of the form $-\Delta u=m\left( x\right) f\left( u\right) $ in $\Omega$, $u=0$ on $\partial \Omega$. author: - 'T. Godoy, U. Kaufmann [^1]' - | \ [FaMAF, Universidad Nacional de Córdoba, (5000) Córdoba, Argentina]{} title: 'Existence of strictly positive solutions for sublinear elliptic problems in bounded domains [^2] [^3] [^4] ' --- Introduction ============ Let $\Omega\subset\mathbb{R}^{N}$, $N\geq1$, be a $C^{1,1}$ bounded domain. Our aim in this paper is to consider the question of existence of solutions for nonlinear problems of the form$$\left\{ \begin{array} [c]{ll}-\Delta u=mf\left( u\right) & \text{in }\Omega\\ u>0 & \text{in }\Omega\\ u=0 & \text{on }\partial\Omega, \end{array} \right. \label{prob}$$ where $m:\Omega\rightarrow\mathbb{R}$ is a function that changes sign in $\Omega$ and $f:\left[ 0,\infty\right) \rightarrow\left[ 0,\infty\right) $ is a continuous function satisfying H1. $f$ is nondecreasing, and there exist $k_{1},k_{2}>0$ and $p\in\left( 0,1\right) $ such that $k_{1}\xi^{p}\leq f\left( \xi\right) \leq k_{2}\xi^{p}$ for all $\xi\geq0$. As pointed out in [@nodea], the existence of strictly positive solutions for sublinear problems with indefinite nonlinearities as (\[prob\]) raises many interesting questions and is intriguing even in the one-dimensional case for various reasons. One of them is that the existence of (nontrivial) nonnegative solutions does not guarantee the strict positivity of such solutions (in contrast for example to superlinear problems, where they even belong to the interior of the positive cone). In fact, there are situations in which there exist nonnegative solutions which actually vanish in a subset of $\Omega$ (see e.g. [@bandle]). Another one is for instance that several *non-comparable* sufficient conditions on $m$ can be established for the existence of solutions for (\[prob\]) in the one-dimensional case under some evenness assumptions on $m$ (see [@nodea], Section 2), and these solutions may not be in the interior of the positive cone. The present work is a natural continuation of the research started in [@nodea], where $m$ was considered (when $N>1$) to be radially symmetric. Let us note that the nonlinearity studied there was $f\left( \xi\right) =\xi^{p}$. One of the most important differences between $\xi^{p}$ and the nonlinearities treated in this paper is that here (\[prob\]) is no longer homogeneous in $m$ (i.e. (\[prob\]) may admit a solution but $km$ may not ($k>0$ constant), and viceversa), and the homogeneity was crucial in the every existence proofs given in [@nodea]. We shall primarily rely on the well-known sub- and supersolution method in the presence of weak sub and supersolutions (see e.g. [@du], Theorem 4.9). One of the reasons is that the existence of supersolutions represent no difficulty, see Remark 2.3 below. In order to supply (strictly positive) subsolutions, we shall divide the domain in parts and construct subsolutions in each of them, and later check that they can be joined appropriately to get a subsolution in the entire domain. This last fact depends on obtaining estimates for the normal derivatives of these subsolutions on the boundaries of the subdomains. In [@nodea] these bounds could be computed rather explicitly making use of the radial symmetry of $m$ (and the fact that $\Omega$ was a ball) but in the present situation those computations cannot be done any more. Let us mention that here the key tool will be an estimate due to Morel and Oswald, see Lemma 2.1 below. In Theorem 3.1 we shall state a sufficient condition on $m$ for the existence of solutions of (\[prob\]), while in Theorem 3.2 we shall provide a local necessary condition and a global one in Corollary 3.3 under an additional assumption on $m$. We observe that this last condition is of similar type to the one in Theorem 3.1. In order to relate these results to others already existing, we mention that two necessary conditions were proved for some particular radial functions in [@nodea], Theorem 3.4 (see also Remark 3.5 there), and as far as we know there are no other results (other than the obvious condition $m^{+}\not \equiv 0$ implied by the maximum principle). Concerning the matter of sufficient conditions, the only theorem we found in the literature, apart from the ones proved in [@nodea] for $m$ radial, is that there exists a solution for (\[prob\]) provided that the solution of the linear problem $-\Delta\phi=m$ in $\Omega$, $\phi=0$ on $\partial\Omega$, satisfies $\phi>0$ in $\Omega$ (see [@jesusultimo], Theorem 4.4, or [@hand], Theorem 10.6). As a matter of fact, this even holds for linear second order elliptic operators with nonnegative zero order coefficient. We note however that the aforementioned condition is far from being necessary in the sense that there are examples of (\[prob\]) having a solution but with the corresponding $\phi$ satisfying $\phi<0$ in $\Omega$ (cf. [@nodea]). Let us finally mention that for $m$ smooth an $p\in\left( 0,1\right) $ it is known that the problem $-\Delta u=mu^{p}$ in $\Omega$, $u=0$ on $\partial\Omega$ admits a (nontrivial) nonnegative solution if and only if $m\left( x_{0}\right) >0$ for some $x_{0}\in\Omega$ (see e.g. [@bandle] or [@publi]). We conclude this introduction with some few words on the case of a general second order elliptic operator. We believe that at least some of the results presented here should still be true when $-\Delta$ is replaced by such differential operators. In fact, one can verify that except the use of Lemma 2.1, the proof of Theorem 3.1 can be carried out exactly as it is done here (with the obvious changes) in the case of a general operator. Hence, if a similar version of the aforementioned lemma holds for these operators (which a priori it is not clear since the proof makes use of the mean value properties for superharmonic functions), then an analogue of Theorem 3.1 can be proved in this case. *Acknowledgments*. The authors are pleased to thank the referee for her-his careful and detailed reading of the paper. Preliminaries ============= The following estimate appeared first in an unpublished work by Morel and Oswald ([@morel]), and a nice proof can be found in the paper of Brezis and Cabré, [@cabre], Lemma 3.2. $\qquad$ **Lemma 2.1.** *Let* $h\in L^{r}\left( \Omega\right) $*,* $r>N$*, and let* $u$ *be the solution of* $$\left\{ \begin{array} [c]{ll}-\Delta u=h & \text{\textit{in} }\Omega\\ u=0 & \text{\textit{on }}\partial\Omega. \end{array} \right. \label{h}$$ *Then there exists some* $c=c\left( \Omega\right) >0$ *such that* $$\begin{gathered} u\left( x\right) \geq c\delta_{\Omega}\left( x\right) \int_{\Omega}h\delta_{\Omega}\qquad\text{\textit{for all }}x\in\Omega,\\ \text{\textit{where }}\delta_{\Omega}\left( x\right) :=dist\left( x,\partial\Omega\right) .\end{gathered}$$ $\qquad$ The next result is also known (see e.g. Theorem 3.4 in [@uriel2]). We present a brief sketch of the proof for the sake of completeness. Let us note that the following proof is much simpler than the one given in [@uriel2]. We set $$P^{\circ}\overset{.}{=}\text{interior of the positive cone of }C^{1,\alpha }\left( \overline{\Omega}\right) \text{, }\alpha\in\left( 0,1\right) \text{.}$$ $\qquad$ **Lemma 2.2.** *Let* $m\in L^{r}\left( \Omega\right) $ *with* $r>N$ *and such that* $0\not \equiv m\geq0$, *and let* $f$ *satisfying H1. Then there exists* $v\in W^{2,r}\left( \Omega\right) \cap P^{\circ}$ *solution of* $$\left\{ \begin{array} [c]{ll}-\Delta v=mf\left( v\right) & \text{in }\Omega\\ v>0 & \text{in }\Omega\\ v=0 & \text{on }\partial\Omega. \end{array} \right. \label{z}$$ *Proof*. Let $\phi>0$ be the solution of $-\Delta\phi=m$ in $\Omega$ and $\phi=0$ on $\partial\Omega$. Then using the second inequality in H1 one can verify that for every $k>0$ large enough it holds that $k\left( \phi+1\right) $ is a supersolution of (\[z\]). On the other side, let $\varphi>0$ with $\left\Vert \varphi\right\Vert _{\infty}=1$ satisfying $$\left\{ \begin{array} [c]{ll}-\Delta\varphi=\lambda_{1}\left( m,\Omega\right) m\varphi & \text{in }\Omega\\ \varphi=0 & \text{on }\partial\Omega, \end{array} \right.$$ where $\lambda_{1}\left( m,\Omega\right) $ denotes the (unique) positive principal eigenvalue for $m$. It is easy to check employing the first inequality in H1 that $\varepsilon\varphi$ is a subsolution of (\[z\]) for all $\varepsilon>0$ sufficiently small, and the lemma follows. $\blacksquare$ $\qquad$ **Remark 2.3.** Let us mention that the construction of the supersolution made in the first part of the above proof still works if $m$ changes sign in $\Omega$, taking there $\phi$ as the solution of $-\Delta\phi=m^{+}$ in $\Omega$ and $\phi=0$ on $\partial\Omega$. (where as usual we write $m=m^{+}-m^{-}$ with $m^{+}=\max\left( m,0\right) $ and $m^{-}=\max\left( -m,0\right) $). Furthermore, this is also true for a strongly uniformly elliptic differential operator with nonnegative zero order coefficient. $\blacksquare$ Main results ============ **Theorem 3.1.** *Let* $\Omega_{0}$ *be a* $C^{1,1}$ *domain with* $\overline{\Omega}_{0}\subset\Omega$*, and let *$m\in L^{r}\left( \Omega\right) $ *with* $r>N$ *and* $0\not \equiv m\geq0$* in* $\Omega_{0}$*. Let* $k_{1}$, $k_{2}$ *be given by H1. There exist some* $C_{0},C_{1}>0$ *depending only on* $\Omega$ *and* $\Omega_{0}$ *such that if*$$\left\Vert m^{-}\right\Vert _{L^{r}\left( \Omega-\overline{\Omega}_{0}\right) }\leq\frac{k_{1}C_{0}}{k_{2}C_{1}^{1-p}}\int_{\Omega_{0}}m\delta_{\Omega_{0}}^{p+1}$$ *then* (\[prob\]) *has a solution* $u\in W^{2,r}\left( \Omega\right) $. *Proof*. Let $\Omega-\overline{\Omega}_{0}:=\Omega_{1}$. For $M>0$, we start constructing some $0\leq w\in W^{2,r}\left( \Omega_{1}\right) $ solution of $$\left\{ \begin{array} [c]{ll}-\Delta w=-m^{-}f\left( w\right) & \text{in }\Omega_{1}\\\begin{array} [c]{l}w=0\\ w=M \end{array} & \begin{array} [c]{l}\text{on }\partial\Omega\\ \text{on }\partial\Omega_{0}. \end{array} \end{array} \right. \label{w}$$ Let us first note that since by H1 $f\left( 0\right) =0$, it holds that $\underline{w}:=0$ is a subsolution of (\[w\]), and also since $f$ is nonnegative we have that $\overline{w}:=M$ is a supersolution of (\[w\]). It follows from Theorem 4.9 in [@du] that there exists some $w$ weak solution of (\[w\]) satisfying $0\leq w\leq M$. Furthermore, by standard arguments we may conclude that $w\in W^{2,r}\left( \Omega_{1}\right) $ (indeed, it is enough to note that if $z\in W^{2,r}\left( \Omega_{1}\right) $ is the unique solution of the problem $-\Delta z=-m^{-}f\left( w\right) $ in $\Omega_{1}$, $z=0$ on $\partial\Omega$ and $z=M$ on $\partial\Omega_{0}$, then the maximum principle implies that $z=w$). We claim now that there exists some $C>0$ depending only on $\Omega_{1}$ such that if $M:=\left[ Ck_{2}\left\Vert m^{-}\right\Vert _{L^{r}}\right] ^{1/\left( 1-p\right) }$ then $w>0$ in $\Omega_{1}$ ($k_{2}$ given by H1). To confirm this, let $\theta,\psi\in W^{2,r}\left( \Omega_{1}\right) $ be the unique solutions of$$\left\{ \begin{array} [c]{ll}\Delta\theta=0 & \text{in }\Omega_{1}\\\begin{array} [c]{l}\theta=0\\ \theta=1 \end{array} & \begin{array} [c]{l}\text{on }\partial\Omega\\ \text{on }\partial\Omega_{0}, \end{array} \end{array} \right. \qquad\quad\left\{ \begin{array} [c]{cc}-\Delta\psi=m^{-} & \text{in }\Omega_{1}\\ \psi=0 & \text{on }\partial\Omega_{1}. \end{array} \right.$$ From the Sobolev imbedding theorems and the $W^{2,r}$-theory for elliptic equations (e.g. [@grisvard], Theorem 2.4.2.5) we derive that $$\left\vert \psi\right\vert \leq\left\Vert \nabla\psi\right\Vert _{L^{\infty}}\delta_{\Omega_{1}}\leq\left\Vert \psi\right\Vert _{C^{1}}\delta_{\Omega_{1}}\leq c_{0}\left\Vert \psi\right\Vert _{W^{2,r}}\delta_{\Omega_{1}}\leq c_{1}\left\Vert m^{-}\right\Vert _{L^{r}}\delta_{\Omega_{1}}$$ for some $c_{1}=c_{1}\left( \Omega_{1}\right) >0$, and we also have that $\theta>c_{2}\delta_{\Omega_{1}}$ in $\Omega_{1}$ for some $c_{2}=c_{2}\left( \Omega_{1}\right) >0$. On the other hand, since $w\leq M$, recalling H1 we get that in $\Omega_{1}$ $$-\Delta\left( M\theta-k_{2}M^{p}\psi\right) =-m^{-}k_{2}M^{p}\leq-m^{-}k_{2}w^{p}\leq-m^{-}f\left( w\right) =-\Delta w$$ and so $$w\geq M\theta-k_{2}M^{p}\psi>\left( c_{2}M-c_{1}k_{2}M^{p}\left\Vert m^{-}\right\Vert _{L^{r}}\right) \delta_{\Omega_{1}}\qquad\text{in }\Omega_{1}$$ and the claim is proved. We fix for rest of the proof $M$ as in the aforementioned claim. Let $\nu$ denote the outward unit normal to $\partial\Omega_{0}$. Let us observe now that $$\begin{gathered} \left\vert \frac{\partial w}{\partial\nu}\right\vert \leq\left\Vert w\right\Vert _{C^{1}}\leq c_{0}\left\Vert w\right\Vert _{W^{2,r}}\leq c_{1}\left( M+\left\Vert m^{-}\right\Vert _{L^{r}}\left\Vert f\left( w\right) \right\Vert _{L^{\infty}}\right) \leq\label{deri}\\ c_{1}\left( M+k_{2}M^{p}\left\Vert m^{-}\right\Vert _{L^{r}}\right) \leq2c_{1}\left[ \max\left\{ 1,C\right\} k_{2}\left\Vert m^{-}\right\Vert _{L^{r}}\right] ^{1/\left( 1-p\right) }:=\nonumber\\ c_{3}\left[ c_{4}k_{2}\left\Vert m^{-}\right\Vert _{L^{r}}\right] ^{1/\left( 1-p\right) },\nonumber\end{gathered}$$ with $c_{3}$ and $c_{4}$ depending only on $\Omega_{1}$. On the other side, let $v>0$ be the solution of (\[z\]) with $\Omega_{0}$ in place of $\Omega$. Taking into account H1 and Lemma 2.1, there exists $c_{5}=c_{5}\left( \Omega_{0}\right) >0$ such that $v\geq c_{5}k_{1}\delta_{\Omega_{0}}\int_{\Omega_{0}}mv^{p}\delta_{\Omega_{0}}$ and so raising this inequality to the power $p$, multiplying by $m\delta_{\Omega_{0}}$ and integrating over $\Omega_{0}$ we obtain $\left( \int_{\Omega_{0}}mv^{p}\delta_{\Omega_{0}}\right) ^{1-p}\geq\left( c_{5}k_{1}\right) ^{p}\int_{\Omega_{0}}m\delta_{\Omega_{0}}^{1+p}$ and hence$$v\geq\left[ c_{5}k_{1}\int_{\Omega_{0}}m\delta_{\Omega_{0}}^{1+p}\right] ^{1/\left( 1-p\right) }\delta_{\Omega_{0}}\text{.}$$ Define now $u:=M+v$. Then $\partial u/\partial\nu\leq-\left[ c_{5}k_{1}\int_{\Omega_{0}}m\delta_{\Omega_{0}}^{p+1}\right] ^{1/\left( 1-p\right) }$ and $u=w$ on $\partial\Omega_{0}$. Hence, if we set $\omega:=u$ in $\overline{\Omega}_{0}$ and $\omega:=w$ in $\overline{\Omega}-\Omega_{0}$ it follows applying the divergence theorem (as stated e.g. in [@cuesta], p. 742) that $\omega$ is a weak subsolution of (\[prob\]) if $\partial u/\partial\nu\leq\partial w/\partial\nu$. Recalling (\[deri\]) this occurs if $$c_{3}^{1-p}c_{4}k_{2}\left\Vert m^{-}\right\Vert _{L^{r}}\leq c_{5}k_{1}\int_{\Omega_{0}}m\delta_{\Omega_{0}}^{p+1}$$ and thus, taking into account Remark 2.3, this ends the proof. $\blacksquare$ We denote with $B_{R}\left( x_{0}\right) $ the open ball in $\mathbb{R}^{N}$ centered at $x_{0}$ with radius $R$, and we write $\left( -\Delta\right) ^{-1}:L^{r}\left( \Omega\right) \rightarrow L^{\infty}\left( \Omega\right) $ for the solution operator of (\[h\]). We also set $$C_{N,p}:=\frac{\left( 1-p\right) ^{2}}{2\left( N\left( 1-p\right) +2p\right) }. \label{cnp}$$ **Theorem 3.2.** *Let* $m\in L^{r}\left( \Omega\right) $ *with* $r>N$*, let* $C_{N,p}$ *be given by* (\[cnp\]) *and let* $k_{1}$, $k_{2}$ *be given by H1. If there exists a solution* $u\in C\left( \overline{\Omega}\right) $ of (\[prob\]), *then* $$\begin{gathered} \frac{C_{N,p}}{\left\Vert \left( -\Delta\right) ^{-1}\right\Vert }\sup_{B_{R}\left( x_{0}\right) \in\mathfrak{B}}\left[ m_{R}R^{2}\right] <\frac{k_{2}}{k_{1}}\left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega\right) }\text{,}\qquad\text{\textit{where}}\label{nec}\\ \mathfrak{B}:=\left\{ B_{R}\left( x_{0}\right) \subset\Omega:m\leq0\text{ \textit{in }}B_{R}\left( x_{0}\right) \right\} \text{,}\qquad m_{R}:=\inf_{B_{R}\left( x_{0}\right) }m^{-}\text{.}\nonumber\end{gathered}$$ *Proof*. We proceed by contradiction. If (\[nec\]) does not hold, then there exists some $B_{R}\left( x_{0}\right) \in\mathfrak{B}$ such that $$\frac{C_{N,p}m_{R}R^{2}}{\left\Vert \left( -\Delta\right) ^{-1}\right\Vert }\geq\frac{k_{2}}{k_{1}}\left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega\right) }.\label{acsur}$$ Let $\beta:=1/\left( 1-p\right) $, and for $x\in\overline{B}_{R}\left( x_{0}\right) $ define $$w\left( x\right) :=\left[ k_{1}C_{N,p}m_{R}\left\vert x-x_{0}\right\vert ^{2}\right] ^{\beta}\text{.}$$ After some computations one can verify that $\Delta w\leq k_{1}m^{-}w^{p}$ in $B_{R}\left( x_{0}\right) $. Let $u$ be a solution of (\[prob\]). In particular, it holds that $\Delta u\geq k_{1}m^{-}u^{p}$ in $B_{R}\left( x_{0}\right) $. Also, taking into account H1, from (\[prob\]) we deduce that $$\left\Vert u\right\Vert _{L^{\infty}\left( \Omega\right) }\leq\left[ k_{2}\left\Vert \left( -\Delta\right) ^{-1}\right\Vert \left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega\right) }\right] ^{\beta}.\label{cita}$$ Moreover, if $x\in\partial B_{R}\left( x_{0}\right) $, employing (\[acsur\]) and (\[cita\]) we derive that$$w\left( x\right) =\left( k_{1}C_{N,p}m_{R}R^{2}\right) ^{\beta}\geq\left[ k_{2}\left\Vert \left( -\Delta\right) ^{-1}\right\Vert \left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega\right) }\right] ^{\beta}\geq\left\Vert u\right\Vert _{L^{\infty}\left( \Omega\right) }\geq u\left( x\right) .$$ It follows by the comparison principle that $w\geq u$ in $B_{R}\left( x_{0}\right) $, but $w\left( x_{0}\right) =0$, contradicting the fact that $u>0$ in $\Omega$. $\blacksquare$ **Corollary 3.3.** *Let* $\Omega_{1}\subset\Omega$ *be a convex domain and let* $m\in L^{r}\left( \Omega\right) $ *with* $r>N$ *and such that in* $\Omega_{1}$ $m$ *is convex and* $m\leq0$. *If there exists a solution* $u\in C\left( \overline{\Omega}\right) $ *of* (\[prob\]), *then* $$\frac{4C_{N,p}}{27\left\vert \Omega_{1}\right\vert \left\Vert \left( -\Delta\right) ^{-1}\right\Vert }\int_{\Omega_{1}}m^{-}\delta_{\Omega_{1}}^{2}<\frac{k_{2}}{k_{1}}\left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega-\overline{\Omega}_{1}\right) }\text{.} \label{coro}$$ *Proof*. Let $\alpha:=2/3$ and let $x_{1}\in\Omega_{1}$. We set $R_{1}:=\alpha\delta_{\Omega_{1}}\left( x_{1}\right) $ and let $y\in B_{R_{1}}\left( x_{1}\right) $. Observe that $z_{y}\left( t\right) :=x_{1}+t\left( y-x_{1}\right) \in\Omega_{1}$ for every $t\in\left[ 0,1/\alpha\right] $ since $\left\vert z_{y}\left( t\right) -x_{1}\right\vert <\delta_{\Omega_{1}}\left( x_{1}\right) $. Define $M\left( t\right) :=m^{-}\left( z_{y}\left( t\right) \right) $. Then $M\left( t\right) $ is concave in $\left[ 0,1/\alpha\right] $ and hence$$m^{-}\left( y\right) =M\left( 1\right) \geq\alpha M\left( 1/\alpha \right) +\left( 1-\alpha\right) M\left( 0\right) \geq\left( 1-\alpha\right) M\left( 0\right) =m^{-}\left( x_{1}\right) /3.$$ It follows that $\inf_{B_{R_{1}}\left( x_{1}\right) }m^{-}\geq m^{-}\left( x_{1}\right) /3$. Now, if (\[prob\]) possesses a solution ** $u\in C\left( \overline{\Omega}\right) $, by Theorem 3.2 we obtain that$$\begin{gathered} \frac{k_{2}}{k_{1}}\left\Vert m^{+}\right\Vert _{L^{r}\left( \Omega\right) }>\frac{C_{N,p}}{\left\Vert \left( -\Delta\right) ^{-1}\right\Vert }\sup_{B_{R}\left( x_{0}\right) \in\mathfrak{B}}\left[ m_{R}R^{2}\right] \geq\\ \frac{C_{N,p}}{3\left\Vert \left( -\Delta\right) ^{-1}\right\Vert }m^{-}\left( x_{1}\right) R_{1}^{2}=\frac{4C_{N,p}}{27\left\Vert \left( -\Delta\right) ^{-1}\right\Vert }m^{-}\left( x_{1}\right) \delta _{\Omega_{1}}^{2}\left( x_{1}\right)\end{gathered}$$ for every $x_{1}\in\Omega_{1}$. Integrating this inequality in $\Omega_{1}$ with respect to $x_{1}$ gives (\[coro\]) and thus the corollary is proved. $\blacksquare$ **Remark 3.4.** We observe that $C_{N,p}\rightarrow0$ when $p\rightarrow 1$ and thus (\[nec\]) and (\[coro\]) are satisfied for any $m$ provided that $p$ is close enough to $1$. Let us mention that this must occur since, at least when $m^{-}\in L^{\infty}\left( \Omega\right) $, $f\left( \xi\right) =\xi^{p}$, and either $N=1$ or $N>1$ and $m$ is radial with $0\not \equiv m\geq0$ in some $B_{r}\left( 0\right) $, it is known that (\[prob\]) has a solution ** if $p$ is sufficiently close to $1$ (cf. [@nodea], Theorems 2.1 (i) and 3.2). $\blacksquare$ [99]{} C. Bandle, M. Pozio, A. Tesei, *The asymptotic behavior of the solutions of degenerate parabolic equations,* Trans. Amer. Math. Soc. **303** (1987), 487-501.* * H. Brezis, X. Cabré, *Some simple nonlinear PDE’s without solutions*, Boll. Unione Mat. Ital. Sez. B Artic. Ric. Mat. (8) **1** (1998), 223–262. M. Cuesta, P. Takáč, *A strong comparison principle for positive solutions of degenerate elliptic equations*, Differential Integral Equations **13** (2000), 721–746. Y. Du, *Order structure and topological methods in nonlinear partial differential equations. Vol. 1. Maximum principles and applications*, World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2006. T. Godoy, U. Kaufmann, *On the existence of positive solutions for periodic parabolic sublinear problems,* Abstr. Appl. Anal. **2003** (2003), 975-984. T. Godoy, U. Kaufmann, *Periodic parabolic problems with nonlinearities indefinite in sign*, Publ. Mat. **51** (2007), 45-57. T. Godoy, U. Kaufmann, *On strictly positive solutions for some semilinear elliptic problems*, NoDEA Nonlinear Differ. Equ. Appl. **20** (2013), 779-795. P. Grisvard, Elliptic problems in nonsmooth domains. Monographs and Studies in Mathematics, 24. Pitman (Advanced Publishing Program), Boston, MA, 1985. J. Hernández, F. Mancebo, *Singular elliptic and parabolic equations*, M. Chipot (ed.) et al., Handbook of differential equations: Stationary partial differential equations. Vol. III. Amsterdam: Elsevier/North Holland. Handbook of Differential Equations, 317-400 (2006). J. Hernández, F. Mancebo, J. Vega, *On the linearization of some singular, nonlinear elliptic problems and applications*, Ann. Inst. H. Poincaré Anal. Non Linéaire **19** (2002), 777–813. J. Morel, L. Oswald, *A uniform formulation for the Hopf maximum principle*, preprint, 1985. [^1]: *E-mail addresses.* godoy@mate.uncor.edu (T. Godoy), kaufmann@mate.uncor.edu (U. Kaufmann, Corresponding Author). [^2]: 2000 *Mathematics Subject Clasification*. 35J25, 35J61, 35B09, 35J65. [^3]: *Key words and phrases*. Elliptic problems, indefinite nonlinearities, sub and supersolutions, positive solutions. [^4]: Partially supported by Secyt-UNC.

--- abstract: 'Results are presented for the electron current in gold chiral nanotubes (AuNTs). Starting from the band structure of $(4,3)$ and $(5,3)$ AuNTs, we find that the magnitude of the chiral currents are greater than those found in carbon nanotubes. We also calculate the associated magnetic flux inside the tubes and find this to be higher than the case of carbon nanotubes. Although (4,3) and (5,3) AuNTs carry transverse momenta of similar magnitudes, the low-bias magnetic flux carried by the former is far greater than that carried by the latter. This arises because the low-bias longitudinal current carried by a (4,3) AuNT is significantly smaller than that of a (5,3) AuNT.' author: - 'D. Zs. Manrique$^1$' - 'J. Cserti$^2$' - 'C. J. Lambert$^1$' title: Chiral currents in gold nanotubes --- [^1] Nanotubes and nanowires are of interest, not only because of their potential for deployment as interconnects, p-n junctions and rectifiers [@1; @2] in future nanoscale circuits, but also because they exhibit fundamental physical properties, such as conductance quantisation[@3; @4], and magic numbers reflecting structural stabilities[@5; @6]. One ubiquitous property associated with nanotubes is chirality, which arises because there is an infinite number of ways of rolling up a two-dimensional periodic lattice to form a cylinder. In addition to widely-studied carbon nanotubes [@7], chiral nanotubes have been formed from a range of other materials, including gold [@5; @8; @9], platinum [@10], silver [@11], alkaline metals [@12; @13; @14] and boron nitride [@15; @16]. These experimental observations have been supported by a range of theoretical investigations [@16b; @17; @18; @19; @20; @21]. It has recently been noted that the presence of intrinsic chiral electron currents in chiral nanotubes can be exploited to yield photogalvanic effects in heteropolar nanotubes [@16b], a new drive mechanism in carbon-nanotube windmills [@22] and to produce internal magnetic fields in carbon nanotube solenoids [@23] . In each of these examples, the underlying lattice is hexagonal, with two atoms per unit cell. In contrast nanotubes formed from gold, silver and platinum are derived from triangular lattices with one atom per unit cell and therefore it is of interest examine whether or not these effects are enhanced or diminished compared with their carbon counterparts. In this paper, to answer these questions, we examine chiral currents in gold nanotubes. Our choice of gold is in part motivated by the fact that chiral currents are expected to scale with the Fermi velocity of the underlying two-dimensional lattice, which in gold is approximately double that of graphene. A nanotube formed from a 2D lattice with periodic boundary conditions can be described by a chiral vector $\mathbf{C} = n \mathbf{a_1} + m \mathbf{a_2}$, which defines the circumference of the nanotube, where $\mathbf{a_1}$, $\mathbf{a_2}$ are the lattice vectors and $n,m$ are integers. The axis of the nanotube lies parallel to the longitudinal translation vector $\mathbf{T}$ (which is perpendicular the the chiral vector), whose magnitude is equal to the length of the nanotube unit cell, along the tube axis. To understand the currents carried by such a nanotube, we first calculate the electron group velocity components in the chiral and longitudinal directions. These are given by $\hbar v_{C} = \partial E(\mathbf{k})/\partial k_{C} $ and $\hbar v_{T} = \partial E(\mathbf{k})/\partial k_{T}$, where $k_C=\mathbf{k}.\mathbf{\hat{C}}$ and $k_T=\mathbf{k}.\mathbf{\hat{T}}$ are components of the wavevector $\mathbf{k}$ parallel to the unit vectors $\mathbf{\hat{C}}$ and $\mathbf{\hat{T}}$ respectively. In the presence of periodic boundary conditions, $k_C$ is quantized and takes the values $$k_C^{(q)}=\frac{2\pi q} {C}, \,\,\, q=N_0,N_0 + 1, N_0 +2, \ldots, N_0+N-1$$ where $N$ is the number of the mini-bands and $C=\vert\mathbf{{C}}\vert$. Since $q$ and $q+N$ are equivalent, $N_0$ is an arbitrary integer. In what follows $N_0$ is chosen such that if $N$ is even, $N_0=1-N/2$, whereas if $N$ is odd, $N_0=\frac{1-N}{2}$. We also choose $N$ to equal to the number of unit cells of the 2D lattice contained within a unit cell of the nanotube and therefore $-\frac{\pi}{T}\leq k_T<\frac{\pi}{T}$, where $T=\vert\mathbf{{T}}\vert$. The simplest model describing electronic properties of a triangular lattice consists of a tight-binding Hamiltonian with one orbital per atom. The dispersion relation of such an infinite sheet takes the form $$E(\mathbf{k})=-2\gamma (\cos \mathbf{k}.\mathbf{a_1} + \cos \mathbf{k}.\mathbf{a_2} + \cos \mathbf{k}.\mathbf{a_3} ), \label{disp1}$$ where $\gamma$ is the nearest neighbour coupling, $\mathbf{a_1}=-\frac{a}{2}(1,\sqrt{3})$ and $\mathbf{a_2}=a(1,0)$ are lattice vectors of the triangular lattice, $\mathbf{a_3}=\mathbf{a_1}+\mathbf{a_2}$ is an auxiliary vector and $a$ is the lattice constant. For a triangular lattice $\mathbf{T} = \frac{2m-n}{d} \mathbf{a_1} + \frac{m-2n}{d}\mathbf{a_2}$, where $d=GCD(2m-n,2n-m)$. ![$(4,3)$ A AuNT and a triangular sheet showing a $(4,3)$ chiral vector. Periodic boundary conditions make the two lattice points at the ends of $\mathbf{C}$ vector identical. The $\mathbf{T}$ vector shows the unit cell of the AuNT. The red lines shows the 1st, 2nd and 3rd nearest neighbours of the highlighted lattice point.[]{data-label="fig:trigraph"}](trigraph.eps) For the case of a nanotube formed by imposing periodic boundary conditions, $E(\mathbf{k})=E(k_C^{(q)},k_T)$ and $N=2(n^2+m^2-nm)/d$. To go beyond this simple model, we performed DFT calculations [@siesta; @dft] on AuNTs and a triangular gold lattice, whose lattice constant (obtained by relaxing the size of the unit cell) was found to be $a=2.73\text{\AA}$. By fitting the resulting band structures to a third-nearest neighbour model, we obtained a more accurate dispersion relation of the form $$E(k_C^{(q)},k_T)=\epsilon_0 - 2 \sum_{i}^{9} \gamma_{i} \cos(\alpha_{i} k_C^{(q)} + \beta_{i} k_T) \label{dispgen}$$ where $\alpha_i = \mathbf{\hat{C}}.\mathbf{a_i}$, $\beta_i = \mathbf{\hat{T}}.\mathbf{a_i}$ and $\mathbf{a_4}=\mathbf{a_1}+\mathbf{a_3}$, $\mathbf{a_5}=\mathbf{a_2}+\mathbf{a_3}$, $\mathbf{a_6}=\mathbf{a_2}-\mathbf{a_1}$, $\mathbf{a_7}=2\mathbf{a_1}$, $\mathbf{a_8}=2\mathbf{a_2}$, $\mathbf{a_9}=2\mathbf{a_3}$ auxiliary vectors. The values obtained for these couplings are shown in Table \[tablegamma\]. To account for curvature effects, nine different couplings are needed for AuNTs, whereas, due to symmetry, only three are needed for a flat sheet. With this choice of parameters, the ordering of the bands in eq (2) follows that obtained from DFT. In contrast, we found it impossible to obtain the correct ordering using the simple nearest-neighbour model of eq. (1). In what follows, we focus on the $(5,3)$ and $(4,3)$ AuNTs, since these are realistic experimental targets [@6; @23b]. As a consequence of curvature, we also expect that the Fermi energy of the tube will be shifted from that of the sheet. This shift is taken into account by an appropriate choice of $\epsilon_0$. Furthermore, since the effect of curvature depends on the choice of $n,m$, the couplings are also allowed to vary with $n,m$. The corresponding band structures are shown in Fig. \[fig:bands\]. [$\mathbf{n,m}$]{} [$\mathbf{\epsilon_{0}}$]{} [$\mathbf{\gamma_{1}}$]{} [$\mathbf{\gamma_{2}}$]{} [$\mathbf{\gamma_{3}}$]{} [$\mathbf{\gamma_{4}}$]{} [$\mathbf{\gamma_{5}}$]{} [$\mathbf{\gamma_{6}}$]{} [$\mathbf{\gamma_{7}}$]{} [$\mathbf{\gamma_{8}}$]{} [$\mathbf{\gamma_{9}}$]{} -------------------- ----------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- [sheet]{} [$\mathbf{0}$]{} [$\mathbf{4,3}$]{} [$\mathbf{0.6}$]{} [$\mathbf{1.13}$]{} [$\mathbf{0.5}$]{} [$\mathbf{1.56}$]{} [$\mathbf{-0.23}$]{} [$\mathbf{-0.14}$]{} [$\mathbf{-0.4}$]{} [$\mathbf{-0.19}$]{} [$\mathbf{-0.2}$]{} [$\mathbf{-0.13}$]{} [$\mathbf{5,3}$]{} [$\mathbf{0.69}$]{} [$\mathbf{0.74}$]{} [$\mathbf{1.07}$]{} [$\mathbf{0.42}$]{} [$\mathbf{-0.61}$]{} [$\mathbf{-0.21}$]{} [$\mathbf{-0.39}$]{} [$\mathbf{0.1}$]{} [$\mathbf{-0.18}$]{} [$\mathbf{0.38}$]{} : Fitted coupling parameters for a 2D triangular-lattice sheet with lattice constant $a=2.73\text{\AA}$ and for perfect $(4,3)$ and $(5,3)$ AuNTs (in eV).[]{data-label="tablegamma"} By differentiating the dispersion relation (\[dispgen\]) with respect to $k_C$ and $k_T$, the longitudinal group velocity (parallel to $\mathbf{\hat{T}}$) is found to be $$\hbar v_{T}^{(q)} (k_T)= 2 \sum_{i=1}^9 \gamma_i \beta_i \sin(\alpha_i k_C^{(q)}+\beta_i k_T) \label{labexactvt}$$ and in the transverse group velocity (parallel to $\mathbf{\hat{C}}$) is $$\hbar v_{C}^{(q)} (k_T)= 2 \sum_{i=1}^9 \gamma_i \alpha_i \sin(\alpha_i k_C^{(q)}+\beta_i k_T) \label{labexactvc}$$ As a reference velocity we note that the Fermi velocity for the triangular sheet(which varies by $\pm 10\%$ around the Fermi surface) has an average value of $v_F=1.8\times 10^6m/s$, when averaged over the Fermi surface. For an infinitely-long AuNT, we now compute the chiral velocities of right-moving electrons, (i.e. with $ v_{T}^{(q)} (k_T) > 0$), by first inverting eq. (\[dispgen\]) to obtain $k_T$ as a function of $E$ and $q$. We denote this inverse $k_T^{(q)+}(E)$, where $+$ sign refers to solutions belonging to branches with $ v_{T}^{(q)} > 0$. Real values of this function arise in the energy range $\varepsilon_<^{(q)} \leq E< \varepsilon_>^{(q)}$, where $\varepsilon_<^{(q)}$ is the bottom of positive-$ v_{T}^{(q)}$ branch of the $q$th mini-band and $\varepsilon_>^{(q)}$ is the top of the positive-slope branch of the $q$th mini-band. Substituting $k_T^{(q)+}(E)$ into eq. (\[labexactvt\]) yields the chiral velocity $v_{C}^{(q)+} (E)=v_{C}^{(q)} (k_T^{(q)+}(E))$ belonging to right-moving electrons of energy $E$. The total chiral velocity for the all right-moving electrons of energy $E$ is obtained by summing up the chiral velocities for each mini-band with a real $k_T^{(q)+}(E)$, to yield $$\begin{gathered} v_{C}^{\mathrm{(tot)}+}(E)= \\= \sum_{q=N_0}^{N_0+N-1} v_{C}^{(q)+} (E) \Theta(E-\varepsilon_<^{(q)})\Theta(\varepsilon_>^{(q)}-E) \label{vctotal}\end{gathered}$$ ![The top, second and third rows show chiral and longitudinal velocities and their ratios for each miniband of the (4,3) and (5,3) AuNTs. The red curves show those chiral velocity values belonging to right moving electrons. Those of left-moving electrons are shown by dotted curves. The thick black curve shows the sum of the corresponding red curves. and the chiral velocities (red lines) for each channel. The velocities are in units of average Fermi velocity $v_F$ of the triangular lattice sheet. The capital letters label individual open channels.[]{data-label="fig:evc"}](gEvfull.eps) To invert eq. (\[dispgen\]), we note that since $\beta_{i} k_T < 2\pi \frac{a}{T}$, and for most of the $n,m$ pairs, $T$ is much longer than $a$, the cosine functions can be approximated by parabolae, from which one can easily obtain an expression for $k_T$ as a function of energy, along with the allowed miniband energy ranges. To obtain an explicit form for the energy dependence of the transverse velocity near an energy band minimum, we note that in the sum on the right-hand-side of eq. (\[vctotal\]), if the condition ${\varepsilon}_<^{(q)} \leq E< {\varepsilon}_>^{(q)}$ is satisfied for a given $q$, then usually it is not satisfied for $-q$, because a mini-band with no local extremum in the Brillouin zone satisfies ${\varepsilon}_<^{(\pm q)}={\varepsilon}_>^{(\mp q)}$. [However]{} if it has local extremum in Brillouin zone, then both $q$ and $-q$ give contributions to the sum because if the miniband has a minimum, then ${\varepsilon}_<^{(q)}={\varepsilon}_<^{(-q)}$, and if it has maximum then ${\varepsilon}_>^{(q)}={\varepsilon}_>^{(-q)}$. Therefore in the case of a minimum, the beginning of the energy ranges of the minibands overlap and in case of maximum the end of the energy ranges overlap. Consequently channels open (close) in pairs when there is a band with local minimum (maximum). Just as a channel pair opens (closes) they give a combined contribution to the right-hand-side of eq. (\[vctotal\]). Noting that $\beta_i K_T \ll 2\pi$ and Taylor expanding (\[labexactvt\]) in $k_T$ around zero, one can find that the contribution from the channel pair $q,-q$ is a square-root function of $E$. The $q=0$ case, which belongs to the first channel and does not have partner, yields a contribution $$v_{C}^{(0)+}(E)\approx \frac{2}{\hbar} \sum_i^9 \gamma_i \alpha_i \beta_i \sqrt{\frac{E-\epsilon_0+2\sum_j^9\gamma_j}{\sum_j^9\gamma_j \beta_j^2}}.$$ ![The forces against bias voltage. The red curve belongs to $(4,3)$ AuNT and the blue curve belongs to the $(5,3)$ AuNT. []{data-label="fig:force"}](FoFigSmall.eps) ![The magnetic field against bias voltage. The red curves belongs to $(4,3)$ AuNT and the blue curve belongs to the $(5,3)$ AuNT. []{data-label="fig:magnetic"}](MFFigSmall.eps) The above square root behavior near the bottom of a miniband is also found in CNTs [@23]. For gold (4,3) and (5,3) AuNTs, the top row of Fig. \[fig:evc\] shows chiral velocities of individual minibands (red curves A-D) as function of energy, along with the total chiral velocity (black curves), obtained by summing the individual velocities. The second row of Fig. \[fig:evc\] shows the longitudinal velocities of individual minibands. The sign of the chiral velocities alternates with successive open channels, leading to oscillations in the total chiral velocity with energy. To illustrate the behaviour of these quantities when bands open and close, results are plotted over a wider energy range than that used in Fig. \[fig:bands\]. However results are only meaningful at low energies and therefore when predicting forces and magnetic fields in figures Fig. \[fig:force\] and Fig. \[fig:magnetic\], we revert to energies within $1$ eV of $E_F$. The chirality of right-moving electrons can play an important role in driving nano motors, because they can exert a torque on the AuNT. An estimate of the maximum possible tangential force is given by the total flux of tangential momentum associated with right-moving electron injected into a bias-voltage window $-\frac{U}{2}\, , \,\frac{U}{2}$. This takes the form $$\begin{gathered} F_C= \frac{2m_e}{h}\int_{-\frac{eU}{2}}^{+\frac{eU}{2}} {v}_{C}^{\mathrm{(tot)}+}(E)d E = \\ = \frac{2m_e}{h}\sum_{q=N_0}^{N_0+N-1} \int_{-\frac{eU}{2}}^{+\frac{eU}{2}} v_{C}^{(q)} (k_T^{(q)+}(E)) \times \\ \times \Theta(E-\varepsilon_<^{(q)+}) \Theta(\varepsilon_>^{(q)+}-E) d E\end{gathered}$$ Fig. \[fig:force\] shows the resulting force for $(4,3)$ and $(5,3)$ AuNTs. As expected, this is higher than in CNTs, because chiral velocity at the Fermi energy is finite and also, because the Fermi velocity of AuNTs (which sets the velocity scale) is greater than that of CNTs. A further consequence of the chirality is the presence of an induced magnetic field, given by [@23] $$\begin{gathered} B= \frac{2e}{h} \frac{\mu_0}{C} \sum_{q=N_0}^{N_0+N-1} \int_{-\frac{eU}{2}}^{+\frac{eU}{2}} \frac{v_C^{(q)+}(E)}{v_T^{(q)+}(E)} \times \\ \times \Theta(E-\varepsilon_<^{(q)+}) \Theta(\varepsilon_>^{(q)+}-E) d E .\end{gathered}$$ The integrand of this expression involves the ratio of the chiral to longitudinal velocities. For individual minibands, these ratios are shown in the third row of Fig. \[fig:evc\]. The resulting magnetic fields are shown in Fig. \[fig:magnetic\]. Comparison between Figs. \[fig:force\] and \[fig:magnetic\] shows that although (4,3) and (5,3) AuNTs carry transverse momenta of similar magnitudes, the low-bias magnetic flux carried by a (4,3) AuNT is far greater than that carried by the (5,3) AuNT. This arises because the velocity ratio of the former is significantly higher than the latter, even though they possess similar chiral velocities near the Fermi energy. We have calculated the chiral velocities carried by electrons in infinitely-long $(4,3)$ and $(5,3)$ AuNTs, which as shown in ref. [@23], can be a guide to the size of chiral currents in finite NTs connected to reservoirs. We have found that in a similar fashion to CNTs, chiral currents are oscillatory functions of energy, but unlike in CNTs, the chiral current has a finite value near $E_F$. Furthermore the tangential force and the induced magnetic field is higher than in comparable CNTs. We have considered perfectly-periodic nanotubes only. For the future it will be of interest to consider the effect of disorder on chiral currents. In this regard, one notes that at least in one dimension, disorder, which preserves the average spatial symmetry of a lattice, does not completely randomize the phase and density of current-carrying states [@cjl1; @cjl2]. Therefore, one expects chiral currents to persist in the presence of disorder, although with a diminished magnitude. Acknowledgements: Supported by the Hungarian Science Foundation OTKA contracts T48782 & 75529, EPSRC and the EU ITNs FUNMOLS & NANOCTM. [99]{} F. Leonard and J. Tersoff, Phys. Rev. Lett. [**83**]{} 5174 (1999) P.G. Collins, M.S. Arnold and Ph. Avouris, Science [**292**]{} 706 (2001) R. Saito et al., Appl. Phys. Lett. [**60**]{} 2204 (1992) N. Hamada, S.I. Sawada and A. Oshiyama, Phys. Rev. Lett. [**68**]{} 1579 (1992) Y. Kondo and K. Takayanagi, Science [**289**]{} 606 (2000) R.T. Senger, S. Dag and S. Ciraci, Phys. Rev. Lett., [**93**]{} 196807 (2004) S. Ijima, Nature (London) [**354**]{} 56 (1991) Y. Kondo and K. Takayanagi, Phys. Rev.Lett. [**79**]{} 606 (1997) Y. Oshima, A. Onga and K. Takayanagi, Phys. Rev. Lett., [**91**]{} 205503 (2003) Y. Oshima et al., Phys. Rev B, [**65**]{} 121401(R) (2002) V. Rodrigues, J. Bettini, A.R. Rocha, L.G.C. Rego and D. Ugarte, Phys. Rev B [**65**]{} 153402 (2002) A.I. Yanson, I.K. Yanson and J.M. van Ruitenbeek, Nature (London) [**400**]{} 144 (1999) A.I. Yanson, I.K. Yanson and J.M. van Ruitenbeek, Phys. Rev. Lett. [**84**]{} 5832 (2000) W. A. de Heer, Rev. Mod. Phys. [**65**]{} 611 (1993) N.G. Chopra et al., Science [**269**]{} 966 (1995) Z. Weng-Sieh et al., Phys. Rev. B [**51**]{} 11229 (1995) P. Kral, E.J. Mele and D. Tomanek, Phys. Rev. Lett. [**85**]{} 1512 (2000) E.Tossati et al., Science 291 [**288**]{} (2001) X. Yang and J. Dong, Phys. Rev. B [**71**]{} 233403 (2005) M. Del Valle, C. Tejedor and G. Cuniberti, Phys. Rev. [**B 74**]{} 045408 (2006) Y. Iguchi, T. Hoshi and T. Fujiwara, Phys. Rev. Lett. [**99**]{} 125507 (2007) S. Konar and B. C. Gupta, Phys. Rev. B [**78**]{} 235414 (2008) S.W.D. Bailey, I. Amanatidis and C.J. Lambert, Phys. Rev. Lett. [**100**]{} 256802 (2008) C.J. Lambert and S.W.D. Bailey and J. Cserti, Phys. Rev. B [**78**]{} 233405 (2008) C-K Yang, Appl. Phys. Lett. [**85**]{} 2923 (2004) J. M. Soler, E. Artacho, J. D. Gale, A. García, J. Junquera, P. Ordejón, and D. Sánchez-Portal, J. Phys.: Condens. Matter [**14**]{}, 2745 (2002) DFT calculation was carried out within GGA, using a DZP basis set and a force tolerance for coordinate relaxation of 0.01 eV/Ang C. Lambert, J. Phys. C. [**17**]{}, 2401 (1984) C. Lambert, Phys. Rev. B. [**29**]{}, 1091 (1984) [^1]: e-mail: c.lambert@lancaster.ac.uk

--- author: - | $^a$[Chuancun Yin ]{} and $^b$[Kam Chuen Yuen ]{}\ [*$^a$ School of Statistics, Qufu Normal University,*]{}\ [*Shandong 273165, P.R. China* ]{}\ [*Corresponding author: E-mail: ccyin@mail.qfnu.edu.cn* ]{}\ \[3mm\] [*$^b$ Department of Statistics and Actuarial Science, The University of Hong Kong,*]{}\ [*Pokfulam Road, Hong Kong* ]{}\ [*E-mail: kcyuen@hku.hk*]{} title: '**Optimal dividend problems for a jump-diffusion model with capital injections and proportional transaction costs**' --- [**Abstract**]{}    In this paper, we study the optimal control problem for a company whose surplus process evolves as an upward jump diffusion with random return on investment. Three types of practical optimization problems faced by a company that can control its liquid reserves by paying dividends and injecting capital. In the first problem, we consider the classical dividend problem without capital injections. The second problem aims at maximizing the expected discounted dividend payments minus the expected discounted costs of capital injections over strategies with positive surplus at all times. The third problem has the same objective as the second one, but without the constraints on capital injections. Under the assumption of proportional transaction costs, we identify the value function and the optimal strategies for any distribution of gains. [*Key words and phrases*]{}. Barrier strategy, dual model, HJB equation, jump-diffusion, optimal dividend strategy, stochastic control. [*Mathematics Subject Classification (2000)*]{}. Primary: 93E20, 91G80 Secondary: 60J75. INTRODUCTION {#intro} ============ For the optimal dividend problem, one may adopt the objective of maximizing the expectation of the discounted dividends until possible ruin. This problem was first addressed by De Finetti \[16\] who considered a discrete time risk model with step sizes $\pm 1$ and showed that the optimal dividend strategy is a barrier strategy. Miyasawa \[21\] generalized the model to the case that periodic gains of a company can take on values $-1,0,1,2,3,\cdots$, and showed that the optimal dividend strategy of the generalized model is a barrier one. Subsequently, the problem of finding the optimal dividend strategy has attracted great attention in the literature of insurance mathematics. For nice surveys on this topic, we refer the reader to Avanzi \[3\] and Schmidli \[22\]. Besides insurance risk models, the optimal dividend problem in the so-called dual model has also been studied extensively in recent years. Among others, Avanzi et al. \[6\] discussed how the expectation of the discounted dividends until ruin can be calculated for the dual model when the gain amounts follow an exponential distribution or a mixture of exponential distributions, and showed how the exact value of the optimal dividend barrier can be determined; and Avanzi and Gerber \[5\] examined the same problem for the dual model that is perturbed by diffusion, and showed that the optimal dividend strategy in the dual model is also a barrier strategy. To make the problem more interesting, the issue of capital injections has also been considered in the study of optimal dividends in the dual model. Yao et al. \[23\] studied the optimal problem with dividend payments and issuance of equity in the dual model with proportional transaction costs, and derived the optimal strategy that maximizes the expected present value of dividend payments minus the discounted costs of issuing new equity before ruin. Yao et al. \[24\] considered the same problem with both fixed and proportional transaction costs. Dai et al. \[14,15\] investigated the same problem as in Yao et al. \[23\] for the dual model with diffusion with bounded gains and exponential gains, respectively. Avanzi et al. \[7\] derived an explicit expression for the value function in the dual model with diffusion when the gains distribution in a mixture of exponentials in the presence of both dividends and capital injections. Specifically, they showed that barrier dividend strategy is optimal, and conjectured that the optimal dividend strategy in the dual model with diffusion should be the barrier strategy regardless of the distribution of gains. Bayraktar et al. \[11\] examined the same cash injection problem, and used the fluctuation theory of spectrally positive Lévy processes to show the optimality of the barrier strategy for all positive Lévy processes. Bayraktar et al. \[12\] extended the study to the case with fixed transaction costs. Other related work can be found in Yin and Wen \[26\], Yin, Wen and Zhao \[28\], Avanzi et al. \[8\], Yao et al. \[25\] and Zhang \[29\]. In this paper, we provide a uniform mathematical framework to analyze the optimal control problem with dividends and capital injections in the presence of proportional transaction costs for the dual model with random return on investment. The associated value function is defined as the expected present value of dividends minus costs of capital injections until ruin. The rest of the paper is organized as follows. In Section 2, we give a rigorous mathematical formulation of the problem. Section 3 works on the model without capital injections, while Section 4 deals with the model with capital injections which never goes bankrupt. Finally, we solve the general stochastic control problem in Section 5. Problem formulation {#math} =================== Assume that the surplus generating process $P_t$ at time $t$ is given by $$P_t=x-pt+\sigma_p W_{p,t}+\sum_{i=1}^{N_{t}}X_{i},\ \ t\ge 0,$$ where $x>0$ is the initial assets, $p$ and $\sigma_p$ are positive constants, $\{W_{p,t}\}_{t\ge 0}$ is a standard Brownian motion independent of the homogeneous compound Poisson process $\sum_{i=1}^{N_{t}}X_{i}$, and $\{X_{i}\}$ is a sequence of independent and identically distributed random variables having common distribution function $F$ with $F(0)=0$. Let $\lambda$ be the intensity of the Poisson process $N_{t}$. We assume throughout the paper that $E[X_{i}]<\infty$ and $\lambda E[X_i]-p>0$. Here, we consider the return on investment generating process $$R_t=rt+\sigma_R W_{R,t},\ \ t\ge 0,$$ where $\{ W_{R,t}\}_{t\ge 0}$ is another standard Brownian motion, and $r$ and $\sigma_R$ are positive constants. It is assumed that $W_{p,t}$ and $W_{R,t}$ are correlated in the way that $$W_{R,t}=\rho W_{p,t} +\sqrt{1-\rho^2}W_{p,t}^0,$$ where $\rho\in [-1,1]$ is constant, and $W_{p,t}^0$ is a standard Brownian motion independent of $W_{p,t}$. Define the risk process $U_t$ as the total assets of the company at time $t$, i.e., $U_t$ is the solution to the stochastic differential equation $$U_t=P_t+\int_0^t U_{s-}\text{d}R_s,\;\; t\ge 0.\label{math-eq0}$$ The solution to (2.3) is given by (see, e.g. Jaschke \[19, Theorem 1\]) $$U_t={\cal E}(R)_t\left(x+\int_0^t{\cal E}(R)_{s-}^{-1}\text{d}P_s-\rho\sigma_p\sigma_R \int_0^t{\cal E}(R)_{s-}^{-1}\text{d}s\right),\nonumber$$ where $${\cal E}(R)_t=\exp\{(r-\frac12\sigma_R^2)t+\sigma_R W_{R,t}\}.$$ Using It$\hat{\text{\rm o}}$’s formula for semimartingale, one can show that the infinitesimal generator $\cal{L}$ of $U=\{U_t,t\ge 0\}$ is given by $$\begin{aligned} {\cal L} g(y)=(ry-p) g'(y) &+& \frac12 \left[(\sigma_p+\rho \sigma_R y)^2+\sigma^2_R (1-\rho^2)y^2\right]g''(y)\nonumber\\ &+&\lambda\int_{0}^{\infty}[g(y+z)-g(y)]F(\text{d}z). \label{math-eq1} \end{aligned}$$ The model (2.3) is a natural extension of the dual model in Avanzi and Gerber \[5\] and Avanzi et al. \[6\]. As was mentioned in Avanzi et al. \[6\], the dual model is appropriate for companies that have deterministic expenses and occasional gains whose amount and frequency can be modelled by the jump process $\sum_{i=1}^{N_{t}}X_{i}$. For example, for companies such as pharmaceutical or petroleum companies, the jump could be interpreted as the net present value of future gains from an invention or discovery. Another example is the venture capital investments or research and development investments. Venture capital funds screen out start-up companies and select some companies to invest in. When there is a technological breakthrough, the jump is generated. More examples can be found in Bayraktar and Egami \[10\] and Avanzi and Gerber \[5\]. In this paper, we denote by $L_t$ the cumulative amount of dividends paid up to time $t$ with $L_{0-}=0$, and by $G_t$ the total amount of capital injections up to time $t$ with $G_{0-}=0$. A dividend control strategy $\xi$ is described by the stochastic process $\xi=(L_t, G_t )$. A strategy is called admissible if both $L$ and $G$ are non-decreasing $\{\cal{F}$$_t\}$-adapted processes, and their sample paths are right-continuous with left limits. We denote by $\Xi$ the set of all admissible dividend policies. The risk process with initial capital $x\ge 0$ and controlled by a strategy $\xi$ is given by $U^{\xi}=\{U_t^{\xi}, t\ge 0\}$, where $U_t^{\xi}$ is the solution to the stochastic differential equation $$\text{d}U^{\xi}_t=\text{d}P_t+ U^{\xi}_{t-}\text{d}R_t-\text{d}L_t^{\xi}+\text{d}G_t^{\xi},\;\; t\ge 0.\nonumber$$ Moreover, $L_{t}^{ \xi}-L_{t-}^{ \xi}\le U_{t-}^{ \xi}$ for all $t$. In words, the amount of dividends is smaller than the size of the available capitals. Let $\tau^{\xi}=\inf\{t\ge 0: U^{\xi}_t=0\}$ be the ruin time. Then, the associated performance function is given by $$V(x;\xi)=E_x\left(\alpha\int_{0-}^{\tau^{ \xi}-}\text{e}^{-\delta t}dL_t^{ \xi}-\beta \int_{0-}^{\tau^{ \xi}-}\text{e}^{-\delta t}dG_t^{ \xi}\right),\label{math-eq4}$$ where $\delta>0$ is the discounted rate, $1-\alpha$ ($0<\alpha\le 1$) is the rate of proportional costs on dividend transactions, $1\le \beta<\infty$ is the rate of proportional transaction costs of capital injections. The notation $E_x$ represents the expectation conditioned on $U^{\xi}_0=x$ and the integral is understood pathwise in a Lebesgue-Stieltjes sense. Our aim is to find the value function $$V_*(x)=\sup_{ \xi\in\Xi}V(x; \xi),\label{math-eq5}$$ and the optimal policy $ \xi^*\in\Xi$ such that $V(x;\xi^*)=V_*(x)$ for all $x\ge 0$. The study of optimal dividends has been around many years. The commonly-used approach to solving these optimal control problems is to proceed by guessing a candidate optimal solution, constructing the corresponding value function, and subsequently verifying its optimality through a verification result. For the model of study, i.e., an upward jump-diffusion process with random return on investment, the optimal control problem remains to be solved. The problem of study can be seen as a natural extension of Bayraktar and Egami \[10\], and Avanzi, Shen and Wong \[7\]. In addition, one can see later that the method used in Bayraktar, Kyprianou and Yamazaki \[11\] cannot be applied to our model since their proof relies on certain characteristics of Lévy process. In order to solve the optimal control problem in this paper, we shall first consider two sub-optimal problems in the next two sections. 0.2cm Optimal dividend problem without capital injections {#without} =================================================== In this section, we first consider the dividend problem without capital injections. We shall show that the barrier strategy solve the optimal dividend problem regardless of the jump distribution. Let $\Xi_d=\{\xi_d=(L^{\xi_d}, G^{\xi_d}): (L^{\xi_d}, G^{\xi_d})\in \Xi\; {\rm and}\; G^{\xi_d}\equiv 0\}$. The associated controlled process is denoted by $U^{\xi_d}=\{U_t^{\xi_d},t\ge 0\}$, where $U_t^{\xi_d}$ is the solution to the stochastic differential equation $$\text{d}U^{\xi_d}_t=\text{d}P_t+ U^{\xi_d}_{t-}\text{d}R_t-\text{d}L_t^{\xi_d},\;\; t\ge 0.\nonumber$$ and the value function is given by $$V_{d}(x)=\sup_{ \xi_d\in\Xi_d}V(x; \xi_d)\equiv \sup_{ \xi_d\in\Xi_d}E_x\left(\alpha\int_{0-}^{\tau_{\xi_d}-} e^{-\delta t}dL_t^{\xi_d}\right),\; x\ge 0, \label{without-eq1}$$ where $\tau_{\xi_d}=\inf\{t: U_t^{\xi_d}=0\}$ is the time of ruin under the strategy $\xi_d$. We next identify the form of the value function $V_d$ and the optimal strategy $\xi_d^*$ such that $V_d(x)=V(x; \xi_d^*)$. [**3.1 HJB equation and verification lemma**]{} For notational convenience, denote $v(x)=V(x; \xi_d^*)$. If $v$ is twice continuously differentiable, then applying standard arguments from stochastic control theory (see Fleming and Soner \[17\]) or an approach similar to that in Azcue and Muler \[9\], we can show that the value function fulfils the dynamic programming principle $$v(x)=\sup_{\xi_d\in\Xi}E_x\left(\int_0^{{\tau_{\xi_d}}\wedge T}e^{-\delta s}dL_s^{\xi_d}+e^{-\delta (\tau_{\xi_d}\wedge T)}v(U_{\tau_{\xi_d}\wedge T}^{\xi_d})\right),$$ for any stopping time $T$, and that the associated Hamilton-Jacobi-Bellman (HJB) equation is $$\max\{{\cal L} v(x)-\delta v(x),\; \alpha-v'(x)\}=0, \; x>0,\label{without-eq2}$$ with $v(0)=0$, where ${\cal L}$ is the the extended generator of $U$ defined in (\[math-eq1\]). The HJB equation (\[without-eq2\]) can also be obtained by the heuristic argument of Avanzi et al. \[7\]. (Verification Lemma) Let $v$ be a solution to (3.2). Then, $v(x)\ge V(x; \xi_d)$ for any admissible strategy $\xi_d\in\Xi_d$, and thus $v(x)\ge V_{d}(x).$ [**Proof.**]{} For any admissible strategy $\xi_d\in\Xi_d$, put $\Lambda=\{s: L^{\xi_d}_{s-}\neq L^{\xi_d}_s\}$. Applying Ito’s formula for semimartingale to $e^{-\delta t}v(U^{\xi_d}_{t})$ gives $$\begin{aligned} E_x[e^{-\delta (t\wedge \tau_{\xi_d}-)}v({U}^{\xi_d}_{t\wedge \tau_{\xi_d}-})] &=&v(x)+ E_x\int_0^{t\wedge \tau_{\xi_d}-} e^{-\delta s}({\cal L}-\delta)v({U}^{\xi_d}_{s-})ds \nonumber\\ &&+ E_x\sum_{s\in \Lambda,s\le t\wedge \tau_{\xi_d}-}e^{-\delta s}\left\{v({U}^{\xi_d}_{s})-v({U}^{\xi_d}_{s-})\right\} \nonumber \\ &&- E_x\int_{0-}^{t\wedge \tau_{\xi_d}-} e^{-\delta s}v'({U}^{\xi_d}_{s-})d{L}^{\xi_d,c}_s,\end{aligned}$$ where ${L}^{\xi_d,c}_s$ is the continuous part of ${L}^{\xi_d}_s$. From (3.2), we see that $({\cal L}-\delta)v({U}^{\xi_d}_{s-})\le 0$ and $v'(x)\ge \alpha$. Thus, for $s\in \Lambda,s\le t\wedge \tau_{\xi_d}$, $$v({U}^{\xi_d}_{s})-v({U}^{\xi_d}_{s-})\le -\alpha ({L}^{\xi_d}_s-{L}^{\xi_d}_{s-}).$$ It follows from (3.3) and (3.4) that $$E_x[e^{-\delta (t\wedge \tau_{\xi_d}-)}v({U}^{\xi_d}_{t\wedge \tau_{\xi_d}-})]\le v(x)-\alpha E_x \int_{0-}^{t\wedge \tau_{\xi_d}-} e^{-\delta s}d{L}^{\xi_d}_s.$$ Letting $t\to\infty$ in (3.5) yields the result. $\Box$ [**3.2 Construction of a candidate solution**]{} It is assumed that dividends are paid according to the barrier strategy $\xi_b$. Such a strategy has a level of barrier $b>0$. When the surplus exceeds the barrier, the excess is paid out immediately as dividends. Let $L_t^b$ be the total amount of dividends up to time $t$. The controlled risk process when taking into account of the dividend strategy $\xi_b$ is $U^{b}=\{U_t^b, t\ge 0\}$, where $U_t^b$ is the solution to the following stochastic differential equation $$\text{d}U^b_t=\text{d}P_t+ U^b_{t-}\text{d}R_t-\text{d}L_t^b,\;\; t\ge 0.\nonumber$$ Denote by $V_b(x)$ the expected discounted dividends function if the barrier strategy $\xi_b$ is applied, that is, $$V_b(x)=\alpha E_x \left( \int_{0-}^{T^x_b-}e^{-\delta t}dL^b_t\right),$$ where $\delta>0$ is the force of interest and $T^x_b=\inf\{t\ge0: U^b_t=0\}.$ The following result shows that $V_b(x)$ as a function of $x$ satisfies an integro-differential equation with certain boundary conditions. \[thrm3-1\] For the risk process $U$ of (2.3) and the infinitesimal generator $\cal{L}$ of (2.4), if $h_b(x)$ solves $${\cal{L}} h_b(x)=\delta h_b(x), \quad 0< x <b,$$ and $h_b(x) = h_b(b) + \alpha(x-b)$, for $x>b$, together with the boundary conditions $$h_b(0) = 0, \quad h_b'(b) = \alpha,$$ then $h_b(x)$ coincides with $V_b(x)$ given by (3.6). [**Proof.**]{}   Applying Ito’s formula for semimartingale to $e^{-\delta t}h_b(U^b_{t-})$ gives $$\begin{aligned} e^{-\delta t}h_b({U}^{b}_{t-})&-&h_b({U}^{b}_0) = \int_{0-}^{t-} e^{-\delta t}dN_s^{b}+\int_0^t e^{-\delta s}({\cal L}-\delta)h_b({U}^{b}_{s-})ds \nonumber \\ &+ &\sum_{s< t}\text{\bf 1}_{\{\triangle {L}_s>0\}}e^{-\delta s}\left\{h_b({U}^{b}_{s-}+\triangle P_s-\triangle {L}_s)-h_b({U}^{b}_{s-}+\triangle P_s)\right\} \nonumber \\ &-& \int_{0-}^{t-} e^{-\delta s}h_b'({U}^{b}_{s-})d{L}_s^{c},\label{main-eq3}\end{aligned}$$ where ${L}_s^{c}$ is the continuous part of ${L}_s$, and $$\begin{aligned} N_t^{b}&=&\sum_{s\le t}\text{\bf 1}_{\{|\triangle P_s|>0\}}\left\{h_b({U}^{b}_{s-}+\triangle P_s)-h_b({U}^{b}_{s-}) \right\}\\ &&-\int_0^t\int_{0}^{\infty}\left\{h_b({U}^{b}_{s-}+y) -h_b({U}^{b}_{s-})\right\}\Pi(dy)ds\\ &&+\sigma \int_0^t h_b'({U}^{b}_{s-})dW_s.\end{aligned}$$ Note that $P(\triangle{L}_s>0, \triangle P_s<0)=0$ and that ${U}^{b}_{s-}+\triangle P_s\ge {U}^{b}_{s-}+\triangle P_s-\triangle {L}_s\ge b $ on $\{\triangle {L}_s>0, \triangle P_s>0\}$. Consequently, $$\begin{aligned} &\sum_{s< t}\text{\bf 1}_{\{\triangle {L}_s>0\}}e^{-\delta s}\left\{h_b({U}^{b}_{s-}+\triangle P_s-\triangle {L}_s)-h_b({U}^{b}_{s-}+\triangle P_s)\right\}\\ &= -\alpha\sum_{s<t}\text{\bf 1}_{\{\triangle {L}_s>0\}}e^{-\delta s}\triangle {L}_s.\end{aligned}$$ Note that $N_t^b$ is a local martingale, and $$\int_{0-}^{t-} e^{-\delta s}h_b'({U}^{b}_{s-})d{L}_s^{c}= \int_{0-}^{t-}e^{-\delta s}h_b'({U}^{b}_{s})d{L}_s^{c}=\alpha \int_{0-}^{t-} e^{-\delta s}h_b'(b)d{L}_s^{c}.$$ Thus, for any appropriate localization sequence of stopping times $\{t_n, n\ge 1\}$, we have $$E_x (e^{-\delta (t_n\wedge T^{b})}h_b({U}^{b}_{t_n\wedge T^{b}}))-E_x h_b({U}^{b}_0)= -\alpha E_x\int_{0-}^{t_n\wedge T^{b}-} e^{-\delta s} d {L}_s.\label{main-eq4}$$ Letting $n\to\infty$ in (\[main-eq4\]) yields the result. $\Box$ \[without-4\] $V_{b}(x)$ is a concave increasing function on $(0, \infty)$. [**Proof.**]{}   To prove the lemma, we use arguments similar to those in Kulenko and Schmidli \[20\]. Let $x>0, y>0$, and $l\in (0,1)$. Consider the strategies $L^x$ and $L^y$ for the initial capitals $x$ and $y$. Define $L_t=l L^x_t+(1-l) L^y_t$. Then, $L_t=L_t^{lx+(1-l)y}.$ Since the processes $\{P_t, t\ge 0\}$ and $\{R_t, t\ge 0\}$ have no negative jumps, we have $\tau_{L}=\tau_{L^x}\vee \tau_{L^y}$. It follows that $$\begin{aligned} V_b(lx+(1-l)y)&=&\alpha E_x \left(\int_{0-}^{{\tau_L}-}e^{-\delta t}dL_t\right)\\ &=& \alpha l E_x \left(\int_{0-}^{{\tau_L}-}e^{-\delta t}dL^x_t\right)+\alpha (1-l) E_x \left(\int_{0-}^{{\tau_L}-}e^{-\delta t}dL^y_t\right)\\ &&\ge \alpha l E_x \left(\int_{0-}^{{\tau_{L^x}}-}e^{-\delta t}dL^x_t\right) + \alpha (1-l) E_x \left(\int_{0-}^{{\tau_{L^y}}-}e^{-\delta t}dL^y_t\right)\\ &=&l V_b(x)+(1-l)V_b (y),\end{aligned}$$ and thus the concavity of $V_b$ follows. The increasingness of $V_{b}(x)$ is trivial $\Box$ [**3.3 Verification of optimality**]{} Define the barrier level by $$b^*=\sup\{b\ge 0: V_{b}'(b-)=\alpha\}.$$ We conjecture that the barrier strategy $\xi_{b^*}$ is optimal. \[without-2\] $b^*=0$ if and only if $\lambda\int_0^{\infty}y F(dy)\le p$. [**Proof.**]{}  Here, we follow the approach of Yao et al. \[23\] to prove the proposition. Suppose that $b^*=0$. Then, the associated value function is $V_d(x)=\alpha x$ which satisfies the HJB equation (3.2). As a result, we obtain $(\Gamma -\delta)V_d(x)\le 0$ which in turn gives $\lambda\int_0^{\infty}y F(dy)\le p$. On the other hand, suppose that $\lambda\int_0^{\infty}y F(dy)\le p$. Then, $w(x)=\alpha x$ satisfies (3.2). By Lemma 3.1, we get $w(x)\ge V_{d}(x)$. However, $w(x)\le V_{d}(x)$ since $w(x)=\alpha x$ is the performance function associated with the strategy that $x$ is paid immediately as dividends. In this case, ruin occurs immediately. Thus, $w(x)= V_{d}(x)$ and the optimal barrier level $b^*=0$. $\Box$ \[without-5\] If $\lambda\int_0^{\infty}y F(dy)>p$, then the function $V_{b^*}$ defined in (3.6) satisfies $$V_{b^*}(x)=V_d (x), \quad x\ge 0,$$ and the optimal barrier strategy $\xi_{d}^{*}$ is the solution to $$\begin{aligned} && \text{d}U_t^{\xi^*_d}=\text{d}P_t+ U_{t-}^{\xi^*_d}\text{d}R_t-\text{d} L_t^{\xi^*_d}, \qquad t\ge 0,\end{aligned}$$ with the conditions $$\begin{aligned} U_t^{\xi^*_d}\le b^*, \quad G_t^{\xi_{d}^{*}}\equiv 0, \quad \int_0^{\infty}{\bf 1}_{\{U_s^{\xi^*_d}<b^*\}}dL_s^{\xi_d^*}=0.\end{aligned}$$ [**Proof.**]{}   Using the method of Avanzi and Gerber \[5\], it can be shown that $V_{b^*}(x)$ is twice continuously differentiable at $x=b^*$. Consequently, $V_{b^*}\in C^2(\Bbb{R}_+)$. Note that $({\cal L}-\delta)V_{b^*}(x)=0$ and $V_{b^*}'(x)\ge \alpha$ for $x\in [0,b^*)$ due to the concavity of $V_{b^*}$ on $[0,b^*)$. Since $V_{b^*}(x)=\alpha(x-b^*)+ V_{b^*}(b^*)$ for $x\ge b^*$, we have $$\begin{aligned} ({\cal L}-\delta)V_{b^*}(x)&=&-p\alpha+\alpha\int_0^{\infty}y F(dy)-\alpha (x-b^*)-\delta V_{b^*}(b^*)\\ &&<-p\alpha+\alpha\int_0^{\infty}y F(dy)-\delta V_{b^*}(b^*)\\ &=&\lim_{x\to b^*+}({\cal L}-\delta)V_{b^*}(x)=\lim_{x\to b^*-}({\cal L}-\delta)V_{b^*}(x)=0,\end{aligned}$$ because of the continuity of $V_{b^*}, V_{b^*}'$, and $V_{b^*}''$ at $x=b^*$. Thus, the function $V_{b^*}$ satisfies the HJB equation (3.2). Then, it follows from Lemma 3.1 that $ V_{b^*}(x)\ge V_{d}(x)$. However, $V_{b^*}(x)\le V_{d}(x)$ by definition, and hence $V_{b^*}(x)=V_{d}(x)$. $\Box$ [**3.4 Two closed-form solutions**]{} Owing to the complexity of the equation, the solution may not be available in explicit form in general. The following two examples show that one can derive closed-form solution in some special cases. [**Example 3.1.**]{}  Assume that $r=0$ and $\sigma_R=0$. Then, $V_{b^*}(x)$ satisfies the following integro-differential equation $${\cal A}V_{b^*}(x)=\delta V_{b^*}(x), \quad 0<x<b^*,$$ and $$V_{b^*}(x)=\alpha(x-b^*)+ V_{b^*}(b^*), \quad x>b^*,$$ with the boundary conditions $$V_{b^*}(0)=0, \quad {V_{b^*}}'(x)|_{x=b^*}=\alpha,$$ where $${\cal{A}} g(x)=\frac{1} {2}\sigma^2_{p} g''(x)-p g'(x)-\lambda g(x) +\lambda\int_{0}^{\infty}g(x+y)F(dy). \nonumber$$ Following the arguments of Laplace transform used in Yin, Wen and Zhao \[28\], one can show that the solution to (3.9)-(3.11) is given by $$V_{b^*}(x)=-\alpha\overline{Z}^{(\delta)}(b^*-x)+\alpha\frac{E[X_1]}{\delta},$$ and $$b^*=(\overline{Z}^{(\delta)})^{-1}\left(\frac{E[X_1]}{\delta}\right),$$ where $$Z^{(\delta)}(x)=1+\delta\int_0^x W^{(\delta)}(y)dy,\; \overline{Z}^{(\delta)}(x)=\int_0^x Z^{(\delta)}(y)dy, \ x\in \Bbb{R}.$$ Here, $W^{(\delta)}$ is the so-called $\delta$-scale function defined in the way that $W^{(\delta)}(x) = 0$ for all $x < 0$ and that its Laplace transform on $[0,\infty)$ is given by $$\int_0^{\infty}\text{e}^{-\theta x}W^{(\delta)}(x)dx=\frac{1}{\Psi(\theta)-\delta},\; \theta >\sup\{\theta\ge 0: \Psi(\theta)=\delta\}, \nonumber$$ where $$\Psi (\theta) =p\theta + \frac1 2\sigma^2_{p}\theta^2 +\lambda\int_{0}^{\infty}(e^{-\theta x}-1)F(dx).$$ For further details, the reader is referred to Yin and Wen \[26\]. $\Box$ [**Example 3.2.**]{} Let $\sigma_R=\sigma_p=0$. Assume that $X_i$ is exponentially distributed with parameter $\mu$. Then, by Theorem 3.1 and Lemma 3.2, it can be shown that $V_{b^*}(x)$ satisfies the following integro-differential equation $$(rx-p)V_{b^*}'(x)+\lambda\mu\int_0^{\infty}V_{b^*}(x+z)e^{-\mu z}dz=(\lambda+\delta) V_{b^*}(x), \quad 0<x<b^*,$$ and $$V_{b^*}(x)=\alpha(x-b^*)+ V_{b^*}(b^*), \quad x>b^*,$$ with the boundary conditions $$V_{b^*}(0)=0, \quad {V_{b^*}}'(x)|_{x=b^*}=\alpha.$$ From equation (3.12), we find that $$zg''(z)+\left(1-\frac{\lambda+\delta}{r}-z\right)g'(z)+\frac{\delta}{r}g(z)=0,$$ where $$g(z)= V_{b^*}(x),\; z=\mu\left(x-\frac{p}{r}\right).$$ Note that this is Kummer’s confluent hypergeometric equation with the solution given by $$g(z)=C_1 M\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r},z\right)+C_2 U\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r},z\right),$$ where $C_1$ and $C_2$ are constants, and $M(a,b,x)$ is the standard confluent hypergeometric function with $U(a,b,x)$ being its second form; see, for example, Abramowitz and Stugen \[1, pp. 504-505\]. Then, it follows that $$V_{b^*}(x)=C_1 M\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r}, \mu(x-\frac{p}{r})\right) +C_2 U\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r},\mu(x-\frac{p}{r})\right).$$ Using the boundary conditions (3.14) and the formulae $$M'(a,b,z)=\frac{a}{b}M(a+1,b+1,z), \quad U'(a,b,z)=-aU(a+1,b+1,z),$$ we obtain the coefficients $$C_1=\frac{\alpha U(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r}, -\frac{\mu p}{r})}{\Delta(b^*)},$$ and $$C_2=-\frac{\alpha M(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r}, -\frac{\mu p}{r})}{\Delta(b^*)},$$ where $$\begin{aligned} \Delta(b^*)&=&-\frac{\mu\delta}{r-\lambda-\delta}U\left(-\frac{\delta}{r},1-\frac{\lambda+\delta}{r}, -\frac{\mu p}{r}\right)M\left(1-\frac{\delta}{r}, 2-\frac{\lambda+\delta}{r},\mu(b^*-\frac{p}{r})\right)\\ &&+\frac{\mu\delta}{r}M\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r}, -\frac{\mu p}{r}\right)U\left(1-\frac{\delta}{r}, 2-\frac{\lambda+\delta}{r},\mu(b^*-\frac{p}{r})\right),\end{aligned}$$ and $b^*$ is the maximizer of term $1/\Delta(b)$ with respect to $b$, i.e., $$b^*=argmax\frac{1}{\Delta(b)}.$$ $\Box$ 0.2cm Optimal dividend problem with capital injections {#with} ================================================ In this section, we consider the optimal dividend problem with capital injections. The set of admissible strategies is given by $$\Xi_c=\{\xi_c=(L^{\xi_c}, G^{\xi_c}): (L^{\xi_c}, G^{\xi_c})\in \Xi\; {\rm and}\; U_t^{\xi_c}\ge 0\}.$$ The controlled surplus process $U^{\xi_c}_t$ satisfies $$\text{d}U^{\xi_c}_t=\text{d}P_t+ U^{\xi_c}_{t-}\text{d}R_t-\text{d}L_t^{\xi_c}+\text{d}G_t^{\xi_c}, \quad t\ge 0,\nonumber$$ and the value function is defined as $$V_{c}(x)=\sup_{ \xi_c\in\Xi_c}V(x; \xi_c) \equiv \sup_{ \xi_c\in\Xi_c}E_x\left(\alpha\int_{0-}^{\infty} e^{-\delta t}dL_t^{\xi_c}-\beta \int_{0-}^{\infty}\text{e}^{-\delta t}dG_t^{ \xi_c}\right), \quad x\ge 0.$$ Since the controlled surplus process always stays positive, the company will never go bankrupt. We shall identify the form of the value function $V_c$ and the optimal strategy $\xi_c^*$ such that $V_c(x)=V(x; \xi_c^*)$. [**4.1 HJB equation and verification lemma**]{} Applying the techniques used in Section 3, we get the HJB equation and the verification Lemma. $$\max\{{\cal L} w(x)-\delta w(x),\; \alpha-w'(x), w'(x)-\beta \}=0, \quad x\ge 0.\label{with-eq2}$$ (Verification Lemma) Let $w$ be a solution to (4.2). Then, $w(x)\ge V(x; \xi_c)$ for any admissible strategy $\xi_c\in\Xi_c$, and thus $w(x)\ge V_{c}(x).$ [**4.2 Construction of a candidate solution**]{} We now construct a concave $C^2$ solution $H$ to the HJB equation (\[with-eq2\]). Due to the effect of the discount factor, it is clear that the optimal strategy is the one that postpone capital injections as long as possible, i.e., we inject capital only when surplus become zero. Consider the barrier strategy with the upper barrier $B^*$ and the lower barrier $0$, and the strategy $\pi^*=(L^{\pi^*}, G^{\pi^*})$ where $(U_t^{\pi^*}, L_t^{\pi^*, x}, G^{\pi^*, x})$ is a solution to the following system $$\begin{aligned} &&\text{d}U^{\pi^*}_t=\text{d}P_t+ U^{\pi^*}_{t-}\text{d}R_t-\text{d}L_t^{\pi^*}+\text{d}G_t^{\pi^*}, \\ &&0\le U_t^{\pi^*}\le B^*, \ t\ge 0, \\ &&L^{\pi^*,x}_t=\max(x-B^*,0)+\int_{0-}^{t-}1(U_s^{\pi^*}=B^*)d L_s^{\pi^*},\ t>0, \\ &&G_t^{\pi^*,x}=\max\left(-\inf_{0\le s\le t}(P_s-L^{\pi^*}_s),0\right), \ t>0. \end{aligned}$$ For the problem of (4.3)-(4.6), if $H(x)$ solves $${\cal{L}} H(x)=\delta H(x), \quad 0< x <B^*,$$ with $H(x)=H(B^*) + \alpha(x-B^*)$ for $x>B^*$ and the boundary conditions $$\begin{aligned} &&H'(0) = \beta, \quad H'(B^*) = \alpha,\end{aligned}$$ where the infinitesimal generator $\cal{L}$ is given by (2.4), then $H(x)$ is given by $$H(x)=V(x; \pi^*)\equiv E_x\left(\alpha\int_{0-}^{\infty} e^{-\delta t}dL_t^{\pi^*,x}-\beta \int_{0-}^{\infty}\text{e}^{-\delta t}dG_t^{\pi^*,x}\right), \quad x\ge 0.$$ [**Proof.**]{}  For the strategy $\pi^*$, define $\Lambda=\{s: L^{\pi^*,x}_{s-}\neq L^{\pi^*,x}_s\}$. Let ${L}^{\pi^*,x,c}_t$ be the continuous part of ${L}^{\pi^*,x}_t$. Since the process is skip-free downward, $G^{\pi^*,x}_{t}$ is continuous. In addition, we see from (4.6) that $G^{\pi^*,x}_{t}\ge 0$ and that the support of the Stieltjes measure $dG^{\pi^*,x}_{t}$ is contained in the closure of the set $\{t: U_t^{\pi^*}=0\}$. Applying Ito’s formula for semimartingale to $e^{-\delta t}H(U^{\pi^*}_{t})$ gives $$\begin{aligned} E_x[e^{-\delta t}H({U}^{\pi^*}_{t-})] &=&H(x)+ E_x\int_0^{t} e^{-\delta s}({\cal L}-\delta)H({U}^{\pi^*}_{s})ds \nonumber\\ &&+ E_x\sum_{s\in \Lambda,s\le t} e^{-\delta s}\left\{H(U^{\pi^*}_{s}) -H(U^{\pi^*}_{s-})\right\} \nonumber \\ &&- E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s-})d{L}^{\pi^*,x,c}_s \nonumber\\ &&+E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s-})d{G}^{\pi^*,x}_s.\end{aligned}$$ Note that $({\cal L}-\delta)H({U}^{\pi^*}_{s})=0$, and that $$E_x\sum_{s\in \Lambda,s\le t} e^{-\delta s}\left\{H(U^{\pi^*}_{s}) -H(U^{\pi^*}_{s-})\right\}=\alpha \sum_{s\le t} e^{-\delta s}({L}^{\pi^*,x}_s-{L}^{\pi^*,x}_{s-}),$$ $$E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s-})d{L}^{\pi^*,x,c}_s=E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s})d{L}^{\pi^*,x,c}_s=\alpha E_x\int_{0-}^{t-} e^{-\delta s}d{L}^{\pi^*,x,c}_s,$$ $$E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s-})d{G}^{\pi^*,x}_s=E_x\int_{0-}^{t-} e^{-\delta s}H'({U}^{\pi^*}_{s})d{G}^{\pi^*,x}_s=\beta E_x\int_{0-}^{t-} e^{-\delta s}d{G}^{\pi^*,x}_s.$$ Then, it follows that $$E_x[e^{-\delta t}H({U}^{\pi^*}_{t-})] =H(x)-\alpha E_x\int_{0-}^{t-} e^{-\delta s}d{L}^{\pi^*,x}_s+\beta E_x\int_{0-}^{t-} e^{-\delta s}d{G}^{\pi^*,x}_s.$$ Since $\lim_{t\to\infty} E_x[e^{-\delta t}H({U}^{\pi^*}_{t-})]\le \lim_{t\to\infty} E_x[e^{-\delta t}H(B^*)]=0$, letting $t\to\infty$ in (4.9) and using the monotone convergence theorem yield $$H(x)=\alpha E_x\int_{0-}^{\infty} e^{-\delta s}d{L}^{\pi^*,x}_s-\beta E_x\int_{0-}^{\infty} e^{-\delta s}d{G}^{\pi^*,x}_s=V(x; \pi^*).$$ $\Box$ $V(x; \pi^*)$ is a concave increasing function on $(0, \infty)$. [**Proof.**]{}  Similar to the proof of Lemma 3.3, we use the arguments of Kulenko and Schmidli \[20\]. Let $x>0$, $y>0$, and $l\in (0,1)$. Consider the strategies $(L^{\pi^*,x},G^{\pi^*,x})$ and $(L^{\pi^*,y},G^{\pi^*,y})$ for the initial capitals $x$ and $y$. Define $L_t=l L^{\pi^*, x}_t+(1-l) L^{\pi^*, y}_t$ and $G_t=l G^{\pi^*,x}_t+(1-l) G^{\pi^*,y}_t$. Then, $L_t=L_t^{\pi^*, lx+(1-l)y}$. So, we have $$\begin{aligned} lx&+&(1-l)y +\int_0^t{\cal E}(R)_{s-}^{-1}\text{d}P_s-\rho\sigma_p\sigma_R \int_0^t{\cal E}(R)_{s-}^{-1}\text{d}s\\ &&-\int_0^t {\cal E}(R)_{s-}^{-1}(l dL^{\pi^*,x}_s+(1-l) dL^{\pi^*,y}_s)\\ &&+\int_0^t {\cal E}(R)_{s-}^{-1}(l dG^{\pi^*,x}_s+(1-l) dG^{\pi^*,y}_s)\\ &=& l\left\{x +\int_0^t{\cal E}(R)_{s-}^{-1}\text{d}P_s-\rho\sigma_p\sigma_R \int_0^t{\cal E}(R)_{s-}^{-1}\text{d}s\right. \\ &&\left.- \int_0^t {\cal E}(R)_{s-}^{-1}dL^{\pi^*,x}_s+{\cal E}(R)_t\int_0^t {\cal E}(R)_{s-}^{-1} dG^{\pi^*,x}_s\right\}\\ &&+ (1-l)\left\{y +\int_0^t{\cal E}(R)_{s-}^{-1}\text{d}P_s-\rho\sigma_p\sigma_R \int_0^t{\cal E}(R)_{s-}^{-1}\text{d}s\right.\\ &&- \left.\int_0^t {\cal E}(R)_{s-}^{-1}dL^{\pi^*,y}_s+{\cal E}(R)_t\int_0^t {\cal E}(R)_{s-}^{-1}dG^{\pi^*,y}_s\right\}\ge 0.\end{aligned}$$ This shows that the strategy $(L_t, G_t)$ is admissible and that $$G_t^{\pi^*,lx+(1-l)y}\le l G^{\pi^*,x}_t+(1-l) G^{\pi^*,y}_t.$$ It follows that $$\begin{aligned} V(lx+(1-l)y,\pi^*)&=&E\left(\alpha \int_{0-}^{\infty}e^{-\delta t}dL_t^{\pi^*,lx+(1-l)y}- \beta\int_{0-}^{\infty}e^{-\delta t}dG_t^{\pi^*,lx+(1-l)y}\right)\\ &&\ge l E \left( \alpha \int_{0-}^{\infty}e^{-\delta t}dL^{\pi^*,x}_t-\beta\int_{0-}^{\infty}e^{-\delta t}dG^{\pi^*,x}_t\right)\\ &&+(1-l) E \left(\alpha \int_{0-}^{\infty}e^{-\delta t}dL^{\pi^*,y}_t-\beta\int_{0-}^{\infty}e^{-\delta t}dG^{\pi^*,y}_t\right)\\ &=&l V(x,\pi^*)+(1-l)V(y,\pi^*),\end{aligned}$$ which implies the concavity of $V$. The proof of increasingness of $V(x; \pi^*)$ is routine. $\Box$ [**4.3 Verification of optimality**]{} Define the barrier level as $$B^*=\sup\{B\ge 0: H'(B-)=\alpha\}.$$ We conjecture that the barrier strategy $\pi^*$ is optimal. \[with-1\] The value function $H$ defined in (4.7) satisfies $$H(x)=V_c (x)=\sup_{\xi_c\in\Xi_c}V_{ \xi_c}(x),$$ and the joint strategy $\pi^*=(L^{\pi^*}, G^{\pi^*})$ is optimal, where $(L^{\pi^*}, G^{\pi^*})$ is given by (4.5) and (4.6). [**Proof.**]{}  Note that $({\cal L}-\delta)H(x)=0$ and $\alpha\le H'(x)\le \beta$ for $x\in [0,B^*)$ due to the concavity of $H$ on $[0,B^*)$. For $x\ge B^*$ and $H(x)=\alpha(x-B^*)+ H(B^*)$, we have $$\begin{aligned} ({\cal L}-\delta)H(x)&=&-p\alpha+\alpha\int_1^{\infty}y\Pi(dy)-\alpha (x-B^*)-\delta H(B^*)\\ &&<-p\alpha+\alpha\int_1^{\infty}y\Pi(dy)-\delta H(B^*)\\ &=&\lim_{x\to b^*+}({\cal L}-\delta)H(x)=\lim_{x\to B^*-}({\cal L}-\delta)H(x)=0.\end{aligned}$$ Due to the continuity of $H, H'$ and $H''$ at $x=B^*$. Thus, the function $H$ satisfies the HJB equation (4.2). By Lemma 4.1, we get $H(x)\ge V_{c}(x)$. On the other hand, $H(x)\le V_{c}(x)$. Thus, $H(x)=V_{c}(x)$. $\Box$ [**4.4 Two closed-form solutions**]{} We now present two examples in which closed-form solution can be derived. [**Example 4.1.**]{}  Assume that $r=0$ and $\sigma_R=0$. Then, $H(x)$ satisfies the following integro-differential equation $${\cal A}H(x)=\delta H(x), \; 0<x<B^*,$$ and $$H(x)=\alpha(x-B^*)+ H(B^*),\ x>B^*,$$ with the boundary conditions $$H'(0)=\beta, \quad H'(B^*)=\alpha,$$ where $${\cal{A}} g(x)=\frac{1} {2}\sigma^2_{p} g''(x)-p g'(x)-\lambda g(x) +\lambda\int_{0}^{\infty}g(x+y)F(dy). \nonumber$$ Again, using the arguments of Laplace transform, one can show that the solution to (4.10) and (4.11) is given by $$H(x)=-\alpha\overline{Z}^{(\delta)}(B^*-x)+\alpha\frac{E[X_1]}{\delta},$$ and $$B^*=({Z}^{(\delta)})^{-1}\left(\frac{\beta}{\alpha}\right),$$ where $Z^{(\delta)}(x)$ and $\overline{Z}^{(\delta)}(x)$ are defined in Example 3.1. In the case of $\alpha=1$, these formulae were obtained in Bayraktar, Kyprianou and Yamazaki \[11\] by using the fluctuation theory of spectrally positive Lévy processes. [**Example 4.2.**]{}  Let $\sigma_R=\sigma_p=0$. Assume that $X_i$ is exponentially distributed with parameter $\mu$. Then, by Theorem 4.1 and Lemma 4.2, $H(x)$ satisfies the following integro-differential equation $$(rx-p)H'(x)+\lambda\mu\int_0^{\infty}H(x+z)e^{-\mu z}dz=(\lambda+\delta) H(x), \quad 0<x<B^*,$$ and $$H(x)=\alpha(x-B^*)+ H(B^*), \quad x>B^*,$$ with the boundary conditions $$H'(0)=\beta,\ \ \ H'(B^*)=\alpha.$$ Repeating the steps in Example 3.2, we obtain $$H(x)=C_3 M\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r}, \mu(x-\frac{p}{r})\right) +C_4 U\left(-\frac{\delta}{r}, 1-\frac{\lambda+\delta}{r},\mu(x-\frac{p}{r})\right).$$ The constants $C_3$ and $C_4$ can be determined from the boundary conditions (4.15). Using the formulae $$M'(a,b,z)=\frac{a}{b}M(a+1,b+1,z), \quad U'(a,b,z)=-aU(a+1,b+1,z),$$ we get $$C_3=\frac{\beta \Delta_4-\alpha \Delta_2}{\Delta_1\Delta_4-\Delta_2\Delta_3},$$ and $$C_4=\frac{\alpha \Delta_1-\beta \Delta_3}{\Delta_1\Delta_4-\Delta_2\Delta_3},$$ where $$\begin{aligned} \Delta_1&=&-\frac{\mu\delta}{r-\lambda-\delta}M\left(1-\frac{\delta}{r},2-\frac{\lambda+\delta}{r},-\frac{\mu p}{r}\right),\\ \Delta_2&=&\frac{\mu\delta}{r}U\left(1-\frac{\delta}{r}, 2-\frac{\lambda+\delta}{r},-\frac{\mu p}{r}\right),\\ \Delta_3&=&-\frac{\mu\delta}{r-\lambda-\delta}M\left(1-\frac{\delta}{r},2-\frac{\lambda+\delta}{r},\mu(B^*-\frac{p}{r})\right),\\ \Delta_3&=&\frac{\mu\delta}{r}U\left(1-\frac{\delta}{r}, 2-\frac{\lambda+\delta}{r},\mu(B^*-\frac{p}{r})\right).\end{aligned}$$ Here, $B^*$ is the unique solution to the following equation with respect to $b$: $$-\frac{\mu\delta}{r-\lambda-\delta} C_3 M\left(1-\frac{\delta}{r}, 2-\frac{\lambda+\delta}{r},\mu(b-\frac{p}{r})\right)+\frac{\mu\delta}{r}U\left(1-\frac{\delta}{r},2-\frac{\lambda+\delta}{r},\mu(b-\frac{p}{r})\right)=\alpha.$$ 0.2cm Solution to the problem without constraints {#general} =========================================== We now consider the control problem (2.6) without any restrictions on capital injections. In this case, ruin can occur and the time of ruin for a control strategy $\xi$ is defined as $$\tau_{\xi}=\inf\{t: U^{\xi}_t=0\},$$ because of the diffusion and the skip-free downward surplus process. Then, it follows from (3.1), (4.1) and (2.5) that for all $x\ge 0$, $V_{\xi}(x)\ge \max\{V_d(x), V_c(x)\}$. We shall determine $V_*$ and the optimal strategy $\xi^*$ such that $V_*(x)=V(x; \xi^*)$. [**5.1 Verification lemma**]{} For the control problem without any restrictions on capital injections, we get the following associated HJB equation: $$\max\{{\cal L} v(x)-\delta v(x), \quad \alpha-v'(x), v'(x)-\beta \}=0, \; x\ge 0,\label{general-eq1}$$ with the boundary condition $$\max\{-v(0), v'(0)-\beta\}=0.$$ \[general-1\] (Verification Lemma) If $v$ satisfies the HJB equation (5.1) with the boundary condition (5.2), then $v(x)\ge V_{\xi}(x)$ for any admissible policy $\xi$. [**Proof.**]{}  For any admissible strategy $\xi\in\Xi$, put $\Lambda=\{s: L^{\xi}_{s-}\neq L^{\xi}_s\}$. Applying Ito’s formula for semimartingale to $e^{-\delta t}v(U^{\xi}_{t})$ gives $$\begin{aligned} E_x[e^{-\delta (t\wedge \tau_{\xi})}v({U}^{\xi}_{t\wedge \tau_{\xi}-})] &=&v(x)+ E_x\int_0^{t\wedge \tau_{\xi}-} e^{-\delta s}({\cal L}-\delta)v({U}^{\xi}_{s-})ds \nonumber\\ &&+ E_x\sum_{s\in \Lambda,s\le t\wedge \tau_{\xi}-}e^{-\delta s}\left\{v({U}^{\xi}_{s})-v({U}^{\xi}_{s-})\right\} \nonumber \\ &&- E_x\int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}v'({U}^{\xi}_{s-})d{L}^{\xi,c}_s\nonumber \\ &&+E_x\int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}v'({U}^{\xi}_{s-})d{G}^{\xi}_s,\end{aligned}$$ where ${L}^{\xi,c}_s$ is the continuous part of ${L}^{\xi}_s$. We see from (5.1) that $({\cal L}-\delta)v({U}^{\xi_d}_{s-})\le 0$ and $\alpha\le v'(x)\le \beta$. Thus, $$E_x\int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}v'({U}^{\xi}_{s-})d{G}^{\xi}_s\le \beta E_x\int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}d{G}^{\xi}_s,$$ and for $s\in \Lambda,s\le t\wedge \tau_{\xi}$, $$v({U}^{\xi}_{s})-v({U}^{\xi}_{s-})\le -\alpha ({L}^{\xi}_s-{L}^{\xi}_{s-}).$$ It follows from (5.3) and (5.5) that $$E_x[e^{-\delta (t\wedge \tau_{\xi})}v({U}^{\xi}_{t\wedge \tau_{\xi}-})]\le v(x)-\alpha E_x \int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}d{L}^{\xi}_s +\beta E_x \int_{0-}^{t\wedge \tau_{\xi}-} e^{-\delta s}d{G}^{\xi}_s.$$ Finally, by letting $t\to\infty$ in (5.6) and noting that (by Fatou’s lemma) $$\liminf_{t\to\infty}E_x[e^{-\delta (t\wedge \tau_{\xi})}v({U}^{\xi}_{t\wedge \tau_{\xi}-})]\ge E_x[\liminf_{t\to\infty} e^{-\delta (t\wedge \tau_{\xi})}v({U}^{\xi}_{t\wedge \tau_{\xi}})]\ge v(0) E_x[e^{-\delta \tau_{\xi})}]\ge 0,$$ we prove the lemma. $\Box$ [**5.2 Construction of a candidate solution**]{} For any $x\ge 0$, we set our candidate strategy to be $$\xi^*=\left\{ \begin{array}{ll} \xi_{d}^{*},& {\rm if}\; V'_{b^*}(0)\le \beta,\\ \xi_{c}^{*}, & {\rm if}\; H(0)\ge 0, \end{array} \right.\label{general-eq2}$$ and our candidate solution to be $$V_{\xi^*}(x)=\left\{ \begin{array}{ll} V_d(x),& {\rm if}\; V'_{b^*}(0)\le \beta,\\ V_c(x), & {\rm if}\;H(0)\ge 0, \end{array} \right.\label{general-eq3}$$ where $V_d$ and $V_c$ are given by (3.1) and (4.1), respectively, and $V_{b^*}$ and $H$ are given by (3.6) and (4.7), respectively. [**5.3 Verification of optimality**]{} \[general-1\] The value function $V_{\xi^*}$ defined in (5.8) satisfies $$V_{\xi^*}(x)=V_*(x)=\sup_{\xi\in\Xi}V(x;\xi),$$ and the joint strategy $\xi^{*}$ defined in (5.7) is optimal. [**Proof.**]{} If $V_{b^*}'(0)\le \beta$, then $V_{b^*}$ satisfies the equation (5.1) with the condition (5.2). Hence, $V_{b^*}(x)\ge V_*(x)$. On the other hand, $V_{b^*}(x)=V(x; \xi_d^*)\le V_d(x)$. It follows that $V_{\xi^*}(x)=V_{b^*}(x)=V_d(x)$. The optimality of $\xi_d^*$ is verified by Theorem 3.1. If $ H(0)\ge 0$, then $H$ satisfies the HJB equation (4.1), so that $H(x)\le V_c(x)$. Since $H$ also satisfies the equation (5.1) with the condition (5.2), $H(x)\ge V(x; \xi_c^*)\ge V_c(x)$. Hence, we have $V_{\xi^*}(x)=V(x; \xi_c^*)=V_c(x)$. The optimality of $\xi_c^*$ is verified by Theorem 4.1. $\Box$ 0.3cm [**Acknowledgements**]{} The authors would like to thank two anonymous referees and the editor for their helpful comments on the previous version of the paper. The research of Chuancun Yin was supported by the National Natural Science Foundation of China (No. 11171179) and the Research Fund for the Doctoral Program of Higher Education of China (No. 20133705110002). The research of Kam C. Yuen was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. HKU 7057/13P). [99]{} M. Abramowitz and A. Stegun, [*Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables,*]{} Dover Publications, New York, 1965. S. Asmussen, F. Avram and M. R. Pistorius, Russian and American put options under exponential phase-type Lévy models, [*Stochastic Processes and their Applications*]{}, [**109**]{} (2004), 79-111. B. Avanzi, Strategies for dividend distribution: A review, [*North American Actuarial Journal*]{}, [**13**]{} (2009), 217-251. B. Avanzi, E. C. K. Cheung, B. Wong and J.-K. Woo, On a periodic dividend barrier strategy in the dual model with continuous monitoring of solvency, [*Insurance: Mathematics and Economics,*]{} [**52**]{} (2013), 98-113. B. Avanzi and H. U. Gerber, Optimal dividends in the dual model with diffusion, [*ASTIN Bulletin*]{}, [**38**]{} (2008), 653-667. B. Avanzi, H. U. Gerber and E. S. W. Shiu, Optimal dividends in the dual model, [*Insurance: Mathematics and Economics*]{}, [**41**]{} (2007), 111-123. B. Avanzi, J. Shen and B. Wong, Optimal dividends and capital injections in the dual model with diffusion, [*ASTIN Bulletin*]{}, [**41**]{} (2011), 611-644. B. Avanzi, V. Tu and B. Wong, On optimal periodic dividend strategies in the dual model with diffusion, [*Insurance: Mathematics and Economics*]{}, [**55**]{} (2014), 210-224. P. Azcue and N. Muler, Optimal reinsurance and dividend distribution policies in the Cramér-Lundberg model, [*Mathematical Finance*]{}, [**15**]{} (2005), 261-308. E. Bayraktar and M. Egami, Optimizing venture capital investments in a jump diffusion model, [*Mathematical Methods of Operations Research*]{}, [**67**]{} (2008), 21-42. E. Bayraktar, A. E. Kyprianou and K. Yamazaki, On optimal dividends in the dual model, [*ASTIN Bulletin*]{}, [**43**]{} (2013), 359-372. E. Bayraktar, A. E. Kyprianou and K. Yamazaki, Optimal dividends in the dual model under transaction costs, [*Insurance: Mathematics and Economics*]{}, [**54**]{} (2014), 133-143. E. C. K. Cheung and S. Drekic, Dividend moments in the dual model: Exact and approximate approaches, [*ASTIN Bulletin*]{}, [**38**]{} (2008), 149-159. H. Dai, Z. Liu and N. Luan, Optimal dividend strategies in a dual model with capital injections, [*Mathematical Methods of Operations Research*]{}, [**72**]{} (2010), 129-143. H. Dai, Z. Liu and N. Luan, Optimal financing and dividend control in the dual model, [*Mathematical and Computer Modelling*]{}, [**53**]{} (2011), 1921-1928. B. De Finetti, Su un’impostazion alternativa dell teoria collecttiva del rischio, [*Transactions of the XVth International Congress of Actuaries*]{}, [**2**]{} (1957), 433-443. W. H. Fleming and H. M. Soner, [*Controlled Markov Processes and Viscosity Solutions*]{}, Applications of Mathematics, Springer-Verlag, New York, 1993. L. He and Z. Liang, Optimal financing and dividend control of the insurance company with fixed and proportional transaction costs, [*Insurance: Mathematics and Economics*]{}, [**44**]{} (2009), 88-94. S. Jaschke, A note on the inhomogeneous linear stochastic differential equation, [*Insurance: Mathematics and Economics*]{}, [**32**]{} (2003), 461-464. N. Kulenko and H. Schmidli, Optimal dividend strategies in a Cramér-Lundberg model with capital injections, [*Insurance: Mathematics and Economics*]{}, [**43**]{} (2008), 270-278. K. Miyasawa, An economic survival game, [*Journal of the Operations Research Society of Japan*]{}, [**4**]{} (1962), 95-113. H. Schmidli, [*Stochastic Control in Insurance*]{}, Springer, New York, 2008. D. J. Yao, H. L. Yang and R. M. Wang, Optimal financing and dividend strategies in a dual model with proportional costs, [*Journal of Industrial and Management Optimization*]{}, [**6**]{} (2010), 761-777. D. J. Yao, H. L. Yang and R. W. Wang, Optimal dividend and capital injection problem in the dual model with proportional and fixed transaction costs, [*European Journal of Operational Research*]{}, [**211**]{} (2011), 568-576. D. J. Yao, R. W. Wang and L. Xu, Optimal dividend and capital injection strategy with fixed costs and restricted dividend rate for a dual model, [*Journal of Industrial and Management Optimization*]{}, [**10**]{} (2014), 1235-1259. C. C. Yin and Y. Z. Wen, Optimal dividends problem with a terminal value for spectrally positive Lévy processes, [*Insurance: Mathematics and Economics*]{}, [**53**]{} (2013), 769-773. C. C. Yin and Y. Z. Wen, An extension of Paulsen-Gjessing’s risk model with stochastic return on investments, [*Insurance: Mathematics and Economics*]{}, [**52**]{} (2013), 469-472. C. C. Yin, Y. Z. Wen and Y. X. Zhao, On the optimal dividend problem for a spectrally positive Lévy process, [*ASTIN Bulletin*]{}, [**44**]{} (2014), 635-651. Z. M. Zhang, On a risk model with randomized dividend-decision times, [*Journal of Industrial and Management Optimization*]{}, [**10**]{} (2014), 1041-1058.

--- abstract: 'We report on measurements of the neutron spin asymmetries $A_{1,2}^n$ and polarized structure functions $g_{1,2}^n$ at three kinematics in the deep inelastic region, with $x=0.33$, $0.47$ and $0.60$ and $Q^2=2.7$, $3.5$ and $4.8$ (GeV/c)$^2$, respectively. These measurements were performed using a $5.7$ GeV longitudinally-polarized electron beam and a polarized $^3$He target. The results for $A_1^n$ and $g_1^n$ at $x=0.33$ are consistent with previous world data and, at the two higher $x$ points, have improved the precision of the world data by about an order of magnitude. The new $A_1^n$ data show a zero crossing around $x=0.47$ and the value at $x=0.60$ is significantly positive. These results agree with a next-to-leading order QCD analysis of previous world data. The trend of data at high $x$ agrees with constituent quark model predictions but disagrees with that from leading-order perturbative QCD (pQCD) assuming hadron helicity conservation. Results for $A_2^n$ and $g_2^n$ have a precision comparable to the best world data in this kinematic region. Combined with previous world data, the moment $d_2^n$ was evaluated and the new result has improved the precision of this quantity by about a factor of two. When combined with the world proton data, polarized quark distribution functions were extracted from the new $g_1^n/F_1^n$ values based on the quark parton model. While results for $\Delta u/u$ agree well with predictions from various models, results for $\Delta d/d$ disagree with the leading-order pQCD prediction when hadron helicity conservation is imposed.' address: ' 2 pt 0.3 cm [The Jefferson Lab Hall A Collaboration]{} 0.1 cm ' author: - 'X. Zheng,$^{{13}}$ K. Aniol,$^{{3}}$ D. S. Armstrong,$^{{22}}$ T. D. Averett,$^{{{8},{22}}}$ W. Bertozzi,$^{{13}}$ S. Binet,$^{{21}}$ E. Burtin,$^{{17}}$ E. Busato,$^{{16}}$ C. Butuceanu,$^{{22}}$ J. Calarco,$^{{14}}$ A. Camsonne,$^{{1}}$ G. D. Cates,$^{{21}}$ Z. Chai,$^{{13}}$ J.-P. Chen,$^{{8}}$ Seonho Choi,$^{{20}}$ E. Chudakov,$^{{8}}$ F. Cusanno,$^{{7}}$ R. De Leo,$^{{7}}$ A. Deur,$^{{21}}$ S. Dieterich,$^{{16}}$ D. Dutta,$^{{13}}$ J. M. Finn,$^{{22}}$ S. Frullani,$^{{7}}$ H. Gao,$^{{13}}$ J. Gao,$^{{2}}$ F. Garibaldi,$^{{7}}$ S. Gilad,$^{{13}}$ R. Gilman,$^{{{8},{16}}}$ J. Gomez,$^{{8}}$ J.-O. Hansen,$^{{8}}$ D. W. Higinbotham,$^{{13}}$ W. Hinton,$^{{15}}$ T. Horn,$^{{11}}$ C.W. de Jager,$^{{8}}$ X. Jiang,$^{{16}}$ L. Kaufman,$^{{12}}$ J. Kelly,$^{{11}}$ W. Korsch,$^{{10}}$ K. Kramer,$^{{22}}$ J. LeRose,$^{{8}}$ D. Lhuillier,$^{{17}}$ N. Liyanage,$^{{8}}$ D.J. Margaziotis,$^{{3}}$ F. Marie,$^{{17}}$ P. Markowitz,$^{{4}}$ K. McCormick,$^{{9}}$ Z.-E. Meziani,$^{{20}}$ R. Michaels,$^{{8}}$ B. Moffit,$^{{22}}$ S. Nanda,$^{{8}}$ D. Neyret,$^{{17}}$ S. K. Phillips,$^{{22}}$ A. Powell,$^{{22}}$ T. Pussieux,$^{{17}}$ B. Reitz,$^{{8}}$ J. Roche,$^{{22}}$ R. Roché,$^{{5}}$ M. Roedelbronn,$^{{6}}$ G. Ron,$^{{19}}$ M. Rvachev,$^{{13}}$ A. Saha,$^{{8}}$ N. Savvinov,$^{{11}}$ J. Singh,$^{{21}}$ S. Širca,$^{{13}}$ K. Slifer,$^{{20}}$ P. Solvignon,$^{{20}}$ P. Souder,$^{{18}}$ D.J. Steiner,$^{{22}}$ S. Strauch,$^{{16}}$ V. Sulkosky,$^{{22}}$ A. Tobias,$^{{21}}$ G. Urciuoli,$^{{7}}$ A. Vacheret,$^{{12}}$ B. Wojtsekhowski,$^{{8}}$ H. Xiang,$^{{13}}$ Y. Xiao,$^{{13}}$ F. Xiong,$^{{13}}$ B. Zhang,$^{{13}}$ L. Zhu,$^{{13}}$ X. Zhu,$^{{22}}$ P.A. Żo[ł]{}nierczuk,$^{{10}}$' title: 'Precision Measurement of the Neutron Spin Asymmetries and Spin-dependent Structure Functions in the Valence Quark Region' --- INTRODUCTION {#ch1:main} ============ Interest in the spin structure of the nucleon became prominent in the 1980’s when experiments at CERN [@exp:cern-emc] and SLAC [@exp:e080e130] on the integral of the proton polarized structure function $g_1^p$ showed that the total spin carried by quarks was very small, $\approx (12\pm 17)\%$ [@exp:cern-emc]. This was in contrast to the simple relativistic valence quark model prediction [@theory:rCQM] in which the spin of the valence quarks carries approximately $75\%$ of the proton spin and the remaining $25\%$ comes from their orbital angular momentum. Because the quark model is very successful in describing static properties of hadrons, the fact that the quark spins account for only a small part of the nucleon spin was a big surprise and generated very productive experimental and theoretical activities to the present. Current understanding [@theory:spin-sr] of the nucleon spin is that the total spin is distributed among valence quarks, $q\bar{q}$ sea quarks, their orbital angular momenta, and gluons. This is called the nucleon spin sum rule: $$\begin{aligned} S_z^N &=& S_z^q+L_z^q+J_z^g = \frac{1}{2}~, \label{equ:spin}\end{aligned}$$ where $S_z^N$ is the nucleon spin, $S_z^q$ and $L_z^q$ represent respectively the quark spin and orbital angular momentum (OAM), and $J_z^g$ is the total angular momentum of the gluons. Only about $(20-30)\%$ of the nucleon spin is carried by the spin of the quarks. To further study the nucleon spin, one thus needs to know more precisely how it decomposes into the three components and to measure their dependence on $x$. Here $x$ is the Bjorken scaling variable, which in the quark-parton model [@theory:partonmodel] can be interpreted as the fraction of the nucleon momentum carried by the quark. For a fixed target experiment one has $x={Q^2}/({2M\nu})$, with $M$ the nucleon mass, $Q^2$ the four momentum transfer squared and $\nu$ the energy transfer from the incident electron to the target. However, due to experimental limitations, precision data have been collected so far only in the low and moderate $x$ regions. In these regions, one is sensitive to contributions from a large amount of $q\bar q$ sea and gluons and the nucleon is hard to model. Moreover, at large distances corresponding to the size of a nucleon, the theory of the strong interaction – Quantum Chromodynamics (QCD) – is highly non-perturbative, which makes the investigation of the roles of quark orbital angular momentum (OAM) and gluons in the nucleon spin structure difficult. Our focus here is the first precise neutron spin structure data in the large $x$ region $x\gtorder 0.4$. For these kinematics, the valence quarks dominate and the ratios of structure functions can be estimated based on our knowledge of the interactions between quarks. More specifically, the virtual photon asymmetry $A_1$, defined as $$\begin{aligned} A_1(x,Q^2) &\equiv& \frac{\sigma_{{1}/{2}}-\sigma_{{3}/{2}}} {\sigma_{{1}/{2}}+\sigma_{{3}/{2}}} \nonumber \end{aligned}$$ (the definitions of $\sigma_{1/2,3/2}$ are given in Appendix \[app:formalism\]), which at large $Q^2$ is approximately the ratio of the polarized and the unpolarized structure functions $g_1/F_1$, is expected to approach unity as $x\to 1$ in perturbative QCD (pQCD). This is a dramatic prediction, not only because this is the only kinematic region where one can give an absolute prediction for the structure functions based on pQCD, but also because all previous data on the neutron asymmetry $A_1^n$ in the region $x\gtorder 0.4$ have large uncertainties and are consistent with $A_1^n\leqslant 0$. Furthermore, because both $q\bar q$ sea and gluon contributions are small in this region, it is a relatively clean region to test the valence quark model and to study the role of valence quarks and their OAM contribution to the nucleon spin.\ Deep inelastic scattering (DIS) has served as one of the major experimental tools to study the quark and gluon structure of the nucleon. The formalism of unpolarized and polarized DIS is summarized in Appendix \[app:formalism\]. Within the quark parton model (QPM), the nucleon is viewed as a collection of non-interacting, point-like constituents, one of which carries a fraction $x$ of the nucleon’s longitudinal momentum and absorbs the virtual photon [@theory:partonmodel]. The nucleon cross section is then the incoherent sum of the cross sections for elastic scattering from individual charged point-like partons. Therefore the unpolarized and the polarized structure functions $F_1$ and $g_1$ can be related to the spin-averaged and spin-dependent quark distributions as  $$\begin{aligned} F_1(x,Q^2) =&&\hspace*{-0.4cm} \frac{1}{2}\sum_i e_i^2 q_i(x,Q^2) \label{equ:F1qpm}\end{aligned}$$ and $$\begin{aligned} g_1(x,Q^2) =&&\hspace*{-0.4cm} \frac{1}{2}\sum_i e_i^2 \Delta q_i(x,Q^2)~, \label{equ:g1qpm}\end{aligned}$$ where $q_i(x,Q^2)=q_i^{\uparrow}(x,Q^2)+q_i^{\downarrow}(x,Q^2)$ is the unpolarized parton distribution function (PDF) of the $i^{th}$ quark, defined as the probability that the $i^{th}$ quark inside a nucleon carries a fraction $x$ of the nucleon’s momentum, when probed with a resolution determined by $Q^2$. The polarized PDF is defined as $\Delta q_i(x,Q^2)=q_i^{\uparrow}(x,Q^2)-q_i^{\downarrow}(x,Q^2)$, where $q_i^\uparrow(x,Q^2)$ ($q_i^\downarrow(x,Q^2)$) is the probability to find the spin of the $i^{th}$ quark aligned parallel (anti-parallel) to the nucleon spin. The polarized structure function $g_2(x,Q^2)$ does not have a simple interpretation within the QPM . However, it can be separated into leading twist and higher twist terms using the operator expansion method [@theory:ope]: $$\begin{aligned} g_2(x,Q^2) &=& g_2^{WW}(x,Q^2) +\bar g_2(x,Q^2)~.\end{aligned}$$ Here $g_2^{WW}(x,Q^2)$ is the leading twist (twist-2) contribution and can be calculated using the twist-2 component of $g_1(x,Q^2)$ and the Wandzura-Wilczek relation [@theory:g2ww] as $$\begin{aligned} g_2^{WW}(x,Q^2) &=& -g_1(x,Q^2) + \int_x^1 \frac{g_1(y,Q^2)}{y}\mathrm{d}y ~. \label{equ:g2ww} \end{aligned}$$ The higher-twist contribution to $g_2$ is given by $\bar g_2$. When neglecting quark mass effects, the higher-twist term represents interactions beyond the QPM, [*e.g.*]{}, quark-gluon and quark-quark correlations [@theory:ope-g2]. The moment of $\bar g_2$ can be related to the matrix element $d_2$ [@theory:d2def]: $$\begin{aligned} d_2 &=& \int_{0}^{1} {\mathrm}{d}x ~x^2 \Big[3g_2(x,Q^2)+2g_1(x,Q^2)\Big] \nonumber\\ &=& 3\int_{0}^{1} {\mathrm}{d}x ~x^2 \bar g_2(x,Q^2)~. \label{equ:d2def}\end{aligned}$$ Hence $d_2$ measures the deviations of $g_2$ from $g_2^{WW}$. The value of $d_2$ can be obtained from measurements of $g_1$ and $g_2$ and can be compared with predictions from Lattice QCD [@theory:d2lattice], bag models [@theory:d2bag], QCD sum rules [@theory:d2QCDSR] and chiral soliton models [@theory:d2chi]. In this paper we first describe available predictions for $A_1^n$ at large $x$. The experimental apparatus and the data analysis procedure will be described in Section \[exp:main\], \[targ:main\] and \[ch5:main\]. In Section \[result:main\] we present results for the asymmetries and polarized structure functions for both $^3$He and the neutron, a new experimental fit for $g_1^n/F_1^n$ and a result for the matrix element $d_2^n$. Combined with the world proton and deuteron data, polarized quark distribution functions were extracted from our $g_1^n/F_1^n$ results. We conclude the paper by summarizing the results for $A_1^n$ and $\Delta d/d$ and speculating on the importance of the role of quark OAM on the nucleon spin in the kinematic region explored. Some of the results presented here were published previously [@A1nPRL]; the present publication gives full details on the experiment and all of the neutron spin structure results for completeness. Predictions for $A_1^n$ at large $x$ {#ch2:main} ==================================== From Section \[ch2:su6\] to \[ch2:other\] we present predictions of $A_1^n$ at large $x$. Data on $A_1^n$ from previous experiments did not have the precision to distinguish among different predictions, as will be shown in Section \[ch2:exp\]. SU(6) Symmetric Non-Relativistic Constituent Quark Model {#ch2:su6} -------------------------------------------------------- In the simplest non-relativistic constituent quark model (CQM) [@theory:cqm_org], the nucleon is made of three constituent quarks and the nucleon spin is fully carried by the quark spin. Assuming SU(6) symmetry, the wavefunction of a neutron polarized in the $+z$ direction then has the form [@theory:su6close]: $$\begin{aligned} \label{equ:su6wvfunc} {\left| n{\uparrow}\right\rangle} = \frac{1}{\sqrt{2}}{\left| d^{\uparrow}(du)_{000} \right\rangle} +\frac{1}{\sqrt{18}}{\left| d^{\uparrow}(du)_{110} \right\rangle}~~~~~~~~~~~~~~~~~~&& \\ -\frac{1}{3} {\left| d^{\downarrow}(du)_{111} \right\rangle} -\frac{1}{3}{\left| u^{\uparrow}(dd)_{110} \right\rangle} +\frac{\sqrt{2}}{3}{\left| u^{\downarrow}(dd)_{111} \right\rangle},&&\nonumber\end{aligned}$$ where the three subscripts are the total isospin, total spin $S$ and the spin projection $S_z$ along the $+z$ direction for the ‘diquark’ state. For the case of a proton one needs to exchange the $u$ and $d$ quarks in Eq. (\[equ:su6wvfunc\]). In the limit where SU(6) symmetry is exact, both diquark spin states with $S=1$ and $S=0$ contribute equally to the observables of interest, leading to the predictions $$\begin{aligned} && A_1^p={5}/{9}~~{\mathrm}{and}~A_1^n=0 ~; \\ &&\Delta u/u \to 2/3~~{\mathrm}{and}~ \Delta d/d\to -1/3.\end{aligned}$$ We define $u(x)\equiv u^p(x)$, $d(x)\equiv d^p(x)$ and $s(x)\equiv s^p(x)$ as parton distribution functions (PDF) for the proton. For a neutron one has $u^n(x)=d^p(x)=d(x)$, $d^n(x)=u^p(x)=u(x)$ based on isospin symmetry. The strange quark distribution for the neutron is assumed to be the same as that of the proton, $s^n(x)=s^p(x)=s(x)$. In the following, all PDF’s are for the proton, unless specified by a superscript ‘n’. In the case of DIS, exact SU(6) symmetry implies the same shape for the valence quark distributions, [*i.e.*]{} $u(x)=2d(x)$. Using Eq. (\[equ:F1qpm\]) and (\[equ:disF1F2\]), and assuming that $R(x,Q^2)$ is the same for the neutron and the proton, one can write the ratio of neutron and proton $F_2$ structure functions as $$\begin{aligned} R^{np}\equiv\frac{F_2^n}{F_2^p} = \frac{u(x)+4d(x)}{4u(x)+d(x)} ~.\end{aligned}$$ Applying $u(x)=2d(x)$ gives $$\begin{aligned} R^{np} = {2}/{3} ~.\end{aligned}$$ However, data on the $R^{np}$ ratio from SLAC [@data:slacRnp], CERN [@data:bcdmsRnp; @data:emcRnp; @data:nmcRnp] and Fermilab [@data:e665Rnp] disagree with this SU(6) prediction. The data show that $R^{np}(x)$ is a straight line starting with $R^{np}\vert_{x\to 0}\approx 1$ and dropping to below $1/2$ as $x\to 1$. In addition, $A_1^p(x)$ is small at low $x$ [@data:a1pg1p-emc; @data:a1pg1p-smc; @data:a1pa1n-e143]. The fact that $R^{np}\vert_{x\to 0}\approx 1$ may be explained by the presence of a dominant amount of sea quarks in the low $x$ region and the fact that $A_1^{p}\vert_{x\to 0}\approx 0$ could be because these sea quarks are not highly polarized. At large $x$, however, there are few sea quarks and the deviation from SU(6) prediction indicates a problem with the wavefunction described by Eq. (\[equ:su6wvfunc\]). In fact, SU(6) symmetry is known to be broken [@theory:su6breaking] and the details of possible SU(6)-breaking mechanisms is an important open issue in hadronic physics. SU(6) Breaking and Hyperfine Perturbed Relativistic CQM {#ch2:cqm} ------------------------------------------------------- A possible explanation for the SU(6) symmetry breaking is the one-gluon exchange interaction which dominates the quark-quark interaction at short-distances. This interaction was used to explain the behavior of $R^{np}$ near $x\to 1$ and the $\approx 300$-MeV mass shift between the nucleon and the $\Delta(1232)$ [@theory:su6breaking]. Later this was described by an interaction term proportional to $\vec{S}_i\cdot\vec{S}_j~\delta^3(\vec{r}_{ij})$, with $\vec{S_i}$ the spin of the $i^{th}$ quark, hence is also called the hyperfine interaction, or chromomagnetic interaction among the quarks [@theory:hyperfine-spin]. The effect of this perturbation on the wavefunction is to lower the energy of the $S=0$ diquark state, causing the first term of Eq. (\[equ:su6wvfunc\]), ${\left| d{\uparrow}(ud)_{000} \right\rangle}^{\mathrm}{n}$, to become more stable and to dominate the high energy tail of the quark momentum distribution that is probed as $x\to 1$. Since the struck quark in this term has its spin parallel to that of the nucleon, the dominance of this term as $x\to 1$ implies $(\Delta d/d)^{\mathrm}{n}\to 1$ and $(\Delta u/u)^{\mathrm}{n}\to -1/3$ for the neutron, while for the proton one has $$\begin{aligned} &&\Delta u/u \to 1~{\mathrm}{and}~ \Delta d/d\to -1/3 \mathrm{~~as~} x\to 1 ~.\end{aligned}$$ One also obtains $$\begin{aligned} && R^{np}\to {1}/{4}~~{\mathrm}{as}~ x\to 1 ~,\end{aligned}$$ which could explain the deviation of $R^{np}(x)$ data from the SU(6) prediction. Based on the same mechanism, one can make the following predictions: $$\begin{aligned} A_1^{p}\to 1~{\mathrm}{and}~ A_1^{n}\to 1 ~~{\mathrm}{as}~ x\to 1~.\end{aligned}$$ The hyperfine interaction is often used to break SU(6) symmetry in the relativistic CQM (RCQM). In this model, the constituent quarks have non-zero OAM which carries $\approx 25\%$ of the nucleon spin [@theory:rCQM]. The use of RCQM to predict the large $x$ behavior of the nucleon structure functions can be justified by the valence quark dominance, [*i.e.*]{}, in the large $x$ region almost all quantum numbers, momentum and the spin of the nucleon are carried by the three valence quarks, which can therefore be identified as constituent quarks. Predictions of $A_1^n$ and $A_1^p$ in the large $x$ region using the hyperfine-perturbed RCQM have been achieved [@theory:cqm]. Perturbative QCD and Hadron Helicity Conservation {#ch2:pqcd} ------------------------------------------------- In the early 1970’s, in one of the first applications of perturbative QCD (pQCD), it was noted that as $x\to 1$, the scattering is from a high-energy quark and thus the process can be treated perturbatively [@theory:farrar]. Furthermore, when the quark OAM is assumed to be zero, the conservation of angular momentum requires that a quark carrying nearly all the momentum of the nucleon ([*i.e.*]{} $x\to 1$) must have the same helicity as the nucleon. This mechanism is called hadron helicity conservation (HHC), and is referred to as the leading-order pQCD in this paper. In this picture, quark-gluon interactions cause only the $S=1$, $S_z=1$ diquark spin projection component rather than the full $S=1$ diquark system to be suppressed as $x\to 1$, which gives $$\begin{aligned} && \Delta u/u \to 1~{\mathrm}{and}~ \Delta d/d\to 1 {\mathrm}{~~as~} x\to 1~; \\ && R^{np}\to \frac{3}{7},~ A_1^{p}\to 1 {\mathrm}{~~and~} A_1^{n}\to 1 {\mathrm}{~~as~} x\to 1~.\end{aligned}$$ This is one of the few places where pQCD can make an absolute prediction for the $x$-dependence of the structure functions or their ratios. However, how low in $x$ and $Q^2$ this picture works is uncertain. HHC has been used as a constraint in a model to fit data on the first moment of the proton $g_1^p$, giving the BBS parameterization [@theory:bbs]. The $Q^2$ evolution was not included in this calculation. Later in the LSS(BBS) parameterization [@theory:lssbbs], both proton and neutron $A_1$ data were fitted directly and the $Q^2$ evolution was carefully treated. Predictions for $A_1^n$ using both BBS and LSS(BBS) parameterizations have been made, as shown in Fig. \[fig:a1nmodel\_all\] and \[fig:a1pmodel\_all\] in Section \[ch2:exp\]. HHC is based on the assumption that the quark OAM is zero. Recent experimental data on the tensor polarization in elastic $e-^2$H scattering [@data:cebaf-t20], neutral pion photo-production [@data:cebaf-gammap] and the proton electro-magnetic form factors [@data:cebaf-F2p; @data:cebaf-Gep] disagree with the HHC predictions [@theory:F2pF1pQCDscaling]. It has been suggested that effects beyond leading-order pQCD, such as quark OAM [@theory:miller; @theory:transpdf; @theory:ji1; @theory:ji2], might play an important role in processes involving quark spin flips. Predictions from Next-to-Leading Order QCD Fits ----------------------------------------------- In a next-to-leading order (NLO) QCD analysis of the world data [@theory:lss2001], parameterizations of the polarized and unpolarized PDFs were performed without the HHC constraint. Predictions of $g_1^p/F_1^p$ and $g_1^n/F_1^n$ were made using these parameterizations, as shown in Fig. \[fig:a1nmodel\_all\] and \[fig:a1pmodel\_all\] in Section \[ch2:exp\]. In a statistical approach, the nucleon is viewed as a gas of massless partons (quarks, antiquarks and gluons) in equilibrium at a given temperature in a finite volume, and the parton distributions are parameterized using either Fermi-Dirac or Bose-Einstein distributions. Based on this statistical picture of the nucleon, a global NLO QCD analysis of unpolarized and polarized DIS data was performed [@theory:stat]. In this calculation $\Delta u/u\approx 0.75$, $\Delta d/d\approx -0.5$ and $A_1^{p,n}<1$ at $x\to 1$. Predictions from Chiral Soliton and Instanton Models ---------------------------------------------------- While pQCD works well in high-energy hadronic physics, theories suitable for hadronic phenomena in the non-perturbative regime are much more difficult to construct. Possible approaches in this regime are quark models, chiral effective theories and the lattice QCD method. Predictions for $A_1^{n,p}$ have been made using chiral soliton models [@theory:chi_weigel; @theory:chi_waka] and the results of Ref.  [@theory:chi_waka] give $A_1^n<0$. The prediction that $A_1^{p}<0$ has also been made in the instanton model [@theory:instanton]. Other Predictions {#ch2:other} ----------------- Based on quark-hadron duality , one can obtain the structure functions and their ratios in the large $x$ region by summing over matrix elements for nucleon resonance transitions. To incorporate SU(6) breaking, different mechanisms consistent with duality were assumed and data on the structure function ratio $R^{np}$ were used to fit the SU(6) mixing parameters. In this picture, $A_1^{n,p}\to 1$ as $x\to 1$ is a direct result. Duality predictions for $A_1^{n,p}$ using different SU(6) breaking mechanisms were performed in Ref. [@theory:dual_new]. There also exist predictions from bag models [@theory:bag], as shown in Fig. \[fig:a1nmodel\_all\] and \[fig:a1pmodel\_all\] in the next section. Previous Measurements of $A_1^n$ {#ch2:exp} -------------------------------- A summary of previous $A_1^n$ measurements is given in ------------------------------ ------------------ ----------------- ------------ ------------- Experiment beam target $x$ $Q^2$ (GeV/c)$^2$ E142 [@data:a1ng1n-e142] 19.42, 22.66, $^3$He 0.03-0.6 2 25.51 GeV; e$^-$ E154 [@data:a1ng1n-e154] 48.3 GeV; e$^-$ $^3$He 0.014-0.7 1-17 HERMES [@data:a1ng1n-hermes] 27.5 GeV; e$^+$ $^3$He 0.023-0.6 1-15 E143 [@data:a1pa1n-e143] 9.7, 16.2, NH$_3$, ND$_3$ 0.024-0.75 0.5-10 29.1 GeV; e$^-$ E155 [@data:g1pg1n-e155] 48.35 GeV; e$^-$ NH$_3$, LiD$_3$ 0.014-0.9 1-40 SMC [@data:g1n-smc] 190 GeV; $\mu^-$ C$_4$H$_{10}$O 0.003-0.7 1-60 C$_4$D$_{10}$O ------------------------------ ------------------ ----------------- ------------ ------------- : Previous measurements of $A_1^n$.[]{data-label="tab:exA1nList"} Table \[tab:exA1nList\]. The data on $A_1^n$ and $A_1^p$ are plotted in Fig. \[fig:a1nmodel\_all\] and \[fig:a1pmodel\_all\] along with theoretical calculations described in previous sections. Since the $Q^2$-dependence of $A_1$ is small and $g_1/F_1\approx A_1$ in DIS, data for $g_1^n/F_1^n$ and $g_1^p/F_1^p$ are also shown and all data are plotted without evolving in $Q^2$. As becomes obvious in Fig. \[fig:a1nmodel\_all\], the precision of previous $A_1^n$ data at $x>0.4$ from SMC [@data:g1n-smc], HERMES [@data:a1ng1n-hermes] and SLAC [@data:a1ng1n-e142; @data:a1pa1n-e143; @data:a1ng1n-e154] is not sufficient to distinguish among different predictions. [THE EXPERIMENT]{} {#exp:main} ================== We report on an experiment [@exp:e99117] carried out at in the Hall A of Thomas Jefferson National Accelerator Facility (Jefferson Lab, or JLab). The goal of this experiment was to provide precise data on $A_1^n$ in the large $x$ region. We have measured the inclusive deep inelastic scattering of longitudinally polarized electrons off a polarized $^3$He target, with the latter being used as an effective polarized neutron target. The scattered electrons were detected by the two standard High Resolution Spectrometers (HRS). The two HRS were configured at the same scattering angles and momentum settings to double the statistics. Data were collected at three $x$ points as shown in Table \[tab:kine\]. Both longitudinal and transverse electron asymmetries were measured, from which $A_1$, $A_2$, $g_1/F_1$ and $g_2/F_1$ were extracted using Eq. (\[equ:a2a1\]–\[equ:a2g2\]). $\langle x\rangle$ 0.327 0.466 0.601 ---------------------------------- ------------ ------------ ------------ $E^\prime$ 1.32 1.72 1.455 $\theta$ $35^\circ$ $35^\circ$ $45^\circ$ $\langle Q^2\rangle$ (GeV/c)$^2$ 2.709 3.516 4.833 $\langle W^2\rangle$ (GeV)$^2$ 6.462 4.908 4.090 : Kinematics of the experiment. The beam energy was $E=5.734$ GeV. $E^\prime$ and $\theta$ are the nominal momentum and angle of the scattered electrons. $\langle x\rangle$, $\langle Q^2\rangle$ and $\langle W^2\rangle$ are values averaged over the spectrometer acceptance.[]{data-label="tab:kine"} Polarized $^3$He as an Effective Polarized Neutron {#ch3:intro} -------------------------------------------------- As shown in Fig. \[fig:a1nmodel\_all\], previous data on $A_1^n$ did not have sufficient precision in the large $x$ region. This is mainly due to two experimental limitations. Firstly, high polarization and luminosity required for precision measurements in the large $x$ region were not available previously. Secondly, there exists no free dense neutron target suitable for a scattering experiment, mainly because of the neutron’s short lifetime ($\approx 886$ sec). Therefore polarized nuclear targets such as $^2\vec{{\mathrm}{H}}$ or $^3\vec{{\mathrm}{He}}$ are commonly used as effective polarized neutron targets. Consequently, nuclear corrections need to be applied to extract neutron results from nuclear data. ![An illustration of $^3$He wavefunction. The $S$, $S^\prime$ and $D$ state contributions are from calculations using the AV18 two-nucleon interaction and the Tucson-Melbourne three-nucleon force, as given in Ref. [@theory:PnPp_nogga]. []{data-label="fig:he3wavefunc"}](fig3.eps) For a polarized deuteron, approximately half of the deuteron spin comes from the proton and the other half comes from the neutron. Therefore the neutron results extracted from the deuteron data have a significant uncertainty coming from the error in the proton data. The advantage of using $^3\vec{{\mathrm}{He}}$ is that the two protons’ spins cancel in the dominant $S$ state of the $^3$He wavefunction, thus the spin of the $^3$He comes mainly ($>87\%$) from the neutron [@theory:PnPp_friar; @theory:PnPp_nogga], as illustrated in Fig. \[fig:he3wavefunc\]. As a result, there is less model dependence in the procedure of extracting the spin-dependent observables of the neutron from $^3$He data. At large $x$, the advantage of using a polarized $^3$He target is more prominent in the case of $A_1^n$. In this region almost all calculations show that $A_1^n$ is much smaller than $A_1^p$, therefore the $A_1^n$ results extracted from nuclear data are more sensitive to the uncertainty in the proton data and the nuclear model being used. In the large $x$ region, the cross sections are small because the parton densities drop dramatically as $x$ increases. In addition, the Mott cross section, given by Eq. \[equ:Mottxsec\], is small at large $Q^2$. To achieve a good statistical precision, high luminosity is required. Among all laboratories which are equipped with a polarized $^3$He target and are able to perform a measurement of the neutron spin structure, the polarized electron beam at JLab, combined with the polarized $^3$He target in Hall A, provides the highest polarized luminosity in the world [@thesis:zheng]. Hence it is the best place to study the large $x$ behavior of the neutron spin structure. The Accelerator and the Polarized Electron Source {#ch3:beam} ------------------------------------------------- JLab operates a continuous-wave electron accelerator that recirculates the beam up to five times through two super-conducting linear accelerators. Polarized electrons are extracted from a strained GaAs photocathode [@exp:polestrain] illuminated by circularly polarized light, providing a polarized beam of $(70-80)\%$ polarization and $\approx 200\mu A$ maximum current to experimental halls A, B and C. The maximum beam energy available at JLab so far is 5.7 GeV, which was also the beam energy used during this experiment. Hall A Overview --------------- The basic layout of Hall A during this experiment is shown in Fig. \[fig:floorplan\]. The major instrumentation [@exp:NIM] includes beamline equipment, the target and two HRSs. ![(Color online) Top-view of the experimental hall A (not to scale).[]{data-label="fig:floorplan"}](fig4.eps) The beamline starts after the arc section of the accelerator where the beam is bent into the hall, and ends at the beam dump. The arc section can be used for beam energy measurement, as will be described in Section \[ch3:beamenergy\]. After the arc section, the beamline is equipped with a Compton polarimeter, two Beam Current Monitors (BCM) and an Unser monitor for absolute beam current measurement, a fast raster, the eP device for beam energy measurement, a [M$\o$ller ]{}polarimeter and two Beam Position Monitors (BPM). These beamline elements, together with spectrometers and the target, will be described in detail in the following sections. Beam Energy Measurement {#ch3:beamenergy} ----------------------- The energy of the beam was measured absolutely by two independent methods - ARC and eP [@exp:NIM; @exp:NIM-52]. Both methods can provide a precision of $\delta E_{beam}/E_{beam}\approx 2\times10^{-4}$. For the ARC method [@exp:NIM; @thesis:marchand], the deflection of the beam in the arc section of the beamline is used to determine the beam energy. In the eP measurement [@exp:NIM; @thesis:ravel] the beam energy is determined by the measurement of the scattered electron angle $\theta_e$ and the recoil proton angle $\theta_p$ in $^1$H$(e,e'p)$ elastic scattering. Beam Polarization Measurement {#ch3:beamPol} ----------------------------- Two methods were used during this experiment to measure the electron beam polarization. The [M$\o$ller ]{}polarimeter [@exp:NIM] measures [M$\o$ller ]{}scattering of the polarized electron beam off polarized atomic electrons in a magnetized foil. The cross section of this process depends on the beam and target polarizations. The polarized electron target used by the [M$\o$ller ]{}polarimeter was a ferromagnetic foil, with its polarization determined from foil magnetization measurements. The [M$\o$ller ]{}measurement is invasive and typically takes an hour, providing a statistical accuracy of about $0.2\%$. The systematic error comes mainly from the error in the foil target polarization. An additional systematic error is due to the fact that the beam current used during a [M$\o$ller ]{}measurement ($\approx 0.5\mu$A) is lower than that used during the experiment. The total relative systematic error was $\approx 3.0\%$ during this experiment. During a Compton polarimeter [@exp:NIM; @thesis:baylac] measurement, the electron beam is scattered off a circularly polarized photon beam and the counting rate asymmetry of the Compton scattered electrons or photons between opposite beam helicities is measured. The Compton polarimeter measures the beam polarization concurrently with the experiment running in the hall. The Compton polarimeter consists of a magnetic chicane which deflects the electron beam away from the scattered photons, a photon source, an electromagnetic calorimeter and an electron detector. The photon source was a 200 mW laser amplified by a resonant Fabry-Perot cavity. During this experiment the maximum gain of the cavity reached $G_{max}=7500$, leading to a laser power of $1500$ W inside the cavity. The circular polarization of the laser beam was $>99\%$ for both right and left photon helicity states. The asymmetry measured in Compton scattering at JLab with a $1.165$ eV photon beam and the $5.7$ GeV electron beam used by this experiment had a mean value of $\approx 2.2\%$ and a maximum of $9.7\%$. For a 12 $\mu$A beam current, one hour was needed to reach a relative statistical accuracy of $(\Delta P_b)_{stat}/P_b\approx 1\%$. The total systematic error was $(\Delta P_b)_{sys}/P_b\approx 1.6\%$ during this experiment. The average beam polarization during this experiment was extracted from a combined analysis of 7 [M$\o$ller ]{}and 53 Compton measurements. A value of $(79.7\pm 2.4)\%$ was used in the final DIS analysis. Beam Helicity {#exp:helicity} ------------- The helicity state of electrons is regulated every $33$ ms at the electron source. The time sequence of the electrons’ helicity state is carried by helicity signals, which are sent to experimental halls and the data acquisition (DAQ) system. Since the status of the helicity signal (H+ or H- pulses) has either the same or the opposite sign as the real electron helicity, the absolute helicity state of the beam needs to be determined by other methods, as will be described later. There are two modes – toggle and pseudorandom – which can be used for the pulse sequence of the helicity signal. In the toggle mode, the helicity alternates every $33$ ms. In the pseudorandom mode, the helicity alternates randomly at the beginning of each pulse pair, of which the two pulses must have opposite helicities in order to equalize the numbers of the H+ and H- pulses. The purpose of the pseudorandom mode is to minimize any possible time-dependent systematic errors. ![Helicity signal and the helicity status of DAQ in toggle (top) and pseudorandom (bottom) modes.[]{data-label="fig:helpulse"}](fig5a.eps "fig:")\ ![Helicity signal and the helicity status of DAQ in toggle (top) and pseudorandom (bottom) modes.[]{data-label="fig:helpulse"}](fig5b.eps "fig:") Fig. \[fig:helpulse\] shows the helicity signals and the helicity states of the DAQ system for the two regulation modes. There is a half-wave plate at the polarized source which can be inserted to reverse the helicity of the laser illuminating the photocathode hence reverse the helicity of electron beam. During the experiment this half-wave plate was inserted for half of the statistics to minimize possible systematic effects related to the beam helicity. The scheme described above was used to monitor the relative changes of the helicity state. The absolute sign of the electrons’ helicity states during each of the H+ and H- pulses were confirmed by measuring a well known asymmetry and comparing the measured asymmetry with its prediction, as will be presented in Section \[ana:elana\] and \[ana:delta\]. Beam Charge Measurement and Charge Asymmetry Feedback {#ch3:bcm} ----------------------------------------------------- The beam current was measured by the BCM system located upstream of the target on the beamline. The BCM signals were fed to scaler inputs and were inserted in the data stream. Possible beam charge asymmetry measured at Hall A can be caused by the timing asymmetry of the DAQ system, or by the timing and the beam intensity asymmetries at the polarized electron source. The beam intensity asymmetry originates from the intensity difference between different helicity states of the circularly polarized laser used to strike the photocathode. Although the charge asymmetry can be corrected for to first order, there may exist unknown non-linear effects which can cause a systematic error in the measured asymmetry. Thus the beam charge asymmetry should be minimized. This was done by using a separate DAQ system initially developed for the parity-violation experiments [@exp:parityQasym], called the parity DAQ. The parity DAQ used the measured charge asymmetry in Hall A to control the orientation of a rotatable half-wave plate located before the photocathode at the source, such that intensities for each helicity state of the polarized laser used to strike the photocathode were adjusted accordingly. The parity DAQ was synchronized with the two HRS DAQ systems so that the charge asymmetry in the two different helicity states could be monitored for each run. The charge asymmetry was typically controlled to be below $2\times 10^{-4}$ during this experiment. Raster and Beam Position Monitor {#ch3:bpm} -------------------------------- To protect the target cell from being damaged by the effect of beam-induced heating, the beam was rastered at the target. The raster consists a pair of horizontal and vertical air-core dipoles located upstream of the target on the beamline, which can produce either a rectangular or an elliptical pattern. We used a raster pattern distributed uniformly over a circular area with a radius of 2 mm. The position and the direction of the beam at the target were measured by two BPMs located upstream of the target [@exp:NIM]. The beam position can be measured with a precision of 200 $\mu$m with respect to the Hall A coordinate system. The beam position and angle at the target were recorded for each event. High Resolution Spectrometers {#ch3:hrs} ----------------------------- The Hall A High Resolution Spectrometer (HRS) systems were designed for detailed investigations of the structure of nuclei and nucleons. They provide high resolution in momentum and in angle reconstruction of the reaction product as well as being able to be operated at high luminosity. For each spectrometer, the vertically bending design includes two quadrupoles followed by a dipole magnet and a third quadrupole. All quadrupoles and the dipole are superconducting. Both HRSs can provide a momentum resolution better than $2\times 10^{-4}$ and a horizontal angular resolution better than 2 mrad with a design maximum central momentum of 4 GeV/c [@exp:NIM]. By convention, the two spectrometers are identified as the left and the right spectrometers based on their position when viewed looking downstream. ![(Color online) Schematic layout of the left HRS and detector package (not to scale).[]{data-label="fig:hrs_side"}](fig6.eps) The basic layout of the left HRS is shown in Fig. \[fig:hrs\_side\]. The detector package is located in a large steel and concrete detector hut following the last magnet. For this experiment the detector package included (1) two scintillator planes S1 and S2 to provide a trigger to activate the DAQ electronics; (2) a set of two Vertical Drift Chambers (VDC) [@exp:vdc] for particle tracking; (3) a gas $\breve{{\mathrm}{C}}$erenkov detector to provide particle identification (PID) information; and (4) a set of lead glass counters for additional PID. The layout of the right HRS is almost identical except a slight difference in the geometry of the gas $\breve{{\mathrm}{C}}$erenkov detector and the lead glass counters.\ Particle Identification ----------------------- For this experiment the largest background came from photo-produced pions. We refer to PID in this paper as the identification of electrons from pions. PID for each HRS was accomplished by a CO$_2$ threshold gas $\breve{{\mathrm}{C}}$erenkov detector and a double-layered lead glass shower detector. The two $\breve{{\mathrm}{C}}$erenkov detectors, one on each HRS, were operated with CO$_2$ at atmospheric pressure. The refraction index of the CO$_2$ gas was 1.00041, giving a threshold momentum of $\approx 17$ MeV/c for electrons and $\approx 4.8$ GeV/c for pions. The incident particles on each HRS were also identified by their energy deposits in the lead glass shower detector. Since $\breve{{\mathrm}{C}}$erenkov detectors and lead glass shower detectors are based on different mechanisms and their PID efficiencies are not correlated [@book:partdet], we extracted the PID efficiency of the lead glass counters by using electron events selected by the $\breve{{\mathrm}{C}}$erenkov detector, and [*vice versa*]{}. Fig. \[fig:leftgcadc\] shows a spectrum of the summed ADC signal of the left HRS gas $\breve{{\mathrm}{C}}$erenkov detector, without a cut on the lead glass signal and after applying such lead glass electron and pion cuts. The spectrum from the right HRS is similar. ![(Color online) Summed ADC signal of the left HRS gas $\breve{{\mathrm}{C}}$erenkov detector: without cuts, after lead glass counters electron cut and after pion cut. The vertical line shows a cut $\sum{{\mathrm}{ADC}_i}>400$ applied to select electrons.[]{data-label="fig:leftgcadc"}](fig7.eps) Fig. \[fig:rightshadc\] shows the distribution of the energy deposit in the two layers of the right HRS lead glass counters, without a $\breve{{\mathrm}{C}}$erenkov cut, and after $\breve{{\mathrm}{C}}$erenkov electron and pion cuts. ![(Color online) Energy deposited in the first layer (preshower) [*vs*]{} that in the second layer (shower) of lead glass counters in the right HRS. The two blobs correspond to the spectrum with a tight gas $\breve{{\mathrm}{C}}$erenkov ADC electron cut and with a pion cut applied. The lines show the boundary of the two-dimensional cut used to select electrons in the data analysis.[]{data-label="fig:rightshadc"}](fig8.eps) Detailed PID analysis was done both before and during the experiment. The PID performance of each detector is characterized by the electron detection efficiency $\eta_{e}$ and the pion rejection factor $\eta_{\pi,{\mathrm}{rej}}$, defined as the number of pions needed to cause one pion contamination event. In the HRS central momentum range of $0.8<p_0<2.0$ (GeV/c), the PID efficiencies for the left HRS were found to be\ $\diamond$ Gas $\breve{{\mathrm}{C}}$erenkov:   [$\eta_{\pi,{\mathrm}{rej}}>770$ at $\eta_{e}=99.9\%$;]{}\ $\diamond$ Lead glass counters: [ $\eta_{\pi,{\mathrm}{rej}} \approx 38$ at $\eta_{e}=98\%$;]{}\ $\diamond$ Combined: $\eta_{\pi,{\mathrm}{rej}} >3\times 10^{4}$ at $\eta_{e}=98\%$.\ \ and for the right HRS were\ $\diamond$ Gas $\breve{{\mathrm}{C}}$erenkov:   [$\eta_{\pi,{\mathrm}{rej}}=900$ at $\eta_{e}=99\%$]{};\ $\diamond$ Lead glass counters: [ $\eta_{\pi,{\mathrm}{rej}} \approx 182$ at $\eta_{e}=98\%$;]{}\ $\diamond$ Combined: $\eta_{\pi,{\mathrm}{rej}} >1.6\times 10^{5}$ at $\eta_{e}=97\%$.\ Data Acquisition System {#ch3:daq} ----------------------- We used the CEBAF Online Data Acquisition (CODA) system [@exp:coda] for this experiment. In the raw data file, data from the detectors, the beamline equipment, and from the slow control software were recorded. The total volume of data accumulated during the two-month running period was about 0.6 TBytes. Data from the detectors were processed using an analysis package called Experiment Scanning Program for hall A Collaboration Experiments (ESPACE) [@exp:espace]. ESPACE was used to filter raw data, to make histograms for reconstructed variables, to export variables into ntuples for further analysis, and to calibrate experiment-specific detector constants. It also provided the possibility to apply conditions on the incoming data. The information from scaler events was used to extract beam charge and DAQ deadtime corrections. The Polarized $\bm{^3\mathrm{He}}$ Target {#targ:main} ========================================= Polarized $^3$He targets are widely used at SLAC, DESY, MAINZ, MIT-Bates and JLab to study the electromagnetic structure and the spin structure of the neutron. There exist two major methods to polarize $^3$He nuclei. The first one uses the metastable-exchange optical pumping technique [@targ:metaexch]. The second method is based on optical pumping [@targ:optpump] and spin exchange [@targ:spinexch]. It has been used at JLab since 1998 [@exp:e94010], and was used here. The $^3\vec{{\mathrm}{He}}$ target at JLab Hall A uses the same design as the SLAC $^3\vec{{\mathrm}{He}}$ target [@thesis:romalis]. The first step to polarize $^3$He nuclei is to polarize an alkali metal vapor (rubidium was used at JLab as well as at SLAC) by optical pumping [@targ:optpump] with circularly polarized laser light. Depending on the photon helicity, the electrons in the Rb atoms will accumulate at either the $F=3,m_F=3$ or the $F=3,m_F=-3$ level (here $F$ is the atom’s total spin and $m_F$ is its projection along the magnetic field axis). The polarization is then transfered to the $^3$He nuclei through the spin exchange mechanism [@targ:spinexch] during collisions between Rb atoms and the $^3$He nuclei. Under operating conditions the $^3$He density is about $10^{20}$ nuclei/cm$^{3}$ and the Rb density is about $10^{14}$ atoms/cm$^{3}$. To minimize depolarization effects caused by the unpolarized light emitted from decay of the excited electrons, N$_2$ buffer gas was added to provide a channel for the excited electrons to decay to the ground state without emitting photons [@targ:optpump]. In the presence of N$_2$, electrons decay through collisions between the Rb atoms and N$_2$ molecules, which is usually referred to as non-radiative quenching. The number density of N$_2$ was about $1\%$ of that of $^3$He. Target Cells {#ch4:cell} ------------ The target cells used for this experiment were 25-cm long pressurized glass cells with $\sim130$-$\mu$m thick end windows. ![(Color online) JLab target cell, geometries are given in mm for cell \#2 used in this experiment.[]{data-label="tfig:targcell"}](fig9.eps){width="230pt"} \[htp\] Name $V_p$ $V_t$ $V_{tr}$ $V_0$ $L_{tr}$ $n_0$ lifetime ------------- ------- ------- ---------- ------- ---------- ------- ---------- Cell \#1 116.7 51.1 3.8 171.6 6.574 9.10 49 Cell \#2 116.1 53.5 3.9 173.5 6.46 8.28 44 uncertainty 1.5 1.0 0.25 1.8 0.020 2% 1 : Target cell characteristics. Symbols are: $V_p$ pumping chamber volume in cm$^3$; $V_t$ target chamber volume in cm$^3$; $V_{tr}$ transfer tube volume in cm$^3$; $V_0$ total volume in cm$^3$; $L_{tr}$ transfer tube length in cm; $n_0$: $^3$He density in amg at room temperature (1 amg $= 2.69\times 10^{-19}/$cm$^3$ which corresponds to the gas density at the standard pressure and $T=0^\circ$C); lifetime is in hours. []{data-label="tab:cellchar"} The cell consisted of two chambers, a spherical upper chamber which holds the Rb vapor and in which the optical pumping occurs, and a long cylindrical chamber where the electron beam passes through and interacts with the polarized $^3$He nuclei. Two cells were used for this experiment. Figure \[tfig:targcell\] is a picture of the first cell with dimensions shown in mm. Table \[tab:cellchar\] gives the cell volumes and densities. Target Setup {#ch4:setup} ------------ Figure \[tfig:overview\] is a schematic diagram of the target setup. There were two pairs of Helmholtz coils to provide a 25 G main holding field, with one pair oriented perpendicular and the other parallel to the beamline (only the perpendicular pair is shown). The holding field could be aligned in any horizontal direction with respect to the incident electron beam. The coils were excited by two power supplies in the constant voltage mode. The coil currents were continuously measured and recorded by the slow control system. ![(Color online) Target setup overview (schematic).[]{data-label="tfig:overview"}](fig10.eps){width="230pt"} The cell was held at the center of the Helmholtz coils with its pumping chamber mounted inside an oven heated to $170^\circ$C in order to vaporize the Rb. The lasers used to polarize the Rb were three 30 W diode lasers tuned to a wavelength of 795 nm. The target polarization was measured by two independent methods – the NMR (Nuclear Magnetic Resonance) [@exp:NIM; @exp:e94010; @thesis:kramer] and the EPR (Electro Paramagnetic Resonance) [@exp:NIM; @exp:e94010; @PRA03004; @thesis:zheng] polarimetry. The NMR system consisted of one pair of pick-up coils (one on each side of the cell target chamber), one pair of RF coils and the associated electronics. The RF coils were placed at the top and the bottom of the scattering chamber, oriented in the horizontal plane, as shown in Fig. \[tfig:overview\]. The EPR system shared the RF coils with the NMR system. It consisted of one additional RF coil to induce light signal emission from the pumping chamber, a photodiode and the related optics to collect the light, and associated electronics for signal processing. Laser System {#ch4:laser} ------------ The laser system used during this experiment consisted of seven diode lasers – three for longitudinal pumping, three for transverse pumping and one spare. To protect the diode lasers from radiation damage from the electron beam, as well as to minimize the safety issues related to the laser hazard, the diode lasers and the associated optics system were located in a concrete laser hut located on the right side of the beamline at $90^\circ$, as shown in Fig. \[fig:floorplan\]. The laser optics had seven individual lines, each associated with one diode laser. All seven optical lines were identical and were placed one on top of the other on an optics table inside the laser hut. Each optical line consisted of one focusing lens to correct the angular divergence of the laser beam, one beam-splitter to linearly polarize the lasers, two mirrors to direct them, three quarter waveplates to convert linear polarization to circular polarization, and two half waveplates to reverse the laser helicity. Figure \[tfig:lasersetup\] shows a schematic diagram ![(Color online) Laser polarizing optics setup (schematic) for the Hall A polarized $^3$He target.[]{data-label="tfig:lasersetup"}](fig11.eps){width="240pt"} of one optics line. Under the operating conditions for either longitudinal or transverse pumping, the original beam of each diode laser was divided into two by the beam-splitter. Therefore there were a total of six polarized laser beams entering the target. The diameter of each beam was about 5 cm which approximately matched the size of the pumping chamber. The target was about 5 m away from the optical table. For the pumping of the transversely polarized target, all these laser beams went directly towards the pumping chamber of the cell through a window on the side of the target scattering chamber enclosure. For longitudinal pumping, they were guided towards the top of the scattering chamber, then were reflected twice and finally reached the cell pumping chamber. NMR Polarimetry {#ch4:nmr} --------------- The polarization of the $^3$He was determined by measuring the $^3$He Nuclear Magnetic Resonance (NMR) signal. The principle of NMR polarimetry is the spin reversal of $^3$He nuclei using the Adiabatic Fast Passage (AFP) [@targ:AFP] technique. At resonance this spin reversal will induce an electromagnetic field and a signal in the pick-up coil pair. The signal magnitude is proportional to the polarization of the $^3$He and can be calibrated by performing the same measurement on a water sample, which measures the known thermal polarization of protons in water. The systematic error of the NMR measurement was about $3\%$, dominated by the error in the water calibration [@thesis:kramer]. EPR Polarimetry {#ch4:epr} --------------- In the presence of a magnetic field, the Zeeman splitting of Rb, characterized by the Electron-Paramagnetic Resonance frequency $\nu_{\mathrm}{EPR}$, is proportional to the field magnitude. When $^3$He nuclei are polarized ($P\approx 40\%$), their spins generate a small magnetic field $B_{^3{\mathrm}{He}}$ of the order of $\approx 0.1$ Gauss, super-imposed on the main holding field $B_{H}=25$ Gauss. During an EPR measurement [@PRA03004] the spin of the $^3$He is flipped by AFP, hence the direction of $B_{^3{\mathrm}{He}}$ is reversed and the change in the total field magnitude causes a shift in $\nu_{\mathrm}{EPR}$. This frequency shift $\delta\nu_{\mathrm}{EPR}$ is proportional to the $^3$He polarization in the pumping chamber. The $^3$He polarization in the target chamber is calculated using a model which describes the polarization diffusion from the pumping chamber to target chamber. The value of the EPR resonance frequency $\nu_{\mathrm}{EPR}$ can also be used to calculate the magnetic field magnitude. The systematic error of the EPR measurement was about $3\%$, which came mainly from uncertainties in the cell density and temperature, and from the diffusion model. Target Performance {#ch4:polr} ------------------ The target polarizations measured during this experiment are shown in Fig. \[tfig:targperf\]. Results from the two polarimetries are in good agreement and the average target polarization in beam was $(40.0\pm 2.4)\%$. In a few cases the polarization measurement itself caused an abrupt loss in the polarization. This phenomenon may be the so-called “masing effect” [@thesis:romalis] due to non-linear couplings between the $^3$He spin rotation and conducting components inside the scattering chamber, [*e.g.*]{}, the NMR pick-up coils, and the “Rb-ring” formed by the rubidium condensed inside the cell at the joint of the two chambers. This masing effect was later suppressed by adding coils to produce an additional field gradient. Data Analysis {#ch5:main} ============= In this section we present the analysis procedure leading to the final results in Section \[result:main\]. We start with the analysis of elastic scattering, the $\Delta(1232)$ transverse asymmetry, and the check for false asymmetry. Next, the DIS analysis and radiative corrections are presented. Finally we describe nuclear corrections which were used to extract neutron structure functions from the $^3$He data. Analysis Procedure {#ch5:procedure} ------------------ The procedure to extract the electron asymmetries from our data is outlined in Fig. \[fig:ana\_procedure\]. ![Procedure for asymmetry analysis.[]{data-label="fig:ana_procedure"}](fig13.eps) From the raw data one first obtains the helicity-dependent electron yield $N^{\pm}$ using acceptance and PID cuts. The efficiencies associated with these cuts are not helicity-dependent, hence are not corrected for in the asymmetry analysis. The yield is then corrected for the helicity-dependent integrated beam charge $Q^\pm$ and the livetime of the DAQ system ${\mathrm}{LT}^\pm$. The asymmetry of the corrected yield is the raw asymmetry $A_{raw}$. Next, to go from $A_{raw}$ to the physics asymmetries $A_{\parallel}$ and $A_{\perp}$, four factors need to be taken into account: the beam polarization $P_b$, the target polarization $P_t$, the nitrogen dilution factor $f_{{\mathrm}{N}_2}$ due to the unpolarized nitrogen nuclei mixed with the polarized $^3$He gas, and a sign based on the knowledge of the absolute state of the electron helicity and the target spin direction: $$\begin{aligned} A_{\parallel,\perp} &=& \pm\frac{A_{raw}}{f_{{\mathrm}{N}_2}P_bP_t} \label{equ:he3raw0} \end{aligned}$$ The results of the beam and the target polarization measurements have been presented in previous sections. The nitrogen dilution factor is obtained from data taken with a reference cell filled with nitrogen. The sign of the asymmetry is described by “the sign convention”. The sign convention for parallel asymmetries was obtained from the elastic scattering asymmetry and that for perpendicular asymmetries was from the $\Delta(1232)$ asymmetry analysis, as will be described in Sections \[ana:elana\] and \[ana:delta\]. The physics asymmetries $A_{\parallel}$ and $A_{\perp}$, after corrections for radiative effects, were used to calculate $A_1$ and $A_2$ and the structure function ratios $g_1/F_1$ and $g_2/F_1$ using Eq. (\[equ:a2a1\]—\[equ:a2g2\]). Then the last step is to apply nuclear corrections in order to extract the neutron asymmetries and the structure function ratios from the $^3$He results, as will be described in Section \[ch5:he3model\].\ Although the main goal of this experiment was to provide precise data on the asymmetries, cross sections were also extracted from the data. The procedure for the cross section analysis is outlined in Fig. \[fig:ana\_xsec\_procedure\]. One first determines the absolute yield of $\vec{e}-\vec{^3{\mathrm}{He}}$ inclusive scattering from the raw data. Unlike the asymmetry analysis, corrections need to be made for the detector and PID efficiencies and the spectrometer acceptance. A Monte-Carlo simulation is used to calculate the spectrometer ![Procedure for cross section analysis.[]{data-label="fig:ana_xsec_procedure"}](fig14.eps) acceptance based on a transport model for the HRS [@exp:NIM] with radiative effects taken into account. One then subtracts the yield of $e-{\mathrm}{N}$ scattering caused by the N$_2$ nuclei in the target. The clean $\vec{e}-\vec{^3{\mathrm}{He}}$ yield is then corrected for the helicity-averaged beam charge and the DAQ livetime to give cross section results. Using world fits for the unpolarized structure functions (form factors) of $^3$He, one can calculate the expected DIS (elastic) cross section from the Monte-Carlo simulation and compare to the data. Elastic Analysis {#ana:elana} ---------------- Data for $\vec{e}-^3\vec{{\mathrm}{He}}$ elastic scattering were taken on a longitudinally polarized target with a beam energy of 1.2 GeV. The scattered electrons were detected at an angle of $20^\circ$. The formalism for the cross sections and asymmetries are summarized in Appendix \[app:el\]. Results for the elastic asymmetry were used to check the product of beam and target polarizations, as well as to determine the sign convention for different beam-helicity states and target spin directions.\ The raw asymmetry was extracted from the data by $$\begin{aligned} A_{raw} &=& \frac{\frac{N^+}{Q^+LT^+}-\frac{N^-}{Q^-LT^-}} {\frac{N^+}{Q^+LT^+}+\frac{N^-}{Q^-LT^-}} \label{equ:he3elaraw} \end{aligned}$$ with $N^{\pm}$, $Q^{\pm}$ and $LT^{\pm}$ the helicity-dependent yield, beam charge and livetime correction, respectively. The elastic asymmetry is $$\begin{aligned} A_{\parallel}^{el} &=& \pm\frac{A_{raw}}{f_{{\mathrm}{N}_2}f_{QE} P_bP_t} \label{equ:he3elaphys} \end{aligned}$$ with $f_{{\mathrm}{N}_2}= 0.975\pm 0.003$ the N$_2$ dilution factor determined from data taken with a reference cell filled with nitrogen, and $P_b$ and $P_t$ the beam and target polarization, respectively. A cut in the invariant mass $\vert{W-M_{^3{\mathrm}{He}}}\vert < 6$ (MeV) was used to select elastic events. Within this cut there are a small amount of quasi-elastic events and $f_{QE}>0.99$ is the quasi-elastic dilution factor used to correct for this effect. ![(Color online) Elastic parallel asymmetry results for the two HRS. The kinematics are $E=1.2$ GeV and $\theta=20^\circ$. A cut in the invariant mass $\vert{W-M_{^3{\mathrm}{He}}}\vert < 6$ (MeV) was used to select elastic events. Data from runs with beam half-wave plate inserted are shown as triangles. The error bars shown are total errors including a $4.5\%$ systematic uncertainty, which is dominated by the error of the beam and target polarizations. The combined asymmetry and its total error from $\approx 20$ runs are shown by the horizontal solid and dashed lines, respectively, as well as the solid circle as labeled [@thesis:zheng].[]{data-label="fig:elasym"}](fig15.eps) The sign on the right hand side of Eq. (\[equ:he3elaphys\]) depends on the configuration of the beam half-wave plate, the spin precession of electrons in the accelerator, and the target spin direction. It was determined by comparing the sign of the measured raw asymmetries with the calculated elastic asymmetry. We found that for this experiment the electron helicity was aligned to the beam direction during H+ pulses when the beam half-wave plate was [*not*]{} inserted. Since the electron spin precession in the accelerator can be well calculated using quantum electro-dynamics and the results showed that the beam helicity during H+ pulses was the same for the two beam energies used for elastic and DIS measurements, the above convention also applies to the DIS data analysis. A Monte-Carlo simulation was performed which took into account the spectrometer acceptance, the effect of the quasi-elastic scattering background and radiative effects. Results for the elastic asymmetry and the cross section are shown in Fig. \[fig:elasym\] and \[fig:elxsec\], respectively, along with the expected values from the simulation. The data show good agreement with the simulation within the uncertainties. ![(Color online) Elastic cross section results for the two HRS. The kinematics were $E=1.2$ GeV and $\theta=20^\circ$. A systematic error of $6.7\%$ was assigned to each data point, which was dominated by the uncertainty in the target density and the HRS transport functions [@thesis:zheng].[]{data-label="fig:elxsec"}](fig16.eps) $\Delta (1232)$ Transverse Asymmetry {#ana:delta} ------------------------------------ Data on the $\Delta(1232)$ resonance were taken on a transversely polarized target using a beam energy of $1.2$ GeV. The scattered electrons were detected at an angle of $20^\circ$ and the central momentum of the spectrometers was set to $0.8$ GeV/c. The transverse asymmetry defined by Eq. (\[equ:Aperpdef\]) was extracted from the raw asymmetry using Eq. (\[equ:he3raw0\]). ![(Color online) Measured raw $\Delta(1232)$ transverse asymmetry, with beam half-wave plate inserted and target spin pointing to the left side of the beamline. The kinematics are $E=1.2$ GeV, $\theta=20^\circ$ and $E^\prime=0.8$ GeV/c. The dashed lines show the expected value obtained from previous $^3$He data extrapolated in $Q^2$.[]{data-label="fig:deltaasym"}](fig17.eps) A cut in the invariant mass $\vert W-1232\vert < 20$ (MeV) was used to select $\Delta(1232)$ events. The sign on the right hand side of Eq. (\[equ:he3raw0\]) depends on the beam half-wave plate status, the spin precession of electrons in the accelerator, the target spin direction, and in which (left or right) HRS the asymmetry is measured. Since data from a previous experiment [@exp:e94010] in a similar kinematic region showed that $A^\Delta_\parallel<0$ and $A^\Delta_\perp>0$ [@thesis:deur], $A^\Delta_\perp$ can be used to determine the sign convention of the measured transverse asymmetries. The raw $\Delta(1232)$ transverse asymmetry measured during this experiment was positive on the left HRS, as shown in Fig. \[fig:deltaasym\], with the beam half-wave plate inserted and the target spin pointing to the left side of the beamline. Also shown is the expected value obtained from previous $^3$He data extrapolated in $Q^2$. Similar to the longitudinal configuration, this convention applied to both the $\Delta(1232)$ and DIS measurements. False Asymmetry and Background {#ana:false} ------------------------------ False asymmetries were checked by measuring the asymmetries from a polarized beam scattering off an unpolarized $^{12}$C target. The results show that the false asymmetry was less than $2\times 10^{-3}$, which was negligible compared to the statistical uncertainties of the measured $^3$He asymmetries. To estimate the background from pair production $\gamma\to e^-+e^+$, the positron yield was measured at $x=0.33$, which is expected to have the highest pair production background. The positron cross section was found to be $\approx 3\%$ of the total cross section at $x=0.33$, and the positron contribution at $x=0.48$ and $x=0.61$ should be even smaller. The effect of pair production asymmetry is negligible compared to the statistical uncertainties of the measured $^3$He asymmetries and is not corrected for in this analysis. DIS Analysis {#ana:disana} ------------ The longitudinal and transverse asymmetries defined by Eq. (\[equ:Apardef\]) and (\[equ:Aperpdef\]) for DIS were extracted from the raw asymmetries as $$\begin{aligned} A_{\parallel,\perp} &=& \pm\frac{A_{raw}}{f_{{\mathrm}{N}_2} P_bP_t} \label{equ:aphysdis} \end{aligned}$$ where the sign on the right hand side was determined by the procedure described in Sections \[ana:elana\] and \[ana:delta\]. The N$_2$ dilution factor, extracted from runs where a reference cell was filled with pure N$_2$, was found to be $f_{{\mathrm}{N}_2} 0.938\pm 0.007$ for all three DIS kinematics. Radiative corrections were performed for the $^3$He asymmetries $A_\parallel^{^3{\mathrm}{He}}$ and $A_\perp^{^3{\mathrm}{He}}$. We denote by $A^{obs}$ the observed asymmetry, $A^{Born}$ the non-radiated (Born) asymmetry, $\Delta A^{ir}$ the correction due to internal radiation effects and $\Delta A^{er}$ the one due to external radiation effects. One has $A^{Born}=A^{obs}+\Delta A^{ir}+\Delta A^{er}$ for a specific target spin orientation. Internal corrections were calculated using an improved version of POLRAD 2.0 [@ana:polrad]. External corrections were calculated with a Monte-Carlo simulation based on the procedure first described by Mo and Tsai [@ana:motsai]. Since the theory of radiative corrections is well established [@ana:motsai], the accuracy of the radiative correction depends mainly on the structure functions used in the procedure. To estimate the uncertainty of both corrections, five different fits [@ana:f2comfst; @ana:f2nmc92; @ana:f2nmc95; @ana:f2pslac94; @ana:f2phallc02] were used for the unpolarized structure function $F_2$ and two fits [@ana:r1990; @ana:r1998] were used for the ratio $R$. For the polarized structure function $g_1$, in addition to those used in POLRAD 2.0 [@ana:f2g1sch; @ana:f2g1grsv96], we fit to world $g_1^p/F_1^p$ and $g_1^n/F_1^n$ data including the new results from this experiment. Both fits will be presented in Section \[result:neutron\]. For $g_2$ we used both $g_2^{WW}$ and an assumption that $g_2=0$. The variation in the radiative corrections using the fits listed above was taken as the full uncertainty of the corrections. $x$ 0.33, 0.48 0.61 0.61 ------------------------ --------------- --------------- --------------- $\theta$ 35$^\circ$ 45$^\circ$ 45$^\circ$ Cell \#2 \#2 \#1 Cell window ($\mu$m) 144 144 132 $X_0$ (before) 0.00773 0.00773 0.00758 $d$ (g/cm$^2$, before) 0.23479 0.23479 0.23317 Cell wall (mm) 1.44/1.33 1.44/1.33 1.34/1.43 $X_0$ (after) 0.0444/0.0416 0.0376/0.0354 0.0356/0.0374 $d$ (g/cm$^2$, after) 0.9044/0.8506 0.7727/0.7293 0.7336/0.7687 : Total radiation length $X_0$ and thickness $d$ of the material traversed by incident (before interaction) and scattered (after interaction) electrons. The cell is made of glass GE180 which has $X_0=7.04$ cm and density $\rho=2.77$ g/cm$^3$. The radiation length and thickness after interaction are given by left/right depending on by which HRS the electrons were detected.[]{data-label="tab:radlen"} $x$ $\Delta A_\parallel^{ir, ^3{\mathrm}{He}}$ ($\times 10^{-3}$) $\Delta A_\perp^{ir, ^3{\mathrm}{He}}$ ($\times 10^{-3}$) ------ --------------------------------------------------------------- ----------------------------------------------------------- 0.33 -5.77 $\pm$ 0.47 2.66 $\pm$ 0.03 0.48 -3.28 $\pm$ 0.13 1.47 $\pm$ 0.05 0.61 -2.66 $\pm$ 0.15 1.28 $\pm$ 0.07 : Internal radiative corrections to $A_\parallel^{^3{\mathrm}{He}}$ and $A_\perp^{^3{\mathrm}{He}}$. []{data-label="tab:aIntRC"} $x$ $\Delta A_\parallel^{er,^3{\mathrm}{He}}$ ($\times 10^{-3}$) $\Delta A_\perp^{er,^3{\mathrm}{He}}$ ($\times 10^{-3}$) ------ -------------------------------------------------------------- ---------------------------------------------------------- 0.33 -0.67 $\pm$ 0.10 -0.05 $\pm$ 0.11 0.48 -1.16 $\pm$ 0.15 0.80 $\pm$ 0.46 0.61 -0.39 $\pm$ 0.03 0.29 $\pm$ 0.04 : External radiative corrections to $A_\parallel^{^3{\mathrm}{He}}$ and $A_\perp^{^3{\mathrm}{He}}$. Errors are from uncertainties in the structure functions and in the cell wall thickness. []{data-label="tab:aExtRC"} For external corrections the uncertainty also includes the contribution from the uncertainty in the target cell wall thickness. The total radiation length and thickness of the material traversed by the scattered electrons are given in Table \[tab:radlen\] for each kinematic setting. Results for the internal and external radiative corrections are given in Table \[tab:aIntRC\] and \[tab:aExtRC\], respectively. By measuring DIS unpolarized cross sections and using the asymmetry results, one can calculate the polarized cross sections and extract $g_1$ and $g_2$ from Eq. (\[equ:polxsec\_long\]) and (\[equ:polxsec\_tran\]). We used a Monte-Carlo simulation to calculate the expected DIS unpolarized cross sections within the spectrometer acceptance. This simulation included internal and external radiative corrections. The structure functions used in the simulation were from the latest DIS world fits [@ana:r1998; @ana:f2nmc95] with the nuclear effects corrected [@theory:wallyEMC]. The radiative corrections from the elastic and quasi-elastic processes were calculated in the peaking approximation [@ana:peaking] using the world proton and neutron form factor data [@ana:ffpro; @ana:ffneu_dipole; @ana:ffneu_galster]. The DIS cross section results agree with the simulation at a level of $10\%$. Since this is not a dedicated cross section experiment, we obtained the values for $g_1$ and $g_2$ by multiplying our $g_1/F_1$ and $g_2/F_1$ results by the world fits for unpolarized structure functions $F_1$ [@ana:f2nmc95; @ana:r1998], instead of the $F_1$ from this analysis. From $^3$He to Neutron {#ch5:he3model} ---------------------- Properties of protons and neutrons embedded in nuclei are expected to be different from those in free space because of a variety of nuclear effects, including that from spin depolarization, binding and Fermi motion, the off-shell nature of the nucleons, the presence of non-nucleonic degrees of freedom, and nuclear shadowing and antishadowing. A coherent and complete picture of all these effects for the $^3$He structure function $g_1^{^3{\mathrm}{He}}$ in the range of $10^{-4}\leq x\leqslant 0.8$ was presented in [@theory:3Hecmplt]. It gives $$\begin{aligned} \label{equ:he3-g1n} g_1^{^3{\mathrm}{He}} &=& P_ng_1^n+2P_pg_1^p-0.014 \big[{g_1^p(x)-4g_1^n(x)}\big]\nonumber \\ && +a(x)g_1^n(x)+b(x)g_1^p(x)\end{aligned}$$ where $P_n$($P_p$) is the effective polarization of the neutron (proton) inside $^3$He [@theory:PnPp_friar]. Functions $a(x)$ and $b(x)$ are $Q^2$-dependent and represent the nuclear shadowing and antishadowing effects. From Eq.(\[equ:a1=g1f1-simplified\]), the asymmetry $A_1$ is approximately the ratio of the spin structure function $g_1$ and $F_1$. Noting that shadowing and antishadowing are not present in the large $x$ region, using Eq. (\[equ:he3-g1n\]) one obtains $$\begin{aligned} \label{equ:he3ton} A_1^n&=&\frac{F_2^{^3{\mathrm}{He}} \big[{A_1^{^3{\mathrm}{He}}}-2\frac{F_2^p}{F_2^{^3{\mathrm}{He}}} P_pA_1^p(1-\frac{0.014}{2P_p})\big]} {P_nF_2^n(1+\frac{0.056}{P_n})}~. \end{aligned}$$ The two terms $\frac{0.056}{P_n}$ and $\frac{0.014}{2P_p}$ represent the corrections to $A_1^n$ associated with the $\Delta(1232)$ component in the $^3$He wavefunction. Both terms cause $A_1^n$ to increase in the $x$ range of this experiment, and to turn positive at lower values of $x$ compared to the situation when the effect of the $\Delta(1232)$ is ignored. For $F_2^n$ and $F_2^{^3{\mathrm}{He}}$, we used the world proton and deuteron $F_2$ data and took into account the EMC effects [@theory:wallyEMC]. We used the world proton asymmetry data for $A_1^p$. The effective nucleon polarizations $P_{n,p}$ can be calculated using $^3$He wavefunctions constructed from N-N interactions, and their uncertainties were estimated using various nuclear models [@theory:PnPp_nogga; @theory:PnPp_friar; @theory:3Heconv; @theory:PnPp_bissey], giving $$\begin{aligned} && P_n=0.86^{+0.036}_{-0.02}~~{\mathrm}{and}~P_p=-0.028^{+0.009}_{-0.004}~.\label{equ:PnPp}\end{aligned}$$ Eq. (\[equ:he3ton\]) was also used for extracting $A_2^n$, $g_1^n/F_1^n$ and $g_2^n/F_1^n$ from our $^3$He data. The uncertainty in $A_1^n$ due to the uncertainties in $F_2^{p,d}$, in the correction for EMC effects, in $A_1^p$ data and in $P_{n,p}$ is given in Table \[tab:errana\]. Compared to the convolution approach [@theory:3Heconv] used by previous $^3$He experiments [@data:a1ng1n-e142; @data:a1ng1n-e154; @data:a1ng1n-hermes], in which only the first two terms on the right hand side of Eq. (\[equ:he3-g1n\]) are present, the values of $A_1^n$ extracted from Eq. (\[equ:he3ton\]) are larger by $(1-2)\%$ in the region $0.2<x<0.7$.\ Resonance Contributions ----------------------- Since there are a few nucleon resonances with masses above 2 GeV and our measurement at the highest $x$ point has an invariant mass close to $2$ GeV, the effect of possible contributions from baryon resonances were evaluated. This was done by comparing the resonance contribution to $g_1^n$ with that to $F_1^n$. For our kinematics at $x=0.6$, data on the unpolarized structure function $F_2$ and $R$ [@data:f2p-hallc] show that the resonance contribution to $F_1$ is less than $5\%$. The resonance asymmetry was estimated using the MAID model [@theory:maid] and was found to be approximately $0.10$ at $W=1.7$ (GeV). Since the resonance structure is more evident at smaller $W$, we took this value as an upper limit of the contribution at $W=2$ (GeV). The resonance contribution to our $A_1^n$ and $g_1^n/F_1^n$ results at $x=0.6$ were then estimated to be at most $0.008$, which is negligible compared to their statistical errors. [RESULTS]{} {#result:main} =========== $^3$He Results {#result:he3} -------------- Results of the electron asymmetries for $\vec{e}-^3\vec{{\mathrm}{He}}$ scattering, $A_\parallel^{^3{\mathrm}{He}}$ and $A_\perp^{^3{\mathrm}{He}}$, the virtual photon asymmetries $A_{1}^{^3{\mathrm}{He}}$ and $A_{2}^{^3{\mathrm}{He}}$, structure function ratios $g_{1}^{^3{\mathrm}{He}}/F_1^{^3{\mathrm}{He}}$ and $g_{2}^{^3{\mathrm}{He}}/F_1^{^3{\mathrm}{He}}$ and polarized structure functions $g_{1}^{^3{\mathrm}{He}}$ and $g_{2}^{^3{\mathrm}{He}}$ are given in Table \[tab:result\_he3\]. Results for $g_{1,2}^{^3{\mathrm}{He}}$ were obtained by multiplying the $g_{1,2}^{^3{\mathrm}{He}}/F_1^{^3{\mathrm}{He}}$ results by the unpolarized structure function $F_1^{^3{\mathrm}{He}}$, which were calculated using the latest world fits of DIS data [@ana:r1998; @ana:f2nmc95] and with nuclear effects corrected [@theory:wallyEMC]. Results for $A_1^{^3{\mathrm}{He}}$ and $g_1^{^3{\mathrm}{He}}$ are shown in Fig. \[fig:result\_a1he3\] along with SLAC [@data:a1ng1n-e142; @data:a1heg1he-e154] and HERMES [@data:hermes_dqq] data.\ ----------------------------------------------- ----------------------------- ----------------------------- ----------------------------- $\langle x \rangle$ 0.33 0.47 0.60 $\langle Q^2\rangle$ (GeV/c)$^2$ 2.71 3.52 4.83 $A_\parallel^{^3{\mathrm}{He}}$ $-0.020\pm 0.005\pm 0.001$ $-0.012\pm 0.005\pm 0.000$ $ 0.007\pm 0.007\pm 0.001$ $A_\perp^{^3{\mathrm}{He}}$ $ 0.000\pm 0.010\pm 0.000$ $ 0.016\pm 0.008\pm 0.001$ $-0.010\pm 0.016\pm 0.001$ $A_1^{^3{\mathrm}{He}}$ $-0.024\pm 0.006\pm 0.001$ $-0.019\pm 0.006\pm 0.001$ $ 0.010\pm 0.009\pm 0.001$ $A_2^{^3{\mathrm}{He}}$ $-0.004\pm 0.014\pm 0.001$ $ 0.020\pm 0.012\pm 0.001$ $-0.013\pm 0.023\pm 0.001$ $g_1^{^3{\mathrm}{He}}/F_1^{^3{\mathrm}{He}}$ $-0.022\pm 0.005\pm 0.001$ $-0.008\pm 0.008\pm 0.001$ $ 0.003\pm 0.009\pm 0.001$ $g_2^{^3{\mathrm}{He}}/F_1^{^3{\mathrm}{He}}$ $ 0.010\pm 0.036\pm 0.002$ $ 0.050\pm 0.022\pm 0.003$ $-0.028\pm 0.038\pm 0.002$ $g_1^{^3{\mathrm}{He}}$ $-0.024\pm 0.006\pm 0.001 $ $-0.004\pm 0.004\pm 0.000 $ $ 0.001\pm 0.002\pm 0.000 $ $g_2^{^3{\mathrm}{He}}$ $ 0.011\pm 0.039\pm 0.001 $ $ 0.026\pm 0.012\pm 0.002 $ $-0.006\pm 0.009\pm 0.001 $ ----------------------------------------------- ----------------------------- ----------------------------- ----------------------------- Neutron Results {#result:neutron} --------------- Results for the neutron asymmetries $A_1^n$ and $A_2^n$, structure function ratios $g_1^n/F_1^n$ and $g_2^n/F_1^n$ and polarized structure functions $g_1^n$ and $g_2^n$ are given in Table \[tab:result\_neutron\]. ---------------------------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- $\langle x \rangle$ 0.33 0.47 0.60 $\langle Q^2\rangle$ (GeV/c)$^2$ 2.71 3.52 4.83 $A_1^{n}$ $ -0.048\pm 0.024^{+ 0.015}_{-0.016}$ $ -0.006\pm 0.027^{+ 0.019}_{-0.019}$ $ 0.175\pm 0.048^{+ 0.026}_{-0.028}$ $A_2^{n}$ $ -0.004\pm 0.063^{+ 0.005}_{-0.005}$ $ 0.117\pm 0.055^{+ 0.012}_{-0.021}$ $ -0.034\pm 0.124^{+ 0.014}_{-0.014}$ $g_1^n/F_1^n$ $ -0.043\pm 0.022^{+ 0.009}_{-0.009}$ $ 0.040\pm 0.035^{+ 0.011}_{-0.011}$ $ 0.124\pm 0.045^{+ 0.016}_{-0.017}$ $g_2^n/F_1^n$ $ 0.034\pm 0.153^{+ 0.010}_{-0.010}$ $ 0.207\pm 0.103^{+ 0.022}_{-0.021}$ $ -0.190\pm 0.204^{+ 0.027}_{-0.027}$ $g_1^{n}$ $ -0.012\pm 0.006^{+ 0.003}_{-0.003}$ $ 0.005\pm 0.004^{+ 0.001}_{-0.001}$ $ 0.006\pm 0.002^{+ 0.001}_{-0.001}$ $g_2^{n}$ $ 0.009\pm 0.043^{+ 0.003}_{-0.003}$ $ 0.026\pm 0.013^{+ 0.003}_{-0.003}$ $ -0.009\pm 0.009^{+ 0.001}_{-0.001}$ ---------------------------------- ---------------------------------------- ---------------------------------------- ---------------------------------------- The $A_1^n$, $g_1^n/F_1^n$ and $g_1^n$ results are shown in Fig. \[fig:a1nresult\_all\], \[fig:result\_a1n1\] and \[fig:result\_g1n\], respectively. In the region of $x>0.4$, our results have improved the world data precision by about an order of magnitude, and will provide valuable inputs to parton distribution function (PDF) parameterizations. Our data at $x=0.33$ are in good agreement with previous world data. For the $A_1^n$ results, this is the first time that the data show a clear trend that $A_1^n$ turns to positive values at large $x$. As $x$ increases, the agreement between the data and the predictions from the constituent quark models (CQM) becomes better. This is within the expectation since the CQM is more likely to work in the valence quark region. It also indicates that $A_1^n$ will go to higher values at $x>0.6$. However, the trend of the $A_1^n$ results does not agree with the BBS and LSS(BBS) parameterizations, which are from leading-order pQCD analyses based on hadron helicity conservation (HHC). This indicates that there might be problem in the assumption that quarks have zero orbital angular momentum which is used by HHC. The sources for the experimental systematic uncertainties are listed in Table \[tab:expsysErr\]. source error ---------------------------------- ----------------------------------------------------------- Beam energy $E_{b}$ ${\Delta E_b}/{E_b}<5\times 10^{-4}$ HRS central momentum $p_0$ ${\Delta E_e}/{E_e}<5\times 10^{-4}$ [@technote:HRSgamma] HRS central angle $\theta_0$ $\Delta \theta_0<0.1^\circ$ [@exp:hallATN02-032] Beam polarization $P_b$ ${\Delta P_b}/{P_b}<3\%$ Target polarization $P_t$ ${\Delta P_t}/{P_t}<4\%$ Target spin direction $\alpha_t$ $\Delta\alpha_t <1^\circ$ : Experimental systematic errors for the $A_1^n$ result.[]{data-label="tab:expsysErr"} Systematic uncertainties for the $A_1^n$ results include that from experimental systematic errors, uncertainties in internal radiative corrections $\Delta A_1^{n,ir}$ and external radiative corrections $\Delta A_1^{n,er}$ as derived from the values in Tables \[tab:aIntRC\] and \[tab:aExtRC\], and that from nuclear corrections as described in Section \[ch5:he3model\]. Table \[tab:errana\] gives these systematic uncertainties for the $A_1^n$ results along with their statistical uncertainties. The total uncertainties are dominated by the statistical uncertainties.\ $\langle x \rangle$ 0.33 0.47 0.60 --------------------- ---------------------- ---------------------- ---------------------- Statistics 0.024 0.027 0.048 Experimental syst. 0.004 0.003 0.004 $\Delta A_1^{n,ir}$ 0.012 0.013 0.015 $\Delta A_1^{n,er}$ 0.002 0.002 0.003 $F_2^p$, $F_2^d$ 0.006 0.008 $^{+0.005}_{-0.010}$ EMC effect 0.001 0.000 0.009 $A_1^p$ 0.001 0.005 0.011 $P_n$, $P_p$ $^{+0.005}_{-0.012}$ $^{+0.009}_{-0.020}$ $^{+0.018}_{-0.037}$ : Total uncertainties for $A_1^n$. []{data-label="tab:errana"} We used five functional forms, $x^\alpha P_n(x) (1+\beta/Q^2)$, to fit our $g_1^n/F_1^n$ results combined with data from previous experiments [@data:a1pa1n-e143; @data:g1pg1n-e155]. Here $P_n$ is the $n^{th}$-order polynomial, $n=1,2$ for a finite $\alpha$ or $n=1,2,3$ if $\alpha$ is fixed to be $0$. The total number of parameters is limited to $\leqslant 5$. For the $Q^2$-dependence of $g_1/F_1$, we used a term $1+\beta/Q^2$ as in the E155 experimental fit [@data:g1pg1n-e155]. No constraints were imposed on the fit concerning the behavior of $g_1/F_1$ as $x\to 1$. The function which gives the smallest $\chi^2$ value is $g_1^n/F_1^n = (a+bx+cx^2) (1+\beta/Q^2)$. The new fit is shown in Fig. \[fig:result\_a1n1\]. Results for the fit parameters are given in Table \[tab:g1f1nfitpar\] and the covariance error matrix is $${\epsilon} = \left [ \begin{array}{cccc} 1.000 & -0.737 & 0.148 & 0.960 \\ -0.737 & 1.000 & -0.752 & -0.581 \\ 0.148 & -0.752 & 1.000 & -0.039 \\ 0.960 & -0.581 & -0.039 & 1.000 \end{array} \right ]~.$$ --------------------------- $a = -0.049 \pm 0.052$ $b = -0.162 \pm 0.217$ $c = 0.698 \pm 0.345$ $\beta = 0.751 \pm 2.174$ --------------------------- : Result of the fit $g_1^n/F_1^n = (a+bx+cx^2) (1+\beta/Q^2)$.[]{data-label="tab:g1f1nfitpar"} Similar fits were performed to the proton world data [@data:g1p-hermes; @data:a1pa1n-e143; @data:g1pg1n-e155] and function $g_1^p/F_1^p = x^\alpha (a+bx) (1+\beta/Q^2)$ was found to give the smallest $\chi^2$ value. The new fit is shown in Fig. \[fig:a1pmodel\_all\] of Section \[ch2:exp\]. Results for the fit parameters are given in Table \[tab:g1f1pfitpar\] and the covariance error matrix is\ $${\epsilon} = \left [ \begin{array}{cccc} 1.000 & 0.908 & -0.851 & 0.723 \\ 0.908 & 1.000 & -0.967 & 0.401 \\ -0.851 & -0.967 & 1.000 & -0.369 \\ 0.723 & 0.401 & -0.369 & 1.000 \end{array} \right ]~.$$ ---------------------------- $\alpha = 0.813 \pm 0.049$ $a = 1.231 \pm 0.122$ $b = -0.413 \pm 0.216$ $\beta = 0.030 \pm 0.124$ ---------------------------- : Result of the fit $g_1^p/F_1^p = x^\alpha (a+bx) (1+\beta/Q^2)$.[]{data-label="tab:g1f1pfitpar"} Figures \[fig:result\_a2n\] and \[fig:result\_xg2n\] show the results for $A_2^n$ and $xg_2^n$, respectively. The precision of our data is comparable to the data from E155x experiment at SLAC [@data:e155x], which is so far the only experiment dedicated to measuring $g_2$ with published results. To evaluate the matrix element $d_2^n$, we combined our $g_2^n$ results with the E155x data [@data:e155x]. The average $Q^2$ of the E155x data set is about $5$ (GeV/c)$^2$. Following a similar procedure as used in Ref. [@data:e155x], we assumed that $\bar g_2(x,Q^2)$ is independent of $Q^2$ and $\bar g_2\propto(1-x)^m$ with $m=2$ or $3$ for $x\gtorder 0.78$ beyond the measured region of both experiments. We obtained from Eq. (\[equ:d2def\]) $$\begin{aligned} d_2^n &=& 0.0062\pm 0.0028~.\end{aligned}$$ Compared to the value published previously [@data:e155x], the uncertainty on $d_2^n$ has been improved by about a factor of two. The large decrease in uncertainty despite the small number of our data points arises from the $x^2$ weighting of the integral which emphasizes the large $x$ kinematics. The uncertainties on the integrand has been improved in the region $x>0.4$ due to our $g_2^n$ results at the two higher $x$ points being more precise than that of E155x. While a negative value was predicted by lattice QCD [@theory:d2lattice] and most other models [@theory:d2bag; @theory:d2QCDSR; @theory:d2chi], the new result for $d_2^n$ suggests that the higher twist contribution is positive. Flavor Decomposition using the Quark-Parton Model {#result:dqq} ------------------------------------------------- Assuming the strange quark distributions $s(x)$, $\bar{s}(x)$, $\Delta s(x)$ and $\Delta \bar{s}(x)$ to be small in the region $x>0.3$, and ignoring any $Q^2$-dependence of the ratio of structure functions, one can extract polarized quark distribution functions based on the quark-parton model as $$\begin{aligned} {{\frac{\Delta u+\Delta\bar u}{u+\bar u}}} &=& \frac{4 g_1^p(4+{\mathit{R^{du}}})}{15 F_1^p} -\frac{g_1^n(1+4{\mathit{R^{du}}})}{15 F_1^n}~ \label{equ:duu}\end{aligned}$$ and $$\begin{aligned} {{\frac{\Delta d+\Delta\bar d}{d+\bar d}}} &=& \frac{4 g_1^n(1+4{\mathit{R^{du}}})}{15 F_1^n{\mathit{R^{du}}}} -\frac{g_1^p(4+{\mathit{R^{du}}})}{15 F_1^p{\mathit{R^{du}}}} ~, \label{equ:ddd}\end{aligned}$$ with ${\mathit{R^{du}}}\equiv ({d+\bar d})/({u+\bar u})$. Results for $({\Delta u+\Delta\bar u})/(u+$ $\bar u)$ and $({\Delta d+\Delta\bar d})/({d+\bar d})$ are given in Table \[tab:result\_dqq\]. As inputs we used our own results for $g_1^n/F_1^n$, the world data on $g_1^p/F_1^p$ [@thesis:zheng], and the ratio ${\mathit{R^{du}}}$ extracted from proton and deuteron unpolarized structure function data [@theory:duratio]. In a similar manner as for Eq.(\[equ:duu\]) and (\[equ:ddd\]) and ignoring nuclear effects, one can also add the world data on $g_1^{^2{\mathrm}{H}}/F_1^{^2{\mathrm}{H}}$ to the fitted data set and extract these polarized quark distributions. The results are, however, consistent with those given in Table \[tab:result\_dqq\] and have very similar error bars because the data on the deuteron in general have poorer precision than the data on the proton and the neutron data from this experiment. The results presented here have changed compared to the values published previously in Ref. [@A1nPRL] due to an error discovered in our fitting of ${\mathit{R^{du}}}$ from Ref. [@theory:duratio]. The analysis procedure is consistent with what was used in Ref. [@A1nPRL]. $\langle x \rangle$ $({\Delta u+\Delta\bar u})/({u+\bar u})$ $({\Delta d+\Delta\bar d})/({d+\bar d})$ --------------------- -------------------------------------------------- -------------------------------------------------- $0.33$ $ 0.545 \pm 0.004 \pm 0.002 _{-0.025 }^{+0.024}$ $-0.352 \pm 0.035 \pm 0.014 _{-0.031 }^{+0.017}$ $0.47$ $ 0.649 \pm 0.006 \pm 0.002 _{-0.058 }^{+0.058}$ $-0.393 \pm 0.063 \pm 0.020 _{-0.049 }^{+0.041}$ $0.60$ $ 0.728 \pm 0.006 \pm 0.002 _{-0.114 }^{+0.114}$ $-0.440 \pm 0.092 \pm 0.035 _{-0.142 }^{+0.107}$ : Results for the polarized quark distributions. The three uncertainties are those due to the $g_1^n/F_1^n$ statistical error, $g_1^n/F_1^n$ systematic uncertainty and the uncertainties of the $g_1^p/F_1^p$ data, the ${\mathit{R^{du}}}$ fit and the correction for $s$ and $c$ quark contributions.[]{data-label="tab:result_dqq"} Figure \[fig:result\_dqq\] shows our results along with semi-inclusive data on $(\Delta q+\Delta\bar q)/(q+\bar q)$ obtained from recent results for $\Delta q$ and $\Delta \bar q$ [@data:hermes_newdqq] by the HERMES collaboration, and the CTEQ6M unpolarized PDF [@theory:cteq]. To estimate the effect of the $s$ and $\bar s$ contributions, we used two unpolarized PDF sets, CTEQ6M [@theory:cteq] and MRST2001 [@theory:mrst], and three polarized PDF sets, AAC2003 [@theory:aac03_polpdf], BB2002 [@theory:bb_polpdf] and GRSV2000 [@theory:grsv2000_polpdf]. For $c$ and $\bar c$ contributions we used the two unpolarized PDF sets [@theory:cteq; @theory:mrst] and the positivity conditions that $\vert{\Delta c/c}\vert\leqslant~1$ and $\vert{\Delta\bar c/\bar c}\vert\leqslant~1$. To compare with the RCQM predictions, which are given for valence quarks, the difference between $\Delta q_V/q_V$ and $(\Delta q+\Delta\bar q)/(q+\bar q)$ was estimated using the two unpolarized PDF sets [@theory:cteq; @theory:mrst] and the three polarized PDF sets [@theory:aac03_polpdf; @theory:bb_polpdf; @theory:grsv2000_polpdf] and is shown as the shaded band near the horizontal axis of Fig. \[fig:result\_dqq\]. Here $q_V$($\Delta q_V$) is the unpolarized (polarized) valence quark distribution for $u$ or $d$ quark. Results shown in Fig. \[fig:result\_dqq\] agree well with the predictions from the RCQM [@theory:cqm] and the LSS 2001 NLO polarized parton densities [@theory:lss2001]. The results agree reasonably well with the statistical model calculation [@theory:stat]. But results for the $d$ quark do not agree with the predictions from the leading-order pQCD LSS(BBS) parameterization [@theory:lssbbs] assuming hadron helicity conservation. [ CONCLUSIONS]{} ================ We have presented precise data on the neutron spin asymmetry $A_1^n$ and the structure function ratio $g_1^n/F_1^n$ in the deep inelastic region at large $x$ obtained from a polarized $^3$He target. These results will provide valuable inputs to the QCD parameterizations of parton densities. The new data show a clear trend that $A_1^n$ becomes positive at large $x$. Our results for $A_1^n$ agree with the LSS 2001 NLO QCD fit to the previous data and the trend of the $x$-dependence of $A_1^n$ agrees with the hyperfine-perturbed RCQM predictions. Data on the transverse asymmetry and structure function $A_2^n$ and $g_2^n$ were also obtained with a precision comparable to the best previous world data in this kinematic region. Combined with previous world data, the matrix element $d_2^n$ was evaluated and the new value differs from zero by more than two standard deviations. This result suggests that the higher twist contribution is positive. Combined with the world proton data, the polarized quark distributions $(\Delta u+\Delta \bar u)/(u+\bar u)$ and $(\Delta d+\Delta\bar d)/(d+\bar d)$ were extracted based on the quark parton model. While results for $(\Delta u+\Delta \bar u)/(u+\bar u)$ agree well with predictions from various models and fits to the previous data, results for $(\Delta d+\Delta\bar d)/(d+\bar d)$ agree with the predictions from RCQM and from the LSS 2001 fit, but do not agree with leading order pQCD predictions that use hadron helicity conservation. Since hadron helicity conservation is based on the assumption that quarks have negligible orbital angular momentum, the new results suggest that the quark orbital angular momentum, or other effects beyond leading-order pQCD, may play an important role in this kinematic region.\ Formalism for Electron Deep Inelastic Scattering {#app:formalism} ================================================ The fundamental quark and gluon structure of strongly interacting matter is studied primarily through experiments that emphasize hard scattering from the quarks and gluons at sufficiently high energies. One important way of probing the distribution of quarks and antiquarks inside the nucleon is electron scattering, where an electron scatters from a single quark or antiquark inside the target nucleon and transfers a large fraction of its energy and momentum via exchanged photons. In the single photon exchange approximation, the electron interacts with the target nucleon via only one photon, as shown in Fig. \[fig:eNscat\] , and probes the quark structure of the nucleon with a spatial resolution determined by the four momentum transfer squared of the photon $Q^2\equiv -q^2$. Moreover, if a polarized electron beam and a polarized target are used, the spin structure of the nucleon becomes accessible. ![(Color online) Electron scattering in the one-photon exchange approximation.[]{data-label="fig:eNscat"}](fig25.eps){width="120pt"} In the following we denote the incident electron energy by $E$, the energy of the scattered electron by $E^\prime$ thus the energy transfer of the photon is $\nu=E-E^\prime$, and the three-momentum transfer from the electron to the target nucleus by $\vec{q}$. Structure Functions {#ch1:unpolstrf} ------------------- In the case of unpolarized electrons scattering off an unpolarized target, the differential cross-section for detecting the outgoing electron in a solid angle ${\mathrm}{d}\Omega$ and an energy range ($E^\prime$, $E^\prime+{\mathrm}{d}E^\prime$) in the laboratory frame can be written as $$\begin{aligned} \label{equ:dis_xsec_f1f2} \frac{\mathrm{d}^2\sigma}{\mathrm{d}\Omega\mathrm{d}E^\prime} \hspace*{-0.0cm}= &&\hspace*{-0.3cm} \Big(\frac{\mathrm{d}\sigma}{\mathrm{d}\Omega}\Big)_{Mott} \cdot \nonumber\\ &&\Big[\frac{1}{\nu}F_2(x,Q^2) +\frac{2}{M}F_1(x,Q^2)\tan^2\frac{\theta}{2}\Big]~,\end{aligned}$$ where $\theta$ is the scattering angle of the electron in the laboratory frame. The four momentum transfer $Q^2$ is given by $$\begin{aligned} Q^2&=&4EE^\prime\sin^2\frac{\theta}{2} \label{equ:qmu2}~,\end{aligned}$$ and the Mott cross section, $$\begin{aligned} {\Big(\frac{\mathrm{d}\sigma}{\mathrm{d}\Omega}\Big)}_{Mott} &=&\frac{\alpha^2\cos^2{\frac{\theta}{2}}} {4E^2\sin^4{\frac{\theta}{2}}} = \frac{\alpha^2\cos^2{\frac{\theta}{2}}} {Q^4}\frac{E^\prime}{E}~ \label{equ:Mottxsec}\end{aligned}$$ with $\alpha$ the fine structure constant, is the cross section for scattering relativistic electrons from a spin-0 point-like infinitely heavy target. $F_1(x,Q^2)$ and $F_2(x,Q^2)$ are the unpolarized structure functions of the target, which are related to each other as $$\begin{aligned} F_1(x,Q^2) &=& \frac{F_2(x,Q^2)(1+\gamma^2)} {2x\Big(1+R(x,Q^2)\Big)} \label{equ:disF1F2}\end{aligned}$$ with $\gamma^2={(2Mx)^2}/{Q^2}$. Here $R$ is defined as $R\equiv {\sigma_L}/{\sigma_T}$ with $\sigma_L$ and $\sigma_T$ the longitudinal and transverse virtual photon cross sections, which can also be expressed in terms of $F_1$ and $F_2$. Note that for a nuclear target, there exists an alternative [*per nucleon*]{} definition ([*e.g.*]{} as used in Ref. [@ana:f2nmc95]) which is $1/A$ times the definition used in this paper, here $A$ is the number of nucleons inside the target nucleus.\ A review of doubly polarized DIS was given in Ref. [@theory:disreview]. When the incident electrons are longitudinally polarized, the cross section difference between scattering off a target with its nuclear (or nucleon) spins aligned anti-parallel and parallel to the incident electron momentum is $$\begin{aligned} &&\hspace*{-0.5cm} \frac{{\mathrm}{d}^2\sigma_{{\uparrow}{\Downarrow}}}{{\mathrm}{d}\Omega {\mathrm}{d}E^\prime}-\frac{{\mathrm}{d}^2\sigma_{{\uparrow}{\Uparrow}}}{{\mathrm}{d}\Omega {\mathrm}{d}E^\prime} = \frac{4\alpha^2 E^\prime}{\nu EQ^2} \nonumber\\ &&\times\Big [{(E+E^\prime\cos{\theta}){{g_1(x,Q^2)}} -2Mx{{g_2(x,Q^2)}}}\Big ]~. \label{equ:polxsec_long}\end{aligned}$$ where $g_1 (x,Q^2)$ and $g_2(x,Q^2)$ are the polarized structure functions. If the target nucleons are transversely polarized, then the cross section difference is given by $$\begin{aligned} \frac{{\mathrm}{d}^2\sigma_{{\uparrow}{\Rightarrow}}}{{\mathrm}{d}\Omega {\mathrm}{d}E^\prime} -&&\hspace*{-0.4cm}\frac{{\mathrm}{d}^2\sigma_{{\uparrow}{\Leftarrow}}} {{\mathrm}{d}\Omega {\mathrm}{d}E^\prime} = \frac{4\alpha^2 E^{^\prime2}}{\nu EQ^2}\sin{\theta}\nonumber\\ &&\hspace*{-0.cm}\times\Big [{{g_1(x,Q^2)}}+\frac{2E}{\nu} {{g_2(x,Q^2)}}\Big ]~. \label{equ:polxsec_tran}\end{aligned}$$ Bjorken Scaling and Its Violation {#ch1:bjscaling} --------------------------------- A remarkable feature of the structure functions $F_1$, $F_2$, $g_1$ and $g_2$ is their scaling behavior. In the Bjorken limit [@theory:bjorken] ($Q^2\to \infty$ and $\nu\to\infty$ at a fixed value of $x$), the structure functions become independent of $Q^2$ [@data:slac-bjscaling]. Moreover, in this limit $\sigma_L$ vanishes , hence $R=0$ and Eq. (\[equ:disF1F2\]) reduces to $F_2(x)=2xF_1(x)$, known as the Callan-Gross relation [@theory:Callan-Gross]. At finite $Q^2$, the scaling of structure functions is violated due to the radiation of gluons by both initial and scattered quarks. These gluon radiative corrections cause a logarithmic $Q^2$-dependence to the structure functions, which has been verified by experimental data [@exp:pdg] and can be precisely calculated in pQCD using the Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP) evolution equations [@theory:dglap]. From Bjorken Limit to Finite $Q^2$ using the Operator Product Expansion {#ch1:ope} ----------------------------------------------------------------------- In order to calculate observables at finite values of $Q^2$, a method called the Operator Product Expansion (OPE) [@theory:ope] can be applied to DIS which can separate the non-perturbative part of an observable from its perturbative part. In the OPE, whether an operator is perturbative or not is characterized by the “twist” of the operator. At large $Q^2$ the leading twist $t=2$ term dominates, while at small $Q^2$ higher-twist operators need to be taken into account, which are sensitive to interactions beyond the quark-parton model, [*e.g.*]{}, quark-gluon and quark-quark correlations [@theory:ope-g2]. Virtual Photon-Nucleon Asymmetries {#ch1:a1ndef} ---------------------------------- Virtual photon asymmetries are defined in terms of a helicity decomposition of the virtual photon absorption cross sections [@theory:drechsel]. For the absorption of circularly polarized virtual photons with helicity $\pm 1$ by longitudinally polarized nucleons, the longitudinal asymmetry $A_1$ is defined as $$\begin{aligned} A_1(x,Q^2) &\equiv& \frac{\sigma_{{1}/{2}}-\sigma_{{3}/{2}}} {\sigma_{{1}/{2}}+\sigma_{{3}/{2}}} ~,\end{aligned}$$ where $\sigma_{1/2(3/2)}$ is the total virtual photo-absorption cross section for the nucleon with a projection of $1/2(3/2)$ for the total spin along the direction of photon momentum. $A_2$ is a virtual photon asymmetry given by $$\begin{aligned} A_2(x,Q^2) &\equiv& \frac{2\sigma_{LT}}{\sigma_{{1}/{2}}+\sigma_{{3}/{2}}} ~.\end{aligned}$$ where $\sigma_{LT}$ describes the interference between transverse and longitudinal virtual photon-nucleon amplitudes. Because of the positivity limit, $A_2$ is usually small in the DIS region and it has an upper bound given by [@theory:a2soffer] $$\begin{aligned} A_2(x,Q^2)\leqslant \sqrt{\frac{R}{2}\Big[1+A_1(x,Q^2)\Big]} ~. \label{equ:a2soffer}\end{aligned}$$ These two virtual photon asymmetries, depending in general on $x$ and $Q^2$, are related to the nucleon structure functions $g_1(x,Q^2)$, $g_2(x,Q^2)$ and $F_1(x,Q^2)$ via $$\begin{aligned} A_1(x,Q^2) &=&\frac{g_1(x,Q^2)-\gamma^2 g_2(x,Q^2)}{F_1(x,Q^2)} ~ \label{equ:a1=g1f1}\end{aligned}$$ and $$\begin{aligned} A_2(x,Q^2) &=&\frac{\gamma\Big[g_1(x,Q^2)+g_2(x,Q^2)\Big]} {F_1(x,Q^2)}\label{equ:a2=g1f1} ~.\end{aligned}$$ At high $Q^2$, one has $\gamma^2\ll 1$ and $$\begin{aligned} A_1(x,Q^2) &\approx&\frac{g_1(x,Q^2)}{F_1(x,Q^2)} ~.\label{equ:a1=g1f1-simplified}\end{aligned}$$ In QCD the asymmetry $A_1$ is expected to have less $Q^2$-dependence than the structure functions themselves because of the similar leading order $Q^2$-evolution behavior of $g_1(x,Q^2)$ and $F_1(x,Q^2)$. Existing data on the proton and the neutron asymmetries $A_1^p$ and $A_1^n$ indeed show little $Q^2$-dependence [@data:g1pg1n-e155]. Electron Asymmetries {#ch1:easym} -------------------- In an inclusive experiment covering a large range of excitation energies the virtual photon momentum direction changes frequently, and it is usually more practical to align the target spin longitudinally or transversely to the incident electron direction than to the momentum of the virtual photon. The virtual photon asymmetries can be related to the measured electron asymmetries through polarization factors, kinematic variables and the ratio $R$ defined in Section \[ch1:unpolstrf\]. The longitudinal electron asymmetry is defined by [@exp:e080] $$\begin{aligned} A_\parallel &\equiv& \frac{\sigma_{\downarrow\Uparrow} -\sigma_{\uparrow\Uparrow}} {\sigma_{\downarrow\Uparrow}+\sigma_{\uparrow\Uparrow}} \nonumber\\ &=& \frac{(1-\epsilon)M^3}{(1-\epsilon R)\nu F_1} \Big[(E+E^\prime\cos\theta)g_1-\frac{Q^2}{\nu} g_2\Big] ~, \label{equ:Apardef} \end{aligned}$$ where $\sigma_{{\downarrow}{\Uparrow}}$($\sigma_{{\uparrow}{\Uparrow}}$) is the cross section of scattering off a longitudinally polarized target, with the incident electron spin aligned anti-parallel (parallel) to the target spin, and $\epsilon$ is the magnitude of the virtual photon’s longitudinal polarization: $$\begin{aligned} \epsilon &=& \Big[1+2(1+1/\gamma^2)\tan^2(\theta/2) \Big]^{-1} ~. \label{equ:def-epsilon}\end{aligned}$$ Similarly the transverse electron asymmetry is defined for a target polarized perpendicular to the beam direction as $$\begin{aligned} A_\perp &\equiv& \frac{\sigma_{\downarrow\Rightarrow} -\sigma_{\uparrow\Rightarrow}} {\sigma_{\downarrow\Rightarrow}+\sigma_{\uparrow\Rightarrow}} \nonumber\\ &=& \frac{(1-\epsilon)E^\prime M^3}{(1-\epsilon R)\nu F_1} \Big[g_1+\frac{2E}{\nu}g_2\Big]\cos\theta ~,~ \label{equ:Aperpdef}\end{aligned}$$ where $\sigma_{{\downarrow}{\Rightarrow}}$($\sigma_{{\uparrow}{\Rightarrow}}$) is the cross section for scattering off a transversely polarized target with incident electron spin aligned anti-parallel (parallel) to the beam direction, and the scattered electrons being detected on the same side of the beam as that to which the target spin is pointing [@A1nPRL]. The electron asymmetries can be written in terms of $A_1$ and $A_2$ as $$\begin{aligned} A_\parallel &=& D(A_1+\eta A_2) \label{equ:apap1}\end{aligned}$$ and $$\begin{aligned} A_\perp &=& d(A_2-\xi A_1) ~, \label{equ:apap2}\end{aligned}$$ where the virtual photon polarization factor is given by $$\begin{aligned} D&=&\frac{1-(1-y)\epsilon}{1+\epsilon R}~,\end{aligned}$$ with $y\equiv\nu/E$ the fractional energy loss of the incident electron. The remaining kinematic variables are given by $$\begin{aligned} \eta &=& (\epsilon\sqrt{Q^2})/(E-E^\prime\epsilon)~,\\ \xi &=& \eta(1+\epsilon)/(2\epsilon)~~~~{\mathrm}{and}\\ d &=& D\sqrt{2\epsilon/(1+\epsilon)} ~.\end{aligned}$$ Extracting Polarized Structure Functions from Asymmetries {#ch1:asym2strf} --------------------------------------------------------- From Eq. (\[equ:apap1\]) and (\[equ:apap2\]) the virtual photon asymmetries $A_1$ and $A_2$ can be extracted from measured electron asymmetries as $$\begin{aligned} A_1 &=& \frac{1}{D(1+\eta\xi)}A_{\parallel}-\frac{\eta} {d(1+\eta\xi)}A_{\perp} \label{equ:a2a1}\end{aligned}$$ and $$\begin{aligned} A_2 &=& \frac{\xi}{D(1+\eta\xi)}A_{\parallel} +\frac{1}{d(1+\eta\xi)}A_{\perp} ~.\label{equ:a2a2}\end{aligned}$$ If the unpolarized structure functions $F_1(x,Q^2)$ and $R(x,Q^2)$ are known, then the polarized structure functions can be extracted from measured asymmetries $A_\parallel$ and $A_\perp$ as [@theory:disreview] $$\begin{aligned} g_1(x,Q^2) &=& \frac{F_1(x,Q^2)}{D^\prime} \Big[{A_\parallel+A_\perp\tan(\theta/2)}\Big] \label{equ:a2g1}\end{aligned}$$ and $$\begin{aligned} g_2(x,Q^2) &=& \frac{yF_1(x,Q^2)}{2D^\prime\sin\theta}\nonumber\\ &&\times \Big[{A_\perp\frac{E+E^\prime\cos\theta} {E^\prime}-A_\parallel\sin\theta }\Big] ~, \label{equ:a2g2}\end{aligned}$$ with $D^\prime$ given by $$\begin{aligned} D^\prime&=& \frac{(1-\epsilon)(2-y)}{y\big[1+\epsilon R(x,Q^2)\big]} ~.\end{aligned}$$ Formalism for e-$^3$He Elastic Scattering {#app:el} ========================================= The cross section for electron elastic scattering off an unpolarized $^3$He target can be written as $$\begin{aligned} && \Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)^{u} = \frac{\sigma_{Mott}}{1-\tau} \Bigl\{\frac{Q^2}{\vert\vec{q}\vert^2}F_c^2(Q) \nonumber\\ &&~~~~~+\frac{\mu_{^{3}{\mathrm}{He}}^2 Q^2}{2M^2}\big[\frac{1}{2}\frac{Q^2} {\vert\vec{q}\vert^2}-\tan^2{(\theta/2)}\big]F_m^2(Q)\Bigr\} \label{equ:elxsec4}\end{aligned}$$ where $\tau\equiv Q^2/(4M_t^2)=\nu/(2M_t)$ is the recoil factor, $M_t$ is the target ($^3$He) mass, $Q^2$ is calculated from Eq. (\[equ:qmu2\]), $\vec q$ is the three momentum transfer, $\mu_{^{3}{\mathrm}{He}}$ is the $^3$He magnetic moment, and $F_c$ and $F_m$ are the $^3$He charge and magnetic form factors, which have been measured to a good precision [@ana:he3ff]. The Mott cross section $\sigma_{Mott}$ for a target of charge $Z$ can be written as $$\begin{aligned} \sigma_{Mott}&\equiv&\Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)_{Mott} = \frac{Z^2\alpha^2\cos^2(\theta/2)}{4E^2\sin^4(\theta/2)}\frac{E^\prime}{E}~.\end{aligned}$$ with $E^\prime$ the energy of the outgoing electrons: $$\begin{aligned} E^\prime &=& \frac{E}{1+\frac{2E}{M_T}\sin^2\frac{\theta}{2}}~.\end{aligned}$$ The elastic cross section for a polarized target can be written as [@theory:he3elasym] $$\begin{aligned} \Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)^h &=& \Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)^{u} + h\Delta(\theta^*,\phi^*,E,\theta,Q^2)~,~~\label{equ:elpolxsec}\end{aligned}$$ where $h$ is the helicity of the incident electron beam, $\Delta(\theta^*,\phi^*,E,\theta,Q^2)$ describes the helicity-dependent cross section, $\theta^*$ is the polar angle and $\phi^*$ is the azimuthal angle of the nucleon spin ![(Color online) Polar and azimuthal angles of the target spin.[]{data-label="fig:elkine"}](fig26.eps) direction, as shown in Fig. \[fig:elkine\]. We write them explicitly for a target with spin parallel to the beam direction as [@theory:he3elasym] $$\begin{aligned} \cos{\theta^*} &=& (E-E^\prime\cos{\theta})/\vert\vec q\vert \\ \phi^* &=& 0~.\end{aligned}$$ The helicity-dependent part of the cross section can be written as $$\begin{aligned} \Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)^{h=+1} \hspace*{-0.3cm}-\Big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega}\Big)^{h=-1} \hspace*{-0.3cm} &=& - \sigma_{Mott}\Big( V_{T^\prime}R_{T^\prime}(Q^2)\cos\theta^* \nonumber \\ && \hspace*{-1.4cm} + V_{TL^\prime}R_{TL^\prime}(Q^2)\sin\theta^*\cos\phi^* \Big)~, \label{equ:Apar-el-dxsec}\end{aligned}$$ with kinematic factors $$\begin{aligned} V_{T^\prime} &\equiv& \tan{\frac{\theta}{2}}\sqrt{\frac{Q^2} {\vert\vec{q}\vert^2}+\tan^2{\frac{\theta}{2}}}\end{aligned}$$ and $$\begin{aligned} V_{TL^\prime}&\equiv& -\frac{Q^2}{\sqrt{2}\vert\vec{q}\vert^2} \tan{\frac{\theta}{2}}~.\end{aligned}$$ $R_{T^\prime}$, $R_{TL^\prime}$ can be related to the $^3$He form factors $F_c$, $F_m$ as: $$\begin{aligned} R_{T^\prime} &=& \frac{2\tau E^\prime}{E}(\mu_{^3{\mathrm}{He}} F_m)^2\end{aligned}$$ and $$\begin{aligned} R_{TL^\prime} &=& -\frac{2\sqrt{2\tau(1+\tau)}E^\prime}{E}(ZF_c)(\mu_{^3{\mathrm}{He}} F_m)~,\end{aligned}$$ The elastic asymmetry, defined by $$\begin{aligned} \label{equ:he3elasymdef} && A^{el}_{\parallel} \equiv \frac{\big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega{\mathrm}{d}E^\prime}\big)^{h=+1} -\big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega{\mathrm}{d}E^\prime}\big)^{h=-1}} {\big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega{\mathrm}{d}E^\prime}\big)^{h=+1} +\big(\frac{{\mathrm}{d}\sigma}{{\mathrm}{d}\Omega{\mathrm}{d}E^\prime}\big)^{h=-1}}~,\end{aligned}$$ can therefore be calculated from Eq. (\[equ:elxsec4\]) and (\[equ:Apar-el-dxsec\]) as $$\begin{aligned} \label{equ:he3elasymcal} && A^{el}_{\parallel} = -(1-\tau) \nonumber\\ && \hspace*{0.2cm}\cdot\frac{\big({V_{T^\prime}R_{T^\prime}\cos\theta^* +V_{TL^\prime}R_{TL^\prime}\sin\theta^*\cos\phi^*}\big)} {\Bigl\{\frac{Q^2}{\vert\vec{q}\vert^2}F_c^2 +\frac{\mu^2 Q^2}{2M^2}\Big(\frac{1}{2}\frac{Q^2}{\vert\vec{q}\vert^2} -\tan^2{\frac{\theta}{2}}\Big)F_m^2\Bigr\}}~~~\end{aligned}$$ **[ACKNOWLEDGMENTS]{}** We would like to thank the personnel of Jefferson Lab for their efforts which resulted in the successful completion of the experiment. We thank S. J. Brodsky, L. Gamberg, N. Isgur, X. Ji, E. Leader, W. Melnitchouk, D. Stamenov, J. Soffer, M. Strikman, A. Thomas, M. Wakamatsu, H. Weigel and their collaborators for theoretical support and helpful discussions. This work was supported by the Department of Energy (DOE), the National Science Foundation, the Italian Istituto Nazionale di Fisica Nucleare, the French Institut National de Physique Nucléaire et de Physique des Particules, the French Commissariat à l’Énergie Atomique and the Jeffress Memorial Trust. The Southeastern Universities Research Association operates the Thomas Jefferson National Accelerator Facility for the DOE under contract DE-AC05-84ER40150.\ J. Ashman [*et al.*]{}, [Phys. Lett. B]{} [**206**]{}, 364 (1988); J. Ashman [*et al.*]{}, [Nucl. Phys. B]{} [**328**]{}, 1 (1989). M.J. Alguard [*et al.*]{}, [Phys. Rev. Lett.]{} [**41**]{}, 70 (1978); G. Baum [*et al.*]{}, [Phys. Rev. Lett.]{} [**51**]{}, 1135 (1983). N.N. Bogoliubov, [Ann. Inst. Henri Poincar$\acute{e}$]{} [**8**]{}, 163 (1968); F.E. Close, [Nucl. Phys. B]{} [**80**]{}, 269 (1974); A. LeYaouanc [*et al.*]{}, [Phys. Rev. D]{} [**9**]{}, 2636 (1974); [**15**]{}, 844 (1977); A. Chodos, R.L. Jaffe, K. Johnson, and C.B. Thorn, [*ibid.*]{} [**10**]{}, 2599 (1974); M.J. Ruiz, [*ibid.*]{} [**12**]{}, 2922 (1975); C. Hayne and N. Isgur, [Phys. Rev. D]{} [**25**]{}, 1944 (1982). Z. Dziembowski, C.J. Martoff and P. Zyla, [Phys. Rev. D]{} [**50**]{}, 5613 (1994). B.-Q. Ma, [Phys. Lett. B]{} [**375**]{}, 320 (1996). B.W. Filippone and X. Ji, [Adv. Nucl. Phys.]{} [**26**]{}, 1 (2001). X. Ji, [Phys. Rev. Lett.]{} [**78**]{}, 610 (1997). R.P. Feynman, [Phys. Rev. Lett.]{} [**23**]{}, 1415 (1969). A.W. Thomas and W. Weise, [*The Structure of the Nucleon*]{}, p.78 and p.100, p.105-106, Wiley-Vch, Germany (2001). K. Wilson, [Phys. Rev.]{} [**179**]{}, 1499 (1969). S. Wandzura and F. Wilczek, [Phys. Lett. B]{}, [**72**]{}, 195 (1977). R.L. Jaffe and X. Ji, [Phys. Rev. D]{}, [**43**]{}, 724 (1991). X. Ji and J. Osborne, Nucl. Phys. B [**608**]{}, 235 (2001). M. Göckeler [*et al.*]{}, Phys. Rev. D [**63**]{}, 074506 (2001). R.L. Jaffe and X. Ji, Phys. Rev. D [**43**]{}, 724 (1991); F.M. Steffens, H. Holtmann and A.W. Thomas, Phys. Lett. B [**358**]{}, 139 (1995); X. Song, Phys. Rev. D [**54**]{}, 1955 (1996). E. Stein [*et al.*]{}, Phys. Lett. B [**343**]{}, 369 (1995); I. Balitsky, V. Braun and A. Kolesnichenko, [*ibid.*]{} [**242**]{}, 245 (1990); [**318**]{}, 648 (1993) (Erratum). H. Weigel, L. Gamberg and H. Reinhardt, Nucl. Phys. A [**680**]{}, 48c (2001); Phys. Rev. D [**55**]{}, 6910 (1997); M. Wakamatsu, Phys. Lett. B [**487**]{}, 118 (2000). X. Zheng [*et al.*]{}, [Phys. Rev. Lett.]{}, [**92**]{}, 012004 (2004). G. Morpurgo, [*Physics*]{} [**2**]{}, 95 (1965); R.H. Dalitz, in [*High Energy Physics; Lectures Delivered During the 1965 Session of the Summer School of Theoretical Physics, University of Grenoble*]{}, edited by C. De Witt and M. Jocab, Gordon and Breach, New York, 1965; and in [*Proceedings of the Oxford International Conference on Elementary Particles*]{}, Oxford, England, 1965; F. Close, [Nucl. Phys. B]{} [**80**]{}, 269 (1974); in [*An introduction to Quarks and Partons*]{}, p.197, Academic Press, New York (1979). A. Bodek [*et al.*]{}, [Phys. Rev. Lett.]{} [**30**]{}, 1087 (1973); E.M. Riordan [*et al.*]{}, [*ibid.*]{} [**33**]{}, 561 (1974); J.S. Poucher [*et al.*]{}, [*ibid.*]{} [**32**]{}, 118 (1974). A. Benvenuti [*et al.*]{}, [Phys. Lett. B]{} [**237**]{}, 599 (1990). J.J. Aubert [*et al.*]{}, [Nucl. Phys. B]{} [**293**]{}, 740 (1987). D. Allasia [*et al*]{}, [Phys. Lett. B]{} [**249**]{}, 366 (1990); P. Amaudruz [*et al.*]{}, [Phys. Rev. Lett.]{} [**66**]{}, 2712 (1991); P. Amaudruz [*et al.*]{}, [Nucl. Phys. B]{} [**371**]{}, 3 (1992). M.R. Adams [*et al*]{}, [Phys. Rev. Lett.]{} [**75**]{}, 1466 (1995). J. Ashman [*et al.*]{}, [Phys. Lett. B]{} [**206**]{}, 364 (1988); J. Ashman [*et al.*]{}, [Nucl. Phys. B]{} [**328**]{}, 1 (1989). B. Adeva [*et al.*]{}, [Phys. Rev. D]{} [**60**]{}, 072004 (1999). K. Abe [*et al.*]{}, [Phys. Rev. D]{} [**58**]{}, 112003 (1998). F.E. Close, [Phys. Lett. B]{} [**43**]{}, 422 (1973); R. Carlitz, [*ibid.*]{} [**58**]{}, 345 (1975). A. De Rujula, H. Georgi, S.L. Glashow, [Phys. Rev. D]{} [**12**]{}, 147 (1975). N. Isgur, [Phys. Rev. D]{} [**59**]{}, 034013 (1999). G.R. Farrar and D.R. Jackson, [Phys. Rev. Lett.]{} [**35**]{}, 1416 (1975). S.J. Brodsky, M. Burkardt and I. Schmidt, [Nucl. Phys. B]{} [**441**]{}, 197 (1995). E. Leader, A.V. Sidorov and D.B. Stamenov, [Int. J. Mod. Phys. A]{} [**13**]{}, 5573 (1998). D. Abbott [*et al.*]{}, [Phys. Rev. Lett.]{} [**84**]{}, 5053 (2000). K. Wijesooriya [*et al.*]{}, [Phys. Rev. C]{} [**66**]{}, 034614 (2002). M.K. Jones [*et al.*]{}, [Phys. Rev. Lett.]{} [**84**]{}, 1398 (2000). O. Gayou [*et al.*]{}, [Phys. Rev. Lett.]{} [**88**]{}, 092301 (2002). G.P. Lepage and S.J. Brodsky, [Phys. Rev. D]{} [**22**]{}, 2157 (1981). G.A. Miller and M.R. Frank, [Phys. Rev. C]{} [**65**]{}, 065205 (2002); R.V. Buniy, P. Jain and J.P. Ralston, [AIP Conf. Proc.]{} [**549**]{}, 302 (2000). A.V. Belitsky, X. Ji and Feng Yuan, [Phys. Rev. Lett.]{} [**91**]{}, 092003 (2003). X. Ji, J.-P. Ma and F. Yuan, [Nucl. Phys. B]{} [**652**]{}, 383 (2003). E. Leader, A.V. Sidorov and D.B. Stamenov, [Eur. Phys. J. C]{} [**23**]{}, 479 (2002); E. Leader, D.B. Stamenov, [*priv. comm.*]{} C. Bourrely, J. Soffer and F. Buccella, [Eur. Phys. J. C]{} [**23**]{}, 487 (2002). C. Bourrely, J. Soffer, [*priv. comm.*]{} H. Weigel, L. Gamberg and H. Reinhardt, [Phys. Lett. B]{} [**399**]{}, 287 (1997); [Phys. Rev. D]{} [**55**]{}, 6910 (1997); O. Schröder, H. Reinhardt and H. Weigel, [Nucl. Phys. A]{} [**651**]{}, 174 (1999). M. Wakamatsu, [Phys. Rev. D]{} [**67**]{}, 034005 (2003); [*ibid.*]{} [**67**]{}, 034006 (2003). N.I. Kochelev, [Phys. Rev. D]{} [**57**]{}, 5539 (1998). E.D. Bloom and F.J. Gilman, [Phys. Rev. Lett.]{} [**16**]{}, 1140 (1970); [Phys. Rev. D]{} [**4**]{}, 2901 (1971); F.E. Close and W. Melnitchouk, [Phys. Rev. C]{} [**68**]{}, 035210 (2003); [*priv. comm.*]{} C. Boros and A.W. Thomas, [Phys. Rev. D]{} [**60**]{}, 074017 (1999); F.M. Steffens and A.W. Thomas, [*priv. comm.*]{} D. Adams [*et al.*]{}, [Phys. Lett. B]{} [**357**]{}, 248 (1995). K. Ackerstaff [*et al.*]{}, [Phys. Lett. B]{} [**404**]{}, 383 (1997). P.L. Anthony [*et al.*]{}, [Phys. Rev. D]{} [**54**]{}, 6620 (1996). K. Abe [*et al.*]{}, [Phys. Rev. Lett.]{} [**79**]{}, 26 (1997); [Phys. Lett. B]{} [**405**]{}, 180 (1997). P.L. Anthony [*et al.*]{}, [Phys. Lett. B]{} [**493**]{}, 19 (2000). A. Airapetian [*et al.*]{}, [Phys. Lett. B]{} [**442**]{}, 484 (1998). JLAB E99117, J.-P. Chen, Z.-E. Meziani, P. Souder [*et al.*]{},\ http://hallaweb.jlab.org/physics/experiments/he3/A1n/ A. Nogga, [Ph. D. thesis]{}, Ruhr-Universität Bochum, Bochum, Germany (2001), p.72. J.L. Friar [*et al.*]{}, [Phys. Rev. C]{} [**42**]{}, 2310 (1990). X. Zheng, [Ph. D. thesis]{}, M.I.T., Cambridge, Massachusetts (2002). R. Prepost and T. Maruyama, [Ann. Rev. of Nucl. and Part. Sci.]{} [**45**]{}, 41 (1995). J. Alcorn [*et al.*]{}, [Nucl. Inst. Meth. A]{} [**522**]{}, 294 (2004). J. Berthot and P. Vernin, [Nucl. Phys. News]{} [**9**]{}, 12 (1990). D. Marchand, [Ph. D. thesis]{}, Université Blaise Pascal, Clermont-Ferrand, France (1997). O. Ravel, [Ph. D. thesis]{}, Université Blaise Pascal, Clermont-Ferrand, France (1997). M. Baylac, [Ph.!D. thesis]{}, Universite Claude Bernard Lyon I, France (2000). G.W. Miller IV, [Ph. D. thesis]{}, Princeton University, Princeton, New Jersey (2001). K.G. Fissum [*et al.*]{}, [Nucl. Instrum. Meth. A]{} [**474**]{}, 108 (2001). C. Grupen, [*Particle Detectors*]{}, p.193, Cambridge Univ., (1996). W.A. Watson [*et al.*]{}, [*Real-time Computer Applications in Nuclear, Particle and Plasma Physics*]{}, p.296-303, Vancouver, (1993). E.A.J.M. Offerman, ESPACE User’s Guide,\ http://www.hallaweb.jlab.org/espace/docs.html F.D. Colegrove, L.D. Shearer and G.K. Walters, [Phys. Rev.]{} [**132**]{}, 2561 (1963). W. Happer, [Rev. Mod. Phys. ]{} [**44**]{}, 169 (1972); S. Appelt [*et al.*]{}, [Phys. Rev. A]{} [**58**]{}, 1412 (1998). T.E. Chupp [*et al.*]{}, [Phys. Rev. C]{} [**36**]{}, 2244 (1987). JLab E94-010, G.D. Cates, J.P. Chen, Z.-E. Meziani [*et al.*]{},\ see http://www.jlab.org/e94010/ for technical notes; M. Amarian [*et al.*]{}, [Phys. Rev. Lett.]{} [**89**]{}, 242301 (2002); M. Amarian [*et al.*]{}, [*ibid.*]{} [**92**]{}, 022301 (2004). M.V. Romalis, [Ph. D. thesis]{}, Princeton University, Princeton, New Jersey (1998). K. Kramer, [Ph. D. thesis]{}, College of William and Mary, Williamsburg, Virginia (2003). M.V. Romalis and G.D. Cates, [Phys. Rev. A]{} [**58**]{}, 3004 (1998). A. Abragam, [*Principles of Nuclear Magnetism*]{}, Oxford Univ., (1961). A. Deur, [Ph. D. thesis]{}, Université Blaise Pascal, Clermont-Ferrand, France (2000). I. Akushevich [*et al.*]{}, [Comp. Phys. Comm.]{} [**104**]{}, 201 (1997). L. Mo and Y.S. Tsai, [Rev. Mod. Phys.]{} [**41**]{}, 205 (1969). S. Stein [*et al.*]{}, [Phys. Rev. D]{} [**12**]{}, 1884 (1975); F.W. Brasse [*et al.*]{}, [Nucl. Phys. B]{} [**110**]{}, 413 (1976); P. Amaudruz [*et al.*]{}, [Phys. Lett. B]{} [**295**]{}, 159 (1992). P. Amaudruz [*et al.*]{}, [Nucl. Phys. B]{} [**371**]{}, 3 (1992). M. Arneodo [*et al.*]{}, [Phys. Lett. B]{} [**364**]{}, 107 (1995). C. Keppel, [Ph. D. thesis]{}, American University, Washington DC (1994). C. Keppel, Y. Liang, [*pri. comm.*]{} L.W. Whitlow [*et al.*]{}, [Phys. Lett. B]{} [**250**]{}, 193 (1990). K. Abe [*et al.*]{}, [Phys. Lett. B]{} [**452**]{}, 194 (1999). A. Schaefer, [Phys. Lett. B]{} [**208**]{}, 175 (1988). M. Gluck, E. Reya, M. Stratmann and W. Vogelsang, [Phys. Rev. D]{} [**53**]{}, 4775 (1996). W. Melnitchouk and A. W. Thomas, [Acta Phys. Polon. B]{} [**27**]{}, 1407 (1996); W. Melnitchouk, [*pri. comm.*]{} Y.S. Tsai, in [*Proceedings of the International Conference on Nuclear Structure*]{}, p.221, Stanford Univ., 1964. S.I. Bilenkaya, Y.M. Kazarinov and L.I. Lapidus, [Pisma ZhETF]{}, [**61**]{}, 2225 (1971); [Sov. Phys. JETP]{} [**34**]{}, 1192 (1972); [Pisma ZhETF]{}, [**19**]{}, 613 (1974). R. Arnold [*et al.*]{}, [Phys. Rev. Lett.]{} [**57**]{}, 174 (1986). S. Galster [*et al.*]{}, [Nucl. Phys. B]{} [**32**]{}, 221 (1971). I. Niculescu [*et al.*]{}, [Phys. Rev. Lett.]{} [**85**]{}, 1186 (2000). MAID 2003, D. Drechsel, S.S. Kamalov and L. Tiator, [Nucl. Phys. A]{} [**645**]{}, 145 (1999); http://www.kph.uni-mainz.de/MAID/maid2003/maid2003.html F. Bissey [*et al.*]{}, [Phys. Rev. C]{} [**65**]{}, 064317 (2002). C. Ciofi degli Atti [*et al.*]{}, [Phys. Rev. C]{} [**48**]{}, R968 (1993). F. Bissey [*et al.*]{}, [Phys. Rev. C]{} [**64**]{}, 024004 (2001); Y. Kolomensky, [*priv. comm.*]{} K. Ackerstaff, [*et al*]{}, [Phys. Lett. B]{} [**464**]{}, 123 (1999). P.L. Anthony [*et al.*]{}, [Phys. Lett. B]{} [**553**]{}, 18 (2003). N. Liyanage, JLAB Hall A Technical Note JLAB-TN-01-049 (2001), (unpublished). H. Ibrahim, P. Ulmer and N. Liyanage, [JLAB Hall A Technical Note JLAB-TN-02-032]{} (2002), (unpublished). W. Melnitchouk, A. W. Thomas, [Phys. Lett. B]{} [**377**]{}, 11 (1996). A. Airapetian [*et al.*]{}, [Phys. Rev. Lett]{} [**92**]{}, 012005 (2004); M. Beckmann, D. Ryckbosch and E.C. Aschenauer, [*priv. comm.*]{} J. Pumplin [*et al.*]{}, [J. High Energy Phys.]{} [**7**]{}, 012 (2002). A.D. Martin, R.G. Roberts, W.J. Stirling and R.S. Thorne, [Eur. Phys. J. C]{} [**23**]{}, 73 (2002). Y. Goto [*et al.*]{}, [Phys. Rev. D]{} [**62**]{}, 034017 (2000). J. Bluemlein and H. Boettcher, [Nucl. Phys. B]{} [**636**]{}, 225 (2002). M. Glueck, E. Reya, M. Stratmann and W. Vogelsang, [Phys. Rev. D]{} [**63**]{}, 094005 (2001). M. Anselmino, A. Efremov and E. Leader, [Phys. Rept.]{} [**261**]{}, 1 (1995); [Erratum-ibid.]{} [**281**]{}, 399 (1997). J.D. Bjorken and E.A. Paschos, [Phys. Rev.]{} [**185**]{}, 1975 (1969). H.W. Kendall, [Rev. Mod. Phys.]{} [**63**]{}, 597 (1991). C.G. Callan, Jr. and D.J. Gross, [Phys. Rev. Lett.]{} [**22**]{}, 156 (1969). B. Foster, A.D. Martin and M.G. Vincter, [Phys. Rev. D.]{} [**66**]{}, 010001 (2002). Y. Dokshitzer, [Sov. Phys. JETP]{} [**46**]{}, 1649 (1977); V.N. Gribov and L.N. Lipatov, [Sov. Nucl. Phys.]{} [**15**]{}, 438 and 675 (1972); G. Altarelli, G. Parisi, [Nucl. Phys. B]{} [**126**]{}, 298 (1977). D. Drechsel, S.S. Kamalov and L. Tiator, [Phys. Rev. D]{}, [**63**]{}, 114010 (2001). J. Soffer, and O.V. Teryaev, [Phys. Lett. B]{} [**490**]{}, 106 (2000). P.S. Cooper [*et al.*]{}, [Phys. Rev. Lett.]{} [**34**]{}, 1589 (1975). A. Amroun [*et al.*]{}, [Nucl. Phys. A]{} [**579**]{}, 596 (1994). T.W. Donnelly and A.S. Raskin, [Ann. Phys. (N.Y.)]{} [**169**]{}, 247 (1986).

--- abstract: 'We reexamine the dynamical evolution of the mass and luminosity functions of globular star clusters (GCMF and GCLF). Fall & Zhang (2001, hereafter FZ01) showed that a power-law MF, as commonly seen among young cluster systems, would evolve by dynamical processes over a Hubble time into a peaked MF with a shape very similar to the observed GCMF in the Milky Way and other galaxies. To simplify the calculations, the semi-analytical FZ01 model adopted the “classical” theory of stellar escape from clusters, and neglected variations in the ${{M}}/L$ ratios of clusters. Kruijssen & Portegies Zwart (2009, hereafter KPZ09) modified the FZ01 model to include “retarded” and mass-dependent stellar escape, the latter causing significant ${{M}}/L$ variations. KPZ09 asserted that their model was compatible with observations whereas the FZ01 model was not. We show here that this claim is not correct; the FZ01 and KPZ09 models fit the observed Galactic GCLF equally well. We also show that there is no detectable correlation between ${{M}}/L$ and $L$ for GCs in the Milky Way and Andromeda galaxies, in contradiction with the KPZ09 model. Our comparisons of the FZ01 and KPZ09 models with observations can be explained most simply if stars escape at rates approaching the classical limit for high-mass clusters, as expected on theoretical grounds.' author: - 'Paul Goudfrooij and S. Michael Fall' title: Evolution of the Mass and Luminosity Functions of Globular Star Clusters --- Introduction {#s:intro} ============ One of the most remarkable properties of globular cluster (GC) systems is the similarity of their luminosity functions from one galaxy to another. These have bell-like shapes and are often modeled as log-normal distributions of luminosities or, equivalently, Gaussian distributions of magnitudes (see, e.g., @harr91). In contrast, the luminosity functions of young cluster systems are always found to be power laws, $\phi(L) = dN/dL \propto L^{\alpha}$ with $\alpha \approx -2$ (@vdblaf84 [@elsfal85; @chrsch88]; see @whit+14 for a recent comprehensive study of 20 star-forming galaxies). The mass functions of cluster systems have greater dynamical significance than their luminosity functions. For systems of old, coeval clusters, variations in the mass-to-light (${{M}}/L$) ratios are relatively small, and the luminosity function is a good proxy for the mass function. Thus, the mass function of globular clusters (GCMF) must have nearly the same bell-like shape as the luminosity function (GCLF). However, systems of young clusters in the process of formation have wide spreads in ${{M}}/L$, and the mass function must be determined separately from the luminosity function, usually by estimating the masses, ages, and reddenings of individual clusters in the sample from multi-band photometry. (See @fall06 for the mathematical relations between the mass, luminosity, and age distributions of cluster systems.) Studies of this kind show that the mass functions of young cluster systems are power laws, $\psi(M) = dN/dM \propto M^{\beta}$ with $\beta \approx -2$ (@zhafal99 [@falcha12], and references therein). As a consequence, larger systems of young clusters will contain more massive clusters, reaching $M \sim 10^6 \;M_{\odot}$ or even $\sim 10^7 \; M_{\odot}$ in some cases [@chan+10]. The most massive of these clusters are often referred to as “young GCs”. There are two possible explanations for the radically different mass functions of young and old cluster systems: either (1) the process of cluster formation, and hence the initial cluster mass function, differed in the distant past from the present, or (2) the mass function of clusters evolves by dynamical processes from an initial power law into a bell-shaped distribution. The second possibility—the one we examine in this paper—has been explored at various levels of approximation over the years [@falree77; @gneost97; @baum98; @vesp98; @prigne08]. We focus here on the semi-analytical model for the evolution of the mass function developed by @falzha01 [hereafter FZ01]. In the FZ01 model, clusters are tidally limited at the pericenters of their galactic orbits and are disrupted by the gradual escape of stars driven by a combination of internal two-body relaxation and external gravitational shocks. For most clusters, shocks are relatively weak, and relaxation is the dominant disruption mechanism. According to the “classical” theory of relaxation-driven stellar escape, as formulated by @spit87 and others, the mass $M$ of a tidally limited cluster decreases at a nearly constant rate: $dM/dt = \mu$ with $\mu \propto \rho_{\rm h}^{1/2}$, where $\rho_{\rm h} = 3 M / (8 \pi r_{\rm h}^3)$ is the mean density within the half-mass radius $r_{\rm h}$ of the cluster. FZ01 showed that, in this approximation, the evolving GCMF at any time $\psi({{M}},\,t)$ is related to the initial GCMF $\psi_0({{M}})$ by $\psi({{M}},\,t) = \psi_0({{M}}\,+\,\mu\,t)$. This has a characteristic bend or peak at $M \sim \mu\, t$ and the limiting forms $\psi(M, t) = \psi_0(\mu\,t)$, independent of $M$ for $M \ll \mu\,t$, and $\psi(M, t) = \psi_0(M)$, independent of $t$ for $M \gg \mu\,t$. FZ01 compared their model to the observed GCLF in the Milky Way in the approximation of constant $M/L$ and found excellent agreement. In particular, they showed both theoretically and observationally, for the first time, that the GCMF and GCLF are approximately constant for $M \la 10^5 \; M_{\odot}$ and $L \la 10^5 \; L_{\odot}$, in stark contradiction to the then-standard practice of fitting log-normal distributions to the data. This behavior of the GCMF at small $M$ and GCLF at small $L$ is strong evidence for the late disruption of clusters by internal two-body relaxation. Furthermore, @goud+04 [@goud+07] showed that the predicted time dependence of the FZ01 model is consistent with the observed luminosity functions of intermediate-age (3–4 Gyr old) cluster systems, and @chan+07, @mclfal08, and @goud12 showed that the predicted density dependence agrees well with the observed mass function for subsamples of clusters defined by different ranges of density. The approximations in the FZ01 model were made for simplicity and to highlight the main physical processes that shape the GCMF and GCLF. In particular, the adoption of classical evaporation and the neglect of $M/L$ variations are not essential features of the model. Both approximations were in standard use at the time (2001) to describe the evolution of individual clusters. The novel feature of the FZ01 model was to show how the evolution of the masses of individual clusters could be combined analytically into the evolution of the mass function of a cluster system. At a higher level of approximation, the stellar escape rate is modified by the fact that some of the stars that are scattered into unbound orbits may be scattered back into bound orbits before they have reached the tidal boundary and escaped from a cluster. In this case, often called “retarded” evaporation, the escape rate has a weak dependence on the crossing time $t_{\rm cr}$ in addition to the stronger dependence on the relaxation time $t_{\rm rlx}$. For tidally limited, low-mass clusters (initial masses $M_0 \la 10^{5} \; M_{\odot}$), the evolution can be approximated by $dM/dt \propto M/t_{\rm dis}$ with $t_{\rm dis} \propto M^{\gamma}$ and $\gamma \approx 0.7$, rather than $\gamma$ = 1 for classical evaporation [@fukheg00; @baum01; @baumak03; @lame+10]. For high-mass clusters, however, the retarded evaporation rate must approach the classical rate (in the limit $t_{\rm cr}/t_{\rm rlx} \rightarrow 0$; see Section \[sub:gamma\]). Two-body relaxation will also cause low-mass stars within a cluster to gain energy and escape faster than high-mass stars, thus reducing the average ${{M}}/L$ of the remaining stars over and above the fading caused by stellar evolution alone. As a result of this effect, in a coeval population of clusters (such as GCs), there should be a positive correlation between $M/L$ and the mass or luminosity of clusters, because those that have smaller relaxation times will have lost larger fractions of their initial mass. Such variation in $M/L$ was neglected in the FZ01 model for two reasons: it is difficult to predict reliably from theory, and it appeared from observations at the time to be weak or non-existent [@mcla00]. @krupor09 [hereafter KPZ09] modified the FZ01 model to include retarded evaporation and a variable ${{M}}/L$ ratio. In particular, they assumed that the rate of mass loss from clusters is given by $dM/dt \propto M/t_{\rm dis}$ with $t_{\rm dis} \propto M^{\gamma}$ and $\gamma = 0.7$ for clusters of all masses. Furthermore, to calculate the escape rates of stars of different masses and hence the variation in $M/L$ of clusters, they employed a semi-analytical model developed by @krui09 [hereafter K09] that includes several questionable assumptions and parameter choices. KPZ09 argued that their model is a significant improvement on the FZ01 model, both in terms of its theoretical validity and in terms of its ability to fit the observed GCLF in the Milky Way and other galaxies. Our main purpose in this paper is to demonstrate that the KPZ09 criticisms of the FZ01 model are not correct. Thus motivated, we also show that the variation in $M/L$ with $M$ or $L$ predicted by the KPZ09 model is much stronger than that allowed by observations. Furthermore, we show that the parameter values required for the KPZ09 model to fit the observed GCLF and the observed ${{M}}/L$ vs. $L$ relation are mutually exclusive. We emphasize that we do not dispute the general physical principles underlying retarded evaporation and $M/L$ variations. The results of this paper indicate, however, that the specific implementation of these effects in the KPZ09 model exaggerates their importance. We find that retarded evaporation and $M/L$ variations can be neglected for clusters massive enough to survive for a Hubble time of dynamical evolution. Therefore, for most practical purposes, the benefits of including these effects are largely offset by the increased complexity of the KPZ09 model relative to the FZ01 model. This paper is organized as follows. In Section \[s:GCLFfits\], we compare the FZ01 and KPZ09 models with the observed GCLF in the Milky Way, and we determine the best-fitting values of their parameters including the characteristic dissolution timescale. Section \[s:dyn\] presents our search for ${{M}}/L$ variations of the kind predicted by the KPZ09 model in a large compilation of recent dynamical measurements of GC masses. In Section \[s:Tdiss\], we compare the dissolution timescale required by the KPZ09 model to match the observed GCLF with the one required by the absence of observed ${{M}}/L$ variations. Section \[s:disc\] interprets the results from Sections \[s:dyn\] and \[s:Tdiss\] along with the observed stellar mass functions in Galactic GCs in terms of key properties and assumptions of the K09 model. Finally, we summarize our conclusions in Section \[s:conc\]. Comparisons of Models with the Observed GCLF {#s:GCLFfits} ============================================ In this Section, we compare the FZ01 and KPZ09 models with the observed GCLF in the Milky Way. We first derive analytical expressions for the evolving GCMFs assuming that the clusters are tidally limited and that stellar escape driven by two-body relaxation is the primary disruption mechanism. We then convert these GCMFs at an age of 12 Gyr into GCLFs adopting ${{M}}/L$ = constant for the FZ01 model and the ${{M}}/L$ vs. $L$ relation derived by KPZ09 for their model. Derivation of Model GCMFs {#sub:GCMFs} ------------------------- In both models considered here, the mass-loss rate of an individual cluster takes the form $$dM/dt \; \equiv \; -M/t_{\rm dis} \; = \; -(\mu/\gamma) \; M^{1-\gamma}\mbox{,} \label{eq:dMdt}$$ where $t_{\rm dis}$ is the dissolution timescale and $\mu$ and $\gamma$ are constants. This integrates to $${{M}}\,(t) \; = \; ({{M}}_0^\gamma - \mu \, t)^{1/\gamma}\mbox{,} \label{eq:massevol}$$ where ${{M}}_0$ is the initial cluster mass. These formulae are intended to represent smooth averages over the abrupt changes in mass caused by the escape of individual stars and over at least one full orbit of the cluster around its host galaxy. For $\gamma = 1$, equation (\[eq:dMdt\]) describes the classical mass-loss rate $\mu$ for stellar escape driven by internal two-body relaxation from a tidally limited cluster (@spit87 and references therein). This is the formula adopted in the FZ01 model. For $\gamma < 1$, equation (\[eq:dMdt\]) approximates the corresponding mass-loss rate for retarded evaporation in relatively low-mass clusters for which crossing times are significant fractions of relaxation times [@fukheg00; @baumak03]. The KPZ09 model assumes $\gamma = 0.7$ for clusters of all masses. FZ01 showed that the evolution of the mass function $\psi\, ({{M}},\,t)$ of a cluster system could be derived from the evolution of the masses ${{M}}(t)$ of individual clusters through a continuity equation. This approach yields $$\psi\, ({{M}},\,t) \; = \; \left({\partial {{M}}_0}/{\partial {{M}}}\right)_t \; \psi_0 \, ({{M}}_0) \; = \; \left({{{M}}}/{{{M}}_0}\right)^{\gamma-1} \; \psi_0\,({{M}}_0)\mbox{,} \label{eq:MF}$$ where $\psi\,({{M}}_0) = \psi\,({{M}}_0, 0)$ is the initial mass function (at $t = 0$). In this expression, the initial mass ${{M}}_0$ must be regarded as a function of the current mass ${{M}}$ and current time $t$ as given by inverting equation (\[eq:massevol\]). Following FZ01, we adopt a @sche76 initial mass function $$\psi_0 \, ({{M}}_0) = A \; {{M}}_0^{\beta} \; \exp\,(- {{M}}_0/{{M}}_{\rm c})\mbox{,} \label{eq:schech}$$ with adjustable parameters $A$, $\beta$, and $M_{\rm c}$. This function has a power-law shape with exponent $\beta$ below the bend at ${{M}}_{\rm c}$ to mimic the observed mass functions of young cluster systems, and it has an exponential decline above ${{M}}_{\rm c}$ as suggested by the observed tail of the GCMF at ${{M}}\ga 10^6 \; M_{\odot}$. Inserting equation (\[eq:schech\]) into equation (\[eq:MF\]) then yields the evolving GCMF: $$\begin{aligned} \psi\,({{M}},\,t) & = & A\;{{M}}^{\gamma - 1} \; ({{M}}^{\gamma} + \mu \, t)^{(\beta - \gamma + 1)/\gamma} \nonumber \\ & & \qquad \qquad \qquad \mbox{} \times \exp\left[- \, ({{M}}^{\gamma} + \mu \, t)^{1/\gamma}/{{M}}_{\rm c} \right]\!\!\mbox{.} \label{eq:finMF}\end{aligned}$$ This function has a bend at ${{M}}\sim (\mu\,t)^{1/\gamma}$; for lower ${{M}}$, it behaves as $\psi\,({{M}},\,t) \propto {{M}}^{\gamma-1}$, characteristic of dissolution by two-body relaxation, while for higher ${{M}}$, it behaves as $\psi\,({{M}},\,t) \propto {{M}}^{\beta}\,\exp\,(-{{M}}/{{M}}_{\rm c})$, independent of $\gamma$. Thus, we expect only minor differences in the shapes of the GCMF between $\gamma = 1$ (FZ01 model) and $\gamma = 0.7$ (KPZ09 model). The GCMF derived above is strictly valid only in the idealized case that all clusters in the GC system dissolve at the same rate $\mu$. In reality, clusters with different internal densities, determined mainly by the galactic tidal field, will dissolve at different rates $\mu_i$. In that case, equation (\[eq:finMF\]) can be re-interpreted as the probability density that an individual cluster with evaporation rate $\mu$ has a mass $M$ at an age $t$. The GCMF of a system of $\cal{N}$ coeval clusters is then the sum of the individual probability densities: $$\begin{aligned} \psi\,(M,\,t) & = & \sum_{i = 1}^{\cal{N}} \; A_i\;{{M}}^{\gamma - 1} \; ({{M}}^{\gamma} + \mu_i \, t)^{(\beta - \gamma + 1)/\gamma} \nonumber \\ & & \qquad \qquad \qquad \mbox{} \times \exp\left[- \, ({{M}}^{\gamma} + \mu_i \, t)^{1/\gamma}/{{M}}_{\rm c} \right]\!\!\mbox{.} \label{eq:multi-mu-MF}\end{aligned}$$ Here $\beta$ and $M_{\rm c}$ are assumed to be the same for all clusters in the GC system, and the normalization factors $A_i$ must be chosen such that the integral over all $M$ is unity for each term in the sum. We now relate the evaporation rate $\mu$ of a cluster to its mean density $\rho_{\rm h}$ within the half-mass radius $r_{\rm h}$ as follows. The $N$-body simulations of @baum01 and @baumak03 [hereafter BM03] showed that the dissolution time $t_{\rm dis}$ of a tidally limited cluster can be approximated by $$t_{\rm dis} \propto t_{\rm rlx} \; (t_{\rm cr}/t_{\rm rlx})^{1-\gamma} \propto M^{\gamma} \rho_{\rm h}^{-1/2}\mbox{,} \label{eq:t_dis_BM03}$$ where $t_{\rm rlx} \propto M^{1/2}\, r_{\rm h}^{3/2}$ is the half-mass relaxation time and $t_{\rm cr} \propto M^{-1/2}\, r_{\rm h}^{3/2}$ is the half-mass crossing time. The parameter $\gamma < 1$ in equation (\[eq:t\_dis\_BM03\]) is the same as that in equations (\[eq:dMdt\])–(\[eq:multi-mu-MF\]) and measures the deviation of the dissolution time from the formula $t_{\rm dis} \propto t_{\rm rlx}$ for classical evaporation.[^1] The extra factor of $(t_{\rm cr}/t_{\rm rlx})^{1-\gamma}$ for retarded evaporation comes about because unbound stars take a finite time, proportional to $t_{\rm cr}$, to cross a cluster before escaping from it. During that time, some of the unbound stars will be scattered back into bound orbits within the cluster, thus retarding its evaporation. From equations (\[eq:t\_dis\_BM03\]) and (\[eq:dMdt\]), we obtain $$\mu \propto M^{\gamma}/t_{\rm dis} \propto \rho_{\rm h}^{1/2}\mbox{,} \label{eq:mu}$$ independent of $\gamma$ for both classical and retarded evaporation.[^2] The evaporation rate of a cluster can also be expressed in terms of its mean density $\rho_{\rm t}$ within the tidal radius $r_{\rm t}$. This density is determined largely by the tidal field at the pericenter $R_{\rm p}$ of the orbit of the cluster within its host galaxy: $\rho_{\rm t} \propto G^{-1}\, \left(V_{\rm c,\,p} / R_{\rm p}\right)^2$, where $G$ is the gravitational constant, and $V_{\rm c,\,p}$ is the galactic circular velocity at $R_{\rm p}$ [@king62; @inna+83]. @mclfal08 showed that, while the densities $\rho_{\rm h}$ and $\rho_{\rm t}$ of GCs in the Milky Way span four or five orders of magnitude, the quantity $(\rho_{\rm t} / \rho_{\rm h})^{1/2}$ varies by less than a factor of two. Thus, to a good approximation, we can rewrite equation (\[eq:mu\]) in the form $$\mu \propto \rho_{\rm t}^{1/2} \propto V_{\rm c,\,p} / R_{\rm p}\mbox{.} \label{eq:mu_V/R}$$ In an idealized static and spherical galactic potential, the pericenters of all orbits remain fixed, and $\rho_{\rm t}$ and hence $\mu$ are constants of motion. Both FZ01 and KPZ09 computed the evaporation rates of clusters on different orbits from equation (\[eq:mu\_V/R\]) and then summed over a realistic distribution of orbits to determine the mass function $\psi(M,\,t)$ of a GC system from equation (\[eq:multi-mu-MF\]) or its integral equivalent. However, as @mclfal08 pointed out, the only role of the orbits in this calculation is to determine the cluster densities, $\rho_{\rm h}$ or $\rho_{\rm t}$, a step that can be eliminated by computing $\mu$ directly from the observed values of $\rho_{\rm h}$ or $\rho_{\rm t}$. Because tidal radii are notoriously uncertain, evaporation rates are much more robust when computed from $\rho_{\rm h}$ than from $\rho_{\rm t}$. This is the approach we take in this paper. Another simplification noted by @mclfal08 is that the mass function $\psi(M,\,t)$ of a GC system computed from equation (\[eq:finMF\]) with the median value of $\mu$ is very similar to that computed from equation (\[eq:multi-mu-MF\]) with a realistic distribution of $\mu$ (see also KPZ09). We refer to the former as single-$\mu$ models and the latter as multiple-$\mu$ models. In this paper, we present results for both types of models, confirming their similarity. The model GCMFs described above assume that evaporation by two-body relaxation is the dominant disruption mechanism. As such, they neglect the effects of stellar evolution and gravitational shocks. Mass loss by stellar evolution is dominated by supernovae and strong winds of massive stars in the first few $10^8$ years. This material is assumed to escape from clusters of all masses, thus leaving the shape of the GCMF unchanged. Meanwhile, for surviving GCs in the Milky Way, FZ01 showed that mass loss due to gravitational shocks is generally much weaker than that due to two-body relaxation for clusters with masses below the peak of the GCMF (see also @gneost97 [@dine+99]). Moreover, the rate of mass loss by gravitational shocks depends only on the densities $\rho_{\rm h}$ of clusters, not their masses, which preserves the shape of the GCMF in the sense that both $\psi$ and $M$ are simply rescaled by time-dependent factors (see FZ01). Fits of Model GCLFs to Observations {#sub:GCLFfit} ----------------------------------- Fitting equation (\[eq:finMF\]) to the observed Galactic GCLF requires a conversion from luminosity to mass. For the FZ01 model ($\gamma = 1$), we simply adopt a mass-independent ${{M}}/L_V$ ratio. For the KPZ09 model ($\gamma = 0.7$), we use their relation between ${{M}}/L_V$ and $L_V$ at an age of 12 Gyr.[^3] The corresponding parameters of the KPZ09 model are: @king66 concentration parameter $W_0 = 7$ (the median value of $W_0$ for the Milky Way GC system, @harr96), metallicity $Z = 0.0004$, and $t_0$ = 1.3 Myr. The last of these parameters is defined in the KPZ09 model as $t_0 \equiv t_{\rm dis} \: ({{M}}/M_{\odot})^{-\gamma}$, i.e., the dissolution timescale for a cluster with ${{M}}= 1\; M_{\odot}$. The conversion to our notation is $t_0 = \gamma/\mu$ (see equation \[eq:dMdt\]). We derive the observed GCLF of the Milky Way from the 2010 version of the @harr96 catalog of GC data. To avoid uncertain luminosities, we exclude 17 GCs with ${\it E}_{{\it B}-{\it V}} > 1.5$, resulting in a catalog with 140 GCs. $V$-band luminosities are derived by assuming the standard Galactic reddening law (with $A_V = 3.1 \; {\it E}_{{\it B}-{\it V}}$) and a solar absolute $V$-band magnitude of $M_V^0 = 4.83$ [@binmer98]. For the single-$\mu$ case, we perform least-squares fits of the model GCLFs to the observed Galactic GCLF using the non-linear Marquardt-Levenberg algorithm [@leve44; @marq63]. This is done for $\gamma = 1$ to represent the FZ01 model and for $\gamma = 0.7$ to represent the KPZ09 model. We adopt $\beta = -2$, as found in several studies of young star cluster systems (@falcha12 and references therein). Table \[t:GCLFfits\] lists the reduced $\chi^2$ values of the two model fits along with the resulting values of the evaporation rate $\mu$ and the cutoff and peak luminosities, $L_{\rm c}$ and $L_{\rm p}$, of the GCLF. Values for $L_{\rm p}$ were derived using equation (7) of @goud12. [@rcclcc@]{}\ \[-2ex\] & log $L_{\rm p}/L_{V,\odot}\!\!\!\!$ & log $L_{\rm c}/L_{V,\odot}\!\!\!\!$ & & rms$\!\!$ & $\chi^2_{\rm red}$\ & (2) & (3) & & (5) & (6)\ \[0.5ex\]\ \[-1.5ex\] FZ01 & 5.24 $\pm$ 0.04$\!\!\!\!$ & 7.15 $\pm$ 0.05$\!\!\!\!\!$ & $(1.80 \pm 0.03) \; 10^5 \times \Upsilon_V \!\!\!\!\!\!\!$ & 0.25$\!\!$ & 0.68\ K09 & 5.25 $\pm$ 0.03$\!\!\!\!$ & 7.21 $\pm$ 0.04$\!\!\!\!\!$ & $(4.94 \pm 0.05) \; 10^5$ & 0.21$\!\!$ & 0.60\ Gaussian & 5.24 $\pm$ 0.04$\!\!\!\!$ & N/A & & 0.80$\!\!$ & 5.14\ \[0.5ex\]\ \[-1.2ex\] We also made fits to the observed GCLF derived from the 2003 version of the Harris catalog as a check for consistency with KPZ09 and to estimate the magnitude of systematic errors in the fitted parameter values. We find that the values for both $\mu\,t$ and $L_{\rm p}$ agree to within 0.5% between the two versions of the Harris catalog, while the values for $L_{\rm c}$ come out $\sim$30% lower for the 2003 version. These differences do not affect any of our conclusions. For ${{M}}/L_V$ = 1.8, which is the overall mean value found from dynamical data in Sect. \[s:dyn\] below, the evaporation rates of the best-fitting single-$\mu$ FZ01 and KPZ09 models match each other within only 6% at $\mu \sim 2.6 \times 10^4 \; M_{\odot}$ Gyr$^{-1}$. As discussed in detail in @mclfal08, the corresponding value of $t_{\rm dis}/t_{\rm rlx} \sim 10 \; (-\beta - 1)^{-1} \sim 10$ where $\beta$ is the power-law slope in equation (\[eq:schech\]).[^4] This value of $t_{\rm dis}/t_{\rm rlx}$ is at the low end of the range typically found in theoretical calculations of single-mass clusters ($t_{\rm dis}/t_{\rm rlx} \sim 10-40$). However, simulations of multi-mass clusters have shown evaporation rates that are significantly larger than those of single-mass clusters [e.g., @leegoo95]. Furthermore, evaporation rates depend on the specific techniques, assumptions, and approximations used for each simulation [see, e.g., @gneost97; @vesheg97; @baumak03; @prigne08; @hegg14]. For the multiple-$\mu$ case, we proceed as follows. We compute FZ01 and KPZ09 models from equation (\[eq:multi-mu-MF\]), adopting the values of $M_c$ listed in Table \[t:GCLFfits\] along with $\beta = -2$ as before. To compute the evaporation rates $\mu_i$ of individual clusters, we again assume an age $t = 12$ Gyr, and we use the values of $\mu\,t$ in Table \[t:GCLFfits\] for a cluster with the median density $\hat{\rho_{\rm h}}$. We then use the current half-mass density of each Galactic GC to estimate its evaporation rate from $\mu_i = (\rho_{{\rm h},\,i}/\hat{\rho_{\rm h}})^{1/2}$ (see equation \[eq:mu\]). Again, we adopt the 2010 version of the @harr96 catalog, $M/L_V = 1.8$ for $\gamma = 1.0$ (i.e., the FZ01 model), and the KPZ09 relation between $M/L_V$ and $L_V$ for $\gamma = 0.7$. {width="13cm"} The resulting FZ01 and KPZ09 models are compared with the observed GCLF in Fig. \[f:GCLFplot1\]. Panels (a) and (c) show the GCLF in the form $dN/d\log L = (L\,\ln 10)\, dN/dL$ vs. $\log\,L$, analogous to the familiar observed GCLFs in magnitude space, whereas panels (b) and (d) show the GCLF in the form $\log\,(dN/dL)$ vs. $\log\,L$. For comparison with the more traditional log-normal representation of the GCLF, we also overplot a best-fit Gaussian whose parameters are $\left< \log\,(L_V/L_{V\,\odot}) \right> = 5.24 \pm 0.04$ and $\sigma_{{\rm log}\,L} = 0.64 \pm 0.05$ in panels (a) and (b). As Fig. \[f:GCLFplot1\] shows, the differences between the FZ01 and KPZ09 models in their ability to fit the observed GCLF are very small relative to the uncertainties. This holds for both the single-$\mu$ and multiple-$\mu$ models. Statistically, the KPZ09 model fits the data slightly better at $\log\,(L_V/L_{V,\,\odot}) \la 3.5$, while the FZ01 model fits better at $L \ga L_{\rm p}$. However, we emphasize that these differences are negligible not only with respect to the poisson errors, but also when compared to the improvement that the FZ01 and KPZ09 models provide over the traditional Gaussian representation of the GCLF. This is due to the asymmetry in the observed GCLF in that there are more GCs at $L\, < \, L_{\rm p}$ than at $L\, > \, L_{\rm p}$. This asymmetry is especially clear in panels (b) and (d) of Fig. \[f:GCLFplot1\] which show that $\psi\,(L) = dN/dL$ is approximately flat for GCs with $L_V \la 10^4\,L_{V,\,\odot}$. This behavior is clearly not consistent with a Gaussian LF, while it is matched very well by both the FZ01 and the KPZ09 models. Fig. \[f:GCLFplot1\] also shows that there is no significant difference between the single-$\mu$ and multiple-$\mu$ models in terms of fitting the observed GCLFs, confirming the findings of FZ01 and KPZ09. Quantitatively, the reduced $\chi^2$ values of the multiple-$\mu$ fits are 0.61 and 0.59 for the FZ01 and KPZ09 models, respectively. In the remainder of this paper, we adopt the single-$\mu$ models for simplicity. It is worth noting that the GCLFs of the FZ01 and KPZ09 models are more similar than are their GCMFs (see Fig. 1 of KPZ09). The reason for this is that the different relations between ${{M}}/L_V$ and $L_V$ largely compensate for the differences in the GCMFs. We conclude that the shape of the Galactic GCLF does not provide any evidence for a GC mass-dependent ${{M}}/L_V$ ratio. Thus, the claim by KPZ09 that “the match between the models and the observations \[of the Galactic GCLF\] exists only for values of $\gamma \approx 0.7$” is not correct. Searches for Dynamical M/L Variations {#s:dyn} ===================================== The most reliable estimates of GC masses are those derived from stellar kinematics, often referred to as “dynamical masses”. KPZ09 argued that the dependence of $M/L$ on $L$ predicted by their model was supported by the dynamical masses compiled by @mand+91 who found a weak correlation between $M/L$ and $M$ due to a few high $M/L$ values at $\log\, (M/M_{\odot}) \ga 5.5$. However, no such correlation appears in the more homogeneous and larger compilation of $M/L$ values by @mcla00, which superseded most of the @mand+91 results. With this in mind, we review recent measurements of dynamical ${{M}}/L$ ratios of ancient GCs in the literature to search for any dependence on GC luminosity similar to that predicted by the KPZ09 model. We test for a dependence of $M/L$ on GC luminosity rather than GC mass because the uncertainties in GC luminosities are typically relatively small and similar among clusters. In contrast, the uncertainties in GC masses vary significantly among clusters, depending on measurement specifics (e.g., numbers of stars measured, radial coverage). Thus, the correlation between measurement errors of $M/L$ and $\log\,L$ is much smaller than between those of $M/L$ and $\log\,M$. We include measurements from five independent studies of Galactic GCs: @mclvdm05 [37 GCs], @luet+13 [14 GCs], @zari+12 [@zari+13; @zari+14 14 GCs], @kimm+15 [25 GCs], and @watk+15 [15 GCs]. In addition, we use the large sample of 178 old GCs in the Andromeda galaxy by @stra+11. Limits on ${{M}}/L_V$ Variations in Milky Way GCs {#sub:Galdata} ------------------------------------------------- The dynamical ${{M}}/L_V$ measurements of the five samples of Galactic GCs are shown as a function of log $L_V$ in panels (a)–(e) of Fig. \[f:MLplot1\]. For each cluster sample, the overall average value of ${{M}}/L_V$ is shown by a horizontal dashed line. For comparison with the predictions of the KPZ09 model, we also overplot their relation between ${{M}}/L_V$ and $L_V$ at an age of 12 Gyr. {width="85.00000%"} Systematic differences among dynamical ${{M}}/L_V$ values derived by the five studies mentioned above are described in detail in the Appendix. Panel (f) of Fig. \[f:MLplot1\] shows our correction for these systematic differences, which are mainly due to the different ways to convert observed ${{M}}/L_V$ to “global” values that apply to each cluster as a whole. We adopt the normalization of @luet+13. Thus, we multiply the ${{M}}/L_V$ values of @mclvdm05, @zari+12 [@zari+13; @zari+14], @kimm+15, and @watk+15 by factors of 1.20, 1.75, 1.26, and 1.07, respectively (see Appendix \[s:AppA\]). [@lrrrrrrrrrrrr@]{}\ \[-2ex\] & & & & & &\ Parameter & FZ01 & K09 & FZ01 & K09 & FZ01 & K09 & FZ01 & K09 & FZ01 & K09 & FZ01 & K09\ & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) & (12) & (13)\ \[0.5ex\]\ \[-1.5ex\] $\sigma/\sqrt{N}$ & 0.73 & 0.81 & 0.57 & 0.61 & 0.27 & 0.49 & 0.56 & 0.72 & 0.41 & 0.71 & 0.70 & 0.79\ $\chi^2_{\rm red}$ & 0.38 & 0.49 & 0.29 & 0.32 & 0.04 & 0.10 & 0.21 & 0.33 & 0.15 & 0.41 & 0.40 & 0.50\ \[0.3ex\]\ \[-2.2ex\] $\mid\!\left<\Delta\chi^2_{\rm red} \right>\!\mid$ & & & & & &\ $\sigma(\mid\!\Delta \chi^2_{\rm red}\!\mid)$ & & & & & &\ \[0.5ex\]\ \[-1.8ex\] As Fig. \[f:MLplot1\] clearly shows, there is no evidence favoring the relation between ${{M}}/L_V$ and $L_V$ in the KPZ09 model over the constant ${{M}}/L_V$ ratio in the FZ01 model. This is quantified in Table \[t:MLfits\], which lists the $\chi^2$ values of the fits for the two models to the five data samples shown in Fig. \[f:MLplot1\].[^5] Specifically, the mass-independent ${{M}}/L_V$ model yields somewhat better fits to the data of @mclvdm05, @zari+12 [@zari+13; @zari+14], @kimm+15, and, more clearly, to the data of @watk+15. To put the $\chi^2$ values in Table \[t:MLfits\] in context, we performed a series of Monte Carlo simulations as follows. For each of the five $M/L$ datasets with individual luminosities $L_i$, we calculate masses $M_i$ under the assumptions of both the FZ01 model (using a constant $M/L_V = 1.8$) and the KPZ09 model (using their $M/L_V$ vs. $L_V$ relation). To each of the $L_i$ and $M_i$ values we then add random measurement errors based on the distribution of such errors in the dataset in question. The synthetic $M_i/L_i$ data that are intrinsically distributed like the FZ01 model are then fitted by the KPZ09 model, and vice versa. We define $\Delta \chi_{\rm red}^2 \equiv \chi_{\rm red,\,KPZ09}^2 - \chi_{\rm red,\,FZ01}^2$, i.e., the difference in $\chi_{\rm red}^2$ between the KPZ09 and FZ01 model fits. These simulations were performed 1000 times for each of the five datasets. The resulting distributions of $\Delta \chi_{\rm red}^2$ thus reflect the expected probabilities to find a given difference in $\chi_{\rm red}^2$ between the two model fits for the dataset in question. Fig. \[f:chi2plot\] shows the distributions of $\Delta \chi_{\rm red}^2$, while Table \[t:MLfits\] lists the corresponding absolute mean values $\mid\!\!\left< \Delta \chi_{\rm red}^2 \right>\!\!\mid$ and standard deviations $\sigma (\mid\!\!\Delta \chi_{\rm red}^2\!\!\mid)$. Comparing the values of $\left< \Delta \chi_{\rm red}^2 \right>$ with the measured differences in $\chi_{\rm red}^2$ between the KPZ09 and FZ01 fits to the five $M/L$ datasets (which are shown in Fig. \[f:chi2plot\] as vertical arrows), we find that the measured differences in $\chi_{\rm red}^2$ are all consistent with the hypothesis that a constant $M/L$ fits the data better than the KPZ09 model. Quantitatively, the KPZ09 model is excluded by the data at confidence levels of 2.9$\sigma$, 1.4$\sigma$, 2.4$\sigma$, 3.8$\sigma$, and 12.3$\sigma$, for the datasets of @mclvdm05, @luet+13, @zari+12, @kimm+15, and @watk+15, respectively. ![[*Panel (a)*]{}: probability densities of $\Delta \chi_{\rm red}^2$ values (see Sect. \[sub:Galdata\] for its definition) for the $M/L_V$ datasets shown in Figs. \[f:MLplot1\] and \[f:M31MLplot\]: @mclvdm05 [black line], @zari+12 [@zari+13; @zari+14 purple line], and @luet+13 [red line]. For all datasets, the solid and dashed curves represent sets of Monte-Carlo simulations in which the $M/L_V$ values of GCs are intrinsically distributed according to the FZ01 and KPZ09 models, respectively. For comparison, the measured differences in $\chi_{\rm red}^2$ between the KPZ09 and FZ01 model fits to the $M/L_V$ values of those datasets are shown by vertical solid arrows in the color of the dataset in question. *Panel (b)*: similar to panel (a), but now for the datasets of @kimm+15 [orange line], @watk+15 [light blue line], and SCS11 (black line; see Sect. \[sub:M31data\])). []{data-label="f:chi2plot"}](f3.eps){width="7.5cm"} We also note the lack of a correlation between dynamical ${{M}}/L_V$ and metallicity \[Fe/H\] at any GC luminosity in Figure \[f:ML\_vs\_FeH\]. This is not consistent with the relation predicted by simple stellar population (SSP) models, as illustrated by the solid line in Fig. \[f:ML\_vs\_FeH\]. This suggests that the $M/L_V$ values of ancient GCs are more affected by their dynamical histories than by their metallicities.[^6] For multi-mass King models of stellar systems, @shagie15 showed that the lack of correlation between ${{M}}/L_V$ and \[Fe/H\] among GCs can be explained by mass segregation, which causes the brighter, more massive stars in the central regions, where the kinematic measurements are typically made, to move with lower velocities than the fainter, less massive stars in the outskirts, and whose effect is stronger at higher metallicities due to the increasing turn-off mass with higher metallicity. ![Measured ${{M}}/L_V$ values of Galactic GCs shown in panel (f) of Fig. \[f:MLplot1\] versus \[Fe/H\]. Symbol colors are the same as in Fig. \[f:MLplot1\]. For comparison, the solid line represents the SSP model predictions of @bc03 for a @chab03 IMF. []{data-label="f:ML_vs_FeH"}](f4.eps){width="6.2cm"} The effect of stellar mass segregation on dynamical mass measurements is also mass-dependent, since present-day low-mass clusters have survived many more relaxation times on average than high-mass clusters. This means that the current level of mass segregation increases with decreasing GC mass, which causes dynamical masses of GCs to be *systematically underestimated for low-mass GCs*.[^7] This reinforces our conclusion that the dynamical ${{M}}/L_V$ data show no evidence for the ${{M}}/L_V$ variations predicted by the KPZ09 model. Limits on ${{M}}/L_V$ Variations in Andromeda GCs {#sub:M31data} ------------------------------------------------- We can also compare the model $M/L$ ratios to the observed ones for GCs in the Andromeda galaxy (M31). @stra+11 [hereafter SCS11] derived dynamical ${{M}}/L$ ratios for a large sample ($N = 178$) of old GCs in M31, covering a wide range of luminosities ($4.7 \la \log\,(L_V/L_{V,\,\odot}) \la 6.5$). We adopt the GC masses that they derived using the virial theorem, as well as their $V$-band luminosities. We also follow SCS11 in discarding GCs whose relative errors in ${{M}}/L_V$ are larger than 25%. To evaluate the luminosity dependence of the ${{M}}/L_V$ ratio from the M31 data, we first consider a subselection in metallicity. As reported by SCS11, the ${{M}}/L$ ratios of their metal-rich GCs ($\mbox{[Fe/H]} \ga -0.5$) are systematically and significantly lower than that of SSP model predictions. This contrasts with the lower-metallicity GCs whose ${{M}}/L_V$ ratios scatter around the SSP model predictions (cf. Fig. 1 of SCS11). As mentioned in the previous subsection, this behavior is consistent with mass segregation, whose effect is especially strong at $\mbox{[Fe/H]} > -0.5$ [@shagie15]. To minimize the bias introduced by mass segregation (i.e., causing underestimates of ${{M}}/L$ whose amplitude is mass-dependent), we therefore subselect GCs with $\mbox{[Fe/H]} < -0.5$. ![*Panel (a)*: dynamical ${{M}}/L_V$ versus $\log\,L_V$ for the sample of GCs in the Andromeda galaxy (M31) from SCS11. For comparison, we overplot linear and cubic fits to the data (long-dashed and short-dashed lines, respectively) and the relation between ${{M}}/L_V$ and log$L_V$ predicted by the KPZ09 model for an age of 12 Gyr (solid line). *Panel (b)*: similar to panel (a), but now for $\log\,{{M}}$ versus $\log\,L_V$. See the discussion in Sect. \[sub:M31data\]. []{data-label="f:M31MLplot"}](f5.eps){width="6.2cm"} Panel (a) of Fig. \[f:M31MLplot\] shows ${{M}}/L_V$ versus $\log\,L_V$ for the resulting sample of 109 GCs in M31. Linear and cubic fits to the data (using inverse variance weighting) are shown as black and red dashed lines, respectively, while the prediction of the KPZ09 model is shown as a blue solid line. Overall, the picture is very similar to that for the Galactic GCs in panel (c) of Fig. \[f:MLplot1\] in that ${{M}}/L_V$ is again independent of $L_V$. Quantitatively, the linear fit to the data has a slope $d({{M}}/L_V)/d(\log\,L_V) = 0.06 \pm 0.13$, while the cubic fit actually shows a marginal upturn of ${{M}}/L_V$ at $\log\,(L_V/L_{V,\odot}) \la 5.2$, where the KPZ09 model predicts a downturn. This result may seem surprising, since SCS11 reported a correlation between $M/L_V$ and $\log\,M$ for the same dataset, similar to that predicted by the KPZ09 model. We find that their apparent correlation results from a strong covariance of $M/L_V$ and $\log\,M$, due to correlated errors, rather than a true physical relationship.[^8] To illustrate this, we also perform linear and cubic fits between the *independent* variables $\log\,L_V$ and $\log\,{{M}}$ for the SCS11 dataset. A glance at panel (b) of Fig. \[f:M31MLplot\] confirms the trends seen in ${{M}}/L_V$ vs. $L_V$ discussed above. Finally, we calculate $\chi^2$ values of fits of the two models (i.e., constant ${{M}}/L$ and the KPZ09 model) to the SCS11 data. These values are listed in Table \[t:MLfits\]. Similar to our results from the Galactic GC samples, we find that the constant ${{M}}/L$ model provides a better fit to the M31 data than the KPZ09 model. We also performed Monte Carlo simulations like those described in Sect. \[sub:Galdata\] for the SCS11 dataset. The results are listed in Table \[t:MLfits\]. We find that the KPZ09 model is excluded by the SCS11 data at a confidence level of 4.5$\sigma$. We conclude that the available data on dynamical ${{M}}/L_V$ ratios for ancient GCs in the Milky Way and M31 provide no evidence for a dependence on GC luminosity of the kind predicted by the KPZ09 model. Specifically, the decline of ${{M}}/L_V$ with decreasing $L_V$ below $\log\,(L_V/L_{V,\,\odot}) \approx 5.2$ predicted by the KPZ09 model is not seen in the data. Constraints from Simultaneous fits to the GCLF and the ${{M}}/L_V$ versus $L_V$ Relation {#s:Tdiss} ======================================================================================== In this Section, we check whether the discrepancy between the observed ${{M}}/L_V$ data at low GC luminosities and the predictions of the KPZ09 model discussed in Sect. \[sub:Galdata\] may be resolved by increasing the characteristic dissolution timescale $t_0$, taking the fit to the GCLF into account as well. To do this, we turn to the full set of K09 models that are available online.[^9] In the following, we will refer to the KPZ09 model with $t_0 = 1.3$ Myr and $W_0 = 7$ discussed above as the “reference” K09 model. The impact of higher values of $t_0$ on the ${{M}}/L_V$ ratios predicted by the K09 models as a function of $L_V$ is shown in panel (b) of Fig. \[f:K09tests\], which is a copy of panel (f) of Fig. \[f:MLplot1\] to which we have added the K09 model predictions for $t_0$ = 3 and 10 Myr as dashed and dash-dotted lines, respectively. Note that the K09 models with higher values of $t_0$ are closer to the observed nearly flat distribution of dynamical ${{M}}/L_V$ data than the reference model with $t_0$ = 1.3 Myr. We then calculate the corresponding GCLFs for an age of 12 Gyr predicted by the K09 models for $t_0$ = 3 and 10 Myr, respectively. To do so, we adopt the same initial GCMF as before, i.e., a Schechter function with $\beta = -2$ and $M_{\rm c} = 9\times10^6 \; M_{\odot}$ (cf. Sect.\[s:GCLFfits\] and Table \[t:GCLFfits\]). The initial GC masses are then converted into masses and $V$-band luminosities at an age of 12 Gyr by means of the K09 model tables, using spline interpolation. Finally, present-day GCLFs are calculated from the GCMFs using $dN/d\log L_V = (dN/d\log M) \; (d\log M/d\log L_V)$ where $d\log M/d\log L_V$ represents the local slope of the relation between $\log L_V$ and $\log M$ at an age of 12 Gyr in the K09 models. The resulting GCLFs are depicted in panel (a) of Fig. \[f:K09tests\]. Note that the GCLFs predicted for the K09 models with $t_0$ = 3 and 10 Myr do not fit the GCLF well at all in that they peak at significantly lower luminosities than do the data and the reference K09 model (with $t_0 = 1.3$ Myr), due to the lower evaporation rates. We have verified that the K09 model GCLFs for $t_0$ = 3 and 10 Myr are not sensitive to the adopted cutoff mass $M_{\rm c}$ for $\log\,(L_V/L_{V,\,\odot}) \la 6$, even when $M_{\rm c}$ is increased by factors up to $10^3$. ![*Panel (a)*: observed Galactic GCLF compared with the K09 model at an age of 12 Gyr for three values of the dissolution timescale: $t_0$ = 1.3, 3, and 10 Myr (blue solid, dashed, and dashed-dotted lines, respectively). *Panel (b)*: dynamical ${{M}}/L_V$ ratios of Galactic GCs as a function of their $V$-band luminosities. This is a copy of panel (f) of Fig. \[f:MLplot1\], to which we have added K09 model predictions for $t_0$ = 3 and 10 Myr in blue dashed and dash-dotted lines, respectively. See the discussion in Section \[s:Tdiss\]. []{data-label="f:K09tests"}](f6.eps){width="7.4cm"} We conclude that there is no value of $t_0$ for the K09 model that is able to fit the GCLF and the ${{M}}/L_V$ data at log $(L_V/L_{V,\,\odot}) \la 5$ simultaneously. In contrast, the lack of a luminosity dependence of ${{M}}/L_V$ seen in panel (b) of Fig. \[f:K09tests\] is fitted naturally by the FZ01 model with a constant ${{M}}/L$. To check the robustness of this conclusion, we compare the distributions of relative evaporation rates $\mu$ of the GCs with log $(L_V/L_{V,\,\odot}) \leq 5$ that have dynamical ${{M}}/L_V$ measurements with those of all Milky Way GCs in the same luminosity range (cf.  panel (b) of Fig. \[f:K09tests\]). Once again, we use the 2010 version of the @harr96 catalog, using $M/L = 1.8$ for $\gamma = 1.0$, and the KPZ09 $M/L_V$ vs. $L_V$ relation for $\gamma = 0.7$ (cf. Section \[sub:GCLFfit\]). This comparison is shown in Fig. \[f:muhist\]. Note that the distributions of relative $\mu$ values of the two samples are very similar to each other, with the median value actually being slightly larger for the sample with dynamical ${{M}}/L_V$ measurements. This holds for both classical and retarded evaporation ($\gamma$ = 1.0 and 0.7). The inability of the K09 model to fit both the GCLF and the ${{M}}/L_V$ data at low luminosities simultaneously is therefore *not* due to a mismatch between dissolution timescales of the low-luminosity Galactic GCs with available ${{M}}/L_V$ data and those of the full sample of Galactic GCs in the same luminosity range. {width="12.cm"} Assessing Key Ingredients of the K09 Model {#s:disc} ========================================== In this Section, we assess our findings from the preceding Sections in the context of effects that were neglected in the FZ01 model but included in the K09 model, and which KPZ09 claim are significant improvements. Mass Dependence of Retarded Evaporation Rates {#sub:gamma} --------------------------------------------- We recall that the reference K09 model assumes $\gamma$ = 0.7 and $W_0 = 7$ for clusters of all masses. This choice was based on approximations by @lame+10, which in turn were based on the $N$-body simulations by BM03 (cf. Sect. \[sub:GCMFs\]). While retarded evaporation can be expected to occur in all GCs at some level, it seems unlikely that a single value of $\gamma$ applies to GCs across the full range of initial masses. Because of the scaling $t_{\rm cr}/t_{\rm rlx} \propto M^{-1}$, stars that reach escape velocities in high-mass GCs leave the cluster quicker relative to the situation in lower-mass GCs, implying that $\gamma$ should increase with GC mass. In fact, if a fixed $\gamma < 1$ applied for all GC masses, one would obtain an unphysical $t_{\rm dis} < t_{\rm rlx}$ at some high GC mass [see also @baum01]. We therefore expect $\gamma$ to approach unity for GCs with sufficiently high initial masses. {width="6.5cm"} See the discussion in Section \[sub:gamma\]. \[f:massevol\] The expected increase of $\gamma$ with initial cluster mass $M_0$ is relevant because @lame+10 derived the value $\gamma = 0.7$ from $N$-body simulations by BM03 with $4 \times 10^3 \la M_0/M_{\odot} \la 7 \times 10^4$ (corresponding to $8192 \leq N_0 \leq 131072$ for a @krou01 IMF). However, the great majority of GCs that survive for a Hubble time were initially much more massive than that. This is illustrated in Fig. \[f:massevol\] which shows the relation between the masses at $t$ = 0 and 12 Gyr computed from equation (\[eq:massevol\]) for $\gamma$ = 1.0 and 0.7 with the best-fit values of $\mu$ from Table \[t:GCLFfits\]. For comparison, we also show the same models with evaporation rates that are factors 0.5 and 2.0 times those of the respective best-fit values. Note that the simulated clusters used to derive $\gamma = 0.7$ by @lame+10, which have $M_0 \leq 10^{4.8}\; M_{\odot}$, do not even survive 12 Gyr of dynamical evolution according to these models. For a moderately low-mass GC with current mass $M \approx 10^{4.5}\;M_{\odot}$ for which the K09 models predict ${{M}}/L_V$ to be about half of that of high-mass GCs (see Sect. \[s:dyn\]), the initial mass indicated by these models is in the range $10^{5.5} - 10^6\;M_{\odot}$, depending on the model. It is thus clear that *the GCs that currently make up the bulk of the Galactic GC system were initially at least one order of magnitude more massive than the simulated clusters used to derive $\gamma = 0.7$* by @lame+10. Since $\gamma$ is expected to increase with increasing GC mass (cf. above), it seems prudent to regard the value $\gamma = 0.7$ adopted by K09 as a lower limit. In summary, $\gamma$ is expected to increase from $\approx 0.7$ for low-mass clusters to $\approx 1.0$ for high-mass clusters. Future $N$-body simulations with substantially more particles ($N$ in the approximate range $10^{5.5}-10^{6.5}$ according to Fig. \[f:massevol\]) will be needed to determine the actual dependence of $\gamma$ on initial cluster mass. Other Assumptions in the K09 Model {#sub:K09_other} ---------------------------------- The semi-analytical model of K09 involves a large number of other parameters, assumptions, and approximations (in addition to $\gamma = 0.7$ and $W_0 = 7$ for the reference K09 model). Some of these ingredients are plausible, but some others are ad hoc and/or not tested against observations, more rigorous theory, or realistic $N$-body simulations. Given these uncertain inputs to the model, it seems likely that the outputs from it will also be uncertain. Examples of assumptions in the K09 models whose quantitative effects are hard to estimate include the following. - The initial-final mass relations for dark remnants (white dwarfs, neutron stars, and stellar-mass black holes). - The distributions of kick velocities of the various types of dark remnants, and the dependence of the retention fractions of such remnants on (initial) cluster escape velocity. - The stellar mass dependence of the escape rate, for which the K09 models adopt the @heno69 rate for *close* stellar encounters in an *isolated* cluster (i.e., not residing in a tidal field) without mass segregation. However, real GCs *are* located in a (time-dependent) tidal field and stars escape mainly through repeated *weak* encounters (i.e., two-body relaxation), for which the dependence of the escape rate on stellar mass may be different. - The assumption of perfect energy equipartition during the mass-segregation phase of dynamical evolution. This assumption, which has an impact on the stellar mass dependence of evaporation, has recently been called into question since energy equipartition is not attained in $N$-body simulations except perhaps in their inner cores (BM03; @trevdm13 [@soll+15; @bian+16; @sper+16]). - The functional dependence of the stellar escape rate on the energy required for escape. - The approximation of a cluster potential by a @plum11 model, all the way out to the tidal radius (which does not exist for a Plummer model). - An assumed relation between half-mass radius and initial cluster mass, specifically $r_{\rm h} \propto M^{0.1}$. Observed protoclusters, however, have a different relation, $r_{\rm h} \propto M^{0.4}$ [@fall+10]. - The assumption that the half-mass radius $r_{\rm h}$ of a cluster remains constant throughout its lifetime. Recent $N$-body simulations show this to be an oversimplification. For example, $r_{\rm h}$ changes by a factor of $\sim 6$ during the lifetime of the $N$-body model for the globular cluster M4 by @hegg14. We refer the reader to the K09 paper for a full description and justification of these and other ingredients of the K09 model. As an illustration of uncertainties in the K09 model, we compare its predicted $M/L$ evolution with that of a corresponding BM03 simulation.[^10] We choose this particular comparison because the K09 model predictions were normalized against the BM03 simulations. As shown in Fig. \[f:MLcompare\], the K09 model predictions follow the BM03 simulation quite well until an age of a few Gyr, after which the ${{M}}/L_V$ decreases significantly in the K09 model whereas it continues to *increase* in the BM03 simulation, especially during the last few Gyr of its lifetime of $\approx 14$ Gyr. The latter increase, which implies an *increasing $M/L$ with decreasing luminosity* at ages $\ga 10$ Gyr, is thought to be due to the accumulation of massive white dwarfs in the cluster (see BM03). ![*Black line*: evolution of ${{M}}/L_V$ in the $N$-body simulation of BM03 for a $W_0 = 5$ cluster with initial mass $M_0 = 7 \times 10^4 \; M_{\odot}$ and $t_0 = 10.7$ Myr discussed in Section \[sub:K09\_other\]. *Dashed lines*: K09 model for a $W_0 = 5$ cluster with $M_0 = 10^5 \; M_{\odot}$, $t_0$ = 10 Myr, and three factors by which the “standard” kick velocities of dark remnants can be multiplied (0.5, 1.0, and 2.5 in purple, blue, and red, respectively). *Dotted lines*: same as dashed lines, but now for a cluster with $M_0 = 6 \times 10^4 \; M_{\odot}$. []{data-label="f:MLcompare"}](f9.eps){width="7.5cm"} The K09 model predictions were normalized against the BM03 simulations by using the same initial conditions (see Sect. 4 in K09). However, the initial conditions in the BM03 simulations differ from those in the published K09 models in one important aspect, namely the upper mass limit of the stellar IMF: BM03 used 15 $M_{\odot}$ (thus excluding progenitors of stellar-mass black holes), whereas the K09 model uses 100 $M_{\odot}$, and includes prescriptions for retention fractions of, and kick velocities applied by, neutron stars and stellar-mass black holes. Fig. \[f:MLcompare\] illustrates the significant effect of those kick velocities to ${{M}}/L_V$ in the K09 model. Note that the significant disagreement between ${{M}}/L_V$ in the BM03 simulation and that in the K09 model predictions at ages $\ga 6$ Gyr exists for all choices of kick velocities. While the analytical implementation of kicks by dark remnants in the K09 model seems plausible at some level, there is significant uncertainty in the retention fractions and kick velocities exerted by white dwarfs, neutron stars, and black holes. Another related simplification in the K09 model is that it applies a given retention fraction of dark remnants throughout the lifetime of a cluster, while $N$-body simulations have shown that the retention fraction of stellar-mass black holes can decrease significantly during the lifetime of the cluster due to multiple encounters [@kulk+93; @sigher93; @pormcm00; @merr+04; @tren+10]. Furthermore, the evolution of the retention fraction of black holes varies widely among repeated simulations with the same initial conditions [@merr+04]. It thus seems fair to conclude that the ${{M}}/L_V$ decrease during the second half of the lifetime of a cluster in the K09 models is not well constrained by observations or simulations. We emphasize that this comparison between the K09 models and the BM03 simulation tests only a few of the many assumptions listed above. Most of the others remain untested; however, it seems likely that they would also have some impact on the resulting ${{M}}/L$ vs. time and ${{M}}/L$ vs. $M$ and $L$ relations. Ambiguous Evidence from Stellar Mass Functions {#sub:alpha} ---------------------------------------------- One argument in favor of variations in cluster $M/L$ ratios is that observed stellar MF slopes $\alpha$ tend to be relatively flat for low-luminosity GCs when compared with those for high-luminosity GCs (@krumie09; KPZ09). We re-examine this argument in this Section by comparing available high-quality data on $\alpha$ in Galactic GCs with the K09 model predictions. Fig. \[f:alphaplot1\] shows $\alpha$ for a mass function $dn/dm \propto m^{\alpha}$ versus log $L_V$ from @dema+07, who compiled global MF slopes of 20 GCs in the stellar mass range of 0.3–0.8 $M_{\odot}$ derived from *Hubble Space Telescope (HST)* data. We assume measurement uncertainties of 0.3 dex for $\alpha$ (G. de Marchi, private communication). The observed values of $\alpha$ are compared with predictions of the reference K09 model in Fig. \[f:alphaplot1\]. ![Stellar mass function slope $\alpha$ versus log$L_V$ for GCs in the sample of @dema+07. Filled squares and open squares represent GCs with King concentration indices $c > 1.4$ and $c < 1.4$ (corresponding to $W_0 > 6$ and $W_0 < 6$), respectively. GCs with current half-mass relaxation times $<$1 Gyr are shown in red while the others are shown in black. The solid blue curve represents the predictions of the reference K09 model. The dashed line indicates $\alpha = -1.7$, the mean slope of the Kroupa IMF in the stellar mass range 0.3–0.8 $M_{\odot}$. See the discussion in Sect. \[sub:alpha\]. []{data-label="f:alphaplot1"}](f10.eps){width="6.5cm"} As mentioned by KPZ09, GCs in the fainter half of the sample studied by @dema+07 are on average more depleted in low-mass stars than those in the brighter half (see Fig. \[f:alphaplot1\]). This trend is, in principle, roughly consistent with the predictions of the reference K09 model. However, it should also be noted that *all* GCs that are significantly depleted in low-mass stars relative to the most luminous GCs feature relatively low King concentration parameters ($c \equiv \log\,(r_{\rm t}/r_{\rm c}) \la 1.4$, corresponding to $W_0 \la 6$).[^11] Conversely, all GCs with $W_0 \ga 6$ show $\alpha$ values consistent with a @krou01 IMF, and thus do not show any significant trend of ${{M}}/L_V$ with $L_V$. The latter does not seem consistent with the predictions of the reference K09 model for the Galactic GC system as a whole, which adopted $W_0 = 7$. However, the strong depletions of low-mass stars seen in GCs with low concentration indices can also be explained in a different way: @baum+08b argued that this observation may be caused by low-concentration GCs having started out as tidally limited clusters with relatively high levels of primordial mass segregation. The low-mass stars in such clusters would initially be located relatively close to the tidal radius, allowing their evaporation from the cluster to start well before the cluster experiences core collapse, thus leading to present-day mass functions that are relatively strongly depleted in low-mass stars. In this context, we also note that the $N$-body simulations by @tren+10 showed that GCs with low initial concentration index ($W_0$ = 5 or 3) gradually evolve to higher concentration indices ($W_0 \sim 7$) within 40–70% of their dissolution time. As such, low-luminosity GCs with *current* low concentration indices likely had even lower concentration indices initially, which might have caused the strong depletion of low-mass stars in such GCs to start even earlier than predicted by the @baum+08b study. This effect is not incorporated in the K09 models, rendering it hard to determine the reason for the significant difference in $\alpha$ between GCs with low and high concentration indices without additional information. In conclusion, the observed depletion of low-mass stars in low-luminosity GCs can be explained in more than one way and therefore is a weak or inconclusive test of the stellar mass dependence of the escape rate in the K09 models. Hence our conclusions from the previous sections remain valid. Summary and Conclusions {#s:conc} ======================= The most promising explanation for the peaked shape of the observed luminosity function (LF) of old GCs is that it is a relic of dynamical processes—primarily stellar escape driven by internal two-body relaxation—operating on an initial power-law or Schechter MF of young clusters over a Hubble time. The semi-analytical models of FZ01 showed quantitative agreement with the present-day observed GCLF for a wide range of initial MF shapes. For the sake of simplicity, and based on theoretical and observational standard practice at the time, FZ01 adopted an evaporation rate independent of cluster mass and a mass-to-light ratio independent of cluster mass. KPZ09 challenged both of these assumptions. They claimed that the predicted GCLF was only consistent with the observed GCLF if the stellar evaporation rate depends significantly on cluster mass: $dM/dt \propto M^{1-\gamma}$ with $\gamma = 0.7$ rather than $\gamma = 1$ for the FZ01 model. To calculate the escape rates of stars of different masses and hence the variation of ${{M}}/L$ among clusters with different masses, they employed the semi-analytical model of K09, which involves a significant number of plausible, but largely untested, assumptions and approximations. In this paper, we performed a quantitative evaluation of the KPZ09 claim that their model could fit the observed GCLF while the FZ01 model could not. We conclude that this claim is not valid, based on the following analysis and results. 1. The FZ01 and KPZ09 models provide equally good fits to the observed GCLF in the Milky Way. Furthermore, both models yield a significantly better fit to the observed GCLF at low luminosities than the traditional Gaussian model, highlighting the importance of mass loss driven by two-body relaxation in shaping the GCLF. 2. The measured ${{M}}/L_V$ values of GCs in the Milky Way and the Andromeda galaxy show no dependence on cluster luminosity. At low GC luminosities ($\log\,(L_V/L_{V,\,\odot}) \la 5$), where the impact of a stellar mass-dependent escape rate is expected to be strongest, the observations are fitted better by a mass-independent ${{M}}/L_V$ than by the KPZ09 model. This result holds for all six independent studies of GCs with dynamical ${{M}}/L_V$ data analyzed here. 3. We find that the discrepancy between the observed ${{M}}/L_V$ data at low GC luminosities and the KPZ09 predictions cannot be resolved by increasing the characteristic dissolution timescale $t_0$ of the K09 model, since such an increase would yield an unacceptable fit to the GCLF. In other words, there is no value of $t_0$ that allows the K09 model to fit simultaneously the GCLF and the observed ${{M}}/L_V$ data at $\log\,(L_V/L_{V,\,\odot}) \la 5$. 4. The parameter $\gamma = 0.7$ adopted by KPZ09 is based on results of $N$-body simulations of GCs with initial masses $M_0 \la 7\times 10^4\; M_{\odot}$, whereas the initial masses of GCs that survive 12 Gyr of dynamical evolution are at least one order of magnitude higher than that. Theory indicates that the value of $\gamma$ will increase toward unity at higher masses. Thus, the appropriate value of $\gamma$ for models of the GCMF and GCLF evolution may be closer to 1.0 than to 0.7. We emphasize again that we do not dispute the physical principles of retarded evaporation and $M/L$ variations. Rather, we claim that these effects add substantially to the complexity of dynamical GCMF and GCLF models and are not needed in practice to match observed GCLFs. We thank Laura Watkins and the anonymous referee for helpful comments. This project was partially supported by *HST* Program GO-11691 which was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5–26555. We acknowledge the use of the $R$ Language for Statistical Computing, see <http://www.R-project.org>. Baumgardt, H. 1998, , 330, 480 Baumgardt, H. 2001, , 325, 1323 Baumgardt, H., de Marchi, G., & Kroupa, P. 2008, , 685, 247 Baumgardt, H., & Makino, J. 2003, , 340, 227 (BM03) Bianchini, P., van de Ven, G., Norris, M. A., Schinnerer, E., & Varri, A. L. 2016, , 458, 3644. Binney, J., & Merrifield, M. 1998, Galactic Astronomy (Princeton, NJ: Princeton University Press) Bruzual, G. A. & Charlot, S. 2003, , 344, 1000 Chabrier, G. 2003, , 115, 763 Chandar, R., Fall, S. M., & McLaughlin, D. E. 2007, , 668, L119 Chandar, R., Whitmore, B. C., & Fall, S. M. 2010, , 713, 1343 Christian, C. A., & Schommer, R. A. 1988, , 95, 704 de Marchi, G., Paresce, F., & Pulone, L. 2007, , 656, L65 Dinescu, D. I., Girard, T. M., & van Altena, W. F. 1999, , 117, 1792 Elson, R. A. W., & Fall, S. M. 1985, , 299, 211 Fall, S. M. 2006, , 652, 1129 Fall, S. M., & Chandar, R. 2012, , 752, 96 Fall, S. M., Krumholz, M. R., & Matzner, C. D. 2010, , 710, L142 Fall, S. M., & Rees, M. J. 1977, , 181, 37P Fall, S. M., & Zhang, Q. 2001, , 561, 751 (FZ01) Fukushige, T., & Heggie, D. C. 2000, , 318, 753 Gnedin, O. Y., & Ostriker, J. P. 1997, , 474, 223 Goudfrooij, P. 2012, ApJ, 750, 140 Goudfrooij, P., Gilmore, D., Whitmore, B. C., & Schweizer, F. 2004, , 613, L121 Goudfrooij, P., Schweizer, F., Gilmore, D., & Whitmore, B. C. 2007, , 133, 2737 Harris, W. E. 1991, , 29, 543 Harris, W. E. 1996, , 112, 1487 Heggie, D. C. 2014, , 445, 3435 Hénon, M. 1969, , 2, 151 Innanen, K. A., Harris, W. E., & Webbink, R. F. 1983, , 88, 338 Kimmig, B., Seth, A., Ivans, I. I., Strader, J., Caldwell, N., Anderton, T., & Gregersen, D. 2015, , 149, 53 King, I. R. 1962, , 67, 471 King, I. R. 1966, , 71, 64 Kroupa, P. 2001, , 322, 231 Kruijssen, J. M. D. 2009, , 507, 1409 (K09) Kruijssen, J. M. D., & Mieske, S. 2009, , 500, 785 Kruijssen, J. M. D., & Portegies Zwart, S. F. 2009, , 698, L158 (KPZ09) Kruijssen, J. M. D., & Portegies Zwart, S. F. 2010, in ASP Conf. Ser. 423, Galaxy Wars: Stellar Populations and Star Formation in Interacting Galaxies, ed. B. J. Smith et al. (San Francisco: Astronomical Society of the Pacific), 151 Kulkarni, S. R., Hut, P., & McMillan, S. L. W. 1993, , 364, 421 Lamers, H. J. G. L. M., Baumgardt, H., & Gieles, M. 2010, , 409, 305 Lee, H. M., & Goodman, J. 1995, , 443, 109 Levenberg, K. 1944, Quarterly of Applied Mathematics, 2, 164 Lützgendorf, N., Kissler-Patig, M., Neumayer, N., et al. 2013, , 555, A26 Mandushev, G., Spassova, N., & Staneva, A. 1991, , 252, 94 Marquardt, M. 1963, SIAM J. Appl. Math., 11, 431 McLaughlin, D. E. 2000, , 539, 618 McLaughlin, D. E., & Fall, S. M. 2008, , 679, 1272 McLaughlin, D. E., & van der Marel, R. P. 2005, , 161, 304 Merritt, D., Piatek, S., Portegies Zwart, S., & Hemsendorf, M. 2004, , 608, 25 Meylan, G., & Heggie, D. C. 1997, , 8, 1 Plummer, H. C. 1911, , 71, 460 Portegies Zwart, S. F., & McMillan, S. L. W. 2000, , 528, L17 Prieto, J. L., & Gnedin, O. Y. 2008, , 689, 919 Pryor, C., & Meylan, G. 1993, in ASP Conf. Ser. 50, Structure and Dynamics of Globular Clusters, ed. S. G. Djorgovski & G. Meylan (San Francisco: Astronomical Society of the Pacific), 357 Schechter, P. 1976, , 203, 297 Shanahan, R. L., & Gieles, M. 2015, , 448, L94 Sigurdsson, S., & Hernquist, L. 1993, , 364, 423 Silverman, B. W. 1986, in [*Density Estimation for Statistics and Data Analysis*]{}, Chap and Hall/CRC Press, Inc. Sollima, A., Baumgardt, H., Zocchi, A., Balbinot, E., Gieles, M., Hénault-Brunet, V., & Varri, A. L. 2015, , 451, 2185 Spera, M., Mapelli, M., & Jeffries, R. D. 2016, , 460, 317 Spitzer, L. Jr. 1987, Dynamical Evolution of Globular Clusters (Princeton: Princeton University Press) Strader, J., Caldwell, N., & Seth, A. C. 2011, , 142, 8 (SCS11) Trenti, M., & van der Marel, R. 2013, , 435, 3272 Trenti, M., Vesperini, E., & Pasquato, M. 2010, , 708, 1598 van den Bergh, S., & Lafontaine, A. 1984, , 89, 1822 Vesperini, E. 1998, , 299, 1019 Vesperini, E., & Heggie, D. C. 1997, , 289, 898 Watkins, L. L., van der Marel, R. P., Bellini, A., & Anderson, J. 2015, , 812, 149 Whitmore, B. C., Chandar, R., Bowers, A. S., Larsen, S., Lindsay, K., Ansari, A., & Evans, J. 2014, , 147, 78 Zaritsky, D., Zabludoff, A. I., & Gonzalez, A. H. 2011, , 727, 116 Zaritsky, D., Colucci, J. E., Pessev, P. M., Bernstein, R. A., & Chandar, R. 2012, , 761, 93 Zaritsky, D., Colucci, J. E., Pessev, P. M., Bernstein, R. A., & Chandar, R. 2013, , 770, 121 Zaritsky, D., Colucci, J. E., Pessev, P. M., Bernstein, R. A., & Chandar, R. 2014, , 796, 71 Zhang, Q., & Fall, S. M. 1999, , 527, L81 Systematic Differences between Dynamical $M/L$ Studies {#s:AppA} ====================================================== In this Appendix, we analyze and quantify systematic differences between the five sources of dynamical ${{M}}/L_V$ measurements used in Section \[sub:Galdata\] so that they can be combined in a useful way. The ${{M}}/L_V$ values from @mclvdm05 were derived from central velocity dispersions from @prymey93 which were then extrapolated to “global” values (for the cluster as a whole) using surface brightness profiles. @mclvdm05 used single-mass King models in this extrapolation, so that any radial gradients of ${{M}}/L_V$ are neglected. Since ancient GCs commonly display radial mass segregation [e.g., @meyheg97], which causes the more massive stars to be more centrally concentrated than the less massive stars (which have higher ${{M}}/L$), we treat ${{M}}/L_V$ values from @mclvdm05 as lower limits. @zari+12 [@zari+13; @zari+14] measured velocity dispersions using a drift-scan technique that moved the spectrograph slit across the target cluster during the exposures, covering roughly the region within the half-light radius. The ${{M}}/L_V$ ratios of Zaritsky et al. were determined using an empirical relation between the half-light radius, the average surface brightness within that radius, and the mass-to-light ratio within that radius. This scaling relation was found to apply to all stellar systems from star clusters to massive elliptical galaxies. However, as discussed in @zari+12, their method produces ${{M}}/L_V$ values that are on average $\sim 40-50$% lower than those of @mclvdm05. This is consistent with the observation that ancient star clusters lie systematically somewhat above the empirical relation used by Zaritsky et al. [see Fig. 2 in @zari+11]. The kinematic data analyzed by @kimm+15 consisted of radial velocities of individual cluster stars, both from new observations and from the literature. GC masses were determined by fitting single-mass King models to the observed radial velocity dispersion profiles. Similar to the case of @mclvdm05, we thus treat the ${{M}}/L_V$ values from @kimm+15 as lower limits. @watk+15 derived ${{M}}/L_V$ ratios by fitting dynamical models to a combination of proper-motion velocity dispersions (from multi-epoch *HST* imaging data) and spectroscopic line-of-sight velocity dispersions. Their fitting involved Jeans models that assume a constant ${{M}}/L$ ratio, which we therefore formally treat as lower limits. However, the dispersion data used by @watk+15 covered a large range of radii, and no assumptions were made regarding the radial luminosity density profile, since they used a Multi-Gaussian Expansion fit to the latter. Hence, their resulting ${{M}}/L$ values can be expected to represent the cluster as a whole relatively well. Finally, the integrated-light kinematics in @luet+13 were derived from integral-field spectroscopy with spatial coverage typically out to the half-light radii of the clusters. Along with surface brightness profiles derived from *HST* data, the ${{M}}/L_V$ values in @luet+13 were determined using Jeans modeling. Their method incorporates a correction for radially varying ${{M}}/L_V$ and as such seems likely to produce results that are more robust relative to mass segregation than the other studies mentioned above. From the 11 GCs in common between the studies of @luet+13 and @mclvdm05, the ratio of the ${{M}}/L_V$ values is 1.20 $\pm$ 0.10 where the quoted uncertainty is the standard error of the mean. Similarly, the mean ratio of the ${{M}}/L_V$ values of @luet+13 and those of @kimm+15 is 1.26 $\pm$ 0.25 for the 4 GCs in common between the two studies, whereas that ratio is 1.07 $\pm$ 0.10 for the 7 GCs in common between @luet+13 and @watk+15. For the purposes of this paper, we suggest that these ratios are useful estimates of the factor by which ${{M}}/L_V$ values may be systematically underestimated in the studies that assumed a constant ${{M}}/L$ throughout the cluster.[^12] Panel (f) of Fig. \[f:MLplot1\] depicts our corrections for the systematic differences between the ${{M}}/L_V$ estimates of the five studies described above. We adopt the normalization of @luet+13. For consistency with this normalization, we multiplied the ${{M}}/L_V$ values of @mclvdm05, @zari+12 [@zari+13; @zari+14], @kimm+15, and @watk+15 by factors of 1.20, 1.75, 1.26, and 1.07, respectively. [^1]: BM03 used the notation $x$ instead of $\gamma$ in equation (\[eq:t\_dis\_BM03\]). The difference between the two is negligible: $x$ and $\gamma$ were derived by evaluating $t_{\rm dis}$ as functions of $N$ (number of particles) and the cluster mass $M$, respectively [see @lame+10]. [^2]: To avoid confusion, we note that $\mu_{\rm ev}$ in @mclfal08 is related to our $\mu$ by $\mu_{\rm ev} = (\mu/\gamma)\;M^{1-\gamma}$. [^3]: This relation was shown in Fig. 4 of KPZ09, and with more clarity, in @krupor10. [^4]: Note that $\beta$ in this paper is equal to $-\beta$ in @mclfal08. [^5]: The fits of the KPZ09 model to the data were performed by allowing for a $L_V$-independent scale factor in ${{M}}/L_V$. [^6]: This also explains why one can safely neglect the metallicity dependence of ${{M}}/L$ when fitting the GCLF by cluster evolution models as in Sect. \[s:GCLFfits\] (see also FZ01 and KPZ09). [^7]: This trend might not extend to the lowest-mass clusters ($\log\,(M/M_{\odot}) \la 4.5$) which have only $\la 10\%$ of their lifetimes remaining and may already have undergone core collapse (see, e.g., BM03). [^8]: A similar situation is found for the dataset of @kimm+15 discussed in Sect. \[sub:Galdata\]. [^9]: The K09 `SPACE` models are available at <http://bit.ly/1Pbttlg>. [^10]: The properties of this BM03 simulation, shown in their Fig. 18, are: $W_0 = 5$ and dissolution timescale $t_0 = 10.7$ Myr (see equation 7 in @krumie09). The properties of the corresponding K09 model are: $W_0 = 5$ and $t_0 = 10$ Myr. [^11]: This tendency was already noted by @dema+07 in a diagram of $\alpha$ versus $c$ (their Fig. 1). Our Fig. \[f:alphaplot1\] illustrates the luminosity dependence of the difference in $\alpha$ between GCs with high and low values of $c$. [^12]: As mentioned in Sect. \[sub:Galdata\], the level of mass segregation is expected to depend on cluster mass to some extent. We neglect this effect, which is likely most significant for studies that use central velocity dispersions (e.g., @mclvdm05).

--- abstract: 'We present optimal and minimal measurements on identical copies of an unknown state of a qubit when the quality of measuring strategies is quantified with the gain of information (Kullback of probability distributions). We also show that the maximal gain of information occurs, among isotropic priors, when the state is known to be pure. Universality of optimal measurements follows from our results: using the fidelity or the gain of information, two different figures of merits, leads to exactly the same conclusions. We finally investigate the optimal capacity of $N$ copies of an unknown state as a quantum channel of information.' address: | e-mail: guifre@ecm.ub.es\ Departament d’Estructura i Constituents de la Matèria\ Universitat de Barcelona\ Diagonal 647, E-08028 Barcelona, Spain author: - Rolf Tarrach and Guifré Vidal title: 'Universality of optimal measurements.' --- Consider an unknown state of a two-level quantum system described by the density matrix $\rho(\vec{b})$, $\vec{b}$ being the Bloch vector, $b \equiv |\vec{b}| \leq 1$. The preparation device provides $N$ identical copies of the system, so that the state at our disposal is $\rho(\vec{b})^{\otimes N}$. In the past few years the optimal measuring strategy, i.e. the most successful at revealing the identity of the unknown state, has been obtained, first for pure states [@MasPop; @DBE] and then for mixed states [@VLPT]. Also the minimal among the optimal strategies, i.e. the ones with the smallest number of outcomes, have been constructed, both for pure states [@LPT] and mixed states [@VLPT]. In the processing of information contained in quantum states, knowing the most efficient read-out procedures, i.e. the optimal and least resource consuming ones, is of course of importance. In all these contributions the quality of the measuring strategy, characterized by a resolution of the identity \_i M\_i = , \[reso\] in terms of positive operators $M_i \geq 0$, has been quantified by the fidelity [@Uhl]. In other words, when outcome $i$ (related to $M_i$) happens one guesses the unknown state to be $\tilde{\rho}_i \equiv \rho(\vec{p}_i)$ and one quantifies the quality of the guess by F((),(\_i)) (\[  \] )\^2. \[fid\] One can arrive at eq. (\[fid\]) from several different starting points. One of them is based on a measure of distinguishability of the probability distributions associated to $\rho$ and $\rho'$ by performing general positive operator valued measurements (as in eq. (\[reso\])) on them [@Woo] and minimizing, F(, ’) = ( \_j )\^2. Another is based on the standard Hilbert space scalar product of the two pure states which belonging to ${\cal C}^2\otimes{\cal C}^2$ lead to $\rho$ and $\rho'$ when reduced [@Jos], F(, ’) = ||\^2, where maximization is performed over $\{{\mbox{$| \psi \rangle$}}, {\mbox{$| \psi' \rangle$}}\} / \rho = \mbox{Tr}_a [{{\mbox{$| \psi \rangle$}}\!{\mbox{$\langle \psi |$}}}],~\rho' = \mbox{Tr}_a [{{\mbox{$| \psi' \rangle$}}\!{\mbox{$\langle \psi' |$}}}].$ These equivalent definitions of the fidelity, plus the following properties which characterize it further, make it a unique quantification of the comparison of two general quantum states: 1. $0 \leq F(\rho,\rho') = F(\rho',\rho) \leq 1$. 2. $F(\rho,\rho')= 1 ~\Leftrightarrow~ \rho = \rho';~~ F(\rho,\rho')= 0 ~\Leftrightarrow~ \rho\rho'=0$. 3. $F(U\rho~ U^{\dagger},U\rho'U^{\dagger})=F(\rho,\rho'), ~~UU^{\dagger}=U^{\dagger}U=\openone$. 4. $F({{\mbox{$| \psi \rangle$}}\!{\mbox{$\langle \psi |$}}},\rho) = {\mbox{$\langle \psi |$}} \rho {\mbox{$| \psi \rangle$}}$. 5. $F(\rho\otimes \sigma,\rho'\otimes\sigma') = F(\rho,\rho')F(\sigma,\sigma')$ 6. $F(\rho,p\rho_1\!+\! (1\!-\!p)\rho_2) \geq pF(\rho,\rho_1) + (1\!-\!p) F(\rho,\rho_2),\\ 0\leq p\leq 1$. In references [@MasPop; @DBE; @LPT] the unknown state was known to be pure, $b=1$, but no knowledge of the direction of the Bloch vector was assumed. In reference [@VLPT] the unknown state was a mixed state drawn stochastically from a known isotropic distribution $f(b)$, and although the best guess $\tilde{\rho}_i$ depended on $f(b)$, the optimal measuring strategy, that is the set $\{ M_i \}$ of positive operators of the different outcomes, did not. For isotropic distributions optimal measurements are thus distribution, i.e. $f(b)$, independent. However, proposing an outcome-dependent guess and evaluating its quality through the fidelity is only one of the criteria that could have been used to define optimal measurements. A sound alternative, the one we shall investigate in this work and probably the most sensible choice in the context of quantum information theory, consists of quantifying the quality of measuring strategies through the gain of information about the unknown state. In fact, information theory already supplies a universally accepted, unambiguous scheme for this purpose, that we shall follow. It is based on Bayes formula, which provides a conditional (outcome-dependent), posterior distribution $f_c(\vec{b}|i)$ from the (here isotropic) prior distribution $f(b)$, and on the Kullback, which quantifies the gain of information acquired when replacing $f(b)$ with $f_c(\vec{b}|i)$. More specifically, if $P_i(\vec{b}) \equiv \mbox{Tr}[\rho(\vec{b})M_i]$ is the probability of outcome $i$ when the unknown state is $\rho(\vec{b})$ and P\_[ap]{}(i) d\^3b f(b) P\_i()   (d\^3b f(b) = 1) \[apriori\] is the a priori probability of outcome $i$, then Bayes formula states that the posterior distribution $f_c(\vec{b}|i)$, the one which collects our knowledge about the unknown state $\rho(\vec{b})$ after measuring when the initial knowledge was given by $f(b)$, reads f\_c(|i) = . \[Bayes\] The gain of information about $\rho(\vec{b})$, $\Delta I$, is then given, in bits, by the Kullback of $f_c(\vec{b}|i)$ relative to $f(b)$ [@Kull] K\_i\[f\_c/f\] d\^3b  f\_c(|i) \_2 . \[kullback\] This expression, the only one satisfying a series of intuitively reasonable conditions [@Hob], is well-defined for continuous distributions (it has no dependence on the measure in the space of quantum states) and its average over possible outcomes, |[K]{}\[f\_c/f\] \_i P\_[ap]{}(i) K\_i\[f\_c/f\], \[avekullback\] is precisely the difference of the a priori and average a posteriori entropies $H$ of the corresponding probability distributions of states, H\[f\]-|[H]{}\[f\_c\] - && d\^3b  f(b) \_2 f(b)\ + \_i P\_[ap]{}(i) && d\^3b  f\_c(|i)\_2 f\_c(|i), \[deltaentro\] as can be checked by considering eqs. (\[Bayes\]-\[avekullback\]) and that $\sum_i P_i(\vec{b})=1$ [@text]. This quantification is therefore equivalent to the one already used in previous works on quantum state estimation with discrete distributions (see, e.g., ref. [@PerWoo]). First, the question of which are the optimal measurements according to this information theoretically based criterion will be addressed. We will check explicitly for $N=1$ and $N=2$, and provide clues for any $N$, that optimal –and also minimal– measuring strategies are universal, i.e. independent of whether the fidelity or the increase of information is used for their quantification, and will compute the corresponding optimal gain of information $\Delta I$. Then we will move to consider which is the isotropic prior $f(b)$ for which optimal measurements extract most information, so that it corresponds to the optimal (isotropic) quantum channel of information. After introducing a reversible compression procedure we conclude that the optimal amount of extractable information is, as $N\rightarrow\infty$, of one bit per effective qubit isotropic distributions. In order to find an optimal measuring strategy, i.e. a set of operators $M_i$ as in eq. (\[reso\]) maximizing the gain of information (eq. (\[avekullback\])), the following theorem and subsequent corollaries, valid for any number of copies $N$, will be very useful. [**Theorem:**]{} Let the positive operator $M_i \geq 0$ be such that its probability $P_i(\vec{b})= \mbox{Tr}[M_i\rho(\vec{b})^{\otimes N}]$ can be written, for any $\vec{b}$, as the sum of two contributions of the form $$P_{i,k}(\vec{b})\equiv$Tr$[M_{i,k}\rho(\vec{b})^{\otimes N}], ~~k=1,2,$$ where the operators $M_{i,1},M_{i,2}$ are also positive (and $M_{i,1}+M_{i,2}$ is not necessarily equal to $M_i$). Let us introduce corresponding prior probabilities $P_{ap}(i,k)$ and posterior distributions $f_c(\vec{b}|i,k)$ as in eqs. (\[apriori\]) and (\[Bayes\]). Then, P\_[ap]{}(i) K\_i\[f\_c/f\] \_[k=1]{}\^2 P\_[ap]{}(i,k) K\_[i,k]{}\[f\_c/f\]. [**Proof:**]{} It follows from the inequality (x\_1+x\_2) x\_1 + x\_2 $\forall~ x_1,x_2,y_1,y_2 \geq 0$. $\Box$ [**Corollary 1:**]{} An optimal measuring strategy with rank-one operators always exists. (cf.[@Dav]) [**Proof:**]{} Indeed, suppose $\sum_i M_i = \openone$ corresponds to an optimal measurement. Then, if $M_i = \sum_{k} {{\mbox{$| i,k \rangle$}}\!{\mbox{$\langle i,k |$}}}$ is the spectral decomposition of $M_i$, it follows from the theorem that the rank-one POV measurement $\sum_{i,k} {{\mbox{$| i,k \rangle$}}\!{\mbox{$\langle i,k |$}}} = \openone$ is also optimal. $\Box$ We can already consider the case $N=1$, that is, when only one copy of the unknown state is available. One can convince oneself immediately that an optimal (and also minimal) measurement is just a standard von Neumann measurement. In fact, any will do because of the isotropy of $f(b)$. Suppose that we measure $\sigma_z$. Then, for $\vec{b}= (b\sin\theta\cos\phi,b\sin\theta\sin\phi, b\cos\theta)$, we have f\_c(|) = (1b) f(b) and the gain of information is I\^[(1)]{}=\_0\^1 db b\^2 f(b)  -  .          \[info1\] The function in square brackets in eq. (\[info1\]) is monotonically increasing, so that the distribution for which the absolute increase in knowledge is maximal is f\_m\^[(1)]{}(b) = (b-1), \[pure\] i.e. an isotropic distribution of pure states. It is interesting to point out that if instead of using in ref. [@VLPT] the mean average fidelity $\bar{F}^{(1)}$ we had used the mean average increase in fidelity, F\^[(1)]{} |[F]{}\^[(1)]{} - F\^[(1)]{}\_[ap]{}, with the optimal guess $\tilde{\rho}_0 \equiv \rho(0)$ if no measurement is performed, so that F\^[(1)]{}\_[ap]{} = + I\_[1/2]{} = F\_[ap]{}\^[(N)]{} with (cf. [@VLPT]) I\_ 4\_0\^1 db b\^2 f(b)()\^ ($I_0 = 1, I_{\alpha} \geq 4I_{\alpha+1}$), we would have obtained F\^[(1)]{} = - I\_[1/2]{}. It is then easily verified that the maximum value of $\Delta F^{(1)}$ also corresponds to the distribution eq. (\[pure\]). Thus, for N=1, quantifying with the fidelity or with the Kullback information leads to the same (for N=1 somewhat obvious) optimal and minimal measuring strategy and to the same distribution which maximizes $\Delta I^{(1)}$ and $\Delta F^{(1)}$. Is this also true for $N=2$? In order to answer this question we need to present a second corollary. Notice first that with the following notation (borrowed from [@VLPT]) for the composite Hilbert space of $N$ copies of the unknown state $\rho(\vec{b})$, $$\label{hilberts} {\cal H}^{(N)} \equiv {\cal H}_A \otimes {\cal H}_B \otimes ... {\cal H}_N,$$ for the corresponding local spin operators, $$\begin{aligned} \label{sabn} \nonumber \vec S_A &\equiv& {1\over 2}\vec {\sigma} \otimes I^{\otimes N-1},\\ \nonumber \vec S_B &\equiv& {1\over 2} I \otimes \vec {\sigma} \otimes I^{\otimes N-2},\\ \vdots \nonumber\\ \vec S_N &\equiv& {1\over 2} I^{\otimes N-1} \otimes \vec {\sigma}, \end{aligned}$$ and for the partial and total spin operators, $$\label{sms} \vec {S}_{(\alpha)} \equiv \sum^{\alpha}_{\beta=A} \vec {S}_{\beta},~~ \alpha = A, B, \cdots, N;~~~\vec {S} \equiv \vec {S}_{(\alpha=N)},$$ the following spin invariances hold [@VLPT]: $$\label{comm} \left[\vec {S}^2_{(\alpha)}, \rho^{\otimes N}\right]=0~~~~~~~ \alpha=A,\cdots,N,$$ and since $$\label{directsum} \left[\vec {S}^2_{(\alpha)}, \vec{S}^2_{(\beta)}\right]=0\qquad \forall \alpha,\beta$$ the total Hilbert space can be written as a direct sum $$\label{totalh} {\cal H}^{(N)}= \oplus_{\{ s_{(\alpha)}\} } E_{\{ s_{(\alpha)}\} }$$ where $E_{\{ s_{(\alpha)}\} }$ are the simultaneous eigenspaces of all the operators $\vec{S}^2_{(\alpha)}, \forall \alpha \not = A$, with corresponding eigenvalues $\{ s_{\alpha}(s_{\alpha} +1)\}$, ordered with decreasing $\alpha$ (see [@VLPT] for more details). For instance, for $N = 2$ only $\vec{S}^2_{(B)}$ ($s_{(B)}$) is relevant, i.e. $E_{\{ s_{(\alpha)}\} } = E_{s_{(B)}}$, and the decomposition reads \^[(N=2)]{}= E\_1 E\_0, where $E_1$ is the triplet or symmetric (under exchange of copies) subspace, with total spin $s\equiv s_{(B)}=1$, whereas $E_0$ is the singlet or antisymmetric subspace, with total spin $s=0$. Then, [**Corollary 2:**]{} There always exists an optimal measuring strategy consisting only of rank-one operators of the form ${{\mbox{$| \{s_{(\alpha)}\} \rangle$}}\!{\mbox{$\langle \{s_{(\alpha)}\} |$}}}$, where the not necessarily normalized vector ${\mbox{$| \{s_{(\alpha})\} \rangle$}}$ is an eigenvector of all partial and total spin operators, i.e. \^2\_[()]{}    =  s\_[()]{}(s\_[()]{}+1)     , and thus it belongs to the subspace $E_{\{s_{(\alpha)}\}}$. [**Proof:**]{} Let $\sum_i M_i=\openone$ correspond to an optimal measurement with rank-one operators $M_i = {{\mbox{$| i \rangle$}}\!{\mbox{$\langle i |$}}}$ (where the do not need to be orthogonal nor normalized) and let $\Pi_{\{s_{\alpha}\}}=\Pi_{\{s_{\alpha}\}}^{2}$ be a projector onto the whole subspace $E_{\{s_{\alpha}\}}$. Then it follows from eq. (\[totalh\]) that \_[{s\_}]{} \_[{s\_}]{} \_[{s\_}]{} = \_[{s\_}]{} \_[{s\_}]{} = , so that if we replace $\openone$ with $\sum_i M_i$ in the LHS of this equation, we obtain a new measurement \_[i,{s\_}]{} = ;      \_[{s\_}]{} . \[optimtoo\] Now, since eq. (\[comm\]) implies that for each ${\mbox{$| i \rangle$}}$, \[()\^[N]{}\] = \_[{s\_}]{} \[()\^[N]{}\], the theorem guarantees that the measurement of eq. (\[optimtoo\]) is also optimal. $\Box$ \[Notice that exactly the same conclusion was also achieved, for any $N$, when the fidelity was used as a criterion for optimality [@VLPT], this being indicative of the universality we are considering here.\] Thus, in order to find an optimal measuring strategy for $N=2$ we can always choose the pure states on which the measurement projects to be symmetric or antisymmetric under the exchange of the two qubits. Let us next compute $\Delta I^{(2)}$ for the optimal strategy of ref. [@VLPT], that is corresponding to a resolution of the identity of the form = + \_[i=1]{}\^4()\^[2]{}, \[opti2\] where ${\mbox{$| \sigma \rangle$}}$ is the (normalized) singlet state, $\vec{\sigma}\cdot\hat{n} {\mbox{$| \hat{n} \rangle$}} = {\mbox{$| \hat{n} \rangle$}}$ (${\mbox{$\langle \hat{n} | \hat{n} \rangle$}}=1$) and the four unitary vectors $\hat{n}_i$ point to the four directions of the vertices of a regular tetrahedron. One readily obtains f\_c(|) =  &;& P\_[ap]{}() = I\_1\ f\_c(|) = (1+)\^2 &;&P\_[ap]{}() = (1-I\_1) so that I\^[(2)]{} = \_0\^1db b\^2f(b)-\ (1-I\_1)( + \_2) -I\_1\_2I\_1 - 2. \[expressio\] Can we do better, i.e. is there another resolution of the identity which leads to a larger $\Delta I^{(2)}$? Let us prove that there is none. Because of corollary 2, the whole question boils down to whether symmetric entangled states could do better than the symmetric product states ${\mbox{$| \hat{n}_i \rangle$}}{\mbox{$| \hat{n}_i \rangle$}}$ used in eq. (\[opti2\]). Consider therefore a general symmetric state of Schmidt decomposition = + ,   p, where the isotropy of $f(b)$ has been taken into account in choosing the basis. One can readily obtain the average Kullback information corresponding to this state, $$\Delta I^{(2)}_{\psi} = \frac{1}{2}\int_0^1db~b^2f(b)\int_0^{2\pi}d\phi\int_{-1}^{1}d\mu ~h\log_2\frac{h}{\frac{1-I_1}{3}},$$ h && k + l2,     l 2b\^2(1-\^2),\   k && 1+b\^2\^2 + (2p-1)2b, which after integration of $\phi$ gives I\^[(2)]{}\_ &=& \_0\^1db b\^2f(b)\_[-1]{}\^[1]{}d{(1+b\^2\^2)\_2\ &+& k\_2 (k + )}. This is a function of $p$ that we want to maximize. Only $k\log_2 (k + \sqrt{k^2 - l^2})$ depends on $p$. The part $- l^2$ is maximized for $p=0$ and $p=1$. The other part too, as one can see easily neglecting the term $l^2$. Thus $\Delta I^{(2)}_{\psi}$ is maximized when ${\mbox{$| \psi \rangle$}}$ is a product state and the resolution of eq. (\[opti2\]) is indeed optimal. As we did for $N=1$, it is interesting to recall, with the help of ref. [@VLPT], the average increase in fidelity for $N=2$ F\^[(2)]{} = + I\_ - I\_. One can now check that both $\Delta I^{(2)}$ and $\Delta F^{(2)}$ are again maximized for the distribution eq. (\[pure\]). For $\Delta I^{(2)}$ this follows by observing that the part in square brackets of eq. (\[expressio\]) is an increasing function of $b$ and that the other part, which depends on $I_1$, increases as $I_1$ goes towards zero. We have thus checked for $N=1$ and $N=2$ that both the fidelity and the Kullback information lead to the same optimal measuring strategy and to the same, pure state, distribution which maximizes their increases. We conjecture, while not foreseeing any feature which could jeopardize extending the proof to $N>2$, that the universality of optimal measurements holds for any number $N$ of copies of the unknown state [@localmax]. Corollary 2 makes this conjecture very plausible. The precise optimal strategy is in fact determined to a great extend by the isotropy of the prior distribution, the symmetries of the state $\rho(\vec{b})^{\otimes N}$ which allow to choose each positive operator $M_i$ to act only on one of the subspaces $E_{\{s(\alpha)\}}$, and the fact that both the fidelity and the Kullback favour strategies with outcomes $i$ whose normalized probability of occurrence Tr$[\rho(\vec{b})^{N} M_i]/$Tr$[M_i]$ spans the largest possible range as a function of the direction of $\vec{b}$. Now, suppose we want to use the $N$ qubits as a quantum channel of classical information. Alice prepares $N$ copies of a given state $\rho(\vec{b})$ (the classical information being encoded in the vector $\vec{b}$) and sends them to Bob, who will perform a collective measurement in order to recover as much information about $\vec{b}$ as possible. The previous results single out using, when restricted to isotropic prior distributions, only pure states $(b=1)$ to encode classical information as the optimal method. We can then easily compute the optimal capacity of this isotropic quantum channel for any $N$, to find that I\^[(N)]{} = \_2 (N+1) - \_2 e, which for large $N$ gives $\frac{\log_2 N}{N}$ bits carried per qubit. Notice that this is a purely quantum channel, no additional flow of classical information being required at any stage. Its poor capacity can be exponentially enhanced without spoiling this fact if we take into account that a pure state $\phi^{\otimes N}$ belongs to the symmetric subspace ${\cal S}^{(N)}$ of the whole Hilbert space ${\cal H}^{(N)}$. Since the dimension of ${\cal S}^{(N)}$ is $N+1$, which corresponds to the dimension of a Hilbert space ${\cal H}^{(M)}$ of $M \equiv \log_2 (N+1)$ qubits, Alice can always compress, by means of a state-independent, unitary (and thus fully reversible) transformation, the state $\phi^{\otimes N}$ to fit in $M$ qubits, that will then be transferred to Bob. In this case the capacity increases up to $1 - O(\frac{1}{\log N})$ bits per qubit, which is asymptotically the classical one (as expectable, since for any two inequivalent states $\phi$ and $\phi'$, $\phi^{\otimes N}$ and $\phi'^{\otimes N}$ become orthogonal as $N \rightarrow \infty$), and which is consistent with the Levitin-Holevo bound [@Holevo] for the classical capacity of a quantum channel. Summarizing, using the gain of information as a guide we have constructed optimal and minimal measurements on $N = 1,2$ identical copies and have shown that for isotropic distributions the maximal gain of information is achieved for pure states. Also universality of optimal measurements has been proven, since these measurements exactly coincide with those obtained in previous work, where the fidelity was taken as figure of merit. We conjecture that also for $N \geq 3$ the most informative measurements are the most faithful ones, and vice versa. G.V. acknowledges a CIRIT grant 1997FI-00068 PG. Financial support from CIRYT, contract AEN98-0431, CIRIT, contract 1998SGR-00026 and from the ESF-QIT programme is also acknowledged. S. Massar, S. Popescu, Phys. Rev. Lett. [**74**]{} (1995) 1259. R. Derka, V. Buzek, A.K. Ekert, Phys. Rev. Lett. [**80**]{} (1998) 1571. G. Vidal, J.I. Latorre, P. Pascual, R. Tarrach, Phys. Rev. A[**59**]{} 1999. J.I. Latorre, P. Pascual, R. Tarrach, Phys. Rev. Lett. [**81**]{} (1998) 1351. A. Uhlmann, Rep. Math. Phys. [**9**]{} (1976) 273. W.K. Wootters, Phys. Rev. D [**23**]{} (1981) 357. R. Josza, Jour. Mod. Optics [**41**]{} (1994) 2315. Furthermore, and due to the symmetry in the $i$ and $\vec{b}$ distributions of Bayes’ formula, eq. (\[Bayes\]), the expressions in eqs. (\[avekullback\]) and (\[deltaentro\]) are also equal to the corresponding expressions for which $f$ and $f_c$ have been traded for $P_{ap}$ and $P$. A. Peres, W.K. Wootters, Phys. Rev. Lett. [**66**]{} (1991) 1119 S. Kullback, R.A. Leibler, Ann. Math. Stat. [**22**]{} (1951) 78. S Kullback, “Information theory and statistics” (Wiley, New York, 195?). A. Hobson, Jour. Stat. Phys. [**1**]{}, No. 3 (1969) 383. E.B. Davies, IEEE Trans. Tnfor. Theory IT-24(1978)596. We have also been able to check, for an arbitrary number $N$ of copies of the unknown state, that the optimal measurements according to the fidelity (as presented in [@VLPT]) are at least locally optimal (that is, better than any other measurement which follows from infinitessimally perturbing the former) for the Kullback. L.B. Levitin, in Proc. Fourth All-Union Conf. on Information and Coding Theory, Tashkent, 1969.\ A.S. Holevo, Probl. Inf. Transmission [**9**]{} (1973) 177.

--- abstract: | The aim of this paper is to establish the asymptotic behavior of the mutual influence of the Gini index and the poverty measures by using the Gaussian fields described in Mergane and Lo(2013). The results are given as representation theorems using the Gaussian fields of the unidimensional or the bidimensional functional Brownian bridges. Such representations, when combined with those already available, lead to joint asymptotic distributions with other statistics of interest like growth, welfare and inequality indices and then, unveil interesting results related to the mutual influence between them. The results are also appropriate for studying whether a growth is fair or not, depending on the variation of the inequality measure. Datadriven applications are also available. Although the variances may seem complicated at a first sight, their computations which are needed to get confidence intervals of the indices, are possible with the help of R codes we provide. Beyond the current results, the provided representations are useful in connection with different ones of other statistics.\ $^{(1)}$ LERSTAD, Gaston Berger University, Saint-Louis, Senegal.$^{(2)}$ LSTA, Pierre and Marie Curie University, Paris VI, France.$^{(3)}$ AUST - African University of Sciences and Technology, Abuja, Nigeria$^{(4)}$ FAST, Kara University, Togo. *Corresponding author*. Gane Samb Lo. Email : gane-samb.lo@edu.ugb.sn, ganesamblo@ganesamblo.net\ Permanent address : 1178 Evanston Dr NW T3P 0J9, Calgary, Alberta, Canada. author: - '$^{(1)}$ Pape Djiby MERGANE' - '$^{(1,2,3)}$Gane Samb LO' - '$^{(4)}$ Tchilabalo Abozou Kpanzou' title: On the joint distribution of variations of the Gini index and Welfare indices --- INTRODUCTION AND MOTIVATION {#sec1} =========================== In this paper, the asymptotic behavior of the Gini inequality index (1921) is jointly studied with a general class of welfare indices within the frame of unified Gaussian fields both for in a one phase frame (fixed time) and in a two phase frame (variation between two periods). Beyond the results themselves, the obtained asymptotic representations allow future couplings of the studied statistics with other indices. These couplings will lead to joint asymptotic distributions, enabling interesting comparison and influence studies between indices.\ We begin by a survey on the Gini index, based on historical and recent works, concerning its statistical properties, its asymptotic distributions and some of its generalizations. In a second step, we will explain the notion of Gaussian fields we mentioned before.\ The Gini index (1921) has played and is playing an important role in the measurement of economic inequality since its development by Corrado Gini in the early 20th century. Besides, this index is also used in many other disciplines, including Biology (Graczyk, 2007), Astronomy (Lisker, 2008), Environment (Druckman and Jackson, 2008; Groves-Kirkby, 2009).\ Various expressions for the Gini index are given by authors such as Davidson (2009), Dorfman (1979), Duclos and Araar (2006). Extended Gini indices are also developed (see e.g., Weymark, 1981; Yitzhaki, 1983; Chakravarty, 1988). Over the years, statistical inference for the Gini index has attracted many researchers. For example, Gastwirth (1972) discussed the estimation of the index from that of the Lorenz curve. Cowell and Flachaire (2007) have developed its influence function and looked at how influenced is its non-parametric estimator to extreme values. Moni (1991) also studied the Gini measure by means of the influence function. On their part, Qin et al. (2010) constructed empirical likelihood confidence intervals for the Gini coefficient and showed that these perform well, but only for large samples. In order to improve inference based on it, Sarno (1998) proposed, in a non-parametric setting, a new stabilizing transformation for the sample Gini coefficient.\ Fakoor et al. (2011) considered non-parametric estimators of the Lorenz curve and Gini index based on a sample from the corresponding length-biased distribution, showed that such estimators are strongly consistent for the Gini index, and derived an asymptotic normality for that index. Davidson (2009) developed a reliable standard error for the plug-in estimator of the Gini index and derived an effective bias correction. Martínez-Camblor and Corral (2009) developed results on exact and an asymptotic distribution of the Gini coefficient. Asymptotic distribution of the S-Gini index is derived by Zitikis and Gastwirth (2002), who provided an explicit formula for the asymptotic variance. More on inference for the extended Gini indices can be found in, e.g., Xu (2000) and Barrett and Donald (2000).\ But the Gini’s index is one of a quite few number of inequality measures that are available in the literature. A considerable number of them has been gathered in a class named Theil-like family and studied jointly with welfare statistics. This study dit not concern the Gini’s index nor the new Zenga’s ([@zenga1]) inequality measure. Because of its great importance, a similar handling for the Gini’s measure seems to be highly recommended alongside comparison investigations.\ As mentioned above, a new approach, that is set to put the asymptotic results of indices related to welfare and inequality analysis in a unified frame of one Gaussian field, was attempted in [@merganelo13]. In that paper, a large class of inequality measures named as the Theil-like family has been jointly studied with an other general class of poverty measures known under the name of General Poverty Index (*GPI*), both with respect to a spatial (horizontal) and a time (vertical) perspective. Such an approach leads to powerful tools when comparing different indicators or their variation over the space or the time scale. Since the joint asymptotic results are expressed with respect to one common Gaussian process, the method makes easy the comparison of the results for one particular index with those for different indices or statistics using the same frame. Our aim is to offer such representations for the Gini’s index and to benefit from them, in order to have insightful relations with the GPI. These representations will be used later in a full study of all available inequality measures. In the coming Subsection \[subsec12\], we will give a full description of the probability spaces holding the representations.\ Our main results start from the complete description of the asymptotic representation of the Gini’s index in a Gaussian field and in a residual Gaussian process $\beta $ already introduced and studied in [@logs] for the fixed time scheme in Theorem \[theoGI\]. These results are extended to the two phase variation scheme in Theorem \[theodeltaGI\]. Finally, their combination with available representations, yields successful descriptions of the mutual influence of the Gini’s index and usual poverty indices including the Sen and Kakawani ones in Theorem \[theoR\]. Unlike former works on the topic, we appeal to the Bahadur Representation Theorem (see [@bahadur66]) as a tool for handling L-statistis in the lines of [@logs]. Datadriven studies are included. But beyond this, the representations will serve in connection with similar ones for different indices of interest.\ We will exclusively limit our study in the field of the welfare analyis and focus on the Gini’s index and the General poverty measure. In future works, extentions of our current results will be extended to extension of the Gini’s measures : the Generalized Gini, S-Gini, E-Gini. (see Barrett and Donald 2009 [@barrett]).\ Let us recall that we may and do measure poverty (or richness) with the help of poverty indices $J$ based on the income variable $X$. To each income, a poverty line $Z>0$ is associated. This poverty line is defined the minimum income under which an individual is declared as *poor*. Over two periods $s=1$ and $t=2,$ we say that we have a gain against poverty when $\Delta J(s,t)=J(t)-J(s) \leq 0$, or simply a growth against poverty. But this variation is not enough to describe the situation of the population, one must be sure that, meanwhile, the income did not become more *unequally* distributed, that is the appropriate inequality coefficient $I$ did not increase. One can achieve this by studying the ratio $R=\Delta J(s,t)/\Delta I(s,t)$, where $\Delta I(s,t) = I(t) - I(s)$ denotes the variation of the distribution of the income variable. To make the ideas more precise, let us suppose that we are monitoring the poverty scene on some population over the period time $[1,2]$ and let $Y=(X^1,X^2)$ be the income variable of that population at periods $1$ and $2$. Let us consider one sample of $n \geq 1$ individuals or households, and observe the income couple $Y_{j}=(X^{(2)}_{j},X^{(2)}_{j})$, $j=1,...,n$. For each period $i \in \left\{1,2\right\}$, we also denoted by $X^i_{1,n} \leq X^i_{2,n} \leq \cdots X^i_{n,n}$ the order statistics. We assume that $X ^ i $ is strictly positive, and we compute the poverty measure $J_{n}(i)$ and the inequality measure $I_{n}(i)$.\ For poverty, we consider the Generalized Poverty Index (GPI) introduced by Lo *et al.* [@loGPI] and Lo [@loGPI2013] as an attempt to gather a large class of poverty measures reviewed in Zheng [zheng]{} defined as follows for period $i$, $$J_{n}(i)=\frac{A(Q_{n}(i),n,Z(i))}{nB(Q_{n}(i))}\sum_{j=1}^{Q_{n}(i)}w(\mu _{1}n+\mu _{2}Q_{n}(i)-\mu _{3}j+\mu _{4})\text{\ }d\left( \frac{% Z(i)-X^i_{j,n}}{Z(i)}\right) \label{gpi01}$$ where $B(Q_n(.))=\sum_{j=1}^{n} w(j),$ $Z(i)$ is the poverty line at time $t=i$, $Q_n(.)$ is the number of poor, $\mu _{1},\mu _{2},\mu _{3}$ and $\mu _{4}$ are constants, $A(u,v,s),$ $w(t),$ and $d(y)$ are measurable functions of $(u,v,s) \in \mathbb{N}\times \mathbb{N} \times \mathbb{R}^{\ast}_{+},$ $t \in \mathbb{R}^{\ast}_{+},$ and $x \in (0,1).$ By particularizing the functions $A$ and $w$ and by giving fixed values to the $\mu _{i}^{\prime }s,$ we may find almost all the available indices, as we will do it later on. *In the sequel, (\[gpi01\]) will be called a poverty index (indices in the plural) or simply a poverty measure according to the economists’ terminology.*.\ This class includes the most popular indices such as those of Sen ([@sen]), Kakwani ([@kakwani]), Shorrocks ([@shorrocks]), Clark-Hemming-Ulph ([@chu]), Foster-Greer-Thorbecke ([@fgt]), etc. See Lo ([@loGPI2013]) for a review of the GPI. From the works of many authors ([@lo1], [@sl] for instance), $J_{n}(i)$ is an asymptotically sufficient estimate of the exact poverty measure $$J(i) = \int_{0}^{Z(i)} L \left(x,F_{(2),i}\right) d\left(\frac{Z(i) - x}{Z(i)}% \right)\,dF_{(2),i}(x)$$ where $F_{(2),i}$ is the distribution function of $X^{(i)} \,(i=1,2)$, and $L$ is some weight function. As for the inequality measure, we only use the Gini index $(GI)$ which is based on the Lorenz curve (1905). And, for a given date $i \in \left\{1, 2\right\}$, we denote by $$GI_n(i) = \frac{1}{\mu_n(i)}\frac{1}{n} \sum_{j=1}^{n} \left(\frac{2 j - 1}{n% } - 1 \right) X^{(i)}_{j,n}$$ the empirical measure of the Gini index (see *Greselin et al.*, [@greselin09]), and its continuous form is defined as follows $$GI(i) = \frac{1}{\mu(i)}\int_{0}^{+\infty}F_{(2),i}^{-1}(x)\left(2 F_{(2),i}(x) - 1 \right)\,dF_{(2),i}(x),$$ where $\mu(i) = \mathbb{E}(X^i)$ is the mathematical expectation of $X^i$ and $F_{(2),i}^{-1}$ denotes the generalized inverse of the *cdf* $F_{(2),i}$.\ The motivations stated above lead to the study of the behavior of $$(\Delta J_{n}(s,t),\Delta GI_{n}(s,t)),$$ defined for two periods $s<t$, as an estimate of the unknown value of $$(\Delta J(s,t),\Delta GI(s,t)).$$ Precisely confidence intervals of $$R(s,t)=\frac{\Delta J(s,t)}{\Delta GI(s,t)}$$ will be an appropriate set of tools for the study of the mutual influence of the Gini index and the poverty measures. To achieve our goal we need a coherent asymptotic theory allowing the handling of longitudinal data as it is the case here and a stochastic process approach leading to asymptotic sub-results with the help of the continuity mapping theorem. The rest of the paper is structured as follows. In the rest of this Section \[sec1\], we describe the probability space on which the asymptotic representations will take place. In Section [ABGI]{} we provide a study on the asymptotic behavior of the Gini index. Then in Section \[VGI\],a complete study of the variation of this index between two given dates is provided. And next, Section \[MI\], we treat the mutual influence of the latter on the Generalized Poverty Index (GPI) introduced by Lo *et al.* [@loGPI] and Lo [@loGPI2013]. Section [App]{} is devoted to applications of the theoretical results with datadriven examples. The paper ends with a conclusion in Section [CR]{}.\ The notation used in the paper may be seen as complicated, but knowing the following simple facts may help in making them very comprehensive. The subscript $(1)$ means that we are working un on dimension, where the randoms variables do not have a superscript. In dimension 2, we always have the subscript (2) to main functions : *cdf*’s, copulas, empirical process,etc. When followed by $i$, like $F_{(2),i}$, it refers to a margin. For example $F_{(2),1}$ is the first marginal *cdf* of $F_{(2)}$. Still in dimension 2, any superscript $i=1,2$ refers to the first coordinate of a couple.\ Notations and Probability Space {#subsec12} ------------------------------- $ $\ In this Subsection, we complete the notations we already gave and precise our probability space.\ **Univariate frame**. We are going to describe the general Gaussian field in which we present our results. Indeed, we use a unified approach when dealing with the asymptotic theories of the welfare statistics. It is based on the Functional Empirical Process (*fep*) and its Functional Brownian Bridge (*fbb*) limit. It is laid out as follows.\ When we deal with the asymptotic properties of one statistic or index at a fixed time, we suppose that we have a non-negative random variable of interest which may be the income or the expense $X$ whose probability law on $(\mathbb{R},\mathcal{B}(\mathbb{R}))$, the Borel measurable space on $\mathbb{R}$, is denoted by $\mathbb{P}_{X}.$ We consider the space $\mathcal{F}_{(1)}$ of measurable real-valued functions $f$ defined on $\mathbb{R}$ such that $$V_{X}(f)=\int (f-\mathbb{E}_{X}(f))^{2}d\mathbb{P}_{X}=\mathbb{E}(f(X)-\mathbb{E}(f(X))^{2}<+\infty ,$$ where $$\mathbb{E}_{X}(f)=\mathbb{E}f(X).$$ On this functional space $\mathcal{F}_{(1)},$ which is endowed with the $L_{2} $-norm $$\left\Vert f\right\Vert _{2}=\left( \int f^{2}d\mathbb{P}_{X}\right) ^{1/2},$$ we define the Gaussian process $\{\mathbb{G}_{(1)}(f),f\in \mathcal{F}_{(1)}\},$ which is characterized by its variance-covariance function $$\Gamma_{(1)}(f,g)=\int^{2}(f-\mathbb{E}_{X}(f))(g-\mathbb{E}_{X}(g))d\mathbb{P% }_{X},(f,g)\in \mathcal{F}_{(1)}^{2}.$$ This Gaussian process is the asymptotic weak limit of the sequence of functional empirical processes (fep) defined as follows. Let $X_{1},X_{2},...$ be a sequence of independent copies of $X$. For each $n\geq 1,$ we define the functional empirical process associated with $X$ by $$\mathbb{G}_{n,(1)}(f)=\frac{1}{\sqrt{n}}\sum_{i=1}^{n}(f(X_{i})-\mathbb{E}% f(X_{i})),f\in \mathcal{F}_{(1)},$$ and denote the integration with respect to the empirical measure by $$\mathbb{P}_{n,(1)}(f)=\frac{1}{n} \sum_{i=1}^{n}(f(X_{i}), \ f\in \mathcal{F}_{(1)},$$ Denote by $\ell^{\infty }(T)$ the space of real-valued bounded functions defined on $T=\mathbb{R}$ equipped with its uniform topology. In the terminology of the weak convergence theory, the sequence of objects $\mathbb{G}_{n,(1)}$ weakly converges to $\mathbb{G}_{(1)}$ in $\ell^{\infty}(\mathbb{R})$, as stochastic processes indexed by $\mathcal{F}_{(1)}$, whenever it is a Donsker class. The details of this highly elaborated theory may be found in Billingsley [@billingsley], Pollard [@pollard], van der Vaart and Wellner [@vaart] and similar sources.\ we only need the convergence in finite distributions which is a simple consequence of the multivariate central limit theorem, as described in Chapter 3 in Lo [@wcia-srv-ang].\ We will use the Renyi’s representation of the random variable $X_i$’s of interest by means (*cdf*) $F_{(1)}$ as follows $$X=_{d}F_{(1)}^{-1}(U),$$ where $U$ is a uniform random variable on $(0,1)$, $=_{d}$ stands for the equality in distribution and $F^{-1}_{(1)}$ is the generalized inverse of $F_{(1)}$, defined by $$F_{(1)}^{-1}(s)=\inf \{x, F_{(1)}(x)\geq s\}, \ s \in (0,1).$$ Based on these representations, we may and do assume that we are on a probability space $(\Omega,\mathcal{A},\mathbb{P})$ holding a sequence of independent $(0,1)$-uniform random variables $U_1$, $U_2$, ..., and the sequence of independent observations of $X$ are given by $$X_{1}=F_{(1)}^{-1}(U_1), \ \ X_{2}=F_{(1)}^{-1}(U_2), \ \ etc. \label{repRenyi}$$ For each $n\geq 1$, the order statistics of $U_1,...,U_n$ and of $X_1,...,X_n$ are denoted respectively by $U_{1,n}\leq \cdots \leq U_{n,n}$ and $X_{1,n}\leq \cdots \leq X_{n,n}$.\ To the sequences of $(U_n)_{n\geq 1}$, we also associate the sequence of real empirical functions $$\mathbb{U}_{n,(1)}(s)=\frac{1}{n} \#\{i,1\leq i \leq n, \ U_i \leq s\}, \ s\in(0,1) \ n\geq 1 \label{empiricalfunctionU}$$ and the the sequence of real uniform quantile functions $$\mathbb{V}_{n,(1)}(s)=U_{1,n}1_{(0\leq s \leq 1/n)}+\sum_{j=1}^{n}U_{j,n}1_{((i-1)/n\leq s \leq (i/n))}, \ s\in(0,1), \ n\geq 1 \label{quantilefunctionU}$$ and next, the sequence of real uniform empirical processes $$\alpha_{n,(1)}(s)=\sqrt{n}(\mathbb{U}_{n,(1)}-s), \ s\in(0,1) \ n\geq 1 \label{empiricalprocessU}$$ and the sequence of real uniform quantile processes $$\gamma_{n,(1)}(s)=\sqrt{n}(s-\mathbb{V}_{n,(1)}), \ s\in(0,1) \ n\geq 1. \label{quantileprocessU}$$ The same can be done for the sequence $(X_n)_{n\geq 1}$, and we obtain the associated sequence of real empirical procecesses a $$\mathbb{G}_{n,r,(1)}(x)=\sqrt{n} \left( \mathbb{F}_{n,(1)}(x)-F_{(1)}(x)\right), \ x\in \mathbb{R}, \ n\geq 1 \label{empiricalprocess}$$ where $$\mathbb{F}_{n,(1)}(x)=\frac{1}{n} \#\{i,1\leq i \leq n, \ X_i \leq x\}, \ x\in \mathbb{R} \ n\geq 1 \label{empiricalfunction}$$ is the associated sequence of empirical functions. We also have the associated sequence of quantile processes $$\mathbb{Q}_{n,(1)}(x)=\sqrt{n} \left( \mathbb{F}^{-1}_{(n),(1)}(s) - F^{-1}(s) \right), \ s\in (0,1), \ n\geq 1 \label{quantileprocesses}$$ where, for $n\geq 1$, $$\mathbb{F}^{-1}_{n,(1)}(s)=X_{1,n}1_{(0\leq s \leq 1/n)}+\sum_{j=1}^{n}X_{j,n}1_{((i-1)/n\leq s \leq (i/n))}, \ s\in(0,1), \label{quantilefunction}$$ is the associated sequence of quantile processes.\ By passing, we recall that $\mathbb{F}^{-1}_{n,(1)}$ is actually the generalized inverse of $\mathbb{F}_{(n),(1)}$ and for the uniform sequence, we have $$\mathbb{V}_{n,(1)}=\mathbb{U}^{-1}_{n,(1)} \label{invCDFEMP}$$ In virtue of Representation (\[repRenyi\]), we have the following remarkable relations $$\mathbb{G}_{n,r,(1)}(x)=\alpha_{n,(1)}(F_{(1)}(x)), \ x\in \mathbb{R} \label{EmpEmpprocess}$$ and $$\mathbb{Q}_{n,(1)}(x)=\sqrt{n}\left( F^{-1}_{(1)}(\mathbb{V}_{n,(1)}(s))- F^{-1}_{(1)}(s)\right) \ s\in (0,1), \ n\geq 1, \label{QQprocess}$$ We also have the following relations between the empirical functions and quantile functions $$\mathbb{F}_{n,(1)}(x)=\mathbb{U}_{n,(1)}(F_{(1)}(x)), \ x\in \mathbb{R} \label{EEFunction}$$ and $$\mathbb{F}^{-1}_{n,(1)}(s)=F^{-1}_{(1)}(\mathbb{V}_{(n),(1)}(s)), \ s\in (0,1), \ n\geq 1. \label{QQFunction}$$ As well, the real and functional empirical processes are related as follows : for $n\geq 1$, $$\mathbb{G}_{n,r,(1)}(x)=\mathbb{G}_{n,(1)}(f_{x}^{\ast}), \ \alpha_{n,(1)}(s)=\mathbb{G}_{n,(1)}(f_s), \ s \in (0,1) \ x \in \mathbb{R}, \label{empiricalprocessRealFonct}$$ where for any $x \in \mathbb{R}$, $f_{x}^{\ast}=1_{]-\infty,x]}$ is the indicator function of $]-\infty,x]$ and for $s \in (0,1)$, $f_s=1_{[0,s]}$.\ To finish the description, a result of Kiefer-Bahadur (See [@bahadur66]) that says that the addition of the sequences of uniform empirical processes and quantiles processes (\[empiricalprocessU\]) and (\[quantileprocessU\]) is asymptotically, and uniformly on $[0,1]$, zero in probability, that is $$\sup_{s\in [0,1]} \left\vert \alpha_{n,(1)}(s)+\gamma_{n,(1)}(s) \right\vert =o_{\mathbb{P}}(1) \text{ as } n\rightarrow +\infty. \label{bahadurRep}$$ This result is a powerful tool to handle the rank statistics when our studied statistics are $L$-statistics.\ **Bivariate frame**. As to the bivariate case, we use the Sklar’s theorem (See [@sklar]). Let us begin to define a copula in $\mathbb{R}^2$ as bivariare probability distribution function $C(u,v)$, $(u,v)\in \mathbb{R}^2$ with support $[0,1]^2$ and with $[0,1]$-uniform margins, that is $$C(u,v)=0 \text{ for } (u,v)\in]-\infty,0[\times \mathbb{R} \text{ for } (u,v)\in \mathbb{R} \times ]-\infty,0[.$$ Let us denote by $F_{(2)}$ the bivariate distribution function of our random couple $Y=(X^{(1)}, X^{(2)})$ and by $F_{(21)}$ and $F_{(22)}$ its margins, which are the *cdf* of $X^{(1)}$ and $X^{(2)}$ respectively. The Sklar’s theorem ([@sklar]) says that there exists a copula $C_{(2)}$ such that we have $$F_{(2)}(x,y)=C_{(2)}(F_{(21)}(x), F_{(22)}(y)), \text{ for any } (x,y)\in \mathbb{R}^2. \label{theoSklar}$$ This copula is unique if the marginal *cdf*’s are continuous. In this paper, we will suppose that the marginal *cdf*’s are continuous and then $C_{(2)}$ is unique andfixed for once. By the Kolmogorov Theorem, there exists a probability space $(\Omega,\mathcal{A},\mathbb{P})$ holding a sequence of independent random couples $(U^{(1)}_i,U^{(2)}_n)$, $n\geq 1$, of common bivariate distribution function $C_{(2)}$. On that space the random couples $(F_{(21)}^{-1}(U^{(1)}_n), \ F_{(22)}^{-1}(U^{(2)}_n))$ are independent and have a common bivariate distribution function equal to $C_{(2)}$, since $$\begin{aligned} &&\mathbb{P}(F_{(21)}^{-1}(U^{(1)}_i)\leq x_{1}, \ F_{(22)}^{-1}(U^{(2)}_i)\leq x_{2})\\ &=&\mathbb{P}(U^{(1)}_i \leq F_{(21)}(x_{1}), \ U^{(2)}_i\leq F_{(22)}(x_{2}))\\ &=&C_{(2)}(F_{(21)}(x_{1}), \ F_{(22)}(x_{2}))\\ &=& F_{(2)}(x_{1}, \ x_{2}),\end{aligned}$$ by (\[theoSklar\]), and where we applied the general formula for generalized inverses functions for a *cdf* : $$F^{-1}(s) \leq y \Leftrightarrow s \leq F(x), \text{ for } (s,x) \in [0,1]\times \mathbb{R}.$$ For more on interesting properties of generalized inverses of monotone functions, see [@wcia-srv-ang], Chapter 4.\ Based on this remark, we place ourselves on the probability space holding the sequence of independent random couples $(U^{(1)},U^{(2)})$, $(U^{(1)}_n,U^{(2)}_n)$, $n \geq 2$, with common distribution function $C_{(2)}$, and the observations from $Y=(X^{(1)}, X^{(2)})=(F_{(2),1}^{-1}(U^{(1)}),F_{(2),2}^{-1}(U^{(2)}))$, are generated as follows : $$\label{renyidim2} Y_n=(F_{(21)}^{-1}(U^{(1)}_n),F_{(22)}^{-1}(U^{(2)}_n)), \ n\geq 1.$$ In this setting, we rather use the the bidimensional functional empirical process based on $\left\{\left((U_{i}^{(1)},U_{i}^{(2)})\right)\right\}_{i=1,\ldots,n}$ and defined by $$\mathbb{T}_{n,(2)}\left(f\right) = \frac{1}{% \sqrt{n}}\,\sum_{j=1}^{n}\,\left( f\left((U_{i}^{(1)},U_{(i)}^{2})\right) - \mathbb{P}% _{\left((U^{(1)},U^{(2)})\right)}\left(f\right)\right), \label{empProcUV}$$ whenever $f$ is a function of $(u,v)\in [0,1]^2$ such that $\mathbb{E}(f(U^{(1)},U^{(2)})^2)$.\ For any Donsker class $\mathcal{F}_{(2)}([0,1]^2)$, the stochastic process $\mathbb{T}_{n,(2)}$ converges to a Gaussian process $\mathbb{T}$ with variance-covariance function, for $(f,g) \in \mathcal{F}_{(2)}^{2}([0,1]^2)$, denoted by $\Gamma^\ast\left(f,g\right)$, is given by $$\label{Gamma} \int_{[0,1]^2}\left(f(u,v) - \mathbb{P}_{\left(U^{(1)},U^{(2)}\right)}\left(f\right)\right)\left(g(u,v) - \mathbb{P}_{\left(U^{(1)},U^{(2)}\right)}\left(g\right)\right)\,dC(u,v)$$ where $$\mathbb{P}_{\left(U^{(1)},U^{(2)}\right)}\left(f\right)=\mathbb{E}\left( f\left(U^{(1)},U^{(2)}\right) \right)=\int_{[0,1]^2}\,f(u,v)\,dC(u,v).$$ Another form of the variance-covariance function $\ref{Gamma}$ is also $$\label{gamma2} \gamma_{(2)}(f,g)= \int_{[0,1]^2}f(s)g(t)\left(C(s,t) - s\,t\right)\,ds\,dt$$ By deciding to use the functional empirical process based on the observations provided by the Copula C, the functional empirical process based on the incomes and defined by $$\mathbb{G}_{n,(2)}\left(g\right) = \frac{1}{% \sqrt{n}}\,\sum_{j=1}^{n}\,\left( g\left((X_{j}^{(1)},X_{(j)}^{2})\right) - \mathbb{P}% _{\left((X^{(1)},X^{(2)})\right)}\left(g\right)\right), \label{empProcXY}$$ for any function $g$, defined on $\mathcal{V}_X^{2}$ such that $\mathbb{E}(g(X^{(1)},X^{(2)})^2)<+\infty$ is not used directly. Instead, by using Representation (\[renyidim2\]), we have $$\large \mathbb{G}_{n,(2)}\left(g\right) = \frac{1}{\sqrt{n}}\,\sum_{j=1}^{n}\,\left( g\left(F^{-1}_{(2),1}(U_{j}^{(1)}),F^{-1}_{(2),1}(U_{j}^{(2)})\right) - \mathbb{P}_{\left((U^{1},U^{2})\right)}\left(g((F^{-1}_{(2),1}(.),F^{-1}_{(2),2}(.))\right)\right).$$ Hence the correspondence between the function $g$ in Formula (\[empProcXY\]) and $f$ in Formula in (\[empProcUV\]) is the following. $$f(s,t)=g\left(F^{-1}_{(2),1}(s),F^{-1}_{(2),2}(t)\right), \ (s,t) \in [0,1]^2, \label{empProcXYUV}.$$ All the needed notation are now complete and will allow the expression of the asymptotic theory we undertake here.\ The asymptotic behavior of the Gini Index {#ABGI} ========================================= Let $X$ denote the income random variable of one given population with a positive mean $\mu=\mathbb{E}(X)$ and let $\mathcal{V}_X$ denote its support. $$\label{formgini1} GI_n = \frac{1}{\mu_n}\left(\frac{1}{n}\sum_{j=1}^{n} \left(\frac{2 j - 1}{n} -1 \right)X_{j,n}\right).$$ Set $$\label{formAn1} A_n = \frac{1}{n}\sum_{j=1}^{n}\frac{j}{n} X_{j,n}.$$ We can write this expression of as $$\label{formAn2} A_n = \frac{1}{n}\sum_{j=1}^{n}\mathbb{F}_{n,(1)}(X_j)X_j.$$ Formula (\[formgini1\]) becomes $$\label{formgini2} GI_n = \frac{2 A_n}{\mu_n} - 1 - \frac{1}{n}.$$ Before tackling this study, let us first introduce some notations: $$\forall x \in \mathcal{V}_X, h(x) = x F_{(1)}(x),\; I_d(x) = x;$$ $$\forall s \in (0,1), \ell(s) = F_{(1)}^{-1}(s), f_{s}(x) = \mathbf{1}_{(0 ,F_{(1)}^{-1}(s))}(x),$$ And finally, set for real-valued measurable functions $f$ and $g$ defined on $\mathbb{R}$ $$\label{gamma1} \gamma_1(f,g)=\int_0^1 \int_0^1 f(s) g(t)\left( \min(s,t) - s\,t \right)\,ds\,dt$$ Now, we have the following theorems for the asymptotic behavior, the first concerns the statistic $A_n$ and the second concerns that of $GI_n. $ Let us state first the following lemma of the representation. \[lemma1\] Define $$\beta_{n,(1)}(\ell) = \int_0^1 \ell(s)\mathbb{G}_{n,(1)}(f_s)\,ds.$$ The, the statistic $A_n$ can be represented as follows $$A_n = \mathbb{P}_{n,1}(h) + \frac{1}{\sqrt{n}}\beta_{n,(1)}(\ell) + o_p(n^{-1/2}).$$ By decomposing the equation (\[formAn2\]), we get $$A_n = \frac{1}{n}\sum_{j=1}^{n)}F_{(1)}(X_j)X_j + \frac{1}{n}\sum_{j=1}^{n}% \left(\mathbb{F}_{n,(1)}(X_j) - F_{(1)}(X_j) \right)X_j.$$ Let us denote the residual term by $$Re_n = \frac{1}{n}\sum_{j=1}^{n}\left(\mathbb{F}_{n,(1)}(X_j) - F_{(1)}(X_j) \right)X_j$$ then $$Re_n = \sum_{j=1}^{n} \int_{\frac{j-1}{n}}^{\frac{j}{n}}\left\{ \mathbb{F}_{n,(1)}(\mathbb{F}_{n,(1)}^{-1}(s)) - F_{(1)}(\mathbb{F}_{n,(1)}^{-1}(s))\right\}\,\mathbb{F}_{n,(1)}^{-1}(s) \,ds,$$ and so $$\label{Rn} Re_n = \int_{0}^{1}\left\{ F_{n,(1)}(F_{n,(1)}^{-1}(s)) -F_{(1)}(F_{n,(1)}^{-1}(s))\right\}\,F_{n,(1)}^{-1}(s) \,ds.$$ By using Formulas (\[invCDFEMP\]), (\[EmpEmpprocess\]) and (\[QQprocess\]), we get $$\sqrt{n} Re_n = - \int_{0}^{1} \sqrt{n} \left\{\mathbb{U}_{n,(1)}\left( \mathbb{V}_{n,(1)}(s) \right) - \mathbb{V}_{n,(1)}(s)\right\}\,F_{n,(1)}^{-1} \left(\mathbb{V}_{n,(1)}(s)\right)\, ds$$ $$= - \int_{0}^{1} \sqrt{n} \left( s - \mathbb{V}_{n,(1)}(s)\right)F_{(1)}^{-1}\left(\mathbb{V}_{n,(1)}(s)\right)\, ds$$ $$- \int_{0}^{1} \sqrt{n} \left( \mathbb{U}_{n,(1)}\left(\mathbb{V}_{n,(1)}(s)\right) -s\right)\,G^{-1}\left(\mathbb{V}_{n,(1)}(s)\right)\, ds.$$ From Shorack and Wellner [@shorackwellner] (page 585), we have $$\sup_{0\leq s\leq 1} \left| \mathbb{U}_{n,(1)}\left(\mathbb{V}_{n,(1)}(s) \right) -s \right| \leq \frac{1}{n}.$$ Using the notations $\alpha_{n,(1)}$ and $\gamma_{n,(1)}$, we get $$\sqrt{n} Re_n = - \int_{0}^{1} \sqrt{n} \left( s - \mathbb{V}_{n,(1)}(s)\right)F_{(1)}^{-1}\left(\mathbb{V}_{n,(1)}(s)\right)\, ds + o_p(1)$$ $$= - \int_{0}^{1}\gamma_{n,(1)}(s)F_{(1)}^{-1}(s)\,ds + o_p(1)$$ By using the following Bahadur’s representation (See Formula \[bahadurRep\]) and by applying Formula \[empiricalprocessRealFonct\], we get $$\sqrt{n} Re_n = \int_{0}^{1} \mathbb{G}_{n,(1)}(f_s)\,\ell(s)\,ds + o_p(1),$$ then by identification we get $$\beta_{n,(1)}(\ell) = \int_{0}^{1} \mathbb{G}_{n,(1)}(f_s)\,\ell(s)\,ds,$$ which closes the proof. Here is the full representation of the asymptotic distribution of Gini’s statistic. \[theoA\] Let $\mathbb{P}_X(h^2)$ is finite and the function $% \ell$ is bounded, then when $n$ tends to $\infty,$ $\sqrt{n}\left(A_n - \mathbb{P}_X(h) \right) \rightarrow_d\; \mathbb{A}(h) = \mathbb{G}_{(1)}(h) + \beta_{(1)}(\ell)$, with $\mathbb{A}(h) \rightsquigarrow \mathcal{N}% \left(0,\sigma^2_A\right)$, $$\sigma^2_A = \Gamma(h,h) + \Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) + 2 \Gamma(h,\beta_{(1)}(\ell)), \label{gammaG}$$ with $$\Gamma(h,h) = \int \left(h(x)-\mathbb{P}_X\left(h\right) \right)^2\,dF_{(1)}(x), \label{gamma}$$ $$\Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) = \gamma_1(\ell,\ell), \label{gamman1},$$ where the function $\gamma_1(.,.)$ is defined in Formula (\[gamma1\]), and $$\Gamma(h,\beta_{(1)}(\ell)) = \int_0^1 \ell(s)\left(\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} h(x)\,dF_{(1)}(x)\right)\,ds - \left(\mathbb{P}_X(h) \right)^2.$$ By using the previous lemma, its easy to see that $$\sqrt{n} \left(A_n - \mathbb{P}_X(h) \right) = \mathbb{G}_{n,(1)}(h) + \beta_{n,(1)}(\ell) + o_p(1)$$ which tends to a centered Gaussian process $\mathbb{A}(h) = \mathbb{G}_{(1)}(h) + \beta_{(1)}(\ell)$. Now, let us find the variance of this centered process. We have $$\sigma^2_A = \mathbb{E}\left(\left(\mathbb{G}_{(1)}(h) + \beta_{(1)}(\ell)\right)^2\right)$$ $$\begin{aligned} \sigma^2_A &=& \mathbb{E}\left(\mathbb{G}_{(1)}(h)^2 \right) + \mathbb{E}\left(\beta_{(1)}(\ell)^2 \right)+ 2 \mathbb{E}\left(\mathbb{G}_{(1)}(h)\beta_{(1)}(\ell)\right)\\ &\equiv& \Gamma(h,h) + \Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) + 2\Gamma(h,\beta_{(1)}(\ell)). \end{aligned}$$ By Equation (\[gamma\]) of the definition of the covariance function, we find $$\Gamma(h,h) = \int \left(h(x)-\mathbb{P}_X\left(h\right) \right)^2\,dG(x).$$ Let us compute now the remaining terms as follows. $$\Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) = \mathbb{E}\left(\int_0^1\int_0^1\ell(s)% \ell(t) \mathbb{G}_{(1)}(f_s)\mathbb{G}_{(1)}(f_t)\,ds\,dt \right)$$ which gives, by applying Fubini’s Theorem, $$\Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) = \int_0^1\int_0^1 \ell(s)\ell(t)\mathbb{E}% \left( \mathbb{G}_{(1)}(f_s)\mathbb{G}_{(1)}(f_t)\right)\,ds\,dt.$$ Since we have $$\mathbb{E}\left( \mathbb{G}_{(1)}(f_s)\mathbb{G}_{(1)}(f_t)\right) = \min\left(s,t \right) - s\,t,$$ we get $$\Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) = \int_0^1\int_0^1 \ell(s)\ell(t)\left(\min\left(s,t \right) - s\,t \right)\,ds\,dt.$$ By Equation (\[gamma1\]), we have $$\Gamma(\beta_{(1)}(\ell),\beta_{(1)}(\ell)) = \gamma_1(\ell,\ell).$$ For $\Gamma(h,\beta_{(1)}(\ell)) = \mathbb{E}\left(\mathbb{G}_{(1)} (h)\beta_{(1)}(\ell)\right),$ we obtain $$\Gamma(h,\beta_{(1)}(\ell)) = \int_0^1 \ell(s) m(h,f_s)\,ds$$ with $$\begin{aligned} m(h,f_s) &=& \mathbb{E}\left(\mathbb{G}_{(1)}(h)\mathbb{G}_{(1)}(f_s)\right)\\ &=&\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} h(x)\,dF_{(1)}(x)- s\,\mathbb{P}_X(h).\end{aligned}$$ Finally, we conclude that $$\label{gammahb} \Gamma(h,\beta_{(1)}(\ell)) = \int_0^1 \ell(s)\left(\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} h(x)\,dF_{(1)}(x)- s\,\mathbb{P}_X(h) \right)\,ds.$$ This completes the proof of Theorem \[theoA\]. For the last part of this section, let’s define the continuous form of the Gini index as follows $$GI = \frac{2\,A}{\mu} - 1 \, \text{ with } A = \mathbb{P}_X(h).$$ Then we are able to expose the following Theorem. \[theoGI\] Assume that $\mu \neq 0,$ $\ell$ is bounded, the quantities $% \mathbb{P}_X(h^2)$ and $\mathbb{P}_X(I_d^2)$ are finite, then the following assertion holds : $$\sqrt{n}\left(GI_n - GI\right) =\frac{2}{\mu}\left(\mathbb{G}_{n,(1)}(h-\frac{A}{\mu}I_d)+\beta_{n,(1)}(\ell) \right) + o_p(1).$$ This quantity tends to a centered Gaussian process with variance $% \sigma^2_{GI}$ which is giving by $$\sigma^2_{GI} = \frac{4}{\mu^2} \left(\sigma^2_A + \frac{A^2}{\mu^2} \Gamma(I_d,I_d) -\frac{2A}{\mu}\left(\Gamma(h,I_d) + \Gamma(I_d,\beta_{(1)}(\ell))\right) \right)$$ where $\sigma^2_A$ is giving in Theorem \[theoA\]; $\Gamma(I_d,I_d) = \mathbb{E}\left((X - \mu)^2\right)$ is the variance of the random variable $X$; $$\Gamma(h,I_d) = \int \left(h(x) - \mathbb{P}_X(h) \right)\left(x - \mu \right)\,dF_{(1)}(x);$$ $$\Gamma(I_d,\beta_{(1)}(\ell)) = \int_0^1 \ell(s)\left(\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} x\,dF_{(1)}(x)\right)\,ds - A\,\mu.$$ $$\sqrt{n}\left(GI_n - GI\right) = 2\left( \frac{\sqrt{n}\left(A_n - A\right)}{% \mu_n} - \frac{A}{\mu\mu_n}\sqrt{n}\left(\mu_n - \mu\right)\right) - \frac{1% }{\sqrt{n}}$$ $$=\frac{2}{\mu}\left(\mathbb{G}_{n,(1)}(h) +\beta_n(\ell) - \frac{A}{\mu}\mathbb{G}_{n,(1)}(I_d) \right) + o_p(1)$$ which tends to a centered Gaussian process $\frac{2}{\mu} \left( \mathbb{A} - \frac{A}{\mu}\mathbb{G}_{(1)}(I_d)\right)$. Compute now the expression of the variance. We get $$\sigma^2_{GI} = \mathbb{E}\left(\left(\frac{2}{\mu} \left( \mathbb{A} - \frac{A}{\mu}\mathbb{G}_{(1)}(I_d)\right)\right)^2 \right)$$ $$= \frac{4}{\mu^2} \left(\sigma^2_A + \frac{A^2}{\mu^2} \Gamma(I_d,I_d) -% \frac{2A}{\mu}\left(\Gamma(h,I_d) + \Gamma(I_d, \beta_{(1)}(\ell) )\right) \right).$$ Applying the equation (\[gamma\]) to the functions $h$ and $I_d,$ we obtain the expression of $\Gamma(h,h),$ $\Gamma(I_d,I_d)$ and $% \Gamma(h,I_d).$ And by replacing the function $h$ by $I_d$ in equation [gammahb]{} we obtain $$\Gamma(I_d, \beta_{(1)}(\ell)) = \int_0^1 \ell(s)\left(\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} I_d(x)\,dF_{(1)}(x)- s\,\mathbb{P}_X(I_d) \right)\,ds ,$$ but, remember that $I_d$ is the identity function and $$\mathbb{P}_X(I_d) \int_0^1 s\,\ell(s)\,ds = \mu\,A$$ then $$\Gamma(I_d,\beta_{\ell}) = \int_0^1 \ell(s)\left(\int_{\left(x\leq F_{(1)}^{-1}(s)\right)} x\,dF_{(1)}(x)\right)\,ds - A\,\mu.$$ This completes the proof of this part. Let us move to the variation of the Gini’s statistics. Variation of the Gini Index between two dates {#VGI} ============================================= We fully use the setting described in Subsection \[subsec12\] regarding the two phase approach. We need to adapt the notation and the results found in Theorems \[theoA\] and \[theoGI\] to follow the consequences of the moving from Formula (\[empProcXY\]) to Formula (\[empProcUV\]) through Formula (\[empProcXYUV\]). Accordingly, define $\forall$ $(u,v)\, \in\, [0,1]^2$ and $% \forall\, j\,=\,1,\,2:$ $$\ell_j(u) = F^{-1}_{(2),j}(s),\; h_j(u) = u\, \ell_j(u); \; L_j(u) = \frac{2}{% \mu(j)} \ell_j(u);$$ $$\tilde{f}_j(u,v) = \ell_j \circ \pi_j(u,v), \; f_{j,h}(u,v) = h_j \circ \pi_j(u,v),$$ $$f_{j,s}(u,v) = \pi_i\left(\mathbf{1}_{(0\leq s)}(u),\mathbf{1}_{(0\leq s)}(v)\right)$$ where $\pi_j$ is the $j^{th}$ projection; $$F^\ast_{j,h}(u,v) = \frac{2}{\mu(j)} f_{j,h}(u,v);$$ $$F^\ast(u,v) = F^\ast_{2,h}(u,v) - F^\ast_{1,h}(u,v);$$ $$\tilde{F}^\ast(u,v) = 2\left(\frac{A(2)}{\mu(2)^2}\tilde{f}_2(u,v) - \frac{% A(1)}{\mu(1)^2}\tilde{f}_1(u,v)\right).$$ Let $$\beta^\ast_{n,(2)}(L) = \int_[0,1] \left(L_2(s) \mathbb{G}_{n,(2)}(f_{2,s}) - L_1(s) \mathbb{G}_{n,(2)}(f_{1,s}) \right)\,ds$$ be the bidimensional residual process.\ We can now expose our main theorem which concerns the variation of the Gini Index. \[theodeltaGI\] Assume that, for all $j = 1,2,$ $\mu(j)$ is finite and not null; $L_j$ is bounded; the functions $f_{j,s},$ $f_{j,h},$ $\tilde{f}% _{j},$ $F^\ast$ and $\tilde{F}^\ast$ are square integrable, then we have the following convergence in distribution as $n$ tends to infinity: $$\sqrt{n}\left( \Delta GI_n(1,2) - \Delta GI(1,2) \right) \rightarrow_d \mathcal{G}_{\Delta GI} \rightsquigarrow \mathcal{N}\left(0,\sigma^2_{\Delta GI} \right)$$ with $$\sigma^2_{\Delta GI} = \Gamma^\ast\left(F^\ast,F^\ast \right) + \Gamma^\ast\left(\tilde{F}^\ast,\tilde{F}^\ast \right) + \Gamma^\ast\left(\beta^\ast_L,\beta^\ast_L\right)$$ $$-2 \left( \Gamma^\ast\left(F^\ast,\tilde{F}^\ast \right) + \Gamma^\ast\left(\tilde{F}^\ast,\beta^\ast_L \right) - \Gamma^\ast\left(F^\ast,\beta^\ast_L \right) \right)$$ where $$\Gamma^\ast\left( \beta^\ast_L,\beta^\ast_L\right) = \gamma_1(L_1,L_1) + \gamma_1(L_2,L_2) - 2 \gamma_2(L_1,L_2);$$ $$\Gamma^\ast\left(F^\ast,\beta^\ast_L \right) = \int_{[0,1]}\left\{L_2(s)\int_{[0,1]\times(0,s)} F^{\ast}(u,v) dC(u,v) \right\}\,ds$$ $$- \int_{[0,1]}\left\{ \L _1(s)\int_{(0,s)\times [0,1]} F^{\ast}(u,v) dC(u,v) \right\} ds$$ $$-2\, \left(\frac{A(2)}{\mu(2)} - \frac{A(1)}{\mu(1)} \right) \, \mathbb{P}% _{(U^{(1)},U^{(2)})}\left(F^{\ast}\right)%\int_{D} s(L_2(s)-L_1(s))ds;$$ $\Gamma^\ast\left(\tilde{F}^\ast,\beta^\ast_L \right)$ is obtained by replacing ${F}^\ast$ by $\tilde{F}^\ast$ in the previous expression. And we get the covariances of $\Gamma^\ast\left(F^\ast,F^\ast \right)$, $\Gamma^\ast\left(\tilde{F}^\ast,\tilde{F}^\ast \right)$ and $% \Gamma^\ast\left(F^\ast,\tilde{F}^\ast \right)$ by the equation ([Gamma]{}). We get $$\sqrt{n}\left( \Delta GI_n(1,2) - \Delta GI(1,2) \right)$$ $$= \frac{2}{\mu(2)}\left( \mathbb{T}_{n,(2)}\left(f_{2,h}\right) + \beta^{\ast}_{n,(2)}\left(\ell_2\right) - \frac{A(2)}{\mu(2)}\mathbb{T}_{n,(2)}\left(% \tilde{f}_2\right)\right)$$ $$-\frac{2}{\mu(1)}\left( \mathbb{T}_{n,(2)}\left(f_{1,h}\right) + \beta^{\ast}_{n,(2)}\left(\ell_1\right) - \frac{A(1)}{\mu(1)}\mathbb{T}_{n,(2)}\left(% \tilde{f}_1\right)\right) + o_p(1)$$ $$= 2\mathbb{T}_{n,(2)}\left(\frac{f_{2,h}}{\mu(2)}-\frac{f_{1,h}}{\mu(1)} \right) + 2\beta^{\ast}_{n,(2)}\left( \frac{\ell_2}{\mu(2)}-\frac{\ell_1}{\mu(1)}\right)$$ $$-2\mathbb{T}_{n,(2)}\left( \frac{A(2)}{\mu^2(2)}\tilde{f}_2-\frac{A(1)}{\mu^2(1)}% \tilde{f}_1 \right) + o_p(1).$$ We find the next expression $$\sqrt{n}\left( \Delta GI_n(1,2) - \Delta GI(1,2) \right) = \mathbb{T}_{n,(2)}\left(F^{\ast}\right) + \beta^{\ast}_{n,(2)}\left(L\right)$$ $$- \mathbb{T}_{n,(2)}\left(\tilde{F}^{\ast}\right) + o_p(1)$$ $$\rightarrow_d \mathcal{T}_{\Delta GI} = \mathbb{T}_{(2)}\left(F^{\ast}\right) + \beta^\ast\left(L\right) - \mathbb{T}_{(2)}\left(\tilde{F}^{\ast}\right) \rightsquigarrow \mathcal{N}\left(0,\sigma^2_{\Delta GI} \right).$$ Now, we search the expression of the variance. $$\sigma^2_{\Delta GI} = \mathbb{E}\left(\left(\mathbb{T}_{(2)}(F^\ast) + \beta^\ast(L)\right)^2 +\mathbb{T}_{(2)}\left(\tilde{F}^{\ast}\right)^2\right)$$ $$-2 \mathbb{E}\left( \left(\mathbb{T}_{(2)}(F^\ast) + \beta^\ast(L)\right) \mathbb{T}_{(2)}\left(\tilde{F}^{\ast}\right)\right)$$ $$= \mathbb{E}\left(\mathbb{T}_{(2)}(F^\ast)^2\right) + \mathbb{E}% \left(\beta^\ast(L)^2\right) + 2 \mathbb{E}\left(\mathbb{T}_{(2)} (F^\ast)\beta^\ast(L)\right)$$ $$+ \mathbb{E}\left(\mathbb{T}_{(2)}(\tilde{F}^\ast)^2\right) - 2 \mathbb{E}\left(% \mathbb{T}_{(2)}(F^\ast) \mathbb{T}_{(2)}(\tilde{F}^\ast) \right) - 2 \mathbb{E}\left(% \mathbb{T}_{(2)}(\tilde{F}^\ast) \beta^\ast(L)\right).$$ Then $$\sigma^2_{\Delta GI} = \Gamma^\ast\left(F^\ast,F^\ast \right) + \Gamma^\ast\left(\tilde{F}^\ast,\tilde{F}^\ast \right) + \Gamma^\ast\left(\beta^\ast_L,\beta^\ast_L\right)$$ $$-2 \left( \Gamma^\ast\left(F^\ast,\tilde{F}^\ast \right) + \Gamma^\ast\left(\tilde{F}^\ast,\beta^\ast_L \right) - \Gamma^\ast\left(F^\ast,\beta^\ast_L \right) \right).$$ And we get the expressions of the covariances of $% \Gamma^\ast\left(F^\ast,F^\ast \right)$, $\Gamma^\ast\left(\tilde{F}% ^\ast,\tilde{F}^\ast \right)$ and $\Gamma^\ast\left(F^\ast,\tilde{F}% ^\ast \right)$ by the equation (\[Gamma\]). For the tree rest, let’s find them. Firstly, compute $\Gamma^\ast\left( \beta^\ast_L,\beta^\ast_L\right).$ $$\Gamma^\ast\left( \beta^\ast_L,\beta^\ast_L\right) = \mathbb{E}% \left(\beta^\ast(L)^2\right)$$ $$= \mathbb{E}\left(\int_D \left(L_2(s)\mathbb{T}_{(2)}(f_{2,s}) - L_1(s)\mathbb{T}_{(2)}(f_{1,s}) \right)\,ds\right)^2$$ $$=\mathbb{E}\left(\int_{[0,1]^2} \left(L_2(s)\mathbb{T}_{(2)}(f_{2,s}) - L_1(s)\mathbb{G% }(f_{1,s}) \right) \left(L_2(t)\mathbb{T}_{(2)}(f_{2,t}) - L_1(t)\mathbb{T}_{(2)} (f_{1,t})\right)\,ds\,dt\right)$$ $$= \int_{[0,1]^2}L_2(s)L_2(t)\Gamma^\ast\left(f_{2,s},f_{2,t}\right)\,ds\,dt + \int_{[0,1]^2}L_1(s)L_1(t)\Gamma^\ast\left(f_{1,s},f_{1,t}\right)\,ds\,dt$$ $$- \int_{[0,1]^2}L_1(t)L_2(s)\Gamma^\ast\left(f_{1,t},f_{2,s}\right)\,ds\,dt - \int_{[0,1]^2}L_1(s)L_2(t)\Gamma^\ast\left(f_{1,s},f_{2,t}\right)\,ds\,dt.$$ But, we have $$\int_{[0,1]^2}L_1(t)L_2(s)\Gamma^\ast\left(f_{1,t},f_{2,s}\right)\,ds\,dt = \int_{[0,1]^2}L_1(s)L_2(t)\Gamma^\ast\left(f_{1,s},f_{2,t}\right)\,ds\,dt$$ then $$\Gamma^\ast\left( \beta^\ast_L,\beta^\ast_L\right) = \int_{[0,1]^2}L_1(s)L_1(t)\Gamma^\ast\left(f_{1,s},f_{1,t}\right)\,ds\,dt$$ $$+ \int_{[0,1]^2}L_2(s)L_2(t)\Gamma^\ast\left(f_{2,s},f_{2,t}\right)\,ds\,dt$$ $$- 2 \int_{[0,1]^2}L_1(t)L_2(s)\Gamma^\ast\left(f_{1,t},f_{2,s}\right)\,ds\,dt$$ $$\equiv \gamma_1(L_1,L_1) + \gamma_1(L_2,L_2) - 2 \gamma_2(L_1,L_2).$$ But $$\Gamma^\ast\left(f_{1,s},f_{1,t}\right) = \mathbb{E}\left(\mathbf{1}% _{(0,s)}\left(U\right)\mathbf{1}_{(0,t)}\left(U\right) \right) - st = \min(s,t) - s\,t,$$ $$\Gamma^\ast\left(f_{2,s},f_{2,t}\right) = \mathbb{E}\left(\mathbf{1}% _{(0,s)}\left(U^{(2)}\right)\mathbf{1}_{(0,t)}\left(U^{(2)}\right) \right) - st = \min(s,t) - s\,t,$$ and $$\Gamma^\ast\left(f_{1,s},f_{2,t}\right) = \mathbb{E}\left(\mathbf{1}% _{(0,s)}\left(U\right)\mathbf{1}_{(0,t)}\left(U^{(2)}\right) \right) - st = C(s,t) - s t.$$ Then by identification we get $$\gamma_1(L_1,L_1)= \int_{[0,1]^2}L_1(s)L_1(t)\left(\min(s,t) - s\,t\right)\,ds\,dt;$$ $$\gamma_1(L_2,L_2)= \int_{[0,1]^2}L_2(s)L_2(t)\left(\min(s,t) - s\,t\right)\,ds\,dt;$$ $$\gamma_2(L_1,L_2)= \int_{[0,1]^2}L_1(s)L_2(t)\left(C(s,t) - s\,t\right)\,ds\,dt.$$ Secondly, let’s find the expression of $\Gamma^\ast\left(F^\ast,% \beta^\ast_L \right).$ $$\Gamma^\ast\left(F^\ast,\beta^\ast_L \right) = \mathbb{E}\left(\mathbb{T}_{(2)}(F^\ast)\beta^\ast_L \right)$$ $$= \int_D L_2(s)\Gamma^\ast\left(F^\ast,f_{2,s} \right)\,ds - \int_D L_1(s)\Gamma^\ast\left(F^\ast,f_{1,s} \right)\,ds$$ or $$\Gamma^\ast\left(F^\ast,f_{2,s} \right) = \mathbb{E}\left(F^\ast(U^{(1)},U^{(2)}) f_{2,s}(U^{(1)},U^{(2)}) \right) - \mathbb{P}_{(U^{(1)},U^{(2)})}\left(F^\ast\right)\mathbb{P}% _{(U^{(1)},U^{(2)})}(f_{2,s})$$ $$\mathbb{P}_{(U^{(1)},U^{(2)})}(f_{2,s}) = \int_{[0,1]^2}f_{2,s}(u,v)\,dC(u,v)$$ $$= \int_{[0,1]^2} \mathbf{1}_{(0,s)}(v)\,dC(u,v) = C(1,s) = s.$$ $$\mathbb{E}\left(F^\ast(U^{(1)},U^{(2)}) f_{2,s}(U^{(1)},U^{(2)} \right) = \mathbb{E}% \left(F^\ast(U^{(1)},U^{(2)})\mathbf{1}_{(0,s)}(U^{(2)}) \right)$$ $$=\int_{[0,1]\times(0,s)} F^\ast(u,v)\,dC(u,v),$$ and so, we arrive at $$\Gamma^\ast\left(F^\ast,f_{2,s} \right) = \int_{[0,1]\times(0,s)} F^\ast(u,v)\,dC(u,v) - s\,\mathbb{P}_{(U^{(1)},U^{(2)})}\left(F^\ast\right).$$ Similarly, we get $$\Gamma^\ast\left(F^\ast,f_{1,s} \right) = \int_{(0,s)\times [0,1]} F^\ast(u,v)\,dC(u,v) - s\,\mathbb{P}_{(U^{(1)},U^{(2)})}\left(F^\ast\right).$$ Therefore, we have $$\Gamma^\ast\left(F^\ast,\beta^\ast_L \right) = \int_{[0,1]}\left\{L_2(s)\int_{[0,1]\times(0,s)} F^{\ast}(u,v) dC(u,v)\right\}\,ds$$ $$- \int_{[0,1]}\left\{\L _1(s)\int_{(0,s)\times [0,1]} F^{\ast}(u,v) dC(u,v) \right\} ds$$ $$-\mathbb{P}_{(U^{(1)},U^{(2)})}\left(F^{\ast}\right)\int_{[0,1]} s(L_2(s)-L_1(s))ds,$$ and $$\int_{[0,1]} s(L_2(s)-L_1(s))ds = 2\, \int_D \left(\frac{s\,\ell_2(s)}{\mu(2)} - \frac{s\,\ell_1(s)}{\mu(1)} \right)$$ $$=2 \left(\frac{A(2)}{\mu(2)} - \frac{A(1)}{\mu(1)} \right).$$ Finally we get the expression of $\Gamma^\ast\left(\tilde{F}% ^\ast,\beta^\ast_L \right)$ by the same way. This achieves the proof of this theorem. Mutual influence with the GPI {#MI} ============================= Remaind on the GPI ------------------ We consider a class of poverty measures called the Generalized Poverty Index (GPI) introduced by Lo *and al.* [@loGPI] as an attempt to gather a large class of poverty measures reviewed in Zheng [@zheng]. This class includes the most popular indices such as those of Sen ([@sen]), Kakwani ([@kakwani]), Shorrocks ([@shorrocks]), Clark-Hemming-Ulph ([chu]{}), Foster-Greer-Thorbecke ([@fgt]), etc. See Lo ([@loGPI2013]) for a review of the GPI. For the variation of the GPI, we need the functions $g_i$ and $% \nu_i$ provided by the theorem of the representation of the GPI [@losall]. Put accordingly with these functions: $$g_i(x) = c\left(F_{(2),i}(x)\right)q_i(x)\;\mathrm{\ and }\;\nu_i(s) =c^{\prime}(s)q_i\left(F_{(2),i}^{-1}(s)\right).$$ We define for all $(u,v)\in [0,1]^2$ $$f_{i,s}(u,v) = \pi_i\left( \mathbf{1}_{(o,s)}(u),\mathbf{1}% _{(o,s)}(v)\right),$$ $$F^{\ast}_{i,J}(u,v)=g_i\circ \tilde{f}_i(u,v)= g_i\circ F_{(2),i}^{-1}\circ \pi_i(u,v),$$ and $$F^{\ast}_{J}(u,v) = F^{\ast}_{2,J}(u,v) - F^{\ast}_{1,J}(u,v).$$ And denote the residual term for the *GPI* by $$\beta^{\ast}_{(2)}(\nu) = \int_{[0,1]} \left(\mathbb{T}_{(2)}\left(f_{2,s}\right)\nu_2(s) - \mathbb{T}_{(2)}\left(f_{1,s}\right)\nu_1(s)\right)ds.$$ \[theodeltaGPI\] Let $\mu(i)$ finite for $i=1,2$. Suppose that $\mathbb{P% }_{\left(U^{(1)},U^{(2)}\right)}\left(\left(f_{1,s}\right)^2\right),$ $\mathbb{P}% _{\left(U^{(1)},U^{(2)}\right)}\left(\left(f_{2,s}\right)^2\right)$ and $\mathbb{P}% _{\left(U^{(1)},U^{(2)}\right)}\left({F^{\ast}_{J}}^2\right)$ are finite.\ Then $\sqrt{n}\left(\Delta J_n(1,2) - \Delta J(1,2)\right)$ converges to $% \mathcal{G}_{\Delta GPI} = \mathbb{T}_{(2)}\left(F^{\ast}_{J}\right) + \beta^\ast(\nu)$ which is a centered Gaussian process of variance-covariance function: $$\sigma^2_{\Delta GPI} = \Gamma^\ast\left(F^\ast_J,F^\ast_J\right) + \Gamma^\ast\left(\beta^\ast_{\nu},\beta^\ast_{\nu} \right) + 2\, \Gamma^\ast\left(F^\ast,\beta^\ast_{\nu} \right)$$ where $$\Gamma^\ast\left(F^\ast_J,F^\ast_J\right) = \int_{[0,1]^2}\left(F^{\ast}_{J}(u,v) - \mathbb{P}_{\left(U^{(1)},U^{(2)}\right)}\left(F^{% \ast}_{J}\right)\right)^2\,dC(u,v);$$ $$\Gamma^\ast\left(\beta^\ast_{\nu},\beta^\ast_{\nu} \right) = \gamma_1(\nu_1,\nu_1) - 2\,\gamma_2(\nu_1,\nu_2) + \gamma_1(\nu_2,\nu_2)$$ with the covariance functions $\gamma_1(.,.)$ and $\gamma_2(.,.)$ are respectively defined in Equation (\[gamma1\]) and Equation ([gamma2]{}); and $$\Gamma^\ast\left(F^\ast,\beta^\ast_{\nu} \right) = \int_{[0,1]}\left\{ \nu_2(s) \int_{[0,1]\times (0,s)} F^{\ast}_J(u,v)\,dC(u,v)\right\}\,ds$$ $$- \int_{[0,1]}\left\{\nu_1(s) \int_{(0,s)\times [0,1]} F^{\ast}_J(u,v)\,dC(u,v)\right\}\,ds$$ $$- \mathbb{P}_{\left(U^{(1)},U^{(2)}\right)}\left(F^{\ast}_{J}\right)\, \int_{[0,1]} s\left(\nu_2(s) - \nu_1(s)\right)\,ds.$$ See Mergane and Lo ([@merganelo13]). We are now able to stable our main results. Mutual influence ---------------- Let $$R = \frac{\Delta J(1,2)}{\Delta GI(1,2)},\; a = \frac{1}{\Delta GI(1,2)}\: \text{ and } \: b = \frac{\Delta J(1,2)}{\left(\Delta GI(1,2)\right)^2}.$$ \[theoR\] Supposing that the above mentioned hypotheses are true, then $$\left(\sqrt{n}\left(\Delta J_n(1,2) - \Delta J(1,2) \right) , \sqrt{n}% \left(\Delta GI_n(1,2) - \Delta GI(1,2) \right) \right)^t \rightarrow_d \mathcal{N}_2\left(0,\Sigma\right),$$ with $$\Sigma = \left(% \begin{array}{ccc} \sigma^2_{\Delta GPI} & & \sigma_{\Delta GPI,\Delta GI} \\ & & \\ \sigma_{\Delta GPI,\Delta GI} & & \sigma^2_{\Delta GI} \\ & & \end{array}% \right)$$ where $$\sigma_{\Delta GPI,\Delta GI} = \Gamma^\ast\left(F^\ast_J,F^\ast\right) + \Gamma^\ast\left(F^\ast_J,\beta^\ast_L\right) - \Gamma^\ast\left(F^\ast_J,\tilde{F}^\ast\right)$$ $$+ \Gamma^\ast\left(F^\ast,\beta^\ast_{\nu}\right) + \Gamma^\ast\left(\beta^\ast_{L},\beta^\ast_{\nu}\right) - \Gamma^\ast\left(\tilde{F}^\ast,\beta^\ast_{\nu}\right)$$ with $$\Gamma^\ast\left(\beta^\ast_{L},\beta^\ast_{\nu}\right) = \gamma_1(L_1,\nu_1) + \gamma_1(L_2,\nu_2) - \gamma_2(L_1,\nu_2) - \gamma_2(L_2,\nu_1);$$ $\sigma^2_{\Delta GPI}$ and $\sigma^2_{\Delta GI}$ are given in the previous theorems. Further, $\sqrt{n}\left\{R_n(1,2) - R(1,2)\right\}$ tends to a functional Gaussian process $$a\,\mathcal{G}_{\Delta GPI}\, -\, b\,\mathcal{G}_{\Delta GI}$$ of variance-covariance function $$\sigma^2_R = a^2\,\sigma^2_{\Delta GPI} + b^2\,\sigma^2_{\Delta GI} - \,2\,a\,b\,\sigma_{\Delta GPI,\Delta GI}.$$ By Theorem \[theodeltaGI\] and Theorem \[theodeltaGPI\], it is clear that from Van der Vaart and Wellner ([@vaart], p. $81$), the bivariate random variable $$\left(\sqrt{n}\left(\Delta J_n(1,2) - \Delta J(1,2) \right) , \sqrt{n}% \left(\Delta GI_n(1,2) - \Delta GI(1,2) \right) \right)$$ is asymptotically Gaussian $\left(\mathcal{G}_{\Delta GPI}, \mathcal{G}_{\Delta GI}\right)$ with $$\sigma_{\Delta GPI,\Delta GI} = \mathbb{E}\left(\mathcal{G}_{\Delta GPI}\, \mathcal{G}_{\Delta GI}\right).$$ But let’s recall that $$\mathcal{G}_{\Delta GPI} = \mathbb{G}_{(2)}\left(F^{\ast}_{J}\right) + \beta^\ast(\nu)$$ and $$\mathcal{G}_{\Delta GI} = \mathbb{G}_{(2)}\left(F^{\ast}\right) + \beta^\ast(L) - \mathbb{G}_{(2)}\left(\tilde{F}^{\ast}\right),$$ then $$\sigma_{\Delta GPI,\Delta GI} = \mathbb{E}\left(\mathcal{G}_{\Delta GPI} \mathcal{G}_{\Delta GI} \right);$$ by expanding this, we find the following expression $$\sigma_{\Delta GPI,\Delta GI} =\Gamma^\ast\left(F^\ast_J,F^\ast\right) + \Gamma^\ast\left(F^\ast_J,\beta^\ast_L\right) - \Gamma^\ast\left(F^\ast_J,\tilde{F}^\ast\right)$$ $$+ \Gamma^\ast\left(F^\ast,\beta^\ast_{\nu}\right) + \Gamma^\ast\left(\beta^\ast_{L},\beta^\ast_{\nu}\right) - \Gamma^\ast\left(\tilde{F}^\ast,\beta^\ast_{\nu}\right),$$ and we can obtain the complete expression for each covariance by using the same procedure as in the theorems \[theodeltaGI\] and \[theodeltaGPI\]. Final Comments {#App} ============== We have shown hat the approach we used here, once set up, leads to powerful asymptotic laws. Besides, the construction we use allow to couple the results on the Gini index with results on aby other index as long as they are expressed in the current frame. We will not need to begin from scratch.\ However, the variances and co-variance may have not simple forms. But this is not a major concern in the modern time of powerful computers. For example the variance and co-variance and co-variance given here may easily be performed with the free software of R.\ To avoid to make more lengthy the paper, we decided to prepare and publish, in a very near future, papers devoted to computational methods and simulations in which we will share the codes and papers with focus on data analysis.\ [99]{} Bahadur, R.R. (1966). A note on quantiles in large samples, *Ann. Math. Stat.* **37**, pp. 577–580. MR :32:6522ZL : 0147.18805. doi:10.1214/aoms/1177699450. euclid.aoms/1177699450 Barrett, G., & Donald, S. (2009). Statistical inference with generalized Gini indices of inequality, poverty, and welfare. *J. Bus. Econom. Statist.*, **27**(1), 1-17. http://dx.doi.org/10.1198/jbes.2009.0001 Billingsley, P.(1968). *Convergence of Probability measures*. John Wiley, New-York. Clark, S., Hemming, R. and Ulph, D.(1981). On Indices for the Measurement of Poverty. *Economic Journal* **91**, 525-526. Druckman, A. and Jackson, T. (2008). Measuring resource inequalities: The concepts and methodology for an area-based Gini coefficient. *Ecological Economics,* **65**, 242-252. Foster, J., Greer, J. and Shorrocks, A.(1984). A class of Decomposable Poverty Measures. *Econometrica* **52**, 761-766. Gini C., (1921). Measurement of Inequality of Incomes. *The Economic Journal,* March: p.124-126. Graczyk, P.P. (2007). Gini Coefficient : A New Way To Express Selectivity of Kinase Inhibitors against a Family of Kinases. *Journal of Medicinal Chemistry*, **50**, 5773U5779. Greselin, F., Puri, M.L. and Zitikis, R. (2009). L-functions, processes, and statistics in measuring economic inequality and actuarial risks. *Statistics and Its Interface,* **2**, 227–245. Groves-Kirkby, C.J., Denman, A.R. and Phillips, P.S. (2009). Lorenz Curve and Gini Coefficient: Novel tools for analysing seasonal variation of environmental radon gas. *Journal of Environmental Management*, **90**, 2480-2487. Kakwani, N.(1980). On a Class of Poverty Measures. *Econometrica*, **48**, 437-446. (MR0560520) http://dx.doi.org/10.2307/1911106. Lisker, T. (2008). Is the Gini coefficient a stable measure of galaxy structure? *The Astrophysical Journal Supplement Series*, **179**, 319-325. doi:10.1086/591795 Lo, G.S., Ngom M. and Kpanzou T. A.(2016). Weak Convergence (IA). Sequences of random vectors. SPAS Books Series.(2016). Doi : 10.16929/sbs/2016.0001. Arxiv : 1610.05415 Lo, G. S.(2013). The Generalized Poverty Index. *far East Journal of Theoretical Statistics* **Vol. 42**, No. 1, pp. 1-22. Available online at http://pphmj.com/journals/fjst.htm Lo, G. S., Sall. S. and Seck, C. T.(2006). Une Théorie asymptotique des indicateurs de pauvreté. *C. R. Math. Acad. Sci. Soc. R. Can.* **31** (2009), no. 2, 45-52. (MR2535867), (2010m:91167) LO, G. S. and Mergane P. D.(2013). On the influence of the Theil-like inequality measure on the growth (2013). Applied Mathematics, 2013, 4, 986-1000. doi:10.4236/am.2013.47136, arXiv:1210.3190. Lo G. S. and Sall, S.T.(2010). Asymptotic Representation Theorems for Poverty Indices. *Afrika Statistika*., **5**, pp. 238-244. (MR2920300) Lo G. S.(2010). A simple note on some empirical stochastic process as a tool in uniform L-statistics weak laws. *Afrika Statistika*, **5**, pp. 437-446. (MR2920301) Lorenz M.O. (1975). “ Methods for Meausuring Concentration of Wealth", *Journal of American Statistical Association*, **Vol 70**, p.209-219. NELSEN, R. B. (2006). An Introduction to Copulas, Springer. Pollard D.(1984). Convergence of Stochastic Processes. Springer-Verlag, Berlin. Sall, S.T. and Lo, G.S., (2007). The Asymptotic Theory of the Poverty Intensity in View of Extreme Values Theory For Two Simple Cases. *Afrika Statistika*, **vol 2** ($n^{0}1$), p.41-55 Sall, S. T. and Lo, G. S., (2010). Uniform Weak Convergence of the time-dependent poverty Measure for Continuous Longitudinal Data. *Braz. J. Probab. Stat.*, **24**, (3), 457-467. (MR2719696)http://dx.doi.org/10.1214/08-BJPS101 Seck, C. T. and Lo, G. S. (2008). Uniform Weak Convergence of the Nonweithed Poverty measure. To appear in *Comm. Statist., A*. Sen, A. K.(1973). *on Economic Inequality.* Oxford: Clarendon Press. Sen Amartya K.(1976). Poverty: An Ordinal Approach  to Measurement. *Econometrica*, **44**, 219-231. Shorack G.R. and Wellner J.A. (1986). Empirical Processes with Applications to Statistics, wiley-Interscience, New-York. http://dx.doi.org/10.1137/1.9780898719017 Shorrocks, A. (1995). Revisiting the Sen Poverty Index. *Econometrica*, **63**, 1225-1230. (doi:10.2307/2171728) Rongve, I. and C.M. Beach (1997). Estimation and inference for normative inequality indices of distributions in the continuum. *International Economic Review,* **38** (1), 83-96. Rothschild, M. and J. E. Stiglitz (1973). Some further results on the measurement of inequality. *Journal of Economic Theory* 6, 188-203. van der Vaart, A. W. and Wellner J. A.(1996) *Weak Convergence and Empirical Processes: With Applications to Statistics*, Springer-Verlag New-York. Zheng, B.(1997). Aggregate Poverty Measures. *Journal of Economic Surveys*, **11** (2), 123-162.\ Barrett G.F. and Donald S.G. (2000). Statistical inference with generalized Gini indices of inequality and poverty. *Discussion paper* 2002/01. Sydney: School of Economics, University of New South Wales. Chakravarty S.R. (1988). Extended Gini Indices of Inequality. *International Economic Review,* 29, 147-156. Cowell F.A. and Flachaire E. (2007). Income Distribution and Inequality Measurement: The Problem of Extreme Values. *Journal of Econometrics,* 141:1044–1072. Davidson R. (2009). Reliable Inference for the Gini Index. *Journal of Econometrics,* 150:30–40. Dorfman R. (1979). A Formula for the Gini Coefficient. *The Review of Economics and Statistics,* 61:146–149. Duclos J.-Y. and Araar A. (2006). *Poverty and Equity: Measurement, Policy and Estimation with DAD.* Springer. Fakoor V., Bolbolian Ghalibaf M. and Azamoosh H.A. (2011). Asymptotic behaviors of the Lorenz curve and Gini index in sampling from a length-biased distribution. *Statistics and Probability Letters,* 81: 1425-1435. Gastwirth J.L. (1972). The estimation of the Lorenz curve and Gini index. *Review of Economics and Statistics* 54, 306-316. Moni A.C. (1991). The study of the Gini concentration ratio by means of the influence function. *Statistica,* 51, 561-577. Martínez-Camblor P. and Corral N. (2009). About the exact and asymptotic distribution of the Gini coefficient. *Revista de Matemática: Teoría y Applicaciones,* 16: 199-204. Qin Y., Rao J.N.K., and Wu C. (2010). Empirical Likelihood Confidence Intervals for the Gini Measure of Income Inequality. *Economic Modelling,* 27:1429–1435. Sarno E. (1998). A variance stabilising transformation for the Gini concentration ratio. *Journal of the Italian Statistical Society,* 1: 77-91. Weymark J.A. (1981). Generalized Gini inequality indices. *Mathematical Social Sciences,* 1: 409-430. Xu K. (2000). Inference for Generalized Gini indices using the iterated-bootstrap method. *Journal of Business and Economic Statistics,* 18: 223-227. Yitzhaki S. (1983). On an extension of the Gini inequality index. *International Economic Review,* 24: 617-628. Zenga, M. (1984). Proposta per un Indice di Concentrazione Basato sui Rapporti tra Quantili di Popolazione e Quantili di Reddito. *Giornale degli Economisti e Annali di Economia*, vol. 43, pp. 301-326. Zitikis R. and Gastwirth J.L. (2002). The asymptotic distribution of the S-Gini index. *Australian and New Zealand Journal of Statistics,* 44: 439-446. Sklar A.(1959) Fonctions de répartition à $n$ dimensions et leurs marges. *Publ. Inst. Statist Univ Paris*, 8:229-231.\

--- abstract: 'The superiorization methodology is intended to work with input data of constrained minimization problems, that is, a target function and a set of constraints. However, it is based on an antipodal way of thinking to what leads to constrained minimization methods. Instead of adapting unconstrained minimization algorithms to handling constraints, it adapts feasibility-seeking algorithms to reduce (not necessarily minimize) target function values. This is done by inserting target-function-reducing perturbations into a feasibility-seeking algorithm while retaining its feasibility-seeking ability and without paying a high computational price. A superiorized algorithm that employs component-wise target function reduction steps is presented. This enables derivative-free superiorization (DFS), meaning that superiorization can be applied to target functions that have no calculable partial derivatives or subgradients. The numerical behavior of our derivative-free superiorization algorithm is illustrated on a data set generated by simulating a problem of image reconstruction from projections. We present a tool (we call it a *proximity-target curve*) for deciding which of two iterative methods is “better” for solving a particular problem. The plots of proximity-target curves of our experiments demonstrate the advantage of the proposed derivative-free superiorization algorithm.' author: - 'Yair Censor[^1]' - 'Edgar Gardu[ñ]{}o[^2]' - 'Elias S. Helou[^3]' - 'Gabor T. Herman[^4]' title: 'Derivative-Free Superiorization: Principle and Algorithm[^5]' --- derivative-free; superiorization; constrained minimization; component-wise perturbations; proximity function; bounded perturbations; regularization 68Q25, 68R10, 68U05 Introduction\[sec:Introduction\] ================================ The superiorization methodology (SM) ------------------------------------ In many applications there exist efficient iterative algorithms for producing *constraints-compatible* solutions. Often these algorithms are *perturbation resilient* in the sense that, even if certain kinds of changes are made at the end of each iterative step, the algorithms still produce a constraints-compatible solution. This property is exploited in *superiorization* by using such perturbations to steer an algorithm to an output that is as constraints-compatible as the output of the original algorithm, but is superior (not necessarily optimal) to it with respect to a given target function. Superiorization has a world-view that is quite different from that of classical constrained optimization. Both in superiorization and in classical constrained optimization there is an assumed domain $\Omega$ and a criterion that is specified by a target function $\phi$ that maps $\Omega$ into $\mathbb{R}$. In classical optimization it is assumed that there is a constraints set $C$ and the task is to find an $\boldsymbol{x}\in C$ for which $\phi(\boldsymbol{x})$ is minimal over $C$. Two difficulties with this approach are: (1) The constraints that arise in a practical problem may not be consistent, so $C$ could be empty and the optimization task as stated would not have a solution. (2) Even for nonempty $C$, iterative methods of classical constrained optimization typically converge to a solution only in the limit and some stopping rule is applied to terminate the process. The actual output at that time may not be in $C$ (especially if the iterative algorithm is initialized at a point outside $C$) and, even if it is in $C$, it is most unlikely to be a minimizer of $\phi$ over $C$. Both issues are handled in the superiorization approach investigated here by replacing the constraints set $C$ by a nonnegative real-valued *proximity* *function* $\mathcal{P}r_{T}$ that indicates how incompatible a given $\boldsymbol{x}\in\Omega$ is with specified constraints $T$. Then the merit of an actual output $\boldsymbol{x}$ of an algorithm is represented by the smallness of the two numbers $\mathcal{P}r_{T}(\boldsymbol{x})$ and $\phi(\boldsymbol{x})$. Roughly, if an iterative algorithm produces an output $\boldsymbol{x}$, then its superiorized version will produce an output $\boldsymbol{x'}$ for which $\mathcal{P}r_{T}(\boldsymbol{x'})$ is not larger than $\mathcal{P}r_{T}(\boldsymbol{x})$, but (as in-practice demonstrated) generally $\phi(\boldsymbol{x'})$ is smaller than $\phi(\boldsymbol{x})$. As an example, let $\Omega=\mathbb{R}^{J}$ and consider a set $T$ of constraints of the form $$\left\langle \boldsymbol{d}^{i},\boldsymbol{x}\right\rangle =h_{i},\thinspace\thinspace\thinspace i=1,2,\ldots,I,\label{eq:equalities}$$ where $\boldsymbol{d}^{i}\in\mathbb{R}^{J}$ and $h_{i}\in\mathbb{R}$, for all $i=1,2,\ldots,I$, and $\left\langle \cdot,\cdot\right\rangle $ is the Euclidean inner product in $\mathbb{R}^{J}$. There may or may not be an $\boldsymbol{x}\in\mathbb{R}^{J}$ that satisfies this set of constraints, but we can always define a proximity function for $T$ as, for example, by $$\mathcal{P}r_{T}(\boldsymbol{x}):={\displaystyle \sum_{i=1}^{I}\left(\left\langle \boldsymbol{d}^{i},\boldsymbol{x}\right\rangle -h_{i}\right)^{2}}.\label{eq:residual}$$ There are several approaches in the literature that attempt to minimize both competing objectives $\mathcal{P}r_{T}(\boldsymbol{x})$ and $\phi(\boldsymbol{x})$ as a way to handle constrained minimization. The oldest one is the penalty function approach, also useful in regularization of inverse problems [@ehn00]. In that approach, the constrained minimization problem is replaced by the unconstrained minimization of the combination $\phi(\boldsymbol{x})+\pi\mathcal{P}r_{T}(\boldsymbol{x})$, in which $\pi\geq0$ is a penalty parameter that governs the relative importance of minimizing the two summands. An inherent difficulty with this is that the penalty parameter needs to be chosen by the user. The filter method approach [@filter02], among others, was developed to avoid this difficulty. Of course, people have also applied multiobjective minimization with two objectives (bi-objective minimization) to the competing objectives $\mathcal{P}r_{T}(\boldsymbol{x})$ and $\phi(\boldsymbol{x})$. None of these approaches are close in their underlying principles to the superiorization methodology employed in this paper. Derivative-free superiorization: Expanding the boundaries of superiorization and competing with derivative-free optimization ---------------------------------------------------------------------------------------------------------------------------- Our motivating purpose in this paper is to investigate the general applicability of derivative-free superiorization (DFS) as an alternative to previously proposed superiorization approaches. These earlier approaches were based on generation of nonascending vectors, for target function reduction steps, that mostly required the ability to calculate gradients or subgradients of the target function. Paralleling the body of knowledge of derivative-free optimization (DFO), see, e.g., [@Conn-book-2009], we explore a specific DFS algorithm and demonstrate its action numerically. The output of a superiorized version of a constraints-compatibility-seeking algorithm will have smaller (but not minimal) target function $\phi$ value than the output by the same constraints-compatibility-seeking algorithm without perturbations, everything else being equal. Even though superiorization is not an exact minimization method, we think of it as an applicable (and possibly, more efficacious) alternative to derivative-free constrained minimization methods applied to the same data for two main reasons: its ability to handle constraints and its ability to cope with very large-size problems. The review paper of Rios and Sahinidis [@Rios] “... addresses the solution of *bound-constrained* optimization problems using algorithms that require only the availability of objective function values but no derivative information,” with bound constraints imposed on the vector $\boldsymbol{x}$. The book by Conn, Scheinberg and Vicente [@Conn-book-2009] deals only with derivative-free unconstrained minimization, except for its last chapter (of 10 pages out of the 275) entitled “Review of constrained and other extensions to derivative-free optimization.” Li *et al*. [@LCLLLL] do not even mention constraints. In [@diniz2011] the numerical work deals with: “The dimension of the problems [\[]{}i.e., the size of the vector $\boldsymbol{x}$[\]]{} varies between 2 and 16, while the number of constraints are between 1 and 38, exceeding 10 in only 5 cases.” In [@dfo-4-oil] the numerical tests are limited to: “The first case has 80 optimization variables [\[]{}i.e., the size of the vector $\boldsymbol{x}$[\]]{} and only bound constraints, while the second example is a generally constrained production optimization involving 20 optimization variables and 5 general constraints.” Similar orders of magnitude for problem sizes appear in the numerical results presented in [@Audet-Dennis-2009] and also in the book of Audet and Hare [@Audet-book-2017]. This indicates that (i) much of the literature on derivative-free minimization is concerned with unconstrained minimization or with bound-constraints on the variables, and (ii) many, if not all, proposed methods were designed (or, at least, demonstrated) only for small-scale problems. In contrast, the DFS method proposed here can handle any type of constraints for which a separate efficient derivative-free constraints-compatibility-seeking algorithm is available and is capable of solving very large problems. In that respect, of problem sizes, we discover here, admittedly with a very preliminary demonstration, that DFS can compete well with DFO on large problems. Since the constraints-compatibility-seeking algorithm forms part of the proposed DFS method, the method can use exterior initialization (that is initializing the iterations at any point in space). Furthermore, very large-scale problems can be accommodated. The progressive barrier (PB) approach, described in Chapter 12 of the book [@Audet-book-2017], originally published in [@Audet-Dennis-2009], is an alternative to the exterior penalty (EP) approach that we mention in Subsection \[subsec:EP-approach\] below. However, the PB differs from our DFS method, in spite of some similarities with it, as we explain in Subsection \[subsec:The-progressive-barrier\] below. Earlier work on superiorization ------------------------------- A comprehensive overview of the state of the art and current research on superiorization appears in our continuously updated bibliography Internet page that currently contains 109 items [@sup-bib]. Research works in this bibliography include a variety of reports ranging from new applications to new mathematical results on the foundations of superiorization. A special issue entitled: “Superiorization: Theory and Applications” of the journal Inverse Problems [@Sup-Special-Issue-2017] contains several interesting papers on the theory and practice of SM, such as [@Cegielski-2017], [@He2017], [@Reich2017], to name but a few. Later papers continue research on perturbation resilience, which lies at the heart of the SM, see, e.g., [@Bargetz2018]. An early paper on bounded perturbation resilience is [@But06], a recent book containing many results on the behavior of algorithms in the presence of summable perturbations is [@ZAS18]. Structure of the paper ---------------------- In Section \[sect:SM\] we present the basics of the superiorization methodology. We present our DFS algorithm in Section \[sec:Specific-superiorization-approac\] and juxtapose it with an existing superiorization algorithm that uses derivative information. In Section \[sec:proximity-target-curve\] we present a tool (we call it a *proximity-target curve*) for deciding which of two iterative methods is “better” for solving a particular problem. The experimental demonstration of our DFS algorithm appears in Section \[sec:Experimental\]. In Section \[sec:Discussion-and-conclusions\] we offer a brief discussion and some conclusions. \[sect:SM\]The basics of the superiorization methodology ======================================================== We follow the approach of [@HGDC12]. $\Omega$ denotes a nonempty set in the Euclidean space $\mathbb{R}^{J}$. $\mathbb{T}$ is a *problem set*; each problem $T\in\mathbb{T}$ is described by a particular set of constraints such as provided, for example, in . $\mathcal{P}r$ is a *proximity function* on $\mathbb{T}$ such that, for every $T\in\mathbb{T}$, $\mathcal{P}r_{T}:\Omega\rightarrow\mathbb{R}_{+}$ (nonnegative real numbers). $\mathcal{P}r_{T}\left(\boldsymbol{x}\right)$ measures how incompatible $\boldsymbol{x}$ is with the constraints of $T$. A *problem structure* is a pair $(\mathbb{T},\mathcal{P}r)$, where $\mathbb{T}$ is a problem set and $\mathcal{P}r$ is a proximity function on $\mathbb{T}$. For an $\boldsymbol{x}\in\Omega$, we say that $\boldsymbol{x}$ is $\varepsilon$*-compatible* with $T$ if $\mathcal{P}r_{T}\left(\boldsymbol{x}\right)\leq\varepsilon$. We assume that we have computer procedures that, for any $\boldsymbol{x}\in\mathbb{R}^{J}$, determine whether $\boldsymbol{x}\in\Omega$ and, for any $\boldsymbol{x}\in\Omega$ and $T\in\mathbb{T}$, calculate $\mathcal{P}r_{T}\left(\boldsymbol{x}\right)$. In many applications, each problem $T\in\mathbb{T}$ is determined by a family of sets $\left\{ C_{i}\right\} _{i=1}^{I}$, where each $C_{i}$ is a nonempty, often closed and convex, subset of $\Omega$ and the problem $T$ is to find a point that is in the intersection of the $C_{i}$. We introduce $\Delta$, such that $\Omega\subseteq\Delta\subseteq\mathbb{R}^{J}$ and a *target function* $\phi:\Delta\rightarrow\mathbb{R}$, which is referred to as an optimization criterion in [@HGDC12]. We assume that we have a computer procedure that, for any $\boldsymbol{x}\in\mathbb{R}^{J}$, determines whether $\boldsymbol{x}\in\Delta$ and, if so, calculates $\phi\left(\boldsymbol{x}\right)$. An *algorithm* $\mathbf{P}$ *for a problem structure* $(\mathbb{T},\mathcal{P}r)$ assigns to each problem $T\in\mathbb{T}$ a computable *algorithmic operator* $\mathbf{P}_{T}:\Delta\rightarrow\Omega$. For any *initial point* $\boldsymbol{x}\in\Omega$, $\mathbf{P}_{T}$ produces the infinite sequence $\left(\left(\mathbf{P}_{T}\right)^{k}\boldsymbol{x}\right)_{k=0}^{\infty}$ of points in $\Omega$. **The** $\varepsilon$**-output of a sequence** For a problem structure $(\mathbb{T},\mathcal{P}r)$, a $T\in\mathbb{T}$, an $\varepsilon\in\mathbb{R}_{+}$ and a sequence $R:=\left(\boldsymbol{x}^{k}\right)_{k=0}^{\infty}$ of points in $\Omega$, we use $O\left(T,\varepsilon,R\right)$ to denote the $\boldsymbol{x}\in\Omega$ that has the following properties: $\mathcal{P}r_{T}(\boldsymbol{x})\leq\varepsilon,$ and there is a nonnegative integer $K$ such that $\boldsymbol{x}^{K}=\boldsymbol{x}$ and, for all nonnegative integers $k<K$, $\mathcal{P}r_{T}\left(\boldsymbol{x}^{k}\right)>\varepsilon$. If there is such an $\boldsymbol{x}$, then it is unique. If there is no such $\boldsymbol{x}$, then we say that $O\left(T,\varepsilon,R\right)$ is *undefined*, otherwise it is *defined*. If $R$ is an infinite sequence generated by a process that repeatedly applies $\mathbf{P}_{T}$, then $O\left(T,\varepsilon,R\right)$ is the *output* produced by that process when we add to it instructions that make it terminate as soon as it reaches a point that is $\varepsilon$-compatible with $T$. Roughly, we refer to $\mathbf{P}$ as a *feasibility-seeking algorithm* for a problem structure $(\mathbb{T},\mathcal{P}r)$ that arose from a particular application if, for all $T\in\mathbb{T}$ and $\varepsilon\in\mathbb{R}_{+}$ of interest for the application, $O\left(T,\varepsilon,R\right)$ is defined for all infinite sequences $R$ generated by repeated applications $\mathbf{P}_{T}$. Each application of $\mathbf{P}_{T}$ is referred to as a *feasibility-seeking step*. **Strong perturbation resilience** An algorithm $\mathbf{P}$ for a problem structure $(\mathbb{T},\mathcal{P}r)$ is said to be *strongly perturbation resilient* if, for all $T\in\mathbb{T}$, 1. there is an $\varepsilon\in\mathbb{R}_{+}$ such that $O\left(T,\varepsilon,\left(\left(\mathbf{P}_{T}\right)^{k}\boldsymbol{x}\right)_{k=0}^{\infty}\right)$ is defined for every $\boldsymbol{x}\in\Omega$; 2. for all $\varepsilon\in\mathbb{R}_{+}$ such that $O\left(T,\varepsilon,\left(\left(\mathbf{P}_{T}\right)^{k}\boldsymbol{x}\right)_{k=0}^{\infty}\right)$ is defined for every $\boldsymbol{x}\in\Omega$, we also have that $O\left(T,\varepsilon',R\right)$ is defined for every $\varepsilon'>\varepsilon$ and for every sequence $R=\left(\boldsymbol{x}^{k}\right)_{k=0}^{\infty}$ of points in $\Omega$ generated by $$\boldsymbol{x}^{k+1}=\mathbf{P}_{T}\left(\boldsymbol{x}^{k}+\beta_{k}\boldsymbol{v}^{k}\right),\:\mathrm{for\:all\:}k\geq0,\label{eq:perturbations}$$ where $\beta_{k}\boldsymbol{v}^{k}$ are *bounded perturbations*, meaning that the sequence $\left(\beta_{k}\right)_{k=0}^{\infty}$ of nonnegative real numbers is *summable* (that is, ${\displaystyle \sum\limits _{k=0}^{\infty}}\beta_{k}\,<\infty$), the sequence $\left(\boldsymbol{v}^{k}\right)_{k=0}^{\infty}$ of vectors in $\mathbb{R}^{J}$ is bounded and, for all $k\geq0$, $\boldsymbol{x}^{k}+\beta_{k}\boldsymbol{v}^{k}\in\Delta$. Sufficient conditions for strong perturbation resilience appeared in [@HGDC12 Theorem 1]. With respect to the target function $\phi:\Delta\rightarrow\mathbb{R}$, we adopt the convention that a point in $\Delta$ for which the value of $\phi$ is smaller is considered *superior* to a point in $\Delta$ for which the value of $\phi$ is larger. The essential idea of the SM is to make use of the perturbations of to transform a strongly perturbation resilient algorithm that seeks a constraints-compatible solution (referred to as the *Basic Algorithm*) into a *superiorized version* whose outputs are equally good from the point of view of constraints-compatibility, but are superior (not necessarily optimal) with respect to the target function $\phi$. This can be done by making use of the following concept. \[def:old-nonascent\][@HGDC12] **Nonascending vector** Given a function $\phi:\Delta\rightarrow\mathbb{R}$ and a point $\boldsymbol{y}\in\mathbb{R}^{J}$, we say that a $\boldsymbol{d}\in\mathbb{R}^{J}$ is a *nonascending vector for* $\phi$ *at* $\boldsymbol{y}$ if $\left\Vert \boldsymbol{d}\right\Vert \leq1$ and there is a $\delta>0$ such that $$\text{for all }\lambda\in\left[0,\delta\right]\text{ we have }\phi\left(\boldsymbol{y}+\lambda\boldsymbol{d}\right)\leq\phi\left(\boldsymbol{y}\right).\label{eq:nonascend-1}$$ Obviously, the zero vector $\mathbf{0}$ (all components are 0) is always such a vector, but for the SM to work we need a strict inequality to occur in frequently enough. Generation of nonascending vectors, used for target function reduction steps, has been based mostly on the following theorem or its variants such as [@GH14 Theorem 1] and [@GTH unnumbered Theorem on page 7], which provide sufficient conditions for a nonascending vector. \[thm:old-thm\][@HGDC12 Theorem 2]. Let $\phi:\mathbb{R}^{J}\rightarrow\mathbb{R}$ be a convex function and let $\boldsymbol{x}\in\mathbb{R}^{J}$. Let $\boldsymbol{g}\in\mathbb{R}^{J}$ satisfy the property: For $1\leq j\leq J$, if the $j$th component $g_{j}$ of $\boldsymbol{g}$ is not zero, then the partial derivative $\frac{\partial\phi}{\partial x_{j}}(\boldsymbol{x})$ of $\phi$ at $\boldsymbol{x}$ exists and its value is $g_{j}$. Define $\boldsymbol{d}$ to be the zero vector if $\left\Vert \boldsymbol{g}\right\Vert =0$ and to be $-\boldsymbol{g}/\left\Vert \boldsymbol{g}\right\Vert $ otherwise. Then $\boldsymbol{d}$ is a nonascending vector for $\phi$ at $\boldsymbol{x}$. In order to use this theorem, $\phi$ must have at least one calculable partial derivative (which is nonzero) at points in the domain of $\phi$. Otherwise, the theorem would apply only to the zero vector, which is a useless nonascending vector because it renders the SM ineffective. To enable application of the SM to target functions that have no calculable partial derivatives or subgradients, we proposed in [@CHS18] to search for a point in the neighborhood of $\boldsymbol{x}$ at which the target function exhibits nonascent by comparing function values at points of a fixed distance from $\boldsymbol{x}$ along the space coordinates. To obtain a sequence of nonascending points without making use of Theorem \[thm:old-thm\], we replaced in [@CHS18] the notion of a nonascending vector by the following alternative notion. \[def:nonascent\][@CHS18] **Nonascending $\delta$-bound direction** Given a target function $\phi:\Delta\rightarrow\mathbb{R}$ where $\Delta\subseteq\mathbb{R}^{J}$, a point $\boldsymbol{y}\in\Delta$, and a positive $\delta\in\mathbb{R}$, we say that $\boldsymbol{d}\in\mathbb{R}^{J}$ is a *nonascending* $\delta$-*bound direction for* $\phi$ *at* $\boldsymbol{y}$ if $\|\boldsymbol{d}\|\leq\delta$, $\boldsymbol{y}+\boldsymbol{d}\in\Delta$ and $\phi(\boldsymbol{y}+\boldsymbol{d})\leq\phi(\boldsymbol{y})$. The collection of all such vectors is called a *nonascending* $\delta$-*ball* and is denoted by $\mathcal{B}_{\delta,\phi}(\boldsymbol{y})$, that is, $$\mathcal{B}_{\delta,\phi}(\boldsymbol{y}):=\left\{ \boldsymbol{d}\in\mathbb{R}^{J}\mid\|\boldsymbol{d}\|\leq\delta,\;(\boldsymbol{y}+\boldsymbol{d})\in\Delta,\ \phi(\boldsymbol{y}+\boldsymbol{d})\leq\phi(\boldsymbol{y})\right\} .\label{eq:ball}$$ The zero vector is contained in each nonascending $\delta$-ball, that is, $\boldsymbol{0}\in\mathcal{B}_{\delta,\phi}(\boldsymbol{y})$ for each $\delta>0$ and $\boldsymbol{y}\in\Delta$. The purpose of this definition is to allow the use, as a direction of target function decrease, of any vector $\boldsymbol{d}\in\mathbb{R}^{J}$ for which $\phi(\boldsymbol{y}+\boldsymbol{d})\leq\phi(\boldsymbol{y})$ holds locally only for $\boldsymbol{d}$, and not throughout a certain interval as in Definition \[def:old-nonascent\]. The vector $\boldsymbol{d}$ depends on the value of $\delta$ and they may be determined simultaneously in the superiorization process, as seen below. This kind of nonascent was referred to as *local nonascent* in [@CHS18 Subsection 2.3]. Obviously, local nonascent is a more general notion since every nonascending vector according to Definition \[def:old-nonascent\] is also a nonascending $\delta$-*bound direction* according to Definition \[def:nonascent\] but not vice versa. The advantage of this notion is that it is detectable by using only function value calculations. The following easily-proved proposition unifies these approaches in the convex case. Let $\phi:\mathbb{R}^{J}\rightarrow\mathbb{R}$ be a convex function and let $\boldsymbol{x}\in\mathbb{R}^{J}$. If $\boldsymbol{d}\in\mathbb{R}^{J}$ is a nonascending $\delta$-bound direction for $\phi$ at $\mathbf{\boldsymbol{x}}$, then either $\boldsymbol{d}=\boldsymbol{0}$ (and hence $\boldsymbol{d}$ is a nonascending vector for $\phi$ at $\boldsymbol{x}$) or $\boldsymbol{d}/\|\boldsymbol{d}\|$ is a nonascending vector for $\phi$ at $\boldsymbol{x}$. The idea of calculating $\delta$ (equivalently, the step-size $\gamma_{\ell}$ in the superiorized algorithms presented in the next section) simultaneously with a direction of nonascent appeared in a completely different way in [@LZZS16], where they use an additional internal loop of a penalized minimization to calculate the direction of nonascent; see also [@LZZSW19]. \[sec:Specific-superiorization-approac\]Specific superiorization approaches =========================================================================== This section presents two specific approaches to superiorizing a Basic Algorithm that operates by repeated applications of an algorithmic operator $\mathbf{P}_{T}$ starting from some initial point. The first approach produces the superiorized version that is named **Algorithm \[alg:NonAsc\_Superiorization\]** below, it has been published in the literature previously [@HGDC12 page 5537]. The second approach, named **Algorithm \[alg:Component-wise\]** below, is novel to this paper. The two superiorized versions have some things in common. They are both iterative procedures in which $k$ is used as the iteration index. The first two steps of both algorithms sets $k$ to $0$ and $\boldsymbol{x}^{0}$ to a given initial vector $\boldsymbol{\bar{x}}\in\Delta$. They both assume that we have available a summable sequence $\left(\gamma_{\ell}\right)_{\ell=0}^{\infty}$ of nonnegative real numbers (for example, $\gamma_{\ell}=a^{\ell}$, where $0<a<1$). In Step 3 of both algorithms, $\ell$ is initialized to $-1$ (this is acceptable since $\ell$ is increased by 1 before the first time $\gamma_{\ell}$ is used). In both algorithms the iterative step that produces $\boldsymbol{x}^{k+1}$ from $\boldsymbol{x}^{k}$, as in , is specified within a **repeat** loop that first performs a user-specified number, $N$, of *perturbation steps* followed by one *feasibility-seeking step* that uses the algorithmic operator $\mathbf{P}_{T}$. In more detail, the **repeat** loop in each of the algorithms has the following form. After initializing the loop index $n$ to $0$ and setting $\boldsymbol{x}^{k,0}$ to $\boldsymbol{x}^{k}$, it produces one-by-one $\boldsymbol{x}^{k,1},\boldsymbol{x}^{k,2}$, $\:\ldots,\:\boldsymbol{x}^{k,N}$ (these are the iterations of the perturbation steps), followed by producing $\boldsymbol{x}^{k+1}=\mathbf{P}_{T}\boldsymbol{x}^{k,N}$ (the feasibility-seeking step). The difference between the two algorithms is in how they perform the perturbations for getting from $\boldsymbol{x}^{k,n}$ to $\boldsymbol{x}^{k,n+1}$. We state an important property of **Algorithm \[alg:NonAsc\_Superiorization\]**; for a proof see [@HGDC12 Section II.E]. \[perturbatuion\_resilience\]Suppose that the algorithm $\mathbf{P}$ for a problem structure $(\mathbb{T},\mathcal{P}r)$ is strongly perturbation resilient. Suppose further that $T\in\mathbb{T}$ and $\varepsilon\in\mathbb{R}_{+}$ are such that $O\left(T,\varepsilon,\left(\left(\mathbf{P}_{T}\right)^{k}\boldsymbol{x}\right)_{k=0}^{\infty}\right)$ is defined for every $\boldsymbol{x}\in\Omega$. It is then the case that $O\left(T,\varepsilon',R\right)$ is defined for every $\varepsilon'>\varepsilon$ and every sequence $R=\left(\boldsymbol{x}^{k}\right)_{k=0}^{\infty}$ produced by **Algorithm \[alg:NonAsc\_Superiorization\].** The pseudo-code of **Algorithm \[alg:NonAsc\_Superiorization\]** does not specify how the nonascending vector in Step 8 is to be selected. In publications using **Algorithm \[alg:NonAsc\_Superiorization\]**, such details are usually based on a variant of Theorem \[thm:old-thm\], resulting in a not derivative-free algorithm. For the specification of **Algorithm \[alg:Component-wise\]** we let, for $1\leq j\leq J$, $\mathbf{e}^{j}$ be the vector in $\mathbb{R}^{J}$ all of whose components are $0$, except for the $j$th component, which is $1$. The set of *coordinate directions* is defined as $\Gamma:=\left\{ \mathbf{e}^{j}\:|\:1\leq j\leq J\right\} \,\cup\:\left\{ -\mathbf{e}^{j}\:|\:1\leq j\leq J\right\} $. We assume that $\left(\mathbf{c}^{m}\right)_{m=0}^{\infty}$ is a given sequence of coordinate directions such that any subsequence of length $2J$ contains $\Gamma$. We make the following comments: 1. Steps 15, 16 and 17 of **Algorithm \[alg:Component-wise\]** implement nonascending $\gamma_{\ell}$-bound directions, as in Definition \[def:nonascent\]. In doing so, **Algorithm \[alg:Component-wise\]** realizes in a component-wise manner the algorithmic framework of [@CHS18] (specifically, as expressed in Steps 7 and 8 of Algorithm 1 in that paper). 2. No partial derivatives are used by **Algorithm \[alg:Component-wise\]**. 3. Step 16 of **Algorithm \[alg:Component-wise\]** is similar to Step 13 of **Algorithm \[alg:NonAsc\_Superiorization\]**. One difference is the use of strict inequality in **Algorithm \[alg:Component-wise\]**, the reason for this is that it was found advantageous in some applications of the algorithm. In addition, the **while** loop due to Step 8 of **Algorithm \[alg:Component-wise\]** is executed at most $2J$ times, but there is no upper bound on the (known to be finite) number of executions of the **while** loop due to Step 10 of **Algorithm \[alg:NonAsc\_Superiorization\]**. Also, it follows from the pseudo-code of **Algorithm \[alg:Component-wise\]** that, for all $k\geq0$ and $0\leq n\leq N$, $\phi\left(\boldsymbol{x}^{k,n}\right)<\phi\left(\boldsymbol{x}^{k}\right)$, even though there is no explicit check for this as in Step 13 of **Algorithm \[alg:NonAsc\_Superiorization\]**. 4. A desirable property of **Algorithm \[alg:Component-wise\]** is that it cannot get stuck in a particular iteration $k$ because the value of $L$ increases in an execution of the **while** loop of Step 13 and the value of $n$ increases in an execution of the **while** loop of Step 8. 5. **Algorithm \[alg:Component-wise\]** shares with **Algorithm \[alg:NonAsc\_Superiorization\]** the important property in Theorem \[perturbatuion\_resilience\]. Stated less formally: “For a strongly perturbation resilient algorithm, if for all initial points from $\Omega$ the infinite sequence produced by an algorithm contains an $\varepsilon$-compatible point, then all perturbed sequences produced by the superiorized version of the algorithm contain an $\varepsilon'$-compatible point, for any $\varepsilon'>\varepsilon$.” 6. At present there is no mathematical proof to guarantee that the output of a superiorized version of a constraints-compatibility-seeking algorithm will have smaller target function $\phi$ value than the output by the same constraints-compatibility-seeking algorithm without perturbations, everything else being equal. A partial mathematical result toward coping with this lacuna, in the framework of weak superiorization, is provided by Theorem 4.1 in [@cz3-2015].[^6] \[sec:proximity-target-curve\]The proximity-target curve ======================================================== We now give a tool for deciding which of two iterative methods is “better” for solving a particular problem. Since an iterative method produces a sequence of points, our definition is based on such sequences. For incarnations of the definitions given in this section, the reader may wish to look ahead to Figure \[fig:Target-function-values\]. That figure illustrates the notions discussed in this section for two particular finite sequences $R:=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ and $S:=\left(\boldsymbol{y}^{k}\right)_{k=L_{lo}}^{L_{hi}}$. The details of how those sequences were produced are given below in Subsection \[subsec:Algorithmic-details-and\]. **Monotone proximity of a finite sequence** For a problem structure $(\mathbb{T},\mathcal{P}r)$, a $T\in\mathbb{T}$, positive integers $K_{lo}$ and $K_{hi}>K_{lo}$, the finite sequence $R:=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ of points in $\Omega$ is said to be of *monotone proximity* if for $K_{lo}<k\leq K_{hi}$, $\mathcal{P}r_{T}\left(\boldsymbol{x}^{k-1}\right)>\mathcal{P}r_{T}\left(\boldsymbol{x}^{k}\right)$. **The proximity-target curve of a finite sequence** For a problem structure $(\mathbb{T},\mathcal{P}r)$, a $T\in\mathbb{T}$, a target function $\phi:\Omega\rightarrow\mathbb{R}$, positive integers $K_{lo}$ and $K_{hi}>K_{lo}$, let $R:=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ be a sequence of monotone proximity. Then the *proximity-target curve* $P\subseteq\mathfrak{\mathbb{R}}^{2}$ associated with $R$ is uniquely defined by: 1. For $K_{lo}\leq k\leq K_{hi}$, $\left(\mathcal{P}r_{T}\left(\boldsymbol{x}^{k}\right)\!,\phi\left(\boldsymbol{x}^{k}\right)\right)\in P$. 2. The intersection $\{(y,x)\in\mathbb{R}^{2}:\mathcal{P}r_{T}\left(\boldsymbol{x}^{k}\right)\leq y\leq\mathcal{P}r_{T}\left(\boldsymbol{x}^{k-1}\right)\}\cap P$ is the line segment from $\left(\mathcal{P}r_{T}\left(\boldsymbol{x}^{k-1}\right),\phi\left(\boldsymbol{x}^{k-1}\right)\right)$ to $\left(\mathcal{P}r_{T}\left(\boldsymbol{x}^{k}\right),\phi\left(\boldsymbol{x}^{k}\right)\right)$. \[def:comparison\]**Comparison of proximity-target curves of sequences** For a problem structure $(\mathbb{T},\mathcal{P}r)$, a $T\in\mathbb{T}$, a target function $\phi:\Omega\rightarrow\mathbb{R}$, positive integers $K_{lo}$, $K_{hi}>K_{lo}$, $L_{lo}$, $L_{hi}>L_{lo}$, let $R:=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ and $S:=\left(\boldsymbol{y}^{k}\right)_{k=L_{lo}}^{L_{hi}}$ be sequences of points in $\Omega$ of monotone proximity for which $P$ and $Q$ are their respective associated proximity-target curves. Define $$\begin{array}{c} t:=\max\left(\mathcal{P}r_{T}\left(\boldsymbol{x}^{K_{hi}}\right),\mathcal{P}r_{T}\left(\boldsymbol{y}^{L_{hi}}\right)\right),\\ u:=\min\left(\mathcal{P}r_{T}\left(\boldsymbol{x}^{K_{lo}}\right),\mathcal{P}r_{T}\left(\boldsymbol{y}^{L_{lo}}\right)\right). \end{array}\label{eq:limits-1}$$ Then $R$ is *better targeted* than $S$ if: 1. $t\leq u$ and 2. for any real number $h$, if $t\leq h\leq u$, $\left(h,v\right)\in P$ and $\left(h,w\right)\in Q$, then $v\leq w$. Let us see how this last definition translates into something that is intuitively desirable. Suppose that we have an iterative algorithm that produces a sequence, $\boldsymbol{y}^{0},\boldsymbol{y}^{1},\boldsymbol{y}^{2},\cdots$, of which $S:=\left(\boldsymbol{y}^{k}\right)_{k=L_{lo}}^{L_{hi}}$ is a subsequence. An alternative algorithm that produces a sequence of points of which $R:=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ is a subsequence that is better targeted than $S$ has a desirable property: Within the range $\left[t,u\right]$ of proximity values, the point that is produced by the alternative algorithm with that proximity value, is likely to have lower (and definitely not higher) value of the target function as the point with that proximity value that is produced by the original algorithm. This property is stronger than what we stated before, namely that superiorization produces an output that is equally good from the point of view of proximity, but is superior with respect to the target function. Here the single output determined by a fixed $\varepsilon$ is replaced by a set of potential outputs for any $\varepsilon\in\left[t,u\right]$. \[sec:Experimental\]Experimental demonstration of derivative-free component-wise superiorization ================================================================================================ \[subsec:Goal\]Goal and general methodology ------------------------------------------- Our goal is to demonstrate that compo-nent-wise superiorization (**Algorithm \[alg:Component-wise\]**) is a viable efficient DFS method to handle data of constrained-minimization problems (that is, a target function and a set of constraints), when the target function has no calculable partial derivatives. To ensure the meaningfulness and worthiness of our experiments, we generate the constraints and choose a target function, that has no calculable partial derivatives, inspired by an application area of constrained optimization, namely image reconstruction from projections in computerized tomography (CT). For the so-obtained data we consider two runs of **Algorithm \[alg:Component-wise\]**, one with and the other without the component-wise perturbation steps. To be exact, by “without perturbation” we mean that Steps 10–18 in **Algorithm \[alg:Component-wise\]** are deleted so that $\boldsymbol{x}^{k,N}=\boldsymbol{x}^{k}$, which amounts to running the feasibility-seeking basic algorithm $\mathbf{P}_{T}$ without any perturbations. Everything else is equal in the two runs, such as the initialization point $\boldsymbol{\bar{x}}$ and all parameters associated with the application of the feasibility-seeking basic algorithm in Step 20. The results are presented below by plots of proximity-target curves that show that the target function values of **Algorithm \[alg:Component-wise\]** when run “with perturbations” are systematically lower than those of the same algorithm without the component-wise perturbations. The numerical behavior of **Algorithm \[alg:Component-wise\]**, as demonstrated by our experiment, makes it a meritorious choice for superiorization in situations involving a derivative-free target function and a set of constraints. To reach the goal described above we proceed in the following stages. 1. Specification of a problem structure $(\mathbb{T},\mathcal{P}r)$ for the experimental demonstration, and generation of constraints, simulated from the application of image reconstruction from projections in computerized tomography. 2. Choice of a $\Delta$ and a derivative-free target function $\phi$ for the experiment. 3. Specification of the algorithmic operator $\mathbf{P}_{T}$ to be used in **Algorithm \[alg:Component-wise\]**. This is chosen so that the Basic Algorithm that operates by repeated applications of $\mathbf{P}_{T}$ is a standard sequential iterative projections method for feasibility-seeking of systems of linear equations; a version of the Algebraic Reconstruction Techniques (ART) [@GTH-book Chapter 11] that is equivalent to Kaczmarz’s projections method [@kacmarcz]. 4. Specification of algorithmic details and parameters, such as $N$ and $\gamma_{\ell}$ in **Algorithm \[alg:Component-wise\]**. \[subsec:The-selected-problem\]Problem selection, constraints generation and choices of $\Delta$ and of the target function --------------------------------------------------------------------------------------------------------------------------- We generate the constraints and chose a target function from the application area of image reconstruction from projections in computerized tomography (CT).[^7] The problem structure $(\mathbb{T},\mathcal{P}r)$ for our demonstration has been used in the literature for comparative evaluations of various algorithms for CT [@GTH; @GTH-book; @HGDC12; @NDH12]. It is of the type described in Section \[sec:Introduction\] by and . Specifically, vectors $\boldsymbol{x}$ in $\Omega=\mathbb{R}^{J}$ represent two-dimensional (2D) images, with each component of $\boldsymbol{x}$ representing the *density* assigned to one of the pixels in the image. We use $J=235,225$, thus each $\boldsymbol{x}$ represents a $485\times485$ image. Our test image (phantom) is represented by the vector $\boldsymbol{\hat{x}}$, that is a digitization of a picture of a cross-section of a human head [@GTH-book Sections 4.1–4.4 and 5.2]. In the problem $T$ that we use for our illustration, each index $i=1,2,\ldots,I$ is associated with a line across the image and the corresponding $\boldsymbol{d}^{i}$ is a vector in $\mathbb{R}^{J}$, whose $j$th component is the length of intersection of that line with the $j$th of the $J$ pixels. There are $I=498,960$ such lines (organized into 720 divergent projections with 693 lines in each; similar to the *standard geometry* in [@GTH-book] but with more lines in each projection). The $h_{i}$ have been calculated by simulating the behavior of CT scanning of the head cross-section [@GTH-book Section 4.5]. All the above was generated using the SNARK14 programming system for the reconstruction of 2D images from 1D projections [@SNARK14], giving rise to a system of linear equations . For the resulting $T$, we calculated that the proximity of the phantom to the generated constraints is $\mathcal{P}r_{T}(\boldsymbol{\boldsymbol{\hat{x}}})=6.4192$. For our demonstration we make the simplest choice for $\Delta$, namely, $\Delta=\Omega=\mathbb{R}^{J}$. Our choice of the target function $\phi$ is as follows. We index the pixels (i.e., the components of a vector $\boldsymbol{x}$) by $j$ and let $\Theta$ denote the set of all indices of pixels that are not in the rightmost column or the bottom row of the 2D pixel array that displays that vector as an image. For any pixel with index $j\in\Theta$, let $r\left(j\right)$ and $b\left(j\right)$ be the index of the pixel to its right and below it in the 2D pixel array, respectively. Denoting by $\mathrm{med}$ the function that selects the median value of its three arguments, we define $$\phi\left(\boldsymbol{x}\right):=\sum_{j\in\Theta}\sqrt{\left|x_{j}-\mathrm{med}\left\{ x_{j},x_{r\left(j\right)},x_{b\left(j\right)}\right\} \right|}.\label{eq:median_target}$$ Finding partial derivatives for this target function is problematic. On the other hand, when only one pixel value (that is, only one component of the vector) is changed in vector $\boldsymbol{x}$ to get another vector $\boldsymbol{y}$, then it is possible to obtain $\phi(\boldsymbol{y})$ from $\phi(\boldsymbol{x})$ by computing only three of the terms in the summation on the right-hand side of . These observations indicate that the use of the derivative-free approach of Steps 10–18 in **Algorithm \[alg:Component-wise\]** is a viable option whereas Step 8 of **Algorithm \[alg:NonAsc\_Superiorization\]** is hard to perform unless the trivial nonascending vector $\boldsymbol{v}^{k,n}=\boldsymbol{0}$ is selected, which is ineffective. For our chosen phantom we calculated $\phi\left(\boldsymbol{\boldsymbol{\hat{x}}}\right)=2,048.57$. The algorithmic operator $\mathbf{P}_{T}$ ----------------------------------------- Our chosen operator, mapping $\boldsymbol{x}$ into $\mathbf{P}_{T}\boldsymbol{x}$, is specified by **Algorithm \[alg:OperatorPT\].** It depends on a real parameter $\lambda$ in its Step 4. **Algorithm \[alg:ART\]** is a special case of the general class of Algebraic Reconstruction Techniques as discussed in [@GTH-book Chapter 11] and is, for $\lambda=1$, equivalent to the original method of Kaczmarz in the seminal paper [@kacmarcz]. For further references on Kaczmarz’s method and the Algebraic Reconstruction Techniques see, e.g., [@cegielski-book page 220], [@annotated15 Section 2] and [@HERM19]. Note that **Algorithm \[alg:ART\]** (ART) can be obtained from either **Algorithm \[alg:NonAsc\_Superiorization\]** or **Algorithm \[alg:Component-wise\]** by removing the perturbation steps in their **while** loops. \[subsec:EP-approach\]A comment about the exterior penalty function (EP) approach to derivative-free constrained minimization ----------------------------------------------------------------------------------------------------------------------------- One possibility for doing a derivative-free constrained minimization algorithm is to follow the option of using the exterior penalty (EP) function approach mentioned in [@Conn-book-2009 Chapter 13, Section 13.1, page 242] as applied to the constrained problem $$\min\text{ }\text{ }\left\{ \phi\left(\boldsymbol{x}\right)\mid\left\langle \boldsymbol{d}^{i},\boldsymbol{x}\right\rangle =h_{i},\thinspace\thinspace\thinspace i=1,2,\ldots,I\right\} ,\label{eq:constrained}$$ where $\phi$ is as in and the constraints are as in . With a user-selected *penalization parameter* $\eta$, the exterior penalty function approach replaces the constrained minimization problem by the penalized unconstrained minimization: $$\min\text{ }\{\psi(\boldsymbol{x})\mid\boldsymbol{x}\in\mathbb{R}^{J}\},\label{eq:reg-prob}$$ $$\psi(\boldsymbol{x}):=\phi\left(\boldsymbol{x}\right)+\eta\mathcal{P}r_{T}(\boldsymbol{x}),\label{eq:unconstrained}$$ with $\phi$ as in and $\mathcal{P}r_{T}(\boldsymbol{x})$ as defined in . By applying the coordinate-search method of [@Conn-book-2009 Algorithm 7.1] to the penalized unconstrained minimization problem , we get the next algorithm. From the point of view of keeping the computational cost low, **Algorithm \[alg:Derivative-free\]** can be much more of a challenge than **Algorithm \[alg:Component-wise\]**. The reason for this has been indicated when we have stated, near the end of Subsection \[subsec:The-selected-problem\], that if only one component is changed in vector $\boldsymbol{x}$ to get another vector $\boldsymbol{y}$, then it is possible to obtain $\phi(\boldsymbol{y})$ from $\phi(\boldsymbol{x})$ by computing only three of the terms in the summation on the right-hand side of . When we use $\psi$ in instead of $\phi$, there seems to be a need for many more computational steps. This is because the number of terms that change on the right-hand side of due to a change in one component of $\boldsymbol{x}$ is of the order of 1,000 for the dataset described in Subsection \[subsec:The-selected-problem\] (in the language of image reconstruction from projections, there is at least one line $i$ in each of of the 720 projections for which there is a change in value of $\left\langle \boldsymbol{d}^{i},\boldsymbol{x}\right\rangle $ due to changing one component of $\boldsymbol{x}$). Thus our advocated approach of component-wise superiorization in **Algorithm \[alg:Component-wise\]** is likely to be orders of magnitude faster for our application area than the more traditional approach of derivative-free constrained minimization by the exterior penalty (EP) approach in **Algorithm \[alg:Derivative-free\].** \[subsec:The-progressive-barrier\]The progressive barrier (PB) approach ----------------------------------------------------------------------- The progressive barrier (PB) approach, described in Chapter 12 of [@Audet-book-2017], was originally published in [@Audet-Dennis-2009] wherein the history of the approach as a development of the earlier filter methods of Fletcher and Leyffer [@filter02] is succinctly described. It is an alternative to the exterior penalty (EP) approach of the previous subsection and, therefore, we briefly describe it here and compare it to DFS. The EP approach bears some similarities to DFS but differs from it in a way that explains why our DFS is likely to be advantageous for large-scale problems. No penalty is used in the PB approach. Instead of combining the constraints with the target function it uses a particular “constraint violation function” $h(x)$ [@Audet-book-2017 Definition 12.1] alongside with the target function $\phi(x)$ so that the iterates appear in an $h$ versus $\phi$ plot called “a filter”, based on the pairs of their $h$ and $\phi$ values. The constraint violation function of PB is similar in nature to our “proximity function” and the $h$ versus $\phi$ filter plot of PB is similar to our proximity-target curve, both specified above. The difference between the PB approach of [@Audet-Dennis-2009] and our DFS lies in how these objects are used. The PB optimization algorithm defines at each iteration what is a “success” or a “failure” of an iterate based on the current filter plot, and decides accordingly what will be the next iterate. We do not provide full details here but, in a nutshell, the PB algorithm performs at each iteration sophisticated searches of both the target function values and the constraint violation function values. In contrast to the PB approach, the DFS investigated here takes the world-view of superiorization. It uses a feasibility-seeking algorithm whose properties are already known and perturbs its iterates without loosing its feasibility-seeking ability and properties. The component-wise derivative-free search for a nonascending direction of the target function is done in a manner that makes it a perturbation to which the feasibility-seeking algorithm is resilient (that is, the underlying feasibility-seeking algorithm retains its feasibility-seeking nature). Thus, our DFS method searches the target function in a derivative-free fashion and automatically produces feasibility-seeking steps that actively reduce the proximity function. This implies an advantage for the DFS method in handling large-scale problems, because no expensive additional time and computing resources are needed for the feasibility-seeking phase of the DFS algorithm proposed here. Validation of this claim requires further experimental work. \[subsec:Algorithmic-details-and\]Algorithmic details and numerical demonstration --------------------------------------------------------------------------------- Our experiments were carried out using the public-domain software package SNARK14 [@SNARK14]. In all experiments the initial vector $\boldsymbol{\bar{x}}$ was the $235,225$-dimensional zero vector (all components 0). The relaxation parameter in **Algorithm \[alg:OperatorPT\]** was $\lambda=0.05$. Another issue that needs specification is the ordering of the constraints in , because the output of **Algorithm \[alg:OperatorPT\]** depends not only on the set of constraints, but also on their order. We used in our experiments the so-called *efficient ordering*, since it has been demonstrated to lead to better results faster when incorporated into ART [@GTH-book page 209]. In **Algorithm \[alg:Component-wise\]**, the number $N$ of perturbation steps (for each feasibility-seeking step) was 100,000 and we used $\gamma_{\ell}=ba^{\ell}$, with $b=0.02$ and $a=0.999,999$. The infinite sequence $\left(\mathbf{c}^{m}\right)_{m=0}^{\infty}$ was obtained by repetitions of the length-$2J$ subsequence $\left(\mathbf{e}^{1},\mathbf{e}^{2},\ldots\mathbf{e}^{J},-\mathbf{e}^{1},\mathbf{-e}^{2},\ldots,-\mathbf{e}^{J}\right)$. ![\[fig:Target-function-values\]Proximity-target curves $P$ and $Q$ of the first 30 iterates of **Algorithm 2** with perturbations () and without perturbations ($\circ)$.](Plots_ONLY_ALG2_060319) We applied **Algorithm \[alg:Component-wise\]** twice, thirty iterations in each case, with and without its component-wise perturbations steps, respectively, under otherwise completely identical conditions. The resulting finite sequences of iterates are both of monotone proximity, the associated proximity-target curves are shown in Figure \[fig:Target-function-values\]. The $\circ$s and s on the plots represent actually calculated values at iterations of each algorithm, that are connected by line segments. For any proximity value on the horizontal axis we can read the target-function value associated with it from the curve. The plots indicate visually the behavior of the algorithms, initialized at the same point denoted by $x^{0}=y^{0}$ that appears in the right-most side of the figure. The V-shaped form of the proximity-target curve for **Algorithm \[alg:Component-wise\]** with perturbations is typical for the behavior of superiorized feasibility-seeking algorithms, showing the initially strong effect of the perturbations that diminishes as the iterations proceed. For a more precise interpretation, consider Definition \[def:comparison\]. In the experiment evaluating the two versions of **Algorithm \[alg:Component-wise\]**, $K_{lo}=L_{lo}=1$ and $K_{hi}=L_{hi}=30$. The $R=\left(\boldsymbol{x}^{k}\right)_{k=K_{lo}}^{K_{hi}}$ and $S=\left(\boldsymbol{y}^{k}\right)_{k=L_{lo}}^{L_{hi}}$ produced by **Algorithm \[alg:Component-wise\]**, with and without perturbations, respectively, are both of monotone proximity. We find that $\mathcal{P}r_{T}\left(\boldsymbol{x}^{K_{lo}}\right)=Pr_{T}\left(\boldsymbol{y}^{L_{lo}}\right)=35.4703$ (and, hence, $u=35.4703$) and that $\mathcal{P}r_{T}\left(\boldsymbol{x}^{K_{hi}}\right)=3.4065$ and $\mathcal{P}r_{T}\left(\boldsymbol{y}^{L_{hi}}\right)=4.7828$ (and, hence, $t=4.7828$). By showing the target curves $P$ and $Q$ associated with $R$ and $S$, respectively, Figure \[fig:Target-function-values\] clearly illustrates that $R$ is better targeted than $S$. \[sec:Discussion-and-conclusions\]Discussion and conclusions ============================================================ In this paper we investigated the general applicability of derivative-free superiorization (DFS) as an alternative to previously proposed superiorization approaches. In our computational demonstration, we generated the constraints and chose the target function from the application area of image reconstruction from projections in computerized tomography (CT). However, we use the demonstration for indicating only the numerical behavior of the algorithms. We do not investigate or comment on the potential usefulness of the resulting reconstructions in CT, since that usefulness depends not so much on the numerical behavior of the algorithms as on the appropriateness of the modeling used to turn a physical problem into a mathematical one (for example, by the specific choice of target function). The numerical results of our demonstration attest, as seen from the proximity-target curves, to the mathematical efficacy of our derivative-free superiorization algorithm, but say nothing about its efficacy for providing an answer to a practical image reconstruction problem. (Nevertheless, we have observed while doing our experiment that, even from the image reconstruction quality point of view, DFS seems to be advantageous. For example, if we consider the distances between the phantom and the reconstructions -defined as the 2-norm between the representing vectors-, the smallest distance that we get as we iterate without perturbations is 0.0922, while with the DFS perturbations it is 0.0863.) Much of the literature on derivative-free minimization is concerned with unconstrained minimization or at most with bound-constraints on the variables, and many, if not all, proposed methods can handle only small-size problems efficiently. In contrast, the DFS method proposed here can handle any type of constraints for which a separate efficient derivative-free constraints-compatibility-seeking algorithm is available. Since the constraints-compatibility-seeking algorithm forms part of the proposed DFS method, the method can use exterior initialization (i.e., initializing the iterations at any point in space). Furthermore, and very importantly, very large-size problems can be accommodated. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Nikolaos Sahinidis and Katya Scheinberg for several informative mail exchanges that helped us see better the general picture. We are grateful to S[é]{}bastien Le Digabel for calling our attention to the work of Charles Audet and coworkers, particularly the Audet and Dennis paper [@Audet-Dennis-2009] and the book of Audet and Hare [@Audet-book-2017]. [28]{} \[1\][\#1]{} \[1\] Audet, C. and Dennis J.E. JR., 2009. A progressive barrier for derivative-free nonlinear programming, *SIAM Journal on Optimization*, 20, 445–472. Audet, C. and Hare, W., 2017. *Derivative-Free and Blackbox Optimization*, Springer International Publishing, Cham, Switzerland. Bargetz, C., Reich, S., and Zalas, R., 2018. Convergence properties of dynamic string-averaging projection methods in the presence of perturbations, *Numerical Algorithms*, 77, 185–209. Butnariu, D., Reich, S., and Zaslavski, A. J., 2006. Convergence to fixed points of inexact orbits of Bregman-monotone and of nonexpansive operators in Banach spaces, *in: Proceedings of Fixed Point Theory and its Applications, Mexico*, Yokohama, 11–32. Cegielski, A., 2012. *Iterative Methods for Fixed Point Problems in Hilbert Spaces*, Springer-Verlag. Cegielski, A. and Al-Musallam, F., 2017. Superiorization with level control, *Inverse Problems*, 33, 044009. Censor, Y., 2015. Weak and strong superiorization: Between feasibility-seeking and minimization, *Analele Stiintifice ale Universitatii Ovidius Constanta-Seria Matematica*, 23, 41–54. Censor, Y., 2019. *Superiorization and Perturbation Resilience of Algorithms: A Bibliography compiled and continuously updated*, <http://math.haifa.ac.il/yair/bib-superiorization-censor.html>, last updated: December 27, 2019. Censor, Y. and Cegielski, A., 2015. Projection methods: An annotated bibliography of books and reviews, *Optimization*, 64, 2343–2358. Censor, Y., Heaton, H., and Schulte, R.W., 2019. Derivative-free superiorization with component-wise perturbations, *Numerical Algorithms*, 80, 1219–1240. Censor, Y., Herman, G.T., and Jiang, M., (Editors) 2017. Special issue on Superiorization: Theory and Applications, *Inverse Problems*, 33, 040301–044014. Censor, Y. and Zaslavski, A., 2015. Strict Fej[é]{}r monotonicity by superiorization of feasibility-seeking projection methods, *Journal of Optimization Theory and Applications*, 165, 172–187. Conn, A.R., Scheinberg, K., and Vicente, L.N., 2009. *Introduction to Derivative-Free Optimization*, Society for Industrial and Applied Mathematics (SIAM). Davidi, R., Gardu[ñ]{}o, E., Herman, G.T., Langthaler, O., Rowland, S.W., Sardana, S., and Ye, Z., 2019. SNARK14: A programming system for the reconstruction of 2D images from 1D projections, available from <http://turing.iimas.unam.mx/SNARK14M/SNARK14.pdf>. Diniz-Ehrhardt, M., Mart[í]{}nez, J., and Pedroso, L., 2011. Derivative-free methods for nonlinear programming with general lower-level constraints, *Computational and Applied Mathematics*, 30, 19–52. Echeverr[í]{}a Ciaurri, D., Isebor, O., and Durlofsky, L., 2012. Application of derivative-free methodologies to generally constrained oil production optimization problems, *Procedia Computer Science*, 1, 1301–1310. Engl, H.W., Hanke, M., and Neubauer, A., 2000. *Regularization of Inverse Problems*, Kluwer Academic Publishers. Fletcher, R. and Leyffer, S., 2002. Nonlinear programming without a penalty function, *Mathematical Programming, Series A*, 91, 239–269. Gardu[ñ]{}o, E. and Herman, G.T., 2014. Superiorization of the ML-EM algorithm, *IEEE Transactions on Nuclear Science*, 61, 162–172. Gardu[ñ]{}o, E. and Herman, G.T., 2017. Computerized tomography with total variation and with shearlets, *Inverse Problems*, 33, 044011. He, H. and Xu, H.K., 2017. Perturbation resilience and superiorization methodology of averaged mappings, *Inverse Problems*, 33, 044007. Herman, G.T., 2009. *Fundamentals of Computerized Tomography: Image Reconstruction from Projections*, Springer-Verlag, 2nd ed. Herman, G.T., 2019. Iterative reconstruction techniques and their superiorization for the inversion of the Radon transform, *in*: R. Ramlau and O. Scherzer, eds., *The Radon Transform: The First 100 Years and Beyond*, De Gruyter, 217–238. Herman, G.T., Gardu[ñ]{}o, E., Davidi, R., and Censor, Y., 2012. Superiorization: An optimization heuristic for medical physics, *Medical Physics*, 39, 5532–5546. Kaczmarz, S., 1937. Angen[ä]{}herte Aufl[ö]{}sung von Systemen linearer Gleichungen, *Bulletin de l’Acad[é]{}mie Polonaise des Sciences et Lettres*, A35, 355–357. Li, L., Chen, Y., Liu, Q., Lazic, J., Luo, W., and Li, Y., 2017. Benchmarking and evaluating MATLAB derivative-free optimisers for single-objective applications, *in*: D.S. Huang, K.H. Jo, and J. Figueroa-Garc[í]{}a, eds., *Intelligent Computing Theories and Application. ICIC 2017*, Springer, 75–88. Luo, S., Zhang, Y., Zhou, T., and Song, J., 2016. Superiorized iteration based on proximal point method and its application to XCT image reconstruction, *ArXiv e-prints*, <https://arxiv.org/abs/1608.03931>. Luo, S., Zhang, Y., Zhou, T., Song, J., and Wang, Y., 2018. XCT image reconstruction by a modified superiorized iteration and theoretical analysis, *Optimization Methods and Software*, to appear. Nikazad, T., Davidi, R., and Herman, G.T., 2012. Accelerated perturbation resilient block-iterative projection methods with application to image reconstruction, *Inverse Problems*, 28, 035005. Reich, S. and Zaslavski, A.J., 2017. Convergence to approximate solutions and perturbation resilience of iterative algorithms, *Inverse Problems*, 33, 044005. Rios, L.M. and Sahinidis, N.V., 2013. Derivative-free optimization: A review of algorithms and comparison of software implementations, *Journal of Global Optimization*, 56, 1247–1293. Zaslavski, A.J., 2018. *Algorithms for Solving Common Fixed Point Problems*, Springer. [^1]: Department of Mathematics, University of Haifa, Mt. Carmel, Haifa 3498838, Israel () [^2]: Departamento de Ciencias de la Computaci[ó]{}n, Instituto de Investigaciones en Matem[á]{}ticas Aplicadas y en Sistemas, Universidad Nacional Aut[ó]{}noma de M[é]{}xico, Cd. Universitaria, C.P. 04510, Mexico City, Mexico () [^3]: Department of Applied Mathematics and Statistics, University of S[ã]{}o Paulo, S[ã]{}o Carlos, S[ã]{}o Paulo 13566-590, Brazil () [^4]: Computer Science Ph.D. Program, The Graduate Center, City University of New York, New York, NY 10016, USA () [^5]: February 3, 2020. [^6]: The approach followed in the present paper was termed *strong superiorization* in [@cz3-2015 Section 6] and [@Cen15] to distinguish it from *weak superiorization*, wherein asymptotic convergence to a point in $C$ is studied instead of $\varepsilon$-compatibility. [^7]: The term projection has in this field a different meaning than in convex analysis. It stands for a set of estimated line integrals through the image that has to be reconstructed, see [@GTH-book page 3].

--- abstract: '[We study the action of substitution maps between power series rings as an additional algebraic structure on the groups of Hasse–Schmidt derivations. This structure appears as a counterpart of the module structure on classical derivations. ]{}' author: - 'L. Narváez Macarro[^1]' title: 'On Hasse–Schmidt derivations: the action of substitution maps' --- *Dedicated to Antonio Campillo on the ocassion of his 65th birthday* Introduction {#introduction .unnumbered} ============ For any commutative algebra $A$ over a commutative ring $k$, the set $\operatorname{Der}_k(A)$ of $k$-derivations of $A$ is an ubiquous object in Commutative Algebra and Algebraic Geometry. It carries an $A$-module structure and a $k$-Lie algebra structure. Both structures give rise to a [*Lie-Rinehart algebra*]{} structure over $(k,A)$. The $k$-derivations of $A$ are contained in the filtered ring of $k$-linear differential operators ${\mathcal{D}}_{A/k}$, whose graded ring is commutative and we obtain a canonical map of graded $A$-algebras $$\tau: \operatorname{Sym}_A \operatorname{Der}_k(A) \longrightarrow \operatorname{\rm gr}{\mathcal{D}}_{A/k}.$$ If ${\mathbb{Q}}\subset k$ and $\operatorname{Der}_k(A)$ is a finitely generated projective $A$-module, the map $\tau$ is an isomorphism ([@nar_2009 Corollary 2.17]) and we can deduce that the ring ${\mathcal{D}}_{A/k}$ is the enveloping algebra of the Lie-Rinehart algebra $\operatorname{Der}_k(A)$ (cf. [@nar_Maringa Proposition 2.1.2.11]). If we are not in characteristic $0$, even if $A$ is “smooth” (in some sense) over $k$, e.g. $A$ is a polynomial or a power series ring with coefficients in $k$, the map $\tau$ has no chance to be an isomorphism. In [@nar_2009] we have proved that, if we denote by $\operatorname{Ider}_k(A) \subset \operatorname{Der}_k(A)$ the $A$-module of [*integrable derivations*]{} in the sense of Hasse–Schmidt (see Definition \[def:HS-integ\]), then there is a canonical map of graded $A$-algebras $$\vartheta: \Gamma_A \operatorname{Ider}_k(A) \longrightarrow \operatorname{\rm gr}{\mathcal{D}}_{A/k},$$ where $\Gamma_A(-)$ denotes the [*divided power algebra*]{} functor, such that: 1. $\tau=\vartheta$ when ${\mathbb{Q}}\subset k$ (in that case $\operatorname{Ider}_k(A) = \operatorname{Der}_k(A)$ and $\Gamma_A = \operatorname{Sym}_A$). 2. $\vartheta$ is an isomorphism whenever $\operatorname{Ider}_k(A) = \operatorname{Der}_k(A)$ and $\operatorname{Der}_k(A)$ is a finitely generated projective $A$-module. The above result suggests an idea: under the “smoothness” hypothesis (ii), can be the ring ${\mathcal{D}}_{A/k}$ and their modules functorially reconstructed from Hasse–Schmidt derivations? To tackle it, we first need to explore the algebraic structure of Hasse–Schmidt derivations. Hasse–Schmidt derivations of length $m\geq 1$ form a group, non-abelian for $m\geq 2$, which coincides with the (abelian) additive group of usual derivations $\operatorname{Der}_k(A)$ for $m=1$. But $\operatorname{Der}_k(A)$ has also an $A$-module structure and a natural questions arises: Do Hasse–Schmidt derivations of any length have some natural structure extending the $A$-module structure of $\operatorname{Der}_k(A)$ for length $=1$? This paper is devoted to study the action of [*substitution maps*]{} (between power series rings) on Hasse–Schmidt derivations as an answer to the above question. This action plays a key role in [@nar_in_prep]. Now let us comment on the content of the paper. In Section 1 we have gathered, due to the lack of convenient references, some basic facts and constructions about rings of formal power series in an arbitrary number of variables with coefficients in a non-necessarily commutative ring. In the case of a finite number of variables many results and proofs become simpler, but we need the infinite case in order to study $\infty$-variate Hasse-Schmidt derivations later. Sections 2 and 3 are devoted to the study of substitution maps between power series rings and their action on power series rings with coefficients on a (bi)module. In Section 4 we study multivariate (possibly $\infty$-variate) Hasse–Schmidt derivations. They are a natural generalization of usual Hasse–Schmidt derivations and they provide a convenient framework to deal with Hasse–Schmidt derivations. In Section 5 we see how substitution maps act on Hasse–Schmidt derivations and we study some compatibilities on this action with respect to the group structure. In Section 6 we show how the action of substitution maps allows us to express any HS-derivation in terms of a fixed one under some natural hypotheses. This result generalizes Theorem 2.8 in [@magda_nar_hs] and provides a conceptual proof of it. Rings and (bi)modules of formal power series ============================================ From now on $R$ will be a ring, $k$ will be a commutative ring and $A$ a commutative $k$-algebra. A general reference for some of the constructions and results of this section is [@bourbaki_algebra_II_chap_4_7 §4]. Let ${{\bf s}}$ be a set and consider the free commutative monoid ${\mathbb{N}}^{({{\bf s}})}$ of maps $\alpha:{{\bf s}}\to {\mathbb{N}}$ such that the set $\operatorname{\rm supp}\alpha := \{s\in {{\bf s}}\ |\ \alpha(s)\neq 0\}$ is finite. If $\alpha \in {\mathbb{N}}^{({{\bf s}})}$ and $s\in {{\bf s}}$ we will write $\alpha_s$ instead of $\alpha(s)$. The elements of the canonical basis of ${\mathbb{N}}^{({{\bf s}})}$ will be denoted by ${{\bf s}}^t$, $t\in {{\bf s}}$: ${{\bf s}}^t_u = \delta_{tu}$ for $t,u\in {{\bf s}}$. For each $\alpha \in {\mathbb{N}}^{({{\bf s}})}$ we have $\alpha = \sum_{\scriptscriptstyle t\in {{\bf s}}} \alpha_t {{\bf s}}^t$. The monoid ${\mathbb{N}}^{({{\bf s}})}$ is endowed with a natural partial ordering. Namely, for $\alpha,\beta\in {\mathbb{N}}^{({{\bf s}})}$, we define $$\alpha \leq \beta\quad \stackrel{\text{def.}}{\Longleftrightarrow}\quad \exists \gamma \in {\mathbb{N}}^{({{\bf s}})}\ \text{such that}\ \beta = \alpha + \gamma\quad \Leftrightarrow\quad \alpha_s \leq \beta_s\quad \forall s\in {{\bf s}}.$$ Clearly, $t\in \operatorname{\rm supp}\alpha$ $\Leftrightarrow$ ${{\bf s}}^t \leq \alpha$. The partial ordered set $({\mathbb{N}}^{({{\bf s}})},\leq)$ is a directed ordered set: for any $\alpha,\beta\in {\mathbb{N}}^{({{\bf s}})}$, $\alpha,\beta \leq \alpha \vee \beta$ where $(\alpha \vee \beta)_t := \max \{\alpha_t,\beta_t\}$ for all $t\in{{\bf s}}$. We will write $\alpha < \beta$ when $\alpha \leq \beta$ and $\alpha \neq \beta$. For a given $\beta \in {\mathbb{N}}^{({{\bf s}})}$ the set of $\alpha \in {\mathbb{N}}^{({{\bf s}})}$ such that $\alpha \leq \beta$ is finite. We define $|\alpha| := \sum_{\scriptscriptstyle s\in {{\bf s}}} \alpha_s = \sum_{\scriptscriptstyle s\in \operatorname{\rm supp}\alpha} \alpha_s \in {\mathbb{N}}$. If $\alpha \leq \beta$ then $|\alpha|\leq |\beta|$. Moreover, if $\alpha \leq \beta$ and $|\alpha| = |\beta|$, then $\alpha=\beta$. The $\alpha \in {\mathbb{N}}^{({{\bf s}})}$ with $|\alpha| = 1$ are exactly the elements ${{\bf s}}^t$, $t\in {{\bf s}}$, of the canonical basis. A [*formal power series*]{} in ${{\bf s}}$ with coefficients in $R$ is a formal expression $\sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} r_\alpha {{\bf s}}^\alpha$ with $r_\alpha \in R$ and ${{\bf s}}^\alpha = \prod_{s\in {{\bf s}}} s^{\alpha_s} = \prod_{s\in \operatorname{\rm supp}\alpha} s^{\alpha_s}$. Such a formal expression is uniquely determined by the family of coefficients $a_\alpha$, $\alpha \in {\mathbb{N}}^{({{\bf s}})}$. If $r=\sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} r_\alpha {{\bf s}}^\alpha$ and $r'=\sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} r'_\alpha {{\bf s}}^\alpha$ are two formal power series in ${{\bf s}}$ with coefficients in $R$, their sum and their product are defined in the usual way $$\begin{aligned} \displaystyle r+r' := \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} S_\alpha {{\bf s}}^\alpha, &\displaystyle S_\alpha := r_\alpha + r'_\alpha,&\\ \displaystyle r r' := \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} P_\alpha {{\bf s}}^\alpha, &\displaystyle P_\alpha := \sum_{\scriptscriptstyle \beta + \gamma=\alpha} r_\beta r'_\gamma. &\end{aligned}$$ The set of formal power series in ${{\bf s}}$ with coefficients in $R$ endowed with the above internal operations is a ring called the [*ring of formal power series in ${{\bf s}}$ with coefficients in $R$*]{} and is denoted by $R[[{{\bf s}}]]$. It contains the polynomial ring $R[{{\bf s}}]$ (and so the ring $R$) and all the monomials ${{\bf s}}^\alpha$ are in the center of $R[[{{\bf s}}]]$. There is a natural ring epimorphism, that we call the [*augmentation*]{}, given by $$\label{eq:augmen-psr} \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} r_\alpha {{\bf s}}^\alpha \in R[[{{\bf s}}]] \longmapsto r_0 \in R,$$ which is a retraction of the inclusion $R\subset R[[{{\bf s}}]]$. Clearly, the ring $R[[{{\bf s}}]]$ is commutative if and only if $R$ is commutative and $R^{\text{opp}}[[{{\bf s}}]] = R[[{{\bf s}}]]^{\text{opp}} $. Any ring homomorphism $f: R \to R'$ induces a ring homomorphism $$\label{eq:f-bar-ring} \overline{f}: \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} r_\alpha {{\bf s}}^\alpha \in R[[{{\bf s}}]] \longmapsto \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} f(r_\alpha) {{\bf s}}^\alpha \in R'[[{{\bf s}}]],$$ and clearly the correspondences $R \mapsto R[[{{\bf s}}]]$ and $f \mapsto \overline{f}$ define a functor from the category of rings to itself. If ${{\bf s}}=\emptyset$, then $R[[{{\bf s}}]] = R$ and the above functor is the identity. \[def:k-algebra-over-A\] A [*$k$-algebra over $A$*]{} is a (non-necessarily commutative) $k$-algebra $R$ endowed with a map of $k$-algebras $\iota:A \to R$. A map between two $k$-algebras $\iota:A \to R$ and $\iota':A \to R'$ over $A$ is a map $g:R\to R'$ of $k$-algebras such that $\iota' = g {{\scriptstyle \,\circ\,}}\iota$. If $R$ is a $k$-algebra (over $A$), then $R[[{{\bf s}}]]$ is also a $k[[{{\bf s}}]]$-algebra (over $A[[{{\bf s}}]]$). If $M$ is an $(A;A)$-bimodule, we define in a completely similar way the set of formal power series in ${{\bf s}}$ with coefficients in $M$, denoted by $M[[{{\bf s}}]]$. It carries an addition $+$, for which it is an abelian group, and left and right products by elements of $A[[{{\bf s}}]]$. With these operations $M[[{{\bf s}}]]$ becomes an $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-bimodule containing the polynomial $(A[{{\bf s}}];A[{{\bf s}}])$-bimodule $M[{{\bf s}}]$. There is also a natural [*augmentation*]{} $M[[{{\bf s}}]] \to M$ which is a section of the inclusion $M \subset M[{{\bf s}}]$ and $M^{\text{opp}}[[{{\bf s}}]] = M[[{{\bf s}}]]^{\text{opp}}$. If ${{\bf s}}=\emptyset$, then $M[[{{\bf s}}]]=M$. The [*support*]{} of a series $m=\sum_\alpha m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]$ is $\operatorname{\rm supp}(x) := \{ \alpha \in {\mathbb{N}}^{({{\bf s}})}| m_\alpha \neq 0\} \subset {\mathbb{N}}^{({{\bf s}})}$. It is clear that $m=0\Leftrightarrow \operatorname{\rm supp}(m) = \emptyset$. The [*order*]{} of a non-zero series $m=\sum_\alpha m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]$ is $ \operatorname{ord}(m) := \min \{ |\alpha|\ |\ \alpha \in \operatorname{\rm supp}(m) \} \in {\mathbb{N}}$. If $m=0$ we define $\operatorname{ord}(0) = \infty$. It is clear that for $a\in A[[{{\bf s}}]]$ and $m,m'\in M[[{{\bf s}}]]$ we have $\operatorname{\rm supp}(m+m') \subset \operatorname{\rm supp}(m) \cup \operatorname{\rm supp}(m')$, $\operatorname{\rm supp}(am), \operatorname{\rm supp}(ma) \subset \operatorname{\rm supp}(m) + \operatorname{\rm supp}(a)$, $\operatorname{ord}(m+m') \geq \min \{\operatorname{ord}(m),\operatorname{ord}(m')\}$ and $\operatorname{ord}(am), \operatorname{ord}(ma) \geq \operatorname{ord}(a) + \operatorname{ord}(m)$. Moreover, if $\operatorname{ord}(m') > \operatorname{ord}(m)$, then $\operatorname{ord}(m+m') = \operatorname{ord}(m)$. Any $(A;A)$-linear map $h:M \to M'$ between two $(A;A)$-bimodules induces in an obvious way and $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-linear map $$\label{eq:h-bar-mod} \overline{h}: \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]] \longmapsto \sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} h(m_\alpha) {{\bf s}}^\alpha \in M'[[{{\bf s}}]],$$ and clearly the correspondences $M \mapsto M[[{{\bf s}}]]$ and $h \mapsto \overline{h}$ define a functor from the category of $(A;A)$-bimodules to the category $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-bimodules. For each $\beta \in M^{({{\bf s}})}$, let us denote by ${\mathfrak{n}}^M_\beta({{\bf s}})$ the subset of $M[[{{\bf s}}]]$ whose elements are the formal power series $\sum m_\alpha {{\bf s}}^\alpha$ with $m_\alpha = 0$ for all $\alpha \leq \beta$. One has ${\mathfrak{n}}^M_\beta({{\bf s}}) \subset {\mathfrak{n}}^M_\gamma({{\bf s}})$ whenever $\gamma \leq \beta$, and $ {\mathfrak{n}}^M_{\alpha \vee\beta}({{\bf s}}) \subset {\mathfrak{n}}^M_\alpha({{\bf s}}) \cap {\mathfrak{n}}^M_\beta({{\bf s}})$. It is clear that the ${\mathfrak{n}}^M_\beta({{\bf s}})$ are sub-bimodules of $M[[{{\bf s}}]]$ and ${\mathfrak{n}}^A_\beta({{\bf s}}) M[[{{\bf s}}]] \subset {\mathfrak{n}}^M_\beta({{\bf s}})$ and $ M[[{{\bf s}}]] {\mathfrak{n}}^A_\beta({{\bf s}}) \subset {\mathfrak{n}}^M_\beta({{\bf s}})$. For $\beta=0$, ${\mathfrak{n}}^M_0({{\bf s}})$ is the kernel of the augmentation $M[[{{\bf s}}]] \to M$. In the case of a ring $R$, the ${\mathfrak{n}}^R_\beta({{\bf s}})$ are two-sided ideals of $R[[{{\bf s}}]]$, and ${\mathfrak{n}}^R_0({{\bf s}})$ is the kernel of the augmentation $R[[{{\bf s}}]] \to R$. We will consider $R[[{{\bf s}}]]$ as a topological ring with $\{{\mathfrak{n}}^R_\beta({{\bf s}}), \beta \in {\mathbb{N}}^{({{\bf s}})}\}$ as a fundamental system of neighborhoods of $0$. We will also consider $M[[{{\bf s}}]]$ as a topological $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-bimodule with $\{{\mathfrak{n}}^M_\beta({{\bf s}}), \beta \in {\mathbb{N}}^{({{\bf s}})}\}$ as a fundamental system of neighborhoods of $0$ for both, a topological left $A[[{{\bf s}}]]$-module structure and a topological right $A[[{{\bf s}}]]$-module structure. If ${{\bf s}}$ is finite, then ${\mathfrak{n}}^M_\beta({{\bf s}})= \sum_{s\in{{\bf s}}} s^{\beta_s +1} M[[{{\bf s}}]] = \sum_{s\in{{\bf s}}} M[[{{\bf s}}]] s^{\beta_s +1}$ and so the above topologies on $R[[{{\bf s}}]]$, and so on $A[[{{\bf s}}]]$, and on $M[[{{\bf s}}]]$ coincide with the $\langle {{\bf s}}\rangle$-adic topologies. Let us denote by ${\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c} \subset M[{{\bf s}}]$ the intersection of ${\mathfrak{n}}^M_\beta({{\bf s}})$ with $M[{{\bf s}}]$, i.e. the subset of $M[{{\bf s}}]$ whose elements are the finite sums $\sum m_\alpha {{\bf s}}^\alpha$ with $m_\alpha = 0$ for all $\alpha \leq \beta$. It is clear that the natural map $ R[{{\bf s}}]/{\mathfrak{n}}^R_\beta({{\bf s}})^{\bf c} \longrightarrow R[[{{\bf s}}]]/{\mathfrak{n}}^R_\beta({{\bf s}})$ is an isomorphism of rings and the quotient $R[[{{\bf s}}]]/{\mathfrak{n}}^R_\beta({{\bf s}})$ is a finitely generated free left (and right) $R$-module with basis the set of the classes of monomials ${{\bf s}}^\alpha$, $\alpha\leq \beta$. In the same vein, the ${\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c}$ are sub-$(A[{{\bf s}}];A[{{\bf s}}])$-bimodules of $M[{{\bf s}}]$ and the natural map $ M[{{\bf s}}]/{\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c} \longrightarrow M[[{{\bf s}}]]/{\mathfrak{n}}^M_\beta({{\bf s}})$ is an isomorphism of bimodules over $(A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c};A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c})$. Moreover, we have a commutative diagram of natural ${\mathbb{Z}}$-linear isomorphisms $$\label{eq:varrho-lambda} \begin{CD} A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c} \otimes_A M @>{\varrho}>{\simeq}> M[{{\bf s}}]/{\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c} @<{\lambda}<{\simeq}< M \otimes_A A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c}\\ @V{\text{nat.} \otimes {{\rm Id}}}V{\simeq}V @VV{\simeq }V @V{\simeq}V{ {{\rm Id}}\otimes \text{nat.} }V\\ A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}}) \otimes_A M @>{\varrho'}>{\simeq}> M[[{{\bf s}}]]/{\mathfrak{n}}^M_\beta({{\bf s}}) @<{\lambda'}<{\simeq}< M \otimes_A A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}}) \end{CD}$$ where $\varrho$ (resp. $\varrho'$) is an isomorphism of $(A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c};A)$-bimodules (resp. of $(A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}});A)$-bimodules ) and $\lambda$ (resp. $\lambda'$) is an isomorphism of bimodules over $(A;A[{{\bf s}}]/{\mathfrak{n}}^A_\beta({{\bf s}})^{\bf c})$(resp. over $(A;A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}}))$. It is clear that the natural map $$R[[{{\bf s}}]] \longrightarrow \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} R[[{{\bf s}}]]/{\mathfrak{n}}^R_\beta({{\bf s}}) \equiv \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} R[{{\bf s}}]/{\mathfrak{n}}^R_\beta({{\bf s}})^{\bf c}$$ is an isomorphism of rings and so $R[[{{\bf s}}]]$ is complete (hence, separated). Moreover, $R[[{{\bf s}}]]$ appears as the completion of the polynomial ring $R[{{\bf s}}]$ endowed with the topology with $\{{\mathfrak{n}}^R_\beta({{\bf s}})^{\bf c}, \beta\in {\mathbb{N}}^{({{\bf s}})}\}$ as a fundamental system of neighborhoods of $0$. Similarly, the natural map $$M[[{{\bf s}}]] \longrightarrow \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} M[[{{\bf s}}]]/{\mathfrak{n}}^M_\beta({{\bf s}}) \equiv \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} M[{{\bf s}}]/{\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c}$$ is an isomorphism of $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-bimodules, and so $M[[{{\bf s}}]]$ is complete (hence, separated). Moreover, $M[[{{\bf s}}]]$ appears as the completion of the bimodule $M[{{\bf s}}]$ over $(A[{{\bf s}}];A[{{\bf s}}])$ endowed with the topology with $\{{\mathfrak{n}}^M_\beta({{\bf s}})^{\bf c}, \beta\in {\mathbb{N}}^{({{\bf s}})}\}$ as a fundamental system of neighborhoods of $0$. Since the subsets $\{\alpha \in {\mathbb{N}}^{({{\bf s}})}\ |\ \alpha \leq \beta\}$, $\beta\in {\mathbb{N}}^{({{\bf s}})}$, are cofinal among the finite subsets of ${\mathbb{N}}^{({{\bf s}})}$, the additive isomorphism $$\sum_{\scriptscriptstyle \alpha \in {\mathbb{N}}^{({{\bf s}})}} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]] \mapsto \{m_\alpha\}_{\alpha \in {\mathbb{N}}^{({{\bf s}})}} \in M^{{\mathbb{N}}^{({{\bf s}})}}$$ is a homeomorphism, where $M^{{\mathbb{N}}^{({{\bf s}})}}$ is endowed with the product of discrete topologies on each copy of $M$. In particular, any formal power series $ \sum m_\alpha {{\bf s}}^\alpha$ is the limit of its finite partial sums $ \sum_{\alpha \in F} m_\alpha {{\bf s}}^\alpha$, over the filter of finite subsets $F\subset {\mathbb{N}}^{({{\bf s}})}$. Since the quotients $A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}})$ are free $A$-modules, we have exact sequences $$0 \longrightarrow {\mathfrak{n}}^A_\beta({{\bf s}}) \otimes_A M \longrightarrow A[[{{\bf s}}]]\otimes_A M \longrightarrow \frac{A[[{{\bf s}}]]}{{\mathfrak{n}}^A_\beta({{\bf s}})} \otimes_A M \longrightarrow 0$$ and the tensor product $A[[{{\bf s}}]]\otimes_A M$ is a topological left $A[[{{\bf s}}]]$-module with $\{{\mathfrak{n}}^A_\beta({{\bf s}})\otimes_A M, \beta \in {\mathbb{N}}^{({{\bf s}})}\}$ as a fundamental system of neighborhoods of $0$. The natural $(A[[{{\bf s}}]];A)$-linear map $$A[[{{\bf s}}]]\otimes_A M \longrightarrow M[[{{\bf s}}]]$$ is continuous and, if we denote by $ A[[{{\bf s}}]]\widehat{\otimes}_A M$ the completion of $ A[[{{\bf s}}]]\otimes_A M$, the induced map $ A[[{{\bf s}}]]\widehat{\otimes}_A M \longrightarrow M[[{{\bf s}}]]$ is an isomorphism of $(A[[{{\bf s}}]];A)$-bimodules, since we have natural $(A[[{{\bf s}}]];A)$-linear isomorphisms $$\left( A[[{{\bf s}}]]\otimes_A M \right)/\left( {\mathfrak{n}}^A_\beta({{\bf s}})\otimes_A M \right) \simeq \left( A[[{{\bf s}}]]/{\mathfrak{n}}^A_\beta({{\bf s}}) \right)\otimes_A M \simeq M[[{{\bf s}}]]/{\mathfrak{n}}^M_\beta({{\bf s}})$$ for $\beta \in {\mathbb{N}}^{({{\bf s}})}$, and so $$\label{eq:A[[s]]wotM} A[[{{\bf s}}]]\widehat{\otimes}_A M = \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} \left(\frac{ A[[{{\bf s}}]]\otimes_A M }{ {\mathfrak{n}}^A_\beta({{\bf s}})\otimes_A M }\right) \simeq \lim_{\stackrel{\longleftarrow}{\beta\in {\mathbb{N}}^{({{\bf s}})}}} \left( \frac{M[[{{\bf s}}]]}{{\mathfrak{n}}^M_\beta({{\bf s}})} \right) \simeq M[[{{\bf s}}]].$$ Similarly, the natural $(A;A[[{{\bf s}}]])$-linear map $M\otimes_A A[[{{\bf s}}]] \to M[[{{\bf s}}]]$ induces an isomorphism $ M \widehat{\otimes}_A A[[{{\bf s}}]] \xrightarrow{\sim} M[[{{\bf s}}]]$ of $(A;A[[{{\bf s}}]])$-bimodules. If $h: M \to M'$ is an $(A;A)$-linear map between two $(A;A)$-bimodules, the induced map $\overline{h}:M[[{{\bf s}}] \to M'[[{{\bf s}}]$ (see (\[eq:h-bar-mod\])) is clearly continuous and there is a commutative diagram $$\begin{CD} A[[{{\bf s}}]]\widehat{\otimes}_A M @>{\simeq}>> M[[{{\bf s}}]] @<{\simeq}<< M \widehat{\otimes}_A A[[{{\bf s}}]] \\ @V{{{\rm Id}}\widehat{\otimes} h }VV @V{\overline{h}}VV @V{h\widehat{\otimes} {{\rm Id}}}VV\\ A[[{{\bf s}}]]\widehat{\otimes}_A M' @>{\simeq}>> M'[[{{\bf s}}]] @<{\simeq}<< M' \widehat{\otimes}_A A[[{{\bf s}}]]. \end{CD}$$ Similarly, for any ring homomorphism $f: R \to R'$, the induced ring homomorphism $ \overline{f}: R[[{{\bf s}}]] \to R'[[{{\bf s}}]]$ is also continuous. We say that a subset $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$ is an [*ideal*]{} of ${\mathbb{N}}^{({{\bf s}})}$ (resp. a [*co-ideal*]{} of ${\mathbb{N}}^{({{\bf s}})}$) if whenever $\alpha\in\Delta$ and $\alpha\leq \alpha'$ (resp. $\alpha'\leq \alpha$), then $\alpha'\in\Delta$. It is clear that $\Delta$ is an ideal if and only if its complement $\Delta^c$ is a co-ideal, and that the union and the intersection of any family of ideals (resp. of co-ideals) of ${\mathbb{N}}^{({{\bf s}})}$ is again an ideal (resp. a co-ideal) of ${\mathbb{N}}^{({{\bf s}})}$. Examples of ideals (resp. of co-ideals) of ${\mathbb{N}}^{({{\bf s}})}$ are the $\beta + {\mathbb{N}}^{({{\bf s}})}$ (resp. the ${\mathfrak{n}}_\beta({{\bf s}}) := \{\alpha \in {\mathbb{N}}^{({{\bf s}})}\ |\ \alpha\leq \beta\}$) with $\beta \in {\mathbb{N}}^{({{\bf s}})}$. The ${\mathfrak{t}}_m({{\bf s}}) := \{\alpha \in {\mathbb{N}}^{({{\bf s}})}\ |\ |\alpha|\leq m\}$ with $m\geq 0$ are also co-ideals. Actually, a subset $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$ is an ideal (resp. a co-ideal) if and only if $\Delta = \cup_{\beta\in \Delta} \left(\beta + {\mathbb{N}}^{({{\bf s}})}\right) = \Delta + {\mathbb{N}}^{({{\bf s}})} $ (resp. $\Delta = \cup_{\beta\in \Delta} {\mathfrak{n}}_\beta({{\bf s}})$). We say that a co-ideal $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$ is bounded if there is an integer $m\geq 0$ such that $|\alpha|\leq m$ for all $\alpha\in\Delta$. In other words, a co-ideal $\Delta\subset {\mathbb{N}}^{({{\bf s}})}$ is bounded if and only if there is an integer $m\geq 0$ such that $\Delta \subset {\mathfrak{t}}_m({{\bf s}})$. Also, a co-ideal $\Delta\subset {\mathbb{N}}^{({{\bf s}})}$ is non-empty if and only if ${\mathfrak{t}}_0({{\bf s}}) = {\mathfrak{n}}_0({{\bf s}}) =\{0\} \subset \Delta$. For a co-ideal $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$ and an integer $m\geq 0$, we denote $\Delta^m := \Delta \cap {\mathfrak{t}}_m({{\bf s}})$.\ For each co-ideal $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$, we denote by $\Delta_M$ the sub-$(A[[{{\bf s}}];A[[{{\bf s}}]])$-bimodule of $M[[{{\bf s}}]]$ whose elements are the formal power series $\sum_{\alpha\in{\mathbb{N}}^{({{\bf s}})}} m_\alpha {{\bf s}}^\alpha$ such that $m_\alpha=0$ whenever $\alpha\in \Delta$. One has $$\begin{aligned} &\displaystyle \Delta_M = \cdots = \left\{m\in M[[{{\bf s}}]]\ |\ \operatorname{\rm supp}(m) \subset \bigcap_{\beta\in \Delta} {\mathfrak{n}}_\beta({{\bf s}})^c\right\}= &\\ &\displaystyle \bigcap _{\beta\in \Delta} \left\{m\in M[[{{\bf s}}]]\ |\ \operatorname{\rm supp}(m) \subset {\mathfrak{n}}_\beta({{\bf s}})^c\right\} = \bigcap_{\beta\in \Delta} {\mathfrak{n}}^M_\beta({{\bf s}}),\end{aligned}$$ and so $\Delta_M$ is closed in $M[[{{\bf s}}]]$. Let $\Delta' \subset {\mathbb{N}}^{({{\bf s}})}$ be another co-ideal. We have $$\Delta_M + \Delta'_M = (\Delta \cap \Delta')_M.$$ If $\Delta \subset \Delta'$, then $\Delta'_M \subset \Delta_M$, and if $a\in \Delta'_A$, $m\in \Delta_M$ we have $$\operatorname{\rm supp}(am) \subset \operatorname{\rm supp}(a) + \operatorname{\rm supp}(m) \subset \left(\Delta'\right)^c + \Delta^c \subset \left(\Delta'\right)^c \cap \Delta^c =\left( \Delta' \cup \Delta\right)^c ,$$ and so $ \Delta'_A \Delta_M \subset (\Delta' \cup \Delta)_M$. Is a similar way we obtain $ \Delta_M \Delta'_A \subset (\Delta' \cup \Delta)_M$.\ Let us denote by $M[[{{\bf s}}]]_\Delta := M[[{{\bf s}}]]/\Delta_M$ endowed with the quotient topology. The elements in $M[[{{\bf s}}]]_\Delta$ are power series of the form $$\sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha,\quad m_\alpha \in M.$$ It is clear that $M[[{{\bf s}}]]_\Delta$ is a topological $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-bimodule. A fundamental system of neighborhoods of $0$ in $M[[{{\bf s}}]]_\Delta$ consist of $$\frac{{\mathfrak{n}}^M_\beta({{\bf s}}) + \Delta_M}{\Delta_M} = \frac{({\mathfrak{n}}_\beta({{\bf s}}) \cap\Delta)_M}{\Delta_M},\quad \beta \in {\mathbb{N}}^{({{\bf s}})},$$ and since the subsets ${\mathfrak{n}}_\beta({{\bf s}}) \cap \Delta, \beta \in {\mathbb{N}}^{({{\bf s}})}$, are cofinal among the finite subsets of $\Delta$, we conclude that the additive isomorphism $$\sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]_\Delta \mapsto \{m_\alpha\}_{\alpha \in \Delta} \in M^\Delta$$ is a homeomorphism, where $M^\Delta$ is endowed with the product of discrete topologies on each copy of $M$.\ For $\Delta \subset \Delta'$ co-ideals of ${\mathbb{N}}^{({{\bf s}})}$, we have natural continuous $(A[[{{\bf s}}]]_{\Delta'};A[[{{\bf s}}]]_{\Delta'})$-linear projections $\tau_{\Delta' \Delta}:M[[{{\bf s}}]]_{\Delta'} \longrightarrow M[[{{\bf s}}]]_\Delta$, that we also call [*truncations*]{}, $$\tau_{\Delta' \Delta} : \sum_{\scriptscriptstyle \alpha\in\Delta'} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]_{\Delta'} \longmapsto \sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]_{\Delta},$$ and continuous $(A;A)$-linear scissions $$\sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]_\Delta \longmapsto \sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha \in M[[{{\bf s}}]]_{\Delta'}.$$ which are topological immersions. In particular we have natural continuous $(A;A)$-linear topological embeddings $M[[{{\bf s}}]]_\Delta \hookrightarrow M[[{{\bf s}}]]$ and we define the [*support*]{} (resp. the [*order*]{}) of any element in $M[[{{\bf s}}]]_\Delta$ as its support (resp. its order) as element of $M[[{{\bf s}}]]$. We have a bicontinuous isomorphism of $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-bimodules $$M[[{{\bf s}}]]_\Delta = \lim_{\stackrel{\longleftarrow}{m\in{\mathbb{N}}}} M[[{{\bf s}}]]_{\Delta^m}.$$ For a ring $R$, the $\Delta_R$ are two-sided closed ideals of $R[[{{\bf s}}]]$, $\Delta_R\Delta'_R \subset (\Delta\cup\Delta')_R$ and we have a bicontinuous ring isomorphism $$R[[{{\bf s}}]]_\Delta = \lim_{\stackrel{\longleftarrow}{m\in{\mathbb{N}}}} R[[{{\bf s}}]]_{\Delta^m}.$$ When ${{\bf s}}$ is finite, ${\mathfrak{t}}_m({{\bf s}})_R$ coincides with the $(m+1)$-power of the two-sided ideal generated by all the variables $s\in {{\bf s}}$. As in (\[eq:A\[\[s\]\]wotM\]) one proves that $A[[{{\bf s}}]]_\Delta \otimes_A M$ (resp. $M\otimes_A A[[{{\bf s}}]]_\Delta$) is endowed with a natural topology in such a way that the natural map $A[[{{\bf s}}]]_\Delta \otimes_A M \to M[[{{\bf s}}]]_\Delta$ (resp. $M\otimes_A A[[{{\bf s}}]]_\Delta \to M[[{{\bf s}}]]_\Delta$) is continuous and gives rise to a $(A[[{{\bf s}}]]_\Delta;A)$-linear (resp. to a $(A;A[[{{\bf s}}]]_\Delta)$-linear) isomorphism $$A[[{{\bf s}}]]_\Delta \widehat{\otimes}_A M \xrightarrow{\sim} M[[{{\bf s}}]]_\Delta\quad\quad \text{(resp. $ M\widehat{\otimes}_A A[[{{\bf s}}]]_\Delta \xrightarrow{\sim} M[[{{\bf s}}]]_\Delta$)}.$$ If $h: M \to M'$ is an $(A;A)$-linear map between two $(A;A)$-bimodules, the map $\overline{h}:M[[{{\bf s}}] \to M'[[{{\bf s}}]$ (see (\[eq:h-bar-mod\])) obviously satisfies $\overline{h}(\Delta_M) \subset \Delta_{M'}$, and so induces another natural $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-linear continuous map $M[[{{\bf s}}]_\Delta \to M'[[{{\bf s}}]_\Delta$, that will be still denoted by $\overline{h}$. We have a commutative diagram $$\begin{CD} A[[{{\bf s}}]]_\Delta\widehat{\otimes}_A M @>{\simeq}>> M[[{{\bf s}}]]_\Delta @<{\simeq}<< M \widehat{\otimes}_A A[[{{\bf s}}]]_\Delta \\ @V{{{\rm Id}}\widehat{\otimes} h }VV @V{\overline{h}}VV @V{h\widehat{\otimes} {{\rm Id}}}VV\\ A[[{{\bf s}}]]_\Delta\widehat{\otimes}_A M' @>{\simeq}>> M'[[{{\bf s}}]]_\Delta @<{\simeq}<< M' \widehat{\otimes}_A A[[{{\bf s}}]]_\Delta. \end{CD}$$ In the same way that the correspondences $M \mapsto M[[{{\bf s}}]]$ and $h \mapsto \overline{h}$ define a functor from the category of $(A;A)$-bimodules to the category of $(A[[{{\bf s}}]];A[[{{\bf s}}]])$-bimodules, we may consider functors $M \mapsto M[[{{\bf s}}]]_\Delta$ and $h \mapsto \overline{h}$ from the category of $(A;A)$-bimodules to the category of $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-bimodules. We may also consider functors $R \mapsto R[[{{\bf s}}]]_\Delta$ and $f \mapsto \overline{f}$ from the category of rings to itself. Moreover, if $R$ is a $k$-algebra (over $A$), then $R[[{{\bf s}}]]_\Delta$ is a $k[[{{\bf s}}]]_\Delta$-algebra (over $A[[{{\bf s}}]]_\Delta$). \[lema:topol-of-M\[\[s\]\]\] Under the above hypotheses, $\Delta_M$ is the closure of $\Delta_{\mathbb{Z}}M[[{{\bf s}}]]$. Any element in $\Delta_M$ is of the form $\sum_{\scriptscriptstyle \alpha\in\Delta} m_\alpha {{\bf s}}^\alpha$, but ${{\bf s}}^\alpha m_\alpha \in \Delta_{\mathbb{Z}}M[[{{\bf s}}]]$ whenever $\alpha\in\Delta$ and so it belongs to the closure of $\Delta_{\mathbb{Z}}M[[{{\bf s}}]]$. \[lemma:unit-psr\] Let $R$ be a ring, ${{\bf s}}$ a set and $\Delta\subset {\mathbb{N}}^{({{\bf s}})}$ a non-empty co-ideal. The units in $R[[{{\bf s}}]]_\Delta$ are those power series $r=\sum r_\alpha {{\bf s}}^\alpha$ such that $r_0$ is a unit in $R$. Moreover, in the special case where $r_0=1$, the inverse $r^*= \sum r^*_\alpha {{\bf s}}^\alpha$ of $r$ is given by $r^*_0 = 1$ and $$r^*_\alpha = \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\alpha|} (-1)^d \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} r_{\alpha^1} \cdots r_{\alpha^d}\quad \text{for\ }\ \alpha\neq 0,$$ where $\operatorname{\mathcal{P}}(\alpha,d)$ is the set of $d$-uples $\alpha^\bullet=(\alpha^1,\dots,\alpha^d)$ with $\alpha^i \in {\mathbb{N}}^{({{\bf s}})}$, $\alpha^i\neq 0$, and $\alpha^1+\cdots + \alpha^d=\alpha$. The proof is standard and it is left to the reader. It is clear that $r_0$ to be a unit in $R$ is a necessary condition for $r$ being a unit in $R[[{{\bf s}}]]_\Delta$. For the opposite, it is enough to treat the case $r_0=1$. Let us consider the element $r^*\in R[[{{\bf s}}]]_\Delta$ defined above and set $u= r r^*$. We obviously have $u_0=1$. For $\alpha\in\Delta, \alpha \neq 0$ we have $$\begin{aligned} & \displaystyle u_\alpha = \sum_{\scriptscriptstyle \beta + \gamma=\alpha} r^{}_\beta r^*_\gamma = r_\alpha + \sum_{\substack{\scriptscriptstyle \beta + \gamma=\alpha\\ \scriptscriptstyle |\gamma|>0 }} r_\beta \left( \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\gamma|} (-1)^d \left( \sum_{\scriptscriptstyle \gamma^\bullet \in \operatorname{\mathcal{P}}(\gamma,d)} r_{\gamma^1} \cdots r_{\gamma^d} \right) \right) = & \\ & \displaystyle r_\alpha + \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\alpha|} (-1)^d \left( \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} r_{\alpha^1} \cdots r_{\alpha^d} \right) + \sum_{\substack{\scriptscriptstyle \beta + \gamma=\alpha\\ \scriptscriptstyle |\beta|,|\gamma|>0 }} r_\beta \left( \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\gamma|} (-1)^d \left( \sum_{\scriptscriptstyle \gamma^\bullet \in \operatorname{\mathcal{P}}(\gamma,d)} r_{\gamma^1} \cdots r_{\gamma^d} \right) \right) = & \\ & \displaystyle r_\alpha + \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\alpha|} (-1)^d \left( \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} r_{\alpha^1} \cdots r_{\alpha^d} \right) + \sum_{\scriptscriptstyle d=1}^{\scriptscriptstyle |\alpha|-1} (-1)^d \left( \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d+1)} r_{\alpha^1} \cdots r_{\alpha^{d+1}} \right) = & \\ & \displaystyle \sum_{\scriptscriptstyle d=2}^{\scriptscriptstyle |\alpha|} (-1)^d \left( \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} r_{\alpha^1} \cdots r_{\alpha^d} \right) + \sum_{\scriptscriptstyle d=2}^{\scriptscriptstyle |\alpha|} (-1)^{d-1} \left( \sum_{\scriptscriptstyle \alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} r_{\alpha^1} \cdots r_{\alpha^d} \right) =0\end{aligned}$$ and so $r r^* =1$. In a similar way one proves that $r^* r=1$. We conclude that $r$ is a unit and its inverse $r^*$ is given by the expression above. \[notacion:Ump\] Let $R$ be a ring, ${{\bf s}}$ a set and $\Delta\subset {\mathbb{N}}^{({{\bf s}})}$ a non-empty co-ideal. We denote by $\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ the multiplicative sub-group of the units of $R[[{{\bf s}}]]_\Delta$ whose 0-degree coefficient is $1$. Clearly, $\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)^{\text{\rm opp}} = \operatorname{\mathcal{U}}^{{\bf s}}(R^{\text{\rm opp}};\Delta)$. For $\Delta\subset \Delta'$ co-ideals we have $\tau_{\Delta'\Delta}\left(\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta')\right) \subset \operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ and the truncation map $\tau_{\Delta'\Delta}:\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta') \to \operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ is a group homomorphisms. Clearly, we have $$\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta) = \lim_{\stackrel{\longleftarrow}{m\in{\mathbb{N}}}} \operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta^m).$$ For any ring homomorphism $f:R\to R'$, the induced ring homomorphism $\overline{f}: R[[{{\bf s}}]]_\Delta \to R'[[{{\bf s}}]]_\Delta$ sends $\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ into $\operatorname{\mathcal{U}}^{{\bf s}}(R';\Delta)$ and so it induces natural group homomorphisms $\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta) \to \operatorname{\mathcal{U}}^{{\bf s}}(R';\Delta)$. \[defi:external-x\] Let $R$ be a ring, ${{\bf s}},{{\bf t}}$ sets and $\nabla\subset {\mathbb{N}}^{({{\bf s}})}, \Delta\subset {\mathbb{N}}^{({{\bf t}})}$ non-empty co-ideals. For each $r\in R[[{{\bf s}}]]_\nabla, r'\in R[[{{\bf t}}]]_\Delta$, the [*external product*]{} $r{{\scriptstyle \,\boxtimes\,}}r'\in R[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$ is defined as $$r{{\scriptstyle \,\boxtimes\,}}r' := \sum_{\scriptscriptstyle (\alpha,\beta)\in \nabla \times \Delta} r_\alpha r'_\beta {{\bf s}}^\alpha {{\bf t}}^\beta.$$ Let us notice that the above definition is consistent with the existence of natural isomorphism of $(R;R)$-bimodules $R[[{{\bf s}}]]_\nabla \widehat{\otimes}_R R[[{{\bf t}}]]_\Delta \simeq R[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}\simeq R[[{{\bf t}}\sqcup {{\bf s}}]]_{\Delta \times \nabla} \simeq R[[{{\bf t}}]]_\Delta \widehat{\otimes}_R R[[{{\bf s}}]]_\nabla$. Let us also notice that $1{{\scriptstyle \,\boxtimes\,}}1 = 1$ and $r{{\scriptstyle \,\boxtimes\,}}r' = (r {{\scriptstyle \,\boxtimes\,}}1) (1 {{\scriptstyle \,\boxtimes\,}}r')$. Moreover, if $r\in \operatorname{\mathcal{U}}^{{\bf s}}(R;\nabla)$, $r'\in \operatorname{\mathcal{U}}^{{\bf t}}(R;\Delta)$, then $r{{\scriptstyle \,\boxtimes\,}}r' \in \operatorname{\mathcal{U}}^{{{\bf s}}\sqcup{{\bf t}}}(R;\nabla\times \Delta)$ and $(r{{\scriptstyle \,\boxtimes\,}}r')^* = {r'}^* {{\scriptstyle \,\boxtimes\,}}r^*$. Let $k \to A$ be a ring homomorphism between commutative rings, $E,F$ two $A$-modules, ${{\bf s}}$ a set and $\Delta\subset {\mathbb{N}}^{({{\bf s}})}$ a non-empty co-ideal, i.e ${\mathfrak{n}}_0({{\bf s}}) = \{0\} \subset \Delta$. \[prop:induced-cont-maps-M\[\[s\]\]\_m\] Under the above hypotheses, let $f:E[[{{\bf s}}]]_\Delta \to F[[{{\bf s}}]]_\Delta$ be a continuous $k[[{{\bf s}}]]_\Delta$-linear map. Then, for any co-ideal $\Delta'\subset {\mathbb{N}}^{({{\bf s}})}$ with $\Delta'\subset \Delta$ we have $$f\left(\Delta'_E/\Delta_E\right) \subset \Delta'_F/\Delta_F$$ and so there is a unique continuous $k[[{{\bf s}}]]_{\Delta'}$-linear map $\overline{f}:E[[{{\bf s}}]]_{\Delta'} \to F[[{{\bf s}}]]_{\Delta'}$ such that the following diagram is commutative $$\begin{CD} E[[{{\bf s}}]]_\Delta @>{f}>> F[[{{\bf s}}]]_\Delta\\ @V{\text{nat.}}VV @VV{\text{nat.}}V \\ E[[{{\bf s}}]]_{\Delta'} @>{\overline{f}}>> F[[{{\bf s}}]]_{\Delta'}. \end{CD}$$ It is a straightforward consequence of Lemma \[lema:topol-of-M\[\[s\]\]\]. Under the above hypotheses, the set of all continuous $k[[{{\bf s}}]]_\Delta$-linear maps from $E[[{{\bf s}}]]_\Delta$ to $F[[{{\bf s}}]]_\Delta$ will be denoted by $$\operatorname{Hom}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta,F[[{{\bf s}}]]_\Delta).$$ It is an $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-bimodule central over $k[[{{\bf s}}]]_\Delta$. For any co-ideals $\Delta'\subset\Delta\subset {\mathbb{N}}^{({{\bf s}})}$, Proposition \[prop:induced-cont-maps-M\[\[s\]\]\_m\] provides a natural $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-linear map $$\operatorname{Hom}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta,F[[{{\bf s}}]]_\Delta) \longrightarrow \operatorname{Hom}_{k[[{{\bf s}}]]_{\Delta'}}^{\text{\rm top}}(E[[{{\bf s}}]]_{\Delta'},F[[{{\bf s}}]]_{\Delta'}).$$ For $E=F$, $\operatorname{End}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta)$ is a $k[[{{\bf s}}]]_\Delta$-algebra over $A[[{{\bf s}}]]_\Delta$. [[**.  **]{}]{}\[nume:tilde\] For each $r= \sum_\beta r_\beta {{\bf s}}^\beta \in \operatorname{Hom}_k(E,F)[[{{\bf s}}]]_\Delta$ we define $\widetilde{r}: E[[{{\bf s}}]]_\Delta \to F[[{{\bf s}}]]_\Delta$ by $$\widetilde{r} \left( \sum_{\scriptscriptstyle \alpha\in \Delta} e_\alpha {{\bf s}}^\alpha \right) := \sum_{\scriptscriptstyle \alpha\in \Delta } \left( \sum_{\scriptscriptstyle \beta + \gamma=\alpha} r_\beta(e_\gamma) \right) {{\bf s}}^\alpha,$$ which is obviously a continuous $k[[{{\bf s}}]]_\Delta$-linear map. Let us notice that $ \widetilde{r} = \sum_\beta {{\bf s}}^\beta \widetilde{r_\beta}$. It is clear that the map $$\label{eq:tilde-map} r\in \operatorname{Hom}_k(E,F)[[{{\bf s}}]]_\Delta \longmapsto \widetilde{r} \in \operatorname{Hom}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta,F[[{{\bf s}}]]_\Delta)$$ is $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-linear. If $f:E[[{{\bf s}}]]_\Delta \to F[[{{\bf s}}]]_\Delta$ is a continuous $k[[{{\bf s}}]]_\Delta$-linear map, let us denote by $f_\alpha:E \to F$, $\alpha\in\Delta $, the $k$-linear maps defined by $$f(e) = \sum_{\scriptscriptstyle \alpha\in \Delta} f_\alpha(e) {{\bf s}}^\alpha,\quad \forall e\in E.$$ If $g:E\to F[[{{\bf s}}]]_\Delta$ is a $k$-linear map, we denote by $g^e:E[[{{\bf s}}]]_\Delta \to F[[{{\bf s}}]]_\Delta$ the unique continuous $k[[{{\bf s}}]]_\Delta$-linear map extending $g$ to $E[[{{\bf s}}]]_\Delta = k[[{{\bf s}}]]_\Delta \widehat{\otimes}_k E$. It is given by $$g^e \left( \sum_\alpha e_\alpha {{\bf s}}^\alpha \right) := \sum_\alpha g(e_\alpha) {{\bf s}}^\alpha.$$ We have a $k[[{{\bf s}}]]_\Delta$-bilinear and $A[[{{\bf s}}]]_\Delta$-balanced map $$\langle -,-\rangle : (r,e) \in \operatorname{Hom}_k(E,F)[[{{\bf s}}]]_\Delta \times E[[{{\bf s}}]]_\Delta \longmapsto \langle r,e\rangle := \widetilde{r}(e) \in F[[{{\bf s}}]]_\Delta.$$ \[lema:tilde-map\] With the above hypotheses, the following properties hold: 1. The map (\[eq:tilde-map\]) is an isomorphism of $(A[[{{\bf s}}]]_\Delta;A[[{{\bf s}}]]_\Delta)$-bimodules. When $E=F$ it is an isomorphism of $k[[{{\bf s}}]]_\Delta$-algebras over $A[[{{\bf s}}]]_\Delta$. 2. The restriction map $$f \in \operatorname{Hom}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta,F[[{{\bf s}}]]_\Delta) \mapsto f|_E \in \operatorname{Hom}_k(E,F[[{{\bf s}}]]_\Delta)$$ is an isomorphism of $(A[[{{\bf s}}]]_\Delta;A)$-bimodules. \(1) One easily sees that the inverse map of $r \mapsto \widetilde{r}$ is $f \mapsto \sum_\alpha f_\alpha {{\bf s}}^\alpha$. \(2) One easily sees that the inverse map of the restriction map $f \mapsto f|_E$ is $g \mapsto g^e$. Let us call $R = \operatorname{End}_k(E)$. As a consequence of the above lemma, the composition of the maps $$\label{eq:comple_formal_1} R[[{{\bf s}}]]_\Delta \xrightarrow{r \mapsto \widetilde{r}} \operatorname{End}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta) \xrightarrow{f \mapsto f|_E} \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)$$ is an isomorphism of $(A[[{{\bf s}}]]_\Delta;A)$-bimodules, and so $\operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)$ inherits a natural structure of $k[[{{\bf s}}]]_\Delta$-algebra over $A[[{{\bf s}}]]_\Delta$. Namely, if $g,h\in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)$ with $$g(e)=\sum_{\scriptscriptstyle \alpha\in \Delta} g_\alpha(e){{\bf s}}^\alpha,\ h(e)=\sum_{\scriptscriptstyle \alpha\in \Delta} h_\alpha(e){{\bf s}}^\alpha,\quad \forall e\in E,\quad g_\alpha, h_\alpha \in \operatorname{Hom}_k(E,E),$$ then the product $h g \in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)$ is given by $$\label{eq:product-HomEE[[s]]} (hg)(e) = \sum_{\scriptscriptstyle \alpha\in \Delta} \left( \sum_{\scriptscriptstyle \beta + \gamma = \alpha} (h_\beta {{\scriptstyle \,\circ\,}}g_\gamma)(e) \right) {{\bf s}}^\alpha.$$ \[defi:external-prod-of-maps\] Let ${{\bf s}},{{\bf t}}$ be sets and $\Delta\subset{\mathbb{N}}^{({{\bf s}})}, \nabla\subset{\mathbb{N}}^{({{\bf t}})}$ non-empty co-ideals. For each $f\in\operatorname{End}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta)$ and each $g\in\operatorname{End}_{k[[{{\bf t}}]]_\nabla}^{\text{\rm top}}(E[[{{\bf t}}]]_\nabla)$, with $$f(e)=\sum_{\scriptscriptstyle \alpha \in \Delta} f_\alpha(e) {{\bf s}}^\alpha, \quad g(e)=\sum_{\scriptscriptstyle \beta \in \nabla} g_\beta(e) {{\bf t}}^\beta\quad \forall e\in E,$$ we define $f{{\scriptstyle \,\boxtimes\,}}g\in \operatorname{End}_{k[[{{\bf s}}\sqcup {{\bf t}}]]_{\Delta\times\nabla}}^{\text{\rm top}}(E[[{{\bf s}}\sqcup {{\bf t}}]]_{\Delta\times\nabla})$ as $f{{\scriptstyle \,\boxtimes\,}}g := h^e$, with: $$h(x) := \sum_{\scriptscriptstyle (\alpha,\beta) \in \Delta\times\nabla} (f_\alpha {{\scriptstyle \,\circ\,}}g_\beta)(x) {{\bf s}}^\alpha {{\bf t}}^\beta\quad \forall x\in E.$$ The proof of the following lemma is clear and it is left to the reader. \[lemma:tilde-otimes\] With the above hypotheses, or each $r\in R[[{{\bf s}}]]_\Delta, r'\in R[[{{\bf t}}]]_\nabla$, we have $ \widetilde{r{{\scriptstyle \,\boxtimes\,}}r'} = \widetilde{r}{{\scriptstyle \,\boxtimes\,}}\widetilde{r'}$ (see Definition \[defi:external-x\]). Let us call $R = \operatorname{End}_k(E)$. For any $r\in R[[{{\bf s}}]]_\Delta$, the following properties are equivalent: 1. $r_0 = {{\rm Id}}$. 2. The endomorphism $\widetilde{r}$ is compatible with the natural augmentation $E[[{{\bf s}}]]_\Delta \to E$, i.e. $\widetilde{r}(e) \equiv e \mod {\mathfrak{n}}^E_0({{\bf s}})/\Delta_E$ for all $e\in E[[{{\bf s}}]]_\Delta$. Moreover, if the above properties hold, then $\widetilde{r}: E[[{{\bf s}}]]_\Delta \to E[[{{\bf s}}]]_\Delta$ is a bi-continuous $k[[{{\bf s}}]]_\Delta$-linear automorphism. The equivalence of (a) and (b) is clear. For the second part, $r$ is invertible since $r_0 = {{\rm Id}}$. So $\widetilde{r}$ is invertible too and $\widetilde{r}^{-1} = \widetilde{r^{-1}} $ is also continuous. \[notacion:pcirc\] We denote: $$\begin{aligned} & \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf s}}]]_\Delta) := &\\ &\left\{ f \in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)\ |\ f(e) \equiv e\!\!\!\!\mod {\mathfrak{n}}^E_0({{\bf s}})/\Delta_E\quad \forall e\in E \right\}, \end{aligned}$$ $$\begin{aligned} & \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(E[[{{\bf s}}]]_\Delta) := &\\ & \left\{ f \in \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf s}}]]_\Delta)\ |\ f(e) \equiv e_0\!\!\!\!\mod {\mathfrak{n}}^E_0({{\bf s}})/\Delta_E\quad \forall e\in E[[{{\bf s}}]]_\Delta \right\}.\end{aligned}$$ Let us notice that a $f \in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\Delta)$, given by $f(e) = \sum_{\alpha\in \Delta} f_\alpha(e) {{\bf s}}^\alpha$, belongs to $\operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf s}}]]_\Delta)$ if and only if $f_0={{\rm Id}}_E$. The isomorphism in (\[eq:comple\_formal\_1\]) gives rise to a group isomorphism $$\label{eq:U-iso-pcirc} r\in \operatorname{\mathcal{U}}^{{\bf s}}(\operatorname{End}_k(E);\Delta) \stackrel{\sim}{\longmapsto} \widetilde{r} \in \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(E[[{{\bf s}}]]_\Delta)$$ and to a bijection $$\label{eq:Aut-iso-pcirc} f\in \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(E[[{{\bf s}}]]_\Delta) \stackrel{\sim}{\longmapsto} f|_E \in \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf s}}]]_\Delta).$$ So, $\operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf s}}]]_\Delta)$ is naturally a group with the product described in (\[eq:product-HomEE\[\[s\]\]\]). Substitution maps {#nume:substitution_maps} ================= In this section we will assume that $k$ is a commutative ring and $A$ a commutative $k$-algebra. The following notation will be used extensively. 1. For each integer $r\geq 0$ let us denote $[r]:= \{1,\dots,r\}$ if $r>0$ and $[0]=\emptyset$. 2. Let ${{\bf s}}$ be a set. Maps from a set $\Lambda$ to ${\mathbb{N}}^{({{\bf s}})}$ will be usually denoted as $\alpha^\bullet: l \in \Lambda \longmapsto \alpha^l\in {\mathbb{N}}^{({{\bf s}})}$, and its [*support*]{} is defined by $\operatorname{\rm supp}\alpha^\bullet := \{l \in \Lambda\ |\ \alpha^l\neq 0\}$. 3. For each set $\Lambda$ and for each map $\alpha^\bullet: \Lambda \to {\mathbb{N}}^{({{\bf s}})}$ with finite support, its [*norm*]{} is defined by $|\alpha^\bullet| := \sum_{l\in \operatorname{\rm supp}\alpha^\bullet} \alpha^l = \sum_{l\in \Lambda} \alpha^l$. When $\Lambda=\emptyset$, the unique map $\Lambda \to {\mathbb{N}}^{({{\bf s}})}$ is the inclusion $\emptyset \hookrightarrow {\mathbb{N}}^{({{\bf s}})}$ and its norm is $0\in {\mathbb{N}}^{({{\bf s}})}$. 4. If $\Lambda$ is a set and $e\in {\mathbb{N}}^{({{\bf s}})}$, we define $$\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,\Lambda) := \{ \alpha^\bullet : \Lambda \to {\mathbb{N}}^{({{\bf s}})}\ |\ \# \operatorname{\rm supp}\alpha^\bullet < +\infty, |\alpha^\bullet| = e\}.$$ If $F$ is a finite set and $e\in {\mathbb{N}}^{({{\bf s}})}$, we define $$\operatorname{\mathcal{P}}(e,F) := \{ \alpha : F \to {\mathbb{N}}^{({{\bf s}})}_*\ |\ |\alpha| = e\} \subset \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,F) .$$ It is clear that $\operatorname{\mathcal{P}}(e,F) = \emptyset$ whenever $\# F > |e|$, $\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,\emptyset)=\emptyset$ if $e\neq 0$, $\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(0,\Lambda)$ consists of only the constant map $0$ and that $\operatorname{\mathcal{P}}(0,\emptyset)=\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(0,\emptyset)$ consists of only the inclusion $\emptyset \hookrightarrow {\mathbb{N}}^{({{\bf s}})}_*$. If $\# F =1$ and $e\neq 0$, then $\operatorname{\mathcal{P}}(e,F)$ also consists of only one map: the constant map with value $e$. The natural map $\displaystyle \coprod_{\scriptstyle F\in \operatorname{\mathfrak{P}}_f(\Lambda)} \operatorname{\mathcal{P}}(e, F) \longrightarrow \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,\Lambda)$ is obviously a bijection. If $r\geq 0$ is an integer, we will denote $\operatorname{\mathcal{P}}(e,r) := \operatorname{\mathcal{P}}(e,[r])$. 5. Assume that $\Lambda$ is a finite set, ${{\bf t}}$ is an arbitrary set and $\pi:\Lambda \to {{\bf t}}$ is map. Then, there is a natural bijection $$\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,\Lambda) \leftrightarrow \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in {{\bf t}}} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t)) = \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in \operatorname{\rm supp}e^\bullet} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t)).$$ Namely, to each $\alpha^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,\Lambda)$ we associate $e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})$ defined by $e^t = \sum_{\pi(l)=t} \alpha^l$, and $\{ \alpha^{t\bullet} \}_{t\in {{\bf t}}} \in \prod_{t\in {{\bf t}}} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t))$ with $\alpha^{t\bullet} =\alpha^\bullet|_{\pi^{-1}(t)}$. Let us notice that if for some $t_0\in{{\bf t}}$ one has $\pi^{-1}(t_0)=\emptyset$ and $e^{t_0}\neq 0$, then $\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^{t_0},\pi^{-1}(t_0))=\emptyset$ and so $\prod_{t\in {{\bf t}}} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t)) =\emptyset$. Hence $$\begin{aligned} &\displaystyle \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in {{\bf t}}} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\Lambda_t) = \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}_{\pi}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in {{\bf t}}} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t)) = &\\ &\displaystyle \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}_{\pi}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in \operatorname{\rm supp}e^\bullet} \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e^t,\pi^{-1}(t)),\end{aligned}$$ where $\operatorname{\mathcal{P}}_{\pi}^{{\scriptstyle \circ}}(e,{{\bf t}})$ is the subset of $\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})$ whose elements are the $e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})$ such that $ e^t = 0 $ whenever $\pi^{-1}(t)=\emptyset$ and $|e^t| \geq \# \pi^{-1}(t)$ otherwise. The preceding bijection induces a bijection $$\label{eq:bijec-par} \operatorname{\mathcal{P}}(e,\Lambda) \longleftrightarrow \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}_{\pi}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in {{\bf t}}} \operatorname{\mathcal{P}}(e^t,\pi^{-1}(t))= \coprod_{\scriptstyle e^\bullet \in \operatorname{\mathcal{P}}_{\pi}^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in \operatorname{\rm supp}e^\bullet} \operatorname{\mathcal{P}}(e^t,\pi^{-1}(t)).$$ 6. If $\alpha\in {\mathbb{N}}^{({{\bf t}})}$, we denote $$[\alpha] := \{ (t,r) \in {{\bf t}}\times {\mathbb{N}}_*\ |\ 1\leq r\leq \alpha_t \}$$ endowed with the projection $\pi:[\alpha] \to {{\bf t}}$. It is clear that $|\alpha|= \# [\alpha]$, and so $\alpha=0$ $\Longleftrightarrow$ $[\alpha] =\emptyset$. We denote $\operatorname{\mathcal{P}}(e,\alpha):= \operatorname{\mathcal{P}}(e,[\alpha])$. Elements in $ \operatorname{\mathcal{P}}(e,\alpha)$ will be written as $$\mathcal{b}^{\bullet\bullet}: (t,r)\in [\alpha] \longmapsto \mathcal{b}^{tr}\in {\mathbb{N}}^{({{\bf s}})},\quad \text{with}\ \sum_{\scriptscriptstyle (t,r)\in [\alpha]} \mathcal{b}^{tr} =e.$$ For each $\mathcal{b}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\alpha)$ and each $t\in{{\bf t}}$, we denote $$\mathcal{b}^{t\bullet}: r\in [\alpha_t] \longmapsto \mathcal{b}^{tr} \in {\mathbb{N}}^{({{\bf s}})},\quad [\mathcal{b}]^\bullet: t\in {{\bf t}}\longmapsto [\mathcal{b}]^t := |\mathcal{b}^{t\bullet}|= \sum_{\scriptscriptstyle r=1}^{\scriptscriptstyle \alpha_t} \mathcal{b}^{tr} \in {\mathbb{N}}^{({{\bf s}})}.$$ Notice that $|[\mathcal{b}]^t| \geq \alpha_t$, $[\mathcal{b}]^t=0$ whenever $\alpha_t=0$ and $\left| [\mathcal{b}]^\bullet \right| = e$. The bijection (\[eq:bijec-par\]) gives rise to a bijection $$\label{eq:bijec-par-alpha} \operatorname{\mathcal{P}}(e,\alpha) \longleftrightarrow \coprod_{\scriptstyle e^\bullet \in {\operatorname{\mathcal{P}}}_\alpha^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in {{\bf t}}} \operatorname{\mathcal{P}}(e^t,\alpha_t) = \coprod_{\scriptstyle e^\bullet \in {\operatorname{\mathcal{P}}}_\alpha^{{\scriptstyle \circ}}(e,{{\bf t}})} \prod_{\scriptstyle t\in \operatorname{\rm supp}e^\bullet} \operatorname{\mathcal{P}}(e^t,\alpha_t),$$ where $ {\operatorname{\mathcal{P}}}_\alpha^{{\scriptstyle \circ}}(e,{{\bf t}})$ is the subset of $\operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})$ whose elements are the $e^\bullet \in \operatorname{\mathcal{P}}^{{\scriptstyle \circ}}(e,{{\bf t}})$ such that $ e^t = 0 $ if $\alpha_t=0$ and $|e^t| \geq \alpha_t$ otherwise. [[**.  **]{}]{}\[nume:explicit-substitution\] Let ${{\bf t}}$, ${{\bf u}}$ be sets and $\Delta\subset {\mathbb{N}}^{({{\bf u}})}$ a non-empty co-ideal. Let $\varphi_0: A[{{\bf t}}] \xrightarrow{} A[[{{\bf u}}]]_\Delta $ be an $A$-algebra map given by: $$\varphi_0(t)=: c^t = \sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta| }} c^t_\beta {{\bf u}}^\beta \in {\mathfrak{n}}^A_0({{\bf u}})/\Delta_A \subset A[[{{\bf u}}]]_\Delta,\ \ t\in {{\bf t}}.$$ Let us write down the expression of the image $\varphi_0(a)$ of any $a\in A[{{\bf t}}]$ in terms of the coefficients of $a$ and the $c^t, t\in {{\bf t}}$. First, for each $r\geq 0$ and for each $t\in {{\bf t}}$ we have $$\varphi_0(t^r)= (c^t)^r = \dots = \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq r }} \left( \sum_{\scriptscriptstyle \beta^\bullet \in \operatorname{\mathcal{P}}(e,r)} \prod_{\scriptstyle k=1}^{\scriptstyle r} c^t_{\beta^k}\right) {{\bf u}}^e.$$ Observe that $$\label{eq:convention} \sum_{\scriptscriptstyle \beta^\bullet \in \operatorname{\mathcal{P}}(e,r)} \prod_{\scriptscriptstyle k=1}^r c^t_{\beta^k} = \left\{ \begin{array}{ll} 1 & \text{if $|e|=r=0$}\\ 0 & \text{if $|e|>r=0$.} \end{array} \right.$$ So, for each $\alpha \in {\mathbb{N}}^{({{\bf t}})}$ we have $$\begin{aligned} &\displaystyle \varphi_0({{\bf t}}^\alpha) = \prod_{\scriptscriptstyle t\in {{\bf t}}} (c^t)^{\alpha_t} = \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} (c^t)^{\alpha_t} = \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \left( \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq \alpha_t }} \left( \sum_{\scriptscriptstyle \beta^\bullet \in \operatorname{\mathcal{P}}(e,\alpha_t)} \prod_{\scriptscriptstyle k=1}^{\scriptscriptstyle \alpha_t} c^t_{\beta^k}\right) {{\bf u}}^e \right)= &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle e^t\in \Delta, t\in \operatorname{\rm supp}\alpha\\ \scriptscriptstyle |e^t|\geq \alpha_t }} \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \left(\left( \sum_{\scriptscriptstyle \beta^\bullet \in \operatorname{\mathcal{P}}(e^t,\alpha_t)} \prod_{\scriptscriptstyle k=1}^{\scriptscriptstyle \alpha_t} c^t_{\beta^k}\right) {{\bf u}}^{e^t}\right) =\end{aligned}$$ $$\begin{aligned} &\displaystyle \sum_{\substack{\scriptscriptstyle e^t\in \Delta, t\in \operatorname{\rm supp}\alpha\\ \scriptscriptstyle |e^t|\geq \alpha_t }} \left( \sum_{\substack{\scriptscriptstyle \beta^{t \bullet} \in \operatorname{\mathcal{P}}(e^t,\alpha_t)\\ \scriptscriptstyle t\in \operatorname{\rm supp}\alpha}} \left( \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \prod_{\scriptscriptstyle k=1}^{\scriptscriptstyle \alpha_t} c^t_{\beta^{tk}}\right) \right) \left( \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} {{\bf u}}^{e^t}\right) = &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} \left( \sum_{\substack{\scriptscriptstyle e^t\in \Delta, t\in \operatorname{\rm supp}\alpha\\ \scriptscriptstyle |e^t|\geq \alpha_t\\ \scriptscriptstyle |e^\bullet|=e }} \left( \sum_{\substack{\scriptscriptstyle \beta^{t \bullet} \in \operatorname{\mathcal{P}}(e^t,\alpha_t)\\ \scriptscriptstyle t\in \operatorname{\rm supp}\alpha}} \left( \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \prod_{\scriptscriptstyle k=1}^{\scriptscriptstyle \alpha_t} c^t_{\beta^{tk}}\right) \right) \right) {{\bf u}}^e = &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} \left( \sum_{\scriptscriptstyle e^\bullet \in \operatorname{\mathcal{P}}_\alpha^{{\scriptstyle \circ}}(e,{{\bf t}})} \left( \sum_{\substack{\scriptscriptstyle \beta^{t \bullet} \in \operatorname{\mathcal{P}}(e^t,\alpha_t)\\ \scriptscriptstyle t\in \operatorname{\rm supp}\alpha}} \left( \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \prod_{\scriptscriptstyle k=1}^{\scriptscriptstyle \alpha_t} c^t_{\beta^{tk}}\right) \right) \right) {{\bf u}}^e= \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} {\bf C}_e(\varphi_0,\alpha) {{\bf u}}^e,\end{aligned}$$ with (see (\[eq:bijec-par-alpha\])): $$\label{eq:explicit-C_e(varphi,alpha)} {\bf C}_e(\varphi_0,\alpha)= \sum_{\scriptscriptstyle \beta^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\alpha)} C_{\beta^{\bullet\bullet}},\quad C_{\beta^{\bullet\bullet}} = \prod_{\scriptstyle t\in \operatorname{\rm supp}\alpha} \prod_{\scriptstyle r=1}^{\scriptstyle \alpha_t} c^t_{\beta^{tr}},\quad \text{for\ }\ |\alpha| \leq |e|. $$ We have ${\bf C}_0(\varphi_0,0) = 1$ and ${\bf C}_e(\varphi_0,0)=0$ for $e\neq 0$. For a fixed $e\in {\mathbb{N}}^{({{\bf u}})}$ the support of any $\alpha\in {\mathbb{N}}^{({{\bf t}})}$ such that $|\alpha|\leq |e|$ and ${\bf C}_e(\varphi_0,\alpha)\neq 0$ is contained in the set $$\bigcup_{\substack{\scriptscriptstyle \beta \in \Delta\\ \scriptscriptstyle \beta \leq e}} \{t\in {{\bf t}}\ |\ c^t_\beta \neq 0\}$$ and so the set of such $\alpha$’s is finite provided that property (\[eq:cond-fini-c\]) holds. We conclude that $$\label{eq:exp-substi} \varphi_0\left( \sum_{\scriptscriptstyle\alpha\in{\mathbb{N}}^{({{\bf t}})}} a_\alpha {{\bf t}}^\alpha \right) = \sum_{\scriptscriptstyle \alpha\in {\mathbb{N}}^{({{\bf t}})}} a_\alpha c^\alpha= \sum_{\scriptscriptstyle e\in \Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptscriptstyle |\alpha|\leq |e| }} {\bf C}_e(\varphi_0,\alpha) a_\alpha \right) {{\bf u}}^e.$$ Observe that for each non-zero $\alpha\in{\mathbb{N}}^{({{\bf t}})}$ we have: $$\label{eq:condition-supp-substi} \operatorname{\rm supp}(\varphi_0({{\bf t}}^\alpha)) = \operatorname{\rm supp}\left( \prod_{\scriptstyle t\in\operatorname{\rm supp}\alpha} \left(c^t\right)^{\alpha_t}\right) \subset \sum_{\scriptscriptstyle t\in \operatorname{\rm supp}(\alpha)} \alpha_t \cdot \operatorname{\rm supp}(c^t).$$ Let us notice that if we assign the weight $|\beta|$ to $c^t_{\beta}$, then ${\bf C}_e(\varphi_0,\alpha)$ is a quasi-homogeneous polynomial in the variables $c^t_\beta$, $t\in\operatorname{\rm supp}\alpha$, $|\beta| \leq |e|$, of weight $|e|$. The proof of the following lemma is easy and it is left to the reader. \[lemma:behavior\_C\_alpha\] For each $e\in\Delta$ and for each $\alpha\in {\mathbb{N}}^{({{\bf t}})}$ with $0<|\alpha|\leq |e|$, the following properties hold: 1. If $|\alpha|=1$, then ${\bf C}_e(\varphi_0,\alpha) = c_e^s$, where $\operatorname{\rm supp}\alpha = \{s\}$, i.e. $\alpha = {{\bf t}}^s$ (${{\bf t}}^s_t = \delta_{st}$). 2. If $|\alpha|=|e|$, then $${\bf C}_e(\varphi_0,\alpha) = \sum_{\substack{\scriptscriptstyle e^t\in \Delta, t\in \operatorname{\rm supp}\alpha\\ \scriptscriptstyle |e^t|= \alpha_t, |e^\bullet|=e }} \left( \prod_{\scriptscriptstyle t\in \operatorname{\rm supp}\alpha} \prod_{\scriptscriptstyle v\in \operatorname{\rm supp}e^t } \left(c^t_{{{\bf u}}^v}\right)^{e^t_v} \right).$$ Let ${{\bf t}}, {{\bf u}}$ be sets and $\Delta\subset {\mathbb{N}}^{({{\bf u}})}$ a non-empty co-ideal. For each family $$c=\left\{c^t = \sum_{\substack{\scriptscriptstyle \beta\in \Delta\\ \scriptscriptstyle \beta\neq 0}} c^t_\beta {{\bf u}}^\beta \in {\mathfrak{n}}^A_0({{\bf u}})/\Delta_A \subset A[[{{\bf u}}]]_\Delta,\ t\in {{\bf t}}\right\}$$ (we are assuming that $c^t_0 = 0$) satisfying the following property $$\label{eq:cond-fini-c} \#\{t\in {{\bf t}}\ |\ c^t_\beta \neq 0\} < \infty\quad\quad \text{for all\ }\ \beta \in \Delta,$$ there is a unique continuous $A$-algebra map $\varphi: A[[{{\bf t}}]] \xrightarrow{} A[[{{\bf u}}]]_\Delta$ such that $\varphi(t) = c^t$ for all $t\in{{\bf t}}$. Moreover, if $\nabla\subset {\mathbb{N}}^{({{\bf t}})}$ is a non-empty co-ideal such that $\varphi(\nabla_A) =0$, then $\varphi$ induces a unique continuous $A$-algebra map $A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta$ sending (the class of) each $t\in{{\bf t}}$ to $c^t$. Let us consider the unique $A$-algebra map $\varphi_0: A[{{\bf t}}] \xrightarrow{} A[[{{\bf u}}]]_\Delta$ defined by $\varphi_0(t) = c^t$ for all $t\in{{\bf t}}$. From (\[eq:explicit-C\_e(varphi,alpha)\]) and (\[eq:exp-substi\]) in \[nume:explicit-substitution\], we know that $$\varphi_0\left( \sum_{\substack{\scriptscriptstyle\alpha\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptscriptstyle \text{finite}}} a_\alpha {{\bf t}}^\alpha \right) = \sum_{\scriptscriptstyle e\in \Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptscriptstyle |\alpha|\leq |e| }} {\bf C}_e(\varphi_0,\alpha) a_\alpha \right) {{\bf u}}^e.$$ Since for a fixed $e\in {\mathbb{N}}^{({{\bf u}})}$ the support of the $\alpha\in {\mathbb{N}}^{({{\bf t}})}$ such that $|\alpha|\leq |e|$ and ${\bf C}_e(\varphi_0,\alpha)\neq 0$ is contained in the finite set $$\bigcup_{\substack{\scriptscriptstyle \beta \in \Delta\\ \scriptscriptstyle \beta \leq e}} \{t\in {{\bf t}}\ |\ c^t_\beta \neq 0\},$$ the set of such $\alpha$’s is always finite and we deduce that $\varphi_0$ is continuous, and so there is a unique continuous extension $\varphi: A[[{{\bf t}}]] \xrightarrow{} A[[{{\bf u}}]]_\Delta$ such that $\varphi(t) = c^t$ for all $t\in{{\bf t}}$. The last part is clear. Let us notice that, after (\[eq:condition-supp-substi\]), to get the equality $\varphi(\nabla_A) =0$ in the above proposition it is enough to have for each $\alpha \in \nabla^c$ (actually, it will be enough to consider the $\alpha \in \nabla^c$ minimal with respect to the ordering $\leq$ in ${\mathbb{N}}^{({{\bf t}})}$): $$\sum_{\scriptscriptstyle t\in \operatorname{\rm supp}(\alpha)} \alpha_t \cdot \operatorname{\rm supp}(c^t) \subset \Delta^c.$$ \[def:substitution\_maps\] Let $\nabla\subset {\mathbb{N}}^{({{\bf t}})}, \Delta\subset {\mathbb{N}}^{({{\bf u}})}$ be non-empty co-ideals. An $A$-algebra map $\varphi:A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta$ will be called a [*substitution map*]{} if the following properties hold: 1. $\varphi$ is continuous. 2. $\varphi(t)\in {\mathfrak{n}}^A_0({{\bf u}})/\Delta_A$ for all $t\in {{\bf t}}$. 3. The family $c=\{\varphi(t), t\in {{\bf t}}\}$ satisfies property (\[eq:cond-fini-c\]). The set of substitution maps $A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta$ will be denoted by $\operatorname{\mathcal{S}}_A({{\bf t}},{{\bf u}};\nabla,\Delta)$. The [*trivial*]{} substitution map $A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta $ is the one sending any $t\in {{\bf t}}$ to $0$. It will be denoted by $\mathbf{0}$. In the above definition, a such $\varphi$ is uniquely determined by the family $c=\{\varphi(t), t\in {{\bf t}}\}$, and will be called the [*substitution map associated*]{} with $c$. Namely, the family $c$ can be lifted to $A[[{{\bf u}}]]$ by means of the natural $A$-linear scission $A[[{{\bf u}}]]_\Delta \hookrightarrow A[[{{\bf u}}]]$ and we may consider the unique continuous $A$-algebra map $\psi: A[[{{\bf t}}]] \to A[[{{\bf u}}]]$ such that $\psi(s) = c^s$ for all $s\in{{\bf s}}$. Since $\varphi$ is continuous, we have a commutative diagram $$\begin{CD} A[[{{\bf t}}]] @>{\psi}>>A[[{{\bf u}}]] \\ @V{\text{proj.}}VV @VV{\text{proj.}}V\\ A[[{{\bf t}}]]_\nabla @>{\varphi}>>A[[{{\bf u}}]]_\Delta, \end{CD}$$ and so $\psi(\nabla_A) \subset \Delta_A$. Then, we may indentify $$\operatorname{\mathcal{S}}_A({{\bf t}},{{\bf u}};\nabla,\Delta) \equiv \left\{\overline{\psi}\in \operatorname{\mathcal{S}}_A({{\bf t}},{{\bf u}};{\mathbb{N}}^{({{\bf t}})},\Delta)\ |\ \overline{\psi}(\nabla_A)=0 \right\}.$$ For $\alpha \in \nabla$ and $e\in \Delta$ with $|\alpha|\leq |e|$ we will write ${\bf C}_e(\varphi,\alpha) := {\bf C}_e(\varphi_0,\alpha)$, where $\varphi_0: A[{{\bf t}}] \to A[[{{\bf u}}]]_\Delta$ is the $A$-algebra map given by $\varphi_0(t) = \varphi(t)$ for all $t\in{{\bf t}}$ (see (\[eq:explicit-C\_e(varphi,alpha)\]) in \[nume:explicit-substitution\]). For any family of integers $\nu =\{\nu_t\geq 1, t\in{{\bf t}}\}$, we will denote $ [\nu]: A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf t}}]]_{\nu\nabla}$ the substitution map determined by $[\nu](t) = t^{\nu_t}$ for all $t\in{{\bf t}}$, where $$\nu\nabla := \{\gamma \in {\mathbb{N}}^{({{\bf t}})}\ |\ \exists \alpha \in\nabla, \gamma\leq \nu \alpha \}.$$ We obviously have $[\nu \nu'] = [\nu] {{\scriptstyle \,\circ\,}}[\nu']$. \[lemma:compos\_subst\_maps\] The composition of two substitution maps $A[[{{\bf t}}]]_\nabla \stackrel{\varphi}{\to} A[[{{\bf u}}]]_\Delta \stackrel{\psi}{\to} A[[{{\bf s}}]]_\Omega$ is a substitution map and we have $${\bf C}_f(\psi {{\scriptstyle \,\circ\,}}\varphi,\alpha) = \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |f|\geq |e|\geq |\alpha| }} {\bf C}_e(\varphi,\alpha) {\bf C}_f(\psi,e),\quad \forall f\in\Omega, \forall \alpha\in\nabla, |\alpha|\leq |f|.$$ Moreover, if one of the substitution maps is trivial, then the composition is trivial too. Properties (1) and (2) in Definition \[def:substitution\_maps\] are clear. Let us see property (3). For each $t\in{{\bf t}}$ let us write: $$\varphi(t) =: c^t = \sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta| }} c^t_\beta {{\bf u}}^\beta \in {\mathfrak{n}}^A_0({{\bf u}})/\Delta_A \subset A[[{{\bf u}}]]_\Delta,$$ and so $$\begin{aligned} &\displaystyle (\psi {{\scriptstyle \,\circ\,}}\varphi)(t) = \psi\left(\sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta| }} c^t_\beta {{\bf u}}^\beta \right) = \sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta| }} c^t_\beta \left( \sum_{\substack{\scriptscriptstyle f\in \Omega\\ \scriptscriptstyle |f|\geq |\beta| }} {\bf C}_f(\psi,\beta) {{\bf s}}^f \right)= \sum_{\substack{\scriptscriptstyle f\in \Omega\\ \scriptscriptstyle |f|>0 }} d^t_f {{\bf s}}^f\end{aligned}$$ with $$d^t_f= \sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta|\leq |f| }} c^t_\beta {\bf C}_f(\psi,\beta)$$ and for a fixed $f\in\Omega$ the set $$\{t\in{{\bf t}}\ |\ d^t_f\neq 0\} \subset \bigcup_{\substack{\scriptscriptstyle \beta\in\nabla, |\beta|\leq|f|\\ \scriptscriptstyle {\bf C}_f(\psi,\beta) \neq 0}} \{t\in{{\bf t}}\ |\ c^t_\beta\neq 0\}$$ is finite. On the other hand $$\begin{aligned} &\displaystyle (\psi{{\scriptstyle \,\circ\,}}\varphi)({{\bf t}}^\alpha) = \psi\left( \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} {\bf C}_e(\varphi,\alpha) {{\bf u}}^e \right) = \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} {\bf C}_e(\varphi,\alpha) \left( \sum_{\substack{\scriptscriptstyle f\in \Omega\\ \scriptscriptstyle |f|\geq |e| }} {\bf C}_f(\psi,e) {{\bf s}}^f \right)= & \\ &\displaystyle \sum_{\substack{\scriptscriptstyle f\in \Omega\\ \scriptscriptstyle |f|\geq |\alpha| }} \left( \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |f|\geq |e|\geq |\alpha| }} {\bf C}_e(\varphi,\alpha) {\bf C}_f(\psi,e) \right) {{\bf u}}^f\end{aligned}$$ and so $${\bf C}_f(\psi{{\scriptstyle \,\circ\,}}\varphi,\alpha) = \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |f|\geq |e|\geq |\alpha| }} {\bf C}_e(\varphi,\alpha) {\bf C}_f(\psi,e),\quad \forall f\in\Omega, \forall \alpha\in\nabla, |\alpha|\leq |f|.$$ If $B$ is a commutative $A$-algebra, then any subtitution map $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ induces a natural substitution map $\varphi_B:B[[{{\bf s}}]]_\nabla \to B[[{{\bf t}}]]_\Delta$ making the following diagram commutative $$\begin{CD} B\widehat{\otimes}_A A[[{{\bf s}}]]_\nabla @>{{{\rm Id}}\widehat{\otimes} \varphi}>> B\widehat{\otimes}_A A[[{{\bf t}}]]_\Delta\\ @V{\text{nat.}}V{\simeq}V @V{\simeq}V{\text{nat.}}V\\ B[[{{\bf s}}]]_\nabla @>{\varphi_B}>> B[[{{\bf t}}]]_\Delta. \end{CD}$$ [[**.  **]{}]{}\[nume:operaciones-con-substitutions\] For any substitution map $\varphi: A[[{{\bf s}}]]_\nabla \xrightarrow{} A[[{{\bf t}}]]_\Delta$ and for any integer $n\geq 0$ we have $\varphi(\nabla^n_A/\nabla_A) \subset \Delta^n_A/\Delta_A$ and so there are induced substitution maps $\tau_{n}(\varphi):A[[{{\bf s}}]]_{\nabla^n} \to A[[{{\bf t}}]]_{\Delta^n}$ making commutative the following diagram $$\begin{CD} A[[{{\bf s}}]]_\nabla @>{\varphi}>> A[[{{\bf t}}]]_\Delta\\ @V{\text{nat.}}VV @VV{\text{nat.}}V\\ A[[{{\bf s}}]]_{\nabla^n} @>{\tau_{n}(\varphi)}>> A[[{{\bf t}}]]_{\Delta^n}. \end{CD}$$ Moreover, if $\varphi$ is the substitution map associated with a family $c=\{c^s, s\in {{\bf s}}\}$, $$c^s = \sum_{\scriptscriptstyle \beta\in \Delta} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}^A_0({{\bf t}})/\Delta_A\subset A[[{{\bf t}}]]_\Delta,$$ then $\tau_{n}(\varphi)$ is the substitution map associated with the family $\tau_{n}(c)=\{\tau_{n}(c)^s, s\in {{\bf s}}\}$, with $$\tau_{n}(c)^s := \sum_{\substack{\scriptscriptstyle \beta\in \Delta\\ \scriptscriptstyle |\beta|\leq n}} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}^A_0({{\bf t}})/\Delta^n_A\subset A[[{{\bf t}}]]_{\Delta^n}.$$ So, we have truncations $\tau_n: \operatorname{\mathcal{S}}_A({{\bf s}},{{\bf t}};\nabla,\Delta) \longrightarrow \operatorname{\mathcal{S}}_A({{\bf s}},{{\bf t}};\nabla^n,\Delta^n)$, for $n\geq 0$. We may also add two substitution maps $\varphi,\varphi':A[[{{\bf s}}]] \xrightarrow{} A[[{{\bf t}}]]_\Delta$ to obtain a new substitution map $\varphi+\varphi':A[[{{\bf s}}]] \xrightarrow{} A[[{{\bf t}}]]_\Delta$ determined by[^2]: $$(\varphi+\varphi')(s) = \varphi(s)+\varphi'(s),\quad \text{for all\ }\ s\in {{\bf s}}.$$ It is clear that $\operatorname{\mathcal{S}}_A({{\bf s}},{{\bf t}};{\mathbb{N}}^{({{\bf s}})},\Delta)$ becomes an abelian group with the addition, the zero element being the trivial substitution map $\mathbf{0}$. If $\psi: A[[{{\bf t}}]]_\Delta \xrightarrow{} A[[{{\bf u}}]]_\Omega$ is another substitution map, we clearly have $$\psi {{\scriptstyle \,\circ\,}}(\varphi+\varphi') = \psi {{\scriptstyle \,\circ\,}}\varphi + \psi {{\scriptstyle \,\circ\,}}\varphi'.$$ However, if $\psi: A[[{{\bf u}}]] \xrightarrow{} A[[{{\bf s}}]]$ is a substitution map, we have in general $$(\varphi+\varphi'){{\scriptstyle \,\circ\,}}\psi \neq \varphi{{\scriptstyle \,\circ\,}}\psi +\varphi'{{\scriptstyle \,\circ\,}}\psi.$$ We say that a substitution map $\varphi: A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta $ has [*constant coefficients*]{} if $c^t_\beta \in k$ for all $t\in {{\bf t}}$ and all $\beta\in\Delta$, where $$\varphi(t)=c^t = \sum_{\substack{\scriptscriptstyle \beta\in\Delta\\ \scriptscriptstyle 0<|\beta| }} c^t_\beta {{\bf u}}^\beta \in {\mathfrak{n}}^A_0({{\bf u}})/\Delta_A \subset A[[{{\bf u}}]]_\Delta.$$ This is equivalent to saying that ${\bf C}_e(\varphi,\alpha)\in k$ for all $e\in\Delta$ and for all $\alpha\in \nabla$ with $0<|\alpha|\leq |e|$. Substitution maps which constant coefficients are induced by substitution maps $k[[{{\bf t}}]]_\nabla \xrightarrow{} k[[{{\bf u}}]]_\Delta$. We say that a substitution map $\varphi: A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_\Delta $ is [*combinatorial*]{} if $\varphi(t) \in {{\bf u}}$ for all $t\in{{\bf t}}$. A combinatorial substitution map has constant coefficients and is determined by (and determines) a map ${{\bf t}}\to {{\bf u}}$, necessarily with finite fibers. If $\iota : {{\bf t}}\to {{\bf u}}$ is such a map, we will also denote by $\iota: A[[{{\bf t}}]]_\nabla \xrightarrow{} A[[{{\bf u}}]]_{\iota_*(\nabla)} $ the corresponding substitution map, with $$\iota_*(\nabla):= \{\beta \in {\mathbb{N}}^{({{\bf u}})}\ |\ \beta {{\scriptstyle \,\circ\,}}\iota \in \nabla\}.$$ [[**.  **]{}]{}\[nume:cont-maps-power-series-rings\] Let $\varphi: A[[{{\bf s}}]]_\nabla \xrightarrow{} A[[{{\bf t}}]]_\Delta $ be a continuous $A$-linear map. It is determined by the family $ K =\{ K_{e,\alpha}, e\in \Delta, \alpha \in\nabla \} \subset A$, with $\displaystyle \varphi( {{\bf s}}^\alpha ) = \sum_{ \scriptscriptstyle e\in \Delta} K_{e,\alpha} {{\bf t}}^e$. We will assume that - $\varphi$ is compatible with the order filtration, i.e. $\varphi (\nabla^n_A/\nabla_A) \subset \Delta^n_A/\Delta_A$ for all $n\geq 0$. - $\varphi$ is compatible with the natural augmentations $A[[{{\bf s}}]]_\nabla \to A$ and $A[[{{\bf t}}]]_\Delta \to A$. These properties are equivalent to the fact that $K_{e,\alpha}=0$ whenever $|\alpha|>|e|$ and $K_{0,0}=1$. Let $ K =\{ K_{e,\alpha}$, $e\in \Delta, \alpha \in\nabla$, $|\alpha|\leq |e| \}$ be a family of elements of $A$ with $$\# \{\alpha\in\nabla\ |\ |\alpha|\leq |e|, K_{e,\alpha}\neq 0\} < +\infty,\ \ \forall e\in \Delta,$$ and $K_{0,0} = 1$, and let $\varphi: A[[{{\bf s}}]]_\nabla \xrightarrow{} A[[{{\bf t}}]]_\Delta $ be the $A$-linear map given by $$\varphi\left( \sum_{\scriptscriptstyle\alpha\in\nabla} a_\alpha {{\bf s}}^\alpha \right) = \sum_{\scriptscriptstyle e\in \Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in\nabla\\ \scriptscriptstyle |\alpha|\leq |e| }} K_{e,\alpha} a_\alpha \right) {{\bf t}}^e.$$ It is clearly continuous and since $\displaystyle \varphi( {{\bf s}}^\alpha ) = \sum_{\substack{ \scriptscriptstyle e\in\Delta\\ \scriptscriptstyle |\alpha|\leq |e|}} K_{e,\alpha} {{\bf t}}^e, $ it determines the family $K$. \[prop:identidad-Cs\] With the above notations, the following properties are equivalent: 1. $\varphi$ is a substitution map. 2. For each $\mu,\nu \in\nabla$ and for each $e\in \Delta$ with $|\mu+\nu|\leq |e|$, the following equality holds: $$K_{e,\mu+\nu} = \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|, |\nu|\leq |\gamma| }} K_{\beta,\mu} K_{\gamma,\nu}.$$ Moreover, if the above equality holds, then $K_{e,0}=0$ whenever $|e|>0$ and $\varphi$ is the substitution map determined by $$\varphi(u) = \sum_{\substack{\scriptscriptstyle e \in \Delta \\ \scriptscriptstyle 0 <|e| }} K_{e,{{\bf s}}^u} {{\bf t}}^e,\quad u\in {{\bf s}}.$$ \(a) $\Rightarrow$ (b) If $\varphi$ is a substitution map, there is a family $$c^s = \sum_{\scriptscriptstyle \beta\in\Delta} c^s_\beta {{\bf t}}^\beta \in A[[{{\bf t}}]]_\Delta,\ \ s\in {{\bf s}},$$ such that $\varphi(s) = c^s$. So, from (\[eq:exp-substi\]), we deduce $$K_{e,\alpha}= {\bf C}_e(\varphi,\alpha)= \sum_{\scriptstyle \mathcal{f}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\alpha)} C_{\mathcal{f}^{\bullet\bullet}} \quad \text{for\ }\ |\alpha| \leq |e|,$$ with $ \displaystyle C_{\mathcal{f}^{\bullet\bullet}} = \prod_{\scriptstyle s\in \operatorname{\rm supp}\alpha} \prod_{\scriptstyle r=1}^{\scriptstyle \alpha_s} c^s_{\mathcal{f}^{sr}}. $ For each ordered pair $(r,s)$ of non-negative integers there are natural injective maps $$i \in [r] \mapsto i \in [r+s],\quad i \in [s] \mapsto r+i \in [r+s]$$ inducing a natural bijection $[r]\sqcup [s] \longleftrightarrow [r+s]$. Consequently, for $(\mu,\nu) \in{\mathbb{N}}^{({{\bf s}})}\times {\mathbb{N}}^{({{\bf s}})}$ there are natural injective maps $ [\mu] \hookrightarrow [\mu+\nu] \hookleftarrow [\nu] $ inducing a natural bijection $[\mu]\sqcup [\nu] \longleftrightarrow [\mu+\nu]$. So, for each $e\in {\mathbb{N}}^{({{\bf t}})}$ and each $\mathcal{f}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\mu + \nu)$, we can consider the restrictions $\mathcal{g}^{\bullet\bullet} = \mathcal{f}^{\bullet\bullet}|_{[\mu]} \in \operatorname{\mathcal{P}}(\beta,\mu)$, $\mathcal{h}^{\bullet\bullet} = \mathcal{f}^{\bullet\bullet}|_{[\nu]} \in \operatorname{\mathcal{P}}(\gamma,\nu)$, with $\beta = | \mathcal{g}^{\bullet\bullet} |$ and $\gamma = | \mathcal{h}^{\bullet\bullet} |$, $\beta+\gamma=e$. The correspondence $ \mathcal{f}^{\bullet\bullet } \longmapsto (\beta,\gamma,\mathcal{g}^{\bullet\bullet}, \mathcal{h}^{\bullet\bullet})$ establishes a bijection between $\operatorname{\mathcal{P}}(e,\mu + \nu) $ and the set of $(\beta,\gamma,\mathcal{g}^{\bullet\bullet}, \mathcal{h}^{\bullet\bullet})$ with $\beta,\gamma\in {\mathbb{N}}^{({{\bf t}})}$, $\mathcal{g}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\beta,\mu)$, $\mathcal{h}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\gamma,\nu)$ and $|\beta|\geq |\mu|, |\gamma|\geq |\nu|,\beta + \gamma =e$. $$\{ (\beta,\gamma,\mathcal{g}^{\bullet\bullet}, \mathcal{h}^{\bullet\bullet})\ |\ \beta,\gamma\in {\mathbb{N}}^{({{\bf t}})}, \mathcal{g}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\beta,\mu), \mathcal{h}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\gamma,\nu), |\beta|\geq |\mu|, |\gamma|\geq |\nu|,\beta + \gamma =e \},$$ Moreover, under this bijection we have $C_{\mathcal{f}^{\bullet\bullet}} = C_{\mathcal{g}^{\bullet\bullet}} C_{\mathcal{h}^{\bullet\bullet}}$ and we deduce $$\begin{aligned} & \displaystyle K_{e,\mu+\nu}={\bf C}_e(\varphi,\mu + \nu)= \sum_{\scriptstyle \mathcal{f}^{\bullet\bullet} } C_{\mathcal{f}^{\bullet\bullet}} = \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|\\ \scriptscriptstyle |\nu|\leq |\gamma| }} \sum_{\substack{\scriptscriptstyle \mathcal{g}^{\bullet\bullet}, \mathcal{h}^{\bullet\bullet} }} C_{\mathcal{g}^{\bullet\bullet}} C_{\mathcal{h}^{\bullet\bullet}} =&\\ & \displaystyle \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|\\ \scriptscriptstyle |\nu|\leq |\gamma| }} \left( \sum_{\scriptscriptstyle \mathcal{g}^{\bullet\bullet} } C_{\mathcal{g}^{\bullet\bullet}} \right) \left( \sum_{\scriptscriptstyle \mathcal{h}^{\bullet\bullet}} C_{\mathcal{h}^{\bullet\bullet}} \right) = \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|\\ \scriptscriptstyle |\nu|\leq |\gamma| }} {\bf C}_\beta(\varphi,\mu ) {\bf C}_\gamma(\varphi,\nu ) = \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|\\ \scriptscriptstyle |\nu|\leq |\gamma| }} K_{\beta,\mu} K_{\gamma,\nu}.\end{aligned}$$ where $\mathcal{f}^{\bullet\bullet}\in \operatorname{\mathcal{P}}(e,\mu + \nu)$, $\mathcal{g}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\beta,\mu)$ and $\mathcal{h}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(\gamma,\nu)$. \(b) $\Rightarrow$ (a) First, one easily proves by induction on $|e|$ that $K_{e,0}=0$ whenever $|e|>0$, and so $\varphi(1) = \varphi({{\bf s}}^0) = K_{0,0}=1$. Let $ a = \sum_\alpha a_\alpha {{\bf s}}^\alpha, b = \sum_\alpha b_\alpha {{\bf s}}^\alpha $ be elements in $A[[t]]_\Delta$, and $c=ab=\sum_\alpha c_\alpha {{\bf s}}^\alpha$ with $c_\alpha= \sum_{\mu+\nu=\alpha} a_{\mu} b_{\nu}$. We have: $$\begin{aligned} & \displaystyle \varphi (ab) = \varphi (c) = \sum_{\scriptscriptstyle e\in\Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in\nabla \\ \scriptscriptstyle |\alpha|\leq |e|}} K_{e,\alpha} c_\alpha\right) {{\bf t}}^e = \sum_{\scriptscriptstyle e\in\Delta} \left( \sum_{\substack{ \scriptscriptstyle \mu,\nu\in\nabla\\ \scriptscriptstyle |\mu+\nu|\leq |e|}} K_{e,\mu+\nu} a_{\mu} b_{\nu}\right) {{\bf t}}^e =&\\ & \displaystyle \sum_e \left( \sum_{\scriptscriptstyle |\mu+\nu|\leq |e|} \sum_{\substack{\scriptscriptstyle \beta+\gamma=e\\ \scriptscriptstyle |\mu|\leq |\beta|, |\nu|\leq |\gamma| }} K_{\beta,\mu} K_{\gamma,\nu} a_{\mu} b_{\nu}\right) {{\bf t}}^e = \dots = \varphi (a) \varphi (b).\end{aligned}$$ We conclude that $\varphi$ is a (continuous) $A$-algebra map determined by the images $$\varphi(u) = \varphi\left({{\bf s}}^{{{\bf s}}^u}\right)=\sum_{\substack{\scriptscriptstyle e \in \Delta \\ \scriptscriptstyle 0 <|e| }} K_{e,{{\bf s}}^u} {{\bf t}}^e,\quad u\in {{\bf s}},$$ (remember that $\{{{\bf s}}^u\}_{u\in{{\bf s}}}$ is the canonical basis of ${\mathbb{N}}^{({{\bf s}})}$) and so it is a substitution map. \[defi:tensor-prod-of-substi\] The [*tensor product*]{} of two substitution maps $\varphi:A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$, $\psi:A[[{{\bf u}}]]_{\nabla'} \to A[[{{\bf v}}]]_{\Delta'}$ is the unique substitution map $$\varphi \otimes \psi: A[[{{\bf s}}\sqcup {{\bf u}}]]_{\nabla \times \nabla'} \longrightarrow A[[{{\bf t}}\sqcup {{\bf v}}]]_{\Delta \times \Delta'}$$ making commutative the following diagram $$\begin{CD} A[[{{\bf s}}]]_\nabla @>>> A[[{{\bf s}}\sqcup {{\bf u}}]]_{\nabla \times \nabla'} @<<< A[[{{\bf u}}]]_{\nabla'} \\ @VV{\varphi}V @VV{\varphi \otimes \psi}V @VV{\psi}V \\ A[[{{\bf t}}]]_\Delta @>>> A[[{{\bf t}}\sqcup {{\bf v}}]]_{\Delta \times \Delta'} @<<< A[[{{\bf v}}]]_{\Delta'}, \end{CD}$$ where the horizontal arrows are the combinatorial substitution maps induced by the inclusions ${{\bf s}}, {{\bf u}}\hookrightarrow {{\bf s}}\sqcup {{\bf u}}$, ${{\bf t}}, {{\bf v}}\hookrightarrow {{\bf t}}\sqcup {{\bf v}}$[^3]. For all $(\alpha,\beta)\in \nabla \times \nabla' \subset {\mathbb{N}}^{({{\bf s}})} \times {\mathbb{N}}^{({{\bf u}})} \equiv {\mathbb{N}}^{({{\bf s}}\sqcup {{\bf u}})}$ we have $$(\varphi \otimes \psi)({{\bf s}}^\alpha {{\bf u}}^\beta) = \varphi ({{\bf s}}^\alpha) \psi({{\bf u}}^\beta) = \cdots = \sum_{\substack{\scriptscriptstyle e\in \Delta, f\in \Delta'\\ \scriptscriptstyle |e|\geq |\alpha| \\ \scriptscriptstyle |f|\geq |\beta| }} {\bf C}_e(\varphi,\alpha) {\bf C}_f(\psi,\beta) {{\bf t}}^e {{\bf v}}^f$$ and so, for all $(e,f)\in \Delta \times \Delta'$ and all $(\alpha,\beta) \in \nabla \times \nabla'$ with $ |e|+|f|= |(e,f)| \geq |(\alpha,\beta)| = |\alpha|+|\beta|$ we have $${\bf C}_{(e,f)}(\varphi\otimes \psi,(\alpha,\beta)) = \left\{ \begin{array}{ll} {\bf C}_e(\varphi,\alpha) {\bf C}_f(\psi,\beta) & \text{if\ }\ |\alpha|\leq |e|\ \text{and\ }\ |\beta|\leq |f|, \\ 0 & \text{otherwise}. \end{array} \right.$$ The action of substitution maps {#sec:action-substi} =============================== In this section $k$ will be a commutative ring, $A$ a commutative $k$-algebra, $M$ an $(A;A)$-bimodule, ${{\bf s}}$ and ${{\bf t}}$ sets and $\nabla\subset{\mathbb{N}}^{({{\bf s}})}$, $\Delta\subset {\mathbb{N}}^{({{\bf t}})}$ non-empty co-ideals. Any $A$-linear continuous map $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ satisfying the assumptions in \[nume:cont-maps-power-series-rings\] induces $(A;A)$-linear maps $$\varphi_M := \varphi \widehat{\otimes} {{\rm Id}}_M : M[[{{\bf s}}]]_\nabla \equiv A[[{{\bf s}}]]_\Delta \widehat{\otimes}_A M \longrightarrow M[[{{\bf t}}]]_\Delta \equiv A[[{{\bf t}}]]_\Delta \widehat{\otimes}_A M$$ and $$\sideset{_M}{}\operatorname{\varphi}:= {{\rm Id}}_M \widehat{\otimes} \varphi : M[[{{\bf s}}]]_\nabla \equiv M \widehat{\otimes}_A A[[{{\bf s}}]]_\nabla \longrightarrow M[[{{\bf t}}]]_\Delta \equiv M \widehat{\otimes}_A A[[{{\bf t}}]]_\Delta.$$ If $\varphi$ is determined by the family $ K =\{ K_{e,\alpha}, e\in \nabla, \alpha \in\Delta, |\alpha|\leq |e| \} \subset A$, with $\displaystyle \varphi( {{\bf s}}^\alpha ) = \sum_{\substack{\scriptscriptstyle e\in \Delta\\ \scriptscriptstyle |e|\geq |\alpha| }} K_{e,\alpha} {{\bf t}}^e$, then $$\begin{aligned} &\displaystyle \varphi_M \left(\sum_{ \scriptscriptstyle \alpha\in\nabla} m_\alpha {{\bf s}}^\alpha \right) = \sum_{\scriptscriptstyle \alpha\in\nabla} \varphi( {{\bf s}}^\alpha ) m_\alpha = \sum_{ \scriptscriptstyle e\in \Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in\nabla\\ \scriptscriptstyle |\alpha|\leq |e| } } K_{e,\alpha} m_\alpha \right) {{\bf t}}^e,\quad m\in M[[{{\bf s}}]]_\nabla, &\\ &\displaystyle \sideset{_M}{}\operatorname{\varphi}\left(\sum_{ \scriptscriptstyle \alpha\in\nabla} m_\alpha {{\bf s}}^\alpha \right) = \sum_{\scriptscriptstyle \alpha\in\nabla} m_\alpha \varphi( {{\bf s}}^\alpha ) = \sum_{ \scriptscriptstyle e\in\Delta } \left( \sum_{ \substack{\scriptscriptstyle \alpha\in\nabla\\ \scriptscriptstyle |\alpha|\leq |e| } } m_\alpha K_{e,\alpha} \right) {{\bf t}}^e,\quad m\in M[[{{\bf s}}]]_\nabla.&\end{aligned}$$ If $\varphi': A[[{{\bf t}}]]_\Delta \to A[[{{\bf u}}]]_\Omega$ is another $A$-linear continuous map satisfying the assumptions in \[nume:cont-maps-power-series-rings\] and $\operatorname{\varphi}'' = \operatorname{\varphi}{{\scriptstyle \,\circ\,}}\operatorname{\varphi}'$, we have $\operatorname{\varphi}''_M = \varphi_M {{\scriptstyle \,\circ\,}}\varphi'_M$, $\sideset{_M}{}\operatorname{\varphi}'' = \sideset{_M}{}\operatorname{\varphi}{{\scriptstyle \,\circ\,}}\sideset{_M}{}\operatorname{\varphi}'$. If $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ is a substitution map and $m\in M[[{{\bf s}}]]_\nabla$, $a \in A[[{{\bf s}}]]_\nabla$, we have $$\varphi_M (a m) = \varphi(a) \varphi_M(m),\ \sideset{_M}{}\operatorname{\varphi}(m a) = \sideset{_M}{}\operatorname{\varphi}(m) \varphi(a),$$ i.e. $\varphi_M$ is $(\varphi;A)$-linear and $\sideset{_M}{}\operatorname{\varphi}$ is $(A;\varphi)$-linear. Moreover, $\varphi_M$ and $\sideset{_M}{}\operatorname{\varphi}$ are compatible with the augmentations, i.e. $$\label{eq:compat-augment} \varphi_M(m) \equiv m_0,\sideset{_M}{}\operatorname{\varphi}(m) \equiv m_0 \!\!\!\!\mod {\mathfrak{n}}^M_0({{\bf t}})/\Delta_M,\quad m\in M[[{{\bf s}}]]_\nabla.$$ If $\varphi$ is the trivial substitution map (i.e. $\varphi(s)=0$ for all $s\in{{\bf s}}$), then $\varphi_M : M[[{{\bf s}}]]_\nabla \to M[[{{\bf t}}]]_\Delta$ and $\sideset{_M}{}\operatorname{\varphi}: M[[{{\bf s}}]]_\nabla \to M[[{{\bf t}}]]_\Delta$ are also trivial, i.e. $$\varphi_M (m) = \sideset{_M}{}\operatorname{\varphi}(m) = m_0,\ m\in M[[{{\bf s}}]]_\nabla.$$ [[**.  **]{}]{}\[nume:def-sbullet\] The above constructions apply in particular to the case of any $k$-algebra $R$ over $A$, for which we have two induced continuous maps, $\varphi_R= \varphi \widehat{\otimes} {{\rm Id}}_R : R[[{{\bf s}}]]_\nabla \to R[[{{\bf t}}]]_\Delta$, which is $(A;R)$-linear, and $\sideset{_R}{}\operatorname{\varphi}= {{\rm Id}}_R \widehat{\otimes} \varphi : R[[{{\bf s}}]]_\nabla \to R[[{{\bf t}}]]_\Delta$, which is $(R;A)$-linear. For $r\in R[[{{\bf s}}]]_\nabla$ we will denote $$\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r := \varphi_R(r),\quad r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi := \sideset{_R}{}\operatorname{\varphi}(r).$$ Explicitely, if $r=\sum_\alpha r_\alpha {{\bf s}}^\alpha$ with $\alpha\in\nabla$, then $$\label{eq:explicit-bullet} \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r = \sum_{\scriptscriptstyle e \in\Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in\nabla\\ \scriptscriptstyle |\alpha|\leq |e|}} {\bf C}_e(\varphi,\alpha) r_\alpha \right) {{\bf t}}^e,\quad r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi = \sum_{\scriptscriptstyle e \in\Delta} \left( \sum_{\substack{\scriptscriptstyle \alpha\in\nabla\\ \scriptscriptstyle |\alpha|\leq |e|}} r_\alpha {\bf C}_e(\varphi,\alpha) \right) {{\bf t}}^e.$$ From (\[eq:compat-augment\]), we deduce that $\varphi_R(\operatorname{\mathcal{U}}^{{\bf s}}(R;\nabla)) \subset \operatorname{\mathcal{U}}^{{\bf t}}(R;\Delta)$ and $\sideset{_R}{}\operatorname{\varphi}(\operatorname{\mathcal{U}}^{{\bf s}}(R;\nabla)) \subset \operatorname{\mathcal{U}}^{{\bf t}}(R;\Delta)$. We also have $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}1 = 1 {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi = 1$. If $\varphi$ is a substitution map with , then $\varphi_R = \sideset{_R}{}\operatorname{\varphi}$ is a ring homomorphism over $\varphi$. In particular, $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r = r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi$ and $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(rr') = (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r) (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r')$. If $\varphi = \mathbf{0}: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ is the trivial substitution map, then $\mathbf{0} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r = r{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\mathbf{0} = r_0$ for all $r\in R[[{{\bf s}}]]_\nabla$. In particular, $\mathbf{0} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r = r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\mathbf{0} = 1$ for all $r\in \operatorname{\mathcal{U}}^{{\bf s}}(R;\nabla)$. If $\psi:R[[{{\bf t}}]]_\Delta \to R[[{{\bf u}}]]_\Omega$ is another substitution map, one has $$\psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r) = (\psi {{\scriptstyle \,\circ\,}}\varphi) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r,\quad (r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\psi = r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(\psi {{\scriptstyle \,\circ\,}}\varphi).$$ Since $\left(R[[{{\bf s}}]]_\nabla\right)^{\text{opp}} = R^{\text{opp}}[[{{\bf s}}]]_\nabla$, for any substitution map $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ we have $\left( \varphi_R \right)^{\text{opp}} = \sideset{_{R^{\text{opp}}}}{}\operatorname{\varphi}$ and $\left( \sideset{_R}{}\operatorname{\varphi}\right)^{\text{opp}} = \varphi_{R^{\text{opp}}}$. The proof of the following lemma is straighforward and it is left to the reader. \[lemma:phi-linearity-of-phi\_R\] If $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ is a substitution map, then: 1. $\varphi_R$ is left $\varphi$-linear, i.e. $\varphi_R(ar) = \varphi(a) \varphi_R(r)$ for all $a\in A[[{{\bf s}}]]_\nabla$ and for all $r\in R[[{{\bf s}}]]_\nabla$. 2. $\sideset{_R}{}\operatorname{\varphi}$ is right $\varphi$-linear, i.e. $\sideset{_R}{}\operatorname{\varphi}(ra)= \sideset{_R}{}\operatorname{\varphi}(r) \varphi(a)$ for all $a\in A[[{{\bf s}}]]_\nabla$ and for all $r\in R[[{{\bf s}}]]_\nabla$. Let us assume again that $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ is an $A$-linear continuous map satisfying the assumptions in \[nume:cont-maps-power-series-rings\]. We define the $(A;A)$-linear map $$\varphi_*: f\in \operatorname{Hom}_k(A,A[[{{\bf s}}]]_\nabla) \longmapsto \varphi_*(f)= \varphi {{\scriptstyle \,\circ\,}}f \in \operatorname{Hom}_k(A,A[[{{\bf t}}]]_\Delta)$$ which induces another one $\overline{\varphi_*}: \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(A[[{{\bf s}}]]_\nabla) \longrightarrow \operatorname{End}_{k[[{{\bf t}}]]_\Delta}^{\text{\rm top}}(A[[{{\bf t}}]]_\Delta)$ defined by $$\overline{\varphi_*}(f):= \left(\varphi_*\left(f|_A\right)\right)^e = \left(\varphi {{\scriptstyle \,\circ\,}}f|_A\right)^e,\quad f\in \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(A[[{{\bf s}}]]_\nabla) .$$ More generally, for a given left $A$-module $E$ (which will be considered as a trivial $(A;A)$-bimodule) we have $(A;A)$-linear maps $$\begin{aligned} &\displaystyle (\varphi_{E})_*: f\in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\nabla) \mapsto (\varphi_{E})_*(f)= \varphi_E {{\scriptstyle \,\circ\,}}f \in \operatorname{Hom}_k(E,E[[{{\bf t}}]]_\Delta), &\\ &\displaystyle \overline{(\varphi_E)_*}: \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(E[[{{\bf s}}]]_\nabla) \rightarrow \operatorname{End}_{k[[{{\bf t}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf t}}]]_\Delta),\quad \overline{(\varphi_E)_*}(f):=\left(\varphi_E {{\scriptstyle \,\circ\,}}f|_A\right)^e.\end{aligned}$$ Let us denote $R=\operatorname{End}_k(E)$. For each $r\in R[[{{\bf s}}]]_\nabla$ and for each $e\in E$ we have $$\widetilde{\varphi_R(r)}(e) = \varphi_E\left(\widetilde{r}(e) \right),$$ or more graphically, the following diagram is commutative (see (\[eq:comple\_formal\_1\])): $$\begin{CD} R[[{{\bf s}}]]_\nabla @>{\sim}>{r\mapsto \widetilde{r}}> \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(E[[{{\bf s}}]]_\nabla) @>{\sim}>{\text{rest.}}> \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\nabla) \\ @V{\varphi_R}VV @VV{\overline{(\varphi_E)_*}}V @V{(\varphi_{E})_*}VV\\ R[[{{\bf t}}]]_\Delta @>{\sim}>{r\mapsto \widetilde{r}}> \operatorname{End}_{k[[{{\bf t}}]]_\Delta}^{\text{\rm top}}(E[[{{\bf t}}]]_\Delta) @>{\sim}>{\text{rest.}}> \operatorname{Hom}_k(E,E[[{{\bf t}}]]_\Delta). \end{CD}$$ In order to simplify notations, we will also write $$\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}f := \overline{(\varphi_E)_*}(f)\quad \forall f \in \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(E[[{{\bf s}}]]_\nabla)$$ and so have $ \widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r} = \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\widetilde{r}$ for all $r\in R[[{{\bf s}}]]_\nabla$. Let us notice that $(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}f)(e) = (\varphi_E {{\scriptstyle \,\circ\,}}f)(e)$ for all $e\in E$, i.e. $$\label{eq:atencion-bullet} \framebox{$(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}f)|_E = (\varphi_E {{\scriptstyle \,\circ\,}}f)|_E$, but in general $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}f \neq \varphi_E {{\scriptstyle \,\circ\,}}f$.}$$ If $\varphi$ is the trivial substitution map, then $(\varphi_E)_*$ (resp. $\overline{(\varphi_E)_*}$) is also trivial in the sense that if $f=\sum_\alpha f_\alpha {{\bf s}}^\alpha \in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\nabla)$ (resp. $f=\sum_\alpha f_\alpha {{\bf s}}^\alpha \in\operatorname{End}_k(E)[[{{\bf s}}]]_\nabla \equiv \operatorname{End}^{\text{\rm top}}_{k[[{{\bf s}}]]_\nabla}(E[[{{\bf s}}]]_\nabla)$), then $(\varphi_E)_*(f)= f_0 \in \operatorname{End}_k(E) \subset \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\nabla)$ (resp. $\overline{(\varphi_E)_*}(f)= f_0^e \in \operatorname{End}^{\text{\rm top}}_{k[[{{\bf s}}]]_\nabla}(E[[{{\bf s}}]]_\nabla)$, with $f_0^e(\sum_\alpha e_\alpha {{\bf s}}^\alpha) = \sum_\alpha f_0(e_\alpha) {{\bf s}}^\alpha$). If $\varphi: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\nabla$ is a substitution map, we have $$\begin{aligned} &\displaystyle (\varphi_{E})_*(af) = \varphi(a) (\varphi_{E})_*(f) \quad \forall a\in A[[{{\bf s}}]]_\nabla, \forall f\in \operatorname{Hom}_k(E,E[[{{\bf s}}]]_\nabla)\end{aligned}$$ and so $$\begin{aligned} &\displaystyle \overline{(\varphi_{E})_*}(af) = \varphi(a) \overline{(\varphi_{E})_*}(f) \quad \forall a\in A[[{{\bf s}}]]_\nabla, \forall f\in \operatorname{End}_{k[[{{\bf s}}]]_\nabla}^{\text{\rm top}}(E[[{{\bf s}}]]_\nabla).\end{aligned}$$ Moreover, the following inclusions hold $$\begin{aligned} &(\varphi_E)_*(\operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,M[[{{\bf s}}]]_\nabla)) \subset \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf t}}]]_\Delta), &\\ &\overline{(\varphi_E)_*} \left( \operatorname{Aut}_{k[[{{\bf s}}]]_\nabla}^{{\scriptstyle \,\circ\,}}(E[[{{\bf s}}]]_\nabla) \right) \subset \operatorname{Aut}_{k[[{{\bf t}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(E[[{{\bf t}}]]_\Delta),\end{aligned}$$ and so we have a commutative diagram: $$\label{eq:diag-funda} \begin{CD} \operatorname{\mathcal{U}}^{{\bf s}}(R;\nabla) @>{\sim}>{r\mapsto \widetilde{r}}> \operatorname{Aut}_{k[[{{\bf s}}]]_\nabla}^{{\scriptstyle \,\circ\,}}(E[[{{\bf s}}]]_\nabla) @>{\sim}>{\text{rest.}}> \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf s}}]]_\nabla) \\ @V{\varphi_R}VV @VV{\overline{(\varphi_E)_*}}V @V{(\varphi_{E})_*}VV\\ \operatorname{\mathcal{U}}^{{\bf t}}(R;\Delta) @>{\sim}>{r\mapsto \widetilde{r}}> \operatorname{Aut}_{k[[{{\bf t}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(E[[{{\bf t}}]]_\Delta) @>{\sim}>{\text{rest.}}> \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(E,E[[{{\bf t}}]]_\Delta). \end{CD}$$ \[lemma:const\_coeff\_subst\_map\] With the notations above, if $\varphi: k[[{{\bf s}}]]_\nabla \to k[[{{\bf t}}]]_\Delta$ is a substitution map with constant coefficients, then $$\langle \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r,\varphi_E(e)\rangle = \varphi_E(\langle r,e\rangle),\quad \forall r \in R[[{{\bf s}}]]_\nabla, \forall e\in E[[{{\bf s}}]]_\nabla.$$ Let us write $r=\sum_\alpha r_\alpha {{\bf s}}^\alpha$, $r_\alpha \in R = \operatorname{End}_k(E)$ and $e=\sum_\alpha e_\alpha {{\bf s}}^\alpha$, $e_\alpha \in E$. We have $$\begin{aligned} & \displaystyle \langle \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r,\varphi_E(e)\rangle = (\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r})(\varphi_E(e)) = \left( \sum_\alpha \varphi({{\bf s}}^\alpha) \widetilde{r_\alpha} \right) \left( \sum_\alpha \varphi({{\bf s}}^\alpha) e_\alpha \right) =& \\ & \displaystyle \sum_{\alpha,\beta}\varphi({{\bf s}}^\alpha) \widetilde{r_\alpha} \left(\varphi({{\bf s}}^\beta) e_\beta\right)= \sum_{\alpha,\beta}\varphi({{\bf s}}^\alpha)\varphi({{\bf s}}^\beta) \widetilde{r_\alpha} \left(e_\beta\right)= \sum_{\alpha,\beta}\varphi({{\bf s}}^{\alpha+\beta}) \widetilde{r_\alpha} (e_\beta)= & \\ & \displaystyle \sum_{\gamma} \varphi({{\bf s}}^\gamma) \left(\sum_{\alpha+\beta=\gamma} \widetilde{r_\alpha} (e_\beta)\right) = \varphi_E \left( \sum_{\gamma} \left(\sum_{\alpha+\beta=\gamma} \widetilde{r_\alpha} (e_\beta)\right) {{\bf s}}^\gamma \right) &\\ & \displaystyle = \varphi_E\left(\widetilde{r}(e)\right) =\varphi_E(\langle r,e\rangle).\end{aligned}$$ Notice that if $\varphi: k[[{{\bf s}}]]_\nabla \to k[[{{\bf t}}]]_\Delta$ is a substitution map with constant coefficients, we already pointed out that $\sideset{_R}{}\operatorname{\varphi}= \varphi_R$, and indeed, $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r = r {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\varphi$ for all $r\in R[[{{\bf s}}]]_\nabla$. [[**.  **]{}]{}\[nume:properties-external-x\] Let us denote $\iota: A[[{{\bf s}}]]_\nabla \to A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$, $\kappa: A[[{{\bf t}}]]_\Delta \to A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$ the combinatorial substitution maps given by the inclusions ${{\bf s}}\hookrightarrow {{\bf s}}\sqcup {{\bf t}}$, ${{\bf t}}\hookrightarrow {{\bf s}}\sqcup {{\bf t}}$. Let us notice that for $r\in R[[{{\bf s}}]]_\nabla$ and $r'\in R[[{{\bf t}}]]_\Delta$, we have (see Definition \[defi:external-x\]) $ r{{\scriptstyle \,\boxtimes\,}}r' = (\iota{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r) (\kappa{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r') \in R[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$. If $\nabla'\subset\nabla\subset{\mathbb{N}}^{({{\bf s}})}$, $\Delta'\subset\Delta\subset {\mathbb{N}}^{({{\bf t}})}$ are non-empty co-ideals, we have $$\tau_{\nabla \times \Delta,\nabla' \times \Delta'}(r{{\scriptstyle \,\boxtimes\,}}r') = \tau_{\nabla,\nabla'}(r) {{\scriptstyle \,\boxtimes\,}}\tau_{\Delta,\Delta'}(r').$$ If we denote by $\Sigma: R[[{{\bf s}}\sqcup {{\bf s}}]]_{\nabla\times\nabla} \rightarrow R[[{{\bf s}}]]_\nabla$ the combinatorial substitution map given by the co-diagonal map ${{\bf s}}\sqcup {{\bf s}}\to {{\bf s}}$, it is clear that for each $r,r'\in R[[{{\bf s}}]]_\nabla$ we have $$\label{eq:prod-in-terms-x} r r' = \Sigma{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(r {{\scriptstyle \,\boxtimes\,}}r').$$ If $\varphi:A[[{{\bf s}}]]_\nabla \to A[[{{\bf u}}]]_\Omega$ and $\psi:A[[{{\bf t}}]]_\Delta \to A[[{{\bf v}}]]_{\Omega'}$ are substitution maps, we have new substitution maps $\varphi \otimes {{\rm Id}}: A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla\times\Delta} \to A[[{{\bf u}}\sqcup {{\bf t}}]]_{\Omega\times\Delta}$ and ${{\rm Id}}\otimes \psi: A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla\times\Delta} \to A[[{{\bf s}}\sqcup {{\bf v}}]]_{\nabla\times\Omega'}$ (see Definition \[defi:tensor-prod-of-substi\]) taking part in the following commutative diagrams of $(A;A)$-bimodules $$\begin{CD} R[[{{\bf s}}]]_\nabla \otimes_R R[[{{\bf t}}]]_\Delta @>{\varphi_R \otimes{{\rm Id}}}>> R[[{{\bf u}}]]_\Omega \otimes_R R[[{{\bf t}}]]_\Delta\\ @V{\text{can.}}VV @VV{\text{can.}}V\\ R[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla\times\Delta} @>{\left(\varphi\otimes{{\rm Id}}\right)_R}>> R[[{{\bf u}}\sqcup {{\bf t}}]]_{\Omega\times\Delta} \end{CD}$$ and $$\begin{CD} R[[{{\bf s}}]]_\nabla \otimes_R R[[{{\bf t}}]]_\Delta @>{{{\rm Id}}\otimes\psi }>> R[[{{\bf s}}]]_\nabla \otimes_R R[[{{\bf v}}]]_{\Omega'}\\ @V{\text{can.}}VV @VV{\text{can.}}V\\ R[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla\times\Delta} @>{\left({{\rm Id}}\otimes\varphi\right)_R}>> R[[{{\bf s}}\sqcup {{\bf v}}]]_{\nabla\times\Omega'}. \end{CD}$$ So $(\varphi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}r) {{\scriptstyle \,\boxtimes\,}}r'= (\varphi\otimes {{\rm Id}}){{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(r{{\scriptstyle \,\boxtimes\,}}r')$ and $r{{\scriptstyle \,\boxtimes\,}}(r'{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\psi)= (r{{\scriptstyle \,\boxtimes\,}}r'){{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}({{\rm Id}}\otimes \psi)$. Multivariate Hasse-Schmidt derivations {#section:HS} ====================================== In this section we study multivariate (possibly $\infty$-variate) Hasse–Schmidt derivations. The original reference for 1-variate Hasse–Schmidt derivations is [@has37]. This notion has been studied and developed in [@mat_86 §27] (see also [@vojta_HS] and [@nar_2012]). In [@hoff_kow_2016] the authors study “finite dimensional” Hasse–Schmidt derivations, which correspond in our terminology to $p$-variate Hasse–Schmidt derivations. From now on $k$ will be a commutative ring, $A$ a commutative $k$-algebra, ${{\bf s}}$ a set and $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$ a non-empty co-ideal. A [*$({{\bf s}},\Delta)$-variate Hasse-Schmidt derivation*]{}, or a [*$({{\bf s}},\Delta)$-variate HS-derivation*]{} for short, of $A$ over $k$ is a family $D=(D_\alpha)_{\alpha\in \Delta}$ of $k$-linear maps $D_\alpha:A \longrightarrow A$, satisfying the following Leibniz type identities: $$D_0={{\rm Id}}_A, \quad D_\alpha(xy)=\sum_{\beta+\gamma=\alpha}D_\beta(x)D_\gamma(y)$$ for all $x,y \in A$ and for all $\alpha\in \Delta$. We denote by $\operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ the set of all $({{\bf s}},\Delta)$-variate HS-derivations of $A$ over $k$ and $\operatorname{HS}^{{{\bf s}}}_k(A)=$ for $\Delta = {\mathbb{N}}^{({{\bf s}})}$. In the case where ${{\bf s}}= \{1,\dots,p\}$, a $({{\bf s}},\Delta)$-variate HS-derivation will be simply called a [*$(p,\Delta)$-variate HS-derivation*]{} and we denote $\operatorname{HS}^p_k(A;\Delta):= \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ and $\operatorname{HS}^p_k(A) := \operatorname{HS}^{{{\bf s}}}_k(A)$. For $p=1$, a $1$-variate HS-derivation will be simply called a [*Hasse–Schmidt derivation*]{} (a HS-derivation for short), or a [*higher derivation*]{}[^4], and we will simply write $\operatorname{HS}_k(A;m):= \operatorname{HS}^1_k(A;\Delta)$ for $\Delta=\{q\in{\mathbb{N}}\ |\ q\leq m\}$[^5] and $\operatorname{HS}_k(A) := \operatorname{HS}^1_k(A)$. [[**.  **]{}]{}\[nume:leibniz-HS\] The above Leibniz identities for $D\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ can be written as $$\label{eq:leibniz-HS} D_\alpha x=\sum_{\beta+\gamma=\alpha}D_\beta(x)D_\gamma,\quad \forall x\in A, \forall \alpha\in \Delta.$$ Any $({{\bf s}},\Delta)$-variate HS-derivation $D$ of $A$ over $k$ can be understood as a power series $$\sum_{\scriptscriptstyle \alpha\in\Delta} D_\alpha {{\bf s}}^\alpha \in \operatorname{End}_k(A)[[{{\bf s}}]]_\Delta$$ and so we consider $\operatorname{HS}^{{\bf s}}_k(A;\Delta) \subset \operatorname{End}_k(A)[[{{\bf s}}]]_\Delta$. Let $D\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ be a HS-derivation. Then, for each $\alpha\in\Delta$, the component $D_\alpha:A \to A$ is a $k$-linear differential operator or order $\leq |\alpha|$ vanishing on $k$. In particular, if $|\alpha|=1$ then $D_\alpha:A \to A$ is a $k$-derivation. The proof follows by induction on $|\alpha|$ from \[eq:leibniz-HS\]. The map $$\label{eq:HS1-Der} D\in \operatorname{HS}^{{\bf s}}_k(A;{\mathfrak{t}}_1({{\bf s}})) \mapsto \{D_\alpha\}_{|\alpha|=1} \in \operatorname{Der}_k(A)^{{\bf s}}$$ is clearly a bijection. The proof of the following proposition is straightforward and it is left to the reader (see Notation \[notacion:Ump\] and \[nume:tilde\]). \[prop:caracter\_HS\] Let us denote $R=\operatorname{End}_k(A)$ and let $D= \sum_\alpha D_\alpha {{\bf s}}^\alpha \in R[[{{\bf s}}]]_\Delta$ be a power series. The following properties are equivalent: 1. $D$ is a $({{\bf s}},\Delta)$-variate HS-derivation of $A$ over $k$. 2. The map $\widetilde{D}: A[[{{\bf s}}]]_\Delta \to A[[{{\bf s}}]]_\Delta$ is a (continuous) $k[[{{\bf s}}]]_\Delta$-algebra homomorphism compatible with the natural augmentation $A[[{{\bf s}}]]_\Delta \to A$. 3. $D\in\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ and for all $a\in A[[{{\bf s}}]]_\Delta$ we have $D a = \widetilde{D}(a) D$. 4. $D\in\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$ and for all $a\in A$ we have $D a = \widetilde{D}(a) D$. Moreover, in such a case $\widetilde{D}$ is a bi-continuous $k[[{{\bf s}}]]_\Delta$-algebra automorphism of $A[[{{\bf s}}]]_\Delta$. Under the above hypotheses, $\operatorname{HS}^{{\bf s}}_k(A;\Delta)$ is a (multiplicative) sub-group of $\operatorname{\mathcal{U}}^{{\bf s}}(R;\Delta)$. If $\Delta' \subset \Delta \subset {\mathbb{N}}^{({{\bf s}})}$ are non-empty co-ideals, we obviously have group homomorphisms $\tau_{\Delta \Delta'}: \operatorname{HS}^{{\bf s}}_k(A;\Delta) \longrightarrow \operatorname{HS}^{{\bf s}}_k(A;\Delta')$. Since any $D\in\operatorname{HS}_k^{{\bf s}}(A;\Delta)$ is determined by its finite truncations, we have a natural group isomorphism $$\operatorname{HS}_k^{{\bf s}}(A) = \lim_{\stackrel{\longleftarrow}{\substack{\scriptscriptstyle \Delta' \subset \Delta\\ \scriptscriptstyle \sharp \Delta'<\infty}}} \operatorname{HS}_k^{{\bf s}}(A;\Delta').$$ In the case $\Delta' = \Delta^1 = \Delta\cap {\mathfrak{t}}_1({{\bf s}})$, since $\operatorname{HS}_k^{{\bf s}}(A;\Delta^1)\simeq \operatorname{Der}_k(A)^{\Delta^1}$, we can think on $\tau_{\Delta \Delta^1}$ as a group homomorphism $\tau_{\Delta\Delta^1}:\operatorname{HS}_k^{{\bf s}}(A;\Delta)\to \operatorname{Der}_k(A)^{\Delta^1}$ whose kernel is the normal subgroup of $\operatorname{HS}_k^{{\bf s}}(A;\Delta)$ consisting of HS-derivations $D$ with $D_\alpha = 0$ whenever $|\alpha|=1$. In the case $\Delta'= \Delta^n = \Delta\cap {\mathfrak{t}}_n({{\bf s}})$, for $n\geq 1$, we will simply write $\tau_n = \tau_{\Delta,\Delta^n}: \operatorname{HS}^{{\bf s}}_k(A;\Delta) \longrightarrow \operatorname{HS}^{{\bf s}}_k(A;\Delta^n)$. \[nota:HS-as-Rees\] Since for any $D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ we have $D_{\alpha}\in \operatorname{\mathcal{D}iff}_{A/k}^{|\alpha|}(A)$, we may also think on $D$ as an element in a generalized Rees ring of the ring of differential operators: $$\widehat{\operatorname{\mathcal{R}}}^{{\bf s}}\left({\mathcal{D}}_{A/k}(A);\Delta\right) := \left\{ \sum_{\scriptscriptstyle\alpha\in\Delta} r_\alpha {{\bf s}}^\alpha\in {\mathcal{D}}_{A/k}(A)[[{{\bf s}}]]_\Delta \ |\ r_\alpha \in \operatorname{\mathcal{D}iff}_{A/k}^{|\alpha|}(A)\right\}.$$ The group operation in $\operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ is explicitely given by $$(D,E) \in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta) \times \operatorname{HS}^{{{\bf s}}}_k(A;\Delta) \longmapsto D {{\scriptstyle \,\circ\,}}E \in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$$ with $$(D {{\scriptstyle \,\circ\,}}E)_\alpha = \sum_{\scriptscriptstyle \beta+\gamma=\alpha} D_\beta {{\scriptstyle \,\circ\,}}E_\gamma,$$ and the identity element of $\operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ is $\mathbb{I}$ with $\mathbb{I}_0 = {{\rm Id}}$ and $\mathbb{I}_\alpha = 0$ for all $\alpha \neq 0$. The inverse of a $D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ will be denoted by $D^*$. Let $D\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$, $E \in \operatorname{HS}^{{{\bf t}}}_k(A;\nabla)$ be HS-derivations. Then their external product $D{{\scriptstyle \,\boxtimes\,}}E$ (see Definition \[defi:external-x\]) is a $({{\bf s}}\sqcup {{\bf t}},\nabla\times\Delta)$-variate HS-derivation. From Lemma \[lemma:tilde-otimes\] we know that $\widetilde{D{{\scriptstyle \,\boxtimes\,}}E} = \widetilde{D}{{\scriptstyle \,\boxtimes\,}}\widetilde{E}$ and we conclude by Proposition \[prop:caracter\_HS\]. \[defi:sbullet-0\] For each $a\in A^{{\bf s}}$ and for each $D\in \operatorname{HS}_k^{{\bf s}}(A;\Delta)$, we define $a {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$ as $$(a{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)_\alpha := a^\alpha D_\alpha,\quad \forall \alpha\in \Delta.$$ It is clear that $a {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\in \operatorname{HS}_k^{{\bf s}}(A;\Delta)$, $a'{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(a{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = (a'a){{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$, $1 {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D=D$ and $0 {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D=\mathbb{I}$. If $\Delta' \subset \Delta \subset {\mathbb{N}}^{({{\bf s}})}$ are non-empty co-ideals, we have $\tau_{\Delta \Delta'}(a{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)=a{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{\Delta \Delta'}(D)$. Hence, in the case $\Delta' = \Delta^1 = \Delta\cap {\mathfrak{t}}_1({{\bf s}})$, since $\operatorname{HS}_k^{{\bf s}}(A;\Delta^1)\simeq \operatorname{Der}_k(A)^{\Delta^1}$, the image of $\tau_{\Delta\Delta^1}:\operatorname{HS}_k^{{\bf s}}(A;\Delta)\to \operatorname{Der}_k(A)^{\Delta^1}$ is an $A$-submodule. The following lemma provides a dual way to express the Leibniz identity (\[eq:leibniz-HS\]), \[nume:leibniz-HS\]. \[lema:dual-leibniz-HS\] For each $D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ and for each $\alpha\in \Delta$, we have $$x D_\alpha =\sum_{\beta+\gamma=\alpha}D_\beta\, D^*_\gamma(x),\quad \forall x\in A.$$ We have $$\begin{aligned} &\displaystyle \sum_{\beta+\gamma=\alpha}D_\beta\, D^*_\gamma(x) = \sum_{\beta+\gamma=\alpha} \sum_{\mu+\nu=\beta} D_\mu(D^*_\gamma(x)) D_\nu = &\\ &\displaystyle \sum_{e+\nu=\alpha} \left( \sum_{\mu+\gamma=e} D_\mu(D^*_\gamma(x))\right) D_\nu = x D_\alpha.\end{aligned}$$ It is clear that the map (\[eq:HS1-Der\]) is an isomorphism of groups (with the addition on $\operatorname{Der}_k(A)$ as internal operation) and so $\operatorname{HS}_k^{{\bf s}}(A;{\mathfrak{t}}_1({{\bf s}}))$ is abelian. \[notacion:pcirc-HS\] Let us denote $$\begin{aligned} &\displaystyle \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta) := &\\ &\displaystyle \left\{ f \in \operatorname{Hom}_{k-\text{\rm alg}}(A,A[[{{\bf s}}]]_\Delta)\ |\ f(a) \equiv a\!\!\!\!\mod {\mathfrak{n}}^A_0({{\bf s}})/\Delta_A\ \forall a\in A \right\}, &\\ &\displaystyle \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta) := &\\ &\displaystyle \left\{ f \in \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{\text{\rm top}}(A[[{{\bf s}}]]_\Delta)\ |\ f(a) \equiv a_0\!\!\!\!\mod {\mathfrak{n}}^A_0({{\bf s}})/\Delta_A\ \forall a\in A[[{{\bf s}}]]_\Delta \right\}.\end{aligned}$$ It is clear that (see Notation \[notacion:pcirc\]) $\operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta) \subset \operatorname{Hom}_k^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta)$ and $\operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta) \subset \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta)$ are subgroups and we have group isomorphisms (see (\[eq:Aut-iso-pcirc\]) and (\[eq:U-iso-pcirc\])): $$\label{eq:HS-funda} \begin{CD} \operatorname{HS}^{{\bf s}}_k(A;\Delta) @>{D \mapsto \widetilde{D}}>{\simeq}> \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta) @>{\text{restriction}}>{\simeq}> \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta). \end{CD}$$ The composition of the above isomorphisms is given by $$\label{eq:HS-iso-Hom(A,A[[s]])} D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta) \stackrel{\sim}{\longmapsto} \Phi_D := \left[a\in A \mapsto \sum_{\scriptscriptstyle \alpha\in\Delta} D_\alpha(a) {{\bf s}}^\alpha\right] \in \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta).$$ For each HS-derivation $D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ we have $$\widetilde{D}\left( \sum_{\scriptscriptstyle \alpha\in\Delta} a_\alpha {{\bf s}}^\alpha \right) = \sum_{\scriptscriptstyle \alpha\in\Delta} \Phi_D(a_\alpha) {{\bf s}}^\alpha,$$ for all $\sum_\alpha a_\alpha {{\bf s}}^\alpha \in A[[{{\bf s}}]]_\Delta$, and for any $E\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ we have $\Phi_{D {{\scriptstyle \circ}}E} = \widetilde{D} {{\scriptstyle \,\circ\,}}\Phi_E$. If $\Delta'\subset \Delta$ is another non-empty co-ideal and we denote by $\pi_{\Delta \Delta'}: A[[{{\bf s}}]]_\Delta \to A[[{{\bf s}}]]_{\Delta'}$ the projection, one has $\Phi_{\tau_{\Delta \Delta'} (D)} = \pi_{\Delta \Delta'}{{\scriptstyle \,\circ\,}}\Phi_D$. \[def:ell-new\] For each HS-derivation $E\in\operatorname{HS}^{{\bf s}}_k(A;\Delta)$, we denote $${\ell}(E) := \min \{r\geq 1 \ |\ \exists \alpha\in \Delta, |\alpha|= r, E_\alpha \neq 0 \} \geq 1$$ if $E\neq \mathbb{I}$ and $ {\ell}(E) = \infty$ if $E= \mathbb{I}$. In other words, $\ell (E) = \operatorname{ord}(E-\mathbb{I})$. Clearly, if $\Delta$ is bounded, then ${\ell}(E)>\max\{|\alpha|\ |\ \alpha\in\Delta\} \Longleftrightarrow {\ell}(E)=\infty \Longleftrightarrow E = \mathbb{I}$. We obviously have ${\ell}(E{{\scriptstyle \,\circ\,}}E') \geq \min \{{\ell}(E),{\ell}(E')\}$ and ${\ell}(E^*) = {\ell}(E)$. Moreover, if ${\ell}(E') > {\ell}(E)$, then ${\ell}(E{{\scriptstyle \,\circ\,}}E')={\ell}(E)$: $${\ell}(E{{\scriptstyle \,\circ\,}}E') = \operatorname{ord}(E{{\scriptstyle \,\circ\,}}E' - \mathbb{I}) = \operatorname{ord}( E {{\scriptstyle \,\circ\,}}(E' - \mathbb{I}) + (E - \mathbb{I}))$$ and since $\operatorname{ord}(E {{\scriptstyle \,\circ\,}}(E' - \mathbb{I})) \geq\footnote{Actually, here an equality holds since the $0$-term of $E$ (as a series) is $1$.} \operatorname{ord}( E' - \mathbb{I}) = {\ell}(E') > {\ell}(E) = \operatorname{ord}( E - \mathbb{I})$ we obtain $${\ell}(E{{\scriptstyle \,\circ\,}}E') = \cdots = \operatorname{ord}( E {{\scriptstyle \,\circ\,}}(E' - \mathbb{I}) + (E - \mathbb{I})) = \operatorname{ord}(E - \mathbb{I}) = {\ell}(E).$$ \[prop:ell-new\] For each $D\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ we have that $D_\alpha$ is a $k$-linear differential operator or order $\leq \lfloor \frac{|\alpha|}{{\ell}(D)}\rfloor$ for all $\alpha\in \Delta$. In particular, $D_\alpha$ is a $k$-derivation if $|\alpha|={\ell}(D)$, whenever ${\ell}(D) < \infty$ ( $ \Leftrightarrow D\neq \mathbb{I}$). We may assume $D\neq \mathbb{I}$. Let us call $n:={\ell}(D) <\infty$ and, for each $\alpha\in \Delta$, $q_\alpha := \lfloor \frac{|\alpha|}{n}\rfloor$ and $r_\alpha := |\alpha| - q_\alpha n $, $0\leq r_\alpha < n$. We proceed by induction on $q_\alpha$. If $q_\alpha=0$, then $|\alpha|< n$, $D_\alpha =0$ and the result is clear. Assume that the order of $D_\beta$ is less or equal than $q_\beta$ whenever $0\leq q_\beta \leq q$. Now take $\alpha\in \Delta$ with $q_\alpha = q+1$. For any $a\in A$ we have $$[D_\alpha,a] = \sum_{\substack{\scriptscriptstyle \gamma+\beta=\alpha\\ \scriptscriptstyle |\gamma|>0}} D_\gamma(a) D_\beta = \sum_{\substack{\scriptscriptstyle \gamma+\beta=\alpha\\ \scriptscriptstyle |\gamma|\geq n}} D_\gamma(a) D_\beta,$$ but any $\beta$ in the index set of the above sum must have norm $\leq |\alpha|-n$ and so $q_\beta < q_\alpha =q+1$ and $D_\beta$ has order $\leq q_\beta$. Hence $[D_\alpha,a]$ has order $\leq q$ for any $a\in A$ and $D_\alpha$ has order $\leq q+1 =q_\alpha$. The following example shows that the group structure on HS-derivations takes into account the Lie bracket on usual derivations. If $D,E\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$, then we may apply the above proposition to $[D,E] = D{{\scriptstyle \,\circ\,}}E{{\scriptstyle \,\circ\,}}D^*{{\scriptstyle \,\circ\,}}E^*$ to deduce that $[D,E]_\alpha \in \operatorname{Der}_k(A)$ whenever $|\alpha|=2$. Actually, for $|\alpha|=2$ we have: $$[D,E]_\alpha = \left\{ \begin{array}{lcl} [D_{{{\bf s}}^t},E_{{{\bf s}}^t}] & \text{if} & \alpha= 2 {{\bf s}}^t \\ {[}D_{{{\bf s}}^t},E_{{{\bf s}}^u}{]} + [D_{{{\bf s}}^u},E_{{{\bf s}}^t}] & \text{if} & \alpha= {{\bf s}}^t + {{\bf s}}^u,\ \text{with}\ t\neq u. \end{array} \right.$$ \[prop:ell-corchete\] For any $D,E\in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ we have $ {\ell}([D,E]) \geq {\ell}(D) + {\ell}(E)$. We may assume $D,E\neq \mathbb{I}$. Let us write $m={\ell}(D)=\ell(D^*)$, $n={\ell}(E)={\ell}(E^*)$. We have $D_\beta=D^*_\beta = 0$ whenever $0<|\beta| < m$ and $E_\gamma=E^*_\gamma = 0$ whenever $0<|\gamma| < n$. Let $\alpha\in\Delta$ be with $0<|\alpha|<m+n$. If $|\alpha|< m$ or $|\alpha|<n$ it is clear that $[D,E]_\alpha=0$. Assume that $m,n\leq |\alpha| < m+n$: $$\begin{aligned} & \displaystyle [D,E]_\alpha = \sum_{\scriptscriptstyle \beta + \gamma + \lambda + \mu =\alpha} D_\beta {{\scriptstyle \,\circ\,}}E_\gamma \, D^*_\lambda \, E^*_\mu = \sum_{\scriptscriptstyle \gamma + \mu =\alpha} \ E_\gamma \, E^*_\mu + &\\ & \displaystyle \sum_{\substack{\scriptscriptstyle \beta + \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\beta+\lambda|> 0}} D_\beta \, E_\gamma \, D^*_\lambda \, E^*_\mu = 0 + \sum_{\substack{\scriptscriptstyle \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\lambda|>0}} E_\gamma \, D^*_\lambda \, E^*_\mu + \sum_{\substack{\scriptscriptstyle \beta +\gamma + \mu =\alpha\\ \scriptscriptstyle |\beta|>0}} D_\beta\, E_\gamma \, E^*_\mu + &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle \beta + \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\beta|, |\lambda| >0 }} D_\beta \, E_\gamma \, D^*_\lambda \, E^*_\mu = \sum_{\substack{\scriptscriptstyle \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\lambda|\geq m}} E_\gamma \, D^*_\lambda \, E^*_\mu + \sum_{\substack{\scriptscriptstyle \beta +\gamma + \mu =\alpha\\ \scriptscriptstyle |\beta|\geq m}} D_\beta\, E_\gamma \, E^*_\mu + &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle \beta + \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\beta|, |\lambda| \geq m }} D_\beta \, E_\gamma \, D^*_\lambda \, E^*_\mu = D^*_\alpha+ \sum_{\substack{\scriptscriptstyle \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\lambda|\geq m, |\gamma+\mu|>0}} E_\gamma \, D^*_\lambda \, E^*_\mu + D_\alpha+ &\\ &\displaystyle \sum_{\substack{\scriptscriptstyle \beta + \mu =\alpha\\ \scriptscriptstyle |\beta|\geq m\\ \scriptscriptstyle |\gamma + \mu|>0 }} D_\beta\, E_\gamma \, E^*_\mu + \sum_{\substack{\scriptscriptstyle \beta + \lambda =\alpha\\ \scriptscriptstyle |\beta|, |\lambda| \geq m }} D_\beta \, D^*_\lambda + \sum_{\substack{\scriptscriptstyle \beta + \gamma + \lambda + \mu =\alpha\\ \scriptscriptstyle |\beta|, |\lambda| \geq m\\ \scriptscriptstyle |\gamma + \mu|>0 }} D_\beta \, E_\gamma \, D^*_\lambda \, E^*_\mu = &\\ &\displaystyle D^*_\alpha+ 0 + D_\alpha + 0 + \sum_{\substack{\scriptscriptstyle \beta + \lambda =\alpha\\ \scriptscriptstyle |\beta|, |\lambda| > 0 }} D_\beta \, D^*_\lambda + 0 = \sum_{\scriptscriptstyle \beta + \lambda =\alpha} D_\beta \, D^*_\lambda = 0.\end{aligned}$$ So, $ {\ell}([D,E]) \geq {\ell}(D) + {\ell}(E)$. Assume that $\Delta$ is bounded and let $m$ be the $\max$ of $|\alpha|$ with $\alpha\in\Delta$. Then, the group $\operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ is nilpotent of nilpotent class $\leq m$, where a central series is[^6] $$\{\mathbb{I}\} = \{E|\ {\ell}(E) >m \} \vartriangleleft \{E|\ {\ell}(E) \geq m\} \vartriangleleft \cdots \vartriangleleft \{E|\ {\ell}(E) \geq 1\} = \operatorname{HS}^{{{\bf s}}}_k(A;\Delta).$$ \[prop:expresion-inverse-multi-HS\] For each $D\in \operatorname{HS}_k^{{\bf s}}(A;\Delta)$, its inverse $D^*$ is given by $D^*_0={{\rm Id}}$ and $$D^*_\alpha = \sum_{d=1}^{|\alpha|} (-1)^d \sum_{\alpha^\bullet \in \operatorname{\mathcal{P}}(\alpha,d)} D_{\alpha^1}{{\scriptstyle \,\circ\,}}\cdots {{\scriptstyle \,\circ\,}}D_{\alpha^d},\quad \alpha\in\Delta.$$ Moreover, $\sigma_{|\alpha|}(D^*_\alpha) = (-1)^{|\alpha|} \sigma_{|\alpha|}(D_\alpha)$. The first assertion is a straightforward consequence of Lemma \[lemma:unit-psr\]. For the second assertion, first we have $D^*_\alpha = - D_\alpha$ for all $\alpha$ with $|\alpha|=1$, and if we denote by $\mathbf{-1} \in A^{{\bf s}}$ the constant family $-1$ and $E= D {{\scriptstyle \,\circ\,}}((\mathbf{-1}) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)$, we have ${\ell}(E)> 1$. So, $D^* = ((\mathbf{-1}) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}E^*$ and $$D^*_\alpha = \sum_{\scriptscriptstyle \beta+\gamma=\alpha} (-1)^{|\beta|} D_\beta E^*_\gamma = (-1)^{|\alpha|} D_\alpha + \sum_{\substack{\scriptscriptstyle \beta+\gamma=\alpha\\ \scriptscriptstyle |\gamma|>0}} (-1)^{|\beta|} D_\beta E^*_\gamma.$$ From Proposition \[prop:ell-new\], we know that $E^*_\gamma$ is a differential operator of order strictly less than $|\gamma|$ and so $\sigma_{|\alpha|}(D^*_\alpha) = (-1)^{|\alpha|} \sigma_{|\alpha|}(D_\alpha)$. The action of substitution maps on HS-derivations ================================================= In this section, $k$ will be a commutative ring, $A$ a commutative $k$-algebra, $R=\operatorname{End}_k(A)$, ${{\bf s}}$, ${{\bf t}}$ sets and $\Delta \subset {\mathbb{N}}^{({{\bf s}})}$, $\nabla \subset {\mathbb{N}}^{({{\bf t}})}$ non-empty co-ideals. We are going to extend the operation $(a,D) \in A^{{\bf s}}\times \operatorname{HS}^{{{\bf s}}}_k(A;\Delta) \mapsto a{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D \in \operatorname{HS}^{{{\bf s}}}_k(A;\Delta)$ (see Definition \[defi:sbullet-0\]) by means of the constructions in section \[sec:action-substi\]. \[prop:equiv-action-subs-HS\] For any substitution map $\varphi:A[[{{\bf s}}]]_\Delta \to A[[{{\bf t}}]]_\nabla$, we have: 1. $\varphi_*\left(\operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta)\right) \subset \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf t}}]]_\nabla)$, 2. $\varphi_R \left( \operatorname{HS}^{{\bf s}}_k(A;\Delta)\right) \subset \operatorname{HS}^{{\bf t}}_k(A;\nabla)$, 3. $ \overline{\varphi_*} \left( \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta) \right) \subset \operatorname{Aut}_{k[[{{\bf t}}]]_\nabla-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf t}}]]_\nabla)$. By using diagram (\[eq:diag-funda\]) and (\[eq:HS-funda\]), it is enough to prove the first inclusion, but if $f\in \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta)$, it is clear that $\varphi_*(f)= \varphi {{\scriptstyle \,\circ\,}}f: A \to A[[{{\bf t}}]]_\nabla$ is a $k$-algebra map. Moreover, since $\varphi({\mathfrak{t}}^A_0({{\bf s}})/\Delta_A) \subset {\mathfrak{t}}^A_0({{\bf t}})/\nabla_A$ (see \[nume:operaciones-con-substitutions\]) and $f(a) \equiv a \!\mod {\mathfrak{t}}^A_0({{\bf s}})/\Delta_A$ for all $a\in A$, we deduce that $\varphi(f(a)) \equiv \varphi(a) \!\mod {\mathfrak{t}}^A_0({{\bf t}})/\nabla_A$ for all $a\in A$, but $\varphi$ is an $A$-algebra map and $\varphi(a) = a$. So $\varphi_*(f)\in\operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf t}}]]_\nabla)$. As a consequence of the above proposition and diagram (\[eq:diag-funda\]) we have a commutative diagram: $$\label{eq:diag-funda-HS} \begin{CD} \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf s}}]]_\Delta) @<{\sim}<{\Phi_D \mapsfrom D}< \operatorname{HS}^{{\bf s}}_k(A;\Delta) @>{\sim}>> \operatorname{Aut}_{k[[{{\bf s}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\Delta)\\ @V{\varphi_*}VV @V{\varphi_R}VV @VV{\overline{\varphi_*}}V\\ \operatorname{Hom}_{k-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A,A[[{{\bf t}}]]_\nabla) @<{\sim}<{\Phi_D \mapsfrom D}< \operatorname{HS}^{{\bf t}}_k(A;\nabla) @>{\sim}>> \operatorname{Aut}_{k[[{{\bf t}}]]_\nabla-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf t}}]]_\nabla). \end{CD}$$ The inclusion 2) in Proposition \[prop:equiv-action-subs-HS\] can be rephrased by saying that for any substitution map $\varphi:A[[{{\bf s}}]]_\Delta \to A[[{{\bf t}}]]_\nabla$ and for any HS-derivation $D\in \operatorname{HS}^{{\bf s}}_k(A;\Delta)$ we have $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\in \operatorname{HS}^{{\bf t}}_k(A;\nabla)$ (see \[nume:def-sbullet\]). Moreover $ \Phi_{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D} = \varphi {{\scriptstyle \,\circ\,}}\Phi_D$. It is clear that for any co-ideals $\Delta' \subset \Delta$ and $\nabla' \subset \nabla$ with $\varphi \left( \Delta'_A/\Delta_A \right) \subset \nabla'_A/\nabla_A$ we have $$\label{eq:trunca_bullet} \tau_{\nabla \nabla'}(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = \varphi' {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{\Delta \Delta'}(D),$$ where $\varphi': A[[{{\bf s}}]]_{\Delta'} \to A[[{{\bf t}}]]_{\nabla'}$ is the substitution map induced by $\varphi$. Let us notice that any $a\in A^{{\bf s}}$ gives rise to a substitution map $\varphi: A[[{{\bf s}}]]_\Delta \to A[[{{\bf s}}]]_\Delta$ given by $\varphi(s)= a_s s$ for all $s\in{{\bf s}}$, and one has $a {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. [[**.  **]{}]{}\[nume:def-sbullet-HS\] Let $\varphi\in\operatorname{\mathcal{S}}_A({{\bf s}},{{\bf t}};\nabla,\Delta)$, $\psi\in\operatorname{\mathcal{S}}_A({{\bf t}},{{\bf u}};\Delta,\Omega)$ be substitution maps and $D,D'\in\operatorname{HS}_k^{{\bf s}}(A;\nabla)$ HS-derivations. From \[nume:def-sbullet\] we deduce the following properties: -) If we denote $E:= \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D \in\operatorname{HS}_k^{{\bf t}}(A;\Delta)$, we have $$\label{eq:expression_phi_D} E_0={{\rm Id}},\quad E_e = \sum_{\substack{\scriptstyle \alpha\in\nabla\\ \scriptstyle |\alpha|\leq |e| }} {\bf C}_e(\varphi,\alpha) D_\alpha,\quad \forall e\in \Delta.$$ -) If $\varphi$ has , then $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D{{\scriptstyle \,\circ\,}}D') = (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D')$. The general case will be treated in Proposition \[prop:varphi-D-main\]. -) If $\varphi = \mathbf{0}$ is the trivial substitution map or if $D=\mathbb{I}$, then $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = \mathbb{I}$. -) $ \psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = (\psi {{\scriptstyle \,\circ\,}}\varphi) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. We recall that a HS-derivation $D\in\operatorname{HS}_k(A)$ is called [*iterative*]{} (see [@mat_86 pg. 209]) if $$D_i {{\scriptstyle \,\circ\,}}D_j = \binom{i+j}{i} D_{i+j}\quad \forall i,j\geq 0.$$ This notion makes sense for ${{\bf s}}$-variate HS-derivations of any length. Actually, iterativity may be understood through the action of substitution maps. Namely, if we denote by $\iota,\iota': s \hookrightarrow s \sqcup s$ the two canonical inclusions and $\iota+\iota': A[[{{\bf s}}]] \to A[[{{\bf s}}\sqcup {{\bf s}}]]$ is the substitution map determined by $$(\iota+\iota') (s) = \iota(s) + \iota'(s),\quad \forall s\in {{\bf s}},$$ then a HS-derivation $D\in \operatorname{HS}^{{\bf s}}_k(A)$ is iterative if and only if $$(\iota+\iota'){{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = (\iota{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}(\iota' {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D).$$ A similar remark applies for any formal group law instead of $\iota+\iota'$ (cf. [@hoff_kow_2015]). \[prop:varphi-D-main\] Let $\varphi:A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ be a substitution map. Then, the following assertions hold: 1. For each $D\in\operatorname{HS}_k^{{\bf s}}(A;\nabla)$ there is a unique substitution map $\varphi^D: A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ such that $\left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\varphi^D = \varphi {{\scriptstyle \,\circ\,}}\widetilde{D}$. Moreover, $\left(\varphi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)^* = \varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D^*$ and $\varphi^{\mathbb{I}} = \varphi$. 2. For each $D,E\in\operatorname{HS}_k^{{\bf s}}(A;\nabla)$, we have $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D {{\scriptstyle \,\circ\,}}E) = (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}(\varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}E)$ and $\left(\varphi^D\right)^E = \varphi^{D {{\scriptstyle \,\circ\,}}E}$. In particular, $\left(\varphi^D\right)^{D^*} = \varphi$. 3. If $\psi$ is another composable substitution map, then $(\varphi {{\scriptstyle \,\circ\,}}\psi)^D = \varphi^{\psi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D} {{\scriptstyle \,\circ\,}}\psi^D$. 4. $\tau_{n}(\varphi^D) = \tau_{n}(\varphi)^{\tau_{n}(D)}$, for all $n\geq 1$. 5. If $\varphi$ has constant coefficients then $\varphi^D = \varphi$. \(i) We know that $$\widetilde{D}\in \operatorname{Aut}_{k[[{{\bf s}}]]_\nabla-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf s}}]]_\nabla)\quad \text{and}\quad \widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\in \operatorname{Aut}_{k[[{{\bf t}}]]_\Delta-\text{\rm alg}}^{{\scriptstyle \,\circ\,}}(A[[{{\bf t}}]]_\Delta).$$ The only thing to prove is that $$\varphi^D := \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right)^{-1} {{\scriptstyle \,\circ\,}}\varphi {{\scriptstyle \,\circ\,}}\widetilde{D}$$ is a substitution map $A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ (see Definition \[def:substitution\_maps\]). Let start by proving that $\varphi^D$ is an $A$-algebra map. Let us write $E=\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. For each $a\in A$ we have $$\begin{aligned} & \varphi^D (a) = \widetilde{E}^{-1} \left( \varphi \left( \widetilde{D}(a)\right)\right) = \widetilde{E}^{-1} \left( \varphi \left( \Phi_D(a)\right)\right) = &\\ & \widetilde{E}^{-1} \left( (\varphi {{\scriptstyle \,\circ\,}}\Phi_D)(a))\right) = \widetilde{E}^{-1} \left( \Phi_{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}(a)\right) = \widetilde{E}^{-1} \left( \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right)(a)\right) = a,\end{aligned}$$ and so $\varphi^D$ is $A$-linear. The continuity of $\varphi^D$ is clear, since it is the composition of continuous maps. For each $s\in{{\bf s}}$, let us write $$\varphi(s) = \sum_{\substack{\scriptscriptstyle \beta \in \Delta\\ \scriptscriptstyle |\beta| > 0}} c^s_\beta {{\bf t}}^\beta.$$ Since $\varphi$ is a substitution map, property (\[eq:cond-fini-c\]) holds: $$\#\{s\in {{\bf s}}\ |\ c^s_\beta \neq 0\} < \infty\quad\quad \text{for all\ }\ \beta \in \Delta.$$ We have $$\varphi^D (s) = \widetilde{E^*} \left( \varphi ( \widetilde{D}(s))\right) = \widetilde{E^*} \left( \varphi (s)\right) = \sum_{\scriptscriptstyle \beta \in \Delta} \left( \sum_{\scriptscriptstyle \alpha + \gamma=\beta} E^*_\alpha (c^s_\gamma) \right) {{\bf t}}^\beta = \sum_{\scriptscriptstyle \beta \in \Delta} d^s_\beta {{\bf t}}^\beta$$ with $ d^s_\beta = \sum_{\scriptscriptstyle \alpha + \gamma=\beta} E^*_\alpha (c^s_\gamma)$. So, for each $\beta \in \Delta$ we have $$\{s\in {{\bf s}}\ |\ c^s_\beta \neq 0\} \subset \bigcup_{\gamma \leq \beta} \{s\in {{\bf s}}\ |\ c^s_\gamma \neq 0\}$$ and $\varphi^D$ satisfies property (\[eq:cond-fini-c\]) too. We conclude that $\varphi^D$ is a substitution map, and obviously it is the only one such that $\left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\varphi^D = \varphi {{\scriptstyle \,\circ\,}}\widetilde{D}$. From there, we have $$\varphi^D {{\scriptstyle \,\circ\,}}\widetilde{D^*} = \varphi^D {{\scriptstyle \,\circ\,}}\widetilde{D}^{-1} = \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right)^{-1} {{\scriptstyle \,\circ\,}}\varphi = \widetilde{\left(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)^*} {{\scriptstyle \,\circ\,}}\varphi,$$ and taking restrictions to $A$ we obtain $ \varphi^D {{\scriptstyle \,\circ\,}}\Phi_{D^*} = \Phi_{\left(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)^*} $ and so $\varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D^* = \left(\varphi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)^*$. On the other hand, it is clear that if $D=\mathbb{I}$, then $\varphi^\mathbb{I} = \varphi$ and if $\varphi=\mathbf{0}$, $\mathbf{0}^D=\mathbf{0}$. \(ii) In order to prove the first equality, we need to prove the equality $ \widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D {{\scriptstyle \,\circ\,}}E)} = \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\left(\widetilde{\varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}E}\right) $. For this it is enough to prove the equality after restriction to $A$, but $$\left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D {{\scriptstyle \,\circ\,}}E)}\right)|_A = \Phi_{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D {{\scriptstyle \,\circ\,}}E)} = \varphi {{\scriptstyle \,\circ\,}}\Phi_{D {{\scriptstyle \,\circ\,}}E} = \varphi {{\scriptstyle \,\circ\,}}\widetilde{D} {{\scriptstyle \,\circ\,}}\Phi_E,$$ $$\left( \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\left(\widetilde{\varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}E}\right) \right)|_A = \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\Phi_{\varphi^D {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}E} = \left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\varphi^D {{\scriptstyle \,\circ\,}}\Phi_E$$ and both are equal by (i). For the second equality, we have $\left(\varphi^D\right)^{D^*} = \varphi^{\mathbb{I}}= \varphi$. \(iii) Since $$\begin{aligned} & \widetilde{\left((\varphi {{\scriptstyle \,\circ\,}}\psi) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)} {{\scriptstyle \,\circ\,}}\left( \varphi^{\psi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D} {{\scriptstyle \,\circ\,}}\psi^D \right) = \widetilde{\left(\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(\psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)\right)} {{\scriptstyle \,\circ\,}}\varphi^{\psi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D} {{\scriptstyle \,\circ\,}}\psi^D = &\\ & \varphi {{\scriptstyle \,\circ\,}}\left(\widetilde{\psi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\psi^D = \varphi {{\scriptstyle \,\circ\,}}\psi {{\scriptstyle \,\circ\,}}\widetilde{D},\end{aligned}$$ we deduce that $(\varphi {{\scriptstyle \,\circ\,}}\psi)^D = \varphi^{\psi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D} {{\scriptstyle \,\circ\,}}\psi^D$ from the uniqueness in (i). Part (iv) is also a consequence of the uniqueness property in (i). \(v) Let us assume that $\varphi$ has constant coefficients. We know from Lemma \[lemma:const\_coeff\_subst\_map\] that $ \langle \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D, \varphi(a) \rangle = \varphi \left( \langle D, a\rangle \right)$ for all $a\in A[[{{\bf s}}]]_\nabla$, and so $\left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\varphi = \varphi {{\scriptstyle \,\circ\,}}\widetilde{D}$. Hence, by the uniqueness property in (i) we deduce that $\varphi^D = \varphi$. The following proposition gives a recursive formula to obtain $\varphi^D$ from $\varphi$. \[prop:varphi-D\] With the notations of Proposition \[prop:varphi-D-main\], we have $${\bf C}_e(\varphi,f+\nu) = \sum_{\substack{\scriptscriptstyle\beta+\gamma=e\\ \scriptscriptstyle |f+g|\leq |\beta|,|\nu|\leq |\gamma| }} {\bf C}_\beta(\varphi,f+g) D_g({\bf C}_\gamma(\varphi^D,\nu))$$ for all $e\in \Delta$ and for all $f,\nu\in\nabla$ with $|f+\nu|\leq |e|$. In particular, we have the following recursive formula $${\bf C}_e(\varphi^D,\nu) := {\bf C}_e(\varphi,\nu) - \sum_{\substack{\scriptscriptstyle\beta+\gamma=e\\ \scriptscriptstyle |g|\leq |\beta|, |\nu|\leq |\gamma| < |e| }} {\bf C}_\beta(\varphi,g) D_g({\bf C}_\gamma(\varphi^D,\nu)).$$ for $e\in \Delta$, $\nu\in\nabla$ with $|e|\geq 1$ and $|\nu|\leq |e|$, starting with ${\bf C}_0(\varphi^D,0) = 1$. First, the case $f=0$ easily comes from the equality $$\sum_{\substack{\scriptscriptstyle e \in\Delta \\ \scriptscriptstyle |\nu|\leq |e|}} {\bf C}_e(\varphi,\nu) {{\bf t}}^e = \varphi ({{\bf s}}^\nu) = (\varphi {{\scriptstyle \,\circ\,}}\widetilde{D})({{\bf s}}^\nu) = \left(\left(\widetilde{\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\varphi^D\right) ({{\bf s}}^\nu) \quad \forall \nu\in\nabla.$$ For arbitrary $f$ one has to use Proposition \[prop:identidad-Cs\]. Details are left to the reader. The proof of the following corollary is a consequence of Lemma \[lema:dual-leibniz-HS\]. \[cor:aux-phi-D\] Under the hypotheses of Proposition \[prop:varphi-D-main\], the following identity holds for each $e\in\Delta$ $$\left(\varphi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)^*_e = \sum_{ |\mu+\nu|\leq |e| } D^*_\mu \cdot D_\nu \left( {\bf C}_e(\varphi^D,\mu+\nu) \right).$$ \[prop:D-circ-varphi\] Let $D\in\operatorname{HS}^{{\bf t}}_k(A;\Delta)$ be a HS-derivation and $\varphi:A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ a substitution map. Then, the following identity holds: $$\widetilde{D} {{\scriptstyle \,\circ\,}}\varphi = \left(D(\varphi) \otimes \pi\right) {{\scriptstyle \,\circ\,}}\left(\widetilde{\kappa {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D}\right) {{\scriptstyle \,\circ\,}}\iota,$$ where: - $D(\varphi): A[[{{\bf s}}]]_\nabla \to A[[{{\bf t}}]]_\Delta$ is the substitution map determined by $D(\varphi)(s)= \widetilde{D}(\varphi(s))$ for all $s\in{{\bf s}}$. - $\pi: A[[{{\bf t}}]]_\Delta \rightarrow A$ is the augmentation, or equivalently, the substitution map[^7] given by $\pi (t) = 0$ for all $t\in{{\bf t}}$. - $\iota: A[[{{\bf s}}]]_\nabla \to A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$ and $\kappa: A[[{{\bf t}}]]_\Delta \to A[[{{\bf s}}\sqcup {{\bf t}}]]_{\nabla \times \Delta}$ are the combinatorial substitution maps determined by the inclusions ${{\bf s}}\hookrightarrow {{\bf s}}\sqcup {{\bf t}}$ and ${{\bf t}}\hookrightarrow {{\bf s}}\sqcup {{\bf t}}$, respectively. It is enough to check that both maps coincide on any $a\in A$ and on any $s\in{{\bf s}}$. Details are left to the reader. Let us notice that with the notations of Propositions \[prop:varphi-D-main\] and \[prop:D-circ-varphi\], we have $ \varphi^D = (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)^*(\varphi)$. The following proposition will not be used in this paper and will be stated without proof. For any HS-derivation $D\in \operatorname{HS}_k^{{\bf s}}(A;\nabla)$ and any substitution map $\varphi \in \operatorname{\mathcal{S}}({{\bf t}},{{\bf u}};\Delta,\Omega) $, there exists a substitution map $D\star \varphi \in \operatorname{\mathcal{S}}({{\bf s}}\sqcup {{\bf t}},{{\bf s}}\sqcup {{\bf u}};\nabla \times \Delta, \nabla \times \Omega)$ such that for each HS-derivation $E\in \operatorname{HS}_k^{{\bf t}}(A;\Delta)$ we have: $$D {{\scriptstyle \,\boxtimes\,}}(\varphi{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}E) = (D\star \varphi){{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}(D{{\scriptstyle \,\boxtimes\,}}E).$$ Generating HS-derivations ========================= In this section we show how the action of substitution maps allows us to express any HS-derivation in terms of a fixed one under some natural hypotheses. We will be concerned with $({{\bf s}},{\mathfrak{t}}_m({{\bf s}}))$-variate HS-derivations, where ${\mathfrak{t}}_m({{\bf s}}) = \{\alpha\in{\mathbb{N}}^{({{\bf s}})}\ |\ |\alpha|\leq m\}$. To simplify we will write $A[[{{\bf s}}]]_m := A[[{{\bf s}}]]_{{\mathfrak{t}}_m({{\bf s}})}$ and $\operatorname{HS}^{{\bf s}}_k(A;m):= \operatorname{HS}^{{\bf s}}_k(A;{\mathfrak{t}}_m({{\bf s}}))$ for any integer $m\geq 1$, and $\operatorname{HS}^{{\bf s}}_k(A;\infty) := \operatorname{HS}^{{\bf s}}_k(A)$. For $m\geq n\geq 1$ we will denote $\tau_{mn}: \operatorname{HS}^{{\bf s}}_k(A;m) \to \operatorname{HS}^{{\bf s}}_k(A;n)$ the truncation map. Assume that $m\geq 1$ is an integer and let $\varphi: A[[{{\bf s}}]]_m \to A[[{{\bf t}}]]_m$ be a substitution map. Let us write $$\varphi(s)=c^s = \sum_{\substack{\scriptstyle \beta\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptstyle 0<|\beta|\leq m }} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}_0({{\bf t}})/{\mathfrak{t}}_m({{\bf t}}) \subset A[[{{\bf t}}]]_m,\quad s\in {{\bf s}}$$ and let us denote by $\varphi_m, \varphi_{<m}:A[[{{\bf s}}]]_m \to A[[{{\bf t}}]]_m$ the substitution maps determined by $$\begin{aligned} & \displaystyle \varphi_m(s) = c_m^s:= \sum_{\substack{\scriptstyle \beta\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptstyle |\beta|= m }} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}_0({{\bf t}})/{\mathfrak{t}}_m({{\bf t}}) \in A[[{{\bf t}}]]_m,\quad s\in {{\bf s}},&\\ & \displaystyle \varphi_{<m}(s) = c_{<m}^s:= \sum_{\substack{\scriptstyle \beta\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptstyle 0<|\beta|< m }} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}_0({{\bf t}})/{\mathfrak{t}}_m({{\bf t}}) \in A[[{{\bf t}}]]_m, \quad s\in {{\bf s}}.&\end{aligned}$$ We have $c^s = c_m^s + c_{<m}^s$ and so $\varphi = \varphi_m + \varphi_{<m}$ (see \[nume:operaciones-con-substitutions\]). \[prop:previa\_main\] With the above notations, for any HS-derivation $D\in \operatorname{HS}^{{\bf s}}_k(A;m)$ the following properties hold: 1. $\left(\varphi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)_{e} = 0$ for $0<|e| < m$ and $\left(\varphi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\right)_{e} = \sum_{t\in {{\bf s}}} c^t_e D_{{{\bf s}}^t}$ for $|e| = m$, where the ${{\bf s}}^t$ are the elements of the canonical basis of ${\mathbb{N}}^{({{\bf s}})}$. 2. $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = (\varphi_{m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}(\varphi_{<m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = (\varphi_{<m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D){{\scriptstyle \,\circ\,}}(\varphi_{m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)$. \(1) Let us denote $E'= \varphi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. Since $\tau_{m,m-1}(E')$ coincides with\ $\tau_{m,m-1}(\varphi_m) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{m,m-1}(D)$ (see (\[eq:trunca\_bullet\])) and $\tau_{m,m-1}(\varphi_m)$ is the trivial substitution map, we deduce that $\tau_{m,m-1}(E') = \mathbb{I}$, i.e. $E_e=0$ whenever $0<|e| < m$. From (\[eq:expression\_phi\_D\]) and (\[eq:explicit-C\_e(varphi,alpha)\]), for $|e|>0$ we have $E'_e = \sum_{\scriptscriptstyle 0<|\alpha|\leq |e|} {\bf C}_e(\varphi_m,\alpha) D_\alpha$, with $${\bf C}_e(\varphi_m,\alpha)= \sum_{\scriptscriptstyle \mathcal{f}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\alpha)} C_{\mathcal{f}^{\bullet\bullet}} \quad \text{for\ }\ |\alpha| \leq |e|,\quad C_{\mathcal{f}^{\bullet\bullet}} = \prod_{\scriptscriptstyle s\in \operatorname{\rm supp}\alpha} \prod_{\scriptstyle r=1}^{\scriptstyle \alpha_s} (c^s_m)_{\mathcal{f}^{sr}}.$$ Assume now that $|e|=m$, $1 < |\alpha|\leq m$ and let $\mathcal{f}^{\bullet\bullet} \in \operatorname{\mathcal{P}}(e,\alpha)$. Since $$\sum_{\scriptscriptstyle s\in \operatorname{\rm supp}\alpha} \sum_{\scriptscriptstyle r=1}^{\scriptscriptstyle \alpha_s} \mathcal{f}^{sr} = e,$$ we deduce that $|\mathcal{f}^{sr}| < |e|=m$ for all $s,r$ and so $(c^s_m)_{\mathcal{f}^{sr}}=0$ and $C_{\mathcal{f}^{\bullet\bullet}}=0$. Consequently, ${\bf C}_e(\varphi_m,\alpha)=0$. If $|\alpha|= 1$, then $\alpha$ must be an element ${{\bf s}}^t$ of the canonical basis of ${\mathbb{N}}^{({{\bf s}})}$ and from Lemma \[lemma:behavior\_C\_alpha\], (1), we know that ${\bf C}_e(\varphi_m,{{\bf s}}^t)=(c^t_m)_e$. We conclude that $$E'_e =\cdots = \sum_{\scriptscriptstyle t\in {{\bf s}}} {\bf C}_e(\varphi_m,{{\bf s}}^t) D_{{{\bf s}}^t} = \sum_{\scriptscriptstyle t\in {{\bf s}}} (c^t_m)_e D_{{{\bf s}}^t}= \sum_{\scriptscriptstyle t\in {{\bf s}}} c^t_e D_{{{\bf s}}^t}.$$ \(2) Let us write $E=\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$, $E'= \varphi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$ and $E''=\varphi_{<m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. We have $$\begin{aligned} &\tau_{m,m-1}(E) = \tau_{m,m-1}(\varphi) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{m,m-1}(D) = &\\ & \tau_{m,m-1}(\varphi_{<m}) {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{m,m-1}(D) = \tau_{m,m-1}(E''). \end{aligned}$$ By property (1), we know that $\tau_{m,m-1}(E')$ is the identity and we deduce that $\tau_{m,m-1}(E) = \tau_{m,m-1}(E'{{\scriptstyle \,\circ\,}}E'') = \tau_{m,m-1}(E''{{\scriptstyle \,\circ\,}}E')$. So $ E_e = (E'{{\scriptstyle \,\circ\,}}E'')_e = (E''{{\scriptstyle \,\circ\,}}E')_e $ for $|e|<m$. Now, let $e\in {\mathbb{N}}^{({{\bf t}})}$ be with $|e|=m$. By using again that $\tau_{m,m-1}(E')$ is the identity, we have $(E'{{\scriptstyle \,\circ\,}}E'')_e = \dots = E'_e + E''_e = \cdots = (E'' {{\scriptstyle \,\circ\,}}E')_e$, and we conclude that $E'{{\scriptstyle \,\circ\,}}E'' = E'' {{\scriptstyle \,\circ\,}}E'$. On the other hand, from Lemma \[lemma:behavior\_C\_alpha\], (1), we have that ${\bf C}_e(\varphi_{<m},\alpha) =0$ whenever $|\alpha|=1$, and one can see that ${\bf C}_e(\varphi,\alpha) = {\bf C}_e(\varphi_{<m},\alpha)$ whenever that $2\leq |\alpha|\leq |e|$. So: $$\begin{aligned} &\displaystyle E_e= \sum_{\scriptscriptstyle 1\leq |\alpha|\leq m} {\bf C}_e(\varphi,\alpha) D_\alpha = \sum_{\scriptscriptstyle |\alpha|=1} {\bf C}_e(\varphi,\alpha) D_\alpha + \sum_{\scriptscriptstyle 2\leq |\alpha|\leq m} {\bf C}_e(\varphi,\alpha) D_\alpha=&\\ &\displaystyle \sum_{\scriptscriptstyle t\in {{\bf s}}} c^t_e D_{{{\bf s}}^t} + \sum_{\scriptscriptstyle 2\leq |\alpha|\leq m} {\bf C}_e(\varphi_{<m},\alpha) D_\alpha= E'_e + \sum_{\scriptscriptstyle 1\leq |\alpha|\leq m} {\bf C}_e(\varphi_{<m},\alpha) D_\alpha= E'_e + E''_e\end{aligned}$$ and $E=E'{{\scriptstyle \,\circ\,}}E'' = E'' {{\scriptstyle \,\circ\,}}E'$. The following theorem generalizes Theorem 2.8 in [@magda_nar_hs] to the case where $\operatorname{Der}_k(A)$ is not necessarily a finitely generated $A$-module. The use of substitution maps makes its proof more conceptual. \[thm:main\_1\] Let $m\geq 1$ be an integer, or $m=\infty$, and $D\in \operatorname{HS}^{{\bf s}}_k(A;m)$ a ${{\bf s}}$-variate HS-derivation of length $m$ such that $\{D_\alpha, |\alpha|=1\}$ is a system of generators of the $A$-module $\operatorname{Der}_k(A)$. Then, for each set ${{\bf t}}$ and each HS-derivation $G\in\operatorname{HS}^{{\bf t}}_k(A;m)$ there is a substitution map $\varphi:A[[{{\bf s}}]]_m \to A[[{{\bf t}}]]_m$ such that $G = \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. Moreover, if $\{D_\alpha, |\alpha|=1\}$ is a basis of $\operatorname{Der}_k(A)$, $\varphi$ is uniquely determined. For $m$ finite, we will proceed by induction on $m$. For $m=1$ the result is clear. Assume that the result is true for HS-derivations of length $m-1$ and consider a $D\in \operatorname{HS}^{{\bf s}}_k(A;m)$ such that $\{D_\alpha, |\alpha|=1\}$ is a system of generators of the $A$-module $\operatorname{Der}_k(A)$ and a $G\in\operatorname{HS}^{{\bf t}}_k(A;m)$. By the induction hypothesis, there is a substitution map $\varphi': A[[{{\bf s}}]]_{m-1} \to A[[{{\bf t}}]]_{m-1}$, given by $\varphi'(s) = \sum_{\scriptscriptstyle |\beta|\leq m-1} c^s_\beta {{\bf t}}^\beta$, $s\in {{\bf s}}$, and such that $\tau_{m,m-1}(G)= \varphi' {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{m,m-1}(D)$. Let $\varphi'':A[[{{\bf s}}]]_m \to A[[{{\bf u}}]]_m$ be the substitution map lifting $\varphi'$ (i.e. $\tau_{m,m-1}(\varphi'') = \varphi'$) given by $\varphi''(s) = \sum_{\scriptscriptstyle |\beta|\leq m-1} c^s_\beta {{\bf t}}^\beta \in A[[{{\bf t}}]]_m$, $s\in {{\bf s}}$, and consider $F= \varphi'' {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. We obviously have $\tau_{m,m-1}(F)= \tau_{m,m-1}(G)$ and so, for $H=G {{\scriptstyle \,\circ\,}}F^*$, the truncation $\tau_{m,m-1}(H)$ is the identity and $H_e=0$ for $0<|e|<m$. We deduce that each component of $H$ of highest order, $H_e$ with $|e|=m$, must be a $k$-derivation of $A$ and so there is a family $\{c^s_e, s\in {{\bf s}}\}$ of elements of $A$ such that $c^s_e=0$ for all $s$ except a finite number of indices and $H_e= \sum_{\scriptscriptstyle s\in {{\bf s}}} c^{{\bf s}}_e D_{{{\bf s}}^s}$, where $\{{{\bf s}}^s, s\in {{\bf s}}\}$ is the canonical basis of ${\mathbb{N}}^{({{\bf s}})}$. To finish, let us consider the substitution map $\varphi:A[[{{\bf s}}]]_m \to A[[{{\bf t}}]]_m$ given by $\varphi(s) = \sum_{\scriptscriptstyle |\beta|\leq m} c^s_\beta {{\bf t}}^\beta$, $s\in {{\bf s}}$. From Proposition \[prop:previa\_main\] we have $$\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = (\varphi_{m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) {{\scriptstyle \,\circ\,}}(\varphi_{<m} {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = H {{\scriptstyle \,\circ\,}}(\varphi'' {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D) = H {{\scriptstyle \,\circ\,}}F = G.$$ For HS-derivations of infinite length, following the above procedure we can construct $\varphi$ as a projective limit of substitution maps $A[[{{\bf s}}]]_m \to A[[{{\bf t}}]]_m$, $m\geq 1$. Now assume that the set $\{D_\alpha, |\alpha|=1\}$ is linearly independent over $A$ and let us prove that $$\label{eq:uni-varphi} \varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = \psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D\quad \Longrightarrow\quad \varphi = \psi.$$ The infinite length case can be reduced to the finite case since $\varphi = \psi$ if and only if all their finite truncations are equal. For the finite length case, we proceed by induction on the length $m$. Assume that the substitution maps are given by $$\begin{aligned} & \displaystyle \varphi(s) = c^s:= \sum_{\substack{\scriptscriptstyle \beta\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptscriptstyle 0<|\beta|\leq m }} c^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}_0({{\bf t}})/{\mathfrak{t}}_m({{\bf t}}) \subset A[[{{\bf t}}]]_m,\quad s\in {{\bf s}}&\\ & \displaystyle \psi(s) = d^s:= \sum_{\substack{\scriptscriptstyle \beta\in{\mathbb{N}}^{({{\bf t}})}\\ \scriptscriptstyle 0<|\beta|\leq m }} d^s_\beta {{\bf t}}^\beta \in {\mathfrak{n}}_0({{\bf t}})/{\mathfrak{t}}_m({{\bf t}}) \subset A[[{{\bf t}}]]_m,\quad s\in {{\bf s}}. &\end{aligned}$$ If $m=1$, then $\varphi = \varphi_1$ and $\psi = \psi_1$ and for each $e\in{\mathbb{N}}^{({{\bf t}})}$ with $|e|=1$ we have from Proposition \[prop:previa\_main\] $$\sum_{\scriptscriptstyle s\in {{\bf s}}} c^s_e D_{{{\bf s}}^s} = (\varphi_1 {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)_e = (\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)_e = (\psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)_e = (\psi_1 {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D)_e = \sum_{\scriptscriptstyle s\in {{\bf s}}} d^s_e D_{{{\bf s}}^s}$$ and we deduce that $c^s_e= d^s_e$ for all $s\in {{\bf s}}$ and so $\varphi=\psi$. Now assume that (\[eq:uni-varphi\]) is true whenever the length is $m-1$ and take $D, \varphi $ and $\psi$ as before of length $m$ with $\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = \psi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$. By considering $(m-1)$-truncations and using the induction hypothesis we deduce that $\tau_{m,m-1}(\varphi)=\tau_{m,m-1}(\psi)$, or equivalently $\varphi_{<m} = \psi_{<m}$. From Proposition \[prop:previa\_main\] we obtain first that $\varphi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D = \psi_m{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$ and second that for each $e\in{\mathbb{N}}^{({{\bf t}})}$ with $|e|=m$ $$\sum_{\scriptscriptstyle s\in {{\bf s}}} c^s_e D_{{{\bf s}}^s} = \sum_{\scriptscriptstyle s\in {{\bf s}}} d^s_e D_{{{\bf s}}^s}.$$ We conclude that $\varphi_m = \psi_m$ and so $\varphi=\psi$. Now we recall the definition of integrability. (Cf. [@brown_1978; @mat-intder-I]) \[def:HS-integ\] Let $m\geq 1$ be an integer or $m=\infty$ and ${{\bf s}}$ a set. 1. We say that a $k$-derivation $\delta:A\to A$ is [*$m$-integrable*]{} (over $k$) if there is a Hasse–Schmidt derivation $D\in \operatorname{HS}_k(A;m)$ such that $D_1=\delta$. Any such $D$ will be called an [*$m$-integral*]{} of $\delta$. The set of $m$-integrable $k$-derivations of $A$ is denoted by $\operatorname{Ider}_k(A;m)$. We simply say that $\delta$ is [*integrable*]{} if it is $\infty$-integrable and we denote $\operatorname{Ider}_k(A):=\operatorname{Ider}_k(A;\infty)$. 2. We say that a ${{\bf s}}$-variate HS-derivation $D'\in\operatorname{HS}^{{\bf s}}_k(A;n)$, with $1\leq n < m$, is [*$m$-integrable*]{} (over $k$) if there is a ${{\bf s}}$-variate HS-derivation $D\in \operatorname{HS}^{{\bf s}}_k(A;m)$ such that $\tau_{mn}D=D'$. Any such $D$ will be called an [*$m$-integral*]{} of $D'$. The set of $m$-integrable ${{\bf s}}$-variate HS-derivations of $A$ over $k$ of length $n$ is denoted by $\operatorname{IHS}^{{\bf s}}_k(A;n;m)$. We simply say that $D'$ is [*integrable*]{} if it is $\infty$-integrable and we denote $\operatorname{IHS}^{{\bf s}}_k(A;n) := \operatorname{IHS}^{{\bf s}}_k(A;n;\infty)$. Let $m\geq 1$ be an integer or $m=\infty$. The following properties are equivalent: 1. $\operatorname{Ider}_k(A;m) = \operatorname{Der}_k(A)$. 2. $\operatorname{IHS}^{{\bf s}}_k(A;n;m) = \operatorname{HS}^{{\bf s}}_k(A;n)$ for all $n$ with $1\leq n <m$ and all sets ${{\bf s}}$. We only have to prove (1) $\Longrightarrow$ (2). Let $\{\delta_t, t\in {{\bf t}}\}$ be a system of generators of the $A$-module $\operatorname{Der}_k(A)$, and for each $t\in{{\bf t}}$ let $D^t\in \operatorname{HS}_k(A;m)$ be an $m$-integral of $\delta_t$. By considering some total ordering $<$ on ${{\bf t}}$, we can define $D\in \operatorname{HS}^{{\bf t}}_k(A;m)$ as the external product (see Definition \[defi:external-x\]) of the ordered family $\{D^t, t\in{{\bf t}}\}$, i.e. $D_0={{\rm Id}}$ and for each $\alpha \in{\mathbb{N}}^{({{\bf t}})}$, $\alpha\neq 0$, $$D_\alpha = D^{t_1}_{\alpha_{t_1}} {{\scriptstyle \,\circ\,}}\cdots {{\scriptstyle \,\circ\,}}D^{t_e}_{\alpha_{t_e}}\quad \text{with\ }\ \operatorname{\rm supp}\alpha = \{t_1 < \cdots < t_e\}.$$ Let $n$ be an integer with $1\leq n <m$, ${{\bf s}}$ a set and $E\in \operatorname{HS}^{{\bf s}}_k(A;n)$. After Theorem \[thm:main\_1\], there exists a substitution map $\varphi:A[[{{\bf t}}]]_n \to A[[{{\bf s}}]]_n$ such that $E=\varphi {{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}\tau_{mn}(D)$. By considering any substitution map $\varphi':A[[{{\bf t}}]]_m \to A[[{{\bf s}}]]_m$ lifting $\varphi$ we find that $\varphi'{{\hspace{.1em}\scriptstyle\bullet\hspace{.1em}}}D$ is an $m$-integral of $E$ and so $E\in \operatorname{IHS}^{{\bf s}}_k(A;n;m)$. [99]{} W. C. Brown. [On the embedding of derivations of finite rank into derivations of infinite rank](https://projecteuclid.org/euclid.ojm/1200771280). 15 (1978), 381–389. N. Bourbaki, Elements of Mathematics. Algebra II. Chapters 4-7. Springer-Verlag, Berlin, 2003. M. Fernández-Lebrón and L. Narváez-Macarro. [ Hasse-[S]{}chmidt derivations and coefficient fields in positive characteristics](https://doi.org/10.1016/S0021-8693(03)00238-2). 265 (1) (2003), 200–210. ([arXiv:math/0206261](https://arxiv.org/abs/math/0206261)). H. Hasse and F. K. Schmidt. [Noch eine Begründung der Theorie der höheren Differrentialquotienten in einem algebraischen Funktionenkörper einer Unbestimmten](https://doi.org/10.1515/crll.1937.177.215). 177 (1937), 223-239. D. Hoffmann and P. Kowalski. [Integrating Hasse–Schmidt derivations](https://doi.org/10.1016/j.jpaa.2014.05.024). , 219 (2015), 875–896. D. Hoffmann and P. Kowalski. [Existentially closed fields with [$G$]{}-derivations](https://doi.org/10.1112/jlms/jdw009). , 93 (3) (2016), 590–618. H. Matsumura. [Integrable derivations](https://projecteuclid.org/euclid.nmj/1118786907). 87 (1982), 227–245. H. Matsumura. Commutative Ring Theory. Vol. 8 of Cambridge studies in advanced mathematics, Cambridge Univ. Press, Cambidge, 1986. L. Narváez Macarro. [Hasse–Schmidt derivations, divided powers and differential smoothness](https://doi.org/10.5802/aif.2513). 59 (7) (2009), 2979–3014. ([arXiv:0903.0246](https://arxiv.org/abs/0903.0246)). L. Narváez Macarro. [On the modules of [$m$]{}-integrable derivations in non-zero characteristic](https://doi.org/10.1016/j.aim.2012.01.015). 229 (5) (2012), 2712–2740. ([arXiv:1106.1391](https://arxiv.org/abs/1106.1391)). L. Narváez Macarro. [Differential Structures in Commutative Algebra](http://personal.us.es/narvaez/DSCA_course_2014.pdf). Mini-course at the XXIII Brazilian Algebra Meeting, July 27 - August 1, 2014, Maringá, Brazil. L. Narváez Macarro. Rings of differential operators as enveloping algebras of Hasse–Schmidt derivations. ([arXiv:1807.10193](https://arxiv.org/abs/1807.10193)). P. Vojta. Jets via [H]{}asse–[S]{}chmidt derivations. In “Diophantine geometry”, CRM Series, vol. 4, Ed. Norm., Pisa, 2007, 335–361. ([arXiv:math/1201.3594](https://arxiv.org/abs/math/0407113)). [L. Narváez Macarro](http://personal.us.es/narvaez/)\ Departamento de Álgebra & Instituto de Matemáticas (IMUS)\ Universidad de Sevilla, Spain\ email: narvaez@us.es [^1]: Partially supported by MTM2016-75027-P, P12-FQM-2696 and FEDER. [^2]: Pay attention that $(\varphi+\varphi')(r) \neq \varphi(r) + \varphi'(r)$ for arbitrary $r\in A[[{{\bf s}}]]_\nabla$. [^3]: Let us notice that there are canonical continuous isomorphisms of $A$-algebras $A[[{{\bf s}}\sqcup {{\bf u}}]]_{\nabla \times \nabla'} \simeq A[[{{\bf s}}]]_\nabla \widehat{\otimes}_A A[[{{\bf u}}]]_{\nabla'}$, $A[[{{\bf s}}\sqcup {{\bf u}}]]_{\Delta \times \Delta'} \simeq A[[{{\bf s}}]]_\Delta \widehat{\otimes}_A A[[{{\bf u}}]]_{\Delta'}$. [^4]: This terminology is used for instance in [@mat_86]. [^5]: These HS-derivations are called of length $m$ in [@nar_2012]. [^6]: Let us notice that $\{E\in\operatorname{HS}^{{\bf s}}_k(A;\Delta)\ |\ {\ell}(E) > r\} = \ker \tau_{\Delta,\Delta_r}$. [^7]: The map $\pi$ can be also understood as the truncation $\tau_{\Delta,\{0\}}: A[[{{\bf t}}]]_\Delta \rightarrow A[[{{\bf t}}]]_{\{0\}} =A $.

--- address: | Institute of Physics, AS CR, Na Slovance 2, Praha 8, 182 21, Czech Republic\ E-mail: ksedlak@fzu.cz\ [On behalf of the H1 and ZEUS Collaborations]{} author: - 'K. SEDLÁK [^1]' title: 'STRUCTURE OF VIRTUAL PHOTONS AT HERA [^2]' --- Dijet Production at HERA {#uvod} ======================== The production of dijet events at HERA is dominated by processes, in which a virtual photon, radiated from the electron, interacts with a parton in the proton. In the region of photon virtuality $Q^2 \gg \Lambda_{\mathrm QCD}^2$, hard collisions of the photons do not necessitate the introduction of the concept of the resolved photon (as for the real photon) and the process can in principle be described by the direct photon contribution alone, in which the photon interacts as a whole with partons from the proton. The analyses presented here explore the region $\Lambda_{\mathrm QCD}^2 \ll Q^2$ $E_t^2$, where different theoretical approaches can be used to take into account higher order corrections. based on the DGLAP evolution equations and parton showers allows the effects of transversally and also longitudinally polarized resolved photon interactions to be studied [@chyla; @Chyla:2000hp]. Since the cross section for longitudinal photons vanishes in the photoproduction regime due to gauge invariance, and the concept of a resolved photon breaks down for $Q^2 > E_t^2$, the only evidence for this phenomena can be observed in the region $0 < Q^2 \ll E_t^2$. The main difference between the longitudinal ($\gamma^*_L$) and transverse ($\gamma^*_T$) virtual photon arises from the dependence of the respective fluxes on $y$. While for $y\rightarrow 0$, both transverse and longitudinal fluxes are approximately same, the longitudinal flux vanishes for $y\rightarrow 1$. Also the dependence of the point-like (i.e. perturbatively calculable) part of the photon Parton Distribution Functions (PDF) on [$Q^2$]{}and $E_t^2$ differs – while the $\gamma^*_T$ PDF is proportional to $\ln (E_t^2/Q^2)$, the $\gamma^*_L$ PDF depends on the scale in a typically hadron-like manner. [**b) $\mathbf k_t$ unordered initial QCD cascades**]{} accompanying the hard process, employed for example in BFKL or CCFM evolution, can lead to final states in which the partons with the largest $k_t$ may come from the cascade, and not, as in DGLAP evolution, from the hard subprocess. Such events may have a similar topology as is observed for the resolved interactions introduced in the previous paragraph. This possibility is investigated using the CASCADE generator [@Cascade; @Cascade2] based on the CCFM evolution equation. [**c) NLO calculations**]{} with and without the concept of the resolved virtual photon offer another natural possibility to be examined. Such comparisons are envisaged. All of these three approaches include higher order corrections to the LO QCD matrix elements; however, each of the approaches treats them differently. Measurement of Dijet Cross-Section ================================== The selection criteria defining the H1 and ZEUS dijet samples are summarized in Table \[tabulka\]: [**H1**]{}    (16.3 pb$^{-1}$, 1999) [**ZEUS**]{}    (38.2 pb$^{-1}$, 1996-97) -------------------------------------------- ------------------------------------------- $\sqrt{s}=318~{\rm GeV}$ $\sqrt{s}=300~{\rm GeV}$ $2~{{\rm GeV}}^2 < Q^2 < 80~{{\rm GeV}}^2$ $0.1 < Q^2 < 10\,000~{{\rm GeV}}^2$ $0.1 < y < 0.85$ $0.2 < y < 0.55$ $E_t^{jet\,1,2} > 5~{\rm GeV}$ $E_t^{jet\,1} > 7.5~{\rm GeV}$ $\overline{E}_t > 6~{\rm GeV}$ $E_t^{jet\,2} > 6.5~{\rm GeV}$ $-2.5 < \eta^{jet\,1,2} < 0$ $-3 < \eta^{jet\,1,2} < 0$ : Selection criteria of the dijet samples.[]{data-label="tabulka"} $Q^2$ denotes the photon virtuality, $y$ is the inelasticity, $s$ is the total electron-proton centre-of-mass energy, $E_t^{\rm jet\,1,2}$ and $\eta^{\rm jet\,1,2}$ are the transverse energy and pseudorapidity of the jet with the highest or second highest $E_t$, and $\overline{E_t}$ is defined as $(E_t^{\rm jet\,1}+E_t^{\rm jet\,2})/2$. The measured data are corrected for detector effects using a bin-to-bin method (ZEUS) or Bayesian unfolding (H1). The largest source of systematic errors arises from the main calorimeter calibration uncertainty and, in the case of H1, also from a model dependence of the detector correction. The ZEUS measurement was presented in more detail at the EPS 2001 [@Zeus]. Results and Discussion ====================== The corrected triple-differential dijet cross-section measured by ZEUS as a function of [$Q^2$]{}, [$\overline{E}_t^2$]{}and [$x_\gamma$]{}is shown in Fig. \[Zeus1\]. A prediction of HERWIG with the SaS1D parameterization of the [$\gamma^*_T$ ]{}PDF, as well as the direct contribution of HERWIG is compared to the data. Since the overall normalization of the LO Monte Carlo simulation is to some extent uncertain, the HERWIG prediction has been normalized to the highest [$x_\gamma$]{}bin $(x_\gamma > 0.75)$ in the data. The normalization is done separately for each $(Q^2,{\overline{E}_t^2})$ bin. In the region where $Q^2 > {\overline{E}_t^2}$, the data are well described by the direct HERWIG component only. Resolved interactions are needed if $Q^2 < {\overline{E}_t^2}$. However, even with the [$\gamma^*_T$ ]{}resolved processes included, HERWIG tends to underestimate the data in the lowest [$Q^2$]{}region for $x_\gamma < 0.75$. This fact is also demonstrated in Fig. \[Zeus2\] by the ratio of $\sigma(x_\gamma < 0.75) / \sigma(x_\gamma > 0.75)$. The slope of this distribution can be interpreted as a suppression of the virtual photon structure with increasing photon virtuality. The corrected triple-differential dijet cross-section measured at H1 as a function of [$Q^2$]{}, [$\overline{E}_t^2$]{}and [$x_\gamma$]{}is shown in Fig. \[H1\]. The H1 measurement is performed in a different phase space (see Table \[tabulka\]) and Monte Carlo predictions are not normalized to the data. A comparison of the H1 measurement with HERWIG leads to similar conclusions as drawn above for ZEUS. In addition, we can see that for the highest [$Q^2$]{}range ($25 < Q^2 < 80 ~{\rm GeV}^2$) and $x_\gamma<0.75$, the HERWIG direct contribution almost describes the data in the lowest [$\overline{E}_t^2$]{}bin, but is too low in the highest [$\overline{E}_t^2$]{}bin. This indicates an importance of the resolved processes even at high [$Q^2$]{}, once the hard scale, [$\overline{E}_t^2$]{}, is high enough. Standard HERWIG with direct and [$\gamma^*_T$ ]{}resolved contributions underestimates the data. The description is improved by adding [$\gamma^*_L$ ]{}resolved photon interactions, which is done using a slightly modified HERWIG with the correct longitudinal photon flux and a recent $\gamma^*_L$ PDF parameterization [@Chyla:2000hp]. As demonstrated in Fig. \[H1\], the [$\gamma^*_L$ ]{}resolved contribution is significant, and brings HERWIG closer to the measurement. On the other hand, a simple enhancement of the PDF of the $\gamma^*_T$ in the resolved contribution could lead to a similar prediction as the introduction of $\gamma^*_L$. To eliminate this ambiguity, the dijet cross-section has also been studied as a function of [$Q^2$]{}, [$x_\gamma$]{} and $y$, which is shown in Fig. \[H2\]. HERWIG is below the data, even if the resolved [$\gamma^*_L$ ]{}is added. This may be due to the uncertainty of the overall normalization of the LO Monte Carlo prediction. In the region where $x_\gamma<0.75$, the slope of the HERWIG prediction depends significantly on whether $\gamma^*_L$ processes are included or not. The $\gamma^*_L$ contributes significantly at low $y$, while it becomes very small compared to $\gamma^*_T$ at high $y$. Unlike a pure enhancement of $\gamma^*_T$ resolved processes by a constant factor, the addition of $\gamma^*_L$ brings the $y$ dependence of HERWIG much closer to the measurement. As motivated in Section \[uvod\], the measured cross-sections in Fig. \[H1\] are also compared to a prediction of the CASCADE MC program based on the CCFM evolution approach. This theoretical concept does not involve any information about the virtual photon structure and involves many fewer free parameters for tuning than the usual DGLAP-like MC programs. Nevertheless, CASCADE describes the data well, except for the [$Q^2$]{}dependence. The [$Q^2$]{}behavior, however, is related to the parameterization of the unintegrated PDFs used in the program, which are not yet constrained unambiguously. Conclusions =========== The dijet cross-sections measured as a function of [$Q^2$]{}, [$\overline{E}_t^2$]{}, [$x_\gamma$]{} and [$Q^2$]{}, [$x_\gamma$]{}, $y$ at H1 and ZEUS have been presented. In the DGLAP evolution scheme, the importance of the [$\gamma^*_T$ ]{}resolved photon interactions is clearly demonstrated in the region where ${\overline{E}_t^2}> Q^2$. Additional [$\gamma^*_L$ ]{}resolved photon contributions further improve the agreement of the HERWIG 5.9 prediction with the measurement. Exploring the CCFM approach, the MC program CASCADE 1.0 gives a qualitative description of the measured differential cross-sections; however, the [$Q^2$]{}dependence is not reproduced. On the other hand, the [$x_\gamma$]{}dependence in CASCADE is comparable to the sum of the direct and resolved contributions in DGLAP-like MC programs. [99]{} J. Chýla and M. Taševský, Eur. Phys. J. [**C18**]{} (2001), 723. J. Chýla, Phys. Lett.  [**B488**]{} (2000), 289. H. Jung and G.P. Salam, Eur. Phys. J. [**C19**]{} (2001), 351-360. H. Jung, “Heavy Quark production at HERA in $k_t$ factorization supplemented with CCFM evolution”, These proceedings. ZEUS Collab., EPS 2001 conference: “The [$Q^2$]{}and [$\overline{E}_t^2$]{}dependence of dijet cross sections in $\gamma^*p$ interactions at HERA”, paper no. 636. [^1]: S GA AVČR B1010005, INGO-LA 116/2000, LN00A006 [^2]: T PHOTON 2001 C, A, S, S 2001

--- author: - Bingbing Suo - 'Yan-Mei Yu' - HuiXian Han title: 'Supplementary materials for “Relativistic configuration interaction calculation on the ground and excited states of iridium monoxide"' --- Compare two lowest states with $\Omega$ value of 5/2 and 7/2 on RASSI level of theory ===================================================================================== Recently, the highly resolution spectra of IrO has been studied by Pang *et al.* (J. Phys. Chem. A, 116, 9379, 2013) and Adam *et al.* (J. Mol. Spectrosc., 286,46,2013). Both of these studies determined that the ground state of IrO had the $\Omega$ value of 5/2. These inspired Suo and Han to reconsider previous calculation (Chem. Phys. Lett., 548, 12, 2012). As pointed out by Adam and coworkers (J. Mol. Spectrosc., 286, 46, 2013), the $^4\Delta_{5/2}$ state may drop below the $^4\Delta_{7/2}$ state when the influence of spin-orbital coupling between the $^4\Delta$ and $^2\Delta$ is included. In previous work by Suo *et. al*, the interaction between the $^4\Delta_{5/2}$ and $^2\Delta_{5/2}$ states is not be considered. Therefore, Suo and Han performed the SOC calculation for $^4\Delta$, $A^2\Delta$ and $B^2\Delta$ states. The results of the $^4\Delta_{7/2}$ and $^4\Delta_{5/2}$ substates are shown in the Table \[RASSIRE\]. In the new calculation, the energy separation of the 7/2 and 5/2 is reduced to 1108 cm$^{-1}$, which is smaller than 2425 cm$^{-1}$ in previous calculation (Chem. Phys. Lett., 548, 12, 2012). However, we still could not assign a 5/2 ground state due to larger energy gap of the 7/2 and 5/2 states. The discrepancy between theoretical and experimental results may come from insufficient treatment the relativistic effect. The large SOC effect in IrO seems to be calculated by the variational method rather than the perturbation theory. Therefore, multi-configurational electron correlation calculation in the frame of fully relativistic theory is highly desired to give the more accurate description to the ground state of IrO. This is our main motivation of the present work. $\Omega$ R$_e$ Å $\Delta$E $\omega$ (cm$^{-1}$) Configurations ---------- --------- ----------- ---------------------- ------------------------------------------------------------------------- 7/2 1.700 0 950 $^4\Delta_{7/2}(100\%)$ 5/2 1.704 1108 937 $^4\Delta_{5/2}(89.6\%),A^2\Delta_{5/2}(9.6\%), B^2\Delta_{5/2}(6.4\%)$ Excited state of IrO ==================== More roots are required in the KRCI implementation in order to get the low-lying excited state of IrO, which increases the computation demand greatly. Therefore, the calculation for the excited states is mainly performed on the small basis set, then a composite scheme is adapted to give the estimated value $T^{Com}$ of the excited energy at larger basis set. The correction due to the finite basis set is determined as $\Delta T^{basis}=T^{Ver}_{cv3z}-T^{Ver}_{cv2z}$, where $T^{Ver}_{cv3z}$ and $T^{Ver}_{cv2z}$ are the vertical excitation energies that are calculated at the basis sets \[dyall.cv3z (Ir) + aug-cc-pCVTZ (O)\] and \[dyall.cv2z (Ir) + aug-cc-pCVDZ (O)\], respectively. Then, $T^{Com}$ is determined as $T^{Com}=T^{Adi}_{cv2z}+\Delta T^{basis}$, where $T^{Adi}_{cv2z}$ is the adiabatic excited energy obtained at the basis set \[dyall.cv2z (Ir) and aug-cc-pCVDZ (O)\]. The correction $\Delta T^{basis}$ is regarded as the corresponding uncertainty. [p[1.2cm]{} p[1.2cm]{} p[1.2cm]{} p[1.2cm]{} p[1.2cm]{} p[1.2cm]{} ]{}state &$T^{Adi}_{cv2z}$ &$T^{Ver}_{cv2z}$ &$T^{Ver}_{cv3z}$ & $\Delta T^{basis}$ & $T^{Com}$\ \ 1 &8481 &8481 &8373 &-108 &8373\ 2 &10350 &10994 &12256 &1262 &11612\ 3 &12691 &13369 &14140 &771 &13462\ 4 &13893 &14260 &14752 &492 &14385\ 5 &14723 &14901 &15673 &772 &15495\ 6 &15448 &15866 &16900 &1035 &16483\ 7 &17231 &17249 &17712 &462 &17694\ 8 &18316 &18762 &19750 &988 &19304\ 9 &20043 &20519 &21511 &992 &21035\ 10 &22022 &22807 &23360 &553 &22575\ 11 &22320 &23086 &23699 &613 &22933\ 12 &22914 &23735 &25247 &1513 &24426\ 13 &23518 &24470 &25660 &1190 &24708\ 14 &24796 &25935 &27176 &1241 &26037\ 15 &25324 &27778 &29215 &1437 &26761\ \ 1 &8662 &9165 &10063 &898 &9560\ 2 &12061 &12815 &13880 &1065 &13126\ 3 &15684 &15684 &15428 &-256 &15428\ 4 &16209 &16771 &17110 &338 &16547\ 5 &17661 &17963 &18887 &924 &18585\ 6 &20028 &21021 &22006 &985 &21013\ 7 &22319 &23674 &24833 &1159 &23478\ 8 &24497 &24965 &26014 &1049 &25546\ 9 &26374 &28992 &29288 &296 &26671\

--- abstract: 'Measurements in magnetic fields applied at small angles relative to the electron plane in silicon MOSFETs indicate a factor of two increase of the frequency of Shubnikov-de Haas oscillations at $H>H_{sat}$. This signals the onset of full spin polarization above $H_{sat}$, the parallel field above which the resistivity saturates to a constant value. For $H<H_{sat}$, the phase of the second harmonic of the oscillations relative to the first is consistent with scattering events that depend on the overlap instead of the sum of the spin-up and spin-down densities of states.' address: - 'Physics Department, City College of the City University of New York, New York, New York 10031' - 'Delft University of Technology, Department of Applied Physics, 2628 CJ Delft, The Netherlands' author: - 'Sergey A. Vitkalov, Hairong Zheng, K. M. Mertes and M. P. Sarachik' - 'T. M. Klapwijk' title: 'Small Angle Shubnikov-de Haas Measurements in Silicon MOSFET’s: the Effect of Strong In-Plane Magnetic Field.' --- [2]{} A great deal of interest has recently been focussed on the anomalous behavior of two-dimensional (2D) systems of electrons[@krav; @popovic] and holes[@coleridge; @shahar; @cambridge] whose resistivities unexpectedly decrease with decreasing temperature, behavior that is generally associated with metals rather than insulators[@NAS]. One of the most intriguing characteristics of these systems is their enormous response to magnetic fields applied in the plane of the electrons[@simonian; @pudalov; @dolgopolov] or holes[@cambridge; @yoon]: the resistivity increases sharply by more than an order of magnitude, saturating to a constant plateau value above a magnetic field $H_{sat}$. In this paper we report studies of the resistivity of the 2D electron system in silicon MOSFETs in magnetic fields applied at small angles $\phi$ with respect to the plane. This allows a study of the Shubnikov-de Haas (SdH) oscillations in perpendicular fields sufficiently small that the orbital motion has a negligible effect on the response to the in-plane component of the magnetic field. At small tilt angles $\phi$, the SdH oscillations plotted versus filling factor have twice the period below $H_{sat}$ compared with the period above $H_{sat}$. This implies that the electron system is fully spin-polarized in high fields, $H>H_{sat}$, where the resistivity has reached saturation. Detailed examination of the oscillations in fields below $H_{sat}$ indicates unusual behavior consistent with electron scattering that depends on the product rather than the sum of the spin-up and spin-down densities of states. Two silicon MOSFETs with mobilities $\mu \approx 20,000\;$V/cm$^2$ s at $T=4.2$ K were used in these studies. Contact resistances were minimized by using samples with a split-gate geometry, which permit high densities in the vicinity of the contacts while allowing independent control of the density of the 2D system under investigation. Standard $AC$ four-probe techniques were used at $3$ Hz to measure the resistance in the linear regime using currents typically below $5$ nA. Data were taken on samples mounted on a rotating platform in a $^3$He Oxford Heliox system at temperatures down to $0.235$ K in magnetic fields up to $12$ T. \[1\] Measurements were first taken with the plane of the sample oriented parallel to the magnetic field[@alignment]. The resistance, $R_{xx}$, is shown in Fig. 1 as a function of field for different fixed gate voltages spanning densities between $n_s = 0.8 \times 10^{11}$ cm$^{-2}$ and $n_s = 2.88 \times 10^{12}$ cm$^{-2}$ (the zero field critical density for the metal-insulator transition is $n_c \approx 0.84 \times 10^{11}$ cm$^{-2}$). Consistent with earlier findings [@simonian; @pudalov; @yoon; @magneto; @okamoto], the in-plane ($\phi=0$) magnetoresistance rises dramatically with increasing field and saturates above a density-dependent field $H_{sat}(n_s)$[@defineH]. The sample was then rotated to make a small angle $\phi$ with respect to the field, so that the in-plane component was almost equal to the total field $H_{\Vert} \approx H$, while the projection in the perpendicular direction, $H_{\perp} \approx \phi H$, remained relatively small even in high fields. $R_{xx}$ and $R_{xy}$ were measured simultaneously as a function of magnetic field for fixed angle $\phi$, temperature T, and density $n_s$. \[1\] For various different densities $n_s$, Figs. 2 (a) and 2 (b) show the resistance $R_{xx}$ vs filling factor $\nu$ at two different angles $\phi$ between the magnetic field and the 2D plane. Similar curves were obtained at other small angles. The filling factor $\nu=n_s\Phi_0/H_\bot$ was calculated using the relation $n_s=H_{\bot}/(R_{xy}ec)$. The Hall resistivity $R_{xy}$ and Hall coefficient $R_H$ were determined from the low-field data, $R_{xy}=R_H\times H_{\bot}$, ([*i. e.*]{} in fields below the onset of quantum oscillations). For the lowest densities shown in Fig. 2 (a) at angle $\phi=6^o$, the system is in the high-field saturated regime above $H_{sat}$ for filling factor $\nu<10$. The large arrow indicates $H_{sat}$ for $n_s=1.54 \times 10^{11}$ cm$^{-2}$. Quantum oscillations are clearly evident superimposed on the large plateau value of $R_{xx}$ at small $\nu$. In this region the period of the SdH oscillations corresponds to a change in filling factor $\Delta \nu_{sat}=2$ (including the two-fold valley degeneracy for a 2D layer of (100) silicon). For higher filling factors $\nu>10$ where the 2D electron system is below saturation ($H<H_{sat}$), the period of the SdH oscillations is twice as long, namely $\Delta \nu=4$. Similar behavior is shown for a bigger angle in Fig. 2 (b), where the larger perpendicular component gives rise to stronger SdH oscillations; here the period doubling is found above $\nu \approx 8$. The period $\Delta \nu=2$ of the oscillations at $H>H_{sat}$ corresponds to complete spin polarization of the electrons. There are several scenarios that could account for a transition to full polarization in strong magnetic fields, depending on the nature of the ground state of the system at $H=0$ [@NAS], an issue that is currently under debate. Here we restrict the discussion to a simple model within a single- particle description. Using this approach we were able to explain the doubling of the frequency of the SdH oscillations at $H>H_{sat}$. However, the detailed behavior of the oscillations in small perpendicular fields is not fully consistent with this model. As shown schematically in the inset to Fig. 2, the spin-up and spin-down electron bands are split by the Zeeman energy $\Delta_Z=g\mu_B H$, while the spacing between the Landau levels, $\hbar \omega_c= \hbar eH_{\bot}/mc$, is determined by the perpendicular component of the field. We consider the progression as electrons are added to the system: for small densities, $E_F<\Delta_Z$, electrons are added to Landau levels in the spin-up band only, corresponding to a SdH periodicity $\Delta \nu=2$ (including a factor 2 for the valley degeneracy in silicon); at high densities, $E_F>\Delta_Z$, twice as many electrons are required to fill both spin-up and spin-down Landau levels, yielding the double period, or $\Delta \nu=4$[@Elihu]. An equivalent argument holds for fixed density as one reduces the magnetic field. Thus, the shorter period $\Delta \nu=2$ at $H>H_{sat}$ signals the onset of full polarization of the electron system[@splitting]. The relationship between $H_{sat}$ and complete spin polarization was also found by Okamoto [*et al*]{} [@okamoto] using a different experimental method. The observed period-doubling is consistent with this simple model only if the spin-up and spin-down levels are degenerate or nearly so, so that $\alpha=\Delta_Z/ \hbar \omega_c = i$ with $i$ an integer. The double period should revert to a single period when $\alpha=i+1/2$, corresponding to a spin-up Landau level between two spin-down Landau levels. The ratio $\alpha$ can be varied experimentally by changing the angle $\phi$, or by using the fact that the electron-electron interaction-enhanced g-factor (and thus $\Delta_Z)$ decreases with increasing electron density in silicon MOSFETs[@okamoto; @Ando]. By taking data over a broad range of densities, we were able to smoothly vary $\alpha$ by more than $1$. Close examination of the data shows that the double period in fields below $H_{sat}$, although stable over a broad range, does break down in a narrow range of densities that is different for different angles $\phi$, as expected within this model. SdH oscillations reflect changes in electron scattering due to periodic oscillations of the density of states at the Fermi level[@abric]. For the weak perpendicular fields used in our expriments, there is strong scattering and the SdH oscillations are small[@Dingle]. Unlike the situation that prevails in high magnetic fields, where the Landau levels are sharp and well-defined, the density of states is best represented in this regime by a harmonic expansion[@Ando]: $$D_{\downarrow, \uparrow}(E)=D_0(1+\epsilon \times cos(2\pi (E \pm \Delta_Z/2)/\hbar \omega_c) +O(\epsilon^2)) \eqno{(1)}$$ Here $E$, $\Delta_Z=g\mu_B H$, and $\hbar \omega_c$ are the energy, Zeeman energy and cyclotron energy, respectively. The small parameter $\epsilon=2 exp(-\pi/(\omega_c \tau)) \ll 1$[@Ando] is proportional to the Dingle factor [@Dingle]. Small variations in the resistivity are proportional to small variations in scattering: $\Delta \rho/ \rho=\Delta W/W$. Using the Born approximation, $W \sim \int \delta (E-E_F)D(E)dE$, one can show that: $$\Delta \rho/ \rho= \cases {\epsilon \times cos(\pi \Delta_Z/ \hbar \omega_c) cos(2\pi E_F^0/\hbar \omega_c) , & $\Delta_Z <2 E_F^0 $\cr \epsilon \times cos(4\pi E_F^0/\hbar \omega_c), & $\Delta_Z >2 E_F^0$} \eqno{(2)}$$ Here the Fermi energy $E_F$ is measured from the bottom of the band at $H=0$, $E_F^0$ is the Fermi energy at $H=0$, and $D(E)$ is the total density of states: $D(E)=D_\uparrow(E)+ D_\downarrow(E)$. We assumed $T=0$ and neglected higher harmonic terms of order $\epsilon^2$ in Eq.(1) as well as higher order corrections due to oscillations of the Fermi energy. This demonstrates that the SdH period changes by a factor of 2 when $\Delta_Z >2 E_F^0$, corresponding to full polarization of the electrons. The term $A = \epsilon cos(\pi \Delta_Z/ \hbar \omega_c)$ of Eq. (2) depends on the ratio $\alpha=\Delta_Z/\hbar\omega_c$, which is fixed for a given angle and electron density and does not vary with magnetic field. It determines the overall amplitude of the oscillations at fields below $H_{sat}$, when $\Delta_Z <2 E_F^0$. This amplitude has a maximum when $\alpha=\Delta_Z/ \hbar \omega_c=i$ is an integer, corresponding to spin up and spin down densities of states oscillating in phase, and vanishes when $\alpha = i + 1/2$ (see Eq.1). Figure 3 shows data over a narrow region near $\alpha=i+1/2$ where the amplitude of the first harmonic is small, allowing detailed examination of higher harmonic terms (see Eq.(1)). Based on the usual assumption that the SdH oscillations are determined by the total density of states, $D(E)=D_\uparrow(E) + D_\downarrow(E)$, one expects and generally observes[@example] the progression illustrated schematically in Fig. 4: a minimum which becomes progressively deeper develops at the center of each maximum (see curves (a) and (b)), gradually splitting it into two separate maxima (curve (c)). Thus, the minima of the second harmonic (curve (c)) are at the positions of the maxima of the first harmonic (curve (a) in Fig. 4). \[3\] \[3\] However, careful examination of the data of Fig. 3 shows that the behavior below $H_{sat}$ observed experimentally in silicon MOSFETs is quite different: no minima develop within the maxima, splitting them into two; instead, the maxima simply diminish in amplitude and new neighboring maxima appear and grow in amplitude. The maxima of the first ($n_s=3.83$ and $4.51\times 10^{11}$ cm$^{-2}$) and second ($n_s=4.14\times 10^{11}$ cm$^{-2}$) harmonics are in phase (as in curves (a) and (d) of Fig. 4) rather than $180^o$ out of phase. The origin of this unusual behavior is not clear and warrants further careful study. Interestingly, the phase relation between first and second harmonics observed in our experiments can be obtained within the single particle model used earlier if one considers the product of spin-up and spin-down densities of states, $W =f(D_\uparrow(E)\times D_\downarrow(E))$, rather than their sum; curve (d) of Fig. 4 is the result of such a calculation. This suggests there is a sizable contribution to the electron scattering from events that depend on the overlap of spin-up and spin-down densities of states, perhaps reflecting enhanced scattering of electrons of opposite spin. In summary, measurements of small-angle Shubnikov-de Haas oscillations indicate that the period of the oscillations changes by a factor of two at the magnetic field $H_{sat}$ above which the resistance has reached saturation. We attribute the abrupt change in period to the onset of full polarization of the electron spins. The period doubling in fields below $H_{sat}$ is stable with respect to the angle between the magnetic field and the 2D plane, and is observed for all electron densities except in a narrow interval, where the amplitude of the first harmonic of the SdH oscillations vanishes and the second harmonic is observable. The phase observed for the second harmonic relative to the first is consistent with SdH oscillations due to scattering events that depend on the overlap instead of the sum of the spin-up and spin-down densities of states at the Fermi level. This unusual behavior may reflect the importance of many-body interactions in the 2D system. We are grateful to S. Bakker and R. Heemskerk for their contributions in developing and preparing the MOSFETs used in this work. We thank E. Abrahams, L. Ioffe, F. Fang, A. Fowler, S. V. Kravchenko, A. Shashkin, X. Si, S. Chakravarty, D. Schmeltzer, F. Stern, M. Raikh and U. Lyanda-Geller for illuminating discussions. We are grateful to A. Shashkin and S. V. Kravchenko for valuable comments on the manuscript. This work was supported by U.  S. Department of Energy grant No. DE-FG02-84ER45153. S. V. Kravchenko, G. V. Kravchenko, J. E.Furneaux, V. M. Pudalov, and M. D’Iorio, Phys. Rev. B [**50**]{}, 8039 (1994); S. V. Kravchenko, W. E. Mason, G. E. Bowker, J. E. Furneaux, V. M. Pudalov, and M. D’Iorio, Phys. Rev. B [**51**]{}, 7038 (1995); S. V. Kravchenko, D. Simonian, M. P. Sarachik, Whitney Mason, and J. Furneaux, Phys.Rev. Lett. [**77**]{}, 4938 (1996). D. Popović, A. B. Fowler, and S. Washburn, Phys.Rev. Lett. [**79**]{}, 1543 (1997). P. M. Coleridge, R. L. Williams, Y. Feng, and P. Zawadzki, Phys. Rev. B [**56**]{}, R12764 (1997). Y. Hanein, U. Meirav, D. Shahar, C. C. Li, D. C. Tsui, and Hadas Shtrikman, Phys. Rev. Lett. [**80**]{}, 1288 (1998). M. Y. Simmons, A. R. Hamilton, M. Pepper, E. H. Linfield, P. D. Rose, D. A. Ritchie, A. K. Savchenko, and T. G. Griffiths, Phys. Rev. Lett. [**80**]{}, 1292 (1998). For a brief review see M. P. Sarachik and S. V. Kravchenko, Proc. Natl. Acad. Sci. USA, [**96**]{}, 5900 (1999). D. Simonian, S. V. Kravchenko, M. P. Sarachik, and V. M. Pudalov, Phys. Rev. Lett. [**79**]{}, 2304 (1997). V. M. Pudalov, G. Brunthaler, A. Prinz, and G. Bauer, Pisma Zh. Eksp. Teor. Fiz. [**65**]{}, 887 (1997) \[JETP Lett. [**65**]{}, 932 (1997)\]. V. Dolgopolov, G. V. Kravchenko, A. A. Shashkin, and S. V. Kravchenko, JETP Lett. [**55**]{}. 733 (1992). J. Yoon, C. C. Li, D. Shahar, D. C. Tsui, and M. Shayegan, cond-mat/9907128 (1999). Parallel alignment was determined by choosing the angle for which $R_{xy}=0$. K. Mertes, D. Simonian, M. P. Sarachik, S. V. Kravchenko, and T. M. Klapwijk, Phys. Rev. B [**60**]{}, R5093 (1999). T. Okamoto, K. Hosoya, S. Kawaji, and A. Yagi, Phys. Rev. Lett. [**82**]{}, 3875 (1999). $H_{sat}$ was chosen as the magnetic field where the resistivity has reached 90% of its constant high-field value. We thank Elihu Abrahams for this suggestion. An alternative explanation is that each Landau level splits into spin-up and spin-down sublevels in high magnetic field. It is unlikely that this would occur in every case at precisely the magnetic field where the resistance reaches saturation. Moreover, one expects the Landau levels to be broader for the larger resistances at high fields, making such splitting less likely. Also at $H>H_{sat}$ we do not see any interference effects corresponding to two spin bands. T. Ando, A. B. Fowler and F. Stern, Rev. Mod. Phys. [**54**]{},550, (1982). A.A. Abrikosov,“Fundamentals of the Theory of Metals”, (North Holland, 1988). R.B. Dingle Proc. Roy. Soc. A [**211**]{}, 517 (1952). A clear example can be found in Figs. 16 and 17 of D. R. Leadley [*et al.*]{}, Phys. Rev. B [**58**]{}, 13 036, (1998).

--- abstract: 'Erdös and Nicolas [@erdos1976methodes] introduced an arithmetical function $F(n)$ related to divisors of $n$ in short intervals $\left] \frac{t}{2}, t\right]$. The aim of this note is to prove that $F(n)$ is the largest coefficient of polynomial $P_n(q)$ introduced by Kassel and Reutenauer [@kassel2015counting]. We deduce that $P_n(q)$ has a coefficient larger than $1$ if and only if $2n$ is the perimeter of a Pythagorean triangle. We improve a result due to Vatne [@vatne2017sequence] concerning the coefficients of $P_n(q)$.' --- \[theorem\][Corollary]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Conjecture]{} \[theorem\][Remark]{} José Manuel Rodríguez Caballero\ Département de Mathématiques\ UQÀM\ Case Postale 8888, Succ. Centre-ville\ Montréal, Québec H3C 3P8 Canada\ <rodriguez_caballero.jose_manuel@uqam.ca>\ .2 in Introduction ============ Erdös and Nicolas introduced in [@erdos1976methodes] the function $$F(n) = \max\{ \mathrm{q}_t(n): \quad t \in \mathbb{R}_+^{\ast}\},\label{eq9je93jf3f}$$ where $\mathrm{q}_t(n) = \#\left\{d : \quad d | n\quad \textrm{ and } \quad \frac{1}{2}\,t < d \leq t \right\}$, and they proved that $$\lim_{x \to +\infty} \frac{1}{x}\,\sum_{n \leq x} F(n) = +\infty.\label{PourCoro2r90j430}$$ Kassel and Reutenauer introduced in [@kassel2015counting] a $q$-analog of the sum of divisors, denoted $P_n(q)$, by means of the generating function $$\prod_{m \geq 1}\frac{\left(1-t^m\right)^2}{\left(1-q\, t^m\right)\left(1-q^{-1}\, t^m\right)} = 1 + \left( q + q^{-1} - 2\right) \,\sum_{n=1}^{\infty} \frac{P_n(q)}{q^{n-1}} \, t^n$$ and they proved that, for $q = \exp\left( \frac{2\,\pi}{k}\,\sqrt{-1}\right)$, with $k \in \{2,3,4, 6\}$, this infinite product can be expressed by means of the Dedekind $\eta$-function (see [@kassel2016fourier]). A consequence of this coincidence is that the corresponding arithmetic functions $n \mapsto P_n(q)$, for each of the above-mentioned values of $q$, are related to the number of ways to express a given integer by means of a quadratic form (see [@kassel2015counting] and [@kassel2016complete]). The aim of this paper is to prove the following theorem. \[Prop1\] For each integer $n \geq 1$, the largest coefficient of $P_n(q)$ is $F(n)$. Using this result, we will derive that $P_n(q)$ has a coefficient larger than $1$ if and only if $2\,n$ is the perimeter of a Pythagorean triangle. Also, we will prove that each nonnegative integer $m$ is the coefficient of $P_n(q)$ for infinitely many positive integers $n$. Proof of the main result ======================== In order to simplify the notation in the proofs, we will consider two functions[^1] $f:\mathbb{R} \longrightarrow \mathbb{R}_+^{\ast}$ and $g:\mathbb{R}_+^{\ast} \longrightarrow \mathbb{R}$, defined by $$\begin{aligned} f(x) &=& \frac{1}{2}\,\left(x + \sqrt{8\,n + x^2} \right),\label{Eqr398r49f3fref} \\ g(y) &=& y - \frac{2\,n}{y}.\label{Eqr3489rj49r}\end{aligned}$$ \[Lemmma43j934jr9\] The functions $f(x)$ and $g(y)$ are well-defined, strictly increasing and mutually inverse. Furthermore, $g(y)$ satisfies the identity $$g\left(y\right) = - g\left( \frac{2 \, n}{y} \right). \label{IdenReflection}$$ It follows in a straightforward way from the explicit expressions (\[Eqr398r49f3fref\]) and (\[Eqr3489rj49r\]) that $f(x)$ and $g(y)$ are well-defined and strictly increasing. In particular, the inequality $|x| < \sqrt{2\,n + x^2}$ guarantees that $f(x)\in \mathbb{R}_+^{\ast}$ for all $x \in \mathbb{R}$. On the one hand, for all $x \in \mathbb{R}$, we have $$g\left( f(x)\right) = \frac{\left(f(x) - x - \sqrt{2\,n + x^2}\right)\,\left(f(x) - x + \sqrt{2\,n + x^2}\right)}{2\,f(x)} + x = x.$$ On the other hand, for all $y \in \mathbb{R}_+^{\ast}$, we have $$f\left(g(y)\right) = \frac{y}{2} - \frac{n}{y} + \sqrt{ \left( \frac{y}{2} + \frac{n}{y} \right)^2 } = \frac{y}{2} - \frac{n}{y} + \frac{y}{2} + \frac{n}{y} = y,$$ where we used the inequality $\frac{y}{2} + \frac{n}{y} > 0$, provided that $y > 0$, for the elimination of the square root. Hence, $f(x)$ and $g(y)$ are mutually inverses. Furthermore, using the identity $$- g \left( 2 \, \frac{n}{y}\right) = - \left( \frac{ 2 \, \frac{n}{y}}{2} - \frac{n}{2 \, \frac{n}{y}}\right) = - \left( \frac{n}{y} - \frac{y}{2}\right) = \frac{y}{2} - \frac{n}{y} = g(y),$$ we conclude that (\[IdenReflection\]) holds for all $y \in \mathbb{R}_+^{\ast}$. \[Lemmmmma1\] For each integer $n \geq 1$, $$\frac{P_n(q)}{q^{n-1}} = \sum_{i \in \mathbb{Z}} a_{n,i} q^i,\label{Eqseriesovertheintegers}$$ where $$a_{n,i} = \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{1}{2} \, g \left(d\right) \leq i < \frac{1}{2} \, g\left(2\,d\right) \right\}.\label{anientermesdeg}$$ By Theorem 1.2 in [@kassel2016complete], $$P_n(q) = a_{n,0}\, q^{n-1} + \sum_{i=1}^{n-1} a_{n,i} \, \left(q^{n-1 + i} + q^{n-1 - i} \right),\label{Eqr489rj983jr}$$ where $$a_{n,i} = \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{f(2\,i)}{2} < d \leq f\left(2\,i \right) \right\}. \label{Eq94j94jr934r}$$ The condition $\frac{f(2\,i)}{2} < d \leq f(2\,i)$ is equivalent to $d \leq f(2\,i) < 2\,d$. So, since $g(y)$ is strictly increasing by Lemma \[Lemmma43j934jr9\], the expression (\[anientermesdeg\]) follows for all $0 \leq i \leq n-1$. We will extend $a_{n,i}$ to any $i \in \mathbb{Z}$ using the expression (\[anientermesdeg\]) as the definition of $a_{n,i}$ for $i < 0$. Applying the identity (\[IdenReflection\]) to (\[anientermesdeg\]), $$a_{n,i} = \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{1}{2} \, g \left(d\right) < -i \leq \frac{1}{2} \, g\left(2\,d\right) \right\}.\label{anientermesdeg1}$$ Substituting $i$ by $-i$ in (\[anientermesdeg\]), $$a_{n,-i} = \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{1}{2} \, g \left(d\right) \leq -i < \frac{1}{2} \, g\left(2\,d\right) \right\}.\label{anientermesdeg2}$$ Now, we will prove that $$\begin{aligned} & & \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{1}{2} \, g \left(d\right) < -i \leq \frac{1}{2} \, g\left(2\,d\right) \right\} \\ \nonumber &=& \# \left\{d: \quad d | n \quad \textrm{ and } \quad \frac{1}{2} \, g \left(d\right) \leq -i < \frac{1}{2} \, g\left(2\,d\right) \right\}.\label{EQ89dj93jd984j}\end{aligned}$$ Suppose that $$\frac{1}{2} \, g \left(d\right) = -i, \label{Eqr98j39r9384hr9}$$ for some $d|n$. Transforming (\[Eqr98j39r9384hr9\]) into $d = 2 \left(\frac{n}{d} - i\right)$, it follows that $d$ is even. So, $-i = \frac{1}{2} \, g \left(2 \, d^{\prime}\right)$, where $d^{\prime} = \frac{d}{2}$ is a divisor of $n$. Conversely, suppose that $$-i = \frac{1}{2} \, g \left(2 \, d\right),\label{Eq983jr93jr}$$ for some $d|n$. Transforming (\[Eq983jr93jr\]) into $\frac{n}{d} = 2 \left(d + i\right)$, it follows that $\frac{n}{d}$ is even. Furthermore, $2 \, d$ divides $n$, because $2\,d \, \frac{\frac{n}{d}}{2} = n$ and $\frac{\frac{n}{d}}{2} \in \mathbb{Z}$. So, $\frac{1}{2} \, g \left(d^{\prime}\right) = -i$, where $d^{\prime} = 2\,d$ is a divisor of $n$. Hence, (\[EQ89dj93jd984j\]) holds. Combining (\[EQ89dj93jd984j\]), (\[anientermesdeg1\]) and (\[anientermesdeg2\]), we obtain that $$a_{n,i} = a_{n,-i} \label{Eq4rj93j}$$ holds for all $0 \leq i \leq n-1$. Furthermore, the bound $-(2\,n - 1) \leq g(y) \leq 2\,n - 1$ for all $1 \leq y \leq 2\,n$ and the equality (\[anientermesdeg\]) imply that $$a_{n,i} = 0 \label{Eqr8943jr983r}$$ for all $i \in \mathbb{Z}$ such that $|i| \geq n$. Using that (\[Eq4rj93j\]) holds for all $0 \leq i \leq n-1$ and that (\[Eqr8943jr983r\]) holds for all $i \in \mathbb{Z}$, with $|i| \geq n$, we conclude that the expression (\[Eqr489rj983jr\]) can be transformed into (\[Eqseriesovertheintegers\]), where $a_{n,i}$ is given by (\[anientermesdeg\]) for all $i \in \mathbb{Z}$. \[LEmmer4390r934jr984j\] Let $y_1$ and $y_2$ be two divisors of $2\,n$. If $y_1 < y_2$ then $$g(y_1) + 2 \leq g(y_2). \label{Ineq9j43j934j}$$ Using the expression (\[Eqr3489rj49r\]) we obtain that, for any real number $y > 0$, $$g(y+1) - g(y) > 1, \label{Ineq04r9934r9r}$$ because $g(y+1) - g(y) = 1 + \frac{2 \, n}{y \, \left(y+1\right)}$. Let $y_1$ and $y_2$ be two positive real numbers satisfying $y_2 - y_1 \geq 1$. By Lemma \[Lemmma43j934jr9\], the function $g(y)$ is strictly increasing. So, (\[Ineq04r9934r9r\]) implies that $$g(y_2) - g(y_1) > 1. \label{Ineq03jf9j349fj3}$$ Furthermore, suppose that $y_1$ and $y_2$ are divisors of $2\,n$. It follows that $g(y_2) - g(y_1)$ is an integer, because of (\[Eqr3489rj49r\]). In this case, the inequality (\[Ineq03jf9j349fj3\]) becomes $$g(y_2) - g(y_1) \geq 2. \label{Ineq349j943fj9}$$ Therefore, (\[Ineq9j43j934j\]) holds. Now, we can prove our main result. By Lemma \[Lemmmmma1\], the coefficient $a_{n,i}$ is defined for all $i \in \mathbb{Z}$ by the expression (\[anientermesdeg\]). First, we will prove that the largest coefficient of $P_n(q)$ is at most $F(n)$. Take some $j \in \mathbb{Z}$ satisfying $a_{n,j} = \max\{a_{n,i}: \quad i\in \mathbb{Z} \}$. By (\[anientermesdeg\]), there are $h = a_{n,j}$ divisors of $n$, denoted $d_1, d_2, ..., d_h$ satisfying $$g\left(d_1\right) < g\left(d_2\right) < ... < g\left(d_h\right) \leq 2\,j < g\left(2\,d_1\right) < g\left(2\,d_2\right) < ... < g\left(2\,d_h\right). \label{INEQtee32984ur98}$$ In particular, $$g\left(d_1\right) < g\left(d_2\right) < ... < g\left(d_h\right) < g\left(2\,d_1\right) < g\left(2\,d_2\right) < ... < g\left(2\,d_h\right). \label{INEQtee0it34t5}$$ Applying $f(x)$ to the inequalities (\[INEQtee0it34t5\]) we obtain $$d_1 < d_2 < ... < d_h < 2\,d_1 < 2\,d_2 < ... < 2\,d_h, \label{Ineq42j89jr43}$$ because $f(x)$ and $g(y)$ are mutually inverses in virtue of Lemma \[Lemmma43j934jr9\]. So, we guarantee that $$\frac{1}{2}\, t < d_1 < d_2 < ... < d_h \leq t, \label{Ine589u594hf}$$ where $t = 2\,d_1 - \varepsilon$ for all $\varepsilon > 0$ small enough. Hence, $a_{n,j} \leq F(n)$, because of (\[eq9je93jf3f\]). Now, we will prove that there is at least one coefficient of $P_n(q)$ which reaches the value $F(n)$. Setting $h = F(n)$ and applying (\[eq9je93jf3f\]), it follows that there are $h$ divisors of $n$ satisfying (\[Ine589u594hf\]) for some $t \in \mathbb{R}_+^{\ast}$. The inequalities (\[Ineq42j89jr43\]) follow. Applying $g(y)$ to (\[Ineq42j89jr43\]) we obtain (\[INEQtee0it34t5\]). Setting $$j := \left\lceil \frac{g\left( d_h\right)}{2} \right\rceil,\label{Eq9jr93jr9r4}$$ we have the inequalities $$\begin{aligned} g(d_h) &\leq& 2 j, \label{ineq834jr893jr9} \\ 2 j &\leq& g(d_h) + 1, \label{Ineq8329j3r489j} \\ g(d_h) + 1 &<& g(d_h) + 2 \label{IneqTrivueiuf33}, \\ g(d_h) + 2 &\leq& g(2\,d_1). \label{Ineq3i4r903jr93j} \end{aligned}$$ The inequality (\[ineq834jr893jr9\]) follows from (\[Eq9jr93jr9r4\]). The inequality $2\,j < g\left( d_h\right)+2$ follows from (\[Eq9jr93jr9r4\]) and the stronger inequality (\[Ineq8329j3r489j\]) is obtained using the fact $g\left( d_h\right) \in \mathbb{Z}$, derived from (\[Eqr3489rj49r\]). The inequality (\[IneqTrivueiuf33\]) is trivial. Finally, the inequality (\[Ineq3i4r903jr93j\]) follows by Lemma \[LEmmer4390r934jr984j\], because $d_h$ and $2\,d_1$ are divisors of $2\,n$ satisfying $d_h < 2\,d_1$. Combining (\[ineq834jr893jr9\]), (\[Ineq8329j3r489j\]), (\[IneqTrivueiuf33\]) and (\[Ineq3i4r903jr93j\]) we obtain that $$g(d_h) \leq 2 j < g(2\,d_1). \label{Ineqqj3r9jr9}$$ The inequalities (\[INEQtee32984ur98\]) holds, because of (\[Ineqqj3r9jr9\]) and (\[INEQtee0it34t5\]). Hence, $a_{n,j} = F(n)$. Therefore, the largest coefficient of $P_n(q)$ is $F(n)$. \[Cor0dsffewfewf\] The largest coefficient of $P_n(q)$ is the largest value of $h$ for which (\[Ineq42j89jr43\]) holds. The polynomial $P_{12}(q)$ was computed in [@kassel2015counting], $$\begin{aligned} & & P_{12}(q) = q^{22} + q^{21} + q^{20} + q^{19} + q^{18} + q^{17} + q^{16} + q^{15} + q^{14} + 2 q^{13} \\ & & + 2 q^{12} + 2 q^{11} + 2 q^{10} + 2 q^{9} + q^{8} + q^{7} + q^{6} + q^{5} + q^{4} + q^{3} + q^{2} + q + 1.\end{aligned}$$ Let us compute $j$ such that $a_{12,j} = a_{12,-j}$ are equal to the largest coefficient of $P_{12}(q)$. $$\begin{array}{|c|c|c|c|c|c|c|} \hline d & 1 & 2 & 3 & 4 & 6 & 12 \\ \hline g(d) & -23 & -10 & -5 & -2 & 2 & 10 \\ \hline g(2\,d) & -10 & -2 & 2 & 5 & 10 & 23 \\ \hline \end{array}$$ The equality $F(12) = 2$ implies the existence of $2$ divisors of $12$, for example $d_1 = 2$ and $d_2 = 3$, satisfying (\[Ineq42j89jr43\]). In our case, $$2 < 3 < 2\cdot 2 < 2\cdot 3.$$ Applying $g(y)$ to the above inequalities, we obtain a particular case of (\[INEQtee0it34t5\]), $$-10 < -5 < -2 < 2.$$ So, taking $2\,j = g(3) + 1 = -5 + 1 = -4$, we obtain $a_{12,-2} = a_{12,2} = 2$, which are the coefficients of $q^{9}$ and $q^{13}$. Some consequences of the main result ==================================== Kassel and Reutenauer observed in [@kassel2015counting] that $P_n(q)$ has a coefficient larger than $1$ provided that $n$ is a perfect number or an abundant number. The corresponding necessary and sufficient condition is given in the following result. The polynomial $P_n(q)$ has a coefficient larger than $1$ if and only if $2\,n$ is the perimeter of a Pythagorean triangle. From the explicit formula for Pythagorean triples (see [@sierpinski2003pythagorean]), it follows in a straightforward way that $2\,n$ is the perimeter of a Pythagorean triangle if and only if $n$ has a pair of divisors $d$ and $d^{\prime}$ satisfying the inequality $d < d^{\prime} < 2 \, d$. So, the result follows from Corollary \[Cor0dsffewfewf\]. Vatne proved in [@vatne2017sequence] that the set of coefficients of $P_n(q)$ is unbounded. The following result is a stronger version of this property. \[Coro2\] Let $a_{n,i}$ be the coefficients of $P_n(q)$ as shown in (\[Eqseriesovertheintegers\]) and (\[anientermesdeg\]). For any integer $m \geq 0$, the equality $a_{n,i} = m$ holds for infinitely many $(n, i) \in \mathbb{Z}^2$, with $n \geq 1$. In the proof of Corollary \[Coro2\] we will use the following auxiliary result. \[Lemma2sdfijfiu4\] Let $h \geq 1$ be an integer. If $h$ is a coefficient of the polynomial $P_n(q)$, then $h-1$ is also a coefficient of the same polynomial. Consider two fixed integers $n \geq 1$ and $h \geq 1$. Let $j$ be the largest integer such that $a_{n,j} \geq h$, where $a_{n,j}$ is given by (\[anientermesdeg\]). The inequalities (\[INEQtee32984ur98\]) hold for $h$ divisors of $n$, denoted $d_1, d_2, ..., d_h$. Setting $$i := \left\lceil \frac{g\left( 2\,d_1\right)}{2} \right\rceil,\label{Eq38hf83h48h3}$$ we have the inequalities $$\begin{aligned} g\left( d_h\right) &<& g\left( 2\,d_1\right), \label{ineq334kd90k30d} \\ g\left( 2\,d_1\right) &\leq& 2 \, i, \label{ineq398j93dd34} \\ 2 \,i &\leq& g\left( 2\,d_1\right) + 1, \label{Ineq3dhj93jd9} \\ g(2\,d_1) + 1 &<& g(2\,d_1) + 2 \label{IneqTrivue34jd9348jd934j}, \\ g(2\,d_1) + 2 &\leq & g(2\,d_2). \label{Ineqjwe9jd923jr93j} \end{aligned}$$ The inequality (\[ineq334kd90k30d\]) follows by (\[INEQtee32984ur98\]). The inequality (\[ineq398j93dd34\]) follows from (\[Eq38hf83h48h3\]). The inequality $2 \,i < g\left( 2\,d_1\right) + 2$ follows from (\[Eq38hf83h48h3\]) and the stronger inequality (\[Ineq3dhj93jd9\]) is obtained using the fact $g\left( 2\,d_1\right) \in \mathbb{Z}$, derived from (\[Eqr3489rj49r\]). The inequality (\[IneqTrivue34jd9348jd934j\]) is trivial. Finally, the inequality (\[Ineqjwe9jd923jr93j\]) follows by Lemma \[LEmmer4390r934jr984j\], because $2\,d_1 < 2\,d_2$ and both are divisors of $2\,n$. Combining (\[ineq334kd90k30d\]), (\[ineq398j93dd34\]), (\[Ineq3dhj93jd9\]), (\[IneqTrivue34jd9348jd934j\]) and (\[Ineqjwe9jd923jr93j\]) we obtain that $$g(d_h) \leq 2 i < g(2\,d_2). \label{Inefrijfie4jf}$$ Combining (\[Inefrijfie4jf\]) with (\[Ineq42j89jr43\]), it follows that $$d_2 < d_3 < ... < d_h \leq 2\,i < 2\,d_2 < 2\,d_3 < ... < 2\,d_h. \label{Ineq390ir93j9jr94}$$ Notice that (\[INEQtee32984ur98\]) and (\[ineq398j93dd34\]) imply $$j < i. \label{Ineq98fj8jf943}$$ In virtue of the expression (\[anientermesdeg\]), the inequalities (\[Ineq390ir93j9jr94\]) imply that $$a_{n,i} \geq h-1.\label{Ineq93r94jr93jr}$$ The inequalities (\[Ineq98fj8jf943\]) and (\[Ineq93r94jr93jr\]) imply that $a_{n,i} = h-1$, because $j$ is the largest integer satisfying $a_{n,j} \geq h$. Using (\[PourCoro2r90j430\]), it follows that the range of $F(n)$ is unbounded. By Theorem \[Prop1\], the set of coefficients of $P_n(q)$, for all $n \geq 1$, is unbounded. Take an integer $m \geq 0$. Consider a polynomial $P_n(q)$ whose largest coefficient is $h > m$. Applying Lemma \[Lemma2sdfijfiu4\] several times, we will obtain that $m$ is a coefficient of $P_n(q)$. As there are infinitely many values of $n$ such that $P_n(q)$ has a coefficient larger than $m$, the equality $a_{n,i} = m$ holds for infinitely many $(n, i) \in \mathbb{Z}^2$, with $n \geq 1$. Acknowledge {#acknowledge .unnumbered} =========== The author thanks S. Brlek and C. Reutenauer for they valuable comments and suggestions concerning this research. [99]{} Paul Erdös, Jean-Louis Nicolas. Méthodes probabilistes et combinatoires en théorie des nombres. Bull. SC. Math [**2**]{} (1976): 301–320. Christian Kassel and Christophe Reutenauer, Counting the ideals of given codimension of the algebra of Laurent polynomials in two variables, <https://arxiv.org/abs/1505.07229>, 2015. Christian Kassel and Christophe Reutenauer, Complete determination of the zeta function of the Hilbert scheme of $ n $ points on a two-dimensional torus, <https://arxiv.org/abs/1610.07793>, 2016. Christian Kassel and Christophe Reutenauer, The Fourier expansion of $\eta(z)$ $\eta(2z)$ $\eta(3z)$ $/$ $\eta(6z)$, [*Archiv der Mathematik*]{} [**108**]{}.5 (2017): 453-463. Waclaw Sierpinski, Pythagorean triangles, Courier Corporation, 2003. Jon Eivind Vatne, The sequence of middle divisors is unbounded. Journal of Number Theory [**172**]{} (2017): 413–415. [^1]: The function $g(y)$ was implicitly used in Proposition 2.2. in [@kassel2016fourier].

--- abstract: 'We study a complex formation between the DNA and cationic amphiphilic molecules. As the amphiphile is added to the solution containing DNA, a cooperative binding of surfactants to the DNA molecules is found. This binding transition occurs at specific density of amphiphile, which is strongly dependent on the concentration of the salt and on the hydrophobicity of the surfactant molecules. We find that for amphiphiles which are sufficiently hydrophobic, a charge neutralization, or even charge inversion of the complex is possible. This is of particular importance in applications to gene therapy, for which the functional delivery of specific base sequence into living cells remains an outstanding problem. The charge inversion could, in principle, allow the DNA-surfactant complexes to approach negatively charged cell membranes permitting the transfection to take place.' author: - | Paulo S. Kuhn, Yan Levin[^1], and Marcia C. Barbosa\ Instituto de Física, Universidade Federal do Rio Grande do Sul\ Caixa Postal 15051, CEP 91501-970, Porto Alegre, RS, Brazil title: 'Charge inversion in DNA–amphiphile complexes: Possible application to gene therapy' --- PACS.05.70.Ce - Thermodynamic functions and equations of state PACS.61.20.Qg - Structure of associated liquids: electrolytes, molten salts, etc. PACS.61.25.Hq - Macromolecular and polymer solutions; polymer melts; swelling Introduction ============ In the last few years gene therapy has received significant attention both from the scientific community and from the general public. The development of new techniques for transferring genes into living cells allowed for the potential treatment of several diseases of genetic origin [@Fried]-[@Hope]. The central problem of gene therapy is the development of safe and efficient gene delivery system. Since both the $DNA$ and the cell membranes are negatively charged, the naked polynucleotides are electrostatically prevented from entering inside cells. Furthermore, the unprotected $DNA$ is rapidly degraded by nucleases present in plasma [@Hope]. Although, much effort has concentrated on viral transfection, non-viral methods have received increased attention. This is mostly due to the possible complications which can arise from recombinant viral structures, and the consequent risk of cancer. In the non-viral category, the $DNA-liposome$ complexes have shown the most promise. Cationic liposomes can associate with the $DNA$ segments, neutralizing or even inverting the electric charge of nucleotides, thus significantly increasing the efficiency of gene adsorption and transfection by cells. In this paper we present a model of $DNA-amphiphile$ solutions. We find that in equilibrium, solution consists of complexes composed of DNA and associated counterions and amphiphiles. As more amphiphiles are added to solution, a cooperative binding transition is found. At the transition point a large fraction of the $DNA's$ charge is neutralized by the condensed surfactants. If the density of surfactant is increased beyond this point, a charge inversion of the DNA becomes possible. The necessary density of amphiphile needed to reach the charge inversion is strongly dependent on the characteristic hydrophobicity of surfactant molecules. In particular, we find that for sufficiently hydrophobic amphiphiles, such as for example some cationic lipids, the charge inversion can happen at extremely low densities. The model ========= Our system consists of an aqueous solution of $DNA$ segments, cationic surfactants, and monovalent salt. Water is modeled as a uniform medium of dielectric constant $D$. In an aqueous solution, the phosphate groups of the $DNA$ molecules become ionized resulting in a net negative charge. The salt is completely ionized, forming an equal number of cations and anions. Similarly the surfactant molecules are assumed to be fully dissociated producing negative anions and polymeric chains with cationic head groups. Following the usual nomenclature, we shall call the ionized $DNA$ molecules the “polyions”, the positively charged ions the “counterions”, and the negatively charged anions the “coions”. To simplify the calculations, all the counterions and coins will be treated as identical, independent of the molecules from which they were derived. The $DNA$ strands will be modeled as long rigid cylinders of length $L$ and diameter $a_p$, with the charge $-Zq$ distributed uniformly, with separation $b \equiv L/Z$, along the major axis. The cations and anions will be depicted as hard spheres of diameter $a_c$ and charge $\pm q$. For simplicity we shall also suppose that each one of the $s$ surfactant monomers is a rigid sphere of diameter $a_c$ with the “head” monomer carrying the charge $+q$. The interaction between the hydrophobic tails is short ranged and characterized by the hydrophobicity parameter $\chi$ (see Fig. $1$). The density of $DNA$ segments is $\rho_{p}=N_p/V$, the density of monovalent salt is $\rho_m=N_m/V$, and the density of amphiphile is $\rho_s=N_s/V$, where $N_i$ is the number of molecules of specie $i$ and $V$ is the volume of the system. The strong electrostatic attraction between the polyions, counterions, and amphiphiles, leads to formation of complexes consisting of [*one*]{} polyion, $n_c$ counterions, and $n_s$ amphiphilic molecules. We shall assume that to each phosphate group of the $DNA$ molecule can be associated at most [*one*]{} counterion [**or**]{} $l \le l_{max}$ surfactants. This assumption seems to be quite reasonable in view of the fact that the electrostatic repulsion between the counterions will prevent more than one counterion from condensing onto a given monomer. On the other hand, the gain in hydrophobic energy resulting from the close packing of the surfactant molecules might be able to overcome the repulsive electrostatic interaction between the surfactant head groups, favoring condensation of more than one surfactant on a given monomer (see Fig. $2$). The $l$ amphiphilic molecules form a “ring” of radius $a$ around the central negative monomer of the $DNA$ (see Fig. $3$). If we assume that most of the hydrocarbon chain of the associated surfactants is hidden inside the $DNA$ molecule, the maximum number of surfactants in a ring can be estimated from the excluded volume considerations, $l_{max}=2 \pi a /a_c$, where $a \equiv (a_p + a_c)/2$ is the radius of the exclusion cylinder around a polyion. At equilibrium, each site (monomer) of a polyion can be free or have one counterion [*or*]{} a ring of $l=1,...,l_{max}$ surfactants associated to it. We define the surface coverage of counterions as $p_c=n_c/Z$, and the surface coverage of surfactant rings as $p_l=n_l/Z$, where $n_c$ is the number of condensed counterions and $n_l$ is the number of rings containing $l$ surfactants. Each polyion has a distribution of rings containing from one to $l_{max}$ surfactants. We shall neglect the polydispersity in the size of the complexes, assuming that all the complexes have $n_c$ counterions and $n_s$ amphiphilic molecules — in rings of $\{p_l\}$ — with $$\label{eq00} n_s=\sum_{l=1}^{l_{max}}Z l p_l \; .$$ The total charge of each polyion is, therefore, renormalized from $-Zq$ to $-Z_{eff}q$, with $Z_{eff}\equiv Z-n_c-n_s$ [@Alex]-[@LevBarb1]. &gt;From overall charge neutrality, the density of free cations is $\rho_+=\rho_m+(Z-n_c)\rho_p$, the density of free anions is $\rho_-=\rho_m+\rho_s$, and the density of free surfactants is $\rho_s^f=\rho_s-n_s\rho_p$. We shall restrict our attention to the limit of low surfactant densities, so as to prevent micellar formation in the bulk. The aim of the theory is to determine the characteristic values of $n_c$, $n_s$, and the surface coverage by rings $\{p_l\}$. To accomplished this, the free energy of the $DNA-surfactant$ solution will be constructed and minimized. The Helmholtz free energy ========================= The free energy is composed of three contributions, $$\label{eq0} F=F_{complex}+F_{electrostatic}+F_{mixing} \; .$$ The first term is the free energy needed to form the isolated complexes. The second term accounts for the electrostatic interaction between the counterions, coions, surfactants and complexes. Finally, the third term is the result of entropic mixing of various species. To calculate the free energy required to construct an isolated complex composed of one polyion, $n_c$ condensed counterions, and $n_s$ condensed surfactants, we employ the following simplified model. Each monomer of a polyion can be free or occupied by a counterion, [*or*]{} by $1\leq l\leq l_{max}$ amphiphiles (see Fig. $2$). Therefore, to each monomer $i$ we associate occupation variables $\sigma_c(i)$ and $\{\sigma_l(i)\}$, which are nonzero if that particular monomer is occupied by a condensed counterion or a ring with $l$ surfactants, respectively. The free energy of $N_p$ isolated complexes can then be written as $$\label{eq1} \beta F_{\rm complex} = - N_p\ln \sum_{\nu}^{} e^{-\beta E_{\nu} } \, ,$$ where the sum is over all possible configurations of counterions and surfactants along a complex. For a particular configuration $\nu$, the energy can be expressed as the sum of three terms, $E_{\nu}=E_1+E_2+E_3$. The first one is the electrostatic contribution arising from the Coulombic interactions between all charged sites of a complex, $$\label{eq2} E_1 = \frac {q^2} {2} \sum_{i \neq j}^Z \frac {[ - 1 + \sigma_c(i) + \sum_l^{l_{max}} l \sigma_l(i)] [ - 1 + \sigma_c(j) + \sum_l^{l_{max}} l \sigma_l(j)]} {D |r(i)-r(j)|} \, ,$$ where we have assumed that the only effect of association is the renormalization of the effective charge of each monomer. The second term $E_2$, is due to hydrophobic interactions between the surfactant molecules, $$\label{eq3} E_2 = \frac{\chi}{2} \sum_{\langle i,j \rangle}^Z\sum_{l,l'=1}^{l_{max}} \frac {(l+l')} {2} \sigma_l(i) \sigma_{l'}(j) \; ,$$ where, in order to simulate the short-ranged nature of hydrophobic interactions, the first sum is constrained to run over the nearest neighbors. The hydrophobicity parameter $\chi$ is negative, representing the tendency of the two adjacent surfactant molecules to expel water. We can estimate its value from the experimental measurement of the energy necessary to remove an amphiphile from a monolayer and place it in the bulk [@Isra]. The third contribution $E_3$, accounts for the internal energy of each ring, $$\label{eq4} E_3 = \sum_{i}^Z \sum_{l=2}^{l_{max}}\sigma_l(i)E_l \; .$$ $E_l$ is the interaction energy between $l$ surfactants forming a ring. Each ring contains a maximum of $l_{max}$ sites, which can be occupied by surfactants. To each one of these sites we associate an occupation variable $\tau(j)$, which is zero if site $j$ is unoccupied by a surfactant and is one if it is occupied ( see Fig. $3$). The interaction energy of surfactants forming a ring can then be written as $$\label{eq5} E_l=\frac{q^2}{2 D}\sum_{i\neq j}^{l_{max}}\frac{\tau(i)\tau(j)} {2 a \sin(\pi|i-j|/l_{max})}+ \frac {\chi} {2} \sum_{\langle i,j \rangle}^{l_{max}} \tau(i) \tau(j) \; .$$ The first term of Eq.(\[eq5\]) is due to electrostatic repulsion between the surfactant head groups, while the second is the result of attraction between the adjacent hydrocarbon tails. The exact solution of even this simpler sub-problem (i.e. evaluation of the sum in Eq.(\[eq1\])) is very difficult due to the long ranged electrostatic interactions. We shall, therefore, resort to mean-field theory, which works particularly well for long-ranged potentials. Evaluating the upper bound for the free energy, given by the Gibbs-Bogoliubov inequality, and neglecting the end effects we obtain, $$\begin{aligned} \label{eq6} \beta F_{\rm complex} & = &\beta N_p[ f_{el}+ f_{hyd}+ f_{ring}+ f_{mix}] \; .\end{aligned}$$ The first term, $$\begin{aligned} \label{eq7} \beta f_{el}&=& \xi S [-1+ p_c + \sum_{l=1}^{l_{max}}l p_l]^2-\xi S N_p \; ,\end{aligned}$$ is the electrostatic interaction between the sites along one rod and is related to $E_1$. $S$ is expressed in terms of the digamma function [@GR], $$\begin{aligned} \label{eq8} S&=&Z[\Psi(Z)-\Psi(1)]-Z+1 \; ,\end{aligned}$$ and $\xi \equiv \beta q^2/Db$ is the Manning parameter [@Man1; @Joan]. The second term in Eq. (\[eq6\]), $$\begin{aligned} \label{eq9} \beta f_{hyd}&=&\beta \chi (Z-1) \sum_{n,m}^{l_{max}} \frac{(n+m)}{2}p_m p_n \; ,\end{aligned}$$ is the hydrophobic attraction between the rings inside a complex. The third term is the free energy due to the electrostatic and hydrophobic interactions between the surfactants forming a ring, $$\begin{aligned} \label{eq10} \beta f_{ring} & = & \frac{2\ln l_{max} + \nu_0}{4\pi T^*} \sum_{l=2}^{l_{max}}Zp_l l^2 + \frac {\beta \chi} {l_{max}} \sum_{l=2}^{l_{max}} Z p_l l^2 + \\ && \sum_{l=1}^{l_{max}} Z p_l l \ln \left( \frac {l} {l_{max}} \right) + \sum_{l=1}^{l_{max}} Zp_l l_{max} \left( 1-\frac{l}{l_{max}} \right) \ln \left( 1-\frac{l}{l_{max}} \right) \nonumber \end{aligned}$$ where $\nu_0\approx 0.25126591$, and the reduced temperature is $T^*=k_BTDa/q^2$. Finally, the free energy of mixing for rings and counterions of a complex is, $$\begin{aligned} \label{eq11} \beta f_{mix}&=&Z (1-p_c-\sum_{l}^{l_{max}}p_l)+ \ln (1-p_c-\sum_{l=1}^{l_{max}}p_l) +Z p_c \ln p_c + \nonumber \\ && Z\sum_{l=1}^{l_{max}}p_l \ln p_l- Z p \ln l_{max} + \\ && Z p l_{max} \left(1 - \frac {1} {l_{max}} \right) \ln \left(1 - \frac {1} {l_{max}} \right) \nonumber \; , \end{aligned}$$ where to be consistent with the expression (\[eq10\]), we have included a contribution to the free energy arising from the azimuthal motion of condensed counterions around the polyion, [*i.e.*]{} the last two terms of Eqn.(\[eq11\]). Once a cluster, constructed in isolation, is introduced into solution, it gains an additional solvation energy due to its interaction with other clusters, free counterions, free coions, and free surfactants. The electrostatic repulsion between the complexes is screened by the ionic atmosphere, producing an effective short ranged potential of DLVO form [@Der]-[@FisLevLi]. The electrostatic free energy due to interactions between various clusters can be estimated from the second virial coefficient, $$\begin{aligned} \label{eq12} \beta F^{cc} = (Z - n_c - n_s)^2 \frac {2 \pi N_p^2 a^3 e^{-2 \kappa a}} {V T^* (\kappa a)^4 K_1^2(\kappa a)} \, ,\end{aligned}$$ where $ (\kappa a)^2 \equiv 4 \pi \rho^*_1/ T^*$ and $\rho_1^* \equiv a^3[\rho_p(Z-n_s-n_c)+2\rho_m+2\rho_s]$ is the reduced density of free ions. The free energy due to interaction between the complexes and free ions and surfactants can be obtained following the general methodology of the Debye-H[ü]{}ckel-Bjerrum theory[@FL1; @FL2], [@DH1]-[@Surfoplex], $$\begin{aligned} \label{eq14} \beta F^{ci} & = & N_p (Z - n_c - n_s)^2 \frac {(a/L)} {T^* (\kappa a)^2} \times \nonumber \\ & & \times \left[ -2 \ln (\kappa a K_1(\kappa a)) + I(\kappa a) - \frac {(\kappa a)^2} {2} \right] \, ,\end{aligned}$$ with $$\label{eq15} I(\kappa a) = \int^{\kappa a}_0 \frac {x K_0^2(x)} {K_1^2(x)} dx \, ,$$ where $K_n$ is the modified Bessel function of order $n$. The contribution to the total free energy arising from the interactions between the free ions and surfactants is given by the usual Debye-Hückel expression [@DH1],[@DH2] $$\label{eq13} \beta F^{ii} = - \frac {V} {4 \pi a_c^3} \left[ \ln(1 + \kappa a_c) - \kappa a _c + \frac {(\kappa a_c)^2} {2} \right] \, .$$ This term is very small and is included only for completeness. The last contribution to the total free energy Eq. (\[eq0\]), results from the entropic mixing of the counterions, coions, surfactant and complexes, $$\label{eq16} F_{mixing}=F_{m+}+F_{m-}+F_{s}+F_{c} \, .$$ The free energy of mixing is obtained following the general ideas introduced by Flory [@Flo], $$\begin{aligned} \label{eq17} \beta F_{m+} & = & N_{m+} \ln \phi_{m+} - N_{m+} \, , \nonumber \\ \beta F_{m-} & = & N_{m-} \ln \phi_{m-} - N_{m-} \, , \nonumber \\ \beta F_{s} & = & N_{s} \ln (\phi_{s}/n_s) - N_{s} \, , \nonumber \\ \beta F_{c} & = & N_{p} \ln \left( \frac {(Z + n_c + n_s) \phi_{c}} {Z + n_c + n_s s} \right) - N_p \, .\end{aligned}$$ In the above expression $m+$ denotes free counterions, $m-$ free coions, $s$ free surfactant molecules, and $c$ complexes. The $$\begin{aligned} \label{eq18} & & \phi_{m+} = \frac {\pi \rho^*_+} {6} \left( \frac {a_c} {a} \right)^3 \, , \nonumber \\ & & \phi_{m-} = \frac {\pi \rho^*_-} {6} \left( \frac {a_c} {a} \right)^3 \, , \nonumber \\ & & \phi_{s} = \frac {s \pi \rho_s^{f*}} {6} \left( \frac {a_c} {a} \right)^3 \, , \nonumber \\ & & \phi_{c} = \pi \rho_{p}^* \left[ \frac {1} {4 (a/L)} \left( \frac {a_p} {a} \right)^2 + \frac {1} {6} (n_c + n_s s) \left( \frac {a_c} {a} \right)^3 \right] \; \end{aligned}$$ are the volume fractions occupied by the free counterions, coions, surfactants, and complexes, respectively. Results and Conclusions ======================= The equilibrium configuration of the polyelectrolyte-surfactant solution is determined by the requirement that the Helmholtz free energy be minimum. Since $F$ is the function of $n_s,n_c$, and the surface coverage by rings $\{p_l\}$, minimization of $F$ implies that $$\begin{aligned} \label{eq19} \delta F=\frac{\partial F}{\partial n_s}\delta n_s+ \frac{\partial F}{\partial n_c}\delta n_c+ \sum_{l=1}^{l_{max}}\frac{\partial F}{\partial p_l}\delta p_l=0\;\; .\end{aligned}$$ Using the constraint Eq. (\[eq00\]), Eq. (\[eq19\]) can be separated into $l_{max}+1$ equations, $$\begin{aligned} \label{eq20} \frac{\partial F}{\partial n_c}=0\end{aligned}$$ and $$\begin{aligned} \label{eq21} \frac{\partial F}{\partial n_s}Zl+ \frac{\partial F}{\partial p_l}=0 \; ; \; \; l=1...l_{max}.\end{aligned}$$ The system of equations (\[eq20\]) and (\[eq21\]) can, in principle, be solved numerically. However, for reasonable values of $l_{max}$ this requires a significant numerical effort. Instead of pursuing this brute force method, we note that to a reasonable accuracy, the surface coverage by rings, $\{p_l\}$, can be approximated by an exponential distribution[@Lev2], $$\begin{aligned} \label{eq22} p_l = \frac {n_s e^{\alpha l}} {Z \sum_{l=1}^{l_{max}} l e^{\alpha l}} \; .\end{aligned}$$ We have checked that this is, indeed, a good approximation by numerically solving Eq. (\[eq21\]) for an isolated complex. Using ansatz (\[eq22\]), the total free energy becomes a function of $n_c$, $n_s$, and $\alpha$. For a fixed volume and number of particles the equilibrium corresponds to the minimum of Helmholtz free energy, $$\begin{aligned} \label{eq23} \frac{\partial F}{\partial n_c}=0 \, , \\ \frac{\partial F}{\partial n_s}=0 \, , \\ \frac{\partial F}{\partial \alpha}=0 \; .\end{aligned}$$ These are three coupled algebraic equations, which can be easily solved numerically to yield the characteristic number of condensed counterions, surfactants, as well as the shape of the distribution of ring sizes $(\alpha)$. In Fig. 4 we present a numerical solution of these equations. As a specific example we consider a cationic surfactant with an alkyl chain of $s=12$ groups. In this case the hydrophobicity parameter can be estimated [@Surf] to be in the range of $\chi \approx -3,5 k_B T$. To explore the dependence of condensation on the hydrophobicity of surfactant, we shall vary this value within reason. The density of monovalent salt and the $DNA$ is taken to be $18\;mM$ and $2 \times 10^{-3}mM$, respectively. The resulting binding isotherms are illustrated in Fig. $4$. The fraction of associated amphiphilic molecules $\beta_s=n_s/Z$, is plotted against the density of surfactant for a fixed amount of monovalent salt, $\rho_m$. For small concentrations of cationic surfactant, few amphiphilic molecules associate with the $DNA$ segments. At the certain critical concentration, however, the system forms [*surfoplexes*]{} [@Surf] [@Surfoplex] — complexes in which the charge of the $DNA$ is almost completely neutralized by the associated amphiphiles. If the density is increased further, on average, more than one surfactant molecule will associate to each phosphate group, leading to charge inversion of the surfoplexes. For highly hydrophobic surfactants the charge inversion can happen very close to the cooperative binding transition. We note that our theory predicts the binding transition to be discontinuous, this, most likely, is an artifact of the mean-field approximation [@Surfoplex]. We have presented a simple theory of $DNA-surfactant$ solutions. Our results should be of direct interest to researchers working on design of improved gene delivery systems. In particular we find that addition of cationic surfactants leads to a strong cooperative binding transition. This transition happens far bellow the critical micell concentration. A further increase of amphiphile density can result in the charge inversion of the $DNA-surfactant$ complexes. This regime should be particularly useful in designing gene or oligonucleotide delivery systems. Until now most of non-viral gene-delivery systems were in the form of lipoplexes — complexes formed by $DNA$ and cationic liposomes. To form the liposomes, however, is required a significant concentration of cationic lipid. Unfortunately, at high concentrations both lipids and surfactants are toxic to organism. Our model suggests that the charge inversion can be achieved with quite small concentration of cationic amphiphile, [*if*]{} it is sufficiently hydrophobic. This should reduce the risk of unnecessary medical complications. [**ACKNOWLEDGMENTS**]{} This work was supported in part by Conselho Nacional de Desenvolvimento Cient[í]{}fico e Tecnol[ó]{}gico and Financiadora de Estudos e Projetos, Brazil. [40]{} T. Friedmann, [*Sci. Am.*]{} [**276**]{}, 80 (1997). P. L. Felgner, [*Sci. Am.*]{} [**276**]{}, 86 (1997). P. L. Felgner and G. M. Ringold, [*Nature*]{} [**337**]{}, 387-388 (1989). P. L. Felgner and G. Rhodes, [*Nature*]{} [**349**]{}, 351 (1991). I. M. Verma and N. Somia, [*Nature*]{} [**389**]{} 239 (1997). J. O. Rädler [*et al.*]{}, [*Science*]{} [**275**]{}, 810 (1997). D. Harries [*et al.*]{}, [*Biophys. J.*]{} [**75**]{}, 159 (1998). W. F. Anderson, [*Nature*]{}, [**392**]{}, 25, (1998). Suppl. A. V. Gorelov [*et al.*]{}, [*Physica*]{} [**A249**]{}, 216 (1998). K. Shirahama [*et al.*]{}, [*Bull. Chem. Soc. Jpn.*]{} [**60**]{}, 43 (1987). M.J. Hope, B. Mui, S. Ansell, and Q. F. Ahkong, [*Mol. Membrane Biol.* ]{} [**15**]{}, 1 (1998). S. Alexander [et. al]{}, [*J. Chem. Phys.*]{} [**80**]{}, 5776 (1984). M. E. Fisher and Y. Levin, [*Phys. Rev. Lett.*]{} [**71**]{}, 3826 (1993). Y. Levin and M. E. Fisher, [*Physica*]{} [**A225**]{}, 164 (1996). Y. Levin, [*Europhys. Lett.*]{} [**34**]{}, 405 (1996). Y. Levin and M.C. Barbosa, [*J. Phys. II (France)*]{} [**7**]{}, 37 (1997). J. N. Israelachvili, D. Mitchell, and B. W. Ninham, [*J. Chem. Soc., Faraday Trans.*]{} [**72**]{}, 1525 (1976). I. S. Gradshteyn and I. M. Ryzhik, [*Table of integrals series and products*]{}, New York: Academic Press (1965). G. S. Manning, [*J. Chem. Phys.*]{} [**51**]{}, 924 (1969). J. L. Barrat and J.F. Joanny, [*Adv. Chem. Phys.*]{} [**94**]{}, 1 (1996). B. V. Derjaguin and L. Landau, [*Acta Phys. (USSR)*]{}, [**14**]{}, 633 (1941). E. J. W. Verwey and J. Th. G. Overbeek, [*Theory of the stability of lyophobic colloids*]{}, Amsterdam: Elsevier (1948). M. Medina-Noyola and D. A. McQuarrie, [*J. Chem. Phys.*]{} [**73**]{}, 6279 (1980). X.-J. Li, Y. Levin, and M. E. Fisher, [*Europhys. Lett.*]{} [**26**]{},683 (1994). M. E. Fisher, Y. Levin, and X.-J. LI, [*J. Chem. Phys.*]{} [**101**]{}, 2273 (1994). P. W. Debye and E. H[ü]{}ckel, [*Phys. Z.*]{} [**24**]{}, 185 (1923). D. A. McQuarrie, [*Statistical mechanics*]{}, New York: Harper and Row (1976). N. Bjerrum, [*Kgl. Dan. Vidensk. Selsk. Mat.-Fys. Medd.*]{}, 7 [1926]{}. P. S. Kuhn, Y. Levin, and M. C. Barbosa, [*Macromolecules*]{} [**31**]{}, 8347, (1998). P. S. Kuhn, Y. Levin, and M. C. Barbosa, [*Chem. Phys. Lett.*]{} [**298**]{}, 51 (1998). P. S. Kuhn, Y. Levin, and M. C. Barbosa, [*Physica*]{} [**A266**]{}, 413, (1999). P. S Kuhn, M. C. Barbosa, and Y. Levin, [*Physica*]{} [**A269**]{}, 278 (1999). Flory, P. [*Principles of Polymer Chemistry*]{} (Cornell University Press, Ithaca, New York, 1971). Y. Levin, [*Phys. Rev. Lett.*]{} [**83**]{}, 1159 (1999). **FIGURE CAPTION** Figure 1. A cylindrical polyion of diameter $a_p$, length $L$, and charge $-Zq$, surrounded by spherical ions of radius $a_c$ and amphiphilic molecules of $s$ monomers. Each monomer of a macroion is free or has [*one*]{} counterion, [**or**]{} a ring made of $l$ amphiphilic molecules associated with it. Figure 2. Schematic representation of a complex. Empty sites (monomers) ($-$), sites with associated counterion ($c$), sites with $l$ associated amphiphiles ($s_l$). Figure 3. Ring composed of $l$ surfactant molecules, $l_{max}=15$. Figure 4. Effective binding fraction of amphiphiles $\beta_{s}\equiv n_s/Z$, as a function of amphiphile concentration $\rho_s$. The concentrations of DNA and of added salt is $2 \times 10^{-6}M$ and $18 mM$, respectively. The length of the DNA segments is $220$ base pairs. The solvent is water at room temperature, so that $\xi=4.17$. Figure 5. Average size of rings in a complex (parameters the same as in Fig. 4). [^1]: Corresponding author: levin@if.ufrgs.br

--- abstract: 'In this contribution, we discuss the possible importance of continuum-coupling (or threshold) effects in heavy quarkonium spectroscopy. Our calculations are carried out in a coupled-channel model, where meson-meson higher Fock (or molecular-type) components are introduced in $Q \bar Q$ bare meson wave functions by means of a pair-creation mechanism. After providing a quick resume of the main characteristics of the coupled-channel model, we briefly discuss its application to the calculation of the masses of heavy quarkonium-like $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$ states with threshold corrections. We show that the introduction of pair-creation effects in the Quark Model (QM) formalism makes it possible to explain the deviations of $\chi_{\rm c}(2P)$ states’ masses from the experimental data, without affecting the good QM description of the properties of $\chi_{\rm b}(3P)$ states.' author: - 'J. Ferretti' title: 'Threshold effects in heavy quarkonium spectroscopy[^1]' --- Introduction ============ The Quark Model (QM) formalism provides a good overall description of meson and baryon observables, including the spectrum (especially the lower-energy part) [@Eichten:1974af; @Godfrey:1985xj; @Capstick:1986bm; @Iachello:1991re; @Bijker:1994yr; @Ferraris:1995ui; @Ferretti:2011zz], the open-flavor strong decay amplitudes [@Bijker:1994yr; @Micu:1968mk; @LeYaouanc:1972vsx; @Kokoski:1985is; @Capstick:1993kb; @Barnes:2005pb; @Strong2015; @Ferretti:2015rsa], the nucleon electromagnetic form factors [@Iachello:1972nu; @Santopinto:1998ma; @Sanctis:2000eg; @DeSanctis:2011zz], and so on. However, some difficulties emerge when one moves to higher energies, specifically to higher-lying meson (baryon) radial excitations. One of the main problems, both from the theoretical and experimental point of view, is related to the emergence of [*exotic*]{} candidates. These are mesons (baryons) characterized by an unconventional (non-$q \bar q$/non-$qqq$) quark structure and/or non-standard quantum numbers. Some examples include the $X(3872)$ \[now $\chi_{\rm c1}(3872)$\] [@Choi:2003ue; @Acosta:2003zx; @Abazov:2004kp; @Tanabashi:2018oca], $X(4140)$ [@Aaltonen] and $X(4260)$ mesons [@Choi:2007wga]. From a theoretical/phenomenological point of view, the previous exotic meson candidates can be described variously. Some of the main interpretations include compact tetraquark states [@Jaffe:1976ih; @Barbour:1979qi; @Weinstein:1983gd; @SilvestreBrac:1993ss; @Brink:1998as; @Maiani:2004vq; @Barnea:2006sd; @Santopinto:2006my; @Ebert:2008wm; @Deng:2014gqa; @Zhao:2014qva; @Anwar:2017toa], meson-meson molecules [@Weinstein:1990gu; @Manohar:1992nd; @Tornqvist:1993ng; @Martins:1994hd; @Swanson:2003tb; @Hanhart:2007yq; @Thomas:2008ja; @Baru:2011rs; @Valderrama:2012jv; @Aceti:2012cb; @Guo:2013sya], the result of kinematic or threshold effects caused by virtual particles [@Heikkila:1983wd; @Pennington:2007xr; @Li:2009ad; @Danilkin:2010cc; @Ortega:2012rs; @Ferretti:2013faa; @Ferretti:2013vua; @Achasov:2015oia; @Kang:2016jxw; @Lu:2016mbb; @Ferretti:2018tco], or hadro-quarkonia (hadro-charmonia) [@Dubynskiy:2008mq; @Guo:2008zg; @Wang:2009hi; @Voloshin:2013dpa; @Li:2013ssa; @Wang:2013kra; @Brambilla:2015rqa; @Alberti:2016dru; @Panteleeva:2018ijz; @Ferretti:2018kzy]. More precise experimental informations on those states, as well as a deeper understanding of the quarkonium spectrum and its patterns, will make it possible to rule out those interpretations which are not compatible with the experimental data [@Tanabashi:2018oca]. For recent reviews on exotics, see Refs. [@Esposito:2016noz; @Olsen:2017bmm; @Guo:2017jvc]. In this contribution, we focus on the interpretation of the previously mentioned $X$-type exotic mesons as the result of threshold effects caused by virtual particles. To do that, we make use of the Unquenched Quark Model (UQM) formalism with some modifications [@Ferretti:2018kzy]. The UQM is an extension of the QM. It makes it possible to include the effects of virtual $q \bar q$ pairs in the naïve QM formalism by means of a $q \bar q$ pair-creation mechanism [@Micu:1968mk; @LeYaouanc:1972vsx; @Heikkila:1983wd; @Ferretti:2012zz; @Geiger:1991qe; @Bijker:2009up; @Bijker:2012zza]. Above meson-meson thresholds, the creation of $q \bar q$ pairs from the vacuum is responsible of the open-flavor strong decays of the meson of interest [@Micu:1968mk; @LeYaouanc:1972vsx; @Kokoski:1985is; @Barnes:2005pb; @Ferretti:2015rsa]. Below threshold, it is responsible of the coupling between the meson of interest and meson-meson continuum (or molecular-type) components [@Heikkila:1983wd; @Ferretti:2012zz; @Geiger:1991qe]. After discussing our modifications to the “standard" version of the UQM, including some extra hypotheses to make the UQM calculations converge and remove unphysical results, we show that the introduction of meson-meson continuum components in the bare wave function of the meson of interest can provide a non-negligible correction to the meson energy and the emergence of significant continuum components in its wave function [@Ferretti:2018kzy]. Other possible applications of the previous coupled-channel formalism are also briefly summarized. Unquenching the quark model =========================== The procedure for “unquenching the Quark Model" consists in the introduction of higher Fock components in quark-antiquark bare meson wave functions, $$\label{eqn:QQ-WF} \left| Q \bar Q \right\rangle \rightarrow \left| Q \bar Q \right\rangle + \left| Q \bar q - q \bar Q \right\rangle + \left| Q \bar Q g \right\rangle + ... \mbox{ }.$$ Here, $g$ is a constituent gluon, $\left| Q \bar q - q \bar Q \right\rangle$ a tetraquark or meson-meson molecular-type component, and $\left| Q \bar Q g \right\rangle$ a hybrid one. The first step, namely the introduction of $\left| Q \bar q - q \bar Q \right\rangle$ components in heavy quarkonium-like meson spectroscopy, has already been carried out and some observables have been calculated [@Heikkila:1983wd; @Pennington:2007xr; @Ortega:2012rs; @Ferretti:2013faa; @Ferretti:2013vua; @Lu:2016mbb; @Ferretti:2018tco; @Ferretti:2012zz; @Ferretti:2014xqa]. For the introduction of $\left| Q \bar Q g \right\rangle$ components in quarkonium spectroscopy, see Ref. [@LeYaouanc:1984gh]. If we restrict the extra terms of Eq. (\[eqn:QQ-WF\]) to molecular-type components, in the Unquenched Quark Model (UQM) formalism the quarkonium-like meson wave function can be written as $$\label{eqn:Psi-A} \footnotesize \begin{array}{l} \left| \psi_A \right\rangle = {\cal N} \left[ \left| A \right\rangle + \displaystyle \sum_{BC \ell J} \int q^2 dq \left| BC q \ell J \right\rangle \frac{ \left\langle BC q \ell J \right| T^{\dagger} \left| A \right\rangle}{M_A - E_B - E_C} \right] ~. \end{array}$$ Here, ${\cal N}$ is a normalization factor, $\left| \psi_A \right\rangle$ is the superposition of a zeroth order quark-antiquark configuration, $\left| A \right\rangle$, plus a sum over all the possible higher Fock components, $\left| BC \right\rangle$, due to the creation of quark-antiquark pairs with vacuum quantum numbers. The sum is extended over a complete set of intermediate meson-meson states, $\left| BC \right\rangle$, with energies $E_{B,C} = \sqrt{M_{B,C}^2 + q^2}$; $M_A$ is the physical mass of the meson $A$; $q$ and $\ell$ are the relative radial momentum and orbital angular momentum of $B$ and $C$, and $J$ is the total angular momentum, with ${\bf J} = {\bf J}_B + {\bf J}_C + {\bm \ell}$. The symbol $T^{\dagger}$ in Eq. (\[eqn:Psi-A\]) stands for the pair-creation operator of Refs. [@Ferretti:2015rsa; @Ferretti:2013faa; @Ferretti:2013vua; @Ferretti:2018tco; @Ferretti:2012zz; @Ferretti:2014xqa; @Strong2015; @Santopinto:2016fgs; @Garcia-Tecocoatzi:2016rcj]. See also Refs. [@Geiger:1991qe; @Bijker:2009up; @Bijker:2012zza]. Below threshold, $T^{\dagger}$ is responsible of the coupling between a bare meson, $\left| A \right\rangle$, and meson-meson continuum components, $\left| BC \right\rangle$; above threshold, it is responsible of $A \rightarrow BC$ open-flavor strong decays, which proceed via the creation of a light $q \bar q$ pair (with $q = u$, $d$ or $s$) from the vacuum. Given this, in the UQM the expectation value of a meson observable, $\hat {\mathcal O}_{\rm m}$, on the quarkonium-like states of Eq. (\[eqn:Psi-A\]) is computed as $$\left\langle \psi_A \right| \hat {\mathcal O}_{\rm m} \left| \psi_A \right\rangle = \left\langle \hat {\mathcal O}_{\rm m} \right\rangle_{\rm val} + \left\langle \hat {\mathcal O}_{\rm m} \right\rangle_{\rm cont} \mbox{ },$$ where $\left\langle \hat {\mathcal O}_{\rm m} \right\rangle_{\rm val}$ and $\left\langle \hat {\mathcal O}_{\rm m} \right\rangle_{\rm cont}$ stand for the expectation values of $\hat {\mathcal O}_{\rm m}$ on the valence, $\left| A \right\rangle$, and continuum components, $\left| BC \right\rangle$, respectively. As an example, we describe the procedure to calculate the masses of quarkonium-like states with self-energy corrections. The physical meson masses are related to the bare and self-energies via $$\label{eqn:Ma-UQM} M_A = E_A + \Sigma(M_A) \mbox{ }.$$ Here, $E_A$ are the bare energies of the meson $A$, which have to be computed in a specific quark model; for example, we make use of the relativized QM of Ref. [@Godfrey:1985xj]. These energies are calculated by considering mesons as the bound states of a constituent quark-antiquark pair bounded by one-gluon-exchange forces. $\Sigma(M_A)$ are the self-energy corrections to the bare meson masses, resulting from the coupling between the bare, $\left| A \right\rangle$, and the continuum components, $\left| BC \right\rangle$. They can be written as $$\label{eqn:self-a} \Sigma(M_A) = \sum_{BC} \int_0^{\infty} q^2 dq \mbox{ } \frac{\left| \left\langle BC q \ell J \right| T^\dag \left| A \right\rangle \right|^2}{M_A - E_B - E_C} \mbox{ },$$ where the sum is extended over a complete set of intermediate meson-meson states $BC$. One can also calculate the norm of the continuum (or molecular-type) component of a quarkonium-like state via [@Ferretti:2018tco; @Ferretti:2012zz] $$\label{eqn:Pa-sea} P_A^{\rm sea} = \sum_{BC\ell J} \int_0^\infty q^2 dq \mbox{ } \frac{\left| \left\langle BC q \, \ell J \right| T^\dag \left| A \right\rangle \right|^2}{(M_A - E_B - E_C)^2} \mbox{ },$$ where the probability to find the meson in its valence component, $P_A^{\rm val}$, is given by $P_A^{\rm val} = 1 - P_A^{\rm sea}$. The UQM formalism has been extensively used in the past to compute both baryon and meson observables, including the calculation of baryon [@Garcia-Tecocoatzi:2016rcj; @SilvestreBrac:1991pw; @Morel:2002vk] and meson [@Heikkila:1983wd; @Pennington:2007xr; @Ferretti:2013faa; @Ferretti:2013vua; @Lu:2016mbb; @Ferretti:2018tco; @Ferretti:2012zz] masses with self-energy corrections, heavy quarkonium hidden flavor strong decays [@Ferretti:2018tco], and the strangeness contribution to the nucleon electromagnetic form factors [@Bijker:2012zza; @Geiger:1996re]. Despite of its merits, including its simplicity and versatility, the UQM calculations do not converge quickly. Indeed, it can be easily shown that, as the tower of meson-meson intermediate states $\left| BC \right\rangle$ is enlarged, the contribution of continuum or sea components to hadron observables keeps growing larger and larger. Below, we discuss a simple procedure to “renormalize" the UQM results. More details can be found in Ref. [@Ferretti:2018tco]. A COUPLED-CHANNEL MODEL FOR HEAVY QUARKONIUM-LIKE STATES ======================================================== After discussing the main features of the UQM formalism for mesons [@Ferretti:2015rsa; @Ferretti:2013faa; @Ferretti:2013vua; @Ferretti:2012zz; @Ferretti:2014xqa], here we show a procedure to “renormalize" it and avoid the production of unphysical results [@Ferretti:2018tco]. In particular, as a first step, we give a brief resume of a simple coupled-channel method to compute the physical masses of quarkonium-like mesons, $M_A$, with threshold corrections [@Ferretti:2018tco]. The method is based on the UQM formalism, plus the following hypotheses and prescriptions: a) The method is not used to perform a global fit to the heavy quarkonium spectrum, but it is applied only to specific meson multiplets, like $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$; b) Only the closest complete set of accessible SU(N)$_{\rm flavor} \otimes$ SU(2)$_{\rm spin}$ open-flavor meson-meson intermediate states (e.g. $1S1S$, $1S1P$ or $1S2S$) can influence the multiplet structure. The other (lower or upper) meson-meson thresholds, which are further in energy, are supposed to give some kind of global or background contribution, which can be subtracted; c) The presence of a certain complete set of open-flavor intermediate states does not affect the properties of a single resonance, but it influences those of all the multiplet members. Thus, the net effect of the intermediate states on a quarkonium-like meson multiplet is similar to that of a spin-orbit or hyperfine splitting. State $E_A$ \[MeV\] $\Sigma(M_A) - \Delta$ \[MeV\] $M_A^{\rm th}$ \[MeV\] $M_A^{\rm exp}$ \[MeV\] --------------------- --------------- -------------------------------- ------------------------ ------------------------- $h_{\rm c}(2P)$ 3956 $-16$ 3940 – $\chi_{\rm c0}(2P)$ 3916 0 3916 3918 $\chi_{\rm c1}(2P)$ 3953 $-65$ 3888 3872 $\chi_{\rm c2}(2P)$ 3979 $-30$ 3949 3927 $h_{\rm b}(3P)$ 10541 $-4$ 10538 10519$^\dag$ $\chi_{\rm b0}(3P)$ 10522 0 10522 10500$^\dag$ $\chi_{\rm b1}(3P)$ 10538 $-2$ 10537 10512 $\chi_{\rm b2}(3P)$ 10550 $-7$ 10543 10528$^\dag$ Threshold mass-shifts --------------------- Under the previous hypotheses, the physical masses of the members of a quarkonium-like meson multiplet are computed as [@Ferretti:2018tco] $$\label{eqn:new-Ma} M_A = E_A + \Sigma(M_A) + \Delta \mbox{ },$$ where $E_A$ is the bare mass of meson $A$, whose value is extracted from the relativized QM predictions of Refs. [@Godfrey:1985xj; @Barnes:2005pb; @Godfrey:2015dia]. It is worth noting that here and in Ref. [@Ferretti:2018tco], contrary to the calculations of Refs. [@Ferretti:2013faa; @Ferretti:2013vua], the relativized QM parameters are not fitted to the reproduction of the physical masses of Eq. (\[eqn:Ma-UQM\]). The bare mass values are directly extracted from the original relativized QM fit of Ref. [@Godfrey:1985xj]. See also Refs. [@Barnes:2005pb; @Godfrey:2015dia]. The second term in Eq. (\[eqn:new-Ma\]) is the self-energy correction of Eq. (\[eqn:self-a\]). In the case of heavy quarkonium-like states around the opening of the first meson-meson decay thresholds, the closest complete set of meson-meson intermediate states is made up of $1S1S$ open-flavor mesons. For example, in the case of the $\chi_{\rm c}(2P)$ multiplet, we consider $D \bar D$, $D \bar D^*$, $D^* \bar D^*$, $D_{\rm s} \bar D_{\rm s}$, $D_{\rm s} \bar D_{\rm s}^*$, $D_{\rm s}^* \bar D_{\rm s}^*$, $\eta_{\rm c} \eta_{\rm c}$, $\eta_{\rm c} J/\psi$, and $J/\psi J/\psi$ meson-meson components [@Ferretti:2018tco]. The pair-creation model parameters, which we need in the calculation of the $\left\langle BC q \ell J \right| T^\dag \left| A \right\rangle$ vertices of Eq. (\[eqn:self-a\]), were fitted to the open-flavor strong decays of charmonia [@Ferretti:2013faa Table II] and bottomonia [@Ferretti:2013vua Table I]; see also [@Ferretti:2015rsa Table II]. Therefore, for each multiplet, there is only one free parameter, $\Delta$. This is the smallest self-energy correction (in terms of absolute value) to the bare mass of a multiplet member; see [@Ferretti:2018tco Sec. 2] and the following section. THRESHOLD MASS-SHIFTS IN $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$ MULTIPLETS ============================================================================= We calculate the threshold mass shifts of the $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$ multiplet members due to a complete set of ground state $1S 1S$ meson loops, like $D \bar D$, $D \bar D^*$ ($B \bar B$, $B \bar B^*$), and so on [@Ferretti:2018tco]. The values of the bare masses, $E_A$, are extracted from the relativized model [@Godfrey:1985xj], those of the physical masses, $M_A$, from the PDG [@Tanabashi:2018oca]. The self-energy corrections, $\Sigma(M_A)$, are computed according to Eq. (\[eqn:self-a\]), using the same pair-creation model parameter values as Refs. [@Ferretti:2015rsa; @Ferretti:2013faa; @Ferretti:2013vua]. In the case of the $\chi_{\rm c}(2P)$ multiplet, we get: $\Sigma(M_{h_{\rm c}(2P)}) = -119$ MeV, $\Sigma(M_{\chi_{\rm c0}(2P)}) = -103$ MeV, $\Sigma(M_{\chi_{\rm c1}(2P)}) = -168$ MeV, $\Sigma(M_{\chi_{\rm c2}(2P)}) = -133$ MeV. Thus, $\Delta = \Sigma(M_{\chi_{\rm c0}(2P)})$. In the case of the $\chi_{\rm b}(3P)$ multiplet, we obtain: $\Sigma(M_{h_{\rm b}(3P)}) = -116$ MeV, $\Sigma(M_{\chi_{\rm b0}(3P)}) = \Delta = -112$ MeV, $\Sigma(M_{\chi_{\rm b1}(3P)}) = -114$ MeV, $\Sigma(M_{\chi_{\rm b2}(3P)}) = -119$ MeV. The values of the calculated physical masses of the $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$ multiplet members via Eq. (\[eqn:new-Ma\]) are reported in Table \[tab:ChiC(2P)-splittings\]. See also Fig. \[fig:ChiC(2P)-splittings\]. It is worth noting that: I) Our theoretical predictions agree with the data within the typical error of a QM calculation ($\sim 30-50$ MeV); II) Among the $\chi_{\rm c}(2P)$ multiplet members, the $\chi_{\rm c1}(2P)$ receives the largest contribution from the continuum. This continuum contribution is necessary to lower the relativized QM prediction, 3.95 GeV, towards the observed value of the meson mass, 3871.69 MeV [@Tanabashi:2018oca]; III) In the $\chi_{\rm c}(2P)$ case, threshold effects break the usual mass pattern of a $\chi$-type multiplet, namely $M_{\chi_0} < M_{\chi_1} \approx M_{\rm h} < M_{\chi_2}$; IV) The threshold effects are negligible in the $\chi_{\rm b}(3P)$ case. Because of this, we interpret $\chi_{\rm b}(3P)$ states as (almost) pure bottomonia; V) Unlike the $\chi_{\rm c}(2P)$ case, the usual mass pattern within a $\chi$-type multiplet, namely $M_{\chi_0} < M_{\chi_1} \approx M_{\rm h} < M_{\chi_2}$, in the $\chi_{\rm b}(3P)$ case is now respected. ![$\chi_{\rm c}(2P)$ multiplet: masses with threshold corrections. Yellow boxes, blue dashed and black continuous lines correspond to the experimental [@Tanabashi:2018oca], calculated bare and physical masses, respectively. The calculated bare and physical mass values are extracted from Table \[tab:ChiC(2P)-splittings\]; see also [@Ferretti:2018tco Fig. 1 and Table 1].[]{data-label="fig:ChiC(2P)-splittings"}](fig1){width="200pt"} CONCLUSIONS =========== We discussed the possible importance of continuum-coupling (or threshold) effects in heavy quarkonium spectroscopy. Our calculations were carried out in a coupled-channel model, where meson-meson higher Fock (or molecular-type) components were introduced in $Q \bar Q$ bare meson wave functions by means of a pair-creation mechanism. After providing a quick resume of the main characteristics of the coupled-channel model, we briefly discussed its application to the calculation of the masses of heavy quarkonium-like $\chi_{\rm c}(2P)$ and $\chi_{\rm b}(3P)$ states with threshold corrections. Our results are compatible with the present experimental data [@Tanabashi:2018oca]. This work was supported by US Department of Energy Grant No. DE-FG-02-91ER-40608. E. Eichten [*et al.*]{}, Phys. Rev. Lett. [**34**]{}, 369 (1975); Phys. Rev. D [**17**]{}, 3090 (1978). S. Godfrey and N. Isgur, Phys. Rev. D [**32**]{}, 189 (1985). S. Capstick and N. Isgur, Phys. Rev. D [**34**]{}, 2809 (1986). F. Iachello, N. C. Mukhopadhyay and L. Zhang, Phys. Rev. D [**44**]{}, 898 (1991); Phys. Lett.  B [**256**]{}, 295 (1991). R. Bijker, F. Iachello and A. Leviatan, Annals Phys.  [**236**]{}, 69 (1994); [**284**]{}, 89 (2000). M. Ferraris, M. M. Giannini, M. Pizzo, E. Santopinto and L. Tiator, Phys. Lett. B [**364**]{}, 231 (1995); M. M. Giannini, E. Santopinto and A. Vassallo, Eur. Phys. J. A [**12**]{}, 447 (2001); M. M. Giannini and E. Santopinto, Chin. J. Phys.  [**53**]{}, 020301 (2015). E. Santopinto, Phys. Rev. C [**72**]{}, 022201 (2005); J. Ferretti, A. Vassallo and E. Santopinto, Phys. Rev. C [**83**]{}, 065204 (2011); E. Santopinto and J. Ferretti, Phys. Rev. C [**92**]{}, 025202 (2015); M. De Sanctis, J. Ferretti, E. Santopinto and A. Vassallo, Eur. Phys. J. A [**52**]{}, 121 (2016). L. Micu, Nucl. Phys. B [**10**]{}, 521 (1969). A. Le Yaouanc, L. Oliver, O. Pene and J. C. Raynal, Phys. Rev. D [**8**]{}, 2223 (1973). R. Kokoski and N. Isgur, Phys. Rev. D [**35**]{}, 907 (1987). S. Capstick and W. Roberts, Phys. Rev. D [**47**]{}, 1994 (1993); [**49**]{}, 4570 (1994). T. Barnes, S. Godfrey and E. S. Swanson, Phys. Rev. D [**72**]{}, 054026 (2005). R. Bijker, J. Ferretti, G. Galatà, H. García-Tecocoatzi and E. Santopinto, Phys. Rev.  D [**94**]{}, 074040 (2016). J. Ferretti and E. Santopinto, Phys. Rev. D [**97**]{}, 114020 (2018). F. Iachello, A. D. Jackson and A. Lande, Phys. Lett.  [**43B**]{}, 191 (1973); R. Bijker and F. Iachello, Phys. Rev. C [**69**]{}, 068201 (2004). E. Santopinto, F. Iachello and M. M. Giannini, Eur. Phys. J. A [**1**]{}, 307 (1998). M. De Sanctis, M. M. Giannini, L. Repetto and E. Santopinto, Phys. Rev. C [**62**]{}, 025208 (2000); M. De Sanctis, M. M. Giannini, E. Santopinto and A. Vassallo, Phys. Rev. C [**76**]{}, 062201 (2007). M. De Sanctis, J. Ferretti, E. Santopinto and A. Vassallo, Phys. Rev. C [**84**]{}, 055201 (2011). S. K. Choi [*et al.*]{} \[Belle Collaboration\], Phys. Rev. Lett.  [**91**]{}, 262001 (2003). D. Acosta [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett.  [**93**]{}, 072001 (2004). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett.  [**93**]{}, 162002 (2004). M. Tanabashi [*et al.*]{} \[Particle Data Group\], Phys. Rev. D [**98**]{}, no. 3, 030001 (2018). T. Aaltonen [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett.  [**102**]{}, 242002 (2009). S. K. Choi [*et al.*]{} \[Belle Collaboration\], Phys. Rev. Lett.  [**100**]{}, 142001 (2008). R. L. Jaffe, Phys. Rev. D [**15**]{}, 281 (1977). I. M. Barbour and D. K. Ponting, Z. Phys. C [**5**]{}, 221 (1980); I. M. Barbour and J. P. Gilchrist, Z. Phys. C [**7**]{}, 225 (1981) Erratum: \[Z. Phys. C [**8**]{}, 282 (1981)\]. J. D. Weinstein and N. Isgur, Phys. Rev. D [**27**]{}, 588 (1983). B. Silvestre-Brac and C. Semay, Z. Phys. C [**57**]{}, 273 (1993). D. M. Brink and F. Stancu, Phys. Rev. D [**57**]{}, 6778 (1998). L. Maiani, F. Piccinini, A. D. Polosa and V. Riquer, Phys. Rev. D [**71**]{}, 014028 (2005). N. Barnea, J. Vijande and A. Valcarce, Phys. Rev. D [**73**]{}, 054004 (2006). E. Santopinto and G. Galatà, Phys. Rev. C [**75**]{}, 045206 (2007). D. Ebert, R. N. Faustov, V. O. Galkin and W. Lucha, Phys. Rev. D [**76**]{}, 114015 (2007); D. Ebert, R. N. Faustov and V. O. Galkin, Phys. Atom. Nucl.  [**72**]{}, 184 (2009). C. Deng, J. Ping and F. Wang, Phys. Rev. D [**90**]{}, 054009 (2014). L. Zhao, W. Z. Deng and S. L. Zhu, Phys. Rev. D [**90**]{}, 094031 (2014). M. N. Anwar, J. Ferretti, F. K. Guo, E. Santopinto and B. S. Zou, Eur. Phys. J. C [**78**]{}, 647 (2018); M. N. Anwar, J. Ferretti and E. Santopinto, Phys. Rev. D [**98**]{}, 094015 (2018). J. D. Weinstein and N. Isgur, Phys. Rev. D [**41**]{}, 2236 (1990). A. V. Manohar and M. B. Wise, Nucl. Phys. B [**399**]{}, 17 (1993). N. A. Törnqvist, Z. Phys. C [**61**]{}, 525 (1994); Phys. Lett. B [**590**]{}, 209 (2004). K. Martins, D. Blaschke and E. Quack, Phys. Rev. C [**51**]{}, 2723 (1995). E. S. Swanson, Phys. Lett. B [**588**]{}, 189 (2004). C. Hanhart, Y. S. Kalashnikova, A. E. Kudryavtsev and A. V. Nefediev, Phys. Rev. D [**76**]{}, 034007 (2007). C. E. Thomas and F. E. Close, Phys. Rev. D [**78**]{}, 034007 (2008). V. Baru [*et al.*]{}, Phys. Rev. D [**84**]{}, 074029 (2011). M. P. Valderrama, Phys. Rev. D [**85**]{}, 114037 (2012). F. Aceti, R. Molina and E. Oset, Phys. Rev. D [**86**]{}, 113007 (2012). F. K. Guo, C. Hidalgo-Duque, J. Nieves and M. P. Valderrama, Phys. Rev. D [**88**]{}, 054007 (2013). K. Heikkila, S. Ono and N. A. Tornqvist, Phys. Rev. D [**29**]{}, 110 (1984) Erratum: \[Phys. Rev. D [**29**]{}, 2136 (1984)\]. M. R. Pennington and D. J. Wilson, Phys. Rev. D [**76**]{}, 077502 (2007). B. -Q. Li, C. Meng and K. -T. Chao, Phys. Rev. D [**80**]{}, 014012 (2009). I. V. Danilkin and Y. A. Simonov, Phys. Rev. Lett.  [**105**]{}, 102002 (2010). P. G. Ortega, J. Segovia, D. R. Entem and F. Fernandez, Phys. Rev. D [**81**]{}, 054023 (2010); P. G. Ortega, D. R. Entem and F. Fernandez, J. Phys. G [**40**]{}, 065107 (2013). J. Ferretti, G. Galatà and E. Santopinto, Phys. Rev. C [**88**]{}, 015207 (2013). J. Ferretti and E. Santopinto, Phys. Rev. D [**90**]{}, 094022 (2014). N. N. Achasov and E. V. Rogozina, Mod. Phys. Lett. A [**30**]{}, 1550181 (2015). X. W. Kang and J. A. Oller, Eur. Phys. J. C [**77**]{}, 399 (2017). Y. Lu, M. N. Anwar and B. S. Zou, Phys. Rev. D [**94**]{}, 034021 (2016); M. N. Anwar, Y. Lu and B. S. Zou, arXiv:1806.01155. J. Ferretti and E. Santopinto, Phys. Lett. B [**789**]{}, 550 (2019). S. Dubynskiy and M. B. Voloshin, Phys. Lett. B [**666**]{}, 344 (2008). F. K. Guo, C. Hanhart and U. G. Mei[ß]{}ner, Phys. Lett. B [**665**]{}, 26 (2008); Phys. Rev. Lett.  [**102**]{}, 242004 (2009). Z. G. Wang and X. H. Zhang, Commun. Theor. Phys.  [**54**]{}, 323 (2010); Eur. Phys. J. C [**66**]{}, 419 (2010). M. B. Voloshin, Phys. Rev. D [**87**]{}, 091501 (2013). X. Li and M. B. Voloshin, Mod. Phys. Lett. A [**29**]{}, 1450060 (2014). Q. Wang, M. Cleven, F. K. Guo, C. Hanhart, U. G. Mei[ß]{}ner, X. G. Wu and Q. Zhao, Phys. Rev. D [**89**]{}, 034001 (2014); M. Cleven, F. K. Guo, C. Hanhart, Q. Wang and Q. Zhao, Phys. Rev. D [**92**]{}, 014005 (2015). N. Brambilla, G. Krein, J. Tarrús Castellà and A. Vairo, Phys. Rev. D [**93**]{}, 054002 (2016). M. Alberti, G. S. Bali, S. Collins, F. Knechtli, G. Moir and W. Söldner, Phys. Rev. D [**95**]{}, 074501 (2017). J. Y. Panteleeva, I. A. Perevalova, M. V. Polyakov and P. Schweitzer, arXiv:1802.09029. J. Ferretti, Phys. Lett. B [**782**]{}, 702 (2018). A. Esposito, A. Pilloni and A. D. Polosa, Phys. Rept.  [**668**]{}, 1 (2016). S. L. Olsen, T. Skwarnicki and D. Zieminska, Rev. Mod. Phys.  [**90**]{}, 015003 (2018). F. K. Guo, C. Hanhart, U. G. Mei[ß]{}ner, Q. Wang, Q. Zhao and B. S. Zou, Rev. Mod. Phys.  [**90**]{}, 015004 (2018). J. Ferretti, G. Galatà, E. Santopinto and A. Vassallo, Phys. Rev. C [**86**]{}, 015204 (2012). P. Geiger and N. Isgur, Phys. Rev. D [**41**]{}, 1595 (1990); Phys. Rev. Lett.  [**67**]{}, 1066 (1991). R. Bijker and E. Santopinto, Phys. Rev. C [**80**]{}, 065210 (2009); E. Santopinto and R. Bijker, Phys. Rev. C [**82**]{}, 062202 (2010). R. Bijker, J. Ferretti and E. Santopinto, Phys. Rev. C [**85**]{}, 035204 (2012). J. Ferretti, G. Galatà and E. Santopinto, Phys. Rev. D [**90**]{}, 054010 (2014). A. Le Yaouanc, L. Oliver, O. Pene, J. C. Raynal and S. Ono, Z. Phys. C [**28**]{}, 309 (1985). E. Santopinto, H. García-Tecocoatzi and R. Bijker, Phys. Lett. B [**759**]{}, 214 (2016) H. García-Tecocoatzi, R. Bijker, J. Ferretti and E. Santopinto, Eur. Phys. J. A [**53**]{}, 115 (2017). B. Silvestre- Brac and C. Gignoux, Phys. Rev. D [**43**]{}, 3699 (1991). D. Morel and S. Capstick, nucl-th/0204014. P. Geiger and N. Isgur, Phys. Rev. D [**55**]{}, 299 (1997). S. Godfrey and K. Moats, Phys. Rev. D [**92**]{}, 054034 (2015). [^1]: Proceedings of the Symposium “Symmetries and Order: Algebraic Methods in Many Body Systems" in honor of Professor Francesco Iachello on the occasion of his retirement, October 5-6 2018, Yale University.

--- abstract: 'We study the first passage statistics to adsorbing boundaries of a Brownian motion in bounded two-dimensional domains of different shapes and configurations of the adsorbing and reflecting boundaries. From extensive numerical analysis we obtain the probability $P(\omega)$ distribution of the random variable $\omega=\tau_1/(\tau_1+\tau_2)$, which is a measure for how similar the first passage times $\tau_1$ and $\tau_2$ are of two independent realisations of a Brownian walk starting at the same location. We construct a chart for each domain, determining whether $P(\omega)$ represents a unimodal, bell-shaped form, or a bimodal, M-shaped behavior. While in the former case the mean first passage time (MFPT) is a valid characteristic of the first passage behavior, in the latter case it is an insufficient measure for the process. Strikingly we find a distinct turnover between the two modes of $P(\omega)$, characteristic for the domain shape and the respective location of absorbing and reflective boundaries. Our results demonstrate that large fluctuations of the first passage times may occur frequently in two-dimensional domains, rendering quite vague the general use of the MFPT as a robust measure of the actual behavior even in bounded domains, in which all moments of the first passage distribution exist.' author: - 'Thiago G. Mattos' - 'Carlos Mejía-Monasterio' - Ralf Metzler - Gleb Oshanin title: 'First passages in bounded domains: When is the mean first passage time meaningful?' --- Introduction {#sec:intro} ============ The concept of first passage underlies diverse stochastic processes in which it is relevant when the value of the random variable reaches a preset value for the first time. A few stray examples across disciplines include chemical reactions [@ol; @lov; @ben; @beni; @mattosreis; @smolu], the firing of a neuron [@gersh; @burkitt], random search of a mobile or immobile target [@alb; @alb1; @ol1; @ralf; @osh; @wio; @carlos; @paul; @port; @evans; @gel], diffusional disease spreading [@Lloyd2001], DNA bubble breathing [@hanke], dynamics of molecular motors [@motor1; @motor2], the triggering of a stock option [@bouch], etc. A variety of first passage time phenomena and different related results have been investigated in Refs. [@katja; @sid]. While for continuous processes the first passage across a given preset value coincides with the first arrival to exactly this value, for L[é]{}vy flights characterized by long-tailed jump length distributions with diverging variance both quantities become different, and large overshoots across a preset value occur [@chech]. The distribution of first passage times in unbounded domains is typically broad, such that not even the mean first passage time exists [@sid]. In particular, in one-dimensional, semi-infinite domains the first passage time distribution of a Markovian process is universally dominated by the $t^{-3/2}$ scaling nailed down by the Sparre Andersen theorem [@sid]. A similar divergence of the mean first passage time occurs in stochastic processes characterized by scale-free distributions of waiting times [@mekla]. In contrast, in many practically important situations first passage processes involve particles which move randomly in bounded domains (see, e.g., Refs. [@ol2; @ol3; @ol4; @fBM-wedge]). In this case the random variable of interest, the first passage time $\tau$ to, e.g., a boundary, a target chemical group, a binding site on the surface of the domain or elsewhere within the domain, etc., has a distribution $\Psi(\tau)$ of the generic, generalized inverse Gaussian form (see [*e.g.*]{}, the discussion in Ref. [@carlos] and references therein) $$\label{eq:Pt-exptrunc} \Psi(\tau)\sim \exp\left(-\frac{a}{\tau}\right)\frac{1}{\tau^{1+\mu}}\exp \left(-\frac{\tau}{b}\right),$$ where $a$ and $b$ are some constants dependent on the shape of the domain, the exact starting point within the domain, etc., and $\mu$ is the so-called persistence exponent [@satya]. When the linear size of the domain (say, the radius $R$ of a circular or a spherical domain) diverges (i.e., $R\to\infty$), the parameter $b$ also diverges such that the long-time asymptotic behavior of the first passage time distribution is of power-law form without a cutoff. In this case, at least some, if not all, of the moments of $\Psi(\tau)$ diverge. The first passage time distribution in Eq. (\[eq:Pt-exptrunc\]) is exact only in the particular case of Brownian motion on a semi-infinite line in presence of a bias pointing towards the target site, or, equivalently, for the celebrated integrate-and-fire model of neuron firing by Gerstein and Mandelbrot [@gersh]. In general, the detailed form of $\Psi(\tau)$ is obviously much more complex than given by Eq. (\[eq:Pt-exptrunc\]), depending on the very shape of the domain under consideration and the exact boundary value problem. Typically $\Psi(\tau)$ is given in terms of an infinite series. Nonetheless, on a *qualitative* level, the approximation (\[eq:Pt-exptrunc\]) provides a clear picture of the actual behavior of the first passage time distribution in bounded domains. Namely, $\Psi(\tau)$ consists of three different parts: a singular decay for small values of $\tau$, which mirrors the fact that the first passage to some point starting from a distant position cannot occur instantaneously. This is followed at intermediate times by a generic power-law decay with exponent $\mu$, depending on the exact type of random motion. Finally, an exponential decay at long $\tau$ cuts off the power-law. A crucial aspect is that the exponential cutoffs at both short and long $\tau$ ensure that in bounded domains $\Psi(\tau)$ possesses moments of arbitrary positive or negative order. Distributions of the form (\[eq:Pt-exptrunc\]) are usually considered *narrow*, as opposed to *broad* distributions, which do not possess all moments [@katja; @sid; @mekla; @chech], e.g., $\Psi(\tau)$ in Eq. (\[eq:Pt-exptrunc\]) with $b=\infty$. Once all moments exist, it is often tacitly assumed that the first moment of this distribution, the mean first passage time (MFPT) $$\label{mfpt} \langle\tau\rangle=\int_0^{\infty}\tau\Psi(\tau)d\tau,$$ is an adequate measure of the first passage behavior. The actual analytical calculation of the MFPT may require a considerable computational effort, and the calculation of higher moments is quite formidable and is not always possible, (compare, e.g., Refs. [@ol2; @ol3; @ol4]). Conversely, it has been demonstrated in, e.g., recent Refs. [@red; @samor; @samor2; @hol; @schehr; @gl] that random variables with truncated power-law distributions behave in several important aspects as those characterized by non-truncated, *broad* distributions, revealing substantial fluctuations between individual realisations and thus rendering the concept of a mean first passage time a bit unsubstantiated. To be more precise, this concerns not the functional form of the MFPT for a given process, but rather its use as a characteristic quantity for the process. The functional form of the MFPT is certainly an important property, providing valuable insights to the scaling behavior, for instance with the system size or the initial distance of starting point and target. In contrast, the very numerical value of the MFPT can significantly differ from the values drawn from individual trajectories. Therefore, the MFPT can be substantially larger than the most probable value for the first passage time. Clearly, an understanding of how representative the MFPT is of the actual behavior and, concurrently, how important fluctuations of $\tau$ between individual realisations indeed are of utmost conceptual importance in many areas, such as, e.g., an interpretation of the first passage data obtained from single particle tracking. In this paper we analyze, via extensive Monte Carlo simulations the role of fluctuations between individual realisations of first passage times for Brownian motion (BM) in two-dimensional bounded domains of different shapes, and with different configurations of the reflective and adsorbing boundaries. Analogous results for three-dimensional systems and for systems with quenched disorder will be presented elsewhere [@thiago2]. Simultaneity concept of first passage ===================================== To quantify the relevance of such fluctuations and the effective *broadness* of the corresponding first passage time distribution $\Psi(\tau)$ we employ a novel diagnostics method based on the concept of simultaneity of first passage events, compare Fig. \[fig:scheme\]. Instead of the original first passage problem with quantifying the statistical outcome for a single Brownian walker, we simultaneously launch two identical, non-interacting Brownian particles at the same position ${\bf r_0}$ (which is identical to two different realisations of the trajectories of a single BM starting at ${\bf r_0}$). The corresponding outcomes are the first passage times $\tau_1$ and $\tau_2$. We now define the random variable $$\label{def:omega} \omega\equiv\frac{\tau_1}{\tau_1+\tau_2},$$ such that $\omega$ ranges in the interval $[0,1]$. The *uniformity index* $\omega$ measures the *likelihood* that both walkers arrive to the adsorbing boundary simultaneously: when $\omega$ is close to 1/2, the process is uniform and the particles behave as if they were almost performing a Prussian *Gleichschritt*. In contrast, values of $\omega$ close to 0 or 1 mean highly non-uniform behavior, implying that the MFPT is not representative of the actual behavior, but is merely the first moment of an *effectively* broad distribution. We note parenthetically that similar random variables have been used in the analysis of random probabilities induced by normalization of self-similar Lèvy processes [@iddo1], of the fractal characterization of Paretian Poisson processes [@iddo2], and of the so-called Matchmaking paradox [@iddo3; @iddo4]. ![(Color online) Trajectories of two Brownian walkers starting at the same initial position $(r_0,\theta_0)$ inside a bounded pie-wedge domain with opening angle $\Theta$ as well as absorbing radial boundaries (dashed lines) and reflecting boundary (solid line) at $r=1$. The values of the first passage times to the adsorbing boundaries are used to construct the random variable $\omega$.[]{data-label="fig:scheme"}](fig1.eps){width="0.75\linewidth"} Within a given bounded domain we evaluate the distribution $P(\omega)$ measuring the uniformity of the first passage dynamics with respect to some fixed starting point $\bf r_0$. This is repeated for a large number of nodes $\bf r_0$ within the domain, thus producing a uniformity chart of first passage. Remarkably, we find that the very shape of this distribution depends delicately on the domain shape, the actual settings of adsorbing and reflecting boundaries, and on the starting location $\bf r_0$. In some starting areas $P(\omega)$ has a characteristic unimodal, bell-shaped form with a maximum at $\omega =1/2$, signaling that most pairs of BMs will arrive to the adsorbing boundary simultaneously. This means, in turn, that in this case the parental first passage time distribution $\Psi(\tau)$ can be considered as sufficiently *narrow* such that the MFPT can be considered as a plausible measure of individual first passage events, providing a rather accurate estimate for the typical value of the first passage time. Conversely, we find that for other starting areas $P(\omega)$ exhibits a completely different behavior and has a characteristic bimodal, M-shaped form with a local *minimum* at $\omega=1/2$ and two maxima close to $\omega=0$ and $\omega=1$. In that case simultaneous arrival of two initially synchronized walkers is unlikely, i.e., any two trajectories will most likely possess distinctly different first passage times. The parental first passage time distribution $\Psi(\tau)$ is consequently *broad* and sample-to-sample fluctuations matter: the MFPT cannot be considered as an adequate measure of the actual behavior. Given that, by definition, the averages $\langle\tau_1\rangle$ and $\langle \tau_2\rangle$ are identical and, moreover, the moments of $\tau_1$ and $\tau_2$ of arbitrary order coincide, one can think of $\omega$ (and, hence, of the distribution $P(\omega)$) as a measure of the symmetry breaking between different realisations of the process. Note also that situations in which the mean value of some pertinent parameter is dominated by the tails of the distribution, and this mean thus has a very different value compared to the most probable value (and may even show a completely different dependence on the system parameters) is most often encountered in disordered systems [@evans; @gl]. Here we observe such a behavior in absence of any disorder. Scanning then over the possible starting points within each bounded domain, we obtain a corresponding phase-chart for $P(\omega)$, distinguishing regions in which $P(\omega)$ has M-shaped or bell-shaped behavior. The demarcation zone between these two phases, depicted by beige color in the following, represents a plateau-like, almost uniform behavior of $P(\omega)$ with zero second derivative at $\omega=1/2$. We proceed by giving a general definition of the first passage time distribution $\Psi(\tau)$ and its corresponding MFPT in Section \[sec:MFPT\], and also establish a relation between $\Psi(\tau)$ and the uniformity distribution $P(\omega)$. In Section \[sec:pie\] we study in detail the problem of Brownian motion in a pie-wedge shaped domain with absorbing and reflecting boundaries. In Sections \[sec:circle\] and \[sec:triangle\] we discuss the forms of $P(\omega)$, as a function of the location of the starting point, for circular domains with small aperture on the boundary, a two-dimensional version of the so-called Narrow Escape Time problem [@NET], and for triangular domains with adsorbing boundaries, respectively. Our results are summarized in Section \[sec:concl\]. First passage distribution, mean first passage time, and the uniformity distribution $P(\omega)$. {#sec:MFPT} ================================================================================================= Consider a BM inside a general two-dimensional domain $\mathcal{S}$, whose boundary $\partial\mathcal{S}\equiv\partial\mathcal{S}_a\cup\partial\mathcal{ S}_r$ comprises reflecting, $\partial\mathcal{S}_r$, and absorbing, $\partial \mathcal{S}_a$, parts. At time $t=0$, the BM initiates at $\mathbf{r}_0\in \mathcal{S}$ and evolves within the domain until the trajectory hits $\partial \mathcal{S}_a$ for the first time at some random instant $\tau$. Furthermore let $P(\mathbf{r},t|\mathbf{r}_0)$ denote the conditional probability distribution for finding the Brownian walker at position $\mathbf{r}$ at time $t$, provided the initial condition was at $\mathbf{r}_0$ at $t=0$. The distribution $P( \mathbf{r},t|\mathbf{r}_0)$ is the solution of the diffusion equation $$\label{diff} \frac{\partial}{\partial t}P(\mathbf{r},t|\mathbf{r}_0)=D\nabla^2_{\mathbf{r}} P(\mathbf{r},t|\mathbf{r}_0)$$ on $\mathcal{S}$, where $\nabla^2_{\mathbf{r}}$ is the two-dimensional Laplacian equivalent to $\partial^2/\partial x^2+\partial^2/\partial y^2$ in Cartesian coordinates. Eq. (\[diff\]) is subject to the initial condition as well as the boundary conditions at $\partial\mathcal{S}$. Here $D$ is the diffusion coefficient. The solution of this boundary value problem is, in the best case, cumbersome, and explicit solutions may be obtained for only few simple geometries, compare Ref. [@Carslaw]. If a finite part of the boundary is absorbing, i.e., $\partial\mathcal{S}_a$ is not empty, then the distribution $P(\mathbf{r},t|\mathbf{r}_0)$ is no longer normalized. The survival probability $\mathscr{S}_{\mathbf{r}_0}(t)$ that the walker has not reached $\partial\mathcal{S}_a$ up to time $t$, is defined by $$\label{surv} \mathscr{S}_{\mathbf{r}_0}(t)=\int_{\mathcal{S}}P(\mathbf{r},t|\mathbf{r}_0) d\mathbf{r}.$$ $\mathscr{S}_{\mathbf{r}_0}(t)$ is a monotonically decreasing function of time, eventually reaching zero value, $\lim_{t\to\infty}\mathscr{S}_{\mathbf{r} _0}(t)=0$. The desired distribution of first passage times to the adsorbing boundary becomes $$\label{FPT} \Psi_{\mathbf{r}_0}(\tau)=-\frac{d\mathscr{S}_{\mathbf{r}_0}(\tau)}{d\tau}.$$ The MFPT associated with the distribution $\Psi(\tau)$ is defined as the first moment $$\label{MFPT} \langle\tau\rangle(\mathbf{r}_0)=\int_0^\infty\tau\Psi_{\mathbf{r}_0}(\tau) d\tau=\int_0^\infty\mathscr{S}_{\mathbf{r}_0}(\tau)d\tau.$$ We note parenthetically that in most of the existing literature, apart of recent Refs.[@ben; @ol2; @ol3], the dependence of the MFPT on the starting position of the walker is either simply neglected, or it is assumed that the starting point is randomly distributed within the domain $\mathcal{S}$. As we proceed to show, the $\mathbf{r}_0$-dependence of the first passage time distribution is a crucial aspect which cannot be neglected. We now turn to the uniformity distribution $P(\omega)$ of the random variable $\omega$, Eq. (\[def:omega\]). Let $$\label{mg} \Phi(\lambda)=\int_0^1P(\omega)\exp\left(-\lambda\omega\right)d\omega,$$ with $\lambda\geq0$, denote the moment generating function of $\omega$. Since $\tau_1$ and $\tau_2$ are independent, identically distributed random variables, expression (\[mg\]) can formally be represented as $$\label{2} \Phi(\lambda)=\int^{\infty}_0\!\!\!\!\int^{\infty}_0\!\!\!\!\Psi(\tau_1)\Psi(\tau_2)\exp\left( -\lambda\frac{\tau_1}{\tau_1+\tau_2}\right)d\tau_1d\tau_2.$$ Integrating over $d\tau_1$ we change the integration variable, $\tau_1\to \omega$, so that Eq. (\[2\]) is rewritten in the form $$\begin{aligned} \label{3} \Phi(\lambda) & = & \int^{1}_0\exp\left(-\lambda\omega\right)\frac{d\omega}{(1- \omega)^2}\times\nonumber \\ & & \int^{\infty}_0\tau_2\Psi(\tau_2)\Psi\left(\frac{\omega}{1-\omega}\tau_2\right) d\tau_2.\end{aligned}$$ From comparison with Eq. (\[mg\]), we readily read off the desired distribution function $$\label{def:Pw} P(\omega)=\frac{1}{(1-\omega)^2}\int^{\infty}_0\tau\Psi(\tau) \Psi\left(\frac{\omega}{1-\omega}\tau\right)d\tau.$$ Therefore, $P(\omega)$ is known for given $\Psi(t)$. To get an idea of the typical behavior of the uniformity distribution $P( \omega)$, we use the generic form (\[eq:Pt-exptrunc\]) for the first passage time distribution. From Eq. (\[def:Pw\]) we find from integration that $$\begin{aligned} \label{eq:Pw-exptrunc} P(\omega) & = & \frac{1}{2K_{\mu}^2\left(2\sqrt{\frac{a}{b}}\right)} \frac{1}{\omega(1-\omega)}\times \nonumber \\ & & K_{2\mu}\left(2\sqrt{\frac{a}{b\omega (1-\omega)}}\right),\end{aligned}$$ where $K_{2 \mu}(\cdot)$ is the modified Bessel function of the second type. It was realized [@carlos; @schehr] that the form of the distribution $P(\omega)$ in Eq. (\[eq:Pw-exptrunc\]) is distinctly sensitive to the value of the persistence exponent $\mu$, which characterizes the scaling behavior of the first passage time distribution $\Psi(\tau)$ at intermediate times. Thus, for $\mu>1$, $P(\omega)$ is always a unimodal, bell-shaped function with a maximum at $\omega=1/2$. For $\mu=1$, $P(\omega)$ is almost uniform, $P(\omega)\approx1$, apart from narrow regions at the corners $\omega =0$ and $\omega=1$, for $b/a\gg1$. Curiously, for $\mu<1$, which corresponds to the most common case, there exists a critical value $p_c$ of the ratio $p=b/a$ such that for $p>p_c$ the distribution $P(\omega)$ has a characteristic M-shaped form with two maxima close to 0 and 1, while at $\omega=1/2$ we find a local minimum. Such a transition from a unimodal, bell-shaped to bimodal, M-shaped form mirrors a significant manifestation of sample-to-sample fluctuations that has been indeed observed in exact calculations of $P(\omega)$ for Brownian search processes for an immobile target in $d$-dimensional spherical geometries [@carlos]. In what follows we further explore this intriguing behavior of the first passage time distribution via extensive Monte Carlo simulations focusing on the effects of the domain shape, the type of the boundary conditions, and the initial position of the walker. Uniformity distribution $P(\omega)$ in a pie-wedge domain {#sec:pie} ========================================================= Consider now the case of a bounded domain of pie-wedge shape with unit radius, $R=1$ and opening angle $\Theta$. The absorbing boundaries correspond to the radial edges, while the outer circular edge is reflective, compare Fig. \[fig:scheme\]. Clearly, for a BM inside such a pie-wedge domain, all moments of the first passage time distribution exist. Before we proceed to investigate this case we first turn to the case when the wedge radius is infinite, $R\rightarrow\infty$. Then the distribution function $P(\mathbf{r},t|\mathbf{r}_0)$ is known exactly, (see, e.g., Refs. [@sid; @metzler]) and is represented by an infinite series whose leading term for $t\to\infty$ is given, up to a normalization constant, by $$\begin{aligned} \label{eq:Pinf} P(\mathbf{r},t|\mathbf{r}_0) & \simeq & \frac{\pi\sin\left(\pi\theta_0/\Theta\right)}{ 4D\Theta t}e^{-(\rho^2+\rho_0^2)/4Dt}\times \nonumber \\ & & I_{\pi/\Theta}\left(\frac{\rho_0 \rho}{2Dt} \right),\end{aligned}$$ where $I_\nu(z)$ is the modified Bessel function of the first kind and ${\bf r}=(\rho,\theta)$ is conveniently represented in polar coordinates. This solution is obtained for the sharp initial condition $P(\mathbf{r},0|\mathbf{r}_0)=\pi\sin(\pi\theta_0/\Theta) \delta({\bf r}-{\bf r_0})/2\Theta\rho_0$. From Eq. (\[eq:Pinf\]) one finds the asymptotic behavior of the survival probability, $$\label{eq:Sinf} \mathscr{S}_{\mathbf{r}_0}(t)\simeq\left(\frac{\rho_0^2}{D}\right)^{\pi/2\Theta} t^{-\pi/2\Theta},$$ such that the first passage time distribution becomes $$\label{eq:FPTDinf} \Psi_{\mathbf{r}_0}(t)\simeq\frac{\pi}{2\Theta}\left(\frac{\rho_0^2}{D}\right)^{ \pi/2\Theta}\frac{1}{t^{1+\pi/2\Theta}}.$$ Note that this distribution is of the generic form (\[eq:Pt-exptrunc\]), where $b=\infty$ due to the infinite domain size. The non-universal persistence exponent is given by $$\mu=\frac{\pi}{2\Theta}.$$ Therefore, the MFPT diverges when $\Theta\geq \pi/2$ and is finite for $\Theta<\pi /2$. According to the qualitative analysis from Section \[sec:MFPT\], $P( \omega)$ will have a bimodal form in the former case and a unimodal one in the latter. We now turn our attention to finite-sized pie-wedges, for which the MFPT and all higher moments of Eq. (\[FPT\]) are finite. In principle, an exact solution for the first passage time distribution in this case can be obtained by solution of the corresponding mixed boundary value problem, but the result will be too cumbersome for our purposes. Instead, we resort to numerical simulations. We now show that for finite pie-wedges the actual behavior is in fact richer than in the case of an infinite wedge. We performed Monte Carlo simulations of a random walk inside a pie-wedge of unit radius and opening angle $\Theta$. The boundary conditions along the radii are absorbing and reflecting along the circular edge, see Fig. \[fig:scheme\]. The random walk is simulated in terms of a standard Pearson walk on a plane (compare Ref. [@Hughes]), which consists of a sequence of steps of fixed length $\lambda=0.001$ and uniform waiting time $v=1/\lambda$. After each step the walker turns by a random angle with uniform distribution. At time $t=0$ the walker is released at $(\rho_0,\theta_0)$, and its trajectory is recorded until it hits a point on the absorbing boundary for the first time. Generating $N$ (we used $N=10^5$) such trajectories, we obtain a set of first passage times $\{\tau_i\}$, from which we construct the first passage time distribution. Since all $\tau_i$ are independent, identically distributed random variables, the uniformity distribution $P(\omega)$ is then readily obtained via Eq.  from distinct pairs $\tau_1$ and $\tau_2$ chosen at random from the set $\{\tau_i\}$. ![(Color online)(a) First passage time distribution $\Psi_{(\rho_0,\theta_0)}(\tau)$ for a Brownian walk in a pie-wedge domain with opening angle $\Theta=\pi/2$. Different colors and symbols correspond to different starting points: $(\rho_0=0.76, \theta_0=-0.38~\Theta$ (dark blue/triangles), $(0.76,-0.23~\Theta)$ (beige/squares), and $(0.76,0)$ (light blue/circles). The dashed straight line indicates the intermediate power-law decay $\Psi(\tau)\sim1/\tau^2$, Eq. (\[eq:FPTDinf\]). (b). The corresponding distribution $P(\omega)$ with the same color and symbol coding. In grey scale, dark blue corresponds to the darkest shade of grey, light blue to dark grey and beige to light grey.[]{data-label="fig:pie-fptd"}](fig2-a.eps "fig:"){width="0.5\linewidth"}![(Color online)(a) First passage time distribution $\Psi_{(\rho_0,\theta_0)}(\tau)$ for a Brownian walk in a pie-wedge domain with opening angle $\Theta=\pi/2$. Different colors and symbols correspond to different starting points: $(\rho_0=0.76, \theta_0=-0.38~\Theta$ (dark blue/triangles), $(0.76,-0.23~\Theta)$ (beige/squares), and $(0.76,0)$ (light blue/circles). The dashed straight line indicates the intermediate power-law decay $\Psi(\tau)\sim1/\tau^2$, Eq. (\[eq:FPTDinf\]). (b). The corresponding distribution $P(\omega)$ with the same color and symbol coding. In grey scale, dark blue corresponds to the darkest shade of grey, light blue to dark grey and beige to light grey.[]{data-label="fig:pie-fptd"}](fig2-b.eps "fig:"){width="0.5\linewidth"} Fig. \[fig:pie-fptd\](a) shows the first passage time distributions corresponding to a fixed $\rho_0=0.76$ and three different starting angles $\theta_0$ for a pie-wedge with opening angle $\Theta=\pi/2$. One notices that, for small and large values of $\tau$, all three distributions $\Psi(\tau)$ significantly deviate from the intermediate power-law behavior, which is due to exponential tempering. On the other hand, at intermediate times the distributions exhibit a slower, power-law like decay within a range that depends significantly on $\theta_0$. The narrowest distribution (light blue) is obtained for a starting position $(0.76,0)$, which is exactly on the symmetry axis of the wedge. Increasing the angle $\theta_0$ away from the symmetry axis results in a broadening of $\Psi(\tau)$, and the intermediate algebraic decay is more pronounced. In panel (b) of Fig. \[fig:pie-fptd\] we plot the corresponding uniformity distributions $P(\omega)$. For the narrowest first passage time distribution $\Psi(\tau)$ (light blue symbols), $P(\omega)$ is bell-shaped with 1/2 representing the most probable value, such that sample-to-sample fluctuations of $\tau$ are less significant. In this case, apparently, the MFPT is a meaningful, reliable measure of the first passage behavior, and any two walkers starting from the same position on the symmetry axis of the wedge will most likely be absorbed at the same instant of time. Strikingly, we find that this is no longer valid for the two other starting positions off the symmetry axis: for ${\bf r_0}=(0.76,-\pi/3)$ the uniformity distribution $P(\omega)$ is almost uniform, except for narrow regions in the vicinity of the edges, meaning that any relation between the first passage times of two walkers is equally probable. Finally, for starting position $(0.76,-\pi/7)$, which is the one closest to the absorbing boundary $P(\omega)$ has a characteristic M-shaped form with maxima close to 0 and 1, and a local minimum at $\omega=1/2$. This signifies that in a such a case the symmetry between any two walkers is most distinctly broken, and they will be absorbed at very different times. Clearly, in this case the MFPT is not representative of the actual behavior. ![Dependence of the parameter $\chi$ on the starting angle $\theta_0$ for fixed $\rho_0=0.76$ in a pie-wedge domain with opening angle $\Theta=\pi/2$. The three insets show the shape of the uniformity distribution $P(\omega)$ for the $\theta_0$ values indicated by the arrows.[]{data-label="fig:trans"}](fig3.eps){width="1.0\linewidth"} To quantify the shape of the distribution of $\omega$ we perform a fit of the numerically obtained $P(\omega)$ to a quadratic polynomial of $\omega$ in the domain $0.05<\omega<0.95$. From this fit we obtain the coefficient $\chi$ of the quadratic term. The sign of $\chi$ thus determines the shape of $P(\omega)$: $\chi<0$ corresponds to the unimodal, bell-shaped distribution, $\chi>0$ signifies that the distribution is bimodal, M-shaped, and a zero value of $\chi=0$ means that $P(\omega)$ is uniform. Such a procedure, of course, has some ambiguities, especially when we deal with the demarcation line $\chi=0$ between regions in which $P(\omega)$ has unimodal and bimodal forms, as it is not always clear how many digits are to be taken into account. This results in a certain broadening of the demarcation line. However, we have checked in several cases that this procedure produces reliable results. In Fig. \[fig:trans\] we show the evolution of $\chi$ versus the starting angle $\theta_0$ for fixed $\rho_0=0.76$ in a pie-wedge domain with the opening angle $\Theta=\pi/2$. We observe a continuous, periodic variation of $P(\omega)$ with the starting position, changing from an M-shaped to a bell-shaped form. The insets of Fig. \[fig:trans\] show the schematic distribution $P(\omega)$ for some specific values of $\theta_0$. The absolute value of $\chi$ indicates how far the distribution $P(\omega)$ deviates from a locally uniform distribution. Finally, we used this approach to create the phase-chart for the shape of the uniformity distribution $P(\omega)$ with respect to the starting position of the walker within the pie-wedge domain. In Fig. \[fig:pie-phase\] we present a systematic scan of the domain for three pie-wedges with different opening angles. One observes that in all three cases, there exists a region in which $P(\omega)$ is bell-shaped (light blue symbols) and a region with M-shaped $P(\omega)$ (dark blue symbols), separated by a small region with nearly uniform distribution (beige symbols). ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ in three different pie-wedge domains: (a) $\Theta= 3\pi/4$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi/3$. The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, where we chose $\chi_\star=0.25$. For each initial location, $P(\omega)$ was computed from a sample of $N=10^5$ random trajectories.[]{data-label="fig:pie-phase"}](fig4-a.eps "fig:") ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ in three different pie-wedge domains: (a) $\Theta= 3\pi/4$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi/3$. The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, where we chose $\chi_\star=0.25$. For each initial location, $P(\omega)$ was computed from a sample of $N=10^5$ random trajectories.[]{data-label="fig:pie-phase"}](fig4-b.eps "fig:") ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ in three different pie-wedge domains: (a) $\Theta= 3\pi/4$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi/3$. The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, where we chose $\chi_\star=0.25$. For each initial location, $P(\omega)$ was computed from a sample of $N=10^5$ random trajectories.[]{data-label="fig:pie-phase"}](fig4-c.eps "fig:") Therefore, as we have already remarked, the actual behavior in a finite pie-wedge appears to be much richer than in an infinite wedge. Consider an experiment in which one aims to find an estimate of the MFPT by tracking the evolution of a few single particle trajectories starting at the same position inside the light blue region. The outcome of such an experiment will be a good estimate of the MFPT, with reliably small error. This will be the case since in the light blue region $P(\omega)$ is bell-shaped, which means that the probability that the two trajectories arrive at the same time is maximal. In contrast, if two single particle trajectories start anywhere inside the dark blue region, then it is most likely that these trajectories will arrive to the adsorbing boundary at very different times, yielding a poor and unreliable estimate for the MFPT. The sample-to-sample fluctuations in this case are very important and, as a consequence, the MFPT is not an adequate measure of the actual behavior. Qualitatively, the sample-to-sample fluctuations of the MFPT increase as the trajectories start closer to the absorbing boundaries. However, this is not always true, as can be observed in Fig. \[fig:pie-phase\] (c) for the pie-wedge with $\Theta=\pi/3$ for which the light blue region extends toward the vertex of the wedge. Circular domain with aperture {#sec:circle} ============================= We now turn our attention to the first passage time problem of a Brownian particle in a circular domain of unit radius and the following boundary conditions: the segment with $|\theta|<\Theta/2$ is absorbing while the remaining part of the outer circle is reflective. The aperture of the circular domain is thus of angle $\Theta$. One often encounters a three-dimensional version of this problem in cellular biochemistry, when one is interested in the time needed for a particle (a ligand, etc.), diffusing within a bounded domain (for instance, a microvesicle) to reach a small escape window or a binding site, which is an aperture in an otherwise reflecting boundary. This is the so-called Narrow Escape Time problem, which attracted considerable attention within the last two decades (see, e.g., Refs.[@ol4; @NET] and references therein). ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ for a Brownian walker in the unit circle with reflective BCs (solid lines) and an aperture of size $\Theta$ (absorbing BC - dashed lines). (a) $\Theta=\pi/18$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi$, and (d) $\Theta=2\pi$. Starting locations are colored light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$ and beige if $|\chi|<\chi_ ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ for a Brownian walker in the unit circle with reflective BCs (solid lines) and an aperture of size $\Theta$ (absorbing BC - dashed lines). (a) $\Theta=\pi/18$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi$, and (d) $\Theta=2\pi$. Starting locations are colored light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$ and beige if $|\chi|<\chi_ \star$, where $\chi_\star=1$. In the lower panel, the relative error of the FPT $\varepsilon$ is shown as function of the initial radial position $r$ along the horizontal diameter of the unit circle with $\Theta=\pi/2$, panel (b): $r=0$ is associated to the center of the circle and the point $r=1$ lies over the absorbing boundary.[]{data-label="fig:circle"}](fig5-a.eps "fig:"){width="0.32\linewidth"} \star$, where $\chi_\star=1$. In the lower panel, the relative error of the FPT $\varepsilon$ is shown as function of the initial radial position $r$ along the horizontal diameter of the unit circle with $\Theta=\pi/2$, panel (b): $r=0$ is associated to the center of the circle and the point $r=1$ lies over the absorbing boundary.[]{data-label="fig:circle"}](fig5-b.eps "fig:"){width="0.32\linewidth"} ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ for a Brownian walker in the unit circle with reflective BCs (solid lines) and an aperture of size $\Theta$ (absorbing BC - dashed lines). (a) $\Theta=\pi/18$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi$, and (d) $\Theta=2\pi$. Starting locations are colored light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$ and beige if $|\chi|<\chi_ ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ for a Brownian walker in the unit circle with reflective BCs (solid lines) and an aperture of size $\Theta$ (absorbing BC - dashed lines). (a) $\Theta=\pi/18$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi$, and (d) $\Theta=2\pi$. Starting locations are colored light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$ and beige if $|\chi|<\chi_ \star$, where $\chi_\star=1$. In the lower panel, the relative error of the FPT $\varepsilon$ is shown as function of the initial radial position $r$ along the horizontal diameter of the unit circle with $\Theta=\pi/2$, panel (b): $r=0$ is associated to the center of the circle and the point $r=1$ lies over the absorbing boundary.[]{data-label="fig:circle"}](fig5-c.eps "fig:"){width="0.32\linewidth"} \star$, where $\chi_\star=1$. In the lower panel, the relative error of the FPT $\varepsilon$ is shown as function of the initial radial position $r$ along the horizontal diameter of the unit circle with $\Theta=\pi/2$, panel (b): $r=0$ is associated to the center of the circle and the point $r=1$ lies over the absorbing boundary.[]{data-label="fig:circle"}](fig5-d.eps "fig:"){width="0.32\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![(Color online) Phase-chart for the shape of the uniformity distribution $P(\omega)$ for a Brownian walker in the unit circle with reflective BCs (solid lines) and an aperture of size $\Theta$ (absorbing BC - dashed lines). (a) $\Theta=\pi/18$, (b) $\Theta=\pi/2$, (c) $\Theta=\pi$, and (d) $\Theta=2\pi$. Starting locations are colored light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$ and beige if $|\chi|<\chi_ \star$, where $\chi_\star=1$. In the lower panel, the relative error of the FPT $\varepsilon$ is shown as function of the initial radial position $r$ along the horizontal diameter of the unit circle with $\Theta=\pi/2$, panel (b): $r=0$ is associated to the center of the circle and the point $r=1$ lies over the absorbing boundary.[]{data-label="fig:circle"}](fig5-e.eps){width="0.7\linewidth"} We analyze the shape of the uniformity distribution $P(\omega)$ as a function of the starting point of a Brownian walker. As in the previous section, we generate $N=10^5$ random walks commencing from the same starting position $(\rho_0,\theta_0)$ inside the unit circle and determine the set $\{\tau_i\}$ of first passage times to the location of the aperture. From these data we obtain $P(\omega)$ and compute the parameter $\chi$. In Fig. \[fig:circle\] we show the phase-chart for the shape of $P(\omega)$ for four different sizes of the aperture: $\Theta=\pi/18$, $\pi/2$, $\pi$, and $2\pi$. Each symbol in the charts is light blue, beige, or dark blue, depending on whether the corresponding starting position leads to a bell-shaped, uniform, or M-shaped distribution $P(\omega)$, respectively. Note that the case shown in Fig. \[fig:circle\] (d) reduces to a one-dimensional problem (see, e.g., Ref. [@carlos]). Similarly to our findings for the pie-wedge domain, we observe that the MFPT is not always a representative measure for the two-dimensional Narrow Escape Time problem. Interestingly, the sub-domain in which the MFPT is the least probable outcome (dark blue coding) is practically the same for small holes of $\Theta\le\pi/2$. It is worthwhile noting that while in this region $P(\omega)$ is always bimodal, the height of its maxima increases (and its value $P(\omega=1/2)$ representative of the MFPT decreases) depending on the distance from the opening. In addition, from the set of first passage times $\{\tau_i\}$ we directly computed the MFPT and the variance $\mathrm{var}(\tau)$. Both statistical indicators grow with the distance from the absorbing boundary. A more sensitive measure is the relative error $\varepsilon$, defined as the ratio $$\label{eq:eps} \varepsilon=\frac{\sqrt{\mathrm{var}(\tau)}}{\langle\tau\rangle}.$$ In the lower panel of Fig. \[fig:circle\] we show the dependence of the relative error on the starting position $\mathbf{r}_0$ for trajectories initiating along the symmetry axis of the domain, namely with respect to $r_0$ for fixed $\theta_0=0$. In agreement with the qualitative results of the phase-chart, $\varepsilon<1$ only when $P(\omega)$ is bell-shaped, and the MFPT is the most probable outcome of a single-particle trajectory, namely, for trajectories starting far enough from the absorbing boundary. Clearly, the closer the starting position is to the absorbing boundary the larger the relative error becomes. Very near the absorbing boundary the standard deviation of the first passage time becomes much larger than its mean. We note that this result is generic irrespectively of the aperture size. However $\varepsilon$ is just a number and it is not clear how to interpret it. For instance, $\varepsilon = 1$ or $\varepsilon = 2$, are these values too small or large enough to allow us to say that trajectory-to-trajectory fluctuations are significant? On the other hand, the distribution of the simultaneity index which we discuss here gives a lucid answer on this question as manifested by the change of modality of $P(\omega )$. Triangular domain with absorbing boundaries {#sec:triangle} =========================================== Finally, as a complementary example we consider a domain whose boundaries are completely absorbing. We consider the triangular domains shown in Fig. \[fig:triangle\]: two symmetric triangles with central angle $\pi/2$, panel (a), and $2\pi/3$, in panel (c), and an asymmetric triangle with angles $2\pi/3,\pi/4,\pi/12$ shown in panel (b). ![(Color online) Phase-chart of the shape of the uniformity distribution $P(\omega)$ for symmetric triangles with central angle $\pi/2$ (a) and $2\pi/3$ (c), and for the asymmetric triangle with angles $2\pi/3,\pi/4,\pi/12$ (b). The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, with $\chi_\star=0.25$.[]{data-label="fig:triangle"}](fig6-a.eps "fig:"){width="0.3\linewidth"} ![(Color online) Phase-chart of the shape of the uniformity distribution $P(\omega)$ for symmetric triangles with central angle $\pi/2$ (a) and $2\pi/3$ (c), and for the asymmetric triangle with angles $2\pi/3,\pi/4,\pi/12$ (b). The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, with $\chi_\star=0.25$.[]{data-label="fig:triangle"}](fig6-b.eps "fig:"){width="0.3\linewidth"} ![(Color online) Phase-chart of the shape of the uniformity distribution $P(\omega)$ for symmetric triangles with central angle $\pi/2$ (a) and $2\pi/3$ (c), and for the asymmetric triangle with angles $2\pi/3,\pi/4,\pi/12$ (b). The color of the symbols is light blue if $\chi<-\chi_\star$, dark blue if $\chi>\chi_\star$, and beige if $|\chi|<\chi_\star$, with $\chi_\star=0.25$.[]{data-label="fig:triangle"}](fig6-c.eps "fig:"){width="0.396\linewidth"} In Fig. \[fig:triangle\] we show the phase-chart of the shape of $P(\omega)$, the results obtained are qualitatively the same as for the previous two examples with mixed boundary conditions. $P(\omega)$ is bell-shaped and the MFPT the most probable outcome only when the trajectory starts far enough from the boundary. Also, similarly to what we observed for the pie-wedge domain, the domain in which $P(\omega)$ is unimodal extends toward the absorbing boundary if the vertex angle is less than $\pi/2$. Conclusions {#sec:concl} =========== We explored the problem of first passage of a Brownian particle to the absorbing boundary of finite, two-dimensional domains. From our study of the characteristic shapes of the associated distribution of the uniformity index $\omega$ we demonstrated that the MFPT represents the most probable outcome (and thus is quite meaningful) only if the trajectories start in a certain subregion of the total domain. For starting points in the complementary region the MFPT becomes the least probable outcome, indicating very large sample-to-sample fluctuations. These observations are generically important for single trajectory analysis of first passage time processes. We showed that the associated separation into bell-shaped and M-shaped forms of the uniformity distribution $P(\omega)$ is a robust property of Brownian motion by studying the problem in different symmetric and asymmetric domains with mixed or fully absorbing boundaries. We found that in general, sample-to-sample fluctuations of the first passage time increase when the trajectories start close to the target boundary, leading to the unexpected conclusion that in such situations the MFPT yields insufficient information, particularly, if the absorption time is extracted from the outcome of very few single-particle trajectories. Next, it is worthwhile mentioning that in many interesting situations the starting position of the trajectories are randomly distributed inside the finite domain. From such analysis the so-called global MFPT is usually derived, see, e.g., Ref. [@ol4]. Here we found that averaging the associated uniformity distribution $P(\omega)$ over the domain, $$P_{\mathrm{av}}(\omega)=\int_{\mathcal{S}} P_{\mathbf{r}_0}(\omega) d\mathbf{r}_0,$$ attains a uniform shape, except near $\omega=0$ and $\omega=1$. This appears to be a general property of $P_{\mathrm{av}}(\omega)$ associated with the probability conservation, and leads to the unexpected conclusion that the global MFPT has little meaning in such situations. As a final remark, we emphasize that the approach outlined here is not limited to first passage phenomena only, but can be quite generally applied to probe the significance of sample-to-sample fluctuations of arbitrary random variables having distributions for which *all* moments exist. Such distributions, as shown in our work, may appear $\omega$-broad, in the sense that the corresponding uniformity distribution $P(\omega)$ is bimodal, or, alternatively, $\omega$-narrow with unimodal $P(\omega)$. We recall that the variable $\omega$ has a very lucid physical meaning and its distribution can be determined if the parental distribution of the random variable is known. Indeed such sort of heterogeneity analysis has very recently started to be used to quantify sample-to-sample fluctuations in mathematical finances [@samor2; @hol], chaotic systems [@schehr], analysis of distributions of the diffusion coefficient of proteins diffusing along DNAs [@boyer] and FPT phenomena [@carlos]. **ACKNOWLEDGMENTS** The authors wish to thank Olivier Bénichou and Satya Majumdar for helpful discussion. The authors acknowledge financial support from the European Science Foundation and the hospitality of NORDITA, Stockholm, where part of this work was performed during the Non-Equilibrium Statistical Mechanics program. The research of TGM, RM and GO is partially supported by a Marie Curie International Research Staff Exchange Scheme Fellowship PIRSES-GA-2010-269139 within the 7th European Community Framework Programme. RM acknowledges funding from the Academy of Finland within the FiDiPro programme. The authors are partially supported by the ESF Research Network “Exploring the Physics of Small Devices”. [99]{} C. Loverdo, O. B[é]{}nichou, M. Moreau and R. Voituriez, Nat. Phys. [**4**]{}, 134 (2008). C. Loverdo, O. B[é]{}nichou, M. Moreau and R. Voituriez, J. Stat. Mech. P02045 (2009). O. Bénichou et al., Nat. Chem. [**2**]{}, 472 (2010). O. Bénichou, M. Moreau and G. Oshanin, Phys. Rev. E [**61**]{}, 3388 (2000). T. G. Mattos and Fábio D. A. Aarão Reis, J. Chem. Phys. [**131**]{}, 014505 (2009) M. Smoluchowski, Z. Phys. Chem. [**92**]{}, 129 (1917). G. L. Gerstein and B. B. Mandelbrot, Biophys. J [**4**]{}, 41 (1964). A. N. Burkitt, Biol. Cybern. [**95**]{}, 1 (2006). G. M. Viswanathan et al., Nature [**401**]{}, 911 (1999). G. M. Viswanathan et al., [*The Physics of Foraging: An Introduction to Random Searches and Biological Encounters*]{}, (Cambridge, Cambridge University Press, 2011). O. Bénichou, M. Coppey, M. Moreau, P. H. Suet and R. Voituriez, Phys. Rev. Lett. [**94**]{}, 198101 (2005); C. Loverdo, O. Bénichou, M. Moreau, and R. Voituriez, Phys. Rev. E [**80**]{}, 031146 (2009); O. Bénichou, C. Loverdo, M. Moreau, and R. Voituriez, Rev. Mod. Phys. [**83**]{}, 81 (2011). M. A. Lomholt, T. Ambj[ö]{}rnsson and R. Metzler, Phys. Rev. Lett. [95]{}, 260603 (2005); M. A. Lomholt, T. Koren, R. Metzler and J. Klafter, Proc. Natl. Acad. Sci. USA [**105**]{}, 11055 (2008); M. A. Lomholt et al., Proc. Natl. Acad. Sci. USA [**106**]{}, 8204 (2009). G. Oshanin, H. S. Wio, K. Lindenberg and S. F. Burlatsky, J. Phys.: Cond. Mat. [**19**]{}, 065142 (2007); J. Phys. A: Math. Theor. [**42**]{}, 434008 (2009); G. Oshanin, O. Vasilyev, P. Krapivsky and J. Klafter, Proc. Natl. Acad. Sci. USA [**106**]{}, 13696 (2009). F. Rojo, C. E. Budde and H. S Wio, J. Phys. A: Math. Theor. [**42**]{}, 125002 (2009); F. Rojo et al., J. Phys. A: Math. Theor. [**43**]{}, 345001 (2009). C. Mejia-Monasterio, G. Oshanin and G. Schehr, J. Stat. Mech. P06022 (2011). J. M. Newby and P. C. Bressloff, J. Stat. Mech. P04014 (2010); P. C. Bressloff and J. M. Newby, Phys. Rev. E [**85**]{}, 031909 (2012). I. G. Portillo, D. Campos and V. Méndez, J. Stat. Mech. P02033 (2011). M. R. Evans and S. N. Majumdar, Phys. Rev. Lett. [**106**]{}, 160601 (2011); J. Phys. A [**44**]{}, 435001 (2011). E. Gelenbe, Phys. Rev. E [**82**]{}, 061112 (2010) A. L. Lloyd and R. M. May, Science [**292**]{}, 1316 (2001). A. Hanke and R. Metzler, J. Phys. A [**36**]{}, L473 (2003); H. C. Fogedby and R. Metzler, Phys. Rev. Lett. [**98**]{}, 070601 (2007). G. Oshanin, J. Klafter and M. Urbakh, Europhys. Lett. [**68**]{}, 26 (2004); J. Phys.: Condens. Matter [**17**]{}, S3697 (2005). V. Palyulin and R. Metzler, J. Stat. Mech. L03001 (2012). J. P. Bouchaud and M. Potters, [*Theory of financial risk and derivative pricing: from statistical physics to risk management*]{}, (Cambridge, Cambridge University Press, 2003). K. Lindenberg and B. J. West, J. Stat. Phys. [**42**]{}, 201 (1986). S. Redner, [*A Guide to First-Passage Processes*]{}, (Cambridge, Cambridge University Press, 2001) A. V. Chechkin et al., J. Phys. A [**36**]{}, L537 (2003); T. Koren, M. A. Lomholt, A. V. Chechkin, J. Klafter, and R. Metzler, Phys. Rev. Lett. [**99**]{}, 160602 (2007). R. Metzler and J. Klafter, Physica A **278**, 107, (2000); H. Scher et al., Geophys. Res. Lett. **29**, 1061 (2002). S. Condamin et al., Nature [**450**]{}, 77 (2007). S. Condamin, O. Bénichou and M. Moreau, Phys. Rev. Lett. [**95**]{}, 260601 (2005); Phys. Rev. [**75**]{}, 021111 (2007). B. Meyer, C. Chevalier, R. Voituriez and O. Bénichou, Phys. Rev. E [**83**]{}, 051116 (2011); C. Chevalier, O. Bénichou, B. Meyer and R. Voituriez, J. Phys. A.: Math. Theor. [**44**]{}, 025002 (2011). J.-H. Jeon, A. V. Chechkin, and R. Metzler, Europhys. Lett. [**94**]{}, 20008 (2011). S. N. Majumdar, Curr. Sci. [**77**]{}, 370 (1999). G. Oshanin and S. Redner, Europhys. Lett. [**85**]{}, 10008 (2009). M. Grabchak and G. Samorodnitsky, Quantitative Finance [**10**]{}, 883 (2010). G. Oshanin and G. Schehr, Quantitative Finance [**12**]{}, 1325 (2012). G. Oshanin, Yu. Holovatch and G. Schehr, Physica A [**390**]{}, 4340 (2011). C. Mej[í]{}a-Monasterio, G. Oshanin and G. Schehr, Phys. Rev. E [**84**]{}, 035203 (2011). G. Oshanin, A. Mogutov and M. Moreau, J. Stat. Phys. [**73**]{}, 379 (1993); G. Oshanin, S. F. Burlatsky, M. Moreau and B. Gaveau, Chem. Phys. [**177**]{}, 803 (1993); C. Monthus, G. Oshanin, A. Comtet, and S. F. Burlatsky, Phys. Rev. E [**54**]{}, 231 (1996). T. G. Mattos, C. Mej[í]{}a-Monasterio, R. Metzler and G. Oshanin, in preparation. I. Eliazar, Physica A, [**356**]{}, 207 (2005). I. Eliazar and I. M. Sokolov, Physica A [**391**]{}, 3043 (2012). I. Eliazar and I. M. Sokolov, J. Phys. A [**43**]{}, 055001 (2010). I. M. Sokolov and I. Eliazar, Phys. Rev. E [**81**]{}, 026107 (2010). H. S. Carslaw and J. C. Jaeger, [*Conduction of Heat in Solids*]{}, (Oxford University Press, Oxford, 1959). A. Metzler, Stat. Prob. Lett. [**80**]{}, 277 (2010). B. D. Hughes, [*Random Walks and Random Environments*]{}, (Oxford University Press, Oxford, 1995). M. J. Ward and J. B. Keller, SIAM J. Appl. Math. [**53**]{}, 770 (1993); I. V. Grigoriev, Y. A. Makhnovskii, A. M. Bereshkovskii and V. Y. Zitserman, J. Chem. Phys. [**116**]{}, 9574 (2002); Y. Levin, M. A. Idiart and J. J. Arenzon, Physica A [**354**]{}, 95 (2005); O. Bénichou and R.  Voituriez, Phys. Rev. Lett. [**100**]{}, 168105 (2008); G. Oshanin, M. Tamm and O. Vasilyev, J. Chem. Phys. [**132**]{}, 235101 (2010); O. Bénichou et al., Phys. Rev. Lett. [**105**]{}, 150606 (2010); A. Singer, Z. Schuss, D. Holcman and R. S. Eisenberg, J. Stat. Phys. [**122**]{}, 437 (2006); A. F. Cheviakov, M. J. Ward and R. Straube, Multiscale Model. Simul. [**8**]{}, 836 (2010). D. Boyer, D.S. Dean, C. Mejia-Monasterio, and G. Oshanin, Phys. Rev. E [**85**]{}, 031136 (2012).

--- abstract: 'A canonical pseudo-Nambu Goldstone Boson (pNGB) can play the role of a dark energy field responsible for present cosmic acceleration. Confronting with the recent cosmological data, we find that the pNGB field requires spontaneous symmetry breaking scale $f$ close to $M_P$ and the initial field value fine-tuned. It is difficult to achieve a large $f$ in a theoretically consistent set-up. A possible resolution can be achieved by increasing the Hubble friction in well motivated particle physics models in the general set-up of modified gravity theories. We show two phenomenological examples of this set-up where the standard pNGB action have been modified by introducing terms motivated from galileon cosmology. We confront those examples with the recent supernovae, PLANCK and BAO data. We find that moderate values of the dimensionless constants that increase the friction, make $f << M_{P}$ and the generic initial conditions also favourable by the data. We also comment how the fifth force constraints arising in these modified theories can be evaded.' author: - Debabrata Adak - Koushik Dutta bibliography: - 'pNGB.bib' title: 'Viable dark energy models using pseudo-Nambu-Goldstone bosons' --- Introduction {#intro} ============ Presently we live in a strange Universe where the recent observations suggest that the Universe has started to accelerate very recently [@Spergel:2006hy; @Riess:1998cb; @Perlmutter:1998np]. With the assumption of the General Theory of Relativity being the correct description of physical phenomenon at the cosmic scale, and the effects of inhomogeneities can be neglected, recent cosmic acceleration can be explained by the addition of some exotic matter with negative pressure (broadly called “dark energy”) that has started to dominate the total energy budget of the Universe. The cosmological constant ($\Lambda$) is the simplest source of dark energy that can solve this problem. Moreover, the single parameter solution of introducing $\Lambda$ in the Einstein-Hilbert action is an excellent fit to the all observational data [@Ade:2013zuv]. But the required value of $\Lambda$ to match the observations is very tiny: $\Lambda\approx(2\times10^{-3} ~eV)^4$. Even if the cosmological constant is the reason behind the cosmic acceleration, its tiny value is extremely difficult to justify from the present theoretical understanding of the particle physics. Additionally, in the context of cosmological constant, there is no dynamical resolution to the question of why the Universe is getting dominated by the cosmological constant very recently. Surely, the anthropic arguments in the context of String Theory landscape scenario is a possibility [@Weinberg:1987dv]. Another well motivated idea of the source of dark energy is the existence of a scalar field $\phi$ (‘quintessence’) whose present energy density is approximately the above mentioned value of $\Lambda$ [@Wetterich:1987fm; @Ratra:1987rm; @Caldwell:1997ii]. For reviews on dark energy see [@Copeland:2006wr], and for its effective field theory approach see [@Bloomfield:2011np; @2013PhRvD..87h3504M]. The scalar field $\phi$ being dynamical, in contrast to the cosmological constant, its pressure $p$ is related to the energy density $\rho$ by an equation of state parameter $w \neq -1$ where $p = \omega \rho$. The fact that the cosmological constant (with $w = -1$) is still a very good fit to the data forces $w$ for the scalar field not being too away from $-1$. This essentially translates to a very flat potential for $\phi$ with its mass around $10^{-33}$ eV. Even though the idea of quintessence is very well motivated, finding a natural candidate for $\phi$ in particle physics is very challenging. It is due to the following two reasons. Firstly, quantum corrections due to the other fields in any theory will generically spoil the flatness of the potential [@Kolda:1998wq]. Secondly, the ultra-light $\phi$ field would carry fifth-force, typically of gravitational strength [@Carroll:1998zi; @Chiba:1999wt]. All the models of dark energy are plagued with these problems, unless a symmetry protects its mass and the fifth-force constraints are avoided by some mechanisms. These problems can be easily solved for the case of a pNGB potential whose mass is protected by a shift symmetry [@Frieman:1995pm]. The breaking of the symmetry in a controlled way allows us to keep the mass of $\phi$ radiatively stable and “naturally” light. A pNGB field can only couple to other fields with its derivative couplings, and it naturally suppresses the fifth-force constraints. A nice feature of the pNGB potential is that it has only two free parameters: parameter $f$ is related to the shift symmetry breaking scale, and $\mu$ is related to the explicit symmetry breaking scale. The present value of the cosmological constant fixes the value of $\mu$. Because of the periodic nature of the pNGB potential, the initial field range is compact ranging from $0$ to $2\pi$. In addition, $f$ can not be larger than the reduced Planck mass $M_P$, allowing us to constrain or possibly rule out the potential completely by confronting with the data. Following this line of thought, a detailed study of the parameter space of single pNGB potential was done in [@Dutta:2006cf]. It was found that unless the initial field values are chosen in a fine-tuned way, the data always prefer $f$ to be being close to $M_P$. In this work, we will reanalyse the situation in the light of the latest available cosmological data and reassure ourselves that the constraints have indeed become much severe. Now it is very difficult to arrange for any realistic theoretical set-up where the $f$ is close to $M_{Pl}$ [@Banks:2003sx][^1]. In our work, we consider this theoretical obstacle seriously and look for its possible resolutions. The $f$ parameter essentially controls the slope of the potential and larger $f$ makes the potential flatter which is preferred by the data. A scalar field with a canonical kinetic energy term rolls down the potential too fast to be observationally consistent unless we let the field roll from the top of the potential. It was suggested that $N$ pNGB fields can also collectively drive the present epoch of cosmic acceleration, where $f$ for individual pNGB can be much lower than the $M_P$, but the effective $f_{eff} = \sqrt{N} f$ can be easily close to $M_P$ making it observationally consistent [@Kaloper:2005aj]. In this work, we propose two well motivated modification to the single field pNGB potential, and show that even smaller values of the $f$ being much smaller than $M_{Pl}$, the parameter is observationally consistent for any reasonable choice of initial condition for the field values. The central idea for both these modifications is to add extra terms (motivated by particle physics considerations) in the action that effectively increases the friction in the expanding Universe governed by Friedman equation. Among several modifications to the Einstein’s gravity, a particularly well motivated class of models is the Galileon gravity based on the symmetry of the Galileon field $\pi$ as $\pi \rightarrow \pi + a + b_{\mu}x^{\mu}$ where $a$ and $b_{\mu}$ are constants [@Nicolis:2008in; @Deffayet:2009wt][^2]. This symmetry is motivated by the effective field theory of the decoupling limit of the DGP model [@Dvali:2000hr]. But in contrast to the DGP model, the Galileon theories are free of ghost as its equations of motions have derivatives up to second order. Based on this symmetry, five different field Lagrangians can be written based on the orders of the field derivatives. In this paper, in addition to the standard kinetic energy term of the pNGB field, we also consider the term that involves fourth derivatives of the field, namely $\mathcal{L}_3 = (\Box \phi)(\nabla \phi)^2/M^3$. As we will show, addition of this term modifies the Hubble equation in such a way that it increases the Hubble friction for the field $\phi$. When the idea of Galieleon invariance is generalised to curved background, an additional term also can be added in the standard Einstein-Hilbert action: $\mathcal{L} = -\frac{1}{2M^2}G^{\mu\nu} \partial_\mu \phi \partial_\nu \phi$, where $G^{\mu \nu}$ is the Einstein tensor [@Germani:2010hd], [@Germani:2011bc]. Again, we will see that this term also effectively increases the Hubble friction. In the present paper we will discuss the effects of these two terms in pNGB dark energy models and confront those with the latest SN, CMB and BAO data. This paper is organised as follows. In the next section we review the standard PNGB dark energy. and latest constraints on its parameter space. In section III, we will outline the procedures of confronting dark energy models with recent SN, CMB and BAO data, and will confront the standard pNGB dark energy with canonical kinetic energy with data. This needs to be compared with the earlier work in [@Dutta:2006cf]. In section IV, we will outline the ideas in resolving the “high-$f"$ problem and confront those with data. It would be clear that even small values of the $f$ parameter is well suited with observations. In the last section we will conclude. pNGB dark energy {#review} ================ Once it is assumed that the source of dark energy is the existence of a quintessence scalar field, there are effectively infinite number of potentials that can serve the job required by observations [@Sahni:2004ai]. Moreover, many of these phenomenological potentials come with several parameters (often non-compact), leaving us with ample opportunities to fit the data. But a very important desirable property of the quintessence potential is that the potential must be stable under quantum corrections originating from the couplings to other quantum fields. To say it other way, it is not easy to write down a potential for a scalar field whose mass can be kept at $10^{-33} eV$. The rescue can come from a symmetry, in this case, it is a shift symmetry: $\phi \rightarrow \phi + c$. Exact shift symmetry allows only the constant potential with zero mass. But the symmetry can be broken in controlled way to give the field a suitably light mass. The above mentioned idea can be realised in the particle physics when a global $U(1)$ symmetry is spontaneously broken, giving rise to a massless Goldstone boson. The spontaneous breaking of a global symmetry gives rise to two modes namely the radial modes that get massive and the angular modes which remains massless at the spontaneous symmetry breaking energy scale. These massless angular modes are called the NGBs. Now these spin-0 massless NGBs acquire masses (making those pseudo-NGBs) when there is another soft explicit breaking of the global symmetry at a lower energy scale compared to the spontaneous symmetry breaking scale. The pNGB potential is characterised by these two symmetry breaking scales, spontaneous global symmetry breaking scale $f$ and explicit global symmetry breaking scale $\mu$, and it is given by $$\begin{aligned} V(\phi)&=& \mu^4\left[1+\cos\left(\frac{\phi}{f}\right)\right]\,\,. \label{pngb-pot}\end{aligned}$$ The value of $f$ determines the steepness of the pNGB potential. As the value of $f$ increases, the pNGB potential becomes flatter. In fact $f \rightarrow \infty$ corresponds to the exact shift symmetric constant potential. pNGBs were first proposed in the context of natural inflation [@Freese:1990rb], and then it was subsequently extended for the case of dark energy [@Frieman:1995pm]. We consider the dynamics of a pNGB quintessence field with potential $V(\phi)$ of Eq. in a flat Friedmann-Robertson-Walker Universe governed by the Einstein’s equations. The Lagrangian for a canonical pNGB field in this case looks like $$\begin{aligned} \cal{L} &=& \frac{M_{pl}^2}{2} R - X -V(\phi)\end{aligned}$$ where $X=\frac{1}{2}\partial_\mu\phi\partial^\mu\phi$. The equations of motion in the late universe containing matter are given by the Friedmann equation $$\begin{aligned} H^2&=& \frac{1}{3 M_P^2}\left(\frac{1}{2}\dot\phi^2+V(\phi) +\rho_m\right) \,\,,\label{fried1} $$ and Klein Gordon equation for the scalar field $$\begin{aligned} \ddot\phi+3H\dot\phi+\frac{dV}{d\phi}&=&0\,\,, \label{pngb-eos}\end{aligned}$$ where $H = \dot a/a$ is the Hubble constant and $\rho_m$ is the energy density of the non relativistic matter. Note that the effects of radiation energy density is negligible for the late time cosmology that we are concerned with. We use these equations to obtain the evolution of the scalar field and Hubble parameter to constrain the model parameters. We will compare the latest observational constraints with the earlier work done in [@Dutta:2006cf]. $\chi^2$ analysis of Observational data {#Obs} ======================================= We use the latest type Ia supernovae data, Cosmic microwave background shift parameter data and baryon acoustic oscillation data to constrain the two different models discussed above. Type Ia supernovae are considered to be the standard candles in astrophysics. Measurements of luminosity distance ($d_L$) of the type Ia supernovae with their redshifts happened to be the first probe [@Perlmutter:1998np] of the discovery that the Universe is undergoing an accelerated phase of expansion in the present epoch. The lateset compilation of Union 2.1 SNe Ia data has been performed by Suzuki [it et. al.]{}[@Suzuki:2011hu] and we use this set of 580 data points to constrain our model parameters space. The distance modulus is defined as $$\begin{aligned} \mu(z)&=& 5 \log_{10}(D_L(z))+\mu_0,\end{aligned}$$ where $D_L(z)=H_0 d_L(z)/c$ ($c$ is the speed of light in vacuum) and $\mu_0=42.38- 5 \log_{10} h$ with $H_0=100h ~{\rm Km Sec^{-1} Mpc^{-1}}$. $\chi^2_{\rm SNe}$ is defined as $$\begin{aligned} \chi^2_{\rm SNe}(p_s)&=& \sum_i \left[\frac{\mu_{\rm obs}(z_i)-\mu_{\rm theo}(z_i,p_s)}{\sigma_i}\right]^2\,\,,\end{aligned}$$ where $p_s$ correspond to the model parameters that are constrained with the type Ia supernovae data. We marginalise the $\chi^2_{\rm SNe}$ over the nuisance parameter $\mu_0$ and use that marginalised $\chi^2_{\rm SNe}$ for the data analysis perpose. Cosmic microwave background shift parameter $R$ is extracted from the first peak in the cosmic microwave background temparature anisotropy plot. This is more or less is a model independent parameter that is greatly used in constraining the dark energy models. Shift parameter is defined as $$\begin{aligned} R(z_*)&=& (\Omega_m^0 H_0^2)^{1/2}\int_0^{z_*}\frac{dz}{H(z)}\,\,,\end{aligned}$$ where $z_*$ is the radiation-matter decoupling redshift.$\chi^2_{\rm CMB}$ is defined as $$\begin{aligned} \chi^2_{\rm CMB}&=&\left[\frac{R(z_*,p_s)-R}{\sigma_R}\right]^2\,\,.\end{aligned}$$ We have used the CMB shift parameter from the Planck results $R=1.7499\pm 0.0088$ at the decoupling redshift $z_*=1090.41$ [@Planck:2013nga]. At the very high energy of the early Universe, the baryons are simultaneously acted upon by the two oppositely directed forces namely attractive gravitational force which tries to take them closer and the radiation pressure which tries to take them away from each other. As a result of these two opposite forces an oscillation comes into play in the baryon photon plasma and the disturbence travel through the baryon photon medium with a sound speed which is close to the speed of light. As the Universe cools down the sound speed drops down and after the radiation matter decoupling the sound speed drops down to zero and the disturbance gets frozen in the large scale structure. This phenomenon is known as the baryon acoustic oscillation (BAO) and is observed as the excess number of galaxies at a certain length scale. This is been being measured by the Sloan Digital Sky Survey (SDSS) in the form of two point galaxy correlation function [@Tegmark:2006az]. We use the BAO data of $\frac{d_A(z_\star)}{D_V(Z_{BAO})}$ [@Percival:2009xn; @Beutler:2011hx; @Jarosik:2010iu; @Blake:2011en], where $z_\star$ is the decoupling redshift given by $z_\star \approx 1091$, $d_A$ is the comoving angular-diameter distance given by $d_A(z)=\int_0^z \frac{dz'}{H(z')}$ and $D_V(z)=\left(d_A(z)^2\frac{z}{H(z)}\right)^{\frac{1}{3}}$. We calculate $\chi_{BAO}^2$ as described in Ref. [@Giostri:2012ek], where it is defined as, $$\chi_{BAO}^2=X_{BAO}^T C_{BAO}^{-1} X_{BAO}\,\,,$$ where $$X_{BAO}=\left( \begin{array}{c} \frac{d_A(z_\star)}{D_V(0.106)} - 30.95 \\ \frac{d_A(z_\star)}{D_V(0.2)} - 17.55 \\ \frac{d_A(z_\star)}{D_V(0.35)} - 10.11 \\ \frac{d_A(z_\star)}{D_V(0.44)} - 8.44 \\ \frac{d_A(z_\star)}{D_V(0.6)} - 6.69 \\ \frac{d_A(z_\star)}{D_V(0.73)} - 5.45 \end{array} \right)$$ and the inverse covariance matrix $C_{\rm BAO}^{-1}$ is given in the Ref. [@Giostri:2012ek]. We perform a combined analysis of all the data sets together by making a combined $\chi^2_{tot}$ given by, $$\begin{aligned} \chi^2_{tot}&=&\chi^2_{SN}+\chi^2_{CMB}+\chi^2_{BAO}\,\,.\end{aligned}$$ We minimize this $\chi^2_{tot}$ and find best fit values as well as the $2\sigma$ C.L.s for the model parameters. Constraining standard pNGB dark energy {#pNGB} ====================================== Using the above mentioned analysis techniques in confronting dark energy models with data, we will now reanalyse the observational viability of a standard pNGB dark energy field using the latest available data sets. In doing so, we solve Eq. and Eq. numerically and calculate the Huuble parameter as a function of redshift. =3.4truein A pNGB dark energy model has two parameters, namely $f$ and $\mu$ and two initial conditions $\phi_{in},~\dot\phi_{in}$. We use $\dot\phi_{in}=0$ in our calculations, and the assumption is reasonable considering that the large Hubble damping would typically make the field roll slowly at the initial stage. The present value of the dark energy density $\Omega_{\phi}^{(0)} \simeq 0.7$ can be traded with the value of $\mu$, and we are left with the parameter $f$ and the initial field value $\phi_{in}$. It is worth to remember that because of the periodic nature of the pNGB potential, $\phi_{in}$ can vary between $0$ and $2\pi f$, where the spontaneous symmetry breaking scale $f$ has a natural theoretical cut-off of the order of $M_P$. In principle, we can thus hope to exclude the whole model. =3.4truein With $\Omega_\phi^{(0)}=0.7$ fixed, we plot the confidence contours in terms of $\phi_{in}~vs~f$ in Fig. (\[fig1\]). The upper left shaded portion of the plot (yellow) is excluded by the fact that for those choice of parameters, the Universe can not evolve to a stage where $\Omega_\phi^{(0)}=0.7$. Looking at the $2 \sigma$ confidence contour marked by the light grey area, it is clear that $f$ close to $M_{Pl}$ is preferred observationally. Unless the initial field value is chosen carefully close to the top of the potential ($\phi_{in}\simeq 0$), $f \lesssim \mathcal{O}(0.5) M_{Pl} $ would be $2\sigma$ excluded. The nature of the confidence contours close to small $f$ and $\phi_{in}$ values tells that the statement remains more or less true with higher significance. In fact, the effective $\chi^2$ is minimum for $\phi_{in} \simeq 0$ for any values of $f$. But, it changes sharply when we increase $\phi_{in}$ value for smaller values of $f$. The field does not experience the steepness of the potential around $\phi_{in} \simeq 0$, and the effective dynamics is independent of $f$. This makes all values of $f$ equally probable for $\phi_{in} \simeq 0$. As the field moves from $\phi_{in} \simeq 0$, it starts to experience the slope of the potential and rolls down accordingly depending on $f$. Therefore if the potential is steep enough the field will quickly evolve to the present epoch excluding the most part of the parameter space in Fig. (\[fig1\]). Note that the initial field values at the top of the potential is also very unlikely considering the quantum fluctuations produced by the earlier inflationary epoch at higher scale. Thus we are observationally pushed to a very high value of the spontaneous symmetry breaking scale $f \simeq M_{Pl}$, and it is difficult to accommodate from our theoretical understanding. This is what we call the “high-$f$” problem of pNGB quintessence [@Kaloper:2005aj]. In the next section, we will outline two possible resolutions of this problem and will discuss the observational implications. We also compare our findings with the previous analysis derived in [@Dutta:2006cf]. We note that the $2\sigma$ region has shrunk considerably where the dotted line marks the boundary of $2\sigma$ contour of previous analysis. In fact, the present $5\sigma$ contour is within the previous $2\sigma$ contour. The comparison between these two figures indicates the improvement of the data points. The authors in the paper [@Dutta:2006cf] used 182 Gold SNe Ia data points [@Riess:2006fw] along with the CMB shift parameter constrain from WMAP 3 year data [@Wang:2006ts], whereas in this work we use 580 SNe data points from Union2.1 compilations [@Suzuki:2011hu], CMB shift parameter from PLANCK data [@Planck:2013nga] and the baryon acoustic oscillation data from SDSS [@Percival:2009xn; @Beutler:2011hx; @Jarosik:2010iu; @Blake:2011en]. =3.4truein It is worth mentioning here that in the parameter space just bellow the theoretical bound, but outside the dotted line, we have a small region for which $\Omega_\phi$ oscillates around $\Omega_\phi^{(0)}$. This corresponds to the case where the scalar field crosses the potential minima and starts climbing up slowly. We can also identify the present epoch when the scalar field is climbing up rather than rolling down. But it is clear from the plot that the parameters that allow this situation is outside the $5\sigma$ confidence limit from the present analysis[^3]. In Fig. (\[fig2\]), we show the $2\sigma$ confidence regions in $\phi_{in}~vs~\Omega_\phi^{(0)}$ plane for $f=M_{pl}$. We plot the confidence contours for each individual observations, and the combined confidence contour (dot-dashed curve) is plotted in Fig. \[fig3\]. Even though all three different observations have vertical contours, but the combined plot produce a very stringent constraints in $\phi_{in}$ and $\Omega_\phi^{(0)}$ plane. This is more evident when the present constraints are compared with the old analysis of [@Dutta:2006cf] as shown in the dotted curve of Fig. \[fig3\]. The preferred value is around $\Omega_\phi^{(0)} \simeq 0.7$ with initial field value close to the top of the potential. There is no sign of evolution of pNGB field. When the similar plot is drawn for smaller values of $f$, the broad feature of the confidence contour (solid curve) remains the same, except that the contour shrinks along the $\phi_{in}$ axis, forcing the field to start closer to the top of the potential. This is in accordance with what we see in Fig. (\[fig1\]). =3.4truein In Fig. (\[fig4\]), we show the variation of the the dark energy equation of state $w$ as a function of the redshift $z$. The plots carry the clear signature that the best fit values mimick the cosmological constant behaviour. The deviations of $w$ from $w=-1$ happens near the present epoch showing the thawing like behaviour of dark energy [@Linder:2007wa; @Adak:2012bv]. Being stuck at the potential due to Hubble damping, the field shows very little evolution, and it is almost mimicking the cosmological constant. This is also clear from the plot in Fig. (\[wz1\]) where the present value of the dark energy equation of state parameter $w_0$ is shown against the present dark energy density. Even though this constraint was understood earlier, with recent data the parameter space has shrunk considerably. Within the $2\sigma$, the present equation of state parameter $w_0$ is constrained to be $w_0 < -0.92 $. =3.4truein The main finding in this section is that if a pNGB field is responsible for dark energy, the data favours large values of the $f$ parameter. Obviously the necessity of $f$ being large can be easily alleviated by choosing fine-tuned initial conditions. But we argue that this choice is unrealistic considering any prior inflationary epoch [@Kaloper:2005aj]. On the other hand, due to lack of theoretical understanding, it is difficult to construct models of pNGB where the decay constant $f$ is close to $M_{Pl}$. In the next sections, we will propose two well motivated particle physics modifications of the standard pNGB that can easily accommodate data even with smaller values of the $f$ parameter. Helping hands {#MP} ============= In the previous section, we have seen that a pNGB quintessence field with canonical kinetic energy term is under strain with observational data. Only a fine-tuned initial conditions for the field with $f$ close to $M_{Pl}$ can mimic the observed expansion history of the Universe. On the other hand, having theoretically consistent model with large values $f$ is difficult to construct. The [*helping hands*]{} can come from the modifications to the Einstein’s gravity. Here we propose two such modifications on the phenomenological ground motivated by the Galileon gravity and its generalisations. As we will see, both these examples will modify the Hubble equation enhancing the friction allowing small values $f$ observationally consistent. We note that a general Galileon field has usually non-derivative couplings to the matter sector. To evade the fifth-force constraints arising from these terms, the applicability of the Veinshtein mechanism is necessary [@Vainshtein:1972sx]. At the same time, the couplings to the matter those are of derivative types (if present) can circumvent the long-range force constraints by the usual Adler decoupling [@Carroll:1998zi] [@Kolda:1998wq] Example 1 {#ex1} --------- We start with the following Lagrangian $$\begin{aligned} \cal{L} &=& \frac{M_{pl}^2}{2} R - X - F(\phi)X \Box \phi -V(\phi),\end{aligned}$$ where the third term is just the $\mathcal{L}_3$ in the Galileon Lagrangian. Here $\Box\phi=\frac{1}{\sqrt{-g}} \partial_\mu\left(\sqrt{-g}\partial^\mu\phi\right)$. The Lagrangian is also a special case of the more general “Kinetic Gravity Braiding" dark energy models [@Deffayet:2010qz] $$\mathcal{L} \supset K(\phi,X)+G(\phi,X)\Box \phi\,\,.$$ These models though contains higher than the second derivative of the scalar fields, the equation of motion of the scalar fields do not contain more than second derivative of the scalar field and thus these theories are free from Ostrogadsky ghosts [@ostro]. For simplicity we explore the situation where $F(\phi)$ is a constant and is given by $F(\phi)=\frac{1}{M^3}$, where $M$ is a new mass scale. =3.4truein The Friedmann equation and the equation of motion for the scalar field are given by, $$3 M_{pl}^2 H^2 = \frac{\dot{\phi}^2}{2} \left(1+\frac{6}{M^3}H\dot\phi\right) +V(\phi) + \rho_m$$ $$\ddot \phi +3 H \dot\phi +3 \frac{1}{M^3} \dot\phi \left(3 H^2 \dot\phi+\dot H \dot\phi +2 H \ddot\phi\right) +V'(\phi) = 0$$ where $\rho_m$ is the matter density. To solve the model numerically, we transform the quantities in dimensionless variables, and we have 5 parameters $(\phi_{in},~\dot\phi_{in},~f,~\mu,~\alpha)$ where $\alpha$ is a dimensionless parameter given by $\alpha=H_0^2 M_{pl}/M^3$, and parametrises the new physics. If the new physics scale $M=M_{pl}$, we find $\alpha \sim 10^{-124}$, without having any effect on the pNGB dynamics. The new mass scale should be such that the enough amount of hubble friction is produced. For reasonable values of the dimensionless constant $\alpha \sim \mathcal{O}(1) - \mathcal{O}(100)$ the hubble friction makes the field roll down along the pNGB potential slowly independent of the steepness of the potential. Assuming $\dot\phi_{in}=0$ and the flatness condition, we are left with only four parameters $(\phi_{in},~f,~\Omega_\phi^0,~\alpha)$ where $\alpha$ parameterize the increased Hubble friction. $\alpha = 0$ corresponds to the standard pNGB case as discussed in section \[pNGB\]. We choose different values of $\alpha$ with $\Omega_{\phi}^{(0)} = 0.7$ and plot the 2-$\sigma$ contours in the $\phi_{in}-f$ plane in Fig. \[fig1beta-nonzero\]. Increasing the value of $\alpha$, the allowed region expands and smaller values of $f$ become equally favoured by the observations. At the same time, the field need not necessarily start to roll from the top of the potential. This implies that those initial values of $\phi_{in}$ which are disfavoured by the observations in standard pNGB case ($\alpha=0$), becomes allowed as the values of $\alpha$ increases. This is also clear from the Fig. \[fig3beta-nonzero\] which has been plotted for $f = M_{Pl}$. =3.4truein Example 2 {#ex2} --------- In the following, we show another example that can do the similar job as like the previous example. The Lagrangian in this case is given by $$\begin{aligned} \cal{L} &=& \frac{M_{pl}^2}{2} R -\frac{1}{2}g^{\mu\nu} \partial_\mu \phi \partial_\nu \phi +\frac{1}{2M^2}G^{\mu\nu} \partial_\mu \phi \partial_\nu \phi -V(\phi)\nonumber\\\end{aligned}$$ where $G^{\mu \nu}$ is the Einstein tensor, and $M$ is a mass scale not necessarily linked to the previous example. The above modifications to the Einstein gravity was proposed in [@Germani:2010hd] in the context of inflation. In the context of modified gravity theories, the effects of the term have been studied in [@Adak:2013vwa], but with different potentials and for different motivations. The phenomenology of the Lagrangian in the context of inflation was discussed in [@Maity:2012dx]. Friedmann equations and the equation of motion for the scalar field are given by, $$\begin{aligned} 3M_{pl}^2 H^2&=&\frac{1}{2}\dot{\phi}^2+V(\phi)+\frac{9}{2M^2}H^2\dot\phi^2 +\rho_m,\\ M_{pl}^2(2\dot H +3 H^2) &=&-\frac{1}{2M^2}\dot\phi^2+V(\phi) \nonumber\\&& +\frac{1}{2M^2}\left((3H^2+2\dot H)\dot\phi^2+4H\dot\phi\ddot\phi\right)\,\,,\end{aligned}$$ $$\begin{aligned} \ddot \phi +3 H \dot\phi + \frac{1}{2M^2}\left(6H^2\ddot\phi+18H^3\dot\phi +12H\dot H\dot\phi\right) +V'(\phi) &=& 0\,\,.\nonumber\\\end{aligned}$$ where $\rho_m$ is the matter density. It is important to note that there are no higher derivative terms in the equations of motions. =3.4truein Like in the previous example, here also we use a dimensionless parameter $\beta=H_0^2/M^2$ that captures the effect of the non-minimal coupling of the Eienstein tensor to the kinetic term. The role of $\beta$ is similar to $\alpha$ in the previous example. The effect of non-zero $\beta$ has been shown in the Fig. \[fig1alpha-nonzero\]. It is clear that for moderate values of $\beta$, a large portion of the parameter space becomes allowed. Discussions and Conclusions =========================== If observed cosmic acceleration is due to a dynamical degree of freedom, namely a quintessence field, a pNGB scalar field is the most suitable candidate. This is due to the the stability of its mass from quantum corrections, as well as its ability to evade the fifth force constraints. But once we confront a pNGB potential with cosmic data, we see that the data favours large value of the spontaneous symmetry breaking $f$-parameter, as well as initial field value at the top of the potential. On the other hand, it has been argued that constructing theoretically consistent model with $f \simeq M_{Pl}$ is very challenging. As we have seen in section IV, with $f <0.1 M_{Pl}$, the field must start to roll exactly from the top of the potential. If we combine it with the physics prior to the dark energy, say inflation, these initial conditions are very unlikely due quantum fluctuations. Combining all these ingredients, it seems that a standard pNGB is under strain. Even though it was understood earlier [@Dutta:2006cf], we see that the recent observational data has severely constrained the parameter space. In particular, the present analysis done in section IV includes the BAO data, and it has a significant effect in constraining the model. In summary, the observation pushes the model to have large values of $f$ being close to $M_{Pl}$ with initial field values close to the top of the potential. But both these requirements are difficult to meet if we take our theoretical prejudices seriously. The way out to the problem may come from the introduction of several pNGB fields driving quintessence. But we propose other alternatives with single pNGB driving quintessence. In particular, we work on two alternatives where the Einstein Hilbert action is modified. These modifications are motivated by the Galieleon cosmology and its generalisations. Even though the the Lagrangian involves higher derivative terms, the equations of motions are of second derivative. For both these examples the main goal is to increase the friction due to non canonical terms. We then confront these modified models with data to show that for $\mathcal{O}(1) - \mathcal{O}(10)$ values of the dimensionless constants, a large part of the parameter space become viable. From the values of $\alpha$ in Sec. \[ex1\] and $\beta$ in Sec. \[ex2\], it is evident that the new mass scale $M$ has to be of the order of $H_0$ which is much smaller than the Planck scale $M_{Pl}$. In fact the modified pNGB reduces effectively to the standard pNGB when $M = M_{Pl}$. Unless we have a more fundamental theory, we can not know what might be the source of this mass scale, but it is clear that to have the significant effect in the cosmic expansion at the present epoch, it must be of the order of $H_{0}$. In our work, even though we had analysed two examples separately for simplicity, but the combined effect can also be analysed. All these modified Lagrangians are just subset of a larger class given by Hordenski, and dark energy phenomenology of these general Lagrangian have been analysed already. In the standard pNGB dark energy case, a pNGB scalar field that gives rise to late-time cosmic acceleration, couples to the matter derivatively [@Carroll:1998zi]. These interactions lead to a fifth-force that is suppressed by the Adler decoupling in the low momentum limit. On the other hand, the couplings of a Galileon scalar field to the matter is of non-derivative types $ S_m[\psi_m; e^{2\beta \phi/M_{pl}} g_{\mu\nu}]$ [@Ali:2012cv], and to suppress the fifth-force in this case, it is required to invoke the Vainshtein mechanism. In our case, motivated by the Galileon theories, we have introduced $\frac{1}{M^3}X\Box \phi$ and $\frac{1}{2M^2}G^{\mu\nu}\nabla_\mu \phi \nabla_\nu \phi$ terms to resolve the “high $f$” issue of axionic decay constant. As these terms are motivated by the Galileon gravity, it is natural to expect that the couplings of the scalar field to the matter sector will contain both derivative or non-derivative interactions. The derivative couplings of the scalar field gives rise to fifth force which is highly suppressed by Adler decoupling mechanism [@Carroll:1998zi] [@Kolda:1998wq]. The non-derivative couplings with matter (can arise from soft breaking of shift symmetry) can give rise to fifth-force constraints, but those are usually suppressed by the Vainshtein mechanism [@Ali:2012cv] [@Kase:2013uja]. In this work we have studied the recent observational constraints on the pNGB dark energy. We find that the low $f$ region ($f<M_{pl}$) is highly constrained and about to be ruled out by the observations. In this situation, we offer two “helping hands” strictly on the phenomenological ground and study the effects of these terms in resolving the “high $f$” issue of the pNGB dark energy. These two terms are well motivated in the context of Galileon cosmology [@Ali:2012cv] as well as in the context of “Kinetic Gravity Braiding” dark energy models [@Deffayet:2010qz] and have been studied earlier in the context of inflation [@Maity:2012dx] [@Germani:2010hd]. We have shown that in presence of these terms, low $f$ values go well with the observations. This is the main result of this work in addition to revising the constraints with the latest available observational data. It would be really interesting to find a concrete theoretical construction where a pNGB can be a Galileon type field. We leave this effort for future work. Acknowledgement =============== D.A is supported by the fellowship from SINP. K.D is partially supported by the Ramanujan Fellowship and Max Planck Society-DST Visiting Fellowship. We thank all the participants of the Friday Cosmology Discussions $@$ SINP for keeping us moving. [^1]: A pNGB potential is used for natural inflation in [@Freese:1990rb]. The situation is similar in the context of inflation due to slow-roll conditions, but the problem is severe as the data requires $f > 3.5 M_P$ [@Ade:2013uln]. [^2]: Galileon theories are subclass of the most general theories proposed by Horndeski [@Horndeski:1974wa] with nonminimal interactions between the scalar field and gravity that give second order field equations in four dimensions. [^3]: We thank Lorenzo Sorbo for making a clarifying point in this regard.

--- abstract: 'Hot cavity resonant ionization laser ion sources (RILIS) provide a multitude of radioactive ion beams with high ionization efficiency and element selective ionization. However, in hot cavity RILIS there still remains isobaric contaminations in the extracted beam from surface ionized species. An ion guide-laser ion source (IG-LIS) has been implemented that decouples the hot isotope production region from the laser ionization volume. A number of IG-LIS runs have been conducted to provide isobar free radioactive ion beams for experiments. Isobar suppression of up to 10$^{6}$ has been achieved, however, IG-LIS still suffers from an intensity loss of 50-100$\times$ as compared to hot cavity RILIS. Operating parameters for IG-LIS are being optimized and design improvements are being implemented into the prototype for robust and efficient on-line operation. Recent SIMION ion optics simulation results and the ongoing development status of the IG-LIS are presented.' author: - 'M. Mostamand' - 'R. Li' - 'J. Romans' - 'F. Ames' - 'P. Kunz' - 'A. Mj$\phi$s' - 'J. Lassen' title: | Production of clean rare isotope beams at TRIUMF\ Ion guide laser ion source --- Introduction {#intro} ============ Modern experiments at isotope separator on-line (ISOL) facilities like TRIUMF’s isotope separator and accelerator (ISAC) often depend critically on the purity of the delivered rare isotope beams. Therefore, highly selective ion sources are required. The mass separator magnet at ISAC was designed for $A/q \leq 30$ and an ion source emittance $\leq 30$$\pi$mm.mrad. As such the $m/\delta m \sim 2000$ with high throughput allows for isobar separation. As ISAC radioactive ion beams delivered have expanded to $A/q \leq 240$, there is need for additional selectivity. Isobaric beam contamination often arises from surface ionization of elements with low ionization potential ($\leq 6$eV) on the hot target or transfer tube surfaces. The resulting background can be especially large for exotic nuclides far from stability and produced at low rates which stems from the fact that the abundance ratio becomes very unfavorable from a production point of view for ’isotope of interest / background isobar’. The electronic structure of elements, gives an avenue to use laser resonance ionization to provide element-selective ionization and enhance the yield of the isotope of interest. Although resonant laser ionization increases the yield of a desired element, delivered beams can still be overwhelmed by surface ionized isobars. In these cases, the production of a sufficiently pure laser ionized beam requires a means of improving the RILIS selectivity. This problem was tackled by developing a new type of ion source, the ion guide laser ion source (IG-LIS). This type of ion source was proposed by Blaum *et al*. [@blaum_novel_2003] in the form of the laser ion source trap (LIST) as a laser resonance ionization within a segmented radio frequency quadrupole (RFQ). A first implementation of an IG-LIS (previously named RFQ-LIS) for off-line tests at ISAC was done in 2007 with a design [@lavoie_production_2010; @lavoie_segmented_2007] for isobar suppression. Along the RFQ axis the electrodes are segmented into six sections. By applying DC potentials of $0-5$V to the individual segments a potential gradient can be created inside the RFQ to assist longitudinal guidance and extraction of the ions. Through detailed thermal and ion optics simulation studies and off-line tests with stable isotopes [@heggen_2013; @raeder_ion_2014] a more robust and higher efficiency design without segmentation was brought on-line in 2013. This ion guide laser ion source is presented in Fig. \[iglis\]. ![Technical realization of the IG-LIS at ISAC. Shown is a cross section of the device, centered on the 3mm diameter transfer tube that connects to the top of the target container. It features from left to right: isotope production target, transfer tube, heat shield, repeller electrode, RFQ structure, exit electrode, and extraction electrode [@heggen_2013]. Potentials are chosen so that surface ionized species from the hot cavity/target transfer tube are prevented from entering the “cold” ionization volume within the ion guide.[]{data-label="iglis"}](iglis){width="\linewidth"} In a first on-line run with a silicon carbide (SiC) target a suppression of surface-ionized sodium (Na) contaminants in the ion beam of up to six orders of magnitude was demonstrated [@raeder_ion_2014]. In this work, operation and further improvements of this unique laser ion source were tested, an IG-LIS will allow key experiments on the exotic ISOL beams. A number of IG-LIS runs have been conducted at ISAC successfully. Details about mechanical design and performance of the currently used on-line IG-LIS are given in [@heggen_2013]. This paper focuses on optimized operating parameters and design improvements for IG-LIS from SIMION [@dahl] simulations for a robust and efficient on-line operation. The energy spread of ions extracted from the ion-guide is investigated for different operation parameters. The influence of laser parameters on the mass separator transmission of extracted laser ionized species was calculated. A systematic study of the ion signal dependence on repetition rate and laser pulse energy has been performed in offline tests in Leuven [@rept_rate_2012; @rept_rate_2013]. The energy spread in an octupole ion-guide in comparison to the standard quadrupole was studied. Simulations and improvements on IG-LIS {#sec:iglis} ====================================== The quadrupole ion guide based IG-LIS prototype operated at ISAC in 2013 aimed to guide ions and therefore operated in RF-only mode. It uses a square wave RF which allows simple adjustment of the driving frequency and duty cycle. With a field radius of r$_{0}$ = 5mm, a frequency range of 1–3MHz and RF-amplitudes up to 100V, singly charged ions with masses from 2 to 240$^{+}$amu can be transmitted. IG-LIS design for on-line operation has to be radiation hard and compatible with high temperature, high vacuum conditions. The technical requirements and constraints are explained in details in [@heggen_2013; @raeder_ion_2014]. To develop a next generation IG-LIS, a better physical understanding of the IG-LIS prototype, ion optics, and its limiting factors is essential and can be arrived through simulations. Further charged particle trajectory simulations of IG-LIS were carried out in SIMION 8 in view of different DC bias configurations and the use of higher multipole ion guides [@dahl]. Potential array calculations along the beam axis {#sec:potential} ------------------------------------------------ The repeller electrode which is located 1.4mm behind the heat shield has a large aperture of 5mm diameter to avoid the heat radiation from the target and material deposition. The RFQ electrodes are 35mm long and are mounted on a single piece copper bracket. The electrodes are insulated from the copper mount by ceramic washers. This mount also holds an electrically insulated exit electrode with an aperture of 3.5mm diameter. There are two different operation modes for IG-LIS: transmission mode, and suppression mode. Fig. \[potentiala\] presents the on axis potential in both operation modes. Positive source potential in the transmission mode helps to extract more ions from the source towards the RFQ and extraction electrode. ![Potential distribution along the z-axis in IG-LIS for the principal operation modes: (i) transmission and (ii) suppression mode either with (a) positively biased suppression electrode, or (b) negatively biased target. Vertical red dash lines represent the locations for source, heat shield, repeller electrode, and beginning of the 35mm long RFQ. V$_{repeller}$: potential on the repeller, V$_{source}$: potential on the source, RFQ offset: dc potential offset of RFQ. The region from which ions can be extracted is larger in transmission mode and allows for an improved ion extraction from the source without much surface ion suppression. The efficiency of the IG-LIS in transmission mode comes close to that achieved by the standard surface ion source - RILIS configuration. Source bias operation in suppression mode (red curve) compared to repeller operation (blue curve), covers a larger ionization region and has lower potential difference along the axis of cold ionization volume and therefore presents the highest efficiency IG-LIS operation mode. Blue rectangles on the top axis represent the thickness of the electrode. All potentials given are with respect to the source bias.[]{data-label="potentiala"}](potentiala){width="\linewidth"} By applying a voltage higher than the source bias to the repeller electrode, mainly neutral atoms will be able to enter the ionization region where they can be ionized by resonant laser light while surface ionized species are repelled. For online operation, beam tuning and optimization in high transmission mode is advantageous, since it provides intense beams for beam tuning. Potential arrays along the z-axis in IG-LIS for the two operation modes with the standard on-line operation parameters are shown in Fig. \[potential\]. ![Effect of RFQ bias operation with potential distribution along the z-axis in IG-LIS. Red dash lines represent the locations of source exit, heat shield, repeller electrode, and start of the RFQ. V$_{repeller}$: potential on the repeller, V$_{source}$: potential on the source, RFQ offset: dc potential offset of RFQ. Shifting down the potential on offset with respect to the repeller helps to extract more ions. Blue rectangles on the top axis represent the thickness of the electrode.[]{data-label="potential"}](potential){width="\linewidth"} The RFQ offset potential in Fig. \[potential\] helps to increase the kinetic energy of ions towards the RFQ and extraction electrode and consequently enhance the transmission efficiency (larger ionization region). However, the effect of this potential difference along the axis on the overall energy spread of the ions must be considered. The mass separator’s mass resolution of $m/\delta m \sim 2000$ has such an effect that ions with an energy spread above $\sim 10$eV will not be able to pass through the mass separator. Therefore the energy spread of ions extracted from the IG-LIS must be considered. Energy spread {#sec:Es} ------------- For on-line IG-LIS operation, the effect of energy spread on the extracted beam and transmission efficiency through the downstream mass separator is an important factor. The resultant simulation studies presented in this section are particularly useful for establishing high resolution, high throughput mass separator tunes for radioactive isotope beams from IG-LIS. Enhancement of the beam transmission (in suppression mode) to some extent with shifting down the offset potential is reasonable, since lowering the RFQ offset with respect to the source causes larger kinetic energy of the ions and therefore pushes more ions from the source towards the extraction electrode. However, shifting the offset potential results in a larger potential difference in the ionization region and consequently higher energy spread. Fig. \[data\_offset\] presents the effect of RFQ-offset on ion transmission and energy spread in suppression mode. This effect saturates at RFQ-offset = -3V after which there is no further increase in transmission efficiency. ![Effect of RFQ-offset on extracted ion’s energy spread and transmission efficiency in suppression mode. Transmission efficiency saturates at RFQ-offset = -3V.[]{data-label="data_offset"}](data_offset){width="\linewidth"} Possible IG-LIS operational and design improvements {#sec:simulations} =================================================== Simulation studies in this section provide possible improvements of energy spread and overall efficiency in the ion guide. Effect of laser parameters on the ion guide efficiency {#sec:laser rep} ------------------------------------------------------ Calulated IG-LIS transmission efficiency as a function of laser repetition rate is shown in Fig. \[laserep\]. ![Calculated laser repetition rate effect on (a) the extracted laser ionized species and (b) energy spread of the extracted laser ionized beam. IG-LIS ionization reaches optimum efficiency at a repetition rate of 50kHz. Whether this maximum point is exactly at 50kHz or not depends on the ionization efficiency of the laser scheme which is considered to be 100% for simulations in this work. Simulation is performed on mass 200amu.[]{data-label="laserep"}](laserep){width="\linewidth"} An increased laser repetition rate helps to ionize more of the neutral atoms that enter the ion guide region, resulting in higher efficiency. Calculations are based on suppression mode operation and mass 200amu. With increasing duty cycle eventually the low energy tail in Fig. \[laserep\](red curve in (b)) can be suppressed and most of the ions are extracted with the same energy. At repetition rates higher than 50kHz in suppression mode, most of the neutral atoms are laser ionized before they pass the repeller and IG-LIS efficiency is reduced. Thus most of the laser ionized species will be rejected by the repeller electrode and the efficiency drops down. An experimental test of the laser repetition rate behavior was not possible due the the performance characteristics of the available pump lasers (LEE LDP100MQg) which are limited in overall output power and repetition rate maximum of 20kHz. A 5 fold increase in repetition rate would also require the equivalent increase in pump power from currently $\sim$10W/Ti:Sa laser to effectively 50W/Ti:Sa laser. The optimized operating conditions at on-line ISOL type facilities are based on pulsed laser operation to make use of the high pulse peak power for hot cavity RILIS the optimum repetition rate is at about 10kHz, which fits well with available laser technology. In the standard hot cavity RILIS atoms are confined in the approximately 40mm long, 3mm diameter transfer tube, with much larger residence time, as well as a re-introduction of atoms into the ionization volume through spatial confinement of atoms such that ion yields are optimal at $\sim$ 10kHz laser repetition rate. Octupole ion guide {#sec:octupole} ------------------ In an octupole RF ion-guide the concentration of the field near the field radius is more pronounced. The octupole field is usually approximated using eight cylindrical rods. An octupole RF ion guide can transport a range of masses four times larger than the quadrupole ion guide of the same field radius and field parameters. Due to concentration of the field close to the field radius the amplitude of particle motion in an octupole guide is larger than in the quadrupole with similar field parameters. From the study of octupole potential distributions for several geometries, each with different rod-to-field-radius ratios an optimum value of r$_{rod}$/r$_{0}$ = 0.355 was derived [@niculae_numerical_2009]. The kinetic energies of ions are modulated much more strongly by the quadrupole RF field than higher multipoles RF field. The ion beam from the ion source has an intrinsic energy, position, and angular spread (phase space in mm.mrad or emittance), any defect in the primary beam would affect beam resolution significantly, due to compression and distortion by the multi-keV axial acceleration and focusing abberations prior to inserting into the magnet mass separator system. The transverse kinetic energy of an ion propagating inside a linear RF ion guide does not stay constant. It varies rapidly following the ion’s rapid micro-motion and the slower macro-motion in the RF field in the radial plane. For higher order multipoles, the RF field in the center is much weaker, and the micro-motion is less pronounced near the axis. For a better acceptance and resolution by the mass separator magnet (ISAC at TRIUMF) with $\delta E \sim 1$eV, more ions can be extracted from an octupole ion guide. ### Octupole versus quadrupole ion guide Simulation studies depicted in Fig. \[quadoct\](a) show that the energy spread of ions extracted from a quadrupole ion-guide configuration only allows to extract a maximum 28% of ions at the magnet mass separator with an energy acceptance of $\sim$1eV. Upgrading to an octupole ion guide, Fig. \[quadoct\](b), would increase the efficiency to 50% with the same 1eV energy spread. ![Simulation studies for kinetic energy of extracted ions in suppression mode from (a) a quadrupole ion guide and (b) an octupole ion guide. Dashed lines represent the $\delta E$ = 1eV region for extracted ions from an ion guide. Due to smaller radial RF field gradient of an octupole, more of the ions are extracted within 1eV transmission window. Simulations are performed for mass 200amu.[]{data-label="quadoct"}](quadoct){width="\linewidth"} An alternative method to suppress surface ions is to bias the source instead of applying a potential on the repeller electrode. Utilizing a source-bias setup results in a reduced ion energy spread over 1eV compared to the suppression by means of the repeller electrode for both quardupole and octupole ion guides (Fig. \[bias\]). It allows for a more efficient transmission of ions through the mass separator. ![Simulation studies for kinetic energy of extracted ions in suppression mode with source bias (the preferred IGLIS operation mode) from a quadrupole ion guide (red curve) and an octupole ion guide (blue curve). The peak in blue curve indicates a reduced energy spread of the extracted ions in an octupole ion guide. Simulations are performed for mass 200amu.[]{data-label="bias"}](bias){width="\linewidth"} Conclusion ========== Operating the IG-LIS (RFQ-offset, RF-amplitude and frequency) with proper settings is critical in order for system performance. Modifying the IG-LIS ion guide to an octupole results in a net reduction of energy spread. Low energy spread becomes important when coupling IG-LIS to the ISAC isobar separator for collinear fast beam laser spectroscopy, or when injecting beams into ion traps. Increasing laser repetition rate from 10kHz to 50kHz can enhance the IG-LIS efficiency while significantly reducing the energy spread of extracted laser ions. Acknowledgments {#acknowledgments .unnumbered} =============== The work has been funded by TRIUMF under a contribution from National Research Council of Canada (NRC) and through a Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (SAP-IN-2017-00039). M. Mostamand acknowledges funding through the University of Manitoba Graduate Fellowship.\ \ K. Blaum, C. Geppert, H. J. Kluge, M. Mukherjee, S. Schwarz and K. Wendt, A novel scheme for a highly selective laser ion source, Nucl. Instrum. Meth. B. 204, 331–335 (2003). J. P. Lavoie, Production of pure ion beams by laser ionization and a fast release [RFQ]{}, PhD thesis, University of Laval (2010). J. P. Lavoie, P. Bricault, J. Lassen and M. R. Pearson, Segmented linear radiofrequency quadrupole/laser ion source project at [TRIUMF]{}, Hyperfine Interact. 174, 33–39 (2007). H. Heggen, Development of a radio frequency quadrupole- laser ion source (RFQ - LIS) for isobar suppression, M.Sc thesis, Technische Universität Darmstadt (2013). S. Raeder, H. Heggen, J. Lassen, F. Ames, D. Bishop, P. Bricault, P. Kunz, A. Mj$\phi$s and A. Teigelhöfer, An ion guide laser ion source for isobar-suppressed rare isotope beams, Rev. Sci. Instrum. 85, 033309 (2014). D. A. Dahl, SIMION for the personal computer in reflection, Int. J. Mass. Spect. 200, (2000). C. Niculae, and M. Niculae, Numerical method for calculating of potential distribution in non-ideal multipole ion guides, Optoelectron. Adv. Mater. Rapid. Commun. 3, 1073–1075 (2009). R. Ferrer, V. T. Sonnenschein, B. Bastin, S. Franchoo and M. Huyse, Performance of a high repetition pulse rate laser system for in-gas-jet laser ionization studies with the Leuven laser ion source @ LISOL, Nucl. Instrum. Meth. B. 291, 29–37 (2012). T. Kron, R. Ferrer, N. Lecesne, V. Sonnenschein, S. Raeder, J. Rossnagel and K. Wendt, Control of RILIS lasers at IGISOL facilities using a compact atomic beam reference cell, Hyperfine Interact. 216, 53–58 (2013).

--- abstract: 'The flow of quantized vortex lines in superfluid $^3$He-B is laminar at high temperatures, but below $0.6\,T_{\rm c}$ turbulence becomes possible, owing to the rapidly decreasing mutual friction damping. In the turbulent regime a vortex evolving in applied flow may become unstable, create new vortices, and start turbulence. We monitor this single-vortex instability with NMR techniques in a rotating cylinder. Close to the onset temperature of turbulence, an oscillating component in NMR absorption has been observed, while the instability generates new vortices at a low rate $\sim 1\,$vortex/s, before turbulence sets in. By comparison to numerical calculations, we associate the oscillations with spiral vortex motion, when evolving vortices expand to rectilinear lines.' author: - 'R. Hänninen' - 'V.B. Eltsov' - 'A.P. Finne' - 'R. de Graaf' - 'J. Kopu' - 'M. Krusius' - 'R.E. Solntsev' title: 'Precessing Vortex Motion and Instability in a Rotating Column of Superfluid $^3$He-B' --- In superfluid dynamics a longstanding goal has been to account for the formation of all vortices. A quantized vortex line is a topologically stable structure of a coherent order parameter field, with fixed circulation of superflow. What is then the mechanism by which sudden burst-like turbulent vortex proliferation is started from one single vortex loop which is evolving in applied flow? The common view holds that turbulence, once started, is sustained via loop formation and reconnections, as seen in well-developed thermal counterflow turbulence of superfluid $^4$He-II [@Vinen]. However, as emphasized by Schwarz [@Schwarz], to start turbulence, for instance in linear flow along a circular pipe, vortices have to be pinned in “vortex-mill” configurations which continuously inject new vortex loops downstream in the applied flow. Nevertheless, a turbulent burst of vortex formation is observed to evolve from a single seed vortex loop in rotating $^3$He-B, with no permanently pinned vortices [@NeutronInjection]. A recent explanation [@Precursor] of this controversy concludes that a vortex evolving in applied flow may become unstable while interacting and reconnecting with the container wall. In $^3$He-B this *single-vortex instability* becomes possible with decreasing temperature below an onset temperature $T_{\rm on}$, where mutual friction damping has dropped to sufficiently low value: $\alpha (T) \lesssim 1$. In the rotating cylinder the instability leads to a sudden transition to the equilibrium vortex state. This process is most conveniently studied in the onset temperature regime $T \sim T_{\rm on}$, by monitoring the number of vortex lines as a function of time at constant rotation velocity, as a response to the injection of a seed vortex [@de; @Graaf:2007]. The events following the instability have been reviewed in Refs. [@ROP; @PLTP] for the case of a long rotating column. Here the end point of an evolving seed vortex describes a spiral trajectory along the cylinder wall while it expands towards its stable state as rectilinear vortex line. In the onset region $T \sim T_{\rm on}$, the spiral vortex motion is sometimes observed to give rise to an oscillating NMR absorption signal which is examined in this report. ![Numerical calculation of vortex evolution in a rotating cylinder of $^3$He-B: Rotation is suddenly increased at $t=0$ from $\Omega_{\rm i} \approx 0.03\,$rad/s to $\Omega_{\rm f} = 0.2\,$rad/s. There are 22 vortices in this sample, of which two in the outermost ring (lying opposite to each other) have been initially bent to the cylindrical wall, to break cylindrical symmetry. In the later snapshots at $\Omega_{\rm f}$, the two short vortices expand towards the top and bottom end plates of the cylinder, while other vortices have contracted to a central vortex cluster. Parameters: $R=3\,$mm, $L=30\,$mm, $P = 29.0\,$bar, and $T=0.4\,T_{\rm c}$ (which corresponds to $\alpha = 0.18$ and $\alpha^{\prime} = 0.16$ [@Bevan]).[]{data-label="EquilVorStateInject"}](fig1.jpg){width="1\linewidth"} [**Numerical illustrations:**]{} To visualize the motion of evolving vortices in rotating flow, a calculation is presented in Fig. \[EquilVorStateInject\] using the vortex filament method described in Ref. [@de; @Graaf:2007]. The initial configuration in Fig. \[EquilVorStateInject\], with two curved vortices which connect at one end to the cylindrical side wall, mimics the equilibrium vortex state in a real rotating experiment. Depending on the rotation velocity $\Omega_{\rm i}$ and the residual misalignment between the rotation and the sample cylinder axes ($ \lesssim 1^{\circ}$ in the setup of Fig. \[Sample\]), a certain fraction of the peripheral vortices ends on the cylindrical wall [@misalignment]. ![Comparison of azimuthal $v_{{\rm L}{\phi}}$ (top) and longitudinal $v_{{\rm L}z}$ (bottom) velocities when different numbers of $2N_{\rm b}$ vortices expand in the setup of Fig. \[EquilVorStateInject\]. With 22 vortices in total, 1, 2, 3, or 9 vortices of equal initial length $0.4\,L$ have been bent both on the top and bottom to the cylindrical wall, to mimic a tilted cylinder, as shown in Fig. \[EquilVorStateInject\]. The spiralling motion both upwards and downwards along the rotating column is calculated as a function of time. The average velocity of the vortex ends on the cylindrical wall in the upward and downward moving bundles is plotted during the time needed to reach the respective end plate. The actual velocity is the number on the vertical scale times $\alpha v(\Omega_{\rm f},R,N)$ or $(1-\alpha^{\prime})v(\Omega_{\rm f},R,N)$, where $N = 22-N_{\rm b}$. The case $N_{\rm b} = 22$ is different: here all 22 vortices are initially bent at one cylinder end to the wall and, after the rotation increase to $\Omega_{\rm f}$, a vortex front starts expanding towards the other vortex-free end of the cylinder. This situation is examined further in Fig. \[PrecessVorFront\].[]{data-label="SpiralVorVelocities"}](fig2.jpg){width="0.8\linewidth"} In Fig. \[EquilVorStateInject\], rotation is increased in step-like manner at $t=0$ from the equilibrium vortex state at $\Omega_{\rm i} \neq 0$ to a final stable value $\Omega_{\rm f} $. The $N= 20$ rectilinear vortices are thereby compressed to a central cluster with an areal density $n_{\rm v} = 2 \Omega_{\rm f} /\kappa$ by the surrounding counterflow (cf) at the velocity $\bm{v} = \bm{v}_{\rm n} - \bm{v}_{\rm s}$, the difference between the velocities of the normal and superfluid fractions. The highly viscous normal component we consider to be in solid-body rotation: $\bm{v}_{\rm n} = \bm{\Omega} \times \bm{r}$. Outside the compressed cluster the cf velocity is then given by $$v(\Omega_{\rm f},r,N) = v_{\rm n} - v_{\rm s} = \Omega_{\rm f} r - \frac{\kappa N} {2\pi r} \, , \label{cf}$$ where $\kappa = h/(2m_3)$ is the quantum of circulation. The two short vortices now expand in this cf in spiral motion towards the top and bottom end plates, respectively. The velocity ${\bf v}_{\rm L}$ of a vortex line element with tangent $\hat{\mathbf{s}}$ is obtained from the equation of motion [@Donnelly] $$\bm{v}_{\rm L}= \bm{v}_{\rm s} +\alpha \hat{\mathbf{s}} \times (\bm{v}_{\rm n}-\bm{v}_{\rm s}) -\alpha' \hat{\mathbf{s}} \times [\hat{\mathbf{s}}\times(\bm{v}_{\rm n}-\bm{v}_{\rm s} )]\, ,\label{vl}$$ where $\alpha$ is the dissipative and $\alpha^{\prime}$ the reactive mutual friction coefficient. Since the vortex end is perpendicular to the cylindrical wall, it has from Eq. (\[vl\]) a longitudinal velocity $v_{{\rm L}z} = \alpha v(\Omega_{\rm f},R,N)$ and an azimuthal component $v_{{\rm L}{\phi}} = (1-\alpha^{\prime}) v(\Omega_{\rm f},R,N)$. In the rotating coordinate system, $v_{{\rm L}{\phi}}$ points in the opposite direction from the rotation of the container. The resulting end point motion gives us a simple picture of the spiral trajectories along the cylindrical wall, although other parts of the vortex also contribute to its motion, in particular its curvature where it connects to the cylindrical wall. During their expansion the two evolving vortices are wound around the central vortex cluster as a helix whose pitch depends on the cf velocity $v(\Omega_{\rm f},R,N)$. Ultimately, the helix will unwind, when the vortex ends slip along the flat end plates of the cylinder, so that the final state is composed of only rectilinear vortex lines. However, since the cf velocity is close to zero at the border of the cluster, the unwinding is a slow process. The calculated velocities of the vortex ends in Fig. \[EquilVorStateInject\] are $v_{{\rm L}z} \approx 0.84 \, \alpha \Omega R \approx 0.96 \, \alpha \, v(\Omega_{\rm f},R,N)$ and $v_{{\rm L}{\phi}} \approx 0.73 (1-\alpha^{\prime}) \Omega R \approx 0.83(1-\alpha^{\prime}) \, v(\Omega_{\rm f},R,N)$. The wave length of the spiral trajectory is thus $\lambda = $ $2\pi R \, v_{{\rm L}z}/v_{{\rm L}{\phi}} \approx 5\,$mm and the period $p = 2\pi R/v_{{\rm L}{\phi}} \approx \,50\,$s. In Fig. \[SpiralVorVelocities\] the spiral vortex motion is analyzed further for the same experimental setup as in Fig. \[EquilVorStateInject\], but when several vortices are expanding simultaneously. The initial starting state for these calculations is one where a specified number ($N_{\rm b}$) of nearest-neighbor vortices is bent to the cylindrical wall at the same distance of $0.4\,L$ from the end plate both at the bottom and top of the cylinder, similar to the configuration in Fig. \[EquilVorStateInject\] (at $t=0$). When rotation is increased from $\Omega_{\rm i}$ to $\Omega_{\rm f}$, one set of $N_{\rm b}$ curved vortices starts expanding towards the top and one towards the bottom end plate. The axial (bottom panel) and azimuthal velocities (top panel) have been plotted as a function of time, for different numbers of $2N_{\rm b}$ vortices in spiral expansion. Two observations can be made from Fig. \[SpiralVorVelocities\]. The vortices, which expand in one direction and initially start off in close proximity of each other, remain a close bundle during their spiral motion. Vortex bundling is a general phenomenon which gives rise to the formation of larger eddies, similar to the eddies in classical viscous fluid motion. Here the bundles do not disperse, but are well preserved especially during the later part of the expansion in a state of steady propagation, since all $N_{\rm b}$ vortices travel roughly at the same speed both axially and azimuthally along the cylinder. Thus the spread in velocities among the different vortices within the bundle and also the changes in the velocities as a function of time remain relatively small. The second note about Fig. \[SpiralVorVelocities\] is that both $v_{{\rm L}z}$ and $v_{{\rm L}{\phi}}$ are rather insensitive to the number of vortices spiralling in the bundle. This is because the total number of vortices $N = 22$ corresponds here to a very low equilibrium rotation velocity $\sim 0.03\,$rad/s compared to the actual rotation velocity of 0.2rad/s. For better comparison with measurements the final rotation velocity $\Omega_{\rm f}$ should be increased even more, which calls for time-consuming calculations. ![Calculated azimuthal front velocity $V_{{\rm f}\phi}$ (in the rotating coordinate system), as a function of the number of vortices $N_{\rm f}$ which compose the front. No prior vortex cluster exists in the center of the cylinder. The zero temperature limit of Eq. \[vm\] [*(solid curve)*]{} is compared to calculations at finite temperatures [*(data points)*]{}. The calculations with $N_{\rm f} < N_{\rm eq}$ are for $R = 3\,$mm, $L = 30\,$mm, $\Omega = 0.2\,$rad/s, where $N_{\rm eq} \approx 170\,$ vortices: ($\circ$) $\alpha = 0.1$, $\alpha^{\prime} =0$; ($\times$) $\alpha = 0.18$, $\alpha^{\prime} = 0.16$, $T \approx 0.4\,T_{\rm c}$. The two calculations with $N_{\rm f} \approx N_{\rm eq}$ are for $R = 1.5\,$mm, $L = 40\,$mm, $\Omega = 1\,$rad/s, where $N_{\rm eq} \approx 210\,$ vortices: ($\bigtriangledown$) $\alpha = 0.040$, $\alpha^{\prime} = 0.030$, $T \approx 0.3\,T_{\rm c}$; ($\bigtriangleup$) $\alpha = 0.18$, $\alpha^{\prime} = 0.16$, $T \approx 0.4\,T_{\rm c}$. On the vertical scale the normalized velocity $V_{{\rm f}\phi}/[(1- \alpha^{\prime}) \Omega R]$ is plotted. $V_{{\rm f}\phi}$ is taken as the front velocity when a stable value has been reached, which generally happens when the front has propagated 2/3 of the length of the cylinder. []{data-label="PrecessVorFront"}](fig3.jpg){width="0.85\linewidth"} A special case is that where all existing vortices are released simultaneously from one end of the rotating cylinder and expand as a vortex front towards the vortex-free end of the cylinder (*i.e.* where no prior central vortex cluster exists). In Fig. \[SpiralVorVelocities\] this case is the example with $N_{\rm b} = 22$. Such a calculation models the experimental situation after a turbulent burst, like in Fig. \[FastCF\_PeakResponse\] (although there the burst occurs higher up in the column, which then starts both an upward and downward moving front). At temperatures below about $0.45\, T_{\rm c}$ the front, once it is formed from $N_{\rm f}$ vortices, maintains its constant narrow width during the spiralling propagation along the column [@PLTP; @ROP]. The longitudinal propagation velocity $V_{\rm fz}$ of the front has been measured and analyzed in Ref. \[\]. The azimuthal velocity $V_{{\rm f}{\phi}}$ (in laboratory coordinates) is obtained in the continuum model at zero temperature (where friction vanishes) from the equation $$V_{{\rm f}{\phi}}= \frac{\kappa N_{\rm f}} {2\pi R} \; \frac{ 1/2\,\ln{(N_{\rm eq} / N_{\rm f})} + 1/4} {1- N_{\rm f} /(2 N_{\rm eq})}\, , \label{vm}$$ where the approximate continuum value for the number of vortices in the equilibrium state is $N_{\rm eq} = \pi R^2 (2\Omega/\kappa)$. When $N_{\rm f} \rightarrow N_{\rm eq}$, this expression gives $V_{{\rm f}{\phi}}= {\frac{1}{2}}\,\Omega R$. This limit was discussed in Ref. [@TwistedPropagation]. In Fig. \[PrecessVorFront\] we compare Eq. (\[vm\]) to numerical calculations at finite temperatures and finite friction for a few examples. The results for the normalized azimuthal front velocity $V_{{\rm f}\phi}/[(1- \alpha^{\prime}) \Omega R]$ drop below the estimate from Eq. (\[vm\]), but by less than 10%. Thus also at finite friction $V_{{\rm f}{\phi}} \approx \frac{1}{2} (1-\alpha^{\prime}) \Omega R$ is a good approximation, which applies when the number of vortices in the front is large ($N \lesssim N_{\rm eq}$). ![[*(Left)*]{} Rotating vortex cluster from Fig. \[EquilVorStateInject\] at $t=60\,$s, viewed at an angle deviating by $3^{\circ}$ from vertical. Only the central part of the sample cross section in the $x$ – $y$ plane is shown. Helical Kelvin waves are induced on the vortex lines in the cluster by the two vortices spiralling outside the cluster. In the early stage of the expansion, the waves start at roughly half-height of the cylinder, *i.e.* from the points where the two short vortices merge with the cluster (see Fig. \[EquilVorStateInject\]). [*(Right)*]{} Side view through the vortex cluster ($t=60\,$s). A vortex spiralling around the cluster induces radial displacements on the outermost vortices within the cluster. The displacements start longitudinally propagating Kelvin waves on these vortices. []{data-label="PrecessVorMotion"}](fig4.jpg){width="1\linewidth"} While spiralling around the cluster in Fig. \[EquilVorStateInject\], the two curved vortices may reconnect with each other or with the outermost lines in the cluster. These reconnections do not increase the vortex number. As can be seen from Figs. \[EquilVorStateInject\] and \[PrecessVorMotion\], a spiralling vortex distorts the cluster by inducing propagating helical Kelvin waves on the vortex lines, which results in oscillations of the cluster around its equilibrium position [@movies]. The amplitude of the waves is comparable to the inter-vortex distance in the cluster, $\sim 1/\sqrt{n_{\rm v}}$, and initially the wavelength and period of the excited Kelvin waves have similar values as the precessing vortex motion outside the cluster. At high friction ($\alpha \geq 0.18$) the number of vortices remains strictly constant in Figs. \[EquilVorStateInject\] – \[PrecessVorMotion\]. At low friction ($\alpha \leq 0.1$) the calculations on the precessing vortex front in Fig. \[PrecessVorFront\] may display some increase in vortex number, when $N < N_{\rm eq}$. Figs. \[EquilVorStateInject\] – \[SpiralVorVelocities\] thus illustrate the situation above $T_{\rm on}$ where no increase in vortex number occurs. Experimentally $T_{\rm on}$ depends on the applied flow velocity ($\sim [\Omega_{\rm f} - \Omega_{\rm i}] R$) and on the number ($N_{\rm f}$) and initial configuration of evolving vortices [@Remnants; @de; @Graaf:2007; @PLTP]. In most cases $T_{\rm on}$ is found to be higher than the temperature of the present calculations, $0.40\, T_{\rm c}$. The reason for this difference appears to be the high stability of rotating flow in a circular cylinder, which according to our calculations [@de; @Graaf:2007] appears to be a particular characteristic of a cylinder with a circular cross section and ideal wall properties, *i.e.* without surface friction or pinning. The most likely section of an evolving vortex to become unstable is the curved piece ending on the cylindrical wall in the maximum flow $v(\Omega_{\rm f},R,N)$, while the most stable parts are the sections which reside in low cf close to the central vortex cluster. The extreme case are rectilinear vortex lines in the cluster which are experimentally found to remain stable even in sinusoidally modulated rotation at temperatures $\gtrsim 0.3\,T_{\rm c}$. ![Sample cylinder with detector coils. []{data-label="Sample"}](fig5.jpg){width="0.45\linewidth"} [**Experiment:**]{} The sample is cooled in a rotating nuclear demagnetization cryostat at 29bar liquid pressure within a fuzed quartz cylinder, with radius $R=3\,$mm and length $L=110\,$mm (Fig. \[Sample\]). The number of vortices is measured non-invasively simultaneously at both ends of the long cylinder, by monitoring the NMR absorption line shape. Both NMR detectors consist of circular split-pair coils, with a separation of 9mm between the coil halves and with their common axis aligned transverse to the cylinder. The coils are wound from superconducting wire and are part of a LC resonator with a Q value of $\sim 1\cdot 10^4$, coupled to a GaAs MESFET preamplifier located within the cryostat at 4.2K temperature. Both NMR spectrometers operate independently at constant frequency in an axially oriented polarization magnetic field with a linear sweep. The NMR setup is described in more detail in Refs. [@NeutronInjection; @ExpSetup] With vortex-free cf, the NMR absorption line shape includes a prominent cf peak which is shifted far from the Larmor frequency, while in the equilibrium vortex state the absorption drops to zero at the location of the cf peak [@Kopu]. The height of the cf peak increases with the magnitude of the azimuthal cf velocity $v(\Omega,R,N)$ and thus decreases with the number of vortices $N$ in the central cluster. The correspondence between peak height and vortex number $N$ can be obtained from calculations of the order parameter texture or from measurements. The measurements which we discuss below use remanent vortices [@Remnants] as the seeds which evolve in the applied flow after the rotation increase to $\Omega_{\rm f}$. Such a measurement is performed at constant temperature, by first decelerating rotation to zero from a state with a large number of vortices. Zero rotation is then maintained for a period $\Delta t$, to allow vortices to annihilate, with the exception of a few remaining dynamic remnants [@Remnants]. The final step is to increase $\Omega$ at a fixed rate (typically $d \Omega/dt \leq 0.02\,$rad/s$^2$) to $\Omega_{\rm f}$, where it is kept constant while the buildup in the number of vortex lines $N(t)$ in the central cluster is recorded as a function of time. [**Single vortex instability:**]{} Well above the onset temperature, the cf peak height settles at $\Omega_{\rm f}$ to a stable value and remains constant. Well below onset the cf peak collapses to zero when the equilibrium vortex state is formed. Around the onset temperature $T \approx T_{\rm on}$, both types of behavior occur randomly. If the equilibrium vortex state is here formed, then in perhaps one third of the cases the collapse of the cf peak can be preceded by an initial slow decrease in peak height. The slow reduction in peak height corresponds to slow generation of new vortices. It is the signature of the single-vortex instability, as described and analyzed in Ref. [@de; @Graaf:2007]. It can only be observed in the onset regime $T \sim T_{\rm on}$; at lower temperatures the instability proceeds too rapidly to be monitored with our techniques. The slow vortex generation is terminated in a turbulent burst which takes place as a localized event in some short section (of length $\sim R$) of the long sample column. In this burst enough vortices are created to start the evolution towards the equilibrium vortex state. From the site of the burst the vortices propagate in spiral motion as a front both up and down along the cylinder, leaving behind a vortex bundle composed of helically twisted vortices [@ROP; @PLTP]. ![NMR response during a rapid transition from vortex-free flow to the equilibrium vortex state at constant externally controlled conditions. The top and bottom detectors monitor the NMR absorption at different values of constant NMR field. The cf peak height is recorded with the top detector, while the bottom detector monitors the peak close to the Larmor edge. Time $t=0$ marks the moment when $\Omega_{\rm f} = 1.61\,$rad/s is reached. The time delay between the sudden rise in the bottom response (at $ t = 40\,$s) and the start of the rapid collapse in the top (at $ t = 78\,$s) is caused by the longitudinal motion of the two vortex fronts along the column, controlled by $\alpha = 0.33$ [@Bevan]. []{data-label="FastCF_PeakResponse"}](fig6.jpg){width="1\linewidth"} The NMR absorption responses of rapid and slow transitions to the equilibrium vortex state are compared in Figs. \[FastCF\_PeakResponse\] and \[OscCF\_Peak\], respectively. These were recorded in two consecutive measuring runs at the same temperature using the same measuring procedure. The order of the two measurements was such that Fig. \[OscCF\_Peak\] was measured first and Fig. \[FastCF\_PeakResponse\] next. The responses prove to be different owing to the different value of the rotation velocity $\Omega_{\rm f}$, which in turn is caused by the stochastic nature of the remanent vortex injection process: in the former case the waiting period at zero rotation prior to the measurements was $\Delta t = 40\,$min, while in the latter case it was 22min. Thus in Fig. \[FastCF\_PeakResponse\] the remanent vortices are expected to be fewer and smaller loops. In both cases the sample was then accelerated to rotation at $\dot{\Omega} = 0.004\,$rad/s$^2$. The acceleration was interrupted occasionally, to check at constant rotation whether vortex-free rotation continued to persist over long periods in time. In Fig. \[FastCF\_PeakResponse\] (Fig. \[OscCF\_Peak\]) rotation could be increased until 1.61rad/s (0.90rad/s), before any indication of vortex formation was noticed. Time $t=0$ coincides with the moment when $\Omega_{\rm f} = 1.61\,$rad/s (0.90rad/s) was reached. After 40s (170s) the first response from vortices is observed in the bottom (top) spectrometer. This should be compared with the time it takes for a small vortex loop to expand in vortex-free flow at $\Omega_{\rm f}$ from one end of the cylinder to the other, to become a rectilinear line, $\approx L/(\alpha \Omega_{\rm f} R) = 70\,$s (120s). In Fig. \[OscCF\_Peak\] the start of vortex formation is thus unusual because of the long 170s delay, before the generation of new vortices starts. The explanation for the long delay must be associated with the properties of a small remanent vortex loop with a radius close to the critical value of $\sim 4\,\mu$m needed to overcome the barrier for spontaneous expansion [@Ruutu]. [**Counterflow peak response :**]{} In Figs. \[FastCF\_PeakResponse\] and \[OscCF\_Peak\] the peak heights of two different absorption maxima of the NMR spectrum are recorded as a function of time at constant externally controlled conditions. These are: (1) The cf peak height, which initially in vortex-free flow is at maximum and zero in the final equilibrium vortex state. (2) The Larmor peak height, which initially is close to zero, but in the final state different from zero. While the Larmor peak response is qualitatively similar in the two examples, the drop in the cf peak height happens differently. Fig. \[FastCF\_PeakResponse\] shows the generic response from a rapid transition, while in Fig. \[OscCF\_Peak\] the decay in the cf peak height starts first with a more gentle linear decrease (dashed line) where the number of vortices gradually increases within the top detector, as generated by the single-vortex instability. This continues for more than 100s, before the rapid peak height decay starts. (Also in Fig. \[FastCF\_PeakResponse\] at higher $\Omega_{\rm f}$ a short vestige of slower peak height reduction can be distinguished, but the rate is faster and it only lasts for $\sim 4\,$s.) However most importantly, in Fig. \[OscCF\_Peak\] the slow decrease in peak height differs from usual examples in that the cf peak height also displays well resolved oscillations of large amplitude. The characterization of these oscillations is the central issue in this report. ![Precursory vortex formation which terminates in a sudden turbulent burst. This is a repetition of the measurement in Fig. \[FastCF\_PeakResponse\], but at a lower rotation velocity $\Omega_{\rm f} = 0.90\,$rad/s. The difference is the prolonged slow vortex generation via the single-vortex instability, denoted by the dashed line. Superimposed comes the large-amplitude quasi-coherent oscillation in the cf peak height at a frequency which is related to the azimuthal vortex motion. The latter is controlled by $1-\alpha^{\prime}= 0.76$ [@Bevan]. []{data-label="OscCF_Peak"}](fig7.jpg){width="1\linewidth"} According to the measured calibration, the linear decrease in the cf peak height, as marked by the dashed line in Fig. \[OscCF\_Peak\], represents a vortex formation rate $\dot{N} \approx 1.2\,$vortices/s. The later rapid collapse is caused by the arrival of the vortex front [@FlightTime] to the bottom end of the top coil and its subsequent travel through the coil. At the arrival of the front, the central cluster within the top coil contains $\sim 140$ vortices (to be compared with $N_{\rm eq} \approx 840$ in the equilibrium vortex state). The front incorporates approximately all the remaining vortices needed to fill the sample with the equilibrium vortex state. The front, composed of these additional vortices, travels in spiral motion along the cylinder around the central cluster. Behind the front the expanding vortices are wound in a helically twisted configuration which later slowly relaxes [@TwistedPropagation]. At that point the macroscopic cf, which was generated by the rotation increase to $\Omega_{\rm f}$, is completely removed. [**Larmor peak response:**]{} The second signal trace in Figs. \[FastCF\_PeakResponse\] and \[OscCF\_Peak\] is recorded by scanning the absorption maximum close to the Larmor edge with a rapidly moving field sweep. Here the sudden sharp absorption increase signals the arrival of the vortex front, as it passes through the top edge of the bottom detector. The increasing absorption is generated by the axially flowing supercurrent which is produced by the helically twisted vortices behind the front [@Kopu-2]. The subsequent exponential decay of this absorption is caused by the unwinding of the twist while the vortex ends slip along the top and bottom end plates of the sample cylinder and the vortices are converted to rectilinear lines [@TwistedPropagation]. Thus the Larmor absorption response forms a transient peak, followed by a stable final absorption level, which is representative of the equilibrium vortex state, after the twist has relaxed [@ROP]. In Fig. \[OscCF\_Peak\] a delay time $t_{\rm d} \approx 144\,$s separates the moments when the vortex fronts moving up and down along the column pass through the bottom edge of the top detector and the top edge of the bottom detector, respectively. This distance is $82\,$mm. The longitudinal propagation velocity $V_{{\rm f}z}$ of the front has been measured to be $V_{{\rm f}z} /(\Omega R) \approx 0.28$ at $0.45\,T_{\rm c}$ [@FlightTime], in a situation when there is no prior central vortex cluster. Assuming that the propagation velocity of the front in the presence of a central vortex cluster is $\approx v(\Omega_{\rm f}, R, N) \; V_{{\rm f}z} /(\Omega R)$, we estimate that, with a vortex formation rate of $\dot{N} \approx 1.2\,$vortices/s, on an average $N \approx 230\,$ vortices populate the central cluster at the time when the front travels from the location of the turbulent burst to the top edge of the bottom detector. The delay time of 144s can then be used to localize the site and the moment of the turbulent burst. Here the burst proves to happen a couple of mm below the top detector at $t \approx 290\,$s, or some five seconds before the absorption response in the top detector starts to collapse. In Fig. \[FastCF\_PeakResponse\] at a higher flow velocity all events occur faster: with a delay of 33s between the first bottom and top responses, the burst happens at $t = 28\,$s some 19mm above the top edge of the bottom coil. ![Time difference between the maxima and minima in the oscillating cf peak height in Fig. \[OscCF\_Peak\]. The line is obtained by inserting the appropriate values from Fig. \[OscCF\_Peak\] in Eq. (\[PrecessionFrequency\]). []{data-label="OscPeriod"}](fig8.jpg){width="0.8\linewidth"} [**Spiral vortex motion:**]{} The generic signal from slow vortex generation in the cf peak height is a relatively smooth decay, while the many periods of well-resolved quasi-coherent oscillation in Fig. \[OscCF\_Peak\] are unusual. If both the top and bottom detectors are tuned to monitor the cf peak height simultaneously, then the two recordings of the oscillations look similar, but neither the amplitudes or the frequencies are exactly identical. Since the distribution and number of evolving vortices varies along the column, this is expected. Obviously the oscillating signal is driven by the spiral vortex motion, which is the only source for precession at a frequency around 0.1Hz. In Fig. \[OscPeriod\] we have extracted the difference in time between the maxima and minima in the oscillation of Fig. \[OscCF\_Peak\]. This data set consists of 13 full periods. The average period is 9.2s, but the period also appears to be increasing with time. The spiral vortex motion is expected to have the frequency $$f_{\phi} = \varepsilon \, (1-\alpha^{\prime}) v(\Omega_{\rm f}, R,N) / (2\pi R)\;,\label{PrecessionFrequency}$$ so that $\dot{f}_{\phi} \propto -\dot{N}$. We assume that the azimuthal vortex velocity is of the form $v_{{\rm L}{\phi}} = \varepsilon (1-\alpha^{\prime}) v(\Omega_{\rm f}, R, N)$, where $\varepsilon < 1$ is a numerical factor, which in the case of Fig. \[SpiralVorVelocities\] is about 0.8. The measuring setup in Fig. \[Sample\] incorporates also an instrument factor: The sensitivity of the detector coil pair is not constant across the cross section of the sample, but increases towards the windings. The cf peak height is regulated by the azimuthal vortex-free flow, *i.e.* the height depends on the velocity $v(\Omega_{\rm f}, R,N)$ at the cylindrical wall. This doubles the frequency of an azimuthally precessing asymmetry in the cf velocity $v(\Omega_{\rm f}, R,N)$. For simplicity we assume that $ \varepsilon = {1 \over 2}$ which then compensates for the frequency doubling owing to the sensitivity pattern. If we take $\dot{N} \approx 1.2\,$vortices/s and $N = 0$ at the moment when vortex formation starts (at $t = 170\,$s) in Fig. \[OscCF\_Peak\], we obtain the line through the data in Fig. \[OscPeriod\]. The best fit would have a twice larger slope, but in view of the large scatter in the signal, which is only quasi-coherent, the agreement with the calculated line appears acceptable. Using this approach, we have extracted $1 - \alpha^{\prime}$ from seven runs with slow oscillatory cf peak decay. In these seven cases the single-vortex instability proceeds at a rate $\dot{N} \sim 0.4$ — 2vortices/s, while $\Omega_{\rm f}$ is in the range 0.8 — 1.2rad/s. As seen in Fig. \[AlphaPrime\], the results agree with the measurements on $1 - \alpha^{\prime} (T)$ in Ref. [@Bevan]. ![Consistency test on the analysis of the cf peak height oscillations from seven different measuring runs. The reactive mutual friction coefficient $1-\alpha^{\prime}$ has been extracted using Eq. (\[PrecessionFrequency\]) and is compared to the measurements in Ref. [@Bevan]. The line is a fit from Ref. [@Bevan].[]{data-label="AlphaPrime"}](fig9.jpg){width="0.9\linewidth"} **Discussion:** The oscillatory response in the cf peak height in Fig. \[OscCF\_Peak\] is a rare event where several preconditions seem to be fulfilled:\ 1) New vortices have to be created at a slow rate $\dot{N} \sim 1$, which requires that $ T \approx T_{\rm on}$.\ 2) In all examples the longitudinal vortical transit time through the 9mm long detector coil pair is roughly $10\,$s and equal to the oscillation period in the cf peak height.\ 3) The oscillations start immediately when the slow vortex generation switches on. In Figs. \[FastCF\_PeakResponse\] and \[OscCF\_Peak\] the resolution in the measurement of $N$ is roughly one vortex, *i.e.* the reduction in cf peak height per one vortex is approximately equal to the amplifier noise (when $N \ll N_{\rm eq}$). This signal/noise resolution applies to the reduction in cf peak height, when one new vortex is placed in the center of the cylinder. Presumably the signal from a single vortex spiralling along the cylinder wall is larger.\ 4) To explain the quasi-coherent oscillation, the spiral motion of several vortices has to be reasonably coherent: the 13 oscillations in Fig. \[OscCF\_Peak\] cannot be attributed to a single vortex since the transit time for a vortex end to pass through the detector coil pair is equal to one period. Three features may help to create and preserve coherence in spiral motion: (i) When a vortex end on the cylindrical wall creates a new loop, one end of the new loop starts spiralling close to the original end. (ii) Both detectors are close to one of the end plates of the cylinder and, assuming that the single-vortex instability occurs randomly everywhere on the cylindrical wall [@de; @Graaf:2007], almost all vortices approach the detector from the direction of the far end of the cylinder. (iii) In numerical calculations the evolving vortices, which start to spiral as a close bundle, tend to remain in a bunch while propagating along the cylinder. A complete explanation has not been worked out for the mechanisms which gives rise to the oscillating signal in the cf peak height. To understand the coherence over many periods of spiral motion we need better agreement between vortex-dynamics calculations and the experiment, in particular concerning the generation of new vortices by the single-vortex instability in a rotating circular cylinder [@de; @Graaf:2007]. As to the amplitude of the oscillations, spiral vortex motion gives rise to oscillatory displacements of the order parameter texture from cylindrical symmetry, as seen in Fig. \[PrecessVorMotion\]. To explain the signal from such oscillations, numerical calculations are needed on distortions of flare-out textures from cylindrical symmetry and on the resulting NMR line shapes. **Conclusions:** We have studied the single-vortex instability of evolving vortices in the turbulent temperature regime of $^3$He-B. In applied flow the single-vortex instability generates new vortices and becomes the precursor mechanism which starts turbulence. In a rotating circular cylinder the instability occurs while remanent vortices, for instance, expand in spiral motion towards their stable state as rectilinear vortex lines. The spiral motion has been examined here with numerical calculations, but we also argued that oscillations in the NMR response bear direct evidence for the precessing motion. Other kinds of oscillating responses from precessing vortex motion have been reported before. A remarkable example is the unwinding of trapped circulation around a thin wire suspended along the axis of a cylindrical container, known as the Vinen vibrating wire experiment [@Zieve]. The spiral motion of an evolving vortex around a central cluster of vortex lines provides another example of precessing vortex signals. [**Acknowledgements:**]{} This work was supported by the Academy of Finland (grants 213496, 124616, 114887), by ULTI research visits (EU Transnational Access Programme FP6, contract RITA-CT-2003-505313), and the ESF research program COSLAB. [9]{} W.F. Vinen, J. Low Temp. Phys. [**145**]{}, 7 (2007). K.W. Schwarz, Physica B [**197**]{}, 324 (1994). A.P. Finne, S. Boldarev, V.B. Eltsov, and M. Krusius, J. Low Temp. Phys. [**135**]{}, 479 (2004). A.P. Finne, V.B. Eltsov, R. Hänninen, J. Kopu, M. Krusius, E.V. Thuneberg, and M. Tsubota, Phys. Rev. Lett. [**96**]{}, 85301 (2006). R. de Graaf, R. Hänninen, T.V. Chagovets, V.B. Eltsov, M. Krusius, and R.E. Solntsev, J. Low Temp. Phys. December (2008); preprint arXiv:0708.3003v2. A.P. Finne, V.B. Eltsov, R. Hänninen, N.B. Kopnin, J. Kopu, M. Krusius, M. Tsubota, and G.E. Volovik, Rep. Prog. Phys. [**69**]{}, 3157 (2006). V.B. Eltsov, R. de Graaf, R. Hänninen, M. Krusius, R.E. Solntsev, V.S. L’vov, A.I. Golov, P.M. Walmsley, Prog. Low Temp. Phys. Vol XVI, ed. M. Tsubota (Elsevier B.V., Amsterdam, December 2008); preprint – arXiv:0803.3225v2. T.D.C. Bevan, A.J. Manninen, J.B. Cook, A.J. Armstrong, J.R. Hook, and H.E. Hall, J. Low Temp. Phys. [**109**]{}, 423 (1997); Phys. Rev. Lett. [**74**]{}, 750 (1995). V.M. Ruutu, J.J. Ruohio, M. Krusius, B. Plaçais, and E.B. Sonin, Physica B [**255**]{}, 27 (1998). R.J. Donnelly, [*Quantized Vortices in Helium II*]{} (Cambridge Univ. Press, Cambridge, UK, 1991). V.B. Eltsov, A.I. Golov, R. de Graaf, R. Hänninen, M. Krusius, V. L’vov, and R.E. Solntsev, Phys. Rev. Lett. **99**, 265301 (2007); A.P. Finne, V.B. Eltsov, R. Blaauwgeers, Z. Janu, M. Krusius, and L. Skrbek, J. Low Temp. Phys. [**134**]{}, 375 (2004). V.B. Eltsov, A.P. Finne, R. Hänninen, J. Kopu, M. Krusius, M. Tsubota, and E.V. Thuneberg, Phys. Rev. Lett. [**96**]{}, 215302 (2006); J. Low Temp. Phys. **150**, 373 (2008). See movies at http://ltl.tkk.fi/research/theory/twist.html R.E. Solntsev, R. de Graaf, V.B. Eltsov, R. Hänninen, and M. Krusius, J. Low Temp. Phys. [**148**]{}, 311 (2007). A.P. Finne, S. Boldarev, V.B. Eltsov, and M. Krusius, J. Low Temp. Phys. [**136**]{}, 249 (2004). J. Kopu, R. Schanen, R. Blaauwgeers, V.B. Eltsov, M. Krusius, J.J. Ruohio, and E.V. Thuneberg, J. Low Temp. Phys. [**120**]{}, 213 (2000). V.M.H. Ruutu, Ü. Parts, J.H. Koivuniemi, N.B. Kopnin, and M. Krusius, J. Low Temp. Phys. [**107**]{}, 93 (1997); Europhys. Lett. [**31**]{}, 449 (1995). J. Kopu, J. Low Temp. Phys. **146**, 47 (2007). R.J. Zieve, Yu.M. Mukharky, J.D. Close, J.C. Davis, and R.E. Packard, J. Low Temp. Phys. [**91**]{}, 315 (1993); *ibid.* [**90**]{}, 243 (1993). [**Keywords:**]{} quantized vortex, vortex formation, vortex dynamics, vortex instability, mutual friction, transition to turbulence, onset temperature of turbulence, precursor of turbulence\

--- author: - 'M. Scodeggio' - 'D. Vergani' - 'O. Cucciati' - 'A. Iovino' - 'P. Franzetti' - 'B. Garilli' - 'F. Lamareille' - 'M. Bolzonella' - 'L. Pozzetti' - 'U. Abbas' - 'C. Marinoni' - 'T. Contini' - 'D. Bottini' - 'V. Le Brun' - 'O. Le Fèvre' - 'D. Maccagni' - 'R. Scaramella' - 'L. Tresse' - 'G. Vettolani' - 'A. Zanichelli' - 'C. Adami' - 'S. Arnouts' - 'S. Bardelli' - 'A. Cappi' - 'S. Charlot' - 'P. Ciliegi' - 'S. Foucaud' - 'I. Gavignaud' - 'L. Guzzo' - 'O. Ilbert' - 'H.J. McCracken' - 'B. Marano' - 'A. Mazure' - 'B. Meneux' - 'R. Merighi' - 'S. Paltani' - 'R. Pellò' - 'A. Pollo' - 'M. Radovich' - 'G. Zamorani' - 'E. Zucca' - 'M. Bondi' - 'A. Bongiorno' - 'J. Brinchmann' - 'S. de la Torre' - 'L. de Ravel' - 'L. Gregorini' - 'P. Memeo' - 'E. Perez-Montero' - 'Y. Mellier' - 'S. Temporin' - 'C.J. Walcher' bibliography: - 'marcos.bib' subtitle: 'Stellar mass segregation and large-scale galaxy environment in the redshift range $0.2<z<1.4$[^1]' title: 'The Vimos VLT Deep Survey:' --- [Hierarchical models of galaxy formation predict that the properties of a dark matter halo depend on the large-scale environment surrounding the halo. As a result of this correlation, we expect massive haloes to be present in larger number in overdense regions than in underdense ones. Given that a correlation exists between a galaxy stellar mass and the hosting dark matter halo mass, the segregation in dark matter halo mass should then result in a segregation in the distribution of stellar mass in the galaxy population.]{} [In this work we study the distribution of galaxy stellar mass and rest-frame optical color as a function of the large-scale galaxy distribution using the VLT VIMOS Deep Survey sample, in order to verify the presence of segregation in the properties of the galaxy population.]{} [We use the VVDS redshift measurements and multi-band photometric data to derive estimates of the stellar mass, rest-frame optical color, and of the large-scale galaxy density, on a scale of approximately 8 Mpc, for a sample of 5619 galaxies in the redshift range $0.2<z<1.4$]{} [We observe a significant mass and optical color segregation over the whole redshift interval covered by our sample, such that the median value of the mass distribution is larger and the rest-frame optical color is redder in regions of high galaxy density. The amplitude of the mass segregation changes little with redshift, at least in the high stellar mass regime that we can uniformely sample over the $0.2<z<1.4$ redshift interval. The color segregation, instead, decreases significantly for z$>$0.7. However, when we consider only galaxies in narrow bins of stellar mass, in order to exclude the effects of the stellar mass segregation on the galaxy properties, we do not observe any more any significant color segregation. ]{} Introduction ============ It is well known that galaxy properties depend on the environment the galaxies are part of. The two best and longest known examples of such an environmental dependence are the galaxy morphology-density relation [@Dressler_80], that describes the increasing fraction of early-type galaxies in the galaxy population with the increase of the local galaxy density, and the galaxy HI content deficiency [@HIdef_9clust] which is observed in cluster late-type galaxies. More recently, with the advent of large galaxy surveys comprising samples of tens of thousands of objects, it has become possible to study even subtler environmental effects on galaxy properties. Galaxy color and star formation history appear to be the two properties which are most strongly correlated with the galaxy local environment [@Blanton_05; @Ball_08], although an accurate determination of what is the typical scale-length over which these effects are generated is still missing. Only over the last few years some agreement is emerging that such a scale-length must be comparable to that of clusters of galaxies, i.e. of the order of 1 to 2 Mpc [@Kauffmann_04_env hereafter Ka04; @Blanton_07]. On the other hand it is also well known that physical properties of galaxies are mostly inter-related [see @Roberts_Haynes_araa for a review], and that the galaxy total stellar mass plays a significant role in determining these properties [@Scodeggio_02_cube; @Kauffmann_03_stellarMass]. The best known example of such a role is certainly the color-magnitude or color-stellar mass relations observed for both early and late-type galaxies, but it is by now equally clear that stellar mass plays an important role in shaping the star formation history of a galaxy [see for example @Gavazzi_Scodeggio_96; @Kauffmann_03_stellarMass; @Heavens_04]. This complex set of inter-relations among environment, galaxy properties, and galaxy stellar mass is certainly part of the reason why the decades old argument about which agent, between “nature” and “nurture”, is the primary driver for galaxy differential evolution, is still far from being settled. A relatively recent addition to the debate on this subject is a scenario where galaxy properties depend exclusively on the mass and formation history of the dark matter halo the galaxies are formed in, but they appear to be correlated with the large scale environment properties purely because of a correlation between halo properties and the large scale environment which surrounds the halo (such a correlation is predicted by hierarchical models, see for example @Mo_White_96 [@Sheth_Tormen_02]). This possibility, discussed explicitely by @Abbas_Sheth_05 [@Abbas_Sheth_06], is also the basic assumption behind the halo-model descriptions of galaxy clustering that have been used quite successfully in the recent past. In particular, hierarchical models predict the ratio of massive to low mass dark matter haloes to be larger in dense environments than it is in under-dense ones [@Mo_White_96], and simulations show that this is indeed the case [@Sheth_Tormen_02; @Abbas_Sheth_05], even when the large scale environment is defined using the standard scale length of 8 Mpc (significantly larger than the 1–2 Mpc scale typical of galaxy clusters, which is the scale length over which true environmental effects are expected to be active). Since a good correlation exists between galaxy stellar mass and dark matter halo mass, spanning a large range of stellar masses and galaxy types [e.g. @Mandelbaum_06; @Yang_07], we can directly translate the dark matter halo mass segregation prediction of hierarchical models into a stellar mass segregation prediction. Observationally, such a segregation has been demonstrated to exist with reasonable certainty only in the local Universe, using SDSS data, by . The magnitude of this effect is rather small, with the median stellar mass of galaxies in the densest environments probed by the SDSS data being only twice as large as the one of galaxies in the most under-dense environments (see Ka04 for the details). At higher redshift the effect has never been thoroughly analyzed, except for a brief mentioning of a qualitatively similar result by @Bundy_06, while discussing the stellar mass function for the galaxies in the DEEP2 sample. Similar conclusions are being obtained by Bolzonella et al. (in preparation), while studying the stellar mass function for the galaxies in the zCOSMOS sample [@zCosmos]. In this paper we use the data from the VIMOS-VLT Deep Survey [@VVDS] to extend the result, discussing the observational evidence for the presence of a stellar mass segregation over the whole redshift interval from $z\sim0.2$ (approximately the upper limit of the SDSS sample) up to $z\sim1.4$ (above this redshift our sample becomes too sparse, and our stellar mass estimates quite uncertain), and for a possible evolution in its strength. We also re-analyze the presence of color segregation in this redshift interval (a topic extensively discussed in @Cucciati_06, hereafter Cu06), and discuss the correlation between mass and color segregation. Sect. \[sec:data\] of the paper briefly summarizes the data used in this work, while Sects. \[sec:mass\_segregation\] and \[sec:color\_segregation\] discuss the evidence for stellar mass segregation and the connection between color and stellar mass segregation, respectively. The data {#sec:data} ======== In this work we use the observations of the VIMOS-VLT Deep Survey [VVDS, @VVDS] to derive stellar masses and local galaxy densities for a large sample of galaxies over an extended redshift interval. The VVDS Deep is a purely magnitude limited redshift survey targeting a random subset of a higly complete sample of galaxies in the magnitude range $17.5 < I_{AB} < 24.0$ (see @McCracken_F02photom for details on the photometric parent sample) in the VVDS 0226-04 field (hereafter VVDS-F02). Spectroscopic observations with the VIMOS multi-object spectrograph were carried out using 1 arcsec wide slits and the LRred grism, covering the spectral range from 5500 to 9400 Å, with an effective resolution of R $\sim$ 230 at 7500 Å. All data have been reduced using the VIMOS Interactive Pipeline and Graphical Interface [VIPGI, @vipgi_mos; @vipgi_ifu]. Highly reliable redshift measurements were obtained for 7528 objects, which corresponds to a sampling rate of approximately 23% of the complete parent photometric sample. In this work we use only galaxies with redshift within the $0.2 < z < 1.4$ interval, for a total sample of 5884 objects. Within this redshift range the spectral coverage of our sample is basically uniform for galaxies of all spectral types. A comparison between the relative aboundance of different spectral types in the spectroscopic and in the parent photometric sample shows that any bias against early-type galaxies in the spectroscopic sample (due to the lack of emission lines, which in turn makes the redshift estimate more difficult to obtain) is limited below the five percent level over the whole redshift range (see the discussion in @Franzetti_07). Stellar mass estimates were obtained for all these objects using the GOSSIP spectral energy distribution modeling software [@GOSSIP], taking advantage of the multi-band photometric observations available in the VVDS-F02 field, including BVRI data from the CFHT [@McCracken_F02photom], U-band data from the ESO-MPI 2.2m telescope [@Radovich_F02photom], ubvrz data from the CFHT Legacy Survey (McCracken et al. 2007), J and Ks-band data from SOFI at the NTT [@Iovino_F02photom; @Temporin_F02infrared] and from the UKIDSS survey [@Lawrence_07], 3.6 micron data from the Spitzer-IRAC SWIRE survey [@SWIRE]. The photometric and spectroscopic data were fitted with a grid of stellar population models, generated using the PEGASE2 population synthesis code [@PEGASE], assuming a set of “delayed” star formation histories (see @Virgo_spec for details), and a @Salpeter_imf initial mass function. Further details on the derivation of the stellar masses are presented in @Vergani_08 [see also @Pozzetti_VVDS_MF]. redshift bin median z Log($M_{limit}/M_\odot$) N($>M_{limit}$) -------------- ---------- -------------------------- ----------------- 0.2–0.5 0.40 9.0 722 0.5–0.7 0.61 9.5 878 0.7–0.9 0.82 9.8 844 0.9–1.4 1.12 10.3 692 : Sample General Properties[]{data-label="tab:summary"} Local galaxy densities were obtained computing the three-dimensional number density contrast for galaxies above a certain luminosity threshold in the spectroscopic sample, within a fixed comoving volume. The point-like galaxy distribution was smoothed with a Gaussian filter with a sigma of 5 Mpc, roughly equivalent in volume to a top-hat spherical filter with a sphere radius of 8 Mpc. Further details on the estimation of the local galaxy density are given in . The choice of this smoothing length has been dictated by the sample properties, mostly by the fact that the mean inter-galaxy separation typical of the VVDS Deep sample is of approximately 4.5 Mpc at the peak of the sample redshift distribution, and even larger at the low and high redshift ends of the distribution. The relatively large value for the smoothing length has also the advantage of mitigating the effects of redshift-space distorsions created by galaxy peculiar motions in overdense regions on the density estimates . To avoid using very uncertain density contrast estimates we have removed from the sample objects that are located at the edges of the volume sampled by the VVDS Deep, for which only one third or less of the comoving volume used to sample the galaxy density contrast is effectively inside the survey volume. Therefore our final sample is composed of 5619 galaxies. Stellar mass segregation {#sec:mass_segregation} ======================== If a correlation exists between stellar mass and dark matter halo mass on one side, and the large scale environment properties on the other side, we can expect the strength of the observed stellar mass segregation to depend on the range of large-scale galaxy densities being considered. Indeed the results presented by for the SDSS sample show exactly a progressive increase in the median stellar mass for samples of galaxies that range from isolated objects to objects in high density environments (see their Fig. 3). The global change in median mass is small, only a factor of 2 when going from the lowermost to the highest density sample, but nonetheless the large statistical sample provided by the SDSS dataset makes this a very robust result. ![The cumulative mass distribution for the galaxies in the lower (blue line) and upper (red line) quartile of the overall density distribution in four redshift bins, identified at the bottom of each plot; only galaxies above the stellar mass completeness limit appropriate for their redshift interval are considered. []{data-label="fig:mass_integral"}](massDist_lowAndHigh.ps){width="9cm"} To extend this result to higher redshift using the VVDS Deep sample, we have to deal with the fact that the magnitude limited nature of the sample produces a stellar mass completeness limit which varies significantly with redshift. As the stellar mass is correlated with galaxy color, and galaxy color is correlated with environment, the only meaningful way we have to search for a possible mass segregation effect in our data is to make sure we are using a mass complete sample at all redshifts. Some discussion about the mass completeness of the VVDS Deep sample has already been presented in @Pozzetti_VVDS_MF, and was further extended in @Vergani_08 and in @Meneux_08 [see their Figure 3)] to better keep into account the differential mass limit as a function of galaxy color in our sample. Using these latter results, we have divided our samples into 4 redshift bins, each one with its own stellar mass completeness limit, as listed in Table \[tab:summary\]. The table lists the limits adopted for the different redshift bins, the median redshift for the objects in the stellar mass complete sample, the logarithm of the stellar mass completeness limit, in solar mass units, and the number of objects in such a sample. Following the results presented by @Meneux_08 on the stellar mass completeness, and the discussion presented by @Franzetti_07 on the uniformity of the VVDS spectral coverage as a function of spectral type, we can be confident that any color or mass-dependent incompleteness in these 4 samples is limited below the 5 percent level, and it cannot therefore have any significant impact on the results presented below. Because of the relatively small number of objects within any redshift bin we are considering, we can only partition the sample in a limited number of large-scale environments. In this work we consider a partition of the various environments according to the estimated galaxy number density contrast. For the total VVDS Deep galaxy sample described in Section \[sec:data\] we have obtained the distribution of the density contrast values, and for this work, unless stated differently, we consider as galaxies in low density environment those objects for which the density contrast value is in the lower quartile of the distribution. Conversely, we consider galaxies in high density environment those objects for which the density contrast value is in the upper quartile of the distribution. As already discussed in , the separation between these two extremes of the density distribution is very robust when using the smoothing length of approximately 8 Mpc which is used in this work. We find that a significant stellar mass segregation is present throughout the VVDS Deep sample used here, up to a redshift of 1.4 (the median z value for our highest redshift bin being 1.12). Fig. \[fig:mass\_integral\] shows such a segregation, plotting the integral distribution of stellar mass values for galaxies in the four redshift intervals listed in Table \[tab:summary\], limited to the mass complete samples of objects with $Log~M_{star} > Log~M_{limit}$; this is the analogous of Figure 3 in . It is quite clear from this plot how the stellar mass distribution in the high density environments is skewed towards higher masses with respect to the distribution in the low density environments. The statistical significance of the observed segregation is at the three-sigma level or above for the three lower redshift bins, while it is only at the two-sigma level for the highest redshift bin, partly because of the lower number of objects we have in that bin. ![Median values of the mass distribution for galaxies in the lower (blue points) and upper (red points) extreme of the overall density distribution, for three different definitions of extreme: upper and lower half of the distribution, or 50% extreme; upper and lower quartile of the distribution, or 25% extreme, upper and lower 10% of the distribution. Only galaxies in the redshift bin $0.5<z<0.7$ and above the mass completeness limit Log($M_{limit}/M_\odot$)=9.5 are considered in this plot. The error bars represent the bootstrap-based uncertainty estimate in the median value determination. As a reference, the amplitude of the median stellar mass offset measured in the SDSS sample is given by the two dashed lines to the right of the plot (see text for details).[]{data-label="fig:mass_contrast"}](massSeparation_extremes.ps){width="8cm"} As expected, given the results, we observe that the strength of this segregation depends mildly on how much the low and high density environments are differentiated. In Fig. \[fig:mass\_contrast\] we show, limited to the redshift bin $0.5<z<0.7$, the median values for the stellar mass distribution for galaxies in low and high density environments as a function of how extreme we take these two environments to be. Therefore in this case we do not consider just objects in the lowermost or highest quartile (i.e. 25% extremes) of the galaxy density contrast distribution, but also those in the lower and upper half (i.e. 50% extreme) of the distribution, and those in the lower and upper 10% of the distribution (10% extremes). Together with the median values we also plot the associated uncertainties, derived using a bootstrap procedure. Although the separation of the median stellar mass values increases as we move towards the extremes of the density distribution (i.e. towards the extremes of the environment), the statistical significance of these differences remains quite constant, according to a two-populations Kolmogorov-Smirnov test applied to the full stellar mass distributions, because of the decreasing number of objects in the more extreme samples. Therefore the significance with which we measure mass segregation between low and high density environments is always approximately 3 sigmas (i.e. the probability that the two stellar mass distributions are actually drawn from the same parent population is always in the 0.1–0.5 percent range). Similar results are obtained for the other redshift bins we are considering in this work. Purely as a reference, the offset in median stellar mass measured by in the two most extreme environments they sampled is indicated in the figure by the two dashed lines on the right of the plot (the mass values where the two lines are drawn are totally arbitrary), although we must remark that the mass completeness limit for that sample, and the definition of environmental densities are different from those used in this work. ![Median values of the mass distribution for galaxies in the lower (blue points) and upper (red points) quartiles of the overall density distribution, in the 4 redshift bins listed in Table \[tab:summary\]. Only galaxies above the common mass completeness limit Log($M_{limit}/M_\odot$)=10.2 are considered. The error bars represent the bootstrap-based uncertainty estimate in the median value determination.[]{data-label="fig:mass_z"}](massSeparation_redshift.ps){width="8cm"} Finally, to examine in detail if and how this mass segregation evolves with cosmic time, we have defined a subsample of 1872 galaxies, limited to objects with Log$M_{star}/M_\odot > 10.3$, which is complete in stellar mass over the full redshift range sampled by our data. Figure \[fig:mass\_z\] shows the median stellar mass value for galaxies in low and high density environments, over four redshift bins (those listed in Table \[tab:summary\]). As in the previous figure, we also plot the uncertainty in the median estimates. Although we observe a small increase in the strength of the mass segregation (as measured from the median value of the mass distribution), from redshift of approximately 1.1 down to redshift of approximately 0.4, this change is not statistically significant: if we compare any two stellar mass distributions for the same environment, but in different redshift bins, we find a probability of more than 50 percent for the two to be drawn from the same parent population (according to a two-populations Kolmogorov-Smirnov test). This is true for both the low and the high density environment samples. A significantly larger mass-complete sample of galaxies is needed before we can reliably detect any significant evolution in the strength of the stellar mass segregation from the local Universe to z=1. ![The cumulative rest-frame B-I color distribution for the galaxies in the lower (blue line) and upper (red line) quartile of the overall density distribution in the redshift bins $0.2<z<0.5$, $0.5<z<0.7$ and $0.7<z<1.4$ respectively.[]{data-label="fig:color_dist"}](colorDist_allMasses.ps){width="9cm"} Rest-frame color segregation {#sec:color_segregation} ============================ Physical properties of galaxies (like their color, dust and gas content, and star formation activity) and their morphologies are significantly inter-related, and it is well documented that the galaxy total stellar mass plays some important role in determining these properties [see for example @Scodeggio_02_cube; @Kauffmann_03_stellarMass]. It is therefore quite natural to expect that some segregation in galaxy properties should be observed, purely as a consequence of the stellar mass segregation we have discussed in the previous section. Rest-frame galaxy colors are well known to show such a segregation in the local Universe [see for example @Blanton_05; @Baldry_06], mirroring the equally well-know morphology-density relation [@Dressler_80]. Moving to higher redshift, both the VVDS and the DEEP2 survey have shown that a significant correlation between galaxy color and the large-scale environment exists at least up to z$\simeq$1 (see and @Cooper_06, respectively), notwithstanding the general trend towards bluer colors which is observed when moving from the local Universe to z$\simeq$1. In particular , by examining the fraction of red and blue galaxies in different large-scale environments sampled by the VVDS dataset, have shown that, up to $z<0.9$, one observes the locally well known correlation between the fraction of red galaxies and the large-scale density of galaxies. At higher redshift ($0.9<z<1.5$) the correlation basically disappears, as more and more massive galaxies become actively star-forming, and it is even possible that the correlation reverses completely, with a predominance of blue galaxies in high-density environments at z$\simeq$1.5. {width="18cm"} In Fig. \[fig:color\_dist\] we confirm and extend the result, by showing the presence of a significant rest-frame color segregation in the VVDS dataset, not only for volume-limited datasets, but also for mass complete ones. Here we plot the integral distributions of rest-frame B-I color values for galaxies in the first two redshift intervals listed in Table \[tab:summary\] separtately, and for galaxies in the last two redshift intervals grouped together. Within each plot, the whole stellar mass complete sample of objects is considered. Here again, because of the completeness in stellar mass and in spectral type coverage already discussed in Section \[sec:mass\_segregation\], we can confidently state that the color distributions for objects in low and high density environments are statistically different at the 3-sigma level (having a probability of being drawn from the same parent population of 0.2–0.4 percent, as estimated from a two-populations Kolmogorov-Smirnov test) for z$<$0.7, while in the redshift interval 0.7$<$z$<$1.4 the significance drops to approximately the 2-sigma level. This result is in complete agreement with the earlier findings of about the disappearance of the correlation between the fraction of red galaxies and the large-scale density of galaxies for z$>$0.9. In Fig. \[fig:color\_massBin\], on the contrary, we show that the rest-frame color segregation is significantly weaker when we consider only galaxies within relatively small stellar mass bins, to the point that we do not observe any statistically significant difference in the color distribution for galaxies in the low and high density environment. This lack of any residual color segregation is in contrast with some findings from the SDSS sample discussed in and in @Baldry_06, that report a significant environmental dependence of galaxy properties, even when considering only objects within relatively small stellar mass bins. One possible explanation for this discrepancy is the very different scale-length over which the local galaxy density is estimated in this work (8 Mpc) and in the SDSS ones (1 Mpc). Another important difference coming into play is the smaller size of the VVDS Deep sample used here, with respect to the SDSS one. It is therefore entirely possible that some segregation in galaxy rest-frame color could still be present in our sample, even when considering narrow mass bins, but the relatively minor strength of this segregation, coupled with the small sample size, could prevent us from measuring this effect with any statistical significance. This topic will be analyzed in further detail using the larger zCOSMOS sample in a forthcoming paper (Cucciati et al., in preparation). This discrepancy aside, the observed significant weakening of the galaxy color segregation which is observed when we consider galaxies in relatively narrow stellar mass bins strongly suggests that the segregation we observe on 8 Mpc scales is mostly (if not entirely) driven by the underlying stellar mass segregation, coupled with the well known color-stellar mass correlation. A rather similar conclusion could be derived from the observed absence of any environmental effects on the locus of the color-stellar mass relation, both for objects in the red sequence and in the blue cloud, which was discussed by @Cassata_07, using the data from the COSMOS survey [@COSMOS]. Conclusions =========== A significant stellar mass segregation as a function of the large-scale galaxy environment, the latter defined from the measured galaxy number density contrast within scales of approximately 8 Mpc, is observed in the VVDS sample over the whole redshift range from z=0.2 up to z$\sim$1.4: the stellar mass distribution of high density regions shows the presence of a larger number of high stellar mass galaxies than the distribution observed in low density regions. The scales over which this segregation is observed are much bigger than the typical group or cluster scale (approximately 1 Mpc) where environmental effects are expected to play a significant role in shaping galaxy evolution. It is however impossible for us, with the present sample, to evaluate the possibility that the mass segregation signature we observe on the 8 Mpc scale could just be a diluted signal produced by a much stronger segregation at the 1 Mpc scale, because the typical mean interparticle separation of the VVDS Deep sample does not allow us to reliably derive galaxy densities on the 1 Mpc scale. It is remarkable however that the observed segregation is in rather good agreement with the expectations from the hierarchical models, which predict a very similar segregation of dark matter haloes (the strength in the mass segregation of the dark matter haloes is expected to be 2–3 times stronger than the observed stellar mass one, but part of this discrepancy could be explained by the presence of a large number of very low mass halos in the underdense regions that do not host any galaxy, see @Abbas_Sheth_05). The strength of the observed stellar mass segregation decreases marginally with increasing redshift, from z=0.4 to z=1.2, but within the relatively small statistics offered by the mass complete VVDS galaxy sample such a weakening of the mass segregation cannot be considered as a statistically robust result. A significant rest-frame galaxy color segregation is observed as well, mirroring the stellar mass one. However we have demonstrated that a large fraction of this color segregation is simply a reflection of the stellar mass segregation, via the well known correlation between stellar mass and galaxy color. In fact, when we compare rest-frame colors for galaxies with stellar mass within a narrow range, we do not find significant differences in the color distribution as a function of the large-scale environment. These results, coupled with the observed differential galaxy clustering reported in @Abbas_Sheth_06, provide strong support to the hypothesis that an important fraction of the observed environmental effects on galaxy properties, like broadband optical color, or star formation history, are just the reflection of the correlation between galaxy stellar mass and the galaxy hosting dark matter halo mass, which in turn correlates with the surrounding large scale environment. Further discussion on the correlation between star formation activity and the large-scale environment where galaxies are located will be presented in a forthcoming paper (Vergani et al., in preparation). MS would like to thank Frank van den Bosch for a very useful discussion on environmental effects, that provided the motivation for this work.\ This research has been developed within the framework of the VVDS consortium. This work has been partially supported by the CNRS-INSU and its Programme National de Cosmologie (France), and by Italian Ministry (MIUR) grants COFIN2000 (MM02037133) and COFIN2003 (num.2003020150) and by INAF grants (PRIN-INAF 2005). DV acknowledges the support through a Marie Curie ERG, funded by the European Commission under contract No. MERG-CT-2005-021704.\ The VLT-VIMOS observations have been carried out on guaranteed time (GTO) allocated by the European Southern Observatory (ESO) to the VIRMOS consortium, under a contractual agreement between the Centre National de la Recherche Scientifique of France, heading a consortium of French and Italian institutes, and ESO, to design, manufacture and test the VIMOS instrument.\ Based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Science de l’Univers of the Centre National de la Recherche Scientifique (CNRS) of France, and the University of Hawaii. This work is based in part on data products produced at TERAPIX and the Canadian Astronomy Data Centre as part of the Canada-France-Hawaii Telescope Legacy Survey, a collaborative project of NRC and CNRS. [^1]: Based on data obtained with the European Southern Observatory Very Large Telescope, Paranal, Chile, program 070.A-9007(A), and on data obtained at the Canada-France-Hawaii Telescope, operated by the CNRS of France, CNRC in Canada and the University of Hawaii.

--- abstract: 'High-order harmonic generation is investigated for H$_2^+$ and D$_2^+$ with and without Born-Oppenheimer approximation by numerical solution of full dimensional electronic time-dependent Schrödinger equation under 4-cycle intense laser pulses of 800 nm wavelength and $I$=4, 5, 7, 10 $\times 10^{14}$ W$/$cm$^2$ intensities. For most harmonic orders, the intensity obtained for D$_2^+$ is higher than that for H$_2^+$, and the yield difference increases as the harmonic order increases. Only at some low harmonic orders, H$_2^+$ generates more intense harmonics compared to D$_2^+$. The results show that nuclear motion, ionization probability and system dimensionality must be simultaneously taken into account to properly explain the isotopic effects on high-order harmonic generation and to justify experimental observations.' author: - 'Hamed Ahmadi$^{1}$' - 'Ali Maghari$^{1}$' - 'Hassan Sabzyan$^{2}$' - 'Ali Reza Niknam$^{3}$' - 'Mohsen Vafaee$^{4}$' bibliography: - 'p7.bib' title: 'Effect of nuclear motion on high-order harmonic generation of H$_2^+$ in intense ultrashort laser pulses ' --- Introduction ============ High-order harmonic generation (HHG) is one of the phenomena observed in the interaction of intense laser pulses with atoms and molecules \[1,2\]. A three-step model for the description of the HHG mechanism has been proposed by Corkum \[3\] and extended by Lowenstein et al. \[4\]. In the first step of this mechanism, an electron wavepacket tunnels from an atom or a molecule into the continuum. In the second step, the electron moves away from the ion core, and after the field reverses, it is driven back to it. The third step arises when the electron recombines with its parent ion in which a high energetic photon is emitted. This model can be used in tunnelling regime which predicts maximum recollision energy of $3.17U_p$, where $U_p=I/{4\omega^2}$, is the pondermotive energy in which $I$ and $\omega$ are laser intensity and angular frequency, respectively. Based on the three-step model, for each harmonic order smaller than the cutoff harmonic order, we have two trajectories that contribute to the HHG with the same kinetic energy. These two trajectories return to their parent ion at different times. In one cycle of laser pulse, the electrons released over time interval $0.3 T_0<t<0.5 T_0$ ($T_0=2\pi/\omega_0$, with $\omega_0$ being laser frequency), return to the core during $0.5 T_0<t<0.95 T_0$. The trajectories travelled by these electrons are called short trajectories. While the electrons released over the $0.25 T_0<t<0.3 T_0$, return to the core during $0.95 T_0<t<1.25T_0$, and their path are called long trajectories because of longer round trip times than those of the short trajectories. The HHG is used to generate attosecond laser pulses \[5,6\] and to get structural information \[7-11\]. Here, we focus on the HHG reported for the H$_2^+$, H$_2$ and their corresponding isotopomers. The HHG process in molecules is more complex than that in atoms because of nuclear motion \[12\], two-center interference \[13\] and different orientations of molecule with respect to the laser field \[13\]. Effects of different initial vibrational states \[14,15\], initial nuclear velocities \[16\] and relation between electronic wavepacket expansion and internuclear distance on the HHG produced by H$_2^+$ \[17\] have been reported. In addition, extraction of nuclear dynamics from HHG spectra \[18\], effect of nuclear motion on the HHG efficiency by varying pump-probe time delay \[19\], on the broadening of cutoff regime \[20\], on the length of generated attosecond laser pulses \[21\] and on the generation of isolated attosencond laser pulses for H$_2^+$ have been investigated \[22,23\]. The HHG for H$_2$ and D$_2$ with higher yield for heavier isotopomer has been theoretically \[12\] and experimentally \[8,24-25\] reported. For most harmonic orders, experiments on H$_2$ and D$_2$ \[8,24-25\], and CH$_4$ and CD$_4$ \[8\] reveal the higher HHG yield for heavier isotopomers. While in theoretical works on H$_2^+$ and D$_2^+$, Feng *et al*. \[22\] reported higher HHG yield in H$_2^+$ but Bandrauk *et al*. \[23\] reported higher HHG yield for D$_2^+$. In most theoretical works mentioned above, electron and nuclei are considered quantum mechanically in a one-dimensional (1D) model. The full-dimensional electron wavepacket expansion during laser interaction and its consequent HHG cannot be described properly by a one dimensional model \[17\]. In the present work, we are interested in the effect of the motion of nuclei on the HHG spectra by considering different isotopomers. As stated above, the theoretical results on the HHG yield on 1D H$_2^+$ and D$_2^+$ are different from experimental resulsts on H$_2$ and D$_2$. We want to address the discrepancy between these experimental and theoretical reports. The HHG is essentially a single-electron phenomenon and the HHG yield for H$_2^+$ and D$_2^+$ has not been experimentally reported because of difficulties in sample preparation. Therefore, in this work, besides comparing to theoretical works, we compare the HHG yield of different isotopomers of our single-electron systems with available experimental results obtained for two-electron systems. In this study, full-dimensional electronic time-dependent Schrödinger equation (TDSE) beyond Born-Oppenheimer approximation (NBO) is solved numerically for H$_2^+$, D$_2^+$ and X$_2^+$ (X is a virtual isotope of H being 10 times heavier). The full-dimensional electronic TDSE within Born-Oppenheimer approximation (BO) for H$_2^+$, which is indicated throughout the article by H$_2^+$(BO), is also solved to compare with NBO results. The equilibrium internuclear distance within BO is set to $R_e=1.96$ a.u. All calculations have been done with 4-cycle laser pulses of 800 nm wavelength with $I=$4, 5, 7 and 10 $\times 10^{14}$ W$/$cm$^2$ intensities. The Morlet-wavelet Fourier transform is used for time-frequency analysis of harmonics. We use atomic units throughout the article unless stated otherwise. Computational Methods ===================== The time-dependent Schrödinger equation for H$_2^+$ (D$_2^+$) with electron cylindrical coordinate $(z,\rho)$ ͒ with respect to the molecular center of mass and internuclear distance $R$, for both $z$ and $R$ parallel to the laser polarization direction, can be read (after elimination of the center-of-mass motion) as \[26-27\] $$\begin{aligned} \label{eq:1} i \frac{\partial \psi(z,\rho, R,t)}{\partial t}={\widehat H}(z,\rho, R,t)\psi(z,\rho, R,t).\end{aligned}$$ In this equation, Ĥ is the total electronic and nuclear Hamiltonian which is given by $$\begin{aligned} \label{eq:2} \widehat{H}(z,\rho, R,t)=&-\frac{2m_N+m_e}{4m_Nm_e}[\frac{\partial^2}{\partial \rho^2}+\frac{1}{\rho}\frac{\partial}{\partial \rho}+\frac{\partial^2}{\partial z^2}] \nonumber \\ & -\frac{1}{m_N}\frac{\partial^2}{\partial R^2}+V_C(z,\rho, R,t),\end{aligned}$$ with $$\begin{aligned} \label{eq:3} \widehat{V}_C&(z,\rho, R,t)=-\frac{1}{\sqrt{(z+\frac{R}{2})^2+\rho^2}}-\frac{1}{\sqrt{(z-\frac{R}{2})^2+\rho^2}} \nonumber \\ &+\frac{1}{R}+(\frac{2m_N+2m_e}{2m_N+me})zE_0f(t)sin(\omega t).\end{aligned}$$ In these equations, $E_0$ is the laser peak amplitude, $m_e$ and $m_N$ are the masses of electron and single nuclei, $\omega$ angular frequency and *f*(t) is the laser pulse envelope which is considered to have Gaussian form as $$\begin{aligned} \label{eq:4} {\textit{f}}(t)=exp[-\frac{4 \ln(2)(t-2T_0)^2}{\tau^2}],\end{aligned}$$ in which $\tau$ is a measure of the laser pulse duration (full width at half maximum (FWHM)) which is set to 3 femtosecond ($\sim$ 124 a.u.). The shape of the electric field of the laser pulse used in this work is shown in Fig. 1. [c]{} The TDSE is solved using unitary split-operator methods \[28-29\] with 11-point finite difference scheme through a general nonlinear coordinate transformation for both electronic and nuclear coordinates which is described in more details in our previous works \[30-32\]. The grid points for z, $\rho$, and R coordinates are 300, 83, and 210, respectively. The finest grid size values in this adaptive grid schemes are 0.13, 0.1, and 0.025, respectively for $z$, $\rho$, and $R$ coordinates. The grids extend up to $z_{max} = 34$, $\rho_{max} = 25$, and $R_{max}= 16$. Only for $I$=1$\times 10^{15}$ W$/$cm$^2$ intensity, the $z$ grid points is set to 500 and $z_{max} = 98$. The size of the simulation boxes is considered large enough so that the loss of norm at the end of laser pulse does not exceed a few percent. The HHG spectra are calculated as square of the windowed Fourier transform of dipole acceleration $a_z(t)$ in the electric field direction (z) as $$\begin{aligned} \label{eq:5} S(\omega)= \vert \int_0^T a_z(t)\,H(t)\,exp[-i\omega t]\,dt\; \vert ^2,\end{aligned}$$ where $$\begin{aligned} \label{eq:6} H(t)= \frac{1}{2}[1-cos(2\pi \frac{t}{T})],\end{aligned}$$ is the Hanning filter and $T$ is the total pulse duration which is set to 4 optical cycles (one optical cycle of 800 nm wavelength equals 2.6 fs). The time dependence of harmonics is obtained by Morlet-wavelet transform of dipole acceleration $a_z(t)$ via \[33-34\] $$\begin{aligned} \label{eq:7} &w(\omega,t)= \sqrt{ \frac{\omega}{\pi^\frac{1}{2}\sigma}}\times \nonumber \\ &\int_{-\infty}^{+\infty}a_z(t^\prime)exp[-i\omega (t^\prime-t)]exp[-\frac{\omega^2 (t^\prime-t)^2}{2\sigma^2}]dt^\prime.\end{aligned}$$ We tried different $\sigma$ values and obtained the best time-frequency resolution by $\sigma=2\pi$ which is used in this work. Results of the calculations with and without Born-Oppenheimer (fixed-nuclei) approximation are denoted by BO and NBO, respectively. Results and Discussion ====================== The HHG spectra of H$_2^+$, D$_2^+$, X$_2^+$ and H$_2^+$(BO) obtained under 4-cycle 800 nm laser pulses of $I$=4, 5, 7, 10 $\times 10^{14}$ W$/$cm$^2$ intensities are shown in Fig. 2. For some low harmonics, the HHG yield is higher for H$_2^+$ than the other two isotopomers and H$_2^+$(BO) (right panels of Fig. 2), but for harmonic orders greater than $\sim$ 25, the HHG yield is higher for H$_2^+$(BO), and for NBO cases, the difference between the HHG yield of lighter and heavier isotopomers increases as the harmonic order increases. Such behaviours have already been observed in recent experiments. The higher HHG yield for H$_2^+$ than D$_2^+$ for some low harmonic orders is observed in experiments on H$_2$ and D$_2$ at 1300 nm wavelength \[25\]. Higher HHG yield for D$_2^+$ than for H$_2^+$ at higher harmonic orders is also reported in experiments on H$_2$ and D$_2$ at 800 and 1300 nm wavelengths \[8,24-25\]. As shown in Fig. 2, for the intensities used in our calculations, the cutoff occurs at harmonic orders smaller than that predicted by the three-step model \[4\]. The cutoff harmonic order $N_c$, according to the three-step model, for H$_2^+$, with ionization potential $I_p=1.1$ a.u., under 800 nm wavelength ($\omega_l=0.057$ a.u.) and different intensities are given in TABLE I. $I (I_e)$ W/cm$^2$ $U_p\, (a.u.)$ $N_c=(3.17U_p+1.32I_p)/\omega_l$ -------------------------- ---------------- ---------------------------------- 4 (3.15)$\times 10^{14}$ 0.69 64 5 (3.9)$\times 10^{14}$ 0.86 73 7 (5.5)$\times 10^{14}$ 1.21 93 10 (7.8)$\times 10^{14}$ 1.71 121 : \[Table\] The cutoff harmonic order $N_c$ according to the three-step model for H$_2^+$, with ionization potential $I_p=1.1$ a.u., and pondermotive energy $U_p$ under 800 nm wavelength ($\omega_l=0.057$ a.u.) and $I$=4, 5, 7, 10 $\times 10^{14}$ W$/$cm$^2$ intensities. The effective intensity (of the two central peaks, $I_e$) due to Gaussian envelope is also given for each laser intensity, $I$. The $N_c$ and $U_p$ are calculated for effective intensities. The plateau region is wider for heavier isotopomers and the NBO cutoff approaches the BO cutoff as the isotopomer becomes heavier. For harmonic orders between 13-25 (Fig. 2), everywhere H$_2^+$ spectrum has valley (peak), the X$_2^+$ spectrum has peak (valley), and also as isotopomer becomes heavier, spectra modulation approaches to the corresponding BO spectra, as we see similar modulation between X$_2^+$ and H$_2^+$(BO). To justify the above observations, we study ionization probability, nuclear motion and compare these full-dimensional results with those reported for 1D models \[22,23\]. First, the time-dependent contribution of ground state population in the evolving wavepacket for the four laser pulses with different intensities are calculated and demonstrated in Fig. 3. [c]{} This figure shows that the ground state population is lower for H$_2^+$ than for D$_2^+$ and X$_2^+$. When ground state population is lower, higher ionization is expected. This higher ionization plays a positive role in the HHG enhancement for all harmonics because the more released electron results in more return of the released electron to the core giving rise to the HHG. Therefore, the higher ionization can justify higher HHG yield observed for H$_2^+$ than those observed for other isotopomers at low harmonic orders. To demonstrate better the effects of nuclear motion on the HHG spectra, the Morlet-wavelet Fourier transform of the HHG spectra of Fig. 2 for BO and NBO cases, and the time-dependent average internuclear distance for NBO cases are derived and depicted in Figs. 4 and 5, respectively. [c]{} Each row of Fig. 4 is related to a specific (labelled) intensity. For all intensities in Fig. 4, the three peaks are observed around 1.75, 2.25 and 2.75 optical cycles (o.c.) which, based on the three-step model, correspond to electron trajectories released at 1.325-1.5, 1.775-2 and 2.225-2.5 o.c., respectively. The very weak peak around 1.75 o.c., which is only present for $I$=4 and 5 $\times 10^{14}$ W$/$cm$^2$ intensities with BO, is related to the short trajectories born during 1.325-1.5 o.c. which is suppressed when NBO is considered as a result of nuclear wavefunction spreading that is justified in Ref. \[23\]. For the weak peak around 2.75 o.c., we see almost both short and long trajectories which are limited to low harmonic orders not of importance in this work. The strong peak around 2.25 o.c. is the most important peak which is extended to high harmonic orders and we concentrate on it. Note that for harmonic orders greater than $\sim$ 30, this peak is only responsible for differences seen and stated for Fig. 2. The birth and return times of this peak are between 1.775-2.675 o.c. in which the electric field has two strong minimum (at 1.775 o.c.) and maximum (at 2.225 o.c.). The electric field strength should be high enough to ionize and accelerate the electron, and to return the released electrons to the core with higher energy. It can be said that for most panels in Fig. 4, the short-trajectory branch dominantly contribute to the peak around 2.25 o.c., and the weak long-trajectory branch is seen only for H$_2^+$(BO) at $I$=7 $\times 10^{14}$ W$/$cm$^2$ intensity. As shown in Fig. 5, the internuclear distance is increased more for the lighter isotopomer H$_2^+$ than D$_2^+$ and X$_2^+$. This increase is higher for higher laser intensities. The sharp increase in the internuclear distance of lighter isotopomer can affect the HHG spectra. We can see the effect of nuclear motion in two cases: harmonic orders 13-25 and above 25. For 13-25 harmonic orders, opposite modulation of X$_2^+$ and H$_2^+$ HHG spectra and similar modulation of X$_2^+$ and H$_2^+$(BO) HHG spectra (Fig. 2), suggest that nuclear motion is responsible for these observations. Note that electronic structures of different isotopomers used in this work are similar and therefore the difference between their HHG spectra can be attributed to their nuclear motion only. It should, however, be mentioned that ionization rate of these isotopomers are different and we showed that H$_2^+$ has higher ionization than other (Fig. 3). Now, we consider nuclear motion for harmonic orders greater than 25. As stated before, there is only one peak around 2.25 o.c which is mainly related to the short trajectories that contribute to the HHG for high harmonic orders. As it can be deduced from Fig. 4 by looking at short-trajectory peaks, higher harmonic orders are produced at longer times (the time between the birth and recollision of the ionized electron increases as harmonic order increases) which is vice versa for long trajectories \[33\]. The HHG attenuation of H$_2^+$ compared to those of heavier isotopomers and H$_2^+$(BO) can be related to more increase in the internuclear distance for lighter isotopomer (Fig. 5). The difference between the HHG yield of H$_2^+$ and those of other isotopomers is increased with the increase of the harmonic order because higher harmonic orders are produced at longer times when there is a larger internuclear distance for lighter isotopomer, giving rise thus to attenuation of the HHG. For all intensities in Fig. 4, the HHG produced in the long trajectory is very weak compared to that of the short trajectory. For low intensities, $I$=4 and 5 $\times 10^{14}$ W$/$cm$^2$, long trajectory is completely suppressed for both BO and NBO cases. As shown in Fig 4, the long-trajectory attenuation is more obvious for NBO than for BO, and for lightest isotopomer is more distinct because of faster and larger nuclear motion. The long-trajectory suppression in NBO has already been observed for the 1D electronic NBO calculations on H$_2^+$ \[23\]. Nevertheless, the 1D BO results for which short and long trajectories have similar HHG strengths \[23\], are not in agreement with the long-trajectory suppression observed with full-dimensional electronic BO calculations carried out in this work. This difference is because of the drift induced in the laser propagation direction, which can not happen in 1D models, and thus reduces the recollision of electrons in different trajectories into their parent ions. [l]{} As stated before, for both NBO and BO cases, the cutoff position is lower than that obtained by the three-step model and 1D NBO calculations \[22\]. Two reasons can be considered here. First, taking into account electron in full dimension causes electron with large return time to lose its chance to recollide with its parent ion (the electron drifts in the propagation direction). Second, nuclear motion also leads to the suppression of the harmonics near the cutoff. Since the short trajectories have higher contributions to the HHG spectra, especially near the cutoff, and for these trajectories, higher harmonics are produced at longer times, and at longer times, the nuclei are more displaced away from their equilibrium positions, the role of nuclear motion in attenuating the HHG spectra is increased with increasing harmonic order. This can also be deduced when the NBO cutoff approaches the BO cutoff as the isotopomer becomes heavier (Fig. 2). For most harmonics (except for the first few low harmonic orders), the HHG is more intense for heavier isotopomers, which is in agreement with experimental reports \[8,24-25\] and is in contrast to the results predicted by the 1D electronic NBO calculation, that is the lighter isotopomer produces higher HHG yield over the whole spectrum \[22\]. Feng *et al*. explained their results based on the ionization probabilities which is valid only for 1D NBO calculations \[22\]. Based on the full-dimensional electronic NBO calculations carried out in this work, it can be stated that more ionization occurs for lighter isotopomers which is responsible for higher HHG yield at low harmonic orders. In 1D electronic model, the recollision is overestimated compared to that in the real full-dimensional calculations. The recollision-recombination occurrence is suppressed for full-dimensional electronic NBO cases because of drift induced in the laser propagation direction and the increase of internuclear distance. Conclusion ========== We solved numerically time-dependent Schrödinger equation for H$_2^+$ and D$_2^+$ with and without Born-Oppenheimer approximation to investigate the effect of the nuclear motion via analysis of the high-order harmonic generation to address the discrepancy between the HHG yield obtained for H$_2^+$ and D$_2^+$ isotopomers by one-dimensional non-Born-Oppenheimer calculations \[22,23\] and those experimentally observed on H$_2$ and D$_2$ \[8,24-25\]. While, our results show that when nuclear motion is taken into account, higher HHG yield is obtained for heavier isotopomer which is compatible with experimental reports \[8,24-25\]. The 1D electronic NBO calculations overestimate the recollision-recombination effect relative to full-dimensional electronic NBO calculations, especially at longer return times of the released electron corresponding to increased internuclear distance, which is detrimental to the recollision-recombination phenomenon, and thus attenuates the HHG production. References ========== [100]{} T. Brabec and F. Krausz, Rev. Mod. Phys. **72**, 545 (2000). C. Winterfeldt, C. Spielmann, and G. Gerber, Rev. Mod. Phys. **80**, 117 (2008). P. B. Corkum, Phys. Rev. Lett. **71**, 1994 (1993). M. Lewenstein, P. Balcou, M. Y. Ivanov, A. L$^{^,}$Huillier and P. A. Corkum, Phys. Rev. A **49**, 2117 (1994). E. Goulielmakis, V. S. Yakovlev, A. L. Cavalieri, M. Uiberacker, V. Pervak, A. Apolonski, R. Kienberger, U. Kleineberg, and F. Krausz, Science **317**, 769 (2007). F. Krausz and M. Y. Ivanov, Rev. Mod. Phys. **81**, 163 (2009). J. Itatani, J. Levesque, D. Zeidler, H. Niikura, H. Pépin, J. C. Kieffer, P. B. Corkum, and D. M. Villeneuve, Nature (London) **432**, 867 (2004). S. Baker, J. S. Robinson, C. A. Haworth, H. Teng, R. A. Smith, C. C. Chiril$\breve{a}$, M. Lein, J. W. G. Tisch, and J. P. Marangos, Science **312**, 424 (2006). S. Haessler, J. Caillat, W. Boutu, C. Giovanetti-Teixeira, T. Ruchon, T. Auguste, Z. Diveki, P. Breger, A. Maquet, B. Carré, R. Taïeb, and P. Salières, Nat. Phys. **6**, 200 (2010). C. Vozzi, M. Negro, F. Calegari, G. Sansone, M. Nisoli, S. D. Silvestri, and S. Stagira, Nat. Phys. **7**, 822 (2011). W. Li, X. Zhou, R. Lock, S. Patchkovskii, A. Stolow, H. C. Kapteyn, and M. M. Murnane, Science **322**, 1207 (2008). M. Lein, Phys. Rev. Lett. **94**, 053004 (2005). M. Lein, N. Hay, R. Velotta, J. P. Marangos, and P. L. Knight, Phys. Rev. Lett. **88**, 183903 (2002). Y. H. Guo, H. X. He, J. Y. Liu and G. Z. He, J. Mol. Struct. **947**, 119 (2010). J. Zhao and Z. Zhao, Phys. Rev. A **78**, 053414 (2008). N. T. Nguyen, V. H. Hoang and V. H. Le, Phys. Rev. A **88**, 023824 (2013). A. D. Bandrauk and H. Yu, J. Phys. B: At. Mol. Opt. Phys. **31**, 4243 (1998). R. Daniele, G. Castiglia, P. Corso, E. Fiordilino, F. Morales and G. Orlando, J. Mod. Opt. **56**, 751 (2009). T. Bredtmann, S. Chelkowski and A. D. Bandrauk, J. Phys. Chem. A **116**, 11398 (2012). W. Qu, Z. Chen, Z. Xu and C. H. Keitel, Phys. Rev. A **65**, 013402 (2001). A. D. Bandrauk, S. Chelkowski and H. Lu, J. Phys. B: At. Mol. Opt. Phys. **42**, 075602 (2009). L. Feng and T. Chu, J. Chem. Phys. **136**, 054102 (2012). X. L. Ge, T. Wang, J. Guo, and X. S. Liu Phys. Rev. A **89**, 023424 (2014). S. Baker, J. S. Robinson, M. Lein, C. C. Chiril$\breve{a}$, R. Torres, H. C. Bandulet, D. Comtois, J. C. Kieffer, D. M. Villeneuve, J. W. G. Tisch, and J. P. Marangos, Phys. Rev. Lett. **101**, 053901 (2008). H. Mizutani, S. Minemoto, Y. Oguchi and H. Sakai, J. Phys. B: At. Mol. Opt. Phys. **44**, 081002 (2011). J. R. Hiskes, Phys. Rev. **122**, 1207 (1961). M. Vafaee, Phys. Rev. A **78**, 023410 (2008). A. D. Bandrauk and H. Shen, J. Chem. Phys. **99**, 1185 (͑1993)͒. M. D. Feit, J. A. Fleck, Jr. , and A. Steiger, J. Comput. Phys. **47**, 412 (1982). M. Vafaee and H. Sabzyan, J. Phys. B **37**, 4143 (2004). M. Vafaee, H. Sabzyan, Z. Vafaee, and A. Katanforoush, e-print arXiv:physics/0509072. M. Vafaee, H. Sabzyan, Z. Vafaee, and A. Katanforoush, Phys. Rev. A **74**, 043416 (2006). C. Chandre, S. Wiggins and T. Uzer, Phys. D **181**, 171 (2003). A. D. Bandrauk, S. Chelkowski and H. Lu, Chem. Phys. **414**, 73 (2013).

--- abstract: 'We derive simple analytical formulae for the renormalization group running of neutrino masses, leptonic mixing angles and CP phases, which allow an easy understanding of the running. Particularly for a small angle $\theta_{13}$ the expressions become very compact, even when non-vanishing CP phases are present. Using these equations we investigate: (i) the influence of Dirac and Majorana phases on the evolution of all parameters, (ii) the implications of running neutrino parameters for leptogenesis, (iii) changes of the mass bounds from WMAP and neutrinoless double $\beta$ decay experiments, relevant for high-energy mass models, (iv) the size of radiative corrections to $\theta_{13}$ and $\theta_{23}$ and implications for future precision measurements.' ---   TUM-HEP-510/03\ DESY 03-065 [ **Running Neutrino Masses, Mixings and CP Phases: Analytical Results and Phenomenological Consequences**]{}\ Stefan Antusch[^1], Jörn Kersten[^2], Manfred Lindner[^3]\ [*Physik-Department T30, Technische Universität München\ James-Franck-Stra[ß]{}e, 85748 Garching, Germany* ]{}\ Michael Ratz[^4]\ [*Deutsches Elektronensynchrotron DESY\ 22603 Hamburg, Germany* ]{} Introduction ============ The Standard Model (SM) agrees very well with experiments and the only solid evidence for new physics consists in the observation of neutrino masses. Compared to quarks and charged leptons they are tiny, for which the see-saw mechanism [@Yanagida:1980; @Glashow:1979vf; @Gell-Mann:1980vs; @Mohapatra:1980ia] provides an attractive explanation. The parameters which enter into the neutrino mass matrix usually stem from model predictions at high energy scales, such as the scale $M_\mathrm{GUT}$ of grand unification. The measurements and bounds for neutrino masses and lepton mixings, on the other hand, determine the parameters at low energy. The high- and low-energy parameters are related by the renormalization group (RG) evolution, so that low-energy data yield only indirect restrictions for mass models or other high-energy mechanisms like leptogenesis [@Fukugita:1986hr]. It is well-known that the model independent RG evolution between low energy and the lowest see-saw scale can have large effects on the leptonic mixing angles and on the mass squared differences, in particular if the neutrinos have quasi-degenerate masses [@Tanimoto:1995bf; @Ellis:1999my; @Casas:1999tp; @Casas:1999ac; @Chankowski:1999xc; @Casas:1999tg; @Balaji:2000au; @Haba:2000tx; @Miura:2000bj; @Chankowski:2000fp; @Chen:2001gk; @Chankowski:2001mx; @Parida:2002gz; @Dutta:2002nq; @Miura:2002nz; @Bhattacharyya:2002aq; @Joshipura:2002xa; @Frigerio:2002in]. RG effects may even serve as an explanation for the discrepancy between the mixings in the quark and the lepton sector [@Mohapatra:2003tw]. The RG equations (RGEs) for the neutrino mass operator and for all the other parameters of the theory have to be solved simultaneously. The mixing angles, phases and mass eigenvalues can then be extracted from the evolved mass matrices. Both steps are, however, non-trivial and can only be performed numerically in practice. In order to determine the change of the parameters under the RG flow in a qualitative and, to a reasonable accuracy, also quantitative way, it is useful to derive analytical formulae for the running of the masses, mixing angles and phases. This was done in [@Chankowski:1999xc] assuming CP conservation and in [@Casas:1999tg] for the general case. We modify the derivation of [@Casas:1999tg] by a step which simplifies the formulae that arise after explicitly writing out the dependence on the mixing parameters. These results are exact, and they make it easier to derive simple approximations in the limit of small $\theta_{13}$. These approximations are very useful in understanding the RG evolution of the phases and the phase dependence of the evolution of other parameters. For example, we find that the phases show significant running. Consequently, vanishing phases at low energy appear unnatural unless exact CP conservation is a boundary condition at high energy, which seems unlikely, since the CP phase in the quark sector is sizable. The presence of CP phases at low energies has significant impact on observations [@Dick:1999ed; @Frigerio:2002rd; @Frigerio:2002fb]. The outline for the paper is: In Sec. \[sec:RGEvolution\] we present analytical formulae for the RG evolution of the neutrino masses, leptonic mixing angles and phases, where an expansion in the small angle $\theta_{13}$ is performed. This leads to very simple and in most cases accurate formulae which are compared with numerical results. Sec. \[sec:Applications\] is devoted to phenomenological consequences for leptogenesis, the WMAP bound, the effective neutrino mass relevant for neutrinoless double beta decay and precision measurements of $\theta_{13}$ and $\theta_{23}$. RG Evolution of Leptonic Mixing Parameters and Neutrino Masses {#sec:RGEvolution} ============================================================== In this study, we will focus on neutrino masses which can be described by the lowest-dimensional neutrino mass operator compatible with the gauge symmetries of the SM. This operator reads in the SM $$\label{eq:Kappa:Babu:1993:1} \mathscr{L}_{\kappa} =\frac{1}{4} \kappa_{gf} \, \overline{\ell_\mathrm{L}^\mathrm{C}}^g_c\varepsilon^{cd} \phi_d\, \, \ell_{\mathrm{L}b}^{f}\varepsilon^{ba}\phi_a +\text{h.c.} \;,$$ and in its minimal supersymmetric extension, the MSSM, $$\label{eq:Kappa-MSSM-s} \mathscr{L}_{\kappa}^{\mathrm{MSSM}} \,=\, \mathscr{W}_\mathrm{\kappa} \big|_{\theta\theta} +\text{h.c.} = -\tfrac{1}{4} {\kappa}^{}_{gf} \, {\bbsymbol{l}}^{g}_c\varepsilon^{cd} {\bbsymbol{h}}^{(2)}_d\, \, {\bbsymbol{l}}_{b}^{f}\varepsilon^{ba} {\bbsymbol{h}}^{(2)}_a \big|_{\theta\theta} +\text{h.c.} \;.$$ $\kappa_{gf}$ has mass dimension $-1$ and is symmetric under interchange of the generation indices $f$ and $g$, $\varepsilon$ is the totally antisymmetric tensor in 2 dimensions, and $\ell_\mathrm{L}^\ChargeC$ is the charge conjugate of a lepton doublet. $a,b,c,d \in \{1,2\}$ are $\mathrm{SU}(2)_\mathrm{L}$ indices. The double-stroke letters ${\bbsymbol{l}}$ and ${\bbsymbol{h}}$ denote lepton doublets and the up-type Higgs superfield in the MSSM. After electroweak (EW) symmetry breaking, a Majorana neutrino mass matrix proportional to $\kappa$ emerges as illustrated in Fig. \[fig:KappaVertex\]. $ {\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{KappaSM.eps}}}}} \xrightarrow[\mbox{{\small breaking}}]{\mbox{{\small EW symmetry}}} {\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{KappaEWSB.eps}}}}} $ The above mass operator provides a rather model-independent way to introduce neutrino masses as there are many possibilities to realize it radiatively or at tree-level within a renormalizable theory (see e.g. [@Ma:1998dn]). The tree-level realizations from integrating out heavy singlet fermions and/or Higgs triplets naturally appear for instance in left-right-symmetric extensions of the SM or MSSM and are usually referred to as type I and type II see-saw mechanisms. The energy dependence of the effective neutrino mass matrix below the scale where the operator is generated (which we will call $M_1$ in the following) is described by its RGE. At the one-loop level, this equation is given by [@Chankowski:1993tx; @Babu:1993qv; @Antusch:2001ck; @Antusch:2001vn] $$\label{eq:BetaKappa} 16\pi^2 \, \frac{\D\kappa}{\D t} \,=\, C\,(Y_e^\dagger Y_e)^T\,\kappa+C\,\kappa\,(Y_e^\dagger Y_e) + \alpha\,\kappa\;,$$ where $t=\ln(\mu/\mu_0)$ and $\mu$ is the renormalization scale[^5] and where $$\begin{aligned} C &=& 1 \hphantom{-\frac{3}{2}} \;\; \text{in the MSSM}\;, \nonumber\\ C &=& -\frac{3}{2} \hphantom{1} \;\; \text{in the SM}\;.\end{aligned}$$ In the SM and in the MSSM, $\alpha$ reads $$\begin{aligned} \alpha_\mathrm{SM} & = & -3 g_2^2 + 2 (y_\tau^2+y_\mu^2+y_e^2) + 6 \left( y_t^2 + y_b^2 + y_c^2 + y_s^2 + y_d^2 + y_u^2 \right) + \lambda \;, \\ \alpha_\mathrm{MSSM} & = & -\frac{6}{5} g_1^2 - 6 g_2^2 + 6 \left( y_t^2 + y_c^2 + y_u^2 \right) \;.\end{aligned}$$ Here $Y_f$ ($f\in\{e,d,u\}$) represent the Yukawa coupling matrices of the charged leptons, down- and up-type quarks, respectively, $g_i$ denote the gauge couplings[^6] and $\lambda$ the Higgs self-coupling in the SM. We work in the basis where $Y_e$ is diagonal. The parameters of interest are the masses, which are proportional to the eigenvalues of $\kappa$ and defined to be non-negative, as well as the mixing angles and physical phases of the MNS matrix [@Maki:1962mu] $$\begin{aligned} \label{eq:StandardParametrizationUMNS} U_\mathrm{MNS} & = & V(\theta_{12},\theta_{13},\theta_{23},\delta) \, \operatorname{diag}(e^{-\I\varphi_1/2},e^{-\I\varphi_2/2},1) \;,\end{aligned}$$ which diagonalizes $\kappa$ in this basis. $V$ is the leptonic analogon to the CKM matrix in the quark sector. The parametrization we use will be explained in more detail in App. \[app:MixingParameters\]. Currently, we learn from experiments that there occur two oscillations with mass squared differences $\Delta m^2_\mathrm{sol}$ and $\Delta m^2_\mathrm{atm}$ and corresponding mixing angles $\theta_{12}$ and $\theta_{23}$, respectively. For the third mixing angle $\theta_{13}$ and the absolute scale of light neutrino masses, there are only upper bounds at the moment (see Tab. \[tab:ExpData\] for the present status). --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $\vphantom{\sqrt{\big|}}$ Best-fit value Range (for $\theta_{ij} \in [0^\circ,45^\circ]$) C.L. ------------------------------------------------------------------- ------------------------------------------ -------------------------------------------------- --------------- $\vphantom{\sqrt{\big|}}$$\theta_{12}$ \[${\:}^\circ$\] $32.6$ $25.6 - 42.0$ $99 \% \;(3\sigma)$ $\vphantom{\sqrt{\big|}}$$\theta_{23}$ \[${\:}^\circ$\] $45.0$ $33.2 - 45.0$ $99 \% \;(3 \sigma)$ $\vphantom{\sqrt{\big|}}$$\theta_{13}$ \[${\:}^\circ$\] $-$ $0.0- 9.2$ $90 \% $ $\vphantom{\sqrt{\big|}}$$\Delta m^2_{\mathrm{sol}}$ \[eV$^2$\] $7.3 \cdot 10^{-5}$ $4\cdot 10^{-5} - 2.8\cdot 10^{-4}$ $99 \% \;(3 \sigma)$ $\vphantom{\sqrt{\big|}}$$|\Delta m^2_{\mathrm{atm}}|$ \[eV$^2$\] $2.5 \cdot 10^{-3}$ $1.2\cdot 10^{-3} - 5\cdot 10^{-3}$ $99 \% \;(3 \sigma)$ --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- : Experimental data for the neutrino mixing angles and mass squared differences. For the solar angle $\theta_{12}$ and the solar mass squared difference, the LMA solution as confirmed by KamLAND is shown. The results stem from the analysis [@deHolanda:2002iv] of the recent KamLAND and the SNO data, the Super-Kamiokande atmospheric data [@Toshito:2001dk] and the CHOOZ experiment [@Apollonio:1999ae]. []{data-label="tab:ExpData"} The Analytical Formulae {#sec:AnalyticFormulae} ----------------------- In this section, we present explicit RGEs for the physical parameters. They determine the slope of the RG evolution at a given energy scale and thus yield an insight into the RG behavior. The derivation will be discussed in App. \[sec:DerivRGEs\]. Note that a naive linear interpolation, i.e.  assuming the right-hand sides of the equations to be constant, will not always give the correct RG evolution. As we will show later, this is mainly due to large changes of $\theta_{12}$ and the mass squared differences. In the following, we will neglect ${y_e}$ and ${y_\mu}$ against $y_\tau$ and introduce the abbreviation $$\zeta \,:=\, \frac{\Delta m^2_\mathrm{sol}}{\Delta m^2_\mathrm{atm}} \;,$$ whose LMA best-fit value is about 0.03. In order to keep the expressions short, we will only show the leading terms in an expansion in the small angle $\theta_{13}$ for the mixing parameters. In almost all cases they are sufficient for understanding the features of the RG evolution.[^7] In all cases except for the running of the Dirac phase $\delta$, the limit $\theta_{13}\to0$ causes no difficulties, the subtleties arising for $\delta$ will be discussed in Sec. \[sec:RunningOfDelta\]. We furthermore define $m_i(t):=v^2\,\kappa_i(t)/4$ with $v=246\,\mathrm{GeV}$ in the SM or $v=246\,\mathrm{GeV}\cdot \sin\beta$ in the MSSM and, as usual, $\Delta m^2_\mathrm{sol} := m_2^2-m_1^2$ and $\Delta m^2_\mathrm{atm} := m_3^2-m_2^2$. Note that our formulae cannot be applied if one of the mass squared differences vanishes. For a discussion of RG effects in this case, see e.g.  [@Ellis:1999my; @Casas:1999tp; @Casas:1999ac; @Joshipura:2002xa; @Joshipura:2002gr]. With these conventions, we obtain the following analytical expressions for the mixing angles: $$\begin{aligned} \label{eq:AnalyticApproxT12} \Dot{\theta}_{12} & = & -\frac{C y_\tau^2}{32\pi^2} \, \sin 2\theta_{12} \, s_{23}^2\, \frac{ | {m_1}\, e^{\I \varphi_1} + {m_2}\, e^{\I \varphi_2}|^2 }{\Delta m^2_\mathrm{sol} } + \mathscr{O}(\theta_{13}) \;, \label{eq:Theta12Dot}\end{aligned}$$ $$\begin{aligned} \label{eq:AnalyticApproxT13} \Dot{\theta}_{13} & = & \frac{C y_\tau^2}{32\pi^2} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \frac{m_3}{\Delta m^2_\mathrm{atm} \left( 1+\zeta \right)} \times \nonumber\\ && \quad \times \left[ m_1 \cos(\varphi_1-\delta) - \left( 1+\zeta \right) m_2 \, \cos(\varphi_2-\delta) - \zeta m_3 \, \cos\delta \right] + \mathscr{O}(\theta_{13}) \;, \label{eq:Theta13Dot}\end{aligned}$$ $$\begin{aligned} \label{eq:AnalyticApproxT23} \Dot{\theta}_{23} & = & -\frac{C y_\tau^2}{32\pi^2} \, \sin 2\theta_{23} \, \frac{1}{\Delta m^2_\mathrm{atm}} \left[ c_{12}^2 \, |m_2\, e^{\I \varphi_2} + m_3|^2 + s_{12}^2 \, \frac{|m_1\, e^{\I \varphi_1} + m_3|^2}{1+\zeta} \right] \nonumber\\ & & {} + \mathscr{O}(\theta_{13}) \;. \label{eq:Theta23Dot}\end{aligned}$$ Note that in order to apply Eq.  to the case $\theta_{13}=0$, where $\delta$ is undefined, the analytic continuation of the latter, which will be given in Eq. , has to be inserted. The $\mathscr{O}(\theta_{13})$ terms in the above RGEs can become important if $\theta_{13}$ is not too small and in particular if cancellations appear in the leading terms. For example, this is the case for $|\varphi_1-\varphi_2|=\pi$ in , as we will discuss below in more detail. The RGE for the Dirac phase is given by $$\label{eq:DeltaPrimeWithNonZeroDelta} \Dot{\delta} \,=\, \frac{C y_\tau^2}{32\pi^2} \frac{\delta^{(-1)}}{\theta_{13}} +\frac{C y_\tau^2}{8\pi^2}\delta^{(0)}+\mathscr{O}(\theta_{13})\;,$$ where $$\begin{aligned} \delta^{(-1)} & = & \sin2\theta_{12}\,\sin2\theta_{23}\, \frac{m_3} {\Delta m^2_\mathrm{atm} \left(1+\zeta\right)} \times \nonumber\\ & & \quad\times \left[ m_1 \sin(\varphi_1-\delta) - \left( 1+\zeta \right) m_2\,\sin(\varphi_2-\delta) + \zeta m_3\,\sin\delta \right] \;,\\ \delta^{(0)} & = & \frac{m_1 m_2\,s_{23}^2\,\sin(\varphi_1-\varphi_2)}{\Delta m^2_\mathrm{sol}} \nonumber\\ & & {} +m_3\,s_{12}^2 \left[ \frac{m_1\,\cos2\theta_{23}\,\sin\varphi_1}{\Delta m^2_\mathrm{atm}(1+\zeta)} +\frac{m_2\,c_{23}^2\,\sin(2\delta-\varphi_2)}{\Delta m^2_\mathrm{atm}} \right] \nonumber\\ & & {} +m_3\,c_{12}^2 \left[ \frac{m_1\,c_{23}^2\,\sin(2\delta-\varphi_1)}{\Delta m^2_\mathrm{atm}(1+\zeta)} +\frac{m_2\,\cos2\theta_{23}\,\sin\varphi_2}{\Delta m^2_\mathrm{atm}} \right] \;.\end{aligned}$$ For the physical Majorana phases, we obtain $$\begin{aligned} \label{eq:AnalyticApproxP1} \Dot\varphi_1 & = & \frac{C y_\tau^2}{4\pi^2} \left\{ m_3 \, \cos 2\theta_{23} \, \frac{ m_1 s_{12}^2\, \sin\varphi_1 + \left( 1+\zeta \right) m_2 \, c_{12}^2 \, \sin\varphi_2 } { \Delta m^2_\mathrm{atm} \left( 1+\zeta \right) } \right. \nonumber\\ && \hphantom{ \frac{C y_\tau^2}{4\pi^2} \left\{ \right.} \left. {} + \frac{ m_1 m_2 \, c_{12}^2\,s_{23}^2\, \sin(\varphi_1-\varphi_2) } { \Delta m^2_\mathrm{sol} } \right\} + \mathscr{O}(\theta_{13}) \;, \label{eq:Phi1Dot} \\ \Dot\varphi_2 \label{eq:AnalyticApproxP2} & = & \frac{C y_\tau^2}{4\pi^2} \left\{ m_3 \, \cos 2\theta_{23} \, \frac{ m_1 s_{12}^2\, \sin\varphi_1 + \left( 1+\zeta \right) m_2 \, c_{12}^2 \sin\varphi_2 } { \Delta m^2_\mathrm{atm} \left( 1+\zeta \right) } \right. \nonumber\\ && \hphantom{ \frac{C y_\tau^2}{4\pi^2} \left\{ \right.} \left. {} + \frac{ m_1 m_2 \, s_{12}^2\,s_{23}^2\, \sin(\varphi_1-\varphi_2) } { \Delta m^2_\mathrm{sol} } \right\} + \mathscr{O}(\theta_{13}) \;. \label{eq:Phi2Dot}\end{aligned}$$ We would like to emphasize that the above expressions do not contain expansions in $\zeta$, i.e. their $\zeta$ dependence is exact. In many cases, they can be further simplified by neglecting $\zeta$ against 1 without losing much accuracy. Note that singularities can appear in the $\mathscr{O}(\theta_{13})$-terms at points in parameter space where the phases are not well-defined. For the masses, the results for $y_e=y_\mu=0$ but arbitrary $\theta_{13}$ are \[eq:EvolutionOfMassEigenvalues\] $$\begin{aligned} 16\pi^2\,\Dot{m}_{1} & = & \left[ \alpha + C y_\tau^2 \left( 2 s_{12}^2 \, s_{23}^2 + F_1 \right) \right] m_1 \;, \\ 16\pi^2\,\Dot{m}_2 & = & \left[ \alpha + C y_\tau^2 \left( 2 c_{12}^2 \, s_{23}^2 + F_2 \right) \right] m_2 \;, \\ 16\pi^2\,\Dot{m}_3 & = & \left[ \alpha +2 C y_\tau^2 \, c_{13}^2 \, c_{23}^2 \right] m_3 \;,\end{aligned}$$ where $F_1$ and $F_2$ contain terms proportional to $\sin\theta_{13}$, \[eq:Fi\] $$\begin{aligned} F_1 &=& -s_{13} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \cos\delta + 2 s_{13}^2 \, c_{12}^2 \, c_{23}^2 \;, \\ F_2 &=& s_{13} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \cos\delta + 2 s_{13}^2 \, s_{12}^2 \, c_{23}^2 \;.\end{aligned}$$ These formulae can be translated into RGEs for the mass squared differences, \[eq:RunningDeltaM2s\] $$\begin{aligned} \! 8\pi^2 \, \frac{\D}{\D t} \Delta m_\mathrm{sol}^2 \, & = & \! \alpha \, \Delta m_\mathrm{sol}^2 + C y_\tau^2 \left[ 2 s_{23}^2 \left( m_2^2\,c_{12}^2 - m_1^2\,s_{12}^2 \right) + F_\mathrm{sol} \right]\; , \label{eq:Dm2solDot} \\ 8\pi^2 \, \frac{\D}{\D t} \Delta m_\mathrm{atm}^2 & = & \! \alpha \, \Delta m_\mathrm{atm}^2 + C y_\tau^2 \left[ 2 m_3^2 \, c_{13}^2 \, c_{23}^2 - 2 m_2^2 \, c_{12}^2 \, s_{23}^2 + F_\mathrm{atm} \right]\; , \label{eq:Dm2atmDot}\end{aligned}$$ where \[eq:FsolAndFatm\] $$\begin{aligned} F_\mathrm{sol} & = & \left( m_1^2+m_2^2 \right) s_{13} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \cos\delta \nonumber\\ && {} + 2 s_{13}^2\, c_{23}^2 \left( m_2^2\,s_{12}^2-m_1^2\,c_{12}^2 \right) \;,\label{eq:Fsol} \\ F_\mathrm{atm} &=& -m_2^2 \,s_{13} \,\sin 2\theta_{12} \,\sin 2\theta_{23} \,\cos\delta -2 m_2^2 \, s_{13}^2 \, s_{12}^2 \, c_{23}^2 \;.\label{eq:Fatm}\end{aligned}$$ Generic Enhancement and Suppression Factors ------------------------------------------- From Eqs. – it follows that there are generic enhancement and suppression factors for the RG evolution of the mixing parameters, depending on whether the mass scheme is hierarchical, partially degenerate or nearly degenerate. We have listed these factors in the approximation of small $\theta_{13}$ in Tab. \[tab:SuppressionEnhancementFactorsNorm\]. They can be compensated by cancellations due to a special alignment of the phases. For example, an opposite CP parity of the first and second mass eigenstate, i.e.  $|\varphi_1-\varphi_2|=\pi$, results in a maximal suppression of the running of the solar mixing angle, which has been pointed out earlier in papers like [@Casas:1999tg; @Balaji:2000gd; @Haba:2000tx; @Chankowski:2001mx]. Nevertheless, Tab. \[tab:SuppressionEnhancementFactorsNorm\] allows to determine which angles or phases have a potential for a strong RG evolution. Obviously, the expressions for $\Dot\delta$ are not applicable for $\theta_{13}=0$. This special case will be discussed at the end of Sec. \[sec:RunningOfDelta\]. [|c|c|c|c|c|c|]{} & $\Dot{\theta}_{12}$ & $\Dot{\theta}_{13}$ & $\Dot{\theta}_{23}$ & $\Dot{\delta}$ & $\Dot{\varphi}_{i}$\ n.h. & 1 & $\displaystyle\sqrt{\zeta}$ & $1$ & $\displaystyle\sqrt{\zeta} \, \theta_{13}^{-1}$ & $\displaystyle\sqrt{\zeta}$\ p.d.(n.) ------------------------------------------------------------------------ & $\displaystyle\frac{m_1^2}{\Delta m^2_\mathrm{sol}}$ & $\displaystyle\frac{m_1}{\sqrt{\Delta m^2_\mathrm{atm}}}$ & $1$ & $\displaystyle\frac{m_1}{\sqrt{\Delta m^2_\mathrm{atm}}} \theta_{13}^{-1} + \frac{m_1^2}{\Delta m^2_\mathrm{sol}}$ & $\displaystyle\frac{m_1^2}{\Delta m^2_\mathrm{sol}}$\ i.h. & $\zeta^{-1}$ & $\mathscr{O}(\theta_{13})$ & $1$ & $\zeta^{-1}$ & $\displaystyle\zeta^{-1}$\ p.d.(i.) ------------------------------------------------------------------------ & $\zeta^{-1}$ & $\displaystyle\frac{m_3}{\sqrt{\Delta m^2_\mathrm{atm}}}$ & $1$ & $\displaystyle\frac{m_3}{\sqrt{\Delta m^2_\mathrm{atm}}}\theta_{13}^{-1}+\zeta^{-1}$ & $\displaystyle\zeta^{-1}$\ d. ------------------------------------------------------------------------ & $\displaystyle\frac{m^2}{\Delta m^2_\mathrm{sol}}$ & $\displaystyle\frac{m^2}{\Delta m^2_\mathrm{atm}}$ & $\displaystyle\frac{m^2}{\Delta m^2_\mathrm{atm}}$ & $\displaystyle\frac{m^2}{\Delta m^2_\mathrm{atm}}\theta_{13}^{-1} +\displaystyle\frac{m^2}{\Delta m^2_\mathrm{sol}}$ & $\displaystyle\frac{m^2}{\Delta m^2_\mathrm{sol}}$\ Let us consider some numerical values in order to estimate the size of RG effects. The SM $\tau$ Yukawa coupling is $y_\tau^\mathrm{SM}=\frac{\sqrt{2}}{v} m_\tau \approx 0.01$. Thus, the typical factor in the formulae for the mixing angles and phases amounts to $$\frac{3 y_\tau^2}{64\pi^2} \approx 0.5 \cdot 10^{-6} \;.$$ In the MSSM it changes to $$\frac{y_\tau^2}{32\pi^2} \approx 0.3 \cdot 10^{-6} \left( 1+\tan^2\beta \right) \;.$$ If the running was purely logarithmic, it would yield a factor of $$\ln \frac{M_1}{M_Z} \approx \ln \frac{10^{13}}{10^2} \approx 25$$ for $M_1=10^{13}\,\mathrm{GeV}$. If we assume that the solar and atmospheric angle are large and that the phases do not cause excessive cancellations, then multiplying the above two contributions with the enhancement factor $\Gamma_\mathrm{enh}$ from Tab. \[tab:SuppressionEnhancementFactorsNorm\] yields a rough estimate for the change of the angles and phases due to the RG evolution, $$\label{eq:NaiveRGChange} \Delta_\mathrm{RG} \sim 10^{-5} \left( 1+\tan^2\beta \right) \Gamma_\mathrm{enh} \;.$$ Of course the factor $1+\tan^2\beta$ has to be omitted in the SM. It is immediately clear that even in the MSSM with very large $\tan\beta$ no significant change occurs if the enhancement factor is 1 or less – except maybe for $\theta_{13}$, where even a change by $1^\circ$ could be interesting. However, for quasi-degenerate neutrinos large enhancement factors are possible. As an example, let us estimate the size of the absolute neutrino mass scale (the ‘amount of degeneracy’) needed for a sizable RG change of $\theta_{12}$, say $0.1 \approx 6^\circ$. In the SM, this requires $\Gamma_\mathrm{enh} \sim 10^4$, corresponding to a neutrino mass of the order of $1\,\mathrm{eV}$, which is excluded by WMAP and double beta decay experiments. On the other hand, in the MSSM this mass scale can easily be lowered to about $0.1\,\mathrm{eV}$ with $\tan\beta$ as small as 8. Discussion and Comparison with Numerical Results ------------------------------------------------ We now study in detail the running of the mixing angles and masses, in particular the influence of the phases. The RG evolution of the phases will be studied separately in Sec. \[sec:RunningPhases\]. We solve the RGEs for the neutrino mass operator and for the other parameters numerically and compare the results with those obtained from the analytical formulae of Sec. \[sec:AnalyticFormulae\]. For the numerics we follow the ‘run and diagonalize’ procedure, i.e. we first compute the running of the mass matrix and then extract the evolving mass eigenvalues and mixing parameters. The algorithm used for this is described in App. \[app:MixingParameters\]. As an example, we consider the MSSM with $\tan\beta=50$, a normal mass hierarchy for the neutrinos, $m_1=0.1\,\mathrm{eV}$ for the mass of the lightest neutrino, and a mass of about $120\,\mathrm{GeV}$ for the light Higgs. These boundary conditions are given at the electroweak scale, i.e.  we calculate the evolution from low to high energies. Below the SUSY-breaking scale, which we take to be $1.5\,\mathrm{TeV}$, we assume the SM to be valid as an effective theory and use the corresponding RGEs. Above, we apply the ones of the MSSM. ### RG Evolution of $\boldsymbol{\theta_{12}}$ {#sec:RGevolutionTheta12} From Tab. \[tab:SuppressionEnhancementFactorsNorm\], we see that the solar angle $\theta_{12}$ generically has the strongest RG effects among the mixing angles. The reason for this is the smallness of the solar mass squared difference associated with it, in particular compared to the atmospheric one, which leads to an enhanced running for quasi-degenerate neutrinos and for the case of an inverted mass hierarchy. Furthermore, it is known that in the MSSM the solar angle always increases when running down from $M_1$ for $\theta_{13}=0$ [@Miura:2002nz]. This is confirmed by our formula . From the term $| {m_1}\, e^{\I \varphi_1} + {m_2}\, e^{\I \varphi_2}|^2$ in Eq. , we see that a non-zero value of the difference $|\varphi_1-\varphi_2|$ of the Majorana phases damps the RG evolution. The damping becomes maximal if this difference equals $\pi$, which corresponds to an opposite CP parity of the mass eigenstates $m_1$ and $m_2$. This is in agreement with earlier studies, e.g.  [@Casas:1999tg; @Balaji:2000gd; @Haba:2000tx; @Chankowski:2001mx]. Let us now compare the analytical approximation for $\Dot{\theta}_{12}$ of Eq.  with the numerical solution for the running in the case of nearly degenerate masses, which is shown in Fig. \[fig:PhasesMSSMtb50\_t12\] in detail. The dark-gray region shows the evolution with LMA best-fit values for the neutrino parameters, $\theta_{13}$ varying in the interval $[0^\circ,9^\circ]$ and all CP phases equal to zero. The medium-gray regions show the evolution for $|\varphi_1-\varphi_2|\in \{0^\circ,90^\circ,180^\circ,270^\circ\}$, $\theta_{13}\in [0^\circ,9^\circ]$ and $\delta\in \{0^\circ,90^\circ,180^\circ,270^\circ\}$, confirming the expectation of the damping influence of $\varphi_1$ and $\varphi_2$. The flat line at low energy stems from the SM running below $M_\mathrm{SUSY}$, which is negligible as we have seen earlier. Note that the numerics never yield negative values of $\theta_{12}$ due to the algorithm used for extracting the mixing parameters from the MNS matrix, which guarantees $0 \leq \theta_{12} \leq 45^\circ$ (see App. \[sec:LeptMixingMatrix\] for further details). [${\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{Theta12Running.eps}}}}}$]{} As can be seen from the relatively broad dark-gray band in the figure, the $\mathscr{O}(\theta_{13})$-term in the RGE is quite important here. The dominant part of this term is $$\begin{aligned} \Upsilon & = & \frac{C y_\tau^2}{32\pi^2} \frac{m_2+m_1}{m_2-m_1} \, \sin 2\theta_{23} \, \cos\frac{\varphi_1-\varphi_2}{2} \times \nonumber\\* && \quad {} \times \Bigl( \cos 2\theta_{12} \,\cos\delta\,\cos\frac{\varphi_1-\varphi_2}{2} +\sin\delta\,\sin\frac{\varphi_1-\varphi_2}{2} \Bigr) \cdot \theta_{13} \;.\label{eq:AnalyticApproxT12OT13}\end{aligned}$$ Clearly, the RG evolution of $\theta_{12}$ is independent of the Dirac phase $\delta$ only in the approximation $\theta_{13}=0$. The largest running, where $\theta_{12}$ can even become zero, occurs for $\theta_{13}$ as large as possible ($9^\circ$), $\delta=\pi$ and $\varphi_1 - \varphi_2 = 0$. In this case the leading and the next-to-leading term add up constructively. It is also interesting to observe that due to $\mathscr{O}(\theta_{13})$ effects $\theta_{12}$ can run to slightly larger values. The damping due to the Majorana phases is maximal in this case, which almost eliminates the leading term. Then, all the running comes from the next-to-leading term . In the inverted scheme, $m_1\gg m_2-m_1$ always holds, so that large RG effects are generic, i.e.  always present except for the case of cancellations due to Majorana phases. For a normal mass hierarchy with a small $m_1$, the running of the solar mixing is of course rather insignificant. Finally, we would like to emphasize that it is not appropriate to assume the right-hand sides of Eq.  and Eq.  to be constant in order to interpolate $\theta_{12}$ up to a high energy scale, since non-linear effects especially from the running of $\sin 2\theta_{12}$ and $\Delta m^2_\mathrm{sol}$ cannot be neglected here. This is easily seen from the curved lines in Fig. \[fig:PhasesMSSMtb50\_t12\]. ### RG Evolution of $\boldsymbol{\theta_{13}}$ {#sec:RunningTheta13} The analytical approximation for $\Dot{\theta}_{13}$ is given in Eq. . As already pointed out, in order to apply it to the case $\theta_{13}=0$, where $\delta$ is undefined, the analytic continuation of the latter has to be inserted. It will be given in Eq.  in section \[sec:RunningOfDelta\], where the phases are treated in detail. The comparison with the numerical results in Fig. \[fig:PhasesMSSMtb50\_t13\] shows that above $M_\mathrm{SUSY}$ the angle runs linearly on a logarithmic scale to a good approximation. Thus, using Eq.  with a constant right-hand side yields pretty accurate results. With $\varphi_1\neq\varphi_2$, significant RG effects can be expected for nearly degenerate masses. This is confirmed by the light-gray region in Fig. \[fig:PhasesMSSMtb50\_t13\]. The fastest running occurs if $\varphi_1-\varphi_2=\pi$ and $\varphi_1-\delta \in \{0,\pi\}$, so that the terms proportional to $m_1$ and $m_2$ in the RGE are maximal and add up. Interestingly, cancellations between the first two terms in the second line of Eq.  appear for $\varphi_1=\varphi_2$, in particular if all phases are zero. If so, the leading contribution to the evolution of $\theta_{13}$ is suppressed by an additional factor of $\zeta$. This suppression is in agreement with earlier studies, for instance [@Balaji:2000gd; @Bhattacharyya:2002aq], where it was discussed for the CP-conserving case $\varphi_1=\varphi_2=\pi$, which implies an opposite CP parity of $m_3$ compared to the other two mass eigenvalues. Such cancellations cannot occur for a strong normal mass hierarchy, since then the evolution is dominated by the term proportional to $m_2$ in Eq. . Besides, $\theta_{13}$ runs towards smaller values in the MSSM with zero phases and a normal hierarchy, because $m_1<m_2$, so that the second line of the RGE is negative. This yields the dark-gray region in Fig. \[fig:PhasesMSSMtb50\_t13\].[^8] As $\theta_{13}$ can always be made positive by a suitable redefinition of parameters, the sign of $\Dot\theta_{13}$ is irrelevant for $\theta_{13}=0$. For an inverted hierarchy, the situation is reversed, since $\Delta m^2_\mathrm{atm}$ is negative then. For a small $m_3$, the running is highly suppressed in this case, because the leading term is proportional to $m_3$. Then the dominant contribution comes from the $\mathscr{O}(\theta_{13})$-term unless $\theta_{13}$ is very small as well. Future experiments will probably be able to probe $\sin^2 2\theta_{13}$ down to $10^{-4}$, corresponding to $\theta_{13} \sim 5\cdot10^{-3} \sim 0.3^\circ$. Consequently, even RG changes of this order of magnitude could be important, since a low-energy value smaller than the RG change would appear unnatural. This will be discussed in more detail in Sec. \[sec:ConstraintsFromRG\]. [${\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{Theta13and23Running.eps}}}}}$]{} ### RG Evolution of $\boldsymbol{\theta_{23}}$ The analytical RGE for $\Dot{\theta}_{23}$ can be found in Eq. . Again, the comparison with the numerical results (see Fig. \[fig:PhasesMSSMtb50\_t13\]) shows that to a good approximation the angle runs linearly on a logarithmic scale above $M_\mathrm{SUSY}$. The sign of $\Delta m^2_\mathrm{atm}$ is very important here. For a normal mass spectrum, the leading term is always negative in the MSSM, so that $\theta_{23}$ decreases with increasing energy, while for an inverse spectrum the situation is exactly reversed, so that $\theta_{23}$ becomes larger than $45^\circ$ if one starts with the LMA best-fit value at low energy. From Eq.  we expect that switching on the phases $\varphi_1$ and $\varphi_2$ always reduces the running of $\theta_{23}$ for nearly degenerate masses. This is confirmed by the light-gray region in Fig. \[fig:PhasesMSSMtb50\_t13\]. The damping is much less severe for a hierarchical mass spectrum, since either $m_1$ and $m_2$ or $m_3$ are very small then. However, in these cases the running is generally expected to be rather insignificant, since according to Tab. \[tab:SuppressionEnhancementFactorsNorm\] the enhancement factor is only 1. ### RG Evolution of the Neutrino Mass Eigenvalues {#sec:RunningOfMasses} The running of the mass eigenvalues is significant even in the SM or for strongly hierarchical neutrino masses due to the factor $\alpha$ in the RGEs . Clearly, the evolution is not directly dependent on the Majorana phases [@Casas:1999tg]. This can be understood from Eqs.  and , which show that only the moduli of the elements of the MNS matrix enter into $\Dot{m}_i$. Besides, $\Dot{m}_3$ does not depend on $\delta$, since only the moduli of the elements of the third column of the MNS matrix are relevant in this case. Of course, there is an indirect dependence on the phases, as these influence the running of the mixing angles. Apart from the MSSM with large $\tan\beta$, the running of the mass eigenvalues is virtually independent of the mixing parameters, since $\alpha$ is usually much larger than $y_\tau^2$. In the SM, the Higgs mass influences the running via the self-coupling $\lambda$ – the heavier the Higgs, the larger the RG effects. Thus, except for large $\tan \beta$ in the MSSM, the running is given by a common scaling of the mass eigenvalues [@Chankowski:2001mx], which is obtained by neglecting $y_\tau$ and integrating Eq. , $$m_i(t) \,\approx\, \exp\left[\frac{1}{16\pi^2} \int_{t_0}^t\D\tau\,\alpha(\tau)\right]\,m_i(t_0) \,=:\, s(t,t_0)\, m_i(t_0) \;.$$ We plot $s$ in the SM and in the MSSM for various parameter combinations in Fig. \[fig:ScalingOfMasses\]. The three SM curves correspond to different Higgs masses in the current experimentally allowed region at 95% confidence level, $114\,\mathrm{GeV}\lesssim m_H \lesssim200\,\mathrm{GeV}$ [@Higgsbound]. $m_H=180\,\mathrm{GeV}$ is the value for which the self-coupling $\lambda$ stays perturbative up to $10^{16}\,\mathrm{GeV}$, i.e. $\lambda\lesssim1$, and $m_H=165\,\mathrm{GeV}$ is the minimal mass for which $\lambda$ is positive up to $10^{16}\,\mathrm{GeV}$, so that the vacuum is stable in this region (see e.g. [@Cabibbo:1979ay; @Lindner:1986uk]).[^9] In the MSSM, we choose $m_H=120\,\mathrm{GeV}$ for the light Higgs mass, since the allowed range is further restricted by the upper limit at about $130\,\mathrm{GeV}$ here, and since it influences the evolution of the RG scaling only marginally as long as $M_\mathrm{SUSY}$ and $M_Z$ differ only by a few orders of magnitude. Moreover, further uncertainties due to threshold corrections and the unknown value of the SUSY-breaking scale can be equally important as the one due to the unknown Higgs mass. The RG enhancement of the masses is smallest if $\tan\beta\approx 10$. ${\ensuremath{\vcenter{\hbox{\includegraphics[scale=0.9]{ScalingOfMasses.eps}}}}}$ As already mentioned, substantial deviations from the common scaling arise in the MSSM for large $\tan\beta$. There is a plethora of effects which can be understood with the aid of and . In order to give an interesting example, we show the evolution of the mass eigenvalues for $m_\mathrm{min}=0.19\,\mathrm{eV}$ (where $m_\mathrm{min}=\min\{m_1,m_2,m_3\}$) in the MSSM with $\tan\beta=50$ in Fig. \[fig:RunningMasses\]. A particular interesting effect is that for an inverted mass spectrum the property $|\Delta m^2_\mathrm{atm}|>\Delta m^2_\mathrm{sol}$ possibly does not survive the RG evolution. In other words, what looks like a normal mass hierarchy at high energies turns out to become an inverted hierarchy at low energies (cf. Fig. \[fig:RunningMassesi\]). From the dependence on the $y_\tau^2$ terms (cf. Eqs. and ), we find that this effect can disappear if $\delta$ is large. ### RG Evolution of $\boldsymbol{\Delta m_\mathrm{sol}^2}$ {#sec:RunningDm2sol} The RGE for the solar mass squared difference is given in Eq. . In the SM and the MSSM with small $\tan\beta$, the running is due to the common scaling of the masses described in the previous section and thus virtually independent of the mixing parameters. For large $\tan\beta$ and nearly degenerate masses, the influence of CP phases, in particular the Dirac phase, is crucial. The numerical example in Fig. \[fig:PhasesMSSMtb50\_m2Dsol\] confirms this expectation and furthermore shows that $\Delta m_\mathrm{sol}^2$ runs dramatically. On the one hand, it can grow by more than an order of magnitude. As we have seen in Fig. \[fig:RunningMasses\], $\Delta m_\mathrm{sol}^2$ can even get larger than $|\Delta m_\mathrm{atm}^2|$. On the other hand, it can run to $0$ at energy scales slightly beyond the maximum of $10^{13}\,\mathrm{GeV}$ shown in the figure. For large $\tan\beta$, $\Delta m_\mathrm{sol}^2\ll m_1^2$ and not too small $\theta_{13}$, the first term in $F_\mathrm{sol}$ is essential for understanding these effects, since it is proportional to the sum of the masses squared rather than the difference. For $\delta=\pi$ and $\theta_{13}$ near the CHOOZ bound, its sign is negative and its absolute value maximal, which causes the evolution of $\Delta m_\mathrm{sol}^2$ towards zero. For $\delta=0$, the sign becomes positive, so that the running towards larger values is enhanced, which explains the upper boundary of the light-gray region in Fig. \[fig:PhasesMSSMtb50\_m2Dsol\]. ${\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{DeltaM2solRunning.eps}}}}}$ ### RG Evolution of $\boldsymbol{\Delta m_\mathrm{atm}^2}$ {#sec:RunningDm2atm} From the numerical example in Fig. \[fig:PhasesMSSMtb50\_m2Datm\], we see that $\Delta m_\mathrm{atm}^2$ can be damped by the phases, but not significantly enhanced. Depending on the CP phases, $\Delta m_\mathrm{atm}^2$ grows by about 50% – 95%. Analogously to above, the maximal damping is mainly due to the first term in $F_\mathrm{atm}$, so that it occurs for large $\theta_{13}$ and $\delta=0$. Compared to the case of the solar mass squared difference, the influence of $\delta$ is generically smaller here, because $\Delta m_\mathrm{atm}^2/m_i^2$ is larger and because the phase-independent terms in the RGE do not nearly cancel. ${\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{DeltaM2atmRunning.eps}}}}}$ RG Running of the Dirac and Majorana Phases {#sec:RunningPhases} ------------------------------------------- Most earlier studies of RG effects either neglected phases or concentrated on the special case of a Majorana parity, where one or both of the Majorana phases are $\pi$. We have seen that they can have a dramatic influence on the running of the masses and mixings. Moreover, many effects are affected by phases, e.g.  neutrinoless double beta decay, or require phases, e.g. leptogenesis.[^10] Of course, if the phases are given at some scale, they also change due to the RG evolution. We now discuss the running of the phases themselves and give numerical examples. In general, a significant evolution of the phases is expected for nearly degenerate and inverted hierarchical mass patterns, since the RGEs – contain the ratios $m_1 m_2 / \Delta m^2_\mathrm{sol}$. ### RG Evolution of the Dirac Phase {#sec:RunningOfDelta} The running of the Dirac phase $\delta$ is given by Eq.  for $y_e=y_\mu=0$. An interesting possibility is the radiative generation of a Dirac phase by Majorana phases [@Casas:1999tg]: A non-zero $\delta$ is produced by RG effects, since some of the terms in the RGE do not vanish for $\delta\rightarrow0$. Fig. \[fig:DiracPhaseEvolution5\] shows an example. The most important term in this context is the first one in $\delta^{(0)}$. As it is proportional to $\sin(\varphi_1-\varphi_2)$, the effect is suppressed for $\varphi_1=\varphi_2$. For small but non-zero values of $\theta_{13}$, the term involving $\delta^{(-1)}$ also contributes significantly because of the factor $\theta_{13}^{-1}$. For $\varphi_1=\varphi_2$, this contribution is suppressed as well, since the parts proportional to $m_1$ and $m_2$, respectively, nearly cancel. [$\vcenter{\hbox{\includegraphics[scale=1]{PhaseEvolution12.eps}}}$]{} In the case of an inverted hierarchy with $\tan\beta$ varying between 30 and 50, Dirac phases of about $15^\circ$ to $30^\circ$ can be generated. Now the term involving $\delta^{(-1)}$ receives an additional suppression from the small value of $m_3$, so that the subleading effects described above become unimportant. Hence, the running of $\delta$ is independent of $\theta_{13}$ and depends only on the difference of the Majorana phases to a very good approximation. Before we turn to the evolution of the Majorana phases, let us discuss some further properties of the RGE for $\delta$ that are also valid beyond the special case of a radiative generation of this phase. To start with, the most important term in $\Dot\delta$ depends only on the difference of the Majorana phases. Consequently, the evolution is expected to stay roughly the same if both phases change by the same value. A comparison with numerical results shows that this is true only to a first approximation. If one starts with $\varphi_2=0$ and increments it step by step, the running of $\delta$ is increasingly damped. The main reason for this is the second term in square brackets in $\delta^{(-1)}$ (the one proportional to $m_2$), whose sign is opposite to that of the leading term for $\delta<\varphi_2$. This term grows with $\varphi_2$, while the previous one (proportional to $m_1$) does not change much as long as $\varphi_1$ is close to $90^\circ$. The situation can be very different for smaller values of $\theta_{13}$. Now the initial rise of $\delta$ is enhanced, so that it can become larger than $\varphi_2$. Then the sign of the aforementioned second term in square brackets changes, so that it no longer damps the evolution but amplifies it. With a strong normal hierarchy, RG effects are usually tiny. The running of the Dirac phase is one of the few examples where this is not always the case. Due to the terms proportional to $\theta_{13}^{-1}$ in the RGE, a significant evolution is possible for small $\theta_{13}$. However, one has to keep in mind that a measurement of $\delta$ is very hard in this case. Regardless of the mass hierarchy, the limit $\theta_{13} \to 0$ is dangerous, because in this case the RGE diverges. However, we can show that $\Dot{\delta}$ remains well-defined: The derivative of the MNS matrix $U$ is given by , $\Dot{U} = U \cdot T$, where $U$ and $T$ are continuous. Hence, $U_{13}(t)$ describes a continuously differentiable curve in the complex plane. Consequently, $\theta_{13}$ and $\delta$ are continuously differentiable even for $\theta_{13}=0$, if $\delta$ is extended continuously at this point. Note that restricting the parameters to certain ranges can nevertheless result in discontinuities. For example, if the RG evolution causes $\theta_{13}$ to change its sign and if we demand $0\leq\theta_{13}<\frac{\pi}{2}$, then there will be a kink in the evolution of $\theta_{13}$ and $\delta$ will jump by $\pi$. However, even in the presence of such artificial discontinuities there must still be finite one-sided limits for $\delta$ and $\Dot{\delta}$ as $\theta_{13}$ approaches 0. The limit for $\delta$ is determined by the requirement that $\Dot{\delta}$ remains finite. Then the divergence of $\theta_{13}^{-1}$ has to be canceled by $\delta^{(-1)}$. For $\varphi_1=\varphi_2=0$, this obviously implies $\delta=0$ or $\delta=\pi$. In the general case, a short calculation yields $$\label{eq:DeltaZeroT13} \cot\delta = \frac{ m_1 \cos\varphi_1 - \left(1+\zeta\right) m_2 \,\cos\varphi_2- \zeta m_3 } { m_1 \sin\varphi_1 - \left(1+\zeta\right) m_2 \, \sin\varphi_2 } \;.$$ Due to the periodicity of $\cot$, there are two solutions differing by $\pi$, corresponding to the different limits on the two sides of a node of $\theta_{13}$. ### RG Evolution of the Majorana Phases While the RGEs for the Majorana phases are somewhat lengthy, there is a simple expression for the running of their difference for small $\theta_{13}$, $$\label{eq:DeltaPhiEvolution} \Dot\varphi_1 - \Dot\varphi_2 = \frac{C y_\tau^2}{4\pi^2} \, \frac{ m_1 m_2 }{ \Delta m^2_\mathrm{sol} } \, \cos 2\theta_{12} \,\sin^2\theta_{23} \, \sin(\varphi_1-\varphi_2) + \mathscr{O}(\theta_{13}) \;.$$ It shows that for $\theta_{13}=0$, the phases remain equal, if they are equal at some scale. Obviously, $\Dot\varphi_1-\Dot\varphi_2 > 0$ for $\varphi_1>\varphi_2$ and vice versa, which means that the difference between the phases tends to increase with increasing energy. In other words, a large difference at the see-saw scale becomes smaller at low energy. An example is shown in Fig. \[fig:MajPhaseEvolution1\]. [$\vcenter{\hbox{\includegraphics[scale=1]{PhaseEvolution2.eps}}}$]{} If $\varphi_1-\varphi_2$ is not too small, a non-zero $\theta_{13}$ tends to damp its running. This is due to a term in the RGE for $\varphi_1$ whose sign is opposite to that of the leading one in Eq.  and which is proportional to $\sin\theta_{13} \, \cot\theta_{12}$. This term can grow important if $\theta_{12}$ becomes small with increasing energy. For $\varphi_1=\varphi_2$ the evolution of the Majorana phases is suppressed, since the leading terms in the RGEs and are zero then. However, for larger $\tan\beta$ RG effects are still important. Non-linear effects caused by the decrease of the solar and atmospheric mixing angles are essential here, as the initial slope of the curves is extremely small due to the suppression by $\sin\theta_{13}$ and $\cos 2\theta_{23}$. For $\theta_{13}=5^\circ$, the second line in the RGE and the terms proportional to $\sin\theta_{13}$ are about equally important for the running of $\varphi_1$. The evolution of $\varphi_2$ is virtually independent of $\theta_{13}$, since the respective terms are not multiplied by $\cot\theta_{12}$, which again can become large as the energy increases because of the diminishing $\theta_{12}$, but by $\tan\theta_{12}$, which remains smaller than 1. In principle, it is also possible to generate Majorana phases radiatively, if the CP phase is non-zero. However, it follows from the discussion in the previous paragraph that this only happens via terms proportional to $\sin\theta_{13}$. Some Applications {#sec:Applications} ================= The discussed RG effects obviously have important implications whenever masses and mixings at different energy scales enter the analysis. Relating the Leptogenesis Parameters to Observations {#sec:Leptogenesis} ---------------------------------------------------- One of the most attractive mechanisms for explaining the observed baryon asymmetry of the universe, $\eta_B = (6.5^{+0.4}_{-0.8}) \cdot10^{-10}$ [@Spergel:2003cb], is leptogenesis [@Fukugita:1986hr]. In this scenario, $\eta_B$ is generated by the out-of-equilibrium decay of the same heavy singlet neutrinos which are responsible for the suppression of light neutrino masses in the see-saw mechanism. The masses of the heavy neutrinos are typically assumed to be some orders of magnitude below the GUT scale. Though the parameters entering the leptogenesis mechanism cannot be completely expressed in terms of low-energy neutrino mass parameters, it is possible to derive bounds on the neutrino mass scale from the requirement of a successful leptogenesis [@Buchmuller:2003gz]. Since, as we demonstrated in Sec. \[sec:RunningOfMasses\], the neutrino masses experience corrections of about 20-25% in the MSSM or more than 60% in the SM, we expect the corrections for such bounds to be sizable. The maximal baryon asymmetry generated in the thermal version of this scenario is given by [@Hamaguchi:2001gw; @Davidson:2002qv; @Buchmuller:2003gz] $$\eta_B^\mathrm{max} \,\simeq\, 0.96\cdot10^{-2}\,\varepsilon_1^\mathrm{max}\,\kappa_\mathrm{f}\;.$$ $\kappa_\mathrm{f}$ is a dilution factor which can be computed from a set of coupled Boltzmann equations (see, e.g.  [@Buchmuller:2000as]). In [@Buchmuller:2003gz], an analytic expression for the maximal relevant CP asymmetry was derived, $$\varepsilon_1^\mathrm{max} (m_1, m_3, \widetilde{m}_1) \,=\, \frac{3}{16\pi}\,\frac{M_1\,m_3}{(v/\sqrt{2})^2} \left[1-\frac{m_1}{m_3} \left(1+\frac{m_3^2-m_1^2}{\widetilde{m}_1^2}\right)^{1/2}\right] ,\label{eq:newEpsilon1max}$$ which refines the older bound $$\varepsilon_1^\mathrm{max} (m_1, m_3) \,=\, \frac{3}{16\pi}\frac{M_1}{(v/\sqrt{2})^2} \frac{\Delta m^2_\mathrm{atm}+\Delta m^2_\mathrm{sol}}{m_3} \label{eq:oldEpsilon1max}$$ and is valid for a normal mass hierarchy in the SM as well as in the MSSM.[^11] $\widetilde{m}_1$ is defined by $$\widetilde{m}_1 = \frac{(m_\mathrm{D}^\dagger m_\mathrm{D})_{11}}{M_1}$$ with $m_\mathrm{D}\sim Y_\mathrm{\nu}$ being the neutrino Dirac mass and typically lies between $m_1$ and $m_3$. It can be constrained by the requirement of successful leptogenesis because it controls the dilution of the generated asymmetry. The authors of [@Buchmuller:2003gz] introduced the ‘neutrino mass window for baryogenesis’ which corresponds to the region in the $\widetilde{m}_1$-$M_1$ plane allowing for successful thermal leptogenesis. The shape and size of the ‘mass window’ depends on $\overline{m}=\sqrt{m_1^2+m_2^2+m_3^2}$, i.e.  it becomes smaller for increasing $\overline{m}$, and $\overline{m}\ge 0.2\,\mathrm{eV}$ is not compatible with thermal leptogenesis. The calculations relevant for leptogenesis, however, refer to processes at very high energies, and therefore the RG evolution of the input parameters has to be taken into account [@Barbieri:1999ma]. The correct procedure would be to assume specific values for the neutrino mass parameters at low energy, taking into account the experimental input, evolve them to the scale $M_1$ and test the leptogenesis mechanism using these values. As the full calculation is beyond the scope of this paper, we present the evolution of the relevant mass parameters, i.e.  the light neutrino masses, to the leptogenesis scale $M_1$ and estimate the size of the error arising if RG effects are neglected. As discussed in Sec. \[sec:RunningOfMasses\], there are basically two cases which have to be distinguished, the case of the SM or the MSSM with small $\tan\beta$, and the case of the MSSM with large $\tan\beta$. In the first case, running effects can be understood to arise due to the rescaling of the light neutrino mass eigenvalues under the renormalization group. From Eq.  it is clear that the maximal CP asymmetry scales like the masses. This statement also holds for the asymmetry from Eq. , if $\widetilde{m}_1$ is a linear combination of the light mass eigenvalues. Hence, the RG yields an enhancement of the CP asymmetry of between 10% and 80%, which can be read off from Fig. \[fig:ScalingOfMasses\]. These effects are almost completely independent of the low-energy CP phases. On the other hand, the dilution factor $\kappa_\mathrm{f}$ is expected to become tiny since larger mass eigenvalues imply larger Yukawa couplings, which makes the washout more efficient. This expectation is substantiated by the fact that $\overline{m}$, which controls an important class of washout processes, also increases under the renormalization group, i.e.  it scales like the masses. As a detailed numerical calculation of the dilution factor is beyond the scope of this paper, we refer to [@Buchmuller:2000as], from which we see that in the region of interest, i.e.  the edge of the mass window, $\kappa_\mathrm{f}$ decreases exponentially. From this behavior, which is also in accordance with the analytic approximations (see, e.g.  [@Nielsen:2002pc; @DiBari:2002wi]), we expect that the neutrino mass window for baryogenesis will rather shrink than become larger when RG effects are properly taken into account. In the second case, i.e.  in the MSSM for large $\tan\beta$, we distinguish between hierarchical and degenerate mass spectra. In the hierarchical spectrum, the running of $\varepsilon_1^\mathrm{max}$ is to a high accuracy given by the running of $m_3$,[^12] so that in this case Fig. \[fig:ScalingOfMasses\] yields the relevant plot. The scaling depends on $\tan\beta$. In order to illustrate this dependence, we pick $M_1=10^{10}\,\mathrm{GeV}$ and plot $\overline{m}_\mathrm{rel}:=\overline{m}(10^{10}\,\mathrm{GeV})/ \overline{m}(M_Z)$ in Fig. \[fig:RunningMrelTanBeta\] as a function of $\tan\beta$, including small values of this parameter as well. It is clear that $\overline{m}\approx m_3$ so that Fig. \[fig:RunningMrelTanBeta\] also shows the scaling of $\varepsilon_1^\mathrm{max}$. Since $\tan\beta=10$ and $\tan\beta=50$ correspond to extreme cases, the scaling factor for different $M_1$ can be read off from Fig. \[fig:ScalingOfMasses\] by interpolation. In the case of a quasi-degenerate mass spectrum (and large $\tan\beta$), the CP asymmetry can run stronger than the average mass scale because, as we already have seen in Sec. \[sec:RunningDm2sol\] and \[sec:RunningDm2atm\], the mass squared differences can experience a stronger RG enhancement than the squares of the mass eigenvalues. We show the evolution of $\varepsilon_\mathrm{rel}:=\varepsilon_1^\mathrm{max}(10^{10}\,\mathrm{GeV})/ \varepsilon_1^\mathrm{max}(M_Z)$ in Fig. \[RunningEps\]. To produce this plot, we employed and inserted the running mass parameters. For this combination of parameters, the low-energy phases do influence the evolution of $\varepsilon_\mathrm{rel}$ by damping its running, and the plot shows the maximal evolution, which means that the phases are simply set to zero. The running effects are even larger for the new bound , since it is more sensitive to the mass splittings than the old one. More precisely, for highly degenerate mass spectra it is much smaller than the old one and the degeneracy can be lifted by running effects. This strong enhancement of the CP asymmetry may even overcompensate the decrease of the dilution factor for large $\tan\beta$, so that the parameter region compatible with thermal leptogenesis grows. Altogether, we have presented the relevant mass parameters at the scale of leptogenesis, thus making it convenient to take into account RG effects in future studies. Moreover, we have estimated the impact of the renormalization effects, and found that there are two effects in opposite directions: The CP asymmetry is enhanced because the mass squared differences increase, and the dilution of the baryon asymmetry is more effective since the overall mass scale rises due to RG effects. As the dependence of the dilution factor on the mass scale is stronger than that of the CP asymmetry, we expect the mass window for baryogenesis to shrink when RG effects are included in the analysis. An exception is the case of large $\tan\beta$, where the situation is more complicated. Note also that there exist different, non-thermal baryogenesis mechanisms [@Kumekawa:1994gx] in which the masses of the light neutrinos may be almost degenerate [@Fujii:2002jw]. In these kinds of scenarios, RG effects increase the baryon asymmetry, since $\varepsilon_1$ increases, while the effects from the expected decrease of the dilution factor do not occur. RG Evolution of Bounds on the Neutrino Mass Scale ------------------------------------------------- The absolute neutrino mass scale at low energy is restricted by low-energy experiments such as searches for $0\nu\beta\beta$ decay and cosmological observations. As usual, the RG evolution of the results has to be taken into account in order to translate the experimental results into constraints on high-energy theories. ### Neutrinoless Double Beta Decay The amplitude of $0\nu\beta\beta$ decay is proportional to the effective neutrino mass $$\begin{aligned} \Braket{m_\nu} & = & (m_\nu)_{11} \ = \ \Bigl| \sum_i U_{1i}^2 \, m_i \Bigr| \nonumber\\ & = & \left| m_1 \, c_{12}^2 c_{13}^2 \, e^{\I \varphi_1} + m_2 \, s_{12}^2 c_{13}^2 \, e^{\I \varphi_2} + m_3 \, s_{13}^2 \, e^{2\I \delta} \right| \;,\end{aligned}$$ where $U$ is the MNS matrix. Instead of inserting the lengthy RGEs for all the quantities in the second line in order to calculate the RG evolution of $\Braket{m_\nu}$, it is much more convenient to use Eq. , which directly yields $$16\pi^2 \: \frac{\D}{\D t}\Braket{m_\nu} \,=\, \left( 2 C\, y_e^2 + \alpha \right) \Braket{m_\nu} \;.$$ As the first term is negligible, the RG change of the effective neutrino mass is basically caused by the universal rescaling of the neutrino masses alone. It is completely independent of the other neutrino mass parameters, since neither the running of $y_e$ nor that of the terms in $\alpha$ is sensitive to them. Besides, the value of $\tan\beta$ is not very important here, because $y_e^2$ is always tiny and $\alpha$ contains only the up-type quark Yukawa couplings in the MSSM. However, there is a dependence on the Higgs mass in the SM. Currently, the best experimental upper limit on the effective neutrino mass is about $\Braket{m_\nu}<0.35\,\mathrm{eV}$ [@Klapdor-Kleingrothaus:2000sn; @Aalseth:2002rf], with some uncertainty due to nuclear matrix elements. Fig. \[fig:EffectiveMassEvolution1\] shows the running of this limit in the SM and the MSSM. As it is very close to the best-fit value of the recently claimed evidence for double beta decay, $\Braket{m_\nu}=0.39\,\mathrm{eV}$ [@Klapdor-Kleingrothaus:2001ke], the evolution of the latter is nearly identical. The SM plot contains three curves corresponding to different Higgs masses in the current experimentally allowed region. In the MSSM, the light Higgs mass is chosen to be about $120\,\mathrm{GeV}$. The running is much more significant in the SM than in the MSSM because of the contribution of the Higgs self-coupling. ### WMAP Bound Combining the observations of the cosmic microwave background by the WMAP satellite with other astronomical data allows to place an upper bound of about $0.7\,\mathrm{eV}$ onto the sum of the light neutrino masses [@Spergel:2003cb]. This implies $$m_i \lesssim 0.23\, \text{eV}$$ for each mass eigenvalue. Analogous to the limit from $0\nu\beta\beta$ decay in the previous section, this bound is modified substantially by the RG evolution. This is shown in Fig. \[fig:WMAPEvolution\] for the eigenvalue $m_3$. As discussed in Sec. \[sec:RunningOfMasses\], the running of the mass eigenvalues is not sensitive to the mixing parameters in the SM, but it depends on the Higgs mass. In the MSSM, the variation of the phases causes a slight modification of the running, but its order of magnitude is only a few percent even for the large $\tan\beta$ used in the plot. The influence of $\theta_{13}$ is negligible. Interestingly, the evolution of the sum of the mass eigenvalues is virtually independent of the mixing parameters for nearly degenerate neutrinos both in the SM and in the MSSM. This can be explained by considering the sum of the RGEs . For $m_1 \sim m_2 \sim m_3$, the terms proportional to $y_\tau^2$ add up to 1, with small corrections of the order of $\frac{\Delta m^2_\mathrm{atm}}{m^2}$ and $\theta_{13}$. Constraints on Neutrino Properties from RG Effects {#sec:ConstraintsFromRG} -------------------------------------------------- One may wonder if deviations from $\theta_{13}=0$ and $\theta_{23}=\pi/4$ exist which are the consequence of radiative corrections. Let us assume therefore that $\theta_{13}=0$ or $\theta_{23}=\pi/4$ are given by some high-energy model. Low-energy deviations from the exact values are then RG effects, which can be compared to the sensitivities of future experiments. Therefore we investigate in a model-independent way the size of RG corrections to $\theta_{13}$ and $\theta_{23}$ from the running of the effective neutrino mass operator between the see-saw scale and the electroweak scale. ### Corrections to $\boldsymbol{\theta_{13}}$ As pointed out in Sec. \[sec:RunningTheta13\], it is a rather good approximation to assume $\Dot{\theta}_{13}\simeq$ const. in Eq. , which leads to an RG evolution with a constant slope depending on the Dirac CP phase $\delta$ and the Majorana phases $\varphi_1$ and $\varphi_2$. Therefore, let us first apply the naive estimate explicitly to the change of $\theta_{13}$ in the MSSM for nearly degenerate neutrinos. In this case, the enhancement factor $m^2/\Delta m^2_\mathrm{atm}$ leads to a generic change of $\theta_{13}$ under the RG that exceeds the detection limit of future experiments even for moderate values of $\tan\beta$. For example, $m_1=0.1\,\mathrm{eV}$ and $\tan\beta=30$ yield a change in $\sin^2 2\theta_{13}$ of $\Delta\!\sin^2 2\theta_{13} \sim 0.5\cdot10^{-2}$, which is further enhanced by a factor of 4 if the Majorana phases are aligned properly. In order to obtain a more detailed picture, we now apply Eq.  to calculate the RG correction to the initial value $\theta_{13}=0$ between some high energy scale $M_1$, where neutrino masses are generated, and low energy, i.e. $10^2\,\mathrm{GeV}$. In this case the initial value of the Dirac phase $\delta$ is determined by the analytic continuation Eq. . For the examples we take $M_1 = 10^{12}\,\mathrm{GeV}$. The approximate size of the RG corrections to $\sin^2 2\theta_{13}$ in the MSSM is shown in Fig. \[fig:RGCorrt13\]. In the upper diagram it is plotted as a function of $\tan\beta$ and the lightest neutrino mass $m_1$ for constant Majorana phases $\varphi_1=0$ and $\varphi_2=\pi$. The lower diagram shows the dependence of the corrections on $\varphi_1$ and $\varphi_2$ for $\tan\beta=50$ and $m_1=0.08\,\mathrm{eV}$ in the case of a normal mass hierarchy. The diagrams look rather similar for an inverted hierarchy. Analytically, the pattern of the upper plot is easy to understand, and for the lower one there is a simple explanation as well. Consider partially or nearly degenerate neutrino masses. Then Eq.  yields to a reasonably good approximation $$\begin{aligned} \label{eq:T13DotApprox} \Dot{\theta}_{13} & \approx & \frac{C y_\tau^2}{32\pi^2} \, \sin 2\theta_{12} \, \sin 2\theta_{23} \, \frac{m^2}{\Delta m^2_\mathrm{atm}} \left[ \cos(\varphi_1-\delta) - \cos(\varphi_2-\delta) \right] \nonumber\\* & \propto & \sin\frac{\varphi_1+\varphi_2-2\delta}{2} \, \sin\frac{\varphi_1-\varphi_2}{2} \;.\end{aligned}$$ Applying an analogous approximation to Eq. , it can easily be shown that the first term in the second line is always $\pm 1$, so that the running is completely determined by the difference of the Majorana phases. This leads to the diagonal bands in Fig. \[fig:RGCorrt13\_Phases\], in particular the white one corresponding to $\varphi_1-\varphi_2=0$. If one starts with a small but non-zero $\theta_{13}$, which allows an arbitrary $\delta$, it turns out that the RG evolution quickly drives $\delta$ to a value satisfying Eq. , so that the final pattern of Fig. \[fig:RGCorrt13\_Phases\] is unchanged. Planned reactor experiments [@Huber:2003pm] and next generation superbeam experiments [@Huber:2002rs; @Minakata:2003ca] are expected to have an approximate sensitivity on $\sin^2 2\theta_{13}$ of $10^{-2}$. From Fig. \[fig:RGCorrt13\] we find that the radiative corrections exceed this value for large regions of the currently allowed parameter space, unless there are cancellations due to Majorana phases, i.e.  $\varphi_1\approx\varphi_2$ (which might be due to some symmetry). If so, the effects are generically smaller than $10^{-2}$ as can be seen from the lower diagram. Future upgraded superbeam experiments like JHF-HyperKamiokande have the potential to further push the sensitivity to about $10^{-3}$ and with a neutrino factory even about $10^{-4}$ might be reached. From the theoretical point of view, one would expect that even if some model predicted $\theta_{13}=0$ at the energy scale of neutrino mass generation, RG effects would at least produce a non-zero value of the order shown in Fig. \[fig:RGCorrt13\]. Consequently, experiments with such a sensitivity have a large discovery potential for $\theta_{13}$. We should point out that this is a conservative estimate, since if neutrino masses are e.g. determined by GUT scale physics, model-dependent radiative corrections in the region between $M_1$ and $M_\mathrm{GUT}$ contribute as well [@Casas:1999tp; @Casas:1999ac; @King:2000hk; @Antusch:2002rr; @Antusch:2002hy; @Antusch:2002fr] and there can be additional corrections from physics above the GUT scale [@Vissani:2003aj]. On the other hand, if experiments do not measure $\theta_{13}$, this will improve the upper bound on $\theta_{13}$. Parameter space regions where the corrections are larger than this bound will then appear unnatural from the theoretical side. [${\ensuremath{\vcenter{\hbox{\includegraphics[scale=0.75]{RGCorrectionTheta13.eps}}}}}$]{}\ ### Corrections to $\boldsymbol{\theta_{23}}$ We now consider the RG corrections which induce a deviation of $\theta_{23}$ from $\pi/4$, even if some model predicted this specific value at high energy. We apply the analytical formula with a constant right-hand side in order to calculate the running in the MSSM between $M_Z$ and the see-saw scale, which we take as $M_1 = 10^{12}\,\mathrm{GeV}$ for our examples. As initial conditions we assume small $\theta_{13}$ at $M_1$ and low-energy best-fit values for the remaining lepton mixings and the neutrino mass squared differences. In leading order in $\theta_{13}$, the evolution is of course independent of the Dirac phase $\delta$. The size of the RG corrections in the MSSM is shown in Fig. \[fig:RGCorrt23\]. From the upper diagram it can be read off for desired values of $\tan\beta$ and the lightest mass eigenvalue $m_1$ in an example with vanishing Majorana phases. The lower diagram shows its dependence on the Majorana phases $\varphi_1$ and $\varphi_2$ for $\tan\beta=50$, $m_1=0.1\,\mathrm{eV}$ and a normal mass hierarchy. The diagrams look rather similar in the case of an inverted hierarchy. The effects of the Majorana phases can easily be understood from Eq. . In the region with $\varphi_1 \approx \varphi_2 \approx \pi$ (again, this might be, e.g., due to some symmetry), both $|m_2\, e^{\I \varphi_2} + m_3|^2$ and $|m_1\, e^{\I \varphi_1} + m_3|^2$ are small for quasi-degenerate neutrinos, which gives the ellipse with small radiative corrections in the center of the lower diagram. Such cancellations are not possible with hierarchical masses, but the RG effects are generally not very large in this case, as shown by the upper plot. Even if a model predicted $\theta_{23}=\pi/4$ at some high energy scale, we would thus expect radiative corrections to produce at least a deviation from this value of the size shown in Fig. \[fig:RGCorrt23\], so that experiments with such a sensitivity are expected to measure a deviation of $\theta_{23}$ from $\pi/4$. The sensitivity to $\sin^2 2\theta_{23}$ of future superbeam experiments like JHF-SuperKamiokande is expected to be approximately 1% (see e.g. [@Itow:2001ee]). This can now be compared with Fig. \[fig:RGCorrt23\]. We find that the radiative corrections exceed this value for large regions of the currently allowed parameter space, where no significant cancellations due to Majorana phases occur. This means that $\varphi_1$ and $\varphi_2$ must not be too close to $\pi$. Otherwise, the effects are generically smaller as can be seen from the lower diagram. Upgraded superbeam experiments or a neutrino factory might even reach a sensitivity of about $0.5$%. As argued for the case of $\theta_{13}$, if experiments measure $\theta_{23}$ rather close to $\pi/4$, parameter combinations implying larger radiative corrections than the measured deviation will appear unnatural from the theoretical point of view. [${\ensuremath{\vcenter{\hbox{\includegraphics[scale=0.75]{RGCorrectionTheta23.eps}}}}}$]{}\ Conclusions {#sec:Conclusions} =========== We have derived compact expressions which allow an analytical understanding of the running of neutrino masses, leptonic mixing angles and CP phases in the SM and MSSM. The results are given directly in terms of these quantities as well as gauge and Yukawa couplings, and especially for a small angle $\theta_{13}$ the expressions become very simple, even when non-vanishing CP phases are present. We have extensively compared those formulae to numerical results and we have found that the RG evolution of the physical parameters is described qualitatively, and to a reasonable accuracy also quantitatively, very well. We have shown that Dirac and Majorana CP phases can have a drastic influence on the RG evolution of the mixing parameters. We have reproduced and illustrated some effects that were previously described in the literature. As a particularly interesting example, we have discussed the radiative generation of the Dirac phase from the Majorana phases. Besides, we have derived new results, for example concerning the running of the CP phases. Even though the RG effects for the mixing parameters in the SM are rather small, the RG effects for the masses are not, and have to be taken into account in any careful analysis which relates high and low energy scales. In the MSSM, especially for large $\tan\beta$, the evolution of the mixings and phases can be large. The RG evolution has interesting phenomenological implications. In the case of leptogenesis, we have estimated the corrections which arise if the running is appropriately taken into account and found that the mass window for baryogenesis is likely to shrink when those corrections are considered. In order to simplify the inclusion of RG effects in future calculations, we provide the relevant information of the mass parameters at the leptogenesis scale. Furthermore, we investigated the extrapolation of the upper bounds on the neutrino mass scale from $0\nu\beta\beta$ decay experiments and WMAP to higher energy scales, where they become restrictions for model building. Experimentally one finds $\theta_{23}\simeq \pi/4$, $\theta_{13}\simeq 0$. The deviations from $\pi/4$ and zero may have a radiative origin and we calculated therefore in a model-independent analysis the RG corrections to $\theta_{23}=\pi/4$, $\theta_{13}=0$. With future precision experiments this may lead to interesting insights into model parameters. To conclude, we have obtained analytic formulae which are a useful tool to understand the RG corrections, relevant whenever parameters at two different energy scales are compared. This has been demonstrated in the phenomenological applications. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported in part by the “Sonderforschungsbereich 375 für Astro-Teilchenphysik der Deutschen Forschungsgemeinschaft”. We would like to thank W. Buchmüller, K. Hamaguchi, P. Huber, R. N. Mohapatra and W. Winter for interesting discussions. Appendix {#appendix .unnumbered} ======== Definition and Extraction of Mixing Parameters {#app:MixingParameters} ============================================== Standard Parametrization ------------------------ In this section we describe our conventions and how mixing angles and phases can be extracted from mass matrices. For a general unitary matrix we choose the so-called standard-parametrization $$\begin{aligned} \label{eq:StandardParametrizationU} U & = &\operatorname{diag}(e^{\I\delta_{e}},e^{\I\delta_{\mu}},e^{\I\delta_{\tau}}) \cdot V \cdot \operatorname{diag}(e^{-\I\varphi_1/2},e^{-\I\varphi_2/2},1)\end{aligned}$$ where $$V=\left( \begin{array}{ccc} c_{12}c_{13} & s_{12}c_{13} & s_{13}e^{-\I\delta}\\ -c_{23}s_{12}-s_{23}s_{13}c_{12}e^{\I\delta} & c_{23}c_{12}-s_{23}s_{13}s_{12}e^{\I\delta} & s_{23}c_{13}\\ s_{23}s_{12}-c_{23}s_{13}c_{12}e^{\I\delta} & -s_{23}c_{12}-c_{23}s_{13}s_{12}e^{\I\delta} & c_{23}c_{13} \end{array} \right)$$ with $c_{ij}$ and $s_{ij}$ defined as $\cos\theta_{ij}$ and $\sin\theta_{ij}$, respectively. Extracting Mixing Angles and Phases {#sec:ExtractingMixingAngles} ----------------------------------- In this standard-parametrization, the mixing angles $\theta_{13}$ and $\theta_{23}$ can be chosen to lie between $0$ and $\frac{\pi}{2}$, and by reordering the masses, $\theta_{12}$ can be restricted to $0\le\theta_{12}\le\frac{\pi}{4}$. For the phases the range between $0$ and $2\pi$ is required. In order to read off the mixing parameters, we use the following procedure: 1. $\theta_{13}=\arcsin(|U_{13}|)$. 2. $\displaystyle \theta_{12}=\left\{\begin{array}{ll} \displaystyle \arctan\left(\frac{|U_{12}|}{|U_{11}|}\right) \quad & \text{if}\;U_{11}\ne0\\ \frac{\pi}{2} & \text{else} \end{array}\right.$ 3. $\displaystyle \theta_{23}=\left\{\begin{array}{ll} \displaystyle \arctan\left(\frac{|U_{23}|}{|U_{33}|}\right) \quad & \text{if}\;U_{33}\ne0\\ \frac{\pi}{2} & \text{else} \end{array}\right.$ 4. $\delta_\mu = \arg(U_{23})$ 5. $\delta_\tau = \arg(U_{33})$ 6. \[step6\]$\displaystyle\delta= -\arg\left(\frac{\displaystyle\frac{U_{ii}^*U_{ij}U_{ji}U_{jj}^*} {c_{12}\,c_{13}^2\,c_{23}\,s_{13}} +c_{12}\,c_{23}\,s_{13}} {s_{12}\,s_{23}}\right)$\ where $i,j\in\{1,2,3\}$ and $i\ne j$. 7. $\delta_e=\arg(e^{\I\delta}\,U_{13})$ 8. $\displaystyle\varphi_1=2\arg(e^{\I\delta_e}\,U_{11}^*)$ 9. \[step9\]$\displaystyle\varphi_2=2\arg(e^{\I\delta_e}\,U_{12}^*)$ Here we used the relation $$\begin{aligned} U_{ii}^*U_{ij}U_{ji}U_{jj}^* & = & c_{12}\,c_{13}^2\, c_{23}\,s_{13} \left(e^{-\I\delta}\,s_{12}\,s_{23} - c_{12}\,c_{23}\,s_{13}\right) \;,\nonumber\end{aligned}$$ which holds for $i,j\in\{1,2,3\}$ and $i\ne j$. Note that this relation is often used in order to introduce the Jarlskog invariants [@Jarlskog:1985ht] $$\begin{aligned} J_\mathrm{CP} & = & \frac{1}{2} \left| \operatorname{Im}(U_{11}^*U_{12}U_{21}U_{22}^*)\right| \,=\, \frac{1}{2} \left| \operatorname{Im}(U_{11}^*U_{13}U_{31}U_{33}^*)\right| \nonumber\\ & = & \frac{1}{2} \left| \operatorname{Im}(U_{22}^*U_{23}U_{32}U_{33}^*)\right| \, = \, \frac{1}{2}\left|c_{12}\,c_{13}^2\, c_{23}\,\sin \delta \, s_{12}\,s_{13}\, s_{23}\right|\;.\end{aligned}$$ For the sake of a better numerical stability, one can choose any of the three combinations. In particular, if the modulus of one of the $U_{ij}$ is very small, it turns out to be more accurate to choose a combination in which this specific $U_{ij}$ does not appear. Leptonic Mixing Matrix {#sec:LeptMixingMatrix} ---------------------- Since the effective neutrino mass matrix is symmetric, it can be diagonalized by a unitary matrix $U_\nu$, $$U_\nu^T\, m_\nu\, U_\nu\,=\, \operatorname{diag}(m_1,m_2,m_3) \;.$$ The form of $U$ depends on a prescription how to order the mass eigenvalues. In order to obtain a mixing matrix which can be compared with the experimental data, the choice of the prescription is somewhat subtle. From experiment we know that there is a small mass difference, called $\Delta m^2_\mathrm{sol}=m_i^2-m_j^2$, and a larger one, referred to as $\Delta m^2_\mathrm{atm}=m_k^2 - m_\ell^2$. By convention, the masses are labeled such that $i,j\ne 3$ while either $k$ or $\ell$ equals 3. The different schemes are depicted in Fig. \[fig:NeutrinoMassSchemes\]. [\[subfig:NeutrinoMassScheme1\] $\begin{array}{c}{\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{NormalHierarchy.eps}}}}}\\[1cm]\phantom{\text{M\"ummel}} \end{array}$ ]{} [\[subfig:NeutrinoMassScheme2\] $\begin{array}{c}{\ensuremath{\vcenter{\hbox{\includegraphics[scale=1]{InvertedHierarchy.eps}}}}}\\[1cm]\phantom{\text{M\"ummel}} \end{array}$ ]{} The mass label 2 is attached to the eigenvector with the lower modulus of the first component. We are doing this since we want to read off a mixing angle $\theta_{12}$ less then $45^\circ$. The neutrino mixing matrix $U_\mathrm{MNS}$ can then be read off in the following way: 1. Diagonalize $Y_e^\dagger Y_e$ by $U_e$, i.e.  $Y_e\to U_e^\dagger \cdot Y_e^\dagger\cdot Y_e\cdot U_e =\operatorname{diag}\left(y_e^2,y_\mu^2,y_\tau^2\right)$ where $y_f^2$ are positive for $f\in\{e,\mu,\tau\}$. 2. Change the basis according to $m_\nu\to m_\nu'=U_e^T \cdot m_\nu \cdot U_e$. 3. Diagonalize $m_\nu'$: $m_\nu'\to U_\mathrm{MNS}^T\cdot m_\nu'\cdot U_\mathrm{MNS}=\operatorname{diag}(m_1,m_2,m_3)$ where $m_i>0$. Then $U_\mathrm{MNS}$ contains the leptonic mixing angles which can be read off as described in Sec. \[sec:ExtractingMixingAngles\]. Note that $m_1<m_2<m_3$ is not necessarily fulfilled, as we already mentioned before (cf. Fig. \[fig:NeutrinoMassSchemes\]). Derivation of the Analytical Formulae {#sec:DerivRGEs} ===================================== To derive the RGEs for the mixing parameters, we follow in general the methods of [@Babu:1987im]. The RGE for $\kappa$ reads $$\label{eq:GeneralRGEforKappa} 16\pi^2\,\frac{\D\kappa}{\D t} \,=\, \alpha\,\kappa + P^T\,\kappa + \kappa\, P\;,$$ where all terms with trivial flavour structure are absorbed in $\alpha$. $\kappa$ can be diagonalized (in the basis where $Y_e$ is diagonal) by a unitary transformation, $$U(t)^T\,\kappa(t)\, U(t) \,=\, D(t) \,=\, \frac{4}{v^2} \,\operatorname{diag}\big(m_1(t),m_2(t),m_3(t)\big)\;.$$ We hence obtain $$\begin{aligned} \lefteqn{ \frac{\D}{\D t}\left(U^*\, D\, U^\dagger\right) =\Dot U^*\, D \, U^\dagger + U^*\, D \, \Dot U^\dagger +U^*\, \Dot D\, U^\dagger } \nonumber\\ & \stackrel{\eqref{eq:GeneralRGEforKappa}}{=} & \frac{1}{16\pi^2}\, \left(\alpha\,U^*\, D\, U^\dagger +P^T\, U^*\, D \, U^\dagger +U^*\, D\, U^\dagger \, P\right) \;.\end{aligned}$$ Multiplying with $U^T$ from the left and with $U$ from the right yields $$U^T\, \Dot U^* \, D +D\, \Dot U^\dagger \, U + \Dot D \,=\, \frac{1}{16\pi^2} \left[\alpha\,D+P^{\prime\,T}\, D + D\, P'\right] \;,$$ where we have introduced $P'=U^\dagger\, P\, U$. The next step is defining an anti-Hermitian matrix $T$ by $$\label{eq:EvolutionOfU} \frac{\D}{\D t} U \,=\, U\, T\;.$$ With this definition, we find $$\label{eq:MasterEquationForDTP} \Dot D \,=\, \frac{1}{16\pi^2}\, \left(\alpha\,D + P^{\prime\,T}\, D + D\, P'\right) -T^*\, D+ D\, T \;,$$ where the anti-hermiticity of $T$ was used. Since the left-hand side of this equation is diagonal and real per definition, the right-hand side has to possess these properties as well, $$\label{eq:DotMi1} \Dot m_i \,=\, \frac{1}{16\pi^2}\, \left(\alpha\, m_i+ 2\,P'_{ii}\,m_i\right) +(T_{ii}-T_{ii}^*)\,m_i \;.$$ Note that here and in the following equations, no sum over repeated indices is implied. The second bracket is purely imaginary, hence it has to cancel with the imaginary part of the first one, $$\label{eq:ImDiagT} 2\,\operatorname{Im}T_{ii} \,=\, \frac{-1}{16\pi^2}\,(\operatorname{Im}\alpha +2\operatorname{Im}P'_{ii}) \;,$$ and we further confirm eq. (15) of [@Casas:1999tg], which translates with our conventions to $$\label{eq:EvolutionOfEffectiveMasses} 16\pi^2\,\Dot m_i \,=\, \left(\operatorname{Re}\alpha+2\,\operatorname{Re}P'_{ii}\right)\,m_i \;.$$ Eq.  differs from Eq. (19) of [@Casas:1999tg], where the imaginary part of $\alpha$ is not present; however, this difference is irrelevant in the SM and the MSSM, where $\alpha$ is real. By comparing the off-diagonal parts of we find $$\label{eq:TPprimeIntermediate} m_i\,T_{ij}-T_{ij}^*\,m_j \,=\, -\frac{1}{16\pi^2}\, \left(P^{\prime\,T}_{ij}\,m_j+m_i\,P'_{ij}\right) \;.$$ Adding and subtracting this equation and its complex conjugate, we obtain for $i\ne j$ \[eq:ExpressTbyPprime\] $$\begin{aligned} 16\pi^2\,\operatorname{Re}T_{ij} & = & -\frac{m_j\,\operatorname{Re}P'_{ji}+m_i\,\operatorname{Re}P'_{ij}}{m_i-m_j} \;, \\ 16\pi^2\,\operatorname{Im}T_{ij} & = & -\frac{m_j\,\operatorname{Im}P'_{ji}+m_i\,\operatorname{Im}P'_{ij}}{m_i+m_j} \;. \label{eq:ImTijBelow}\end{aligned}$$ Let us now focus on Hermitian $P$, which implies Hermitian $P'$, for a moment. Using $\operatorname{Re}P'_{ji}=\operatorname{Re}P^{\prime\,*}_{ij} =\operatorname{Re}P'_{ij}$ and an analogous relation for $\operatorname{Im}P'_{ij}$, we obtain in this case $$\begin{aligned} 16\pi^2\,\operatorname{Im}T_{ij} & = & -\frac{m_i-m_j}{m_i+m_j}\,\operatorname{Im}P'_{ij} \;, \\ 16\pi^2\,\operatorname{Re}T_{ij} & = & -\frac{m_i+m_j}{m_i-m_j}\,\operatorname{Re}P'_{ij} \;.\end{aligned}$$ In order to obtain the renormalization group equations for the mixing angles, we use , $$\label{eq:EvolutionOfU2} U^\dagger\, \Dot U = T\;.$$ Inserting the standard parametrization , we can express the left-hand side of in terms of the mixing parameters and their derivatives. Now we can solve for the derivatives of the mixing parameters. Note that due to the separation of the evolution of the mass eigenvalues in equation , we have reduced the number of parameters from 12 to 9. The discussion so far has been very similar to the one of [@Casas:1999tg]. There, the RG evolution of the mixing parameters is expressed in terms of the mixing matrix elements and $P'$. In order to obtain rather short and more explicit formulae, which are e.g. useful for deriving the approximations of Sec. \[sec:AnalyticFormulae\], we now consider and label the mixing parameters as $$\{\xi_k\} = \{\theta_{12},\theta_{13},\theta_{23}, \delta,\delta_e,\delta_\mu,\delta_\tau, \varphi_1,\varphi_2\} \;.$$ We observe that the left-hand side of is linear in $\Dot\xi_k$. Therefore, by solving the corresponding system of linear equations, we can express the derivatives of the mixing parameters by the mixing parameters, the mass eigenvalues and the Yukawa couplings. The resulting formulae are still too long to be presented here but can be obtained from the web page `http://www.ph.tum.de/~mratz/AnalyticFormulae/`. Finally, let us record that only the moduli of $U_{ij}$ enter into the diagonal elements of $P'$, if $P$ is diagonal, $P = \operatorname{diag}(P_1,P_2,P_3)$ (which is the case in the SM and MSSM in the basis we have used in the main part), since $$\label{eq:PiiPrimeReal} P_{ii}' = \sum_{jk} (U^\dagger)_{ij} P_{jk} U_{ki} = \sum_{jk} U^*_{ji} P_j \delta_{jk} U_{ki} = \sum_j |U_{ji}|^2 P_j \;.$$ Consequently, the evolution of the mass eigenvalues does not directly depend on the Majorana phases, as claimed in Sec. \[sec:RunningOfMasses\]. [10]{} T. Yanagida, in *Proceedings of the Workshop on the Unified Theory and the Baryon Number in the Universe* (O. Sawada and A. Sugamoto, eds.), KEK, Tsukuba, Japan, 1979, p. 95. S. L. Glashow, *The future of elementary particle physics*, in *Proceedings of the 1979 Carg[è]{}se Summer Institute on Quarks and Leptons* (M. L[é]{}vy et al., eds.), Plenum Press, New York, 1980, pp. 687–713. M. Gell-Mann, P. Ramond, and R. Slansky, *Complex spinors and unified theories*, in *Supergravity* (P. van Nieuwenhuizen and D. Z. Freedman, eds.), North Holland, Amsterdam, 1979, p. 315. R. N. Mohapatra and G. Senjanovi[ć]{}, *Neutrino mass and spontaneous parity violation*, Phys. Rev. Lett. **44** (1980), 912. M. Fukugita and T. Yanagida, *Baryogenesis without grand unification*, Phys. Lett. **174B** (1986), 45. M. Tanimoto, *Renormalization effect on large neutrino flavor mixing in the minimal supersymmetric standard model*, Phys. Lett. **B360** (1995), 41–46, `hep-ph/9508247`. J. R. Ellis and S. Lola, *Can neutrinos be degenerate in mass?*, Phys. Lett. **B458** (1999), 310–321, `hep-ph/9904279`. J. A. Casas, J. R. Espinosa, A. Ibarra, and I. Navarro, *Naturalness of nearly degenerate neutrinos*, Nucl. Phys. **B556** (1999), 3–22, `hep-ph/9904395`. J. A. Casas, J. R. Espinosa, A. Ibarra, and I. Navarro, *Nearly degenerate neutrinos, supersymmetry and radiative corrections*, Nucl. Phys. **B569** (2000), 82–106, `hep-ph/9905381`. P. H. Chankowski, W. Krolikowski, and S. Pokorski, *Fixed points in the evolution of neutrino mixings*, Phys. Lett. **B473** (2000), 109, `hep-ph/9910231`. J. A. Casas, J. R. Espinosa, A. Ibarra, and I. Navarro, *General [RG]{} [E]{}quations for [P]{}hysical [N]{}eutrino [P]{}arameters and their [P]{}henomenological [I]{}mplications*, Nucl. Phys. **B573** (2000), 652, `hep-ph/9910420`. K. R. S. Balaji, A. S. Dighe, R. N. Mohapatra, and M. K. Parida, *Radiative magnification of neutrino mixings and a natural explanation of the neutrino anomalies*, Phys. Lett. **B481** (2000), 33–38, `hep-ph/0002177`. N. Haba, Y. Matsui, and N. Okamura, *The effects of [M]{}ajorana phases in three-generation neutrinos*, Eur. Phys. J. **C17** (2000), 513–520, `hep-ph/0005075`. T. Miura, E. Takasugi, and M. Yoshimura, *Quantum effects for the neutrino mixing matrix in the democratic-type model*, Prog. Theor. Phys. **104** (2000), 1173–1187, `hep-ph/0007066`. P. H. Chankowski, A. Ioannisian, S. Pokorski, and J. W. F. Valle, *Neutrino unification*, Phys. Rev. Lett. **86** (2001), 3488–3491, `hep-ph/0011150`. M.-C. Chen and K. T. Mahanthappa, *Implications of the renormalization group equations in three neutrino models with two-fold degeneracy*, Int. J. Mod. Phys. **A16** (2001), 3923–3930, `hep-ph/0102215`. P. H. Chankowski and S. Pokorski, *Quantum corrections to neutrino masses and mixing angles*, Int. J. Mod. Phys. **A17** (2002), 575–614, `hep-ph/0110249`. M. K. Parida, C. R. Das, and G. Rajasekaran, *Radiative stability of neutrino-mass textures*, (2002), `hep-ph/0203097`. G. Dutta, *Stable bimaximal neutrino mixing pattern*, (2002), `hep-ph/0203222`. T. Miura, T. Shindou, and E. Takasugi, *Exploring the neutrino mass matrix at [M(R)]{} scale*, Phys. Rev. **D66** (2002), 093002, `hep-ph/0206207`. G. Bhattacharyya, A. Raychaudhuri, and A. Sil, *Can radiative magnification of mixing angles occur for two-zero neutrino mass matrix textures?*, (2002), `hep-ph/0211074`. A. S. Joshipura, S. D. Rindani, and N. N. Singh, *Predictive framework with a pair of degenerate neutrinos at a high scale*, Nucl. Phys. **B660** (2003), 362–372, `hep-ph/0211378`. M. Frigerio and A. Yu. Smirnov, *Radiative corrections to neutrino mass matrix in the standard model and beyond*, JHEP **02** (2003), 004, `hep-ph/0212263`. R. N. Mohapatra, M. K. Parida, and G. Rajasekaran, *High scale mixing unification and large neutrino mixing angles*, (2003), `hep-ph/0301234`. K. Dick, M. Freund, M. Lindner, and A. Romanino, *[CP]{}-violation in neutrino oscillations*, Nucl. Phys. **B562** (1999), 29–56, `hep-ph/9903308`. M. Frigerio and A. Yu. Smirnov, *Structure of neutrino mass matrix and [CP]{} violation*, Nucl. Phys. **B640** (2002), 233–282, `hep-ph/0202247`. M. Frigerio and A. Yu. Smirnov, *Neutrino mass matrix: Inverted hierarchy and [CP]{} violation*, Phys. Rev. **D67** (2003), 013007, `hep-ph/0207366`. E. Ma, *Pathways to naturally small neutrino masses*, Phys. Rev. Lett. **81** (1998), 1171–1174, `hep-ph/9805219`. P. H. Chankowski and Z. Pluciennik, *Renormalization group equations for seesaw neutrino masses*, Phys. Lett. **B316** (1993), 312–317, `hep-ph/9306333`. K. S. Babu, C. N. Leung, and J. Pantaleone, *Renormalization of the neutrino mass operator*, Phys. Lett. **B319** (1993), 191–198, `hep-ph/9309223`. S. Antusch, M. Drees, J. Kersten, M. Lindner, and M. Ratz, *Neutrino mass operator renormalization revisited*, Phys. Lett. **B519** (2001), 238–242, `hep-ph/0108005`. S. Antusch, M. Drees, J. Kersten, M. Lindner, and M. Ratz, *Neutrino mass operator renormalization in two [Higgs]{} doublet models and the [MSSM]{}*, Phys. Lett. **B525** (2002), 130–134, `hep-ph/0110366`. S. Antusch and M. Ratz, *Supergraph techniques and two-loop beta-functions for renormalizable and non-renormalizable operators*, JHEP **07** (2002), 059, `hep-ph/0203027`. Z. Maki, M. Nakagawa, and S. Sakata, *Remarks on the unified model of elementary particles*, Prog. Theor. Phys. **28** (1962), 870. P. C. de Holanda and A. Yu. Smirnov, *[LMA]{} [MSW]{} solution of the solar neutrino problem and first [KamLAND]{} results*, JCAP **0302** (2003), 001, `hep-ph/0212270`. Super[K]{}amiokande, T. Toshito, *Super-[K]{}amiokande atmospheric neutrino results*, (2001), `hep-ex/0105023`. CHOOZ, M. Apollonio et al., *Limits on neutrino oscillations from the [CHOOZ]{} experiment*, Phys. Lett. **B466** (1999), 415–430, `hep-ex/9907037`. A. S. Joshipura and S. D. Rindani, *Radiatively generated $\nu_e$ oscillations: [G]{}eneral analysis, textures and models*, Phys. Rev. **D67** (2003), 073009, `hep-ph/0211404`. K. R. S. Balaji, A. S. Dighe, R. N. Mohapatra, and M. K. Parida, *Generation of large flavor mixing from radiative corrections*, Phys. Rev. Lett. **84** (2000), 5034–5037, `hep-ph/0001310`. The [LEP]{} Electroweak Working Group, D. Abbaneo et al., *A combination of preliminary electroweak measurements and constraints on the [S]{}tandard [M]{}odel* (2003), `http://lepewwg.web.cern.ch/LEPEWWG/`. N. Cabibbo, L. Maiani, G. Parisi, and R. Petronzio, *Bounds on the fermions and Higgs boson masses in Grand Unified Theories*, Nucl. Phys. **B158** (1979), 295. M. Lindner, *Implications of triviality for the Standard Model*, Zeit. Phys. **C31** (1986), 295. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, *Dynamical electroweak symmetry breaking by a neutrino condensate*, Nucl. Phys. **B658** (2003), 203–216, `hep-ph/0211385`. G. C. Branco, T. Morozumi, B. M. Nobre, and M. N. Rebelo, *A bridge between CP violation at low energies and leptogenesis*, Nucl. Phys. [**B617**]{} (2001), 475, `hep-ph/0107164`. S. Pascoli, S. T. Petcov, and W. Rodejohann, *On the connection of leptogenesis with low energy CP violation and LFV charged lepton decays*, `hep-ph/0302054`. P. H. Frampton, S. L. Glashow, and T. Yanagida, *Cosmological sign of neutrino [CP]{} violation*, Phys. Lett. **B548** (2002), 119–121, `hep-ph/0208157`. D. N. Spergel et al., *First year [W]{}ilkinson [M]{}icrowave [A]{}nisotropy [P]{}robe [(WMAP)]{} observations: [D]{}etermination of cosmological parameters*, (2003), `astro-ph/0302209`. W. Buchm[ü]{}ller, P. Di Bari, and M. Pl[ü]{}macher, *The neutrino mass window for baryogenesis*, (2003), `hep-ph/0302092`. K. Hamaguchi, H. Murayama, and T. Yanagida, *Leptogenesis from sneutrino-dominated early universe*, Phys. Rev. **D65** (2002), 043512, `hep-ph/0109030`. S. Davidson and A. Ibarra, *A lower bound on the right-handed neutrino mass from leptogenesis*, Phys. Lett. **B535** (2002), 25–32, `hep-ph/0202239`. W. Buchm[ü]{}ller and M. Pl[ü]{}macher, *Neutrino masses and the baryon asymmetry*, Int. J. Mod. Phys. **A15** (2000), 5047–5086, `hep-ph/0007176`. R. Barbieri, P. Creminelli, A. Strumia, and N. Tetradis, *Baryogenesis through leptogenesis*, Nucl. Phys. **B575** (2000), 61–77, `hep-ph/9911315`. H. B. Nielsen and Y. Takanishi, *Baryogenesis via lepton number violation and family replicated gauge group*, Nucl. Phys. **B636** (2002), 305–337, `hep-ph/0204027`. P. Di Bari, *News on leptogenesis*, AIP Conf. Proc. **655** (2003), 208–219, `hep-ph/0211175`. K. Kumekawa, T. Moroi, and T. Yanagida, *Flat potential for inflaton with a discrete [R]{} invariance in supergravity*, Prog. Theor. Phys. **92** (1994), 437–448, `hep-ph/9405337`. M. Fujii, K. Hamaguchi, and T. Yanagida, *Leptogenesis with almost degenerate [M]{}ajorana neutrinos*, Phys. Rev. **D65** (2002), 115012, `hep-ph/0202210`. H. V. Klapdor-Kleingrothaus et al., *Latest results from the [H]{}eidelberg-[M]{}oscow double-beta-decay experiment*, Eur. Phys. J. **A12** (2001), 147–154, `hep-ph/0103062`. 16EX Collaboration, C. E. Aalseth et al., *The [IGEX Ge-76]{} neutrinoless double-beta decay experiment: [P]{}rospects for next generation experiments*, Phys. Rev. **D65** (2002), 092007, `hep-ex/0202026`. H. V. Klapdor-Kleingrothaus, A. Dietz, H. L. Harney, and I. V. Krivosheina, *Evidence for neutrinoless double beta decay*, Mod. Phys. Lett. **A16** (2001), 2409–2420, `hep-ph/0201231`. P. Huber, M. Lindner, T. Schwetz, and W. Winter, *Reactor neutrino experiments compared to superbeams*, (2003), `hep-ph/0303232`. P. Huber, M. Lindner, and W. Winter, *Synergies between the first-generation [JHF-SK]{} and [NuMI]{} superbeam experiments*, Nucl. Phys. **B654** (2003), 3–29, `hep-ph/0211300`. H. Minakata, H. Nunokawa, and S. Parke, *The complementarity of eastern and western hemisphere long- baseline neutrino oscillation experiments*, (2003), `hep-ph/0301210`. S. F. King and N. N. Singh, *Renormalisation group analysis of single right-handed neutrino dominance*, Nucl. Phys. **B591** (2000), 3–25, `hep-ph/0006229`. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, *Neutrino mass matrix running for non-degenerate see-saw scales*, Phys. Lett. **B538** (2002), 87–95, `hep-ph/0203233`. S. Antusch, J. Kersten, M. Lindner, and M. Ratz, *The [LMA]{} [S]{}olution from [B]{}imaximal [L]{}epton [M]{}ixing at the [GUT]{} [S]{}cale by [R]{}enormalization [G]{}roup [R]{}unning*, Phys. Lett. **B544** (2002), 1–10, `hep-ph/0206078`. S. Antusch and M. Ratz, *Radiative generation of the [LMA]{} solution from small solar neutrino mixing at the [GUT]{} scale*, JHEP **11** (2002), 010, `hep-ph/0208136`. F. Vissani, M. Narayan, and V. Berezinsky, *U(e3) from physics above the [GUT]{} scale*, (2003), `hep-ph/0305233`. Y. Itow et al., *The [JHF-Kamioka]{} neutrino project*, in *Proceedings of the 3rd Workshop on Neutrino Oscillations and their Origin (NOON 2001)* (Y. Suzuki et al., eds.), World Scientific, Singapore, 2003, p. 239, `hep-ex/0106019`. C. Jarlskog, *Commutator of the quark mass matrices in the standard electroweak model and a measure of maximal [CP]{} violation*, Phys. Rev. Lett. **55** (1985), 1039. K. S. Babu, *[R]{}enormalization-[G]{}roup [A]{}nalysis of the [K]{}obayashi-[M]{}askawa [M]{}atrix*, Z. Phys. **C35** (1987), 69. [^1]: E-mail: `santusch@ph.tum.de` [^2]: E-mail: `jkersten@ph.tum.de` [^3]: E-mail: `lindner@ph.tum.de` [^4]: E-mail: `mratz@mail.desy.de` [^5]: In the MSSM, the RGE is known at two-loop [@Antusch:2002ek]. In this study, we will, however, focus on the one-loop equation. [^6]: We are using GUT charge normalization for $g_1$. [^7]: The exact formulae, from which we have derived the analytical approximations presented here, can be obtained from the web page `http://www.ph.tum.de/~mratz/AnalyticFormulae/`. [^8]: The relatively large slope of its upper boundary is due to the $\mathscr{O}(\theta_{13})$ contribution to the RGE. [^9]: In some models (see, e.g.  [@Antusch:2002xh] for a viable model) $\lambda$ can be larger, in particular if $M_1\ll 10^{16}\,\mathrm{GeV}$. A negative value of $\lambda$ at high energy implies a metastable vacuum. [^10]: Clearly, the phases relevant for leptogenesis are those of the ‘right-handed’ sector and therefore in general not directly related to the phases considered here [@Branco:2001pq; @Pascoli:2003uh]. However, as the left-handed sector with its – in principle – observable phases is related to the right-handed one by the see-saw relation, it is reasonable to assume that non-vanishing right-handed phases imply non-zero $\delta$, $\varphi_1$ and/or $\varphi_2$. An explicit relation which supports this point of view is specified in, e.g., [@Frampton:2002qc]. [^11]: To use these formulae in our conventions for the inverted scheme, one would have to replace $(m_1,m_2,m_3)\to(m_3,m_1,m_2)$. [^12]: For an inverted hierarchy, $m_1$ has to be used instead, whose evolution is approximately the same as that of $m_3$ here.

--- abstract: 'The magnetorotational instability (MRI) may dominate outward transport of angular momentum in accretion disks, allowing the disk material to fall onto the central object. Previous work has established that the MRI can drive a mean-field dynamo, which could lead to a self-sustaining accretion system. Recently, however, simulations of the scaling of the angular momentum transport parameter ${\ensuremath{\alpha_{SS}}}$ with the magnetic Prandtl number ${\ensuremath{\mathrm{Pm}}}$ have cast doubt on the ability of the MRI to transport astrophysically relevant amounts of angular momentum in real disk systems. Here, we use simulations including explicit physical viscosity and resistivity to show that when vertical stratification is taken into account, mean field dynamo action operates, driving the system to a configuration in which the magnetic field is not fully helical. This relaxes the strict constraints on the generated field provided by magnetic helicity conservation. Our models demonstrate the existence of a critical magnetic Reynolds number ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$, below which transport becomes strongly ${\ensuremath{\mathrm{Pm}}}$-dependent and chaotic, but above which the transport is steady and ${\ensuremath{\mathrm{Pm}}}$-independent. Scaling to realistic astrophysical parameters suggests that accretion disks around both protostars and stellar mass black holes have magnetic Reynolds numbers that easily exceed ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$. Thus, we suggest that the strong ${\ensuremath{\mathrm{Pm}}}$ dependence seen in recent simulations is not applicable to real systems.' author: - 'Jeffrey S. Oishi' - 'Mordecai-Mark Mac Low' bibliography: - 'bibliography.bib' title: Magnetorotational turbulence transports angular momentum in stratified disks with low magnetic Prandtl number but magnetic Reynolds number above a critical value --- Introduction ============ Accretion disks form in a wide variety of astrophysical objects, from active galactic nuclei to dwarf novae and protostars. In the former two, disks are responsible for the tremendous luminosities, while in the latter they are the site of planet formation. Understanding the structure and evolution of disks is therefore critical to understanding each of these objects. The successful model of accretion disks relies on viscous transport of angular momentum outward through the disk, allowing mass to spiral inwards. The amount of viscosity necessary to explain observations requires turbulence, as molecular viscosity is far too small. parameterized the turbulent stress providing this viscosity by ${\ensuremath{\alpha_{SS}}}= (\left< \rho u_x u_y\right> - \left<B_x B_y\right>)/P_0$ where $P_0$ is the midplane pressure of the disk. The magnetorotational instability (MRI) offers the most viable mechanism for providing these enhanced levels of angular momentum transport [@1998RvMP...70....1B]. It is known to saturate in a magnetohydrodynamic turbulent state that transports angular momentum outward through the disk. Aside from its central role in accretion disk theory, the MRI also provides an interesting system in which to study mean field dynamo theory. Starting from a zero-net flux magnetic field configuration, the MRI acts as a mean field dynamo, generating strong, fluctuating fields with order at the size of the simulation box [@1995ApJ...446..741B; @1996ApJ...464..690H]. The system, fed by the free energy from Keplerian shear, can sustain angular momentum transport with only a weak (sub-equipartition) field and the presence of a negative radial gradient in angular velocity. Because the turbulence is driven by the MRI itself, the Lorentz force is essential to the operation of the dynamo, ensuring that the system is never in a kinematic phase [@1996ApJ...464..690H]. In recent years, the MRI has become the target of intense numerical investigation owing to recent studies showing a decline of the angular momentum transport rate with increasing resolution in the simplest 3D systems that demonstrate MRI turbulence: unstratified shearing boxes. @2007MNRAS.378.1471L and showed that this can be understood physically as a rather steep power-law dependence of transport on the magnetic Prandtl number, ${\ensuremath{\mathrm{Pm}}}\equiv \nu/\eta$, the ratio of viscous momentum diffusion to resistive magnetic diffusion. Given that real astrophysical disks, especially protoplanetary disks, are at extremely low ${\ensuremath{\mathrm{Pm}}}\sim 10^{-9}$ [@2008ApJ...674..408B], this would imply that the MRI cannot be responsible for angular momentum transport in such systems. Subsequently, several authors have proposed solutions to this problem. These include the idea that unstratified shearing boxes lack a characteristic outer scale [@2009ApJ...696.1021V]; that the initial magnetic field strength in the simulations is too weak and they are thus stable to non-axisymmetric MRI modes that are essential to sustained transport ; or that an unstratified shearing box with periodic boundary conditions cannot sustain large scale dynamo action, while small-scale dynamo action is known to be ${\ensuremath{\mathrm{Pm}}}$-dependent, thus rendering the magnetic field necessary for the MRI susceptible to similar ${\ensuremath{\mathrm{Pm}}}$-dependence [@2010arXiv1004.2417K]. A pair of recent papers has demonstrated that even without explicit viscosity and resistivity, *stratified* MRI simulations *do* converge to a consistent value of $\alpha_{SS}$ with increasing resolution [@2010ApJ...713...52D; @2010ApJ...708.1716S]. In this paper, we broaden the examination of the stratified MRI to consider the viscous and magnetic Reynolds numbers ${\ensuremath{\mathrm{Re}}}$ and ${\ensuremath{\mathrm{Rm}}}$. We find evidence for the existence of a critical magnetic Reynolds number that may represent a boundary for sustained MRI dynamo activity. We note two dynamo behaviors, one corresponding to a sustained, organized dynamo, and the other corresponding to a transient, chaotic dynamo. We then attempt to connect the properties of the stratified MRI to other, better studied dynamo systems at high and low ${\ensuremath{\mathrm{Pm}}}$. This study is most similar to the work of [@2010arXiv1004.2417K], in that we consider the effects of boundary conditions on the transport of magnetic helicity and its relation to the generation of large-scale magnetic fields. However, we consider the effect of varying ${\ensuremath{\mathrm{Pm}}}$ on stratified shearing boxes. Furthermore, we attempt to reconcile the convergence in the simulations of @2010ApJ...713...52D [@2010ApJ...708.1716S] with the fact that they are periodic and thus cannot support the same type of flux-transport dynamo action posited by @2010arXiv1004.2417K. It is our conclusion that, despite the presence of a large scale dynamo, the large scale field is not helicity limited, though the MRI does eject helicity through open boundaries if they are present. The dynamo period does not show a discernible dependence on ${\ensuremath{\mathrm{Pm}}}$. Most importantly, we find that the angular momentum transport coefficient appears independent of ${\ensuremath{\mathrm{Pm}}}$ *as long as the magnetic Reynolds number ${\ensuremath{\mathrm{Rm}}}$ remains above a critical value, ${\ensuremath{\mathrm{Rm}}}_{crit} \sim 3000$* . In § \[s:methods\], we review our methods. We present the main results in § \[s:results\], followed by a discussion of the importance of magnetic helicity conservation on these results in § \[s:discussion\]. We conclude and note several avenues for future work in  § \[s:conclusions\]. Methods {#s:methods} ======= Using the shearing box formalism , we study a stratified patch of a Keplerian accretion disk threaded by an initial magnetic field of the form $\mathbf{B_0} = B_0 \sin(x) \mathrm{\mathbf{e_z}}$ with maximum midplane value of plasma $\beta = 2 c_s^2/v_A^2 \simeq 89$ ($B_0 = 0.15$). We solve the equations of isothermal, compressible magnetohydrodynamics using the Pencil Code[^1] [@2002CoPhC.147..471B; @2009ApJ...697.1269J], a spatially sixth-order, temporally third-order finite difference method. The constraint $\mathbf{\nabla \cdot B} = 0$ is enforced by solving the evolution equation for the magnetic vector potential, $$\label{e:induction} \frac{\partial \mathbf{A}}{\partial t} + u_y^0 \frac{\partial \mathbf{A}}{\partial y} = \mathbf{u \times B} + \frac{3}{2} \Omega A_y \mathbf{\hat{e}_x+ \eta \nabla^2 A + \eta_3 \nabla^6 A} + \mathbf{\nabla} {\ensuremath{\phi_{\mathrm{kep}}}},$$ where $\eta_3 \mathbf{\nabla^6 A}$ is a hyperdiffusion operator we apply to all dynamical equations in order to dissipate excess energy at the grid scale, and the second terms on each side are from the shearing box formalism. In order to study the effects of ${\ensuremath{\mathrm{Pm}}}$ on the saturated MRI turbulence, we vary the viscosity $\nu$ and resistivity $\eta$. We define ${\ensuremath{\mathrm{Re}}}\equiv S H^2/\nu$ and ${\ensuremath{\mathrm{Rm}}}\equiv S H^2/\eta$, where $S = \Omega d \log \Omega/ d \log r = q \Omega = -3/2$ is the (Keplerian) shear rate in the box in units of rotation $\Omega$ and $H^2 = c_s^2 / \Omega^2$ is the scale height of the disk. We set $c_s = H = \Omega = \mu_0 = 1$, choosing our units to minimize the number of values we need to remember. In these units, ${\ensuremath{\mathrm{Re}}}= 1.5/\nu$ and ${\ensuremath{\mathrm{Rm}}}= 1.5/ \eta$. The advantage of defining ${\ensuremath{\mathrm{Re}}}$ and ${\ensuremath{\mathrm{Rm}}}$ this way is that they can be set *a priori*, though they do not measure the relative effects of advection and dissipation in the saturated state of the MRI. Because our ${\ensuremath{\mathrm{Re}}}$ and ${\ensuremath{\mathrm{Rm}}}$ are *a priori* parameters, we have checked that they correlate with the ratio of turbulent advection to dissipation, ${\ensuremath{\mathrm{Re}}}' = u_{rms}/k_1 \nu$ and ${\ensuremath{\mathrm{Rm}}}' = u_{rms}/k_1 \eta$, respectively. Here, $k_1 = 2\pi/L_z$ is the largest integer wavenumber in the box. Note that $k = k_1/2$ is consistent with the VF boundary conditions, and large-scale fields of this size can also fit within the box. This data is presented in table \[t:runs\]. Indeed, in figure \[f:re\_rm\_corr\] ${\ensuremath{\mathrm{Rm}}}$ is particularly well correlated with ${\ensuremath{\mathrm{Rm}}}'$. Table \[t:runs\] summarizes the parameters of our models. We report resolution in terms of zones per scale height, with a standard resolution of $64\ \mathrm{zones/H}$, though we also ran three simulations with $128\ \mathrm{zones/H}$ to confirm the convergence of our results. All of our runs use cubic zones, $dx = dy = dz$. Our standard box size is $1 \mathrm{H} \times 4 \mathrm{H} \times 4 \mathrm{H}$. Our simulations are periodic in $y$ (azimuthal), shearing periodic in $x$ (radial), and one of three different choices for $z$ (axial): periodic, perfect conductor, or vertical field (hereafter VF). Among these choices, the first does not allow a flux of magnetic helicity out of the simulation domain, while the others do. This has a significant effect on the resulting dynamo action and turbulence, though not nearly as dramatic as in the unstratified results of @2010arXiv1004.2417K. [lllllllll]{} \[t:runs\] ![(left) Reynolds number parameter ${\ensuremath{\mathrm{Re}}}$ versus the ratio of advection to viscous dissipation in the simulations. (right) magnetic Reynolds number parameter ${\ensuremath{\mathrm{Rm}}}$ versus the ratio of advection to resistive dissipation in the simulations. \[f:re\_rm\_corr\]](re_rm_vs_reprime_rmprime.pdf){width="0.95\columnwidth"} Results {#s:results} ======= In order to place our results in the context of other recent studies, we begin by comparing our results to the unstratified results presented by @2010arXiv1004.2417K. Figure \[f:alpha\_vs\_pm\] shows that the dependence of $\alpha_{SS}$ on ${\ensuremath{\mathrm{Pm}}}$ is not well described by a single power-law at a given ${\ensuremath{\mathrm{Re}}}$. The points are the mean $\alpha_{SS}$ for all times $t > 20 t_{orb}$, and the error bars represent the standard deviation over that range. These runs all use the VF boundary conditions. The figure shows evidence of a cut-off that moves to higher ${\ensuremath{\mathrm{Pm}}}$ as ${\ensuremath{\mathrm{Re}}}$ decreases. This is indicative of a critical ${\ensuremath{\mathrm{Rm}}}$, most visible for ${\ensuremath{\mathrm{Re}}}= 3200$ (center left) and ${\ensuremath{\mathrm{Re}}}= 9600$ (lower left). Above ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$, it appears that ${\ensuremath{\alpha_{SS}}}$ is consistent with a constant value. We recast these results in the $({\ensuremath{\mathrm{Rm}}}, {\ensuremath{\mathrm{Re}}})$ plane in figure \[f:re\_rm\_alpha\_grid\]. This figure makes clear the fact that along with a reduced but still-present ${\ensuremath{\mathrm{Pm}}}$-dependence, there appears to be a fairly clear critical ${\ensuremath{\mathrm{Rm}}}$: transport is significantly reduced left of the vertical line near ${\ensuremath{\mathrm{Rm}}}\sim 3000$. As the simulations approach ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$, the $\alpha_{SS}$ variance increases significantly. ![Volume and time averaged value of the angular momentum transport parameter ${\ensuremath{\alpha_{SS}}}$ as a function of magnetic Prandtl number ${\ensuremath{\mathrm{Pm}}}$. Each subfigure is plotted on the same axes, but shows models with different Reynolds number ${\ensuremath{\mathrm{Re}}}$. Angular momentum transport increases with ${\ensuremath{\mathrm{Pm}}}$ up to a threshold value of ${\ensuremath{\mathrm{Pm}}}$ beyond which it remains constant. This threshold occurs at lower ${\ensuremath{\mathrm{Pm}}}$ for increasing ${\ensuremath{\mathrm{Re}}}$, suggesting a constant threshold at some critical magnetic Reynolds number ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$. The superposed points in the ${\ensuremath{\mathrm{Re}}}= 6400$, ${\ensuremath{\mathrm{Re}}}=9600$, and ${\ensuremath{\mathrm{Re}}}= 12800$ series represent resolution studies with one point representing a model with 128 zones/$H$, double the standard resolution.[]{data-label="f:alpha_vs_pm"}](pm_sat_alpha_grid_2.pdf){width="0.95\columnwidth"} ![The log of $\alpha_{SS}$ as a function of both viscous and magnetic Reynolds numbers ${\ensuremath{\mathrm{Rm}}}$ and ${\ensuremath{\mathrm{Re}}}$. Darker shades represent weaker angular momentum transport. Below a critical magnetic Reynolds number ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$ slightly above $10^3$, transport falls off significantly.[]{data-label="f:re_rm_alpha_grid"}](re_rm_alpha_grid.pdf){width="0.95\columnwidth"} Two Dynamo Behaviors -------------------- Figure \[f:bym\_bxm\_vs\_t\] shows that the MRI appears to operate in two distinct dynamo states, one in which regular $\left< B_y \right>$ cycles appear with a characteristic period of $\tau_B \sim 10 t_{orb}$, and one with irregular variations with timescales on the order of $\sim 50 t_{orb}$. The key control parameter appears to be ${\ensuremath{\mathrm{Rm}}}$, as the second, third, and fourth panels from the top show the regular behavior despite being at different ${\ensuremath{\mathrm{Pm}}}$. There is no discernible trend in cycle period with ${\ensuremath{\mathrm{Pm}}}$, though the top panel, with ${\ensuremath{\mathrm{Rm}}}= 12800, {\ensuremath{\mathrm{Pm}}}=2$ seems to show an intermediate behavior. However, all runs with ${\ensuremath{\mathrm{Rm}}}\lesssim 3200$ conclusively show the irregular behavior. There appears to be a secondary ${\ensuremath{\mathrm{Pm}}}$ effect as well, since the two ${\ensuremath{\mathrm{Rm}}}= 3200$ models show different behavior depending on ${\ensuremath{\mathrm{Pm}}}$ (third and fifth panels from the top). {width="75.00000%"} These two drastically different behaviors complicate efforts to ascertain scaling properties of the MRI as a function of dimensionless parameters. The range of available ${\ensuremath{\mathrm{Pm}}}$ is limited by numerical resolution both from above and below: large values of ${\ensuremath{\mathrm{Pm}}}$ imply a viscous scale much larger than the resistive one, while small ${\ensuremath{\mathrm{Pm}}}$ implies the opposite. Since both length scales must fit on (and be resolved by) the grid, and indeed neither can be so large as to stabilize the largest linear MRI modes that fit in the simulation box, our parameter range is necessarily limited. Furthermore, since low ${\ensuremath{\mathrm{Rm}}}$ runs transition to a very different behavior with much larger cycle period and very different turbulent properties in a discontinuous fashion, our range of available ${\ensuremath{\mathrm{Pm}}}$ space for the regular dynamo mode is further limited. Thus, in what follows, we briefly describe the irregular dynamo before focusing on connecting the details of the MRI turbulence in the regular region of parameter space to better-established results on small and large-scale dynamo action. The MRI is known to produce both small and large scale dynamo action, the latter typically requiring stratification or open boundary conditions . The small-scale dynamo for driven, isotropic, homogeneous, incompressible turbulence has ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$ that *increases* with *decreasing* ${\ensuremath{\mathrm{Pm}}}$ [@2005ApJ...625L.115S], and thus it behooves us to understand how MRI turbulence fits into this picture. Irregular Regime ---------------- When ${\ensuremath{\mathrm{Rm}}}$ drops below ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$, the mean field dynamo switches to a irregular regime, showing quasi-periodic magnetic cycles, with cycle lengths (one hesitates to call them periods) $\sim 40 t_{orb}$ (see the lower two panels of figure \[f:bym\_bxm\_vs\_t\]). In this regime, the turbulence often appears only in one half-plane, either $z > 0$ or $z < 0$, sometimes staying that way for nearly the duration of the simulation. This kind of behavior at first appears unphysical, as the stratified shearing box system, while having odd parity about the midplane, should not necessarily damp perturbations of one helicity more than the other. Indeed, in all simulations, the kinetic helicity shows erratic fluctuations of sign with respect to the midplane. However, there are signs that this is a physical effect. First, it is causal, and easily explainable in a simple $\alpha - \Omega$ treatment. The only prerequisite for this explanation is that the two half-planes be dynamically decoupled from one another. For the case of VF boundary conditions, $B_y(z = \pm L_z) \mapsto 0$. Reasoning spectrally, the largest wavenumber modes compatible with this boundary condition are $B_y \propto \sin(2 \pi/L_z z)$ and $B_y \propto \sin(\pi/L_z z)$, corresponding to $k= 1$ and $1/2$, respectively. If during a cycle period, the MRI turbulence shuts off, the $\alpha$ effect will cease, leaving a mean induction equation that looks like $$\partial_t \left<B_x\right> = \eta \nabla^2 \left<B_x\right>$$ for the $x$ component and $$\partial_t \left< B_y \right> = -q \Omega \left< B_x \right> +\eta \nabla^2 \left< B_y \right>$$ for the $y$ component. Because the decay time for modes is roughly $t_{decay} \simeq 1/k^2 \eta$, small scale structure will undergo selective decay, leaving only the largest scale modes, the decay time of which is $\sim 108 t_{orb}$ for ${\ensuremath{\mathrm{Rm}}}= 1600$. At this stage, $\left< B_y \right>$ has about this much time to grow linearly via stretching of any residual $\left< B_x \right>$ to amplitudes at which non-axisymmetric MRI can reestablish fluid turbulence and hence an $\alpha$ effect. If the two half-planes are not strongly coupled, it is possible that the turbulence will die out in one half of the box first, leading to an $\alpha$ effect only in the upper or lower midplane. This scenario is essentially the same as demonstrated in the midplane by @2010arXiv1010.0005S. Dynamo coefficients ------------------- Understanding the origin of the scaling of angular momentum transport with magnetic Prandtl number requires understanding the underlying field generation mechanism. @2010MNRAS.405...41G has demonstrated that the field patterns present in the MRI can be explained in terms of a mean-field dynamo model that includes dynamical $\alpha$-quenching. In the standard $\alpha-\Omega$ dynamo mechanism, the $\alpha$ effect of isotropic, helical turbulence generates poloidal field from toroidal fields, which are in turn sheared out by differential rotation $\Omega$ and regenerate toroidal field, thus leading to exponential amplification. Based on the early work of @1995ApJ...446..741B, the traditional $\alpha-\Omega$ scenario can explain the observed periodic dynamo generation and propagation of $\left< B_y \right>$ away from the disk midplane if the $\alpha$ term has the opposite sign of that expected from rotating, stratified turbulence (the strong Keplerian shear has no problem stretching poloidal field to generate toroidal field; the issue at hand is generating the former). However, that expected sign was derived from an analysis that only assumed the presence of a kinetic $\alpha_{K}$ effect from the helicity of the fluid turbulence. Once the magnetic $\alpha_{M}$ from the Lorentz force is also included, the total $\alpha = \alpha_{K} + \alpha_{M}$ is dominated by $\alpha_{M} = 1/3 \tau \mathbf{\left< J \cdot B \right>}$, where $\tau$ is a typical turbulent correlation time, and has the required sign [@2010MNRAS.405...41G]. While we similarly find that $\alpha_{M} \simeq 10 \alpha_{K}$, our results for the $z$ profile of $\alpha_{M}$ (and thus the total $\alpha$) contradict those of [@2010MNRAS.405...41G]. Figure \[f:jbm\_z\_sat\] shows the profile for $\left< J \cdot B \right>_{xy}(z)$ for three simulations with ${\ensuremath{\mathrm{Pm}}}= 1, 4$ and ${\ensuremath{\mathrm{Re}}}= 3200, 6400, 12800$. There is no monotonic trend with ${\ensuremath{\mathrm{Pm}}}$, and there are some differences in shape, but the overall profile is negative in the upper plane ($z > 0$) and positive in the lower, opposite to that found by @2010MNRAS.405...41G. The only significant difference between his study and ours are the vertical boundary conditions on the fluid, which are outflow in his case. We see a very similar profile for all boundary conditions. ![Radial and aziumthially averaged total current helicity $\left< J \cdot B \right>_{xy}(z)$ for several runs averaged over $t > 5 t_{orb}$. Note the profile is opposite to that presented in @2010MNRAS.405...41G.[]{data-label="f:jbm_z_sat"}](jbm_z_sat.pdf){width="0.95\columnwidth"} Discussion {#s:discussion} ========== We interpret our results in light of the ability of the system to build up a large scale magnetic flux. Our results suggest that, with stratification, the MRI can act as a mean-field dynamo with any boundary conditions, unlike the unstratified case. In order to clarify terms, we will refer to a mean-field dynamo as any system capable of building magnetic fields at the lowest possible wavenumber in the box, either $k = 1$ in the periodic case or $k = 0$ (a true mean field) in the case including a VF boundary condition. Current Helicity ---------------- In the modern picture of mean-field dynamo theory, the conservation of magnetic helicity, $$H \equiv \left< \mathrm{A \cdot B} \right>,$$ is an important constraint. The evolution equation for $H$ can be written $$\label{e:helicity} \frac{dH}{dt} = -2 \eta \left< \mathbf{J \cdot B} \right> - 2 \oint \left( \mathbf{ A \times E} + \phi \mathbf{B} \right) \cdot \mathbf{\hat{n}} dS,$$ and $\phi$ is an arbitrary gauge term in the definition of the vector potential [@2005PhR...417....1B]. In our simulations we use the Kepler Gauge, $\phi = \mathrm{{\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\cdot A}$; see § \[s:gauge\] for more details [@1995ApJ...446..741B]. The first term is due to resistivity acting on the current helicity $C \equiv \left< \mathrm{J \cdot B} \right>$, while the second is a boundary flux term that can entirely determine the behavior of the MRI in unstratified simulations [@2010arXiv1004.2417K]. We note that our gauge choice eliminates all possible horizontal magnetic helicity fluxes [see @2011ApJ...727...11H for a detailed explanation], leaving only the possibility of vertical fluxes, which could be driven by a turbulent diffusivity or shear, as in the case of @2001ApJ...550..752V. In the stratified case, the vertical boundary conditions are considerably less important. This is evident in the periodic, stratified simulations of @2010ApJ...713...52D, who find sustained turbulence for lower ${\ensuremath{\mathrm{Pm}}}$ than similar unstratified boxes. @2010MNRAS.405...41G has emphasized the importance of current helicity in his mean-field model for the MRI. We note that the current helicity drives the magnetic $\alpha_M$ effect @1976JFM....77..321P, which in turn leads to an inverse cascade of magnetic energy, and thus the build up of large scale magnetic field. That inverse cascade can be seen simply as a result of the conservation of magnetic helicity. We outline this process following @2005PhR...417....1B. Consider a fully helical field, $\mathcal{H} = 1$, which in turn implies $k H(k) = 2 M(k)$. Because the non-linear terms allow coupling only among modes $k$ and $p$ satisfying $p + q = k$, and helicity is conserved for each mode, $p H(p) = 2 M(p)$ and $q H(q) = 2 M(q)$. Since energy is conserved in the interaction, $M(k) = M(p) + M(q)$, and thus $p H(p) + q H(q) = k H(k)$. Because helicity must also be conserved, $H(k) = H(p) + H(q)$, and $$k = \frac{p H(p) + q H(q)}{H(p) + H(q)}. \label{e:invcas}$$ If $k$ is the final wavenumber, one of $p$ or $q$ is the starting wavenumber, and equation \[e:invcas\] demands that $k$ is less than or equal to the maximum of $p$ or $q$ (the other is an arbitrary mediating wavenumber in the three-wave interaction). Since $k$ is smaller than or equal to the parent wavenumber, this corresponds to an inverse cascade of magnetic energy. @1976JFM....77..321P explicitly relate this to the current helicity, and construct a magnetic $\alpha_M$ term that back-reacts on the flow as this cascade proceeds. ![Magnetic helicity $H = \left< \mathbf{A \cdot B} \right>$ (blue solid line) and current helicity $C = \left< \mathbf{J \cdot B} \right>$ (red dashed line), as a function of time for two run with ${\ensuremath{\mathrm{Re}}}= 3200$, ${\ensuremath{\mathrm{Pm}}}= 2$, one with periodic boundary conditions (upper panel; ensuring that $H$ is gauge independent and thus physically meaningful) and one with VF boundary conditions. The green triangles are a simple time integration of $-2\eta C$, plotted every 100 timesteps. Horizontal lines mark the zero points for each axis. For periodic boundary conditions, the integration nearly overlies $H$, demonstrating that magnetic helicity is indeed constrained by resistive action on current helicity. For VF boundary conditions, the integration (green triangles) does not track the helicity at all, presumably because the flux terms in equation \[e:helicity\] are not zero in this case. \[f:abm\]](abm_jbm_vs_t_B100Re3200Pm2_64_pbc.pdf "fig:"){width="0.95\columnwidth"} ![Magnetic helicity $H = \left< \mathbf{A \cdot B} \right>$ (blue solid line) and current helicity $C = \left< \mathbf{J \cdot B} \right>$ (red dashed line), as a function of time for two run with ${\ensuremath{\mathrm{Re}}}= 3200$, ${\ensuremath{\mathrm{Pm}}}= 2$, one with periodic boundary conditions (upper panel; ensuring that $H$ is gauge independent and thus physically meaningful) and one with VF boundary conditions. The green triangles are a simple time integration of $-2\eta C$, plotted every 100 timesteps. Horizontal lines mark the zero points for each axis. For periodic boundary conditions, the integration nearly overlies $H$, demonstrating that magnetic helicity is indeed constrained by resistive action on current helicity. For VF boundary conditions, the integration (green triangles) does not track the helicity at all, presumably because the flux terms in equation \[e:helicity\] are not zero in this case. \[f:abm\]](abm_jbm_vs_t_B100Re3200Pm2_64.pdf "fig:"){width="0.95\columnwidth"} For a system with periodic boundary conditions, the second term in equation \[e:helicity\] is zero, and the only contribution to changes in helicity can come from resistively limited current helicity fluctuations. In order to establish that helicity conservation is robust in our simulations, we ran a simulation with ${\ensuremath{\mathrm{Re}}}= 3200$ and ${\ensuremath{\mathrm{Pm}}}= 2$, with periodic boundary conditions, and tracked the time evolution of magnetic and current helicities (upper panel of figure \[f:abm\]). We used a simple forward Euler scheme to integrate $dH/dt = -2\eta C$ with volume average data from the simulations. The $H(t)$ that results from this integration is shown in figure \[f:abm\] as the green triangles, while the $H$ calculated directly during the run is given by the blue solid line. The agreement is quite good, considering the crudeness of the integration method and the sparseness of the data (it is sampled at intervals of 100 timesteps). As a result, we are reasonably confident that our simulations are accurately tracking magnetic helicity. Moving to the case with VF boundary conditions, the lower panel of figure \[f:abm\] again shows the current helicity and the now-gauge dependent quantity $H = \left< A \cdot B\right>$. Because of the presence of the gauge in the second term of equation \[e:helicity\], this quantity is not physically meaningful itself. However, it is well-defined and provides a clue as to the behavior of the system. Once again, we integrate $-2\eta C$ using $C$ in the same way. We expect that if magnetic helicity ejection is important in the MRI, the $H$ resulting from this integration will *not* track the actual $H$ from the simulation, and indeed the lower panel of figure \[f:abm\] shows that it does not. Because of our gauge choice, we know that the flux of magnetic helicity flux density must be vertical, and this figure confirms that significant amounts of helicity escape. Furthermore, the $H$ fluctuates considerably more frequently in the VF case than in the run with periodic boundary conditions, showing that the timescale for variation of the global magnetic helicity is considerably shorter when a helicity flux is allowed. {width="0.95\columnwidth"} \[f:alpha\_vs\_t\_bcs\] How does transport work in the periodic case? The only way for *net* helicity to change in this case is via resistive effects, leading to catastrophic quenching of the dynamo effect, where saturated magnetic energy is $\propto {\ensuremath{\mathrm{Rm}}}^{-1/2}$ [@1992ApJ...393..165V]. We have shown that this is not the case in the stratified MRI: boundary conditions do not make significant differences in the transport (figure \[f:alpha\_vs\_t\_bcs\]) and for vertical field boundaries, figure \[f:bym\_bxm\_vs\_t\] demonstrates that the saturated large-scale field strength is not declining strongly with increasing ${\ensuremath{\mathrm{Rm}}}$. @2001ApJ...550..752V suggested a potential solution to this situation: helicity may not need to be ejected *entirely* across a system boundary (as in, say, a coronal mass ejection in the Solar dynamo). It could instead be transported *spectrally*, transferring from small scales to large. Indeed, @2001ApJ...550..824B show that this is exactly the case for a helically driven turbulence simulation with no shear or rotation–the prototypical $\alpha^2$ dynamo. However, in this case, while equipartition fields are built even for periodic boundary conditions, the timescale required to do so is $t_{sat} \propto \eta^{-1}$: instead of catastrophic quenching, there is a catastrophic timescale problem instead [@2001ApJ...550..824B]. While our data is not conclusive on the relationship between ${\ensuremath{\mathrm{Rm}}}$, ${\ensuremath{\mathrm{Pm}}}$, and the cycle period of large scale magnetic fields, it certainly does not suggest an inverse relationship between $t_{sat}$ and ${\ensuremath{\mathrm{Rm}}}$. However, the MRI is difficult to analyze in these terms: it does not present a single energy injection scale at which we could expect small scale helicity to be generated [@2010ApJ...713...52D]. This lack of clear scale separation makes it difficult to extract dynamo coefficients via formalisms such as the test field method. Nonetheless, we should be able to see a crude difference between the current helicity at the small and large scales if this is in fact what allows dynamo action, and thus transport. We can follow the recent work by @2010GApFD.104..577H and attempt to understand the transport of small and large scale magnetic helicity directly. A significant difference between the unstratified results of @2010arXiv1004.2417K and ours is that their simulations are likely to be some form of an incoherent-$\alpha$ dynamo, in which case despite having a zero-mean $\alpha$, the presence of a shear-driven small scale helicity flux across the boundaries allows mean field growth. In our case, small-scale helicity annihilation may be occurring at the midplane, and this process may be weakly-${\ensuremath{\mathrm{Pm}}}$ dependent. Furthermore, the correlated dynamo waves present in stratified but absent in unstratified MRI turbulence may lead to expulsion of both large and small scale helicity, leading again to reduced dynamo action. The two runs in figure \[f:abm\], taken together, tell an intriguing tale: the stratified MRI can transport angular momentum and build large scale magnetic energy even without a global helicity flux, but if one is allowed, the system takes advantage of the possibility. Ultimately, as any boundary condition is in some sense an approximation of reality, we must understand the details of the accretion disk dynamo independent of this choice. However, it is also clear comparing figures \[f:abm\] and \[f:bym\_bxm\_vs\_t\] that the timescale for building mean fields for the vertical field boundary condition case is not related in any obvious way to the flux of magnetic helicity. We address this in the next section. Power spectra ------------- In order to understand the dependence of the magnetic field on its magnetic helicity, we measure its scale-dependent relative helicity, $$\mathcal{H}(k) = \frac{k H(k)}{2 M(k)},$$ where $H(k)$ and $M(k)$ are the Fourier transforms of magnetic helicity and magnetic energy, respectively. A fully helical field has $\mathcal{H}(k) = 1$. We compute this spectrum on spherical shells in $k$-space, only on runs with closed boundary conditions (i.e., periodic or perfect conductor). This ensures magnetic helicity is gauge independent by removing the surface terms in equation \[e:helicity\]. In figure \[f:helicty\_spectrum\], we show the helicity spectrum for ${\ensuremath{\mathrm{Re}}}= 3200$, ${\ensuremath{\mathrm{Pm}}}=2$. The field is not strongly helical at any scale, reaching only $\mathcal{H} \sim 0.2$ at the smallest scales and remaining much lower on the largest scales. By contrast, the $\alpha^2$ dynamo driven by a fully helical fluid forcing function has $\mathcal{H} \sim 1$ at all scales [@2001ApJ...550..824B]. In that case, something similar to the classic inverse cascade of magnetic energy due to helicity conservation [@1975JFM....68..769F; @1976JFM....77..321P] occurs. In our case, however, we do not see a significant difference in the properties of the dynamo when the boundary conditions change between those that do not allow a flux of magnetic helicity and those that do, despite the fact that our results suggest that the MRI does in fact eject helicity when given the chance (see figure \[f:abm\]). This resolves that observation: the MRI generated field is not strongly helical, even when both kinetic and magnetic $\alpha$ effects occur. Thus, the constraints placed on the field by the conservation of magnetic helicity do not dominate its formation. Furthermore, the fact that the relative helicity is peaked at small scales suggests that helicity constraints might be more important in the unstratified MRI, as suggested by @2010arXiv1004.2417K. ![Relative magnetic helicity as a function of scale for a run with ${\ensuremath{\mathrm{Re}}}= 3200$, ${\ensuremath{\mathrm{Pm}}}= 2$ and periodic boundary conditions, averaged over a $10 t_{orb}$ period in saturation. Light lines show individual timesteps, the dark line is the average. The value remains well below unity at all scales, showing that the field is not significantly helical at any scale.[]{data-label="f:helicty_spectrum"}](B100Re3200Pm2_64_pbc_var90-100_h_spectra_f.pdf){width="0.95\columnwidth"} Open Issues and Some Speculations --------------------------------- A few issues remain unresolved. Chief among them is the incongruity between our results for $\alpha_M(z)$ and those of @2010MNRAS.405...41G. The only significant difference between his simulations and ours is his use of outflow boundary conditions on the fluid (the magnetic field boundary conditions are identical). The MRI coupled with these boundary conditions leads to a magnetized wind, which could saturate the dynamo itself [@2001ApJ...550..752V]. If this were the case, then it would not be surprising that our results differ, given that losses from a wind are not an available saturation mechanism here. We have demonstrated that understanding angular momentum transport via the MRI is strongly tied to the detailed physics of MHD turbulence and dynamo action. It is worth commenting briefly on some results for isotropic (non-shearing), non-rotating MHD turbulence, a much better studied system. Recently, several groups have demonstrated that non-local interactions in $k$-space are important at large scales in MHD turbulence, cross-coupling large scale velocity fields with small scale magnetic fields [e.g. @2005PhRvE..72d6302M; @2009PhRvE..79f6307L; @2010ApJ...725.1786C]. If such an analysis holds for the stratified MRI dynamo, it could explain the existence of a critical magnetic Reynolds number ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$: if the large scale velocity fields are coupled to the magnetic dissipation range, a much more efficient energy sink appears than if they are coupled to a magnetic inertial range. In the latter case, the large scale velocities could contribute to small scale helicity production, continuing to drive large scale dynamo action. This could be verified by a shell-transfer analysis of the stratified MRI, which we will pursue in a future publication. We do see some residual ${\ensuremath{\mathrm{Pm}}}$ dependence for the stratified MRI, despite the presence of boundaries identical to those of @2010arXiv1004.2417K. We speculate that these differences to a difference in the physical transport of magnetic helicity with and without stratification. In the unstratified case, the dynamo essentially acts as an incoherent-$\alpha$ dynamo, but in the stratified case, $\alpha$ shows long-time coherence. In the latter case two distinct possibilities remain to explain the residual ${\ensuremath{\mathrm{Pm}}}$ dependence: first that small-scale helicity is being ejected through the boundaries, but it is accompanied by significant amounts of total flux, and that total flux is ${\ensuremath{\mathrm{Pm}}}$ dependent. Another possibility is that the stratified case has gradients down which helicity can be transported diffusively, which would also be sensitive to the details of the turbulence and hence ${\ensuremath{\mathrm{Pm}}}$. Furthermore, the @2010arXiv1004.2417K results may have ${\ensuremath{\mathrm{Rm}}}>> {\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$, and our results, if numerical resolution allowed, would also show a ${\ensuremath{\mathrm{Pm}}}$-independent $\alpha$ for sufficiently high ${\ensuremath{\mathrm{Re}}}$. It is also possible that our hyperdiffusive terms might be biasing our results. However, in a series of detailed small-scale dynamo simulations, groups led by Schekochihin and Brandenburg have found no evidence that hyperviscosity plays a significant role [@2005ApJ...625L.115S; @2007NJPh....9..300S]. Nevertheless, this should be explicitly checked and we cannot rule out any hyperdiffusive influence. Conclusions {#s:conclusions} =========== The strength of the angular momentum transport parameter ${\ensuremath{\alpha_{SS}}}$ in MRI driven turbulence appears to scale only weakly, if at all, with the magnetic Prandtl number ${\ensuremath{\mathrm{Pm}}}$ above some critical magnetic Reynolds number ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}\sim 3000$. Our models suggest that the ${\ensuremath{\mathrm{Pm}}}$ at which the flattening of the ${\ensuremath{\mathrm{Pm}}}- {\ensuremath{\alpha_{SS}}}$ relation occurs is a function of ${\ensuremath{\mathrm{Rm}}}$. @2010arXiv1004.2417K demonstrated that in the unstratified case, $\alpha_{SS}$ is entirely independent of ${\ensuremath{\mathrm{Pm}}}$ if boundary conditions allow for the ejection of magnetic helicity. If these conclusions are indicative of the asymptotic state in real disks, then concerns about ${\ensuremath{\mathrm{Pm}}}$-dependent scaling of the MRI in real disks, which are certainly stratified, would be entirely alleviated: astrophysical disks have tremendous values of $Re$ and $Rm$, so even with a ${\ensuremath{\mathrm{Pm}}}\sim 10^{-8}$, ${\ensuremath{\mathrm{Rm}}}>> {\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$ if ${\ensuremath{\mathrm{Re}}}\sim 10^{12}$. Using the @2008ApJ...674..408B estimates for the viscosity and resistivity, we can estimate the Reynolds number for a typical protoplanetary disk. Using a fiducial midplane density $\rho \simeq 10^{-10} \mathrm{g\ cm^{-3}}$, scale height $H \simeq 0.05 \mathrm{AU}$, and temperature $T \sim 500 \mathrm{K}$ , we arrive at ${\ensuremath{\mathrm{Re}}}\simeq 4 \times 10^{16}$. The same estimates give ${\ensuremath{\mathrm{Pm}}}\simeq 1 \times 10^{-8}$, easily fulfilling ${\ensuremath{\mathrm{Rm}}}>> {\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$ and so we expect the transport to be independent of ${\ensuremath{\mathrm{Pm}}}$ in such disks. A major caveat to this, of course, is that the ionization state of such disks is still not well established. Our results do not bear on the question of dead zones, which result from poor ionization. Nevertheless, our point here is to establish the dynamo state of a protoplanetary disk, *not* to compute detailed predictions for ${\ensuremath{\alpha_{SS}}}$ in such systems. Finally, black hole accretion systems should also be well in excess of ${\ensuremath{\mathrm{Rm}_\mathrm{crit}}}$: the same computation yields ${\ensuremath{\mathrm{Re}}}\simeq 6 \times 10^{12}$ and ${\ensuremath{\mathrm{Pm}}}\simeq 0.1$ at roughly $100$ Schwarzschild radii for a $10 M_{\odot}$ black hole. We thank Oliver Gressel, Jake Simon, Eliot Quataert, Ian Parrish, Axel Brandenburg, and Marie Oishi for helpful discussions. Computations were performed on the Kraken machine at NICS under grant TG-AST090072, the Big Ben and Pople systems at Pittsburgh Supercomputing Center under grant TG-MCA99S024, both supported by NSF, and the NASA Advanced Supercomputing Division’s Pleiades system under grant SMD-19-1846. The Kepler Gauge {#s:gauge} ================ We here derive the Kepler gauge used in our definition of the magnetic vector potential. Beginning from the induction equation for the vector potential, $$\frac{\partial \mathbf{A}}{\partial t} = \mathbf{v \times B} - \eta \mu_0 \mathbf{J}$$ where $\mathbf{J = \nabla \times B}$, we expand the velocity $\mathbf{v = u + {\ensuremath{\mathbf{U}^{\mathrm{kep}}}}}$, where ${\ensuremath{\mathbf{U}^{\mathrm{kep}}}}= q \Omega_0 x \hat{\mathbf{y}}$ is the linearized Keplerian shear velocity. Making this substitution, we have $$\frac{\partial \mathbf{A}}{\partial t} = \mathbf{u} \times \mathbf{B} + [{\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\times \mathbf{B}] - \eta \mu_0 \mathbf{J}$$ Expanding the second term on the right hand side, $${\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\times \mathbf{B} = {\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\times (\mathbf{\nabla \times A}) = \mathbf{\nabla({\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\cdot A)} - \mathbf{A \times (\nabla \times {\ensuremath{\mathbf{U}^{\mathrm{kep}}}})} - \mathbf{{\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\cdot \nabla A} - \mathbf{A \cdot \nabla {\ensuremath{\mathbf{U}^{\mathrm{kep}}}}}.$$ The first term on the right hand side is the gradient of a scalar function ${\ensuremath{\phi_{\mathrm{kep}}}}= \mathbf{{\ensuremath{\mathbf{U}^{\mathrm{kep}}}}\cdot A}$, which we term the Kepler gauge. The remainder of the terms represent shear and advection, and correspond to the second terms on either side of equation \[e:induction\]. We refer the reader to @2011ApJ...727...11H for a generalization of this gauge to any advective velocity . [^1]: `http://www.nordita.org/software/pencil-code/`

--- abstract: 'Coupled arrays of Andronov-Hopf oscillators are investigated. These arrays can be diffusively or repulsively coupled, and can serve as central pattern generator models in animal locomotion and robotics. It is shown that repulsive coupling generates out-of-phase oscillations, while diffusive coupling generates synchronous oscillations. Specifically, symmetric solutions and their corresponding amplitudes are derived, and contraction analysis is used to prove global stability and convergence of oscillations to either symmetric out-of-phase or synchronous states, depending on the coupling constant. Next, the two mechanisms are used jointly by coupling multiple arrays. The resulting dynamics is analyzed, in a model inspired by the CPG-motorneuron network that controls the heartbeat of a medicinal leech.' address: - 'ETH Zurich, Department of Physics, Zurich, Switzerland (email: landsmanster@gmail.com)' - 'MIT, Nonlinear Systems Lab, Boston, MA, USA (e-mail: jjs@mit.edu)' author: - 'Alexandra S. Landsman' - 'Jean-Jacques Slotine' title: 'Global stability of synchronous and out-of-phase oscillations in central pattern generators' --- I. Introduction =============== Central pattern generators (CPGs) are often modeled as coupled nonlinear oscillators delivering phase-locked signals. Some of their applications include animal locomotion, robotics (see [@Seo:07], [@Ijs:07], and [@Ijs:08]), and other biological rhythmic behaviors such as e.g. the generation of a heartbeat, [@Buo:04]. In this paper, we explore a system of coupled Andronov-Hopf oscillators that can be used to model such phenomena. Repulsive coupling between the oscillators creates a traveling wave that simulates salamander gate ([@Ijs:07]), and with additional coupling architecture, the heartbeat of a leech ([@Buo:04]). Explicit rotational coupling between neighbouring oscillators has been used in earlier work (see [@Pha:07], [@Seo:07]) to generate a traveling wave or out-of-phase state. Here, we use inhibitory (repulsive) coupling to achieve a similar effect in perhaps a more physical way, allowing indeed the system itself to compute couplings achieving the out-of-phase behavior. While all our derivations aim at establishing global convergence results, we found some significant qualitative differences between these two types of couplings, such as the existence of two degenerate out-of-phase states in repulsive coupling, but not in rotational coupling. In addition, while rotational coupling preserves the amplitude of the coupled oscillators, repulsive coupling increases the amplitude of oscillation. In fact, this increase in amplitude is the key factor behind generating high frequency oscillation patterns in the heartbeat of a leech model, to be analyzed in the paper. The paper is organized as follows. In Part II, we obtain out-of-phase and synchronous solutions for nearest neighbor coupled Andronov-Hopf oscillators, and derive the condition for the existence of an out-of-phase state. In Part III, either the out-of-phase or the synchronous solution is shown to be globally stable using techniques from contraction theory (for original derivation of contraction theory see [@Loh:98]). The stability of the two different types of behavior is determined by the values of the coupling constant, with a positive coupling constant generating the out-of-phase state and a negative constant resulting in a globally stable synchronous solution. In Part IV, the synchronous and the out-of-phase arrays are globally coupled to each other in a model inspired by the CPG in the heartbeat of a leech. Building on prior results, the bifurcation value for the onset of high frequency oscillations in the synchronous array is derived. II. Steady-state dynamics ========================= In this section we solve for the steady-state dynamics of the nearest neighbor coupled Andronov-Hopf oscillators, showing that both synchronous and out-phase solutions exist. In the next section, we consider the global stability properties of these solutions applying techniques from contraction theory. We first consider a ring of nearest neighbor coupled Andronov-Hopf oscillators, of the form: $${\bf{\dot{x}_j = F (x_j)}} + k \left(\bf{x_j - x_{j+1} - x_{j-1}} \right) \label{eq:osc}$$ where ${\bf{x_j}}=\{x_j, y_j\}$ is a two dimensional vector describing the dynamics of the $j$th oscillator, with $\bf{F(x)}$ given by: $$\bf{F}\left(\begin{array}{c} x \\ y \end{array} \right)\ = \left(\begin{array}{c} x-y-x^3 -x y^2 \\ x+y-y^3 - y x^2 \end{array} \right)\ \label{eq:fx}$$ In complex form, Eqns. (\[eq:osc\]) and (\[eq:fx\]) can be expressed as: $$\dot{z}_j = (\alpha + i \omega) z_j - |z_j|^2 z_j + k \left(z_j - z_{j-1} - z_{j+1} \right) \label{eq:complex}$$ where $z$ is a complex variable, given by: $z = x + i y$. In the absence of coupling, the dynamics are that of a limit cycle, of amplitude, $ |\bf{x_j}| = \sqrt{\alpha}$ and frequency $\omega$. Due to symmetry considerations, the synchronous state is one possible solution to the above equation, resulting in the amplitude of oscillation given by: $ |{\bf{x_j}}| = \sqrt{\alpha - k}$. This solution will be shown in the next section to be globally stable for diffusive type of coupling, given by: $k < 0$. For the “repulsive coupling”, given by $k > 0$, the system tends to an out-of-phase state, whereby the neighboring oscillators are maximally out of phase with each other, [@Lan:06]. For oscillators coupled in a ring, this results in two different types of dynamics, depending on whether the number of oscillators, $N$ is even or odd. When $N$ is even, the array oscillates with a difference of $\pi$ between nearest neighbors, splitting into two equal synchronous groups, that are $180$ degrees out of phase with each other (see Figure \[fig:2\], plotted for $N=4$). The phase difference between nearest neighbors is thereby given by:$$\triangle \phi_{j,j+1}^{even} = \pi \label{eq:Neven}$$ In the case of oscillators in a line, coupled with $k > 0$, the above phase difference is the only globally stable solution. The situation becomes more complicated in a ring coupled model when $N$ is odd. In this case the phase difference of $\pi$ between nearest neighbors is not a symmetric or a stable solution. ![Repulsively coupled array, $k > 0$, for even number of oscillators, $N=4$. Shows that the neighboring oscillators are maximally out-of-phase (by $\pi$)[]{data-label="fig:2"}](AH_Fig1.pdf){height="7cm"} For example, imagine a ring of 3 oscillators, where the 2nd oscillator is out of phase with the 1st and the 3rd by $\pi$. Then the 1st and the 3rd oscillator will actually be in-phase, which is an unstable state. Requiring the neighboring oscillators to be maximally out of phase, while preserving the symmetry of the system leads to two possible degenerate solutions given by: $$\triangle \phi_{j,j+1}^{odd} = \pi \pm \pi/N \label{eq:Nodd}$$ The smallest phase difference between the two oscillators is given by the next to nearest neighbor phase difference: $\triangle \phi_{j,j+2}^{odd} \pm 2 \pi/N$. The vectors $\bf{x_j}$ therefore fall on a circle where they are spaced with an equal phase difference of $2 \pi/N$, forming a symmetric out-of-phase solution. Figure \[fig:3\] illustrates the stable out-of-phase dynamics for $N=5$. Each oscillator in the Figure is shifted in phase from its neighbor by $\pi + \pi/5$. Comparing Eq. (\[eq:Nodd\]) to Eq. (\[eq:Neven\]), we can see that as $N \rightarrow \infty$, the steady-state dynamics of the $N$-odd array approach that of $N$-even. This should be expected, since for large $N$, adding one more oscillator to the ring (and thereby changing $N$ from odd to even or vice versa) will not significantly change the energy function, which is minimized when Eqs. (\[eq:Nodd\]) and (\[eq:Neven\]) for $N$-odd and $N$-even, respectively, are satisfied. ![Repulsively coupled array for odd number of oscillators, $N=5$. The neighboring oscillators are out of phase by $\pi+\pi/5$ (for example, compare the first and the second oscillator).[]{data-label="fig:3"}](AH_Fig2.pdf){height="7cm"} To solve for the amplitude of the out-of-phase solution, which is stable for repulsive coupling, we use Eqn. (\[eq:Nodd\]), writing $\{z_{j-1} = e^{\pm i \left(\pi + \pi/N \right)} z_j, z_{j+1} = e^{\mp i \left(\pi + \pi/N \right)} z_j\}$ and substituting for $\{z_{j-1}, z_{j+1}\}$ into Eqn. (\[eq:complex\]). After grouping the linear terms, we can now solve for the amplitude of oscillation, getting: $$|z_{j}|^{N-odd} = \lbrack \alpha + k \left( 1 + 2 cos(\pi/N)\right) \rbrack ^ {1/2} \label{eq:AmpOdd}$$ where the above equation is valid for all oscillators with odd $N$. Performing the same type of analysis for even number of oscillators, where the nearest neighbors are $180$ degrees out of phase, we get: $$|z_{j}|^{N-even} = \left(\alpha + 3 k \right)^{1/2} \label{eq:AmpEven}$$ Note that in the above equation the amplitude of the oscillation is independent of the total number of oscillators, while in Eqn. (\[eq:AmpOdd\]), this amplitude increases with increasing $N$, asymptotically approaching $\left(\alpha + 3 k \right)^{1/2}$ as $n \rightarrow \infty$. III. Contraction analysis ========================= This section uses partial contraction analysis to analyze the stability of the synchronous and out-of-phase states for the system in Eqns. (\[eq:osc\]) - (\[eq:complex\]), for the $N=3$ case. Here we use the partial contraction results first derived in the paper by Pham and Slotine, see [@Pha:07]. The results state a simple sufficient condition for global exponential stability on a flow-invariant linear subspace $\mathcal{M}$ (i.e. a linear subspace $\mathcal{M}$ such that $\forall t: {\bf{F}}(\mathcal{M}, t) \subset \mathcal{M}$) as given by: $$- \lambda_{min} \left(\bf{V L V^T} \right) > sup \lambda_{max} \left(\frac{\partial \bf{F}}{\partial \bf{x}}\right) \label{eq:condition}$$ where $\bf{V}$ forms a basis of the linear subspace, $\mathcal{M^\perp}$ (orthogonal to $\mathcal{M}$), $\partial \bf{F}/\partial \bf{x}$ is the Jacobian of the uncoupled system, with $\bf{F}$ given by Eqn. (\[eq:fx\]) and $\bf{L}$ is the coupling matrix, to be given below. The terms $\lambda_{min}$ and $\lambda_{max}$ indicate the minimum and maximum eigenvalues of the symmetric parts of the matrices $\bf{L}$ and $\partial \bf{F}/\partial \bf{x}$, respectively. Intuitively, the above condition insures that the system is contracting in the orthogonal subspace, $\mathcal{M^\perp}$, thereby insuring that the dynamics converge exponentially to $\mathcal{M}$. From Eqn. (\[eq:osc\]), the coupling matrix, $\bf{L}$ for the $N=3$ case is given by: $$\bf{L} = k \left(\begin{array}{ccc} I & -I & -I \\ -I & I & -I \\ -I & -I & I \end{array}\right) \label{eq:something}$$ where $I$ above is a $2 \times 2$ identity matrix. The entire phase space, $\mathcal{M} \oplus \mathcal{M^\perp}$, is spanned by a total of six vectors: the two vectors spanning the synchronous solution, plus the four vectors spanning a linear vector subspace formed by the two degenerate out-of-phase solutions. For convenience, let’s define the subspace corresponding to the synchronous case as: $$\mathcal{M}_{sync} = \{\left(\bf{x, x, x} \right) : \bf{x} \in \mathcal{R}^2 \} \label{eq:Msync}$$ and the subspace corresponding to the out-of-phase states as: $$\mathcal{M}_{phase} =\{ \left(\bf{x, R_{\frac{2 \pi}{3}} x, R_{\frac{4 \pi}{3}} x} \right), \left(\bf{x, R_{\frac{4 \pi}{3}} x, R_{\frac{2 \pi}{3}} x} \right) \} \label{eq:Mphase}$$ Where $\bf{R}$ is a $2 \times 2$ rotation matrix, with the rotation angles of $\{2 \pi/3, 4 \pi/3 \}$ obtained from Eqn. (\[eq:Nodd\]) for the $N=3$ case. The corresponding eigenvectors can be obtained by substituting ${\bf{x}} =\{1, 0 \}$ and $\{0, 1 \}$ into Eqns. (\[eq:Msync\]) and (\[eq:Mphase\]), resulting in six orthogonal eigenvectors. In complex form the two out-of-phase solutions given in Eqn. (\[eq:Mphase\]) can also be written as: $z \cdot \left(1, e^{i 2 \pi/3}, e^{i 4 \pi/3} \right)$ and $z \cdot \left(1, e^{i 4 \pi/3}, e^{i 2 \pi/3} \right)$. As before, the entire phase-space of solutions is spanned by the sum: $\mathcal{M}_{sync} \oplus \mathcal{M}_{phase}$. Having obtained all the eigenvectors, we can now use Eqn. (\[eq:condition\]) to analyze the stability of either the synchronous or the out-of-phase states, represented by $\mathcal{M}_{sync}$ and $\mathcal{M}_{phase}$, respectively. As will be shown shortly, this stability depends on the value of the coupling constant, $k$, with the synchronous solution being stable for negative $k$ (or diffusive type of of coupling) and the out-of-phase subspace being stable for positive values of $k$ (representing repulsive coupling). To analyze the stability of the synchronous state, we equate: $\mathcal{M}_{sync} \equiv \mathcal{M}$ and $\mathcal{M}_{phase} \equiv \mathcal{M^\perp}$, thereby obtaining the $6 \times 4$ matrix $\bf{V}$ from Eqn. (\[eq:Mphase\]). The four eigenvalues of $ \bf{V L V^T}$, with $\bf{L}$ given in Eqn. (\[eq:something\]), are all identical and given by: $\lambda_{1,2,3,4} = 2 k$. The eigenvalues of the symmetric part of the Jacobian, $\partial \bf{F}/\partial \bf{x}$, are given by: $\alpha - |x|^2$ and $\alpha - 3 |x|^2$, which are upper-bounded by $\alpha$. It follows from Eqn. (\[eq:condition\]) that the solution converges exponentially to $\mathcal{M}_{sync}$ when, $$\qquad k < -\alpha/2 \label{eq:synchstable}$$ Next, doing the reverse by equating: $\mathcal{M}_{phase} \equiv \mathcal{M}$ and $\mathcal{M}_{sync} \equiv \mathcal{M^\perp}$, we again obtain the eigenvalues of $ \bf{V L V^T}$ (with $\bf{V}$ now given by Eqn. (\[eq:Msync\])): $\lambda_{5,6} = -k $. Combining the above results and again using Eqn. (\[eq:condition\]), we now get the condition for the global stability of $\mathcal{M}_{phase}$: $$k > \alpha \label{eq:kstable}$$ Equations (\[eq:synchstable\]) and (\[eq:kstable\]) give conditions for the global stability of synchronous and out-of-phase dynamics, respectively. It is instructive to compare the model in Eqns. (\[eq:osc\]) and (\[eq:fx\]) to another coupling architecture which directly uses rotational matrices, see [@Seo:07] and [@Pha:07], to create globally stable out-of-phase state, in the following way (for a network of $N$ oscillators): $${\bf{\dot{x}_j = {\bf{F}} (x_j)}} + k \left(\bf{x_j - R_{\frac{2 \pi}{N}} x_{j-1}} \right) \label{eq:oscR}$$ The above model results in a globally stable state where the nearest neighbors are displaced out-of-phase by $2 \pi/N$. Unlike the model in Eqn. (\[eq:oscR\]), the coupling term in Eqn. (\[eq:osc\]) does not need to be adjusted to get the $2 \pi/N$ out-of-phase state as more oscillators are added, provided that the number, $N$, is always increased by 2, so that $N$ remains odd. As described in the previous section, for $N$-even, the system splits into two identical synchronous groups $180$ degrees out-of-phase with each other. The other significant differences of the out-of-phase state created by the coupling term in Eqn. (\[eq:oscR\]) from the system analyzed in the present work include: I. Existence of out-of-phase solution for any number, $N$, of oscillators. This is in contrast to the system analyzed here, where (as mentioned above), an odd $N$ is needed for an out-of-phase state, II. The existence of a single (non-degenerate) out-of-phase solution, unlike the two solutions given in Eqn. (\[eq:Mphase\]), III. Different phase difference between nearest neighbors. Thus, the phase difference given by Eqn. (\[eq:Nodd\]) is such that the nearest neighbors are maximally out-of-phase, while the phase difference from rotational coupling is such that oscillator $j$ is advanced from $j-1$ by a phase of $2 \pi/N$, and IV. Preservation of uncoupled amplitude, $\sqrt{\alpha}$ in the out-of-phase state. This can be seen directly by substituting the out-of-phase solution $\bf{x_j = R_{\frac{2 \pi}{N}} x_{j-1}}$ into Eqn. (\[eq:oscR\]), whereby the coupling term drops out, Points I-III can be explained by pointing out that in Eqn. (\[eq:oscR\]), the rotational coupling creates an out-of-phase state in the same way that a synchronous state is created in diffusive coupling. In other words, the oscillators all try to be in synch with the rotated by $2 \pi/N$ solution of their nearest neighbor. Therefore the out-of-phase state in rotational coupling is actually analogous to the synchronous state in diffusive coupling, which also has uniqueness, global stability (for appropriate values of $k$), and existence for any value of $N$. Point IV, that is the increase in the amplitude of oscillation (given by Eqn. (\[eq:AmpOdd\])) in the out-of-phase state, is actually essential for the creation of $N \omega$ frequency oscillations when the two arrays are globally coupled. This model, inspired by the nervous system of a leech is analyzed in the following section. IV. Globally coupled arrays and the onset of high frequency oscillations ======================================================================== Here we globally couple two arrays, each described by Eqns. (\[eq:osc\]) and (\[eq:fx\]), with the only difference being that the first array has repulsive coupling, given by $k_r>0$ and the second array has diffusive coupling, given by $k_d<0$. The dynamics were inspired by the CPG-motorneuron network that controls the heartbeat of a medicinal leech, [@Buo:04]. The heartbeat of the leech is driven by direct contact between two arrays of motorneurons, such that on one side of the leech the heart beats in a rear-to-front (peristaltic) fashion, well described by the out-of-phase state of coupled limit cycle oscillators. On the other side, the heart beats synchronously and is therefore represented by the diffusively coupled array, where the synchronous state is stable. The total system has the following form: $${\bf{\dot{x^r}_j = F (x^r_j)}} + k_r \left({\bf{x^r_j - x^r_{j+1} - x^r_{j-1} }}\right) + c \sum_{k=1}^N |\bf{x^d_k}| \label{eq:osc2}$$ $${\bf{\dot{x^d_j} = {\bf{F}} (x^d_j)}} + k_d \left({\bf{x^d_j - x^d_{j+1} - x^d_{j-1} }}\right) + c \sum_{k=1}^N |\bf{x^r_k}| \label{eq:osc3}$$ where as before, $\bf{F(x)}$ is defined in Eqn. (\[eq:fx\]). Based on the results of the previous section, we know that the system in Eqn. (\[eq:osc2\]) has a stable out-of-phase state (corresponding to peristaltic motion), while the system in Eqn. (\[eq:osc3\]) has a stable synchronous state. It has been shown both computationally (see [@Pal:05]) and analytically (see [@Lan:06]) that a system of this type undergoes a bifurcation, as the global coupling constant, $c$, increases. For $c > c_{bif}$, the $\{\bf{x^d}\}$ array begins to oscillate in phase at the ultra-harmonic frequency given by $N \omega$. Following the method in Landsman and Schwartz, [@Lan:06], the bifurcation value of $c$ that leads to high frequency oscillations can be calculated by solving for the value of the parameter $P$ that causes a Hopf bifurcation in the following equation: $$\dot{z}_j = (\alpha - k_d + i \omega) z_j - |z_j|^2 z_j + P \label{eq:complex2}$$ where we have used the complex formulation, $z=x + i y$, of Eqn. (\[eq:complex\]) for the diffusively coupled array and substituted the synchronous solution: $z_j = z_{j+1} = z_{j-1}$. The bifurcation diagram for Eqn. (\[eq:complex2\]) as a function of $P$ is plotted in Figure (\[fig:3b\]), where bold lines at higher $|P|$ indicate stable equilibria, with the broken line in the center showing an unstable equilibria. ![Nullclines for Eqn. (\[eq:complex2\]) for different values of $P$ ($\alpha=0.9$, $k_d= -0.1$, $\omega=1/2$). The green nullclines are the $\dot{x}=0$ nullclines for various values of $P$. The red nullcline is the $\dot{y}=0$ nullcline. Higher nullclines corresponding to higher values of $P$. The circular orbit is a limit cycle for $P =0$.[]{data-label="fig:4"}](AH_Fig4.pdf){height="7cm"} ![Bifurcation diagram as a function of $P$, ($\alpha=0.9$, $k_d= -0.1$, $\omega=1/2$). Bold lines correspond to the high-frequency oscillations after the system in Eqn. (\[eq:complex2\]) has been driven through a bifurcation for $P>P_{bif}$.](AH_Fig6.pdf){height="7cm"} \[fig:3b\] The bifurcation diagram in Fig. (\[fig:3b\]), in fact also corresponds to the bifurcation digram for the onset on high frequency oscillations in the diffusively coupled array, given by Eqn. (\[eq:osc3\]). Namely the solid lines in the figure correspond to the $N \omega$ frequency oscillations in the $\bf{x^d}$ array, while the broken line corresponds to oscillations close to the limit cycle frequency, $\omega$ in the same array. We can therefore obtain the bifurcation value of the global coupling, $c$ for the onset of $N \omega$ frequency oscillations by first solving for the bifurcation value of $P$ (given by $P_{bif}$), and then using: $P_{bif} \approx c _{bif} N |z^r|$ to calculate $c_{bif}$. Here $|z^r|$ is the amplitude of the out-of-phase oscillation in the repulsively coupled array. It is given by Eqn. (\[eq:AmpOdd\]), with $k \rightarrow k_r$. Solving for $c_{bif}$, we obtain the value of the bifurcation to $N \omega$ oscillations as a function of $P$: $$c_{bif} \approx \frac{P_{bif}}{N |z^r|} = \frac{P_{bif}}{N \lbrack \alpha + k_r \left( 1 + 2 cos(\pi/N)\right) \rbrack ^ {1/2}} \label{eq:general}$$ where Eqn. (\[eq:AmpOdd\]) was used in the denominator. The effect of the constant $P$ on the dynamics of Eqn. (\[eq:complex2\]) can be seen by referring to the nullclines diagram, in Figure \[fig:4\]. The various green curves correspond to the $\dot{x} =0$ nullclines plotted for the corresponding values of $P$, with the $P=0$ nullcline crossing the origin. Since $P$ is a real constant in Eqn. (\[eq:complex2\]), the $\dot{y}=0$ nullcline (shown in red) does not change with the parameter $P$. As $|P|$ increases, the system bifurcates to a steady-state with a stable fixed point given by the intersection of the corresponding nullclines. The bifurcation value of $|P|$ is found at a point where the intersection of the vertical line with the $\dot{x}=0$ nullcline no longer has three real roots (see for example [@Guck:83]), and given by: $$P_{bif} = \frac{\left(8 \gamma^2 + \omega^2 \right)}{4 \gamma^{1/2}} \label{eq:Cbnew}$$ where $\gamma = \left(\alpha - k_d \right)/3$. Note that $\gamma > 0$, since $k_d < 0$ in diffusive coupling. Substituting Eqn. (\[eq:Cbnew\]) into Eqn. (\[eq:general\]), we have the bifurcation value as a function of $\alpha$, $N$, $\omega$, and the repulsive and diffusive coupling constants, given by $k_r$ and $k_d$, respectively. The mechanism behind the onset of ultraharmonics for a similar type of coupling architecture was analyzed in [@Lan:06]. Here we briefly summarize the mechanism behind the onset. The onset of high frequency oscillations hinges on the amplitude of the repulsively coupled array being higher than the amplitude of the diffusively coupled array. This results in a diffusively coupled array being driven through a bifurcation first, when $c > c_{bif}$, with $c_{bif}$ given in Eqn. (\[eq:general\]), and thereafter being driven by the out-of-phase dynamics of the repulsively coupled array, which causes the high-frequency synchronous oscillation in ${\bf{x^d}}$. In the language of contraction theory, [@Loh:98], for $c > c_{bif}$, the diffusively coupled array becomes a contracting system and can therefore be driven at the ultraharmonic frequency provided by the repulsively coupled array, which after the bifurcation acts like a drive. If $c$ is increased even higher, beyond the bifurcation value of the repulsively coupled array, then total oscillator death in both arrays results. For $c$ below this critical value, but above $c_{bif}$, given by Eqns. (\[eq:general\]) (with $P_{bif}$ given by Eqn. (\[eq:Cbnew\])), high frequency synchronous oscillations are produced in the diffusively coupled array. [xx]{} P.L. Buono, and A. Palacios. A mathematical model of motorneuron dynamics in the heartbeat of the leech. *Physica D*, 188:0 292–313, 2004. J. Guckenheimer, and P. Holmes. Nonlinear Oscillations, Dynamical Systems and Bifurcations of Vector Fields. Volume 2, Springer-Verlag, New York, 1983. A.J. Ijspeert, A. Crespi, D. Ryczko, and J.M. Cabelguen. From swimming to walking with a salamander robot driven by a spinal cord model. *Science*, 315(5817):0 1416–1420, 2007. A.J. Ijspeert. Central pattern generators for locomotion control in animals and robots: a review. *Neural Networks*, 21(4):0 642–653, 2008. A.S. Landsman and I. B. Schwartz. Predictions of ultraharmonic oscillations in coupled arrays of limit cycle oscillators. *Phys. Rev. E*, 74 (036204):0 1–8, 2006. W. Lohmiller and J. J. Slotine. On contraction analysis for nonlinear systems. *Automatica*, 34(6), 1998. A. Palacios et. al. Multifrequency synthesis using two coupled nonlinear oscillator arrays. *Phys. Rev. E*, 72 (026211), 2005. Q.C. Pham and J. J. Slotine. Stable concurrent synchronization in dynamic system network *Neural Networks*, 20(1), 2007. K. Seo and J. J. Slotine. Model for global synchronization in CPG-based locomotion. *IEEE International Conference on Robotics and Automation*, (ISSN:1050-4729):0 281–286, 2007.

--- author: - 'K. Beuermann' - 'F. V. Hessman' - 'S. Dreizler' - 'T. R. Marsh' - 'S. G. Parsons' - 'D. E. Winget' - 'G. F. Miller' - 'M. R. Schreiber' - 'W. Kley' - 'V. S. Dhillon' - 'S. P. Littlefair' - 'C. M. Copperwheat' - 'J. J. Hermes' date: 'Received 26 July 2010 / accepted 6 October 2010' title: 'Two planets orbiting the recently formed post-common envelope binary NN Serpentis' --- Introduction ============ [^1] is a short-period ($P_\mathrm{orb}$=$3.12$hr) eclipsing binary at a distance of 500 pc. The detached system contains a hot hydrogen-rich white dwarf NNSera of spectral type DAO1 and an M4 dwarf star NNSerb with masses of 0.535[M$_\odot$]{} and 0.111[M$_\odot$]{}, respectively [@parsonsetal10a]. With an effective temperature of 57000K [@haefneretal04], the white dwarf has a cooling age of only $10^6$yrs [@wood95]. The present system resulted from a normal binary with a period of $\sim$1 year when the more massive component evolved to a giant and engulfed the orbit of its companion. The subsequent common envelope (CE) phase led to the expulsion of the envelope, laying bare the newly born white dwarf and substantially shortening the orbital period. Some eclipsing post-CE binaries display long-term eclipse time variations, among them V471Tau [@kaminskietal07], QSVir and NNSer [@parsonsetal10b and references therein]. The latter possesses deep and well-defined eclipses, which allow measurements of the mid-eclipse times to an accuracy of 100ms and better [@brinkworthetal06; @parsonsetal10b]. The processes advanced to explain them include the long-term angular momentum loss by gravitational radiation and magnetic braking, possible quasi-periodicities caused, e.g., by Applegate’s (1992) mechanism, and the strict periodicities produced by apsidal motion or the presence of a third body in the system. Finding the correct interpretation requires measurements of high precision and a coordinated effort over a wide range of time scales. The existence of a third body orbiting [NNSerab]{} was previously considered by @qianetal09, but the orbital parameters suggested by them are incompatible with more recent data [@parsonsetal10b]. In this Letter, we present an analysis of the eclipse time variations of [NNSerab]{}, based on published data, the reanalysis of published data, and new measurements obtained over the first half of 2010. The data ======== After their 1988 discovery of deep eclipses in [NNSer]{}, @haefneretal04 acquired a series of accurate mid-eclipse times in 1989. After a hiatus of ten years, they added a potentially very accurate trailed CCD imaging observation using the ESO VLT. From 2002 on, the Warwick group systematically secured a total of 22 mid-eclipse times of high precision [@brinkworthetal06; @parsonsetal10b this work]. @parsonsetal10b list all published mid-eclipse times by other authors until the end of 2009. These are included in our analysis that weights them by their statistical errors. Since the individual Warwick mid-eclipse times between 2002 and 2009 were separated by about one year, information on eclipse time variations on a shorter time scale is lacking. We, therefore, organized a collaborative effort of the Göttingen, McDonald, and Warwick groups to monitor  over the first half of 2010. We used the remotely controlled MONET/North 1.2-m telescope at McDonald Observatory via the MONET internet remote-observing interface, the McDonald 2.1-m telescope, and the ESO 3.5-m NTT. The MONET data were taken in white light, the McDonald data with a BG40 filter, and the NTT observations were acquired with the ULTRACAM high-speed CCD camera equipped with Sloan filters. The mid-eclipse times measured in Sloan u’$\!$, g’$\!$, and i’$\!$ are consistent, and we used the g’ data as the most accurate set for the present purpose. The mid-eclipse time derived by @haefneretal04 from the trailed VLT image of 11 June 1999[^2] is the most variant of the published eclipse time measurements and was assigned a large error of 17s, although this should be a very precise measurement, given the very simple form of the eclipses in NNSer and the use of an 8.2m telescope. We reanalysed the image of 11 June 1999, which started 04:53:05.537 UT with an exposure of 1125.7462s and was taken in good atmospheric conditions. The key issue is the conversion of the track from pixel space to time. Using two independent methods, we found that the original analysis by @haefneretal04 was in error and that the mid-eclipse time can be determined with an accuracy of 0.20s (cycle $E$=30721). We also reanalysed the less accurate data of @pigulskimachalska02 (cycle $E$=33233) by including the effects of the finite integration times. Table 2 lists all previously published, the reanalysed, and the new mid-eclipse times shifted to the solar system barycenter and corrected for leap seconds. The table also gives the statistical errors and the residuals relative to our Model2a, as shown in Fig.2 and discussed in Sect.4, below. The light-travel-time effect in NN Ser ====================================== All measurements of mid-eclipse times of  are displayed in Figs.1 and 2 as $O-C$ values relative to the model-dependent linear ephemerides of the respective fits. Data points with errors $<$1s and $>$1s are shown as green and yellow dots, respectively. The eclipse time measurements dominating the fit are the 1989 data points of @haefneretal04 near the abscissa value JD’= JD-2450000=–2295, the reevaluated VLT point on JD’=1340, the 2002–2009 series of Warwick eclipse times since JD’=2411 [@parsonsetal10b], and the data of this work since JD’=5212. In particular, the revised VLT mid-eclipse time implies a twofold change in the time derivative of $O-C$ and excludes the simple quadratic ephemerides used by @brinkworthetal06 and @parsonsetal10b. The available data do not exclude abrupt period changes or an ultimate aperiodicity, but there is no physical process that predicts such behavior. We consider a periodic behavior the most promising assumption and proceed to explore this possibility. Strictly periodic $O-C$ variations may result from apsidal motion of the binary orbit or an additional body orbiting the binary. Given the parameters of [NNSerab]{}, classical apsidal motion for small eccentricities $e_\mathrm{bin}$ produces a sinusoidally varying time shift with an amplitude $P_\mathrm{bin}e_\mathrm{bin}/\pi$=3577$e_\mathrm{bin}$s [@todoran72]. As a result, $e_\mathrm{bin}$$\sim$0.01 would suffice to produce the observed amplitude. However, the likewise predicted variation of the FWHM of the eclipse and the time shift of the secondary eclipse are not observed [@parsonsetal10b this work]. Furthermore, the observed variation is not sinusoidal and, given an apsidal motion constant for the secondary star NNSerb of $k_{22}\simeq0.11$, the period of the apsidal motion would be as short as $\sim$0.4 years. Such periodicity is not detected (see Fig.2, bottom panel). This leaves us with the third-body hypothesis, at least for the major fraction of the observed eclipse time variations. In general, it would be possible that different physical processes combine to produce the observed signal. We find, however, that a perfect fit within the very small statistical errors can be obtained for a signal that consists of the periodicities produced by two objects orbiting [NNSerab]{}. Guided by Ockham’s razor and the history of discoveries in the Solar system, we consider that a fourth body in the presence of a third one is a natural assumption. One-planet and two-planet fits to the data ========================================== Including the light-travel-time effect of the objects NNSer(ab)c and NNSer(ab)d, the times of mid-eclipse become $$T=T_0\,+\,P_\mathrm{bin}E + \sum_\mathrm{k=c,d} \frac{K_\mathrm{bin,k}\,(1-e_\mathrm{k}^2)}{(1+e_\mathrm{k}\,\mathrm{cos}\,\upsilon_\mathrm{k})}\,\mathrm{sin}\,(\upsilon_\mathrm{k}-\varpi_\mathrm{k}),$$ where time is measured from a fiducial mid-eclipse time $T_0$. A linear binary ephemeris is assumed with $P_\mathrm{bin}$ the orbital period and $E$ the cycle number. The five free parameters for planet $k$ are the orbital period $P_\mathrm{k}$, the eccentricity $e_\mathrm{k}$, the longitude of periastron $\varpi_\mathrm{bin,k}$ measured from the ascending node in the plane of the sky, the time $T_\mathrm{k}$ of periastron passage, and the amplitude of the eclipse time variation $K_\mathrm{k}= a_\mathrm{bin,k}$sin$i_\mathrm{k}$/c, with $a_\mathrm{bin,k}$ the semi-major axis of the orbit of the center of mass of the binary about the common center of mass of the system, $i_\mathrm{k}$ the inclination, and c the speed of light. In the denominator, $\upsilon_\mathrm{k}$ is the true anomaly, which progresses through 2$\pi$ over the orbital period $P_\mathrm{k}$. We explored the multi-dimensional $\,\chi^2$ space of the two-planet model, using the Levenberg-Marquardt routine implemented in IDL and an independent code. The search showed that compensation effects render some parameters ill defined. This uncertainty results, in particular, from the long hiatus between the accurate measurements of 1989 [@haefneretal04] and 1999 (VLT, this work). We selected the best model, therefore, by imposing the additional requirement that the derived orbits be secularly stable. We investigated all solutions permitted by the data with numerical N-body simulations with a variable time step Runge-Kutta integrator, following the orbits over $10^5$yrs, and find that only a narrow range in parameter space corresponds to stable solutions. In what follows, we consider the one-planet and the two-planet models in turn. *Model1* with seven free parameters describes a single planet with eccentricity $e$. The fit requires $e\!\ga$0.60 and is bad for any value of $e$, with a reduced $\chi^2_\nu\!\ge$23.3 ($\chi^2$=1052 for 45 degrees of freedom). The top panel of Fig1 shows the case $e\!=\!0.65$. The residuals based on the statistical errors of the data points (center panel) reach 23 standard deviations and indicate that there is an additional modulation at about half the orbital period. The residuals of the 2010 data (bottom panel) demonstrate the lack of $O-C$ fluctuations on a short time scale. *Model 2* for two planets requires some restriction in parameters, because the grid search yields good fits for a range of eccentricities of the outer planet [$e_\mathrm{c}$]{}, including zero, and for a period ratio $r_\mathrm{p}\!=\!P_\mathrm{c}/P_\mathrm{d}\!=\!1.90\pm0.30$ or $r_\mathrm{p}\!=\!2.50\pm0.15$ (1-$\sigma$ errors), with the former slightly preferred. The dichotomy in [$r_\mathrm{p}$]{} arises from the uncertain phasing of the singular 1989 point relative to the train of the 1999–2010 data. Further minima at still larger [$r_\mathrm{p}$]{} do not exist. Only a small fraction of the parameter space allowed by the fits corresponds to secularly stable orbits, however. Near $r_\mathrm{p}\!\simeq\!2$, orbits with $e_\mathrm{c}\!>\!0.1$ tend to be unstable, while the stability region is broad in the remaining parameters for $e_\mathrm{c}\!\approx\!0.02$. Furthermore, all solutions with $r_\mathrm{p}\!\la\!1.9$ are unstable, with only some solutions stable at $r_\mathrm{p}\!=\!1.9$. The solutions near $r_\mathrm{p}\!=\!2.5$ are more generally stable. We consider Models 2a and 2b, representing the cases of $r_\mathrm{p}\!\simeq\!2.0$ and 2.5, respectively, both with $e_\mathrm{c}\!\equiv\!0$. Model 2a provides the slightly better fit and is shown in Fig.2. It yields $K_\mathrm{c}\!=\!27.4$s, $K_\mathrm{d}\!=\!5.7$s, $P_\mathrm{c}\!=\!15.5$yrs, $P_\mathrm{d}\!=\!7.75$yrs, and $e_\mathrm{d}\!=\!0.20$ with $\chi^2_\nu=0.78~(\,\chi^2\!=\!32.9$ for 42 d.o.f.). Periastron passage of NNSer(ab)d time NNSer(ab)c was at longitude $213^\circ$. For the low value of $e_\mathrm{c}\!=\!0.03$, a shallow minimum of [$\,\chi^2$]{} is attained for aligned apses. From the present data, we cannot infer the true value of [$r_\mathrm{p}$]{} with certainty, but it is intriguing that objects c and d may be locked in either the 2:1 resonance, found also in other planetary systems, or the 5:2 resonance. The parameters for Models 2a and 2b are listed in Table1, together with their 1-$\sigma$ errors. A simpler model with two circular orbits reaches only $\chi^2_\nu=1.96~(\,\chi^2\!=\!86.2$ for 44 d.o.f.) at $r_\mathrm{p}\!=\!2.46$ and can be excluded. ---------------- --------- ----------- --------------------------- --------------------------- ---------------------------------------- --------------------- --------------------------- --------------------------- --------------------------- -------------------------- -------------------------- ------------------------------------- ------------------------------------- ---------------- --------------------- \[-1ex\] \[-1ex\] Model Planets Number [$P_\mathrm{c}$]{}  [$P_\mathrm{d}$]{}  [$P_\mathrm{c}$]{}/[$P_\mathrm{d}$]{}  [$e_\mathrm{c}$]{}  [$e_\mathrm{d}$]{}  [$a_\mathrm{c}$]{}  [$a_\mathrm{d}$]{}  [$\varpi_\mathrm{c}$]{}  [$\varpi_\mathrm{d}$]{}  [$M_\mathrm{c}$]{}sin$i_\mathrm{c}$ [$M_\mathrm{d}$]{}sin$i_\mathrm{d}$ [$\,\chi^2$]{} [$\,\chi^2_\nu$]{}  free par. (yrs) (yrs) (AU) (AU) ($^\circ$) ($^\circ$) ([M$_\mathrm{Jup}$]{}) ([M$_\mathrm{Jup}$]{}) \[1ex\] \[-1ex\] 1 1 2+5 22.60 $\ga\!0.65$ 6.91 8.0 8.36 1052.3 23.38 \[0.5ex\] 2a 2 2+8 15.50 7.75 $2.00$ $\equiv 0.0$ 0.20 5.38 3.39 74 6.91 2.28 32.9 0.78 $\hspace{-0.8mm}\pm 0.45$ $\hspace{-1.8mm}\pm 0.35$ $\hspace{-1.8mm}\pm 0.15$ $\hspace{-1.8mm}\pm 0.02$ $\hspace{-1.8mm}\pm 0.20$ $\hspace{-1.5mm}\pm 0.10$ $\hspace{-0.5mm}\pm 4$ $\hspace{-2.0mm}\pm 0.54$ $\hspace{-1.8mm}\pm 0.38$ \[0.5ex\] 2b 2 2+8 16.73 6.69 2.50 $\equiv 0.0$ 0.23 5.66 3.07 73 5.92 1.60 33.8 0.80 $\hspace{-0.8mm}\pm 0.26$ $\hspace{-1.8mm}\pm 0.40$ $\hspace{-1.8mm}\pm 0.15$ $\hspace{-1.8mm}\pm 0.04$ $\hspace{-2.2mm}\pm 0.06$ $\hspace{-1.5mm}\pm 0.13$ $\hspace{-0.5mm}\pm 7$ $\hspace{-2.0mm}\pm 0.40$ $\hspace{-1.8mm}\pm 0.27$ \[1ex\] \[-3ex\] ---------------- --------- ----------- --------------------------- --------------------------- ---------------------------------------- --------------------- --------------------------- --------------------------- --------------------------- -------------------------- -------------------------- ------------------------------------- ------------------------------------- ---------------- --------------------- Using Model2a as input to our N-body simulations, we find that [$e_\mathrm{c}$]{} and [$e_\mathrm{d}$]{} oscillate around 0.02 and 0.22 with amplitudes of 0.02 and 0.05, respectively. The difference $\Delta \varpi$ of the periastron longitudes circulates on a time scale of 400yrs. The periods perform small-amplitude anti-phased oscillations, which cause [$r_\mathrm{p}$]{} to oscillate between 1.9 and 2.2. Even if the two planets are secularly locked in the 2:1 mean motion resonance, therefore, the observed period ratio at any given time may deviate slightly from its nominal value. For Model 2a, the best-fit binary ephemeris is $T$= BJED 2,447344.524425(40)+0.1300801419(10)$E$, where the errors refer to the last digits. Adding a quadratic term $BE^2$ to the ephemeris does not improve the two-planet fit and yields a 1-$\sigma$ limit of $\mid\!B\!\mid\,<\,1.5\,10^{-13}$ days, leaving room for a period change by gravitational radiation or a long-term activity-related effect [@brinkworthetal06; @parsonsetal10b]. Discussion ========== The large amplitude of the $O-C$ eclipse time variations in [NNSer]{} can only be explained by a third body in the system, while the still substantial residuals from a single-planet fit could, in principle, have a different origin from that of a fourth body. The two-planet model, however, possesses the beauty of simplicity, and the fact that the residuals for the entire data set vanish simultaneously imposes tight restrictions on any other mechanism. In particular, the lack of short-term variability of the residuals in the first half of 2010 argues against any process that acts on a short time scale or leads to erratic eclipse time variations. Hence, there is strong evidence for two planets orbiting [NNSerab]{}. With masses $M_\mathrm{c}\mathrm{sin}\,i_\mathrm{c}\!\simeq\!6$[M$_\mathrm{Jup}$]{} and $M_\mathrm{d}\mathrm{sin}\,i_\mathrm{d}\!\simeq\!2$[M$_\mathrm{Jup}$]{}, NNSer(ab)c and NNSer(ab)d both qualify as giant planets for all inclinations $i_\mathrm{\,c}>$28$^\circ$ and $i_\mathrm{\,d}>$9$^\circ$, respectively. The probable detection of resonant motion with a period ratio of either 2:1 or 5:2 is a major bonus, which adds to the credence of the two-planet model. It is the second planetary system found by eclipse timing, after HWVir [@leeetal09]. Given a pair of planets orbiting a post-CE binary, two formation scenarios are possible. They could either be old first-generation planets that formed in a circumbinary protoplanetary disk or they could be young second-generation planets formed $\la\!10^6$ yrs ago in a disk that resulted from the CE [@perets10]. To evaluate both scenarios, we have reconstructed the CE evolution of NNSerab using the improved algorithm by @zorotovicetal10, who constrain the CE efficiency to a range $\alpha\simeq0.2 - 0.3$. Possible solutions for the progenitor binary of NNSerab are not very sensitive to $\alpha$: for $\alpha=0.25$, the progenitor was a giant of 2.08[M$_\odot$]{} and radius 194[R$_\odot$]{} with the present secondary star at a separation of 1.44AU. When the CE engulfed the secondary star, dynamic friction caused the latter to spiral in rapidly, thereby dramatically decreasing the binary separation to the current 0.0043AU. Stability arguments imply that any planet from the pre-CE phase must have formed with semi-major axes exceeding 3.5AU [@holmanwiegert99]. With three quarters of the central mass expelled in the CE event, pre-existing planets would move outward or may even be lost from the system. However, given a sufficiently dense and slowly expanding CE, the dynamical force experienced by them may have ultimately moved them inward [@alexanderetal76]. Since the drag primarily affects the more massive and more slowly moving outer planet, such a scenario could lead to resonant orbits, so a first-generation origin appears possible. The alternative post-CE origin in a second-generation of planet formation is also possible, since the formation of circumbinary disks is a common phenomenon among post-AGB binary stars and the concentration of a slow, dusty wind to the orbital plane of the binary is thought to favor the formation of planets [e.g. @vanwinckeletal09; @perets10]. In particular the tiny separation of the present binary poses no problem for stable orbits of second-generation planets even at significantly shorter distances than the inner planet that we have detected [@holmanwiegert99]. A particularly intriguing aspect of a second-generation origin of the planets in NN Ser would be their extreme youth, equal to or less than the $10^6$yrs cooling age of the white dwarf [@wood95]. This feature would distinguish them from all known exoplanets and may ultimately lead to their direct detection. While we cannot presently prove a second-generation origin for these planets, modeling the CE event may allow us to distinguish between the two scenarios. We would like to thank Dr. Reinhold Haefner for information concerning the original VLT observations and analyses and Dr. Andrzej Pigulski for sending us his original photometry. This work is based on data obtained with the MONET telescopes funded by the “Astronomie & Internet” program of the Alfried Krupp von Bohlen und Halbach Foundation, Essen, on observations with the ESO NTT under ESO programme 085.D-0541, and on data obtained from the ESO/ST-ECF Science Archive Facility. TRM, VSD, CMC, and SPL acknowledge grant support from the UK’s STFC. MRS acknowledges support from FONDECYT under grant number 1061199 (MRS) and the Centre of Astrophysics Valpara[í]{}so. DEW acknowledges the support of the Norman Hackerman Advanced Research Program under grant 003658-0255-2007. [r@c@c@c@r@r@c@l]{}\ E & BJD(TT) & Error & Residual & Error & Residual & References & Comment\ & JD2400000$+$&(days) & (days) & (s) & (s)\ \ 0 & 47344.5246635 & 0.0003500 & 0.0000290 & 30.00 & 2.51 & (1) & Reanalysed\ 2760 & 47703.5457436 & 0.0000020 & 0.0000012 & 0.17 & 0.10 & (2) &\ 2761 & 47703.6758326 & 0.0000060 & 0.0000101 & 0.52 & 0.87 & (2) &\ 2769 & 47704.7164596 & 0.0000030 & $-$0.0000038 & 0.26 & $-$0.33 & (3) &\ 2776 & 47705.6270226 & 0.0000030 & $-$0.0000016 & 0.26 & $-$0.14 & (3) &\ 2777 & 47705.7571046 & 0.0000070 & 0.0000003 & 0.60 & 0.03 & (3) & Corrected\ 2831 & 47712.7815836 & 0.0001500 & 0.0001534 & 12.96 & 13.25 & (2) &\ 2839 & 47713.8222336 & 0.0001500 & 0.0001625 & 12.96 & 14.04 & (2) &\ 7360 & 48301.9141954 & 0.0001500 & $-$0.0000627 & 12.96 & -5.42 & (2) &\ 28152 & 51006.5405495 & 0.0002000 & 0.0000605 & 17.28 & 5.23 & (2) &\ 30721 & 51340.7165402 & 0.0000023 & $-$0.0000004 & 0.20 & $-$0.03 & (2) & Reanalysed\ 33233 & 51667.4780058 & 0.0000960 & 0.0000041 & 8.29 & 0.35 & (4) & Reanalysed\ 38960 & 52412.4470566 & 0.0000006 & $-$0.0000006 & 0.05 & $-$0.05 & (5) &\ 38961 & 52412.5771382 & 0.0000005 & 0.0000008 & 0.04 & 0.07 & (5) &\ 38968 & 52413.4876977 & 0.0000009 & $-$0.0000006 & 0.08 & $-$0.05 & (5) & Corrected\ 38976 & 52414.5283389 & 0.0000007 & $-$0.0000004 & 0.06 & $-$0.03 & (5) &\ 38984 & 52415.5689804 & 0.0000007 & 0.0000000 & 0.06 & 0.00 & (5) &\ 41782 & 52779.5331703 & 0.0000015 & 0.0000001 & 0.13 & 0.01 & (5) &\ 41798 & 52781.6144523 & 0.0000007 & 0.0000002 & 0.06 & 0.02 & (5) &\ 41806 & 52782.6550927 & 0.0000008 & $-$0.0000004 & 0.07 & $-$0.03 & (5) &\ 41820 & 52784.4762150 & 0.0000008 & 0.0000003 & 0.07 & 0.03 & (5) &\ 44472 & 53129.4486808 & 0.0000040 & 0.0000008 & 0.35 & 0.07 & (5) &\ 44473 & 53129.5787632 & 0.0000028 & 0.0000031 & 0.24 & 0.27 & (5) &\ 44474 & 53129.7088370 & 0.0000017 & $-$0.0000032 & 0.15 & $-$0.28 & (5) &\ 44480 & 53130.4893234 & 0.0000030 & 0.0000025 & 0.26 & 0.22 & (5) &\ 49662 & 53804.5644567 & 0.0000025 & 0.0000001 & 0.22 & 0.01 & (5) &\ 49663 & 53804.6945350 & 0.0000012 & $-$0.0000017 & 0.10 & $-$0.15 & (5) &\ 49671 & 53805.7351781 & 0.0000006 & 0.0000005 & 0.05 & 0.04 & (5) &\ 53230 & 54268.6903114 & 0.0000006 & 0.0000008 & 0.05 & 0.07 & (5) &\ 53237 & 54269.6008713 & 0.0000002 & $-$0.0000001 & 0.02 & $-$0.01 & (5) &\ 56442 & 54686.5076279 & 0.0000009 & $-$0.0000001 & 0.08 & $-$0.01 & (5) &\ 58638 & 54972.1634971 & 0.0000800 & $-$0.0000380 & 6.91 & -3.28 & (6) &\ 58645 & 54973.0740553 & 0.0001000 & $-$0.0000406 & 8.64 & -3.51 & (6) &\ 58684 & 54978.1471791 & 0.0001200 & $-$0.0000408 & 10.37 & -3.53 & (6) &\ 58745 & 54986.0820789 & 0.0001200 & $-$0.0000274 & 10.37 & -2.37 & (6) &\ 58753 & 54987.1228359 & 0.0001300 & 0.0000887 & 11.23 & 7.66 & (6) &\ 58796 & 54992.7161925 & 0.0000015 & 0.0000008 & 0.13 & 0.07 & (6) &\ 60489 & 55212.9418187 & 0.0000069 & 0.0000027 & 0.60 & 0.23 & (7,8) &\ 60505 & 55215.0230961 & 0.0000066 & $-$0.0000017 & 0.57 & $-$0.15 & (7,8) &\ 60528 & 55218.0149380 & 0.0000043 & $-$0.0000024 & 0.37 & $-$0.21 & (7,8) &\ 60735 & 55244.9415254 & 0.0000029 & 0.0000012 & 0.25 & 0.10 & (7,8) &\ 60743 & 55245.9821654 & 0.0000032 & 0.0000003 & 0.28 & 0.03 & (7,8) &\ 60751 & 55247.0228063 & 0.0000034 & 0.0000002 & 0.29 & 0.02 & (7,8) &\ 60774 & 55250.0146469 & 0.0000034 & $-$0.0000018 & 0.29 & $-$0.16 & (7,8) &\ 60927 & 55269.9169047 & 0.0000014 & $-$0.0000018 & 0.12 & $-$0.16 & (7,9) &\ 60950 & 55272.9087487 & 0.0000013 & $-$0.0000005 & 0.11 & $-$0.04 & (7,9) &\ 61219 & 55307.9003015 & 0.0000010 & 0.0000005 & 0.09 & 0.04 & (7,10) &\ 61426 & 55334.8268834 & 0.0000018 & $-$0.0000025 & 0.16 & $-$0.22 & (7,9) &\ 61440 & 55336.6480059 & 0.0000018 & $-$0.0000017 & 0.16 & $-$0.15 & (7,9) &\ 61441 & 55336.7780894 & 0.0000015 & 0.0000017 & 0.13 & 0.15 & (7,9) &\ 61564 & 55352.7779443 & 0.0000016 & 0.0000017 & 0.14 & 0.15 & (7,9) &\ 61579 & 55354.7291448 & 0.0000009 & 0.0000004 & 0.08 & 0.03 & (7,10) &\ \[1.5ex\]\ \ (1) Haefner, R., 1989, ESO Msngr, 55, 61, reanalysed using up-to-date eclipse profile; (2) Haefner et al. (2004), misprint for E=2777 corrected, VLT trailed imaging observation (E=30721) reanalysed using the original data; (3) Wood, J. H. & Marsh, T. R., 1991, ApJ, 381, 551; (4) Pigulski & Michalska (2002), reanalysed using the original data; (5) Parsons et al. (2010b), timing for E=38968 corrected for misprint; (6) Qian et al. (2009); (7) This work, (8) MONET/North 1.2-m white light photometry, (9) McDonald 2.1-m photometry with Schott BG40 filter, (10) ESO NTT 3.5-m ULTRACAM Sloan g’ photometry. [^1]: On recommendation by the Editor of A&A, we refer to the system as NNSer, to the binary explicitly as NNSerab, and to the objects orbiting the binary as NNSer(ab)c and NNSer(ab)d. [^2]: http://www.eso.org/public/images/eso9936b/

--- abstract: 'Autonomous vehicles commonly rely on highly detailed birds-eye-view maps of their environment, which capture both static elements of the scene such as road layout as well as dynamic elements such as other cars and pedestrians. Generating these map representations on the fly is a complex multi-stage process which incorporates many important vision-based elements, including ground plane estimation, road segmentation and 3D object detection. In this work we present a simple, unified approach for estimating maps directly from monocular images using a single end-to-end deep learning architecture. For the maps themselves we adopt a semantic Bayesian occupancy grid framework, allowing us to trivially accumulate information over multiple cameras and timesteps. We demonstrate the effectiveness of our approach by evaluating against several challenging baselines on the NuScenes and Argoverse datasets, and show that we are able to achieve a relative improvement of 9.1% and 22.3% respectively compared to the best-performing existing method. [^1]' author: - | Thomas Roddick\ University of Cambridge\ [tr346@cam.ac.uk]{} - | Roberto Cipolla\ University of Cambridge\ [rc10001@cam.ac.uk]{} bibliography: - 'egbib.bib' title: | Predicting Semantic Map Representations from Images\ using Pyramid Occupancy Networks --- Introduction ============ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![An example prediction from our algorithm. Given a set of surround-view images, we predict a full 360[$^{\circ}$ ]{}birds-eye-view semantic map, which captures both static elements like road and sidewalk as well as dynamic actors such as cars and pedestrians.[]{data-label="fig:headline"}](figures/images_anno.png "fig:"){width="\linewidth"} ![An example prediction from our algorithm. Given a set of surround-view images, we predict a full 360[$^{\circ}$ ]{}birds-eye-view semantic map, which captures both static elements like road and sidewalk as well as dynamic actors such as cars and pedestrians.[]{data-label="fig:headline"}](figures/headline-right.pdf "fig:"){width="\linewidth"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Autonomous vehicles and other robotic platforms require a rich, succinct and detailed representation of their environment which captures both the geometry and layout of the static world as well as the pose and dimensions of other dynamic agents. Such representations often provide the foundation for all decision making, including path planning, collision avoidance and navigation. Rather than capturing the full 3D world in its entirety, one popular solution is to represent the world in the form of a birds-eye-view (BEV) map, which provide a compact way to capture the spatial configuration of the scene. Such maps are convenient in that they are simple to visualise and process, exploiting the fact that in many scenarios the essential information for navigation is largely confined to the ground plane. Construction of birds-eye-view maps is however at present a complex multistage processing pipeline, involving the composition of multiple fundamental machine vision tasks: structure from motion, ground plane estimation, road segmentation, lane detection, 3D object detection, and many more. Intuitively, all these tasks are related: knowing the layout of the road ought to inform us about where in the image we should look for cars; and similarly a car emerging from behind a building may indicate the presence of a hidden side road beyond. There seems to be a clear impetus towards replacing this complicated pipeline with a simple end-to-end approach which is able to reason holistically about the world and predict the desired map representation directly from sensor observations. In this work we focus on the particularly challenging scenario of BEV map estimation from monocular images alone. Given the high cost and limited resolution of LiDAR and radar sensors, the ability to build maps from image sensors alone is likely to be crucial to the development of robust autonomous vehicles. Whilst a number of map representations are possible, we choose to represent the world using a probabilistic occupancy grid framework. Occupancy grid maps [@elfes1990occupancy] are widely used in robotics, and allow us to trivially incorporate information over multiple sensors and timesteps. Unlike other map representations, their grid-based structure also makes them highly agreeable to processing by convolutional neural networks, allowing us to take advantage of powerful developments from the deep learning literature. In this work we extend the traditional definition of occupancy grids to that of a semantic occupancy grid [@lu2019monocular], which encodes the presence or absence of an object category at each grid location. Our objective is then to predict the probability that each semantic class is present at each location in our birds-eye-view map. The contributions of this paper are as follows: 1. We propose a novel dense transformer layer which maps image-based feature maps into the birds-eye-view space. 2. We design a deep convolutional neural network architecture, which includes a pyramid of transformers operating at multiple image scales, to predict accurate birds-eye-view maps from monocular images. 3. We evaluate our approach on two large-scale autonomous driving datasets, and show that we are able to considerably improve upon the performance of leading works in the literature. We also qualitatively demonstrate how a Bayesian semantic occupancy grid framework can be used to accumulate map predictions across multiple cameras and timesteps to build a complete model of a scene. The method is fast enough to be used in real time applications, processing 23.2 frames per second on a single GeForce RTX 2080 Ti graphics card. Related Work ============ #### Map representations for autonomous driving High definition birds-eye-view maps have been shown to be an extremely powerful representation across a range of different driving tasks. In 3D object detection, [@yang2018hdnet] use ground height prior information from maps to improve the quality of input LiDAR point clouds. [@ma2019exploiting] correlate visual observations with sparse HD map features to perform highly accurate localisation. Birds-eye-view maps are particularly valuable in the context of prediction and planning given their metric nature: [@djuric2018motion] and [@casas2018intentnet] render the local environment as a rasterised top-view map representation, incorporating road geometry, lane direction, and traffic agents, and use this representation to predict future vehicle trajectories. A similar representation is used by [@bansal2018chauffeurnet] as input to their imitation learning pipeline, allowing an autonomous agent to drive itself by recursively predicting its future state. [@hecker2018end] augment their camera-based end-to-end driving model with a rendered map view from a commercial GPS route planner and show that this significantly improves driving performance. #### Top-down representations from images A number of prior works have tackled the difficult problem of predicting birds-eye-view representations directly from monocular images. A common approach is to use inverse perspective mapping (IPM) to map front-view image onto the ground plane via a homography [@ammar2019geometric; @lin2012vision]. [@zhu2018generative] use a GAN to refine the resulting predictions. Other works focus on the birds-eye-view object detection task, learning a to map 2D bounding box detections to the top-down view[@palazzi2017learning; @wang2019monocular], or predicting 3D bounding boxes directly in the birds-eye-view space [@roddick2019orthographic]. Relatively few works however have tackled the more specific problem of generating semantic maps from images. Some use the IPM approach mentioned above to map a semantic segmentation of the image plane into the birds-eye-view space [@deng2019restricted; @samann2018efficient], an approach which works well for estimating local road layout but which fails for objects such as cars and pedestrians which lie above the ground plane. [@henriques2018mapnet] take advantage of RGB-D images to learn an implicit map representation which can be used for later localisation. The VED method of [@lu2019monocular] uses a variational encoder-decoder network to predict a semantic occupancy grid directly from an image. The use of a fully-connected bottleneck layer in the network however means that much of the spatial context in the network is lost, leading to an output which is fairly coarse and is unable to capture small objects such as pedestrians. [@pan2019cross] adopt a similar approach, predicting a birds-eye-view semantic segmentation from a stack of surround view images, via a fully-connected view-transformer module. [@schulter2018learning] propose to use an in-painting CNN to infer the semantic labels and depth of the scene behind foreground objects, and generate a birds-eye-view by projecting the resulting semantic point cloud onto the ground plane. Unfortunately, given the lack of available ground truth data, many of the above methods are forced to rely on weak supervision from stereo [@lu2019monocular], weakly-aligned map labels [@schulter2018learning] or synth-to-real domain transfer [@schulter2018learning; @pan2019cross]. Training on real data is crucial to performance in safety critical systems, and we believe we are the first to do so using a directly supervised approach. {width="\linewidth"} Semantic occupancy grid prediction ================================== In this work we represent the state of the world as a birds-eye-view semantic occupancy grid map. Occupancy grid maps [@elfes1990occupancy] are a type of discrete random field where each spatial location $x_i$ has an associated state $m_i$, which may be either occupied ($m_i=1$), or free ($m_i=0$). In practice, the true state of the world is unknown, so we treat $m_i$ as a random variable, and estimate the probability of occupancy $p(m_i|z_{1:t})$, conditioned on a set of observations $z_t$. The occupancy grid formulation may be further extended to that of a semantic occupancy grid, where instead of generic cell occupancy, the state $m_i^c$ represents the presence or absence of an object of class $c$ in a given grid cell. These occupancies are non-exclusive: for example road, crossing and vehicle classes may conceivably coexist at the same location. Traditionally in occupancy grid mapping, the occupancy probabilities $p(m_i|z_{t}$) are estimated using an inverse sensor model, often a simple hand-engineered function which maps from range sensor readings to occupancy probabilities based on sensor characteristics. In our application, observations take the form of images and cell occupancies capture high-level semantic knowledge of the scene. We therefore propose to train a deep CNN-based inverse sensor model $p(m_i^c|z_t) = f_\theta(z_t, x_i)$ which learns to predict occupancy probabilities from a single monocular input image. Our objective is therefore to predict a set of multiclass binary labels at each location on a 2D birds-eye-view image. This scenario bears many similarities to the widely-studied computer vision problem of semantic segmentation. What makes this task particularly challenging however is the fact that the input and output representations exist within entirely different coordinate systems: the former in the perspective image space, and the latter in the orthographic birds-eye-view space. We therefore propose a simple transformer layer, which makes use of both camera geometry and fully-connected reasoning to map features from the image to the birds-eye-view space. We incorporate this dense transformer layer as part of our deep Pyramid Occupancy Network (PyrOccNet). The pyramid occupancy network consists of four main stages. A **backbone feature extractor** generates multiscale semantic and geometric features from the image. This is then passed to an FPN [@lin2017feature]-inspired **feature pyramid** which upsamples low-resolution feature-maps to provide context to features at higher resolutions. A stack of **dense transformer layers** together map the image-based features into the birds-eye view, which are processed by the **topdown network** to predict the final semantic occupancy grid probabilities. An overview of the approach is shown in [Figure \[fig:arch\]]{}. Losses {#sec:loss} ------ We train our network using a combination of two loss functions. The binary cross entropy loss encourages the predicted semantic occupancy probabilities $p(m_i^c|z_t)$ to match the ground truth occupancies $\hat{m}_i^c$. Given that our datasets includes many small objects such as pedestrians, cyclists and traffic cones, we make use of a balanced variant of this loss, which up-weights occupied cells belonging to class $c$ by a constant factor $\alpha^c$: $${\mathcal{L}_{xent}} = \alpha^c \hat{m}_i^c \log p(m_i^c|z_t) + (1-\alpha^c)(1-\hat{m}_i^c)\log\left(1-p(m_i^c|z_t)\right)$$ Neural networks are however renowned for routinely predicting high probabilities even in situations where they are highly uncertain. To encourage the networks to predict high uncertainty in regions which are known to be ambiguous, we introduce a second loss, which maximises the entropy of the predictions, encouraging them to fall close to 0.5: $${\mathcal{L}_{uncert}} = 1 - p(m_i^c|z_t)\log_2 p(m_i^c|z_t)$$ We apply this maximum entropy loss only to grid cells which are not visible to the network, either because they fall outside field of view of the image, or because they are completely occluded (see [Section \[sec:data\]]{} for details). We ignore the cross entropy loss in these regions. The overall loss is given by the sum of the two loss functions: $${\mathcal{L}_{total}} = {\mathcal{L}_{xent}} + \lambda{\mathcal{L}_{uncert}}$$ where $\lambda=0.001$ is a constant weighting factor. Temporal and sensor data fusion {#sec:fusion} ------------------------------- The Bayesian occupancy grid formulation provides a natural way of combining information over multiple observations and multiple timesteps using a Bayesian filtering approach [@thrun2005probabilistic]. Consider an image observation $z_t$ taken by a camera with extrinsic matrix $M_t$. We begin by converting our occupancy probabilities $p(m_i^c|z_t)$ into a log-odds representation $$l_{i, t}^c = \log\frac{p(m_i^c|z_t)}{1-p(m_i^c|z_t)}$$ which conveniently is equivalent to the network’s pre-sigmoid output activations. The combined log-odds occupancies over observations $1$ to $t$ is then given by $$l_{i, 1:t}^c = l_{i, 1:t-1}^c + l_{i, t}^c - l_0^c$$ from which the occupancy probabilities after fusion can be recovered by applying the standard sigmoid function $$p(m_i^c|z_{1:t}) = \frac{1}{1 + \exp\left(-l_{i, 1:t}^c\right)}$$ The log-odds value $l_0^c$ represents the prior probability of occupancy for class $c$: $$l_0^c = \frac{p(m_i^c)}{1-p(m_i^c)}$$ To obtain the occupancy probabilities in the global coordinate system, we resample the output from our network, which predicts occupancies in the local camera-frame coordinate system, into the global frame using the extrinsics matrix $M_t$, i.e. $p(m_i|z_t) = f_\theta(z_t, M_t^{-1}x_i)$. This approach is used in [Section \[sec:fusion-results\]]{} both to combine sensory information from a set of surround view cameras, and also to fuse occupancy grids over a 20s duration sequence of observations. Dense transformer layer {#sec:transform} ----------------------- One of the fundamental challenges of the occupancy grid prediction task is that the input and output exist in two entirely disparate coordinate systems: the perspective image space and the orthographic birds-eye-view space. To overcome this problem, we introduce a simple transformation layer, which is depicted in [Figure \[fig:transform\]]{}. Our objective is to convert from an image plane feature map with $C$ channels, height $H$ and width $W$, to a feature map on the birds-eye-view plane with $C$ channels, depth $Z$ and width $X$. The dense transformer layer is inspired by the observation that while the network needs a lot of vertical context to map features to the birds-eye-view (due to occlusion, lack of depth information, and the unknown ground topology), in the horizontal direction the relationship between BEV locations and image locations can be established using simple camera geometry. Therefore, in order to retain the maximum amount of spatial information, we collapse the vertical dimension and channel dimensions of the image feature map to a bottleneck of size $B$, but preserve the horizontal dimension $W$. We then apply a 1D convolution along the horizontal axis and reshape the resulting feature map to give a tensor of dimensions $C\times Z\times W$. However this feature map, which is still in image-space coordinates, actually corresponds to a trapezoid in the orthographic birds-eye-view space due to perspective, and so the final step is to resample into a Cartesian frame using the known camera focal length $f$ and horizontal offset $u_0$. ![Our dense transformer layer first condenses the image-based features along the vertical dimension, whilst retaining the horizontal dimension. We then predict a set of features along the depth axis in a polar coordinate system, which are then resampled to Cartesian coordinates.[]{data-label="fig:transform"}](figures/Transformer.pdf){width="\linewidth"} Multiscale transformer pyramid {#sec:multiscale} ------------------------------ The resampling step described in [Section \[sec:transform\]]{} involves, for a row of grid cells a distance $z$ away from the camera, sampling the polar feature map at intervals of $$\Delta u = \frac{f\Delta x}{sz}$$ where $\Delta x$ is the grid resolution and $s$ is the downsampling factor of input feature map with respect to the image. The use of a constant factor for $s$ however is problematic: features corresponding to grid cells far from the camera will be blurred whilst those close to the camera will be undersampled and aliasing can occur. We therefore propose to apply multiple transformers, acting on a pyramid of feature maps with downsampling factors $s_k = 2^{k+3}, k\in{\{0, ..., 4\}}$. The $k^{th}$ transformer generates features for a subset of depth values, ranging from $z_k$ to $z_{k-1}$, where $z_k$ is given by $$z_k = \frac{f\Delta x}{s_k}.$$ Values of $z_k$ for a typical camera and grid setting are given in [Table \[tab:depth\]]{}. The final birds-eye-view feature map is then constructed by concatenating the outputs of each individual transformer along the depth axis. One downside of this approach is that at high resolutions the height of the feature maps $H_k$ can become very large, which leads to an excessive number of parameters in the corresponding dense transformer layer. In practice however, we can crop the feature maps to a height $$H_k = f\frac{y_{max} - y_{min}}{s_kz_k}$$ corresponding to a fixed vertical range between $y_{min}$ and $y_{max}$ in the world space. This means that the heights of the cropped feature maps stay roughly constant across scales. The feature maps are taken from the outputs of each residual stage in our backbone network, from conv3 to conv7. To ensure that the high resolution feature maps still encompass a large spatial context, we add upsampling layers from lower resolutions in the style of [@lin2017feature]. $k$ 0 1 2 3 4 -------------- ------- ------- ------- ------- ------- $s_k$ 8 16 32 64 128 $z_{k}$ (m) 39.0 19.5 9.0 4.5 1.0 ResNet layer conv3 conv4 conv5 conv6 conv7 : Depth intervals for each layer of the feature pyramid. \[tab:depth\] Experimental Setup ================== Datasets -------- We evaluate our approach against two large-scale autonomous driving datasets. The NuScenes dataset [@caesar2019nuscenes] consists of 1000 short video sequences captured from four locations in Boston and Singapore. It includes images captured from six calibrated surround-view cameras, 3D bounding box annotations for 23 object categories and rich semantic map annotations which include vectorised representations of lanes, traffic lights, sidewalks and more. From these we select a subset of four map categories which can feasibly be estimated from images, along with ten object categories. The Argoverse 3D dataset [@chang2019argoverse] is comprised of 65 training and 24 validation sequences captured in two cities, Miami and Pittsburg, using a range of sensors including seven surround-view cameras. Like NuScenes, the Argoverse dataset provides both 3D object annotations from 15 object categories, as well as semantic map information including road mask, lane geometry and ground height. From these we choose 7 object categories which contain sufficient training examples, along with the driveable road mask. As both NuScenes and Argoverse are predominantly object detection rather than map prediction datasets, the default dataset splits contain multiple road segments which appear in both the training and validation splits. We therefore redistribute the train/val sequences to remove any overlapping segments, taking care to ensure a balanced distribution over locations, objects and weather conditions. Method Drivable\* Vehicle\* Pedest.\* Large veh. Bicycle\* Bus\* Trailer Motorcy.\* Mean CS Mean ------------------------ ------------ ----------- ----------- ------------ ----------- ---------- --------- ------------ ---------- ---------- IPM 43.7 7.5 1.5 - 0.4 7.4 - 0.8 - 10.2 Depth Unproj. 33.0 12.7 3.3 - 1.1 **20.6** - 1.6 - 12.1 VED [@lu2019monocular] 62.9 14.0 1.0 3.9 0.0 12.3 1.3 0.0 11.9 15.0 VPN [@pan2019cross] 64.9 23.9 6.2 9.7 0.9 3.0 0.4 1.9 13.9 16.8 Ours - baseline 58.5 23.4 3.9 5.2 0.5 11.0 0.4 1.9 13.1 16.5 Ours - D 63.8 27.9 4.8 8.8 1.0 11.0 0.0 3.4 15.1 18.7 Ours - D+P **65.9** 30.7 7.3 10.2 1.7 9.3 **1.7** 2.2 16.1 19.5 Ours - D+P+T 65.4 **31.4** **7.4** **11.1** **3.6** 11.0 0.7 **5.7** **17.0** **20.8** Data generation {#sec:data} --------------- The NuScenes and Argoverse datasets provide ground truth annotations in the form of vectorised city-level map labels and 3D object bounding boxes. We convert these into ground truth occupancy maps by first mapping all vector annotations into the coordinate system of the $t^{th}$ sample using the camera extrinsic matrix $M_t$ provided by the datasets. We then rasterise each annotation to a binary image in the birds-eye-view, which lies on a grid extending 50m in front of the given camera and 25m to either side, at a resolution of 25cm per pixel. For the case of object annotations, we first project the 3D bounding box onto the xz-plane to obtain a 2D polygon. The result of this process is a stack of binary images, which represent the ground truth occupancies for each semantic category $c$ as observed from camera $t$. The resulting labels however represent a close to impossible task for the network, since some grid cell locations lie outside the camera field of view (FoV) or are completely occluded by other objects. We therefore generate an additional binary mask indicating whether each grid cell is visible. A cell is treated as visible if it is within the FoV and has at least one LiDAR ray passing through it (i.e. not blocked by a closer object). Baselines {#sec:baseline} --------- #### Published methods In order to demonstrate the effectiveness of our approach, we compare against two previously published works: the Variational Encoder-Decoder (VED) of Lu [ *et al.*]{} [@lu2019monocular], and the View Parsing Network (VPN) of Pan [ *et al.*]{} [@pan2019cross]. These networks presume different input and output dimensions, so we make minor architectural changes which we detail in Section A of the supplementary material. #### Inverse Perspective Mapping (IPM) We present a simple baseline inspired by other works [@deng2019restricted; @samann2018efficient] of mapping an image-based semantic segmentation to the ground plane via a homography. The image-level segmentation is computed using a state-of-the-art DeepLabv3 [@chen2017rethinking] network, pretrained on Cityscapes [@cordts2016cityscapes], which shares many classes in common with both NuScenes and Argoverse. The ground planes are obtained either by fitting a plane to LiDAR points in the case of NuScenes, or using the precomputed ground heights provided by Argoverse. Note that this information would not be available to a real monocular system at test time, making this baseline additionally competitive. #### Depth-based unprojection Another intuitive solution to this problem would be to use a monocular depth estimator to generate a 3D point cloud from the image, and then drop the z-axis to transfer image-based semantic labels onto the ground plane. As an upper-bound on the performance of this type of approach, we use ground truth depth computed by densifying LiDAR points using the algorithm adopted in the NYU depth dataset [@silberman2012indoor; @levin2004colorization]. We use the same DeepLabv3 to predict image level labels as before. Architecture and training details --------------------------------- For the backbone and feature pyramid components of our network, we use a pretrained FPN network [@lin2017feature], which incorporates a ResNet-50 [@he2016deep] front-end. The topdown network consists of a stack of 8 residual blocks, including a transposed convolution layer which upsamples the birds-eye-view features from a resolution of 0.5m to 0.25m per pixel. For the balanced loss weighting $\alpha^c$, we use the square root of the inverse class frequency, as we found that using inverse frequency directly leads to a tendancy to overpredict on small classes. The uncertainty loss weighting $\lambda$ is taken as 0.001. We train all networks until convergence using SGD with a learning rate of 0.1, batch size 12 and a momentum of 0.9. Evaluation ---------- Our primary evaluation metric is the Intersection over Union (IoU) score, which we compute by binarising the predictions according to a Bayesian decision boundary ($p(m_i^c|z_t) > 0.5$). To account for the arbitrary nature of this threshold, we also provide precision-recall curves as part of the supplementary material. Non-visible grid cells (see [Section \[sec:data\]]{}) are ignored during evaluation. Results ======= Ablation study -------------- Before comparing against other methods, we validate our choice of architecture by performing an ablation study on the Argoverse dataset. We begin from a simple baseline, consisting of the backbone network, an inverse perspective mapping to geometrically map features to the birds-eye-view, and a sigmoid layer to predict final occupancy probabilities. We then incrementally reintroduce each of the key components of our approach: the dense transformer layer (**D**), transformer pyramid (**P**), and topdown network (**T**). The results of this ablation study are shown in the second half of [Table \[tab:argoverse\]]{}. Each successive component improves the performance by a consistent factor of roughly 1% mean IoU, with the addition of the dense transformer having a particularly pronounced effect on the results, which we argue is one of the key novelties of our approach. The topdown network provides no advantage for large classes such as driveable area, but significantly improves performance for small, rare classes such as motorbike and bicycle. Method [Drivable\*]{} [Ped. crossing]{} [Walkway\*]{} [Carpark]{} [Car\*]{} [Truck]{} [Bus\*]{} [Trailer]{} [Constr. veh.]{} [Pedestrian\*]{} [Motorcycle\*]{} [Bicycle\*]{} [Traf. Cone]{} [Barrier]{} [Mean]{} [CS Mean]{} ------------------------ ---------------- ------------------- --------------- ------------- ----------- ----------- ----------- ------------- ------------------ ------------------ ------------------ --------------- ---------------- ------------- ---------- ------------- IPM 40.1 - 14.0 - 4.9 - 3.0 - - 0.6 0.8 0.2 - - - 9.1 Depth Unproj. 27.1 - 14.1 - 11.3 - 6.7 - - 2.2 2.8 1.3 - - - 9.4 VED [@lu2019monocular] 54.7 12.0 20.7 13.5 8.8 0.2 0.0 7.4 0.0 0.0 0.0 0.0 0.0 4.0 8.7 12.0 VPN [@pan2019cross] 58.0 27.3 29.4 12.9 **25.5** **17.3** 20.0 **16.6** 4.9 7.1 5.6 4.4 4.6 **10.8** 17.5 21.4 Ours **60.4** **28.0** **31.0** **18.4** 24.7 16.8 **20.8** **16.6** **12.3** **8.2** **7.0** **9.4** **5.7** 8.1 **19.1** **23.1** Comparison to other methods {#sec:argoverse} --------------------------- In addition to the ablation experiments described above, we evaluate our final architecture to a number of baseline methods described in [Section \[sec:baseline\]]{}. It can be seen from [Table \[tab:argoverse\]]{} that we outperform all previous approaches by a significant margin. The two prior works, VPN and VED, achieve a comparable IoU on the drivable area class (representing the road surface), but across the smaller classes such as vehicle, pedestrian etc., we are able to obtain considerably better results. We suggest that this improvement is explained by the fact that our dense transformer layer preserves more spatial information compared to the fully connected bottlenecks of [@lu2019monocular] and [@pan2019cross]. This hypothesis is supported by the qualitative results illustrated in [Figure \[fig:argoverse\]]{}, which show that our method is much more able to resolve fine details such as the separation between individual cars (rows 1 and 2) or crowds of pedestrians (row 3). Both VPN and in particular VED on the other hand are only capable of making relatively coarse predictions and often miss important features, such as the car in row 3. The IPM baseline achieves reasonably good performance on the drivable area class but fails across all other classes because the predictions are elongated along the camera rays, as can be seen from [Figure \[fig:argoverse\]]{}. The success of the depth unprojection method meanwhile is limited by the inherent sparsity of the lidar point clouds beyond a range of about 25m. Evaluation on the NuScenes dataset {#sec:nuscenes} ---------------------------------- Having justified our approach on the relatively small Agoverse dataset, we move to the more challenging evaluation scenario of the NuScenes dataset. We report quantitative results in [Table \[tab:nuscenes\]]{}, and visualise our predictions in [Figure \[fig:nuscenes\]]{}. Despite the greater diversity of this dataset, we are able to outperform the next-best approach, the VPN method of [@pan2019cross], by a relative factor of 9.1%. As with Argoverse, our method is consistently able to capture finer details in the scene, such as the shape of the bus in row 2 and the geometry of the crossroads in row 3. On this dataset, the VED method completely breaks down for the cases of small (pedestrian, cyclist, etc.) or infrequently occurring (construction vehicle, bus) classes. Temporal and sensor fusion {#sec:fusion-results} -------------------------- Predicting BEV maps from a single viewpoint as discussed in [Section \[sec:nuscenes\]]{} and [Section \[sec:argoverse\]]{} is typically insufficient for driving purposes; in general we want to build a complete picture of our environment taking into account multiple sensors and historical information. In [Figure \[fig:headline\]]{} we show an example of how the occupancy grids from six surround-view cameras can be combined using the Bayesian fusion scheme described in [Section \[sec:fusion\]]{}. We assume a prior probability of $p(m_i^c) = 0.5$ for all classes. For static elements of the scene, such as road, sidewalk etc., we can go a step further by combining predictions over multiple timesteps to build a complete model of the geometry of a given scene. [Figure \[fig:multiframe\]]{} shows several examples of accumulating occupancy probabilities over 20s long sequences from the NuScenes dataset. The network is able to utilise information from multiple views to resolve ambiguities, resulting in a smoother overall prediction. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Scene-level occupancy grid maps generated by accumulating occupancy probabilities over 20s sequences. White lines indicate the ego-vehicle trajectory. Note that only static classes (drivable, crossing, walkway, carpark) are visualised.[]{data-label="fig:multiframe"}](figures/multiframe/scene-0010.png "fig:"){height="4.3cm"} ![Scene-level occupancy grid maps generated by accumulating occupancy probabilities over 20s sequences. White lines indicate the ego-vehicle trajectory. Note that only static classes (drivable, crossing, walkway, carpark) are visualised.[]{data-label="fig:multiframe"}](figures/multiframe/scene-0963e.png "fig:"){height="4.3cm"} ![Scene-level occupancy grid maps generated by accumulating occupancy probabilities over 20s sequences. White lines indicate the ego-vehicle trajectory. Note that only static classes (drivable, crossing, walkway, carpark) are visualised.[]{data-label="fig:multiframe"}](figures/multiframe/scene-0962e.png "fig:"){height="4.3cm"} ![Scene-level occupancy grid maps generated by accumulating occupancy probabilities over 20s sequences. White lines indicate the ego-vehicle trajectory. Note that only static classes (drivable, crossing, walkway, carpark) are visualised.[]{data-label="fig:multiframe"}](figures/multiframe/scene-0054e.png "fig:"){height="4.3cm"} ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ Conclusions =========== We have proposed a novel method for predicting birds-eye-view maps directly from monocular images. Our approach improves on the state-of-the-art by incorporating dense transformer layers, which make use of camera geometry to warp image-based features to the birds-eye-view, as part of a multiscale transformer pyramid. As well as predicting maps from a single image, our method is able to effortlessly combine information across multiple views to build an exhaustive model of the surrounding environment. We believe that this work provides a broad framework for future work into other tasks which operate in the birds-eye-view, such as lane instance detection and future prediction. {width="\linewidth"} Modifications to competing networks =================================== In Section 5 of the main paper, we compare our approach to two recent works from the literature: the Variational Encoder-Decoder of Lu[ *et al.*]{}[@lu2019monocular] and the View Parsing Network of Pan[ *et al.*]{}[@pan2019cross]. Both works tackle a closely related task to ours, but use other datasets and presume slightly different input dimensions and output map resolutions. In order to compare our work directly, we must therefore make minor architectural changes, which we consider to be the minimum possible for compatibility with our datasets. In the interest of transparency, we detail these changes below: VED: : We modify the bottleneck to use dimensions of 36128 for NuScenes and 47128 for Argoverse to account for the different input aspect ratios. We add an additional decoder layer (identical to previous layers) to increase the resolution from 6464 to 128128 and then bilinearly upsample to our output size of 196200. VPN: : We increase the transformer module bottleneck dimension to 2950 for NuScenes and 3850 for Argoverse, and then upsample the output to 196200 using the authors’ existing code. Since we consider a multilabel prediction setting unlike the single label prediction task addressed in the original works, we train both methods using the balanced cross entropy loss described in Section 3.1. Precision-Recall curves for Argoverse and NuScenes experiments ============================================================== Figures \[fig:pr-argoverse\] and \[fig:pr-nuscenes\] show the precision-recall trade-off on the Argoverse and NuScenes datasets respectively. Across almost all classes it can be seen that our method represents an upper envelope on the precision achievable for a given recall setting. {width="\linewidth"} [^1]: Source code and dataset splits will be made available at [github.com/tom-roddick/mono-semantic-maps](github.com/tom-roddick/mono-semantic-maps).

--- abstract: | The proton matrix element of the isovector-scalar density, $\langle p|\overline{u}u-\overline{d}d|p\rangle/2M_p$, is calculated by evaluating the nucleon current correlation function in an external isovector-scalar field using the QCD sum-rule method. In addition to the usual chiral and gluon condensates of the QCD vacuum, the response of the chiral condensates to the external isovector field enters the calculation. The latter is determined by two independent methods. One relates it to the difference between the up and down quark chiral condensates and the other uses the chiral perturbation theory. To first order in the quark mass difference $\delta m=m_d-m_u$, the non-electromagnetic part of the neutron-proton mass difference is given by the product $\delta m \langle p|\overline{u}u-\overline{d}d|p\rangle/2M_p$; the resulting value is in reasonable agreement with the experimental value. address: - | Department of Physics and Center for Theoretical Physics\ University of Maryland, College Park, Maryland 20742 - | Center for Theoretical Studies, Institute of Science\ Bangalore, India 560 012 author: - 'Xuemin Jin and Marina Nielsen[^1]' - 'J. Pasupathy' title: 'Calculation of $\langle p|\overline{u}u-\overline{d}d|p\rangle$ from QCD sum rule and the neutron-proton mass difference' --- INTRODUCTION {#intro} ============ Understanding the properties of the nucleon is of obvious importance in hadron physics. A variety of nucleon matrix elements of bilinear quark operators have been evaluated in the past using various approaches. The QCD sum-rule method originally proposed by Shifman, Vainstein, and Zakharov [@svz] and its extension by Ioffe and Smilga [@ioffe2] for external field problems, provide a tractable framework for the study of these nucleon matrix elements. These include the matrix element of electromagnetic current to determine the magnetic moments [@ioffe2; @balitsky1; @chiu1], the matrix element of the axial vector current to find the renormalization of nucleon axial coupling constant [@belyaev2; @chiu2], the matrix element of the quark part of the energy momentum tensor, which gives the momentum fraction carried by the up and down quarks in deep inelastic scattering [@kolesinchenko1; @belyaev3], and the matrix element of isoscalar-scalar current for evaluating the nucleon Sigma term [@jin1]. In the present work, we calculate the matrix element of the isovector-scalar density, $\langle p|\overline{u}u- \overline{d}d|p\rangle/2M_p$, following the same external field approach. The appearance of the external field leads to specific new features in QCD sum rules which distinguish them from those in the absence of the external field. Thus the phenomenological representation of the correlation function written in terms of physical intermediate states contains a double pole at the nucleon mass whose residue contains the matrix element of interest. In addition there are single pole terms which arise from the transition matrix element between the ground state nucleon and excited states. These later contributions are not exponentially damped after Borel transformation relative to the double pole term and should be retained in a consistent analysis of the sum rules. In the theoretical side of the sum rules expressed in terms of an operator product expansion (OPE) the external field contributes in two different ways–by directly coupling to the quark fields in the nucleon current and by polarizing the QCD vacuum. Since for our problem the external field is a Lorentz scalar, non-scalar correlators cannot be induced in the QCD vacuum. However the external field does modify the quark and gluon condensates already present in the QCD vacuum. It turns out that for the problem under study the most important one is the response of the up and down quark chiral condensates to the external field which can be described by a susceptibility $\chi$. This can be determined by writing a spectral representation, which can be evaluated for example by using chiral perturbation theory. Alternatively, since in the real world up and down quark masses are different, $\delta m = m_d - m_u $ itself can be regarded as an external isovector scalar field and $\chi$ can be related to the isospin breaking in quark condensates $\gamma\equiv (\langle 0|\overline{d}d-\overline{u}u|0\rangle)/\langle 0|\overline{u}u|0\rangle$ and is given to first order in $\delta m $ by $\chi = -\gamma / \delta m$. At the current level of accuracy where these two determinations can be done, they are mutually consistent. In Sec. \[sumrule\] we derive the sum rules for the nucleon current correlation function in an external isovector-scalar field and describe the analysis of these sum rules, which leads us to determine $\langle p|\overline{u}u-\overline{d}d|p\rangle/2M_p$. We also study its dependence on the $\chi$ value. In Sec. \[npd\] we show that an evaluation of the matrix element $\langle p|\overline{u}u - \overline{d}d|p\rangle/2M_p$ enables us to compute the non-electromagnetic part of the neutron-proton mass difference. Since the 1970 ’s it has been recognized that the empirical mass difference arises from two sources. One is purely electromagnetic and yields a contribution of $-0.76\pm 0.30\,\text{MeV}$[@gasser1] to the neutron-proton mass difference. The second source is the difference between up and down quark masses, which is of the order of the quark masses themselves. This difference leads to larger contribution which overrides the electromagnetic contribution making neutron heavier than proton. To first order in $\delta m $ this later contribution is given by the product of $\delta m$ and $\langle p|\overline{u}u-\overline{d}d|p\rangle /2M_p$. Recently several authors [@hatsuda1; @yang1; @adami1; @eletsky1] have applied the QCD sum rules for the nucleon mass to extract the neutron-proton mass difference by including the up and down quark masses in the mass sum rules for proton and neutron and the isospin breaking in the condensates. The relation of these works to ours is described in Sec. \[discussion\], where we also briefly comment on the Nolen-Schiffer anomaly. Sum-rule calculation {#sumrule} ==================== The procedure that we use for calculating the matrix element $\langle p|\overline{u}u-\overline{d}d|p\rangle/2M_p$ follows the same pattern as used in Refs. [@ioffe2; @balitsky1; @chiu1; @belyaev2; @chiu2; @kolesinchenko1; @belyaev3; @jin1]. We first couple the quarks to an external isovector-scalar field $S_V(x)$ which is described by adding a term $\Delta {\cal L}$ to the usual QCD Lagrangian $$\Delta{\cal L} \equiv - S_V(x) [\overline u (x) u(x) -\overline d(x) d(x)]\; . \label{lag}$$ We can take $S_V(x)$ to be a constant $S_V$ since we are only interested in the zero momentum transfer matrix element. The correlation functions of the nucleon current in QCD vacuum in the presence of $S_V$ will be computed. The term linear in $S_V$ gives the matrix element of interest. QCD sum rules for $\langle p|\overline{u}u-\overline{d} d|p\rangle$ ------------------------------------------------------- Consider the correlation function of the proton interpolating field in the presence of a [*constant*]{} external isovector-scalar field $S_{\mbox{\tiny{\rm V}}}$ which can be taken to be arbitrarily small $$\Pi(S_{\mbox{\tiny{\rm V}}},q)\equiv i\int d^4{x}e^{iq\cdot x} \langle 0|{{\rm T}[\eta_p(x)\overline{\eta}_p(0)]}\rangle_{S_ {\mbox{\tiny{\rm V}}}}\ , \label{corr}$$ where $\eta_p$ is the proton interpolating field introduced in Ref. [@ioffe1] $$\eta_p(x)=\epsilon_{abc} \left[{u^T_a}(x)C\gamma_\mu u_b(x)\right] \gamma_5\gamma^\mu d_c(x)\ , \label{eta}$$ where $u_a(x)$ and $d_c(x)$ stand for the up and down quark fields, $a,b$ and $c$ are the color indices, and $C=-C^T$ is the charge conjugation matrix. Since $S_V$ is a scalar, Lorentz covariance and parity allow one to decompose $\Pi(S_{\mbox{\tiny{\rm V}}},q)$ into two distinct structures $$\Pi(S_{\mbox{\tiny{\rm V}}},q)\equiv \Pi^1(S_{\mbox{\tiny{\rm V}}},q^2)+ \Pi^q(S_{\mbox{\tiny{\rm V}}},q^2)\rlap{/}{q}\ .$$ To the first order in the external field $S_{\mbox{\tiny{\rm V}}}$, the two invariant functions can be written as $$\Pi^i(S_{\mbox{\tiny{\rm V}}},q^2)= \Pi^i_0(q^2)+S_{\mbox{\tiny{\rm V}}}\Pi^i_1(q^2)$$ for $\{i=1,q\}$, where $\Pi^i_0$ are the invariant functions in the absence of the external field which give rise to the mass sum rules discussed extensively in Refs. [@ioffe1; @belyaev1; @leinweber1; @reinders1]. We are concerned here with the linear response to the external field given by $\Pi^i_1(q^2)$. To derive a QCD sum rule, one first carries out an OPE, which will express $\Pi^i_1(q^2)$ in terms of various vacuum correlators, and then matches it to an expansion in terms of physical intermediate states. Now the external field contributes to $\Pi({S_{\mbox{\tiny{\rm V}}}},q)$ in two ways: it couples directly to the quark fields in the propagating nucleon current and it also polarizes the QCD vacuum. The chiral condensates of up and down quarks change as follows $$\begin{aligned} \langle\overline{u}u\rangle_{S_{\mbox{\tiny{\rm V}}}}&=&\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}-\chi S_{\mbox{\tiny{\rm V}}}\langle\overline{u}u\rangle_{\mbox {\tiny{\rm 0}}} , \label{uc} \\*[7.2pt] \langle\overline{d}d\rangle_{S_{\mbox{\tiny{\rm V}}}}&=&\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}}+\chi S_{\mbox{\tiny{\rm V}}}\langle\overline{d}d\rangle_{\mbox{\tiny {\rm 0}}} , \label{dc}\end{aligned}$$ where $\langle \hat{O}\rangle_{\mbox{\tiny{\rm 0}}}\equiv \langle 0|\hat{O}|0\rangle$. Using Eq. (\[lag\]) it is easy to see that $\chi$ is related to the correlation function $$\chi\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}\equiv {i\over 2}\int d^4x \langle{\rm T}\{\overline{u}(x)u(x)-\overline{d}(x)d(x), \overline{u}(0)u(0)-\overline{d}(0)d(0)\}\rangle_{\mbox {\tiny{\rm 0}}}\ . \label{chi}$$ Similarly the mixed quark-gluon condensates change as follows $$\begin{aligned} \langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{S_{\mbox{\tiny{\rm V}}}} &=&\langle g_s\overline{u}\sigma\cdot {\cal G} u\rangle_{\mbox{\tiny {\rm 0}}}-\chi_m S_{\mbox{\tiny{\rm V}}}\langle g_s\overline{u}\sigma\cdot {\cal G} u\rangle_{\mbox {\tiny{\rm 0}}} , \label{uqc} \\*[7.2pt] \langle g_s\overline{d}\sigma\cdot {\cal G} d \rangle_{S_{\mbox{\tiny{\rm V}}}} &=&\langle g_s\overline{d}\sigma\cdot {\cal G} d\rangle_{\mbox{\tiny {\rm 0}}}+\chi_m S_{\mbox{\tiny{\rm V}}}\langle g_s\overline{d}\sigma\cdot {\cal G} d\rangle_{\mbox {\tiny{\rm 0}}} , \label{dqc}\end{aligned}$$ where $\sigma\cdot {\cal G}\equiv \sigma_{\mu\nu}{\cal G}^{\mu\nu}$ with ${\cal G}^{\mu\nu}$ the gluon field tensor and $\chi_{\mbox{\tiny m}}$, the susceptibility corresponding to the quark-gluon mixed condensate, can also be expressed in terms of a spectral representation. Since isospin is a good symmetry for hadron matrix elements we have assumed that the response of the up and down quarks is the same, apart from sign, and that we can disregard the vchange in the gluon condensate $\langle (\alpha_s/\pi)G^2\rangle$ due to the external isovector field. To calculate the Wilson coefficients of the OPE, we need the coordinate-space quark propagators in the presence of the external field and the vacuum condensates. To first order in the external field $S_{\mbox{\tiny{\rm V}}}$, the propagators in the fixed-point gauge [@fock1; @schwinger1; @jin4] take the form $$\begin{aligned} \langle {\rm T}[u_{i}^a(x)\overline{u}_{j}^b(0)]\rangle_{S_ {\mbox{\tiny{\rm V}}}}&= &{i\over 2\pi^2}\delta^{ab}{1\over x^4}\left[\hat{x}\right]_{ij}-\delta^{ab}{S_{\mbox{\tiny {\rm V}}}\over 4\pi^2}{\delta_{ij}\over x^2}- {1\over 12}\delta^{ab}\langle\overline{u}u\rangle_ {\mbox{\tiny{\rm 0}}}\delta_{ij} \nonumber\\*[7.2pt] & & +{1\over 12}\delta^{ab}\chi S_{\mbox{\tiny{\rm V}}}\langle\overline{u}u\rangle_{\mbox {\tiny{\rm 0}}}\delta_{ij} +\delta^{ab}{i S_{\mbox{\tiny{\rm V}}}\over 48}\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}} \left[\rlap{/}{x}\right] \nonumber\\*[7.2pt] & & -\delta^{ab}{x^2\over 192 }\langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{\mbox{\tiny{\rm 0}}}\delta_{ij}-\delta^{ab}\chi_ {\mbox{\tiny m}}S_{\mbox{\tiny{\rm V}}}{x^2\over 192 }\langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{\mbox{\tiny{\rm 0}}}\delta_{ij} \nonumber\\*[7.2pt] & & +\delta^{ab}{i S_{\mbox{\tiny{\rm V}}}x^2\over 9\cdot 128}\langle g_s\overline{u}\sigma\cdot {\cal G}u \rangle_{\mbox{\tiny{\rm 0}}}\left[\rlap{/}x\right]_{ij} \nonumber\\*[7.2pt] & & -{ig_s\over 32\pi^2}\left(G_{\mu\nu}(0)\right)^{ab} {1\over x^2}[\rlap{/}x\sigma^{\mu\nu}+\sigma^{\mu\nu} \rlap{/}x]_{ij}+ \cdot\cdot\cdot\ , \label{propu}\end{aligned}$$ $$\begin{aligned} %\\*[7.2pt] \langle {\rm T}[d_{i}^a(x)\overline{d}_{j}^b(0)]\rangle_ {S_{\mbox{\tiny{\rm V}}}}&= &{i\over 2\pi^2}\delta^{ab}{1\over x^4}\left[\hat{x}\right]_{ij}+\delta^{ab}{S_{\mbox{ \tiny{\rm V}}}\over 4\pi^2}{\delta_{ij}\over x^2}- {1\over 12}\delta^{ab}\langle\overline{d}d\rangle_ {\mbox{\tiny{\rm 0}}}\delta_{ij} \nonumber\\*[7.2pt] & & -{1\over 12}\delta^{ab}\chi S_{\mbox{\tiny{\rm V}}}\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}}\delta_{ij} -\delta^{ab}{i S_{\mbox{\tiny{\rm V}}}\over 48}\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}} \left[\rlap{/}{x}\right] \nonumber\\*[7.2pt] & & -\delta^{ab}{x^2\over 192 }\langle g_s\overline{d}\sigma\cdot {\cal G} d \rangle_{\mbox{\tiny{\rm 0}}}\delta_{ij}+\delta^{ab}\chi_ {\mbox{\tiny m}}S_{\mbox{\tiny{\rm V}}}{x^2\over 192 } \langle g_s\overline{d}\sigma\cdot {\cal G} d \rangle_{\mbox{\tiny{\rm 0}}}\delta_{ij} \nonumber\\*[7.2pt] & & -\delta^{ab}{i S_{\mbox{\tiny{\rm V}}}x^2\over 9\cdot 128}\langle g_s\overline{d}\sigma\cdot {\cal G}d \rangle_{\mbox{\tiny{\rm 0}}}\left[\rlap{/}x\right]_{ij} \nonumber\\*[7.2pt] & & -{ig_s\over 32\pi^2}\left(G_{\mu\nu}(0)\right)^{ab} {1\over x^2}[\rlap{/}x\sigma^{\mu\nu}+\sigma^{\mu\nu} \rlap{/}x]_{ij}+ \cdot\cdot\cdot\ . \label{propd}\end{aligned}$$ Here we are interested in terms linear in $S_{\mbox{\tiny{\rm V}}}$. We can then disregard the current quark masses because they make negligible contributions. The correlation function $\Pi(S_{\mbox{\tiny{\rm V}}},q)$ can be computed using the Eqs. (\[propu\]) and (\[propd\]) above. We have computed the contributions corresponding to the diagrams listed in Fig. \[diag\]. The results of our calculations for the invariant functions $\Pi^q_1$ and $\Pi^1_1$ are $$\begin{aligned} \Pi^q_1(q^2)&=&{1\over 4\pi^2}\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}} \ln(-q^2)+{4\over 3}\chi\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}^2 {1\over q^2}-{1\over 24\pi^2}\left(2\langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{\mbox{\tiny{\rm 0}}}-\langle g_s\overline{d}\sigma\cdot {\cal G} d\rangle_{\mbox{\tiny{\rm 0}}}\right){1\over q^2}\nonumber \\*[7.2pt] & & + {\chi\over 6}\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}} \langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{\mbox{\tiny{\rm 0}}}{1\over q^4} + {\chi_{\mbox{\tiny m}}\over 6}\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}} \langle g_s\overline{u}\sigma\cdot {\cal G} u \rangle_{\mbox{\tiny{\rm 0}}}{1\over q^4}\; , \label{q} \\*[7.2pt] \Pi^1_1(q^2)&=&{1\over 32 \pi^4}(q^2)^2\ln(-q^2)+{\chi\over 4\pi^2}\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}} q^2\ln(-q^2)-{2\over 3}\left(3 \langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}\langle\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}}-2\langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}^2 \right){1\over q^2}\ . \label{1}\end{aligned}$$ We now turn to the phenomenological side of the sum rules, which is obtained by expanding $\Pi(S_{\mbox{\tiny{\rm V}}},q)$ in terms of physical hadronic intermediate states. There are three types of contributions. Firstly the matrix element of interest is contained in the term $$\langle 0|\eta_p|p\rangle\langle p|\overline{u}u-\overline{d}d|p \rangle\langle p|\overline{\eta}_p |0\rangle\ , \label{double}$$ where the current $\overline{\eta}_p$ creates a proton that interacts with the external field $S_{\mbox{\tiny{\rm V}}}$ and is then annihilated by the current $\eta_p$. Defining $$\langle 0|\eta_p|p\rangle=\lambda_p v\ ,$$ where $\lambda_p$ denotes the coupling between $\overline{\eta}_p(0)|0\rangle$ and the physical proton state and $v$ is the usual Dirac spinor ($\overline{v}v=2M_p$), and introducing the notation $$H\equiv {\langle p|\overline{u}u - \overline{d}d |p \rangle\over 2 M_p}\; , \label{defH}$$ one can write the contribution of Eq. (\[double\]) to $\Pi^i_1(q^2)$ as $$-\lambda_p^2{\rlap{/}q+M_p\over q^2-M_p^2}\hspace*{0.15cm} H \hspace*{0.15cm} {\rlap{/}q+M_p\over q^2-M_p^2}\ . \label{dia_con}$$ It is seen that the above term has a double pole at the nucleon mass. The external field $S_{\mbox{\tiny{\rm V}}}$ can also cause transition between the proton and an excited state which can have either positive or negative parity relative to the proton. When the relative parity is positive, the contribution can be written as $$-\lambda_p\lambda_{p^*}{\rlap{/}q+M_p\over q^2-M_p^2}\hspace*{0.15cm} H^* \hspace*{0.15cm} {\rlap{/}q-M^*\over q^2-{M^*}^2}\ .$$ where now $H^*$ is the transition matrix element between $|p\rangle$ and $|p^* \rangle$. This term has a simple pole at the proton mass as well as at the mass $M^*$ of the excited state. It is easy to see that after a Borel transformation this contribution is not exponentially damped as compared to the double pole contribution Eq. (\[dia\_con\]). Therefore, one has to retain these simple pole terms in the analysis of the sum rules through the introduction of a phenomenological parameter to be determined along with the diagonal matrix element $H$ (see Refs. [@ioffe2; @balitsky1; @chiu1; @belyaev2; @chiu2; @kolesinchenko1; @belyaev3; @jin1]). The third type of contributions comes from transitions involving only the excited states. These are of course exponentially damped after a Borel transformation and can be approximated in the usual manner [@ioffe1; @belyaev1; @leinweber1; @reinders1], by equating them to the perturbative contributions starting from an effective threshold. Equating the OPE results Eqs. (\[1\]) and (\[q\]) and the physical intermediate state expansion discussed above and applying the Borel transformation [@svz], we obtain the following sum rules $$\begin{aligned} {M^2\over 4\pi^2}\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}} E_0 L^{-4/9}+{4\over 3}\chi\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}}^2 L^{4/9}- {1\over 24\pi^2}\langle g_s\overline{q}\sigma&\cdot&G q \rangle_{\mbox{\tiny{\rm 0}}} L^{-8/9} - {\chi\over 6M^2}\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}} \langle g_s\overline{q}\sigma\cdot G q \rangle_{\mbox{\tiny{\rm 0}}}L^{-2/27} \nonumber\\*[7.2pt] -{\chi_m\over 6M^2}\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}} \langle g_s\overline{q}\sigma\cdot G q \rangle_{\mbox{\tiny{\rm 0}}}L^{-2/27} &=&\left[2\lambda_p^2{M_p\over M^2}H +A_q\right]e^{-M_p^2/M^2} \; , \label{sum_q}\end{aligned}$$ $$\begin{aligned} {M^6\over 16\pi^4}E_2L^{-8/9}+{\chi\over 4\pi^2}\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}}M^4E_1&-&{2 \over 3}\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}}^2 \nonumber\\*[7.2pt] &=&\left[2\lambda_p^2{M_p^2\over M^2}H +A_1\right]e^{-M_p^2/M^2} \; , \label{sum_1}\end{aligned}$$ where $A_1$ and $A_q$ are the phenomenological parameters that represent the sum over the contributions from all off-diagonal transitions between the proton and the excited states. Here we have defined $E_0\equiv 1-e^{-s_0/M^2}$, $E_1\equiv 1-e^{-s_0/M^2}\left({s_0\over M^2}+1\right)$ and $E_2\equiv 1-e^{-s_0/M^2}\left({s_0\over 2M^4} +{s_0\over M^2}+1\right)$, which account for the sum of the contributions involving excited states only, where $s_0$ is an effective continuum threshold. In Eqs. (\[sum\_q\]) and (\[sum\_1\]), we have ignored the isospin breaking in the vacuum condensates. We have also taken into account the anomalous dimension of the various operators through the factor $L\equiv\ln(M^2/\Lambda_{\rm QCD}^2)/\ln(\mu^2/\Lambda_{\rm QCD}^2)$[@svz; @ioffe1]. We take the renormalization scale $\mu$ and the QCD scale parameter $\Lambda_{\rm QCD}$ to be $500\,\text{MeV}$ and $150\,\text{MeV}$. Estimate of $\chi$ ------------------ It is clear from Eq. (\[chi\]) that $\chi$ is determined once the isovector-scalar two point function is known. The latter has been studied by Gasser and Leutwyler using chiral perturbation theory to one loop [@gasser2]. They found $$i\int d^4x \langle 0|{\rm T}\{\overline{u}(x)u(x)-\overline{d} (x)d(x), \overline{u}(0)u(0)-\overline{d}(0)d(0)\}|0\rangle = 8\left({m_\pi^2\over m_u+m_d}\right)^2 h_3 \ . \label{chi2}$$ Since two pions cannot form an isovector-scalar, the $|\pi \pi\rangle$ intermediate state does not contribute in Eq. (\[chi2\]). The other possible pseudoscalar two particle states are $| K \overline K\rangle$ and $|\eta \pi\rangle$ which means that an extension to $SU (3)$ flavor symmetry is necessary. This has been done by Gasser and Leutwyler in Ref. [@gasser3]. Using Eq. (11.6) of Ref. [@gasser3], and $\langle\overline{s}s\rangle_{\mbox{\tiny {\rm 0}}}/\langle\overline{q}q\rangle_{\mbox {\tiny{\rm 0}}}=0.8$, we get $h_3\simeq -0.003$. Combining Eqs. (\[chi\]) and (\[chi2\]), we obtain $$\chi = -4{m_\pi^2\over (m_u+m_d)f_\pi^2} h_3\simeq 2.2 \, \text{GeV}^{-1} \; , \label{chiva}$$ where we have used $(m_u+m_d) \langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}}=-m_\pi^2 f_\pi^2$, and the experimental values $m_\pi=138\,\text{MeV}$, $f_\pi=93\,\text{MeV}$, and a median value $\langle\overline{q}q\rangle_{\mbox{\tiny{\rm 0}}}= -(240\text{MeV})^3$ which corresponds to $m_u+m_d=11.8\,\text{MeV}$. Alternatively, $\chi$ can be determined as follows. The terms proportional to the current quark masses in the QCD Lagrangian can be written as $${\cal L}_{\mbox{mass}} =-\hat{m}(\overline{u}u + \overline{d}d)+{1\over 2}\delta m (\overline{u}u - \overline{d}d)-m_s\overline{s}s-\cdots\ ,$$ where $\hat{m}\equiv {1\over 2}(m_u+m_d)$ and the ellipses denote the terms due to heavier quarks. Treating $\delta m(\overline{u}u-\overline{d}d)$ as a source term one obtains using Eq. (\[chi\]) $$\chi \langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}={d \over d \delta m}\langle\overline{u}u-\overline{d}d\rangle_{\mbox {\tiny{\rm 0}}}\ .$$ On the other hand, one can expand $\langle\overline{u}u-\overline{d}d\rangle_{\mbox{\tiny {\rm 0}}}$ and ${d \over d \delta m} \langle\overline{u}u-\overline{d}d \rangle_{\mbox{ \tiny{\rm 0}}}$ into the Taylor Series in $\delta m$. Using $\langle\overline{u}u-\overline{d}d\rangle_{\mbox{\tiny {\rm 0}}}|_{\delta m=0}=0$, we find $$\langle\overline{u}u-\overline{d}d\rangle_{\mbox{\tiny{\rm 0}}}= \chi\delta m \langle\overline{u}u\rangle_{\mbox{\tiny{\rm 0}}}+ O[(\delta m)^2]\ ,$$ which implies $$\chi\delta m=-\gamma+O[(\delta m)^2]\ . \label{chiper}$$ Therefore, to the lowest order in $\delta m$, the susceptibility $\chi$ is determined by the ratio of the isospin breaking parameters $\gamma$ and $\delta m$. The value of $\gamma$ has been estimated previously in various approaches [@gasser3; @paver1; @pascual1; @bagan1; @dominguez1; @dominguez2; @narison1], with results ranging from $-1\times 10^{-2}$ to $-3\times 10^{-3}$. Gasser and Leutwyler [@gasser2] have determined the ratio $\delta m /(m_u + m_d ) = 0.28 \pm 0.03 $. Since we have used a median value of 11.8 MeV for the sum of the up and down quark masses we adopt a median value for $\delta m=3.3\,\text{MeV}$. Here we consider the $\gamma $ values to be in the range $-1\times 10^{-2}$ to $-3\times 10^{-3}$, which, upon using Eq. (\[chiper\]), corresponds to $$0.9\,\text{GeV}^{-1}\le\chi\le 3\,\text{GeV}^{-1}\ . \label{chi-gamma}$$ The susceptibility $\chi_{m}$ can also be determined using a spectral representation for Eqs. (\[uqc\]) and (\[dqc\]). However, given the uncertainty in the value of $\chi$, we assume, in this work, $\chi_{m}\simeq\chi$. Sum-rule analysis ----------------- Defining $a\equiv-4\pi^2\langle\overline{q}q \rangle_{\mbox{\tiny{\rm 0}}}$, $\tilde{\lambda}_p^2\equiv 32\pi^4\lambda_p^2$, $m_0^2\equiv\langle\overline{q}g_s\sigma\cdot G q\rangle_ {\mbox{\tiny{\rm 0}}}/ \langle\overline{q}q \rangle_{\mbox{\tiny{\rm 0}}}$, $\tilde{A}_q=(2\pi)^4 A_q$, and $\tilde{A}_1=(2\pi)^4 A_1$, we can rewrite the sum rules Eqs. (\[sum\_1\]) and (\[sum\_q\]) as $$\begin{aligned} & &e^{M_p^2/M^2}\left[-M^4aE_0 L^{-4/9}+{4M^2\over 3}\chi a^2 L^{4/9}+{M^2\over 6}m_0^2 a L^{-8/9}-{\chi\over3}a^2 m_0^2 L^{-2/27}\right] \nonumber\\*[7.2pt] & &\hspace*{2in}=\tilde\lambda_p^2 M_p H +\tilde A_q M^2 \; , \label{fiq}\end{aligned}$$ $$\begin{aligned} & &e^{M_p^2/M^2}\left[M^8E_2L^{-8/9}-M^6\chi a E_1-{2\over 3} M^2 a^2\right] \nonumber\\*[7.2pt] & &\hspace*{2in}=\tilde\lambda_p^2 M_p^2 H +\tilde A_1 M^2 \; . \label{fi1}\end{aligned}$$ To extract $H = \langle p|\overline{u}u-\overline{d}d|p\rangle /2M_p$, $\tilde{A}_q$ and $\tilde{A}_1$ from the above sum rules, we use the experimental value for the proton mass $M_p$ and extract $\tilde{\lambda}_p^2$ from the proton [*mass*]{} sum rules. Using $a=0.55\,\text{GeV}^3 (m_u+m_d=11.8\,\text{MeV})$ and $m_0^2=0.8\,\text{GeV}^2$, one finds from the best fit of the proton mass sum rules that $\tilde{\lambda}_p^2=2.1\,\text{GeV}^6$ corresponding to $s_0=2.3\,\text{GeV}^2$[@ioffe2].We use only the first or the chiral-odd sum rule of Ioffe which is more accurate than the second or the chiral-even sum rule. First consider the sum rule Eq. (\[fiq\]), which is obtained from $\Pi^q_1(q^2)$. For definiteness we take for $\chi$ the value $\chi=2.2\,\text{GeV}^{-1}$ given in Eq. (\[chiva\]). In Fig. \[sides1\], we plot the individual terms in the LHS of Eq. (\[fiq\]) as well as their sum, as functions of $M^2$ in the interval $0.8\leq M^2\leq 1.4\,\text{GeV}^2$. It can be seen that the LHS follows a linear behavior in $M^2$ and we can match it with the RHS. To find the best values for the constant and the coefficient of the linear term in the RHS we follow the numerical optimization procedure used in Refs.[@leinweber1; @furnstahl1]. We sample the sum rules in the fiducial region of $M^2$, where the contributions from the highest-dimensional condensates included in the sum rule remain small and the continuum contribution is controllable. Here we choose $0.8\leq M^2\leq 1.4\,\text{GeV}^2$ as the optimization region, which is identified by Ioffe and Smilga[@ioffe2] as the fiducial region for the nucleon mass sum rules. To quantify the fit of the left- and right-hand sides, we use the logarithmic measure $$\delta(M^2)=\ln\left[{{\mbox{maximum}}\{LHS,RHS\}\over{\mbox {minimum}}\{ LHS,RHS\}}\right]\ ,$$ which is averaged over $100$ points evenly spaced within the fiducial region of $M^2$, where $LHS$ and $RHS$ denote the left- and right-hand sides of the sum rules respectively. The sum-rule predictions are obtained by minimizing $\delta$. Using this procedure we obtain $$H \simeq 0.54\; , \; \tilde A_q \simeq 0.29\,\text{GeV}^{5}\ .$$ The RHS of Eq. (\[fiq\]) with these optimized values is also plotted in Fig.2. One can see that the LHS and RHS have a very good overlap. We now turn to the other sum rule Eq. (\[fi1\]), which is obtained from $\Pi^1_1$. In Fig. \[sides2\], the individual terms in the LHS as well as their sum are shown as functions of $M^2$ for $\chi=2.2\,\text{GeV}^{-1}$. The LHS is fairly linear in $M^2$ in the interval $0.8 \leq M^2 \leq 1.4 \text{GeV}^{-2}$. The RHS of Eq. (\[fi1\]), with the optimized values $$H \simeq 0.01\; , \; \tilde A_1 \simeq -1.82\,\text{GeV}^{6}\ ,$$ is also shown in Fig. \[sides2\]. It is worth noting that in the sum rule Eq. (\[fiq\]) the double pole term is more important than the single pole term, which is also qualitatively evident from the near constancy of the LHS as a function of $M^2$. In contrast, the single pole term in the sum rule Eq. (\[fi1\]) clearly dominates, leading us to suspect that it is not reliable to determine the double pole term in which we are interested. This is confirmed by the following. Let us consider relaxing our tacit assumption that the continuum threshold used in our external field sum rules should be the same as the one occurring in Ioffe’s mass sum rules. If $s_0$ is varied from $2.3 \text{GeV}^2$ to $2\text{GeV}^2$ or $2.6 \text{GeV}^2$, the result for the matrix element $H$ extracted from the sum rule Eq. (\[fi1\]) changes from $0.01$ to $-0.07$ ($s_0 =2\text{GeV}^2$) or $+0.08$ ($s_0=2.6\text{GeV}^2$) while the result from the sum rule Eq. (\[fiq\]) changes only from $0.54$ to $0.52$ or $0.56$. It is clear from the above analysis that the chiral-odd sum rule Eq. (\[fiq\]) is extremely stable, while the chiral-even sum rule Eq. (\[fi1\]) is not. So, in this paper we shall disregard the results based on the sum rule Eq. (\[fi1\]) and consider only the results from the stable sum rule Eq. (\[fiq\]). The fact that one of the sum rules works well, while the other fails, is not peculiar to the problem under study. This pattern is seen also in a study of the isoscalar-scalar matrix element as well as in the sum rules for the matrix elements of electromagnet current and axial vector current [@ioffe2; @balitsky1; @chiu1; @belyaev2; @chiu2; @kolesinchenko1; @belyaev3; @jin1]. As discussed extensively in Ref. [@chiu2], the different asymptotic behavior of various sum rules can be traced to the fact that even and odd parity states contribute with different sign and kinematical factors. If chiral symmetry is realized in the Wigner-Weyl mode at high energies, i.e., by parity doubling, it is possible to have either cancellation or reinforcement between excited state contributions. Irrespective of the exact manner in which these cancellations take place, it is clear that the sum rule in which the continuum contributions are weak is more reliable. This is the case for the chiral-odd sum rule Eq. (\[fiq\]), where the continuum factor $E_0$ appears, to be contrasted with the chiral-even sum rule Eq. (\[fi1\]), where the factor $E_2$ occurs. Let us now consider the effect of varying $\chi$ , which as we have seen earlier is not precisely known. We find that the quality of the overlap of the two sides of the chiral-odd sum rule remains good (as measured by $\delta$) as we change the value of $\chi$, and gets better as $\chi$ increases and worse as $\chi$ decreases. We also find that the continuum contribution gets larger as $\chi$ decreases and smaller as $\chi$ increases. For $\chi$ values in the range $1.4\,\text{GeV}^{-1}\le\chi\le 3.0\,\text{GeV}^{-1}$, we find $$\ H =0.32 -0.76\; .$$ For $\chi<1.4\,\text{GeV}^{-1}$, we find that the continuum contribution is larger than $50\%$. This implies that the sum rule is dominated by continuum and the predictions are not reliable for $\chi$ values smaller than $1.4\,\text{GeV}^{-1}$. The neutron-proton mass difference {#npd} ================================== In this section we consider the relation between the matrix element evaluated in the last section and the neutron-proton mass difference. First let us disregard electromagnetism and work to the first order in the quark mass difference $m_d-m_u$. Consider the quark mass term in the QCD Hamiltonian density ${\cal H}_{\mbox{\tiny QCD}}$, as given by $$\begin{aligned} {\cal H}_{\mbox{\rm mass}}&=&m_u\overline{u}u+m_d\overline{d}d+m_s\overline{s}s+ \cdots \nonumber\\*[7.2pt] &=&\hat{m}(\overline{u}u+\overline{d}d)-{1\over 2}\delta m (\overline{u}u-\overline{d}d)+m_s\overline{s}s+\cdots\ . \label{xbreaking}\end{aligned}$$ The isospin symmetry is explicitly broken by the term proportional to $\delta m$. Using covariant normalization for the hadron state labeled by $h$ and momentum $k$ $$\langle k^\prime, h \mid k, h \rangle~~=~~ (2\pi)^3 {k^0} \delta^{(3)} (\vec k^\prime - \vec k)\ ,$$ and regarding the $\delta m$ term in Eq. (\[xbreaking\]) as a small parameter, we can make a perturbation expansion and write the shift in the mass $M_h$ of the hadron as [@gasser1] $$\delta (M_h)={- \delta m\over 2} {\langle h| (\overline u u - \overline d d)|h\rangle\over 2 M_h}\; . \label{delm}$$ The matrix element occurring in Eq. (\[delm\]) is to be computed at $\delta m = 0$, which means isospin can be taken to be exact. We can then write for the difference between the neutron and proton mass to first order in $\delta m$ as $$(M_n-M_p)_q =\delta m {\langle p | \overline u u - \overline d d | p \rangle\over 2 M_p}\; .$$ The subscript $q$ in the left-hand side denotes the fact that we are considering only the non-electromagnetic part of the mass difference. The effect of turning on electromagnetism is described by an effective Lagrangian $${\cal L}_{\rm e.m.} ~~=~~ - {1 \over 2} e^2 \int d^4y D(x-y) T j ^\mu (x) j_\mu (y) \; ,$$ where $D(x) = [i 4 \pi^2 (z^2-i \epsilon)]^{-1}$  is the photon propagator and $j^\mu(x)$  is the electromagnetic current. To remove the divergence arising from electromagnetism one must add counter terms, and these, of course, depend on renormalization prescription. However, this dependence is extremely weak. For example the change in the up quark mass due to a change in the renormalization scale by a factor of two is less than 0.01 MeV. It is therefore meaningful to separate the contribution from the quark mass difference, from that due to electromagnetism. The latter is estimated to be [@gasser1] $$(M_n-M_p)_{\mbox\tiny{\rm elec}}=-0.76\pm 0.30 \; .$$ The experimental mass difference is 1.29 MeV, which then gives $$(M_n-M_p)_q^{\mbox\tiny{\rm exp}}=2.05\pm 0.30 \; .$$ In the last section we saw that uncertainty in our knowledge of $\chi$ leads to a corresponding uncertainty in our determination of $H$. For the $\chi$ value obtained from chiral perturbation theory \[see Eq. (\[chiva\])\], we get $(M_n-M_p)_q\simeq 1.8\,\text{MeV}$. For $\chi$ values in the range $2.15\,\text{GeV}^{-1}\le\chi\le 2.80\,\text{GeV}^{-1}$, we find $$1.75\text{MeV}\le (M_n-M_p)_q \le 2.35\text{MeV} \,,$$ which is consistent with experimental data. Smaller and larger values of $\chi$ outside the range considered above lead to correspondingly smaller and larger values for the neutron-proton mass difference. Discussion ========== One of our main objectives in this paper has been to extract the proton matrix element $H=\langle p|\overline{u}u-\overline{d}d|p\rangle/2M_p$. We have seen that the chiral-odd sum rule Eq. (\[fiq\]) is reliable for determining this matrix element. The major limiting factor has been the uncertainty in the value of the susceptibility $\chi$. A more accurate evaluation of the two point function Eq. (\[chi\]) should reduce the uncertainty in the value of $\chi$ and hence help to pin down the value of the matrix element $H$. We also saw that the non-electromagnetic part of the neutron-proton mass difference is essentially given by the matrix element $H$ multiplied by the light quark mass difference $\delta m$. If we use a median value $\delta m = 3.3$ MeV and $H \simeq 0.54$ as obtained using a value of $\chi = 2.2 \text{GeV}^{-1}$ we get $(M_n-M_p)_q\simeq 1.8\,\text{MeV}$, which indeed has the right sign and magnitude. This suggest that our approach to extract $H$ and the neutron-proton mass difference is reliable. However, since $\chi$ and $\delta m $ are not precisely known we cannot make a critical comparison with data at present. We now turn to a comparison of our method with those of earlier authors. The nucleon mass was originally extracted by Ioffe [@ioffe1] with a [*combined*]{} use of both the chiral-odd and chiral-even mass sum rules. It is necessary to use these two sum rules together since one must eliminate the coupling constant $\lambda_N^2$. Belyaev and Ioffe [@belyaev1] extended this method to determine the mass splitting between hyperon and nucleon by treating the strange quark mass as a perturbation. In addition to the mass shift, $M_Y - M_N$, one must also take into account the change in the coupling constant $\lambda_Y ^2 - \lambda_N ^2$ and the change in the continuum threshold $s_0$. The authors of Refs. [@adami1; @eletsky1] used the same procedure to determine the mass splittings within an isospin multiplet by treating $m_d$ and $m_u$ as perturbation parameters. In Ref. [@hatsuda1] the neutron-proton mass was extracted directly from the difference between the neutron and proton mass sum rules, but the continuum contributions were disregarded. In Ref. [@adami1], Adami [*et.al.*]{} retained continuum corrections but regarded $\gamma $ as a parameter to be determined by a fit to all isospin splittings in the baryon octet. In Ref. [@yang1], apart from the perturbation due to the quark masses, an attempt was made to incorporate the electromagnetic contribution also phenomenologically in the sum rules. The sum rules derived by us in Sec. \[sumrule\] can also be derived directly from the mass sum rules. Writing $m_u =\hat m - \delta m /2$, $m_d =\hat m + \delta m /2$ and using $\chi = -\gamma /\delta m $ one can differentiate Eqs. (16) and (17) of Ref. [@yang1] with respect to $\delta m$. One can then identify our sum rules Eqs. (\[fiq\]) and (\[fi1\]) with Eqs. (29) and (30) of Ref. [@yang1]. (An assumption about the mixed chiral condensate, equivalent to our assumption, $\chi_m = \chi$, was made in Ref. [@yang1]). This coincidence between the sum rules is not surprising, since the quark mass term in the QCD Lagrangian can also be regarded as a constant external scalar field. The double pole term in the RHS of our sum rules clearly arises from the simple pole term in the mass sum rules after the differentiation of the proton mass with respect to $\delta m$. We note that the term proportional to the mixed quark-gluon condensate \[the fourth term in the LHS of our Eq. (\[fiq\])\] has not been included in Ref. [@adami1]. We have seen in our analysis of the sum rules (see Fig.2) that this term is numerically significant. There is also a minor discrepancy in the coefficient of the four-quark condensate term between Ref. [@adami1] and this work (or Ref. [@yang1]). This discrepancy comes from the fact that the authors of Ref. [@adami1] directly used the $\Sigma$ and $\Xi$ mass sum rules from Ref. [@belyaev1], where [*not*]{} all the quark mass terms were taken into account. It is now easy to see, as discussed earlier in Sec. \[sumrule\], why Eq. (\[fiq\]) is a better sum rule as compared to the chiral-even sum rule Eq. (\[fi1\]). In the chiral-odd sum rule, the perturbation due to the finite quark mass does not affect the leading asymptotic behavior and hence when a differentiation with respect to $\delta m$ (alternately when the neutron and proton sum rule difference is taken) the leading asymptotic behavior reduces to a smaller power of the Borel Mass $M^2$, which in turn means weaker dependence on the continuum contribution. On the other hand, in the chiral-even sum rule the introduction of quark mass leads to the term $m_d M^6$ in the proton and $m_u M^6$ in the neutron sum rule. Consequently in the difference the leading asymptotic behavior is now $M^6$. In other words, the continuum contributions are enhanced. This feature is also clearly reflected in our analysis in Sec. \[sumrule\]. For the chiral-odd sum rule the double pole term residue was stable when $s_0$ was varied while in the chiral-even sum rule the corresponding residue was unstable. Further, in the chiral-even sum rule the single pole term corresponding to transition between proton and the excited states dominated over the double pole term. We would also like to point out that inclusion of instanton contributions improves the chiral-even mass sum rule [@hil]. In the light of our observations above regarding the relation of our sum rules to the mass sum rules it suggests that such instanton contributions can also be significant in our sum rule Eq. (\[fi1\]). It is also worth emphasizing that in our work two stages are involved. We first calculate a hadronic matrix element $H$. This involves the susceptibility $\chi$, which being essentially the ratio $-\gamma$/$\delta m$ is relatively insensitive to errors in our knowledge of $\delta m$. The quark mass part of the neutron-proton mass difference was obtained as the product $H \delta m $. As a final remark we shall comment on the Nolen-Schiffer anomaly [@nsa]. We saw in the last section that the empirical neutron-proton mass difference can be written as $$(M_n-M_p)^{\mbox\tiny{\rm exp}}~~=-0.76 + \delta m H \; . \label{mnp}$$ Now, it is well known that the matrix element of the axial vector current is quenched [@ris] inside the nuclear medium by about 30 percent. It may be reasonable to assume that the isovector-scalar matrix element $\langle p| \overline u u - \overline d d |p\rangle/2M_p$ is also quenched in a similar fashion. Assuming then for example a 30 percent reduction in the value of $H$, it follows form Eq. (\[mnp\]) that the effective mass difference in the nuclear medium is $(M_n-M_p)^{\mbox\tiny{\rm exp}}_{\mbox\tiny{\rm{med}}} \simeq 0.49\text{ MeV}$. Understanding the Nolen-Schiffer anomaly is then reduced to explain the quenching of $\langle p | \overline u u - \overline d d | p \rangle/2M_p$ in nuclear medium. This can be handled either by traditional nuclear structure calculations or again by use of QCD sum rules as in the quenching of nucleon axial coupling [@pp; @dl]. We thank T. D. Cohen and H. Forkel for useful conversations, and S. L. Adler for a useful correspondence. One of us (J.P.) would like to thank D. Sen and A. D. Patel for discussions. X.J.acknowledges support from the National Science Foundation under Grant No. PHY-9058487 and from the Department of Energy under Grant No. DE–FG05–93ER–40762. M.N. acknowledges the warm hospitality and congenial atmosphere provided by the University of Maryland Nuclear Theory Group and support from FAPESP BRAZIL. M. A. Shifman, A. I. Vainshtein, and V. I. Zakharov, Nucl. Phys. [**B147**]{}, 385 (1979); [**B147**]{}, 448 (1979); [**B147**]{}, 519 (1979). B. L. Ioffe and A. V. Smilga, Nucl. Phys. [**B232**]{}, 109 (1984). V. M. Balitsky and A. V. Yung, Phys.Lett., 328 (1983). C. B. Chiu, J. Pasupathy, and S. L. Wilson, Phys. Rev. [**D33**]{}, 1961 (1986). V. M. Belyaev and Y. I. Kogan, Phys. Lett., 27 (1984). C. B. Chiu, J. Pasupathy, and S. L. Wilson, Phys. Rev. [**D32**]{}, 1786 (1985). A. V. Kolesnichenko, Sov. J. Nucl. Phys. [**39**]{}, 968 (1984). V. M. Belyaev and B. Y. Blok, Phys.Lett. [**B167**]{}, 99 (1986). X. Jin, M. Nielsen, and J. Pasupathy, Phys. Lett. B [**314**]{}, 163 (1993). J. Gasser and H. Leutwyler, Phys. Rep., 77 (1982). T. Hatsuda, H. H[ø]{}gaasen, and M. Prakash, Phys. Rev. [**C42**]{}, 2212 (1990). K.-C. Yang, W.-Y.P. Hwang, E.M. Henley, and L.S. Kisslinger, Phys. Rev. [**D48**]{}, 3001 (1993). C. Adami, E. G. Drukarev, and B. L. Ioffe, Phys. Rev. [**D48**]{}, 2304 (1993). V.L. Eletsky and B. L. Ioffe, Phys. Rev. [**D48**]{}, 1441 (1993). B. L. Ioffe, Nucl. Phys. [**B188**]{}, 317 (1981); [**B191**]{}, 591(E) (1981). V. M. Belyaev and B. L. Ioffe, Zh. Eksp. Teor. Fiz. [**83**]{}, 876 (1982) \[Sov. Phys. JETP [**56**]{}, 493 (1982)\]; [**84**]{}, 1236 (1983) \[[**57**]{}, 716 (1983)\]. D. B. Leinweber, Ann. Phys. (N.Y.) [**198**]{}, 203 (1990). For a review, see L. J. Reinders, H. Rubinstein, and S. Yazaki, Phys. Rep. [**127**]{}, 1 (1985), and references therein. V. Fock, Physikalische Zeitschrift der Sowjetunion [**12**]{}, 404 (1937). J. Schwinger, [*Particles, Sources, and Fields*]{} (Addison-Wesley, Reading, 1970), Vol. I. X. Jin, T. D. Cohen, R. J. Furnstahl, and D. K. Griegel, Phys. Rev. C [**47**]{}, 2882 (1993), and references therein. J. Gasser and H. Leutwyler, Ann. Phys., 142 (1984). J. Gasser and H. Leutwyler, Nucl. Phys., 465 (1985). N. Paver, Riazzudin, and M. D. Scadron, Phys.Lett. [**B197**]{}, 430(1987). P. Pascual and R. Tarrach, Phys. Lett., 443(1982). E. Bagan [**et al.**]{}, Phys. Lett. [**135B**]{}, 463(1984). C. A. Dominguez and M. Loewe, Phys. Rev., 2930(1985). C. A. Dominguez and E. de Rafael, Ann.Phys. (N.Y.) [**174**]{}, 372(1987). S. Narison, Rev. Nuovo Cimento [**10**]{}, 1(1987); [*QCD Spectral Sum Rules*]{}, World Scientific Lecture Notes in Physics Vol. 26(World Scientific, Singapore, 1989). R. J. Furnstahl, D. K. Griegel, and T. D. Cohen, Phys. Rev. C [**46**]{}, 1507 (1992); X. Jin, M. Nielsen, T. D. Cohen, R. J. Furnstahl, and D. K. Griegel, Phys. Rev. C [**49**]{}, 464 (1994); X. Jin and R. J. Furnstahl, Phys. Rev. C [**49**]{}, 1190 (1994). H. Forkel and M. K. Banerjee, Phys. Rev. Lett. [**71**]{}, 484(1993). J. A. Nolen and J. P. Schiffer, Ann. Rev. Nuc. Sci. [**19**]{}, 414 (1969). D. O. Riska and K. Tsushima, Nucl. Phys. [**A553**]{}, 684c (1993). R. Parthasarathy and J. Pasupathy, Phys. Rev. C [**37**]{}, 2140 (1988). E. G. Drukarev and E. M. Levin, Prog. Part. Nucl. Phys. [**27**]{}, 77 (1991); Nucl. Phys. [**A511**]{}, 679 (1990); [**A516**]{}, 715(E) (1990). [^1]: Permanent address: Instituto de Física, Universidade de São Paulo, 01498 - SP- Brazil.

--- abstract: 'We consider the *Navier-Stokes* equations in a channel with a narrowing and walls of varying curvature. By applying the *empirical interpolation method* to generate an affine parameter dependency, the offline-online procedure can be used to compute reduced order solutions for parameter variations. The reduced order space is computed from the steady-state snapshot solutions by a standard POD procedure. The model is discretised with high-order spectral element *ansatz* functions, resulting in 4752 degrees of freedom. The proposed reduced order model produces accurate approximations of steady-state solutions for a wide range of geometries and kinematic viscosity values. The application that motivated the present study is the onset of asymmetries (i.e., symmetry breaking bifurcation) in blood flow through a regurgitant mitral valve, depending on the Reynolds number and the valve shape. Through our computational study, we found that the critical Reynolds number for the symmetry breaking increases as the wall curvature increases.' author: - title: 'Reduced Basis Model Order Reduction for Navier-Stokes equations in domains with walls of varying curvature' --- Navier–Stokes equations; Reduced order methods; Reduced basis methods; Parametric geometries; Symmetry breaking bifurcation Introduction and Motivation =========================== We consider the flow of an incompressible fluid through a planar channel with a narrowing, where the walls creating the narrowing have variable curvature. An application that motivated the present study is the flow of blood through a regurgitant mitral valve. Mitral regurgitation (MR) is a valvular disease characterized by abnormal leaking of blood through the mitral valve from the left ventricle into the left atrium of the heart. See Fig. \[MitralValve\]. In certain cases, the regurgitant jet “hugs" the wall of the heart’s atrium as shown in Fig. \[MitralValve\] (right). These wall-hugging, non-symmetric regurgitant jets have been observed at low Reynolds numbers ([@regurgRe1000; @regurgRe50]) and are said to undergo the Coanda effect ([@Hess:wille_fernholz_1965]). Such jets represent one of the biggest challenges in echocardiographic assessment of MR ([@Hess:Ginghina2007]). In ([@Hess:Quaini2016; @Hess:Pitton2017534; @Hess:Pitton2017; @Hess:Wang2017; @Hess:max_bif; @Hess:ICOSAHOM2018]), we made a connection between the cardiovascular and bioengineering literature reporting on the Coanda effect in MR and the fluid dynamics literature with the goal of identifying and understanding the main features of the corresponding flow conditions. .7cm In ([@Hess:Quaini2016; @Hess:Pitton2017534; @Hess:Pitton2017; @Hess:Wang2017; @Hess:max_bif; @Hess:ICOSAHOM2018]), we studied what triggers the Coanda effect in a simplified setting by reformulating the problem in terms of the hydrodynamic stability of solutions of the incompressible Navier-Stokes equations in contraction-expansion channels with straight walls. Such channels have the same geometric features of MR and the wall-hugging effect is nothing but a symmetry breaking bifurcation. Here, we extend the study to contraction-expansion channels with curved walls as a first step towards more realistic geometries and eventually fluid-structure interaction. For numerical treatment of the incompressible Navier–Stokes problem, we apply the spectral element method (SEM), which uses high-order polynomial *ansatz* functions such as Legendre polynomials. See, e.g., ([@Hess:PateraSEM; @Hess:CHQZ1; @Hess:CHQZ2]), and ([@Hess:Sherwin:2005; @Hess:Taddei]) for applications in fluid dynamics. With a coarse partitioning of the computational domain into spectral elements, the high-order *ansatz* functions are prescribed over each element. The *ansatz* functions are modified for numerical stability and to enable continuity across element boundaries. Let $p$ be the order of the polynomial. Typically, an exponential error decay under $p$-refinement can be observed, which provides computational advantages over more standard finite element methods. Varying wall curvature and kinematic viscosity are considered for the parametric model order reduction. From a set of sampled high-order solves, a reduced order model is generated, which approximates the high-order solutions and allows fast parameter sweeps of the two-dimensional parameter domain. The offline-online decomposition required for fast reduced order parameter evaluations is established with the *empirical interpolation method* ([@Hess:Barrault], [@Hess:Chaturantabut], [@Hess:Quarteroni2007], [@Hess:Rozza2009], [@Hess:Maday2015]). The reduced-order modeling (ROM) techniques described in this work are implemented in open-source project ITHACA-SEM[^1]. This extends our previous work ([@Hess:ICOSAHOM2018], [@Hess:ENUMATH17_me]) to bifurcations in geometries with non-affine geometry variations. The outline of the paper is as follows. In sec. 2, the model problem is defined and the parametric variations are explained. Sec. 3 provides details on the spectral element discretization and sec. 4 explains the model order reduction with the empirical interpolation. Numerical results are presented in sec. 5, while in sec. 6 conclusions are drawn and future perspectives and developments are discussed. Problem Formulation =================== Let $\Omega \in \mathbb{R}^2$ be the computational domain. Incompressible, viscous fluid motion in spatial domain $\Omega$ over a time interval $(0, T)$ is governed by the incompressible *Navier-Stokes* equations: $$\begin{aligned} \frac{\partial \mathbf{u}}{\partial t} + \mathbf{u} \cdot \nabla \mathbf{u} &=& - \nabla p + \nu \Delta \mathbf{u} + \mathbf{f}, \label{Hess:NSE0} \\ \nabla \cdot \mathbf{u} &=& 0, \label{Hess:NSE1}\end{aligned}$$ where $\mathbf{u}$ is the vector-valued velocity, $p$ is the scalar-valued pressure, $\nu$ is the kinematic viscosity and $\mathbf{f}$ is a body forcing. Boundary and initial conditions are prescribed as $$\begin{aligned} \mathbf{u} &=& \mathbf{d} \quad \text{ on } \Gamma_D \times (0, T), \\ \nabla \mathbf{u} \cdot \mathbf{n} &=& \mathbf{g} \quad \text{ on } \Gamma_N \times (0, T), \\ \mathbf{u} &=& \mathbf{u}_0 \quad \text{ in } \Omega \times 0, \label{Hess:NSE_boundaryCond}\end{aligned}$$ with $\mathbf{d}$, $\mathbf{g}$ and $\mathbf{u}_0$ given and $\partial \Omega = \Gamma_D \cup \Gamma_N$, $\Gamma_D \cap \Gamma_N = \emptyset$. The *Reynolds* number $Re$, which characterizes the flow ([@Hess:Holmes]), depends on $\nu$, a characteristic velocity $U$, and a characteristic length $L$: $$\label{eq:re} Re = \frac{UL}{\nu}.$$ We are interested in the steady states, i.e., solutions where $\frac{\partial \mathbf{u}}{\partial t}$ vanishes. The high-order simulations are obtained through time-advancement, while the reduced order solutions are computed through fixed-point iterations. Non-linear solver ----------------- The *Oseen*-iteration is a secant modulus fixed-point iteration, which in general exhibits a linear rate of convergence ([@Hess:Oseen]). It solves for a steady-state solution, i.e., $\frac{\partial \mathbf{u}}{\partial t} = 0$ is assumed. Given a current iterate (or initial condition) $\mathbf{u}^k$, the next iterate $\mathbf{u}^{k+1}$ is found by solving the following linear system: $$\begin{aligned} -\nu \Delta \mathbf{u}^{k+1} + (\mathbf{u}^k \cdot \nabla) \mathbf{u}^{k+1} + \nabla p &=& \mathbf{f} \text{ in } \Omega, \label{Hess:eq_Oseen_main} \cl \nabla \cdot \mathbf{u}^{k+1} &=& 0 \text{ in } \Omega, \cl \mathbf{u}^{k+1} &=& \mathbf{d} \text{ on } \Gamma_D, \cl \nabla \mathbf{u}^{k+1} \cdot \mathbf{n} &=& \mathbf{g} \text{ on } \Gamma_N. \el\end{aligned}$$ Iterations are stopped when the relative difference between iterates falls below a predefined tolerance in a suitable norm, like the $L^2(\Omega)$ or $H^1_0(\Omega)$ norm. Model Description ----------------- We consider the channel flow through a narrowing created by walls of varying curvature and with variable kinematic viscosity. See Fig. \[Hess:FOM\_straight\] and \[Hess:FOM\_curved\] for the steady-state velocity components for $\nu = 0.15$ in a geometry with straight walls and curved walls, respectively. In all the cases under consideration, the spectral element expansion uses modal Legendre polynomials of order $p = 10$ for the velocity. The pressure *ansatz* space is chosen of order $p-2$ to fulfill the inf-sup stability condition ([@Hess:infsup; @Hess:BBF]). A parabolic inflow profile is prescribed at the inlet (i.e., $x = 0$) with horizontal velocity component $u_x(0,y) = y(3-y)$ for $y \in [0, 3]$. At the outlet (i.e., $x = 18$) we impose a stress-free boundary condition, while everywhere else a no-slip condition is prescribed. We consider symmetric boundary conditions, because we want to study the symmetry breaking due to the nonlinearity in problem - . For a more realistic setting one would have to account for different inlet velocity profiles and the pulsatility of the flow (i.e., include the Strouhal number among the parameters). ![Full order, steady-state solution in the geometry with straight walls and for $\nu = 0.15$: velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_straight"}](straight_walls_u.png "fig:") $\quad$ ![Full order, steady-state solution in the geometry with straight walls and for $\nu = 0.15$: velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_straight"}](straight_walls_u_legend.png "fig:")\ ![Full order, steady-state solution in the geometry with straight walls and for $\nu = 0.15$: velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_straight"}](straight_walls_v.png "fig:") $\quad$ ![Full order, steady-state solution in the geometry with straight walls and for $\nu = 0.15$: velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_straight"}](straight_walls_v_legend.png "fig:") ![Full order, steady-state solution in the geometry with curved walls with the largest considered curvature and for $\nu = 0.15$ : velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_curved"}](curved_walls_u.png "fig:") $\quad$ ![Full order, steady-state solution in the geometry with curved walls with the largest considered curvature and for $\nu = 0.15$ : velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_curved"}](curved_walls_u_legend.png "fig:")\ ![Full order, steady-state solution in the geometry with curved walls with the largest considered curvature and for $\nu = 0.15$ : velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_curved"}](curved_walls_v.png "fig:") $\quad$ ![Full order, steady-state solution in the geometry with curved walls with the largest considered curvature and for $\nu = 0.15$ : velocity in x-direction (top) and y-direction (bottom).[]{data-label="Hess:FOM_curved"}](curved_walls_v_legend.png "fig:") Each curved wall is defined by a second order polynomial, interpolating three prescribed points. While the points at the domain boundary $y=0$ and $y=3$ are kept fixed, the inner points are moved towards $x=0$ in order to create an increasing curvature. The viscosity varies in the interval $\nu \in [0.15, 0.2]$. We recall that the Reynolds number $Re$ depends on the kinematic viscosity. As $Re$ is varied for each fixed geometry, a supercritical pitchfork bifurcation occurs: for $Re$ higher than the critical bifurcation point, three solutions exit. Two of these solutions are stable, one with a jet towards the top wall and one with a jet towards the bottom wall, and one is unstable. The unstable solution is symmetric to the horizontal centerline at $y=1.5$, while the jet of the stable solutions is said to undergo the Coanda effect. In this investigation, we do not deal with recovering all bifurcation branches, but limit our attention to the stable branch of solutions with jets hugging the bottom wall. However, we remark that recovering all bifurcating solutions with model reduction methods is also possible. See, e.g., ([@Hess:Herrero2013132]). Spectral Element Full Order Discretization ========================================== The spectral/hp element software framework we use for the numerical solution of problem is Nektar++, version 4.4.0[^2]. The large-scale discretized system that has to be solved at each step of the *Oseen*-iteration can be written as $$\begin{aligned} \begin{bmatrix} \begin{array}{ccc} A & -D^T_{bnd} & B \\ -D_{bnd} & 0 & -D_{int} \\ \tilde{B}^T & -D^T_{int} & C \end{array} \end{bmatrix} \begin{bmatrix} \begin{array}{ccc} \mathbf{v}_{bnd} \\ \mathbf{p} \\ \mathbf{v}_{int} \end{array} \end{bmatrix} &= \begin{bmatrix} \begin{array}{ccc} \mathbf{f}_{bnd} \\ \mathbf{0} \\ \mathbf{f}_{int} \end{array} \end{bmatrix}, \label{Hess:fully_expanded}\end{aligned}$$ for fixed parameter vector $\boldsymbol{\mu}$, which denotes the geometrical and physical parameters. In , $\mathbf{v}_{bnd}$ and $\mathbf{v}_{int}$ denote the arrays of the velocity degrees of freedom on the boundary and in the interior of the domain, respectively. The array of the pressure degrees of freedom is denoted by $\mathbf{p}$. The forcing terms on the boundary and interior are $\mathbf{f}_{bnd}$ and $\mathbf{f}_{int}$, respectively. Next, we explain the matrix blocks. Matrix $A$ assembles the boundary-boundary velocity coupling, $B$ the boundary-interior velocity coupling, $\tilde{B}$ the interior-boundary velocity coupling, and $C$ assembles the interior-interior velocity degree of freedom coupling. The matrices $D_{bnd}$ and $D_{int}$ assemble the pressure-velocity boundary and pressure-velocity interior contributions. Due to the varying geometry, each matrix is dependent on the parameter $\boldsymbol{\mu}$. The linear system is assembled in local degrees of freedom, i.e., *ansatz* functions with support extending over spectral element boundaries are treated seperately for each spectral element. See ([@Hess:Sherwin:2005]) for detailed explanations. As a result, matrices $A, B, \tilde{B}, C, D_{bnd}$ and $D_{int}$ have a block structure, with each block corresponding to a spectral element. This allows for an efficient matrix assembly since each spectral element is independent from the others, but the local degrees of freedom need to be gathered into the global degrees of freedom in order to obtain a non-singular system. The boundary-boundary global element coupling is achieved with the rectangular assembly matrix $M$, which gathers the local boundary degrees of freedom. Multiplication of the first row of by $M^T M$ sets the boundary-boundary coupling in local degrees of freedom: $$\begin{aligned} \begin{bmatrix} \begin{array}{ccc} M^T M A & -M^T M D^T_{bnd} & M^T M B \\ -D_{bnd} & 0 & -D_{int} \\ \tilde{B}^T & -D^T_{int} & C \end{array} \end{bmatrix} \begin{bmatrix} \begin{array}{ccc} \mathbf{v}_{bnd} \\ \mathbf{p} \\ \mathbf{v}_{int} \end{array} \end{bmatrix} &= \begin{bmatrix} \begin{array}{ccc} M^T M \mathbf{f}_{bnd} \\ \mathbf{0} \\ \mathbf{f}_{int} \end{array} \end{bmatrix} . \label{Hess:fully_expanded_MtM}\end{aligned}$$ The action of the matrix in on the prescribed Dirichlet boundary conditions is computed and added to the source term. Since the Dirichlet boundary conditions are known, the corresponding equations are removed from the system. Let $N_\delta$ denote the system size after removal of the known boundary conditions. The resulting system of high-order dimension $N_\delta \times N_\delta$ is composed of the block matrices and depends on the parameter $\boldsymbol{\mu}$. For simplicity of notation, we will write such system in compact form as: $$\begin{aligned} \mathcal{A}(\boldsymbol{\mu}) \mathbf{x}(\boldsymbol{\mu}) = \mathbf{f}(\boldsymbol{\mu}). \label{Hess:final_to_be_proj}\end{aligned}$$ Reduced Order Space Generation ============================== The ROM computes an approximation to the full order model using a few modes of the POD as *ansatz* functions ([@Hess:Lassila2014]). To achieve a computational speed-up, the matrix assembly for a new parameter of interest is independent of the large-scale discretization size $N_\delta$. The *empirical interpolation method* (see sec. \[sec:EIM\]) computes an affine parameter dependency, which enables an offline-online decomposition (see sec. \[sec:off-on\]). After a time-intensive offline phase, reduced order solves can be evaluated quickly over the parameter range of interest. The POD of $N$ (typically small) uniformly sampled full-order solves, called *snapshots*, is performed. The most dominant modes corresponding to $99.99\%$ of the POD energy (as suggested in [@Hess:Lassila2014]) form the projection matrix $U \in \mathbb{R}^{N_\delta \times N}$ and implicitly define the low-order space $V_N = \text{span} (U)$. The large-scale system is then projected onto the reduced order space: $$\begin{aligned} U^T \mathcal{A}(\boldsymbol{\mu}) U \mathbf{x}_N(\boldsymbol{\mu}) = U^T \mathbf{f}(\boldsymbol{\mu}) . \label{Hess:final_proj}\end{aligned}$$ The low order solution $\mathbf{x}_N(\boldsymbol{\mu})$ approximates the large-scale solution as $\mathbf{x}(\boldsymbol{\mu}) \approx U \mathbf{x}_N(\boldsymbol{\mu})$. The stability properties of the full-order model do not necessarily carry over to the reduced-order model, which can introduce instabilities. In particular, the reduced order inf-sup stability constant might approach zero for some parameter value, while the full-order inf-sup stability constant is bounded away from zero. One way to alleviate this problem is by using inf-sup supremizers ([@Hess:Lassila2014]) or considering space-time variational approaches ([@Hess:Yano2013]). Empirical Interpolation Method {#sec:EIM} ------------------------------ The discrete *empirical interpolation method* (EIM) ([@Hess:Barrault], [@Hess:Chaturantabut], [@Hess:Quarteroni2007], [@Hess:Rozza2009]) computes an approximate affine parameter dependency. During the snapshot computation, the parameter-dependent matrices are collected. The matrix discrete empirical interpolation ([@Hess:Negri]) allows to decompose $\mathcal{A}(\boldsymbol{\mu})$ as follows: $$\begin{aligned} \mathcal{A}(\boldsymbol{\mu}) \approx \sum_{i=1}^{Q_a} \tau_i(\boldsymbol{\mu}) A_i, \label{Hess:MDEIM}\end{aligned}$$ where $\tau_i(\boldsymbol{\mu})$ are scalar parameter-dependent coefficient functions and $A_i$ are parameter-independent matrices. Each coefficient function $\tau_i(\boldsymbol{\mu})$ corresponds to a single matrix entry of $\mathcal{A}(\boldsymbol{\mu})$. Since the assembly of only a few (here $Q_a < 30$) matrix entries can be implemented efficiently, an approximation of $\mathcal{A}(\boldsymbol{\mu})$ is readily available for each new $\boldsymbol{\mu}$. As for the projection space $U$, $99.99\%$ of the POD energy is used to approximate the system matrices from the collected matrices during snapshot computation. The EIM is preformed for each submatrix identified in and the actual system matrix $\mathcal{A}(\boldsymbol{\mu})$ is then composed of the separately approximated block matrices. Offline-Online Decomposition {#sec:off-on} ---------------------------- The offline-online decomposition ([@Hess:RBref]) enables the computational speed-up of the ROM approach in many-query scenarios. It relies on an affine parameter dependency, such that all computations depending on the high-order model size $N_\delta$ can be performed in a parameter-independent offline phase. Then, the input-output evaluation performed online is independent of $N_\delta$ and thus fast. After applying the *empirical interpolation method* in the geometry parameter, the parameter dependency is cast in an affine form. Therefore, there exists an affine expansion of the system matrix $\mathcal{A}(\boldsymbol{\mu})$ in the parameter $\boldsymbol{\mu}$ given by . To achieve fast reduced order solves, the offline-online decomposition computes the parameter independent projections offline, which are stored as small-sized matrices of the order $N \times N$. Since in an *Oseen*-iteration each matrix is dependent on the previous iterate, the submatrices corresponding to each basis function are assembled and then formed online using the affine expansion computed from the EIM and a fast evaluation of a single matrix entry as required by the EIM coefficient functions $\tau_i$. Numerical Results ================= Snapshot solutions are sampled over a uniform $8 \times 9$ grid from the full-order model, with $8$ samples along the $\nu$ parameter direction and $9$ samples along the geometry parameter direction. The number of required snapshot computations might potentially be reduced when using a greedy sampling, which requires error indicators or error estimators. See [@Hess:RBref]. Error estimation does even allow a certification of the ROM accuracy, but it requires an estimation of the inf-sup constant as well as a bound on the *empirical interpolation* error. The vertical velocity at the point $(2, 1.5)$ is used to generate the bifurcation diagram reported in Fig. \[Hess:FOM\_sampled\]. As expected, in a fixed geometry the symmetry breaking bifurcation occurs when the $Re$ exceeds a critical value. It is very interesting to observe that such critical value increases as the wall curvature increases, i.e. as the walls get more curved, the stronger the inertial forces need to be to break the symmetry of the solution. This means that the estimates for the critical $Re$ in 3D geometries with straight walls found in [@Hess:Pitton2017534; @Hess:Wang2017] provide a lower bound for the critical $Re$ at which the Coanda effect is observed in vivo (see Fig. \[MitralValve\] (right)). We remark that the 2D case can be seen as a limit of the 3D case for channel depth tending to infinity. In [@Hess:Pitton2017534], the influence of the channel depth on the flow pattern is investigated. It is shown that non-symmetry jets appear at higher $Re$ as the channel depth gets smaller. Thus, we expect that the critical $Re$ for the symmetry breaking in a 3D channel with curved walls to be higher than the values reported here for a 2D channel. This indicates the Coanda effect occurs in mitral valves with elongated orifices (corresponding to deeper channels). ![Bifurcation diagram of the full order model over a uniform $8 \times 9$ grid.[]{data-label="Hess:FOM_sampled"}](FOMdata_added_insets.png) The accuracy of the ROM is assessed using $N = 72$ snapshots for the POD to recover the original snapshot data. Fig. \[Hess:PODdecay\] shows the decay of the energy of the POD modes. To reach the typical threshold of $99.99\%$ on the POD energy, $N = 33$ POD modes are required as RB ansatz functions. ![Decay of POD energy of the sampled snapshot solutions.[]{data-label="Hess:PODdecay"}](PODdecay.pdf) Fig. \[Hess:ROM20\] shows the bifurcation diagram of the reduced order model with $N = 20$ basis functions. The absolute error at the point value is less than $0.01$ at $46$ parameter locations and less than $0.1$ at $63$ parameter locations. This indicates, that the high-order solutions have been well-resolved at these configurations. There are a few outliers, where the iteration scheme did not converge and the value for the last iterate is shown. Most likely this can be resolved by taking a finer snapshot sampling into account or by using a localized reduced-order modeling approach ([@Hess:max_bif]). ![Bifurcation diagram of the reduced order model with reduced model size $20$.[]{data-label="Hess:ROM20"}](ROMdata20.pdf) The *empirical interpolation method* relies on the fast computation of a few matrix entries during the online phase. Since the spectral element *ansatz* functions have support over a whole spectral element, this operation cannot be performed as fast as with a finite element or finite volume method for instance, where *ansatz* functions have a local support. Nevertheless, the computational gain is significant after the affine form has been established. The time requirement for a single fixed point iteration step reduces from about $10$ s to $0.1$ s. Conclusion and Outlook ====================== We proposed a reduced order model that combines *empirical interpolation method* and a POD reduced basis technique to recover full-order solutions of the Navier–Stokes equations in domains with walls of varying curvature (non-affine variation). The non-linear geometry changes allow to simulate more realistic scenarios in the context of the Coanda effect in cardiology, but also require a fine sampling at the snapshot and *empirical interpolation* level. Since the model problem studied here undergoes a supercritical pitchfork bifurcation, introducing non-unique solutions, further numerical techniques are required to recover all bifurcation branches. The spectral element method is a suitable method for these tasks. However, the computational gain that one can expect is not as significant as in the case of methods using *ansatz* functions with a local support, such as the finite element method. As a next step, we will enhance the reduced order model proposed here by using localizes bases in order to recover every solutions in the considered parameter domain with high accuracy. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by European Union Funding for Research and Innovation through the European Research Council (project H2020 ERC CoG 2015 AROMA-CFD project 681447, P.I. Prof. G. Rozza). This work was also partially supported by NSF through grant DMS-1620384 (A. Quaini). Albers, J., Nitsche, T., Boese, J., Simone, R. D., Wolf, I., Schroeder, A., Vahl, F.: Regurgitant jet evaluation using three-dimensional echocardiography and magnetic resonance. Ann. Thorac. Surg., [**78**]{} (2004), 96–102. Vermeulen, M., Kaminsky, R., Smissen, B. V. D., Claessens, T., Segers, P., Verdonck, P., Ransbeeck, P. V.: In vitro flow modelling for mitral valve leakage quantification. In: Proc. 8th Int. Symp. Particle Image Velocimetry (2009), page 4. Ginghina, C.: The Coanda Effect in Cardiology. Journal of Cardiovascular Medicine, [**8**]{} (2007), 411–413. Quaini, A., Glowinski, R., Canic, S.: Symmetry breaking and preliminary results about a Hopf bifurcation for incompressible viscous flow in an expansion channel. International Journal of Computational Fluid Dynamics, [**30**]{}:1 (2016), 7–19. Pitton, G., Quaini, A., Rozza, G.: Computational Reduction Strategies for the Detection of Steady Bifurcations in Incompressible Fluid-Dynamics: Applications to *Coanda* Effect in Cardiology. Journal of Computational Physics, [**344**]{} (2017), 534–557. Wang, Y., Quaini, A., Canic, S., Vukicevic, M., Little, S.H.: 3D Experimental and Computational Analysis of Eccentric Mitral Regurgitant Jets in a Mock Imaging Heart Chamber", Cardiovascular Engineering and Technology, [**8**]{}:4, (2017), 419–438. Canuto, C., Hussaini, M.Y., Quarteroni, A., Zhang, Th.A.: Spectral [M]{}ethods [F]{}undamentals in [S]{}ingle [D]{}omains. In: Springer – Scientific [C]{}omputation, (2006). Lassila, T., Manzoni, A., Quarteroni, A., Rozza, G.: Model Order Reduction in Fluid Dynamics: Challenges and Perspectives. In: Reduced Order Methods for Modelling and Computational Reduction, Springer International Publishing, MS&A, Vol. 9, A. Quarteroni, G.Rozza eds. (2014), 235–273. Canuto, C., Hussaini, M.Y., Quarteroni, A., Zhang, Th.A.: Spectral [M]{}ethods [E]{}volution to [C]{}omplex [G]{}eometries and [A]{}pplications to [F]{}luid [D]{}ynamics. In: Springer – Scientific [C]{}omputation, (2007). Karniadakis, G., Sherwin, S. 2005. *Spectral/hp Element Methods for Computational Fluid Dynamics.* Oxford University Press, 2nd ed. (2005). Boffi, D., Brezzi F., Fortin, M. 2013. *Mixed Finite Element Methods and Applications.* Springer Series in Computational Mathematics. Hess, M.W., Quaini, A., Rozza, G. 2018. *A Spectral Element Reduced Basis Method for Navier-Stokes Equations with Geometric Variations.* ICOSAHOM 2018 conference proceeding. Submitted. ArXiv preprint 1812.11051. Quarteroni, A., Valli, A. 1994. *Numerical Approximation of Partial Differential Equations.* Springer-Verlag, Berlin-Heidelberg. Wille, R., Fernholz, H. 1965. *Report on the first European Mechanics Colloquium, on the Coanda effect.* Journal of Fluid Mechanics, [**23**]{}:4, 801–819. Burger, M. 2010. *Numerical Methods for Incompressible Flow.* Lecture Notes, UCLA, `http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.9.5470`. Holmes, P., Lumley, J., Berkooz, G. 1996. *Turbulence, Coherent Structures, Dynamical Systems and Symmetry.* Cambridge University Press. Fick, L., Maday, Y., Patera A., Taddei T.: A stabilized POD model for turbulent flows over a range of Reynolds numbers: Optimal parameter sampling and constrained projection. Journal of Computational Physics, [**371**]{} (2018), 214–243. Patera, A.T.: A Spectral Element Method for Fluid Dynamics; Laminar Flow in a Channel Expansion. Journal of Computational Physics, [**54**]{}:3 (1984), 468–488. Herrero, H., Maday, Y., Pla, F.: RB (Reduced Basis) for RB (Rayleigh–Bénard). Computer Methods in Applied Mechanics and Engineering, [**261–262**]{}, (2013), 132–141. Barrault, M., Maday, Y., Nguyen, N.C., Patera, A.T.: An ‘empirical interpolation’ method: application to efficient reduced-basis discretization of partial differential equations. Comptes Rendus Mathematique, [**339**]{}:9, (2004), 667–672. Negri, F., Manzoni, A., Amsallam, D.: Efficient model reduction of parametrized systems by matrix discrete empirical interpolation. Journal of Computational Physics, [**303**]{}, (2015), 431–454. Chaturantabut, S., Sorensen, D.C.: Nonlinear Model Reduction via Discrete Empirical Interpolation. [SIAM]{} Journal on Scientific Computing, [**32**]{}:5, (2010), 2737–2764. Hesthaven, J.S., Rozza, G., Stamm, B.: Certified Reduced Basis Methods for Parametrized Partial Differential Equations. In: SpringerBriefs in Mathematics, (2016). Hess, M.W., Rozza, G.: A Spectral Element Reduced Basis Method in Parametric [CFD]{}. Numerical Mathematics and Advanced Applications - ENUMATH 2017, Springer, in press, ArXiv e-print 1712.06432, (2018). Hess, M.W., Alla, A., Quaini, A., Rozza, G., Gunzburger, M.: A Localized Reduced-Order Modeling Approach for PDEs with Bifurcating Solutions. ArXiv e-print 1807.08851, (2018). Pitton, G., Rozza, G.: On the Application of Reduced Basis Methods to Bifurcation Problems in Incompressible Fluid Dynamics. Journal of Scientific Computing, (2017). Quarteroni, A., Rozza, G.: Numerical solution of parametrized Navier-Stokes equations by reduced basis methods. Numerical Methods for Partial Differential Equations, [**23**]{}:4, (2007), 923–948. Rozza, G.: Reduced basis methods for Stokes equations in domains with non-affine parameter dependence. Computing and Visualization in Science, [**12**]{}:1, (2009), 23–35. Maday, Y., Mula, O., Patera, A.T., Yano, M.: The generalized Empirical Interpolation Method: stability theory on Hilbert spaces with an application to the Stokes equation. Computational Methods in Applied Mechanics and Engineering, [**287**]{}, 2015, 310–334. Yano, M., Patera, A.T.: A space-time variational approach to hydrodynamic stability theory. Proceedings of the Royal Society, A, 469(2155): Article Number 20130036, 2013. [^1]: `https://github.com/mathLab/ITHACA-SEM` and `https://mathlab.sissa.it/ITHACA-SEM` [^2]: See **www.nektar.info**.

--- abstract: 'We give an example of a Morita algebra $A$ with a tilting module $T$ such that the algebra $End_A(T)$ has dominant dimension at least two but is not a Morita algebra. This provides a counterexample to a conjecture by Chen and Xi from [@CX].' address: - 'Institute of algebra and number theory, University of Stuttgart, Pfaffenwaldring 57, 70569 Stuttgart, Germany' - 'Institute of algebra and number theory, University of Stuttgart, Pfaffenwaldring 57, 70569 Stuttgart, Germany' author: - 'Bernhard B[ö]{}hmler' - René Marczinzik title: On a conjecture about Morita algebras --- Introduction {#introduction .unnumbered} ============ In this article we assume that all rings are finite dimensional algebras over a field $K$ and all modules are finitely generated right modules unless stated otherwise. Recall that the dominant dimension $domdim(M)$ of a module $M$ with minimal injective coresolution $(I^i)$ is defined as zero in case $I^0$ is not projective and $domdim(M):= \sup \{ n \geq 0 | I^i$ is projective for $i=0,1,...,n \}+1$ otherwise. The dominant dimension of an algebra is defined as the dominant dimension of the regular module. It is well known that an algebra has dominant dimension at least one if and only if there is a minimal faithful projective-injective right module $eA$ for some idempotent $e$ of $A$. All Nakayama algebras have dominant dimension at least one and therefore have a minimal faithful projective-injective module given by the direct sum of all indecomposable projective-injective modules, see for example chapter 32 of [@AnFul]. For more on the dominant dimension we refer to [@Yam]. In [@KY] the authors introduced Morita algebras as algebras $A$ that are algebras with dominant dimension at least two and a minimal faithful projective-injective module $eA$ such that $eAe$ is selfinjective. Morita algebras contain several important classes of algebras such as Schur algebras $S(n,r)$ for $n \geq r$ or blocks of category $\mathcal{O}$ and provide a useful generalisation of selfinjective algebras. At the end of the article [@CX] the authors provided three conjectures related to the dominant dimension of algebras. Their third conjecture states the following: Suppose two algebras $A$ and $B$ are derived equivalent. If $A$ is a Morita algebra and the dominant dimension of $B$ is at least two then also $B$ is a Morita algebra. In [@CX] several special cases of this conjecture were proven. In this article we give a counterexample to this conjecture. Let $A$ be the Nakayama algebra with Kupisch series \[4,5,4,5\] with vertices numbered from 0 to 3. Let $M$ be the module $e_0 A \oplus e_1 A \oplus e_3 A \oplus e_1 A/e_1 J^4$. Then $A$ is a Morita algebra and $M$ is a tilting module of projective dimension two such that the algebra $B:=End_A(M)$ is an algebra of dominant dimension equal to 4 that is not a Morita algebra. Note that $B$ is derived equivalent to $A$, since endomorphism algebras of tilting modules are derived equivalent to the original algebra. Therefore, our theorem gives a counterexample to the conjecture. We found the counterexample to the conjecture while experimenting with the GAP-package QPA, see [@QPA]. We thank Hongxing Chen and Changchang Xi for useful discussions in Stuttgart and Changchang Xi for proofreading and useful suggestions. Proof of the theorem ==================== In this section we give a proof of the theorem that we group into several smaller lemmas. We assume that the reader is familiar with the basics of the representation theory of finite dimensional algebras as explained for example in [@ARS] or [@ASS]. We use the conventions of [@ASS]. Thus we use right modules and write arrows in quiver algebras from left to right. For background on Nakayama algebras and how to calculate projective or injective resolutions for modules in such algebras we refer to [@Mar]. All algebras will be given by quiver and relations and are connected. Recall that the Kupisch series of a Nakayama algebra is just the sequence $[a_0,a_1,...,a_r]$ when $a_i$ denotes the dimension of the indecomposable projective modules corresponding to point $i$. Let $A$ always be the Nakayama algebra with Kupisch series \[4,5,4,5\]. Thus $A$ is a quiver algebra with a cyclic quiver. We assume that the vertices are numbered from 0 to 3. The quiver of $A$ looks as follows: $$\xymatrix@1{ \circ^{0}\ar [r]^{\alpha} & \circ^{1}\ar [d]^{\beta} & & & & & & \\ \circ^{3} \ar [u]^{\delta} & \circ^{2} \ar[l]^{\gamma}}.$$ We denote the idempotents corresponding to the points $i$ by $e_i$ and the simple modules corresponding to $i$ by $S_i$. By $J$ we denote the Jacobson radical of an algebra. \[lemma 1\] $A$ is a Morita algebra with dominant dimension equal to two. The projective-injective indecomposable $A$-modules are $e_1 A$ and $e_3 A$. Thus the minimal faithful projective-injective $A$-module is $eA$ with $e=e_1 + e_3$ and we have that $eAe$ is the symmetric Nakayama algebra with Kupisch series $[3,3]$. The minimal injective coresolution of $e_0 A$ is as follows: $$0 \rightarrow e_0 A \rightarrow e_3 A \rightarrow e_3 A \rightarrow e_3 A / e_3 J^4 \rightarrow 0. \ \ (*)$$ As $e_3 A$ is projective-injective and $e_3 A / e_3 J^4$ is not projective but injective, this shows that $e_0A$ has dominant dimension equal to two. The minimal injective coresolution of $e_2 A$ looks as follows: $$0 \rightarrow e_2 A \rightarrow e_1 A \rightarrow e_1 A \rightarrow e_1 A / e_1 J^4 \rightarrow 0. \ \ (**)$$ As $e_1 A$ is projective-injective and $e_1 A / e_1 J^4$ is not projective but injective, this shows that $e_2A$ has dominant dimension equal to two. Since the dominant dimension of an algebra is equal to the minimum of the dominant dimensions of the indecomposable projective modules, we conclude that $A$ has dominant dimension equal to two and thus is a Morita algebra. Now let $M:=e_0 A \oplus e_1 A \oplus e_3 A \oplus e_1 A/e_1 J^4$. Recall that a tilting module is a module $T$ over an algebra $\Lambda$ that has finite projective dimension and $Ext_{\Lambda}^i(T,T)=0$ for all $i>0$ such that the regular module $\Lambda$ has a finite coresolution in $add(T)$. $M$ is a tilting module of projective dimension two. Note that $M$ has three indecomposable projective modules as direct summands where only $e_0A$ is not injective and one indecomposable injective non-projective module, namely $e_1 A/e_1 J^4$. The following minimal projective resolution of $e_1 A / e_1 J^4$ shows that the projective dimension of $M$ is equal to two: $$0 \rightarrow e_2 A \rightarrow e_1 A \rightarrow e_1 A \rightarrow e_1 A/ e_1 J^4 \rightarrow 0.$$ Now the exact sequence $(**)$ in the proof of \[lemma 1\] shows that $A$ has a coresolution in $add(M)$. What is left to show is that $Ext_A^i(M,M)=0$ for $i=1$ and $i=2$ because $Ext_A^i(M,M)=0$ for $i>2$ since $M$ has projective dimension 2. Note that $\Omega^1(M) = e_1 J^4 \cong S_1$. We have $Ext_A^1(M,M)=Ext_A^1(e_1 A / e_1 J^4, e_0 A)$ and $Ext_A^2(M,M)=Ext_A^1(\Omega^1(M),M)=Ext_A^1(S_1 , e_0 A)$. Note that in general for a simple module $S$ and a module $N$ over an algebra $\Lambda$, we have $Ext_{\Lambda}^1(S,N)=0$ iff the socle of $I^1(N)$ does not have $S$ as a direct summand when $(I^i(N))$ denotes a minimal injective coresolution of $N$, see for example [@Ben] corollary 2.5.4. This gives us that $Ext_A^1(S_1 , e_0 A)=0$ when looking at the minimal injective coresolution of $e_0 A$ in $(*)$ in the proof of \[lemma 1\]. Now we show that $Ext_A^1(e_1 A / e_1 J^4, e_0 A)=0$. Look at the following short exact sequence: $$0 \rightarrow e_1 J^4 \rightarrow e_1 A \rightarrow e_1 A / e_1 J^4 \rightarrow 0.$$ We apply the functor $Hom_A(-,e_0 A)$ to this short exact sequence and obtain the following exact sequence: $$0 \rightarrow Hom_A(e_1 A / e_1 J^4, e_0 A) \rightarrow Hom_A(e_1 A, e_0 A) \rightarrow Hom_A(S_1, e_0 A) \rightarrow Ext_A^1(e_1 A/ e_1 J^4, e_0 A) \rightarrow 0.$$ This gives us that $Ext_A^1(e_1 A/ e_1 J^4, e_0 A)=0$ iff $dim(Hom_A(e_1 A, e_0 A))=dim(Hom_A(e_1 A / e_1 J^4, e_0 A))+dim(Hom_A(S_1, e_0 A))$, which is true since $dim(Hom_A(e_1 A, e_0 A))=1$ and $dim(Hom_A(e_1 A / e_1 J^4, e_0 A))=1$ but $dim(Hom_A(S_1, e_0 A))=0$. This proves that $Ext_A^i(M,M)=0$ for all $i>0$ and thus that $M$ is a tilting module of projective dimension two. Now let $B:=End_A(M)$ be the endomorphism ring of $M$. $B$ is a Nakayama algebra given by quiver and relations with Kupisch series $[4,4,5,5]$. By the main theorem of [@Yam2], the endomorphism ring of a module over a Nakayama algebra which only has indecomposable projective or injective modules as a direct summands is again a Nakayama algebra. Also note that $B$ is a basic algebra since $M$ is a basic module and $B$ has simple modules isomorphic to $End_A(M_i)/rad(End_A(M_i))$, which are one-dimensional modules when $M_i$ denote the indecomposable direct summands of $M$. A basic algebra with all simple modules of dimension equal to one is given by quiver and relations. We therefore just have to determine the Kupisch series of $B$. We have $$B=End_A(e_0 A \oplus e_1 A \oplus e_3 A \oplus e_1 A/e_1 J^4)=$$ $$\begin{pmatrix}[1] e_0 A e_0 & e_0 A e_1 & e_0 A e_3 & e_0 J e_1 \\ e_1 A e_0 & e_1 A e_1 & e_1 A e_3 & e_1 J e_1 \\ e_3 A e_0 & e_3 A e_1 & e_3 A e_3 & e_3 J e_1 \\ (e_1 A/ e_1 J^4)e_0 & (e_1 A/ e_1 J^4)e_1 & (e_1 A/ e_1 J^4)e_3 & (e_1 A/ e_1 J^4)e_1 \end{pmatrix} .$$ Noting that $e_0 J e_0=0$ and $(e_1 J/ e_1 J^4)e_1=0$, the radical of $B$ is then equal to $$\begin{pmatrix}[1] 0 & e_0 A e_1 & e_0 A e_3 & e_0 J e_1 \\ e_1 A e_0 & e_1 J e_1 & e_1 A e_3 & e_1 J e_1 \\ e_3 A e_0 & e_3 A e_1 & e_3 J e_3 & e_3 J e_1 \\ (e_1 A/ e_1 J^4)e_0 & (e_1 A/ e_1 J^4)e_1 & (e_1 A/ e_1 J^4)e_3 & 0 \end{pmatrix} .$$ We have $rad^2(B)=rad(B) rad(B)$ and multiplication gives that the (1,4)-entry of $rad^2(B)$ is equal to $e_0 A e_1 J e_1 +e_0 A e_3 J e_1=0$. Thus the (1,4)-entry in $rad(B)/rad^2(B)$ is $e_0 J e_1$, which is non-zero. This gives us that there is an arrow in the quiver of $B$ from the first point to the fourth point. Now the projective indecomposable $B$-modules are given by $Hom_A(M,M_i)$. We have $dim(Hom_A(M,e_0A))=4, dim(Hom_A(M,e_1 A))=5 , dim(Hom_A(M,e_3A))=5 $ and $dim(Hom_A(M,e_1A/e_1 J^4))=4$ and we saw that there is an arrow in the quiver of $B$ from a point whose corresponding indecomposable projective module has dimension 4 and a point whose corresponding indecomposable projective module has dimension 4. This gives us that the Kupisch series can only be $[4,4,5,5]$. After renumbering we can assume that he quiver of the Nakayama algebra $B$ with Kupisch series \[4,4,5,5\] looks as follows: $$\xymatrix@1{ \circ^{0}\ar [r]^{a} & \circ^{1}\ar [d]^{b} & & & & & & \\ \circ^{3} \ar [u]^{d} & \circ^{2} \ar[l]^{c}}.$$ $B$ has dominant dimension equal to 4 but is not a Morita algebra. The projective-injective indecomposable $B$-modules are $e_1 B, e_2 B$ and $e_3B$. Thus the minimal faithful projective-injective $B$-module is $eB$ with $e=e_1 +e_2 +e_3$. The algebra $eBe$ is the Nakayama algebra with Kupisch series $[3,4,4]$, which is not selfinjective. What is left to show it that $B$ has dominant dimension equal to 4. We give the minimal injective coresolution of $e_0 B$: $$0 \rightarrow e_0 B \rightarrow e_3 B \rightarrow e_3 B \rightarrow e_2 B \rightarrow e_2 B \rightarrow D(B e_1) \rightarrow 0.$$ This shows that $e_0 B$ has dominant dimension equal to 4 and also that $B$ has dominant dimension equal to 4 since the dominant dimension of the regular module equals the minimum of the dominant dimensions of the indecomposable projective modules. Combining all the results of this section we obtain the following theorem: Let $A$ be the Nakayama algebra with Kupisch series \[4,5,4,5\] with vertices numbered from 0 to 3. Let $M$ be the module $e_0 A \oplus e_1 A \oplus e_3 A \oplus e_1 A/e_1 J^4$. Then $A$ is a Morita algebra and $M$ is a tilting module of projective dimension two such that the algebra $B:=End_A(M)$ is an algebra of dominant dimension equal to 4 that is not a Morita algebra. [Gus]{} Anderson, Frank W.; Fuller, Kent R.: [*Rings and Categories of Modules.*]{}Graduate Texts in Mathematics, Volume 13, Springer-Verlag, (1992). Assem, I.; Simson, D.; Skowronski, A.: [*Elements of the Representation Theory of Associative Algebras, Volume 1: Techniques of Representation Theory.*]{} London Mathematical Society Student Texts, 2007. Auslander, M.; Reiten, I.; Smalo, S.: [*Representation Theory of Artin Algebras.*]{} Cambridge Studies in Advanced Mathematics, Volume 36, Cambridge University Press, 1997. Benson, D.: [*Representations and cohomology I: Basic representation theory of finite groups and associative algebras.*]{} Cambridge Studies in Advanced Mathematics, Volume 30, Cambridge University Press, 1991. Chen, H.; Xi, C.: [*Dominant dimensions, derived equivalences and tilting modules.*]{} Israel Journal of Mathematics, September 2016, Volume 215, Issue 1, pp 349-395. Kerner, O.; Yamgata, K.: [*Morita algebras.*]{} Journal of Algebra Volume 382, 15 May 2013, Pages 185-202. Marczinzik, R.: [*Upper bounds for the dominant dimension of Nakayama and related algebras.*]{} Journal of Algebra Volume 496, 15 February 2018, Pages 216-241. The QPA-team, QPA - Quivers, path algebras and representations - a GAP package, Version 1.25; 2016 (https://folk.ntnu.no/oyvinso/QPA/) Skowronski, A.; Yamagata, K.: [*Frobenius Algebras I: Basic Representation Theory.*]{} EMS Textbooks in Mathematics, (2011). Yamagata, K.: [*Frobenius Algebras*]{} [Hazewinkel, M. (editor): Handbook of Algebra, North-Holland, Amsterdam, Volume I, 841-887, (1996).]{} Yamagata, K.:[*Modules with serial Noetherian endomorphism rings.*]{} Journal of Algebra Volume 127, Issue 2, December 1989, Pages 462-469.

--- abstract: 'Recently, Earl, Vander Meulen, and Van Tuyl characterized some families of Cohen-Macaulay or Buchsbaum circulant graphs discovered by Boros-Gurvich-Milani$\check{\text{c}}$, Brown-Hoshino, and Moussi. In this paper, we will characterize those families of circulant graphs which satisfy Serre’s condition $S_2$. More precisely, we show that for some families of circulant graphs, $S_2$ property is equivalent to well-coveredness or Buchsbaumness, and for some other families it is equivalent to Cohen-Macaulayness. We also give examples of infinite families of circulant graphs which are Buchsbaum but not $S_2$, and vice versa.' address: | Amir Mousivand\ Department of Mathematics, Firoozkooh Branch, Islamic Azad University (IAU), Firoozkooh, Iran. author: - Amir Mousivand title: 'Circulant $S_2$ graphs' --- Introduction ============ Let $G$ be a finite simple undirected graph with the vertex set $V(G)$ and the edge set $E(G)$. A subset $C$ of $V(G)$ is called a [*vertex cover*]{} of $G$ if $C \cap e \ne \emptyset $ for any $e \in E(G)$. A vertex cover $C$ of $G$ is called *minimal* if there is no proper subset of $C$ which is a vertex cover. A graph $G$ is said to be [*well-covered*]{} if all its minimal vertex covers have the same cardinality. A simplicial complex $\Delta$ on the vertex set $V=\{x_1 , \ldots, x_n \}$ is a collection of subsets of $V$, with the properties: (1) $\{x_i\}\in\Delta$ for all $i$, and (2) if $F\in\Delta$, then all subsets of $F$ are also in $\Delta$ (including the empty set). An element $F$ of $\Delta$ is called a [*face*]{} of $\Delta$, and maximal faces of $\Delta$ (with respect to inclusion) are called [*facets*]{} of $\Delta$. We denote the simplicial complex $\Delta$ with facets $F_1 , \ldots , F_t$ by $\Delta = \langle F_1 , \ldots , F_t \rangle$. The [*dimension*]{} of a face $F\in\Delta$ is defined by $\dim F= | F | -1$, and The dimension of $\Delta $ is defined by $\dim \Delta = \max \{ \dim F ~:~ F \in \Delta \}$. A simplicial complex is called [*pure*]{} if all its facets have the same cardinality. $\Delta$ is called [*Cohen-Macaulay*]{} (resp. [*Buchsbaum*]{}) over a field $k$ if its Stanley-Reisner ring $k[\Delta]$ is Cohen-Macaulay (resp. Buchsbaum), and called Cohen-Macaulay (resp. Buchsbaum) if it has the same property over any field $k$. For a face $F$ of $\Delta$, the [*link*]{} of $F$ is the simplicial complex $$\textnormal{link}_\Delta(F)=\{G\in\Delta ~ ~:~ ~ G\cap F=\emptyset ~ \text{and} ~ G\cup F\in \Delta\}.$$ By Reisner’s criterion (see e.g. [@V Theorem 5.3.5]), A simplicial complex $\Delta$ is Cohen-Macaulay over a field $k$, if and only if $\tilde{H}_i(\textnormal{link}_\Delta(F);k)=0$ for all $F\in \Delta$ and $i<\rm{dim}\rm{link}_\Delta(F)$. The [*independence complex*]{} of a graph $G$, denoted by $\text{Ind}(G)$, is the simplicial complex whose faces correspond to independent (or stable) sets of $G$, where a subset $F$ of $V(G)$ is called an independent set if any subsets of $F$ with cardinality two do not belong to $E(G)$. Since the complement of a vertex cover is an independent set, it follows that a graph $G$ is well-covered if and only if $\text{Ind}(G)$ is a pure simplicial complex. A graph $G$ is called Cohen-Macaulay (resp. Buchsbaum) if the independence complex $\text{Ind}(G)$ is Cohen-Macaulay (resp. Buchsbaum). Given an integer $n\geq 1$ and a generating set $S\subseteq \{1,2\ldots,\lfloor\frac{n}{2}\rfloor\}$, the [*circulant graph*]{} $C_n(S)$ is the graph with the vertex set $V=\{0,1,\ldots,n-1\}$ whose edge set is $$E=\{\{i,j\} ~ ~ : ~ ~ |i-j|\in S ~ ~ \text{or} ~ ~ n-|i-j|\in S\}.$$ For $S=\{a_1,\ldots,a_t\}$, we abuse the notation and use $C_n(a_1,\ldots,a_t)$ to denote $C_n(S)$. Circulant graphs belong to the family of cayley graphs and may be considered as a generalization of cycles because $C_n=C_n(1)$. In recent years, there have been a flurry of work identifying circulant graphs which are also well-covered (see e.g. [@BGM; @BH1; @BH2; @H; @M]). Since a well-covered graph has the property that its independence complex is pure, and a pure complex can have some extra combinatorial (e.g. vertex decomposable and shellable) or topological (e.g. Cohen-Macaulay and Buchsbaum) structure, i.e., for a pure simplicial complex the following hierarchy is known: $$\mbox {vertex~decomposability} \Longrightarrow \mbox {shellability} \Longrightarrow \mbox {Cohen-Macaulayness} \Longrightarrow \mbox {Buchsbaum},$$ so it is natural to ask what more structures $\text{Ind}(C_n(S))$ entertains? Recently, Vander Meulen, Van Tuyl, and Watt [@VVW] characterized Cohen-Macaulay (vertex decomposable, shellable, or buchsbaum) circulant graphs of the form $C_n(1,2,\ldots,d)$ and Cohen-Macaulay cubic circulant graphs. Earl, Vander Meulen, and Van Tuyl [@EVV] determined when circulant graphs of the form $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$, $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$, and one-paired circulants have these structures. Also Vander Meulen and Van Tuyl [@VV] investigated when the independence complex of the lexicographical product of two graphs is either vertex decomposable or shellable. They also constructed an infinite family of graphs in which the independence complex of each graph is shellable, but not vertex decomposable. A finitely generated graded module $M$ over a Noetherian graded ring $R$ is said to satisfy the Serre’s condition $S_r$ (or simply say $M$ is an $S_r$ module) if $$\text{depth}(M_p)\geq \text{min}\{r, \text{dim}(M_p)\},$$ for all $p\in \text{Spec}(R)$. Since $M$ is Cohen-Macaulay if $\text{depth}(M_p)=\text{dim}(M_p)$ for every $p\in \text{Spec}(R)$, it follows that $M$ is Cohen-Macaulay if and only if it satisfies the Serre’s condition $S_r$ for all $r\geq 1$. A simplicial complex $\Delta$ is said to satisfy Serre’s condition $S_r$ over a field $k$ (or simply say $\Delta$ is an $S_r$ complex) if the Stanley-Reisner ring $k[\Delta]$ satisfies Serre’s condition $S_r$. Terai [@T] presented the following analogue of Reisner’s criterion for $S_r$ simplicial complexes.\ \ [**Theorem.**]{} [*A simplicial complex $\Delta$ satisfies Serre’s condition $S_r$ over a field $k$ if and only if for every face $F\in\Delta$ (including the empty face), $\tilde{H}_i(\textnormal{link}_\Delta(F);k)=0$ for all $i<\rm{min} \{{\it r}-1,\rm{dimlink}_\Delta(F)\}$.*]{}\ There are some basic facts related to $S_r$ simplicial complex. Every simplicial complex satisfies $S_1$. On the other hand, for $r\geq 2$, simplicial complexes satisfying $S_r$ (over a field k), are pure ([@MT Lemma 2.6]) and strongly connected ([@Ha Corollary 2.4]). Recall that a pure simplicial complex $\Delta$ is called strongly connected if, for every pair of facets $F$ and $F'$ of $\Delta$, there exists a sequence of facets $F=F_0,F_1,\ldots,F_t =F'$ such that $|F_k\cap F_{k+1}|=|F_k|-1$ for $k=0,1,\ldots,t-1$. We refer the reader to [@GPSY; @HTYZ; @MT; @PSTY; @T; @TY] for more details on the properties of $S_r$ (and sequentially $S_r$) simplicial complexes. Also as an immediate consequence of previous Theorem, we have the following corollary on $S_2$ simplicial complexes.\ \ [**Corollary.**]{} [*A simplicial complex $\Delta$ satisfies Serre’s condition $S_2$ over a field $k$ if and only if $\textnormal{link}_\Delta(F)$ is connected for every face $F\in\Delta$ with $\rm{dim}\rm{link}_\Delta(F)\geq 1$. In particular, $S_2$ property of a simplicial complex does not depend on the characteristic of the field $k$.*]{}\ Although Cohen-Macaulay (or Buchsbaum) property of stanley-Reisner rings depend on the base field and hence it is a topological property, $S_2$ property does not depend on the base field and $S_2$ Stanley-Reisner rings can be characterized combinatorially. We say a graph $G$ satisfies the Serre’s condition $S_n$, or simply is an $S_n$ graph, if its independence complex $\text{Ind}(G)$ satisfies this condition. As an immediate consequence, it follows that $S_n$ graphs are well-cowered. Haghighi, Yassemi, and Zaare-Nahandi [@HYZ] showed that $S_2$ property for bipartite graphs and chordal graphs is equivalent to Cohen-Macaulayness. Recall that a graph $G$ is bipartite if there exists a partition $V(G)=V \bigcup {V'}$ with $V \bigcap {V'} = \emptyset$ such that each edge of $G$ is of the form $\{i,j\}$ with $i \in V$ and $j \in {V'}$, and $G$ is called chordal if every cycle of length at least four has a chord, where a chord is an edge joining two nonadjacent vertices of the cycle. In this paper we characterize some families of circulant $S_2$ graphs. In section 2, we consider $S_2$ property of powers of cycles. More precisely, we show that if $n\geq 2d\geq 2$, then $C_n(1,2,\ldots,d)$ is $S_2$ if and only if $n\leq 3d+2$ and $n\neq 2d+2$, or $n=4d+3$. Comparing this with [@VVW Theorem 3.7] implies that $C_n(1,2,\ldots,d)$ is Buchsbaum but not $S_2$ if and only if $n=2d+2$. In section 3, we consider circulants of the form $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$. We show that such a graph satisfies serre’s condition $S_2$ if and only if it is well-covered, i.e., $n>3d$ or $n= 2d + 2$. This is equivalent to say that $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is Buchsbaum. In this case, we will show that the only non-shellable (connected) link in $\Delta=\text{Ind}(C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor))$ is the link of $\emptyset$ which is $\Delta$ (except for $d=1$), and the link of all non-empty faces of $\Delta$ are shellable. In section 4, we investigate circulants of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$. We show that cirulant $S_2$ graphs in this family are all Cohen-Macaulay. Using [@EVV Theorem 4.2] it is equivalent to say that $\text{gcd}(i,n)=1$. Since circulants in this family are all Buchsbaum (see [@EVV Theorem 4.2]), it follows that circulant graphs of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$ for which $\text{gcd}(i,n)> 1$, is an infinite family of Buchsbaum graphs that are not $S_2$. Section 5 is dedicated to one-paired circulant graph, where we show that one-paired circulant graph $C(n;a,b)$ is $S_2$ if and only if $n=ab$, i.e., it is Cohen-Macaulay. Combining this together with [@EVV Corollary 5.10] implies that, if $m>1$, then $C(mb;1,b)$ is Buchsbaum but not $S_2$. Finally, in section 6 we identify which circulant cubic graphs are $S_2$. More precisely, we first show that the only non-$S_2$ connected circulant cubic graph is $C_6(1,3)$. Combining this together with a result of Davis and Domke [@DD] enables us to prove that the circulant cubic graph $C_{2n}(a,n)$ is $S_2$ if and only if $\frac{2n}{t}=3,4,5,8$, where $t={\rm gcd}(a,2n)$. As the final result of this paper, we present infinite families of circulant graphs which are $S_2$ but not Buchsbaum, namely, circulants of the form $C_{8t}(t,4t)$, $C_{10t}(2t,5t)$, and $C_{10t}(4t,5t)$, where $t>1$. Circulants of the form $C_n(1,2,\ldots,d)$ ========================================== In this section we identify which circulants of the form $C_n(1,2,\ldots,d)$ satisfy serre’s condition $S_2$. We need the following result of Brown and Hoshino. \[THM2\] [([@BH2 Theorem 4.1])]{} Let $n$ and $d$ be integers with $n\geq 2d\geq 2$. Then $C_n(1,2,\ldots,d)$ is well-covered if and only if $n\leq 3d+2$ or $n=4d+3$. Using the above result we get the next characterization of circulant $S_2$ graphs of the form $C_n(1,2,\ldots,d)$. \[THM3\] Let $n$ and $d$ be integers with $n\geq 2d\geq 2$. Then $G=C_n(1,2,\ldots,d)$ is $S_2$ if and only if $n\leq 3d+2$ and $n\neq 2d+2$, or $n=4d+3$. If $G$ is $S_2$, then $G$ is well-covered and hence by Theorem \[THM2\] we get $n\leq 3d+2$ or $n=4d+3$. In the case where $n=2d+2$, $\text{Ind}(G)$ comprises of $d+1$ disjoint $1$-faces (edges) and hence $G$ is not $S_2$. Now we prove the converse. The idea is inspired by the proof of [@VVW Theorems 3.4 and 3.5]. By [@BH2 Theorem 3.1] one has $\text{dimInd}(G)=\lfloor \frac{n}{d+1}\rfloor-1$. If $n=2d$ or $n=2d+1$, then $\text{dimInd}(G)=0$ and $G$ is Cohen-Macaulay. If $2d+3\leq n\leq 3d+2$, then $\text{dimInd}(G)=1$. In this case for $0\leq i<j\leq n-1$, there exists the path $i,i+d+2,i+1,i+d+3,i+2,\ldots,j$ of $1$-faces (facets or edges) of $\text{Ind}(G)$. In particular, $\text{Ind}(G)$ is connected, and again it is Cohen-Macaulay. Now assume $n=4d+3$. One has $\text{dimInd}(G)=2$. First note that $$\textit{\bf{0}},d+1,2d+2,3d+3,\textit{\bf{1}},d+2,2d+3,3d+4,\textit{\bf{2}},\ldots,\textbf{\textit{d}-1},2d,3d+1,4d+2,\textbf{\textit{d}},2d+1,3d+2,\textit{\bf{0}}$$ is a path (cycle) of $1$-faces of $\text{Ind}(G)$, i.e., $\text{Ind}(G)$ is connected. To complete the proof, by symmetry, it suffices to show that $\rm{link}_{\rm{Ind}(G)}(0)$ is connected. One can directly check that\ $\{0,d+1,2d+2\}, \hspace{3mm} \{0,d+1,2d+3\}, \hspace{3mm} \ldots, \hspace{3mm} \{0,d+1,3d+1\}, \hspace{3mm} \{0,d+1,3d+2\},$ $\{0,d+2,2d+3\}, \hspace{2mm} \{0,d+2,2d+4\}, \hspace{3mm} \ldots, \hspace{2mm} \{0,d+2,3d+2\},$ $\hspace{9mm} \vdots \hspace{28mm} \vdots$ $\{0,2d,3d+1\}, \hspace{3mm} \{0,2d,3d+2\},$ $\{0,2d+1,3d+2\}$\ is a complete list of facets of $\text{Ind}(G)$. From this description quickly follows that $\rm{link}_{\rm{Ind}(G)}(0)$ is connected, and $G$ is $S_2$. As an immediate consequence, we get the following corollary which is the content of [@HYZ Proposition 1.6]. The cyclic graph $C_n$ of length $n\geq 3$ is $S_2$ if and only if $n = 3,5,$ or $7$. In particular, $C_7$ is the only cyclic graph which is $S_2$ but not Cohen-Macaulay. Vander Meulen [*et al.*]{} in [@VVW Theorems 3.4 and 3.5] showed that for $n\geq 2d\geq 2$ one has - $C_n(1,2,\ldots,d)$ is vertex decomposable/shellable/Cohen-Macaulay if and only if $n\leq 3d+2$ and $n\neq 2d+2$. - $C_n(1,2,\ldots,d)$ is Buchsbaum but not Cohen-Macaulay if and only if $n=2d+2$ or $n=4d+3$. Comparing these with Theorem \[THM3\] yields that $C_n(1,2,\ldots,d)$ is Buchsbaum but not $S_2$ if and only if $n=2d+2$. Circulants of the form $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ ============================================================================ In this section we investigate $S_2$ circulant graphs of the form $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$. To do this, we need the following result of Brown and Hoshino on well-coveredness of these circulant graphs. \[THM1\] [([@BH2 Theorem 4.2])]{} Let $n $ and $d$ be integers with $n\geq 2d+2$ and $d\geq1$. Then $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is well-covered if and only if $n>3d$ or $n= 2d + 2$. [A simplicial complex $\Delta$ is called [*shellable*]{} if there is a linear order $F_1 , \ldots, F_s$ of all the facets of $\Delta $ such that for all $1\leq i<j\leq s$, there exists some $v\in F_j\setminus F_i$ and some $l\in \{1 , \ldots, j-1\}$ with $F_j\setminus F_l=\{v\}$. $F_1 , \ldots, F_s$ is called a shelling order of $\Delta$.]{} We will make use the next two Lemmas to prove Theorem \[MAIN1\] which is the main result of this section. \[LEM1\] Let $\Delta$ be a $d$-dimensional pure simplicial complex on the vertex set $V=\{x_1,x_2,\ldots,x_n\}$ whose facets are given by $$F_i=\{x_i,x_{i+1},\ldots,x_{i+d}\}\hspace{3mm}:\hspace{3mm}i=1,2,\ldots,n-d.$$ Then $\Delta$ is pure shellable. It is easy to see that $F_1,F_2,\dots,F_{n-d}$ is the desired shelling order on the facets of $\Delta$. \[LEM2\] Let $\Delta$ be a shellable simplicial complex and $F_1,F_2,\dots,F_s$ a shelling order of $\Delta$. Also let $F\in\Delta$ be such that $F\subseteq F_i$ for all $i=1,2,\ldots,s$. Then $\rm{link}_\Delta(F)$ is shellable. It is clear that $\text{link}_\Delta(F)=\langle F_1\setminus F,F_2\setminus F,\dots,F_s\setminus F\rangle$, and that this order is the desired shelling order of the facets of $\text{link}_\Delta(F)$. We also need the following result on circulants of the form $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$. \[THM4\] [([@EVV Theorem 3.3])]{} Let $n $ and $d$ be integers with $n\geq 2d+2$ and $d\geq1$. The following are equivalent: - $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is Buchsbaum; - $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is well-covered; - $n>3d$ or $n= 2d + 2$. Furthermore, $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is vertex decomposable/shellable/Cohen-Macaulay if and only if $n=2d+2$ and $d\geq 1$ or $d=1$ and $n\geq 3$. Now we are ready to state and prove the main result of this section. \[MAIN1\] Let $n $ and $d$ be integers with $n\geq 2d+2$ and $d\geq1$. Then $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is $S_2$ if and only if $n>3d$ or $n= 2d + 2$. Let $G=C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ with $n\geq 2d+2$ and $d\geq1$. If $G$ is $S_2$, then $G$ is well-covered and the result follows from Theorem \[THM1\]. Now we prove the converse. If $n= 2d + 2$, then $G=C_{2d+2}(d + 1)$ is the union of $(d+1)$ disjoint edges. It follows that $G$ is Cohen-Macaulay, and hence it is $S_2$. So assume $n>3d$ and $d\geq 1$, and let $\Delta=\text{Ind}(G)$. By the proof of [@EVV Theorem 3.3] one has $\text{Ind}(G)=\langle F_0,\ldots,F_{n-1}\rangle$, where $F_i=\{i,i+1,\ldots,i+d\}$ for all $i=0,\ldots,n-1$, and the indices computed modulo $n$. If $\text{dim}\text{link}_\Delta(F)=d$, then $\text{link}_\Delta(F)=\Delta$, which is a connected simplicial complex that is not shellable except for $d=1$ (see Theorem \[THM4\]) and there is nothing to prove. So it is enough to show that $\text{link}_\Delta(F)$ is connected for each face $F\in\Delta$ with $0<\text{dim}\text{link}_\Delta(F)<d$. We show that $\text{link}_\Delta(F)$ is indeed shellable.\ If $d=1$, then the only case where $\text{dim}\text{link}_\Delta(F)>0$ is for $F=\emptyset$ which in this case $\text{link}_\Delta(F)=\Delta$ is shellable. So suppose $d>1$. Assume that $\text{dim}\text{link}_\Delta(F)=d-t$ where $0<t<d$. It follows that $|F|=t$. Without loos of generality we may assume $F\subseteq F_0$. Suppose $F=\{j_1,j_2,\ldots,j_t\}$ with $0\leq j_1<j_2<\ldots<j_t\leq d$. We claim that $$F_{n-d+j_t},F_{n-d+j_t+1},\ldots,F_{j_1-1},F_{j_1}$$ is a complete list of the facets of $\text{Ind}(G)$ that contain $F$, where the indices computed modulo $n$. To see this, It suffices to notice that one has $F_{n-d+j_t}=\{n-d+j_t,n-d+j_t+1,\ldots,j_t-1,j_t\},$ $F_{n-d+j_t+1}=\{n-d+j_t+1,\ldots,j_t,j_t+1\},$ $~~~\vdots$ $F_{j_1-1}=\{j_1-1,j_1,\ldots,j_1+d-1\},$ $F_{j_1}=\{j_1,j_1+1,\ldots,j_1+d\}.$\ It follows from Lemma \[LEM1\] that $\Delta'=\langle F_{n-d+j_t},F_{n-d+j_t+1},\ldots,F_{j_1-1},F_{j_1}\rangle$ is a pure shellable simplicial complex on the vertex set $V'=\{n-d+j_t,n-d+j_t+1,\ldots,j_1+d\}$ (Note that $|V'|=2d-(j_t-j_1)+1\leq 2d+1<n$, and that $n-d+j_t,n-d+j_t+1,\ldots,j_1+d$ are all distinct since $j_1+d\leq 2d$ and $n-d+j_t>2d+j_t\geq 2d$). Set $$F'_0=F_{n-d+j_t}\setminus F \hspace{3mm},\hspace{3mm} F'_1=F_{n-d+j_t+1}\setminus F \hspace{3mm},\hspace{3mm}\ldots \hspace{3mm},\hspace{3mm}F'_{d-(j_t-j_1)+1}=F_{j_1}\setminus F.$$ One can easily check that $$\text{link}_\Delta(F)=\langle F'_0,F'_1,\ldots,F'_{d-(j_t-j_1)+1}\rangle.$$ Now it follows from Lemma \[LEM2\] that $\text{link}_\Delta(F)$ is shellable, as desired. [Let $G=C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$, where $n>3d$ and $d>1$. Proof of Theorem \[MAIN1\] shows that if $\Delta =\rm{Ind}(G)$, then the only non-shellable (connected) link in $\Delta$ is the link of $\emptyset$ which is $\Delta$, and the link of all non-empty faces of $\Delta$ are shellable.]{} Let $n $ and $d$ be integers with $n\geq 2d+2$ and $d\geq1$. For $G=C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ and $\Delta=\rm{Ind}(G)$ the following are equivalent: - $G$ is $S_2$; - $G$ is Buchsbaum; - $G$ is well-covered; - $n>3d$ or $n= 2d + 2$; - $\Delta$ is strongly connected and $\rm{link}_\Delta(F)$ is (pure) shellable for all $F\in\Delta$ with $\rm{dim}\rm{link}_\Delta(F)<d$; - $\rm{link}_\Delta(F)$ is strongly connected for all $F\in\Delta$ with $0<\rm{dim}\rm{link}_\Delta(F)\leq d$. The equivalent of (ii), (iii), and (iv) is Theorem \[THM4\]. Also (i) and (iv) are equivalent by Theorem \[MAIN1\]. On the other hand, (v) and (vi) follow from the description of facets of $\Delta$ and $\rm{link}_\Delta(F)$ in the proof of Theorem \[MAIN1\]. Finally, if (v) or (vi) hold, then $\Delta$ is pure and $\rm{link}_\Delta(F)$ is connected for all $F\in\Delta$ with $\rm{dim}\rm{link}_\Delta(F)>0$, i.e., $\Delta$ is $S_2$. Circulants of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$ =============================================================================== Moussi [@M Theorem 6.4] proved that Circulants of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$ are well-covered. Earl, Vander Meulen, and Van Tuyl [@EVV Section 4] showed that these circulants are always Buchsbaum, and they are Cohen-Macaulay if and only if $\text{gcd}(i,n)=1$. In this section we show that this condition is equivalent to $S_2$ property. Let $G=C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$. The following are equivalent: - $G$ is $S_2$; - $G$ is Cohen-Macaulay; - ${\rm gcd}(i,n)=1$. The equivalent of (ii) and (iii) is [@EVV Theorem 4.2]. Also (ii) $\Longrightarrow (i)$ always hold. Now we show that (i) $\Longrightarrow (ii)$. By [@M Theorem 6.4] one has $\text{dimInd}(G)=1$ except if $i=\frac{n}{3}$, in which case, $\text{dimInd}(G)=2$. If $i\neq\frac{n}{3}$, then by the proof of [@EVV Theorem 4.2] $\text{Ind}(G)$ is connected and hence it is Cohen-Macaulay. Now assume $i=\frac{n}{3}$. Again by the proof of [@EVV Theorem 4.2] we have $$\text{Ind}(G)=\langle\{0,i,2i\},\{1,i+1,2i+1\},\ldots,\{i-1,2i-1,3i-1\}\rangle.$$ Since $G$ is $S_2$, $\text{Ind}(G)$ is connected. This yields that $i=1, n=3$, and $\text{Ind}(G)=\langle\{0,i,2i\}\rangle$, i.e., $G$ is Cohen-Macaulay. [While $S_2$ property for the family $C_n(d + 1, d + 2,\ldots ,\lfloor\frac{n}{2}\rfloor)$ is equivalent to Buchsbaumness (and hence there exist circulants in this family that are $S_2$ but not Cohen-Macaulay (see Theorem \[THM4\])), but $S_2$ graphs of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$ are those which are Cohen-Macaulay.]{} [It follows form [@EVV Theorem 4.2] that circulant graphs of the form $C_n(1,\ldots,\hat{i},\ldots,\lfloor\frac{n}{2}\rfloor)$ for which ${\rm gcd}(i,n)\neq 1$, is an infinite family of Buchsbaum graphs that are not $S_2$.]{} One-Paired Circulants ===================== In this section we consider $S_2$ property of one-paired circulant graphs introduced by Boros, Gurvich, and Milani$\check{\text{c}}$ [@BGM] as a subfamily of CIS circulant graphs. one-paired circulant graphs define as follows. [Let $(a,b)$ a pair of positive integers such that $ab|n$ and let $S=\{d\in\{1,\ldots,\lfloor\frac{n}{2}\rfloor\}~ ~ : ~ ~ a|d ~ \text{and} ~ ab\nmid d\}$ . The circulant graph $C_n(S)$ is called [*one-paired*]{} and will be denoted by $C(n;a,b)$]{} Earl, Vander Meulen, and Van Tuyl characterized the structure of one-paired circulant graphs, and using that, they identified when a one-paired circulant graph is Cohen-Macaulay or Buchsbaum. More precisely, they showed the followings. \[THM\] [([@EVV Theorem 5.4])]{} Let $G=C(n;a,b)$ be a one-paired circulant. Then $$G=\bigcup_{i=1}^a\left(\bigvee_{j=1}^b\overline{K_{\frac{n}{ab}}}\right).$$ \[COR\] [([@EVV Corollary 5.10])]{} Let $G$ be the one-paired circulant graph $G=C(n;a,b)$. Then - $G$ is vertex decomposable/shellable/Cohen-Macaulay if and only if $n=ab$. - $G$ is Buchsbaum but not Cohen-Macaulay if and only if $a=1$ and $ab<n$. - $\text{Ind}(G)$ is pure but not Buchsbaum if and only if $1<a$ and $ab<n$. Using the above results we show that one-paired circulant $S_2$ graphs are Cohen-Macaulay. Let $G$ be the one-paired circulant graph $G=C(n;a,b)$. Then $G$ is $S_2$ if and only if $n=ab$, i.e., $G$ is Cohen-Macaulay. By Theorem \[THM\] one has $$G=\bigcup_{i=1}^a\left(\bigvee_{j=1}^b\overline{K_{\frac{n}{ab}}}\right).$$ It concludes that $$\text{Ind}(G)=\text{Ind}(G_1)\ast\text{Ind}(G_2)\ast\cdots\ast\text{Ind}(G_a),$$ where $\ast$ denotes the join of complexes, and $G_i=\bigvee_{j=1}^b\overline{K_{\frac{n}{ab}}}$ for all $i=1,\ldots,a$. Therefore each $\text{Ind}(G_i)$ is a pure complex consists of disjoint union of $b$ simplices with cardinality $k=\frac{n}{ab}$. Thus we may assume $$\text{Ind}(G)=\langle F_{11},\ldots,F_{1b}\rangle\ast\langle F_{21},\ldots,F_{2b}\rangle\ast\cdots\ast\langle F_{a1},\ldots,F_{ab}\rangle,$$ where $|F_{ij}|=k$ and $F_{ij}\cap F_{kl}=\emptyset$ for all $1\leq i,k\leq a$ and $1\leq j,l\leq b$.\ If $k>1$, then $F=F_{11}\cup F_{21}\cup\ldots\cup F_{(a-1)1}$ is a face of $\text{Ind}(G)$ whose link is $\rm{link}_\Delta(F)=\text{Ind}(G_a)=\langle F_{a1},\ldots,F_{ab}\rangle.$ This yields that $\rm{dim}\rm{link}_\Delta(F)=k-1>0$ and $\rm{link}_\Delta(F)$ is disconnected since $b>1$. Thus, if $G$ is $S_2$, then $k=1$, i.e., $n=ab$. Corollary \[COR\].(i) now completes the proof. As an immediate consequence, we get the following which is the $S_2$ analogue of [@EVV Theorem 5.9]. \[COR1\] Let $G=C(n;1,b)=C(mb;1,b)$ be a one-paired circulant graph. Then $G$ is $S_2$ if and only if $m=1$, i.e., $G$ is Cohen-Macaulay. [It follows from Corollaries \[COR1\] and \[COR\].(ii) that, if $m>1$, then $C(mb;1,b)$ is Buchsbaum but not $S_2$.]{} circulant cubic graphs ====================== A graph in which every vertex has degree $3$, is called a cubic graph. It is easy to see that a circulant cubic graph is of the form $C_{2n}(a,n)$ with $1\leq a <n$. Brown and Hoshino [@BH2 Theorem 4.3] characterized which connected circulant cubic graphs are well-covered. They showed that a connected circulant cubic graph $G$ is well-covered if and only if it is isomorphic to one of the following graphs: $C_4(1,2)$, $C_6(1,3)$, $C_6(2,3)$, $C_8(1,4)$, or $C_{10}(2,5)$ (see, [@BH2 Theorem 4.3]). On the other hand, Vander Meulen [*et al.*]{} in [@VVW Theorem 5.2] proved that a connected circulant cubic graph $G$ is Cohen-Macaulay if and only if it is isomorphic to $C_4(1,2)$ or $C_6(2,3)$. In the next Proposition we examine which connected circulant cubic graphs are $S_2$. \[PROP1\] The only non-$S_2$ connected circulant cubic graph is $C_6(1,3)$. We know that $C_4(1,2)$ and $C_6(2,3)$ are Cohen-Macaulay. If $G=C_8(1,4)$, then $\text{Ind}(G)=\langle 025,035,036,136,147,247,257\rangle$. Thus $\text{Ind}(G)$ and $\text{link}_{\text{Ind}(G)}(0)$ are connected, i.e., $G$ is $S_2$. If $G=C_{10}(2,5)$, then $$\text{Ind}(G)=\langle 0147,0347,0367,0369,1258,1458,1478,2369,2569,2589\rangle.$$ Again $\text{Ind}(G)$ and $\text{link}_{\text{Ind}(G)}(0)$ are connected. One can easily check that $\text{link}_{\text{Ind}(G)}(F)$ is connected for each $F\in\rm{Ind}(G)$ with $|F|=1$. Therefore $C_{10}(2,5)$ is also $S_2$. Finally, note that the independence complex of $C_6(1,3)$ is a disconnected two dimensional simplicial complex, and so it is not $S_2$. [Earl, Vander Meulen, and Van Tuyl showed that all connected circulant cubic graphs are Buchsbaum (see [@EVV Table 1]). By Proposition \[PROP1\], $C_6(1,3)$ is the only connected circulant cubic Buchsbaum graph which is not $S_2$.]{} We need some preliminaries to generalize Proposition \[PROP1\] to all circulant graphs. \[PROP2\] Let $\Delta_1$ and $\Delta_2$ be simplicial complexes with disjoint vertex sets. Then $\Delta=\Delta_1\ast\Delta_2$ is $S_2$ if and only if $\Delta_1$ and $\Delta_2$ are $S_2$. First note that if $\Delta_1$ and $\Delta_2$ are nonempty simplicial complexes, then $\Delta_1\ast\Delta_2$ is connected. On the other hand, it is not difficult to check that for $F_1\in\Delta_1$ and $F_2\in\Delta_2$ one has $\text{link}_\Delta(F_1\cup F_2)=\text{link}_{\Delta_1}(F_1)\ast\text{link}_{\Delta_2}(F_2)$.\ Now assume $\Delta$ is $S_2$, and $F_1\in\Delta_1$ be such that $\text{dimlink}_{\Delta_1}(F_1)>0$. Then for a facet $F_2\in\Delta_2$ one has $F=F_1\cup F_2\in \Delta$ and $\text{link}_\Delta(F)=\text{link}_{\Delta_1}(F_1)\ast\text{link}_{\Delta_2}(F_2)=\text{link}_{\Delta_1}(F_1)\ast\{\emptyset\}=\text{link}_{\Delta_1}(F_1)$. Since $\Delta$ is $S_2$, $\text{link}_\Delta(F)$ is connected, and so is $\text{link}_{\Delta_1}(F_1)$, i.e., $\Delta_1$ is $S_2$. Similarly $\Delta_2$ is $S_2$.\ Conversely, assume $\Delta_1$ and $\Delta_2$ are $S_2$. Also assume $F\in\Delta$ be such that $\text{dimlink}_{\Delta}(F)>0$. If $F\in\Delta_1$ (similarly for $F\in\Delta_2$), then $\text{link}_\Delta(F)=\text{link}_{\Delta_1}(F)\ast\text{link}_{\Delta_2}(\emptyset)=\text{link}_{\Delta_1}(F)\ast\Delta_2$. Now if $\text{link}_{\Delta_1}(F)\neq\emptyset$, then $\text{link}_{\Delta_1}(F)\ast\Delta_2$ is always connected, and if $\text{link}_{\Delta_1}(F)=\emptyset$, then $\text{link}_{\Delta}(F)=\Delta_2$ which is connected because it is $S_2$ (with positive dimension) and we are done. Now assume $F=F_1\cup F_2$ where $F_1\in\Delta_1$ and $F_2\in\Delta_2$, and that $F_1\neq\emptyset$ and $F_2\neq\emptyset$. Thus $\text{link}_\Delta(F)=\text{link}_{\Delta_1}(F_1)\ast\text{link}_{\Delta_2}(F_2)$. If $\text{link}_{\Delta_1}(F_1)$ and $\text{link}_{\Delta_2}(F_2)$ are nonempty, then $\text{link}_{\Delta}(F)$ is connected. If $\text{link}_{\Delta_1}(F_1)=\emptyset$, then $\text{link}_\Delta(F)=\text{link}_{\Delta_2}(F_2)$ which is connected because $\Delta_2$ is $S_2$. \[COR2\] Let $G$ be a simple graph which is a disjoint union of two graphs $H$ and $K$. Then $G$ is $S_2$ if and only if $H$ and $K$ are $S_2$. It is enough to notice that $\text{Ind}(G)=\text{Ind}(H)\ast\text{Ind}(K)$, and apply Lemma \[PROP2\]. We also need the following result of Davis and Domke [@DD] to extend Proposition \[PROP1\] to all circulant cubic graphs. \[THM5\] Let $G=C_{2n}(a,n)$ with $1\leq a<n$, and let $t={\rm gcd}(a,2n)$. - If $\frac{2n}{t}$ is even, then $G$ is isomorphic to $t$ copies of $C_{\frac{2n}{t}}(1,\frac{n}{t})$. - If $\frac{2n}{t}$ is odd, then $G$ is isomorphic to $\frac{t}{2}$ copies of $C_{\frac{4n}{t}}(2,\frac{2n}{t})$. Now we bring the main result of this section. Let $G=C_{2n}(a,n)$ be a circulant cubic graph and let $t={\rm gcd}(a,2n)$. Then $G$ is $S_2$ if and only if $\frac{2n}{t}=3,4,5,8$. First assume $\frac{2n}{t}$ is even. By Theorem \[THM5\] and Corollary \[COR2\], $G$ is $S_2$ if and only if $\frac{2n}{t}=4,8$. Now assume $\frac{2n}{t}$ is odd. Again by Theorem \[THM5\] and Corollary \[COR2\], $G$ is $S_2$ if and only if $\frac{4n}{t}=6,10$, or equivalently, $\frac{2n}{t}=3,5$. [$G=C_{16}(2,8)$ is an example of a circulant graph which is $S_2$ but not Buchsbaum. Indeed, $C_{16}(2,8)$ is isomorphic to $2$ copies of $C_{8}(1,4)$, and hence it is $S_2$, but as it is shown in [@VVW Table 1], it is not Buchsbaum.]{} As the final result of this paper, we present infinite families of circulant graphs which are $S_2$ but not Buchsbaum. To do this, we need the following lemma. \[LEM3\] [([@EVV Lemma 2.5])]{} Let $G$ and $H$ be two disjoint graphs that are both Buchsbaum, but not Cohen-Macaulay. Then $G\cup H$ is not Buchsbaum. For $t>1$, the followings are infinite families of circulant graphs which are $S_2$ but not Buchsbaum: - $C_{8t}(t,4t)$. - $C_{10t}(2t,5t)$ and $C_{10t}(4t,5t)$. \(i) By Theorem \[THM5\].(i), $C_{8t}(t,4t)$ is isomorphic to $t>1$ copies of $C_{8}(1,4)$, and hence it is $S_2$. On the other hand, $C_{8}(1,4)$ is Buchsbaum but not Cohen-Macaulay. Thus Lemma \[LEM3\] implies that $C_{8t}(t,4t)$ is not Buchsbaum.\ (ii) By Theorem \[THM5\].(ii), $C_{10t}(2t,5t)$ and $C_{10t}(4t,5t)$ are isomorphic to $t>1$ copies of $C_{10}(2,5)$, and hence they are $S_2$. Again, $C_{10}(2,5)$ is Buchsbaum but not Cohen-Macaulay. Now Lemma \[LEM3\] yields that $C_{8t}(t,4t)$ is not Buchsbaum. [10]{} E. Boros, V. Gurvich, and M. Milani$\check{\text{c}}$, On CIS circulants, Discrete Math. 318 (2014), 78–95. J. Brown and R. Hoshino, Independence polynomials of circulants with an application to music, Discrete Math. 309 (2009), 2292–2304. J. Brown and R. Hoshino, Well-covered circulant graphs, Discrete Math. 311 (2011), 244–251. G. Davis and G. Domke, $3$-Circulant graphs, J. Combin. Math. Comput. 40 (2002), 133–142. J. Earl, K. N. Vander Meulen, and A. Van Tuyl, Independence complexes of well-covered circulant graphs, preprint, arXiv:1505.02837 \[math.CO\]. A. Goodarzi, M. R. Pournaki, S. A. Seyed Fakhari, and S. Yassemi, On the $h$-vector of a simplicial complex with Serre’s condition, J. Pure Appl. Algebra 216 (2012), no. 1, 91–94. H. Haghighi, N. Terai, S. Yassemi, and R. Zaare-Nahandi, Sequentially $S_r$ simplicial complexes and sequentially $S_2$ graphs, Proc. Amer. Math. Soc. 139 (2011), no. 6, 1993–2005. H. Haghighi, S. Yassemi, and R. Zaare-Nahandi, Bipartite $S_2$ graphs are Cohen–Macaulay, Bull. Math. Soc. Sci. Math. Roumanie (N.S.) 53(101) (2010), no. 2, 125–132. R. Hartshorne, Complete intersections and connectedness, Amer. J. Math. 84 (1962), 497–508. R. Hoshino, Independence polynomials of circulant graphs, Ph.D. Thesis, Dalhousie University, 2007. R. Moussi, A characterization of certain families of well-covered circulant graphs, M.Sc. Thesis, St. Mary’s University, 2012. S. Murai and N. Terai, $h$-Vectors of simplicial complexes with Serre’s conditions, Math. Res. Lett. 16 (2009), no. 6, 1015–1028. M.R. Pournaki, S.A. Seyed Fakhari, N. Terai, and S. Yassemi, Simplicial complexes satisfying Serre’s condition: A survey with some new results, J. Commut. Algebra 6 (2014), No. 4, 455-483. N. Terai, Alexander duality in Stanley–Reisner rings, Affine Algebraic Geometry, Osaka Univ. Press, Osaka, 2007, 449–462. N. Terai and K. I. Yoshida, A note on Cohen–Macaulayness of Stanley–Reisner rings with Serre’s condition ($S_2$), Comm. Algebra 36 (2008), no. 2, 464–477. K. N. Vander Meulen and A. Van Tuyl, Shellability, vertex decomposability, and lexicographical products of graphs, preprint, arXiv:1505.02838 \[math.CO\]. K. N. Vander Meulen, A. Van Tuyl, and C. Watt, Cohen-Macaulay Circulant Graphs, Comm. Algebra 42 (2014), 1896–1910. R.H. Villarreal, [*Monomial algebras*]{}, Monographs and Textbooks in Pure and Applied Mathematics, 238. Marcel Dekker, Inc., New York, 2001.

--- abstract: 'Employing the path integral approach, we calculate the semiclassical equilibrium density matrix of a particle moving in a nonlinear potential field for coordinates near the top of a potential barrier. As the temperature is decreased, near a critical temperature $T_c$ the harmonic approximation for the fluctuation path integral fails. This is due to a caustic arising at a bifurcation point of the classical paths. We provide a selfconsistent scheme to treat the large quantum fluctuations leading to a nonlinear fluctuation potential. The procedure differs from methods used near caustics of the real time propagator. The semiclassical density matrix is determined explicitly for the case of asymmetric barriers from high temperatures down to temperatures somewhat below $T_c$.' author: - Franz Josef Weiper - Joachim Ankerhold - Hermann Grabert title: | Semiclassical Density Matrix Near the Top\ of a Potential Barrier --- , and

--- abstract: 'We analyze the present-day structure and assembly history of a high resolution hydrodynamic simulation of the formation of a Milky Way (MW)-like disk galaxy, from the “Eris” simulation suite, dissecting it into cohorts of stars formed at different epochs of cosmic history. At $z=0$, stars with ${\ensuremath{t_\mathrm{form}}}< 2$ Gyr mainly occupy the stellar spheroid, with the oldest (earliest forming) stars having more centrally concentrated profiles. The younger age cohorts populate disks of progressively longer radial scale length and shorter vertical scale height. At a given radius, the vertical density profiles and velocity dispersions of stars vary smoothly as a function of age, and the superposition of old, vertically-extended and young, vertically-compact cohorts gives rise to a double-exponential profile like that observed in the MW. Turning to formation history, we find that the trends of spatial structure and kinematics with stellar age are largely imprinted at birth, or immediately thereafter. Stars that form during the active merger phase at $z>3$ are quickly scattered into rounded, kinematically hot configurations. The oldest disk cohorts form in structures that are radially compact and relatively thick, while subsequent cohorts form in progressively larger, thinner, colder configurations from gas with increasing levels of rotational support. The disk thus forms “inside-out” in a radial sense and “upside-down” in a vertical sense. Secular heating and radial migration influence the final state of each age cohort, but the changes they produce are small compared to the trends established at formation. The predicted correlations of stellar age with spatial and kinematic structure are in good qualitative agreement with the correlations observed for mono-abundance stellar populations in the MW.' author: - | Jonathan C. Bird, Stelios Kazantzidis, David H. Weinberg,\ Javiera Guedes, Simone Callegari, Lucio Mayer, & Piero Madau title: 'Inside Out and Upside Down: Tracing the Assembly of a Simulated Disk Galaxy Using Mono-Age Stellar Populations' --- Introduction {#sec:intro} ============ Eggen, Lynden-Bell, and Sandage ([-@Eggen62]; hereafter, ELS) presented one of the earliest theoretical accounts of Milky Way (MW) formation: stars formed during the gravitational contraction of a rotating gas cloud, with successive generations forming successively more flattened, rotationally supported populations. They argued that this picture could explain the observed correlation between orbital eccentricity and chemical composition. Subsequent analytic models have emphasized the dark matter (DM) halo’s role as primary potential well and baryon angular momentum as a governor of disk scale length, with SN-driven outflows having a major impact on baryonic mass [[[e.g.]{}]{}, @White_Rees78; @Fall80; @White91; @Kauffmann93; @Cole94; @Mo98]. In the cold dark matter (CDM) scenario, hierarchical structure formation leads to complex assembly histories whose effects can only be fully assessed with numerical simulations, which have grown steadily in computational and physical sophistication [[[e.g.]{}]{}, @Katz92; @Navarro97; @Brook12 to give just three snapshots of a voluminous field]. In this paper, we analyze one of the highest resolution simulations ever run of a MW-like galaxy formed from cosmological initial conditions to investigate its detailed assembly history. The physics governing the construction of each major galactic component is likely to be different. Mergers and cannibalism of dwarf satellites may play a major role in the formation of the stellar halo, perhaps accounting for most of its mass [@Searle78; @Bullock01; @Bullock05]. The bulge could be the remnant of early mergers [[[e.g.]{}]{}, @Hopkins10] or the result of internal processes such as bar and bending instabilities that heat the inner disk over the life of the galaxy [[[e.g.]{}]{}, @Debattista04; @Debattista05]. Thin disk formation is likely driven by gas dissipation and angular momentum conservation as in the classic picture. However, bimodal distributions in vertical star counts [[[e.g.]{}]{}, @Gilmore83], kinematics [[[e.g.]{}]{}, @Bensby03], and chemistry [[[e.g.]{}]{}, @Lee11] in the solar neighborhood suggest a distinct thick disk. Many theories of its origin have been proposed. The thick disk may arise from the accretion of stars stripped from satellites [@Abadi03], from a satellite dynamically heating a pre-existing stellar disk [[[e.g.]{}]{}, @Kazantzidis08; @Villalobos08; @Kazantzidis09], from a gas rich merger at early times [@Brook04], from a clumpy and turbulent ISM at high redshift [[[e.g.]{}]{} @Bournaud09], or from stars migrating outwards from the hot, inner disk [@Schonrich09; @Loebman11]. Classic chemical evolution models describe disk galaxy formation using an inside-out star formation and chemical enrichment history [[[e.g.]{}]{}, @Matteucci89]. This approach is sensible; the inner disk is thought to assemble first and form stars faster because of the high density of accreted gas in the center of the galaxy’s potential well. However, both theoretical and observational studies over the last decade have emphasized the potential role of radial migration as a mechanism to redistribute stars and diversify the distribution of stellar chemical composition as a function of radius [[[e.g.]{}]{}, @Sellwood02; @Haywood08; @Roskar08b; @Schonrich09; @Casagrande11; @Minchev12b]. Radial migration alleviates tension between models and observations of the local stellar metallicity distribution and super metal rich stars in the solar annulus [@Schonrich09]; it also naturally reproduces both the vertical density profile of stars and the chemical bimodality of the thin and thick disk [@Schonrich09; @Schonrich09b]. Still, the contribution of radial migration to the dynamical evolution of the disk, specifically thick disk dynamics, is currently debated in the literature [[[e.g.]{}]{}, @Schonrich09b; @Loebman11; @Minchev12a; @Roskar12b]. In this paper, we dissect a galaxy simulated in a fully cosmological context into individual age cohorts of stars formed at different epochs. We study the present-day spatial structure and kinematics of these cohorts, as well as their assembly history and radial migration patterns. Our results are less directly comparable to observations than chemistry-based investigations, but the age cohorts are physically simpler than chemically-selected tracer populations and avoid the uncertainties associated with numerical chemical enrichment implementation [[[e.g.]{}]{}, @Shen10]. In this paper, age cohorts reveal the expected inside-out growth of the disk. The simulated galaxy’s formation can also be described as “upside-down” in the sense that old stars form in a relatively thick component, or are kinematically heated very quickly after their birth. Younger populations form in successively thinner disks. Stars radially migrate throughout the galaxy’s assembly, but migrators do not disrupt the general kinematics-age trends put into place by the “upside-down” construction. There is some overlap of our results with those of @Brook12, who study the buildup of a chemically-defined thin disk, thick disk, and halo in the solar annulus of a less massive simulated disk galaxy; we discuss a comparison with their work in Section \[sec:discussion\]. Motivation for our study comes partly from analysis of the SDSS Sloan Extension for Galactic Understanding and Exploration (SEGUE) [@Yanny09] G dwarf sample by @Bovy12b [@Bovy12a; @Bovy12c]. They dissect the sample into groups spanning narrow ranges of [$[\mathrm{Fe/H}]$]{} and [$[\alpha\mathrm{/Fe}]$]{} and discover that the physical configuration of each mono-abundance population is well described by a single exponential function in both the radial and vertical directions; specifically, more alpha-rich, iron-poor (old) populations have shorter scale lengths but larger scale heights than alpha-poor, iron-rich (young) populations (see their Figure 5). @Bovy12b use this result to contend that the MW’s thick disk is not a distinct structural component, a possibility raised by the mixture models of @Nemec91 [@Nemec93]. The mono-abundance populations also have simple, isothermal kinematics in the vertical direction [@Bovy12c]. The assembly histories of the age-cohorts (Section \[sec:evolution\]) illustrate the simulation’s fully self-consistent galaxy formation scenario that culminates in galactic structure (Section \[sec:current\]) qualitatively similar to the mono-abundance description of the MW. The Simulation {#sec:sim} ============== The cosmological simulation employed in the present study is part of the “Eris” campaign [@Guedes11] of hydrodynamical simulations of the formation of Milky Way (MW)-sized disk galaxies. It was performed with the parallel, TreeSPH $N$-body code <span style="font-variant:small-caps;">gasoline</span> [@Wadsley_etal04] in a [*Wilkinson Microwave Anisotropy Probe*]{} 3-year cosmology [$\Omega_M=0.24$, $\Omega_\Lambda=0.76$, $\Omega_b=0.042$, $H_0=73\,\kmsmpc$, $n=0.96$, $\sigma_8=0.76$; @Spergel07]. [{width="\textwidth"}]{} All simulation parameters are identical to those of @Guedes11, except for the star formation efficiency parameter, which was equal to $\epsilon_{\rm SF} = 0.1$ in the latter study, but it is $\epsilon_{\rm SF} = 0.05$ in the simulation employed here (see below). Specifically, we started from a low-resolution ($300^3$ particles), DM-only simulation corresponding to a periodic box of $90$ Mpc on a side. The target halo was identified at $z=0$ and was chosen to have a mass similar to that of the MW and a rather quiet late merging history, namely to have experienced no major mergers (defined as encounters with mass ratios $\ge 1/8$) after $z=3$. Three halos with masses between $6\times 10^{11}$ and $8\times 10^{11}\,\Mo $ satisfied these conditions in the (90 Mpc)$^3$ box, and we chose to re-simulate the one which appeared to be most “regular” and isolated, with a virial mass of $M_{\rm vir}=7.9\times 10^{11}\,\Mo$ at $z=0$. New initial conditions were then generated with improved mass resolution (by a factor $20^3$) on a nested grid in a Lagrangian sub-region of size $1$ Mpc, centered around the chosen halo, using the standard “zoom-in” technique to add small-scale perturbations. High-resolution particles were further split into 13 million dark matter particles and an equal number of gas particles, for a final dark and gas particle mass of $m_{\rm DM}=9.8\times 10^4\,\Mo$ and $m_{\rm SPH}=2\times 10^4\,\Mo$, respectively. The gravitational softening length was fixed to $120$ pc for all particle species from $z=9$ to $z=0$, and evolved as $1/(1+z)$ from $z=9$ to the starting redshift of $z=90$. At the present epoch, the simulation contains $\sim 7$, $\sim 3$, and $\sim 8.6$ million dark matter, gas, and star particles, respectively, for a total of $N \sim 18.6$ million particles within the virial radius ($R_{\rm vir} \sim 239$ kpc). The version of <span style="font-variant:small-caps;">gasoline</span> used in this study includes Compton cooling, atomic cooling, and metallicity-dependent radiative cooling at low temperatures [@Mashchenko_etal06]. A uniform UV background modifies the ionization and excitation state of the gas and is implemented using a modified version of the @Haardt_Madau96 spectrum. Our star formation recipe follows that of @Stinson_etal06, which is based on that of @Katz92. Star formation occurs when cold ($T<T_{\rm max}$), virialized gas that is part of a converging flow reaches a threshold density ($n_{\rm SF}$). Gas particles spawn stellar particles with a given efficiency $\epsilon_{\rm SF}$ at a rate proportional to the local dynamical time, $t_{\rm dyn}$, $$d\rho_*/dt=\epsilon_{\rm SF} \rho_{\rm gas}/t_{\rm dyn} \propto \rho_{\rm gas}^{1.5}$$ (i.e., locally enforcing a Schmidt law), where $\rho_*$ and $\rho_{\rm gas}$ are the stellar and gas densities, respectively. In our simulation, $T_{\rm max}=3\times 10^4$ K, $n_{\rm SF}=5$ atoms cm$^{-3}$, and $\epsilon_{\rm SF}=0.05$. Star particles are created stochastically with an initial mass $m_* \sim 6.2\times 10^3\,\Mo$, and the gas particle that spawns the new star has its own mass reduced accordingly. Gas particles are deleted from the simulation once their masses fall below $\sim 4\times 10^3\,\Mo$ (with their masses and metals being distributed in neighboring gas particles). A newly formed star particle represents a simple stellar population with its own age and metallicity and a @Kroupa93 initial stellar mass function. Feedback from supernovae (SNe) is treated using the blast-wave model described by @Stinson_etal06 , which is based on the analytic treatment of blastwaves described by @McKee_Ostriker77. Each Type-II SN deposits metals and a net thermal energy of $\epsilon_{\rm SN} \times 10^{51}\,$ergs into the nearest neighboring gas particles, and the fraction of energy that couples to the interstellar medium was set to $\epsilon_{\rm SN}=0.8$ (the same value was adopted in previous cosmological simulations; see, for example, @Governato_etal07). The heated gas has its cooling shut off for a timescale set by the local gas density and temperature and by the total amount of energy injected [@Stinson_etal06]. The energy deposited by many SNe adds up to create larger, hot bubbles and longer cooling shutoff times. No feedback from an active galactic nucleus was included in the simulation. Our high mass and force resolution enables us to adopt a density threshold for star formation that is $50$ times higher compared to those employed in previous lower-resolution studies. The local Jeans length corresponding to our density threshold and $T = 10^3 K$ is resolved with more than $5$ SPH smoothing lengths, and thus artificial fragmentation is prevented [@Bate_Burkert97]; we note that very few gas particles inside the virial radius ever cool below $10^3 K$. Similarly, the Jeans mass at our threshold density is resolved with at least $100$ $m_{\rm SPH}$, where $m_{\rm SPH}$ denotes the mass of the SPH gas particle. Lastly, we stress that the star formation density threshold adopted here is high enough to allow the development of a clumpy, inhomogeneous ISM, which serves as a vital element for the formation of realistic disk galaxies [@Guedes11]. Our simulation produces a galaxy that at $z=0$ has structural properties nearly identical to those of the base “Eris” run described by @Guedes11 (for example the bulge-to-disk ratio and the radial scale length of the disk differ by less than $10\%$ in the two simulations; see @Mayer12). Although we defer a detailed comparison of the two simulations to future work, we stress that the rotation curve of our resulting galaxy is even flatter than that of @Guedes11, perhaps because the nuclear bar is weaker and thus less dense, providing an even better match to the rotation curve of a typical late-type Sb/Sbc spiral galaxy. ![\[fig:400\_pos\] The radial profiles of surface mass density (left panel) and median height above the disk plane (right panel) for all age cohorts at redshift zero. Note [$r_\mathrm{gc}$]{}, used throughout this paper, refers to galactocentric radius. The color of each radial profile indicates cohort formation time; colors progress from red for old stars to blue for young stars (see legend for details; in addition to color, line types help differentiate the oldest cohorts). We also plot the surface mass density and median height of all stars (thick, black lines). All radial profiles in this paper have $0.37$ kpc bins; bins with $n<10$ particles are left blank. Younger cohorts populate more disk-like configurations and reside closer to the plane than their older counterparts. ](figs/400_tf_rp_spatial.eps){width="3.5in"} The Present-Day Galaxy {#sec:current} ====================== {width="\textwidth"} We denote stellar populations grouped by age (age cohorts) with a script $\mathcal{S}$ and subscripts corresponding to the population’s range in formation time; [[e.g.]{}]{}, stars born between $0.0$ and $0.5$ Gyr after the big bang will be collectively referred to as [$\mathcal{S}_{0.0,0.5}$]{}. Cohort age is the difference between the age of the universe in the simulation ($13.327$ Gyr) and the cohort’s formation time. Figure \[fig:400\_den\] shows the stellar surface density of the galaxy as a function of position and age at redshift zero. The specific age cohorts are defined by formation times ([$t_\mathrm{form}$]{}) spanning $0.0$ to $0.5$ Gyr, $0.5$ to $1.0$ Gyr, $1.0$ to $2.0$ Gyr, $2.0$ to $4.0$ Gyr, $4.0$ to $8.0$ Gyr, $8.0$ to $12.8$ Gyr, and the final $\sim500$ Myr of star formation, $12.8$ to $13.4$ Gyr [^1]. Each panel denotes the range of [$t_\mathrm{form}$]{} considered and shows face-on and edge-on views of the galaxy. Following the density distributions from left to right and top to bottom, younger stellar populations show longer, thinner structure than older populations. The oldest cohort ([$\mathcal{S}_{0.0,0.5}$]{}) is concentrated in the central region of the galaxy while stars born during the next 500 Myr ([$\mathcal{S}_{0.5,1.0}$]{}) have a spheroidal mass density profile out to a galactocentric radius ([$r_\mathrm{gc}$]{}) of $\sim 10$ kpc. Compared to stars with [$t_\mathrm{form}$]{}$<1$ Gyr, [$\mathcal{S}_{1.0,2.0}$]{} extends further in radius, and its edge-on view shows a distinct yet subdominant flattened component indicative of an infant disk. [$\mathcal{S}_{2.0,4.0}$]{} is the first cohort to display primarily non-spheroidal spatial structure; the edge-on view shows a vertically extended and relatively compact disk. Edge-on views of [$\mathcal{S}_{4.0,8.0}$]{} and [$\mathcal{S}_{8.0,12.8}$]{} reveal that younger populations inhabit progressively longer and thinner disks. [$\mathcal{S}_{12.8,13.4}$]{} is confined predominantly to the spiral arms and has a very sharp break at ${\ensuremath{r_\mathrm{gc}}}\approx10$ kpc (Figure \[fig:400\_den\], face-on view). In time, the [$\mathcal{S}_{12.8,13.4}$]{} stars in these spiral arms will heat up and increase their random motions, eventually leading to the dissolution of the cohort’s present-day structure (though newly formed stars would enable the spiral arms to persist). The reduced azimuthal structure in the density distributions of [$\mathcal{S}_{4.0,8.0}$]{} and [$\mathcal{S}_{8.0,12.8}$]{} are suggestive of how the spiral overdensity can fade over time. As descriptions of galaxy morphology, the terms “early” and “late” have largely lost their evolutionary connotations, but it is intriguing that the configurations of the age cohorts within this individual simulated galaxy map almost perfectly onto the Hubble sequence from early-type elliptical to late-type spiral. ![\[fig:400\_rp\] The present-day radial profiles of vertical velocity dispersion (top left), radial velocity dispersion (top right), median circularity (bottom left, [$\epsilon$]{}$={\ensuremath{j_z}}/{\ensuremath{j_\mathrm{circ}}}$, where [$j_z$]{} is a particle’s angular momentum and [$j_\mathrm{circ}$]{} the angular momentum for a circular orbit with the particle’s energy $E$), and dispersion of circularity (bottom right) for each age cohort. Younger cohorts are kinematically colder than older cohorts. The color of each radial profile indicates the cohort age as in Figure \[fig:400\_pos\] (see legend).](figs/400_tf_rp_kinematics.eps){width="3.7in"} The age cohorts’ surface mass density profiles quantify their mass contribution to the galaxy as a function of radius (Figure \[fig:400\_pos\], left panel). The color scheme for this figure is repeated throughout the rest of the paper: red colors for the oldest cohorts, progressing through orange, yellow, green, cyan, and dark blue for younger populations. A bulge-like component dominates the inner 2 kpc of the surface mass density profile for all stars (black line); starting at [$r_\mathrm{gc}$]{}$\sim 3.5$ kpc, the profile becomes exponential and shows no break out to ${\ensuremath{r_\mathrm{gc}}}=20$ kpc. As seen in Figure \[fig:400\_den\], [$\mathcal{S}_{0.0,0.5}$]{} contributes only to the central region of the galaxy. The mass density profiles of [$\mathcal{S}_{0.5,1.0}$]{} and [$\mathcal{S}_{1.0,2.0}$]{} extend to larger radii and are intermediate in shape between a power law and exponential characterization. In addition to a central power law distribution, [$\mathcal{S}_{2.0,4.0}$]{} has a qualitatively exponential component at $3<{\ensuremath{r_\mathrm{gc}}}<10$ kpc. Younger populations show dominant exponential components: the mass density profile of [$\mathcal{S}_{4.0,8.0}$]{} is exponential at $3<{\ensuremath{r_\mathrm{gc}}}<20$ kpc while that of [$\mathcal{S}_{8.0,12.8}$]{} has two exponential components over the same radial range with a break at ${\ensuremath{r_\mathrm{gc}}}\sim 12.5$ kpc. Stars born in the last $500$ Myr have the mass profile of a exponential disk with a break radius at $9$ kpc. The cohorts also show trends in their vertical structure: older populations are systematically at greater heights above the plane than younger cohorts at all galactocentric radii (right panel, Figure \[fig:400\_pos\]). Several groups have measured a similar correlation at the solar annulus in the MW and in external galaxies [[[e.g.]{}]{}, @Bensby05; @Yoachim08]. Note that the precipitous rise in the median height above the plane in [$\mathcal{S}_{12.8,13.4}$]{} occurs past the break radius of this population; only a few percent of this cohort’s mass is found in the disk outskirts, and these stars may be kinematically distinct from those inside the break. The median height of all stars is correlated with [$r_\mathrm{gc}$]{}, partly because of an increasing halo to disk ratio in the outskirts of the galaxy (see Section \[sec:discussion\] for the same analysis applied to a kinematically selected disk sample). [llrrrrr]{}\[ht\] (0.0,0.5) & [$\mathcal{S}_{0.0,0.5}$]{} & $0.002$ & $0.018$ & $0.010$ & $0.223$ & $0.749$\ (0.5,1.0) & [$\mathcal{S}_{0.5,1.0}$]{} & $0.058$ & $0.021$ & $0.053$ & $0.182$ & $0.743$\ (1.0,2.0) & [$\mathcal{S}_{1.0,2.0}$]{} & $0.186$ & $0.103$ & $0.111$ & $0.250$ & $0.536$\ (2.0,4.0) & [$\mathcal{S}_{2.0,4.0}$]{} & $0.313$ & $0.438$ & $0.143$ & $0.211$ & $0.208$\ (4.0,8.0) & [$\mathcal{S}_{4.0,8.0}$]{} & $0.260$ & $0.725$ & $0.045$ & $0.097$ & $0.134$\ (8.0,12.8) & [$\mathcal{S}_{8.0,12.8}$]{} & $0.165$ & $0.852$ & $0.015$ & $0.073$ & $0.060$\ (12.8,13.4) & [$\mathcal{S}_{12.8,13.4}$]{} & $0.017$ & $0.907$ & $0.023$ & $0.051$ & $0.018$\ Total & & $1.000$ & $0.501$ & $0.083$ & $0.161$ & $0.254$\ \[tab:decomp\] We show each age cohort’s vertical density profile at three different galactocentric radii in Figure \[fig:400\_vp\]. The profile of all stars in the solar annulus ($7<{\ensuremath{r_\mathrm{gc}}}<9$ kpc, middle panel, black line) is consistent with the classic two exponential profile shape observed in the MW [@Gilmore83] and recently measured to have scale heights $z_{{ \mathrm h}_{thin}}=0.3$ kpc and $z_{{ \mathrm h}_{thick}}=0.9$ kpc [@Juric_etal08]; the simulated galaxy’s scale heights are a factor of $1.6$-$1.8$ larger at $z_{{\mathrm h}_1}=0.55$ kpc and $z_{{\mathrm h}_2}=1.47$ kpc but maintain the same thick to thin disk ratio. Traditionally, this double-exponential vertical density profile has been interpreted as two distinct structural components, but Figure \[fig:400\_vp\] clearly shows that the superposition of all cohort vertical density profiles, spanning a continuum of scale heights (increasing with age), naturally yields the double-exponential profile (see Section \[sec:discussion\]). Incorporating data from the inner and outer disk (left and right panels, respectively), we find two general trends governing the relationship between age, scale height, and radius: older stellar populations have more extended vertical density profiles than younger age cohorts, while each individual cohort’s vertical profile becomes more extended at larger radii. The galaxy has dynamical trends with age as well: the orbits of older age cohorts are systematically hotter than their younger counterparts in the radial and vertical directions at nearly all radii (top row of Figure \[fig:400\_rp\]). The same general trend is seen in studies based on the Geneva Copenhagen Survey in the MW [@Holmberg07; @Casagrande11], though some explorations of smaller data sets find a plateau in the age-velocity relationship for stars of intermediate age [[[e.g.]{}]{}, @Wielen77; @Quillen01; @Soubiran08]. Each individual age cohort’s velocity dispersion profiles have note-worthy features. The velocity dispersions of [$\mathcal{S}_{0.5,1.0}$]{} and [$\mathcal{S}_{1.0,2.0}$]{} are locally maximal at the edges of their bulge-like mass distributions (${\ensuremath{r_\mathrm{gc}}}\sim 3$ and $6$ kpc, respectively). Each cohort of stars with [$t_\mathrm{form}$]{}$>4$ Gyr shows little internal variation in velocity dispersion as a function of radius in the galaxy’s disk out to [$r_\mathrm{gc}$]{}$\sim 12.5$ kpc; thereafter, there is a positive gradient in [$\sigma_{\mathrm{v_r}}$]{} and [$\sigma_{\mathrm{v_z}}$]{} with radius. Interestingly, the mass density profile of [$\mathcal{S}_{2.0,4.0}$]{} is consistent with that of an exponential disk at [$r_\mathrm{gc}$]{}$>5$ kpc (Figure \[fig:400\_rp\]), yet this radius is precisely the starting point for a large, positive radial gradient in its [$\sigma_{\mathrm{v_r}}$]{} and (to a lesser degree) [$\sigma_{\mathrm{v_z}}$]{} profiles, which is not seen in the younger cohorts except for the sparse outskirts of [$\mathcal{S}_{12.8,13.4}$]{}. It is plausible that the vertically extended disk of [$\mathcal{S}_{2.0,4.0}$]{} experienced a significantly different evolutionary history than the younger disk populations. The bottom row of Figure \[fig:400\_rp\] examines the orbital shape (circularity) of cohort members as a function of radius. To define orbital circularity ([$\epsilon$]{}), we first choose a coordinate system such that the origin is coincident with the stellar particles’ center of mass and the $z$-axis is aligned with the total angular momentum vector of the stars. Each particle’s angular momentum in the plane of the disk is then [$j_z$]{}, and positive [$j_z$]{}denotes a corotating orbit. Circular orbits have the maximal [$j_z$]{} (hereafter [$j_\mathrm{circ}$]{}) given a particle’s total energy ($E_\mathrm{tot}$, see below). A particle’s circularity ([$\epsilon$]{}) is defined as [$j_z$]{}$/$[$j_\mathrm{circ}$]{}$(E_\mathrm{tot})$. Following the trends in velocity dispersion, older stars are on less circular orbits and show more variation in their orbital shape than younger populations. The median circularity within the disk components of [$\mathcal{S}_{2.0,4.0}$]{}, [$\mathcal{S}_{4.0,8.0}$]{}, and [$\mathcal{S}_{8.0,12.8}$]{} remains constant over a radial range that increases when younger populations are considered. The dispersion in circularity, however, grows with radius for these three cohorts; the radial onset of the increase is inversely correlated with age. Stars formed during the last $500$ Myr are generally born on very circular orbits and only have significant circularity dispersion past the [$\mathcal{S}_{12.8,13.4}$]{} disk break radius. ![\[fig:tf0.0\_den\] The surface density of [$\mathcal{S}_{0.0,0.5}$]{} as a function of position and time. Edge-on and face-on views are shown at the time of each snapshot (labeled in gigayears at the bottom right hand corner of each panel). Pixel color represents the logarithmic surface mass density at the pixel position (see colorbar). The coordinate system is such that the disk lies in the $xy$ plane and the galaxy’s angular momentum vector is aligned with the $z$ axis. Note the changing spatial scale: for clarity, our axis limits either correspond to the box encompassing $95\%$ of the cohort’s mass or the region considered in Figure \[fig:400\_den\] ($|x|,|y|,|z| < 17.5,17.5,4.0$ kpc), whichever is larger. [$\mathcal{S}_{0.0,0.5}$]{} assembles in the center of the galaxy quickly. Snapshots at later times show no qualitative differences with $t=4.0$ Gyr.](figs/tf_0.0_0.5_density.eps){width="3.7in"} {width="\textwidth"} To aid in our description of the current state of the galaxy, we kinematically separate the galaxy into structural components using particle energy and angular momentum and following a modified version of the prescription outlined in @Abadi03. We adhere to convention and describe these components as the “thin disk”, “pseudobulge”, [[etc.]{}]{}, though this nomenclature implies, to some degree, sharp distinctions in the observed continuous distribution. The principal discriminants in our decomposition are circularity and total specific energy defined by $E_\mathrm{tot}=\frac{1}{2}v^2 + \Phi$; where $v$ is the total velocity of the particle and $\Phi$ is the value of the potential (an output of the simulation code) at the position of the particle. We first assign all particles with ${\ensuremath{\epsilon}}>0.8$ to the thin disk as it should contain mostly circular orbits (${\ensuremath{\epsilon}}=1$ for a circular orbit). We assign to the spheroid all particles with circularity below a critical threshold ${\ensuremath{\epsilon_\mathrm{crit}}}=0.33$, which is chosen such that the ensemble of all particles with [$\epsilon$]{}$<{\ensuremath{\epsilon_\mathrm{crit}}}$ exhibit zero net rotation (all counter-rotating particles are assigned to the spheroid). The remaining population has [$\epsilon_\mathrm{crit}$]{}$<{\ensuremath{\epsilon}}<0.8$ and is subdivided into the thick disk (${\ensuremath{E_\mathrm{tot}}}\ge E_\star$) and pseudobulge (${\ensuremath{E_\mathrm{tot}}}< E_\star$) based on the median energy of all stellar particles ($E_\star$). Note that this definition of pseudobulge is not identical to that used in @Guedes12, where the pseudobulge was first identified photometrically as the stellar mass contained in the inner $2$ kpc, which has a higher Sec profile index than the disk. Their pseudobulge was then constrained further by including only stars with circularity below $0.8$ to exclude disk stars. A fraction of stars that we identify as belonging to the thick disk via the energy cut or to the spheroid via the circularity threshold would have been classified as pseudobulge using the @Guedes12 definition. Our kinematic distinctions are somewhat arbitrary and may not perfectly describe the galaxy. The caveats to this decomposition are similar to those in @Abadi03: the spheroidal component could have a modest amount of rotation in reality, the circularity cutoff for the thin disk may not precisely match that in the MW, and the energy cuts do not ensure that our structural distinctions are complete and free of contamination. In addition, our choice to uniquely assign all regions of the $E_\mathrm{tot}$, [$\epsilon$]{} plane to specific structures demands that the [$\epsilon$]{} distribution of the spheroid is not symmetric about [$\epsilon$]{}$=0$, extending to [$\epsilon$]{}$=-1$ for counter-rotating stars but stopping at [$\epsilon_\mathrm{crit}$]{} for co-rotating orbits. We stress that our subsequent analysis and conclusions are not based on these kinematic selections. However, the decomposition gives us a qualitative description of the galaxy’s kinematic structure that provides a familiar context for our analysis. In Table \[tab:decomp\], we show how each age cohort contributes to the galactic structures identified in our kinematic decomposition. The results corroborate much of what we see in Figure \[fig:400\_den\]. Old stellar populations with [$t_\mathrm{form}$]{}$<2.0$ Gyr predominantly contribute to the spheroid. After $t=4$ Gyr, new stars overwhelmingly populate the disk of the galaxy. [$\mathcal{S}_{2.0,4.0}$]{} is an intermediate population: significant fractions of its members have kinematics consistent with classical disk and spheroid orbits. The total fractions confirm that the simulated galaxy has a strong disk component. The majority of all particles have high circularity ($f_{{\ensuremath{\epsilon}}>0.7}=0.6$), similarly prominent high circularity populations are found in the most disk-dominated galaxies simulated with the latest implementation of the Tree-SPH code GADGET-3 [@Aumer13]. ![\[fig:tf0.5\_den\] The surface density of [$\mathcal{S}_{0.5,1.0}$]{} as a function of position and time. The orientation of the galaxy, calculation of axis limits, and color scheme are the same as in Figure \[fig:tf0.0\_den\]. The time of each snapshot is labeled in gigayears at the bottom right hand corner of each panel. Later outputs are omitted as they show no qualitative changes since $t=6.0$ Gyr.](figs/tf_0.5_1.0_density.eps){width="3.7in"} Evolution of Age Cohorts {#sec:evolution} ======================== {width="\textwidth"} We now examine the assembly history of the galaxy using the individual age cohorts as tracer populations. In this section we adjust our choice of age cohorts so that each represents at most a 1 Gyr range of formation times. The first three [$t_\mathrm{form}$]{} bins ([$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{}, and [$\mathcal{S}_{1.0,2.0}$]{}) are illustrative of the relatively complex assembly of stellar populations at early times. [$\mathcal{S}_{3.0,4.0}$]{} is an “intermediate” population born just after the major merger epoch concludes. The final age cohort described here, [$\mathcal{S}_{7.0,8.0}$]{}, is representative of populations born and evolved [*in situ*]{} in a quiescent disk. ![\[fig:tf1.0\_den\] The surface density of [$\mathcal{S}_{1.0,2.0}$]{} as a function of position and time. The orientation of the galaxy, calculation of axis limits, and color scheme are the same as in Figure \[fig:tf0.0\_den\]. The time of each snapshot is labeled in gigayears at the bottom right hand corner of each panel. Later outputs are omitted as they show no qualitative changes since $t=8.1$ Gyr.](figs/tf_1.0_2.0_density.eps){width="3.7in"} **The First 500 Myr** We consider the surface density distribution of [$\mathcal{S}_{0.0,0.5}$]{} as a function of time, specifically at $t = 0.6$, $1.1$, $2.0$, and $4.0$ Gyr, in Figure \[fig:tf0.0\_den\]. These stars populate a number of different subhalos at early times, but their dominant concentration is in the parent halo of the galaxy (note the changing scale on each panel in the figure). The final merger event involving a perceptible fraction of [$\mathcal{S}_{0.0,0.5}$]{} is well underway at $t=2$ Gyr; by $t=4$ Gyr, [$\mathcal{S}_{0.0,0.5}$]{} has assembled into a configuration much like what we see today (compare the bottom right of Figure \[fig:tf0.0\_den\] with the top left of Figure \[fig:400\_den\]). We omit snapshots at later times as [$\mathcal{S}_{0.0,0.5}$]{} shows no further qualitative changes in its density distribution. {width="\textwidth"} Physical parameters describing spatial structure and kinematics confirm that the [$\mathcal{S}_{0.0,0.5}$]{} population assembles fairly rapidly and evolves quiescently thereafter (Figure \[fig:tf0.0\_rp\]). Both the satellite accretion and the decreasing central density seen in Figure \[fig:tf0.0\_den\] broaden the cohort’s radial surface mass density profile with time. The radial profiles of vertical velocity dispersion (right panel) and median height above the plane (middle panel) show that [$\mathcal{S}_{0.0,0.5}$]{} kinematically heats up as it assembles but only makes moderate energy gains after $t=4$ Gyr. [$\mathcal{S}_{0.0,0.5}$]{} congregates early in the galaxy’s history, remains centrally concentrated, and evolves quiescently after its last substantial merger event. **$\mathbf{ {\ensuremath{t_\mathrm{form}}}=[0.5,1.0]}$ Gyr** Similar to [$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{} is scattered amongst several subhalos at early times but most of its mass is still born in the parent halo (Figure \[fig:tf0.5\_den\]). The merger event at $t=2$ Gyr is more significant for [$\mathcal{S}_{0.5,1.0}$]{} than [$\mathcal{S}_{0.0,0.5}$]{}. Still, after another $2$ Gyr, the morphological features associated with the encounter have dissipated. From $t=4$ Gyr to $t=6$ Gyr, the density distribution of stars outside [$r_\mathrm{gc}$]{}$>6$ kpc becomes more symmetric. Later snapshots do not show any qualitative changes. The surface mass density radial profile of [$\mathcal{S}_{0.5,1.0}$]{} resembles its current state throughout most of the galaxy by $t\sim4$ Gyr; however, a perceptible decrease in mass within the innermost kiloparsec and corresponding increase in $1<{\ensuremath{r_\mathrm{gc}}}<3$ kpc persists until $t=8$ Gyr (Figure \[fig:tf0.5\_rp\], left panel). [$\mathcal{S}_{0.5,1.0}$]{} is born with more random vertical motion than [$\mathcal{S}_{0.0,0.5}$]{} initially had; like [$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{} also experiences an early, dramatic increase in its vertical energy (Figure \[fig:tf0.5\_rp\], right panel). Like [$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{} shows almost no evolution of the vertical velocity dispersion after $t=4$ Gyr. Only the central kiloparsec shows any appreciable change in the median height of the cohort after this time, with growth of the median height by about 100 pc (Figure \[fig:tf0.5\_rp\], middle panel). [$\mathcal{S}_{0.5,1.0}$]{}, like [$\mathcal{S}_{0.0,0.5}$]{}, quickly becomes the kinematically hot, vertically extended population seen at redshift zero (Section \[sec:current\]). We also note that a central bulge-like component with steeper density profile slope has already formed at early times as a result of bar-like instabilities in the early disk [see @Guedes12]. **$\mathbf{ {\ensuremath{t_\mathrm{form}}}=[1.0,2.0]}$ Gyr** The surface density distribution of [$\mathcal{S}_{1.0,2.0}$]{} is shown as a function of time in Figure \[fig:tf1.0\_den\]. [$\mathcal{S}_{1.0,2.0}$]{} has a significant [*in situ*]{} population that forms either during or just prior to the merger event at $t=2.0$ Gyr. This merger strongly impacts the cohort’s morphology. In spite of these turbulent beginnings, almost all of the [$\mathcal{S}_{1.0,2.0}$]{} stars have been assimilated into a coherent structure by $t=4$ Gyr. [$\mathcal{S}_{1.0,2.0}$]{}’s configuration is the first to show an ellipsoidal disk component; the disk is evident just after the cohort forms ($t=2$ Gyr) and is obvious by $t=4$ Gyr (see edge-on views). From $t=4$-$8$ Gyr, the density distribution at [$r_\mathrm{gc}$]{}$>5$ kpc becomes more smooth, and any discernible tilt to the ellipsoid vanishes. Despite differences in their assembly history, [$\mathcal{S}_{1.0,2.0}$]{}’s kinematic properties are similar to those of older populations. The cohort’s radial surface density profile remains constant after $t=4$ Gyr, except for some emigration away from the innermost kiloparsec (Figure \[fig:tf1.0\_rp\], left panel). The initial, [*in situ*]{} population is born with relatively large [$\sigma_{\mathrm{v_z}}$]{}, yet there is still a substantial increase in the random vertical motions of this cohort as its satellite-born populations merge (Figure \[fig:tf1.0\_rp\], right panel), including the satellite still visible in the $t=4$ Gyr snapshot. As with older cohorts, there is little heating at late times, and both the median [$\sigma_{\mathrm{v_z}}$]{} and median height profiles show essentially no evolution after $t=6$ Gyr. ![\[fig:tf3.0\_den\] The surface density of [$\mathcal{S}_{3.0,4.0}$]{} as a function of position and time. The orientation of the galaxy, calculation of axis limits, and color scheme are the same as in Figure \[fig:tf0.0\_den\]. The time of each snapshot is labeled in gigayears at the bottom right hand corner of each panel. Later outputs are omitted as they show no qualitative changes since $t=9.0$ Gyr.](figs/tf_3.0_4.0_density.eps){width="3.7in"} {width="\textwidth"} **$\mathbf{ {\ensuremath{t_\mathrm{form}}}=[3.0,4.0]}$ Gyr** Many of the evolutionary trends established for stars with [$t_\mathrm{form}$]{}$< 2$ Gyr are no longer evident when we examine [$\mathcal{S}_{3.0,4.0}$]{} (Figure \[fig:tf3.0\_den\]). Here, the vast majority of the cohort is born [*in situ*]{} in the galactic disk; the early [$\mathcal{S}_{3.0,4.0}$]{} disk shows a prominent $m=2$ mode spiral (top left of Figure \[fig:tf3.0\_den\]). The initial snapshot also reveals two significant satellites that will soon interact with the [*in situ*]{} population. At $t=5$ Gyr (top right), one of the two satellites is fully disrupted and the cohort’s strong spiral structure has dissipated. Over the next two billion years, the second satellite merges, the spiral structure dissolves, and there is a small but perceptible thickening of the disk. Stars at large radii continue to phase-mix over time ($t=9$ Gyr, bottom right). [$\mathcal{S}_{3.0,4.0}$]{} is the oldest cohort to have a dominant exponential component in its surface density radial profile (Figure \[fig:tf3.0\_rp\], left panel). As seen in the density plots, this cohort’s disk was in place at early times. The density profile shows some mass redistribution from the interior of the disk ([$r_\mathrm{gc}$]{}$<6$ kpc) to the outer disk over time, moderately increasing the cohort’s disk scale length. We investigate radial migration patterns of this cohort and others in Section \[sec:radmix\_tf\]. The radial profile of [$\sigma_{\mathrm{v_z}}$]{} (right panel) shows that [$\mathcal{S}_{3.0,4.0}$]{}, while kinematically hot for a disk population, is born with lower internal energy than the older cohorts. [$\mathcal{S}_{3.0,4.0}$]{} members within [$r_\mathrm{gc}$]{}$<6$ kpc become modestly more energetic with time (Figure \[fig:tf3.0\_rp\], right panel); over the last $10$ Gyr, the population’s heating rate ($\Delta{\ensuremath{\sigma_{\mathrm{v_z}}}}/ \Delta t$) is twice that of [$\mathcal{S}_{1.0,2.0}$]{}. While the older populations became colder at increasing radius, [$\mathcal{S}_{3.0,4.0}$]{} has a positive [$\sigma_{\mathrm{v_z}}$]{} radial gradient; we suspect that this is most likely due to relatively cold, disk orbits at smaller radii giving way to hotter halo kinematics past the break radius of the [$\mathcal{S}_{3.0,4.0}$]{} disk (Section \[sec:discussion\] examines a kinematically selected disk sample). The median height of the cohort increases between $t=4$ and $t=5$ Gyr at [$r_\mathrm{gc}$]{}$<5$ kpc, while the outer disk settles to a thinner configuration. ![\[fig:tf7.0\_den\] The surface density of [$\mathcal{S}_{7.0,8.0}$]{} as a function of position and time. The orientation of the galaxy, calculation of axis limits, and color scheme are the same as in Figure \[fig:tf0.0\_den\]. The time of each snapshot is labeled in gigayears at the bottom right hand corner of each panel. Later outputs are omitted as they show no qualitative changes since $t=11.5$ Gyr.](figs/tf_7.0_8.0_density.eps){width="3.7in"} {width="\textwidth"} **$\mathbf{ {\ensuremath{t_\mathrm{form}}}=[7.0,8.0]}$ Gyr** [$\mathcal{S}_{7.0,8.0}$]{} is illustrative of stellar populations born after redshift one: it forms almost exclusively in the disk and secularly evolves (Figure \[fig:tf7.0\_den\]). The initial disk of [$\mathcal{S}_{7.0,8.0}$]{} is thin and shows obvious warping in its outskirts (Figure \[fig:tf7.0\_den\], top left panel). During this cohort’s evolution, a small bulge-like component re-grows in the central region of the galaxy [see @Guedes12], the initial outer disk warps weaken, and the cohort’s disk becomes more vertically and radially extended. The cohort’s surface density radial profile clearly shows a central component giving way to an exponential disk at [$r_\mathrm{gc}$]{}$\sim 2$ kpc (Figure \[fig:tf7.0\_rp\], left panel). Like [$\mathcal{S}_{3.0,4.0}$]{}, there is a net transfer of mass outwards; here, the effect is more dramatic and the donor region extends out to ${\ensuremath{r_\mathrm{gc}}}\sim11$ kpc. The inner disk ([$r_\mathrm{gc}$]{}$\la 10$ kpc), born cold, experiences significant vertical heating, with [$\sigma_{\mathrm{v_z}}$]{} at the solar radius growing from $\approx 20$ [km s$^{-1}$]{} to $\approx 30$ [km s$^{-1}$]{}(Figure \[fig:tf7.0\_rp\], right panel). More cohort members populate the sparse outer disk, beyond [$r_\mathrm{gc}$]{}$\approx 12$ kpc, over time and reduce the vertical velocity dispersion there. The central mass concentration seems distinct from the disk; stars in the central few hundred parsecs continue to move further from the plane throughout the cohort’s history. In a departure from previously examined cohorts, [$\mathcal{S}_{7.0,8.0}$]{}’s entire disk slowly grows more vertically extended as a function of time (middle panel). The gradual increases in vertical velocity dispersion and median particle height are fractionally large compared to the late-time evolution of older cohorts, but they are still modest on an absolute scale. Radial Migration {#sec:radmix_tf} ================ Radial migration, [[i.e.]{}]{}, stellar orbital angular momentum changes brought about by resonant interactions, can have a profound impact on the evolution of disk galaxies [@Schonrich09; @Roskar12a and references therein]. We now investigate the age cohorts’ migratory patterns. Stars are grouped into the same age cohorts used in Section \[sec:current\], except that all stars with [$t_\mathrm{form}$]{}$<2$ Gyr are now labeled as a single cohort. We limit contamination from halo stars by constraining our migration analysis to particles within $4$ kpc of the galactic plane at $z=0$ [^2]. Figure \[fig:rform\] shows the birth radii for cohort members currently residing in three different areas of the disk. The middle panel examines the present day solar annulus ($7\leq{\ensuremath{r_\mathrm{gc}}}\leq9$ kpc). The oldest inhabitants of the solar annulus ([$\mathcal{S}_{0.0,0.2}$]{}) come from the broadest range of formation sites. Stars born at large radii likely formed in a satellite halo before merging with the [*in situ*]{} population. Members of [$\mathcal{S}_{0.0,2.0}$]{} are predominantly associated with the spheroid (Section \[sec:current\]) and only contribute $7\%$ of the total mass at this annulus (see Figure legend). For current solar annulus stars with [$t_\mathrm{form}$]{}$>2$ Gyr, progressively younger stars are less centrally concentrated at birth; the median [$r_\mathrm{form}$]{} is $5.2$, $6.6$, $7.8$, and $7.9$ kpc for [$\mathcal{S}_{2.0,4.0}$]{}, [$\mathcal{S}_{4.0,8.0}$]{}, [$\mathcal{S}_{8.0,12.8}$]{}, and [$\mathcal{S}_{12.8,13.4}$]{}, respectively. Inside-out growth ensures that the same correlation between age and formation radius found in the solar annulus exists both in the inner (left panel) and outer (right panel) disk. Excluding [$\mathcal{S}_{0.0,2.0}$]{} and young stars with little time to migrate ([$\mathcal{S}_{12.8,13.4}$]{}), the median radial excursion ($r_{\mathrm{final}}-$[$r_\mathrm{form}$]{}) of a cohort increases with cohort age and the final radius considered. Though the galaxy was built up in an inside-out fashion, it is important to note that the birth sites of old stars are not restricted to the central galaxy. For example, $20\%$ of [$\mathcal{S}_{2.0,4.0}$]{} members currently in the solar annulus have $6<{\ensuremath{r_\mathrm{form}}}<10$ kpc. The radial excursions seen in Figure \[fig:rform\] are the result of two distinct scattering processes. Stars can be born on, or non-resonantly scattered to, eccentric orbits with large epicyclic amplitude, causing large in-plane motions that sweep through wide ranges in radius without increasing the orbit’s angular momentum. Stars can also can gain or lose angular momentum, thereby changing their guiding centers, through resonant interactions [^3] with spiral waves [[[e.g.]{}]{}, @Sellwood02; @Roskar12a], bar resonance overlap [@Minchev10], and satellite perturbations [@Quillen09; @Bird12]. Figure \[fig:gcr\] is similar to Figure \[fig:rform\] but it shows the distribution of current guiding center radii ([$R_\mathrm{g}$]{}) for each age cohort; a star with angular momentum [$j_z$]{} has a guiding center equal to the radius of a circular orbit with the same angular momentum. If a star’s guiding center radius lies outside the indicated annulus, then the star appears in the annulus only because its eccentricity allows it to spend some of its orbit there. For example, the [$\mathcal{S}_{2.0,4.0}$]{} population in the solar annulus (${\ensuremath{r_\mathrm{gc}}}=7.0-9.0$ kpc, yellow curve in the middle panel of Figure \[fig:gcr\]) are predominantly stars with $R_g < 6$ kpc, which are present near the solar radius because they are on the outer excursions of their eccentric orbits. The trends with age in Figure \[fig:gcr\] reflect the combination of inside-out formation and the age-kinematics correlation. Older cohorts have hotter kinematics either at birth or after early heating by mergers, so they are more likely to spend time well outside their guiding center radius. Furthermore, because older stars form preferentially at small [$r_\mathrm{gc}$]{}, the members of old cohorts found in outer annuli are primarily stars on outward radial excursions. By contrast, the young stars ([$\mathcal{S}_{12.8,13.4}$]{}) in each annulus tend to be stars that were born in that annulus and have guiding centers in that annulus. The oldest stars ([$\mathcal{S}_{0.0,2.0}$]{}) in each annulus are generally on highly eccentric, low angular momentum orbits with low $R_g$. Scattering by the corotational resonances (CR) of spiral density waves is a particularly interesting mechanism of radial migration because it allows stars to shift their guiding centers (change angular momentum) without increasing orbital eccentricity [@Sellwood02]. Figure \[fig:dgcr\] shows the distribution of $\Delta R_g$, the change in guiding center radius between formation and $z=0$, for the same three annuli illustrated in Figures \[fig:rform\] and \[fig:gcr\]. At ${\ensuremath{r_\mathrm{gc}}}=4.0-5.0$ kpc, these distributions are symmetric and fairly compact for all age cohorts. At the solar annulus, the [$\mathcal{S}_{2.0,4.0}$]{} and [$\mathcal{S}_{4.0,8.0}$]{} distributions are positively skewed and substantially broader, a trend that is still stronger at ${\ensuremath{r_\mathrm{gc}}}=11.0-12.0$ kpc. Even the [$\mathcal{S}_{8.0,12.8}$]{} cohort has a positively skewed $\Delta R_g$ distribution at this outer radius. The imposed condition of older stars at large radii preferentially selects stars that have experienced positive $\Delta R_g$. In sum, radial migration is an important effect in the outer galaxy. At the solar annulus and beyond, most stars with $t_{\rm form} < 8.0$ Gyr formed well inside their current galactocentric radius (Figure \[fig:rform\]). This migration reflects the combined impact of the higher orbital eccentricities in older populations, which allow stars to spend much of their orbits well beyond their guiding center radii (Figure \[fig:gcr\]), and angular momentum changes that shift guiding center radii outwards (Figure \[fig:dgcr\]). The importance of radial migration in the outskirts of disk galaxies was made clear by @Roskar08a and @SanchezBlazquez09; studies cite radial migration as the probable origin for the inverse age-radius gradients seen outside the break radius of nearby disk galaxies [[[e.g.]{}]{}, @RadburnSmith12; @Yoachim12]. Our findings here, in conjunction with Figure \[fig:400\_den\], suggest that the older, migrating stars travel past the break radius of the galaxy’s gas reservoir at late times as traced by the density profile of the youngest stars ([$\mathcal{S}_{12.8,13.4}$]{}). We discuss possible implications for thick disk evolution in the Section \[sec:discussion\]. {width="\textwidth"} {width="\textwidth"} {width="\textwidth"} Discussion {#sec:discussion} ========== The shared lifetime of mono-age populations makes them physically intuitive descriptors of the galaxy formation process. A cohort’s members experience relatively similar dynamical histories. Their spatial origin tracks the star formation process, and their initial kinematics traces that of the gas at the cohort’s formation epoch. The oldest three cohorts ([$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{}, [$\mathcal{S}_{1.0,2.0}$]{}) describe the galaxy’s turbulent youth. The first significant star formation occurs in the parent halo; the [*in situ*]{} component of the old cohorts show the classic signs of inside-out galaxy formation. As gas accumulates, cools, and collapses within the main halo, both the rate and radial extent of star formation contributing to the [*in situ*]{} population increases (Figures \[fig:tf0.0\_den\],  \[fig:tf0.5\_den\],  \[fig:tf1.0\_den\]; top left panels). The spatial distribution of formation sites reveals a similar growth process for satellite-born stellar populations within their host halos (same Figures as above). During the major merger epoch, the energy in the velocity difference of the interacting satellite and parent halos is converted into the internal energy of the remnant [@BinneyTremaine]. All cohorts that exist at this time show dramatic and rapid increases in their internal energies (see kinematics of [$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{}, [$\mathcal{S}_{1.0,2.0}$]{}; Figures \[fig:tf0.0\_rp\],  \[fig:tf0.5\_rp\],  \[fig:tf1.0\_rp\]). The present day spatial configuration and kinematic description of the aforementioned cohorts are in place once the last of their satellite-born members have merged with the parent halo ($t\approx2$-$4$ Gyr; see Figures \[fig:tf0.0\_den\] -  \[fig:tf1.0\_rp\]). Almost $90\%$ of stars with [$t_\mathrm{form}$]{}$<2$ Gyr are associated with the spheroid in the present-day galaxy (Table \[tab:decomp\]). The vast majority of stars born after $t=2$ Gyr form in the disk (top left panel of Figures \[fig:tf3.0\_den\] and  \[fig:tf7.0\_den\]). Monotonic trends of position and velocity with age, consistent with stars forming from a cooling gas reservoir, characterize the disk’s assembly. [$\mathcal{S}_{3.0,4.0}$]{} and [$\mathcal{S}_{7.0,8.0}$]{} are illustrative examples; progressively younger cohorts form with larger radial break radii, colder vertical velocity distributions, and more compact vertical structure (see initial outputs in Figures \[fig:tf3.0\_rp\] and  \[fig:tf7.0\_rp\]). Younger cohorts are more susceptible to secular heating mechanisms (Section \[sec:evolution\]), but any changes in population thickness and velocity dispersion with time are not sufficient to disrupt the strong trends between cohort kinematics and age established by each cohort’s initial properties. The galaxy retains much of its assembly history: old cohorts are born thick (or quickly become so), while increasingly younger cohorts form thinner structures with longer radial scales. The galaxy forms inside-out and upside-down. ![\[fig:vrot\_vzdisp\] The ratio of mean rotational velocity to vertical velocity dispersion as a function of redshift for gas under the star formation threshold temperature of the simulation ($T<3\times10^4$ K) in the disk and close to the plane ($4<{\ensuremath{r_\mathrm{gc}}}<8$ kpc and $|z|<2$ kpc, filled circles). The rotated crosses mark the corresponding median height of the cold gas reservoir. Gas viable for star formation becomes increasingly rotation-dominated and vertically compact throughout the simulation. To highlight trends over noise, each marker represents a three-point time-weighted moving average. ](figs/vrot_sigmavz_gas_0.0_3.0K.eps){width="3.7in"} It is important to distinguish between a cohort’s internal energy at birth and subsequent evolution due to mergers or other scattering events. The former is wholly dependent on the properties of the gas when the cohort forms, while the latter is affected by gravitational interactions following the cohort’s inception. Figure \[fig:vrot\_vzdisp\] shows the ratio of mean rotational velocity to vertical velocity dispersion and median height for gas below the star formation threshold temperature of the simulation ($T<3\times10^4$ K) and spatially coincident with the growing disk ($4<{\ensuremath{r_\mathrm{gc}}}<8$ kpc, $|z|<2$ kpc). Gas eligible to form stars becomes steadily more rotationally dominated as time progresses, leading to increasingly cooler, more vertically compact stellar populations. [@Forbes12] find similar results in one-dimensional models of the vertical structure of an evolving, star-forming disk. In our simulation, the star-forming gas rapidly contracts despite mergers and SNe feedback at $t<3$ Gyr due to its high physical density (and short cooling time) within the small halo in the early universe. By $t\approx3$ Gyr, the cooling time is long enough that strong perturbations by several mergers (Figures \[fig:tf0.5\_den\], \[fig:tf1.0\_den\], \[fig:tf3.0\_den\]) and stirring by SN feedback (outflows are more powerful owing to higher star formation rates that peak at $t\approx3.5$ Gyr) can prevent further collapse as suggested by the plateau at $t\approx3$-$5$ Gyr. After $t\approx 5$ Gyr, the merger rate declines rapidly, and the gaseous disk undergoes another period of collapse. The last minor merger, at $t\approx7$ Gyr, may temporarily hinder further collapse, but the slow rate of contraction at $t>8$ Gyr is more difficult to characterize. At present, the cold gas has [$\sigma_{\mathrm{v_z}}$]{}$\approx 15$ [km s$^{-1}$]{}over the radial range ($0.5<{\ensuremath{r_\mathrm{gc}}}<9$ kpc). The late-time evolution is the most uncertain regime of Figure \[fig:vrot\_vzdisp\], as it is most likely to depend on details of the SN feedback scheme and on poorly understood aspects of ISM physics that can influence the level of turbulent motion in the star-forming gas (see, e.g., @Pilkington11). The absence of molecular cooling in the simulation could artificially suppress vertical contraction at late times. However, the evolution at $t < 8$ Gyr reflects the dynamic assembly history of the galaxy and the impact of vigorous supernova feedback at early times, so we expect the trends to remain robust to details of the ISM modeling. The evolving kinematics of the parent gas component indicate that older stellar cohorts are truly born hotter than their younger counterparts, yielding a smooth transition from thick to thin disk over time. In detail, the upside-down construction of the disk and the final stellar age-velocity relationship (AVR) are the integrated response of the age cohorts to the contracting gas disk and dynamical heating processes. In Figure \[fig:temporal\_vzdisp\], we show the temporal evolution of [$\sigma_{\mathrm{v_z}}$]{}for a series of cohorts with 1-Gyr formation time bins. To compare with Figure \[fig:vrot\_vzdisp\], we further restrict our analysis to likely disk stars (see figure caption for selection criteria). The initial [$\sigma_{\mathrm{v_z}}$]{} of each cohort closely follows the properties of the star-forming gas. The plateau seen in Figure \[fig:vrot\_vzdisp\] from $t\approx3$-$5$ Gyr is echoed by the similar initial [$\sigma_{\mathrm{v_z}}$]{} of [$\mathcal{S}_{2.0,3.0}$]{}, [$\mathcal{S}_{3.0,4.0}$]{}, and [$\mathcal{S}_{4.0,5.0}$]{}. Similarly, initial [$\sigma_{\mathrm{v_z}}$]{} rapidly decreases from [$\mathcal{S}_{4.0,5.0}$]{} to [$\mathcal{S}_{6.0,7.0}$]{} as the gas reservoir contracts following the merger epoch, shows little change during the last minor merger (compare [$\mathcal{S}_{6.0,7.0}$]{} and [$\mathcal{S}_{7.0,8.0}$]{}), and slowly declines thereafter. For all cohorts, the heating rate ($\delta{\ensuremath{\sigma_{\mathrm{v_z}}}}/\delta t$, the slope of each line) measured over any one Gyr time interval ($t_0,t_1$) is anti-correlated with [$\sigma_{\mathrm{v_z}}$]{} at $t_0$, confirming that both mergers and secular scattering processes are more efficient heating mechanisms within colder stellar populations. Rapid increases in [$\sigma_{\mathrm{v_z}}$]{}($>5$ [km s$^{-1}$]{}Gyr$^{-1}$) are temporally coincident with merger activity, suggesting that the precise $z=0$ stellar AVR may correlate with the galaxy’s accretion history. Secular heating does affect the AVR, increasing the dispersion of all cohorts over time, but it is clear from Figure \[fig:temporal\_vzdisp\] that the overall trend of [$\sigma_{\mathrm{v_z}}$]{} with age is dominated by differences in initial dispersion at birth rather than by differential heating. There is remarkable qualitative similarity between the age-kinematics trends measured in the $z=0$ simulated galaxy and those observed in the MW. Cohort age is positively correlated with the population’s vertical extent but anti-correlated with its radial extent (Figure \[fig:400\_den\]). Recent MW studies, using chemical composition as a proxy for stellar age, suggest similar structure in our own Galaxy [[[e.g.]{}]{}, @Bovy12a]. The alpha-enhanced, old, thick disk has a shorter scale length than the alpha-poor, young, thin disk [@Cheng12b]. Stars far from the plane ($|z|>1$ kpc) have similar [$[\mathrm{Fe/H}]$]{} over a broad range in radius [@Cheng12a], suggesting that these stars may share a temporal origin. Stellar composition gradients in the vertical direction are even more well established; observations in the solar neighborhood and beyond find that [$[\alpha\mathrm{/Fe}]$]{} increases and [$[\mathrm{Fe/H}]$]{} decreases with height above the disk plane [[[e.g.]{}]{}, @Bensby03; @Ruchti10; @Lee11; @Schlesinger12]. In addition to these geometrical differences, older cohorts in the simulated galaxy form with higher random motions[^4] (Figure \[fig:400\_rp\]). This general age-velocity relationship (AVR) exists in the MW as well [[[e.g.]{}]{}, @Freeman08], and it is often explained by older stars having more time to be heated by various perturbing events ([[e.g.]{}]{}, molecular clouds, transient waves, satellites). Here we find that while the disk-forming cohorts show a modest increase in their random motions with time, older cohorts *are born* kinematically hotter than younger stars. Our analysis of the mono-age populations makes it clear that the galaxy’s violent youth and [*in situ*]{} star-formation are responsible for the positively-sloped AVR, in addition to the age-density gradients present in both the radial and vertical dimensions. ![\[fig:temporal\_vzdisp\] The temporal evolution of the vertical velocity dispersion for individual cohorts. For this figure, we select likely disk orbits ([$\epsilon$]{}$>0.5$) that are spatially coincident with the disk ($4<{\ensuremath{r_\mathrm{gc}}}<8$ kpc and $|z|<2$ kpc) to directly compare with Figure \[fig:vrot\_vzdisp\]. Color-coding differentiates the age cohorts and refers to their formation time; line types help distinguish amongst similar colors (see legend). ](figs/temporal_sigvz_circcut.eps){width="3.7in"} {width="\textwidth"} The present-day age cohorts show near one to one correspondence with the mono-abundance populations of @Bovy12a [@Bovy12c], warranting a more direct comparison. As SEGUE’s G-dwarf sample primarily probes disk populations [@Bovy12a], we identify current disk members of the simulated galaxy and examine their kinematic properties as a function of age (Figure \[fig:circ\_cut\], see caption for selection criteria). Progressively older cohorts have shorter scale lengths (steeper slopes in left panel of Figure \[fig:circ\_cut\]) and larger vertical extents (middle panel); both monotonic trends are remarkably smooth over the G-dwarf sample’s radial and presumed age range ($6\leq{\ensuremath{r_\mathrm{gc}}}\leq12$ kpc and $\tau<10$ Gyr, respectively). [@Bovy12a] find that each MW mono-abundance population can be globally fit by a single exponential in the radial and vertical directions, but in the simulated galaxy each age cohort’s vertical density profile is itself a function of radius (Figure \[fig:circ\_cut\], middle panel). The observed mono-abundance populations show trends between vertical kinematics and age proxy: more alpha-rich, iron-poor populations have larger vertical velocity dispersions than their alpha-poor, iron-rich counterparts [@Bovy12c]. The equivalent trend is found in the simulated galaxy, where progressively older cohorts have larger [$\sigma_{\mathrm{v_z}}$]{} (Figure \[fig:circ\_cut\], right panel). @Bovy12c find a shallow, but negative, radial gradient in [$\sigma_{\mathrm{v_z}}$]{} for all mono-abundance populations. Both [$\mathcal{S}_{2.0,4.0}$]{} and [$\mathcal{S}_{4.0,8.0}$]{} show a decrease in [$\sigma_{\mathrm{v_z}}$]{} with radius, but the vertical velocity dispersion of stars born in the last $5$ Gyr ([$\mathcal{S}_{8.0,12.8}$]{} and [$\mathcal{S}_{12.8,13.4}$]{}) slightly increases with radius. These results suggest that age is a good proxy for chemical composition, a notion independently confirmed in a recent study [@Stinson13]. @Bovy12b [@Bovy12a; @Bovy12c] favor internal, secular evolution mechanisms rather than discrete, external heating events to explain the smooth evolution of MW kinematic properties as a function of chemical abundance. Our simulated galaxy shows analogously smooth correlations of stellar kinematics with age, and the trends are largely imprinted at birth or immediately thereafter. Thick disks are thought to be fundamental in understanding the balance of external and internal influences in disk galaxy formation (Section \[sec:intro\]). @Bovy12b showed that the mono-abundance populations span a continuous range of scale heights at the solar neighborhood and thereby argued that the MW does not have a distinct thick disk component. @Rix13 find that integrated vertical space density of all mono-abundance populations masquerades as observations classically identifying the thin and thick disk. At the solar annulus in the simulated galaxy, the vertical mass density profiles of individual age cohorts are progressively steeper for younger populations (Figure \[fig:400\_vp\], middle panel). The superposition of all age cohorts in the solar annulus results in the familiar double-exponential profile observed in MW star counts [[[e.g.]{}]{}, @Gilmore83; @Juric_etal08]. The two exponentially-fitted components are not distinct in origin; for example, the [$\mathcal{S}_{2.0,4.0}$]{} and [$\mathcal{S}_{4.0,8.0}$]{} cohorts populate both the thin and thick components at the solar annulus and vary their fractional contributions to these structures as a function of radius (Figure \[fig:400\_vp\]). Overall, our simulation leads to a “continuous” view of the thick disk similar to that advocated by @Bovy12b. We note that radial migration can produce a similar vertical structure [@Schonrich09b], but in our simulation the double-exponential emerges principally from the superposition of trends imprinted by upside-down and inside-out formation. The mechanisms primarily responsible for the main components of galactic structure (determined by kinematic decomposition, Table \[tab:decomp\]) can be inferred from the histories of corresponding member cohorts. The pseudobulge, predominantly made of stars with [$t_\mathrm{form}$]{}$<4$ Gyr, formed [*in situ*]{} and at early times (Figures \[fig:tf0.5\_den\] and  \[fig:tf1.0\_den\]; for a full investigation of bulge formation in a similar cosmological simulation, see @Guedes12). Over half the mass in the spheroid was born just prior to or during the major merger epoch ($t=0.5$-$2.0$ Gyr), and another $25\%$ formed immediately thereafter ([$\mathcal{S}_{2.0,4.0}$]{}), suggesting stellar halo formation by stars scattered during the merger epoch, with a moderate contribution of stars stripped from satellites. The kinematically defined thick disk begins to form just before the end of the major merger epoch ([$\mathcal{S}_{1.0,2.0}$]{} accounts for $25\%$ of the thick disk’s mass at $z=0$). Still, the majority of the thick disk forms in the parent halo after the merger epoch ends; over half the thick disk’s mass is in [$\mathcal{S}_{2.0,4.0}$]{}, which has a strong [*in situ*]{} component (Figure \[fig:tf3.0\_rp\]). MW observations may call for a significant contribution of $t_{\rm form}=4-6$ Gyr stars in a chemically defined thick disk [@Bensby04], and in our simulation such stars are formed overwhelmingly in the parent halo. Even as thick disk formation is well underway, stars with thin disk kinematics begin to appear. $26\%$ of the stars on cold, thin disk orbits at $z=0$ are members of [$\mathcal{S}_{2.0,4.0}$]{}, but many of these stars would likely be members of a chemically defined thick disk. The fraction of new stars that are kinematically associated with the thin disk increases with time in all subsequently formed cohorts. Kinematic vs. chemical definitions of disk components likely involve significant mixing of the thin and thick populations [[[e.g.]{}]{}, @Navarro11; @Lee11]. However, our kinematics-based results are broadly consistent with the chemistry-based results of [@Brook12], who study a cosmological zoom-in simulation of a disk galaxy that is $< 1/5$ the virial mass of the MW. They investigate the history of the gas reservoirs eventually forming the chemically-selected thin disk, thick disk and halo to describe galactic component formation in the annulus ${\ensuremath{r_\mathrm{gc}}}=7$-8 kpc. Their halo stars predominantly form [*in situ*]{} and are scattered to the halo during an early merger epoch, with a smaller fraction directly accreted from satellite galaxies. After the major merger epoch in their simulation ends, half of the thick disk forms from hot gas that was already in the central galaxy during the merger epoch. @Brook12 report a smooth transition in star formation from thick to thin disk stars, with the majority of their thin disk stars forming from smoothly accreted gas after the merger epoch. @Brook12 also find that old stars must move away from their formation radii to populate the solar vicinity. Our analysis in §\[sec:radmix\_tf\] leads to a similar conclusion and differentiates between kinematically hot orbits and guiding center changes (discussed further below). Our simulated galaxy and that of @Brook12 form their disk components in qualitatively similar fashions despite nearly an order-of-magnitude difference in galaxy mass and many differences of detail in the simulations. Both experiments have relatively quiescent merger histories [^5] by construction, so we cannot say whether more active merger histories would substantially change the picture and/or lead to disagreement with observed MW structure. Stars radially migrate throughout our simulation, and the redistribution of these stars impacts the galaxy’s construction to an extent. For simplicity, we focus on the role of radial migration in the creation of the thick disk at the solar annulus, which has been a topic of debate in the recent literature [[[e.g.]{}]{}, @Schonrich09; @Sales09; @Loebman11; @Brook12; @Minchev12a; @Roskar12b]. The basic idea is that the hotter populations of the inner galaxy will increase their scale heights as they move to the weaker vertical potential of the outer disk, but there are complications because of adiabatic cooling and dynamical selection of the stars that migrate. We now compare the kinematics of outwardly migrating (${\ensuremath{\Delta R_\mathrm{g}}}>1.0$ kpc, designated “om”) and non-migrating ($|{\ensuremath{\Delta R_\mathrm{g}}}|<0.5$ kpc, “non”) disk populations [^6] in the solar annulus. When all disk stars currently residing at $7<{\ensuremath{r_\mathrm{gc}}}<9$ kpc are considered, outwardly migrating stars and non-migrators have virtually the same vertical velocity dispersion ($\sigma_{v_{\mathrm{z,om}}}=29.7$ [km s$^{-1}$]{}vs. $\sigma_{v_{\mathrm{z,non}}}=29.5$ [km s$^{-1}$]{}). These results are consistent with those of @Minchev12a, who find that outwardly migrating stars regulate themselves such that they are only marginally ($<5\%$ level) hotter than the stellar populations at their destination radius, suggesting that migrators do not “thicken” the disk. A more complex picture emerges when we further dissect the sample and examine two older, thick-disk contributing cohorts ([$\mathcal{S}_{2.0,4.0}$]{} and [$\mathcal{S}_{4.0,6.0}$]{}). In [$\mathcal{S}_{2.0,4.0}$]{}, non-migrators have more vertical energy ($\sigma_{v_{\mathrm{z,non}}}=46.6$ [km s$^{-1}$]{}) than their outwardly migrating counterparts ($\sigma_{v_{\mathrm{z,om}}}=37.0$ [km s$^{-1}$]{}). [$\mathcal{S}_{4.0,6.0}$]{} shows the same relationship between migration and approximate vertical energy: non-migrators are hotter than outward migrators ($\sigma_{v_{\mathrm{z,non}}}=36.5$ [km s$^{-1}$]{} and $\sigma_{v_{\mathrm{z,om}}}=29.1$ [km s$^{-1}$]{}, respectively). The relationship between horizontal and vertical motions in an orbit is well-described using vertical action ($J_z$) invariance [[[e.g.]{}]{} @Binney10; @Binney11; @Schonrich12a], and simulations show that migrating populations conserve their vertical action on average [@Solway12]. Within an idealized, one-dimensional disk galaxy, the vertical velocity dispersion of an outwardly migrating population with constant $J_z$ scales with the surface density of the disk, [[i.e.]{}]{}, $\sigma_{v_{\mathrm{z,om}}} \propto\Sigma^{1/4}$ (assuming $\Sigma\propto e^{- {\ensuremath{r_\mathrm{gc}}}/\mathrm{r}_d}$ where r$_d$ is the scale length of the disk; see @Minchev12a or @Roskar12b for derivation in the 1D case, @Schonrich12a for a more general treatment). In the same model galaxy, the vertical velocity dispersion of an isothermal, non-migrating population also scales with disk surface density but has an additional dependence on the scale height of the population as a function of radius ($\mathrm{h}_z$), [[i.e.]{}]{}, $\sigma_{v_{\mathrm{z,non}}} \propto\sqrt{\Sigma\mathrm{h}_z}$ [[[e.g.]{}]{}, @Kruit11]. Comparing the radial dependence of [$\sigma_{\mathrm{v_z}}$]{}for migrating and non-migrating populations, we find $$\frac{{\ensuremath{\sigma_{\mathrm{v}_{z,om}}}}}{{\ensuremath{\sigma_{\mathrm{v}_{z,non}}}}}\propto \frac{\Sigma^{1/4}}{\sqrt{\Sigma\mathrm{h}_z}} \label{eq:vz}$$ . This 1D scaling relationship cannot be used to make quantitative predictions of [$\sigma_{\mathrm{v}_{z,om}}$]{}$/$[$\sigma_{\mathrm{v}_{z,non}}$]{}, but it suggests that non-migrating stars are likely to have more vertical energy than outward migrators at a given radius when the disk scale length is short or h$_z$ is a strong, positively-sloped function of radius. The simulated galaxy presents a more complicated, multi-component scenario, but our measurements within the solar annulus disk sample (previous paragraph) follow the relationships established in equation \[eq:vz\]. There, [$\sigma_{\mathrm{v}_{z,non}}$]{} and [$\sigma_{\mathrm{v}_{z,om}}$]{} are nearly the same for the ensemble population, but [$\sigma_{\mathrm{v}_{z,om}}$]{}$/$[$\sigma_{\mathrm{v}_{z,non}}$]{} decreases for [$\mathcal{S}_{2.0,4.0}$]{} and [$\mathcal{S}_{4.0,6.0}$]{}, both of which have relatively short scale lengths and steep $\mathrm{h}_z$[^7] (Figure \[fig:circ\_cut\]). The trend of scale height with radius may therefore play an important role in the relative influence of migrators and non-migrators at large scale heights in MW-like galaxies. The simulated disk sample shows increasing median height ([$z_\mathrm{med}$]{}) with radius as $\frac{\Delta{\ensuremath{z_\mathrm{med}}}}{\Delta{\ensuremath{r_\mathrm{gc}}}}=0.07$ over the radial range $5<{\ensuremath{r_\mathrm{gc}}}<14$ kpc ($71$ pc per kpc; middle panel, black line, Figure \[fig:circ\_cut\]). Radial migration may contribute to this radially increasing scale height. We find that migrating stellar populations become more vertically extended as they move outward from the inner disk [in agreement with @Roskar12b], and some old, migratory groups currently have velocity dispersions exceeding the typical dispersion of thick disk stars in the MW [${\ensuremath{\sigma_{\mathrm{v_z}}}}=35$[km s$^{-1}$]{}, @Bensby03]. These old stars populate both the traditional thin and thick disk components at their destination radius, so they are important when discussing population demographics and chemodynamic trends. The radial dependence of $\mathrm{h}_z$ observed here is contrary to the traditional notion of constant scale height in disk galaxies [[[e.g.]{}]{}, @Kruit82], but radially increasing stellar scale heights have been found in more recent measurements of a large sample of nearby disk galaxies [@deGrijs97; $\frac{\Delta\mathrm{h}_z}{\Delta{\ensuremath{r_\mathrm{gc}}}}=0.05$ for MW Hubble type] and in the MW itself [@Narayan02; $\frac{\Delta\mathrm{h}_z}{\Delta{\ensuremath{r_\mathrm{gc}}}}=0.02$]. Simulation improvements, including incorporating missing ISM physics such as molecular cooling, will likely flatten $\mathrm{h}_z$ and decrease the tension between simulations and observations. Individual age cohorts may still show steep $\mathrm{h}_z$ within the disk because of heating events like mergers, as even minor perturbations can cause flaring due to the low restorative force in the outer disk [[[e.g.]{}]{}, @Kazantzidis08; @Bournaud09]. Old, thick disk contributing cohorts ([[e.g.]{}]{}, [$\mathcal{S}_{2.0,6.0}$]{}) in MW-sized halos are particularly susceptible to flaring: these stars are typically born during a significant merger period (Section \[sec:evolution\]) and non-migrators at the solar annulus would form at the edge of a young, relatively compact disk . If this scenario is generic to older age cohorts in $M_\star$ halos, the maximal vertical extent of the thick disk (and similar structures in external galaxies) is likely set by the initially hot, [*in situ*]{} stars, as measured in the simulated galaxy. Conclusions {#sec:summary_tf} =========== We have investigated the structure, kinematics, and evolutionary history of stellar age cohorts in one of the highest resolution cosmological simulations ever run of the formation of a MW-like galaxy. This simulation produces good (though not perfect) agreement with many observed properties of the MW and similar galaxies [@Guedes11], so it may provide useful guidance to the origin of trends found in the MW. We find remarkably good qualitative agreement between the trends for our mono-age cohorts and the observed trends for mono-abundance stellar populations in the MW found by @Bovy12b [@Bovy12a; @Bovy12c]. Within the disk, older cohorts have shorter radial scale lengths, larger vertical scale heights, and hotter kinematics. The trends with age are smooth, but at a given radius the superposition of old, vertically-extended populations and young, vertically-compact populations gives rise to a total vertical profile that has the double-exponential form usually taken to represent the superposition of “thin” and “thick” disks in the MW. Tracing back the formation history of each age cohort, we find that these present-day trends of spatial structure and kinematics are largely imprinted at birth. The oldest stars ([$\mathcal{S}_{0.0,0.5}$]{}, [$\mathcal{S}_{0.5,1.0}$]{}, and, to a lesser extent, [$\mathcal{S}_{1.0,2.0}$]{}) form during the active merging phase of the galaxy’s assembly, and these mergers scatter stars into rounded configurations and kinematically hot orbits, with the oldest stars being the most centrally concentrated. Even in these early cohorts, the great majority of stars form in the main halo rather than satellites, and phase-mixing erases differences in their distributions in any case. Younger cohorts form in disk-like configurations that are progressively more extended, thinner, and colder as $t_{\rm form}$ increases. The larger scale heights at earlier times are mostly inherited from the star-forming gas disk, which has a greater degree of turbulent support relative to rotational support (Figure \[fig:vrot\_vzdisp\]), presumably because of stronger SN feedback from more vigorous star formation and more dynamical stirring by mergers and halo substructure. A disk cohort typically shows additional heating for another $1-2$ Gyr after $t_{\rm form}$, but thereafter the growth of velocity dispersion or vertical thickness is minimal. The “inside-out” radial growth of the disk is a long-standing theoretical expectation — even if the gas disk were fully present at $t=0$, star-formation would proceed more rapidly in the higher surface density, central regions. The resulting shape of the radial abundance gradient early in a galaxy’s evolution can therefore directly constrain feedback models [@Pilkington12; @Gibson13]. At redshift zero, inside-out growth readily explains the direction of radial abundance gradients measured both locally and in external disk galaxies and is a central component of many Galactic chemical evolution models [[[e.g.]{}]{} @Chiappini97; @Prantzos00]. Radial migration of stars is also an important ingredient in understanding chemical evolution [@Wielen96; @Sellwood02; @Roskar08b; @Schonrich09], and it arises naturally in our simulation. In the outer disk, progressively older stars have progressively smaller formation radii on average, though the distribution of $r_{\rm form}$ is usually broad and includes some old stars born at large radius. The appearance of old stars at $r_{\rm gc} > r_{\rm form}$ is partly a consequence of increases in angular momentum (and thus guiding center radius) and partly a consequence of the larger orbital eccentricities of older stellar populations. The “upside-down” evolution that we find for the disk’s vertical structure in our simulation is more novel, and it highlights a view of the MW’s thick disk that has growing support from analytic and numerical galaxy formation studies. In our simulation, the thick disk arises from continuous trends between stellar age and the evolving structure of the star-forming disk, not from a discrete merger event or from secular heating or radial migration after formation. The simulation clearly reproduces the basic phenomenology of the MW’s thin and thick disks: a double-exponential vertical profile, and an increasing fraction of old stars at large distance from the plane and/or large random velocity. @Forbes12 have proposed a similar model for the origin of the thin/thick transition, based on one-dimensional simulations that track the evolution of turbulence in the star-forming gas. “Upside-down” formation is also observed to varying degrees in gas-rich merger simulations [@Brook04], zoom-in cosmological simulations [@Stinson13], and in clumpy, unstable gas-rich disks [@Bournaud09]. The disk heating study of @House11 found, much like our study, that upside-down formation can naturally lead to a smooth stellar AVR at redshift zero. Quantitative tests of the simulation predictions will require using chemistry in place of age and carefully matching the selection criteria and measurement uncertainties of observational samples, a task that we reserve to future work. Any simulation of galaxy formation has numerical limitations and physical uncertainties associated with the treatment of star formation and feedback. The specifics of our star formation treatment may matter for some of our conclusions, as we find that the thickness of the stellar disk is largely inherited from that of the star-forming gas. The Eris simulations have many successes in reproducing disk galaxy properties at high redshift [@Shen12] and low redshift [@Guedes11], but the adopted star formation model may predict galaxies that are too luminous for their halo mass relative to abundance matching constraints [@Moster13; @Munshi13]. Solving this problem may require simulations with more vigorous feedback, and it will be important to confirm that the trends found here persist in such simulations. Furthermore, this simulation is a single realization selected to have a quiescent merger history, which limits our ability to draw statistical conclusions. However, we find good qualitative agreement with the simulation of [@Brook12] despite a roughly order-of-magnitude difference in galaxy mass, which suggests that our conclusions are fairly robust, and upside-down formation also arises in the one-dimensional calculations of [@Forbes12] as well as a small but diverse set of numerical experiments [[[e.g.]{}]{} @Brook04; @Bournaud09; @House11; @Stinson13]. The good agreement between the properties of mono-age populations and observed trends in the MW further suggests that this recipe for galaxy formation has most of the right ingredients. ELS envisioned a monolithic collapse of the proto-Galaxy, a scenario at odds with the hierarchical formation predicted in the CDM model. Our simulated galaxy forms self-consistently via violent, hierarchical accretion and cold flows at early times [@Shen12] followed by a nearly quiescent phase in which the disk grows predominantly from cold flow accretion and, to a lesser extent, from cooling of a diffuse warm/hot corona [@Brooks09; @Guedes11]. Nonetheless, there are some qualitative similarities between our simulation results and the ELS scenario: the bulge [@Guedes12] and the thin and thick disks (this paper) form the overwhelming majority of their stars [*in situ*]{}, and these stars form in progressively more flattened, more rotationally supported configurations. The global distributions and correlations of galaxy properties — luminosities, colors, rotation speeds, morphologies, etc. — provide critical tests of galaxy formation simulations in the $\Lambda$CDM cosmology. As the realism and resolution of simulations continues to improve, and as chemodynamical surveys of the MW grow in scope and detail, the close-to-home tests of simulation predictions against the observed correlations of the age, composition, spatial structure, and kinematics of stellar populations will provide steadily more powerful constraints on the theory of galaxy formation. We thank Ralph Schönrich and the referee, Brad Gibson, for detailed comments on the original manuscript that led to significant improvements in the paper. We acknowledge fruitful conversations with Tom Quinn. We made use of pynbody (http://code.google.com/p/pynbody) in our analysis for this paper. J.B. acknowledges the support of the Vanderbilt Office of the Provost through the Vanderbilt Initiative in Data-intensive Astrophysics (VIDA) and the Distinguished University Fellowship awarded by the Ohio State University. S.K. is supported by the Center for Cosmology and Astro-Particle Physics at The Ohio State University. This research was partially funded by the Marie Curie Actions for People COFUND Program through the ETH Zurich Postdoctoral Fellowship awarded to J.G. Support for this work was provided by the NSF through grants AST-0908910, AST-1009505, AST-1211853, and OIA-1124453, and by NASA through grant NNX12AF87G. We acknowledge the Ohio Supercomputer Center (http://www.osc.edu) for computing time. [102]{} natexlab\#1[\#1]{} , M. G., [Navarro]{}, J. F., [Steinmetz]{}, M., & [Eke]{}, V. R. 2003, , 597, 21 , M., [White]{}, S., [Naab]{}, T., & [Scannapieco]{}, C. 2013, ArXiv e-prints , M. R., & [Burkert]{}, A. 1997, , 288, 1060 , T., [Feltzing]{}, S., & [Lundstr[ö]{}m]{}, I. 2003, , 410, 527 —. 2004, , 421, 969 , T., [Feltzing]{}, S., [Lundstr[ö]{}m]{}, I., & [Ilyin]{}, I. 2005, , 433, 185 , J. 2010, , 401, 2318 , J., & [McMillan]{}, P. 2011, , 413, 1889 , J., & [Tremaine]{}, S. 2008, [Galactic Dynamics: Second Edition]{}, ed. J. [Ostriker]{} & D. [Spergel]{} (Princeton University Press) , J. C., [Kazantzidis]{}, S., & [Weinberg]{}, D. H. 2012, , 420, 913 , F., [Elmegreen]{}, B. G., & [Martig]{}, M. 2009, , 707, L1 , J., [Rix]{}, H.-W., & [Hogg]{}, D. W. 2012a, , 751, 131 , J., [Rix]{}, H.-W., [Hogg]{}, D. W., [et al.]{} 2012c, , 755, 115 , J., [Rix]{}, H.-W., [Liu]{}, C., [et al.]{} 2012b, , 753, 148 , C. B., [Kawata]{}, D., [Gibson]{}, B. K., & [Freeman]{}, K. C. 2004, , 612, 894 , C. B., [Stinson]{}, G. S., [Gibson]{}, B. K., [et al.]{} 2012, , 426, 690 , A. M., [Governato]{}, F., [Quinn]{}, T., [Brook]{}, C. B., & [Wadsley]{}, J. 2009, , 694, 396 , J. S., & [Johnston]{}, K. V. 2005, , 635, 931 , J. S., [Kravtsov]{}, A. V., & [Weinberg]{}, D. H. 2001, , 548, 33 , L., [Sch[ö]{}nrich]{}, R., [Asplund]{}, M., [et al.]{} 2011, , 530, A138 , J. Y., [Rockosi]{}, C. M., [Morrison]{}, H. L., [et al.]{} 2012a, , 746, 149 —. 2012b, , 752, 51 , C., [Matteucci]{}, F., & [Gratton]{}, R. 1997, , 477, 765 , S., [Aragon-Salamanca]{}, A., [Frenk]{}, C. S., [Navarro]{}, J. F., & [Zepf]{}, S. E. 1994, , 271, 781 , R., & [Peletier]{}, R. F. 1997, , 320, L21 , V. P., [Carollo]{}, C. M., [Mayer]{}, L., & [Moore]{}, B. 2004, , 604, L93 —. 2005, , 628, 678 , O. J., [Lynden-Bell]{}, D., & [Sandage]{}, A. R. 1962, , 136, 748 , S. M., & [Efstathiou]{}, G. 1980, , 193, 189 , J., [Krumholz]{}, M., & [Burkert]{}, A. 2012, , 754, 48 , K. C. 2008, in Astronomical Society of the Pacific Conference Series, Vol. 396, Formation and Evolution of Galaxy Disks, ed. J. G. [Funes]{} & E. M. [Corsini]{}, 3 , B. K., [Pilkington]{}, K., [Brook]{}, C. B., [Stinson]{}, G. S., & [Bailin]{}, J. 2013, , in press , G., & [Reid]{}, N. 1983, , 202, 1025 , F., [Willman]{}, B., [Mayer]{}, L., [et al.]{} 2007, , 374, 1479 , J., [Callegari]{}, S., [Madau]{}, P., & [Mayer]{}, L. 2011, , 742, 76 , J., [Mayer]{}, L., [Carollo]{}, M., & [Madau]{}, P. 2012, ArXiv e-prints , F., & [Madau]{}, P. 1996, , 461, 20 , M. 2008, , 388, 1175 , J., [Nordstr[ö]{}m]{}, B., & [Andersen]{}, J. 2007, , 475, 519 , P. F., [Bundy]{}, K., [Croton]{}, D., [et al.]{} 2010, , 715, 202 , E. L., [Brook]{}, C. B., [Gibson]{}, B. K., [et al.]{} 2011, , 415, 2652 , M., [et al.]{} 2008, , 673, 864 , N. 1992, , 391, 502 , G., [White]{}, S. D. M., & [Guiderdoni]{}, B. 1993, , 264, 201 , S., [Bullock]{}, J. S., [Zentner]{}, A. R., [Kravtsov]{}, A. V., & [Moustakas]{}, L. A. 2008, , 688, 254 , S., [Zentner]{}, A. R., [Kravtsov]{}, A. V., [Bullock]{}, J. S., & [Debattista]{}, V. P. 2009, , 700, 1896 , P., [Tout]{}, C. A., & [Gilmore]{}, G. 1993, , 262, 545 , Y. S., [Beers]{}, T. C., [An]{}, D., [et al.]{} 2011, , 738, 187 , S. R., [Ro[š]{}kar]{}, R., [Debattista]{}, V. P., [et al.]{} 2011, , 737, 8 , S., [Couchman]{}, H. M. P., & [Wadsley]{}, J. 2006, , 442, 539 , F., & [Francois]{}, P. 1989, , 239, 885 , L. 2012, in Astronomical Society of the Pacific Conference Series, Vol. 453, Advances in Computational Astrophysics: Methods, Tools, and Outcome, ed. R. [Capuzzo-Dolcetta]{}, M. [Limongi]{}, & A. [Tornamb[è]{}]{}, 289 , C. F., & [Ostriker]{}, J. P. 1977, , 218, 148 , I., [Chiappini]{}, C., & [Martig]{}, M. 2012b, ArXiv e-prints , I., & [Famaey]{}, B. 2010, , 722, 112 , I., [Famaey]{}, B., [Quillen]{}, A. C., [et al.]{} 2012a, , 548, A127 , H. J., [Mao]{}, S., & [White]{}, S. D. M. 1998, , 295, 319 , B. P., [Naab]{}, T., & [White]{}, S. D. M. 2013, , 428, 3121 , F., [Governato]{}, F., [Brooks]{}, A. M., [et al.]{} 2013, , 766, 56 , C. A., & [Jog]{}, C. J. 2002, , 394, 89 , J. F., [Abadi]{}, M. G., [Venn]{}, K. A., [Freeman]{}, K. C., & [Anguiano]{}, B. 2011, , 412, 1203 , J. F., & [Steinmetz]{}, M. 1997, , 478, 13 , J., & [Nemec]{}, A. F. L. 1991, , 103, 95 , J. M., & [Nemec]{}, A. F. L. 1993, , 105, 1455 , K., [Gibson]{}, B. K., [Calura]{}, F., [et al.]{} 2011, , 417, 2891 , K., [Few]{}, C. G., [Gibson]{}, B. K., [et al.]{} 2012, , 540, A56 , N., & [Boissier]{}, S. 2000, , 313, 338 , A. C., & [Garnett]{}, D. R. 2001, in Astronomical Society of the Pacific Conference Series, Vol. 230, Galaxy Disks and Disk Galaxies, ed. J. G. [Funes]{} & E. M. [Corsini]{}, 87–88 , A. C., [Minchev]{}, I., [Bland-Hawthorn]{}, J., & [Haywood]{}, M. 2009, , 397, 1599 , D. J., [Ro[š]{}kar]{}, R., [Debattista]{}, V. P., [et al.]{} 2012, , 753, 138 , H.-W., & [Bovy]{}, J. 2013, , 21, 61 , R., [Debattista]{}, V. P., & [Loebman]{}, S. R. 2012b, ArXiv e-prints , R., [Debattista]{}, V. P., [Quinn]{}, T. R., [Stinson]{}, G. S., & [Wadsley]{}, J. 2008, , 684, L79 , R., [Debattista]{}, V. P., [Quinn]{}, T. R., & [Wadsley]{}, J. 2012a, , 426, 2089 , R., [Debattista]{}, V. P., [Stinson]{}, G. S., [et al.]{} 2008, , 675, L65 , G. R., [Fulbright]{}, J. P., [Wyse]{}, R. F. G., [et al.]{} 2010, , 721, L92 , L. V., [Helmi]{}, A., [Abadi]{}, M. G., [et al.]{} 2009, , 400, L61 , P., [Courty]{}, S., [Gibson]{}, B. K., & [Brook]{}, C. B. 2009, , 398, 591 , K. J., [Johnson]{}, J. A., [Rockosi]{}, C. M., [et al.]{} 2012, , 761, 160 , R., & [Binney]{}, J. 2009a, , 396, 203 —. 2009b, , 399, 1145 —. 2012, , 419, 1546 , L., & [Zinn]{}, R. 1978, , 225, 357 , J. A., & [Binney]{}, J. J. 2002, , 336, 785, (SB02) , S., [Madau]{}, P., [Aguirre]{}, A., [et al.]{} 2012, , 760, 50 , S., [Wadsley]{}, J., & [Stinson]{}, G. 2010, , 407, 1581 , M., [Sellwood]{}, J. A., & [Sch[ö]{}nrich]{}, R. 2012, , 422, 1363 , C., [Bienaym[é]{}]{}, O., [Mishenina]{}, T. V., & [Kovtyukh]{}, V. V. 2008, , 480, 91 , D. N., [Bean]{}, R., [Dor[é]{}]{}, O., [et al.]{} 2007, , 170, 377 , G., [Seth]{}, A., [Katz]{}, N., [et al.]{} 2006, , 373, 1074 , G. S., [Bovy]{}, J., [Rix]{}, H.-W., [et al.]{} 2013, ArXiv e-prints , P. C., & [Freeman]{}, K. C. 2011, , 49, 301 , P. C., & [Searle]{}, L. 1982, , 110, 61 , [Á]{}., & [Helmi]{}, A. 2008, , 391, 1806 , J. W., [Stadel]{}, J., & [Quinn]{}, T. 2004, New Astronomy, 9, 137 , S. D. M., & [Frenk]{}, C. S. 1991, , 379, 52 , S. D. M., & [Rees]{}, M. J. 1978, , 183, 341 , R. 1977, , 60, 263 , R., [Fuchs]{}, B., & [Dettbarn]{}, C. 1996, , 314, 438 , B., [Rockosi]{}, C., [Newberg]{}, H. J., [et al.]{} 2009, , 137, 4377 , P., & [Dalcanton]{}, J. J. 2008, , 682, 1004 , P., [Ro[š]{}kar]{}, R., & [Debattista]{}, V. P. 2012, , 752, 97 [^1]: The bin edge is at $13.4$ Gyr as it is the nearest $100$ Myr increment that was inclusive of all stars. [^2]: Relatively few particles have heights $z>4$ kpc; our main results are independent of this selection cut. [^3]: @Schonrich09 refer to the latter process as “churning” and the former as “blurring.” [^4]: The trend flattens for stars with [$t_\mathrm{form}$]{}$<2$ Gyr. [^5]: The last major merger in @Brook12 occurs at $z=2.7$. [^6]: Disk stars are identified according to the criteria used for Figure \[fig:circ\_cut\]. [^7]: The slope of $\mathrm{h}_z$ and the median height as a function of radius are equal if the if the vertical density distribution is exponential.

--- abstract: 'The astrophysical factor for the proton weak capture on $^3$He is calculated with correlated-hyperspherical-harmonics bound and continuum wave functions corresponding to a realistic Hamiltonian consisting of the Argonne $v_{18}$ two-nucleon and Urbana-IX three-nucleon interactions. The nuclear weak charge and current operators have vector and axial-vector components, that include one- and many-body terms. All possible multipole transitions connecting any of the $p\,^3$He S- and P-wave channels to the $^4$He bound state are considered. The $S$-factor at a $p\,^3$He center-of-mass energy of $10$ keV, close to the Gamow-peak energy, is predicted to be $10.1 \times 10^{-20}$ keV b, a factor of five larger than the standard-solar-model value. The P-wave transitions are found to be important, contributing about 40 % of the calculated $S$-factor.' address: - '$^{({\rm a})}$Department of Physics, Old Dominion University, Norfolk, VA 23529' - '$^{({\rm b})}$Jefferson Lab, Newport News, VA 23606' - '$^{({\rm c})}$INFN, Sezione di Pisa, I-56100 Pisa, Italy' - '$^{({\rm d})}$Department of Physics, University of Pisa, I-56100 Pisa, Italy' author: - 'L.E. Marcucci$^{({\rm a})}$, R. Schiavilla$^{({\rm a,b})}$, M. Viviani$^{({\rm c})}$, A. Kievsky$^{({\rm c})}$, S. Rosati$^{({\rm c,d})}$' title: Realistic Calculation of the $hep$ Astrophysical Factor --- Recently, there has been a revival of interest in the reaction $^3$He($p$,$e^+ \nu_e$)$^4$He [@BK98]. This interest has been spurred by the Super-Kamiokande collaboration measurements of the energy spectrum of electrons recoiling from scattering with solar neutrinos [@Fuk99]. At energies larger than 14 MeV more recoil electrons have been observed than expected on the basis of standard-solar-model (SSM) predictions [@BBP98]. The $hep$ process, as the proton weak capture on $^3$He is known, is the only source of solar neutrinos with energies larger than 15 MeV–their end-point energy is about 19 MeV. This fact has naturally led to questions about the reliability of the currently accepted SSM value for the astrophysical factor at zero energy, $2.3 \times 10^{-20}$ keV b [@Sch92]. In particular, Bahcall and Krastev [@BK98] have shown that a large enhancement, by a factor in the range 20–30, of the SSM $S$-factor value given above would essentially fit the observed excess [@Fuk99] of recoiling electrons. The theoretical description of the $hep$ process, as well as that of the neutron and proton radiative captures on deuteron and $^3$He, constitute a challenging problem from the standpoint of nuclear few-body theory. Its difficulty can be appreciated by comparing the measured values for the cross section of thermal neutron radiative capture on $^1$H, $^2$H, and $^3$He. Their respective cross sections are: $334.2 \pm 0.5$ mb [@CWC65], $0.508 \pm 0.015$ mb [@JBB82], and $0.055 \pm 0.003$ mb [@Wol89]. Thus, in going from $A$=2 to 4 the cross section has dropped by almost four orders of magnitude. These processes are induced by magnetic-dipole transitions between the initial two-cluster state in relative S-wave and the final bound state. In fact, the inhibition of the $A$=3 and 4 captures has been understood for a long time [@Sch37]: the $^3$H and $^4$He states are approximate eigenstates of the magnetic dipole operator ${\bbox \mu}$, and consequently matrix elements of $\mu_z$ between $nd$ ($n\,^3$He) and $^3$H ($^4$He) vanish (approximately) due to orthogonality. This orthogonality argument fails in the case of the deuteron, since then $\mu_z$ can connect the large S-wave component of the deuteron to an isospin $T$=1 $^1$S$_0$ $np$ state. This quasi-orthogonality, while again invalid in the case of the proton weak capture on protons [@Sch98], is also responsible for inhibiting the $hep$ process. Both these reactions are induced by the Gamow-Teller operator, which differs from the (leading) isovector spin part of the magnetic dipole operator essentially by an isospin rotation. As a result, the $hep$ weak capture and $nd$, $pd$ and $n\,^3$He radiative captures are extremely sensitive to: i) the D-state admixtures generated by tensor interactions, and ii) many-body terms in the electroweak current operator. For example, many-body current contributions provide, respectively, 50 % and over 90 % of the calculated $pd$ [@Viv00] and $n\,^3$He [@Sch92; @Car90] cross sections at very low energies. In this respect, the $hep$ weak capture is a particularly delicate reaction, for two additional reasons: firstly and most importantly, the one- and many-body current contributions are comparable in magnitude, but of opposite sign [@Sch92; @Car91]; secondly, many-body axial currents, specifically those arising from excitation of $\Delta$ isobars which give the dominant contribution, are model dependent [@Car91]. This destructive interference between one- and many-body currents also occurs in the $n\,^3$He ($hen$) radiative capture [@Sch92; @Car90], with the difference that there the leading components of the many-body currents are model independent, and give a much larger contribution than that associated with the one-body current. The cancellation in the $hep$ process between the one- and two-body matrix elements has the effect of enhancing the importance of P-wave capture channels. Indeed, one of the results of the present work is that these channels give about 40 % of the $S$-factor calculated value. That the $hep$ process could proceed as easily through P- as S-wave capture was not not sufficiently appreciated [@WB73] in all earlier studies of this reaction we are aware of, with the exception of Ref. [@BK98], in which Horowitz suggested, on the basis of a very simple one-body reaction model, that the $^3$P$_0$ channel may be important. Most of the earlier studies [@Wol89; @WB73; @TEG83] had attempted to relate the matrix element of the axial current occurring in the $hep$ capture to that of the electromagnetic current in the $hen$ capture, exploiting (approximate) isospin symmetry. This approach led, however, to $S$-factor values ranging from $3.7$ to $57$, in units of $10^{-20}$ keV b. In an attempt to reduce the uncertainties in the predicted values for both the radiative and weak capture rates, [*ab initio*]{} microscopic calculations of these reactions were performed in the early nineties [@Sch92; @Car90; @Car91], using variational wave functions corresponding to a realistic Hamiltonian, and a nuclear electroweak current consisting of one- and many-body components. These studies showed that inferring the $hep$ $S$-factor from the measured $hen$ cross section can be misleading, because of different initial-state interactions in the $n\,^3$He and $p\,^3$He channels, and because of the large contributions associated with the two-body components of the electroweak current operator, and their destructive interference with the one-body current contributions. The significant progress made in the last few years in the modeling of two- and three-nucleon interactions and the nuclear weak current, and the description of the bound and continuum four-nucleon wave functions, have prompted us to re-examine the $hep$ reaction. In the present work we briefly summarize the salient points in the calculation, and report our results for the $S$-factor in the energy range 0–10 keV. An exhaustive account of this study [@Mar00], however, will be published elsewhere. The cross section for the $^{3}$He($p$,$e^{+}\nu_{e}$)$^{4}$He reaction at a c.m. energy $E$ is written as $$\begin{aligned} \sigma(E)=\int&& 2\pi \, \delta\left (\Delta m + E - \frac{q^{2}}{2 m_{4}} - E_e - E_\nu\right )\frac{1}{v_{\rm rel}} \nonumber \\ &&\times \frac{1}{4}\sum_{s_e s_\nu}\sum_{s_1 s_3} |\langle f\,|\,H_{W}\,|\,i\rangle|^{2} \frac{d{\bf{p}}_{e}}{(2\pi)^3} \frac{d{\bf{p}}_{\nu}}{(2\pi)^3} \ , \label{eq:xsc1}\end{aligned}$$ where $\Delta m = m + m_3 - m_4 $ = 19.8 MeV ($m$, $m_3$, and $m_4$ are the proton, $^3$He, and $^4$He rest masses, respectively), $v_{\rm rel}$ is the $p\,^{3}$He relative velocity, and the transition amplitude is given by $$\langle f|H_{W}|i\rangle= \frac{G_{V}}{\sqrt{2}}\,l^{\sigma} \langle -{\bf{q}}; ^{4}\!{\rm{He}}| j_{\sigma}^{\dag}({\bf{q}})|{\bf{p}}; p\,^{3}{\rm{He}}\rangle \ . \label{eq:tra}$$ Here $G_V$ is the Fermi constant, ${\bf{q}}={\bf{p}}_{e}+{\bf{p}}_{\nu}$, $|{\bf{p}}; p\,^{3}{\rm{He}}\rangle$ and $|-\!{\bf{q}}; ^{4}\!{\rm{He}} \rangle$ represent, respectively, the $p\,^{3}$He scattering state with relative momentum ${\bf{p}}$ and $^{4}$He bound state recoiling with momentum $-{\bf{q}}$, $l_{\sigma}$ is the leptonic weak current, $l_{\sigma} = \overline{u}_{\nu}\gamma_{\sigma}(1-\gamma_5)v_{e}$ (the lepton spinors are normalized as $v_{e}^{\dag}v_{e}={u}_{\nu}^{\dag}{u}_{\nu}=1$), and $j^\sigma({\bf q})$ is the nuclear weak current, $j^\sigma({\bf q}) =( \rho({\bf q}), {\bf j}({\bf q}))$. The dependence of the amplitude upon the spin projections of the leptons, proton and $^3$He has been omitted for ease of presentation. Since the energies of interest are of the order of 10 keV or less–the Gamow-peak energy is 10.7 keV for the $hep$ reaction–it is convenient to expand the $p\,^3$He scattering state into partial waves, and perform a multipole decomposition of the nuclear weak charge, $\rho({\bf q})$, and current, ${\bf j}({\bf q})$, operators. Standard manipulations lead to [@Mar00] $$\frac{1}{4}\sum_{s_e s_\nu}\sum_{s_1 s_3} |\langle f\,|\,H_{W}\,|\,i\rangle|^{2} = (2 \pi)^2\> G_V^2\> L_{\sigma \tau} \> N^{\sigma\tau} ,$$ where the lepton tensor $L^{\sigma \tau}$ is written in terms of electron and neutrino four-velocities as $L^{\sigma \tau}={\rm v}_{e}^{\sigma}{\rm v}_{\nu}^{\tau}+ {\rm v}_{\nu}^{\sigma}{\rm v}_{e}^{\tau}-g^{\sigma\tau} {\rm v}_{e} \cdot {\rm v}_{\nu}+ {\rm{i}}\>\epsilon^{\sigma\alpha\tau\beta}{\rm v}_{e,\alpha} {\rm v}_{\nu,\beta}$, while the nuclear tensor is defined as $$N^{\sigma \tau} \equiv \sum_{s_1 s_3} W^\sigma({\bf q};s_1 s_3) W^{\tau *}({\bf q}; s_1 s_3) \ , \label{nuclt}$$ with $$\begin{aligned} W^{\sigma=0,3}({\bf q}; s_1 s_3)&=&\sum_{LSJ} X^{LSJ}_0(\hat{\bf q};s_1 s_3) T^{LSJ}_J(q) \ , \\ W^{\sigma=\lambda}({\bf q}; s_1 s_3)&=& -\frac{1}{\sqrt{2}} \sum_{LSJ} X^{LSJ}_{-\lambda}(\hat{\bf q};s_1 s_3) \nonumber \\ & &\left [\lambda \, M^{LSJ}_J(q) + E^{LSJ}_J(q) \right ] \ ,\end{aligned}$$ where $\lambda = \pm 1$ denote spherical components. The functions $X_{\lambda=0,\pm 1}$ depend upon the direction $\hat{\bf q}$, and the proton and $^3$He spin projections $s_1$ and $s_3$ [@Mar00] (note that the quantization axis for the hadronic states is taken along $\hat{\bf p}$, the direction of the $p\,^3$He relative momentum), while $T^{LSJ}_J$=$C_{J}^{LSJ}$ or $L_{J}^{LSJ}$ for $\sigma$=0 or 3. The quantities $C_J^{LSJ}$, $L_J^{LSJ}$, $E_{J}^{LSJ}$ and $M_{J}^{LSJ}$ are the reduced matrix elements (RMEs) of the Coulomb $(C_{\ell \ell_z})$, longitudinal $(L_{\ell \ell_z})$, transverse electric $(E_{\ell \ell_z})$, and transverse magnetic $(M_{\ell \ell_z})$ multipole operators between the initial $p\,^3$He state with orbital angular momentum $L$, channel spin $S$ ($S$=0,1), and total angular momentum $J$, and final $^4$He state. The present study includes S- and P-wave capture channels, i.e. the $^1$S$_0$, $^3$S$_1$, $^3$P$_0$, $^1$P$_1$, $^3$P$_1$, and $^3$P$_2$ states, and retains all contributing multipoles connecting these states to the $J^\pi$=0$^+$ ground state of $^4$He. The bound- and scattering-state wave functions are obtained variationally with the correlated-hyperspherical-harmonics (CHH) method, developed in Refs. [@VKR95; @VRK98]. The nuclear Hamiltonian consists of the Argonne $v_{18}$ two-nucleon [@WSS95] and Urbana-IX three-nucleon [@Pud95] interactions. This realistic Hamiltonian, denoted as AV18/UIX, reproduces the experimental binding energies and charge radii of the trinucleons and $^4$He in exactGreen’s function Monte Carlo (GFMC) calculations [@Pud97]. The binding energy of $^4$He calculated with the CHH method [@Mar00; @VKR95] is within 1 % of that obtained with the GFMC method. The accuracy of the CHH method to calculate scattering states has been successfully verified in the case of three-nucleon systems, by comparing results for a variety of $Nd$ scattering observables obtained by a number of groups using different techniques [@Kie98]. Studies along similar lines [@Viv98] to assess the accuracy of the CHH solutions for the four-nucleon continuum have already begun. The CHH predictions [@VRK98] for the $n\,^3$H total elastic cross section and coherent scattering length have been found to be in excellent agreement with the corresponding experimental values. The $n\,^3$H cross section is known over a rather wide energy range, and its extrapolation to zero energy is not problematic [@Phi80]. The situation is different for the $p\,^3$He channel, for which the singlet and triplet scattering lengths $a_{\rm s}$ and $a_{\rm t}$ have been determined from effective range extrapolations of data taken above 1 MeV, and are therefore somewhat uncertain, $a_{\rm s}=(10.8 \pm 2.6)$ fm [@AK93] and $a_{\rm t}=(8.1 \pm 0.5)$ fm [@AK93] or $(10.2 \pm 1.5)$ fm [@TEG83]. Nevertheless, the CHH results are close to the experimental values above: the AV18/UIX Hamiltonian predicts [@VRK98] $a_{\rm s}=11.5$ fm and $a_{\rm t}=9.13$ fm. At low energies (below 4 MeV) $p\,^3$He elastic scattering proceeds mostly through S- and P-wave channels, and the CHH predictions, based on the AV18/UIX model, for the differential cross section [@VKR00] are in good agreement with the experimental data. The nuclear weak current has vector and axial-vector parts, with corresponding one- and many-body components. The one-body components have the standard expressions obtained from a non-relativistic reduction of the covariant single-nucleon vector and axial-vector currents, including terms proportional to $1/m^2$. The two-body weak vector currents are constructed from the isovector two-body electromagnetic currents in accordance with the conserved-vector-current hypothesis, and consist [@Mar00] of model-independentand model-dependentterms. The model-independent terms are obtained from the nucleon-nucleon interaction, and by construction satisfy current conservation with it. The leading two-body weak vector current is the $\pi$-likeoperator, obtained from the isospin-dependent spin-spin and tensor nucleon-nucleon interactions. The latter also generate an isovector $\rho$-likecurrent, while additional isovector two-body currents arise from the isospin-independent and isospin-dependent central and momentum-dependent interactions. These currents are short-ranged, and numerically far less important than the $\pi$-like current. With the exception of the $\rho$-like current, they have been neglected in the present work. The model-dependent currents are purely transverse, and therefore cannot be directly linked to the underlying two-nucleon interaction. The present calculation includes the currents associated with excitation of $\Delta$ isobars which, however, are found to give a rather small contribution in weak-vector transitions, as compared to that due to the $\pi$-like current. The $\pi$-like and $\rho$-like contributions to the weak vector charge operator [@Mar00] have also been retained in the present study. The leading many-body terms in the axial current due to $\Delta$-isobar excitation are treated non-perturbatively in the transition-correlation-operator (TCO) scheme, originally developed in Ref. [@Sch92] and further extended in Ref. [@Mar98]. In the TCO scheme–essentially, a scaled-down approach to a full $N$$+$$\Delta$ coupled-channel treatment–the $\Delta$ degrees of freedom are explicitly included in the nuclear wave functions. The axial charge operator includes, in addition to $\Delta$-excitation terms (which, however, are found to be unimportant [@Mar00]), the long-range pion-exchange term [@Kub78], required by low-energy theorems and the partially-conserved-axial-current relation, as well as the (expected) leading short-range terms constructed from the central and spin-orbit components of the nucleon-nucleon interaction, following a prescription due to Riska and collaborators [@Kir92]. The largest model dependence is in the weak axial current. To minimize it, the poorly known $N$$\Delta$ transition axial coupling constant $g_A^*$ has been adjusted to reproduce the experimental value of the Gamow-Teller matrix element in tritium $\beta$-decay [@Sch98; @Mar00]. While this procedure is model dependent, its actual model dependence is in fact very weak, as has been shown in Ref. [@Sch98]. The calculation proceeds in two steps [@Mar00]: first, the matrix elements of $\rho({\bf q})$ and ${\bf j}({\bf q})$ between the initial $p\,^3$He $LSJJ_z$ states and final $^4$He are calculated with Monte Carlo integration techniques; second, the contributing RMEs are extracted from these matrix elements, and the cross section is calculated by performing the integrations over the electron and neutrino momenta in Eq. (\[eq:xsc1\]) numerically, using Gauss points. The results for the $S$-factor, defined as $S(E) = E\, \sigma(E)\, {\rm exp}( 4\, \pi \, \alpha/v_{\rm rel})$ ($\alpha$ is the fine structure constant), at $p\,^3$He c.m. energies of $0$, $5$, and $10$ keV are reported in Table \[tb:sfact\]. In the table, the column labelled S includes both the $^1$S$_0$ and $^3$S$_1$ channel contributions, although the former are at the level of a few parts in $10^{3}$. The energy dependence is rather weak: the value at $10$ keV is only about 4 % larger than that at $0$ keV. The P-wave capture states are found to be important, contributing about 40 % of the calculated $S$-factor. However, the contributions from D-wave channels are expected to be very small. We have verified explicitly that they are indeed small in $^3$D$_1$ capture. The many-body axial currents associated with $\Delta$ excitation play a crucial role in the (dominant) $^3$S$_1$ capture, where they reduce the $S$-factor by more than a factor of four. Thus the destructive interference between the one- and many-body current contributions, first obtained in Ref. [@Car91], is confirmed in the present study. The (suppressed) one-body contribution comes mostly from transitions involving the D-state components of the $^3$He and $^4$He wave functions, while the many-body contributions are predominantly due to transitions connecting the S-state in $^3$He to the D-state in $^4$He, or viceversa. It is important to stress the differences between the present and all previous studies. Apart from ignoring, or at least underestimating, the contribution due to P-waves, the latter only considered the long-wavelength form of the weak multipole operators, namely, their $q$=$0$ limit. In $^3$P$_0$ capture, for example, only the $C_0$-multipole, associated with the weak axial charge, survives in this limit, and the corresponding $S$-factor is calculated to be $2.2 \times 10^{-20}$ keV b, including two-body contributions. However, when the transition induced by the longitudinal component of the axial current (via the $L_0$-multipole, which vanishes at $q$=$0$) is also taken into account, the $S$-factor becomes $0.82 \times 10^{-20}$ keV b, because of a destructive interference between the $C_0$ and $L_0$ RMEs. Thus use of the long-wavelength approximation in the calculation of the $hep$ $S$-factor leads to inaccurate results. Finally, besides the differences listed above, the present calculation also improves that of Ref. [@Sch92] in a number of other important respects: firstly, it uses accurate CHH wave functions, corresponding to the last generation of realistic interactions; secondly, the model for the nuclear weak current has been extended to include the axial charge as well as the vector charge and current operators. Thirdly, the one-body operators now take into account the $1/m^2$ relativistic corrections, which were previously neglected. In $^3$S$_1$ capture, for example, these terms increase by 25 % the dominant (but suppressed) $L_1$ and $E_1$ RMEs calculated with the (lowest order) Gamow-Teller operator. These improvements in the treatment of the one-body axial current indirectly affect also the contributions of the $\Delta$-excitation currents [@Mar00], because of the procedure used to determine the coupling constant $g_A^*$. To conclude, we have carried out a realistic calculation of the $hep$ reaction, predicting a value for the $S$-factor five times larger than that used in the SSM. This enhancement, while very significant, is far smaller than that required by fits to the Super-Kamiokande data. Although the present result is inherently model dependent, it is unlikely the model dependence be so large to accomodate a drastic increase in the prediction obtained here. The authors wish to thank J.F. Beacom, J. Carlson, V.R. Pandharipande, D.O. Riska, P. Vogel, and R.B. Wiringa for useful discussions. The support of the U.S. Department of Energy under contract number DE-AC05-84ER40150 is gratefully acknowledged by L.E.M. and R.S. J.N. Bahcall and P.I. Krastev, Phys. Lett. B [**436**]{}, 243 (1998); G. Fiorentini, V. Berezinsky, S. Degl’Innocenti, and B. Ricci, Phys. Lett. B [**444**]{}, 387 (1998); C.J. Horowitz, Phys. Rev. C [**60**]{}, 022801 (1999); W.M. Alberico, S.M. Bilenky, and W. Grimus, hep-ph/0001245. Y. Fukuda [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 2430 (1999); M.B. Smy, hep-ex/9903034. J.N. Bahcall, S. Basu, and M.H. Pinsonneault, Phys. Lett. B [**433**]{}, 1 (1998). R. Schiavilla, R.B. Wiringa, V.R. Pandharipande, and J. Carlson, Phys. Rev. C [**45**]{}, 2628 (1992). A.E. Cox, S.A.R. Wynchank, and C.H. Collie, Nucl. Phys. [**74**]{}, 497 (1965). E.T. Jurney, P.J. Bendt, and J.C. Browne, Phys. Rev. C [**25**]{}, 2810 (1982). F.L.H. Wolfs, S.J. Freedman, J.E. Nelson, M.S. Dewey, and G.L. Greene, Phys. Rev. Lett. [**63**]{}, 2721 (1989); R. Wervelman, K. Abrahams, H. Postma, J.G.L. Booten, and A.G.M. Van Hees, Nucl. Phys. [**A526**]{}, 265 (1991). L.I. Schiff, Phys. Rev. [**52**]{}, 242 (1937). R. Schiavilla [*et al.*]{}, Phys. Rev. C [**58**]{}, 1263 (1998). M. Viviani, A. Kievsky, L.E. Marcucci, S. Rosati, and R. Schiavilla, Phys. Rev. C in press. J. Carlson, D.O. Riska, R. Schiavilla, and R.B. Wiringa, Phys. Rev. C [**42**]{}, 830 (1990). J. Carlson, D.O. Riska, R. Schiavilla, and R.B. Wiringa, Phys. Rev. C [**44**]{}, 619 (1991). C. Werntz and J.G. Brennan, Phys. Rev. C [**8**]{}, 1545 (1973). P.E. Tegnér and C. Bargholtz, Astrophys. J. [**272**]{}, 311 (1983). L.E. Marcucci, R. Schiavilla, M. Viviani, A. Kievsky, and S. Rosati, in preparation. M. Viviani, A. Kievsky, and S. Rosati, Few-Body Syst. [**18**]{}, 25 (1995). M. Viviani, S. Rosati, and A. Kievsky, Phys. Rev. Lett. [**81**]{}, 1580 (1998). R.B. Wiringa, V.G.J. Stoks, and R. Schiavilla, Phys. Rev. C [**51**]{}, 38 (1995). B.S. Pudliner, V.R. Pandharipande, J. Carlson, and R.B. Wiringa, Phys. Rev. Lett. [**74**]{}, 4396 (1995). B.S. Pudliner, V.R. Pandharipande, J. Carlson, S.C. Pieper, and R.B. Wiringa, Phys. Rev. C [**56**]{}, 1720 (1997). A. Kievsky [*et al.*]{}, Phys. Rev. C [**58**]{}, 3085 (1998). M. Viviani, Nucl. Phys. [**A631**]{}, 111c (1998). T.W. Phillips, B.L. Berman, and J.D. Seagrave, Phys. Rev. C [**22**]{}, 384 (1980). M.T. Alley and L.D. Knutson, Phys. Rev. C [**48**]{}, 1901 (1993). M. Viviani, A. Kievsky, and S. Rosati, in preparation. L.E. Marcucci, D.O. Riska, and R. Schiavilla, Phys. Rev. C [**58**]{}, 3069 (1998). K. Kubodera, J. Delorme, and M. Rho, Phys. Rev. Lett. [**40**]{}, 755 (1978). M. Kirchbach, D.O. Riska, and K. Tsushima, Nucl. Phys. [**A542**]{}, 616 (1992). ----------- ------ ------ ------ ------ ------ ------ $E$ (keV) S S+P S S+P S S+P one-body 26.4 29.0 25.9 28.7 26.2 29.3 full 6.39 9.64 6.21 9.70 6.37 10.1 ----------- ------ ------ ------ ------ ------ ------ : The $hep$ $S$-factor, in units of $10^{-20}$ keV b, calculated with CHH wave functions corresponding to the AV18/UIX Hamiltonian model, at $p\,^3$He c.m. energies $E$=$0$, $5$, and $10$ keV. The rows labelled one-bodyand fulllist the contributions obtained by retaining the one-body only and both one- and many-body terms in the nuclear weak current. The contributions due to the S-wave channels only and S- and P-wave channels are listed separately. The Monte Carlo statistical error is at the 5% level on the total $S$-factor. \[tb:sfact\]

--- abstract: 'We consider a class of ordinary differential equations describing one-dimensional quasi-periodically forced systems in the presence of large damping. We give a fully constructive proof of the existence of response solutions, that is quasi-periodic solutions which have the same frequency vector as the forcing. This requires dealing with a degenerate implicit function equation: we prove that the latter has a unique solution, which can be explicitly determined. As a by-product we obtain an explicit estimate of the minimal size of the damping coefficient.' author: - | **Guido Gentile\ Dipartimento di Matematica, Università di Roma Tre, Roma, I-00146, Italy.\ E-mail: gentile@mat.uniroma3.it** title: | **Construction of quasi-periodic response solutions\ in forced strongly dissipative systems** --- Introduction {#sec:1} ============ In this paper we continue the analysis started in [@GBD1; @GBD2; @G2] on the existence and properties of quasi-periodic motions in one-dimensional strongly dissipative forced systems. We consider one-dimensional systems with a quasi-periodic forcing term in the presence of strong damping, described by ordinary differential equations of the form $${\varepsilon}\ddot x + \dot x + {\varepsilon}g(x) = {\varepsilon}f({\boldsymbol{\omega}}t) , \label{eq:1.1}$$ where $g\!:{\mathds{R}}\to{\mathds{R}}$ and $f\!:{\mathds{T}}^{d}\to{\mathds{R}}$ are real analytic functions and ${\mathds{T}}={\mathds{R}}/2\pi{\mathds{Z}}$. We call $g(x)$ the *mechanical force*, $f({\boldsymbol{\omega}}t)$ the *forcing term*, ${\boldsymbol{\omega}}\in{\mathds{R}}^{d}$ the *frequency vector* of the forcing, and ${\gamma}=1/{\varepsilon}>0$ the *damping coefficient*. The function $f$ is quasi-periodic in $t$, i.e. $$f({\boldsymbol{\psi}}) = \sum_{{\boldsymbol{\nu}}\in {\mathds{Z}}^{d}} {\rm e}^{{{\rm i}}{\boldsymbol{\nu}}\cdot {\boldsymbol{\psi}}} f_{{\boldsymbol{\nu}}} , \qquad {\boldsymbol{\psi}}\in {\mathds{T}}^{d} , \label{eq:1.2}$$ with average ${\langle}f {\rangle}= f_{{\boldsymbol{0}}}$, and $\cdot$ denoting the scalar product in ${\mathds{R}}^{d}$. By the analyticity assumption on $f$ and $g$, one has $|f_{{\boldsymbol{\nu}}}| \le \Phi {\rm e}^{-\xi|{\boldsymbol{\nu}}|}$ and $|{{\rm d}}^{s}g(c_{0})/{{\rm d}}x^{s}| \le s!\Gamma^{s}$ for suitable positive constants $\Phi$, $\xi$, and $\Gamma$. A Diophantine condition is assumed on ${\boldsymbol{\omega}}$. Define the *Bryuno function* $${{\mathfrak B}}({\boldsymbol{\omega}}) = \sum_{n=0}^{\infty} \frac{1}{2^{n}} \log \frac{1}{{\alpha}_{n}({\boldsymbol{\omega}})} , \qquad {\alpha}_{n}({\boldsymbol{\omega}}) = \inf \{ |{\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}| : {\boldsymbol{\nu}}\in {\mathds{Z}}^{d} \hbox{ such that } 0<|{\boldsymbol{\nu}}|\le 2^{n} \} . \label{eq:1.3}$$ \[hyp:1\] The frequency vector ${\boldsymbol{\omega}}$ satisfies the Bryuno condition ${{\mathfrak B}}({\boldsymbol{\omega}})<\infty$. The following assumption will be made on the functions $g$ and $f$ (for given force $g(x)$ this can read as a condition on the forcing term). \[hyp:2\] There exists $c_{0}\in{\mathds{R}}$ such that $x=c_{0}$ is a zero of odd order ${{\mathfrak n}}$ of the equation $$g(x) - f_{{\boldsymbol{0}}} = 0 , \label{eq:1.4}$$ that is $g_{0}={{\mathfrak n}}!^{-1} {{\rm d}}^{{{\mathfrak n}}}g/{{\rm d}}x^{{{\mathfrak n}}}(c_{0})\neq0$ and, if ${{\mathfrak n}}>1$, ${{\rm d}}^{k} g/{{\rm d}}x^{k}(c_{0})=0$ for $k=1,\ldots,{{\mathfrak n}}-1$. In [@G2] we proved the following result about the existence of quasi-periodic solutions with the same frequency vector ${\boldsymbol{\omega}}$ as the forcing (response solutions). \[thm:1.1\] Under Assumptions \[hyp:1\] and \[hyp:2\], for ${\varepsilon}$ small enough there exists at least one quasi-periodic solution $x_{0}(t)=c_{0}+u({\boldsymbol{\omega}}t,{\varepsilon})$, with frequency vector ${\boldsymbol{\omega}}$, reducing to $c_{0}$ as ${\varepsilon}$ tends to $0$. Such a solution is analytic in $t$. Note that the condition ${\varepsilon}>0$ could be eliminated: indeed, the proof in [@G2] works for all ${\varepsilon}$ small enough, and the request ${\varepsilon}>0$ only aims to interpret ${\gamma}=1/{\varepsilon}$ as the damping coefficient. Analyticity on $t$ is not explicitly stated in [@G2], but follows immediately from the proof therein. We also proved in [@G2] that the condition that $c_{0}$ be a zero of odd order of (\[eq:1.4\]) is a necessary and sufficient condition for a quasi-periodic solution around $c_{0}$ to exists. However, as pointed out in [@G2], except for the non-degenerate case ${{\mathfrak n}}=1$ (where the implicit function theorem applies), in general the proof ultimately relies on continuity arguments, which do not provide a quantitative constructive estimate on the maximal size of the perturbation parameter ${\varepsilon}$. Moreover, the quasi-periodic solution was constructed in terms of two parameters, that is ${\varepsilon}$ and $c$: the latter is defined as the constant part of the quasi-periodic solution itself and is fixed in terms of ${\varepsilon}$ so as to solve a certain implicit function equation (the so-called bifurcation equation). In particular, a quasi-periodic solution was showed to exist for any solution $c({\varepsilon})$ to the bifurcation equation, but the problem of studying how many such solutions exist and how do they depend on ${\varepsilon}$ was not investigated. In this paper we show that for all odd ${{\mathfrak n}}$ the bifurcation equation admits one and only one solution. Furthermore, we explicitly construct the quasi-periodic solution in terms of the only parameter ${\varepsilon}$, and we also give a quantitative estimate on the maximal size of ${\varepsilon}$. So, with respect to [@G2], the proof of existence of the quasi-periodic solution is fully constructive. We can summarise our results in the following statement. \[thm:1.2\] Under Assumptions \[hyp:1\] to \[hyp:2\] on the ordinary differential equation (\[eq:1.1\]), there exists an (explicitly computable) constant ${\varepsilon}_{0}>0$ such that for all $|{\varepsilon}|<{\varepsilon}_{0}$ there exists a quasi-periodic solution $x_{0}(t)=c_{0}+u({\boldsymbol{\omega}}t,{\varepsilon})$, with frequency vector ${\boldsymbol{\omega}}$, such that $u=O({\varepsilon})$ as ${\varepsilon}\to0$. Such a solution is analytic in $t$ and depends $C^{{\infty}}$-smoothly on ${\varepsilon}$. The result solves a problem left as open in [@G2]. The problem remains whether other quasi-periodic solutions exist. Numerical algorithms to construct response solutions of quasi-periodically forced dissipative systems are provided in [@CH; @KP1; @KP2], based either on a generalised harmonic balance method or on a fixed point method for a suitable Poincaré map. Also the method described in the present paper is well suited for numerical implementations, and allows a completely rigorous control of the approximation error for the solution. Moreover it applies also in degenerate cases where one cannot apply directly the implicit function theorem (of course it requires for the damping coefficient to be large enough). The bifurcation equation {#sec:2} ======================== The existence of quasi-periodic solutions $x_{0}(t)$ has been proved in [@G2]. The proof proceeds as follows. First of all, write $$x_{0}(t)=c_{0}+u({\boldsymbol{\omega}}t,{\varepsilon})=c+X({\boldsymbol{\omega}}t;{\varepsilon},c) , \qquad X({\boldsymbol{\psi}};{\varepsilon},c)=\sum_{{\boldsymbol{\nu}}\in{\mathds{Z}}^{d}_{*}} {\rm e}^{{{\rm i}}{\boldsymbol{\nu}}\cdot{\boldsymbol{\psi}}} X_{{\boldsymbol{\nu}}}({\varepsilon},c) , \label{eq:2.1}$$ with ${\mathds{Z}}^{d}_{*}={\mathds{Z}}^{d}\setminus\{{\boldsymbol{0}}\}$ (so that $c$ is the average of $x_{0}$ on the $d$-dimensional torus). The Fourier coefficients $X_{{\boldsymbol{\nu}}}=X_{{\boldsymbol{\nu}}}({\varepsilon},c)$ of the function $X({\boldsymbol{\psi}};{\varepsilon},c)$ are obtained by solving the *range equation* $${{\rm i}}{\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}\left( 1 + {{\rm i}}{\varepsilon}{\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}\right) X_{{\boldsymbol{\nu}}} + {\varepsilon}\left[ g(c+X(\cdot;{\varepsilon},c) \right]_{{\boldsymbol{\nu}}} = {\varepsilon}f_{{\boldsymbol{\nu}}} , \qquad {\boldsymbol{\nu}}\neq {\boldsymbol{0}}, \label{eq:2.2}$$ where $$\left[ g(c+X(\cdot;{\varepsilon},c) \right]_{{\boldsymbol{\nu}}} = \sum_{p=0}^{\infty} \frac{1}{p!} \frac{{\rm d}^{p}}{{\rm d}x^{p}} g(c) \sum_{\substack{{\boldsymbol{\nu}}_{1},\ldots,{\boldsymbol{\nu}}_{p}\in{\mathds{Z}}^{d}_{*} \\ {\boldsymbol{\nu}}_{1}+\ldots+{\boldsymbol{\nu}}_{p}={\boldsymbol{\nu}}}} X_{{\boldsymbol{\nu}}_{1}}({\varepsilon},c) \ldots X_{{\boldsymbol{\nu}}_{p}}({\varepsilon},c) . \label{eq:2.3}$$ The analysis in [@G2] shows that for all $c$ close enough to $c_{0}$ and all ${\varepsilon}$ small enough there exists a function $X({\boldsymbol{\psi}};{\varepsilon},c)$ which solves (\[eq:2.2\]). Moreover the map $({\varepsilon},c) \mapsto X(\cdot;{\varepsilon},c)$ is $C^{\infty}$ in a neighbourhood of $(0,c_{0})$, and $$X({\boldsymbol{\psi}};{\varepsilon},c) = {\varepsilon}X^{(1)}({\boldsymbol{\psi}}) + o({\varepsilon}) , \qquad \dot X^{(1)}({\boldsymbol{\omega}}t) = f({\boldsymbol{\omega}}t) . \label{eq:2.4}$$ Then, for any solution $c=c({\varepsilon})$ to the *bifurcation equation* $$F({\varepsilon},c) := [g(c+X(\cdot;{\varepsilon},c)]_{{\boldsymbol{0}}} - f_{{\boldsymbol{0}}} = 0 , \label{eq:2.5}$$ there exists a quasi-periodic solution $x_{0}(t)=c({\varepsilon})+X({\boldsymbol{\omega}}t;{\varepsilon},c({\varepsilon}))$. In principle, there could be several solutions to (\[eq:2.5\]) when ${{\mathfrak n}}>1$. On the contrary, we shall prove that the solution to (\[eq:2.5\]) is unique. Unfortunately, this does not implies that the response solution $c_{0}+u({\boldsymbol{\omega}}t,{\varepsilon})$ of Theorem \[thm:1.1\] is the only quasi-periodic solution reducing to $c_{0}$ as ${\varepsilon}\to0$, because no result ensures that the function $X({\boldsymbol{\psi}};{\varepsilon},c)$ which solves (\[eq:2.2\]) at fixed $c$ and ${\varepsilon}$ is unique. We shall write $c=c({\varepsilon})=c_{0}+\zeta$, with $\zeta=\zeta({\varepsilon})$ such that $\zeta(0)=0$, so that (\[eq:2.3\]) becomes $$\left[ g(c+X(\cdot;{\varepsilon},c) \right]_{{\boldsymbol{\nu}}} = \sum_{p={{\mathfrak n}}}^{\infty} \frac{1}{p!} \frac{{\rm d}^{p}}{{\rm d}x^{p}} g(c_{0}) \sum_{k=0}^{p} \left( \begin{matrix} p \\ k \end{matrix} \right) \zeta^{p-k} \sum_{\substack{{\boldsymbol{\nu}}_{1},\ldots,{\boldsymbol{\nu}}_{k}\in{\mathds{Z}}^{d}_{*} \\ {\boldsymbol{\nu}}_{1}+\ldots+{\boldsymbol{\nu}}_{k}={\boldsymbol{\nu}}}} X_{{\boldsymbol{\nu}}_{1}}({\varepsilon},c) \ldots X_{{\boldsymbol{\nu}}_{k}}({\varepsilon},c) . \label{eq:2.6}$$ \[lem:2.1\] Let $c=c({\varepsilon})$ a solution of the bifurcation equation (\[eq:2.5\]). Then $c=c_{0}+O({\varepsilon})$, and $c=c_{0}+o({\varepsilon})$ requires $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$. [[*Proof.* ]{}]{}Assume that $\zeta/{\varepsilon}\to{\infty}$ as ${\varepsilon}\to0$. Then, by using (\[eq:2.4\]), one has $$[g(c+X(\cdot;{\varepsilon},c)]_{{\boldsymbol{0}}} - f_{{\boldsymbol{0}}} = g_{0} \left[ \left( \zeta + X(\cdot,{\varepsilon},c) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} + O(\zeta^{{{\mathfrak n}}+1}) = g_{0} \zeta^{{{\mathfrak n}}} + O(\zeta^{{{\mathfrak n}}+1}) = 0 , \nonumber$$ which leads to a contradiction. Hence $\zeta=O({\varepsilon})$. On the other hand if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ and $\zeta=o({\varepsilon})$ one has $$[g(c+X(\cdot;{\varepsilon},c)]_{{\boldsymbol{0}}} - f_{{\boldsymbol{0}}} = g_{0} \left[ \left( \zeta + X(\cdot;{\varepsilon},c) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} + O({\varepsilon}^{{{\mathfrak n}}+1}) = g_{0} {\varepsilon}^{{{\mathfrak n}}} [(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}} + O({\varepsilon}^{{{\mathfrak n}}+1}) = 0 , \nonumber$$ which once more leads to a contradiction. Note that a sufficient condition for $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}$ to vanish is that $f$ in (\[eq:1.2\]) is even. The function $F({\varepsilon},c)$, defined in (\[eq:2.5\]), is $C^{{\infty}}$ in both ${\varepsilon}$ and $c$; see [@G2]. In order to identify the leading orders to $F({\varepsilon},c)$, we consider the carrier $$\Delta(F) = \left\{ (k,j) \in {\mathds{Z}}_{+}\times{\mathds{Z}}_{+} : F_{k,j} \neq 0 \right\} , \qquad F_{k,j} = \frac{{\rm d}^{k}}{{\rm d}{\varepsilon}^{k}} \frac{{\rm d}^{j}}{{\rm d}c^{j}}F(0,c_{0}) , \label{eq:2.7}$$ and draw the Newton polygon in the $(k,j)$ plane; see [@BK] (see also [@CG]). If we denote by $\{(k_{1},j_{1}),(k_{2},j_{2})\}$ the segment with endpoints $(k_{1},j_{1})$ and $(k_{2},j_{2})$ in the $(k,j)$ plane, by construction for ${{\mathfrak n}}>1$ the Newton polygon consists of (cf. figure \[fig:1\]) 1. either the only segment $\{(0,{{\mathfrak n}}),({{\mathfrak n}},0)\}$, if $F_{{{\mathfrak n}},0}=[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$, 2. or the only segment $\{(0,{{\mathfrak n}}),({{\mathfrak n}}-1,1)\}$, if $F_{k,0}=0$ for all $k\ge0$, 3. or the two segments $\{(0,{{\mathfrak n}}),({{\mathfrak n}}-1,1)\}$ and $\{({{\mathfrak n}}-1,1),({{\mathfrak p}},0)\}$, with ${{\mathfrak p}}\ge {{\mathfrak n}}+1$, if $F_{k,0}=0$ for all $k\le {{\mathfrak p}}-1$ and $F_{{{\mathfrak p}},0}\neq0$. .3truecm ![Newton polygon (corresponding to cases 1 to 3, respectively)[]{data-label="fig:1"}](newton.eps "fig:") Therefore, in principle, a solution to the bifurcation equation $F({\varepsilon},c)=0$ is - either $c=c_{0}+\zeta_{1}{\varepsilon}+o({\varepsilon})$ for some $\zeta_{1}\neq0$, corresponding to the segment $\{(0,{{\mathfrak n}}),({{\mathfrak n}},0)\}$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ (case 1) and to the segment $\{(0,{{\mathfrak n}}),({{\mathfrak n}}-1,1)\}$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$ (cases 2 and 3), - or $c=c_{0}+\zeta_{0} {\varepsilon}^{{{\mathfrak p}}-{{\mathfrak n}}+1}+o({\varepsilon}^{{{\mathfrak p}}-{{\mathfrak n}}+1})$ for some $\zeta_{0} \neq 0$, corresponding to the segment $\{({{\mathfrak n}}-1,1),({{\mathfrak p}},0)\}$ (case 3), - or $c=c_{0}+\zeta({\varepsilon})$, with $\zeta({\varepsilon})$ decaying to zero faster than any power of ${\varepsilon}$ (case 2). The case ${{\mathfrak n}}=1$ should be discussed apart, but since it is much easier and moreover has already been discussed in [@GBD1; @GBD2], here we concentrate on ${{\mathfrak n}}>1$. \[lem:2.2\] If $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$ then the bifurcation equation (\[eq:2.5\]) admits one and only one solution $c({\varepsilon})=o({\varepsilon})$. Moreover such a solution is smooth in ${\varepsilon}$, and it can be explicitly computed. [[*Proof.* ]{}]{}If $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$ one has $$\begin{aligned} \left[ (\zeta + {\varepsilon}X^{(1)}(\cdot) )^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} & \!\!\! = \!\!\! & \left[ (\zeta + {\varepsilon}X^{(1)}(\cdot) )^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} - \left[ ( {\varepsilon}X^{(1)}(\cdot) )^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} = \left[ (\zeta + {\varepsilon}X^{(1)}(\cdot) )^{{{\mathfrak n}}} - ({\varepsilon}X^{(1)}(\cdot))^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} \nonumber \\ & \!\!\! = \!\!\! & {{\mathfrak n}}\zeta \left[ \int_{0}^{1} {\rm d}s \left( {\varepsilon}X^{(1)}(\cdot) + s\zeta \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} = {{\mathfrak n}}\zeta \int_{0}^{1} {\rm d}s \left[ \left( {\varepsilon}X^{(1)}(\cdot) + s\zeta \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} , \nonumber \end{aligned}$$ which is non-zero because ${{\mathfrak n}}-1$ is even. Therefore, by using that $\zeta=O({\varepsilon})$ by Lemma \[lem:2.1\], one obtains $$F({\varepsilon},c) = \zeta {\varepsilon}^{{{\mathfrak n}}-1} a + G({\varepsilon},c) , \qquad a = g_{0} {{\mathfrak n}}\int_{0}^{1} {\rm d}s \left[ \left( X^{(1)}(\cdot) + s\zeta{\varepsilon}^{-1} \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} , \nonumber$$ with $a>0$ and $G({\varepsilon},c) =O({\varepsilon}^{{{\mathfrak n}}+1})$. Furthermore either all derivatives of $G({\varepsilon},c_{0})$ vanish at ${\varepsilon}=0$ (and hence $F_{k,0}=0$ for all $k\ge 0$) or $$G({\varepsilon},c_{0}) = b {\varepsilon}^{{{\mathfrak p}}} + O({\varepsilon}^{{{\mathfrak p}}+1}) , \qquad b = F_{{{\mathfrak p}},0} , \nonumber$$ for some ${{\mathfrak p}}>{{\mathfrak n}}$. In both cases, there is no solution corresponding to the segment $\{(0,{{\mathfrak n}}),({{\mathfrak n}}-1,1)\}$. Indeed, the bifurcation equation can be written as $${\varepsilon}^{{{\mathfrak n}}-1} \left( a_{0} \zeta + \Gamma({\varepsilon},\zeta) \right) = 0 , \qquad a_{0} = \left[ \left( X^{(1)}(\cdot) \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} , \nonumber$$ with the function $\Gamma({\varepsilon},\zeta)$ which is $C^{{\infty}}$ in both ${\varepsilon}$ and $\zeta$. Hence we can apply the implicit function theorem to deduce that for ${\varepsilon}\neq0$ there is one and only one solution $c=c_{0}+\zeta({\varepsilon})$ to the bifurcation equation, with $\zeta({\varepsilon})$ smooth in ${\varepsilon}$ and such that $\zeta({\varepsilon})=G({\varepsilon},c_{0})/a_{0}{\varepsilon}^{{{\mathfrak n}}-1}+ \tilde\zeta({\varepsilon})$ and $\tilde\zeta({\varepsilon})/\zeta({\varepsilon})\to0$ as ${\varepsilon}\to0$. Moreover one can explicitly estimate the interval $U$ around ${\varepsilon}=0$ in which such a solution exists (again by using the implicit function theorem and the bounds in [@G2])). If $F_{k,0}=0$ for all $k\ge0$ then the only solution to the bifurcation equation is $c=c_{0}+\zeta({\varepsilon})$, where $\zeta({\varepsilon})$ is a $C^{{\infty}}$ function which goes to zero faster than any power as ${\varepsilon}$ tends to $0$. If there exists ${{\mathfrak p}}\ge{{\mathfrak n}}+1$ such that $F_{k,0}=0$ for $k\le {{\mathfrak p}}$ and $F_{p,0}\neq0$, the only solution is of the form $c=c_{0} + \zeta_{0}{\varepsilon}^{{{\mathfrak p}}-{{\mathfrak n}}+1} + o({\varepsilon}^{{{\mathfrak p}}-{{\mathfrak n}}+1})$, where $\zeta_{0}$ is the (unique) solution of the equation $a \zeta + b = 0$. In the case $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ the following result holds. \[lem:2.3\] If $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq 0$ then all solutions of the bifurcation equation (\[eq:2.5\]) are of order ${\varepsilon}$ and have multiplicity 1. In particular, there exists a constant $a_{0}>0$ such that $|c_{i}({\varepsilon})-c_{j}({\varepsilon})|>a_{0}|{\varepsilon}|$ for all pairs of such solutions $c_{i}({\varepsilon})$ and $c_{j}({\varepsilon})$. [[*Proof.* ]{}]{}If $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ then the equation $$\left[ \left( \zeta + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} = 0 , \label{eq:2.8}$$ with $X^{(1)}$ defined in (\[eq:2.4\]), admits at least one non-zero real solution $\zeta_{1}$ of order ${\varepsilon}$. Thus, one can write $$\left( \zeta + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} = \left( \zeta - \zeta_{1} + \zeta_{1} + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} = \sum_{k=0}^{{{\mathfrak n}}} \left( \begin{matrix} {{\mathfrak n}}\\ k \end{matrix} \right) \left( \zeta - \zeta_{1} \right)^{{{\mathfrak n}}-k} \left( \zeta_{1} + {\varepsilon}X^{(1)}(\cdot) \right)^{k} , \nonumber$$ so that, by using that $[(\zeta_{1} + {\varepsilon}X^{(1)} (\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}} = 0$, one has $$\left[ \left( \zeta + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} = \left( \zeta - \zeta_{1} \right) \sum_{k=0}^{{{\mathfrak n}}-1} \left( \begin{matrix} {{\mathfrak n}}\\ k \end{matrix} \right) \left( \zeta - \zeta_{1} \right)^{{{\mathfrak n}}-1-k} \left[ \left( \zeta_{1} + {\varepsilon}X^{(1)}(\cdot) \right)^{k} \right]_{{\boldsymbol{0}}} . \nonumber$$ For $\zeta=\zeta_{1}$ one has $$\sum_{k=0}^{{{\mathfrak n}}-1} \left( \begin{matrix} {{\mathfrak n}}\\ k \end{matrix} \right) \left( \zeta - \zeta_{1} \right)^{{{\mathfrak n}}-1-k} \left[ \left( \zeta_{1} + {\varepsilon}X^{(1)}(\cdot) \right)^{k} \right]_{{\boldsymbol{0}}} = \left( \begin{matrix} {{\mathfrak n}}\\ {{\mathfrak n}}-1 \end{matrix} \right) \left[ \left( \zeta_{1} + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} > 0 , \nonumber$$ which shows that $\zeta_{1}$ is a simple root of the equation (\[eq:2.8\]), and hence also of the bifurcation equation (\[eq:2.5\]). The following result extends Lemma \[lem:2.2\], and shows that the bifurcation equation admits a unique solution for any $f$. \[lem:2.4\] The bifurcation equation (\[eq:2.5\]) admits one and only one solution $c({\varepsilon})$. One has $c({\varepsilon})=\zeta_{1}{\varepsilon}+O({\varepsilon}^{2})$, with $\zeta_{1}\neq0$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ and $c({\varepsilon})=O({\varepsilon}^{2})$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$. [[*Proof.* ]{}]{}By Lemma \[lem:2.2\] we can confine ourselves to the case $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$. In that case we know by Lemma \[lem:2.3\] that all the solutions of the bifurcation equation are of order ${\varepsilon}$ and separated by order ${\varepsilon}$. Hence we can set $\zeta=\zeta_{1}{\varepsilon}+ \zeta'$, with $\zeta'=o({\varepsilon})$, so that one has $$\begin{aligned} \frac{\partial}{\partial c} F({\varepsilon},c_{0}+\zeta) & \!\!\! = \!\!\! & g_{0} \frac{\partial}{\partial \zeta} \left[ \left( \zeta + {\varepsilon}X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} + O({\varepsilon}^{{{\mathfrak n}}+1}) \nonumber \\ & \!\!\! = \!\!\! & g_{0} {\varepsilon}^{{{\mathfrak n}}} {{\mathfrak n}}\left[ \left( \zeta_{1} + X^{(1)}(\cdot) \right)^{{{\mathfrak n}}-1} \right]_{{\boldsymbol{0}}} + O( {\varepsilon}^{{{\mathfrak n}}+1}) , \nonumber \end{aligned}$$ which is strictly positive for ${\varepsilon}>0$ small enough. Thus, at fixed ${\varepsilon}$, $F({\varepsilon},c)$ is increasing in $c$: this yields that the solution $c=c({\varepsilon})$ of $F({\varepsilon},c)=0$ is unique for ${\varepsilon}$ small enough. Multiscale analysis and diagrammatic expansion {#sec:3} ============================================== In this section we develop a diagrammatic representation for the quasi-periodic solution. There are a few differences with respect to [@G2], as we expand simultaneously both $X$ and $c$ in terms of the perturbation parameter ${\varepsilon}$. As in [@G2], the expansions we shall find are not power series expansions. See also [@GGG; @CG] for analogous situations; note, however, that in our case, as in [@GCB], no fractional expansions arise. We can summarise the results of the previous section as follows. The bifurcation equation (\[eq:2.5\]) admits a unique solution $c({\varepsilon})=c_{0}+\zeta({\varepsilon})$ such that $\zeta(0)=0$. Let $X^{(1)}$ the zero-average function such that $\dot X^{(1)}=f$. If $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ then one has $\zeta({\varepsilon})=\zeta_{1}{\varepsilon}+ o({\varepsilon})$, where ${\varepsilon}\zeta_{1}$ is the unique solution of the equation (\[eq:2.8\]). To make notation uniform we can set in the following $\zeta_{1}=0$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$. By defining $$P_{{{\mathfrak n}}}(\zeta)= \left[ \left( \zeta + X^{(1)}(\cdot) \right)^{{{\mathfrak n}}} \right]_{{\boldsymbol{0}}} , \qquad P_{{{\mathfrak n}}}'(\zeta) = \frac{{\rm d}}{{\rm d}\zeta} P_{{{\mathfrak n}}}(\zeta) , \label{eq:3.0}$$ if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}\neq0$ one has $P_{{{\mathfrak n}}}'(\zeta_{1}) \neq 0$ by Lemma \[lem:2.3\], and if $[(X^{(1)}(\cdot))^{{{\mathfrak n}}}]_{{\boldsymbol{0}}}=0$ one has $P_{{{\mathfrak n}}}'(\zeta_{1})=P_{{{\mathfrak n}}}'(0) ={{\mathfrak n}}[(X^{(1)}(\cdot))^{{{\mathfrak n}}-1}]_{{\boldsymbol{0}}}\neq0$ because ${{\mathfrak n}}$ is odd. Set $a=g_{0}P_{{{\mathfrak n}}}'(\zeta_{1})$ in both cases; then $a \neq 0$. In the following we give the diagrammatic rules in detail to make the exposition self-contained, and we only stress where the main differences lie with respect to the expansion of [@G2]. From a technical point of view, the diagrammatic analysis turns out a bit more involved, as it requires further expansions, and hence more labels to be assigned to the diagrams. On the other hand, eventually there is the advantage that one has a completely constructive algorithm to determine the solution within any fixed accuracy and an explicitly computable value for the maximal size of the allowed perturbation. A graph is a connected set of points and lines. A *tree* $\theta$ is a graph with no cycle, such that all the lines are oriented toward a unique point (*root*) which has only one incident line (root line). All the points in a tree except the root are called *nodes*. The orientation of the lines in a tree induces a partial ordering relation ($\preceq$) between the nodes. Given two nodes $v$ and $w$, we shall write $w \prec v$ every time $v$ is along the path (of lines) which connects $w$ to the root. We call $E(\theta)$ the set of *end nodes* in $\theta$, that is the nodes which have no entering line, and $V(\theta)$ the set of *internal nodes* in $\theta$, that is the set of nodes which have at least one entering line. Set $N(\theta)= E(\theta) \amalg V(\theta)$. With each end node $v$ we associate a *mode* label ${\boldsymbol{\nu}}_{v}\in{\mathds{Z}}^{d}$. For all $v\in N(\theta)$ denote with $s_{v}$ the number of lines entering the node $v$; for $v\in V(\theta)$ one has $s_{v} \ge {{\mathfrak n}}$. With respect to [@G2] the mode label of the end nodes can be ${\boldsymbol{0}}$. This only occurs when $\zeta_{1}\neq0$. Hence we define $E_{0}(\theta)=\{v\in E(\theta) : {\boldsymbol{\nu}}_{v}={\boldsymbol{0}}\}$ and $E_{1}(\theta)=\{v\in E(\theta) : {\boldsymbol{\nu}}_{v} \neq {\boldsymbol{0}}\}$; we can set $E_{0}(\theta)=\emptyset$ if $\zeta_{1}=0$. Define also $L_{0}(\theta)=\{\ell\in L(\theta) : \ell \hbox{ exits a node } v\in E_{0}(\theta)\}$. If $|N(\theta)|=1$ one requires $E_{1}(\theta)=N(\theta)$ and hence $E_{0}(\theta)=\emptyset$. We denote with $L(\theta)$ the set of lines in $\theta$. Since a line $\ell$ is uniquely identified with the node $v$ which it leaves, we may write $\ell = \ell_{v}$. With each line $\ell$ we associate a *momentum* label ${\boldsymbol{\nu}}_{\ell} \in {\mathds{Z}}^{d}$ and a *scale* label $n_{\ell}\in{\mathds{Z}}_{+} \cup \{-1\}$. We set $n_{\ell}=-1$ when ${\boldsymbol{\nu}}_{\ell}={\boldsymbol{0}}$ (note that ${\boldsymbol{\nu}}_{\ell}={\boldsymbol{0}}$ was not allowed in [@G2]). The modes of the end nodes and the momenta of the lines are related as follows: if $\ell = \ell_{v}$ one has $${\boldsymbol{\nu}}_{\ell} = \sum_{w \in E(\theta) : w \preceq v} {\boldsymbol{\nu}}_{w} . \label{eq:3.1}$$ Given a tree $\theta$ we set $\Lambda_{0}(\theta)= \{\ell\in L(\theta): {\boldsymbol{\nu}}_{\ell}={\boldsymbol{0}}\} \setminus L_{0}(\theta) = \{\ell\in L(\theta): {\boldsymbol{\nu}}_{\ell}={\boldsymbol{0}}$ and $\ell$ does not exit an end node$\}$, and define the *order* of $\theta$ as $$k(\theta) = |N(\theta)| - {{\mathfrak n}}\left| \Lambda_{0}(\theta) \right| , \label{eq:3.2}$$ and the *total momentum* of $\theta$ as ${\boldsymbol{\nu}}(\theta)= {\boldsymbol{\nu}}_{\ell_{0}}$, if $\ell_{0}$ is the root line of $\theta$. We call *equivalent* two trees which can be transformed into each other by continuously deforming the lines in such a way that they do not cross each other. Let ${{\mathcal T}}_{k,{\boldsymbol{\nu}}}$ be the set of inequivalent trees of order $k$ and total momentum ${\boldsymbol{\nu}}$, that is the set of inequivalent trees $\theta$ such that $k(\theta)=k$ and ${\boldsymbol{\nu}}(\theta)={\boldsymbol{\nu}}$. A *cluster* $T$ on scale $n$ is a maximal set of nodes and lines connecting them such that all the lines have scales $n'\le n$ and there is at least one line with scale $n$. The lines entering the cluster $T$ and the possible line coming out from it (unique if existing at all) are called the external lines of the cluster $T$. Given a cluster $T$ on scale $n$, we shall denote by $n_{T}=n$ the scale of the cluster. We call $V(T)$, $E(T)$, and $L(T)$ the set of internal nodes, of end nodes, and of lines of $T$, respectively; the external lines of $T$ do not belong to $L(T)$. We call *self-energy cluster* any cluster $T$ such that $T$ has only one entering line $\ell_{T}^{2}$ and one exiting line $\ell_{T}^{1}$, and one has $\sum_{v\in E(T)} {\boldsymbol{\nu}}_{v} = {\boldsymbol{0}}$ (and hence ${\boldsymbol{\nu}}_{\ell_{T}^{1}}={\boldsymbol{\nu}}_{\ell_{T}^{2}}$ by (\[eq:3.1\])). Set $x_{T}={\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}_{\ell_{T}^{2}}$. We call *excluded* a node $v$ such that ${\boldsymbol{\nu}}_{\ell_{v}}= {\boldsymbol{0}}$, $s_{v}={{\mathfrak n}}$, at least ${{\mathfrak n}}-1$ lines entering $v$ do exit end nodes, and the other line $\ell'$ entering $v$ either also exits an end node or has momentum ${\boldsymbol{\nu}}_{\ell'}={\boldsymbol{0}}$. Let ${{\mathfrak T}}_{k,{\boldsymbol{\nu}}}$ be the set of *renormalised trees* in ${{\mathcal T}}_{k,{\boldsymbol{\nu}}}$, i.e. of trees in ${{\mathcal T}}_{k,{\boldsymbol{\nu}}}$ which contain neither any self-energy clusters nor any excluded nodes. Define also ${{\mathfrak R}}_{n}$ as the set of renormalised self-energy clusters on scale $n$, i.e. of self-energy clusters on scale $n$ which contain neither any further self-energy clusters nor any excluded nodes. \[lem:3.1\] One has $k(\theta) \ge 1$ if ${\boldsymbol{\nu}}(\theta)\neq{\boldsymbol{0}}$ and $k(\theta) \ge 2$ if ${\boldsymbol{\nu}}(\theta)={\boldsymbol{0}}$. [[*Proof.* ]{}]{}By induction on $|N(\theta)|$. If $|N(\theta)|=1$ then $N(\theta)=E_{1}(\theta)$ and hence $\Lambda_{0}(\theta)=\emptyset$, so that $k(\theta)=1$. If $|N(\theta)|>1$ let $v_{0}$ be the last node of $\theta$, that is the node which the root line of $\theta$ exits, and let $\theta_{1},\ldots,\theta_{s}$ the trees with root in $v_{0}$. Note that $s\ge {{\mathfrak n}}$. Then one has $|N(\theta)|=1+|N(\theta_{1})|+\ldots+|N(\theta_{s})|$, while $|\Lambda_{0}(\theta)|=|\Lambda_{0}(\theta_{1})|+\ldots +|\Lambda_{0}(\theta_{s})|$ if ${\boldsymbol{\nu}}(\theta)\neq{\boldsymbol{0}}$ and $|\Lambda_{0}(\theta)|=1+|\Lambda_{0}(\theta_{1})|+\ldots +|\Lambda_{0}(\theta_{s})|$ if ${\boldsymbol{\nu}}(\theta)={\boldsymbol{0}}$. In the first case one has $k=k(\theta)=1+k(\theta_{1})+\ldots+ k(\theta_{s})\ge 1+{{\mathfrak n}}$. In the second case one has $k=k(\theta)=1+k(\theta_{1})+\ldots+ k(\theta_{s}) - {{\mathfrak n}}\ge 1+s-{{\mathfrak n}}$, so that $k\ge1$. Moreover $k=1$ would be possible only if $s={{\mathfrak n}}$ and $k(\theta_{1})=\ldots=k(\theta_{s})=1$. However, in such a case the lines entering $v_{0}$ would all exit end nodes and hence $v_{0}$ would be an excluded node. Thus, $k\ge2$ if ${\boldsymbol{\nu}}(\theta)={\boldsymbol{0}}$. \[lem:3.2\] There exists a positive constant $\kappa$ such that $|N(\theta)| \le \kappa k(\theta)$ for any renormalised tree $\theta$. One can take $\kappa=3{{\mathfrak n}}$. [[*Proof.* ]{}]{}We give the proof for ${{\mathfrak n}}>1$ (the case ${{\mathfrak n}}=1$ being much easier; see [@GBD1]). For $k(\theta)=1$ the bound is trivially satisfied. We prove that for all $k\ge 2$, all ${\boldsymbol{\nu}}\in{\mathds{Z}}^{d}$ and all trees $\theta\in{{\mathfrak T}}_{k,{\boldsymbol{\nu}}}$ one has $|N(\theta)| \le a k(\theta)-b$, with $a=3{{\mathfrak n}}$ and $b=4{{\mathfrak n}}$. The proof is by induction on $k$. For $k=2$ it is just a check: if ${\boldsymbol{\nu}}\ne{\boldsymbol{0}}$ one has $N(\theta)=2$, while if ${\boldsymbol{\nu}}={\boldsymbol{0}}$ one has $N(\theta)=2+{{\mathfrak n}}\le 2{{\mathfrak n}}$ for ${{\mathfrak n}}>1$. By assuming that the bound holds for all $k< h$, one has for $\theta\in{{\mathfrak T}}_{h,{\boldsymbol{\nu}}}$ $$|N(\theta)| = s_{1} + \sum_{i=1}^{s_{0}} |N(\theta_{i})| , \qquad |\Lambda_{0}(\theta)| \le 1 + \sum_{i=1}^{s_{0}} |\Lambda_{0}(\theta_{i})| , \nonumber$$ where $\ell_{1},\ldots,\ell_{s_{0}}$ are the lines in $\Lambda_{0}(\theta)$ closest to the root line of $\theta$, $\theta_{1},\ldots,\theta_{s_{0}}$ are the trees with root lines $\ell_{1},\ldots,\ell_{s_{0}}$, respectively, and $s_{1}$ is the number of nodes which precede the root line of $\theta$ but not the root lines of $\theta_{1},\ldots,\theta_{s_{0}}$. By using the definition of order (\[eq:3.2\]) one has $$h = k(\theta) \ge s_{1} - {{\mathfrak n}}+ \sum_{i=1}^{s_{0}} k(\theta_{i}) . \nonumber$$ Note that $k(\theta_{i})\ge 2$ by Lemma \[lem:3.2\], and hence $|N(\theta_{i})| \le a k(\theta_{i})-b$ by the inductive hypothesis. If $s_{0}=0$ then $\Lambda_{0}(\theta) \le 1$, so that $|N(\theta)| \le k(\theta) + {{\mathfrak n}}= h + {{\mathfrak n}}\le 3{{\mathfrak n}}h - 4{{\mathfrak n}}$ for ${{\mathfrak n}}>1$ and $h>1$. If $s_{0} \ge 1$ the inductive hypothesis yields $$\begin{aligned} |N(\theta)| & \!\!\! \le \!\!\! & s_{1} + a \sum_{i=1}^{s_{0}} k(\theta_{i}) - s_{0} b \nonumber \\ & \!\!\! \le \!\!\! & a k(\theta) - b - \Big( (a-1) s_{1} + ( s_{0} - 1 ) b - a{{\mathfrak n}}\Big) . \nonumber \end{aligned}$$ Thus, the assertion follows if $$\left( a-1 \right) s_{1} + \left( s_{0} - 1 \right) b - a {{\mathfrak n}}= a \left( s_{1} + s_{0} - {{\mathfrak n}}\right) + \left( b-a \right) \left( s_{0}-1 \right) - s_{1} - a \ge 0 . \nonumber$$ A key remark is that $s_{0}+s_{1} = {{\mathfrak n}}+ 1 + p$, with $p \ge 0$, because the root line of $\theta$ exits a node $v_{0}\in V(\theta)$, so that $s_{v_{0}} \ge {{\mathfrak n}}$. Then we can rewrite $$a \left( s_{1} + s_{0} - {{\mathfrak n}}\right) + \left( b-a \right) \left( s_{0}-1 \right) - s_{1} - a = a p - {{\mathfrak n}}- p - 1 + s_{0} + \left( b - a \right) \left( s_{0}-1 \right) . \nonumber$$ If $p=0$ then $s_{1}+s_{0}={{\mathfrak n}}+1$; this yields $s_{0} \ge 2$ (otherwise $v_{0}$ would be an excluded node), so that $a p - {{\mathfrak n}}- p - 1 + s_{0} + \left( b - a \right) \left( s_{0}-1 \right) \ge 1 + \left( b - a \right) - {{\mathfrak n}}\ge 1$. If $p \ge 1$ then $a p - {{\mathfrak n}}- p - 1 + s_{0} + \left( b - a \right) \left( s_{0}-1 \right) \ge a p - {{\mathfrak n}}- p = {{\mathfrak n}}p + {{\mathfrak n}}\left( p - 1 \right) + p \left( {{\mathfrak n}}- 1 \right) \ge 2{{\mathfrak n}}-1$. Let $\psi$ be a non-decreasing $C^{\infty}$ function defined in ${\mathds{R}}_{+}$, such that $$\psi(u) = \left\{ \begin{array}{ll} 1 , & \text{for } u \geq 1 , \\ 0 , & \text{for } u \leq 1/2 , \end{array} \right. \label{eq:3.3}$$ and set $\chi(u) := 1-\psi(u)$. For all $n \in {\mathds{Z}}_{+}$ define $\chi_{n}(u) := \chi(u/4{\alpha}_{n}({\boldsymbol{\omega}}))$ and $\psi_{n}(u) := \psi(u/4{\alpha}_{n}({\boldsymbol{\omega}}))$, and set $\Xi_{0}(x)=\chi_{0}(|x|)$, $\Psi_{0}(x)=\psi_{0}(|x|)$, and $$\Xi_{n}(x)=\chi_{0}(|x|)\ldots \chi_{n-1}(|x|) \chi_{n}(|x|) , \qquad \Psi_{n}(x)=\chi_{0}(|x|)\ldots \chi_{n-1}(|x|) \psi_{n}(|x|) , \label{eq:3.4}$$ for $n\ge 1$. Then we define the node factor as $$F_{v} = \begin{cases} - \displaystyle{ \frac{1}{s_{v}!} \frac{{\rm d}^{s_{v}}}{{\rm d}x^{s_{v}}} g(c_{0}) } , & v \in V(\theta) , \\ f_{{\boldsymbol{\nu}}_{v}} , & v \in E_{1}(\theta) , \\ \zeta_{1} , & v \in E_{0}(\theta) , \end{cases} \label{eq:3.5}$$ and the propagator as $$G_{\ell} = \begin{cases} G^{[n_{\ell}]}({\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}_{\ell};{\varepsilon},c_{0}) , & {\boldsymbol{\nu}}_{\ell} \neq {\boldsymbol{0}}, \\ - \displaystyle{ \frac{1}{a} } , & {\boldsymbol{\nu}}_{\ell} = {\boldsymbol{0}}, \quad \ell \in \Lambda_{0}(\theta) , \\ 1 , & {\boldsymbol{\nu}}_{\ell} = {\boldsymbol{0}}, \quad \ell \in L_{0}(\theta) , \end{cases} \label{eq:3.6}$$ with $\zeta_{1}$ and $a$ defined before and after (\[eq:3.0\]), respectively, and $G^{[n]}(x;{\varepsilon},c_{0})$ recursively defined for $n\ge 0$ as $$\begin{aligned} & \hskip-.3truecm G^{[n]}(x;{\varepsilon},c) = \frac{\Psi_{n}(x)}{ {{\rm i}}x(1+{{\rm i}}{\varepsilon}x) - {{\mathcal M}}^{[n-1]}(x;{\varepsilon},c)} , \label{eq:3.7a} \\ & \hskip-.3truecm {{\mathcal M}}^{[n]}(x;{\varepsilon},c) = {{\mathcal M}}^{[n-1]}(x;{\varepsilon},c) + \Xi_{n}(x) M^{[n]}(x;{\varepsilon},c) , \quad M^{[n]}(x;{\varepsilon},c) = \!\! \sum_{T\in{{\mathfrak R}}_{n}}\operatorname{Val}(T,x;{\varepsilon},c) , \label{eq:3.7b}\end{aligned}$$ \[eq:3.7\] where ${{\mathcal M}}^{[-1]}(x;{\varepsilon},c_{0})={\varepsilon}g_{1}(c_{0})=0$ and $$\operatorname{Val}(T,x;{\varepsilon},c_{0}) = \Big(\prod_{\ell \in L(T)} G_{\ell} \Big) \Big( \prod_{v \in N(T)} F_{v} \Big) \label{eq:3.8}$$ is called the value of the self-energy cluster $T$. Note that, with respect to the analogous formulae (3.5) and (3.6) of [@G2], both $F_{v}$ and $G_{\ell}$ are computed at $c=c_{0}$. We prefer writing explicitly the dependence on $c_{0}$ (even if not necessary, as $c_{0}$ is fixed once and for all as in Assumption \[hyp:2\]) simply in order to use the same notations as in [@G2]. Set $$X^{[k]}_{{\boldsymbol{\nu}}} = \sum_{\theta\in {{\mathfrak T}}_{k,{\boldsymbol{\nu}}}} \operatorname{Val}(\theta;{\varepsilon},c_{0}) , \qquad {\boldsymbol{\nu}}\neq {\boldsymbol{0}}, \quad k \ge 1 , \label{eq:3.9}$$ and $$\zeta^{[k]}= \sum_{\theta\in {{\mathfrak T}}_{k+{{\mathfrak n}},{\boldsymbol{0}}}} \operatorname{Val}(\theta;{\varepsilon},c_{0}) , \qquad k \ge 2 , \label{eq:3.10}$$ where the tree value $\operatorname{Val}(\theta;{\varepsilon},c_{0})$ is defined as $$\operatorname{Val}(\theta,x;{\varepsilon},c_{0}) = \Big(\prod_{\ell \in L(\theta)} G_{\ell} \Big) \Big( \prod_{v \in N(\theta)} F_{v} \Big) . \label{eq:3.11}$$ Call ${{\mathfrak N}}_{n}(\theta)$ the number of lines $\ell\in L(\theta)$ such that $n_{\ell}\ge n$, and ${{\mathfrak N}}_{n}(T)$ the number of lines $\ell\in L(T)$ such that $n_{\ell}\ge n$, and set $$M(\theta) = \sum_{v \in E(\theta)} |{\boldsymbol{\nu}}_{v}| , \qquad M(T) = \sum_{v \in E(T)} |{\boldsymbol{\nu}}_{v}| . \label{eq:3.12}$$ Define the renormalised series $$\overline X({\boldsymbol{\psi}};{\varepsilon})= \sum_{{\boldsymbol{\nu}}\in{\mathds{Z}}^{d}_{*}} {\rm e}^{{{\rm i}}{\boldsymbol{\nu}}\cdot{\boldsymbol{\psi}}} \overline X_{{\boldsymbol{\nu}}} , \qquad \overline X_{{\boldsymbol{\nu}}} = \sum_{k=1}^{{\infty}} {\varepsilon}^{k} X_{{\boldsymbol{\nu}}}^{[k]} . \label{eq:3.14}$$ and $$\overline \zeta({\varepsilon}) = {\varepsilon}\zeta_{1} + \sum_{k=2}^{{\infty}} {\varepsilon}^{k} \zeta^{[k]} . \label{eq:3.15}$$ In the next section we shall prove first that the series (\[eq:3.14\]) and (\[eq:3.15\]) converge, then that the function $x_{0}(t) = c_{0} + \overline \zeta({\varepsilon}) + \overline X({\boldsymbol{\omega}}t;{\varepsilon})$ solves the equations $$\begin{aligned} & {{\rm i}}{\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}\left( 1 + {{\rm i}}{\varepsilon}{\boldsymbol{\omega}}\cdot{\boldsymbol{\nu}}\right) X_{{\boldsymbol{\nu}}} + {\varepsilon}\left[ g(c_{0}+\zeta({\varepsilon})+X(\cdot;{\varepsilon}) \right]_{{\boldsymbol{\nu}}} = {\varepsilon}f_{{\boldsymbol{\nu}}} , \qquad {\boldsymbol{\nu}}\neq {\boldsymbol{0}}, \label{eq:3.16a} \\ & [g(c_{0}+\zeta({\varepsilon})+X(\cdot;{\varepsilon})]_{{\boldsymbol{0}}} - f_{{\boldsymbol{0}}} = 0 , \label{eq:3.16b}\end{aligned}$$ \[eq:3.16\] -.3truecm and hence the equation (\[eq:1.1\]). Bounds {#sec:4} ====== The proof of the convergence of the renormalised series proceeds as in [@G2]. We confine ourselves to state the basic steps of the proof, without giving the details, except when the discussion departs from [@G2]. \[lem:4.1\] For any renormalised tree $\theta$, one has ${{\mathfrak N}}_{n}(\theta) \le 2^{-(n-2)}M(\theta)$. [[*Proof.* ]{}]{}The same as the proof of Lemma 3.1 in [@G2]. \[lem:4.2\] Assume there exists a constant $C_{0}$ such that $|G^{[n]}(x;{\varepsilon},c_{0})|\le C_{0}/{\alpha}_{n}({\boldsymbol{\omega}})$ for all $n\in{\mathds{Z}}_{+}$. Then there exists ${\varepsilon}_{0}>0$ such that, for all all $|{\varepsilon}|<{\varepsilon}_{0}$, the series (\[eq:3.14\]) and (\[eq:3.15\]) converge. Moreover the series (\[eq:3.14\]) is analytic in ${\boldsymbol{\psi}}$. [[*Proof.* ]{}]{}Let $\theta$ be a tree in ${{\mathfrak T}}_{k,{\boldsymbol{\nu}}}$ and set $K=|N(\theta)|$. One reasons as in [@G2] to prove that $$\left| \operatorname{Val}(\theta;{\varepsilon},c) \right| \le \widetilde C_{0}^{K} D_{0}^{K} {\alpha}_{n_{0}}^{-K}({\boldsymbol{\omega}}) {\rm e}^{-\xi M(\theta)/2} , \nonumber$$ where $D_{0}=\max\{\Gamma,\Phi\}$, $\widetilde C_{0}=\max\{C_{0},a\}$, and $n_{0} \in {\mathds{N}}$ such that $$4 \sum_{n=n_{0}+1}^{{\infty}} \frac{1}{2^{n}} \log \frac{1}{{\alpha}_{n}({\boldsymbol{\omega}})} \le \frac{\xi}{2} . \nonumber$$ By Lemma \[lem:3.2\] one has $K\le \kappa k$. The number of trees of order $k$ with fixed mode labels is hence bounded by $C_{2}^{k}$ for a suitable constant $C_{2}$, and the sum over the mode labels can be performed by using half the exponent in the dacaying factor ${\rm e}^{-\xi M(\theta)/2}$. This gives $$\left| X^{[k]}_{{\boldsymbol{\nu}}} \right| \le C^{k} {\rm e}^{-\xi|{\boldsymbol{\nu}}|/4} , \nonumber$$ for some constant $C$, which yields the convergence of the series (\[eq:3.14\]) to a function analytic in ${\boldsymbol{\psi}}$. In the same way one obtains $$\left| \zeta^{[k]} \right| \le C^{k} \nonumber$$ with the same constant $C$, and this yields the convergence of the series (\[eq:3.15\]). \[lem:4.3\] For any self-energy cluster $T\in{{\mathfrak R}}_{n}$ such that $\Xi_{n}(x_{T}) \neq 0$, one has $M(T) \ge 2^{n-1}$ and ${{\mathfrak N}}_{p}(T) \le 2^{-(p-2)}M(T)$ for all $p \le n$. [[*Proof.* ]{}]{}The same as the proof of Lemma 3.3 of [@G2]. \[lem:4.4\] Assume the propagators $G^{[p]}(x;{\varepsilon},c_{0})$ are differentiable in $x$ and there exist constants $C_{0}$ and $C_{1}$ such that $|G^{[p]}(x;{\varepsilon},c_{0})| \le C_{0}/{\alpha}_{p}({\boldsymbol{\omega}})$ and $|\partial_{x} G^{[p]}(x;{\varepsilon},c_{0})| \le C_{1}/{\alpha}_{p}^{3}({\boldsymbol{\omega}})$ for all $p<n$. Then there exists ${\varepsilon}_{0}>0$ such that, for all $|{\varepsilon}|<{\varepsilon}_{0}$, the function $x \mapsto M^{[n]}(x;{\varepsilon},c_{0})$ is differentiable, and one has $$\left| M^{[n]} (x;{\varepsilon},c_{0}) \right| , \left| \partial_{x} M^{[n]} (x;{\varepsilon},c_{0}) \right| \le D_{1} |{\varepsilon}|^{2} {\rm e}^{-D_{2}2^{n}} , \nonumber$$ for some positive constants $D_{1}$ and $D_{2}$. [[*Proof.* ]{}]{}Proceed as in the proof of Lemma 3.4 of [@G2], by taking into account the differences already pointed out in the proof of Lemma \[lem:4.2\]. \[lem:4.5\] Assume there exists a constant $C_{0}$ such that $|G^{[p]}(x;{\varepsilon},c_{0})|\le C_{0}/{\alpha}_{p}({\boldsymbol{\omega}})$ for all $p<n$. Then one has $({{\mathcal M}}^{[p]}(x;{\varepsilon},c_{0}))^{*} = {{\mathcal M}}^{[p]}(-x;{\varepsilon},c_{0})$ for all $p \le n$. [[*Proof.* ]{}]{}As the proof of Lemma 3.5 of [@G2]. \[lem:4.6\] Let ${\varepsilon}_{0}$ as in Lemma \[lem:4.1\]. For all $n\in{\mathds{Z}}_{+}$ the function $x \mapsto {{\mathcal M}}^{[n]}(x;{\varepsilon},c_{0}))$ is differentiable and one has $|{{\rm i}}x (1 + {{\rm i}}{\varepsilon}x) - {{\mathcal M}}^{[n]}(x;{\varepsilon},c_{0})|\ge |x|/2$ for all $|{\varepsilon}|<{\varepsilon}_{0}$. [[*Proof.* ]{}]{}As the proof of Lemma 3.6 of [@G2]. By collecting together the results above we can prove that the series (\[eq:3.14\]) and (\[eq:3.15\]) converge. The proof proceeds as follows. Convergence holds if the hypotheses of Lemma \[lem:4.2\] are verified. So we have to prove that the propagators satisfy the bounds $|G^{[n]}(x;{\varepsilon},c_{0})| \le C_{0}/{\alpha}_{n}({\omega})$. More precisely, we can prove by induction on $n$ that the propagators satisfy the bounds $$\left| G^{[n]}(x;{\varepsilon},c_{0}) \right| \le \frac{C_{0}}{{\alpha}_{n}({\boldsymbol{\omega}})} , \qquad \left| \partial_{x} G^{[n]}(x;{\varepsilon},c_{0}) \right| \le \frac{C_{1}}{{\alpha}_{n}^{3}({\boldsymbol{\omega}})} , \label{eq:4.1}$$ for suitable constants $C_{0}$ and $C_{1}$. For $n=0$ the check is trivial. Then we assume that the bounds are satisfied up to scale $n-1$, and we write $G^{[n]}(x;{\varepsilon},c_{0})$ according to (\[eq:3.7\]). Now, $G^{[n]}(x;{\varepsilon},c_{0})$ depends on the quantities ${{\mathcal M}}^{[p]}(x;{\varepsilon},c_{0})$ with $p < n$, which in turn depend on the propagators $G^{[p']}(x;{\varepsilon},c_{0})$ with $p'\le p$. Thus, by the inductive hypothesis we can apply Lemma \[lem:4.4\] and Lemma \[lem:4.5\] to deduce Lemma \[lem:4.6\]. This implies that also $G^{[n]}(x;{\varepsilon},c_{0})$ satisfies the bounds (\[eq:4.1\]). Smoothness in ${\varepsilon}$ of the series is discussed in the same way, proving by induction that the derivatives of the propagators satisfy the bounds $|\partial_{{\varepsilon}}^{m}G^{[n]}(x;{\varepsilon},c_{0})| \le K_{m}/{\alpha}_{n}^{m}({\boldsymbol{\omega}})$ for all $m\ge1$ and for suitable constants $K_{m}$ (we refer to [@G2] for further details). Finally, by proceeding as in the proof of Lemma 3.7 of [@G2], we can prove that the function $x_{0}(t)=c_{0}+\zeta({\varepsilon})+ X({\boldsymbol{\omega}}t;{\varepsilon})$ solves the equations (\[eq:3.16\]). As far as equation (\[eq:3.16a\]) is concerned the proof proceeds as in the proof of Lemma 3.7 of [@G2]. To deal with equation (\[eq:3.16b\]), we write $[g(c_{0}+\zeta({\varepsilon})+X(\cdot;{\varepsilon})]_{{\boldsymbol{0}}}$ according to (\[eq:2.6\]) and expand $X(\cdot;{\varepsilon})=\overline X(\cdot;{\varepsilon})$ according to (\[eq:3.14\]) and (\[eq:3.9\]). Then we obtain $$g_{0} P_{{{\mathfrak n}}}'(\zeta_{1}) \left( \zeta_{k} + \sum_{\theta\in {{\mathfrak T}}_{k+{{\mathfrak n}},{\boldsymbol{0}}}} \operatorname{Val}(\theta;{\varepsilon},c_{0}) , \right) \qquad k \ge 2 , \label{eq:4.2}$$ which is identically zero if $\zeta_{k}$ is defined by (\[eq:3.10\]). [99]{}

--- abstract: 'We establish a general “boundedness implies convergence” principle for a family of evolving Riemannian metrics. We then apply this principle to collapsing Calabi-Yau metrics and normalized Kähler-Ricci flows on torus fibered minimal models to obtain convergence results.' address: - 'School of Mathematical Sciences, Peking University, Beijing, China 100871' - 'Department of Mathematics, Nanjing University, Nanjing, China 210093' author: - Wangjian Jian - Yalong Shi title: 'A “boundedness implies convergence” principle and its applications to collapsing estimates in Kähler geometry' --- Introduction {#intro} ============ In differential geometry, we usually try to find “canonical” geometric objects, e.g. Einstein metrics, minimal surfaces, etc., through a deformation process. Starting with a given object, we deform it through a suitable geometric flow or continuity path to get the desired metric or submanifold. To obtain convergence, we need a priori bounds for the higher order covariant derivatives of the evolving functions, metrics or curvature tensors. In general, such bounds are invalid and singularities may form. However, if we have curvature bounds or non-collapsing condition, general theories from Riemannian geometry and PDE usually gave us a lot of information about the formation of singularities and in the end we still get good geometric results. On the other hand, there are some important cases in which we lose curvature bounds and the metrics may collapse to lower dimensional spaces. In such cases, a priori bounds play a more decisive role. It is still interesting to know whether certain tensor fields converge to some degenerate or singular one. In this paper, we prove a simple but useful “Boundedness Implies Convergence" lemma, which says that in certain situation if we have the convergence of an evolving tensor field together with uniform estimates of its covariant derivatives, then we automatically have the convergence of its covariant derivatives. In spirit, this is similar to the simple fact in calculus that if a smooth function on ${\mathbb{R}}^n$ has bounded partial derivatives of all orders and if the function converges to 0 at infinity, then all its partial derivatives converge to 0 at infinity. Indeed, if we have uniform equivalence of metrics, this follows easily from interpolation type inequalities. Instead, here we assume the existence of good cut-off functions. In applications, such cut-off functions usually arise as the pull back of cut-off functions from the base manifold of a fibration. To be precise, we have: \[Thm: BIC principle\] Let $X$ be an n-dimension Riemannian manifold (not necessarily to be compact or complete) and $U$ be an open subset. Let ${\tilde{g}}(t)$ be a family of Riemannian metrics on $X$, $t\in {\mathbb{R}}$ and let $\eta(t)$ be a family of smooth functions or general tensor fields on $X$, satisfying the following conditions: There exists positive constants $A_1, A_2, \dots$, and a positive function $h_0(t)$ which tends to zero as $t\to \infty$ such that 1.  $\|\eta(t)\|_{C^0(U,{\tilde{g}}(t))}\leq h_0(t).$ 2.  $\|\eta(t)\|_{C^k(U,{\tilde{g}}(t))}\leq A_k, ~for~ $k=1,2,…$~. $ 3.  For any compact subset $K\subset\subset U$, there exists smooth cut-off function $\rho$ with compact support $\hat{K}\subset\subset U$ such that $0\leq\rho\leq 1$, and $\rho\equiv 1$ in a neighborhood of $K$, satisfying $$\label{Equ: cutoff bound} \left|\nabla{\rho}\right|_{{\tilde{g}}(t)}^2+|\Delta_{{\tilde{g}}(t)}\rho|\leq B_{K}.$$ on $\hat{K}\times [0,\infty)$, for some constant $B_{K}$ independent of $t$ (but may depend on the geometry of $K$). Then we have: For any compact subset $K\subset\subset U$ the estimates $$\label{Equ: convergence in general} \|\eta(t)\|_{C^k(K,{\tilde{g}}(t))}\leq h_{K,k}(t).$$ where $h_{K,k}(t)$ are positive functions which tend to zero as $t\to \infty$, depenging on the constants $A_0$, $A_1$, $\dots$, $A_{k+2}$, $B_K$ and the function $h(t)$. 1. Lemma \[Thm: BIC principle\] also holds true for discrete sequences of metrics and tensors. This is easily seen from the proof in section \[sec: pf of BIC\]. 2. In Condition $(A)$ if $h_0(t)$ is of the form $Ce^{-ct}$ for some positive constants $C<\infty$ and $c>0$, then the functions $h_{K,k}(t)$ can be chosen to be $C_{K,k}e^{-c_{K,k} t}$ for some constants $C_{K,k}, c_{K,k}$, i.e. covariant derivatives of $\eta(t)$ also decay exponentially. 3. In Lemma \[Thm: BIC principle\] if we only have Condition $(B)$ for $1\leq k\leq N+2$, then we still have the estimate for $1\leq k\leq N$. Note that in Lemma \[Thm: BIC principle\] we do not require the metrics to be uniformly non-degenerate or have bounded curvature and do not require the metrics and tensors satisfy differential equations. So it applies even when the metrics collapse to lower dimensional spaces without curvature bounds. In particular, we shall apply this principle to two collapsing problems in Kähler geometry. We hope this simple “BIC principle” may find other applications in geometric analysis, especially in collapsing problems.\ The first application in this paper is on Calabi-Yau degenerations. Given a Calabi-Yau manifold $X$ with a holomorphic fiber space structure (i.e. there is a holomorphic surjection $f$ onto a lower dimensional variety, which is a holomorphic submersion outside a subvariety), the Calabi-Yau theorem [@Yau] assures the existence of unique Ricci-flat Kähler metrics on the total space in every Kähler classes. Now let the Kähler classes approach the pull back of a Kähler class from the base, then the volume of these Ricci-flat metrics go to zero. One would like to understand the asymptotic behaviors of these degenerating metrics. This problem has been much studied in recent years, starting from the pioneering work of Gross-Wilson [@GW] on elliptic $K3$ surfaces fibered over the 2-sphere, to more recent works in general dimensions by Tosatti [@To], Tosatti-Zhang [@TZ; @TZ2], Gross-Tosatti-Zhang [@GTZ; @GTZ2], Hein-Tosatti [@HT], Tosatti-Weinkove-Yang [@TWY], Li [@Li], and elsewhere. From these works, we know that the Ricci-flat metrics collapse, in some sense, to the pullback of a canonical Kähler metric on the base, uniformly on compact sets away from the singular fibers. In this general setting, a $C^0_{\rm loc}$ estimate was proved by Tosatti-Weinkove-Yang in [@TWY], and this estimate can be improved to $C^\infty_{\rm loc}$ estimate when the smooth fibers are tori or finite étale quotients of tori by Gross-Tosatti-Zhang [@GTZ] and Hein-Tosatti [@HT]. Certain components of the first order derivatives were bounded by Tosatti [@To] and Tosatti-Weinkove-Yang in [@TWY]. An even stronger partial estimate was proved by Tosatti-Zhang in [@TZ]: the restriction of $e^t{\omega}(t)$ to $f^{-1}(U)$ converges in the pointed $C^\infty$ Cheeger-Gromov topology to the product of a flat $\mathbb{C}^m$ with a fiber equipped with Ricci-flat metric. In [@HT2], with a systematic use of iterated blow-up-and-contradiction type arguments, Hein-Tosatti substantially improved the estimate to $C^\infty_{{\rm loc}}$ if the regular fibers are pairwise bi-holomorphic to each other. In the general fibration case, they can improve the $C^0$ convergence of [@TWY] to $C^\alpha$ convergence. In a more recent preprint [@STZ], Song-Tian-Zhang proved the uniform diameter bound and the Gromov-Hausdorff convergence of this family of collapsing metrics. In this paper, we derive from Hein-Tosatti’s $C^\infty_{{\rm loc}}$ estimate the corresponding convergence results, and such estimates also imply the $C^\infty_{{\rm loc}}$ asymptotic behavior of the curvature tensor. To state our results, let $X$ be a compact Kähler $(n+m)$-manifold with $c_1(X)=0$ in $H^2(X,\mathbb{R})$ (i.e. a Calabi-Yau manifold), and let ${\omega}_X$ be a Ricci-flat Kähler metric on $X$. Suppose that we have a holomorphic map $f:X\to Z$ with connected fibers, where $(Z,{\omega}_Z)$ is a compact Kähler manifold, with image $B=f(X)\subset Z$ an irreducible normal subvariety of $Z$ of dimension $m>0$. Then the induced surjective map $f:X\to B$ is a fiber space, and if $S'\subset B$ denotes the singular set of $B$ together with the set of critical values of $f$, and $S=f^{-1}(S')$, then $S'$ is a proper analytic subvariety of $B$, $S$ is a proper analytic subvariety of $X$, and $f:X\backslash S\to B\backslash S'$ is a submersion between smooth manifolds. The fibers $X_b$ for $b \in B \backslash S'$ are Calabi-Yau $n$-folds. Write $\chi=f^*\omega_Z$, which is a smooth nonnegative $(1,1)$ form on $X$, and we will also write $\chi$ for the restriction of $\omega_Z$ to $B\backslash S'$. Let ${\omega}_B$ be the current in $[\chi]$ that is smooth and positive on $B\backslash S'$ satisfying $${\mathrm{Ric}}({\omega}_B)={\omega}_{{\rm WP}},$$ where ${\omega}_{{\rm WP}}$ is a smooth semipositive Weil-Petersson form on $B\backslash S'$. Its construction will be briefly recalled in section \[sec:CY degeneration 2\]. We also have the semi-Ricci flat metric ${\omega_{\textrm{SRF}}}$ on $X\backslash S$, such that for each $b \in B \backslash S'$ , ${\omega_{\textrm{SRF}}}|_{X_b}$ is the unique Ricci-flat metric on $X_b$ in the Kähler class $[{\omega}_X|_{X_b}]$. Define the reference metrics ${\tilde{{\omega}}}_t$ on the regular part $X\backslash S$ by $$\begin{aligned} {\tilde{{\omega}}}_t={\omega}_B+e^{-t}{\omega_{\textrm{SRF}}}.\end{aligned}$$ Let ${\omega}(t)$ be the unique Ricci-flat metric in $[{\tilde{{\omega}}}_t]= [\chi] + e^{-t} [{\omega}_X]$. Then Tosatti-Weinkove-Yang proved in [@TWY] that $\|{\omega}(t)-{\tilde{{\omega}}}_t\|_{C^0(K,{\tilde{{\omega}}}_t)} \rightarrow 0, \quad \textrm{as } t\rightarrow \infty$ on any compact set $K\subset X\setminus S$. Denote the (1,3)-curvature tensor of a Kähler metric ${\omega}$ (associated to the Riemannian metric $g$) by $R^\sharp({\omega})$, i.e. $$\label{Equ: definition of R-sharp tensor} {R^\sharp({\omega})_{i\bar{j}k}}^l=g^{l\bar{q}}R({\omega})_{i\bar{j}k\bar{q}}.$$ If all the regular fibers are bi-holomorphic to each other, then ${\omega}_{\textrm{SRF,b}}$ is independent of the base regular point $b$ (see Equation \[Equ: identity of reference metric on Y\]), and at this time we define the Kähler metric ${\omega}_Y$ on $Y$ by $${\omega}_Y={\omega}_{\textrm{SRF,b}},$$ for any $b\in B \backslash S'$. Based on the higher estimates of Hein-Tosatti [@HT2], we have: \[Thm: convergence of higher order estimate for CY\] Assume all the regular fibers are bi-holomorphic to a fixed Calabi-Yau manifold $Y$. Let $U\subset B\setminus S'$ be an open set such that the fibration is holomorphically trivial over $U$. Identify $f^{-1}(U)$ with $U\times Y$. Define another reference metrics ${\tilde{{\omega}}}(t)$ on $U\times Y$ by $$\begin{aligned} {\tilde{{\omega}}}(t)={\omega}_B+e^{-t}{\omega}_Y.\end{aligned}$$ Then for each compact set $K \subset U$, for any $k \in {\mathbb{N}}$, we have $$\label{Equ: convergence of higher order estimate for CY} \|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^k(U\times Y,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ and $$\label{Equ: convergence of Rm in CY case} \|R^\sharp({\omega}(t))-R^\sharp({\tilde{{\omega}}}(t))\|_{C^k(K\times Y,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ where $h_{K,k}(t)$ are positive functions which tends to zero as $t\to \infty$, depenging only on $k$ and the domain $K$. Note that since we do not have curvature bounds, we can not derive the decay estimate of $\|Rm({\omega}(t))-Rm({\tilde{{\omega}}}(t))\|_{C^k}$ from (\[Equ: convergence of Rm in CY case\]).\ Another special case is when the smooth fibers $X_b$ are all complex torus by a holomorphic free action of a finite group, but we allow the complex structure to change. By [@GTZ; @HT; @TZ], we have $C^\infty$ estimates as well as local curvature bounds on $X\backslash S$. We can apply Lemma \[Thm: BIC principle\] to obtain: \[Thm: convergence Rm of CCY with torus fiber\] Assume that for some $b\in B\backslash S'$ the fiber $X_b=f^{-1}(b)$ is bi-holomorphic to a finite quotient of a torus. Let $K\subset X\backslash S$ be any compact subset. Then we have $$\label{Equ: convergence of CCY with torus fiber} \|{\omega}(t)-{\tilde{{\omega}}}_t\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq h_{K,k}(t).$$ and $$\label{Equ: convergence Rm of CCY with torus fiber} \|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}_t)\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq h_{K,k+2}(t).$$ where $h_{K,k}(t)$ are positive functions which tends to zero as $t\to \infty$, depending only on $k$ and the domain $K$. In particular, when $S=\emptyset$, the estimates are globally true and each $h_k(t)$ is of exponential fast decay. The second application is on the normalized Kähler-Ricci flow on torus fibered minimal models. Let $(X,{\omega}_0)$ be a compact Kähler $(n+m)$-manifold with semiample canonical bundle and Kodaira dimension $m$. Here we assume $m>0, n>0$. The sections of $K_X^\ell,$ for $\ell$ large, give rise to a fiber space $f:X\to B$ called the [*Iitaka fibration*]{} of $X$, with $B$ a normal projective variety of dimension $m$ and the smooth fibers $X_b=f^{-1}(b), b\in B\backslash S'$ are all Calabi-Yau $n$-manifolds, diffeomorphic to each other. Let $\chi$ be the restriction of $\frac{1}{\ell}\omega_{FS}$ to $B$, as well as its pullback to $X$. This time we consider the solution ${\omega}={\omega}(t)$ of the [*normalized*]{} Kähler-Ricci flow $$\label{Equ: normalized KRF} \frac{{\partial}}{{\partial}t}\omega=-{\mathrm{Ric}}(\omega)-\omega,\quad\omega(0)=\omega_0.$$ which exists for all $t\geq 0$. Thanks to [@ST1; @ST2; @ST3; @FZ; @TWY; @TZ; @Gi; @GTZ] we have that the evolving metrics have uniformly bounded scalar curvature globally and collapse locally uniformly on $X\backslash S$ to a canonical Kähler metric on $B\backslash S'$, and moreover the rescaled metrics along the fibers $e^t\omega|_{X_b}$ converge in $C^\infty$ to a Ricci-flat metric on $X_b$. This is the collapsing phenomenon in the Kähler-Ricci flow case. Now, assume the smooth fibers are the quotient of a complex torus by a holomorphic free action of a finite group, then we have smooth collapsing to the generalized Kähler-Einstein metrics defined by Song-Tian [@ST2] on the regular part with respect to a fixed metric. An immediate corollary of Lemma \[Thm: BIC principle\] is the smooth convergence of the solution and its curvatures. \[Thm: convergence of KRF with torus fiber\] Let $(X^{n+m},\omega_0)$ with $n>0$ be a compact Kähler manifold with $K_X$ semiample and $\kappa(X)=m>0$, and let $f:X\to B$ be the fibration as described above. Assume that for some $y\in B\backslash S'$ the fiber $X_b=f^{-1}(b)$ is bi-holomorphic to a finite quotient of a torus. Let ${\omega}(t),t\in[0,\infty)$ be the solution of the Kähler-Ricci flow starting at ${\omega}_0$. Let $K\subset X\backslash S$ be any compact subset. Then we have $$\label{Equ: convergence of KRF with torus fiber 1} \|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ where ${\tilde{{\omega}}}(t)=e^{-t}{\omega}_{{\mathrm{SRF}}}+(1-e^{-t}){\omega}_B$ with ${\omega}_B$ the Song-Tian’s generalized Kähler-Einstein metric current on $B$. (See section \[sec:flow\] for its definition) Moreover, we have the smooth convergence of the curvature tensors $$\label{Equ: convergence Rm curvature of KRF with torus fiber} \|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ $$\label{Equ: convergence Ric curvature of KRF with torus fiber} \|{\mathrm{Ric}}({\omega}(t))-{\mathrm{Ric}}({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ $$\label{Equ: convergence scalar curvature of KRF with torus fiber} \|R({\omega}(t))-R({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t),$$ where $h_{K,k}(t)$ are positive functions which tends to zero as $t\to \infty$, depending only on $k$ and the domain $K$. In particular, when $S=\emptyset$, the estimates are globally true and each $h_k(t)$ is of exponential fast decay. Our next result is to exhibit the relation between the Ricci curvature and scalar curvature of the solution ${\omega}(t)$ and the generalized Kähler-Einstein metric ${\omega}_B$. We have \[Thm: convergence Ric curvature to GKE of KRF with torus fiber\] Assume the same set-up as in Theorem \[Thm: convergence of KRF with torus fiber\]. Let $K\subset X\backslash S$ be any compact subset. Then we have $$\label{Equ: convergence Ric curvature to GKE of KRF with torus fiber} \|{\mathrm{Ric}}({\omega}(t))+{\omega}_B\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k+2}(t),$$ and the convergence of scalar curvature $$\label{Equ: convergence scalar curvature to GKE of KRF with torus fiber} \|R({\omega}(t))+m\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k+2}(t),$$ where $h_{K,k}(t)$ are positive functions which tends to zero as $t\to \infty$, depending only on $k$ and the domain $K$. In particular, when $S=\emptyset$, the estimate is globally true and each $h_k(t)$ is of exponential fast decay. In [@J], the first author showed that the scalar curvature converges to $-m$ in the $C^0_{loc}$ topology in the general fibration case. Here we improved the topology to $C^{\infty}_{loc}$ in the special case when the fibers are flat. We do expect that Theorem \[Thm: convergence Ric curvature to GKE of KRF with torus fiber\] holds for the general case.\ This paper is arranged as follows: In section \[sec: pf of BIC\], we prove Lemma \[Thm: BIC principle\] by maximum principle. Then we apply it to Calabi-Yau degenerations in section \[sec:CY degeneration\], where Theorem 1.3, 1.4 are proved. Finally in section \[sec:flow\], we prove Theorem 1.5 and 1.6 for normalized Kähler-Ricci flow on minimal models whose regular fibers are all finite quotients of complex tori.\ In this paper, we use the following notations and conventions. We always denote by $h(t)$ a positive function which tends to zero as $t\to \infty$, and by $h_{K,k}(t)$ we mean that this function also depends on the domain $K$ and the order $k$. We allow these functions change from line to line. When we compute on a product manifold $X=B\times Y$, we always use a product coordinate system and, we call $B$ the base space and the corresponding indices the base directions, and we call $Y$ the fiber space and the corresponding indices the fiber directions. We will denote any complex $(1,0)$ “base” $\mathbb{C}^m$ direction by a subscript $\mathbf{b}$ and any complex $(1,0)$ “fiber” $Y$ direction by a subscript $\mathbf{f}$. By $\nabla^{k,g}$ we means all the possible covariant derivatives with respect to the metric $g$, including holomorphic and anti-holomorphic covariant derivatives when $g$ is a Kähler. [**Acknowledgements.**]{} The first author would like to thank his Ph.D advisor Gang Tian for introducing him to geometric analysis, and for his innumerable encouragements and supports. The first author also would like to thank Yashan Zhang and Dongyi Wei for many helpful conversations. The second authors thanks Gang Tian for his constant supports. Both authors would like to thank Zhenlei Zhang and Jian Song for many helpful discussions. Proof of the “Boundedness Implies Convergence" principle {#sec: pf of BIC} ======================================================== We use the maximum principle to prove Lemma \[Thm: BIC principle\]. Let $K$ be any compact subset of $U$, and choose compact subsets $\hat{K}\subset\subset U$ and smooth cut-off function $\rho$ as in the Condition $(C)$ of Lemma \[Thm: BIC principle\]. Let $C_k$ denote constants which depend on $A_1, \dots, A_{k+2}$ which may change from line to line. The proof is by induction on the order $k$. The case $k=0$ for Equation is just the Condition $(A)$. Suppose we have established for $0, \dots, k-1$ for $k\geq 1$. Now we prove the estimate for $k$. First, using induction hypothesis, we can find some positive function $h(t)$ converging to 0 such that $$\label{Equ: induction hypothesis for k-1} \left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\leq h(t)$$ holds on $\hat{K}$. Next, using Condition $(B)$, we can compute for every $k\geq 0$ on $U$ $$\label{Equ: evolution inequality of eta for general k 1} \begin{split} &\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right)\\ =& -{\tilde{g}}(t)^{a\bar{b}}\cdot\nabla^{{\tilde{g}}(t)}_a\nabla^{{\tilde{g}}(t)}_{\bar{b}}\left(\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right)\\ =& -2\left|\nabla^{k+1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+\nabla^{k+2,{\tilde{g}}(t)}{\eta}*\nabla^{k,{\tilde{g}}(t)}{\eta}\\ \leq& -2\left|\nabla^{k+1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+C\cdot\left|\nabla^{k+2,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}\cdot\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}\\ \leq& -2\left|\nabla^{k+1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+C\cdot\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)} \end{split}$$ where $C=C_k$ and $*$ denotes the tensor contraction by the metric ${\tilde{g}}(t)$. Note that in the above inequalities we do not use Bochner-type formulas since we do not have curvature bounds. We use instead the assumption of higher order estimates. Applying for $k-1\geq 0$, we have on $U$ $$\label{Equ: evolution inequality of eta for general k-1} \left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right)\leq -2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+C\cdot\left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}.$$ Also by and Condition $(B)$ , we have on $U$ $$\label{Equ: evolution inequality of eta for general k 2} \begin{split} &\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\right)\\ =& 2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\cdot\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right)-2\left|\nabla\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right|_{{\tilde{g}}(t)}^2\\ \leq& 2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\cdot\left(-2\left|\nabla^{k+1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+C\cdot\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}\right)-2\left|\nabla\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right|_{{\tilde{g}}(t)}^2\\ \leq& 2C\cdot\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^3-2\left|\nabla\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right|_{{\tilde{g}}(t)}^2 \end{split}$$ where $C=C_k$. Now, take the cut-off function $\rho$ into consideration. By and Condition $(B)$ we have $$\label{Equ: evolution inequality of eta for general k 3} \begin{split} &\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\rho^2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\right)\\ =& \rho^2\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\right)+\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\rho^2\right)-2\left<\nabla\rho^2, \nabla \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\right>_{{\tilde{g}}(t)}\\ \leq& \rho^2\left(C\cdot\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2-2\left|\nabla\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right|_{{\tilde{g}}(t)}^2\right)+C\cdot \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\\ &+C\cdot\rho|\nabla \rho|_{{\tilde{g}}(t)}\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\left|\nabla \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\right|_{{\tilde{g}}(t)}\\ \leq& C\cdot \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2+C\cdot \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4\\ \leq& C\cdot \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2,\\ \end{split}$$ where $C=C_k$. Now, put , and together, we conclude that we can find some $h(t)$ such that on $\hat{K}$ $$\label{Equ: inequality for higher order} \left\{ \begin{aligned} &\left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1}\leq 1,\\ &\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1}\right)\leq-2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1}+1,\\ &\left(-\Delta_{{\tilde{g}}(t)}\right)\left(\rho^2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4h(t)^{-1}\right)\leq C\cdot \left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1},\\ \end{aligned} \right.$$ where $C=C_k$. Define an auxiliary function $$Q:=\rho^2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^4h(t)^{-1}+C\cdot \left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1}.$$ Using , on $\hat{K}$ we have $$\left(-\Delta_{{\tilde{g}}(t)}\right)\left(Q\right)\leq -\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2h(t)^{-1}+C.$$ Now, at a given time $t\geq 0$, assume $Q$ achieves it’s maximum in $U$ at a point $x_0\in\bar U$. If $x_0\notin \mathrm{Int}\hat K$, then $\rho\equiv 0$, , implies that $Q$ has a uniform upper bound $C$, and we are done. Otherwise $x_0\in \mathrm{Int}\hat{K}$, we have $$0\leq \left(-\Delta_{{\tilde{g}}(t)}\right)\left(Q\right)(x_0)\leq -\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2(x_0)h(t)^{-1}+C.$$ which gives $$\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2(x_0)h(t)^{-1}\leq C.$$ Then by Condition $(B)$ and we have on $\hat{K}$ $$Q\leq Q(x_0)\leq A_{k}^2\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2(x_0)h(t)^{-1}+C\cdot \left|\nabla^{k-1,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2(x_0)h(t)^{-1}\leq C.$$ Since $\rho\equiv 1$ on $K$, we obtain $$\left|\nabla^{k,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{g}}(t)}^2\leq Ch(t)^{\frac{1}{2}}$$ on $K$, where $h(t)$ depends on the constants $A_1, \dots, A_{k+2}, B_K$ and the function $h_0(t)$. This establishes for $k$ and hence completes the proof. In applications, it is crucial to have Condition (C). If there is a fibration structure, then we can find such cut-off functions by pulling back a cut-off function from the base manifold, as shown by the following Lemma: \[Lem:Condition C\] If $f:X^{m+n}\to B$ is a proper holomorphic submersion onto a ball in ${\mathbb{C}}^m$, and if ${\tilde{{\omega}}}(t)$ is of the form ${\omega}_B+e^{-t}{\omega}_0$, where ${\omega}_B$ is a Kähler form on $B$ and ${\omega}_0$ is real closed $(1,1)$ form on $X$ whose restriction to each fiber is positive, then for any compact subset $K\subset X$, we can find cut-off function $rho$ satisfying Condition (C) of Lemma \[Thm: BIC principle\]. Since $K$ is compact, so is $f(K)\subset B$. Then we can find a ball $B_1\subset\subset B$ such that $f(K)\subset B_1$. Choose a cut-off function $\rho_0\in C_0^\infty(B)$ such that $\textrm{supp} \rho_0\subset B_1$ and $\rho_0|_{f(K)}\equiv 1$. Then we can take $\rho:=f^*\rho_0$ and $\hat K:=f^{-1}(\overline{B_1})$. Since $$\sqrt{-1}\partial\rho_0\wedge{\overline{\partial}}\rho_0\leq C{\omega}_B,\qquad -C{\omega}_B\leq{\sqrt{-1} \partial \overline{\partial}}\rho_0\leq C{\omega}_B,$$ for some $C>0$ on $B$, we have on $\hat{K}$ $$\begin{split} &|\nabla\rho|_{{\tilde{{\omega}}}(t)}^2=\sum_{i,j=1}^{m}{\tilde{g}}(t)^{i\bar{j}}\partial_i\rho\partial_{\bar{j}}\rho\leq C,\\ &|\Delta_{{\tilde{{\omega}}}(t)}\rho|=\left|{\mathrm{tr}_{{\tilde{{\omega}}}(t)}{{\sqrt{-1} \partial \overline{\partial}}\rho}}\right|\leq C{\mathrm{tr}_{{\tilde{{\omega}}}(t)}{{\omega}_B}}\leq C\\ \end{split}$$ with some constant $C$ depending on the domain $\hat K$. This verifies Condition $(C)$. Applications to collapsing Calabi-Yau metrics {#sec:CY degeneration} ============================================= Metric and curvature convergence on locally trivial Calabi-Yau fibrations ------------------------------------------------------------------------- We recall basic definitions in the general fibration case. Let $X$ be a compact Kähler $(n+m)$-manifold with $c_1(X)=0$ and let ${\omega}_X$ be a Ricci-flat Kähler metric on $X$. Let $f:X\to Z$ be the fibration map with $B=f(X)$. Write $\chi=f^*\omega_Z$, where ${\omega}_Z$ is a smooth Kähler form on $Z$. Then $\chi$ is a smooth nonnegative $(1,1)$ form on $X$, and we will also write $\chi$ for the restriction of $\omega_Z$ to $B$. Note that $\int_{B\backslash S'}\chi^m$ is finite. We define a semi Ricci-flat form ${\omega_{\textrm{SRF}}}$ on $X\backslash S$ in the usual way. Namely, for each $b \in B \backslash S'$ there is a smooth function $\rho_b$ on $X_b$ so that ${\omega}_X|_{X_b} + {\sqrt{-1} \partial \overline{\partial}}\rho_b = {\omega}_{\textrm{SRF,b}}$ is Ricci-flat, normalized by $\int_{X_b} \rho_b ({\omega}_X|_{X_b})^n=0$. As $b$ varies, this defines a smooth function $\rho$ on $X\backslash S$ and we define ${\omega_{\textrm{SRF}}}= {\omega}_X + {\sqrt{-1} \partial \overline{\partial}}\rho.$ Let $F$ be the function on $X\backslash S$ given by $$F=\frac{{\omega}_X^{n+m}}{\binom{n+m}{n}{\omega_{\textrm{SRF}}}^n\wedge\chi^m}.$$ It is easy to see that $F$ is constant along the fibers $X_b$, $b\in B\backslash S'$, so it descends to a smooth function, also denoted by $F$, on $B\backslash S'$. We see that $F$ satisfies $\int_{B\backslash S'}F\chi^m=\int_X{\omega}_X^{n+m}/\binom{n+m}{n}\int_{X_y}{\omega}_X^n$ (see [@ST2 Section 3] and [@To Section 4]). Here note that $\int_{X_b}{\omega}_X^n$ is independent of $b\in B\backslash S'$. Then [@ST2 Section 3] shows that the Monge-Ampère equation $$\label{Equ: MA of GKE on the base} (\chi+{\sqrt{-1} \partial \overline{\partial}}v)^m=\frac{\binom{n+m}{n}\int_{X} {\omega}_X^n\wedge\chi^m}{\int_{X}{\omega}_X^{n+m}}F\chi^m,$$ has a unique solution $v$ which is a bounded $\chi$-plurisubharmonic function on $B$, smooth on $B\backslash S'$, with $\int_{X}v\omega_X^{n+m}=0$, where here and henceforth we write $v$ for $\pi^*v$. Define $${\omega}_B=\chi+{\sqrt{-1} \partial \overline{\partial}}v,$$ for $v$ solving (\[Equ: MA of GKE on the base\]). Note that we have $$\label{Equ: ct3} {\omega_{\textrm{SRF}}}^n\wedge{\omega}_B^m=\frac{\binom{n+m}{n}\int_{X} {\omega}_X^n\wedge\chi^m}{\int_{X}{\omega}_X^{n+m}} F{\omega_{\textrm{SRF}}}^n\wedge\chi^m=\frac{\int_{X} {\omega}_X^n\wedge\chi^m}{\int_{X}{\omega}_X^{n+m}}{\omega}_X^{n+m}.$$ Moreover, ${\omega}_B$ is a smooth Kähler metric on $B\backslash S'$, and satisfies $${\mathrm{Ric}}({\omega}_B)={\omega}_{{\rm WP}},$$ where ${\omega}_{{\rm WP}}$ is the semipositive Weil-Petersson form on $B\backslash S'$, characterizing the change of complex structures of the fibers. If on a domain $U\subset B\backslash S'$ the bundle is holomorphically trivial, then we have ${\omega}_{{\rm WP}}\equiv 0$ on $U$. For $t\geq 0$, let ${\omega}_t$ be the global reference metrics defined by $${\omega}_t=\chi+e^{-t}\omega_X \in \alpha_t = [\chi] + e^{-t} [{\omega}_X],$$ and let ${\omega}(t)={\omega}_t+{\sqrt{-1} \partial \overline{\partial}}{\varphi}(t)$ be the unique Ricci-flat Kähler metric on $X$ cohomologous to ${\omega}_t$, with the normalization $\int_X{\varphi}\omega_X^{n+m}=0$. Then ${\omega}(t)$ solves the Calabi-Yau equation $$\label{Equ: MA of collapsing CY} ({\omega}(t))^{n+m}=c_t e^{-nt}\omega_X^{n+m},$$ where $c_t$ is the constant given by $$\label{Equ: c_t} c_t=\frac{\int_X e^{nt}{\omega}_t^{n+m}}{\int_X{\omega}_X^{n+m}}=\frac{1}{\int_X{\omega}_X^{n+m}}\sum_{k=0}^m\binom{n+m}{k}e^{-(m-k)t}\int_X {\omega}_X^{n+m-k}\wedge\chi^k=\binom{n+m}{n} \frac{\int_X{\omega}_X^{n}\wedge\chi^m }{\int_X{\omega}_X^{n+m}}+O(e^{-t}),$$ which has a positive limit as $t \to \infty$. Now, define the reference metrics ${\tilde{{\omega}}}_t$ on the regular part $X\backslash S$ by $$\begin{aligned} {\tilde{{\omega}}}_t={\omega}_B+e^{-t}{\omega_{\textrm{SRF}}}.\end{aligned}$$ In [@TWY], Tosatti-Weinkove-Yang proved the following $C^0$ convergence theorem: \[Thm: convergence of TWY\] Let ${\omega}={\omega}(t) \in \alpha_t$ be Ricci-flat Kähler metrics on $X$ as described above. Then the following holds: For each compact set $K \subset X \backslash S$, $$\label{Equ: 0-th convergence of TWY} \|{\omega}(t)-{\tilde{{\omega}}}_t\|_{C^0(K,{\tilde{{\omega}}}_t)} \rightarrow 0, \quad \textrm{as } t\rightarrow \infty.$$ In particulr, if S=$\emptyset$, then the convergence is global and exponentially fast. In [@HT2], Hein-Tosatti obtained higher-order estimate when the smooth fibers are pairwise bi-holomorphic: \[Thm: bounded estimate for CY of HT\] Assume that all the fibers $X_b$ $(b\in U\subset B\setminus f(S))$ are bi-holomorphic to the same Calabi-Yau manifold $Y$. Over any small coordinate ball $U$ compactly contained in $B\setminus f(S)$, use [@FG] to trivialize $f$ holomorphically to a product $U\times Y\to U$. Define another Ricci-flat reference Kähler forms on $U\times Y$ by $\hat{\omega}(t)={\omega}_{\mathbb{C}^m}+e^{-t}{\omega}_Y$. Then for any $k \in {\mathbb{N}}$, there exists a constant $C_{U,k}$ such that $$\label{Equ: bounded estimate for CY of HT} \|{\omega}(t)\|_{C^k(U\times Y,\hat{\omega}(t))}\leq C_{U,k}.$$ holds uniformly for all $t \in [0,\infty)$. Here ${\omega}_Y$ is defined as follows. Let $f:X\to B$ be as in Theorem \[Thm: bounded estimate for CY of HT\]. By [@FG], $f$ is a holomorphic fiber bundle over $U$. Fix any small coordinate ball in $U$ over which this holomorphic fiber bundle is trivial. We may assume that $U$ is a ball in $\mathbb{C}^m$ and $f:U\times Y\to U$ is the projection, with $Y=X_b$ a compact Calabi-Yau manifold. Then we need to apply a gauge transformation. By [@He2 Prop 3.1] (cf. [@GTZ Prop 3.1], [@GW Lemma 4.1], [@He Claim 1, p.382], [@TZ p.2936–2937]), there is a unique Ricci-flat Kähler metric ${\omega}_Y$ on $Y$ such that, we can find a bi-holomorphism $T$ of $U\times Y$ (over $U$) such that $T^*{\omega}_X=S^*{\omega}_{\mathbb{C}^m}+{\omega}_Y$ for some $S\in \mathrm{GL}(m,\mathbb{C})$. Note that [@He2 Prop 3.1] is stated with $U=\mathbb{C}^m$, but the proof applies also if $U$ is a ball in $\mathbb{C}^m$. Let us also note that $T$ takes the form $T(z,y) = (z, y + \sigma(z))$, where $\sigma$ is a holomorphic function from $U$ to the space of $g_Y$-parallel $(1,0)$-vector fields on $Y$, and where the addition $y+\sigma(z)$ has the same meaning as in [@He2 (1.1)]. We should note that for each $b\in U$ the metrics ${\omega}_{\textrm{SRF,b}}$ and ${\omega}_Y$ are in the same Kähler class and both are Ricci-flat, so they are equal by the uniqueness part of Calabi-Yau theorem, i.e., we have $$\label{Equ: identity of reference metric on Y} {\omega}_{\textrm{SRF,b}}={\omega}_Y,$$ on $Y=X_b$ for all $b\in U$. Before we prove Theorem \[Thm: convergence of higher order estimate for CY\], we need some useful lemmas. \[Lem: bounded estimate between reference metric\] Let $(Y,{\omega}_Y)$ be a Kähler manifold and $B$ the unit ball in $\mathbb{C}^m$ ($m\geq 1$). Let ${\omega}_1, {\omega}_2$ be any two Kähler metrics on $B$, and define two families of product metrics $\hat{\omega}(t)$ and ${\tilde{{\omega}}}(t)$ for $t\in [0,\infty)$ on $X=B\times Y$ as $$\label{} \left\{ \begin{aligned} &\hat{\omega}(t)={\omega}_1+e^{-t}{\omega}_Y,\\ &{\tilde{{\omega}}}(t)={\omega}_2+e^{-t}{\omega}_Y,\\ \end{aligned} \right.$$ Then $$\label{Equ: bounded estimate between reference metric} \|\hat{\omega}(t)\|_{C^k(X,{\tilde{{\omega}}}(t))}\leq C_k.$$ for all $k\geq 0$, where the constant $C_k$ depends only on ${\omega}_1$ and ${\omega}_2$, but independent of $t$. Denote by $f$ the projection $B\times Y\to B$, and compute under product coordinates. We prove by induction that: $$\label{Equ: gradient of product} \nabla^{k,{\tilde{g}}(t)}{\hat{g}(t)}=f^*(\alpha_k),$$ where $\alpha_k=\nabla^{k,g_1}{g_2}$ is a well-defined covariant tensor on the base space $B$. Indeed, since $\hat{\omega}(t)$ is a product metric, we have $$\label{Equ: Christoffel of product metric hat} \Gamma(\hat{g}(t))^{k}_{ip}= \left\{ \begin{aligned} &(g_1)^{k\bar{l}}\nabla^{E}_i(g_1)_{p\bar{l}}, ~~ i,k,p\in \mathbf{b},\\ &(g_Y)^{k\bar{l}}\nabla^{E}_i(g_Y)_{p\bar{l}}, ~~ i,k,p\in \mathbf{f},\\ &0, ~~otherwise.\\ \end{aligned} \right.$$ and similarly $$\label{Equ: Christoffel of product metric tilde} \Gamma({\tilde{g}}(t))^{k}_{ip}= \left\{ \begin{aligned} &(g_2)^{k\bar{l}}\nabla^{E}_i(g_2)_{p\bar{l}}, ~~ i,k,p\in \mathbf{b},\\ &(g_Y)^{k\bar{l}}\nabla^{E}_i(g_Y)_{p\bar{l}}, ~~ i,k,p\in \mathbf{f},\\ &0, ~~otherwise.\\ \end{aligned} \right.$$ So $\hat{g}(t)^{k\bar{l}}\nabla^{{\tilde{g}}(t)}_i\hat{g}(t)_{p\bar{l}}$ is nonzero only if $i,k,p\in \mathbf{b}$, and when $i,k,p\in \mathbf{b}$, we have $$\label{} \hat{g}(t)^{k\bar{l}}\nabla^{{\tilde{g}}(t)}_i\hat{g}(t)_{p\bar{l}}=(g_1)^{k\bar{l}}\nabla^{E}_i(g_1)_{p\bar{l}}-(g_2)^{k\bar{l}}\nabla^{E}_i(g_2)_{p\bar{l}}=f^*\left((g_1)^{k\bar{l}}\nabla^{g_2}_i(g_1)_{p\bar{l}}\right),$$ which gives that $$\nabla^{{\tilde{g}}(t)}_i\hat{g}(t)_{k\bar{l}}=f^*\left(\nabla^{g_2}_i(g_1)_{k\bar{l}}\right).$$ This is also true for all directions, and hence establishes for $k=1$. Now assume that we have for $1, 2, \dots, k-1$ with $k\geq 2$. Then we have $$\nabla^{k-1,{\tilde{g}}(t)}\hat{g}(t)=f^*\left(\alpha_{k-1}\right),$$ with $\alpha_{k-1}=\nabla^{k-1,g_2}(g_1)$ which is a covariant tensor. Then $f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}$ is nonzero only if $i_2, \dots, i_{k+2}$ are all of $\mathbf{b}$ or $\overline{\mathbf{b}}$ drections. Now suppose $i_1$ is of the $\mathbf{b}$ or $\mathbf{f}$ directions, then we have $$\nabla^{{\tilde{g}}(t)}_{i_1}f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}=\nabla^{E}_{i_1}f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}-\sum_{2\leq j\leq k+2,i_j\in \mathbf{b}}\Gamma\left({\tilde{g}}(t)\right)^{p}_{i_1i_j}f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{j-1}, p, i_{j+1}, \dots,i_{k+2}}.$$ If $i_1$ is of the $\mathbf{f}$ directions, then since $f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}$ is a function only depends on the base invariant, and hence $\nabla^{E}_{i_1}f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}\equiv 0$, and using we have $\Gamma({\tilde{g}}(t))^{p}_{i_1i_j}\equiv 0$ since $i_j\in \mathbf{b}$. Hence this covariant derivative is nonzero only if $i_1$ is of the $\mathbf{b}$ directions, and the second terms is nonzero only if $p$ is of the $\mathbf{b}$ directions. Using again we obtain $$\nabla^{{\tilde{g}}(t)}_{i_1}f^*\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}=f^*\left(\nabla^{g_2}_{i_1}\left(\alpha_{k-1}\right)_{i_2, \dots, i_{k+2}}\right).$$ This is also true for all directions. The case for $i_1$ of the $\overline{\mathbf{b}}$ or $\overline{\mathbf{f}}$ directions is similar. Using induction, we obtain . Now, we have on $X$ the estimate $$\left|\nabla^{k,{\tilde{g}}(t)}{\hat{g}(t)}\right|_{{\tilde{{\omega}}}(t)}=\left|f^*(\alpha_k)\right|_{{\tilde{{\omega}}}(t)}=\left|\alpha_k\right|_{{\omega}_2}\leq C_k.$$ where the constant is independent the time $t$. This completes the proof of the Lemma. Next, we compare covariant derivatives of two Kähler metrics. \[Lem: relation between two metrics\] Let $X$ be a Kähler manifold. Let $\hat{\omega}, {\tilde{{\omega}}}$ be any two Kähler metrics on $X$ and $\alpha$ be any tensor on $X$. Then we have for any $k\geq 1$ $$\label{Equ: relation between three metrics} \nabla^{k,{\tilde{g}}}{\alpha}=\sum_{j\geq 1, i_1+\dots +i_j=k,i_1,\dots ,i_j\geq 0}\nabla^{i_1,\hat{g}}{\beta}*\dots*\nabla^{i_j,\hat{g}}{\beta}.$$ where $\beta$ means either the metric ${\tilde{g}}$ or the tensor $\alpha$, and $*$ denotes the tensor contraction by ${\tilde{g}}$. We prove by induction on $k$. We compute under any given coordinate system $\{z_i\}$ around a given point. First consider the $k=1$ case. For example, if $\alpha=\alpha_{k\bar{l}}$ is a two tensor, then we have $$\begin{split} \nabla^{{\tilde{g}}}_i\alpha_{k\bar{l}}-\nabla^{\hat{g}}_i\alpha_{k\bar{l}} &=\left(\nabla^{E}_i\alpha_{k\bar{l}}-\Gamma({\tilde{g}})^{p}_{ik}\cdot \alpha_{p\bar{l}}\right)-\left(\nabla^{E}_i\alpha_{k\bar{l}}-\Gamma(\hat{g})^{p}_{ik}\cdot \alpha_{p\bar{l}}\right)\\ &=-\left(\Gamma({\tilde{g}})^{p}_{ik}-\Gamma(\hat{g})^{p}_{ik}\right)\cdot \alpha_{p\bar{l}}\\ &=-{\tilde{g}}^{p\bar{q}}\cdot\nabla^{\hat{g}}_i{\tilde{g}}_{k\bar{q}}\cdot \alpha_{p\bar{l}}.\\ \end{split}$$ which gives $$\nabla^{{\tilde{g}}}\alpha=\nabla^{\hat{g}}\alpha+\nabla^{\hat{g}}{\tilde{g}}*\alpha.$$ where $*$ is tensor contraction by ${\tilde{g}}$. It’s easy to see that the same argument holds for any tensor field $\alpha$, this proves for $k=1$. Now assume that we have established for $1, 2, \dots, k-1$ with $k\geq 2$. Then we have $$\begin{split} \nabla^{k,{\tilde{g}}}\alpha &=\nabla^{{\tilde{g}}}\left(\nabla^{k-1,{\tilde{g}}}\alpha\right)\\ &=\nabla^{\hat{g}}\left(\nabla^{k-1,{\tilde{g}}}\alpha\right)+\nabla^{\hat{g}}{\tilde{g}}*\nabla^{k-1,{\tilde{g}}}\alpha\\ &=\left(\nabla^{\hat{g}}+\nabla^{\hat{g}}{\tilde{g}}*\right)\left(\sum_{j\geq 1, i_1+\dots +i_j=k-1,i_1,\dots ,i_j\geq 0}\nabla^{i_1,\hat{g}}{\beta}*\dots*\nabla^{i_j,\hat{g}}{\beta}\right)\\ &=\sum_{j\geq 1, i_1+\dots +i_j=k,i_1,\dots ,i_j\geq 0}\nabla^{i_1,\hat{g}}{\beta}*\dots*\nabla^{i_j,\hat{g}}{\beta},\\ \end{split}$$ where $\beta$ still denotes either the metric ${\tilde{g}}$ or the tensor $\alpha$, and $*$ still denotes the tensor contraction by ${\tilde{g}}$. This completes the inductive step and establish this lemma. As a corollary, we can change the reference metric in Hein-Tosatti’s estimate: \[Cor: bounded estimate on local base 2\] For all compact sets $K\subset B$ and all $k \in {\mathbb{N}}$, there exists a constant $C_{K,k}$ independent of $t$ such that for all $t \in [0,\infty)$ we have that $$\label{Equ: bounded estimate 2} \|{\omega}(t)\|_{C^k(K\times Y,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ Denote $\beta(t)$ the Kähler form ${\omega}(t)$ or ${\tilde{{\omega}}}(t)$, using Theorem \[Thm: bounded estimate for CY of HT\] and Lemma \[Lem: bounded estimate between reference metric\], we have for any compact subset $K\subset B$ $$\|\beta(t)\|_{C^k(K\times Y,\hat{{\omega}}(t))}\leq C_{K,k}.$$ Hence, using Lemma \[Lem: relation between two metrics\] with $\alpha={\omega}(t)$, $\hat{{\omega}}=\hat{{\omega}}(t)$ and ${\tilde{{\omega}}}={\tilde{{\omega}}}(t)$, we have the following estimate on $K\times Y$ for $k\geq 1$ $$\begin{split} &\left|\nabla^{k,{\tilde{g}}(t)}g(t)\right|_{\hat{{\omega}}(t)}\\= &\left|\sum_{j\geq 1, i_1+\dots +i_j=k,i_1,\dots ,i_j\geq 1}\nabla^{i_1,\hat{g}(t)}{\beta(t)}*\dots*\nabla^{i_j,\hat{g}(t)}{\beta(t)}\right|_{\hat{{\omega}}(t)}\\\leq &C\cdot\sum_{j\geq 1, i_1+\dots +i_j=k,i_1,\dots ,i_j\geq 1}\left|\nabla^{i_1,\hat{g}(t)}{\beta(t)}\right|_{\hat{{\omega}}(t)}\cdot \dots \cdot\left|\nabla^{i_j,\hat{g}(t)}{\beta(t)}\right|_{\hat{{\omega}}(t)}\\\leq &C_{K,k}.\\ \end{split}$$ where $*$ denotes tensor contraction by ${\tilde{g}}(t)$. Here we have used the uniformly equivalent relations between ${\omega}(t)$, $\hat{{\omega}}(t)$ and ${\tilde{{\omega}}}(t)$, and hence completes the proof of Corollary \[Cor: bounded estimate on local base 2\]. Suppose we have two uniformly equivalent families of Kähler metrics ${\omega}(t)$ and ${\tilde{{\omega}}}(t)$, it doesn’t matter which metric we use to measure the norm. Also, assume we have any quantity of the form $$A_1*A_2*\dots *A_k,$$ where each $A_i$ is a tensor, and $*$ is the tensor contraction given by ${\omega}(t)$ or ${\tilde{{\omega}}}(t)$, then by the uniformly equivalent relations between ${\omega}(t)$ and ${\tilde{{\omega}}}(t)$, we have $$\left|A_1*A_2*\dots *A_k\right|\leq C\cdot\left|A_1\right|_{{\tilde{{\omega}}}(t)}\cdot \dots \cdot \left|A_k\right|_{{\tilde{{\omega}}}(t)}.$$ for some uniform constant $C$ independent of $t$, since here we have only finitely many contractions (depending only on k and the degrees of the $A_i$’s). The case for three or more uniformly equivalent metrics is similar. We will use such principle to take norms throughout this paper. The following lemma is a standard result in Kähler geometry, and follows easily by direct computations. So we omit the proof. \[Lem: relation of Rm between two metrics\] Given $X$ be a Kähler manifold, and ${\omega}$, ${\tilde{{\omega}}}$ be any two Kähler forms on $X$. Define the tensor $\Psi$ on $X$ by $$\Psi^k_{ip}:=\Gamma(g)^{k}_{ip}-\Gamma({\tilde{g}})^{k}_{ip}=g^{k\bar{l}}\nabla^{{\tilde{g}}}_ig_{p\bar{l}}.$$ Then we have $$\label{Equ: relation of Rm between two metrics 1} R({\omega})_{i\bar{j}k\bar{l}}={\tilde{g}}^{s\bar{v}}g_{k\bar{v}}R({\tilde{{\omega}}})_{i\bar{j}s\bar{l}}-\nabla^{{\tilde{g}}}_i\nabla^{{\tilde{g}}}_{\bar{j}}g_{k\bar{l}}+g_{u\bar{v}}\Psi^u_{ik}\overline{\Psi^v_{jl}}.$$ In particular, we have $$\label{Equ: relation of Rm between two metrics 2} {R^\sharp({\omega})_{i\bar{j}k}}^l={R^\sharp({\tilde{{\omega}}})_{i\bar{j}k}}^l-g^{l\bar{v}}\nabla^{{\tilde{g}}}_{\bar{j}}\nabla^{{\tilde{g}}}_ig_{k\bar{v}}+g^{l\bar{v}}g_{s\bar{t}}\Psi^s_{ik}\overline{\Psi^t_{jv}}.$$ This lemma shows that we can express the difference of the ${\mathrm{Rm}}$ curvature tensor of two Kähler metrics as the tensor contraction of the first covariant derivatives and second covariant derivatives. Now we can prove Theorem \[Thm: convergence of higher order estimate for CY\]: We still just need to verify the Conditions $(A)-(C)$ of Lemma \[Thm: BIC principle\] with $$\eta(t)={\omega}(t)-{\tilde{{\omega}}}(t).$$ We have already trivialize $f$ holomorphically to a product $U\times Y\to U$. Let $K\subset\subset U$ be any compact subset. Condition $(C)$ follows from Lemma \[Lem:Condition C\]. Condition $(B)$: Replacing $U$ by a slightly smaller subset. With Theorem \[Thm: bounded estimate for CY of HT\] at hand, Corollary \[Cor: bounded estimate on local base 2\] implies the estimate $$\|{\omega}(t)\|_{C^k(U\times Y,{\tilde{{\omega}}}(t))}\leq C_{U,k}$$ for any $k\geq 0$. Condition $(A)$: Replacing $U$ by a smaller subset, we only need to verify the condition $$\|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^0(U\times Y,{\tilde{{\omega}}}(t))}\leq h_0(t).$$ From the result of Tosatti-Weinkove-Yang, say Equation of Theorem \[Thm: convergence of TWY\], we have $$\|{\omega}(t)-{\tilde{{\omega}}}_t\|_{C^0(U\times Y,{\tilde{{\omega}}}_t)}^2\leq h(t).$$ where ${\tilde{{\omega}}}_t={\omega}_B+e^{-t}{\omega_{\textrm{SRF}}}$. Choosing product coordinates, say $\{z_{\alpha}, 1\leq\alpha\leq m; y_i, 1\leq i\leq n\}$ with $z_{\alpha}$ being base coordinates and $y_i$ being fiber coordinates. Then the above estimate implies on $U\times Y$ $$\label{Equ: local estimate of metric 1} \left\{ \begin{aligned} & \left|g(t)_{\alpha\bar{\beta}}-\left[(1-e^{-t})(g_B)_{\alpha\bar{\beta}}+e^{-t}(g_{\mathrm{SRF}})_{\alpha\bar{\beta}}\right]\right|^2\leq h(t),~~\alpha, \beta\in \mathbf{b},\\ & e^t\cdot\left|g(t)_{\alpha\bar{j}}-e^{-t}(g_{\mathrm{SRF}})_{\alpha\bar{j}}\right|^2\leq h(t),~~\alpha\in \mathbf{b},j\in \mathbf{f},\\ & e^{2t}\cdot\left|g(t)_{i\bar{j}}-(g_{\mathrm{SRF}})_{i\bar{j}}\right|^2\leq h(t),~~i, j\in \mathbf{f}.\\ \end{aligned} \right.$$ The first inequality of implies that $$\left|g(t)_{\alpha\bar{\beta}}-(g_B)_{\alpha\bar{\beta}}\right|^2\leq 2h(t)+2\left|e^{-t}(g_B)_{\alpha\bar{\beta}}-e^{-t}(g_{\mathrm{SRF}})_{\alpha\bar{\beta}}\right|^2\leq h(t),~~\alpha, \beta\in \mathbf{b},$$ and the second inequality of implies $$e^t\cdot\left|g(t)_{\alpha\bar{j}}\right|^2\leq 2h(t)+e^{-t}\cdot\left|(g_{\mathrm{SRF}})_{\alpha\bar{j}}\right|^2\leq h(t),~~\alpha\in \mathbf{b},j\in \mathbf{f}.$$ Since ${\omega}_{\textrm{SRF,b}}={\omega}_Y$ for any $b\in U$, we have $(g_{\mathrm{SRF}})_{i\bar{j}}=(g_Y)_{i\bar{j}}$ with $i,j\in \mathbf{f}$ under the product coordinates, hence the third inequality of implies $$e^{2t}\cdot\left|g(t)_{i\bar{j}}-(g_Y)_{i\bar{j}}\right|^2\leq h(t),~~i, j\in \mathbf{f}.$$ So we conclude that on $U\times Y$ under product coordinates $$\label{Equ: local estimate of metric 2} \left\{ \begin{aligned} & \left|g(t)_{\alpha\bar{\beta}}-(g_B)_{\alpha\bar{\beta}}\right|^2\leq h(t),~~\alpha, \beta\in \mathbf{b},\\ & e^t\cdot\left|g(t)_{\alpha\bar{j}}\right|^2\leq h(t),~~\alpha\in \mathbf{b},j\in \mathbf{f},\\ & e^{2t}\cdot\left|g(t)_{i\bar{j}}-(g_Y)_{i\bar{j}}\right|^2\leq h(t),~~i, j\in \mathbf{f}.\\ \end{aligned} \right.$$ This implies that $$\|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^0(U\times Y,{\tilde{{\omega}}}(t))}\leq h(t).$$ This verifies Condition $(A)$ since we already have local uniform equivalence between ${\omega}(t)$ and ${\tilde{{\omega}}}(t)$. Now applying Lemma \[Thm: BIC principle\], we get the desired estimate . It remains to prove curvature convergence estimates . Applying Lemma \[Lem: relation of Rm between two metrics\] with ${\omega}={\omega}(t)$ and ${\tilde{{\omega}}}={\tilde{{\omega}}}(t)$, and define the tensor $$\label{Equ: definition of first order derivative 1} \Psi(t)^k_{ip}:=\Gamma(g(t))^{k}_{ip}-\Gamma({\tilde{g}}(t))^{k}_{ip}=g(t)^{k\bar{l}}\nabla^{{\tilde{g}}(t)}_ig(t)_{p\bar{l}},$$ we get $$\begin{split} {R^\sharp({\omega}(t))_{i\bar{j}k}}^l-{R^\sharp({\tilde{{\omega}}}(t))_{i\bar{j}k}}^l &=g(t)^{l\bar{v}}\nabla^{{\tilde{g}}(t)}_{\bar{j}}\nabla^{{\tilde{g}}(t)}_ig(t)_{k\bar{v}}+g(t)^{l\bar{v}}g(t)_{s\bar{t}}\Psi(t)^s_{ik}\overline{\Psi(t)^t_{jv}}\\ &=g(t)^{l\bar{v}}\nabla^{{\tilde{g}}(t)}_{\bar{j}}\nabla^{{\tilde{g}}(t)}_ig(t)_{k\bar{v}}+g(t)^{l\bar{v}}g(t)^{p\bar{q}}\nabla^{{\tilde{g}}(t)}_ig(t)_{k\bar{q}}\nabla^{{\tilde{g}}(t)}_{\bar{j}}g(t)_{p\bar{v}}.\\ \end{split}$$ Then by induction we can show that for all $k\geq 0$ $$\label{Equ: higher-order derivatives of Q} \nabla^{k,{\tilde{g}}(t)}\left[R^\sharp({\omega}(t))-R^\sharp({\tilde{{\omega}}}(t))\right]=\sum_{j\geq 1,i_1+\dots +i_j=k+2,i_1,\dots ,i_j\geq 1}\nabla^{i_1,{\tilde{g}}(t)}g(t)*\dots *\nabla^{i_j,{\tilde{g}}(t)}g(t).$$ where $*$ is the tensor contraction given by $g(t)$. In fact, $k=0$ case already follows from . Suppose we have for $0, \dots, k-1$ for $k\geq 1$. Then for $k$ we have $$\begin{split} \nabla^{k,{\tilde{g}}(t)}\left[R^\sharp({\omega}(t))-R^\sharp({\tilde{{\omega}}}(t))\right] &=\nabla^{{\tilde{g}}(t)}\left\{\sum_{j\geq 1,i_1+\dots +i_j=k+1,i_1,\dots ,i_j\geq 1}\nabla^{i_1,{\tilde{g}}(t)}g(t)*\dots *\nabla^{i_j,{\tilde{g}}(t)}g(t)\right\}\\ &=\sum_{j\geq 1,i_1+\dots +i_j=k+2,i_1,\dots ,i_j\geq 1}\nabla^{i_1,{\tilde{g}}(t)}g(t)*\dots *\nabla^{i_j,{\tilde{g}}(t)}g(t).\\ \end{split}$$ This proves for $k$. Now, taking norms with respect to ${\tilde{g}}(t)$ and using the equivalence of ${\omega}(t)$ and ${\tilde{{\omega}}}(t)$ we have on $K\times Y$ $$\begin{split} \left|\nabla^{k,{\tilde{g}}(t)}\left[R^\sharp({\omega}(t))-R^\sharp({\tilde{{\omega}}}(t))\right]\right|_{{\tilde{{\omega}}}(t)} &\leq C_k\cdot\sum_{j\geq 1,i_1+\dots +i_j=k+2,i_1,\dots ,i_j\geq 1}\left|\nabla^{i_1,{\tilde{g}}(t)}g(t)\right|_{{\tilde{{\omega}}}(t)}\cdot\dots\cdot\left|\nabla^{i_j,{\tilde{g}}(t)}g(t)\right|_{{\tilde{{\omega}}}(t)}\\ &\leq h_{K,k}(t).\\ \end{split}$$ So we get , and this completes the proof. Metric and curvature convergence on torus-fibered Calabi-Yau manifolds {#sec:CY degeneration 2} ---------------------------------------------------------------------- Now we consider the case when the smooth fibers $X_b$ are finite quotients of complex tori, but we allow the complex structure to change. We denote the semi-Ricci flat form ${\omega}_{{\mathrm{SRF}}}$ by ${\omega}_{{\mathrm{SF}}}$ now, since in this case, its restriction to each smooth fiber is actually flat. We have the following higher-order estimates. \[Lemma: basic lemma 1\] For all compact sets $K\subset X\backslash S$ and all $k\geq 0$, there exists constants $C_{K,k}$ independent of $t$ such that for all $t\in [0,\infty)$ we have that $$\label{Equ: higher-order bounded 2} \|{\omega}(t)\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq C_{K,k}.$$ and the curvature bound $$\label{Equ: Rm curvature bounded 2} \|{\mathrm{Rm}}({\omega}(t))\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq C_{K,k}.$$ Now, we can prove Theorem \[Thm: convergence Rm of CCY with torus fiber\]. For all compact sets $K\subset X\backslash S$, we choose disks $B_1\subset\subset B_2\subset\subset B\backslash S'$ such that $K\subset f^{-1}(B_1)$. Set $U=f^{-1}(B_2)$. Then, as the proof of Theorem \[Thm: convergence of higher order estimate for CY\], with the help of Theorem \[Thm: convergence of TWY\] and Lemma \[Lemma: basic lemma 1\], we can similarly verify the Conditions $(A)-(C)$ of Lemma \[Thm: BIC principle\] with $$\eta(t)={\omega}(t)-{\tilde{{\omega}}}_t.$$ So we immediately get the higher-order convergence $$\|{\omega}(t)-{\tilde{{\omega}}}_t\|_{C^k(K,{\tilde{{\omega}}}_t)}\le h_{K,k}(t),$$ where $h_{K,k}(t)$ are positive functions which tends to zero as $t\to \infty$, depending only on $k$ and the domain $K$. It remains to show that $$\|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}_t)\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq h_{K,k}(t).$$ Applying Lemma \[Lem: relation of Rm between two metrics\] with ${\omega}={\omega}(t)$ and ${\tilde{{\omega}}}={\tilde{{\omega}}}_t$, we get $$\label{Equ: relation of the Rm tensor} R({\omega}(t))_{i\bar{j}k\bar{l}}=({\tilde{g}}_t)^{s\bar{v}}g(t)_{k\bar{v}}R({\tilde{{\omega}}}_t)_{i\bar{j}s\bar{l}}-\nabla^{{\tilde{g}}_t}_i\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{k\bar{l}}+g(t)_{u\bar{v}}\Psi(t)^u_{ik}\overline{\Psi(t)^v_{jl}}.$$ which gives $$\label{Equ: relation of the Rm tensor 2} R({\tilde{{\omega}}}_t)_{i\bar{j}k\bar{l}}=g(t)^{s\bar{v}}({\tilde{g}}_t)_{k\bar{v}}R({\omega}(t))_{i\bar{j}s\bar{l}}+g(t)^{s\bar{v}}({\tilde{g}}_t)_{k\bar{v}}\nabla^{{\tilde{g}}_t}_i\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{s\bar{l}}-g(t)^{s\bar{t}}({\tilde{g}}_t)_{k\bar{t}}g(t)_{u\bar{v}}\Psi(t)^u_{is}\overline{\Psi(t)^v_{jl}}.$$ This is equivalent to $$\label{Equ: relation of the Rm tensor 3} R({\tilde{{\omega}}}_t)_{i\bar{j}k\bar{l}}=g(t)^{s\bar{v}}({\tilde{g}}_t)_{k\bar{v}}R({\omega}(t))_{i\bar{j}s\bar{l}}+g(t)^{s\bar{v}}({\tilde{g}}_t)_{k\bar{v}}\nabla^{{\tilde{g}}_t}_i\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{s\bar{l}}-g(t)^{s\bar{t}}g(t)^{p\bar{q}}({\tilde{g}}_t)_{k\bar{t}}\nabla^{{\tilde{g}}_t}_ig(t)_{s\bar{q}}\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{p\bar{l}}.$$ Hence we have $$\label{Equ: relation of the Rm tensor 4} \begin{split} &R({\tilde{{\omega}}}_t)_{i\bar{j}k\bar{l}}-R({\omega}(t))_{i\bar{j}k\bar{l}}\\ =&g(t)^{s\bar{v}}[({\tilde{g}}_t)_{k\bar{v}}-g(t)_{k\bar{v}}]R({\omega}(t))_{i\bar{j}s\bar{l}}+g(t)^{s\bar{v}}({\tilde{g}}_t)_{k\bar{v}}\nabla^{{\tilde{g}}_t}_i\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{s\bar{l}}-g(t)^{s\bar{t}}g(t)^{p\bar{q}}({\tilde{g}}_t)_{k\bar{t}}\nabla^{{\tilde{g}}_t}_ig(t)_{s\bar{q}}\nabla^{{\tilde{g}}_t}_{\bar{j}}g(t)_{p\bar{l}}.\\ \end{split}$$ Taking norms with respect to ${\tilde{{\omega}}}_t$ and applying and Lemma \[Lemma: basic lemma 1\], we have that on $U$ $$\begin{split} &\left|{\mathrm{Rm}}({\tilde{{\omega}}}_t)-{\mathrm{Rm}}({\omega}(t))\right|_{{\tilde{{\omega}}}_t}\\ \leq& C\cdot \left|{\tilde{{\omega}}}_t-{\omega}(t)\right|_{{\tilde{{\omega}}}_t}\cdot \left|{\mathrm{Rm}}({\omega}(t)\right|_{{\tilde{{\omega}}}_t}+C\cdot \left|\nabla^{2,{\tilde{g}}_t}g(t)\right|_{{\tilde{{\omega}}}_t}+C\cdot \left|\nabla^{{\tilde{g}}_t}g(t)\right|^2_{{\tilde{{\omega}}}_t}\\ \leq& C\cdot h_K(t)\cdot C_K+C\cdot h_{K}(t)+C\cdot h_{K}(t)^2\\ \leq&h_K(t).\\ \end{split}$$ Also from , by induction we can easily obtain that for all $k\geq 0$ $$\label{Equ: higher-order derivatives of Rm} \begin{split} &\nabla^{k,{\tilde{g}}_t}{\mathrm{Rm}}({\tilde{{\omega}}}_t)\\ =&\sum_{j\geq 1,i_1+\dots +i_j=k,i_1,\dots ,i_j\geq 0}\nabla^{i_1,{\tilde{g}}_t}g(t)*\dots *\nabla^{i_{j-1},{\tilde{g}}_t}g(t)*\nabla^{i_j,{\tilde{g}}_t}{\mathrm{Rm}}({\omega}(t)),\\ \end{split}$$ where $*$ denotes tensor contraction by $g(t)$ or ${\tilde{g}}_t$. This implies that on $U$ $$\begin{split} &\left|\nabla^{k,{\tilde{g}}_t}{\mathrm{Rm}}({\tilde{{\omega}}}_t)\right|_{{\tilde{{\omega}}}_t}\\ \leq& C\cdot \sum_{j\geq 1,i_1+\dots +i_j=k,i_1,\dots ,i_{j}\geq 0}\left|\nabla^{i_1,{\tilde{g}}_t}g(t)\right|_{{\tilde{{\omega}}}_t}\cdot\dots\cdot\left|\nabla^{i_{j-1},{\tilde{g}}_t}g(t)\right|_{{\tilde{{\omega}}}(t)}\cdot\left|\nabla^{i_j,{\tilde{g}}_t}{\mathrm{Rm}}({\omega}(t))\right|_{{\tilde{{\omega}}}_t}\\ \leq&C_{K,k}.\\ \end{split}$$ Hence if we set $${\tilde{\eta}}(t)={\mathrm{Rm}}({\tilde{{\omega}}}_t)-{\mathrm{Rm}}({\omega}(t)),$$ then we have $$\label{Equ: conditions of Rm 1} \left\{ \begin{aligned} &\|{\tilde{\eta}}(t)\|_{C^0(U,{\tilde{{\omega}}}_t)}=\|{\mathrm{Rm}}({\tilde{{\omega}}}_t)-{\mathrm{Rm}}({\omega}(t))\|_{C^0(U,{\tilde{{\omega}}}_t)}\leq h_K(t),\\ &\|{\tilde{\eta}}(t)\|_{C^k(U,{\tilde{{\omega}}}_t)}=\|{\mathrm{Rm}}({\tilde{{\omega}}}_t)-{\mathrm{Rm}}({\omega}(t))\|_{C^k(U,{\tilde{{\omega}}}_t)}\leq C_{K,k},\\ \end{aligned} \right.$$ Hence Conditions $(A)$ and $(B)$ of Lemma \[Thm: BIC principle\] are satisfied for $\eta(t)={\tilde{\eta}}(t)$ and ${\tilde{g}}(t)={\tilde{g}}_t$. Hence from Lemma \[Thm: BIC principle\] we conclude the local convergence $$\|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}_t)\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq h_{K,k}(t).$$ This establish . Applications to normalized Kähler-Ricci flow on torus-fibered minimal models {#sec:flow} ============================================================================ Let $(X^{m+n},\omega_0)$ be a compact Kähler manifold with semi-ample canonical bundle $K_X$ and assume its Kodaira dimension to be $0<m:={\textrm{Kod}}(X)<m+n$. Then the pluricanonical system $|\ell K_X|$ for sufficiently large $\ell \in \mathbb{Z^+}$ induces the so called “Iitaka fibration" map $$\label{Equ: canonical map} f:X \to B \subset \mathbb{C}\mathbb{P}^N:=\mathbb{P}H^0(X,K_X^{\otimes \ell}),$$ where $B$ is the canonical model of $X$ with $dimB=m$. Let $S'$ be the singular set of $B$ together with the set of critical values of $f$, and we define $S=f^{-1}(S') \subset X$. From , we have $f^*\mathcal{O}(1)=K_X^{\otimes \ell}$, hence if we let $\chi=\frac{1}{\ell}\omega_{\mathrm{FS}}$ on $\mathbb{P}H^0(X,K_X^{\otimes \ell})$, we have that $f^*\chi$ (denoted by $\chi$ afterwards) is a smooth semi-positive representative of $-c_1(X)$. Here, $\omega_{\mathrm{FS}}$ denotes the Fubini-Study metric. Given the Kähler metric ${\omega}_0$ on $X$, since $X_b:=f^{-1}(b)$ is a Calabi-Yau manifold for each $b\in B\backslash S'$, there exists a unique smooth function $\rho_b$ on $X_b$ with $\int_{X_b}\rho_b\omega_0^{n}=0$, such that ${\omega}_0|_{X_b}+{\sqrt{-1} \partial \overline{\partial}}\rho_b=:{\omega}_b$ is the unique Ricci-flat Kähler metric on $X_b$ in the class $[{\omega}_0|_{X_b}]$. Moreover, $\rho_b$ depends smoothly on $b$, and so define a global smooth function on $X\backslash S$. We define $${\omega}_{{\mathrm{SRF}}}={\omega}_0+{\sqrt{-1} \partial \overline{\partial}}\rho,$$ which is a closed real $(1,1)$-form on $X\backslash S$, called the “semi-Ricci flat metric”. Let ${\Omega}$ be the smooth volume form on $X$ with $$\label{Equ: background volume form} {\sqrt{-1} \partial \overline{\partial}}\log{\Omega}=\chi,~\int_X{\Omega}=\binom{m+n}{m}\int_X{\omega}_0^{n}\wedge\chi^m.$$ Define a function $F$ on $X\backslash S$ by $$F:=\frac{{\Omega}}{\binom{m+n}{m}\chi^m\wedge{\omega}_{{\mathrm{SRF}}}^{n}},$$ then $F$ is constant along the fiber $X_b$, $b\in B\backslash S'$, so it descends to a smooth function on $B\backslash S^\prime$. Then [@ST2] showed that the Monge-Ampére equation $$\label{Equ: MAE on canonical model} (\chi +{\sqrt{-1} \partial \overline{\partial}}v)^m=Fe^v\chi^m,$$ has a unique solution $v\in {\mathrm{PSH}}(\chi)\cap C^0(B)\cap C^\infty(B\backslash S')$. Define $${\omega}_B=\chi+{\sqrt{-1} \partial \overline{\partial}}v ,$$ which is a smooth Kähler metric on $B\backslash S'$, satisfying the twisted Kähler-Einstein equation $${\mathrm{Ric}}({\omega}_B)=-{\omega}_B+{\omega}_{\mathrm{WP}},$$ where ${\omega}_{\mathrm{WP}}$ is the smooth Weil-Petersson form on $B\backslash S'$. Now let ${\omega}={\omega}(t)$ be the solution of the normalized Kähler-Ricci flow $$\label{Equ: normalized KRF 2} {\frac{\partial}{\partial t}}{\omega}=-{\mathrm{Ric}}({\omega})-{\omega},~{\omega}(0)={\omega}_0,$$ whose solution exists for all time. Define the global reference metrics $$\hat{{\omega}}(t)=e^{-t}{\omega}_0+(1-e^{-t})\chi,$$ then it is Kähler for all $t\ge 0$, and we can write ${\omega}(t)=\hat{{\omega}}(t)+{\sqrt{-1} \partial \overline{\partial}}{\varphi}(t)$. Then the Kähler-Ricci flow is equivalent to the parabolic Monge-Ampére equation $$\label{scalar MAE} {\frac{\partial}{\partial t}}{\varphi}=\mathrm{log}\frac{e^{(n-m)t}\left(\hat{{\omega}}(t)+{\sqrt{-1} \partial \overline{\partial}}{\varphi}(t)\right)^n}{{\Omega}}-{\varphi},~{\varphi}(0)=0.$$ We denote by $T_0={\mathrm{tr}_{{\omega}(t)}{{\omega}_B}}$ and $u=\dot{{\varphi}}+{\varphi}-v$ on $X\backslash S$. Define on $X\backslash S$ the reference metrics $${\tilde{{\omega}}}(t)=e^{-t}{\omega}_{{\mathrm{SRF}}}+(1-e^{-t}){\omega}_B.$$ We always set $K=f^{-1}(K')$， where $K'\subset B\backslash S'$ is a compact subset. Then we can choose some open subset $U'\subset\subset B\backslash S'$ such that $K'\subset U'$. Set $U=f^{-1}(U')$, then $K\subset \subset U\subset \subset X\backslash S$. First, we have the following lemma in the general fibration case. See [@ST2; @ST3; @FZ; @TWY; @To4; @J]. \[Lemma: basic lemma 2\] There exist some constant $C=C(K)>0$ and positive functions $h(t)$ which tends to zero as $t\to \infty$, depending on the domain $K$, such that 1. $C^{-1}\hat{{\omega}}(t)\le{\omega}(t)\le C\hat{{\omega}}(t)$, on $K\times[0,\infty).$ 2. $|{\varphi}-v|+|\dot{{\varphi}}+{\varphi}-v|\le h(t),$ on $K\times[0,\infty).$ 3. $\|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^0(K,{\omega}(t))}\le h(t).$ 4. $|T_0-m|+\left|\|{\omega}_B\|_{{\omega}(t)}^2-m\right|\le h(t),~on~K\times[0,\infty).$ 5. There exists a uniform constant $C_0>0$ such that $$|R|\le C_0,~on~X\times[0,\infty).$$ 6. Along the normalized Kähler-Ricci flow , we have on $X\backslash S$ $$\label{Equ: evolution equality of u} \left({\partial}_t-\Delta\right)u={\mathrm{tr}_{{\omega}(t)}{{\omega}_B}}-m.$$ 7. We have $$\label{first derivative convergence2} |\nabla u|^2\leq h(t),~on~K\times[0,\infty).$$ 8. $\left|R(t)+m\right|\le h(t),~on~K\times[0,\infty).$ Especially, if $S=\emptyset$, then all of the above estimates hold with $K$ replaced by $X$ and $h(t)$ replaced by $Ce^{-\eta t}$ for some constants $\eta,C>0$ depending on $(X,{\omega}_0)$. From now on, assume the smooth fibers are the quotients of complex tori by holomorphic free action of a finite group. In this case, the semi-Ricci flat metric ${\omega}_{\mathrm{SRF}}$ we constructed above is actually flat when restricted to any smooth fiber $X_b$, $b\in B\backslash S^\prime$, and we denote ${\omega}_{\mathrm{SRF}}$ by ${\omega}_{\mathrm{SF}}$ to indicate such semi-flat property. We have the following estimates. \[Lemma: basic lemma 3\] For all compact sets $K\subset X\backslash S$ and all $k\geq 0$, there exists constants $C_{K,k}$ independent of $t$ such that for all $t\in [0,\infty)$ we have the higher-order derivatives bound $$\label{Equ: higher-order bounded} \|{\omega}(t)\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ and the curvatures bound $$\label{Equ: Rm curvature bounded} \|{\mathrm{Rm}}({\omega}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ $$\label{Equ: Ric curvature bounded} \|{\mathrm{Ric}}({\omega}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ $$\label{Equ: R curvature bounded} \|R({\omega}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ Now we can prove Theorem \[Thm: convergence of KRF with torus fiber\]. For all compact sets $K\subset X\backslash S$, we choose disks $B_1\subset\subset B_2\subset\subset B\backslash S'$ such that $K\subset f^{-1}(B_1)$. Set $U=f^{-1}(B_2)$. Then, as the proof of Theorem \[Thm: convergence Rm of CCY with torus fiber\], with the help of Lemma \[Lemma: basic lemma 2\] and Lemma \[Lemma: basic lemma 3\], we can similarly verify the Conditions $(A)-(C)$ of Lemma \[Thm: BIC principle\] with $$\eta(t)={\omega}(t)-{\tilde{{\omega}}}(t).$$ Hence from Lemma \[Thm: BIC principle\] we have the convergence estimate $$\|{\omega}(t)-{\tilde{{\omega}}}(t)\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t).$$ Also, with Lemma \[Lemma: basic lemma 1\] being replaced by Lemma \[Lemma: basic lemma 2\], the same argument as in the proof of Theorem \[Thm: convergence Rm of CCY with torus fiber\] gives the convergence estimates $$\label{Equ: bound of Rm tensor} \|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t).$$ Next we consider the Ricci curvature. First from and Lemma \[Lemma: basic lemma 3\], we have $$\label{Equ: bound of Rm tensor 2} \|{\mathrm{Rm}}({\tilde{{\omega}}}(t))\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ and hence naturally $$\label{Equ: bound of Ric tensor 1} \|{\mathrm{Ric}}({\tilde{{\omega}}}(t))\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ Combining and Lemma \[Lemma: basic lemma 3\] we obtain $$\label{Equ: bound of Ric tensor 3} \|{\mathrm{Ric}}({\omega}(t))\|_{C^k(U,{\tilde{{\omega}}}(t))}+\|{\mathrm{Ric}}({\tilde{{\omega}}}(t))\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{K,k}.$$ By definiton, we have $$\label{Equ: relation of Ric tensor} \begin{split} &R({\omega}(t))_{i\bar{j}}-R({\tilde{{\omega}}}(t))_{i\bar{j}}\\ =&g(t)^{k\bar{l}}R({\omega}(t))_{i\bar{j}k\bar{l}}-{\tilde{g}}(t)^{k\bar{l}}R({\tilde{{\omega}}}(t))_{i\bar{j}k\bar{l}}\\ =&\left[g(t)^{k\bar{l}}-{\tilde{g}}(t)^{k\bar{l}}\right]R({\omega}(t))_{i\bar{j}k\bar{l}}+{\tilde{g}}(t)^{k\bar{l}}\left[R({\omega}(t))_{i\bar{j}k\bar{l}}-R({\tilde{{\omega}}}(t))_{i\bar{j}k\bar{l}}\right].\\ \end{split}$$ Taking norms with respect to ${\tilde{{\omega}}}(t)$ gives that on $U$ $$\begin{split} &\left|{\mathrm{Ric}}({\omega}(t))-{\mathrm{Ric}}({\tilde{{\omega}}}(t))\right|_{{\tilde{{\omega}}}(t)}\\ \leq& C\cdot \left|{\tilde{{\omega}}}(t)^{-1}-{\omega}(t)^{-1}\right|_{{\tilde{{\omega}}}(t)}\cdot \left|{\mathrm{Rm}}({\omega}(t)\right|_{{\tilde{{\omega}}}(t)}+C\cdot \left|{\mathrm{Rm}}({\omega}(t))-{\mathrm{Rm}}({\tilde{{\omega}}}(t))\right|_{{\tilde{{\omega}}}(t)}\\ \leq& C\cdot h_K(t)\cdot C_{K}+C\cdot h_{K}(t)\\ \leq&h_K(t).\\ \end{split}$$ Hence the Conditions $(A)-(C)$ of Lemma \[Thm: BIC principle\] with $$\eta(t)={\mathrm{Ric}}({\omega}(t))-{\mathrm{Ric}}({\tilde{{\omega}}}(t))$$ are satisfied, and we conclude from Lemma \[Thm: BIC principle\] the convergence estimates $$\label{Equ: bound of Ric tensor} \|{\mathrm{Ric}}({\omega}(t))-{\mathrm{Ric}}({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t).$$ Finally, for the scalar curvature, from Equation and Lemma \[Lemma: basic lemma 3\], we naturally have $$\label{Equ: bound of scalar curvature} \|R({\omega}(t))\|_{C^k(U)}+\|R({\tilde{{\omega}}}(t))\|_{C^k(U)}\leq C_{K,k}.$$ By definition we have $$\label{Equ: relation of scalar curvature} \begin{split} &R({\omega}(t))-R({\tilde{{\omega}}}(t))\\ =&g(t)^{k\bar{l}}R({\omega}(t))_{k\bar{l}}-{\tilde{g}}(t)^{k\bar{l}}R({\tilde{{\omega}}}(t))_{k\bar{l}}\\ =&\left[g(t)^{k\bar{l}}-{\tilde{g}}(t)^{k\bar{l}}\right]R({\omega}(t))_{k\bar{l}}+{\tilde{g}}(t)^{k\bar{l}}\left[R({\omega}(t))_{k\bar{l}}-R({\tilde{{\omega}}}(t))_{k\bar{l}}\right].\\ \end{split}$$ As before, we have on $U$ $$\begin{split} &\left|R({\omega}(t))-R({\tilde{{\omega}}}(t))\right|\\ \leq& C\cdot \left|{\tilde{{\omega}}}(t)^{-1}-{\omega}(t)^{-1}\right|_{{\tilde{{\omega}}}(t)}\cdot \left|{\mathrm{Ric}}({\omega}(t))\right|_{{\tilde{{\omega}}}(t)}+C\cdot \left|{\mathrm{Ric}}({\omega}(t))-{\mathrm{Ric}}({\tilde{{\omega}}}(t))\right|_{{\tilde{{\omega}}}(t)}\\ \leq& C\cdot h_K(t)\cdot C_{K}+C\cdot h_{K}(t)\leq h_K(t).\\ \end{split}$$ Hence the Conditions $(A)-(C)$ of Lemma \[Thm: BIC principle\] with $$\eta(t)=R({\omega}(t))-R({\tilde{{\omega}}}(t))$$ are satisfied, and we conclude from Lemma \[Thm: BIC principle\] that $$\|R({\omega}(t))-R({\tilde{{\omega}}}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t).$$ This completes the proof of Theorem \[Thm: convergence of KRF with torus fiber\]. Finally, we prove Theorem \[Thm: convergence Ric curvature to GKE of KRF with torus fiber\]. First, we need the following Proposition. \[Prop: higher-order derivative bound of GKE\] For all compact sets $U\subset\subset X\backslash S$ and all $k\geq 0$, there exists constants $C_{U,k}$ independent of $t$ such that for all $t\in [0,\infty)$ we have that $$\label{Equ: higher-order derivative bound of GKE} \|{\omega}_B\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{U,k}.$$ We adopt the notations of [@To4 Theorem 5.24, p363-368]. We just need to consider the case when $X_b$ is in fact bi-holomorphic to a torus for some $b\in B\backslash S'$. Then we have $$\lambda_t^*p^*f^*{\omega}_B=p^*f^*{\omega}_B,$$ and $$\lambda_t^*p^*{\omega}_{{\mathrm{SF}}}=\lambda_t^*{\sqrt{-1} \partial \overline{\partial}}\eta={\sqrt{-1} \partial \overline{\partial}}(\eta\circ\lambda_t)=e^t{\sqrt{-1} \partial \overline{\partial}}\eta=e^tp^*{\omega}_{{\mathrm{SF}}}.$$ Hence we have $$\lambda_t^*p^*({\tilde{{\omega}}}(t))=\lambda_t^*p^*((1-e^{-t})f^*{\omega}_B+e^{-t}{\omega}_{{\mathrm{SF}}})=(1-e^{-t})p^*f^*{\omega}_B+p^*{\omega}_{{\mathrm{SF}}}=:\Check{{\omega}}(t).$$ Now we may assume that $t\geq 1$ so that the factor $(1-e^{-t})\in (\frac{1}{2},1)$ would be a harmless factor. Then we have for every compact set $K\subset B'\times\mathbb{C}^{n-m}$ (here $B'$ is the $B$ in the notations of [@To4 Theorem 5.24, p363-368], which is a small ball in the regular part of the base space $B$ here) there are constants $C_{K,k}$ such that $$\|p^*f^*{\omega}_B\|_{C^k(K,\Check{{\omega}}(t))}\leq C_{K,k},$$ for all $t\geq 1$. We can rewrite this as $$\|\lambda_t^*p^*f^*{\omega}_B\|_{C^k(K,\lambda_t^*p^*({\tilde{{\omega}}}(t)))}\leq C_{K,k}.$$ Given any open subset $U\subset X\backslash S$, if $K'\subset\subset U\subset X\backslash S$ is a compact set which is small enough so that $K=p^{-1}(K')\subset B'\times\mathbb{C}^{n-m}$ is compact and $p$ is a bi-holomorphism on $K$, (note that such compact sets $K'$ cover $U$) then we have $$\|f^*{\omega}_B\|_{C^k(K',{\tilde{{\omega}}}(t))}=\|p^*f^*{\omega}_B\|_{C^k(K,p^*f^*{\tilde{{\omega}}}(t))}=\|\lambda_t^*p^*f^*{\omega}_B\|_{C^k(\lambda_{1/t}(K),\lambda_t^*p^*({\tilde{{\omega}}}(t)))},$$ where $\lambda_{1/t}$ is the inverse map of $\lambda_t$. But the compact sets $\lambda_{1/t}(K)$ ($t\geq 1$) are all contained in a fixed compact set of $B'\times\mathbb{C}^{n-m}$, hence we have $$\|f^*{\omega}_B\|_{C^k(K',{\tilde{{\omega}}}(t))}=\|\lambda_t^*p^*f^*{\omega}_B\|_{C^k(\lambda_{1/t}(K),\lambda_t^*p^*({\tilde{{\omega}}}(t)))}\leq C_{K',k},$$ and so by a covering argument we easily obtain $$\|{\omega}_B\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{U,k}.$$ This finishes the proof of Proposition \[Prop: higher-order derivative bound of GKE\]. The same conclusion of Proposition \[Prop: higher-order derivative bound of GKE\] holds with ${\omega}_B$ being replaced by any other fixed Kähler metric on the regular part of the base space $B$. Now we can prove Theorem \[Thm: convergence Ric curvature to GKE of KRF with torus fiber\]. Along the normalized Kähler-Ricci flow , we have on $X\backslash S\times[0,\infty)$ $$\label{Equ: relation of Ric and u} \begin{split} {\mathrm{Ric}}({\omega}(t)) &=-{\sqrt{-1} \partial \overline{\partial}}(\dot{{\varphi}}+{\varphi})-\chi\\ &=-{\sqrt{-1} \partial \overline{\partial}}(\dot{{\varphi}}+{\varphi}-v)-(\chi+{\sqrt{-1} \partial \overline{\partial}}v)\\ &=-{\sqrt{-1} \partial \overline{\partial}}u-{\omega}_B.\\ \end{split}$$ We define $\eta(t)$ to be the $(1,1)$ form $$\eta(t)={\mathrm{Ric}}({\omega}(t))+{\omega}_B=-{\sqrt{-1} \partial \overline{\partial}}u.$$ Combining Lemma \[Lemma: basic lemma 3\] with Proposition \[Prop: higher-order derivative bound of GKE\], we have $$\label{Equ: higher-order bound for eta} \|\eta(t)\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq \|{\mathrm{Ric}}({\omega}(t))\|_{C^k(K,{\tilde{{\omega}}}(t))}+\|{\omega}_B\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq C_{K,k},$$ for any compact subset $K\subset\subset X\backslash S$. We need to prove that $$\label{Equ: general order convergence for eta} \|\eta(t)\|_{C^0(K,{\tilde{{\omega}}}(t))}\leq h_K(t),$$ for any compact subset $K\subset\subset X\backslash S$. To this end, we choose disks $B_1\subset\subset B_2\subset\subset B\backslash S'$ such that $K\subset f^{-1}(B_1)$. Set $U=f^{-1}(B_2)$. According to Lemma \[Lemma: basic lemma 2\], we have the estimate $$|\nabla u|_{{\tilde{{\omega}}}(t)}^2\leq h_K(t)$$ on $U\times[0,\infty)$. Combining Equations and (with $K$ being replaced by $U$ in Equation ), we have on $U\times[0,\infty)$ $$\begin{split} &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|\nabla u\right|_{{\tilde{{\omega}}}(t)}^2\right)\\ =&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2-{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_a\nabla^{{\tilde{g}}(t)}_{\bar{b}}\nabla^{{\tilde{g}}(t)}_i{u}\nabla^{{\tilde{g}}(t)}_{\bar{j}}{u}-{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_i{u}\nabla^{{\tilde{g}}(t)}_a\nabla^{{\tilde{g}}(t)}_{\bar{b}}\nabla^{{\tilde{g}}(t)}_{\bar{j}}{u}\\ =&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2-{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_a\nabla^{{\tilde{g}}(t)}_i\nabla^{{\tilde{g}}(t)}_{\bar{b}}{u}\nabla^{{\tilde{g}}(t)}_{\bar{j}}{u}\\ &-{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_i{u}\left\{\nabla^{{\tilde{g}}(t)}_{\bar{b}}\nabla^{{\tilde{g}}(t)}_a\nabla^{{\tilde{g}}(t)}_{\bar{j}}{u}+{{\mathrm{Rm}}({\tilde{{\omega}}}(t))^{\bar{q}}}_{\bar{j}a\bar{b}}\nabla^{{\tilde{g}}(t)}_{\bar{q}}{u}\right\}\\ =&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2+{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_a\eta_{i\bar{b}}\nabla^{{\tilde{g}}(t)}_{\bar{j}}{u}-{\tilde{g}}(t)^{a\bar{b}}{\tilde{g}}(t)^{i\bar{j}}\cdot\nabla^{{\tilde{g}}(t)}_i{u}\left\{-\nabla^{{\tilde{g}}(t)}_{\bar{b}}\eta_{a\bar{j}}+{{\mathrm{Rm}}({\tilde{{\omega}}}(t))^{\bar{q}}}_{\bar{j}a\bar{b}}\nabla^{{\tilde{g}}(t)}_{\bar{q}}{u}\right\}\\ =&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2+\nabla^{{\tilde{g}}(t)}{\eta}*\nabla^{{\tilde{g}}(t)}{u}+{\mathrm{Rm}}({\tilde{{\omega}}}(t))*\nabla^{{\tilde{g}}(t)}{u}*\nabla^{{\tilde{g}}(t)}{u}\\ \leq&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot\left|\nabla^{{\tilde{g}}(t)}{\eta}\right|_{{\tilde{{\omega}}}(t)}\cdot\left|\nabla^{{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}+C\cdot\left|{\mathrm{Rm}}({\tilde{{\omega}}}(t))\right|_{{\tilde{{\omega}}}(t)}\cdot\left|\nabla^{{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2\\ \leq&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot\left|\nabla^{{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}\\ \leq&-2\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2+h(t),\\ \end{split}$$ where $*$ denotes tensor contraction by ${\tilde{g}}(t)$. Also, we naturally have $$\left|\nabla^{2,{\tilde{g}}(t)}{u}\right|_{{\tilde{{\omega}}}(t)}^2\geq \left|\eta\right|_{{\tilde{{\omega}}}(t)}^2.$$ Hence we conclude that $$\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|\nabla u\right|_{{\tilde{{\omega}}}(t)}^2\right)\leq-2\left|\eta\right|_{{\tilde{{\omega}}}(t)}^2+h(t).\\$$ Next, using we have $$\label{Equ: evolution inequality of eta for k=0} \begin{split} &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right)\\ =& -2\left|\nabla^{{\tilde{g}}(t)}{\eta}\right|_{{\tilde{{\omega}}}(t)}^2+\nabla^{2,{\tilde{g}}(t)}{\eta}*{\eta}\\ \leq& C\cdot\left|\nabla^{2,{\tilde{g}}(t)}{\eta}\right|_{{\tilde{{\omega}}}(t)}\cdot\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}\\ \leq& C\cdot\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}.\\ \end{split}$$ where $*$ denotes tensor contraction by ${\tilde{g}}(t)$. Hence using we can further compute on $U$ $$\label{Equ: evolution inequality of eta for k=0 2} \begin{split} &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\right)\\ \leq& 2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\cdot\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right)-2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2\\ \leq& C\cdot\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^3-2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2\\ \leq& C\cdot\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2-2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2.\\ \end{split}$$ Choose cut-off function $\rho$ as in the proof of Theorem \[Thm: convergence of higher order estimate for CY\], and use again we can compute on $U$ $$\begin{split} &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\rho^2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\right)\\ =& \rho^2\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\right)+\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\rho^2\right)-2\mathrm{Re}\left\{\left<\nabla\rho^2, \nabla \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\right>_{{\tilde{{\omega}}}(t)}\right\}\\ \leq& \rho^2\left(C\cdot\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2-2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2\right)+C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4+C\cdot\rho|\nabla \rho|_{{\tilde{{\omega}}}(t)}\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\left|\nabla \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}\\ \leq& -2\rho^2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot\left(\rho\left|\nabla \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}\right)\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\\ \leq& -2\rho^2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2+\rho^2\left|\nabla\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\right|_{{\tilde{{\omega}}}(t)}^2+C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4\\ \leq& C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2.\\ \end{split}$$ Now we conclude that we can find some $h(t)$ such that on $U$ $$\label{Equ: inequality for higher order 2} \left\{ \begin{aligned} &|\nabla u|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}\leq 1,\\ &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(|\nabla u|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}\right)\leq-2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}+1,\\ &\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(\rho^2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4h(t)^{-1}\right)\leq C\cdot \left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}.\\ \end{aligned} \right.$$ Set $$Q:=\rho^2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^4h(t)^{-1}+C\cdot |\nabla u|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}.$$ Using , on $U\times[0,\infty)$ we have $$\left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(Q\right)\leq -\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2h(t)^{-1}+C.$$ Now, at a given time $t$, assume $Q$ achieves it’s maximum at point $x_0$. If $x_0\in {\partial}U$, where $\rho\equiv 0$, using , $Q$ has an upper bound $C$ at time $t$, and we are done. Otherwise $x_0\in U$ and by maximum principle, we have $$0\leq \left(-\Delta_{{\tilde{{\omega}}}(t)}\right)\left(Q\right)(x_0)\leq -\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2(x_0)h(t)^{-1}+C,$$ which gives $$\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2(x_0)h(t)^{-1}\leq C.$$ Then by and we have on $U$ $$Q\leq Q(x_0)\leq C_{K}^2\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2(x_0)h(t)^{-1}+C\cdot |\nabla u|_{{\tilde{{\omega}}}(t)}^2(x_0)h(t)^{-1}\leq C.$$ Since $\rho\equiv 1$ on $K$, we obtian the estimate $$\left|{\eta}\right|_{{\tilde{{\omega}}}(t)}^2\leq Ch(t)^{\frac{1}{2}}$$ on $K$. Hence we conclude $$\label{Equ: conditions of Ric} \left\{ \begin{aligned} &\|\eta(t)\|_{C^0(U,{\tilde{{\omega}}}_t)}\leq h_K(t),\\ &\|\eta(t)\|_{C^k(U,{\tilde{{\omega}}}_t)}\leq C_{K,k}.\\ \end{aligned} \right.$$ Now Conditions $(A)$ and $(B)$ of Lemma \[Thm: BIC principle\] are satisfied for $\eta(t)$. Hence from Lemma \[Thm: BIC principle\] we conclude the local convergence $$\label{Equ: convergence of Ric} \|{\mathrm{Ric}}({\omega}(t))+{\omega}_B\|_{C^k(K,{\tilde{{\omega}}}(t))}\leq h_{K,k}(t).$$ Finally, for the scalar curvature, Lemma \[Lemma: basic lemma 3\] gives that $$\|R({\omega}(t))\|_{C^k(U,{\tilde{{\omega}}}(t))}\leq C_{K,k}(t).$$ The estimate or the main result of [@J] implies that $$\|R({\omega}(t))+m\|_{C^0(U)}\leq h(t).$$ Hence set $${\tilde{\eta}}(t)=R({\omega}(t))+m$$ and apply Lemma \[Thm: BIC principle\], we conclude the local convergence $$\|R({\omega}(t))+m\|_{C^k(K,{\tilde{{\omega}}}_t)}\leq h_{K,k}(t).$$ This establish and hence completes the proof of Theorem \[Thm: convergence Ric curvature to GKE of KRF with torus fiber\]. [99]{} T. Aubin, [*Équations du type Monge-Ampère sur les variétés kähleriennes compactes*]{}, C. R. Acad. Sci. Paris Sér. A-B [**283**]{} (1976), no. 3, Aiii, A119–A121. W. Fischer, H. Grauert, [*Lokal-triviale Familien kompakter komplexer Mannigfaltigkeiten*]{}, Nachr. Akad. Wiss. Göttingen Math.-Phys. Kl. II (1965), 89–94. F.T.-H. Fong, Y.S. Zhang, [*Local curvature estimates of long-time solutions to the Kähler-Ricci flow*]{}, preprint, arXiv:1903.05939. F.T.-H. Fong, Z. Zhang, *The collapsing rate of the Kähler-Ricci flow with regular infinite time singularity*, J. reine angew. Math. [**703**]{} (2015), 95–113. M. Gill, [*Convergence of the parabolic complex Monge-Ampère equation on compact Hermitian manifolds*]{}, Comm. Anal. Geom. [**19**]{} (2011), no. 2, 277–303. M. Gill, [*Collapsing of products along the Kähler-Ricci flow*]{}, Trans. Amer. Math. Soc. [**366**]{} (2014), no. 7, 3907–3924. B. Greene, A. Shapere, C. Vafa, S.-T. Yau, [*Stringy cosmic strings and noncompact Calabi-Yau manifolds*]{}, Nuclear Phys. B [**337**]{} (1990), no. 1, 1–36. M. Gross, V. Tosatti, Y. Zhang, [*Collapsing of abelian fibered Calabi-Yau manifolds*]{}, Duke Math. J. [**162**]{} (2013), no. 3, 517–551. M. Gross, V. Tosatti, Y. Zhang, [*Gromov-Hausdorff collapsing of Calabi-Yau manifolds*]{}, Comm. Anal. Geom. [**24**]{} (2016), no. 1, 93–113. M. Gross, P.M.H. Wilson, [*Large complex structure limits of $K3$ surfaces*]{}, J. Differential Geom. [**55**]{} (2000), no. 3, 475–546. H.-J. Hein, [*Gravitational instantons from rational elliptic surfaces*]{}, J. Amer. Math. Soc. [**25**]{} (2012), no. 2, 355–393. H.-J. Hein, [*A Liouville theorem for the complex Monge-Ampère equation on product manifolds*]{}, to appear in Comm. Pure Appl. Math. H.-J. Hein, V. Tosatti, [*Remarks on the collapsing of torus fibered Calabi-Yau manifolds*]{}, Bull. Lond. Math. Soc. [**47**]{} (2015), no. 6, 1021–1027. H.-J. Hein, V. Tosatti, [*Higher-order estimates for collapsing Calabi-Yau metrics*]{}, preprint, arXiv:1803.06697. W. Jian, [*Convergence of scalar curvature of Kähler-Ricci flow on manifolds of positive Kodaira dimension*]{}, preprint, arXiv:1805.07884. W. Jian, Y. Shi, and J. Song, [*A remark on constant scalar curvature Kähler metrics on minimal models*]{}, arXiv:1805.06863v1, to appear at Proceedings of A.M.S. Y. Li, [*On collapsing Calabi-Yau fibrations*]{}, preprint, arXiv:1706.10250. C. Li, J. Li, X. Zhang, [*A mean value formula and a Liouville theorem for the complex Monge-Ampère equation*]{}, preprint, arXiv:1709.05754. J. Song, G. Tian, [*The Kähler-Ricci flow on surfaces of positive Kodaira dimension*]{}, Invent. Math. [**170**]{} (2007), no. 3, 609–653. J. Song, G. Tian, [*Canonical measures and Kähler-Ricci flow*]{}, J. Amer. Math. Soc. [**25**]{} (2012), no. 2, 303–353. J. Song, G. Tian, [*Bounding scalar curvature for global solutions of the Kähler-Ricci flow*]{}, Amer. J. Math. [**138**]{} (2016), no. 3, 683–695. J. Song, G. Tian, [*The Kähler-Ricci flow through singularities*]{}, Invent. Math. [**207**]{} (2017), 519-595. J. Song, G. Tian and Z. Zhang, [*Collapsing behavior of Ricci-flat Kahler metrics and long time solutions of the Kahler-Ricci flow*]{}, preprint, arXiv:1904.08345. J. Song, B. Weinkove, [*Contracting exceptional divisors by the Kähler-Ricci flow*]{}, Duke Math. J. [**162**]{} (2013), no. 2, 367–415. J. Song, B. Weinkove, [*Contracting exceptional divisors by the Kähler-Ricci flow II*]{}, Proc. Lond. Math. Soc. (3) [**108**]{} (2014), no. 6, 1529–1561. J. Song, B. Weinkove, [*Introduction to the Kähler-Ricci flow*]{}, Chapter 3 of ‘Introduction to the Kähler-Ricci flow’, eds S. Boucksom, P. Eyssidieux, V. Guedj, Lecture Notes Math. 2086, Springer 2013. V. Tosatti, [*Adiabatic limits of Ricci-flat Kähler metrics*]{}, J. Differential Geom. [**84**]{} (2010), no. 2, 427–453. V. Tosatti, [*Degenerations of Calabi-Yau metrics*]{}, in [*Geometry and Physics in Cracow,*]{} Acta Phys. Polon. B Proc. Suppl. [**4**]{} (2011), no. 3, 495–505. V. Tosatti, [*Calabi-Yau manifolds and their degenerations*]{}, Ann. N.Y. Acad. Sci. [**1260**]{} (2012), 8–13. V. Tosatti, [*KAWA lecture notes on the Kähler-Ricci flow*]{}, to appear in Ann. Fac. Sci. Toulouse Math. V. Tosatti, Y. Zhang, [*Infinite time singularities of the Kähler-Ricci flow*]{}, Geom. Topol. [**19**]{} (2015), no. 5, 2925–2948. V. Tosatti, Y. Zhang, [*Collapsing hyperkähler manifolds*]{}, preprint, arXiv:1705.03299. V. Tosatti, Y. Zhang, [*Triviality of fibered Calabi-Yau manifolds without singular fibers*]{}, Math. Res. Lett. [**21**]{} (2014), no.4, 905–918. V. Tosatti, B.Weinkove, and X.K. Yang, [*The Kähler-Ricci flow, Ricci-flat metrics and collapsing limits*]{}, arXiv:1408.0161, to appear in Amer. J. Math. G. Tian, Z.L. Zhang, [*Convergence of Kähler-Ricci flow on lower-dimensional algebraic manifolds of general type*]{}, Int. Math. Res. Not. IMRN 2016, no. 21, 6493-6511 G. Tian, Z.L. Zhang, [*Relative volume comparison of Ricci flow and its applications*]{}, arXiv:1802.09506 G. Tian, Z. Zhang, [*On the Kähler-Ricci flow on projective manifolds of general type*]{}, Chinese Ann. Math. Ser. B [**27**]{} (2006), no. 2, 179–192. S.T. Yau, [*On the Ricci curvature of a compact Kähler manifold and the complex Monge-Ampère equation. I.*]{}, Comm. Pure Appl. Math. 31 (1978), no. 3, 339-411. S.-T. Yau, [*A general Schwarz lemma for Kähler manifolds*]{}, Amer. J. Math. [**100**]{} (1978), no. 1, 197–203. Y. Zhang, [*Infinite-time singularity type of the Kähler-Ricci flow*]{}, J. Geom. Anal. (2017), https://doi.org/10.1007/s12220-017-9949-2. Y. Zhang, [*Collapsing limits of the Kähler-Ricci flow and the continuity method*]{}, Math. Ann. (2018), https://doi.org/10.1007/s00208-018-1676-x. Y. Zhang, [*Infinite-time singularity type of the Kähler-Ricci flow II*]{}, Math. Res. Lett. (2019) (to appear), arXiv:1809.01305. Z. Zhang, [*Scalar curvature bound for Kähler-Ricci flows over minimal manifolds of general type*]{}, Int. Math. Res. Not. 2009; doi: 1093/imrn/rnp073 Z. Zhang, [*Scalar curvature behavior for finite time singulairty of Kähler-Ricci flow*]{}, Michigan Math. J. [**59**]{} (2010), no. 2, 419–433

--- abstract: 'Two identical 1D autocatalytic systems with Gray–Scott kinetics—driven towards convectively unstable regimes and submitted to independent spatiotemporal Gaussian white noises—are coupled unidirectionally, but otherwise linearly. Numerical simulation then reveals that (even when perturbed by noise) the slave system replicates the convective patterns arising in the master one to a very high degree of precision, as indicated by several measures of synchronization.' address: - | Departamento de Física, Facultad de Ciencias Exactas y Naturales, Universidad Nacional de Mar del Plata,\ Deán Funes 3350, 7600 Mar del Plata, Argentina. - | Grupo de Física Non Lineal, Facultade de Física, Universidade de Santiago de Compostela,\ E-15782 Santiago de Compostela, Spain. author: - 'Gonzalo Izús$^*$' - Roberto Deza$^+$ - 'Luis Bernal$^*$' - 'Vicente Pérez-Villar' title: 'Complete synchronization of convective patterns between Gray–Scott systems' --- $^*$ Member of CONICET, Argentina.\ $^+$ Invited talk at the Workshop on “*Complex Systems: New Trends and Expectations*”, Santander (Spain) 5–9 June 2006. Introduction {#intro} ============ The beautiful talk in this Workshop by Jürgen Kurths [@Kurths] (as well as others dealing to some extent with the subject) exempts us to introduce the field of synchronization at large. Thus, hereafter we shall restrict our scope to the less explored subfield of the synchronization between continuous systems [@Parmananda; @Kocarev; @Junge; @Boccaletti]—concretely, to non-delayed synchronization between systems of stochastic partial differential equations. In particular, a topic that has been hardly addressed is the synchronization between noise-sustained structures (NSS) in systems undergoing a convective instability [@opo2]. A *convectively* unstable regime is characterized by the fact that local perturbations to the steady state are advected more rapidly than their spreading rate [@Deissler]. When seen in a Lagrangian framework, the system is unstable; from an Eulerian description, however, perturbations are “washed out by the flow”. Macroscopic patterns named *noise sustained structures* (NSS) emerge in this regime if noise is present at all times. It is through dynamical amplification of random fluctuations that the system is driven out of its linearly unstable steady state towards the state sustaining NSS. Thus, if noise (or any external deterministic forcing) were not present, nonequilibrium structures could not arise. In fluid dynamics the NSS are a spatial macroscopic manifestation of amplified thermal fluctuations. NSS have been observed in fluid convection experiments (both in open flow configuration [@ahlers] and Taylor–Couette flows [@Bab; @Tsam]), and their precursors have been also observed in nematic liquid crystals [@Rehberg]. They have also been numerically shown to exist in optical systems [@opo2; @Marco] (driven in this case by quantum noise) and recently, in a model autocatalytic chemical reaction—the Gray–Scott (GS) model—taking place in a differential-flow reactor [@GS]. The paper is organized as follows: In Sec. \[sec:model\] a brief sketch is made of the GS model, and the rationale and features of the chosen master–slave coupling are pointed out. Section \[sec:numsim\] introduces the details of the numerical integration scheme and discusses the features of the NSS arising in the uncoupled systems. Section \[sec:sync1\] is devoted to a fairly thorough numerical characterization of the replication of NSS through complete synchronization. In particular, the behavior of the synchronization measures as functions of the parameters in the model is studied, and a numerical estimation is made of the robustness of the phenomenon. Finally, the main conclusions are summarized in Sec. \[sec:concl\]. The model {#sec:model} ========= The GS model proposes three steps for the conversion of the precursor species $P$ into the inert product $C$ $$\begin{aligned} P&\stackrel{k_0}{\rightarrow}&A\\ A+2B&\stackrel{k_1}{\rightarrow}&3B\\ B&\stackrel{k_2}{\rightarrow}&C\end{aligned}$$ The intermediate step has cubic autocatalytic kinetics. In the case we consider, the reaction takes place in a differential-flow reactor where $A$ is immobilized, whereas $B$ is free to diffuse and is also advected by the flow. Moreover, the reaction is maintained out of equilibrium by keeping the concentration of $P$ constant ($p=p_0$) and that of $C$ zero ($c=0$). Hence, the present model describes the irreversible decay of $P$ towards a product $C$ that is immediately removed from the reactor. After scaling concentrations by $(k_2/k_1)^{1/2}$, time by $k_2^{-1}$ and length by $(D_B/k_2)^{1/2}$, the rate equations for system 1 read $$\begin{aligned} \label{eq:master1} \frac{\partial a_1}{\partial t}&=&\mu-a_1\,b_1^2+\sqrt{\sigma_1}\, \xi_1(\mathbf{r},t),\nonumber\\ \frac{\partial b_1}{\partial t}&=&\nabla^2b_1-\phi\,\nabla b_1-b_1+a_1\,b_1^2,\end{aligned}$$ where $\mu$ stands for the scaled version of $k_0p_0$ and $\phi$ for that of the fluid velocity $v$. The real Gaussian noise $\xi_1$ in the rate equation for $a_1$—with zero mean, variance $\sigma_1$, and delta-correlated in space and time—accounts for fluctuations (either thermal or in $p_0$). For $\mu>1$, the uniform steady state $(a_1=\mu^{-1},b_1=\mu)$ becomes convectively unstable at some $\phi_c(\mu)$, yielding to traveling waves with $\pm q_c$. Further details are found in [@GS] and references therein. Now we assume that system 1 drives another system (called thereafter system 2 and lying in a second differential-flow reactor) in a master–slave configuration. System 2 has the same values of $\mu$ and $\phi$ but its $A$–component is submitted to a spatiotemporal Gaussian white noise $\xi_2(\mathbf{r},t)$ with a possibly different noise variance $\sigma_2$: $$\begin{aligned} \label{eq:master2} \frac{\partial a_2}{\partial t}&=&\mu-a_2\,b_2^2+\sqrt{\sigma_2}\,\xi_2(\mathbf{r},t),\nonumber \\ \frac{\partial b_2}{\partial t}&=&\nabla^2b_2-\phi\,\nabla b_2-b_2+a_2\,b_2^2+\epsilon\,(b_1-b_2).\end{aligned}$$ $\epsilon$ denotes the strength of the unidirectional *linear* coupling between both reactions. Besides being the simplest coupling that enables synchronization, it facilitates an approach to the stability analysis of the synchronization manifold [@IDB]. Numerical simulation and unsynchronized NSS {#sec:numsim} =========================================== We shall restrict hereafter to the 1D case (the specificities found in higher spatial dimensions will be published elsewhere [@IDB]). Equations (\[eq:master1\]) and (\[eq:master2\]) have been integrated using an Euler stochastic scheme in a grid of $16384$ sample points with a grid space $\Delta x=0.1$ and time step $\Delta t=0.0001$. The parameters have been chosen as $\mu=2.0$, $\phi=9.5$, $\sigma_1=10^{-7}$. For each system, Dirichlet BC is assumed at the inlet of the reaction domain \[$a_i(0,t)=\mu^{-1},\,b_i(0,t)=\mu,\,(i=1,2)$\] and Neumann BC at the outlet ($x=L$). The length $L$ is chosen in such a way that spatiotemporal patterns develop well before they reach the outlet. For $\epsilon=0$, Eqs. (\[eq:master1\])–(\[eq:master2\]) describe two uncoupled reactions, identical with regard to the deterministic parameters but submitted to independent spatiotemporal noises which produce non-correlated NSS in both systems. These patterns have been characterized in Ref. [@GS]: they are dynamical structures that drift with the flow, disappearing on the right, whereas new wave excitations are continuously regenerated by dynamical amplification of noise. \ \ Synchronization of noise-sustained structures {#sec:sync1} ============================================= When $\epsilon\neq0$, some correlation is expected between the NSS in both systems. A handy measure of correlation for these snapshots are the deviation fields $$\begin{aligned} \label{eq:dif} \alpha(x,t)&=&a_1(x,t)-a_2(x,t),\nonumber\\ \beta(x,t)&=&b_1(x,t)-b_2(x,t).\end{aligned}$$ Figures \[fig:1\]a and \[fig:1\]b are respectively snapshots of typical $a_1$ and $b_1$ profiles. The deviation fields are also depicted (in solid lines and in the same scales as $a_1$, $b_1$ respectively) in Figs. \[fig:1\]a and \[fig:1\]b. The result is surprising, given that system 2 is also submitted to an independent spatiotemporal noise source: In the scales of Figs. \[fig:1\], *system 2 synchronizes completely to system 1*. In other words, an effective replication of the NSS arising in the first reactor takes place at the second one, and time evolution—even under the influence of noise—does not spoil the high degree of synchronization. If we regard the system’s evolution as a succession of snapshots like those of Figs. \[fig:1\], natural quantifiers for this phenomenon (as functions of $t$) are the variances of $\alpha$ and $\beta$: $$\begin{aligned} \label{eq:var} \sigma_\alpha(t)&\equiv& \sqrt{\frac{1}{L}\int_0^L[\alpha^2(x,t)-\langle\alpha\rangle^2]\,dx},\nonumber\\ \sigma_\beta(t)&\equiv&\sqrt{\frac{1}{L}\int_0^L [\beta^2(x,t)-\langle\beta\rangle^2]\,dx},\end{aligned}$$ with $\langle\varphi\rangle_{(t)}\equiv\frac{1}{L}\int_0^L\varphi(x,t)\,dx$ ($\varphi$ stands for $\alpha$ and $\beta$ respectively), and the global synchronization error $$\label{eq:global} E(t)=\sqrt{\frac{1}{L}\int_0^L[\alpha^2(x,t)+\beta^2(x,t)]\,dx}.$$ Figure \[fig:2\] is a plot of $E$, $\sigma_\alpha$ and $\sigma_\beta$ vs $\epsilon$ for typical realizations (as stated before, the time evolution preserves the degree of synchrony). In the numerical simulation, $\langle\alpha\rangle$ and $\langle\beta\rangle$ remain below $\sim 5 \times 10^{-5}$, i.e., $E^2\sim\sigma_\alpha^2+\sigma_\beta^2$ and basically $E$ accumulates the information of both variances. On the other hand, correlations between the NSS remain during time evolution above $.9999$, indicating a very high degree of structure replication. In other words, the coupling in Eqs. (\[eq:master2\]) synchronizes the whole stochastic processes. A dependence of $\sigma_\alpha$, $\sigma_\beta$ (and thus of $E$) on the noise intensity $\sigma_2$ is to be expected. Figure \[fig:3\] shows (here $\sigma_1= 10^{-7}$) that this is indeed the case, and maximum synchronization corresponds to $\sigma_2=0$. In other words, only in system 1 does the noise play a constructive role (by pushing the system out of its unstable steady state). \ \ As usual, the synchronization between the stochastic fields $a_1$ and $a_2$ (resp. $b_1$ and $b_2$) can also be viewed in the corresponding phase planes. As an illustration, Fig. \[fig:4\] shows the dynamical correlation between $b_1$ and $b_2$ during the complete time history of a numerical simulation. \ One might wonder whether the proposed coupling is general enough, or whether the reported phenomenon is robust. To elucidate (at least partially) on these questions, we have performed numerical experiments where the coupling is switched on only after two well-developed and independent NSS are formed in each reactor. The results are shown in Fig. \[fig:5\]. The synchronization error decreases as soon as the coupling is switched on, a replication of system 1’s NSS takes place after a transient, regardless of the initial condition on reactor 2. In particular, there is no need to stabilize the second reactor prior to synchronizing it with the first one. This fact shows explicitly the robustness of the observed phenomenon, and suggests that the attraction basin of the synchronization manifold ($a_2=a_1$, $b_2=b_1$) is large enough. \ In the same way, if the coupling is only switched on in a part of the reactors, numerical simulations (not shown) indicate that (even for $\sigma_2=0$) a replication of the “master structure” takes place in the coupled region of the slave system, opening the possibility of local replication (see below). We have also explored the synchronization of deterministic non-equilibrium structures (i.e, no noise in both systems). In particular, we have considered two cases of pattern generation: (a) structures generated by the time evolution of an inhomogeneous initial condition (Fig. \[fig:6\]); and (b) structures generated by an oscillatory external forcing (Fig. \[fig:7\]). In both cases, the numerical results (discused in the captions) clearly confirm the robustness of the proposed mechanism of synchronization. \ \ Conclusions {#sec:concl} =========== By coupling unidirectionally (but otherwise linearly) corresponding points of two samples of the convectively unstable system under study, complete synchronization of macroscopic structures has been achieved, both for deterministic and stochastic dynamics. Figure \[fig:5\] suggests that the synchronization attractor is very extended and a full replication of the structures is to be expected under very general conditions. This is a strong indication that the synchronization manifold ($a_2=a_1$, $b_2=b_1$) is at least linearly stable. A complete stability analysis of the synchronization manifold will be published elsewhere [@IDB]. We remark that the coupling synchronizes completely both systems (after a transient) regardless of the initial condition in the “slave” system. Even more, the coupling may be defined in part of the system’s extension. Thus, for $\sigma_2=0$, a (synchronized) structure arises in the slave system *because* of the coupling. We expect the phenomenon to have technological applications in the control of differential-flow chemical reactors, and eventually in the case where the convective structures in the master system are not noise-sustained ones, but carry useful information. [99]{} J. Kurths *et al.*, *Synchronization: from two coupled oscillators to complex networks*, this Workshop. P. Parmananda, Phys. Rev. E **56**, (1997) 1595. L. Kocarev, Z. Tasev, and U. Parlitz, Phys. Rev. Lett. **79**, (1997) 51. L. Junge and U. Parlitz, Phys. Rev. E **61**, (1999) 3736. S. Boccaletti, J. Bragard, F. T. Arecchi, and H. Mancini, Phys. Rev. Lett. **83**, (1999) 536. G. Izús, M. Santagiustina, M. San Miguel, and P. Colet, J. Opt. Soc. Am. B **16**, (1999) 1592; Phys. Rev. E **68**, (2003) 036201. R. Deissler, J. Stat. Phys. [**40**]{}, (1985) 386; [**54**]{}, (1989) 1459; Physica D [**56**]{}, (1992) 303. M. Scherer and G. Ahlers, Phys. Rev. E **65**, (2002) 051101. K. L. Babcock, G. Ahlers, and D. S. Cannell, Phys. Rev. Lett. **67**, (1991) 3388. A. Tsameret and V. Steinberg, Phys. Rev. Lett. **67**, (1991) 3392. I. Rehberg *et al.*, Phys. Rev. Lett. **67**, (1991) 596. M. Santagiustina, P. Colet, M. San Miguel, and D. Walgraef, Phys. Rev. Lett. **79**, (1997) 3633; M. Santagiustina, P. Colet, M. San Miguel, and D. Walgraef, Phys. Rev. E **58**, (1998) 3843; M. Santagiustina, P. Colet, M. San Miguel, and D. Walgraef, Opt. Express **3**, (1998) 63. B. von Haeften and G. Izús, Phys. Rev. E **67**, (2003) 056207. L. Bernal, R. Deza and G. Izús, in preparation.

**Central limit theorem for mesoscopic eigenvalue statistics\ of the free sum of matrices** Zhigang Bao\ \ [*mazgbao@ust.hk*]{} Kevin Schnelli\ \ [*schnelli@kth.se*]{} Yuanyuan Xu\ \ [*yuax@kth.se*]{} [\ ]{} Introduction {#sec_introduction} ============ In a seminal work Voiculescu [@free_Vol] showed that two large Hermitian matrices are [*asymptotically free*]{} if their eigenvectors are in general relative position. In particular, asymptotic freeness identifies the law of the sum of such large Hermitian matrices. A fundamental mechanism to generate asymptotic freeness is conjugation by independent unitary matrices that are distributed according to Haar measure. To be more specific, if $A_N$ and $B_N$ are two sequences of (deterministic and uniformly bounded) Hermitian matrices, and $U_N$ is a sequence of Haar unitaries, then $A_N$ and $U_NB_NU_N^*$ are asymptotically free and the eigenvalue distribution of the [*free sum*]{} $H_N:=A_N+U_NB_NU_N^*$ is given by the [*free additive convolution*]{}, $\mu_A\boxplus\mu_B$, of the eigenvalue distributions of $A_N$ and $B_N$ for large $N$. One way of rephrasing this result is a law of large numbers: For sufficiently regular test functions $g$, $$\begin{aligned} \label{le global law} \frac{1}{N}\sum_{i=1}^N g(\lambda_i)-\int_\R g(x){\mathrm{d}}\mu_A\boxplus\mu_B(x)\end{aligned}$$ converges in probability to zero, as $N\rightarrow\infty$. Having identified the free additive convolution as the limiting eigenvalue distribution, it is a natural question to consider fluctuations of such [*linear eigenvalue statistics*]{}. The theory of second order freeness developed by Collins, Mingo, Śniady, Speicher [@second; @order; @3; @second; @order; @1; @second; @order; @2] shows for analytic test functions $g$ that $$\begin{aligned} \label{le first clt} \sum_{i=1}^N g(\lambda_i)-\sum_{i=1}^N \E g(\lambda_i)\end{aligned}$$ converges in distribution to a centered Gaussian random variable, whose variance depends in an intricate way on the free additive convolution measure; see  below for an explicit expression for the variance. Using an analytic approach based on resolvent and characteristic function techniques similar results were obtained by Pastur and Shcherbina in [@random_book]. Conspicuously different from standard central limit theorem (CLT), the linear statistics in  is not rescaled by $N^{-1/2}$ which is explained by the strong correlations among the eigenvalues. In the present paper we are interested in the mesoscopic linear eigenvalue statistics for the free sum of matrices. We choose an energy $E$ inside the support of the free additive convolution measure, a test function $g \in C_c^2(\R)$ and consider the statistics $$\begin{aligned} \label{observable} \sum_{i=1}^N g\Big(\frac{\lambda_i-E}{\eta}\Big)-\sum_{i=1}^N \E g\Big(\frac{\lambda_i-E}{\eta}\Big)\,,\end{aligned}$$ where $\eta$ is an $N$-dependent spectral scale. The mesoscopic regime ranges over $N^{-1}\ll \eta\ll 1$, where the sum in  includes order $\eta^{-1}$ eigenvalues. For $\eta\sim 1$ we recover the macroscopic or global observable in , while for $\eta\sim N^{-1}$ the sum in  is governed by single eigenvalues, where the statistics is determined by Dyson’s sine kernel; see [@CL]. Our main result shows that in the bulk spectrum the mesoscopic linear statistic  converges to a centered Gaussian random variable with variance given by $$\begin{aligned} \frac{1}{4 \pi^2} \int_{\R} \int_{\R} \frac{(g(x_1)-g(x_2))^2}{(x_1-x_2)^2} {\mathrm{d}}x_1 {\mathrm{d}}x_2=\frac{1}{2 \pi} \int_{\R} |\xi| |\widehat{g}(\xi)|^2 {\mathrm{d}}\xi\,,\end{aligned}$$ which is the universal variance found in many other random matrices models, e.g. classical compact groups [@Soshnikov], and invariant ensembles [@meso1; @FKS]. Our approach is based on an analysis of the characteristic function of  and establishes its convergence to the characteristic function of the limiting Gaussian distribution [@character; @lytova+pastur]. Our proof has three main ingredients: Invariance properties of the Haar measure in the form of so-called Ward identities; analytic subordination for the free convolution measure; and local laws for two-point product functions of the resolvent. Ward identities are used to compute the derivative of the characteristic function and will allow us to connect the variance of fluctuations to the analytic subordination phenomenon of free probability. The Stieltjes transform of the free convolution measure can be described by an analytic change of variables from the Stieltjes transforms of the measures $\mu_A$ or $\mu_B$. This is referred to as analytic subordination [@Bia98; @Voi93], and, in fact, may be used to give an analytic definition of the free additive convolution [@belin_H; @CG]; see Theorem \[subordination\_function\] below. The subordination phenomenon carries over to the random matrix model $H_N$. The Green function or resolvent of $H_N$ is determined not only on global scales [@invariant] but also on local scales just above the microsopic scale by the Stieltjes transform of $\mu_A\boxplus\mu_B$. Such local laws giving strong rigidity estimates for the eigenvalues were established in [@eta1; @eta2] down to the optimal scale, see also [@stability; @concentrate] for previous results on some mesoscopic scales. Local laws also yield optimal speed of convergence estimates for  inside the bulk spectrum and at regular spectral edges [@edge]. In our proof we use stability properties for the subordination equations established in [@stability] and the local laws of [@eta2] to bound various error terms. A main technical difficulty in this paper is to derive systems of self-consistent equations for two-point product functions of resolvents appearing in the variance term for the linear statistics. We rely on a partial randomness decomposition of the Haar measure that was a previously used to derive local laws for the resolvent in [@eta1]. This technique allows us to exploit fluctuations on the whole mesoscopic scales and surpasses more conventional approaches where concentration with respect to the full Haar measure and Ward identities are used. Mesoscopic linear statistics were studied for Wigner matrices [@meso2; @moment; @character; @mesowigner] and many other random matrix models such as the orthogonal polynomial ensembles [@Breuer+Duits], Dyson Brownian motion [@Duits+Johansson; @character2], invariant $\beta$-ensembles [@Bekerman+lodhia; @BEYY] and random band matrix [@erdos+knowles; @erdos+knowles2]. In recent years mesoscopic central limit theorems turned out to be important tools in the theory of homogenization for Dyson’s Brownian motion (DBM) introduced by Bourgade, Erdős, Yau and Yin [@BEYY] to prove fixed energy universality of the local eigenvalue statistics of Wigner matrices. Landon, Sosoe and Yau [@character2] subsequently derived a mesoscopic CLT to show fixed energy universality of the DBM. Mesoscopic central limit theorems combined with DBM were also used in [@Bourgadegap; @logarithm; @character] to derive Gaussian fluctuations of single eigenvalues. We expect that the results derived in this paper will find similar applications for the free sum of matrices. This paper is organized as follows. In Section \[sec\_main\_results\], we introduce the model in more detail and state our main results. We also give an outline of the proof in Subsection \[subsection outline of proof\]. We collect some preliminary results, e.g., local stability of the subordination equations and local laws for the Green function, in Section \[sec\_preliminaries\]. In Sections \[sec\_proof\_main\_theorem\] and \[sec\_proof\_main\_lemma\] the main arguments of the proofs are given. In the short Section \[sec\_expectation\] we complement the results by computing the so-called bias. All our methods carry over to the orthogonal setup where one of the matrices is conjugated by Haar orthogonal matrices. The orthogonal case is analyzed separately in Section \[sec\_orthogonal\]. The proofs of some technical results used in Sections \[sec\_proof\_main\_theorem\]–\[sec\_orthogonal\] are postponed to the Appendix. We conclude this introductory section by collecting some notational conventions used throughout the paper. We use the following notion for high-probability estimates: \[definition of stochastic domination\] Let $\mathcal{X}\equiv \mathcal{X}^{(N)}$ and $\mathcal{Y}\equiv \mathcal{Y}^{(N)}$ be two sequences of nonnegative random variables. We say $\mathcal{Y}$ stochastically dominates $\mathcal{X}$ if, for all (small) $\epsilon>0$ and (large) $D>0$, $$\begin{aligned} \mathbb{P}\big(\mathcal{X}^{(N)}>N^{\epsilon} \mathcal{Y}^{(N)}\big)\le N^{-D},\end{aligned}$$ for sufficiently large $N\ge N_0(\epsilon,D)$, and we write $\mathcal{X} \prec \mathcal{Y}$ or $\mathcal{X}=O_\prec(\mathcal{Y})$. For any vector ${\bm y} \in \C^N$, denoted by bold font, we use $\|{\bm y}\|_2$ to denote the Euclidean norm. We write ${\bm g}=(g_i)_{i=1}^N \sim N_{\R}(0, \sigma^2 I_N)$ if $g_1, \cdots, g_N$ are i.i.d. centered Gaussian random variables $N(0,\sigma^2)$. In the complex case, ${\bm g} \sim N_{\C}(0, \sigma^2 I_N)$ means that ${\mathrm{Re}\,}g_i$ and ${\mathrm{Im}\,}g_i$ are i.i.d. Gaussian random variables $N(0,\frac{1}{2}\sigma^2)$. For a general random variable $\mathcal{X}$, we denote by $$\label{notation_E} \<\mathcal{X}\>:=\mathcal{X}-\E[\mathcal{X}]$$ its centering. For a matrix $X \in \C^{N \times N}$, we denote by $\|X\|_{\mathrm{op}}$ its operator norm and by $\|X\|_{\mathrm{HS}}$ its Hilbert-Schmidt norm. Moreover we use the convention $|X|^2=X^* X$. The normalized trace of $X$ is denoted by $$\label{notation_X} \ud{X}:=\frac{1}{N} \Tr X\,.$$ Finally, we use $c$ and $C$ to denote strictly positive constants that are independent of $N$. Their values may change from line to line. We write $X \ll Y$ or $X=o(Y) $ if $\frac{X}{Y} \rightarrow 0$ as $N \rightarrow \infty$. We write $X=O(Y) $ if there exists a constant $C>0$ such that $|X| \leq C |Y|$. We write $X \sim Y$ if there exist constants $c, C>0$ such that $c |Y| \leq |X| \leq C |Y|$. We denote the complex upper half-plane by $\C^+:=\{z\in\C\,:\,{\mathrm{Im}\,}z>0\}$. Main results {#sec_main_results} ============ Setup ----- Consider a sequence of random Hermitian matrices of the form $$\label{H_N} H\equiv H_N= A+U B U^*\,,$$ where $A \equiv A_N=\mathrm{diag}(a_i)$ and $B \equiv B_N=\mathrm{diag}(b_i)$ are two sequences of $N$ by $N$ deterministic real diagonal matrices, and $U \equiv U_N$ are $N$ by $N$ random unitary matrices distributed according to the Haar measure on the unitary group of order $N$, $U(N)$. Without loss of generality, by shifting with multiples of the identity matrix, we may assume that $\Tr A=\Tr B=0.$ We assume that for a constant $M$, independent of $N$, $$\label{assumption_1} \sup_{N}\|A\|_{\mathrm{op}} \leq M\,,\quad \sup_N \|B\|_{\mathrm{op}} \leq M\,.$$ The eigenvalues of $H$ are denoted by $(\lambda_i)_{i=1}^N$ in non-decreasing order. The empirical spectral measures of $A$, $B$ and $H$ are denoted by $\mu_A$, $\mu_B$ and $\mu_N$ respectively, i.e., $$\mu_A:=\frac{1}{N} \sum_{i=1}^N \delta_{a_i}; \qquad \mu_B:=\frac{1}{N} \sum_{i=1}^N \delta_{b_i}; \qquad \mu_N:=\frac{1}{N} \sum_{i=1}^N \delta_{\lambda_i}.$$ We will assume that $\mu_A$ and $\mu_B$ have weak limits as $N$ tends to infinity: \[assumption\_2\] There are deterministic compactly supported Borel probability measures $\mu_{\alpha}$ and $\mu_{\beta}$ on $\R$, neither of them being a single point mass and at least one of them being supported at more than two points, such that $\mu_A$ and $\mu_B$ converge weakly to $\mu_{\alpha}$ and $\mu_{\beta}$, respectively, as $N \rightarrow \infty$. More precisely, we assume that $$\begin{aligned} \label{levy distance assumption} d_{\mathrm{L}}(\mu_A, \mu_\alpha) +d_{\mathrm{L}}(\mu_B, \mu_\beta) \rightarrow 0, \qquad N \rightarrow \infty,\end{aligned}$$ where $d_{\mathrm{L}}$ denotes the Lévy distance. Bercovici and Voiculescu [@BeV93] showed that the free additive convolution is continuous with respect to weak convergence of measures, more specifically  implies $$\begin{aligned} \label{le difference levy distance} d_{\mathrm{L}}(\mu_A \boxplus \mu_B, \mu_\alpha \boxplus \mu_\beta) \leq C\big( d_{\mathrm{L}}(\mu_A, \mu_\alpha)+d_{\mathrm{L}}(\mu_B,\mu_\beta)\big)\,.\end{aligned}$$ The assumption that neither of $\mu_\alpha$, $\mu_\beta$ is a single point mass excludes trivial shifts by multiples of identities. The additional condition in Assumption \[assumption\_2\] that at least one of them is supported at more than two points is related to the stability of the subordination equations and the arguments of Section \[sec\_proof\_main\_lemma\]. Yet, the special case when $\mu_\alpha$ and $\mu_\beta$ are both two-point masses can be treated by combining our methods and results in Section 7 of [@stability] and Appendix B of [@eta1]. We next recall the analytic definition of the free additive convolution. For a probability measure $\mu$ on $\R$ denote by $m_\mu$ its [*Stieltjes transform*]{}, i.e. $$\begin{aligned} \label{stieltjes} m_\mu(z):=\int_\R\frac{{\mathrm{d}}\mu(x)}{x-z}\,,\qquad z\in\C^+\,.\end{aligned}$$ Note that $m_{\mu}\,:\C^+\rightarrow\C^+$ is analytic and can be analytically extended to the real line outside the support of $\mu$. Moreover, $m_{\mu}$ satisfies $$\label{m_condition} \lim_{\eta\nearrow\infty}{\mathrm{i}}\eta {m_{\mu}}({\mathrm{i}}\eta)=-1.$$ Conversely, if $m\,:\,\C^+\rightarrow\C^+$ is analytic and satisfies $\lim_{\eta\nearrow\infty}{\mathrm{i}}\eta m({\mathrm{i}}\eta)=-1$, then $m$ is the Stieltjes transform of a probability measure $\mu$, i.e., $m(z) = m_{\mu}(z)$, for all $z\in\C^+$; see e.g. [@Ak]. For notational simplicity we further introduce the negative [*reciprocal Stieltjes transform*]{} of $\mu$ by setting $$\label{stieltjes_inverse} F_{\mu}(z):=-\frac{1}{m_{\mu}(z)}.$$ Note that $F_{\mu}: \C^+ \rightarrow \C^+$ is analytic and satisfies $$\label{F_condition} \lim_{\eta\nearrow\infty} \frac{F_{\mu}({\mathrm{i}}\eta)}{{\mathrm{i}}\eta} =1.$$ The free additive convolution of two probability measures on the real line is characterized by the following result. \[subordination\_function\] Given two Borel probability measures $\mu_{\alpha}$ and $\mu_{\beta}$ on $\R$, there exist unique analytic functions, $\omega_{\alpha}, ~\omega_\beta: \C^+ \rightarrow \C^+$ such that 1. for all $z \in \C^+$, ${\mathrm{Im}\,}\omega_\alpha(z), {\mathrm{Im}\,}\omega_\beta(z) \geq {\mathrm{Im}\,}z$, and $$\label{omega_far} \lim_{\eta\nearrow\infty} \frac{\omega_{\alpha}({\mathrm{i}}\eta)}{{\mathrm{i}}\eta} =\lim_{\eta\nearrow\infty} \frac{\omega_{\beta}({\mathrm{i}}\eta)}{{\mathrm{i}}\eta}=1;$$ 2. for all $z \in \C^+$ $$\label{omega_limit} F_{\mu_\alpha}(\omega_\beta(z))=F_{\mu_\beta}(\omega_\alpha(z)); \qquad \omega_{\alpha}(z)+\omega_{\beta}(z)-z=F_{\mu_\alpha}(\omega_\beta(z)).$$ Hence, by (\[omega\_far\]) the function $${F}_{\mu_\alpha\boxplus \mu_\beta}(z):=F_{\mu_\alpha}(\omega_\beta(z))=F_{\mu_\beta}(\omega_\alpha(z))$$ satisfies (\[F\_condition\]) and thus is the negative reciprocal Stieltjes transform of a probability measure, the free additive convolution of $\mu_\alpha$ and $\mu_\beta$. The functions $\omega_\alpha$ and $\omega_\beta$ are referred to as [*subordination functions*]{}. It was shown by Belinschi [@belin_1; @belin_2] that if both $\mu_\alpha$ and $\mu_\beta$ are compactly supported probability measures on $\R$ and are supported on more than one point, then ${F}_{\mu_\alpha\boxplus \mu_\beta}$, $\omega_{\alpha}$ and $\omega_\beta: \C^+ \rightarrow \C^+$ can be extended continuously to $\R$. The singular continuous part of $\mu_{\alpha} \boxplus \mu_{\beta}$ is always zero while the absolutely continuous part is always non-zero. The corresponding density, denoted by $\rho_{\mu_\alpha \boxplus \mu_\beta}$, is real analytic whenever positive and finite. Atoms in the free additive convolution measure are identified as follows [@BeV93]. A point $c\in\R$ is an atom of $\mu_\alpha\boxplus\mu_\beta$, if and only if there exist $a,b\in\R$ such that $c=a+b$ and $\mu_\alpha(\{a\})+\mu_\beta(\{b\}) > 1$. In fact, it was shown in [@belin_3] that the density of $\mu_\alpha\boxplus\mu_\beta$ is always bounded if $\mu_\alpha(\{a\})+\mu_\beta(\{b\})<1$, for all $a,b\in\R$. Returning to the free sum of matrices, we first consider the linear eigenvalue statistics in  on the global scale. Gaussian fluctuations of the linear eigenvalue statistics for analytic test functions were derived in [@second; @order; @3] within the framework of second order freeness (we refer also to the monograph [@free_prob_book]) and in [@random_book] using resolvent based methods. Let $H_N$ be of the form (\[H\_N\]) and satisfy (\[assumption\_1\]) and (\[levy distance assumption\]). Let $g \in C(\R)$ be analytic in a neighborhood of $[-2M,2M]$, where $M$ is the constant in . Then the linear eigenvalue statistics, $$\label{linear_stat} \Tr g(H_N)-\E \Tr g(H_N),$$ converges in distribution to a centered Gaussian random variable of variance $$\label{vf_global} -\frac{1}{4 \pi^2} \int_{\mathcal{C}_2} \int_{\mathcal{C}_1} g(z_1) g(z_2) S(z_1,z_2) {\mathrm{d}}z_1 {\mathrm{d}}z_2,$$ where the kernel $S(z_1,z_2)$ is given by $$S(z_1,z_2)=\pzab \log \Big( \frac{(\omega_\alpha(z_1)-\omega_\alpha(z_2)) (\omega_\beta(z_1)-\omega_\beta(z_2)) }{(z_1-z_2)({{F}_{\mu_\alpha\boxplus \mu_\beta}(z_1)-{F}_{\mu_\alpha\boxplus \mu_\beta}(z_2)})} \Big),$$ and where $\mathcal{C}_{1,2}$ are contours enclosing $[-2M,2M]$ and are lying in the domain of analyticity of $g$. In the present paper, we prove Gaussian fluctuations for the linear eigenvalue statistics (\[observable\]) on the mesoscopic scales and establish a universal mesoscopic CLT inside the [*regular bulk*]{}: The regular bulk of $\mu_{\alpha}\boxplus \mu_{\beta}$ is the open set on which $\mu_\alpha \boxplus \mu_\beta$ has a continuous density that is strictly positive and bounded from above, i.e., $$\label{omega} \mathcal{B}_{\mu_\alpha \boxplus \mu_\beta}:=\Big\{ x \in \mbox{supp}(\rho_{\mu_\alpha \boxplus \mu_\beta})\,:\, \rho_{\mu_\alpha \boxplus \mu_\beta} (x)> 0\,; \quad \lim_{\eta \searrow 0} F_{\mu_\alpha \boxplus \mu_\beta}(x+{\mathrm{i}}\eta) \neq 0 \Big\}.$$ By the remarks after Theorem \[subordination\_function\], the regular bulk is always non-empty under Assumption \[assumption\_2\]. The convergence rate in  of Assumption \[assumption\_2\] may be very slow. Yet, by working with the finite-N deterministic measures $\mu_A \boxplus \mu_B$ instead of $\mu_\alpha \boxplus \mu_\beta$ we avoid issues related to this. Theorem \[subordination\_function\] ensures that there exist unique analytic functions, $\omega_{A}, ~\omega_B: \C^+ \rightarrow \C^+$ such that $$\label{omega_finite} F_{\mu_A}(\omega_B(z))=F_{\mu_B}(\omega_A(z)); \qquad \omega_{A}(z)+\omega_{B}(z)-z=F_{\mu_A}(\omega_B(z)),$$ and $$\label{F_definition} F_{\mu_A\boxplus\mu_B}(z):=F_{\mu_A}(\omega_B(z))=F_{\mu_B}(\omega_A(z))$$ is the negative reciprocal Stieltjes transform of the free additive convolution of $\mu_A$ and $\mu_B$. Besides the unitary conjugation in (\[H\_N\]), we also consider orthogonal conjugations, i.e., the matrix $$\label{beta} H=A+O B O^T\,,$$ where $O \equiv O_N$ is Haar distributed on the orthogonal group $O(N)$ and obtain the corresponding results. We will use the conventional parameter $\beta$ as indicator for the symmetry class; $\beta=2$ for unitary and $\beta=1$ for orthogonal conjugations. Main results {#main-results} ------------ Choose a nonempty compact interval $\mathcal I$ within the regular bulk $\mathcal{B}_{\mu_\alpha \boxplus \mu_\beta}$ (see (\[omega\])) and fix $E_0 \in \mathcal{I}$. Choose an $N$-dependent $\eta_0$ such that $ N^{-1} \ll \eta_0 \ll 1$. We then consider a mesoscopic test function $$\label{fn} f(x) \equiv f_{N}(x):=g\Big( \frac{x-E_0}{\eta_0}\Big), \qquad g \in C_c^2(\R), \qquad x \in \R.$$ Following [@character; @lytova+pastur; @random_book], we study the characteristic function $$\label{mme} \phi(\lambda):=\E[e(\lambda)], \quad \mbox{where }e(\lambda):=\exp \Big\{ \i \lambda(\Tr f(H_N)-\E \Tr f(H_N)) \Big\}, \qquad \lambda \in \R.$$ We have the following result for the characteristic function $\phi$. \[prop\] Let $H_N$ be of the form (\[H\_N\]), satisfying (\[assumption\_1\]) and Assumption \[assumption\_2\]. Let $N^{-1+c_0}\leq \eta_0\leq N^{-c_0}$, for some small $c_0>0$. Assume in addition that there is a small $c>0$, such that $ |m'_{\mu_\alpha \boxplus \mu_\beta}(E_0+ {\mathrm{i}}0)|>c$. Then there exists $0 < \tau < c_0/6$, such that the characteristic function $\phi$ satisfies $$\begin{aligned} \label{le phi prime}\phi'(\lambda)=-\lambda \phi(\lambda) V(f)+\tilde{\mathcal{E}}, \end{aligned}$$ where $$\begin{aligned} \label{vf} V(f)=-\frac{1}{2 \beta \pi^2} \int_{\Gamma_{1}} \int_{\Gamma_{2}} \tf(z_1) \tf(z_2) \mathcal K(z_1,z_2) {\mathrm{d}}z_1 {\mathrm{d}}z_2, \end{aligned}$$ and $\tilde{\mathcal{E}}$ is an error term. The integral kernel $\mathcal K$ in (\[vf\]) is given by $$\begin{aligned} \label{kernel} \mathcal K(z_1,z_2)= \pzab \log \bigg( \frac{(\omega_A(z_1)-\omega_A(z_2)) (\omega_B(z_1)-\omega_B(z_2)) }{(z_1-z_2)({F_{\mu_A\boxplus\mu_B}(z_1)-F_{\mu_A\boxplus\mu_B}(z_2)})} \bigg); \end{aligned}$$ the function $\tf$ is an almost analytic extension of $f$ given in Lemma \[helffler\] below; the contours $\Gamma_1$, $\Gamma_2$ are $\Gamma_1=\{ z_1 \in \C\,:\, |{\mathrm{Im}\,}z_1| = N^{-\tau} \eta_0 \}$ and $\Gamma_2=\{ z_2 \in \C\,:\, |{\mathrm{Im}\,}z_2| = \frac{1}{2}N^{-\tau} \eta_0 \}$ with counterclockwise orientation; and $\beta=1,2$ is the symmetry parameter. The error term $\tilde{\mathcal{E}}$ in  is bounded as $$\begin{aligned} |\tilde{\mathcal{E}}|=O_{\prec}\bigg( |\lambda| (\log N) N^{- \tau} \bigg)+O_{\prec}\bigg( \frac{(1+|\lambda|)N^{3 \tau}}{\sqrt{N \eta_0 }}\bigg)\,, \label{20011501} \end{aligned}$$ provided that $ V(f) \prec 1$. The condition $|m'_{\mu_\alpha \boxplus \mu_\beta}(E_0+ {\mathrm{i}}0)|>0$ helps us to control in Propositions \[variance\_bfbg\] and \[gbg\] some error terms effectively. It ensure that $m_{\mu_\alpha\boxplus\mu_\beta}$ is locally injective in a neighborhood of $E_0$, yet it may not be a necessary condition for the results to hold. The condition is satisfied for familiar distributions of random matrix theory such as Wigner’s semicircle law or the Marchenko-Pastur law. The expectation of $\Tr f(H_N)$ has the following asymptotic expansion for the so-called bias. \[prop2\] Under the same assumptions and notations as in Proposition \[prop\], the bias is given by $$\label{bf} \E \Tr f(H_N)-N \int_{\R} f(x) {\mathrm{d}}\mu_A\boxplus\mu_B(x)= \frac{1}{2 \pi {\mathrm{i}}} \int_{\Gamma_1} \tf(z) b(z) {\mathrm{d}}z+O(N^{-\tau})+O_{\prec}\bigg( \frac{N^{2\tau}}{\sqrt{N \eta_0}}\bigg)\,,$$ with $$\label{bz} b(z):=\frac{1}{2} \bigg( \frac{2}{\beta} -1\bigg) \frac{{\mathrm{d}}}{{\mathrm{d}}z} \log \bigg( \frac{\omega'_A(z) \omega'_B(z)}{F'_{\mu_A\boxplus\mu_B}(z)} \bigg)\,.$$ Proposition \[prop\] and \[prop2\] imply the following universal mesoscopic CLT in the regular bulk. \[meso\] Under the same assumptions as in Proposition \[prop\], for any test function $g \in C^2_c(\R)$, the mesoscopic linear statistics $$\label{linear_stat} \sum_{i=1}^N g \bigg( \frac{\lambda_i-E_0}{\eta_0} \bigg)-N \int_{\R} g \bigg(\frac{x-E_0}{\eta_0} \bigg) {\mathrm{d}}\mu_A\boxplus\mu_B(x)\,,$$ converges in distribution to a centered Gaussian random variable of variance $$\begin{aligned} \label{bulk_variance} \frac{1}{2\beta \pi^2} \int_{\R} \int_{\R} \frac{(g(x_1)-g(x_2))^2}{(x_1-x_2)^2} {\mathrm{d}}x_1 {\mathrm{d}}x_2=\frac{1}{\beta \pi} \int_{\R} |\xi| |\widehat{g}(\xi)|^2 {\mathrm{d}}\xi, \end{aligned}$$ where $\widehat{g}(\xi):=(2 \pi)^{-1/2} \int_{\R} g(x) e^{-\i \xi x} {\mathrm{d}}x$, and $\beta=1,2$ is the symmetry parameter. In particular, the bias vanishes inside the regular bulk on mesoscopic scales. Propositions \[prop\] and \[prop2\] can be extended using the Gromov-Milman concentration estimate to the regular spectral edges, where the density of the free convolution measure shows a square root behavior, under the restriction $ {N^{-2/5}}\ll \eta_0 \ll 1$. As a consequence, Theorem \[meso\] holds true on these scales but the limiting Gaussian law becomes ${N}_\R\Big(\Big(\frac{2}{\beta}-1\Big)\frac{g(0)}{4}, \frac{1}{\beta \pi} \int_{\R} |\xi| |\hat{h}(\xi)|^2 {\mathrm{d}}\xi \Big)$, where $h(x)=g(\mp x^2)$. Variance and bias agree with the expressions found for the Gaussian unitary and orthogonal ensembles [@basor+widom; @Min+Chen]; see also [@huang; @Li+Schnelli+Xu]. However, the mesoscopic scale at the regular edges ranges over ${N^{-2/3}}\ll \eta_0\ll 1$. The extension of these results to the full mesoscopic range remains an open problem. Outline of proof {#subsection outline of proof} ---------------- In this subsection we give an outline of the proof which is essentially split into two parts carried out in Sections \[sec\_proof\_main\_theorem\] and \[sec\_proof\_main\_lemma\]. Let $$\begin{aligned} \label{H_tilde_definition} H:=A+U B U^*; \qquad \mathcal{H}:= U^* H U= U^* A U+B \end{aligned}$$ and denote their resolvents or Green functions by $$\begin{aligned} \label{tilde quantities} G(z):=(H-zI)^{-1}, \qquad \tg(z):=(\mathcal{H}-zI)^{-1}, \qquad z \in \C^+.\end{aligned}$$ Note that the Stieltjes transform of the empirical spectral measure $\mu_{N}$ of $H$ is given by $$\begin{aligned} \label{m_N_definition} m_N(z) \equiv m_{\mu_N}(z)=\frac{1}{N}\sum_{i=1}^N \frac{1}{\lambda_i-z}=\frac{1}{N} \Tr G(z)=\frac{1}{N} \Tr \mathcal G(z).\end{aligned}$$ To simplify the notation, we let $$\tb:=U B U^*; \qquad \ta:=U^* A U.$$ In the first part, we study the characteristic function $\phi(\lambda)$ of the linear statistics $\Tr f(H_N)-\E \Tr f(H_N)$, see . Using the Helffer-Sjöstrand formula, Lemma \[helffler\], we link the derivative of $\phi(\lambda)$ to the resolvent of $H_N$ as follows $$\label{phimini} \phi'(\lambda)=\i \E\Big[ e(\lambda) (\Tr f(H_N)-\E \Tr f(H_N)) \Big]=\frac{\i}{\pi}\int_{\C} \frac{\partial}{\partial \overline{z_1}} \tf(z_1) \E \Big[ e(\lambda)(\Tr G(z_1)-\E \Tr G(z_1)) \Big] {\mathrm{d}}^2z_1\,,$$ where $\tf$ is a quasi-analytic continuation of $f$; see . We further use the Helffer-Sjöstrand formula to rewrite $e(\lambda)$ as $$\begin{aligned} \label{emini} e(\lambda)=\exp\Big\lbrace \frac{{\mathrm{i}}\lambda}{\pi}\int_\C\frac{\partial}{\partial \overline{z_2}}\tf(z_2) (\Tr G(z_2)-\E \Tr G(z_2)){\mathrm{d}}^2 z_2\Big\rbrace\,.\end{aligned}$$ The integration domains of the spectral parameters $z_1$ in  and $z_2$ in  are the whole complex plane. Thanks to the mesoscopic scaling in the test function $f$, recall , the contributions from local scales are negligible to the integral on the right sides of  and . More precisely, following [@character], contributions from spectral parameters with imaginary parts much smaller in absolute value than $\eta_0$ are negligible and we restrict the integration to the domains $\Omega_1\ni z_1$ defined  and $\Omega_2\ni z_2$ defined in . Moreover, we can replace $e(\lambda)$ by the regularized quantity $\ea$ of . The details are presented in Subsection \[subsection char\]. We also mention that on the domains $\Omega_1$ and $\Omega_2$ we have optimal control of the resolvent $G(z)$ and its normalized trace $m_{N}(z)$ in terms of local laws in Theorem \[local\], which will enable us to control various error terms. From  and  we are led to study the correlation $$\begin{aligned} \label{le rock} \E[\ea (\Tr G(z_1)-\E \Tr G(z_1))]\,.\end{aligned}$$ This is accomplished by using the left translation invariance of the Haar measure. Let $X=X^*$ be a deterministic $N$ by $N$ matrix and let $t\in\R$, then $\mathrm{e}^{\i t X}$ belongs to $U(N)$ and $U_t:= \mathrm{e}^{\i t X}U$ is also Haar distributed as $U$ by the translation invariance. Let $M\,:\,\C^{N \times N}\rightarrow \C$ be a differentiable map and introduce $H_t:= A+U_tBU_t^*$. Then we have that $\E [M(H_t)]$ is constant in $t$ and thence $$\begin{aligned} \label{le how to get wardmini} \frac{{\rm d}}{{\rm d} t} \bigg|_{t=0}\E [M(H_t)]=0\,.\end{aligned}$$ With different choices of functions $M$ and matrices $X$ we can generate identities among correlations functions of Green functions. In the physics literature such relations are often referred to as [*Ward identities*]{}. We can produce further Ward identities by considering the matrix $\mathcal{H}_t= U_t^*AU_t+B$ and proceed as above. We will treat the matrices $H=A+UBU^*$ and $\mathcal{H}=U^*AU+B$ in tandem, the deeper reason for this is that the subordination equations in  form a two-by-two system. Combining different Ward identities with the subordination equations, we show in Subsection \[subsection ward\] that $$\begin{aligned} \label{le rock 2} \E[\ea (\Tr G(z_1)-\E \Tr G(z_1)]&\approx\frac{ {\mathrm{i}}\lambda}{\pi} \E \bigg[ \ea \int_{ \Omega_2} \frac{\partial }{\partial \overline{z_1}} \tf(z_2) \frac{\partial }{\partial {z_2}}K(z_1,z_2) {\mathrm{d}}^2 z_2 \bigg]\,,\end{aligned}$$ up to a negligible error term. The kernel $K$ is explicitly given in . It is a linear combination of the quantities $$\begin{aligned} \label{le rock 3} K_{B,1}(z_1,z_2)&:=\frac{1}{N}\sum_{j=1}^N\frac{1}{a_j-\omega_B(z_1)}({\widetilde}BG(z_2)G(z_1))_{ii}\,,\nonumber\\ K_{B, 2}(z_1,z_2)&:=\frac{1}{N}\sum_{j=1}^N\frac{1}{a_j-\omega_B(z_1)}(G(z_2){\widetilde}B G(z_1))_{ii}\,,\end{aligned}$$ and respective counterparts involving the matrix $A$ and Green functions of the matrix $\mathcal{H}$. We remark at this point that we heavily relied on the local laws in Theorem \[local\] to control the error term in . Identity  is the main outcome of the first step in the proof; see Lemma \[lemma\_useful\]. In the second step of the proof we study the quantities in . The first term in , $K_{B,1}(z_1,z_2)$, is easy to understand thanks to the following identity for resolvents, $$\begin{aligned} \label{le rock 4} G(z_1)-G(z_2)=(z_1-z_2)G(z_1)G(z_2)\end{aligned}$$ which reduces the first term in  to a one-point function that can be well-understood by the local laws. However the second term in , $K_{B,2}(z_1,z_2)$ is harder to understand as we cannot use cyclicity and the resolvent identity  to reduce it to a one-point function. This term in fact constitutes one of the main technical difficulty of this paper. We have singled out the analysis in Section \[sec\_proof\_main\_lemma\]. Recently, similar two-point product functions have been studied for ensembles with independent entries in [@Cipolloni; @et; @all] for the Hermitization of non-Hermitian random matrices. To analyze $K_{B,2}$ it is not enough to rely only on Ward identities and local laws for the resolvent. We derive a local law for the two-point quantities $K_{B,2}(z_1,z_2)$ and $K_{A,2}(z_1,z_2)$. One strategy for that is to use the Gromov-Milman concentration inequality (see e.g. Section 4.4. of [@AGZ]) to estimate $K_{B,2}-\E K_{B,2}$. However, it turns out that $K_{B,2}$ is not sufficiently regular in $z_1$ and $z_2$ to obtain an effective estimate for all mesoscopic scales. (More precisely, for $\eta_0\gg N^{-1/2}$ this method works out.) Instead, we follow the approach of [@eta1]. It relies on a decomposition of Haar measure on the unitary groups given, e.g., in [@DiaSha; @Mezzadri]. For any fixed $1 \leq i\leq N$, any Haar unitary $U$ can be written as $$\begin{aligned} \label{le first decomposition} U=-\mathrm{e}^{\mathrm{i}\theta_i}R_i\,U^{\langle i\rangle}\,.\end{aligned}$$ Here $R_i$ is the [*Householder reflection*]{} (up to a sign) sending the vector ${\boldsymbol}{e}_i$ to ${\boldsymbol}{v}_i$, where ${\boldsymbol}{v}_i\in\C^N$ is a random vector distributed uniformly on the complex unit $(N-1)$-sphere, and $\theta_i\in [0,2\pi)$ is the argument of the $i$th coordinate of ${\boldsymbol}{v}_i$. The unitary matrix $U^{\langle i\rangle}$ has ${\boldsymbol}{e}_i$ as its $i$th column and its $(i,i)$-matrix minor is Haar distributed on $U(N-1)$. The gist of the decomposition in  is that $R_i$ and the unitary $U^{\langle i\rangle}$ are independent, for each fixed $1 \leq i\leq N$. Hence, the decomposition in  allows one to split off the partial randomness of the vector ${\boldsymbol}{v}_i$ from $U$. The analysis of $K_{B,2}$ is split into two parts: For each index $i$, we establish a concentration estimate for $(G(z_2){\widetilde}B G(z_1))_{ii}$ around the partial average $\E_{{\boldsymbol}{v}_i}[(G(z_2){\widetilde}B G(z_1))_{ii}]:=\E[(G(z_2){\widetilde}B G(z_1))_{ii}|U^{\langle i\rangle}]$. This concentration is stronger than in the conventional Gromov-Milman inequality, as we are integrating out order $N$ variables (the entries of ${\boldsymbol}{v}_i$) rather than order $N^2$ variables when taking the full expectation with respect to Haar measure. The details are given in Lemmas \[remove\] and \[concentrate\_2\]. In the second part, we identify $\E_{{\boldsymbol}{v}_i}(G(z_2){\widetilde}B G(z_1))_{ii}$. Using the decomposition  and the notation ${\widetilde}{B}^{{\langle}i{\rangle}}:= U^{{{\langle i\rangle}}}B(U^{{{\langle i\rangle}}})^*$, one works out, using concentration estimates with respect to ${\boldsymbol}{v}_i$, that $$\begin{aligned} \label{le rock 5} \E_{{\boldsymbol}{v}_i}(G(z_2){\widetilde}B G(z_1))_{ii}\approx-\E_{{\boldsymbol}{v}_i}\mathrm{e}^{{\mathrm{i}}\theta_i}{\boldsymbol}{v}_i^* {\widetilde}{B}^{{\langle}i{\rangle}}G(z_2){\widetilde}BG(z_1)\ei\,.\end{aligned}$$ We then introduce the two-point product functions $$\begin{aligned} \label{le S2t} S_i^{[2]}(z_1,z_2):=\mathrm{e}^{{\mathrm{i}}\theta_i}{\boldsymbol}{v}_i^* {\widetilde}{B}^{{\langle}i{\rangle}}G(z_2){\widetilde}BG(z_1)\ei\,, \qquad T_i^{[2]}(z_1,z_2):=\mathrm{e}^{{\mathrm{i}}\theta_i}{\boldsymbol}{v}_i^* G(z_2){\widetilde}BG(z_1)\ei\,,\end{aligned}$$ as well as the one-point functions $$\begin{aligned} \label{le S1t} S_i^{[1]}(z_1):=\mathrm{e}^{{\mathrm{i}}\theta_i}{\boldsymbol}{v}_i^* {\widetilde}{B}^{{\langle}i{\rangle}} G(z_1)\ei\,, \qquad T_i^{[1]}(z_1):=\mathrm{e}^{{\mathrm{i}}\theta_i}{\boldsymbol}{v}_i^* G(z_1)\ei\,, \end{aligned}$$ where the latter were already used in [@eta1]. (The definitions of the quantities used in Section \[sec\_proof\_main\_lemma\] are for technical reasons slightly different, but for simplicity we use here the versions above.) Next, approximating $\mathrm{e}^{-\mathrm{i}\theta_i}{\boldsymbol}{v}_i$ by a Gaussian vector and using integration by parts in $\E_{{\boldsymbol}{v}_i}S_i^{[\sharp]}$ and $\E_{{\boldsymbol}{v}_i}T_i^{[\sharp]}$, with $\sharp=1,2$, we obtain a system of equations linking $\E_{{\boldsymbol}{v}_i}S_i^{[2]}$ and $\E_{{\boldsymbol}{v}_i}T_i^{[2]}$, which can approximately be solved. This step involves local laws for the quantities $S_i^{[1]}(z)$ and $T_i^{[1]}(z)$ alongside with a further Ward identity (Lemma \[mathcal\_Y\]) that were established in [@eta1]. Interestingly, it suffices to monitor the four quantities in  and  to close that system and no higher order correlation functions appear. Once we have found an expression for $\E_{{\boldsymbol}{v}_i}S_i^{[2]}$, we can identify $(G(z_2){\widetilde}B G(z_1))_{ii}$ via . In this argument, we require the condition $m_{fc}'(E_0+{\mathrm{i}}\eta_0)\not=0$ of Theorem \[meso\] to control some error terms. The results of this analysis are summarized in Proposition \[variance\_bfbg\]. This will conclude Section 5. With Proposition \[variance\_bfbg\], we can compute the kernel $K(z_1,z_2)$ in . In Section \[sec\_proof\_main\_theorem\] we then prove Proposition \[prop\] and Theorem \[meso\] based on this result. Along the way, we require some deterministic stability estimates on the subordination functions and the Jacobian associated with the subordination equations . Those are all collected in Section \[sec\_preliminaries\]. The computation of the bias in the mesoscopic bulk is done in Section \[sec\_expectation\]. In Section \[sec\_orthogonal\], we extend the analysis to the orthogonal setting. Finally, some technical estimates, in particular related to the concentration estimates with respect the vectors $({\boldsymbol}{v}_i)$, are postponed to the Appendix. Preliminaries {#sec_preliminaries} ============= In this section, we collect some preliminary results: stability estimates and local laws for the Green function. Recall (\[stieltjes\]) and (\[stieltjes\_inverse\]). To simplify the notation, we introduce the shorthands $$m_{A}(z):=m_{\mu_A}(z), \quad m_{B}(z):=m_{\mu_B}(z), \quad F_{A}(z):=-\frac{1}{m_{\mu_A}(z)}, \quad F_{B}(z):=-\frac{1}{m_{\mu_B}(z)},$$ we abbreviate ${\mu}_{fc}:=\mu_A \boxplus \mu_B$, and denote the corresponding Stieltjes transform, negative reciprocal Stieltjes transform and density by $m_{fc}$, $F_{fc}$, and $\rho_{fc}$ respectively. They are $N$ dependent but deterministic. In addition, we write $\widetilde{\mu}_{fc}=\mu_\alpha \boxplus \mu_\beta$ and use $\widetilde{m}_{fc}$, $\widetilde{F}_{fc}$ and $\widetilde \rho_{fc}$ to denote the corresponding limiting Stieltjes transform, negative reciprocal Stieltjes transform and density, as $N \rightarrow \infty$. Properties of the subordination functions ----------------------------------------- Recall the regular bulk $\mathcal{B}_{\mu_\alpha \boxplus \mu_\beta}$ defined in (\[omega\]). We introduce a corresponding domain for the spectral parameter $z$, $$\label{bulk_domain} D_{bulk}:=\big\{ z=E +\i \eta: E \in \mathcal{I}, \; N^{-1+\epsilon} < \eta \leq 1 \big\},$$ where $\mathcal{I} \subset \mathcal{B}_{\mu_\alpha \boxplus \mu_\beta}$ is a nonempty compact interval and $\epsilon>0$ is a small constant. It was shown in [@stability; @kargin_annal_1] that $$\label{difference_m} \max_{z \in D_{bulk}}\Big( |\omega_A(z)-\omega_\alpha(z)|+ |\omega_B(z)-\omega_\beta(z)|+ |m_{fc}(z)-\widetilde m_{fc}(z)|\Big)\leq C \Big( d_{\mathrm{L}}(\mu_A, \mu_\alpha) +d_{\mathrm{L}}(\mu_B, \mu_\beta)\Big),$$ for $N$ sufficiently large, which directly implies  by (\[levy distance assumption\]). Thanks to these convergence results, the qualitative properties of $\omega_A(z)$, $\omega_B(z)$ and $m_{fc}(z)$ asymptotically agree with the limiting $\omega_\alpha(z)$, $\omega_\beta(z)$ and $\widetilde m_{fc}(z)$ respectively, and we obtain the following estimates: \[bound\_bulk\] Under Assumption \[assumption\_2\], we have the following estimates. 1. There exists $C>0$ such that $$\label{m_bound} |m_{fc}(z)| \leq C\,; \qquad |\omega_A(z)| \leq C\,; \qquad |\omega_B(z)| \leq C\,,$$ uniformly for $ z \in D_{bulk}$, for sufficient large $N$. 2. There exists $c>0$ such that $$\label{imaginary_bound} |{\mathrm{Im}\,}m_{fc}(z)| \geq c\,; \qquad |{\mathrm{Im}\,}\omega_A(z)| \geq c\,; \qquad | {\mathrm{Im}\,}\omega_B(z)| \geq c\,,$$ uniformly for $ z \in D_{bulk}$, for sufficient large $N$. 3. There exist $c, C>0$ such that $$\label{delta_bound} c \leq |1-(F'_{A}(\omega_B(z))-1)(F'_{B}(\omega_A(z))-1)|\leq C\,,$$ uniformly for $ z \in D_{bulk}$, for sufficient large $N$. 4. There exists $C>0$ such that $$\label{omega_prime_bound} |\omega'_A(z)| \leq C\,; \qquad |\omega'_B(z)| \leq C\,, \qquad |m'_{fc}(z)| \leq C\,,$$ uniformly for $ z \in D_{bulk}$, for sufficient large $N$. Note that the quantity estimated in (\[delta\_bound\]), henceforth denoted $$\label{delta} \Delta(z):=1-(F'_{A}(\omega_B(z))-1)(F'_{B}(\omega_A(z))-1)\,,$$ is the Jacobian of the following linear system obtained by differentiating the subordination equations (\[omega\_finite\]), $$\begin{aligned} \label{le system for omega primes} \begin{pmatrix} 1& 1- F'_{A}(\omega_B(z))\\ 1-F'_{B}(\omega_A(z)) & 1 \end{pmatrix} \begin{pmatrix} \omega_A'(z)\\ \omega_B'(z) \end{pmatrix} = \begin{pmatrix} 1\\ 1 \end{pmatrix}\,,\qquad z\in \C^+\,.\end{aligned}$$ Since the Jacobian $\Delta(z)$ is non-vanishing for $z \in D_{bulk}$, we hence get from  that $$\label{system} \begin{pmatrix} \omega_A'(z)\\ \omega_B'(z) \end{pmatrix} =\frac{1}{\Delta(z)} \begin{pmatrix} F'_{A}(\omega_B(z))\\ F'_{B}(\omega_A(z)) \end{pmatrix} =\frac{1}{\Delta(z) m^2_{fc}(z)} \begin{pmatrix} m'_{A}(\omega_B(z))\\ m'_{B}(\omega_A(z)) \end{pmatrix},$$ where $$\label{m_A_m_B} m'_A(\omega_B(z))=\frac{1}{N} \sum_{j=1}^N \frac{1}{(a_j-\omega_B(z))^2};\qquad m'_B(\omega_A(z))=\frac{1}{N} \sum_{j=1}^N \frac{1}{(b_j-\omega_A(z))^2}.$$ In combination with the lower bounds in (\[imaginary\_bound\]) and (\[delta\_bound\]), the linear system (\[system\]) yields that $\omega'_A(z), \omega'_B(z)$ are of constant order in the regular bulk, as stated in (\[omega\_prime\_bound\]). Furthermore, by the subordination equations , we have $$\label{self_m_fc} m_{fc}(z)=m_A(\omega_B(z))=m_B(\omega_A(z)), \qquad \omega_A+\omega_B-z=F_{fc}(z)=-\frac{1}{m_{fc}(z)}.$$ Differentiating (\[self\_m\_fc\]) with respect to $z$, we find that $$\label{differential_relation} m_{fc}'(z)=\omega'_A(z) m'_B(\omega_A(z))=\omega'_B(z) m'_A(\omega_B(z)), \qquad \omega_A'(z)+\omega_B'(z)-1=F_{fc}'(z)=\frac{m_{fc}'(z)}{m_{fc}(z)^2}.$$ The first relation in (\[differential\_relation\]) implies that $m'_{fc}(z)$ is also of constant order as are $\omega'_A(z)$ and $\omega'_B(z)$. If $m_{fc}'(z)$ is not zero, neither are the factors $\omega_A'(z)$ and $\omega_B'(z)$. Thus $$\label{LMprime} m'_A(\omega_B(z))=\frac{m_{fc}'(z)}{\omega'_B(z)};\qquad m'_B(\omega_A(z))=\frac{m_{fc}'(z)}{\omega'_A(z)}.$$ In addition, we have the following lemma whose proof is postponed to the Appendix. \[difference\] Under Assumption \[assumption\_2\], we have $$\begin{aligned} \max_{z \in D_{bulk}}\Big( |\omega'_A(z)-\omega'_\alpha(z)|+ |\omega'_B(z)-\omega'_\beta(z)|+ |m'_{fc}(z)-\widetilde m'_{fc}(z)|\Big) \leq C \Big( d_{\mathrm{L}}(\mu_A, \mu_\alpha) +d_{\mathrm{L}}(\mu_B, \mu_\beta)\Big)\,,\end{aligned}$$ for $N$ sufficiently large. Variance kernel $\mathcal{K}(z_1,z_2)$ {#another_formula} -------------------------------------- In this subsection, we define some functions in terms of the subordination functions for later purpose and then rewrite the kernel (\[kernel\]) of the variance expression (\[vf\]) of the linear statistics in a form without singularities. Generalizing (\[delta\]) and (\[m\_A\_m\_B\]), we introduce the following functions of two spectral parameters $z_1,z_2 \in \C \setminus \R$, $$\begin{aligned} L_A(z_1,z_2)&:=\frac{1}{N} \sum_{j=1}^N \frac{1}{(b_j-\omega_A(z_1))(b_j-\omega_A(z_2))}\,;\nonumber\\ L_B(z_1,z_2)&:= \frac{1}{N} \sum_{j=1}^N \frac{1}{(a_j-\omega_B(z_1))(a_j-\omega_B(z_2))}\,;\label{LM_def}\end{aligned}$$ and $$\begin{aligned} \Delta(z_1,z_2)&:=1-\Big(\frac{L_A(z_1,z_2)}{m_{fc}(z_1)m_{fc}(z_2)}-1\Big)\Big(\frac{L_B(z_1,z_2)}{m_{fc}(z_1)m_{fc}(z_2)}-1\Big)\,.\label{delta2}\end{aligned}$$ Note that if $z_1=z_2=z$, then $$\label{LM_def_zz} L_A(z):=L_A(z,z)=m'_B(\omega_A(z)), \qquad L_B(z):=L_B(z,z)=m'_A(\omega_B(z)); \qquad \Delta(z)=\Delta(z,z).$$ As an analogue of (\[differential\_relation\]), for $z_1\not= z_2$, we have from that $$\begin{aligned} m_{fc}(z_1)-m_{fc}(z_2)=\frac{1}{N} \sum_{j=1}^N \frac{\omega_B(z_1)-\omega_B(z_2)}{(a_j-\omega_B(z_1))(a_j-\omega_B(z_2))}=(\omega_B(z_1)-\omega_B(z_2) )L_B(z_1,z_2)\\ =\frac{1}{N} \sum_{j=1}^N \frac{\omega_A(z_1)-\omega_A(z_2)}{(b_j-\omega_A(z_1))(b_j-\omega_A(z_2))}=(\omega_A(z_1)-\omega_A(z_2)) L_A(z_1,z_2).\end{aligned}$$ If we choose $z_1 \neq z_2$ such that the difference $m_{fc}(z_1) - m_{fc}(z_2)$ is nonzero, then by the above relation neither are $\omega_A(z_1)-\omega_A(z_2)$ and $\omega_B(z_1) - \omega_B(z_2)$. Therefore, dividing these two factors on both sides, we have $$\label{LM} L_A(z_1,z_2)=\frac{m_{fc}(z_1)-m_{fc}(z_2)}{\omega_A(z_1)-\omega_A(z_2)}; \qquad L_B(z_1,z_2)=\frac{m_{fc}(z_1)-m_{fc}(z_2)}{\omega_B(z_1)-\omega_B(z_2)}.$$ Using (\[LM\]), we obtain the analogue of (\[system\]), i.e., $$\label{system2} \begin{pmatrix} \omega_A(z_1)-\omega_A(z_2)\\ \omega_B(z_1)-\omega_B(z_2) \end{pmatrix} =\frac{z_1-z_2}{\Delta(z_1,z_2) m_{fc}(z_1)m_{fc}(z_2)} \begin{pmatrix} L_B(z_1,z_2)\\ L_A(z_1,z_2) \end{pmatrix}.$$ Therefore, the kernel $\mathcal{K}$ in (\[kernel\]) of the variance expression can be written as $$\mathcal{K}(z_1,z_2)=\pzab \log \Big( \frac{(\omega_A(z_1)-\omega_A(z_2)) (\omega_B(z_1)-\omega_B(z_2)) }{(z_1-z_2)(F_{fc}(z_1)-F_{fc}(z_2))} \Big)=- \pzab \log (\Delta(z_1,z_2)).$$ The benefit of this form is to avoid singularities caused by $z_1=z_2$ or $m_{fc}(z_1) = m_{fc}(z_2)$, since $L_A(z_1,z_2)$, $L_B(z_1,z_2)$ as well as $\Delta(z_1,z_2)$ are well-defined functions for all $z_1,z_2 \in D_{bulk}$. Similarly, using (\[system\]) and (\[LMprime\]), we can rewrite $b(z)$ in the bias formula (\[bz\]) as $$b(z)=\frac{1}{2} \Big( \frac{2}{\beta} -1\Big) \frac{{\mathrm{d}}}{{\mathrm{d}}z} \log \Big( \frac{\omega'_A(z) \omega'_B(z)m^2_{fc}(z)}{m'_{fc}(z)}\Big)=-\frac{1}{2} \Big( \frac{2}{\beta} -1\Big) \frac{{\mathrm{d}}}{{\mathrm{d}}z} \log \Delta(z).$$ Local law for the Green function -------------------------------- We end this section by stating the local laws for the Green functions of $H$ and $\mathcal{H}$ in (\[tilde quantities\]) and . For this purpose, we introduce the deterministic control parameter $$\label{control} \Psi \equiv \Psi(z):=\frac{1}{\sqrt{N \eta}}\,,\qquad \qquad z=E+{\mathrm{i}}\eta\in\C^+\,.$$ \[local\] Under Assumption \[assumption\_2\] and (\[assumption\_1\]), the following estimates $$\begin{aligned} &\max_{i,j}\left|G_{ij}(z)-\frac{1}{a_i-\omega_B(z)} \delta_{ij} \right| \prec \Psi(z) ,\qquad & |m_{N}(z)-m_{fc}(z) | \prec \Psi^2(z),&\\ &\max_{i,j}\left|(\tb G(z))_{ij}-\frac{z-\omega_B(z)}{a_i-\omega_B(z)} \delta_{ij} \right| \prec \Psi(z),\qquad &\Big|\frac{1}{N}\mathrm{Tr}(\tb G(z))-(z-\omega_B)m_{fc}(z)\Big| \prec \Psi^2(z),&\end{aligned}$$ hold uniformly for all $z\in D_{bulk}$. Furthermore, for any deterministic and uniformly bounded $d_1, \cdots, d_N \in \C$, we have $$\label{average} \Big| \frac{1}{N} \sum_{i=1}^N d_i\Big(G_{ii}(z)-\frac{1}{a_i-\omega_B(z)} \Big) \Big| \prec \Psi^2(z)\,,$$ uniformly for all $z\in D_{bulk}$. The same estimates hold for the Green function $\mathcal{G}$ in (\[tilde quantities\]) with the roles of $A$ and $B$ interchanged. Proof of Proposition \[prop\] and Theorem \[meso\] {#sec_proof_main_theorem} ================================================== In this section, we give the proof of Proposition \[prop\] and Theorem \[meso\] for the unitary case $\beta=2$. The orthogonal case $\beta=1$ is proved similarly in Section \[sec\_orthogonal\]. Characteristic function and its derivative {#subsection char} ------------------------------------------ The idea is to study the derivative of the characteristic function of the linear eigenvalue statistics and link it with the resolvent of $H_N$ via the Helffer-Sjöstrand calculus. Let $f$ be the mesoscopic test function introduced in . Recall from  the characteristic function $$\label{mmeZG} \phi(\lambda):=\E[e(\lambda)], \quad \mbox{with }e(\lambda):=\exp \big\{ \i \lambda(\Tr f(H_N)-\E \Tr f(H_N)) \big\}\,,\quad \qquad\lambda \in \R\,.$$ The following lemma is a version of the well-known Helffer-Sjöstrand formula. \[helffler\](Helffer-Sjöstrand formula) Let $\chi(y)$ be a smooth cutoff function with support in $[-2,2]$, with $\chi(y)=1$ for $|y| \leq 1$. Define the almost-analytic extension of $f$ by $$\begin{aligned} \label{le almost analytic extension} \tilde{f}(x+{\mathrm{i}}y):=(f(x)+{\mathrm{i}}y f'(x)) \chi(y)\,. \end{aligned}$$ Then, for any $w\in\R$, $$\begin{aligned} \label{le helffer sjostrand formula} f(w)=\frac{1}{\pi} \int_{\C}\pzz \frac{ \tilde{f}(z)}{w-z} {\mathrm{d}}^2z=\frac{1}{2 \pi} \int_{\R^2} \frac{{\mathrm{i}}y f''(x) \chi(y)+{\mathrm{i}}\big( f(x)+{\mathrm{i}}y f'(x) \big) \chi'(y)}{w-x-{\mathrm{i}}y} {\mathrm{d}}x {\mathrm{d}}y\,, \end{aligned}$$ where $z=x+{\mathrm{i}}y$, $\pzz=\frac{1}{2}(\px+{\mathrm{i}}\py)$, and ${\mathrm{d}}^2z$ denotes Lebesgue measure on $\C$. Using  and the definition of the Green function in , the spectral calculus yields the following representation of the linear eigenvalue statistics, $$\label{fw} \Tr f(H_N) -\E \Tr f(H_N) =\frac{1}{\pi} \int_{\C} \pzz \tf(z) (\Tr G(z)-\E \Tr G(z)){\mathrm{d}}^2z\,.$$ Then taking the derivative of the characteristic function $\phi(\lambda)$, we obtain $$\label{phi} \phi'(\lambda)=\i \E\Big[ (\Tr f(H_N)-\E \Tr f(H_N)) e(\lambda)\Big]=\frac{\i}{\pi}\int_{\C} \pzz \tf(z) \E \Big[ e(\lambda)(\Tr G(z)-\E \Tr G(z)) \Big] {\mathrm{d}}^2z.$$ As an observation in [@character], instead of working on the $\C$, we can remove the ultra-local, or sub-mesoscopic, scales and restrict the integration domain in  to $$\label{domain} \Omega_1 := \big\{ z_1:=x_1 + \i y_1 \in \C: |y_1| \geq N^{-\tau} \eta_0 \big\},$$ with $\tau>0$ and $\eta_0$ as in Proposition \[prop\], without effecting the mesoscopic linear eigenvalue statistics. Indeed, using that $y_1 \rightarrow {\mathrm{Im}\,}m_N(z_1)y_1$ is increasing, we can extend the local law as follows: $$\label{mmp} \left| \Tr G(x_1+\i y_1)-\E \Tr G(x_1+\i y_1) \right| = O_{\prec}\Big(\frac{1}{|y_1|}\Big),$$ uniformly in $|y_1|>0$ and $x_1 \in \mathcal{I}$; see (\[bulk\_domain\]). In addition, due to (\[fn\]), there exists some $C>0$ such that $$\label{assumpf} \int_{\R} |f(x)| {\mathrm{d}}x \leq C \eta_0; \qquad \int_{\R} |f'(x)| {\mathrm{d}}x \leq C'; \qquad \int_{\R} |f''(x)| {\mathrm{d}}x \leq \frac{C''}{\eta_0},$$ and thus we have $$\label{fw2} \Tr f(H_N) -\E \Tr f(H_N)=\frac{1}{ \pi} \int_{\Omega_1} \frac{\partial }{\partial \overline{z_1}} \tf(z_1) (\Tr(G(z_1))-\E \Tr G(z_1)) {\mathrm{d}}^2 z_1+O_{\prec}(N^{-\tau}).$$ Similarly, we remove the ultra-local scales in the integral domain in the expression of $e(\lambda)$ and define $$\label{e22} \ea:=\exp\Big\{ \frac{\i \lambda}{ \pi} \int_{\Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) (\Tr(G(z_2))-\E \Tr G(z_2)) {\mathrm{d}}^2 z_2 \Big\},$$ where $$\label{domain2} \Omega_2 := \Big\{ z_2:=x_2 + \i y_2 \in \C: |y_2| \geq \frac{1}{2} N^{-\tau} \eta_0 \Big\}\,.$$ It is straightforward to check that $\ea$ approximates $e(\lambda)$ as $$\label{e2} |e(\lambda)-\ea|=O_{\prec}\big( |\lambda| N^{ -\tau} \big)\,.$$ Summarizing the above, we have the following lemma. Under the assumptions of Proposition \[prop\], we have the representation $$\label{newphi} \phi'(\lambda)=\frac{\i}{\pi}\int_{\Omega_1} \frac{\partial }{\partial \overline{z_1}} \tf(z_1) \E \Big[ \ea (\Tr(G(z_1))-\E \Tr G(z_1)) \Big] {\mathrm{d}}^2z_1+O_{\prec}\Big( |\lambda| (\log N) N^{ -\tau} \Big).$$ In view of  we are led to study $$\begin{aligned} \label{le semi expression} \E [ \ea (\Tr(G(z_1))-\E \Tr G(z_1))]\end{aligned}$$ in the next subsection. Invariance of Haar measure: Ward identities {#subsection ward} ------------------------------------------- In this subsection, we study  further. For simplicity, we recall the shorthands in (\[notation\_E\]) and (\[notation\_X\]). With these notations, we write  as $\E[\ea \< \underline{G(z_1)}\>]$. We have the following estimate for . \[lemma\_useful\] Let $z_1=E_1+{\mathrm{i}}\eta_1\in D_{bulk}$. Then $$\begin{aligned} \label{sum1bis} \E [\ea \<\Tr G(z_1)\>]=&\frac{ {\mathrm{i}}\lambda}{ \pi} \E \bigg[ \ea \int_{ \Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \frac{\partial }{\partial {z_2}}K(z_1,z_2) {\mathrm{d}}^2 z_2 \bigg]+O_{\prec}\Big( \frac{1}{N \eta_1^2 }\Big),\end{aligned}$$ where the kernel $K(z_1,z_2)$ is given by $$\label{kernel_K_1} K(z_1,z_2):= \frac{\omega_B'(z_1)}{m_{fc}(z_1)} (K_{B_1}(z_1,z_2)-K_{B_2}(z_1,z_2))+\frac{\omega_A'(z_1)}{m_{fc}(z_1)} (K_{A_1}(z_1,z_2)-K_{A_2}(z_1,z_2))\,,$$ with$$\begin{aligned} K_{B,1}(z_1,z_2)&:=\frac{1}{N} \sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} (\tb G(z_2) G(z_1))_{jj}; \quad K_{B,2}(z_1,z_2):=\frac{1}{N} \sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} (G(z_2) \tb G(z_1))_{jj};\label{K_B_formula}\\ K_{A,1}(z_1,z_2):&=\frac{1}{N} \sum_{j=1}^N \frac{1}{b_j-\omega_A(z_1)} (\ta \mathcal{G}(z_2) \mathcal{G}(z_1))_{jj}; \quad K_{A,2}(z_1,z_2):=\frac{1}{N} \sum_{j=1}^N \frac{1}{b_j-\omega_A(z_1)} (\mathcal G(z_2) \ta \mathcal G(z_1))_{jj}.\label{K_A_formula}\end{aligned}$$ The left side of  involves the full expectation with respect the Haar measure on $U(N)$. This suggest to make use of its left invariance: Let $X=X^*$ be any deterministic $N$ by $N$ matrix and let $t\in\R$, then $\mathrm{e}^{\i t X}$ belongs to $U(N)$ and $U_t:= \mathrm{e}^{\i t X}U$ is also Haar distributed as $U$ is by assumption. Let $M\,:\,\C^{N \times N}\rightarrow \C$ be a differentiable map and introduce $H_t:= A+U_tBU_t^*$. Then by the above we must have that $\E [M(H_t)]$ is constant in $t$ and hence $$\begin{aligned} \label{le how to get ward} \frac{{\rm d}}{{\rm d} t} \bigg|_{t=0}\E [M(H_t)]=0\,.\end{aligned}$$ In view of , we will first choose $$\label{function_invariant} M(H_t)=e_0(\lambda,t) \big( G_t(z_1)\big)_{ij}=\exp \Big\{ \frac{\i \lambda}{\pi} \int_{\Omega_2}\frac{\partial }{\partial \overline{z_2}} \tf(z_2) \Big( \Tr G_t(z_2)-\E \Tr G_t(z_2)\Big) {\mathrm{d}}^2 z_2 \Big\} \big(G_t(z_1)\big)_{ij},$$ where $G_t:=(H_t-zI)^{-1}$ and $e_0(\lambda,t)$ is given by with $G$ replaced by $G_t$. Using $$\label{derivative} \frac{{\rm d}}{{\rm d} t}\bigg|_{t=0} H_t= \frac{{\rm d}}{{\rm d} t}\bigg|_{t=0} \ U_t B U_t^*=\i [X, \tb],\qquad \frac{{\rm d}}{{\rm d} t}\bigg|_{t=0} G_t=-\i G [X, \tb] G,$$ we obtain from  with the choice of $M$ in  the relation $$\E\Big[ \Big( \ea \frac{\i \lambda}{\pi} \int_{\Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \sum_{l=1}^N \Big( -\i G(z_2) [X, \tb] G(z_2)\Big)_{ll} {\mathrm{d}}^2 z_2 \Big) (G(z_1))_{ij}\Big] =\E \Big[\ea \Big( \i G(z_1) [X, \tb] G(z_1) \Big)_{ij}\Big],$$ where $X$ is an arbitrary deterministic self-adjoint matrix. Let now first $X=\ei \ej^*+\ej \ei^*$ and then $X=\i \ei \ej^* -\i \ej \ei^*$. Using linearity and averaging over the index $i$, we obtain, for fixed $j$ and $z_1\in D_{bulk}$, the following identity $$\label{equation_2} \E \Big[ \ea \Big(\underline{\tb G} G_{jj}-\g (\tb G)_{jj} \Big) \Big]=I_j(z_1)\,,$$ with $$\begin{aligned} \label{le Ij} I_j(z_1):=&\frac{1}{N}\E \Big[ \ea \frac{\i \lambda}{\pi} \int_{\Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \frac{\partial }{\partial {z_2}} \Big( (\tb F G)_{jj} -(F \tb G)_{jj}\Big) {\mathrm{d}}^2 z_2 \Big]\,\end{aligned}$$ where we further introduced the short hands $$\begin{aligned} \label{le shorthands} F \equiv G(z_2)\,,\qquad G \equiv G(z_1)\,.\end{aligned}$$ We first work on the left side of . Repeating the above invariance argument with $M(H_t)=(G_t(z_1))_{ij}$, we obtain after averaging over the index $i$ the identity $$\begin{aligned} \label{le another ward} \E \Big(\underline{\tb G} G_{jj}-\g (\tb G)_{jj} \Big)=0\,,\end{aligned}$$ and we can write $$\begin{aligned} \label{le table} \E \Big[ \ea \Big(\underline{\tb G} G_{jj}-\g (\tb G)_{jj} \Big) \Big]&=\E \Big[ \ea\Big( \<\underline{\tb G} G_{jj}\>-\<\g (\tb G)_{jj} \>\Big) \Big]\,.\end{aligned}$$ Next, recalling the definition of the resolvent in , we write $(a_j-z_1)G_{jj}+({\widetilde}{B}G)_{jj}=1$, which implies $$\label{resolvent_def} \E [\ea \<\g(\tb G)_{jj}\>]=(z_1-a_j) \E [\ea \<\g G_{jj}\>]+\E [\ea \<\g\>]\,.$$ We then rewrite  as $$\begin{aligned} \label{le table 2} \E \Big[ \ea \Big(\underline{\tb G} G_{jj}-\g (\tb G)_{jj} \Big) \Big]&=\E \Big[ \ea\Big( \<\underline{\tb G} G_{jj}\>+(a_j-z_1)\<\g G_{jj} \>-\<\g\>\Big) \Big]\,.\end{aligned}$$ Next, in view of the local laws Theorem \[local\], we write the right side of  as $$\begin{aligned} \E \Big[ \ea\Big(& \<\underline{\tb G} G_{jj}\>+(a_j-z_1)\<\g G_{jj} \>-\<\g\>\Big) \Big]\nonumber\\ &=(z_1-\omega_B(z_1))m_{fc}(z_1)\E \big[ \ea \<G_{jj}\>\big]+(a_j-z_1)m_{fc}(z_1)\E\big[\ea\< G_{jj} \> \big]\nonumber\\ &\qquad+\E \Big[ \ea\Big( \<\big(\underline{\tb G}-(z_1-\omega_B(z_1))m_{fc}(z_1)\big) G_{jj}\> \Big]\nonumber\\ &\qquad+(a_j-z_1)\E \Big[ \ea\<\big(\g-m_{fc}(z_1)) G_{jj} \>\Big) \Big]-\E [\ea \<\g\>]\nonumber\\ &=(a_j-\omega_B(z_1))m_{fc}(z_1)\E \big[ \ea \<G_{jj}\>\big]-\E [\ea \<\g\>]\nonumber\\ &\qquad+\E \Big[ \ea \<\big(\underline{\tb G}-(z_1-\omega_B(z_1))m_{fc}(z_1)\big) G_{jj}\> \Big]\nonumber\\ &\qquad+(a_j-z_1)\E \Big[ \ea\<\big(\g-m_{fc}(z_1)) G_{jj} \>\Big]\,.\end{aligned}$$ Returning to , we hence obtain, after rearranging, $$\begin{aligned} &(a_j-\omega_B(z_1))m_{fc}(z_1)\E \big[ \ea \<G_{jj}\>\big]-\E [\ea \<\g\>] \nonumber\\ &\qquad\qquad=I_j(z_1)-\E \Big[ \ea \<\big(\underline{\tb G}-(z_1-\omega_B(z_1))m_{fc}(z_1)\big) G_{jj}\> \Big]\nonumber\\ &\qquad\qquad\qquad\qquad-(a_j-z_1)\E \Big[ \ea\<\big(\g-m_{fc}(z_1)) G_{jj} \> \Big]\,.\end{aligned}$$ Dividing by $(a_j-\omega_B(z_1))$ and then summing over the index $j$, the left side of the above equation vanishes by (\[self\_m\_fc\]), and we thus obtain $$\begin{aligned} \sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)}&=\E \Big[ \ea \<\big(\underline{\tb G}-(z_1-\omega_B(z_1))m_{fc}(z_1)\big)\sum_{j=1}^N\frac{1}{a_j-\omega_B(z_1)} G_{jj}\> \Big]\nonumber\\ &\qquad+\E \Big[ \ea\<\big(\g-m_{fc}(z_1))\sum_{j=1}^N\frac{a_j-z_1}{a_j-\omega_B(z_1)} G_{jj} \> \Big]\,.\end{aligned}$$ At this point, we invoke the local laws in Theorem \[local\], to get the estimate $$\begin{aligned} \sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)}&=\E \Big[ \ea \<\big(\underline{\tb G}-(z_1-\omega_B(z_1))m_{fc}(z_1)\big)\sum_{j=1}^N\frac{1}{(a_j-\omega_B(z_1))^2} \> \Big]\nonumber\\ &\qquad+\E \Big[ \ea\<\big(\g-m_{fc}(z_1))\sum_{j=1}^N\frac{a_j-z_1}{(a_j-\omega_B(z_1))^2} \>\Big]+O_\prec\big(\frac{1}{N\eta_1^2}\big)\nonumber\\ &=\frac{1}{N}\sum_{j=1}^N\frac{1}{(a_j-\omega_B(z_1))^2}\E \Big[ \ea \< \Tr({\tb G}) \> \Big]\nonumber\\ &\qquad+\frac{1}{N}\sum_{j=1}^N\frac{a_j-z_1}{(a_j-\omega_B(z_1))^2} \E \Big[ \ea\<\Tr G \> \Big]+O_\prec\big(\frac{1}{N\eta_1^2}\big)\,,\end{aligned}$$ where the second equality follows from the property that $\<\mathcal{X}\>=0$ if $\mathcal{X}$ is deterministic. Next, note that $$\begin{aligned} \frac{1}{N}\sum_{j=1}^N\frac{a_j-z_1}{(a_j-\omega_B(z_1))^2}&=\frac{1}{N}\sum_{j=1}^N\frac{a_j-\omega_B(z_1)}{(a_j-\omega_B(z_1))^2}+\frac{1}{N}\sum_{j=1}^N\frac{\omega_B(z_1)-z_1}{(a_j-\omega_B(z_1))^2}\nonumber\\ &=m_{fc}(z_1)+m_A'(\omega_B(z_1))(\omega_B(z_1)-z_1)\,.\end{aligned}$$ Hence we obtain $$\begin{aligned} \label{sum11} &\big( m_{fc}(z_1)-(z_1-\omega_B(z_1))m_A'(\omega_B(z_1)) \big)\E \big[ \ea \<\Tr G\>\big]\nonumber\\ &=\sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)}-m_A'(\omega_B(z_1))\E \Big[ \ea \< \Tr({\tb G}) \> \Big]+O_\prec\big(\frac{1}{N\eta_1^2}\big)\,.\end{aligned}$$ Next, we treat $\tg$ in (\[tilde quantities\]) similarly and obtain $$\begin{aligned} \label{sum22} &\big( m_{fc}(z_1)-(z_1-\omega_A(z_1))m_B'(\omega_A(z_1)) \big)\E \big[ \ea \<\Tr \mathcal{G}\>\big]\nonumber\\ &=\sum_{j=1}^N\frac{\mathcal{I}_j(z_1)}{b_j-\omega_A(z_1)}-m_B'(\omega_A(z_1))\E \Big[ \ea \< \Tr({\ta \mathcal{G}}) \> \Big]+O_\prec\big(\frac{1}{N\eta_1^2}\big)\,,\end{aligned}$$ where we wrote $\mathcal{F} \equiv \mathcal{G}(z_2)$, $\tg \equiv \tg(z_1)$ for short and introduced $$\begin{aligned} \label{le Ij tilde} \mathcal{ I}_j(z_1):= \frac{1}{N}\E \Big[ \ea \frac{{\mathrm{i}}\lambda}{ \pi} \int_{\Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \frac{\partial }{\partial {z_2}} \Big( (\ta \mathcal{F} \tg)_{jj} -(\mathcal{F} \ta \tg)_{jj} \Big) {\mathrm{d}}^2 z_2 \Big].\end{aligned}$$ Combining (\[sum11\]) and (\[sum22\]) with the definition of the resolvent  and the subordination equations (\[self\_m\_fc\]), we obtain $$\begin{aligned} m^3_{fc}(z_1) \Delta(z_1) \E [\ea \<\Tr G\>] =m'_B(\omega_A(z_1)) \sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)} +m'_A(\omega_B(z_1)) \sum_{j=1}^N\frac{\mathcal{I}_j(z_1)}{b_j-\omega_A(z_1)} +O_{\prec} \Big( \frac{1}{N \eta_1^2} \Big),\label{sum33}\end{aligned}$$ with $\Delta(z)$ given in (\[delta\]). Recall that we have $z_1\in D_{bulk}$, hence by (\[imaginary\_bound\]), (\[delta\_bound\]) and (\[differential\_relation\]), we can divide (\[sum33\]) by $m^3_{fc}(z_1) \Delta(z_1)$ to obtain $$\begin{aligned} \label{sum1} \E [\ea \<\Tr G\>]=&\frac{\omega_B'(z_1)}{m_{fc}(z_1)}\sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)} +\frac{\omega_A'(z_1)}{m_{fc}(z_1)} \sum_{j=1}^N\frac{\mathcal{I}_j(z_1)}{b_j-\omega_A(z_1)} +O_{\prec}\Big( \frac{1}{N \eta_1^2 }\Big)\nonumber\\ =&\frac{ {\mathrm{i}}\lambda}{ \pi} \E \bigg[ \ea \int_{ \Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \frac{\partial }{\partial {z_2}}K(z_1,z_2) {\mathrm{d}}^2 z_2 \bigg]+O_{\prec}\Big( \frac{1}{N \eta_1^2 }\Big),\end{aligned}$$ where $K(z_1,z_2)$ is given by $$\label{kernel_K_1} K(z_1,z_2):= \frac{\omega_B'(z_1)}{m_{fc}(z_1)} (K_{B,1}-K_{B,2})+\frac{\omega_A'(z_1)}{m_{fc}(z_1)} (K_{A,1}-K_{A,2}),$$ with $$K_{B,1}:=\frac{1}{N} \sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} (\tb F G)_{jj}; \quad K_{B,2}:=\frac{1}{N} \sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} (F \tb G)_{jj};$$ $$K_{A,1}:=\frac{1}{N} \sum_{j=1}^N \frac{1}{b_j-\omega_A(z_1)} (\ta \mathcal{F} \mathcal{G})_{jj}; \quad K_{A,2}:=\frac{1}{N} \sum_{j=1}^N \frac{1}{b_j-\omega_A(z_1)} (\mathcal F \ta \mathcal G)_{jj}.$$ This completes the proof of Lemma \[lemma\_useful\]. Proof of Proposition \[prop\] ----------------------------- The two terms $K_{A,1}$ and $K_{B,1}$ are easy to identify: Using the resolvent identity , we have $$\begin{aligned} K_{B,1}&=\frac{1}{z_1-z_2}\frac{1}{N}\sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} \Big( (\tb G)_{jj}(z_1)-(\tb G)_{jj}(z_2) \Big).\end{aligned}$$ Recall the local law (\[average\]) and choose $d_j(z)=\frac{1}{a_j-\omega_B(z)}$. The uniform bound of $(d_j(z))_j$ follows from (\[imaginary\_bound\]). Though $(d_j(z))_j$ depends on $z$, we can use a continuity argument to show that the local law still holds, i.e., $$\label{argument_continuity} \Big| \frac{1}{N}\sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} \Big( (\tb G(z_l))_{jj} -\frac{z_l-\omega_B(z_l)}{a_j-\omega_B(z_l)} \Big) \Big| \prec \frac{1}{N \eta_l}, \qquad l=1,2.$$ For notational simplicity, we define $$h_j(z):=(\tb G(z))_{jj} -\frac{z-\omega_B(z)}{a_j-\omega_B(z)}.$$ Next, we consider two cases to show that $$\label{argument_1} \Big|\frac{1}{z_1-z_2}\frac{1}{N}\sum_{j=1}^N \Big( d_j(z_1) (h_j(z_1)-h_j(z_2)) \Big)\Big| \prec \frac{1}{N \eta^2_1}+\frac{1}{N \eta_1 \eta_2}.$$ **Case 1:** If $z_1$ and $z_2$ are in different half-planes, then $\frac{1}{|z_1-z_2|} \leq \frac{1}{| {{\mathrm{Im}\,}}z_1 |}.$ Thus (\[argument\_1\]) follows directly from (\[argument\_continuity\]). **Case 2:** If $z_1$ and $z_2$ are in the same half-plane, without loss of generality, we can assume they both belong to the upper half plane. If $|{{\mathrm{Im}\,}}z_1-{{\mathrm{Im}\,}}z_2| \geq \frac{1}{2} {{\mathrm{Im}\,}}z_1$, then we can use the same argument as in Case 1. Thus it is sufficient to consider $|{{\mathrm{Im}\,}}z_1-{{\mathrm{Im}\,}}z_2| \leq \frac{1}{2}{{\mathrm{Im}\,}}z_1$, which means $ \frac{2}{3} {{\mathrm{Im}\,}}z_2 \leq {{\mathrm{Im}\,}}z_1 \leq 2 {{\mathrm{Im}\,}}z_2$. The left side of (\[argument\_1\]) can be bounded as $$\begin{aligned} \Big| \frac{1}{N}\sum_{j=1}^N \frac{d_j(z_1)}{z_1 -z_2} (h_j(z_1)-h_j(z_2)) \Big| & \leq \Big| \frac{1}{N}\sum_{j=1}^N \frac{d_j(z_1)-d_j(z_2)}{z_1-z_2} h_j(z_2) \Big|\nonumber\\ &\qquad\qquad +\Big| \frac{\frac{1}{N}\sum_{j=1}^N d_j(z_1) h_j(z_1)-\frac{1}{N}\sum_{j=1}^N d_i(z_2) h_j(z_2)) }{z_1-z_2} \Big|. \end{aligned}$$ The coefficients of the first term on the right side have the following upper bound $$\Big| \frac{d_j(z_1)-d_j(z_2)}{z_1-z_2} \Big| = |d_j(z_1) || d_{j}(z_2) | \Big| \frac{\omega_{B}(z_1)-\omega_{B}(z_2)}{z_1-z_2} \Big|\leq C.$$ The last step follows from the fact that $\omega_B(z)$ is analytic in the neighborhood of the segment connecting $z_1$ and $z_2$ and (\[omega\_prime\_bound\]). Using the arguments in proving (\[argument\_continuity\]), one shows from the local law (\[average\]) that the first term is bounded by $O_{\prec}\Big(\frac{1}{N \eta_2}\Big).$ The second term is bounded by $O_{\prec}\Big(\frac{1}{N \eta^2_1}+\frac{1}{N \eta_1\eta_2}\Big)$ from (\[argument\_continuity\]) using the Cauchy integral formula. Thus the error term has the same upper bound as in Case 1. Therefore, by direct computation, we obtain that $$\begin{aligned} K_{B,1}=&\frac{(z_1-\omega_B(z_1)) L_B(z_1,z_1) -(z_2-\omega_B(z_2)) L_B(z_1,z_2)}{z_1-z_2}+O_{\prec}\Big( \frac{1}{N \eta^2_1}\Big)+O_{\prec}\Big( \frac{1}{N \eta_1 \eta_2}\Big), \label{right_first_a}\end{aligned}$$ and similarly, $$\begin{aligned} K_{A,1}=&\frac{(z_1-\omega_A(z_1)) L_A(z_1,z_1) -(z_2-\omega_A(z_2)) L_A(z_1,z_2)}{z_1-z_2}+O_{\prec}\Big( \frac{1}{N \eta^2_1}\Big)+O_{\prec}\Big( \frac{1}{N \eta_1 \eta_2}\Big),\label{right_first_b}\end{aligned}$$ with $L_A(z_1,z_2)$ and $L_B(z_1,z_2)$ given in (\[LM\_def\]). Next, we estimate the rest two terms $K_{B,2}$ and $K_{A,2}$. Note that the Gromov-Milman concentration inequality (see, e.g., [@concentrate; @stability]) is not sufficiently strong to obtain the optimal mesoscopic CLT in the regular bulk. We will use the random partial decomposition used in [@eta1] to prove the following lemma in the next subsection. \[variance\_bfbg\] Under the same assumptions as in Proposition \[prop\], there exists a small neighborhood of $E_0$, denoted by $D_0$, such that for all $z_{1}=E_1+{\mathrm{i}}\eta_1, z_2=E_2+{\mathrm{i}}\eta_2 \in D_0 \cap D_{bulk}$, we have the following estimate for every $j$: $$\label{fbg_ii} (G(z_2) \tb G(z_1))_{jj} = \frac{1}{(z_1-z_2) (a_j-\omega_B(z_1))(a_j-\omega_B(z_2))} \frac{T_B(z_1,z_2)}{L_B(z_1,z_2)}+E_{B,j}(z_1,z_2),$$ for sufficiently large $N$, where $L_{B}(z_1,z_2)$ is given in (\[LM\_def\]), and $$\label{T_B} T_{B}(z_1,z_2):=(z_1-\omega_{B}(z_1))m_{fc}(z_1)- (z_2-\omega_{B}(z_2))m_{fc}(z_2),$$ and the error function $E_{B,j}(z_1,z_2)$ is analytic in $z_1, z_2 \in \C \setminus \R$ with the following estimate: $$\label{bound_notation} E_{B,j}(z_1,z_2)= O_{\prec}\Big(\frac{1}{\sqrt{N \eta_1} \eta_2} +\frac{1}{\sqrt{N \eta_2} \eta_1}+\frac{1}{N \eta_1\eta_2}+\frac{1}{N \eta_1^2}\Big).$$ The same holds true for $(\mathcal{G}(z_2) \ta \mathcal{G}(z_1))_{jj}$ with $\tg$ in (\[tilde quantities\]) by interchanging the roles of $A$ and $B$. Recalling (\[sum1bis\]) and applying Stokes’ formula, we obtain $$\begin{aligned} \E [\ea \<\Tr G(z_1)\>]=&\frac{ \lambda}{2 \pi} \E \Big[ \ea \int_{ \partial \Omega_2} \tf(z_2) \frac{\partial}{\partial z_2} K(z_1,z_2) {\mathrm{d}}z_2 \Big]+O_{\prec}\Big( \frac{1}{N \eta_1^2 }\Big)\,.\end{aligned}$$ We further apply the above equation to (\[newphi\]), using Stokes’ formula and (\[assumpf\]), we have $$\label{plug_to} \phi'(\lambda)=\frac{\lambda}{4\pi^2}\E \Big[ \ea \int_{\partial \Omega_1} \tf(z_1) \int_{ \partial \Omega_2} \tf(z_2) \frac{\partial}{\partial z_2} K(z_1,z_2) {\mathrm{d}}z_2 {\mathrm{d}}z_1\Big]+O_{\prec}\Big( \frac{N^{2\tau}}{N \eta_0 }\Big)+O_{\prec}\Big( |\lambda| (\log N) N^{ -\tau} \Big).$$ If $z_1 \in \partial \Omega_1$ and $z_2 \in \partial \Omega_2$, then by (\[fn\]), $z_1,z_2 \in D_0 \cap D_{bulk}$ (see Proposition \[variance\_bfbg\]) for $N$ sufficiently large. Combining with the fact that $|a_j-\omega_B(z)| \geq {\mathrm{Im}\,}\omega_B(z)>c>0$ by (\[imaginary\_bound\]), we hence obtain $$\label{right_second_a} K_{B,2}=\frac{1}{z_1-z_2} \frac{1}{N} \sum_{j=1}^N \frac{1}{ (a_j-\omega_B(z_1))^2(a_j-\omega_B(z_2))} \frac{T_B(z_1,z_2)}{L_B(z_1,z_2)}+E_{B}(z_1,z_2),$$ and similarly, $$\label{right_second_b} K_{A,2}=\frac{1}{z_1-z_2} \frac{1}{N} \sum_{j=1}^N \frac{1}{ (b_j-\omega_A(z_1))^2(b_j-\omega_A(z_2))} \frac{T_A(z_1,z_2)}{L_A(z_1,z_2)}+E_{A}(z_1,z_2),$$ where the error functions $E_{A}(z_1,z_2), E_{B}(z_1,z_2)$ are analytic in $z_1, z_2 \in \C \setminus \R$ with the same upper bound as in (\[bound\_notation\]). Note that by (\[LM\_def\]) and (\[LM\]), we have $$\begin{aligned} \frac{1}{N}\sum_{j=1}^N \frac{1}{(a_j-\omega_B(z_1))^2(a_j-\omega_B(z_2))}&=\frac{1}{\omega'_B(z_1)}\frac{\partial}{\partial z_1} L_B(z_1,z_2)\nonumber\\ &= \frac{m'_{fc}(z_1)}{\omega'_B(z_1)} \frac{1}{\omega_B(z_1)-\omega_B(z_2)}-\frac{m_{fc}(z_1)-m_{fc}(z_2)}{(\omega_B(z_1)-\omega_B(z_2))^2}.\end{aligned}$$ Plugging (\[right\_first\_a\]), (\[right\_first\_b\]), (\[right\_second\_a\]) and (\[right\_second\_b\]) into (\[kernel\_K\_1\]), together with (\[LM\]), (\[T\_B\]) and (\[LMprime\]), we have $$K(z_1,z_2)= \frac{\omega'_A(z_1)}{\omega_A(z_1)-\omega_A(z_2)}+\frac{\omega'_B(z_1)}{\omega_B(z_1)-\omega_B(z_2)}-\frac{1}{z_1-z_2}-\frac{m_{fc}'(z_1) m_{fc}(z_2)}{m_{fc}(z_1)(m_{fc}(z_1)-m_{fc}(z_2))}+E(z_1,z_2),$$ where the error term $E(z_1,z_2)$ has an upper bound from (\[m\_bound\]), (\[imaginary\_bound\]), (\[omega\_prime\_bound\]) and (\[bound\_notation\]): $$|E(z_1,z_2)|=O_{\prec}\Big( \frac{1}{\sqrt{N \eta_1} \eta_2} +\frac{1}{\sqrt{N \eta_2} \eta_1}+\frac{1}{N \eta^2_1}+\frac{1}{N \eta_1\eta_2} \Big).$$ Since $E(z_1,z_2)$ is analytic in $z_1, z_2 \in \C \setminus \R$, the Cauchy integral formula yields $$\begin{aligned} \label{plug} \frac{\partial }{\partial {z_2}}K(z_1,z_2)=\mathcal{K}(z_1,z_2)+\mathcal{E}(z_1,z_2),\end{aligned}$$ with the kernel function $$\begin{aligned} \mathcal K(z_1,z_2)&:= \frac{\omega'_A(z_1)\omega'_A(z_2)}{(\omega_A(z_1)-\omega_A(z_2))^2}+\frac{\omega'_B(z_1)\omega'_B(z_2)}{(\omega_B(z_1)-\omega_B(z_2))^2}-\frac{1}{(z_1-z_2)^2}-\frac{m_{fc}'(z_1) m'_{fc}(z_2)}{(m_{fc}(z_1)-m_{fc}(z_2))^2}\nonumber\\ &=\pzab \log \Big( \frac{(\omega_A(z_1)-\omega_A(z_2)) (\omega_B(z_1)-\omega_B(z_2)) }{(z_1-z_2)(\frac{1}{m_{fc}(z_2)}-\frac{1}{m_{fc}(z_1)})} \Big)=- \pzab \log (\Delta(z_1,z_2)),\end{aligned}$$ with $\Delta(z_1,z_2)$ in (\[delta2\]), and the error function $$\label{error_bound} \mathcal{E}(z_1,z_2)=O_{\prec} \Big(\frac{1}{\sqrt{N \eta_1} \eta^2_2} +\frac{1}{\sqrt{N \eta_2} \eta_1 \eta_2}+\frac{1}{N \eta^2_1\eta_2}+\frac{1}{N \eta_1\eta^2_2}\Big).$$ Plugging (\[plug\]) into (\[plug\_to\]), by (\[error\_bound\]) and (\[assumpf\]), we have $$\phi'(\lambda)=\lambda \E [ \ea] \frac{1}{4 \pi^2} \int_{\partial \Omega_1}\int_{\partial \Omega_2} \tf(z_1) \tf(z_2) \mathcal{K}(z_1,z_2) {\mathrm{d}}z_2 {\mathrm{d}}z_1+O_{\prec}\Big(\frac{N^{3 \tau}}{\sqrt{N \eta_0}} \Big)+O_{\prec}( |\lambda| N^{ -\tau}).$$ The last step is to replace $\ea$ by $e(\lambda)$ with difference (\[e2\]), provided that $V(f) \prec 1$. Thus we finish the proof of Proposition \[prop\]. Proof of Theorem \[meso\] ------------------------- We end this section by computing the explicit formula of $V(f)$ in (\[vf\]), where the test function $f$ is given in (\[fn\]). If $z_1 \in \Gamma_1$ and $z_2 \in \Gamma_1$ are in the same half plane, using the expansions of $\omega_A(z)$, $\omega_B(z)$ and $m_{fc}(z)$ in a neighborhood of $E_0$ inside the regular bulk such that $|m_{fc}'(z)|, |\omega_A'(z)|, |\omega_{B}'(z)| \sim 1$, we have $$\mathcal K(z_1,z_2) =O_{\prec}(\eta_0^{-1}).$$ Combining with (\[assumpf\]), the integral for $z_1$ and $z_2$ belonging to the same half plane only contributes $O_{\prec}(\eta_0)$. If $z_1$ and $z_2$ are in different half planes, since $E_0$ is in the regular bulk, we have $|m_{fc}(z_1)-m_{fc}(z_2)| \geq c>0$, as well as $|\omega_A(z_1)-\omega_A(z_2)|$, $|\omega_B(z_1)-\omega_B(z_2)|$. Hence $$\mathcal K(z_1,z_2) =-\frac{1}{(z_1-z_2)^2}+O_{\prec}(1).$$ Since the computation in the following is similar as the proof of Lemma 6.1 in [@Li+Schnelli+Xu], we omit it here. Therefore, we have $$V(f)=- \frac{1}{4 \pi^2} \int_{ \Gamma_1}\int_{ \Gamma_2} \tf(z_1) \tf(z_2) \mathcal K(z_1,z_2) {\mathrm{d}}z_2 {\mathrm{d}}z_1=\frac{1}{4 \pi^2} \int_{\R} \int_{\R} \frac{(g(x)-g(y))^2}{(x-y)^2} {\mathrm{d}}x {\mathrm{d}}y+O_{\prec}(\eta_0)+O(N^{-\frac{\tau}{2}}).$$ Thus $V(f)$ converges to some positive constant since $g \in C^2_c(\R)$. Theorem \[meso\] is a direct result of Proposition \[prop\] after integrating $\phi'(\lambda)$ and using the Lévy continuity theorem. Hence we finish the proof of Theorem \[meso\]. Proof of Proposition \[variance\_bfbg\] {#sec_proof_main_lemma} ======================================= In this section, we use a partial randomness decomposition to prove Proposition \[variance\_bfbg\]. We remark that this decomposition was a key ingredient in [@eta1; @eta2] to derive the local laws in Theorem \[local\]. For any Haar unitary matrix $U \equiv U_N \in U(N)$, there exists a random vector $\vi$, the $i$-th column of the matrix $U$ and an independent Haar unitary matrix $U^{i}\in U(N-1)$, such that $$U=-e^{\i \theta_i} R_i U^{\<i\>};\qquad R_i:=I-\ri \ri^*;\qquad \ri:=\sqrt{2} \frac{\ei+e^{-{\mathrm{i}}\theta_i} \vi}{\|\ei+e^{-{\mathrm{i}}\theta_i} \vi\|_2},$$ where $\theta_i$ is the argument of $i$-th entry of $\vi$, denoted by $v_{ii}$, $R_i$ is the Householder transform sending $\ei$ to $-e^{-\i \theta_i}{\bm v_i}$, and $U^{\<i\>}$ is a unitary matrix with $\ei$ as its $i$-th column and $U^{i}$ as its $(i,i)$-minor. Thus we can write $$\tb=U B U^* =R_i \tb^{\<i\>} R_i, \quad \mbox{with} \quad \tb^{\<i\>}:=U^{\<i\>} B (U^{\<i\>})^*.$$ Note that $\tbhat$ is independent of $\vi$, and define $$\label{hat_definition} H^{\<i\>}:=A+\tbhat; \qquad \ghat(z):=(A+\tbhat-z)^{-1}.$$ It is well-known that $\vi$ is a uniformly distributed unit vector in $\C^{N}$, and there exists a Gaussian vector $\widetilde \gi \sim {N}_{\C}(0,N^{-1}I_N)$, such that $$\label{uniform_vector} \vi=\frac{\widetilde \gi}{\|\widetilde \gi\|_2}.$$ Hence we can write $$\hi:=e^{-{\mathrm{i}}\theta_i}\vi=e^{-{\mathrm{i}}\theta_i} \frac{\widetilde{\gi}}{\|\widetilde \gi\|_2}; \qquad \ri:=l_i (\ei+{\bm h_i}), \qquad \mbox{with } l_i :=\frac{\sqrt{2}}{\|\ei+{\bm h_i}\|_2}=1+O_{\prec}(\frac{1}{\sqrt{N}}).$$ Note that $\hi$ is independent of $ \tb^{\<i\>}$, and $$R_i \ei=-{\bm h_i}; \qquad R_i {\bm h_i}=-\ei; \qquad {\bm h_i}^* \tb^{\<i\>} R_i=-\ei^* \tb; \qquad \ei^* \tb^{\<i\>} R_i=-b_i {\bm h_i}^*=-{\bm h_i}^* \tb .$$ Set $g_{ik}:=e^{-{\mathrm{i}}\theta_i} \widetilde{g}_{ik}$ for $k \neq i$ where $g_{ik} \sim {N}_{\C}(0,N^{-1})$ are independent Gaussian random variables, and we further introduce an independent Gaussian random variable $g_{ii} \sim {N}_{\C}(0,N^{-1})$. To simplify the proof, we use the Gaussian vector $\gi:=(g_{i1}, \cdots g_{iN})\sim {N}_{\C}\Big(0,\frac{1}{N}I_N\Big)$ to approximate $\hi$, and define $$\tb^{(i)}=W_i \tb^{\<i\>} W_i; \qquad \mbox{with }W_i:=I-{\bm w_i} {\bm w_i}^*, \qquad {\bm w_i}:=\ei+\gi,$$ and $$\label{bra_definition} H^{(i)}:=A+B^{(i)}; \qquad G^{(i)}(z):=(A+B^{(i)}-z)^{-1}.$$ Note that we have $$\label{h_approximate_r} \hi=\frac{|\widetilde g_{ii}|-g_{ii}}{\|\widetilde \gi\|_2} \ei+ \frac{1}{\|\widetilde \gi\|_2} \gi; \qquad \ri={\bm w_i}+d_{1} \ei +d_2 \gi,$$ where $$\label{bound_d} d_1:=( l_i -1)+l_i \frac{|\widetilde g_{ii}|-g_{ii}}{\|\widetilde \gi\|_2}=O_{\prec} \Big(\frac{1}{\sqrt{N}}\Big); \qquad d_2 := \frac{l_i}{\|\widetilde \gi\|_2} -1=O_{\prec}\Big( \frac{1}{\sqrt{N}}\Big).$$ Because of this, $G^{(i)}(z)$ is a good approximation of $G(z)$, see (\[removei\_1\]) below. For any $1 \leq i,j, k\leq N$ and $z=E+{\mathrm{i}}\eta \in D_{bulk}$, we have $$\label{removei_1} \Big| G_{jk}(z)-G^{(i)}_{jk}(z)\Big| , \qquad \Big| (\tb G)_{jk}-(\tbbra G^{(i)})_{jk}\Big| \prec \frac{1}{\sqrt{N \eta}}.$$ Before we proceed with the proof, we first introduce some previous results that will be used later. \[previous\_bound\] For $j \neq i$, define $$\label{define_S_one} S_i^{[1]}:=\gi^* \tb^{\<i\>}G^{(i)} \ei; \qquad S_{i,j}^{[1]}:=\gi^* \tb^{\<i\>}G^{(i)} \ej; \qquad T_i^{[1]}:=\gi^*G^{(i)} \ei; \qquad T_{i,j}^{[1]}:=\gi^* G^{(i)} \ej.$$ Then for all $z=E+{\mathrm{i}}\eta \in D_{bulk}$, we have the following estimates $$S_{i}^{[1]}=-\frac{z-\omega_B(z)}{a_i-\omega_B(z)}+O_{\prec}\Big(\frac{1}{\sqrt{N \eta}}\Big); \qquad T_i^{[1]},\quad T_{i,j}^{[1]}, \quad S_{i,j}^{[1]}, \quad \gbra_{ij} =O_{\prec}\Big(\frac{1}{\sqrt{N \eta}}\Big).$$ Hence we obtain the following local laws: $$\label{local_law_lemma} (\tbbra \gbra)_{ii}=-S_{i}^{[1]}+O_{\prec}\Big(\frac{1}{\sqrt{N}}\Big)=\frac{z-\omega_B}{a_i-\omega_B}+O_{\prec}\Big(\frac{1}{\sqrt{N\eta}}\Big); \qquad \gbra_{ii}=\frac{1}{a_i-\omega_B}+O_{\prec}\Big(\frac{1}{\sqrt{N\eta}}\Big).$$ In addition, for $\xii, \yi$ either $\gi$ or $\ei$, and $Q^{\<i\>}_1,Q_2^{\<i\>}$ either $\tbhat$ or $I$, we have an upper bound: $$\max_{i} |\xii^* Q^{\<i\>}_1 \gbra Q^{\<i\>}_2\yi | \prec 1.$$ Moreover, for all $1 \leq i,j \leq N$, we have $$\label{trivial} |\gi^* \ej|, \qquad |\gi^* \tbhat \ej|, \qquad |\gi^* \tbhat \gi| \prec \frac{1}{\sqrt{N}}.$$ In the proof of the local laws in (\[local\_law\_lemma\]), the following two lemmas were introduced in [@eta1]. Recall the shorthand notation . \[lemma\_concentrate\] For all $z=E+{\mathrm{i}}\eta \in D_{bulk}$ and $1 \leq i \leq N$, we have $$\label{concentrate_1} |\IE S_i^{[1]}|,\quad |\IE T_i^{[1]}|,\quad |\IE G_{ii}^{(i)}| \prec \frac{1}{\sqrt{N \eta}},$$ where we use the notation that for any general random variable $\mathcal{X}$, $$\IE \mathcal{X}:=\mathcal{X}-\E_{\gi} \mathcal{X},$$ denoting by $\E_{\gi}$ the partial expectation with respect to the Gaussian vector $\bm{g_i}$. For all $z=E+{\mathrm{i}}\eta \in D_{bulk}$ and $1 \leq i \leq N$, we have $$\label{finite_rank_1} |\ud{\tb G}-\ud{\tbhat \gbra}|, \quad |\ud{\tbhat \gbra \tbhat}-\ud{\tb G \tb}| \prec \frac{1}{N \eta}.$$ Furthermore, we have the following upper bounds $$\label{trivial_4} |\ud{\tbhat \gbra}|, \qquad |\ud{\tbhat \gbra \tbhat}| \prec 1.$$ Moreover, we have $$\label{concentrate_4} |\IE \ud{\tbhat \gbra}| ,\quad |\IE \ud{\tbhat \gbra \tbhat}| \prec \frac{1}{N \eta}.$$ The strategy in [@eta1] to prove the local laws (\[local\_law\_lemma\]) is to use Gaussian integration by parts in combination with Lemma \[lemma\_concentrate\] to find a pair of equations for $S_i^{[1]}$ and $T_i^{[1]}$, and thus obtain the following estimates: $$\label{bound_S_1} S_i^{[1]}=-\frac{z-\omega_B(z)}{a_i-\omega_B(z)}+O_{\prec}(\Psi); \qquad T_i^{[1]}=O_{\prec}(\Psi).$$ In this section, we extend this technique to deal with the quantity in (\[K\_B\_formula\]), $$K_{B,2}(z_1,z_2):=\frac{1}{N} \sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)} (G(z_2) \tb G(z_1))_{jj} .$$ It suffices to find a local law for the two point function $(G(z_2) \tb G(z_1))_{jj}$, as well as $(\tb G(z_2) \tb G(z_1))_{jj}$. For simplicity, recall the shorthands (\[le shorthands\]), $$F \equiv G(z_2)\,,\qquad G \equiv G(z_1)\,.$$ In addition, for $z_1=E_1+{\mathrm{i}}\eta_1$, $z_2=E_2+{\mathrm{i}}\eta_2$, we define two control parameters $$\Xi_1\equiv \Xi_1(z_1,z_2):= \frac{1}{\sqrt{N \eta_1} \eta_2} +\frac{1}{\sqrt{N \eta_2} \eta_1}; \qquad \Xi_2 \equiv \Xi_2(z_1,z_2):=\frac{1}{N \eta_1^2}+\frac{1}{N \eta_1\eta_2}.$$ The following lemma, whose proof is given in the Appendix, ensures that one can replace $F, G$ by $\fbra$ and $\gbra$ respectively with affordable price. \[remove\] For all $1 \leq i,j \leq N$ and $z_1=E_1+{\mathrm{i}}\eta_1, z_2=E_2+{\mathrm{i}}\eta_2 \in D_{bulk}$, we have $$\label{removei} \Big| (F \tb G)_{jj}-( F^{(i)}\tb^{(i)} G^{(i)})_{jj}\Big| , \quad \Big| (\tb F \tb G)_{jj}-(\tbbra F^{(i)}\tb^{(i)} G^{(i)})_{jj}\Big| \prec \Xi_1.$$ With Lemma \[remove\], we can reduce the problem to study the approximation $$\widehat{K}_{B,2}(z_1,z_2):= \frac{1}{N} \sum_{i=1}^N \frac{1}{a_i-\omega_B(z_1)} (\fbra \tbbra \gbra)_{ii}.$$ Coming in pair with $(\fbra \tbbra \gbra)_{ii}$, we will also study $(\tbbra \fbra \tbbra \gbra)_{ii}$. Note that $$\begin{aligned} \label{approximate_qq} \Big( \tb^{(i)} F^{(i)}\tb^{(i)} G^{(i)}\Big)_{ii}=&\ei^* (1-\ei^*\ei-\gi^*\ei-\ei^*\gi-\gi^*\gi) \tb^{\<i\>} (1-\wi^*\wi) F^{(i)} \tbbra G^{(i)}\ei \nonumber\\ =&-\gi^* \tb^{\<i\>}F^{(i)}\tb^{(i)} G^{(i)} \ei +O_{\prec}\Big( \frac{1}{\sqrt{N \eta_1 \eta_2}} \Big).\end{aligned}$$ The last step follows from Lemma \[previous\_bound\] and from the Cauchy-Schwarz inequality, i.e., $$\label{trivial_1} \wi^* \fbra \tbbra \gbra \ei \leq \|\tbbra\|_{\mathrm{op}} \|\wi^* \fbra \|_2 \|\gbra \ei\|_2 =O_{\prec}\Big( \frac{1}{\sqrt{\eta_1 \eta_2}}\Big).$$ Combining with Lemma \[remove\], we have $$\label{approximate} (\tb F \tb G)_{ii}=(\tbbra \fbra \tbbra \gbra)_{ii}+O_{\prec}(\Xi_1):=-S_i^{[2]}+O_{\prec}(\Xi_1),$$ where we define the following two point functions for simplicity $$\label{define_S_two} S_i^{[2]}:=\gi^* \tb^{\<i\>}F^{(i)}\tb^{(i)} G^{(i)} \ei, \qquad T_i^{[2]}:=\gi^* F^{(i)}\tbbra G^{(i)} \ei.$$ It is hence enough to look at $S^{[2]}_i$. Using Lemma \[previous\_bound\], it is straightforward to check the crude bound $$\label{trivial_2} |S_i^{[2]}|, \quad |T_i^{[2]}| \prec \frac{1}{\sqrt{\eta_1\eta_2}}.$$ As an analogue of Lemma \[lemma\_concentrate\], we have the following concentration results for the two point functions $S^{[2]}_i$, $T_i^{[2]}$ and $(\fbra \tbbra \gbra )_{ii}$. \[concentrate\_2\] The following hold uniformly for $z_1,z_2 \in D_{bulk}$: $$|\IE S_i^{[2]}| \prec \Xi_1; \qquad |\IE T_i^{[2]}| \prec \Xi_1; \qquad |\IE (\fbra \tbbra \gbra )_{ii}| \prec \Xi_1.$$ The proof can be found in the Appendix. Together with these concentration results, we use Gaussian integration by parts to find a pair of linear equations of $\E_\gi S_i^{[2]}$ and $\E_\gi T_i^{[2]}$, and then solve for $\E_\gi S_i^{[2]}$. Note that if $g \sim {N}_\C(0,\sigma^2)$, since $g$ and $\bar{g}$ are independent, then we have the formula of integration by parts, $$\label{integration_by_parts_complex} \int_{\C} \bar{g} f(g,\bar{g}) e^{-\frac{|g|^2}{\sigma^2}} {\mathrm{d}}g \wedge {\mathrm{d}}\bar{g}=\sigma^2 \int_{\C} \partial_{g} f(g,\bar{g}) e^{-\frac{|g|^2}{\sigma^2}} {\mathrm{d}}g \wedge {\mathrm{d}}\bar{g}.$$ By direct computation and $$\label{derivative_R_complex} \frac{\partial W_i}{\partial g_{ik} }= -\ek (\ei+\gi)^*,$$ we obtain that $$\begin{aligned} \label{counterpart} \E _{\gi} S^{[2]}_i=&\E_{\gi}\Big[ \underline{\tbhat \fbra} \Big( S_i^{[2]} -b_i T_i^{[2]} +(b_i \gi^* \ei+\gi^* \tbhat \gi) ( (\fbra \tbbra \gbra)_{ii}+T^{[2]}_i) \Big)\nonumber \\ &+\underline{\tbhat \fbra W_i \tbhat} \Big( (\fbra \tbbra \gbra)_{ii} +T_i^{[2]} \Big)\nonumber\\ &+(\underline{\tbhat \fbra \tbbra \gbra} - \underline{\tbhat \fbra}) \Big( S_i^{[1]} -b_i T_i^{[1]} +(b_i \gi^* \ei+\gi^* \tbhat \gi) ((\gbra)_{ii} +T_i^{[1]}) \Big) \nonumber\\ & +(\underline{\tbhat \fbra \tbbra W_i \gbra \tbhat}-\underline{\tbhat \fbra W_i \tbhat}) \Big( (\gbra)_{ii} +T_i^{[1]} \Big)\Big].\end{aligned}$$ Note that $\|\fbra\|_{\mathrm{op}} \leq \frac{1}{\eta_2}$, which implies that $$|\ud{\tbhat \fbra W_i \tbhat} -\ud{\tbhat \fbra \tbhat} |=\frac{1}{N} | \wi^* (\tbhat)^2 \fbra \wi| \leq \frac{1}{N} \|\tbhat\|^2_{\mathrm{op}} \|\fbra\|_{\mathrm{op}} \|\wi\|_2\prec \frac{1}{N \eta_2};$$ and similarly $$|\ud{\tbhat \fbra \tbbra \gbra W_i \tbhat} -\ud{\tbhat \fbra \tbbra \gbra \tbhat} |=\frac{1}{N} |\wi^* (\tbhat)^2 \fbra \tbbra \gbra \wi| \prec \frac{1}{N \eta_1 \eta_2}.$$ In addition, we have the following lemma, with proof provided in the Appendix. \[finite\_rank\_2\] The following hold uniformly for $z_{1}, z_2 \in D_{bulk}$: $$\label{finite_rank_3} |\ud{\tbhat \fbra \tbbra \gbra}-\underline{\tb F \tb G}|, \quad |\ud{\tbhat \fbra \tbbra \gbra \tbhat}-\underline{\tb F \tb G \tb}| \prec \Xi_2.$$ Furthermore, we have $$\label{trivial_bounds} |\ud{\tbhat \fbra \tbbra \gbra}|, \quad |\ud{\tbhat \fbra \tbbra \gbra \tbhat} |\prec \frac{1}{\sqrt{\eta_1\eta_2}}.$$ In addition, we have $$\label{concentrate_5} | \IE \underline{\tbhat \fbra \tbbra \gbra}|, \quad | \IE \underline{\tbhat \fbra \tbbra \gbra \tbhat}| \prec \Xi_2.$$ Combining with Lemma \[previous\_bound\], (\[trivial\_1\]), (\[trivial\_2\]), (\[trivial\_4\]) and (\[trivial\_bounds\]), we obtain that $$\E _{\gi} S^{[2]}_i=\E_{\gi}\Big[ \underline{\tbhat \fbra} \Big( S_i^{[2]} -b_i T_i^{[2]} \Big) +\underline{\tbhat \fbra \tbhat} \Big( (\fbra \tbbra \gbra)_{ii} +T_i^{[2]} \Big)$$ $$-\Big( \underline{\tbhat \fbra \tbbra \gbra} - \underline{\tbhat \fbra} \Big) (\tbbra \gbra)_{ii}$$ $$+\Big(\underline{\tbhat \fbra \tbbra \gbra \tbhat} -\underline{\tbhat \fbra \tbhat} \Big) (\gbra)_{ii} \Big]+O_{\prec}(\Xi_1+\Xi_2).$$ Using the concentration results (\[concentrate\_4\]) and (\[concentrate\_5\]), together with approximation results (\[removei\]), (\[removei\_1\]), (\[finite\_rank\_1\]) and (\[finite\_rank\_3\]), we have $$\begin{aligned} \label{use_later_S} \E _{\gi} S^{[2]}_i =&\underline{\tb F} \E_{\gi} \Big( S_i^{[2]} -b_i T_i^{[2]} \Big) +\underline{\tb F \tb} \E_{\gi} \Big( (F \tb G)_{ii} +T_i^{[2]}\Big)\nonumber\\ &+(\underline{\tb F \tb G \tb} -\underline{\tb F \tb}) \E_{\gi} G_{ii} - (\underline{\tb F \tb G}- \underline{\tb F}) \E_{\gi} (\tb G)_{ii}+O_{\prec}(\Xi_1+\Xi_2).\end{aligned}$$ We use integration by parts on $T_i^{[2]}$ and obtain similarly that $$\begin{aligned} \label{use_later_T} \E _{\gi} T^{[2]}_i =&\underline{ F} \E_{\gi} \Big( S_i^{[2]} -b_i T_i^{[2]} \Big) +\underline{ F \tb} \E_{\gi} \Big( (F \tb G)_{ii} +T_i^{[2]}\Big) \nonumber\\ &+(\underline{ F \tb G \tb} -\underline{F \tb} ) \E_{\gi} G_{ii} - (\underline{ F \tb G}- \underline{ F}) \E_{\gi} (\tb G)_{ii}+O_{\prec}(\Xi_1+\Xi_2).\end{aligned}$$ Combining them together, we have $$\begin{aligned} \underline{F} \E_{\gi} S^{[2]}_i =& -\underline{\tb F} \E_{\gi}(F \tb G)_{ii} +\mathcal{Y} \E_{\gi}\Big( (F \tb G)_{ii}+T_i^{[2]}\Big)\nonumber\\ &+\Big( \underline{ F} (\underline{\tb F \tb G \tb} -\underline{\tb F \tb} )- \underline{\tb F} (\underline{ F \tb G \tb} -\underline{F \tb}) \Big) \E_{\gi} G_{ii} \nonumber\\ & -\Big( \ud{F} (\underline{\tb F \tb G}- \underline{\tb F}) - \ud{\tb F}(\underline{ F \tb G}- \underline{ F})\Big) \E_{\gi} (\tb G)_{ii}+O_{\prec}(\Xi_1+\Xi_2),\label{eq_FS}\end{aligned}$$ where $\mathcal{Y}:=\ud{\tb F}-(\ud{\tb F})^2+\ud{\tb F \tb} \cdot \ud{F}$. We recall the following result: \[mathcal\_Y\] Under the assumptions in Proposition \[prop\], the estimate $$\ud{\tb G(z)}-(\ud{\tb G(z)})^2+\ud{\tb G(z) \tb} =O_{\prec}\Big( \frac{1}{N \eta}\Big),$$ holds uniformly for $z=E+{\mathrm{i}}\eta \in D_{bulk}$. The smallness of $\mathcal{Y}$ in Lemma \[mathcal\_Y\] is a consequence of the optimal concentration estimates established in [@eta2] and the fact that $\E[\mathcal{Y}]=0$. The latter can be seen from Ward identities similar to Subsection \[subsection ward\]. Summing over $i$, using the concentration results Lemma \[lemma\_concentrate\], \[concentrate\_2\] and (\[approximate\]), we have $$\ud{F} \cdot \underline{\tb F \tb G} = \underline{\tb F} \cdot \underline{F \tb G} - \underline{ F} \cdot \Big(\underline{\tb F \tb G \tb} \cdot \underline{ G} -\underline{\tb F \tb} \cdot \underline{ G} -(\underline{\tb F \tb G}- \underline{\tb F}) \underline{\tb G} \Big)$$ $$+\underline{\tb F} \Big(\underline{ F \tb G \tb} \cdot \underline{ G} -\underline{F \tb} \cdot \underline{ G} - (\underline{ F \tb G}- \underline{ F}) \underline{\tb G} \Big)+O_{\prec}(\Xi_1+\Xi_2) .$$ We can solve for $\underline{\tb F \tb G \tb}$ and $$\underline{\tb F \tb G \tb}=\Big( -\frac{1}{\g} +\frac{\underline{\tb G}}{\underline{G}} + \frac{\underline{\tb F}}{\underline{F}}\Big) \underline{\tb F \tb G} +\frac{\underline{\tb F}}{\underline{F}\cdot \underline{G}} \Big( 1-\underline{\tb G}\Big) \underline{F \tb G} +\underline{\tb F \tb}-\frac{(\underline{\tb F})^2}{\underline{F}}+O_{\prec}(\Xi_1+\Xi_2).$$ Returning to (\[eq\_FS\]), we obtain $$\E_{\gi} S_i^{[{2}]}=-\frac{\underline{\tb F}}{\underline{F}}\E_{\gi} (F \tb G)_{ii}+\Big( -\frac{ \underline{\tb F \tb G}}{\g} +\frac{\underline{\tb F} \cdot \underline{F \tb G}}{ \underline{F}\cdot \g}\Big) \E_{\gi} G_{ii}+O_{\prec}(\Xi_1+\Xi_2).$$ The resolvent definition (\[tilde quantities\]) and (\[approximate\]) imply that $$S_i^{[2]}=-(\tb F \tb G)_{ii}+O_{\prec}(\Xi_1)=(a_i-z_2) (F \tb G)_{ii} -(\tb G)_{ii} +O_{\prec}(\Xi_1).$$ Therefore, we obtain that $$(a_i-z_2+\frac{\underline{\tb F}}{\underline{F}}) \E_{\gi}(F \tb G)_{ii} =\Big( -\frac{ \underline{\tb F \tb G}}{\g} +\frac{\underline{\tb F} \cdot \underline{F \tb G}}{ \underline{F}\cdot \g}\Big) \E_{\gi} G_{ii} +\E_{\gi}(\tb G)_{ii}+O_{\prec}(\Xi_1+\Xi_2).$$ Dividing $a_i-z_2+\frac{\underline{\tb F}}{\underline{F}} \approx a_i-\omega_B(z_2)$ which is away from zero, combining with the local law Theorem \[local\], we have $$\begin{aligned} \E_{\gi} (F \tb G)_{ii} =&\Big( -\frac{ \underline{\tb F \tb G}}{\g} +\frac{\underline{\tb F} \cdot \underline{F \tb G}}{ \underline{F}\cdot \g}\Big) \frac{\E_{\gi}G_{ii}}{a_i-\omega_B(z_2)} +\frac{\E_{\gi}(\tb G)_{ii}}{a_i-\omega_B(z_2)}+O_{\prec}(\Xi_1+\Xi_2)\nonumber\\ =&\frac{z_1-\omega_B(z_1)}{(a_j-\omega_B(z_1))(a_j-\omega_B(z_2))}-\frac{1}{m_{fc}(z_1)(a_j-\omega_B(z_1)) (a_j-\omega_B(z_2))} \Big( \ud{\tb F \tb G} -(z_2-\omega_B(z_2)) \ud{F \tb G} \Big)\label{partial_exp_fbg}\\ &+O_{\prec}(\Xi_1+\Xi_2).\nonumber\end{aligned}$$ Armed with Lemma \[concentrate\_2\], it suffices to estimate $\ud{\tb F \tb G}$ and $\ud{F \tb G}$. The second one is easy to find. It follows from the resolvent identity (\[le rock 4\]), the local law Theorem \[local\] and the arguments in proving (\[argument\_1\]), i.e., $$\label{fbg} \underline{F \tb G}= \frac{1}{z_1-z_2} \Big( \underline{\tb G} -\underline{\tb F} \Big) = \frac{T_B}{z_1-z_2}+O_{\prec}(\Xi_2),$$ with $T_B \equiv T_B(z_1,z_2)$ given in (\[T\_B\]). Next, we will use (\[partial\_exp\_fbg\]) to solve for $\ud{\tb F \tb G}$. Averaging over $i$ and using the concentration results Lemma \[concentrate\_2\], we obtain that $$\label{ieq} L_B \underline{\tb F \tb G} = (z_1-\omega_B(z_1))m_{fc}(z_1)L_B +(z_2-\omega_B(z_2)) L_B \ud{F \tb G}-m_{fc}(z_1) \ud{F \tb G}+O_{\prec}(\Xi_1+\Xi_2),$$ where $L_B \equiv L_B(z_1,z_2)$ is given in (\[LM\_def\]). Since $|\widetilde m'_{fc}(E_0+ {\mathrm{i}}0)| \geq c>0$, due to Lemma \[difference\], there exists a neighborhood of $E_0$, denoted by $D_0$, such that for all $z \in D_0$, $|m'_{fc}(z)| \geq c>0$ for large $N$. Hence $m_{fc}(z)$ is locally injective in such neighborhood, so are $\omega_A(z)$ and $\omega_B(z)$. Thus, $L_B(z_1,z_2) \neq 0$, and (\[ieq\]) implies that $$\label{bfbg} \underline{\tb F \tb G} = (z_1-\omega_B(z_1))m_{fc}(z_1) +(z_2-\omega_B(z_2)) \frac{T_B}{z_1-z_2}-\frac{m_{fc}(z_1)}{z_1-z_2} \frac{T_B}{L_B}+\frac{1}{|L_B|}O_{\prec}(\Xi_1+\Xi_2).$$ Note that if $z_1$ and $z_2$ are in different half planes, $\frac{1}{L_B(z_1,z_2)} =\frac{\omega_B(z_1)-\omega_B(z_2)}{m_{fc}(z_1)-m_{fc}(z_2)} \sim 1$ because of (\[imaginary\_bound\]) and (\[m\_bound\]). If $z_1$ and $z_2$ are in the same half plane, then $\omega_B(z)$ is analytic in a neighborhood encircling $z_1,z_2$. Due to (\[imaginary\_bound\]) and (\[m\_bound\]), we have $$|L_B(z_1,z_2)-L_B(z_1,z_1)|=\Big|\sum_{j=1}^N \frac{1}{a_j-\omega_B(z_1)}\Big( \frac{1}{a_j-\omega_B(z_1)}-\frac{1}{a_j-\omega_B(z_2)} \Big) \Big|=O_{\prec}(|z_1-z_2|).$$ In addition, by (\[LMprime\]), (\[omega\_prime\_bound\]) and the fact that $|m'_{fc}(E_0+{\mathrm{i}}\eta_0)| > c>0$, we have $|L_B(z_1,z_1)|=\Big| \frac{m'_{fc}(z_1)}{\omega'_B(z_1)} \Big| \geq c'>0$. Hence $|L_B(z_1,z_2)| \sim 1$ for sufficiently large $N$. Plugging (\[bfbg\]) and (\[fbg\]) into (\[partial\_exp\_fbg\]), we obtain that $$(F \tb G)_{jj} = \frac{1}{(z_1-z_2) (a_j-\omega_B(z_1))(a_j-\omega_B(z_2))} \frac{T_B}{L_B}+O_{\prec}(\Xi_1+\Xi_2).$$ Hence we complete the proof of Proposition \[variance\_bfbg\]. Expectation of the linear statistics {#sec_expectation} ==================================== In this section, we compute the expectation of the linear statistics $\Tr f(H_N)$ and prove that the bias for mesoscopic linear statistics in the regular bulk vanishes. Via the Helffer-Sjöstrand Calculus, we have $$\label{expectation_fw2} \E \Tr f(H_N) -N \int_{\R} f(x) {\mathrm{d}}\mu_{fc}(x)=\frac{1}{ \pi} \int_{\Omega_1} \pzz \tf(z) \Big( \E \Tr G(z) -N m_{fc}(z) \Big) {\mathrm{d}}^2z+O_{\prec}(N^{-\tau}).$$ We then reduce the bias on the left side to $\E \Tr G(z)-N m_{fc}(z)$. The translation-invariance of Haar measure yields $$\label{expectation_invariant} \E [\underline{\tb G} G_{ii}]=\E [\underline{G} (\tb G)_{ii}].$$ Then we have $$\E [\underline{\tb G}] \E[ G_{ii}] +\E \<\underline{\tb G} \> \<G_{ii}\>=\E [\underline{G}] \E[ (\tb G)_{ii}]+\E \< \underline{G}\> \< (\tb G)_{ii}\>,$$ where the local law Theorem \[local\] implies that $$\E [(\tb G)_{ii}] =\frac{\E [\underline{\tb G}]}{\E [\underline{G}]} \E[ G_{ii}] +O_{\prec}(\Psi^3).$$ Combining with the definition of the resolvent (\[tilde quantities\]), we have $$\Big(z-a_j -\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}\Big) \E G_{ii}=-1+O_{\prec}(\Psi^3).$$ Since $z-a_j -\frac{\E [\underline{\tb G}]}{\E [\underline{G}]} \approx \omega_B(z)-a_j$ is alway from zero by (\[imaginary\_bound\]), we write $$\E G_{ii}=\frac{1}{a_i-z +\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}}+O_{\prec}(\Psi^3).$$ Note that Theorem \[local\] yields $$\label{expansion_error} \frac{\E [\underline{\tb G}]}{\E [\underline{G}]}=(z-\omega_B(z))+O_{\prec}(\Psi^2).$$ Hence, we obtain the following expansion: $$\label{expansion_term} \E G_{ii}=\frac{1}{a_i-\omega_B(z)} -\frac{1}{(a_i-\omega_B(z))^2} \Big( \frac{\E [\underline{\tb G}]}{\E [\underline{G}]}- z+\omega_B(z)\Big)+O_{\prec}(\Psi^3).$$ Summing over the index $i$, we have $$\E \Tr G(z)=N m_{fc}(z)-N m'_A(\omega_B(z)) \Big( \frac{\E [\underline{\tb G}]}{\E [\underline{G}]}- z+\omega_B(z)\Big)+O_{\prec}\Big(\frac{1}{\sqrt{ N\eta^3}}\Big).$$ Multiplying $\E[ \underline{G}]$ on both sides, by direct computation, we have $$\begin{aligned} \Big( (z-\omega_B(z))m'_A(\omega_B(z)) -m_{fc}(z)\Big) \Big(\E \Tr G-&N m_{fc}(z) \Big)=m'_A(\omega_B(z)) \E [\Tr \tb G]\\ &-m'_A(\omega_B(z)) (z-\omega_B(z)) N m_{fc}(z)+O_{\prec}\Big(\frac{1}{\sqrt{ N\eta^3}}\Big).\end{aligned}$$ We treat $\mathcal{G}$ in (\[tilde quantities\]) similarly and obtain corresponding relation by interchanging $A$ with $B$. Combining them together, using the resolvent definition (\[tilde quantities\]) and subordination equations (\[self\_m\_fc\]), we obtain that $$m^3_{fc}(z) \Delta(z) \Big(\E \Tr G-N m_{fc}(z) \Big) =O_{\prec}\Big(\frac{1}{\sqrt{ N\eta^3}}\Big),$$ with $\Delta(z)$ in (\[delta\]). Therefore, by (\[delta\_bound\]) and (\[imaginary\_bound\]), we have $$\E \Tr G(z) -N m_{fc}(z)=O_{\prec}\Big(\frac{1}{\sqrt{ N\eta^3}}\Big).$$ Plugging this into (\[expectation\_fw2\]), using Stokes’ formula and (\[assumpf\]), we obtain $$\begin{aligned} \E \Tr f(H_N) -N \int_{\R} f(x) {\mathrm{d}}\mu_{fc}(x)&=\frac{1}{2 \pi \i} \int_{\partial \Omega_1} \tf(z) \Big( \E \Tr G(z) -N m_{fc}(z) \Big) {\mathrm{d}}z+O_{\prec}(N^{-\tau})\\ &=O_{\prec}\Big( \frac{N^{2 \tau}}{\sqrt{N \eta_0} }\Big)+O_{\prec}(N^{-\tau})\,.\end{aligned}$$ Thus, in the bulk, the bias vanishes on mesoscopic scales. This concludes the proof of Proposition \[prop2\]. Orthogonal case {#sec_orthogonal} =============== In this section, we prove that Theorem \[meso\] holds true for the orthogonal conjugation $H_N=A+O B O^{T}$, where $O \in O(N)$ with Haar measure ($\beta=1$). Proof of Proposition \[prop\] for $\beta=1$ ------------------------------------------- Let $X^T=-X \in \R^{N \times N}$ be deterministic and define $O_t:=e^{t X} O$, $H_{t}:=A+O_t^T B O_{t}$, and $G_t=(H_t-zI)^{-1}$ for $t \in \R$. The function in (\[function\_invariant\]) has constant expectation in $t$ by the translation-invariance of the Haar orthogonal measure. Choose $X=\ei \ej^*-\ej \ei^*$ and average over $i$, since $G$ is symmetric, we then obtain an analogue of (\[equation\_2\]), i.e., $$\begin{aligned} \E \Big[ \ea \Big(\g (\tb G)_{jj}-\frac{1}{N} (G \tb G)_{jj}-\underline{\tb G} G_{jj} + \frac{1}{N} (G^2 \tb)_{jj}\Big) \Big] +2 I_j(z)=0,\end{aligned}$$ with $I_j$ given in (\[le Ij\]). Proceeding to the arguments in Section \[sec\_proof\_main\_theorem\] in combination with the local law Theorem \[local\], as an analogue of (\[sum11\]), we obtain $$\begin{aligned} & \Big( (z_1-\omega_B(z_1))m'_A(\omega_B(z_1)) -m_{fc}(z_1)\Big) \E [\ea \<\Tr G\>]=m'_A(\omega_B(z_1)) \E [\ea \<\Tr{\tb G}\>] -2 \sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)} \nonumber\\ & \qquad + \E \Big[ \ea \Big\< \frac{1}{N} \sum_{j=1}^N \frac{1}{a_j -\omega_B(z_1)} (G \tb G)_{jj}\Big\> \Big]-\E \Big[ \ea \Big\< \frac{1}{N} \sum_{j=1}^N \frac{1}{a_j -\omega_B(z_1)} (G^2 \tb)_{jj}\Big\> \Big] +O_{\prec}\Big(\frac{1}{N \eta_1^2} \Big).\label{sum_orth_1}\end{aligned}$$ Next, we will show that the last line of (\[sum\_orth\_1\]) is negligible. For the second term on the last line, using the local law (\[average\]) and Cauchy integral formula, we have $$\E \Big[ \ea \Big\< \frac{1}{N} \sum_{j=1}^N \frac{1}{a_j -\omega_B(z_1)} (G^2 \tb)_{jj}\Big\> \Big] =\E \Big[ \ea \Big\< \frac{1}{N} \sum_{j=1}^N \frac{1}{a_j -\omega_B(z_1)} \frac{{\mathrm{d}}}{{\mathrm{d}}z_1} (G \tb)_{jj}\Big\> \Big] =O_\prec \Big(\frac{1}{N \eta_1^2}\Big).$$ For the first term in the last line of (\[sum\_orth\_1\]), we need the following analogue of Proposition \[variance\_bfbg\] for the orthogonal case \[gbg\] Under the same assumptions and notations from Proposition \[prop\], Proposition \[variance\_bfbg\] holds true for $\beta=1$. In particular, if $z_1=z_2=z$, we have $$\label{gbg_eq} (G \tb G)_{jj} = \frac{ [(z-\omega_B)m_{fc}(z)]' }{(a_j-\omega_B(z))^2} \frac{1}{L_B(z)}+O_{\prec}(\Xi_1+\Xi_2),$$ for sufficiently large $N$, with $L_B(z)$ given in (\[LM\_def\_zz\]). The proof of Proposition \[gbg\] is provided in the next subsection. Therefore, $$\E \Big[ \ea \Big\< \frac{1}{N} \sum_{j=1}^N \frac{1}{a_j -\omega_B} (G \tb G)_{jj}\Big\> \Big] =O_{\prec}(\Xi_1+\Xi_2).$$ Returning to (\[sum\_orth\_1\]), we have $$\begin{aligned} \label{Io} &\Big( (z_1-\omega_B(z_1))m'_A(\omega_B(z_1)) -m_{fc}(z_1)\Big) \E [\ea \<\Tr G\>]\nonumber\\ &=m'_A(\omega_B(z_1)) \E [\ea \<\Tr{\tb G}\>]-2 \sum_{j=1}^N\frac{I_j(z_1)}{a_j-\omega_B(z_1)} +O_{\prec}(\Xi_1+\Xi_2).\end{aligned}$$ We treat similarly for $\tg$ in (\[tilde quantities\]) and obtain corresponding relation by interchanging $A$ with $B$. Therefore, as an analogue of (\[sum1\]), we obtain that $$\E [\ea \<\Tr G\>]=\frac{ {\mathrm{i}}2 \lambda}{\pi} \E \bigg[ \ea \int_{\Omega_2} \frac{\partial }{\partial \overline{z_2}} \tf(z_2) \frac{\partial }{\partial {z_2}}K(z_1,z_2) {\mathrm{d}}^2 z_2 \bigg]+O_{\prec}(\Xi_1+\Xi_2),$$ with $K(z_1,z_2)$ given in (\[kernel\_K\_1\]). Following the arguments in Section \[sec\_proof\_main\_theorem\], we complete the proof of Proposition \[prop\] for $\beta=1$. Proof of Proposition \[gbg\] ---------------------------- In this subsection, we will use partial randomness decomposition to prove Proposition \[gbg\]. For any Haar orthogonal matrix $O \equiv O_N \in O(N)$, there exists a random vector $\vi$ uniformly distributed on the unit sphere in $\R^{N}$ and an independent Haar orthogonal matrix $O^{i}\in O(N-1)$, such that $$O=-\mbox{sgn}(v_{ii}) R_i O^{\<i\>};\qquad R_i:=I-\ri \ri^*;\qquad \ri:=\sqrt{2} \frac{\ei+\mbox{sgn}(v_{ii}) \vi}{\|\ei+\mbox{sgn}(v_{ii}) \vi\|_2},$$ where $R_i$ is the Householder transform sending $\ei$ to $-\mbox{sgn}(v_{ii}) {\bm v_i}$, and $O^{\<i\>}$ is an orthogonal matrix with $\ei$ as its $i$-th column and $O^{i}$ as its $(i,i)$-minor. Thus we can write $$\tb=O B O^* =R_i \tb^{\<i\>} R_i, \quad \mbox{with} \qquad \tb^{\<i\>}:=O^{\<i\>} B (O^{\<i\>})^*.$$ Note that $\tbhat$ is independent of $\vi$, and define $$H^{\<i\>}:=A+\tbhat; \qquad \ghat:=(A+\tbhat-z)^{-1}.$$ Similarly as (\[uniform\_vector\]), there exists $\widetilde \gi \sim {N}_{\R}(0,N^{-1}I_N)$, so that $$\hi:=\mbox{sgn}(v_{ii})\vi=\mbox{sgn}(v_{ii}) \frac{\widetilde{\gi}}{\|\widetilde \gi\|_2}; \qquad l_i :=\frac{\sqrt{2}}{\|\ei+{\bm h_i}\|_2}=1+O_{\prec}(\frac{1}{\sqrt{N}}); \qquad \ri:=l_i (\ei+{\bm h_i}),$$ We define $g_{ik}:=\mbox{sgn}(v_{ii}) g_{ik} \sim {N}_{\R}(0,N^{-1})$ for $k \neq i$ that are independent and further introduce an independent random variable $g_{ii} \sim {N}_{\R}(0,N^{-1})$. To simplify the proof, we use the Gaussian vector $\gi \sim {N}_{\R}(0, N^{-1}I)$ to approximate $\hi$, and define $$\tb^{(i)}=W_i \tb^{\<i\>} W_i; \qquad W_i:=I-{\bm w_i} {\bm w_i}^*; \qquad {\bm w_i}:=\ei+\gi; \qquad G^{(i)}:=(A+B^{(i)}-z)^{-1}.$$ Most results in the unitary case still hold true for orthogonal setup, like Lemma \[remove\]-\[finite\_rank\_2\]. As the analogue of (\[integration\_by\_parts\_complex\]) and (\[derivative\_R\_complex\]), we have $$\label{integration_by_parts_real} \int_{\R} g f(g) e^{-\frac{g^2}{\sigma^2}} {\mathrm{d}}g =\sigma^2 \int_{\R} \partial_{g} f'(g) e^{-\frac{g^2}{\sigma^2}} {\mathrm{d}}g;$$ $$\label{derivative_R_complex} \frac{\partial W_i}{\partial g_{ik} }= -\ek (\ei+\gi)^*-(\ei+\gi) \ek^*.$$ Recall the definitions of $S_i^{[2]}, T_i^{[2]}, S_i^{[1]}, T_i^{[1]}$ in (\[define\_S\_two\]) and (\[define\_S\_one\]). By direct computation, as the counterpart (\[counterpart\]) in unitary case, we obtain that $$\begin{aligned} \E _{\gi} S^{[2]}_i=&\E_{\gi}\Big[ \underline{\tbhat \fbra} \Big( S_i^{[2]} -b_i T_i^{[2]} \Big) +\underline{\tbhat \fbra W_i \tbhat } \Big( (\fbra \tbbra \gbra)_{ii} +T_i^{[2]} \Big)\\ &+(\underline{\tbhat \fbra \tbbra \gbra}- \underline{\tbhat \fbra}) \Big( S_i^{[1]} -b_i T_i^{[1]} \Big) +(\underline{\tbhat \fbra \tbbra \gbra \tbhat}-\underline{\tbhat \fbra \tbhat}) \Big( G^{(i)}_{ii} +T_i^{[1]} \Big) \Big]\\ &+\frac{1}{N} \E_{\gi} \Big[ (\ei+\gi)^* \fbra (\tbhat)^2 W_i \fbra \tbbra \gbra \ei + (\ei+\gi)^* \tbhat W_i \fbra \tb^{\<i\>} \fbra \tbbra \gbra \ei\Big]\\ &+\frac{1}{N} \E_{\gi} \Big[ (\ei+\gi)^* \gbra\tbbra \fbra (\tbhat)^2 W_i \gbra \ei+ (\ei+\gi)^* \tbhat W_i \gbra \tbbra \fbra \tbhat \gbra \ei \Big] \\ &-\frac{1}{N} \E_{\gi} \Big[ (\ei+\gi)^* \fbra (\tbhat)^2 W_i \gbra \ei+ (\ei+\gi)^* \tbhat W_i \fbra \tb^{\<i\>} \gbra \ei \Big]+O_{\prec}(\Xi_1+\Xi_2).\end{aligned}$$ Recalling Lemma A.1 in [@eta1], we have the following rough estimates: $$\xii^* \fbra (\tbhat)^2 \fbra \tbbra \gbra \ei,\quad \xii^* \tbhat \fbra \tb^{\<i\>} \fbra \tbbra \gbra \ei,\quad \xii^* \fbra (\tbhat)^2 \gbra \ei =O_\prec( \Xi_2);$$ $$\xii^* \tbhat \fbra \tb^{\<i\>} \gbra \ei,\quad \xii^* \gbra\tbbra \fbra (\tbhat)^2 \gbra \ei,\quad \xii^* \tbhat \gbra \tbbra \fbra \tbhat \gbra \ei =O_{\prec}( \Xi_2),$$ where $\xii$ stands for either $\ei$ or $\gi$. Therefore, we obtain the same results, e.g., (\[use\_later\_S\]) and (\[use\_later\_T\]), as in the unitary case with additional affordable error $\Xi_2$. Using the arguments in the proof of Proposition \[variance\_bfbg\], one extends Proposition \[variance\_bfbg\] for $\beta=1$. In particular, if $z_1=z_2=z$, we have $$\begin{aligned} & (G \tb G)_{ii} =\frac{z-\omega_B(z)}{(a_j-\omega_B(z))^2}-\frac{1}{m_{fc}(z)(a_j-\omega_B(z))^2} \Big( \ud{\tb G \tb G} -(z-\omega_B(z)) \ud{G \tb G} \Big)+O_{\prec}(\Xi_1+\Xi_2)\label{partial_exp_fbg_real}.\end{aligned}$$ Averaging over $i$ and by (\[LM\_def\_zz\]), we obtain that $$L_B(z) \underline{\tb G \tb G} = (z-\omega_B(z))m_{fc}(z) L_B(z) +(z-\omega_B(z)) L_B(z) \ud{G \tb G}-m_{fc}(z) \ud{G \tb G}+O_{\prec}(\Xi_1+\Xi_2).$$ By (\[LMprime\]), (\[omega\_prime\_bound\]) and $\widetilde m'_{fc}(E_0+\i \eta_0) \neq 0$, there exists some $c>0$ such that $|L_B(z) |= \Big|\frac{m_{fc}'(z)}{\omega_B'(z)} \Big| >c$ for large $N$. Hence, dividing $L_B(z)$ on both sides, we have $$\underline{\tb G \tb G} = (z-\omega_B)m_{fc}+(z-\omega_B) \ud{G \tb G}-\frac{m_{fc}(z) }{L_B(z)}\ud{G \tb G} +O_{\prec}(\Xi_1+\Xi_2).$$ Note that the local law Theorem \[local\] implies that $$\ud{G \tb G}=\ud{\tb G^2}=\frac{{\mathrm{d}}}{{\mathrm{d}}z} \ud{\tb G}=[(z-\omega_B)m_{fc}(z)]'+O_{\prec}(\Xi_2).$$ Combining them together, we complete the proof of Proposition \[gbg\]. Expectation of the linear statistics and bias for $\beta=1$ ----------------------------------------------------------- Theorem \[meso\] follows directly from Proposition \[prop\]. The last step is to show that the bias in the regular bulk vanishes. As the analogue of (\[expectation\_invariant\]), we have $$\label{ortho} \E \underline{G} (\tb G)_{ii}-\frac{1}{N} \E (G \tb G)_{ii}= \E \underline{\tb G} G_{ii} -\frac{1}{N} \E (G^2 \tb)_{ii}.$$ Combining with the local law Theorem \[local\] and the definition of resolvent, we have $$\Big(z-a_i -\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}\Big) \E G_{ii}=-1-\frac{1}{N} \frac{1}{\E \underline{G}} \Big( \E (G^2 \tb)_{ii}- \E (G \tb G)_{ii} \Big)+O_{\prec}(\Psi^3).$$ Dividing $z-a_i -\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}$ on both sides, we obtain that $$\E G_{ii}=\frac{1}{a_i-z +\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}}+\frac{1}{N} \frac{1}{\E \underline{G}} \Big(\E ( G^2 \tb)_{ii}- \E (G \tb G)_{ii} \Big)\frac{1}{a_j-z +\frac{\E [\underline{\tb G}]}{\E [\underline{G}]}}+O_{\prec}(\Psi^3).$$ Using (\[expansion\_error\]), we have the following expansion: $$\E G_{ii}=\frac{1}{a_i-\omega_B(z)} -\frac{1}{(a_i-\omega_B(z))^2} \Big( \frac{\E [\underline{\tb G}]}{\E [\underline{G}]}- z+\omega_B(z)\Big)+\frac{1}{N} \frac{1}{\E \underline{G} (a_i-\omega_B(z))} \Big( \E ( G^2 \tb)_{ii}-\E (G \tb G)_{ii} \Big)+O_{\prec}(\Psi^3).$$ Summing over the index $i$, by direct computation, we have $$\begin{aligned} \label{eq1_ortho} \Big( (z-\omega_B(z))m'_A(\omega_B(z)) -m_{fc}(z)\Big)& \Big(\E \Tr G-N m_{fc}(z) \Big)=m'_A(\omega_B(z)) \E [\Tr \tb G]\nonumber\\ &-m'_A(\omega_B(z)) (z-\omega_B(z)) N m_{fc}-S_B(z)+O_{\prec}(N \Psi^3)\,,\end{aligned}$$ where $$\label{I_s} S_B(z):=\frac{1}{N} \sum_{j=1}^N \frac{1}{(a_j-\omega_B)} \Big(\E ( G^2 \tb)_{ii}- \E (G \tb G)_{ii} \Big)=:S_{B,1}+S_{B,2}.$$ We treat $\mathcal{G}$ in (\[tilde quantities\]) similarly and obtain that $$\begin{aligned} \label{eq2_ortho} &\Big( (z-\omega_A(z))m'_B(\omega_A(z)) -m_{fc}(z)\Big) \Big(\E \Tr \mathcal{G}-N m_{fc}(z) \Big) \nonumber\\ &\qquad=m'_B(\omega_A(z)) \E [\Tr \ta \mathcal{G}]-m'_B(\omega_A(z)) (z-\omega_A(z)) N m_{fc}(z) -S_A(z)+O_{\prec}(N \Psi^3),\end{aligned}$$ where $$\label{J_s} S_A(z):=\frac{1}{N} \sum_{j=1}^N \frac{1}{(b_j-\omega_A)} \Big(\E ( \mathcal G^2 \ta)_{ii}- \E (\mathcal G \ta \mathcal G)_{ii} \Big)=:S_{A,1}+S_{A,2}.$$ Therefore, combining (\[eq1\_ortho\]) with (\[eq2\_ortho\]), we obtain that $$m^3_{fc}(z)\Delta(z) \Big(\E \Tr G-N m_{fc}(z) \Big) =m'_B(\omega_A(z)) S_B(z)+m'_A(\omega_B(z)) S_A(z)+O_{\prec}(N \Psi^3),$$ with $\Delta(z)$ in (\[delta\]). Dividing $m^3_{fc}(z) \Delta(z)$ on both sides and by (\[delta\_bound\]) (\[imaginary\_bound\]) and (\[self\_m\_fc\]), we obtain $$\begin{aligned} \label{linear_statistics_formula} \E \Tr G(z) -N m_{fc}(z)=\frac{\omega_B'(z)}{m_{fc}(z)}S_B(z)+\frac{\omega_A'(z)}{m_{fc}(z)}S_A(z)+O_{\prec}(N \Psi^3).\end{aligned}$$ The first term of $S_B$ is easy to estimate using the local law and Cauchy integral formula, i.e., $$\begin{aligned} S_{B,1}=\frac{1}{N}\sum_{j=1}^N \frac{1}{a_j-\omega_B(z)} \frac{{\mathrm{d}}}{{\mathrm{d}}z} (G \tb)_{jj} =(1-\omega'_B(z) ) L_B(z)+ \frac{1}{N}\sum_{j=1}^N \frac{(z-\omega_B(z))\omega_B'(z)}{(a_j -\omega_B(z))^3} +O_{\prec}\Big(\frac{1}{N \eta^2}\Big),\label{ggb_average} \end{aligned}$$ with $L_B(z)$ given in (\[LM\_def\_zz\]). By Proposition \[gbg\], we have $$S_{B,2}= \frac{ [(z-\omega_B)m_{fc}(z)]' }{L_B(z)} \frac{1}{N} \sum_{j=1}^N \frac{1}{(a_j-\omega_B(z))^3} .$$ Note that by differentiating (\[LM\_def\_zz\]) with respect to $z$, we obtain that $$\frac{1}{N} \sum_{j=1}^N \frac{1}{(a_j-\omega_B(z))^3}=\frac{L'_B(z)}{2 \omega'_B(z)}.$$ Combining with (\[ggb\_average\]), we have $$S_B(z) =(1-\omega'_B(z)) L_B(z)+\frac{1}{2}(z-\omega_B(z)) L_B'(z)-\frac{[(z-\omega_B(z))m_{fc}(z)]' L_B'(z)}{ 2 \omega_B'(z)L_B(z)}+O_{\prec}\Big( \frac{1}{\sqrt{N \eta^3}} +\frac{1}{N \eta^3} \Big).$$ Similarly, we obtain the corresponding estimate for $S_A$ by interchanging $B$ with $A$. Therefore, by direct computation from (\[linear\_statistics\_formula\]) and (\[differential\_relation\]), we have $$\E \Tr G(z) -N m_{fc}(z)=b(z)+O_{\prec}\Big( \frac{1}{\sqrt{N \eta^3}} +\frac{1}{N \eta^3} \Big),$$ where $b(z)$ is given by $$\label{easy_b} b(z):=-\frac{1-\omega_B'(z)}{2} \frac{1}{L_B(z)} L'_B(z)-\frac{1-\omega_A'(z)}{2} \frac{1}{L_A(z)} L'_A(z)+\frac{m_{fc}'(z)}{m_{fc}(z)}-\frac{m_{fc}'^2(z)}{m_{fc}^3(z)}.$$ By (\[LMprime\]) and (\[differential\_relation\]), we can rewrite the $b(z)$ as $$\begin{aligned} b(z)&=-\frac{(1-\omega_B'(z)) \omega'_B(z)}{2 m_{fc}'(z)} \Big(\frac{m_{fc}'(z)}{\omega_B'(z)}\Big)' -\frac{(1-\omega_A'(z)) \omega'_A(z)}{2 m_{fc}'(z)} \Big(\frac{m_{fc}'(z)}{\omega_A'(z)}\Big)' +\frac{m_{fc}'(z)}{m_{fc}(z)}-\frac{m_{fc}'^2(z)}{m_{fc}^3(z)}\\ &=\frac{1}{2} \Big(\frac{\omega''_A(z)}{\omega'_A(z)} + \frac{\omega''_B(z)}{\omega'_B(z)} -\frac{m''_{fc}(z)}{m'_{fc}(z)}+\frac{2m'_{fc}(z)}{m_{fc}(z)}\Big)=\frac{1}{2} \frac{{\mathrm{d}}}{{\mathrm{d}}z} \log\Big( \frac{\omega'_A(z) \omega'_B(z)}{F'_{fc}(z)}\Big).\end{aligned}$$ Using Lemma \[bound\_bulk\] and (\[easy\_b\]), if $z \in D_{bulk}$, we have $|L_B'(z)|=O_{\prec}(1)$ and thus $|b(z)|=O_{\prec}(1)$. Plugging it into (\[expectation\_fw2\]), using the Stokes’ formula and (\[assumpf\]), we have $$\E \Tr f(H_N) -N \int_{\R} f(x) d \mu_{fc}(x)=\frac{1}{ \pi} \int_{\Omega_1} \pzz \tf(z) b(z) {\mathrm{d}}^2z+O_{\prec}(N^{-\tau})+O_{\prec}\Big( \frac{1}{\sqrt{N \eta_0}} +\frac{1}{N \eta_0^2} \Big)$$ $$=O_{\prec}\Big( \frac{1}{\sqrt{N \eta}} +\frac{1}{N \eta_0} \Big)+O_{\prec}(\eta_0)+O_{\prec}(N^{-\tau}).$$ Therefore, inside the regular bulk, the bias vanishes on mesoscopic scales, hence we prove Proposition \[prop2\] and Theorem \[meso\] for $\beta=1$. In this Appendix, we provide proofs for some auxiliary results used in this paper. We start by proving Lemma \[difference\]. We will first look at $|\omega'_B(z)-\omega'_\beta(z)|$. By replacing $F'_{A}(z), F'_{B}(z), \omega_A(z), \omega_B(z), m_{fc}(z)$ by $F'_{\alpha}(z), F'_{\beta}(z), \omega_\alpha(z), \omega_\beta(z), \widetilde{m}_{fc}(z)$ respectively, one defines analogously $\tilde{\Delta}(z)$ as in (\[delta\]), and (\[self\_m\_fc\]), (\[system\]) also holds true. Thus we have $$\begin{aligned} \label{temp_eq} |\omega_B'(z)-\omega_\beta'(z)|&=\Big|\frac{m'_{B}(\omega_A(z))}{ \Delta(z) m_{fc}^2(z)}-\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) \widetilde m_{fc}^2(z)}\Big| \nonumber\\ &\leq \Big|\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) \widetilde m_{fc}^2(z)}-\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) m_{fc}^2(z)}\Big|+\Big|\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) m_{fc}^2(z)}-\frac{m'_{\beta}(\omega_\alpha(z))}{ \Delta(z) m_{fc}^2(z)}\Big|\nonumber\\ &\qquad+\Big|\frac{m'_{\beta}(\omega_\alpha(z))}{ \Delta(z) m_{fc}^2(z)}-\frac{m'_{\beta}(\omega_A(z))}{ \Delta(z) m_{fc}^2(z)}\Big|+\Big|\frac{m'_{\beta}(\omega_A(z))}{ \Delta(z) m_{fc}^2(z)}-\frac{m'_{B}(\omega_A(z))}{ \Delta(z) m_{fc}^2(z)}\Big|.\end{aligned}$$ Note that $|\widetilde m_{fc}(z)|, |\widetilde \Delta(z)| \sim 1, |m'_{\beta}(\omega_\alpha(z))| =O(1)$ for all $z\in D_{bulk}$, and the same bounds $|m_{fc}(z)|$, $|\Delta(z)| \sim 1$, $|m'_{B}(\omega_A(z))|, |m'_{\beta}(\omega_A(z))|=O(1)$ hold for large $N$, because of (\[difference\_m\]). Thus in combination of (\[difference\_m\]), the first term on the right side can be bounded by $$\Big|\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) \widetilde m_{fc}^2(z)}-\frac{m'_{\beta}(\omega_\alpha(z))}{ \widetilde \Delta(z) m_{fc}^2(z)}\Big| \leq C|m_{fc}(z)-\widetilde m_{fc}(z)| \leq C'(d_{\mathrm{L}}(\mu_{B}, \mu_{\beta})+d_{\mathrm{L}}(\mu_{A}, \mu_{\alpha})).$$ The third term on the right side of (\[temp\_eq\]) can be treated similarly together with $ |{\mathrm{Im}\,}\omega_A(z)|, |{\mathrm{Im}\,}\omega_\alpha(z)|>c>0$ by (\[imaginary\_bound\]). In addition, using $ |{\mathrm{Im}\,}\omega_A(z)|>c>0$ by (\[imaginary\_bound\]) and the property of Lévy distance, one checks that $$\max_{z\in D_{bulk}}\{ |m'_{\beta}(\omega_A(z))-m'_{B}(\omega_A(z))| \}\leq C d_{\mathrm{L}}(\mu_{B}, \mu_{\beta}).$$ We hence obtain the estimate of the last term of (\[temp\_eq\]). Finally, we estimate the second term of (\[temp\_eq\]). Note that $$|\Delta(z) -\widetilde{\Delta}(z)|\leq \Big| \frac{m_A'(\omega_B(z))m_B'(\omega_A(z))}{m^4_{fc}(z)}-\frac{m_\alpha'(\omega_\beta(z))m_\beta'(\omega_\alpha(z))}{\widetilde m^4_{fc}(z)} \Big|$$ $$+\Big|\frac{m_A'(\omega_B(z))}{m^2_{fc}(z)}-\frac{m_\alpha'(\omega_\beta(z))}{\widetilde m^2_{fc}(z)}\Big|+\Big|\frac{m_B'(\omega_A(z))}{m^2_{fc}(z)}-\frac{m_\beta'(\omega_\alpha(z))}{\widetilde m^2_{fc}(z)}\Big|.$$ These terms can be treated similarly as above using (\[difference\_m\]). Hence we obtain the estimate of $|\omega_B'(z)-\omega_\beta'(z)|$. The same holds for $|\omega'_A(z)-\omega'_\alpha(z)|$. Furthermore, by (\[differential\_relation\]) we have $$|m'_{fc}(z)-\widetilde m'_{fc}(z)|=|\omega'_B(z)m_A'(\omega_B(z))- \omega'_\beta(z) m'_{\alpha}(\omega_\beta(z))| \leq |\omega'_\beta(z) m'_{\alpha}(\omega_\beta(z))-\omega'_\beta(z) m'_{\alpha}(\omega_B(z))|$$ $$+|\omega'_\beta(z) m'_{\alpha}(\omega_B(z))-\omega'_\beta(z) m'_{A}(\omega_B(z)) |+|\omega'_\beta(z) m'_{A}(\omega_B(z))-\omega'_B(z) m'_{A}(\omega_B(z))|.$$ These terms can also be treated similarly using (\[difference\_m\]) in combination with the estimates on $|\omega_B'(z)-\omega_\beta'(z)|$ and $|\omega_A'(z)-\omega_\alpha'(z)|$. Thus we complete the proof of Lemma \[difference\]. Next, we will check Lemma \[remove\]. Note that the definition of resolvent (\[tilde quantities\]) yields that $$(\tb F \tb G)_{jj}-(\tbbra F^{(i)}\tb^{(i)} G^{(i)})_{jj} =(z_2-a_j) \Big( ( F \tb G)_{jj}- ( F^{(i)}\tb^{(i)} G^{(i)})_{jj}\Big) +( \tb G)_{jj}-(\tb^{(i)} G^{(i)})_{jj}.$$ Together with (\[removei\_1\]), the second inequality in (\[removei\]) follows directly from the first one. It is then sufficient to show the first inequality of (\[removei\]). From (\[h\_approximate\_r\]) and (\[bound\_d\]), we have $$\begin{aligned} &\ri=\wi+d_1 \ei+d_2 \gi; \qquad R_i-W_i={\bm w_i} {\bm w_i}^*-\ri \ri^*:=\Delta_i; \\ &\tb-\tbbra=R_i \tbhat R_i-W_i \tbhat W_i=\Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i .\end{aligned}$$ Using the second resolvent identity, we have $$G-\gbra=-G \Big( \tb-\tbbra \Big) \gbra =-G \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \gbra :=\delta(G),$$ $$F-\fbra=-F \Big( \tb-\tbbra \Big) \fbra =-F \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \fbra:=\delta(F).$$ Combining the definition of the resolvent with the resolvent identity (\[le rock 4\]), we write $$\begin{aligned} &{F \tb G}- { F^{(i)}\tb^{(i)} G^{(i)}}= z_1 F G- F A G+F -z_1 \fbra \gbra + \fbra A \gbra- \fbra\nonumber\\ &=\frac{z_1}{z_2-z_1}\Big( F-\fbra -G +\gbra \Big) + (F-\fbra)- (F A G- \fbra A \gbra). \label{20021401}\end{aligned}$$ The second term can be estimated easily by using the first inequality of (\[removei\_1\]). Furthermore, using the arguments in proving (\[argument\_1\]), one shows that the first term can be bounded by $O_{\prec}\Big( \frac{1}{\sqrt{N \eta_1} \eta_1}+\frac{1}{\sqrt{N \eta_2} \eta_1}\Big)$. It is sufficient to study the last term in (\[20021401\]). $$\begin{aligned} &(F A G)_{jj}-( \fbra A \gbra)_{jj}=( \fbra A \delta(G))_{jj}+( \delta(F) A \gbra)_{jj}+( \delta(F) A \delta(G))_{jj}\nonumber\\ &=-\ej^* \fbra A G \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \gbra \ej\nonumber\\& - \ej^* F \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \fbra A \gbra \ej \nonumber\\ &+\ej^* F \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \fbra A G \Big( \Delta_i \tbhat W_i +W_i \tbhat \Delta_i+\Delta_i \tbhat \Delta_i\Big) \gbra \ej \label{20011402}\end{aligned}$$ The first term of the right side of (\[20011402\]) is a polynomial of the terms of the following form: $$\xii^* \tbhat \gbra \ej =O_\prec(1); \qquad \xii^* \tbhat \xii =O_\prec(1); \qquad \xii^* \gbra \ej =O_\prec(1);$$ $$\ej^* \fbra A G \xii \prec \|A\|_{\mathrm{op}}\|G\|_{\mathrm{op}} \|\fbra \ej\|_2 =O_\prec \Big(\frac{1}{\sqrt{\eta_1}\eta_2}\Big) ; \qquad \ej^* \fbra A G \tbhat \xii=O_\prec \Big( \frac{1}{\sqrt{\eta_1} \eta_2 }\Big),$$ where $\xii$ is either $\gi$ or $\ei$, and the coefficients are in the form of $d_1^{k_1}d_2^{k_2} \prec \frac{1}{\sqrt{N}}$ with $k_1+k_2 \geq 1$, $k_1, k_2 \in \N$. Combining (\[bound\_d\]) with the above bounds, we obtain an upper bound of the first term. The second term can be treated similarly. As for the last term, it is a polynomial of the terms above and additional ones: $$\xii^* \tbhat \fbra A G \yi, \quad \xii^* \tbhat \fbra A G \tbhat \yi, \quad \xii^* \fbra A G \tbhat \yi =O_\prec \Big(\frac{1}{\eta_1\eta_2}\Big),$$ with coefficients in the form of $d_1^{k_1}d_2^{k_2} \prec \frac{1}{N}$ with $k_1+k_2 \geq 2$. Combining with (\[bound\_d\]), it is easy to obtain the first bound in (\[removei\]) and conclude the proof of Lemma \[remove\]. In the following, we will prove the concentration results for $S_i^{[2]}$ and $T_i^{[2]}$ in Lemma \[concentrate\_2\]. We will use the following large deviation bounds of Gaussian vectors, whose proof is standard. \[gaussian\_deviation\] Let $X=(x_{ij}) \in \C^{N \times N}$ be a deterministic matrix and let ${\bm y}=(y_i) \in \C^N$ be a deterministic vector. For a Gaussian random vector ${\bm g}=(g_1, \cdots, g_N) \sim {N}_{\R}(0,\sigma^2 I_N)$ or ${N}_{\C}(0,\sigma^2 I_N)$, we have $$|{\bm y}^* {\bm g}| \prec \sigma \|{\bm y}\|_2; \qquad |{\bm g}^* X {\bm g}-\sigma^2 \Tr X| \prec \sigma^2\|X\|_{\mathrm{HS}}.$$ Before we proceed to prove Lemma \[concentrate\_2\], we introduce the following rank-one perturbation formula: $$\label{rank_one_formula} (D+{\bm\alpha} {\bm\beta}^*)^{-1}=D^{-1} -\frac{D^{-1} {\bm\alpha} {\bm\beta}^* D^{-1}}{1+{\bm\beta}^* D^{-1} {\bm\alpha}},$$ for any ${\bm\alpha}, {\bm\beta} \in \C^N$ and invertible $D \in \C^{N \times N}$. As an application of the rank-one perturbation formula (\[rank\_one\_formula\]), we obtain the following trace formula. \[trace\_rank\] Let $Q, D \in \C^{N \times N}$ and $D$ be Hermitian. Then for any finite rank Hermitian matrix $R \in \C^{N \times N}$ with rank $r$, we have $$\Big| \frac{1}{N} \Tr \Big( Q(D+R-z)^{-1} \Big)-\frac{1}{N} \Tr \Big( Q(D-z)^{-1} \Big) \Big| \leq \frac{r \|Q\|_{\mathrm{op}}}{N \eta}.$$ We will next use Lemma \[trace\_rank\] to prove Lemma \[finite\_rank\_2\]. We start by showing the first line of (\[finite\_rank\_3\]). We first estimate the error of removing the upper index $(i)$. Using the definition of resolvent (\[tilde quantities\]) and the resolvent identity (\[le rock 4\]), we have $$\ud{\tbhat (\fbra \tbbra \gbra- F \tb G)}=z_1 \ud{\tbhat (\fbra \gbra- F G)}-\ud{\tbhat (\fbra A \gbra- F A G)}+\ud{\tbhat (\fbra - F )}$$ $$=\frac{z_1}{z_2-z_1} \ud{\tbhat (\fbra- \gbra- F +G)}-\ud{\tbhat (\fbra A \gbra- F A G)}+\ud{\tbhat (\fbra - F )}.$$ It is easy to check that the last term is bounded by $O_{\prec}\Big(\frac{1}{N \eta_2}\Big)$ using Lemma \[trace\_rank\], since $H^{(i)}$ is a rank-two perturbation of $H$. Furthermore, the first term is bounded by $O_{\prec}\Big(\frac{1}{N \eta_1^2}+\frac{1}{N \eta_1 \eta_2}\Big)$ using the arguments in the proof of (\[argument\_1\]). For the second term, iterating Lemma \[trace\_rank\] twice, we obtain that $$\ud{\tbhat (\fbra A \gbra- F A G)} =O_\prec\Big( \frac{1}{N \eta_1\eta_2}\Big).$$ Thus we have $$|\ud{\tbhat (\fbra \tbbra \gbra- F \tb G)}| \prec \frac{1}{N \eta_1^2}+\frac{1}{N \eta_1 \eta_2}+\frac{1}{N \eta_2}.$$ Next, we notice that $$\begin{aligned} \ud{\tbhat F \tb G}-\underline{\tb F \tb G}=\ud{R_i \tb R_i F \tb G}-\underline{\tb F \tb G}&=\frac{1}{N}\Big(-\ri^* F \tb G \tb \ri-\ri^* \tb F \tb G \ri +(\ri^* \tb \ri )(\ri^* F \tb G \ri)\Big)\nonumber\\ &\prec \frac{1}{N\eta_1\eta_2}\,.\end{aligned}$$ The last step follows the fact that $\|F \tb G \tb\|_{\mathrm{op}} \leq \|\tb\|^2_{\mathrm{op}} \|F\|_{\mathrm{op}}\|G\|_{\mathrm{op}} \leq \frac{C}{\eta_1 \eta_2}$. Hence we prove the first inequality of (\[finite\_rank\_3\]). The second one of (\[finite\_rank\_3\]) can be treated similarly. In addition, (\[trivial\_bounds\]) is implied by (\[finite\_rank\_3\]) and $$\label{bfbg_later} \ud{\tb F \tb G}=O_\prec \Big(\frac{1}{\sqrt{\eta_1\eta_2}}\Big);\qquad \ud{\tb F \tb G \tb}=O_\prec \Big(\frac{1}{\sqrt{\eta_1\eta_2}}\Big).$$ Finally, we will prove the concentration inequalities (\[concentrate\_4\]). We will only prove the first one for simplicity. Note that $\tbhat$ is independent of $\gi$, and thus $$\IE \underline{\tbhat \fbra \tbbra \gbra}=\IE \Big[ \underline{\tbhat \fbra \tbbra \gbra} -\ud {\tbhat \fhat \tbhat \ghat}\Big].$$ Since $H^{\<i\>}$ is a rank-two perturbation of $H^{(i)}$, using previous arguments and Lemma \[trace\_rank\] repeatedly, we obtain the desired result. Now, we are ready to prove Lemma \[concentrate\_2\]. We will only prove the last two inequalities by studying $\IE \xii^* \fbra \tbbra \gbra \ei$, where $\xii$ equals either $\gi$ or $\ei$ for simplicity. Note that by (\[approximate\_qq\]) and the definition of resolvent, we have $$\begin{aligned} S^{[2]}_i=&(a_i-z_2) (\fbra \tbbra \gbra)_{ii}+( a_i-z_1)\gbra_{ii}-1+O_{\prec}\Big(\frac{1}{\sqrt{N \eta_1 \eta_2}}\Big).\end{aligned}$$ Combining with (\[concentrate\_1\]), the last inequality in Lemma \[concentrate\_2\] implies the first one. Note that $$\begin{aligned} \label{approximate_later} \xii^* \fbra \tbbra \gbra \ei=&\xii^* \fbra (1-\wi \wi^*) \tbhat (1-\wi \wi^*) \gbra \ei\nonumber\\ =&\xii^* \fbra \tbhat \gbra \ei-\xii^* \fbra \wi \wi^* \tbhat \gbra \ei \nonumber\\ &-\xii^* \fbra \tbhat \wi \wi \gbra \ei+\xii^* \fbra \wi \wi^* \tbhat \wi \wi^* \gbra \ei.\end{aligned}$$ Using the property of $\IE$, $$\label{triangle} \IE[XY]=\IE[\IE X \IE Y]+\IE X \E_{\gi} Y+\E_{\gi} X \IE Y,$$ in combination with (\[concentrate\_1\]) and Lemma \[previous\_bound\], it is enough to estimate $\IE[ \xii^* \fbra \tbhat \gbra \ei]$. Recalling that $G^{\<i\>}$ in (\[hat\_definition\]) is independent of $\gi$, it is hence natural to expand $G^{(i)}$ around $G^{\<i\>}$ and then use Lemma \[gaussian\_deviation\] to obtain the concentration results. However, from the construction in (\[hat\_definition\]), $$G^{\<i\>}_{ii}(z)=\frac{1}{a_i+b_i-z}$$ is not stable for $z \in D_{bulk}$. To overcome this problem, we further introduce $$\begin{aligned} H^{{\lbrace i \rbrace}}:=A+{\widetilde}B^{(i)}-(b_i+\omega_B(z)-z)\ei\ei^*\,,\qquad G^{{\lbrace i \rbrace}}(z):=\frac{1}{H^{{\lbrace i \rbrace}}-z}\,,\qquad z\in\C^+\,,\end{aligned}$$ to enhance the stability in the $\ei$ direction. Note that from the rank-one perturbation formula Lemma \[rank\_one\_formula\], for $1 \leq i,j \leq N$ we have $$\begin{aligned} \label{hello} G_{ij}^{{{\lbrace i \rbrace}}}(z)=G^{(i)}_{ij}(z)+\frac{(b_i+\omega_B(z)-z)G^{(i)}_{ii}(z) G^{(i)}_{ij}(z)}{1-(b_i+\omega_B(z)-z)G^{(i)}_{ii}(z)}=\frac{G^{(i)}_{ij}(z)}{1-(b_i+\omega_B(z)-z)G^{(i)}_{ii}(z)}.\end{aligned}$$ For $z\in D_{bulk}$, we have from the local laws of $G^{(i)}_{ii}(z)$, $$\begin{aligned} \label{le pseudo local laws} \frac{1}{{1-(b_i+\omega_B(z)-z)G^{(i)}_{ii}}}= \frac{a_i-\omega_B(z)}{a_i-b_i-2\omega_B(z)+z} +O_{\prec}(\Psi)=O_\prec (1)\,,\end{aligned}$$ because of (\[m\_bound\]), (\[imaginary\_bound\]) and also $|2 {\mathrm{Im}\,}\omega_B(z_1)-{\mathrm{Im}\,}z_1| >c >0$ for $z_1 \in D_{bulk}$. Together with Lemma \[previous\_bound\], we obtain that $$\label{le pseudo local laws} G^{\{i\}}_{ii}(z)= \frac{1}{a_i-b_i-2\omega_B(z)+z}+O_{\prec}(\Psi(z))=O_\prec(1); \qquad G^{\{i\}}_{ij} =O_{\prec}(\Psi(z)), \qquad j \neq i.$$ Next, we will replace $\fbra$, $\gbra$ by the regularized $F^{{\lbrace i \rbrace}}$ and $G^{{{\lbrace i \rbrace}}}$. As a consequence of the rank-one perturbation formula (\[rank\_one\_formula\]), we get for general $\yone$ and $\ytwo$, $$\begin{aligned} \label{le 2} \yone^* G^{(i)} \ytwo &=\yone^* G^{{\lbrace i \rbrace}}\ytwo-\frac{(b_i+\omega_B(z_1)-z_1)\yone^* G^{{{\lbrace i \rbrace}}}\ei \ei^*G^{{{\lbrace i \rbrace}}}\ytwo}{1+(b_i+\omega_B(z_1)-z_1)G_{ii}^{{\lbrace i \rbrace}}}\,. \end{aligned}$$ Set $\yone^*=\xii^* \fbra \tbhat$ and $\ytwo=\ei$, then $$\begin{aligned} \label{temp_111} \xii^* \fbra \tbhat \gbra \ei=& \xii^* \fbra \tbhat G^{{\lbrace i \rbrace}}\ei-\frac{(b_i+\omega_B(z_1)-z_1) \xii^* \fbra \tbhat G^{{{\lbrace i \rbrace}}} \ei G^{{{\lbrace i \rbrace}}}_{ii}}{1+(b_i+\omega_B(z_1)-z_1)G_{ii}^{{\lbrace i \rbrace}}}\nonumber\\ =&\frac{ \xii^* \fbra \tbhat G^{{\lbrace i \rbrace}}\ei}{1+(b_i+\omega_B(z_1)-z_1)G_{ii}^{{\lbrace i \rbrace}}}:= \xii^* \fbra \tbhat G^{{\lbrace i \rbrace}}\ei \Lambda_1(z_1), \end{aligned}$$ where we have the estimate from , $$\label{Lambda_1} \Lambda_1(z_1):=\frac{1}{1+(b_i+\omega_B(z_1)-z_1)G_{ii}^{{\lbrace i \rbrace}}}=\frac{a_i-b_i-2\omega_B(z_1)+z_1}{a_i-\omega_B(z_1)}+O_\prec(\Psi(z_1)),$$ Note the $ \Lambda_1(z_1)$ is asymptotically deterministic and $| \Lambda_1(z_1)| \sim 1$, because of (\[m\_bound\]) and (\[imaginary\_bound\]) for $z_1 \in D_{bulk}$. Applying the Cauchy-Schwarz inequality, using the estimates for $F^{(i)}$ in Lemma \[previous\_bound\] and the local law of $G^{\{i\}}_{ij}$ in (\[le pseudo local laws\]), we have $$\label{CS_ward} |\xii^* \fbra \tbhat G^{\{i\}} \ei| \leq \|\tbhat\|_{\mathrm{op}} \|\xii^* \fbra \|_2 \| G^{\{i\}} \ei \|_2 \leq C\|\xii^* \fbra \|_2 \| G^{\{i\}} \ei \|_2\prec \frac{1}{\sqrt{\eta_1\eta_2}}.$$ Combining (\[triangle\]), (\[temp\_111\]), (\[Lambda\_1\]) and (\[CS\_ward\]), it suffices to study $\IE [\xii^* \fbra \tbhat G^{{\lbrace i \rbrace}}\ei]$ in order to estimate $\IE [\xii^* \fbra \tbhat \gbra \ei]$. We use the rank-one perturbation formula (\[rank\_one\_formula\]) again by letting $\yone=\xii$ and $\ytwo=\tbhat G^{{\lbrace i \rbrace}}\ei$, and have $$\xii^* \fbra \tbhat G^{{\lbrace i \rbrace}}\ei=\xii^* F^{{\lbrace i \rbrace}}\tbhat G^{{\lbrace i \rbrace}}\ei-\frac{(b_i+\omega_B(z_2)-z_2) \xii^* F^{{{\lbrace i \rbrace}}} \ei \ei^* F^{{{\lbrace i \rbrace}}}\tbhat G^{{\lbrace i \rbrace}}\ei}{1+(b_i+\omega_B(z_2)-z_2)F_{ii}^{{\lbrace i \rbrace}}}.$$ From , we have the estimate $$\Lambda_2(z_2):=-\frac{(b_i+\omega_B(z_2)-z_2) }{1+(b_i+\omega_B(z_2)-z_2)F_{ii}^{{\lbrace i \rbrace}}}=\frac{(b_i+\omega_B(z_2)-z_2)(a_i-b_i-2\omega_B(z_2)+z_2)}{-a_i+\omega_B(z_2)}+O_\prec(\Psi(z_2)).$$ Note the first term on the right side is deterministic and at constant order since $z'$ in the regular bulk. Similarly as (\[CS\_ward\]), we have $$|\ei^* F^{{{\lbrace i \rbrace}}} \tbhat G^{{{\lbrace i \rbrace}}} \ei| \leq \|\tbhat\|_{\mathrm{op}} \|F^{{\lbrace i \rbrace}}\ei \|_2 \| G^{{\lbrace i \rbrace}}\ei \|_2 \prec \frac{1}{\sqrt{\eta_1\eta_2}}.$$ Combining with (\[triangle\]) and the concentration results in Lemma \[lemma\_concentrate\], it is hence enough to estimate $$\IE \xii^* F^{{\lbrace i \rbrace}}\tbhat G^{{\lbrace i \rbrace}}\ei.$$ Introduce now the block-diagonal matrix $$\begin{aligned} \label{final_G} H^{{[i]}}:=A+{\widetilde}B^{{{\langle i\rangle}}}-(b_i+\omega_B(z)-z)\ei\ei^*\,, \qquad G^{{[i]}}=( H^{{[i]}}-z)^{-1}.\end{aligned}$$ Since $G^{{{[i]}}}$ is independent of $\gi$, we expand $G^{{{\lbrace i \rbrace}}}$ around $G^{{{[i]}}}$ and then apply Lemma \[gaussian\_deviation\]. Note that $$\begin{aligned} \label{resolvent_entry_Grir} G_{ii}^{{[i]}}(z)=\frac{1}{a_i-\omega_B(z)}=O(1)\,,\qquad G_{ij}^{{[i]}}(z)=0\,,\qquad j\not=i\,.\end{aligned}$$ It is straightforward to check that $$\begin{aligned} H^{{\lbrace i \rbrace}}-H^{{[i]}}=\wi\si^*+\ti\wi^*\,,\end{aligned}$$ with $$\begin{aligned} \wi=\ei+\gi\,,\qquad \si=-{\widetilde}B^{{{\langle i\rangle}}}\wi\,,\qquad \ti=-({\widetilde}B^{{\langle i\rangle}}-\wi^*{\widetilde}B^{{\langle i\rangle}}\wi I)\wi\,.\end{aligned}$$ Iterating the rank-one perturbation formula  twice, we obtain the following lemma: \[le lemma twice rank one\] $$\begin{aligned} \label{le twice rank one} G^{{\lbrace i \rbrace}}=G^{{[i]}}+\frac{\Pi}{1+\Xi_i(z)}\,,\end{aligned}$$ with $\Xi_i(z)$ is given by $$\begin{aligned} \Xi_i(z)=\si^*G^{{[i]}}\wi+\wi^*G^{{[i]}}\ti+\si^*G^{{[i]}}\wi \wi^*G^{{[i]}}\ti-\wi^*G^{{[i]}}\wi \si^*G^{{[i]}}\ti\,.\end{aligned}$$ and $\Pi$ is the matrix $$\begin{aligned} \label{le Pi} \Pi(z)\equiv\Pi&=(\si^*G^{{[i]}}\ti)G^{{[i]}}\wi\wi^* G^{{[i]}}+(\wi^* G^{{[i]}}\wi)G^{{[i]}}\ti\si^*G^{{[i]}}\nonumber\\ &\qquad\qquad-(1+(\si G^{{[i]}}\wi))G^{{[i]}}\ti\wi^* G^{{[i]}}-(1+(\wi^*G^{{[i]}}\ti))G^{{[i]}}\wi \si^*G^{{[i]}}\,.\end{aligned}$$ Furthermore, we have $$\begin{aligned} \label{le bound from eta1 paper} |\IE [\Xi_i(z)]|\prec\frac{1}{\sqrt{N\eta}}\,,\qquad \big|\frac{1}{1+\E_\gi [\Xi_i(z)]}\big|\prec 1\,.\end{aligned}$$ Moreover, for $Q_1^{{{\langle i\rangle}}}, Q_2^{{{\langle i\rangle}}}$ either $\tbhat$ or $I$, $$\label{eta1_bound} |\IE [\wi^* Q_1^{{{\langle i\rangle}}} G^{{{[i]}}} Q_2^{{{\langle i\rangle}}} \wi] |\prec \frac{1}{\sqrt{N \eta}}; \qquad |\wi^* Q_1^{{{\langle i\rangle}}} G^{{{[i]}}} Q_2^{{{\langle i\rangle}}} \wi| \prec 1;\qquad \wi^* \tbhat \wi =b_i+O_{\prec}\Big(\frac{1}{\sqrt{N}}\Big).$$ From Lemma \[le lemma twice rank one\] we have $$\begin{aligned} \xii^* F^{{{\lbrace i \rbrace}}}{\widetilde}B^{{{\langle i\rangle}}}G^{{{\lbrace i \rbrace}}}\ei&=\xii^*(F^{{[i]}}+\frac{\Pi(z_2)}{1+\Xi_i(z_2)}){\widetilde}B^{{{\langle i\rangle}}}(G^{{[i]}}+\frac{\Pi(z_1)}{1+\Xi_i(z_1)})\ei\nonumber\\ &=b_i \xii^* F^{{[i]}}\ei G_{ii}^{{[i]}}+ b_i\frac{\xii^*\Pi(z_2)\ei G_{ii}^{{[i]}}}{1+\Xi_i(z_2)}+ \frac{\xii^*F^{{[i]}}\tbhat \Pi(z_1)\ei}{1+\Xi_i(z_1)}+\frac{\xii^*\Pi(z_2){\widetilde}B^{{\langle i\rangle}}\Pi(z_1)\ei}{(1+\Xi_i(z_2))(1+\Xi_i(z_1))}\label{20011410}\end{aligned}$$ Take $\IE$ on both sides. Since $G^{[i]}(z)$ is independent of $\gi$, we apply Lemma \[gaussian\_deviation\] and (\[resolvent\_entry\_Grir\]) to obtain that $$|\IE[\gi F^{[i]} \ei]|= |\gi F^{[i]} \ei| =|F^{[i]}_{ii}| |\gi^* \ei| \prec \frac{1}{\sqrt{N}}.$$ Thus the concentration of the first term on the right side of (\[20011410\]) can be bounded by $O_{\prec}\Big(\frac{1}{\sqrt{N}}\Big)$. The rest terms can also be treated similarly, using the property of $\IE$ (\[triangle\]), the estimate of the denominator (\[le bound from eta1 paper\]) and the estimates of numerators of each one as following. Observe that for $z=E+{\mathrm{i}}\eta$, $$\begin{aligned} \label{Pi_ii} \xii^*\Pi(z) \ei&=(\si^*G^{{[i]}}\ti) \xii^* G^{{[i]}}\wi\wi^* G^{{[i]}}\ei +(\wi^* G^{{[i]}}\wi) \xii^*G^{{[i]}}\ti\si^*G^{{[i]}}\ei \nonumber\\ &\qquad\qquad-(1+(\si G^{{[i]}}\wi)) \xii^* G^{{[i]}}\ti\wi^* G^{{[i]}}\ei-(1+(\wi^*G^{{[i]}}\ti)) \xii^* G^{{[i]}}\wi \si^*G^{{[i]}}\ei \,.\end{aligned}$$ Combining with (\[eta1\_bound\]), we hence obtain from the property of $\IE$ (\[triangle\]) that $$\label{estimate_111} \IE [\xii^*\Pi(z)\ei] \prec \frac{1}{\sqrt{N \eta}}, \qquad |\xii^* \Pi(z) \ei|\prec 1.$$ Next, we look at $$\begin{aligned} \label{hahaha} \xii^*F^{{[i]}}\tbhat \Pi(z_1)\ei &=(\si^*G^{{[i]}}\ti)(\xii^* F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}\wi)(\wi^* G^{{[i]}}\ei) \nonumber\\ &\qquad+(\wi^* G^{{[i]}}\wi)(\xii^* F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}\ti)(\si^*G^{{[i]}}\ei)\nonumber\\ &\qquad-(1+(\si G^{{[i]}}\wi))(\xii^* F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}\ti)(\wi^* G^{{[i]}}\ei)\nonumber\\ &\qquad-(1+(\wi^*G^{{[i]}}\ti))(\xii^* F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}\wi)( \si^*G^{{[i]}}\ei)\,.\end{aligned}$$ The right side can be written in terms of $$\wi^*Q^{{\langle i\rangle}}G^{{[i]}}Q^{{\langle i\rangle}}\wi; \qquad F^{{[i]}}_{ii} G_{ii}^{{[i]}}; \qquad F^{{[i]}}_{ii} G_{ii}^{{[i]}}\ei Q^{{\langle i\rangle}}\gi; \qquad \gi^* F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{\langle i\rangle}}\gi.$$ We will only look at the last term for simplicity. Using Lemma \[gaussian\_deviation\] for $\gi$, we obtain that $$\begin{aligned} |\IE (\gi^*Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}\gi )|\prec\frac{1}{\sqrt{N}}\|Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{\langle i\rangle}}\|_{\mathrm{HS}}\,.\end{aligned}$$ Observe that $$\begin{aligned} \|Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{\langle i\rangle}}\|_{\mathrm{HS}}&\le \| Q^{{\langle i\rangle}}\|_{\mathrm{op}} \|Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}\|_{\mathrm{op}}\,\|G^{{[i]}}\|_{\mathrm{HS}} \prec \frac{1}{\eta_2}\big(\frac{\mathrm{Tr}{{\mathrm{Im}\,}G^{{[i]}}}}{\eta_1}\big)^{1/2}.\end{aligned}$$ The last step follows from the modified ward identities $$\label{ward_identity_1} |G^{[i]}(z)|_{jj}^2=\frac{{\mathrm{Im}\,}(G^{[i]}(z))_{jj}}{(1-\delta_{ij})\eta+\delta_{ij} {\mathrm{Im}\,}\omega_B(z)}, \qquad z=E+{\mathrm{i}}\eta.$$ Thus, by (\[resolvent\_entry\_Grir\]), we have $$\begin{aligned} |\IE (\gi^*Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}\gi )| \prec \frac{1}{\sqrt{N\eta_1}\eta_2}\,.\end{aligned}$$ Note that $$\E_{\gi}[\gi^*Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{\langle i\rangle}}\gi] =\frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{\langle i\rangle}})$$ Due to the construction (\[final\_G\]) and (\[resolvent\_entry\_Grir\]), we have $$\begin{aligned} \frac{1}{N}\big|\mathrm{Tr} Q^{{{\langle i\rangle}}} G^{{[i]}}Q^{{{\langle i\rangle}}}-\mathrm{Tr} Q^{{{\langle i\rangle}}} \ghat Q^{{{\langle i\rangle}}}\big|&= \frac{1}{N}\big|\mathrm{Tr} Q^{{{\langle i\rangle}}} G^{{[i]}}(b_i+\omega_B(z)-z) \ei \ei^* \ghat Q^{{{\langle i\rangle}}} \big| \nonumber\\ &\prec \frac{1}{N} \| Q^{{{\langle i\rangle}}} Q^{{{\langle i\rangle}}} G^{[i]} \ei \|_2 \| \ghat\ei\|_2 \prec\frac{1}{N\eta_1}.\end{aligned}$$ Furthermore, we have $$\begin{aligned} \Big| \frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}) -\frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}\fhat {\widetilde}B^{{{\langle i\rangle}}} \ghat Q^{{{\langle i\rangle}}}) \Big| \prec \frac{1}{N \eta_1 \eta_2}.\end{aligned}$$ In addition, since $H^{\<i\>}$ is a Hermitian finite-rank perturbation of $H$, by (\[trace\_rank\]), we have $$\Big| \frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}) -\frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}F {\widetilde}B^{{{\langle i\rangle}}} G Q^{{{\langle i\rangle}}}) \Big| \prec \frac{1}{N \eta_1 \eta_2}.$$ Combining with (\[finite\_rank\_3\]) and (\[bfbg\_later\]), we hence obtain a priori bound: $$\label{priori_bound} \Big|\frac{1}{N}\mathrm{Tr}(Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}) \Big| \prec \frac{1}{\sqrt{\eta_1 \eta_2}}.$$ Thus, we have the following estimate: $$|\gi^*Q^{{\langle i\rangle}}F^{{[i]}}{\widetilde}B^{{{\langle i\rangle}}}G^{{[i]}}Q^{{{\langle i\rangle}}}\gi | \prec \frac{1}{\sqrt{\eta_1 \eta_2}}.$$ Therefore, using the property of $\IE$ (\[triangle\]) on (\[hahaha\]), we have $$\label{estimate_222} | \IE [\xii^*F^{{[i]}}\tbhat \Pi(z_1)\ei ] |\prec \frac{1}{\sqrt{N\eta_1}\eta_2}; \qquad |\xii^*F^{{[i]}}\tbhat \Pi(z_1)\ei | \prec \frac{1}{\sqrt{\eta_1\eta_2}}.$$ Finally, we treat $\xii^* \Pi(z_2) \tbhat \Pi(z_1)\ei$ in the same way. With above bounds, using the property of $\IE$ (\[triangle\]) on (\[20011410\]), we have $$|\IE \xii^* F^{{\lbrace i \rbrace}}\tbhat G^{{\lbrace i \rbrace}}\ei | \prec \frac{1}{\sqrt{N\eta_1}\eta_2}+\frac{1}{\sqrt{N\eta_2}\eta_1},$$ and we hence complete the proof of Lemma \[concentrate\_2\]. [99]{} A. Adhikari, J. Huang: [*Dyson Brownian Motion for General $\beta$ and Potential at the Edge*]{}, preprint, arXiv:1810.08308, (2018). N. I. Akhiezer: [*The classical moment problem: and some related questions in analysis*]{}, Hafner Publishing Co., New York (1965). G. Anderson, A. Guionnet, O. Zeitouni: *An introduction to random matrices*, Cambridge Stud. Adv. Math. [**118**]{}, Cambridge Univ. Press, Cambridge, 2010. Z. Bao, L. Erdős, K. Schnelli: [*Local stability of the free additive convolution*]{}, J. Funct. Anal. [**271(3)**]{}, 672-719 (2016). Z. Bao, L. Erdős, K. Schnelli: [*Local law of addition of random matrices on optimal scale*]{}, Comm. Math. Phys. [**349(3)**]{}, 947-990 (2016). Z. Bao, L. Erdős, K. Schnelli: [*Convergence rate for spectral distribution of addition of random matrices*]{}, Adv. Math. [**319**]{}, 251-291 (2017). Z. Bao, L. Erdős, K. Schnelli: [*On the support of the free addictive convolution*]{}, preprint arXiv:1804.11199 (2018). Z. Bao, L. Erdős, K. Schnelli: [*Spectral rigidity for addition of random matrices at the regular edge*]{}, preprint, arXiv:1708.01597 (2017). E. L. Basor, H. Widom: [*Determinants of Airy Operators and Applications to Random Matrices*]{}, J. Stat. Phys. [**96**]{}, 1-20 (1999). F. Bekerman, A. Lodhia: [*Mesoscopic central limit theorem for general $\beta$-ensembles*]{}, Ann. Inst. H. Poincaré Probab. Statist. [**54**]{}, 1917-1938 (2018). S.T. Belinschi: [*A note on regularity for free convolutions*]{}, Ann. Inst. Henri Poincaré Probab. Stat. [**42(5)**]{}, 635-648 (2006). S.T. Belinschi: [*The Lebesgue decomposition of the free additive convolution of two probability distributions*]{}, Probab. Theory Relat. Fields [**142**]{}(1-2), 125-150 (2008). S.T. Belinschi: [*$L^{\infty}$-boundedness of density for free additive convolutions*]{}, Rev. Roumaine Math. Pures Appl. [**59(2)**]{}, 173-184 (2014). S.T. Belinschi, H. Bercovici: [*A new approach to subordination results in free probability*]{}, J. Anal. Math. [**101**]{}, 357-366 (2007). H. Bercovici, D. Voiculescu: [*Free convolution of measures with unbounded support*]{}, Indiana Univ. Math. J. **42**, 733-773 (1993). P. Biane: [*Process with free increments*]{}, Math. Z. **227(1)**, 143-174 (1998). P. Bourgade: [*Extreme gaps between eigenvalues of Wigner matrices*]{}, preprint arXiv:1812.10376, (2018). P. Bourgade, L. Erdős, H.-T. Yau, J. Yin: [*Fixed energy universality for generalized Wigner matrices*]{}, Comm. Pure Appl. Math. [**69**]{} (2015). P. Bourgade, K. Mody: [*Gaussian fluctuations of the determinant of Wigner matrices*]{}, Electron. J. Probab. [**24**]{}(96), 1-28 (2019). A. Boutet de Monvel, A. Khorunzhy: [*Asymptotic distribution of smoothed eigenvalue density. I. Gaussian random matrices*]{}, Random Oper. and Stoch. Equ. [**7(1)**]{}, 1-22 (1999). A. Boutet de Monvel, A. Khorunzhy: [*Asymptotic distribution of smoothed eigenvalue density. II. Wigner random matrices*]{}, Random Oper. and Stoch. Equ. [**7(2)**]{} 149-168 (1999). J. Breuer, M. Duits: [*Universality of mesoscopic fluctuations for orthogonal polynomial ensembles*]{}, Comm. Math. Phys. [**342**]{} (2), 491-531 (2016). Z. Che, B. Landon: [*Local spectral statistics of the addition of random matrices*]{}, preprint, arXiv:1701.00513 (2017). G. Cipolloni, L. Erdős, D. Schröder: [*Central limit theorem for linear eigenvalue statistics of non-Hermitian random matrices*]{}, preprint arXiv:1912.04100 (2019). G. P. Chistyakov, F. Götze: [*The arithmetic of distributions in free probability theory*]{}, Cent. Euro. J. Math. **9**, 997-1050 (2011). B. Collins, J. A. Mingo, P. Śniady, R. Speicher: [*Second order freeness and fluctuations of random matrices. III. Higher order freeness and free cumulants*]{}, Doc. Math [**12**]{}, 1-70 (2007). P. Diaconis, M. Shahshahani: [*The subgroup algorithm for generating uniform random variables*]{}, Probab. Engrg. Inform. Sci. [**1(01)**]{}, 15-32 (1987). M. Duits, K. Johansson: [*On mesoscopic equilibrium for linear statistics in Dyson’s Brownian Motion*]{}, Mem. Amer. Math. Soc. [**255**]{} (1222), (2018). Y. Fyodorov, B. Khoruzhenko, N, Simm: [*Fractional Brownian motion with Hurst index $ H= 0$ and the Gaussian unitary ensemble*]{}, The Annals of Probability, 2016, 44(4): 2980-3031 (2016). L. Erdős, A. Knowles: [*The Altshuler-Shklovskii formulas for random band matrices I: the unimodular case*]{}, Comm. Math. Phys. [**333**]{}, 1365-1416 (2015). L. Erdős, A. Knowles: [*The Altshuler-Shklovskii formulas for random band matrices II: the general case*]{}, Ann. H. Poincaré [**16**]{}, 709-799 (2015). Y. He, A. Knowles: [*Mesoscopic eigenvalue statistics of Wigner matrices*]{}, Ann. Appl. Probab. [**27**]{}(3), 1510-1550 (2017). V. Kargin: [*A concentration inequality and a local law for the sum of two random matrices*]{}, Probab. Theory Relat. Fields [**154**]{}, 677-702 (2012). V. Kargin: [*An inequality for the distance between densities of free convolutions*]{}, Ann.Probab. [**41**]{}(5), 3241-3260 (2013). V. Kargin: [*Subordination for the sum of two random matrices*]{}, Ann.Probab. [**43(4)**]{}, 2119-2150 (2015). B. Landon, P. Sosoe: [*Applications of mesoscopic CLTs in random matrix theory*]{}, preprint, arXiv:1811.05915 B. Landon, P. Sosoe, H.-T. Yau: [*Fixed energy universality of Dyson Brownian motion*]{}, Advances in Math, [**346**]{}, 1137-1332, 2019. Y. Li , K. Schnelli, Y. Xu: [*Central limit theorem for mesoscopic eigenvalue statistics of deformed Wigner matrices and sample covariance matrix*]{}, preprint, arXiv:1909.12821. Y. Li, Y. Xu: [*On fluctuations and global and mesoscopic linear statistics of generalized Wigner matrices*]{}, preprint (2020). A. Lodhia, N. Simm: [*Mesoscopic linear statistics of Wigner matrices*]{}, Preprint, arXiv:1503.03533. A. Lytova, L. Pastur: [*Central limit theorem for linear eigenvalue statistics of random matrices with independent entries*]{}, Ann. Prob. [**37(5)**]{}, 1778-1840 (2009). F. Mezzadri: [*How to generate random matrices from the classical compact groups*]{}, Notices Amer. Math. Soc. [**54(5)**]{}, 592-604 (2007). C. Min, Y. Chen: [*Linear Statistics of Random Matrix Ensembles at the Spectrum Edge Associated with the Airy Kerne*]{}l, Nucl. Phys. B [**950**]{}, (2020). J. A. Mingo, P. Śniady, R. Speicher: [*Second order freeness and fluctuations of random matrices: II. Unitary random matrices*]{}, Adv. Math. [**209.1**]{}, 212-240 (2007). J. A. Mingo, R. Speicher: [*Second order freeness and fluctuations of random matrices: I. Gaussian and Wishart matrices and cyclic Fock spaces*]{}, J. Funct. Anal. [**235.1**]{}, 226-270 (2006). J. A. Mingo, R. Speicher: [*Free probability and random matrices*]{}, Fields Institute Monographs 35, Springer (2017). L. Pastur, M. Shcherbina. [*Eigenvalue distribution of large random matrices*]{}, Mathematical Surveys and Monographs 171, American Mathematical Society, (2011). L. Pastur, V. Vasilchuk: [*On the Law of Addition of Random Matrices*]{}, Commun. Math. Phys. [**214**]{}, 246-286 (2000). A. Soshnikov: [*Central Limit Theorem for Linear Statistics in Classical Compact Groups and Related Combinatorial Identities*]{}, Ann. Probab. [**28**]{}, 1353-1370 (2000). D. Voiculescu: [*Limit laws for Random matrices and free products*]{}, Invent. Math. [**104**]{}(1), 201-220 (1991). D. Voiculescu: [*The analogues of entropy and of Fisher’s information theory in free probability theory, I*]{}, Comm. Math. Phys. **155**, 71-92 (1993).

--- abstract: 'Using Lindblad approach we develop a general formalism for theoretical description of a spatially inhomogenous bosonic system with dissipation provided by the interaction of bosons with a phonon bath. We apply our results to model the dynamics of an interacting one dimensional polariton system in real space and time, analyzing in detail the role of polariton-polariton and polariton- phonon interactions.' author: - 'I. G. Savenko' - 'E.B. Magnusson' - 'I. A. Shelykh' title: 'Density-matrix approach for an interacting polariton system' --- I. Introduction =============== A semiconductor microcavity is a photonic structure designed to enhance the light-matter interaction. In a planar microcavity photons are confined between two mirrors and resonantly interact with the excitonic transition of a two-dimensional semiconductor quantum well (QW). If the quality factor of the cavity is sufficiently high, it is possible to achieve the regime of strong coupling between the cavity photon and QW exciton. In this case the elementary excitations in the system, which are called cavity polaritons, have a hybrid, half-light, half-matter nature. The peculiar properties of polaritons make them a unique laboratory for studying of various collective phenomena interesting from the point of view of basic physics, which range from polariton BEC [@KasprzakNature], superfluidity [@AmoNature] and Josephson effect [@LagoudakisJosephson] to polariton-mediated superconductivity [@LaussySupercond]. Besides fundamental interest, quantum microcavities in the strong coupling regime can be used for a variety of optoelectronic applications [@Imamoglu]. Recently, it was proposed that the peculiar spin structure and possibility to achieve lateral confinement of polaritons [@Confinement] opens a way for creation of optical analogs of spintronic components (so-called spinoptronic devices), based on transport of cavity polaritons in real space. In this context, the analysis of one-dimensional (1D) polariton transport is of particular importance [@WertzNature], as 1D polariton channels are fundamental building blocks of such future spinoptronic devices as polariton neurons [@LiewNeuron] and polariton integrated circuits [@LiewCircuit]. On the theoretical side, transport properties of exciton-polaritons in real space have not yet been studied in detail. Early works [@SpinHall; @Glazov; @ShelykhBerry] treated the case of polaritons interacting with external potentials only neglecting both polariton-polariton and polariton-phonon interactions. However, in the most interesting regime of polariton BEC neither of the two interactions can in principle be neglected. Indeed, coupling of the polaritons with a reservoir of acoustic phonons leads to thermalization of the polariton subsystem, which is dramatically speeded up by polariton- polariton interactions which are known to be responsible for overcoming the so-called “bottleneck effect” [@Bottleneck]. Besides, in the regime of polariton BEC polariton-polariton interactions are responsible for the onset of superfluidity [@AmoNature]. Currently, there are two ways of describing a system of interacting polaritons. Assuming full coherence of the polariton system, polariton-polariton interactions can be accounted for using nonlinear Gross- Pitaevskii (GP) equation, satisfactorily describing the dynamics of inhomogeneous polariton droplets in real space and time [@Carusotto; @ShelykhGP]. The approach, however, does not include interactions with a phonon bath, responsible for thermalization of the system and leading to dephasing. In the opposite limit, when the polariton system is supposed to be fully incoherent, its dynamics can be described using a system of semiclassical Boltzmann equations [@Porras2002; @Kasprzak2008; @Haug2005; @Cao], which provides information about time dependence of the occupation numbers in reciprocal space but fails to describe real space dynamics in the inhomogeneous system. Recently, there appeared theoretical attempts to combine the two mentioned approaches introducing dissipation terms into GP equation in a phenomenological way [@Wouters2007]. To our mind, however, these attempts, although being interesting and leading to a rich phenomenology [@BerloffVortex; @Tim1D], are not fully satisfactory, as they lack any microscopic justification. In the present paper we give an alternative way of describing the dynamics of an inhomogeneous polariton system in real space and time accounting for dissipation effects. Our consideration is based on Lindblad approach for density matrix dynamics. We use our results for modeling of the propagation of polariton droplets in 1D channels. It should be noted, that the method we develop is rather general and can in principle be applied to any system of interacting bosons in contact with a phonon reservoir (for example, indirect excitons [@Butov]). We have previously applied this formalism to model Josephson oscillations between two spatially separated condensates, but in that context the spatial dependence was trivial [@Josephson2010]. II. Formalism ============= We describe the state of the system (polaritons plus phonons) by its density matrix $\rho$, for which we apply Born approximation factorizing it into the phonon part which is supposed to be time-independent and corresponds to the thermal distribution of acoustic phonons $\rho_{ph}=\texttt{exp}\left\{-\beta\widehat{H}_{ph}\right\}$ and the polariton part $\rho_{pol}$ whose time dependence should be determined, $\rho=\rho_{ph}\otimes \rho_{pol}$. Our aim is to find dynamic equations for the time evolution of the single-particle polariton density matrix in real space. $$\begin{aligned} \rho(\textbf{r},\textbf{r}',t)=Tr\left\{\widehat{\psi}^\dagger(\textbf{r},t)\widehat{\psi}(\textbf{r}',t)\rho\right\} \equiv\langle\widehat{\psi}^\dagger(\textbf{r},t)\widehat{\psi}(\textbf{r}',t)\rangle\end{aligned}$$ where $\widehat{\psi}^\dagger(\textbf{r},t),\widehat{\psi}(\textbf{r},t)$ are operators of the polariton field, and the trace is performed by all degrees of freedom of the system. The particularly interesting quantities are the diagonal matrix elements which give the density of the polariton field in real space and time $n(\textbf{r},t)=\rho(\textbf{r},\textbf{r},t)$. In our consideration we neglect the spin of the cavity polaritons for simplicity, as our main goal here is to account for dissipative dynamics coming from interaction with phonons, and spin degree of freedom is not expected to introduce any qualitatively new physics from this point of view. However, that introduction of spin into the model is straightforward and corresponding work is currently underway. It should also be noted that our model is to some extent similar to one proposed in Ref. . Differently from that paper, however, we assume that all decoherence in the system comes from the interaction with acoustic phonons which is neglected in Ref.  and thus do not perform the separation of the polariton ensemble into coherent low energy part and incoherent high- energy reservoir. It is convenient to go to reciprocal space, making a Fourier transform of the one-particle density matrix, $$\begin{aligned} \rho(\textbf{k},\textbf{k}',t)&=(2\pi)^d/L^d\int e^{i(\textbf{kr}-\textbf{k}'\textbf{r}')}\rho(\textbf{r},\textbf{r}',t)d\textbf{r}d\textbf{r}'=\\ \nonumber&=Tr\left\{a_{\textbf{k}}^+a_\textbf{k+q}\right\}\equiv\langle a_{\textbf{k}}^+a_{\textbf{k}'}\rangle\end{aligned}$$ where $d$ is the dimensionality of the system ($d=2$ for non-confined polaritons, $d=1$ for the polariton channel), L is its linear size, $a_{\textbf{k}}^+$, $a_{\textbf{k}}$ are creation and annihilation operators of the polaritons with momentum **k**. Note, that we have chosen the prefactor in a Fourier transform in such a way, that the values of $\rho(\textbf{k},\textbf{k}',t)$ are dimensionless, and diagonal matrix elements give occupation numbers of the states in discretized reciprocal space. Knowing the density matrix in reciprocal space, we can find the density matrix in real space straightforwardly applying inverse Fourier transform. The total Hamiltonian of the system can be represented as a sum of two parts, $$H=H_1+H_2$$ where the first term $H_1$ describes the “coherent” part of the evolution, corresponding to free polariton propagation and polariton-polariton interactions and the second term $H_{phon}$ corresponds to the dissipative interaction with acoustic phonons. The two terms affect the polariton density matrix in a qualitatively different way. A. Polariton-polariton interactions ----------------------------------- The part of the evolution corresponding to $H_1$ is given by the following expression $$H_1=\sum_{\textbf{k}} E_\textbf{k} a_{\textbf{k}}^+a_\textbf{k}+ \frac{U}{2}\sum_{\textbf{k}_1,\textbf{k}_2,\textbf{p}}a_{\textbf{k}_1}^+a_{\textbf{k}_2}^+a_{\textbf{k}_1+\textbf{p}}a_{\textbf{k}_2-\textbf{p}}$$ where $E_\textbf k$ is the energy dispersion of the polaritons, $U$ is the matrix element of the polariton-polariton interactions. In the current paper we neglect the **p**-dependence of the polariton- polariton interaction constant coming from Hopfield coefficients [@polpol]. We do this approximation because the goal of the manuscript is to present a novel formalism for description of the relaxation effects and not detailed modeling of a particular experiment, and we want to keep our formalism as simple as possible. Besides, this approximation is widely used in current description of polaritonic systems based on modifications of the Gross-Pitaevskii equations [@Wouters2007; @BerloffVortex; @Tim1D]. However, for modeling of realistic experiments the **p**-dependence of the interaction constant can easily be introduced into the equations. The effect of $H_1$ on the evolution of the density matrix is described by the Liouville-von Neumann equation, $$i\left(\partial_t\rho\right)^{(1)}=\left[H_1;\rho\right] \label{liouville}$$ which after use of mean field approximation leads to the following dynamic equations for the elements of the single-particle density matrix in reciprocal space (see Appendix I for details of the derivation): $$\begin{aligned} &\left\{\partial_t\rho(\textbf{k},\textbf{k})\right\}^{(1)}= 2U\sum_{\textbf{k}_1,\textbf{p}}\textrm{Im}\left\{\rho(\textbf{k}_1,\textbf{k}_1-\textbf{p})\rho(\textbf{k},\textbf{k}+\textbf{p})\right\} \label{EqOccupancyPolPol}\\ &\left\{\partial_t\rho(\textbf{k},\textbf{k}')\right\}^{(1)} =i(E_\textbf{k}-E_\textbf{k}')\rho(\textbf{k},\textbf{k}')+ \label{EqCoherencePolPol} \\ \nonumber &+iU\sum_{\textbf{k}_1,\textbf{p}}\rho(\textbf{k}_1,\textbf{k}_1-\textbf{p})\left[\rho(\textbf{k}-\textbf{p},\textbf{k}')-\rho(\textbf{k},\textbf{k}'+\textbf{p})\right]\end{aligned}$$ These expressions represent an analog of the Gross-Pitaevskii equation written for the density matrix. B. Scattering with acoustic phonons ----------------------------------- Polariton-phonon scattering corresponds to the interaction of the quantum polariton system with a classical phonon reservoir. It is of dissipative nature, and thus straightforward application of the Liouville-von Neumann equation is impossible. Introduction of dissipation into quantum systems is an old problem, for which there is no single well established solution. In the domain of quantum optics, however, there are standard methods based on the Master Equation techniques [@Carmichael]. In the following we give a brief outline of this approach applied to a dissipative polariton system. The Hamiltonian of interaction of polaritons with acoustic phonons in Dirac representation can be represented as $$\begin{aligned} &\widehat{H}_2(t)=H^-(t)+H^+(t)=\\ \nonumber&=\sum_{\textbf{k},\textbf{q}}D(\textbf{q})e^{i(E_{\textbf{k}+\textbf{q}}-E_{\textbf{k}})t}a_{\textbf{k}+\textbf{q}}^+a_\textbf{k}(b_\textbf{q}e^{-i\omega_{\textbf{q}}t}+b_{-\textbf{q}}^+e^{i\omega_{\textbf{q}}t})\end{aligned}$$ where $a_\textbf{k}$ are operators for polaritons, $b_\textbf{q}$ operators for phonons, $E_\textbf{k}$ and $\omega_\textbf{q}$ are dispersion relations for polaritons and acoustic phonons respectively, $D(\textbf{q})$ is the polariton-phonon coupling constant. In the last equality we separated the terms $H^+$ where a phonon is created, containing the operators $b^+$, from the terms $H^-$ in which it is destroyed, containing operators $b$. Now, one can consider a hypothetical situation when polariton-polariton interactions are absent, and all redistribution of the polaritons in reciprocal space is due to the scattering with a thermal reservoir of acoustic phonons. One can rewrite the Liouville-von Neumann equation in an integro-differential form and apply the so called Markovian approximation, corresponding to the situation of fast phase memory loss (see Ref.  for the details and discussion of limits of validity of the approximation) $$\begin{aligned} \left(\partial_t\rho\right)^{(2)}&=-\int_{-\infty}^t dt'\left[H_{2}(t); \left[H_{2}(t');\rho(t)\right]\right]= \label{Liouville_int}\\ \nonumber &= \delta_{\Delta E}\left[2\left(H^+\rho H^-+H^-\rho\right. H^+\right)- \nonumber \\ &\left.-\left(H^+H^-+H^-H^+\right)\rho-\rho\left(H^+H^-+H^-H^+\right)\right] \nonumber\end{aligned}$$ where the coefficient $\delta_{\Delta E}$ denotes energy conservation and has dimensionality of inverse energy and in the calculation taken to be equal to the broadening of the polariton state [@KavokinMalpuech]. For time evolution of the mean value of any arbitrary operator $ \langle \widehat{A}\rangle=Tr(\rho \widehat{A})$ due to scattering with phonons one thus has (derivation of this formula is represented in Appendix II): $$\label{eqM} \left\{\partial_t\langle \widehat{A}\rangle\right\}^{(2)}=\delta_{\Delta E}\left(\langle[H^-;[\widehat{A};H^+]]\rangle+\langle[H^+;[\widehat{A};H^-]]\rangle\right).$$ Putting $\widehat{A}=a_\textbf{k}^+a_{\textbf{k}'}$ in this equation we get the contributions to the dynamic equations for the elements of the single-particle density matrix coming from polariton-phonon interaction (see Appendix III): $$\begin{aligned} \label{BoltzmannOcc} \left\{\partial_t\rho(\textbf{k},\textbf{k})\right\}^{(2)}&=\\ \nonumber \sum_{\textbf{q}',E_\textbf{k}<E_{\textbf{k+q}'}}&2W(\textbf{q}')\left[\rho(\textbf{k+q}',\textbf{k+q}')(n_{\textbf{q}'}^{ph}+1)(\rho(\textbf{k},\textbf{k})+1) \right.\\ \nonumber &\left.-\rho(\textbf{k},\textbf{k})n_{\textbf{q}'}^{ph}(\rho(\textbf{k+q}',\textbf{k+q}')+1)\right]+ \\ \nonumber +\sum_{\textbf{q}',E_\textbf{k}>E_{\textbf{k+q}'}}&2W(\textbf{q}')\left[\rho(\textbf{k+q}',\textbf{k+q}')n_{-\textbf{q}'}^{ph}(\rho(\textbf{k},\textbf{k})+1)-\right. \\ \nonumber &\left.-\rho(\textbf{k},\textbf{k})(n_{-\textbf{q}'}^{ph}+1)(\rho(\textbf{k+q}',\textbf{k+q}')+1)\right]\end{aligned}$$ and $$\begin{aligned} \label{BoltzmannCorr} &\left\{\partial_t\rho(\textbf{k},\textbf{k}')\right\}^{(2)}=\rho(\textbf{k},\textbf{k}') \times \\\nonumber &\times \left\{\sum_{\textbf{q}',E_\textbf{k}<E_{\textbf{k+q}'}}W(\textbf{q}')\left[\rho(\textbf{k+q}',\textbf{k+q}')-n_{\textbf{q}'}^{ph}\right]+ \right. \\\nonumber +&\sum_{\textbf{q}',E_\textbf{k}>E_{\textbf{k+q}'}}W(\textbf{q}')\left[-\rho(\textbf{k+q}',\textbf{k+q}')-n_{\textbf{q}'}^{ph}-1\right]+\\ \nonumber +&\sum_{\textbf{q}',E_\textbf{k}'<E_{\textbf{k}'+\textbf{q}'}}W(\textbf{q}')\left[\rho(\textbf{k}'+\textbf{q}',\textbf{k}'+\textbf{q}')-n_{\textbf{q}'}^{ph}\right] +\\\nonumber +&\left.\sum_{\textbf{q}',E_\textbf{k}'>E_{\textbf{k}'+\textbf{q}'}}W(\textbf{q}')\left[-\rho(\textbf{k}'+\textbf{q}',\textbf{k}'+\textbf{q}')-n_{\textbf{q}'}^{ph}-1\right]\right\}\end{aligned}$$ where the transition rates are given by $W(\textbf{q})=D^2(\textbf{q})\delta_{\Delta E_\textbf{q}}$ and $n_\textbf{q}^{ph}$ denote the occupation numbers of phonons with wavevector $\textbf{q}$ given by Bose distribution. Eq.  is nothing but a standard Boltzmann equation for the phonon-assisted polariton relaxation, while Eq.  describes the decay of the off-diagonal matrix elements of a single-particle density matrix due to interaction with classical phonon reservoir. Together these equations thus describe thermalization of the polariton system. To account for the effects of free polariton propagation, polariton-polariton and polariton-phonon interactions one should combine together expressions , , and . After finding the single-particle density matrix in reciprocal space by solving the corresponding dynamic equations, one can determine the dynamic of the system in real space simply performing a Fourier transformation by **k** and **k**’ variables. III. Results and discussion =========================== Our formalism is suitable for the description of both 2D polaritons and polaritons confined within 1D channels. In this paper we present the results of numerical modeling for the latter case only, as solving of the dynamic equations for 2D case requires the use of supercomputing facilities. ![Time evolution of polariton distribution in $k$-space. The vertical plane denotes $k=0$ state. a) Only polariton-phonon scattering is accounted for. The relaxation towards $k=0$ state is blocked due to the bottleneck effect. b) Both polariton-phonon and polariton-polariton scattering are accounted for. The maximum of the polariton concentration is developed at $k=0$, signifying the overcoming of the bottleneck effect. Another maximum appears at higher $k$, due to the energy-conserving nature of the polariton-polariton interactions. []{data-label="fig:1"}](Fig1.eps){width="0.85\linewidth"} The results of modeling are shown on Figures 1-3. We consider a $2$ $\mu$m wide polariton channel in GaAs microcavity with Rabi splitting 15 meV at temperature $T=1$ K. The polaritons are created by a short coherent localized laser pulse. We account for the finite lifetime of cavity polaritons $\tau=5$ ps adding the term $-\rho(\textbf{k},\textbf{k}')/\tau$ into the dynamic equations. Figure 1 shows the dynamics of the polariton system in reciprocal space and demonstrates the roles of the polariton-phonon and polariton-polariton interactions. If only the former are included, the system demonstrates the bottleneck effect shown on Fig. 1a. Due to the energy relaxation coming from the interactions with phonons, polaritons have a tendency to move towards the ground state in $k$-space. However, this process is dramatically slowed down in the inflection region of the polariton dispersion, where polariton-phonon interaction becomes inefficient. Consequently, there is no remarkable increase of the population of $k=0$ state [@Bottleneck]. The bottleneck effect can be overcome by the polariton-polariton interactions, as shown on Fig. 1b. One sees that in this case the particles accumulate quickly in $k=0$ state. At the same time, the second maximum of the polariton distibution appears at higher $k$ due to the energy conservative nature of the polariton-polariton interactions (analogically to the formation of the idler mode in polariton parametric oscillator [@PPO]). {width="0.9\linewidth"} Our results for the dynamics of the polariton distribution in reciprocal space are in good agreement with those obtained by using Boltzmann equations. In addition, our approach allows consideration of the dynamics of the dissipative polariton system in real space. This is illustrated on Figures 2 and 3. Figure 2 shows the effect of the various types of interactions on the real space dynamics of the localized polariton wavepackage. We compare the cases of the ballistic propagation with those where only polariton-phonon interactions are included, only polariton-polariton interactions are included and both polariton-polariton and polariton-phonon interactions are included. As one sees, the dynamics are very different for these four cases. Polariton-polariton interactions lead to splitting of the wavepackage into two soliton- like peaks, which is in good qualitative agreement with the results given by Gross-Pitaevskii equation [@ShelykhGP]. On the other hand, polariton-phonon interactions lead to damping of the package, contributing to the recovering of the homogenious distribution of the polaritons in real space as it is expected from the classical diffusion equation. {width="0.9\linewidth"} Naturally, the effect of the phonon damping strongly depends on temperature, as shown on Fig. 3. One sees, that at low temperatures the propagating package is split into two due to the polariton-polariton interactions. Increasing the temperature smoothes down the polariton distibution and at $T=20K$ one has just a single peak. IV. Conclusion ============== In conclusion, we developed a formalism for the description of the dissipative dynamics of an inhomogeneous polariton system in real space and time accounting for polariton-polariton interactions and polariton-phonon scattering. The formalism was applied for numerical modeling of the propagation of a polariton droplet in a 1D channel. Our approach can also be used for modeling the dissipative dynamics of other bosonic systems (e.g. indirect excitons). Acknowledgements ================ We thank G. Malpuech, D.D. Solnyshkov and T.C.H. Liew for useful discussions. The work was supported by Rannis “Center of excellence in polaritonics” and FP7 IRSES projects “SPINMET” and “POLAPHEN”. Appendix I: derivation of kinetic equations for polariton- polariton scattering =============================================================================== The evaluation of the dynamic equations of $\langle a_\textbf k ^\dagger a_\textbf k\rangle$ and $\langle a_\textbf k ^\dagger a_\textbf {k+q}\rangle$ using Eq.  and applying the mean field approximation gives: $$\begin{aligned} \textrm{Tr}\left\{a_\textbf{k}^+a_{\textbf{k+q}}\left[\rho;H_{p-p}\right]\right\}&= U\sum_{\textbf{k}_1,\textbf{k}_2,\textbf{p}}\rho(\textbf{k}_2,\textbf{k}_2-\textbf{p})\textrm{Tr}\left\{a_\textbf{k}^+a_{\textbf{k+q}}\left[\rho;a_{\textbf{k}_1}^+a_{\textbf{k}_1+\textbf{p}}\right]\right\}=\\ &=U\sum_{\textbf{k}_2,\textbf{p}}\rho(\textbf{k}_2,\textbf{k}_2-\textbf{p})\left[-\rho(\textbf{k},\textbf{k+q+p})+\rho(\textbf{k-p},\textbf{k+q})\right] \nonumber\end{aligned}$$ $$\begin{aligned} \textrm{Tr}\left\{a_\textbf{k}^+a_{\textbf{k}}\left[\rho;H_{p-p}\right]\right\}= U\sum_{\textbf{k}_1,\textbf{k}_2,\textbf{p}}\rho(\textbf{k}_2,\textbf{k}_2-\textbf{p})\textrm{Tr}\left\{a_\textbf{k}^+a_{\textbf{k}}\left[\rho;a_{\textbf{k}_1}^+a_{\textbf{k}_1+\textbf{p}}\right]\right\}=\\ \nonumber =U\sum_{\textbf{k}_1,\textbf{k}_2,\textbf{p}}\rho(\textbf{k}_2,\textbf{k}_2-\textbf{p})\textrm{Tr}\left\{\rho \left[a_{\textbf{k}_1}^+a_{\textbf{k}_1+\textbf{p}}a_\textbf{k}^+a_{\textbf{k}}- a_\textbf{k}^+a_{\textbf{k}}a_{\textbf{k}_1}^+a_{\textbf{k}_1+\textbf{p}}\right]\right\}=-2iU\sum_{\textbf{k}_2,\textbf{p}}\textrm{Im}\left\{\rho(\textbf{k}_2,\textbf{k}_2-\textbf{p})\rho(\textbf{k},\textbf{k+p})\right\} \nonumber\end{aligned}$$ From this one straightforwardly gets Eqs. \[EqOccupancyPolPol\], \[EqCoherencePolPol\]. Appendix II: Derivation of expression for dynamics of mean values in Born- Markov approximation =============================================================================================== Now, consider the evolution of a mean value of any arbitrarty operator $A, \langle A\rangle=Tr(\rho A)$ (energy conserving delta-function omitted) if dynamics of a density matrix is given by Eq. \[Liouville\_int\]: $$\begin{aligned} \partial_t\langle A\rangle&=\texttt{Tr}\left(2H^+\rho H^-A+2H^-\rho H^+A-\left(H^+H^-+H^-H^+\right)\rho A-\rho\left(H^+H^-+H^-H^+\right)A\right)=\\ \nonumber &=\texttt{Tr}\left[\rho\left(2H^-AH^++2H^+AH^--A\left(H^+H^-+H^-H^+\right)-\left(H^+H^-+H^-H^+\right)A\right)\right]\end{aligned}$$ where we used the property of the invariance of the trace as regards cyclic permutations of operators. The latter expression can be simplified as $$\begin{aligned} &2H^-AH^++2H^+AH^--A\left(H^+H^-+H^-H^+\right)-\left(H^+H^-+H^-H^+\right)A=\\ \nonumber={}&2H^-([A;H^+]+H^+A)+2H^+([A;H^-]+H^-A)-A\left(H^+H^-+H^-H^+\right)-\left(H^+H^-+H^-H^+\right)A=\\ \nonumber={}&2H^-[A;H^+]+2H^+[A;H^-]-H^+[A;H^-]-[A;H^+]H^--H^-[A;H^+]-[A;+H^-]H^+=\\ \nonumber={}&[H^-;[A;H^+]]+[H^+;[A;H^-]]\end{aligned}$$ where we used the following property of commutators: $[A;BC]=B[A;C]+[A;B]C$. Thus $$\partial_t\langle A\rangle=\texttt{Tr}\left(\rho[H^-;[A;H^+]]\right)+\texttt{Tr}\left(\rho[H^+;[A;H^-]]\right)$$ For the important case of the Hermitian operator A corresponding to a physical observable one has $$\begin{aligned} &[H^-;[A;H^+]]^+=-[[A;H^+]^+;H^+]=\\ \nonumber ={}&[[H^-;A^+];H^+]=[H^+;[A;H^-]]\end{aligned}$$ and $$\partial_t\langle A\rangle=2\texttt{Re}\{\texttt{Tr}\left(\rho[H^-;[A;H^+]]\right)\}$$ This formula can be applied for calculation of occupation numbers. Appendix III: Derivation of dynamic equations with acoustic phonons =================================================================== To get explicit expressions for dynamics of $\rho(\textbf{k},\textbf{k},t)$ let us consider a simple case when only states **k**, **k**+**q** are present. We have two cases: a\) $E(\textbf k+\textbf q)>E(\textbf k)$. In this case, leaving energy- conserving terms only one gets $$\begin{aligned} H^+&=Da_\textbf k^+a_{\textbf k+\textbf q}b_\textbf q^+\\ H^-&=Da_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q \end{aligned}$$\ The application of Eq.\[eqM\] gives the following results: $$\begin{aligned} &\delta^{-1}\partial_t\rho(\textbf k,\textbf k)=\delta^{-1}\partial_tn_\textbf k=\texttt{Tr}\left(\rho[H^-;[a_\textbf k^+a_\textbf k;H^+]]\right)+\\ \nonumber&\texttt{Tr}\left(\rho[H^+;[a_\textbf k^+a_\textbf k;H^-]]\right)=2\texttt{Tr}\left(\rho[H^+;[a_\textbf k^+a_\textbf k;H^-]]\right)\end{aligned}$$ After some straightforward algebra one gets $$\begin{aligned} [a_\textbf k^+a_\textbf k;H^-]=D[a_\textbf k^+a_\textbf k;a_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q]=Da_{\textbf k+\textbf q}^+b_\textbf q[a_\textbf k^+a_\textbf k;a_\textbf k]=-Da_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q,\\ \nonumber [H^+;[a_\textbf k^+a_\textbf k;H^-]]=-D^2[a_\textbf k^+a_{\textbf k+\textbf q}b^+_\textbf q;a_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q]=\\ \nonumber=D^2\left(a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}(a_\textbf k^+a_\textbf k+1)(b_\textbf q^+b_\textbf q+1)-(a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}+1)a_\textbf k^+a_\textbf kb^+_\textbf qb_\textbf q\right)\end{aligned}$$ and $$\partial_t\rho(\textbf k,\textbf k)=2\delta D^2\left[\rho(\textbf k+\textbf q,\textbf k+\textbf q)(n_\textbf q^{ph}+1)(\rho(\textbf k,\textbf k)+1)-\rho(\textbf k,\textbf k)n_\textbf q^{ph}(\rho(\textbf k+\textbf q,\textbf k+\textbf q)+1)\right]$$ This is nothing but ordinary Boltzmann equation, accounting for transition from state $\textbf{k+q}$ to state $\textbf{k}$ accompanied by the emission of the phonon (spontaneous or stimulated) and transition from state $\textbf{k}$ to state $\textbf{k+q}$ due to the phonon absorption.\ b) $E(\textbf k+\textbf q)<E(\textbf k)$. Treating this case in a similar way we find: $$\partial_t\rho(\textbf k,\textbf k)=2\delta D^2\left[\rho(\textbf k+\textbf q,\textbf k+\textbf q)n_{-\textbf q}^{ph}(\rho(\textbf k,\textbf k)+1)-(\rho(\textbf k+\textbf q,\textbf k+\textbf q)+1)(n_{-\textbf q}^{ph}+1)\rho(\textbf k,\textbf k)\right]$$ Again we obtained Boltzmann equation, but differently to the previous case the transition from state $\textbf{k+q}$ to state $\textbf{k}$ goes with absorption of the phonon and from state $\textbf{k}$ to state $\textbf{k+q}$ with phonon emission. Performing summation over all reciprocal space one gets Eq. . Now let us consider the dynamics of the off-diagonal part of the density matrix $\rho(\textbf{k},\textbf{k}',t)$. $$\begin{aligned} &\delta^{-1}\partial_t\rho(\textbf k,\textbf k')=\label{offdiag_phon}\\ \nonumber={}&\texttt{Tr}\left(\rho[H^+;[a_\textbf k^+a_{\textbf k'};H^-]]\right)+\texttt{Tr}\left(\rho[H^-;[a_\textbf k^+a_{\textbf k'};H^+]]\right)\end{aligned}$$ Here we should consider different orderings of the energies corresponding to the states $\textbf k$, $\textbf k'$ and state $\textbf k+\textbf q$. As an example, let us consider the case when $E_\textbf k<E_{\textbf k+\textbf q},\ E_{\textbf k'}<E_{\textbf k+\textbf q}$. Leaving energy-conserving terms only one gets $$\begin{aligned} H^+&=Da_\textbf k^+a_{\textbf k+\textbf q}b_\textbf q^++a_{\textbf k'}^+a_{\textbf k+\textbf q}b_{\textbf q'}^+\\ H^-&=Da_\textbf ka_{\textbf k+\textbf q}^+b_\textbf q+a_{\textbf k'}a_{\textbf k+\textbf q}^+b_{\textbf q'}\end{aligned}$$ where $\textbf q'=\textbf k+\textbf q-\textbf k'$. In this case the first term in Eq.  gives: $$\begin{aligned} [a_{\textbf k}^+a_{{\textbf k}'};H^-]&=D[a_{\textbf k}^+a_{{\textbf k}'};a_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q+a_{\textbf k'}a_{\textbf k+\textbf q}^+b_{\textbf q'}]=-Db_{\textbf q'}a_{\textbf k'}a_{\textbf k+\textbf q}^+\\\ \nonumber [H^+;[a_\textbf k^+a_{\textbf k'};H^-]]&=D^2[a_{\textbf k+\textbf q}a_\textbf k^+b^+_\textbf q+a_{\textbf k+\textbf q}a_{\textbf k'}^+b^+_{\textbf q'};-b_\textbf qa_{\textbf k+\textbf q}^+a_{\textbf k'}]=\\ \nonumber &=D^2\left\{a_\textbf k^+a_{\textbf k'}(a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}+1)b_\textbf q^+b_\textbf q-(b_\textbf q^+b_\textbf q+1)a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}a_\textbf k^+a_\textbf k\right\}\\ \nonumber\end{aligned}$$ and $$\begin{aligned} [a_{\textbf k}^+a_{{\textbf k}'};H^+]&=D[a_{\textbf k}^+a_{{\textbf k}'};a_{\textbf k+\textbf q}a_\textbf k^+b_\textbf q^++a_{\textbf k'}^+a_{\textbf k+\textbf q}b_{\textbf q'}^+]=Db_{\textbf q'}^+a_{\textbf k}^+a_{\textbf k+\textbf q}\\ \nonumber [H^-;[a_\textbf k^+a_{\textbf k'};H^+]]&=D^2[a_{\textbf k+\textbf q}^+a_\textbf kb_\textbf q+a_{\textbf k+\textbf q}^+a_{\textbf k'}b_{\textbf q'};b_{\textbf q'}^+a_{\textbf k}^+a_{\textbf k+\textbf q}]=\\ \nonumber &=D^2\left\{a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}a_\textbf k^+a_{\textbf k'}(b_{\textbf q'}^+b_{\textbf q'}+1) - (a_{\textbf k+\textbf q}^+a_{\textbf k+\textbf q}=1)a_\textbf k^+a_{\textbf k'}b_{\textbf q'}^+b_{\textbf q'}\right\}\\ \nonumber\end{aligned}$$ Finally for the case $E_\textbf k<E_{\textbf k+\textbf q},\ E_{\textbf k'}<E_{\textbf k+\textbf q}$ one obtains equation in the form $$\begin{aligned} \partial_t\rho(\textbf k,\textbf k')={}&\delta D^2\left(\rho(\textbf k+\textbf q,\textbf k+\textbf q)(n_\textbf q^{ph}+1)-(\rho(\textbf k+\textbf q,\textbf k+\textbf q)+1)n_\textbf q^{ph}\right)\rho(\textbf k,\textbf k')+\\ & \left(\rho(\textbf k+\textbf q,\textbf k+\textbf q)(n_{\textbf q'}^{ph}+1)-(\rho(\textbf k+\textbf q,\textbf k+\textbf q)+1)n_{\textbf q'}^{ph}\right)\rho(\textbf k,\textbf k')\end{aligned}$$ The same procedure is applied for all other cases. Performing again summation over all reciprocal space one gets Eq. \[BoltzmannCorr\]. [99]{} J. Kasprzak, M. Richard, S. Kundermann, A. Baas, P. Jeambrun, J. M. J. Keeling, F. M. Marchetti, M. H. Szy- manska, R. Andre, J. L. Staehli, V. Savona, P. B. Lit- tlewood, B. Deveaud and Le Si Dang, Nature **443**, 409 (2006). A. Amo, D. Sanvitto, F. P. Laussy, D. Ballarini, E. del Valle, M. D. Martin, A. Lemaitre, J. Bloch, D. N. Krizhanovskii, M. S. Skolnick, C. Tejedor and L. Vina, Nature 457, 291 (2009) I.A. Shelykh, D. D. Solnyshkov, G. Pavlovic, and G. Malpuech, Phys. Rev. B **78**, 041302 (2008) K. G. Lagoudakis, B. Pietka, M. Wouters, R. Andre and B. Deveaud- Pledran, Phys. Rev. Lett. 105, 120403 (2010) F.P. Laussy, A.V. Kavokin, I.A. Shelykh, Phys. Rev. Lett. **104**, 106402 (2010) A. Imamoglu and J. R. Ram, Phys. Lett. A **214**, 193, (1996) A. T. Hammack, M. Griswold, L. V. Butov, L. E. Smallwood, A. L. Ivanov, and A. C. Gossard, Phys. Rev. Lett. **96**, 227402 (2006); R. B. Balili, D. W. Snoke, L. Pfeiffer, and K. West, Appl. Phys. Lett. **88**, 031110. (2006); O. El Daif, A. Baas, T. Guillet, J.-P. Brantut, R. Idrissi Kaitouni, J. L. Staehli1, F. Morier-Genoud, and B. Deveaud, Appl. Phys. Lett. **88**, 061105 (2006); R. I. Kaitouni, O. El Daif, A. Baas, M. Richard, T. Paraiso, P. Lugan, T. Guillet, F. Morier-Genoud, J. D. Ganiere, J. L. Staehli, V. Savona, and B. Deveaud, Phys. Rev. B **74**, 155311 (2006); M. M. Kaliteevskii, S. Brand, R. Abram, I. Iorsh, A. Kavokin, and I. Shelykh, Appl. Phys. Lett. **95**, 251108 (2009) E. Wertz, L. Ferrier, D. Solnyshkov, R. Johne, D. Sanvitto, A. Lemaitre, I. Sagnes, R. Grousson, A. V. Kavokin, P. Senellart, G. Malpuech, and J. Bloch, Nature Physics **6**, 860 (2010) T.C.H. Liew, A.V. Kavokin and I.A. Shelykh, Phys. Rev. Lett. **101**, 016402 (2008) T. C. H. Liew, A. V. Kavokin, T. Ostatnicky, M. Kaliteevski, I. A. Shelykh, and R. A. Abram Phys. Rev. B **82**, 033302 (2010) A.V. Kavokin, G. Malpuech and M. Glazov, Phys. Rev. Lett. **95**, 136601 (2005) M.M. Glazov and L.E. Golub, Phys. Rev. B **77**, 165341 (2008) I.A. Shelykh, G. Pavlovic, D. D. Solnyshkov, and G. Malpuech, Phys. Rev. Lett. **102**, 046407 (2009) F. Tassone, C. Piermarocchi, V. Savona, A. Quattropani and P. Schwendimann, Phys. Rev. B **56**, 7554 (1997) F. Tassone and Y. Yamamoto, Phys. Rev. B **59**, 10830 (1999) I. Carusotto and C. Ciuti, Phys. Rev. Lett. **93**, 166401 (2004). I.A. Shelykh, Yuri G. Rubo, G. Malpuech, D. D. Solnyshkov, and A. Kavokin, Phys. Rev. Lett. **97**, 066402 (2006) D. Porras, C. Ciuti, J. J. Baumberg, and C. Tejedor, Phys. Rev. B **66**, 085304 (2002) J. Kasprzak, D. D. Solnyshkov, R. Andre, Le Si Dang, and G. Malpuech, Phys. Rev. Lett. **101**, 146404 (2008) T.D. Doan, H.T. Cao, D.B.Tran Thoai and H. Haug, Phys. Rev. B **72**, 085301 (2005) H.T. Cao, T. D. Doan, D. B. Tran Thoai, and H. Haug, Phys. Rev. B **77**, 075320 (2008) M. Wouters and I. Carusotto, Phys. Rev. Lett. **99**, 140402 (2007). M. O. Borgh, J. Keeling, and N. G. Berloff, Phys. Rev. B **81**, 235302 (2010) M. Wouters, T. C. H. Liew, and V. Savona, Phys. Rev. B **82**, 245315 (2010) L. V. Butov, *J. Phys.: Condens. Matter* **19**, 295202 (2007). E.B. Magnusson, H. Flayac, G. Malpuech and I.A. Shelykh, Phys. Rev. B **82**, 195312 (2010) M. Wouters and V. Savona, Phys. Rev. B **79**, 165302 (2009) H. Carmichael, Quantum Optics 1: Master Equations And Fokker-Planck Equations, Springer, New York, 2007. A. Kavokin and G. Malpuech, Cavity Polaritons (Elsevier Academic Press, Amsterdam, 2003). P.G. Savvidis, J. J. Baumberg, R. M. Stevenson, M. S. Skolnick, D. M. Whittaker, and J. S. Roberts, Phys. Rev. Lett. **84**, 1547 (2000)

--- abstract: 'A comprehensive theory for robust PID control in continuous-time and discrete-time domain is reviewed in this paper. For a given finite set of linear time invariant plants, algorithms for fast computation of robustly stabilizing regions in the ($k_P,~k_I,~k_D$)-parameter space are introduced. The main impetus is given by the fact that non-convex stable regions in the PID parameter space can be built up by convex polygonal slices. A simple and an elegant theory evolved in the last few years up to a quite mature level.' author: - 'Naim Bajcinca[^1]' title: '\' --- INTRODUCTION {#sec:Introduction} ============ It is a well-known fact that by far the most applied control law for SISO systems in nearly all industries (process control, motion control, aerospace etc) is the PID control. No other controller can compete to PID when it comes to performance per simplicity of the control structure, this being the reason for its absolute dominance in the practice. Traditionally, tuning of a PID controller has been the overwhelming usability approach in research and applications. The design technique presented here is in some sense an opposing one. We want to compute the total set of PID-stabilizers. While it turns out to be interesting in theoretical terms, its impact on practical applications is difficult to predict. Advanced software tools based on this technology (e.g. [<span style="font-variant:small-caps;">Robsin</span>]{}, [@bajc:cca2004]) suggest to the user a 3D-region in ($k_P,~k_I,~k_D$)-parameter space, where he can select a controller from. By doing so, he would additionally have an idea how robust (i.e. how far from instability) his design is. A further good news is that the same technique applies for time-delay systems and in discrete-time domain. This paper covers theoretical design issues only. It has been shown that the stabilizing region for continuous-time PID ($=k_P+k_I/s+k_D s$) controllers is defined by a set of convex polygonal slices normal to $k_P$ axis in the $(k_P,~k_I,~k_D)$ parameter space, see [@bhatta:book2000], [@kaes:ecc2001], [@bajc:auto:2006]. The method followed by the author uses the decoupling effect of PID parameter space at singular frequencies, [@bajc:auto:2006]. Thereby the characteristic polynomial decouples into two frequency parameterized equations, one involving $k_I$ and $k_D$, and another one with $k_P$ only. As a consequence, the computation of all stable PID controllers may be divided into two subproblems: assertion of stable intervals of parameter $k_P$ (so-called $k_P$-problem), and detection of stable polygons on the plane ($k_I,~k_D$) for a given $k_P$. In [@bajc:ifac2002] and [@bajc:ecc07] the design approach was transferred to discrete-time systems, and in [@bajc:tds2004] to time-delay systems. A general algorithm for automatic detection of stable polygons is based on the analysis of the motion of eigenvalues in the vicinity of the singular frequencies. This algorithm was first presented in [@bajc:med2001]. It detects the inner polygons and selects the one with maximal stable eigenvalues, which is finally checked for stability. It is important to emphasize, that this algorithm can be applied equally well for PID quasi-polynomials describing the PID control loops for time-delay systems. Preserving simplicity has been a basic motivation in searching for a solution to the $k_P$-problem. A simple necessary condition was firstly understood in [@bajc:acc2005]. It turns out that for a given plant a sufficient number of singular frequencies must be available for its stabilization. Since $k_P$ uniquely determines the number of singular frequencies, one can directly read $k_P$-intervals from the $k_P$-plot, which may host stable polygons. This simple criterion can be extended to discrete-time domain and time-delay quasi-polynomials, for instance, see [@bajc:ecc07] and [@bajc:tds2004]. The presented methods may be directly applied for the computation of the total robust set of PID controllers which stabilize a finite set of plants (multi-model). This is exactly what is meant by robust design of PID controllers. In this paper methods for PID control in continuous- and discrete-time domain, as well as for time-delay systems, are reviewed. Therefore, we will have to switch between the $s$ domain (continuous-time) and $z$-domain (discrete-time) while presenting results. The paper is organized as follows. In Section \[sec:problemDefinition\] we postulate the stabilizability problem. Section \[sec:veryBasicIdea\] presents all steps of the design process, and simultaneously introduces the formulates technical problems, in a simple-case study. Theoretical fundamentals of the methods are introduced in Section \[sec:basicDefinitions\]. Automatic detection of convex stable polygons and $k-P$-problems are discussed in Section \[sec:stablePolygons\] and \[sec:kp-Porblem\], respectively. Finally, in Section \[sec:TimeDelaySystems\] the ideas are extended for time-delay systems. For illustration purposes, throughout the paper examples are provided. The reader may follow them by using the toolbox [<span style="font-variant:small-caps;">Robsin</span>]{}, which can be downloaded from *http://www.robotic.dlr.de/robsin*. This article primarily reviews the original work of the author. It has not been an intention to refer to the all work in the area. Still, the key contributions (and contributors!) to the theory are referred to and they read as follows. The role of convex polygons was firstly understood by Battacharyya and his co-workers, see the monograph [@bhatta:book2000]. Their derivation bases on Hermite-Biehler theorem. For computation of stable polygons they proposed linear programming techniques, but they did not really address the $k_P$-problem. Munro’s computation is based on the real-axis intersections of the Nyquist plot, see, [@munro:2000], [@munro:2003]. The relationship to singular frequencies was noticed firstly by Ackermann and Kaesbauer, see [@kaes:ecc2001]. Soylemez proposed a solution to the $k_P$-problem, however in author’s opinion his criterion is not as simple and usable as the one presented in this article. The remainder theory reviewed here is fairly developed by the author, including the algorithms for automatic detection of stable polygons and solutions to the $k_P$-problem, as well as generalizations to discrete-time domain and time-delay systems. Problem definition {#sec:problemDefinition} ================== Discrete-time domain --------------------  Consider a simple closed curve $\partial\Gamma~=\{z\mid z=\tau(\alpha)+j~\eta(\alpha),~\alpha~\textrm{is a parameter}\}$, in $z-$plane, which is symmetric to the real axis $\tau$ and can be expressed in the form $$\label{eq:gammaeq} \partial\Gamma:~~F(\tau, \eta)=0.$$ Let $$\label{eq:basicEquation} p=A(z)Q(z,r_1,r_2,r_3)+B(z)=0$$ be a three-term algebraic equation in ($r_1,~r_2,~r_3$), with given polynomials $A$ and $B$ $$\begin{aligned} \label{eq:polynA} A(z)&=&a_0+a_1 z+\cdots+a_m z^m\\ \label{eq:polynB} B(z)&=&b_0+b_1 z+\cdots+b_{n} z^{n},\end{aligned}$$ and the second-order polynomial $Q$ in the form $$\label{eq:QEquation} Q=\delta_1(z)r_1+\delta_2(z)r_2+\delta_3(z)r_3.$$ In this article we want to compute the set of all parameters $r_1,~r_2,~r_3$, s.t. the polynomial (\[eq:basicEquation\]) is $\Gamma-$stable, that is, all its eigenvalues must lie within the $\Gamma$-region (enclosed by $\partial\Gamma$). Of main interest are circles with center on the real axis $\tau$ and an arbitrary radius, which will be referred to as $\Gamma-$circles. For discrete-time systems particularly important is the unity Schur-circle. It may be easily shown that (\[eq:basicEquation\]) describes the characteristic equation of a feedback loop with a PID or a three-term controller. Indeed, a discrete-time equivalent of the PID controller has the transfer function $$\label{eq:PID} K(z)=\frac{c_1 + c_2z + c_3z^2}{(z + z_1)(z - 1)}.$$ Its structure follows in the quasi-continuous consideration by applying the rectangular integration rule $(s \rightarrow (z -1)/Tz)$ to the ideal PID controller $k_I/s + k_P + k_Ds$, resulting in $z_1 = 0$, or by the trapezoidal integration rule $(s \rightarrow 2(z - 1)/T(z + 1))$, resulting in $z_1 = 1$. Also the realizable PID controller $k_I/s + k_P + k_Ds/(1 + T_1s)$ converts by the trapezoidal integration rule to the controller structure (\[eq:PID\]) with a pole at $z_1 = -(2T_1 - T)/(2T_1 + T)$. Equation (\[eq:basicEquation\]) covers also a three-term controller with an arbitrary second order denominator polynomial $$\label{eq:threeTerm1} K(z)=\frac{n(z)(c_1 + c_2z + c_3z^2)}{d(z)}.$$ For both control structures (\[eq:PID\]) and (\[eq:threeTerm1\]), the polynomial $Q$ is of the form $$\label{eq:SchurQ'Equationc1c2c3} Q=c_1+c_2 z +c_3 z^2.$$ It is clear that (\[eq:QEquation\]) and (\[eq:SchurQ’Equationc1c2c3\]) are connected by a linear transformation $$\label{eq:Tmatrix} c=T~r,~~\textrm{det~}T~\neq 0,$$ with $c=[c_1,c_2,c_3]^T,~~r=[r_1,r_2,r_3]^T$. The matrix $T$ is determined by the polynomials $\delta_1(z),~\delta_2(z),~\delta_3(z)$. Continuous-time domain ----------------------  The pendant equation (\[eq:basicEquation\]) for a feedback-loop with a PID controller in continuous-time domain is $$\label{eq:basicCharacteristicPolynomial} p = {A}(s)( k_I+ k_P s+ k_D s^2)+ {B}(s)$$ where polynomials $A(s)$ and $B(s)$ are as in (\[eq:polynA\]) and (\[eq:polynB\]). As in the discrete-time case, we want to compute the set of all parameters $k_P$, $k_I$ and $k_D$ for which the polynomial (\[eq:basicCharacteristicPolynomial\]) is Hurwitz-stable, i.e. all of its eigenvalues lie on the left-hand side of the imaginary axis $s=j\omega$. In other words, here $\partial\Gamma=\{j\omega:~\forall \omega\in \mathbb{R}\}$. In view of the definition (\[eq:basicEquation\]), the polynomial $Q$ is also of the form (\[eq:SchurQ’Equationc1c2c3\]) $$\label{eq:HurwitzQ} Q=k_I+k_P s +k_D s^2.$$ Obviously discrete-time consideration is more general, with $z,\alpha,r_1,r_2,r_3$ corresponding to $s,\omega,k_I,k_D,~k_P$, respectively. The very basic idea {#sec:veryBasicIdea} =================== Consider the special case with $A(s)=1$ in (\[eq:basicCharacteristicPolynomial\]). Substitution $s=j\omega$, decouples the latter into two equations $$\label{eq:basicIdea} k_I -\omega^2 k_D=-R_B, ~~k_P=-I_B/\omega$$ where $R$ and $I$ stand for the real and imaginary part of polynomial $B(s)$ at $s=j\omega$. Notice that PID parameters appear decoupled in tow equations. Computation of the stable PID region should be quite obvious for this case study. (1): First, for a fixed parameter $k_P=k'_P$ solve for the frequencies $\omega'$ from the second equation in (\[eq:basicIdea\]), representing the [$k_P$-plot]{} and shown in Fig. \[fig:kpplot\]. Such frequencies are known as *singular frequencies*. (2): Map all singular frequencies using the first equation in (\[eq:basicIdea\]) into the $(k_I,k_D)-$plane as shown in Fig. \[fig:kikdplane\], and compute the stable polygons (gray area in Fig. \[fig:basicIdea\]). Hereby, each pair of singular frequencies $\pm j\omega'$ maps to a straight boundary line. Note that, as parameters $k_I$ and $k_D$ are varied, with $k_P$ kept fixed, the eigenvalues of (\[eq:basicEquation\]) can cross over imaginary axis $j\omega$ at singular frequencies only. This procedure is repeated for other $k_P$ parameters, which yield other stable polygons. Thus a tomographic 3-D picture results, as that in Fig. \[fig:3DsingLinesExample1\]. \ \ In the sequel, we want to generalize the above algorithm and provide solutions to the problems: \[P1\] For what values of the parameter $k_P$ stable polygons should be searched for - indeed one would like to exclude *a priori* $k_P$-intervals with no stable polygons. This problem is referred to as the *$k_P$-problem*. \[P2\] For a given set of boundary lines, how to automate the computation of the stable polygons. Basic definitions and theorems {#sec:basicDefinitions} ============================== Let $H$ and $G$ represent the real and imaginary part of the characteristic polynomial $p(z,r_1,r_2,r_3)$ in (\[eq:basicEquation\]) on $\partial\Gamma$. ** \[def:SingulaeresGammaGebiet\] $\Gamma$ is said to be singular with respect to parameters $r_1$ and $r_2$ in (\[eq:basicEquation\]) if $$\label{eq:RangebdingungFuerSingulaereGammaGebiete} \textrm{Rank}\frac{\partial{}(H,G)}{\partial{}(r_1,r_2)}=1\qquad \textrm{for all } z\in{}\partial\Gamma{}.$$ The latter equation is referred to as the *rank-condition*. From now on, we only consider $\Gamma$s which satisfy (\[eq:RangebdingungFuerSingulaereGammaGebiete\]). A zero of the polynomial (\[eq:basicEquation\]) that fulfills (\[eq:RangebdingungFuerSingulaereGammaGebiete\]) is referred to as *singular (or critical) frequency*. ** \[def:decouplingFunction\] A function $E_{\Gamma}(z)$ defined as $$\label{eq:decouplingFunction} Q(z,r_1,r_2,r_3)=E_{\Gamma}(z)~q(z,r_1,r_2,r_3)$$ with $$\label{eq:Iq1} I_q=r_3 g_3(\alpha)+g_0(\alpha),~~~\alpha\in [a,~b]$$ where $I_q$ stands for the imaginary part of $q$, will be referred to as decoupling function of $Q$ on $\partial\Gamma$. In other words, by introducing $E_{\Gamma}$, a function $q$ is extracted from $Q$, whose imaginary part depends on one parameter $r_3$ only. It can be easily checked that if $\Gamma$ satisfies (\[eq:RangebdingungFuerSingulaereGammaGebiete\]), then $$\label{eq:derr1r2} \frac{\partial I_q}{\partial r_1}=0 \Leftrightarrow \frac{\partial I_q}{\partial r_2}=0,~~~\forall z\in \partial\Gamma.$$ Furthermore, two trivial decoupling functions on $\partial\Gamma$ of $Q$ in (\[eq:QEquation\]) are $$\label{eq:decouplingFuntion} E_\Gamma(z) = \delta_1(z),~~~E_\Gamma(z) = \delta_2(z).$$ Using these two facts, the next statement follows directly. ** \[thm:decoupling\]Consider the function $$\label{eq:Fz} F(z):=\frac{p(z)}{A(z)E_\Gamma(z)}.$$ The equation $F(z)=0$ for $z\in\partial\Gamma$ decouples the parameters $r_1, r_2$ and $r_3$ into two equations $$\begin{aligned} \label{eq:singFreqsLines} r_1 h_1(\alpha)+r_2 h_2(\alpha)+h_0(\alpha)&=&0,\\ \label{eq:singFreqs}r_3 g_3(\alpha)+g_0(\alpha)&=&0.\end{aligned}$$ Note that the latter equations reveal decoupling of the parameter space ($r_1,r_2,r_3$). They represent generalizations of the simple equations (\[eq:basicIdea\]). Finally, let $\{p_{\nu}\}$ represent a finite set of polynomials of the form (\[eq:basicEquation\]), i.e. $$\begin{aligned} \label{eq:basicEquationFamily} p_{\nu}=A_{\nu}(z)Q(z,r_1,r_2,r_3)+B_{\nu}(z).\end{aligned}$$ For instance, this may be a multi-model of a continuum of plants or Kharitonov polynomials of an interval uncertainty. It can be simply proven that the rank-condition (\[eq:RangebdingungFuerSingulaereGammaGebiete\]) does not depend on the polynomials $A_{\nu}(z)$ and $B_{\nu}(z)$. Hence, a singular $\Gamma$ is completely determined by the polynomial $Q$ in (\[eq:QEquation\]). The polynomials $A_{\nu}(z)$ and $B_{\nu}(z)$ rather determine the singular frequencies on $\partial\Gamma$. Hurwitz-stability ----------------- Consider the Hurwitz-region $\partial\Gamma~=\{j\omega:~\omega\in \mathbb{R}\}$. Then condition (\[eq:RangebdingungFuerSingulaereGammaGebiete\]) $$\label{eq:rankConditionHurwitz} \textrm{Rank}\frac{\partial{}(H,G)}{\partial{}(k_I, k_D)}=1$$ is satisfied everywhere on the imaginary axis, since $$\begin{aligned} \label{eq:H} H&=&R_A k_I-\omega^2 R_A k_D-\omega I_A k_P +R_B\\ \label{eq:G} G&=&I_A k_I-\omega^2 I_A k_D +\omega R_A k_P +I_B.\end{aligned}$$ Referring to equations (\[eq:decouplingFuntion\]) and (\[eq:HurwitzQ\]), a simple choice for the decoupling function is $E_\Gamma(z) = 1$. Indeed the function (\[eq:Fz\]) $$\label{eq:FzHurwitz} F(s)=k_I+ k_P s+ k_D s^2 + \frac{B(s)}{A(s)}$$ on the imaginary axis $s=j\omega$ decouples into two equations of the form (\[eq:singFreqsLines\]) and (\[eq:singFreqs\]), similar (but not equal) to equations (\[eq:basicIdea\]). Note that the rank-condition (\[eq:rankConditionHurwitz\]) applies also on any line parallel $\sigma=\sigma_0$ to $s=j\omega$. Thus all derivations hold also for $\partial\Gamma~=\{\sigma_0+j\omega:~\omega\in \mathbb{R}\}$, see [@bajc:auto:2006]. Schur-stability --------------- Consider the Schur-circle $\partial\Gamma_1~=\{e^{j\alpha}:~\alpha\in [-\pi,~\pi]\}$. It can be easily checked that the rank-condition (\[eq:RangebdingungFuerSingulaereGammaGebiete\]) is satisfied on $\partial\Gamma_1$ for $$\label{eq:SchurQEquation} Q=(1+z^2) r_1+z r_2+ r_3$$ and that the matrix $T$, as defined in (\[eq:Tmatrix\]), reads $$\begin{aligned} \label{eq:Tcircle} {T} = \left[ \begin{array}{ccc} 1 & 0 & 1\\0 & 1 & 0\\1 & 0 & 0 \end{array} \right].\end{aligned}$$ Following (\[eq:decouplingFuntion\]), the trivial decoupling functions of (\[eq:SchurQEquation\]) on the Schur-circle $\Gamma_1$ are $$\label{eq:decouplingFunctionScur} E_\Gamma(z) = 1+z^2~~\textrm{or}~~E_\Gamma(z) = z.$$ For $E_\Gamma(z) = z$, the imaginary part of the function on $\Gamma_1$ $$\label{eq:FzSchur} F(z)=\frac{1+z^2}{z} r_1+r_2 + \frac{B(z)}{z A(z)}r_3$$ is of the form (\[eq:Iq1\]), since imaginary part of $\frac{1+z^2}{z}$ is null on $\partial\Gamma_1$. $\Gamma$-stability ------------------ Consider a $\Gamma-$circle with center on real axis $\partial\Gamma~=\{m+\rho e^{j\alpha}$, $\alpha\in [-\pi,~\pi]\}$, Fig. \[fig:circles\]. Now $$\label{eq:CirclesQEquation} Q=(\rho^2-m^2+z^2) r_1+(z-m) r_2+ r_3.$$ For $\Gamma-$circles with center at $\tau=m$ and radius $\rho$ a transformation matrix from $c-$ to $r-$parameter space is $$\begin{aligned} \label{eq:12} {T} = \left[ \begin{array}{ccc} \rho^2 - m^2 & -m & 1\\0 & 1 & 0\\1 & 0 & 0 \end{array} \right].\end{aligned}$$ Furhtermore $$\label{eq:decouplingFunctionCircle} E_\Gamma(z) = z-m,$$ and $$\label{eq:FzCircle} F(z)=\frac{\rho^2-m^2+z^2}{z-m} r_1+r_2 + \frac{B(z)}{(z-m) A(z)}r_3.$$ ![Schur- and $\Gamma-$circles[]{data-label="fig:circles"}](tex_circ_pic){width="7cm"} Stable convex polygons {#sec:stablePolygons} ====================== This section recalls briefly a solution to problem \[P2\] as formulated in Section \[sec:veryBasicIdea\]. For details the reader is referred to [@bajc:med2001] and [@bajc:auto:2006]. The algorithm is motivated by the concept of inner polygons, which claim a necessary condition for stability: A polygon $\Pi$ is said to be an inner polygon if entering the polygon in ($r_1,r_2$)- i.e. ($k_I,k_D$)-plane, causes an eigenvalue-pair to enter the $\Gamma-$region at the corresponding singular frequency. Discrete-time domain -------------------- Let $\{z'\}$ be the set of singular frequencies on $\partial\Gamma$ determined by the equation (\[eq:singFreqs\]) for a fixed $r_3$, and let $\{\lambda=\lambda(z')\}$ be the corresponding straight lines determined by (\[eq:singFreqsLines\]), see Fig. \[fig:e1e2defs\]. In order to automate the detection of an inner polygon each boundary line $\lambda$, will be assigned a “transition” $e$: it is negative if the transition $[\delta{}r_1,\delta{}r_2]$ (see Fig. \[fig:e1e2defs\]) over the singular line causes an eigenvalue to become stable, i.e. enter the $\Gamma$-region, otherwise it is positive. Let $e_1$ correspond to $\delta{}r_1>0,~\delta{}r_2=0$ and $e_2$ to $\delta{}r_2>0,~ \delta{}r_1=0$. [![Definition of $e_1$ and $e_2$: the motion of eigenvalues in the vicinity of a singular frequency $z'$. The shaded part refers to the stable side of $\Gamma$ and the normal $N$ points outside the $\Gamma$-region.[]{data-label="fig:e1e2defs"}](tex_fig3_pic "fig:")]{} To compute the functions $e_1$ and $e_2$, define the normal vector ${\mbox{\boldmath$N$}}$ on $\partial\Gamma$ at the singular frequency $z'=\tau(\alpha')\pm j\eta(\alpha')$ by its complex associate as $${\mbox{\boldmath$N$}}=\left({\partial F}/{\partial\tau}+j {\partial F}/{\partial\eta} \right)_{\normalsize z'}.$$ For tracking the motion of an eigenvalue due to small variations in $r_1$ and $r_2$ introduce the vectors $$\label{eq:eid=n*muid} {\mu_1}=\frac{\displaystyle\textrm{d}{z}}{\displaystyle\textrm{d}{r_1}} =\frac{\displaystyle\textrm{d}\tau}{\displaystyle\textrm{d} r_1}+j\frac{\displaystyle\textrm{d}\eta}{\displaystyle\textrm{d}r_1} ~~\textrm{and}~~\mu_2=\frac{\displaystyle\textrm{d}{z}}{\displaystyle\textrm{d}{r_2}}.$$ Assuming ${\partial{p}}/{\partial{z}} \neq 0$[^2], it can be easily shown that $$\label{eq:mu1_complex} {\mu_1}=-{\displaystyle\frac{\partial{p}}{\partial {r_1}}}/{\displaystyle\frac{\partial{p}}{\partial {z}}}.$$ Now transitions functions can be computed by $$\label{eq:eid=n*muid} e_{1/2}=\mathrm{Re}\left(\mu_{1/2}^{\ast}N\right)_{\normalsize z'}.$$ Using this information, an algorithm for the detection of the *inner polygons* (those with maximal number of $\Gamma$-stable eigenvalues) is developed in [@bajc:med2001]. Such polygons are the only stability candidates, that can be proved by checking any point therein. [*Example 1:* Consider the discrete-time model of the plant $$\label{eq:GsDigit} G=10^{-6}\frac{4.165 z^3 + 45.77 z^2 + 45.77 z + 4.165}{z^4 - 3.985z^3 + 5.97z^2 - 3.985z + 1}$$ and a three-term stabilizer $$\label{eq:threeTerm} C(z)=10^4\frac{(z^2 - 1.541 z +0.5992)(c_1+c_2 z+c_3 z^2)}{z(z+0.4047)(z+0.2162)(z-0.4934)},$$ whose parameters $c_1, c_2, c_3$ are to be synthesized. The synthesis is done in $(r_1,r_2,r_3)$-parameter space. Therefore the transformation (\[eq:Tcircle\]) can be used. Then $$\begin{aligned} \label{eq:Azexmaple} A(z)&=& {z}^{5}+9.44z^4-5.34 z^3-9.34z^2+5.04 z+0.59\\ \label{eq:Bzexmaple} B(z)&=&0.19 z^8-0.73 z^7+z^6-0.45 z^5-0.12 z^4+\cdots\nonumber\\ {} &{} & 0.14 z^3-0.009z^2-0.008z.\end{aligned}$$ For $r_3=-0.26118$, the singular frequencies lying on the Schur-circle are computed to be $z'_1=1, z'_{2}=0.9172\pm j 0.3983, z'_{3}=0.5628\pm j 0.8266, z'_{4}=-1$. The resulting stable polygon is shown in Fig. \[fig:astabpolyg1\]. It is enclosed by the straight lines frequencies $\lambda_1,\lambda_2$ and $\lambda_3$, corresponding to $z'_1, z'_2 ~\textrm{and}~ z'_3$.]{} [![The stable polygon for $r_3=-0.26118$, Example 1[]{data-label="fig:astabpolyg1"}](tex_fig4_pic "fig:"){width="7.5cm"}]{} Hurwitz conditions ------------------ Let $e_I$ correspond to $\delta{}k_I>0,~ \delta{}k_D=0$ and $e_D$ to $\delta{}k_D>0,~ \delta{}k_I=0$. It can be shown that (\[eq:eid=n\*muid\]) yields $$\label{eq:cros} e_{I/D}=\Bigg\vert\frac{\partial(H,G)}{\partial(\omega,k_{I/D})}\Bigg\vert_{\normalsize \omega'}.$$ Expressions for $e_{I/D}$ do not depend on where a boundary line is crossed at. Indeed, check that (\[eq:cros\]) is equivalent to $$\label{eq:signsEquation2} e_{I/D}=\frac{\partial k_P}{\partial\omega}\Bigg\vert\frac{\partial(H,G)}{\partial(k_P,k_{I/D})}\Bigg\vert_ {\normalsize\omega'},$$ where $k_P=k_P(\omega)$ stands for the $k_P$-plot. Since $\omega'$ is a singular frequency, the determinant in the latter equation is shown to be independent on $k_I$ and $k_D$. Furthermore, the following holds $$\label{eq:signsEquation3} \mathrm{sign}~e_{I}=-\mathrm{sign}~\frac{\partial k_P}{\partial\omega}\Bigg\vert_{\omega'},~~~~~~ \mathrm{sign}~e_{D}= \mathrm{sign}~\frac{\partial k_P}{\partial\omega}\Bigg\vert_{\omega'}.$$ These expressions prove again that the transitions $e_I$ and $e_D$ are independent on parameters $ k_I$ and $ k_D$, and of opposite sign. Their sign is determined by the slope of the $k_P$-plot at the corresponding singular frequency, see Fig. \[fig:kpplot\]. ![Stable polygons for $k_P=-2$, Example 2[]{data-label="fig:singLinesExample1"}](tex_fig5_pic){width="7.5cm"} [*Example 2:* Consider the polynomial (\[eq:basicCharacteristicPolynomial\]) with $$\begin{aligned} \label{eq:cpExample1} A(s)&=& -0.5 s^4-7 s^3-2 s+1\\ \label{eq:cpExample11} B(s)&=&{s}^{7}+11\,{s}^{6}+46\,{s}^{5}+95\,{s}^{4}+109\,{s}^{3}+74\,s^2+24 s.\end{aligned}$$ The singular frequencies for $k_P=-2$ are computed from its $k_P$-plot (see Fig. \[fig:kpwGraphExample1\]): $s'_0=0,~s'_1=\pm j 0.3530,~s'_2=\pm j 0.6638,~s'_3=\pm j 0.7742,~s'_4=\pm j 3.3473.$ Fig. \[fig:singLinesExample1\] depicts the corresponding straight lines and the stable polygons. Note that the stable polygons need not to be connected.]{} $k_P$-Problem {#sec:kp-Porblem} ============= Hurwitz-stability {#sec:kp-Porblem:Hurwitz} ----------------- This section focuses on the problem \[P1\], as defined in Section \[sec:veryBasicIdea\]: a rule is sought to discriminate $k_P-$intervals with stable PID controllers. Intuitively, it must be tightly related to the $k_P$-plot. Indeed, it is clear that at maxima and minima of the $k_P-$plot - compare Fig. \[fig:kpplot\] and \[fig:kikdplane\] - convex polygons close at an edge due to the overlapping of two straight boundary lines. As $k_P$-intervals with maximal number of singular frequencies are most likely to host stable polygons. The following result renders this idea precisely. ** \[thm:r2GriddingIntervalsHurwitz\] Consider the polynomial (\[eq:basicCharacteristicPolynomial\]). Assume that polynomial $A(s)$ has no zeros on the imaginary axis and let\ ------ ---------------------------------------------------------------- $N$: order of the polynomial (\[eq:basicCharacteristicPolynomial\]) $M$: order of the polynomial $A(s)$ $P$: number of RHP zeros of $A(s)$ $Z$: number of singular frequencies in the interval $0 < \omega < +\infty$. ------ ---------------------------------------------------------------- A necessary condition for stability of (\[eq:basicCharacteristicPolynomial\]) is $$\label{eq:kPConditionHurwitz} Z\geq \frac{E(N-M+2P-1)}{2}.$$ Here, the function $E:\mathbb{N}\mapsto\mathbb{N}$ selects the nearest smaller even number. The proof of the theorem can be found in [@bajc:auto:2006]. It is important to observe, that we exclude the zero singular frequency from $Z$. Using this criterion one can directly read from the $k_P$-plot (Fig. \[fig:kpplot\]) the $k_P-$interval(s) with (potentially) host stable polygons. However, in some cases a polygon may close when three boundary lines intersect at a single point in ($k_P, k_I, k_D$)-parameter space. This situation destroys the sufficiency of the condition (see Lemma \[thm:stabPeaks\]) is not sensed by the above criterion and will be discussed in Section \[sec:stabilitypeaks\]. The following is an extension of Theorem \[thm:r2GriddingIntervalsHurwitz\] to the cases when $A(s)$ possesses zeros on the imaginary axis. ** \[thm:r2GriddingIntervalsHurwitz2\] Suppose $A(s)$ has $J$ zeros on the imaginary axis. Then, for stability of the polynomial (\[eq:basicCharacteristicPolynomial\]) $Z$, singular frequencies are required within the interval $\omega\in(0,+\infty)$, where (a)  if $s=0$ is not a zero of $A(s)$\ $$\label{eq:kPConditionHurwitz_1} Z \geq \frac{E(N-M+2P-J-1)}{2}$$ (b)  if $s=0$ is a zero of order $J_0$ of $A(s)$\ $$\label{eq:kPConditionHurwitz_2} Z \geq \frac{E(N-M+2P-J-1)-E(J_0)}{2}.$$ ![The $k_P$-plot, Example 2[]{data-label="fig:kpwGraphExample1"}](tex_fig6_pic){width="8cm"} ![ The region of PID stabilizers, Example 2[]{data-label="fig:3DsingLinesExample1"}](tex_fig7_pic){width="8.25cm"} [*Example 2: (cont).*  The $k_P$-plot is depicted in Fig. \[fig:kpwGraphExample1\]. Notice that it is very easy to read the number of available singular frequencies from Fig. \[fig:kpwGraphExample1\] for a given $k_P$. For the polynomial $A(s)$ we have $N=7,~M=4,~P=1,~J=0,~J_0=0$. According to Theorem \[thm:r2GriddingIntervalsHurwitz\] for stability at least $Z \geq {E(N-M+2P-J-1)}/{2}=2$ singular frequencies are required for $\omega > 0$. By checking the Fig. \[fig:kpwGraphExample1\] it is obviously that this condition is fulfilled for $-24<k_P<6.1565$. More precisely, for $-24 < k_P < -2.7614$ and $3.7664 < k_P < 6.1565$, two non-zero singular frequencies exist, and for $-2.7614 < k_P < 3.7664$ four ones. Finally, by gridding $k_P$ within these intervals, stable polygonal slice can be computed. The result is shown in Fig. \[fig:3DsingLinesExample1\].]{} ![The $k_P$-plot, Example 3[]{data-label="fig:kpw2Intervals"}](tex_fig8_pic){width="8cm"} [*Example 4: Missing stability. * Let $$\begin{aligned} A(s)&=& 1\\ B(s)&=&s^5+s^4-3s^3-s^2+2 s.\end{aligned}$$ Theorem \[thm:r2GriddingIntervalsHurwitz\] requires at least $2$ singular frequencies in $\omega>0$, however for $-2>k_P$, $1$ singular frequency exists, otherwise none. Thus, polynomial (\[eq:basicCharacteristicPolynomial\]) is always unstable, no matter what $k_P,~k_I,~k_D$.]{} Schur-stability {#sec:kp-Porblem:Schur} --------------- Without loss of generality, we consider just the Schur-circle. The generalizations for other $\Gamma-$circles are straightforward. ** \[thm:r1GrddingIntervals\] Consider the characteristic polynomial (\[eq:basicEquation\]) and the Schur-circle $\Gamma_1$. Let\ -------- ------------------------------------------------------------------------------ $N$: order of the polynomial (\[eq:basicEquation\]) $R$: number of zeros of $A(z)E_\Gamma(z)$ lying inside $\partial\Gamma_1$ $J$: number of zeros $\neq\pm 1$ of $A(z)E_\Gamma(z)$ lying on $\partial\Gamma_1$ $J_+$: order of the zero $+1$ of $A(z)E_\Gamma(z)$ $J_-$: order of the zero $-1$ of $A(z)E_\Gamma(z)$ $Z$: number of singular frequencies in the interval $0 < \alpha < +\pi$. -------- ------------------------------------------------------------------------------ A necessary condition for stability of (\[eq:basicEquation\]) is $$\label{eq:kPCondition} Z \geq N-R-\frac{J+E(J_+)+E(J_-)+2}{2}.$$ *Example 1: (cont).* It can be checked that $A(z)$ has three zeros inside the Schur-circle, one zero at $z=-1$ and one zero outside the Schur-circle. Thus if the decoupling function $E_\Gamma(z)=z$ is used, it follows that $N=8,~R=3+1=4, ~J=1,~J_+=0$, and $J_-=1$. Hence, for stability, $Z\geq 3$ singular frequencies are required in the interval $0<\alpha<+\pi$. In order to discriminate stable $r_3$ intervals check the $k_P$-plot in Fig. \[fig:kpwexample\]. The stable interval is indicated by the grayed strip in Fig. \[fig:kpwexample\] $-0.52236 < r_3 <0.00290$. Notice that the zoomed plot in Fig. \[fig:kpwexample\] reveals that for $0< r_3 <0.00290$ four additional singular frequencies appear. On the other side if the decoupling function $E_\Gamma(z)=1+z^2$ is used, then $N=8,~R=3,~J=3,~J_+=0$, and $J_-=1$ i.e. again for stability $Z\geq 3$ singular frequencies are required in the interval $0<\alpha<+\pi$. [![The $k_P$-plot, Example 1[]{data-label="fig:kpwexample"}](tex_fig9_pic "fig:"){width="8cm"}]{} *Example 5: PID control.* Consider PID control of the plant (\[eq:GsDigit\]) now using the control law (\[eq:PID\]). It can be shown that in this case $$\begin{aligned} \label{eq:A1B1} A(z)&=&{z}^{3}+ 10.98\,{z}^{2}+10.98\,z+ 1\\ B(z)&=&0.1 z^6-0.5 z^5+z^4-z^3+0.5 z^2-0.1 z.\end{aligned}$$ By using $E_\Gamma(z)=z$, it is easily checked that $N=6,~R=1+1=2, ~J=0,~J_+=0$ and $J_-=1$. Hence, $Z\geq 3$ singular frequencies within $0<\alpha<+\pi$ are required. However, the maximal number of singular frequencies within $0<\alpha<+\pi$ is $2$, so no PID controller can stabilize the plant (\[eq:GsDigit\]). Stability peaks {#sec:stabilitypeaks} --------------- Note that at a peak, the three-term polynomial (\[eq:basicCharacteristicPolynomial\]) must have at least three different eigenvalues on the imaginary axis, that is $$\label{eq:peaks} A(s)(k_I+k_Ps+k_D s^2)+B(s)=R(s)\prod_{i=1,2,3}(s^2+\omega'^2_i),$$ where $R(s)$ is a remainder polynomial, which has to be stable, otherwise the peak is irrelevant. This is a nonlinear set of $N$ equations with $N$ unknowns. The left-hand side of the equation provides the three unknowns $k_I,k_P,k_D$ and the right-hand side the rest $N-3$ ones, including three singular frequencies $\omega'_i,~i=1,2,3$ and $N-6$ coefficients of the polynomial $R(s)$. Hence by elimination of the latter $N-3$ variables, a system of three nonlinear equations with the three $k_I,k_P,k_D$ variables results. Its solution provides the required peaks. In general, finitely many stability peaks exist. [*Example 6: (cont).* It can be shown that for the three-term polynomial defined by (\[eq:Example4\]) a stability peak appears at $k_P\approx-9.0023$. The three straight lines corresponding to $\omega'_1\approx 0.2581$, $\omega'_2\approx 0.44261$ and $\omega'_3\approx 9.7621$, intersect at $k_D\approx 21.4958, k_I\approx 3.0195$. A stability peak appears also in Example 7, see Fig. \[fig:det6\].]{} ** \[thm:stabPeaks\] The condition in Theorem \[thm:r2GriddingIntervalsHurwitz\] is also sufficient for $N\leq 6$. Theorem \[thm:r2GriddingIntervalsHurwitz\] provides necessary and sufficient conditions for any PID feedback loop with plants of 1st, 2nd and 3rd order. PID for time-delay systems {#sec:TimeDelaySystems} ========================== In this section we extend the PID control theory to systems with time-delay. Such feedback loops always result in quasi-polynomials of the form $$\begin{aligned} \label{eq:basicEquationTD} p =A(s)( k_I+k_P s +k_D s^2)+ B(s)e^{L s},\end{aligned}$$ where $L>0$ is the time-delay. Fundamental stability conditions regarding quasi-polynomials are provided in [@pon:55]. E.g. a simple necessary condition is the existence of principal term, i.e. $e^{Ls}$ in (\[eq:basicEquationTD\]) must be multiplied by the highest power in $s$. Thus, in this section we assume $n\geq m+2$, i.e. the quasi-polynomials of the *retarded* ($n>m+2$) and *neutral* type ($n=m+2$) are considered only. It is easy to check that decoupling conditions hold for the quasi-polynomial (\[eq:basicEquationTD\]), too. With definitions $$\label{eq:amp_phase_exp} \alpha(\omega)=\sqrt {{\frac {{{R_{B}}}^{2}+ {{I_{B}}}^{2}}{{{{R}_{A}}}^{2}+{{I_{A}}}^{2}}}},~~ \tan\phi(\omega) = {{\frac {{R_{A}}\,{I_{B}}-{I_{A}}\,{R_{B}}}{{R_{A}}\,{R_{B}}+ {I_{A}}\,{I_{B}}}}},$$ equations (\[eq:singFreqsLines\]) and (\[eq:singFreqs\]) take the form $$\begin{aligned} \label{eq:singularLineGeneratorTD} k_I -\omega^2 k_D & = & \alpha(\omega) \cos(\omega L+\phi(\omega)),\\ \label{eq:singularFrequencyGeneratorTD} \omega k_P &=& {\alpha(\omega)}\sin(\omega L+\phi(\omega)).\end{aligned}$$ Obviously $k_P$-plot is now a sinusoidal function and we have to deal with infinitely many singular frequencies. High-frequency behavior {#sec:HighfrequencyBehaviour} ----------------------- Consider equation (\[eq:amp\_phase\_exp\]). Given that $n\geq m+ 2$ for high singular frequencies, i.e. as $\omega\to\infty$ $$\label{eq:alpha_infty} \alpha(\omega)\sim\omega^{n-m}$$ and $$\label{eq:phase_infty} \tan \phi(\omega)\to \left\{ \begin{array}{l} \pm\infty \\ 0 \end{array} \right. ~~\Rightarrow~~ \phi(\omega)\to \left\{ \begin{array}{l} \pm\frac{\pi}{2}\\ 0~~\textrm{or}~~\pi. \end{array} \right.$$ For a fixed $k_P$-grid, equation (\[eq:singularFrequencyGeneratorTD\]) implies $$\label{eq:sf_infty} \sin(\omega L +\phi(\omega)) \to \frac{1}{ k_P}{\omega^{1-n+m}} \to 0,$$ that is, all higher singular frequencies tend to $$\label{eq:sing_freqs_infty_3} \omega\to \left\{ \begin{array}{ll} k \pi/ L, & {}\\ (k+\frac{1}{2}) \pi/ L,& k\in \mathbb{N}, {k}\gg 1, \end{array} \right.$$ and $$\label{eq:cos_freqs_infty} \cos(\omega L +\phi(\omega)) \to \pm 1.$$ To investigate the behavior of boundary lines for high frequencies, applying $\omega\to\infty$ to (\[eq:singularLineGeneratorTD\]) reads $$\label{eq:irb} k_D = \pm{\alpha(\omega)}/{\omega^2}\big\vert_{\omega\to\infty}.$$ For quasi-polynomials of neutral type, $n=m+2$ in (\[eq:alpha\_infty\]), and boundary lines converge to so-called *infinity root boundaries* $$\label{eq:irb1} k_D = \pm{b_n}/{a_m},$$ For quasi-polynomials of retarded type, straight lines diverge. Note that infinity root boundaries describe the situation with infinitely many eigenvalues arbitrarily close to the imaginary axis. Relevant frequency range {#sec:theRelevantFrequencyRange} ------------------------ According to (\[eq:signsEquation3\]) the sign of the transition function $e_{I/D}$ at a singular frequency $\omega'$ is determined by the slope of the function $k_P=k_P(\omega)$ at $\omega'$. Hence, at successive singular frequencies corresponding to a fixed $k_P$, $e_{I/D}$ takes opposite signs. This motivates the definition of the set of odd $\Omega_o$ and even $\Omega_e$ singular frequencies $$\Omega_o=\{\omega'_1,~\omega'_3,~\omega'_5,~\cdots\},~~\Omega_e=\{\omega'_0,\omega'_2,~\omega'_4,~\cdots\}$$ with $$0=\omega'_0<\omega'_1<\omega'_2<\omega'_3<\cdots<\infty.$$ The transitions $e_{I/D}$ have the same sign for all even (odd) singular frequencies, which is opposite to that of odd (even) singular frequencies. The intersection of a singular line with $k_I=0$ and fixed $k_P$ $$k_D(0)=-\frac{ k_P}{\omega\tan(\omega L +\phi(\omega))}$$ discriminates between even and odd singular lines, too. For quasi-polynomials of retarded type (no infinity root boundaries), as $\omega\to\infty$, one group of the boundary straight lines diverges toward $+\infty$, while the other towards $-\infty$. Since for high frequencies, $k_D$ and $e_D$ share the same sign, the stable region must lie between even and odd boundary lines. Thus, starting from a sufficiently large singular frequency, the boundary lines become irrelevant. The same holds for the quasi-polynomials of neutral type. However, the infinity root boundaries (\[eq:irb\]) may impact the stable polygons: indeed these must lie between the two infinity root boundaries (\[eq:irb1\]). As a conclusion, stable inner polygons are determined by the low frequency boundaries and infinity root boundaries (if any). For instance, it was shown that for PID control of a first order proportional term $G=K/(T s+1)e^{-Ls}$, the relevant singular frequencies are the first two ones and the two infinity root boundaries. It is difficult to state a rule, which would precisely discriminate the relevant frequency range in the general case. However, some estimates are still thinkable. The relevant frequency range should comprise the singular frequencies, which correspond to the minima and maxima of the stable $k_P$-intervals. Another helpful rule of thumb is to discriminate the region, where $k_P(\omega)$, Fig. \[fig:kpplot\], oscillates with an almost fixed period as settled in (\[eq:phase\_infty\]). Stable $k_P$ intervals {#sec:theStableKPIntervals} ---------------------- In this section the solution of $k_P-$problem is extended to time-delay systems. ** \[thm:kpGriddingIntervalsTD\] Consider the quasi-polynomial (\[eq:basicEquationTD\]). Assume that $A(s)$ has $J$ non-zero zeros on the imaginary axis, $P$ right-hand-side zeros and a zero of order $J_0$ at $s=0$. If (\[eq:basicEquationTD\]) is Hurwitz-stable for a fixed $k_P=k'_P$, then a $k \in \mathbb{N}$ exists, such that for $l \geq k$, $l \in \mathbb{N}$, the number of singular frequencies $Z$ in the interval $0 < \omega < (2 l \pi+\delta)/L$ corresponding to $k_P=k'_P$ is $$\label{eq:r1ConditionQuasipolynomial} Z \geq \frac{E(4l+N-M+2P-J-1)+E(J_0)}{2},$$ where $\delta$ is chosen such that the principal term does not vanish at $\omega = (\pm 2l \pi+\delta)/L$. For the proof of this theorem is refer to [@bajc:tds2004]. [*Example 7. * Consider PID control of the plant $$\label{eq:norbertsBeispiel} G(s)= \frac{-s^4-7s^3-2 s+1}{(s+1)(s+2)(s+3)(s+4)(s^2+s+1)}e^{-0.05 s}.$$ Note that $N=7$, $M=4$ and $P=1$ (since a right-hand-sided zero of $A(s)$ exists at $s=0.3483$). According to Theorem \[thm:kpGriddingIntervalsTD\], we need to find a sufficiently large $k$, such that within any interval $0 < \omega <40 l\pi +\delta$, with $l\geq k$, at least $E(4 l+ 6)/2=2r+3$ singular frequencies are available. It can be easily checked that already for $k=1$ and $\delta=\pi$ the condition is fulfilled within $-24<k_P<6.0693$. Fig. \[fig:det6\] shows the set of all PID stabilizers for the plant (\[eq:norbertsBeispiel\]). This example illustrates two interesting situations: first, for $-3.7671 < k_P <4.6807$ the stable region includes two separated polygons, and second, one of the polygons closes at a vertex.]{} ![PID stability region, Example 7[]{data-label="fig:det6"}](tex_fig11_pic) Conclusion ========== Fast computational methods of the set of all PID controllers for linear continuous-time, discrete-time and time-delay systems are proposed in this article. The driving force of the theory is the fact that non-convex stability regions can be built up easily by convex polygonal slices. The high computational speed results due to inspection of conditions at a relatively low number of (singular) frequencies. Only the results of the control group at DLR are surveyed here. A software tool called [<span style="font-variant:small-caps;">Robsin</span>]{} originated on that basis. In author’s opinion, the proposed design approach is especially elegant in the discrete-time domain, and of particular interest for time-delay systems. A powerful feature is the fact that, in principle, in all cases all design lines apply also when simultaneous stabilization of a set of plants is considered. This paves the basis for robust design of PID controllers. Indeed, while in all derivations of the article just a single representant is assumed, the extensions to the situation with a finite number of representants is straightforward. When applying the $k_P$-criterion, one would have to search for the intersection of stable $k_P-$intervals of each representant. And, inner-polygon candidates should provide the simultaneous stability for all representants. Yet some important issues, particularly those involving time-delay systems, remain open. For instance, it is not clear how to discriminate the frequency range with relevant singular frequencies for quasi-polynomials. [99]{} Bajcinca, N.:  [*On the computation of the total set of robust discrete-time PID controllers*]{},  ECC 2007, Kos, Greece. Bajcinca, N.:  [*Design of robust PID controllers using decoupling at singular frequencies*]{}, Automatica, 2006. Bajcinca, N.:  [*A necessary stabilization condition for PID control*]{},  ACC 05, Portland, 2005 USA. Bajcinca, N. and T. Hulin:  [*Robsin: A new tool for robust design of PID and Three-term controllers based on singular frequencies*]{}, CACSD/ISIC/CCA 2004, Taipei. , N.:   *Computation of stable regions in PID parameter space for time-delay systems*,  IFAC Workshop on TDS, Leuven, 2004. Soylemez, M.T., N. Munro and H. Baki:  [*Fast calculation of stabilizing PID controllers*]{},  Automatica 39, 121-126. <span style="font-variant:small-caps;">Silva</span> G.J, A. <span style="font-variant:small-caps;">Datta</span> and S. P. <span style="font-variant:small-caps;">Bhattacharryya</span>: *New results on the synthesis of [PID]{} controllers*. IEEE Trans. on Automatic Control, pp. 241-252. Ackermann, J., D. Kaesbauer and N. Bajcinca:  [*Discrete-time robust PID and Three-Term control*]{},  XV IFAC World Congress, 2002 Barcelona. Ackermann, J. and D. Kaesbauer:  [*Design of robust PID controllers*]{},  Proc. European Control Conference, 2001 Porto. Bajcinca, N.:  [*The method of singular frequencies for robust design in an affine parameter space*]{},  9th Mediterranean Conference on Control and Automation, 2001 Dubrovnik, Croatia. Ho, M.T. , A. Datta and S. P. Bhattacharryya:  [*Structure and synthesis of PID controllers*]{},  Springer, 2000 London. Munro, N, M.T. Soylemez:  [*Fast calculation of stabilizing PID controllers for uncertain parameter systems*]{},  In Proceed. ROCOND 2000, Prague. <span style="font-variant:small-caps;">Pontryagin</span>, L.S: *On the Zeros of Some Elementary Transcendental Functions*. American Mathematical Society Translations, pp. 95-110. [^1]: N. Bajcinca is with the Max-Planck Institute for Dynamics of Complex Technical Systems in Magdeburg, Germany. The work presented here was done while working with the group of Prof. J. Ackermann at German Aerospace Research Center (DLR) in Oberpfaffenhofen, Germany. [^2]: For situations with ${\partial{p}}/{\partial{z}} = 0$, refer to [@bajc:auto:2006].

--- abstract: 'Suppose $G$ is a real reductive group. The determination of the irreducible unitary representations of $G$ is one of the major unsolved problem in representation theory. There is evidence to suggest that every irreducible unitary representation of $G$ can be constructed through a sequence of well-understood operations from a finite set of building blocks, called the unipotent representations. These representations are ‘attached’ (in a certain mysterious sense) to the nilpotent orbits of $G$ on the dual space of its Lie algebra. Inside this finite set is a still smaller set, consisting of the unipotent representations attached to non-induced nilpotent orbits. In this paper, we prove that in many cases this smaller set generates (through a suitable kind of induction) all unipotent representations.' author: - 'Lucas Mason-Brown' bibliography: - 'bibliography.bib' title: Upper Triangularity for Unipotent Representations --- Suppose $G$ is a real reductive group. There are three powerful techniques for producing irreducible unitary representations of $G$: parabolic induction, cohomological induction, and the formation of complementary series. These techniques produce most, but not all, irreducible unitary representations. If $G = SL_2(\mathbb{R})$, there are exactly three missing representations: the two limit of discrete series representations and the trivial representation. These missing representations are called *unipotent*. They are linked in a mysterious way to the nilpotent orbits of $G$ on the dual space of its Lie algebra. In the case of $SL_2(\mathbb{R})$, there are exactly three such orbits: two principal nilpotent orbits and $\{0\}$. The limit of descrete series representations correspond (in a precise sense) to the principal nilpotent orbits, and the trivial representation corresponds to $\{0\}$. The search for a general theory of unipotent representations has a long and colorful history. Highlights include character formulas, due to Barbasch and Vogan ([@BarbaschVogan1985]), for unipotent representations of complex reductive groups and a construction, due to Torasso ([@Torasso1997]), of the unipotent representations attached to minimal nilpotent orbits. Despite these results (and others), the general theory remains elusive. In this article, we propose a blueprint. If $\mathbf{G}$ is a complex reductive algebraic group, and $\mathbf{L} \subset \mathbf{G}$ is a Levi subgroup, there is a correspondence $$\mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}}: \{\text{nilpotent co-adjoint }\mathbf{L}-\text{orbits}\} \to \{\text{nilpotent co-adjoint }\mathbf{G}-\text{orbits}\}$$ called the induction of nilpotent orbits. We propose the following general strategy for understanding the unipotent representations of $G$: 1. Understand the unipotents ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ attached to a non-induced $\mathbf{G}$-orbit $\mathcal{O}_{\mathfrak{g}}$ 2. Find a recipe for building the unipotents ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ attached to an induced $\mathbf{G}$-orbit $\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}}\mathcal{O}_{\mathfrak{l}}$ from the unipotents ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{l}})$ attached to $\mathcal{O}_{\mathfrak{l}}$ In a previous paper ([@MasonBrown2018]), we have made some encouraging progress towards $(1)$. In this paper, we will turn our attention towards $(2)$. Our main result in this direction is that the set ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ of unipotent representations attached to $\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}}\mathcal{O}_{\mathfrak{l}}$ is related in the K-theory of finite-length $(\mathfrak{g},\mathbf{K})$-modules by an upper triangular matrix to a set of standard classes related to ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{l}})$. This result is strong evidence in favor of the strategy outlined above: the unipotent representations attached to non-induced orbits are indeed the building blocks from which all others can be formed. There is some hope that the relationship between ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{l}})$ and ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ can be made extremely precise in certain special cases (in [@MasonBrown2019], this is achieved for the principal nilpotent orbit). In Section \[sec:preliminaries\], we will review some preliminary facts and constructions. In Section \[sec:unipotentrepresentations\], we will provide an operational definition of unipotent representations. In Section \[sec:uppertriangularity\] we will prove our main result (Theorem \[thm:unipotentsanddegenerates\]). This will require brief digressions into the theory of twisted $\mathcal{D}$-modules and translation functors (which we define in a rather general setting). In Section \[sec:examples\], we apply our main result to the special case of the $2^n$-orbit of $Sp(2n,\mathbb{R})$ (for several small values of $n$). preliminaries {#sec:preliminaries} ============= Harish-Chandra Isomorphism -------------------------- Let $\mathfrak{g}$ be a complex reductive Lie algebra and let $\mathfrak{h}$ be a Cartan subalgebra of $\mathfrak{g}$. Write $\Delta(\mathfrak{g},\mathfrak{h}) \subset \mathfrak{h}^*$ for the corresponding root system, $\Delta(\mathfrak{g},\mathfrak{h})^{\vee} = \{\alpha^{\vee}: \alpha \in \Delta(\mathfrak{g},\mathfrak{h})\}$ for the co-roots, and $W(\mathfrak{g},\mathfrak{h}) \subset GL(\mathfrak{h}^*)$ for the Weyl group. Since $W(\mathfrak{g},\mathfrak{h})$ is independent (up to canonical isomorphism) of $\mathfrak{h}$, we will sometimes write simply $W(\mathfrak{g})$. Recall that the center $Z(\mathfrak{g})$ of the universal enveloping algebra $U(\mathfrak{g})$ of $\mathfrak{g}$ is identified by the Harish-Chandra isomorphism $$\zeta: Z(\mathfrak{g}) \cong \mathbb{C}[\mathfrak{h}^*]^{W(\mathfrak{g})}$$ with the algebra of $W(\mathfrak{g})$-invariant polynomial functions on $\mathfrak{h}^*$. If $\lambda \in \mathfrak{h}^*$, there is a character $\gamma_{\lambda}$ of $Z(\mathfrak{g})$ defined by $$\gamma_{\lambda}(X) := \zeta(X)(\lambda) \in \mathbb{C}$$ Since $\zeta$ is an isomorphism, every character of $Z(\mathfrak{g})$ arises in this fashion and $\gamma_{\lambda} = \gamma_{\mu}$ if and only if $\lambda \in W(\mathfrak{g})\mu$. $(\mathfrak{g},\mathbf{K})$-Modules ----------------------------------- Let $\mathfrak{g}$ be a (finite-dimensional) complex Lie algebra and let $\mathbf{K}$ be a complex algebraic group. We say that $(\mathfrak{g},\mathbf{K})$ is a *pair* if 1. The Lie algebra $\mathfrak{k}$ of $\mathbf{K}$ is a subalgebra of $\mathfrak{g}$ 2. $\mathbf{K}$ acts on $\mathfrak{g}$ by Lie algebra automorphisms ${\operatorname{Ad}}(k) \in \mathrm{Aut}(\mathfrak{g})$ extending the adjoint action of $\mathbf{K}$ 3. The Lie algebra of ${\operatorname{Ad}}(\mathbf{K})$ is the subalgebra ${\operatorname{ad}}(\mathfrak{k}) \subset {\operatorname{ad}}(\mathfrak{g})$ \[def:gkmodule\] Let $(\mathfrak{g},\mathbf{K})$ be a pair. A $(\mathfrak{g}, K)$-module is a complex vector space $V$ with a Lie algebra action of $\mathfrak{g}$ and an algebraic group action of $\mathbf{K}$ such that 1. \[cond:ugk1\] The action map $$\mathfrak{g} \otimes_{\mathbb{C}} V \to V$$ is $\mathbf{K}$-equivariant, and 2. \[cond:ugk2\] The $\mathfrak{g}$-action, restricted to the subspace $\mathfrak{k} \subset \mathfrak{g}$, coincides with the differentiated action of $\mathbf{K}$. A morphism of $(\mathfrak{g},\mathbf{K})$-modules is a linear map which commutes with the actions of $\mathfrak{g}$ and $\mathbf{K}$. We will write $M(\mathfrak{g},\mathbf{K})$ the abelian category of $(\mathfrak{g},\mathbf{K})$-modules. An object $M \in M(\mathfrak{g},\mathbf{K})$ has finite-length if it has a finite composition series $$0 = M_0 \subset M_1 \subset ... \subset M_n = M$$ with irreducible quotients. We will write $M^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$ for the full subcategory of finite-length $(\mathfrak{g},\mathbf{K})$-modules and $KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$ for its Grothendieck group. Dixmier Algebras ---------------- Let $\mathbf{G}$ be a complex algebraic group and let $\mathfrak{g}$ be its Lie algebra. \[def:dixmier\] A Dixmier algebra for $\mathbf{G}$ is a triple $(B,{\operatorname{Ad}}, \phi)$ consisting of a complex algebra $B$, an algebraic group action ${\operatorname{Ad}}: \mathbf{G} \to {\operatorname{Aut}}(B)$, and an algebra homomorphism $$\phi: U(\mathfrak{g}) \to B$$ satisfying three properties: 1. $\phi$ commutes with the adjoint actions of $\mathbf{G}$ 2. The differential of the action ${\operatorname{Ad}}$ of $\mathbf{G}$ on $B$ (denoted ${\operatorname{ad}}$) coincides with the difference of the left and right $\mathfrak{g}$-actions defined by $\phi$: $${\operatorname{ad}}(X)b = \phi(X)b - b\phi(X) \qquad X \in \mathfrak{g}, b \in B$$ 3. \[cond:dixmier3\] $B$ is a finite-length $U(\mathfrak{g})$-bimodule Almost always, ${\operatorname{Ad}}$ and $\phi$ will be clear from the context. When this is the case, we will omit them from our notation. The prototypical example of a Dixmier algebra is the quotient of $U(\mathfrak{g})$ by a primitive ideal. More generally, \[prop:dixmier1\] Let $I \subset U(\mathfrak{g})$ be a two-sided ideal such that $I \cap Z(\mathfrak{g})$ has finite-codimension in $Z(\mathfrak{g})$. Then $A=U(\mathfrak{g})/I$ is a Dixmier algebra. Now let $B$ be a Dixmier algebra and let $(F,\rho)$ be a finite-dimensional representation of $\mathbf{G}$. We will define the structure of a Dixmier algebra on $B \otimes \mathrm{End}(F)$. The $\mathbf{G}$-action on $F$ lifts to a Lie algebra map $\rho: \mathfrak{g} \to \mathrm{End}(F)$, which lifts to an algebra homomorphism $\rho: U(\mathfrak{g}) \to \mathrm{End}(F)$. Let $$\Delta: U(\mathfrak{g}) \to U(\mathfrak{g}) \otimes U(\mathfrak{g}) \qquad \Delta(X) = X \otimes 1 + 1 \otimes X$$ be the co-multiplication on $U(\mathfrak{g})$. $\Delta$ is an algebra homomorphism. Consider the composite $$(\varphi \otimes \rho) \circ \Delta: U(\mathfrak{g}) \to B \otimes \mathrm{End}(F)$$ This map will play the role of $\phi$ for the algebra $B \otimes \mathrm{End}(F)$. The action of $\mathbf{G}$ on $B \otimes \mathrm{End}(F)$ is the diagonal one. \[prop:dixmier2\] Let $B$ be a Dixmier algebra and let $F$ be a finite-dimensional representation of $\mathbf{G}$. Then $B \otimes \mathrm{End}(F)$ is a Dixmier algebra (with $\phi$ and ${\operatorname{Ad}}$ defined as above). In analogy with Definition \[def:gkmodule\], we define \[def:dixmiermodule\] Let $\mathbf{K} \subset \mathbf{G}$ be an algebraic subgroup, and let $B$ be a Dixmier algebra for $\mathbf{G}$. A $(B,\mathbf{K})$-module is a (left) $B$-module $V$ with an algebraic action of $\mathbf{K}$ such that 1. The action map $$B \otimes V \to V$$ Is $\mathbf{K}$-equivariant, and 2. The $\mathfrak{k}$-action on $V$ coming from the $B$-action on $V$ and the homomorphism $\phi: \mathfrak{k} \subset U(\mathfrak{g}) \to B$ coincides with the differentiated action of $\mathbf{K}$. A morphism of $(B,\mathbf{K})$-modules is a $\mathbf{K}$-equivariant $B$-module homomorphism. We will write $M(B,\mathbf{K})$ for the abelian category of $(B,\mathbf{K})$-modules and $M^{\mathrm{fl}}(B,\mathbf{K})$ for the full subcategory of finite-length modules. Parabolic Induction {#sec:parabolicinduction} ------------------- Let $G$ be a real reductive group (for us this will mean the real points of a connected reductive algebraic group defined over $\mathbb{R}$). Fix a maximal compact subgroup $K \subset G$ and write $\theta: G \to G$ for the associated Cartan involution. Write $\mathbf{G}$ and $\mathbf{K}$ for the complexifications (of $G$ and of $K$) and $\mathfrak{g}$ and $\mathfrak{k}$ for their Lie algebras. Then $(\mathfrak{g},\mathbf{K})$ is a pair. If $\mathfrak{q} \subset \mathfrak{g}$ is a parabolic subalgebra and $\mathbf{Q} \subset \mathbf{G}$ is the corresponding (connected) subgroup of $\mathbf{G}$, then $(\mathfrak{q},\mathbf{Q} \cap \mathbf{K}) \subset (\mathfrak{g},\mathbf{K})$ is a subpair. We will define a left-exact functor $$\label{eqn:zuckermanfunctor} \mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}: M(\mathfrak{h},\mathbf{L}) \to M(\mathfrak{g},\mathbf{K})$$ This functor reduces to the usual functor of parabolic induction (when $\mathfrak{q}$ is germane), cohomological induction (when $\mathfrak{q}$ is $\theta$-stable), and real parabolic induction (when $\mathfrak{q}$ is the complexification of a real parabolic subalgebra). Roughly speaking, if $W \in M(\mathfrak{l},\mathbf{Q} \cap \mathbf{K})$, $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{l},\mathbf{Q}\cap \mathbf{K})}W$ is the $(\mathfrak{g},\mathbf{K})$-module $$\mathbf{K}-\text{finite vectors in } \mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W \otimes \det(\mathfrak{u}))$$ where $\det(\mathfrak{u})$ is the top exterior power of $\mathfrak{u}$, the nilradical of $\mathfrak{q}$. This definition is not quite correct (or meaningful, strictly speaking) if $\mathbf{K}$ is disconnected. We review the correct definition below. Let $\mathbf{K}^0$ denote the identity component of $\mathbf{K}$, and let $\mathbf{K}^1 = \mathbf{L}\mathbf{K}^0$. Then $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap\mathbf{K})}W$ is defined in stages 1. Let $$W' := W \otimes \det(\mathfrak{u})$$ 2. Form $$\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W')$$ This vector space has the structure of a $(\mathfrak{g}, \mathbf{Q} \cap \mathbf{K})$-module 3. Take $\mathbf{K}^0$-finite vectors $$\Gamma^0\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W') := \mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W')_{\mathbf{K}^0}$$ This vector space has the structure of a $\mathfrak{g}$-module with algebraic actions of $\mathbf{Q} \cap \mathbf{K}$ and $\mathbf{K}^0$. These group actions are both compatible with $\mathfrak{g}$, but not necessarily with eachother. They restrict to two (often distinct) actions of $\mathbf{Q} \cap \mathbf{K}^0$. 4. Form the subspace of $\Gamma^0\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W')$ on which both $\mathbf{Q} \cap \mathbf{K}^0$-actions coincide: $$\Gamma^1\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W') := \{v \in \Gamma^0\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W'): g \cdot_1 v = g \cdot_2 v \quad \forall g \in \mathbf{Q} \cap \mathbf{K}^0\}$$ This is a $\mathfrak{g}$-module with compatible algebraic actions of $\mathbf{K}^0$ and $\mathbf{Q} \cap \mathbf{K}$, and hence of $\mathbf{K}^1 = (\mathbf{Q} \cap \mathbf{K})\mathbf{K}^0$ 5. Perform a finite induction $$\Gamma \mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W'):= \mathrm{Ind}^{\mathbf{K}}_{\mathbf{K}^1} \Gamma^1\mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W')$$ This vector space has the structure of a $(\mathfrak{g},\mathbf{K})$-module. The assignment $$W \mapsto \Gamma \mathrm{Hom}_{\mathfrak{q}}(U(\mathfrak{g}),W')$$ defines a left-exact functor, since each of its consituents (described in $(1)-(5)$ above) are exact or left-exact. Since the category $M(\mathfrak{g},\mathbf{K})$ has enough injectives, we can define the right derived functors $$R^i\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}: M(\mathfrak{q},\mathbf{Q}\cap \mathbf{K}) \to M(\mathfrak{g},\mathbf{K})$$ If we fix a Levi decomposition $\mathfrak{q} = \mathfrak{l} \oplus \mathfrak{u}$ (and hence a Levi decomposition $\mathbf{Q} = \mathbf{L} \mathbf{U}$), then $(\mathfrak{l},\frac{\mathbf{Q} \cap \mathbf{K}}{\mathbf{U} \cap \mathbf{K}})$ is a pair (with $\mathbf{Q} \cap \mathbf{K}$ acting on $\mathfrak{l}$ via the isomorphism $\mathfrak{l} \cong \mathfrak{q}/\mathfrak{u}$), and there is a surjective morphism of pairs $$(\mathfrak{q}, \mathbf{Q} \cap \mathbf{K}) \twoheadrightarrow (\mathfrak{l},\frac{\mathbf{Q} \cap \mathbf{K}}{\mathbf{U} \cap \mathbf{K}})$$ Pulling back along this morphism defines a fully faithful embedding $$M(\mathfrak{l},\frac{\mathbf{Q} \cap \mathbf{K}}{\mathbf{U} \cap \mathbf{K}}) \subset M(\mathfrak{q}, \mathbf{Q} \cap \mathbf{K})$$ The restriction of $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}$ to this subcategory has many favorable properties, which we summarize below. \[thm:propertiesofI\] Let $W \in M(\mathfrak{l}, \frac{\mathbf{Q}\cap \mathbf{K}}{\mathbf{U} \cap \mathbf{K}})$ and let $\mathfrak{h} \subset \mathfrak{l}$ be a Cartan subalgebra. Then 1. If $W$ has finite-length, then $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}W$ has finite-length, for every $i \geq 0$ 2. If $W$ has infinitesimal character $\gamma_{\lambda}$, then $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}W$ has infinitesimal character $\gamma_{\lambda+\rho(\mathfrak{u})}$, for every $i \geq 0$ 3. There is an integer $s$ (depending only on $\mathfrak{q}$, $\mathbf{K}$, and $\mathfrak{g}$) such that $\mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}W = 0$ for every $i > s$ For proofs, we direct the reader to [@Vogan1981], Section 6.3. Vogan proves these statements under the assumption that $\mathfrak{q}$ is germane, but his arguments can be (slightly and easily) modified to accomodate our more general setting. In light of Theorem \[thm:propertiesofI\], we can define a group homomorphism $$I(\mathfrak{l},\mathfrak{q},\cdot): KM^{\mathrm{fl}}(\mathfrak{l}, \frac{\mathbf{Q}\cap \mathbf{K}}{\mathbf{U} \cap \mathbf{K}}) \to KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K}) \qquad I(\mathfrak{l},\mathfrak{q},[W]) = \sum_i (-1)^i [R^i \mathbf{I}^{(\mathfrak{g},\mathbf{K})}_{(\mathfrak{q},\mathbf{Q}\cap \mathbf{K})}W]$$ This homomorphism will be our primary tool for constructing unipotent representations. Associated Varieties -------------------- Let $\mathbf{G}$ be a complex reductive algebraic group and let $\mathfrak{g}$ be its Lie algebra. Write $\mathcal{N} \subset \mathfrak{g}$ for the $\mathbf{G}$ and $\mathbb{C}^{\times}$-invariant subset of nilpotent elements of $\mathfrak{g}$. $\mathbf{G}$ acts on $\mathcal{N}$ with finitely many orbits, which are called the *nilpotent orbits of* $\mathbf{G}$. If we fix a (non-degenerate, symmetric, bilinear $\mathbf{G}$-invariant) form $B(- , -)$ on $\mathfrak{g}$ (as we can, since $\mathfrak{g}$ is reductive), we get a $\mathbf{G}$-invariant isomorphism $\mathfrak{g} \cong \mathfrak{g}^*$. Let $\mathcal{N}^*$ be the image of $\mathcal{N}$ under this isomorphism. This set is independent of $B$—its elements are precisely the functionals $\lambda \in \mathfrak{g}^*$ which annihilate their centralizers in $\mathfrak{g}$. Now suppose $I \subset U(\mathfrak{g})$ is a two-sided ideal. If we equip $U(\mathfrak{g})$ with its usual filtration, then there is a canonical $\mathbf{G}$-invariant isomorphism (of graded commutative algebras) ${\operatorname{gr}}U(\mathfrak{g}) \cong S(\mathfrak{g})$. There is also an obvious isomoprhism $S(\mathfrak{g}) \cong \mathbb{C}[\mathfrak{g}^*]$. Hence, ${\operatorname{gr}}(I)$ is identified with a graded ideal in $\mathbb{C}[\mathfrak{g}^*]$. Make the following The *associated variety* of $I$ is the Zariski-closed subset of $\mathfrak{g}^*$ defined by the graded ideal ${\operatorname{gr}}(I) \subset \mathbb{C}[\mathfrak{g}^*]$ $$\begin{aligned} {\operatorname{AV}}(I) &:= V({\operatorname{gr}}(I))\\ &\subset \mathfrak{g}^*\end{aligned}$$ Since ${\operatorname{gr}}(I) \subset \mathbb{C}[\mathfrak{g}^*]$ is $\mathbf{G}$-invariant and graded, ${\operatorname{AV}}(I) \subset \mathfrak{g}^*$ is $\mathbf{G}$ and $\mathbb{C}^{\times}$-invariant. \[thm:finitecodimensionAV\] Suppose $I \cap Z(\mathfrak{g}) \subset Z(\mathfrak{g})$ is an ideal of finite codimension. Then ${\operatorname{AV}}(I) \subset \mathcal{N}^*$. In the setting of Theorem \[thm:finitecodimensionAV\], ${\operatorname{AV}}(I)$ is a finite union of $\mathbf{G}$-orbits on $\mathcal{N}^*$. If we write $\mathcal{O}_1,...,\mathcal{O}_n$ for the open $\mathbf{G}$-orbits on ${\operatorname{AV}}(I)$, then $\overline{\mathcal{O}}_i$ are the irreducible components of ${\operatorname{AV}}(I)$ and $${\operatorname{AV}}(I) = \bigcup \overline{\mathcal{O}}_i$$ When $I$ is primitive, ${\operatorname{AV}}(I)$ is irreducible. \[thm:Josephirreducibility\] If $I$ is primitive, there is a $\mathbf{G}$-orbit $\mathcal{O} \subset \mathcal{N}^*$ such that $${\operatorname{AV}}(I) = \overline{\mathcal{O}}$$ Unipotent Representations {#sec:unipotentrepresentations} ========================= Let $\mathbf{G}$ be a complex connected reductive algebraic group and let $\mathfrak{g}$ be its Lie algebra. Fix a Cartan subalgebra $\mathfrak{h} \subset \mathfrak{g}$ and write $\Lambda \subset \mathfrak{h}^*$ for the lattice of integral weights. If $I \subset U(\mathfrak{g})$ is a primitive ideal, then by Schur’s lemma the intersection $I \cap Z(\mathfrak{g})$ is a maximal ideal in $Z(\mathfrak{g})$ and hence the kernel of an infinitesimal character $\gamma_{\lambda}$ for some $\lambda \in \mathfrak{h}^*$ well-defined up to $W(\mathfrak{g})$. If $\lambda \in \mathfrak{h}^*$, write ${\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ for the set of primitive ideals in $U(\mathfrak{g})$ of infinitesimal character $\gamma_{\lambda}$. Primitive ideals in universal enveloping algebras have been studied extensively by Dixmier, Duflo, Joseph, Vogan, Barbasch, and others. We will need only the following handful of results (proofs can be found in [@Dixmier1974]). \[thm:primitiveideals\] Let $\lambda \in \mathfrak{h}^*$. Then 1. ${\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ is a finite set containing a unique minimal and unique maximal element. 2. If $I, J \in {\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ and there is a strict inclusion $I \subset J$, then there is a strict inclusion ${\operatorname{AV}}(J) \subset {\operatorname{AV}}(I)$. 3. If $I \subset U(\mathfrak{g})$ is any prime two-sided ideal of infinitesimal character $\lambda$, then $I \in \mathrm{Prim}^{\lambda}U(\mathfrak{g})$. In [@BarbaschVogan1985], Barbasch and Vogan use the following definition: \[def:unipotent\] Let $\lambda \in \mathbb{Q} \otimes_{\mathbb{Z}} \Lambda \subset \mathfrak{h}^*$. A primitive ideal $I \in {\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ is unipotent if the following conditions are satisfied 1. If $\mu \in \lambda + \Lambda$ and $J \in {\operatorname{Prim}}_{\mu}U(\mathfrak{g})$ satisfies $${\operatorname{AV}}(J) \subseteq {\operatorname{AV}}(I)$$ then $$\label{eqn:inequalityofinfchar} B(\mu, \mu) \geq B(\lambda,\lambda)$$ 2. If, in addition, \[eqn:inequalityofinfchar\] is an equality, then $\mu \in W(\mathfrak{g})\lambda$ and $${\operatorname{AV}}(J) = {\operatorname{AV}}(I)$$ In the following proposition, we catalog some of the elementary properties of unipotent ideals: \[prop:propsofunipotentideals\] For every $\lambda \in \mathfrak{h}^*$, there is at most one unipotent ideal in ${\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$. If $I \in {\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ is unipotent, then $I$ is a maximal ideal and for every $J \in {\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$, $J \neq I$, there is a proper inclusion $${\operatorname{AV}}(I) \subset {\operatorname{AV}}(J)$$ Choose a maximal ideal $J \subset U(\mathfrak{g})$ with $I \subseteq J$. Note that $J$ is primitive (in any associative algebra, every maximal ideal is primitive). The inclusion $I \subseteq J$ induces an inclusion ${\operatorname{AV}}(J) \subseteq {\operatorname{AV}}(I)$, which must be an equality by the second part of Definition \[def:unipotent\]. By part 2 of Theorem \[thm:primitiveideals\], this implies $I=J$. Hence, $I$ is a maximal ideal. By part 1 of Theorem \[thm:primitiveideals\], ${\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ contains a unique maximal element. In particular, if $J \in {\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$, $J \neq I$, then there is a strict inclusion $J \subset I$ and hence a strict inclusion ${\operatorname{AV}}(I) \subset {\operatorname{AV}}(J)$, as desired. We will say that an infinitesimal character $\gamma_{\lambda}$ is *unipotent* if ${\operatorname{Prim}}^{\lambda}U(\mathfrak{g})$ contains a unipotent ideal $I_{\lambda} \subset U(\mathfrak{g})$. In this case, $I_{\lambda}$ is unique (by Proposition \[prop:propsofunipotentideals\]) and is in fact the (unique) maximal ideal of infinitesimal character $\gamma_{\lambda}$. By Theorem \[thm:Josephirreducibility\], there is a (unique) $\mathbf{G}$-orbit $\mathcal{O} \subset \mathcal{N}^*$ such that $${\operatorname{AV}}(I_{\lambda}) = \overline{\mathcal{O}}$$ This defines a map from unipotent infinitesimal characters to $\mathbf{G}$-orbits on $\mathcal{N}^*$. Write ${\operatorname{unip}}(\mathcal{O})$ for the fiber of this map over a $\mathbf{G}$-orbit $\mathcal{O} \subset \mathcal{N}^*$. Usually, ${\operatorname{unip}}(\mathcal{O})$ will be infinite. If we have in mind a pair $(\mathfrak{g},\mathbf{K})$, we will write ${\operatorname{Unip}}^{\lambda}(\mathcal{O})$ for the set of (isomorphism classes of) irreducible $(\mathfrak{g},\mathbf{K})$-modules $X$ with $\mathrm{Ann}(X) = I_{\lambda}$, and $${\operatorname{Unip}}(\mathcal{O}) := \bigsqcup_{\lambda \in {\operatorname{unip}}(\mathcal{O})} {\operatorname{Unip}}^{\lambda}(\mathcal{O})$$ Special Unipotent Representations --------------------------------- In [@BarbaschVogan1985], Barbasch and Vogan study a distinguished class of unipotent representations related to the Arthur Conjectures ([@Arthur1983],[@Arthur1989]) called (by Barbasch and Vogan) the *special* unipotent representations. For completeness, we will review their definition below. In the setting of Section \[sec:unipotentrepresentations\], write $\mathbf{G}^{\vee}$ for the Langlands dual of $\mathbf{G}$ and $\mathfrak{g}^{\vee}$ for its Lie algebra. By construction, $\mathfrak{g}^{\vee}$ contains a distinguished Cartan subalgebra $\mathfrak{h}^{\vee}$ which is naturally identified with $\mathfrak{h}^*$. There is an order-reversing map $$\psi: \mathcal{N}^{\vee}/\mathbf{G}^{\vee} \to \mathcal{N}/\mathbf{G}$$ first defined by Spaltenstein ([@Spaltenstein1982]). If $\mathcal{O}^{\vee} \subset \mathcal{N}^{\vee}$ is a nilpotent $\mathbf{G}^{\vee}$-orbit, we can find a Lie algebra homomorphism (far from unique) $$\phi_{\mathcal{O}^{\vee}}: \mathfrak{sl}_2(\mathbb{C}) \to \mathfrak{g}^{\vee} \qquad \phi(E) \in \mathcal{O}^{\vee}$$ Then $d^{\vee} := \frac{1}{2}\phi_{\mathcal{O}^{\vee}}(D)$ is a semisimple element of $\mathfrak{g}^{\vee}$. Conjugating by $\mathbf{G}^{\vee}$ if necessary, we can assume that $d^{\vee} \in \mathfrak{h}^{\vee} \cong \mathfrak{h}^*$. This element of $\mathfrak{h}^*$ is well-defined modulo $W(\mathfrak{g})$ and therefore determines a character $$\gamma_{\mathcal{O}^{\vee}} := \gamma_{d^{\vee}}: Z(\mathfrak{g}) \to \mathbb{C}$$ by the Harish-Chandra isomorphism. If $\mathcal{O} \subset \mathcal{N}$ is a $\mathbf{G}$-orbit, let $$\mathrm{arth}(\mathcal{O}) := \{\gamma_{\mathcal{O}^{\vee}}: \psi(\mathcal{O}^{\vee}) = \mathcal{O}\}$$ This is a finite (sometimes empty) set of small (often singular) infinitesimal characters associated to $\mathcal{O}$. If $\mathcal{O}'$ is a $\mathbf{G}$-orbit on $\mathcal{N}^*$, the set $\mathrm{arth}(\mathcal{O}')$ is defined by first replacing $\mathcal{O}'$ with its image in $\mathcal{N}$ under the $\mathbf{G}$-invariant identification $B:\mathfrak{g} \cong \mathfrak{g}^*$. Barbasch and Vogan prove For every $\mathbf{G}$-orbit $\mathcal{O} \subset \mathcal{N}^*$, there is an inclusion $$\mathrm{arth}(\mathcal{O}) \subseteq {\operatorname{unip}}(\mathcal{O})$$ If we have a pair $(\mathfrak{g},\mathbf{K})$ in mind, we will write $${\operatorname{Unip}}^{\mathrm{arth}}(\mathcal{O}) := \bigsqcup_{\lambda \in \mathrm{arth}(\mathcal{O})} {\operatorname{Unip}}^{\lambda}(\mathcal{O})$$ If $G = SL_2(\mathbb{R})$, there are two $\mathbf{G}$-orbits on $\mathcal{N}^*$: $\{0\}$ and the principal nilpotent orbit, $\mathcal{O}^{\mathrm{prin}}$. Infinitesimal characters are parameterized by (weakly) dominant weights $\lambda \in \mathfrak{h}^*$. In standard coordinates $${\operatorname{unip}}(\{0\}) = \{1\} \qquad {\operatorname{unip}}(\mathcal{O}^{\mathrm{prin}}) = [0,\frac{1}{2}] \cap \mathbb{Q}$$ and $$\mathrm{arth}(\{0\}) = \{1\} \qquad \mathrm{arth}(\mathcal{O}^{\mathrm{prin}}) = \{0\}$$ Upper Triangularity for Unipotent Representations {#sec:uppertriangularity} ================================================= Let $\mathfrak{q}$ be a parabolic subalgebra of $\mathfrak{g}$. Fix a Levi decomposition $\mathfrak{q} = \mathfrak{l}\oplus \mathfrak{u}$ and a nilpotent $\mathbf{L}$-orbit $\mathcal{O}_{\mathfrak{l}} \subset \mathcal{N}_{\mathfrak{l}}$. Lusztig and Spaltenstein observed ([@LusztigSpaltenstein1979]) that the set $\mathbf{G} \cdot \left(\overline{\mathcal{O}}_{\mathfrak{l}} + \mathfrak{u}\right) \subset \mathcal{N}_{\mathfrak{g}}$ contains a unique open $\mathbf{G}$-orbit $\mathcal{O}_{\mathfrak{g}} \subset \mathcal{N}_{\mathfrak{g}}$, which (they proved) depends only on $\mathfrak{l}$ and $\mathcal{O}_{\mathfrak{l}}$. The correspondence $\mathcal{O}_{\mathfrak{l}} \mapsto \mathcal{O}_{\mathfrak{g}}$ is called the *induction of nilpotent orbits* and is denoted by $$\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}}\mathcal{O}_{\mathfrak{l}}$$ There is a parallel notion of induction for $(\mathfrak{g},\mathbf{K})$-modules. In Section \[sec:parabolicinduction\], we defined the category $M^{\mathrm{fl}}(\mathfrak{l},\frac{\mathbf{Q}\cap \mathbf{K}}{\mathbf{U}\cap \mathbf{K}})$ and the group homomorphism $$I(\mathfrak{l},\mathfrak{q},\cdot): KM^{\mathrm{fl}}(\mathfrak{l},\frac{\mathbf{Q}\cap \mathbf{K}}{\mathbf{U}\cap \mathbf{K}}) \to KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$$ In Section \[sec:unipotentrepresentations\], we defined sets ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{l}})$ and ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ consisting of irreducible $(\mathfrak{l},\frac{\mathbf{Q}\cap \mathbf{K}}{\mathbf{U}\cap \mathbf{K}})$-modules and $(\mathfrak{g},\mathbf{K})$-modules, respectively. The following question arises immediately \[question:induction\] What is the relationship between ${\operatorname{Unip}}(\mathcal{O}_{\mathfrak{g}})$ and $I(\mathfrak{l},\mathfrak{q},{\operatorname{Unip}}(\mathcal{O}_{\mathfrak{l}}))$? A bit of thought reveals that this is not the right question to ask. The orbit $\mathcal{O}_{\mathfrak{g}}$ can be obtained by induction from a whole range of parabolics, of which $\mathfrak{q}$ is only one. Each of these parabolics contributes its own set of induced representations, which should be taken into account. To fix (and ultimately answer) Question \[question:induction\], we will consider the partial flag variety $\mathcal{Q}$ of parabolic subalgebras, $\mathbf{G}$-conjugate to $\mathfrak{q}$. If $\mathfrak{q}' \in \mathcal{Q}$, there is a group element $g \in \mathbf{G}$ such that ${\operatorname{Ad}}(g)\mathfrak{q} = \mathfrak{q}'$. If we fix a Levi decomposition $\mathfrak{q}' = \mathfrak{l}'\oplus \mathfrak{u}'$ and require that ${\operatorname{Ad}}(g)\mathfrak{l}=\mathfrak{l}'$, then $g$ is unique up to right multiplication by $\mathbf{L}$ and left multiplication by $\mathbf{L}'$. In particular, ${\operatorname{Ad}}(g)$ induces a *canonical* isomorphism $Z(\mathfrak{l}) \cong Z(\mathfrak{l}')$. If $\gamma$ is a character of $Z(\mathfrak{l})$, we will write $\gamma'$ for the corresponding character of $Z(\mathfrak{l}')$. \[def:degenerate\] Let $\gamma \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{l}})$ and denote by $X^{\vee}$ the contragradient of a $(\mathfrak{g},\mathbf{K})$-module $X$ (or its class). A degenerate representation for the data $(\mathfrak{q},\gamma)$ is a class in $KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$ of the form $$I(\mathfrak{l}',\mathfrak{q}',W^{\vee})^{\vee}$$ where $\mathfrak{q}' \in \mathcal{Q}$, $\mathfrak{l}' \subset \mathfrak{q}'$ is a Levi subalgebra, and $W$ is an irreducible $(\mathfrak{l}',\frac{\mathbf{Q}'\cap \mathbf{K}}{\mathbf{U}'\cap \mathbf{K}})$-module with annihilator equal to the unipotent ideal $I_{\mu'} \subset U(\mathfrak{l}')$ (cf Section \[sec:unipotentrepresentations\]). Write ${\operatorname{Deg}}^{\gamma}(\mathfrak{q})$ for the set of all such representations. Fix, once and for all, a Cartan subalgebra $\mathfrak{h} \subset \mathfrak{l}$ and a positive system $\Delta^+(\mathfrak{g},\mathfrak{h})$ compatible with $\mathfrak{q}$. Suppose the infinitesimal character $\gamma$ in Definition \[def:degenerate\] is represented by a functional $\mu \in \mathfrak{h}^*$. Then a degenerate representation $[M] \in {\operatorname{Deg}}^{\gamma_{\mu}}(\mathfrak{q})$ has infinitesimal character $\gamma_{\mu - \rho(\mathfrak{u})}$. Suppose that $\gamma_{\mu-\rho(\mathfrak{u})} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{g}})$. A more reasonable formulation of Question \[question:induction\] is the following What is the relationship between ${\operatorname{Unip}}^{\mu-\rho(\mathfrak{u})}(\mathcal{O}_{\mathfrak{g}})$ and ${\operatorname{Deg}}^{\mu}(\mathfrak{q})$? Our main result provides an answer to this question under some conditions on $\mu$ and on $\mathcal{O}_{\mathfrak{g}}$: \[thm:unipotentsanddegenerates\] Let $\mu \in \mathfrak{h}^*$. Suppose 1. $\gamma_{\mu} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{l}})$ 2. $\gamma_{\mu - \rho(\mathfrak{u})} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{g}})$ 3. $\mu - \rho(\mathfrak{u})$ is antidominant for $\mathfrak{u}$ 4. The moment map $$\eta: \mathbf{G} \times_{\mathbf{Q}} \left(\overline{\mathcal{O}}_{\mathfrak{l}} + \mathfrak{u}\right) \to \overline{\mathcal{O}}_{\mathfrak{g}}$$ is birational Then the sets ${\operatorname{Unip}}^{\mu-\rho(\mathfrak{u})}(\mathcal{O}_{\mathfrak{g}})$ and ${\operatorname{Deg}}^{\mu}(\mathfrak{q})$ have the same $\mathbb{Z}$-span in $KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$ and are related therein by an upper triangular matrix with $\pm 1$’s along the diagonal. $\mathcal{D}$-modules on $\mathcal{Q}$ {#sec:Dmodules} -------------------------------------- We begin by recalling some basic facts from the theory of algebraic $\mathcal{D}$-modules. For details and proofs, see [@BeilinsonBernstein1981] or [@HechtMilicicSchmidWolf]. Let $X$ be a smooth complex algebraic variety. Let $\mathcal{O}_X$ be its structure sheaf and let $\mathcal{D}_X$ be its sheaf of (local, algebraic) differential operators. A twisted sheaf of differential operators (a TDO, for short) is any sheaf of algebras on $X$ which is locally isomorphic to $\mathcal{D}$. If $\mathcal{D}'_X$ is a TDO, we can consider the abelian category $\mathcal{M}(\mathcal{D}'_X)$ of coherent (left) $\mathcal{D}'_X$-modules. Every object $\mathcal{M} \in \mathcal{M}(\mathcal{D}'_X)$ has a good filtration (by quasi-coherent subsheaves) and ${\operatorname{gr}}(\mathcal{M})$ is a coherent sheaf on the cotangent bundle $T^*X$. The support of this sheaf is a closed, conical subset $\mathrm{CV}(\mathcal{M}) \subset T^*X$ called the characteristic variety of $\mathcal{M}$. It is independent of the filtration used to define it. If we denote the support of $\mathcal{M}$ (in the usual sense) by $\mathrm{Supp}(\mathcal{M})$, then $\mathrm{Supp}(\mathcal{M})$ is the image of $\mathrm{CV}(\mathcal{M})$ under the projection map $T^*X \to X$. If the characteristic variety of $\mathcal{M}$ has the same dimension as $X$ (i.e. half the dimension of $T^*X$), we say that $\mathcal{M}$ is holonomic. Holonomic $\mathcal{D}'_X$-modules have finite length. They form a subcategory of $\mathcal{M}(\mathcal{D}_X')$ which we will denote by $\mathcal{M}^{\mathrm{hol}}(\mathcal{D}'_X)$. If it is the sheaf of sections of a finite-rank vector bundle and its characteristic variety is the zero section of $T^*X$. In particular, $\mathcal{M}$ is holonimic. Now suppose $f: Y \hookrightarrow X$ is an open or closed embedding of smooth complex algebraic varieties. The restriction $f^*\mathcal{D}'_X$ of $\mathcal{D}'_X$ to $Y$ (in the sense of quasi-coherent sheaves) is a TDO, which we will denote by $\mathcal{D}'_Y$. There are two pullbacks $$f^*,f^!: \mathcal{M}^{\mathrm{hol}}(\mathcal{D}'_X) \to \mathcal{M}^{\mathrm{hol}}(\mathcal{D}'_Y)$$ and two pushforwards $$f_*,f_!: \mathcal{M}^{\mathrm{hol}}(\mathcal{D}'_Y) \to \mathcal{M}^{\mathrm{hol}}(\mathcal{D}'_X)$$ In general, $f_*$ is right adjoint to $f^*$ and $f_!$ is left adjoint to $f^!$. In particular, the functors $f_*$ and $f^!$ are left exact and the functors $f^*$ and $f_!$ are right exact. If $f$ is an open embedding, then 1. $f^*=f^!$ is the usual (exact) restriction functor (of quasi-coherent sheaves). 2. $f_*$ is the usual pushforward (of quasi-coherent sheaves) If $f$ is a closed embedding, then 1. $f_*=f_!$ is an (exact) embedding of categories. 2. $f^!$ is the usual restriction functor (of quasi-coherent sheaves) 3. If $\mathcal{M}_Y(\mathcal{D}'_X)$ is the subcategory of $\mathcal{M}(\mathcal{D}'_X)$ consisting of objects supported in $Y$, then $f_*=f_!$ defines an equivalence $\mathcal{M}(\mathcal{D}'_Y) \cong \mathcal{M}(\mathcal{D}'_X)$ with inverse given by $f^!$. This fact, due to Kashiwara ([@Kashiwara1983]) is somewhat deeper than the rest. If $\mathbf{K}$ is an algebraic group acting (algebraically) on $X$, then $\mathcal{D}$ is a sheaf of complex algebras with $\mathbf{K}$-action. A $\mathbf{K}$-equivariant TDO is a sheaf of algebras with $\mathbf{K}$-action which is locally isomorphic to $\mathcal{D}$. If $\mathcal{D}'_X$ is such a sheaf, we can form the abelian category $\mathcal{M}(\mathcal{D}'_X,\mathbf{K})$ of $\mathbf{K}$-equivariant coherent $\mathcal{D}'_X$-modules. If $\mathbf{K}$ acts on $X$ with finitely many orbits, then objects in this category are automatically holonomic (and hence automatically finite-length). The four functors defined above have $\mathbf{K}$-equivariant analogs which satisfy ($\mathbf{K}$-equivariant analogs of) properties $(1)-(5)$. Return now to the setting of Theorem \[thm:unipotentsanddegenerates\]. The $\mathbf{K}$-equivariant TDOs on $\mathcal{Q}$ are parameterized by one-dimensional representations of $\mathfrak{l}$. We will write $\mathcal{D}_{\mathcal{Q}}^{\lambda}$ for the TDO corresponding to a representation $\lambda$. If $\lambda$ is the trivial representation, then $\mathcal{D}_{\mathcal{Q}}^{\lambda} = \mathcal{D}_{\mathcal{Q}}$. If $\lambda$ integrates to a character of $\mathbf{L}$, then $\mathcal{D}_{\mathcal{Q}}^{\lambda} = \mathcal{L}^{\lambda} \otimes \mathcal{D}_{\mathcal{Q}} \otimes (\mathcal{L}^{\lambda})^{-1}$, where $\mathcal{L}^{\lambda}$ is (the sheaf of sections of) the $\mathbf{G}$-equivariant line bundle with fiber $\lambda$ over $\mathfrak{q} \in \mathcal{Q}$. For a definition of $\mathcal{D}_{\mathcal{Q}}^{\lambda}$ in the general case, see e.g. [@HechtMilicicSchmidWolf]. For any $\lambda$, $\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$ acquires the structure of an algebraic $\mathbf{G}$-module. An element $X \in \mathfrak{g}$ determines a right-invariant vector field $\xi_X \in \Gamma(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathcal{Q})$, and the correspondence $X \mapsto \xi_X$ induces a $\mathbf{G}$-equivariant map $$\phi: U(\mathfrak{g}) \to \Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$$ which turns $\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$ into a Dixmier algebra for $\mathbf{G}$. We can consider the category $M(\Gamma(\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$ of $\mathbf{K}$-equivariant modules for the Dixmier algebra $\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$ (cf Definition \[def:dixmiermodule\]). Taking global sections defines a left-exact functor $$\label{eqn:globalsections} \Gamma': \mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K}) \to M(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$$ Pulling back along $\phi$ defines an exact functor $$\label{eqn:pullbackfunctor} \phi^* :M(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K}) \to M(\mathfrak{g},\mathbf{K})$$ We reserve the symbol $\Gamma$ for the composition $\phi^* \circ \Gamma'$. A module in the image of this functor has infinitesimal character $\gamma_{\lambda - \rho_{\mathfrak{g}}}$. When $\lambda- \rho_{\mathfrak{g}}$ is strictly antidominant, the functor $\Gamma'$ is particularly well-behaved. The following theorem was proved in the special case $\mathfrak{q}=\mathfrak{b}$ by Beilinson and Bernstein ([@BeilinsonBernstein1981]) and in the general case by Kitchen ([@Kitchen2011]) Suppose $\lambda -\rho_{\mathfrak{g}}$ is strictly antidominant, i.e. $$\label{eqn:weaklyantidominant} \langle \lambda -\rho_{\mathfrak{g}}, \alpha^{\vee}\rangle < 0 \qquad \forall \alpha \in \Delta^+(\mathfrak{g},\mathfrak{h})$$ Then the functor $$\Gamma': \mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K}) \to M(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$$ is an exact equivalence of categories. The work of Beilinson and Bernstein also leads to a classification of the irreducible objects in $M(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ (without restrictions on $\lambda$). A *standard datum* for ($\mathfrak{q}$, $\lambda$) is a pair $(Z, \tau)$ consisting of a $\mathbf{K}$-orbit $Z \subset \mathcal{Q}$ and a $\mathbf{K}$-equivariant line bundle $\tau$ on $Z$ which is suitably compatible with $\lambda$. To describe this compatibility, choose a point $\mathfrak{q}' \in Z$. The fiber of $\tau$ over $\mathfrak{q}'$ is a character $\tau_{\mathfrak{q}'}$ of $\mathbf{Q}'\cap \mathbf{K}$. To be compatible with $\lambda$, the differential of this character should coincide with the character of $\mathfrak{l}'$ determined by $\lambda$. This condition on $d\tau_{\mathfrak{q}'}$ is precisely what is required to define on $\tau$ the structure of a $\mathbf{K}$-equivariant $\mathcal{D}^{\lambda}_Z$-module. If $\mathcal{M}$ is an irreducible $(\mathcal{M}^{\lambda}_{\mathcal{Q}},\mathbf{K})$-module, then $\mathrm{Supp}(\mathcal{M})$ is an irreducible subset of $\mathcal{Q}$ and, consequently, the closure in $\mathcal{Q}$ of a unique $\mathbf{K}$-orbit $Z$. Let $\mathcal{Q}' = \mathcal{Q} \setminus \partial Z$. Write $$j: \mathcal{Q}' \hookrightarrow \mathcal{Q} \qquad i: Z \hookrightarrow \mathcal{Q}'$$ for the open and closed embeddings and let $k = j \circ i$. By Kashiwara’s equivalence, there is a unique $(\mathcal{D}^{\lambda}_Z,\mathbf{K})$-module $\tau = k^!\mathcal{M}$ such that $i_!\tau = j^!\mathcal{M}$. On the other hand, $k_!\tau$ contains a unique irreducible $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-submodule, which is isomorphic to $\mathcal{M}$. This proves \[thm:bbequivalence1\] The map $$\mathcal{M} \mapsto (Z,\tau)$$ characterized by the requirements $\mathrm{Supp}(\mathcal{M}) = \overline{Z}$ and $\tau = k^!\mathcal{M}$ defines a bijective correspondence between irreducible $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-modules and standard data for ($\mathfrak{q}$, $\lambda$). The inverse map takes the standard datum $(Z,\tau)$ to the unique irreducible $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-submodule of $k_!\tau$. To elaborate on the relationship between irreducibles in $\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ and standard data for $(\mathfrak{q},\lambda)$, we will pass to the Grothendieck group $K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$. For every $p \geq 0$, we can define the left-derived functor $$L^pk_!: \mathcal{M}(\mathcal{D}_Z^{\lambda},\mathbf{K}) \to \mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$$ Furthermore, $L^pk_!\tau = 0$ for $p >>0$. Hence, we can define $$I(Z,\tau) := \sum_p (-1)^p [L^pk_!\tau] \in K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$$ We will call this class the standard $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-module associated to $(Z,\tau)$. \[thm:bbequivalence2\] The irreducible and standard $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-modules are related in $K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ by an upper triangular matrix with $\pm 1$’s along the diagonal. More precisely: 1. Both sets form $\mathbb{Z}$-bases for $K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ 2. There is an ordering of the irreducible modules $\mathcal{M}_1,...,\mathcal{M}_n$ so that $$\label{eqn:uppertriangularity}[\mathcal{M}_p] = \pm I(Z_p,\tau_p) + \sum_{q >p} c_{pq} I(Z_q,\tau_q) \qquad c_{pq} \in \mathbb{Z}$$ where $(Z_p,\tau_p)$ is the standard datum associated to $\mathcal{M}_p$ under the bijection described in Theorem \[thm:bbequivalence1\]. Choose an ordering $(Z_1,\tau_1), ..., (Z_n,\tau_n)$ of the standard data for $(\mathfrak{q},\lambda)$ so that $$(Z_k,\tau_k) \leq (Z_l,\tau_l) \Leftrightarrow \dim(Z_k) \geq \dim(Z_l)$$ and induce an ordering $\mathcal{M}_1,...,\mathcal{M}_n$ on the irreducible objects using the bijection of Theorem \[thm:bbequivalence1\]. Since $\mathbf{K}$ acts on $\mathcal{Q}$ with finitely many orbits, the $\mathbf{K}$-orbits of minimal dimension are necessarily closed. If $(Z_p,\tau_p)$ is a standard datum and $Z_p$ is closed, then by Kashiwara’s theorem $I(Z_p,\tau_p) = [k_!\tau_p]$ is the class of an irreducible $(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$-module. In particular $$[\mathcal{M}_p] = I(Z_p,\tau_p)$$ as desired. Now let $(Z_p,\tau_p)$ be a standard datum for $(\mathfrak{q},\lambda)$ with the property that Equation \[eqn:uppertriangularity\] holds for every $(Z_q,\tau_q)$ with $\dim(Z_q) < \dim(Z_p)$. We will show that Equation \[eqn:uppertriangularity\] holds for $(Z_p,\tau_p)$. By Kashiwara’s theorem, we can assume that $Z_p$ is an open subset of $\mathcal{Q}$. Write $l: \partial Z \hookrightarrow \mathcal{Q}$ for the closed embedding (recall that $i=j=k: Z \hookrightarrow \mathcal{Q}$ is the open embedding). There is a distinguished triangle $$k_!k^!\mathcal{M}_p \to \mathcal{M}_p \to l_*l^*\mathcal{M}_p$$ in the bounded derived category $D^b\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$. Taking Euler characteristics, we obtain an equation in $K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ $$[\mathcal{M}_p] = \chi k_!k^! \mathcal{M}_p + \chi l_*l^*\mathcal{M}_p$$ By definition, $\chi k_!k^!\mathcal{M}_p = I(Z_p,\tau_p)$. Furthermore $$k^*l_*l^*\mathcal{M}_k = k^!l_*l^*\mathcal{M}_k = 0$$ Hence, $\chi l_*l^*\mathcal{M}_p$ is contained in $K\mathcal{M}_{\partial Z_p}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$, and is therefore a linear combination of classes $I(Z_q,\tau_q)$ with $q > p$. This proves $(2)$. $(1)$ follows immediately. Even when $\Gamma'$ fails to be exact, there is a well-defined class $$\Gamma' I(Z,\tau):= \sum_p (-1)^p [\Gamma L^pk_!\tau] \in KM^{\mathrm{fl}}(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$$ Up to a natural duality, this is exactly the class defined in Section \[sec:parabolicinduction\]. \[thm:hmswduality\] Let $(Z,\tau)$ be a standard datum for $(\mathfrak{q},\lambda)$. Choose a parabolic subalgebra $\mathfrak{q}' \in Z$ and a Levi decomposition $\mathfrak{q}' = \mathfrak{l}' \oplus \mathfrak{u}'$. There is an equality in $KM^{\mathrm{fl}}(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$ $$\Gamma I(Z,\tau) = \pm I(\mathfrak{l}',\mathfrak{q}', \tau_{\mathfrak{q}'}^{\vee})^{\vee}$$ In the special case $\mathfrak{q}=\mathfrak{b}$, this is a direct consequence of the duality theorem of Hecht, Milicic, Schmid, and Wolf ([@HechtMilicicSchmidWolf]). The general case was handled by Kitchen (Theorem 1.2, [@Kitchen2011]). Translation Functors {#sec:translationfunctors} -------------------- In this section, we will define translation functors in a very general setting. This approach will illuminate some general properties of these functors which we will need in Section \[sec:inducedmainresults\]. Peirce Decompositions {#sec:peirce} --------------------- Let $A$ be an associative ring with unit. An idempotent $e \in A$ is any element satisfying $e^2=e$. Two idempotents $e,e' \in A$ are orthogonal if $ee' = e'e = 0$. Suppose $\{e_i\}_{i=1}^n \subset A$ is a collection of mutually orthogonal idempotents and assume $$\sum_{i=1}^n e_i = 1$$ If we write $A_{ij}$ for the subgroup $e_iAe_j \subset A$, there is a decomposition $$A = 1A1 = \left(\sum_{i=1}^ne_i\right)A\left(\sum_{i=1}^ne_i\right) = \sum_{i,j} A_{ij}$$ which is direct by the orthogonality condition on $\{e_i\}$. This is called the Perice decomposition of $A$ associated to the set $\{e_i\}$. The pieces $A_{ij}$ satisfy the obvious relations $$A_{ij}A_{kl} \subseteq \delta_{jk} A_{il}$$ In particular, each $A_{ii}$ is a ring (although not a subring of $A$. Its unit $e_i$ is *not* the unit of $A$). If $M$ is an $A$-module (all modules will be left modules unless otherwise noted), then $e_iM$ is an $A_{ii}$-module and there is a decomposition $$\label{eqn:decompofM} M = \bigoplus_i e_iM$$ The assignment $M \mapsto e_iM$ defines a covariant functor $$P_i: A-\mathrm{mod} \to A_{ii}-\mathrm{mod} \qquad P_iM = e_iM$$ which is exact by \[eqn:decompofM\]. Define a functor in the opposite direction $$Q_i: A_{ii}-\mathrm{mod} \to A-\mathrm{mod} \qquad Q_iN = A \otimes_{A_{ii}} N$$ \[prop:QiPi\] In the setting described above, $Q_i$ is right inverse and right adjoint to $P_i$. Note that $$Q_iN = A \otimes_{A_{ii}} N = Ae_i \otimes_{A_{ii}}N$$ Hence, $$P_iQ_iN = e_i(Ae_i \otimes_{A_{ii}} N) = A_{ii} \otimes_{A_{ii}} N = N$$ This proves that $Q_i$ is right inverse to $P_i$. The functor $P_i$ is naturally equivalent to $\mathrm{Hom}_A(Ae_i, \cdot)$. The equivalence is implemented by the natural transformation $$\eta_M: \mathrm{Hom}_A(Ae_i, M) \to e_iM \qquad \eta_M(\varphi) = \varphi(e_i)$$ with inverse $$\eta_M^{-1}: e_iM \to \mathrm{Hom}_A(Ae_i,M) \qquad \eta_M^{-1}(e_im)(ae_i) = ae_im$$ Hence the asserted adjunction is a special case of the the usual tensor-hom adjunction (for the $A-A_{ii}$-bimodule $Ae_i$). We will also need \[prop:uniquesimplequotient\] In the setting of Proposition \[prop:QiPi\], let $N$ be a nonzero simple $A_{ii}$-module. Then the $A$-module $Q_iN$ has a unique simple quotient, $\tilde{Q}_iN$, and there is an isomorphism of $A_{ii}$-modules $P_i\tilde{Q}_iN \cong N$. Left multiplication by $e_i$ defines an $A_{ii}$-linear endomorphism $s_i$ of $Q_iN = A \otimes_{A_{ii}} N$ with image equal to $N$. Let $L$ be a proper $A$-submodule of $Q_iN$, and suppose $s_i(L) \neq 0$. Then by the simplicity of $N$, $s_i(L) = N$ and therefore $N \subseteq L$. But $N$ generates $Q_iN$ as an $A$-module, so in fact $Q_iN \subseteq L$ which contradicts the properness of $L \subset Q_iN$. Hence, $L \subseteq \ker{s_i}$. Let $J \subset Q_iN$ be the sum of all proper submodules of $Q_iN$. By the argument above, $J \subseteq \ker{s_i}$. In particular, $J$ is the unique maximal proper submodule of $Q_iN$ and $e_iJ = 0$. Let $$\tilde{Q}_iN := Q_iN/J$$ Then $\tilde{Q}_iN$ is the unique simple quotient of $Q_iN$ and $$P_i\tilde{Q}_iN = e_i(Q_iN/J) \cong e_i(Q_i) = N$$ as desired. Translation Functors: Abstract Definition {#sec:abstracttranslation} ----------------------------------------- Now let $B$ be a (complex) associative algebra and let $F$ be a finite-dimensional vector space. Let $A = B \otimes_{\mathbb{C}} \mathrm{End}(F)$ and choose a family $\{e_i\} \subset A$ of mutually orthogonal idempotents with sum equal to $1$. Write $$A = \bigoplus_{ij} A_{ij}$$ for the associated Peirce decomposition of $A$. If $M$ is a $B$-module, then $M \otimes_{\mathbb{C}} F$ is an $A$-module. The assignment $M \mapsto M \otimes_{\mathbb{C}}F$ defines a covariant functor $$t_F: B-\mathrm{mod} \to A-\mathrm{mod} \qquad t_F(M) = M \otimes_{\mathbb{C}}F$$ It is a standard fact that $t_F$ is an (exact) equivalence of categories, although its inverse is a bit harder to describe. Choose an idempotent $e_i \in A$. Let $T_{F,i}B := A_{ii}$. The abstract translation functor associated to the data of $B,F$, and $e_i$, is the composite $$T_{F,i} := P_i \circ t_F: B-\mathrm{mod} \to T_{F,i}B-\mathrm{mod}$$ Since both $P_i$ and $t_F$ are exact, so is $T_{F,i}$. If $N$ is a nonzero simple $T_{F,i}B$-module, there is a canonically defined $B$-module $t_F^{-1}\tilde{Q}_iN$. By Proposition \[prop:uniquesimplequotient\], this $B$-module is simple and maps to $N$ under $T_{F,i}$. Translation Functors: Dixmier Algebras {#sec:translationdixmier} -------------------------------------- We will apply the elementary results of Sections \[sec:peirce\] and \[sec:abstracttranslation\] in the following special case: - $I \subset U(\mathfrak{g})$ is a two-sided ideal such that $I \cap Z(\mathfrak{g})$ has finite codimension in $Z(\mathfrak{g})$ - $B = U(\mathfrak{g})/I$ - $F$ is a finite-dimensional representation of $\mathbf{G}$ - $A = B \otimes \mathrm{End}(F)$ By Propositions \[prop:dixmier1\] and \[prop:dixmier2\], both $B$ and $A$ are Dixmier algebras. Condition \[cond:dixmier3\] of Definition \[def:dixmier\] implies that $A$ is $Z(\mathfrak{g})\times Z(\mathfrak{g})$-finite. Hence, $A$ admits a Peirce decomposition by generalized left and right infinitesimal characters: $$A = \bigoplus_{\mu\nu} A_{\mu \nu} \qquad A_{\mu \nu} := \{a \in A: \mathfrak{m}_{\mu}^pa = a\mathfrak{m}_{\nu}^q = 0 \text{ for } p,q >>0\}$$ where $\mathfrak{m}_{\mu}$ is the kernel in $Z(\mathfrak{g})$ of the infinitesimal character $\gamma_{\mu}$. If we fix $\mu$ appearing in the decomposition above, we get (as in Section \[sec:abstracttranslation\]) an algebra $T_{F,\mu}B = A_{\mu\mu}$ (which is in fact a Dixmier algebra, since $B$ is) and a functor $$T_{F,\mu}: B-\mathrm{mod} \to F_{F,\mu}B-\mathrm{mod}$$ As in Section \[sec:abstracttranslation\], $T_{F,\mu}$ is an exact functor and every nonzero simple $T_{F,\mu}B$-module has a canonically defined simple preimage. If we introduce a $\mathbf{K}$-action, almost everything remains true. $T_{F,\mu}$ defines a functor $$T_{F,\mu}: M(B,\mathbf{K}) \to M(T_{F,\mu}B,\mathbf{K})$$ Like its non-equivariant counterpart, this functor is exact and every nonzero simple $(T_{F,\mu}B,\mathbf{K})$-module has a canonical simple preimage. Main Results {#sec:inducedmainresults} ------------ In this section, we analyze the effect of the global sections functor $\Gamma: \mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K}) \to M(\mathfrak{g},\mathbf{K})$ on standard modules $I(Z,\tau)$ in the case of unipotent infinitesimal character. To simplify the arguments, we will assume that $\mathcal{O}_{\mathfrak{g}}$ is induced from $\{0\} \subset \mathfrak{l}$. The general case is proved by a similar argument. Our first observation is the following \[prop:phiforunipotentlambda\] Let $\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}} \{0\}$ and suppose 1. $\gamma_{\lambda -\rho_{\mathfrak{g}}} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{g}})$ 2. The moment map $$\eta: T^*\mathcal{Q} \to \overline{\mathcal{O}}_{\mathfrak{g}}$$ is birational Then the map $$\phi: U(\mathfrak{g}) \to \Gamma(\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$$ defined in Section \[sec:Dmodules\] has the following properties: 1. $\ker{\phi} = I_{\lambda -\rho_{\mathfrak{g}}}$ (cf Section \[sec:unipotentrepresentations\]). 2. $\phi$ is surjective Most of the necessary ingredients are contained in [@BorhoBrylinski1982]. There is a degree filtration on the Dixmier algebra $\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$. With respect to this filtation, there is a natural isomorphism of graded commutative algebras $${\operatorname{gr}}\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}) \cong \Gamma(T^*\mathcal{Q},\mathcal{O}_{T^*\mathcal{Q}})$$ The algebra on the right is an integral domain (since the scheme $T^*\mathcal{Q}$ is reduced and irreducible). It follows for general reasons that $\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda})$ is zero-divisor free. Since $U(\mathfrak{g})/\ker{\phi}$ is included in this algebra, $\ker{\phi}$ is a completely prime ideal. Using Theorem \[thm:primitiveideals\], we deduce that $$\ker{\phi} \in \mathrm{Prim}^{\lambda-\rho_{\mathfrak{g}}}U(\mathfrak{g})$$ We also have $${\operatorname{AV}}(\ker{\phi}) = \eta(T^*\mathcal{Q}) = \overline{\mathcal{O}}_{\mathfrak{g}}$$ by, e.g. Proposition 4.3 in [@BorhoBrylinski1982]. Hence, $\ker{\phi} = I_{\lambda -\rho_{\mathfrak{g}}}$ by Proposition \[prop:propsofunipotentideals\]. Let $J$ be annihilator in $U(\mathfrak{g})$ of the cokernel of $\phi$. Under the birationality assumption, there is a strict inclusion $${\operatorname{AV}}(J) \subset \eta(T^*\mathcal{Q}) = \overline{\mathcal{O}}_{\mathfrak{g}}$$ by, e.g. Corollary 5.12 in [@BorhoBrylinski1982]. Since $J$ has infinitesimal character $\gamma_{\lambda-\rho_{\mathfrak{g}}}$, this implies that $J = U(\mathfrak{g})$ and hence that $\phi$ is surjective. Recall the factorization $\Gamma = \phi^* \circ \Gamma'$. Proposition \[prop:phiforunipotentlambda\] tells us that $\phi^*$ is easy to understand when $\gamma_{\lambda-\rho_{\mathfrak{g}}}$ is unipotent. Next, we turn our attention to $\Gamma'$. If $\lambda-\rho_{\mathfrak{g}}$ is strictly antidominant, then $\Gamma'$ is an equivalence by Theorem \[thm:bbequivalence1\]. Unfortunately, this condition is rarely satisfied when $\gamma_{\lambda-\rho_{\mathfrak{g}}}$ is unipotent. To apply Theorems \[thm:bbequivalence1\] and \[thm:bbequivalence2\], we will first need to translate to a more regular infinitesimal character. The following proposition is critical \[prop:translationweaklyfair1\] Suppose $\lambda-\rho(\mathfrak{u})$ is antidominant for $\mathfrak{u}$, i.e. $$\langle \lambda - \rho(\mathfrak{u}),\alpha^{\vee}\rangle \leq 0 \qquad \alpha \in \Delta^+(\mathfrak{u},\mathfrak{h})$$ Let $\xi \in \mathfrak{h}^*$ be the weight of a one-dimensional representation $\mathbb{C}_{\xi}$ of $\mathbf{Q}$, and let $F_{\xi}$ be the irreducible representation of $\mathbf{G}$ of highest weight $\xi$. Then there is a natural isomorphism of Dixmier algebras $$\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}) \cong T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}}\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda-\xi})$$ In the setting of Proposition \[prop:translationweaklyfair1\], $T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}}$ defines a surjective group homomorphism $$T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}}: KM(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda-\xi}),\mathbf{K}) \twoheadrightarrow KM(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$$ Write $\mathcal{L}^{\xi} \to \mathcal{Q}$ for the $\mathbf{G}$-equivariant line bundle corresponding to $\mathbb{C}_{\xi}$. If $(Z,\tau)$ is a standard datum for $(\mathfrak{q},\lambda-\xi)$, then $(Z,\tau \otimes \mathcal{L}^{\xi}|_Z)$ is a standard datum for $(\mathfrak{q},\lambda)$. This defines a bijection between standard data for $(\mathfrak{q},\lambda-\xi)$ and standard data for $(\mathfrak{q},\lambda)$ (with inverse $(Z,\tau) \mapsto (Z, \tau \otimes \mathcal{L}^{-\xi}|_Z)$). By the proof of Proposition \[prop:phiforunipotentlambda\] in [@Vogan1988b] \[prop:translationweaklyfair2\] In the setting of Proposition \[prop:translationweaklyfair1\], let $(Z,\tau)$ be a standard datum for $(\mathfrak{q},\lambda)$. Then there is an equality in $KM(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda},\mathbf{K})$ $$\Gamma' I(Z,\tau) = T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}} \Gamma' I(Z,\tau \otimes \mathcal{L}^{-\xi}|_Z)$$ We deduce \[thm:inductionmainresult1\] Let $\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}} \{0\}$, and suppose 1. $\gamma_{\lambda - \rho_{\mathfrak{g}}} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{g}})$ 2. $\lambda - \rho(\mathfrak{u})$ is antidominant for $\mathfrak{u}$ 3. The moment map $$\eta: T^*\mathcal{Q} \to \overline{\mathcal{O}}_{\mathfrak{g}}$$ is birational Then the sets ${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}})$ and $\mathrm{Deg}^{\lambda + \rho_{\mathfrak{l}}}(\mathfrak{q})$ are related by an upper triangular matrix. More precisely 1. ${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}})$ and $\mathrm{Deg}^{\lambda + \rho_{\mathfrak{l}}}(\mathfrak{q})$ have the same $\mathbb{Z}$-span in $KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$ 2. There is a natural injection $${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}}) \hookrightarrow \mathrm{Deg}^{\lambda + \rho_{\mathfrak{l}}}(\mathfrak{q})$$ and an ordering of the unipotent modules $U_1,...,U_n$ such that $$[U_p] = \pm [D_p] + \sum_{q >p} c_{pq} [D_q] \qquad c_{pq} \in \mathbb{Z}$$ for an ordering $D_1,...,D_m$ of the degenerate modules compatible with this injection. Suppose $[D] \in {\operatorname{Deg}}^{\lambda+\rho_{\mathfrak{l}}}(\mathfrak{q})$. By Proposition \[prop:phiforunipotentlambda\] and the definition of ${\operatorname{Deg}}^{\lambda+\rho_{\mathfrak{l}}}(\mathfrak{q})$, $D$ is contained in the subgroup $KM^{\mathrm{fl}}(U(\mathfrak{g})/I_{\lambda-\rho_{\mathfrak{g}}},\mathbf{K}) \subset KM^{\mathrm{fl}}(\mathfrak{g},\mathbf{K})$. Hence, $[D]$ is a linear combination of ${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}})$. Conversely, suppose $[U] \in {\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}})$. For every $\alpha \in \Delta(\mathfrak{u},\mathfrak{h})$, there is a strict inequality $$\langle \rho(\mathfrak{u}),\alpha^{\vee}\rangle = \langle \rho_{\mathfrak{g}}, \alpha^{\vee}\rangle - \langle \rho_{\mathfrak{l;}}, \alpha^{\vee}\rangle = \langle \rho_{\mathfrak{g}},\alpha^{\vee}\rangle > 0$$ Choose $N >>0$ so that $$\label{eqn:bigN}\langle \lambda - 2N\rho(\mathfrak{u}) - \rho_{\mathfrak{g}}\rangle < 0 \qquad \forall \alpha \in \Delta(\mathfrak{u},\mathfrak{h})$$ Then if $\xi = 2N\rho(\mathfrak{u})$, $ \lambda - \xi$ satisfies the antidominance condition of Proposition \[thm:bbequivalence1\]. Note that $\xi$ is the weight of the character $\left(\wedge^{\mathrm{top}}(\mathfrak{u})\right)^N$ of $\mathbf{Q}$. By Proposition \[prop:phiforunipotentlambda\], there is a unique irreducible $(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$-module $U'$ such that $U = \phi^*U'$. By the results of Section \[sec:translationdixmier\], there is a canonically-defined irreducible $(\Gamma(\mathcal{Q},\mathcal{D}_{\mathcal{Q}}^{\lambda}),\mathbf{K})$-module $\tilde{U}'$ such that $U' = T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}}\tilde{U}'$. And by Theorem \[thm:bbequivalence1\], there is a unique irreducible $(\mathcal{D}_{\lambda-\xi},\mathbf{K})$-module $\tilde{\mathcal{U}}'$ such that $\tilde{U}' = \Gamma'\tilde{\mathcal{U}}'$. The assignment $[U] \mapsto \tilde{\mathcal{U}}'$ defines an injection $${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}}) \hookrightarrow \mathrm{Irr}(\mathcal{D}_{\mathcal{Q}^{\lambda-\xi}},\mathbf{K})$$ By Theorem \[thm:bbequivalence2\], there is an ordering $\mathcal{M}_1,...\mathcal{M}_n$ of $\mathrm{Irr}(\mathcal{D}_{\mathcal{Q}^{\lambda-\xi}},\mathbf{K})$ such that $$[\mathcal{M}_p] = \pm I(Z_p,\tau_p) + \sum_{q >p} c_{pq} I(Z_q,\tau_q) \qquad c_{pq} \in \mathbb{Z}$$ where $(Z_1,\tau_1),...,(Z_n,\tau_n)$ is the induced ordering on the standard data for $(\mathfrak{q},\lambda-\xi)$. Define an ordering $U_1,...,U_r$ on ${\operatorname{Unip}}^{\lambda-\rho_{\mathfrak{g}}}(\mathcal{O}_{\mathfrak{g}})$ using the injection described above. Then there is an ordering on the standard data $(Z_i, \tau_i)$ so that $$\label{eqn:uppertriangularity2}[\tilde{\mathcal{U}}'_p] = \pm I(Z_p,\tau_p) + \sum_{q >p} c_{pq} I(Z_q,\tau_q) \qquad c_{pq} \in \mathbb{Z}$$ The composite $\phi^* \circ T_{F_{\xi},\lambda-\rho_{\mathfrak{g}}} \circ \Gamma': \mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda-\xi},\mathbf{K}) \to M(\mathfrak{g},\mathbf{K})$ is exact, and therefore defines a group homomorphism $K\mathcal{M}(\mathcal{D}_{\mathcal{Q}}^{\lambda-\xi},\mathbf{K}) \to KM(\mathfrak{g},\mathbf{K})$. If we apply this homomorphism to both sides of \[eqn:uppertriangularity2\] and use Proposition \[prop:translationweaklyfair2\] to simplify, we obtain equalities in $KM(\mathfrak{g},\mathbf{K})$ $$[U_p] = \pm \Gamma I(Z_p,\tau_p \otimes \mathcal{L}^{\xi}|_Z) + \sum_{q >p} d_{pq} \Gamma I(Z_q,\tau_q \otimes \mathcal{L}^{\xi}|_Z) \qquad d_{pq} \in \mathbb{Z}$$ The classes appearing on the right hand side are the degenerate modules by Theorem \[thm:hmswduality\]. A similar argument shows Let $\mathcal{O}_{\mathfrak{g}} = \mathrm{Ind}^{\mathfrak{g}}_{\mathfrak{l}}\mathcal{O}_{\mathfrak{l}}$ and let $\mu \in \mathfrak{h}^*$. Suppose 1. $\gamma_{\mu} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{l}})$ 2. $\gamma_{\mu - \rho(\mathfrak{u})} \in {\operatorname{unip}}(\mathcal{O}_{\mathfrak{g}})$ 3. $\mu - \rho(\mathfrak{u})$ is antidominant for $\mathfrak{u}$ 4. The moment map $$\eta: \mathbf{G} \times_{\mathbf{Q}} \left(\overline{\mathcal{O}}_{\mathfrak{l}} + \mathfrak{u}\right) \to \overline{\mathcal{O}}_{\mathfrak{g}}$$ is birational Then the sets ${\operatorname{Unip}}^{\mu-\rho(\mathfrak{u})}(\mathcal{O}_{\mathfrak{g}})$ and $\mathrm{Deg}^{\mu}(\mathfrak{q})$ are related by an upper triangular matrix (in the sense of Theorem \[thm:inductionmainresult1\]). Examples {#sec:examples} ======== Let $G = Sp(2n,\mathbb{R})$. The $\mathbf{G}$-orbits on $\mathcal{N}_{\mathfrak{g}}$ are parameterized by *even* partitions of $2n$. Let $\mathcal{O}$ be the orbit corresponding to the partition $2+2+...+2$. With respect to the usual symplectic form on $\mathbb{C}^{2n}$ $$\mathfrak{g} = \left\{\left( \begin{array}{c|c} A & B\\ \hline C & -A^t \end{array}\right): B,C \text{ symmetric}\right\}$$ and $\mathcal{O}$ is the $\mathbf{G}$-orbit of the matrix $$\left( \begin{array}{c|c} 0 & I_n\\ \hline 0 & 0 \end{array}\right)$$ It is not hard to see that $\mathcal{O}$ is birationally induced from the $\{0\}$-orbit of the Segal parabolic $$\mathfrak{l} = \left\{\left( \begin{array}{c|c} A & 0\\ \hline 0 & -A^t \end{array}\right)\right\} \qquad \mathfrak{u} = \left\{\left( \begin{array}{c|c} 0 & B\\ \hline 0 & 0 \end{array}\right): B \text{ symmetric}\right\} \qquad \mathfrak{q} = \mathfrak{l} \oplus \mathfrak{u}$$ Choose $\theta$ so that $\mathfrak{k} \subset \mathfrak{g}$ is the diagonal copy of $\mathfrak{gl}_2(\mathbb{C})$ (i.e. so that $\mathfrak{k} = \mathfrak{l}$). Then $\mathcal{O} \cap \mathfrak{p}$ decomposes into $n+1$ $\mathbf{K}$-orbits, which are parameterized by pairs $(a,b)$ of nonnegative integers with $a+b =n$. Write $\mathcal{O}_{(a,b)}$ for the nilpotent $\mathbf{K}$-orbit corresponding to the pair $(a,b)$. Then $\mathcal{O}_{(a,b)}$ is the $\mathbf{K}$-orbit of the matrix $$e_{(a,b)}:=\left( \begin{array}{c|c} 0 & \begin{array}{cc}I_a & 0\\0 & 0 \end{array}\\ \hline \begin{array}{cc}0 & 0\\0 & I_b\end{array} & 0 \end{array}\right)$$ Note that the centralizer in $\mathbf{K}$ of $e_{a,b}$ is the subgroup $$\mathbf{K}^{e_{(a,b)}} = \left\{\left( \begin{array}{c|c}O_a(\mathbb{C}) & \ast \\ \hline 0 & O_b(\mathbb{C}) \end{array}\right)\right\} \hookrightarrow Sp(2n,\mathbb{C})$$ Hence, if $a,b \geq 1$, there are four $\mathbf{K}$-equivariant line bundles on $\mathcal{O}_{a,b}$, parameterized by pairs $(x,y) \in \{0,1\}^2$. The line bundle $\mathcal{L}_{(x,y)}$ corresponding to $(x,y)$ has fiber over $e_{(a,b)}$ equal to the character $\det^x \otimes \det^y$ of $\mathbf{K}^{e_{(a,b)}}$. If either $a$ or $b$ is equal to $0$, then there are only two line bundles $\mathcal{L}_0$ and $\mathcal{L}_1$ corresponding to the trivial and determinant characters of $O_n(\mathbb{C})$. Choose the standard (diagonal) Cartan subalgebra $\mathfrak{h} \subset \mathfrak{g}$ and the usual positive system $\Delta^+(\mathfrak{g},\mathfrak{h}) = \{e_i-e_j\}_{i <j} \cup \{e_i+e_j\}_{i<j} \cup \{2e_i\}$. Compute $$\rho_{\mathfrak{l}} = \frac{1}{2}(n-1,n-3,...,3-n,1-n) \qquad \rho(\mathfrak{u}) = \frac{1}{2}(n+1,...,n+1) \qquad \rho_{\mathfrak{g}} = (n,n-1,...,1)$$ Note that $\lambda = \rho(\mathfrak{u})$ is a one-dimensional representation of $\mathfrak{l}$ and satisfies the conditions of Theorem \[thm:inductionmainresult1\]: $\gamma_{\lambda-\rho_{\mathfrak{g}}}= \gamma_{\rho_{\mathfrak{l}}}$ is a unipotent infinitesimal character associated to $\mathcal{O}$ (in fact, it is one of the special unipotent infinitesimal characters defined by Barbasch and Vogan in [@BarbaschVogan1985]), and $\lambda - \rho(\mathfrak{u}) = 0$ is antidominant for $\mathfrak{u}$. Hence, Theorem \[thm:inductionmainresult1\] implies that the sets ${\operatorname{Unip}}^{\rho_{\mathfrak{l}}}(\mathcal{O})$ and ${\operatorname{Deg}}^{\rho_{\mathfrak{g}}}(\mathfrak{q})$ are related by an upper triangular matrix. Using the Atlas software, we can compute its entries explicitly. We do so below for $n=2,3$, and $5$. For each group, we have included two tables. The first table lists the elements of ${\operatorname{Unip}}^{\rho(\mathfrak{l})}(\mathcal{O})$. For each representation, we record the associated variety (i.e. the set of open $\mathbf{K}$-orbits $\mathcal{O}_{(a,b)}$ therein) and the associated vector bundles. The second table lists the degenerate representations. For each representation, we record the support $\overline{Z}$ of the corresponding standard $(\mathcal{D}^{\lambda}_{\mathcal{Q}},\mathbf{K})$-module (by indicating the open KGB element in the preimage of $Z$ in the full flag variety) and its decomposition into unipotent representations. The KGB element is indicated in the Atlas notation. Note that some degenerates are $0$. Others are duplicated (when they arise as the global sections of several different standards). $Sp(4,\mathbb{R})$ ------------------ #### Unipotents Associated Variety Associated Vector Bundles ----- ---------------------------------------------------------- ------------------------------------------------------------- $1$ $\mathcal{O}_{2,0},\mathcal{O}_{1,1}, \mathcal{O}_{0,2}$ $\mathcal{L}_{(0)}, \mathcal{L}_{(0,0)}, \mathcal{L}_{(0)}$ $2$ $\mathcal{O}_{2,0},\mathcal{O}_{1,1}, \mathcal{O}_{0,2}$ $\mathcal{L}_{(1)}, \mathcal{L}_{(1,1)}, \mathcal{L}_{(1)}$ #### Degenerates Support (KGB) Decomposition into Unipotents --------------- ------------------------------- $10$ $1$ $10$ $2$ $Sp(6,\mathbb{R})$ ------------------ #### Unipotents Associated Variety Associated Vector Bundles ----- -------------------------------------------- ------------------------------------------- $1$ $\mathcal{O}_{(3,0)}$ $\mathcal{L}_{(0)}$ $2$ $\mathcal{O}_{(2,1)}$ $\mathcal{L}_{(1,0)}$ $3$ $\mathcal{O}_{(1,2)}$ $\mathcal{L}_{(0,1)}$ $4$ $\mathcal{O}_{(0,3)}$ $\mathcal{L}_{(0)}$ $5$ $\mathcal{O}_{(3,0)}, \mathcal{O}_{(2,1)}$ $\mathcal{L}_{(1)}, \mathcal{L}_{(1,1)}$ $6$ $\mathcal{O}_{(2,1)},\mathcal{O}_{(1,2)}$ $\mathcal{L}_{(0,0)},\mathcal{L}_{(0,0)}$ $7$ $\mathcal{O}_{(1,2)},\mathcal{O}_{(0,3)}$ $\mathcal{L}_{(1,1)},\mathcal{L}_{(1)}$ #### Degenerates Support (KGB) Decomposition into Unipotents --------------- ------------------------------- -- $7$ $1$ $18$ $2$ $31$ $2+1$ $17$ $3$ $5$ $4$ $30$ $4+3$ $31$ $5$ $44$ $6+4+1$ $34$ $6-2-3$ $40$ $6+4-3$ $40$ $6+4-3$ $41$ $6+1-2$ $41$ $6+1-2$ $30$ $7$ $44$ $7+5$ $34$ $0$ $Sp(10,\mathbb{R})$ ------------------- Associated Varieties Associated Vector Bundles ---- -------------------------------------------- ------------------------------------------- 1 $\mathcal{O}_{(5,0)}$ $\mathcal{L}_{(1)}$ 2 $\mathcal{O}_{(4,1)}$ $\mathcal{L}_{(0,1)}$ 3 $\mathcal{O}_{(3,2)}$ $\mathcal{L}_{(1,0)}$ 4 $\mathcal{O}_{(2,3)}$ $\mathcal{L}_{(0,1)}$ 5 $\mathcal{O}_{(1,4)}$ $\mathcal{L}_{(1,0)}$ 6 $\mathcal{O}_{(0,5)}$ $\mathcal{L}_{(1)}$ 7 $\mathcal{O}_{(5,0)},\mathcal{O}_{(4,1)}$ $\mathcal{L}_{(0,0)},\mathcal{L}_{(0,0)}$ 8 $\mathcal{O}_{(4,1)},\mathcal{O}_{(3,2)}$ $\mathcal{L}_{(1,1)},\mathcal{L}_{(1,1)}$ 9 $\mathcal{O}_{(3,2)},\mathcal{O}_{(2,3)}$ $\mathcal{L}_{(0,0)},\mathcal{L}_{(0,0)}$ 10 $\mathcal{O}_{(2,3)},\mathcal{O}_{(1,4)}$ $\mathcal{L}_{(1,1)},\mathcal{L}_{(1,1)}$ 11 $\mathcal{O}_{(1,4)}, \mathcal{O}_{(0,5)}$ $\mathcal{L}_{(0,0)},\mathcal{L}_{(0,0)}$ #### Degenerates Support (KGB) Decomposition Into Unipotents --------------- ------------------------------- $28$ $1$ $304$ $2$ $414$ $-2-1$ $469$ $3$ $469$ $3$ $779$ $-3-1-1$ $470$ $4$ $470$ $4$ $877$ $-4-2-2$ $740$ $-4-3$ $306$ $5$ $878$ $-5-5-3$ $956$ $5+3+1$ $31$ $6$ $781$ $-6-6-4$ $416$ $-6-5$ $957$ $6+4+2$ $414$ $7$ $690$ $8-3-2$ $779$ $-8+2-1-1$ $877$ $-8-4+3-2$ $921$ $8+4+2+1+1$ $740$ $-9$ $921$ $9+7+7$ $691$ $10-5-4$ $781$ $-10-6-6+5$ $878$ $-10-5+4-3$ $922$ $10+6+6+5+3$ $941$ $-10-8+5+4+3+2$ $956$ $10+8-4-2+1$ $957$ $10+8-3-5+6$ $962$ $-10-8-6-1$ $416$ $11$ $922$ $11+11+9$ $962$ $-11-9-7$ $690$ $0$ $691$ $0$ $941$ $0$

--- author: - 'S. Hernàndez-Navarro' - 'P. Tierno' - 'J. Ignés-Mullol' - 'F. Sagués' date: 'Received: date / Revised version: date' title: 'Liquid-crystal enabled electrophoresis: Scenarios for driving and reconfigurable assembling of colloids' --- Liquid crystals as dispersing media for colloids ================================================ One of the most distinctive properties of soft matter systems is that they are extremely labile, and as such they may be readily controllable by means of appropriately chosen external fields. However, this capability is useless if one does not guarantee a sufficient control over the elicited responses, particularly when these responses are prone to be highly degenerated. This is particularly true when referring to two of the most studied categories of soft materials: liquid crystals and colloids. In particular, interest has increased enormously in recent years to bridge the potentialities of both systems through the use of liquid crystals as dispersing media for solid or liquid colloidal inclusions. Such composite systems, where we benefit of the discrete nature of the dispersed components and, at the same time, of the orientational properties of the liquid crystal medium, constitute nowadays familiar mixtures under intense scrutiny to unveil new fundamental concepts and original applications [@1; @2; @3]. In particular, the capacity of the elastic and anisotropic host matrix to mediate interactions between suspended inclusions was already discovered many years ago [@4], and has been profusely used since then [@5; @6; @7; @8; @9; @10; @11; @12; @13; @14]. In contrast to research on liquid crystal dispersions of sub-micron or nano-scale particles [@15; @16], we concentrate in what follows on systems that contain colloidal inclusions at the micrometer scale, which enable a real discrete control and observation of the dispersed components. It is also worth emphasizing at this point that, in spite of very recently investigated dispersions prepared with lyotropic (i.e. water-based) liquid crystals [@17; @18], all what we are going to review in this contribution refers to thermotropic oily materials in their nematic phases [@19]. Nematic liquid crystals (NLCs) are complex fluids mostly used in display technologies. They are characterized by rod shaped organic molecules which tend to align their long axis on average along a common direction, called the director field. The behavior of colloids dispersed in NLCs is largely determined by considering the state of the elastic LC matrix around the dispersed inclusion. In this respect, the most important characteristics is the anchoring conditions of the liquid crystal at the surface of the colloid. The simplest case is that of a spherical solid particle, around which the nematic material may adopt either a tangential (parallel to the surface) or a homeotropic (perpendicular to the surface) arrangement, depending on the particular system and surface preparation. In either case, placing a colloidal unit into an otherwise aligned nematic phase leads to topological frustration and to the appearance of defects (i.e. singularities in the distribution of the director field), that are tightly bound to the inclusion. For homeotropic anchoring, two main types of defects may arise: a point-defect known as a hyperbolic hedgehog or a disclination loop extending as an equatorial band and known as Saturn-ring [@20; @21; @22]. For tangential anchoring, the defect structure is that of a double-boojum [@20]. Schematic representations of these defect configurations are presented in Fig. 1. Of singular importance on what follows is the elucidation of the symmetries of these defects. In this respect, Poulin et al., already pointed out in their seminal paper  [@4] the convenience to map the colloidal-induced elastic distortion of the far field distribution of the nematic director onto an electrostatic context. According to this analogy, a (ideal) spherical colloid with its corresponding hedgehog defect acts as an electrostatic dipole, whereas the structures of Saturn-rings and doubleboojums are better understood as giving rise to quadrupolar distortions of the director field. From non-linear electrophoresis to liquid-crystal enabled electrophoresis ========================================================================= Another more recently investigated feature of nematic colloids, this term used as a short denomination for colloidal dispersions in a continuous nematic phase, is the possibility to observe non-trivial manifestations of phoretic driving of the dispersed inclusions under the application of electric fields. Indeed, what makes this possibility particularly striking is that, contrarily to normal phoresis under steady fields, the motion is here induced by oscillating electric fields. This confers to the phenomenon a distinctive non-linear character that justifies the use of the abbreviated denomination non-linear electrophoresis [@23]. An alternative way to qualify this electrically-driven transport, more appropriate to its physical origin, is known as induced-charge electroosmosis  [@24]. Focusing on liquid crystal based systems, we will employ on what follows the composed form liquid-crystal enabled electrophoresis LCEEP, as was recently coined by Lavrentovich et al. [@25]. For normal electrically induced phoresis, particles are transported under uniform DC fields with a velocity that depends linearly on the applied field. Obviously this means that oscillating fields which average to zero over a period would cause a zero displacement. However the possibility to use alternate rather than steady driving modes has the advantage that concurrent electrochemical processes, that take place in the normally aqueous medium employed in electrophoretic cells, are avoided. In fact, nonlinear versions of electrophoresis were first reported for isotropic fluids more than twenty years ago [@26], as mentioned in the recent review [@25], and were exhaustively analyzed in some theoretical papers by Bazant and Squires [@24]. In this latter study, the main conclusion is that a nonlinear electro-osmotic slip occurs when an applied field acts on the ionic charge it induces around a polarizable surface. Notice from this very first statement that nonlinear electrokinetic phenomena apply to charged and non-charged inclusions as well, in striking contrast with normal electrophoresis. As a matter of fact, the possibility to further break the symmetries of the resulting flow and obtain direct motion for Janus-like particles (metallo-dielectric) was first published by Velev et al. [@27]. In the context of liquid crystals, the breaking of the fore-aft symmetry that guarantees direct transport may be achieved even for perfectly spherical particles, bearing in mind our above considerations relative to the symmetries of the nematics director field around the inclusions. This is quite evident for defects with dipolar symmetry, but also applies to distorted quadrupolar symmetries. The latter situation would occur, for instance, in the case of a parallel anchoring of the liquid crystal material wrapping around non-perfectly symmetric, i.e., anisometric, particles. Both possibilities have been largely exploited in our recent experiments that will be reviewed in the second part of this contribution. Notice finally that a quadratic dependence of the phoretic velocity with respect to the applied field has another striking implication: the dependence of the velocity on the applied electric field, both vectorial quantities, must be indeed of a tensorial nature. This, in turn, permits that the direction of motion and that of the electric field are non-necessarily parallel [@23]. A general overview on the different possibilities involved in the transport of particles in liquid crystals is provided in  [@25]. Phoresis of nematic colloids: An active matter-based perspective and applications in materials science ====================================================================================================== Active matter is a presently very celebrated keyword under which we recognize many forms of elementary living matter, altogether with remarkable non-living soft matter analogs [@28]. A defining feature commonly exhibited by the units composing these discrete systems is their ability for autonomous or driven motion. More precisely, we should properly reserve the term active for the first kind of (self-driven) behavior and employ the term actuated to refer to the second situation, which is in fact what applies to our electrophoretic colloids dispersed in a liquid crystal material. Under whatever acception, much attention has been devoted to elucidate the mechanisms behind complex multiparticle phenomena that self-organize into characteristic patterns with long range order, either of static or dynamic nature. In this respect, terms like bands, asters, vortices are very much familiar to practitioners in this field. Diverse aspects of this topic are examined in this Proceedings volume while a comprehensive update is provided in the review by Marchetti et al. [@28]. Although experimental realizations are becoming more and more impressive particularly when restricting to a bio-inspired context, see  [@29; @30; @31] for striking observations of active self-organization in a biophysical context, we feel that theoretical and numerical work is leading the activity in the field. We thus claim for the need to provide the interested community with experimentally robust realizations of new scenarios of self-organized patterns in soft-matter systems, either active or driven. More important, as it is going to be much emphasized on the last part of this contribution, is the further need to gain control on the assembly process and to demonstrate the eventual possibility to reconfigure these aggregates at will. We would like to mention another completely different and more materials based perspective from where our experimental research on phoretic colloids could be addressed. The assembly and transport of colloidal entities, such as particles, droplets or microorganisms have direct applications in fields such as photonics  [@32], lab-on-a-chip technologies  [@33] and biomedicine  [@34]. During this last decade the focus on self-assembling started to shift from equilibrium characteristics to dynamic aspects [@35], seeking enhanced functionalities of the resulting materials [@36]. Optical trapping techniques are often employed to achieve direct control over the placement of colloidal inclusions, and holographic tweezers allow to extend such control to a few hundreds of particles [@37], although this technique is limited by the field of view of the optical system. Going beyond earlier work on physically and chemically activated colloids [@38; @39; @40], a recent realization explores the idea of phoretic colloids driven by osmotic pressure imbalances [@41]. These results have opened new possibilities for reconfigurable self-assembly, enabling the massive transport of inclusions [@42], although the precise and selective control on such dynamical patterns lacked experimental evidence. AC electrophoresis of microdroplets in liquid crystals: Transport and reaction ============================================================================== The second part of this contribution is prepared to briefly summarize our recent research on the phoresis of inclusions dispersed in nematic liquid crystals. It is itself divided into two sections. The first one is devoted to the transport of (water) microdroplets, while we reserve the second part to present a new development that permits the reconfigurable assembly of colloidal particles into clusters (swarms). Water-based microemulsions prepared in oil phases are of fundamental importance in chemical and analytical sciences, due to the possibility of encapsulating and delivering chemical compounds otherwise immiscible in the dispersion medium. The possibility to remotely transport microdroplets is thus singularly appealing. Electrically induced phoresis is undoubtedly the first considered alternative, and particularly its non-linear version based on the use of oscillating electric fields offer great advantages as mentioned previously in Sect. 2 Our experimental system, thoroughly described in Hernández-Navarro et al.  [@44], permits to transport water microdroplets dispersed in a nematic liquid crystal with negative dielectric anisotropy, i.e. it allows to observe phoretic motion perpendicular to the applied electric field. Additionally, we decouple in this way AC electrophoresis from any residual linear DC contribution. As reported in the reference mentioned, we were able to demonstrate not only droplet motion but to show in addition that these droplets can be used as microreactors to transport sub-micrometric particles or to mix tiny volumes of chemicals. The used cell was composed of two 0.7$mm$ thick microscope slides of size $15 \times 25 \rm{mm^2}$ coated with a layer of indium-tin oxide (ITO). The two slides were cleaned, dried and further chemically treated to obtain a planar alignment of the liquid crystal on the bounding plates. These were separated by a spacer of nominal thickness $23$ microns and glued together with the ITO layers facing inwards. Dispersions of aqueous microdroplets (from $1$ to $20$ micron diameter sizes) were prepared by vortex agitation using MLC-7029 as a nematic phase, using as stabilizing agent sodium dodecyl sulphate (SDS). This protocol guarantees a homeotropic alignment of the nematics on the droplets surface. The experimental cells were filled by capillarity. Sinusoidal electric fields were applied by using a function generator, within a range of $0$ to $30$ volts peak-to-peak, while the applied range of frequencies varied from $0$ to $100 $Hz. Experimental observations were performed with an optical microscope, and images were captured with a recording camera controlled with the appropriate software AVT SmartView 1.10.2, and further treated using software packages ImageJ and IgorPro. For more details see the original paper  [@44]. Figure 2(a) shows the experimental cell. In the scenarios we are reviewing here we found that the defects were predominantly of the hedgehog type, while only a small fraction of inclusions (less than 1%) featured Saturn-ring defects. For isolated droplets it is known [@45] that the hedgehog is the usual defect configuration for large inclusions, whereas Saturn-ring distortions are preferred in smaller ones. However the situation is more complicated for confined systems, and in this case Saturn-ring defects are, in general, limited to large droplets, as observed in our experiments. Motion was always observed in the direction towards the point defect as clearly observed in Fig. 2(b). Motion of the droplet in relation to the location of the point defect may, however, depend on the used liquid crystal, as reported in [@44]. The measured velocities are typically on the range of a few microns per second, featuring the expected quadratic dependence on the amplitude of the electric field and a peak at intermediate frequencies (tens of Hertz). To further demonstrate the possibilities of this phoretic motion, a localized cargo release operation is shown in Fig. 3(a). A microdroplet loaded with polystyrene particles is driven towards a larger droplet bearing a Saturn-ring defect, and thus stays at rest under the applied electric field. The larger droplet elastically attracts the small one, and the colloidal cargo is released by droplet coalescence. In Fig. 3(b), two droplets loaded with specific reactants and featuring antiparallel configurations of their respective dipolar defect configuration attract each other, and their coalescence triggers a chemical precipitation reaction (see original paper [@44] for further details). Reconfigurable swarming ======================= Reconfigurability is a genuine property of living active matter. As laboratory-model systems, collections of driven colloids have a potential for large-scale addressability. Here we review our very recent research that indeed demonstrate real-time reconfigurable clustering of phoretic colloids dispersed in thin nematic layers in response to illumination patterns. Swarms of particles are shown to reversible organize in submillimeter ensembles that can be individually or collectively addressed to change their position or dynamic behavior. The strategy we follow permits to separate particle steering, achieved through photoelastic modulation of the host nematic, and particle driving realized through LCEEP. A complete account of these experiments can be found in Hernàndez-Navarro et al. [@46]. We used the same nematic liquid crystal as in the previously reported experiments, thus favoring motion of the dispersed colloids perpendicular to the applied electric field. However, the chemical treatments of the confining plates was different in this case and, in fact they were specifically aimed at achieving the purposed control on the assembly of the dispersed particles. One of the plates was chemically functionalized with an azosilane photosensitive self-assembled monolayer that allows to alternate between perpendicular (homeotropic) and tangential (planar) anchoring of the nematic [@47]. The counter plate was coated with a polyimide compound to guarantee strong and permanent perpendicular contact. Without any external influence, such boundary conditions lead to a uniform homeotropic texture within the whole cell. By irradiating with UV ($365$nm) light from an incoherent source the azosylane is forced to adopt the cis configuration, and consequent planar boundary conditions are attained. The azosylane is easily reverted to trans form using blue ($455$nm) illumination, and corresponding homeotropic anchoring is regained. Light-induced reorientations typically take a few seconds, a much shorter time scale than that set by particle motion. As colloidal inclusions we used pear-shaped microparticles made of polystyrene, a material that promotes planar orientation of the NLC on the particle surface. The chosen particle shape guaranteed a distorted quadrupolar symmetry arising from the defect distribution around the particles. In the absence of irradiation or electric field, particles align perpendicular to the cell plates, following the uniform director field. Illuminating the cell for a few seconds with a spot of UV light forces the NLC in contact with the azosylane-treated surface to transit to a planar configuration by locally adopting a local splayed (radially-spread) texture emanating from a central defect. Application of an external AC field makes the bulk NLC to adopt the splay configuration that now extends for several millimeters, well beyond the area of the irradiated spot, thanks to the homeotropic anchoring degeneracy relative to any planar direction. This configuration is stable for days under AC field, well past the half-life for thermal relaxation of the azosylane film, which is about $30$ minutes. The region with radial alignment will be the basin of attraction for dispersed particles, which tumble instantaneously following the NLC director so that their long axis lays, on average, parallel to the cell plates. Simultaneously the LCEEP sets the particles into motion at a constant speed. All particles moved following the local NLC director, with roughly half of them being attracted by the photoinduced radial defect and the rest being repelled from it. This is direct consequence of the random tumbling of the particle following the reorientation of the liquid crystal matrix. Most of the particles propel with the large lobe ahead, as reported in [@46]. See Fig.4 for a combined schematics of the different aspects of the optical and electrical forcing of the cell. During the experiment, particles accumulate as they follow the nematic field lines. A growing aster-like cluster was typically observed after the progressive arrival and subsequent jamming of the particles leading to high density assemblies, see Fig. 5 (panels a, b, c). We could easily switch to a dynamic structure, a rotating mill-like cluster, by profiting from the elastic properties of the nematic material (smaller value of the bend elastic constant with respect to the splay counterpart) and the fact that the particles are totally slave to the director. The experimental protocol proceeds by erasing the central region of an imprinted UV area with a smaller spot of blue light, prior to the application of the electric field. Following the inverse isomerization to the trans form of the azosylane anchoring derivative, the NLC director features a homeotropic configuration both outside and inside a corona with planar alignment. Upon application of the AC field, the NLC director adopts a degenerate planar alignment both inside and outside the ring. The energy cost of the large splay distortion inside the ring prompts the director to adopt a spiral bend-splay texture and director field lines conform to a spiral-like geometry. As a consequence particles assemble into a rotating mill structure, preceding around the central defect with a constant linear velocity, see Fig. 5 (panels d, e, f). Both assembly modes can be reversibly interconverted in real time via the photoativation control just described (see original reference [@46]). It is worth empasizing that, as expected, there is no preferent sign in the cluster rotation, since there is no chiral influence in the system. In this respect particle aggregation here is very different with respect to assembling of particles under a forced imposed vortical flow (see for instance [@48]). The reversibility and quick response of the photoalignment layer enables straightforward cluster addressability. A preformed aggregate of arbitrary size, either aster or vortex-like, can be relocated to a pre-designed place anywhere within the experimental cell with minimum dismantlement of the cluster structure by changing the location of the UV spot. An example of this process is shown in Fig. 6. After blocking the LCEEP mechanism by increasing the field frequency above 50 Hz, the center of attraction is translated 600 μm. Once LCEEP is reactivated, the swarm of particles moves towards the new position developing a leading edge around which the particles assemble. The space-time plot illustrates the resulting collective behavior. Alternatively, by the same principle one can imprint predesigned arbitrary paths connecting distant locations inside the cell or draw circuits with complex topologies as a simple way to accumulate colloidal swarms in the irradiated area and further entrain them collectively. We thank Patrick Oswald for the polyimide compound. We acknowledge financial support by MICINN (FIS2010-21924C02, FIS2011-15948-E) and DURSI (2009 SGR 1055). S.H.-N. acknowledges support through an FPU Fellowship (AP2009-0974). P.T. further acknowledges support from the ERC through the starting grant DynaMO (335040) and from the Ramon y Cajal program (RYC-2011-07605). A. Agarwal, et al., Small **9**, 2785 (2013) B. Senyuk, et al., Nature **493**, 200 (2013) T. A. Wood, et al., Science **334**, 79 (2011) P. Poulin, et al., Science **275**, 1770 (1997) J.C. Loudet, et al., Nature **407**, 6111 (2000) J. Yamamoto, et al., Nature **409**, 322 (2001) C. Lapointe, et al., Science **303**, 652 (2004) M. Yada, et al., Phys. Rev. Lett. **92**, 185501 (2004) I. Musevic, et al., Science **313**, 954 (2006) O.P. Pishnyak, et al., Phys. Rev. Lett. **99**, 127802 (2007) C.P. Lapointe, et al., Science **326**, 1083 (2009) G.M. Koening, et al., Proc. Natl. Acad. Sci. USA **107**, 3998 (2010) U. Tkalec, et al., Science **333**, 62 (2011) R. P. Trivedi, et al., Proc. Natl. Acad. Sci. USA **109**, 4744 (2012) H. Qi, et al., Adv. Funct. Mat. **18**, 212 (2008) S. Acharya, et al., Adv. Mat. **21**, 989 (2009) S. Zhou, et al., Proc. Natl. Acad. Sci. USA **111**, 1265 (2014) P.C. Mushenheim, et al., Soft Matter **10**, 88 (2014) P. Oswald, et al., Nematic and cholesteric liquid crystals: Concepts and physical properties illustrated by experiments (Taylor and Francis, Boca Raton, 2005) P. Poulin, et al., Phys. Rev. E **57**, 626 (1998) T.C. Lubensky, et al., Phys. Rev. E **57**, 610 (1998) Y. Gu, et al., Phys. Rev. Lett. **85**, 4719 (2000) O.D. Lavrentovich, et al., Nature **467**, 947 (2010) T.M. Squires, et al., J. Fluid. Mech. **509**, 217 (2004) O.D. Lavrentovich, et al., Soft Matter **10**, 1264 (2014) V.A. Murtsokvin, et al., Colloid J. **52**, 933 (1990) S. Gangwal, et al., Phys. Rev. Lett. **100**, 058302 (2008) M.C. Marchetti, et al., Rev. Mod. Phys. **85**, 1143 (2013) V. Schaller, et al., Nature **467**, 73 (2010) Y. Sumino, et al., Nature **483**, 448 (2012) T. Sanchez, et al., Nature **491**, 431 (2012) A. Yethiraj, et al., Adv. Mat. **16**, 596 (2004) A. Terray, et al., Science **296**, 1841 (2002) E.C. Dreaden, et al., Chem. Soc. Rev. **41**, 2740 (2012) G.M. Whitesides, et al., Science **295**, 2418 (2002) N.I. Zheludev, et al., Nat. Mat. **11**, 917 (2012) D.G. Grier, Nature **424**, 810 (2003) W.F. Paxton, et al., J. Am. Chem. Soc. **126**, 13424 (2004) J.R. Howse, et al., Phys. Rev. Lett. **99**, 048102 (2007) I. Buttinoni, et al., Phys. Rev. Lett. **110**, 238301 (2013) J. Palacci, et al., Science **339**, 936 (2013) A. Bricard, et al., Nature **503**, 95 (2013) J. Guzowski, et al., Soft Matter **8**, 7269 (2012) S. Hernàndez-Navarro, et al., Soft Matter **9**, 7999 (2013). H. Stark, Eur. Phys. J. B **10**, 311 (1999) S. Hernàndez-Navarro, et al., Angew. Chem. Int. Ed. **53**, 10696 (2014) J. Ignés-Mullol, et al., Langmuir **21**, 2948 (2005) T. Nagatani, et al., J. Phys. Soc. Jpn. **59**, 3447 (1990)

\[theorem\][Proposition]{} \[theorem\][Corollary]{} \[theorem\][Lemma]{} \[theorem\][Observation]{} \[theorem\][Conjecture]{} \[theorem\][Question]{} \[theorem\][Questions]{} \[theorem\][Problem]{} \[theorem\][Problems]{} \[theorem\][Open Problem]{} \[theorem\][Definition]{} \[theorem\][Example]{} \[theorem\][Exercise]{} \[theorem\][Remark]{} \[theorem\][Remarks]{} \[theorem\][Fact]{} \[theorem\][Claim]{}

--- abstract: 'The interaction mechanism of the $\Sigma$(1385) photoproduction from proton $\gamma p\to K^+\Sigma^0(1385)$ is investigated within a Regge-plus-resonance approach based on the experimental data released by the CLAS Collaboration recently. The $t$ channel and the $u$ channel are responsible to the behaviors of differential cross sections at forward and backward angles, respectively. The contributions from nucleon resonances including $N^*$ and $\Delta^*$, which are determined by the predicted decay amplitudes in the constituent quark model, are found small, but the $F_{35}$ state $\Delta(2000)$ is essential to reproduce differential cross section.' author: - Jun He title: 'The $\Sigma$(1385) photoproduciton from proton within a Regge-plus-resonance approach' --- Introduction ============ The study of nucleon resonances is an important topic of hadron physics. The information about nucleon resonance is mainly extracted from the pion-nucleon scattering especially in the early stage of the study of nucleon resonance [@Vrana:1999nt]. Based on a large number of nucleon resonances found in the experiment, the constituent quark model (CQM) was developed and achieved great success in the explanation about the property of nucleon resonance [@Isgur:1978xj; @Capstick:1986bm]. However, the predicted nucleon resonances in the CQM are much more than the ones found in the experiment, which is the so-called “missing resonance” problem. One explanation about this problem is that the decay ratio of “missing resonance” is very small in the usual experimental detected channels, such as the pion-nucleon channel. Hence, the channels more than the pion-nucleon channel, such as $\eta N$ and strange channels, attract much attentions. A mount of experimental data of the kaon photoproduction companied by a ground strange baryon $\Lambda$ or $\Sigma$ have been accumulated in the recent years [@Klempt:2009pi]. However, the study about the kaon photoproduction with a strange baryon resonance is scarce. Very recently, the CLAS Collaboration released their experimental data about the kaon photoproduction with $\Sigma$(1385), $\Lambda(1520)$ and $\Lambda(1405)$ with high precision [@Moriya:2013hwg], which provide an opportunity to study nucleon resonances in these channels. The strong decays of nucleon resonances to $\Sigma$(1385), $\Lambda(1520)$ or $\Lambda(1405)$ with kaon meson have been studied in the CQM [@Capstick:1998uh]. Combined with the theoretical prediction about the radiative decay [@Capstick:1992uc], one can make a rough estimation which nucleon resonances play important roles in a certain photoproduction process. For example, the large decay widths to $N\gamma$ and $K\Lambda(1520)$ suggest that $N(2120)$ should be easy to be found in the kaon photoproduction with $\Lambda(1520)$, which has been confirmed by many theoretical analysis of the $\Lambda(1520)$ photoproduction data [@Xie:2010yk; @He:2012ud]. The CQM prediction suggests a large decay ratio of the $D_{13}$ state $N(2095)$ and the $F_{35}$ state $\Delta$(2000) in $\Sigma(1385)K$ decay channel[@Capstick:1998uh]. Hence it is interesting to study the roles played by such states in the $\Sigma$(1385) photoproduction. There only exist some old experimental data about the $\Sigma$(1385) photoproduction with low precision, which were obtained before last seventies [@CambridgeBubbleChamberGroup:1967zzb; @Erbe:1970cq]. The LEPS Collaboration also released some results in this channel but only at extreme forward angles [@Niiyama:2008rt]. Correspondingly, the theoretical studies are also scarce. In Ref. [@Oh:2007jd] the $\Sigma$(1385) photoproduction has been studied in an effective Lagrangian approach based on the preliminary data from CLAS Collaboration. However, due to the absence of the constraint of the precise data large discrepancies at low energies between the experimental data and the theoretical predictions can be found in the differential cross section released by the CLAS Collaboration [@Moriya:2013hwg]. In this work, we will analyze the new CLAS data within a Regge-plus-resonance approach and investigate the roles played by nucleon resonances. This paper is organized as follows. After introduction, we will present the effective Lagrangian used in this work and Reggeized treatment for $t$ channel. The gauge invariance will be also discussed in this section. The numerical results for the cross section will be given and compared with the experimental data in Section \[Sec: Results\]. Finally, the paper ends with a brief summary. FORMALISM ========= The four types of interaction mechanism, the $s$ channel, the $u$ channel, the $t$ channel and the contact term for the $\Lambda(1520)$ photoproduction from nucleon with $K$ are presented in Fig. \[pic:dia\]. ![(Color online) The diagrams for the $s$, $u$ and $t$ channels and contact term for $\gamma p\to K^+\Sigma^0(1385)$.[]{data-label="pic:dia"}](diagram) The Born terms contain the $N$, $Y$, $K$ intermediate states and the contact term. For the Born $s$ channel, $t$ channel and contact term, the Lagrangians involved are given as below, $$\begin{aligned} \mathcal{L}_{\gamma KK} &=& ie A_\mu \left( K^- \partial^\mu K^+ - \partial^\mu K^- K^+ \right),\\ \mathcal{L}_{KN\Sigma^*} &=& \frac{f_{KN\Sigma^*}}{m_K} \partial_\mu \overline{K} \overline{\bm \Sigma}^{*\mu} \cdot \bm{\tau} N + \mbox{ H.c.}, \\ \mathcal{L}_{\gamma NN} &=& -e \bar{N} \left( e_N\gamma^\mu - \frac{\kappa_N^{}}{2M_N} \sigma^{\mu\nu} \partial_\nu\right) A_\mu N, \\ \mathcal{L}_{\gamma KN\Sigma^*} &=& -ie \frac{f_{KN\Sigma^*}}{m_K} A^\mu K^- \left( \bar{\Sigma}_\mu^{*0} p + \sqrt{2} \bar{\Sigma}_\mu^{*+} n \right) + \mbox{ H.c.}, \label{eq:lag1}\end{aligned}$$ where $A^\mu$, $N$, $K$, $\Sigma^{*\mu}$ are the photon, nucleon, kaon and $\Sigma^*(1385)$ fields and the charge of nucleon $e_N=1,0$ for proton and neutron in the unit of $e=\sqrt{4\pi \alpha}$ with $\alpha$ being the fine-structure constant . The anomalous magnetic moment $\kappa_N=1.79$ for proton. $m_K$ and $M_N$ are the masses of kaon and nucleon. The coupling constant for $KN\Sigma^*$ vertex can be related to the $\pi N \Delta$ coupling by the SU(3) flavor symmetry relation, and the value $f_{KN\Sigma^*} = -3.22$ can be obtained [@Oh:2007jd; @Gao:2010hy]. The $t$ channel for the $\Sigma$(1385) photoproduction occurs through both $K$ and $K^*$ exchanges. As shown in Ref. [@Oh:2007jd], the contribution from the $K^*$ exchange is negligible at energies $E_\gamma=3\sim4$ GeV with the reasonable coupling constant. Hence only $K$ exchange will be considered in this work. The $u$ channel diagram shown in Fig. \[pic:dia\] contains intermediate hyperons. The effective Lagrangians for these diagrams are $$\begin{aligned} \mathcal{L}_{\Sigma^*Y\gamma} &=& -\frac{ief_1}{2M_Y} \overline{Y} \gamma_\nu \gamma_5 F^{\mu\nu} \Sigma_\mu^* - \frac{ef_2}{(2M_Y)^2} \partial_\nu \overline{Y} \gamma_5 F^{\mu\nu} \Sigma_\mu^* + \mbox{ H.c.},\nonumber\\ \mathcal{L}_{KNY} &=& \frac{g_{KNY}^{}}{M_N + M_Y} \overline{N} \gamma^\mu \gamma_5^{} Y \partial_\mu K + \mbox{ H.c.},\end{aligned}$$ where $F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu$ and $Y$ stands for a hyperon with spin-$1/2$ carrying a mass $M_Y$ . For the intermediate $\Lambda(1116)$ state, $f_1 = 4.52$ and $f_2 = 5.63$ [@Oh:2007jd]. The coupling constant $g_{KN\Lambda}$ can be determined by flavor SU(3) symmetry relation, which gives value $g_{KN\Lambda}= -13.24$ [@Oh:2007jd]. The $\Sigma$ exchange is negligible due to to the small coupling constant determined from SU(3) symmetries [@Oh:2007jd]. Besides, the effect of higher resonances can be included through Reggeized treatment. Due to the large uncertainty at backward angles of the CLAS experimental data and much larger masses of $\Lambda$ and $\Sigma$ baryons compared with the mass of $K$ in the $t$ channel we do not consider the Reggeized treatment in the $u$ channel in this work as Ref. [@Corthals:2005ce]. The amplitude for the $u$ channel is gauge invariant itself while the amplitudes for the Born $s$ channel, $t$ channel and contact term are not gauge invariant. After summing up the amplitudes from the $s$ channel, the $t$ channel and the contact term of the $\Sigma$(1385) photoproduction, the gauge invariance can be guaranteed as the $\pi$ and $\Lambda(1520)$ photoproductions [@Ohta:1989ji; @Nam:2010au]. The effect of the hadron internal structure can be reflected by the form factor added at each vertex. Unfortunately, it will violate the gauge invariance. To restore the gauge invariance, a generalized contact term is introduced as [@Haberzettl:2006bn] $$\begin{aligned} {\cal M}^{\mu\nu}_c&=&\frac{ief_{KN\Sigma^*}}{m_K}\left[g^{\mu\nu}F_t+ k_2^\mu (2k_2-k_1)^\nu\frac{(F_t-1)[1-h(1-F_s)]}{t-m_K^2}\right.\nonumber\\ &+&\left.k_2^\mu (2p_1-k_1)^\nu\frac{(F_s-1)[1-h(1-F_t)]}{s-M_N^2}\right], \label{Eq: contact}\end{aligned}$$ where the $h$ is a free parameter and will be fitted and $F_i$ with $i=s,t$ is the form factor. In this work, for the Born $s$ channel and the $u$ channel we choose the form factor in the from, $$\begin{aligned} F_i(q^2)&=&\left(\frac{n\Lambda_i^4}{n\Lambda_i^4+(q^2-M^2)^2}\right)^n,\label{Eq: FF}\end{aligned}$$ which goes to Gaussian form as $n\to \infty$ and for $t$ channel $K$ exchange, $$\begin{aligned} F_i(q^2)&=&\frac{\Lambda_i^2-M^2}{\Lambda_i^2-q^2},\end{aligned}$$ where $M$ and $q$ are the mass and momentum of the off-shell intermediate particle. The cutoff $\Lambda_i$ for $s$, $u$ or $t$ channel should be about 1 GeV and will be set as free parameter in this work. We introduce $K$ Reggeized treatment as following to describe the behavior of differential cross section of the $\Sigma(1385)$ photoproduction at high photon energies  [@Guidal:1997hy; @Corthals:2005ce; @Titov:2005kf], $$\begin{aligned} \label{eq:RT} \frac{1}{t-m^{2}_{K}}\to\mathcal{D}_{K} &=&\left(\frac{s}{s_{scale}} \right)^{\alpha_{K}} \frac{\pi\alpha'_{K}}{\Gamma(1+\alpha_{K})\sin(\pi\alpha_{K})},\end{aligned}$$ where $\alpha'_{K}$ is the slope of the trajectory and the scale factor $s_{scale}$ is fixed at 1 GeV$^2$. $\alpha_{K}$ is the linear trajectory of the $K$ meson, which is a function of $t$ assigned as follows, $\alpha_{K}=0.70\,\mathrm{GeV}^{-2}(t-m^{2}_{K})$. The $K^*$ Reggeized treatment is analogous. There is no reason a priori that the coupling constants for Reggeized treatment $f^{Reg}_{KN\Sigma^*}$ and $f^{Reg}_{K^*N\Sigma^*}$ are same as those for the real $K$ and $K^*$ exchange [@oai:arXiv.org:0711.3536]. The same observation applies to the Reggeized K\* coupling. In this work we set them as free parameters. We expect the difference should not be very large, so the $K^*$ exchange is still very small and omitted in this work as Ref. [@Oh:2007jd]. The Reggeized treatment should work completely at high photon energies and interpolate smoothly to low energies. It is implemented by Toki $et\ al.$ [@Toki:2007ab] and Nam and Kao [@Nam:2010au] by introducing a weighting function ${\cal R}$. Here we adopt the treatment as, $$\begin{aligned} \label{eq:R} &&\frac{F_t}{t-m^{2}_{K}}\to \frac{F_t}{t-m^{2}_{K}} {\cal R}=\mathcal{D}_{K}R+\frac{F_t}{t-m^{2}_{K}}(1-{R}),\,\,\,\,\end{aligned}$$ where $R={R}_{s}{R}_{t}$ with $$\begin{aligned} \label{eq:RSRT} {R}_{s}&=& \frac{1}{2} \left[1+\tanh\left(\frac{s-s_{\mathrm{Reg}}}{s_{0}} \right) \right],\nonumber\\ {R}_{t}&=& 1-\frac{1}{2} \left[1+\tanh\left(\frac{|t|-t_{\mathrm{Reg}}}{t_{0}} \right) \right].\end{aligned}$$ The free parameters $s_{Reg}$, $s_0$, $t_{Reg}$ and $t_0$ will be fitted with the differential cross section. As inclusion of the from factor $F_i$, the Reggeized treatment will violate the gauge invariance and current conservation also. To restore the current conservation, we redefine the relevant amplitudes , $$\begin{aligned} \label{eq:WT1} i\mathcal{M}^{\mu\nu}&=&i\mathcal{M}^{\mu\nu}_{t}+i\mathcal{M}^{\mu\nu}_{s}+i\mathcal{M}^{\mu\nu}_{c} \to (i\mathcal{M}^{\mu\nu}_{t}+i\mathcal{M}^{\mu\nu}_{s} +i\mathcal{M}^{\mu\nu}_{c})\mathcal{R} \nonumber\\ &\equiv&i\mathcal{M}^{\mathrm{Regge}\ \mu\nu}_{t}+(i\mathcal{M}^{\mu\nu}_{s} +i\mathcal{M}^{\mu\nu}_{c})\mathcal{R} .\ \ \ \ \\end{aligned}$$ With such definition, the relation $k_1^\mu\mathcal{M}^{\mu\nu}=0$ is satisfied. For the nucleon resonance contributions, we adopt the Lagrangians for the radiative decay, $$\begin{aligned} \mathcal{L}_{\gamma N R(\frac{1}{2}^{\pm})} &=&\frac{e f_2}{2M_N} \bar{N} \Gamma^{(\mp)}\sigma_{\mu\nu}F^{\mu\nu} R \,+{\rm H.c.},\nonumber \\ \mathcal{L}_{\gamma N R(J^{\pm})} &=&\frac{-i^{n}f_1}{(2M_N)^{n}} \bar{N}~\gamma_\nu \partial_{\mu_2}\cdots\partial_{\mu_{n}} F_{\mu_1\nu}\Gamma^{\pm(-1)^{n+1}}R^{\mu_1\mu_2\cdots\mu_{n}}\nonumber\\ &+&\frac{-i^{n+1}f_2}{(2M_N)^{n+1}} \partial_{\nu}\bar{N} ~ \partial_{\mu_2}\cdots\partial_{\mu_{n}} F_{\mu_1\nu}\Gamma^{\pm(-1)^{n+1}}R^{\mu_1\mu_2\cdots\mu_{n}}\nonumber\\ &+&{\rm H.c.},\label{Eq:Lg}\end{aligned}$$ where $F^{\mu\nu}=\partial^\mu A\nu-\partial^\nu A_\mu$ with $R_{\mu_1\cdots\mu_n}$ is the field for the nucleon resonance with spin $J=n+1/2$, and $\Gamma^{(\pm)}=(i\gamma_5,1)$ for the different resonance parity. The Lagrangians here are also adopted from the previous works on nucleon resonances with spins $3/2$ or $5/2$ [@Nam:2010au; @Xie:2010yk; @Oh:2007jd]. The Lagrangians for the strong decay can be written as $$\begin{aligned} \mathcal{L}_{R(\frac{1}{2}^{\pm})K\Sigma^*} &=& \frac{h_2}{2m_{K}} \partial_\mu K\bar{\Sigma}^*_{\mu} \Gamma^{(\pm)}R,+{\rm h.c.}, \nonumber\\ \mathcal{L}_{R(J^\pm)K\Sigma^*} &=&\frac{-i^{n+1}h_1}{m_K^{n}} \bar{\Sigma}^*_{\mu_1}~\gamma_\nu\partial_\nu \partial_{\mu_2}\cdots\partial_{\mu_{n}} K\Gamma^{\pm(-1)^{n}}~R^{\mu_1\mu_2\cdots\mu_{n}}\nonumber\\ &+&\frac{-i^{n}h_2}{m_K^{n+1}} \bar{\Sigma}^*_{\alpha}~\partial_{\alpha}\partial_{\mu_1} \partial_{\mu_2}\cdots\partial_{\mu_{n}} K\Gamma^{\pm(-1)^{n}}~R^{\beta\mu_1\mu_2\cdots\mu_{n}}\nonumber\\ &+&{\rm H.c.}.\label{Eq:Ls}\end{aligned}$$ In this work the coupling constants $f_1$, $f_2$, $h_1$ and $h_2$ will be determined by the helicity amplitudes $A_{1/2}$ and $A_{3/2}$ and the decay amplitudes $G(\ell_1)$ and $G(\ell_2)$, which are obtained in the CQM. The interested reader is referred to Refs. [@Oh:2007jd; @He:2012ud] for further information. In this work the nucleon resonances $R$ including $N^*$ and $\Delta^*$ will be considered. The resonance field $R$ carries either isospin-$1/2$ or isospin-$3/2$. By omitting the space-time indices, the isospin structure of $RK\Sigma^*$ vertex reads as, $$\begin{aligned} \overline{R} \bm{\Sigma}^* \cdot \bm{\tau} K=p\Sigma^0{K}^+ -n{\Sigma}^0{K}^0 +\sqrt{2}n{\Sigma}^- K^+ +\sqrt{2}p\Sigma^+ K^0,\ \\end{aligned}$$ for resonance $R$ with isospin-$1/2$. If the resonance $R$ carries isospin-$3/2$, the effective Lagrangian has the isospin structure as $$\begin{aligned} \overline{R} \bm{\Sigma}^*\cdot\bm{T} K&=&\sqrt{3}\Delta^{++}\Sigma^+K^+-\sqrt{2}\Delta^+\Sigma^0 K^+ -\Delta^0\Sigma^- K^+\nonumber\\&-&\sqrt{2}\Delta^0\Sigma^0 K^0 +\Delta^+\Sigma^+ K^0-\sqrt{3}\Delta^-\Sigma^- K^0.\ \ \ \ \ \\end{aligned}$$ Results {#Sec: Results} ======= As shown in the previous section, we consider the $s$ channel with intermediate nucleon, the Reggeized $t$ channel with $K$ exchange, the $u$ channel with intermediate $\Lambda$, the contact term and the nucleon resonance intermediate $s$ channel in the $\Sigma(1385)$ photoproduction. By using the MINUIT code the differential cross section released by the CLAS Collaboration recently will be fitted with the help of the Lagrangians presented in the previous section. The free parameters involved and their fitted values are listed in Table \[Tab: Para\]. Here we exclude total cross section in the fitting procedure because it can be obtained by integrating the differential cross section. ------------------ --------------- ------------------------ ---------------- ------------------ --------------- $\Lambda_s$ $0.77\pm0.03$ $\Lambda_t$ $1.48\pm0.04$ $\Lambda_u$ $0.98\pm0.01$ $\Lambda_R$ $1.19\pm0.03$ $h$ $1.66\pm0.07$ $\sqrt{s_{Reg}}$ $2.10\pm0.01$ $s_0$ $0.29\pm0.27$ $\sqrt{t_{Reg}}$ $4.95\pm3.62$ $t_0$ $0.88\pm0.48$ $f^{Reg}_{KN\Sigma^*}$ $-4.74\pm0.02$ ------------------ --------------- ------------------------ ---------------- ------------------ --------------- : The free parameters used in fittting. The cut-offs $\Lambda_i$ are in the units of GeV, the parameters $s_{Reg}$, $s_0$, $t_{Reg}$ and $t_0$ for Reggeized treatment are in the units of GeV$^2$.\[Tab: Para\] As expected, the fitted values of cutoffs $\Lambda_i$ for the $s$ channel, the $t$ channel, the $u$ channel and the nucleon resonance contributions are close to 1 GeV. The $s_{Reg}$ is about 2.1 GeV, which indicte the Reggeized treatment play an important role even at the energies not so high as the $K$ photoprodcution with $\Lambda$ baryon in Ref. [@Corthals:2005ce]. The coupling constant $g^{Reg}_{KN\Sigma^*}$ for the Reggeized $t$ channel is about half larger than the values obtained by SU(3) symmetry, which is consistent with our expectation. As will be presented, the experimental CLAS data are well reproduced with $\chi^2=0.8$ per degree of freedom. If the systematic uncertainties are excluded, the best fitted $\chi^2$ per degree of freedom is $2.5$. The results of the cross section are similar to those with systematic uncertainty. Contributions from nucleon resonances ------------------------------------- First, we will present the contributions from nucleon resonances. As predicted in the CQM, for the $\Sigma(1385)K$ channel the decay widthsof nucleon resonances, such as $N(2095)$ and $\Delta(2000)$, are large and expected to play more important roles than other nucleon resonances [@Capstick:1998uh; @Capstick:1992uc]. In this work we use following criterion to select the resonances which will be considered in the fitting, $$\begin{aligned} \lambda=(A_{1/2}^2+A_{3/2}^2)(G(\ell_1)^2+G(\ell_2)^2)I^2\cdot10^{5}>\lambda_0,\end{aligned}$$ where helicity amplitudes $A_{1/2,3/2}$ and partial wave decay amplitudes $G(\ell)$ are in the units of $10^{-3}/\sqrt{\rm{GeV}}$ and $\sqrt{\rm{MeV}}$. A factor $10^5$ is introduced to make the largest value of $\lambda$ in the order of $10^0$. The isospin factor $I=1$ for $N^*$ and $\sqrt{2}$ for $\Delta^*$. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- State PDG $A^p_{1/2}$ $A^p_{3/2}$ $G(\ell_1)$ $G(\ell_2)$ $\lambda$ $\delta\chi^2$ $\delta\chi^2_r$ ------------------------------------------ ------------------------ ------------- ------------- ------------------------- --------------------------- ----------- ---------------- ------------------ $[N\textstyle{3\over 2}^-]_3(1960)$ $N(2120)D_{13}**$ 36 $-43$ $1.3 ^{+ 0.4}_{- $1.4 ^{+ 1.3}_{- 1.3}$ 1.1 0.2 \[0.8\] 0.1 \[0.4\] 0.4}$ $[N\textstyle{3\over 2}^-]_4(2055)$ 16 0 $-2.5 ^{+ $-2.5$ $^{+ 2.3}_{- 1.9}$ 0.3 0.0 \[0.0\] 0.0 \[0.0\] 1.0}_{-1.0}$ $[N\textstyle{3\over 2}^-]_5(2095)$ $-9$ $-14$ $7.7 ^{+1.2}_{-1.2}$ $-0.8$ $^{+0.7}_{-1.0}$ 1.7 0.4 \[3.3\] 0.1 \[0.1\] $[N\textstyle{3\over 2}^+]_3(1910)$ $-21$ $-27$ $-1.9 ^{+ 1.9}_{- 7.3}$ 0.0 $^{+ 0.0}_{- 0.4}$ 0.4 0.0 \[0.5\] 0.0 \[0.1\] $[N\textstyle{3\over 2}^+]_5(2030)$ $-9$ 15 2.2 $^{+ 1.0}_{- 1.9}$ $-0.2 ^{+ 0.2 0.0 \[0.0\] 0.0 \[0.0\] 0.1}_{- 0.3}$ $[N\textstyle{5\over 2}^+]_2(1980)$ $-11$ $-6$ $-3.6 ^{+ 2.5}_{- 3.0}$ $-0.1 ^{+ 0.1}_{- 0.2 0.2 \[0.5\] 0.1 \[0.3\] 0.3}$ $[\Delta\textstyle{3\over 2}^-]_2(2080)$ $\Delta(1940)D_{33}**$ $-20$ $-6$ $-4.1 ^{+ 4.0}_{- 1.5}$ $-0.5 ^{+ 0.5}_{- 2.2}$ 1.5 0.1 \[1.1\] 0.0 \[0.0\] $[\Delta\textstyle{3\over 2}^-]_3(2145)$ 0 10 5.2 $\pm $ 0.4 $-1.9 ^{+ 1.2}_{- 4.0}$ 0.6 0.2 \[1.0\] 0.0 \[0.0\] $[\Delta\textstyle{5\over 2}^+]_2(1990)$ $\Delta(2000)F_{35}**$ $-10$ $-28$ 4.0 $^{+ 4.5}_{- $-0.1 ^{+ 0.1}_{- 0.4}$ 2.8 1.5 \[14.7\] 0.7 \[4.5\] 4.0}$ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- \[Tab: Resonances\] First we choose several nucleon resonances in descending order of $\lambda$. If the contributions and influences of the nucleon resonances with small $\lambda$ is negligible, we stop here. If not, more resonances would be added. According to such criterion nine resonances survived with $\lambda_0=0.1$ are listed in Table. \[Tab: Resonances\]. For the masses of the nucleon resonances, the values suggested in the Particle Data Group (PDG) [@PDG] are adopted and for the nucleon resonances not listed in PDG, the prediction by the CQM will be adopted [@Capstick:1998uh; @Capstick:1992uc]. To prevent the proliferation of the free parameters, the Breit-Wigner widths for all nucleon resonances are set to 500 MeV, which is consistent to the widths for the $\Delta(2000)$ and $\Delta(1940)$ obtained in the multichannel partial-wave analysis [@Shrestha:2012ep; @Anisovich:2011fc]. As shown in Table \[Tab: Para\], the fitted value of cutoff for the nucleon resonances $\Lambda_R=1.19$ GeV with Gaussian from factor, that is, the form of from factor in Eq. (\[Eq: FF\]) with $n\to \infty$. In Fig. \[Fig: TCSR\], we present the total cross section of each nucleon resonance listed in Table \[Tab: Resonances\] to give an image of the magnitude of the corresponding nucleon resonance. Generally, the contributions from the nucleon resonances are smaller in the $\Sigma(1385)$ photoproduction compared with the contributions of nucleon resonances in the $\Lambda(1520)$ photoproduction [@He:2012ud]. The largest contribution is from $\Delta(2000)$ which have largest $\lambda$ about 3. The three nucleon resonances listed in the PDG, $N(2120)$, $\Delta(1940)$ and $\Delta(2000)$ have relatively large contributions among all nucleon resonances considered. The $N(2095)$ with largest decay width in the $\Sigma(1385)K$ channel has much smaller contribution than $\Delta(2000)$ due to its relative small radiatively decay width. In Table \[Tab: Resonances\], we list $\delta\chi^2$ and $\delta\chi^2_r$, which are the variations of the $\chi^2$ after turning off the corresponding resonance without and with refitting, respectively. It reflects the influence of the corresponding resonance on the reproduction of the experimental differential cross section. Generally, the variation of the $\chi^2$ is consistent with the value of $\lambda$. The resonances with $\lambda>1$, $N(2120)$, $N(2095)$ and $\Delta(2000)$, give $\delta\chi^2_r$ about or larger than 0.1 \[0.7\]. The $\Delta(2000)$ not only provides largest contribution to total cross section as shown in Fig. \[Fig: TCSR\], but also has largest influence on the $\chi^2$ with $\delta\chi^2_r=0.7~[4.5]$ after refitting. The influences of other resonances including the $N(2095)$ are much smaller than $\Delta(2000)$. The $\lambda$ of $N(2095)$ is large, about 1.7, while the $\delta\chi^2_r$ after refitting is about 0.1 which is much smaller than $\Delta(2000)$. It is due to the compensation effect from other resonances and (even) the Born terms in refitting. After a nucleon resonance turned off, the variation of the parameters after refitting will lead to the variation of the contributions from other resonances even the Born terms. The absence of the $N(2095)$ is smeared by such variation. The Contact term and the Reggeized treatment -------------------------------------------- In this section we will present more explicit information about the contact term and Reggeized treatment. As shown in Fig. \[Fig: tc\], the first term of the contact term in Eq. (\[Eq: contact\]), which comes from the Lagrangians given by Eq. (\[eq:lag1\]), play most dominant role at energies upto about 3 GeV. For $t$ channel, the $K$ exchange is dominant in low energies while the Regge contribution becomes dominant at energies higher than 2.5 GeV as we expected. Differential cross section -------------------------- With the nucleon resonance contributions and the Born terms given in the previous subsections, the results of differential cross section for the $\Sigma$(1385) photoproduction from proton compared with the CLAS data are shown in Fig. \[Fig: dcs\]. As shown in the figure, the experimental data are well reproduced in our model. The contributions from the $u$ channel and the contact term are dominant and is responsible for the behaviors of the differential cross section at backward and forward angles, respectively. The $t$ channel contribution is smaller but give considerable contribution at forward angles. The Born $s$ channel contribution is very small. Compared with the plausible results at forwards angles, the results at backward angles are not so satisfactory. We have tried to introduce the Reggeized treatment $u$ channel contribution. But as mentioned in Section II, the large uncertainty at backward angles make it difficult to give an meaningful determination of extra five parameters required by Reggeized treatment. Hence, we keep the $\Lambda$ intermediate $u$ channel in this work. The further experimental data at extreme backward with high precision will be helpful to deepen the understanding about the interaction mechanism in the $u$ channel. Total cross section ------------------- We also present the theoretical results of total cross section compared with the CLAS data in Fig. \[Fig: TCS\]. One can find that our result is well comparable with the CLAS data. At all energies, the contact term provides most important contribution, and the Reggeized $t$ channel contribution is large near threshold and decreases rapidly at higher energies. The $u$ channel contribution becomes important at higher energies. The contributions from the nucleon resonances are small. But as shown in Table \[Tab: Resonances\], it is essential to reproduce the differential cross section. The $\Delta$(2000) has magnitude comparable to the $t$ channel and the $u$ channel at $E_\gamma$ about 2.1 GeV. Summary ======= The $\Sigma$(1385) photoproduction in the $\gamma p \to K^+\Sigma^0(1385)$ reaction is investigated within a Regge-plus-resonance approach. The contact term is dominant in the interaction mechanism and Reggeized $t$ channel is important at energies near threshold at forward angles. The $u$ channel is responsible for the behavior of differential cross section at backward angles. The contributions of nucleon resonances are determined by the radiative and strong decay amplitudes predicted from the CQM. The results show that the contributions from nucleon resonances are small compared with the contact term, $u$ and $t$ channel contributions but essential to reproduce the experimental data. The $D_{13}$ state $N(2095)$ which is expected to be important in $\Sigma(1385)$ photoproduction have much smaller contribution for total cross section and smaller influence on the reproduction of differential cross section than $F_{35}$ state $\Delta(2000)$. The resonance $\Delta(2000)$ is the most important nucleon resonance in $\Sigma(1385)$ photoproduction as suggested by CQM [@Capstick:1998uh; @Capstick:1992uc], which is also consistent with the results in Ref. [@Oh:2007jd]. Acknowledgement {#acknowledgement .unnumbered} =============== The author wants to thank Prof. Schumacher for valuable comments on the manuscript. This project is partially supported by the Major State Basic Research Development Program in China (No. 2014CB845405), the National Natural Science Foundation of China (Grants No. 11275235, No. 11035006) and the Chinese Academy of Sciences (the Knowledge Innovation Project under Grant No. KJCX2-EW-N01). [38]{} T. P. Vrana, S. A. Dytman and T. S. H. Lee, Phys. Rept.  [**328**]{}, 181 (2000) N. Isgur and G. Karl, Phys. Rev. D [**18**]{}, 4187 (1978) S. Capstick and N. Isgur, Phys. Rev. D [**34**]{}, 2809 (1986) E. Klempt and J. -M. Richard, Rev. Mod. Phys.  [**82**]{}, 1095 (2010) K. Moriya [*et al.*]{} \[CLAS Collaboration\], Phys. Rev. C [**88**]{}, 045201 (2013) S. Capstick and W. Roberts, Phys. Rev. D [**58**]{}, 074011 (1998) S. Capstick, Phys. Rev. D [**46**]{}, 2864 (1992) J. J. Xie and J. Nieves, Phys. Rev.  C [**82**]{}, 045205 (2010) J. He and X. -R. Chen, Phys. Rev. C [**86**]{}, 035204 (2012) Cambridge Bubble Chamber Group, Phys. Rev.  [**156**]{}, 1426 (1967) R. Erbe [*et al.*]{} , Phys. Rev.  [**188**]{}, 2060 (1969) M. Niiyama, [*et al.*]{}, Phys. Rev. C [**78**]{}, 035202 (2008) Y. Oh, C. M. Ko, K. Nakayama, Phys. Rev.  [**C77**]{}, 045204 (2008) P. Gao, J. -J. Wu and B. S. Zou, Phys. Rev. C [**81**]{}, 055203 (2010) T. Corthals, J. Ryckebusch and T. Van Cauteren, Phys. Rev. C [**73**]{}, 045207 (2006) K. Ohta, Phys. Rev. C [**40**]{}, 1335 (1989) S. I. Nam and C. W. Kao, Phys. Rev.  C [**81**]{}, 055206 (2010) H. Haberzettl, K. Nakayama and S. Krewald, Phys. Rev. C [**74**]{}, 045202 (2006) A. I. Titov, B. Kampfer, S. Date and Y. Ohashi, Phys. Rev.  C [**72**]{}, 035206 (2005) \[Erratum-ibid.  C [**72**]{}, 049901 (2005)\] M. Guidal, J. M. Laget and M. Vanderhaeghen, Nucl. Phys.  A [**627**]{}, 645 (1997) H. Toki, C. Garcia-Recio and J. Nieves, Phys. Rev. D [**77**]{}, 034001 (2008) \[arXiv:0711.3536 \[hep-ph\]\]. H. Toki, C. Garcia-Recio and J. Nieves, Phys. Rev.  D [**77**]{}, 034001 (2008) J. Beringer et al. (Particle Data Group), Phys. Rev. [**D**]{} 86, 010001 (2012) and 2013 partial update for the 2014 edition M. Shrestha and D. M. Manley, Phys. Rev. C [**86**]{}, 055203 (2012) A. V. Anisovich, [*et al.*]{}, Eur. Phys. J. A [**48**]{}, 15 (2012)

--- abstract: 'We develop a simple model for the formation energies (FEs) of alkali and alkaline earth alanates and boranates, based upon ionic bonding between metal cations and AlH$_4^-$ or BH$_4^-$ anions. The FEs agree well with values obtained from first principles calculations and with experimental FEs. The model shows that details of the crystal structure are relatively unimportant. The small size of the BH$_4^-$ anion causes a strong bonding in the crystal, which makes boranates more stable than alanates. Smaller alkali or alkaline earth cations do not give an increased FE. They involve a larger ionization potential that compensates for the increased crystal bonding.' author: - 'Michiel J. van Setten' - Geert Brocks - 'Gilles A. de Wijs' title: A model for the formation energies of alanates and boranates --- The large scale utilization of hydrogen as a fuel crucially depends on the development of compact storage materials with a high mass content of hydrogen.[@bog_rev] Over the last decade alanates and boranates have been studied extensively because of their potential use as hydrogen storage materials.[@bog_rev; @zuttel04] These materials consist of a lattice of metal cations and AlH$_4^-$ or BH$_4^-$ anions, respectively. The ideal hydrogen storage material should have a high gravimetric hydrogen density, which requires the use of light metals. Moreover, the formation energy (FE) of such a material has to be such that it is stable at room temperature, yet it has to decompose at low temperature to release its hydrogen. In principle a large variety of alanates and boranates can be synthesized by changing the metal cations, which can be used to tune the formation energy.[@nakamori06] Since synthesis is a very time consuming effort, there is a need for a materials specific theory with a predictive power for the FE. At present the state of the art is formed by first principles calculations based upon density functional theory (DFT). Several papers have been dedicated to trends in the DFT FEs of alanates and boranates.[@nakamori06; @chun; @vajeeston04; @lovvik05rev; @vajeeston05] There exists a surprising variety of crystal structures among these compounds. In DFT calculations the crystal structure with the lowest energy has to be searched for each compound, and the cell parameters and the atomic positions have to be optimized. This procedure also makes DFT calculations a very time consuming effort. A simple theory would help to understand the trend in the FEs of alanates and boranates. Our aim is to construct a simple model for the FEs at 0 K of alkali alanates and boranates (MAH$_4$, M = Li, Na, K; A = Al, B) and of their alkaline earth counterparts (M$'$(AH$_4$)$_2$, M = Mg, Ca), avoiding the use of the actual crystal structure. We assume that these compounds can be described by ionic bonding between M$^+$ or M$'^{2+}$ cations and AH$_4^-$ anions. Our model for the FE, $\Delta E_f$, is based upon a Born-Haber cycle,[@born19] $$\Delta E_f = E_\mathrm{elem} + E_\mathrm{ions} + E_\mathrm{crys}. \label{eq:hof}$$ Starting from bulk elemental solids and H$_2$ molecules, $E_\mathrm{elem}$ is the energy required to atomize the solids and the molecules. The $E_\mathrm{elem}$ of the bulk solids are listed in Table \[param\]. We use a value of 4.48 eV for the dissociation energy of H$_2$.[@hcp] The second step is to create M$^+$, M$'^{2+}$ and AH$_4^-$ ions from the atoms, represented by the energy $E_\mathrm{ions}$. The contribution to $E_\mathrm{ions}$ from the M$^+$ ions is simply the first ionization potential (IP) and from the M$'^{2+}$ ions it is the sum of the first and second IPs. The numbers $\Sigma IP$ are given in Table \[param\]. We calculate the contribution to $E_\mathrm{ions}$ from the AH$_4^-$ anions as follows. First an electron is added to an Al or B atom, which lowers the energy by the atomic electron affinity (EA). These atoms then have four valence electrons that are used to form covalent bonds with four hydrogen atoms. Using EAs of 0.44 eV and 0.28 eV for Al and B, and 2.91 eV and 3.45 eV for the Al-H and B-H bond strengths,[@hcp] we calculate FEs of $-12.08$ eV and $-14.08$ eV for the AlH$_4^-$ and BH$_4^-$ anions. The final step consists of constructing the crystal from the M$^+$ (or M$'^{2+}$) and AH$_4^-$ ions, which is represented by the energy $E_\mathrm{crys}$. We use a simple Born model for the potential between cations and anions. It consists of an attractive Coulomb potential between point charges at the centers of the ions plus a repulsive short-range potential $\propto r^{-\bar{n}}$, where $\bar{n}$ is the average Born exponent. $E_\mathrm{crys}$ is then given by[@ssp16] $$E_\mathrm{crys} = \frac{M_c Z_A Z_C e^2}{4\pi \varepsilon_0 r_0}\left(1-\frac{1}{\bar{n}}\right), \label{eq:born}$$ where $Z_A=-1$ and $Z_C=+1,+2$ are the valencies of the anions and the cations, respectively, and $r_0$ is the shortest cation-anion distance in the lattice. $M_c$ represents the Madelung constant, which depends upon the type of lattice.[@ssp16] Li Na K Mg Ca B Al -------------------- ------ ------ ------ ------- ------- ------ ------ -- $R_{\textrm{ion}}$ 0.90 1.16 1.52 0.86 1.14 $\Sigma IP$ 5.39 5.14 4.34 22.67 17.98 $E_{\textrm{dis}}$ 1.64 1.08 0.93 1.48 1.81 5.81 3.38 $n$ 5 7 9 7 9 7 9 : \[param\] Ionic radii, $R_{\textrm{ion}}$ (Å), summed ionization potentials, $\Sigma IP$ (eV), dissociation energies for the elemental bulks, $E_{\textrm{dis}}$ (eV), and the Born exponents, $n$. Values are taken from Refs.  and   Note that in all compounds considered here the AlH$_4^-$ and BH$_4^-$ anions have a tetrahedral geometry. However, in the Born model of Eq. (\[eq:born\]), we have approximated these tetrahedra by spheres. Our motivation for this is that we are interested in a simple model of $E_\mathrm{crys}$ without having to take into account the full details of the crystal structure. The cation-anion distance then is the sum of the ionic radii of the cation and the anion, $r_0=r_C+r_A$. Since the cations we consider are mostly octahedrally coordinated, we use standard ionic radii $r_C$ of 6-fold coordinated alkaline and alkaline earth ions, see Table \[param\].[@shannon76] As the radius of the anions $r_A$ we use the Al-H and B-H bond lengths, which are 1.62 Å and 1.20 Å, respectively. $r_A$ then roughly corresponds to the average of the maximum radius of an AH$_4^-$ ion (A = Al, B) and the minimum radius, which is the radius of the central atom A. The values obtained for the cation-anion distance $r_0$ then correspond to the average of the cation-Al/B and cation-H distances in the crystal. Avoiding the full details of the crystal structure also leads to using average values for the Madelung constants $M_c$ in Eq. (\[eq:born\]). The alkali alanates and boranates have an AB type lattice, where A is the alkali cation, and B is the boranate or alanate anion. The variation of the Madelung constant over different AB lattices is relatively small, so we use an average value $M_c=1.76$. The root mean square deviation (rms) averaged over all AB lattice types is 4%. A similar reasoning holds for the alkaline earth alanates and boranates. They have an AB$_2$ lattice, whose average Madelung constant is $M_c=2.40$ with a rms deviation of 4%. Fig. \[hofvaspmodexp\] shows the most important results, i.e. the FE calculated with the model represented by Eqs. (\[eq:hof\]) and (\[eq:born\]), compared to experimental values.[@smith63] For some of the materials the experimental FE is not known. Therefore we have also performed first principles DFT calculations. We use the projector augmented wave (PAW) method,[@paw; @blo] and the PW91 generalized gradient approximation (GGA),[@pw91] as implemented in the Vienna *ab initio* simulation package (VASP).[@vasp1; @vasp2; @vasp3; @vasp4] To integrate the Brillouin zone we apply the tetrahedron scheme. The $\mathbf{k}$-point mesh and the plane wave kinetic energy cutoff (700 eV) are chosen such, that total energies are converged to a numerical accuracy of 1 meV per formula unit.[@zpe] The structures of LiAlH$_4$,[@vansetten06] NaAlH$_4$,[@vansetten06] Mg(AlH$_4$)$_2$,[@vansetten06] Ca(AlH$_4$)$_2$,[@lovvik05] Ca(BH$_4$)$_2$,[@miwa:155122] KAlH$_4$,[@vajeeston04-2] and of the alkali boranates[@vajeeston05] are taken from the literature. We additionally relaxed the atomic positions, but the relaxations were small and had only a minor effect on the total energies. Our calculated values compare well to those obtained in previous calculations.[@nakamori06; @vajeeston05; @lovvik05rev; @chun; @chou] Details on Mg(BH$_4$)$_2$ will be published elsewhere.[@vansetten06-2]** ![\[hofvaspmodexp\] Model formation energies (eV/AH$_4$) compared to DFT and experimental values.[@smith63]](hofvaspmodexp.ps){width="8.5cm"} Fig. \[hofvaspmodexp\] shows that, despite its simplicity, the model gives FEs that are in quantitative agreement with both the experimental, and the calculated DFT values. The rms deviation of the model with the experimental and the DFT values is 0.27 and 0.33 eV/AH$_4$, respectively. Note that these numbers are comparable to the rms deviation between the experimental and the first principles values, 0.19 eV/AH$_4$, which represents the state-of-the-art. An obvious source of error is our neglect of the details of the crystal structure, e.g., by using an average Madelung constant in Eq. (\[eq:born\]). Changing the Madelung constant by 5% changes the FE of the alkali compounds by 0.4 eV/AH$_4$ and that of the alkaline earth compounds by 0.7 eV/AH$_4$. As these numbers are larger than the rms deviation of the model, one can conclude that the details of the crystal structure are relatively unimportant. The model also seems to work reasonably well for some other boranates. The model FE for Sc(BH$_4$)$_3$ is 0.37 eV/BH$_4$ higher than the DFT value calulated by Nakamori *et. al*,[@nakamori06] which is within the rms error bar given above. The model FE for Zn(BH$_4$)$_2$ and CuBH$_4$ are 0.6 and 0.8 eV/BH$_4$ higher than the DFT values, respectively. For cations with a nominal charge $Z_C=4$, such as Zr or Hf, the model breaks down. The model FE then deviates by 2.5 eV/BH$_4$ from the DFT values.[@nakamori06] As can be seen from Eq. (\[eq:born\]), for large $Z_C$ $E_\mathrm{crys}$ becomes sensitive to small changes in the Madelung constant and the ionic radius of the cation, or in other words, to the details of the crystal structure.** Fig. \[hofvaspmodexp\] shows that boranates are generally more stable than alanates. The origin of this stability can be analyzed by decomposing the FE into the contributions according to Eq. (\[eq:hof\]), which is shown Fig. \[bar\]. The differences in formation energy of the elements $E_\mathrm{elem}$ are to a large degree compensated by the differences in the formation energies $E_\mathrm{ions}$ of the ions from the atoms. The ionic crystal energy $E_\mathrm{crys}$ of the boranates is however significantly larger than that of the alanates, which results in a larger stability of the latter. This is a size effect since the BH$_4^-$ anions are significantly smaller than the AlH$_4^-$ anions. ![\[bar\]The contributions to the formation energy (eV/AH$_4$) according to Eq. (\[eq:hof\]).](bar.ps){width="8.5cm"} It has been observed that the dissociation energies of complex alkali hydrides into simple alkali hydrides increase with the atomic number of the alkali atom.[@arr] For the FEs from the elements the overall trend is not that clear. $E_\mathrm{elem}$ and $E_\mathrm{ions}$ both decrease with increasing atomic number, see Fig. \[bar\], which increases the stability. However, this is almost compensated by $E_\mathrm{crys}$, which increases with the cation radius $r_C$. In the alkaline earth series the FE decreases with the atomic number. The dominant effect is a decreasing $E_\mathrm{ions}$, which is due to a decrease in the ionization potentials of the cations, see Table \[param\]. To summarize we constructed a model for the formation energies (FEs) of alkali and alkaline earth alanates and boranates from the elemental solids and H$_2$ molecules. The model is based upon ionic bonding between metal cations and AlH$_4^-$ or BH$_4^-$ anions. It can be constructed using simple energy values that are available in the literature and it does not make use of explicit crystal structure information. Compared to experimental values, the model FEs have a similar accuracy as calculated DFT values. The trends in the FEs over the series of compounds can be analyzed in terms of the individual contributions to the model. The authors wish to thank R. A. de Groot for useful and stimulating discussions. This work is part of the research programs of ‘Advanced Chemical Technologies for Sustainability (ACTS)’ and the ‘Stichting voor Fundamenteel Onderzoek der Materie FOM)’, both financially supported by the ‘Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO)’. [10]{} F. Schüth, B. Bogdanovic, and M. Felderhoff, Chem. Commun. [**20**]{}, 2249 (2004). A. Züttel, P. Wenger, P. Sudan, P. Mauron, and S. I. Orimo, Mater. Sci. Eng. B-Solid State Mater. Adv. Technol. [**108**]{}, 9 (2004). Y. Nakamori et al., Phys. Rev. B [**74**]{}, 045126 (2006). S. C. Chung and H. Morioka, J. Alloys Compd. [**372**]{}, 92 (2004). P. Vajeeston, P. Ravindran, R. Vidya, H. [Fjellvåg]{}, and A. Kjekshus, Cryst. Growth Des. [**4**]{}, 471 (2004). O. M. [Løvvik]{}, O. Swang, and S. M. Opalka, J. Mater. Res. [**20**]{}, 3199 (2005). P. Vajeeston, P. Ravindran, A. Kjekshus, and H. [Fjellvåg]{}, J. Alloy. Compd. [**387**]{}, 97 (2005). M. Born, Verhandl. Deut. Physik. Ges. [**21**]{}, 679 (1919). R. C. Weast (ed.) and M. J. Astle (ed.), , (CRC press, New York, 62 edition, 1982), F 190-201. M.P. Tosi, F. Seitz (ed.) and D. Turnbull (ed.), , (Academic press, New York and London, 1965), Chap. 1, 1-113. R. Shannon, Acta. Cryst. A [**32**]{}, 751 (1976). M. Smith and G. [Bass Jr.]{}, J. Chem. Eng. Data [**8**]{}, 342 (1963). P. E. Blöchl, Phys. Rev. B [**50**]{}, 17953 (1994). J. P. Perdew and J. A. Chevary and S. H. Vosko and K. A. Jackson and M. R. Pederson and D. J. Singh and C. Fiolhais, Phys. Rev. B, [**46**]{}, 6671 (1992). G. Kresse and D. Joubert, Phys. Rev. B [**59**]{}, 1758 (1999). G. Kresse and J. Hafner, Phys. Rev. B [**49**]{}, 14251 (1994). G. Kresse and J. Hafner, Phys. Rev. B [**47**]{}, 558 (1993). G. Kresse and J. Furthmüller, Comput. Mater. Sci. [**6**]{}, 15 (1996). G. Kresse and J. Furthmüller, Phys. Rev. B [**54**]{}, 11169 (1996). In both the DFT calculations and the model no zero point energies are taken into account. M. J. [van Setten]{}, G. A. [de Wijs]{}, V. A. Popa, and G. Brocks, Phys. Rev. B [**74**]{}, in press, cont-math/0609189 (2006). O. M. [Løvvik]{}, Phys. Rev. B [**71**]{}, 144111 (2005). K. Miwa et al., Phys. Rev. B [**74**]{}, 155122 (2006). P. Vajeeston, P. Ravindran, A. Kjekshus, and H. [Fjellvåg]{}, J. Alloy. Compd. [**363**]{}, L7 (2004). A. Peles, J. A. Alford, [Zhu Ma]{}, [Li Yang]{}, and M. Y. Chou, Phys. Rev. B [**70**]{}, 165105 (2004). M. J. [van Setten]{}, G. A. [de Wijs]{}, and G. Brocks, in preparation (2006). M. E. [Arroyo y de Dompablo]{} and G. Ceder, J. Alloys Compd. [**364**]{}, 6 (2003).

--- abstract: 'We examine the prospects for discovering new physics through muon dipole moments. The current deviation in $g_{\mu}-2$ may be due entirely to the muon’s [*electric*]{} dipole moment. We note that the precession frequency in the proposed BNL muon EDM experiment is also subject to a similar ambiguity, but this can be resolved by up-down asymmetry measurements. We then review the theoretical expectations for the muon’s electric dipole moment in supersymmetric models.' author: - 'Jonathan L. Feng' - 'Konstantin T. Matchev' - Yael Shadmi date: 'October 11, 2001' title: 'Muon Dipole Moment Experiments: Interpretation and Prospects' --- Introduction {#intro} ============ The Standard Model of particle physics provides an extremely successful description of all known particles and their interactions, but fails to address many deeper questions concerning their physical origin. Among the least understood phenomena is CP violation. At present, the only observed source of CP violation in the Standard Model is the phase of the CKM matrix. Its fundamental origins are unknown. Further, while CP violation is an essential ingredient of almost all attempts to explain the matter-antimatter asymmetry of the universe [@Sakharov:1967dj] (alternative explanations are subject to stringent bounds: see, [[*e.g.*]{}]{}, Ref. [@Cohen:1998ac]), the amount of CP violation present in the CKM matrix is insufficient to explain the observed asymmetry [@asymmetry]. Searches for CP violation beyond the CKM matrix are necessary to shed light on this puzzle and are also probes of physics beyond the Standard Model. Electric dipole moments (EDMs) violate both parity (P) and time reversal (T) invariance. If CPT is assumed to be an unbroken symmetry, a permanent EDM is, then, a signature of CP violation [@bb]. A non-vanishing permanent EDM has not been measured for any of the known elementary particles. In the Standard Model, EDMs are generated only at the multi-loop level and are predicted to be many orders of magnitude below the sensitivity of foreseeable experiments [@hk]. A non-vanishing EDM therefore would be unambiguous evidence for CP violation beyond the CKM matrix, and searches for permanent EDMs of fundamental particles are powerful probes of extensions of the Standard Model. In fact, current EDM bounds are already some of the most stringent constraints on new physics, and they are highly complementary to many other low energy constraints, since they require CP violation, but not flavor violation. The field of precision muon physics will be transformed in the next few years [@Hawaiiproc]. The EDM of the muon is therefore of special interest. A new BNL experiment [@Semertzidis:1999kv] has been proposed to measure the muon’s EDM at the level of $${d_{\mu}}\sim 10^{-24}~{e~\text{cm}}\ , \label{proposedEDM}$$ more than five orders of magnitude below the current bound [@Bailey:1979mn] $${d_{\mu}}= (3.7 \pm 3.4) \times 10^{-19}~{e~\text{cm}}\ , \label{currentEDM}$$ and even higher precision might be attainable at a future neutrino factory complex [@Aysto:2001zs]. The interest in the muon’s EDM is further heightened by the recent measurement of the muon’s anomalous magnetic dipole moment (MDM) ${a_{\mu}}= (g_{\mu}-2)/2$, where $g_{\mu}$ is the muon’s gyromagnetic ratio. The current measurement ${a_{\mu}^{\text{exp}}}= 11\ 659\ 202\, (14)\, (6) \times 10^{-10}$ [@Brown:2001mg] from the Muon $(g-2)$ Experiment at Brookhaven differs from the Standard Model prediction ${a_{\mu}^{\text{SM}}}$ [@Davier:1998si; @Marciano:2001qq] by $2.6 \sigma$: $$\Delta a_\mu \equiv {a_{\mu}^{\text{exp}}}- {a_{\mu}^{\text{SM}}}= (43 \pm 16) \times 10^{-10} \ . \label{currentamu}$$ The muon’s EDM and MDM arise from similar operators, and this tentative evidence for a non-Standard Model contribution to ${a_{\mu}}$ also motivates the search for the muon’s EDM [@Feng:2001sq]. In fact, the deviation of [Eq. (\[currentamu\])]{} may be partially, or even entirely attributed to a muon EDM! [@Feng:2001sq] In Section \[sec:experimental\] we discuss the interplay between the new physics contributions to the muon MDM and EDM, and their manifestation in muon dipole moment experiments. Then in Section \[sec:theoretical\] we present model-independent predictions for the muon EDM, based on the current $g_\mu-2$ measurement. Finally in Section \[sec:susy\] we review the theoretical expectations for the size of the muon EDM in supersymmetry. Interpretation of Muon Dipole Experiments {#sec:experimental} ========================================= Modern measurements of the muon’s MDM exploit the equivalence of cyclotron and spin precession frequencies for $g=2$ fermions circulating in a perpendicular and uniform magnetic field. Measurements of the anomalous spin precession frequency are therefore interpreted as measurements of ${a_{\mu}}$. The spin precession frequency also receives contributions from the muon’s EDM, however. For a muon traveling with velocity ${{\text{\normalsize\boldmath $\beta$}}}$ perpendicular to both a magnetic field ${{\text{\normalsize\boldmath $B$}}}$ and an electric field ${{\text{\normalsize\boldmath $E$}}}$, the anomalous spin precession vector is $$\begin{aligned} {{\text{\normalsize\boldmath $\omega$}}}_a &=& -a_{\mu} \frac{e}{m_{\mu}} {{\text{\normalsize\boldmath $B$}}} - d_{\mu} \frac{2c}{\hbar} {{\text{\normalsize\boldmath $\beta$}}} \times {{\text{\normalsize\boldmath $B$}}} - d_{\mu} \frac{2}{\hbar} {{\text{\normalsize\boldmath $E$}}} \nonumber \\ &&- \frac{e}{m_{\mu}c} \left(\frac{1}{\gamma^2-1} - a_{\mu}\right) {{\text{\normalsize\boldmath $\beta$}}} \times {{\text{\normalsize\boldmath $E$}}} \ . \label{omega}\end{aligned}$$ In recent experiments, the last term of [Eq. (\[omega\])]{} is removed by running at the ‘magic’ $\gamma \approx 29.3$, and the third term is negligible. For highly relativistic muons with $|{{\text{\normalsize\boldmath $\beta$}}}| \approx 1$, then, the anomalous precession frequency is found from $${|{{\text{\normalsize\boldmath $\omega$}}}_a| \over |{{\text{\normalsize\boldmath $B$}}}|} \approx \left[ \left( \frac{e}{m_{\mu}} \right)^2 \left({a_{\mu}^{\text{SM}}}+ {a_{\mu}^{\text{NP}}}\right)^2 + \left(\frac{2c}{\hbar}\right)^2 {{d_{\mu}^{\text{NP}}}}^2 \right]^{1/2} \ , \label{both}$$ where NP denotes new physics contributions, and we have assumed ${d_{\mu}^{\text{NP}}}\gg {d_{\mu}^{\text{SM}}}$. The observed deviation from the Standard Model prediction for $|{{\text{\normalsize\boldmath $\omega$}}}_a|$ has been assumed to arise entirely from a MDM and has been attributed to a new physics contribution of size $\Delta a_\mu$. However, from [Eq. (\[both\])]{}, we see that, more generally, it may arise from some combination of magnetic and electric dipole moments from new physics. More quantitatively, the effect can also be due to a combination of new physics MDM and EDM contributions satisfying $$\begin{aligned} \left| {d_{\mu}^{\text{NP}}}\right| &\approx& \frac{\hbar e}{2 m_{\mu} c} \, \sqrt{\ 2\, {a_{\mu}^{\text{SM}}}\, \left(\Delta a_\mu - {a_{\mu}^{\text{NP}}}\right)} \nonumber \\ &\approx& 3.0 \times 10^{-19}~{e~\text{cm}}\ \sqrt{1 - \frac{{a_{\mu}^{\text{NP}}}}{43 \times 10^{-10}}} \ , \label{mdmisedm}\end{aligned}$$ where we have taken into account that ${a_{\mu}^{\text{NP}}}\ll {a_{\mu}^{\text{SM}}}$ and normalized ${a_{\mu}^{\text{NP}}}$ to the current central value given in [Eq. (\[currentamu\])]{}. In Fig. \[fig:amu\_dmu\] we show the regions in the $({a_{\mu}^{\text{NP}}},{d_{\mu}^{\text{NP}}})$ plane that are consistent with the observed deviation in $|{{\text{\normalsize\boldmath $\omega$}}}_a|$. The current 1$\sigma$ and 2$\sigma$ upper bounds on ${d_{\mu}^{\text{NP}}}$ [@Bailey:1979mn] are also shown. We see that a large fraction of the region allowed by both the current $g_\mu-2$ measurement [Eq. (\[currentamu\])]{} and the $d_\mu$ bound [Eq. (\[currentEDM\])]{} is already within the sensitivity of phase I of the newly proposed experiment (with sensitivity $\sim 10^{-22}\ {e~\text{cm}}$). ![Regions in the $({a_{\mu}^{\text{NP}}},{d_{\mu}^{\text{NP}}})$ plane that are consistent with the observed $|{{\text{\normalsize\boldmath $\omega$}}}_a|$ at the 1$\sigma$ and 2$\sigma$ levels. The current 1$\sigma$ and 2$\sigma$ bounds on ${d_{\mu}^{\text{NP}}}$ [@Bailey:1979mn] are also shown.[]{data-label="fig:amu_dmu"}](amu_dmu.eps){height="2.3in"} In fact, the observed anomaly may, in principle, be due entirely to the muon’s EDM! This is evident from Eqs. (\[currentEDM\]) and (\[mdmisedm\]), or from Fig. \[fig:amu\_dmu\]. Alternatively, in the absence of fine-tuned cancellations between ${a_{\mu}^{\text{NP}}}$ and ${d_{\mu}^{\text{NP}}}$, [*the results of the Muon $(g-2)$ Experiment also provide the most stringent bound on ${d_{\mu}}$ to date*]{}, with $1\sigma$ and $2\sigma$ upper limits $$\begin{aligned} \Delta a_{\mu} &<& 59 \ (75) \times 10^{-10} \Longrightarrow \nonumber \\ \left| {d_{\mu}^{\text{NP}}}\right| &<& 3.5 \ (3.9) \times 10^{-19}~{e~\text{cm}}\ . \label{newdmubound}\end{aligned}$$ Of course, the effects of ${d_{\mu}}$ and ${a_{\mu}}$ are physically distinguishable: while ${a_{\mu}}$ causes precession around the magnetic field’s axis, ${d_{\mu}}$ leads to oscillation of the muon’s spin above and below the plane of motion. This oscillation is detectable in the distribution of positrons from muon decay, and further analysis of the recent ${a_{\mu}}$ data should tighten the current bounds on ${d_{\mu}}$ significantly. Such analysis is currently in progress [@lee] and should be able to further restrict the allowed region depicted in Fig. \[fig:amu\_dmu\]. The proposed dedicated muon EDM experiment will use a different setup from the one described above, by applying a constant radial electric field. As can be seen from [Eq. (\[omega\])]{}, the anomalous precession frequency will then have both a radial component, $$- d_{\mu} \frac{2c}{\hbar} {{\text{\normalsize\boldmath $\beta$}}} \times {{\text{\normalsize\boldmath $B$}}} - d_{\mu} \frac{2}{\hbar} {{\text{\normalsize\boldmath $E$}}} \ , \label{omega_rad}$$ and a vertical component, $$-a_{\mu} \frac{e}{m_{\mu}} {{\text{\normalsize\boldmath $B$}}} - \frac{e}{m_{\mu}c} \left(\frac{1}{\gamma^2-1} - a_{\mu}\right) {{\text{\normalsize\boldmath $\beta$}}} \times {{\text{\normalsize\boldmath $E$}}} \ . \label{omega_ver}$$ Then for any given $\gamma$, and [*assuming the SM value for ${a_{\mu}}$*]{}, the electric field can be tuned to cancel the precession from [Eq. (\[omega\_ver\])]{} due to $a_\mu$. The remaining radial component of ${{\text{\normalsize\boldmath $\omega$}}}_a$ will lead to an oscillating up-down asymmetry in the counting rate. Measurements of both the asymmetry and the spin precession frequency can be used to deduce a limit on ${d_{\mu}^{\text{NP}}}$. As in the $g_{\mu}-2$ experiment, however, the measurement of the spin precession frequency in the muon EDM experiment receives, in principle, contributions from both the muon EDM and MDM. In the presence of a sizable new physics contribution to ${a_{\mu}}$, the cancellation in [Eq. (\[omega\_ver\])]{} is not perfect, leaving a residual radial component $$-{a_{\mu}^{\text{NP}}}\frac{e}{m_{\mu}} \left( {{\text{\normalsize\boldmath $B$}}} - \frac{1}{c}\ {{\text{\normalsize\boldmath $\beta$}}} \times {{\text{\normalsize\boldmath $E$}}} \right) \ . \label{omega_res}$$ From Eqs. (\[omega\_rad\]) and (\[omega\_res\]) we then obtain for the magnitude of the anomalous precession frequency $$\begin{aligned} &&|{{\text{\normalsize\boldmath $\omega$}}}_a|^2 = |{{\text{\normalsize\boldmath $B$}}}|^2 \left[ \left( {a_{\mu}^{\text{NP}}}\frac{e}{m_{\mu}} \right)^2 \left( 1 - \frac{{a_{\mu}^{\text{SM}}}}{{a_{\mu}^{\text{SM}}}- \frac{1}{\gamma^2 - 1}} \right)^2 \right. \nonumber \\ &&+ \left. \left( {d_{\mu}^{\text{NP}}}\frac{2}{\hbar} \right)^2 \left( c |{{\text{\normalsize\boldmath $\beta$}}}| + \frac{{a_{\mu}^{\text{SM}}}}{\frac{|{{\text{\normalsize\boldmath $\beta$}}}|}{c} \left( {a_{\mu}^{\text{SM}}}- \frac{1}{\gamma^2 - 1} \right)} \right)^2 \right] ,\end{aligned}$$ where we have used the tuning condition for [Eq. (\[omega\_ver\])]{} to eliminate the electric field. In the setup of the proposed experiment, $\gamma \approx 5$, and we can approximate $|{{\text{\normalsize\boldmath $\beta$}}}| \approx 1 \gg 1/(\gamma^2-1) \gg {a_{\mu}^{\text{SM}}}$ to get $$|{{\text{\normalsize\boldmath $\omega$}}}_a|^2 \approx |{{\text{\normalsize\boldmath $B$}}}|^2 \left[ \left( \frac{e}{m_{\mu}} \, {a_{\mu}^{\text{NP}}}\right)^2 + \left( \frac{2c}{\hbar} \, {d_{\mu}^{\text{NP}}}\right)^2 \right]\ .$$ We see that the measurement of ${{\text{\normalsize\boldmath $\omega$}}}_a$ again constrains only a combination (albeit a different one — cf. [Eq. (\[both\])]{}) of ${a_{\mu}^{\text{NP}}}$ and ${d_{\mu}^{\text{NP}}}$. This time, the constraint contours are ellipses centered on the origin in Fig. \[fig:amu\_dmu\]. Only by combining both measurements can the muon EDM and MDM be determined unambiguously. Of course, the up-down asymmetry is CP-violating, and so provides unambiguous information about ${d_{\mu}^{\text{NP}}}$ without contamination from ${a_{\mu}^{\text{NP}}}$. The measurement of the up-down asymmetry is therefore extremely valuable. Implications of the result for the Muon’s EDM {#sec:theoretical} ============================================= The muon’s EDM and anomalous MDM are defined through (here and below we set $\hbar = c = 1$) $$\begin{aligned} \label{EDMoperator} {\cal L}_{\text{EDM}} &=& -\frac{i}{2} {d_{\mu}}\, \bar{\mu} \sigma^{mn} \gamma_5 \mu \, F_{mn} \\ {\cal L}_{\text{MDM}} &=& {a_{\mu}}\frac{e}{4m_\mu} \, \bar{\mu} \sigma^{mn} \mu \, F_{mn} \ ,\end{aligned}$$ where $\sigma^{mn} = \frac{i}{2} \left[ \gamma^m, \gamma^n \right]$ and $F$ is the electromagnetic field strength. These operators are closely related. Assuming that they have the same origin, it is useful to write the new physics contributions to their coefficients as $$\begin{aligned} {d_{\mu}^{\text{NP}}}&=& \frac{e}{2m_\mu}\ {{\cal I}m}A \label{imA}\ , \\ {a_{\mu}^{\text{NP}}}&=& {{\cal R}e}A \ , \label{reA}\end{aligned}$$ with $A \equiv |A|e^{i{\phi_{\text{CP}}}}$. This defines an experimentally measurable quantity ${\phi_{\text{CP}}}$ which quantifies the amount of CP violation in the new physics, independently of its energy scale. Upon eliminating $|A|$, we find $${d_{\mu}^{\text{NP}}}= 4.0 \times 10^{-22}~{e~\text{cm}}\ \frac{{a_{\mu}^{\text{NP}}}}{43 \times 10^{-10}} \ \tan{\phi_{\text{CP}}}\ . \label{phicp}$$ The measured discrepancy in $|{{\text{\normalsize\boldmath $\omega$}}}_a|$ then constrains ${\phi_{\text{CP}}}$ and ${d_{\mu}^{\text{NP}}}$. Eliminating ${a_{\mu}^{\text{NP}}}$ from Eqs. (\[both\]) and (\[phicp\]), we find $$\begin{aligned} &&\left| {d_{\mu}^{\text{NP}}}\right| = \frac{e}{2 m_{\mu}} \, {a_{\mu}^{\text{SM}}}\sin{\phi_{\text{CP}}}\Biggl[ \ - \ \cos{\phi_{\text{CP}}}\nonumber \\ &&+ \left( \cos^2{\phi_{\text{CP}}}+ {(2{a_{\mu}^{\text{SM}}}+\Delta a_\mu) \Delta a_\mu \over ({a_{\mu}^{\text{SM}}}) ^2} \right)^{1/2} \Biggr] \ , \end{aligned}$$ The preferred regions of the $({\phi_{\text{CP}}},{d_{\mu}^{\text{NP}}})$ plane are shown in Fig. \[fig:dmu\_phi\]. For ‘natural’ values of ${\phi_{\text{CP}}}\sim 1$, ${d_{\mu}^{\text{NP}}}$ is of order $10^{-22}~{e~\text{cm}}$. With the proposed ${d_{\mu}^{\text{NP}}}$ sensitivity of [Eq. (\[proposedEDM\])]{}, all of the 2$\sigma$ allowed region with ${\phi_{\text{CP}}}> 10^{-2} \ {\rm rad}$ yields an observable signal. ![Regions of the $({\phi_{\text{CP}}}, {d_{\mu}^{\text{NP}}})$ plane allowed by the measured central value of $|{{\text{\normalsize\boldmath $\omega$}}}_a|$ (solid) and its 1$\sigma$ and 2$\sigma$ preferred values (shaded). The horizontal dot-dashed line marks the proposed experimental sensitivity to ${d_{\mu}^{\text{NP}}}$. The red horizontal solid lines denote the current 1$\sigma$ and 2$\sigma$ bounds on ${d_{\mu}^{\text{NP}}}$ [@Bailey:1979mn].[]{data-label="fig:dmu_phi"}](dmu_phi2.eps){height="2.3in"} At the same time, while this model-independent analysis indicates that natural values of ${\phi_{\text{CP}}}$ prefer ${d_{\mu}^{\text{NP}}}$ well within reach of the proposed muon EDM experiment, very large values of ${d_{\mu}^{\text{NP}}}$ also require highly fine-tuned ${\phi_{\text{CP}}}$. For example, we see from Fig. \[fig:dmu\_phi\] that values of ${d_{\mu}^{\text{NP}}}{ \mathop{}_{\textstyle \sim}^{\textstyle >} }10^{-20}\ {e~\text{cm}}$ are possible only if $|\pi/2 - {\phi_{\text{CP}}}| \sim 10^{-3}$. This is a consequence of the fact that EDMs are CP-odd and ${d_{\mu}^{\text{SM}}}\approx 0$, and so ${d_{\mu}^{\text{NP}}}$ appears only quadratically in $|{{\text{\normalsize\boldmath $\omega$}}}_a|$. Without a strong motivation for ${\phi_{\text{CP}}}\approx \pi/2$, it is therefore natural to expect the EDM contribution to $|{{\text{\normalsize\boldmath $\omega$}}}_a|$ to be negligible. Theoretical Expectations for in Supersymmetry {#sec:susy} ============================================= Our discussion up to now has been completely model-independent. In specific models, however, it may be difficult to achieve values of ${d_{\mu}}$ large enough to saturate the bound of [Eq. (\[newdmubound\])]{}. For example, in supersymmetry, assuming flavor conservation and taking extreme values of superparticle masses ($\sim 100~{\text{GeV}}$) and ${\tan\beta}$ (${\tan\beta}\sim 50$) to maximize the effect, the largest possible value of ${a_{\mu}}$ is $a_{\mu}^{\text{max}} \sim 10^{-7}$ [@Feng:2001tr]. Very roughly, one therefore expects a maximal $d_{\mu}$ of order $(e \hbar / 2 m_{\mu} c) a_{\mu}^{\text{max}} \sim 10^{-20}~{e~\text{cm}}$ in supersymmetry. With additional model assumptions, however, it is possible to further narrow down the expected range of ${d_{\mu}^{\text{NP}}}$ in supersymmetry. The EDM operator of [Eq. (\[EDMoperator\])]{} couples left- and right-handed muons, and so requires a mass insertion to flip the chirality. The natural choice for this mass is the lepton mass. On dimensional grounds, one therefore expects $$\label{massscaling} {d_{\mu}^{\text{NP}}}\propto \frac{m_{\mu}}{\tilde{m}^2}\ ,$$ where $\tilde{m}$ is the mass scale of the new physics. If the new physics is flavor blind, $d_f \propto m_f$ for all fermions $f$, which we refer to as ‘naive scaling.’ In particular, $$\label{naive} d_\mu \approx {m_\mu\over m_e}\, d_e \ .$$ The current bound on the electron EDM is $d_e = 1.8\, (1.2)\, (1.0) \times 10^{-27}~{e~\text{cm}}$ [@Commins:1994gv]. Combining the statistical and systematic errors in quadrature, this bound and [Eq. (\[naive\])]{} imply $$d_\mu \alt 9.1\times 10^{-25}~{e~\text{cm}}\ , \label{muedmlimit}$$ at the 90% CL, barely below the sensitivity of [Eq. (\[proposedEDM\])]{}. Naive scaling must be violated if a non-vanishing ${d_{\mu}}$ is to be observable at the proposed experiment. On the other hand, the proximity of the limit of [Eq. (\[muedmlimit\])]{} to the projected experimental sensitivity of [Eq. (\[proposedEDM\])]{} implies that even relatively small departures from naive scaling may yield an observable signal. Is naive scaling violation well-motivated, and can the violation be large enough to produce an observable EDM for the muon? To investigate these questions quantitatively, we consider supersymmetry [@recentwork]. (For violations of naive scaling in other models, see, for example, Ref. [@Babu:2000cz].) Many additional mass parameters are introduced in supersymmetric extensions of the Standard Model. These are in general complex and are new sources of CP violation, leading to a separate, major challenge for SUSY model building along with flavor violation. For a recent discussion of the supersymmetric CP problem in various supersymmetry breaking schemes, see Ref. [@Dine:2001ne]. In the minimal supersymmetric model, naive scaling requires $\bullet$ Degeneracy: Generation-independent slepton masses. $\bullet$ Proportionality: The $A$ terms must scale with the corresponding fermion mass. $\bullet$ Flavor conservation: Vanishing off-diagonal elements for the sfermion masses and the $A$-terms. We now briefly discuss violations of each of these properties in turn. Scalar degeneracy is the most obvious way to reduce flavor changing effects to allowable levels. Therefore many schemes for mediating supersymmetry breaking try to achieve degeneracy. However, in many of these, with the exception of simple gauge mediation models, there may be non-negligible contributions to scalar masses that are generation-dependent. For example, scalar non-degeneracy is typical in alignment models [@Nir:1993mx] or models with anomalous U(1) contributions to the sfermion masses where the sfermion hierarchy is often inverted relative to the fermion mass hierarchy [@models]. We now consider a simple model-independent parameterization to explore the impact of non-degenerate selectron and smuon masses. We set $m_{\tilde{e}_R} = m_{\tilde{e}_L} = {m_{\tilde{e}}}$ and $m_{\tilde{\mu}_R} = m_{\tilde{\mu}_L} = {m_{\tilde{\mu}}}$ and assume vanishing $A$ parameters. For fixed values of $M_1$, $M_2$, $|\mu|$, and large ${\tan\beta}$, then, to a good approximation both ${d_e}$ and ${d_{\mu}}$ are proportional to $\sin {\phi_{\text{CP}}}\, {\tan\beta}$, and we assume that $\sin {\phi_{\text{CP}}}\, {\tan\beta}$ saturates the ${d_e}$ bound. ![Contours of ${d_{\mu}}\times 10^{24}$ in ${e~\text{cm}}$ for varying $m_{\tilde{e}_R} = m_{\tilde{e}_L} = {m_{\tilde{e}}}$ and $m_{\tilde{\mu}_R} = m_{\tilde{\mu}_L} = {m_{\tilde{\mu}}}$ for vanishing $A$ terms, fixed $|\mu| = 500~{\text{GeV}}$ and $M_2 = 300~{\text{GeV}}$, and $M_1 = (g_1^2/g_2^2) M_2$ determined from gaugino mass unification. The CP-violating phase is assumed to saturate the bound ${d_e}< 4.4 \times 10^{-27}~{e~\text{cm}}$. The shaded regions are preferred by ${a_{\mu}}$ at $1\sigma$ and $2\sigma$ for ${\tan\beta}=50$. []{data-label="fig:msel_msmu"}](msel_msmu.eps){height="2.3in"} Contours of ${d_{\mu}}$ are given in Fig. \[fig:msel\_msmu\]. Observable values of ${d_{\mu}}$ are possible even for small violations of non-degeneracy; for example, for ${m_{\tilde{\mu}}}/{m_{\tilde{e}}}\alt 0.9$, muon EDMs greater than $10^{-24}~{e~\text{cm}}$ are possible. The current value of ${a_{\mu}}$ also favors light smuons and large EDMs. The smuon mass regions preferred by the current ${a_{\mu}}$ anomaly are given in Fig. \[fig:msel\_msmu\] for ${\tan\beta}= 50$. Within the $1\sigma$ preferred region, ${d_{\mu}}$ may be as large as $4\ (10) \times 10^{-24}~{e~\text{cm}}$ for ${m_{\tilde{e}}}< 1\ (2)~{\text{TeV}}$. Our assumed value of ${\tan\beta}$ is conservative; for smaller ${\tan\beta}$, the preferred smuon masses are lower and the possible ${d_{\mu}}$ values larger. Naive scaling is also broken if the $A$-terms are not proportional to the corresponding Yukawa couplings. Just as in the case of non-degeneracy, deviations from proportionality are found in many models. Although for large ${\tan\beta}$, the $A$ term contribution to the EDM is suppressed relative to the typically dominant chargino contribution, there are many possibilities that may yield large effects. In Ref. [@Ibrahim:2001jz], for example, it was noted that $A_e$ may be such that the chargino and neutralino contributions to $d_e$ cancel, while, since $A_e\neq A_\mu$, there is no cancellation in $d_\mu$, and observable values are possible. Finally, most models of high-scale supersymmetry breaking [@Dine:2001ne] typically contain flavor violation as well. In particular, smuon-stau mixing leads to a potentially significant enhancement in ${d_{\mu}}$, because it breaks naive scaling by introducing contributions enhanced by ${m_\tau\over m_\mu}$. In order to evaluate the significance of this enhancement, we must first determine how large the flavor violation may be. Taking into account the current $\tau \to \mu \gamma$ constraint, we found that values of ${d_{\mu}^{\text{NP}}}$ as large as $10^{-22}{e~\text{cm}}$ are possible [@Feng:2001sq]. In conclusion, the proposal to measure the muon EDM at the level of $10^{-24}~{e~\text{cm}}$ potentially improves existing sensitivities by five orders of magnitude. While the existing deviation in $g_{\mu}-2$ may be interpreted as evidence for new physics in either the muon’s MDM or EDM, the proposed experiment will definitively resolve this ambiguity, and may also uncover new physics in a wide variety of superysmmetric extensions of the Standard Model. The work of J.L.F. is supported in part by the U. S. Department of Energy under cooperative research agreement DF–FC02–94ER40818. K.T.M. thanks the Fermilab Theory Group, the ANL Theory Group and the Aspen Center for Physics for hospitality during the completion of this work. [99]{} A. D. Sakharov, Pisma Zh. Eksp. Teor. Fiz.  [**5**]{}, 32 (1967) \[JETP Lett.  [**5**]{}, 24 (1967)\]. A. G. Cohen, A. De Rujula and S. L. Glashow, Astrophys. J.  [**495**]{}, 539 (1998) \[astro-ph/9707087\]. G. R. Farrar and M. E. Shaposhnikov, Phys. Rev. D [**50**]{}, 774 (1994) \[hep-ph/9305275\]; M. B. Gavela, P. Hernandez, J. Orloff and O. Pene, Mod. Phys. Lett. A [**9**]{}, 795 (1994) \[hep-ph/9312215\]; M. B. Gavela, P. Hernandez, J. Orloff, O. Pene and C. Quimbay, Nucl. Phys. B [**430**]{}, 382 (1994) \[hep-ph/9406289\]; P. Huet and E. Sather, Phys. Rev. D [**51**]{}, 379 (1995) \[hep-ph/9404302\]. S. M. Barr and W. J. Marciano, BNL-41939, in “CP Violation”, ed. C. Jarlskog, World Scientific, Singapore (1989); W. Bernreuther and M. Suzuki, Rev. Mod. Phys.  [**63**]{}, 313 (1991) \[Erratum-ibid.  [**64**]{}, 633 (1991)\]. F. Hoogeveen, Nucl. Phys. B [**341**]{}, 322 (1990); I. B. Khriplovich, Phys. Lett. B [**173**]{}, 193 (1986) \[Sov. J. Nucl. Phys.  [**44**]{}, 659.1986 YAFIA,44,1019 (1986)\]. See, [[*e.g.*]{}]{}, [*Proceedings of the Joint U.S./Japan Workshop on New Initiatives in Muon Lepton Flavor Violation and Neutrino Oscillation with High Intense Muon and Neutrino Sources*]{}, Honolulu, Hawaii, 2–6 Oct 2000, eds. Y. Kuno and W. R. Molzon (World Scientific). Y. K. Semertzidis [*et al.*]{}, hep-ph/0012087; see also [http://www.bnl.gov/edm]{}. J. Bailey [*et al.*]{} \[CERN-Mainz-Daresbury Collaboration\], Nucl. Phys.  [**B150**]{}, 1 (1979). J. Aysto [*et al.*]{}, hep-ph/0109217. H. N. Brown [*et al.*]{} \[Muon g-2 Collaboration\], Phys. Rev. Lett.  [**86**]{}, 2227 (2001) \[hep-ex/0102017\]. M. Davier and A. Hocker, Phys. Lett. B [**435**]{}, 427 (1998) \[hep-ph/9805470\]. For recent reviews of standard model contributions to ${a_{\mu}}$, see, [[*e.g.*]{}]{}, W. J. Marciano and B. L. Roberts, hep-ph/0105056; S. Narison, Phys. Lett. B [**513**]{}, 53 (2001) \[hep-ph/0103199\]; K. Melnikov, hep-ph/0105267. J. L. Feng, K. T. Matchev and Y. Shadmi, Nucl. Phys. B [**613**]{}, 366 (2001) \[hep-ph/0107182\]. B. L. Roberts, private communication. J. L. Feng and K. T. Matchev, Phys. Rev. Lett.  [**86**]{}, 3480 (2001) \[hep-ph/0102146\]. E. D. Commins, S. B. Ross, D. DeMille and B. C. Regan, Phys. Rev. A [**50**]{}, 2960 (1994). For other recent work, see K. S. Babu, B. Dutta and R. N. Mohapatra, Phys. Rev. Lett.  [**85**]{}, 5064 (2000) \[hep-ph/0006329\]. T. Ibrahim, U. Chattopadhyay and P. Nath, Phys. Rev. D [**64**]{}, 016010 (2001) \[hep-ph/0102324\]. A. Bartl, T. Gajdosik, E. Lunghi, A. Masiero, W. Porod, H. Stremnitzer and O. Vives, hep-ph/0103324; M. Graesser and S. Thomas, hep-ph/0104254; Z. Chacko and G. D. Kribs, Phys. Rev. D [**64**]{}, 075015 (2001) \[hep-ph/0104317\]; T. Blazek and S. F. King, Phys. Lett. B [**518**]{}, 109 (2001) \[arXiv:hep-ph/0105005\]. R. Arnowitt, B. Dutta and Y. Santoso, hep-ph/0106089; A. Romanino and A. Strumia, hep-ph/0108275. K. S. Babu, S. M. Barr and I. Dorsner, hep-ph/0012303. M. Dine, E. Kramer, Y. Nir and Y. Shadmi, Phys. Rev. D [**63**]{}, 116005 (2001) \[hep-ph/0101092\]. Y. Nir and N. Seiberg, Phys. Lett. B [**309**]{}, 337 (1993) \[hep-ph/9304307\]. E. Dudas, S. Pokorski and C. A. Savoy, Phys. Lett. B [**369**]{}, 255 (1996) \[hep-ph/9509410\]; E. Dudas, C. Grojean, S. Pokorski and C. A. Savoy, Nucl. Phys. B [**481**]{}, 85 (1996) \[hep-ph/9606383\]; P. Brax and C. A. Savoy, JHEP [**0007**]{}, 048 (2000) \[hep-ph/0004133\]. T. Ibrahim and P. Nath, hep-ph/0105025.

--- abstract: 'The long-standing Alekseevskii conjecture states that a connected homogeneous Einstein space $G/K$ of negative scalar curvature must be diffeomorphic to $\RR^n$. This was known to be true only in dimensions up to $5$, and in dimension $6$ for non-semisimple $G$. In this work we prove that this is also the case in dimensions up to $10$ when $G$ is not semisimple. For arbitrary $G$, besides $5$ possible exceptions, we show that the conjecture holds up to dimension $8$.' address: - 'FaMAF $\&$ CIEM, Universidad Nacional de Córdoba, Córdoba, Argentina' - 'Mathematisches Institut, Universität Münster, Einsteinstr.  62, 48149 Münster, Germany' author: - 'Romina M. Arroyo' - 'Ramiro A. Lafuente' bibliography: - 'aleklow.bib' title: The Alekseevskii conjecture in low dimensions --- Introduction ============ A Riemannian manifold $(M^n, g)$ is called Einstein if its Ricci tensor satisfies $\Ricci(g) = c\, g$, for some $c\in \RR$. This is a very subtle condition, since it is too strong to allow general existence results, and at the same time too weak for obtaining obstructions in dimensions above $4$. It is therefore natural to consider the Einstein equation for a special class of metrics such as Kähler, Sasakian, with special holonomy, or with some symmetry assumption, among others (see [@LeBWang; @cruzchica; @Sparkssurvey; @Wang2012] for further details and examples). We study this equation on homogeneous manifolds. The classification of homogeneous Einstein spaces is naturally divided into cases according to the sign of the scalar curvature. Ricci-flat homogeneous manifolds are flat by [@AlkKml]. If the scalar curvature is positive, the manifold must be compact by Bonet-Myers’ theorem, while a theorem of Bochner [@Bochner1948] implies that if it is negative, the manifold is non-compact. In the latter case, the following fundamental problem remains unsolved [@Bss 7.57] Any connected homogeneous Einstein space of negative scalar curvature is diffeomorphic to a Euclidean space. The purpose of the present article is to investigate this conjecture in low-dimensional spaces. Recall that in dimensions $2$ and $3$, Einstein metrics have constant sectional curvature. Simply-connected homogeneous Einstein $4$-manifolds were classified by G. Jensen in his thesis [@Jns], and they are all isometric to symmetric spaces. In dimension $5$, non-compact homogeneous Einstein spaces $G/K$ were studied in [@Nkn1], where it was shown that if $G\neq \Sl_2(\CC)$ then they are isometric to simply-connected Einstein solvmanifolds, and in particular diffeomorphic to $\RR^5$. In the recent work [@semialglow] the authors proved that the conjecture holds in dimension $6$, provided there exists a non-semisimple transitive group of isometries (a shorter proof of this fact was recently obtained in [@JblPtr]). Our first main result is the following \[main6\] Let $(M^6,g)$ be a $6$-dimensional connected homogeneous Einstein space of negative scalar curvature, on which neither $\Sl_2(\CC)$ nor $\widetilde{\Sl_2(\RR)}\times \widetilde{\Sl_2(\RR)}$ acts transitively by isometries. Then, $M^6$ is diffeomorphic to $\RR^6$. Remarkably, the question of whether the $6$-dimensional simple Lie groups $\Sl_2(\CC)$ and $\widetilde{\Sl_2(\RR)}\times \widetilde{\Sl_2(\RR)}$ admit a left-invariant Einstein metric is still open. This is however not surprising if one recalls that the total number of homogeneous Einstein metrics on its compact counterpart $S^3 \times S^3$ is still unknown, even though the compact case has been much more investigated in the literature. Our second main result confirms the conjecture in dimension $7$. \[main7\] Any $7$-dimensional connected homogeneous Einstein space of negative scalar curvature is diffeomorphic to $\RR^7$. Besides the case of left-invariant metrics on two simple Lie groups and one very special homogeneous space, we show that the conjecture also holds in dimension $8$. \[main8\] Let $(M^8,g)$ be an $8$-dimensional connected homogeneous Einstein space of negative scalar curvature which is de Rham irreducible. Assume that $(M^8,g)$ is not an invariant metric on the simply connected homogeneous space $\left(\Sl_2(\RR)\times \Sl_2(\CC)\right)/\Delta\U(1)$, and that neither $\widetilde{\Sl_3(\RR)}$ nor $\widetilde{\SU(2,1)}$ acts transitively by isometries. Then, $M^8$ is diffeomorphic to $\RR^8$. It is important to remark that in Theorems \[main6\], \[main7\] and \[main8\] we actually obtain a stronger conclusion, namely that the spaces admit a simply-transitive solvable group of isometries (i.e.  they are isometric to a *solvmanifold*). We mention here that there is a stronger version of the conjecture, which is obtained by replacing the conclusion “diffeomorphic to a Euclidean space” by “isometric to a simply-connected solvmanifold” (this is commonly referred to as the *strong Alekseevskii conjecture* in the literature, see [@JblPtr]). Both statements turn out to be equivalent when the isometry group is linear, and in fact at the present time all known-examples of homogeneous Einstein spaces with negative scalar curvature are isometric to simply-connected solvmanifolds. Finally, we focus on the case where the presentation group is not semisimple. Our main result in this direction is the following \[mainnonuni\] Let $(M^n,g)$ be a simply-connected non-compact homogeneous Einstein space of dimension less than or equal to $10$, which is de Rham irreducible. If $(M^n,g)$ admits a non-semisimple transitive group of isometries, then $M^n$ is diffeomorphic to $\RR^n$. Using a close link relating non-compact homogeneous Einstein spaces and expanding homogeneous Ricci solitons (cf. [@HePtrWyl; @alek] and [@Jbl13b]), Theorem \[mainnonuni\] immediately implies the following result. Let $(M^n,g)$ be a simply-connected expanding homogeneous Ricci soliton which is not Einstein, of dimension less than or equal to $9$, and which is de Rham irreducible. Then, $M^n$ is diffeomorphic to $\RR^n$. With regards to other previous known results on low-dimensional homogeneous Einstein spaces, we mention that the classification of simply-connected compact homogeneous Einstein manifolds was obtained in [@AleDotFer] in dimension $5$, and in [@Nkn04] in dimension $7$. Partial results in dimension $6$ may be found in [@NknRod03]. In [@BhmKrr] it was proved that all simply-connected compact homogeneous spaces of dimension less than $12$ admit a homogeneous Einstein metric. In the non-compact case, the classification of Einstein solvmanifolds in low dimensions was studied in [@finding; @Wll03; @NikitenkoNikonorov; @FC13]. The starting point for the proof of our main results are the structural results for non-compact homogeneous Einstein spaces given in [@alek], and specially its more recent refinements proved in [@JblPtr]. Roughly speaking, these results state that the simply-connected cover of such a space admits a very special presentation of the form $G/K$, where $G = \left( G_1 A\right)\ltimes N$ is a semi-direct product of a nilpotent normal Lie subgroup $N$ and a reductive Lie subgroup $U = G_1 A$, with center $A$ and whose semisimple part $G_1 = [U,U]$ has no compact simple factors and contains the isotropy $K$. Moreover, the orbits of $U$ and $N$ are orthogonal at $eK$, the induced metric on $N$ is a homogeneous Ricci soliton, and the induced metric on $U/K$ satisfies an Einstein-like condition in which the action of $U$ on $N$ comes into play (see below). In the present article we further improve those results by showing that the orbits of $A$ and $G_1$ are also orthogonal at $eK$ (Theorem \[thm\_lemadimn\]). This allows us to reduce the problem to solving the generalized Einstein equation on $G_1/K$, which turns out to be a homogeneous space of dimension at most $7$ with semisimple transitive group. Moreover, as an application of our new structure refinements we present a short proof of a result of Jablonski [@Jbl13b] which states that homogeneous Ricci solitons are algebraic (Corollary  \[cor\_algebraic\]). The reduction to the simply-connected case is possible in dimensions $8$ and lower because we show that those spaces are isometric to solvmanifolds, thus allowing us to apply the results in [@Jab15]. In order to study the Einstein equation (and its generalized version) in the semisimple case, we give in Table \[tabla\] a complete classification of non-compact homogeneous spaces with a semisimple transitive group without compact simple factors, in dimensions up to $8$. The classification is based on that of the compact case, mainly given in [@BhmKrr], and a duality procedure [@Nkn1]. It includes some infinite families, such as the non-compact analogs of the Aloff-Wallach spaces, and some other examples in dimension $8$. To solve the Einstein equation for homogeneous metrics on these spaces we proceed case by case, studying the isotropy representations, and in many cases the results from [@Nkn2] can be applied to conclude that there is no solution. However, in some cases –mostly in higher dimensions– this is not enough, and a very detailed analysis of the Ricci curvature is carried out. As a by-product of this analysis, a general non-existence result for some cases where $\Sl_2(\RR)$ is one of the simple factors of the transitive group is given in Proposition \[Propsl2RxG1\]. One of the reasons why we are not able to extend Theorem \[mainnonuni\] to dimensions $11$ and higher is that already in dimension $11$, examples such as $( \Sl_2(\CC) \cdot \RR )\ltimes \RR^4$ appear, with $N = \RR^4$ and $\Sl_2(\CC)$ acting non-trivially on it. The homogeneous Einstein equation for such a space reduces to an equation for left-invariant metrics on $\Sl_2(\CC)$ which is even more general than the Einstein equation. The article is organized as follows. In Section \[prelimstruct\] we state the structure theorems for non-compact homogeneous Einstein spaces, since they will be repeatedly used along the paper, and prove the new refinements metioned above. In Section \[semisimple\] we prove Theorem \[thmsemisimple\], which deals with the semisimple case, and in order to do that we give the classification of non-compact semisimple homogeneous spaces up to dimension $8$. This, together with previously known results, already implies Theorem \[main6\]. In Section \[sectionnonuni\] we prove Theorem \[mainnonuni\], and then in Section \[strong\] we focus on the strong Alekseevskii conjecture and complete the proofs of Theorems \[main7\] and \[main8\]. [*Acknowledgements.*]{} It is our pleasure to thank Jorge Lauret for fruitful discussions, and Christoph Böhm for providing useful comments on a draft version of this article. Part of this research was carried out while the first author was a visitor at McMaster University. She is very grateful to the Department of Mathematics, the Geometry and Topology group and especially to McKenzie Wang for his kindness and hospitality. Structure of non-compact homogeneous Einstein spaces {#prelimstruct} ==================================================== In this section we review the most important known facts about the algebraic structure of non-compact homogeneous Einstein spaces, since they will be crucial in the proof of our main results. Here and throughout the rest of the article, all manifolds under consideration are connected and all homogeneous spaces are almost-effective, unless otherwise stated. \[structure\] Let $(M,g)$ be a simply-connected homogeneous Einstein space with negative scalar curvature. Then, there exists a transitive Lie group of isometries $G$ whose isotropy at some point $p\in M$ is $K$, with the following properties: - $G = \left(G_1 A\right) \ltimes N$, where $N$ is a nilpotent normal Lie subgroup, $U = G_1 A$ is a reductive Lie group with center $A = Z(U)$, and $G_1 = [U,U]$ is semisimple without any compact simple factors and contains the isotropy $K$. - The orbits of $U$ and $N$ are orthogonal at $p$. - The induced left-invariant metric $g_N$ on $N$ is a Ricci soliton (i.e.  $(N, g_N)$ is a *nilsoliton*). - The Ricci curvature of the induced $U$-invariant metric $g_{U/K}$ on $U/K$ is given by $$\label{eqRicU/K} \ricci_{U/K}(Y,Y) = c \cdot g_{U/K}(Y,Y) + \tr \left(S\left(\theta(Y)\right)^2\right),$$ for some $c<0$, where $\theta : \ug \to \Der(\ngo)$ is the corresponding infinitesimal action ($\ug = \Lie(U), \ngo = \Lie(N)$), $S(A) = \unm\left(A + A^t\right)$, and the transpose is taken relative to the nilsoliton inner product on $\ngo$. - The infinitesimal action $\theta$ and the nilsoliton metric satisfy the following compatibility condition: $$\label{eqmmtheta} \sum_i [\theta(Y_i), \theta(Y_i)^t] = 0,$$ where the sum is taken over an orthonormal basis for $\ug$[^1]. Moreover, $\theta(Y) = \theta(Y)^t$ for every $Y\in \zg(\ug)$. Conversely, if a simply-connected homogeneous manifold admits a transitive group of isometries $G$ satisfying $(i)-(v)$, then it is Einstein, with negative scalar curvature. \[remarks\] (a) \[remarkinfinitesimal\] Conditions (i) and (ii) may also be interpreted at the infinitesimal level, as follows: Let $\ggo, \ug, \ngo, \kg$ be the Lie algebras of the groups $G, U, N, K$, respectively. We have that $\ggo = \ug \ltimes_\theta \ngo$, with $\ug$ a reductive subalgebra and $\ngo$ the nilradical of $\ggo$ (the maximal nilpotent ideal). Consider the reductive decomposition $\ggo = \kg \oplus \pg$ for $G/K$, where $\pg$ is the orthogonal complement of $\kg$ relative to the Killing form of $\ggo$. This induces a reductive decomposition $\ug = \kg \oplus \hg$ for the homogeneous space $U/K$, by letting $\hg := \pg \cap \ug$. The $G$-invariant metric $g$ on $G/K$ is thus identified with an $\Ad(K)$-invariant inner product $\ip$ on  $\pg$, and one has that $$\langle \hg, \ngo \rangle = 0.$$ For technical reasons, it is sometimes convenient to extend this inner product to an inner product on $\ggo$, which we will also denote $\ip$, by letting $\kg \perp\pg$ and choosing on $\kg$ some $\Ad(K)$-invariant inner product. By doing so, we clearly obtain $\langle \ug, \ngo \rangle = 0$. In fact, in condition (v), by an orthonormal basis of $\ug$ we mean that it is orthonormal with respect to the inner product extended as explained above. (b) \[remarktheta\] $\theta : \ug \to \Der(\ngo)$ is nothing but the adjoint representation of $\ggo$ co-restricted to act on the nilradical, that is, $$\theta(Y) X = [Y,X] \in \ngo, \qquad X\in \ngo, \,\, Y\in \ug.$$ It was noticed by J. Lauret that condition is equivalent to $\theta$ being a zero of the moment map associated with the natural $\Gl(\ngo)$-action on the vector space $\End(\ug,\End(\ngo))$ (see [@semialglow Appendix] and [@JblPtr $\S 2.1$] for more details on this fact). (c) The Einstein constant of $(G/K,g)$ and the cosmological constant of the nilsoliton $(N,g_N)$ both coincide with the scalar $c<0$ in condition (iv). (d) According to the construction procedure for expanding algebraic solitons described in [@alek §5], it is easy to see that given any non-compact Einstein homogeneous space $G/K$, we can always build another one with the same $U/K$ but with abelian nilradical. (e) The simply-connected hypothesis is not necessary for obtaining the results at the infinitesimal level. However, it turns out to be necessary for the converse assertion to hold. In particular, one question which still remains unanswered is whether any homogeneous Einstein space with negative scalar curvature is simply connected. This is known to be true when the universal cover is a solvmanifold, by the results in [@AC99; @Jab15]. (f) \[rmkDotti\] If $G$ is a unimodular Lie group, then by [@Dtt88 Theorem 2] it must in fact be semisimple, and hence it equals $G_1$. In this case, the only information that Theorem \[structure\] provides is that $G_1$ has no compact simple factors. (g) On the other hand, if $(M,g)$ admits a non-semisimple transitive group of isometries, it follows from [@alek; @JblPtr] that the group $G$ in Theorem \[structure\] may be chosen to be non-unimodular. In this case, the so called “mean curvature vector” $H$, implicitly defined by $$\langle H, X \rangle = \tr \left( \ad X\right), \qquad \forall \, X\in\ggo,$$ is non-zero. By using that under the hypothesis of Theorem \[structure\], $G/K$ is diffeomorphic to the product manifold $G_1/K \times AN$, with $S = AN$ a simply-connected solvable Lie group, one obtains the following \[reductionG1/K\] Let $(M,g)$ be a simply-connected homogeneous Einstein space with negative scalar curvature, and let $G/K$ be the presentation given in Theorem \[structure\]. Then, $M$ is diffeomorphic to a Euclidean space if and only if $G_1/K$ is so. It is important to notice that Theorem \[structure\] does not state that the orbits of $Z(U)$ and $G_1$ are orthogonal at $p$. In other words, it is not known whether $\zg(\ug) \perp \ggo_1$ (where $\ggo_1= \Lie(G_1) = [\ug,\ug]$). This would be the most natural result to expect, since it would imply that there is a Levi decomposition $G = G_1 \ltimes S$ which is adapted to the geometry of $(M^n,g)$, in the sense that the orbits of $G_1$ and $S$ at $p$ are orthogonal. In what follows we prove that in fact one always has this nicer structure. \[thm\_lemadimn\] Let $(M,g)$ be a simply-connected homogeneous Einstein space of negative scalar curvature, and consider for it the presentation $G/K$ given in Theorem \[structure\]. Then, $\zg(\ug)$ is orthogonal to $\ggo_1$. \[rmkn1\] If furthermore one has that $\theta|_{\ggo_1} = 0$, then $G/K$ is isometric to a Riemannian product $G_1/K \times S$ of Einstein homogeneous spaces of negative scalar curvature. Notice that condition $\theta|_{\ggo_1} = 0$ is trivially satisfied when $\dim \ngo = 1$. Following the notation from Remark \[remarks\], , equation may be rewritten as an equation for endomorphisms of $\hg \simeq T_{eK} U/K$ as $$\label{eqn_RicUK} \Ricci_{U/K} = c \cdot I + C_\theta.$$ Here, $\Ricci_{U/K} \in \End(\hg)$ denotes the Ricci operator of the homogeneous space $(U/K, g_{U/K})$, and $C_\theta \in \End(\hg)$ is the symmetric endomorphism given by $$\langle C_\theta X, Y \rangle = \tr S(\theta(X))S(\theta(Y)), \qquad X,Y\in \hg.$$ Since $\theta$ is defined on $\ug$ and not just on $\hg$, we may of course extend $C_\theta$ to a symmetric endomorphism of $\ug$, where $C_\theta (\kg) = 0$ (recall that the action of the isotropy is by skew-symmetric operators). We have $\theta: \ug \to \End(\ngo)$, and by Theorem \[structure\] $S(\theta(\zg(\ug)))$ is a family of pairwise commuting, symmetric operators in $\End(\ngo)$, which commute also with all of $\theta(\ug)$ (recall that $\theta$ is a Lie algebra representation). We may thus consider an orthogonal decomposition of $\ngo$ into common eigenspaces for the family $S(\theta(\zg(\ug)))$ (i.e.  a weight-space decomposition): $$\label{rootdec} \ngo = \ngo_1 \oplus \ldots \oplus \ngo_l,$$ with $\alpha_1, \ldots, \alpha_l \in \zg(\ug)^*$ the corresponding weights. The restricted representation $\theta_{\ggo_1} = \theta|_{\ggo_1}: \ggo_1 \to \End(\ngo)$ must preserve this weight-space decomposition. For each $k = 1,\ldots,l$ we haveß a *co-restricted* representation of the semisimple Lie algbera $\ggo_1$, given by $$\theta_{\ggo_1}^k := \pi_k \circ \theta |_{\ggo_1} : \ggo_1 \to \End(\ngo_k),$$ where $\pi_k:\ngo \to \ngo_k$ is the orthogonal projection. Observe that, in particular, $\theta_{\ggo_1}^k(Y)$ is traceless for each $Y\in \ggo_1$ and $k=1,\ldots,l$. Now we claim that for $Y\in \ggo_1$, $X\in \zg(\ug)$ one has that $\la C_\theta X, Y\ra = 0$. Indeed, using the orthogonality of the decomposition , and the fact that it is preserved by $\theta(\ug)$, we obtain $$\begin{aligned} \left\langle C_\theta \, Y, X \right\rangle &= \sum_{k=1}^l \tr \left( S\left(\theta_{\ggo_1}^k (Y)\right) \left(\alpha_k(X) \cdot I\right)\right) = \sum_{k=1}^l \alpha_k(X) \, \tr \theta_{\ggo_1}^k(Y) = 0.\end{aligned}$$ Consider in $U/K$ the reductive decomposition $\ug = \kg \oplus \hg$. We may also assume that $\ggo_1 = \kg \oplus \hg_1$ is a reductive decomposition for $G_1/K$, where $\hg_1 \subseteq \hg$. Let $\qg$ be the orthogonal complement of $\hg_1$ in $\hg$, and let us show that $\qg = \zg(\ug)$. To that end, take $Y \in \qg$ and write it as $Y = Y_1 + Y_\zg$, where $Y_1 \in \hg_1$, $Y_\zg \in \zg(\ug)$. We now look at the Ricci curvature in the directions $Y_1$, $Y_\zg$. First, by  and the above claim we obtain $$\Ricci_{U/K}(Y_1, Y_\zg) = c \, \langle Y_1, Y_\zg\rangle = c\, \langle Y_1, Y-Y_1\rangle = - c\, \| Y_1\|^2 \geq 0,$$ since $c<0$. On the other hand, we use that $Y_\zg \in \zg(\ug)$, $Y\perp [\ug,\ug]$, and the explicit formula for the Ricci curvature in the unimodular case (see [@Bss 7.38]), to get $$\begin{aligned} \Ricci_{U/K}(Y_1, Y_\zg) =& -\unm\sum_{i,j}\langle [Y_1, X_i]_{\hg}, X_j \rangle \langle [Y_\zg, X_i]_{\hg}, X_j \rangle \\ & + \unc \sum_{i,j} \langle [X_i,X_j]_{\hg}, Y_1\rangle \langle [X_i,X_j]_{\hg}, Y_\zg\rangle - \unm \tr \ad_\ug Y_1 \ad_\ug Y_\zg \\ = &\, \unc \sum_{i,j} \langle [X_i,X_j]_{\hg}, Y_1\rangle \langle [X_i,X_j]_{\hg}, Y - Y_1\rangle \\ = & -\unc \sum_{i,j} \langle [X_i,X_j]_{\hg}, Y_1\rangle^2 \leq 0,\end{aligned}$$ where $\{ X_i\}$ is an orthonormal basis for $\hg.$ Hence, we must have equality, and $Y_1 = 0$. Therefore, $\qg = \zg(\ug)$, and the proof is finished. \[rmk\_centerorthogonal\] The previous theorem holds more generally for expanding homogeneous Ricci solitons. More precisely, if $(M^n, g)$ is an expanding (i.e. $c<0$) homogeneous Ricci soliton, and $G$ is the full isometry group, then by [@Jbl] the soliton is *semi-algebraic*. Therefore by [@alek] the homogeneous space $G/K$ satisfies all the nice properties stated in Theorem \[structure\], but possibly without the additional conditions proven in [@JblPtr] for Einstein spaces (namely, $G_1$ might have compact simple factors, and the action of $\zg(\ug)$ on $\ngo$ might not be by symmetric endomorphisms). Nevertheless, Lemma 3.5 from [@JblPtr] still assures that by the compatibility condition one has that the family $\theta(\zg(\ug)) \subset \End(\ngo)$ consists of normal operators, whose transposes commute with all of $\theta(\ug)$. Thus, one can also consider the decomposition as in the proof of Theorem  \[thm\_lemadimn\], and proceed in exactly the same way to conclude that $\zg(\ug) \perp \ggo_1$. As a quick application we get an alternative proof of the following result of Jablonski [@Jbl13b]. \[cor\_algebraic\] Homogeneous Ricci solitons are algebraic. As is well-known, the only non-trivial examples (that is, not locally isometric to the product of an Einstein homogeneous space and a flat factor $\RR^k$) occur in the expanding case (see the discussion in [@solvsolitons §2] and the references therein for more details). Let $(M,g)$ be an expanding homogeneous Ricci soliton. For the presentation $G/K$, where $G$ is the full isometry group, we have by Theorem \[thm\_lemadimn\] and Remark \[rmk\_centerorthogonal\] that $\zg(\ug) \perp \ggo_1$. Now recall that the mean curvature vector $H \in \hg \subset \ug$ is always orthogonal to $\ggo_1 = [\ug,\ug]$, since any representation of a semi-simple Lie algebra consists of traceless endomorphisms. Thus, $H \in \zg(\ug)$, and in particular $$S(\ad H|_\hg) = 0.$$ By applying Proposition 4.14 from [@alek] we conclude that the soliton is indeed algebraic. Another application of our new structural results is the reduction of the classification problem (in the non-unimodular case) to the so called “rank one” case (cf.  [@Heb Theorem D]). \[cor\_rankone\] Let $(M^n, g)$, $G/K$ be as in Theorem \[thm\_lemadimn\], with $G$ non-unimodular. Consider $U_0$, $G_0$ the connected Lie subgroups of $U$, $G$ with Lie algebras $\ug_0 := [\ug,\ug]\oplus \RR H \subset \ug$, $\ggo_0 := \ug_0 \oplus \ngo$, respectively. Then, there is a diffeomorphism $$M^n \simeq \RR^a \times G_0 / K, \qquad a = \dim Z(U) - 1,$$ and the induced $G_0$-invariant metric on $G_0/K$ is Einstein with the same Einstein constant $c<0$ as $g$. Recall the following formula for the Ricci curvature of a homogeneous space, whose proof follows immediately from the proof of Proposition 6.1 in [@alek]. Let $(G/K,g)$ be a Riemannian homogeneous space with reductive decomposition $\ggo = \kg \oplus \pg$, and assume there exists $X\in \pg$ such that $[H,X] = 0$, and the subspace $\tilde\ggo := \{X\}^\perp$ is a codimension-one ideal of $\ggo$ that contains $H$ and $\kg$. Let $\widetilde G$ be the connected Lie subgroup of $G$ with Lie algebra $\tilde\ggo$, and consider the induced metric on the orbit $\widetilde G \cdot (eK) \simeq \widetilde G / K$. Then, the corresponding Ricci operators satisfy $$\ricci_{G/K} |_{ \, \, \tilde \pg} = \ricci_{\widetilde G / K} + \unm \left[A,A^t\right],$$ where $\tilde \pg = \pg \cap \tilde\ggo$ and $A := \ad X|_{\tilde \pg} \in \End(\pg)$. Theorem \[thm\_lemadimn\] implies that $H\in \zg(\ug)$, and that any $X \in \zg(\ug)$ with $X \perp H$ satisfies the conditions of the above lemma. Moreover, the corresponding endomorphism $A$ is symmetric by Theorem  \[structure\],  (v), thus the term $\unm [A,A^t]$ in the formula vanishes. By applying the lemma to any such $X$ we obtain a codimension-one submanifold $\tilde G / K$ in $G/K$ which with the induced metric is Einstein, with the same Einstein constant as $G/K$. Since the spaces are simply-connected, as differentiable manifolds we have that $G/K \simeq \RR \times \tilde G/ K$. After applying this procedure $a$ times, where $a = \dim Z(U)-1$, the corollary follows. To conclude this section we prove the following simple but useful formula for the Ricci curvature of a homogeneous space, which is in some way a generalization of [@Mln Lemma 2.3]. \[lem\_formulaRicci\] Let $(U/K,g)$ be a Riemannian homogeneous space with $U$ a unimodular Lie group, and consider a reductive decomposition $\ug = \kg \oplus \mg$. If $X, Y\in \mg$ are such that $[\kg,X]= [\kg,Y] = 0$, then $$\Ricci(X,Y) = \unc \sum_{i,j} \langle [X_i,X_j]_\mg, X\rangle \langle [X_i,X_j]_\mg, Y\rangle - \tr S(\ad_\mg X) S(\ad_\mg Y),$$ where $\{ X_i\}$ is any orthonormal basis for $\mg$ (here, $\ad_\mg X \in \End(\mg)$ stands for the restriction of $\ad X$ to $\mg$, projected onto $\mg$). Moreover, if $Y$ is orthogonal to the commutator ideal $[\ug,\ug]$ (i.e.  to its projection onto $\mg$), then $$\Ricci(X,Y) = - \tr S(\ad_\mg X) S(\ad_\mg Y), \qquad \forall \, X\in \mg \mbox{ such that } [\kg,X]=0.$$ From the formula [@Bss 7.38] for the Ricci curvature of a homogeneous space, and using that $\ug$ is unimodular, we see that $$\begin{aligned} \Ricci(X,Y) =& -\unm \sum_{i,j} \langle [X,X_i]_\mg, X_j \rangle \langle [Y,X_i]_\mg, X_j \rangle \\ & + \unc \sum_{i,j} \langle [X_i,X_j]_\mg, X\rangle \langle [X_i,X_j]_\mg, Y\rangle - \unm B(X,Y) \\ =& -\unm \tr \left(\ad_\mg X\right) \left(\ad_\mg Y \right)^t - \unm \tr (\ad X)(\ad Y) \\ & + \unc \sum_{i,j} \langle [X_i,X_j]_\mg, X\rangle \langle [X_i,X_j]_\mg, Y\rangle,\end{aligned}$$ where $\{ X_i\}$ is an orthonormal basis for $\mg$. Notice that conditions $[\kg,X] = 0$ and $[\kg,Y] =~ 0$ imply that $\tr(\ad X)(\ad Y) = \tr(\ad_\mg X)(\ad_\mg Y)$. Then, the first formula follows. If moreover $Y\perp [\ug,\ug]_\mg$, then it is easy to see that $[\kg,Y]=0$, so the first formula applies, and the sum term in it disappears. Semisimple transitive group {#semisimple} =========================== The main purpose of this section is to prove the following \[thmsemisimple\] Let $G$ be a semisimple Lie group and consider a homogeneous Einstein space $\left(G/H,g\right)$ which is de Rham irreducible. Assume that $\dim G/H \leq 8$, $\dim H \geq 1$ and that $G/H \neq$ $(\Sl_2(\RR) \times \Sl_2(\CC))\slash \Delta \U(1)$. Then, $(G/H,g)$ is an irreducible symmetric space of the non-compact type. The proof will follow from a case-by-case analysis. We warn the reader that, in contrast with the rest of the article, throughout this section the group $G$ will always be a semisimple Lie group. \[defsshomogspace\] We call a homogeneous space $G/H$ *semisimple of the non-compact type* if $G$ is a semisimple Lie group without compact simple factors. In view of Theorem \[structure\], we are reduced to studying the cases where $G/H$ is semisimple of the non-compact type. Moreover, we may restrict ourselves to the simply-connected case. Indeed, the universal cover of an Einstein manifold is still Einstein, and it is a classical result that symmetric spaces of the non-compact type do not admit non-trivial homogeneous quotients  [@Car27]. Following [@Al11; @Nkn1], we use the duality between compact and non-compact symmetric spaces to obtain the classification of semisimple homogeneous spaces of the non-compact type from the classification of compact homogeneous spaces in low dimensions (\[BK\]), as follows: Given $G/H$ a semisimple homogeneous space of the non-compact type, let $\ggo = \ggo_1 \oplus \ldots \oplus \ggo_s$ be the decomposition of $\ggo$ into simple ideals –which are all of the non-compact type– let $\kg\subseteq \ggo$ be a maximal compactly embedded subalgebra such that $\hg \subseteq \kg$, and for each $i=1,\ldots,s$ let $\kg_i = \ggo_i \cap \kg$, which is a maximal compactly embedded subalgebra of $\ggo_i$. The pairs $(\ggo_i, \kg_i)$ are symmetric pairs of the non-compact type (at the Lie algebra level), and its corresponding dual symmetric pairs $(\hat\ggo_i, \hat\kg_i)$ are of the compact type. If $\hat \ggo := \hat \ggo_1 \oplus \ldots\oplus \hat\ggo_s$, $\hat \kg := \hat\kg_1 \oplus \ldots \oplus \hat\kg_s$, then $\hat \kg$ and $\kg$ are isomorphic Lie algebras, and via this isomorphism we can consider the subalgebra $\hat \hg \subseteq \hat \kg$ corresponding to $\hg \subseteq \kg$. The effective homogeneous space $\hat G/ \hat H$ associated with $\hat \ggo, \hat\hg$ is compact. Therefore, in order to obtain all possible spaces $G/H$ as above one can argue as follows: - Classify all compact homogeneous spaces in “canonical presentation” (in the sense of [@BhmKrr]). - For each compact homogeneous space $(\hat G/ \hat H)$ in the previous classification, consider all possible compact symmetric pairs $(\hat \ggo,\hat\kg)$ with $\hat \hg \subseteq \hat \kg$, where $\Lie(\hat G) = \hat\ggo$, $\Lie(\hat H) = \hat \hg$ (there may be none at all). - For each such pair, let $(\ggo,\kg)$ be its dual, obtained by dualizing each simple factor to its non-compact counterpart. The isomorphism $\kg \simeq \hat \kg$ defines a subalgebra $\hg\subseteq \kg$ isomorphic to $\hat \hg$, and from $\ggo, \hg$ one obtains a non-compact homogeneous space $G/H$ as desired. We note that if a non-compact $G/H$ is obtained from a compact $\hat G/ \hat H$, then the Lie groups $H$ and $\hat H$ are isomorphic, and moreover the isotropy representations are equivalent. To obtain the classification of all non-compact homogeneous spaces with semisimple transitive group (i.e.  taking into account that $G$ may have compact simple factors), the duality procedure works in the very same way. One only needs to dualize the symmetric pairs which are of the non-compact type. We give in Table \[tabla\] the classification of simply-connected, semisimple homogeneous spaces of the non-compact type (cf.  Definition \[defsshomogspace\]), together with its corresponding compact duals, the compact symmetric space used in each case for the dualization procedure, and the decomposition of the isotropy representation into irreducible summands. Notice also that, for notational purposes, some of the non-compact spaces in the table are not simply connected, but still they are to be read as their universal covers. Symmetric spaces are not included, since a list of all irreducible symmetric spaces can be found for instance in [@Bss p. 200]. We also do not include cases which are product of lower dimensional homogeneous spaces, unless the space admits non-product invariant metrics (see Proposition \[prodRiem\] below). Our notation follows that of [@Bss], with the only exception of $\SU(1,1)$, which we call $\Sl_2(\RR)$. Regarding the list of compact homogeneous spaces in canonical presentation, we refer the reader to [@BhmKrr]. All the embedings of the isotropy subgroup are clear once the corresponding compact symmetric space used for the dualization is taken into account. The precise meaning of the parameters corresponding to abelian subgroups in the isotropy may be found in [@Nkn04  §1]. The information on the isotropy representation is to be understood as follows: for a space $G/H$, consider $\hg \subseteq \kg \subseteq \ggo$ as above, where $\kg$ is a maximal compactly embedded subalgebra with corresponding connected subgroup $K$. Take the corresponding Cartan decomposition $\ggo = \kg \oplus \pg$ (cf. [@Helgason pp. 182]), and let $\qg$ be an $\Ad(K)$-invariant complement for $\hg$ in $\kg$. Setting $\mg = \qg \oplus \pg$, we obtain a reductive decomposition $\ggo = \hg \oplus \mg$ for the homogeneous space $G/H$. Whenever we write $\sum_i \qg_i^{(a_i)} \oplus \sum_j \pg_j^{(b_j)}$ we mean that $$\qg = \sum_i \qg_i^{(a_i)}, \qquad \pg = \sum_j \pg_j^{(b_j)}, \qquad \dim \qg_i^{(a_i)} = a_i, \quad \dim \pg_j^{(b_j)} = b_j,$$ and for $i,j \neq 0$, each summand $\qg_i^{(a_i)}$, $\pg_j^{(b_j)}$ is an irreducible $\Ad(H)$-module, where any two such modules are inequivalent unless otherwise stated. The $0$ sub-index stands for trivial modules (i.e.  $[\hg,\qg_0^{(a_0)}] = [\hg,\pg_0^{(b_0)}] = 0$). *Notes on Table \[tabla\]:* (a) \[Aloff-Wallach\] $p,q\in \ZZ$, $0\leq p \leq q $, $\operatorname{gcd}(p,q) = 1$. See [@Wng82]. (b) \[S2xS3\] $p,q\in \ZZ-\{0\}$, $p\leq q$, $\operatorname{gcd}(p,q) = 1$. (c) \[isotropyirred\] These compact spaces are isotropy irreducible but non-symmetric (see [@Bss pp.  203]). Clearly, they do not admit any non-compact counterpart. (d) \[a1a2a3\] $a_1,a_2,a_3 \in \ZZ-\{0\}$, $a_1\leq a_2 \leq a_3$, $\operatorname{gcd}(a_1,a_2,a_3) = 1$ (the order may be assumed up to equivariant diffeomorphism, by using outer automorphisms given by the Weyl group; the parameters are all nonzero since otherwise the space splits as a product, and this are considered as a separate case). The space $(\Sl_2(\RR) \times \Sl_2(\CC))\slash \Delta_{p,q} \U(1)$ is obtained only when $a_2 = a_3$. For convenience, we have renamed the parameters as $p=a_1$, $q = a_2 = a_3$. Recall that by [@Nkn2 Theorem 1], if a $G$-invariant metric makes the chosen Cartan decomposition orthogonal (i.e. it is such that $\langle \qg, \pg \rangle = 0$), then $(G/H,g)$ is not Einstein. In particular, if we consider a decomposition of $\mg$ into irreducible $\Ad(H)$-modules given by $\qg = \qg_1 \oplus \ldots \oplus \qg_u$, $\pg = \pg_1 \oplus \ldots \oplus \pg_v$ (recall that $\qg$ and $\pg$ are $\Ad(H)$-invariant), and none of the $\qg_i$ is equivalent to any of the $\pg_j$, then every $G$-invariant metric on $G/H$ satisfies $\langle \qg,\pg\rangle=0$, and thus none of them is Einstein [@Nkn2 Corollary]. It may be the case that a single Cartan decomposition is not orthogonal with respect to *every* $G$-invariant metric, and still every metric makes *some* Cartan decomposition orthogonal (recall that a Cartan decomposition is only unique up to the action of inner automorphisms). In [@Nkn2 Theorem  2], necessary and sufficient conditions are given for this to happen. The following result is well-known, but we include a proof of it for the sake of completeness. \[prodRiem\] Let $G_1\slash H_1$, $G_2\slash H_2$ be two homogeneous spaces such that the isotropy representation of $G_1/H_1$ acts non-trivially on every invariant subspace. Then, any $\left(G_1 \times G_2\right)$-invariant metric on $\left(G_1 \times G_2\right)\slash \left(H_1 \times H_2\right)$ is a Riemannian product of invariant metrics on each factor. If $\ggo_1=\hg_1\oplus\mg_1$ and $\ggo_2=\hg_2\oplus\mg_2$ are reductive decompositions of $G_1/H_1$ and $G_2/H_2$ respectively, then $\ggo_1\oplus\ggo_2=(\hg_1\oplus\hg_2)\oplus(\mg_1\oplus\mg_2)$ is a reductive decomposition of $G_1 \times G_2/H_1 \times H_2.$ Let $\pg_i\subseteq \mg_i$ be $\ad(\hg_i)$-irreducible subspaces, $i=1,2$. We know that $\ad(\hg_1)|_{\pg_1}$ is non-trivial. If there was an intertwining operator $T:\pg_1\rightarrow\pg_2$, i.e, $$T\circ\ad(Z)|_{\pg_1}=\ad(Z)|_{\pg_2}\circ T, \quad \mbox{for all } Z\in\hg_1\oplus\hg_2,$$ we could take $Z=(Z_1,0) \in \hg_1\oplus\hg_2$ and would have that $T\circ\ad(Z_1)|_{\pg_1}=0,$ for all $Z_1 \in \hg_1,$ so $\ad(Z_1)|_{\pg_1}=0$, for all $Z_1 \in \hg_1,$ which is a contradiction. We are now in a position to start the case-by-case analysis. $\dim G/H \leq 7$ ----------------- $ $ After having computed all the isotropy representations, we see that in most of the spaces of dimension up to $7$ in Table \[tabla\], the Cartan decomposition we have chosen is such that $\qg$ and $\pg$ share no equivalent modules, and thus these spaces admit no $G$-invariant Einstein metric. The exceptions are the following: $$\Sl_2(\CC)/\U(1), \quad \SO(4,1)/\SO(3), \quad \SU(2,1)/\Delta_{p,q}\U(1), \quad \Sl_3(\RR)/\SO(2).$$ Non-existence of homogeneous Einstein metrics on $\Sl_3(\RR)/\SO(2)$ was established in [@Nkn2  Example 4] by finding, for every $G$-invariant metric, a suitable Cartan decomposition which is orthogonal. By applying the same methods and a straightforward computation, it can be shown that the space $\SO(4,1)/\SO(3)$ also satisfies the hypotheses of [@Nkn2 Theorem 2], and hence it admits no homogeneous Einstein metric. ### $\Sl_2(\CC)/\U(1)$ {#sectionsl2C} Unfortunately, this space admits invariant metrics for which there is no orthogonal Cartan decomposition. Consider the following ordered basis $\mathcal{B}$ for $\slg_2(\CC)$ $$\begin{aligned} \label{matricessl2C} Z &= \twomatrix{i}{0}{0}{-i}, \quad Y_0 = \twomatrix{1}{0}{0}{-1}, \quad Y_1 = \twomatrix{0}{1}{1}{0},\\ Y_2 &= \twomatrix{0}{i}{-i}{0}, \quad X_1 = \twomatrix{0}{1}{-1}{0}, \quad X_2 = \twomatrix{0}{i}{i}{0}.\nonumber\end{aligned}$$ The isotropy subalgebra is given by $\hg = \RR Z$, $\mg = \operatorname{span}_\RR \{ Y_0, Y_1, Y_2, X_1, X_2\}$ is a reductive complement, and it decomposes into irreducible modules as $\mg = \pg_0 \oplus \pg_1 \oplus \qg_1$, where $\pg_0 = \RR Y_0$, $\pg_1 = \RR Y_1 \oplus \RR Y_2$, $\qg_1 = \RR X_1 \oplus \RR X_2$. Also, if $\kg = \hg \oplus \qg_1 \simeq \sug(2)$, $\pg = \pg_0 \oplus \pg_1$, then $\slg_2(\CC) = \kg \oplus \pg$ is a Cartan decomposition. Let us fix an inner product $\ip_B$ on $\slg_2(\CC)$ that makes $\mathcal{B}$ orthonormal (this inner product is, up to a scalar multiple, the one given by the Killing form of $\slg_2(\CC)$, after reversing its sign on the subalgebra  $\kg$). Finally, let $\ip_0 = \ip_B \big|_{\mg \times \mg}$, which is of course $\Ad(\U(1))$-invariant. \[lemmasl2C\] Up to isometry, $\Sl_2(\CC)$-invariant metrics on $\Sl_2(\CC)\slash\U(1)$ can be parameterized by $\Ad(\U(1))$-invariant inner products on $\mg$ of the form $$\langle \cdot, \cdot \rangle_h = \langle h\, \cdot\, , h\, \cdot \rangle_0,$$ where $h\in \Gl_5(\RR)$ is given by $$h = \left[\begin{array}{ccccc} e &0 &0 &0 & 0\\ 0& a &0 &0 &0 \\0 & 0& a & 0&0 \\0 &0 & -d& b &0 \\ 0&d &0 &0 & b \end{array}\right], \qquad a,b,d,e \in \RR, \quad a,b,e\neq 0.$$ Moreover, for each such metric, the Ricci curvature satisfies $$\Ricci(h^{-1 }Y_1, h^{-1} X_2) = 4 \, d \cdot \left( (a^2 - e^2)^2 + a^2(b^2 + d^2) \right)a^{-3} b^{-2} e^{-2}.$$ Since the modules $\pg_1$ and $\qg_1$ are the only equivalent modules, and they are of complex type, it is clear that the metrics are parameterized by inner products $\ip_h$ on $\mg$, where $h$ is as in the statement, but with a $2\times 2$ block of the form $\minimatrix{c}{-d}{d}{c}$ mapping $\pg_1$ to $\qg_1$. Using that $\ip_h$ and $\ip_{h \cdot T}$ give rise to isometric metrics for any $T = \Ad\left(\exp{t Y_0}\right) \in \Aut(\slg_2(\CC))$, it is easy to find $t$ so that the matrix $h \cdot T$ has the desired form. The formula for the Ricci curvature follows from a routine (though somewhat lengthy) computation. The importance of the previous formula for the Ricci curvature is that this off-diagonal entry vanishes if and only if $d=0$ (recall also that $\langle h^{-1} Y_1, h^{-1} X_2 \rangle_h = 0$). But if $d=0$ then the Cartan decomposition is orthogonal, and the metric is non-Einstein. ### $\SU(2,1)/\Delta_{p,q}\U(1)$ These spaces are the non-compact analogous of the well-known Aloff- Wallach spaces [@AW75]. As long as $p\neq 0$, the Cartan decomposition is orthogonal with respect to any $\SU(2,1)$-invariant metric, hence none of them is Einstein by [@Nkn2]. However, the space corresponding to $p=0$, $q=1$ admits $\SU(2,1)$-invariant metrics which make no Cartan decomposition orthogonal. Let us have a closer look at the Lie algebra $\sug(2,1)$: an $\Ad(\Delta_{0,1}\U(1))$-invariant decomposition is given by $\sug(2,1) = \hg_{0,1} \oplus \qg_0 \oplus \qg_1 \oplus \pg_1 \oplus \pg_2$, where $$\begin{aligned} \hg_{0,1} &= \RR \threematrix{0}{0}{0}{0}{i}{0}{0}{0}{-i}, \quad \qg_0 = \RR \threematrix{2i}{0}{0}{0}{-i}{0}{0}{0}{-i}, \quad \qg_1 = \left\{ \threematrix{0}{z}{0}{-\bar{z}}{0}{0}{0}{0}{0} : \, z\in \CC \right\}, \\ & \pg_1 = \left\{ \threematrix{0}{0}{z_1}{0}{0}{0}{\bar{z_1}}{0}{0} : \, z_1\in \CC \right\}, \quad \pg_2 = \left\{ \threematrix{0}{0}{0}{0}{0}{z_2}{0}{\bar{z_2}}{0} : \, z_2\in \CC \right\},\end{aligned}$$ and the modules $\qg_1$ and $\pg_1$ are equivalent. Any invariant metric would then make the subspaces $\qg_0$, $\pg_2$ and $\qg_1\oplus \pg_1$ orthogonal. But observe that $\ad(\qg_0)$ acts trivially on $\qg_0$ and $\pg_2$, and it acts precisely as the isotropy $\hg_{0,1}$ on $\qg_1\oplus \pg_1$. This immediately implies that for any invariant metric, $\ad(\qg_0)$ consists of skew-symmetric endomorphisms, and hence by Lemma \[lem\_formulaRicci\] the Ricci curvature is non-negative in this directions. Therefore, $\SU(2,1)\slash \Delta_{0,1}\U(1)$ admits no invariant metrics of negative Ricci curvature. $\dim G/H = 8$ -------------- $ $ The first two spaces of dimension $8$ in Table \[tabla\] are Lie groups and will be omitted. The next two cases correspond to homogeneous spaces $G/H$ where $\rank G = \rank H$. As is well known, this implies that the isotropy representation decomposes a sum of pairwise inequivalent modules. Clearly, any Cartan decomposition will be orthogonal with respect to any $G$-invariant metric, and thus none of those can be Einstein by [@Nkn2]. For the infinite family of homogeneous spaces $\left(\Sl_2(\RR) \times \Sl_2(\RR) \times \Sl_2(\RR)\right) / \Delta_{a_1,a_2,a_3} \U(1)$ ($6$-th line in the table), the isotropy representation may have some equivalent modules in some special cases, but they are all contained in the subspace $\pg$ of the Cartan decomposition. This implies that for any $G$-invariant metric one still has $\qg\perp\pg$, and none of them can be Einstein. Let us now consider the following spaces: $$\Sl_2(\RR)\times \Sl_2(\CC)/\U(1),\quad \Sl_2(\RR) \times \left(\Sl_2(\RR)\times \Sl_2(\RR)\right)/\Delta_{p,q} \U(1), \quad \Sl_2(\RR) \times \SU(2,1)/\SU(2).$$ They are all of the form $\Sl_2(\RR) \times G_1 / H$, for some semisimple Lie group $G_1$. Notice that all of them admit metrics which are not Cartan-orthogonal for *any* Cartan decomposition, and hence [@Nkn2] can not be applied. Another problem that arises when studying the Einstein equation in these spaces is that whenever the isotropy representation of the space $G_1/H$ has some trivial modules, then the space $G/H$ admits non-product $G$-invariant metrics (cf.  Proposition \[prodRiem\]). However, there is still *some* control on such trivial modules. Namely, an easy computations with Lie brackets shows that for the spaces under consideration we have $[\mg_0,\mg_0] \subseteq \hg$, where $\mg_0$ represents the trivial module in $G_1/H$. By looking at the Ricci curvature of $G/H$ at $eH$ in directions tangent to the orbit of $G_1$, and in directions orthogonal to this orbit, we were able to show that if the Ricci curvature preserves this orthogonality, then this forces the metric to be a product (clearly, for an Einstein metric such orthogonality would automatically be preserved by the Ricci curvature). Since $\Sl_2(\RR)$ does not admit any left-invariant Einstein metric, this proves that $G/H$ does not either. \[Propsl2RxG1\] Let $G/H = \Sl_2(\RR)\times \left( G_1/ H\right)$ be a homogeneous space with $G_1$ semisimple, and assume that $N_{G_1}(H) / H$ is abelian. Then, $G/H$ admits no $G$-invariant Einstein metric. Assume that there exists a $G$-invariant Einstein metric $g$ on $G/H$. Let $\ggo_1 = \hg \oplus \mg$ be an $\Ad(H)$-invariant decomposition, and further decompose $\mg = \mg_0 \oplus \mg_1$, where $\hg \oplus \mg_0 = \Lie(N_{G_1}(H))$. Then $\mg_0 = \{X\in \mg : [\hg,X] = 0 \}$ is the trivial $\Ad(H)$-module, $\mg_1$ is the sum of all non-trivial $\Ad(H)$-modules of $\mg$, and our assumption implies that $[\mg_0,\mg_0]\subseteq \hg$. By setting $\pg_0 = \slg_2(\RR) \oplus \mg_0$, $\pg = \pg_0 \oplus \mg_1$, we have a reductive decomposition for $G/H$ given by $\ggo = \hg \oplus \pg$, and $\pg_0$ corresponds to the trivial module. In particular, $\pg_0 \perp \mg_1$. Setting $\lgo := \mg_0^\perp \subseteq \pg_0$, we obtain the orthogonal decomposition $$\pg = \rlap{$\overbrace{\phantom{\lgo\overset{\perp}\oplus\mg_0}}^{\pg_0}$} \lgo \overset{\perp}\oplus \underbrace{\mg_0\overset{\perp}\oplus\mg_1}_\mg.$$ Our assumption $[\mg_0,\mg_0]\subseteq \hg$ implies that the following bracketing relations are satisfied: $$\begin{aligned} [\hg,\pg_0] &= 0, & [\hg,\mg_1] &\subseteq \mg_1, & [\lgo,\lgo] &\subseteq [\pg_0,\pg_0] \subseteq \hg\oplus\pg_0, \label{bracketrelations}\\ [\lgo,\mg_1]&\subseteq \hg\oplus\mg, & [\mg_0,\pg_0]&\subseteq \hg, & [\mg,\mg_1]&\subseteq \hg\oplus \mg. \nonumber\end{aligned}$$ Since $[\hg,\pg_0]=0$ we may use Lemma \[lem\_formulaRicci\] to obtain $$\langle \Ricci X, Y \rangle = \unc \sum_{r,s} \langle [U_r, U_s]_\pg, X\rangle \langle [U_r, U_s]_\pg, Y\rangle - \unm \tr S\left( \ad_\pg X\right) S\left( \ad_\pg Y\right),$$ where $X,Y \in \pg_0$ and $\{U_r\}$ is any orthonormal basis for $\pg$. Assume from now on that $X\in \mg_0$, $Y\in \lgo$, and that $\{ U_r\}$ is the union of orthonormal basis for $\lgo, \mg_0$ and $\mg_1$. Noticing that by one has that $[\pg,\mg] \perp \lgo$ and that $\ad_\pg X$ only acts nontrivially on $\mg$, the above formula simplifies as $$\langle \Ricci X, Y \rangle = \unc \sum_{U_r,U_s\in \lgo} \langle [U_r, U_s]_\pg, X\rangle \langle [U_r, U_s]_\pg, Y\rangle - \unm \tr S\left( \ad_\mg X\right) S\left( \ad_\mg Y\right).$$ Choose an orthonormal basis $\{Y_i \}_{i=1}^3$ for $\lgo$, with $Y_i = A_i + B_i$, $A_i\in \slg_2(\RR)$, $B_i\in \mg_0$, and such that $\{A_i \}$ is a *Milnor basis* for $\slg_2(\RR)$, with brackets $$[A_2,A_3] = \alpha A_1, \qquad [A_3,A_1] = \beta A_2, \qquad [A_1, A_2] = \gamma A_3, \qquad\alpha,\beta,\gamma \neq 0.$$ Also, choose an orthonormal basis $\{ X_j^0\}_{j=1}^d$ for $\mg_0$ so that $\tr S(\ad_\mg X_i^0) S(\ad_\mg X_j^0) = 0$ if $i\neq j$ (this is indeed possible since the application $(X,Y) \mapsto \tr S(\ad_\mg X)S(\ad_\mg Y)$ is a symmetric bilinear form on $\mg_0$). Then, a straightforward calculation shows that $$\langle \Ricci X_j^0, Y_3 \rangle = -\langle B_3, X_j^0\rangle \left( \gamma^2 + \tr S(\ad_\mg X_j^0)^2 \right),$$ and using the Einstein condition we conclude that $\langle B_3, X_j^0\rangle = 0$. Analogously, we obtain that $\langle B_i, X_j^0\rangle = 0,$ for all $i=1,2,3$, $j=1,\ldots,d$, thus $\langle \slg_2(\RR), \mg_0\rangle = 0$. Therefore, $\slg_2(\RR)\perp \mg$, and this implies that the metric is locally a Riemannian product. But this is a contradiction, since $\widetilde{\Sl_2(\RR)}$ does not admit any left-invariant Einstein metric. Finally, we study the family $\left(\Sl_2(\RR) \times \Sl_2(\CC)\right)/\Delta_{p,q}\U(1)$. Unfortunately, we are not able to deal with the case where $p = q$. Notice though that this missing case represents just one homogeneous space from the above infinite family. With respect to the inclusions $\hg_1:= \sog(2) \subseteq \slg_2(\RR), \hg_2 := \ug(1) \subseteq \sug(2) \subseteq \slg_2(\CC)$ we have that $$\hg := \Delta_{p,q} \ug(1) \subseteq \hg_1 \oplus \hg_2 \subseteq \slg_2(\RR) \oplus \slg_2(\CC) =: \ggo.$$ Given an $\Ad(H)$-invariant inner product on some reductive complement $\mg$, we extend it in the usual way to an $\Ad(H)$-invariant inner product $\ip$ on $\ggo$. By looking at the decomposition of the isotropy representation from Table \[tabla\] in the case when $p\neq q$, $$\mg = \qg_0^{(1)} \oplus \qg_1^{(2)} \oplus \pg_0^{(1)} \oplus \pg_1^{(2)} \oplus \pg_2^{(2)}, \qquad \qg_1 \simeq \pg_1 \not\simeq \pg_2,$$ we see that the ideals $\slg_2(\RR), \slg_2(\CC)$ are orthogonal (notice that $\pg_2^{(2)}$, $\qg_1^{(2)} \oplus \pg_0^{(1)} \oplus \pg_1^{(2)}$ correspond to reductive complements for the homogeneous spaces $\Sl_2(\RR)/\SO(2)$, $\Sl_2(\CC)/\U(1)$, respectively, and $\hg \oplus \qg_0 = \hg_1 \oplus \hg_2$). This easily gives that $\ip$ is $\ad(\hg)$-invariant if and only if it is both $\ad(\hg_1)$- and $\ad(\hg_2)$-invariant. Thus, $\qg_0$ acts by skew-symmetric endomorphisms on $\ggo$, and by Lemma \[lem\_formulaRicci\] we have that $$\Ricci(Y,Y) = \unc \sum_{i,j} \langle [X_i,X_j]_\mg, Y \rangle^2 \geq 0, \qquad Y\in \qg_0.$$ It is worth pointing out that when $p\neq q$, all homogeneous metrics in the above homogeneous space can be approximated by *strictly locally homogeneous metrics* (cf.  [@Tric92]), namely, metrics which are locally homogeneous but are not locally isometric to any globally homogeneous manifold. This is simply done by considering irrational slopes approximating the rational slope $p/q$ of the given space. It was proved in [@Spiro] (see also [@Bhm15]) that strictly locally homogeneous metrics do not have non-positive Ricci curvature. On the other hand, if $p=q$ there exist homogeneous metrics that cannot be approximated in that way. Non-unimodular transitive group {#sectionnonuni} =============================== In this section we study Einstein homogeneous spaces $G/K$ of negative scalar curvature with $G$ non-unimodular and $G/K$ as in Theorem \[structure\]. Following the discussion of Section  \[prelimstruct\] we can assume that $$\label{decom} \ggo=(\ggo_1 + \zg(\ug)) \ltimes \ngo,$$ where $\ug=\ggo_1 + \zg(\ug)$ is a reductive Lie algebra, $\ggo_1$ is semisimple with no compact ideals, $\kg \subset \ggo_1$ and $\zg(\ug)=\RR H ,$ with $H$ the mean curvature vector (see Corollary \[cor\_rankone\]). Before starting the proof of our main result in the non-unimodular case, we state two lemmas which yield information about semisimple homogeneous spaces in low dimensions. Their proof follows immediately from Table \[tabla\], and the fact that irreducible symmetric spaces of the non-compact type are diffeomorphic to Euclidean spaces. \[diff5\] Let $G_1/K$ be a simply-connected semisimple homogeneous space of the non-compact type[^2] with $n = \dim G_1/K \leq 5.$ Then, either $G_1/K=\Sl_2(\CC)/\U(1)$ or $G_1/K \simeq \RR^n$. \[diff6\] Let $G_1/K$ be a $6$-dimensional simply-connected semisimple homogeneous space of the non-compact type such that $G_1$ contains $\widetilde{\Sl_2(\RR)}$ as a simple factor. Then, $G_1/K$ is diffeomorphic to  $\RR^6.$ We now focus on the proof of Theorem \[mainnonuni\]. Given a homogeneous Einstein space $(G/K,g)$ with negative scalar curvature and $G$ non-unimodular, we consider for it the decomposition given in . By virtue of Corollary \[reductionG1/K\], our goal will be to prove that $G_1/K$ is diffeomorphic to a Euclidean space. Observe that we may assume $\dim G_1/K \leq \dim G/K -3.$ Indeed, we always have $\dim \zg(\ug) =1$ and $\dim \ngo \geq \dim \zg(\ug)$ because the representation $\theta|_{\zg(\ug)} : \zg(\ug) \to \ngo$ is faithful (its kernel must be in the nilradical). If $\dim \ngo = 1$, we know by Theorem \[thm\_lemadimn\] and Remark  \[rmkn1\] that $(G/K, g)$ is a Riemannian product, and thus not de Rham irreducible. More generally, for this reason we can also assume that $\theta|_{\ggo_1} \neq 0.$ We now proceed with the proof, considering different cases according to the dimension of $G/K.$ $\dim G/K \leq 8$ {#dimG/Kleq8} ----------------- $ $ We have that $\dim G_1/K \leq 5.$ By Lemma \[diff5\], either $G_1/K$ is diffeomorphic to $\RR^n$ for some $n \leq 5,$ or $G_1/K=\Sl_2(\CC)/\U(1).$ In the latter case, $\theta|_{\ggo_1}=0,$ since $\dim \ngo=2$ and there exists no nontrivial $2$-dimensional representation of the simple Lie algebra $\slg_2(\CC)$. Thus, this is a product case. $\dim G/K=9$ ------------ $ $ ### $\dim G_1/K=6$ {#G1/K=6} As $\dim \ngo=2$ we know that $\theta(\ggo_1)\subseteq \End(\RR^2)$ is semisimple, so $\ggo_1$ must have an ideal isomorphic to $\slg_2(\RR).$ We conclude that $G_1/K$ is diffeomorphic to $\RR^6$ by using Lemma \[diff6\]. ### $\dim G_1/K \leq 5$ By Lemma \[diff5\], we only consider the case where $G_1/K=\Sl_2(\CC)/\U(1).$ We have that $\dim \ngo = 3$, and it is easy to see that $\theta|_{\ggo_1}=0$ since there is no subalgebra of $ \slg_3(\RR)$ isomorphic to $\slg_2(\CC)$. Hence this is also a product case. $\dim G/K=10$ ------------- ### $\dim G_1/K=7$ We have that $\dim \ngo=2$ and $\theta(\ggo_1)$ is semisimple, that is, $\ggo_1$ has an ideal isomorphic to $\slg_2(\RR)$. The list of all possible homogeneous spaces $G_1/K$ to consider is very long. However, if $G_1/K$ is a product of lower dimensional homogeneous spaces, then it must be a product of some irreducible symmetric spaces of non-compact type and some of the spaces in Table \[tabla\]. Since its dimension is $7$, it easily follows that there is at most one factor which is non-symmetric, and Proposition \[prodRiem\] implies that on $G_1/K$ the metric is a product of corresponding invariant metrics on each of the factors. Moreover, since the kernel of $\theta|_{\ggo_1}$ has codimension $3$, $\theta$ must necessarily vanish on some of these factors. This implies at once that the whole space $G/K$ splits as a Riemannian product. By the preceeding discussion, we may now assume that $G_1/K$ is non-product, i.e.  it is one of the spaces listed in Table \[tabla\]. Since $\ggo_1$ has an ideal isomorphic to $\slg_2(\RR)$ there are actually only two possibilities: the simply connected covers of $\Sl_2(\RR)^3/\Delta T^2_{a,b,c}$ and of $\SU(2,1) \times \Sl_2(\RR)/\Delta_{p,q}\U(1)(\SU(2)\times \{e\})$. But both of them are diffeomorphic to $\RR^7$, hence we are done. ### $\dim G_1/K=6$ {#dim-g_1k6} Here $\dim \ngo=3$ and $\theta|_{\ggo_1} \neq 0.$ Since $\theta|_{\ggo_1}$ maps $\ggo_1$ into $\slg_3(\RR)$, for it to be non-trivial it is necessary that $\ggo_1$ contains at least one simple ideal of dimension at most $8$. By Lemma \[diff6\], we may further assume that $G_1$ contains no $\widetilde{\Sl_2(\RR)}$ factor. Thus it is clear from Table \[tabla\] that if $G_1/K$ were a product of lower dimensional homogeneous spaces then each factor would be symmetric space, and the result would follow. On the other hand, if $G_1/K$ is non-product, it also follows from Table \[tabla\] that the only possibilities are $$\begin{aligned} \SU(2,1)/T_{max}, \quad \Sl_2(\CC).\end{aligned}$$ But then we must have $\theta|_{\ggo_1} = 0$, because there exist no nontrivial $3$-dimensional representations of the simple Lie algebras $\sug(2,1)$ or $\slg_2(\CC)$. ### $\dim G_1/K \leq 5$ {#section514sl2C} By using Lemma \[diff5\] we have that either $G_1/K \simeq \RR^n,$ for some $n \leq 5,$ or $G_1/K=\Sl_2(\CC)/\U(1).$ We need to analyze the latter case. We are reduced to showing that equation has no solutions for $\Sl_2(\CC)$-invariant metrics on $\Sl_2(\CC)/\U(1)$, for any $\theta: \ggo_1 \to \End(\RR^4)$ that satisfies . We now use the notation of Section \[sectionsl2C\]. Let us assume that $\theta(\ggo_1) \neq 0$, since otherwise we are in a product case, and consider an arbitrary $\Ad(\U(1))$-invariant inner product on $\slg_2(\CC)$ written in the form $\ip_h$ given in Lemma \[lemmasl2C\], $h\in \Gl_5(\RR)$. An orthonormal basis for the reductive complement $\mg$ is given by $\mathcal{B}_h = \{h^{-1} Y_0, \ldots, h^{-1} X_2 \}$. Up to equivalence of representations, there are two $4$-dimensional real faithful representations of $\slg_2(\CC)$: the tautological representation, and its conjugate, and they are both irreducible. Let us consider the case where $\theta$ is equivalent to the tautological representation (the other case is completely analogous). There exists $h_2 \in \Gl(\ngo)$ such that with respect to the inner product $h_2 \cdot \ip_\ngo$ the matrices of $\theta(Z)$, …, $\theta(X_2)$ have the forms (after identifying a complex number $a+b i$ with a $2\times 2$ real matrix $\minimatrix{a}{b}{-b}{a}$). This is equivalent to saying that the matrices of $(h_2^{-1} \cdot \theta)(Z), \ldots, (h_2^{-1} \cdot \theta)(X_2)$ have such a form with respect to the inner product $\ip_\ngo$. But now an easy computation shows that $$\sum_{Y \in \, \mathcal{B}_h} \left[\left(h_2^{-1} \cdot \theta\right)(Y), \left(h_2^{-1} \cdot \theta\right)(Y)^t \right] = 0,$$ that is, $h_2^{-1}\cdot\, \theta$ is also a zero of the moment map for the natural $\Gl(\ngo)$-action on $\End(\slg_2(\CC),\End(\ngo))$ (see Remark \[remarks\], ). From the rigidity imposed by Geometric Invariant Theory for such zeros [@RS90 Theorem 4.3], we can conclude that in fact $h_2\in \Or\left(\ngo, \ip_\ngo\right)$, and thus the matrices of $\theta(Z), \ldots, \theta(X_2)$ have the form with respecto to $\ip_\ngo$ (see also the proof of [@semialglow Proposition A.1] for a more detailed application of this argument). We now plug this information into equation , and since $\theta(X_2)$ is skew-symmetric, by looking at the Ricci curvature and using Lemma \[lemmasl2C\] we obtain that $d=0$. In other words, $h$ is diagonal, and in particular the metric associated to $\ip_h$ leaves orthogonal the Cartan decomposition $\slg_2(\CC) = \kg \oplus \pg$, $\kg = \hg \oplus \qg_1$, $\pg = \pg_0 \oplus \pg_1$. Notice that the operator $C_\theta \in \End(\mg)$ given by $$\left\langle C_\theta X, Y\right\rangle = \tr S \left( \theta(X)\right) S\left(\theta(Y)\right)$$ is a positive multiple of the identity on $\pg$. Hence, by following the same arguments used in the proof of [@Nkn2 Theorem 1] we can conclude that the equation $$\label{eqRictheta} \Ricci_{\ip} = c \, I + C_\theta$$ can not be satisfied for $\ip_h$. Therefore, there are no Einstein metrics in this case. The fact that the proof of [@Nkn2 Theorem 1] could be adapted for the more general equation as long as $G_1$ is simple was kindly communicated to us by Jorge Lauret [@Lauretpersonalcom]. Strong Alekseevskii’s conjecture {#strong} ================================ This section is devoted to studying the strong Alekseevskii conjecture and showing that it holds up to dimension $8$, with the possible exceptions of invariant metrics on non-compact semisimple Lie groups or on the space $\left(\Sl_2(\RR)\times \Sl_2(\CC)\right)/\Delta\U(1)$. As in the previous section, we consider Einstein homogeneous spaces $G/K$ with $G$ chosen as in Theorem  \[structure\], which are not Riemannian products. We also assume the simply-connected hypothesis, which is non-restrictive since it suffices to prove the strong Alekseevskii conjecture in the simply-connected case (see Remark \[remarks\] (e) and [@AC99; @Jab15]). Regarding the semisimple case, according to Theorem \[thmsemisimple\] all semisimple homogeneous spaces in Table \[tabla\] are either Lie groups, or the space $\left(\Sl_2(\RR)\times \Sl_2(\CC)\right)/\Delta\U(1)$, or they do not admit an Einstein metric. The remaining semisimple homogeneous spaces are symmetric spaces, and it is well-known that they are isometric to solvmanifolds. Therefore, by [@Dtt88], in the following we will only study the cases where the transitive group is non-unimodular (see Remark \[remarks\], ). We proceed case by case, according to the dimension of $G/K.$ Our goal will be to show that $G_1/K$ is isometric a solvmanifold. $\dim G/K=6$ {#strong6} ------------ $ $ These spaces were analyzed by Jablonski and Petersen in [@JblPtr §4]. $\dim G/K=7$ ------------ $ $ First we state the following lemma, which follows easily from Table \[tabla\] and the well-known fact that irreducible symmetric spaces of the non-compact type are isometric to solvmanifolds. \[solv5\] Let $(G_1/K, g)$ be a simply-connected semisimple homogeneous space of the non-compact type with a $G_1$-invariant metric $g$, and $\dim G_1/K \leq 5.$ Then, either $(G_1/K,g)$ is isometric to a solvmanifold, or $G_1/K$ is one of the following spaces $$\begin{gathered} \widetilde{\Sl_2(\RR)}, \quad \Sl_2(\CC)/\U(1),\quad \left(\Sl_2(\RR)\times \Sl_2(\RR)\right)/\Delta_{p,q} \SO(2), \\ \SU(2,1)/\SU(2), \quad \Sl_2(\RR) \times \Sl_2(\RR)/ \SO(2).\end{gathered}$$ By using Theorem \[thm\_lemadimn\], Remark \[rmkn1\] and Corollary \[cor\_rankone\], we can assume that $\dim G_1/K \leq 4,$ $\dim\zg(\ug)=1$ and $\dim \ngo \geq 2.$ We divide into cases according to the dimension of $G_1/K.$ ### $\dim G_1/K=4$ By Lemma \[solv5\] we know that $G_1/K$ is a solvmanifold. ### $\dim G_1/K=3$ {#G/K7G_1/K3} The only case in which $G_1/K$ is not a solvmanifold is when $G_1/K=\widetilde{\Sl_2(\RR)}.$ The existence of an example in this case would imply that there is a $6$-dimensional unimodular expanding algebraic soliton, by using that non-unimodular Einstein spaces are one-dimensional extensions of unimodular algebraic solitons (see [@alek §6]). Since for that soliton one would have $\ug=\slg_2(\RR)$, we arrive at a contradiction by using [@semialglow Appendix]. $\dim G/K=8$ ------------ $ $ We assume that $\dim G_1/K \leq 5$ and $\dim \ngo \geq 2.$ ### $\dim G_1/K=5$ We have that $\dim \ngo=2$. Then, by using Lemma \[solv5\], we consider the following possibilities to $G_1/K:$ $$\Sl_2(\CC)/\U(1), \quad \SU(2,1)/\SU(2), \quad \left(\Sl_2(\RR)\times \Sl_2(\RR)\right)/\Delta_{p,q}\SO(2), \quad \Sl_2(\RR) \times \Sl_2(\RR)/ \SO(2).$$ In the first two cases, we must have $\theta|_{\ggo_1} = 0$ since there exist no nontrivial $2$-dimensional representations of the simple Lie algebras $\slg_2(\CC)$ and $\sug(2,1).$ By Theorem \[thm\_lemadimn\] and Remark \[rmkn1\], these are product cases. In the latter case, we also are in a product case. Indeed, the metric restricted to $G_1/K$ is a product metric. In addition, since $\dim \ngo=2,$ $\theta$ must necessarily vanish on some of these factors. This implies at once that the whole space $G/K$ splits as a Riemannian product. We now deal with the case $G_1/K =\left(\Sl_2(\RR)\times \Sl_2(\RR)\right)/\Delta_{p,q}\SO(2)$. This case is similar in nature to the one in Section \[section514sl2C\]: we are reduced to solving equation  for invariant metrics on $G_1/K$, for every possible representation $\theta: \ggo_1 = \slg_2(\RR)\oplus \slg_2(\RR) \to \End(\RR^2)$. Let $H = \minimatrix{1}{0}{0}{-1}$, $X = \minimatrix{0}{1}{-1}{0}$, $Y = \minimatrix{0}{1}{1}{0}$ be a basis for $\slg_2(\RR)$, and consider the ordered basis $\mathcal{B}$ for $\slg_2(\RR) \oplus \slg_2(\RR)$ given by $$\begin{aligned} Z = \left(p \,H, q\, H\right), \quad X_0 &= \left(q \,H, -p\, H\right), \quad Y_1 = \left( X, 0\right), \\ Y_2 = \left( Y,0\right), \quad X_1 &= \left( 0, X\right), \quad X_2 = \left( 0, Y\right).\end{aligned}$$ The isotropy subalgebra is $\hg_{p,q} = \RR Z$, $\mg = \operatorname{span_\RR}\{X_0, Y_1, Y_2, X_1, X_2 \}$ is a reductive complement, and the decomposition into irreducible submodules is given by $\mg = \qg_0 \oplus \pg_1 \oplus \pg_2$, where $\qg_0 = \RR X_0$, $\pg_1 = \RR Y_1 \oplus \RR Y_2$, $\pg_2 = \RR X_1 \oplus \RR X_2$, and $\pg_1 \simeq \pg_2$ if and only if $p=q$. First notice that $\theta$ must have a kernel, which we may assume without loss of generality to be the first $\slg_2(\RR)$ factor. This implies that $\theta(H,0) = 0$, and since $\theta(Z)$ is skew-symmetric, it is clear that $\theta(X_0)$ must also be skew-symmetric. Thus, using equation we obtain that $$\Ricci_{G_1/K}(X_0,X_0) < 0.$$ This is already enough to rule out the cases $p\neq q$, since in those cases we have $\pg_1 \not\simeq \pg_2$, which forces $X_0$ to act skew-symmetrically on $\ggo_1$, and by Lemma \[lem\_formulaRicci\] we get a contradiction. Let us now consider the remaining case $p = q = 1$, which is considerably more difficult. The following is the analogous of Lemma \[lemmasl2C\] for this situation, and can be proved in the very same way. Up to isometry, $\Sl_2(\RR)\times \Sl_2(\RR)$-invariant metrics on $\left(\Sl_2(\RR)\times \Sl_2(\RR)\right) / \Delta_{1,1} \SO(2)$ can be parameterized by $\Ad(\SO(2))$-invariant inner products on $\mg$ of the form $$\langle \cdot, \cdot \rangle_h = \langle h\, \cdot\, , h\, \cdot \rangle_0,$$ where $h\in \Gl_5(\RR)$ is given by $$h = \left[\begin{array}{ccccc} e &0 &0 &0 & 0\\ 0& a &0 &0 &0 \\0 & 0& a & 0&0 \\0 &d & 0& b &0 \\ 0&0 &d &0 & b \end{array}\right], \qquad a,b,d,e \in \RR, \quad a,b,e\neq 0.$$ Reasoning as in Section \[section514sl2C\], we see that condition implies that $\theta$ restricted to the second $\slg_2(\RR)$ factor is nothing but the tautological representation of this Lie algebra. Let us consider the operator $\Ricci^\theta \in \End(\mg)$ given by $$\left\langle \Ricci^\theta \, X, Y\right\rangle_h = \Ricci_{G_1/K}(X,Y) - \tr S(\theta(X)) S(\theta(Y)).$$ Equation for a metric $\ip_h$ can be rephrased as $$\label{eqRiccithetaop} \Ricci^\theta = c I, \qquad c<0.$$ Since we now know $\theta$ and $\ip_h$ explicitly, we can actually compute $\Ricci^\theta$ in terms of $a,b,d,e$. Let us call $r_{i,j}^\theta$, $1\leq i,j \leq 5$, the entries of the matrix of $\Ricci^\theta$ with respect to the $\ip_h$-orthonormal ordered basis $\mathcal{B}_h = \left\{ h^{-1} Y \mid Y\in \mathcal{B} \right\}$ . Then, assuming that $\det h = 1$, we have that $$\begin{aligned} r_{1,1}^\theta &= \unm \left( a^4 e^4 + \left(b^2 - d^2\right)^2 \right) + a^2 d^2 \left( e^2- 4b^2\right) \left( e^2 + 4 b^2\right), \\ r_{1,1}^\theta + 2\, r_{4,4}^\theta + 2\, a \cdot d^{-1} \, r_{2,4}^\theta & = \unm \left(a^2 - b^2 + d^2 \right)^2 e^4 + 4 a^4 b^2 \left( 4 b^2 - e^2 \right).\end{aligned}$$ Despite the ugliness of these formulas, we see that all the terms and factors on the right hand side are positive except for $ e^2 - 4 b^2 $, which appears with a different sign in both of them. For a solution of , both expressions should be negative (they would equal $c$ and $3\, c$, respectively). It is now clear that such a solution does not exist. ### $\dim G_1/K=4$ By Lemma \[solv5\] we know that $G_1/K$ is isometric to a solvmanifold. ### $\dim G_1/K=3$ {#dim-g_1k3} Here, the only case to consider is $G_1/K=\widetilde{\Sl_2(\RR)}.$ Similarly to the case in Section  \[G/K7G\_1/K3\], we have a contradiction. [^1]: See Remark \[remarks\], below. [^2]: see Definition \[defsshomogspace\].

--- abstract: 'We performed a new analysis of $B$ and $R$ light curves of a sample of PG quasars. We confirm the variability-redshift correlation and its explanation in terms of spectral variability, coupled with the increase of rest-frame observing frequency for quasars at high redshift. The analysis of the instantaneous spectral slope for the whole quasar samples indicates both an inter-QSO and intra-QSO variability-luminosity correlation. Numerical simulations show that the latter correlation cannot be entirely due to the addition of the host galaxy emission to a nuclear spectrum of variable luminosity but constant shape, implying a spectral variability of the nuclear component. Changes of accretion rate are also insufficient to explain the amount of spectral variation, while hot spots possibly caused by local disk instabilities can explain the observations.' author: - Dario Trèvese - Fausto Vagnetti title: Quasar Spectral Slope Variability in the Optical Band --- INTRODUCTION ============ Although variability plays a key role in constraining the size of the central engine of active galactic nuclei, yet its physical origin remains substantially unknown. Even restricting to the class of non-Blazar objects, the most diverse mechanism have been proposed in recent years, including gravitational lensing due to intervening matter [@haw96], supernovae explosions [@are97], instabilities in the accretion disk [@kaw98] and star collisions [@tor00]. For a small number of low redshift objects, a multi-wavelength monitoring with adequate time sampling and resolution allows the interpretation of changes of the spectral energy distribution (SED) in terms of an interplay of emission components with different spectral and variability properties (see [@umu97] for a general review and [@cou98] for the case of 3C273). In the near infrared-optical-UV bands, variability studies indicate a hardening of the spectrum in the bright phase [@cut85; @ede90; @kin91; @pal94]. So far, however, most of the statistical information on AGN variability, derives from single-band light curves of magnitude limited samples of objects [@ang72; @bon79; @haw83; @tre89; @cri90; @tre94; @hoo94; @ber98]. In the case of magnitude limited samples the analysis is complicated by the strong luminosity-redshift (L-z) correlation, caused by the crowding of objects towards the limiting flux. As a consequence it is difficult to isolate any intrinsic variability-luminosity (v-L) and variability-redshift (v-z) correlation. Moreover, the results of these analyses depend on the specific variability index adopted, as shown by [@gia91] who found a positive v-z correlation through a variability index defined on the basis of the rest-frame structure function. The existence of an average increase of variability with redshift was later confirmed by [@cri96]. It turns out to be consistent with the suggestion of [@gia91] that QSOs at high redshift appear more variable since they are observed at a higher rest-frame frequency, where the variability is stronger (hardening in the bright phase and vice versa). A direct statistical evidence of the spectral variability, in terms of an average change of the B-R color with the variation of the B magnitude, was found by [@giv99] as part of a statistical analysis of the B and R light curves of a sample of PG QSOs. A similar evidence has been found by [@tre01] for the faint QSO sample of SA 57 [@kkc]. Unfortunately, most (if not all) of the variability mechanisms proposed so far imply a hardening of optical-UV spectrum in the bright phase, so that they cannot be discriminated on purely qualitative grounds. As a first step towards constraining variability models quantitatively [@tre01] compared with the observations a simple model where both spectral slope changes and brightness variations are due to temperature changes of an emitting black body. In the present paper, we present a new analysis of the B and R light curves of a sample of PG QSOs made available by [@giv99], we analyze the consistency with previous results, and consider possible different sources of spectral variability. The paper is organized as follows: in section 2 we summarize the characteristics of the [@giv99] data, in section 3 we analyze the v-z correlation, in section 4 we discuss the parameters adopted for the spectral variation analysis, in section 5 we consider the effect of the host-galaxy on SED variations, in section 6 we compare the observed spectral variations with those produced by changes of the accretion rate, and discuss the consistency of the observations with a simple model consisting of hot spots on the accretion disk. Section 7 contains a summary and the conclusions. We adopt $H_o= 50 {\rm ~km~s^{-1}~Mpc^{-1},~q_o=0.5}$, unless otherwise stated. THE DATA ======== The present analysis is based on the light curves made available to the community by the Wise Observatory group [@giv99]. The sample consist of 42 PG QSOs selected to be nearby, i.e. $z < 0.4$, and bright, i.e. $B < 16$ mag. The observations were made in the Johnson-Cousins B and R bands with the 1 m Wise Observatory telescope. The total duration of the campaign was 7 years and the median observing interval of the objects was 39 days. The r.m.s. photometric uncertainty is $\sim 0.01$ mag and $\sim 0.02$ mag in the B and R bands respectively. We refer to the original paper of [@giv99] for all details concerning observations, calibration etc. The full B and R light curves of the entire sample were made available in electronic form. [@giv99] present the analysis of the correlation of different variability properties with other properties like luminosity, redshift, radio power, various line intensities and the X-ray spectral slope. In particular they show a correlation between the color changes $\Delta (B-R)$ and the brightness variations $\Delta B$ and $\Delta R$, corresponding to an average hardening of the spectrum in the bright phase (and vice versa). They do not find a correlation between variability and redshift, at variance with [@gia91], [@tre94], [@cri96], but they ascribe this to the difficulty of disentangling v-z, v-L and L-z correlations in the sample. We stress that this is made particularly difficult by the small redshift interval spanned by the sample. We perform a new analysis of both the v-z correlation and the spectral slope changes in the following sections. For these purposes we have dereddened the data, using the extinction calculator of the NASA Extragalactic Database (NED). VARIABILITY-REDSHIFT CORRELATION ================================ To measure the amplitude of variability we define, for each object, the first order structure function, as in [@dic96] (but omitting the subscript “1”): $$S(\tau,\Delta \tau)=[({\pi\over 2}\overline{|m(t+\tau)-m(t)|}^2 -\sigma_n^2)]^{1\over 2}$$ where $m(t)$ is either the $B$ or the $R$ magnitude, $t$ is the rest-frame time, $\tau$ is the time lag between the observations, $\sigma_n$ is the relevant r.m.s. noise and the bar indicates the average taken over all the pairs of observations lying in the time interval $\tau \pm \Delta \tau$. In this definition we adopt the square average of the absolute values of the difference instead of the average of the square difference as in [@dic96], since the former quantity is less sensitive to outliers [@hoo94]. The $\pi/2$ factor normalizes $S$ to the r.m.s. value in the case of a Gaussian distribution. The adopted value of $\Delta \tau$ is the result of a trade-off between time resolution and statistical uncertainty. In the following we define four variability indices $S_i(\tau \pm \Delta \tau)$, with $i=B,R$, and $\tau=0.3 \pm 0.09$ yr, $2.0 \pm 0.6$ yr. The subscripts $B$ and $R$ refer to the observing band and the values of $\tau$ and $\Delta \tau$ have been chosen for comparison with previous analyses [@dic96]. None of these four indices shows a significant correlation with redshift when the whole sample is considered, confirming the result of [@giv99]. However, to disentangle the v-L, v-z and L-z correlations, we can restrict the analysis to a magnitude bin $-23.5 < M_B < -22.5$, around the average absolute magnitude of the sample $<M_B>=-22.75$. The result is shown in Figure 1. In this case we find a v-z correlation coefficient $r_{v,z}=0.39$, which is marginally significant ($P(>r)=0.09$) despite the small number of objects (19) in the bin. To examine the dependence of variability on redshift we take the ensemble averages of the four variability indices defined above, over the same subsample of 19 objects. For each observing band we compute the average rest-frame observing frequency of the sample. To compare the result with [@dic96] we must take into account the dependence of variability on magnitude. For this purpose, we reduce $S$ by an amount $\Delta S = (\partial S / \partial M_B)~\Delta M_B$ where for $S(M_B)$ we adopt model A of [@cri96] and $\Delta M_B$ is the difference between the average absolute magnitudes of the present sample and the sample of [@dic96]. Figure 2 shows the increase of variability with the observing frequency. At each frequency variability is larger for larger time lag $\tau$ (due to the increase of the structure function from $\tau=0.3$ yr to $\tau=2$ yr). The new points, obtained from the present analysis, are consistent with the previous results, which were obtained from different samples and observing frequencies. The general trend can be quantified as: $$\partial S_i / \partial \log \nu_{rest} = \partial S_i / \partial \log (1+z) \simeq 0.25 - 0.3$$ and is consistent with [@gia91], [@cri96] and [@dic96]. Therefore the increase of variability with redshift and its interpretation in terms of an average increase of the amplitude of variability with frequency are confirmed by the present results despite their poor statistical basis. [@haw96] finds that the variability timescale is independent of redshift, i.e. it is not affected by cosmological time dilation. He suggests that the effect can be explained in the framework of gravitational microlensing caused by intervening matter. [@haw01] estimates that a possible decrease of variability [*timescale*]{} with frequency is not sufficient to compensate for the cosmological time dilation. The frequency dependence of variability [*amplitude*]{} can also compensate, at least in part, for time dilation. We note that from Eq. 2 we can estimate a vertical shift of $\sim 0.1$ between $B$ and $R$ power spectra, approximately consistent with the horizontal shift of $\sim 0.06$ found by [@haw01] (see his Figure 3). However, a detailed evaluation of the amplitude shift of the variability power spectrum as a function of redshift would require the knowledge of the redshift, luminosity, and observing time distributions of the sample. We also stress that the frequency dependence of microlensing [@ale95] should be compared with the results shown in Fig. 2 and Eq. 2. THE SPECTRAL SLOPE $\alpha$ AND THE SPECTRAL VARIABILITY PARAMETER $\beta$ ========================================================================== The average increase of variability with frequency can be interpreted in terms of a change of the spectral slope with luminosity. However the statistical evidence discussed in the previous section, and in [@gia91], [@dic96], is only indirect and must be confirmed by observations of individual objects in at least two bands. This was done by [@giv99] in terms of correlation of color changes $\Delta(B-R)$ with brightness variations $\Delta B$ or $\Delta R$ and by [@tre01] in terms of changes of the spectral slope $\alpha$, defined by $L_{\nu} \sim \nu^{\alpha}$, where $L_{\nu}$ is the intrinsic power per unit frequency. The latter work is based on $U$,$B_J$,$F$, $N$ observations at two epochs of the sample of QSOs of the Selected Area 57. The increase of spectral slope is computed as an average value over the ensemble of objects, and compared with a simple model consisting of a black body subject to small temperature changes. The data of [@giv99], which contain on average 40 $B$ and $R$ observations of each QSO, allow the statistical analysis of spectral slope changes of each object, for which we compute the instantaneous slope: $$\alpha(t) \equiv \log (L_{\nu_B}/L_{\nu_R})/\log(\nu_B/\nu_R)= -\frac{0.4[(B-R)-(B_o-R_o)]} {\log \frac {\lambda_R}{\lambda_B}} - 2,$$ where $m_o=-2.5 \log f_o$ are the zero points of B and R photometric bands respectively [@cox00]. Taking into account that typical variability time scales are $\approx 1$ yr (see e.g. [@tre94]) we regard as simultaneous the observations within a time interval of 9 hours. In Figure 3 we report the instantaneous value of the spectral slope $\alpha$ as a function of the intrinsic luminosity $L_{\nu_B}$. Each small cloud of points represents a single QSO in different luminosity states. The relevant regression lines are reported on each cloud. They show, with a few exceptions, a positive correlation between the spectral slope and the intrinsic luminosity, which we call intra-QSO $\alpha-L$ correlation. The distribution of clouds in Figure 3 also shows that the average slope of each QSO tends to be larger for brighter objects, forming a sort of QSO main sequence in the $\alpha-L$ plane [@tre01a]. This inter-QSO $\alpha-L$ correlation can be quantified considering for each QSO the average values $\langle\alpha\rangle$ and $\langle L_{\nu_B}\rangle$ of the slope and luminosity. The correlation $r_{\alpha-L}=0.58$ is highly significant: $P(>r)=6 \cdot 10^{-5}$. Figure 3 might suggest that both correlations are produced by the same physical mechanism, e.g. an increase of the temperature of the emitting gas ([@tre01], see also [@pal94]). However, in order to try any comparison with possible models, a precise quantification of the above effects is needed. In fact, a variety of mechanisms might in principle be responsible for a hardening of the spectrum in the bright phases. For instance, if AGNs are powered by supernovae explosions and variability is caused by a “Christmas tree” effect, then there will be an excess of blue emission in the bright phase [@are97; @cid00]. Even gravitational lensing [@haw96], usually thought of as achromatic, can produce a stronger variability in the blue than in the red band since the amplification depends on the size of the accretion disk which is larger in the red than in the blue band [@ale95]. Also hot spots due to instabilities of the accretion disk [@kaw98] are likely to enhance the blue emission. Finally the spectral energy distribution of the host galaxy, which is redder than the AGN and gives a stronger contribution in the faint AGN phase, can produce a similar effect. In order to quantify the spectral variations we define a spectral variability parameter (SVP) representing the spectral slope changes per unit log-luminosity change: $$\beta(\tau)\equiv\frac{\alpha(t+\tau)-\alpha(t)}{\log L_B(t+\tau)-\log L_B(t)},$$ where $L_B(t)$ is the luminosity in the $B$ band and $\tau$ is an appropriate time delay. In fact we expect that different physical phenomena are causing different SED changes on the relevant time scales, which go at least from days to more than ten years. With the available light curves we have computed, for each QSO, $\beta(\tau_{ij})$ with $\tau_{ij}\equiv t_i-t_j$ ,$i,j=1,N$ representing all the possible time differences between the N points of the light curve. Figure 4 represents $\beta(\tau_{ij})$ for two of the 42 QSOs considered, PG0804+762 and PG1354+213, taken as examples of good and poor time sampling respectively. For each bin, the mean value of the SVP is also shown. The uncertainty is the r.m.s. variation of the mean. No obvious trend is seen looking at similar plots for the entire sample: the average $\beta$ values stay almost constant, within the uncertainty, at least for $\tau \mincir 1000$ d. Clearly, bins at large $\tau$ are less populated. An increase or decrease of $\beta$ appears for some objects at larger $\tau$ where, however, the statistic becomes poor. Apparently, an analysis of possible systematic $\beta$ changes on time scales $\tau \magcir 3$ yr requires a time base larger than the $7$ yr of the present one. For the following analysis we define the SVP of each QSO as the mean value $\beta_m$ in a single bin $0 < \tau < 1000$ d. Since $\beta(\tau_{ij})$ values are not independent, the above computation of the uncertainty represents an overestimate of the standard deviation of $\beta_m$. In Figure 6 $\beta_m$ of the 42 QSOs are reported versus the relevant time average $\bar{\alpha}$. The curves are described in the following paragraphs, except the dot-dashed line representing a black body, which is reported for comparison with the results of [@tre01]. In the case of a black body (of fixed area), defining $x \equiv h\nu/kT$, with $T$, $h$, $k$ equal to temperature, Planck and Boltzmann constants respectively, the spectral slope is $\alpha_{BB}(x)= 3-xe^x/(e^x-1)$ and the SVP is $(d\alpha/dT)/(d\log B_{\nu}/dT)=(\ln 10)[1-x /(e^x-1)]\equiv\beta_{BB}(x)$. We stress that the dotted line, defined by the above expression, does not represent a fit to the points (no free parameters). For a fixed frequency $\log \bar{\nu} \equiv \log (\nu_B \nu_R)^{\frac{1}{2}}$ and increasing temperature, a point moves on the curve from left to right (Rayleigh-Jeans limit $\alpha=2, \beta=0$). The “average QSO”, $\langle\bar{\alpha}\rangle = -0.2\pm 1.0, \langle\beta_m\rangle = 2.2\pm 0.9$, can be represented by temperature changes of a black body of $T\approx 10^4$ K, in approximate agreement with the result of [@tre01], which was obtained with different QSO sample and mean rest-frame frequency. THE EFFECT OF THE HOST GALAXY ============================= An independent analysis of the same light curves, performed by [@cid00] in the framework of Poissonian model of variability, implies the existence of an underlying spectral component, redder than the variable one, which could be identified either with the non-flaring part of the QSO spectrum or with the host galaxy considered by [@rom98]. We evaluate the effect of the host galaxy through numerical simulations based on templates of the QSO and host galaxy SEDs [@tre01b]. Both SEDs are derived from the atlas of normal QSO continuum spectra of [@elv94]. We compute a synthetic QSO+host spectrum, adding to the fixed host galaxy template SED the average QSO spectrum with a relative weight characterized by the parameter $\eta \equiv \log({L_H^Q}/{L_H^g})$, where ${L_H^Q}$ and ${L_H^g}$ are the total $H$ band luminosities of the QSO and the host galaxy respectively. In the [@elv94] sample, $\eta$ ranges from -1 to 2. Figure 5 shows an example of composite spectrum. We want to test (disprove) the hypothesis that the QSO SED maintains its shape during brightness changes, while the variation of the spectral shape is entirely due to the contribution of the (constant) galaxy SED. We know that the effect of the host galaxy should be small in the case of [@giv99] data since: i) magnitudes were computed using point spread function fitting of the images, ii) the images were limited by an aperture with the diameter depending on the seeing conditions, but smaller than 3 arcsec. However we don’t know the appropriate value of $\eta$. For this reason we perform simulations for a range of $\eta$ values. Variability is represented by small changes $\Delta \eta$, around each $\eta$ value, with an amplitude corresponding to a r.m.s. variability $\sigma_B = 0.16$ mag in the blue band. For each synthetic spectrum, representing the QSO plus host SED at a given time, we compute $\alpha(\bar{\nu},t) \equiv (\partial \log L_{\nu}/\partial\log\nu)_{\nu=\bar{\nu}}$, $\bar {\nu} = \sqrt {\nu_B \nu_R}$, then we derive the SVP $\beta$. The result, for $-3 < \eta < 3$, is shown in Figure 6. To check the dependence of the result on QSO redshift, the same computation is repeated for the maximum redshift of the [@giv99] sample, $z=0.4$, and shown in the same figure. Although an appropriate choice of $\eta$ can reproduce the observed $\beta$ or $\alpha$ separately, the curves are clearly shifted respect to the distribution of the observational points. This means that the effect of the host galaxy is not sufficient to account for the observed changes of the spectral shape. Thus the spectral variability is intrinsic of the active nucleus. This also implies that the constant red continuum, resulting from the analysis of [@cid00], cannot be identified with the host galaxy. It must be, at least in part, due to the nucleus, and can be identified with the spectrum of the non-flaring part of the accretion disk. CHANGES OF $\dot{M}$ AND “HOT SPOTS” ==================================== Once the spectral variability is ascribed to the nuclear component, we can ask whether a change of any of the parameters defining an emission model can account for the observed variations of the spectral shape . We considered the accretion disk model of [@sie95], [@fio95], corresponding to a Kerr metric and modified black body SED, which depend on the black hole mass $M$, the accretion rate $\dot{M}$ and the inclination $\theta$ ($\theta=0 \rightarrow$ face-on). A grid of models has been considered for $\log M/M_{\odot}=7.0,8.0,9.0,10.0$, $\dot{m} \equiv \dot{M} c^2/L_E=0.1, 0.3, 0.8$ (where $L_E$ is the Eddington luminosity $L_E= \frac{4 \pi G c m_p}{\sigma_e M}$ with the usual meaning of symbols) and $\mu \equiv \cos \theta = 1, 0.75, 0.5, 0.25, 0.1$. A change of $\dot{M}$ produces a variation of both luminosity and the SED shape. On purely theoretical grounds, we know that the time required for the accretion disk to reach a new equilibrium condition with a different $\dot{M}$ value is at least of the order of the sound crossing time $t_{sound}\approx 10^3-10^5$ d [@cou91], i.e. much longer than typical variability time scales. Still, it is interesting to see how the spectral changes between two $\dot{M}$ states compare with the observed ones, as done by [@tri94], and by [@sie95]. In Figure 6 the two curves on the bottom right represent $\beta$ versus $\alpha$ for varying $M$ (from right to left) as computed for two different values of $\dot{m}$ and for $\mu=1$, from the Kerr metric, modified black body model of [@sie95] (their table 4). The spectral variations are clearly smaller, on average, than the observed ones. This means that a transition e.g. from a lower to a higher $\dot{M}$ regime implies a larger luminosity change for a given slope variation. Notice that a black body of varying $T$ and fixed area provides larger $\beta$ values consistent with the observations. Thus, $\beta$ values computed for an increase of $\dot{m}$ would better compare with the case of a transition to a hotter disk with larger area. Observed slope variations of two objects, NGC 5548 and NGC 3783, have been compared with the predictions of an accretion disk model by [@tri94], in the UV range. They conclude that the observed points lie roughly on curves of constant black hole mass, giving confidence in the accretion disk models. However we notice that also in this case the distribution of observed points is steeper than the iso-mass lines, specially in the case of NGC 3783. Our result on the statistical sample of 42 objects indicates that this is indeed a systematic effect. This result suggests that transient phenomena, like hot spots produced on the accretion disk by instability phenomena [@kaw98], instead of a transition to a new equilibrium state, may better explain the relatively large changes of the local spectral slope, The available models of instability phenomena do not provide a spectrum of the hot spot. Thus we try a simple “model” based on the addition of a black body flare to the disk SED, represented by the average QSO SED of [@elv94] (shown in Figure 5). The free parameters are temperature $T_{BB}$, and emitting area $A$, while the constraints are the amplitude of the luminosity change (e.g. in the $B$ band) and the relevant $\beta$ value (or the relevant $\Delta \alpha$). The solution is not univocal, given the spread of the observed $\beta$ values. However, $\Delta B = 0.16$ mag (corresponding to the r.m.s. variability of the sample) can be obtained by a hot spot of $T_{BB} \approx 2\cdot 10^5$ K and $A=5 \cdot 10^{30}$ cm$^2$, producing $\beta= 3.2$, or $T_{BB} \approx 2\cdot 10^4$ K, $A=1,3 \cdot 10^{32}$ cm$^2$, giving $\beta= 2.2$. This is shown by the large filled squares in Figure 6. A sudden heating of a fraction of the disk surface is thus capable of producing the observed change of the SED in the $B$ and $R$ bands and the relevant intra-QSO $\alpha$-L correlation. SUMMARY AND CONCLUSIONS ======================= We have performed a new analysis of the $B$ and $R$ light curves of a sample of 42 PG QSOs made available by the Wise Observatory group [@giv99]. We have shown the existence of a positive v-z correlation when the analysis is restricted to a small luminosity bin. This correlation disappears in the sample as a whole due to the interplay of v-L, L-z and v-z correlations and the small redhift range. The slope of the v-z relation is consistent with the previous findings of [@dic96] and [@tre01], thus confirming that the v-z correlation can be entirely explained by the increase of variability with frequency, coupled with the increase of (rest-frame) observing frequency for higher redshift objects. The dependence of variability on redshift has been questioned in the past (see table 1 of [@giv99] for a summary of previous variability studies). We stress that: i) it is a relatively weak effect ($\partial v / \partial z \approx 0.1$) so that it cannot be detected unless other competing effects (including v-L and z-L correlations) are properly taken into account; ii) to get rid of spurious effects connected with cosmological time dilation and finite sampling time, variability must be quantified by an intrinsic index defined on the basis of the rest frame structure function. Once these prescriptions are adopted, the v-z correlations as measured in different optical-UV bands and different QSO samples appear consistent. This was already found by [@dic96] and [@tre01]. The present evidence, though marginal, is again quantitatively consistent and confirms previous results. The two-color light curves of the [@giv99] sample allow for the first time the statistical analysis of SED variability of individual QSOs and the study of the spectral slope changes among QSOs of different luminosity, leading to the evidence of an intra-QSO and an inter-QSO correlation. We have analyzed the spectral variability by the distribution of the parameter $\beta(\tau)\equiv \frac{\alpha(t+\tau)-\alpha(t)}{\log L_B(t+\tau)-\log L_B(t)}$ as a function of the spectral slope $\alpha$. We have compared with the observed distribution the spectral slope changes produced by the contribution of the host galaxy to the QSO SED, under the assumption that the nuclear spectrum maintains its shape while changing its brightness. We conclude that the host galaxy alone cannot be responsible for the observed spectral changes. Thus the spectral variation must be intrinsic of the nuclear component. The $\beta$-$\alpha$ distribution has been also compared with the SED changes of a disk model for $\dot{M}$ variations. The latter appear insufficient to explain the observed spectral changes. Bright spots on the disk, likely produced by local instabilities, are able to represent the observed spectral variability. Since it is likely that different physical phenomena are causing variability in different bands and time scales, multi-frequency analyses will be ultimately needed to obtain a complete description. However even a two optical bands analysis, once performed on a statistical sample, provides valuable constraints on the origin of variability. This strongly suggests to extend the work of the Wise Observatory group both in frequency and sampling time to allow a more detailed comparison with possible models. We are grateful to the Wise Observatory Group for promptly making available their data to the community, and for providing us with details about the observations. We are indebted to Fabrizio Fiore for his help in the use of disk models and for clarifying discussions. This research has made use 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. Alexander, T., 1995, , 274, 909 Angione, R. J., & Smith, H. J., 1972, in IAU Symp 44, External Galaxies and Quasi-Stellar Objects, ed. D. S. Evans, (New York: Springer-Verlag), p.171 Aretxaga, I., Cid Fernades, & Terlevich, R., 1997, , 286, 271 Bershady, M. A., Trèvese, D., & Kron, R. G. 1998, , 496, 103 Bonoli, F., Braccesi, A., Federici, L., & Zitelli, V., 1979, , 35, 391 Cid Fernandes, R., Sodrè, L., Jr. Vieira da Silva, L., Jr., 2000 , 544, 123 Courvoisier, T. J.-L., 1991, Clavel, J., , 248, 389 Courvoisier, T. J.-L., 1998, , 9, 1 Cox, A. N., 2000 “Allen’s Astrophysical Quantities” (Berlin: Springer-Verlag) Cristiani, S., Vio, R., & Andreani, P. 1990, , 100, 56 Cristiani, S., Trentini, S., La Franca F., Aretxaga, I., Andreani, P, Vio, R., & Gemmo A. 1996, , 306, 395 Cutri, R. M., Wisniewski, W. Z., Rieke, G. H., & Lebofski, H. J. 1985, , 296, 423 Di Clemente A., Natali G., Giallongo E., Trévese D., Vagnetti F. 1996, , 463, 466 Edelson, R.A., Krolik, J. H.,& Pike, G. F. 1990, , 359, 86 Elvis, M., Wilkes, B. J., McDowell, J. C., Green, R. F., Bechtold, J., Willner, S. P., Oey, M. S., Polomski, E., Cutri, R., 1994, , 95, 1 Fiore, F., Elvis, M., Siemiginowska, A., Wilkes, B.J., McDowell, J.C., & Mathur, S.: 1995, , 449, 74 Giallongo E., Trèvese D., Vagnetti F.: 1991, , 377, 345 Giveon, U., Maoz, D., Kaspi, S., Netzer, H., & Smith P. S. 1999, , 306, 637 Kawaguchi, T., Mineshige, S., Umemura, M., & Turner, E. L. 1998, , 504, 671 Kinney, A. L., Bohlin, R.C., Blades, J. C., & York, D. G. 1991, , 75, 645 Hawkins, M. R. S., 1983, , 202, 571 Hawkins, M. R. S., 1996, , 278, 787 Hawkins, M. R. S., 2001, , 553, L97 Hook, I. M.,McMahon R. G., Boyle, B. J., & Irwin, M. J. 1994, , 268, 305 Koo, D.C., Kron, R.G., Cudworth K.M.: 1986, , 98, 285 Kaspi, S., Smith, P. S., Netzer, H., Maoz, D., Jannuzi, B. T., & Giveon, U., 2000, , 533, 631 Paltani, S., & Courvoisier, T. J.-L. 1994, , 291, 74 Romano, P., & Peterson, B. M., 1998, ASP Conf. Ser. 175, 55 Siemiginowska, A. Kuhn, O., Elvis, M., Fiore, F., McDowell, J., & Wilkes, B., 1995, , 454, 77 Torricelli-Ciamponi, G., Foellmi, C., Courvoisier, T. J.-L., & Paltani, S., 2000, , 358, 57 Trèvese, D., Kron, R. G., & Bunone A., 2001, , 551, 103 Trèvese, D., Kron, R. G., Majewski, S. R., Bershady, M. A., Koo, D. C. 1994, , 433, 494 Trèvese D., Pittella G., Kron R. G., Koo D. C., & Bershady M. A. 1989, , 98, 108 Trèvese, D., & Vagnetti, F., 2001a in ’Guillermo Haro Advanced Lectures on the Starburst-AGN Connection’, R. Mujica et al. Eds., INAOE Electronic Edition (http://www.inaoep.ms/$\sim$agn00/posters.html), pp. 101-103 Trèvese, D., & Vagnetti, F., 2001b, in ’QSO Hosts and their Environments’, I. Marquez et al. Eds., in press (astro-ph/0102252) Tripp, T. M., Bechtold, J., & Green, R. F., 1994, , 433, 533 Ulrich, M. H., Maraschi L., & Urry C. M. 1997, , 35, 445

--- abstract: 'It is believed that planets are formed by aggregation of dust particles suspended in the turbulent gas forming accretion disks around developing stars. We describe a mechanism, termed Stokes trapping’, by which turbulence limits the growth of aggregates of dust particles, so that their Stokes number (defined as the ratio of the damping time of the particles to the Kolmogorov dissipation timescale) remains close to unity. We discuss possible mechanisms for avoiding this barrier to further growth. None of these is found to be satisfactory and we introduce a new theory which does not involve the growth of small clusters of dust grains.' author: - 'M. Wilkinson$^{1}$, B. Mehlig$^{2}$ and V. Uski$^{1}$' title: Stokes trapping and planet formation --- Background {#sec: 1} ========== It is widely believed that planets are formed by aggregation of dust particles in an accretion disk surrounding a growing star; the fact that solar planets have orbits which are roughly circular and coplanar with the Sun’s equator is readily explained by this model. According to this picture, planet formation is a two-stage process [@Saf69; @Arm07]. The final stage must be driven by gravitational forces, but initially the density of dust particles is not sufficiently high for gravitational instability to overcome centrifugal forces. It is believed that the dust particles initially aggregate by random collisions, and that when the aggregates have grown to a sufficient size, they sink to the mid-plane of the accretion disk. Gravitational collapse to form planets can commence if the density at the mid-plane of the disk becomes sufficiently high. In this paper we are concerned with the first (kinetic’) phase of planet formation in this standard model. The kinetic aggregation process in protoplanetary disks (reviewed by @Dom07 [@Hen06; @Arm07]) has distinctive features. The dust particles are surmised to be in a highly turbulent environment (at least during periods when the star is accreting material at a substantial rate), because laminar flows could not dissipate energy sufficiently rapidly to account for observed accretion rates. Unless some mechanism enables the particles to fuse together chemically, they must form very weak aggregates (bound by van der Waals and possibly also electrostatic forces) until gravitational collapse is well advanced: this implies that the aggregates are very fragile unless they can be heated to high temperatures. Also, the aggregates of dust particles may be fractal structures, with a fractal dimension $D$. The principal purpose of this paper is to argue that the standard model for planet formation by aggregation of dust grains is highly problematic, because the dust grains are unable to aggregate in the turbulent environment of the accretion disc. Furthermore, we argue that the turbulence would prevent even very large aggregates from settling to the mid-plane of the disc. The difficulties are so severe that it is necessary to develop an alternative theory. A note about terminology is in place here. We describe the condensed material in the interstellar medium as consisting of dust [*grains*]{}, which may form [*aggregates*]{}, bound by van der Waals or electrostatic forces. We argue that these aggregates may [*cluster*]{} together in space, without coming into contact. We also describe properties of small fragments of solid material in a turbulent gas in a more general context, without needing to specify whether they are dust grains or aggregates: in such cases we refer to [*particles*]{} suspended in the gas. Summary of results {#sec: 2} ================== This paper uses recently acquired insights [@Wil06; @Fal02; @Wil05] into the dynamics of turbulent aerosols to draw conclusions about the aggregation of dust particles in a standard model for a protostellar accretion system. An important parameter is the Stokes number, ${\rm St}$, which is the ratio of the time scale for an aggregate of dust grains to be slowed by drag forces (the stopping time’) to the dissipative correlation time of the turbulent forcing. Note that our definition of the Stokes number differs from that used in some other works on planet formation (for example, [@You03; @Bau07; @You07]), which define the Stokes number as the ratio of the stopping time to the orbital period, although our definition is the one which is used in the fluid dynamics literature, for example in [@Fal02; @Dun05; @Bec06]. Our principal conclusions may be summarised as follows. We argue that the relative velocity of the dust aggregates, and therefore their rate of collision, suddenly increases by several orders of magnitude when aggregates grow to a size such that their Stokes number is of order unity. One consequence of this effect is the formation of a Stokes trap’: aggregates with Stokes numbers significantly larger than unity will collide with sufficiently high relative velocity that they fragment upon collision, thereby replenishing the population of aggregates with smaller Stokes number. This process would result is a stable population of aggregates with a Stokes number which has a distribution spanning ${\rm St}\approx 1$. The production of planets therefore depends upon the particle aggregation process being able to escape from this Stokes trap. We consider four scenarios by which escape from the Stokes trap might happen: [*1. Clustering.*]{} Particles in a turbulent flow with Stokes numbers of order unity have been shown to cluster together. This clustering effect has been ascribed to particles being centrifuged’ away from vortices [@Max87]. More recent work has shown that the particles cluster onto a fractal set [@Som93; @Bec06; @Dun05] of dimension $D_{\rm cl}$. Note that there are two fractal dimensions in this problem: the dust particle aggregates may be fractals of dimension $D$, and these aggregates cluster onto a set of dimension $D_{\rm cl}$. It is conceivable that this clustering effect might create a means of escaping the Stokes trap, but we have not been able to identify a mechanism which would be sufficiently effective in the model which we have considered. [*2. Temperature variations.*]{} The temperature of the dust aggregates could vary dramatically as they are swept back and forth between the surface and the mid plane of the accretion disk by turbulent motion of the gas. High temperatures can facilitate chemical reactions, as can the condensation of water and other volatile materials when the dust aggregate encounters low temperature regions. Chemical processes could cause the dust aggregates to become sufficiently tightly bound that that are less easily fragmented in collisions. Also, condensation tends to occur more readily in corners, favouring the formation of compact aggregates. Below we argue that the accretion disk is sufficiently optically thick at the crucial stage of its development that there can be a marked difference between the surface and mid-plane temperatures. However, unless the star is accreting at a high rate, at distances significantly greater than $1\,{\rm AU}$ the temperatures are low everywhere across the profile of the disk. We conclude that chemical processes are probably not significant unless accretion is occurring at a very high rate. [*3. Fractal aggregates.*]{} The Stokes trap can also be avoided if the dust particles form aggregates with a fractal dimension $D$ less than two: in this case we find that the Stokes number does not increase as the size of the aggregate increases. Some experiments on dust grains [@Blu00; @Kra04] have shown evidence that they form aggregates with a fractal dimension approximately equal to $1.4$, others [@Wur98] obtained values close to $1.9$. It should be noted that these experiments only observe the dust particle aggregates for a short time after their creation, whereas in protoplanetary processes the aggregates have plenty of time to relax to a more compact state, which would be energetically favoured due to increased contact area. The structure of dust aggregates in real protoplanetary systems is therefore highly uncertain. The hypothesis that the dust aggregates are tenuous structures with $D\le 2$ does not appear to provide a satisfactory resolution, because these tenuous aggregates will always be advected with the gas, and will not sink to the midplane. We also show that tenuous aggregates may be torn apart by shearing forces. [*4. Quiescence.*]{} It might be expected that the relative velocities of the dust particle aggregates are reduced if the rate of accretion is reduced. We write the rate of accretion onto the star as $\dot M=10^{-7}\Lambda M{\rm yr}^{-1}$, where $M$ is the solar mass, and the coefficient is chosen so that $\Lambda=1$ corresponds to a typical rate of accretion [@Arm07]. We investigate the scaling of the relative velocity of suspended particles with the scaling parameter $\Lambda$, and find that it increases as $\Lambda$ decreases. We conclude that the Stokes trap is still present in relatively quiescent disks, where $\Lambda$ is small. Our calculations are based upon estimates using Kolmogorov’s theory for turbulence [@Fri97] in combination with a standard model for an accretion disk [@Sha73]. In our application of the Kolmogorov theory we assume that the size scale of the largest eddies, $L$ is proportional to the scale height of the disk, $H$, writing $L=\ell H$ (we show that this is equivalent to the $\alpha$-prescription’ [@Sha73]). We also consider how the results depend upon the additional scaling parameter $\ell$. Because of the large uncertainties in the parameters of the model, our calculations are to be interpreted as illustrations of the Stokes trapping principle, rather than as quantitative predictions. Order of magnitude changes in the uncertain parameters will not alter our principle conclusion about the effectiveness of the Stokes trap in limiting dust aggregation. For simplicity we confine attention to the case where the heating of the disk is active (that is dominated by dissipation in the disk rather than by direct illumination from the star). Our calculations lead to the estimate that typical particle sizes of Stokes-trapped aggregates at 10 AU from the star are of the order of $25\mu{\rm m}$. In principle it is possible to gain information about the sizes of particles suspended in accretion disks around young stars from spectroscopic studies, but this is an ill-conditioned problem [@Car04], with uncertain spectroscopic data. Studies of the silicate feature’ (at approximately $10\mu{\rm m}$ wavelength) in a survey of spectra of accretion disks around young stars by @Kes07 are interpreted as indicating typical particle sizes between $3$ and $10\mu$m, between $10^{-3}$ AU and $10$AU from the star. Similar estimates were obtained by @vBo03 and @Apa05. By contrast, spectral analysis in the mm-range (probing predominantly the outer regions of the disk) has been interpreted as indicating that particles larger than $50\mu{\rm m}$ may be present [@Woo02; @Bau07]. Even if the particle growth process can circumvent the Stokes trap, the particle aggregates must collapse to the mid-plane of the disk for gravitational instability to become effective. If we make the assumption that the integral scale of the turbulence is comparable to the thickness of the accretion disk (that is, if $\ell$ is of order one), we show that advection of particles by the largest turbulent eddies presents another very significant barrier to planet formation: turbulence will prevent aggregates of centimetre size from settling to the mid-plane, unless the rate of accretion is very substantially smaller than $10^{-7}M{\rm yr}^{-1}$. We conclude that there is no satisfactory theory for the formation of planets by aggregation of sub-micron sized dust grains. An alternative theory is therefore required. In the concluding section we introduce an alternative hypothesis. Collision processes in turbulent aerosols {#sec: 3} ========================================= Throughout we assume that the drag force on a particle is proportional to the difference in velocity between the particle and the surrounding gas, so that the equation of motion is $$\label{eq: 3.1} \ddot{\mbox{\boldmath$r$}}=\gamma (\mbox{\boldmath$u$}(\mbox{\boldmath$r$},t)-\dot{\mbox{\boldmath$r$}})$$ where $\mbox{\boldmath$r$}$ is the position of the particle and $\mbox{\boldmath$u$}(\mbox{\boldmath$r$},t)$ is the fluid velocity field (until the particles come into contact). This equation is familiar in the context of Stokes’s law for the drag on a sphere, where the damping rate $\gamma$ is proportional to the kinematic viscosity $\nu$. We assume that the gas has a multi-scale turbulent flow, which is described by the Kolmogorov theory of turbulence [@Fri97]. The length scale and time scale associated with the smallest eddies where energy is dissipated are denoted by $\eta$ and $\tau$ respectively. According to the Kolmogorov theory of turbulence, these quantities are determined by the rate of dissipation per unit mass, ${\cal E}$, and the kinematic viscosity $\nu $: $$\label{eq: 3.2} \eta = \biggl(\frac{\nu^3}{{\cal E}}\biggr)^{1/4} \ ,\ \ \ \tau = \biggl(\frac{\nu}{{\cal E}}\biggr)^{1/2}\ .$$ If the turbulent motion is driven by forces acting on a length scale $L\gg \eta$, the velocity fluctuations of the fluid have a power-law spectrum for wavenumbers between $1/L$ and $1/\eta$ [@Fri97]. We define the Stokes number to be ${\rm St}=1/(\gamma\tau)$. Very small particles are advected with the gas flow, but in addition they have small random velocity due to Brownian diffusion, which can cause them to collide at a rate ${\cal R}_{\rm d}$. They can also collide due to the shearing motion of the flow, an effect discussed by @Saf56 in the context of rainfall from turbulent clouds. These advective collisions occur at a rate ${\cal R}_{\rm a}$ which turns out to be negligible compared to Brownian diffusion in the present context (see tables \[tab:2\] and \[tab:3\]). ![\[fig: 1\] Illustration of fold caustics. Shown are results from numerical simulations in one spatial dimension of the equation of motion $\ddot x_i = \gamma (u(x_i,t)-\dot x_i)$, where $u(x,t)$ is a random velocity field which is a smooth function of $x$ and $t$. Here $x_i$ denotes the position of particle number $i$. The top panel shows particle trajectories $x_i(t)$ as a function of time $t$. The bottom panels show a different representation of the dynamics: the particle velocities $\dot x=v(x,t)$ versus their positions $x$ at three different times. The velocity $v(x,t)$ of the suspended particles is initially single-valued, but becomes triple-valued in the region between two fold caustics, which are singularities of the projection of the phase-space manifold onto coordinate space. In the region between the caustics the suspended particles have a relative velocity and their trajectories cross.](f1.eps){width="14cm"} When the Stokes number approaches unity, there is a dramatic increase in the relative velocity of suspended particles due to the formation of fold caustics in their velocity field, illustrated for the case of one spatial dimension in figure \[fig: 1\]. When faster particles overtake slower ones, the manifold representing the phase-space distribution of the particles develops folds, and the velocity field of the particles goes from being single valued to multi-valued (three-valued, in this illustration). Before the caustics form the relative velocity of the particles is due to their Brownian diffusion. After the folds have formed the relative velocity may be orders of magnitude higher. The rate of collision ${\cal R}$ for a suspended particle in a turbulent flow is approximated by a formula presented by @Meh07: $$\label{eq: 3.3} {\cal R}={\cal R}_{\rm d}+{\cal R}_{\rm a}+\exp(-A/{\rm St}){\cal R}_{\rm g}\ .$$ Here ${\cal R}_{\rm g}$ is the collision rate predicted by a model introduced by @Abr75, often termed the gas-kinetic model’, in which the suspended particles move with velocities which become uncorrelated with each other and with the gas flow. A less precise formula was suggested by @Fal02, and the idea that there is a dramatic change in the relative velocity of particles at ${\rm St}\approx 1$ appears to have been originally proposed by @Mar91. The exponential term in (\[eq: 3.3\]) describes the fraction of the coordinate space for which the velocity field is multi-valued: $A$ is a universal’ dimensionless constant. The non-analytical dependence of the rate of caustic production on ${\rm St}$ was noted by @Wil05 and recent simulations of Navier-Stokes turbulence suggest that $A\approx 2$ [@Pum06]. The rate ${\cal R}_{\rm g}$ greatly exceeds ${\cal R}_{\rm d}$, but the gas-kinetic theory is only applicable when the velocity field of the suspended particles is multi-valued. The exponential term arises because the formation of caustics is determined by a process involving escape from an attractor by a diffusion process, similar to the Kramers model for a chemical reaction. The exponential term is therefore analogous to the Arrhenius term $\exp(-E/k_{\rm B}T)$ in the expression for the rate of an activated chemical reaction [@Wil05]. The abrupt increase of the collision rate as the Stokes number exceeds a threshold value was first noted in numerical experiments by @Sun97. At large Stokes number the rate of collision of particles of radius $a$ is $$\label{eq: 3.4} {\cal R}_{\rm g}=4\pi a^2 n \langle \Delta v \rangle$$ where $n$ is the number density of particles and $\Delta v $ is the relative speed of two suspended particles at the same position in space (and angular brackets denote averages throughout). In a multi scale turbulent flow, when $\gamma \tau\ll 1$ the motion of the suspended particles is underdamped relative to the motion on the dissipative scale, but it is overdamped relative to slower long-wavelength motions in the fluid. The relative velocity of two nearby particles is a result of the different histories of the particles. If we follow the particles far back in time to when they had a large separation, their velocities were very different, but these velocity differences are damped out when the particles approach each other. @Meh07 surmised the variance of the relative velocities using the Kolmogorov scaling principle. In the following we consider, for simplicity, only the case of a symmetric collision, where both particles have the same damping rate, $\gamma$ (in the case of collisions between particles of very unequal sizes, we may still use the estimate below, using the smaller of the two values of $\gamma$). When the particles are underdamped relative to the smallest dissipative scale, but overdamped relative to the integral’ (driving) scale, we can apply the Kolmogorov cascade principle [@Fri97], that motion in the inertial range is independent of the mechanism of dissipation. It is therefore determined by the rate of dissipation ${\cal E}$ but it does not depend on $\nu$. The moments of the relative velocity therefore depend only upon ${\cal E}$ and $\gamma$. For the second moment, dimensional considerations imply that [@Meh07] $$\label{eq: 3.5} \langle \Delta v^2\rangle =K\frac{{\cal E}}{\gamma}$$ where Kolmogorov’s 1941 theory of turbulence suggests that $K$ a universal constant, but in practice $K$ should have a weak dependence upon Reynolds number due to intermittency effects [@Fri97]. The constant $K$ can be determined by simulation of Navier-Stokes turbulence, but this has not yet been done. We use $K=1$ in the remainder of this article. @Vol80 (see also [@Miz88], @Mar91) have discussed relative velocities in the large Stokes number limit, expressing their results in terms of properties of the integral velocity and timescales: their results are equivalent to (\[eq: 3.5\]), but are harder to apply because the precise form of the integral scale motion is highly uncertain, whereas the rate of dissipation in (\[eq: 3.5\]) is determined directly from the accretion rate and the mass per unit area in the accretion disk. Properties of dust-particle aggregates {#sec: 4} ====================================== Damping rate {#sec: 4.1} ------------ During the kinetic phase of planet formation, the size $a$ of the suspended particles is usually smaller than the mean free path of the gas, $\lambda$. In this case the drag force is proportional to $\rho_{\rm g}c_{\rm s}\bar A ({\mbox{\boldmath$u$}}-\dot{ \mbox{\boldmath$r$}})$, where $\rho_{\rm g}$ is the gas density, $\bar A$ is the angular average of the projected area of the particle and $c_{\rm s}$ is the mean molecular speed of the gas [@Eps24]. In this case the damping rate is $$\label{eq: 4.1} \gamma\sim \frac{\rho_{\rm g}c_{\rm s}\bar A}{m}$$ where $m$ is the mass of the particle. In the case of spherical particles, the coefficient in eq. (\[eq: 4.1\]) is $3/4$ [@Eps24]. Structure of aggregates {#sec: 4.2} ----------------------- The suspended particles are expected to consist of dust particles, of typical size $a_0$, which are weakly bound together by electrostatic and van der Waals interactions. These aggregates may have a fractal structure, such that the number of particles in an aggregate of linear dimension $a$ is $$\label{eq: 4.2} N\sim \biggl(\frac{a}{a_0}\biggr)^D$$ where $D$ is a fractal dimension. The damping rate for an aggregate of $N$ dust grains with size $a_0$ composed of material of density $\rho_{\rm p}$ is therefore estimated as $$\label{eq: 4.3} \gamma \sim \frac{\rho_{\rm g}c_{\rm s}}{\rho_{\rm p}a_0}N^{\frac{2-D}{D}}=\gamma_0\left(\frac{a}{a_0}\right)^{2-D}\ .$$ Since ${\rm St} = 1/(\gamma \tau)$ we have ${\rm St} \propto (a/a_0)^{D-2}$ and conclude that if $D\le 2$, the Stokes number does not increase as the mass of the dust aggregate increases. In the early stages of aggregation, we expect that the mean free path of the suspended particles is very large compared to their size, and the relevant fractal growth mechanism is ballistic aggregation (where particles incident with random linear trajectories adhere to the aggregate on contact) rather than diffusion-limited aggregation (where an incoming particle executes a random walk). Numerical experiments [@Nak94] on ballistic limited aggregation suggest $D\approx 2.98$ for a ballistic particle cluster aggregate’ (BPCA) model (particles adhering to a stationary aggregate) and $D\approx 1.93$ for a ballistic cluster cluster aggregate’ (BCCA) model (where aggregates of particles move and collide with each other). (Note that the use of the term cluster’ in the designations of the BPCA and BCCA differs from that used in this paper, as explained at the end of section \[sec: 1\]). However, experiments on agitated dust grains [@Blu00; @Kra04] have shown evidence of aggregates with a lower fractal dimension, $D\approx 1.4$. The lower fractal dimension observed in these experiments may be related to electrostatic effects: for example if agitation of dust particles causes them to acquire dipolar charges, they will tend to form strings (dimension $D=1$), as illustrated in figure \[fig: 2\]. The timescales on which these aggregates are observed is very short compared to the astrophysical context which we consider (and also the number of particles in the aggregates are not very large). The appropriate value for $D$ is, therefore, highly uncertain. It is probably not universal throughout the accretion disk structure. Our interest is in aggregates which undergo energetic collisions (which are sufficient to cause fragmentation in some cases), and which have a long time to relax into a compact configuration with lowest energy. In the following we therefore usually assume that the aggregates are predominantly in compact configurations, with $D\approx 3$, but in many cases below we continue to give formulae for general values of $D$. ![\[fig: 2\] Dust grains in a low gravity environment are found to form aggregates with a low fractal dimension. This may be due to particles having an electric dipole, causing them to form chains.](f2.eps){width="8cm"} Strength of aggregates {#sec: 4.3} ---------------------- The other distinctive aspect of the dust-grain aggregates concerns their fragility under collisions. There is a substantial literature on this topic: see @Dom97, @You03 and references therein. In the following we introduce a new treatment, valid in the limit where the grains are very small. We mentioned earlier that the aggregates are bound together by van der Waals forces, which means that they must be very fragile. For a collision between two aggregates of size $a_1$ and $a_2$, there will be a velocity $v_{\rm cr}$ above which the average change in the mass of the larger particle is negative. It is extremely hard to estimate this critical velocity for collisions of aggregates. We estimate this critical velocity below for the collision of two grains. We will assume that the critical velocity is approximately independent of size, so that the two-grain estimate is sufficient for the general case. In the following we therefore use $v_{\rm cr}$ to refer to the mean value of the relative velocity below which [*pairs*]{} of grains remain bound. We motivate this assumption by the following argument. When two aggregates collide, initially the collision only influences the grains on the surface of the aggregates in the vicinity of the point of collision. If the aggregates collide with a relative velocity $v$, some of the grains will acquire relative velocities which are comparable to $v$. If $v/v_{\rm cr}$ is significantly greater than unity, grains on the periphery of the colliding aggregates will be able to escape. In order to estimate the critical relative collision velocity above which fragmentation occurs we therefore estimate the critical velocity $v_{\rm cr}$ above which two dust grains do not remain bound after a binary collision. This will be estimated by equating the kinetic energy of the collision process between two dust grains, treated as spheres of radius $a_0$, to the binding energy. The binding energy is usually estimated by a theory proposed by @Cho93 (and discussed in simpler terms by @You03), which assumes that the spheres deform so that they are in contact on a small circular patch. By balancing the force required to deform the spheres against a derivative of the surface energy released by bringing them into contact, the equilibrium deformation and hence the binding energy are estimated in terms of a Young’s modulus for the spheres, $Y$, and their surface energy of attraction, $\sigma$. Omitting the dimensionless prefactors (which depend upon making more precise definitions of $Y$, $\sigma$) the surface deformation $\delta$ and binding energy $E_{\rm b}$ are estimated to be $$\label{eq: 4.4} \delta\sim \left(\frac{\sigma^2a_0}{Y^2}\right)^{1/3} \ ,\ \ \ E_{\rm b}\sim \left(\frac{\sigma^5a_0^4}{Y^2}\right)^{1/3}\ .$$ The latter expression gives an estimate for the critical velocity $v_{\rm cr}$, obtained in @Cho93: $$\label{eq: 4.5} v_{\rm cr}\sim\left(\frac{\sigma^5}{\rho_{\rm p}^3Y^2a_0^5}\right)^{1/6}\ .$$ Using values for quartz of $Y=5.4\times 10^{10}\,{\rm N\,m}^{-2}$, $\sigma=2.5\times 10^{-2}\,{\rm N\,m}^{-1}$, $\rho_{\rm p}=2.6\times 10^3\,{\rm kg\,m}^{-3}$ quoted in [@Cho93], we estimate $\delta\approx 3\times 10^{-11}{\rm m}$ and $v_{\rm cr}\approx 0.2\,{\rm m\,s}^{-1}$ for grains of size $a_0=10^{-7}{\rm m}$. The validity of this approach is questionable when the value of the deformation is small compared to the size of an atom, as is found in this example. We therefore propose an alternative estimate of the binding energy. We assume that there is a van der Waals binding energy $\delta E_1$ between atoms, which operates over a range $l$, which we equate with the typical interatomic distance. We assume that there is a binding energy $\delta E_1/l^2$ per unit area for all parts of two spheres which are within a distance $2l\ll a_0$ of each other, where $\delta E$ is the van der Waals binding energy between two atoms, and $l$ is the range over which it acts. Two spheres in contact have a surface area $\pi l a_0$ with separation less than $2l$ (see figure \[fig:4\]), so that the binding energy is $E_{\rm b}\sim \pi \delta E a_0/l$. Unless the grains are made of very plastic material, particles are not expected to aggregate if their (centre of mass frame) kinetic energy is large compared to the binding energy. Writing $E_{\rm b}\sim m v_{\rm cr}^2\sim \rho_{\rm p}a_0^3 v_{\rm cr}^2$, we estimate a critical velocity $$\label{eq: 4.6} v_{\rm cr}=\sqrt{\delta E_1 /\rho_{\rm p}l a_0^2}\ .$$ Taking $\delta E_1\approx 10^{-21}{\rm J}$, $l=3\times 10^{-10}{\rm m}$ and $a_0=10^{-7}{\rm m}$, we find that $v_{\rm cr}\approx 3\times 10^{-1}{\rm m\,s}^{-1}$ for binding of microscopic grains. Thus collisions between aggregates with relative velocities of the order of $30\,{\rm cms}^{-1}$ have the potential to dislodge particles from an aggregate. It is instructive to compare the microscopic and the macroscopic models, equations (\[eq: 4.5\]) and (\[eq: 4.6\]). The surface energy and Young’s modulus can be related to microscopic parameters by writing $\sigma\sim \delta E_1/l^2$, where $\delta E_1\approx 10^{-21}\,{\rm J}$ is the van der Waals interaction energy and $Y\sim \delta E_2/l^3$, where $\delta E_2\approx 10^{-18}\,{\rm J}$ is the characteristic energy scale for covalent bonds: these expressions give values which are consistent with the values for $\sigma $ and $Y$ for quartz quoted above. Expressing equation (\[eq: 4.5\]) in terms of microscopic quantities gives $$\label{eq: 4.7} v_{\rm cr}\sim \sqrt{\delta E_1 /\rho_{\rm p}l a_0^2}\left(\frac{\delta E_1}{\delta E_2}\right)^{1/3}\left(\frac{a_0}{l}\right)^{1/6}\ .$$ We see that the two estimates (\[eq: 4.5\]) and (\[eq: 4.6\]) differ by dimensionless ratios with small exponents, and the values obtained from the two formulae are therefore typically quite similar. Whichever of equation (\[eq: 4.6\]) or (\[eq: 4.5\]) gives the larger value should be used to estimate the critical velocity. The material parameters can make a significant difference, in particular ice has a much larger surface energy and is softer than quartz, and the predicted critical velocity is approximately an order of magnitude larger for ice [@Cho93]. There are significant sources of uncertainty in these estimates for the critical velocity for binding single grains, which are very hard to quantify. One is the effect of surface roughness. This will reduce the effective contact area between two grains, and therefore reduce $v_{\rm cr}$. The effect of electrostatic charging of dust grains and aggregates is harder to quantify. It is possible that the microscopic grains will all tend to acquire a positive charge, due to ionisation by energetic photons. This would reduce the binding energy. Another possibility is that grains acquire random charges through frictional charge transfer in collisions. This may increase the critical velocity for pairs of grains, because of electrostatic attraction of oppositely charged particles. This effect will be reduced by cancellation of charges for larger aggregates. Experiments to determine the critical velocity for collisions of grains have been performed by projecting small silica spheres at a glass surface [@Pop97]. These experiments have usually showed that the spheres adhere to the surface at velocities which are approximately ten times higher than those predicted by equation (\[eq: 4.5\]). The reasons for this discrepancy are not understood. Although the effects of charges on the particles were quantified in these experiments, it is possible that effects of electrostatic charges on the surface might account for the discrepancy. [![\[fig:4\] Two spheres in contact have a surface area $\pi l a_0$ with separation less than $2l$.](f3.eps "fig:"){width="5cm"}]{} Turbulence in protoplanetary disks {#sec: 5} ================================== We now consider estimates relating to dust particles suspended in gas surrounding a growing star. Observational evidence suggests that there is usually sufficient angular momentum that the gas forms an accretion disk and we use the steady-state disk model described by @Sha73. The kinetic and potential energy of the material in this cloud must be dissipated into heat and radiated away in order to allow the material to fall into the growing star: we consider the case where the disk is primarily heated by dissipation, rather than by radiation from the star. It is known that stars can form quite rapidly (over a timescale of $t_{\rm acc}\approx 10^6{\rm yr}$). The process is so fast that either shocks or turbulent processes must play a role in the dissipation of energy: the rate of dissipation by laminar flow is too small by many orders of magnitude. It has been argued that magneto-hydrodynamic mechanisms provide the necessary large-scale instability [@Arm07], but the gas is only weakly ionised and the small scale dissipation is expected to be described by conventional hydrodynamics, for which the Kolmogorov theory of turbulence is applicable. Shocks dissipate a finite fraction of the kinetic energy of a gas almost instantaneously, so that their role must be short lived and throughout most of the accretion process turbulence is the dissipation mechanism. In the following we consider a star of mass $M$ (taken to be one solar mass) surrounded by a protoplanetary nebula, which is collapsing into the star at a rate $\dot M$. We assume $$\label{eq: 5.0} \dot M=10^{-7}\Lambda M{\rm yr}^{-1}\ .$$ Most discussions of protoplanetary accretion systems assume that $\Lambda \approx 1$. We assume that the accretion zone is predominantly molecular hydrogen (we take the mean molecular mass of the gas molecules $m_{\rm g}=\mu m_{\rm H}$, with $\mu=2.34$ and $m_{\rm H}=1.67\times 10^{-27}{\rm kg}$ being the mass of atomic hydrogen). For estimating the mean-free path of the gas, we assume that the molecular collision cross section is $S_{\rm col}=2\times 10^{-19}{\rm m}^2$. These and other assumed parameter values are collected in table \[tab:tab0\]. We use $$\label{eq: 5.1}c_{\rm s}^2=k_{\rm B}T/m_{\rm g}$$ to relate $c_{\rm s}$, a characteristic speed which is of the order of the speed of sound, to temperature. [lcc]{} Mass of gas molecules &$m_{\rm g} = \mu m_{\rm H}$ &$ \mu = 2.34 $ Molecular collision cross section &$S_{\rm col}$& $2\times 10^{-19}{\rm m}^2$ Mass of the star &$M$ & $1.99 \times 10^{30}{\rm kg}$ Typical size of interstellar dust grains & $a_0$& $10^{-7}{\rm m}$ Density of suspended particles & $\rho_{\rm p}$& $2\times 10^3{\rm kg}/{\rm m}^3$Mass fraction of suspended particles to gas molecules&$\kappa$& $3/170$ The accretion disk is characterised by its mass density per unit area $\Sigma(R)$ at a distance $R$ from the star, its scale height $H(R)$, its rate of dissipation per unit area $Q(R)$, and its surface temperature $T(R)$. Using the continuity equations for flow of mass and angular momentum within a thin accretion disk which is in a quasi-stationary state, the rate of dissipation per unit area $Q$ at radius $R$ can be obtained in terms of the rate of accretion, $\dot M$ [@Sha73]: $$\label{eq: 5.2} Q(R)=\frac{3}{8\pi}\frac{GM\dot M} {R^3}\biggl[1-\biggl(\frac{R_{\rm c}}{R}\biggr)^{1/2}\biggr]\,.$$ Here $R_{\rm c}$ is the radius of the core of the accretion system. In the following we concentrate on the region $R/R_{\rm c}\gg 1$, and do not carry forward the final factor of the above expression. Note that this formula is independent of the actual mechanism of dissipation. The rate of dissipation per unit mass is obtained from the mass per unit area $\Sigma(R)$ of the accretion disk at radius $R$, and the density of gas $\rho_g(R)$ at the mid-plane of the disk is related to the scale height: $$\label{eq: 5.3} {\cal E}(R)=\frac{Q(R)}{\Sigma(R)}\ ,\ \ \ \rho_{\rm g}(R)=\frac{\Sigma(R)}{\sqrt{2\pi}H(R)}$$ where in the second expression we assume a Gaussian density profile with variance $H$. Consider next how to estimate the height of the disk, $H(R)$. The gas is in Maxwell-Boltzmann equilibrium in the gravitational potential and we assume that the contribution to the potential from the disk itself is negligible. The potential energy at a distance $z$ from the mid plane is $\Phi(z,R)=GMm_{\rm g}z^2/2R^3$, so that the gas density at distance $z$ from the mid plane of the disk is $\rho(z,R)=\rho_{\rm g}(R)\exp[-z^2/2H^2(R)]$, the scale height of the disk being $$\label{eq: 5.4} H(R)=\biggl({kT(R)R^3\over{GMm_{\rm g}}}\biggr)^{1/2}=c_{\rm s}(R)/\Omega(R)$$ where $\Omega(R)=\sqrt{GM/R^3}$ is the Keplerian orbital angular frequency for a mass $M$ at radius $R$. We assume that the surface temperature $T(R)$ is determined by the rate at which the thermal energy created by dissipation of turbulent motion can be radiated away. This assumption is justifiable as long as $\Lambda$ in equation (\[eq: 5.0\]) is not too small, but when $\Lambda\ll 1$ the rate of heating by radiation from the star is expected to be significant [@Arm07]. Assuming that the surface of the disk behaves as a black body with emissivity $\varepsilon$, one obtains $$\label{eq: 5.5} Q(R)=2\varepsilon \sigma T^4(R)\sim {3\over{8\pi}}{GM\dot M\over{R^3}}$$ where $\sigma$ is the Stefan-Boltzmann constant. Assuming $\varepsilon \approx 1$, one obtains the radial temperature profile $$\label{eq: 5.6} T(R)=\left(\frac{3\Omega^2\dot M}{16\pi \sigma}\right)^{1/4}=T(R_0)\Lambda^{1/4}\left(\frac{R}{R_0}\right)^{-3/4}\ .$$ where $R_0$ is a convenient reference radius, which we take to be $R_0=1\,{\rm AU}=1.5\times 10^{11}\,{\rm m}$, and $T(R_0)$ is the temperature at $R_0$ with $\Lambda$ set equal to unity. Given the radial dependence $T(R)$ of the temperature, the radial dependence of the velocity of sound and of the disk thickness can be determined. Using equations (\[eq: 5.5\]) and (\[eq: 5.4\]) we find $$\label{eq: 5.7}c_{\rm s}(R)=c_{\rm s}(R_0)\Lambda^{1/8}\biggl({R\over{R_0}}\biggr)^{-3/8}$$ $$\label{eq: 5.7a} H(R)=H(R_0)\Lambda^{1/8}\biggl({R\over{R_0}}\biggr)^{9/8}\ .$$ (Again, the dependence on $\Lambda$ is not shown explicitly in the list of arguments.) In order to estimate the radial dependence of other quantities it is necessary to determine the radial dependence of the density. This is achieved as follows. Applying the continuity equation for angular momentum leads to a relation between $\Sigma$ and an effective kinematic viscosity (termed the eddy viscosity), $\nu_{\rm eff}$, that is the diffusion coefficient characterising the transport of angular momentum across the accretion disk [@Sha73]: $$\label{eq: 5.8} \dot M=3\pi \nu_{\rm eff}(R)\Sigma(R)\ .$$ The eddy viscosity in turn is determined by the size $L$ of the largest eddies in the turbulent flow (the integral scale). We estimate $\nu_{\rm eff}\sim L^2/t_L$, where $t_L$ is the eddy turnover time for an eddy of size $L$. The Kolmogorov scaling argument implies that $t_L\sim (L^2/{\cal E})^{1/3}$, so that we estimate $$\label{eq: 5.9} \nu_{\rm eff}\sim L^{4/3}Q^{1/3}(R)\Sigma^{-1/3}(R)\ .$$ It is natural to assume that $L$ is of the order of $H(R)$ and we write $$\label{eq: 5.9a} L = \ell\, H(R)$$ where $\ell$ is a parameter. Taking (\[eq: 5.4\]), (\[eq: 5.5\]),(\[eq: 5.8\]) and (\[eq: 5.9\]) together, we find $$\label{eq: 5.10} \Sigma(R) = \frac{2\sqrt{2}}{9\pi \ell^2}\frac{\dot M}{H^2\Omega} .$$ Usually it is assumed [@Sha73] that $\nu_{\rm eff}\sim \alpha c_{\rm s}H$ (where $\alpha < 1$ is a dimensionless coefficient). Our own approach is equivalent to this $\alpha$-prescription’: using (\[eq: 5.10\]) in equation (\[eq: 5.8\]), we obtain $$\label{eq: 5.10a} \nu_{\rm eff}= \frac{3}{2\sqrt{2}}\ell^2 c_{\rm s}H\ .$$ Thus we see that the $\alpha$-presription is equivalent to assuming that the integral scale of the turbulence is smaller than $H$ by a factor $\ell \approx \sqrt{\alpha}$. It is widely accepted that may many observations are compatible with $\alpha\approx 10^{-2}$ [@Har98], corresponding to $L\approx H/10$. Accordingly, wherever we quote numerical values for quantities without specifying how they scale with $\ell$, we have set $\ell=0.1$. An advantage of our $\ell$-prescription’ is that it makes the physical nature of the adjustable parameter clearer than for the standard $\alpha$-prescription’. The rather small values of $\alpha$ indicated by observations could be indicative of a fundamental problem with a theory. Our alternative approach is reassuring because it indicates that a more physically transparent parameter, $\ell=L/H$, is not in fact a very small number. These results allow us to determine the power-law radial dependence of other quantities. The results for those determining the gaseous component of the accretion disk are summarised in table \[tab:tab1\], writing a generic variable in the form $$\label{eq: 5.11} X=X(R_0)\left(\frac{R}{R_0}\right)^{\delta_R}\Lambda^{\delta_\Lambda}\ell^{\delta_\ell}\,.$$ where $R_0=1\,{\rm AU}=1.5\times 10^{11}\,{\rm m}$, and $X(R_0)$ is the value of $X$ at radius $R_0$ with $\Lambda=\ell=1$. [lcccccc]{} Surface temperature&$T$& eq.(\[eq: 5.6\])&$130\,{\rm K}$&$-3/4$&$1/4$&$0$\ Speed of sound&$c_{\rm s}$ & eq.(\[eq: 5.7\])&$670\,{\rm m\,s}^{-1}$&$-3/8$&$1/8$&$0$\ Disk height&$H$& eq.(\[eq: 5.7a\])&$3.4\times10^9\,{\rm m}$&$9/8$&$1/8$&$0$\ Surface density&$\Sigma$&eq.(\[eq: 5.10\]) & $280\,{\rm kg\,m}^{-2}$& $-3/4$&$3/4$&$-2$\ Gas density&$\rho_{\rm g}$&eq.(\[eq: 5.3\])&$3.3\times 10^{-6}\,{\rm kg\,m}^{-3}$& $-15/8$&$5/8$&$-2$\ Dissipation rate&${\cal E}$&eq.(\[eq: 5.3\])&$0.11\,{\rm m}^2{\rm s}^{-3}$&$-9/4$&$1/4$&$2$\ Gas mean-free path&$\lambda$&$\mu m_{\rm H}/(\sqrt{2}\rho_{\rm g}S)$&$0.42\,{\rm m}$& $15/8$&$-5/8$&$2$\ Kinematic viscosity&$\nu$ & $\nu=\lambda c_{\rm s}$ & $280\,{\rm m}^2{\rm s}^{-1}$&$3/2$&$-1/2$&$2$\ Kolmogorov length&$\eta$ & eq.(\[eq: 3.2\]) & $120\,{\rm m}$ & $27/16$&$-7/16$&$1$\ Kolmogorov time &$\tau$ & eq.(\[eq: 3.2\]) & $52\,{\rm s}$ & $15/8$&$-3/8$&$0$\ Kolmogorov velocity&$u_{\rm K}$ & $\eta/\tau$ & $2.3\, {\rm m\,s}^{-1}$ & $-3/16$&$-1/16$&$1$\ Integral velocity&$u_L$ & $u_L=({\cal E}L)^{1/3}$ & $710\, {\rm m\,s}^{-1}$ & $-3/8$&$1/8$&$1$\ Integral timescale&$t_L$ & $t_L=L/u_L$ & $4.6\times 10^6\, {\rm s}$ & $3/2$&$0$&$0$\ Relative velocities, collision rates and Stokes trapping {#sec: 6} ======================================================== Next we consider the behaviour of dust grains suspended in the gas forming the accretion disk. We consider the case where the dust grains are almost all sub-micron sized particles: we assume they are composed of material with density $\rho_{\rm p} = 2\times 10^3{\rm kgm}^{-3}$, and that these are initially spherical particles of radius $a_0=10^{-7}{\rm m}$. Following @Hay81, we assume that the mass ratio $\kappa$ of suspended particles to gas molecules is $\kappa =3/170\approx 0.018$. The results described above can be used to estimate collision rates. We consider two cases. First, we estimate the collision rate when the Stokes number is small, and when the collision mechanism is Brownian diffusion (we shall see that advective collisions [@Saf56] make a negligible contribution). Second, we also require the rate of collision for large Stokes number. Both estimates require the number density of dust grains. We write $$\label{eq: 6.1} n\sim {\kappa \rho_{\rm g}\over{m_0}}\left(\frac{a}{a_0}\right)^{-D}=n_0\left(\frac{a}{a_0}\right)^{-D}$$ where $D$ is a fractal dimension and $\kappa\approx 0.018$ is the ratio of the mass densities of condensed to gaseous matter and $m_0=4\pi\rho_{\rm p}a_0^3/3$ is the mass of a microscopic dust grains; the second equality defines $n_0$. The mean speed of particles due to Brownian motion is $$\label{eq: 6.2} \langle v_{\rm d}\rangle\sim c_{\rm s}\sqrt{\mu m_{\rm H}\over{{N m_0}}}=v_0\left(\frac{a}{a_0}\right)^{-D/2}\nonumber$$ where the second equality defines $v_0$. Our estimate of the collision rate due to Brownian diffusion for a general value of the fractal dimension of the aggregates is $$\begin{aligned} \label{eq: 6.3} \nonumber {\cal R}_{\rm d}&\sim& 4\pi\sqrt{2}na^2\langle v_{\rm d}\rangle\sim 4\pi\sqrt{2}n_0a_0^2v_0\left(\frac{a}{a_0}\right)^{4-3D\over 2}\\&=&{\cal R}_{{\rm d}0}\left(\frac{a}{a_0}\right)^{4-3D\over 2}\\end{aligned}$$ and the factor ${\cal R}_{{\rm d}0}$ is given in table \[tab:2\]. The collision rate due to advective shearing is approximately $$\label{eq: 6.8} {\cal R}_{\rm a}\sim \frac{na^3}{\tau}\ .$$ The gas-kinetic collision rate (taking $K=1$ in equation (\[eq: 3.5\])) is $$\begin{aligned} \label{eq: 6.4} \nonumber {\cal R}_{\rm g}&=&4\pi n a^2\sqrt{\langle \Delta v^2\rangle}=4\pi n_0a_0^2\sqrt{{\cal E}\over{\gamma_0}}\left(\frac{a}{a_0}\right)^{2-D\over 2}\\&=&{\cal R}_{{\rm g}0}\left(\frac{a}{a_0}\right)^{2-D\over 2}\ .\end{aligned}$$ The Stokes number is $$\label{eq: 6.5} {\rm St}={1\over{\gamma \tau}}={\rm St}_0\left(\frac{a}{a_0}\right)^{D-2}\ .$$ All of these quantities have a power-law dependence upon radius and particle size: for the generic quantity $X$ we write $$\label{eq: 6.6} X =X(R_0,a_0)\left(\frac{R}{R_0}\right)^{\delta_R} \Lambda^{\delta_\Lambda}\, \ell^{\delta_\ell}\left(\frac{a}{a_0}\right)^{\delta_a}$$ (again, it is to be understood that $X(R_0,a_0)$ is evaluated for $\Lambda=\ell=1$). The quantities determining the collision rates are collected in table \[tab:2\]. [lccccccc]{} Damping rate&$\gamma$&eq.(\[eq: 4.3\])&$0.11\,{\rm s}^{-1}$&$-9/4$&$3/4$&$-2$&$2-D$\ Number density&$n$ &eq.(\[eq: 6.1\])&$7.0\times 10^{7}\,{\rm m}^{-3}$&$-15/8$&$5/8$&$-2$&$-D$\ Diffusive velocity&$v_d$&eq.(\[eq: 6.2\])&$1.4\times 10^{-2}\,{\rm m\,s}^{-1}$&$-3/8$&$1/8$&$0$&$-D/2$\ Diffusive collision rate&${\cal R}_{{\rm d}}$&eq.(\[eq: 6.3\])&$1.8\times10^{-7}\,{\rm s}^{-1}$&$-9/4$&$3/4$&$-2$&$(4-3D)/2$\ Advective collision rate&${\cal R}_{{\rm a}}$&eq.(\[eq: 6.8\])&$1.4\times 10^{-15}\,{\rm s}^{-1}$&$-15/4$&$1$&$-2$&$2-D$\ Relative velocity&$\sqrt{\langle \Delta v^2\rangle}$&eq.(\[eq: 3.5\]), $K\!=\!1$&$0.98\,{\rm ms}^{-1}$&$0$&$-1/4$&$2$&$(D-2)/2$\ Gas-kinetic collision rate&${\cal R}_{{\rm g}}$&eq.(\[eq: 6.4\])&$3.6\times 10^{-6}\,{\rm s}^{-1}$ &$-27/16$&$3/16$&$1$&$(2-D)/2$\ Stokes number&${\rm St}$&eq.(\[eq: 6.5\])&$0.17$&$3/8$&$-3/8$&$2$&$D-2$\ [lcccccc]{} Stokes-trapped particle size&$a^\ast$&eq.(\[eq: 6.7\])&$5.7\times 10^{-7}\,{\rm m}$&$-3/8$&$3/8$&$-2$\ Diffusive collision rate in Stokes-trap&${\cal R}_{{\rm d}}^\ast$ &eq.(\[eq: 6.3\])&$2.3\times10^{-9}\,{\rm s}^{-1}$ &$-21/16$&$3/16$&$3$\ Advective collision rate in Stokes trap&${\cal R}_{{\rm a}}^\ast$&eq.(\[eq: 6.8\]) &$2.5\times 10^{-13}\,{\rm s}^{-1}$&$-39/8$&$17/8$&$-8$\ Gas-kinetic collision rate in Stokes trap&${\cal R}_{{\rm g}}^\ast$&eq.(\[eq: 6.4\]) & $3.6\times 10^{-6}\,{\rm s}^{-1}$& $-27/16$&$3/16$&$1$\ Turbulent relative velocity at ${\rm St}=1$&$\langle \Delta v\rangle^\ast$ &$({\cal E}\tau)^{1/2}$&$2.3\,{\rm m\,s}^{-1}$&$-3/16$&$-1/16$&$1$\ The growth of the aggregates due to collisions could be be limited by two factors. Firstly, the rate of collisions usually decreases as the size of the aggregates increases and the collision rate could become negligible when the particles grow to a sufficiently large size. Secondly, large aggregates could be more vulnerable to being fragmented upon collision. To illuminate a discussion of whether these limitations occur in practice, we now consider the properties of particles with size $a$ chosen so that ${\rm St}=1$. For compact particles ($D=3$) we find that their size is $$\label{eq: 6.7} a^\ast=a_0\left(\frac{1}{{\rm St}_0}\right)^{1\over{D-2}}\sim a^\ast(R_0)\Lambda^{3/8}\,\ell^{-2}\left(\frac{R}{R_0}\right)^{-3/8}$$ with $a^\ast(R_0)\approx 6\times 10^{-7}{\rm m}$: for $\ell=0.1$ this leads to particles of size $a^\ast\approx 25\,\mu{\rm m}$ at $R=10\,{\rm AU}$. It is also of interest to evaluate the collision rate of these particles. Below we describe a mechanism which will lead to the aggregates becoming compact, so that we concentrate on the predictions of this calculation when $D=3$. The approximate values are listed in table \[tab:3\]. Note that the diffusive collision rate is very small, but it increases abruptly (by approximately five orders of magnitude, for the case where $\ell=0.1$) when the size of the particles increases such that the Stokes number exceeds unity (see figure \[fig: 3\](a)). Also, note that the relative velocity for collisions of particles with Stokes number unity is comparable to the critical velocity for microscopic dust aggregates to be fragmented. For compact aggregates ($D\approx 3$), the collision rate for the gas-kinetic mechanism is the much larger when ${\rm St}\gg 1$, and it also decreases less rapidly as the size of the particles increases. If the lifetime of the accretion system is $t_{\rm acc}\approx 10^6{\rm yr}$, the growth process may be assumed to be ended when ${\cal R}_{\rm g}t_{\rm acc}=1$. The estimate for ${\cal R}_{\rm g}$ above indicates that the gas-kinetic collision mechanism is sufficiently rapid that it does not limit the growth of aggregates, at least until they are so large that the approximations in equation (\[eq: 6.5\]), such as assuming that the drag is determined by the Epstein formula, break down. We conclude that kinetic factors do not limit the growth of aggregates, once particles become sufficiently large that ${\rm St}>1$. Note, however that for compact particles at the size where ${\rm St}=1$, the diffusive collision rate is extremely slow (of the order of one collision per $10^5$ years when $\ell=0.1$). The slowing down of the collision rate as the particle size increases is therefore quite close to becoming an insurmountable bottleneck. As the size $a$ of the aggregates increases, so does their relative velocity $\langle \Delta v\rangle$, for example compact aggregates of size $a=10\,{\rm cm}$ collide with a relative velocity of approximately $10\,{\rm ms}^{-1}$ if we take $\ell=0.1$. Such a collision would certainly cause the aggregates to fragment upon impact. We have arrived at the following picture of the initial stages of the growth of dust particle aggregates. Initially, dust particles will aggregate by Brownian diffusion. This process continues until the Stokes number of the larger particles is of order unity. Then the velocity of these particles starts to separate from that of the gas due to the formation of fold caustics. This greatly increases the relative velocity between these particles and smaller ones in their vicinity (figure \[fig: 3\](a)): the larger particles will rapidly sweep up’ the remaining small particles in their vicinity. However, when these larger particles collide with each other they have sufficient kinetic energy that they cause each other to fragment, so that particles with large Stokes number fragment to replenish the supply of particles with small Stokes number (figure \[fig: 3\](b)). This results in a steady-state distribution of particle sizes, corresponding to Stokes numbers of order unity (illustrated in figure \[fig: 3\](c)): we call this the Stokes trap’. Mechanism for avoiding the Stokes trap {#sec: 7} ====================================== In the following we propose four mechanisms by which the Stokes trap might be avoided. Escaping the Stokes trap by clustering {#sec: 7b} -------------------------------------- It is known that small particles suspended in a turbulent flow can cluster, even if the fluid flow is incompressible, due to effects of the inertia of the particles. This effect was proposed by @Max87, who suggested that heavy particles would be centrifuged’ out of vortices and would tend to cluster in regions of low vorticity: this effect is often referred to as preferential concentration’ because it is thought that the particles cluster in regions of low vorticity. The argument by Maxey does not give a quantitative description of the nature of the clustering effect at long times. @Som93 suggested that particles in random fluid flows cluster onto a fractal set of dimension $D_{\rm cl}$ and proposed using the Lyapunov dimension (first defined by @Kap79) to characterise this set. Recent numerical experiments [@Bec06] have studied how the (Lyapunov) fractal dimension $D_{\rm L}$ of particles in a three-dimensional turbulent flow varies as a function of Stokes number: the minimum value is $D_{\rm L}\approx 2.6$ for ${\rm St}=0.55$, $D_{\rm L}=3$ for ${\rm St}>{\rm St}_{\rm c}\approx 1.7$ and $D_{\rm L}\to 3$ as ${\rm St}\to 0$. Theoretical work [@Dun05] has shown that the fractal clustering may be explained without the centrifuge (or preferential concentration’) effect: a model for which it is absent gives a minimum dimension $D_{\rm L}\approx 2.65$, in good agreement with simulations. It has previously been suggested that this clustering effect may play a role in planet formation [@Cuz01], but there is a difficulty. The clustering effect only occurs when ${\rm St}=O(1)$, but most aggregation processes produce a broad (and time-dependent) distribution of sizes. The clustering mechanism would then only act on a small proportion of the particles. Our Stokes-trapping mechanism avoids this difficulty. [![\[fig: 3\] Schematic illustration of the Stokes trap. (a): Relative collision velocity of equal-sized aggregates versus their size $a$. (b): Small aggregates grow by collisions, but larger aggregates tend to fragment upon collision. (c): Corresponding stable particle-size distribution.](f4.eps "fig:"){width="5cm"}]{} The Stokes-trapped dust aggregates will be clustered, but if the clustering is confined to length scales below the Kolmogorov length $\eta$, the effect is not sufficiently strong to initiate gravitational collapse: the mass of solid matter within a Kolmogorov length is $$\label{eq: 7b.1} M\sim \kappa \rho_{\rm g}\eta^3\approx 1.0\times 10^{-3}\,{\rm kg}\,\Lambda^{-11/16}\,\ell\left(\frac{R}{R_0}\right)^{51/16}\ .$$ The size of the clusters created by the preferential concentration’ effect cannot significantly exceed this value, and this mass is much too small for gravitational effects to become significant. We thus conclude that although the clustering effect is promoted by the Stokes trap, clustering of dust aggregates onto a fractal set cannot provide inhomogeneities which are sufficient to trigger gravitational instability in protoplanetary systems. However, it is possible that this clustering effect could influence the scattering of electromagnetic radiation by the dust grains. Some other mechanisms for clustering have been proposed, discussed in @Bar95, @Joh07. We argue that these mechanisms can be discounted because they assume the existence of much larger aggregates than the Stokes-trapped size. Escaping the Stokes trap by chemical processes {#sec: 7a} ---------------------------------------------- We propose that the mid-plane of the accretion disk can be at a much higher temperature than its surface. The dust aggregates circulate between regions with different temperatures. In the lower temperature regions, water and organic compounds can condense onto the aggregates. When the particles circulate to higher temperature regions chemical reactions may occur. This is expected to both compactify the aggregate and to make it much more resistant to being fragmented in collisions. It is also possible that the temperature in regions close to the star is sufficiently high that at least some of their component materials are melted. This will also produce much more robust and compact aggregates, able to withstand collisions at much higher relative velocities. We have assume that the temperature of the accretion disk is determined by dissipative heating, and we have estimated the surface temperatures, which are quite low. However, if the disk is optically thick in the spectral region corresponding to the Wien wavelength, the mid-plane temperature may be considerably higher than the surface temperature. The optical thickness $W$ is the ratio of the height of the disk to the photon mean free path. The optical properties are dominated by the dust particles if $W\gg 1$, and when the particles are larger than the wavelength we may assume that their optical cross-section $S_{\rm opt}$ is approximately the square of their linear dimension: $S_{\rm opt}\approx a^2$. If the density of particles is $n$, the optical mean-free path is $\lambda_{\rm opt}\sim 1/(nS_{\rm opt})$. Thus for compact particles ($D=3$), the optical thickness is $$\label{eq: 7a.1} W\sim Hna^2\ .$$ The ratio of the interior temperature $T_{\rm int}$ of the accretion disk to its surface temperature $T$ is $$\label{eq: 7a.2} \frac{T_{\rm int}}{T}\sim W^{1/4}\ .$$ Very small particles, with dimension small compared to the Wien wavelength corresponding to temperature $T$, do not absorb or scatter radiation effectively. If the aggregates are very large, their number is correspondingly reduced, so that they make a smaller contribution to the optical thickness. From the data above, we see that the Stokes-trapped particles have sizes which are comparable to, but somewhat larger than, the Wien wavelength corresponding to the surface temperature. The Stokes-trapped particles are therefore of approximately optimal size to increase the optical thickness. The optical thickness corresponding to Stokes-trapped particles is therefore (for compact particles, $D=3$): $$\label{eq: 7a.3} W^\ast\sim \frac{H n_0 a_0^3}{a^\ast}\sim 410\, \Lambda^{3/8}\left(\frac{R}{R_0}\right)^{-3/8}\ .$$ The temperature at the centre of the disk is therefore expected to exceed the surface temperature by a factor of approximately four in the range $R=1-10{\rm AU}$. The chondrules, grains of typically submillimetre size which are found in many meteorites, show evidence of having been heated to very high temperatures, causing melting. The temperatures required are rather higher than those which we predict above (except very close to the star). Also, the exposure of chondrules to high temperatures may have been brief compared to the circulation time $t_L$, otherwise the chondrules may have evaporated. Escaping the Stokes trap by forming tenuous aggregates {#sec: 7c} ------------------------------------------------------ We have already remarked that if the fractal dimension $D$ of the dust aggregates is $D\le 2$, then the Stokes number does not increase as the size of the aggregate increases. Such aggregates would avoid the Stokes trap’, but there are three reasons why these tenuous aggregates might not provide a satisfactory solution to the problem of planet formation. First, it might be thought that as such aggregates grow their collision rate would become so small that growth would, for all practical purposes, cease. In section \[sec: 4\] it was determined how the collision rate for fractal aggregates with dimension $D$ scales as a function of their characteristic size, $a$: we found ${\cal R}_{\rm d} \propto a^{(4-3D)/{2}}$. The collision rate therefore increases with the aggregate size when $D<4/3$, because such very tenuous aggregates are so extended that they entangle each other. Experiments on growing aggregates in negligible gravity have produced objects with very low fractal dimensions of roughly $D=1.4$ [@Blu00; @Kra04], rather smaller than the values predicted for kinetic aggregation processes. The dust aggregates observed in low-gravity experiments are only observed for a short period compared to the time between collisions between dust aggregates, so that the structure of the aggregates in protostellar accretion systems is uncertain, but there is probably no kinetic restriction on the growth of very tenuous clusters. A second possibility is that the tenuous aggregates are torn apart by the effect of the shearing motion of the fluid. We estimate this effect as follows. Consider the force required to split an aggregate of fractal dimension $D$, size $a$, composed of particles of size $a_0$. The number of bonds at an equatorial plane of the aggregate is approximately $$\label{eq: 7c.2} N_{\rm b}\sim \left(\frac{a}{a_0}\right)^{D-1}\ .$$ Each bond has a breaking strain $F_{\rm b}$. Grains in contact have an energy per unit area $\delta E_1/l^2$ per unit area, acting over a range $l$. The binding energy is $E_{\rm b}=\pi\delta E_1a_0/l$ (see section \[sec: 4.3\]). Writing $E_{\rm b}=F_{\rm b}l$, we estimate $$\label{eq: 7c.3} F_{\rm b}\sim \frac{\delta E_1}{l^2}a_0\ .$$ The drag force on a single grain due to fluid motion with a speed $u$ relative to the fluid is $F=m_0\gamma_0 u$. When $D\le 2$, the drag force pulling the aggregate apart is comparable to the sum of the forces which are predicted by applying this formula to each grain. The relative velocity of the gas at opposite sides of the aggregate is $a/\tau$, so that the force acting to break up the aggregate is $$\label{eq: 7c.4} F_{\rm br}\sim \frac{N\gamma_0m_0a}{\tau}\sim \frac{m_0\gamma_0 a_0}{\tau}\left(\frac{a}{a_0}\right)^{D+1}\ .$$ The ratio of the applied force to the critical force is $F_{\rm br}/N_{\rm b}F_{\rm b}$; note that this ratio is independent of $D$ (provided $D\le 2$). Setting this ratio equal to unity we estimate the size $a_{\rm max}$ of aggregates which will be torn apart by shearing forces: $$\label{eq: 7c.5} a_{\rm max}\approx 78\,{\rm m}\,\Lambda^{-9/16}\,\ell\left(\frac{R}{R_0}\right)^{33/16}\ .$$ The estimates leading to this result are valid provided the aggregate is small compared to the Kolmogorov length: $a\ll \eta$: this condition is only marginally satisfied. It is not clear how the shear forces increase with aggregate size once the aggregate is large compared to the Kolmogorov correlation scale, but it is hard to escape the conclusion that tenuous aggregates are torn apart by shear forces when they become sufficiently large. The mass contained in these tenuous aggregates is clearly comparable to that estimated in equation (\[eq: 7b.1\]) and is too small to initiate gravitational collapse directly. A third problem is that aggregates with Stokes number less than unity are always advected with the gas. Even if they could become very massive, they would never collapse to the mid-plane of the disk, as long as the motion of the gas remains turbulent. Escaping the Stokes trap by reducing the accretion rate {#sec: 7d} ------------------------------------------------------- If the accretion rate $\dot M$ is reduced (by choosing a small value for the parameter $\Lambda$ in equation (\[eq: 5.0\])), there is a corresponding reduction in the turbulence intensity ${\cal E}$. We might expect that there would be also be a reduction in the relative velocities of suspended particles. This latter expectation is false, because a reduction in the rate of accretion results in a reduced density of the gas, and the suspended particles are more lightly damped by the gas (note that if $\gamma$ decreases the relative velocity increases: see equation (\[eq: 3.5\])). This principle is illustrated by results in table \[tab:3\], which show that the Stokes-trapped particle size decreases as the accretion rate parameter $\Lambda$ decreases (because $\delta_\Lambda>0$ for $a^\ast$). A further demonstration comes from considering the size of particles which are not fragmented on collision. If the relative velocity is less than the critical velocity $v_{\rm cr}$ for disintegration of dust aggregates, then they continue to grow on collision. The size of particles for which the turbulent relative velocity exceeds $v_{\rm cr}$ is $$\label{eq: 7d.1} a_{\rm cr}=1.0\times 10^{-7}\ {\rm s}^2{\rm m}^{-1}v_{\rm cr}^2\,\ell^{-4}\,\Lambda^{1/2}\ .$$ We see that, contrary to intuition, the attainable particle size decreases as the rate of accretion decreases. Settling and gravitational collapse {#sec: 8} =================================== Even if the Stokes trap is circumvented, the particles must grow to a substantial size in order for them to settle to the mid-plane of the accretion disk. In the following we illustrate this fact by estimating the size at which particles start to settle to the mid-plane, according to our model. We find that settling only starts at large particle sizes (approximately $10\,{\rm cm}$, where the use of the Epstein formula for the damping rate becomes questionable). We also estimate the critical height of the dust layer for the onset of gravitational collapse, and find that this is a small fraction of the height of the gas layer. Critical size for settling {#sec: 8.1} -------------------------- The integral scale fluctuations with velocity $u_L$ will advect aggregates away from the mid plane. This effect is opposed by gravitational attraction to the mid plane. The gravitational acceleration at a distance $z$ from the mid plane is $$\label{eq: 8.1} g(R,z)={GM\over R^3}z=\Omega^2z\ .$$ If the integral timescale of the turbulence, $t_L$, is large compared the the relaxation time $\gamma^{-1}$, then the aggregates only settle to the mid-plane if their terminal speed $u_{\rm t}=g(R,z)/\gamma$ exceeds $u_L$. There is thus no settling at all unless $\gamma<\gamma_{\rm s}$, where $$\label{eq: 8.2} \gamma_{\rm s}={\Omega^2H\over{u_L}}=1.8\times 10^{-7}\,{\rm s}^{-1}\ell^{-1}\left(\frac{R}{R_0}\right)^{-3/2}\ .$$ In the case of compact aggregates ($D=3$), we conclude that particles start to settle to the mid-plane when their size exceeds $$\label{eq: 8.3} a_{\rm s}\approx a_0\left(\frac{\gamma_0}{\gamma_{\rm s}}\right)\approx 6.0\times 10^{-2}{\rm m}\,\Lambda^{3/4}\,\ell^{-1}\left(\frac{R}{R_0}\right)^{-3/4}\ .$$ When considering very large aggregates, the damping rate $\gamma $ may be sufficiently small that $\gamma t_L\ll 1$. In this case, the effect of the turbulence must be modelled as a sequence of random impulses on the aggregate [@Cuz93; @You07]. We argue that, because of the Stokes trapping effect, the aggregates never grow to a sufficient size for this alternative approach to become relevant. Critical height for gravitational collapse {#sec 8.2} ------------------------------------------ A stationary gas of solid particles with mass density $\rho_{\rm sol}$ is unstable against gravitational collapse on a timescale $t_{\rm coll}\sim (G\rho_{\rm sol})^{-1/2}$. If the gas consists of particles with typical relative speed $\Delta v$, then gravitational collapse (the Jeans instability) occurs on length scales greater than $\Delta v t_{\rm coll}$. In the case of an accretion disk, the question of gravitational instability can be complicated by the rotational motion and the finite thickness of the disk. The gravitational stability of a uniform thin disk of area density $\Sigma$ may be described by giving a dispersion relation for the frequency $\omega$ of a density perturbation with wavenumber $k$ $$\label{eq: 8.4} \omega^2=\Omega^2-2\pi G\Sigma \vert k\vert+v_{\rm s}k^2$$ where $v_{\rm s}$ is a two-dimensional sound velocity [@Bin88]. The system becomes gravitationally unstable when $\omega^2<0$ for any value of $k$. For a collisionless system, we have $v_{\rm s}=0$, and the instability arises when the term containing the gravitational constant is larger in magnitude that the centrifugal term. The instability is favoured by choosing a large value for $k$, but if we consider very large values of $k$, the approximation of treating the mass distribution as two-dimensional fails. We therefore assume that the largest possible value of $k$ is $2\pi/H$, where $H$ is the height of the disk. The solid material in the disk (with area density $\kappa \Sigma$) therefore becomes gravitationally unstable when its scale height, $H_{\rm sol}$ is less than a critical value which is given (approximately) by $$\label{eq: 8.5} H_{\rm cr}=\frac{4\pi^2 \kappa G\Sigma}{\Omega^2}\approx 3.3\times 10^5{\rm m}\,\Lambda^{3/4}\,\ell^{-2}\left(\frac{R}{R_0}\right)^{9/4}\ .$$ Note that $H_{\rm cr}/H\ll 1$, implying that to trigger a gravitational instability, particles must be very much heavier than particles of the size $a_{\rm cr}$ (given by equation (\[eq: 8.3\])). Concluding remarks {#sec: 9} ================== In this paper we have analysed the consequences of turbulence for a standard model of the formation of planets by aggregation of sub-micron sized dust. This led us to introduce the concept of the Stokes trap, a mechanism which limits the growth of dust aggregates in turbulent protostellar accretion disks and thus constitutes a barrier to planet formation. The particle sizes predicted by the Stokes trap mechanism are so small that turbulent fluctuations would never allow the particles to settle to the mid-plane of the disk and achieve sufficient density to trigger gravitational instability. The Stokes trap must therefore be avoided if the planets are to form. We have proposed four possible ways of avoiding the Stokes trap and assessed these mechanisms in the light of Kolomogorov’s scaling principle for turbulence in combination with a standard model for the accretion disk. Particles clustering due to preferential concentration’ cannot produce sufficiently large clusters. Chemical processes which consolidate aggregates will strengthen them, but not to the extent that large aggregates would be invulnerable to fragmentation by collision. If the aggregates are very tenuous fractals, with dimension $D\le 2$, their Stokes number remains very small, but the aggregates may be torn apart by shearing forces when they become sufficiently large, and they will not settle to the mid-plane of the disk. Finally, we observed that as the disk becomes less active (that is, as $\dot M$ decreases) the relative velocity of the suspended aggregates increases, so that the difficulty is not removed as the disk becomes quiescent. Even if the aggregates escape from the Stokes trap and continue growing, our estimates indicate that they must reach very large sizes before they can overcome the effects of turbulence and start to settle to the mid-plane. Our estimates indicate that the turbulence intensity would have to be reduced by many orders of magnitude before this could happen. We conclude that a model for planet growing by aggregation of small dust grains is highly problematic. It is therefore necessary to examine other possibilities. The difficulties with the theory are avoided if the protoplanetary accretion disk already contains large objects with a gravitational escape velocity which is large enough to prevent them from being disrupted by collisions with dust grains and aggregates. A star forms when a cloud of interstellar gas satisfies conditions which allow it to undergo gravitational collapse. It is likely that at the same time as the star forms, other regions of the cloud in the vicinity of the nascent star will condense by gravitational attraction. Condensed objects much smaller than the star itself may be drawn into the accretion disk of the star, and eventually become planets lying on roughly circular orbits in the equatorial plane of the star. Thus we propose an alternative picture for the formation of planets which we term concurrent collapse’, in which the planets are formed as gravitationally bound systems at the same time as the star. During the life of the accretion disk they become coupled by friction and mass transfer to the circular motion of the accretion disk. While in the accretion disk their structure may be radically transformed by processes such as further accretion of material. Their mass might also be reduced by ablation or evaporation of lighter elements. One possible criticism of this scenario is that the density $\rho$ of the gas cloud at temperature $T$ which collapses to form the star will be associated with a mass scale, the Jeans mass, $$\label{eq: 9.1} M_{\rm J}\sim \left[\frac{1}{\rho}\left(\frac{k_{\rm B}T}{G m_{\rm H}}\right)^3\right]^{1/2}\ .$$ It may be argued that the Jeans mass is a lower limit to the mass of objects that could form by gravitational collapse. However, as gravitational collapse proceeds, the density of some regions will increase, thus lowering the Jeans mass estimate which applies for further collapse. Images of gas clouds suggest that they are typically non-uniform in their density, and that they have density fluctuations spanning a wide range of spatial scales. The distribution of sizes of dense objects produced by gravitational collapse from such a non-uniform initial condition might be expected to be very broad. For the time being our concurrent collapse’ hypothesis must be supported by the implausibility of the standard dust aggregation model. Because the additional hypothesis only concerns conditions at early stages of the life of the stellar accretion system, its implications for the eventual structure of a solar system are unclear. Support from Vetenskapsrådet and the platform Nanoparticles in an interactive environment’ at Göteborg university are gratefully acknowledged. J. Abrahamson, Collision rates of small particles in a vigorously turbulent fluid, [*Chem. Eng. Sci.*]{}, [**30**]{}, 1371-9, (1975) D. Apai [*et al.*]{}, The onset of planet formation in brown dwarf disks, [*Science*]{}, [**310**]{}, 834, (2005) P. J. Armitage, Lecture notes on the formation and early evolution of planetary systems, arXiv:astro-ph/0701485 P. Barge and J. Sommeria, Did planet formation begin inside persistent gaseous vortices?, [*Astron. Astrophys.*]{}, [**295**]{}, L1-4, (1995). F. Bauer, C. P. Dullemond, A. Johansen, Th. Henning, H. Klahr and A. Natta, Survival of the mm-cm size grain population observed in protoplanetary disks, arXiv:astro-ph/0704.2332 J. Bec, L. Biferale, G. Boffetta, M. Cencini, S. Musachchio and F. Toschi, Lyapunov exponents of heavy particles in turbulence, [*Phys. Fluids*]{}, [**18**]{}, 091702, (2006) J. Binney and S. Tremaine, [*Galactic Dynamics*]{}, Princeton University Press, (1988) J. Blum [*et al*]{}, Growth and form of planteary seedlings: results from a microgravity aggregation experiment, [*Phys. Rev. Lett.*]{}, [**85**]{}, 2426-9, (2000) A. C. Carciofi, J. E. Bjorkman and A. M. Magalhaes, Effects of grain size of the spectral energy distribution of dusty circumstellar envelopes, [*Astrophysical J.*]{}, [**604**]{}, 238-51, (2004) A. Chokshi, A. G. G. M. Tielens and D. Hollenbach, Dust coagulation, [*Astrophys. J.*]{}, [**407**]{}, 806-819, (1993) J. N. Cuzzi, A. R. Dobrovolskis and J. M. Champney, Particle-gas dynamics in the mid-plane of a protoplanetary nebula, [*Icarus*]{}, [**106**]{}, 102-34, (1993) J. N. Cuzzi, R. C. Hogan, J. M. Paque, A. R. Dobrovolskis, Size-selective concentration of chondrules and other small particles in protoplanetary nebula turbulence, [*Astrohys. J.*]{}, [**546**]{}, 469-508, (2001) C. Dominik, J. Blum, J. N. Cuzzi and G. Wurm, Growth of dust as the initial step toward planet formation, [*Protostars and protoplanets V*]{}, B. Reipurth, D. Jewitt and K. Keil (eds.), 783-800, (2007) C. Dominik and A. G. G. M. Tielens, The physics of dust coagulation and the structure of dust aggregates in space, [*Astrophys. J.*]{}, [**480**]{}, 647-73, (1997) K. P. Duncan, B. Mehlig, S. " Ostlund and M. Wilkinson, Clustering in mixing flows, [*Phys. Rev. Lett.*]{}, [**95**]{}, 240602, (2005) P. S. Epstein, On the resistance experienced by spheres moving through gases, [*Phys. Rev.*]{}, [**23**]{}, 710, (1924) G. Falkovich, A. Fouxon and G. Stepanov, Acceleration of rain initiation by cloud turbulence, [*Nature*]{}, [**419**]{}, 151-154, (2002) G. Falkovich and A. Pumir, Sling effect in collisions of water droplets in clouds, arXiv:nlin/0605040, (2006) U. Frisch, [*Turbulence*]{}, Cambridge University Press, (1997) L. Hartmann, N. Calvet, E. Gullbring and P D’Aessio, Accretion and evolution of T-Tauri disks, [*Astrophysical J.*]{}, [**495**]{}, 385, (1998) C. Hayashi, Structure of the solar nebula, growth and decay of magnetic fields and effects of magnetic and turbulent viscosities of the nebula, [*Suppl. Prog. Theor. Phys.*]{}, [**70**]{}, 35-53, (1981) T. Henning, C. P. Dullemond, S. Wolf, Dust coagulation in protoplanetary disks, in [*Planet Formation*]{}, ed. H. Klar and W. Brandner, in press, (2006) A. Johansen, J. S. Oishi, M-M. M. Low, H. Klahr, T. Henning and A. Youdin, Rapid Planetesimal Formation in Turbulent Circumstellar Discs, [*Nature*]{}, [**448**]{}, 1022-5, (2007) J. L. Kaplan and J. A. Yorke, in: [*Functional differential equations and approximations of fixed points*]{}, eds.: H.-O. Peitgen and H.-O. Walter, Lecture notes in mathematics, [**730**]{}, Springer, Berlin, (1979), p. 204 J. E. Kessler-Silacci, C. P. Dullemond and J. C. Augereau, Probing protoplanetary disks with silicate emission: where is the silicate emission zone?, [*Astrophysical J.*]{}, [**659**]{}, 680-4, (2007) M. Krause and J. Blum, Growth and form of planteary seedlings: results from a sounding rocket microgravity aggregation experiment, [*Phys. Rev. Lett.*]{}, [**93**]{}, 021103, (2004) W. J. Markiewicz, H. Mizuno and H. J. Völk, Turbulence induced relative velocity between two grains, [*Astron. Astropkys.*]{}, [**242**]{}, 286, (1991) M. R. Maxey, The gravitational settling of aerosol paerticles in homogeneous turbulence and random flow field, [*J. Fluid Mech.*]{}, [**174**]{}, 441, (1987) B. Mehlig, V. Uski and M. Wilkinson, Colliding particles in highly turbulent flows, [*Phys. Fluids*]{}, [**19**]{}, 098107, (2007) H. Mizuno, W. J. Markiewicz and H. J. Völk, Grain growth in turbulent protoplanetary accretion disks, [*Astron. Astrophys.*]{}, [**195**]{}, 183-92, (1988) R. Nakamura, Y. Kitada and T. Mukai, Gas drag forces on fractal aggregates, [*Planet. Space Sci.*]{}, [**42**]{}, 771-6, (1994) T. Poppe and J. Blum, Experiments on pre-planetary grain growth, [*Adv. Space Res.*]{}, [**20**]{}, 1595-1604, (1997) P. G. Saffman and J. S. Turner, On the collision of drops in turbulent clouds, [*J. Fluid Mech.*]{}, [**1**]{}, 16-30, (1956) V. S. Safranov, Evoliutsiia Doplanetnogo Oblaka, (1969) (English transl.: Evolution of the protoplanetary cloud and formation of the Earth and planets, NASA Tech. Transl. F-677, Jerusalem: Israel Sci. Transl., 1972) N. I. Shakura and R. A. Sunyaev, Black holes in binary systems. Observational appearance, [*Astronomy and Astrophysics*]{}, [**24**]{}, 337, (1973) J. Sommerer and E. Ott, Particles floating on a fluid - a dynamically comprehensible physical fractal, [*Science*]{}, [**259**]{}, 335, (1993) S. Sundaram and L. R. Collins, Collision statistics in a isotropic particle-laden turbulent suspension. Part 1. Direct numerical simulations, [*J. Fluid Mech.*]{}, [**335**]{}, 75-109, (1997) R. van Boekel [*et al.*]{}, Grain growth in the inner regions of Herbig Ae/Be star disks, [*Astron. & Astrophys.*]{}, [**400**]{}, L21, (2003) H. J. Völk, F. C. Jones, G. E. Morfill and S. Röser, Collisions between grains in a turbulent gas, [*Astron. & Astrophys.*]{}, [**85**]{}, 316, (1980) M. Wilkinson and B. Mehlig, Caustics in turbulent aerosols, [*Europhysics Lett.*]{}, [**71**]{}, 186-92, (2005) M. Wilkinson, B. Mehlig and V. Bezuglyy, Caustic activation of rain showers, [*Phys. Rev. Lett.*]{}, [**97**]{}, 048501, (2006) K. Wood, M. J. Wolff, J. E. Bjorkman and B. Whitney, The spectral energy distribution of HH 30 IRS: constraining the circumstellar dust size distribution, [*Astrophysical J.*]{}, [**564**]{}, 887, (2002) G Wurm and J. Blum, Experiments on preplanetary dust aggregation, [*Icarus*]{}, [**132**]{}, 125-36, (1998) A. N. Youdin, Obstacles to the collisional growth of planetesimals, arXiv:astro-ph/0311191 A. N. Youdin and Y. Lithwick, Particle stirring in turbulent gas disks: including orbital oscillations, arXiv:astro-ph/0707.2975

--- abstract: 'Using computer simulations, we identify the mechanisms causing aggregation and structural arrest of colloidal suspensions interacting with a short-ranged attraction at moderate and high densities. Two different non-ergodicity transitions are observed. As the density is increased, a glass transition takes place, driven by excluded volume effects. In contrast, at moderate densities, gelation is approached as the strength of the attraction increases. At high density and interaction strength, both transitions merge, and a logarithmic decay in the correlation function is observed. All of these features are correctly predicted by mode coupling theory.' address: 'Department of Physics and Astronomy, The University of Edinburgh, EH9 3JZ, UK' author: - 'Antonio M. Puertas [@puertas], Matthias Fuchs and Michael E. Cates' title: Comparative simulation study of colloidal gels and glasses --- [2]{} Colloidal dispersions aggregate into various non-equilibrium structures depending on density, interaction strength and range. The accompanying rheology and structure are among the key properties desired for their technological applications [@Russel89]. Moreover, thanks to the possibility to tailor effective interactions by e.g. addition of salt and polymer, colloids allow us to study the fundamental mechanisms of kinetic arrest. Whereas colloidal hard spheres have become a model system for the study of structural arrest at a glass transition [@Megen98], colloidal gelation has only recently been associated with glassy behavior [@Verduin95; @Bergenholtz99; @Segre01]. Colloidal gelation is ubiquitous in suspensions driven by attractions of quite short range and moderate to high strength [@Poon97]. At low packing fractions, it entails the formation of heterogeneous and often self-similar networks; there, an interplay of phase separation kinetics and percolation often are considered responsible for its existence [@Poon97]. At higher densities, the gelation boundary extends into the homogeneous fluid region [@Grant93; @Verduin95], where it also lies well separated from estimates of percolation [@Grant93; @Verduin95; @Rueb98; @Mallamace00]. Crossing into the gelled state anywhere along the transition line results in qualitatively the same phenomena, like flow properties that indicate solidification [@Grant93; @Rueb98], and non-ergodic dynamics according to light scattering [@Verduin95; @Poon99; @Pham01]. We present simulations designed to identify the mechanism of colloidal gelation driven by attractions of only moderate strength. Because of the distance of the gel boundary from other boundaries (percolation and phase separation) at higher densities, we concentrate on these, where we sweep out the region between gel and glass transition lines. We show that both non-equilibrium transitions are caused by a slowing down of local rearrangements, as predicted well by mode coupling theory (MCT) [@Fabbian99; @Bergenholtz99; @Dawson01]. We contrast the glass transition, caused by caging of particles owing to steric hindrance, with attraction-driven gelation caused by bonding between particles. We verify that the simultaneous presence of two non-ergodic states results in anomalous non-exponential (logarithmic) time dependences, as recently conjectured to explain observations in micellar systems [@Mallamace00] or microgel suspensions [@Bartsch94]. The simulated system comprises 1000 soft-core ($V(r=|{\bf r}_{i}-{\bf r}_j|) \propto(a_{ij}/r)^{36}$, $a_{ij}=a_i+a_j$) particles of mean radius $a$ with polydispersity in size (flat distribution with 10% width) to prevent crystallisation. Densities are reported as packing fractions $\phi_c=\frac{4\pi}{3} n a^3$. A short range attraction, mimicking the polymer induced depletion attraction in experimental systems [@Russel89; @Poon97; @Poon99; @Pham01], is given by an Asakura-Oosawa (AO) form generalized to polydisperse systems [@mendez00; @potshape]. The range of the attraction, $2 \xi$, is set to $0.2 a$, and its strength is proportional to the polymer concentration $\phi_p$. To help avoid liquid-gas separation, a weak long range barrier is added to the potential. The barrier extends from $a_{12}+2\xi$ to $4a$, and is described by a fourth order polynomial matched to give a continuous force. Its maximal height is $1k_BT$, which equals the depth at contact of the AO potential at $\phi_p=0.0625$. In all states studied, the barrier is much smaller than the attraction, and in the purely repulsive case ($\phi_p=0$) it is omitted. We will measure lengths in units of $a$ and time in units of $\sqrt{4a^2/3v^2}$, where the thermal velocity, $v$, is set to $2/\sqrt{3}$. Equations of motion were integrated using the velocity-Verlet algorithm, with a time step of $0.0025$. Colloidal dynamics (neglecting hydrodynamic interactions) were mimicked by running the simulations in the canonical (constant NTV) ensemble, where the thermostat plays the role of the surrounding liquid. Every $N$ time steps, the velocity of the particles was rescaled to assure constant $v$. No effect of $N$ on the results was observed for well equilibrated samples. The central quantity of our study will be the self part of the intermediate scattering function, $\Phi_q^s(t)=\langle\exp i {\bf q}\cdot \left( {\bf r}_j(t)- {\bf r}_j(0) \right) \rangle$, for wave-vector ${\bf q}$, where $\langle \dots \rangle$ denotes an average over particles and time origin. $\Phi_q^s(t)$ allows us to probe and identify the nature of the dominant dynamical mechanism because of $(i)$ its $q$-dependence and $(ii)$ the detailed predictions that are available from MCT. Indeed, if a structural arrest at a non-ergodicity transition is approached, $\Phi_q^s(t)$ reveals a two-step process, where the decay from the plateau is given by the von Schweidler power-law series [@Franosch97]: $$\Phi_q^s(t)\:=\:f_q^s\,-\,h_q^{(1)} (t/\tau)^b\,+\,h_q^{(2)} (t/\tau)^{2b}\,+\,O(t^{3b}) \; . \label{alpha-decay}$$ Here $f_q^s$ is the non-ergodicity parameter, $h_q^{(1)}$ and $h_q^{(2)}$ are amplitudes, and $b$ is known as the von Schweidler exponent. On the one hand, the observation of this (universal) von Schweidler behavior – and tests of further relations, as done below – establishes that a feed-back mechanism in the structural relaxation causes arrest. On the other hand, the (non-universal) wave-vector dependence of the amplitudes, like $f_q^s$, allows us to identify the specific kinetic process which freezes out. As the transition is approached, the characteristic time $\tau$ diverges as $\tau \propto |\phi-\phi^c|^{-\gamma}$, where $\gamma$ is determined by $b$, see e.g. [@Franosch97]. Figure 1 presents evidence for both the repulsion and attraction driven glass transitions, as identified by a diverging $\tau$ [@taucom]. Upon increasing the packing fraction $\phi_c$ (inset in figure 1), the system approaches a glass transition caused by steric hindrance, which we have studied including only the $r^{-36}$ repulsion ($\phi_p=0$ and no barrier) [@barrier]. The transition correlates well with observations at the colloidal glass transition [@Megen98] and previous simulations of e. g. a glassy Lennard-Jones mixture [@Kob]. We have analysed it using the concepts of idealized MCT, but will present only a few results here for comparison with gelation. Gelation itself is induced by strengthening the attraction at intermediate packing fractions (figure 1). There, far from equilibrium or percolation transitions (as we tested by monitoring the static structure factor), arrest again is of kinetic origin, and occurs at lower attraction strengths the higher $\phi_c$. To shed light on this transition, the correlation functions at different wave-vectors were studied. The slowest state at $\phi_c=0.40$ is presented in Fig \[fig2\]. The upper panel shows the self intermediate scattering functions for different wave-vectors, and the fits using (\[alpha-decay\]) up to second order. A common exponent $b$ was taken in the fitting, yielding a value of $b=0.38$, appreciably lower than the hard spheres value $b=0.53$ (which we found for our soft sphere glass at $\phi_p=0$). As predicted by MCT, we can calculate from $b$ the divergence of the relaxation times in Fig. \[fig1\]. The resulting value $\gamma=3.03$ fits the data, while $\gamma=2.63$ at the soft sphere glass transition [@barrier]. The gel transition is estimated to occur at $\phi_p=0.431$. Because we find the universal properties predicted by MCT, we conclude that at $\phi_c=0.40$ the gel transition is a regular non-ergodicity transition in the structural dynamics. Their very different $q$-width (Fig. \[fig2\]b) for the non-ergodicity parameters and amplitudes brings out a major difference in the two underlying mechanisms. Whereas repulsions localize the particle within a cage, which it can explore up to mean squared displacements $r_l^2$ of the order of $r_l^2=0.13$ (from our simulations, not shown; $r_l^2=0.134$ from MCT [@Fuchs92]), attractions bind the particle to its neighbors and thus localize it much more tightly. At $\phi_c=0.40$, we find $r_l^2=0.018$ by simulations, which is of the order of a low-density estimate [@Bergenholtz99] for our interaction range, $2\xi=0.2$. The corresponding high amplitudes $f_q^s$ of density fluctuations are consistent with light scattering observations at fixed $q$ [@Poon99; @Pham01] and with MCT calculations [@Bergenholtz99; @Fabbian99; @Dawson01]. These fluctuations extend to large $q$ and relax only when the particles break free from their bonds. The comparison with the Gaussian approximation, $f^{s {\rm G}}_q=\exp{\{-q^2 r_l^2/6\}}$ evidences stronger non-Gaussian effects at gelation than at the glass transition. We stress the cooperativity of the structural relaxation at both transitions. Holding all particles fixed, except for one, leads to mean squared displacement for the tracer (as it explores the frozen enviroment) much smaller than in the free system (before the start of the structure relaxation of the free system, the ratio is $\approx 6$ for both cases). The cage or network of bonds around an arrested particle thus necessarily fluctuates with it. To test further the nature of the gel transition, the scaling of the final (or $\alpha$-) decay was studied. In Fig. \[fig3\] we present the rescaled ($\Phi_q^s(t/\tau=1)=0.25$) correlation functions at $q=9.9$ for different attraction strengths, close to the gel transition. In the inset to this figure, a similar plot deals with the glass transition ($q=3.9$ in this case). In both cases, the curves clearly collapse during the $\alpha$-decay indicating an unique mechanism which dominates the slowing down at the transitions. For the purely repulsive case, the MCT master curve for the rescaled decay of hard spheres [@Fuchs92] at a slightly larger wave-vector ($q=4.3$) is also presented, confirming the quantitative agreement between MCT and our results. In the gel case, no master function is available, but the fit to (\[alpha-decay\]) is presented. The different stretching in the two cases is clear. We study now the gel transition at a higher density $\phi_c=0.55$, where it lies closer to the glass transition. Within MCT, the simultaneous existence of two different non-ergodicity transition branches opens the possibility for end-point singularities, where the branches merge [@gotze89]. In systems with short range attractions, such a singularity has been predicted close to the crossing of the two transition lines (the actual distance depending on the details of the potential) [@Fabbian99; @Bergenholtz99; @Dawson01]. Close to the singularity, a logarithmic decay around the plateau in the correlation function, $\Phi_q^s(t)=f_q^{sA}-C_q \ln (t/t_1)$, is a proposed signature [@Fabbian99; @Dawson01], and intriguingly is observed experimentally in more complicated systems [@Bartsch94; @Mallamace00]. Having identified two different non-ergodicity transitions in our system, we now test this prediction. In Fig. \[fig4\], the correlation functions at the same wave-vectors as in Fig. \[fig2\] are presented for the state $\phi_p=0.375$. Logarithmic decays are observed in all of the correlators (linear traces in the plot) for up to three decades in time, signalling a higher order singularity nearby. It is interesting to note that the logarithmic trace of the correlator has different extents, depending on the wave-vector. To make it clearer, $(\Phi_q^s-f_q^{sA})/C_q$ vs. time has been plotted as an inset to Fig. \[fig5\], where $f_q^{sA}$ is determined at $t_1=5$ (vertical line in Fig. \[fig4\]). Deviations from the logarithmic decay are stronger, the higher $f_q^s$. This is in complete agreement with the theoretical expectations in [@Dawson01]. Since both the von Schweidler decay (associated with the gel transition) and the logarithmic trace now take place in the same window, important corrections to the $\alpha$-scaling of the curves are expected. This is seen in the main graph of Fig. \[fig5\]: the long time dynamics at different states cannot be collapsed onto a master curve by time rescaling. This shows that two mechanisms are responsible for the structural slowing down and that changes in the control parameters change the relative distance to gelation but also to the higher order singularity. The intermediate isochore, $\phi_c=0.50$, shows a mixed behavior: a logarithmic decay in a smaller window than for $\phi_c=0.55$, and followed by an apparent power law decay. In summary, using MD simulations, we have deduced from the wave-vector dependence of the dynamical density fluctuations, that repulsion and short-ranged attraction lead to two different structural arrests at high enough density or attraction strength, respectively. At the merging of both glassy states, subtle logarithmic time variations appear. Comparing with the recent MCT predictions of these phenomena we find perfect agreement. We thank W. Kob for valuable discussions. M.F. was supported by the Deutsche Forschungsgemeinschaft, grant Fu 309/3, and A.M.P. by the Ministerio de Educación y Cultura. [99]{} Permanent address: Department of Applied Physics, University of Almería, 04.120 Almería, Spain. W. B. Russel, D. A. Saville, and W. R. Schowalter, *Colloidal Dispersions* (Cambridge University Press, New York, 1989). W. van Megen, T.C. Mortensen, S.R. Williams and J. Müller, Phys. Rev. E [**58**]{}, 6073 (1998); and references therein. H. Verduin and J. K. G. Dhont, J. Colloid Interface Sci. [**172**]{}, 425 (1995). J. Bergenholtz and M. Fuchs, Phys. Rev. E [**59**]{}, 5706 (1999). P. N. Segr[è]{}, V. Prasad, A. B. Schofield and D. A. Weitz, Phys. Rev. Lett. [**86**]{}, 6042 (2001). W. C. K. Poon and M. D. Haw, Adv. Colloid Interface Sci. [**73**]{}, 71 (1997); and references therein. M. C. Grant and W. B. Russel, Phys. Rev. E [**47**]{}, 2606 (1993). C. J. Rueb and C. F. Zukoski, J. Rheology [**42**]{}, 1451 (1998). F. Mallamace [*et al.*]{}, Phys. Rev. Lett. [**84**]{}, 5431 (2000). W. C. K. Poon [*et al.*]{}, Faraday Discuss. [**112**]{}, 143 (1999). K. Pham [*et al.*]{}, in preparation (2001). L. Fabbian [*et al.*]{}, Phys. Rev. E [**59**]{}, R1347 (1999); [**60**]{}, 2430 (1999). K. Dawson [*et al.*]{}, Phys. Rev. E [**63**]{}, 011401 (2001). E. Bartsch, V. Frenz, and H. Sillescu, J. Non-Cryst. Solids [**172-174**]{}, 88 (1994). Private communication. J.M. Méndez-Alcaraz and R. Klein, Phys. Rev. E [**61**]{}, 4095 (2000). The potential minimum was set to $a_{12}$ using a parabola for $r<a_{12}+\xi/5$, smoothly connected to the AO. T. Franosch [*et al.*]{}, Phys. Rev. E [**55**]{}, 7153 (1997); M. Fuchs, W. Götze, and M. R. Mayr, Phys. Rev. E [**58**]{}, 3384 (1998); and references therein. The relaxation time was defined by $\Phi_q^s(t=\tau)=0.25$ at $q=3.9$ for the glass transition and $q=9.9$ for gelation. The glass transition was shifted from $\phi_c=0.594$ to $\phi_c=0.55$, upon inclusion of the barrier. No qualitative differences were observed. W. Kob, in [*Soft and Fragile Matter*]{}, edited by M.E. Cates and M.R. Evans (Institute of Physics Publishing, Bristol, 2000), p. 259. W. Götze and L. Sjögren, J. Phys. Condens. Matter [**1**]{}, 4203 (1989). J. L. Barrat, W. Götze and A. Latz, J. Phys. Condens. Matter [**1**]{}, 7163 (1989); M. Fuchs, I. Hofacker and A. Latz, [*Phys. Rev. A*]{} [**45**]{}, 898 (1992).

--- abstract: 'We present experimental and numerical studies of broad-area semiconductor lasers with chaotic ray dynamics. The emission intensity distributions at the cavity boundaries are measured and compared to ray tracing simulations and numerical calculations of the passive cavity modes. We study two different cavity geometries, a D-cavity and a stadium, both of which feature fully chaotic ray dynamics. While the far-field distributions exhibit fairly homogeneous emission in all directions, the emission intensity distributions at the cavity boundary are highly inhomogeneous, reflecting the non-uniform intensity distributions inside the cavities. The excellent agreement between experiments and simulations demonstrates that the intensity distributions of wave-chaotic semiconductor lasers are solely determined by the cavity geometry. This is in contrast to conventional Fabry-Perot broad-area lasers for which the intensity distributions are to a large degree determined by the nonlinear interaction of the lasing modes with the semiconductor gain medium.' author: - Stefan Bittner - Kyungduk Kim - Yongquan Zeng - Qi Jie Wang - Hui Cao title: 'Spatial structure of lasing modes in wave-chaotic semiconductor microcavities' --- Introduction ============ Broad-area semiconductor lasers are commonly employed for high-power applications such as machining, material processing or medical surgery. The typical geometry is a Fabry-Perot cavity with broad cross section of the order of $100~\mu$m, which is necessary to achieve high powers but leads to lasing in several spatial (transverse) modes. The emission intensity distributions are not simply determined by the passive cavity resonances since the nonlinear interactions of the optical field with the gain medium lead to lensing and self-focusing that create spots of high intensity, so-called filaments [@Mehuys1987; @Abraham1990; @Lang1991]. Since the filaments are intrinsically unstable, the lasing emission patterns exhibit spatio-temporal fluctuations on a sub-nanosecond timescale [@Fischer1996; @Hess1996; @Marciante1997; @OhtsuboBook2013]. Intensive efforts [@Takimoto2009; @Simmendinger1999; @Mandre2005; @Gehrig1999; @Adachihara1993] have been made to stabilize the lasing dynamics because a temporally stable beam profile is required for many applications. Recently it has been shown that broad-area semiconductor lasers with D-shaped cavities can suppress the spatio-temporal instabilities from which the conventional Fabry-Perot type broad-area lasers suffer [@Bittner2018a]. In contrast to the regular ray dynamics in a Fabry-Perot cavity, the D-shaped cavity features fully chaotic ray dynamics. Instead of propagating mainly along one axis as in a Fabry-Perot cavity, the rays in the D-shaped cavity travel into all possible directions. Following the principle of ray-wave correspondence, the resonant modes consist of plane wave components with all possible propagation directions, and the resulting complex interference prevents self-focusing and filamentation. However, the lasing emission intensity distributions on the cavity boundary are very inhomogeneous with regions of high as well as very low intensity [@Bittner2018a]. This experimental observation raises the question to what extent the structure of the lasing modes is influenced by the asymmetric cavity geometry and the nonlinear light-matter interaction, respectively. Asymmetric dielectric microcavities have been intensively studied for laser applications [@Tureci2005; @Harayama2010; @Xiao2010b; @Cao2015]. Most asymmetric resonators feature at least partially chaotic ray dynamics and are hence called wave-chaotic cavities. Dielectric resonators are leaky systems because rays can escape refractively, and their properties consequently differ significantly from those of closed cavities. Semiclassical methods [@Brack2003] and ray tracing simulations [@Altmann2013; @Cao2015] have proven very effective to understand and predict their spectral properties and emission directions. Most studies concentrate on the far-field intensity distributions and the phase space representations of the modes (so-called Husimi distributions [@Husimi1940; @Hentschel2003]) instead of the intensity distributions inside the cavities or at the cavity boundaries. Here we focus on the lasing intensity distributions inside the fully chaotic dielectric microcavities and at the cavity boundaries. The degree of spatial localization of lasing modes determines the strength of modal competition for gain and thus the number of lasing modes, with important consequences for the spatial coherence of the emission [@Cao2019; @Cerjan2019]. Although the intra-cavity intensity distributions cannot be easily measured experimentally, the emission profiles at the cavity boundaries allow to draw conclusions about the spatial structure of the lasing modes inside asymmetric cavities [@Lafargue2014; @Bittner2016; @Bittner2018; @Alekseev2018]. Furthermore, knowing the locations of intense emission at the cavity boundary enables efficient coupling into a local waveguide. Our aim is to understand the roles that the cavity geometry and the nonlinear modal interactions play in determining the lasing intensity distributions of wave-chaotic cavities. We fabricate and investigate GaAs quantum well lasers with two different cavity shapes, D-cavities and stadia, both featuring fully chaotic ray dynamics. Although the D-cavities and stadia emit fairly homogeneously in all directions, the intensity distributions inside the cavities and at the cavity boundaries are very inhomogeneous above lasing threshold. Furthermore the coarse structure (i.e., envelope) of the emission intensity distributions at the cavity boundaries is independent of the pump current above lasing threshold and scales with the cavity size, which indicates that the coarse structure is dictated by the cavity shape rather than by the nonlinear light-matter interactions. This is additionally confirmed by numerical calculations of the passive cavity modes and ray tracing simulations which show that the structure of the intensity distributions results from refractive escape of light from the cavity and is completely determined by the passive cavity modes with high quality ($Q$) factor. Moreover, the excellent agreement with ray tracing simulations demonstrates that the principle of ray-wave correspondence [@Tureci2005; @Cao2015] holds for fully chaotic cavities even in the presence of nonlinear interactions between lasing modes and gain medium. The ray tracing calculations accurately predict not only the intensity distributions inside and outside of the cavities but also the quality factors of the most long-lived modes. Such predictions are particularly valuable for cavities that are much larger than the wavelength and are thus not accessible for wave simulations. The article is organized as follows. In Section \[sec:exp\], we present the experiments and the measurement results. Section \[sec:sim\] describes the wave and ray simulations, and in Section \[sec:comp\] we compare the results of wave calculations, ray tracing and experimental measurements. We conclude with a summary and outlook in Section \[sec:conc\]. Experiments {#sec:exp} =========== The edge-emitting semiconductor microlasers are fabricated from a commercial GaAs/AlGaAs quantum well epiwafer (Q-Photonics QEWLD-808) with photolithography and inductively coupled plasma dry etching (see Ref. [@Bittner2018a] for details). The etching depth is about $4~\mu$m to ensure a strong refractive index contrast at the cavity boundary for good optical confinement. The effective refractive index of the cavity is ${n}= 3.37$. We investigate two types of wave-chaotic microcavities. The first one is a D-cavity, which is a circle with a segment cut off \[see [Fig. \[fig:cavGeom\]]{}(a)\]. A D-cavity larger than a semicircle has completely chaotic ray dynamics [@Bunimovich1979; @Ree1999]. Here we consider the D-cavity with the cut $R/2$ away from the center of the circle with radius $R$ because the average Lyapunov exponent of the ray trajectories is approximately the largest for this geometry. The second cavity has the shape of a stadium, which comprises a rectangle between two semicircles \[see [Fig. \[fig:cavGeom\]]{}(b)\]. The ray dynamics in a stadium is completely chaotic as well [@Bunimovich1979], and we consider the stadium with a square between the semicircles so that the average Lyapunov exponents of the ray trajectories is approximately maximized. Experimentally, D-cavities with radii $R = 100$ and $200~\mu$m as well as stadia with $a = 119~\mu$m are investigated, where the stadia with $a = 119~\mu$m have the same area as the D-cavities with $R = 100~\mu$m (cf. appendix \[sec:cavArea\]). The microlasers are pumped electrically with $2~\mu$s-long pulses at a repetition rate of $10-50$ Hz to reduce heating. All experiments are performed at ambient temperature. A $20\times$ microscope objective ($\mathrm{NA} = 0.40$) is used to collect the emission from one of the cavity sidewalls. The emission is coupled into a multimode fiber bundle connected to a spectrometer for measuring the lasing spectrum. For spatial measurements, the objective is used in conjunction with a second lens with $f = 150$ mm in a $2f$-$2f$ configuration to image the emission intensity distributions on the sidewalls on a CCD camera (Allied Vision Mako G125-B, see Ref. [@Bittner2018a] for more details of the setup). The image planes used for D-cavities and stadia are indicated by the blue dashed lines in [Fig. \[fig:cavGeom\]]{}. A long working-distance objective ($\mathrm{NA} = 0.42$) is used to make top view images of the lasers with a second CCD camera (Allied Vision Mako G234-B). Figures \[fig:spectraLI\](a) and \[fig:spectraLI\](b) show typical lasing spectra of a D-cavity and stadium, respectively. The lasers operate in a multimode regime even close to threshold. The polarization of the laser emission is purely transverse electric (TE, electric field parallel to the plane of the cavity). The light-current (LI) curves in Figs. \[fig:spectraLI\](c) and \[fig:spectraLI\](d) show a clear threshold, which is at about $I_{th} = 130$ mA ($100$ mA) for the D-cavity (stadium). These values of the threshold currents are typical and are confirmed for multiple cavities. The top view microscope images of the cavities in Figs. \[fig:topViewNF\](a) and \[fig:topViewNF\](b) show the lasing emission that is scattered in the vertical direction just outside the cavities. The images indicate that the emission intensity distributions of the cavity are very inhomogeneous and exhibit the same mirror symmetries as the cavities. For example, the D-cavity shows almost no emission from the middle of its straight sidewall, whereas its top and bottom third feature strong emission. For the stadium, almost the complete emission originates from the semicircular parts of the boundary, whereas the emission from the straight sidewalls is negligible. The straight sidewall of a D-cavity and the plane touching the semicircle of a stadium (blue dashed lines in [Fig. \[fig:cavGeom\]]{}) are imaged onto a CCD camera to enable a more quantitative measurement of the emission intensity distributions. Figure \[fig:sweid\](a) shows the CCD images of a D-cavity and a stadium pumped well below threshold. Both cavities feature a fairly homogeneous emission intensity distribution. When pumped above threshold, however, the emission profiles of both cavities become very inhomogeneous as shown in [Fig. \[fig:sweid\]]{}(b). Figures \[fig:sweid\](c)–(e) show the emission distributions obtained from the CCD images by integrating in the direction perpendicular to the cavity plane. Figure \[fig:sweid\](c) shows that the emission profiles below threshold are in fact not completely homogeneous, but have very little variation along $x$. However, already just above threshold, the emission distributions are very inhomogeneous as shown in [Fig. \[fig:sweid\]]{}(d). The emission intensity distributions above threshold feature two different length scales: sharp peaks with widths of the order of $1~\mu$m, and a coarse structure (i.e., envelope) that varies on a length scale of several $10~\mu$m. Most notably, the emission distributions exhibit a region of low intensity in the middle of the sidewall for both D-cavity and stadium, a feature which already starts to develop below the lasing threshold \[see [Fig. \[fig:sweid\]]{}(c)\]. When increasing the pump, the fine features of the emission intensity distributions change, whereas the coarse structure stays the same as shown in [Fig. \[fig:sweid\]]{}(e). The intensity distribution of a lasing mode in a wave-chaotic cavity evidently features variations on the scale of the wavelength which results in the sharp peaks. In the experiments, however, their width is limited by the finite numeric aperture of the objective, $\mathrm{NA} = 0.40$, which yields the length scale of $1~\mu$m. The differences in the fine structure that appear with increasing pump are due to the changes of the lasing modes and their relative intensities (cf. [Fig. \[fig:spectraLI\]]{}). The fact that the coarse structure does not change as a function of the pump current indicates that it is determined by a mechanism that is independent of specific lasing modes. The sidewall emission intensity distributions of various D-cavities with different sizes ($R = 100$ and $200~\mu$m) are shown in [Fig. \[fig:sweidsDiffCavs\]]{}. All measured emission profiles exhibit the same coarse structure with a region of very low intensity in the middle flanked by regions of high intensity. Moreover, this coarse structure scales linearly with the cavity size so that the patterns match when plotted as a function of $x/R$ as in [Fig. \[fig:sweidsDiffCavs\]]{}. Analogous results are obtained for the stadia (not shown). While the measurements presented here are integrated over a single current pulse, time-resolved measurements (see Ref. [@Bittner2018a]) demonstrate that the same coarse structure is observed at any given time during a pulse with fluctuations of the fine structure only. In order to complement the information of the intensity distributions measured at the cavity sidewalls which show the origins of intense emission, the directions of emission are measured by scattering the light escaping from the cavities with a ring surrounding them at a distance. The ring has a radius of $300~\mu$m in the case of the D-cavity with $R = 100~\mu$m shown in [Fig. \[fig:MidFieldDistr\]]{}(a) and a radius of $319~\mu$m in the case of the stadium with $a = 119~\mu$m shown in [Fig. \[fig:MidFieldDistr\]]{}(b). Because the radius of the scattering rings is of the same order of magnitude as the cavity sizes for technical reasons, the scattering rings are *not* in the far field. Figures \[fig:MidFieldDistr\](c) and \[fig:MidFieldDistr\](d) shows the top view emission images of the D-cavity and the stadium pumped with $500$ mA, respectively. The image for the D-cavity (stadium) was integrated over $250$ ($300$) pump pulses, and hence the camera is saturated by the emission scattered directly near the cavity sidewalls (cf. [Fig. \[fig:topViewNF\]]{}). The emission intensity scattered at the outer ring is plotted as a function of the azimuthal angle $\varphi$ in Figs. \[fig:MidFieldDistr\](e) and \[fig:MidFieldDistr\](f), respectively. The intensity distributions in [Fig. \[fig:MidFieldDistr\]]{} exhibit the symmetries expected from the cavity geometries, however, there are some perturbations due to the needles (indicated by the gray areas). Furthermore, the presence of scatterers and other defects near the rings leads to artificial peaks in the distributions, e.g., at $\varphi = 75^\circ$ and in the region $\varphi = -30^\circ$ to $0^\circ$ in [Fig. \[fig:MidFieldDistr\]]{}(f). For the D-cavity, the majority of the emission intensity is in the range of $\varphi = \pm (120^\circ$–$150^\circ)$ and in the range of $\varphi = -60^\circ$ to $+60^\circ$. For the stadium, the emission is concentrated in two broad regions around $0^\circ$ and $180^\circ$. Therefore, even though the emission from the cavities is not completely homogeneous in all directions as in the case of a circle cavity, it lacks strong directionality. Like the emission intensity distributions at the cavity boundaries in [Fig. \[fig:sweid\]]{}, the intensity distributions at the scattering rings show little dependence on the pump current above the lasing threshold. Wave and ray simulations {#sec:sim} ======================== The experimental observations in the previous section strongly indicate that the emission intensity distributions are determined not by nonlinear interactions but by the cavity geometry. First, the coarse structure of the emission intensity distributions does not depend on the pump current above the lasing threshold. Second, the same coarse structure is found for different cavities of the same size but different realizsations of surface roughness. Third, the coarse structure scales linearly with the cavity size for the same resonator shape. If, in contrast, the structure of the intensity distributions resulted from the nonlinear interaction with the active medium, it would change with the pump current which increases the strength of the interaction. Moreover, the length scales of the structure would be mainly determined by the details of the light-matter interaction [@Marciante1997] rather than by the cavity size. In order to understand the structure of the lasing modes and how it is influenced by the cavity geometry, we compare the intensity distributions of lasing emission with calculations of the passive cavity modes. Because calculations of the passive modes are only feasible for cavity sizes significantly smaller than the experimental ones, we perform ray tracing simulations in order to obtain the relation between the cavity geometry and the structure of the intensity distributions in the semiclassical limit. This relation will be directly applicable to the experimental cavities as ray tracing simulations are scale free. Simulations of passive cavity modes ----------------------------------- We calculate the passive cavity modes (also called resonances) of the D-cavities and stadia by solving the two-dimensional scalar Helmholtz equation $$[ \Delta + n^2(x, y) k^2 ] \Psi(x, y) = 0$$ with outgoing wave boundary conditions [@Tureci2005], where ${n}$ is the effective refractive index and $k = 2 \pi / \lambda$ is the wave number with $\lambda$ the free-space wavelength. Here the wave function $\Psi$ corresponds to the $z$-component of the magnetic field, $H_z$, as experimentally the lasing emission is TE polarized. The modes are calculated numerically using the Comsol eigenmode solver. The refractive index of the cavity is set to ${n}= 3.37$ and the free-space wavelength is around $800$ nm as in the experiments. Calculations are performed for a D-cavity with $R = 20~\mu$m, which corresponds to $kR \simeq 157$, and a stadium with $a = 23.8~\mu$m, which corresponds to $ka \simeq 187$. Both cavities have the same area, but have $5$–$10$ times smaller linear dimensions than those used in the experiments due to the restrictions of available computing power. However, they are quite far in the semiclassical limit $kR \gg 1$, and it was checked that simulations with cavities with half the linear dimensions yield the same qualitative results. Moreover, the wave calculation results agree well with the ray tracing simulations which confirms that they can be compared to experimental data. Figure \[fig:spectraDecay\] shows the calculated spectra, which consist of $61$ resonances for the D-cavity and $106$ for the stadium. Only the modes with the highest quality factors in the given wavelength range are calculated, whereas additional modes with lower $Q$-factors exist but are of no relevance here since they have no realistic chance of lasing. The most long-lived mode of the D-cavity has ${Q_\mathrm{max}}= 3699$ \[${\mathrm{Im}}(kR) = -0.0212$\], whereas the modes of the stadium can have significantly higher $Q$-factors up to ${Q_\mathrm{max}}= 7137$ \[${\mathrm{Im}}(ka) = -0.0131$\]. This raises the question what determines and limits the highest $Q$-factor for a given cavity and how ${Q_\mathrm{max}}$ depends on the cavity geometry, size and refractive index, which is hard to determine with wave simulations alone. Ray dynamics simulations ------------------------ In order to answer these questions, we performed ray tracing simulations of dielectric billiards with D-cavity and stadium shape. The ray simulations allow to explore the semiclassical regime beyond the cavity sizes that can be treated with wave simulations, and moreover the comparison of wave and ray simulations enables us to distinguish between effects due to the classical ray dynamics and wave effects such as tunneling and scarring [@Heller1984]. As shown in the following, the ray simulations can explain both the intensity distributions and the different $Q$-factors (or lasing thresholds) of the D-cavities and stadia. For closed cavities, the semiclassical eigenfunction hypothesis states that the wave functions are supported by the regions of phase space explored by typical classical trajectories in the semiclassical limit, that is, when the cavity size is much larger than the wavelength [@Berry1977]. This implies that the average intensity distributions of resonators with fully chaotic (and thus ergodic) ray dynamics become uniform [@Shnirelman1974; @ColindeVerdiere1985; @Zelditch1996]. However, the dielectric resonators considered here are open systems since rays can escape by refraction at the dielectric interfaces that form the cavity sidewalls. Hence, in the classical limit of ray optics, they are leaking Hamiltonian systems [@Altmann2013], which results in non-uniform intensity distributions. At the moment, the generalization of the semiclassical eigenfunction hypothesis to this case remains an unsolved problem [@Claus2018]. Nonetheless, some properties of the wave functions are known for dielectric resonators with integrable or chaotic ray dynamics. For the integrable case, the resonant modes localize on classical tori [@Bittner2013b; @Bittner2014a; @Sukharevsky2019]. For chaotic ray dynamics, it has been shown that the modes with the highest $Q$-factors are based on a particular set of trajectories, the unstable manifold of the chaotic saddle [@Altmann2013; @Casati1999b; @Schwefel2004; @Lee2004; @Wiersig2008]. The chaotic saddle consists of the trajectories that stay confined forever in the cavity both for forward and backward propagation in time, and the unstable manifold is the set converging to the chaotic saddle for backward propagation[^1]. In essence, the unstable manifold consists of the most long-lived trajectories that eventually escape from the cavity (for forward propagation in time), and for this reason is closely related to the high-$Q$ modes. Ray tracing simulations of the unstable manifold have been used successfully to predict for example the far-field intensity distributions of wave-chaotic cavities, and here we extend this method for the first time to the intra-cavity and emission intensity distributions. It should be noted that the most general and frequent case of partially chaotic and regular (i.e., mixed) ray dynamics presents additional complications and is beyond the scope of this article. We calculate the unstable manifold of the chaotic saddle using the so-called sprinkler method [@Schneider2002]. A large ensemble of trajectories with unit intensity that are distributed uniformly in phase space are launched and their evolution inside the billiard as well as the decay of their intensity due to refractive escape is calculated as a function of time. The algorithm is further explained in Appendix \[sec:algoRayTrace\]. The decay of the total intensity $I(t)$ of the trajectories remaining inside the billiard is shown in Figs. \[fig:spectraDecay\](c) and \[fig:spectraDecay\](d) for the D-cavity and the stadium, respectively. After an initial transient period, the total intensity in both cavities decays exponentially with $I(t) \propto \exp(-t / {\tau_\mathrm{cl}})$. We call ${\tau_\mathrm{cl}}$ the classical photon lifetime[^2] since it represents the average lifetime of photons in the cavity in the long-time limit and thus the maximal lifetime of resonances in the semiclassical limit. In this regime, the intensity distribution of the rays inside the billiard becomes conditionally invariant, that is, it no longer changes in time except for an overall exponential decay [@Claus2018; @Lee2004; @Ryu2006a], and the remaining trajectories represent a good approximation of the unstable manifold of the chaotic saddle. A representation of the unstable manifold in a Poincaré surface of section (PSOS) of phase space is shown in Appendix \[app:PSOS+FF\]. The intra-cavity and emission intensity distributions during the regime of exponential decay are calculated by averaging over all rays during the time interval $29.7$–$44.5~nR/c$ for the D-cavity ($29.7$–$44.5~na/c$ for the stadium). The exact time interval is not important since the ray distributions are conditionally invariant during the exponential decay regime. However, since the evolution of the ray trajectories between reflections instead of just the reflections at the boundaries needs to be tracked to obtain the intra-cavity intensity distributions, the time interval should be longer than several mean scattering times $\left< t_s \right> = \left< l_s \right> n/c$, where $\left< l_s \right>$ is the average distance between two reflections at the boundary. The mean scattering times are $\left< t_s \right> = 1.341~nR/c$ for the D-cavity and $\left< t_s \right> = 1.091~na/c$ for the stadium (see Appendix \[sec:cavArea\]). In addition, the finite $\mathrm{NA}$ of the imaging system is taken into account to properly compare the simulated emission profiles with the measured ones. More details are given in Appendix \[sec:algoRayTrace\]. Comparison of experiments, wave and ray simulations {#sec:comp} =================================================== Photon lifetimes and thresholds ------------------------------- First, we consider the lifetimes of the high-$Q$ modes and the lasing thresholds. The photon lifetime predicted by the ray tracing simulations is ${\tau_\mathrm{cl}}= 6.3507~{n}R / c$ (${\tau_\mathrm{cl}}= 7.1736~{n}a /c$) for the D-cavity (stadium) as shown in [Fig. \[fig:spectraDecay\]]{}. The predicted classical quality factor, ${Q_\mathrm{cl}}= k c {\tau_\mathrm{cl}}$ where $c$ is the speed of light in vacuum, for a D-cavity with $R = 20~\mu$m (stadium with $a = 23.8~\mu$m) at $\lambda = 800$ nm is ${Q_\mathrm{cl}}^\mathrm{(D)} = 3362$ (${Q_\mathrm{cl}}^\mathrm{(Stad)} = 4519$). The lifetimes predicted by ray tracing are compared to the calculated resonance wavenumbers in Figs. \[fig:spectraDecay\](a) and \[fig:spectraDecay\](b), where the corresponding values of ${\mathrm{Im}}(k) = -1 / (2 c {\tau_\mathrm{cl}})$ are indicated as dashed horizontal lines. Since the ray trajectories remaining in the cavity converge to the set of most long-lived trajectories in time, their lifetime ${\tau_\mathrm{cl}}$ should be an upper limit for the lifetimes of the high-$Q$ modes of the resonator in the semiclassical limit [@Harayama2015]. For the same reason, the intensity distribution of this set of ray trajectories should predict those of the high-$Q$ modes. For the D-cavity, ${\tau_\mathrm{cl}}$ is indeed in good agreement with the lifetimes of the most long-lived modes \[see [Fig. \[fig:spectraDecay\]]{}(a)\], with only a few modes slightly exceeding ${\tau_\mathrm{cl}}$. In the case of the stadium shown in [Fig. \[fig:spectraDecay\]]{}(b), however, there are several modes with clearly longer lifetimes than ${\tau_\mathrm{cl}}$. Such modes are hence called supersharp resonances [@Novaes2012]. The existence of supersharp resonances is explained by a higher degree of localization compared to other high-$Q$ modes [@Casati1999b; @Novaes2012], for example due to scarring (i.e., localization) [@Heller1984; @Gmachl2002; @Harayama2003] on short unstable periodic orbits (UPOs) confined by total internal reflection. However, since the average scarring strength decreases in the semiclassical limit of large resonators [@Vergini2012; @Vergini2015], the lifetimes of supersharp resonances converge to ${\tau_\mathrm{cl}}$ in this limit [@Novaes2012]. This has been verified numerically for the stadium billiard [@Cerjan2019], and we do not expect any supersharp resonances for the five to ten times larger cavities investigated experimentally. It is interesting to note that there are no supersharp resonances for the D-cavity in [Fig. \[fig:spectraDecay\]]{}(a), which does not feature modes with strong scarring either. Experimentally we observe that the stadium cavities have consistently lower lasing thresholds than the D-cavities with the same area. For the cavities presented in [Fig. \[fig:spectraLI\]]{}, the threshold of the stadium is about $1.3$ times lower than that of the D-cavity, and the average over several D-cavities with $R = 100~\mu$m and stadia with $a = 119~\mu$m yields a ratio of $1.25$ [@Cerjan2019]. This agrees with the ray tracing simulations that predict a longer photon lifetime for the stadium, but the ratio of ${Q_\mathrm{cl}}^\mathrm{(Stad)} / {Q_\mathrm{cl}}^\mathrm{(D)} \simeq 1.34$ is somewhat higher than the ratio of the measured thresholds. In practice, additional loss mechanisms such as surface roughness play a role. While the quality of the cavity sidewalls is very good (see Refs. [@Bittner2018a; @Cerjan2019]), surface roughness is not negligible. Since surface roughness tends to reduce the difference in the $Q$-factors for different cavity shapes, it can explain the slightly lower ratio of the thresholds of D-cavities and stadia found experimentally. Interior intensity distributions -------------------------------- Next we consider the interior intensity distributions and compare the ray tracing with the wave simulations. The intra-cavity intensity distributions of several high-$Q$ modes of the D-cavity and the stadium are shown in Figs. \[fig:IID-Dcav\](a)–(d) and Figs. \[fig:IID-Stad\](a)–(d), respectively. While their fine structure is clearly different, their coarse structure shows common features. In particular, the modes of the D-cavity in Figs. \[fig:IID-Dcav\](a)–(d) feature a roughly circular region in the center of the cavity with significantly lower intensity. Similarly, two circular regions of low intensity at the centers of the circular arcs are found for the modes of the stadium in Figs. \[fig:IID-Stad\](a)–(d). We add the intra-cavity intensity distributions of high-$Q$ modes because experimentally we observe the total emission from all lasing modes. The modes are added in intensity since the lasing modes are phase incoherent. The summation highlights the common features of the intensity distributions of individual modes, and better agreement with ray tracing simulations is obtained [@Choi2008; @Shinohara2008; @Harayama2015]. We sum the field intensities of all modes with $Q \geq 0.9~{Q_\mathrm{cl}}$, which are $11$ modes for the D-cavity and $24$ modes for the stadium, respectively. These modes are highlighted by the gray areas in Figs. \[fig:spectraDecay\](a) and \[fig:spectraDecay\](b). The threshold of $0.9~{Q_\mathrm{cl}}$ was chosen since ${Q_\mathrm{cl}}$ gives the scale of what can be considered to be a high-$Q$ mode for a given cavity geometry, but the results do not depend sensitively on the chosen threshold. The intensity distributions are normalized such that the integral over the interior of the cavity, $\int_S |\Psi(x, y)|^2 dA$, is equal to $1$ before adding them. The total intra-cavity intensity distributions are presented in [Fig. \[fig:IID-Dcav\]]{}(e) and [Fig. \[fig:IID-Stad\]]{}(e) for the D-cavity and the stadium, respectively. The circular regions of low intensity are even more pronounced in the sum of intensity distributions. Even outside these regions, the total intensity distributions are not homogeneous and exhibit points and lines of high intensity. The intra-cavity intensity distributions resulting from the ray-tracing simulations are shown in [Fig. \[fig:IID-Dcav\]]{}(f) and [Fig. \[fig:IID-Stad\]]{}(f), respectively, and show very good agreement with the total intensity distributions of the calculated modes. The circular regions of low intensity as well as many other features are accurately predicted by the ray tracing simulations. The most prominent feature of the intra-cavity intensity distributions, the circular regions of lower intensity, are naturally explained by the classical ray dynamics. Trajectories with an angle of incidence larger than the critical angle for total internal reflection, ${\chi_\mathrm{cr}}$, are obviously among the most long-lived orbits and contribute significantly to the unstable manifold. The trajectories that hit the circular boundaries exactly at the critical angle form a caustic with radius $R / {n}$ indicated by the dashed white circles in Figs. \[fig:IID-Dcav\](f) and \[fig:IID-Stad\](f). They delimit precisely the regions of low intensity, and it was checked that the radius of the low-intensity regions scales indeed as $1 / {n}$ as a function of the refractive index ${n}$. Even though no experimental data for the intra-cavity intensity distributions are available for comparison, the good agreement of ray and wave simulations of the intensity distribution validates the ray tracing approach. Moreover, the structure of the intra-cavity distributions and in particular the caustics allow us to better understand the emission intensity distributions discussed in the following. Emission intensity distributions -------------------------------- Wave simulations of the emission intensity distributions at the straight sidewall of a D-cavity are shown in [Fig. \[fig:SWEID-Dcav-sim\]]{} together with the sum of the interior intensity distributions for all modes with $Q \geq 0.9~{Q_\mathrm{cl}}$. The emission intensity distributions are obtained from the calculated wave functions by applying a Fourier filter with width corresponding to $\mathrm{NA} = 0.4$ in order to account for the imaging optics (cf. Ref. [@Bittner2018a]). Figure \[fig:SWEID-Dcav-sim\](b) shows the emission profiles of the $11$ individual modes. While there are clear differences between their intensity distributions, all of them exhibit a region of low intensity in the center, and most have peaks at the border of the center region and near the corners at $x = \pm \sqrt{3} R / 2$. These common features are responsible for the structure of the total emission intensity distribution in [Fig. \[fig:SWEID-Dcav-sim\]]{}(c), which features a region of low intensity surrounded by two peaks and two further peaks near the corners as observed experimentally. It is due to the common features of the modes that the coarse structure of the measured intensity distributions is independent of the pump current and the active lasing modes. The region of low intensity in the center results from the caustic with radius $R / {n}$, shown as dashed white circle in [Fig. \[fig:SWEID-Dcav-sim\]]{}(a). The projection of the caustic is indicated by the horizontal dashed lines in [Fig. \[fig:SWEID-Dcav-sim\]]{} and delimits the center region of low intensity quite precisely. The caustic can be seen so clearly in the emission intensity distributions because rays impinging on the straight sidewall with near normal incidence contribute most to the emission distribution due the finite angle of collection of the objective with $\mathrm{NA} = 0.4$, and there are practically no long-lived rays that go through the caustic and hit the straight segment perpendicularly. Figure \[fig:SWEID-Dcav-sim\](a) also shows that the caustic is surrounded by regions with high intensity above and below, which explains the two peaks surrounding the center region in the emission profiles. In the same way, the features with relatively high intensity near the corners in [Fig. \[fig:SWEID-Dcav-sim\]]{}(a) explain the peaks of emission intensity at the two ends of the straight segment. Similar arguments apply to the emission intensity distributions of the stadium which also exhibit a region of low intensity in the middle even though it is not as pronounced as for the D-cavity. Next we compare the emission intensity distributions from ray and wave simulations with the experimentally measured ones. Figures \[fig:sweid\](e), \[fig:sweid\](f) and \[fig:sweid\](g) show the emission profiles measured experimentally well above threshold, calculated by ray tracing and obtained from wave simulations, respectively. The ray and wave simulations show very good agreement with the measured profiles for both cavity geometries. It should be noted that the smallest feature sizes of the emission intensity distributions from wave calculations with a width approximately given by the resolution limit, $\lambda / (2 \, \mathrm{NA})$, appear broader than those of the measured intensity distributions because the distributions are presented as function of the transverse coordinate normalized by the cavity size, and the experimental cavities are five times larger. Finally we compare the top view emission intensity distributions just outside of the cavities and at the rings surrounding the cavities with the ray tracing simulations. Figures \[fig:topViewNF\](a) and \[fig:topViewNF\](b) show the top view emission images around a D-cavity and a stadium respectively, and Figs. \[fig:topViewNF\](c) and \[fig:topViewNF\](d) the corresponding ray simulations of the intensity just outside the cavities. These images highlight from which parts of the boundary the emission originates and are in good agreement. Figures \[fig:MidFieldDistr\](e) and \[fig:MidFieldDistr\](f) show the emission scattered at the rings surrounding the D-cavity and stadium, respectively. Measurements and ray simulations agree well, demonstrating that also the emission directions of the cavities are accurately predicted by our simulations. Summary and outlook {#sec:conc} =================== The excellent agreement of the ray tracing results with the experimental data and passive cavity wave simulations confirms that the ray simulations can predict precisely the intensity distributions inside and outside the D-cavity and stadium microlasers. Because the ray and wave simulations — which do not take nonlinear effects into account — agree so well with experiments, we conclude that the nonlinear interaction of modes and active medium does not have a perceivable influence on the structure of the intensity distributions, in contrast to the case of broad-area Fabry-Perot lasers [@Mehuys1987; @Abraham1990; @Lang1991]. Also the influence of surface roughness appears to be negligible, and the inhomogeneous intensity distributions are dictated by the geometry of the cavities alone. The ability of ray tracing simulations to comprehensively predict the intra-cavity and exterior intensity distributions as well as photon lifetimes of cavities with fully chaotic ray dynamics is very useful to estimate, e.g., the emission directionality, the lasing thresholds or the strength of modal interaction. They hence provide a computationally efficient tool for the design of asymmetric microcavity lasers, in particular in the semiclassical limit where cavities are often too large for a numerical solution of the Helmholtz equation. However, ray tracing simulations can only predict the intensity distribution of a sum of many lasing modes, not of individual modes. Furthermore, the cavity needs to be much larger than the wavelength so that the high-$Q$ modes are indeed based on the unstable manifold of the chaotic saddle as assumed for the ray tracing simulations. Especially for smaller cavities, significant variations from mode to mode can occur [@Choi2008], and in particular when strong scarring on UPOs is involved [@Fang2007b]. More work is needed to understand the regime of validity of the ray tracing predictions as far as the photon lifetimes are concerned. In the semiclassical limit, the photon lifetimes of the fully chaotic stadium and D-cavities are predicted by the time constant ${\tau_\mathrm{cl}}$ of the exponential decay of the rays remaining in the cavities. Exponential decay in ray tracing simulations was also observed in Refs. [@Lee2004; @Ryu2006a; @Shinohara2007], and good agreement of ${\tau_\mathrm{cl}}$ with the lifetimes of the high-$Q$ modes was demonstrated in Ref. [@Shinohara2007]. However, chaotic dielectric billiards do not always exhibit exponential decay, and we observe non-exponential decay for D-cavities with a smaller section cut off (i.e., a cut farther away from the center). This is attributed to the existence of families of marginally unstable periodic orbits (MUPOs) [@Altmann2013; @Dettmann2009]. For smaller cuts than at $R/2$, the D-cavities feature an increasingly large family of equilateral triangle orbits as well as other polygonal MUPO families. Since these orbits are confined by total internal reflection, they and nearby trajectories contribute significantly to the long-term decay of the system. For the D-cavity with cut at $R/2$ considered here, however, the only MUPO family consists of the orbits along the diameter, which have a short lifetime due to their normal incidence at the cavity boundary. Similarly, for the stadium the only MUPO family consists of the so-called bouncing ball orbits between the straight sides of the stadium which are very short-lived as well. In conclusion, for the D-cavity and the stadium geometries considered here, the MUPO families have no contribution to the exponential decay of the rays remaining inside the cavities. In general, the estimation of the photon lifetime for fully chaotic billiards with long-lived MUPO families will not be as straightforward as for the cases considered here. Because the ray tracing simulations only take geometric effects and refractive escape into account, they predict that the $Q$-factors scale linearly with the linear cavity size. The very good agreement of the ray simulations with the experimental results shows that refractive escape is indeed the dominant decay mechanism for the optical field in the cavities considered here. In general, however, various wave effects can also contribute significantly to the losses, and their importance depends sensitively on the ratio of cavity size to wavelength. Hence a quantitative understanding of all possible effects involved requires further studies. For example, scarring can result in unusually long-lived resonances [@Novaes2012; @Casati1999b] and interference between scar contributions from different UPOs may yield significant fluctuations of the $Q$-factors as a function of cavity size and geometry [@Fang2005a; @Fang2007a]. However, the scarring strength decreases in the semiclassical limit [@Vergini2012; @Vergini2015], and also the contributions from other wave effects such as tunneling loss at curved interfaces diminish for increasingly large cavities. While we consider only cavities with fully chaotic ray dynamics here, most asymmetric microcavities feature mixed ray dynamics, i.e., they exhibit both chaotic and integrable regions in phase space. Dielectric billiards with mixed dynamics usually feature non-exponential decay of the ray trajectories inside [@Ryu2006a]. Moreover, contributions from dynamical tunneling between the integrable and chaotic regions of phase space need to be taken into account to predict the lifetimes of their resonances [@Baecker2009; @Gehler2015]. Therefore, while ray tracing simulations have proven themselves as a powerful tool to understand and predict the properties of asymmetric microcavities, a complete understanding of the physical mechanisms that determine the intensity distributions and lifetimes for arbitrary cavity geometries and accurate predictions based on semiclassical methods remain an important future challenge. The authors thank A. D. Stone, M. Constantin, O. Hess and T. Harayama for fruitful discussion. The work at Yale is supported partly by the Office of Naval Research (ONR) with MURI grant N00014-13-1-0649, and the Air Force Office of Scientific Research (AFOSR) under grant FA9550-16-1-0416. Y. Z. and Q. J. W. acknowledge support from the Ministry of Education, Singapore (grants MOE2016-T2-1-128 and MOE2016-T2-2-159) and the National Research Foundation, Competitive Research Program (NRF-CRP18-2017-02). Cavity area and mean free path length {#sec:cavArea} ===================================== The area of the D-cavity shown in [Fig. \[fig:cavGeom\]]{}(a) is $S = R^2 ( 2 \pi / 3 + \sqrt{3} / 4) \simeq 2.527 \, R^2$ and its circumference $\partial S = R (4 \pi / 3 + \sqrt{3}) \simeq 5.921 \, R$. This yields an area of $S = 25,274~\mu\mathrm{m}^2$ for $R = 100~\mu$m. The stadium shown in [Fig. \[fig:cavGeom\]]{}(b) has an area of $S = a^2 (1 + \pi/4)$ and a circumference of $\partial S = a (2 + \pi)$, yielding an area of $S = 25,283~\mu\mathrm{m}^2$ for $a = 119~\mu$m. The mean free path length in a two-dimensional billiard with ergodic dynamics is given by (see Ref. [@Chernov1997] and references therein) $$\left< l_s \right> = \frac{\pi S}{\partial S} \, .$$ This yields $\left< l_s \right> \simeq 1.341 \, R$ for the D-cavity and $\left< l_s \right> \simeq 1.091 \, a$ for the stadium. Ray-tracing algorithm {#sec:algoRayTrace} ===================== For the ray tracing simulations, an ensemble of $10^7$ trajectories are launched with random initial positions uniformly distributed inside the cavity (not just at the boundary) and random initial directions uniformly distributed in the interval $[0, 2 \pi)$. All rays start with unit intensity, and their intensity is reduced according to the Fresnel coefficients for ${n}= 3.37$ and p-polarization at each reflection. The actual time of flight is tracked (instead of just the number of reflections) as recommended in Ref. [@Altmann2013]. The intensity profiles inside and outside of the cavities are calculated by sampling the positions and intensities of all rays during the time interval $[29.7$–$44.5]~nR/c$ ($[29.7$–$44.5]~na/c$) for the D-cavity (stadium) with a spatial resolution of $10^{-3}~R$ ($10^{-3}~a$). The far-field intensity distributions and the intensity distributions at the ring surrounding the cavities are sampled analogously as a function of the azimuthal angle $\varphi$. When calculating the emission intensity distributions, only rays with $|\sin(\alpha)| \leq \mathrm{NA}$ are considered, where $\alpha$ is the angle of the outgoing ray with respect to the optical axis of the imaging optics and $\mathrm{NA} = 0.4$ is the numerical aperture of the objective used in the experiments. Phase space and far-field distributions {#app:PSOS+FF} ======================================= The unstable manifold of the chaotic saddle in a PSOS of phase space is shown in [Fig. \[fig:PSOS\]]{}. The PSOS is parametrized in the so-called Birkhoff coordinates [@Birkhoff1927] defined in Figs. \[fig:PSOS\](a) and \[fig:PSOS\](b), where $s$ is the position along the boundary of the billiard and $\chi$ the angle of incidence. The boundaries of the lossy region in phase space with $|\sin(\chi)| < 1 / {n}$ are indicated by the horizontal white lines. Only the reflections in the lossy region contribute to the emission intensity distributions, and the intensity in it is significantly lower compared to the regions confined by total internal reflection, $|\sin(\chi)| \geq 1 / n$. The ray tracing calculations of the far-field intensity distributions are shown in [Fig. \[fig:farfield\]]{}. While the far-field distributions are not uniform since the cavities are asymmetric, their emission is far from being directional, and both D-cavity and stadium exhibit significant emission into almost all directions. This can be partially attributed to the fact that the unstable manifolds shown in [Fig. \[fig:PSOS\]]{} cover a fairly large part of the leaky region and thus contribute to many different emission angles. [61]{}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 10.1109/JQE.1987.1073261) [****, ()](\doibase 10.1364/JOSAB.7.000951) [****,  ()](\doibase 10.1109/3.81329) [****,  ()](\doibase 10.1209/epl/i1996-00154-7) [****,  ()](\doibase 10.1103/PhysRevA.54.3360) [****,  ()](\doibase 10.1109/3.594881) [**](\doibase 10.1007/978-3-642-30147-6),  ed. (, , ) [****, ()](\doibase 10.1109/LPT.2009.2022181) [****, ()](\doibase 10.1364/OE.5.000048) [****, ()](\doibase 10.1016/j.optcom.2004.09.058) [****,  ()](\doibase 10.1109/3.748837) [****,  ()](\doibase 10.1364/JOSAB.10.000658) [****,  ()](\doibase 10.1126/science.aas9437) [****,  ()](\doibase 10.1016/S0079-6638(05)47002-X) [****,  ()](\doibase 10.1002/lpor.200900057) [****,  ()](\doibase 10.1007/s12200-010-0003-2) [****,  ()](\doibase 10.1103/RevModPhys.87.61) @noop [**]{} (, , ) [****, ()](\doibase 10.1103/RevModPhys.85.869) [****,  ()](\doibase 10.11429/ppmsj1919.22.4_264) [****,  ()](\doibase 10.1209/epl/i2003-00421-1) [****,  ()](\doibase 10.1038/s42254-018-0010-6) @noop [ ()]{} [****,  ()](\doibase 10.1103/PhysRevE.90.052922) [****,  ()](\doibase 10.1209/0295-5075/113/54002) [****,  ()](\doibase 10.1103/PhysRevA.97.043826) [****,  ()](\doibase 10.1364/OE.26.014433) [****,  ()](\doibase 10.1007/BF01197884) [****, ()](\doibase 10.1103/PhysRevE.60.1607) [****,  ()](\doibase 10.1103/PhysRevLett.53.1515) @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1007/BF01209296) [****,  ()](\doibase 10.1007/BF02099513) @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevE.88.062906) [****,  ()](\doibase 10.1103/PhysRevE.90.052909) [****,  ()](\doibase 10.1117/1.OE.58.1.016115) [****, ()](\doibase 10.1016/S0167-2789(98)00265-6) [****,  ()](\doibase 10.1364/JOSAB.21.000923) [****,  ()](\doibase 10.1103/PhysRevLett.93.164102) [****,  ()](\doibase 10.1103/PhysRevLett.100.033901) [****,  ()](\doibase 10.1103/PhysRevE.66.066218) [****,  ()](\doibase 10.1103/PhysRevE.73.036207) [****,  ()](\doibase 10.1103/PhysRevE.92.042916) [****,  ()](\doibase 10.1103/PhysRevE.85.036202) [****,  ()](\doibase 10.1364/OL.27.000824) [****,  ()](\doibase 10.1103/PhysRevE.67.015207) [****,  ()](\doibase 10.1103/PhysRevLett.108.264101) [****,  ()](\doibase 10.1209/0295-5075/110/10010) [****,  ()](\doibase 10.1364/OE.16.017554) [****,  ()](\doibase 10.1103/PhysRevA.77.033807) @noop [ ()]{} [****,  ()](\doibase 10.1103/PhysRevE.75.036216) [****,  ()](\doibase 10.1016/j.physd.2009.09.019) [****,  ()](\doibase 10.1103/PhysRevA.72.023815) [****,  ()](\doibase 10.1063/1.2762285) [****,  ()](\doibase 10.1103/PhysRevA.79.063804) [****,  ()](\doibase 10.1103/PhysRevLett.115.104101) [****, ()](\doibase 10.1007/BF02508462) [****, ()](\doibase 10.1007/BF02421325) [^1]: The unstable manifold is hence also called backward-trapped set. [^2]: The inverse, $1 / {\tau_\mathrm{cl}}$, is called natural decay rate in Ref. [@Claus2018].

--- abstract: | We show that if the Kobayashi–Royden metric of a complex manifold is continuous and positive at a given point and any non-zero tangent vector, then the “derivatives” of the higher order Lempert functions exist and equal the respective Kobayashi metrics at the point. It is a generalization of a result by M. Kobaya-\ shi for taut manifolds. address: - | Institute of Mathematics and Informatics\ Bulgarian Academy of Sciences\ Acad. G. Bonchev 8, 1113 Sofia, Bulgaria - | Carl von Ossietzky Universität Oldenburg\ Institut für Mathematik, Fakultät V\ Postfach 2503\ D-26111 Oldenburg, Germany author: - Nikolai Nikolov and Peter Pflug title: On the derivatives of the Lempert functions --- [This note was written during the stay of the first named author at the Universität Oldenburg supported by a grant from the DFG, Az. PF 227/8-2 (November – December 2006). He likes to thank both institutions for their support.]{} Introduction and results ======================== Let $\Bbb D\subset\Bbb C$ be the unit disc. Let $M$ be an $n$-dimensional complex manifold. Recall first the definitions of the Lempert function $\tilde k_M$ and the Kobayashi–Royden pseudometric $\kappa_M$ of $M$: $$\aligned \tilde k^\ast_M(z,w)&=\inf\{|\alpha|:\exists f\in\mathcal O(\Bbb D,M):f(0)=z,f(\alpha)=w\},\\ \tilde k_M&=\tanh^{-1}\tilde k^\ast_M,\\ \kappa_M(z;X)&=\inf\{|\alpha|:\exists f\in\mathcal O(\Bbb D,M): f(0)=z,\alpha f_\ast(d/d\zeta)=X\}, \endaligned$$ where $X$ is a complex tangent vector to $M$ at $z$. Note that such an $f$ always exists (cf. [@Win]; according to [@Din], page 49, this was already known by J. Globevnik). The Kobayashi pseudodistance $k_M$ can be defined as the largest pseudodistance bounded by $\tilde k_M$. Note that if $k_M^{(m)}$ denotes the $m$-th Lempert function of $M$, $m\in\Bbb N$, that is, $$k_M^{(m)}(z,w)=\inf\{\sum_{j=1}^m\tilde k_M(z_{j-1},z_{j}):z_0,\dots,z_{m}\in M, z_0=z,z_m=w\},$$ then $$k_M(z,w)=k_M^{(\infty)}:=\inf_m k_M^{(m)}(z,w).$$ By a result of M.-Y. Pang (see [@Pang]), the Kobayashi–Royden metric is the “derivative” of the Lempert function for taut domains in $\Bbb C^n;$ more precisely, if $D\subset\Bbb C^n$ is a taut domain, then $$\kappa_D(z;X)=\lim_{t\nrightarrow 0}\frac{\tilde k_D(z,z+tX)}{t}.$$ In [@KobS], S. Kobayashi introduces a new invariant pseudometric, called the Kobayashi–Buseman pseudometric in [@Jar-Pfl]. One of the equivalent ways to define the Kobayashi–Buseman pseudometric $\hat\kappa_M$ of $M$ is just to set $\hat\kappa_M(z;\cdot)$ to be largest pseudonorm bounded by $\kappa_M(z;\cdot)$. Recall that $$\hat\kappa_M(z;X)=\inf\{\sum_{j=1}^m\kappa_M(z;X_j):m\in\Bbb N,\ \sum_{j=1}^mX_j=X\}.$$ Thus it is natural to consider the new function $\kappa_M^{(m)}$, $m\in\Bbb N$, namely, $$\kappa_M^{(m)}(z;X)=\inf\{\sum_{j=1}^m\kappa_M(z;X_j): \sum_{j=1}^mX_j=X\}.$$ We call $\kappa_M^{(m)}$ the [*$m$-th Kobayashi pseudometric*]{} of $D$. It is clear that $\kappa_M^{(m)}\ge\kappa_M^{(m+1)}$ and if $\kappa_M^{(m)}(z;\cdot)=\kappa_M^{(m+1)}(z;\cdot)$ for some $m$, then $\kappa_M^{(m)}(z;\cdot)\\=\kappa_D^{(j)}(z;\cdot)$ for any $j>m$. It is shown in [@Nik-Pfl] that $\kappa_M^{(2n-1)}=\kappa_M^{(\infty)}:=\hat\kappa_M$, and $2n-1$ is the optimal number, in general. We point out that all the introduced objects are upper semicontinuous. Recall that this is true for $\kappa_M$ (cf. [@KobS2]). It remains to check this for $\tilde k_M$. We shall use a standard reasoning. Fix $r\in(0,1)$ and $z,w\in M$. Let $f\in\mathcal O(\Bbb D,M)$, $f(0)=z$ and $f(\alpha)=w$. Then $\tilde f=(f,\hbox{id}):\Delta\to\tilde M=M\times\Delta$ is an embedding. Setting $\tilde f_r(\zeta)=\tilde f(r\zeta)$, by [@Roy], Lemma 3, we may find a Stein neighborhood $S\subset\tilde M$ of $\tilde f_r(\Bbb D)$. Embed $S$ as a closed complex manifold in some $\Bbb C^N$ and denote by $\psi$ the respective embedding. Moreover, there is an open neighborhood $V\subset \Bbb C^N$ of $\psi(S)$ and a holomorphic retraction $\theta:V\to\psi(S)$. Then, for $z'$ near $z$ and $w'$ near $w$, we may find, as usual, $g\in\mathcal O(\Bbb D,V)$ such that $g(0)=\psi(z',0)$ and $g(\alpha/r)=\psi(w',\alpha)$. Denote by $\pi$ the natural projection of $\tilde M$ onto M. Then $h=\pi\circ\psi^{-1}\circ\theta\circ g\in\mathcal O(\Bbb D,M)$, $h(0)=z'$ and $h(\alpha/r)=w'$. So $r\tilde k^\ast_M(z',w')\le\alpha$, which implies that $\ds\limsup_{z'\to z,w'\to w}\tilde k_M(z',w')\le\tilde k_M(z,w)$. To extend Pang’s result on manifolds, we have to define the “derivatives” of $k^{(m)}_M$, $m\in\Bbb N^\ast=\Bbb N\cup\{\infty\}$. Let $(U,\varphi)$ be a holomorphic chart near $z$. Set $$\mathcal D k^{(m)}_M(z;X)=\limsup_{t\nrightarrow 0,w\to z,Y\to\varphi_\ast X} \frac{k^{(m)}_M(w,\varphi^{-1}(\varphi(w)+tY))}{|t|}.$$ Note that this notion does not depend on the chart used in the definition and $$\mathcal D k^{(m)}_M(z;\lambda X)=|\lambda|\mathcal D k^{(m)}_M(z;X),\quad \lambda\in\Bbb C.$$ Replacing $\limsup$ by $\liminf$, we define $\underline{\mathcal D} k_M^{(m)}$. From M. Kobayashi’s paper [@KobM] it follows that, if $M$ is a taut manifold, then $$\hat\kappa_M(z;X)=\mathcal D k_M(z;X)=\underline{\mathcal D} k_M(z;X),$$ that is, the Kobayashi–Buseman metric is the “derivative” of the Kobayashi distance. The proof there also leads to $$\kappa_M^{(m)}(z;X)=\mathcal D k^{(m)}_M(z;X)=\underline{\mathcal D} k_M^{(m)}(z;X),\quad m\in\Bbb N^{\ast}.\leqno{(\ast)}$$ We say that a complex manifold $M$ is hyperbolic at $z$ if $k_M(z,w)>0$ for any $w\neq z$. We point out that the following conditions are equivalent: \(i) M is hyperbolic at $z;$ \(ii) $\ds\liminf_{z'\to z,w\in M\setminus U}\tilde k_M(z',w)>0$ for any neighborhood $U$ of $z;$ \(iii) $\ds\underline{\kappa}_M(z;X):=\liminf_{z'\to z,X'\to X}\kappa_M(z';X')>0$ for any $X\neq 0;$ The implication (i)$\Rightarrow$(ii) $\Rightarrow$(iii) are almost trivial (cf. [@Jar-Pfl]) and the implication (iii)$\Rightarrow$(i) is a consequence of the fact that $k_M$ is the integrated form of $\kappa_M$. In particular, if $M$ is hyperbolic at $z$, then it is hyperbolic at any $z'$ near $z$. Since if $M$ is taut, then it is k-hyperbolic and $\kappa_M$ is a continuous function, the following theorem is a generalization of $(\ast)$. \[th1\] Let $M$ be a complex manifold and $z\in M$. [(i)]{} If $M$ is hyperbolic at $z$ and $\kappa_M$ is continuous at $(z,X)$, then $$\kappa_M(z;X)=\mathcal D\tilde k_M(z;X)=\underline{\mathcal D} \tilde k_M(z;X).$$ [(ii)]{} If $\kappa_M$ is continuous and positive at $(z,X)$ for any $X\neq 0$, then $$\kappa_M^{(m)}(z;\cdot)=\mathcal D k_M^{(m)}(z;\cdot)=\underline{\mathcal D} k_M^{(m)}(z;\cdot),\quad m\in\Bbb N^\ast.$$ The first step in the proof of Theorem \[th1\] is the following \[pr2\] For any complex manifold $M$ one has that $$\kappa_M^{(m)}\ge\mathcal D k_M^{(m)}, \quad m\in\Bbb N^\ast.$$ Note that when $M$ is a domain, a weaker version of Proposition \[pr2\] can be found in [@Jar-Pfl], namely, $\hat\kappa_M\ge\mathcal D k_M$ (the proof is based on the fact that $\mathcal D k_M(z;\cdot)$ is a pseudonorm). Examples ======== The following examples show that the assumption on continuity in Theorem \[th1\] is essential. $\bullet$ Let $A$ be a countable dense subset of $\Bbb C_\ast.$ In [@Die-Sib] (see also [@Jar-Pfl]), a pseudoconvex domain $D$ in $\Bbb C^2$ is constructed such that: \(a) $(\Bbb C\times\{0\})\cup(A\times\Bbb C)\subset D;$ \(b) if $z_0=(0,t)\in D,$ $t\neq 0,$ then $\kappa_D(z_0;X)\ge C||X||$ for some $C=C_t>0$. (One can be shown that even $\mathcal D\tilde k_D(z_0;X)\ge C||X||.$) Then it is easy to see that $\underline{ \kappa}_D(\cdot;e_2)=\mathcal D k^{(3)}_D(\cdot;e_2)=k^{(5)}_D=0$ and $\hat\kappa_D(z_0;X)\ge c||X||,$ where $e_2=(0,1)$ and $c>0.$ Thus $$\hat\kappa_D(z_0;X)>\underline{ \kappa}_D(z_0;e_2)=\mathcal D k^{(3)}_D(z_0;e_2)=\mathcal D k^{(5)}_D(z_0;X),\quad X\neq 0.$$ This phenomena obviously extends to $\Bbb C^n$, $n>2$ (by considering $D\times\Bbb D^{n-2}$). So the inequalities in Proposition \[pr2\] are strict in general. $\bullet$ If $D$ is a pseudoconvex balanced domain with Minkowski function $h_D$, then (cf. [@Jar-Pfl]) $$h_D=\kappa_D(0;\cdot)=\mathcal D{\tilde k}_D(0;\cdot).$$ Therefore, $\mathcal D{\tilde k}_D(0;X)>\underline{\mathcal D}\tilde k_D(0;X)$ if $\kappa_D(0;\cdot)$ is not continuous at $X$. On the other hand, if $\hat D$ denotes the convex hull of $D$, then $$h_{\hat D}=\hat\kappa_D(0;\cdot)=\mathcal D{k}_D(0;\cdot)=\underline{\mathcal D}k_D(0;\cdot)= \underline{\hat\kappa}_D(0;\cdot).$$ $\bullet$ Modifying the first example leads to a pseudoconvex domain $D\subset\Bbb C^2$ with $$L_{\mathcal Dk_D}(\gamma)>0=L_{k_D}(\gamma)= L_{\underline{\mathcal D}k_D}(\gamma),$$ where $\gamma:[0,1]\to\Bbb C^2$, $\gamma(t):=(ti/2,1/2)$, and $L_\bullet(\gamma)$ denotes the respective length. Indeed, choose a dense sequence $(r_j)$ in $[0,i/2].$ Put $$u(\lambda)=\sum_{k=1}^\infty \frac{1}{k^2}\log\tfrac{|\lambda-1/k|}{4},\quad v(\lambda)=\sum_{j=1}^\infty\frac{u(\lambda/2-r_j)}{2j^2},\quad\lambda\in\Bbb C,$$ and $$D=\{z\in\Bbb C^2:\psi(z)=|z_2|e^{\|z\|^2+v(z_1)}<1\}.$$ It is easy to see that $v$ is a subharmonic function on $\Bbb C$. Hence $D$ is a pseudoconvex domain with $(\Bbb C \times\{0\})\cup(\bigcup_{j,k=1}^\infty \{r_j+1/k\}\times\Bbb C)\subset D.$ Observe that $u|_{\Bbb D}<-1$ and so $D$ contains the unit ball $\Bbb B_2.$ Note also that $$k_D(a,b)=0,\quad a,b\in\gamma([0,1]).$$ Set $\widehat \psi(z)=\|z\|^2/2-\log\psi(z)$. Fix $z^0\in\Bbb B_2$ with $\Ree z^0_1\le 0,\;\Imm z^0_2\ge1/e$. Since $u(\lambda)\ge u(0)$ for $\Ree \lambda\le 0,$ we have $$||z^0||/2<\widehat\psi(z^0)<1-u(0)=:8C.$$ Let $\varphi\in\mathcal O(\Bbb D,D)$, $\varphi(0)=z^0.$ Following the estimates in the proof of Example 3.5.10 in [@Jar-Pfl], we see that $\|\varphi'(0)\|<C$. Hence, $\kappa_D(z^0;X)\geq C\|X\|$, $X\in\Bbb C^2$. Since $k_D$ is the integrated form of $\kappa_D,$ it follows that $$k_D(a,a-te_1)\geq Ct,\quad a\in\gamma([0,1]),\;0\le t\le 1/2-1/e,\;e_1=(1,0).$$ Hence $\mathcal D k_D(a;e_1)\geq C$ and therefore, $L_{\mathcal Dk_D}(\gamma)\ge C/2>0,$ which completes the proof of this example. Note that it shows that, with respect to the lengths of curves, $\mathcal D k_D$ behaves different than the “real” derivative of $k_D$ (cf. [@Ven] or [@Jar-Pfl2], page 12). Moreover, it implies that, in general, $\mathcal D k_D\neq\underline{\mathcal D}k_D$. [**Questions.**]{} It will be interesting to know examples showing that, in general, $\kappa_D\neq\mathcal D\tilde k_D$. It remains also unclear whether $\mathcal D k_D$ is holomorphically contractible (see [@Jar-Pfl]). Recall that $\int\mathcal D k_D=k_D;$ but we do not know if $\int\underline{\mathcal D}k_D=k_D.$ Proofs ====== [*Proof of Proposition \[pr2\].*]{} First, we shall consider the case $m=1$. The key is the following \[th4\][^1] [@Roy] Let $M$ be an $n$-dimensional complex manifold and $f\in\mathcal O(\Bbb D,M)$ regular at $0$. Let $r\in(0,1)$ and $D_r=r\Bbb D\times\Bbb D^{n-1}$. Then there exists $F\in\mathcal O(D_r,M)$, which is regular at $0$ and $F|_{r\Bbb D\times\{0\}}=f$. Since $\kappa_M(z;0)=\mathcal D\tilde k_M(z;0)=0$, we may assume that $X\neq 0$. Let $\alpha>0$ and $f\in\mathcal O(\Bbb D,M)$ be such that $f(0)=z$ and $\alpha f_\ast(d/d\zeta)=X$. Let $r\in(0,1)$ and $F$ as in Theorem \[th4\]. Since $F$ is regular at $0$, there exist open neighborhoods $U=U(z)\subset M$ and $V=V(0)\subset D_r$ such that $F|_V:V\to U$ is biholomorphic. Hence $(U,\varphi)$ with $\varphi=(F|_V)^{-1}$, is a chart near $z$. Note that $\varphi_\ast(X)=\alpha e_1$, where $e_1=(1,0,\dots,0).$ If $w$ and $Y$ are sufficiently near $z$ and $\alpha e_1$, respectively, then $$g(\zeta):=F(\varphi(w)+\zeta Y/\alpha),\quad \zeta\in r^2\Bbb D,$$ belongs to $\mathcal O(r^2\Bbb D,M)$ with $g(0)=w$ and $g(t\alpha)=\varphi^{-1}(\varphi(w)+tY),$ $t<r^2/\alpha.$ Therefore, $r^2\widetilde k^\ast_M(w,\varphi^{-1}(\varphi(w)+tY))\leq t\alpha$. Hence $r^2\mathcal D\tilde k_M(z;X)\le\alpha$. Letting $r\to 1$ and $\alpha\to\kappa_M(z;X)$ we get that $\mathcal D\tilde k_M(z;X)\le\kappa_M(z;X).$ Let now $m\in\Bbb N$. By definition, $\kappa_M^{(m)}(z;\cdot)$ is the largest function with the following property: For any $X=\sum_{j=1}^mX_j$ one has that $\kappa_M^{(m)}(z;X)\le\sum_{j=1}^m\kappa_M(z;X_j)$. To prove that $\kappa_M^{(m)}\ge\mathcal D k_M^{(m)}$ it suffices to check that $\mathcal D k^{(m)}_M(z;\cdot)$ has the same property. Following the above notation and choosing $Y_j\to\varphi_\ast X_j$ with $\sum_{j=1}^m Y_j=Y$, we set $w_0=w$ and $w_j=\varphi^{-1}(\varphi(w)+t\sum_{k=1}^{j}Y_j)$. Since $$k^{(m)}_M(w,w_q)\le\sum_{j=1}^m\tilde k_M(w_{j-1},w_j),$$ it follows by the case $m=1$ that $$\mathcal D k^{(m)}_M(z;X)\le\sum_{j=1}^m\mathcal D k_M(z;X_j)\le\sum_{j=1}^m\kappa_M(z;X_j).$$ Finally, let $m=\infty$ and $n=\dim M$. Since $\hat\kappa_M=\kappa_M^{(2n-1)}$ and $k_M\le k_M^{(2n-1)}$, we get that $\mathcal D k_M\le\hat\kappa_M$ using the case $m=2n-1$. [*Proof of Theorem \[th1\].*]{} We may assume that $X\neq 0$. In virtue of Proposition \[pr2\], we have to show that $$\kappa_M^{(m)}(z;X)\le\underline{\mathcal D}k^{(m)}_M(z;X).$$ For simplicity we assume that $M$ is a domain in $\Bbb C^n$. \(i) Fix a neighborhood $U=U(z)\Subset M.$ Applying the hyperbolicity of $M$ at $z$, there are a neighborhood $V=V(z)\subset U$ and a $\delta\in (0,1)$ such that, if $h\in\mathcal O(\Bbb D,M)$ with $h(0)\in V$, then $h(\delta\Bbb D)\subset U$. Hence, by the Cauchy inequalities, $||h^{(k)}(0)||\leq c/\delta^k$, $k\in\Bbb N.$ Now choose sequences $M\ni w_j\to z$, $\Bbb C_\ast\ni t_j\to 0$, and $\Bbb C^n\ni Y_j\to X$ such that $$\frac{\widetilde k_M(w_j,w_j+t_jY_j)}{|t_j|}\to\underline{\mathcal D}\widetilde k_M(z;X).$$ There are holomorphic discs $g_j\in\mathcal O(\Bbb D,M)$ and $\beta_j\in (0,1)$ with $g_j(0)=w_j$, $g_j(\beta_j)=w_j+t_jY_j,$ and $\beta_j\le\widetilde k^\ast_M(w_j,w_j+t_jY_j)+|t_j|/j.$ Note that $\widetilde k^\ast_M(w_j,w_j+t_jY_j)\le c_1||t_jY_j||\le c_2 |t_j|.$ Write $$w_j+t_jY_j=g_j(\beta_j)=w_j+g_j'(0)\beta_j+ h_j(\beta_j).$$ Then $$||h_j(\beta_j)||\leq c\sum_{k=2}^\infty \left(\tfrac{\beta_j}{\delta}\right)^k\leq c_3|\beta_j|^2\le c_4|t_j|^2,\quad j\ge j_0.$$ Put $\widehat Y_j=Y_j-h_j(\beta_j)/t_j.$ We have that $g_j(0)=w_j$ and $\beta_jg_j'(0)/t_j=\widehat Y_j\to X$. Therefore, $$\kappa_M(w_j;\widehat Y_j)\leq\frac{\beta_j}{|t_j|}\leq\frac{\widetilde k^\ast_M(z_j,w_j+t_jY_j)}{|t_j|}+\frac{1}{j}.$$ Hence with $j\to\infty$, we get that $\kappa_M(z;X)=\underline{\kappa}_M(z;X)\leq \underline{\mathcal D}\widetilde k_M(z;X)$. \(ii) The proof of the case $m\in\Bbb N$ is similar to the next one and we omit it. Now, we shall consider the case $m=\infty$. Note first that our assumption implies that $M$ is hyperbolic at $z$ and, by the contrary, $$\forall\varepsilon>0\ \exists\delta>0:||w-z||<\delta,||Y-X||<\delta||X||$$ $$\Rightarrow|\kappa_M(w;Y)-\kappa_M(z;X)|<\varepsilon\kappa_M(z;X).\leqno{(1)}$$ Moreover, the proof of (i) shows that $$\tilde k_M(a,b)\ge\kappa_M(a;b-a+o(a,b)),\hbox{ where }\lim_{a,b\to z}\frac{o(a,b)}{||a-b||}=0.\leqno{(2)}$$ Choose now sequences $M\ni w_j\to z$, $\Bbb C_\ast\ni t_j\to 0$, and $\Bbb C^n\ni Y_j\to X$ such that $$\frac{k_M(w_j,w_j+t_jY_j)}{|t_j|}\to\underline{\mathcal D}k_M(z;X).$$ There are points $w_{j,0}=w_j,\dots,w_{j,m_j}=w_j+t_jX_j$ in $M$ such that $$\sum_{k=1}^{m_j}\tilde k_M(w_{j,k-1},w_{j,k})\le k_M(w_j,w_j+t_jY_j)+\frac{1}{j}.\leqno{(3)}$$ Set $w_{j,k}=w_j$ for $k>m_j$. Since $$k_M(w_j,w_{j,l})\le\sum_{j=1}^l\tilde k_M(w_{j,k-1},w_{j,k})\le k_M(w_j,w_j+t_jY_j)+\frac{1}{j}\le c_2|t_j|+\frac{1}{j},$$ then $k_M(w_j,w_{j,l})\to 0$ uniformly in $l.$ Then the hyperbolicity of $M$ at $z$ implies that $w_{j,l}\to z$ uniformly in $l.$ Indeed, assuming the contrary and passing to a subsequence, we may suppose that $w_{j,l_j}\not\in U$ for some $U=U(z).$ Then $$0=\lim_{j\to\infty}k_M(w_j,w_{j,l})\ge\ds\liminf_{z'\to z,w\in M\setminus U}\tilde k_M(z',w)>0,$$ a contradiction. Fix now $R>1.$ Then (1) implies that $$\kappa_M(z;w_{j,k}-w_{j,k-1})\le R\kappa_M(w_{j,k};w_{j,k}-w_{j,k-1} +o(w_{j,k},w_{j,k-1})),\ j\ge j(R).$$ It follows by this inequality, (2) and (3) that $$\sum_{k=1}^{m_j}\kappa_M(z;w_{j,k}-w_{j,k-1}) \le Rk_M(w_j,w_j+t_jYj)+\frac{R}{j}.$$ Since $\hat\kappa_M(z;t_jY_j)$ is bounded by the first sum, we obtain that $$\hat\kappa_M(z;Y_j)\le R\frac{k_M(w_j,w_j+t_jYj)+1/j}{|t_j|}.$$ Note that $\hat\kappa_M(z;\cdot)$ is a continuous function. Hence with $j\to\infty$ and $R\to 1,$ we get that $\hat\kappa_M(z;X)\le\underline{\mathcal D}k_M(z;X)$. [**Remark.**]{} It follows by the above proofs and a standard diagonal process that $\underline{\kappa}_M(z;\cdot)=\underline{\mathcal D}\tilde k(z;\cdot)$ if $M$ is hyperbolic at $z$. K. Diederich, N. Sibony, [*Strange complex structures on Euclidian space*]{}, J. Reine Angew. Math. 311/312 (1979), 397–407. S. Dineen, [*The Schwarz lemma*]{}, Oxford Math. Monographs, Clarendon Press, Oxford, 1989. M. Jarnicki, P. Pflug, [*Invariant distances and metrics in complex analysis*]{}, de Gruyter Exp. Math. 9, de Gruyter, Berlin, New York, 1993. M. Jarnicki, P. Pflug, [*Invariant distances and metrics in complex analysis–revisited*]{}, Dissertationes Math. 430 (2005). M. Kobayashi, [*On the convexity of the Kobayashi metric on a taut complex manifold*]{}, Pacific J. Math. 194 (2000), 117–128. S. Kobayashi, [*A new invariant infinitesimal metric*]{}, International J. Math. 1 (1990), 83–90. S. Kobayashi, [*Hyperbolic complex spaces*]{}, Grundlehren Math. Wiss. 318, Springer, Berlin, 1998. N. Nikolov, P. Pflug, [*On the definition of the Kobayashi-Buseman pseudometric*]{}, International J. Math. (to appear). M.-Y. Pang, [*On infinitesimal behavior of the Kobayashi distance*]{}, Pacific J. Math. 162 (1994), 121–141. H.-L. Royden, [*The extension of regular holomorphic mapps*]{}, Proc. Amer. Math. Soc. 43 (1974), 306–310. S. Venturini, [*Pseudodistances and pseudometrics on real and complex manifolds*]{}, Ann. Mat. Pura Appl. 154 (1989), 385–402. J. Winkelmann, [*Non-degenerate maps and sets*]{}, Math. Z. 249 (2005), 783–795. [^1]: We may replace Theorem \[th4\] by the approach used in the proof of the upper semicontinuity of $\tilde k_M$.

--- abstract: 'A key aspect for any greedy pursuit algorithm used in compressed sensing is a good support-set detection method. For distributed compressed sensing, we consider a setup where many sensors measure sparse signals that are correlated via the existence of a signals’ intersection support-set. This intersection support-set is called the joint support-set. Estimation of the joint support-set has a high impact on the performance of a distributed greedy pursuit algorithm. This estimation can be achieved by exchanging local support-set estimates followed by a (consensus) voting method. In this paper we endeavor for a probabilistic analysis of two democratic voting principle that we call majority and consensus voting. In our analysis, we first model the input/output relation of a greedy algorithm (executed locally in a sensor) by a single parameter known as probability of miss. Based on this model, we analyze the voting principles and prove that the democratic voting principle has a merit to detect the joint support-set.' author: - 'Dennis Sundman,  Saikat Chatterjee,  and Mikael Skoglund, [^1]' bibliography: - 'references/IEEEfull.bib' - 'references/myconffull.bib' - 'references/compressed\_sensing.bib' title: - Hard Decisions by Probabilistic Modeling - 'Analysis of Voting Principles for Support-set Estimation used in Greedy Pursuits' - 'Analysis of Voting Principles for Support-set Estimation used in Distributed Greedy Pursuits' - Analysis of Democratic Voting Principles used in Distributed Greedy Algorithms --- [Shell : Bare Demo of IEEEtran.cls for Journals]{} Greedy algorithms, distributed detection, hard decision. Introduction ============ sensing ([<span style="font-variant:small-caps;">cs</span>]{}) [@Donoho:compressed_sensing; @Candes:stable_signal_recovery] typically considers a single-sensor scenario, where the main task is reconstruction of a large-dimensional signal-vector from a small-dimensional measurement-vector by using a-priori knowledge that the signal is sparse in a known domain. Several [<span style="font-variant:small-caps;">cs</span>]{}reconstruction algorithms have been considered in the literature, for example convex optimization- [@Mota:distributed_basis_pursuit; @Bazerque:distributed_spectrum_sensing], Bayesian- [@Ji:bayesian_compressive_sensing; @Donoho:message_passing_for_cs] and greedy pursuit ([<span style="font-variant:small-caps;">gp</span>]{}) algorithms. The greedy pursuit ([<span style="font-variant:small-caps;">gp</span>]{}) algorithms are popular due to their low complexity and good performance. From a measurement vector, the [<span style="font-variant:small-caps;">gp</span>]{}algorithms use linear algebraic tools to estimate the underlying *support-set* of the sparse signal-vector followed by estimating associated signal values; here we mention that good support-set estimation is a key aspect for the [<span style="font-variant:small-caps;">gp</span>]{}algorithms. A few examples of typical [<span style="font-variant:small-caps;">gp</span>]{}algorithms include: matching pursuit [@Mallat:matching_pursuit_with_time_frequency_dictionaries], orthogonal matching pursuit (<span style="font-variant:small-caps;">omp</span>) [@Tropp:signal_recovery], [$\textsc{c}\text{o}\textsc{s}\text{a}\textsc{mp}$]{}[@Needell:cosamp], subspace pursuit ([$\textsc{sp}$]{}) [@Dai:subspace_pursuit], but there are many others [@Donoho:sparse_solution_of_underdetermined_systems_stomp; @Chatterjee:projection_based_look_ahead; @Sundman:frogs; @Needell:signal_recovery_from_incomplete_and_inaccurate_measurements_via_romp; @Sundman:look_ahead_parallel_pursuit]. For the [<span style="font-variant:small-caps;">gp</span>]{}algorithms, just as for any [<span style="font-variant:small-caps;">cs</span>]{}reconstruction algorithm, providing analytical performance guarantees is an important yet challenging task. These guarantees are typically done through worst case analysis based on restricted isometry property [@Candes:restricted_isometry_property] and mutual coherence inequalities. Distributed (or de-centralized) [<span style="font-variant:small-caps;">cs</span>]{}([<span style="font-variant:small-caps;">dcs</span>]{}) refers to a problem of multiple sensors connected over a network, where the sensors observe correlated sparse signals through [<span style="font-variant:small-caps;">cs</span>]{}measurements. By the term [<span style="font-variant:small-caps;">dcs</span>]{}we refer both to simultaneous estimation in a distributed network [@Tropp:simultaneous_sparse_approx_part1; @Rakotomamonjy:surveying; @Leviatan:simultaneous; @Cotter:sparse_solutions; @Chen:theoretical_results_on_sparse; @Sundman:greedy_pursuit_for_jointly] and to multiple measurement vector setups in some fusion center [@Mota:distributed_basis_pursuit; @Bazeraque:distributed_spectrum_sensing; @Feng:distributed_compressive_spectrum_sensing; @Ling:decenteralized_support_detection]. Recently we developed several [<span style="font-variant:small-caps;">gp</span>]{}algorithms for [<span style="font-variant:small-caps;">dcs</span>]{}, called distributed greedy pursuits [@Sundman:diprsp; @Sundman:a_greedy_pursuit_algorithm; @Sundman:distributed_gp_algorithms; @Sundman:dipp_arxiv]. In [<span style="font-variant:small-caps;">dcs</span>]{}, two (of many) models for signal correlations are the common support-set model and the mixed support-set model [@Sundman:methods_for_dcs]. In the common support-set model, the same (joint) *full* support-set is assumed for all signals measured at different sensors, while in the mixed support-set model a joint *partial* support-set is assumed for all sensors. Based on these models, a prominent approach for distributed [<span style="font-variant:small-caps;">gp</span>]{}algorithms is to let the sensors in the network exchange (or transmit to a centralized point) full support-set estimates and then, using only support-set knowledge, estimate the joint support-set. A better estimate of the joint support-set can then be used to improve [<span style="font-variant:small-caps;">dcs</span>]{}reconstruction performance. In general, theoretical performance analysis of distributed [<span style="font-variant:small-caps;">gp</span>]{}algorithms is non-trivial and we recently developed [$\textsc{dipp}$]{}(distributed parallel pursuit) – a distributed greedy pursuit algorithm – with such theoretical guarantees in [@Sundman:dipp_arxiv]. Through analysis and simulations we have shown that [$\textsc{dipp}$]{}performs better than local [<span style="font-variant:small-caps;">gp</span>]{}algorithms, such as [$\textsc{sp}$]{}. In [$\textsc{dipp}$]{}and other distributed greedy pursuits, the joint support-set is estimated by a consensus voting method. In several of our earlier works [@Sundman:a_greedy_pursuit_algorithm; @Sundman:distributed_gp_algorithms], we assumed that democratic based voting is suitable for consensus, and in [@Sundman:dipp_arxiv] we proved theoretical reconstruction guarantees based on this assumption. The advantage of voting has earlier been studied in politics and finance as early as 1785 [@Young:optimal_voting_rules; @Ledyard:the_approximation_of_efficient]. In this paper, we endeavor to prove that the assumption of democratic voting for support-set estimation, based on [<span style="font-variant:small-caps;">gp</span>]{}algorithms, indeed has a merit. In our approach, we assume that support-sets estimated from [<span style="font-variant:small-caps;">gp</span>]{}algorithms executed locally in several sensors likely contain independent errors. Therefore, based on probability of detection, miss, and false alarms, we first model the input/output relation of relevant [<span style="font-variant:small-caps;">gp</span>]{}algorithms by using standard detection theory framework. Using the input/output relation, we provide probability results for consensus strategies based on democratic voting applied in scenarios that employ the common and mixed support-set models. The main contributions of this paper can be summarized as: - Defining the input/output relation of relevant [<span style="font-variant:small-caps;">gp</span>]{}algorithms. - Probabilistic analysis of democratic voting used in distributed [<span style="font-variant:small-caps;">gp</span>]{}algorithms for both common and mixed support-set models. The outline of the paper is as follows: We first introduce some notation in the next subsection. Then, in Section \[sec:system\_model\], we introduce the signal model, the common support-set, and mixed support-set models. In Section \[sec:prob\_model\], we introduce an input/output relation to model single sensor performance, which is then used for analyzing different voting strategies for the the common support-set model in Section \[sec:common\_model\] and the mixed support-set model in Section \[sec:mixed\_model\]. Then, in Section \[sec:experiment\], we provide experimental verification of the results achieved. Notation -------- Sets are denoted by calligraphic capital letters, in particular; ${\mathcal{T}}$, ${\mathcal{I}}$ and ${\ensuremath{\mathcal{J}}}$ are support-sets or partial support-sets. We define the full set $\Omega = \{ 1,2, \dots, N \}$ and the set complement ${\ensuremath{\mathcal{J}}}^{{\ensuremath{\mathsf{c}}}} = \Omega \setminus {\ensuremath{\mathcal{J}}}$, where ‘$\setminus$’ is the set-minus. We denote the event of an index $i$ residing in the support-set ${\mathcal{T}}$ by $i \in {\mathcal{T}}$. If $i$ resides in two support-sets – ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ and ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}$ – we use either $(i\in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}})$ or $i\in ({\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}})$; where the one providing most insight will be used. The probability of an index $i$ residing inside the support-set ${\mathcal{T}}$ is denoted by ${\ensuremath{\mathbb{P}\,\!\bigl(i\in{\mathcal{T}}\bigr)}}$. Lastly, we denote the conditional probability, where by ${\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}|i\in{\mathcal{T}}\bigr)}}$ we refer to the probability that an index $i$ be in ${\ensuremath{\mathcal{J}}}$, given that $i$ is (randomly) in ${\mathcal{T}}$. Lastly we introduce two algorithmic notations $$\begin{aligned} {\texttt{vote}_1}(\mathbf{z}, {\mathcal{T}}) & \triangleq \{ \forall i \in {\mathcal{T}}, \textit{ perform } z_{i} = z_{i}+1 \}.\end{aligned}$$ $$\begin{aligned} \texttt{max\_indices}(\mathbf{z}, T) \triangleq & \{\textit{select the $T$ largest amplitude} \nonumber \\ & \,\,\, \textit{indices of $\mathbf{z}$} \}.\end{aligned}$$ System Model {#sec:system_model} ============ In this section we define the distributed compressed sensing ([<span style="font-variant:small-caps;">dcs</span>]{}) problem, the common support-set model and the mixed support-set model. Distributed Compressed Sensing ------------------------------ In distributed compressed sensing ([<span style="font-variant:small-caps;">dcs</span>]{}), each ${\ensuremath{\mathsf{p}}}$’th sensor measures a signal $\mathbf{x}_{{\ensuremath{\mathsf{p}}}} \in \mathbb{R}^{N}$ through the following linear relation $$\begin{aligned} \mathbf{y}_{{\ensuremath{\mathsf{p}}}} = \mathbf{A}_{{\ensuremath{\mathsf{p}}}} \mathbf{x}_{{\ensuremath{\mathsf{p}}}} + \mathbf{e}_{{\ensuremath{\mathsf{p}}}}, ~~~~~~ \forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}, \label{eqn:model}\end{aligned}$$ where $\mathbf{y}_{{\ensuremath{\mathsf{p}}}} \in \mathbb{R}^{M}$ is a measurement vector, $\mathbf{A}_{{\ensuremath{\mathsf{p}}}} \in \mathbb{R}^{M\times N}$ is a measurement matrix, $\mathbf{e}_{{\ensuremath{\mathsf{p}}}} \in \mathbb{R}^{M}$ is some measurement noise and $\mathcal{L}$ is a global set containing all sensors (nodes) in the network ($|\mathcal{L}| = L$). The signal vector $\mathbf{x}_{{\ensuremath{\mathsf{p}}}} = [x_{{\ensuremath{\mathsf{p}}}}(1) \,\, x_{{\ensuremath{\mathsf{p}}}}(2) \, \ldots \, x_{{\ensuremath{\mathsf{p}}}}(N)]$ is $T$-sparse, meaning it has $T$ elements that are non-zero. Thus, the setup describes an under-determined system, where $T < M < N$. The element-indices corresponding to non-zero values are collected in the support-set ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$; that means ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}= \{ i : x_{{\ensuremath{\mathsf{p}}}}(i) \neq 0 \}$ and $| {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| = T$. A dense vector containing only the non-zero values of ${\mathbf{x}}_{{\ensuremath{\mathsf{p}}}}$ is represented by $\mathbf{v}_{{\ensuremath{\mathsf{p}}}} = [x_{{\ensuremath{\mathsf{p}}}}({\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}(1)), x_{{\ensuremath{\mathsf{p}}}}({\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}(2)), \dots, x_{{\ensuremath{\mathsf{p}}}}({\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}(T))]$, which may also be independent locally and across the network. Throughout this paper we use measurement matrices that have unit $\ell_2$-norm columns and characterize the signal- to-noise ratio using the signal-to-measurement-noise-ratio ([<span style="font-variant:small-caps;">smnr</span>]{}), which is defined for sensor ${\ensuremath{\mathsf{p}}}$ as $$\begin{aligned} {\textsc{smnr}\xspace}\triangleq \frac{ \mathcal{E} \{ \| \mathbf{x}_{{\ensuremath{\mathsf{p}}}} \|_{2}^{2} \} }{ \mathcal{E} \{ \| \mathbf{e}_{{\ensuremath{\mathsf{p}}}} \|_{2}^{2} \} }. \label{eqn:smnr}\end{aligned}$$ Furthermore, $\mathbf{A}_{{\ensuremath{\mathsf{p}}}}$ and $\mathbf{e}_{{\ensuremath{\mathsf{p}}}}$ are independent both locally and across the network. In order to benefit from cooperation in the network, some correlation in the signal vector ${\mathbf{x}}_{{\ensuremath{\mathsf{p}}}}$ must be present. In the following two subsections we present these correlations by introducing the common and mixed support-set models. Common Support-set Model {#sec:common_support_signal_model} ------------------------ In the common support-set model [@Duarte:distributed_compressed_sensing; @Sundman:methods_for_dcs], the support-sets of all signals in the network ${\mathbf{x}}_{{\ensuremath{\mathsf{p}}}}$ are identical. That is $$\begin{aligned} {\mathcal{T}}_{{\ensuremath{\mathsf{p}}}} = {\ensuremath{\mathcal{J}}}~~~~~ \forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}, \label{eqn:common}\end{aligned}$$ where we refer to ${\ensuremath{\mathcal{J}}}$ as the *joint* support-set. Mixed Support-set Model {#sec:mixed_support_signal_model} ----------------------- A natural extension to the common support-set model is the mixed support-set model, proposed by us in [@Sundman:greedy_pursuit_for_jointly; @Sundman:a_greedy_pursuit_algorithm; @Sundman:methods_for_dcs]. In this case there exists an intersection between all support-sets ${\mathcal{T}}_{{\ensuremath{\mathsf{p}}}}$. Denoting ${\ensuremath{\mathcal{J}}}= \cap_{\forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}} {\mathcal{T}}_{{\ensuremath{\mathsf{p}}}}$, we have $${\mathcal{T}}_{{\ensuremath{\mathsf{p}}}} = {\mathcal{I}}_{{\ensuremath{\mathsf{p}}}} \cup {\ensuremath{\mathcal{J}}}~~~~~ \forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}. \label{eqn:mixed}$$ Here, we refer to ${\ensuremath{\mathcal{J}}}$ as the joint part of the support-set (or simply joint support-set) and ${\mathcal{I}}_{{\ensuremath{\mathsf{p}}}} \triangleq {\mathcal{T}}_{{\ensuremath{\mathsf{p}}}} \setminus {\ensuremath{\mathcal{J}}}$ is the individual part. Denoting $|{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|=I ~ \forall ~ {\ensuremath{\mathsf{p}}}$ and $|{\ensuremath{\mathcal{J}}}|=J$, the following assumptions are used throughout the paper: 1. Elements of support-sets are uniformly distributed, $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} = \frac{|{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|}{|\Omega|} = \frac{T}{N}, ~~ \forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}. \label{eqn:cons:uniformity} \end{aligned}$$ 2. ${\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\mathcal{J}}}= \emptyset, ~~~ \forall {\ensuremath{\mathsf{p}}}\in \mathcal{L}$. 3. ${\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}= \emptyset, ~~~ \forall {\ensuremath{\mathsf{p}}}, {\ensuremath{\mathsf{q}}}\in \mathcal{L}, {\ensuremath{\mathsf{p}}}\neq {\ensuremath{\mathsf{q}}}$. 4. Hence, $T = I+J$.[$\square$]{} Modeling the input/output Relation for Greedy Pursuits {#sec:prob_model} ====================================================== A [<span style="font-variant:small-caps;">gp</span>]{}algorithm in sensor ${\ensuremath{\mathsf{p}}}$ will attempt to find the true support-set ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$. Influencing the chances of success are a number of factors: signal amplitudes (i.e., $\mathbf{v}_{{\ensuremath{\mathsf{p}}}}$), measurement noise ${\mathbf{e}}_{{\ensuremath{\mathsf{p}}}}$, sparsity $T$ and measurement matrix $\mathbf{A}_{{\ensuremath{\mathsf{p}}}}$ realization. Illustrated in [Fig.]{} \[fig:full\_system\], is the whole procedure from an underlying ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$, signal acquisition according to , to recovered support-set estimate ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$. In [Fig.]{} \[fig:system1\], we have simplified the previous figure in one box, referred to as the *System*. Borrowing terms from detection theory we model the system (see Definition \[ass\]), where the idea is to replace the factors influencing support-set recovery performance with one single parameter $\epsilon_{{\ensuremath{\mathsf{p}}}}$, shown in [Fig.]{} \[fig:system2\]. Introduction of this single parameter helps to bring analytical tractability, which we will witness in Sections \[sec:common\_model\] and \[sec:mixed\_model\]. (nature) [$\{ \cdot \}_0$]{}; (mix1) [$\times$]{}; (add1) ; (cs) [[<span style="font-variant:small-caps;">gp</span>]{}]{}; (out) [$\hat{{\mathbf{x}}}_{{\ensuremath{\mathsf{p}}}}$, ${\hat{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}}$]{}; (e) [${\mathbf{e}}_{{\ensuremath{\mathsf{p}}}}$]{}; (in) [${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$]{}; (v) [$\mathbf{v}_{{\ensuremath{\mathsf{p}}}}$]{}; (A) [${\mathbf{A}}_{{\ensuremath{\mathsf{p}}}}$]{}; (acs) ; (aA) ; (v) – (nature); (A) – (mix1); (aA) – (acs) – (cs.north); (in) – (nature); (e) – (add1); (nature) – node\[midway,label=[above:${\mathbf{x}}_{{\ensuremath{\mathsf{p}}}}$]{}\] (mix1); (mix1) – (add1); (add1) – node\[midway,label=[above:${\mathbf{y}}_{{\ensuremath{\mathsf{p}}}}$]{}\] (cs); (cs) – (out); \[ass\] The support-set estimate ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ of any unbiased [<span style="font-variant:small-caps;">gp</span>]{}algorithm described in [Fig.]{} \[fig:system2\] follows $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & & & = \frac{T}{N} \label{ass1} \\ {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl(\text{``detect''}\bigr)}} & & = 1 - \epsilon_{{\ensuremath{\mathsf{p}}}} \label{ass2} \\ {\ensuremath{\mathbb{P}\,\!\bigl(i \notin {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl(\text{``miss''}\bigr)}} && = \epsilon_{{\ensuremath{\mathsf{p}}}} \label{ass3} \\ {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl(\text{``false alarm''}\bigr)}} & & = \frac{T}{N-T}\epsilon_{{\ensuremath{\mathsf{p}}}}\label{ass4}, \end{aligned}$$ where $0 \leq \epsilon_{{\ensuremath{\mathsf{p}}}} \leq \frac{N-T}{N}$. Observe that $i \notin {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}= i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}$. These probabilities should be read as, for example in : *“The probability that index $i$ is part of the output ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ from the system, provided that this index is already part of the true underlying support-set ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$”.* For the remainder of the paper, we assume that all sensors in the network have statistically identical system and signal conditions, meaning that $\epsilon_{{\ensuremath{\mathsf{p}}}} = \epsilon ~ \forall ~ {\ensuremath{\mathsf{p}}}$. [$\square$]{} *Discussion:* The input/output relation in Definition \[ass\] follows from symmetry arguments. Since the system is symmetric and ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ is uniformly random, any unbiased (fair) reconstruction algorithm will produce ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ which is also uniformly random ; an unbiased algorithm should not favor any correct index over than any other correct index, resulting in . Similarly, the algorithm should not favor any missed index over another missed index . Furthermore, whenever a support-index is missed, a false alarm has occurred; therefore the false alarm can be parametrized by the same parameter as the probability of miss and detect . Using $\epsilon$ to specify the system behavior, we see that the worst possible system will select indices for the support-set uniformly at random. Thus the upper-bound on $\epsilon$ is $\epsilon_{\max} = \frac{N-T}{N}$, which means that the worst ${\ensuremath{\mathbb{P}\,\!\bigl(\text{``detect''}\bigr)}} = \frac{T}{N}$. [$\square$]{} At this point, there is no closed-form expression of the parameter $\epsilon$ as it would require complete characterization of ${\mathbf{A}}_{{\ensuremath{\mathsf{p}}}}$, ${\mathbf{x}}_{{\ensuremath{\mathsf{p}}}}$, ${\mathbf{e}}$ and of the present [<span style="font-variant:small-caps;">gp</span>]{}algorithm. Such an analysis is outside the scope of this paper; instead we estimate $\epsilon$ through experiments. This can in practice be performed, for example, by using pilot signals. We now present the first result. \[prop:disconnect\] The probability that an index ‘$i$’ is correct for sensor ${\ensuremath{\mathsf{p}}}$ provided that it is found by the ${\ensuremath{\mathsf{p}}}$’th system is given by $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}. \label{eqn:disconnect} \end{aligned}$$ $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & \overset{(a)}{=} \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}} \nonumber \\ & \overset{(b)}{=} \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}\frac{T}{N}}{\frac{T}{N}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \nonumber \end{aligned}$$ where we in $(a)$ have used Bayes’ rule and in $(b)$ have used and . Numerical Verification of the System Model {#sec:evaluation_prob_model} ------------------------------------------ In order to verify the system model in Definition \[ass\], we perform two different tests. As [<span style="font-variant:small-caps;">gp</span>]{}algorithm we have used the well known subspace pursuit ([$\textsc{sp}$]{}) algorithm, however; similar results can be obtained with other [<span style="font-variant:small-caps;">gp</span>]{}algorithms that are based on fixed support-set size. In the first test, presented in [Fig.]{} \[fig:ass1\], we verify . The test is based on $10^5$ random: support-sets ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$, signal realizations $\mathbf{v}_{{\ensuremath{\mathsf{p}}}}$, measurement matrices $\mathbf{A}_{{\ensuremath{\mathsf{p}}}}$ and noises $\mathbf{e}_{{\ensuremath{\mathsf{p}}}}$ (generated such that ${\textsc{smnr}\xspace}= 20$ dB). In [Fig.]{} \[fig:ass1\], $N = 50$ and $T = 2$ to make the outcome observable ($M = 7$). Along the x-axis we show the support-set index $i$, and on the y-axis, we show how many times each index appears in the output ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$. From this figure, we see that by using the proposed setup, the output from the algorithm is uniform, which verifies of the definition. ![Figure showing how often each index occurs in the output of the system, based on of the system model.[]{data-label="fig:ass1"}](figures/model2.eps){width="\columnwidth"} In the second test, presented in [Fig.]{} \[fig:ass2\]; , and are verified where $N = 50$ and $T = 2$ (and $M = 7$). Here, there are $10^5$ random: signal realizations $\mathbf{v}_{{\ensuremath{\mathsf{p}}}}$, measurement matrices ${\mathbf{A}}_{{\ensuremath{\mathsf{p}}}}$ and noises $\mathbf{e}_{{\ensuremath{\mathsf{p}}}}$ (such that ${\textsc{smnr}\xspace}= 20$ dB). The support-set ${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}= [14, 26]$ is fixed in order to produce an informative figure. Along the x-axis is the support-set index $i$, and on the y-axis, we show how many times each index appears in the output ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$. We can now directly identify the three equations , and . First, we estimate $\epsilon$ by counting the number of false alarms ; in this case $\hat{\epsilon} = 0.267$. Then and are found directly from $\hat{\epsilon}$. We will now apply the input/output relation model to more complex scenarios. ![Simulation verification that indeed there exist an underlying $\epsilon$ such that the proposed system model holds.[]{data-label="fig:ass2"}](figures/model1.eps){width="\columnwidth"} Voting Based Detection for\ the Common Support-set Model {#sec:common_model} ============================ In this section we introduce the concept of voting based on support-set estimates from a number of nodes. We model signal correlation according to the common support-set model (see Section \[sec:common\_support\_signal\_model\]). Throughout this section we use ${\mathcal{T}}$ and ${\ensuremath{\mathcal{J}}}$ interchangeably for the same purpose, since they are equivalent in the common support-set model. Consider a setup with multiple sensor nodes where each sensor gathers [<span style="font-variant:small-caps;">cs</span>]{}measurements and runs a local [<span style="font-variant:small-caps;">gp</span>]{}algorithm to find a local support-set estimate. The support-set estimates from all nodes are then sent to a fusion center (or exchanged distributively) for estimation of ${\ensuremath{\mathcal{J}}}$. Algorithm {#sec:common_algo} --------- We propose a fusion center strategy based on democratic voting where, assuming $T$ is known, the strategy for the final estimate is to choose the $T$ indices with most votes. This is a [$\textup{\texttt{majority}}$]{}voting strategy and a detailed description is presented in Algorithm \[alg:majority\]. *Input:* $\{ \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}} \}_{{\ensuremath{\mathsf{p}}}\in \mathcal{L}}$, $T$\ *Initialization:* $\mathbf{z} \leftarrow \mathbf{0}_{N\times 1}$ \[alg:majority:cons:0\]\ *Algorithm:* $\mathbf{z} \leftarrow {\texttt{vote}_1}(\mathbf{z}, \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}})$ (The estimate of sensor ${\ensuremath{\mathsf{p}}}$) \[alg:majority:cons:others\] $\hat{{\ensuremath{\mathcal{J}}}} \leftarrow \texttt{max\_indices}(\mathbf{z}, J)$ Studying Algorithm \[alg:majority\], we see that the inputs are the support-set estimates from all sensors in the network, and the support-set cardinality. In the initialization phase, a large $N$-sized vector $\mathbf{z}$ is created; where the votes of the sensors are collected. Then, the estimate ${\ensuremath{\hat{\mathcal{J}}}}$ is chosen based on the highest $T$ values in $\mathbf{z}$, which corresponds to majority voting. Observe that when knowledge is available about $J$, the [$\textup{\texttt{majority}}$]{}may be used also for the mixed support-set model, which we did (under another name) in [@Sundman:a_greedy_pursuit_algorithm]. Analysis -------- When the nodes have found the support-set estimates by [<span style="font-variant:small-caps;">gp</span>]{}algorithms, we use the input/output relation in Definition \[ass\] to provide some fundamental results valid for the [$\textup{\texttt{majority}}$]{}algorithm. \[prop:common\_cons\_l\] In a setup with $L = h+m$ sensors with signal support-sets ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l$ for $l = 1,2, \dots, h$ and ${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l$ for $l = 1,2, \dots, m$, let us assume that the index $i \in \left( \bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, \bigcap_{l = 1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} \right)$. Then, the [$\textup{\texttt{majority}}$]{}algorithm finds the estimate ${\ensuremath{\hat{\mathcal{J}}}}$ such that $i \in {\ensuremath{\hat{\mathcal{J}}}}$. In this case, the probability of detection is $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in \bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \phantom{=} = \frac{(1-\epsilon)^h \epsilon^m J}{(1-\epsilon)^h \epsilon^m J + (\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m (N-J)}, \label{eqn:common_cons_l} \end{aligned}$$ where $J = T$. Proof in Appendix \[app:majority\]. Getting any insight for the behavior of [$\textup{\texttt{majority}}$]{}from Proposition \[prop:common\_cons\_l\] is a non-trivial since is a complicated function of $m$, $h$, $J$, $T$ and $N$. For better understanding, we provide an example where some parameters are fixed. \[example1\] Using $N = 1000$, $T = J = 20$, we provide [Fig.]{} \[fig:common\_support\] where several pairs of $\{h,m\}$ are tested via . The black curve corresponds to the disconnected performance of Proposition  and the black dot corresponds to the probability of detect at $\epsilon_{\max} = \frac{N-T}{N}$ which is the biggest value $\epsilon$ can take. Worth noticing in this figure is the interplay between hits and misses which may cause the performance to be very good at some parts, while being poor at other parts. This is illustrated in the curve for $h=3$, $m = 7$. An observation we found is that whenever $h>m$ we get good performance. [$\square$]{} ![A few examples of probability of detect for the [$\textup{\texttt{majority}}$]{}algorithm using the common support-set model.[]{data-label="fig:common_support"}](figures/common_support.eps){width="\columnwidth"} Using [$\textup{\texttt{majority}}$]{}voting, it is intuitively clear that more votes are always better (for a constant number of total sensors in the network). We show this explicitly with the following proposition. \[prop:common\_cons\_ineq\] For the same setup as in Proposition \[prop:common\_cons\_l\] the following relation holds $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \hspace{2cm} \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l=1}^{h-1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \label{eqn:common_cons_ineq} \end{aligned}$$ Proof in Appendix \[app:majority\]. By [proposition]{} \[prop:common\_cons\_ineq\], it is clear that in a network of sensors, under the common support-set model, the [$\textup{\texttt{majority}}$]{}voting has a merit to detect the support-set. Mixed Support-set:\ Distributed Parallel Pursuit {#sec:mixed_model} ============================ We now consider the voting approach in a scenario where the signal correlation is modeled with the mixed support-set model. Assuming $T$ to be known (but not $J$), we previously developed such an algorithm in [@Sundman:dipp_arxiv], where it is called [$\textup{\texttt{consensus}}$]{}voting. In this case, there is no fusion center; instead the sensors exchange support-set estimates and apply the [$\textup{\texttt{consensus}}$]{}algorithm locally, based on information from the neighboring sensors. Algorithm {#algorithm} --------- The [$\textup{\texttt{consensus}}$]{}algorithm differs from [$\textup{\texttt{majority}}$]{}since it has no knowledge of the support-set size of the joint support $J$. Instead it performs a threshold operation by selecting components with at least two votes. We have provided [$\textup{\texttt{consensus}}$]{}in Algorithm \[alg:consensus\]. *Input:* $\{ \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{q}}}} \}_{{\ensuremath{\mathsf{q}}}\in \mathcal{L}_{{\ensuremath{\mathsf{p}}}}^{\text{in}}}$, $\hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}}$, $T$\ *Initialization:* $\mathbf{z} \leftarrow \mathbf{0}_{N\times 1}$ \[alg:cons:0\]\ *Algorithm:* $\mathbf{z} \leftarrow {\texttt{vote}_1}(\mathbf{z}, \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}})$ (${\ensuremath{\mathsf{p}}}$-th node’s estimate) \[alg:cons:own\] $\mathbf{z} \leftarrow {\texttt{vote}_1}(\mathbf{z}, \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{q}}}})$ (The neigbors’ estimates) \[alg:cons:others\] Choose $\hat{{\ensuremath{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ such that $(z(i) \geq 2) ~ \forall i \in {\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$\ and $| {\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}} | \leq T$ \[alg:cons:j\] Studying [algorithm]{} \[alg:consensus\], the inputs are: a set of estimated support-sets $\{ \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{q}}}} \}_{{\ensuremath{\mathsf{q}}}\in \mathcal{L}^{\text{in}}}$ from the neighbors, the locally estimated support-set $\hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}}$, and the sparsity level $T$. The estimate of $\hat{{\ensuremath{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ is formed (step \[alg:cons:j\]) such that no index in $\hat{{\ensuremath{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ has less than two votes (i.e., each index in $\hat{{\ensuremath{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ is present in at least two support-sets from $\{ \{ \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{q}}}} \}_{{\ensuremath{\mathsf{q}}}\in \mathcal{L}_{{\ensuremath{\mathsf{p}}}}^{\text{in}}}, \hat{{\mathcal{T}}}_{{\ensuremath{\mathsf{p}}}}\}$)[^2]. If the number of indices with at least two votes exceed the cardinality $T$, we pick the $T$ largest indices. Thus, the [$\textup{\texttt{consensus}}$]{}strategy can be summarized as: - Pick indices for ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ that have two votes - If $|{\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}| > T$, choose the $T$ largest indices In the following we will analyze the [$\textup{\texttt{consensus}}$]{}strategy using the input/output relation of Definition \[ass\]. Analysis {#sec:cons:analysis} -------- Assuming the nodes use [<span style="font-variant:small-caps;">gp</span>]{}algorithms to find the support-set estimates, we obtain the following results. \[cons:prop:pi\] The probability that an index ‘$i$’ is correct for sensor ‘${\ensuremath{\mathsf{p}}}$’, provided that this index is detected by the sensor ‘${\ensuremath{\mathsf{p}}}$’ itself and additionally ‘$h$’ neighbors, but not detected by ‘$m$’ neighbors is given by . Proof in Appendix \[app:consensus\]. \[cons:prop:pic\] The probability that an index ‘$i$’ is correct for sensor ‘${\ensuremath{\mathsf{p}}}$’, provided that this index is detected by ‘$h$’ neighbors, but not detected by the sensor ‘${\ensuremath{\mathsf{p}}}$’ itself and additionally ‘$m$’ neighbors is given by . Proof in Appendix \[app:consensus\]. $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} = \label{eqn:pi_phi_phii_phiiic} \\ & \frac{(1\!-\!\epsilon)^{h\!+\!1} \epsilon^{m} \frac{J}{N} \!+\! (1\!-\!\epsilon) (\frac{T}{N\!-\!T}\epsilon)^{h} (1 \!-\! \frac{T}{N\!-\!T}\epsilon)^{m} \frac{I}{N} } {(1\!-\!\epsilon)^{h\!+\!1} \epsilon^{m} \frac{J}{N} \!+\! (h+1) (1\!-\!\epsilon) (\frac{T}{N\!-\!T}\epsilon)^{h}(1\!-\!\frac{T}{N\!-\!T}\epsilon)^m \frac{I}{N} \!+\! m \epsilon (\frac{T}{N\!-\!T}\epsilon)^{h\!+\!1} (1\!-\!\frac{T}{N\!-\!T}\epsilon)^{m\!-\!1} \frac{I}{N} \!+\! (\frac{T}{N\!-\!T}\epsilon)^{h\!+\!1} (1\!-\!\frac{T}{N\!-\!T}\epsilon)^{m} \frac{N\!-\!J\!-\!(m\!+\!h\!+\!1)I}{N}}. \nonumber\end{aligned}$$ $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} = \label{eqn:pi_phic_phii_phiiic} \\ & \frac{(1\!-\!\epsilon)^h\epsilon^{m\!+\!1}\frac{J}{N} \!+\! \epsilon (\frac{T}{N\!-\!T}\epsilon)^h (1\!-\!\frac{T}{N\!-\!T}\epsilon)^m\frac{I}{N}}{(1\!-\!\epsilon)^{h} \epsilon^{m\!+\!1} \frac{J}{N} \!+\! h (1\!-\!\epsilon) (\frac{T}{N\!-\!T}\epsilon)^{h\!-\!1}(1\!-\!\frac{T}{N\!-\!T}\epsilon)^{m\!+\!1} \frac{I}{N} \!+\! (m\!+\!1) \epsilon (\frac{T}{N\!-\!T}\epsilon)^{h} (1\!-\!\frac{T}{N\!-\!T}\epsilon)^{m} \frac{I}{N} \!+\! (\frac{T}{N\!-\!T}\epsilon)^{h} (1\!-\!\frac{T}{N\!-\!T}\epsilon)^{m\!+\!1} \frac{N \!-\! J \!-\!(m\!+\!h\!+\!1)I}{N}}. \nonumber\end{aligned}$$ Getting any insight from these propositions is difficult. Therefore, we provide the following numerical example. \[example2\] In [Fig.]{} \[fig:mixed\_support\] we provide examples for the mixed support-set model using Proposition \[cons:prop:pi\] and Proposition \[cons:prop:pic\]. In this system $N = 1000$ and $\epsilon$ is varied. Notice in [Fig.]{} \[fig:mixed\_support\], that there are two curves for each configuration. The top-most curve corresponds to , where the sensor ‘${\ensuremath{\mathsf{p}}}$’ itself found the index, and the lower-most curve corresponds to , where the sensor ‘${\ensuremath{\mathsf{p}}}$’ itself missed the index. By testing it seems that, similarly to [$\textup{\texttt{majority}}$]{}voting, when $h > m$, the performance is good. ![Analytical results for the mixed support-set model. Observe here that there is always a total number of $h+m+1$ nodes present for a [$\textup{\texttt{consensus}}$]{}algorithm.[]{data-label="fig:mixed_support"}](figures/mixed_support.eps){width="\columnwidth"} Derivation of further general results based on Proposition \[cons:prop:pi\] and Proposition \[cons:prop:pic\], for example providing general precise requirements under which [$\textup{\texttt{consensus}}$]{}provides higher probability than a single sensor case is non-trivial. Instead, we assume a limited number of neighbors and fix a number of parameters according to [@Sundman:dipp_arxiv]. In particular, we assume that each local node ${\ensuremath{\mathsf{p}}}$ has two independent neighbors. Using information obtained by the neighbors and the local node, the [$\textup{\texttt{consensus}}$]{}endeavors to estimate the joint support part as ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$. Following [algorithm]{} \[alg:consensus\], we note that in this case $$\begin{aligned} {\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}} = \Bigl\{ i: i\in & \left( ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}) \cup ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}) \right. \nonumber \\ & \left. \phantom{=} \cup ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}) \right) \,\, \forall {\ensuremath{\mathsf{q}}},{\ensuremath{\mathsf{r}}}\in \mathcal{L}_{{\ensuremath{\mathsf{p}}}}^{\text{in}}, {\ensuremath{\mathsf{q}}}\neq {\ensuremath{\mathsf{r}}}\Bigr\}. \label{eqn:jh}\end{aligned}$$ Using , we provide the following remark in numerical manner. \[prop:ineq\_example\] When $N = 1000$, $T = 20$, $J = 15$, $V = 2$ and $0.0140 \leq \epsilon \leq \epsilon_{\max} = 0.98$, then the probability of $i\in{\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ to be correct is always bigger than or equal to the probability of $i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ to be correct, that is $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}\bigr)}} \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}. \end{aligned}$$ Proof in Appendix \[app:consensus\]. We note that although Remark \[prop:ineq\_example\] strongly suggest that the [$\textup{\texttt{majority}}$]{}voting provides for a good result, we can consider the typical [<span style="font-variant:small-caps;">cs</span>]{}condition that $N$ is very large. Then an even stronger result can be formulated as in the following corollary. \[cons:coro:inequalities\] If $J \geq 1$, $T$ grows sublinearly in $N$, and ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ is the output of [$\textup{\texttt{consensus}}$]{}, then $$\begin{aligned} \lim_{N \rightarrow \infty} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}\bigr)}} = 1. \end{aligned}$$ For this proof, we show that tends to one when $h \geq 1$ (follows from ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$) and that tends to one when $h \geq 2$ (also follows from ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$). First consider and note that it can never happen that $m > 0$ when $\epsilon = 0$. Then it is straight-forward to see that $(\frac{T}{N-T}\epsilon)^h \rightarrow 0$ since $h \geq 1$, hence the whole expression tends to 1. Similarly for , note that it can never happen that $m > 0$ when $\epsilon = 0$. Then it is straight-forward to see that $(\frac{T}{N-T}\epsilon)^{h-1} \rightarrow 0$ since $h \geq 2$, consequently the whole expression tends to 1. Experimental Evaluation {#sec:experiment} ======================= In this section we perform two experiments to illustrate the three results: Proposition \[prop:common\_cons\_l\], Proposition \[cons:prop:pi\] and Proposition \[cons:prop:pic\]. The goal of these experiments is to compare the analytical results with observations from a simulation process. Since there is no closed form result for $\epsilon$; this entity has to be estimated. We estimate $\epsilon$ in the same way as in the second test of Section \[sec:evaluation\_prob\_model\], and by averaging over all signals. To find the performance of the different voting strategies, we count how many times ‘$h$’ hits and ‘$m$’ misses correspond to a correct support-set index estimate and divide this number with the number of times ‘$h$’ hits and ‘$m$’ misses occurs in total. Thus the procedure is as follows: 1. Estimate $\hat{\epsilon}$. 2. For each $\hat{\epsilon}$, count the actual accuracy of the voting procedure and put a mark at this point. 3. Compare to the theoretical expression in the respective equation. In [Fig.]{} \[fig:common\_support\], $\epsilon$ is plotted vs the probability of detection for the results of the common support-set model. A total number of $10^6$ random simulations are performed, using parameters $N = 1000$, $T = 20$ and [<span style="font-variant:small-caps;">gp</span>]{}algorithm subspace pursuit ([$\textsc{sp}$]{}). To find different $\epsilon$, $M$ and ${\textsc{smnr}\xspace}$ are varied. In the case where $h = 2, m = 1$, the $\epsilon$ from left to right are found by: $M = 96, 85, 76, 64, 50, 41, 34, 28$ with corresponding ${\textsc{smnr}\xspace}= 20, 20, 20, 10, 10, 10, 10, 0$ and for the case where $h = 3, m = 7$, $M = 101, 96, 92, 88, 50, 41, 34, 28$ with corresponding ${\textsc{smnr}\xspace}= 20, 20, 20, 20, 10, 10, 10, 0$. Observe that largest possible $\epsilon_{\max} = \frac{N-T}{N}$ (marked with a small black dot). The equations used for the analytical results are found in . When we compare the simulations to the result predicted by analysis, we notice an almost perfect match. We argue that the slight mismatch for some points is due to noise and will average out using a larger simulation ensemble. For example, it is a rare event that $h = 3$ $m= 7$ occurs when the algorithms are very good (i.e., the simulation point at $(0.05, 0.05)$). ![Analytical and simulation results for the voting performance based on the common support-set model.[]{data-label="fig:common_support_simulation"}](figures/common_support_simulation.eps){width="\columnwidth"} In [Fig.]{} \[fig:mixed\_support\], $\epsilon$ is plotted vs the probability of detection for the mixed support-set model. A total number of $10^6$ random simulations are performed, using parameters $N = 1000$ and [<span style="font-variant:small-caps;">gp</span>]{}algorithm subspace pursuit ([$\textsc{sp}$]{}). To find different $\epsilon$, $M$ and ${\textsc{smnr}\xspace}$ are varied. Here we used the same data-points for all curves; the $\epsilon$ from left to right: $M = 96, 90, 85, 76, 64, 50, 41, 34, 28$ with corresponding ${\textsc{smnr}\xspace}= 20, 20, 20, 20, 10, 10, 10, 10, 0$. We observe that also here, the simulation points match closely to the predicted values. The equations used for the analytical results are found in and . ![Analytical and simulation results for the [$\textup{\texttt{consensus}}$]{}performance based on the mixed support-set model.[]{data-label="fig:mixed_support_simulation"}](figures/mixed_support_simulation.eps){width="\columnwidth"} Conclusion ========== In this paper, we have analyzed democratic based voting strategies for support-sets estimation using greedy algorithms. We have characterized the input/output relation of any typical [<span style="font-variant:small-caps;">gp</span>]{}algorithm based on four relations. Using these relations we shown the merit of voting for two particular examples: the [$\textup{\texttt{majority}}$]{}algorithm and the [$\textup{\texttt{consensus}}$]{}algorithm, both which has been presented in the literature earlier. With several experiments, we validated both the input/output relation and the results derived; in all cases the experiments closely matched the theoretical prediction. Proofs ====== Here, we present proofs for the propositions provided in the paper. First, we introduce some lemmas used in the proofs. Then, in [appendix]{} \[app:majority\] we present the proofs for [proposition]{} \[prop:common\_cons\_l\] and [proposition]{} \[prop:common\_cons\_ineq\]; in [appendix]{} \[app:consensus\] we present the proofs for [proposition]{} \[cons:prop:pi\], [proposition]{} \[cons:prop:pic\] and remark \[prop:ineq\_example\]. \[lemma:inverted\] For any $\mathcal{A} \subseteq {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$ and for any $\mathcal{B} \subseteq {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}$, the following holds $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl( i\in\mathcal{A} | i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & = \frac{|\mathcal{A}|}{T}{\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \label{lemma:inverted_true1} \\ {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in\mathcal{A} \bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \label{lemma:inverted_true2} \\ {\ensuremath{\mathbb{P}\,\!\bigl( i\in\mathcal{B} | i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & = \frac{|\mathcal{B}|}{T}{\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \bigr)}}, \label{lemma:inverted_false1} \\ {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in\mathcal{B} \bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \bigr)}}. \label{lemma:inverted_false2} \end{aligned}$$ $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl( i\in\mathcal{A} | i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & \overset{(a)}{=} \frac{|\mathcal{A}|}{T} {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & = \frac{|\mathcal{A}|}{T} {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i \in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}} \nonumber \\ & = \frac{|\mathcal{A}|}{T} {\ensuremath{\mathbb{P}\,\!\bigl(i \in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i \in {\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \end{aligned}$$ Here, $(a)$ follows directly from Definition \[ass\]. $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in\mathcal{A} \bigr)}} & = {\ensuremath{\mathbb{P}\,\!\bigl( i\in\mathcal{A} | i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \frac{ {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} }{ {\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{A} \bigr)}} } \nonumber \\ & \overset{(a)}{=} \frac{|\mathcal{A}|}{T}{\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}\frac{ {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} }{ {\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{A} \bigr)}} } \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \nonumber \end{aligned}$$ where we in $(a)$ applied . This proof is similar to the proof for , $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl( i\in\mathcal{B} | i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} & = \frac{|\mathcal{B}|}{|{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & = \frac{|\mathcal{B}|}{|{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}\frac{|{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|}{|{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|} \nonumber \\ & = \frac{|\mathcal{B}|}{T}{\ensuremath{\mathbb{P}\,\!\bigl( i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \bigr)}}. \nonumber \end{aligned}$$ This proof is similar to the proof for and follows directly by applying . \[lemma:indep\] The local results from different sensor nodes are independent over certain regions. Assume there are a total of $h+m$ nodes in the system and that we denote different nodes by sub-indices ${\ensuremath{\mathsf{p}}}_k \neq {\ensuremath{\mathsf{q}}}_l ~ \forall k,l$ and ${\ensuremath{\mathsf{p}}}\neq {\ensuremath{\mathsf{p}}}_k, {\ensuremath{\mathsf{p}}}\neq {\ensuremath{\mathsf{q}}}_l ~ \forall k,l $. Then, for $\mathcal{A} \subseteq {\ensuremath{\mathcal{J}}}$, $\mathcal{B} \subseteq {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h$, $\mathcal{C} \subseteq {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_m$, and $\mathcal{D} \subseteq ({\ensuremath{\mathcal{J}}}\cup \bigcup_{l=1}^h {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l \cup \bigcup_{l=1}^m {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l)^{{\ensuremath{\mathsf{c}}}}$ the following relations hold: $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{A}\bigr)}} \label{lemma:indep1} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^m, \nonumber \\ & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{B}\bigr)}} \label{lemma:indep2}\\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h-1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^m, \nonumber \\ & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{C}\bigr)}} \label{lemma:indep3} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{m-1}, \nonumber \\ & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{D}\bigr)}} \label{lemma:indep4} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^m. \nonumber \end{aligned}$$ $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{A}\bigr)}} \nonumber \\ & \overset{(a)}{=} \prod_{l=1}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l| i\in\mathcal{A}\bigr)}} \prod_{l=1}^m {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{A}\bigr)}}, \nonumber \\ & \overset{(b)}{=} \prod_{l=1}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l\bigr)}} \prod_{l=1}^m {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l\bigr)}}, \nonumber \\ & \overset{(c)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^m, \nonumber \end{aligned}$$ In $(a)$ we applied the chain rule on all intersections and applying the Markov property, which is illustrated in [Fig.]{} \[fig:app:flow\]. In $(b)$ we have used Lemma \[lemma:inverted\] and in $(c)$ we have used that all probabilities are equal. (Ip) [${\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$]{}; (J) [${\ensuremath{\mathcal{J}}}$]{}; (Iq) [${\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}$]{}; (add1) ; (divider) ; (add2) ; (s1) [ System ${\ensuremath{\mathsf{p}}}$ ]{}; (s2) [ System ${\ensuremath{\mathsf{q}}}$ ]{}; (Tp) [${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$]{}; (Tq) [${\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}$]{}; (Ip) – (add1); (Iq) – (add2); (J) – (divider); (divider) – (add1); (divider) – (add2); (add1) – node\[midway,above\] [${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}$]{} (s1); (add2) – node\[midway,above\] [${\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}$]{} (s2); (s1) – (Tp); (s2) – (Tq); (-1cm,1cm) – node\[midway,above\] [Mixed support-set model]{} (2cm, 1cm) – (2cm,-2.3cm) – (-1cm, -2.3cm) – (-1cm,1cm); (2cm,1cm) – node\[midway,above\] [System model]{} (6cm,1cm) – (6cm,-2.3cm) – (2cm,-2.3cm); $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{B}\bigr)}} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h, i\in\bigcap_{l=1}^{h-1} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{B}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h | i\in\mathcal{B}\bigr)}} \prod_{l=1}^{h-1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l| i\in\mathcal{B}\bigr)}} \nonumber \\ & \phantom{=} \cdot \prod_{l=1}^{m} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in\mathcal{B}\bigr)}} \nonumber \\ & \overset{(c)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h\bigr)}} \prod_{l=1}^{h-1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \phantom{=} \cdot \prod_{l=1}^{m}{\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \overset{(d)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h-1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^m. \nonumber \end{aligned}$$ In $(a)$, Lemma \[lemma:inverted\] is used, and in $(b)$ we used that all probabilities are equal. The proofs is similar to the proof for . The proofs is similar to the proof for . \[lemma:joint\_prob\] Assume there are $h+m$ nodes in the system, that ${\ensuremath{\mathsf{p}}}_k \neq {\ensuremath{\mathsf{q}}}_l ~ \forall k,l$ are different nodes. Then, the following holds: $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = (1-\epsilon)^{h} \epsilon^{m} \frac{J}{N} \nonumber \\ & \phantom{=} + h (1-\epsilon) (\frac{T}{N-T}\epsilon)^{h-1}(1-\frac{T}{N-T}\epsilon)^m \frac{I}{N} \nonumber \\ & \phantom{=} + m \epsilon (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m-1} \frac{I}{N} \nonumber \\ & \phantom{=} + (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m} \frac{N - J -(m+h)I}{N} \nonumber \end{aligned}$$ For this proof, we first introduce a notational simplification, define $\mathcal{U} = {\ensuremath{\mathcal{J}}}\cup {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_1 \cup \dots \cup {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_h \cup {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_1 \cup \dots \cup {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_m$. Observe that the sub-sets in $\mathcal{U}$ are non-overlapping. Then, $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \overset{(a)}{=} \sum_{ \hspace{-8pt}\mathcal{A} = \substack{{\ensuremath{\mathcal{J}}}, \\ {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_1, {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_2, \dots, {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_{h}, \\ {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_1, {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_2, \dots, {\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_{m}, \\ \mathcal{U}^{{\ensuremath{\mathsf{c}}}}}} \hspace{-20pt} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in\mathcal{A}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{A}\bigr)}} \nonumber \\ & \overset{(b)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{J}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}} \nonumber \\ & \phantom{=} + h {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}},i\in\bigcap_{l = 1}^{h-1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & \phantom{=} + m {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}},i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m-1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & \phantom{=} + {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in\mathcal{U}^{{\ensuremath{\mathsf{c}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{U}^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \overset{(c)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\bigr)}}^{m} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}} \nonumber \\ & \phantom{=} + h {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h-1} \nonumber \\ & \phantom{=} \phantom{=} \cdot {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^m {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & \phantom{=} + m {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h} \nonumber \\ & \phantom{=} \phantom{=} \cdot {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{m-1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & \phantom{=} + {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{m} {\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{U}^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \overset{(d)}{=} (1-\epsilon)^{h} \epsilon^{m} \frac{J}{N} \nonumber \\ & \phantom{=} + h (1-\epsilon) (\frac{T}{N-T}\epsilon)^{h-1}(1-\frac{T}{N-T}\epsilon)^m \frac{I}{N} \nonumber \\ & \phantom{=} + m \epsilon (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m-1} \frac{I}{N} \nonumber \\ & \phantom{=} + (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m} \frac{N - J -(m+h)I}{N}. \nonumber\end{aligned}$$ In $(a)$, the probability is marginalized over all individual and joint support-sets, and over $\mathcal{U}$. In $(b)$, we extend the sum. In $(c)$ we apply Lemma \[lemma:indep\]. Lastly, in $(d)$ we plug the values from Definition \[ass\]. Proofs of the Results for [$\textup{\texttt{majority}}$]{} {#app:majority} ---------------------------------------------------------- Here, we prove Proposition \[prop:common\_cons\_l\] and Proposition \[prop:common\_cons\_ineq\], which are stated based on the common support-set model. Recall that in the common support-set model, ${\ensuremath{\mathcal{J}}}= {\mathcal{T}}$ and $J = T$. $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in \bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l=1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l =1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}}}{\sum_{\mathcal{A} = {\ensuremath{\mathcal{J}}}, {\ensuremath{\mathcal{J}}}^{{\ensuremath{\mathsf{c}}}}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m} {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}| i\in\mathcal{A}\bigr)}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{A}\bigr)}}} \nonumber \\ & \overset{(a)}{=} \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^h {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^m {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}}}{\sum_{\mathcal{A} = {\ensuremath{\mathcal{J}}}, {\ensuremath{\mathcal{J}}}^{{\ensuremath{\mathsf{c}}}}}{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in\mathcal{A}\bigr)}}^h{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|i\in\mathcal{A}\bigr)}}^m{\ensuremath{\mathbb{P}\,\!\bigl(i\in\mathcal{A}\bigr)}}} \nonumber \\ & \overset{(b)}{=} \frac{(1-\epsilon)^h \epsilon^m \frac{J}{\cancel{N}}}{(1-\epsilon)^h \epsilon^m \frac{J}{\cancel{N}} + (\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m \frac{N-J}{\cancel{N}}} \end{aligned}$$ In $(a)$ Lemma \[lemma:indep\] is applied, and lastly for $(b)$, Definition \[ass\] is used. This proposition states that he following inequality holds $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l=1}^{h-1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \label{proof:prop:common_cons_ineq} \end{aligned}$$ Using Proposition \[prop:common\_cons\_l\], the LHS of is: $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l=1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \phantom{=}= \frac{(1-\epsilon)^h \epsilon^m J}{(1-\epsilon)^h \epsilon^m J + (\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m (N-J)}. \label{eqn:lhs1} \end{aligned}$$ Similarly, the RHS of is $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in\bigcap_{l=1}^{h-1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l=1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \phantom{=}= \frac{(1-\epsilon)^{h\!-\!1} \epsilon^{m\!+\!1} J}{(1\!-\!\epsilon)^{h\!-\!1} \epsilon^{m\!+\!1} J + (\frac{T}{N-T}\epsilon)^{h\!-\!1} (1\!-\!\frac{T}{N-T}\epsilon)^{m\!+\!1} (N\!-\!J)}. \label{eqn:rhs1} \end{aligned}$$ By plugging the and into the inequality we get: $$\begin{aligned} & \frac{(1-\epsilon)^{\cancel{h}} \cancel{\epsilon^m} \cancel{J}}{(1-\epsilon)^h \epsilon^m J + (\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m (N-J)} \nonumber \\ & \geq \nonumber \\ & \frac{\cancel{(1-\epsilon)^{h\!-\!1}} \epsilon^{\,\,\cancel{m\!+\!1}} \cancel{J}}{(1\!-\!\epsilon)^{h\!-\!1} \epsilon^{m\!+\!1} J + (\frac{T}{N-T}\epsilon)^{h\!-\!1} (1\!-\!\frac{T}{N-T}\epsilon)^{m\!+\!1} (N\!-\!J)} \nonumber \end{aligned}$$ Multiplying the denominators gives $$\begin{aligned} & (1-\epsilon) \left( (1\!-\!\epsilon)^{h\!-\!1} \epsilon^{m\!+\!1} J \right. \nonumber \\ & \phantom{(1-\epsilon)} \left. + (\frac{T}{N-T}\epsilon)^{h\!-\!1} (1\!-\!\frac{T}{N-T}\epsilon)^{m\!+\!1} (N\!-\!J) \right) \nonumber \\ & \geq \nonumber \\ & \epsilon \left( (1-\epsilon)^h \epsilon^m J + (\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m (N-J) \right),\nonumber \end{aligned}$$ which by simplifying gives $$\begin{aligned} & \cancel{(1\!-\!\epsilon)^{h} \epsilon^{m\!+\!1} J} \nonumber \\ & \phantom{=} + (1-\epsilon)(\frac{T}{N-T}\epsilon)^{h\!-\!1} (1\!-\!\frac{T}{N-T}\epsilon)^{m\!+\!1} (N\!-\!J) \nonumber \\ & \geq \nonumber \\ & \cancel{(1-\epsilon)^h \epsilon^{m+1} J} + \epsilon(\frac{T}{N-T}\epsilon)^h (1 - \frac{T}{N-T}\epsilon)^m (N-J).\nonumber \end{aligned}$$ Further simplifications give $$\begin{aligned} & (1-\epsilon)\cancel{(\frac{T}{N-T}\epsilon)^{h\!-\!1}} (1\!-\!\frac{T}{N-T}\epsilon)^{\cancel{m\!+\!1}} \cancel{(N\!-\!J)} \nonumber \\ & \geq \nonumber \\ & \epsilon(\frac{T}{N-T}\epsilon)^{\cancel{h}} \cancel{(1 - \frac{T}{N-T}\epsilon)^m} \cancel{(N-J)}.\nonumber \end{aligned}$$ Thus, we arrive at $$\begin{aligned} (1-\frac{T}{N-T}\epsilon) (1-\epsilon) \geq \frac{T}{N-T}\epsilon^2, \nonumber \end{aligned}$$ which can be simplified to $$\begin{aligned} 1 - \frac{N}{N-T}\epsilon + \cancel{\frac{T}{N-T}\epsilon^2} \geq \cancel{\frac{T}{N-T}\epsilon^2}. \nonumber \end{aligned}$$ This is in turn equivalent to $$\begin{aligned} \frac{N}{N-T}\epsilon \leq 1, \nonumber \end{aligned}$$ where the expression reaches its maximum at $\epsilon_{\max} = \frac{N-T}{N}$ $$\begin{aligned} \frac{N}{N-T}\epsilon \leq \frac{N}{N-T}\frac{N-T}{N} = 1. \nonumber \end{aligned}$$ Thus, we conclude that the sought inequality holds true. Proof of the Results for [$\textup{\texttt{consensus}}$]{} {#app:consensus} ---------------------------------------------------------- We now prove Proposition \[cons:prop:pi\], Proposition \[cons:prop:pic\] and Remark \[prop:ineq\_example\], which are based on the mixed support-set model. We first notice that the problem can be split into two parts, $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \label{eqn:pi_j_thi_thii_thiii} \\ & \phantom{=} + {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}, \label{eqn:pi_i_thi_thii_thiii} \end{aligned}$$ where we consider each part separately. - First we study $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}}. \label{eqn:pi_part1} \end{aligned}$$ We now consider each probability in separately, beginning with $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} \nonumber \\ & \overset{(a)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{h+1} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{m} \nonumber \\ & \overset{(b)}{=} (1-\epsilon)^{h+1} \epsilon^{m}. \label{eqn:pi_part2} \end{aligned}$$ Here, Lemma \[lemma:indep\] was used in $(a)$ and Definition \[ass\] was used in $(b)$. We have from the uniformity of the support-sets that $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}} = \frac{J}{N}. \end{aligned}$$ Finally we have $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & \overset{(a)}{=} (1-\epsilon)^{h+1} \epsilon^{m} \frac{J}{N} \nonumber \\ & \phantom{=} + (h+1) (1-\epsilon) (\frac{T}{N-T}\epsilon)^{h}(1-\frac{T}{N-T}\epsilon)^m \frac{I}{N} \nonumber \\ & \phantom{=} + m \epsilon (\frac{T}{N-T}\epsilon)^{h+1} (1-\frac{T}{N-T}\epsilon)^{m-1} \frac{I}{N} \nonumber \\ & \phantom{=} + (\frac{T}{N-T}\epsilon)^{h+1} (1-\frac{T}{N-T}\epsilon)^{m} \frac{N-J-(m+h+1)I}{N}, \label{eqn:pi_part3} \end{aligned}$$ where $(a)$ is achieved by Lemma \[lemma:joint\_prob\]. - We now study $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}}. \label{eqn:pi_part4} \end{aligned}$$ We now consider each probability in separately, beginning with $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & \overset{(a)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{h} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}|i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^{m} \nonumber \\ & \overset{(b)}{=} (1-\epsilon) (\frac{T}{N-T}\epsilon)^{h} (1 - (\frac{T}{N-T}\epsilon))^{m}, \nonumber \end{aligned}$$ where we just as for , used Lemma \[lemma:indep\] for $(a)$ and Definition \[ass\] for $(b)$. We have from the uniformity of support-sets that, $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} = \frac{I}{N}. \end{aligned}$$ Finally we notice for the third probability that the denominator is identical to . Now plugging the parts together gives . This proof is similar to the proof of Proposition \[cons:prop:pi\]. First, split the problem into two parts, $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \label{eqn:pic_j_thi_thii_thiii} \\ & \phantom{=} + {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}, \label{eqn:pic_i_thi_thii_thiii} \end{aligned}$$ and we study each part separately. - First study $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}}. \label{eqn:pic_part1} \end{aligned}$$ This was achieved using Bayes’ rule. We now consider each probability in separately, beginning with $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{J}}}\bigr)}} \\ & \overset{(a)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^h {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}^{m+1} \\ & \overset{(b)}{=} (1-\epsilon)^h\epsilon^{m+1} \end{aligned}$$ Here, Lemma \[lemma:indep\] was used in $(a)$ and Definition \[ass\] was used in $(b)$. We have from the uniformity of the support-sets that $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{J}}}\bigr)}} = \frac{J}{N}. \end{aligned}$$ Finally we have $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \\ & \overset{(a)}{=} (1-\epsilon)^{h} \epsilon^{m+1} \frac{J}{N} \nonumber \\ & \phantom{=} + h (1-\epsilon) (\frac{T}{N-T}\epsilon)^{h-1}(1-\frac{T}{N-T}\epsilon)^{m+1} \frac{I}{N} \nonumber \\ & \phantom{=} + (m+1) \epsilon (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m} \frac{I}{N} \nonumber\\ & \phantom{=} + (\frac{T}{N-T}\epsilon)^{h} (1-\frac{T}{N-T}\epsilon)^{m+1} \frac{N - J -(m+h+1)I}{N}, \label{eqn:pic_part3} \end{aligned}$$ where $(a)$ is obtained by Lemma \[lemma:joint\_prob\]. - For we have that $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}} \nonumber \\ & = \frac{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}}{{\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}}\bigr)}}}, \label{eqn:pic_part2} \end{aligned}$$ which is achieved with Bayes’ rule. We now consider each probability in separately, beginning with the first probability $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}, i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} \nonumber \\ & = {\ensuremath{\mathbb{P}\,\!\bigl(i\in\bigcap_{l = 1}^{h}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}_l, i\in\bigcap_{l = 1}^{m+1}{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_l^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{r}}}}}_{{m+1}}\bigr)}} \nonumber \\ & \overset{(a)}{=} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}| i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{q}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^h {\ensuremath{\mathbb{P}\,\!\bigl({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}} | i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}\bigr)}}^m \nonumber \\ & \overset{(b)}{=} \epsilon (\frac{T}{N-T}\epsilon)^h (1-\frac{T}{N-T}\epsilon)^m, \nonumber \end{aligned}$$ where we used Lemma \[lemma:indep\] for $(a)$ and Definition \[ass\] for $(b)$. We have from the uniformity of support-sets that, $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{I}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} = \frac{I}{N}. \nonumber \end{aligned}$$ Finally we notice for the third probability that the denominator is identical to . Now plugging all the parts together gives . We first study and notice that any index in $i\in{\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ fulfills one of the following: $i\in({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})$, or $i\in({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}})$, or $i\in({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})$. Thus, we will show the remark by proving each of the following inequalities: $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})\bigr)}} & \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \label{prop:ineq_example_1} \\ {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}})\bigr)}} & \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}, \label{prop:ineq_example_2} \\ {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})\bigr)}} & \geq {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}}. \label{prop:ineq_example_3} \end{aligned}$$ First, recall from Proposition \[prop:disconnect\] and that $$\begin{aligned} {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in{\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\bigr)}} = 1 - \epsilon. \label{eqn:ex1_1} \end{aligned}$$ - We now consider . By plugging $N = 1000$, $T = 20$, $J = 15$, $I = 5$, $m = 0$ and $h = 2$ into Proposition \[cons:prop:pi\], we obtain $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})\bigr)}} \label{eqn:ex1_2} \\ & = \frac{49(\epsilon - 1)(7204\epsilon^2 - 14406\epsilon + 7203)}{352900\epsilon^3 - 1058988\epsilon^2 + 1058841\epsilon - 352947}. \nonumber \end{aligned}$$ We multiply the denominator of to get the following inequality $$\begin{aligned} & 49(\epsilon - 1)(7204\epsilon^2 - 14406\epsilon + 7203) \nonumber \\ & \geq (1-\epsilon)(352900\epsilon^3 - 1058988\epsilon^2 + 1058841\epsilon - 352947), \nonumber \end{aligned}$$ which equivalently can be simplified to $$\begin{aligned} 0 \geq (\epsilon(50\epsilon - 49)(7058\epsilon - 7203)(\epsilon - 1))/23529800. \label{eqn:zeros1} \end{aligned}$$ The roots to the polynomial of are: $\epsilon_1 = 0$, $\epsilon_2 = \frac{49}{50} = 0.98$, $\epsilon_3 = \frac{7203}{7058} = 1.0205...$ and $\epsilon_4 = 1$. Thus, the interesting region is $\epsilon \in [\epsilon_1, \epsilon_2]$, for which the inequality holds. - We now consider . By plugging $N = 1000$, $T = 20$, $J = 15$, $I = 5$, $m = 1$ and $h = 1$ into Proposition \[cons:prop:pi\], we obtain $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}})\bigr)}} \nonumber \\ & = \frac{196\epsilon(1801\epsilon^2 - 3614\epsilon + 1813)}{\epsilon(352900\epsilon^2 - 701288\epsilon + 357749)}. \label{eqn:ex3_1} \end{aligned}$$ Observe that also here, $\epsilon = 0$ is undefined. We multiply the denominator of to and get the following inequality $$\begin{aligned} & 196\epsilon(1801\epsilon^2 - 3614\epsilon + 1813) \nonumber \\ & \geq \epsilon(1-\epsilon)(352900\epsilon^2 - 701288\epsilon + 357749), \nonumber \end{aligned}$$ which can be simplified to $$\begin{aligned} 0 \geq -(\epsilon(50\epsilon - 49)(7058\epsilon - 49)(\epsilon - 1))/23529800. \label{eqn:zeros2} \end{aligned}$$ The roots to the polynomial of are: $\epsilon_1 = 0$, $\epsilon_2 = \frac{49}{50} = 0.98$, $\epsilon_3 = \frac{49}{7058} = 0.0069...$ and $\epsilon_4 = 1$. Thus, the interesting region is $\epsilon \in [\epsilon_3, \epsilon_4]$, for which the inequality holds. - We now consider . By plugging $N = 1000$, $T = 20$, $J = 15$, $I = 5$, $m = 0$ and $h = 2$ into Proposition \[cons:prop:pic\], we obtain $$\begin{aligned} & {\ensuremath{\mathbb{P}\,\!\bigl(i\in{\ensuremath{\mathcal{T}}\sc@sub{{\ensuremath{\mathsf{p}}}}}| i\in ({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}^{{\ensuremath{\mathsf{c}}}} \cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}})\bigr)}} \nonumber \\ & = \frac{49\epsilon(7204\epsilon^2 - 14406\epsilon + 7203)}{\epsilon(352900\epsilon^2 - 701288\epsilon + 357749)}. \label{eqn:ex2_1} \end{aligned}$$ Observe that in this expression, $\epsilon = 0$ is undefined. This is naturally true[^3] and follows directly from Lemma \[lemma:joint\_prob\]. We multiply the denominator of to and get the following inequality $$\begin{aligned} & 49\epsilon(7204\epsilon^2 - 14406\epsilon + 7203) \nonumber \\ & \geq \epsilon(1-\epsilon)(352900\epsilon^2 - 701288\epsilon + 357749), \nonumber \end{aligned}$$ which can be simplified to $$\begin{aligned} 0 \geq -(\epsilon(50\epsilon - 49)(7058\epsilon^2 - 7107\epsilon + 98))/23529800. \label{eqn:zeros3} \end{aligned}$$ The roots to the polynomial of are: $\epsilon_1 = 0$, $\epsilon_2 = \frac{49}{50} = 0.98$, $\epsilon_3 = \frac{7107}{2\cdot 7058} + \sqrt{\left(\frac{7107}{2\cdot 7058}\right)^2 - \frac{98}{7058}} = 0.9930...$ and $\epsilon_4 = \frac{7107}{2\cdot 7058} - \sqrt{\left(\frac{7107}{2\cdot 7058}\right)^2 - \frac{98}{7058}} = 0.0140...$. Thus, the interesting region is $\epsilon \in [\epsilon_4, \epsilon_3]$, for which the inequality holds. From the above calculations, we find the interesting region is the region that lies between $\epsilon_4 \geq 0.0140$ for and $\epsilon_{\max} = \frac{N-T}{N} = 0.98$. Since all the inequalities , or hold true in this region (directly verified by plugging in any $0.0140 \leq \epsilon \leq 0.98$), we conclude the proof. [^1]: The authors are with Communication Theory Laboratory, School of Electrical Engineering, KTH - Royal Institute of Technology, Sweden. Emails: $\{$denniss, sach, skoglund$\}$@kth.se [^2]: For node ${\ensuremath{\mathsf{p}}}$, this is equivalent to let algorithm choose ${\ensuremath{\hat{\mathcal{J}}}}_{{\ensuremath{\mathsf{p}}}}$ as the union of all pair-wise intersections of support-sets (see the analysis section \[sec:cons:analysis\] for details). [^3]: If all algorithms are perfect, the cut $({\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{p}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{q}}}}}\cap {\ensuremath{\hat{\mathcal{T}}}\sc@sub{{\ensuremath{\mathsf{r}}}}}^{{\ensuremath{\mathsf{c}}}}) = \emptyset$.

--- abstract: | We construct space-time stationary solutions of the $1$D Burgers equation with random forcing in the absence of periodicity or any other compactness assumptions. More precisely, for the forcing given by a homogeneous Poissonian point field in space-time we prove that there is a unique global solution with any prescribed average velocity. These global solutions serve as one-point random attractors for the infinite-dimensional dynamical system associated to solutions to the Cauchy problem. The probability distribution of the global solutions defines a stationary distribution for the corresponding Markov process. We describe a broad class of initial Cauchy data for which the distribution of the Markov process converges to the above stationary distribution. Our construction of the global solutions is based on a study of the field of action-minimizing curves. We prove that for an arbitrary value of the average velocity, with probability 1 there exists a unique field of action-minimizing curves initiated at all of the Poissonian points. Moreover action-minimizing curves corresponding to different starting points merge with each other in finite time. address: - | School of Mathematics\ Georgia Institute of Technology\ 686 Cherry Street\ Atlanta, GA, 30332-0160, USA - | Delft University of Technology\ Mekelweg 4, 2628 CD Delft, The Netherlands - | Department of Mathematics\ University of Toronto\ 40 St George Street\ Toronto, Ontario, M5S 2E4, Canada author: - Yuri Bakhtin - Eric Cator - Konstantin Khanin bibliography: - 'Burgers.bib' title: 'Space-time stationary solutions for the Burgers equation' --- [^1] [^2] Introduction {#sec:intro} ============ The Burgers equation is one of the basic hydrodynamic models. It describes the evolution of velocity field of sticky particles that interact with each other only when they collide. In one space dimension, the inviscid Burgers equation is $$\label{eq:Burgers} \partial_t u(t,x)+u(t,x)\partial_xu(t,x)=f(t,x).$$ Here $u(t,x)$ is the velocity of the particle located at $x\in{{\mathbb R}}$ at time $t\in{{\mathbb R}}$. The external “forcing” term $f(t,x)$ describes the accelerations of particles. Although typically solutions of this equation develop discontinuities (shocks) in finite time, one can work with generalized solutions. So called entropy solutions, or viscosity solutions, are globally well-defined and unique for a broad class of initial velocity profiles and forcing terms. In this paper we study the long term behavior of the Burgers dynamics for the situation where the right-hand side $f(t,x)$ is a space-time stationary random process. In particular, we construct space-time stationary global solutions for the Burgers equation on the real line and show that they can be viewed as one-point attractors. The results that we present here are connected with two big streams of research developed in the last twenty years. The first one concerns stationary solutions and invariant measures for the randomly forced Burgers equation in a compact setting such as periodic forcing that effectively reduces the system to dynamics on a circle, or torus in the multidimensional version of equation , see [@ekms:MR1779561], [@Iturriaga:MR1952472], [@Gomes-Iturriaga-Khanin:MR2241814], or Burgers dynamics with random boundary conditions, see [@yb:MR2299503]. The main tool in the proof of these results is the Lax–Oleinik variational principle that allows for an efficient analysis of the system via studying the minimizers of the corresponding random Lagrangian system. Namely, the velocity field can be represented as $u(t,x)=\partial_xU(t,x)$, where the potential $U(t,x)$ is a solution of the Hamilton–Jacobi equation $$\label{eq:Hamilton-Jacobi-forced} \partial_t U(x,t)+\frac{(\partial_x U(x,t))^2}{2}+F(x,t)=0,$$ $F$ being the forcing potential: $\partial_x F(x,t)=-f(x,t)$. The entropy solution to Cauchy problem for this equation with initial data $U(\cdot,t_0)=U_0(\cdot)$ can be written as $$\label{eq:LO} U(t,x)=\inf_{\gamma: [t_0,t]\to{{\mathbb R}}}\left\{U_0(\gamma(t_0))+{\int_{t_0}^t\left[\frac{1}{2}\dot\gamma^2(s)-F(s,\gamma(s))\right]} ds\right\},$$ where the infimum is taken over all absolutely continuous curves $\gamma$ satisfying $\gamma(t)=x$. The long-term behavior of the Burgers dynamics thus can be understood by studying this variational problem over long time intervals. For example, to construct stationary solutions, one has to study the behavior of Lagrangian action minimizers on an interval $[-T,t]$ as $T\to\infty$. In particular, it was proved in [@ekms:MR1779561] for the one-dimensional case and in [@Iturriaga:MR1952472] for the multi-dimensional case that in the compact situation, for each space point $x$, the finite-time minimizers on $[-T,t]$ with endpoint $x$ stabilize to a unique infinite one-sided minimizer. Moreover, there is a unique global minimizer which is a hyperbolic trajectory of the stochastic Lagrangian flow. The uniqueness of one-sided minimizers implies the following One Force – One Solution (1F1S) principle for the Burgers equation: for almost every realization of the forcing, there is a unique global solution with given average velocity, and at any given time it depends only on the history of the forcing up to that time. The compactness condition in the above papers was extremely important. Very few results have been available in noncompact setting. These results were dealing with the case where the forcing was spatially nonhomogeneous and effectively compact, see [@Khanin-Hoang:MR1975784] and [@Bakhtin-quasicompact]. The aim of our paper is to consider forcing with homogeneous probability distribution in space and time. In other words, our goal is to replace the exact periodicity (compactness) with statistical homogeneity, namely, translational invariance in distribution. Although the 1F1S principle is a simple conceptual fact in the compact situation, it is far from obvious that stabilization of solutions occurs in the noncompact setting. Indeed, one can easily imagine a scenario where the solutions corresponding to different $T$ will be mostly influenced by the realization of the forcing in completely different spatio-temporal patches. The main challenge in the problem, is to show that this is not the case. To tackle this difficulty, we use the second stream of research mentioned above, the theory of one-sided infinite geodesics for first passage and last passage percolation models. A simple model of that kind can be introduced in the following way: consider a random potential on the lattice ${{\mathbb Z}}^2$. For a finite path on the lattice its action is defined as the sum of all values of the potential along the path. A path is called a geodesic between two points if it minimizes the action among all paths connecting these two points. An infinite path is called an infinite geodesic if every finite part of it is a geodesic between its endpoints. The theory developed by Howard and Newman in [@Ne],[@HoNe3],[@HoNe2],[@HoNe] describes the structure of such infinite length geodesics. It turns out that each infinite geodesic has an asymptotic direction and for a fixed direction the geodesic *tree* spans the entire lattice. These type of results can be extended to a Poissonian point field setting where the disorder is due to random distances between configuration points. The problem is also closely related to the study of Hammersley process (see [@Aldous-Diaconis:MR1355056],[@Johansson:MR1757595],[@Wu],[@CaPi],[@CaPi-ptrf]) which is based on optimal upright paths through Poissonian clouds. In all cases the existence of infinite geodesics is proved using techniques going back to the work of Kesten [@Kesten:MR1221154]. In the present paper we use in a systematic way both Lagrangian methods developed in Aubry–Mather theory for action-minimizing trajectories and weak KAM theory for Hamilton–Jacobi equation (see [@Fathi]), and probabilistic techniques related to the first passage percolation problem. We consider equation  and assume that the random forcing $f(x,t)=-\partial_xF(x,t)$ is associated with a space-time homogeneous Poisson point field. One can think of every Poissonian point as a source of a localized delta-type potential. This means that $\int_{t_0}^t F(s,\gamma(s))ds$ in is equal to the number of Poissonian points visited by the curve $\gamma(s), s \in [t_0,t]$. Such singular random potentials were first introduced in the context of Burgers equation in [@Bakhtin-quasicompact]. For systems with this type of forcing, shocks are created at the Poissonian points. One can say that the main focus of analysis is to study forward dynamics of shocks and their merging and the backward dynamics of minimizers (action-minimizing curves) and their coalescence. The crucial difference between [@Bakhtin-quasicompact] and the present paper is that the Poissonian field in [@Bakhtin-quasicompact] is assumed to be very nonhomogeneous in space with density decaying to zero at infinity. This results in quasi-compactness for the problem, and the asymptotic analysis of the model is related to the one in the compact case. In particular there exists a unique global minimizer and all other minimizers coalesce with it (see also [@Khanin-Hoang:MR1975784]). By imposing the homogeneity condition on the density of the Poissonian point field, we are forced to deal with a completely noncompact situation. Despite the fact that the global behavior here is quite different (in particular, a global minimizer does not exist), one still can prove almost sure existence and uniqueness of global solutions for fixed “average” velocity. The global solutions are stationary in time and space which reflects the translation-invariance of the probability distribution for the forcing. The construction is based on a study of a global tree of action-minimizing curves associated with every Poissonian point. We finish the introduction with a discussion of future directions. We believe that the theory we discuss in this paper opens a wide research area and below we formulate the most interesting and important open problems. We plan to address some of the problems listed below in the future. The first step in developing a general theory is elimination of the singular character of the forcing. There are several settings where such elimination looks plausible. Consider a smooth nonnegative potential $F(t,x)$ vanishing outside of a disk of small radius ${{\varepsilon }}>0$ and equal to ${{\varepsilon }}^{-1}$ everywhere in the disk except for a thin boundary layer. We then consider a homogeneous Poissonian point field and assume that every Poissonian point $(t_i,x_i)$ contributes $F(t-t_i,x-x_i)$ to the external potential, so that the total potential is given by the sum of these contributions over all configuration points. In a certain sense the system considered in the present paper corresponds to the limit ${{\varepsilon }}\to 0$. We believe that similar results hold for the model with small positive ${{\varepsilon }}>0$. Namely, for ${{\varepsilon }}$ small enough, there exists a stationary global solution that approaches the global solution constructed in our paper as ${{\varepsilon }}\to 0$. Another interesting generalization concerns systems with kick forcing. The corresponding variational problem will look the following way. $$U_v(x_0,0)=\min_{(x_i)_{i=-\infty}^0}\left[ \sum_{i}A_i(x_i,x_{i+1})+\frac{(x_{i+1}-x_i-v)^2}{2}\right].$$ Here the right-hand side can be defined up to an additive constant. This means that the increments $U(x_0,0)-U(0,0)$ are well-defined provided that minimizing orbits for different starting points $x_0$ are asymptotic to each other. We shall assume that random kernels $A_i$ are independent copies of a double stationary process $A(x,y)$ (i.e., the distribution of the process $A(x+a,y+b)$ does not depend on $a,b$) with fast enough decay of correlations. The condition of double stationarity is important since it guarantees that the shape function (see Section \[sec:subadd\]) is quadratic. However, it is natural to conjecture that existence and uniqueness of a global solution only requires that the distribution of $A_i(x+a,y+a)$ does not depend on $a$. Of course, the most general result will apply to any reasonable stationary forcing potential $F(t,x)$ in equation with fast decay of correlations. However, this problem is technically too hard at present. It is also not clear how far this program can be pushed in higher dimensions. We believe that extension of our results in this direction will use Lagrangian methods in a very essential way. Another set of problems is connected to the viscous case. In the case of singular potential corresponding to Poissonian points, positive viscosity makes them irrelevant. However, for the smoothed problems discussed above, the interplay between the viscosity and potential is nontrivial. It will be interesting to construct global stationary solutions for positive viscosity $\nu$ and show that they approach the inviscid ones as $\nu\to 0$. In the case of positive viscosity, the solution on a finite time interval is determined not by a single action minimizing path, but by a random probability distribution on the space of paths of fixed length (directed polymers). It is tempting to make a conjecture that with probability 1 these random probability distributions have a random limit as the time interval $[-T,t]$ converges to $(-\infty, t]$ and the endpoint of the paths $\gamma(t)$ stays fixed. The limiting probability distributions can be considered as Gibbs measures on the interval $(-\infty,t]$. These Gibbs measures most probably cannot be extended to $(-\infty, +\infty)$. The obstruction to such an extension is related to strong fluctuations of Gibbs measure as $t\to \infty$. This is similar to the mechanism preventing existence of global minimizers. Finally we briefly discuss connection with a very active area of KPZ scalings. The global solution to the random Hamilton–Jacobi equation $U(x,t)$ constructed in the present paper is a process with stationary increments. If the increments are also weakly dependent for distant intervals then one should expect to have a Gaussian behavior with usual CLT scalings for increments on large intervals. Namely, for some $\sigma>0$, $$\label{eq:g} \frac{U(x,0) - U(0,0) }{\sqrt{|x|}} \stackrel{distr}{\longrightarrow} \mathcal{N}(0,\sigma^2).$$ In fact, an even stronger statement of the type of the invariance principle should hold true. It is quite easy to see that such a “non-exotic” asymptotic behavior will imply KPZ scalings for the limiting minimizer. To make a precise statement, denote by $\gamma(t), t \in (-\infty,0]$ the one-sided minimizer with zero average velocity which originates at the origin at time zero. Then fluctuations of $\gamma(-t)$ and $U(\gamma(-t), -t) - U(0, 0)$ must be of the order $t^{2/3}$ and $t^{1/3}$ respectively. This follows from (\[eq:g\]) and the balance condition between $U(\gamma(-t), -t) - U(0, -t)$ and the contribution coming from the shape function. It is expected that the probability distribution for $$\left(\frac{\gamma(-t)}{t^{2/3}}, \, \frac{U(\gamma(-t), -t) - U(0, 0)}{t^{1/3}}\right)$$ converges to the universal limit which is related to Tracy–Widom distribution for the GOE random matrix ensemble. The exact form of the limiting distribution cannot be derived from the qualitative argument above. However, the universality of the limit law may be a more realistic target. We believe that the limiting behavior is determined by the asymptotic properties of stationary random point field of shocks, i.e., the points where $U(x,0)$ is not smooth. Notice, however, that the field of shocks at a given time is equipped with a random parameter, namely age of a shock, attached to every point. The age is a time interval for which a given shock can be traced in the past. It looks plausible that asymptotic statistical properties of the point field of “aged” shock determines the KPZ-type limits completely. The setting and notation {#sec:notation_and_results} ======================== The forcing in our system is given by a Poissonian point field $\omega$ on space-time ${{\mathbb R}}\times{{\mathbb R}}={{\mathbb R}}^2$ with Lebesgue intensity measure. Throughout the paper we adopt the picture where the space axis of ${{\mathbb R}}^2$ is horizontal and the time axis is vertical and directed upward. The configuration space $\Omega_0$ is the space of all locally finite point sets in space-time. For a Borel set $A\subset{{\mathbb R}}^2$, we use $\omega(A)$ to denote the number of Poissonian points in $A$, and introduce the $\sigma$-algebra ${{\mathcal F}}_0$ generated by maps $\omega\mapsto \omega(A)$ with $A$ running through all bounded Borel sets. The probability measure ${{\mathbb P}}_0$ is such that for any bounded Borel set $A$, $\omega(A)$ is a Poisson random variable with mean equal to the Lebesgue measure of $A$, and for disjoint bounded Borel sets $A_1,\ldots,A_m$, the random variables $\omega(A_1),\ldots,\omega(A_m)$ are independent. Often, we treat the point configuration $\omega$ as a locally bounded Borel measure with a unit atom at each point of the configuration. For background on Poisson point fields, also called Poisson processes, we refer to [@Daley:MR1950431]. There is a natural family of time-shift operators $(\theta^t)_{t\in{{\mathbb R}}}$ on the Poisson configurations: the configuration $\theta^t\omega$ is obtained from $\omega$ by shifting each point $(x,s)\in\omega$ to $(x,s-t)$. The space of velocity potentials that we will consider will be ${{\mathbb H}}$, the space of all locally Lipschitz functions $W:{{\mathbb R}}\to{{\mathbb R}}$ satisfying $$\begin{aligned} \liminf_{x\to+\infty}\frac{W(x)}{x}&>-\infty,\\ \limsup_{x\to-\infty}\frac{W(x)}{x}&<+\infty.\\\end{aligned}$$ Although it is possible to work with weaker conditions, some restrictions on the growth rate of $W(x)$ as $x\to\pm\infty$ are necessary to control velocities of particles coming from $\pm\infty$. Let us define random Hamilton–Jacobi–Hopf–Lax–Oleinik (HJHLO) dynamics on ${{\mathbb H}}$. For a function $W\in{{\mathbb H}}$, a Poissonian configuration $\omega$, and an absolutely continuous trajectory (path) $\gamma$ defined on $[s,t]$, we introduce the action $$\label{eq:action} A_\omega^{s,t}(W,\gamma)=W(\gamma(s))+S^{s,t}(\gamma)-\omega^{s,t}(\gamma),$$ where $$S^{s,t}(\gamma)=\frac{1}{2}\int_s^t\dot\gamma^2(r)dr$$ is the kinetic action, and $\omega^{s,t}(\gamma)=\omega(\{(\gamma(r),r):\ r\in[s,t)\})$ denotes the number of configuration points that $\gamma$ passes through. The last term in  is responsible for the interaction with the external forcing potential corresponding to the realization of the Poissonian field. We now consider the following minimization problem: $$\begin{aligned} \label{eq:optimization_problem} A_\omega^{s,t}(W,\gamma)&\to \inf,\\ \gamma(t)&=x.\notag\end{aligned}$$ Notice that the optimal trajectories are given by straight lines for any time interval on which the trajectory stays away from the configuration points. Since Poissonian configurations are locally finite, it is sufficient to take the minimum over broken lines with vertices at configuration points. \[lem:full\_measure\_set\_where\_dynamics\_is\_defined\] There is a set $\Omega_1\in{{\mathcal F}}_0$ with ${{\mathbb P}}(\Omega_1)=1$ such that for all $t\in{{\mathbb R}}$ $\theta^t\Omega_1=\Omega_1$, and for any $\omega\in\Omega_1$, for any $W\in{{\mathbb H}}$, any $x\in{{\mathbb R}}$ and any $s,t$ with $s<t$ there is a path $\gamma^*$ that realizes the minimum in . The path $\gamma^*$ is a broken line with finitely many segments, all its vertices belong to $\omega$. The proofs of this and other statements of this section are given in Section \[sec:proofs\_of\_basic\_burgers\_facts\]. We denote the restrictions of ${{\mathcal F}}_0$ and ${{\mathbb P}}_0$ onto $\Omega_1$ by ${{\mathcal F}}_1$ and ${{\mathbb P}}_1$. Since this restriction ${{\mathbb P}}_1$ still defines a Poisson point field with the same intensity measure, from now on for convenience we remove from $\Omega_0$ the zero measure complement to $\Omega_1$ and work with the probability space $(\Omega_1,{{\mathcal F}}_1,{{\mathbb P}}_1)$. We denote the infimum (minimum) value in  by $\Phi^{s,t}_\omega W(x)$. The family of random nonlinear operators $(\Phi^{s,t}_\omega)_{s\le t}$ is the main object in this paper. Our main goal is to understand the asymptotics of $\Phi^{s,t}_\omega$ as $t-s\to\infty$. We will need several properties of the random nonlinear operator $\Phi^{s,t}_\omega$ defined on ${{\mathbb H}}$. We begin with a lemma that shows that $\Phi^{s,t}_\omega W$ can be understood as the potential of the velocity field given by the terminal velocities of minimizers and that it produces a tessellation of space-time into the domains of influence of configuration points. For a Borel set $B\subset {{\mathbb R}}^2$, we denote the restriction of $\omega$ to $B$ by $\omega\bigr|_B$. \[lem:evolution\_on\_Burgers\_potentials\] For any $\omega\in\Omega_1$, $W\in{{\mathbb H}}$, $s,t\in{{\mathbb R}}$ with $s<t$, the following holds true: 1. \[it:open-domains-of-influence\] For any $p\in\omega|_{{{\mathbb R}}\times[s,t)}$, the set $O_p$ of points $x\in{{\mathbb R}}$ such that $p$ is the last configuration point visited by a unique minimizer for problem , is open. Also, the set of points with a unique minimizer that does not pass through any configuration points is open. The union of these open sets is dense in ${{\mathbb R}}$. 2. \[it:two-ways-to-compute-velocity\] If $x_0$ belongs to one of these open sets, $\gamma(t)=x_0$, and $A^{s,t}_\omega(W,\gamma)=\Phi_\omega^{s,t}W(x_0)$, then $\Phi_\omega^{s,t}W(x)$ is differentiable at $x_0$ w.r.t. $x$, and $$\frac{d}{dx}\Phi_\omega^{s,t}W(x)\bigr|_{x=x_0}=\dot\gamma(t).$$ At a boundary point $x_0$ of any of the open sets introduced above, the right and left derivatives of $\Phi_\omega^{s,t}W(x)$ w.r.t. $x$ are well defined. They are equal to the slope of, respectively, the leftmost and rightmost minimizers realizing $\Phi_\omega^{s,t}W(x_0)$. 3. \[it:piecewise\_linearity\] For any $p\in\omega|_{{{\mathbb R}}\times[s,t)}$, the function $x\mapsto \frac{d}{dx}\Phi_\omega^{s,t}W(x)$ is linear in $O_p$. 4. \[it:continuity\_of\_HJ\_solution\]The function $x\mapsto\Phi_\omega^{s,t}W(x)$ is locally Lipschitz. The following statement is the cocycle property for the operator family $(\Phi^{s,t}_\omega)$. It is a direct consequence of Bellman’s principle of dynamic programming. \[lem:cocycle\] If $\omega\in\Omega_1$, then for any $W\in{{\mathbb H}}$, any $s,r,t$ satisfying $s<r<t$, $\Phi_\omega^{r,t}\Phi_\omega^{s,r}W$ is well-defined and equals $\Phi_\omega^{s,t}W$. If $\gamma$ is an optimal path realizing $\Phi_\omega^{s,t}W(x)$, then the restrictions of $\gamma$ on $[s,r]$ and $[r,t]$ are optimal paths realizing $\Phi_\omega^{s,r}W(\gamma(r))$ and $\Phi_\omega^{r,t}(\Phi_\omega^{s,r}W)(x)$. Introducing $\Phi^t_{\omega}=\Phi^{0,t}_{\omega}$ we can rewrite the cocycle property as $$\Phi^{0,t+s}_{\omega}W=\Phi^{s}_{\theta^t\omega} \Phi^{t}_\omega W,\quad s,t> 0,\quad\omega\in\Omega_1.$$ Note that potentials are naturally defined up to an additive constant. It is thus convenient to work with $\hat {{\mathbb H}}$, the space of equivalence classes of potentials from ${{\mathbb H}}$. The cocycle $\Phi$ can be projected on $\hat{{\mathbb H}}$ in a natural way. We denote the resulting cocycle on $\hat{{\mathbb H}}$ by $\hat\Phi$. Let us now explain how the dynamics that we consider is connected to the classical Burgers equation. One way to describe this connection is to introduce a mollification of the Poisson integer-valued measure. We give the corresponding statement without a proof. Let us take smooth kernels $\phi,\psi:{{\mathbb R}}\to[0,\infty)$ with bounded support, satisfying $\int_{{{\mathbb R}}}\phi(t)dt=1$ and $\max_{x\in{{\mathbb R}}}\psi(x)=1$, and for each ${{\varepsilon }}>0$ consider the potential of shot-noise type: $$F_{{\varepsilon }}(x,t)=-\frac{1}{{{\varepsilon }}}\sum_{(y,s)\in\omega}\phi\left(\frac{t-s}{{{\varepsilon }}}\right)\psi\left(\frac{x-y}{\alpha({{\varepsilon }})} \right),$$ where $\alpha$ is any function satisfying $\lim_{{{\varepsilon }}\downarrow 0}\alpha({{\varepsilon }})=0$. With probability 1, for all $s,t,x\in{{\mathbb R}}$ and $W\in{{\mathbb H}}$, the entropy solution $U_{{\varepsilon }}(x,t)$ of the Cauchy problem for the Hamilton–Jacobi equation with smooth forcing potential $F_{{\varepsilon }}(\cdot,\cdot)$ converges, as ${{\varepsilon }}\to 0$, to $U(x,t)=\Phi^{s,t}_\omega W(x)$. The next statement shows that away from the Poissonian points the system we consider behaves as unforced Burgers dynamics. \[lem:Hamilton-Jacobi\] For all $\omega\in\Omega_1$, $s\in{{\mathbb R}}$, $W\in{{\mathbb H}}$, the function $U(x,t)=\Phi^{s,t}_\omega W(x)$ is an entropy solution of the Hamilton–Jacobi equation $$\label{eq:Hamilton-Jacobi} \partial_t U+\frac{(\partial_x U)^2}{2}=0.$$ in $((s,\infty)\times{{\mathbb R}})\setminus\omega$. Equivalently, $u(x,t)=\partial_x U(x,t)$ is an entropy solution of the Burgers equation  with $f\equiv 0$ in $((s,\infty)\times{{\mathbb R}})\setminus\omega$. Of course, for each $t$, $u(x,t)$ is a piecewise continuous function of $x$, and at each of the countably many discontinuity points it makes a negative jump. Since a velocity field determines its potential uniquely up to an additive constant we can also introduce dynamics on velocity fields. We can introduce the space ${{\mathbb H}}'$ of functions $w$ (actually, classes of equivalence of functions since we do not distinguish two functions coinciding almost everywhere) such that for some function $W\in{{\mathbb H}}$ and almost every $x$, $w(x)=W'(x)$. This allows us to introduce the Burgers dynamics. We will say that $ w_2=\Psi^{s,t}_\omega w_1$ if $w_1=W'_1$, $w_2=W'_2$, and $W_2=\Phi^{s,t}_\omega W_1$ for some $W_1,W_2\in{{\mathbb H}}$. Often in the context of the Burgers dynamics, the functions in ${{\mathbb H}}'$ will have negative jump discontinuities, so called shocks. Although it is not essential, we can require the functions in ${{\mathbb H}}'$ to be right-continuous. Let us denote ${{\mathbb H}}(v_-,v_+)=\{W\in{{\mathbb H}}:\ \lim_{x\to \pm\infty} (W(x)/x)=v_\pm \}$. The spaces $\hat{{\mathbb H}}(v_-,v_+)$ are defined as classes of potentials in ${{\mathbb H}}(v_-,v_+)$ coinciding up to an additive constant. The following result shows that these spaces are invariant under HJHLO dynamics. Along with Lemma \[lem:cocycle\] it allows to treat the dynamics as a random dynamical system with perfect cocycle property (see, e.g., [@Arnold:MR1723992 Section 1.1]). \[lem:invariant\_spaces\]There is a set $\Omega\in{{\mathcal F}}_1$ with the following properties: ${{\mathbb P}}_1(\Omega)=1$; for any $t\in{{\mathbb R}}$, $\theta^t\Omega=\Omega$;if $\omega\in\Omega$ then for any $s,t$ with $s<t$, 1. If $W\in{{\mathbb H}}$, then $\Phi^{s,t}_\omega W\in{{\mathbb H}}$. 2. If $W\in{{\mathbb H}}(v_-,v_+)$ for some $v_-,v_+$, then $\Phi^{s,t}_\omega W\in{{\mathbb H}}(v_-,v_+)$. We denote the restrictions of ${{\mathcal F}}_1$ and ${{\mathbb P}}_1$ onto $\Omega$ by ${{\mathcal F}}$ and ${{\mathbb P}}$. Since this restriction ${{\mathbb P}}$ still defines a Poisson point field with the same intensity measure, from now on for convenience we remove from $\Omega_1$ the zero measure complement to $\Omega$ and work with probability space $(\Omega,{{\mathcal F}},{{\mathbb P}})$. In nonrandom setting the family of operators $(\Phi^{s,t})$ constructed via a variational problem of type  is called a HJHLO evolution semi-group, see [@Villani Definition 7.33], [@Fathi], but in our setting it would be more precise to call it a HJHLO cocycle. Having defined the dynamics, we now turn to the main results. Main results {#sec:main_results} ============ We say that $u(t,x)=u_\omega(t,x)$ is a global solution for the cocycle $\Psi$ if there is a set $\Omega'$ with ${{\mathbb P}}(\Omega')=1$ such that for all $\omega\in\Omega'$, all $s$ and $t$ with $s<t$, we have $\Psi^{s,t}_\omega u_\omega(s,\cdot)= u_\omega(t,\cdot)$. We can also introduce the global solution as a skew-invariant function: $u_\omega(x)$ is called skew-invariant if there is a set $\Omega'$ with ${{\mathbb P}}(\Omega')=1$ such that for any $t\in{{\mathbb R}}$, $\theta^t\Omega'=\Omega'$, and for any $t>0$ and $\omega\in\Omega'$, $\Psi^t_\omega u_\omega =u_{\theta^t\omega}$. If $u_\omega(x)$ is a skew-invariant function, then $u_\omega(x,t)=u_{\theta^t\omega}(x)$ is a global solution. One can naturally view the potentials of $u_\omega(x)$ and $u_\omega(s,x)$ as a skew-invariant function and global solution for the cocycle $\hat\Phi$. Our first result is the description of global solutions. \[thm:global\_solutions\] For every $v\in{{\mathbb R}}$ there is a unique (up to zero-measure modifications) skew-invariant function $u_v:\Omega\to{{\mathbb H}}'$ such that for almost every $\omega\in\Omega$, the potential $U_{v,\omega}$ defined by $U_{v,\omega}(x)=\int^x u_{v,\omega}(y)dy$ belongs to $\hat {{\mathbb H}}(v,v)$. The potential $U_{v,\omega}$ is a unique skew-invariant potential in $\hat{{\mathbb H}}(v,v)$. The skew-invariant functions $U_{v,\omega}$ and $u_{v,\omega}$ are measurable w.r.t. ${{\mathcal F}}|_{{{\mathbb R}}\times(-\infty,0]}$, i.e., they depend only on the history of the forcing. With probability 1, the realizations of $(u_{v,\omega}(y))_{y\in{{\mathbb R}}}$ are piecewise linear with negative jumps between linear pieces. The spatial random process $(u_{v,\omega}(y))_{y\in{{\mathbb R}}}$ is stationary and mixing. \[rem:stationary-solution\]Notice that this theorem can be interpreted as a 1F1S Principle: for any velocity value $v$, the solution at time $0$ with mean velocity $v$ is uniquely determined by the history of the forcing: $u_{v,\omega}\stackrel{\rm a.s.}{=}\chi_v(\omega|_{{{\mathbb R}}\times (-\infty,0]})$ for some deterministic functional $\chi_v$ of the point configurations on half-plane ${{\mathbb R}}\times (-\infty,0]$ (we actually describe $\chi_v$ in the proof). Since the forcing is stationary in time, we obtain that $u_{v,\theta^t\omega}$ is a stationary process in $t$, and the distribution of $u_{v,\omega}$ is an invariant distribution for the corresponding Markov semi-group, concentrated on ${{\mathbb H}}'(v,v)$. The next result shows that each of the global solutions constructed in Theorem \[thm:global\_solutions\] plays the role of a one-point pullback attractor. To describe the domains of attraction we will make assumptions on initial potentials $W\in{{\mathbb H}}$. Namely, we will assume that there is $v\in{{\mathbb R}}$ such that $W$ and $v$ satisfy one of the following sets of conditions: $$\begin{aligned} v&=0,\notag\\ \liminf_{x\to+\infty} \frac{W(x)}{x}&\ge 0, \label{eq:no_flux_from_infinity}\\ \limsup_{x\to-\infty} \frac{W(x)}{x}&\le 0,\notag\end{aligned}$$ or $$\begin{aligned} v&> 0,\notag \\ \lim_{x\to-\infty} \frac{W(x)}{x}&= v,\label{eq:flux_from_the_left_wins}\\ \liminf_{x\to+\infty} \frac{W(x)}{x}&> -v,\notag\end{aligned}$$ or $$\begin{aligned} v&< 0,\notag\\ \lim_{x\to+\infty} \frac{W(x)}{x}&= v,\label{eq:flux_from_the_right_wins}\\ \limsup_{x\to-\infty} \frac{W(x)}{x}&< -v.\notag\end{aligned}$$ Condition  means that there is no macroscopic flux of particles from infinity toward the origin for the initial velocity profile $W'$. In particular, any $W\in{{\mathbb H}}(0,0)$ or any $W\in{{\mathbb H}}(v_-,v_+)$ with $v_-\le 0$ and $v_+\ge 0$ satisfies . It is natural to call the arising phenomenon a rarefaction fan. We will see that in this case the long-term behavior is described by the global solution $u_0$ with mean velocity $v=0$. Condition  means that the initial velocity profile $W'$ creates the influx of particles from $-\infty$ with effective velocity $v\ge 0$, and the influence of the particles at $+\infty$ is not as strong. In particular, any $W\in{{\mathbb H}}(v,v_+)$ with $v\ge 0$ and $v_+> -v$ (e.g., $v_+=v$) satisfies . We will see that in this case the long-term behavior is described by the global solution $u_v$. Condition  describes a situation symmetric to , where in the long run the system is dominated by the flux of particles from $+\infty$. The following precise statement supplements Theorem \[thm:global\_solutions\] and describes the basins of attraction of the global solutions $u_v$ in terms of conditions –. \[thm:pullback\_attraction\] There is a set $\Omega''\in{{\mathcal F}}$ with ${{\mathbb P}}(\Omega'')=1$ such that if $\omega\in\Omega''$, $W\in {{\mathbb H}}$, and one of conditions ,, holds: then $w=W'$ belongs to the domain of pullback attraction of $u_v$ in the following sense: for any $t\in{{\mathbb R}}$ and any $R>0$ there is $s_0=s_0(\omega)<t$ such that for all $s<s_0$ $$\Psi^{s,t}_\omega w(x) = u_{v,\omega}(x,t),\quad x\in[-R,R].$$ In particular, $${{\mathbb P}}\left\{\Psi^{s,t}_\omega w\big|_{[-R,R]}=u_{v,\omega}(\cdot,t)\big|_{[-R,R]}\right\}\to 1,\quad s\to-\infty.$$ The last statement of the theorem implies that for every $v\in{{\mathbb R}}$, the invariant measure on ${{\mathbb H}}'(v,v)$ described in Remark \[rem:stationary-solution\] is unique and for any initial condition $w=W'\in{{\mathbb H}}'$ satisfying one of conditions ,, and , the distribution of the random velocity profile at time $t$ converges to the unique stationary distribution on ${{\mathbb H}}'(v,v)$ as $t\to\infty$. However, our approach does not produce any convergence rate estimates. Using space-time Galilean transformations, it is easy to obtain a version of Theorem \[thm:pullback\_attraction\] for attraction in a coordinate frame moving with constant velocity, but we omit it for brevity. The proofs of Theorems \[thm:global\_solutions\] and \[thm:pullback\_attraction\] are given in Sections \[sec:global\_solutions\] and \[sec:attractor\], but most of the preparatory work is carried out in Sections \[sec:subadd\], \[sec:concentration\], and \[sec:one-sided\_minimizers\]. The long-term behavior of the cocycles $\Phi$ and $\Psi$ defined through the optimization problem  depends on the asymptotic behavior of the action minimizers over long time intervals. The natural notion that plays a crucial role in this paper is the notion of backward one-sided infinite minimizers or geodesics. A curve $\gamma:(-\infty,t]\to{{\mathbb R}}$ with $\gamma(t)=x$ is called a backward minimizer if its restriction onto any time interval $[s,t]$ provides the minimum to the action $A_\omega^{s,t}(W,\cdot)$ defined in  among paths connecting $\gamma(s)$ to $x$. It can be shown (see Lemma \[lem:geodesic direction\]) that any backward minimizer $\gamma$ has an asymptotic slope $v=\lim_{t\to -\infty}(\gamma(t)/t)$. On the other hand, for every space-time point $(x,t)$ and every $v\in {{\mathbb R}}$ there is a backward minimizer with slope $v$ and endpoint $(x,t)$. The following theorem describes the most important properties of backward minimizers associated with the Poisson point field. For every $v\in{{\mathbb R}}$ there is a set of full measure $\Omega'$ such that for all $\omega\in\Omega'$ and any $(x,t)\in\omega$ there is a unique backward minimizer with asymptotic slope $v$. For any $(x_1,t_1),(x_2,t_2)\in\omega$ there is a time $s\le \min\{t_1,t_2\}$ such that both minimizers coincide before $s$, i.e., $\gamma_1(r)=\gamma_2(r)$ for $r\le s$. The proof of this core statement of this paper is spread over Sections \[sec:subadd\] through \[sec:one-sided\_minimizers\]. In Section \[sec:subadd\] we apply the sub-additive ergodic theorem to derive the linear growth of action. In Section \[sec:concentration\] we prove quantitative estimates on deviations from the linear growth. We use these results in Section \[sec:one-sided\_minimizers\] to analyze deviations of optimal paths from straight lines and deduce the existence of infinite one-sided optimal paths and their properties. Optimal action asymptotics and the shape function {#sec:subadd} ================================================= In this section we study the asymptotic behavior of the optimal action between space-time points $(x,s)$ and $(y,t)$ denoted by $$\begin{aligned} \label{eq:optimal_action_between_two_points} A^{s,t}(x,y)=A^{s,t}_\omega(x,y)&=\min_{\gamma:\gamma(s)=x,\gamma(t)=y}(A^{s,t}_\omega(0,\gamma)) \\ &= \min_{\gamma:\gamma(s)=x,\gamma(t)=y}\left(S^{s,t}(\gamma) - \omega^{s,t}(\gamma)\right) \notag\end{aligned}$$ (the minimum is taken over all absolutely continuous paths $\gamma$ or, equivalently, over all piecewise linear paths with vertices at configuration points). Although to construct stationary solutions for the Burgers equation, we will need the asymptotic behavior as $s\to{-\infty}$, it is more convenient and equally useful (due to the obvious symmetry in the variational problem) to formulate most results for the limiting behavior as $t\to\infty$, and so we will here and in the next two sections. We begin with some simple observations on Galilean shear transformations of the point field. \[lem:shear\] Let $a,v\in{{\mathbb R}}$ and let $L$ be a transformation of space-time defined by $L(x,s)=(x+a+vs,s)$. 1. Suppose that $\gamma$ is a path defined on a time interval $[t_0,t_1]$ and let $\bar \gamma$ be defined by $(\bar\gamma(s),s)=L(\gamma(s),s)$. Then $$S^{t_0,t_1}(\bar\gamma)=S^{t_0,t_1}(\gamma)+(\gamma(t_1) -\gamma(t_0))v+\frac{(t_1-t_0)v^2}{2}.$$ 2. Let $L(\omega)$ be the point configuration obtained from $\omega\in\Omega$ by applying $L$ pointwise. Then $L(\omega)$ is also a Poisson process with Lebesgue intensity measure. 3. Let $\omega\in\Omega$. For any time interval $[t_0,t_1]$ and any points $x_0,x_1,\bar x_0, \bar x_1$ satisfying $L(x_0,t_0)=(\bar x_0, t_0)$ and $L(x_1,t_1)=(\bar x_1, t_1)$, $$A_{L(\omega)}^{t_0,t_1}(\bar x_0, \bar x_1)= A_{\omega}^{t_0,t_1}(x_0, x_1)+(x_1-x_0)v+\frac{(t_1-t_0)v^2}{2},$$ and $L$ maps minimizers realizing $A_{\omega}^{t_0,t_1}(x_0, x_1)$ onto minimizers realizing $A_{L(\omega)}^{t_0,t_1}(\bar x_0, \bar x_1)$. 4. For any points $x_0,x_1,\bar x_0,\bar x_1$ and any time interval $[t_0,t_1]$, $$A^{t_0,t_1}(\bar x_0, \bar x_1)\stackrel{distr}{=} A^{t_0,t_1}(x_0, x_1)+(x_1-x_0)v+\frac{(t_1-t_0)v^2}{2},$$ where $$v=\frac{(\bar x_1-x_1)-(\bar x_0-x_0)}{t_1-t_0}.$$ [[[Proof: ]{}]{}]{}The first part of the Lemma is a simple computation: $$\begin{aligned} S^{t_0,t_1}(\bar\gamma)& =\frac{1}{2}\int_{t_0}^{t_1}(\dot\gamma(s)+v)^2ds\\ &=\frac{1}{2}\int_{t_0}^{t_1}\dot\gamma^2(s)ds+\int_{t_0}^{t_1}\dot\gamma(s)vds+\frac{1}{2}\int_{t_0}^{t_1}v^2ds.\end{aligned}$$ The second part holds since $L$ preserves the Lebesgue measure. The third part follows from the first one since the images of paths transformed by $L$ are also paths passing through the $L$-images of configuration points. The last part is a consequence of the previous two parts, since the appropriate Galilean transformation sending $(x_0,t_0)$ to $(\bar x_0, t_0)$ and $(x_1,t_1)$ to $(\bar x_1, t_1)$ preserves the Lebesgue measure and the distribution of the Poisson process. [[[$\Box$ ]{}]{}]{} The next useful property is the sub-additivity of action along any direction: for any velocity $v\in {{\mathbb R}}$, and any $t,s\geq 0$, we have $$A^{0,t+s}(0,v(t+s))\leq A^{0,t}(0,vt) + A^{t,t+s}(vt,v(t+s)).$$ This means that we can apply Kingman’s sub-additive ergodic theorem to the function $t\mapsto A^{0,t}(0,vt)$ if we can show that $-{{\mathbb E}}A^{0,t}(0,vt)$ grows at most linearly in $t$. We claim this linear bound in the following proposition: \[lem:linbound\] Let $v\in {{\mathbb R}}$. There exists a constant $C=C(v)>0$ such that for all $t\geq 0$ $${{\mathbb E}}|A^{0,t}(0,vt)| \leq Ct.$$ [[[Proof: ]{}]{}]{}Lemma \[lem:shear\] implies that it is enough to prove this for $v=0$. So in this proof we work with $A^t=A^t(0,0)$. Let $\gamma:[0,t]\to {{\mathbb R}}$ be a path realizing $A^t$. We have $\gamma(0)=\gamma(t)=0$. Let us split up ${{\mathbb R}}^2$ into unit blocks $B_{i,j}=[i,i+1)\times [j,j+1)$, for $i,j\in {{\mathbb Z}}$. We define ${{\mathcal A}}$ as the union of all indices $(i,j)$ such that $\gamma$ passes through $B_{i,j}$. The set ${{\mathcal A}}$ is a lattice animal, i.e., it is a connected set that contains the origin $(0,0)\in{{\mathbb Z}}^2$ (see, e.g.,[@Gandolfi-Kesten:MR1258174]). Let us introduce the event $E_{n,t} = \{ \#{{\mathcal A}}= n\}$. \[lem:latani\] There are constants $C_1, C_2, R, t_0>0$ such that if $t\geq t_0$ and $n\geq R t$, then $${{\mathbb P}}(E_{n,t}) \leq C_1\exp(-C_2n^2/t).$$ [[[Proof: ]{}]{}]{}We define $X_{i,j}=\omega(B_{i,j})$, the number of Poisson points in $B_{i,j}$. Define the weight of the animal ${{\mathcal A}}$ as $$N_{{\mathcal A}}= \sum_{\nu\in {{\mathcal A}}} X_\nu.$$ Clearly, the number of Poisson points picked up by $\gamma$ between $0$ and $t$, is upper bounded by $N_{{\mathcal A}}$. Define $k_j=\#\{i\in{{\mathbb Z}}\ :\ (i,j)\in{{\mathcal A}}\}$, the number of blocks hit on the $j^{\rm th}$ row. These blocks will form a connected row of length $k_j$, and the kinetic action accumulated between $j$ and $(j+1)\wedge t$ can therefore be bounded by $$\frac12\int_j^{(j+1)\wedge t} \dot{\gamma}^2(s)\,ds \geq \frac12\left(k_j-2\right)_+^2.$$ Here, $a_+=\max(0,a)$. This leads to the following bound on the action: $$A^t \geq \frac12\sum_{0\le j<t}\left(k_j-2\right)_+^2 - N_{{\mathcal A}}.$$ On $E_{n,t}$ we have $\sum_{0\le j<t} k_j = n$. Since $a\mapsto (a-2)_+^2$ is convex, we can use Jensen’s inequality to see that $$\frac12\sum_{0\le j< t}\left(k_j-2\right)_+^2 \geq \frac12 {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2.$$ Therefore, $$A^t \geq \frac12 {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2 - N_{{\mathcal A}}. \label{eq:upper_bound_on_animal_weight}$$ We also know that $A^t\leq 0$ since we can use the identical zero path on $[0,t]$. Hence, on $E_{n,t}$ we have $$N_{{\mathcal A}}\geq \frac12 {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2.$$ Furthermore, if $N_n$ is the weight of the greedy animal of size $n$ (i.e., the animal of size $n$ with greatest weight), then $N_n\ge N_{{\mathcal A}}$, and $$E_{n,t} \subset \left\{N_n \geq \frac12 {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2\right\}.$$ Let us recall that the reasoning in [@Cox-Gandolfi-Griffin-Kesten:MR1241039] after equation (2.12) implies that, due to standard large deviation estimates and the exponential growth of the number of lattice animals as a function of size $n$, there are constants $K_1, K_2, y_0>0$, such that if $$\label{eq:cond_on_y} y\ge y_0,$$ then $$\label{eq:latanibound} {{\mathbb P}}\{N_n \geq yn\} \leq K_1\exp(-K_2ny).$$ We now need to make sure that  holds for $y= \frac{1}{2n} {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2$. If we require $n\geq \max(4, 8 y_0) {\lceil t\rceil}$, then $$\frac{1}{2n} {\lceil t\rceil}\left(\frac{n}{{\lceil t\rceil}} - 2\right)_+^2 = \frac{1}{2n} {\lceil t\rceil}\left(\frac{n-2{\lceil t\rceil}}{{\lceil t\rceil}}\right)_+^2 \ge \frac{1}{2n} {\lceil t\rceil}\left(\frac{n-\frac{n}{2}}{{\lceil t\rceil}}\right)^2 \geq \frac{1}{8} \frac{n}{{\lceil t\rceil}}\ge y_0,$$ and the lemma follows from . [[[$\Box$ ]{}]{}]{} We will choose the constant $R$ to be an integer, making it larger if needed. From  we already know that on $E_{n,t}$ we have $0\geq A^t \geq -N_n$. We wish to use this to estimate ${{\mathbb E}}|A^t|$, but we need an extension of . \[lem:ENn\] For any $k\geq 1$, there is $c_k>0$ such that for all $n\geq 1$, $${{\mathbb E}}N_n^k \leq c_kn^k.$$ [[[Proof: ]{}]{}]{}Clearly, $${{\mathbb E}}N_n^k = \sum_{i=0}^{\lfloor y_0n\rfloor} i^k{{\mathbb P}}\{N_n=i\} + \sum_{i=\lfloor y_0n\rfloor+1}^{\infty} i^k{{\mathbb P}}\{N_n=i\}.$$ We can bound the first term simply by $$\sum_{i=0}^{\lfloor y_0n\rfloor} i^k{{\mathbb P}}\{N_n=i\} \leq (y_0n)^k.$$ For the second term we can use : $$\sum_{i=\lfloor y_0n\rfloor+1}^{\infty} i^k{{\mathbb P}}\{N_n=i\}\leq \sum_{i=\lfloor y_0n\rfloor+1}^{\infty} K_1i^k\exp(-K_2i).$$ The right-hand side is bounded in $n$ and the proof is complete. [[[$\Box$ ]{}]{}]{} Lemma \[lem:linbound\] now follows from Lemmas \[lem:latani\] and \[lem:ENn\]: $$\begin{aligned} {{\mathbb E}}|A^t| & = & \sum_{n\le Rt} {{\mathbb E}}|A^t|{{\mathbf{1}}}_{E_{n,t}} + \sum_{n> Rt } {{\mathbb E}}|A^t|{{\mathbf{1}}}_{E_{n,t}}\\ & \leq & {{\mathbb E}}N_{[ Rt ]} + \sum_{n> Rt } {{\mathbb E}}N_n{{\mathbf{1}}}_{E_{n,t}}\\ & \leq & Rc_1t + \sum_{n>Rt} \sqrt{{{\mathbb E}}N_n^2}\sqrt{{{\mathbb P}}(E_{n,t})}\\ & \leq & Rc_1t + \sqrt{c_2}\sum_{n>Rt} \sqrt{C_1} n\exp(-C_2n^2/(2t))\\ & \leq & Ct,\end{aligned}$$ for $C$ big enough. [[[$\Box$ ]{}]{}]{} In fact, we can use the last calculation to obtain the following generalization of Lemma \[lem:linbound\] for higher moments of $A^t$: \[lem:moments\_of\_At\] Let $k\in{{\mathbb N}}$. Then there are constants $C(k),t_0(k)>0$ such that $${{\mathbb E}}(|A^t|^k)\le C(k) t^k,\quad t\ge t_0(k).$$ Now a standard application of the sub-additive ergodic theorem shows that there exists a *shape function* $\alpha(v)$ such that $$\frac{A^{0,t}(0,vt)}{t}\to\alpha(v),\quad \text{a.s.\ and in\ } L^1,\quad t\to\infty. \label{eq:def_of_shape_function}$$ Furthermore, $\alpha(0)<0$, since $A^t\leq 0$ and $\alpha(0)\leq {{\mathbb E}}(A^t)<0$. It turns out that the shape function $\alpha(v)$ is quadratic in $v$: \[lem:shape-function\] The shape function satisfies $$\alpha(v)=\alpha(0)+\frac{v^2}{2},\quad v\in{{\mathbb R}}.$$ [[[Proof: ]{}]{}]{}The Galilean shear map $(x,t)\mapsto(x+vt,t)$ transforms the paths connecting $(0,0)$ to $(0,t)$ into paths connecting $(0,0)$ to $(vt,t)$. Lemma \[lem:shear\] implies that under this map the optimal action over these paths is altered by a deterministic correction $v^2t/2$. Since $\alpha$ is a constant almost surely we obtain the statement of the lemma. [[[$\Box$ ]{}]{}]{} We know now from  that $A^{0,t}(0,vt)\sim\alpha(v) t$ as $t\to\infty$ with probability 1. However, this is not enough for our purposes since we need quantitative estimates on deviations of $A^{0,t}(0,vt)$ from $\alpha(v) t$. This is the material of the next section. Concentration inequality for optimal action {#sec:concentration} =========================================== The goal of this section is to prove a concentration inequality for $A^t(vt)=A^t(0,vt)=A^t_\omega(0,vt)=A^{0,t}_\omega(0,vt)$: \[thm:concentration\_around\_alphat\] There are positive constants $c_0,c_1,c_2,c_3,c_4$ such that for any $v\in{{\mathbb R}}$, all $t>c_0$, and all $u\in(c_3t^{1/2}\ln^2 t, c_4t^{3/2}\ln t]$, $${{\mathbb P}}\{|A^t(0,vt)-\alpha(v) t |>u\}\le c_1\exp\left\{-c_2\frac{u}{t^{1/2}\ln t}\right\}.$$ Due to the invariance under shear transformations (Lemmas \[lem:shear\] and \[lem:shape-function\]), it is sufficient to prove this theorem for $v=0$. We will first derive a similar inequality with $\alpha(0)t$ replaced by ${{\mathbb E}}A^t$, and then we will have to estimate the corresponding approximation error. We recall that $A^t\leq 0$. \[lem:concentration\_around\_mean\] There are positive constants $b_0,b_1,b_2,b_3$ such that for all $t>b_0$ and all $u\in(0, b_3t^{3/2}\ln t]$, $${{\mathbb P}}\{|A^t-{{\mathbb E}}A^t|>u\}\le b_1\exp\left\{-b_2\frac{u}{t^{1/2}\ln t}\right\}.$$ The method of proof is derived from that for the generalized Hammersley’s process in [@CaPi], but we have to take into account that the optimal paths are allowed to travel arbitrarily far within any bounded time interval in search for areas rich with configuration points. However, the situation where they decline too far from the kinetically most efficient path is not typical. In the remaining part of this section we will often use the following lemma showing that with high probability the minimizer $\gamma$ connecting $(0,0)$ to $(0,t)$ stays within distance $Rt$ from the origin, where $R$ was introduced in Lemma \[lem:latani\]. \[lem:path\_in\_wide\_rectangle\_whp\] There is a constant $C_3$ such that if $t\ge t_0$ and $u\ge Rt$ then $${{\mathbb P}}\left\{\max_{s\in[0,t]}|\gamma(s)|>u\right\}\le C_3\exp(-C_2 u^2/t),$$ where constants $C_2,R,t_0$ were introduced in Lemma \[lem:latani\]. [[[Proof: ]{}]{}]{}If $\max\{|\gamma(s)|:\ s\in[0,t] \}>u$, then the size of the lattice animal ${{\mathcal A}}$ traced by $\gamma$ is at least $u$. Lemma \[lem:latani\] implies $${{\mathbb P}}\left\{\max_{s\in[0,t]}|\gamma(s)|>u\right\}\le\sum_{n\ge u} C_1\exp(-C_2n^2/t)\le C_3\exp(-C_2 u^2/t)$$ for a constant $C_3$, since the first term of the series is bounded by $C_1\exp(-C_2u^2/t)$ and the ratio of two consecutive terms is bounded by $\exp(-C_2 R)$. [[[$\Box$ ]{}]{}]{} Having Lemma \[lem:path\_in\_wide\_rectangle\_whp\] in mind, we define $\tilde A^t$ to be the optimal action over all paths connecting $(0,0)$ to $(0,t)$ and staying within $[-Rt,Rt]$. \[lem:probability\_that\_Kesten\_action\_worse\] Let constants $t_0,R,C_2,C_3$ be defined in Lemmas \[lem:latani\] and \[lem:path\_in\_wide\_rectangle\_whp\]. For any $t>t_0$, $${{\mathbb P}}\{A^t\ne \tilde A^t\}\le C_3\exp(-R^2C_2 t).$$ [[[Proof: ]{}]{}]{}It is sufficient to notice that $${{\mathbb P}}\{A^t\ne \tilde A^t\}\le{{\mathbb P}}\left\{\max_{s\in[0,t]}|\gamma(s)|>Rt\right\}$$ and apply Lemma \[lem:path\_in\_wide\_rectangle\_whp\]. [[[$\Box$ ]{}]{}]{} \[lem:difference\_of\_rectangle\_expectation\_from\_true\] There is a constant $D_1$ such that for all $t>t_0$, $$0\le {{\mathbb E}}\tilde A^t-{{\mathbb E}}A^t\le -{{\mathbb E}}A^t{{\mathbf{1}}}_{\left\{\sup_{s\in[0,t]}|\gamma(s)|>Rt\right\}}\le D_1.$$ [[[Proof: ]{}]{}]{}The first two inequalities are obvious, since we have that $0\geq \tilde A^t\geq A^t$. For the last one, we have $$\begin{aligned} -{{\mathbb E}}A^t{{\mathbf{1}}}_{ \left\{\sup_{s\in[0,t]}|\gamma(s)|>Rt\right\}} &\le \sum_{n>Rt} {{\mathbb E}}( N_n{{\mathbf{1}}}_{E_{n,t}}) \\&\le \sum_{n>Rt} \sqrt{{{\mathbb E}}N^2_n}\sqrt{{{\mathbb P}}(E_{n,t})} \\&\le \sum_{n>Rt} \sqrt{c_2}n\sqrt{C_1}\exp(-C_2n^2/(2t)),\end{aligned}$$ where we used Lemmas \[lem:latani\] and \[lem:ENn\]. The statement follows since the last series is uniformly convergent for $t>t_0$. [[[$\Box$ ]{}]{}]{} To obtain a concentration inequality for $\tilde A$, we will apply the following lemma by Kesten [@Kesten:MR1221154]: Let $({{\mathcal F}}_k)_{0\le k\le N}$ be a filtration and let $(U_k)_{0\le k\le N}$ be a family of nonnegative random variables measurable with respect to ${{\mathcal F}}_N$. Let $(M_k)_{0\le k\le N}$ be a martingale with respect to $({{\mathcal F}}_k)_{0\le k\le N}$. Assume that for some constant $c>0$ the increments $\Delta_k=M_k-M_{k-1}$ satisfy $$|\Delta_k|<c,\quad k=1,\ldots,N,$$ and $${{\mathbb E}}(\Delta_k^2|\ {{\mathcal F}}_{k-1}) \le {{\mathbb E}}(U_k|\ {{\mathcal F}}_{k-1}).$$ Assume further that for some positive constants $c_1,c_2$ and some $x_0\ge e^2c^2$ we have $${{\mathbb P}}\left\{\sum_{k=1}^N U_k>x\right\}\le c_1\exp(-c_2x),\quad x\ge x_0.$$ Then $${{\mathbb P}}\{M_N-M_0\ge x\}\le c_3\left(1+c_1+\frac{c_1}{c_2 x_0}\right)\exp\left(-c_4\frac{x}{x_0^{1/2}+c_2^{-1/3}x^{1/3}}\right),\quad x>0,$$ where $c_3,c_4$ are universal positive constants that do not depend on $N,c,c_1,c_2,x_0$, nor on the distribution of $(M_k)_{0\le k\le N}$ and $(U_k)_{0\le k\le N}$. In particular, $${{\mathbb P}}\{M_N-M_0\ge x\}\le c_3\left(1+c_1+\frac{c_1}{c_2 x_0}\right)\exp\left(-c_4\frac{x}{2\sqrt{x_0}}\right),\quad x\le c_2x_0^{3/2}.$$ To use this lemma in our framework, we must introduce an appropriate martingale. For a given $t$ we consider the rectangle $Q(t)=[-Rt,Rt]\times [0,t]$ and partition it into $N=2Rt\cdot t=2Rt^2$ disjoint unit squares: $Q(t)=\bigcup_{k=1}^{N} B_k$. The order of enumeration is not important. Here we assume that $t\in{{\mathbb N}}$, but it is easy to adapt the reasoning to the case of non-integer $t$. We introduce a filtration $({{\mathcal F}}_k)_{0\le k\le N}$ in the following way. We set ${{\mathcal F}}_0=\{\emptyset,\Omega\}$ and $${{\mathcal F}}_k=\sigma\left(\omega\bigr|_{\bigcup_{j=1}^k B_j} \right),\quad k=1,\ldots,N.$$ We introduce a martingale $(M_k,{{\mathcal F}}_k)_{0\le k\le N}$ by $$M_k={{\mathbb E}}(\tilde A^t|{{\mathcal F}}_k),\quad 0\le k\le N.$$ We denote by $P_k$ the distribution of $\omega\bigr|_{B_k}$ on the sample space $\Omega_k$ of finite point configurations in $B_k$. For $\omega,\sigma\in\prod_{k=1}^{N}\Omega_k$ we write $$[\omega,\sigma]_k=(\omega_1,\ldots,\omega_k,\sigma_{k+1},\ldots,\sigma_{N})\in\prod_{k=1}^N\Omega_k.$$ Then $$\begin{aligned} \Delta_k(\omega_1,\ldots,\omega_k)&:=M_k-M_{k-1}\\ &= \int \tilde A^t_{[\omega,\sigma]_k}\prod_{j=k+1}^N dP_j(\sigma_j)-\int \tilde A^t_{[\omega,\sigma]_{k-1}}\prod_{j=k}^N dP_j(\sigma_j)\\ &=\int \left(\tilde A^t_{[\omega,\sigma]_k}- \tilde A^t_{[\omega,\sigma]_{k-1}}\right)\prod_{j=k}^N dP_j(\sigma_j).\end{aligned}$$ \[lm:estimating\_integrand\_in\_Kesten\_martingale\_difference\] Let $I_k$ denote the indicator that the minimizer connecting $(0,0)$ to $(0,t)$ and staying in $[-Rt,Rt]$ passes through a Poissonian point in $B_k$. Then $$|\tilde A^t_{[\omega,\sigma]_k}- \tilde A^t_{[\omega,\sigma]_{k-1}}| \le \max\{I_k([\omega,\sigma]_k),I_k([\omega,\sigma]_{k-1})\}\max\{\omega(B_k),\sigma(B_k)\}.$$ [[[Proof: ]{}]{}]{}Suppose we delete the points of $\omega$ in $B_k$. When we then consider the minimizer for $[\omega,\sigma]_k$, we decrease the number of Poissonian points contributing to the action by at most $\omega(B_k)$, and only decrease the kinetic action. Comparing the resulting path with the minimizer for $[\omega,\sigma]_{k-1}$, we obtain $$\tilde A^t_{[\omega,\sigma]_{k-1}}\le \tilde A^t_{[\omega,\sigma]_k}+\omega(B_k).$$ Similarly, we get $$\tilde A^t_{[\omega,\sigma]_{k}}\le \tilde A^t_{[\omega,\sigma]_{k-1}} + \sigma(B_k).$$ This shows that $$|\tilde A^t_{[\omega,\sigma]_k}- \tilde A^t_{[\omega,\sigma]_{k-1}}| \le \max\{\omega(B_k),\sigma(B_k)\}.$$ Now remark that if none of the two minimizers (for $[\omega,\sigma]_k$ and $[\omega,\sigma]_{k-1}$) passes through a Poissonian point inside $B_k$, then $\tilde A^t_{[\omega,\sigma]_k}$ and $\tilde A^t_{[\omega,\sigma]_{k-1}}$ coincide. This completes the proof. [[[$\Box$ ]{}]{}]{} The next step is to define a truncated Poissonian configuration $\bar \omega$ by erasing all Poissonian points of $\omega$ in each block $B_j$ with $\omega(B_j)> b\ln t$, where $b>0$ will be chosen later. The restrictions of $\bar \omega$ to $B_j$, $j=1,\ldots,N$ are jointly independent. Lemma \[lm:estimating\_integrand\_in\_Kesten\_martingale\_difference\] applies to truncated configurations as well and we obtain $$| \tilde A^t_{[\bar\omega,\bar\sigma]_k}- \tilde A^t_{[\bar\omega,\bar\sigma]_{k-1}}| \le b \ln t \max\{I_k([\bar\omega,\bar\sigma]_k),I_k([\bar\omega,\bar\sigma]_{k-1})\},$$ where $\bar\sigma$ is obtained from $\sigma$ in the same way as $\bar \omega$ from $\omega$. Therefore, $$|\Delta_k(\bar\omega_1,\ldots,\bar\omega_k)|\le b\ln t \int \max\{I_k([\bar\omega,\bar\sigma]_k),I_k([\bar\omega,\bar\sigma]_{k-1})\} \prod_{j=k}^N dP_j(\sigma_j)\le b\ln t.$$ We must now estimate the increments of the martingale predictable characteristic. This estimate is a straightforward analogue of Lemma 4.3 of [@CaPi]. Let $U_k=2(b\ln t)^2I_k$. Then, with probability $1$, $|U_k(\bar\omega)|\le 2(b\ln t)^2$ and $${{\mathbb E}}(\Delta_k^2(\bar\omega_1,\ldots,\bar\omega_k)|{{\mathcal F}}_{k-1})\le {{\mathbb E}}(U_k(\bar\omega)|{{\mathcal F}}_{k-1}).$$ [[[Proof: ]{}]{}]{}$$\begin{aligned} {{\mathbb E}}&(\Delta_k^2(\bar\omega_1,\ldots,\bar\omega_k)|{{\mathcal F}}_{k-1}) =\int\left(\int\left(\tilde A^t_{[\bar\omega,\bar\sigma]_k}- \tilde A^t_{[\bar\omega,\bar\sigma]_{k-1}}\right)\prod_{j=k}^N dP_j(\sigma_j)\right)^2 dP_k(\omega_k)\\ &\le\int\left(\int \max\{I_k([\bar\omega,\bar\sigma]_k),I_k([\bar\omega,\bar\sigma]_{k-1})\}\cdot b\ln t \prod_{j=k}^N dP_j(\sigma_j)\right)^2 dP_k(\omega_k)\\ &\le\int\int \max\{I_k([\bar\omega,\bar\sigma]_k),I_k([\bar\omega,\bar\sigma]_{k-1})\}\cdot (b\ln t)^2 \prod_{j=k}^N dP_j(\sigma_j) dP_k(\omega_k)\\ &\le\int\int (I_k([\bar\omega,\bar\sigma]_k)+I_k([\bar\omega,\bar\sigma]_{k-1}))\cdot (b\ln t)^2 \prod_{j=k}^N dP_j(\sigma_j) dP_k(\omega_k)\\ &= {{\mathbb E}}(U_k(\bar\omega)|{{\mathcal F}}_{k-1}).\end{aligned}$$ [[[$\Box$ ]{}]{}]{} We have $$\sum_{k=1}^N U_k(\bar\omega) = 2(b\ln t)^2\sum_{k=1}^{N} I_k(\bar\omega).$$ Since $$\sum_{k=1}^{N} I_k(\bar\omega)\le \# {{\mathcal A}}(\bar\omega),$$ we can write $${{\mathbb P}}\left\{\sum_{k=1}^N U_k(\bar\omega)>x\right\}\le {{\mathbb P}}\left\{\#{{\mathcal A}}(\bar\omega)>\frac{x}{2(b\ln t)^2}\right\}\le \sum_{n>x/(2(b\ln t)^2) } {{\mathbb P}}\{\bar\omega\in E_{n,t}\}.$$ It is easy to see that Lemma \[lem:latani\] applies to $\bar\omega$ as well as to $\omega$, since its proof depends only on the tail estimate for the number of configuration points in each block. We can conclude that $$\begin{aligned} {{\mathbb P}}\{\bar\omega\in E_{n,t}\} \leq C_1\exp(-C_2n^2/t),\quad n\ge Rt,\ t\ge t_0,\end{aligned}$$ where $C_1,C_2,R,t_0$ were introduced in Lemma \[lem:latani\]. Combining the last two inequalities and choosing $x_0=2Rt(b\ln t)^2$, we can write for $x>x_0$ $$\begin{aligned} {{\mathbb P}}\left\{\sum_{k=1}^N U_k(\bar\omega)>x\right\} &\le C_1\sum_{n>x/(2(b\ln t)^2)}\exp(-C_2 n^2/t) \\&\le C_4 \exp(-C_2 x^2/(4t(b\ln t)^4))\\ &\le C_4 \exp(-C_2 x x_0/(4t(b\ln t)^4)) \\&\le C_4 \exp(-C_5 (b\ln t)^{-2}x).\end{aligned}$$ The above estimates on $\Delta_k(\bar\omega)$ and $U_k(\bar\omega)$ allow to apply Kesten’s lemma with $c=2b\ln t$, $c_1=C_4$, $c_2=C_5(b\ln t)^{-2}$, $x_0=2Rt(b\ln t)^2$ and obtain the following statement: \[lem:outcome\_of\_Kestens\_lemma\]There are constants $C_6,C_7,C_8,t_0>0$ such that for $t>t_0$ and $x\le C_8bt^{3/2}\ln t$, $${{\mathbb P}}\{|\tilde A^t(\bar\omega)-{{\mathbb E}}\tilde A_t(\bar\omega)|>x\}\le C_6\exp\left(-C_7\frac{x}{b t^{1/2}\ln t}\right).$$ \[lem:deviation\_of\_truncation\_from\_original\] With probability 1, $$\tilde A^t(\omega)\le \tilde A^t(\bar \omega).$$ Also, we can choose $b$ and $t_0$ such that for all $t>t_0$ and $x>0$, $${{\mathbb P}}\{\tilde A^t(\bar \omega) -\tilde A^t(\omega)>x\}\le 2e^{-x}.$$ [[[Proof: ]{}]{}]{}The first statement of the lemma is obvious, and we have $$0\le \tilde A^t(\bar \omega) -\tilde A^t(\omega)\le \sum_{k=1}^N \omega(B_k){{\mathbf{1}}}_{\{\omega(B_k)>b\ln t\}}.$$ By Markov’s inequality and mutual independence of $\omega|_{B_k}$, $k=1,\ldots,N$, $${{\mathbb P}}\left\{\sum_{k=1}^N \omega(B_k){{\mathbf{1}}}_{\{\omega(B_k)>b \ln t\}}>x\right\}\le e^{-x}\left[{{\mathbb E}}e^{\omega(B_1){{\mathbf{1}}}_{\{\omega(B_1)>b\ln t\}}}\right]^N.$$ The lemma will follow from $$\label{eq:lim_factor_in_Markov_equals_1} \lim_{t\to\infty}\left[{{\mathbb E}}e^{\omega(B_1){{\mathbf{1}}}_{\{\omega(B_1)>b \ln t\}}}\right]^{2Rt^2}=1,$$ which is implied by $${{\mathbb E}}e^{\omega(B_1){{\mathbf{1}}}_{\{\omega(B_1)>b \ln t\}}}\le 1+\frac{{{\mathbb E}}e^{2\omega(B_1)}}{e^{b\ln t}}\le 1+\frac{{{\mathbb E}}e^{2\omega(B_1)}}{t^b},$$ if we choose $b>2$. [[[$\Box$ ]{}]{}]{} The only missing part in the proof of Lemma \[lem:concentration\_around\_mean\] is the following corollary of Lemma \[lem:deviation\_of\_truncation\_from\_original\]: \[lem:expectation\_of\_truncation-expectation\_over\_Kesten\_rectangle\] There is a constant $D_2$ such that for all $t>t_0$, $$0\le {{\mathbb E}}\tilde A^t(\bar\omega)-{{\mathbb E}}\tilde A^t(\omega) <D_2.$$ [[[Proof of Lemma \[lem:concentration\_around\_mean\]: ]{}]{}]{}Lemmas \[lem:difference\_of\_rectangle\_expectation\_from\_true\] and \[lem:expectation\_of\_truncation-expectation\_over\_Kesten\_rectangle\] imply that for $u>D_1+D_2$ $$\begin{aligned} {{\mathbb P}}\{|A^t(\omega)-{{\mathbb E}}A^t(\omega)|>u\}\le& {{\mathbb P}}\{|A^t(\omega)-\tilde A^t(\omega)|>(u-D_1-D_2)/3 \} \\&+{{\mathbb P}}\{|\tilde A^t(\omega)-\tilde A^t(\bar\omega)|>(u-D_1-D_2)/3\} \\&+{{\mathbb P}}\{|\tilde A^t(\bar\omega)-{{\mathbb E}}\tilde A^t(\bar\omega)|>(u-D_1-D_2)/3\}.\end{aligned}$$ The Lemma follows from the estimates of the three terms provided by Lemmas \[lem:probability\_that\_Kesten\_action\_worse\], \[lem:outcome\_of\_Kestens\_lemma\], and \[lem:deviation\_of\_truncation\_from\_original\].[[[$\Box$ ]{}]{}]{} The following lemma quantifies how the growth of $-{{\mathbb E}}A^t$ deviates from the linear one under argument doubling. We will use this lemma to find an estimate on ${{\mathbb E}}A^t-\alpha(0)t$ which makes it possible to fill the gap between Lemma \[lem:concentration\_around\_mean\] and Theorem \[thm:concentration\_around\_alphat\]. \[lem:doubling\] There is a number $b_0>0$ such that for any $t>t_0$, $$0\le 2{{\mathbb E}}A^t - {{\mathbb E}}A^{2t} \le b_0t^{1/2}\ln^2t.$$ [[[Proof: ]{}]{}]{}The first inequality follows from $A^{0,2t}(0,0)\le A^{0,t}(0,0)+A^{t,2t}(0,0)$. Let us prove the second one. Let $\gamma$ be the minimizer from $(0,0)$ to $(0,2t)$. Then $$A^{2t}\ge \min_{|x|\le 2Rt} A^{0,t}(0,x)+\min_{|x|\le 2Rt}A^{t,2t}(x,0)+A^{2t}{{\mathbf{1}}}_{\left\{\max_{s\in[0,2t]}|\gamma(s)|>2Rt\right\}}.$$ Therefore, by symmetry with respect to $t$ and Lemma \[lem:difference\_of\_rectangle\_expectation\_from\_true\], $${{\mathbb E}}A^{2t}\ge 2{{\mathbb E}}\min_{|x|\le 2Rt} A^{t}(0,x) - D_1. \label{eq:doubling+constant}$$ For $k\in I_t=\{-2Rt,\ldots,2Rt-2,2Rt-1\}$, we define a unit square $B_k=[k,k+1]\times[t-1,t]$. Let now $\gamma$ be the minimizer from $(0,0)$ to $(x,t)$, with $x\in[k,k+1]$ for some $k\in I_t$. Denote $t'=\sup\{s\le t: \gamma(s)\notin B_k\}$ and $x'=\gamma(t')$. If $x'< k+1$, then by reconnecting $(x',t')$ to $(k,t)$ we obtain $$A^t(k)\le A^{t'}(x')+1/2\le A^t(x)+\omega(B_k)+1/2.$$ If $x'= k+1$, then by reconnecting $(x',t')$ to $(k+1,t)$ we obtain $$A^t(k+1)\le A^{t'}(x')\le A^t(x)+\omega(B_k).$$ Therefore, $$A^t(x)\ge \min\{A^t(k),A^t(k+1)\}-\omega(B_k)-1/2,$$ and  implies $${{\mathbb E}}A^{2t}\ge 2 {{\mathbb E}}\min_{k\in I_t} A^t(k)-{{\mathbb E}}\max_{k\in I_t}\omega(B_k)-1/2-D_1.$$ The second term grows logarithmically in $t$. Hence, for some constant $c>0$, $${{\mathbb E}}A^{2t}\ge 2 {{\mathbb E}}\min_{k\in I_t} A^t(k)-c(\ln t+1).$$ Lemma \[lem:shear\] implies $$\min_{x} {{\mathbb E}}A^t(x)={{\mathbb E}}A^t(0).$$ Therefore, $$\begin{aligned} {{\mathbb E}}A^{2t}&\ge 2 {{\mathbb E}}\min_{k\in I_t} A^t(k)-c(\ln t+1) \notag \\&\ge 2\min_{k\in I_t}{{\mathbb E}}A^t(k) - 2{{\mathbb E}}X_t - c(\ln t+1) \notag \\&\ge 2{{\mathbb E}}A^t - 2{{\mathbb E}}X_t - c(\ln t+1), \label{eq:doubling1}\end{aligned}$$ where $$X_t=\max_{k\in I_t} \{({{\mathbb E}}A^t(k)-A^t(k))_+\}.$$ For a constant $r$ to be determined later, we introduce the event $$E=\{X_t\le r(\ln^2 t)\sqrt{t}\}.$$ Then $$X_t\le r(\ln^2 t)\sqrt{t}{{\mathbf{1}}}_{E}+ X_t{{\mathbf{1}}}_{E^c}.$$ Therefore, $${{\mathbb E}}X_t\le r(\ln^2 t)\sqrt{t}+\sqrt{{{\mathbb E}}(X_{t})^2{{\mathbb P}}(E^c)}. \label{eq:M_t}$$ Let us estimate the second term. According to Lemma \[lem:shear\], the random variables $A^t(k)-{{\mathbb E}}A^t(k)$, $k\in I_t$ have the same distribution, so replacing the maximum in the definition of $X_t^2$ with summation we obtain $${{\mathbb E}}X_{t}^2\le 4Rt {{\mathbb E}}(A^t-{{\mathbb E}}A^t)_+^2\le 4Rt {{\mathbb E}}(A^t)^2\le Ct^3,$$ for some $C>0$ and all $t$ exceeding some $t_0$, where we used Lemma \[lem:moments\_of\_At\] in the last inequality. Also, Lemma \[lem:concentration\_around\_mean\] shows that $$\begin{aligned} {{\mathbb P}}(E^c)&\le \sum_{k\in I_t}{{\mathbb P}}\left\{A^t(k)-{{\mathbb E}}A^t(k) > r(\ln^2 t)\sqrt{t}\right\}\notag \\ &\le 4R t b_1\exp\{-b_2r \ln t\}.\label{eq:probability_of_D_complement}\end{aligned}$$ We can now finish the proof by choosing $r$ to be large enough and combining estimates –. [[[$\Box$ ]{}]{}]{} With this lemma at hand we can now use the following statement ([@HoNe Lemma 4.2]): Suppose the functions $a: {{\mathbb R}}_+\to{{\mathbb R}}$ and $g:{{\mathbb R}}_+\to{{\mathbb R}}_+$ satisfy the following conditions: $a(t)/t\to \nu\in{{\mathbb R}}$ and $g(t)/t \to 0$ as $t\to\infty$, $a(2t)\ge 2a(t)-g(t)$ and $\psi\equiv \limsup_{t\to\infty} g(2t) /g(t)< 2$. Then, for any $c > 1/(2-\psi)$, and for all large $t$, $$a(t) \leq \nu t + cg(t).$$ Taking $a(t)={{\mathbb E}}A^t$, $\nu=\alpha(0)$, $g(t)=b_0t^{1/2}\ln^2t$, $\psi=\sqrt{2}$, $c=2$, we conclude that for $b'_0=2b_0$ and large $t$, $$0 \le {{\mathbb E}}A^t-\alpha(0) t\le b_0't^{1/2}\ln^2t,$$ and Theorem \[thm:concentration\_around\_alphat\] follows from this estimate, Lemma \[lem:concentration\_around\_mean\], and the shear invariance established in Lemma \[lem:shear\].[[[$\Box$ ]{}]{}]{} Existence and uniqueness of semi-infinite minimizers {#sec:one-sided_minimizers} ==================================================== In this section we will study the properties of geodesics: these are paths $\gamma:[t_1,t_2]\to {{\mathbb R}}$ such that for all $s\leq t \in [t_1,t_2]$, $\gamma|_{[s,t]}$ is the path that minimizes the action $A^{s,t}(\gamma(s),\gamma(t))$. We will closely follow ideas by Howard and Newman in [@HoNe] and by Wüthrich in [@Wu], adapting them to our specific situation, as was done in [@CaPi]. $\delta$-Straightness --------------------- The goal of this section is to estimate deviations of geodesics from straight lines. We will need the curvature of the shape function found in Section \[sec:subadd\]. Remember that $$\lim_{t\to \infty} \frac{A^{0,t}(0,vt)}{t} = \alpha(v) = \alpha(0) + \frac12 v^2.$$ Define $\alpha_0=\alpha(0)$. We will extend $\alpha$ to a function of ${{\mathbb R}}^2$: $$\alpha(p) := \alpha_0p_2 + \frac12 \frac{p_1^2}{p_2}=p_2\left(\alpha_0+\frac{1}{2}\left(\frac{p_1}{p_2}\right)^2\right)$$ (in this section we often denote the space and time coordinates of a space-time point $p\in{{\mathbb R}}^2$ by $p_1$ and $p_2$, respectively). This means that for all $p\in{{\mathbb R}}^2$ with $p_2>0$, $$\lim_{t\to \infty} \frac{A^{0,tp_2}(0,tp_1)}{t} \stackrel{\rm a.s.}= \alpha(p).$$ We need a convexity estimate of this function $\alpha$. Define for $p\in {{\mathbb R}}\times {{\mathbb R}}^+$ and for $L>0$ $${{\mathcal C}}(p,L) := \left\{ q\in{{\mathbb R}}^2:\ q_2\in (p_2,2p_2]\ \mbox{and}\ \left|\frac{q_2}{p_2}\,p_1 - q_1\right| \leq L\right\}.$$ So ${{\mathcal C}}(p,L)$ is a parallelogram of width $2L$ along $[p,2p]$ (for any two points $p,q$ on the plane, $[p,q]$ denotes the straight line segment connecting these two points). We need to consider the side-edges of this parallelogram: $$\partial_S{{\mathcal C}}(p,L) := \left\{ q\in{{\mathbb R}}^2:\ q_2\in (p_2,2p_2]\ \mbox{and}\ \left|\frac{q_2}{p_2}\,p_1 - q_1\right| = L\right\}.$$ The following lemma will play the role of Lemma 2.1 in [@Wu]. \[lem:convest\] For all $p\in {{\mathbb R}}\times {{\mathbb R}}^+$ and $\delta\in(0,1)$, and all $q\in \partial_S{{\mathcal C}}(p,p_2^{1-\delta})$, we have $$\label{eq:convest} \alpha(q-p) + \alpha(p) \geq \alpha(q) + \frac14\,p_2^{1-2\delta}.$$ [[[Proof: ]{}]{}]{}$$\begin{aligned} \alpha(q-p) + \alpha(p) & = & \alpha_0 (q_2-p_2) + \frac12\frac{(q_1-p_1)^2}{q_2-p_2} + \alpha_0 p_2 + \frac12\frac{p_1^2}{p_2}\\ & = & \alpha_0 q_2 + \frac12\frac{(q_1-p_1)^2}{q_2-p_2} + \frac12\frac{p_1^2}{p_2}\\ & = & \alpha(q) - \frac12\frac{q_1^2}{q_2} + \frac12\frac{(q_1-p_1)^2}{q_2-p_2} + \frac12\frac{p_1^2}{p_2}.\end{aligned}$$ This shows that $\alpha(q-p)+\alpha(p)-\alpha(q)$ is a quadratic function in $q_1$. The minimum of this function equals $0$, and is attained at $\tilde{q}_1$, the number defined by $$\frac{\tilde{q}_1-p_1}{q_2-p_2} = \frac{\tilde{q}_1}{q_2}.$$ This means that $(\tilde{q}_1,q_2)$ is a multiple of $p$, so by the definition of $\partial_S{{\mathcal C}}(p,p_2^{1-\delta})$, we know that $|q_1-\tilde{q}_1|=p_2^{1-\delta}$. It follows that $$\begin{aligned} \alpha(q-p) + \alpha(p) & = & \alpha(q) + \frac12\frac{p_2}{q_2(q_2-p_2)}\,(q_1-\tilde{q}_1)^2\\ & = & \alpha(q) + \frac12\frac{p_2^{3-2\delta}}{q_2(q_2-p_2)}.\end{aligned}$$ Now note that $q_2(q_2-p_2)\leq 2p_2^2$ to conclude that $$\alpha(q-p) + \alpha(p) \geq \alpha(q) + \frac14 p_2^{1-2\delta}.$$ [[[$\Box$ ]{}]{}]{} This deterministic convexity lemma, together with the concentration bound of Section \[sec:concentration\], will help us to show that geodesics cannot make large deviations from a straight line. For $p\in {{\mathbb R}}^2$, we define $$K(p,R):= \left\{q\in{{\mathbb R}}^2:\ |q_1-p_1|\leq R\ \mbox{and}\ |q_2-p_2|\leq R\right\}. \label{eq:K-square}$$ For $p,q\in {{\mathbb R}}^2$ satisfying $p_2<q_2$, we denote by $\gamma_{p,q}$ the optimal path from $p$ to $q$. For $p,z\in {{\mathbb R}}^2$ with $0<p_2<z_2$, we define the event $$G(p,z) = \left\{ \exists \tilde{0}\in K((0,0),1), \tilde{z}\in K(z,1):\ \gamma_{\tilde{0},\tilde{z}}\cap K(p,1) \neq \emptyset\right\}.$$ This says that a geodesic starting close to $(0,0)$ and ending near $z$, passes close to $p$. To bound the probability of the event $G(p,z)$, we first have to control the action to and from points close to $p$. For $p,z\in{{\mathbb R}}^2$ with $p_2<z_2$, we denote $$A(p,z)=A^{p_2,z_2}(p_1,z_1).$$ \[lem:boundaction\] Suppose $p,z\in {{\mathbb R}}^2$. Let $\tilde{p}\in K(p,1)$, $\tilde{z}\in K(z,1)$ with $\tilde{p}_2<\tilde{z}_2$. Define $\underline{p}=(p_1,p_2-2)$, $\bar{p}=(p_1,p_2+2)$ and similarly $\underline{z}$ and $\bar{z}$. Then $$A(\tilde{p},\tilde{z})\geq A(\underline{p},\bar{z}) - 1.$$ If $p_2+2<z_2-2$, then $$A(\bar{p},\underline{z})+1 \geq A(\tilde{p},\tilde{z}).$$ [[[Proof: ]{}]{}]{}Let $\tilde{\gamma}$ be the optimal path (i.e., the path that picks up the least action) from $\tilde{p}$ to $\tilde{z}$. Define $\gamma$ as the path that starts at $\underline{p}$, moves at constant speed to $\tilde{p}$, then follows $\tilde{\gamma}$, and then moves at constant speed to $\bar{z}$. Then, denoting $A[\gamma]$ for the action picked up by a path $\gamma$, we get $$\begin{aligned} A(\underline{p},\bar{z}) & \leq & A[\gamma]\\ & \leq & \frac12 + A(\tilde{p},\tilde{z}) + \frac12.\end{aligned}$$ We use that the first part of this path picks up at most $v^2/2$ action, where the speed $v\leq 1$. For the path from $\tilde{z}$ to $\bar{z}$ we get the same upper bound. For the second inequality, note that $$A(\tilde{p},\tilde{z}) \leq A(\tilde{p},\bar{p}) + A(\bar{p},\underline{z}) + A(\underline{z},\tilde{z}).$$ Clearly, as described above, we have $A(\tilde{p},\bar{p})\leq 1/2$ by taking a similar path from $\tilde{p}$ to $\bar{p}$. Likewise, $A(\underline{z},\tilde{z})\leq 1/2$. This proves the lemma. [[[$\Box$ ]{}]{}]{} For a speed $v>0$ we define $${\rm Co}(v) = \{ p\in{{\mathbb R}}\times {{\mathbb R}}^+:\ |p_1|\leq p_2v\}.$$ \[lem:dstraight1\] Fix $\delta\in(0,1/4)$ and $v>0$. There exist constants $c_1, c_2, M>0$ (independent of $\delta$), such that for all $p\in {\rm Co}(v)$ with $p_2>M$ and $z\in \partial_S{{\mathcal C}}(p,p_2^{1-\delta})$, we have $${{\mathbb P}}(G(p,z))\leq c_1\exp\left(-c_2p_2^{1/2-2\delta}/\log(p_2)\right).$$ [[[Proof: ]{}]{}]{}Let us choose any $M>8$. Take $z\in \partial_S{{\mathcal C}}(p,p_2^{1-\delta})$. Define $\underline{p}=(p_1,p_2-2)$, $\bar{p}=(p_1,p_2+2)$, and likewise $\underline{0}$, $\bar{0}$, $\underline{z}$ and $\bar{z}$. The event $G(p,z)$ implies that there exist three points $\tilde{0}\in K(0,1)$, $\tilde{p}\in K(p,1)$ and $\tilde{z}\in K(z,1)$ with $\tilde{p}_2<\tilde{z}_2$ such that $$A(\tilde{0},\tilde{z}) = A(\tilde{0},\tilde{p}) + A(\tilde{p},\tilde{z}).$$ It follows from Lemma \[lem:boundaction\] that $$A(\bar{0},\underline{z})\geq A(\tilde{0},\tilde{z}) - 1,\ \ A(\underline{0},\bar{p})\leq A(\tilde{0},\tilde{p}) + 1\ \mbox{and}\ A(\underline{p},\bar{z})\leq A(\tilde{p},\tilde{z}) + 1.$$ Therefore, $$A(\bar{0},\underline{z})\geq A(\underline{0},\bar{p}) + A(\underline{p},\bar{z}) - 3.$$ Furthermore, consider $\alpha(\underline{z}-\bar{0})$. Since $p_2^{1-\delta}<p_2$, we have that $z\in {\rm Co}(v+1)$. $$\begin{aligned} \alpha(\underline{z}-\bar{0}) & = & \alpha_0(z_2-4) + \frac12\,\frac{z_1^2}{z_2-4}\\ & = & \alpha(z) - 4\alpha_0 + \frac{2z_1^2}{z_2(z_2-4)}\\ & \leq & \alpha(z) -4\alpha_0 + 4(v+1)^2.\end{aligned}$$ Here we use that $8<M\leq z_2$ and $|z_1|\leq (v+1)z_2$. Therefore, following the same reasoning, we can choose $L>0$ independent of $p\in {\rm Co}(v)$ such that $$\alpha(\underline{z}-\bar{0})\leq \alpha(z) + L,\ \ \alpha(\bar{p}-\underline{0})\geq \alpha(p) - L\ \mbox{and}\ \alpha(\bar{z}-\underline{p})\geq \alpha(z-p) - L.$$ Combined with Lemma \[lem:convest\] this leads to $$| A(\bar{0},\underline{z}) - \alpha(\underline{z}-\bar{0})| + |A(\underline{0},\bar{p}) - \alpha(\bar{p}-\underline{0})| + |A(\underline{p},\bar{z}) - \alpha(\bar{z}-\underline{p})| \geq -3 - 3L + \frac14\,p_2^{1-2\delta}.$$ Define the event corresponding to the third term on the l.h.s. $$E_3 = \left\{ |A(\underline{p},\bar{z}) - \alpha(\bar{z}-\underline{p})| \geq \frac1{15}\,p_2^{1-2\delta}\right\},$$ and likewise $E_1$ and $E_2$. By enlarging $M$, we can make sure that $G(p,z)$ implies at least one of these three events. We will bound the probability of $E_3$, which is slightly more complicated than the other two, since $\bar{z}_2-\underline{p}_2$ cannot be made arbitrarily large by increasing $M$. Using the shear transformation we know that $$A(\underline{p},\bar{z}) - \alpha(\bar{z}-\underline{p})\stackrel{distr}{=} A((0,0),(0,\bar{z}_2-\underline{p}_2) = A^{\bar{z}_2-\underline{p}_2}.$$ Clearly, if $s\geq t$, then $|A^s|\geq |A^t|$. Therefore, $${{\mathbb P}}(E_3)\leq {{\mathbb P}}\left\{|A^{2p_2+4}|\geq \frac1{15}\,p_2^{1-2\delta}\right\}.$$ Using $t=2p_2+4$ and $u=\frac1{15}\,p_2^{1-2\delta}$, we see that Theorem \[thm:concentration\_around\_alphat\] guarantees the existence of constants $C_1,C_2>0$ such that for $M$ big enough and all $p$ with $p_2\geq M$, $${{\mathbb P}}(E_3) \leq C_1\exp\left\{-C_2 |p_2|^{1/2-2\delta}/\log(p_2)\right\}.$$ Similar bounds hold for ${{\mathbb P}}(E_1)$ and ${{\mathbb P}}(E_2)$, proving the lemma. [[[$\Box$ ]{}]{}]{} The above Lemma can be used to show that a minimal path starting close to the origin and passing close to $p$, with high probability will not exit the slanted cylinder ${{\mathcal C}}(p,p_2^{1-\delta})$ through the sides. Define the event $$G(p) = \left\{ \exists\ \tilde{0}\in K((0,0),1)\ \exists\ z\in \partial_S{{\mathcal C}}(p,p_2^{1-\delta}):\ \gamma_{\tilde{0},z}\cap K(p,1) \neq \emptyset\right\}.$$ \[lem:dstraight2\] Fix $\delta\in(0,1/4)$ and $v>0$. There exist constants $c_1, c_2, \kappa, M>0$, such that for all $p\in {\rm Co}(v)$ with $p_2>M$ we have $${{\mathbb P}}(G(p))\leq c_1\exp(-c_2p_2^\kappa).$$ [[[Proof: ]{}]{}]{}Suppose $p\in{{\mathbb R}}^2$ with $p_2>M$. There exists a constant $c>0$ (depending on $v$) and points $z_1,\ldots, z_L \in \partial_S{{\mathcal C}}(p,p_2^{1-\delta})$ with $L\leq cp_2$, such that $$\partial_S{{\mathcal C}}(p,p_2^{1-\delta}) \subset \bigcup_{i=1}^L K(z_i,1).$$ This implies that $$G(p)\subset \bigcup_{i=1}^L G(p,z_i).$$ Therefore, by choosing $M$ large enough, $\kappa < 1/2-2\delta$ and using Lemma \[lem:dstraight1\], there exist $C_1, C_2, c_1, c_2>0$ such that $${{\mathbb P}}(G(p)) \leq LC_1\exp(C_2p_2^{1/2-2\delta}/\log(p_2)) \leq c_1\exp(-c_2p_2^\kappa).$$ [[[$\Box$ ]{}]{}]{} Now we are ready to prove $\delta$-straightness of geodesics, as was introduced by Newman in [@Ne]. For a path $\gamma$ and $t\in {{\mathbb R}}$, we define $$\gamma^{\rm out}(t) = \{ (s,\gamma(s))\ :\ s\geq t\},$$ which is the set of all points in the path $\gamma$ (more precisely, the *graph* of $\gamma$), that are reached after time $t$. We also consider the following cone for all $x\in {{\mathbb R}}\times {{\mathbb R}}^+$ and $\eta>0$: $$\label{eq:Co} {\rm Co}(x,\eta) = \{ z\in {{\mathbb R}}\times {{\mathbb R}}^+\ :\ |z_1/z_2 - x_1/x_2| \leq \eta \},$$ which is the cone starting in the origin of all points $z$ that have a corresponding speed closer than $\eta$ to the speed of $x$. \[lem:delta\_straightness\] For $\delta\in(0,1/4)$ and $v>0$ we have with probability one that there exists $M>0$ (depending on $v$ and $\delta$) and $R>0$ (depending only on $\delta$), such that for all $\tilde{0}\in K((0,0),1)$, for all $z\in {{\mathbb R}}\times {{\mathbb R}}^+$ and for all $p\in \gamma(\tilde{0},z)\cap {\rm Co}(v)$ with $p_2>M$, we have $$\gamma^{\rm out}(p_2) \subset {\rm Co}(p,Rp_2^{-\delta}),$$ for $\gamma=\gamma_{\tilde{0},z}$. This lemma states that if a geodesic starting near $(0,0)$ passes through a remote point $p$, it has to stay in a narrow cone around the ray ${{\mathbb R}}^+\cdot p$. [[[Proof: ]{}]{}]{}Consider the events $G(\bar{p})$ for all $\bar{p}\in {{\mathbb Z}}\times {{\mathbb Z}}^+\cap {\rm Co}(v')$, with $v'>v$. Using Lemma \[lem:dstraight2\] and Borel–Cantelli Lemma, we can choose $M$ big enough, such that for all $\bar{p}\in {{\mathbb Z}}\times {{\mathbb Z}}^+\cap {\rm Co}(v')$ with $\bar{p}_2\geq M$ the event $G(\bar{p})$ does not happen. Increase $M$ if necessary to ensure that if $p_2\geq M$, $$K(p,1+(p_2+1)^{1-\delta})\subset {\rm Co}(p,2p_2^{-\delta}).$$ So for any $p\in{{\mathbb R}}\times{{\mathbb R}}^+$ with $p_2\geq M$, we now know that if $\bar{p}\in K(p,1)$, then $${{\mathcal C}}(\bar{p},\bar{p}_2^{1-\delta})\subset {\rm Co}(p,2p_2^{-\delta}).$$ Let $\bar{0}\in K((0,0),1)$. Now suppose there exists $z\in{{\mathbb R}}\times{{\mathbb R}}^+$ and $p\in \gamma_{\bar{0},z}\cap {\rm Co}(v)$ with $p_2>M+1$. Define $\bar{p}=(\lfloor p_1\rfloor,\lfloor p_2\rfloor)$. Suppose $z$ lies outside of the slanted tube ${{\mathcal}C}(\bar{p},\bar{p}_2^{1-\delta})$. We know that $G(\bar{p})$ does not happen, and since $p\in K(\bar{p},1)$, this implies that we can define $p^{(1)}$ as the crossing of $\gamma^{\rm out}(p_2)$ with the top edge of the tube ${{\mathcal}C}(\bar{p},\bar{p}_2^{1-\delta})$, and that $\gamma_{p,p^{(1)}}$ lies inside this tube. Define $\bar{p}^{(1)}=(\lfloor p^{(1)}_1\rfloor,\lfloor p^{(1)}_2\rfloor)$. We can proceed in a similar way to construct $p^{(2)},\bar{p}^{(2)},p^{(3)},\ldots,\bar{p}^{(m)}$, where we finish whenever $z\in {{\mathcal}C}(\bar{p}^{(m)},(\bar{p}^{(m)}_2)^{1-\delta})$. We have to check that for all $k\leq m$, $\bar{p}^{(k)}\in {\rm Co}(v')$, but this will follow from the following considerations. Note that for $1\leq k\leq m$, where we define $p^{(0)}=p$ and $\bar{p}^{(0)}=\bar{p}$, $${{\mathcal}C}(\bar{p}^{(k)},(\bar{p}^{(k)}_2)^{1-\delta})\subset {\rm Co}(p^{(k)},2(p_2^{(k)})^{-\delta})\ \ {\rm and}\ \ p_2^{(k)}\geq 2(p_2^{(k-1)}-1)\geq \frac32 p_2^{(k-1)}.$$ This implies that the average speed of any vector in $y\in \gamma^{\rm out}(p_2)$ satisfies $$|y_1/y_2 - p_1/p_2| \leq \sum_{k=0}^m 2(p_2^{(k)})^{-\delta} \leq \sum_{k=0}^m 2\left(\frac32\right)^{-\delta k}p_2^{-\delta}\leq Rp_2^{-\delta},$$ if we choose $R>0$ large enough (depending only on $\delta$). This also shows that we have to choose $v'>v+RM^{-\delta}$. [[[$\Box$ ]{}]{}]{} \[cor:delta-straightness-prob\] For $\delta\in(0,1/4)$ and $v>0$ there exists $M,R,\kappa,C_1,C_2>0$, such that when we define the event $$\begin{aligned} G_n & = & \Big\{\exists \tilde{0}\in K((0,0),1), z\in {{\mathbb R}}\times {{\mathbb R}}^+, p\in\gamma(\tilde{0},z)\ \mbox{with}\ p_2>n\ \mbox{and}\ p\in {\rm Co}(v)\ :\\ & & \ \ \gamma^{\rm out}(p_2) \not\subset {\rm Co}(p,R p_2^{-\delta})\Big\},\end{aligned}$$ we have for $n\geq M$ $${{\mathbb P}}(G_n)\leq C_1e^{-C_2n^\kappa}.$$ [[[Proof: ]{}]{}]{}It follows directly from the proof of Lemma \[lem:delta\_straightness\] that the event $G_{n}$ is a subset of the event that there exists $\bar{p}^{(n)}\in {{\mathbb Z}}\times {{\mathbb Z}}^+\cap {\rm Co}(v')$ with $\bar{p}^{(n)}_2\geq n-1$ such that the event $G(\bar{p}^{(n)})$ does happen. Here we choose $v'>v+RM^{-\delta}$. The probability of this event is clearly bounded by $C_1e^{-C_2n^\kappa}$, for an appropriate choice of constants, simply by Lemma \[lem:dstraight2\]. [[[$\Box$ ]{}]{}]{} Existence and uniqueness of semi-infinite minimizers {#existence-and-uniqueness-of-semi-infinite-minimizers} ---------------------------------------------------- With $\delta$-straightness in hand, we can prove some important properties of minimizing paths. A semi-infinite minimizer starting at $(x,t)\in{{\mathbb R}}^2$ is a path $\gamma:[t,\infty)\to {{\mathbb R}}$ such that $\gamma(t)=x$ and the restriction of $\gamma$ to any finite time interval is a minimizer. We call $(x,t)$ the endpoint of $\gamma$. \[lem:geodesic direction\] With probability one, all semi-infinite minimizers have an asymptotic slope (velocity, direction): for every minimizer $\gamma$ there exists $v\in {{\mathbb R}}\cup\{\pm \infty\}$ depending on $\gamma$ such that $$\lim_{t\to\infty} \frac{\gamma(t)}{t} = v.$$ [[[Proof: ]{}]{}]{}Let us fix a sequence $v_n\to \infty$. Using the translation invariance of the Poisson point field, with probability one, for any $q\in {{\mathbb Z}}^2$ we can choose a corresponding sequence of constants $M_n(q)>0$ such that the statement in Lemma \[lem:delta\_straightness\] holds for the entire sequence, for paths starting in $K(q,1)$. Let us take some one-sided minimizer $\gamma$. If $\gamma(t)/t\to+\infty$ or $-\infty$, then the desired statement is automatically true. In the opposite case we have $$\liminf_{t\to\infty} \frac{|\gamma(t)|}{t} < \infty.$$ This implies that there exist $n\geq 1$ and a sequence $t_m\to \infty$ such that $|\gamma(t_m)|/t_m \leq v_n$. We define $y_m=(\gamma(t_m),t_m)$ and choose $q\in{{\mathbb Z}}^2$ such that $y_1\in K(q,1)$. For $m$ large enough, we will have that $t_m>M_n(q)$ and, therefore, $$\gamma^{\rm out}(y_m)\subset q + {\rm Co}(y_m-q,R|y_m-q|^{-\delta}),$$ for some constant $R>0$ and $m$ large enough. Clearly this implies that $\gamma$ must have a finite asymptotic slope. [[[$\Box$ ]{}]{}]{} \[lem:exist-geodesic\] With probability one, for every $v\in{{\mathbb R}}$ and for every sequence $(y_n,t_n)\in {{\mathbb R}}^2$ with $t_n\to \infty$ and $$\lim_{n\to \infty} \frac{y_n}{t_n} = v,$$ and for every $x\in {{\mathbb R}}^2$, there exists a subsequence $(n_k)$ such that the minimizing paths $\gamma_{x,(y_{n_k},t_{n_k})}$ are an increasing collection of paths that converge to a semi-infinite minimizer starting at $x$ and with asymptotic speed equal to $v$. [[[Proof: ]{}]{}]{}Without loss of generality, we can assume that $x\in K((0,0),1)$. We take a sequence $v_m\to \infty$ and then choose $M_m\to \infty$ and $R>0$ such that the statement of Lemma \[lem:delta\_straightness\] holds for every triplet $(v_m,M_m,R)$. From these triplets we choose a triplet $(v_0,M,R)$ with $|v|<v_0-2RM^{-\delta}$ (note that $R$ only depends on $\delta$). By going to a subsequence, we make sure that for all $n\geq 1$, $t_n\geq M$, $t_n\uparrow \infty$ and for all $k>n$ we have $(y_k,t_k)\in {\rm Co}((v,1),Rt_{n}^{-\delta})$. Consider the paths $\gamma_n = \gamma_{x,(y_n,t_n)}$. We claim that for each $n\geq 1$ and $k>n$, the path $\gamma_k$ lies in the cone ${\rm Co}((v,1),2Rt_n^{-\delta})$ for times larger than $t_n$. In fact, if $\gamma_k$ visits a point $p$ outside this cone (but inside ${\rm Co}(v_0)$) at time $p_2\geq t_n$, then $\gamma_k$ violates the $\delta$-straightness condition (the relevant cone through $p$ will not intersect ${\rm Co}((v,1),Rt_n^{-\delta})$, and therefore it does not contain $y_k$). In particular, this means that there exists a $C>0$ (independent of $n$) such that all paths $\gamma_k$ with $k>n$ cross the segment $$J_n=[vt_n-Ct_n^{1-\delta},vt_n+Ct_n^{1-\delta}]\times \{t_n\}.$$ We claim that there exists a point $x_n\in J_n$ visited by an infinite number of paths $\gamma_k$. With this claim in hand, we first choose $x_1$ and a subsequence $k_1(n)$ such that every $\gamma_{k_1(n)}$ passes through $x_1$, then we choose $x_2$ in the segment $$[vt_{k_1(1)}-Ct_{k_1(1)}^{1-\delta},vt_{k_1(1)}+Ct_{k_1(1)}^{1-\delta}]\times \{t_{k_1(1)}\}$$ such that an infinite number (a subsequence $k_2(n)\subset k_1(n)$) of the paths $\gamma_{k_1(n)}$ pass through $x_2$, and so on. The paths $\{\gamma_{x,x_n}\ :\ n\geq 1\}$ are then an increasing collection of paths, and their union will be a semi-infinite minimizer starting at $x$, with asymptotic speed $v$.\ What remains is to show that an infinite number of the paths $\gamma_k, k>n$, pass through the same point of the segment $$[vt_n-Ct_n^{1-\delta},vt_n+Ct_n^{1-\delta}]\times \{t_n\}.$$ We already know that all these paths remain in the cone ${{\mathcal}C} = {\rm Co}((v,1),Rt_n^{-\delta})$. We use the following fact about the Poisson process, which is easily checked using the Borel–Cantelli Lemma: with probability one, for all $K>0$ there exists $T_0>0$ depending on $K$, such that for all $T\geq T_0$ and all $x\in [-KT,KT]\times [-1,2T]$, there is at least one Poisson point in the ball of radius $T^{1/4}$ around $x$. Let us choose $K>0$ and $T>t_n$ such that ${{\mathcal}C}\cap ({{\mathbb R}}\times [0,t_n+T])\subset [-KT,KT]\times [-1,2T]$. Then we choose $T_0>t_n$ according to this $K$ and take $T\geq T_0$. Consider $k$ such that $t_k>t_n+2T$. Suppose $\gamma_k$ did not pick up any Poisson point in the time interval $[t_n,t_n+2T]$. That implies that it would have some constant speed $u$ in that interval. We also know that within distance $T^{1/4}$ of $(\gamma_k(t_n+T),t_n+T)$ there will be some Poisson point, since $(\gamma_k(t_n+T),t_n+T)\in {{\mathcal}C}$. Define the path $\tilde{\gamma}_k$ that picks up this Poisson point by moving to this point with constant velocity from the point $(\gamma_k(t_n),t_n)$ and then moving with constant speed to $(\gamma_k(t_n+2T),t_n+2T)$; the rest of the time $\tilde{\gamma}_k$ coincides with $\gamma_k$. When we consider the difference in action picked up by the two paths, we note that they start and end at the same point in the time interval $[t_n,t_n+2T]$. This means that if we define $\delta(t) = \dot{\tilde{\gamma}}_k(t) - \dot\gamma_k(t)$, then $\int_{t_n}^{t_n+2T} \delta(s)\,ds = 0$. Furthermore, there exists a constant $d>0$ depending only on $v$ and the cone ${\mathcal}C$ such that $|\delta(s)|\leq dT^{-3/4}$. If we choose $T>d^4$, this leads to $$\begin{aligned} A[\tilde{\gamma}_k] - A[\gamma_k] & = & -1 + \frac12 \int_{t_n}^{t_n+2T} ((u+\delta(s))^2 - u^2)\,ds\\ & = & -1 + \frac12 \int _{t_n}^{t_n+2T} \delta(s)^2\,ds\\ & \geq & -1 + d^2 T^{-1/2}\\ & < & 0.\end{aligned}$$ This contradicts the optimality of $\gamma_k$, and we conclude that $\gamma_k$ must pick a Poisson point in the time interval $[t_n,t_n+2T]$. The number of paths $\gamma_k$ is infinite, and there are only finitely many Poisson points in the set ${{\mathcal}C}\cap {{\mathbb R}}\times [t_n,t_n+2T]$. Therefore, at time $t_n$, an infinite number of paths $\gamma_k$ cross $J_n$ at the same point. Here we use that when two minimizers meet at two Poisson points at distinct times, they actually will coincide for all times between these two times (since minimizing paths between two Poisson points are almost surely unique). [[[$\Box$ ]{}]{}]{} \[lem:geodesics-do-not-share-more-than-one\] With probability 1, the following statement holds: if $\gamma_1$ and $\gamma_2$ are two (finite-time) geodesics, starting at the same Poisson point $p$, and for some $t>p_2$ we have $\gamma_1(t)<\gamma_2(t)$, then for all (relevant) $s>t$ we have $\gamma_1(s)<\gamma_2(s)$. [[[Proof: ]{}]{}]{}The probability that there are two Poisson point connected by two distinct geodesics is zero. So we only have to consider the situation where two paths with vertices $p,p_1,\ldots,p_n$ and $p,q_1,\ldots,q_m$ intersect transversally, i.e., for some $k,j$, $[p_j,p_{j+1}]\cap [q_k,q_{k+1}]=\{x\}$ for some point $x\notin\omega$. In this case, the total actions of the two paths can be improved by switching to paths with vertices $p,p_1,\ldots,p_j,q_{k+1},\ldots,q_m$ and, respectively, $p,q_1,\ldots,q_{k},p_{j+1},\ldots,p_n.$ Therefore, we obtain a contradiction with the optimality of the original paths. [[[$\Box$ ]{}]{}]{} \[lem:uniqueness-of-geodesic\] Let $v\in {{\mathbb R}}$. With probability one, every Poisson point belongs to at most one semi-infinite minimizer with asymptotic slope $v$. [[[Proof: ]{}]{}]{}Let ${{\mathcal}U}(v)$ be the event that two distinct semi-infinite minimizers with asymptotic speed $v$ pass through a common Poissonian point $p$. Let $${{\mathcal}S}=\{v\in {{\mathbb R}}\ :\ {{\mathbb P}}({{\mathcal}U}(v))>0\}.$$ It is sufficient to show that ${{\mathcal}S}=\emptyset$. The invariance under shear transformations implies that either ${{\mathcal}S}=\emptyset$ or ${{\mathcal}S}={{\mathbb R}}$. The latter will be excluded as soon as we prove that ${{\mathcal}S}$ is at most countable. A triple of distinct Poisson points $(p,q_1,q_2)$ is called a bifurcation triple for $v$ if there exist two distinct semi-infinite minimizers $\gamma_1$ and $\gamma_2$ with asymptotic slope $v$ that both start at $p$, then one goes directly (at constant velocity) to $q_1$ and the other goes directly to $q_2$. We choose $q_1$ such that $\gamma_1$ lies to the left of $\gamma_2$. Lemma \[lem:geodesics-do-not-share-more-than-one\] implies that $\gamma_1$ and $\gamma_2$ will not meet again after $p$ with probability one. Clearly, ${{\mathcal}U}(v)$ implies the existence of such a bifurcation triple for $v$. If ${{\mathbb P}}({{\mathcal}U}(v))>0$, then there exists $L=L(v)\in {\mathbb N}$ such that the event $${{\mathcal}T}_L(v) = \{ \exists \mbox{ bifurcation triple}\ (p,q_1,q_2)\ \mbox{for } v\mbox{ inside the cube } K((0,0),L)\}$$ has positive probability (using translation invariance). Now suppose that ${{\mathcal}S}$ contains an uncountable number of asymptotic slopes. This implies that there exist $m,L\in {\mathbb N}$ such that for uncountably many $v$ we would have that $$\label{eq:biftriple} {{\mathbb P}}({{\mathcal}T}_L(v))> 1/m.$$ Now note that three Poisson points can only form a bifurcation triple for one $v\in {{\mathbb R}}$, since otherwise two distinct semi-infinite minimizers, both starting at some $p$, would have to split up at $p$ and cross again at a later time, which contradicts Lemma \[lem:geodesics-do-not-share-more-than-one\]. Suppose $v_1,v_2,\ldots$ satisfy . Denote by $N_L$ the number of Poisson points in the cube $K((0,0),L)$. Then $$\sum_{n\geq 1} 1_{{{\mathcal}T}_L(v_n)} \leq N_L^3.$$ Taking the expectation on both sides and using leads to a contradiction. Therefore, there can be only countably many $v$’s in ${\mathcal}S$ which completes the proof. [[[$\Box$ ]{}]{}]{} With uniqueness in hand we can strengthen Lemma \[lem:exist-geodesic\]: \[lem:convergence-of-geodesics\] With probability one, for every $v\in{{\mathbb R}}$ and for every sequence $(y_n,t_n)\in {{\mathbb R}}^2$ with $t_n\to \infty$ and $$\lim_{n\to \infty} \frac{y_n}{t_n} = v,$$ and for every Poissonian point $p\in {{\mathbb R}}^2$, the minimizing paths $\gamma_{p,(y_{n},t_{n})}$ converge to a unique semi-infinite minimizer $\gamma_{p,v}$ starting at $p$ and with asymptotic speed equal to $v$. [[[Proof: ]{}]{}]{}Let us assume that the convergence does not hold, i.e., there is a sequence $(n')$ such that the restrictions of $\gamma_{p,(y_{n'},t_{n'})}$ and $\gamma_{p,v}$ on some finite time interval $I$ do not coincide for all $n'$. Lemma \[lem:exist-geodesic\] allows to choose a subsequence $(n'')$ from $(n')$ such that for sufficiently large $n''$ the restrictions of $\gamma_{p,(y_{n''},t_{n''})}$ on $I$ coincide with the restrictions of some infinite one-sided geodesics $\gamma'$. The uniqueness established in Lemma \[lem:uniqueness-of-geodesic\] guarantees that $\gamma'$ coincides with $\gamma_{p,v}$, and the resulting contradiction shows that our assumption was false, completing the proof. [[[$\Box$ ]{}]{}]{} Coalescence of minimizers ------------------------- Here we prove that any two one-sided minimizers with the same asymptotic slope coalesce. \[lem:minimizersdontcross\] With probability one it holds that for every $v\in {{\mathbb R}}$ and for every pair of semi-infinite minimizers, starting at different Poisson points, with asymptotic speed $v$, these minimizers either do not touch, or they coalesce at some Poisson point. [[[Proof: ]{}]{}]{}Suppose for some $v\in {{\mathbb R}}$, there do exist two semi-infinite minimizers with asymptotic speed $v$ that touch, but do not coalesce. If the two minimizers $\gamma_1$ and $\gamma_2$ contain the same Poisson point $p$, then they must stay together for all times above $p$ according to Lemma \[lem:uniqueness-of-geodesic\]. Therefore, the only option is that $\gamma_1$ and $\gamma_2$ cross, i.e., they consecutively visit Poissonian points $p_1,p_2,\ldots$, and, respectively, $q_1,q_2,\ldots$, and $[p_1,p_2]\cap [q_1,q_2] = \{x\}$, for some $x\in {{\mathbb R}}^2$. The sequence $\{q_m:m\geq 1\}$ satisfies the conditions of Lemma \[lem:convergence-of-geodesics\], which means that the minimizers $\gamma_{p_1,q_{m}}$ converge to $\gamma_1$. However, we claim that with probability 1, none of the minimizers $\gamma_{p_1,q_m}$ contain any of the $p_n$ $(n\geq 2)$. In fact, if this claim is violated for some $m,n$, then due to a.s.-uniqueness of a geodesic between any two Poisson points, we know that $\gamma_{p_1,q_m}$ passes through $p_2$ and $x$. This implies that action picked up by $\gamma_2$ between $x$ and $q_m$ must be equal to the action picked up by $\gamma_{p_1,q_m}$ between $x$ and $q_m$. However, this contradicts the optimality of $\gamma_2$ as the comparison with the path connecting $q_1$ directly to $p_2$ and then following $\gamma_{p_1,q_m}$ shows. The proof is complete.[[[$\Box$ ]{}]{}]{} For every $v\in{{\mathbb R}}$, coalescence of one-sided geodesics with asymptotic slope $v$ generates an equivalence relation on Poissonian points. We call each equivalence class a coalescence component. \[lem:unbounded\_trees\] Let $v\in{{\mathbb R}}$. With probability 1, every coalescence component is unbounded below in time. [[[Proof: ]{}]{}]{}Due to shear invariance, it is sufficient to consider only $v=0$. Suppose that the probability of existence of a coalescence component bounded from below is positive. Then there is $(i,j)\in{{\mathbb Z}}^2$ such that with positive probability the earliest (i.e., having the minimal time coordinate) point in some coalescence component belongs to $[i,i+1]\times[j,j+1]$. Due to stationarity, this probability is positive for any $(i,j)\in{{\mathbb Z}}^2$. The Tempelman multi-parametric ergodic theorem (see, e.g., [@Krengel:MR797411 Chapter 6]) implies that there is a constant $c>0$ and a family of random variables $(n_r)_{r\in{{\mathbb N}}}$ such that with probability 1, $$\label{eq:quadratic_growth_of_number_of_trees} \sum_{i,j\in \Delta(r,n)}I_{ij}> cn^2,\quad n\ge n_r.$$ where $\Delta(r,n)=\{(i,j):\ r\le j \le n,\ -j\le i \le j\}$ and $I_{ij}$ is the indicator of event that there is a tree with earliest point within $[i,i+1]\times[j,j+1]$. Let $\gamma_+=\gamma_{3,0}$ and $\gamma_-=\gamma_{-3,0}$ be the one-sided optimal paths emitted from the origin $0$ with asymptotic slopes $3$ and $-3$, respectively. Then there is a random variable $m$ such that for all $t>m$, one has $-4t<\gamma_-(t)< -2t$ and $2t<\gamma_+(t)<4t$. In particular, the set $$\bigcup_{n\ge m} \bigcup_{(i,j)\in \Delta(m,n)} [i,i+1]\times[j,j+1]$$ is bounded on the left by $\gamma_-$ and on the right by $\gamma_+$. Let us denote the coalescence component giving rise to $I_{ij}=1$ by $C_{ij}$. If there are several such components we choose the one containing the earliest point. All the components $C_{ij}$, $j>m$ are disjoint and contained in the set $\{(x,t): t>m,\ -2t\le x\le 2t\}$. Combining this with , we can conclude that for every $n\ge n_m$, the segment $J_n=[\gamma_-(n),\gamma_+(n)]\times \{n\}$ is crossed by geodesics from at least $cn^2$ disjoint coalescence components. Each of these components has to have an edge connecting two Poissonian points and crossing $J_n$. Therefore, either at least $cn^2$ Poissonian points are contained in $R_n=[-4n,4n]\times[n,n-\sqrt{n}]$, or there is an edge connecting two vertices of one component and passing through a point on $H_n=[-4n,4n]\times\{n-\sqrt{n}\}$ and a point on $L_n=[-4n,4n]\times\{n\}$. Let us denote the former event by $D_n$ and the latter by $E_n$. We have $$\begin{aligned} {{\mathbb P}}(D_n)\le e^{-cn^2}{{\mathbb E}}e^{\omega(R_n)}\le e^{-cn^2} e^{8n^{3/2}(e-1)},\end{aligned}$$ and the Borel–Cantelli Lemma implies that with probability 1 only finitely many events $D_n$ occur. To prove an analogous statement for the events $E_n$, we need the following lemma: \[lem:no\_points\_in\_parallelogram\] Let $x,y\in{{\mathbb R}}$. Suppose there is an optimal path $\gamma$ connecting a point $x'\in [x,x+n^{1/5}]\times\{n-\sqrt{n}\}$ straight to a point $y'\in[y,y+n^{1/5}]\times\{n\}$, avoiding all Poissonian points between $n-\sqrt{n}$ and $n$. Then the parallelogram $\Pi_n(x,y)$ that is obtained by intersection of ${{\mathbb R}}\times [n-\sqrt{n}-1,n-\sqrt{n}+1]$ and the parallelogram with vertices $x,x+n^{1/5},y+n^{1/5},y$, does not contain any Poissonian points. [[[Proof: ]{}]{}]{}Suppose that $\Pi_n(x,y)$ contains a Poissonian point $p=(z,s)$. Then the original straight path is not optimal, which follows from the comparison with the path $\tilde \gamma$ consisting of two segments connecting $x'$ to $p$ and $p$ to $y'$. In fact, an elementary calculation (see also the proof of Lemma 6.5 in [@Bakhtin-quasicompact]) shows that $$S^{n-\sqrt{n},n}(\tilde\gamma)-S^{n-\sqrt{n},n}(\gamma)=\frac{\sqrt{n}}{2(s-n+\sqrt{n})(n-s)}\delta^2,$$ where $\delta$ is the distance from $p$ to $\gamma$ measured along the spatial axis. Noticing that $\delta\le n^{1/5}$ and $s\in[n-\sqrt{n}-1,n-\sqrt{n}+1]$, we obtain that for large $n$, the right-hand side is less than $1$, so increasing the number of points that the path passes through by $1$ overweighs the increase of kinetic action. [[[$\Box$ ]{}]{}]{} To finish the proof of Lemma \[lem:unbounded\_trees\], we notice that each $H_n$ and $L_n$ can be covered by $8n^{1-1/5}+1$ intervals of length $n^{1/5}$. The probability of $E_n$ can be then bounded by the sum over all pairs of these intervals such that there is a straight line minimizer connecting points from these two intervals. Lemma \[lem:no\_points\_in\_parallelogram\] implies that $$\begin{aligned} {{\mathbb P}}(E_n)&\le (8n^{1-1/5}+1)^2{{\mathbb P}}\{\omega(\Pi_n(0,0))=0\}\\ &\le (8n^{1-1/5}+1)^2 e^{-2n^{1/5}}.\end{aligned}$$ since the area of $\Pi_n(x,y)$ equals $2n^{1/5}$ and does not depend on $x,y$. The Borel–Cantelli Lemma implies that with probability $1$, only finitely many events $E_n$ can happen. This completes the proof of Lemma \[lem:unbounded\_trees\]. [[[$\Box$ ]{}]{}]{} \[lem:coalescense of geodesics\] Let $v\in{{\mathbb R}}$. With probability one, every two semi-infinite minimizers with asymptotic slope $v$ coalesce. [[[Proof: ]{}]{}]{}Due to shear invariance, it is enough to prove the lemma for $v=0$. Also, it is enough to prove it for minimizers starting at Poisson points. Lemma \[lem:minimizersdontcross\] shows that if two of these semi-infinite minimizers do not coalesce, then they must be disjoint. We call the event that this happens $E$, and suppose ${{\mathbb P}}(E)>0$. Now we define $E_k$ $(k\in \mathbb Z)$ as the event that there exist two disjoint minimizers with asymptotic slope $0$ that both start at Poisson points with times smaller than $k$. Clearly, there exists $k\in {{\mathbb Z}}$ with ${{\mathbb P}}(E_k)>0$, and using translation invariance, we conclude that ${{\mathbb P}}(E_{-1})>0$. Let us now define, for $x_1,x_2\in {{\mathbb R}}\times (-\infty,-1)$ and $\delta\in(0,1)$, $E(x_1,x_2;\delta)$ as the event that there exist two disjoint minimizers $\gamma_1$ and $\gamma_2$ with asymptotic slope $0$, such that $\gamma_1$ starts within distance $\delta>0$ from $x_1$ and $\gamma_2$ starts within distance $\delta$ from $x_2$. Again it follows that there exist $x_1,x_2\in {{\mathbb R}}\times (-\infty,-1)$ and $\delta\in(0,1)$ such that ${{\mathbb P}}(E(x_1,x_2;\delta))>0$. Let $W_i$ be the crossing of the minimizer starting near $x_i$ with the $x$-axis. Clearly, there exists $K>0$ such that the event $E(x_1,x_2;\delta,K)$, consisting of all $\omega\in E(x_1,x_2;\delta)$ with $|W_i-x_{i1}|<K$, has positive probability. Now we choose $h>|x_1-x_2|+2\delta+2K$ and define $z=(h,0)$. Using the ergodicity of the Poisson point field with respect to translations, we obtain that there are $m,n\in {{\mathbb Z}}$, $m\neq n$ such that $${{\mathbb P}}(E(x_1+mz,x_2+mz;\delta,K)\cap E(x_1+nz,x_2+nz;\delta,K))>0.$$ Again using translation invariance, we can take $m=0$ and $n>0$. Let $x_3=x_1+nz$ and $x_4=x_2+nz$. Because of our choice of $h$, we know that $\max(W_1,W_2)<\min(W_3,W_4)$. Now relabel (if necessary) to ensure that $W_1<W_2<W_3<W_4$. We denote the minimizer starting near $x_i$ by $\gamma_i$. By construction, $\gamma_1$ and $\gamma_2$ do not cross, nor do $\gamma_3$ and $\gamma_4$. This implies that the three minimizers $\gamma_1,\gamma_2$ and $\gamma_4$ do not cross for times larger than $0$, since $\gamma_3$ lies between $\gamma_2$ and $\gamma_4$ (in principle, $\gamma_2$ and $\gamma_3$ could still coalesce). Our conclusion is that there is a positive probability of having three non-intersecting minimizers with asymptotic speed $0$, all three starting below time $0$.\ We can now conclude that for $\delta>0$ small enough and $K,N>0$ big enough, there exist $R>0$, $x\in {{\mathbb R}}\times \{-\delta\}$, $y_1,\ldots y_n \in {{\mathbb R}}\times (-\infty, 0)$ and $z_1,\ldots ,z_m\in{{\mathbb R}}\times (-\infty, 0)$ with $y_{n,2}\leq -K-\delta$ and $z_{m,2}\leq -K-\delta$, such that with positive probability: - There exist three disjoint semi-infinite minimizers $\gamma_1,\gamma_2$ and $\gamma_3$ with asymptotic slope $0$ and $\gamma_1(0)<\gamma_2(0)<\gamma_3(0)$, where $\gamma_1$ starts at a Poisson point $p_1$ in $y_n+[0,\delta]^2$, $\gamma_3$ starts at a Poisson point $p_3$ in $z_m+[0,\delta]^2$ and $\gamma_2$ starts at a Poisson point $p_2$ in $x+[0,\delta]^2$. - If $p_{1,2}<p_{3,2}$, then $|p_{3,1}-\gamma_1(p_{3,2})|<N$; if $p_{1,2}\ge p_{3,2}$, then $|p_{1,1}-\gamma_3(p_{1,2})|<N$. - There is exactly one Poisson point in $x+[0,\delta]^2$ and the next Poisson point of $\gamma_2$ (above $x$) has a strictly positive time. - There is exactly one Poisson point in $y_i+[0,\delta]^2$ and in $z_j+[0,\delta]^2$, for all $1\leq i\leq n$ and $1\leq j\leq m$. - All Poisson points of $\gamma_1$ with negative time lie in $\cup_{i=1}^n( y_i+[0,\delta]^2)$. - All Poisson points of $\gamma_3$ with negative time lie in $\cup_{j=1}^m (z_j+[0,\delta]^2)$. - $\gamma_1$ and $\gamma_3$ cross the $x$-axis in the interval $[-R,R]$. Let us call this event $A_{\delta,K,N}$. We choose $M>R$ and $L>K+\delta$ such that $[-M,M-\delta]\times [-L,0]$ contains all $y_i$’s, all $z_j$’s and $x$. If $\omega\in A_{\delta,K,N}$, then we can modify the Poisson configuration $\omega$ by deleting all Poisson points in $$[-M,M]\times [-L,0] \setminus \left(\bigcup_{i=1}^n (y_i+[0,\delta]^2)\,\cup\, \bigcup_{j=1}^m (z_j+[0,\delta]^2)\, \cup\, (x+[0,\delta]^2)\right),$$ and call this new configuration $\tilde{\omega}$. Clearly, $\gamma_1, \gamma_2$ and $\gamma_3$ are still semi-infinite minimizers under $\tilde{\omega}$. Furthermore, Lemma 8 of [@HoNe3] shows that there exists a measurable event $\tilde{A}_{\delta,K,N}$ with positive probability, such that $$\tilde{A}_{\delta,K,N} \subset \{ \tilde{\omega}\ :\ \omega \in A_{\delta,K,N}\}.$$ We claim that on the event $\tilde{A}_{\delta,K,N}$, any semi-infinite minimizer $\gamma$, starting at a Poisson point $p$ in the set $({{\mathbb R}}\times (-\infty,0] )\setminus ([-M,M]\times [-L,0])$ cannot coalesce with $\gamma_2$. Therefore, the coalescence component of $\gamma_2$ is bounded below by $-L$ which contradicts Lemma \[lem:unbounded\_trees\] and finishes the proof. Suppose our claim is wrong and $\gamma$ connects to $\gamma_2$. Since $\gamma$ cannot cross $\gamma_1$ or $\gamma_3$, this implies that $\gamma$ passes *between* $\gamma_1$ and $\gamma_3$, and thus at most at distance $N$ from either $p_1$ or $p_3$, when going with constant velocity from the boundary of $[-M,M]\times [-L,0]$ to $p_2$ or a point on the $x$-axis between $\gamma_1$ and $\gamma_3$. Therefore, fixing $N$, choosing $K$ large enough, and increasing $M$ and $L$ if necessary, we can guarantee that if we modify $\gamma$ and let it pick up one additional Poissonian point $p_1$ or $p_3$, we will decrease its action which contradicts the optimality of $\gamma$. The proof is completed. [[[$\Box$ ]{}]{}]{} Domains of influence of Poisson points -------------------------------------- Let us fix a mean velocity value $v\in{{\mathbb R}}$ for the rest of this section. By a one-sided optimal path we will always mean a one-sided optimal path of asymptotic slope $v$. In this section we describe the random partition of the real line $R\times\{0\}$ (or, equivalently, $R\times\{t\}$ for any $t\in{{\mathbb R}}$) into domains of influence of Poisson points. The domain of influence of $p\in\omega$ consists of all points $x\in{{\mathbb R}}$ such that the first Poisson point visited by a unique one-sided minimizer with endpoint $(x,0)$ is $p$. With a little more effort one can give a description of a space-time tessellation of ${{\mathbb R}}^2$ of the same kind, but we omit it for brevity. \[lem:limit-of-minimizers\] With probability one the following holds true. For any convergent sequence of real numbers $x_n\to x_\infty$, and any sequence $\gamma_{n}$ of one-sided minimizers with endpoints $(x_n,0)$, if there is a point $q$ such that the first Poissonian point visited by each $\gamma_{n}$ is $q$, then (i) after $q$ all these paths are uniquely defined and coincide, and (ii) the path $\gamma_{\infty}$ starting at $(x_\infty,0)$ going straight to $q$ and coinciding with paths $\gamma_{n}$ after $q$, is also a one-sided optimal path. [[[Proof: ]{}]{}]{}After $q$, the one-sided optimal path is uniquely defined due to Lemma \[lem:uniqueness-of-geodesic\]. Assuming that there is a finite improvement of $\gamma_\infty$, i.e., another path starting at $(x_\infty,0)$ connecting to $\gamma_1$ above $q$, one can also construct an improvement of a path $\gamma_n$ for sufficiently large $n$, i.e., for $(x_n,0)$ sufficiently close to $(x_\infty,0)$, a contradiction. [[[$\Box$ ]{}]{}]{} \[lem:finitely-many-Poissonian-directions\] With probability one, for every bounded set $S\in{{\mathbb R}}^2$ there is a finite set $P$ of Poissonian points such that for every one-sided optimal path originating at any $p\in S$, the first Poissonian point it visits belongs to $P$. [[[Proof: ]{}]{}]{}It is easily verified that for any $v>0$ there there is a random $t_0>0$ such that no Poissonian point $(x,t)$ satisfying $|x|\le vt$ and $t\ge t_0$ can be the first point visited by a one-sided optimal path originating within $K((0,0),1)$. Also, for any $R,T>0$, there are only finitely many Poissonian points $(x,t)$ satisfying $|x|\le R$ and $t<T$. Therefore, if there are countably many one-sided minimizers originating from a point within $K((0,0),1)$, then for any $R>0$ and any $v>0$ there is a Poissonian point $(x,t)$ satisfying $|x|>\max\{R,vt\}$ visited first by a one-sided optimal path originating in $K((0,0),1)$. Since one-sided minimizers cannot cross each other, we see that on the above event, there is a point $p=(y,s)\in K((0,0),1) $ such that one of the rays $[y,+\infty)\times \{s\}$, or $(-\infty,y]\times \{s\}$ cannot be crossed by one-sided minimizers originating at negative times. So, if this happens with positive probability, then, due to stationarity and ergodicity, with probability one, there are infinitely many Poissonian points $(y_n,s_n)_{n\in{{\mathbb Z}}}$ satisfying $s_n\in[0,1]$ and $\lim_{n\to\pm\infty} y_n=\pm\infty$ such that one of the rays $[y_n,+\infty)\times \{s_n\}$, or $(-\infty,y_n]\times \{s_n\}$ cannot be crossed by any one-sided optimal path starting at a negative time. Therefore, no one-sided optimal path can originate at a negative time. This contradiction proves the lemma for $S=K((0,0),1)$. The lemma in full generality follows due to stationarity, since one can cover any bounded $S$ by finitely many translates of $K((0,0),1)$. [[[$\Box$ ]{}]{}]{} \[lem:continuity-and-uniqueness-of-geodesics\] The following holds with probability one for all $x\in{{\mathbb R}}$ simultaneously. Suppose there are $n\in{{\mathbb N}}$ one-sided optimal paths originating at $(x,0)$. Let us denote by $(x_1,t_1)$,…,$(x_n,t_n)$ the first Poissonian points visited by these paths, ordering them so that $(x_1-x)/t_1<\ldots<(x_n-x)/t_n$ (note that we still allow $n=1$). Then there is ${{\varepsilon }}>0$ such that (i) for all $y\in(x-{{\varepsilon }},x)$, there is a unique one-sided optimal path originating at $(y,0)$, and the first Poissonian point visited by it is $(x_1,t_1)$, and (ii) for all $y\in(x,x+{{\varepsilon }})$, there is a unique one-sided optimal path originating at $(y,0)$, and the first Poissonian point visited by it is $(x_n,t_n)$. [[[Proof: ]{}]{}]{}We will prove only part (ii), since the proof of part (i) is the same. Suppose that there is a sequence $y_k\downarrow x$ and a sequence $(\gamma_k)$ of one-sided minimizers originating at $y_k$ with first visited Poissonian point $q_k\ne (x_n,t_n)$. Lemma \[lem:finitely-many-Poissonian-directions\] implies that the set $\{q_k\}$ is finite, and we can choose a Poissonian point $q=(x',t')$ from this set and a subsequence $y_{k'}$ such that the first Poissonian point visited by the corresponding paths $\gamma_{k'}$ is $q$. According to Lemma \[lem:limit-of-minimizers\], there is an optimal path connecting $x=\lim {y_{k'}}$ to $q$. However, optimal paths cannot cross, and it follows from our construction that $(x'-x)/t'>(x_n-x)/t_n$, and thus $q$ cannot belong to the set of $n$ Poissonian points in the statement of the Theorem. The resulting contradiction finishes the proof. [[[$\Box$ ]{}]{}]{} \[lem:domains-on-real-line\] With probability 1, there is an increasing doubly infinite sequence of points $(x_k)_{k\in{{\mathbb Z}}}$ on the real line ${{\mathbb R}}$ satisfying $\lim_{k\to\pm\infty}x_k=\pm\infty$ with the following properties: 1. For every $k\in{{\mathbb Z}}$ there is a Poisson point $p_k$ such that for all $x\in (x_k,x_{k+1})$, there is a unique one-sided minimizer $\gamma_{(x,0),v}$ originating at $(x,0)$ and the first Poissonian point it visits is $p_k$. 2. For every $k\in{{\mathbb Z}}$ there are at least two one-sided optimal paths originating at $x_k$, they pass through $p_k$ and $p_{k-1}$, respectively. 3. For any $k\in{{\mathbb Z}}$ any $x\in(x_k,x_{k+1})$ and any sequence $(y_n,t_n)\in {{\mathbb R}}^2$ with $t_n\to \infty$ and $$\lim_{n\to \infty} \frac{y_n}{t_n} = v,$$ and for every $x\in (x_k,x_{k+1})$, the minimizing paths $\gamma_{(x,0),(y_{n},t_{n})}$ converge to $\gamma_{(x,0),v}$. [[[Proof: ]{}]{}]{}The first two parts follow from Lemma \[lem:continuity-and-uniqueness-of-geodesics\]. In particular, that lemma shows that each $x\in{{\mathbb R}}$ has a neighborhood of points (excluding $x$) with a unique minimizer. The set of all $x\in{{\mathbb R}}$ such that $(x,0)$ has two or more one-sided minimizers must therefore be discrete (it has no accumulation points); this set will be $\{x_k:\ k\in {{\mathbb Z}}\}$. The proof of the last part repeats that of Lemma \[lem:convergence-of-geodesics\]. [[[$\Box$ ]{}]{}]{} Busemann functions and stationary solutions of the Burgers equation {#sec:global_solutions} =================================================================== In this Section we use the one-sided minimizers to construct global solutions of the Burgers equation, thus proving the existence part of Theorem \[thm:global\_solutions\]. Let us summarize some facts on one-sided backward minimizers that follow from Section \[sec:one-sided\_minimizers\]. For any velocity $v\in{{\mathbb R}}$, the following holds with probability 1. For every point $p=(x,t)$ there is a non-empty set $\Gamma_{v,p}$ of one-sided action minimizers $\gamma:(-\infty,t]\to{{\mathbb R}}$ with asymptotic slope $v$ $$\lim_{s\to-\infty}\frac{\gamma(s)}{s}=v,$$ ending at $p$. They all coalesce, i.e., they coincide on $(-\infty,t_{v,p}]$ for some $t_{v,p}<t$. For most points $p\in{{\mathbb R}}^2$, $\Gamma_{v,p}$ consists of a unique minimizer $\gamma_{v,p}$, but even if the uniqueness does not hold, there is the right-most minimizer $\gamma_{v,p}\in\Gamma_{v,p}$ such that $\gamma_{v,p}(s)\ge \gamma(s)$ for $s\le t$ and any other minimizer $\gamma\in\Gamma_{v,p}$. For every two points $p_1=(x_1,t_1)$ and $p_2=(x_2,t_2)$, all their one-sided minimizers coalesce, i.e., there is a time $t_v=t_v(p_1,p_2)$ such that $\gamma_{v,p_1}(s)=\gamma_{v,p_2}(s)$ for all $s\le t_v$. This allows us to define Busemann functions for slope $v$: $$B_v(p_1,p_2)= B_{v,\omega}(p_1,p_2)= A_\omega^{t_v(p_1,p_2),t_2}(\gamma_{v,p_2}) - A_\omega^{t_v(p_1,p_2),t_1}(\gamma_{v,p_1}),\ p_1,p_2\in{{\mathbb R}}^2.$$ Although $t_v$ is not defined uniquely, the definition clearly does not depend on a concrete choice of $t_v$ or $\gamma_{v,p_1},\gamma_{v,p_2}$. One can also choose $t_v$ to be the maximal of all possible coalescence times. Some properties of Busemann functions are summarized in the following lemma: \[lem:busemann\_properties\] Let $B_v$ be defined as above for $v\in{{\mathbb R}}$. 1. The distribution of $B_v$ is translation invariant: for any $\Delta \in{{\mathbb R}}^2$, $$B_v(\cdot+\Delta,\cdot+\Delta)\stackrel{distr}{=} B_v(\cdot,\cdot).$$ 2. $B_v$ is antisymmetric: $$B_v(p_1,p_2)=-B_v(p_2,p_1),\quad p_1,p_2\in{{\mathbb R}}^2,$$ in particular $B_v(p,p)=0$ for any $p\in{{\mathbb R}}^2$. 3. $B_v$ is additive: $$B_v(p_1,p_3)=B_v(p_1,p_2)+B_v(p_2,p_3),\quad p_1,p_2,p_3\in{{\mathbb R}}^2.$$ 4. For any $p_1,p_2\in{{\mathbb R}}^2$, ${{\mathbb E}}|B_v(p_1,p_2)|<\infty$. \[item:expectation\_finite\] [[[Proof: ]{}]{}]{}First three parts of the Lemma are straightforward. Let us prove part \[item:expectation\_finite\]. Using additivity and translation invariance of Busemann functions property, we see that it is sufficient to consider points $(0,0)$ and $(x,t)$ with $t< 0$. Since the effect of shear transformations on Busemann functions is easily computable, it is sufficient to assume that $v=0$. If $s<t<0$, and $x,y\in{{\mathbb R}}$ we have $$A^{s,0}(y,0)\le A^{s,t}(y,x)+\frac{x^2}{2|t|},$$ so $$B_0((0,0),(x,t))\ge - \frac{x^2}{2|t|},$$ and all we need is an upper bound on ${{\mathbb E}}B_0((0,0),(x,t))$. Since for any $s<t-1$ and any $y\in{{\mathbb R}}$ we have $$A^{s,t}(y,x) \le A^{s,t-1}(y,0)+\frac{x^2}{2},$$ it is sufficient to assume $x=0$. Also, if $t\in(-1,0)$, then $$A^{s,t}(y,0)\le A^{s,-1}(y,0),$$ so it is sufficient to consider $t\le -1$. If we prove finiteness of expectation for $t=-1$, then it will also follow for $t=-2,-3,\ldots$ by additivity, and for all intermediate times by $$A^{s,[t]+1}(y,0)\le A^{s,t}(y,0)\le A^{s,[t]}(y,0).$$ Therefore, it remains to prove that $$\label{eq:need_to_estimate_expectation_from_above} {{\mathbb E}}(A^{s,-1}(y,0)-A^{s,0}(y,0))<\infty.$$ where $(y,s)$ is the space-time point of coalescence of one-sided minimizers with zero asymptotic slope for the points $(0,0)$ and $(0,-1)$. Let $H=\{((x,t): t\le -1-|x|)\}$. Notice that $(0,-1)$ is the vertex of the right angle formed by $\partial H$. Since we are considering minimizers with zero asymptotic slope, we can define $$\tau=\inf\{t\le 0:\ (\gamma^*(t),t)\not\in H\},$$ where we introduced $\gamma^*=\gamma_{0,(0,0)}$ for brevity. Let us denote $z=\gamma^*(\tau)=\pm (\tau+1)$. If $s\in[\tau,0]$, then $$A^{s,0}(y,0)-A^{s,-1}(y,0)=A^{\tau,0}(z,0)-A^{\tau,-1}(z,0).$$ If $s<\tau$, then $$\begin{aligned} A^{s,-1}(y,0)-A^{s,0}(y,0) & \leq & A^{s,\tau}(y,z)+A^{\tau,-1}(z,0)-(A^{s,\tau}(y,z)+A^{\tau,0}(z,0))\\ & \le& A^{\tau,-1}(z,0)-A^{\tau,0}(z,0).\end{aligned}$$ Combining last two relations with , we see that we need to establish $$\label{eq:finiteness_of_expectation1} {{\mathbb E}}A^{\tau,0}(z,0)>-\infty$$ and $$\label{eq:finiteness_of_expectation2} {{\mathbb E}}A^{\tau,-1}(z,0)<\infty.$$ To prove  it is sufficient to show that $$\sum_{r=1}^\infty{{\mathbb P}}\left\{-A^{\tau,0}(z,0)>r\right\}<\infty.$$ Let us choose a small number ${{\varepsilon }}>0$ and write $$\label{eq:z-far-or-deviation-big} {{\mathbb P}}\left\{-A^{\tau,0}(z,0)>r\right\}\le {{\mathbb P}}\{|z|>{{\varepsilon }}r\}+{{\mathbb P}}\left\{|z|\le{{\varepsilon }}r;\ -A^{-1-|z|,0}(z,0) >r\right\}.$$ The second term can be estimated via Theorem \[thm:concentration\_around\_alphat\]. Estimating the action of the motion with unit speed from $(\pm {{\varepsilon }}r, -{{\varepsilon }}r-1)$ to $(z,\tau)$, we obtain that for $|z|\leq {{\varepsilon }}r$, $$A^{-1-{{\varepsilon }}r,0}(\pm {{\varepsilon }}r,0) \leq \frac12 (-|z|-1 - (-{{\varepsilon }}r -1)) + A^{-1-|z|,0}(z,0),$$ and therefore $$\begin{aligned} &{{\mathbb P}}\left\{|z|\le{{\varepsilon }}r;\ A^{-1-|z|,0}(z,0)<-r\right\} \\ \le & {{\mathbb P}}\left\{ A^{-1-{{\varepsilon }}r,0}({{\varepsilon }}r,0)<-r + \frac{{{\varepsilon }}r}{2} \right\}+ {{\mathbb P}}\left\{ A^{-1-{{\varepsilon }}r,0}(-{{\varepsilon }}r,0)<-r + \frac{{{\varepsilon }}r}{2} \right\}\\ = & 2 {{\mathbb P}}\left\{ A^{{{\varepsilon }}r+1} + \frac12 \frac{({{\varepsilon }}r)^2}{1+{{\varepsilon }}r} < -r + \frac{{{\varepsilon }}r}{2} \right\} \\ \le & 2 {{\mathbb P}}\left\{ A^{{{\varepsilon }}r+1}- \alpha(0)({{\varepsilon }}r+1) < -r + \frac{{{\varepsilon }}r}{2} - \alpha(0)({{\varepsilon }}r+1) \right\} \\ \le & 2 c_1({{\varepsilon }})\exp\left\{-c_2({{\varepsilon }})\frac{r^{1/2}}{\ln r}\right\},\end{aligned}$$ for $r>c_0({{\varepsilon }})$, where ${{\varepsilon }}$ is chosen so small that $1-{{\varepsilon }}/2 +{{\varepsilon }}\alpha(0)>0$. The right-hand side is summable in $r$. Let us now estimate ${{\mathbb P}}\{z>{{\varepsilon }}r\}$, one half of the first term in the r.h.s. of . Let $p_s=(s,-2s)$, $s>0$. Let us find $n_0>0$ such that ${{\rm Co}}(p_s,(2s)^{-\delta})$ does not intersect $\{(x,-3x):\ x>0\}$ for all $s>n_0$. Since the asymptotic slope of $\gamma^*$ is zero, it will eventually cross $\{(x,-3x):\ x>0\}$. Therefore, if $z>{{\varepsilon }}r>n_0$, then there must be $s>{{\varepsilon }}r$ and a point $q=(q_1,q_2)$ outside of the cone ${{\rm Co}}(p_s,(2s)^{-\delta})$ such that $\gamma^*$ connects $q$ to $(0,0)$ and passes through $p_s$. Using Corollary \[cor:delta-straightness-prob\], we see that if ${{\varepsilon }}r>n_0\wedge M$, $$\label{eq:tail_of_z} {{\mathbb P}}\{z>{{\varepsilon }}r\}\le c_1e^{-c_2({{\varepsilon }}r)^{\kappa}},$$ which is summable in $r$. This finishes the proof of . To prove , it is sufficient to notice that $${{\mathbb E}}A^{\tau,-1}(z,0)\le {{\mathbb E}}\frac{|z|}{2}$$ and apply . The proof of part \[item:expectation\_finite\] of Lemma \[lem:busemann\_properties\] is completed. [[[$\Box$ ]{}]{}]{} Having the Busemann function at hand, one can define $$U_v(x,t)=B((0,0),(x,t)),\quad (x,t)\in{{\mathbb R}}^2.$$ The main claim of this Section is that thus defined $U_v$ is skew invariant under of the HJHLO cocycle, and its space derivative is the global solution of the Burgers equation. Let us recall that the HJHLO evolution is given by $$\label{eq:Burgers_dynamics_on_potentials} \Phi^{s,t}W(y)=\inf_{x\in{{\mathbb R}}} \{W(x)+A^{s,t}(x,y)\}, \quad s\le t,\quad y\in{{\mathbb R}},$$ where $A^{s,t}(x,y)$ has been defined in . \[lem:global\_solution\_of\_HJ\] Function $U_v$ defined above is a global solution of the Hamilton–Jacobi equation. If $s\le t$, then $$\Phi^{s,t} U_v(\cdot,s) (x)=U_v(x,t).$$ [[[Proof: ]{}]{}]{}Let $\gamma_v$ be a minimizer through $(x,t)$ with slope $v$. Then $$\begin{aligned} U_v(x,t)=&U_v(\gamma_{v}(s),s)+(U_v(x,t)-U_v(\gamma_{v}(s),s))\\ =&U_v(\gamma_{v}(s),s) + A^{s,t}(\gamma_v(s),x).\end{aligned}$$ We need to show that the right-hand side is the infimum of $U_v(y,s)+A^{s,t}(y,x)$ over all $y\in{{\mathbb R}}$. Suppose that for some $y\in{{\mathbb R}}$, $$U_v(y,s) + A^{s,t}(y,x) < U_v(\gamma_{v}(s),s) + A^{s,t}(\gamma_v(s),x). \label{eq:suppose_not_infimum}$$ Let us take any minimizer $\bar \gamma_v$ originating at $(y,s)$ and denote by $\tau<s$ the time of coalescence of $\bar \gamma_v$ and $\gamma_v$. We claim that $$A^{\tau,s}(\bar \gamma_v)+A^{s,t}(y,x)< A^{\tau,t}(\gamma_v(\tau),x)= A^{\tau,t}(\gamma_v), \label{eq:gamma_bar_better}$$ which contradicts the minimizing property of $\gamma$. In fact, is a consequence of $$\begin{aligned} A^{\tau,s}(\bar \gamma_v)-A^{\tau,s}(\gamma_v)&=U_v(y,s) - U_v(\gamma_{v}(s),s)\\ &< A^{s,t}(\gamma_v(s),x) -A^{s,t}(y,x).\end{aligned}$$ where the second inequality follows from . [[[$\Box$ ]{}]{}]{} Another way to approach the Burgers equation is to consider, for $p=(x,t)$, $$u_v(x,t)=\dot\gamma_{v,p}(t).$$ Then $U_v(x,t)-U_v(0,t)=\int_0^x u_v(y,t)dy$. We recall that $\Psi^{s,t}w$ denotes the solution at time $t$ of the Burgers equation with initial condition $w$ imposed at time $s$. The function $u_v$ defined above is a global solution of the Burgers equation. If $s\le t$, then $$\Psi^{s,t} u_v (\cdot,s) = u_v(\cdot,t),\quad s\le t.$$ [[[Proof: ]{}]{}]{}This statement is a direct consequence of Lemmas \[lem:evolution\_on\_Burgers\_potentials\], \[lem:global\_solution\_of\_HJ\], and the definition of the Burgers cocycle $\Psi$.[[[$\Box$ ]{}]{}]{} The function $u_v(\cdot,t)$ is clearly piecewise linear with respect to the space coordinate, with downward jumps, each linear regime corresponding to the configuration point visited last by one-sided minimizers, see Lemma \[lem:domains-on-real-line\]. To prove that $U_v(\cdot,t)\in {{\mathbb H}}(v,v)$ for all $t$, we will compute the expectation of its spatial increments (we already know that it is well defined due to part \[item:expectation\_finite\] of Lemma \[lem:busemann\_properties\]), and prove that $u_v(\cdot,t)$ is mixing with respect to the spatial variable. \[lem:mean\_increment\] For any $(x,t)\in{{\mathbb R}}^2$, $${{\mathbb E}}(U_v(x+1,t)-U_v(x,t))={{\mathbb E}}B_v((x,t),(x+1,t))=v.$$ [[[Proof: ]{}]{}]{}First, we consider the case $v=0$. Due to the distributional invariance of Poisson process under reflections, $${{\mathbb E}}B_0((x,t),(x+1,t)) = {{\mathbb E}}B_0((x+1,t),(x,t)).$$ Combining this with the anti-symmetry of $B_0$, we obtain ${{\mathbb E}}B_0((x+1,t),(x,t))=0$, as required. In the general case, we can apply the shear transformation $L:(y,s)\mapsto (y+(t-s)v,s)=(y+vt-vs,s)$. Due to Lemma \[lem:shear\], the one-sided minimizers of slope $v$ will be mapped onto one-sided minimizers of slope $0$ for the new Poissonian configuration $L(\omega)$. We already know that $${{\mathbb E}}B_{0,L(\omega)}((x+1,t),(x,t))=0,$$ and a direct computation based on Lemma \[lem:shear\] gives $$B_{0,L( \omega)}((x,t),(x+1,t))=B_{v,\omega}((x,t),(x+1,t))+v,$$ and our statement follows since $L$ preserves the distribution of Poisson process. [[[$\Box$ ]{}]{}]{} \[lem:mixing\] Let $v\in{{\mathbb R}}$. For any $t$, the process $u_v(\cdot,t)$ is mixing. [[[Proof: ]{}]{}]{}Due to translation invariance it is sufficient to consider $t=0$. Using the shear invariance we can restrict ourselves to the case $v=0$. Notice that the values of $u_v(\cdot,0)$ on an interval are determined by the increments of $U_v(\cdot,0)$ on that interval. Therefore, in this proof we can work with these increments. Let us fix $a>0$ and ${{\varepsilon }}>0$. For any $h>2a$, we consider the processes $$L_0(x) = B_0((0,0),(x,0))=U_0(x,0)-U_0(0,0), \quad x\in [-a,a],$$ and $$L_h(x) = B_0((h,0),(h+x,0))=U_0(x+h,0)-U_0(h,0),\quad x\in [-a,a].$$ Next we define the processes $$L^t_0(x) = A_*^{-t,0}(0,0) - A_*^{-t,0}(0,x),\quad x\in [-a,a],$$ and $$L^t_h(x) = A_*^{-t,0}(h,h) - A_*^{-t,0}(h,x+h),\quad x\in [-a,a],$$ where $A_*$ is defined as the usual optimal action $A$, with the restriction that the path cannot cross the vertical line $\chi$ through $(h/2,0)$. This means that $L^t_0$ and $L^t_h$ are by definition independent, since the first depends on the Poisson process left of the vertical line, and $L^t_h$ depends on the Poisson process right of the vertical line. Note that if the one-sided minimizers for $(-a,0)$ and $(a,0)$ coalesce above $-t$, and if the minimizing paths connecting $(0,-t)$ to points $(-a,0)$ and $(a,0)$ coincide with the respective one-sided minimizers above their coalescing point and do not cross $\chi$, then $L_0(x)=L^t_0(x)$ for all $x\in [-a,a]$. By choosing $t$ large enough and then $h$ large enough, we can ensure that $${{\mathbb P}}\{\forall\ x\in[-a,a]\ :\ L_0(x)=L^t_0(x)\} \ge 1-{{\varepsilon }}.$$ Since our set-up is symmetric around $\chi$, we also have $${{\mathbb P}}\{\forall\ x\in[-a,a]\ :\ L_h(x)=L^t_h(x)\} \ge 1-{{\varepsilon }}.$$ Define $$C=C_{t,h} = \{ \forall\ x\in[-a,a]\ :\ L_0(x)=L^t_0(x)\ \mbox{and}\ L_h(x)=L^t_h(x)\}.$$ For events $E,F$ for the process $L_0$, we define $\tau_h(F)$ as the “translated” event for $L_h$, and $E^t$ and $\tau_h(F)^t$ as the corresponding events for $L^t_0$ and $L^t_h$. Then, $$\begin{aligned} |{{\mathbb P}}(E\cap \tau_h(F)) - {{\mathbb P}}(E){{\mathbb P}}(F)| &\leq & |{{\mathbb P}}(E\cap \tau_h(F)\cap C) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 2{{\varepsilon }}\\ & = & |{{\mathbb P}}(E^t\cap \tau_h(F)^t\cap C) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 2{{\varepsilon }}\\ & \leq & |{{\mathbb P}}(E^t\cap \tau_h(F)^t) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 4{{\varepsilon }}\\ & = & |{{\mathbb P}}(E^t){{\mathbb P}}(\tau_h(F)^t) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 4{{\varepsilon }}\\ & \leq & |{{\mathbb P}}(E^t\cap C){{\mathbb P}}(\tau_h(F)^t\cap C) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 8{{\varepsilon }}\\ & = & |{{\mathbb P}}(E\cap C){{\mathbb P}}(\tau_h(F)\cap C) - {{\mathbb P}}(E){{\mathbb P}}(F)| + 8{{\varepsilon }}\\ & \leq & 12{{\varepsilon }},\end{aligned}$$ and mixing follows due to the arbitrary choice of $\epsilon$. [[[$\Box$ ]{}]{}]{} Combining Lemmas \[lem:mean\_increment\] and \[lem:mixing\], we conclude that the Birkhoff space averages of $u_v$ have a well-defined, deterministic limit $v$, so $U_v\in {{\mathbb H}}(v,v)$. Stationary solutions: uniqueness and basins of attraction {#sec:attractor} ========================================================= In this section we prove Theorem \[thm:pullback\_attraction\] and the uniqueness part in Theorem \[thm:global\_solutions\] The key step in the proof of Theorem \[thm:pullback\_attraction\] is the following observation. \[lem:asymptotic\_slope\_in\_pullback\_attraction\] Let $t\in{{\mathbb R}}$ and suppose that an initial condition $W$ satisfies one of the conditions ,,. With probability one, the following holds true for every $y\in{{\mathbb R}}$. Let $y^*(s)$ be a solution of the optimization problem  . Then $$\lim_{s\to-\infty}\frac{y^*(s)}{s}=v.$$ [[[Proof of Theorem \[thm:pullback\_attraction\]: ]{}]{}]{} Let us take any rectangle $Q=[-R,R]\times [t_0,t_1]$ and set $t=t_1$. We use Lemma \[lem:domains-on-real-line\] to find points $a,b\in{{\mathbb R}}$ satisfying $a<-R<R<b$, not coinciding with any of the points $x_k,k\in{{\mathbb Z}}$ and such that one-sided backward minimizers $\gamma_{(a,t),v}$ and $\gamma_{(b,t),v}$ do not cross $Q$. Applying Lemma \[lem:asymptotic\_slope\_in\_pullback\_attraction\] to $x=a,b$, we see that the corresponding points $a^*(s)$ and $b^*(s)$ satisfy $a^*(s)/s\to v$ and $b^*(s)/s\to v$ as $s\to-\infty$. Let $p=(x_0,\tau_0)$ be the point of coalescence of the one-sided minimizers $\gamma_{(a,t),v}$ and $\gamma_{(b,t),v}$. We automatically have $\tau_0<t_0$. Lemma \[lem:domains-on-real-line\] then implies that there is $\tau_1<\min\{\tau_0,0\}$ such that for $s<\tau_1$, the restrictions of the finite minimizers connecting $(a^*(s),s)$ to $(a,t)$ and $(b^*(s),s)$ to $(b,t)$ on $[\tau_0,t]$ coincide with the restrictions of $\gamma_{(a,t),v}$ and $\gamma_{(b,t),v}$ (this also implies that we can choose $a^*(s)=b^*(s)$). Since $Q$ is trapped between $\gamma_{(a,t),v}$ and $\gamma_{(b,t),v}$, and minimizing paths cannot cross each other, we conclude that for any $s<\tau_1$, and any $(x,t)\in Q$, the minimizers connecting $(x^*(s),s)$ to $(x,t)$ (where $x^*$ is a solution of the optimization problem ) have to pass through $p$. In particular, the slopes of these minimizers determining the evolution of the Burgers velocity field in $[-R,R]$ throughout $[t_0,t_1]$ do not change (and coincide with the slopes of one-sided backward minimizers) as long as $s<\tau_1$, which completes the proof. [[[$\Box$ ]{}]{}]{} [[[Proof of Lemma \[lem:asymptotic\_slope\_in\_pullback\_attraction\]: ]{}]{}]{} We will only prove the sufficiency of condition . The proof of sufficiency of conditions and follows the same lines and we omit it. Let us also restrict ourselves to $t=0$ for simplicity. The proof does not change for other values of $t$. Since $y^*$ is increasing in $y$, it is sufficient to show that the conclusion of the lemma holds with probability 1 for fixed $y$. The stationarity of Poisson point field implies that we can assume $y=0$. We must show that for any ${{\varepsilon }}>0$ it is extremely unlikely for a path $\gamma$ with $\gamma(0)=0$ and $|\gamma(-r)|> {{\varepsilon }}r$ to be optimal if $r$ is large. For definiteness, let us work with paths satisfying $\gamma(-r)>{{\varepsilon }}r$. For any $\delta_0\in(0,-\alpha(0)/3)$ and for sufficiently large $r$, $$W(0) + A^{-r,0}(0,0)< (\alpha(0)+\delta_0)r.$$ Let us introduce $$Q_{ij}=[{{\varepsilon }}j +i,{{\varepsilon }}j +i+1]\times [-j-1,-j],\quad i,j\ge 0.$$ If $(x,-r)\in Q_{ij}$ and $W(x) + A^{-r,0}(x,0) < (\alpha(0)+\delta_0)r$, which would be necessary if $x=0^*(-r)$, then $$\inf_{z\in[{{\varepsilon }}j +i,{{\varepsilon }}j +i+1]} W(z)+ A^{-j-2,0}({{\varepsilon }}j +i+1,0)< (\alpha(0)+\delta_0)j+\frac{1}{2}.$$ The condition  at $+\infty$ implies that there is $j_0$ such that for $j>j_0$ and all $i$, $$\inf_{z\in[{{\varepsilon }}j +i,{{\varepsilon }}j +i+1]} W(z)> -(j+i) \delta_0.$$ so there is $j_1$ such that for $j>j_1$, $$A^{-j-2,0}({{\varepsilon }}j +i+1,0) < (\alpha(0) + 2\delta_0)j +\delta_0i + \frac{1}{2} < j\left(\alpha(0)+3\delta_0+\delta_0\frac{i}{j}\right).$$ Let us denote by $B_{ji}$ the event described by this inequality. Due to the Borel–Cantelli lemma, to show that with probability 1, events $B_{ji}$ can happen only for finitely many values of $j$, it suffices to show that for some $\beta>0$ and $c>0$, $$\label{eq:at_most_linear_growth_BC} \sum_{j\ge c}\sum_{i\le \beta j}{{\mathbb P}}(B_{ji})<\infty,$$ and $$\label{eq:super_linear_growth_BC} \sum_{j\ge c}\sum_{i>\beta j}{{\mathbb P}}(B_{ji})<\infty.$$ Denoting $\alpha_{ji}=\alpha\left(\frac{{{\varepsilon }}j+i+1}{j+2}\right)$, using shear and translation invariance, we obtain $${{\mathbb P}}(B_{ji}) = {{\mathbb P}}\left\{A^{j+2}-\alpha(0)(j+2) < j\left(\alpha(0)+3\delta_0+\delta_0 \frac{i}{j}- \frac{j+2}{j}\alpha_{ji}\right)\right\}.$$ If $\delta_0$ was chosen sufficiently small, then, using Lemma \[lem:shape-function\], we can find $j_2$ such that for all $j>j_2$ and all $i$, $$j\left(\alpha(0)+3\delta_0+\delta_0 \frac{i}{j}- \frac{j+2}{j}\alpha_{ji}\right) < -(j+2)\left(\frac{{{\varepsilon }}^2}{3}+\frac{i^2}{2j^2}\right),$$ so that $$\label{eq:B_ji} {{\mathbb P}}(B_{ji})\le {{\mathbb P}}\left\{A^{j+2}-\alpha(0)(j+2)<-(j+2)\left(\frac{{{\varepsilon }}^2}{3}+\frac{i^2}{3j^2}\right)\right\}.$$ Now  follows (with any $c\ge j_2$ and with arbitrary choice of $\beta$) from Theorem \[thm:concentration\_around\_alphat\]. To prove  we need an auxiliary lemma. In its statement and proof we use the notation introduced in Section \[sec:subadd\]. \[lem:tails\_of\_action\] There are constants $c_1,c_2,c_3,X_0,T_0>0$ such that for $t>T_0$, $x>X_0$, $${{\mathbb P}}\{A^t\le - xt\}\le c_2 e^{-c_3 xt}.$$ Notice that this lemma directly implies  and thus completes the proof of Lemma \[lem:asymptotic\_slope\_in\_pullback\_attraction\] if we choose $c>T_0$ and $\beta$ satisfying $$\frac{{{\varepsilon }}^2}{3}+\frac{\beta^2}{3}-\alpha(0)>X_0.$$ It remains to prove Lemma \[lem:tails\_of\_action\]. [[[Proof of Lemma \[lem:tails\_of\_action\]: ]{}]{}]{} It is sufficient to prove the lemma for $t\in{{\mathbb N}}$, although the values of constants may need adjustment for general $t$. Let us take $c_4>0$ and write $${{\mathbb P}}\{A^t\le -xt\}\le {{\mathbb P}}\{\#{{\mathcal A}}\le c_4 xt,\ A^t\le -xt\}+{{\mathbb P}}\{\#{{\mathcal A}}> c_4 xt\}. \label{eq:estimating_action_linear_tails}$$ To estimate the first term on the r.h.s., we can use  and derive that if $c_4$ is chosen small enough to ensure $y_0c_4<2$, where $y_0$ was introduced before condition , then for some constants $c_5,c_6>0$, any $x>1$, and sufficiently large $t$, $$\begin{aligned} {{\mathbb P}}\{\#{{\mathcal A}}\le c_4 xt,\ A^t\le -xt\}\le {{\mathbb P}}\{N_{\lceil c_4 xt\rceil}\ge xt\} \le c_5e^{-c_6 xt}.\end{aligned}$$ The second term on the r.h.s. of  can be estimated using Lemma \[lem:latani\]. If $c_4x\geq R$ and $t$ is sufficiently large, then $$\begin{aligned} {{\mathbb P}}\{\#{{\mathcal A}}> c_4 xt\}&\le \sum_{n\ge c_4xt} {{\mathbb P}}(E_{n,t}) \\ &\le \sum_{n\ge c_4xt} C_1\exp(-C_2n^2/t)\\ &\le C'_1\exp(-C'_2x^2t).\end{aligned}$$ for some constants $C'_1,C'_2>0$, which completes the proof.[[[$\Box$ ]{}]{}]{} [[[Proof of uniqueness in Theorem \[thm:global\_solutions\]: ]{}]{}]{} We will prove that any skew-invariant function $u$ with average velocity $v$ coincides with the global solution $u_v$ at time $0$. Let us take an arbitrary interval $I=(a,b)$. Lemma \[lem:asymptotic\_slope\_in\_pullback\_attraction\] implies that for any $W$ satisfying ${{\mathbb H}}(v,v)$, there is a time $T_0(a,b,W)\ge 0$ such that if $s<-T_0$, then there is a point $a^*\in{{\mathbb R}}$ that solves the optimization problem  for $t=0$ and for all points $y\in I$ at once, and the respective finite minimizers on $[s,0]$ have the same velocity at time 0 as the infinite one-sided minimizers of asymptotic slope $v$. Suppose now that $U_\omega(x,t)=U_{\theta^t\omega}(x)$ is a global solution in $\hat {{\mathbb H}}(v,v)$. Then $T_0(a,b,U_{\theta^t\omega}(\cdot))>0$ is a stationary process. In particular, this means that with probability 1, there is $R>0$ and a sequence of times $s_n\downarrow -\infty$ such that $T_0(a,b,U_{\theta^{s_n}\omega}(\cdot))<R$ for all $n\in{{\mathbb N}}$. Therefore, there is $n$ such that $s_n <- T_0(a,b,U_{\theta^{s_n}\omega}(\cdot))$. This and the fact that $U$ at time $0$ is the solution of problem   for $t=0$, $s=s_n$ and initial condition $W=U_{\theta^{s_n}\omega}(\cdot)$, we conclude that $U$ and the global solution $U_v$ coincide on $I$ at time $0$, and the proof is complete. [[[$\Box$ ]{}]{}]{} Basics on Burgers with Poissonian forcing {#sec:proofs_of_basic_burgers_facts} ========================================= [[[Proof of Lemma \[lem:full\_measure\_set\_where\_dynamics\_is\_defined\]: ]{}]{}]{}Let us take $N\in{{\mathbb N}}$, $x\in[-N,N]$, $W\in{{\mathbb H}}$, and $M\in{{\mathbb N}}$ satisfying $ W(y)\ge -M(|y|+1)$ for all $y\in {{\mathbb R}}$ and $|W(0)|<M$. Let us take any $m,n\in{{\mathbb Z}}$ satisfying $m<n$ and any $t,s\in{{\mathbb R}}$ satisfying $m<s<t<n$. Suppose now that a path $\gamma:[s,t]\to{{\mathbb R}}$ and a constant $L\in{{\mathbb N}}$ satisfy $\gamma(t)=x$ and $\sup_{r\in[s,t]}|\gamma(r)|\in[L,L+1)$. Let us compare $\gamma$ with the straight path $\gamma_0(r)=x(r-s)/(t-s)$, $r\in[s,t]$. Assuming that $$A_\omega^{s,t}(W,\gamma)\le A_\omega^{s,t}(W,\gamma_0),$$ we obtain $$\begin{aligned} M+\frac{x^2}{2(t-s)}&\ge W(0)+\frac{x^2}{2(t-s)}\\ &\ge A_\omega^{s,t}(W,\gamma_0)\\ &\ge A_\omega^{s,t}(W,\gamma)\\ &\ge -M(L+1+1)+\frac{(x-L)^2}{2(t-s)}-\omega([-(L+1),(L+1)]\times[s,t]).\end{aligned}$$ Therefore, for sufficiently large $L$, $$\omega([-(L+1),(L+1)]\times[m,n])\ge -M(L+3)+\frac{L^2-2Lx}{2(t-s)}\ge -M(L+3)+\frac{L^2-2LN}{2(n-m)}.$$ Since the r.h.s. is quadratic in $L$, and the l.h.s. grows linearly in $L$ with probability 1 due to the strong law of large numbers, we conclude that this inequality can be true only for finitely many values of $L\in{{\mathbb N}}$. Therefore, under the imposed restrictions on $x,W,s,t$, there is $L_0=L_0(\omega,m,n)<\infty$ such that the variational problem  will not change if supplied with an additional restriction $\sup_{r\in[s,t]}|\gamma(r)|\le L_0$. Since with probability 1 there are finitely many Poissonian points in $[-L_0,L_0]\times[m,n]$, it is useful to split the variational problem into two parts: optimization over paths that do not pass through any Poisson points and over paths passing through some configuration points. The existence of an optimal path among those not passing through any Poisson points is guaranteed by the theory of unforced Burgers equation $u_t+uu_x=0$. Also, there are only finitely many broken line paths between Poissonian points, and, for the same reason as above, for each Poisson point $p$ in $[-L_0,L_0]\times[s,t)$ there is an optimal path among paths not passing through any other Poisson points and terminating at $p$. Gluing these paths together, we see that the extremum in the optimization problem for paths containing Poissonian points is also attained. Combining the two cases leads to our claim holding a.s. for all $x,W,s,t$, satisfying constraints specified by integers $M,N,m,n$. The countable intersection of full measure sets over all $M,N,m,n$ still produces a full measure set. Its time invariance follows since if $L_0(\omega,m,n)$ is finite for all $m,n$, then $L_0(\theta^t\omega,m,n)$ is finite for all $m,n$ and all $t\in {{\mathbb R}}$. The proof is complete. [[[$\Box$ ]{}]{}]{} [[[Proof of Lemma \[lem:evolution\_on\_Burgers\_potentials\]: ]{}]{}]{} Part \[it:open-domains-of-influence\] follows from the observation that the action depends on the path continuously if the family of configuration points visited by the path is fixed. In the situation where $\gamma$ does not pass through any configuration points, Part \[it:two-ways-to-compute-velocity\] follows from the theory for unforced Burgers equation, see e.g., [@Lions-book:MR667669 Section 1.3]. If the minimizer $\gamma$ for $x$ has a straight line segment connecting a configuration point $p=(y,r)$ to $(x,t)$, then $$\frac{d}{dx}(\Phi_\omega^{s,t}W)(x)=\frac{d}{dx}\frac{(x-y)^2}{2(t-r)}=\frac{x-y}{t-r}=\dot\gamma(t), \label{eq:two-ways-to-compute-velocity}$$ and Part \[it:two-ways-to-compute-velocity\] is proven for points with unique minimizers. The case of the boundary points where the minimizers are not unique is considered similarly. Part \[it:piecewise\_linearity\] is a direct consequence of Part \[it:two-ways-to-compute-velocity\]. Part \[it:continuity\_of\_HJ\_solution\] follows from local boundedness of the slope of minimizers. For minimizers not passing through any configuration points, this is a consequence of the classical theory of unforced Burgers equation, and for minimizers passing through some configuration points it follows from . [[[$\Box$ ]{}]{}]{} [[[Proof of Lemma \[lem:Hamilton-Jacobi\]: ]{}]{}]{} Let us take a point $(x,t)\in ({{\mathbb R}}\times(s,\infty))\setminus\omega$ and a small neighborhood $O\subset({{\mathbb R}}\times ((s+t)/2,\infty))\setminus\omega$ of $(x,t)$. There is a time $s_0>s$ such that for all $(x',t')\in O$, no minimizers realizing $\Phi^{s,t'}W(x')$ pass through any configuration point after $s_0$. Lemma \[lem:cocycle\] implies now that the restriction of $U$ to $O$ coincides with the entropy solution of  with initial data $\Phi_\omega^{s,s_0}W$ imposed at time $s_0$. [[[$\Box$ ]{}]{}]{} [[[Proof of Lemma \[lem:invariant\_spaces\]: ]{}]{}]{} The Lipschitz property required in the definition of ${{\mathbb H}}$ follows from Lemma \[lem:evolution\_on\_Burgers\_potentials\], so we only have to control the behavior of $\Phi_\omega^{s,t}W(x)$ as $x\to\pm\infty$. It is easy to see that for our almost sure statement, it is sufficient to construct exceptional sets for integer times $s,t$ and the limits in definition of ${{\mathbb H}}$ and ${{\mathbb H}}(v_-,v_+)$ only along integer values of $x$. We will prove that if there is a constant $M$ such that if $W(x)>-M(|x|+1)$ for all $x\in{{\mathbb R}}$, then for any ${{\varepsilon }}>0$ and for sufficiently large integer values of $|n|$, $\Phi_\omega^{s,t}W(n)> -(M+{{\varepsilon }})|n|$. Suppose, for some $n$ this inequality is violated and for some $L\ge 0$, an optimal path $\gamma$ realizing $\Phi_\omega^{s,t}W(n)$ satisfies $\sup_{r\in[s,t]}|\gamma(r)-n|\in[L,L+1]$. Since $$-(M+{{\varepsilon }}))|n|\ge\Phi_\omega^{s,t}W(n) \ge -M(|n-L|+1)+\frac{L^2}{2(t-s)}-\omega([n-L,n+L]\times[s,t]),$$ we obtain $$\omega([n-L,n+L]\times[s,t])\ge {{\varepsilon }}|n|-M(L+1)+\frac{L^2}{2(t-s)}.$$ The l.h.s. has the Poisson distribution with mean $2L(t-s)$, and the Borel–Cantelli Lemma implies that this can hold only for finitely many values of $n$ and $L$, which proves the first part of the lemma. The proof of the second part follows the same lines, and we omit it. [[[$\Box$ ]{}]{}]{} [^1]: Yuri Bakhtin was supported by NSF CAREER Award DMS-0742424 and grant 040.11.264 from the Netherlands Organisation for Scientific Research (NWO). He is grateful for hospitality of the Fields Institute in Toronto, Delft Technical University, and CRM in Barcelona where parts of this work have been written. [^2]: Konstantin Khanin is supported by NSERC Discovery Grant RGPIN 328565

--- abstract: 'Let $Q$ be an affine quartic which does not intersect transversely with the line at infinity $L_{\infty}$. In this paper, we show the existence of a $(2,3)$ torus decomposition of the defining polynomial of $Q$ and its uniqueness except for one class.' author: - Masayuki Kawashima and Kenta Yoshizaki title: 'On $(2,3)$ torus decompositions of $QL$-configurations' --- Introduction {#s0 .unnumbered} ============ Let $C=\{f=0\}\subset \Bbb{C}^2$ be an irreducible affine plane curve and let $a, b$ be coprime positive integers with $a,b\ge 2$. We say that $C$ is [*a quasi torus curve of type $(a,b)$*]{} (c.f [@Ku-Albanese]) if there exist polynomials $f_r$, $f_p$ and $f_q$ such that they satisfy the following condition: $$\tag{$*$} f_r(x,y)^{ab}f(x,y)=f_p(x,y)^a+f_q(x,y)^b,\quad \deg f_j = j,\ \ j=r,p,q$$ where $r\ge 0$ and $p$, $q>0$. Under using affine equations, we also add conditions that any two polynomials of $f$, $f_r$, $f_p$ and $f_q$ are coprime. We say that such a decomposition $(*)$ is [*a quasi torus decomposition of $C$*]{}. A quasi torus curve $C$ is called [*torus curve*]{} if $f_r(x,y)$ is a non-zero constant. In [@Tokunaga-Kyoto], H. Tokunaga studied $D_{2p}$ covers of $\Bbb{P}^2$ branched along a quintic $Q+L_{\infty}$ where $Q$ is a quartic and $L_{\infty}$ is the line at infinity and their relative positions are the following: - $Q\cap L_{\infty}$ consists of two points. 1. $L_{\infty}$ is bi-tangent to $Q$ at two distinct smooth points. 2. $L_{\infty}$ is tangent to a smooth point and passes through a singular point of $Q$. 3. $L_{\infty}$ passes through two distinct singular points of $Q$. - $Q\cap L_{\infty}$ consists of a point. 1. $L_{\infty}$ is tangent to $Q$ at a smooth point with intersection multiplicity $4$. 2. $L_{\infty}$ intersects $Q$ at a singular point with intersection multiplicity $4$. Table 1 is the list of the possible configurations $Q+L_{\infty}$ which is given in [@Tokunaga-Kyoto]. No. ${{\rm{Sing}}}(Q)$ $Q\cap L_{\infty}$ No. ${{\rm{Sing}}}(Q)$  $Q\cap L_{\infty}$ ------ -------------------- -------------------- ------ ----------------------------- -------------------- -- -- (1) $2a_2$ (i) (11) $e_6$ (i) (2) $2a_2$ (iv) (12) $e_6$ (iv) (3) $2a_2+a_1$ (i) (13) $a_4+a_2^{\infty}$ (ii) (4) $2a_2+a_1$  (iv) (14) $a_3^{\infty}+a_2+a_1$ (ii) (5) $3a_2$ (i) (15) $a_5+a_1$ (i) (6) $a_2+a_3^{\infty}$ (ii) (16) $a_5^{\infty}+a_2^{\infty}$ (iii) (7) $a_5$ (i) (17) $2a_3^{\infty}$ (iii) (8) $a_5$ (iv) (18) $a_7^{\infty}$ (v) (9) $a_6^{\infty}$ (ii) (19) $2a_3^{\infty}+a_1$ (iii) (10) $a_4^{\infty}+a_2$ (v) \ Table 1.\ where the singularities of types $a_n$ and $e_6$ are defined by $$a_n:x^2+y^{n+1}=0\ (n\ge 1),\quad e_6:x^3+y^4=0$$ and the notation $*^\infty$ express singularities on the line $L_{\infty}$. We call such configurations [*$QL$-configurations*]{}. Note that $Q$ is irreducible for the cases $(1),\dots, (13)$ and $Q$ is not irreducible for the cases $(14),\dots, (19)$. We call configurations for the cases $(1),\dots, (13)$ (respectively for the cases $(14),\dots, (19)$) [*irreducible $QL$-configurations*]{} (resp. [*non-irreducible $QL$-configurations*]{}). In [@Y], the second author studied two topological invariants of $QL$-configurations, the fundamental group $\pi_1(\Bbb{C}^2\setminus Q)$ and the Alexander polynomial $\Delta_Q(t)$. In particular, he showed that $$\tag{$\star$} {\text{$t^2-t+1$ divides $\Delta_Q(t)$ except for the case $(13)$ and $(16)$. }}$$ In [@okatan], M. Oka studied a special type of degeneration family $\{C_\tau\}$ of irreducible torus sextics which degenerates into $C_0:=D+2L_{\infty}$ where $D$ is a quartic and $L_{\infty}$ is a line. We call such a degeneration [*[a line degeneration of order $2$]{}*]{}. (We will give the definition for general situation in §1). He showed the divisibility of the Alexander polynomials $\Delta_{C_{\tau}}(t)\,|\, \Delta_{D}(t)$ for $\tau \ne 0$. (Theorem 14 of [@okatan]) In this paper, we study the possibilities of quasi torus decompositions of $QL$-configurations so that the above divisibility $(\star)$ also follows from the results of the line degeneration by M. Oka. This paper consists of 9 sections. In section 1, we recall the definition of line degenerated torus curves of type $(p,q)$. In section 2, we classify the singularities of line degenerated torus curves of type $(2,3)$. In section 3, we state our main theorem. In sections 4, 5, 6 and 7, we prove the theorems which are stated in §3 using line degenerated torus curves of type $(2,3)$. For the cases $(14),\dots, (19)$, $Q$ is not irreducible but we can also consider torus decomposition of non-irreducible $QL$-configurations without irreducibility and the condition $\gcd(a,b)=1$ in the definition of quasi torus curves. In section 8, we consider torus decomposition of above non-irreducible $QL$-configurations. In section 9, we will show that if a plane curve has a $(2,3)$ torus decomposition, then there exist infinite $(2,3)$ quasi torus decompositions. Line degenerated torus curves {#s1} ============================= Let $U$ be an open neighborhood of $0$ in $\Bbb{C}$ and let $\{C_s\,|\,s\in U\}$ be an analytic family of irreducible curves of degree $d$ which degenerates into $C_0:=D+j\,L_{\infty}$ $(1\le j < d)$ where $D$ is an irreducible curve of degree $d-j$ and $L_{\infty}$ is a line. We assume that there is a point $B\in L_{\infty}\setminus L_{\infty}\cap D$ such that $B\in C_s$ and the multiplicity of $C_s$ at $P$ is $j$ for any non-zero $s \in U$. We call such a degeneration [*a line degeneration of order $j$*]{} and we call $L_{\infty}$ [*the limit line*]{} of the degeneration. $B$ is called [*the base point*]{} of the degeneration. In [@okatan], M. Oka showed that there exists a canonical surjection: $$\varphi:\pi_1(\Bbb{C}^2\setminus D)\to \pi_1(\Bbb{C}^2\setminus C_s),\quad s: {\text{sufficiently small,}}$$ where $\Bbb{C}^2=\Bbb{P}^2\setminus L_{\infty}$ and as a corollary he showed the divisibility among the Alexander polynomials of a line degeneration family: $$\Delta_{C_{s}}(t)\mid \Delta_{D_0}(t).$$ Line degenerated torus curves {#line-degenerated-torus-curves} ----------------------------- Let $C_{p,q}=\{F_{p,q}=0\}$ be a $(p,q)$ torus curve $(p > q\ge 2)$ where $F_{p,q}$ is defined by $$\tag{$1$} F_{p,q}(X,Y,Z)=F_p(X,Y,Z)^q+F_q(X,Y,Z)^p,\quad \deg F_k=k,\ k=p,q$$ where $(X,Y,Z)$ is a homogeneous coordinates system of $\Bbb{P}^2$. Suppose that $F_{p,q}$ is written by the following form: $$\tag{$2$} F_{p,q}(X,Y,Z)=Z^j\,G(X,Y,Z)$$ where $G(X,Y,Z)$ is a homogeneous polynomial of degree $pq-j$. We call a curve $D=\{G=0\}$ [*a line degenerated torus curve of type $(p,q)$ of order $j$*]{} and the line $L_{\infty}=\{Z=0\}$ [*the limit line of the degeneration*]{}. We divide the situations $(2)$ into two cases. [**[First case.]{}**]{} Suppose that the defining polynomials of associated curves are written as follows: $$F_p(X,Y,Z)= F_{p-r}'(X,Y,Z)Z^r,\quad F_q(X,Y,Z)=F_{q-s}'(X,Y,Z)Z^s$$ where $r$ and $s$ are positive integers such that $r < p$ and $s < q$. We assume that $sp\ge rq$. Factoring $F_{p,q}$ as $F_{p,q}(X,Y,Z)=Z^{rq}G(X,Y,Z)$, we can see that $G$ is defined as $$\tag{$3$} G(X,Y,Z)=F_{p-r}'(X,Y,Z)^q+F_{q-s}'(X,Y,Z)^pZ^{sp-rq}.$$ We call such a factorization [*visible factorization*]{} and $D$ is called [*a visible degeneration of torus curve of type $(p,q)$*]{}. By the definition, $D\cap L_{\infty}=\{F_{p-r}'(X,Y,Z)=Z=0\}$ and thus the limit line $L_{\infty}$ is singular with respect to the visible degeneration of torus curve $D$. In [@okatan], M. Oka showed that a visible degeneration of torus curve of type $(p,q)$ can be expressed as a line degeneration of irreducible torus curves of degree $pq$. [**[Second case.]{}**]{} Neither $F_p$ or $F_q$ factors through $Z$ but $F$ can be written as $(2)$. Then $D$ is called [*an invisible degeneration of torus curve of type $(p,q)$*]{}. Line degenerated $(2,3)$\[s2\] torus curves of degree $4$ ========================================================= In this section, we consider a $(2,3)$ sextic of torus type which is a visible factorization. Visible factorization --------------------- Let $D=\{G=0\}$ be a quartic associated with a visible factorization $(3)$: $$D:\quad G(X,Y,Z)=F_{2}'(X,Y,Z)^2+F_{1}'(X,Y,Z)^3Z=0.$$ We consider the associated curves $C_2:=\{F_2'=0\}$, $L:=\{F_1'=0\}$ and $L_{\infty}:=\{Z=0\}$. Let $P$ be an inner singularity of $D$, namely $P$ is on the intersection $C_2\cap L$, $C_2\cap L_{\infty}$ or $C_2\cap L\cap L_{\infty}$. Then the topological type $(D,P)$ depends only on the intersection multiplicities of $C_2$, $L$ and $L_{\infty}$. To describe singularities of $D$, we put the intersection multiplicities $\iota_1:=I(C_2,L;P)$ and $\iota_2:=I(C_2,L_{\infty};P)$. Note that $0\le \iota_i \le 2$ for $i=1,2$ and $(\iota_1,\iota_2)\ne (2,2)$ as $L\ne L_{\infty}$. \[l1\] Suppose that $C_2$ is smooth at $P$. Then we have the following descriptions: 1. If $P\in C_2\cap L\setminus L_{\infty}$, then we have $(D,P)\sim a_{3\iota_1-1}$. 2. If $P\in C_2\cap L_{\infty}\setminus L$, then $D$ is smooth at $P$ and is tangent to $L_{\infty}$ with $I(D,L_{\infty};P)=2\iota_2$. 3. If $P\in C_2\cap L\cap L_{\infty}$, then we have $(D,P)\sim a_{3\iota_1+\iota_2-1}$. The assertion $(1)$ is showed in [@Pho] and [@BenoitTu] for general cases. We consider the case (2) and (3). We use affine coordinates $(x,z)=(X/Y,Z/Y)$ on $\Bbb{C}^2=\Bbb{P}^2\setminus \{Y=0\}$. Then the defining polynomial $g$ of $D$ in the affine coordinates is given as follows. $$g(x,z):=G(x,1,z)=f_2'(x,z)^2+f_1'(x,z)^3z,\quad f_j'(x,z):=F_j'(x,1,z),\,j=1,2.$$ As $C_2$ is smooth at $P$, we can take local coordinates $(u,v)$ so that $L_{\infty}=\{v=0\}$ and $f'_2(u,v)=c_1\,v-\varphi(u)$ where $c_1\ne 0$. Then ${{\rm{ord}}}_u \varphi(u)=\iota_2$ and $$\begin{split} g(u,v)&=(c_1 v-\varphi(u))^2+f'_1(u,v)^3v\\ &=c\,u^{2\iota_2}+f'_1(P)^3\,v+{\text{(higher terms)}}, \ c\ne 0. \end{split}$$ Here “higher terms” are linear combinations of monomials $u^{\alpha}v^{\beta}$ such that $2\iota_2\beta+\alpha > 2\iota_2$. They do not affect the topology of $D$ at $P$. As $P$ is not on $L$, $f'_1(P)$ is not $0$. Hence we have the assertion (2). Now we show the assertion (3). As $C_2$ is smooth at $P$, we can take local coordinates $(u,v)$ so that $f'_2(u,v)=v$, $f'_1(v,v)=c_1\,v-\varphi_1(u)$ and $L_{\infty}=\{c_2\,v-\varphi_2(u)=0\}$ where $c_1$ and $c_2$ are non-zero constant. Then ${{\rm{ord}}}_u \varphi_1(u)=\iota_1$ and ${{\rm{ord}}}_u \varphi_2(u)=\iota_2$ and $$\begin{split} g(u,v)&=v^2+(c_1\,v-\varphi_1(u))^3(c_2\, v-\varphi_2(u))\\ &=v^2-c\,u^{3\iota_1+\iota_2}+{\text{(higher terms)}},\ c\ne 0. \end{split}$$ Here “higher terms” are linear combinations of monomials $u^{\alpha}v^{\beta}$ such that $2\alpha+(3\iota_1+\iota_2)\beta > 2(3\iota_1+\iota_2)$ if $\gcd(2,3\iota_1+\iota_2)=1$ and $\alpha+(3\iota_1+\iota_2)\beta/2 > (3\iota_1+\iota_2)$ if $\gcd(2,3\iota_1+\iota_2)=2$. In particular, $v\varphi_1(u)^3$ is in [[(higher terms)]{}]{}. This shows the assertion (3). Next we consider the case that $C_2$ is singular at $P$. Then $C_2$ consists of two lines $\ell_1$ and $\ell_2$ such that $\ell_1\cap \ell_2=\{P\}$. \[l2\] Suppose that $C_2$ is singular at $P$. Then singularities of $D$ at $P$ are described as follows: 1. If $P\in C_2 \cap L\setminus L_{\infty}$, then we have $(D,P)\sim e_6$. 2. If $P\in C_2 \cap L_{\infty} \setminus L$, then $D$ is smooth at $P$ and is tangent to $L_{\infty}$ with $I(D,L_{\infty};P)=4$. 3. If $P\in C_2 \cap L \cap L_{\infty}$, then $D$ consists of four lines which intersect at $P$. The assertion $(1)$ is showed in [@Pho]. We show the assertion (2). We take a suitable local coordinates $(u,v)$ at $P$ so that $f_2(u,v)=u(b_1u-b_2v)$ and $L_{\infty}=\{v=0\}$ where $b_i\ne 0$ for $i=1,2$. Then $$\begin{split} g(u,v)&=u^2(b_1u-b_2v)^2+b_3\,f'_1(P)^3v\\ &=b_1^2u^4+b_3\,f'_1(P)^3v+{\text{(higher terms)}}. \end{split}$$ As $P$ in not on $L$, we have $f'_1(P)\ne 0$ and $I(D,L_{\infty};P)=4$. This shows the assertion (2). The assertion $(3)$ is obvious form the defining polynomial of $D$. We say that $P$ is [*an outer singularity of $D$*]{} if $P\in {{\rm{Sing}}}(D)\setminus C_2$. We consider possible outer singularities of $D$. \[l3\] If $P\in D$ is an outer singularity, then $(D,P)$ is either $a_1$ or $a_{2}$. Our proof is computational and it is done in the same way as in [@Oka-Pho2]. Invisible factorization ----------------------- Let $D=\{G=0\}$ be an invisible factorization of a $(2,3)$ torus curve which satisfies the following equations. $$\tag{$2$} F_{2,3}(X,Y,Z)= F_2(X,Y,Z)^3-F_3(X,Y,Z)^2=Z^2\,G(X,Y,Z)$$ We assume that $G$ is not divided by $Z$ and $G$ is reduced. We rewrite $F_2$ and $F_3$ as follows: $$\begin{split} F_2(X,Y,Z)&=F_2^{(2)}(X,Y)+F_2^{(1)}(X,Y)Z+F_2^{(0)}(X,Y)Z^2,\\ F_3(X,Y,Z)&=F_3^{(3)}(X,Y)+F_3^{(2)}(X,Y)Z+F_3^{(1)}(X,Y)Z^2 +F_3^{(0)}(X,Y)Z^3 \end{split}$$ where $F_j^{(i)}$ is a homogeneous polynomial of degree $i$. By an easy calculation, we observe that there exists a linear form $\ell_1(X,Y)$ so that $$\begin{cases} F_2^{(2)}(X,Y)=\ell_1(X,Y)^2,\\ F_3^{(3)}(X,Y)=\varepsilon\,\ell_1(X,Y)^3,\quad F_3^{(2)}(X,Y)=\dfrac{3\varepsilon}{2}\ell_1(X,Y)F_2^{(1)}(X,Y) \end{cases}$$ where $\varepsilon=1$ or $-1$. We put $\ell_2(X,Y):=F_2^{(1)}(X,Y)$ and $\ell_3(X,Y):=F_3^{(1)}(X,Y)$. Then we may assume the defining polynomials of $C_2$ and $C_3$ as the following: $$\begin{aligned} (\sharp)\quad \begin{cases} \begin{split} F_2(X,Y,Z)&=\ell_1(X,Y)^2+\ell_2(X,Y)\,Z+a_{00}\,Z^2,\\ F_3(X,Y,Z)&=\ell_1(X,Y)^3+\frac{3}{2}\,\ell_1(X,Y)\,\ell_2(X,Y)\,Z +\ell_3(X,Y)\,Z^2+b_{00}\,Z^3. \end{split} \end{cases}\end{aligned}$$ Then $F_{2,3}$ is factorized as $$\begin{split} F_{2,3}(X,Y,Z)=F_2(X,Y,Z)^3-F_3(X,Y,Z)^2=Z^2G(X,Y,Z). \end{split}$$ To see the local geometry of $D$ at a intersection point $D$ and $L_{\infty}$, we may assume $\ell_1(X,Y)=X$ and we take the affine coordinates $(x,z)=(X/Y,Z/Y)$ at $O^*:=[0:1:0]$. Let $g(x,z)=G(x,1,z)$, $f_2(x,z)=F_2(x,1,z)$ and $f_3(x,z)=F_3(x,1,z)$ be the local equations of $D$, $C_2$ and $C_3$ respectively. In the affine coordinates $(x,z)$, $f_2$ and $f_3$ are written as $$\begin{split} f_2(x,z)&=x^2+\ell_2(x,1)\,z+a_{00}z^2,\\ f_3(x,z)&=x^3+\frac{3}{2}\ell_2(x,1)\,xz+\ell_3(x,1)z^2+b_{00}z^3. \end{split}$$ We can see the local geometries of $C_2$ and $C_3$ at $O^*$. First we consider the case $\ell_2(0,1)\ne 0$. Then we have 1. $C_2$ is smooth at $O^*$ and is tangent to the limit line $L_{\infty}$ at $O^*$. 2. $C_3$ has an $a_1$ singularity at $O^*$. 3. The intersection multiplicity $I(C_2,C_3;O^*)$ is $3$. Then, putting $c_1=\ell_2(0,1)$, $g(x,z)$ is given as $$g(x,z)=c_1^3\,z+\frac{3}{4}\,c_1^2\,x^2+{\text{(higher terms)}}.$$ Thus $D$ is simply tangent to $L_{\infty}$ at $O^*$. We write $g(x,0)=x^2\,(x-\alpha)(x-\beta)$ for some $\alpha,\beta$ such that $\alpha \beta=3c_1^2/4\ne 0$. Then if $\alpha\ne \beta$, then $L_{\infty}$ is tangent to $D$ at $O^*$ and intersects transversely with $D$ at other $2$ points. If $\alpha=\beta$, then $L_{\infty}$ is a bi-tangent line of $D$. \[l4\] If $c_1\ne 0$, then the set of singularities ${{\rm{Sing}}}(D)$ is $\{3a_2\}$ or $\{a_2+a_5\}$. As the intersection $C_2\cap C_3\cap L_{\infty}=\{O^*\}$ and $I(C_2,C_3;O^*)=3$, the sum of the intersection numbers of $C_2\cap C_3$ is $3$ in the affine space $\Bbb{P}^2\setminus L_{\infty}$. The possible configurations of ${{\rm{Sing}}}(D)$ are $\{3a_2\}$, $\{a_5+a_2\}$ and $\{a_8\}$. The singularity $a_8$ is locally irreducible but the Milnor number of an irreducible quartics is less then or equal to $6$. Hence the configuration $\{a_8\}$ does not occur. [[ If $D$ is bi-tangent to $L_{\infty}$, then the configuration $\{a_5+a_2\}$ does not exist. Indeed, if the configuration $\{a_5+a_2\}$ exists, then $D$ can not be irreducible. If $D$ is a union of a line and a cubic, then a cubic can not have a bi-tangent line. If $D$ is a union of two conics which are tangent to $L_{\infty}$, then $D$ can not have any $a_2$ singularity. ]{}]{} Now we consider the case $c_1=\ell_2(0,1)=0$. Then putting $c_2=\ell_3(0,1)$, their defining polynomials are given as $$\begin{split} f_2(x,z)&=a_{00}\,z^2+\ell_2(1,0)\,xz+x^2,\\ f_3(x,z)&=c_2\,z^2+x^3 +\left(c_3\,x^2z +c_4\,xz^2+b_{00}z^3\right),\ \ c_3,c_4\in \Bbb{C}. \end{split}$$ Thus $C_2$ consists of two lines $\ell_1$ and $\ell_2$ such that $\ell_1\cap \ell_2=\{O^*\}$ and $C_3$ has an $a_2$ singularity at $O^*$ and $I(C_2,C_3;O^*)=4$. Then after an easy calculation, we have $$g(x,z)=-c_2^2\,z^2-2\,c_2\,x^3$$ Hence $D$ has an $a_2$ singularity at $O^*$. Thus we have: \[l5\] If $c_1=0$, then $D$ has an $a_2$ singularity on $L_{\infty}$ and ${{\rm{Sing}}}(D)=\{2a_2+a_2^{\infty}\}$ or $\{a_5+a_2^{\infty}\}$. As $C_2\cap C_3\cap L_{\infty}=\{O^*\}$ and $I(C_2,C_3;O^*)=4$, the intersection $C_2\cap C_3$ generically consists of two points in the affine space $\Bbb{P}^2\setminus L_{\infty}$. By a similar argument of Lemma $\ref{l4}$, we have the assertion. Statement of the Theorem. {#s3} ========================= Let $Q$ be a quartic in $QL$-configurations. For a quartic $Q$ in one of the $(1),\dots, (13)$ of Table $1$, $Q$ is irreducible and $Q$ is not irreducible for $(14),\dots, (19)$ of Table $1$. Now our main results are the followings: Let $Q$ be an irreducible quartic in one of the $QL$-configurations. 1. For $Q$ in the case $(13)$, there exists no $(2,3)$ torus decomposition. 2. For $Q$ in the case $(5)$, there exist five torus decompositions of type $(2,3)$ whose three decompositions are visible decompositions and two are invisible decompositions. 3. For $Q$ in the remaining cases, there exists a unique $(2,3)$ torus decomposition for each case. Note that a quartic $Q$ for the case $(5)$ is a $3$ cuspidal quartic. [[ In section 9, we will show that if a plane curve has a $(2,3)$ torus decomposition, then there exist infinite $(2,3)$ quasi torus decompositions. ]{}]{} For each quartic $Q$ of $(1),\dots, (12)$, there exists a line degeneration family of sextic $C(s):H_3(X,Y,Z,s)^2+H_2(X,Y,Z,s)^3=0$ which are $(2,3)$ torus curves such that $C(0)=Q+2L_{\infty}$. In particular, we have the divisibility $\Delta_{C(s)}(t)\mid \Delta_{Q}(t)$. The divisibility $(\star)$ in Introduction also follows from Theorem 2 and Corollary 15 of [@okatan]. \[p2\] For non-irreducible quartics $(14),\dots, (19)$ in Table $1$, we have the following: 1. There exist unique $(2,3)$ torus decompositions for quartics $(14)$ and $(15)$ and their decompositions are represented as visible decompositions. 2. The quartic $(16)$ does not admit any torus decompositions. 3. There exist unique $(2,4)$ torus decompositions for the quartics $(17)$, $(18)$ and $(19)$. Their decompositions are represented as invisible decompositions. [[ For the quartics $(13)$ and $(16)$, there are not torus decomposition. By the classifications of singularities in §2, their singularities do not occur as the quartics with visible or invisible $(2,3)$ torus decompositions. ]{}]{} The proof of Theorem 1 of $(3)$ {#s4} =============================== Strategy -------- There are $13$ configurations of singularities of the quintic $Q+ L_{\infty}$ as in $(1)$, $\dots$, $(13)$ in Table 1. We divide these quintic into $5$ cases ${\rm{(i)}},\dots, {\rm{(v)}}$ as in Introduction. Note that the case ${\rm{(iii)}}$ does not appear when $Q$ is irreducible. By the classification of the singularities for the visible and invisible factorizations in §2, for the quartics $(1),\dots, (12)$ except the case $(5)$, the possible torus decomposition must be visible and unique if it exists. The quartic $(5)$ has an exceptional property. It has both visible and invisible torus decompositions. Thus we treat this case in the next section. First, we construct explicit quartics $Q:=\{F=0\}$ with the prescribed properties at infinity. By the action of ${\rm{PSL}}(3,\Bbb{C})$ of $\Bbb{P}^2$, we can put the singularities at fixed locations. Then we construct the respective torus decompositions in §2.\ [**[Step 1. Construction of an explicit quartic $Q$.]{}**]{} By the classification of the singularities for invisible decomposition case (Lemma 4 and Lemma 5), the quartics in cases (1)$\sim$ (12) except the case (5) can not have invisible torus decomposition. So we only need the possible visible decomposition for these quartics. As the computations are boring and easy, we explain the quartic $(1)$ in Table 1 in detail and for the other cases we simply give the result of the computations. [**[The quartic $(1)$ in Table 1.]{}**]{} In this case, $L_{\infty}$ is a bi-tangent line of $Q$ and the singularity is ${{\rm{Sing}}}(Q)=\{2a_2\}$. We construct a quartic $Q$ with $2a_2$ which $L_{\infty}$ is a bi-tangent line. Let $\Sigma(Q):=\{P_1,P_2\}$ be the singular locus of $Q$ and let $Q\cap L_{\infty}:=\{R_1,R_2\}$ be the bi-tangent points. By the action of $\rm{PSL}(3,\Bbb{C})$ on $\Bbb{P}^2$, we can put the locations of points: $$P_1=[1:0:1],\quad P_2=[-1:0:1],\quad R_1=[1:1:0]$$ and we may assume that the tangent directions at $P_1$ and $P_2$ are given as $$\{x-1=0\},\quad \{x+1=0\} {\text{ \ respectively. }}$$ We start from the generic quartic $F(X,Y,Z)=\sum_{\nu}c_\nu X^{\nu_1}Y^{\nu_2}Z^{\nu_3}$ with $\nu=(\nu_1,\nu_2,\nu_3)$ with $\nu_1+\nu_2+\nu_3=4$. The necessary conditions are $$\begin{split} &F(P_1)=\frac{\partial F}{\partial X}(P_1) =\frac{\partial F}{\partial Y}(P_1) =\frac{\partial F}{\partial YY}(P_1) =\frac{\partial F}{\partial XY}(P_1) =0,\\ &F(P_2)=\frac{\partial F}{\partial X}(P_2) =\frac{\partial F}{\partial Y}(P_2) =\frac{\partial F}{\partial YY}(P_2) =\frac{\partial F}{\partial XY}(P_2) =0,\\ &F(R_1)=\frac{\partial F}{\partial X}(R_1)=0. \end{split}$$ Under the above conditions, we have $F(X,Y,0)=(X-Y)^2(X-\alpha\,Y)(X-\beta\, Y)$. As $L_{\infty}$ is bi-tangent to $Q$, we must have $\alpha=\beta$. Thus we have $13$ equations of the coefficients of $F$. By solving these equations, $F$ has the following form: $$Q:\quad F(X,Y,Z)=Z^4+2(Y^2-X^2)Z^2+t\,Y^3Z+(Y^2-X^2)^2=0,\quad t\ne 0.$$ Then another bi-tangent point $R_2$ is $[-1:1:0]$.\ [**[Step 2. Torus decompositions.]{}**]{} Now we consider the possibilities of visible torus decompositions of $Q$. Thus we assume that $F$ is written as follows: $$F(X,Y,Z)=F_2'(X,Y,Z)^2+F_1'(X,Y,Z)^3Z.$$ Two $a_2$ singularities are inner singularities of $Q$. Hence we assume that $C_2\cap L=\{P_1,P_2\}$ and $C_2$ is smooth at $P_1$ and $P_2$ where $C_2=\{F_2'=0\}$ and $L=\{F_1'=0\}$. Then we have $$\begin{split} & F_1'(P_1)=F_1'(P_2)=0,\qquad F_2'(P_1)=F_2'(P_2)=0. \end{split}$$ Then $L$ is the line pass through $P_1$ and $P_2$. Hence we get $F_1'(X,Y,Z)=s_1\,Y$ where $s_1\in {\Bbb C^*}$. As $C_2$ is smooth at $P_i$, the tangent directions of $C_2$ at $P_i$ must coincide with that of $Q$ for $i=1,2$. Hence $F_2'$ also satisfies the following conditions: $$\dfrac{\partial F_2'}{\partial Y}(P_1)= \dfrac{\partial F_2'}{\partial Y}(P_2)=0.$$ Then ${{\rm{Sing}}}(Q)=\{2a_2\}$ by Lemma \[l1\]. As $R_1\in Q$, $C_2$ passes through $R_1$. Namely $F_2'$ satisfies the condition $F_2'(R_1)=0$. Then $F_2'$ takes the following form: $$F_2'(X,Y,Z)=s_2\,(X^2-Y^2-Z^2),\ \ s_2\in \Bbb{C}^*.$$ Note that $C_2$ also passes through another bi-tangent point $R_2$. Hence $Q$ satisfies the condition (i) by Lemma \[l1\]. Therefore we get the family of quartics with visible factorizations: $$F(X,Y,Z)=s_2^2\,(X^2-Y^2-Z^2)^2+s_1^3\, Y^3\,Z=0.$$ Finally, we put $s_1^3=t$ and $s_2^2=1$. Then we can see easily $$\begin{split} (X^2-Y^2-Z^2)^2+t\,Y^3Z &=Z^4-2 \left(X^2+Y^2 \right)Z^2+t\,Y^3Z+ (X^2+Y^2)^2\\ &=F(X,Y,Z). \end{split}$$ [**[Step 3. Uniqueness.]{}**]{} By the classification of the singularities for the visible and invisible factorizations in §2, we can see easily that the possible torus decompositions are visible. Then two $a_2$ singularities must be inner singularities and the corresponding curves are uniquely determined by the above arguments. ![The quartic (1) in Table 1[]{data-label="s1"}](./s1.eps) For the other quartics, we only give the results of calculations. [|c|c|]{}\ Singularities & $2a_2$ at $[1:0:1]$, $[-1:0:1]$\ $Q\cap L_\infty$ & bi-tangent at $[1:1:0]$, $[-1:1:0]$\ Torus decomposition & $(X^2-Y^2-Z^2)^2+t\,Y^3Z=0$\ \ Singularities & $2a_2$ at $[1:0:1]$, $[-1:0:1]$\ $Q\cap L_\infty$ & tangent multiplicity 4 at $[0:1:0]$.\ Torus decomposition & $(X^2-Z^2)^2+t\,Y^3Z=0$.\ \ Singularities & $2a_2+a_1$ at $[0:0:1]$, $[1:1:1]$, $[-1:1:0]$.\ $Q\cap L_\infty$ & bi-tangent at $[\sqrt{3}:1:0]$, $[-\sqrt{3}:1:0]$.\ Torus decomposition & $(4Z^2-6YZ-X^2+3Y^2)^2+16\,(Y-Z)Z=0$. \ \ Singularities & $2a_2+a_1$ at $[0:0:1]$, $[1:1:1]$, $[-1:1:0]$.\ $Q\cap L_\infty$ & tangent multiplicity 4 at $[0:1:0]$\ Torus decomposition & $(2Z^2+X^2-3YZ)^2+4\,(Y-Z)Z=0$. \ \ Singularities & $a_2+a_3^{\infty}$ at $[0:0:1]$, $[-1:1:0]$.\ $Q\cap L_\infty$ & singular at $[-1:1:0]$ and tangent at $[1:1:0]$.\ Torus decomposition & $(X^2-2\,XZ-Y^2)^2+t\,(X+Y)Z=0$. \ \ Singularities & $a_5$ at $[0:0:1]$.\ $Q\cap L_\infty$ & bi-tangent at $[1:1:0]$, $[-1:1:0]$.\ Torus decomposition & $(XZ+s(X^2-Y^2))^2+s(ts-2)\,X^3Z=0$. \ \ Singularities & $a_5$ at $[0:0:1]$.\ $Q\cap L_\infty$ & tangent multiplicity 4 at $[1:0:0]$.\ Torus decomposition & $(XZ-s\,Y^2)^2+t\,X^3Z=0$\ \ Singularities & $a_6^{\infty}$ at $[0:1:0]$.\ $Q\cap L_\infty$ & singular at $[0:1:0]$ and tangent at $[-1:1:0]$.\ Torus decomposition & $(t_2^2Z^2+t_3XZ-t_2X(X+Y))^2$\  & $+t_2(2t_3+t_1t_2)\,X^3Z=0$.\ \ Singularities & $a_4^{\infty},a_2$ at $[0:0:1]$, $[0:1:0]$.\ $Q\cap L_\infty$ & singular at $[0:1:0]$.\ Torus decomposition & $(YZ-t_2X^2)^2+t\,X^3Z=0$\ \ Singularities & $e_6$ at $[0:0:1]$.\ $Q\cap L_\infty$ & bi-tangent at $[1:1:0]$, $[-1:1:0]$.\ Torus decomposition & $(X^2+Y^2)^2+t\,X^3Z=0$\ \ Singularities & $e_6$ at $[0:0:1]$.\ $Q\cap L_\infty$ & tangent multiplicity 4 at $[1:0:0]$.\ Torus decomposition & $Y^4+t\,X^3Z=0$.\ Proof of Theorem 1 of $(2)$. {#s5} ============================ In this section, we consider the exceptional case (5) in Table $1$. This case (5) is exceptional as the classification of the singularities tell us that it may have a invisible case as well as visible decompositions. Let $Q=\{F=0\}$ be a quartic with ${{\rm{Sing}}}(Q)=\{3a_2\}$. Then, by the class formula ([@Okadual]), $Q$ has a unique bi-tangent line and we take $L_{\infty}$ as the bi-tangent line of $Q$. We put the singular locus $\Sigma(Q)=\{P_1,P_2,P_3\}$ and the intersection $Q \cap L_{\infty}:=\{R_1, R_2\}$. By the action of $\rm{PSL}(3,\Bbb{C})$ on $\Bbb{P}^2$, we can put the locations: $$P_1=[1:0:1],\quad P_2=[-\frac{1}{2}:\frac{\sqrt{3}}{2}:1],\quad P_3=[-\frac{1}{2}:-\frac{\sqrt{3}}{2}:1],\quad R_1=[I,1,0]$$ where $I=\sqrt{-1}$. By direct computations, the defining polynomial $F$ of $Q$ is obtained by $$F(X,Y,Z)=Z^4-6(X^2+Y^2)Z^2+8(X^2-3Y^2)XZ-3(X^2+Y^2)^2.$$ Then another bi-tangent point $R_2$ is $[-I:1:0]$. Note that there is no free parameters left. Now we consider two transformations $\sigma$, $\tau:\Bbb{P}^2\to \Bbb{P}^2$ which are defined as $$\sigma(X:Y:Z):=(X:-Y:Z),\quad \tau(X:Y:Z):=(X:Y:Z)A,$$ where $$A= \begin{pmatrix} \cos \theta & -\sin \theta & 0 \\ \sin \theta & \cos \theta & 0\\ 0&0&1 \end{pmatrix}, \quad \theta=-\frac{2}{3}\pi.$$ Consider the subgroup $G$ of ${\rm{PSL(3;\Bbb{C})}}$ generated by $\sigma$ and $\tau$. Observe that $G\cong S_3$ and $\sigma^2=\tau^3=(\sigma\tau)^2=e$. Then we can see that $L_{\infty}$ and $Q=\{F=0\}$ are stable under the action of $G$: $$F(X,Y,Z)=F(\sigma(X,Y,Z))=F(\tau(X,Y,Z)).$$ We observe also the following. $$\begin{split} & \sigma(R_1)=R_2,\ \sigma(R_2)=R_1,\ \sigma(P_1)=P_1, \ \sigma(P_2)=P_3, \ \sigma(P_3)=P_2.\\ & \tau(R_i)=R_i,\ i=1,2,\qquad \tau(P_i)= \begin{cases} P_{i+1} & {\text{if $i=1,2$}}\\ P_1 & {\text{if $i=3$.}} \end{cases} \end{split}$$ [**[Visible factorization.]{}**]{} Now we consider the possibilities of $(2,3)$ visible factorization of $Q=\{F=0\}$. We assume that $F$ is written as follows: $$F(X,Y,Z)=F_2'(X,Y,Z)^2+F_1'(X,Y,Z)^3Z.$$ In this case, two of $P_1$, $P_2$, $P_3$ must be inner singularities and the rest is an outer singularity. Thus we have three possible cases for these choices: 1. $P_1$ is an outer singularity and $P_2$, $P_3$ are inner singularities. 2. $P_2$ is an outer singularity and $P_1$, $P_3$ are inner singularities. 3. $P_3$ is an outer singularity and $P_1$, $P_2$ are inner singularities. First we assume the case $(1)$. Then $L=\{F_1'=0\}$ and $C_2=\{F_2'=0\}$ are satisfy the following. - $P_1$ is an outer singularity. - $L$ is the line passing $P_2$ and $P_3$. - $C_2$ passes through $P_2$, $P_3$, $R_1$ and $R_2$. Then the defining polynomials $F_1'$ and $F_2'$ are obtained by $$F_1'(X,Y,Z)=-\frac{1}{3}t^2(Z+2X),\quad F_2'(X,Y,Z)=\frac{t^3}{6}(Z^2+4XZ+Y^2+X^2).$$ We take $t$ as one of the solutions $t^6+108=0$. Then $F$ is decomposed into $$\tag{V-1} F(X,Y,Z)=-3(Z^2+4XZ+Y^2+X^2)^2+4(Z+2X)^3Z.$$ Note that $C_2$, $L$, $P_1$ and $\{P_2, P_3\}$ are stable by the action of $\sigma$. Next we consider the case $(2)$. The singular locus $\Sigma(Q)$ is stable by the action of $\tau$ and $\tau(C_2\cap L)=\{P_3,P_1\}$. Hence $P_2$ is the outer singularity. Thus we have $$\tag{V-2} \begin{split} F(X,Y,Z)&= F(\tau(X,Y,Z))=F_2(\tau(X,Y,Z))^2+F_1(\tau(X,Y,Z))^3Z\\ & =-3(Z^2-2XZ-2\sqrt{3}YZ+X^2+Y^2)^2+4(Z-X-\sqrt{3}Y)^3. \end{split}$$ By the same argument, we have one more different torus decomposition: $$\tag{V-3} \begin{split} F(X,Y,Z)&=F(\tau^2(X,Y,Z))=F_2'(\tau^2(X,Y,Z))^2+F_1'(\tau^2(X,Y,Z))^3Z\\ & =-3(Z^2-2XZ+2\sqrt{3}YZ+X^2+Y^2)^2+4(Z-X+\sqrt{3}Y)^3. \end{split}$$ Thus we have three different torus decompositions (V-1), (V-2) and (V-3).\ [**[Invisible factorization.]{}**]{} Next we consider $(2,3)$ invisible factorization (§2): $$Z^2F(X,Y,Z)=F_2(X,Y,Z)^3-F_3(X,Y,Z)^2$$ where $F_2$ and $F_3$ are defined by $$\begin{aligned} (\sharp)\quad \begin{split} F_2(X,Y,Z)&=\ell_1(X,Y)^2+\ell_2(X,Y)\,Z+a_{00}\,Z^2,\\ F_3(X,Y,Z)&=\ell_1(X,Y)^3+\frac{3}{2}\,\ell_1(X,Y)\,\ell_2(X,Y)\,Z +\ell_3(X,Y)\,Z^2+b_{00}\,Z^3 \end{split}\end{aligned}$$ where $\ell_i$ is a linear form for $i=1,2,3$. By the argument in §2.2, the singularity locus $P_1$, $P_2$ and $P_3$ are inner singularities. Hence we have the conditions: $$(*_1)\quad F_2(P_i)=F_3(P_i)=0,\quad i=1,2,3.$$ Moreover one of the bi-tangent points is obtained by the intersection point $\{\ell_1=0\}\cap L_{\infty}$. First we assume that $\{\ell_1=0\}\cap L_{\infty}=\{R_1\}$. By solving conditions $(*_1)$ and $\ell_1(R_1)=0$, we have $a_{00}=0$, $b_{00}=t^3/2$ and $$\ell_1(X,Y)=t\,(X-IY),\quad \ell_2(X,Y)=-t^2\,(X+IY),\quad \ell_3(X,Y)=0.$$ Thus $F_2$ and $F_3$ are given by $$\begin{split} F_2(X,Y,Z)&=t^2(X-IY)^2-t^2(X+IY)Z,\\ F_3(X,Y,Z)&=\frac{t^3}{2}\left( Z^3-3(X^2+Y^2)Z+2(X-IY)^3 \right) \end{split}$$ Note that $C_2=\{F_2=0\}$ and $C_3=\{F_3=0\}$ are stable by the action $\sigma$. Then we have $$\frac{t^6}{4}Z^2F(X,Y,Z)=F_2(X,Y,Z)^3-F_3(X,Y,Z)^2.$$ Hence taking $t$ as one of the solutions $t^6=4$. Thus we have an invisible torus decomposition: $$\tag{In-1} \begin{split} Z^2F(X,Y,Z)&=((X+IY)Z-(X-IY)^2)^3+(Z^3-3(X^2+Y^2)Z+2(X-IY)^3)^2. \end{split}$$ Next we consider the case $\{\ell_1=0\}\cap L_{\infty}=\{R_2\}$. As the singular locus $\Sigma(Q)$ is stable by the action of $\sigma$ and $\sigma(R_1)=R_2$, we have another invisible torus decomposition from (In-1): $$\tag{In-2} Z^2F(\sigma(X,Y,Z))=((X-IY)Z-(X+IY)^2)^3+(Z^3-3(X^2+Y^2)Z+2(X+IY)^3)^2.$$ Thus we have two different invisible torus decompositions (In-1) and (In-2). ![Invisible factorization (In-1) of the quartic (5)[]{data-label="a3"}](./a3.eps) We have shown in the above argument that the three visible decompositions move each other by the action of $\sigma$: $$\begin{matrix} {\rm{(V{\text{-}}1)}}&\overset{\sigma}{\to} & {\rm{(V{\text{-}}2)}}&\overset{\sigma}{\to} & {\rm{(V{\text{-}}3)}}. \end{matrix}$$ We also showed that the two invisible decompositions move each other by the action of $\tau$: $$\begin{matrix} & {\rm{(In{\text{-}}1)}}&\overset{\tau}{\to} &{\rm{(In{\text{-}}2)}}. \end{matrix}$$ In this case, there exist other quasi torus decompositions. We discuss in section 9. Proof of Theorem 2 {#s6} ================== Let $U$ be an open neighborhood of $0$ and let $C(t)=\{F_{2,3}(X,Y,Z,t)=0\}$ $(t\in U)$ be a $(2,3)$ torus curve which is defined by $$F_{2,3}(X,Y,Z,t)=H_3(X,Y,Z,t)^2+H_2(X,Y,Z,t)^3.$$ Fix $B=[0:1:0]$ in $\Bbb{P}^2$. We consider the family $\{C(t)\}_{t\in U}$ which has $B$ as the base point with multiplicity $2$. We assume that the defining polynomials have the following form: $$\tag{$*_2$} \begin{cases} \begin{split} H_2(X,Y,Z,t)&=F_1'(X,Y,Z)Z+t\,X K_1(X,Y),\\ H_3(X,Y,Z,t)&=F_2'(X,Y,Z)Z+t\,X K_2(X,Y) \end{split} \end{cases}$$ where $F_i'(X,Y,Z)$ and $K_i(X,Y)$ homogeneous polynomials of degree $i$ such that $K_i(0,1)\ne 0$ for $i=1,2$. Then $B$ is in $C(t)$ with multiplicity $2$ for $t\ne 0$ and $\{C(t)\}_{t\in U}$ degenerates into $C(0)=D+2L_{\infty}$ where $D=\{G=0\}$ is a visible factorization of torus curve which is defined as $$G(X,Y,Z)=F_2'(X,Y,Z)^2+F_1'(X,Y,Z)^3Z.$$ Thus we can construct a line degeneration of $(2,3)$ torus sextic. For example, we give line degeneration for the case $(1)$. Let $Q=\{F=0\}$ be the quartic which is satisfies the condition of $(1)$. In §4, we obtained the defining polynomials $F$, $F_1$ and $F_2$ as $$\begin{split} Q:&\quad F(X,Y,Z)=Z^4+2(Y^2-X^2)Z^2+t\,Y^3Z+(Y^2-X^2)^2=0, \\ L:&\quad F_1'(X,Y,Z)=s_1\,Y=0,\\ C_2:&\quad F_2'(X,Y,Z)=s_2\,(X^2+Y^2-Z^2)=0 \end{split}$$ where $s_1$ and $s_2$ are one of the solution $s_1^3=t$ and $s_2^2=1$ respectively. Let $K_i(X,Y)$ be any homogeneous polynomial of degree $i$ such that $K_i(0,1)\ne 0$ for $i=1$, $2$. We take $H_2$ and $H_3$ as $(*_2)$ and then the family $\{C(t)\}_{t\in U}$ degenerates into $C(0)=Q+2L_{\infty}$. Degenerate families of $QL$-configurations. {#s7} =========================================== Let $\mathcal{QL}$ be the set of quaritcs of $QL$-configurations and let $L_{\infty}$ be the fixed line at infinity. We consider the subset $\mathcal{QL}(n)$ of $\mathcal{QL}$ which is the set of quartic $Q$ whose configuration $Q\cup L_{\infty}$ is of the type $(n)$ in Table 1 for $n=1,\dots, 12$. It is sesy to see that $\mathcal{QL}(n)$ is a connected subspace of the space of quartic under the canonical topology. This implies that for any $Q$, $Q'\in \mathcal{QL}(n)$ the topology of $\Bbb{C}^2\setminus Q$ and $\Bbb{C}^2\setminus Q'$ are homeomorphic. For the comparison of the topology of $\Bbb{C}^2\setminus Q$ and $\Bbb{C}^2\setminus Q'$ for $Q\in \mathcal{QL}(n)$ and $Q'\in \mathcal{QL}(m)$, we consider the degeneration problem among these subsets. Suppose that there exists an analytic family $Q(s)$ $(s\in U)$ of quartic for an open set $U$ of the origin $0\in \Bbb{C}$ such that $Q(s)\in \mathcal{QL}(n)$ for $s\in U\setminus \{0\}$ and $Q(0)\in \mathcal{QL}(m)$ for some $n$, $m$ with $n\ne m$. In particular, $Q(s)\cup L_{\infty}\to Q(0)\cup L_{\infty}$. Then, by an degenerate properties ([@OkaSurvey]), we have the surjectivity $\pi_1(\Bbb{C}\setminus Q(0))\twoheadrightarrow \pi_1(\Bbb{C}\setminus Q(s))$ $(s\ne 0)$ and the divisibility of the Alexander polynomials $\Delta_{Q(s)}(t) \mid \Delta_{Q(0)}(t)$. We denote this situation as $\mathcal{QL}(n) \to \mathcal{QL}(m)$. There exist the following degeneration families among the set $\mathcal{QL}(n)$, $n=1,\dots, 12$. $$\begin{matrix} && && &\mathcal{QL} (10)&&&&\\ & &&&&\uparrow\\ & && \mathcal{QL} (4)&\leftarrow &\mathcal{QL} (2) &\to &\mathcal{QL} (8)& \to &\mathcal{QL} (12)\\ &&& \uparrow&& \uparrow && \uparrow && \uparrow\\ & \mathcal{QL} (5)&\leftarrow &\mathcal{QL} (3)&\leftarrow &\mathcal{QL} (1) &\to &\mathcal{QL} (7)& \to &\mathcal{QL} (11)\\ &&&&&\downarrow \\ & &&&& \mathcal{QL} (6)& \to &\mathcal{QL} (9) \end{matrix}$$ This follows from the following explicit degenerations: 1. There exist a family of quintic $\{C_{s,t,u}\} \subset \mathcal{QL}(1)$ where $C_{s,t,u}=Q_{s,t,u}+L_{\infty}$ $(stu\ne 0)$ with $3$ parameters such that $$\begin{split} & C_{0,t,u}\in \mathcal{QL}(7),\quad C_{0,t,0}\in \mathcal{QL}(11) {\text{\ \ for \ $t\ne 0$,}}\\ & C_{s,0,u}\in \mathcal{QL}(2), \quad C_{0,0,u}\in \mathcal{QL}(8), \quad C_{0,0,0}\in \mathcal{QL}(12). \end{split}$$ 2. There exist a family of quintic $\{C_{s,t,u}\} \subset \mathcal{QL}(1)$ where $C_{s,t,u}=Q_{s,t,u}+L_{\infty}$ $(stu\ne 0)$ with $3$ parameters such that $$\begin{split} & C_{0,t,u}\in \mathcal{QL}(3),\quad C_{0,t,0}\in \mathcal{QL}(5) {\text{\ \ for \ \ $t\ne 0$,}}\\ & C_{s,0,u}\in \mathcal{QL}(2), \quad C_{0,0,u}\in \mathcal{QL}(4){\text{\ \ for \ $u\ne 0$}}. \end{split}$$ 3. There exist a family of quintic $\{C_{s,t}\} \subset \mathcal{QL}(1)$ where $C_{s,t}=Q_{s,t}+L_{\infty}$ $(st\ne 0)$ with $2$ parameters such that $$C_{0,t}\in \mathcal{QL}(6),\quad C_{0,0}\in \mathcal{QL}(9).$$ 4. There exist a family of quintic $\{C_{s,t}\} \subset \mathcal{QL}(1)$ where $C_{s,t}=Q_{s,t}+L_{\infty}$ $(st\ne 0)$ with $2$ parameters such that $$C_{0,t}\in \mathcal{QL}(2),\quad C_{0,0}\in \mathcal{QL}(10).$$ For a proof, we give explicit defining equations of quartic for each case. $$\tag{1} \begin{split} F_{s,t,u}(X,Y,Z)&=s^4Z^4-2s^2uYZ^3 +(2s^2(t^2Y^2-X^2)+u^2Y^2)Z^2\\ & \hspace{3cm} +(Y^2+2u(X^2-t^2Y^2))YZ +(X^2-t^2Y^2)^2. \end{split}$$ $$\tag{2} \begin{split} F_{s,t,u}(X,Y,Z)&=s(2+s)Z^4-3sXZ^3 +\frac{1}{4}(x^2(3s+8u+8su)-y^2(s+1)(8u+3))Z^2\\ & \hspace{1cm} -\frac{1}{8}(u(x^2-t^2y^2)+(x^2-9t^2y^2))XZ +\frac{1}{64}(8u+3)^2(X^2-t^2Y^2)^2. \end{split}$$ $$\tag{3} \begin{split} F_{s,t}(X,Y,Z)&=(t^3+1)Z^4+3t^2\ell_1(X,Y,s)Z^3 +(3\ell_1(X,Y,s)^2t-2(X^2-Y^2))Z^2\\ & \hspace{4cm} +\ell_1(X,Y,s)^3Z +(X^2-Y^2)^2 \end{split}$$ where $\ell_1(X,Y,s)=X-sY-Y$. $$\tag{4} \begin{split} F_{s,t}(X,Y,Z)&=(X+Y)Z^2+ ((X-sY)^3-2(X+Y)(t^2Y^2-X^2))Z+(t^2Y^2-X^2)^2. \end{split}$$ [**[Example.]{}**]{} Let $Q_1\in \mathcal{QL}(1)$ and $Q_5\in \mathcal{QL}(5)$ be $QL$-configurations of types $(1)$ and $(5)$ whose fundamental gruops and Alexander polynomials are given as follows ([@Y]): $$\begin{split} & \pi_1(\Bbb{C}^2\setminus Q_1)\cong \langle a, b \mid R_1 \rangle, \qquad \pi_1(\Bbb{C}^2\setminus Q_5)\cong \langle a, b,c \mid R_1,\, R_2, \, R_3 \rangle\\ &R_1: aba=bab,\ R_2: bcb = cbc, \ R_3: c(b^{-1}ab)c=(b^{-1}ab)c(b^{-1}ab)\\ & \Delta_{Q_1}(t)=t^2-t+1,\qquad \Delta_{Q_5}(t)=(t^2-t+1)^2. \end{split}$$ By above Proposition, we have $\mathcal{QL}(1)\to \mathcal{QL}(5)$. Hence we have the surjectivity $\pi_1(\Bbb{C}^2\setminus Q_5)\twoheadrightarrow \pi_1(\Bbb{C}^2\setminus Q_1)$ and the divisibility $\Delta_{Q_1}(t)\mid \Delta_{Q_5}(t)$. Non-irreducible $QL$-configurations {#s8} =================================== In this section, we consider torus decompositions of non-irreducible $QL$-configurations for the quartics $(14),\dots, (19)$. Then we will get defining polynomials and torus decompositions by the same argument of irreducible case. Recall that the situations of each case: No. ${{\rm{Sing}}}(Q)$ $Q\cap L_{\infty}$ irreducible components ------ ----------------------------- -------------------- ----------------------------- -- -- -- -- (14) $a_3^{\infty}+a_2+a_1$ (ii) a cuspidal cubic and a line (15) $a_5+a_1$ (i) two conics (16) $a_5^{\infty}+a_2^{\infty}$ (iii) a cuspidal cubic and a line (17) $2a_3^{\infty}$ (iii) two conics (18) $a_7^{\infty}$ (v) two conics (19) $2a_3^{\infty}+a_1$ (iii) a conic and two lines \ Table 3.\ Invisible factorization of (2,4) torus curves {#s9} --------------------------------------------- Let $D$ be an invisible factorization of $(2,4)$ torus curve which satisfies the following. $$\begin{split} F_{2,4}(X,Y,Z)&=F_2(X,Y,Z)^4-F_4(X,Y,Z)^2\\ &=(F_2(X,Y,Z)^2-F_4(X,Y,Z))(F_2(X,Y,Z)^2+F_4(X,Y,Z))\\ & =Z^4G(X,Y,Z). \end{split}$$ By the same argument in §2.2, we can assume that the forms of $F_2$ and $F_4$ are $$\begin{split} F_2(X,Y,Z)&=F_2^{(2)}(X,Y)+F_{2}^{(1)}(X,Y)Z+a_{00}Z^2,\quad \deg F_2^{(i)}=i,\\ F_4(X,Y,Z)&=F_2(X,Y,Z)^2-c\,Z^4,\quad c=b_{00}-a_{00}^2\ne 0. \end{split}$$ Then $D=\{G=0\}$ is defined by $$D:\quad G(X,Y,Z)=F_2(X,Y,Z)^2-c'Z^4=0,\quad c'\ne 0.$$ By the form of the defining polynomial of $D$, the inner singularities of $D$ is on $L_{\infty}$. Singularities of $D$ are described as follows. Under the above notations, $D$ has the following singularities. 1. If $C_2$ is smooth at $P\in C_2\cap L_{\infty}$, then $(D,P)\sim a_{3\iota-1}$ where $\iota=I(C_2,L_{\infty};P)$. 2. If $C_2$ is singular at $P\in C_2\cap L_{\infty}$, then $D$ consists of four lines. 3. If $P\in D$ is an outer singularity, then $(D,P)\sim a_1$. Our proof is done in the same way as Lemma \[l1\] and [@Oka-Pho2]. Torus decompositions of non-irreducible $QL$-configurations ----------------------------------------------------------- In this section, we show the possibilities of torus decompositions for non-irreducible $QL$-configurations. Our proof is similar to the cases of irreducible $QL$-configurations. For the quartics $(14)$ and $(15)$, we use $(2,3)$ visible factorizations. For the quartics $(17)$, $(18)$ and $(19)$, we use $(2,4)$ invisible factorizations.\ [|c|c|c|c|c|c|c|c|]{}\ Singularities & $a_3^{\infty}+a_2+a_1$ at $[1:1:0]$, $[0:0:1]$, $[-1:0:1]$.\ $Q\cap L_\infty$ & singular at $[1:1:0]$ and tangent at $[-1:1:0]$.\ Torus decomposition & $(X^2-Y^2-XZ+3YZ)^2+4\,(X-Y)^3Z=0$.\ \ Singularities & $a_5+a_1$ at $[0:0:1]$, $[-1:0:1]$.\ $Q\cap L_\infty$ & bi-tangent at $[1:1:0]$ and $[-1:1:0]$.\ Torus decomposition & $(X^2-Y^2+XZ)^2+4\,X^3Z=0$.\ \ Singularities & $2a_3$ at $[1:1:0]$ and $[-1:0:0]$.\ $Q\cap L_\infty$ & singular at $[1:1:0]$ and $[-1:0:0]$.\ Torus decomposition & $\frac{1}{64}(2X^2-2Y^2-t_2Z^2)^4$\  & $- (\frac{1}{8}(2X^2-2Y^2-t_2Z^2)^2+\frac{1}{4}(4t_1-t_2^2)Z^4)^2$.\ \ Singularities & $a_7$ at $[0:1:0]$.\ $Q\cap L_\infty$ & singular at $[0:1:0]$.\ Torus decomposition & $\left(Z^2-\frac{2}{a_{01}c_2}(c_3X-c_2Y)Z-2\frac{c_2}{a_{01}}X^2\right)^4$ \  & $- \left(\left(Z^2-\frac{2}{a_{01}c_2}(c_3X-c_2Y)Z -2\frac{c_2}{a_{01}}X^2\right)^2 +2\frac{4a_{00}-a_{01}^2}{a_{01}^2}Z^4\right)^2$. \ \ Singularities & $2a_3+a_1$ at $[1:1:0]$, $[-1:1:0]$, $[0:0:1]$.\ $Q\cap L_\infty$ & singular at $[1:1:0]$ and $[-1:1:0]$.\ Torus decomposition & $\frac{1}{t^2}\Big{(} (-X^2+Y^2-\frac{1}{2}tZ^2)^4$\  & $ -\left((-X^2+Y^2-\frac{1}{2}tZ^2)^2 -\frac{1}{2}t^2Z^4\right)^2\Big{)}$.\ Infiniteness of $(2,3)$ quasi torus decompositions ================================================== In this section, we consider the possibilities of $(2,3)$ quasi torus decompositions of a plane curve which admits a $(2,3)$ torus decomposition. We assert: \[p1\] Let $C=\{f=0\}\subset \Bbb{C}^2$ be a $(2,3)$ torus curve of any degree. Then $C$ has infinitely many $(2,3)$ quasi torus decompositions. Suppose that $f(x,y)$ can be written as $f(x,y)=h_0(x,y)^2-g_0(x,y)^3$. We put inductively $$\begin{split} g_{i+1}(x,y)&=-\frac{4}{3}\,h_{i}(x,y)^2+g_{i}(x,y)^3,\\ h_{i+1}(x,y)&=\frac{\sqrt{-3}}{9}h_i(x,y)(-8\,h_i(x,y)^2+9\,g_i(x,y)^3) \end{split}$$ for $i\ge 0$. Then we claim that they satisfy the following equality: $$\tag{$*_3$} \left(\prod_{k=0}^ig_k(x,y)\right)^6f(x,y)=h_{i+1}(x,y)^2-g_{i+1}(x,y)^3,\quad i\ge 0.$$ Indeed, by a simple calculation, we have $$h_{i+1}(x,y)^2-g_{i+1}(x,y)^3=g_i(x,y)^6\left(h_{i}(x,y)^2-g_{i}(x,y)^3\right).$$ The assertion follows immediately from this equality. [[ Let $C(t)$ be a family of curves given by $$C(t):t\,h_{i+1}(x,y)^2-g_{i+1}(x,y)^3=0, \quad t\in \Bbb{C}.$$ We thank Professor J. I. Cogolludo for informing us the generic fiber of this pencil is not irreducible.]{}]{} Now we study the location of singularities of a family of $(2,3)$ quasi torus decompositions which has the form $(*_3)$.\ We put $r_i(x,y):=\prod_{k=0}^ig_k(x,y)$ and $\Sigma_i:=\{h_i=0\}\cap \{g_i=0\}$ for $i\ge 0$. Then, by the definitions, we have the followings: 1. $\Sigma_0$ is the set of inner singularities of $\{f=0\}$. 2. $\Sigma_{i}\subset \{r_i=0\}$ for all $i \ge 0$. 3. $\Sigma_0\subset \Sigma_1 \subset \cdots \subset \Sigma_i \subset \cdots$. In particular, $\{r_i=0\}$ contains the inner singularities of $\{f=0\}$ for all $i\ge 0$. By Proposition \[p1\] and above observations, it is important to study the existence of $(2,3)$ torus decompositions which is obtained by visible or invisible degenerations. We are also interested in quasi torus decompositions which does not come from a torus decomposition as in $(*_3)$. We will give such an example $(2,3)$ quasi torus decomposition. Let $Q=\{f=0\}$ be the three cuspidal quartic which has three $(2,3)$ torus decompositions (V-1), (V-2) and (V-3) as in §5. Recall that the $(2,3)$ torus decomposition (V-1) and locations of singularities: $$\begin{split} &\qquad f(x,y)=-3(x^2+y^2+4x+1)^2+4(2x+1)^3,\\ & P_1=(1,0),\quad P_2=\left(-\frac{1}{2},\frac{\sqrt{3}}{2}\right),\quad P_3=\left(-\frac{1}{2},-\frac{\sqrt{3}}{2}\right) \end{split}$$ where $P_2,P_3$ are the inner singularities of torus decomposition (V-1). Now we take following three polynomials $s_1$ $s_3$ and $s_5$ of degree $1$, $3$ and $5$ respectively: $$\begin{split} s_1(x,y)&=x-Iy-1,\\ s_3(x,y)&=\sqrt[3]{4}(3Iy^3-(5x+7)y^2-I(x-1)^2y-(x-1)^3),\\ s_5(x,y)&=\sqrt{3}(y^5+3I(x+5)y^4-2 (x^2+13x+10)y^3+2I(x-4)(x-1)^2y^2\\ &\hspace{5cm} -3(x+1)(x-1)^3y -I(x-1)^5). \end{split}$$ Then they satisfy the following equality: $$s_1(x,y)^6f(x,y)=s_5(x,y)^2+s_3(x,y)^3.$$ Note that $\{s_1=0\}$ does not pass through the inner singularities of $Q$. Thus this decomposition is example of $(2,3)$ quasi torus decomposition which does not come from as in $(*_3)$. ![[]{data-label="a21"}](./a21.eps) [**[Acknowledgment.]{}**]{} I would like to express our deepest gratitude to Professor Hiro-o Tokunaga who has proposed this problem. I also express our deepest gratitude to Professor Mutsuo Oka for his various advices during the preparation of this paper. [1]{} B. Audoubert, T. C. Nguyen, and M. Oka. On [A]{}lexander polynomials of torus curves. , 57(4):935–957, 2005. Y. Kenta. On the topology of the complements of quartic and line configurations. , 44:125–152, 2008. V. S. Kulikov. On plane algebraic curves of positive [Albanese]{} dimension. , 59(6):1173–1192, 1995. M. Oka. Geometry of cuspidal sextics and their dual curves. In [*Singularities—Sapporo 1998*]{}, pages 245–277. Kinokuniya, Tokyo, 2000. M. Oka. A survey on [Alexander]{} polynomials of plane curves. , 10:209–232, 2005. M. Oka. Tangential [A]{}lexander polynomials and non-reduced degeneration. In [*Singularities in geometry and topology*]{}, pages 669–704. World Sci. Publ., Hackensack, NJ, 2007. M. Oka and D. Pho. Classification of sextics of torus type. , pages 399–433, 2002. D. T. Pho. Classification of singularities on torus curves of type $(2,3)$. , 24(1):259–284, 2001. H.-o. Tokunaga. Dihedral covers and an elementary arithmetic on elliptic surfaces. , 44(2):255–270, 2004.

--- abstract: 'Generalized Berwald manifolds are Finsler manifolds admitting linear connections such that the parallel transports preserve the Finslerian length of tangent vectors (compatibility condition). By the fundamental result of the theory [@V5] such a linear connection must be metrical with respect to the averaged Riemannian metric given by integration of the Riemann-Finsler metric on the indicatrix hypersurfaces. Therefore the linear connection (preserving the Finslerian length of tangent vectors) is uniquely determined by its torsion. If the torsion is zero then we have a classical Berwald manifold. Otherwise, the torsion is a strange data we need to express in terms of the intrinsic quantities of the Finsler manifold. In the paper we consider the extremal compatible linear connection of a generalized Berwald manifold by minimizing the pointwise length of its torsion tensor. It is a conditional extremum problem involving functions defined on a local neighbourhood of the tangent manifold. In case of a given point of the manifold, the reference element method provides that the number of the Lagrange multipliers equals to the number of the equations providing the compatibility of the linear connection to the Finslerian metric. Therefore the solution of the conditional extremum problem with a reference element can be expressed in terms of the canonical data. The solution of the conditional extremum problem independently of the reference elements can be constructed algorithmically at each point of the manifold. The pointwise solutions constitute a section of the torsion tensor bundle for testing the compatibility of the corresponding linear connection to the Finslerian metric. In other words, we have an intrinsic algorithm to check the existence of compatible linear connections on a Finsler manifold because it is equivalent to the existence of the extremal compatible linear connection.' address: | Inst. of Math., Univ. of Debrecen\ H-4002 Debrecen, P.O.Box 400\ Hungary author: - Csaba Vincze title: On the extremal compatible linear connection of a generalized Berwald manifold --- Introduction {#introduction .unnumbered} ============ The notion of the generalized Berwald manifolds goes back to V. Wagner [@Wag1]. They are Finsler manifolds admitting linear connections such that the parallel transports preserve the Finslerian length of tangent vectors (compatibility condition). We are interested in the unicity of the compatible linear connection and its expression in terms of the canonical data of the Finsler manifold (intrinsic characterization). If the torsion is zero (classical Berwald manifolds), the intrinsic characterization is the vanishing of the mixed curvature tensor of the canonical horizontal distribution. The problem of the intrinsic characterization is solved in the more general case of Finsler manifolds admitting semi-symmetric compatible linear connections [@V10], see also [@V11]. We also have a unicity statement because the torsion tensor of the semi-symmetric compatible linear connection can be explicitly expressed in terms of metrics and differential forms given by averaging. Especially, the integration of the Riemann-Finsler metric on the indicatrix hypersurfaces (the so-called averaged Riemannian metric) provides a Riemannian environment for the investigations. The fundamental result of the theory [@V5] states that a linear connection satisfying the compatibility condition must be metrical with respect to the averaged Riemannian metric. Therefore the compatible linear connection is uniquely determined by its torsion tensor. Unfortunately, the unicity statement for the compatible linear connection of a generalized Berwald manifold is false in general [@Vin1]. To avoid the difficulties coming from different possible solutions, the idea is to look for the extremal solution in some sense: the extremal compatible linear connection of a generalized Berwald manifold keeps its torsion as close to the zero as possible. It is a conditional extremum problem involving functions defined on a local neighbourhood of the tangent manifold. In case of a given point of the manifold, the reference element method provides that the number of the Lagrange multipliers equals to the number of the equations providing the compatibility of the linear connection to the Finslerian metric. Therefore the solution of the conditional extremum problem with a reference element can be expressed in terms of the canonical data. The solution of the conditional extremum problem independently of the reference elements can be constructed algorithmically at each point of the manifold. The pointwise solutions constitute a continuous section of the torsion tensor bundle. The continuity of the components of the torsion tensor implies the continuity of the connection parameters. Using parallel translations with respect to such a connection we can conclude that the Finsler metric is monochromatic. By the fundamental result of the theory [@BM] it is sufficient and necessary for a Finsler metric to be a generalized Berwald metric. Therefore we have an intrinsic algorithm to check the existence of compatible linear connections on a Finsler manifold because it is equivalent to the existence of the extremal compatible linear connection. Notations and terminology ========================= Let $M$ be a differentiable manifold with local coordinates $u^1, \ldots, u^n.$ The induced coordinate system of the tangent manifold $TM$ consists of the functions $x^1, \ldots, x^n$ and $y^1, \ldots, y^n$. For any $v\in T_pM$, $x^i(v):=u^i\circ \pi (v)=p$ and $y^i(v)=v(u^i)$, where $i=1, \ldots, n$ and $\pi \colon TM \to M$ is the canonical projection. A Finsler metric is a continuous function $F\colon TM\to \mathbb{R}$ satisfying the following conditions: $\displaystyle{F}$ is smooth on the complement of the zero section (regularity), $\displaystyle{F(tv)=tF(v)}$ for all $\displaystyle{t> 0}$ (positive homogeneity) and the Hessian $$g_{ij}=\frac{\partial^2 E}{\partial y^i \partial y^j}$$ of the energy function $E=F^2/2$ is positive definite at all nonzero elements $\displaystyle{v\in T_pM}$ (strong convexity), $p\in M$. The so-called *Riemann-Finsler metric* $g$ is constituted by the components $g_{ij}$. It is defined on the complement of the zero section. The Riemann-Finsler metric makes the complement of the origin a Riemannian manifold in each tangent space. The canonical objects are the [*volume form*]{} $\displaystyle{d\mu=\sqrt{\det g_{ij}}\ dy^1\wedge \ldots \wedge dy^n}$, the *Liouville vector field* $\displaystyle{C:=y^1\partial /\partial y^1 +\ldots +y^n\partial / \partial y^n}$ and the [*induced volume form*]{} $$\mu=\sqrt{\det g_{ij}}\ \sum_{i=1}^n (-1)^{i-1} \frac{y^i}{F} dy^1\wedge\ldots\wedge dy^{i-1}\wedge dy^{i+1}\ldots \wedge dy^n$$ on the indicatrix hypersurface $\displaystyle{\partial K_p:=F^{-1}(1)\cap T_pM\ \ (p\in M)}$. The averaged Riemannian metric is defined by $$\label{averagemetric1} \gamma_p (v,w):=\int_{\partial K_p} g(v, w)\, \mu=v^i w^j \int_{\partial K_p} g_{ij}\, \mu \ \ (v, w\in T_p M, p\in U).$$ A linear connection $\nabla$ on the base manifold $M$ is called *compatible* to the Finslerian metric if the parallel transports with respect to $\nabla$ preserve the Finslerian length of tangent vectors. Finsler manifolds admitting compatible linear connections are called generalized Berwald manifolds. Suppose that the parallel transports with respect to $\nabla$ (a linear connection on the base manifold) preserve the Finslerian length of tangent vectors and let $X$ be a parallel vector field along the curve $c\colon [0,1]\to M$. We have that $$\label{eq:5} (F \circ X)'=(x^k\circ X)'{\frac{\partial F}{\partial x^k}}\circ X+(y^k \circ X)'{\frac{\partial F}{\partial y^k}}\circ X$$ where $\displaystyle{(x^k\circ X)'={c^k}'}$ and $\displaystyle{{X^k}'=-{c^i}' X^j \Gamma_{ij}^k\circ c}$ because of the differential equation for parallel vector fields. Therefore $$\label{eq:55} (F \circ X)'={c^i}'\bigg(\frac{\partial F}{\partial x^i}-y^j {\Gamma}_{ij}^{k}\circ \pi \frac{\partial F}{\partial y^k}\bigg)\circ X.$$ This means that the parallel transports with respect to $\nabla$ preserve the Finslerian length of tangent vectors (compatibility condition) if and only if $$\label{cond1} \frac{\partial F}{\partial x^i}-y^j {\Gamma}^k_{ij}\circ \pi \frac{\partial F}{\partial y^k}=0,$$ where $i=1, \ldots,n$. The vector fields of type $$\label{eq:6} X_i^{h}:=\frac{\partial}{\partial x^i}-y^j {\Gamma}^k_{ij}\circ \pi \frac{\partial}{\partial y^k}$$ span the horizontal distribution belonging to $\nabla$. In a similar way, we can introduce the horizontal vector fields $X_i^{h^*}$ ($i=1, \ldots, n$) with respect to the Lévi-Civita connection of the averaged Riemannian metric $F^*(v):=\sqrt{\gamma_p(v,v)}.$ \[heritage\] *[@V5]* If a linear connection on the base manifold is compatible to the Finslerian metric function then it must be metrical with respect to the averaged Riemannian metric. In what follows we are going to substitute the connection parameters with the components of the torsion tensor in the equations of the compatibility condition (\[cond1\]). Since the torsion tensor bundle can be equipped with a Riemannian metric in a natural way, we can measure the length of the torsion to formulate an extremum problem for the compatible linear connection keeping its torsion as close to the origin as possible. The extremal compatible linear connection of a generalized Berwald manifold =========================================================================== Let $F$ be the Finslerian metric of a connected generalized Berwald manifold and suppose that $\nabla$ is a compatible linear connection. Taking vector fields with pairwise vanishing Lie brackets on a local neighbourhood of the base manifold, the Christoffel process implies that $$X\gamma(Y,Z)+Y\gamma(X,Z)-Z\gamma(X,Y)=$$ $$2\gamma(\nabla_X Y, Z)+\gamma(X, T(Y,Z))+\gamma(Y, T(X,Z))-\gamma(Z, T(X,Y))$$ and, consequently, $$\label{Cproc} \gamma(\nabla^*_X Y,Z)=\gamma(\nabla_X Y, Z)+\frac{1}{2}\left(\gamma(X, T(Y,Z))+\gamma(Y, T(X,Z))-\gamma(Z, T(X,Y))\right),$$ where $\nabla^*$ denotes the Lévi-Civita connection of the averaged Riemannian metric $\gamma$ and $T$ is the torsion tensor of $\nabla$. In terms of the connection parameters $$\label{torsion} \Gamma_{ij}^r=\Gamma_{ij}^{*r}-\frac{1}{2}\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)$$ and the compatibility condition (\[cond1\]) can be written into the form $$\label{cond2} X_i^{h^*}F+\frac{1}{2}y^j \left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)\circ \pi \frac{\partial F}{\partial y^r}=0\ \ (i=1, \ldots,n).$$ Formula (\[Cproc\]) shows that the correspondence $\nabla \rightleftharpoons T$ preserves the affine combinations of the linear connections, i.e. for any real number $\lambda\in \mathbb{R}$ we have $$\lambda \nabla_1+(1-\lambda)\nabla_2 \rightleftharpoons \lambda T_1+(1-\lambda)T_2.$$ Additionally, if $\nabla_1$ and $\nabla_2$ satisfy the compatibility condition (\[cond1\]) then so does $$\lambda \nabla_1+(1-\lambda)\nabla_2.$$ This means that the set containing the restrictions of the torsion tensors of the compatible linear connections to the Cartesian product $\displaystyle{T_pM\times T_pM}$ is an affine subspace in the linear space $\wedge^2 T_p^*M \otimes T_pM$ for any $p\in M$. As the point is varying we have an affine distribution of the torsion tensor bundle $\wedge^2 T^*M \otimes TM$. In terms of local coordinates, the bundle is spanned by $$du^i \wedge du^j \otimes \frac{\partial}{\partial u^k} \quad (1\leq i < j\leq n, k=1, \ldots, n)$$ and its dimension is $\displaystyle{\binom{n}{2}n}$. The products $$du^i \wedge du^j \otimes \frac{\partial}{\partial u^k} \quad (1\leq i< j \leq n, k=1, \ldots, n)$$ form an orthonormal basis at the point $p\in M$ if the coordinate vector fields $\partial/\partial u^1, \ldots, \partial/\partial u^n$ form an orthonormal basis with respect to the averaged Riemannian metric at the point $p\in M$. The norm of the torsion tensor is defined by $$\label{torsionnorm} \|T_p\|^2=\sum_{1\leq i < j \leq n} \sum_{k=1}^n {T_{ij}^k(p)}^2$$ provided that $\displaystyle{T=\sum_{1\leq i < j \leq n} \sum_{k=1}^nT_{ij}^k du^i \wedge du^j \otimes \frac{\partial}{\partial u^k}}$ and the products form an orthonormal basis at the point $p\in M$. The corresponding inner product is $$\langle T_p, S_p \rangle=\sum_{1\leq i < j \leq n} \sum_{k=1}^n T_{ij}^k(p)S_{ij}^k(p).$$ Let a point $p\in M$ be given and consider the affine subspace $A_p$ in $\wedge^2 T_p^*M \otimes T_pM$ defined by $$\label{condextb} X_i^{h^*}F(v)+\frac{1}{2}y^j (v)\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)(p) \frac{\partial F}{\partial y^r}(v)=0,$$ where $i=1, \ldots, n$ and $v\in T_pM$. Note that $A_p$ is nonempty because it contains the restrictions of the torsion tensors of the compatible linear connections to the Cartesian product $\displaystyle{T_pM\times T_pM}$. If $T_p\in A_p$ then we can write $A_p$ as the translate $\displaystyle{A_p=T_p+H_p}$, where the linear subspace $H_p \subset \wedge^2 T_p^*M \otimes T_pM$ is defined by $$\label{condextc} \frac{1}{2}y^j (v)\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)(p) \frac{\partial F}{\partial y^r}(v)=0,$$ where $i=1, \ldots, n$ and $v\in T_pM$. \[lemmakey\] If $\ \nabla$ is a compatible linear connection then the linear subspace $H_p$ is invariant under the action $$(\varphi T)_p(v,w):=\varphi^{-1} T_p(\varphi(v), \varphi(v))$$ of the holonomy group of $\nabla$ at the point $p\in M$. Suppose that the coordinate vector fields $\partial/\partial u^1, \ldots, \partial/\partial u^n$ form an orthonormal basis with respect to the averaged Riemannian metric at the point $p\in M$ and let $Q_i^j$ be the matrix representation of $\varphi$, $P_j^i=\left(Q_i^j\right)^{-1}$. Evaluating (\[condextc\]) at $\varphi(v)$ we have an equivalent system of equations because $v$ runs through the non-zero elements in $T_pM$. So does $\varphi(v)$. Therefore (\[condextc\]) is equivalent to $$\frac{1}{2}y^j \circ \varphi (v)\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)(p) \frac{\partial F}{\partial y^r}\circ \varphi(v)=0,$$ $$\frac{1}{2}y^b (v)\left(T^{l}_{jk}Q_b^j P_r^c \gamma^{kr}\gamma_{il}+T^{l}_{ik}Q_b^j P_r^c\gamma^{kr}\gamma_{jl}-T_{ij}^rQ_b^j P_r^c\right)(p) \frac{\partial F}{\partial y^c}\circ \varphi(v)=0,$$ where $i=1, \ldots, n$ and $v\in T_pM$. Indeed, $\displaystyle{\varphi^j=y^j\circ \varphi=Q_b^j y^b}$ and the invariance property $F\circ \varphi=F$ implies that $$\frac{\partial F}{\partial y^c}(v)=\frac{\partial F\circ \varphi}{\partial y^c}(v)=Q_c^r\frac{\partial F}{\partial y^r}\circ \varphi(v) \ \Rightarrow \ P_r^c\frac{\partial F}{\partial y^c}(v)=\frac{\partial F}{\partial y^r}\circ \varphi(v).$$ Since the identities $\displaystyle{Q_a^i Q_b^j\gamma_{ij}(p)=\gamma_{ab}(p)}$ and $\displaystyle{P_i^a P_j^b \gamma^{ij}(p)=\gamma^{ab}(p)}$ give that $\displaystyle{Q_b^j \gamma_{jl}(p)=P_l^j \gamma_{jb}(p)}$ and $\displaystyle{P_j^b \gamma^{jk}(p)=Q_j^k\gamma^{jb}(p)}$, we have $$\frac{1}{2}y^b (v)\left(T^{l}_{jk}Q_b^j Q_r^k \gamma^{rc}\gamma_{il}+T^{l}_{ik}Q_b^j Q_r^k\gamma^{rc}\gamma_{jl}-T_{ij}^r Q_b^j P_r^c\right)(p) \frac{\partial F}{\partial y^c}(v)=0,$$ where $i=1, \ldots, n$ and $v\in T_pM$. Taking the product by the matrix $Q^i_a$ the equivalent system of equations is $$\frac{1}{2}y^b (v)\left(T^{l}_{jk}Q_a^i Q_b^j Q_r^k \gamma^{rc}\gamma_{il}+T^{l}_{ik}Q_a^i Q_b^j Q_r^k\gamma^{rc}\gamma_{jl}-T_{ij}^r Q_a^i Q_b^j P_r^c\right)(p) \frac{\partial F}{\partial y^c}(v)=0,$$ $$\frac{1}{2}y^b (v)\left(T^{l}_{jk}P_l^i Q_b^j Q_r^k \gamma^{rc}\gamma_{ia}+T^{l}_{ik}Q_a^i P_l^j Q_r^k\gamma^{rc}\gamma_{jb}-T_{ij}^r Q_a^i Q_b^j P_r^c\right)(p) \frac{\partial F}{\partial y^c}(v)=0,$$ $$\frac{1}{2}y^b (v)\left(\left(\varphi T\right)_{br}^i \gamma^{rc}\gamma_{ia}+\left(\varphi T\right)_{ar}^j \gamma^{rc}\gamma_{jb}-\left(\varphi T\right)_{ab}^c\right)(p) \frac{\partial F}{\partial y^c}(v)=0,$$ where $a=1, \ldots, n$ and $v\in T_pM$. [*The previous argument is obviously working for any element of the group $G$ containing orthogonal transformations of the tangent space $T_pM$ with respect to the averaged Riemannian metric such that the Finslerian indicatrix is invariant: $F\circ \varphi=F$ ($\varphi\in G$). It is also clear that $G$ is a compact subgroup in the orthogonal group and $\textrm{Hol}_p \nabla \subset G$.*]{} \[smoothness\] If we have a connected generalized Berwald manifold then the mapping $p\in M\to A_p\subset \wedge^2 T_p^*M \otimes T_pM$ is a smooth affine distribution of constant rank of the torsion tensor bundle. Let $\ \nabla$ be a compatible linear connection, $T$ be its torsion tensor and the point $p\in M$ be given. According to Lemma \[lemmakey\] we have that $\displaystyle{A_q=T_q+\tau_{pq}(L_p)}$ for any $q\in M$, where $\tau_{pq}$ is the parallel transport along an arbitrary curve joining $p$ and $q$. In what follows we introduce the extremal compatible linear connection of a generalized Berwald manifold in terms of its torsion $T^0$ by taking the closest point $T^0_q \in A_q$ to the origin for any $q\in M$. The extremal compatible linear connection of a generalized Berwald manifold is the uniquely determined compatible linear connection minimizing the norm of its torsion by taking the values of the pointwise minima. A conditional extremum problem for the extremal compatible linear connection ---------------------------------------------------------------------------- Let a point $p\in M$ be given and consider the conditional extremum problem $$\label{condext} \min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad T_p \in A_p,$$ where the affine subspace $A_p\subset \wedge^2 T_p^*M \otimes T_pM$ defined by (\[condextb\]). First of all note that the coefficient of $T_{ab}^c$ ($1\leq a < b \leq n$, $c=1, \ldots, n$) is $$\label{coeff} \sigma_{c; i}^{ab}=\frac{1}{2}\left(\left(y^a \gamma^{br}-y^b\gamma^{ar}\right)\frac{\partial F}{\partial y^r}\gamma_{ic}+\left(\delta_i^a \gamma^{br}-\delta_i^b\gamma^{ar}\right)\frac{\partial F}{\partial y^r}y^j\gamma_{jc}-\left(\delta_i^a y^b-\delta_i^b y^a\right)\frac{\partial F}{\partial y^c}\right),$$ where the index $i=1, \ldots, n$ refers to the corresponding equation in (\[cond2\]). The symmetric differences in formula (\[coeff\]) is due to $a< b$. If the coordinate vector fields $\displaystyle{\partial/\partial u^1, \ldots, \partial/\partial u^n}$ form an orthonormal basis at $p\in M$ with respect to the averaged Riemannian metric $\gamma$, then $$\label{coeffort} \sigma_{c; i}^{ab}=\frac{1}{2}\left(\delta_i^c\left(y^a \frac{\partial F}{\partial y^b}-y^b\frac{\partial F}{\partial y^a}\right)+\delta_i^a \left(y^c\frac{\partial F}{\partial y^b}-y^b\frac{\partial F}{\partial y^c}\right)-\delta_i^b \left(y^c\frac{\partial F}{\partial y^a}-y^a\frac{\partial F}{\partial y^c}\right)\right).$$ Since the vector fields $$y^a \frac{\partial }{\partial y^b}-y^b\frac{\partial }{\partial y^a}, \ y^c\frac{\partial}{\partial y^b}-y^b\frac{\partial}{\partial y^c}, \ y^c\frac{\partial}{\partial y^a}-y^a\frac{\partial}{\partial y^c}$$ come from the Liouville vector field (the outer unit normal to the Finslerian indicatrix) by an Euclidean quarter rotation in the corresponding $2$-planes, their actions on $F$ are automatically zero at the contact points of the Finslerian and the Riemannian spheres. Vertical and horizontal contact points -------------------------------------- \[contactpoints\] The nonzero element $v\in T_pM$ satisfying $$\frac{\partial \log F}{\partial y^i}(v)=\frac{\partial \log F^*}{\partial y^i}(v) \ \ (i=1, \ldots, n)$$ is called a vertical contact point of the Finslerian and the averaged Riemannian metric functions. The nonzero element $v\in T_pM$ is a horizontal contact point of the Finslerian and the averaged Riemannian metric functions if $$X_i^{h^*} F(v)=0 \ \ (i=1, \ldots, n).$$ The tangent space at $p\in M$ is vertical/horizontal contact if all nonzero elements $v\in T_pM$ are vertical/horizontal contact. The vertical/horizontal contact vector fields can also be defined in a similar way: the vector field $X$ on the base manifold is vertical/horizontal contact if either $X(p)\in T_pM$ is vertical/horizontal contact or $X(p)={\bf 0}$. [*First of all note that the vertical contact points are independent of the choice of the coordinate system. Geometrically, the tangent hyperplanes of the Finslerian and the Riemannian spheres passing through a vertical contact point are the same in the corresponding tangent space. This is because their Euclidean gradient vectors in $T_pM$ are proportional:*]{} $$\label{vertcont1} \frac{\partial F}{\partial y^i}(v)=\frac{F}{F^*}(v)\frac{\partial F^*}{\partial y^i}(v) \ \ (i=1, \ldots, n),$$ [*where the coordinate vector fields $\displaystyle{\partial/\partial u^1, \ldots, \partial/\partial u^n}$ form an orthonormal basis at $p\in M$ with respect to the averaged Riemannian metric $\gamma$.*]{} \[contactpoints01\] The vertical contact points of a generalized Berwald manifold are horizontal contact points. Suppose that the coordinate vector fields $\displaystyle{\partial/\partial u^1, \ldots, \partial/\partial u^n}$ form an orthonormal basis at $p\in M$ with respect to the averaged Riemannian metric $\gamma$. It can be easily seen that $$\label{vertcont2} y^a \frac{\partial F^*}{\partial y^b}(v)-y^b\frac{\partial F^*}{\partial y^a}(v)=y^c\frac{\partial F^*}{\partial y^b}(v)-y^b\frac{\partial F^*}{\partial y^c}(v)=y^c\frac{\partial F^*}{\partial y^a}(v)-y^a\frac{\partial F^*}{\partial y^c}(v)=0.$$ Therefore, by formula (\[vertcont1\]), $$\label{vertcont3} y^a \frac{\partial F}{\partial y^b}(v)-y^b\frac{\partial F}{\partial y^a}(v)=y^c\frac{\partial F}{\partial y^b}(v)-y^b\frac{\partial F}{\partial y^c}(v)=y^c\frac{\partial F}{\partial y^a}(v)-y^a\frac{\partial F}{\partial y^c}(v)=0.$$ This means that $\sigma^{ab}_{c;i}(v)=0$ and the corresponding equations in (\[condextb\]) reduce to $\displaystyle{X_i^{h^*} F (v)=0}$ for any $i=1, \ldots, n$. \[contactpoints02\] Let $p\in M$ be a given point of a connected generalized Berwald manifold. If any nonzero element $v\in T_pM$ is a vertical contact point of the Finslerian and the averaged Riemannian metric functions, i.e. $T_pM$ is a vertical contact tangent space then we have a Riemannian manifold. If any nonzero element in $T_pM$ is a vertical contact point of the Finslerian and the averaged Riemannian metric functions then we have that $$\log \frac{F}{F^*} (v)=\textrm{const.}\ \ (v\in T_pM),$$ i.e. $\displaystyle{F(v)=e^{\textrm{const.}} F^*(v)}$ for any nonzero element $v\in T_pM$. This means that the Finslerian indicatrix is a quadratic hypersurface at a single point $p\in M$. Since we have linear parallel transports between different tangent spaces, the same is true for the Finslerian indicatrix at any point of the (connected) base manifold, i.e. we have a Riemannian manifold. \[contactpoints03\] Let $p\in M$ be a given point of a generalized Berwald manifold. If any nonzero element $v\in T_pM$ is a horizontal contact point of the Finslerian and the averaged Riemannian metric functions, i.e. $T_pM$ is a horizontal contact tangent space then the torsion of the extremal compatible linear connection is identically zero at the point $p\in M$. The statement is trivial because the zero element in $\wedge^2 T_p^*M \otimes T_pM$ solves the equations in (\[condextb\]) under the conditions $X_i^{h^*} F(v)=0$ $(i=1, \ldots, n)$ for any nonzero $v\in T_pM$. \[contactpoints04\][*If we have a classical Berwald manifold then any nonzero element $v\in TM$ is a horizontal contact point of the Finslerian and the averaged Riemannian metric functions and vice versa. The extremal compatible linear connection is $\nabla^*$ (the Lévi-Civita connection of the averaged Riemannian metric) with vanishing torsion.*]{} The reference element method ---------------------------- In what follows we are going to find the Lagrange multipliers for the torsion tensor of the extremal compatible linear connection. They transform the compatibility condition to a system of linear equations containing at most $n$ unknown parameters instead of $\displaystyle{\binom{n}{2}n}$. Let a point $p\in M$ and the reference element $v\in T_pM\setminus \{{\bf 0}\}$ be given and consider the conditional extremum problem $$\label{condextref} \min \frac{1}{2} \|T_p\|^2 \ \ \ \textrm{subject to} \ \ \ T_p\in A_p(v);$$ the affine subspace $A_p(v)\subset \wedge^2 T_p^*M \otimes T_pM$ is defined by $$\label{condextd} X_i^{h^*}F(v)+\frac{1}{2}y^j (v)\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)(p) \frac{\partial F}{\partial y^r}(v)=0$$ where $i=1, \ldots, n$. It is clear that $\displaystyle{A_p = \bigcap_{v\in T_pM\setminus \{{\bf 0}\}} A_p(v) \subset A_p(v)}$. Introducing the notations $$g_i(T_p, v):=X_i^{h^*}F(v)+\frac{1}{2}y^j(v)\left(T^{l}_{jk}\gamma^{kr}\gamma_{il}+T^{l}_{ik}\gamma^{kr}\gamma_{jl}-T_{ij}^r\right)(p) \frac{\partial F}{\partial y^r}(v) \quad (i=1, \ldots, n),$$ $$\textrm{\ rank}\ \frac{\partial g_i}{\partial T_{bc}^a}(T_p, v)=\left\{ \begin{array}{rl} n&\textrm{if $v$ is not a vertical contact point in $T_pM$}\\ 0&\textrm{otherwise}. \end{array} \right.$$ If $v$ is a vertical contact point of the Finslerian and the averaged Riemannian metric functions then $\sigma_{c; i}^{ab}(v)=0$ $(i=1, \ldots, n)$ because of (\[coeffort\]) - (\[vertcont3\]). Otherwise $$\begin{array}{|c|c|c|c|c|} \hline &&&&\\ a=c=i, & b=c=i, & a=c=i,&b=c=i,&a\neq i,\\ i< b, i=j& a< i, i=j& i< b, i\neq j&a < i, i\neq j&b\neq i, c\neq i\\ &&&&\\ \hline &&&&\\ \sigma_{i;i}^{i b}=y^i \frac{\partial F}{\partial y^b}-y^b\frac{\partial F}{\partial y^i}&\sigma_{i;i}^{a i}=y^a \frac{\partial F}{\partial y^i}-y^i\frac{\partial F}{\partial y^a}&\sigma_{i;j}^{i b}=0&\sigma_{i;j}^{a i}=0&\sigma_{c; i}^{a b}=0\\ &&&&\\ \hline \end{array}$$ and we can find at least one non-zero value among the coefficients $\sigma_{i;i}^{i b}(v)$, where $i< b$ and $\sigma_{i;i}^{a i}(v)$, where $a<i$ in each row[^1]. The subsequent rows contain zeros in the corresponding positions: $\displaystyle{\sigma_{i;i+1}^{i b}(v)=\sigma_{i;i+1}^{a i}(v)=\sigma_{i;i+2}^{i b}(v)=\sigma_{i;i+2}^{a i}(v)=\ldots=0}$. Let a point $p\in M$ of a connected generalized Berwald manifold be given. - If any nonzero element $v\in T_pM$ is a vertical contact point of the Finslerian and the averaged Riemannian metric functions, i.e. $T_pM$ is a vertical contact tangent space then we have a Riemannian manifold (see Corollary \[contactpoints02\]) and the extremal compatible linear connection is $\nabla^*$. Otherwise, let $p\in M$ be given and suppose that $v\in T_pM$ is not a vertical contact reference element. Using the notations in subsection 2.1 the conditional extremum problem can be written into the form $$\min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad X_i^{h^*}F(v)+\langle T_p, \sigma_i(v) \rangle=0 \quad (i=1, \ldots, n),$$ where $$\sigma_i=\sum_{1\leq a < b \leq n} \sum_{c=1}^n \sigma_{c; i}^{ab}\frac{\partial}{\partial y^a}\wedge \frac{\partial}{\partial y^b} \otimes dy^c$$ and the inner product on the vertical subspace $V_v \wedge^2 T^*_pM \otimes T_pM$ is induced by the Riemannian metric of the torsion tensor bundle: $$\langle T_p, \sigma_i(v) \rangle =\sum_{1\leq a < b \leq n} \sum_{c=1}^n T_{ab}^c(p) \sigma_{c; i}^{ab}(v);$$ note that the torsion tensor $T$ is identified with its vertically lifted tensor if necessary, i.e. $$T=\sum_{1\leq a < b \leq n} \sum_{c=1}^n T_{ab}^c \frac{\partial}{\partial u^a}\wedge \frac{\partial}{\partial u^b} \otimes du^c \rightleftharpoons \sum_{1\leq a < b \leq n} \sum_{c=1}^n T_{ab}^c\circ \pi \frac{\partial}{\partial y^a}\wedge \frac{\partial}{\partial y^b} \otimes dy^c.$$ The Lagrange method says that $$\frac{\mathcal{\partial L}(v)}{\partial T_{ab}^c}(T_p, \lambda_1(v), \ldots, \lambda_n(v))=0,$$ where $$\mathcal{L}(v)(T_p, \lambda_1(v), \ldots, \lambda_n(v))=\frac{1}{2}\|T_p\|^2-\sum_{j=1}^{n} \lambda_{j}(v)g_{j}(T_p, v)=$$ $$\frac{1}{2}\|T_p\|^2-\sum_{j=1}^{n} \lambda_{j}(v)\left(X_j^{h^*}F(v)+\langle T_p, \sigma_j(v) \rangle \right),$$ i.e. $$T_{ab}^c(p)-\sum_{j=1}^{n} \lambda_{j}(v)\sigma_{c; j}^{ab}(v)=0 \ \ \Rightarrow \ \ T_{ab}^c(p)=\sum_{j=1}^{n} \lambda_{j}(v)\sigma_{c; j}^{ab}(v).$$ Subtituting into the conditional equations we have that $$\label{multiplierseq} X_{i}^{h^*}F(v)+\sum_{j=1}^n \lambda_{j}(v)\sum_{1\leq a < b \leq n} \sum_{c=1}^n \sigma_{c; i}^{ab}(v)\sigma_{c; j}^{ab}(v)=0 \quad (i=1, \ldots, n)$$ and $$\label{formula01} -\mathcal{G}^{-1}\left(\sigma_1(v), \ldots, \sigma_n(v)\right)\left(X_{1}^{h^*}F(v), \ldots, X_n^{h^*}F(v)\right)^T=\left(\lambda_1(v), \ldots, \lambda_n(v)\right)^T,$$ where $i=1, \ldots, n$ and $\displaystyle{\mathcal{G}\left(\sigma_1(v), \ldots, \sigma_n(v)\right)}$ is the Gramian with respect to the inner product $$\langle \sigma_i(v), \sigma_j(v)\rangle=\sum_{1\leq a < b \leq n} \sum_{c=1}^n \sigma_{c; i}^{ab}(v)\sigma_{c; j}^{ab}(v)$$ on the vertical subspace $V_v \wedge^2 T_p^*M \otimes T_pM$. The uniquely determined solutions $\lambda_{1}(v), \ldots, \lambda_n(v)$ give the closest element $T_p^0(v)\in A_p(v)$ to the origin in the linear space $\wedge^2 T_p^*M \otimes T_pM$. Therefore $$A_p(v)=T_p^0(v)+\mathcal{L}^{\bot}\left(\sigma_1(v), \ldots, \sigma_n(v)\right),$$ where $$\label{formula02} T_p^0(v)=\sum_{j=1}^n \lambda_j(v)\sigma_j(v)$$ and the coefficients are given by formula (\[formula01\]). Since $$A_p=\bigcap_{v\ \in \ T_pM\setminus \{{\bf 0}\} \ \textrm{is not a vertical contact point}}T_p^0(v)+\mathcal{L}^{\bot}\left(\sigma_1(v), \ldots, \sigma_n(v)\right)$$ it follows that $$\dim A_p \leq \dim A_p(v)=\binom{n}{2}n-n,$$ where $v$ is not a vertical contact element. Note that the intersection can be taken with respect to Finslerian (or Riemannian) unit vectors because of the homogeneity properties: $$A_p=\bigcap_{v\ \in \ \partial K_p \ \textrm{is not a vertical contact point}}T_p^0(v)+\mathcal{L}^{\bot}\left(\sigma_1(v), \ldots, \sigma_n(v)\right)=$$ $$\ \ \ \ \ \ \bigcap_{v\ \in \ \partial K^*_p \ \textrm{is not a vertical contact point}}T_p^0(v)+\mathcal{L}^{\bot}\left(\sigma_1(v), \ldots, \sigma_n(v)\right).$$ - In case of a horizontal but not vertical contact reference element $T_p^0(v)={\bf 0}\in \wedge^2 T_p^*M \otimes T_pM$ because of Definition \[contactpoints\] and formula (\[formula01\]). If any nonzero element $v\in T_pM$ is a horizontal contact point of the Finslerian and the averaged Riemannian metric functions, i.e. $T_pM$ is a horizontal contact tangent space then the solution of the extremum problem (\[condextref\]) is $T_p^0:={\bf 0}$ independently of the reference elements (see Corollary \[contactpoints03\]). - Otherwise, let $p\in M$ be given and suppose that $v\in T_pM$ is not a horizontal contact reference element, i.e. $T_p^0(v)\neq {\bf 0}$. Since $A_p\subset A_p(v)$ it follows that $A_p$ consisting of the solutions of the conditional equation without any reference element must be contained in the hyperplane $$\label{hyperplane} \langle T_p^0(v), T_p-T_p^0(v)\rangle=0$$ of dimension $\displaystyle{\binom{n}{2}n-1}$ in $\wedge^2 T_p^*M \otimes T_pM$. Therefore the process in case (C) can be completed as follows: - Let $T^0_p(v_1)$ be the uniquely determined, not identically zero solution of the conditional extremum problem $$\label{condext:01} \min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad T_p\in A_p(v_1),$$ where $v_1\in T_pM$ is not a horizontal contact point. If $g_i(T^0_p(v_1), v)=0$ ($i=1, \ldots, n$) for any $v\in T_pM$ then $T^0_p=T_p^0(v_1)$ and we are done. - Otherwise, let us choose a nonzero element $v_2$ such that $g_i(T^0_p(v_1), v_2)\neq 0$ for at least one of the indices $i=1, \ldots, n$. Taking $T_p^0(v_1)$ as the origin, let $T_p^0(v_1, v_2)$ be the uniquely determined, not identically zero solution of the conditional extremum problem $$\label{condext:02} \min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad T_p \in A_p(v_2)$$ and $$\label{extracond:01} \langle T_p^0(v_1), T_p-T_p^0(v_1) \rangle=0.$$ Geometrically, $T_p^0(v_1, v_2)$ is the orthogonal projection of $T_p^0(v_1)$ onto $A_p(v_2)$ in the hyperplane (\[extracond:01\]) of dimension $\displaystyle{\binom{n}{2}n-1}$. If $$\mathcal{G}\left(\sigma_1(v_2), \ldots, \sigma_n(v_2), T_p^0(v_1)\right)=0$$ then $T_p^0(v_1, v_2)=T_p^0(v_2)$. Otherwise $$\label{formula03} -\mathcal{G}^{-1}\left(\sigma_1(v_2), \ldots, \sigma_n(v_2), T_p^0(v_1)\right)\left(X_{1}^{h^*}F(v_2), \ldots, X_n^{h^*}F(v_2), -\|T_p^0(v_1)\|^2\right)^T=$$ $$\left(\lambda_1(v_2), \ldots, \lambda_n(v_2), \lambda_{n+1}(v_2)\right)^T.$$ Especially, equation (\[extracond:01\]) is equivalent to $$\lambda_{n+1}(v_2)=1-\sum_{j=1}^n \lambda_j(v)\frac{\langle \sigma_j(v_2), T^0_p(v_1)\rangle}{\|T^0_p(v_1)\|^2}.$$ If $g_i(T_p^0(v_1, v_2), v)=0$ ($i=1, \ldots, n$) for any $v\in T_pM$ then $T^0_p=T_p^0(v_1, v_2)$ and we are done. - Otherwise, let us choose a nonzero element $v_3$ such that $g_i(T^0_p(v_1, v_2), v_3)\neq 0$ for at least one of the indices $i=1, \ldots, n$. Taking $T_p^0(v_1, v_2)$ as the origin, let $T_p^0(v_1, v_2, v_3)$ be the uniquely determined, not identically zero solution of the conditional extremum problem $$\label{condext:03} \min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad T_p\in A_p(v_3)$$ and $$\label{extracond:02} \langle T_p^0(v_1), T_p-T_p^0(v_1, v_2)\rangle=0, \ \langle T_p^0(v_1, v_2), T_p-T_p^0(v_1, v_2)\rangle=0.$$ Geometrically, $T_p^0(v_1)$, $T_p^0(v_1, v_2)$ and $T_p^0(v_1, v_2, v_3)$ form an orthogonal chain in the sense that $$T_p^0(v_1) \ \bot\ \ T_p^0(v_1, v_2)-T_p^0(v_1) \ \textrm{and} \ \ T_p^0(v_1), \ T_p^0(v_1, v_2)-T_p^0(v_1) \ \bot \ T_p^0(v_1, v_2, v_3)-T_p^0(v_1, v_2).$$ Therefore $T_p^0(v_1, v_2, v_3)$ is the element of a hyperplane of dimension $\displaystyle{\binom{n}{2}n-2}$. In general, if we have $T_p^0(v_1), \ldots, T_p^0(v_1, \ldots, v_{m-1})$ then - $T_p^0=T_p^0(v_1, \ldots, v_{m-1})$ in case of $g_i(T_p^0(v_1, \ldots, v_{m-1}),v)=0$ ($i=1, \ldots, n$) for any $v\in T_pM$. - Otherwise, let us choose a nonzero element $v_m$ such that $g_i(T_p^0(v_1, \ldots, v_{m-1}), v_m)\neq 0$ for at least one of the indices $i=1, \ldots, n$. Taking $T_p^0(v_1, \ldots, v_{m-1})$ as the origin, let $T_p^0(v_1, \ldots, v_m)$ be the uniquely determined, not identically zero solution of the conditional extremum problem $$\label{condext:0m} \min \frac{1}{2} \|T_p\|^2 \quad \textrm{subject to} \quad T_p\in A_p(v_m)$$ and $$\label{extracond:0m2} \langle T_p^0(v_1), T_p-T_p^0(v_1, \ldots, v_{m-1})\rangle=0, \ \langle T_p^0(v_1, v_2), T_p-T_p^0(v_1, \ldots, v_{m-1})\rangle=0, \ \ldots,$$ $$\langle T_p^0(v_1, \ldots, v_{m-1}), T_p-T_p^0(v_1, \ldots, v_{m-1})\rangle=0.$$ \[alg\] For any $m\leq l$ $$T_p^0(v_1), \ T_p^0(v_1, v_2)-T_p^0(v_1), \ \ldots,$$ $$T_p^0(v_1, \ldots, v_{m-1})-T_p^0(v_1, \ldots, v_{m-2}) \ \bot \ T_p^0(v_1, \ldots,v_m)-T_p^0(v_1, \ldots, v_{m-1}),$$ i.e. $T_p^0(v_1), \ldots, T_p^0(v_1, \ldots,v_m)$ is an orthogonal chain, where $\displaystyle{l\leq \binom{n}{2}n}$ is its maximal length and $T_p^0=T_p^0(v_1, \ldots, v_l)$ is the torsion tensor of the extremal compatible linear connection. Let $M$ be a connected manifold equipped with a Finslerian metric $F$. It is a generalized Berwald manifold if and only if it is a classical Berwald manifold including the case of the Riemannian manifolds or - the vertical contact elements of the tangent manifold $TM$ are horizontal contact, - for any non-horizontal contact tangent space $T_pM$, the form $$T^0_p(v_1, \ldots, v_l)\in \wedge^2 T_p^*M \otimes T_pM$$ is independent of the choice of the non-horizontal contact elements, where $\displaystyle{l\leq \binom{n}{2}n}$ is the maximal length of the orthogonal chain $\displaystyle{T_p^0(v_1), T_p^0(v_1, v_2), \ldots, T_p(v_1, v_2, \ldots, v_l)}$, - $\displaystyle{T^0\colon p\in M\to T^0_p:=\left\{\begin{array}{rl} {\bf 0}&\textrm{in case of a horizontal contact tangent space,}\\ T^0_p(v_1, \ldots, v_l)&\textrm{otherwise} \end{array} \right.}$ is a continuous section of the torsion tensor bundle and the corresponding linear connection is compatible to the Finslerian metric. Suppose that we have a connected generalized Berwald manifold. If it is not a classical Berwald manifold including the case of the Riemannian manifolds then we can refer to Corollary \[contactpoints01\]. On the other hand, $$T^0\colon p\in M\to T^0_p:=\left\{\begin{array}{rl} {\bf 0}&\textrm{in case of a horizontal contact tangent space,}\\ T^0_p(v_1, \ldots, v_l)&\textrm{otherwise} \end{array} \right.$$ is the torsion of the extremal compatible linear connection. Using Corollary \[smoothness\], it is a continuous section of the torsion tensor bundle. Conversely, suppose that we have a non-Riemannian and non-classical Berwald Finsler manifold such that the vertical contact elements of the tangent manifold $TM$ are horizontal contact. Then the compatibility equations are satisfied at each vertical contact element. Otherwise we can get the algorithm started by formulating the conditional extremum problem (\[condextref\]) for any non vertical contact $v\in TM\setminus \{{\bf 0}\}$. The algorithm gives a uniquely determined output due to the independency of the form $$T^0_p(v_1, \ldots, v_l)\in \wedge^2 T_p^*M \otimes T_pM$$ of the choice of the non horizontal contact elements $v_1, \ldots, v_l\in T_pM$ ($p\in M$). Finally, if the section is continuous then we have a compatible linear connection to the Finslerian metric with continuous connection coefficients because of (\[torsion\]). Using parallel translations with respect to such a connection we can conclude that the Finsler metric is monochromatic, i.e. there exists a linear mapping preserving the Finslerian norm of tangent vectors between the tangent spaces $T_pM$ and $T_qM$ for any pair of points $p, q\in M$. By the fundamental result of the theory [@BM] it is sufficient and necessary for a Finsler metric to be a generalized Berwald metric. Summary ======= The idea of the paper is to find a distinguished compatible linear connection of a (connected) generalized Berwald manifold. Generalized Berwald manifolds are Finsler manifolds admitting linear connections such that the parallel transports preserve the Finslerian length of tangent vectors (compatibility condition). By the fundamental result of the theory [@V5] such a linear connection must be metrical with respect to the averaged Riemannian metric given by integration of the Riemann-Finsler metric on the indicatrix hypersurfaces. The averaged Riemannian metric induces Riemannian metrics on the torsion tensor bundle $\wedge^2 T^*M \otimes TM$ and its vertical bundle in a natural way. Therefore we are looking for the extremal compatible linear connection in the sense that we want to minimize the norm of the torsion point by point. It is a conditional extremum problem involving functions defined on a local neighbourhood of the tangent manifold. In case of a given point of the manifold, the reference element method helps us to start the application of the Lagrange multiplier rule. The finiteness of the rank of the torsion tensor bundle provides that we can compute the minimizer independently of the reference elements in finitely many steps at each point of the manifold. The pointwise solutions constitute a continuous section of the torsion tensor bundle. The continuity of the components of the torsion tensor implies the continuity of the connection parameters. Using parallel translations with respect to such a connection we can conclude that the Finsler metric is monochromatic. By the fundamental result of the theory [@BM] it is sufficient and necessary for a Finsler metric to be a generalized Berwald metric. Therefore we have an intrinsic algorithm to check the existence of compatible linear connections on a Finsler manifold because it is equivalent to the existence of the extremal compatible linear connection. [99]{} N. Bartelme[ß]{} and V. Matveev, Monochromatic metrics are generalized Berwald, J. Diff. Geom. Appl. [**58**]{} (2018), pp. 264-271. Cs. Vincze, *A new proof of Szabó’ s theorem on the Riemann-metrizability of Berwald manifolds*, J. AMAPN, 21 (2005), pp. 199-204. Cs. Vincze, *On a scale function for testing the conformality of Finsler manifolds to a Berwald manifold*, Journal of Geometry and Physics. 54 (2005), 454-475. Cs. Vincze, *Generalized Berwald manifolds with semi-symmetric linear connections*, Publ. Math. Debrecen 83 (4) (2013), pp. 741-755. Cs. Vincze, [*On Randers manifolds with semi-symmetric compatible linear connections*]{}, Indagationes Mathematicae [**[26]{}**]{} (2), 2014, pp. 363 - 379. Cs. Vincze, *On a special type of generalized Berwald manifolds: semi-symmetric linear connections preserving the Finslerian length of tangent vectors*, “Finsler geometry: New methods and Perspectives”, European Journal of Mathematics, December 2017, Volume 3, Issue 4, pp. 1098 - 1171 . V. Wagner, *On generalized Berwald spaces*, CR Dokl. Acad. Sci. USSR (N.S.) [**39**]{} (1943), pp. 3-5. [^1]: Suppose, in contrary, that (for example) $i=1$ and $\sigma_{1;1}^{1 2}=\ldots=\sigma_{1;1}^{1 n}=0,$ i.e. $\displaystyle{v^1\frac{\partial F}{\partial y^b}(v)-v^b \frac{\partial F}{\partial y^1}(v)=0}$, where $b=2, \ldots, n$. Therefore $$v^1\sum_{b=2}^n v^b \frac{\partial F}{\partial y^b}(v)-\frac{\partial F}{\partial y^1}(v)\sum_{b=2}^n \left(v^b\right)^2=0$$ and the homogeneity property of the metric function implies that $$v^1\left(F(v)-v^1\frac{\partial F}{\partial y^1}(v)\right)-\frac{\partial F}{\partial y^1}(v)\left(\left(F^*(v)\right)^2-\left(v^1\right)^2\right)=0 \ \ \Rightarrow\ \ \frac{1}{F(v)}\frac{\partial F}{\partial y^1}(v)=\frac{v^1}{\left(F^*(v)\right)^2}=\frac{1}{F^*(v)}\frac{\partial F^*}{\partial y^1}(v).$$

--- abstract: 'In this paper, we consider the problem of choosing disks (that we can think of as corresponding to wireless sensors) so that given a set of input points in the plane, there exists no path between any pair of these points that is not intercepted by some disk. We try to achieve this separation using a minimum number of a given set of unit disks. We show that a constant factor approximation to this problem can be found in polynomial time using a greedy algorithm. To the best of our knowledge we are the first to study this optimization problem.' author: - '[Matt Gibson]{}' - '[Gaurav Kanade]{}' - '[Kasturi Varadarajan]{}' bibliography: - 'barcov.bib' title: On Isolating Points using Disks --- Introduction ============ Wireless sensors are being extensively used in applications to provide barriers as a defense mechanism against intruders at important buildings, estates, national borders etc. Monitoring the area of interest by this type of coverage is called *barrier* coverage [@KumarLA05]. Such sensors are also being used to detect and track moving objects such as animals in national parks, enemies in a battlefield, forest fires, crop diseases etc. In such applications it might be prohibitively expensive to attain blanket coverage but sufficient to ensure that the object under consideration cannot travel too far before it is detected. Such a coverage is called *trap* coverage [@BalisterZKS09; @SankararamanERT09]. Inspired by such applications, we consider the problem of isolating a set of points by a minimum-size subset of a given subset of unit radius disks. A unit disk crudely models the region sensed by a sensor, and the work reported here readily generalizes to disks of arbitrary, different radii. [c@c]{} & \ (a) & (b) #### Problem Formulation. The input to our problem is a set ${I}$ of $n$ unit disks, and a set $P$ of $k$ points such that ${I}$ [*separates*]{} $P$, that is, for any two points $p,q \in P$, every path between $p$ and $q$ intersects at least one disk in ${I}$. The goal is to find a minimum cardinality subset of ${I}$ that separates $P$. See Figure \[fig:separate\] for an illustration of this notion of separation. There has been a lot of recent interest on geometric variants of well-known NP-hard combinatorial optimization problems, and our work should be seen in this context. For several variants of the geometric set cover problem, for example, approximation algorithms have been designed [@ClarksonV05; @AronovES10; @MustafaR09] that improve upon the best guarantees for the combinatorial set cover problem. For the problem of covering points by the smallest subset of a given set of unit disks, we have approximation algorithms that guarantee an $O(1)$ approximation and even a PTAS [@bg95; @MustafaR09]. These results hold even for disks of arbitrary radii. Our problem can be viewed as a set cover problem where the elements that need to be covered are not points, but paths. However, known results only imply a trivial $O(n)$ approximation when viewed through this set cover lens. Another example of a problem that has received such attention is the [*independent set*]{} problem. For many geometric variants [@ChalermsookC09; @ChanH09; @FP], approximation ratios that are better than that for the combinatorial case are known. Our problem is similar to the node multi-terminal cut problem in graphs [@GargVY04]. Here, we are given a graph $G = (V,E)$ with costs on the vertices and a subset $U \subseteq V$ of $k$ vertices, and our goal is to compute a minimum cost subset of vertices whose removal disconnects every pair of vertices in $U$. This problem admits a poly-time algorithm that guarantees an $O(1)$ approximation. We note however that the problem we consider does not seem to be a special case of the multi-terminal cut problem. #### Contribution and Related Work. Our main result is a polynomial time algorithm that guarantees an $O(1)$ approximation for the problem. To the best of our knowledge, this is the first non-trivial approximation algorithm for this problem. Our algorithm is simple and combinatorial and is in fact a greedy algorithm. We first present an $O(1)$ approximation for the following two-point separation problem. We are given a set of unit disks $G$, and two points $s$ and $t$, and we wish to find the smallest subset $B \subseteq G$ so that $B$ separates $s$ and $t$. Our greedy algorithm to the overall problem applies the two-point separation algorithm to find the cheapest subset $B$ of $I$ that separates some pair of points in $P$. Suppose that $P$ is partitioned into sets $P_1,P_2,\ldots,P_{\tau}$ where each $P_i$ is the subset of points in the same “face” with respect to $B$. The algorithm then recursively finds a separator for each of the $P_i$, and returns the union of these and $B$. The analysis to show that this algorithm has the $O(1)$ approximation guarantee relies on the combinatorial complexity of the boundary of the union of disks. It uses a subtle and global argument to bound the total size of all the separators $B$ computed in each of the recursive calls.[^1] Our approximation algorithm for the two-point separation problem, which is a subroutine we use in the overall algorithm, is similar to fast algorithms for finding minimum $s$-$t$ cuts in undirected planar graphs, see for example [@Reif83]. Our overall greedy algorithm has some resemblance to the algorithm of Erickson and Har-Peled [@EricksonH04] employed in the context of approximating the minimum cut graph of a polyhedral manifold. The details of the our algorithm and the analysis, however, are quite different from these papers since we do not have an embedded graph but rather a system of unit disks. Sankararaman et al. [@SankararamanERT09] investigate a notion of coverage which they call [*weak coverage*]{}. Given a region ${{\cal R}}$ of interest (which they take to be a square in the plane) and a set $I$ of unit disks (sensors), the region is said to be $k$-weakly covered if each connected component of ${{\cal R}}- \bigcup_{d \in I} d$ has diameter at most $k$. They consider the situation when a given set $I$ of unit disks [*completely*]{} covers ${{\cal R}}$, and address the problem of partitioning $I$ into as many subsets as possible so that ${{\cal R}}$ is $k$-weakly covered by every subset. Their work differs in flavor from ours mainly due to the assumption that $I$ completely covers ${{\cal R}}$. #### Organization. In Section \[sec:prelims\], we discuss standard notions we require, and then reduce our problem to the case where none of the points in input $P$ are contained in any of the input disks. In Section \[sec:two-point\], we present our approximation algorithm for separating two points. In Section \[sec:improved\], we describe our main result, the constant factor approximation algorithm for separating $P$. We conclude in Section \[sec:conclusions\] with some remarks. Preliminaries {#sec:prelims} ============= We will refer to the standard notions of vertices, edges, and faces in arrangements of circles [@AgarwalS98]. In particular, for a set $R$ of $m$ disks, we are interested in the faces in the complement of the union of the disks in $R$. These are the connected components of the set $\Re^2 - \bigcup_{d \in R} d$. We also need the combinatorial result that the number of these faces is $O(m)$. Furthermore, the total number of vertices and edges on the boundaries of all these faces, that is, the combinatorial complexity of the boundary of the union of disks in $R$, is $O(m)$ [@AgarwalS98]. We make standard general position assumptions about the input set ${I}$ of disks in this article. This helps simplify the exposition and is without loss of generality. \[claim:many\] Let $R$ be a set of disks in the plane, and $Q$ a set of points so that (a) no point from $Q$ is contained in any disk from $R$, and (b) no face in the complement of the union of the disks in $R$ contains more than one point of $Q$. Then $|R| = \Omega(|Q|)$. The number of faces in the in the complement of the union of the disks in $R$ is $O(|R|)$. #### Covering vs. Separating. The input to our problem is a set ${I}$ of $n$ unit disks, and $P$ a set of $k$ points such that ${I}$ separates $P$. Let $P_c \subseteq P$ denote those points contained in some disk in $I$; and $P_s$ denote the remaining points. We compute an $\alpha$-approximation to the smallest subset of ${I}$ that covers $P_c$ using a traditional set-cover algorithm; there are several poly-time algorithms that guarantee that $\alpha = O(1)$. We compute a $\beta$-approximation to the smallest subset of ${I}$ that separates $P_s$, using the algorithm developed in the rest of this article. We argue below that the combination of the two solutions is an $O(\alpha + \beta)$ approximation to the smallest subset of ${I}$ that separates $P$. Let ${OPT}\subseteq {I}$ denote an optimal subset that separates $P$. Suppose that ${OPT}$ covers $k_1$ of the points in $P_c$ and let $k_2 = |P_c| - k_1$. By Lemma \[claim:many\], $|{OPT}| = \Omega(k_2)$. Now, by picking one disk to cover each of the $k_2$ points of $P_c$ not covered by ${OPT}$, we see that there is a cover of $P_c$ of size at most $|{OPT}| + k_2 = O(|{OPT}|)$. Thus, our $\alpha$-approximation has size $O(\alpha) \cdot |{OPT}|$. Since ${OPT}$ also separates $P_s$, our $\beta$-approximation has size $O(\beta) \cdot |{OPT}|$. Thus the combined solution has size $O(\alpha + \beta) \cdot |{OPT}|$. In the rest of the article, we abuse notation and assume that no point in the input set $P$ is contained in any disk in ${I}$, and describe a poly-time algorithm that computes an $O(1)$-approximation to the optimal subset of ${I}$ that separates $P$. Separating Two Points {#sec:two-point} ===================== Let $s$ and $t$ be two points in the plane, and $G$ a set of disks such that no disk in $G$ contains either $s$ or $t$, but $G$ separates $s$ and $t$. See Figure \[fig:fig1\]. Our goal is to find the smallest cardinality subset $B$ of $G$ that separates $s$ and $t$. We describe below a polynomial time algorithm that returns a constant factor approximation to this problem. [c@c]{} & \ (a) & (b) Without loss of generality, we may assume that the intersection graph of $G$ is connected. (Otherwise, we apply the algorithm to each connected component for which the disks in the component separate $s$ and $t$. We return the best solution obtained.) Let $f_s$ and $f_t$ denote the faces containing $s$ and $t$, respectively, in the arrangement of $G$. We augment the intersection graph of $G$ with vertices corresponding to $s$ and $t$, and add an edge from $s$ to each disk that contributes an edge to the boundary of the face $f_s$, and an edge from $t$ to each disk that contributes an edge to the boundary of the face $f_t$. We assign a cost of $0$ to $s$, $t$, and a cost of $1$ to each disk in $G$. We then find the shortest path from $s$ to $t$ in this graph, where the length of a path is the number of the [*vertices*]{} on it that correspond to disks in $G$. Let $\sigma$ denote the sequence of disks on this shortest path. Note that any two disks that are not consecutive in $\sigma$ do not intersect. Using $\sigma$, we compute a path $\pi$ in the plane, as described below, from $s$ to $t$ so that (a) there are points $s'$ and $t'$ on $\pi$ so that the portion of $\pi$ from $s$ to $s'$ is in $f_s$, and the portion from $t'$ to $t$ is in $f_t$; (b) every point on $\pi$ from $s'$ to $t'$ is contained in some disk from $\sigma$; (c) the intersection of $\pi$ with each disk in $\sigma$ is connected. See Figure \[fig:fig1\]. Suppose that the sequence of disks in $\sigma$ is $d_1, \ldots, d_{|\sigma|}$. Let $s'$ (resp. $t'$) be a point in $d_1$ (resp. $d_{\sigma}$) that lies on the boundary of $f_s$ (resp. $f_t$). For $1 \leq i \leq |\sigma| - 1$, choose $x_i$ to denote an arbitrary point in the intersection of $d_i$ and $d_{i+1}$. The path $\pi$ is constructed as follows: Take an arbitrary path from $s$ to $s'$ that lies within $f_s$, followed by the line segments $\overline{s' x_1}, \overline{x_1 x_2}, \ldots, \overline{x_{|\sigma| -2} x_{|\sigma| - 1}}, \overline{x_{|\sigma|-1}t'}$, followed by an arbitrary path from $t'$ to $t$ that lies within $f_t$. Properties (a) and (b) hold for $\pi$ by construction. Property (c) is seen to follow from the fact that disks that are not consecutive in $\sigma$ do not overlap. Notice that $\pi$ “cuts” each disk in $\sigma$ into two pieces. (Formally, the removal of $\pi$ from any disk in $\sigma$ yields two connected sets.) The path $\pi$ may also intersect other disks and cut them into two or more pieces, and we refer to these pieces as disk pieces. For a disk that $\pi$ does not intersect, there is only one disk piece, which is the disk itself. We consider the intersection graph $H$ of the disk pieces that come from disks in $G$. Observe that a disk piece does not have points on $\pi$, since $\pi$ is removed; so two disk pieces intersecting means there is a point outside $\pi$ that lies in both of them. In this graph, each disk piece has a cost of $1$. In this graph $H$, we compute, for each disk $d \in \sigma$, the shortest path between the two pieces corresponding to $d$. Suppose $d' \in \sigma$ yields the overall shortest path $\sigma'$; let $D$ denote the set of disks that contribute a disk piece to this shortest path. Our algorithm returns $D$ as its computed solution. See Figure \[fig:fig2\]. We note that $D$ separates $s$ and $t$ – in particular, the union of the disk pieces in $\sigma'$ and the set $\pi \cap d'$ contains a cycle in the plane that intersects the path $\pi$ between $s$ and $t$ exactly once. Bounding the Size of the Output ------------------------------- Let $B^*$ denote the smallest subset of $G$ that separates $s$ and $t$. We will show that $|D| = O(|B^*|)$. Let $f^*$ denote the face containing $s$ in the arrangement of $B^*$. Due to the optimality of $B^*$, we may assume that the boundary of $f^*$ has only one component. Let $a$ (resp. $b$) denote the first (resp. last) point on path $\pi$ where $\pi$ leaves $f^*$. It is possible that $a = b$. We find a minimum cardinality contiguous subsequence $\overline{\sigma}$ of $\sigma$ that contains the subpath of $\pi$ from $a$ to $b$; let $d_a$ and $d_b$ denote the first and last disks in $\overline{\sigma}$. See Figure \[fig:fig4\]. [c@c]{} & \ (a) & (b) We claim that $|\overline{\sigma}| \leq |B^*| + 2$; if this inequality does not hold, then we obtain a contradiction to the optimality of $\sigma$ by replacing the disks in the $\overline{\sigma} \setminus \{d_a, d_b\}$ by $B^*$. Consider the face $f$ containing $s$ in the arrangement with $B^* \cup \overline{\sigma}$. Each edge that bounds this face comes from a single disk piece, except for one edge corresponding to $d_a$ that may come from two disk pieces. (This follows from the fact that the portion of $\pi$ between $a$ and $b$ is covered by the disks in $\overline{\sigma}$.) These disk pieces induce a path in $H$ in between the two pieces from $d_a$, and their cost therefore upper bounds the cost of $D$. We may bound the cost of these disk pieces by the number of edges on the boundary of $f$ (with respect to $B^* \cup \overline{\sigma}$). The number of such edges is $O(|B^*|+|\overline{\sigma}|)=O(|B^*|)$. \[thm:two-point\] Let $s$ and $t$ be two points in the plane, and $G$ a set of disks such that no disk in $G$ contains either $s$ or $t$, but $G$ separates $s$ and $t$. There is a polynomial time algorithm that takes such $G$, $s$, and $t$ as input, and outputs a subset $B \subseteq G$ that separates $s$ and $t$; the size of $B$ is at most a multiplicative constant of the size of the smallest subset $B^* \subseteq G$ that separates $s$ and $t$. Separating Multiple Points {#sec:improved} ========================== We now present a polynomial time algorithm that yields an $O(1)$ approximation to the problem of finding a minimum subset of ${I}$ that separates the set $P$ of points. The algorithm is obtained by calling ${\mbox{recSep}}(P)$, where ${\mbox{recSep}}(Q)$, for any $Q \subseteq P$ is the following recursive procedure: 1. If $|Q| \leq 1$, return $\emptyset$. 2. For every pair of points $s,t \in Q$, invoke the algorithm of Theorem \[thm:two-point\] (with $G \leftarrow {I}$) to find a subset $B_{s,t} \subseteq {I}$ such that $B_{s,t}$ separates $s$ and $t$. 3. Let $B$ denote the minimum size subset $B_{s,t}$ over all pairs $s$ and $t$ considered. 4. Consider the partition of $Q$ into subsets so that each subset corresponds to points in the same face (with respect to $B$). Suppose $Q_1, \ldots, Q_{\tau}$ are the subsets in this partition. Note that $\tau \geq 2$, since $B$ separates some pair of points in $Q$. 5. Return $B \cup \bigcup_{j=1}^{\tau} {\mbox{recSep}}(Q_j)$. Clearly, ${\mbox{recSep}}(P)$ yields a separator for $P$. To bound the size of this separator, let us define a set ${\rm \bf \cal{Q}}$ that contains as its element any $Q \subseteq P$ such that $|Q| \geq 2$ and ${\mbox{recSep}}(Q)$ is called somewhere within the call to ${\mbox{recSep}}(P)$. For any $Q \in {\rm \bf \cal{Q}}$, define $B_Q$ to be the set $B$ that is computed in the body of the call to ${\mbox{recSep}}(Q)$. Notice that ${\mbox{recSep}}(P)$ returns $\cup_{Q \in {\rm \bf \cal{Q}}} B_Q$. Now we “charge” each such $B_Q$ to an arbitrary point within $p_Q \in Q$ in such a way that no point in $P$ is charged more than once. A moment’s thought reveals that this is indeed possible. (In a tree where each interval node has degree at least $2$, the number of leaves is greater than the number of internal nodes.) Let ${OPT}$ denote the optimal separator for $P$ and let $F_Q \subseteq {OPT}$ denote the disks that contribute to the boundary of the face (in the arrangement of ${OPT}$) containing $p_Q$. We claim that $|B_Q| = O(|F_Q|)$; indeed $F_Q$ separates $p_Q \in Q$ from any point in $P$, and thus any point in $Q$. Thus for any $t \in Q \setminus \{p_Q\}$, we have $|B_Q| \leq B_{p_Q,t} = O(|F_Q|)$. We thus have $$\bigcup_{Q \in {\rm \bf \cal{Q}}} |B_Q| \leq \sum_{Q \in {\rm \bf \cal{Q}}} O(|F_Q|) = O(|{OPT}|),$$ where the last equality follows from union complexity. We have derived the main result of this paper: \[thm: main\] Let ${I}$ be a set of $n$ unit disks and $P$ a set of $k$ points such that ${I}$ separates $P$. There is a polynomial time algorithm that takes as input such ${I}$ and $P$, and returns a subset ${O}\subseteq {I}$ of disks that also separates $P$, with the guarantee that $|{O}|$ is within a multiplicative $O(1)$ of the smallest subset of ${I}$ that separates $P$. Conclusions {#sec:conclusions} =========== We have a presented an $O(1)$-approximation algorithm for finding the minimum subset of a given set ${I}$ of disks that separates a given set of points $P$. One way to understand our contribution is as follows. Suppose we had at our disposal an efficient algorithm that optimally separates a single point $p \in P$ from every other point in $P$. Then applying this algorithm for each point in $P$, we get a separator for $P$. That the size of this separator is within $O(1)$ of the optimal is an easy consequence of union complexity. However, we only have at our disposal an efficient algorithm for a weaker task: that of approximately separating two given points in $P$. What we have shown is that even this suffices for the task of obtaining an $O(1)$ approximation to the overall problem. It is easy to see that our algorithm and the approximation guarantee generalize, for example, to the case when the disks have arbitrary and different radii. #### Acknowledgements. We thank Alon Efrat for discussions that led to the formulation of the problem, and Sariel Har-Peled for discussions that led to the algorithm described here. [^1]: In the earlier version of this paper, a similar algorithm was analyzed in a more “local” fashion. The basic observation was that the very first separator $B$ that is computed has size $O(|{OPT}|/k)$, where ${OPT}$ is the optimal solution for the problem. Subsequent separators computed in the recursive calls may be more expensive, but it was shown that the overall size is $O((\log k)\cdot |{OPT}|)$. In contrast, the present analysis does not try to bound the size of the individual separators, but just the sum of their sizes. As a consequence, the analysis also turns out to be technically simpler.

--- abstract: 'We extend known results regarding the maximum rectilinear crossing number of the [*cycle graph*]{} ($C_n$) and the [*complete graph*]{} ($K_n$) to the class of general [*d-regular graphs*]{} $R_{n,d}$. We present the [*generalized star*]{} drawings of the $d$-regular graphs $S_{n,d}$ of order $n$ where $n+d\equiv 1 \pmod 2 $ and prove that they maximize the maximum rectilinear crossing numbers. A [*star-like*]{} drawing of $S_{n,d}$ for $n \equiv d \equiv 0 \pmod 2$ is introduced and we conjecture that this drawing maximizes the maximum rectilinear crossing numbers, too. We offer a simpler proof of two results initially proved by Furry and Kleitman [@cycle1] as partial results in the direction of this conjecture.' author: - 'Matthew Alpert[^1], Elie Feder[^2], and Heiko Harborth[^3]' title: 'The Maximum of the Maximum Rectilinear Crossing Numbers of $d$-regular Graphs of Order $n$' --- Introduction ============ Let $G$ be an abstract graph with vertex set $V(G)$ and edge set $E(G)\subset V(G) \times V(G)$. The *order* of a graph $G$ is defined as the cardinality of $V(G)$. A *drawing* of the graph $G$ is a representation of $G$ in the plane such that the elements of $V(G)$ correspond to points in the plane, and the elements of $E(G)$ correspond to continuous arcs connecting two vertices and having at most one point in common, either a vertexpoint or a crossing. A *rectilinear drawing* is a drawing of a graph in which all edges are represented as straight line segments in the plane. The *degree* of a vertex $v \in V(G)$ is defined as the number of edges in $E(G)$ containing $v$ as an endpoint. If all vertices of a graph have the same degree, then the graph is called *regular*. Specifically, if all the vertices have degree $d$, the graph is called $d$*-regular*. The *cycle* $C_n$ is a connected $2$-regular graph. The *complete graph* $K_n$ is a graph on $n$ vertices, in which any two vertices are connected by an edge, or equivalently an $(n-1)$-regular graph. The class of $d$-regular graphs of order $n$ will be denoted $R_{n,d}$. In a drawing of a graph, a *crossing* is defined to be the intersection of exactly two edges not at a vertex. The *crossing number* of an abstract graph, $G$, denoted ${\rm cr}(G)$, is defined as the minimum number of edge crossings over all nonisomorphic drawings of $G$. The *minimum rectilinear crossing number* of a graph $G$, denoted $\overline{\rm cr}(G)$, is defined as the minimum number of edge crossings over all nonisomorphic rectilinear drawings of $G$. Analogously, the *maximum crossing number*, denoted by ${\rm CR}(G)$, is defined as the maximum value of edge crossings over all nonisomorphic drawings of $G$. The *maximum rectilinear crossing number* of a graph $G$, denoted by $\overline{\rm CR}(G)$, is defined to be the maximum number of edge crossings over all nonisomorphic rectilinear drawings of $G$. Throughout this paper we will also define $\overline{\rm CR}(R_{n,d})$ to be the maximum of the maximum rectilinear crossing numbers throughout the class of graphs. The maximum crossing number and maximum rectilinear crossing number have been studied for several classes of graphs (see [@maxrect1], [@max1], [@max2], [@max3], [@max4]). Most relevant to this paper are studies of the maximum rectilinear crossing number of $C_n$ (a $2$-regular graph) and of $K_n$ ($(n-1)$-regular graph). In [@complete1] it is shown that $$\overline{\rm CR}(K_n)=\overline{\rm CR}(R_{n,n-1})=\binom{n}{4}.$$ In [@cycle1], [@cycle2] it is proved that $$\overline{\rm CR}(C_n)= \begin{cases} \frac{1}{2}n(n-3) & \text{if $n$ is odd,} \\ \frac{1}{2}n(n-4) + 1 & \text{if $n$ is even.} \end{cases}$$ This paper makes a natural generalization from these two results. Namely, it finds an expression for the maximum $\overline{\rm CR}(R_{n,d})$ of all maximum rectilinear crossing numbers for the class $R_{n,d}$ of all $d$-regular graphs of order $n$, where $2 \leq d \leq n-1$. We present a [*star-like*]{} drawing of a $d$-regular graph $S_{n,d}$ for $n$ and $d$ of different parity and prove that it maximizes the maximum rectilinear crossing numbers. A [*star-like*]{} drawing of the $d$-regular graph $S_{n,d}$ for even $n$ and $d$ is introduced and we conjecture that this drawing maximizes the maximum rectilinear crossing numbers offering proofs for $d=2$ and $d=n-2$ as partial results in the direction of this conjecture. We present here an interesting method of generalizing the maximum rectilinear crossing number of $C_n$ and $K_n$ to the more general class $R_{n,d}$ of $d$-regular graphs of order $n$. Finding the minimum rectilinear crossing number of the complete graph, $K_n$, is a well-known and widely-investigated open problem in computational geometry. For $n<17$, $\overline{\rm cr}(K_n)$ is known, and for $n \geq 18$ only bounds are known (see [@min1], [@min2], [@min3]). Perhaps future research can investigate $\overline{\rm cr}(R_{n,d})$ where $d<n-1$ as a tool to gain insight into the minimum rectilinear crossing number of $K_n$. In Sections 2.1 and 2.2 we outline the construction of the [*generalized star-like*]{} drawings of $S_{n,d}$ and present a lower bound for $\overline{\rm CR}(R_{n,d})$. In Section 3.1 we present an upper bound for $\overline{\rm CR}(R_{n,d})$, where $n+d\equiv 1 \pmod 2$, and note that the [*star-like*]{} drawing of $S_{n,d}$ attains this maximum. In Section 3.2 we conjecture the upper bound of $\overline{\rm CR}(R_{n,d})$ where $n \equiv d \equiv 0 \pmod 2$ and offer a partial result in the direction of this conjecture by proving its validity for the case $d=2$. In Section 3.3 we offer simpler proofs of the maximum crossing number of $C_n$ and of $\overline{\rm CR}(R_{n,2})$ where $n \equiv 0 \pmod 2$ than those of Furry and Kleitman [@cycle1] and in Section 3.4 we remark on this paper’s generalization of previous results. Section 3.5 contains some computational results regarding $\overline{\rm CR}(R_{n,d})$. Lower Bounds of $\overline{\rm CR}(R_{n,d})$ ============================================ We first note that there is no $d$-regular graph of order $n$ where $n$ and $d$ are both odd since the number $nd$ of endvertices cannot twice count the number of edges. Thus, we will only consider the two cases $n+d \equiv 1 \pmod 2$ and $n \equiv d \equiv 0 \pmod 2$. Lower bound of $\overline{CR}(R_{n,d})$ where $n+d \equiv 1 \pmod 2$ -------------------------------------------------------------------- The number of crossings in a special rectilinear [*star-like*]{} drawing implies the following lower bound. $$\overline{CR}(R_{n,d}) \geq \frac{nd}{24}(3nd-2d^2-6d+2)$$ if $n+d \equiv 1 \pmod 2$. Consider a rectilinear drawing of $K_n$ where the vertices are arranged as those of a convex $n$-gon. Step by step we delete all diagonals of lengths $1,2, \ldots, k-1$. We proceed by counting the number of crossings we remove from the drawing by now deleting the $n$ diagonals of length $k$. There are $k-1$ vertices in one of the halfplanes each of the diagonals of length $k$ divides the drawing into. Each of these vertices will have $n-1-(2(k-1))=n-2k+1$ edges emanating from it which intersect the original diagonal of length $k$. However, each diagonal of length $k$ intersects $2(k-1)$ other diagonals of length $k$. Since these crossings are counted twice, we find that there are $n(k-1)$ crossings between diagonals of length $k$. These are also counted twice in the sum $n(k-1)(n-2k+1)$ and thus we only remove $n(k-1)(n-2k+1)-n(k-1)=n(k-1)(n-2k)$ crossings in deleting all diagonals of length $k$ provided all shorter diagonals have been previously deleted. Therefore we obtain $$\overline{CR}(R_{n,d}) \geq \binom{n}{4} - \displaystyle\sum_{i=1}^{k-1} n(i-1)(n-2i)$$ $$=\binom{n}{4} - \frac{1}{6}n(k-1)(k-2)(3n-4k).$$ After these deletions there are $d=n-1-2(k-1)$ edges emanating from each vertex. Substituting $k=\frac{1}{2}(n-d+1)$ into the closed form of the sum above we obtain the desired result. We call this drawing of $K_n$ without the diagonals of lengths $1$ through $k-1=\frac{1}{2}(n-d-1)$ the *generalized star* drawing of $S_{n,d}$ in $R_{n,d}$ (see Figure 1).  Lower bound of $\overline{CR}(R_{n,d})$ where $n \equiv d \equiv 0 \pmod 2$ --------------------------------------------------------------------------- The number of crossings in the special rectilinear [*star-like*]{} drawing implies the following lower bound for $\overline{CR}(R_{n,d})$ where $n \equiv d \equiv 0 \pmod 2$. $$\overline{\rm CR}(R_{n,d}) \geq \begin{cases} \frac{1}{24}nd(3nd-2d^2-6d-1) & \text{if $n \equiv d \equiv n/(n,k) \equiv 0 \pmod 2, $ }\\ \\ \frac{1}{24}nd(3nd-2d^2-6d-1)\\ - \frac{1}{4}(n,k)(2d-3) & \text{if $n \equiv d \equiv 0 \pmod 2$} \\ & \text{and $n/(n,k)\equiv 1 \pmod 2 $} \end{cases}$$ For $n \equiv d \equiv 0 \pmod 2$ we use the generalized star drawing of $S_{n,d+1}$ for $d+1 = n-2k+1$ and delete one edge at each vertex to obtain a star-like drawing of $S_{n,d}$ with $d=n-2k$ (an even number). The diagonals of length $k$ in $S_{n,d+1}$ determine $(n,k)$ cycles each of order $n/(n,k)$. If $n/(n,k) \equiv 0 \pmod 2$ we can delete every second edge of every cycle (see Figure 2). In removing these edges we remove $$\frac{1}{2}n(k-1)(n-2k+1) - \frac{1}{4}n(k-1)=\frac{1}{2}n(k-1)(n-2k+\frac{1}{2})$$ edge crossings from the drawing. Subtracting this from the bound in Proposition 2.1 it follows that $$\overline{CR}(R_{n,d}) \geq \binom{n}{4} - \frac{1}{6}n(k-1)(k-2)(3n-4k)- \frac{1}{2}n(k-1)(n-2k+\frac{1}{2}).$$ Substituting $k=\frac{1}{2}(n-d)$ gives the desired result.   If $n/(n,k) \equiv 1 \pmod 2$ then the diagonals of length $k$ determine $(n,k) \equiv 0 \pmod 2$ cycles of odd order. We partition these cycles into $\frac{1}{2}(n,k)$ pairs. For each pair we delete a diagonal of length $k+1$ connecting two vertices of these cycles. For each of these diagonals of length $k+1$ we keep the diagonals of length $k$ which emanate from their endpoints and then delete their neighbor edges and every second of the remaining edges within the cycles of order $n/(n,k)$ (see Figure 3). Thus we remove $\frac{1}{2}(n,k)$ edges of length $k+1$ and $$\frac{1}{2}({n/(n,k)-1})(n,k)=\frac{1}{2}(n-(n,k))$$ diagonals of length $k$. In removing these edges we remove $$\frac{1}{2}(n-(n,k))(k-1)(n-2k+1)+\frac{1}{2}(n,k)(k)(n-2k+1)$$ $$- \frac{1}{2}(n,k)2 - \frac{1}{2}[\frac{1}{2}(n-(n,k))(k-1)+\frac{1}{2}(n,k)k]$$ $$=\frac{1}{2}(n-2k+\frac{1}{2})(kn-n+(n,k))-(n,k)$$ crossings from the drawing of $S_{n,d+1}$. It follows that $$\overline{CR}(R_{n,d}) \geq \binom{n}{4} - \frac{1}{6}n(k-1)(k-2)(3n-4k)$$ $$-\frac{1}{2}(n-2k+\frac{1}{2})(kn-n+(n,k))-(n,k).$$ Substituting $k=\frac{1}{2}(n-d)$ yields the desired inequality.   Upper Bounds of $\overline{\rm CR}(R_{n,d})$ ============================================ In this section we prove that the lower bound obtained in Proposition 2.1 is also an upper bound for $\overline{\rm CR}(R_{n,d})$ where $n+d \equiv 1 \pmod 2$. In addition we conjecture that the lower bound obtained in Proposition 2.2 is an upper bound and offer a partial result in the direction of this conjecture. Upper bound of $\overline{CR}(R_{n,d})$ where $n+d \equiv 1 \pmod 2$ -------------------------------------------------------------------- The following exact value of $\overline{\rm CR}(R_{n,d})$ will be proved. $$\overline{\rm CR}(R_{n,d}) = \frac{1}{24}nd(3nd-2d^2-6d+2) \ \rm{if}\ n+d \equiv 1 \pmod 2.$$ The lower bound follows from Proposition 2.1, so we proceed by proving that this expression is an upper bound. Every $d$-regular graph of order $n$ has $\frac{1}{2}nd$ edges. Every edge can intersect at most $\frac{1}{2}nd-(2d-1)$ other edges. Thus, a first upper bound is $$\overline{\rm CR}(R_{n,d}) \leq \frac{1}{2} (\frac{1}{2}nd) (\frac{1}{2}nd-2d+1)=\frac{1}{24}nd(3nd-12d+6).$$ Every vertex in a $d$-regular graph is an endvertex for $d$ edges. Let an endvertex be of type $i$ if the edge incident to it divides the drawing of the graph into two halfplanes, one containing $i$ edges emanating from one vertex, and the other containing $d-i-1$ edges emanating from the same vertex (see Figure $4$). By symmetry we only consider $0 \leq i \leq \lfloor \frac{1}{2}(d-1) \rfloor = D. $  Let $y_i$ be the number of endvertices of type $i$. Thus, we have $y_0 + y_1+ \ldots +y_D = dn$. We call an edge with $i$ edges in a halfplane at one endvertex and $j$ edges in the same halfplane at the other endvertex a type $i,j$ edge. Let $x_{i,j}$ count the number of type $i,j$ edges (see Figure $5$).  Thus, $y_i$ is related to $x_{i,j}$ by the following equation: $$y_i=2x_{i,i}+\displaystyle\sum_{k=0}^{i-1} x_{k,i} + \displaystyle\sum_{k=i+1}^{D} x_{i,k}.$$ Now, for a type $i,j$ edge, the $i$ edges in the halfplane of one endvertex cannot intersect the $d-j-1$ edges in the opposite halfplane emanating from the other endvertex. The same holds true for the $j$ edges in the halfplane of one endvertex and the $d-i-1$ edges in the opposite halfplane emanating from the other endvertex. Therefore, a given type $i,j$ edge determines $i(d-j-1)+j(d-i-1)$ pairs of nonintersecting edges. A drawing which maximizes the number of edge crossings should minimize the number $M$ of pairs of nonintersecting edges. Note that it is true that for a given type $i,j$ edge it may be that the $i$ edges from one endvertex and the $j$ edges from the other endvertex will be in different halfplanes. This will yield $ij + (d-j-1)(d-i-1)$ nonintersecting edges. However, $i(d-j-1)+j(d-i-1) \leq ij + (d-j-1)(d-i-1)$ when $0 \leq i \leq j \leq D $. Therefore, the minimum number $M$ of pairs of nonintersecting edges over a drawing of the graph occurs when the $i$ and $j$ edges are arranged so that they lie in the same halfplane. Thus, we assume that $0 \leq i \leq j \leq D $ and a given type $i,j$ edge always determines $i(d-j-1)+j(d-i-1)$ pairs of nonintersecting edges. Summing this quantity over all edges of a drawing we obtain $$M=\displaystyle\sum_{i=0}^{D}\displaystyle\sum_{j=i}^{D} [i(d-j-1)+j(d-i-1)]x_{i,j}$$ pairs of nonintersecting edges.\ \ In order to minimize $M$, we begin by multiplying equation ($1$) by $i(d-i-1)$ and subtracting it from $M$ for all values of $i$, yielding $$M= \displaystyle\sum_{i=1}^{D} i(d-i-1)y_i + \displaystyle\sum_{i=0}^{D-1} \displaystyle\sum_{j=i+1}^{D} (j-i)^2 x_{i,j}.$$ Let $p_{s,t}$ count the number of vertices having endvertices of type $s$ as the smallest type ($ 0 \leq s \leq D$). The index $t$ counts the number of distinct sequences of endvertex types for a given vertex counted in $p_{s,t}$ ($t \geq 1$). For example, in a convex drawing of $S_{n,d}$, $p_{0,1}=n$, $p_{0,t}=0$ for $t \geq 2 $, and $p_{s,t}=0$ for $s \geq 1 $. Then, $$n= \displaystyle\sum_{s=0}^{D} \displaystyle\sum_{t \geq 1} p_{s,t}.$$ Note that if the smallest type $s$ of an endvertex is $0$ then the point must be on the convex hull and all such points will have one distinct sequence of endvertex types. Thus, $p_{0,t}=0$ for $t \geq 2 $.\ Let $z_{s,t,i}$ denote the number of endvertices of type $i$ for the $p_{s,t}$ vertices. It follows that $$y_i=2p_{0,1} + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{i} z_{s,t,i}p_{s,t}$$\ and for odd $d$ we have $$y_D=p_{0,1} + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} z_{s,t,D}p_{s,t}.$$ Additionally, since every vertex has $d$ edges, for a fixed $s$ and $t$ it holds that $$\displaystyle\sum_{i=s}^{D} z_{s,t,i}=d.$$\ Using equations ($3$) and ($4$) we obtain $$y_i= 2n + \displaystyle\sum_{t \geq 1}[ \displaystyle\sum_{s=1}^{i} (z_{s,t,i}-2)p_{s,t} - 2\displaystyle\sum_{s=i+1}^{D}p_{s,t}]$$\ and respectively, for $d$ odd we have $$y_D=n+ \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} (z_{s,t,i}-1)p_{s,t}.$$\ We proceed for $d$ even. Using equation ($6$) we can rewrite the first part of the expression for $M$ in equation ($2$) as $$\displaystyle\sum_{i=1}^{D} i(d-i-1)y_i=2n \displaystyle\sum_{i=1}^{D} i(d-i-1) + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{i=1}^{D} i(d-i-1)[\displaystyle\sum_{s=1}^{i} (z_{s,t,i}-2)p_{s,t} - 2\displaystyle\sum_{s=i+1}^{D}p_{s,t}].$$ Following a change in the indices of the sums, the right term can be rewritten as $$2n \displaystyle\sum_{i=1}^{D} i(d-i-1) + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} p_{s,t}[\displaystyle\sum_{i=s}^{D} i(d-i-1)(z_{s,t,i}-2)-2\displaystyle\sum_{i=1}^{s-1} i(d-i-1)].$$ This can again be rewritten as\ $$2n \displaystyle\sum_{i=1}^{D} i(d-i-1)+\displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} p_{s,t}[s(d-s-1)\displaystyle\sum_{i=s}^{D} (z_{s,t,i}-2)+$$\ $$\displaystyle\sum_{i=s+1}^{D}(i(d-i-1)-s(d-s-1))(z_{s,t,i}-2)-2\displaystyle\sum_{i=1}^{s-1} i(d-i-1)].$$ Using equation ($5$), it follows that this term is also equal to $$2n \displaystyle\sum_{i=1}^{D} i(d-i-1) + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} p_{s,t}[C(s,d)+ \displaystyle\sum_{i=s+1}^{D}(i(d-i-1)-s(d-s-1))(z_{s,t,i}-2)]$$ where $$\begin{aligned} C(s,d)&=& s(d-s-1)(d - \displaystyle\sum_{i=s}^{D} 2) - 2\displaystyle\sum_{i=1}^{s-1} i(d-i-1) \\ &=& s(d-s-1)(d-2(D-s+1))- 2\displaystyle\sum_{i=1}^{s-1} i(d-i-1).\end{aligned}$$ We now show that $C(s,d)$ is nonnegative for all $s$ and $d$. First, we have $$s(d-s-1)(d-2(D-s+1) \geq s(d-s-1)(2s-1).$$ Then $$\begin{aligned} 2\displaystyle\sum_{i=1}^{s-1} i(d-i-1) &\leq& 2\displaystyle\sum_{i=1}^{s-1}(s-1)(d-s) \\ &=& 2(s-1)^2(d-s) \\ &<& s(d-s-1)(2s-2). \end{aligned}$$ Therefore $$C(s,d)>s(d-s-1)(2s-1)-s(d-s-1)(2s-2)=s(d-s-1) \geq 0.$$ Additionally, $(i(d-i-1)-s(d-s-1)) \geq 0 $ for $i,s \leq D=\frac{1}{2}(d-1)$ and $i \geq s+1$. Assuming $ z_{s,t,i} -2 \geq 0 $ (which we will prove in the following lemma) then the first half of the expression for $M$ is minimized when $p_{s,t}=0$ for all $s \geq 1$.\ Also, accounting for the discrepancy in $y_D$ when $d$ is odd, an analogous summation can be carried out. Since the term $ z_{s,t,D} -1$ must be carried throughout this summation the expression for $d$ odd is also minimized for $p_{s,t}=0$ for all $s \geq 1$, provided $ z_{s,t,D} -1 \geq 0 $. $z_{s,t,i} \geq 2$ for all $s,t,i$, and $z_{s,t,D} \geq 1$ for $d$ odd. For a given vertex, we begin by proving there is at least one endvertex of type $\frac{1}{2}(d-1)$ for $d$ odd and there are at least two endvertices of type $\frac{1}{2}(d-2)$ for $d$ even. This statement can be proved by induction from $d$ to $d+1$. This statement is obvious for $d=2$ and $d=3$, so we begin with the inductive step. Also, note that in traversing the $d$ edges incident to a given vertex in a clockwise or counterclockwise manner in moving from edge to edge, edge to extension, extension to edge, and extension to extension, the number of edges in the clockwise following halfplane may change by at most one. This fact will be used numerous times throughout the proof.\ **Case I:** From odd $d$ to $d+1$.\ We consider the edge whose endvertex is of type $\frac{1}{2}(d-1)$ in the $d$-regular drawing. When the ($d+1$)st edge is added, this original endvertex will be the first endvertex of type $\frac{1}{2}[(d+1)-2]$. If the ($d+1$)st edge is added in this edge’s clockwise following halfplane then an immediately following edge or edge extension’s endvertex will have type $\frac{1}{2}[(d+1)-2]$. Thus, either this edge or the edge corresponding to this extension’s endvertex will be the second endvertex of type $\frac{1}{2}(d-1)$.\ **Case II:** From even $d$ to $d+1$.\ Consider an edge whose endvertex is of type $\frac{1}{2}(d-2)$ which has $\frac{1}{2}d$ edges in one of its halfplanes and $\frac{1}{2}(d-2)$ in the other. If the ($d+1$)st edge is added in the halfplane with $\frac{1}{2}(d-2)$ edges then the considered endvertex is of type $\frac{1}{2}[(d+1)-1]$. If the ($d+1$)st edge is added in the halfplane with $\frac{1}{2}d$ edges then there are $\frac{1}{2}[(d+1)+1]$ edges in this halfplane and $\frac{1}{2}[(d+1)-3]$ edges in the clockwise following halfplane of this edge’s extension. Since the number of edges in the clockwise following halfplane can change by at most one when moving from edge line to edge line (edge ray and edge extension), we find that traversing the graph from the edge with $\frac{1}{2}[(d+1)+1]$ edges in the clockwise following halfplane to the extension with $\frac{1}{2}[(d+1)-3]$ there must occur an edge or extension with $\frac{1}{2}[(d+1)-1]$ edges in the clockwise following halfplane. Thus, this edge or the edge corresponding to the extension’s endvertex is of type $\frac{1}{2}[(d+1)-1]$.\ \ Using this result and the fact that in moving from edge line to adjacent edge line, the number of edges in the clockwise following halfplane may change by at most one, we can prove that there are two endvertices of each type from the minimal type $s$ to the maximal type $D$. For $d$ odd, we have one endvertex of maximal type $D=\frac{1}{2}(d-1)$. Traversing the $d$ edges starting and ending with the edge of type $D$ from edge line to edge line we must go down to an edge or an extension with $s$ edges in the clockwise following halfplane, and then back up to one with $D$. Thus, we find there are at least two of edges or extensions whose endvertices are of each type from $s$ to $D$. For $d$ even, we have two edges of maximal type $D=\frac{1}{2}(d-2)$. Traversing the $d$ edges from one of the type $D$ edges to the other must go down to an edge or extension with $s$ edges in the clockwise following halfplane and back up to one with $D$. Thus, there are at least two edges or extensions whose endvertices are of each type from $s$ to $D$. It follows that $z_{s,t,i} \geq 2$. Going back to the final expression for equation ($2$) we have $$M=2n \displaystyle\sum_{i=1}^{D} i(d-i-1) + \displaystyle\sum_{t \geq 1} \displaystyle\sum_{s=1}^{D} p_{s,t} [C(s,d)+ \displaystyle\sum_{i=s+1}^{D}(i(d-i-1)-s(d-s-1))(z_{s,t,i}-2)]$$ $$+ \displaystyle\sum_{i=0}^{D-1} \displaystyle\sum_{j=i+1}^{D} (j-i)^2 x_{i,j}.$$ Since $C(s,d)$, $z_{s,t,i} -2 $, and $(j-i)^2$ are greater than or equal to $0$ we find that this expression is minimized when $p_{s,t}=0$ for $s \geq 1$ and $x_{i,j}=0$ for $i<j$. Evaluating the initial sum using these conditions we find that for even $d$ we have $$M=2n \displaystyle\sum_{i=1}^{D} i(d-i-1)=\frac{1}{6}nd(d-1)(d-2)$$ pairs of nonintersecting edges, and for odd $d$ we have $$M=2n \displaystyle\sum_{i=1}^{D-1} i(d-i-1) +nD(d-D-1)= \frac{1}{6}nd(d-1)(d-2)$$ pairs of nonintersecting edges. Since every pair of nonintersecting edges can count twice for two intersecting edges (see Figure 6) we can subtract at least $\frac{1}{12}nd(d-1)(d-2)$ from the initial upper bound to obtain the asserted bound.  Conjecture on the upper bound of $\overline{\rm CR}(R_{n,d})$ ------------------------------------------------------------- For $n \equiv d \equiv 0 \pmod 2$ we have the following conjecture. The bound in Proposition 2.2 is sharp. We offer a an alternate proof of $ \overline{CR}(R_{n,2})$, originally proven by Furry and Kleitman [@cycle1], as a partial result in the direction of this conjecture. $$\overline{CR}(R_{n,2}) = \lfloor \frac{1}{4}n(2n-7) \rfloor \ \rm{where}\ n\equiv 0 \pmod 2.$$ The lower bound follows from Proposition 2.2. Therefore, we proceed by proving the upper bound. For each edge there is a maximum of $n-3$ nonadjacent edges which it can intersect. Since $n \equiv 0 \pmod 2$, those edges which have $n-3$ crossings must have neighbor edges in different halfplanes. The two neighbor edges cannot have $n-3$ crossings since these edges cannot intersect each other. Thus there are at most $\frac{1}{2}n$ disjoint edges which may have $n-3$ crossings. It follows that $$\overline{CR}(R_{n,2}) \leq \frac{1}{2} [ \frac{1}{2}n (n-3) + \frac{1}{2}n(n-4)]= \lfloor \frac{1}{4}n(2n-7) \rfloor.$$ Note that only for $d=2$ and $n$ even there occur disconnected graphs $S_{n,2}$ in the extremal cases, that is, there are copies of $C_4$ if $n \equiv 0\pmod 4$ and there are copies of $C_4$ and one copy of $C_6$ if $n \equiv 2\pmod 4$. Alternate proof of $\overline{CR}(C_n)$ --------------------------------------- In the same vein as the above proof for $ \overline{CR}(R_{n,2})$ we now offer a simpler proof of $\overline{CR}(C_n)$ than that of Furry and Kleitman. Note that for both $R_{n,2}$ and $C_n$ where $n\equiv 1 \pmod 2$ the proof of the maximum rectilinear crossing number is trivial as both achieve the thrackle bound. $$\overline{CR}(C_n) = \frac{1}{2}(n^2-4n+2) \rm{where} n \equiv 0 \pmod 2.$$ The lower bound follows from [@cycle1]. Therefore, we proceed by proving the upper bound. In an even cycle an edge with $n-3$ crossings must have its neighbor edges in different halfplanes. Assume that we have three such edges. We label these three pairwise intersecting edges $A_1A_2$, $B_1B_2$, and $C_1C_2$ as shown in Figure $7$.  We start at $A_1$. The other edge incident to $A_1$ must intersect both $B_1B_2$ and $C_1C_2$. Thus, without a loss of generality, we may assume that the termination of this edge, $P_1$, must lie in region from $A_2$ to $B_2$. The next edge must terminate at $P_2$ which must lie in the region from $A_1$ to $B_1$, and $P_3$ must lie again in the region from $A_2$ to $B_2$, and so on, since all three original edges must be intersected by all edges except their neighbor edges. Eventually, the cycle must close up on itself and $P_{2i-1}$ or $P_{2i}$ must terminate at $B_1$ or $B_2$, respectively, since edges $A_1A_2$ and $C_1C_2$ must be intersected. Note that the cycle cannot close on $A_2$ or any $P_k$ where $k < i$, because this will result in a disconnected two-regular drawing. It follows that edge $B_2P_{2i+1}$ or $B_1P_{2i+2}$ must have $P_{2i+1}$ or $P_{2i+2}$ in the region from $A_1$ to $B_1$ or $A_2$ to $B_2$, respectively. Thus, $C_1$ or $C_2$ can never be reached without forcing one of the edges $A_1A_2$ and $B_1B_2$ to have less than $n-3$ crossings, a contradiction. It follows that at most two edges can have $n-3$ crossings and thus we obtain $$\overline{CR}(C_n) \leq \frac{1}{2} [2(n-3)+(n-2)(n-4)]=\frac{1}{2}(n^2-4n+2).$$ Another partial result in the direction of Conjecture 3.3 is the following proposition. $$\overline{CR}(R_{n,n-2})=\binom{n}{4}$$ for $n$ even. The lower bound follows from Proposition 2.2 where every second side is deleted from a rectilinear drawing of $K_n$ as a convex $n$-gon. This bound is sharp since every $4$-tuple of vertices can determine at most one crossing. A generalization of previous results ------------------------------------ Theorem 3.1 extends known results regarding the maximum rectilinear crossing number of the cycle and complete graph to the more general class $R_{n,d}$ of $d$-regular graphs where $2\leq d \leq n-1$. We remark here that when we substitute $d=2$ into Theorem 3.1 we have $$\overline{\rm CR}(R_{n,2})=\overline{\rm CR}(C_n)=\frac{1}{4}(2n)(3(2)n - 2(2)^2-6(2)+2)=\frac{1}{2}n(n-3).$$ This is the same result obtained in [@cycle1] for $\overline{\rm CR}(C_n)$ where $n \equiv 1 \pmod 2$.\ Additionally, we can substitute $d=n-1$ into Theorem 3.1 yielding $$\begin{aligned} \overline{\rm CR}(R_{n,n-1})&=&\overline{\rm CR}(K_n)=\\ &=&\frac{1}{24}n(n-1)[3n(n-1)-2(n-1)^2-6(n-1)+2] =\\ &=&\frac{1}{24}n(n-1)(n-2)(n-3)=\binom{n}{4}.\end{aligned}$$\ This is the same result obtained in [@complete1] regarding $\overline{\rm CR}(K_n)$. Computational results --------------------- The following table shows the values of $\overline{\rm CR}(R_{n,d})$ for various $n$ and $d$. Note that the values in bold are the conjectured results. $$ $d \backslash n$ 4 5 6 7 8 9 10 ------------------ --- --- ---- ---- -------- ----- --------- 2 - 5 7 14 18 27 32 3 1 - 15 - 38 - 70 4 - 5 15 35 **52** 81 **105** 5 - - 15 - 70 - 150 6 - - - 35 70 126 **133** 7 - - - - 70 - 210 8 - - - - - 126 210 9 - - - - - - 210 $$ A general conjecture -------------------- For the determination of the maximum rectilinear crossing number of any graph $G$ it would be very helpful if the following conjecture can be proved. The maximum rectilinear crossing number of any graph can be realized in a drawing where all the vertices are vertexpoints of a convex polygon. Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to thank David Garber for fruitful discussions. [1]{} Aichholzer, O., Aurenhammer, F. and Krasser, H., *On the crossing number of complete graphs*, In: Proc. 18th Ann. ACM Symp. Computational Geometry, 19-24, Barcelona, Spain, 2002. Aichholzer, O., and Krasser., H., *Abstract order type extension and new results on the rectilinear crossing number*, Computational Geometry: Theory and Applications, Special Issue on the 21st European Workshop on Computational Geometry, **36**(1), 2-15, 2006. Aichholzer, O., Orden, D. and Ramos, P.A., *On the structure of sets attaining the rectilinear crossing number*, In: Proc. 22nd European Workshop on Computational Geometry EuroCG ’06, pp. 43-46, Delphi, Greece, 2006. Bienstock, D., Dean, N., [*Bounds for rectilinear crossing numbers*]{}, J. Graph Theory [**17**]{} (1993), 333-348. Eggleton, R.B., [*Rectilinear drawings of graphs*]{}, Utilitas Math. [**29**]{} (1986), 149-172. Furry, W.H., Kleitman,D.J., [*Maximal Rectilinear Crossings of Cycles*]{}, Studies in Appl. Math. [**56**]{} (1977), 159-167. Gan, C.S., Koo, V.C., [*Enumerations of the maximum rectilinear crossing number of complete and complete multi-partite graphs*]{}, J. of Discrete Mathematical Sciences and Cryptography [**9**]{} (2006), 583-590. Green, J.E., Ringeisen, R.D., *Lower bound for the maximum crossing number using certain subgraphs*, Congr. Numer. **90** (1992), 193-203. Harborth, H., [*Drawing of the cycle graph*]{}, Congr. Numer. [**66**]{} (1988), 15-22. Harborth, H., [*Maximum number of crossings for the cube graph*]{}, Congr. Numer. [**82**]{} (1991), 117-122. Harborth, H., Thuermann, C., [*Number of edges without crossings in rectilinear drawings of the complete graph*]{}, Congr. Numer. [**119**]{} (1996), 76-83. Piazza, B., Ringeisen, R.D., Stueckle, S., *Subthrackle graphs and maximum crossings*, Discrete Math. **127** (1994), 265-276. Ringeisen, R.D., Stueckle, S., Piazza, B.L., *Subgraphs and bounds on maximum crossings*, Bull. Inst. Combin. Appl. 2 (1991), 33-46. Ringel, G., *Extremal problems in the theory of graphs*, in: Fiedler, M. (ed.), Theory of Graphs and its Applications, Proc. Symposium Smolenice 1963, Prague, 1964, 85-90. Thomassen, C., [*Rectilinear drawings of graphs*]{}, J. Graph Theory [**12**]{} (1988), 335-341. [^1]: Lawrence High School, Cedarhurst, NY, USA. *E-mail address:* `mna851@aol.com` [^2]: Department of Mathematics and Computer Science, Kingsborough Community College-CUNY, Brooklyn, NY, USA. *E-mail address:* `efeder@kbcc.cuny.edu` [^3]: Diskrete Mathematik, Technische Universitaet, Braunschweig, Germany. *E-mail address:* `H.Harborth@tu-bs.de`

--- abstract: 'We develop a global theory for complete hypersurfaces in $\R^{n+1}$ whose mean curvature is given as a prescribed function of its Gauss map. This theory extends the usual one of constant mean curvature hypersurfaces in $\R^{n+1}$, and also that of self-translating solitons of the mean curvature flow. For the particular case $n=2$, we will obtain results regarding a priori height and curvature estimates, non-existence of complete stable surfaces, and classification of properly embedded surfaces with at most one end. [^1]' --- ------------------------------------------------------------------------ [**The global geometry of surfaces with\ prescribed mean curvature in $\R^3$**]{}\ \ ------------------------------------------------------------------------ $\mbox{}^a,^b$ Departamento de Geometría y Topología, Universidad de Granada, E-18071 Granada, Spain.\ e-mail: jabueno@ugr.es, jagalvez@ugr.es $\mbox{}^c$ Departamento de Matemática Aplicada y Estadística, Universidad Politécnica de Cartagena, E-30203 Cartagena, Murcia, Spain.\ e-mail: pablo.mira@upct.es Introduction ============ Let $\cH$ be a $C^1$ function on the sphere $\S^n$. We will say that an immersed oriented hypersurface $\Sigma$ in $\R^{n+1}$ has *prescribed mean curvature $\cH$* if its mean curvature function $H_{\Sigma}$ is given by $$\label{presH} H_{\Sigma}=\cH\circ \eta,$$ where $\eta:\Sigma\flecha \S^n$ is the Gauss map of $\Sigma$. For short, we will simply say that $\Sigma$ is an $\cH$-hypersurface. Obviously, when $\cH$ is constant, $\Sigma$ is a hypersurface of constant mean curvature (CMC). The study of hypersurfaces in $\R^{n+1}$ defined by a prescribed curvature function in terms of the Gauss map goes back, at least, to the famous Christoffel and Minkowski problems for ovaloids (see e.g. [@C]). The existence and uniqueness of ovaloids of prescribed mean curvature in $\R^{n+1}$ was studied among others by Alexandrov and Pogorelov in the 1950s (see [@Al; @Po]), and more recently in [@GG]. However, the geometry of complete hypersurfaces of prescribed mean curvature $\cH$ in $\R^{n+1}$ remains largely unexplored. The most studied situation is, obviously, the case of complete CMC hypersurfaces in $\R^{n+1}$. Recently, the global geometry of self-translating solitons of the mean curvature flow (which correspond to $\cH$-hypersurfaces for the particular choice $\cH(x)=\esiz x,e_{n+1}\esde$) is being studied in detail, see e.g. [@DDN; @Il; @IR; @MPGSHS; @MSHS; @SX]. Our objective in this paper is to develop a global theory of complete $\cH$-hypersurfaces in $\R^{n+1}$, taking as a starting point the well-studied global theory of CMC hypersurfaces in $\R^{n+1}$. We will be specially interested in the case of complete $\cH$-surfaces in $\R^3$. The following are three trivial but fundamental properties of $\cH$-hypersurfaces in $\R^{n+1}$, for a given $\cH\in C^1(\S^n)$: (1) any translation of an $\cH$-hypersurface is also an $\cH$-hypersurface; (2) $\cH$-hypersurfaces are locally modeled by a quasilinear elliptic PDE when viewed as local graphs in $\R^{n+1}$ over each tangent hyperplane, and consequently they obey the maximum principle; and (3) any symmetry of $\cH$ in $\S^n$ induces a linear isometry of $\R^{n+1}$ that preserves $\cH$-hypersurfaces in $\R^{n+1}$. These three properties make many ideas of classical CMC hypersurface theory in $\R^{n+1}$ readily available for the case of $\cH$-hypersurfaces, for a non-constant function $\cH\in C^1(\S^n)$. Nonetheless, the class of $\cH$-hypersurfaces is indeed much wider and richer, and new ideas are needed for its study. For example, $\cH$-hypersurfaces in $\R^{n+1}$ do not come in general associated to a variational problem, and this makes many of the most useful techniques of CMC theory unavailable. Also, in contrast with CMC surfaces in $\R^3$, there is no holomorphic object associated to $\cH$-surfaces in $\R^3$ for non-constant $\cH\in C^1(\S^2)$, and so, no clear integrable systems approach to their study. Moreover, in jumping from the CMC condition to the equation of prescribed non-constant mean curvature, one also needs to account for the loss of symmetries and isotropy of the resulting equation. For instance, one can only apply the Alexandrov reflection principle for directions with respect to which $\cH$ is symmetric. Some of these difficulties can already be seen in the case of rotational $\cH$-surfaces in $\R^3$. In the paper [@BGM] we present many examples of rotational $\cH$-surfaces in $\R^3$ with no CMC counterpart: for instance, complete, convex $\cH$-graphs converging to a cylinder, or properly embedded disks asymptotically wiggling around a cylinder. For the case of rotational $\cH$-surfaces of non-trivial topology, we show in [@BGM] examples with a *wing-like* shape, or with two strictly convex ends pointing in opposite directions. Many of these examples can also be constructed so that they self-intersect. All this variety, just for the very particular class of rotational $\cH$-surfaces in $\R^3$, shows that the class of $\cH$-hypersurfaces in $\R^{n+1}$ is indeed very large, and rich in what refers to possible examples and geometric behaviors. It also accounts for the difficulty of obtaining classification results for the case where $\cH$ is not constant, and justifies the need of imposing some additional regularity, positivity or symmetry properties to $\cH$ in order to derive such classification theorems. Despite these difficulties, the results in the present paper show that, under mild assumptions, the class of $\cH$-surfaces in $\R^3$ can sometimes be given a homogeneous treatment for general classes of prescribed mean curvature functions $\cH\in C^1(\S^2)$. In the rest of this introduction we will explain the organization of the paper, and state some of our main results. In [**Section \[sec:2\]**]{} we will study some basic properties of $\cH$-hypersurfaces in $\R^{n+1}$. In Proposition \[hpos\] we prove that if $\cH(x_0)=0$ for some $x_0\in \S^n$, then there are no compact $\cH$-hypersurfaces in $\R^{n+1}$. In Proposition \[ale\] and Corollary \[3planes\] we give some Alexandrov-type results using moving planes in the case that the function $\cH$ has enough symmetries. In Section \[sec:constru\] we will classify flat $\cH$-hypersurfaces. In Section \[sec:compac\] we will prove a compactness theorem for the space of $\cH$-surfaces in $\R^3$ with bounded second fundamental form, with $\cH$ not necessarily fixed. In [**Section \[sec:structure\]**]{} we study the geometry of properly embedded $\cH$-surfaces of finite topology in $\R^3$, motivated by some well-known results by Meeks [@Me] and Korevaar-Kusner-Solomon [@KKS] for the case where $\cH$ is a positive constant. In [@Me], Meeks proved that there are no properly embedded surfaces of positive CMC in $\R^3$ with finite topology and exactly one end. For that, he first obtained universal height estimates for (not necessarily compact) CMC graphs in $\R^3$ with planar boundary. In [@KKS] it was proved that any CMC surface in $\R^3$ in the conditions above, but this time with *two* ends, is a rotational surface. For the general case of $\cH$-surfaces in $\R^3$ with $\cH\in C^1(\S^2)$, $\cH>0$, these statements are not true in general. However, our main results in Section \[sec:structure\] give natural additional conditions on $\cH$ under which Meeks’ results hold for $\cH$-surfaces. Specifically, in Theorem \[th:hees\] we will give some very general sufficient conditions on $\cH$ for the existence of a universal height estimate for $\cH$-graphs with respect to a given direction $v\in \S^2$ in $\R^3$. The way of proving these height estimates is completely different from the one used by Meeks, and relies on a previous curvature estimate for surfaces in $\R^3$ in the spirit of a result on CMC surfaces in Riemannian three-manifolds by Rosenberg, Sa Earp and Toubiana [@RST]; see Theorem \[th:curv\]. As corollaries to Theorem \[th:hees\], in Section \[estru\] we will provide several results about the geometry of properly embedded $\cH$-surfaces in $\R^3$ of finite topology and one end, in the case that $\cH>0$ is invariant under one, two or three reflections in $\S^2$. For instance, in the case that $\cH$ is invariant with respect to three linearly independent reflections in $\S^2$, we will prove (Theorem \[3sim\]) that any properly embedded $\cH$-surface in $\R^3$ of finite topology and at most one end is the Guan-Guan sphere $S_{\cH}$ associated to $\cH$, see [@GG]. In [**Section \[sec:stable\]**]{} we study *stability* of $\cH$-hypersurfaces. As we already mentioned, except for some very particular cases, the equation describing $\cH$-hypersurfaces in $\R^{n+1}$ is *not* the Euler-Lagrange equation of some variational problem. Thus, the way in which we introduce the concept of stability for $\cH$-hypersurfaces (Definition \[def:stabH\]) is not in a variational way, but in connection with the *linearized equation* of the $\cH$-graph equation in $\R^{n+1}$; see Proposition \[varcar\]. In this way we obtain, for any $\cH$-hypersurface $\Sigma$ in $\R^{n+1}$ a linear *stability operator* of the form $$\label{opL} \cL=\Delta + \esiz \nabla \cdot, X\esde + q$$ for some $X\in \X(\Sigma)$ and $q\in C^{2}(\Sigma)$. After defining stability associated to $\cL$ in the natural way, we will focus on stable $\cH$-surfaces $\Sigma$ in $\R^3$, and will seek radius and curvature estimates for $\Sigma$. In the CMC case, Ros and Rosenberg proved in [@RoRos] that the intrinsic distance from a point $p$ in a stable CMC surface $\Sigma$ to $\partial \Sigma$ is less than $\pi/H$, where $H>0$ is the mean curvature of $\Sigma$. In [@Maz], Mazet improved the previous result obtaining the optimal estimate $\pi /(2H)$. Our approach to proving this type of radius estimates for stable $\cH$-surfaces will be different, and based on arguments introduced by Fischer-Colbrie [@FC] and López-Ros [@LR] for non-negative *Schrodinger operators*, i.e. operators of the form for $X=0$. We will also make use of an argument by Galloway and Schoen [@GaSc], which will let us bound the (non-Schrodinger) stability operator $\cL$ in by a non-negative Schrodinger-type one; see Lemma \[lemgs\]. By using these ideas, we will prove (Theorem \[deste\]) that there exists a uniform radius estimate for stable $\cH$-surfaces in $\R^3$, in the case that $\cH\in C^2(\S^2)$ is positive and satisfies a certain additional condition, see inequality . As a trivial consequence we obtain (Corollary \[cor:estrella\]): *if $\cH\in C^2(\S^2)$ satisfies condition , there are no complete stable $\cH$-surfaces in $\R^3$.* It is important to observe here that some condition on $\cH>0$ is needed in order to prove a radius estimate for stable $\cH$-surfaces. For instance, for some radially symmetric choices of $\cH>0$, there exist complete, non-entire, strictly convex $\cH$-graphs in $\R^3$. Also, in Theorem \[testicu\] we will prove a curvature estimate, of the form $$|\sigma(p)|d_{\Sigma}(p,\partial \Sigma)\leq C,$$ for stable $\cH$-surfaces $\Sigma$ in $\R^3$, where $\cH\in C^2(\S^2)$ satisfies ; here $\sigma$ is the second fundamental form of $\Sigma$. For the case of minimal surfaces and CMC surfaces in $\R^3$, such estimates were obtained by Schoen [@Sc] and Berard-Hauswrith [@BH], respectively. [**Acknowledgements:**]{} This work is part of the PhD thesis of the first author. The authors are grateful to José M. Espinar, Francisco Martín, Joaquín Pérez and Francisco Torralbo for helpful comments during the preparation of this manuscript. Basic properties of $\cH$-hypersurfaces {#sec:2} ======================================= Maximum and tangency principles ------------------------------- Locally, $\cH$-hypersurfaces in $\R^{n+1}$ for some given $\cH\in C^1(\S^n)$ are governed by an elliptic, second order quasilinear PDE. Specifically, let $\cH\in C^1(\S^n)$, let $\Sigma$ denote an $\cH$-hypersurface, and take any $q\in \Sigma$. Then, if we view $\Sigma$ around $q$ as an upwards-oriented graph $x_{n+1}=u(x_1,\dots, x_n)$ with respect to coordinates $x_i$ in $\R^{n+1}$ such that $\{\parc_1,\dots,\parc_{n+1}\}$ is a positively oriented orthonormal frame of $\R^{n+1}$, equation shows that the function $u$ is a solution to $$\label{eqH} {\rm div}\left(\frac{Du}{\sqrt{1+|Du|^2}}\right) = n \cH (Z_u), \hspace{1cm} Z_u:= \frac{(-Du,1)}{\sqrt{1+|Du|^2}},$$ where $\div, D$ denote respectively the divergence and gradient operators on $\R^n$. In particular, $\cH$-hypersurfaces satisfy the Hopf maximum principle (both in its interior and boundary versions), a property that we will use geometrically in the following way: \[tan\] Given $\cH\in C^1(\S^n)$, let $\Sigma_1,\Sigma_2$ be two embedded $\cH$-hypersurfaces, possibly with smooth boundary. Assume that one of the following two conditions holds: 1. There exists $p\in {\rm int}(\Sigma_1)\cap {\rm int}(\Sigma_2)$ such that $\eta_1(p)=\eta_2(p)$, where $\eta_i:\Sigma_i\flecha \S^n$ is the unit normal of $\Sigma_i$, $i=1,2$. 2. There exists $p\in \parc \Sigma_1 \cap \parc \Sigma_2$ such that $\eta_1(p)=\eta_2(p)$ and $\nu_1(p)=\nu_2(p)$, where $\nu_i$ denotes the interior unit conormal of $\parc \Sigma_i$. Assume moreover that $\Sigma_1$ lies around $p$ at one side of $\Sigma_2$. Then $\Sigma_1=\Sigma_2$. In the case that $\cH$ is an odd function, i.e. $\cH(-x)=-\cH(x)$ for every $x\in \S^n$, any $\cH$-hypersurface is also an $\cH$-hypersurface with the opposite orientation. Hence, the tangency principle stated in Lemma \[tan\] can be formulated in a stronger way, similar to the usual tangency principle for minimal hypersurfaces. In particular, we have: \[tangency\] Let $\cH\in C^1(\S^n)$ satisfy $\cH(-x)=-\cH(x)$ for every $x\in \S^n$, and let $\Sigma_1,\Sigma_2$ denote two immersed $\cH$-hypersurfaces in $\R^{n+1}$. Assume that there exists $p\in \Sigma_1\cap \Sigma_2$ with $T_p\Sigma_1=T_p\Sigma_2$ and such that $\Sigma_1$ lies around $p$ at one side of $\Sigma_2$. Then $\Sigma_1=\Sigma_2$. It is immediate that any translation of an $\cH$-hypersurface is again an $\cH$-hypersurface. Moreover, the possible symmetries of the function $\cH$ induce further invariance properties of the class of $\cH$-hypersurfaces, as detailed in the next lemma: \[sime\] Given $\cH\in C^1(\S^n)$, let $\Sigma$ be an $\cH$-hypersurface, and let $\Phi$ be a linear isometry of $\R^{n+1}$ such that $\cH \circ \Phi =\cH$ on $\S^n\subset \R^{n+1}$. Then $\Sigma'=\Phi \circ \Sigma$ is an $\cH$-hypersurface in $\R^{n+1}$ with respect to the orientation given by $\eta':=d\Phi(\eta)$. \[lefti\] The notion of a surface of prescribed mean curvature $\cH$ can be generalized to the context of oriented hypersurfaces in $n$-dimensional Lie groups with a left invariant metric. Specifically, if $\psi:M^n\flecha G^{n+1}$ is a hypersurface in such a Lie group $G$, we can define its *left invariant Gauss map* $\eta: M\flecha \S^n = \{v\in \mathfrak{g} : |v|=1\}$ by left translating to the identity element $e$ of $G$ the unit normal vector of $\psi$ at every point. Thus, given $\cH\in C^1(\S^n)$, we may define a *hypersurface of prescribed mean curvature $\cH$* by equation . See e.g. [@GM1; @GM3]. A particular case: self-translating solitons of mean curvature flows {#solit} -------------------------------------------------------------------- We consider next the particular case in which $\cH\in C^1(\S^n)$ is a linear function, i.e. $\cH(x)= a \esiz x,v\esde +b$ for $a,b\in \R$, $a\neq 0$, and $v\in \S^n$. Up to a homothety and a change of Euclidean coordinates in $\R^{n+1}$, we can assume that $a=1$ and that $v=e_{n+1}$, and so $$\label{hlin} \cH(x)=\esiz x,e_{n+1}\esde +b.$$ Let $f:\Sigma\flecha \R^{n+1}$ be an immersed oriented $\cH$-hypersurface for $\cH$ as in , let $\eta:\Sigma\flecha \S^n$ be its unit normal, and consider the family of translations of $f$ in the $e_{n+1}$ direction, given by $F(p,t)= f(p)+t e_{n+1}$. Then, $F(p,t)$ is a solution to the geometric flow below, which corresponds to the mean curvature flow with a constant forcing term: $$\label{geoflo} \left(\frac{\parc F}{\parc t}\right)^{\perp}= ( H - b)\eta,$$ where $H=H(\cdot,t)$, $\eta=\eta(\cdot,t)$ denote the mean curvature and unit normal of the hypersurface $F(\cdot, t):\Sigma\flecha \R^{n+1}$. In other words, $f:\Sigma\flecha \R^{n+1}$ is a *self-translating soliton* of the geometric flow . Note that when $b=0$, is the usual mean curvature flow (in this paper we use the convention for $H$ that the mean curvature of the unit sphere in $\R^{n+1}$ with its inner orientation is equal to $1$). The converse also holds, i.e. any self-translating soliton to is an $\cH$-hypersurface in $\R^{n+1}$ for $\cH$ as in . There is a second characterization for $\cH$-hypersurfaces with $\cH$ given by . For that, recall following Gromov [@Gr] (see also [@BCMR]), that the *weighted mean curvature* $H_{\phi}$ of an oriented hypersurface $\Sigma$ in $\R^{n+1}$ with respect to the density $e^{\phi}\in C^1 (\R^{n+1})$ is given by $$\label{wmc} H_{\phi}= n H_{\Sigma} - \esiz \eta,D \phi\esde,$$ where $H_{\Sigma}$ is the mean curvature of $\Sigma$ in $\R^{n+1}$, $\eta$ is its unit normal, and $D$ denotes the gradient in $\R^{n+1}$. So, by and the previous discussion we arrive at: \[carwe\] Let $\Sigma$ be an immersed oriented hypersurface in $\R^{n+1}$. The following three conditions are equivalent: 1. $\Sigma$ is an $\cH$-hypersurface, for $\cH$ given by . 2. $\Sigma$ is a self-translating soliton of the mean curvature flow with a constant forcing term . 3. $\Sigma$ has constant weighted mean curvature $H_{\phi}= nb \in \R$ for the density $e^{\phi}$ in $\R^{n+1}$, where $\phi(x):=n \esiz x,e_{n+1}\esde$. As explained in the introduction, there are many relevant works on self-translating solitons of the mean curvature flow (which corresponds to the case $b=0$ in ). The situation when $b\neq 0$ in is much less studied. Some particular references for this situation are [@Es; @Lo]. Compact $\cH$-hypersurfaces in $\R^{n+1}$ ----------------------------------------- \[hpos\] Let $\cH\in C^1(\S^n)$, and assume that $\cH(x_0)=0$ for some $x_0\in \S^n$. Then, there are no compact $\cH$-hypersurfaces in $\R^{n+1}$. Take $\cH\in C^1(\S^n)$ with $\cH(x_0)=0$ for some $x_0\in \S^n$, and let $\Sigma$ denote a compact immersed $\cH$-hypersurface in $\R^{n+1}$. Let $\Pi$ denote an oriented hyperplane in $\R^{n+1}$ with unit normal $x_0$, and such that $\Sigma$ is contained in the half-space $\Pi^+:=\cup \{\Pi_{\landa} : \landa>0\}$, where $\Pi_{\landa}:=\Pi + \landa x_0$. Let $\landa_0$ denote the smallest $\landa>0$ for which $\Sigma\cap \Pi_{\landa} \neq \emptyset$, and let $p\in \Sigma\cap \Pi_{\landa_0}$ be any of such *first contact points*. Note that $\eta(p)=\pm x_0$, where $\eta:\Sigma\flecha \S^n$ is the unit normal to $\Sigma$. If $\eta(p)=x_0$, then both $\Sigma,\Pi_{\landa_0}$ are $\cH$-hypersurfaces and we obtain a contradiction with the maximum principle (Lemma \[tan\]). Hence, $\eta(p)=-x_0$. Consequently, the principal curvatures $\kappa_i(p)$ of $\Sigma$ at $p$ are $\leq 0$, and hence $\cH(-x_0)\leq 0$. As a matter of fact, $\cH(-x_0)<0$, for if $\cH(-x_0)=0$, then $\Pi_{\landa_0}$ with its opposite orientation is also an $\cH$-hypersurface, and we contradict again Lemma \[tan\]. Let now $\landa_1$ be the largest $\landa>0$ for which $\Sigma\cap \Pi_{\landa}\neq \emptyset$, and take $q\in \Sigma\cap \Pi_{\landa_1}$. By previous arguments, $\eta(q)=-x_0$. Since $H_{\Sigma}(q)<0$ (because $\cH(-x_0)<0$), one of the principal curvatures $\kappa_i(q)$ of $\Sigma$ at $q$ is negative. This is a contradiction with the fact that $\Sigma\cap \Pi_{\landa}$ is empty for all $\landa>\landa_1$. This contradiction proves Proposition \[hpos\] By Proposition \[hpos\], we see that given $\cH\in C^1(\S^n)$, a necessary condition for the existence of a compact $\cH$-hypersurface $\Sigma$ is that $\cH>0$ or $\cH<0$ on $\S^n$ (up to a change of orientation, we can assume $\cH>0$ in these conditions). However, this is *not* a sufficient condition. To see this, let $\psi:\Sigma\flecha \R^{n+1}$ denote a compact immersed hypersurface of prescribed mean curvature $\cH\in C^1(\S^n)$, take $v\in \S^n$ and denote $h=\esiz \psi,v\esde$. It is then well known that $\Delta h = n H_{\Sigma} \esiz \eta,v\esde$, where $\eta$ denotes the unit normal of $\Sigma$ and $\Delta$ is the Laplacian on $\Sigma$. By the divergence theorem, and since $H_{\Sigma}=\cH\circ \eta$, we have $$\label{intcon} \int_\Sigma \langle\eta,v\rangle \, \cH(\eta) =0.$$ From , and noting that $\int_{\Sigma} \esiz \eta,v\esde =0$ for every $v\in \S^n$ if $\Sigma$ is compact, we have: \[cornex\] Let $\cH\in C^1(\S^n)$ be of the form $\cH(x)=h_0(x)+b$, where $b\in \R$ and $h_0\in C^1(\S^n)$ is not identically zero and satisfies $h_0(x)\esiz x,v\esde \geq 0$ for every $x\in \S^n$ and for some $v\in \S^n$. Then, there are no compact $\cH$-hypersurfaces in $\R^{n+1}$. For the choice $\cH(x)=\esiz x,v\esde +b$, Corollary \[cornex\] is due to López, see [@Lo]. Lemmas \[tan\] and \[sime\] allow to use the Alexandrov reflection principle in any direction in $\R^{n+1}$ with respect to which the function $\cH$ is symmetric. We singularize the following consequence. \[ale\] Let $\cH\in C^1(\S^n)$ be invariant with respect to $n$ linearly independent geodesic reflections $T_1,\dots, T_n$ of $\S^n$. Then, any compact embedded $\cH$-hypersurface in $\R^{n+1}$ is diffeomorphic to $\S^n$. Moreover, $\Sigma$ is a symmetric bi-graph with respect to hyperplanes $\hat{\Pi}_1,\dots, \hat{\Pi}_n$ parallel to the hyperplanes $\Pi_i$ of $\R^{n+1}$ fixed by $T_i$. Let $v_1,\dots, v_n$ be unit normal vectors to $\Pi_1,\dots, \Pi_n$. By applying the Alexandrov reflection technique with respect to these directions, we conclude using Lemma \[tan\] and Lemma \[sime\] that, for each $i$, $\Sigma$ is a symmetric bi-graph with respect to some hyperplane $\hat{\Pi}_i$ of $\R^{n+1}$ parallel to $\Pi_i$. Define now $\Gamma:=\cap_{i=1}^n \hat{\Pi}_i$, which is a straight line in $\R^{n+1}$. For definiteness, we will assume that $\Gamma$ is the $x_{n+1}$-axis, and so each $v_i$ is horizontal, i.e. $\esiz v_i,e_{n+1}\esde =0$. Let $P_t =\{x_{n+1}=t\}\subset \R^{n+1}$. By compactness of $\Sigma$, there is a smallest interval $[a,b]\subset \R$ such that $P_t\cap \Sigma=\emptyset$ if $t\not\in [a,b]$. Also, as $\Sigma$ is a (connected) symmetric bi-graph with respect to the horizontal directions $v_1,\dots, v_n$, it is clear that $P_a\cap \Sigma=\{({\bf 0},a)\}$ and $P_b\cap \Sigma= \{({\bf 0},b)\}$, and that $P_t\cap \Sigma$ is non-empty and transverse for all $t\in (a,b)$. Let $\Omega\subset \R^{n+1}$ be the domain bounded by $\Sigma$, and define $\Omega_t:=\Omega\cap P_t$ for $t\in (a,b)$. Note that for $t<b$ sufficiently close to $b$, $\Omega_t$ is diffeomorphic to an $n$-dimensional ball $\mathbb{B}_n$. As the intersection $P_t\cap \Sigma$ is transverse for all $t\in (a,b)$, we deduce that $\Omega_t$ is diffeomorphic to $\mathbb{B}_n$ for all $t\in (a,b)$. This implies directly that $\Omega$ is diffeomorphic to $\mathbb{B}_{n+1}$ and so, that $\Sigma$ is diffeomorphic to $\S^n$. Proposition \[ale\] is an extension to $\cH$-hypersurfaces of Alexandrov’s theorem according to which compact embedded CMC hypersurfaces in $\R^{n+1}$ are round spheres. The other fundamental result for compact CMC surfaces in $\R^3$ is Hopf’s theorem (i.e. CMC surfaces diffeomorphic to $\S^2$ immersed in $\R^3$ are round spheres). An extension of Hopf’s theorem to $\cH$-surfaces in $\R^3$ follows from work by the second and third authors [@GM1; @GM2]: \[hopf\] Let $\cH\in C^1(\S^2)$, $\cH>0$, and assume that there exists a strictly convex $\cH$-sphere $S$ in $\R^3$. Then any compact immersed $\cH$-surface of genus zero is a translation of $S$. In [@GG] B. Guan and P. Guan proved the following result: \[tgg\] Let $\cH\in C^2(\S^n)$ be positive and invariant under a group of isometries of $\S^n$ without fixed points. Then there exists a closed strictly convex $\cH$-hypersurface in $\R^{n+1}$. In particular, if $\cH\in C^2(\S^n)$ satisfies $\cH(x)=\cH(-x)>0$ for all $x\in \S^n$, then there exists a closed, strictly convex $\cH$-hypersurface $S_{\cH}$ in $\R^{n+1}$. When $n=2$, the Guan-Guan sphere $S_{\cH}$ is actually the unique immersed $\cH$-sphere in $\R^3$, by Theorem \[hopf\]. By combining Proposition \[ale\], Theorem \[hopf\] and Theorem \[tgg\], we then have: \[3planes\] Let $\Sigma$ be a compact embedded $\cH$-surface in $\R^3$, where $\cH\in C^2(\S^2)$ is positive and invariant under a group of isometries of $\S^2$ without fixed points that contains two geodesic reflections of $\S^2$. Then $\Sigma$ is a translation of the Guan-Guan sphere $S_{\cH}$. Examples of groups of isometries of $\S^2$ in the conditions of Corollary \[3planes\] are those generated by reflections with respect to three linearly independent geodesics of $\S^2$. In particular, the groups of isometries of $\S^2$ that leave invariant a Platonic solid inscribed in $\S^2$ are in the conditions of Corollary \[3planes\]. Construction of flat $\cH$-hypersurfaces {#sec:constru} ---------------------------------------- Let $\Sigma$ be a complete flat $\cH$-hypersurface, i.e. $\Sigma$ is the form $\Sigma = \alfa\times \R^{n-1}$, where $\alfa$ is a complete regular curve in a two-dimensional plane $\Pi\equiv \R^2\subset \R^{n+1}$ (here, we denote $\R^{n-1}\equiv \Pi^{\perp}$). It follows from that $\alfa$ satisfies $$\label{planmin} \kappa_{\alfa} = \frac{1}{n} \cH({\bf n}),$$ where $\kappa_{\alfa}$, ${\bf n}$ denote, respectively, the geodesic curvature and unit normal of the planar curve $\alfa$. This equation might be seen as an analogous of the planar Minkowski problem; note, however, that in our case $\cH$ is not assumed to be positive, and $\alfa$ is not necessarily closed. Let us consider then $\cH\in C^1(\S^n)$, let $\{e_1,e_2\}$ be a positively oriented orthonormal basis of $\Pi$, and define a $2\pi$-periodic function $\hat{\cH} \in C^1(\R)$ by $$\hat{\cH}(\theta):= \frac{1}{n} \cH(-\sin \theta e_1 + \cos \theta e_2).$$ Fix $v=\cos \theta_0 e_1 + \sin \theta_0 e_2\in \Pi$, for some $\theta_0$. If $\hat{\cH}(\theta_0)=0$, the straight line generated by $v$ solves , and the corresponding $\cH$-hypersurface $\Sigma$ is a hyperplane in $\R^{n+1}$. Suppose now that $\hat{\cH}(\theta_0)\neq 0$, and let $I_0\subset \R$ denote the largest interval containing $\theta_0$ where $\hat{\cH}$ does not vanish. Then we may consider $F(x)\in C^2(I_0)$ to be a primitive of $1/\hat{\cH}(x)$ in $I_0$, with $F(\theta_0)=0$. By periodicity of $\hat{\cH}$, we have $F'\geq c>0$ for some $c$. If $\hat{\cH}>0$ everywhere, then $I_0=\R$ and we can define the inverse function $F^{-1}$ of $F$, which is globally defined on $\R$. If $\hat{\cH}=0$ somewhere, then $I_0$ is a bounded open interval $(a,b)$ in $\R$, and $F'(x)\to \8$ as $x\to \{a,b\}$. Thus, the same conclusion for $F^{-1}$ holds. If we write now $\alfa'(s)=\cos \theta(s) e_1 + \sin \theta(s) e_2$ for an arclength parametrization $\alfa(s)$ of $\alfa$, equation is rewritten as $$\label{planmin2} \theta'(s)= \hat{\cH} (\theta(s)),$$ that, with the initial condition $\theta(0)=\theta_0$, has as unique solution $\theta(s)= F^{-1}(s):\R\flecha I_0$. By construction, $\alfa(s)$ is complete. Thus, we arrive at the following result. \[cch\] Let $\cH\in C^1(\S^n)$. Fix $v\in \S^n$ and a two-dimensional linear subspace $\Pi\equiv \R^2$ in $\R^{n+1}$ with $v\in \Pi$. Then, there exists a unique (up to translation) complete regular curve $\alfa=\alfa_{v}$ in $\R^2$ with the following properties: 1. $\Sigma_{v,\Pi}:=\alfa \times \R^{n-1}$ is a (complete, flat) $\cH$-hypersurface in $\R^{n+1}$. 2. $v$ is tangent to $\Sigma_{v,\Pi}$ at some point. Conversely, any complete flat $\cH$-hypersurface in $\R^{n+1}$ is one of the examples $\Sigma_{v,\Pi}$. The hypersurface $\Sigma_{v,\Pi}=\alfa \times \R^{n-1}$ is diffeomorphic to $\S^1\times \R^{n-1}$ or to $\R^n$, depending on whether $\alfa$ is a closed curve or not. By construction, a necessary condition for $\alfa$ to be closed is that $\cH$ restricted to $\S^n\cap \Pi$ never vanishes. Thus, the next result follows directly from the classical solution to Minkowski problem for planar curves: \[difci\] Given $\cH\in C^1(\S^n)$, let $\Sigma_{v,\Pi}=\alfa\times \R^{n-1}$ be one of the $\cH$-hypersurfaces in $\R^{n-1}$ constructed in Proposition \[cch\]. The next two conditions are equivalent: 1. $\Sigma_{v,\Pi}$ is diffeomorphic to $\S^1\times \R^{n-1}$ (i.e. $\alfa$ is closed). 2. If we denote $S^1:= \S^n\cap \Pi$, then $\cH(\xi) \neq 0$ for every $\xi \in S^1$, and $$\int_{S^1} \frac{\xi}{\cH(\xi)} d\xi =0.$$ A compactness theorem {#sec:compac} --------------------- We introduce next a compactness theorem for the space of $\cH$-surfaces in $\R^3$ with bounded second fundamental form that will be used in later sections. The argument in the next theorem is well-known for CMC surfaces, see e.g. Section 2 in [@RST]. We sketch it here for the case where $\cH$ is not constant. \[compa\] Let $(\Sigma_n)_n$ be a sequence of $\cH_n$-surfaces in $\R^3$ for some sequence of functions $\cH_n\in C^k(\S^2)$, $k\geq 1$, and take $p_n\in \Sigma_n$. Assume that the following conditions hold: 1. There exists a sequence of positive numbers $r_n\to \8$ such that the geodesic disks $D_n= D_{\Sigma_n} (p_n, r_n)$ are contained in the interior of $\Sigma_n$, i.e. $d_{\Sigma_n} (p_n,\parc \Sigma_n) \geq r_n$. 2. $(p_n)_n \to p$ for a certain $p\in \R^3$. 3. If $|\sigma_n |$ denotes the length of the second fundamental form of $\Sigma_n$, then there exists $C>0$ such that $| \sigma_n | (x) \leq C$ for every $n$ and every $x \in \Sigma_n$. 4. $\cH_n \to \cH$ in the $C^k$ topology to some $\cH \in C^k(\S^n)$. Then, there exists a subsequence of $(\Sigma_n)_n$ that converges uniformly on compact sets in the $C^{k+2}$ topology to a complete, possibly non-connected, $\cH$-surface $\Sigma$ of bounded curvature that passes through $p$. By conditions (i), (iii) and by virtue of a well-known result in surface theory (see e.g. Proposition 2.3 in [@RST]), there exist positive constants $\delta,M$ that only depend on $C$ (and not on $n$, $\cH_n$ or $\Sigma_n$), such that, if $n$ is large enough: 1. An open neighbourhood of $p_n$ in $D_n\subset \Sigma_n$ is the graph of a function $u_n$ over the Euclidean disk $D_{\delta}:=D(0,\delta)$ of radius $\delta$ in $T_{p_n}\Sigma_n$. 2. The $C^2$ norm of $u_n$ in $D_{\delta}$ is not greater than $M$. Since $\Sigma_n$ is an $\cH_n$-surface, it follows that, in adequate Euclidean coordinates $(x^n,y^n,z^n)$ with respect to which $T_{p_n}\Sigma_n= \{z^n=0\}$, each function $u_n$ is a solution in $D_{\delta}$ to the quasilinear elliptic equation $$\label{eqhn} {\rm div}\left(\frac{Du}{\sqrt{1+|Du|^2}}\right) = 2 \cH_n (Z_u), \hspace{1cm} Z_u:= \frac{(-Du,1)}{\sqrt{1+|Du|^2}}.$$ Note that for $u_n$ can be rewritten as a linear elliptic PDE $L[u_n]=f_n$ where the coefficients of $L$ depend $C^{\8}$-smoothly on $Du_n$, and $f_n$ depends $C^k$-smoothly on $Du_n$. By condition b) above, we have $u_n\in C^{1,\alfa}(D_{\delta})$ for all $n$. Thus, all these coefficients are bounded in the $C^{0,\alfa}(D_{\delta})$ norm. Then, by the usual Schauder estimates (see Gilbarg-Trudinger, [@GT] Chapter 6), for any $\delta' \in (0,\delta)$ we conclude that there exists a constant $C'$ (again independent of $n$) such that $||u_n||\leq C'$ in the $C^{2,\alfa}(D_{\delta'})$ norm. In particular, all coefficients of $L[u_n]=f_n$ are uniformly bounded in the $C^{1,\alfa}(D_{\delta'})$ norm. By repeating this argument we eventually obtain $$||u_n||_{C^{k+2,\alfa}(D_{\delta'})} \leq C'', \hspace{1cm} 0<\alfa <1,$$ for some constant $C''$ independent of $n$. In these conditions, we may apply the Arzela-Ascoli theorem, and deduce by (ii), (iv) that a subsequence of the functions $u_n$ converge on $D_{\delta'}$ in the $C^{k+2}$ topology to a solution $u\in C^{k+2}(D_{\delta'})$ to $$\label{eqhn2} {\rm div}\left(\frac{Du}{\sqrt{1+|Du|^2}}\right) = 2 \cH (Z_u), \hspace{1cm} Z_u:= \frac{(-Du,1)}{\sqrt{1+|Du|^2}}.$$ Thus, the graph of $u$ is an $\cH$-surface in $\R^3$ that by construction passes through $p$, and has second fundamental form bounded by $C$. Consider next some $y\in D_{\delta'}$, and let $q$ be the corresponding point in the graph of $u$. It is then clear that there exist points $q_n\in \R^n$ in the graphs of $u_n$, all corresponding to $y$, and such that $q_n\to q$. Thus, passing to a subsequence if necessary so that condition (i) is fulfilled, we can repeat the same process above, this time with respect to the points $q_n$ and $q$. In this way, we obtain an $\cH$-surface $\Sigma$ in $\R^3$ that extends the graph of $u$ over $D_{\delta'}$. Again by (i), it follows by a standard diagonal process that $\Sigma$ can be extended to a complete $\cH$-surface (which will also be denoted by $\Sigma$), that passes through $p$, and whose second fundamental form is bounded by $C$. Moreover, $\Sigma$ is by construction a limit in the $C^{k+1}$ topology on compact sets of the sequence of surfaces $(\Sigma_n)_n$; note that some other limit connected components could also appear in this process. This completes the proof. Properly embedded $\cH$-surfaces {#sec:structure} ================================ In this section we will study properly embedded $\cH$-surfaces of finite topology in $\R^3$, for $\cH\in C^1(\S^2)$, $\cH>0$. In Section \[diacur\] we will recall a diameter estimate for horizontal sections of graphs with positive mean curvature by Meeks (Lemma \[lemi\]), and we will obtain a curvature estimate for $\cH$-surfaces away from their boundary, see Theorem \[th:curv\] and Remark \[rem:curv\]. In Section \[hees\] we will provide several a priori height estimates for $\cH$-graphs with zero boundary values over closed, not necessarily bounded, planar domains. These estimates will be used in Section \[estru\] to study properly embedded $\cH$-surfaces in $\R^3$. Curvature and horizontal diameter estimates {#diacur} ------------------------------------------- The next result is essentially due to Meeks [@Me]: \[lemi\] Let $\Sigma\subset \R^3$ be a graph $z=u(x,y)$ over a closed (not necessarily bounded) domain of $\R^2$, with zero boundary values. Assume that the mean curvature $H_{\Sigma}$ of $\Sigma$ satisfies $H_{\Sigma}>H_0$ for some $H_0>0$. Then, for every $t>2/H_0$, the diameter of each connected component of $\Sigma\cap \{|z|=t\}$ is at most $2/H_0$. In particular, all connected components of $\Sigma\cap \{|z|\geq t\}$ for $t>2/H_0$ are compact. The argument follows the ideas of the proof of Lemma 2.4 in [@Me], so we will only give here a sketch of it, following a slightly simplified version of Meeks’ original proof, that can be found in [@AEG Theorem 4], or [@EGR Theorem 6.2]. Without loss of generality, we may assume that $u\geq 0$ and that, if $\cU$ denotes the unique connected component of $\R^3$ determined by the plane $z=0$ and the graph $\Sigma$, the mean curvature vector of $\Sigma$ points towards $\cU$. Assume that there exist $t>2/H_0$ and a connected component of $\Sigma\cap\{z=t\}$ with a diameter greater than $2/H_0$. Let $\Omega\subset \R^2$ denote the domain where the graph $\Sigma$ is defined. Then, there exists a simple arc $\Gamma\subset\Omega$ such that the Euclidean distance between its extrema $p_1,p_2$ is greater than $2/H_0$, and so that $u(p)\geq t$ for all $p\in \Gamma$. Besides, there is no restriction in assuming that the Euclidean distance between any other two points of $\Gamma$ is smaller than the distance from $p_1$ to $p_2$. Up to a horizontal isometry, we can take $p_1=(-x_0,0),p_2=(x_0,0)$, with $x_0>1/H_0$. In this way, the “rectangle” surface with boundary $S=\Gamma\times[0,t]$ lies entirely in $\cU$. Besides, $S$ divides the solid region $${\cal C}=\{(x,y,z)\in\R^3:\ |x|\leq x_0,\ 0\leq z\leq t\}$$ into two connected components ${\cal C}_1,{\cal C}_2$. Hence, we can place a sphere of radius $1/H_0$ inside ${\cal C}_1$, and move it continuously towards ${\cal C}_2$ without leaving the interior of ${\cal C}$. Consider now just the piece of sphere that passes through $S$ into ${\cal C}_2$. It its clear that this piece cannot touch $\Sigma$, by the mean curvature comparison principle. Hence, the sphere could go through $S$ completely, and end up being contained in ${\cal C}_2\cap{\cal U}$. But, obviously, a sphere of radius $1/H_0$ cannot be contained in the connected component $\cU$, since we could then move it upwards until reaching a first contact point with $\Sigma$, and this would contradict again the mean curvature comparison principle. The next result is a curvature estimate inspired by [@RST]. \[th:curv\] Let $\Lambda,d,\rho$ be positive constants. Then there exists $C=C(\Lambda,d,\rho)>0$ such that the following assertion is true: Let $\Sigma$ be any immersed oriented surface in $\R^3$, possibly with non-empty boundary, let $\sigma, H_{\Sigma},\eta$ denote, respectively, its second fundamental form, its mean curvature and its unit normal, and assume that: $$\label{asscur1} |H_{\Sigma}| + | \nabla H_{\Sigma} | \leq \Lambda \hspace{1cm} \text{on $\Sigma$}.$$ $$\label{asscur2} \eta(\Sigma)\subset \S^2 \text{ omits a spherical disk of radius $\rho$.}$$ Then, for any $p\in \Sigma_d :=\{q\in \Sigma: d_{\Sigma}(q,\parc \Sigma) \geq d\}$, we have $$|\sigma(p)|\leq C.$$ Arguing by contradiction, assume that the statement is not true, i.e. there is a sequence $f_n:\Sigma_n \flecha \R^3$ of immersed oriented surfaces in $\R^3$ satisfying , , and points $p_n\in \Sigma_n$ such that $d_{\Sigma_n}(p_n,\parc \Sigma_n)\geq d$ and $|\sigma_n(p_n)|>n$ for all $n$, where $\sigma_n$ is the second fundamental form of $f_n$. Note that by rotating each $f_n(\Sigma_n)$ adequately in $\R^3$, we may assume that the Gaussian images of *all* the $f_n:\Sigma_n\flecha \R^3$ omit the open spherical disk of radius $\rho$ centered at the north pole of $\S^2$. Consider the compact intrinsic metric disk $D_n=B_{\Sigma_n}(p_n,d/2)$ in $\Sigma_n$, which by construction is at a positive distance from $\parc \Sigma_n$. Let $q_n$ be the maximum on $D_n$ of the function $$h_n(x)=|\sigma_n(x)|d_{\Sigma_n}(x,\partial D_n)$$ Clearly, $q_n$ lies in the interior of $D_n$, as $h_n$ vanishes on $\parc D_n$. Let $\lambda_n=|\sigma_n(q_n)|$ and $r_n=d_{\Sigma_n}(q_n,\partial D_n)$. Then, $$\label{lann} \lambda_n r_n=|\sigma_n(q_n)| d_{\Sigma_n}(q_n,\partial D_n)=h_n(q_n)\geq h_n(p_n)>\frac{d\,n}{2}.$$ Note that this implies that $(\landa_n)_n\to \8$ as $n \to \8$. Also, note that for every $z_n\in B_{\Sigma_n}(q_n,r_n/2)$ we have $$\label{destr} d_{\Sigma_n}(q_n,\partial D_n)\leq 2 d_{\Sigma_n}(z_n,\partial D_n).$$ Consider now the immersed oriented surfaces $g_n:B_{\Sigma_n}(q_n,r_n/2)\flecha \R^3$ obtained by applying a rescaling of factor $\landa_n$ to the restriction of $f_n$ to $B_{\Sigma_n}(q_n,r_n/2)$; that is, $g_n=\landa_n f_n$ restricted to $B_{\Sigma_n}(q_n,r_n/2)$. For short, we will sometimes write $M_n$ to denote this immersed surface given by $g_n$. By , we have the following estimate for the second fundamental form $\hat{\sigma}_n$ of $M_n$ at any point $z_n$ in $B_{\Sigma_n}(q_n,r_n/2)$: $$\label{unisec} |\hat{\sigma}_{n} (z_n)|= \frac{|\sigma_n(z_n)|}{\lambda_n} =\frac{h_n(z_n)}{\lambda_n d_{\Sigma_n}(z_n,\partial D_n)}\leq\frac{h_n(q_n)}{\lambda_n d_{\Sigma_n}(z_n,\partial D_n)}=\frac{d_{\Sigma_n}(q_n,\partial D_n)}{d_{\Sigma_n}(z_n,\partial D_n)}\leq 2.$$ In particular, the norms of the second fundamental forms of the surfaces $M_n$ are uniformly bounded. Also note that, by construction, $|\hat{\sigma}_n(q_n)|=1$. By , the radii of $M_n$ diverge to infinity (recall that the *radius* of a compact Riemannian surface with boundary is the maximum distance of points in the surface to its boundary). Let now $\widetilde{M}_n$ denote the translation of $M_n$ that takes the point $g_n(q_n)$ to the origin of $\R^3$, and let $\xi_n\in \S^2$ denote the Gauss map image of $M_n$ at $q_n$. After passing to a subsequence, we may assume that $(\xi_n)_n\to \xi$ as $n\to \8$, for some $\xi\in \S^2$. By construction, the norm of the second fundamental form of $\widetilde{M_n}$ is at most $2$, and it is equal to $1$ at the origin. We use next an argument similar to the one in the proof of Theorem \[compa\] to show that a subsequence of the surfaces $\widetilde{M}_n$ converges uniformly on compact sets to a complete minimal surface $M_{\8}$. First, Proposition 2.3 in [@RST] ensures that there exist positive constants $\delta_0,\mu$ (independent of $n$) such that, for any $n$ large enough, we can view a neighborhood of the origin in $\widetilde{M_n}$ as a graph of a function $u_n$ over a disk $D^0_n$ of radius $\delta_0$ of its tangent plane $T_0 \widetilde{M}_n= \xi_n^{\perp}$, and such that $||u_n||_{C^2(D^0_n)} \leq \mu$. Since the vectors $\xi_n$ converge to $\xi$ in $\S^2$, after making if necessary $\delta_0$ (resp. $\mu$) smaller (resp. larger), and for every $n$ large enough, we have that the same properties hold with respect to the $\xi$ direction; that is: 1. An open neighborhood of the origin in $\widetilde{M}_n$ is the graph $x_3=u_n(x_1,x_2)$ of a function $u_n$ over the Euclidean disk $\mathcal{D}_0:=D(0,\delta_0)$ of radius $\delta_0$ in $\Pi_0=\xi^{\perp}$; here $(x_1,x_2,x_3)$ are orthonormal Euclidean coordinates centered at the origin, with $\frac{\parc}{\parc x_3}=\xi$. 2. The $C^2$ norm of $u_n$ in $\mathcal{D}_0$ is at most $\mu$. Let $H_n(x_1,x_2)$ denote the mean curvature function of $\widetilde{M}_n$ in these coordinates. Note that, by and the fact that the factors $\landa_n$ diverge to $\8$, the functions $H_n$ are uniformly bounded in the $C^1(\mathcal{D}_0)$ norm, and as a matter of fact they converge uniformly to zero in that norm. Also note that, since the graph of $u_n$ has mean curvature $H_n$, then $u_n$ is a solution to the linear elliptic PDE for $u$ $$\label{linpde} a_{11}(Du_n) u_{11} + 2 a_{12}(Du_n) u_{12} + a_{22} (Du_n) u_{22} = 2 H_n (1+|Du_n|^2)^{3/2},$$ where $u_{ij}$ denotes second derivatives of $u$ with respect to the variables $x_i,x_j$, and the coefficients $a_{ij}$ are smooth functions. As, by condition b) above, the functions $u_n$ are uniformly bounded in the $C^{1,\alfa}$ norm in $\mathcal{D}_0$, we conclude that all coefficients of are bounded in the $C^{0,\alfa}(\mathcal{D}_0)$ norm. By Schauder theory, the $C^{2,\alfa}$-norms in any $D(0,\delta)\subset\subset \mathcal{D}_0$ of the functions $u_n$ are uniformly bounded. Once here, we may repeat the last part of the proof in Theorem \[compa\] using the Arzela-Ascoli theorem and a diagonal argument, and conclude that a subsequence of the surfaces $\widetilde{M}_n$ converges uniformly on compact sets in the $C^2$ topology to a complete minimal surface $M_{\8}$ of bounded curvature that passes through the origin (note that $M_{\8}$ is minimal since the mean curvatures of $\widetilde{M}_n$ converge by construction to zero). Moreover, the norm of the second fundamental form of $M_{\8}$ at the origin is equal to $1$. Also, since all the surfaces $\widetilde{M}_n$ have been obtained by translations and homotheties in $\R^3$ of the original immersions $f_n:\Sigma_n\flecha \R^3$, and since all the Gauss map images of the $f_n$ omit an open spherical disk of radius $\rho$ of the north pole in $\S^2$, it follows that $M_{\8}$ also omits such an open disk. By a classical result of Osserman, according to which the Gauss map image of a complete non-planar minimal surface in $\R^3$ is dense in $\S^2$, we deduce that $M_{\8}$ is a plane. This contradicts the fact that the norm of the second fundamental form of $M_{\8}$ at the origin is equal to $1$. This contradiction proves Theorem \[th:curv\]. \[rem:curv\] It is clear from the proof that, in Theorem \[th:curv\], one can remove Assumption and ask instead that $\Sigma$ is an $\cH$-surface for some fixed, prescribed, $\cH\in C^1(\S^2)$. In that case, the constant $C$ only depends on $d,\rho$ and the $C^1$ norm of $\cH$ in $\S^2$. It is interesting to compare Theorem \[th:curv\] with the family of catenoids $C_{\ep}$ in $\R^3$, where $\ep>0$ is the necksize. When $\ep\to 0$, the curvature of $C_{\ep}$ blows up at its *waist*. Moreover, if we consider, for $d_0>0$ fixed, the piece $C_{\ep}(d_0)$ of $C_{\ep}$ of all points that are at a distance less than $d_0$ from the waist, then the Gaussian image in $\S^2$ of $C_{\ep}(d_0)$ converges as $\ep \to 0$ to $\S^2$ minus two antipodal points. This shows that condition cannot be avoided in Theorem \[th:curv\]. Height estimates for $\cH$-graphs {#hees} --------------------------------- In Definition \[uhai\], $\parc \Sigma$ is not necessarily bounded, and $\Sigma$ is not compact in general. \[uhai\] Let $\cH\in C^1(\S^2)$, and choose some $v\in \S^2$. We will say that there exists a *uniform height estimate for $\cH$-graphs in the $v$-direction* if there exists a constant $C=C(\cH,v)>0$ such that the following assertion is true: For any graph $\Sigma$ in $\R^3$ of prescribed mean curvature $\cH$ oriented towards $v$ (i.e. $\esiz \eta,v\esde>0$ on $\Sigma$ where $\eta$ is the unit normal of $\Sigma$), and with $\parc \Sigma$ contained in the plane $\Pi=v^{\perp}$, it holds that the height of any $p\in \Sigma$ over $\Pi$ is at most $C$. Clearly, minimal graphs in $\R^3$ do not have a uniform height estimate (e.g., half-catenoids are counterexamples). If $\cH$ is a positive constant, Meeks showed in [@Me] that $\cH$-graphs admit uniform height estimates. However, for a general $\cH\in C^1(\S^2)$ the situation is more complicated. For instance, there exist complete, strictly convex, rotational $\cH$-graphs converging to a cylinder for adequate rotationally symmetric positive functions $\cH\in C^1(\S^2)$; see [@BGM]. The existence of such graphs shows that there are no uniform height estimates for arbitrary choices of $\cH\in C^1(\S^2)$, $\cH>0$. We should also point out that Meeks’ proof uses that the CMC equation is invariant by reflections with respect to tilted Euclidean planes, and this is not the case anymore for a general $\cH\in C^1(\S^2)$, not even in the rotationally symmetric case. Thus, our approach to provide uniform height estimates for $\cH$-graphs relies on different ideas. Let us fix some notation. In the next theorem we will assume after choosing new coordinates $(x_1,x_2,x_3)$ that $v=e_3$. We will denote $\S_+^2 =\S^2 \cap \{x_3> 0\}$, $S^1 =\S^2\cap \{x_3=0\}$, and $\overline{\S_+^2}=\S_+^2\cup S^1$. Let $\cH\in C^1(\overline{\S_+^2})$, $\cH>0$. By an *$\cH$-hemisphere* in the $e_3$-direction we will mean a compact, strictly convex $\cH$-surface $\Sigma_{\cH}$ with boundary, such that ${\rm int}(\Sigma_{\cH})$ is an upwards-oriented graph $x_3=u(x_1,x_2)$ over a $C^2$ regular convex disk in $\R^2$, and whose Gauss map image is $\eta(\Sigma_{\cH})=\overline{\S_+^2}$. Given $\cH\in C^1(\overline{\S_+^2})$, recall that a necessary and sufficient condition for the existence of a closed curve $\gamma\subset \R^2$ such that the cylinder $\gamma \times \R\subset \R^3$ is an $\cH$-surface (see Corollary \[difci\]) is that $\cH(\xi) \neq 0$ for every $\xi \in S^1$, and $$\label{conci}\int_{S^1} \frac{\xi}{\cH(\xi)} d\xi =0.$$ In that case, $\gamma$ is strictly convex and bounds a compact domain in $\R^2$, that we will denote by $\Omega_{\cH}$. \[th:hees\] Given $\cH\in C^1(\overline{\S_+^2})$, $\cH>0$, *any* of the following conditions on $\cH$ imply that there exists a uniform height estimate for $\cH$-graphs in the $e_3$-direction: 1. Condition does *not* hold. 2. Condition holds, and there exists a graph $\Sigma_0$ in the $e_3$-direction, oriented towards $e_3$, over a domain $\Omega\subset \R^2$ that contains $\Omega_{\cH}$, and with the property that $H_{\Sigma_0}(p)>\cH(\eta(p))$ for all $p\in \Sigma_0\cap (\Omega_{\cH}\times \R)$. 3. There exists an $\cH$-hemisphere in the $e_3$-direction. 4. There is some $\cH^*\in C^1(\overline{\S_+^2})$, with $\cH=\cH^*$ in $S^1$ and $\cH^*>\cH$ in $\S_+^2$, for which there exists an $\cH^*$-hemisphere in the $e_3$-direction. 5. ${\rm max} \, \cH < 2\, {\rm min} \, \cH|_{S^1}.$ We will prove Theorem \[th:hees\] in Section \[profhes\], and devote the rest of the present Section \[hees\] to discuss the sufficient conditions described in Theorem \[th:hees\], and to deduce some corollaries from it. The first condition in Theorem \[th:hees\] indicates that for generic, non-symmetric choices of $\cH\in C^1(\overline{\S_+^2})$, $\cH>0$, there exist uniform height estimates for $\cH$-graphs. The second condition gives, for the remaining cases of $\cH$, a very general sufficiency property for the existence of uniform height estimates for $\cH$-graphs, in terms of the existence of an adequate *barrier* $\Sigma_0$. The rest of sufficient conditions in Theorem \[th:hees\] are obtained by applying the second condition to situations where we can construct the barrier $\Sigma_0$. One particular consequence of Theorem \[th:hees\] of special interest is: \[cor:esfe\] Let $\cH\in C^1(\S^2)$, $\cH>0$, and assume that there exists a strictly convex $\cH$-sphere $S_{\cH}$ in $\R^3$. Then, there exist uniform height estimates for $\cH$-graphs with respect to any direction $v\in \S^2$. Simply observe that $\Sigma_{\cH,v}:=\{p\in S_{\cH} : \esiz \eta_S(p),v\esde \geq 0\}$ is an $\cH$-hemisphere in the $v$-direction, and apply item 3 of Theorem \[th:hees\]; here, $\eta_S$ is the unit normal of $S_{\cH}$. The next general result is immediate from item 5 of Theorem \[th:hees\]. Let $\cH\in C^1(\overline{\S_+^2})$, $\cH>0$. Choose any $H_0> {\rm max} \, \cH - 2 {\rm min} \, \cH |_{S^1}$. Then there exists a uniform height estimate for graphs of prescribed mean curvature $\cH+H_0$ in the $e_3$-direction. If we impose some additional symmetry to the function $\cH$, we can obtain from Theorem \[th:hees\] more definite sufficient conditions for the existence of a height estimate. For instance: Let $\cH\in C^2(\overline{\S_+^2})$, $\cH>0$, be rotationally symmetric, i.e. $\cH(x)= \mathfrak{h}(\esiz x,e_3\esde)$ for some $\mathfrak{h}\in C^2([0,1])$. Assume that $\mathfrak{h}'(0)\leq 0$. Then, there is a uniform height estimate for $\cH$-graphs in the $e_3$-direction. Assume first that $\mathfrak{h}'(0)=0$. Then, we can extend $\mathfrak{h}$ to a positive, even, $C^2$ function on $[-1,1]$. Hence, $\cH$ can also be extended to a positive $C^2$ function on $\S^2$, also denoted $\cH$, so that $\cH(x)=\cH(-x)$ for all $x\in \S^2$. By Theorem \[tgg\], there exists a strictly convex $\cH$-sphere $S_{\cH}$. Thus, the result follows from Corollary \[cor:esfe\]. Assume now that $\mathfrak{h}'(0)<0$. Then, we can construct a function $h^*\in C^2([-1,1])$, with $h^*(t)=h^*(-t)$ for all $t$, and such that $h^*(0)=\mathfrak{h}(0)$ and $h^*(t)>\mathfrak{h}(t)$ for all $t\in (0,1]$. Let $\cH^*\in C^2(\S^2)$ be defined by $\cH^*(x):= h^*(\esiz x,e_3\esde)$, and note that $\cH^*(x)=\cH^*(-x)>0$ for all $x\in \S^2$. Arguing as above, there exists an $\cH^*$-hemisphere in the $e_3$-direction. Since it is clear that $\cH^*=\cH$ in $S^1$ and $\cH^*>\cH$ in $\S_+^2$, we concluded the desired result by the fourth sufficient condition in Theorem \[th:hees\]. For the case in which $\cH$ is invariant under the symmetry with respect to a geodesic of $\S^2$, we have an estimate for compact embedded $\cH$-surfaces, not necessarily graphs. \[cor:hees\] Let $\cH\in C^1(\S^2)$, $\cH>0$, satisfy: 1. $\cH(x_1,x_2,x_3)=\cH(x_1,x_2,-x_3)$ for all $x=(x_1,x_2,x_3)\in \S^2$. 2. There exists a uniform height estimate for $\cH$-graphs in the directions of $e_3$ and $-e_3$. Then, there is some constant $C=C(\cH)>0$ such that the following assertion is true: any compact embedded $\cH$-surface $\Sigma$ in $\R^3$, with $\parc \Sigma$ contained in the plane $x_3=0$, lies in the slab $|x_3|\leq C$ of $\R^3$. It is an immediate consequence of the Alexandrov reflection principle applied to the family of horizontal planes in $\R^3$. Note that, due to the symmetry condition imposed on $\cH$, we can apply this reflection method in that particular direction; see Lemma \[sime\]. Proof of Theorem \[th:hees\] {#profhes} ---------------------------- We start by proving the first two items. Arguing by contradiction, assume that $\cH\in C^1(\overline{\S_+^2})$, $\cH>0$, is a function for which there is no uniform height estimate for $\cH$-graphs in the $e_3$-direction. So, there exists a sequence $(\Sigma_n)_n$ of $\cH$-graphs with respect to the $e_3$-direction, oriented towards $e_3$, and with boundary $\parc \Sigma$ contained in $\Pi=e_3^{\perp}=\{x_3=0\}$, and points $p_n\in \Sigma_n$ such that the height of $p_n$ over $\Pi$ is greater than $n$. Note that $\Sigma_n\subset \{x_3\leq 0\}$. Take now $H_0\in (0,{\rm min}\, \cH)$, and denote $\Sigma_n^*:=\Sigma_n\cap \{|x_3| \geq 4/H_0\}$. By hypothesis, $p_n\in \Sigma_n^*$ for $n$ large enough. By Lemma \[lemi\], the connected component $\Sigma_n^0$ of $\Sigma_n^*$ that contains $p_n$ is compact, and contained inside a vertical solid cylinder of radius $2/H_0$. Let $q_n\in \Sigma_n^0$ be a point of maximum height of $\Sigma_n^0$, and let $\Sigma_n^1:=\Sigma_n^0-q_n$ denote the translation of $\Sigma_n^0$ that takes $q_n$ to the origin in $\R^3$. By Theorem \[th:curv\] and Remark \[rem:curv\], the norms of the second fundamental form of the graphs $\Sigma_n^1$ are uniformly bounded by some positive constant $C>0$ that only depends on $H_0$ and $||\cH||_{C^1(\S^2)}$, and not on $n$. Moreover, the distances in $\R^3$ of the origin to $\parc \Sigma_n^1$ diverge to $\8$. By Theorem \[compa\], we deduce that, up to a subsequence, there are smooth compact sets $K_n$ of the graphs $\Sigma_n^1$, all of them containing the origin and with horizontal tangent plane at it, and that converge uniformly in the $C^2$ topology to a complete $\cH$-surface $\Sigma_{\8}$ of bounded curvature that passes through the origin. Let us define next $\nu_{\8}:=\esiz \eta_{\8},e_3\esde$, where $\eta_{\8}$ is the unit normal of $\Sigma_{\8}$. Note that $\nu_{\8}({\bf 0})=1$. As all graphs $\Sigma_n^1$ are oriented towards $e_3$, we deduce that $\nu_{\8}\geq 0$ on $\Sigma_{\8}$. Furthermore, it follows directly from Corollary \[nueq\] (to be proved in Section \[sec:stable\]) that $\nu_{\8}$ is a solution to a linear elliptic equation on $\Sigma_{\8}$ of the form $$\label{saei} \Delta \nu_{\8} + \esiz X,\nabla \nu_{\8}\esde + q \nu_{\8}=0,$$ where $\Delta,\nabla$ denote the Laplacian and gradient operators on $\Sigma_{\8}$, $X\in \X(\Sigma_{\8})$, and $q\in C^2(\Sigma_{\8})$. By the maximum principle for , and the condition $\nu_{\8}\geq 0$, we conclude that either $\nu_{\8} \equiv 0$ on $\Sigma_{\8}$ (which cannot happen since $\nu_{\8}({\bf 0})=1$), or $\nu_{\8}>0$ on $\Sigma_{\8}$. Therefore, $\Sigma_{\8}$ is a local vertical graph, i.e. for every $p\in \Sigma_{\8}$ it holds that $T_p \Sigma_{\8}$ is not a vertical plane in $\R^3$. Once here, and since $\Sigma_{\8}$ is a limit of compact pieces of the graphs $\Sigma_n^1$, it is clear that $\Sigma_{\8}$ is itself a proper $\cH$-graph in $\R^3$ oriented towards $e_3$. By construction, this graph has horizontal tangent plane at the origin, it has bounded second fundamental form, and lies entirely in the closed half-space $\{x_3\geq 0\}$. Moreover, since each $\Sigma_n^0$ lies inside a vertical solid cylinder in $\R^3$ of radius $2/H_0$, we deduce that all points of $\Sigma_{\8}$ lie at a distance in $\R^3$ at most $4/H_0$ from the $x_3$-axis. Since $\Sigma_{\8}$ is a complete proper graph, and at a distance at most $4/H_0$ from the $x_3$-axis, it is clear that for any $q_0\in \parc \Omega$ there exists a diverging sequence of points $(a_n)_n\in \Sigma_{\8}$, whose horizontal projections converge to $q_0$, and such that $\nu_{\8}(a_n)\to 0$. Express now $\Sigma_{\8}$ as $x_3=f(x_1,x_2)$ over an open bounded domain $\Omega\subset\R^2$. Let $L$ denote a straight line in $\R^2$ far away from $\Omega$, and let us start moving it towards $\Omega$ until it reaches a first contact point $q_0\in \parc \Omega$ with the compact set $\overline{\Omega}$. Choose a diverging sequence $(a_n)_n$ in $\Sigma_{\8}$ whose horizontal projections converge to $q_0 \in \partial \Omega$, and consider the vertical translations $\Sigma_\8^n=\Sigma_\8-(0,0,a_n^{3})$, where $a_n^3$ denotes the third coordinate of $a_n$. Up to a subsequence we can suppose that the unit normals of $\Sigma_{\8}$ at $a_n$ converge to a fixed vector $\eta_0\in \S^2$, which is horizontal. Again, a similar compactness argument to the ones above ensures that a subsequence of the graphs $\Sigma_\8^n$ converges to a complete $\cH$-surface $\Sigma_\8^*$, that passes through $q_0\in \parc \Omega$. Note that, by construction, $\Sigma_{\8}^*$ is contained in $\overline{\Omega}\times \R$. Moreover, it is clear from the fact that the sequence $a_n^3$ diverges to $\8$ that $(\Omega\times\R)\cap\Sigma_\8^*=\emptyset$ (observe that $\Sigma_{\8}^*$ is constructed from divergent vertical translations of compact pieces of $\Sigma^*$). This implies that $\Sigma_\8^*$ is contained in $\partial\Omega\times\R$, and that the connected component of $\Sigma_\8^*$ that contains the point $(q_0,0)$ is contained in $(\partial_{0}\Omega)\times\R$, where $\partial_0 \Omega$ denotes the connected component of $\parc \Omega$ that contains $q_0$. If we keep calling this connected component as $\Sigma_{\8}^*$, then we have $\Sigma_{\8}^*=\alfa \times \R$, where $\alfa\subset \parc_{0}\Omega$ is a regular curve in $\R^2$ that verifies equation , since $\Sigma_{\8}^*$ has prescribed mean curvature $\cH$. Note that $\alfa$ is a closed curve, since $\parc\Omega$ is compact, and so $\Sigma_{\8}^*$ is a complete flat cylinder with vertical rulings, diffeomorphic to $\S^1\times \R$, i.e. one of the surfaces in Corollary \[difci\]. Since $\alfa\cap \Omega=\emptyset$, it follows by connectedness that $\Omega$ is contained in one of the two regions of $\R^2$ separated by $\alfa$. Moreover, from the way that the point $q_0\in \parc \Omega$ was chosen, it is clear that $\Omega$ is actually contained in the *inner* region bounded by $\alfa$. To see this, one should observe that the $\Sigma_{\8}^n$ are graphs over $\Omega$, that they converge uniformly on compact sets to $\Sigma_{\8}^*=\alfa\times \R$, and that $\Omega$ does not intersect neither $\alfa$ nor $L$. We are now in the conditions to prove the first two items stated in Theorem \[profhes\]. It is important to recall here that our argument was by contradiction. Since $\alfa$ is a closed curve, it follows from Corollary \[difci\] that equation holds. This proves item 1 of Theorem \[profhes\]. We next prove item 2. Assume that there exists a graph $\Sigma_0$ in the conditions of that item; that is, the domain $\Omega_0\subset \R^2$ over which $\Sigma_0$ is defined contains $\alfa$, and $H_{\Sigma_0}(p) > \cH(\eta(p))$ for all $p\in \Sigma_0\cap (\Omega_{\cH} \times \R)$. In particular, $\overline{\Omega}\subset\Omega_0$, where $\Omega$ is the domain of the graph $\Sigma_{\8}$, which has prescribed mean curvature $\cH$. Recall that $\Sigma_{\8}$ is contained in $\{x_3\geq 0\}$. Hence, we can move $\Sigma_0$ downwards by vertical translations so that its restriction to the compact set $\overline{\Omega}\subset \Omega_0$ is contained in $\{x_3<0\}$, and then start moving it upwards until reaching a first contact point with $\Sigma_\8$. This contradicts the mean curvature comparison principle, since $H_{\Sigma_0} > \cH\circ \eta$ on $ \Sigma_0\cap (\Omega_{\cH} \times \R)$ by hypothesis. This contradiction proves item 2 of Theorem \[profhes\]. \[2item2\] The statement of item 2 also holds for the case that $\Sigma_0$ is an upwards-oriented $\cH$-graph in the $e_3$-direction, defined over a domain that contains $\Omega_{\cH}$ in its interior. The only difference in the proof with this new hypothesis is that the desired contradiction is reached by using the maximum principle of $\cH$-surfaces, and not the mean curvature comparison principle. We next prove item 4, as an application of item 2. Let $\cH, \cH^*$ be in the conditions of item 4, and let $\Sigma_{\cH}^*$ denote an $\cH^*$-hemisphere in the $e_3$-direction. Note that $\Sigma_{\cH}^*$ can be seen as an upwards-oriented graph $x_3=u(x_1,x_2)$ over a closed strictly convex disk $\Omega_0$ with $C^2$ regular boundary $\Gamma=\parc \Omega_0$. By item 1, we may assume that condition holds, and in particular we can consider the closed curve $\gamma$ in $\R^2$ such that $\gamma\times \R$ is an $\cH$-surface in $\R^3$ (see the comments before the statement of Theorem \[th:hees\]). If $\kappa_1^*,\kappa_2^*$ denote the (positive) principal curvatures of $\Sigma_{\cH}^*$, then at any $p\in \Sigma_{\cH}^*$ at which the unit normal $\eta^*(p)$ is a horizontal vector $\xi\in S^1$, we have (since $\cH=\cH^*$ in $S^1$) that $$\label{dit32}\kappa_1^*(p)+\kappa_2^*(p)= 2\cH^*(\xi) = 2\cH(\xi) = \kappa_{\gamma}(p'),$$ where $p'$ is the point of $\gamma$ with unit normal equal to $\xi$, and $\kappa_{\gamma}$ denotes the (positive) geodesic curvature of $\gamma$. This implies that for any unit vector $v\in T_p \Sigma_{\cH}^*$, the second fundamental form $\sigma^*$ of $\Sigma_{\cH}^*$ satisfies $$\label{dit31} 0<\sigma^*_p (v,v)<\kappa_{\gamma}(p').$$ Since $\Sigma_{\cH}^*$ is an $\cH$-hemisphere, the points in $ \parc\Sigma_{\cH}^*=\{p\in \Sigma_{\cH}^* : \esiz \eta^*(p), e_3\esde =0\}$ project regularly onto the convex curve $\Gamma$. Hence, $\gamma,\Gamma$ are two closed, strictly convex planar curves that, by , satisfy the following condition: if $n_{\Gamma}$ and $n_{\gamma}$ denote the inner unit normals of $\Gamma, \gamma$, then: $$\label{comcur} \kappa_{\Gamma}(p)<\kappa_{\gamma}(p') \text{ whenever $n_{\Gamma}(p)=n_{\gamma}(p')$. }$$ We will need at this point the following classical property of convex curves in $\R^2$, whose proof we omit: [*Fact: If $\Gamma,\gamma$ are two closed, strictly convex regular planar curves that satisfy condition , then $\gamma$ is contained in the interior region bounded by some translation of $\Gamma$.*]{} It follows then from the Fact above that, up to a translation of $\Sigma_{\cH}^*$, the convex disk $\Omega_0$ of $\R^2$ over which $\Sigma_{\cH}^*$ is a graph contains $\gamma=\parc \Omega_{\cH}$ in its interior. Hence, we can conclude directly the existence of a uniform height estimate for $\cH$-graphs by applying the already proved item 2 of the theorem. This proves item 4. The proof of item 3 is analogous to the one of item 4, using the alternative formulation of item 2 that is explained in Remark \[2item2\], instead of item 2 itself. We omit the details. We next prove item 5. First, observe again that, by item 1, we may assume that condition holds, and so, we can consider again the closed planar curve $\gamma$ for which $\gamma\times \R$ is an $\cH$-surface in $\R^3$. Denote $H_0:={\rm max} \, \cH$, and let $\S^2(1/H_0)$ be the round sphere of constant mean curvature $H_0$. From the condition on $\cH$ in item 5, we get $2\cH(\xi)>H_0$ for all $\xi\in S^1$. This implies that the geodesic curvature of $\gamma$ satisfies $\kappa_{\gamma}>H_0$ at every point. So, this means that, up to a translation, $\gamma$ is contained in the open disk $D(0,1/H_0)$ of $\R^2$. Once here, we conclude the proof by applying item 2 to the lower hemisphere $\Sigma_0$ of $\S^2(1/H_0)$. This finishes the proof of Theorem \[th:hees\]. Properly embedded $\cH$-surfaces with one end {#estru} --------------------------------------------- The following result is due to Meeks [@Me], and will play a key role in this section. See also Lemma 1.5 of [@KKS]. \[seplem\] Let $\Sigma$ be a surface with boundary in $\R^3$, diffeomorphic to the punctured closed disk $\overline{\D}-\{0\}$. Assume that $\Sigma$ is properly embedded, and that its mean curvature $H_{\Sigma}$ satisfies $H_{\Sigma}(p)\geq H_0>0$ for every $p\in \Sigma$, and for some $H_0$. Let $P_1,P_2$ be two parallel planes in $\R^3$ at a distance greater than $2/H_0$, and let $P_+,P_-$ be the connected components of $\R^3-[P_1,P_2]$, where $[P_1,P_2]$ is the open slab between both planes. Then, all the connected components of either $\Sigma\cap P^+$ or $\Sigma\cap P^-$ are compact. In what follows, we say that a surface $\Sigma$ has *finite topology* if it is diffeomorphic to a compact surface (without boundary) with a finite number of points removed. If $\Sigma$ is properly embedded, each of such removed points corresponds then to an *end* of the surface. Any such end is of *annular type*, that is, the surface can be seen in a neighborhood of such punctures as a proper embedding of the punctured closed disk $\overline{\D}-\{0\}$ into $\R^3$, and thus is in the conditions of Theorem \[seplem\]. In the next theorem one should recall that, by our analysis in Section \[hees\], the existence of uniform height estimates for $\cH$-graphs holds in any direction $v\in \S^2$ for generic choices of $\cH\in C^1(\S^2)$, $\cH>0$. Thus, the second hypothesis in the theorem below, although a necessary one, is relatively weak. \[t1pe\] Let $\cH\in C^1(\S^2)$, $\cH>0$, satisfy: 1. $\cH$ is invariant under the reflection in $\S^2$ that fixes a geodesic $\S^2\cap v^{\perp}$, for some $v\in \S^2$. 2. There exists a uniform height estimate for $\cH$-graphs in the directions of $v$ and $-v$. Then, there is some constant $D=D(\cH,v)>0$ such that the following assertion is true: any properly embedded $\cH$-surface in $\R^3$ of finite topology and one end is contained in a slab of width at most $D$ between two planes parallel to $\Pi=v^{\perp}$. Let $\Pi_1,\Pi_2$ be two planes parallel to $\Pi=v^{\perp}$, at a distance $2d$ greater than $2/H_0$, where $H_0:={\rm min} \, \cH$, and assume that both $\Pi_1,\Pi_2$ intersect $\Sigma$ (if such a pair of planes does not exist, then $\Sigma$ lies in a slab of $\R^3$ of width $2d$, and the result follows). After a change of Euclidean coordinates, we may assume that $v=e_3$, that $\Pi_1=\{x_3=d\}$, and $\Pi_2=\{x_3=-d\}$ for some $d>1/H_0$. Since $\Sigma$ is properly embedded and only has one end, we can write $\Sigma=\Sigma_0 \cup \cA$, where $\Sigma_0$ is a compact surface with boundary, and $\cA$ is a proper embedding of $\overline{\D}-\{0\}$ into $\R^3$. Using this decomposition and the plane separation lemma (Theorem \[seplem\]), we deduce that either $\Sigma\cap \{x_3\geq d\}$ or $\Sigma\cap \{x_3\leq -d\}$ only has compact connected components. Say, for definiteness that $\Sigma\cap \{x_3\geq d\}$ has this property. By Corollary \[cor:hees\], $\Sigma\cap \{x_3\geq d\}$ is contained in the slab $d\leq x_3 \leq d+C$, where $C=C(\cH)$ is the constant appearing in that corollary. In particular, $\Sigma$ is contained in the half-space $\{x_3\leq d+C\}$. Consider next the planes $x_3=d-2C$ and $x_3=-d-2C$. By the same arguments, at least one of $\Sigma\cap \{x_3\geq d-2C\}$ or $\Sigma\cap \{x_3\leq -d-2C\}$ only has compact connected components. In case $\Sigma\cap \{x_3\geq d-2C\}$ had this property, the previous arguments show that $\Sigma$ would lie in the half-space $\{x_3 \leq d-C\}$, which is not possible since $\Sigma$ intersects by hypothesis the plane $x_3=d$. Thus, $\Sigma\cap \{x_3\leq -d-2C\}$ only has compact connected components, and arguing as in the previous paragraph we deduce that $\Sigma$ is contained in the half-space $\{x_3 \geq -d-3C\}$. Hence, $\Sigma$ is contained in a slab of width $D=2d+4C$ between two planes parallel to $v^{\perp}$. This proves Theorem \[t1pe\]. We provide next some corollaries of this result. \[2sim\] Let $\cH\in C^1(\S^2)$, $\cH>0$, satisfy properties (a) and (b) of Theorem \[t1pe\] with respect to two linearly independent directions $v,w\in \S^2$. Then there is some $\alfa=\alfa(\cH,v,w)>0$ such that the following holds: any properly embedded $\cH$-surface of finite topology and one end lies inside a solid cylinder $C(v\wedge w,\alfa)$ of radius $\alfa$ and axis orthogonal to both $v,w$. As a particular case of Corollary \[2sim\], we have: \[cor:pdd\] Let $\cH\in C^1(\S^2)$, $\cH>0$, be rotationally symmetric, i.e. $\cH(x)=\mathfrak{h}(\esiz x,v\esde)$ for some $v\in \S^2$ and some $\ch\in C^1([-1,1])$. Then, any properly embedded $\cH$-surface in $\R^3$ of finite topology and one end lies inside a solid cylinder of $\R^3$ with axis parallel to $v$. By Theorem \[t1pe\] (or by Corollary \[2sim\]), it suffices to show that there exist uniform height estimates for $\cH$-graphs in any direction of $\R^3$ orthogonal to $v$. In order to prove this, we can use the argument in the proof of Theorem 6.2 of [@EGR], which we sketch next. Take $w\in \S^2$ orthogonal to $v$. First observe that, by Lemma \[lemi\], in order to prove existence of uniform height estimates for $\cH$-graphs in the $w$-direction, it suffices to do so for *compact* graphs $\Sigma\subset \R^3$ with $\parc \Sigma$ contained in the plane $\Pi=w^{\perp}$, and with the diameter of each connected component of $\parc \Sigma$ being less than $2/H_0$, where $0<H_0<{\rm min}_{\S^2} \cH$. Thus, let $\Sigma$ be an such graph, and let $\xi\in \S^2$ be another vector orthogonal to $v$, and which makes an angle of $\pi/4$ with $w$. Note that, as $\cH$ is rotationally symmetric, we can apply the Alexandrov reflection principle with respect to the family of planes of $\R^3$ orthogonal to $\xi$. By the argument in [@EGR Theorem 6.2] using reflections in this *tilted direction* $\xi$, it can be shown that there exists $C=C(\cH)>0$ (independent of $\Sigma$) such that the distance of each $p\in \Sigma$ to the plane $\Pi$ is bounded by $C$; we omit the specific details. This proves the desired existence of height estimates, and hence, also Corollary \[cor:pdd\]. For the case of three reflection symmetries for $\cH$, we obtain a more definite classification result. \[3sim\] Let $\cH\in C^2(\S^2)$, $\cH>0$. Assume that $\cH$ is invariant under three linearly independent geodesic reflections of $\S^2$. Then, any properly embedded $\cH$-surface $\Sigma_{\cH}$ in $\R^3$ of finite topology and at most one end is the Guan-Guan sphere $S_{\cH}$ associated to $\cH$. By Theorem \[tgg\] we know that the Guan-Guan strictly convex $\cH$-sphere $S_{\cH}$ exists. By Corollary \[cor:esfe\], there exist uniform height estimates for $\cH$-graphs in any direction $v\in \S^2$. Thus, by Theorem \[t1pe\] applied to the three directions of symmetry of $\cH$, we conclude that $\Sigma_{\cH}$ lies in a compact region of $\R^3$. Since $\Sigma_{\cH}$ is proper in $\R^3$, $\Sigma_{\cH}$ is compact. By Corollary \[3planes\], $\Sigma_{\cH}=S_{\cH}$ (up to translation), what proves the result. \[cor:3sim\] Let $\cH\in C^2(\S^2)$, $\cH>0$, be rotationally symmetric and even, i.e. $\cH(x)=\mathfrak{h}(\esiz x,v\esde)$ for some $v\in \S^2$ and some $\mathfrak{h}\in C^2([-1,1])$ with $\mathfrak{h}(x)=\mathfrak{h}(-x)$. Then, any properly embedded $\cH$-surface in $\R^3$ of finite topology and at most one end is the convex rotational $\cH$-sphere $S_{\cH}$ associated to $\cH$, with rotation axis parallel to $v$. The regularity in Collorary \[cor:3sim\] can be weakened to $\cH\in C^1(\S^2)$, by using the existence of rotational $\cH$-spheres proved in [@BGM]. Moreover, as mentioned earlier, in [@BGM] we show that there exists complete non-entire rotational convex $\cH$-graphs in $\R^3$ for some rotationally symmetric functions $\cH\in C^1(\S^2)$. These examples show that Theorem \[3sim\] is not true if we only assume that $\cH$ is invariant by two geodesic reflections in $\S^2$ and that Corollary \[cor:3sim\] is not true if we do not assume that $\cH$ is even. Stability of $\cH$-surfaces {#sec:stable} =========================== The stability operator of $\cH$-hypersurfaces {#sec:scmc} --------------------------------------------- This section is devoted to the study of stability of $\cH$-hypersurfaces in $\R^{n+1}$. We start by recalling some basic notions about stability of CMC hypersurfaces. Let $\Sigma^n$ be an immersed, oriented hypersurface in $\R^{n+1}$ with constant mean curvature $H\in \R$. We define its *stability operator* (or *Jacobi operator*) as $\mathcal{L}:=\Delta + |\sigma|^2$, where $\Delta$ denotes the Laplacian of $(\Sigma,g)$ and $|\sigma|$ is the norm of the second fundamental form of $\Sigma$. As $\cL$ is a Schrodinger operator, it is $L^2$ self-adjoint. The stability operator $\cL$ appears when considering the second variation of the functional ${\rm Area} - n H \, {\rm Volume}$ of which $\Sigma$ is a critical point. We say that $\Sigma$ is *strongly stable* (or simply *stable*) is $-\cL$ is a non-negative operator, that is, $-\int_{\Sigma} f \cL f \geq 0$ for all $f\in C_0^{\8}(\Sigma)$. By a classical criterion by Fischer-Colbrie [@FC], $\Sigma$ is stable if and only if there exists a positive function $u\in C^{\8}(\Sigma)$ with $\cL u \leq 0$. The stability operator $\cL$ also appears as the *linearized mean curvature operator*. Specifically, consider a normal variation $$\label{normalvar} (p,t)\in \Sigma\times (-\ep,\ep) \mapsto p+ t f(p) \eta(p),$$ of an immersed oriented hypersurface $\Sigma$ in $\R^{n+1}$, where $\eta:\Sigma\flecha \S^n$ is the unit normal of $\Sigma$ and $f\in C_0^{\8}(\Sigma)$. Then, if for each $t\in(-\ep,\ep)$ we denote by $H(t)$ the mean curvature function of the corresponding surface $\Sigma_t$ given by , we have $$\label{limifo}\cL f = n H'(0).$$ Let us consider next the case of $\cH$-hypersurfaces in $\R^{n+1}$, for some $\cH\in C^1(\S^n)$, not necessarily constant. In this general situation, there is no known variational characterization of $\cH$-hypersurfaces similar to the one of the CMC case explained above. Still, we can define a stability operator associated to $\cH$-hypersurfaces by considering the linearization of , just as in the above characterization of the CMC case. We do this next. \[varcar\] Let $\Sigma$ be an $\cH$-hypersurface in $\R^{n+1}$ for some $\cH\in C^1(\S^n)$, let $f\in C_0^{2}(\Sigma)$, and for each $t$ small enough, let $\Sigma_t$ denote the hypersurface given by , where $\eta:\Sigma\flecha \S^n$ is the unit normal of $\Sigma$. Denote $\hat{\cH}(t):= H(t)-\cH(\eta^t):\Sigma\times (-\ep,\ep)\flecha \R$, where $H(t)$ and $\eta^t$ stand, respectively, for the mean curvature and the unit normal of $\Sigma_t$. Then, $$\label{stabh} n \hat{\mathcal{H}}'(0) = \cL f, \hspace{1cm} \cL f:= \Delta f+ \esiz X_{\cH},\nabla f\esde + |\sigma|^2 f,$$ where $X_{\cH}\in \X(\Sigma)$ is given for each $p\in \Sigma$ by $X_{\cH}(p):=n\nabla_{\S} \cH(\eta(p))$; here $\nabla_{\S}$ denotes the gradient in $\S^n$, and $\Delta,\nabla, |\sigma |$ denote, respectively, the Laplacian, gradient and norm of the second fundamental form of $\Sigma$ with its induced metric. From now on we work at some fixed but arbitrary $p\in \Sigma$, which will be omitted for clarity reasons. From $\hat{\cH}(t)=H(t)-\cH(\eta^t)$ and we have $$\label{vc1} \def\arraystretch{1.6}\begin{array}{lll} n\hat{\mathcal{H}}'(0)&=& \Delta f + |\sigma|^2 f - n \left. \frac{d}{dt}\right|_{t=0} (\cH(\eta^t))\\ & = & \Delta f + |\sigma|^2 f - n \left \esiz \nabla_{\S} \cH(\eta(p)),\left. \frac{d}{dt}\right |_{t=0} (\eta^t) \right\esde . \end{array}$$ Let $\{e_1,\dots, e_n\}$ be a positively oriented orthonormal basis of principal directions in $\Sigma$ at $p$. Then, for $t\in (-\ep,\ep)$, this basis transforms via to the positively oriented basis $\{e_1^t,\dots, e_n^t\}$ on $\Sigma_t$ given by $$\label{vc2} e_i^t= (1-t f \kappa_i) e_i + t df(e_i) \eta,$$ where $\kappa_i$ is the principal curvature of $\Sigma$ at $p$ associated to $e_i$. From we get $$\left. \frac{d}{dt}\right|_{t=0} (e_1^t\wedge \cdots \wedge e_n^t)=-nHf \eta + \sum_{i=1}^n df(e_i)e_i = -n H f \eta + \nabla f.$$ Noting that $\eta^t=(e_1^t\wedge \cdots \wedge e_n^t)/|e_1^t\wedge \cdots \wedge e_n^t |$, we have $$\label{vc3} \left. \frac{d}{dt}\right|_{t=0} (\eta^t) = \nabla f.$$ From , , we obtain . This finishes the proof of Proposition \[varcar\]. Proposition \[varcar\] justifies the following definition: \[def:stabH\] Let $\Sigma$ be an $\cH$-hypersurface in $\R^{n+1}$ for some $\cH\in C^1(\S^n)$. The *stability operator* of $\Sigma$ is defined as the operator $\cL$ on $\Sigma$ given for each $f\in C_0^{2}(\Sigma)$ by $$\label{stabo} \cL f:= \Delta f+ \esiz X_{\cH},\nabla f\esde + |\sigma|^2 f, \hspace{1cm} X_{\cH}(p):=n \nabla_{\S} \cH(\eta(p)).$$ Note that when $\cH$ is constant, this definition coincides with that of the standard stability operator of CMC hypersurfaces in $\R^{n+1}$ described above. When $\cH(x)=\esiz x,e_{n+1}\esde$, this notion is also consistent with the usual definition of the stability operator of self-translating solitons of the mean curvature flow, which correspond to the previous choice of $\cH$; see, e.g. [@Es; @IR; @Gu; @SX; @Gr]. Since the property of being an $\cH$-hypersurface is invariant by translations of $\R^{n+1}$, we have: \[nueq\] Let $\Sigma$ be an $\cH$-hypersurface in $\R^{n+1}$ for some $\cH\in C^1(\S^n)$, let $\eta:\Sigma\flecha \S^n$ denote its unit normal, choose $a\in \S^n+$, and define $\nu=\esiz \eta,a\esde \in C^{2}(\Sigma)$. Then $\cL \nu=0$, where $\cL$ is the stability operator of $\Sigma$. Consider the variation of $\Sigma$ $$(p,\landa)\in \Sigma\times \R \mapsto p + \landa a$$ and call $\Sigma_{\landa}:=\Sigma + \landa a$. By the implicit function theorem, we can write this variation in a neighborhood of each $(p_0,0)\in \Sigma\times \R$ as a *normal variation* of the form . Specifically, for any $p$ near $p_0$ in $\Sigma$ and any $t\in (-\ep,\ep)$ with $\ep>0$ small enough we can write $$\label{comparvar} p + t f(p) \eta (p)= \phi(p,t) + \landa(t) a$$ for some smooth function $f$ defined near $p_0$ on $\Sigma$, and where $\phi(p,t)\in \Sigma$ for all $(p,t)$, and $\landa(t)$ is smooth with $\landa(0)=0$ and $\landa'(0)\neq 0$. By taking the normal component of the derivative of with respect to $t$ at $t=0$ we obtain using $\esiz \frac{\parc \phi}{\parc t} ,\eta\esde =0$ that $$\label{landap} f=\landa'(0) \esiz \eta,a\esde = \landa'(0) \nu.$$ By Proposition \[varcar\], $\cL f=0$, since all the surfaces $\Sigma_{\landa}$ (and consequently all surfaces given by $t={\rm constant}$ in ) have the same prescribed mean curvature $\cH\in C^1(\S^n)$. Thus, by , we obtain $\cL \nu =0$ as claimed. Generalized elliptic operators and stable $\cH$-hypersurfaces ------------------------------------------------------------- The following terminology is taken from [@Es]: Let $(\Sigma,\esiz,\esde)$ be a Riemannian manifold. A *generalized Schrodinger operator* $L$ on $\Sigma$ is an elliptic operator of the form $$\label{gensc} L=\Delta + \esiz X, \nabla \cdot \esde + q,$$ where $q\in C^{2}(\Sigma)$, $X\in \X(\Sigma)$, and $\Delta,\nabla$ stand for the Laplacian and gradient operators on $\Sigma$. If $X=\nabla \phi$ for some $\phi\in C^{2}(\Sigma)$, we say that $L$ is a *gradient Schrodinger operator*. If $X=0$, we get a standard Schrodinger operator. Note that the stability operator for hypersurfaces of prescribed mean curvature $\cH\in C^1(\S^n)$ is a generalized Schrodinger operator. As explained in Section \[sec:scmc\], the stability operator for CMC hypersurfaces is a Schrodinger operator, and so it is $L^2$ self-adjoint. For self-translating solitons of the mean curvature flow (i.e. for the choice $\cH(x):=\esiz x,e_{n+1}\esde$), the stability operator is a *gradient* Schrodinger operator, and so it is $L^2$ self-adjoint with respect to an adequate weighted structure. In both cases, the stability operator comes associated to the second variation of an adequate functional. There also exist some special cases of (non-gradient) generalized Schrodinger operators that appear associated to the second variation of some geometric functionals. This is the case of the stability operator of marginally outer trapped surfaces (usually called, in short, MOTS). See e.g. [@AEM; @AMS; @AM; @GaSc]. We should observe here two key difficulties in working with the stability operator $\cL$ of $\cH$-hypersurfaces in , namely: (i) that $\cL$ is not in general $L^2$ self-adjoint, and (ii) that $\cL$ does not come in general from a variational problem. Nonetheless, the following definition is natural, as will be discussed below. \[def:stable\] Let $\Sigma$ be an $\cH$-hypersurface in $\R^{n+1}$ for some $\cH\in C^1(\S^n)$, and let $\cL$ denote its stability operator. We say that $\Sigma$ is *stable* if there exists a positive function $u\in C^{2}(\Sigma)$ such that $\cL u \leq 0$. By the Fischer-Colbrie theorem [@FC], this notion agrees in the CMC case with the standard definition of stability. The definition also agrees when $\cH(x)=\esiz x,e_{n+1}\esde$ with the usual stability notion for self-translating solutions to the mean curvature flow, see e.g. Proposition 2 in [@SX]. Also, the stability notion in Definition \[def:stable\] is also consistent with the notion of *outermost stability* in MOTS theory, see [@AEM Definition 3.1]. Let us give some further motivation for Definition \[def:stable\]. Let $L$ be a generalized Schrodinger operator in a Riemannian manifold $\Sigma$, and let $\Omega\subset \Sigma$ denote a smooth bounded domain. Even though $L$ is not formally self-adjoint, it is well known that there exists an eigenvalue $\landa_0$ of $-L$ in $\Omega$ (with Dirichlet conditions) with the smallest real part, that $\landa_0$ is real, and that its corresponding eigenfunction $\psi$ (unique up to multiplication by a non-zero constant) is nowhere zero; see e.g. Section 5 in [@AM]. We call $\landa_0$ the *principal eigenvalue* of the operator $L$ on $\Omega$. By the argument in Remark 5.2 of [@AM], the existence of a positive function $u>0$ on $\Sigma$ such that $Lu\leq 0$ implies that the principal value $\landa_0$ of $-L$ is non-negative on any smooth bounded domain $\Omega\subset \Sigma$. This justifies again the chosen definition for the stability of $\cH$-hypersurfaces in $\R^{n+1}$. We provide next a geometric interpretation for the stability of $\cH$-hypersurfaces. Let $\Sigma$ denote a stable $\cH$-hypersurface in $\R^{n+1}$, and let $\Omega$ denote a smooth bounded domain of $\Sigma$ whose closure is contained in a larger smooth bounded domain $\Omega'\subset \Sigma$. Then, the principal eigenvalue $\landa_0(\Omega')$ of the stability operator $-\cL$ on $\Omega'$ is non-negative. By the monotonicity of the principal eigenvalue with respect to the inclusion of domains (see e.g. [@Pa]), we deduce that $\landa_0(\Omega)>0$. Let $\{\Omega_t\}_{t\in (-\ep,\ep)}$ be a smooth variation in $\R^{n+1}$ of the compact (with boundary) $\cH$-hypersurface $\Omega_0:=\overline{\Omega}$, so that the boundary is fixed by the variation, i.e. $\parc \Omega_t =\parc \Omega$ for every $t$. Assume moreover that all the hypersurfaces $\Omega_t$ are also compact $\cH$-hypersurfaces with boundary (for the same $\cH\in C^1(\S^n)$). By our arguments in Proposition \[varcar\], this implies the existence of a function $u\in C^{2}(\overline{\Omega})$ satisfying $\cL u=0$ on $\Omega$ with $u=0$ on $\parc \Omega$. However, since $\landa_0(\Omega)>0$ by our previous discussion, this function $u$ cannot exist. In other words: *it is not possible to deform a (strictly) stable $\cH$-hypersurface with boundary $\R^{n+1}$ by fixing its boundary, and in a way that all the deformed hypersurfaces in $\R^{n+1}$ have the same prescribed mean curvature $\cH\in C^1(\S^n)$.* This justifies again the notion of stability for $\cH$-hypersurfaces we have introduced. By Corollary \[nueq\], any $\cH$-graph in $\R^{n+1}$ is stable as an $\cH$-hypersurface. More generally, we have: Let $\Sigma$ be an $\cH$-hypersurface in $\R^{n+1}$ for some $\cH\in C^1(\S^n)$, let $\eta:\Sigma\flecha \S^n$ denote its unit normal, and assume that $\esiz \eta,a\esde >0$ for some $a\in \S^n$. Then $\Sigma$ is stable. In particular, any $\cH$-graph is stable. For the case of compact $\cH$-hypersurfaces, we have another trivial consequence: \[conoco\] There are no compact (without boundary) stable $\cH$-hypersurfaces in $\R^{n+1}$. Let $\Sigma$ be a compact $\cH$-hypersurface in $\R^{n+1}$, and let $\cL$ denote its stability operator. By Corollary \[nueq\], the kernel of $\cL$ has dimension at least $n+1$, so in particular $0$ is an eigenvalue for $-\cL$ that is not simple (we do not fix Dirichlet conditions here, as $\parc \Sigma$ is empty). Hence, the principal eigenvalue $\landa_0(\Sigma)$ of $-\cL$ in $\Sigma$ is negative. Thus, $\Sigma$ is not stable. Radius and curvature estimates for stable $\cH$-surfaces -------------------------------------------------------- The next lemma is essentially due to Galloway and Schoen [@GaSc]: \[lemgs\] Let $(\Sigma,\esiz,\esde)$ be a Riemannian manifold, and let $L=\Delta +\esiz X,\nabla \cdot\esde + q$ denote a generalized Schrodinger operator on $\Sigma$; here $X\in \X(\Sigma)$ and $q\in C^{0}(\Sigma)$. Assume that there exists $u\in C^{2}(\Sigma)$, $u>0$, with $Lu \leq 0$. Then the Schrodinger operator $$\label{simsc} \overline{L}:= \Delta + Q, \hspace{1cm} Q:=q-\frac{1}{2} {\rm div}(X) - \frac{|X|^2}{4},$$ satisfies that $-\overline{L}$ is non-negative, i.e. $-\int_{\Sigma} f \overline{L}f \geq 0$ for every $f\in C_0^{2}(\Sigma)$. The condition $Lu\leq 0$ for $u>0$ can be rewritten as $$\label{gas1} \Delta u+\frac{u}{4}|X+2 \nabla\log u |^2-\frac{u}{4} |X|^2- u |\nabla\log u|^2+q u \leq 0.$$ As $u>0$ is positive, writing $u=e^{\phi}$ we have from $$\Delta \phi+\frac{1}{4}|X+2\nabla \phi|^2-\frac{|X|^2}{4}+q\leq 0.$$ Adding and subtracting $\frac{1}{2} \div X$ we have $$\label{gas2} \div(\nabla \phi + X/2)+|\nabla \phi + X/2|^2-\frac{|X|^2}{4}-\frac{1}{2} \div X+q\leq 0.$$ So, if we define now $Y:=\nabla \phi + X/2$ and $Q:=q-\frac{|X|^2}{4}-\frac{1}{2}\div X$, we have from $$\label{gas3} |Y|^2+Q\leq -\div Y.$$ If we consider now $f\in C_0^2(\Sigma)$, then we obtain from $$\def\arraystretch{1.3}\begin{array}{rcl} f^2|Y|^2+f^2Q&\leq& -f^2 \div Y\\ &=&-\div (f^2 Y)+2f\langle\nabla f,Y\rangle\\ &\leq&-\div(f^2 Y)+2|f||\nabla f||Y|\\ &\leq&-\div(f^2 Y)+|\nabla f|^2+f^2|Y|^2. \end{array}$$ By integrating the above inequality and using that $f$ has compact support, we obtain $$\int_{\Sigma}|\nabla f|^2-Qf^2\geq 0,\ \forall f\in C_0^2(\Sigma),$$ which is equivalent to $-\overline{L}$ being a non-negative operator. This completes the proof. Let us also recall the next result for $n=2$, implicitly proved in [@LR] as a variation of the arguments introduced by Fischer-Colbrie in [@FC]; see Theorem 2.8 in [@MPR]. \[lemlo\] Let $\Sigma$ denote a Riemannian surface, and assume that the Schrodinger operator $-(\Delta -K+c)$ is non-negative for some constant $c>0$, where $K$ denotes the Gaussian curvature of $\Sigma$. Then for every $p\in \Sigma$ we have $$d(p,\parc \Sigma)\leq \frac{2\pi}{\sqrt{3c}},$$ i.e. the radius of any (open) geodesic ball centered at $p$ that is relatively compact in $\Sigma$ is at most $2\pi/\sqrt{3c}$. As a consequence of Lemma \[lemgs\] and Lemma \[lemlo\] we can deduce a distance estimate for stable $\cH$-surfaces in $\R^3$, with $\cH\in C^2(\S^2)$. In Theorem \[deste\] below, $\nabla_{\S}$, $\Delta_{\S}$ and $\nabla^2_{\S}$ denote, respectively, the gradient, Laplacian and Hessian operators on the unit sphere $\S^2$. We should emphasize regarding Theorem \[deste\] below that condition in Theorem \[deste\] cannot be eliminated altogether, or just substituted by the weaker condition $\cH>0$. Indeed, there exist complete strictly convex $\cH$-graphs in $\R^3$ for some choices of $\cH>0$; see [@BGM]. \[deste\] Let $\cH\in C^2(\S^2)$ satisfy on $\S^2$ the inequality $$\label{estrella} 3\cH^2 + {\rm det} (\nabla_{\S}^2 \cH) + \cH \Delta_{\S} \cH - |\nabla_{\S} \cH|^2 - \frac{1}{4}(\Delta_{\S} \cH)^2 \geq c>0$$ for some constant $c>0$. Then, for every stable $\cH$-surface $\Sigma$ in $\R^3$, and for every $p\in \Sigma$, we have $$d(p,\parc \Sigma)\leq \frac{2\pi}{\sqrt{3c}}.$$ Define $$\label{xah} Q_{\cH}:= |\sigma|^2 -\frac{1}{2}{\rm div}_{\Sigma} (X_{\cH}) - \frac{|X_{\cH}|^2}{4}, \hspace{1cm} X_{\cH}:=2 \nabla_{\S} \cH(\eta) \in \X(\Sigma).$$ Here, as usual, $|\sigma|$ denotes the norm of the second fundamental form of $\Sigma$ and $\eta$ its unit normal. Let $\cL$ denote the stability operator of $\Sigma$, defined in . Since $\Sigma$ is stable, it follows by Lemma \[lemgs\] that the operator $-\overline{\cL}:=-(\Delta + Q_{\cH})$ is non-negative on $\Sigma$. Assume for the moment that $$\label{desiQ} Q_{\cH}\geq -K+c$$ holds, where $K$ is the Gaussian curvature of $\Sigma$. In that case, the operator $-(\Delta -K+c)$ will also be non-negative, and Theorem \[deste\] will follow directly from Lemma \[lemlo\]. Thus, it only remains to prove . To do this, we first compute ${\rm div}_{\Sigma}(X_{\cH})$. First, note that if $V\in \X(\Sigma)$ and $\nabla,\overline\nabla$ denote the Riemannian connections of $\Sigma$ and $\R^3$, we have $$\label{dive1}\def\arraystretch{1.5}\begin{array}{lll} \esiz \nabla_V X_{\cH},V\esde &= & 2 \esiz \nabla_V (\nabla_{\S} \cH(\eta)),V\esde = 2 \esiz \overline\nabla_V (\nabla_{\S} \cH(\eta)),V\esde\\ & = & 2 \esiz d(\nabla_{\S} \cH)_{\eta} (d\eta(V)),V\esde = 2 (\nabla_{\S}^2 \cH)_{\eta} (V, d\eta (V)).\end{array}$$ Here, one should observe for the last equality that, since $\esiz \eta,V\esde=0$, then both $V,d\eta(V)$ are tangent to $\S^2$ at the point $\eta$. Consider now at any $p\in \Sigma$ an orthonormal basis $\{e_1,e_2\}$ of principal directions of $\Sigma$, and let $\kappa_1,\kappa_2$ denote their associated principal curvatures. It follows then from that, at $p$, $$\label{dive2} {\rm div}_{\Sigma} (X_{\cH})= \sum_{i=1}^2 \esiz \nabla_{e_i} X_{\cH}, e_i\esde =-2 \sum_{i=1}^2 \kappa_i\alfa_i, \hspace{1cm} \alfa_i:= (\nabla_{\S}^2 \cH)_{\eta(p)} (e_i, e_i).$$ Then, by , and the identity $|\sigma|^2= 4 H_{\Sigma}^2-2K$, we obtain at $p$: $$\label{calQ} \def\arraystretch{2.2} \begin{array}{lll} Q_{\cH}&=& 4H_{\Sigma}^2-2K - |\nabla_{\S} \cH(\eta)|^2+\kappa_1\alfa_1 +\kappa_2\alfa_2 \\ & = & 3H_{\Sigma}^2 - K +\displaystyle \frac{(\kappa_1 - \kappa_2)^2}{4} - |\nabla_{\S} \cH(\eta)|^2 \\ & & + \displaystyle\frac{(\kappa_1+\kappa_2)(\alfa_1+\alfa_2)}{2} + \frac{(\kappa_1-\kappa_2)(\alfa_1-\alfa_2)}{2} \\ & = & 3H_{\Sigma}^2 - K - |\nabla_{\S} \cH(\eta)|^2 + \displaystyle\left(\frac{\kappa_1-\kappa_2}{2} + \frac{\alfa_1-\alfa_2}{2}\right) ^2 \\ & & + H_{\Sigma}(\alfa_1+\alfa_2) - \displaystyle \left(\frac{\alfa_1-\alfa_2}{2}\right)^2\end{array}$$ Next, note that, by definition of $\alfa_i$ in , we have $\alfa_1+\alfa_2 = \Delta_{\S} \cH(\eta)$, and $$\label{desir} \left(\frac{\alfa_1-\alfa_2}{2}\right)^2 \leq \frac{(\Delta_{\S} \cH (\eta))^2}{4} - {\rm det}(\nabla_{\S}^2 \cH)_{\eta}.$$ Thus, it follows from , and the identity $H_{\Sigma}=\cH(\eta)$ that $$\label{calQ2} Q_{\cH} + K \geq 3 \cH(\eta)^2 + \cH(\eta) \Delta_{\S} \cH(\eta) -\frac{(\Delta_{\S} \cH (\eta))^2}{4} + {\rm det}(\nabla_{\S}^2 \cH)_{\eta} - |\nabla_{\S} \cH(\eta)|^2.$$ Since by hypothesis, $\cH$ satisfies on $\S^2$, we conclude from that $Q_{\cH}+K \geq c$, which is . Hence, the proof of Theorem \[deste\] is complete. Let us state two simple remarks that clarify the nature of condition . Let $\cH\in C^2(\S^2)$. Then, there is some $H_0>0$ such that for any $\alfa>H_0$, the function $\cH^*:=\cH+\alfa \in C^2(\S^2)$ satisfies equation . Let $\cH\in C^2(\S^2)$ satisfy . Then $\cH(x)\neq 0$ at every $x\in \S^2$. Indeed, if $\cH(x)=0$, equation at $x$ turns into $${\rm det}(\nabla_{\S}^2 \cH)_{x}-\frac{(\Delta_{\S} \cH (x))^2}{4} \geq c+ |\nabla_{\S} \cH(x)|^2>0 ,$$ which is not possible since the left hand-side of this inequality is always non-positive. The next result follows directly from Theorem \[deste\]. It generalizes the well known fact that there are no complete, stable, non-minimal CMC surfaces in $\R^3$. \[cor:estrella\] Let $\cH\in C^2(\S^2)$ satisfy condition . Then, there are no complete stable $\cH$-surfaces in $\R^3$. Recall that any $\cH$-graph in $\R^3$ is stable. Thus, we get as an immediate consequence of Theorem \[deste\]: Let $\cH\in C^2(\S^2)$ satisfy condition . Then, for any $v\in \S^2$ there exist uniform height estimates for $\cH$-graphs in $\R^3$ in the $v$-direction (see Definition \[uhai\] for this notion). We next obtain a curvature estimate for stable $\cH$-surfaces in $\R^3$, for the case that $\cH$ satisfies condition . It generalizes to $\cH$-surfaces the classical curvature estimate obtained by Schoen in [@Sc] for minimal surfaces; see also Bérard and Hauswirth [@BH] for the constant mean curvature case. \[testicu\] Let $a,c>0$. Then, exists a constant $C=C(a,c)>0$ such that the following statement is true: if $\cH\in C^2(\S^2)$ satisfies $$\label{boundH} |\nabla_{\S} \cH | + |\nabla_{\S}^2 \cH| \leq a$$ on $\S^2$, and $\Sigma$ is any stable $\cH$-surface in $\R^3$ satisfying for the constant $c$, then for any $p\in \Sigma$ the following estimate holds: $$|\sigma(p)|d_{\Sigma}(p,\partial \Sigma)\leq C.$$ Arguing by contradiction, assume that there exists a sequence of functions $\cH_n\in C^2(\S^2)$ satisfying for the constant $a$, a sequence $\Sigma_n$ of stable $\cH_n$-surfaces that satisfy for the constant $c$, and points $p_n\in\Sigma_n$ satisfying $$\label{cest1} |\sigma_n(p_n)|d_{\Sigma_n}(p_n,\partial\Sigma_n)>n,$$ where $\sigma_n$ denotes the second fundamental form of $\Sigma_n$. First, note that implies that $|\sigma_n(p_n)|$ diverges to $\8$, since by Theorem \[deste\] we have for all $n$ a uniform upper bound $d_{\Sigma_n}(p_n,\partial\Sigma_n)\leq d$ for some $d>0$. At this point, we can use the arguments in the proof of Theorem \[th:curv\], with some modifications. Let $d_n:=d_{\Sigma_n}(p_n,\partial\Sigma_n)$, and consider the compact intrinsic balls $D_n=B_{\Sigma_n}(p_n,d_n/2)$. Let $q_n\in D_n$ be the maximum of the function $$h_n(x)=|\sigma_n(x)|d_{\Sigma_n}(x,\partial D_n)$$ on $\overline{D_n}$. Clearly $q_n$ is an interior point of $D_n$, as $h_n$ vanishes on $\parc D_n$. Then, if we denote $\lambda_n=|\sigma_n(q_n)|$ and $r_n=d_{\Sigma_n}(q_n,\partial D_n)$, we have by : $$\label{cest3} \landa_n r_n = h_n(q_n) \geq h_n(p_n) \to \8 \hspace{0.5cm} \text{(as $n\to \8$)}.$$ Let now $H_n$ denote the maximum of the mean curvature function of $\Sigma_n$ restricted to $B_{\Sigma_n}(q_n,r_n/2)$, and define $\lambda^*_n:=\max\{\lambda_n, H_n\}$. Consider the rescalings by $\landa_n^*$ of the immersed surfaces $B_{\Sigma_n}(q_n,r_n/2)\subset\Sigma_n$, in a similar fashion to the argument in Theorem \[th:curv\], and denote these rescalings by $M_n:=\lambda^*_n B_{\Sigma_n}(q_n,r_n/2)$. Let $q_n^*$ be the point in $M_n$ that corresponds to $q_n\in \Sigma_n$, and consider the translated surfaces $M_n^*:=M_n-q_n^*$ which take the points $q_n^*$ to the origin. Note that the distance in $M_n^*$ from the origin to $\parc M_n^*$ is equal to $\landa^*_n r_n/2$, and so it diverges to $\8$ as $n\to \8$, by . The surfaces $M_n^*$ have uniformly bounded second fundamental form, since for any $z_n\in B_{\Sigma_n}(q_n,r_n/2)$ we have $$\frac{|\sigma_n(z_n)|}{\lambda^*_n}\leq \frac{|\sigma_n(z_n)|}{\lambda_n} =\frac{h_n(z_n)}{\lambda_n d_{\Sigma_n}(z_n,\partial D_n)}\leq\frac{d_{\Sigma_n}(q_n,\partial D_n)}{d_{\Sigma_n}(z_n,\partial D_n)}\leq 2,$$ where in the last inequality we have used , that also holds in the present context. Also, each $M_n^*$ is a surface of prescribed mean curvature $\cH_n^*\in C^2(\S^2)$, where $\cH_n^*(x):=\cH_n(x)/\landa_n^*$. By definition of $\landa_n^*$, we have $\cH_n^* \leq 1$ on the Gauss map image $\Omega_n\subset \S^2$ of $M_n^*$ in $\S^2$, for all $n$. Also, note that as $n\to \8$ we have by that $$\label{boundH2} |\nabla_{\S} \cH_n^* | + |\nabla_{\S}^2 \cH_n^* | \to 0$$ globally on $\S^2$. Consequently, a subsequence of the $\cH_n^*$ converges in the $C^1(\S^2)$ topology to a constant $\cH_{\8} \in [0,1]$. It follows then from Theorem \[compa\] that a subsequence of the $M_n^*$ converges uniformly on compact sets in the $C^3$ topology to a complete surface $\Sigma^*$ in $\R^3$ of constant mean curvature $\cH_{\8}$ that passes through the origin. We consider the connected component of $\Sigma^*$ that passes through the origin, which will still be denoted by $\Sigma^*$. Since each $\Sigma_n$ is stable, it follows that each $M_n^*$ is a stable $\cH_n^*$-surface in $\R^3$. So, by Lemma \[lemgs\] (see also the beginning of the proof of Theorem \[deste\]), we see that the Schrodinger operator $-\overline{\cL}_n :=-(\Delta_n + Q_{\cH^*_n})$ is non-negative on $M_n^*$; here $\Delta_n$ denotes the Laplacian operator in $M_n^*$, and $Q_{\cH^*_n}$ is given by . Clearly, the Schrodinger operators $\overline{\cL}_n$ converge to the Jacobi operator $\cL_{\8}$ of the CMC surface $\Sigma^*$. It follows then that $-\cL_{\8}$ is non-negative, i.e. $\Sigma^*$ is stable. As $\Sigma^*$ is also complete, then necessarily $\cH_{\8}=0$ and $\Sigma^*$ is a plane. In particular, the norm of the second fundamental forms of the $M_n^*$ at the origin converge to zero, which implies that $\landa_n/\landa_n^*\to 0$. In particular, by the definition of $\landa_n^*$, for $n$ large enough we have $\landa_n^*=H_n$. This implies that $\cH_n^*(x_n)=1$ for some $x_n\in \S^2$, for each $n$ large enough. This is a contradiction with $\cH_{\8}=0$, which completes the proof of Theorem \[testicu\]. [9]{} The authors were partially supported by MICINN-FEDER, Grant No. MTM2016-80313-P, Junta de Andalucía Grant No. FQM325, and Programa de Apoyo a la Investigacion, Fundacion Seneca-Agencia de Ciencia y Tecnologia Region de Murcia, reference 19461/PI/14. [^1]: Mathematics Subject Classification: 53A10, 53C42

--- abstract: 'We study orbifolds by permutations of two identical N=2 minimal models within the Gepner construction of four dimensional heterotic strings. This is done using the new N=2 supersymmetric permutation orbifold building blocks we have recently developed. We compare our results with the old method of modding out the full string partition function. The overlap between these two approaches is surprisingly small, but whenever a comparison can be made we find complete agreement. The use of permutation building blocks allows us to use the complete arsenal of simple current techniques that is available for standard Gepner models, vastly extending what could previously be done for permutation orbifolds. In particular, we consider $(0,2)$ models, breaking of $SO(10)$ to subgroups, weight-lifting for the minimal models and B-L lifting. Some previously observed phenomena, for example concerning family number quantization, extend to this new class as well, and in the lifted models three family models occur with abundance comparable to two or four.' author: - | M. Maio$^{1}$ and A.N. Schellekens$^{1,2,3}$\  \  \ \ $^1$Nikhef Theory Group, Amsterdam, The Netherlands\  \ $^2$IMAPP, Radboud Universiteit Nijmegen, The Netherlands\  \ $^3$Instituto de Física Fundamental, CSIC, Madrid, Spain title: Permutation orbifolds of heterotic Gepner models --- -19cm Introduction ============ Heterotic string theory [@Gross:1984dd] is the oldest approach towards the construction of the standard model in string theory. It owes its success to the fact that the gross features of the standard model appear to come out nearly automatically: families of chiral fermions in representations that are structured as in $SO(10)$-based GUT models. Presumably we know only a very small piece of this landscape of perturbative heterotic strings. Several computational methods are available, based either on space-time geometry or world-sheet conformal field theory. The latter category can be further subdivided into free and interacting field theory methods. But all these methods have obvious but inestimable limitations, making it difficult to determine how much of the heterotic landscape has really been explored so far. In the case of the conformal field theory (CFT) methods [@Belavin:1984vu], these limitations are due to the constraints imposed by world-sheet supersymmetry and modular invariance, for which we only know some classes of special solutions. A general heterotic CFT consists of a right-moving sector that has $N=2$ world-sheet supersymmetry and a non-supersymmetric left-moving sector. Most existing work has been limited either to free CFTs (bosons, fermions or orbifolds) for these two sectors, or to interacting CFTs where the bosonic sector is essentially a copy of the fermionic one. Furthermore the interacting CFTs themselves have mostly been limited to tensor products of $N=2$ minimal models [@Gepner:1987qi; @Gepner:1987vz]. Already in the late eighties of last century ideas were implemented to reduce some of these limitations of interacting CFTs. Instead of minimal models, Kazama-Suzuki models were used [@Kazama:1988qp]. Another extension was to consider permutation orbifolds of $N=2$ minimal models [@Klemm:1990df; @Fuchs:1991vu]. But both of these ideas could only be analyzed in a very limited way themselves. The real power of interacting CFT construction comes from the use of simple current invariants [@Schellekens:1989am; @Schellekens:1989dq; @Intriligator:1989zw; @GatoRivera:1991ru; @Kreuzer:1993tf], which greatly enhance the number and scope of the possible constructions. In particular the left-right symmetry of the original Gepner models could be broken by considering asymmetric simple current invariants [@Schellekens:1989wx], allowing for example a breaking of the canonical $E_6$ subgroup to $SO(10)$, $SU(5)$, Pati-Salam models or even just the standard model (with some additional factors in the gauge group). However, precisely this powerful tool is not available at present in either Kazama-Suzuki models or permutation orbifolds. The original computations were limited to diagonal invariants, where with a combination of a variety of tricks the spectrum could be obtained. Up to now, all that is available in the literature is a very short list of Hodge numbers and singlets for $(2,2)$ spectra with $E_6$ gauge groups [@Klemm:1990df; @Fuchs:1991vu; @Font:1989qc; @Schellekens:1991sb]. To use the full power of simple current methods we need to know the exact CFT spectrum and the fusion rules of the primary fields of the building blocks. The former has never been worked out for Kazama-Suzuki models, and the latter was not available for permutation orbifolds until last year. Using pioneering work by Borisov, Halpern and Schweigert [@Borisov:1997nc] and an extension of these results to fixed point resolution matrices [@Maio:2009kb; @Maio:2009cy; @Maio:2009tg] we have constructed the $\mathbb{Z}_2$ permutation orbifolds of $N=2$ minimal models [@Maio:2010eu]. These can now be used as building blocks in heterotic CFT constructions, on equal footing, and in combination with all other building blocks, such as the minimal models themselves and free fermions. Furthermore we can now for the first time apply the full simple current machinery in exactly the same way as for the minimal models. Meanwhile, another method was added to this toolbox, allowing us to advance a bit more deeply into the heterotic landscape, and away from free or symmetric CFTs. This is called “heterotic weight lifting" [@GatoRivera:2009yt], a replacement of $N=2$ building blocks in the bosonic sector by isomorphic (in the sense of the modular group) $N=0$ building blocks (more precisely, replacing $N=2$ building blocks together with the extra $E_8$ factor). This method requires knowledge of the exact CFT spectrum, which indeed we have. A variant of this idea is the replacement of the $U(1)_{B-L}$ factor ($\times E_8$) by an isomorphic CFT. This has been called “B-L lifting". In a series of papers [@GatoRivera:2010gv; @GatoRivera:2010xn; @GatoRivera:2010fi; @GatoRivera:2010xnb], all combinations of the aforementioned tools were applied to Gepner models. The focus of this work was on three important features of the spectra of these models: do the chiral fermions really form standard model families, do the non-chiral fermions respect standard model charge quantization, and do three families occur reasonably often. The answers to these questions can be summarized as “sometimes, no, and yes". A substantial fraction of the chiral spectra with broken GUT symmetry is organized in standard model families, but the rest has chiral exotics. In the exact spectrum there are almost always vector-like states which violate the charge quantization rules of the standard model and hence would lead to light fractionally charged particles, unless they are lifted by moving into moduli space, away from the special RCFT points. The conclusion on the number of families was a bit more optimistic. Despite the difficulties of standard Gepner models to yield three families, this problem disappears if one moves away from symmetric CFTs. Just using asymmetric MIPFs is not sufficient, but if one really makes the left and right CFTs distinct by using heterotic weight lifting or B-L lifting one obtains a normal, more or less exponential distribution $\propto e^{-\alpha N}$ with $\alpha$ in the range $.2 - .3$, where $N$ is the number of families. The number three is not strongly suppressed in these distributions, and occurs for a total of about $15\%$ of the $N$-family models. The purpose of this paper is to put all these ingredients together using permutation orbifolds of $N=2$ minimal models as building blocks in combination with minimal models. We want to do this for the following reasons: - [Check the consistency of the permutation orbifold CFTs we presented in [@Maio:2010eu]. Chiral heterotic spectra are very sensitive to the correctness of conformal weights and ground state dimensions of the CFT, as well as the correctness of the simple current orbits. This is especially true for weight-lifted spectra, because they have non-trivial Green-Schwarz anomaly cancellations.]{} - [Compare our results with those of previous work on permutation orbifolds [@Klemm:1990df; @Fuchs:1991vu]. These results were obtained using a rather different method, by applying permutations directly to complete heterotic string spectra.]{} - [Check if the aforementioned trends on fractional charged and family number are confirmed also in the class of permutation orbifolds.]{} - [Add a few more items to the growing list of potentially interesting three-family interacting CFT models.]{} The key ingredient of the present work is is our previous paper [@Maio:2010eu], where we studied permutations, together with extensions in all possible order, and found very interesting novelties. For example, we determined how to construct a supersymmetric permutation of minimal models: in particular, the world-sheet-supersymmetry current in the supersymmetric orbifold turns out to be related to the anti-symmetric representation of the world-sheet-supersymmetry current of the original minimal model. When the symmetric representation is used, instead, one ends up with a conformal field theory, which is isomorphic to the supersymmetric orbifold, but it is not supersymmetric itself. In the extended permuted orbifolds so-called exceptional simple currents appear, which originate from off-diagonal representations. Generically, there are many of them, depending on the particular model under consideration, and they do not have fixed points. However, when and only when the “level” is equal to $k=2$ mod $4$, four of all these exceptional currents do admit fixed points. As a consequence, in those cases the knowledge of the modular $S$ matrix is plagued by the existence of non-trivial and unknown $S^J$ matrices (one $S^J$ matrix for each exceptional current $J$). The full set of $S^J$ matrices is available for standard $\mathbb{Z}_2$ orbifolds (see [@Maio:2009kb; @Maio:2009cy; @Maio:2009tg]), but not for their (non)supersymmetric extensions, due to these four exceptional currents with fixed points. This is known as the fixed point resolution problem [@Schellekens:1989uf; @Schellekens:1990xy; @Fuchs:1996dd; @Fuchs:1995zr; @Schellekens:1999yg; @Fuchs:1995tq]. The knowledge of the full set of the $S^J$ matrices is more important in orientifold models [@Pradisi:1996yd; @Pradisi:1995pp; @Fuchs:2000cm] than in heterotic strings, because in the former case one cannot even compute spectra if these matrices are not known. In this paper we consider permutations in Gepner models. One starts with Gepner’s standard construction where the internal CFT is a product of $N=2$ minimal models. Sometimes there are (at least) two $N=2$ identical factors in the tensor product. When it is the case, we can replace these two factors with their permutation orbifold. Moreover, one also has to impose space-time and world-sheet supersymmetry, which is achieved by suitable simple-current extensions. The plan of the paper is as follows. In section \[Section: Permutation orbifold\] we start reviewing the permutation orbifold, which is the main idea used in this paper. In section \[Section: Heterotic Gepner models\] we review the standard constructions of Heterotic Gepner models. In section \[Section: Permutation orbifold of N=2 minimal models\] we review the main ingredients and the most relevant results of $\mathbb{Z}_2$ permutation orbifolds when applied to $N=2$ minimal models. In section \[Section: Lifts\] we describe the heterotic weight lifting and the B-L lifting procedures, which allow us to replace the trivial $E_8$ factor plus either one $N=2$ minimal model or the $U(1)_{B-L}$ with a different CFT, which has identical modular properties, in the bosonic (left) sector. In section \[Section: Comparison\] we compare our results on (2,2) spectra with the known literature. In section \[Section: Results\] we present our phenomenological results concerning the family number distributions, gauge groups, fractional charges and other relevant data. In appendix \[Section: Simple current invariants\] we derive a few facts about simple current invariants. Appendix A.2 contains tables summarizing the main results for the four cases (standard Gepner models and the three kinds of lifts). Permutation orbifold {#Section: Permutation orbifold} ==================== In this section let us recall a few properties of the generic permutation orbifold [@Borisov:1997nc], restricted to the $\mathbb{Z}_2$ case: $$\mathcal{A}_{\rm perm}\equiv \mathcal{A}\times \mathcal{A}/\mathbb{Z}_2\,.$$ If $c$ is the central charge of $\mathcal{A}$, then the central charge of $\mathcal{A}_{\rm perm}$ is $2c$. The typical[^1] weights of the fields are: - $h_{(i,\xi)}=2 h_i$ - $h_{\langle i,j\rangle}=h_i+h_j$ - $h\widehat{(i,\xi)}=\frac{h_i}{2}+\frac{c}{16}+\frac{\xi}{2}$ for diagonal, off-diagonal and twisted representations. Sometimes it can happen that the naive ground state has dimension zero: then one must go to its first non-vanishing descendant whose weight is incremented by integers. The permutation orbifold characters $X$ are related to the characters $\chi$ of the mother theory $\mathcal{A}$ in the following way [@Borisov:1997nc]: \[BHS characters\] $$\begin{aligned} X_{\langle i,j \rangle}(\tau)&=&\chi_{i}(\tau)\cdot\chi_{j}(\tau) \\ X_{(i,\xi)}(\tau)&=&\frac{1}{2}\chi_{i}^2(\tau)+e^{i\pi\xi}\frac{1}{2}\chi_{i}(2\tau) \\ X_{\widehat{(i,\xi)}}(\tau)&=&\frac{1}{2}\chi_{i}(\frac{\tau}{2})+e^{-i\pi\xi}\,T_i^{-\frac{1}{2}}\,\frac{1}{2}\chi_{i}(\frac{\tau+1}{2})\end{aligned}$$ where $T_i^{-\frac{1}{2}}=e^{-i\pi(h_i-\frac{c}{24})}$. The character expansion: $$\chi(\tau)=q^{h_{\chi}-\frac{c}{24}}\,\sum_{n=0}^\infty d_n q^n \qquad \qquad ({\rm with}\,\,q=e^{2i\pi\tau})$$ (the $d_n$’s are non-negative integers) implies a similar expansion for the $X$: $$X(\tau)=q^{h_X-\frac{c}{12}}\,\sum_{n=0}^\infty D_n q^n$$ where the $D_n$’s are expressed in terms of the $d_n$’s as follows: \[d-D relations\] $$\begin{aligned} D^{\langle i,j \rangle}_k &=& \sum_{n=0}^k d_n^{(i)}\,d_{k-n}^{(j)} \\ D^{(i,\xi)}_k &=& \frac{1}{2} \sum_{n=0}^k d_n^{(i)}\,d_{k-n}^{(i)}+ \left\{ \begin{array}{lr} 0 & {\rm if}\,\,k={\rm odd} \\ \frac{1}{2}\,e^{i\pi\xi}\,d_{\frac{k}{2}}^{(i)} & {\rm if}\,\,k={\rm even} \end{array} \right. \\ D^{\widehat{(i,\xi)}}_k &=& d_{2k+\xi}^{(i)}\end{aligned}$$ Using these characters, one can compute their modular transformation and find the orbifold $S$ matrix. It was determined in [@Borisov:1997nc] and will be referred to as $S^{BHS}$. In theories where the permutation orbifold is extended by means of simple currents, one has to face the problem of fixed point resolution, namely one has to determine a set of “$S^J$ matrices”, one for each simple current $J$ [@Schellekens:1990xy]. For standard permutation orbifolds, it was shown in [@Maio:2009kb] that the simple currents can be only of diagonal type, i.e. $(J,\psi)$ where $J$ is a simple current of the mother theory and $\psi=0$ or $1$ for $\mathbb{Z}_2$ orbifolds. This set of matrices was determined in [@Maio:2009tg]. This fixed point resolution plays a crucial rôle in determining the fusion rules of the supersymmetric permutation orbifold, as explained in [@Maio:2010eu]. However, we will not need the explicit formulas for the fixed point resolution matrices in the present paper, and therefore we will not present them here. Heterotic Gepner models {#Section: Heterotic Gepner models} ======================= In this section we review the construction of four-dimensional heterotic string theory. The starting point is a set of bosons $X^\mu$ ($\mu=0,\dots,3$) for both the right and left movers, a right-moving set of NSR fermions $\psi^\mu$, plus corresponding ghosts, and an internal CFT with central charges $(c_L,c_R)=(22,9)$, that we denote by $\mathscr{C}_{22,9}=\mathscr{C}_{22}\times\mathscr{C}_{9}$. Observe that the right-moving superconformal field theory $(X,\psi)$+ghosts has central charge $c=3$. Equivalently, one can think of it as the conformal field theory of two bosons $X^i$ and their fermionic superpartners $\psi^i$ in light-cone gauge, which form an $SO(2)_1$ abelian algebra, with central charge $c=1$. The next step is to replace the NSR $SO(2)_1$ fermions by a set of $13$ bosonic fields living in the maximal torus of an $(E_8)_1\times SO(10)_1$ affine Lie algebra. This is the bosonic string map [@LLS], which transforms the fermionic CFT into a bosonic one with same modular properties. The total right-moving CFT has now central charge equal to $c_R=2+9+13=24$, as the left-moving bosonic theory. Hence, all four-dimensional heterotic strings correspond to all compactified bosonic strings with an internal sector: $$\mathscr{C}_{22,9}\times \left( (E_8)_1\times SO(10)_1 \right)_R \,.$$ To summarize: $$\begin{aligned} \hbox{Left-moving}&&(X^\mu,{\rm ghost})\times \mathscr{C}_{22}\nonumber\\ \hbox{Right-moving}&&(X^\mu,{\rm ghost})\times \mathscr{C}_{9}\times (E_8)_1\times SO(10)_1\nonumber\end{aligned}$$ with $\mu=0,\dots,3$. Equivalently, in light-cone gauge one uses $X^i$ instead of $(X^\mu,{\rm ghost})$. In the right-moving sector all CFT building blocks have $N=2$ worldsheet supersymmetry. This implies the existence of two operators with simple current fusion rules: the worldsheet supercurrent $T_F$ and the spectra flow operator $S_F$. In general, the internal CFT in the fermionic sector is itself built out of $N=2$ building blocks, that have such currents as well. In order to preserve right-moving world-sheet supersymmetry, the total supercurrent $T^{\rm st}_F+T_F^{\rm int}$ must have a well-defined periodicity, since it couples to the gravitino. Here, $T^{\rm st}_F=\psi^\mu \partial X_\mu$ is the world-sheet supercurrent in space-time and $T_F^{\rm int}$ is the supercurrent of the internal sector. Hence the allowed states will have the same spin structure in all the subsectors of the tensor product, namely the R (NS) sector of $SO(10)_1$ must be coupled to the R (NS) sector of the internal CFT. This result is achieved by an integer-spin simple current extension of the full right-moving algebra by the current given by the product of the supercurrents $T^{\rm st}_F\cdot T_F^{\rm int}$: it corresponds to projecting out all the combinations of mixed spin structures. When the internal CFT is a product of many sub-theories, as in the case of Gepner models, each with its own world-sheet supercurrent $T_{F,i}$, then one has to extend the full right-moving algebra by all the currents $T^{\rm st}_F\cdot T_{F,i}$. In simple current language this means that we extend the chiral algebra by all currents $$\label{WS} W_i= (0,\ldots,0,T_{F,i},0,\ldots,0;V)\ ,$$ where we use a semi-colon to separate the internal and space-time part, and we use the standard notation $0,V,S,C$ for $SO(10)_1$ simple currents (or conjugacy classes). A sufficient and necessary condition for space-time supersymmetry is the presence of a right-moving spin-$1$ chiral current transforming as an $SO(10)_1$ spinor. Hence this current must be equal to the product of the spinor $S$ of the $SO(10)_1$, which has spin $h=\frac{5}{8}$, times an operator $S^{\rm int}$ from the Ramond sector of the internal CFT $\mathscr{C}_9$, which must then have spin $h=\frac{3}{8}$. This last value saturates the chiral bound $h\geq\frac{c}{24}$ for the internal right-moving CFT which has central charge $c=9$, hence $S^{\rm int}$ corresponds to a Ramond ground state. Among the Ramond ground states, one is very special. $N=2$ supersymmetry possesses a one-parameter continuous automorphism of the algebra, known as spectral flow, which, when restricted to half-integer values of the parameter, changes the spin structures and maps Ramond fields to NS fields, hence uniquely relating fermionic to bosonic fields. In particular, under spectral flow, the NS field corresponding to the identity is mapped to a Ramond ground state which has $h=\frac{c}{24}$ and is called the spectral-flow operator. Not surprisingly, the spectral flow operator is related to the $N=1$ space-time supersymmetry charge. We will denote it as $S_F$. In our set-up of four dimensional heterotic string theories, $N=1$ space-time supersymmetry is achieved again by a simple current extension. The current in question is the product of the space-time spin field $S_F^{\rm st}$ with $S_F^{\rm int}$, where $S_F^{\rm int}$ is the spectral-flow operator. If the internal CFT is built out of many factors, then $S_F^{\rm int}=\bigotimes_i S_{F,i}$, where $S_{F,i}$ is the spectral-flow operator in each factor. In simple current language, the space-time supersymmetry condition amounts to extending the chiral algebra of the CFT by the simple current $$\label{SS} S_{\rm susy}=(S_{F,1},\ldots,S_{F,r};S)\ ,$$ where $r$ denotes the number of factors. Obviously these simple current extensions must be closed under fusion, in combination with all world-sheet supersymmetry extensions discussed above. Modular invariance of the final theory is then guaranteed by the simple current construction. So far everything holds for any combination of superconformal $N=2$ building blocks. The only ones available in practice (prior to this paper) are suitable combinations of free bosons and/or fermions, and $N=2$ minimal models. These are unitary finite-dimensional representations of the $N=2$ superconformal algebra, which exist only for $c\leq3$. They are labelled by an integer $k$, in terms of which the central charge is $$c=\frac{3k}{k+2}\,.$$ It is not a coincidence that $c$ is equal to the central charge of the $SU(2)_k$ affine Lie algebra. In fact, the $N=2$ minimal models admit a description in terms of the coset[^2] $$\frac{SU(2)_k \times U(1)_4}{U(1)_{2(k+2)}}\,.$$ According to this decompositions, $N=2$ primary fields are products of parafermionic representations (which are primaries of the $\frac{SU(2)}{U(1)}$ algebra) times a bosonic exponential and the $S$ matrix is a product of the matrices $S^{SU(2)_k}$, $S^{U(1)_4}$ and $\left(S^{U(1)_{2(k+2)}}\right)^{-1}$. Representations are labelled by three integers $(l,m,s)$, where $l$ is an $SU(2)_k$ quantum number and $m$ and $s$ are $U(1)$ labels, satisfying the constraint $l+m+s=$even. The range is: $l=0,\dots,k$, $m=-k-1,\dots,k+2$, $s=-1,\dots,2$ ($s=0,2$ for NS sector, $s=\pm1$ for R sector). Moreover, fields are pairwise identified according to: $$\phi_{l,m,s}\sim\phi_{k-l,m+k+2,s+2}\,,$$ which is realized as a formal simple current extension. Now consider the right-moving algebra of the heterotic string. The internal CFT $\mathscr{C}_9$ can be realized as a product of a sufficient number of $N=2$ minimal models such that $$\label{MinSum} \sum_i^r\frac{3k_i}{k_i+2}=9\,,$$ so the full algebra is $$\bigotimes_i (N=2)_i \otimes (E_8)_1\otimes SO(10)_1$$ and representations are labelled by $$\bigotimes_i (l_i,m_i,s_i) \otimes (1) \otimes (s_0)\,.$$ Observe that the $(E_8)_1$ algebra has only one representation, i.e. the identity, and it is often omitted in the product. Here $s_0$ denotes one of the four $SO(10)_1$ representations, $s_0=O,V,S,C$. As discussed above, we impose world-sheet and space-time supersymmetry by simple-current extensions. The world-sheet supercurrent for each $N=2$ minimal model is labelled by $T_{F,i}=(0,0,2)$ and the spectral-flow operator is $S_{F,i}=(0,1,1)$. These are used in the world-sheet and space-time chiral algebra extensions (\[WS\]) and (\[SS\]). These chiral algebra extensions are mandatory only in the fermionic sector. However, modular invariance does not allow an extension in just one chiral sector. The most common way of dealing with this is to use exactly the same CFT in the left-moving sector, with exactly the same extensions. Of course any $N=2$ CFT is a special example of an $N=0$ CFT. This construction leads to $(2,2)$ theories, with spectra analogous to Calabi-Yau compactifications, characterized by Hodge number pairs and with a certain number of families in the $(27)$ of $E_6$. On the other hand, modular invariance is blind to most features of the CFT spectrum. It only sees the modular group representations. This makes it possible to use in the left, bosonic, sector a different set of extension currents than on the right. In particular one can replace the image of the space-time current by something else, thus breaking $E_6$ to $SO(10)$. Furthermore one can break world-sheet supersymmetry in the bosonic sector. One can even go a step further and break $SO(10)$ and $E_8$ to any subgroup, as long as this breaking can be restored by means of simple currents. Those currents are then mandatory in the fermionic sector (since otherwise the bosonic string map cannot be used), but can be replaced by isomorphic alternatives in the left sector. In general, we will call this class $(0,2)$ models. All the aforementioned possibilities will be considered in this paper, except $E_8$ breaking. The $SO(10)$ breaking we consider is to $SU(3)\times SU(2) \times U(1)_{30} \times U(1)_{20}$, where the first three factors are the standard model gauge groups with the standard $SU(5)$-GUT normalization for the $U(1)$. The fourth factor corresponds in certain cases to $B-L$. Permutation orbifold of $N=2$ minimal models {#Section: Permutation orbifold of N=2 minimal models} ============================================ In [@Maio:2010eu] the permutation orbifold of $N=2$ minimal models was studied. Extensions and permutations were performed in all possible orders and a nice structure was seen to arise, together with exceptional off-diagonal simple currents appearing in the extended orbifolds. In this section we recall the procedure of how to build a supersymmetric permutation orbifolds starting from $N=2$ minimal models. We will restrict ourselves to $\mathbb{Z}_2$ permutations, because a formalism to build permutation orbifold CFTs for higher cyclic orders is not yet available. Consider the internal CFT $\mathscr{C}_9$ to be a tensor product of $r$ minimal models such that the total central charge is equal to $9$. We denote such a theory as[^3] $$(k_1,k_2,k_3\dots,k_r)\,,$$ each $k_i$ parametrizing the $i^{\rm th}$ minimal model. Suppose that two of the $k_i$’s are equal: then the two corresponding minimal models are also identical and one can apply the orbifold mechanism to interchange them. We will label with brackets the block corresponding to the orbifold CFT: e.g. if $k_2=k_3$, then the permutation orbifold will be denoted by $$(k_1,\langle k_2,k_3\rangle\dots,k_r)\,.$$ Multiple permutations are of course also possible. For convenience, we will follow the standard notation, used extensively in literature, of ordering the minimal models according to increasing level, namely $k_i\leq k_{i+1}$. Consequently, identical factors will always appear next to each other. The orbifolded theory has the same central charge of the original one, namely $\sum_i^r c_i=9$, and hence can be used to build four dimensional string theories. Note that by $\langle k,k\rangle$ we mean the [*supersymmetric*]{} permutation orbifold, which, as explained in [@Maio:2010eu], is obtained from the minimal model with level $k$ by first constructing the non-supersymmetric BHS orbifold (which we will denote as $[ k,k]$), extending this CFT by the anti-symmetric combination of the world-sheet supercurrent $(T_F,1)$, and resolving the fixed points occurring as a result of that extension. This fixed point resolution promotes some fields to simple currents. All these simple currents will be used to build MIPFs, using the general formalism presented in [@Kreuzer:1993tf]. Fixed point resolution enters the discussion at various points, and to prevent confusion we summarize here some relevant facts. In the following we consider chains of extensions of the chiral algebra of a [CFT]{}, and denote them as $({\rm CFT})_n$. Here $({\rm CFT})_0$ is the original [CFT]{}, $({\rm CFT})_1$ a first extension, $({\rm CFT})_2$ a second extension etc. In this process the chiral algebra is enlarged in each step. The number of primary fields can decrease because some are projected out and others are combined into new representations, but it can also increase due to fixed point resolution (apart from some special cases the decrease usually wins over the increase). We are not assuming that each extension is itself “indecomposable" ([*i.e.*]{} not the result of several smaller extensions), but in practice the case of most interest will be a chain of extensions of order 2. The following facts are important. - [Simple currents $J$ are characterized by the identity $S_{0J}=S_{00}$, where $S$ is the modular transformation matrix. For all other fields $i$, $S_{0i} > S_{00}$.]{} - [In an extension by a simple current of order $N$, the matrix elements $S_{0f}$ of fixed point fields are reduced by a factor of $N$. For this reason a fixed point field of $({\rm CFT})_n$ can be a simple current of $({\rm CFT})_{n+1}$. We will call these “exceptional simple currents".]{} - [Exceptional simple currents can be used to build new MIPFs in $({\rm CFT})_{n+1}$, but such MIPFs are not simple current MIPFs of $({\rm CFT})_n$. They are exceptional MIPFs.]{} - [If the fixed point resolution matrices of $({\rm CFT})_n$ are known, we can promote the exceptional simple currents of $({\rm CFT})_{n+1}$ to ordinary ones . This makes it possible to treat them on equal footing with all other simple currents of $({\rm CFT})_{n+1}$.]{} - [Obviously, this process can be iterated: exceptional simple currents of $({\rm CFT})_{n+1}$ can themselves have fixed points, which can become simple currents of $({\rm CFT})_{n+2}$.]{} - [If we know the fixed point resolution matrices of $({\rm CFT})_n$, we also know all the fixed point resolution matrices of the ordinary simple currents of $({\rm CFT})_{n+1}$, but if the exceptional simple currents have fixed points, there is currently no formalism available to determine their fixed point resolution matrices.]{} In [@Maio:2009kb; @Maio:2009cy; @Maio:2009tg] we have developed a formalism for all fixed point resolution matrices of the BHS permutation orbifolds. This plays the rôle of $({\rm CFT})_0$ in the foregoing. The supersymmetric permutation orbifold $\langle k, k\rangle$ is $({\rm CFT})_1$. It always has exceptional simple currents, but only for $k=2~\hbox{mod}~4$ they have fixed points. As explained above, we cannot resolve these fixed points, but in heterotic spectrum computations this is not necessary. This would be necessary if we want to go beyond spectrum computations to determine couplings. In spectrum computations, fixed point fields $f$ appear in the partition function as character combinations of the form $$N \bar\chi_f(\bar\tau)\chi_f(\tau), \ \ \ \ \ N>1 ,$$ which is resolved into a certain number of distinct representations $(f,x)$ that contribute to the partition function as $$\sum_x \bar\chi_{f,x}(\bar\tau)\chi_{f,x}(\tau), \ \ \ \ \chi_{f,x}(\tau) = m_x \chi_{f}(\tau)\ , \ \ \ \ \ \sum_x (m_x)^2=N \ .$$ Note that for $N\geq4$ the last condition has several solutions, and to find out which one is the right one the twist on the stabilizer of the fixed point must be determined [@Fuchs:1996dd]. However, here we merely want to add up the values of $N$ for a left-right combination of interest, and the individual values of $m_x$ do not matter. A few fields of the supersymmetric orbifold will be relevant in the following, all of untwisted type. They are: - The symmetric representation of the spectral flow operator $(S_F,0)$, with $S_F=(0,1,1)$. It will be relevant to make the whole theory supersymmetric; - The world-sheet supercurrent of the supersymmetric orbifold, that we denote by $\langle 0,T_F \rangle$[^4]; - The anti-symmetric representation of the identity, denoted by $(0,1)$. We will call it the “un-orbifold current" since the extension by this current undoes the orbifold, giving back the original tensor product. The un-orbifold current exists in the BHS orbifold $[k,k]$ as well as in the supersymmetric orbifold $\langle k, k\rangle$. Denoting extension currents by means of a subscript, we have the following CFT relations $$\begin{aligned} (k,k) &= [ k, k]_{\rm unorb}\\ (k,k)_{(T_F,T_F)} &= \langle k, k\rangle_{\rm unorb}\end{aligned}$$ In general, the full set of simple current MIPFs obtained from the permutation orbifold CFT $(k_1,\langle k_2,k_3\rangle\dots,k_r)$ will have a partial overlap with those of straight tensor product $(k_1, k_2,k_3,\ldots,k_r)$. Since the set of simple currents of $(k_1,\langle k_2,k_3\rangle\dots,k_r)$ includes the un-orbifold current one might expect that the latter set is entirely included in the former. However, this is not quire correct, since the supersymmetric permutation orbifold has fewer simple currents than the tensor product from which it originates, as explained above. In the extension chain, $\langle k, k\rangle_{\rm unorb}$ is $({\rm CFT})_2$. In both steps in the chain $$\begin{aligned} \hfill ({\rm CFT})_0 &= & [k,k] \\ &\downarrow\\ ({\rm CFT})_1 &=& \langle k, k\rangle\\ & \downarrow\\ ({\rm CFT})_2& =& \langle k, k\rangle_{\rm unorb} = (k,k)_{(T_F,T_F)}\end{aligned}$$ exceptional simple currents appear. Those of the first step are promoted to ordinary simple currents using fixed point resolution in the BHS orbifold. We then work directly with $\langle k, k\rangle$ as a building block, but by doing so we cannot use the exceptional simple currents emerging in the second step. In this case the exceptional simple currents could be used by working with $(k,k)_{(T_F,T_F)}$ directly, but then we are back in the unpermuted theory. So the point is not that these MIPFs are unreachable, just that they cannot be reached using the simple currents of $\langle k, k\rangle $. Obviously, if we were to use a different exceptional simple current in the second extension, such that $({\rm CFT})_2$ is a new, not previously known CFT with exceptional simple currents, some of its MIPFs cannot be reached using simple current methods neither from $({\rm CFT})_1$ nor from $({\rm CFT})_2$. In all cases, one can try to derive such MIPFs explicitly as exceptional invariants, and they can then be taken into account in heterotic spectrum computations, but this requires tedious and strongly case-dependent calculations. But in this paper we only consider simple current invariants, without any claim regarding completeness of the set of MIPFs we obtain. The phenomenon of exceptional simple currents is nothing new, and occurs for example in the D-invariant of $A_{1,4}$ (which is isomorphic to $A_{2,3}$), or the extension of the tensor products of two Ising models extended by the product of the fermions (turning it into a free boson). The simplest explicit example occurs for $k=1$. In this case the discussion can be made a bit more explicit, since the permutation orbifold is itself a minimal model, namely the one with $k=4$: $$\begin{aligned} \langle 1,1\rangle &=& (4)\\ (1,1)_{(T_F,T_F)} &=&(4)_{\rm unorb} = (4)_D\end{aligned}$$ The minimal $k=1$ model has 12 simple currents, and hence the tensor product $(1,1)$ has 144. To make the tensor product world-sheet supersymmetric we have to extend it by $(T_F,T_F)$, reducing the number of simple currents by a factor of four to 36. The $k=4$ minimal model has 24 simple currents. If we extend the $k=4$ minimal model by the un-orbifold current (which can be identified as such in the $\langle 1,1\rangle$ interpretation), these 24 original simple currents are reduced to 12. Since the resulting CFT is isomorphic to $(1,1)_{(T_F,T_F)}$ there must be 24 additional simple currents. Indeed there are, but they are exceptional. They are related to the aforementioned exceptional currents in the D-invariant of $A_{1,4}$. This is also the only example of exceptional simple currents in $N=2$ minimal models, and clearly in this case no MIPFs are missed, since we can explicitly consider $(1,1)$ as well as $(4)_D$. There might exist additional examples of exceptional simple currents in tensor products of $N=2$ minimal models. If the chiral algebra contains the un-orbifold current of a permutation orbifold, we obviously get nothing new. Therefore we demand that this current is not in the chiral algebra. In general, it would be possible to forbid it in either the left or the right chiral algebra. This is already sufficient to find new cases. We do this, for example, with the $SU(5)$ extension currents of the standard model, which are required in the right (fermionic string) chiral algebra, but not in the left one. However, it turns out that the un-orbifold current is local with respect to all other simple currents. In appendix \[Section: Simple current invariants\] we prove a small theorem about simple current invariants. Consider a simple current modular invariant partition function $$Z(\tau,\bar{\tau})=\sum_{k,\,l}\bar{\chi}_k(\bar{\tau})M_{kl}\chi_l(\tau)\,.$$ In the theorem it is shown that: if a current $J$ that is local with respect to all currents used to construct the modular invariant appears on the right hand side (holomorphic sector) of the algebra, then it will also appear on the left hand side (anti-holomorphic sector): $$M_{0J}\neq 0 \qquad\Leftrightarrow \qquad M_{J0}\neq 0\,.$$ Furthermore we show that the un-orbifold current is local with respect to all other currents. Therefore the existence of the un-orbifold current on one side implies its existence also on the other side. Hence it is sufficient to forbid its occurrence in either the left or the right sector. However, there are a few cases where it cannot be forbidden at all, because it is generated by combinations of world-sheet and space-time supersymmetry in the right (fermionic) sector, where such chiral algebra extensions are required. In general, a tensor product is extended by the currents $S_{\rm susy}$ and $W_i$, as explained in the previous section. If $k$ is even, the un-orbifold current does not appear on the orbit of the Ramond spinor current $S_F$, and hence can never be generated. For arbitrary $k$ we have in the supersymmetric permutation orbifold $$(S_F,0)^{2(k+2)}= \left\{ \begin{array}{lc} (0,0) & {\rm if}\,\,k\,\,{\rm even}\\ (0,1) & {\rm if}\,\,k\,\,{\rm odd} \end{array} \right. \,,$$ so that for $k$ odd one can obtain the un-orbifold current as a power of $S_F$. Note that instead of $2(k+2)$ one could use any odd multiple of $2(k+2)$. In the tensor product $S_F$ is combined with the spinor currents of all the other factors, which will be raised to the same power. Now note that in minimal models of level $k$ the following is true $$(S_F)^{2(k+2)}= \left\{ \begin{array}{lc} 0 & {\rm if}\,\,k\,\,{\rm even}\\ T_F & {\rm if}\,\,k\,\,{\rm odd} \end{array} \right. \,.$$ Furthermore, the value $2(k+2)$ is the first non-trivial power for which either the identity or the world-sheet supercurrent is reached. It follows that if the tensor product contains a factor with $k_i$ even, the complete susy current $(S_{F,1},\ldots,S_{F,r},S)$ must be raised to a power that is a multiple of four in order to reach either the identity or a world-sheet supercurrent. This is true for minimal model factors as well as supersymmetric permutation orbifolds $\langle k_i, k_i\rangle$. Consider then a tensor product $(k_1,\ldots,k_{m-1},\langle k_m,k_m\rangle,k_{m+1},\ldots,k_r)$. Take the susy current $$(S_{F,1},\ldots ,S_{F,m-1},(S_F,0),S_{F,m+1},\ldots S_{F,r};S)$$ to the power $2 M$, where $M$ is the smallest common multiple of $k_i+2$, for all $i$ (including $i=m$). If all $k_i$ are odd, this yields $$\label{Power} (T_{F,1},\ldots,T_{F,m-1},(0,1),T_{F,m-1},\ldots T_{F,r};V)$$ Since this is a power of an integer spin current, the susy current, it must have integer spin. Therefore the number of $T_{F,i}$ must be odd. Indeed, it is not hard to show that eqn. (\[MinSum\]) can only be satisfied with all $k_i$ odd if the total number of factors, $r$, is odd. It then follows that all entries $T_{F,i}$ as well as the representation $V$ of $SO(10)_1$ can be nullified by world-sheet supersymmetry. Hence it follows that the un-orbifold current of $\langle k_m,k_m\rangle$ is automatically in the chiral algebra. It also follows that if one of the $k_i$ is even the un-orbifold current is [*not*]{} in the chiral algebra generated by $S_{\rm susy}$ and $W_i$. The same reasoning can be applied to tensor products containing more than one permutation orbifold. The conclusion is that the un-orbifold currents of each factor separately are not generated by $S_{\rm susy}$ and $W_i$, but if all $k_i$ (in minimal models as well as the permutation orbifolds) are odd, the combination $(0,\ldots,(0,1),\ldots,(0,1),\ldots,0;0)$, with an un-orbifold component in each permutation orbifold, will automatically appear. Obviously, if there is more than one permutation orbifold factor this does not undo the permutation. The set of tensor combinations with only odd factors is rather limited, namely $$\begin{aligned} &(1,1,1,1,1,1,1,1,1)\\ &(3,3,3,3,3)\\ &(1,3,3,3,13)\\ &(1,1,7,7,7)\\ &(1,1,5,5,19)\\ &(1,1,3,13,13) \end{aligned}$$ We will not consider permutations of $k=1$, because $\langle1,1\rangle=4$, and hence nothing new can be found by allowing $\langle1,1\rangle$. Furthermore there is no need to consider any single permutations in the foregoing tensor products. However, we do expect the combinations $(3,\langle 3,3\rangle,\langle 3,3\rangle)$, $(4,\langle 7,7\rangle,7)$ and $(4,\langle 5,5\rangle,19)$ to yield something new. For technical reasons in this paper we consider only permutations of minimal models having level $k \leq 10$: computing time and memory use become just too large for large $k$. Nevertheless, the interval $k\in[2,10]$ still covers almost all the standard Gepner models where at least two factors can be permuted. Permutations of permutations ---------------------------- An additional thing that one can try to do (and which we can in principle do with our formalism, since we know all the relevant data that are needed) is to consider permutations of permutations. Permutations of permutations are possible only for a few Gepner models, because one would need to have a number of factors in the tensor product which is larger than four and with at least four identical minimal models. Out of the 168 possibilities, there are only a few combinations that have these properties. They are: $$\begin{aligned} &(6,6,6,6)&\nonumber\\ &(1,4,4,4,4)&\nonumber\\ &(3,3,3,3,3)&\nonumber\\ &(1,2,2,2,2,4)&\nonumber\\ &(2,2,2,2,2,2)&\end{aligned}$$ As before we restrict the $k>1$. Observe that the maximal level is $k=6$, so these cases are actually all the possibilities that one can consider and one can relax here our previous restriction to $k\leq10$. The approach one should take is the following. Consider a block of four identical minimal models. As before we can permute the factors pairwise and obtain a tensor product of two larger blocks, but again identical. Hence we can permute them again and end up with only one big block which replaces the four ones that we started with: $$\xymatrix{ (k,k,k,k) \ar[d] \\ (\langle k,k \rangle, \langle k,k\rangle) \ar[d]\\ \big\langle \langle k,k \rangle, \langle k,k\rangle\big\rangle }$$ Although straightforward, we will not perform this calculation in this paper. There are only very few cases to analyze, namely the five listed above, but, on the one hand, it is a pretty lengthy computation and on the other hand we do not expect drastically different spectra in comparison with normal permutations. Lifts {#Section: Lifts} ===== In [@GatoRivera:2009yt] the authors describe a new method for constructing heterotic Gepner-like four-dimensional string theories out of $N=2$ minimal models. The method consists of replacing one $N=2$ minimal model together with the $E_8$ factor by a non-supersymmetric CFT with identical modular properties. Generically this method produces a spectrum with fewer massless states. Surprisingly, it is possible to get chiral spectra and gauge groups such as $SO(10)$, $SU(5)$ and other subgroups including the Standard Model. However, the most interesting feature is probably the abundant appearance of three-family models, which are very rare in standard Gepner models [@GatoRivera:2010gv]. Let us review how it is done in more detail, at least in the simplest case. Start from the coset representation of the minimal model: $$\frac{SU(2)_k\times U(1)_4}{U(1)_{2(k+2)}}\,,$$ subject to field identification by the simple current $(J,2,k+2)$. Here $J$ is the simple current of the $SU(2)_k$ factor and the $U_N$ fields are labelled by their charges as $0,\dots,N-1$. The product of the $N=2$ minimal model and the $E_8$ factor is then $$\left(\frac{SU(2)_k\times U(1)_4}{U(1)_{2(k+2)}}\right)_{(J,2,k+2)}\times E_8\,,$$ where the brackets denote this identification. The next step is to remove the identification and mod out $E_8$ by $U(1)_{2(k+2)}$: the new CFT is then $$SU(2)_k\times U(1)_4\times \frac{E_8}{U(1)_{2(k+2)}}\,.$$ Finally we restore the identification by a standard order-2 current extension of the resulting CFT. This procedure works provided we can embed the $U(1)_{2(k+2)}$ factor into $E_8$. Some examples of how to embed $U(1)_{2(k+2)}$ into $E_8$ are given in [@GatoRivera:2009yt]. Finally, one can check explicitly that the modular $S$ and $T$ matrices are the same as for the $N=2$ minimal model times $E_8$, as they must be by construction. The resulting CFT is $SU(2)_k\times U(1)_4\times X_7$, where $X_7$ is the reminder of $E_8$ after dividing out $U(1)_{2(k+2)}$. $X_7$ has central charge $7$ and modular matrices $S$ and $T$ given by the complex conjugates of those of $U(1)_{2(k+2)}$ (since the ones of $E_8$ are trivial). Generically, this procedure raises the weights of the primaries in the new CFT, hence the name “weight lifting”. As it appears from above, the lifting of Gepner models is achieved by only a slight modification of standard Gepner models. All one has to do is to shift the weights of some fields in the left-moving CFT by a certain integer, and replace the ground state dimensions by another, usually larger, value. In [@GatoRivera:2009yt] a list of possible lifts is given for $N=2$ minimal models at level $k$. Not for any level there exists a lift and sometimes for fixed $k$ there are more lifts. When applied to standard Gepner models, a lot of new “lifted” Gepner models are generated. Notationally, if a Gepner model is denoted by $(k_1,\dots,k_i,\dots, k_r)$, the corresponding lifted model will be denoted by $(k_1,\dots,\hat{k}_i,\dots, k_r)$, where the lift is done on the $i^{\rm th}$ $N=2$ factor. If for a given $k$ there exists more than one lift, we use a tilde to denote it. In [@GatoRivera:2010fi] a different class of lifts was considered, the so called B-L lifts. In this case one replaces the $U(1)_{20}$ (with 20 primaries), that is the remainder of $SU(3)\times SU(2)\times U(1)$ embedded in $SO(10)$. In the Standard Model the abelian factor is the $U(1)_{Y}$ hypercharge (denoted also as $U(1)_{30}$, with 30 primaries). The $U(1)_{20}$ that we replace here corresponds to $B-L$, hence the name “B-L lifting”. It is not possible to simply replace the $U(1)_{20}$ by an isomorphic CFT with 20 primaries, central charge $c=1$ and same modular $S$ and $T$ matrices, since all the $c=1$ CFTs are classified. Again, what one can do is to add the $E_8$ factor and replace the $E_8\times U(1)_{20}$ block, which has central charge $c=9$. As it turns out, there are only two possible B-L lifts, that we denote by $A$ and $B$. In terms of compactifications from ten dimensions these possibilities can be distinguished as follows. If one compactifies the $E_8\times E_8$ heterotic string one gets $SO(10)\times E_8$ in four dimensions. The standard model can be embedded in $SO(10)$ (trivial lift, [*i.e.*]{} standard, unlifted $B-L$) or $E_8$ (lift A). If one compactifies the $SO(32)/\mathbb{Z}_2$ heterotic string, one gets $SO(26)$, in which the standard model can then be embedded via an $SO(10)$ subgroup; this yields lift B. As explained in [@GatoRivera:2010fi] both lift A and lift B yield, perhaps counter-intuitively, chiral spectra. In the unlifted case, the number of families is typically a multiple of 6, and sometimes 2; for lift A, the family number quantization unit was found to be usually 1, whereas for lift B it was usually 2. In this paper we will apply all these kinds of lifts to permuted Gepner models. This means that we make, when possible, all sorts of known lifts (namely, standard weight lifting and B-L lifting) for the $N=2$ factors that do *not* belong to the sub-block(s) of the permutation orbifold. Note that permutations and lifting act independently: a given minimal model factor is either unchanged, or lifted, or interchanged with another, identical factor. It may well be possible to construct lifted CFTs for the permutation orbifolds themselves, but no examples are known, and they are in any case not obtainable by the methods of [@GatoRivera:2009yt], because there only a single minimal model factor is lifted. There is one exception to this: there is one known simultaneous lift of two minimal model factors with $k=1$. Conceivably one could apply a permutation to those two identical factors in combination with this lift. We have however not investigated this possibility. Comparison with known results {#Section: Comparison} ============================= To compare our results with previous ones on permutation orbifolds [@Klemm:1990df; @Fuchs:1991vu], it is important to understand the differences in these approaches. These authors first construct the basic Gepner model with all world-sheet and space-time supersymmetry projections already in place in the left- as well as the right-moving sector. They start from either the diagonal (A-type) invariants of all the minimal models, or the D and E-type (exceptional) invariants. They then apply a cyclic permutation to the minimal model factors that are identical. They allow for additional phase symmetries occurring in combination with the permutations. This combined operation is applied to the full partition function. By contrast, we first build an $N=2$ permutation orbifold, then tensor it with other building blocks (either minimal models or other $N=2$ permutation orbifolds), then impose world-sheet and space-time supersymmetry, but only on the fermionic sector, and consider general simple current modular invariants. So the differences can be summarized as follows - [In [@Klemm:1990df; @Fuchs:1991vu] general cyclic $\mathbb{Z}_L$ permutations are considered, while our results are limited to $L=2$.]{} - [In [@Klemm:1990df; @Fuchs:1991vu] extra phases are modded out in combination with the permutations.]{} - [In [@Klemm:1990df; @Fuchs:1991vu] permutations of D and E-invariants are considered.]{} - [We only consider permutations of factors with $2 \leq k \leq 10$.]{} - [We consider general simple current invariants.]{} - [We consider not only $(2,2)$ but also $(0,2)$ invariants and breaking of $SO(10)$.]{} In order to make a comparison we will ignore the last point and focus on $(2,2)$ models. Since simple current invariants include D-invariants as special cases, and because they involve monodromy phases of currents with respect to fields, one might expect that at least part of the limitations in the second and third point are overcome. Exceptional invariants can be taken into account in our method by multiplying the simple current modular matrix with an explicit exceptional modular matrix. Indeed, in standard Gepner models we [*have*]{} taken them into account, and analysed the class of $(1,16^*,16^*,16^*)$ three-family models [@Gepner:1987hi]. In the present case one could easily use exceptional invariants in those factors that are not permuted. To use permutations of exceptional invariants we would first have to construct the exceptional MIPF explicitly in the permutation orbifold CFT, which can be done in principle with a tedious computation. The first point is, however, much harder to overcome, because it would involve extending the BHS construction to higher cyclic orders. Now let us see how the comparison works out in practice. In [@Fuchs:1991vu] a table is presented with all models where cyclic permutations, phase symmetries and cyclic permutations together with phase symmetries have been modded out. For each model the authors give the number of generations $\rm n_{27}$, anti-generations $\rm n_{\overline{27}}$ and singlets $\rm n_{1}$. The first two numbers are equal to Hodge numbers of Calabi-Yau manifolds, namely $h_{21}= \rm n_{27}$ and $h_{11}=\rm n_{\overline{27}}$. These quantities are first obtained by using modular invariance of the partition function of the cyclically-orbifolded Gepner models and are then compared with the same quantities derived by using topological arguments applied to the smooth Calabi-Yau manifold after that the singularities have been resolved. The number of families is specified by $\rm n_{gen}=n_{27}-n_{\overline{27}}$. The total number of singlets is strongly dependent on the multiplicities of the (descendants) states of the $N=2$ minimal models, which can be read off directly from the character expansions. The singlet number $\rm n_{1}$ turns out to be crucial for differentiating different models with equal $\rm n_{27}$ and $\rm n_{\overline{27}}$. Our comparison is based on these three numbers. In table (\[HodgeNumbers\]) we list the values we obtained for these three numbers in the cases we considered. Note that these are the numbers obtained without any simple current extensions or automorphisms. The cases marked with a $*$ are $K_3 \times T_2$ type compactifications with an $E_7$ spectrum; the numbers that are indicated are the ones obtained after decomposing $E_7$ to $E_6$. In comparing the A-type invariants without phase symmetries, we get agreement, but in a somewhat unexpected way. In [@Fuchs:1991vu] one-permutation models are not considered, because the authors argue that they always produce the same spectra as unpermuted Gepner models. However, in this paper we do manage to build one-permutation models as explained in section \[Section: Permutation orbifold of N=2 minimal models\]. The only Gepner combinations for which the one-permutation models yield nothing new are the purely-odd combinations. Furthermore, the one-permutation orbifolds do indeed yield new results. For example, for the combinations $(2,2,2,2,\langle 2,2\rangle)$ the three numbers are $(90,0,284)$ as opposed to $(90,0,285)$ for the unpermuted case; for $(6,6,\langle6,6\rangle)$ we find $(106,2,364)$ as opposed to $(149,1,503)$; for $(\langle 3,3\rangle,10,58)$ we get $(75,27,392)$ as opposed to $(85,25,425)$. These three example illustrate three distinct situations. In the first example, the only difference with the unpermuted case is that the number of singlets is reduced by one. In the second example, the Hodge pair $(106,2)$ does occur for a non-trivial simple current invariant of the tensor product $(6,6,6,6)$, namely $(6_A,6_A,6_A,6_D)$, but with 365 singlets instead of 364. In the last example the Hodge pair $(75,27)$ does not occur for any simple current MIPF of $(3,3,10,58)$ (the only other combination that occurs is $(53,41,401)$ plus the mirrors, so that even the Euler number of the permutation orbifold is new).[^5] In order to make a non-trivial comparison between our spectra and those of [@Fuchs:1991vu] we have to look at Gepner models with [*two*]{} permutations. It turns out that our spectra (specified by $\rm n_{27}$, $\rm n_{\overline{27}}$ and $\rm n_{1}$) do agree with those of [@Fuchs:1991vu]. However, to get the full match, we always have to extend the model by one current. This current is (see section \[Section: Permutation orbifold of N=2 minimal models\]) the double un-orbifold current, which has the un-orbifold current in each of the two factors corresponding to the permutation orbifold and the identity current in the remaining factors. Also in this case we already get new spectra even if we do [*not*]{} extend by this current. Consider for example $(\langle 6,6\rangle,\langle 6,6\rangle)$. As mentioned above, the $(6,6,6,6)$ gives $(149,1,503)$; the completely unextended spectrum we get for the $(\langle 6,6\rangle,\langle 6,6\rangle)$ yields $(77,1,269)$; if we extend the two permutation orbifold CFTs by the current combination $((0,1),(0,1))$ (where $(0,1)$ is the un-orbifold current) we find $(83,3,301)$, which is precisely the result quoted in [@Fuchs:1991vu] for the permutation orbifold. It is noteworthy that [@Fuchs:1991vu] lists a triplet $(77,1,271)$ for the combination $(6_D,6_D,6_A,6_A)$, which from our perspective is a simple current invariant of $(6,6,6,6)$. Again we see two spectra with a minor difference only in the number of singlets, which we will comment on below. In one case we could not make a comparison, because in [@Fuchs:1991vu] no result is listed for $(2,2,\langle 2,2\rangle,\langle 2,2\rangle)$ without extra phases. In all other cases our results agree with [@Fuchs:1991vu]. The need for extending by a combination of un-orbifold currents suggests that such currents are automatically generated or implicitly present in the formalism used in [@Fuchs:1991vu], for reasons we do not fully understand, but which are presumably related to an interchange in the order of two operations: permutation and simple current extension. This is also consistent with the fact that these authors find no new results for single permutations: if an un-orbifold current is automatically present in that case, one inevitably returns to the unpermuted case. Note that for $(3,\langle3,3\rangle,\langle3,3\rangle)$ we have seen before that the separate un-orbifold current of each permutation orbifold is automatically present in the chiral algebra, and hence so is the combination of the two. Therefore in this case we do not have to extend by $(0,(0,1),(0,1))$ to find agreement with [@Fuchs:1991vu] because the extension is already automatically present. Let us now compare the cases with extra phase symmetries. In almost all cases, using the simple-current formalism, we recover for a given suitably-extended model the same Hodge numbers and the same singlet number as in those spectra where both the phase symmetry and the permutation symmetry have been modded out. In a sense, these phase symmetries correspond to simple current extensions or automorphisms. The only two exceptions, out of the many successful instances, both coming from the $2^6$ Gepner model (nr. 21 of Table II in [@Fuchs:1991vu]) with two permutations and phase symmetries, are - (21)(43)56, 111100 ($\rm n_{27}=21$, $\rm n_{\overline{27}}=21$, $\rm n_{1}=180$, $\chi=0$), - (21)3(54)6, 333111 ($\rm n_{27}=44$, $\rm n_{\overline{27}}=8$, $\rm n_{1}=199$, $\chi=-72$), where the first entry is the permutation orbifold and the second one is the phase symmetry. We were not able to find these two cases using our procedure. There are a few other cases that we do not have, but for reasons that are easy to understand. Consider model nr. 168 in the same table. It correspond to the $6^4$ Gepner model. The double permutation that we reproduce is the one labelled as - $6_A 6_A 6_A 6_A$: (21)(43) ($\rm n_{27}=83$, $\rm n_{\overline{27}}=3$, $\rm n_{1}=301$, $\chi=-160$). The other two, with D invariants - $6_A 6_A 6_D 6_D$: (21)(43) ($\rm n_{27}=45$, $\rm n_{\overline{27}}=1$, $\rm n_{1}=181$, $\chi=-88$), - $6_D 6_D 6_D 6_D$: (21)(43) ($\rm n_{27}=35$, $\rm n_{\overline{27}}=3$, $\rm n_{1}=154$, $\chi=-64$), are not present. However these are not comparable with our $6^4$ since they come out of a different construction. In fact, the D invariant is obtained as a simple current automorphism of the $k=6$ Gepner models by the $SU(2)_k$ current $(k,0,0)$ (with $k=6$). This current has spin $h=\frac{k}{4}=\frac32$. In [@Fuchs:1991vu] the authors consider the permutation of two such $k=6$ models, each with such a simple current automorphism. This is different from what happens in this paper. Here, we immediately replace the block by its permutation orbifold; moreover, when we extend it by the current $(T_F,1)$ to build the supersymmetric permutation orbifold, the off-diagonal field $\langle (0,0,0)(6,0,0)\rangle$ with spin $h=\frac{3}{2}$ (the obvious candidate for creating the automorphism invariant) is not a simple current. We expect that the permutation orbifold of two $6_D$ models is present as an exceptional invariant of $\langle6,6\rangle$. The spectra mentioned in the last two paragraphs, that were present in [@Fuchs:1991vu] but absent in our results, might also be understood as follows. As explained in \[Section: Permutation orbifold of N=2 minimal models\], one may consider simple current extension chains of the form $${\rm (CFT)}_0 \rightarrow {\rm (CFT)}_1 \rightarrow {\rm (CFT)}_2 \rightarrow \ldots$$ In this chain, the supersymmetric permutation orbifold is ${\rm (CFT)}_1$. We can use all its simple currents to build MIPFs, and in particular we find all simple current extensions ${\rm (CFT)}_2$. However there are situations where ${\rm (CFT)}_2$ itself has new simple currents that are exceptional, and whose orbits cannot be fully resolved because we do not have the complete fixed point resolution formalism for ${\rm (CFT)}_1$ available. Therefore MIPFs generated by such second order exceptional simple currents cannot be obtained. At best, one could try to get them by explicit computation as exceptional MIPFs of ${\rm (CFT)}_1$. The problem of unresolvable fixed points occurs precisely for supersymmetric permutation orbifolds when $k=2\mod 4$, and therefore might be relevant precisely in these examples. =14truecm .7truecm [|c||c|c|c|]{} \ & & &\ [[** **]{}]{}\ & & &\ \ \[HodgeNumbers\] $(1,1,1,1,1,1,\langle 2,2\rangle)$ & $23^*$& $23^*$& 177\ $(1,1,1,1,1,\langle 4,4\rangle)$ & 84 & 0 & 249\ $(1,1,1,1,\langle 10,10\rangle)$ & 57 & 9 & 248\ $(1,1,1,1,\langle 2,2\rangle,4)$ & 35 & 11 & 229\ $(1,1,1,\langle 2,2\rangle,2,2)$ & $23^*$& $23^*$& 175\ $(1,1,1,2,\langle 6,6\rangle)$ & $23^*$& $23^*$& 173\ $(1,1,1,\langle 4,4\rangle,4)$ & 73 & 1 & 242\ $(1,1,1,\langle 3,3\rangle,8)$ & $23^*$& $23^*$& 173\ $(1,1,1,\langle 2,2\rangle,\langle 2,2\rangle)$ & $23^*$& $23^*$& 173\ $(1,1,2,2,\langle 4,4\rangle)$ & 35 & 11 & 211\ $(1,1,\langle 2,2\rangle,2,10)$ & 46 & 10 & 234\ $(1,1,4,\langle 10,10\rangle)$ & 75 & 3 & 279\ $(1,1,\langle 6,6\rangle,10)$ & 37 & 13 & 211\ $(1,1,\langle 2,2\rangle,4,4)$ & 51 & 3 & 250\ $(1,1,\langle 2,2\rangle,\langle 4,4\rangle)$ & 35 & 11 & 209\ $(1,2,2,\langle 10,10\rangle)$ & 61 & 1 & 251\ $(1,\langle 2,2\rangle,2,2,4)$ & 61 & 1 & 260\ $(1,2,4,\langle 6,6\rangle)$ & 51 & 3 & 235\ $(1,2,\langle 4,4\rangle,10)$ & 62 & 2 & 241\ $(1,2,\langle 3,3\rangle,58)$ & 41 & 17 & 273\ $(1,\langle 4,4\rangle,4,4)$ & 84 & 0 & 279\ $(1,\langle 2,2\rangle,10,10)$ & 89 & 5 & 343\ $(1,\langle 3,3\rangle,4,8)$ & 41 & 5 & 219\ $(1,\langle 2,2\rangle,5,40)$ & 35 & 35 & 329\ $(1,\langle 2,2\rangle,6,22)$ & 68 & 8 & 297\ $(1,\langle 2,2\rangle,7,16)$ & 43 & 19 & 289\ $(1,\langle 2,2\rangle,8,13)$ & 27 & 27 & 249\ $(1,\langle 2,2\rangle,\langle 2,2\rangle,4)$ & 61 & 1 & 259\ $(\langle 2,2\rangle,2,2,2,2)$ & 90 & 0 & 284\ $(2,2,2,\langle 6,6\rangle)$ & 73 & 1 & 251\ $(2,2,\langle 4,4\rangle,4)$ & 51 & 3 & 242\ $(2,2,\langle 3,3\rangle,8)$ & 41 & 5 & 218\ $(2,2,\langle 2,2\rangle,\langle 2,2\rangle)$ & 90 & 0 & 283\ $(2,\langle 10,10\rangle,10)$ & 105 & 3 & 380\ $(2,\langle 8,8\rangle,18)$ & 79 & 7 & 322\ $(\langle 2,2\rangle,2,3,18)$ & 65 & 5 & 279\ $(2,\langle 7,7\rangle,34)$ & 63 & 15 & 312\ $(\langle 2,2\rangle,2,4,10)$ & 69 & 3 & 265\ $(\langle 2,2\rangle,2,6,6)$ & 86 & 2 & 297\ $(2,\langle 2,2\rangle,\langle 6,6\rangle)$ & 73 & 1 & 250\ $(3,\langle 6,6\rangle,18)$ & 51 & 11 & 254\ $(3,\langle 5,5\rangle,68)$ & 53 & 29 & 328\ $(3,\langle 8,8\rangle,8)$ & 99 & 3 & 346\ $(3,\langle 3,3\rangle,\langle 3,3\rangle)$ & 59 & 3 & 228\ $(4,4,\langle 10,10\rangle)$ & 94 & 4 & 334\ $(4,\langle 6,6\rangle,10)$ & 55 & 7 & 238\ $(4,\langle 5,5\rangle,19)$ & 41 & 17 & 238\ $(4,\langle 7,7\rangle,7)$ & 66 & 6 & 270\ $(\langle 5,5\rangle,5,12)$ & 83 & 5 & 308\ $(\langle 6,6\rangle,6,6)$ & 106 & 2 & 364\ $(\langle 4,4\rangle,10,10)$ & 101 & 5 & 370\ $(\langle 3,3\rangle,10,58)$ & 75 & 27 & 392\ $(\langle 3,3\rangle,12,33)$ & 47 & 31 & 306\ $(\langle 3,3\rangle,13,28)$ & 97 & 13 & 404\ $(\langle 3,3\rangle,18,18)$ & 125 & 9 & 490\ $(\langle 2,2\rangle,3,3,8)$ & 39 & 15 & 249\ $(\langle 2,2\rangle,4,4,4)$ & 60 & 6 & 285\ $(\langle 4,4\rangle,5,40)$ & 65 & 17 & 334\ $(\langle 4,4\rangle,6,22)$ & 70 & 10 & 304\ $(\langle 4,4\rangle,7,16)$ & 79 & 7 & 308\ $(\langle 4,4\rangle,8,13)$ & 48 & 12 & 242\ $(\langle 3,3\rangle,9,108)$ & 69 & 49 & 466\ $(\langle 6,6\rangle,\langle 6,6\rangle)$ & 77 & 1 & 269\ $(\langle 2,2\rangle,\langle 4,4\rangle,4)$ & 51 & 3 & 240\ $(\langle 2,2\rangle,\langle 3,3\rangle,8)$ & 41 & 5 & 216\ $(\langle 2,2\rangle,\langle 2,2\rangle,\langle 2,2\rangle)$ & 90 & 0 & 282\ $(1,\langle 2,2\rangle,\langle 10,10\rangle)$ & 61 & 1 & 250\ $(\langle 4,4\rangle,\langle 10,10\rangle)$ & 75 & 3 & 273\ $(1,\langle 4,4\rangle,\langle 4,4\rangle)$ & 73 & 1 & 234\ As already mentioned, the list of Hodge numbers and singlets in table (\[HodgeNumbers\]) is obtained without any simple current extensions other than those required to get a $(2,2)$ model. The complete list obtained with arbitrary simple currents will be posted on the website of one of us [@hodge]. Although the results in table (\[HodgeNumbers\]) are for $(2,2)$ models, the focus of the present paper was on $(0,2)$ models. We can compare the results with those of [@GatoRivera:2010gv] and ask what permutation orbifolds add. Consider first the set of $(0,2)$ models closest to $(2,2)$ models, namely those with an $E_6$ gauge symmetry. They are characterized by the same three numbers ${\rm n}_{27}, {\rm n}_{\overline{27}}$ and ${\rm n}_1$, but since there is not necessarily a world-sheet supersymmetry in the bosonic sector they may not have a Calabi-Yau interpretation. For simplicity we will refer to these as “pseudo Hodge pairs" and “pseudo Hodge triplets". In the complete set of standard Gepner models without exceptional invariants we obtained a total of 1418[^6] different pseudo Hodge pairs and 9604 different pseudo Hodge triplets. For the genuine permutation orbifolds (without extensions by un-orbifold currents) these numbers are respectively 498 and 3830. Note that some permutation orbifolds with $k > 10$ were not considered. How many of the permutation orbifold numbers are new? If we combine the data for pseudo Hodge pairs and remove identical ones, we obtain a total of 1447 pseudo Hodge pairs, so that the total has increased by a mere 29. But if we look at pseudo Hodge triplets, the increase is much more substantial. This number increases from 9604 to 12145, an increase of 2541 or about $26\%$. We tentatively conclude that permutation orbifolds mainly give new points in existing moduli spaces. The following observation is further evidence in that direction. One remarkable feature of the permutation orbifold spectra is the occurrence of identical Hodge numbers and a number of singlets that is almost the same. For example, in the set of permutation orbifolds obtained from the $(2,2,2,2,2,2)$ tensor product we find spectra with (genuine) Hodge numbers $(90,0)$, and either 282, 283, 284 or 285 singlets. A closer look at the spectrum reveals what is going on here. We also compute the number of massless vector bosons in these spectra, and it turns out that this is respectively 2,3,4 and 5 (in addition to those of $E_6$) in these cases. This is consistent with the occurrence of a Higgs mechanism that has made one or more of the vector bosons heavy by absorbing the corresponding number of singlets. So apparently we are finding points in the same moduli space, but with a vev for certain moduli fields so that some of the $U(1)$’s are removed. This is expected to occur in Gepner models, but it is nice to see this happen entirely within RCFT. The same observation was made in [@Fuchs:1991vu]. The reduction of the number of $U(1)$’s by itself has a straightforward reason: each $N=2$ model has an intrinsic $U(1)$, and replacing two minimal models by a permutation reduces the number of $U(1)$’s by 1. Hence the $(2,2,2,2,2,2)$ model generically has five $U(1)$’s (six, minus one combination that becomes an $E_6$ Cartan-subalgebra generator), and $(\langle 2,2\rangle,\langle 2,2\rangle,\langle 2,2\rangle)$ generically has only two. However, the number of vector bosons can be larger than that because the simple current MIPFs add extra generators to the chiral algebra. Indeed, among the MIPFs of $(\langle 2,2\rangle,\langle 2,2\rangle,\langle 2,2\rangle)$ we do not only find $(90,0,282,2)$ (where the last entry is the number of $U(1)$’s), but also $(90,0,283,3)$ and $(90,0,284,4)$. Results {#Section: Results} ======= The CFT approach, based on simple currents extensions, turns out to be extremely powerful. Although we have considered in this paper only order-two permutations, the number of new modular invariant partition functions or, equivalently, the number of new spectra for each model is huge, in the order of a few thousands. Simple currents allow us to generate a huge number of four dimensional spectra. Here we discuss the more phenomenological aspects of our results. Conceptually this is very similar to work on unpermuted Gepner models presented in [@GatoRivera:2010gv; @GatoRivera:2010xn; @GatoRivera:2010fi], to which we refer for more detailed descriptions. In these papers several, mostly empirical, observations were made regarding the resulting spectra. The main question of interest here is if these observations continue to hold as we extend the scope of RCFTs considered. Gauge groups ------------ We allow as gauge groups $SO(10)$ and seven rank 5 subgroups, namely the Pati-Salam group $SU(4)\times SU(2)\times SU(2)$, the Georgi-Glashow GUT group $SU(5)\times U(1)$, two global realizations of left-right symmetric algebra $SU(3)\times SU(2)\times SU(2)\times U(1)$, and three global realizations of the standard model algebra $SU(3)\times SU(2)\times U(1)\times U(1)$. Counted as Lie-algebras there are just five of them, but the last two come in several varieties when we describe them as CFT chiral algebras. These are distinguished by the fractionally charged (here “charge" refers to unconfined electric charge) representations that are allowed. For the left-right algebra this can be either $\frac13$ or $\frac16$, (we call these “LR, Q=1/3" and “LR, Q=1/6" respectively) and for the standard model this can be $\frac12$, $\frac13$ or $\frac16$ (SM, Q=1/2, 1/3 or 1/6). In the string chiral algebra these different global realizations are distinguished by the presence of certain integer spin currents. If these currents have conformal weight one, they manifest themselves in the massless spectrum as extra gauge bosons. This happens in particular for the highly desirable global group corresponding to the standard model with only integer unconfined electric charge. In this class of heterotic strings this necessarily implies an extension of the standard model to (at least) $SU(5)$. Furthermore, if the standard model gauge group is extended to $SU(5)$, this group cannot be broken by a field-theoretic Higgs mechanism, because the required Higgs scalar, a $(24)$, cannot be massless in the heterotic string spectrum. A heterotic string spectrum contains either these massless vector bosons, or fractionally charged states that forbid the former because they are non-local with respect to them [@Schellekens:1989qb] (see also [@Wen:1985qj; @Athanasiu:1988uj]). These eight gauge groups are obtained as extensions of the affine Lie algebra $SU(3)_1\times SU(2)_1\times U_{30}$, with a $U(1)$ normalization that gives rise to $SU(5)$-GUT type unification. In general, there is an additional $U(1)$ factor that corresponds to a gauged $B-L$ symmetry in certain cases. In B-L lifted spectra this $U(1)$ is replaced by a non-abelian group. In addition, the gauge group consists out of a $U(1)$ factor for each superconformal building block, which sometimes is extended to a larger group, depending on the MIPF considered. There may also be extensions of the standard model gauge group outside $SO(10)$, such as $E_6$ or trinification, $SU(3)^3$. In standard Gepner models there is furthermore an unbroken $E_8$ factor, which in lifted Gepner models is replaced by certain combinations of abelian and non-abelian groups. In scanning spectra we focus only on the aforementioned eight (extended) standard model groups. To identify their contribution in the figures we use the same color-coding as in [@GatoRivera:2010xn], and the same names. These codes and names are shown in each figure. MIPF scanning ------------- Since it is essentially impossible to construct the complete set of distinct MIPFs, we use a random scan. This is done by choosing 10.000 randomly chosen simple current subgroups ${\cal H}$ (see appendix A) generated by at most three simple currents. Furthermore, if the number of distinct torsion matrices $X$ is larger than 100, we make 100 random choices. The entire set is guaranteed to be mirror symmetric, because for every given spectrum one can always construct a mirror by multiplying the MIPF with a simple current MIPF of $SO(10)_1$ that flips the chirality of all spinors. Note that this does not imply anything about mirror symmetry of an underlying geometrical interpretation. It is a trivial operation on the spectrum that can however be used to get some idea on the completeness of the scan. Fractional Charges ------------------ Fractional charges can appear either in the form of chiral particles, or as vector-like particles (where “vector-like" is defined with respect to the standard model gauge group) or only as massive particles, with masses of order the string scale. If a spectrum has chiral fractionally charged particles, we reject it after counting it. In nearly all remaining cases the spectrum contains massless vector-like fractional charges (unless there is GUT unification). We regard such spectra as acceptable at this stage. Since no evidence for fractionally charged particles exists in nature, with a limit of less than $10^{-20}$ in matter [@Perl:2009zz], clearly these vector-like particles will have to acquire a mass. Furthermore this will almost certainly have to be a huge (GUT scale or string scale) mass, since otherwise their abundance cannot be credibly expected to be below the experimental limit. This can in principle happen if the vector-like particles couple to moduli that get a vev. An analysis of existence of couplings is in principle doable in this class of models, although there may be some technical complications in those cases where no fixed point resolution procedure is available at present (namely the permutation orbifolds with $k=2\ {\rm mod}\ 4$). However, this analysis is beyond the scope of this paper, and we treat spectra with vector-like fractional charges as valid candidates, for the time being. Just as in previous work [@GatoRivera:2010gv; @GatoRivera:2010xn; @GatoRivera:2010fi; @FF2], there are extremely rare occurrences of spectra without any massless fractionally charged particles at all, but we only found examples with an even number of families. Examples with three families were found in [@FF2] by scanning part of the free-fermion landscape. In the context of orbifold models and Calabi-Yau compactifications, it is known that GUT breaking by modding out freely acting discrete symmetries leads to spectra without massless fractional charges ([@Witten:1985xc]; see [@Blaszczyk:2009in] for a recent implementation of this idea in the context of the “heterotic mini-landscape" [@Buchmuller:2005jr; @Lebedev:2006kn; @Lebedev:2008un]). While these models do fit the data on fractional charges, the question remains for which fundamental reasons such vacua are preferred over all others, especially if they are much rarer. In table (\[FreqTable\]) we display how often four mutually exclusive types of spectra occur in the total sample, before distinguishing MIPFs. The types are: spectra with chiral, fractionally charged exotics, chiral spectra with a GUT gauge group $SU(5)$ or $SO(10)$, non-chiral spectra (no exotics and no families), spectra with $N$ families and, massless $SU(5)$ vector bosons and vector-like fractionally charged exotics, and the same without massless fractionally charged exotics. For comparison, we include some results based on the data of [@GatoRivera:2010gv; @GatoRivera:2010xn; @GatoRivera:2010fi].[^7] All lines refer to Gepner models, except the one labelled “free fermions". The results on free fermions are based on a special class that can be analysed with simple current in a way analogous to Gepner models, as explained in [@GatoRivera:2010gv]. It does not represent the entire class of free fermionic models. For other work on this kind of construction, including three family models, see [@FF1; @FF2] and references therein. 1.truecm                  Type Chiral Exotics GUT Non-chiral $N > 0$ fam. No frac. ------------------------------------------------------- ---------------- -------- ------------ -------------- ---------- -- -- Standard$^*$ 37.4% 32.7% 20.5 % 9.3% 0 Standard, perm. 29.7% 33.4 % 27.9 % 8.9% 0 Free fermionic 1.5% 2.9% 94.4% 1.1% 0.072% Lifted 28.3% 18.7% 51.9% 1.1% 0.00051% Lifted, perm. 26.8% 8.9% 62.7 % 1.6% 0.00078% $({\hbox{B-L}})_{\hbox{\footnotesize Type-A}}^*$ 21.3% 28.0% 50.4 % 0.3% 0.00017% $({\hbox{B-L}})_{\hbox{\footnotesize Type-A}}$, perm. 22.8% 8.1 % 69.1 % 0.03% 0 $({\hbox{B-L}})_{\hbox{\footnotesize Type-B}}^*$ 38.5% 8.7% 52.1% 0.6% 0 $({\hbox{B-L}})_{\hbox{\footnotesize Type-B}}$, perm. 27.6% 7.3 % 65.0 % 0.1% 0 : Relative frequency of various types of spectra. An asterisk indicates that exceptional minimal model MIPFs are included.[]{data-label="FreqTable"} In table (\[MIPFTable\]) we specify the absolute number of distinct MIPFs (more precisely, distinct spectra, based on the criteria spelled out in [@GatoRivera:2010gv; @GatoRivera:2010xn; @GatoRivera:2010fi]) with non-chirally-exotic spectra. The column marked “Total" specifies the total number of distinct spectra without chiral exotics, the third column lists the number of distinct 3-family spectra and the last column the number of distinct $N$-family spectra, in both cases regardless of gauge group and without modding out mirror symmetry. 1.truecm                  Type Total 3-family $N$ family, $N>0$ ------------------------------------------------------- ----------- ---------- ------------------- -- Standard$^*$ 927.100 1.220 369.089 Standard, perm. 245.821 0 64.085 Free fermionic 504.312 0 19.655 Lifted 3.177.493 85.864 537.581 Lifted, perm. 601.452 4.702 54.926 $({\hbox{B-L}})_{\hbox{\footnotesize Type-A}}^*$ 445.978 24.203 155.425 $({\hbox{B-L}})_{\hbox{\footnotesize Type-A}}$, perm. 155.784 778 6.758 $({\hbox{B-L}})_{\hbox{\footnotesize Type-B}}^*$ 206949 0 55917 $({\hbox{B-L}})_{\hbox{\footnotesize Type-B}}$, perm. 156.309 0 6.861 : Total numbers of distinct spectra. []{data-label="MIPFTable"} Family number ------------- In this subsection we would like to say something about the distributions of the number of families emerging from the spectra of permuted Gepner models. The common features of all the different cases is that an even number of families is always more favourable than an odd one and these distributions decrease exponentially when the number of families increases. Figure \[famplot\_standard\] shows the distribution of the number of families for permutation orbifolds of standard Gepner Models. All family numbers are even, as is the case for unpermuted Gepner models (we did not include exceptional MIPFs, which provides the only way to get three families in standard Gepner models). The greatest common denominator $\Delta$ of the number of families for a given tensor combination displays a similar behavior as observed in [@GatoRivera:2010gv]. Two classes can be distinguished. Either $\Delta=6$ (or in a few cases a multiple of $6$), or $\Delta=2$ (sometimes 4), but there are no MIPFs with a number of families that is a multiple of three. In other words, the set of family numbers occurring in these two classes have no overlap whatsoever. It follows that in the second class there are no spectra with zero families. An interesting example is$(3,\langle6,6\rangle,18)$. It has no spectra with chiral exotics, all spectra are chiral and have 4, 8, 16, 20, 28, 32, 40 or 56 families, of types $SO(10)$, Pati-Salam, $SU(5)\times U(1)$ or $SM, Q=1/2$. If we compare this with the unpermuted Gepner model we find some striking differences. In that case the same group types occur, but now there are spectra with chiral exotics, and the family quantization is in units of 2, not 4. In [@GatoRivera:2010gv] an intriguing observation was made regarding the occurrence of these two classes. The second class was found to occur if all values $k_i$ of the factors in the tensor product are divisible by 3. This observation also holds for permutation orbifolds, if one uses the values of $k_i$ of the unpermuted theory. In figure \[famplot\_lift\] we show the family distribution for lifted Gepner models. As expected, this distribution looks a lot more favourable for three family models. The number three appears with more or less the same order of magnitude as two or four. However, there are some clear peaks at even family numbers, which were not visible in the analogous distribution for unpermuted Gepner models presented in [@GatoRivera:2010xn]. For this reason three families are still suppressed by a factor of 3 to 4 with respect to 2 or 4 families, which is however a lot less than the two to three orders of magnitude observed in orientifold models [@Dijkstra:2004cc]. B-L lifts give similar results to those presented in [@GatoRivera:2010fi]. Figure \[famplot\_liftA\] contains the distribution of the number of families for permutation orbifolds of B-L lifted (lift A) Gepner models. Figure \[famplot\_liftB\] contains the same, but for the lift B. Here, odd numbers are completely absent. Note that certain group types (namely those without a “$B-L$" type $U(1)$ factor) cannot occur in chiral spectra in these models, and that in the type that do occur the $U(1)$ is replaced by a non-abelian group. Conclusions =========== In this paper we have considered $\mathbb{Z}_2$ permutation orbifolds of heterotic Gepner models. This should be viewed as an application of our previous paper [@Maio:2010eu] where $\mathbb{Z}_2$ permutations were studied for $N=2$ minimal models, which are the building blocks in Gepner construction. Our main conclusion is that these new building blocks work as they should. They can be used on completely equal footing with all other available ones, which are the $N=2$ minimal models and some free-fermionic building blocks. We have checked the combination with minimal models and found full agreement with previous results on permutation orbifolds whenever they were available. The comparison did bring a few surprises, especially the fact that we were able to get new spectra for single permutations, where the old method of [@Fuchs:1991vu] gave nothing new. We were able to go far beyond the old approach by finding many more $(2,2)$ models, as well as new $(0,2)$ models with $SO(10)$ breaking. We combined all this with heterotic weight lifting and B-L lifting. The main conclusion is that in most respects all observations concerning family number and fractional charges made for minimal models continue to hold in this new class. Also in this case weight lifting greatly enhances the set of three family models in comparison to neighboring numbers. Although this appears to give some entirely new models (Hodge number pairs that were not seen before), we found additional evidence for the observation of [@Fuchs:1991vu] that many of these models look like additional rational points in existing moduli spaces. A next step in the exploration of these models is the study of couplings. An important problem to address is the possibility of lifting the vector-like states, especially the omnipresent fractionally charged ones, by giving vevs to moduli. In some cases, namely the permutation orbifolds with $k=2\mod4$, one may run into the problem of unresolved fixed points for certain couplings. This problem did not arise here because we only studied spectra. But there are plenty of cases where this restriction does not apply. The fact that all building blocks with $k\not=\!2\mod4$ are on equal footing with minimal models even with respect to fixed points should make it possible to develop an algorithm for determining couplings all at once for all cases. This is certainly a problem we intend to study in the near future. Another application of our permutation orbifold CFTs could be to study the case of orientifold models. However, there we will face the same problem with fixed point resolution at an earlier stage, since the fixed point resolution matrices enter in the formula for boundary coefficients. We hope to address this issue in the future. Acknowledgments {#acknowledgments .unnumbered} =============== This research is supported by the Dutch Foundation for Fundamental Research of Matter (FOM) as part of the program STQG (String Theory and Quantum Gravity, FP 57). This work has been partially supported by funding of the Spanish Ministerio de Ciencia e Innovación, Research Project FPA2008-02968, and by the Project CONSOLIDER-INGENIO 2010, Programme CPAN (CSD2007-00042). Simple current invariants {#Section: Simple current invariants} ========================= Consider a simple current $J$ or order $N$, i.e. $J^N=1$. Define the *monodromy parameter* $r$ as $$h_J=\frac{r(N-1)}{2N}\quad {\rm mod}\,\,\mathbb{Z}\,.$$ Also, define the *monodromy charge* $Q_J(\Phi)$ of $\Phi$ w.r.t. $J$ as $$Q_J(\Phi)=h_J+h_{\Phi}-h_{J\phi}\quad {\rm mod}\,\,\mathbb{Z}\,.$$ The monodromy charge takes values $t/N$, with $t\in\mathbb{Z}$. The current $J$ organizes fields into orbits $(\Phi,J\Phi,\dots,J^d\Phi)$, where $d$ (not necessarily equal to $N$) is a divisor of $N$. On each orbit the monodromy charge is $$Q_J(J^n\Phi)=\frac{t+rn}{N}\quad {\rm mod}\,\,\mathbb{Z}\,.$$ If a simple current $J$ exists in a (rational) CFT, and if it satisfies the condition that $N$ times its conformal weight is an integer,[^8] then it is known how to associate a modular invariant partition function to it. Suppose that the current $J$ has integer spin and order $N$. Then a MIPF is given by $$\label{MIPF} Z(\tau,\bar{\tau})=\sum_{k,\,l}\bar{\chi}_k(\bar{\tau})M_{kl}(J)\chi_l(\tau)\,.$$ One way of expressing $M_{kl}(J)$ is [@Kreuzer:1993tf; @Schellekens:1989wx]: $$\label{M_kl} M_{kl}(J)=\sum_{p=1}^N \delta(\Phi_k,J^p \Phi_l)\cdot\delta^1(\hat{Q}_J(\Phi_k)+\hat{Q}_J(\Phi_l))\,$$ where $\delta^1(x)=1$ for $x=0$ mod $\mathbb{Z}$ and $\hat{Q}$ is defined on $J$-orbits as $$\hat{Q}_J(J^n\Phi)=\frac{(t+rn)}{2N} \quad{\rm mod}\,\,\mathbb{Z}\,.$$ Morally speaking, $\hat{Q}$ is half the monodromy charge. Formula (\[M\_kl\]) defines a modular invariant partition functions, since it commutes with the $S$ and $T$ modular matrices, as shown in [@Schellekens:1989dq]. The set of all the simple currents forms an abelian group $\mathcal{G}$ under fusion multiplication. It is always a product of cyclic factors generated by a (conventionally chosen) complete subset of independent simple currents. The foregoing associates a modular invariant partition function with a single simple current. One can construct even more of them by multiplying the matrices $M$. The most general simple current MIPF associated with a given subgroup of $\mathcal{G}$ can be obtained as follows [@Kreuzer:1993tf; @GatoRivera:2010gv]. Choose a subgroup of $\mathcal{G}$ denoted $\mathcal{H}$, such that each element satisfies the effective center condition $N h_J \in \mathbb{Z}$. Its generators are simple currents $J_s$, $s=1,\dots,k$ for some $k$. They have relative monodromies $Q_{J_s}(J_t)=R_{st}$. Take any matrix $X$ that satisfies the equation $$X+X^T=R\,.$$ The matrix $X$ (called the [*torsion matrix*]{}) determines the multiplicities $M_{ij}$ according to $$\label{M=nr of sols} M_{ij}(\mathcal{H},X)={\rm nr.\,\,of\,\,solutions}\,\,K\,\,{\rm to\,\,the\,\,conditions:}$$ - $j=Ki$, $K\in\mathcal{H}$. - $Q_M(i)+X(M,K)=0$ mod 1 for all $M\in\mathcal{H}$. Here $X(K,J)$ is defined in terms of the generating current $J_s$ as $$X(K,J)\equiv \sum_{s,t}n_s m_t X_{st}\,,$$ with $K=(J_1)^{n_1}\dots(J_k)^{n_k}$ and $J=(J_1)^{m_1}\dots(J_k)^{m_k}$. A small theorem --------------- In this subsection we prove the following theorem. The following statements are true.\ i) If a simple current $J$ is local w.r.t. any other current $K$, i.e. $Q_K(J)=0$ (mod $\mathbb{Z}$), then $M_{JJ}(K)\neq 0$.\ ii) For a simple current $J$, which is local w.r.t. any other current $K$, $M_{J0}(K)=M_{0J}(K)$. In particular, if $M_{J0}(K)\neq 0$, then also $M_{0J}(K)\neq 0$. For the proof we use the statement (\[M=nr of sols\]).\ Let us start with [*i)*]{} and consider $M_{JJ}(K,X)$. The first condition in (\[M=nr of sols\]) has only one solution, namely $K=0$. The second condition reads $$Q_M(J)+X(M,0)=0 \qquad \forall M$$ and is always true, because the two terms vanish separately. This proves that $M_{JJ}(K)\neq 0$. Point [*ii)*]{} goes as follows. Consider $M_{0J^c}(K,X)$. There is again only one solution to the first condition, namely $K=J^c$. The second condition reads $$Q_M(0)+X(M,J^c)=0\,.$$ The first term vanishes by hypothesis, while the second is either zero (in which case $M_{0J^c}(K,X)\neq0$) or non-zero (in which case $M_{0J^c}(K,X)=0$).\ Similarly, look at $M_{J0}(K,X)$. There is again only one solution to the first condition, namely $K=J^c$. The second condition reads $$Q_M(0)+X(M,J^c)=0\,.$$ The first term vanishes by hypothesis, while the second is either zero or non-zero. In any case, the same condition holds for both $M_{0J^c}(K,X)$ and $M_{J0}(K,X)$. This implies that $M_{J0}(K,X)=M_{0J^c}(K,X)$ (note that these matrix elements can only be 0 or 1). By closure of the algebra, and because $J^c$ is always a power of $J$, we may replace $J^c$ by $J$ in this relation. Consider now the permutation orbifold. This theorem applies in particular to the un-orbifold current $J=(0,1)$ when coupled to any other current $K$, which is either a standard (diagonal) or an exceptional (off-diagonal) one. In fact, using the same procedure as we did in [@Maio:2009kb] to compute the simple current and fixed point structure of the permutation orbifold, one can show that $$\begin{aligned} N_{(J,\psi)\langle pq\rangle}^{\phantom{(J,\psi)\langle pq\rangle}\langle p'q'\rangle} &=& N_{Jp}^{\phantom{Jp}p'}N_{Jq}^{\phantom{Jq}q'}+N_{Jp}^{\phantom{Jp}q'}N_{Jq}^{\phantom{Jq}p'}\,,\nonumber\\ N_{(J,\psi)(i,\chi)}^{\phantom{(J,\psi)(i,\chi)}(i',\chi')} &=& \frac{1}{2} N_{Ji}^{\phantom{Ji}i'} (N_{Ji}^{\phantom{Ji}i'}+e^{i\pi(\psi+\chi-\chi')})\,.\nonumber\end{aligned}$$ Hence, for the current $(J,\psi)=(0,1)$, we have $$N_{(0,1)\langle pq\rangle}^{\phantom{(0,1)\langle pq\rangle}\langle p'q'\rangle} = \delta_p^{p'}\delta_q^{q'}+\delta_p^{q'}\delta_q^{p'}=\delta_p^{p'}\delta_q^{q'}\,,\nonumber$$ namely $\langle pq\rangle$ must be fixed by $(0,1)$ in order for this to be non-zero (recall that $p<q$ and $p'<q'$), and $$N_{(0,1)(i,\chi)}^{\phantom{(0,1)(i,\chi)}(i',\chi')} = \frac{1}{2} \delta_i^{i'}(\delta_i^{i'}-e^{i\pi(\chi-\chi')})\,\nonumber$$ which is non-zero only if $i=i'$ and $\chi\neq\chi'$ (recall that we can think of $\chi$ as defined mod $2$). Equivalently, in the fusion language: $$(0,1)\cdot \langle pq\rangle=\langle pq\rangle\quad,\qquad (0,1)\cdot (i,\chi)=(i,\chi+1)\,.$$ This implies that $(0,1)$ has zero monodromy charge w.r.t. any other current, since $$\begin{aligned} Q_{\langle pq\rangle}\big( (0,1) \big) &=& h_{\langle pq\rangle}+h_{(0,1)}-h_{\langle pq\rangle}=0\quad{\rm mod}\,\,\mathbb{Z}\,,\nonumber\\ Q_{(i,\chi)}\big( (0,1) \big) &=& h_{(i,\chi)}+h_{(0,1)}-h_{(i,\chi+1)}=0\quad{\rm mod}\,\,\mathbb{Z}\,.\nonumber\end{aligned}$$ Now, the un-orbifold current $(0,1)$ has order two, hence $J^c=J$ and $M_{J0}(K,X)=M_{0J}(K,X)$. This also implies that its existence on left-moving sector is guaranteed by its existence on the right-moving sector (and vice-versa). Summary of results ------------------ Here we present four tables summarizing the results on the number of families for the standard, heterotic weight lifted, B-L lifted (lift A) and B-L lifted (lift B) cases. These tables contain information about spectra in which the un-orbifold current is not allowed in the chiral algebra. This means that these are genuine permutation orbifold spectra. By inspection, we do indeed find that these spectra are usually different than those obtained in the unpermuted case. In the columns we specify respectively the tensor combination the greatest common divisor $\Delta$ of the number of families for all MIPFs of the tensor product and the maximal net number of families encountered. In the next column we indicate which of the seven $SO(10)$ subgroups occur, with the labelling - [0: SM, Q=1/6]{} - [ 1: SM, Q=1/3]{} - [ 2: SM, Q=1/2]{} - [ 3: LR, Q=1/6]{} - [4: $SU(5)\times U(1)$]{} - [ 5: LR, Q=1/3]{} - [ 6: Pati-Salam. ]{} Since $SO(10)$ can always occur there is no need to indicate it. In [@GatoRivera:2010gv] a simple criterion was derived to determine which subgroups can occur in each standard Gepner model. The allowed subgroups for permutation orbifolds of Gepner models are a subset of these. In some cases, such as $(\langle 5,5\rangle,5,12)$ some of the subgroups cannot be realized. In the column labelled “Exotics" we indicate if, for a given tensor product, spectra with chiral exotics occur. Note that in most cases the absence of such spectra is a consequence of the fact that only GUT gauge groups occur. In the next columns we list the number of distinct three family and in column 6 the number of distinct $N$-family $(N>0)$ spectra. In these tables only cases with $\Delta > 0$ are shown. If a permutation orbifold seems to be missing, than either it is a permutation for $k > 10$, or it is a purely odd tensor product for which all permutations are trivial, or it has only non-chiral spectra and hence $\Delta=0$ and there are no chiral exotics. In column 1 of the second table, $\langle \cal{A},\cal{A}\rangle$ denotes the permutation orbifold of CFT ${\cal A}$, a hat indicates the lifted CFT, and a tilde indicates the second lift of a CFT. It turns out that in the only permutation orbifold with two distinct lifts of the same factor, $(\hat 5,\langle5,5\rangle,12)$ and $(\tilde 5,\langle5,5\rangle,12)$, $\Delta=0$ in both cases, which is why a tilde never occurs in the tables. The last column indicates which percentage of the spectra has no mirror. Since mirror symmetry is exact in the full set, this gives an indication of how close our random scan is to a full enumeration. =14truecm .7truecm [|c||c|c|c|c|c|c|c|]{} \ & & & & & & &\ [[** **]{}]{}\ & & & & & & &\ \ \[StandardSummary\] $(1,1,1,1,1,\langle 4,4\rangle)$ & 6 & 84 & 3,5,6 & Yes & 0 & 342 & 6.14%\ $(1,1,1,1,\langle 10,10\rangle)$ & 6 & 48 & 3,5,6 & Yes & 0 & 124 & 4.84%\ $(1,1,1,1,\langle 2,2\rangle,4)$ & 6 & 48 & 3,5,6 & Yes & 0 & 75 & 6.67%\ $(1,1,1,\langle 4,4\rangle,4)$ & 6 & 84 & 3,5,6 & Yes & 0 & 2717 & 22.89%\ $(1,1,2,2,\langle 4,4\rangle)$ & 6 & 24 & 3,5,6 & Yes & 0 & 106 & 0.00%\ $(1,1,\langle 2,2\rangle,2,10)$ & 6 & 48 & 3,5,6 & Yes & 0 & 662 & 7.70%\ $(1,1,4,\langle 10,10\rangle)$ & 6 & 72 & 3,5,6 & Yes & 0 & 493 & 7.10%\ $(1,1,\langle 6,6\rangle,10)$ & 12 & 24 & 3,5,6 & Yes & 0 & 63 & 0.00%\ $(1,1,\langle 2,2\rangle,4,4)$ & 6 & 48 & 3,5,6 & Yes & 0 & 226 & 6.19%\ $(1,1,\langle 2,2\rangle,\langle 4,4\rangle)$ & 12 & 24 & 3,5,6 & Yes & 0 & 73 & 6.85%\ $(1,2,2,\langle 10,10\rangle)$ & 6 & 60 & 3,5,6 & Yes & 0 & 191 & 4.71%\ $(1,\langle 2,2\rangle,2,2,4)$ & 12 & 60 & 3,5,6 & Yes & 0 & 363 & 3.31%\ $(1,2,4,\langle 6,6\rangle)$ & 12 & 48 & 3,5,6 & Yes & 0 & 57 & 3.51%\ $(1,2,\langle 4,4\rangle,10)$ & 6 & 60 & 3,5,6 & Yes & 0 & 1605 & 14.08%\ $(1,2,\langle 3,3\rangle,58)$ & 6 & 24 & 0,1,2,3,4,5,6 & Yes & 0 & 102 & 0.00%\ $(1,\langle 4,4\rangle,4,4)$ & 6 & 84 & 3,5,6 & Yes & 0 & 5605 & 6.57%\ $(1,\langle 2,2\rangle,10,10)$ & 6 & 84 & 3,5,6 & Yes & 0 & 989 & 6.47%\ $(1,\langle 3,3\rangle,4,8)$ & 6 & 36 & 0,1,2,3,4,5,6 & Yes & 0 & 37 & 0.00%\ $(1,\langle 2,2\rangle,6,22)$ & 6 & 60 & 3,5,6 & Yes & 0 & 985 & 3.25%\ $(1,\langle 2,2\rangle,7,16)$ & 12 & 48 & 3,5,6 & Yes & 0 & 41 & 0.00%\ $(1,\langle 2,2\rangle,\langle 2,2\rangle,4)$ & 12 & 60 & 3,5,6 & Yes & 0 & 165 & 0.61%\ $(\langle 2,2\rangle,2,2,2,2)$ & 6 & 90 & 6 & Yes & 0 & 1849 & 5.19%\ $(2,2,2,\langle 6,6\rangle)$ & 12 & 72 & 6 & Yes & 0 & 245 & 0.00%\ $(2,2,\langle 4,4\rangle,4)$ & 6 & 48 & 3,5,6 & Yes & 0 & 250 & 0.00%\ $(2,2,\langle 3,3\rangle,8)$ & 6 & 36 & 2,4,6 & Yes & 0 & 55 & 0.00%\ $(2,2,\langle 2,2\rangle,\langle 2,2\rangle)$ & 6 & 90 & 6 & Yes & 0 & 1580 & 1.58%\ $(2,\langle 10,10\rangle,10)$ & 6 & 102 & 3,5,6 & Yes & 0 & 328 & 0.00%\ $(2,\langle 8,8\rangle,18)$ & 6 & 72 & 2,4,6 & Yes & 0 & 316 & 0.00%\ $(\langle 2,2\rangle,2,3,18)$ & 6 & 60 & 2,4,6 & Yes & 0 & 780 & 4.36%\ $(2,\langle 7,7\rangle,34)$ & 12 & 48 & 3,5,6 & Yes & 0 & 9 & 0.00%\ $(\langle 2,2\rangle,2,4,10)$ & 6 & 66 & 3,5,6 & Yes & 0 & 1550 & 3.81%\ $(\langle 2,2\rangle,2,6,6)$ & 6 & 84 & 6 & Yes & 0 & 1735 & 3.80%\ $(2,\langle 2,2\rangle,\langle 6,6\rangle)$ & 12 & 72 & $SO(10)$ only & No & 0 & 219 & 0.00%\ $(3,\langle 6,6\rangle,18)$ & 4 & 56 & 2,4,6 & No & 0 & 232 & 0.00%\ $(3,\langle 5,5\rangle,68)$ & 24 & 24 & 4 & No & 0 & 18 & 0.00%\ $(3,\langle 8,8\rangle,8)$ & 6 & 96 & 2,4,6 & Yes & 0 & 1909 & 1.41%\ $(3,\langle 3,3\rangle,\langle 3,3\rangle)$ & 2 & 56 & 4 & No & 0 & 126 & 0.00%\ $(4,4,\langle 10,10\rangle)$ & 6 & 90 & 5 & Yes & 0 & 188 & 0.00%\ $(4,\langle 6,6\rangle,10)$ & 12 & 48 & 5 & Yes & 0 & 70 & 0.00%\ $(4,\langle 5,5\rangle,19)$ & 12 & 24 & 5 & Yes & 0 & 6 & 0.00%\ $(4,\langle 7,7\rangle,7)$ & 12 & 60 & 5 & Yes & 0 & 11 & 0.00%\ $(\langle 5,5\rangle,5,12)$ & 6 & 78 & $SO(10)$ only & No & 0 & 44 & 0.00%\ $(\langle 6,6\rangle,6,6)$ & 2 & 104 & 6 & Yes & 0 & 1230 & 0.00%\ $(\langle 4,4\rangle,10,10)$ & 6 & 96 & 3,5,6 & Yes & 0 & 693 & 0.72%\ $(\langle 3,3\rangle,10,58)$ & 6 & 60 & 0,1,2,3,4,5,6 & Yes & 0 & 97 & 0.00%\ $(\langle 3,3\rangle,12,33)$ & 2 & 20 & 4 & No & 0 & 30 & 0.00%\ $(\langle 3,3\rangle,13,28)$ & 6 & 84 & 1,4,5 & Yes & 0 & 587 & 0.00%\ $(\langle 3,3\rangle,18,18)$ & 2 & 116 & 2,4,6 & Yes & 0 & 681 & 0.00%\ $(\langle 2,2\rangle,3,3,8)$ & 6 & 48 & 2,4,6 & Yes & 0 & 332 & 3.61%\ $(\langle 2,2\rangle,4,4,4)$ & 6 & 54 & 5 & Yes & 0 & 75 & 0.00%\ $(\langle 4,4\rangle,5,40)$ & 6 & 48 & 3,5,6 & Yes & 0 & 98 & 0.00%\ $(\langle 4,4\rangle,6,22)$ & 6 & 60 & 3,5,6 & Yes & 0 & 440 & 0.00%\ $(\langle 4,4\rangle,7,16)$ & 6 & 72 & 3,5,6 & Yes & 0 & 271 & 0.00%\ $(\langle 4,4\rangle,8,13)$ & 6 & 48 & 0,1,2,3,4,5,6 & Yes & 0 & 180 & 0.00%\ $(\langle 3,3\rangle,9,108)$ & 2 & 28 & 4 & No & 0 & 30 & 0.00%\ $(\langle 6,6\rangle,\langle 6,6\rangle)$ & 4 & 80 & $SO(10)$ only & No & 0 & 152 & 0.00%\ $(\langle 2,2\rangle,\langle 4,4\rangle,4)$ & 6 & 48 & 5 & Yes & 0 & 103 & 0.00%\ $(\langle 2,2\rangle,\langle 3,3\rangle,8)$ & 6 & 36 & 4 & No & 0 & 37 & 0.00%\ $(\langle 2,2\rangle,\langle 2,2\rangle,\langle 2,2\rangle)$ & 6 & 90 & $SO(10)$ only & No & 0 & 224 & 1.34%\ $(1,\langle 2,2\rangle,\langle 10,10\rangle)$ & 12 & 60 & 3,5,6 & Yes & 0 & 155 & 0.00%\ $(\langle 4,4\rangle,\langle 10,10\rangle)$ & 6 & 72 & 5 & Yes & 0 & 142 & 0.00%\ $(1,\langle 4,4\rangle,\langle 4,4\rangle)$ & 6 & 84 & 3,5,6 & Yes & 0 & 848 & 0.83%\ =14truecm .7truecm [|c||c|c|c|c|c|c|c|]{} \ & & & & & & &\ [[** **]{}]{}\ & & & & & & &\ \ \[HWLSummary\] $(\widehat{1},1,1,1,1,\langle 4,4\rangle)$ & 3 & 33 & 3,5,6 & Yes & 45 & 205 & 16.10%\ $(\widehat{1},1,1,1,\langle 10,10\rangle)$ & 3 & 24 & 3,5,6 & Yes & 0 & 39 & 2.56%\ $(\widehat{1},1,1,1,\langle 2,2\rangle,4)$ & 3 & 18 & 3,5,6 & Yes & 16 & 50 & 14.00%\ $(\widehat{1},1,1,\langle 4,4\rangle,4)$ & 3 & 33 & 3,5,6 & Yes & 549 & 1016 & 28.54%\ $(\widehat{1},1,2,2,\langle 4,4\rangle)$ & 3 & 12 & 3,5,6 & Yes & 17 & 60 & 0.00%\ $(\widehat{1},1,\langle 2,2\rangle,2,10)$ & 3 & 24 & 3,5,6 & Yes & 123 & 283 & 7.42%\ $(\widehat{1},1,4,\langle 10,10\rangle)$ & 3 & 24 & 3,5,6 & Yes & 33 & 206 & 7.77%\ $(\widehat{1},1,\langle 6,6\rangle,10)$ & 6 & 6 & 3,5,6 & Yes & 0 & 15 & 0.00%\ $(\widehat{1},1,\langle 2,2\rangle,4,4)$ & 3 & 24 & 3,5,6 & Yes & 34 & 237 & 4.64%\ $(\widehat{1},1,\langle 2,2\rangle,\langle 4,4\rangle)$ & 6 & 12 & 3,5,6 & Yes & 0 & 38 & 0.00%\ $(\widehat{1},2,2,\langle 10,10\rangle)$ & 12 & 24 & 3,5,6 & Yes & 0 & 18 & 0.00%\ $(\widehat{1},\langle 2,2\rangle,2,2,4)$ & 6 & 24 & 3,5,6 & Yes & 0 & 71 & 7.04%\ $(\widehat{1},2,\langle 3,3\rangle,58)$ & 1 & 8 & 0,1,2,3,4,5,6 & Yes & 2 & 40 & 0.00%\ $(\widehat{1},\langle 2,2\rangle,10,10)$ & 6 & 24 & 3,5,6 & Yes & 0 & 105 & 5.71%\ $(\widehat{1},\langle 3,3\rangle,4,8)$ & 2 & 8 & 0,1,2,3,4,5,6 & Yes & 0 & 23 & 0.00%\ $(\widehat{1},\langle 2,2\rangle,6,22)$ & 3 & 24 & 3,5,6 & Yes & 58 & 281 & 5.69%\ $(\widehat{1},\langle 2,2\rangle,7,16)$ & 6 & 12 & 3,5,6 & Yes & 0 & 13 & 0.00%\ $(\widehat{2},\langle 2,2\rangle,2,2,2)$ & 1 & 36 & 6 & Yes & 587 & 10481 & 6.91%\ $(\widehat{2},2,2,\langle 6,6\rangle)$ & 2 & 36 & 6 & Yes & 0 & 595 & 0.17%\ $(\widehat{2},2,\langle 3,3\rangle,8)$ & 1 & 10 & 2,4,6 & Yes & 3 & 51 & 0.00%\ $(\widehat{2},2,\langle 2,2\rangle,\langle 2,2\rangle)$ & 2 & 24 & 6 & Yes & 0 & 807 & 1.73%\ $(\widehat{2},\langle 10,10\rangle,10)$ & 4 & 8 & 3,5,6 & Yes & 0 & 24 & 0.00%\ $(\widehat{2},\langle 8,8\rangle,18)$ & 1 & 12 & 2,4,6 & Yes & 6 & 85 & 0.00%\ $(\widehat{2},\langle 2,2\rangle,3,18)$ & 1 & 24 & 2,4,6 & Yes & 26 & 225 & 4.00%\ $(\widehat{2},\langle 2,2\rangle,4,10)$ & 2 & 24 & 3,5,6 & Yes & 0 & 89 & 3.37%\ $(\widehat{2},\langle 2,2\rangle,6,6)$ & 1 & 24 & 6 & Yes & 9 & 305 & 1.97%\ $(\widehat{3},\langle 6,6\rangle,18)$ & 2 & 20 & 2,4,6 & No & 0 & 85 & 0.00%\ $(\widehat{3},\langle 5,5\rangle,68)$ & 12 & 12 & 4 & No & 0 & 4 & 0.00%\ $(\widehat{3},\langle 8,8\rangle,8)$ & 1 & 48 & 2,4,6 & Yes & 146 & 1709 & 0.06%\ $(\widehat{3},\langle 3,3\rangle,\langle 3,3\rangle)$ & 1 & 28 & 4 & No & 11 & 80 & 0.00%\ $(\widehat{4},4,\langle 10,10\rangle)$ & 2 & 16 & 5 & Yes & 0 & 47 & 0.00%\ $(\widehat{4},\langle 7,7\rangle,7)$ & 1 & 3 & 5 & Yes & 2 & 6 & 0.00%\ $(\widehat{6},\langle 6,6\rangle,6)$ & 1 & 8 & 6 & Yes & 7 & 85 & 0.00%\ $(\langle 3,3\rangle,10,\widehat{58})$ & 1 & 6 & 0,1,2,3,4,5,6 & Yes & 2 & 11 & 0.00%\ $(\langle 3,3\rangle,\widehat{12},33)$ & 2 & 6 & 4 & No & 0 & 5 & 0.00%\ $(\langle 3,3\rangle,\widehat{13},28)$ & 1 & 30 & 1,4,5 & Yes & 29 & 170 & 0.00%\ $(\langle 2,2\rangle,\widehat{3},3,8)$ & 1 & 24 & 2,4,6 & Yes & 14 & 479 & 4.38%\ $(\langle 2,2\rangle,\widehat{4},4,4)$ & 2 & 24 & 5 & Yes & 0 & 111 & 0.90%\ $(\langle 4,4\rangle,\widehat{6},22)$ & 8 & 8 & 3,5,6 & Yes & 0 & 5 & 0.00%\ $(\langle 4,4\rangle,\widehat{8},13)$ & 4 & 16 & 0,1,2,3,4,5,6 & Yes & 0 & 17 & 0.00%\ $(\langle 3,3\rangle,\widehat{9},108)$ & 2 & 6 & 4 & No & 0 & 6 & 0.00%\ $(\widehat{4},\langle 2,2\rangle,\langle 4,4\rangle)$ & 4 & 20 & 5 & Yes & 0 & 53 & 0.00%\ $(\langle 2,2\rangle,\langle 3,3\rangle,\widehat{8})$ & 2 & 6 & 4 & No & 0 & 11 & 0.00%\ $(1,1,1,\widehat{4},\langle 4,4\rangle)$ & 1 & 24 & 3,5,6 & Yes & 78 & 859 & 20.84%\ $(1,1,\widehat{2},2,\langle 4,4\rangle)$ & 1 & 6 & 3,5,6 & Yes & 0 & 75 & 4.00%\ $(1,1,\widehat{2},\langle 2,2\rangle,10)$ & 1 & 24 & 3,5,6 & Yes & 20 & 323 & 8.05%\ $(1,1,\widehat{4},\langle 10,10\rangle)$ & 2 & 16 & 3,5,6 & Yes & 0 & 191 & 4.71%\ $(1,1,\langle 2,2\rangle,\widehat{4},4)$ & 2 & 32 & 3,5,6 & Yes & 0 & 297 & 5.05%\ $(1,\widehat{2},2,\langle 10,10\rangle)$ & 1 & 16 & 3,5,6 & Yes & 28 & 262 & 0.00%\ $(1,\langle 2,2\rangle,2,2,\widehat{4})$ & 4 & 32 & 3,5,6 & Yes & 0 & 118 & 8.47%\ $(1,\widehat{2},4,\langle 6,6\rangle)$ & 2 & 24 & 3,5,6 & Yes & 0 & 77 & 0.00%\ $(1,\widehat{2},\langle 4,4\rangle,10)$ & 1 & 16 & 3,5,6 & Yes & 147 & 1160 & 8.71%\ $(1,\widehat{2},\langle 3,3\rangle,58)$ & 1 & 8 & 0,1,2,3,4,5,6 & Yes & 3 & 56 & 0.00%\ $(1,\widehat{4},\langle 4,4\rangle,4)$ & 1 & 24 & 3,5,6 & Yes & 425 & 3219 & 6.28%\ $(1,\langle 3,3\rangle,\widehat{4},8)$ & 2 & 10 & 0,1,2,3,4,5,6 & Yes & 0 & 27 & 0.00%\ $(1,\langle 2,2\rangle,\widehat{6},22)$ & 1 & 48 & 3,5,6 & Yes & 46 & 645 & 3.41%\ $(2,2,\widehat{4},\langle 4,4\rangle)$ & 4 & 20 & 3,5,6 & Yes & 0 & 93 & 0.00%\ $(2,2,\langle 3,3\rangle,\widehat{8})$ & 2 & 8 & 2,4,6 & Yes & 0 & 29 & 0.00%\ $(\langle 2,2\rangle,2,\widehat{3},18)$ & 2 & 30 & 2,4,6 & Yes & 0 & 380 & 2.63%\ $(\langle 2,2\rangle,2,\widehat{4},10)$ & 4 & 24 & 3,5,6 & Yes & 0 & 107 & 0.93%\ $(\langle 2,2\rangle,2,\widehat{6},6)$ & 2 & 48 & 6 & Yes & 0 & 477 & 2.73%\ $(3,\widehat{8},\langle 8,8\rangle)$ & 1 & 32 & 2,4,6 & Yes & 24 & 480 & 0.00%\ $(\langle 5,5\rangle,5,\widehat{12})$ & 1 & 6 & $SO(10)$ only & No & 2 & 8 & 0.00%\ $(\langle 2,2\rangle,3,3,\widehat{8})$ & 1 & 18 & 2,4,6 & Yes & 6 & 116 & 0.00%\ $(\langle 4,4\rangle,8,\widehat{13})$ & 2 & 14 & 0,1,2,3,4,5,6 & Yes & 0 & 20 & 0.00%\ $(1,\widehat{2},\langle 2,2\rangle,2,4)$ & 2 & 24 & 3,5,6 & Yes & 0 & 1092 & 4.85%\ $(1,2,\langle 3,3\rangle,\widehat{58})$ & 2 & 12 & 0,1,2,3,4,5,6 & Yes & 0 & 11 & 0.00%\ $(1,\langle 3,3\rangle,4,\widehat{8})$ & 2 & 6 & 0,1,2,3,4,5,6 & Yes & 0 & 10 & 0.00%\ =14truecm .7truecm [|c||c|c|c|c|c|c|c|]{} \ & & & & & & &\ [[** **]{}]{}\ & & & & & & &\ \ \[BLASummary\] $(1,2,\langle 3,3\rangle,58)$ & 1 & 6 & 0,1,2,3,4,5,6 & Yes & 4 & 33 & 0.00%\ $(1,\langle 3,3\rangle,4,8)$ & 2 & 6 & 0,1,2,3,4,5,6 & Yes & 0 & 12 & 0.00%\ $(2,2,\langle 3,3\rangle,8)$ & 2 & 8 & 2,4,6 & Yes & 0 & 25 & 0.00%\ $(2,\langle 8,8\rangle,18)$ & 1 & 12 & 2,4,6 & Yes & 14 & 90 & 0.00%\ $(\langle 2,2\rangle,2,3,18)$ & 2 & 18 & 2,4,6 & Yes & 0 & 364 & 3.85%\ $(3,\langle 6,6\rangle,18)$ & 2 & 14 & 2,4,6 & No & 0 & 84 & 0.00%\ $(3,\langle 5,5\rangle,68)$ & 6 & 6 & 4 & No & 0 & 12 & 0.00%\ $(3,\langle 8,8\rangle,8)$ & 1 & 30 & 2,4,6 & Yes & 238 & 1799 & 0.11%\ $(3,\langle 3,3\rangle,\langle 3,3\rangle)$ & 1 & 18 & 4 & No & 15 & 84 & 0.00%\ $(\langle 3,3\rangle,10,58)$ & 1 & 10 & 0,1,2,3,4,5,6 & Yes & 1 & 19 & 0.00%\ $(\langle 3,3\rangle,12,33)$ & 2 & 4 & 4 & No & 0 & 6 & 0.00%\ $(\langle 3,3\rangle,13,28)$ & 1 & 9 & 1,4,5 & Yes & 57 & 346 & 0.00%\ $(\langle 3,3\rangle,18,18)$ & 1 & 14 & 2,4,6 & Yes & 30 & 200 & 0.00%\ $(\langle 2,2\rangle,3,3,8)$ & 1 & 12 & 2,4,6 & Yes & 30 & 246 & 0.00%\ $(\langle 4,4\rangle,8,13)$ & 2 & 8 & 0,1,2,3,4,5,6 & Yes & 0 & 49 & 0.00%\ $(\langle 3,3\rangle,9,108)$ & 2 & 4 & 4 & No & 0 & 6 & 0.00%\ $(\langle 2,2\rangle,\langle 3,3\rangle,8)$ & 2 & 6 & 4 & No & 0 & 12 & 0.00%\ =14truecm .7truecm [|c||c|c|c|c|c|c|c|]{} \ & & & & & & &\ [[** **]{}]{}\ & & & & & & &\ \ \[BLBSummary\] $(1,2,\langle 3,3\rangle,58)$ & 2 & 10 & 0,1,2,3,4,5,6 & Yes & 0 & 32 & 0.00%\ $(1,\langle 3,3\rangle,4,8)$ & 2 & 10 & 0,1,2,3,4,5,6 & Yes & 0 & 10 & 0.00%\ $(2,2,\langle 3,3\rangle,8)$ & 2 & 14 & 2,4,6 & Yes & 0 & 34 & 0.00%\ $(2,\langle 8,8\rangle,18)$ & 2 & 16 & 2,4,6 & Yes & 0 & 108 & 0.00%\ $(\langle 2,2\rangle,2,3,18)$ & 2 & 36 & 2,4,6 & Yes & 0 & 476 & 3.99%\ $(3,\langle 6,6\rangle,18)$ & 4 & 28 & 2,4,6 & No & 0 & 82 & 0.00%\ $(3,\langle 5,5\rangle,68)$ & 8 & 16 & 4 & No & 0 & 12 & 0.00%\ $(3,\langle 8,8\rangle,8)$ & 2 & 56 & 2,4,6 & Yes & 0 & 1781 & 0.00%\ $(3,\langle 3,3\rangle,\langle 3,3\rangle)$ & 2 & 32 & 4 & No & 0 & 81 & 0.00%\ $(\langle 3,3\rangle,10,58)$ & 2 & 18 & 0,1,2,3,4,5,6 & Yes & 0 & 18 & 0.00%\ $(\langle 3,3\rangle,12,33)$ & 2 & 8 & 4 & No & 0 & 6 & 0.00%\ $(\langle 3,3\rangle,13,28)$ & 2 & 18 & 1,4,5 & Yes & 0 & 322 & 0.00%\ $(\langle 3,3\rangle,18,18)$ & 2 & 26 & 2,4,6 & Yes & 0 & 191 & 0.00%\ $(\langle 2,2\rangle,3,3,8)$ & 2 & 24 & 2,4,6 & Yes & 0 & 226 & 0.00%\ $(\langle 4,4\rangle,8,13)$ & 4 & 16 & 0,1,2,3,4,5,6 & Yes & 0 & 45 & 0.00%\ $(\langle 3,3\rangle,9,108)$ & 2 & 10 & 4 & No & 0 & 6 & 0.00%\ $(\langle 2,2\rangle,\langle 3,3\rangle,8)$ & 2 & 10 & 4 & No & 0 & 10 & 0.00%\ [99]{} D. J. Gross, J. A. Harvey, E. J. Martinec, R. Rohm, “The Heterotic String", Phys. Rev. Lett.  [**54** ]{} (1985) 502-505. A. A. Belavin, A. M. Polyakov and A. B. Zamolodchikov, “Infinite conformal symmetry in two-dimensional quantum field theory,” Nucl. Phys.  B [**241**]{} (1984) 333. D. Gepner, “Space-Time Supersymmetry in Compactified String Theory and Superconformal Models,” Nucl. Phys.  B [**296**]{} (1988) 757. D. Gepner, “Exactly Solvable String Compactifications on Manifolds of SU(N) Holonomy,” Phys. Lett.  B [**199**]{} (1987) 380. Y. Kazama and H. Suzuki, “New N=2 Superconformal Field Theories and Superstring Compactification,” Nucl. Phys.  B [**321**]{} (1989) 232. A. Klemm and M. G. Schmidt, “Orbifolds by cyclic permutations of tensor product conformal field theories,” Phys. Lett.  B [**245**]{} (1990) 53. J. Fuchs, A. Klemm and M. G. Schmidt, “Orbifolds by cyclic permutations in Gepner type superstrings and in the corresponding Calabi-Yau manifolds,” Annals Phys.  [**214**]{} (1992) 221. A. N. Schellekens and S. Yankielowicz, “Extended Chiral Algebras and Modular Invariant Partition Functions” Nucl. Phys.  B [**327**]{} (1989) 673. A. N. Schellekens and S. Yankielowicz, “Modular invariants from simple currents: an explicit proof,” Phys. Lett.  B [**227**]{} (1989) 387. K. A. Intriligator, “Bonus Symmetry in Conformal Field Theory” Nucl. Phys.  B [**332**]{} (1990) 541. B. Gato-Rivera and A. N. Schellekens, “Complete classification of simple current modular invariants for (Z(p))\*\*k", Commun. Math. Phys.  [**145**]{} (1992) 85. M. Kreuzer and A. N. Schellekens, “Simple currents versus orbifolds with discrete torsion: A Complete classification,” Nucl. Phys.  B [**411**]{} (1994) 97 \[arXiv:hep-th/9306145\]. A. N. Schellekens and S. Yankielowicz, “New Modular Invariants for N=2 Tensor Products and Four-Dimensional Strings,” Nucl. Phys.  B [**330**]{} (1990) 103. A. Font, L. E. Ibanez, F. Quevedo, “String Compactifications And N=2 Superconformal Coset Constructions,” Phys. Lett.  [**B224** ]{} (1989) 79. A. N. Schellekens, “Field identification fixed points in N=2 coset theories,” Nucl. Phys.  B [**366**]{} (1991) 27. L. Borisov, M. B. Halpern and C. Schweigert, “Systematic approach to cyclic orbifolds,” Int. J. Mod. Phys.  A [**13**]{} (1998) 125 \[arXiv:hep-th/9701061\]. M. Maio and A. N. Schellekens, “Fixed Point Resolution in Extensions of Permutation Orbifolds,” Nucl. Phys.  B [**821**]{} (2009) 577 \[arXiv:0905.1632 \[hep-th\]\]. M. Maio and A. N. Schellekens, “Complete Analysis of Extensions of $D(n)_1$ Permutation Orbifolds,” arXiv:0907.3053 \[hep-th\]. M. Maio and A. N. Schellekens, “Formula for Fixed Point Resolution Matrix of Permutation Orbifolds,” Nucl. Phys.  B [**830**]{} (2010) 116 \[arXiv:0911.1901 \[hep-th\]\]. M. Maio and A. N. Schellekens, “Permutation Orbifold of N=2 Supersymmetric Minimal Models,” arXiv:1011.0934 \[hep-th\]. B. Gato-Rivera and A. N. Schellekens, “Heterotic Weight Lifting,” Nucl. Phys.  B [**828**]{} (2010) 375 \[arXiv:0910.1526 \[hep-th\]\]. B. Gato-Rivera and A. N. Schellekens, “Asymmetric Gepner Models (Revisited),” Nucl. Phys.  B [**841**]{} (2010) 100 \[arXiv:1003.6075 \[hep-th\]\]. B. Gato-Rivera, A. N. Schellekens and A. N. Schellekens, “Asymmetric Gepner Models II. Heterotic Weight Lifting,” arXiv:1009.1320 \[hep-th\]. B. Gato-Rivera and A. N. Schellekens, “Asymmetric Gepner Models III. B-L Lifting,” arXiv:1012.0796 \[hep-th\]. B. Gato-Rivera, A. N. Schellekens, “Asymmetric Gepner Models IV. E8 breaking,” (to appear). A. N. Schellekens and S. Yankielowicz, “Field identification fixed points in the coset construction,” Nucl. Phys.  B [**334**]{} (1990) 67. A. N. Schellekens and S. Yankielowicz, “Simple currents, modular invariants and fixed points,” Int. J. Mod. Phys.  A [**5**]{} (1990) 2903. J. Fuchs, A. N. Schellekens and C. Schweigert, “A matrix S for all simple current extensions,” Nucl. Phys.  B [**473**]{} (1996) 323 \[arXiv:hep-th/9601078\]. J. Fuchs, B. Schellekens and C. Schweigert, “From Dynkin diagram symmetries to fixed point structures,” Commun. Math. Phys.  [**180**]{} (1996) 39 \[arXiv:hep-th/9506135\]. A. N. Schellekens, “Fixed point resolution in extended WZW-models,” Nucl. Phys.  B [**558**]{} (1999) 484 \[arXiv:math/9905153\]. J. Fuchs, B. Schellekens, C. Schweigert, “The resolution of field identification fixed points in diagonal coset theories,” Nucl. Phys.  [**B461** ]{} (1996) 371-406. \[hep-th/9509105\]. G. Pradisi, A. Sagnotti, Y. .S. Stanev, “Completeness conditions for boundary operators in 2-D conformal field theory", Phys. Lett.  [**B381** ]{} (1996) 97-104. \[hep-th/9603097\]. G. Pradisi, A. Sagnotti, Y. .S. Stanev, “The Open descendants of nondiagonal SU(2) WZW models", Phys. Lett.  [**B356** ]{} (1995) 230-238. \[hep-th/9506014\]. J. Fuchs, L. R. Huiszoon, A. N. Schellekens, C. Schweigert, J. Walcher, “Boundaries, crosscaps and simple currents", Phys. Lett.  [**B495** ]{} (2000) 427-434. \[hep-th/0007174\]. W. Lerche, D. Lüst and A. N. Schellekens, “Chiral Four-Dimensional Heterotic Strings From Selfdual Lattices", Nucl. Phys. B [**287**]{} (1987) 477. D. Gepner, “String Theory On Calabi-Yau Manifolds: The Three Generations Case", arXiv:hep-th/9301089. www.nikhef.nl/$\sim$t58, page “Hodge numbers" A. N. Schellekens, “Electric charge quantization in String Theory", Phys. Lett.  B [**237**]{} (1990) 363. X. G. Wen and E. Witten, “Electric And Magnetic Charges In Superstring Models”, Nucl. Phys.  B [**261**]{} (1985) 651. G. G. Athanasiu, J. J. Atick, M. Dine and W. Fischler, “Remarks on Wilson line, Modular Invariance and Possible String Relics in Calabi-Yau Compactifications", Phys. Lett.  B [**214**]{} (1988) 55. M. L. Perl, E. R. Lee and D. Loomba, “Searches For Fractionally Charged Particles", Ann. Rev. Nucl. Part. Sci.  [**59**]{} (2009) 47. B. Assel, K. Christodoulides, A. E. Faraggi, C. Kounnas and J. Rizos, “Exophobic Quasi-Realistic Heterotic String Vacua,” Phys. Lett.  B [**683**]{}, 306 (2010) \[arXiv:0910.3697 \[hep-th\]\].\ J. Rizos, “The ’landscape’ of Pati–Salam heterotic superstring vacua,” arXiv:1003.0458 \[hep-th\].\ B. Assel, K. Christodoulides, A. E. Faraggi, C. Kounnas and J. Rizos, “Classification of Heterotic Pati-Salam Models,” arXiv:1007.2268 \[hep-th\]. E. Witten, “Symmetry Breaking Patterns in Superstring Models", Nucl. Phys.  [**B258** ]{} (1985) 75. M. Blaszczyk, S. G. Nibbelink, M. Ratz, F. Ruehle, M. Trapletti, P. K. S. Vaudrevange, “A Z2xZ2 standard model”, Phys. Lett.  [**B683** ]{} (2010) 340-348. \[arXiv:0911.4905 \[hep-th\]\]. W. Buchmuller, K. Hamaguchi, O. Lebedev and M. Ratz, “Supersymmetric standard model from the heterotic string", Phys. Rev. Lett.  [**96**]{} (2006) 121602 \[arXiv:hep-ph/0511035\]. O. Lebedev, H. P. Nilles, S. Raby, S. Ramos-Sanchez, M. Ratz, P. K. S. Vaudrevange and A. Wingerter, “A mini-landscape of exact MSSM spectra in heterotic orbifolds", Phys. Lett.  B [**645**]{} (2007) 88 \[arXiv:hep-th/0611095\]. O. Lebedev, H. P. Nilles, S. Ramos-Sanchez, M. Ratz and P. K. S. Vaudrevange, “Heterotic mini-landscape (II): completing the search for MSSM vacua in a $Z_6$ orbifold", Phys. Lett.  B [**668**]{} (2008) 331 \[arXiv:0807.4384 \[hep-th\]\]. A. E. Faraggi, C. Kounnas and J. Rizos, “Chiral family classification of fermionic Z(2) x Z(2) heterotic orbifold models,” Phys. Lett.  B [**648**]{} (2007) 84 \[arXiv:hep-th/0606144\]. T. P. T. Dijkstra, L. R. Huiszoon and A. N. , “Chiral supersymmetric Standard Model spectra from orientifolds of Gepner models,” Phys. Lett.  B [**609**]{} (2005) 408;\ “Supersymmetric Standard Model spectra from RCFT orientifolds,” Nucl. Phys.  B [**710**]{} (2005) 3. ![Distribution of the number of families for permutation orbifolds of standard Gepner Models.[]{data-label="famplot_standard"}](FamiliesStandardGepner){width="17cm"} ![Distribution of the number of families for permutation orbifolds of lifted Gepner Models.[]{data-label="famplot_lift"}](FamiliesLiftedGepner){width="17cm"} ![Distribution of the number of families for permutation orbifolds of B-L lifted (lift A) Gepner Models.[]{data-label="famplot_liftA"}](FamiliesBLLifted1Gepner){width="17cm"} ![Distribution of the number of families for permutation orbifolds of B-L lifted (lift B) Gepner Models.[]{data-label="famplot_liftB"}](FamiliesBLLifted2Gepner){width="17cm"} [^1]: See [@Maio:2010eu] for exceptions. [^2]: According to this notation, $U(1)_N$ has $N$ primary fields, with $N$ always even. [^3]: Note that here we mean the unextended tensor product. In particular, world-sheet supersymmetry extensions are not implied. [^4]: Actually, since $\langle 0,T_F \rangle$ is a fixed point of $(T_F,1)$ in the unextended orbifold, there exist two fields $\langle 0,T_F \rangle_\alpha$ (with $\alpha=0,1$) in the supersymmetric orbifold corresponding to the two resolved fixed points. One can use any of them, since they produce the same CFT. [^5]: For a complete list of Hodge number and singlets of Gepner models see [@hodge]. [^6]: For the standard, unpermuted Gepner models, the number of genuine Hodge number pairs with world-sheet supersymmetry in both sectors is 906. A list can be found on the website [@hodge]. This includes Hodge numbers from asymmetric simple current MIPFs not suitable for orientifold models. [^7]: We thank the authors for making their raw data available to us. [^8]: This is sometimes called the “effective center condition", and eliminates for example the odd level simple currents of $A_1$, which have order two, but quarter-integer spins.

--- abstract: 'This paper presents, evaluates, and discusses a new software tool to automatically build Dynamic Bayesian Networks (DBNs) from ordinary differential equations (ODEs) entered by the user. The DBNs generated from ODE models can handle both data uncertainty and model uncertainty in a principled manner. The application, named PROFET, can be used for temporal data mining with noisy or missing variables. It enables automatic re-estimation of model parameters using temporal evidence in the form of data streams. For temporal inference, PROFET includes both standard fixed time step particle filtering and its extension, adaptive-time particle filtering algorithms. Adaptive-time particle filtering enables the DBN to automatically adapt its time step length to match the dynamics of the model. We demonstrate PROFET’s functionality by using it to infer the model variables by estimating the model parameters of four benchmark ODE systems. From the generation of the DBN model to temporal inference, the entire process is automated and is delivered as an open-source platform-independent software application with a comprehensive user interface. PROFET is released under the Apache License 2.0. Its source code, executable and documentation are available at <http:://profet.it.nuigalway.ie>.' author: - | Hamda Ajmal, Michael Madden and Catherine Enright\ School of Computer Science, National University of Ireland Galway\ h.ajmal1@nuigalway.ie, michael.madden@nuigalway.ie, cathenright@gmail.com title: 'PROFET: Construction and Inference of DBNs Based on Mathematical Models [^1]' --- =1 Introduction ============ Mathematical models of physical systems are widely available in many domains [@ottesen2004applied]. These embody existing expert knowledge and can be considered sufficient statistics of all prior experimentation in the domain. ODE models are generally available in mathematical, engineering and biological textbooks and research publications. Although these models have proven to be quite useful in many domains, they typically describe general population-level behaviors. For example, a large number of ODE models have been published to investigate various dynamic aspects of cancer tumor growth and treatment [@eisen2013mathematical]. However, parameterization of these models is generic, typically done in a theoretical manner or based on laboratory data or literature-derived data. Therefore, they often fail to capture specific clinical scenarios. An individual patient’s unique parameters can be considerably different from those of a population based model. To describe individuals, model parameters must be re-calibrated using observations of the individual. However, the observed data may be missing or noisy, or it could sparse or infrequent relative to the dynamics of the underlying system thus, making individualisation a challenging task. While ODE models are useful to understand the general treatments concepts, they are still not used for prescribing *personalized* treatment regimes at the clinical level [@agur2014personalizing]. Moreover, ODE models are completely deterministic with respect to their behavior, given a certain set of initial conditions [@vodovotz2009mechanistic]. Most real-world systems are rarely completely deterministic and they always contain some level of randomness or noise in them. The stochastic behavior of many real world systems indicates the need to account for stochasticity in ODE models. On the other hand, DBNs are well suited to handle uncertainty in a probabilistic fashion. However, discovering the structure of a DBN is a challenging task. While parameters are often learned from data, most DBN structures are constructed by hand, using knowledge elicited from domain experts [@chatterjee2010dbns], which is a difficult and a time-consuming process [@lucas2004bayesian]. Previous research work in our research group @enright2013 [@enright2011; @enright2013] has shown that mathematical models in the form of ODEs can be encapsulated into DBNs that incorporate a first order Euler solver, and can infer model values in future time slices. DBNs can reason efficiently with this powerful combination of domain knowledge and real-time data. They explicitly model measurement uncertainty and parameter uncertainty, allowing model parameters to be adjusted from initial approximate values to their correct values using real-time evidence. By doing this, the knowledge elicitation bottleneck is bypassed. The technique was previously applied to the problem of modeling glycaemia in patients in an Intensive Care Unit [@enright2010], producing promising results. This paper makes three key contributions. Firstly, we present a software application, PROFET, that automates the process of: (a) converting an ODE model to a DBN incorporating a first order Euler solver; and (b) inferring model values in future time slices using both standard fixed time step particle filtering [@gordon1993novel] and adaptive-time particle filtering [@enright2011] algorithms. Secondly, we evaluate PROFET by inferring model variables of four benchmark ODE models. Finally, we discuss various factors that impact the overall accuracy and performance of DBN inference. Without PROFET, the process of generating a DBN from an ODE model is time-consuming and requires a detailed understanding of the underlying mathematics, as it involves the laborious and error-prone tasks of parsing the mathematical equations, creating the DBN manually, and writing code to execute the inference algorithms. PROFET automates the complete methodology and it can be used by researchers to carry out probabilistic inference from DBNs derived from mathematical models in any application domain. DBNs for Mathematical Models ============================ Mathematical models of different physical systems are generally available in the form of ODEs in domains such as mathematics, engineering, biology and bio-medicine. These mathematical models embody existing expert knowledge. However, as discussed earlier, they are usually idealizations as they attempt to describe a system’s general dynamics, which can result in over-simplification and invite exception [@matos2001]. Building the DBN Structure -------------------------- In order to *individualize* a general mathematical model to a specific case, PROFET automatically maps it into a DBN. The DBN framework explicitly models noise as measurement and parameter uncertainty and then reduces the uncertainty over time by individualizing model parameters using temporal evidence. We automate the methodology we proposed in @enright2013 to map a system of ODEs to a DBN which can also be expanded to reason effectively with noisy and temporal data. This methodology is described below with the help of an example. Consider an Initial Value Problem: find [$X(t), Y(t),Z(t)$]{} such that [$X(t_0),Y(t_0),Z(t_0)$]{} are given and, [$$\begin{aligned} \label{eq:2} \tag{2} \begin{split} \frac{dX}{dt} &= a.X(t) + Y(t).Z(t) \\ \frac{dY}{dt} &= b(Y(t) - Z(t)) \\ \frac{dZ}{dt} &= c.Y(t) - Z(t) - X(t).Y(t) \end{split}\end{aligned}$$]{} for all $t > t_0$. This first order ODE system is known as the Lorenz model. We will revisit this model in Section \[lor\] for details. Here, $X$, $Y$ and $Z$ are the model variables and $a,b,c,d$ are the constant parameters. With PROFET the user simply enters the mathematical equations and the DBN structure is automatically built. The DBN approximates the values of model variables at times $t_1, t_2, t_3, ... $ using Euler’s method, that is: [$$X_{i+1} = X_{i} + h_{i}\frac{dX_i}{dt}\tag{3} \label{eq:3}$$]{} for $i = 0,1,...$, where $h_i = t_{i+1} - t_{i}$. Thus, the rate of change of $X$ at step $i$ is [$$\frac{dX_i}{dt} = \frac{X_{i+1} - X_{i}}{h_{i}} =:\Delta X_i \tag{4} \label{eq:4}$$ ]{} and we can rewrite (\[eq:3\]) as [$$X_{i+1} = X_{i} + h_{i}\Delta X_{i} \tag{5} \label{eq:5} \text{for all $i = 0,1,...$}$$ ]{} The same steps are followed to evaluate the values of model variables $Y$ and $Z$. Figure \[fig:2\] shows a graph of a DBN derived by PROFET from the Lorenz ODE model (\[eq:2\]). The differentials are represented using an Euler approximation by mapping (\[eq:4\]) and (\[eq:5\]) directly to the six deterministic nodes, $\Delta X$, $X$,$\Delta Y$, $Y$, $\Delta Z$, $Z$ . The parents of the nodes $\Delta X, \Delta Y, \Delta Z $ are set to be all the terms needed to evaluate them using (\[eq:4\]), in the same time slice of the DBN. In each time slice, the values of $X_{i+1}$, $Y_{i+1}$ and $Z_{i+1}$ are evaluated using (\[eq:5\]); hence the parents of nodes $X,Y,Z$ are set to be themselves and their corresponding $\Delta$ nodes from the previous time slice. Extra inter-slice arcs on nodes $a,b,c,d$ allow parameters to be tuned to the evidence over time. Model parameters are modeled as continuous nodes. The probability distributions on these nodes depend on the individual ODE model being incorporated. PROFET automates this methodology for any system of ODEs. It creates a sub-net for each equation in the model and adds dependencies between them, as dictated by the terms appearing in the right hand side of the equations. Measurement Uncertainty of Observed Data ---------------------------------------- The DBN provides a natural framework to deal with noisy data. The observed value of a variable that is to be approximated may contain an amount of measurement error. Following the algorithm we proposed in @enright2013, PROFET creates a separate node in the DBN for each observed value, which is dependent on its corresponding true value node. The observed variable nodes are continuous nodes; for each one, its mean ($\mu$) is the value of its parent node representing a true value and its standard deviation ($\sigma$) represents measurement uncertainty. Similarly, the actual inputs to a system may differ from the intended inputs, which are observed, and so a separate node for the intended (observed) input is added to the DBN. The true value node is conditionally dependent on its corresponding intended-value node. In the Lorenz model example, we assume that the values of $X$ are observed. It can be seen in Figure \[fig:2\] that an extra evidence node (black) is present in the model to represent the relationship between observed value of the variable $X$ and its true value. Model Parameter Re-Estimation ----------------------------- The ODE model parameters are represented as continuous nodes in the DBN. They are allowed to vary in each time step. The value of a model parameter at each time step is conditionally dependent on its previous value, as shown by the dashed arcs in Figure \[fig:2\]; they can therefore converge to values appropriate to the individual case over time, based on evidence from the temporal data streams. A user begins by entering some initial estimate of the ODE model parameters distributions. The values of the distribution of the model parameters can often be obtained from published literature. If there are no published values or if the published population differs from the current population under study, a reasonable guess can be made or the ODE model parameters can be learned from the data from a group of observations using standard ODE model fitting techniques. As currently implemented, model parameters can have Uniform, Truncated Gaussian or Linear Gaussian distributions. Evidence -------- In PROFET, evidence can be defined as either continuous (which remains constant until a new value is reported) or instantaneous (where a value is specified for each moment in time). To account for noise, as was discussed in Section 3.2, PROFET models observed nodes as continuous distributions whose mean ($\mu$) is the parent’s node value, the true variable value and whose standard deviation ($\sigma$) represents measurement uncertainty. This $\sigma$ can be configured by the user. Users can add continuous/instantaneous temporal data streams as evidence before running inference. Temporal Inference ------------------ Two types of inference algorithms are implemented, fixed time step particle filtering, originally proposed by @gordon1993novel and adaptive-time particle filtering as we proposed in @enright2013. In fixed time step particle filtering, the step size is chosen by the user. To minimize the numerical error, the step size chosen must be sufficiently small. Results are to be reported at each step while filtering and prediction are carried out. But reducing the step size increases computational cost, so a trade-off must be made. This can be a challenge for stiff problems where very small step sizes are required, so inference quickly becomes inefficient. To overcome this limitation, PROFET also implements the adaptive-time particle filtering. The user specifies the intervals at which the results should be reported, and the adaptive-time inference algorithm automatically adopts a suitable time step that may be smaller than this. It aims to control the numeric error introduced at each time step, while minimizing run-times. To make this possible, the local error must be estimated which is done using delta nodes. This estimated error is compared to a prescribed tolerance. If the tolerance is met, the current step is accepted and a new step size is proposed for the next step, which may be bigger. If the tolerance is exceeded, the current step is rejected and a reduced step size is proposed. Both algorithms are described in detail in @enright2013. ![ Screenshots of PROFET GUI. Left: Drag and drop equation editor. Right: Graphical depiction of resultant DBN. Centre: Graph of inferred data.[]{data-label="fig:3"}](Screenshots1){width="50.00000%"} Results of the inference are saved and can be displayed in form of a graph inside PROFET, making it easier for users to comprehend and analyze. Other Features of PROFET ======================== Figure \[fig:3\] shows some screenshots of the PROFET GUI. Users can enter ODEs via a drag-and-drop equation editor which is based on DragMath [@sangwin2011], modified with the authors’ permission. The equations are automatically parsed into different types of DBN nodes and the DBN is constructed. The user has complete control over configuring model parameters. PROFET uses GraphViz [@ellson2001] to draw DBNs in graphical form. The inferred data is summarized and can be displayed in form of chart, to facilitate analysis. PROFET has been tested on Windows 7 and Ubuntu 14.04. Since it is developed in Java, it can run on any operating system that supports the Java Virtual Machine. The underlying algorithm for DBN structure construction and inference is written in Lisp. Evaluation on Benchmark ODE Models ================================== In this section, we present four small to medium sized ODE models that we will use to benchmark inference and model parameter estimation in PROFET. Three out of these four models are biological systems. We chose them because @dondelinger2013ode applied their ODE model parameter inference methods to them. We evaluate the results of PROFET’s ODE model variable inference and model parameter re-estimation on these ODE models; however our experimental setup is different from that of @dondelinger2013ode. The essence of their work is to infer the ODE model parameters from multiple noisy time series. Our aim is to *individualize* the model parameters on a single time series data stream. Therefore, for each ODE model, our benchmark data is obtained using the R package deSolve. The values of the model parameters and the initial state of model variables to generate the benchmark data are taken from @dondelinger2013ode. Instead of adding noise to the data, we assume that the true values of the ODE model parameters are unknown and must be inferred from the population values. In order for PROFET to discover the correct model parameters, evidence (which in these cases are the true values taken from the benchmark solutions) is incorporated. PROFET automatically converts the ODE models to DBNs. In all of our experiments, the model parameter nodes are modeled as continuous nodes with linear Gaussian distributions. The initial state model distribution can be viewed as the distribution of the population values. To simulate a situation where we do not know the true values of model parameters and/or initial state model variables of the data, we assume that we have a rough idea of the population parameters. In Section 6, we discuss how users can set the population mean ($\mu$) and the standard deviation ($\sigma_i$) of the model parameters for the initial state model. From one time step to the next, every model parameter is allowed to vary by setting $\mu$ equal to its value at the previous time step and its standard deviation $\sigma_t$ on the transition model. Evidence for observed nodes is sampled from the benchmark data at different time points. To simulate a real world situation where evidence is sparse and infrequent, we deliberately sample evidence at sparse intervals. However, the evidence does not contain any noise. In our experiments, we set the natural time step of the DBN equal to the step size used for ODE simulation to generate the benchmark data. For fixed time inference, we set the step size equal to the natural time step of the DBN. Particles are re-sampled and summarized at each step. For adaptive-time inference, our framework automatically selects an appropriate step size to match the dynamics of the system. For each experiment, we calculate Root Mean Square Error (RMSE) and Mean Absolute error (MAE) after an initial run-in period. [.22]{} [.22]{} -------------------------------------------------------------------------- ------ ------- ------ ---------- --------- ---------- ---------- ------- ------- ------- ------- ------- ------- -------- ----- ------ (lr)[2-4]{} (lr)[5-8]{} (lr)[9-14]{} (lr)[15-17]{} **Natural time step** **Model Parameters** s $k_d$ d $\alpha$ $\beta$ $\gamma$ $\delta$ $k_1$ $k_2$ $k_3$ $k_4$ $k_m$ $V$ $a$ $b$ $c$ (lr)[2-4]{} (lr)[5-8]{} (lr)[9-14]{} (lr)[15-17]{} 0.7 0.50 1.3 2.2 1.15 4.2 0.8 0.09 0.3 0.08 0.3 0.1 0.04 -8/3 -10 28 0.1 0.05 0.1 0.1 0.1 0.1 0.1 0.01 0.1 0.01 0.1 0.1 0.01 0 0 1.12 0.01 0.01 0.01 0.05 0.05 0.05 0.05 0.001 0.01 0.001 0.01 0.01 0.001 0 0 0.1 1 0.46 1 2 1 4 1 0.07 0.6 0.05 0.3 0.017 0.3 $-8/3$ -10 28 **Initial State** $S$ $Sd$ $R$ $RS$ $X$ $Y$ $Z$ (lr)[2-4]{} (lr)[5-8]{} (lr)[9-14]{} (lr)[15-17]{} 1 0 1 0 2 2 2 1 0 1 0 1 1 1 **Inference** **No. of samples** **Evidence times** **RMSE, MAE** -------------------------------------------------------------------------- ------ ------- ------ ---------- --------- ---------- ---------- ------- ------- ------- ------- ------- ------- -------- ----- ------ The PIF4/5 Model ---------------- The DBN created from the PIF4/5 ODE model is shown in Figure \[fig5:sfig1\]. Initial values of model variables are set to the true values that are used to generate the benchmark solution. Benchmark data is generated by simulating the complete Locke 2-loop model. All other variables are discarded except $PIF4/5$ and $TOC1$. Evidence for the concentration of $TOC1$ is sampled from the benchmark data at each time step and for $PIF4/5$ at the time points shown in Table \[Setup\]. As the evidence does not contain any noise, we set the standard deviation ($\sigma$) to a very small value, 0.01 to quantify the measurement uncertainty between the true value and the intended/observed values evidence nodes. The mean and the standard deviation of the likely values of the intended-$TOC1$ node are calculated from data from the benchmark solution. We run the inference on the PIF4/5 model in PROFET from time $t=0$ to $t=24$ using standard fixed time step particle filtering using 100,000 particles. Complete details of experimental configuration and results are shown in Table \[Setup\]. [.22]{} ![[Predicted values of model variables in the ODE system. Dashed lines represent the benchmark solution. Solid lines are the predicted trajectories. Error bars show one standard deviation. Gray dots are the points where evidence is received. (a) The PIF4/5 model (b) Lotka-Volterra Model (c) The signal transduction cascade model (d) The Lorenz model]{}](PIF45Final "fig:"){width="\textwidth"} \[fig6:sfig1\] [.22]{} ![[Predicted values of model variables in the ODE system. Dashed lines represent the benchmark solution. Solid lines are the predicted trajectories. Error bars show one standard deviation. Gray dots are the points where evidence is received. (a) The PIF4/5 model (b) Lotka-Volterra Model (c) The signal transduction cascade model (d) The Lorenz model]{}](LotkaFinal "fig:"){width="\textwidth"} \[fig6:sfig2\] [.22]{} ![[Predicted values of model variables in the ODE system. Dashed lines represent the benchmark solution. Solid lines are the predicted trajectories. Error bars show one standard deviation. Gray dots are the points where evidence is received. (a) The PIF4/5 model (b) Lotka-Volterra Model (c) The signal transduction cascade model (d) The Lorenz model]{}](STCFinal "fig:"){width="\textwidth"} [.22]{} ![[Predicted values of model variables in the ODE system. Dashed lines represent the benchmark solution. Solid lines are the predicted trajectories. Error bars show one standard deviation. Gray dots are the points where evidence is received. (a) The PIF4/5 model (b) Lotka-Volterra Model (c) The signal transduction cascade model (d) The Lorenz model]{}](LorFinal "fig:"){width="\textwidth"} To validate the results, we compare the prediction accuracy of PROFET with the true benchmark solution. The graph in Figure \[fig6:sfig1\] shows the comparison of benchmark solution and the predicted solution. It can be seen in the graph that the accuracy of the predicted values of $PIF4/5$ concentration begins to improve as evidence is incorporated into the system. Even though incorrect model parameters are chosen at the outset, they are re-estimated at every time step and eventually converge to their true values. The RMSE and MAE of the predicted values of $PIF4/5$ is calculated after first evidence is received at time $t=4$ and are shown in Table \[Setup\]. The Lotka-Volterra Predator Prey Model -------------------------------------- We follow similar steps to those described in the previous section. The DBN constructed by PROFET is shown in Figure \[fig5:sfig2\]. We run the inference on the Lotka-Volterra model in PROFET from time $t=0$ to $t=2$. Sparse evidence is provided from data of model variable $X$ sampled from the benchmark solution. $\sigma$ is set to 0.01. In Figure \[fig6:sfig2\], we can see that as evidence is incorporated into the system, the prediction accuracy of PROFET improves. Even though incorrect values of model parameters were chosen at the start, the predicted results of PROFET are very close to those of benchmark solution. Full details of experimental configuration and results are shown in Table \[Setup\]. The Signal Transduction Cascade ------------------------------- Table \[Setup\] again shows the experimental setup for the Signal Transduction Cascade model. We run the inference on the model in PROFET from time $t=0$ to $t=100$. Following @dondelinger2013ode, we sample evidence of $Rpp$ at more time points during the earlier part of the time series, where the dynamics tend to be faster. As before, $\sigma$ is set to 0.01. In Figure \[fig6:sfig3\], we plot the data predicted by PROFET along with the benchmark solution. PROFET was able to predict the values of model variable $Rpp$ with high accuracy, even though incorrect model parameters are chosen at the outset. The Lorenz Model {#lor} ---------------- In addition to the above three ODE models, we also evaluate the functionality of PROFET by using it to infer the values of a more challenging model, the Lorenz system of ODEs. The Lorenz model is a simplified mathematical model for atmospheric convection [@lorenz1963deterministic] developed by Edward Lorenz. It can have a chaotic solution under a certain set of model parameters and initial conditions. The model is a system of three ODEs (\[eq:2\]). The Lorenz model is known to be extremely sensitive to the initial conditions of the model variables. In many real-world models, not all model variables have good initial values available, so these must be estimated from experience. This can lead to situations where the model must adjust rapidly over the first few time steps. To simulate such a scenario, we infer the model variables of the Lorenz model assuming that their initial values are unknown. We generate the benchmark data using ODE45 solver in R with values of model parameters as shown in Table \[Setup\], over the time interval $t=0$ to $t=5$. The resulting system is stiff: there are regions where the solution varies rapidly and standard numerical schemes often fail to yield a physically meaningful approximation to the solution unless extremely small step sizes are used. The time step used for ODE model simulation is 0.01. PROFET converts the Lorenz ODEs to a DBN, shown in Figure \[fig:2\]. As this solution is stiff, we use the adaptive-time algorithm for inference. A tolerance of 0.1 is chosen to limit the numeric error introduced at each time step. Sparse evidence is sampled from the benchmark data. It can be seen in the graph in Figure \[fig6:sfig4\] that PROFET manages to find a good solution even though incorrect initial values of the model variables are used. It takes the system first few time steps to adjust to the true trajectory. There are some spurious spikes at time points 0.4 and 0.5 because the inferred value of $X$ is far from the evidence (benchmark), but the accuracy improves as more evidence is incorporated into the system. The RMSE and MAE are shown in Table \[Setup\]. Factors Affecting Performance ============================= We have demonstrated how PROFET can be utilized to run inference over ODE models whose parameters or the initial values of the model variables are unknown. However it must be noted that there are a few factors that impact the outcome of the inference. It is important that reasonable distribution means ($\mu$) and standard deviations ($\sigma_i$) are chosen for model parameters, as they can hugely impact the accuracy of inferred values. The $\sigma_i$ of the initial state model parameter quantifies the amount of variation possible between the true model parameter value and the initially provided (potentially incorrect) value. The $\sigma_t$ on the transition model reflects how much the value of the model parameter is allowed to vary from its value at the previous time slice. A low transition $\sigma_t$ would make the model parameter at each time step close to its value in the previous time step, while a high $\sigma_t$ will allow bigger changes in the values of the model parameters. Thus, a large transition $\sigma_t$ is suitable for non-stationary problems. Similarly, the $\sigma$ on the observed nodes quantifies the measurement uncertainty of the observed data. If it is suspected that the evidence data is noisy, $\sigma$ should kept large enough to reflect that. The choice of natural time step of the DBN and the step size used for inference is also very important, especially in stiff ODE models where they must be extremely small to capture the dynamics of the underlying system. E.g., if ODE model variables tend to change rapidly within one time unit, it is reasonable to select a smaller step size. Increasing the number of time steps increases the numerical accuracy of the solution. Of course, this increases the computational cost, but our adaptive-time inference algorithm seeks to mitigate this. In our experiments with the Lorenz model, we use a very small natural time step of the DBN (0.01) and the adaptive-time algorithm automatically adjusts the step size during inference. The number of samples selected for particle filtering also plays an important role. As a rule of thumb, a higher number will improve accuracy, but at a price of increased computational cost. If a high variance is allowed for each node from the value at its previous node, it is reasonable to increase the number of samples, so that the sample space can be densely populated and there is a higher chance of filtering the values closer to the true value at each time slice. As would be expected, we also observe that the accuracy of the predicted results drops with an increase in the dimensionality of the model, keeping the number of samples fixed. As the number of model parameters whose values are to be estimated increases, the search space becomes sparse, and a larger number of samples are needed to find the best solution. The time required for inference increases linearly with the number of variables in the ODE and the number of particles. We do not anticipate scaling problems because real-world ODE models are usually formulated by domain experts and they do not generally involve a very large number of variables. Related Work ============ [\*[5]{}[ p[0.06]{} p[0.03]{}p[0.1]{}p[0.11]{}p[0.11]{}]{}]{} **Software** & **GUI** & **Structure Learning** & **Parameter Learning** & **Inference**\ GMTk & No & EM & GEM & Frontier Algorithm\ SMILE & GeNIe & Yes & No & No & Yes (by unrolling DBN into a BN)\ Mocapy++ & No & No & EM & Gibbs Sampling\ PNL & No & Hill Climbing & EM & 1.5 Slice J-Tree algorithm\ libDAI & No & No & EM, MLE & Various\ BayesiaLab^a^ & Yes & SopLEQ, Taboo Search & Spiegelhalter and Lauritzen Parameterization Algorithm & J-Tree, Gibbs Sampling\ Hugin^a^ & Yes & No & EM & J-Tree\ Netica^a^ & Yes & No & Spiegelhalter and Lauritzen Parameterization Algorithm, EM and Gradient Descent & Eliminiation J-Tree method\ BayesServer^a^ & Yes & PC Algorithm & EM & Exact-Relevance Tree, Exact-Variable Elimination, Loopy Belief Propagation, Likelihood Sampling\ **PROFET** & Yes & Automatic mapping from ODEs to DBN incorporating first order Euler Solver & Model Parameters adjusted in real time by incorporating evidence& Particle Filtering, adaptive-time Particle Filtering\ ^a^ Commercial Software \[table1\] Here, we present a review of somewhat related current software applications. In our survey, we have narrowed them down ten software applications that provide structure learning, parameter learning and inference on DBNs. Table \[table1\] presents a feature comparison of these software applications. There are also a few software packages for inference methods for ODE models. **BioBayes** [@vyshemirsky2008biobayes] is a software package that provides a framework for Bayesian parameter estimation of biochemical systems and evidential model ranking over models of biochemical systems defined using ODE models. For model parameter inference from experimental data, it implements a variant of Metropolis Hastings sampler [@hastings1970monte] and a population-based MCMC sampler [@jasra2007population]. Similarly, **ABC-SysBio** [@liepe2010abc] implements approximate Bayesian computation (ABC) for parameter inference and model selection in deterministic and stochastic ODE models. It combines ABC rejection sampler and ABC scheme based on Sequential Monte Carlo (ABC-SMC) [@toni2009approximate] for parameter inference. **GNU MCSim** [@bois2009gnu] is a numerical simulation and Bayesian statistical inference tool for algebraic or differential equation systems. **WinBUGS** [@lunn2000winbugs] Differential Interface **(WBDiff)** [@lunn2004wbdiff] is an extension of WinBUGS that allows the numerical solution of any arbitrary systems of ODEs within WinBUGS models. The Runge-Kutta algorithm is used to solve the equations and Metropolis-Hastings [@hastings1970monte] samplers are used for sampling unknown inputs. **NIMROD** [@prague2013nimrod] facilitates the user to make approximate Bayesian inference in models with random effects based on ODEs. It is based on penalized maximum likelihood [@guedj2007maximum]. The **Stan** programming language [@carpenter2016stan] can be used to fit the parameters of complex ODE models. It is a strongly-typed modeling language. Users can specify complex ODE models with minimal effort. It implements gradient-based MCMC algorithms for Bayesian inference, and gradient-based optimization for penalized MLE. Bayesian Logic (**BLOG**) [@blog] is a probabilistic programming language with a declarative syntax. It is designed to describe probabilistic graphical models and then perform inference in those models. Assumed Parameter Filter [@APF] is an approximate inference algorithm that implements a hybrid of particle filtering for state variables and assumed density filtering for parameter variables in State Space Models (SSMs). As it is integrated into BLOG, it can be used within the framework. **LibBi** [@murray2013] is a software package for Bayesian inference specialized for SSMs designed for parallel computing on high-performance computer hardware. It implements SMC, particle Markov chain Monte Carlo (PMCMC) and SMC$^{2}$ for inference. In our survey, we did not find any software application that facilitates probabilistic ODE model variable inference using particle filtering by converting them to DBNs. Conclusions =========== We have presented a user-friendly Java-based software application, called PROFET, that automatically converts first order ODE models to DBNs and performs temporal inference on them. The parameters of the DBN model are individualized as real time evidence is incorporated into the system. The software can be used by researchers in various domains interested in individualizing the general mathematical ODE models which are based on population level behavior. We have evaluated PROFET by using it to infer model variable values of four benchmark ODE systems. PROFET can predict data with high accuracy and can deal with noisy, missing, sparse or infrequent evidence, incorrect model parameters and/or incorrect initial state values. We have also discussed some factors that affect the performance of the DBN inference. PROFET is free and open source. It is licensed under Apache License 2.0. The project website is <http://profet.it.nuigalway.ie/> and source code and executable are maintained on our GitHub account <https://github.com/HamdaBinteAjmal/PROFET> . No installation is required to run it. However the Java Run Time Engine must be installed on the user’s machine. Detailed instructions and a comprehensive user manual, describing the software in a step-by-step approach through an example, are available on the website. [^1]: To cite this paper, please use this full conference citation: Hamda Ajmal, Michael Madden and Catherine Enright. PROFET: Construction and Inference of DBNs Based on Mathematical Models. In *Proceedings of the 14th UAI Bayesian Modelling Applications Workshop (BMAW 2017)*, co-located with the 33rd Conference on Uncertainty in Artificial Intelligence (UAI 2017) Sydney, Australia, 2017.

--- abstract: 'Non-linear gravitational clustering in a universe dominated by dark energy, modelled by a ‘quintessence’ scalar field, and cold dark matter with space-time varying mass is studied. Models of this type, where the variable mass is induced by dependence on the scalar field, as suggested by string theory or extra-dimensions, have been proposed as a viable solution of the coincidence problem. A general framework for the study of the non-linear phases of structure formation in scalar field cosmologies is provided, starting from a general relativistic treatment of the combined dark matter-dark energy system. As a first application, the mildly non-linear evolution of dark matter perturbations is obtained by a straightforward extension of the Zel’dovich approximation. We argue that the dark energy fluctuations may play an active role in cosmological structure formation if the scalar field effective potential develops a temporary spinodal instability during the evolution.' address: - |  Dipartimento di Fisica ‘G. Galilei’, Università di Padova,\ via Marzolo 8, I-35131 Padova. - |  INFN - Sezione di Padova,\ via Marzolo 8, I-35131 Padova. - | ¶ Dipartimento di Fisica ‘M. Melloni’, Università di Parma,\ INFN - Gruppo Collegato di Parma,\ Parco Area delle Scienze 7/A, I-43100 Parma author: - 'Sabino Matarrese, Massimo Pietroni and Carlo Schimd¶' title: 'Non-linear gravitational clustering in scalar field cosmologies' --- Introduction ============ There is increasing agreement on the picture of the Universe on large scales: the analysis of the magnitude-redshift relation for type Ia Supernovae [@sn], measurements of the anisotropies of the Cosmic Microwave Background (CMB) [@Spergel:2003cb], and analyses of the galaxy distribution in large redshift catalogs [@Verde:2001sf] all converge to show that the Universe is spatially flat and composed by roughly 30% of cold dark matter (DM), responsible for the structure we see via the small amount of baryons, and, for the remaining 70% by an unclustered form of dark energy (DE), responsible for its present-day accelerated expansion. In order to alleviate or avoid fine-tuning of initial conditions, an alternative solution to the usual cosmological constant has been proposed in the form of a dynamical vacuum energy component with negative equation of state, the so-called quintessence [@Caldwell:1998]. The most common candidate for quintessence is a self-interacting scalar field $\phi$. Different choices for the effective potential have been considered in the literature. In order to avoid fine tuning on the initial conditions, particular attention has been devoted to potentials leading to scaling or tracker behavior, like those depending on the scalar field exponentially [@Ferreira] or with a negative power-law [@Ratra]. In addition to self-interactions, the DE scalar field might in principle be coupled to any other particle or field present in Nature. However, given the extremely low mass typically expected for $\phi$ today, $m_\phi\sim H_0\sim 10^{-33} $ eV, its couplings to common particles, such as baryons or photons, are severely bounded by experimental tests on the equivalence principle and on the time variation of coupling constants. On the other hand, the coupling to non-baryonic DM is not so tightly constrained and might play a significant role in the cosmological evolution [@Damour]. In this paper we will take into account such a possibility by allowing the DM particles’ mass to depend on the DE scalar field. The DM particles are then Varying Mass Particles (VAMPS), a possibility that was discussed in the past in connection to the age problem of the Universe [@Anderson:1997un]. The VAMPS scenario was considered more recently in the quintessence context in [@CPR], where it was shown that it may lead to a solution of the so called ‘cosmic coincidence’ problem. Indeed, assuming an exponential form both for the coupling and for the effective potential, the late time cosmology may be such that the universe accelerates and the energy densities of DE and DM scale at the same rate. As expected, in this scenario the predictions for the DM abundance are drastically altered with respect to the standard freeze-out scenario [@CPR]. A coupling of this type may arise for radii moduli in braneworld scenarios [@Pietroni]. Couplings between DE and DM may also emerge in strongly coupled string theory [@GPV]. It is generally believed that perturbations of the quintessence scalar field play a negligible dynamical role in cosmic structure formation, because of the extreme smallness of the quintessence mass ($m_\phi\sim H$) which implies that its spatial fluctuations only appear on very large scales and are bound to be linear. Nevertheless, a number of models have been proposed where DE perturbations might in principle grow non-linear and play an active role down to galactic scales. These include coupled-quintessence [@Amendola:1999er], extended quintessence [@Perrotta:1999am; @Perrotta:2002sw], inhomogeneous Chaplygin gas [@Bilic:2001cg], k-essence [@Armendariz-Picon:2000ah]. A key question in the context of scalar field cosmologies is whether fluctuations of the quintessence field can influence the galaxy formation process in a sensible way, affecting the dynamics of structure formation during the non-linear phases e.g. in the cores of dark matter halos. This problem has been recently addressed by Wetterich [@Wetterich:2001], but the issue of whether quintessence fluctuations can grow non-linear on small scales and actively influence the non-linear clustering of the CDM component is still completely open. The VAMPS scenario considered here provides an useful bench mark to study situations where non-linearities in the scalar field may be induced either by the field’s self interaction or by its couplings to other cosmic components. The plan of the paper is as follows. In Section 2 we give a full derivation of the equations which govern the non-linear evolution of DM and DE perturbations on sub-horizon scales. Then, in order to gain some intuitive insight on the content of these equations, we specialize to the VAMP model discussed in [@CPR]. A simple solution of the equations in the mildly non-linear regime for this model is obtained in Section 3, in terms of first-order lagrangian perturbation theory. In Section 4 we discuss how the simple model considered here could be modified to give rise to non-negligible scalar field fluctuations on scales relevant for galaxy or galaxy-cluster formation. General formalism ================= As a starting point, we will expand the metric tensor $g_{\mu\nu}$ (greek indices run from 0 to 3, while latin ones from 1 to 3) to linear order around a flat Robertson-Walker (RW) metric, while keeping the full – [*i.e.*]{} non-linear – energy momentum tensor on the RHS of Einstein equations. In the conformal Newtonian gauge, vector and tensor perturbations of the metric decouple from the scalar ones as long as they are treated linearly; by taking the spatial covariant divergence of the $0-i$ Einstein equations and the trace of the $i-j$ equations, we get rid of vector and tensor modes and single out scalar ones. Thus, it is not restrictive to consider the metric [@Ma:1995] $$ds^2 = a^2(\eta)\left[ (1+2\Phi)d\eta^2 -(1-2\Psi)\delta_{ij}dx^i dx^j \right] \, ,$$ where the scalar perturbations $\Phi$ and $\Psi$ are related to the gravitational potential, and are of order $|{\bf u}|^2\ll 1$ (${\bf u}$ being the typical velocity of DM particles) even in the presence of highly non-linear DM overdensities. Here $a(\eta)$ is the background scale-factor and $\eta$ is the conformal time. In this approximation, Einstein’s equations read $$\label{eqn:hyb} \delta R^\mu_\nu = 8\pi G (S^\mu_\nu - \bar{S}^\mu_\nu )\, ,$$ where $\delta R_{\mu\nu}$ is the linear perturbation to the Ricci tensor, while $S^\mu_{\;\nu} \equiv T^\mu_{\;\nu} - \frac{1}{2}\delta^\mu_{\;\nu} T$ is the fully non-linear source and $\bar{S}^\mu_\nu$ is the background one. The Friedmann equations for the background RW metric read $$\begin{aligned} \mathcal{H}^2 &=& \frac{8\pi G}{3}a^2 \left( \bar{\rho}_{m} + \bar{\rho}_{\phi} \right) \label{fri1} \\ \dot{\mathcal{H}} &=& - \frac{4\pi G}{3}a^2 \left(\bar{ \rho}_{m} + \bar{\rho}_\phi + 3\bar{p}_\phi \right)\label{fri2} \, ,\end{aligned}$$ where $\mathcal{H}\equiv\dot{a}/a$, and the dot denotes differentiation with respect to $\eta$. We will take into account the contribution to the energy-momentum tensor coming from the dark matter and the scalar field. The former is made up of a discrete set of particles with field-dependent mass $m(\phi(\eta,{\bf x}))$ and coordinates ${\bf x}_a(\eta)$ ($a=1,2,\ldots$), described by the energy–momentum tensor $$T^\mu_{\;\nu} \simeq a^{-2} \, \rho_m(\phi) \,u^\mu u_\nu \,, \label{Tm}$$ where $$\rho_m(\phi)=m(\phi) \,a^{-3}\sum_{a} \delta^{(3)}({\bf x} - {\bf x}_a)\,,$$ $u^\mu\equiv dx^\mu/d\eta$ ($x^0\equiv\eta$), and we have neglected $O(|{\bf u}|^2)$ corrections. The scalar field energy-momentum tensor reads $$T^\mu_{\;\nu} = M_p^2 \phi^{;\mu}\phi_{\; ;\nu} - \delta^\mu_{\;\nu}\left( \frac{M_p^2}{2}\phi^{\; ;\alpha}\phi_{;\alpha} - V(\phi) \right) \;, \label{Tphi}$$ where $\phi$ is dimensionless because of the factorization of $M_p^2 = (8\pi G)^{-1}$. After replacing the background quantities $\bar{g}_{\mu\nu}$, $\bar{\phi}$, and $\bar{x}^\mu_a$ in (\[Tm\]) and (\[Tphi\]) one can read out the background energy density and pressure for our two components, $\bar{ \rho}_{m}$, $\bar{\rho}_\phi$, and $\bar{p}_\phi$. The Einstein equations provide a redundant set of equations for the evolution of the linearized metric perturbations; we choose to use the $0-0$ equation and the spatial covariant derivative of the $0-i$’s. They read, respectively, $$\begin{aligned} \fl \nabla^2\Phi +3 \mathcal{H} (\dot{\Phi}+\dot{\Psi})+ 3 \ddot{\Psi} =\frac{a^2}{2 M_p^2} \left[ \rho_m(\phi) - \bar{\rho}_m + \frac{2 M_p^2}{a^{2}} (\dot{\phi}^2 - \dot{\bar{\phi}}^2) - 2 ( V(\phi) - V(\bar{\phi}) ) \right] \label{eqn:EC} , \\ \fl \mathcal{H}\nabla^2\Phi + \nabla^2\dot{\Psi} = \frac{a^2}{2 M_p^2} \left[ \partial_i \left(\rho_m(\phi)\, u^i \right)+ \frac{M_p^2}{a^{2}} (\partial_i\dot{\phi}\,\partial^i\phi + \dot{\phi}\, \nabla^2\phi)+\frac{2 M_p^2}{a^{2}} \dot{\phi}\,\partial_i\Psi\,\partial^i\phi \right] , \label{eqn:MC}\end{aligned}$$ where primes denote differentiation with respect to $\phi$ and, from now on, latin indices are raised with a Kronecker delta. The evolution of the scalar field $\phi(\eta,{\bf x})$ is given by the Klein–Gordon equation, which, for $|\Phi|, |\Psi|, |u^i| \ll 1$, reads $$\label{eqn:KG} \ddot{\phi} + 2\mathcal{H}\dot{\phi} - \nabla^2\phi + \frac{a^2}{M_p^{2}} U^\prime(\phi) = \nabla(\Phi - \Psi) \cdot \nabla\phi \,,$$ where the effective potential $U(\phi)$ is the sum of the scalar field potential and the field-dependent dark matter energy density, $$U(\phi) \equiv V(\phi) + \rho_{m}(\phi)$$ Finally, we need the equation of motion for the dark matter particles, $$\frac{d^2 x^\mu}{d s^2}+ \Gamma^\mu_{\sigma\beta} \frac{d x^\sigma}{ds} \frac{d x^\beta}{ds} - \frac{m^\prime}{m} \partial^\mu\phi = 0 \,,$$ where the last term comes from the field dependence of the particle mass and makes the trajectories deviate from geodesics. The spatial coordinates of the $a$-th particle satisfy $$\label{eqn:EoM} \ddot{x}_a^i + \dot{x}_a^i \left[ \, \mathcal{H} + \frac{m^\prime}{m}\dot{\phi} \, \right] = - \partial^i \Phi - \frac{m^\prime}{m} \partial^i\phi \,,$$ where $\mathcal{H}$, $\Phi$, and $\phi$ are evaluated at $(\eta,{\bf x}_a(\eta))$. Equations (\[eqn:EC\]), (\[eqn:MC\]), (\[eqn:KG\]) and (\[eqn:EoM\]) form a close system in the variables $\Phi$, $\Psi$, $\phi$ and $x_a^i$’s. As we are interested in sub-horizon scales, there are further approximations which can be made. First of all, in Eq. (\[eqn:EC\]), only the first term on the LHS will be kept, as the others are ${\cal O}(\mathcal{H}^2 \Phi\,,\mathcal{H}^2 \Psi)$. Furthermore, the RHS of Eq. (\[eqn:KG\]) is suppressed by at least an ${\cal O}(\Phi,\Psi)$ factor with respect to the Laplacian on the LHS, and can then be dropped. As a result, the $\Psi$ perturbation is not needed anymore, $\Phi$ can be identified with the gravitational potential, and we can concentrate on the reduced system of Eqs. (\[eqn:EC\]), (\[eqn:KG\]), and (\[eqn:EoM\]) in their approximated forms. To illustrate the role of the scalar field we split it into background, ‘non-relativistic’ and ‘relativistic’ contributions, $$\phi(\eta,{\bf x}) = \bar{\phi}(\eta) + \phi_{NR}(\eta,{\bf x}) + \phi_{R}(\eta,{\bf x}) \;.$$ $\bar{\phi}(\eta)$ is the solution of the background Klein-Gordon equation, $$\ddot{\bar{\phi}} + 2\mathcal{H}\dot{\bar{\phi}} + a^2M_p^{-2}U^\prime(\bar{\phi}) = 0 \label{KGback} \,,$$ while the non-relativistic part $\phi_{NR}(\eta,{\bf x})$ is defined as the solution of the Poisson-like equation $$\nabla^2\phi_{NR} = a^2M_p^{-2}\Delta U^\prime(\bar{\phi} + \phi_{NR})\, ,$$ where $\Delta U(\phi) \equiv U(\phi) - V(\bar{\phi}) - \bar{\rho}_m$. Thus, $\phi_{NR}$ accounts for the response of the scalar field to the localized DM distribution, as induced by the field dependence of the DM particle mass. It behaves similarly to the gravitational potential $\Phi$ and, as we will see, in specific cases it may turn out to be just proportional to the latter. The remaining ‘relativistic’ part solves the equation $$\fl\ddot{\phi}_R + 2\mathcal{H}\dot{\phi}_R - \nabla^2\phi_R + \ddot{\phi}_{NR} + 2\mathcal{H}\dot{\phi}_{NR} + a^2M_p^{-2} [\Delta U^\prime(\bar{\phi} + \phi_{NR} + \phi_{R}) - \Delta U^\prime(\bar{\phi} + \phi_{NR})]=0\,. \label{phirel}$$ The effective mass for $\phi_R$, as read out from Eq. (\[phirel\]) is given by $\Delta U^{\prime\prime}(\bar{\phi}+\phi_{NR})$, which, for the scales of interest for the formation of cosmic structures is much smaller than the corresponding $k^2/a^2$. Moreover, since the time-scale for $\bar{\phi}$ and $\phi_{NR}$ variation is given by $\mathcal{H}$, the $\dot{\phi}_{NR}$ and $\ddot{\phi}_{NR}$ terms in (\[phirel\]) provide ${\cal O}(\mathcal{H}^2)$ sources. As a consequence, the $\phi_R$ component behaves essentially as the superposition of approximately massless plane waves on the scales of interest, with ${\cal O}(a^2\bar{\rho}_m/k^2 M_p^2)\ll1$ amplitudes, and can then be neglected. The relevant dynamics can then be described in terms of the equations $$\begin{aligned} &&\nabla^2\Phi = \frac{a^2}{2 M_p^2} \,\left[\rho_m(\bar{\phi}+\phi_{NR})- \bar{\rho}_m -2 \left(V(\bar{\phi}+\phi_{NR})-V(\bar{\phi})\right)\right]\,, \label{POISSON} \\ &&\nabla^2\phi_{NR} = \frac{a^2}{M_p^{2}}\Delta U^\prime(\bar{\phi} + \phi_{NR}) \,, \label{KGNR}\\ && \ddot{x}_a^i + \left[ \mathcal{H} +\frac{m^{\prime}}{m}(\dot{\bar{\phi}}+\dot{\phi}_{NR})\right] \,\dot{x}_a^i = - \partial^i \Phi - \frac{m^{\prime}}{m} \partial^i (\bar{\phi} + \phi_{NR}) \,, \label{eqn:EoMS}\end{aligned}$$ and the background ones, Eqs. (\[fri1\]), (\[fri2\]), and (\[KGback\]). The equations above extend the standard Newtonian approximation from the pure DM case to the coupled system of DM and quintessence DE. All non-background DE effects are given by $\phi_{NR}$, which solves the Poisson-like eq. (\[KGNR\]). In order to get a flavour of the dynamical content of the system, we specialize to a simple model which was proposed as a solution of the coincidence problem [@CPR; @Pietroni]. The scalar potential and the DM particle mass are both exponentially dependent on $\phi$, $$V(\phi)= \hat{V} \exp(\beta \phi)\,,\,\,\,\,m=\hat{m} \exp(-\lambda \phi)\,,\label{expo}$$ with $\beta , \,\lambda>0$. The model has a solution such that $$\bar{\phi} = \frac{-3}{\lambda+\beta}\;\log a\,,\;\;\;\;\;\;\;\; \Omega_\phi = 1-\Omega_m = \frac{3+\lambda(\lambda+\beta)}{(\lambda+\beta)^2}\, , \label{attractor}$$ which is an [*attractor*]{} in field space if $\beta>(-\lambda+\sqrt{\lambda^2+12})/2$. From the moment the attractor is reached, the energy densities in DM and in DE evolve at a constant ratio depending only on $\lambda$ and $\beta$, thus solving the cosmic coincidence problem. Indeed, the solution (\[attractor\]) implies $$\bar{\rho}_m = m(\bar{\phi}) \bar{n}_m \sim \bar{\rho}_\phi \sim a^{-3(1+w)}\,, \;\;\mbox{with}\;\;\;w=-\frac{\lambda}{\beta+\lambda}\,,$$ where $\bar{n}_m$ is the DM background number density. The negative $w$ leads, if $w<-1/3$, to accelerated expansion of the universe. The $\phi$ dependence of the DM mass modifies the usual scaling $a^{-3}$ of non relativistic matter. Since the mass increases with the expansion, the effect is analogous to that of a fluid with negative equation of state, though DM is, as in the standard scenario, a pressureless fluid made up by non-interacting particles. The impact on the CMB of this model is similar to the one considered in [@Amendola:2000uh]. It was shown there that in order to get results in agreement with the CMB anisotropy spectrum a time-dependent $\lambda$ should be invoked, such that the early time dependence is the standard CDM one, and the attractor behavior of Eq. (\[attractor\]) starts only recently. Since we use the model only for illustrative purposes, we will stick to a constant $\lambda$ in the following. The exponential dependences on $\phi$ in $\Delta U$ and Eq. (\[KGNR\]) imply that also $\phi_{NR}={\cal O}(a^2\bar{\rho}_m/k^2 M_p^2)\ll1$. Then, as dark matter fluctuations approach the non linear regime, they start dominating over those of the effective potential, [*i.e.*]{} $$\Delta U \sim m(\bar{\phi})\left(a^{-3} \sum_a \delta^{(3)}({\bf x }- {\bf x}_a)-\bar{n}_m\right)\,.$$ In this approximation, the RHS of the Poisson’s equation (\[POISSON\]) and that of Eq.  (\[KGNR\]) are proportional to each other, so that $\phi_{NR} = -2\lambda \Phi=O(|\bf{u}|^2)$. Using the properties of the attractor solution, Eq. (\[attractor\]), we get the further reduced system, $$\begin{aligned} \nabla^2\Phi& =& \frac{\hat{m} }{2 M_p^2} \,a^{2-3w}\left(a^{-3}\sum_{a} \delta^{(3)}({\bf x} - {\bf x}_a)-\bar{n}_m\right)\,, \label{eqn:PoissonGrav} \\ \ddot{x}_a^i &+& (1-3 w) \mathcal{H} \,\dot{x}_a^i = - (1+2\lambda^2)\, \partial^i \Phi \label{geomod}\end{aligned}$$ The coupling between DM and DE modifies the equations from the standard ones in two ways. It changes the background evolution through the $w$-terms, and it gives an extra contribution to the force between two DM particles, due to the exchange of a light $\phi$ quantum. The last effect is responsible for the $2 \lambda^2$ correction to the gravitational force in the LHS of Eq. (\[geomod\]). Mildly non-linear evolution =========================== The set of equations (\[eqn:PoissonGrav\]), (\[geomod\]) provides the required extension of the standard Newtonian approximation (e.g. [@Peebles:1980]), describing the non-linear evolution of matter perturbations, to the case where a quintessence scalar field component is present. As a preliminary application, we can study the effect of the scalar field on the DM evolution in the midly non-linear regime, using first-order Lagrangian perturbation theory. Standing to standard notation [@zel], one can look for a time variable $\tau = \tau(\eta)$ by which the equation of motion derived from (\[eqn:EoMS\]) does not contain the Hubble drag term. We can then write an equation for the displacement from the initial (Lagrangian) position $\mathbf{q}$ to the final (Eulerian) position $\mathbf{x} = \mathbf{q} + \mathbf{S}(\tau,\mathbf{q})$ subjected to a force per unit mass proportional to $\partial^i \Phi$. After taking the divergence of the resulting equation of motion, one can use the Poisson Eq. (\[eqn:PoissonGrav\]) to obtain an equation for the [*deformation tensor*]{} $\mathcal{S}_{ij}(\tau,\mathbf{q}) \equiv \partial S^i/\partial q^j$ and then reconstruct the density field $\rho(\tau , \mathbf{x}) = \rho(\mathbf{q})J^{-1}$, where $J=\det(\delta_{ij} + \mathcal{S}_{ij})$ is the determinant of the Jacobian $\partial x^i/\partial q^j$. On the attractor (\[attractor\]), for $\tau = a^n$ and $n=-\frac{1}{2}(1+9w)$, the DM equation of motion is simplified to $$\label{eqn:EoMcv} \frac{d^2 \mathbf{S}}{d\tau^2} = - \frac{(1+2\lambda^2)}{n^2\tau^2\mathcal{H}^2}\nabla_\mathbf{x} \Phi \equiv - \nabla_\mathbf{x} \varphi$$ and the Poisson equation (\[eqn:PoissonGrav\]) becomes $$\label{eqn:LPT_first} \nabla^2_\mathbf{x}\varphi = \left( \frac{3}{2}\Omega_m \frac{(1+2\lambda^2)}{n^2\tau^2} \right) \delta_m (\mathbf{x}(\tau,\mathbf{q})) \equiv b(\tau) \delta \, ,$$ where $\delta = \delta_m(\mathbf{x}(\tau,\mathbf{q}),\tau) = [\rho_m (\tau,\mathbf{x}) - \bar{\rho}_m (\tau)]/ \bar{\rho}_m (\tau)$ is the density contrast. Expanding the resulting equation to first order in the displacement vector $\mathbf{S}$ and making the ansatz $\mathbf{S}(\tau,\mathbf{q}) = f(\tau)\mathbf{\tilde{S}}(\mathbf{q})$ we obtain $$\frac{d^2}{d\tau^2} f(\tau) - b(\tau) f(\tau) = 0 \, ,$$ whose solutions read $$f(\tau) = \tau^p \, , \quad p = \frac{1}{2} \pm \frac{1}{2}\sqrt{1+\frac{24 \Omega_m(1+2\lambda^2)}{(1-9w)^2}}\, .$$ We also get $\nabla_{\bf q} \cdot \mathbf{\tilde{S}}(\mathbf{q}) = - \delta_0$, where $\delta_0$ is the VAMPS density fluctuation linearly extrapolated to the present time. The matter density field follows: $$\rho_m(z,\mathbf{x}) = \frac{\bar{\rho}_m (1+z)^{3(1+w)}} {\prod_{i=1}^{3} \left[ 1 + \lambda_i(\mathbf{q})(1+z)^{-m_+} \right] } \, ,$$ where $\lambda_i(\mathbf{q})$ are the eigenvalues of $\partial \tilde{S}_j/ \partial q^i$ and $$m_\pm = -\frac{1}{4}(1-9w)\left( \mp \sqrt{1+\frac{24 \Omega_m (1+2\lambda^2)}{(1-9w)^2}} \right)$$ (in agreement with [@Amendola:2000uh]): the effect of the scalar field and a non-vanishing coupling constant $\lambda$ is to anticipate (if $m_+ > 1$) or delay (if $m_+ < 1$) the formation of pancakes with respect to scenarios with only DM with constant mass ($m_+ = 1$), depending on the value of $\lambda$ and $\beta$. Because the DM-DE coupling in our specific model affects the motion of DM particles only through a $2\lambda^2$ correction of the gravitational force, the eingenvectors of the deformation tensor $\partial \tilde{S}_j/ \partial q^i$ turn out to be exactly aligned with those of the gravitational force itself. As a consequence, no other scalar field effects appear on small scales. Conclusions =========== In this paper we have set the stage for future studies of non-linear gravitational clustering in scalar field cosmologies, extending the Newtonian approximation to the case of a coupled DM-DE system. The non-linearity in the scalar field may emerge either as a result of its self interactions or by its coupling to DM. It obeys a Poisson-like equation, Eq. (\[KGNR\]), quite similar to that for the gravitational potential $\Phi$. In order to get some analytic insight, we have specialized to the VAMP scenario discussed in [@CPR], in which the DM particle’s mass depends exponentially on the DE field. In this particular model, we found that the scalar field perturbations remain linear inside the horizon, although the coupling plays a non-trivial dynamical role. The mildly non-linear regime has been tested here, using first-order Lagrangian perturbation theory. The DM-DE coupling weakly affects the gravitational field, acting on the onset of pancake formation without any further consequence on the spatial distribution of the DM density field on the scales of interest. The constant coupling $\lambda$ considered here is, however, disfavored by comparison with the CMB anisotropy spectrum. A time- or field-dependent $\lambda$ – smaller in the past and approximately constant in the recent epoch – would take care of this problem, as discussed in [@Amendola:2000uh]. In this case, our results would not change qualitatively. Namely, the scalar field fluctuations would still remain well inside the linear regime and the onset of structure formation would be changed with respect to ordinary quintessence, though of a lesser extent than for a constant $\lambda$. Scalar field non-linearities are, however, not precluded in the VAMP scenario. For instance, modifying the field dependence of the DM mass in Eq. (\[expo\]), the full potential $U(\phi)$ may temporarily develop a spinodal instability (negative second derivative). During this time, perturbations at scales $k/a \le \sqrt{-U^{\prime\prime}}$ grow exponentially and easily exit the linear regime. When the spinodal epoch ends, the system can be described by our equations (\[POISSON\]), (\[KGNR\]), (\[eqn:EoMS\]), assigning the developed non-linearities to the initial conditions for the non-relativistic field fluctuation $\phi_{NR}$. The approach developed in this paper can be straightforwardly applied to the numerical study of models where the quintessence field fluctuates non-linearly, as those proposed in refs. [@Amendola:1999er; @Perrotta:1999am; @Perrotta:2002sw; @Bilic:2001cg; @Armendariz-Picon:2000ah; @Wetterich:2001]. Acknowledgments --------------- We thank F. Perrotta for useful discussions. References {#references .unnumbered} ========== [99]{} S. Perlmutter [*et al.*]{} \[Supernova Cosmology Project Collaboration\], Astrophys. J.  [**517**]{}, 565 (1999); A. G. Riess [*et al.*]{} \[Supernova Search Team Collaboration\], Astron. J.  [**116**]{}, 1009 (1998). D. N. Spergel [*et al.*]{}, astro-ph/0302209 (2003). L. Verde [*et al.*]{}, Mon. Not. R. Astr. Soc., [**335**]{}, 432 (2002). R. R. Caldwell, R. Dave, P. J. Steinhardt Phys. Rev. Lett. [**80**]{}, 1582 (1998). C. Wetterich, Astron. and Astrophy. [**301**]{}, 321; P. G. Ferreira and M. Joyce, Phys. Rev. D [**58**]{}, 023503 (1998). P. J. Peebles and B. Ratra, Astrophys. J.  [**325**]{}, L17 (1988). T. Damour, G. W. Gibbons and C. Gundlach, Phys. Rev. Lett.  [**64**]{}, 123 (1990). J. A. Casas, J. Garcia-Bellido and M. Quiros, Class. Quant. Grav.  [**9**]{}, 1371 (1992); G. W. Anderson and S. M. Carroll, astro-ph/9711288 (1997). D. Comelli, M. Pietroni and A. Riotto, hep-ph/0302080 (2003), to appear on Phys.  Lett.  B. M. Pietroni, Phys. Rev. D [**67**]{} (2003) 103523. M. Gasperini, F. Piazza and G. Veneziano, Phys. Rev. D [**65**]{} (2002) 023508. L. Amendola, Phys. Rev. D [**62**]{}, 043511 (2000). F. Perrotta, C. Baccigalupi and S. Matarrese, Phys. Rev. D [**61**]{}, 023507 (2000). F. Perrotta and C. Baccigalupi, Phys. Rev. D [**65**]{}, 123505 (2002). N. Bilic, G. B. Tupper and R. D. Viollier, Phys. Lett. B [**535**]{}, 17 (2002). C. Armendariz-Picon, V. Mukhanov and P. J. Steinhardt, Phys. Rev. D [**63**]{}, 103510 (2001). C. Wetterich, Phys. Rev. D [**65**]{}, 123512 (2002); Phys. Lett. B [**522**]{}, 5 (2001). C. P. Ma and E. Bertschinger, Astrophys. J. [**455**]{}, 7 (1995). L. Amendola and D. Tocchini-Valentini, Phys. Rev. D [**64**]{}, 043509 (2001). P. J. E. Peebles, [*Large-Scale Structure of the Universe,*]{} Princeton Univ. Press (1980). F. R. Bouchet, S. Colombi, E. Hivon and R. Juszkiewicz, Astron. Astrophys.  [**296**]{}, 575 (1995); P. Catelan, Mon. Not. R. Astr. Soc., [**276**]{}, 115 (1995).

--- author: - 'Thanh Huy Nguyen,  Sylvie Daniel,  Didier Guériot,  Christophe Sintès,  and Jean-Marc Le Caillec, [^1][^2] [^3]' bibliography: - 'IEEEabrv.bib' - 'reference.bib' title: 'Coarse-to-Fine Registration of Airborne LiDAR Data and Optical Imagery on Urban Scenes' --- [Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{} [^1]: T. H. Nguyen is with the Department of Geomatics, Université Laval, Québec City, QC G1V 0A6, Canada, and also with IMT Atlantique, Lab-STICC, UMR CNRS 6285, F-29238 Brest, France (e-mail: thanh.nguyen@imt-atlantique.fr). [^2]: S. Daniel is with the Department of Geomatics, Université Laval, Québec City, QC G1V 0A6, Canada. [^3]: D. Guériot, C. Sintès and J.-M. Le Caillec are with IMT Atlantique, Lab-STICC, UMR CNRS 6285, F-29238 Brest, France.

--- abstract: | We introduce a truncated $M$-fractional derivative type for $\alpha$-differentiable functions that generalizes four other fractional derivatives types recently introduced by Khalil et al., Katugampola and Sousa et al., the so-called conformable fractional derivative, alternative fractional derivative, generalized alternative fractional derivative and $M$-fractional derivative, respectively. We denote this new differential operator by $_{i}\mathscr{D}_{M}^{\alpha,\beta }$, where the parameter $\alpha$, associated with the order of the derivative is such that $ 0 <\alpha<1 $, $\beta>0$ and $ M $ is the notation to designate that the function to be derived involves the truncated Mittag-Leffler function with one parameter. The definition of this truncated $M$-fractional derivative type satisfies the properties of the integer-order calculus. We also present, the respective fractional integral from which emerges, as a natural consequence, the result, which can be interpreted as an inverse property. Finally, we obtain the analytical solution of the $M$-fractional heat equation and present a graphical analysis. .5cm *Keywords*: Alternative Fractional Derivative, Conformable Fractional Derivative, $M$-fractional heat equation, Truncated $M$-Fractional Derivative type, Truncated Mittag-Leffler Function. MSC 2010 subject classifications. 26A06; 26A24; 26A33; 26A42. address: '$^1$ Department of Applied Mathematics, Institute of Mathematics, Statistics and Scientific Computation, University of Campinas – UNICAMP, rua Sérgio Buarque de Holanda 651, 13083–859, Campinas SP, Brazile-mail: [*`ra160908@ime.unicamp.br, capelas@ime.unicamp.br `*]{}' author: - 'J. Vanterler da C. Sousa$^1$' - 'E. Capelas de Oliveira$^1$' bibliography: - 'ref.bib' title: 'A new truncated $M$-fractional derivative type unifying some fractional derivative types with classical properties' --- Introduction ============ The non integer-order calculus or fractional calculus, as it is largely diffused, is as important and ancient as the integer-order calculus, and for many years the scientific community didn’t know it. Currently, there are numerous and important definitions of fractional derivatives types, each one of them with its peculiarity and application [@RHM; @AHMJ; @IP]. Although, to be highlighted only from 1974, after the first international conference on fractional calculus, it has shown to be important and with great applicability in modeling problems, more precisely natural phenomena. For this growth, it was necessary the contribution of famous mathematicians, such as Lagrange, Abel, Euler, Liouville, Riemann, as well as recently Caputo and Mainardi, among others. It is possible to define various integrals and fractional derivatives. Each definition has its own peculiarity and thus makes the fractional calculus fruitful in the sense of theory and applications. We report some types of fractional derivatives that have been introduced so far, among which we mention: Riemann-Liouville, Caputo, Hadamard, Caputo-Hadamard, Riesz, among others [@IP; @ECJT]. Many of these derivatives are defined from the fractional integral in the Riemann-Liouville sense. Recently, Katugampola [@UNT], has presented a new fractional integral unifying six existing fractional integrals, named Riemann-Liouville, Hadamard, Erdélyi-Kober, Katugampola, Weyl and Liouville [@IP; @RCEC]. On the other hand, also recent, Khalil et al. [@RMAM] proposed the so-called compatible fractional derivative of order $\alpha$ with $0<\alpha<1$ in order to generalize the classical properties of calculus. More recently, in 2014, Katugampola [@UNK] has also proposed an alternative fractional derivative with classical properties, which refers to the Leibniz and Newton calculus, similar to the conformable fractional derivative. In 2017, Sousa and Oliveira [@JE], introduced an $M$-fractional derivative involving a Mittag-Leffler function with one parameter [@GKAM] that also satisfies the properties of integer-order calculus. In this sense, we are going to introduce a truncated $M$-fractional derivative type that unifies four existing fractional derivative types mentioned above and which also satisfies the classical properties of integer-order calculus. The study of differential equations has proved very useful over time. One of the main reasons is that the simplest differential equations have the ability to model more complex natural systems [@RHM; @IP; @kilbas]. There are a larger number of application involving differential equations of which we mention: the problem of population dynamics, brachistochronous problem, wave equation, heat equation and others [@RHM; @ZE1]. However, natural systems over time, become more complex and more than differential equations, provides a rough and simplified description of the actual process, it is necessary that new and more refined mathematical tools are presented and studied. In this sense, fractional derivatives are used to propose modeling in order to obtain more precise results in the studies and applications involving differential equations [@kilbas]. Then through the use of properties of a truncated $M$-fractional derivative type, we will present an analytical study of the heat equation and through graph, we will analyze the behavior of the solution in relation to other types of fractional derivatives the so-called local derivatives. This paper is organized as follows: in section 2, our main result, we introduce the concept of truncated $M$-fractional derivative type involving a truncated Mittag-Leffler function, as well as several theorems. Also, we introduce the respective $M$-fractional integral for which we demonstrate the inverse property. In section 3, we present the relationship between a truncated $M$-fractional derivative type, introduced here, and the conformable fractional derivative, generalized and alternative fractional derivative and $M$-fractional derivative. In section 4, we perform an analytical study of the $M$-fractional heat equation in order to obtain the analytical solution and present some graphs. Concluding remarks close the paper. Truncated $M$-fractional derivative type ======================================== In this section, we define a truncated $M$-fractional derivative type and obtain several results that have a great similarity with the results found in the classical calculus. From the definition, we present a theorem showing that this truncated $M$-fractional derivative type is linear, obeys the product rule and the composition of two $\alpha$-differentiable functions, the quotient rule and the chain rule. It is also shown that the derivative of a constant is zero, as well as versions for Rolle’s theorem, the mean value theorem and an extension of the mean value theorem. Further, the continuity of this truncated $M$-fractional derivative type is shown as in integer-order calculus. Also, we introduce the concept of $M$-fractional integral of a $f$ function. From the definition, we shown the inverse theorem. We define the truncated Mittag-Leffler function of one parameter by: $$\label{2A} _{i}\mathbb{E}_{\beta }\left(z\right) =\overset{i}{\underset{k=0}{\sum }}\frac{z^{k}}{\Gamma \left( \beta k+1\right) },$$ with $\beta>0$ and $z\in\mathbb{C}$. From Eq.(\[2A\]), we define a truncated $M$-fractional derivative type that unifies other four fractional derivatives that refer to classical properties of the integer-order calculus. In this work, if a truncated $M$-fractional derivative type of order $\alpha$ as defined in Eq.(\[J\]) of a function $f$ exists, we say that the function $f$ is $\alpha$-differentiable. Thus, let us begin with the following definition, which is a generalization of the usual definition of integer order derivative. Let $f:\left[ 0,\infty \right) \rightarrow \mathbb{R}$. For $0<\alpha <1$ a truncated $M$-fractional derivative type of $f$ of order $\alpha$, denoted by $_{i}\mathscr{D}_{M}^{\alpha,\beta }$, is $$\label{J} _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) :=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon },$$ $\forall t>0$ and $_{i}\mathbb{E}_{\beta }\left(\cdot\right) $, $\beta >0$ is a truncated Mittag-Leffler function of one parameter, as defined in .(\[2A\]). Note that, if $f$ is $\alpha$-differentiable in some open interval $(0,a)$, $a>0$, and $\underset{t\rightarrow 0^{+}}{\lim }\left( _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left(t\right) \right) $ exist, then we have $$_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( 0\right) =\underset{t\rightarrow 0^{+}}{\lim }\left( _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left(t\right) \right).$$ If a function $f:\left[ 0,\infty \right) \rightarrow \mathbb{R}$ is $\alpha$-differentiable for $t_{0}>0$, with $0<\alpha \leq 1$, $\beta>0$, then $f$ is continuous at $t_{0}$. In fact, let us consider the identity $$\label{B} f\left(t_{0}\; _{i}\mathbb{E}_{\beta }\left( \varepsilon t_{0}^{-\alpha }\right) \right) -f\left( t_{0}\right) =\left( \frac{f\left(t_{0}\; _{i}\mathbb{E}_{\beta }\left( \varepsilon t_{0}^{-\alpha }\right) \right) -f\left( t_{0}\right) }{\varepsilon }\right) \varepsilon .$$ Applying the limit $\varepsilon \rightarrow 0$ on both sides of [.(\[B\])]{}, we get $$\begin{aligned} \underset{\varepsilon \rightarrow 0}{\lim }f\left( t_{0}\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t_{0}^{-\alpha }\right) \right) -f\left( t_{0}\right) &=&\underset{\varepsilon \rightarrow 0}{\lim }\left( \frac{f\left(t_{0}\; _{i}\mathbb{E}_{\beta }\left( \varepsilon t_{0}^{-\alpha }\right) \right) -f\left( t_{0}\right) }{\varepsilon }\right) \underset{\varepsilon \rightarrow 0}{\lim }\varepsilon \notag \\ &=&_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) \underset{\varepsilon \rightarrow 0}{\lim }\varepsilon \notag \\ &=&0.\end{aligned}$$ Then, $f$ is continuous at $t_{0}$. Using the definition of truncated Mittag-Leffler function of one parameter, we have $$\label{A} f\left(t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) \right) =f\left( t \overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( \beta k+1\right) }\right) .$$ Applying the limit $\varepsilon \rightarrow 0$ on both sides of Eq.(\[A\]) and since $f$ is a continuous function, we have $$\begin{aligned} \underset{\varepsilon \rightarrow 0}{\lim }f\left(t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right)\right) &=&\underset{\varepsilon \rightarrow 0}{\lim }f\left( t\overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( \beta k+1\right) }\right) \notag \\ &=&f\left( t\underset{\varepsilon \rightarrow 0}{\lim }\overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( \beta k+1\right) }\right) . \end{aligned}$$ Further, we have $$\begin{aligned} \label{C1} t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) &=&t\overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( \beta k+1\right) }\notag \\ &=& t+\frac{\varepsilon t^{1-\alpha }}{\Gamma \left( \beta +1\right) }+\frac{t\left( \varepsilon t^{-\alpha }\right) ^{2}}{\Gamma \left( 2\beta +1\right) }+\frac{t\left( \varepsilon t^{-\alpha }\right) ^{3}}{\Gamma \left( 3\beta +1\right) }+\cdot \cdot \cdot +\frac{t\left( \varepsilon t^{-\alpha }\right) ^{i}}{\Gamma \left( i\beta +1\right) }. \notag \\\end{aligned}$$ Applying the limit $\varepsilon \rightarrow 0$ on both sides of .(\[C1\]), we have $$\underset{\varepsilon \rightarrow 0}{\lim }\overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( \beta k+1\right) }=1.$$ In this way, we conclude that $$\begin{aligned} \underset{\varepsilon \rightarrow 0}{\lim }f\left(t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) \right) &=& f\left(t\right) . \end{aligned}$$ Here, we present the theorem that encompasses the main classical properties of integer order calculus. For the chain rule, it is verified through an example, as we will see next. We will do here, only the demonstration of the chain rule, for other items, follow the same steps of Theorem 2 found in the paper by Sousa and Oliveira [@JE]. \[JOSE\] Let $0<\alpha \leq 1$, $\beta>0$, $a,b\in\mathbb{R}$ and $f, g$ $\alpha$-differentiable, at a point $t>0$. Then: 1. $ _{i}\mathscr{D}_{M}^{\alpha,\beta }\left( af+bg\right) \left( t\right) =a\; _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) +b\; _{i}\mathscr{D}_{M}^{\alpha,\beta }g\left( t\right) $. 2. $ _{i}\mathscr{D}_{M}^{\alpha,\beta }\left( f\cdot g\right) \left( t\right) =f\left( t\right)\; _{i}\mathscr{D}_{M}^{\alpha,\beta }g\left( t\right) +g\left( t\right)\; _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)$. 3. $\displaystyle _{i}\mathscr{D}_{M}^{\alpha,\beta }\left( \frac{f}{g}\right) \left( t\right) =\frac{g\left(t\right)\; _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) -f\left( t\right) \;_{i}\mathscr{D}_{M}^{\alpha,\beta }g\left( t\right) }{\left[ g\left( t\right) \right] ^{2}} $. 4. $_{i}\mathscr{D}_{M}^{\alpha,\beta }\left( c\right) =0 $, where $f(t)=c$ is a constant. 5. If $f$ is differentiable, then $_{i}\mathscr{D}_{M}^{\alpha,\beta } \left( f\right)(t)= \displaystyle\frac{t^{1-\alpha}}{\Gamma \left(\beta +1\right) } \frac{df(t)}{dt}. $ From [.(\[C1\])]{}, we have $$\begin{aligned} t\;_{i}\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) &=&t+\frac{\varepsilon t^{1-\alpha }}{\Gamma \left(\beta +1\right) }+O\left( \varepsilon^{2}\right),\end{aligned}$$ and introducing the following change, $$h=\varepsilon t^{1-\alpha }\left( \frac{1}{\Gamma \left(\beta +1\right) }+O\left( \varepsilon \right) \right) \Rightarrow \varepsilon =\frac{h}{t^{1-\alpha }\left( \frac{1}{\Gamma \left(\beta +1\right) }+O\left( \varepsilon \right) \right) },$$ we conclude that $$\begin{aligned} _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) &=&\underset{\varepsilon \rightarrow 0}{\lim }\frac{\displaystyle\frac{f\left(t+h\right) -f\left( t\right) }{ht^{\alpha -1}}}{\frac{1}{\Gamma \left(\beta +1\right) }\left(1+{\Gamma \left(\beta +1\right) } O\left( \varepsilon \right) \right) } \notag \\ &=&\frac{t^{1-\alpha}}{\Gamma \left(\beta +1\right) }\underset{\varepsilon \rightarrow 0}{\lim }\frac{\displaystyle\frac{f\left( t+h\right) -f\left( t\right) }{h}}{1+{\Gamma \left(\beta +1\right) } O\left( \varepsilon \right) } \notag \\ &=&\frac{t^{1-\alpha}}{\Gamma \left(\beta +1\right) }\frac{df\left( t\right) }{dt},\end{aligned}$$ with $\beta>0$ and $t>0$. 6. $_{i}\mathscr{D}_{M}^{\alpha,\beta }\left(f\circ g\right)(t)=f'(g(t)) _{i}\mathscr{D}_{M}^{\alpha,\beta }g(t)$, for $f$ differentiable at $g(t)$. Now, it is necessary to know if, in addition to the previous Theorem 2 that contains important properties similar to integer-order calculus, this truncated $M$-fractional derivative type Eq.(\[J\]) also has important theorems related to the classical calculus. We shall now see that: the Rolle’s theorem, the mean value theorem and its extension coming from the integer-order calculus can be extended to $\alpha$-differentiable functions, i.e., that admit truncated $M$-fractional derivative as introduced in Eq.(\[J\]). Let $a>0$, and $f:\left[ a,b\right] \rightarrow \mathbb{R}$ be a function with the properties: 1. $f$ is continuous on $[a,b]$. 2. $f$ is $\alpha$-differentiable on $(a,b)$ for some $\alpha\in(0,1)$. 3. $f(a)=f(b)$. Then, $\exists c\in(a,b)$, such that $_{i}\mathscr{D}^{\alpha,\beta}_{M}f(c)=0$, with $\beta>0$. Since $f$ is continuous on $[a,b]$ and $f(a)=f(b)$, there exist $c\in(a,b)$, at which the function has a local extreme. Then, $$\begin{aligned} _{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( c\right) &=&\underset{\varepsilon \rightarrow 0^{-}}{\lim }\frac{f\left(c\;_{i}\mathbb{E}_{\beta }\left( \varepsilon c^{-\alpha }\right) \right) -f\left( c\right) }{\varepsilon }=\underset{\varepsilon \rightarrow 0^{+}}{\lim }\frac{f\left( c\;_{i}\mathbb{E}_{\beta }\left( \varepsilon c^{-\alpha }\right) \right) -f\left( c\right) }{\varepsilon }.\end{aligned}$$ But, the two limits have opposite signs. Hence, $_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( c\right) =0$. The proof of Theorem 4 and Theorem 5, will be omitted, but follow the same reasoning of the respective theorems demonstrated in Sousa and Oliveira [@JE]. Let $a>0$ and $f:\left[ a,b\right] \rightarrow \mathbb{R}$ be a function with the properties: 1. $f$ is continuous on $[a,b]$. 2. $f$ is $\alpha$-differentiable on $(a,b)$ for some $\alpha\in(0,1)$. Then, $\exists c\in(a,b)$, such that $$_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( c\right) =\frac{f\left( b\right) -f\left( a\right) }{\displaystyle\frac{b^{\alpha }}{\alpha }-\frac{a^{\alpha }}{\alpha }},$$ with $\beta>0$. Let $a>0$, and $f,g:\left[ a,b\right] \rightarrow \mathbb{R}$ functions that satisfy: 1. $f, g$ are continuous on $[a,b]$. 2. $f, g$ are $\alpha$-differentiable for some $\alpha\in(0,1)$. Then, $\exists c\in(a,b)$, such that $$\frac{_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( c\right) }{_{i}\mathscr{D}_{M}^{\alpha,\beta }g\left( c\right) }=\frac{f\left( b\right) -f\left( a\right) }{g\left( b\right) -g\left( a\right) },$$ with $\beta>0$. Let $\alpha\in(n,n+1]$, for some $n\in\mathbb{N}$, $\beta>0$ and $f$ $n$-differentiable for $t>0$. Then the $\alpha$-fractional derivative of $f$ is defined by $$_{i}\mathscr{D}_{M}^{\alpha,\beta;n }f\left( t\right) :=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f^{(n)}\left(t\; _{i}\mathbb{E}_{\beta }\left( \varepsilon t^{n-\alpha }\right) \right) -f^{(n)}\left( t\right) }{\varepsilon },$$ since the limit exist. From Definition 2 and the chain rule, that is, from item 5 of Theorem 2, by induction on $n$, we can prove that $_{i}\mathscr{D}_{M}^{\alpha,\beta;n }f\left( t\right)= \displaystyle\frac{t^{n+1-\alpha}}{\Gamma \left(\beta +1\right) } f^{(n+1)}(t)$, $\alpha\in(n,n+1]$ and $f$ is $(n+1)$-differentiable for $t>0$. Now, we know that: this truncated $M$-fractional derivative type Eq.(\[J\]) has a corresponding $M$-fractional integral. Then, we will present the definition and a theorem that corresponds to the inverse property. For other results involving integrals, one can consult [@JE; @OSEN]. Let $a\geq 0$ and $t\geq a$. Also, let $f$ be a function defined in $(a,t]$ and $0<\alpha <1$. Then, the $M$-fractional integral of order $\alpha$ of function $f$ is defined by [@JE] $$\label{ZA} _{M}\mathcal{I}_{a}^{\alpha,\beta }f\left( t\right) ={\Gamma \left(\beta +1\right) }\int_{a}^{t}\frac{f\left( x\right) }{x^{1-\alpha }}dx,$$ with $\beta>0$. Let $a\geq 0$ and $0<\alpha < 1$. Also, let $f$ be a continuous function such that exist $_{M}\mathcal{I}_{a}^{\alpha,\beta }f$. Then $$_{i}\mathscr{D}_{M}^{\alpha,\beta }(_{M}\mathcal{I}_{a}^{\alpha,\beta }f\left( t\right)) =f(t),$$ with $t\geq a$ and $\beta>0$. In fact, using the chain rule as seen in \[JOSE\], we have $$\begin{aligned} \label{Z1} _{i}\mathscr{D}_{M}^{\alpha,\beta }\left( _{M}\mathcal{I}_{a}^{\alpha,\beta }f\left( t\right) \right) &=&\frac{t^{1-\alpha }}{\Gamma \left( \beta +1\right) }\frac{d}{dt}(_{M}\mathcal{I}_{a}^{\alpha,\beta }f\left( t\right)) \notag \\ &=&\frac{t^{1-\alpha }}{\Gamma \left( \beta +1\right) }\frac{d}{dt}\left( {\Gamma \left( \beta +1\right) }\int_{a}^{t}\frac{f\left( x\right) }{x^{1-\alpha }}dx\right) \notag \\ &=&\frac{t^{1-\alpha }}{\Gamma \left( \beta +1\right) }\left( \frac{\Gamma \left( \beta +1\right) }{t^{1-\alpha }}f\left( t\right) \right) \notag \\ &=&f\left( t\right) .\end{aligned}$$ With the condition $f(a)=0$, by Theorem 6, that is, Eq.(\[Z1\]), we have $_{M}\mathcal{I}_{a}^{\alpha,\beta } \left[ {_{i}\mathscr{D}_{M}^{\alpha,\beta }}f(t)\right]=f(t)$. Relation with other fractional derivatives types ================================================ In this section, we will discuss the relationship between the fractional conformable derivative proposed by Khalil et al. [@RMAM], the alternative fractional derivative and the generalized alternative fractional derivative proposed by Katugampola [@UNK] and the $M$-fractional derivative proposed by Sousa and Oliveira [@JE], with our truncated $M$-fractional derivative type. Khalil et al. [@RMAM] proposed a definition of a fractional derivative, called conformable fractional derivative that refers to the classical properties of integer order calculus, given by $$\label{K1} f^{(\alpha) }\left( t\right) =\underset{\varepsilon \rightarrow 0}{\lim }\frac{f\left( t+\varepsilon t^{1-\alpha}\right) -f\left( t\right) }{\varepsilon },$$ with $\alpha\in(0,1)$ and $t>0$. In 2014, Katugampola [@UNK] proposed another definition of a fractional derivative, called an alternative fractional derivative which also refers to the classical properties of integer-order calculus, given by $$\label{K2} \mathcal{D}^{\alpha }f\left( t\right) =\underset{\varepsilon \rightarrow 0}{\lim }\frac{f\left( te^{\varepsilon t^{-\alpha }}\right) -f\left( t\right) }{\varepsilon },$$ with $\alpha\in(0,1)$ and $t>0$. In the same paper, Katugampola [@UNK] by means of a truncated exponential function, that is, $_{k}e^{x}$, proposed another generalized fractional derivative, given by $$\label{K3} \mathcal{D}^{\alpha }_{k}f\left( t\right) =\underset{\varepsilon \rightarrow 0}{\lim }\frac{f\left(_{k}e^{\varepsilon t^{-\alpha }}t\right) -f\left( t\right) }{\varepsilon },$$ with $\alpha\in(0,1)$ and $t>0$. Recently, Sousa and Oliveira [@JE] introduced the $M$-fractional derivative $\mathscr{D}_{M}^{\alpha,\beta}$ where the parameter $\beta>0$ and $M$ is the notation to designate that the function to be derived involves the Mittag-Leffler function of one parameter, given by $$\label{K4} \mathscr{D}_{M}^{\alpha,\beta }f\left( t\right) :=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left( t\;\mathbb{E}_{\beta }\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon },$$ with $\alpha\in(0,1)$ and $t>0$. It is clear that our definition of truncated $M$-fractional derivative type Eq.(\[J\]) is more general than the fractional derivatives Eq.(\[K1\]), Eq.(\[K2\]), Eq.(\[K3\]) and Eq.(\[K4\]). We will now study particular cases involving such fractional derivatives. Choosing $\beta=1$ and applying the limit $i\rightarrow 0$ on both sides of Eq.(\[J\]), we have $$_{1}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;_{1}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon }.$$ But, it is know that $$_{1}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) =\overset{1}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( k+1\right) }=1+\varepsilon t^{-\alpha }.$$ Thus, we conclude that $$_{1}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t+\varepsilon t^{1-\alpha }\right) -f\left( t\right) }{\varepsilon }=f^{(\alpha)}(t),$$ which is exactly the conformable fractional derivative Eq.$(\ref{K1})$. Choosing $\beta=1$ and applying the limit $i\rightarrow \infty$ on both sides of Eq.(\[J\]), we have $$_{\infty}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;_{\infty}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon }.$$ But, as we have $$_{\infty}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) =\overset{\infty}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( k+1\right) }=e^{\varepsilon t^{-\alpha }},$$ we conclude that $$_{\infty}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t e^{\varepsilon t^{-\alpha }}\right) -f\left( t\right) }{\varepsilon }=\mathcal{D}^{\alpha }f\left( t\right),$$ which is exactly the alternative fractional derivative, Eq.$(\ref{K2})$. Choosing $\beta=1$ in Eq.(\[J\]), we have $$_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;_{i}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon }.$$ Remembering that $$_{i}\mathbb{E}_{1}\left( \varepsilon t^{-\alpha }\right) =\overset{i}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( k+1\right) }=e_{i}^{\varepsilon t^{-\alpha }},$$ we have $$_{i}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(e_{i}^{\varepsilon t^{-\alpha }}t\right) -f\left( t\right) }{\varepsilon }=\mathcal{D}_{i}^{\alpha }f\left( t\right),$$ exactly the generalized fractional derivative, Eq.$(\ref{K3})$. Finally, applying the limit $i\rightarrow \infty$ on both sides of Eq.(\[J\]), we have $$_{\infty}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;_{\infty}\mathbb{E}_{\beta}\left( \varepsilon t^{-\alpha }\right) \right) -f\left( t\right) }{\varepsilon },$$ since $$_{\infty}\mathbb{E}_{\beta}\left( \varepsilon t^{-\alpha }\right) =\overset{\infty}{\underset{k=0}{\sum }}\frac{\left( \varepsilon t^{-\alpha }\right) ^{k}}{\Gamma \left( k+1\right) }=\mathbb{E}_{\beta}\left( \varepsilon t^{-\alpha }\right),$$ we conclude that $$_{\infty}\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right)=\underset{\varepsilon \rightarrow 0}{\lim }\frac {f\left(t\;\mathbb{E}_{\beta}\left( \varepsilon t^{-\alpha }\right)\right) -f\left( t\right) }{\varepsilon }=\mathscr{D}_{M}^{\alpha,\beta }f\left( t\right),$$ exactly the $M$-fractional derivative, Eq.$(\ref{K4})$. Application =========== In this section, we obtain the solution of the heat equation using a truncated $M$-fractional derivative type with $0<\alpha <1 $ and present some graphs about the behavior of the solution. Consider the heat equation in one dimension given by $$\frac{\partial u\left( x,t\right) }{\partial t}=k\frac{\partial ^{2}u\left( x,t\right) }{\partial x^{2}},\text{ }0<x<L,\text{ }t>0,$$ where $k$ is a positive constant. Using a $M$-fractional derivative type, we propose an $M$-fractional heat equation given by $$\label{B} \frac{\partial ^{\alpha }u\left( x,t\right) }{\partial t^{\alpha }}=k\frac{\partial ^{2}u\left( x,t\right) }{\partial x^{2}},\text{ }0<x<L,\text{ }t>0,$$ where $0<\alpha <1$ and with the initial condition and boundary conditions given by $$\begin{aligned} \label{1} u\left( 0,t\right) &=&0\,, t\geq 0, \\\notag u\left( L,t\right) &=&0\,, t\geq 0, \\\notag u\left( x,0\right) &=&f\left( x\right) \,, 0\leq x\leq L.\notag\end{aligned}$$ We start, considering the so-called $M$-fractional linear differential equation with constant coefficients $$\label{C} \frac{\partial ^{\alpha }v\left( x,t\right) }{\partial t^{\alpha }}\pm \mu ^{2}v\left( x,t\right) =0,$$ where $\mu^{2}$ is a positive constant. Using the item 5 in Theorem \[JOSE\], the [Eq.(\[B\])]{} can be written as follows $$\frac{t^{1-\alpha }}{\Gamma \left( \beta +1\right) }\frac{dv\left( x,t\right) }{dt}\pm \mu ^{2}v\left( x,t\right) =0,$$ whose solution is given by $$\label{D} v\left( t\right) =ce^{\pm \Gamma \left( \beta +1\right) \frac{\mu ^{2}t^{\alpha }}{\alpha }},$$ with $0<\alpha<1$ and $\beta>0$. Now, we will use separation of variables method to obtain the solution of the $M$-fractional heat equation. Then, considering $u\left( x,t\right) =P\left( x\right) Q\left( t\right) $ and replacing in Eq.(\[B\]), we get $$\frac{d^{\alpha }}{dt^{\alpha }}Q\left( t\right) P\left( x\right) =k\frac{d^{2}}{dx^{2}}P\left( x\right) Q\left( t\right)$$ which implies $$\label{E} \frac{1}{kQ\left( t\right) }\frac{d^{\alpha }}{dt^{\alpha }}Q\left( t\right) =\frac{1}{P\left( x\right) }\frac{d^{2}}{dx^{2}}P\left( x\right) =\xi .$$ From Eq.(\[E\]), we obtain a system of differential equations, given by $$\label{F} \frac{d^{\alpha }}{dt^{\alpha }}Q\left( t\right) -k\xi Q\left( t\right) =0$$ and $$\label{G} \frac{d^{2}}{dx^{2}}P\left( x\right) -\xi P\left( x\right) =0.$$ First, let’s find the solution of Eq.(\[G\]). For this, we must study three cases, that is, $\xi =0,$ $\xi =-\mu ^{2}$ e $\xi =\mu ^{2}.$ Case 1: $\xi =0$. Substituting $ \xi = 0 $ into Eq.(\[G\]), we have $$\frac{d^{2}}{dx^{2}}P\left( x\right) -\xi P\left( x\right) =0,$$ whose solution is given by $P\left( x\right) =c_{1}x+c_{2}$, with $c_{1}$ and $c_{2}$ arbitrary constant. Using the initial conditions given by Eq.(\[1\]), we obtain that $ c_{1}=c_{2}=0$. Like this, $P\left( x\right) =0$, which implies $u(x,t)=0$ trivial solution. Case 2: $\xi =-\mu ^{2}$. Substituting $\xi =-\mu ^{2}$ into Eq.(\[G\]), we get $$\frac{d^{2}}{dx^{2}}P\left( x\right) +\mu ^{2}P\left( x\right) =0,$$ whose solution is given by $P\left( x\right) =c_{2}\sin \left( \mu x\right) +c_{1}\cos \left( \mu x\right)$, with $c_{1}$ and $c_{2}$ arbitrary constant. Using the initial conditions Eq.(\[1\]), we obtain $c_{1}=0$ and $0=c_{2}\sin \left( \mu x\right)$ which implies that $\mu =\frac{n\pi }{L},$ with $n=1,2,...$. Then, we obtain $$P_{n}\left( x\right) =a_{n}\sin \left( \frac{n\pi x}{L}\right) \text{ and } \mu =\frac{n\pi }{L}.$$ Case 3: $\xi =\mu ^{2}$. Substituting $\xi =\mu ^{2}$ into Eq.(\[G\]), we get $$\frac{d^{2}}{dx^{2}}P\left( x\right) -\mu ^{2}P\left( x\right) =0$$ whose solution is given by $P\left( x\right) =c_{1}$ $e^{\mu x}+c_{2}$ $e^{-\mu x}=A\cosh \left( \mu x\right) +B\sinh \left( \mu x\right)$, with $c_{1}$, $c_{2}$, $A$, $B$ arbitrary constant. Using the boundary conditions Eq.(\[1\]), we have $A=0$ and $0=B\sinh \left( \mu x\right)$. As $\lambda =-\mu ^{2}<0$ and $\lambda L\neq 0$ then $\sinh \left( \mu x\right) \neq 0.$ Like this, we get $B=0$ and then $P_{n}\left( x\right) =0$, which implies $u(x,t)=0$, trivial solution. Therefore, the solution of Eq.(\[G\]) is given by $$\label{H} P_{n}\left( x\right) =a_{n}\sin \left( \frac{n\pi x}{L}\right) \text{ and} \mu =\frac{n\pi }{L}.$$ Using the Eq.( \[C\]) and Eq.(\[D\]), we have $$\label{I} Q_{n}\left( t\right) =b_{n}\exp \left( -\Gamma \left( \beta +1\right) \left( \frac{n\pi }{L}\right) ^{2}\frac{k}{\alpha }t^{\alpha }\right),$$ where $b_{n}$ are constant coefficients. So, using the Eq.(\[H\]) and Eq.(\[I\]), the partial solutions of Eq.(\[B\]), is given by $$u_{\beta}\left( x,t\right) =\underset{n=1}{\overset{\infty }{\sum }}c_{n}\sin \left( \frac{n\pi x}{L}\right) \exp \left( -\Gamma \left( \beta +1\right) \left( \frac{n\pi }{L}\right) ^{2}\frac{k}{\alpha }t^{\alpha }\right) .$$ Using Eq.(\[1\]), we get $$u\left( x,0\right) =f\left( x\right) =\underset{n=1}{\overset{\infty }{\sum }}c_{n}\sin \left( \frac{n\pi x}{L}\right)$$ which provides $c_{n}$ through $$c_{n}=\frac{2}{L}\int_{0}^{L}f\left( x\right) \sin \left( \frac{n\pi x}{L}\right) dx.$$ So, we conclude that the solution of $M$-fractional heat equation Eq.(\[B\]), satisfying the conditions Eq.(\[1\]), is given by $$\label{fig1} u_{\beta}\left( x,t\right) =\underset{n=1}{\overset{\infty }{\sum }}\sin \left( \frac{n\pi x}{L}\right) \exp \left( -\Gamma \left( \beta +1\right) \left( \frac{n\pi }{L}\right) ^{2}\frac{k}{\alpha }t^{\alpha }\right) \left( \frac{2}{L}\int_{0}^{L}f\left( x\right) \sin \left( \frac{n\pi x}{L}\right) dx\right) .$$ Taking the limit $\beta \rightarrow 1$ in the last equation and using Eq.(\[J\]), we have $$\label{fig2} u\left( x,t\right) =\underset{n=1}{\overset{\infty }{\sum }}\sin \left( \frac{n\pi x}{L}\right) \exp \left( -\left( \frac{n\pi }{L}\right) ^{2}\frac{k}{\alpha }t^{\alpha }\right) \left( \frac{2}{L}\int_{0}^{L}f\left( x\right) \sin \left( \frac{n\pi x}{L}\right) dx\right) ,$$ which is exactly the solution of the fractional heat equation proposed by Çenesiz et al. [@YCA]. On the other hand, taking the limit $\beta \rightarrow 1$ and $\alpha \rightarrow 1$, using Eq.(\[J\]), we recover the solution of heat equation of integer order. $$\label{fig3} u\left( x,t\right) =\underset{n=1}{\overset{\infty }{\sum }}\sin \left( \frac{n\pi x}{L}\right) \exp \left( -\left( \frac{n\pi }{L}\right) ^{2}kt\right) \left( \frac{2}{L}\int_{0}^{L}f\left( x\right) \sin \left( \frac{n\pi x}{L}\right) dx\right) .$$ Next, we will present some plots by choosing values for the parameters $\beta$ and $\alpha$, to see the behavior of the solution presented in Eq (\[J\]). The graphics were plotted using MATLAB 7:10 software (R2010a). For the elaboration of the following plots, we choose the function $f(x)=50x(1-x)$ and for each fixed $\beta$, we vary the $\alpha$ parameter. ![Analytical solution of the $M$-fractional heat equation [Eq.(\[fig1\])]{}. We consider the values $ \beta = 0.5 $, $L = 1$, $k = 0.003$ and at time $t = 150$.](calor1.eps "fig:"){width="14.0cm"} \[fig:eryt1\] ![Analytical solution of the $M$-fractional heat equation [Eq.(\[fig1\])]{}. We take the values $ \beta = 1.0 $, $L = 1$, $k = 0.003$ and at time $t = 150$.](calor2.eps "fig:"){width="14.0cm"} \[fig:eryt1\] ![Analytical solution of the $M$-fractional heat equation [Eq.(\[fig1\])]{}. We chose the values $ \beta = 2.0 $, $L = 1$, $k = 0.003$ and at time $t = 150$.](calor3.eps "fig:"){width="14.0cm"} \[fig:eryt1\] Concluding remarks ================== We introduced a new truncated $M$-fractional derivative type for $\alpha$-differentiable functions and consequently its $M$-fractional integral, we obtained important results with respect to the properties of the integer-order derivative. For a class of $\alpha$-differentiable functions, in the context of fractional derivatives, we conclude that this truncated $M$-fractional derivative type proposed here behaves well with respect to the classical properties of integer-order calculus. Using the truncated Mittag-Leffler function, it was possible to introduce a truncated $M$-fractional derivative type associated with $\alpha$-differentiable functions and consequently we obtained a very important relation with the fractional derivatives mentioned in the paper as seen in section 3. We conclude that the presented results contain as particular cases the derivatives proposed by Khalil et al. [@RMAM], Katugampola [@UNK] and Sousa and Oliveira [@JE]. In addition, using our truncated $M$-fractional derivative type we performed and discussed the analytical solution of the heat equation.

--- abstract: 'We present our recent results for the $\rho$ and $\sigma$ mesons considered as resonances in pion-pion scattering in a thermal bath. We use chiral perturbation theory to order $p^4$ for the low energy behaviour, then extend the analysis via the unitarization method of the Inverse Amplitude into the resonance region. The width of the rho broadens about twice the amount required by phase space considerations alone, its mass staying practically constant up to temperatures of order 150 MeV. The sigma meson behaves in accordance to chiral symmetry restoration expectations.' address: | Deptos. Física Teórica I y II, Univ. Complutense, 28040 Madrid, Spain.\ $*$ Also Dip. di Fisica, Universita’ degli Studi and INFN, Firenze, Italy. author: - 'F.Llanes-Estrada, A. Dobado, A. Gómez Nicola, J.R.Peláez\*' title: '$\rho$ and $\sigma$ Mesons in Unitarized Thermal $\chi$PT' --- ![ Various existing calculations on the thermal (at about T=150 MeV) vector meson mass and width. The numbers refer to the bibliography.[]{data-label="estanque"}](rhofiniteT.eps){height=".3\textheight"} Triggered by the possibility of exploiting the dilepton spectrum as a signal of the Quark and Gluon Plasma in Relativistic Heavy Ion Collisions, there have been numerous studies of the thermal behaviour of the $\rho$ resonance in a hot hadron medium. Some of these theoretical approaches are summarized in figure (\[estanque\]). The wide spread of predictions signals a strong model dependence in the various calculations, although a general (not universal) trend points to a natural broadening of the width due to a larger available phase space. In this work we apply chiral perturbation theory [@GaLe1984]($\chi$PT) complemented with the Inverse Amplitude Method [@JRPAD] (IAM) to obtain predictions guided only by the principles of chiral symmetry and unitarity. These combined methods provide outstanding fits to low energy pion scattering data [@JRPAD] thus fixing the low energy constants of the chiral lagrangian. We use this same constants in our thermal amplitudes and all finite temperature results fall as predictions. The starting point is a calculation of the thermal pion scattering amplitude to one loop in chiral perturbation theory, which includes calculating pion loops and tadpoles (where temperature dependence is included through Matsubara sums in the imaginary time formalism) and polynomial counterterms with arbitrary coefficients (the low energy constants) which are fitted to known pion scattering phase shifts. Since the thermal bath induces a preferred reference frame, our results [@ourplb] can be best written in terms of the $\pi_+ \ \pi_- \longrightarrow \pi_0 \ \pi_0$ scattering amplitude $A({\bf S},{\bf T},{\bf U},\beta)$ which depends on the four vectors ${\bf S}=p_{\pi_+}+p_{\pi_-}$... generalizing the Mandelstam variables and allowing us to generate the other possible pion scattering amplitudes by crossing symmetry. At finite temperature and for $s>4m_\pi^2$ and below other thresholds, the partial wave projections obtained from our amplitude satisfy the unitarity relation perturbatively, that is $$\label{unitarity} {\rm Im}\ a_2(s)=0 \ , \ \ {\rm Im}\ a_4(s;\beta)=\sigma_\beta({\bf S}_0) \arrowvert a_2(s) \arrowvert^2 \ , \ {\bf S}_0>2m_\pi \ ,$$ with a thermal phase space $$\label{phasespace} \sigma_{\beta}(E)=\sqrt{1-\frac{4m_\pi^2}{E^2}}\left[ 1+ \frac{2}{{\rm exp}(\beta \arrowvert E \arrowvert /2)-1} \right] \ .$$ The temperature is of order $p$, $M_\pi$ in the chiral counting, and thus for small temperatures and momenta all corrections scale at least as the fourth power of the small quantity $M_\pi/(4\pi f_\pi)$, providing very good accuracy in this regime. Alas, the chiral expansion is intrinsically a polynomial with chiral logarithms (from pion loops) providing the imaginary parts in (\[unitarity\]) only perturbatively, but violating severely exact unitarity already around the $\rho$ resonance region. Still, from the behaviour of the phase shifts at low energy, we find a model independent thermal enhancement of the low energy interactions, conserving their attractive or repulsive nature. This effect is consistent with a thermal increase of the rho width and an almost constant mass[@ourplb]. To extend this approach to momenta and temperatures beyond the reach of one loop $\chi$PT, several unitarization methods have been proposed. Here we employ the well tested IAM (see [@littlebook] for a thorough introduction). In terms of the $\chi$PT momentum expansion, one approximates $a(s)$ by $$\label{IAM} a^{\rm IAM}(s;\beta)=\frac{a_2^2(s)}{a_2(s)-a_4(s;\beta)}$$ satisfying $${\rm Im}\ a^{\rm IAM}=\sigma_\beta(s) \arrowvert a^{\rm IAM}(s; \beta) \arrowvert^2 \ .$$ One can think of a particular unitarization method giving an exact treatment of the imaginary part of the inverse amplitude but just an approximation to the real part; the IAM uses the $O(p^4)$ $\chi$PT amplitude for the approximation and can be considered as the \[1,1\] Padé approximant for the $\chi$PT series in inverse powers of $(4\pi f_\pi)^2$. Further, the rational formula (\[IAM\]) allows the treatment of one pole resonance in each $IJ$ channel which can be tracked down in the second Riemann sheet for complex $s$. If we think of the doubly subtracted dispersion relation satisfied by $a(s)$, the left cut corresponding to $t$-channel pion loops is only approximately considered, whereas the right $s$-channel cut is exactly taken into account. This induces a small crossing violation, but its impact on the resonance region is numerically controlled because of the large distance of this cut to the $\rho$ or $\sigma$ poles in the complex plane. ![Zero and finite temperature pion scattering phase shifts in the $\rho$ channel.[]{data-label="phaseshifts"}](figT1.eps){height=".3\textheight"} ![How dinamically generated resonance parameters evolve with the temperature.[]{data-label="polos"}](poles.eps){height=".3\textheight"} Figure (\[phaseshifts\]) displays the IAM $\rho$ channel phase shift in pion scattering, at zero temperature (compared to experimental data) and the finite temperature prediction. The results are consistent with the evolution of the $\sigma$ and $\rho$ poles in the complex plane shown in fig.(\[polos\]) (the analytical extension of the loop and tadpole integrals to the complex plane is described in [@ourprc]). The behaviour of the $\sigma$ hints at chiral symmetry restoration: observe how its mass decreases as the temperature raises and how its width, initially increasing, dramatically drops when approaching the relatively stable two pion threshold. On the other hand, the width of the $\rho$ increases in all the considered temperature range. This calculation, performed ab initio in the chirally broken phase, does not have direct access to the phase transition, but serves to pin down models which do have this capability providing them with a general behaviour at low temperatures. Since in our approach resonances are not explicitly included, but arise dynamically, we can predict the temperature evolution of their masses and couplings. If we specifically concentrate on $\rho$ meson properties, replotted for convenience in figure \[rho\], we find that the thermal width grows faster than the naive expectation from the well known [@4] space phase in eqn.(\[phasespace\]), the effective $g_{\rho \pi \pi}$ acquires a temperature dependence which was ignored in most of previous works, but in agreement at low temperatures with an old calculation[@Song] and significantly the $\rho$ mass, which stays practically constant, in agreement with [@Ko] and specially [@4] (both papers offering an explanation of the dilepton excess below $M_\rho$). ![Thermal evolution of the $\rho$ mass, width and $\rho \pi \pi$ effective vertex.[]{data-label="rho"}](figmw.ps){height=".3\textheight"} [**Acknowledgements**]{} Work supported by grants FPA2000-0956, PB98-0782 and BFM2000-1326. J. R. P. acknowledges also support from the CICYT-INFN collaboration grant 003P 640.15. [20]{} C. A. Dominguez, M. Loewe and J. C. Rojas, Z. Phys. [**C59**]{}, 63 (1993). R. Pisarski, Nucl. Phys. [**A590**]{}, 553c (1995), employing Vector Meson Dominance. Same, if VMD does not hold. V. L. Eletsky [*et al.*]{} Phys. Rev. [**C64**]{} 035202 (2001). Only collisions with pions. collisions with both pions and nucleons in the medium included. Haglin, K. Nucl. Phys. [**A584**]{}, 719 (1995). C. M. Ko [*et al.*]{} Nucl. Phys. [**A610**]{}, 342c (1996). Y. B. He [*et al.*]{} Nucl. Phys. [**A630**]{}, 719 (1998). D. Blaschke [*et al*]{}, Int. J. Mod. Phys. [**A16**]{}, 2267, (2001). J. Gasser and H. Leutwyler, Ann. Phys. (N.Y.)[**158**]{}, 142 (1984). T. N. Truong, Phys. Rev. Lett. [**61**]{}, 2526 (1988); [*id.*]{} [**67**]{}, 2260 (1991); A. Dobado, M.J.Herrero and T.N. Truong, Phys. Lett. [**B235**]{}, 134 (1990); A. Dobado and J.R. Peláez, Phys. Rev. [**D47**]{}, 4883 (1993); Phys. Rev. [**D56**]{}, 3057 (1997). A. Gomez Nicola, F. J. Llanes-Estrada, J. R. Pelaez, Phys. Lett. [**B550**]{}:55-64 (2002). A. Dobado [*et al*]{}, Effective Lagrangians for the Standard Model, Springer-Verlag, Berlin, 1997. A. Dobado [*et al*]{}, Phys. Rev. [**C66**]{}, 055201 (2002). A. Gomez Nicola [*et al*]{} hep-ph/0212121. C. Song and V. Koch, Phys. Rev. [**C54**]{}, 3218 (1996).

--- abstract: 'C. Markett proved a Cohen type inequality for the classical Laguerre expansions in the appropriate weighted $L^{p}$ spaces. In this paper, we get a Cohen type inequality for the Fourier expansions in terms of discrete Laguerre–Sobolev orthonormal polynomials with an arbitrary (finite) number of mass points. So, we extend the result due to B. Xh. Fejzullahu and F. Marcellán.' author: - | A. Peña [^1], M. L. Rezola$^{*}$\ \ [Departamento de Matemáticas and IUMA.]{}\ [Universidad de Zaragoza (Spain).]{} title: | Discrete Laguerre-Sobolev expansions:\ A Cohen type inequality --- 2000MSC: 42C05 Key words: Laguerre polynomials; Laguerre–Sobolev type polynomials; Cohen type inequality; Bessel functions. Corresponding author: Ana Peña Departamento de Matemáticas. Universidad de Zaragoza 50009-Zaragoza (Spain) e-mail: anap@unizar.es Phone: 349761328, Fax: 34 976761338 Introduction and notations ========================== Littlewood conjectured in 1948 that for any trigonometric polynomial $F_N(x)=\sum_{k=1}^N a_k e^{in_k x}$ where $0< n_1 < n_2 < \cdots <n_N$, $\, N\ge 2$, and $\vert a_k\vert \ge 1$ for $1\le k \le N$, there holds the estimate from below $$\int_0^{2\pi} \vert F_N(x) dx \vert \ge C \,{\rm log}N$$ where $C$ is an absolute constant (see [@H-L]). Cohen’s inequality [@C] was the first result on the way to the solution of this conjecture. Later, inequalities of this type have been established in various other contexts, e.g., on compact group (see [@G-S-T]). In [@M2] Markett proved such inequalities for classical orthogonal polynomial expansions in the appropriate weighted $L^{p}$ spaces, here in terms of the highest coefficient. The main purpose of this paper is to extend these results to discrete Laguerre-Sobolev expansions. More precisely, we obtain such inequalities, in the appropriate weighted $L^{p}$ spaces, for Fourier expansions in terms of orthonormal polynomials with respect to an inner product of the form $$\label{innerproduct} \langle p,q \rangle_S = \frac{1}{\Gamma(\alpha +1)}\int_0^{\infty} p(x)q(x) \, x^{\alpha} e^{-x} \,dx + \sum_{j=0}^N M_j p^{(j)}(0) q^{(j)}(0),$$ where $\alpha > -1$ and $\, M_j \ge 0, \, j = 0, \dots, N.$ Such inner products are called of discrete Sobolev type. Recently in [@B-P], the authors Fejzullahu and Marcellán obtained Cohen type inequalities for orthonormal expansions with respect to the above inner product in the case $N=1$, i.e. at most two masses in the discrete part. In this particular case, the authors benefit from the fact that there are explicit formulas for the connection coefficients which appear in the representation of discrete Laguerre-Sobolev type polynomials in terms of three standard Laguerre polynomials (see [@km]). For a general discrete Laguerre-Sobolev inner product, we only know that these coefficients are a nontrivial solution of a system of $N+1$ equations on $N+2$ unknowns (see [@k1]). If the system is solved, we get an intricate expression with which it is difficult to work. Our contribution in this paper is that we can assure that there exists limit of these connection coefficients and this is enough for our purpose. Let $\{L_n^{\alpha}(x) \}_{n\ge0}$ be the sequence of Laguerre polynomials, orthogonal on $[0, \infty)$ with respect to the probability measure $\displaystyle d\mu(x)=\frac{1}{\Gamma(\alpha+1)}x^{\alpha}e^{-x}\,dx$ where $\alpha > -1$ and normalized by $L_n^{\alpha}(0)=\begin{pmatrix} n+\alpha \\ n \\ \end{pmatrix}$ . We denote the orthonormal Laguerre polynomial of degree $n$ by $$l_n^{\alpha}(x)=\frac{L_n^{\alpha}(x)}{\Vert L_n^{\alpha} \Vert }$$ where $\Vert L_n^{\alpha} \Vert ^2=\int_0^{\infty}{L_n^{\alpha}(x)}^2\,d\mu(x)$. Let $\{Q_n^{\alpha} \}_{n\ge0}$ be the sequence of discrete Laguerre–Sobolev orthogonal polynomials with respect to the inner product (\[innerproduct\]) and such that $Q_n^{\alpha}(x)$ and $L_n^{\alpha}(x)$ have the same leading coefficient. We denote by $$q_n^{\alpha}(x)={\langle Q_n^{\alpha},Q_n^{\alpha} \rangle }_S^{-1/2}Q_n^{\alpha}(x)$$ the orthonormal discrete Laguerre-Sobolev polynomials. From now on, for simplicity we write $Q_n(x)=Q_n^{\alpha}(x)$ and $q_n(x)=q_n^{\alpha}(x)$. Laguerre expansions have been investigated mainly in the following two sets of weighted Lebesgue spaces, namely in the classical spaces ([@A-W], [@Muc]) $$L_p(x^{\alpha}dx)=\left\{\begin{array} {ll} \{f; \quad \int_0^{\infty} |f(x)e^{-x/2}|^p x^{\alpha}dx< \infty \}, \quad & \hbox{if $1 \le p < \infty ;$} \\ \{f; \quad \displaystyle{\operatornamewithlimits{ess\,sup}}_{0<x<\infty} \,|f(x)e^{-x/2}|< \infty \}, \quad & \hbox{if $p = \infty ,$} \end{array} \right.$$ for $\alpha > -1$ as well as in the spaces $$L_p(x^{\alpha p/2}dx)=\left\{\begin{array} {ll} \{f; \quad \int_0^{\infty} |f(x)e^{-x/2} x^{\alpha/2}|^p dx< \infty \}, \,\, & \hbox{if $1 \le p < \infty ;$} \\ \{f; \quad \displaystyle{\operatornamewithlimits{ess\,sup}}_{0<x<\infty} \,|f(x)e^{-x/2}x^{\alpha/2}|< \infty \}, \,\, & \hbox{if $p = \infty ,$} \end{array} \right.$$ for $\alpha > -\frac{2}{p}$ if $1 \le p < \infty$ and $\alpha \ge 0$ if $p = \infty$. In order to unify the two results we are going to prove, we introduce an auxiliary parameter $\beta$ which means either $\alpha$ or $\alpha p/2$. We consider the class $S_p^{\beta}$, $1 \le p \le \infty$, defined as the space of measurable functions $f$ defined on $[0, \infty)$, such that there exits $f^{(k)}(0)$ for $k=0, \dots ,N$ and if $1 \le p < \infty$ $$\Vert f \Vert _{S_p^{\beta}}^p=\Vert f \Vert _{L_p(x^{\beta}dx)}^p+ \sum_{i=0}^{N} M_j\,|f^{j}(0)|^p< \infty ,$$ where $$\Vert f \Vert _{L_p(x^{\beta}dx)}^p=\int_0^{\infty} |f(x)e^{-x/2}|^p x^{\beta}dx, \quad 1 \le p <\infty,$$ and if $p = \infty$ $$\Vert f \Vert _{S_{\infty}^{\beta}}=\text{max} \{ \Vert f \Vert _{L_{\infty}(x^{\beta}dx)}, |f(0)|, \dots, |f^{(N)}(0)| \}< \infty,$$ where $$\Vert f \Vert _{L_{\infty}(x^{\beta}dx)}=\left\{ \begin{array}{ll} \displaystyle{\operatornamewithlimits{ess\,sup}}_{0<x<\infty} \,|f(x)e^{-x/2}|, & \hbox{if $\beta=\alpha ;$} \\ \displaystyle{\operatornamewithlimits{ess\,sup}}_{0<x<\infty} \,|f(x)e^{-x/2}x^{\alpha/2}|, & \hbox{if $\beta=\alpha p/2 .$} \end{array} \right.$$ (If some $M_j=0$ the corresponding derivative does not appear in the maximum.) Let $f\in S_p^{\beta}$, $1 \le p \le \infty$, then the Fourier expansion in terms of orthonormal discrete Laguerre-Sobolev polynomials $\{q_n\}_{n \ge 0}$, is $$\sum_{k=0}^{\infty}\hat{f}(k)\,q_k(x)$$ where $\hat{f}(k)=\langle f,q_k \rangle_S$. In the following, $[S_p^{\beta}]$ denotes the space of all bounded linear operators $T$ from the space $S_p^{\beta}$ into itself, endowed with the usual operator norm, $$\Vert T \Vert_{[S_p^{\beta}]}=\sup_{0 \not=f \in S_p^{\beta}}\frac{\Vert Tf \Vert_{S_p^{\beta}}}{\Vert f \Vert_{S_p^{\beta}}}.$$ Let $1 \le p \le \infty$. For a family of complex numbers $\{c_{k,n} \}_{k=0}^n, \, n \in \mathbb{N} \cup \{0 \}$, with $|c_{n,n}|>0$ we define the operators $T_n^{\alpha,S}:S_p^{\beta}\rightarrow S_p^{\beta}$ by $$T_n^{\alpha,S}(f)= \sum_{k=0}^n \, c_{k,n}\hat{f}(k)\,q_k.$$ Let us denote $q_0=\frac{4\alpha+4}{2\alpha+1}$ for $\beta=\alpha$ and $q_0=4$ for $\beta=p \alpha/2$, and let $p_0$ be the conjugate of $q_0$, i.e. $1/p_0+1/q_0=1$. Now, we can state our main theorem, which extends the ones given in [@M2] and [@B-P]. \[cohen\] Let $1 \le p \le \infty$. There exists a positive constant $C$, independent of $n$, such that: For $\alpha>-1/2$ $$\Vert T_n^{\alpha,S} \Vert_{[S_p^{\alpha}]}\ge C\,|c_{n,n}|\left\{ \begin{array}{ll} n^{\frac{2\alpha+2}{p}-\frac{2\alpha+3}{2}}, & \hbox{if $1 \le p <p_0$;} \\ (\log (n+1))^\frac{2\alpha+1}{4\alpha+4}, & \hbox{if $p=p_0,\, p=q_0$;} \\ n^{\frac{2\alpha+1}{2}-\frac{2\alpha+2}{p}}, & \hbox{if $q_0<p \le \infty$.} \end{array} \right.$$ For $\alpha>-2/p$ if $1 \le p < \infty$ and $\alpha \ge 0$ if $p=\infty$ $$\Vert T_n^{\alpha,S} \Vert_{[S_p^{p\alpha/2}]}\ge C\,|c_{n,n}|\left\{ \begin{array}{ll} n^{\frac{2}{p}-\frac{3}{2}}, & \hbox{if $1 \le p <p_0$;} \\ (\log (n+1))^\frac{1}{4}, & \hbox{if $p=p_0,\, p=q_0$;} \\ n^{\frac{1}{2}-\frac{2}{p}}, & \hbox{if $q_0<p \le \infty$.} \end{array} \right.$$ This theorem will be proved in Section 3. In Section 2, we obtain some new results for discrete Laguerre-Sobolev polynomials, which we will use to establish Theorem \[cohen\]. More concretely, we prove a technical lemma that will be used to deduce a Mehler-Heine type formula for Laguerre-Sobolev polynomials and a sharp estimation for their norm in the appropriate weighted $L_p$ spaces. In the sequel we use the following notation, $a_n \sim b_n$ means that there exist positive constants $c_1$ and $c_2$, such that $c_1a_n \le b_n \le c_2a_n$ for $n$ large enough, while $a_n \cong b_n$ means that the sequence $\frac{a_n}{b_n}$ converges to $1$. Throughout the paper, the values of the constants may change from line to line. Estimates for discrete Laguerre-Sobolev polynomials =================================================== Consider the standard Laguerre polynomials $L_n^{\alpha}$ and the Laguerre-Sobolev polynomials $Q_n$ with the same leading coefficient. Let us recall some properties of Laguerre polynomials for $\alpha > -1$ (see [@sz]). The evaluation at $x=0$ of the polynomials $L_n^{\alpha}$ and its successive derivatives are given by $$(L_n^{\alpha})^{(k)}(0)=\frac{(-1)^k \Gamma(n+\alpha+1)}{(n-k)!\,\Gamma(\alpha+k+1)}, \,k \in \mathbb{N}\cup \{0 \},$$ and their $L_2$-norm is $$\label{norma Laguerre} \Vert L_n^{\alpha} \Vert^2=\frac{1}{\Gamma(\alpha +1)}\int_0^{\infty} (L_n^{\alpha}(x))^2 \, x^{\alpha} e^{-x} \,dx=\frac{\Gamma(n+\alpha+1)}{n!\,\Gamma(\alpha+1)}.$$ As usual, we denote the derivatives of the $n$th kernels of Laguerre polynomials by $$K_n^{(k,h)}(x,y)=\displaystyle\frac{\partial^{k+h}}{\partial{x^k} \partial{y^h}}K_n(x,y) =\displaystyle \sum _{i=0}^n \frac{(L_i^{\alpha})^{(k)}(x)(L_i^{\alpha})^{(h)}(y)}{\Vert L_i^{\alpha} \Vert ^2}$$ with $k,h \in \mathbb{N}\cup \{0\}$ and the convention $K_n^{(0,0)}(x,y)=K_n(x,y)$. In the next lemma, we obtain an asymptotic estimate for $Q_n^{(k)}(0)$, that will play an important role along this paper. \[cocientederivadas\] Let $ Q_n $ be the polynomials orthogonal with respect to the inner product (\[innerproduct\]). Then the following statements hold: - $$\frac{Q_{n}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)}\cong \left\{ \begin{array}{ll} \displaystyle{\frac{C_k}{n^{\alpha+2k+1}}}, & \hbox{for k such that $ M_k>0$;} \\ \quad C_k, & \hbox{otherwise,} \end{array} \right.$$ where $C_k$ is a nonzero constant independent of $n$. - $$\langle Q_n, Q_n \rangle_S \cong \Vert L_n^{\alpha} \Vert ^2 \,.$$ **Proof.** If all the masses in the inner product (\[innerproduct\]) are zero the result is trivial because $Q_n=L_n^{\alpha}$. We will prove the result by induction concerning the number of positive masses in the inner product (\[innerproduct\]). We take the first mass which is positive, namely $M_{j_1}$ ($j_1 \ge 0$), and consider the sequence of polynomials $\{Q_{n,1}\}_{n \ge0}$ orthogonal with respect to the inner product $$(p,q )_{1} = \frac{1}{\Gamma(\alpha +1)}\int_0^{\infty} p(x)q(x) \, x^{\alpha} e^{-x} \,dx + M_{j_1} p^{(j_1)}(0) q^{(j_1)}(0).$$ The Fourier expansion of the polynomial $Q_{n,1}$ in the orthogonal basis $\{L_n^{\alpha}\}_{n\ge0}$ leads to $$Q_{n,1}(x) = L_n^{\alpha}(x) - M_{j_1} Q_{n,1}^{(j_1)}(0) K_{n-1}^{(0,j_1)}(x,0) \,.$$ Therefore $$\label{Qn,1(x)} Q_{n,1}(x) = L_n^{\alpha}(x) - \frac{M_{j_1} (L_n^{\alpha})^{(j_1)}(0)}{1+ M_{j_1} K_{n-1}^{(j_1,j_1)}(0,0)} K_{n-1}^{(0,j_1)}(x,0) \,,$$ and $$\label{equivnormas1} (Q_{n,1},Q_{n,1})_{1} =\Vert L_n^{\alpha} \Vert ^2+ M_{j_1}\, \frac{ (({L_n^{\alpha})}^{(j_1)}(0))^2}{1+ M_{j_1} K_{n-1}^{(j_1,j_1)}(0,0)} \,.$$ These relationships are very well known in the literature of discrete Sobolev type orthogonal polynomials. Taking derivatives $k$ times in (\[Qn,1(x)\]) and evaluating at $x=0$, we obtain $$\label{Qn,1/Ln} \frac{Q_{n,1}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)} =1- \frac{M_{j_1}K_{n-1}^{(k,j_1)}(0,0)}{1+M_{j_1}K_{n-1}^{(j_1,j_1)}(0,0)}\frac{(L_n^{\alpha})^{(j_1)}(0)}{(L_n^{\alpha})^{(k)}(0)} \,.$$ Applying the Stolz criterion (see, e.g. [@k]), we have $$\label{Kn(k,j1)} \lim_n \frac{K_{n-1}^{(k,j_1)}(0,0)}{n^{\alpha+k+j_1+1}}=\lim_n \frac{(L_{n-1}^{\alpha})^{(k)}(0)(L_{n-1}^{\alpha})^{(j_1)}(0)}{\Vert L_{n-1}^{\alpha} \Vert ^2 (\alpha +k+j_1+1) n^{\alpha+k+j_1}}\not=0,$$ and therefore $$\label{cociente nucleos} \frac{K_{n-1}^{(k,j_1)}(0,0)}{K_{n-1}^{(j_1,j_1)}(0,0)}\frac{(L_n^{\alpha})^{(j_1)}(0)}{(L_n^{\alpha})^{(k)}(0)}\cong \frac{(\alpha+2j_1+1)}{(\alpha+k+j_1+1)}\frac{(L_{n-1}^{\alpha})^{(k)}(0)}{(L_{n-1}^{\alpha})^{(j_1)}(0)} \frac{(L_n^{\alpha})^{(j_1)}(0)}{(L_n^{\alpha})^{(k)}(0)}$$ $$\cong \frac{\alpha+2j_1+1}{\alpha+k+j_1+1}.$$ Thus, from (\[Qn,1/Ln\]), (\[Kn(k,j1)\]) and (\[cociente nucleos\]), we have $$\frac{Q_{n,1}^{(j_1)}(0)}{(L_n^{\alpha})^{(j_1)}(0)}=\frac{1}{1+M_{j_1}K_{n-1}^{(j_1,j_1)}(0,0)}\cong \frac{C_{j_1}}{n^{\alpha+2j_1+1}}$$ and for $k\not=j_1$ $$\frac{Q_{n,1}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)} \cong 1-\frac{\alpha +2j_1+1}{\alpha+k+j_1+1}\not=0.$$ So, we achieve (a) for $Q_{n,1}$. Besides, taking limits in (\[equivnormas1\]) and using again the size of derivatives of Laguerre polynomials, we get (b) for the polynomials $Q_{n,1}$. If there are no more positive masses, since $Q_{n,1}=Q_n$ we have concluded the proof. Otherwise, suppose that the results (a) and (b) hold for the sequence of polynomials $\{Q_{n,s-1}\}_{n \ge0}$ orthogonal with respect to the inner product $$\begin{aligned} (p,q )_{s-1} &= \frac{1}{\Gamma(\alpha +1)}\int_0^{\infty} p(x)q(x) \, x^{\alpha} e^{-x} \,dx \\ \nonumber &+ M_{j_1} p^{(j_1)}(0) q^{(j_1)}(0)+ \dots +M_{j_{s-1}} p^{(j_{s-1})}(0) q^{(j_{s-1})}(0),\end{aligned}$$ where $j_1 <j_2 < \dots <j_{s-1}$ and all these masses are positive. Now, we have to prove the result for the polynomials $Q_{n,s}$ orthogonal with respect to $$\begin{aligned} (p,q )_{s} &= \frac{1}{\Gamma(\alpha +1)}\int_0^{\infty} p(x)q(x) \, x^{\alpha} e^{-x} \,dx \\ \nonumber &+ M_{j_1} p^{(j_1)}(0) q^{(j_1)}(0)+ \dots +M_{j_{s}} p^{(j_{s})}(0) q^{(j_{s})}(0),\end{aligned}$$ where $M_{j_{s}}>0$. Since $(p,q )_{s}=(p,q )_{s-1}+M_{j_s} p^{(j_s)}(0) q^{(j_s)}(0)$ we can work as before. Then the Fourier expansion of the polynomial $Q_{n,s}$ in the orthogonal basis $\{Q_{n,s-1}\}_{n\ge0}$ leads to $$Q_{n,s}(x) = Q_{n,s-1}(x) - M_{j_s} Q_{n,s}^{(j_s)}(0) K_{n-1, s-1}^{(0,j_s)}(x,0) \,,$$ where $K_{n,s-1}$ denotes the corresponding $n$th kernel for the sequence $\{Q_{n,s-1}\}$ and $$K_{n,s-1}^{(k,h)}(x,y)= \displaystyle \sum _{i=0}^n \frac{Q_{i,s-1}^{(k)}(x)Q_{i,s-1}^{(h)}(y)}{( Q_{i,s-1},Q_{i,s-1})_{s-1} }, \quad k,h \in \mathbb{N}\cup \{0\}.$$ Therefore, in the same way as in (\[Qn,1(x)\]) and (\[equivnormas1\]), we get $$\label{Qn,s(x)} Q_{n,s}(x) = Q_{n,s-1}(x) - \frac{M_{j_s} Q_{n,s-1}^{(j_s)}(0)}{1+ M_{j_s} K_{n-1,s-1}^{(j_s,j_s)}(0,0)} K_{n-1,s-1}^{(0,j_s)}(x,0) \,,$$ and $$\label{equivnormas s} (Q_{n,s},Q_{n,s})_{s} =(Q_{n,s-1},Q_{n,s-1})_{s-1} + M_{j_s} \frac{ ({Q_{n,s-1}}^{(j_s)}(0))^2}{1+ M_{j_s} K_{n-1,s-1}^{(j_s,j_s)}(0,0)} \,.$$ Taking derivatives $k$ times in (\[Qn,s(x)\]) and evaluating at $x=0$, we obtain $$\label{Qn,s/Ln} \frac{Q_{n,s}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)} =\frac{Q_{n,s-1}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)}\left[1 - \frac{M_{j_s}K_{n-1,s-1}^{(k,j_s)}(0,0)}{1+M_{j_s}K_{n-1,s-1}^{(j_s,j_s)}(0,0)} \frac{Q_{n,s-1}^{(j_s)}(0)}{Q_{n,s-1}^{(k)}(0)} \right] \,.$$ Applying the Stolz criterion and the hypotheses (a) and (b) for $\{ Q_{n,s-1} \}_{n\ge0}$, we can obtain $$\label{K_{n-1,s-1}} K_{n-1,s-1}^{(k,j_s)}(0,0) \cong \left\{ \begin{array}{ll} C_k \, n^{\alpha+k+j_s+1}, & \hbox{if $k \not=j_1,\dots, j_{s-1}$;} \\ C_k \, n^{j_s-k}, & \hbox{if $ k=j_1,\dots, j_{s-1}$,} \end{array} \right.$$ where $C_k$ is a nonzero constant. Indeed, for $k \not=j_1,\dots, j_{s-1}$, $$\begin{aligned} \label{K_{n-1,s-1}^{(k,js}(0,0)} \nonumber & \lim_n \frac{K_{n-1,s-1}^{(k,j_s)}(0,0)}{n^{\alpha+k+j_s+1}} =\lim_n \frac{Q_{n-1,s-1}^{(k)}(0)Q_{n-1,s-1}^{(j_s)}(0)}{(Q_{n-1,s-1},Q_{n-1,s-1})_{s-1}(\alpha+k+j_s+1) \,\, n^{\alpha+k+j_s}} \\ \nonumber &=\lim_n \frac{Q_{n-1,s-1}^{(k)}(0)}{(L_{n-1}^{\alpha})^{(k)}(0)}\lim_n\frac{Q_{n-1,s-1}^{(j_s)}(0)}{(L_{n-1}^{\alpha})^{(j_s)}(0)}\lim_n \frac{(L_{n-1}^{\alpha})^{(k)}(0)(L_{n-1}^{\alpha})^{(j_s)}(0)}{\Vert L_{n-1}^{\alpha} \Vert^2 \,(\alpha+k+j_s+1)\, n^{\alpha+k+j_s}}, \\\end{aligned}$$ and, for $k=j_1,\dots, j_{s-1}$, $$\begin{aligned} \label{K_{n-1,s-1}^{(k,js)}(0,0)} & \lim_n \frac{K_{n-1,s-1}^{(k,j_s)}(0,0)}{n^{j_s-k}} =\lim_n \frac{Q_{n-1,s-1}^{(k)}(0)Q_{n-1,s-1}^{(j_s)}(0)}{(Q_{n-1,s-1},Q_{n-1,s-1})_{s-1}(j_s-k) \,\, n^{j_s-k-1}} \\ \nonumber &=\lim_n \frac{(L_{n-1}^{\alpha})^{(k)}(0)(L_{n-1}^{\alpha})^{(j_s)}(0)}{\Vert L_{n-1}^{\alpha} \Vert^2(j_s-k) \,\, n^{\alpha+k+j_s}} \lim_n n^{\alpha+2k+1} \frac{Q_{n-1,s-1}^{(k)}(0)}{(L_{n-1}^{\alpha})^{(k)}(0)} \lim_n \frac{Q_{n-1,s-1}^{(j_s)}(0)}{(L_{n-1}^{\alpha})^{(j_s)}(0)}.\end{aligned}$$ Then, from (\[Qn,s/Ln\]), (\[K\_[n-1,s-1]{}\]) and the hypothesis for $Q_{n,s-1}$, we have $$\frac{Q_{n,s}^{(j_s)}(0)}{(L_n^{\alpha})^{(j_s)}(0)}=\frac{Q_{n,s-1}^{(j_s)}(0)}{(L_n^{\alpha})^{(j_s)}(0)} \frac{1}{1+M_{j_s}K_{n-1,s-1}^{(j_s,j_s)}(0,0)}\cong \frac{C_{j_s}}{n^{\alpha+2j_s+1}},$$ with $C_{j_s}$ a nonzero constant. Moreover, for $k\not=j_s$, taking into account (\[K\_[n-1,s-1]{}\^[(k,js]{}(0,0)\]), (\[K\_[n-1,s-1]{}\^[(k,js)]{}(0,0)\]) and the hypothesis for $Q_{n,s-1}$, we can deduce $$\begin{aligned} \nonumber & \frac{K_{n-1,s-1}^{(k,j_s)}(0,0)}{K_{n-1,s-1}^{(j_s,j_s)}(0,0)}\frac{Q_{n,s-1}^{(j_s)}(0)}{Q_{n,s-1}^{(k)}(0)}= \frac{K_{n-1,s-1}^{(k,j_s)}(0,0)}{K_{n-1,s-1}^{(j_s,j_s)}(0,0)}\frac{Q_{n,s-1}^{(j_s)}(0)}{(L_{n}^{\alpha})^{(j_s)}(0)} \frac{(L_{n}^{\alpha})^{(k)}(0)}{Q_{n,s-1}^{(k)}(0)}\frac{(L_{n}^{\alpha})^{(j_s)}(0)}{(L_{n}^{\alpha})^{(k)}(0)}\\ \nonumber & \cong \frac{(L_{n}^{\alpha})^{(j_s)}(0)}{(L_{n-1}^{\alpha})^{(j_s)}(0)} \frac{(L_{n-1}^{\alpha})^{(k)}(0)}{(L_{n}^{\alpha})^{(k)}(0)}\left\{ \begin{array}{ll} \frac{\alpha+2j_s+1}{\alpha+k+j_s+1}, & \hbox{if $k\not=j_1,\dots, j_{s-1}$;} \\ \frac{\alpha+2j_s+1}{j_s-k}, & \hbox{if $k=j_1,\dots, j_{s-1}$,} \end{array} \right.\\ \nonumber & \cong \left\{ \begin{array}{ll} \frac{\alpha+2j_s+1}{\alpha+k+j_s+1}, & \hbox{if $k\not=j_1,\dots, j_{s-1}$;} \\ \frac{\alpha+2j_s+1}{j_s-k}, & \hbox{if $k=j_1,\dots, j_{s-1}$.} \end{array} \right.\end{aligned}$$ Thus, taking limits in (\[Qn,s/Ln\]) and (\[equivnormas s\]), we get $(a)$ and $ (b)$ for the polynomials $Q_{n,s}$, i.e. $$\frac{Q_{n,s}^{(k)}(0)}{(L_n^{\alpha})^{(k)}(0)}\cong \left\{ \begin{array}{ll} \displaystyle{\frac{C_k}{n^{\alpha+2k+1}}}, & \hbox{if $ k=j_1,\dots, j_{s}$;} \\ \quad C_k, & \hbox{otherwise,} \end{array} \right.$$ and $$(Q_{n,s},Q_{n,s})_{s} \cong \Vert L_n^{\alpha} \Vert ^2 .$$ Hence the result follows.$\quad \Box$ Observe that the part $(a)$ of Lemma \[cocientederivadas\] is also true for the ratio of the corresponding orthonormal polynomials, and therefore there exists $$\label{limitecocientederivadas} \lim_n\frac{q_n^{(k)}(0)}{(l_n^{\alpha})^{(k)}(0)}=\left\{ \begin{array}{ll} 0, & \hbox{for $k$ such that $M_k>0$;} \\ C_k\not=0, & \hbox{otherwise.} \end{array} \right.$$ Consider the following representation of the orthonormal polynomials $q_n$ in terms of the orthonormal Laguerre polynomials $l_n^{\alpha}$ (see [@k1 Section 9] ) $$\label{formulaKoekoek} q_n(x)=\sum_{j=0}^{N+1} b_j(n)x^jl_{n-j}^{\alpha+2j}(x).$$ For the inner product (\[innerproduct\]) with $N=1$, the coefficients $b_j(n)$ was explicitly obtained in [@km], and their estimation was essential to obtain the result in [@B-P]. Now in the general case, using Lemma \[cocientederivadas\], we can prove that there is always limit of the connection coefficients $b_j(n)$ for an arbitrary $N$. \[limitecoeficientesKoekoek\] Let $\{b_j(n)\}_{0}^{N+1}$ be the coefficients in formula (\[formulaKoekoek\]). Then, there exists $$\lim_n b_j(n)=b_j \in \mathbb{R}, \quad j \in \{0, \dots, N+1\}.$$ Moreover, the first index $j$ such that $b_j\not=0$ corresponds with the first $j$ such that $M_j=0$ in the inner product (\[innerproduct\]). (We understand that if all the masses are positive, then the unique coefficient $b_j$ different from zero is the last one). **Proof.** Taking derivatives $k$ times in (\[formulaKoekoek\]) and evaluating at $x=0$, we deduce $$\label{cocientederivadasortonormales} \frac{q_n^{(k)}(0)}{(l_n^{\alpha})^{(k)}(0)}=\sum_{j=0}^{k} b_j(n) \left( \begin{array}{c} k \\ j \\ \end{array} \right) j!\,A_j(k,n), \quad k\in \{0, \dots, N+1\},$$ where $A_0(k, n)=1$ and $$\label{expresionA} A_j(k, n)=\displaystyle \frac{(l_{n-j}^{\alpha+2j})^{(k-j)}(0)}{(l_n^{\alpha})^{(k)}(0)}\cong \frac{(-1)^j\Gamma(\alpha+k+1)}{\Gamma(\alpha+k+j+1)}\left(\frac{\Gamma(\alpha+2j+1)}{\Gamma(\alpha+1)}\right)^{1/2}$$ Since there exists $\lim_n A_j(k, n)\not=0$ , applying recursively (\[limitecocientederivadas\]) and (\[cocientederivadasortonormales\]) we can assure there exists $\lim_n b_j(n)=b_j, \, j \in \{0, \dots, N+1 \}$. More precisely, for $k=0$ we have $$\lim_n b_0(n)=\lim_n \frac{q_n(0)}{l_n^{\alpha}(0)}=b_0=\left\{ \begin{array}{ll} 0, & \hbox{if $M_0>0$;} \\ C\not=0, & \hbox{if $M_0=0$.} \end{array} \right.$$ Now, from (\[cocientederivadasortonormales\]) for $k=1$, (\[limitecocientederivadas\]) and (\[expresionA\]) we get $$\lim_n b_1(n)=\lim_n \frac{1}{A_1(1,n)}\left(\frac{q'_n(0)}{(l_n^{\alpha})'(0)}-b_0(n)\right)=b_1$$ Observe that $$b_1=\left\{ \begin{array}{ll} 0, & \hbox{if $M_0>0$ and $M_1>0$ ;} \\ C\not=0, & \hbox{if $M_0>0$ and $M_1=0$ .} \end{array} \right.$$ In this way, recursively, if $M_0M_1 \dots M_i>0$ and $M_{i+1}=0$ we can assure that $$b_j=\left\{ \begin{array}{ll} 0, & \hbox{if $0 \le j \le i$;} \\ C\not=0, & \hbox{if $j=i+1$,} \end{array}\right.$$ and we obtain the result. $\quad \Box$ As a consequence of the above lemma, we can establish a Mehler-Heine type formula for general discrete Laguerre-Sobolev orthonormal polynomials. This formula shows how the presence of the masses in the discrete part of the inner product changes the asymptotic behavior around the origin. Moreover, it supplies information on the location and asymptotic distribution of the zeros of the polynomials in terms of the zeros of known special functions. We recall the corresponding formula for orthonormal Laguerre polynomials (see [@sz]) $$\label{MHLaguerre} \lim_n\frac{l_n^{\alpha}(x/(n+k))}{n^{\alpha /2}}= \sqrt{\Gamma (\alpha +1)} x^{-\alpha /2} J_{\alpha}(2 \sqrt{x})$$ uniformly on compact subsets of $\mathbb{C}$ and uniformly for $k \in \mathbb{N}\cup \{0 \}$, where $J_{\alpha}$ is the Bessel function of the first kind. \[Mehler-Heine\] The polynomials $q_n$ satisfy the following Mehler-Heine type formula: $$\label{MHLaguerre-Sobolev} \lim_n\frac{q_n(x/n)}{n^{\alpha /2}}= \sqrt{\Gamma (\alpha +1)} \sum_{j=0}^{N+1} b_j\, x^{-\alpha /2} J_{\alpha +2j}(2 \sqrt{x})$$ uniformly on compact subsets of $\mathbb{C}$. **Proof.** The proof is a straightforward consequence of formula (\[formulaKoekoek\]), Lemma \[limitecoeficientesKoekoek\] and (\[MHLaguerre\]). $\quad \Box$ **Remark.** According to Lemma \[limitecoeficientesKoekoek\], the first Bessel function which appears in (\[MHLaguerre-Sobolev\]) corresponds with the first index $j$ such that $M_j=0$, in the inner product (\[innerproduct\]). We want to highlight that this result generalizes the one obtained in [@A-B-P-R Theorem 3], where the authors only deal with inner products with a unique $\lq\lq$gap" in the discrete part. The above proposition allows us to deduce a lower estimate of $\Vert q_n \Vert _{L_p(x^{\beta}dx)}$, for $\beta=\alpha$ and $\beta=\alpha p/2$, that will play an important role in the proof of Theorem \[cohen\]. \[acotacionnormasSobolev\] Let $1 \le p \le \infty$. Then, the following statements hold: For $\alpha>-1/2$ $$\Vert q_n \Vert _{L_p(x^{\alpha}dx)} \ge C \left\{ \begin{array}{ll} n^{-1/4}(\log (n+1))^{1/p}, & \hbox{if $p=\frac{4\alpha+4}{2\alpha+1}$;} \\ n^{\alpha/2-(\alpha+1)/p}, & \hbox{if $\frac{4\alpha+4}{2\alpha+1}<p \le \infty ,$} \end{array} \right.$$ and for $\alpha>-2/p$ if $1 \le p < \infty$ and $\alpha \ge 0$ if $p = \infty$ $$\Vert q_n \Vert _{L_p(x^{\alpha p/2}dx)} \ge C\left\{ \begin{array}{ll} n^{-1/4}(\log (n+1))^{1/p}, & \hbox{if $p=4$;} \\ n^{-1/p}, & \hbox{if $4<p \le \infty $,} \end{array} \right.$$ where $C$ is an absolute positive constant. **Proof.** Assume $1 \le p < \infty$. Then, $$\begin{aligned} \nonumber &\Vert q_n \Vert^p _{L_p(x^{\beta}dx)}= \int_0^{\infty}|q_n(x)e^{-x/2}|^px^{\beta}dx \\ \nonumber &> \int_0^{1/\sqrt{n}}|q_n(x)e^{-x/2}|^px^{\beta}dx \ge C n^{- \beta-1}\int_0^{\sqrt{n}}|q_n(t/n)|^pt^{\beta}dt \\ \nonumber\end{aligned}$$ According to formula (\[MHLaguerre-Sobolev\]), $\exists \, n_0\in \mathbb{N}$ such that $\forall n\ge n_0$ $$\int_0^{\sqrt{n}}|q_n(t/n)|^pt^{\beta}dt \ge C n^{p\alpha/2} \int_0^{\sqrt{n}} | \sum_{j=0}^{N+1} b_j\, t^{-\alpha/2}\, J_{\alpha +2j}(2\sqrt{t})|^p t^{\beta} dt$$ and therefore $\forall n\ge n_0$ $$\Vert q_n \Vert^p _{L_p(x^{\beta}dx)} \ge Cn^{p\alpha/2- \beta-1}\int_0^{2{n}^{1/4}}u^{2\beta-p\alpha+1} | \sum_{j=0}^{N+1} b_j\, J_{\alpha +2j}(u)|^pdu .$$ Working as Stempak in [@S Lemma 2.1], we can prove that for $\alpha >-1$, and $\lambda>-1 -\alpha p$ $$\int_0^{2{n}^{1/4}}u^{\lambda} | \sum_{j=0}^{N+1} b_j \, J_{\alpha +2j}(u)|^pdu \sim \left\{ \begin{array}{ll} 1, & \hbox{if $\lambda<p/2-1$;} \\ \log (n+1), & \hbox{if $\lambda=p/2-1$.} \end{array} \right.$$ Thus, if $1 \le p < \infty$, we obtain the first and the second result for $\beta=\alpha$ and $\beta=p\alpha/2$ respectively. The results for $p=\infty$ can be deduced from the previous one by passing to the limit when $p$ goes to $\infty$. $\quad\Box$ It is worth to noticing that these lower bounds are sharp in the following sense. Let $1 \le p \le \infty$. Then: For $\alpha \ge 0$, $$\Vert q_n \Vert _{L_p(x^{\alpha}dx)} \sim \left\{ \begin{array}{ll} n^{-1/4}(\log (n+1))^{1/p}, & \hbox{if $p=\frac{4\alpha+4}{2\alpha+1}$;} \\ n^{\alpha/2-(\alpha+1)/p}, & \hbox{if $\frac{4\alpha+4}{2\alpha+1}<p\le {\infty} $,} \end{array} \right.$$ and for $\alpha>-2/p$ if $1 \le p < \infty$ and $\alpha \ge 0$ if $p = \infty$, $$\Vert q_n \Vert _{L_p(x^{\alpha p/2}dx)} \sim \left\{ \begin{array}{ll} n^{-1/4}(\log (n+1))^{1/p}, & \hbox{if $p=4$;} \\ n^{-1/p}, & \hbox{if $4<p \le {\infty} $.} \end{array} \right.$$ **Proof.** From Lemma 1 of [@M1] it can be deduced that for $\alpha \ge 0$ $$\int_0^{\infty}|x^jl_n^{\alpha +2j}(x)e^{-x/2}|^px^{\alpha}dx \sim \left\{ \begin{array}{ll} n^{-p/4}\log (n+1), & \hbox{if $p=\frac{4\alpha+4}{2\alpha+1}$;} \\ n^{\alpha p/2-(\alpha+1)}, & \hbox{if $\frac{4\alpha+4}{2\alpha+1}<p\le {\infty} $,} \end{array} \right.$$ and for $\alpha>-2/p$ if $1 \le p < \infty$ and $\alpha \ge 0$ if $p = \infty$ $$\int_0^{\infty}|x^jl_n^{\alpha +2j}(x)e^{-x/2}x^{\alpha/2}|^pdx \sim \left\{ \begin{array}{ll} n^{-p/4}\log (n+1), & \hbox{if $p=4$;} \\ n^{-1}, & \hbox{if $4<p \le {\infty} $.} \end{array} \right.$$ Thus, using the representation formula for the polynomials $q_n $ (see (\[formulaKoekoek\])), and the fact that the connection coefficients are bounded (see Lemma \[limitecoeficientesKoekoek\]), we get one of the two inequalities. The other one has been proved in Proposition \[acotacionnormasSobolev\] and therefore the result follows. $\quad\Box$ A Cohen type inequality ======================= In this section we prove a Cohen type inequality for the Fourier expansions in terms of discrete Laguerre-Sobolev orthonormal polynomials with an arbitrary (finite) number of mass points. So we extend the result due to Fejzullahu and Marcellán which deals with a discrete Laguerre-Sobolev inner product with at most two masses in the discrete part (see [@B-P]). **Proof of Theorem \[cohen\].** Let us consider the following test functions which were already used in [@M2] and later in [@B-P] $$\label{funcionestest} g_n^{\alpha,j}(x)=x^j\left[L_n^{\alpha+j}(x)-\sqrt{\frac{(n+1)(n+2)}{(n+\alpha+j+1)(n+\alpha+j+2)}}\,L_{n+2}^{\alpha+j}(x) \right],$$ with $j \in \mathbb{N}\setminus \{1, \dots, N \}$. Notice that $$\label{derivadasfuncionestest} (g_n^{\alpha,j})^{(i)}(0)=0, \, i=0, \dots, N.$$ These functions can be written as (see formula (2.15) in [@M2]) $$\label{desarrollofuncionestest} g_n^{\alpha,j}(x)=\sum_{m=0}^{j+2}a_{m,j}(\alpha, n)L_{n+m}^{\alpha}(x)$$ with $$\label{estimacionaoj} a_{0,j}(\alpha, n)=\frac{\Gamma(n+\alpha+j+1)}{\Gamma(n+\alpha+1)}\cong n^j.$$ From (\[derivadasfuncionestest\]), (\[desarrollofuncionestest\]), and $0 \le k \le n$, we have $$\begin{aligned} &\widehat{g_n^{\alpha,j}}(k)=\langle g_n^{\alpha,j}, q_k \rangle_S =\frac{1}{\Gamma(\alpha+1)}\int_0^{\infty}g_n^{\alpha,j}(x)q_k(x)e^{-x}x^{\alpha}dx \\ &=\frac{1}{\Gamma(\alpha+1)}\sum_{m=0}^{j+2}a_{m,j}(\alpha, n)\int_0^{\infty}L_{n+m}^{\alpha}(x)q_{k}(x)e^{-x}x^{\alpha}dx\,.\end{aligned}$$ By the orthogonality of Laguerre polynomials, we obtain $$\widehat{g_n^{\alpha,j}}(k)= \left\{ \begin{array}{ll} 0, &\hbox{\, if $0 \le k \le n-1$;} \\ \frac{1}{\Gamma(\alpha+1)}a_{0,j}(\alpha, n)\int_0^{\infty}L_{n}^{\alpha}(x)q_{n}(x)e^{-x}x^{\alpha}dx , &\hbox{\, if $k=n$.} \end{array} \right.$$ Thus, from Lemma \[cocientederivadas\] (b), the estimate of $a_{0,j}(\alpha, n)$ and the value of the norm of Laguerre polynomials (see (\[norma Laguerre\])), we can deduce $$\begin{aligned} & \widehat{g_n^{\alpha,j}}(n)=\frac{1}{\Gamma(\alpha+1)}a_{0,j}(\alpha, n)\int_0^{\infty}L_{n}^{\alpha}(x)\frac{Q_{n}(x)}{\langle Q_{n}, Q_{n} \rangle_S^{1/2} }e^{-x}x^{\alpha}dx =\\ &a_{0,j}(\alpha, n) \, \frac{\Vert L_n^{\alpha}\Vert^2}{\langle Q_{n}, Q_{n} \rangle_S^{1/2}} \cong a_{0,j}(\alpha, n) \, \Vert L_n^{\alpha}\Vert \cong \frac{n^{j +\alpha/2}}{\sqrt{\Gamma(\alpha+1)}}\\\end{aligned}$$ Observe that $Q_{n}$ and $L_{n}^{\alpha}$ have always equivalent norms, and, therefore this estimation does not depend neither on the number of positive masses, nor on the existence or non-existence of any gap in the inner product. Applying the operator $T_n^{\alpha,S}$ to the functions $g_n^{\alpha,j}$, we get $$\label{Tfunciones test} T_n^{\alpha,S}(g_n^{\alpha,j})= c_{n,n}\, \widehat{g_n^{\alpha,j}}(n)q_n,$$ and therefore $$\begin{aligned} \Vert T_n^{\alpha,S}\Vert_{[S_p^{\beta}]} & \ge (\Vert g_n^{\alpha,j}\Vert_{S_p^{\beta}})^{-1}\Vert T_n^{\alpha,S}(g_n^{\alpha,j}) \Vert_{S_p^{\beta}}=(\Vert g_n^{\alpha,j}\Vert_{S_p^{\beta}})^{-1} |c_{n,n}||\widehat{g_n^{\alpha,j}}(n)|\,\Vert q_n \Vert_{S_p^{\beta}}\\ &\ge (\Vert g_n^{\alpha,j}\Vert_{S_p^{\beta}})^{-1} |c_{n,n}||\widehat{g_n^{\alpha,j}}(n)|\,\Vert q_n \Vert_{L_p(x^{\beta}dx)}.\\\end{aligned}$$ On the other hand, for $j>\alpha-1/2-2(\alpha +1)/p$ we have $$\Vert g_n^{\alpha,j}\Vert_{S_p^{\beta}} \le c \left\{ \begin{array}{ll} n^{j-1/2+(\alpha +1)/p}, & \hbox{if $\beta=\alpha$;} \\ n^{\alpha/2+j-1/2+1/p}, & \hbox{if $\beta=p\alpha/2$,} \end{array} \right.$$ (see formula (3.3) and formula (1.19), (2.12) in [@M2] respectively). Thus, by Proposition \[acotacionnormasSobolev\] we get: For $\beta=\alpha$ with $\alpha >-1/2$ $$\Vert T_n^{\alpha,S}\Vert_{[S_p^{\alpha}]} \ge C |c_{n,n}|\left\{ \begin{array}{ll} (\log (n+1))^\frac{2\alpha +1}{4\alpha +4}, & \hbox{if $p=q_0$;} \\ n^{\alpha+1/2-2(\alpha+1)/p}, & \hbox{if $q_0<p \le \infty$.} \end{array} \right.$$ For $\beta=p\alpha/2$ with $\alpha>-2/p$ if $1 \le p<\infty$ and $\alpha\ge 0$ if $p=\infty$, $$\Vert T_n^{\alpha,S}\Vert_{[S_p^{p\alpha/2}]} \ge C|c_{n,n}|\left\{ \begin{array}{ll} (\log (n+1))^{1/4}, & \hbox{if $p=4$;} \\ n^{1/2-2/p}, & \hbox{if $4<p \le \infty$.} \end{array} \right.$$ Hence, by duality the theorem follows. $\quad \Box$ **Remark.** In particular, for $M_i=0$, $i=0, \dots, N$, the above theorem extends Theorem 1 in [@M2] to negative values of $\alpha$. In the particular case of $c_{k,n}=1$, $k=0, \dots, n$, the operator $T_n^{\alpha,S}$ is the $n$th partial sum of the Fourier expansion, so, we can assure the following result. If $p$ is outside the Pollard interval $(p_0, q_0)$, we have $$\Vert S_n \Vert_{[S_p^{\beta}]} \rightarrow \infty,\quad n \rightarrow \infty$$ where $S_n$ denotes the $n$th partial sum of the Fourier expansion. [10]{} M. Alfaro, J.J. Moreno-Balcázar, A. Peña, M.L. Rezola, A new approach to the asymptotics of Sobolev type orthogonal polynomials, J. Approx. Theory 163 (2011) 460-480. R. Askey, S. Wainger, Mean convergence of expansions in Laguerre and Hermite series, Amer. J. Math. 87 (1965) 695-708. P. J. Cohen, On a conjecture of Littlewood and idempotent measures,Amer. J. Math. 82 (1960) 191-212. B. Xh. Fejzullahu, F. Marcellán, A Cohen type inequality for Laguerre–Sobolev expansions, J. Math. Anal. Appl. 352 (2009) 880-889. S. Giulini, P. M. Soardi, G. Travaglini, A Cohen type inequality for compact Lie groups, Proc. Amer. Math. Soc. 77 (1979) 359-364. G. H. Hardy, J. E. Littlewood, A new proof of a theorem on rearrangements, J. London Math. Soc., 23 (1948) 163-168. G. Klambauer, Aspects of Calculus, Springer–Verlag, New York 1986. R. Koekoek, Generalizations of Laguerre polynomials, J. Math. Anal. Appl. 153 (1990) 576-590. R. Koekoek, H.G. Meijer, A generalization of Laguerre polynomials, SIAM J. Math. Anal. 24 (1993) 768-782. B. Muckenhoupt, Mean convergence of Hermite and Laguerre series II, Trans Amer. Math. Soc. 147 (1970) 433-460. C. Markett, Mean Cesàro summability of Laguerre expansions and norm estimates with shifted parameter, Anal. Math. 8 (1982) 19-37. C. Markett, Cohen type inequalities for Jacobi, Laguerre and Hermite expansions, Siam J. Math. Anal. 14 (1983) 819-833. K. Stempak, On convergence and divergence of Fourier-Bessel series, Electron. Trans. Numer. Anal. 14 (2002) 223-235. G. Szegő, Orthogonal Polynomials, Amer. Math. Soc. Colloq. Publ. vol. 23, Amer. Math. Soc., Providence R.I., 1975. Fourth Edition. [^1]: Both authors partially supported by MICINN of Spain under Grant MTM2009-12740-C03-03, FEDER funds (EU), and the DGA project E-64 (Spain)

--- abstract: | Coherence in an open quantum system is degraded through its interaction with a bath. This decoherence can be avoided by restricting the dynamics of the system to special decoherence-free subspaces. These subspaces are usually constructed under the assumption of spatially symmetric system-bath coupling. Here we show that decoherence-free subspaces may appear without spatial symmetry. Instead, we consider a model of system-bath interactions in which to first order only multiple-qubit coupling to the bath is present, with single-qubit system-bath coupling absent. We derive necessary and sufficient conditions for the appearance of decoherence-free states in this model, and give a number of examples. In a sequel paper we show how to perform universal and fault tolerant quantum computation on the decoherence-free subspaces considered in this paper. PACS numbers: 03.67.Lx, 03.65.Bz, 03.65.Fd, 89.70.+c address: | Departments of Chemistry$^{1}$, Physics$^{2}$ and Mathematics$^{3}$,\ University of California, Berkeley, CA 94720\ École Nationale Superieure des Télécommunications,\ Paris, France$^4$ author: - 'Daniel A. Lidar,$^{1}$ Dave Bacon,$^{1,2}$ Julia Kempe$^{1,3,4}$ and K.B. Whaley$^{1}$' title: 'Decoherence-Free Subspaces for Multiple-Qubit Errors: (I) Characterization' --- Introduction ============ Quantum information must be protected against the detrimental effects of decoherence [@Lo:book; @Williams:book98]. To this end Decoherence-Free Subspaces (DFSs) [@Duan:98; @Duan:98c; @Zanardi:97c; @Zanardi:97a; @Zanardi:98a; @Lidar:PRL98; @Lidar:PRL99; @Bacon:99] have recently been proposed, alongside Quantum Error Correcting Codes (QECCs) [@Shor:95; @Steane:96a; @Gottesman:97; @Knill:97b] and “dynamical decoupling” and symmetrization schemes [@Viola:98; @Viola:99; @Duan:98e; @Zanardi:98b]. A DFS is a “quiet corner” of the system’s Hilbert space, where the evolution is decoupled from the bath and thus is entirely unitary. DFSs are a special class of (fully degenerate) QECCs [@Lidar:PRL99], so in order to properly distinguish between DFSs and all other QECCs we note that DFSs are [*passive*]{} codes, in that the information encoded in them may not require any active stabilization procedures [@Bacon:99a; @Kempe:00]. All other QECCs, in contrast, always involve an [*active*]{} error-detection/correction process. Examples of DFSs have so far focused almost exclusively on the presence of a [*permutation symmetry*]{} of some sort in the system-bath coupling. The most often used example is that of  “collective decoherence” [@Duan:98; @Duan:98c; @Zanardi:97c; @Lidar:PRL98; @Lidar:PRA99Exchange], where the bath couples in an identical fashion to all qubits, implying that all qubits undergo the [*same*]{} error. In this case four physical qubits suffice to encode a logical qubit against any collective error, and the code efficiency (number of encoded per physical qubits) approaches unity asymptotically [@Zanardi:97c]. It was shown that the requirement of an exact symmetry can be lifted by allowing for a symmetry-breaking perturbation, without spoiling the DFS property significantly [@Lidar:PRL98; @Bacon:99]. Moreover, by concatenation with an active QECC, a symmetry-broken DFS can be stabilized completely [@Lidar:PRL99]. While these results indicate that a small departure from the exact symmetry condition for the system-bath coupling is admissible, they leave unanswered the question of whether a DFS may exist when no assumptions are made regarding the spatial symmetry of this coupling. In this paper, the first out of two, it will be shown that under conditions which do not relate to a spatially symmetric system-bath coupling, DFSs may still exist. This result is exact, i.e., it is not of a perturbative nature as in Refs. [@Lidar:PRL98; @Lidar:PRL99; @Bacon:99]. Instead, it relies on the assumption that errors affecting single qubits are absent, and to lowest order only multiple-qubit errors are possible instead. Formally, the condition is that the qubit register is not affected by the full Pauli group of errors, but only by a subgroup thereof. One may then proceed to find DFSs with respect to this subgroup. The interesting class of system-bath interaction Hamiltonians which allow for such processes generally involve only multiple-qubit operators. Relevant physical systems are therefore those where the bath can only couple to multiple system excitations as is the case for decoherence due to dipolar coupling, e.g., in NMR [@Slichter:96]. Another interesting class of examples are composite particles, such as bi-excitons in quantum dots/wells [@Chemla:98], or Cooper pairs in superconductors [@March:85vol2]. The structure of this paper is as follows. In section \[Hamiltonians\] we briefly review the structure of Hamiltonians pertinent to systems that may function as quantum computers, coupled to a decohering environment. Using these Hamiltonians, we recall in section \[Evolution\] the derivation of the operator sum representation evolution equation for the system density matrix. We show in particular that for a qubit system, the evolution can be expressed entirely in terms of linear combinations of tensor products of Pauli matrices. We then use this in section \[DFS\] to derive the DFS condition under the assumption that decoherence is the result of a subgroup of the Pauli group. In section \[Examples\] we illustrate our general analysis with some examples, and find decoherence-free states for a number of subgroups. We derive the dimension of these DFSs in section \[dimension\]. Conclusions and a summary are presented in section \[Summary\]. Finally, some important properties of the Pauli group are summarized in Appendix \[app1\], and some examples of “non-generic” DFSs are presented in Appendix \[app2\]. We show in a sequel paper [@Lidar:00b] how to perform universal fault tolerant quantum computation using at most 2-body Hamiltonians on the DFSs derived here. Structure of the Hamiltonian for a Universal Quantum Computer Coupled to a Bath {#Hamiltonians} =============================================================================== This section provides a brief review of the structure of Hamiltonians relevant for a qubit system allowing for universal quantum computation and coupled to a decohering bath. The dynamics of a quantum system $S$ coupled to a bath $B$ (which together form a closed system) evolves unitarily under the combined Hamiltonian $${\bf H}={\bf H}_{S}{\bf \otimes }{\bf I}_{B}+{\bf I}_{S}{\bf \otimes H}_{B}+ {\bf H}_{I},$$ where ${\bf H}_{S}$, ${\bf H}_{B}$ and ${\bf H}_{I}$ are the system, bath and interaction Hamiltonians, respectively; ${\bf I}$ is the identity operator. Let ${\bf \sigma }_{i}^{\alpha }$ denote the $\alpha ^{{\rm th}}$ Pauli matrix, $\alpha =\{0,x,y,z\}$, acting on qubit $i$. The $2\times 2$ identity matrix is denoted ${\bf \sigma }_{i}^{0}$. For $K$ qubits the components of ${\bf H}$ can often be written as follows: $${\bf H}_{S}=\sum_{i=1}^{K}\sum_{\alpha =x,z}\varepsilon _{i}^{\alpha }{\bf \sigma }_{i}^{\alpha }+\sum_{i\neq j}^{K}J_{ij}{\bf \sigma }_{i}^{+}{\bf \sigma }_{j}^{-}+{\rm h.c.},$$ where ${\bf \sigma }_{i}^{\pm }=\left( {\bf \sigma }_{i}^{x}\mp i{\bf \sigma }_{i}^{y}\right) /2$. The first sum contains the qubit energies ($ \varepsilon _{i}^{z}$) and tunneling elements ($\varepsilon _{i}^{x}$) [@Leggett:87], and the second sum expresses tunneling between sites $i$ and $ j $. Other forms are also possible, e.g., as in an anisotropic dipolar medium such as in solid state NMR [@Slichter:96], where one would typically encounter an Ising $J_{ij}^{z}{\bf \sigma }_{i}^{z}{\bf \sigma } _{j}^{z}$ term. A Hamiltonian of the form above is sufficiently general to allow for universal quantum computation by satisfying the following two requirements [@DiVincenzo:95; @Lloyd:95; @Barenco:95a]: (i) Arbitrary single-qubit operations are made possible by the presence of ${\bf \sigma } _{i}^{x}$, which allows for the implementation of a continuous $SU(2)$ rotation in the $i^{{\rm th}}$ qubit Hilbert space, while the ${\bf \sigma } _{i}^{z}$ term allows for the introduction of an arbitrary phase-shift between the $|0\rangle $ and $|1\rangle $ states. When ${\bf \sigma }_{i}^{x} $ and ${\bf \sigma }_{i}^{z}$ are exponentiated, they can be combined, using the Lie sum and product formulae [@Bhatia] $$\begin{aligned} \lim_{n\rightarrow \infty }\left( e^{i\alpha A/n}e^{i\beta B/n}\right) ^{n}& =&e^{i(\alpha A+\beta B)} \nonumber \\ \lim_{n\rightarrow \infty }\left( e^{iA/\sqrt{n}}e^{iB/\sqrt{n}}e^{-iA/\sqrt{ n}}e^{-iB/\sqrt{n}}\right) ^{n} &=&e^{[A,B]}, \label{eq:Lie}\end{aligned}$$ to close the Lie algebra $su(2)$, and thus to construct any evolution in the Lie group $SU(2)$ of all possible operations on a single qubit [@Lloyd:95]. (ii) The second ingredient needed for universal quantum computation is the controlled-not ([CNOT]{}) gate, which is made possible through the ability to implement each of the (nearest neighbor) $J_{ij}{\ \sigma }_{i}^{+}{\sigma }_{j}^{-}+{\rm h.c.}$ terms. When exponentiated, such a term yields: $${\bf U}_{\theta }=\left( \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & \cos \theta & i\sin \theta & 0 \\ 0 & i\sin \theta & \cos \theta & 0 \\ 0 & 0 & 0 & 1 \end{array} \right)$$ with $\theta \propto J_{ij}t$. For $\theta =\pi /4$ this is (up to a phase) the “square-root-swap” operation, which when combined with single-qubit rotations allows for the implementation of [CNOT]{}. Alternatively, a $ J_{ij}^{z}{\bf \sigma }_{i}^{z}{\bf \sigma }_{j}^{z}$ term alone is sufficient, since it can be used to implement a controlled-phase-shift, as is done routinely in NMR [@Gershenfeld:97]. It is important to emphasize that the universal gates construction just described is but one of many different ways to achieve universal quantum computation. In fact, universal gates implementing logic operations directly on physical qubits (as above) are generally inappropriate for the purpose of [*fault tolerant*]{} computation [@Preskill:99]. We consider a different gate construction in the sequel paper [@Lidar:00b], operating instead on “encoded” qubits, which can be used to implement universal fault tolerant quantum computation. For a useful survey of different universal and fault tolerant sets of gates see Ref. [@Boykin:99]. The bath Hamiltonian can be written as $$\begin{aligned} {\bf H}_{B}=\sum_{k}\omega _{k}{\bf B}_{k}\end{aligned}$$ where, e.g., for the spin-boson Hamiltonian, ${\bf B}_{k}={\bf b} _{k}^\dagger {\bf b}_{k}$ [@Leggett:87], and ${\bf b}_{k}^{\dagger }$, $ {\bf b}_{k} $ are respectively creation and annihilation operators of bath mode $k$. Finally, the system-bath interaction Hamiltonian is $$\begin{aligned} {\bf H}_{I}=\sum_{i=1}^{K}\sum_{\alpha =+,-,z}\sum_{k}g_{ik}^{\alpha }\sigma _{i}^{\alpha }\otimes {\tilde{{\bf B}}}_{k}^{\alpha }, \label{eq:H_I1}\end{aligned}$$ where $g_{ik}^{\alpha }$ is a coupling coefficient. In the spin-boson model one would have ${\tilde{{\bf B}}}_{k}^{+}={\bf b}_{k}$, ${\tilde{{\bf B}}} _{k}^{-}={\bf b}_{k}^{\dagger }$ and ${\tilde{{\bf B}}}_{k}^{z}={\bf b} _{k}^{\dagger }+{\bf b}_{k}$. Thus $\sigma _{i}^{\pm }\otimes {\tilde{{\bf B} }}_{k}^{\pm }$ expresses a dissipative coupling (in which energy is exchanged between system and environment), and $\sigma _{i}^{z}\otimes {\ \tilde{{\bf B}}}_{k}^{z}$ corresponds to a phase damping process (in which the environment randomizes the system phases, e.g., through elastic collisions). An interesting limiting case arises when the coupling constants are independent of the qubit index: $g_{ik}^{\alpha }\equiv g_{k}^{\alpha }$. This situation, known as “collective decoherence”, arises when there is full permutational symmetry of qubit positions, and implies the existence of a large DFS [@Zanardi:97c; @Lidar:PRA99Exchange]. Defining collective system operators $S^{\alpha }\equiv \sum_{i=1}^{K}\sigma _{i}^{\alpha }$, one can then express the interaction Hamiltonian in greatly simplified form as $${\bf H}_{I}^{{\rm coll.}}=\sum_{\alpha =+,-,z}S^{\alpha }\otimes \left( \sum_{k}g_{k}^{\alpha }{\tilde{{\bf B}}}_{k}^{\alpha }\right) .$$ A case of intermediate symmetry arises when the coupling constants are equal not over the entire qubit register but rather only over finite clusters $ j=1..C$. One can then define cluster system operators $S_{j}^{\alpha }\equiv \sum_{i_{j}=1}^{K_{j}}\sigma _{i_{j}}^{\alpha },$ where $K_{j}$ is the number of qubits in cluster $j$. The interaction Hamiltonian becomes $${\bf H}_{I}^{{\rm clus.}}=\sum_{j=1}^{C}\sum_{\alpha =+,-,z}S_{j}^{\alpha }\otimes \left( \sum_{k}g_{jk}^{\alpha }{\tilde{{\bf B}}}_{k}^{\alpha }\right) .$$ In this case too, DFSs can be found. The point we wish to emphasize presently is that the underlying assumption in cluster decoherence is that of [*spatial symmetry*]{} in the system-bath coupling. This is to be contrasted with the decoherence models studied in this paper, where DFSs will be shown to arise without the need for spatial symmetry. Returning to the general case, ${\bf H}_{I}$ can be rewritten as $$\begin{aligned} {\bf H}_{I}=\sum_{i=1}^{K}\sum_{\alpha =x,y,z}\sum_{k}\sigma _{i}^{\alpha }\otimes {\bf B}_{ik}^{\alpha }, \label{eq:H_I}\end{aligned}$$ where ${\bf B}_{ik}^{z}\equiv {\tilde{{\bf B}}}_{k}^{z}$ and ${\bf B} _{ik}^{x}$ ,${\bf B}_{ik}^{y}$ are appropriate linear combinations of ${\ \tilde{{\bf B}}}_{k}^{+}$ and ${\tilde{{\bf B}}}_{k}^{-}$: $$\begin{aligned} {\bf B}_{ik}^{x} &=&\frac{1}{2}\left( g_{ik}^{-}{\tilde{{\bf B}}} _{k}^{-}+g_{ik}^{+}{\tilde{{\bf B}}}_{k}^{+}\right) \\ {\bf B}_{ik}^{y} &=&\frac{i}{2}\left( g_{ik}^{-}{\tilde{{\bf B}}} _{k}^{-}-g_{ik}^{+}{\tilde{{\bf B}}}_{k}^{+}\right)\end{aligned}$$ The qubit-coupling term in ${\bf H}_{S}$ can also be expressed entirely in terms of $\sigma _{i}^{\alpha }$, where $\alpha =x,y$ or $z$. Thus all system components of the Hamiltonian ${\bf H}$ can be expressed in terms of tensor products of the single qubit [*Pauli matrices*]{}. Time Evolution of the Density Matrix {#Evolution} ==================================== The purpose of this section is to show that the evolution of the density matrix of an open system can be expanded in terms of tensor products of the Pauli matrices (the Pauli group), and that this follows from the structure of the Hamiltonians assumed above for a qubit register. This result is obvious from a formal mathematical point of view (since the elements of the Pauli group of order $K$ form a complete orthogonal set for the $2^{K}\times 2^{K}$ matrices) [@Chuang:97c], so that the reader for whom this type of argument is satisfactory may safely skip ahead to the next section. We present the derivation of this result here in order to motivate the appearance of the multiple-qubit errors that are the subject of this paper. We first transform to the interaction picture [@Gardiner:book] defined by the system and bath Hamiltonians: $$\begin{aligned} {\bf H}\rightarrow {\bf H}(t)={\bf U}_{SB}(t){\bf HU}_{SB}^{\dagger }(t)= {\bf H}_{S}{\bf \otimes }{\bf I}_{B}+{\bf I}_{S}{\bf \otimes H}_{B}+{\bf H} _{I}(t)\end{aligned}$$ where $$\begin{aligned} {\bf U}_{SB}(t) &=&\exp \left[ -\left( {\bf H}_{S}{\bf \otimes }{\bf I}_{B}+ {\bf I}_{S}{\bf \otimes H}_{B}\right) it/\hbar \right] \\ &=&\exp \left[ -it{\bf H}_{S}/\hbar \right] {\bf \otimes }\exp \left[ -it {\bf H}_{B}/\hbar \right] ={\bf U}_{S}(t){\bf \otimes U}_{B}(t).\end{aligned}$$ Because the system and bath operators commute, the interaction picture interaction Hamiltonian can be written as: $$\begin{aligned} {\bf H}_{I}(t)={\bf U}_{SB}(t){\bf H}_{I}{\bf U}_{SB}^{\dagger }(t)=\sum_{i=1}^{K}\sum_{\alpha =x,y,z}\sum_{k}\sigma _{i}^{\alpha }(t)\otimes {\bf B}_{ik}^{\alpha }(t),\end{aligned}$$ where $$\begin{aligned} \sigma _{i}^{\alpha }(t) &=&{\bf U}_{S}(t)\sigma _{i}^{\alpha }{\bf U} _{S}^{\dagger }(t)=\sum_{j,\beta }\lambda _{ij}^{\alpha \beta }(t)\sigma _{j}^{\beta }, \nonumber \\ {\bf B}_{ik}^{\alpha }(t) &=&{\bf U}_{B}(t){\bf B}_{ik}^{\alpha }{\bf U} _{B}^{\dagger }(t) \label{eq:sigma_I}\end{aligned}$$ (see, e.g., Ref. [@Gardiner:book] for an explicit calculation of the $ \lambda _{ij}^{\alpha \beta }(t)$ for some examples). The system-bath density matrix is transformed accordingly from the Schrödinger into the interaction picture (denoted by a prime): $$\begin{aligned} \rho _{SB}(t)\longmapsto \rho _{SB}^{\prime }(t)={\bf U}_{SB}^{\dagger }(t)\rho _{SB}(t){\bf U}_{SB}(t),\end{aligned}$$ and the full dynamics is $$\begin{aligned} \rho _{SB}^{\prime }(t)={\bf U}(t)\rho _{SB}^{\prime }(0){\bf U}^{\dagger }(t),\end{aligned}$$ where $$\begin{aligned} {\bf U}(t)={\rm T}\exp \left[ -\frac{i}{\hbar }\int_{0}^{t}{\bf H} _{I}(\tau)d\tau \right]\end{aligned}$$ and ${\rm T}$ is the Dyson time-ordering operator (defined explicitly below). From now on we work in the interaction picture only, so for notational simplicity the prime is dropped from the density matrices. At $ t=0 $ the Schrödinger and interaction pictures coincide. Thus, assuming that system and bath are initially decoupled so that $\rho _{SB}(0)=\rho (0)\otimes \rho _{B}(0)$, where $\rho $ and $\rho _{B}$ are, respectively, the system and bath density matrices, the system dynamics is described by the reduced density matrix: $$\rho (0)\longmapsto \rho (t)=\text{{\rm Tr}}_{B}[{\bf U}(t)(\rho (0)\otimes \rho _{B}(0)){\bf U}^{\dagger }(t)].$$ Here [Tr]{}$_{B}$ is the partial trace over the bath. By using a spectral decomposition for the bath, $\rho _{B}(0)=\sum_{\nu }p_{\nu }|\nu \rangle \langle \nu |$,[^1] this can be rewritten in the “operator sum representation” [@Bacon:99; @Kraus:83; @Schumacher:96a; @Peres:99]: $$\begin{aligned} \rho (t)=\sum_{d}{\bf A}_{d}(t)\,\rho (0)\,{\bf A}_{d}^{\dagger }(t) \label{eq:OSR}\end{aligned}$$ where $$\begin{aligned} {\bf A}_{d}(t)=\sqrt{p_{\nu }}\langle \mu |{\bf U}(t)|\nu \rangle \;;\qquad d=(\mu ,\nu ) \label{eq:Amunu}\end{aligned}$$ Also, by unitarity of ${\bf U}$, one derives the normalization condition, $$\begin{aligned} \sum_{d}{\bf A}_{d}^{\dagger }{\bf A}_{d}={\bf I}_S \label{eq:OSRnorm}\end{aligned}$$ which guarantees preservation of the trace of $\rho $: $$\begin{aligned} {\rm Tr}[\rho (t)]={\rm Tr}[\sum_{d}{\bf A}_{d}\,\rho (0)\,{\bf A} _{d}^{\dagger }]={\rm Tr}[\rho (0)\sum_{d}{\bf A}_{d}^{\dagger }{\bf A} _{d}]= {\rm Tr}[\rho (0)].\end{aligned}$$ The $\{{\bf A}_{d}\}$, called the [*Kraus operators*]{}, belong to the (Banach, or Hilbert-Schmidt) space ${\cal B}({\cal H})$ of bounded operators acting on the system Hilbert space, and for $K$ qubits are represented by $ 2^{K}\times 2^{K}$ matrices, just like $\rho $.[^2] Consider now a formal Taylor expansion of the propagator: $$\begin{aligned} {\bf U}(t) &=&\sum_{n=0}^{\infty }\frac{\left( -i\right) ^{n}}{n!}{\rm T} \left( \int^{t}{\bf H}_{I}(\tau )d\tau \right) ^{n} \nonumber \\ &=&{\bf I}+\sum_{n=1}^{\infty }\frac{\left( -i\right) ^{n}}{n!} \int_{0}^{t}dt_{n}\int_{0}^{t}dt_{n-1}...\int_{0}^{t}dt_{1}{\rm T}\left\{ {\bf H}_{I}(t_{1})\cdots {\bf H}_{I}(t_{n})\right\} \nonumber \\ &\equiv &{\bf I}+\sum_{n=1}^{\infty }\frac{\left( -i\right) ^{n}}{n!}{\bf U} _{n}(t). \label{eq:U}\end{aligned}$$ The Dyson time-ordered product is defined with respect to any set of operators ${\bf O}_{i}(t_{i})$ as [@March:book] $${\rm T}\left\{ {\bf O}_{1}(t_{1})\cdots {\bf O}_{n}(t_{n})\right\} ={\bf O} _{\tau _{1}}(t_{\tau _{1}})\cdots {\bf O}_{\tau _{n}}(t_{\tau _{n}})\qquad \left( t_{\tau _{1}}>t_{\tau _{2}}>...>t_{\tau _{n}}\right) .$$ Using Eq. (\[eq:H\_I\]) we have for the terms in the above sum: $$\prod_{j=1}^{n}{\bf H}_{I}(t_{j})=\sum_{{\bf i}=1}^{K}\sum_{{ \mbox{\boldmath $\alpha$}}=x,y,z}\sum_{{\bf k}}\bigotimes_{j=1}^{n}\sigma _{i_{j}}^{\alpha _{j}}(t_{j})\bigotimes_{j=1}^{n}{\bf B}_{i_{j}k_{j}}^{\alpha _{j}}(t_{j}),$$ where ${\bf i}=\{i_{1},i_{2},...,i_{n}\}$, ${\mbox{\boldmath $\alpha$}} =\{\alpha _{1},\alpha _{2},...,\alpha _{n}\}$, and ${\bf k} =\{k_{1},k_{2},...,k_{n}\}$. The important point to notice in this complicated expression is that after taking the bath matrix elements $ \langle \mu |\cdots |\nu \rangle $ \[because of Eq. (\[eq:Amunu\])\], one is left with all possible tensor products $\bigotimes_{j=1}^{n}\sigma _{i_{j}}^{\alpha _{j}}(t_{j})$ over $n$ out of $K$ qubits. The integration and time-ordering operation will not change this conclusion. Thus, using, the expansion of $\sigma _{i}^{\alpha }(t)$ in Eq. (\[eq:sigma\_I\]), after a time $O(t^{K})$ one finds the tensor product $\bigotimes_{j=1}^{K}\sigma _{i_{j}}^{\alpha _{j}}$, i.e., [*all*]{} qubits are involved (here $\alpha _{j}=0$, corresponding to the identity matrix, is allowed). At this point the entire [*Pauli group*]{} $P_{K}$ appears (all possible $4^{K+1}$ tensor products of the $3$ Pauli matrices and the identity matrix, and the four roots of unity {$\pm ,\pm i$} – see Appendix \[app1\]), and one has “complete decoherence”, i.e., multiple-qubit errors over the entire system Hilbert space. In the usual approach to QECC one does not consider such high orders in time since one assumes that error correction can be done quickly enough. Instead the error-analysis is usually confined to time evolution to $ O(t)$ only, which leads to “independent decoherence”, i.e., single-qubit errors affecting only one qubit at a time.[^3] It is possible to use multiple-error-correcting quantum codes for $O(t^{n})$ with arbitrary $n$, but these codes are rather unwieldy (i.e., the number of encoding qubits becomes large). In the case of “burst errors” (a spatially contiguous cluster of errors such as $I\cdots IX\cdots XI\cdots I$) some particularly efficient codes are known [@Bandyopadhyay:98]. On the other hand, a DFS that exists by virtue of a spatially symmetric system-bath coupling, is not affected by this proliferation of errors, which all occur in the subspace orthogonal to the DFS [@Lidar:PRL99]. The assumption of spatial symmetry manifests itself in restrictions on the coefficients $g_{ik}^{\alpha }$ appearing in the interaction Hamiltonian \[Eq. $~$(\[eq:H\_I1\])\]. For example, as mentioned above, collective decoherence corresponds to the condition $g_{ik}^{\alpha }=g_{k}^{\alpha }$ $ \forall i$, i.e., the bath cannot distinguish between the qubits [@Zanardi:97c]. In this paper no such spatial symmetry assumptions will be made. Instead, only [*multiple-qubit*]{} errors will be allowed to lowest order instead of single-qubit errors. This condition will be defined more precisely in the next section. As for the Kraus operators, it can be seen from the calculations above that they may be expanded as sums over tensor products of the Pauli matrices: $$\begin{aligned} {\bf A}_{d}(t)=\sum_{n=1}^{4^{K+1}}a_{d,n}(t){\bf p}_{n} \label{eq:Ai}\end{aligned}$$ where ${\bf p}_{n}\in P_{K}$. The Kraus operators thus belong to the [ *group algebra*]{} (the space of linear combinations of group elements) of $ P_{K} $ [@Schensted]. As alluded to in the beginning of this section, that this expansion is possible actually follows simply from the fact that the Pauli group forms a complete orthogonal set (with respect to the trace inner product) for the expansion (with complex coefficients) of arbitrary $ 2^{K}\times 2^{K}$ matrices. However, here we have seen how the expansion in terms of the Pauli group (rather than some other basis) is physically motivated by virtue of the structure of the Hamiltonian. A simple example will now serve to illustrate the point made above about multiple-qubit errors. Consider an interaction Hamiltonian of the form ${\bf H}_{I}=\sum_{i=1}^{2}\sigma _{i}^{z}\otimes B_{i}$ (on two qubits). Some algebra suffices to show that then ${\bf A}_{d}(t)=c_{0}(t){\bf I} _{S}+c_{1}(t)\sigma _{1}^{z}+c_{2}(t)\sigma _{2}^{z}+c_{12}(t)\sigma _{1}^{z}\otimes \sigma _{2}^{z}$. In this case the single-qubit errors $ \sigma _{1}^{z},\sigma _{2}^{z}$ appear, as well as the multiple-qubit error $\sigma _{1}^{z}\otimes \sigma _{2}^{z}$. This situation does not allow for the appearance of DFSs (unless spatial symmetry is present). Alternatively, consider the interaction Hamiltonian ${\bf H}_{I}=(\sigma _{1}^{z}\otimes \sigma _{2}^{z})\otimes B_{12}+(\sigma _{3}^{z}\otimes \sigma _{4}^{z})\otimes B_{34}$ (on four qubits). In this case one finds ${\bf A} _{d}(t)=c_{0}(t){\bf I}_{S}+c_{12}(t)\sigma _{1}^{z}\otimes \sigma _{2}^{z}+c_{34}(t)\sigma _{3}^{z}\otimes \sigma _{4}^{z}+c_{1234}(t)\sigma _{1}^{z}\otimes \sigma _{2}^{z}\otimes \sigma _{3}^{z}\otimes \sigma _{4}^{z} $. Thus only multiple-qubit terms appear, and as will be shown below, this allows for the existence of non-trivial DFSs, even though no spatial symmetry assumptions were made. An important example of this correlated type of system-bath interaction is the dipolar-coupling Hamiltonian, relevant, e.g., to decoherence resulting from spin-rotation coupling in NMR [@Slichter:96].[^4] The dipolar Hamiltonian for a system of spins interacting with a bath of rotations is: $$H_{I}=\sum_{j,k}\frac{\gamma _{j}\gamma _{k}}{r_{jk}^{3}}\left[ {{ \mbox{\boldmath $\sigma$}}}_{j}\cdot {{\mbox{\boldmath $\sigma$}}} _{k}-3\left( {{\mbox{\boldmath $\sigma$}}}_{j}\cdot {\bf r}_{jk}\right) \left( {{\mbox{\boldmath $\sigma$}}}_{k}\cdot {\bf r}_{jk}\right) \right] ,$$ where $\gamma _{j}$ is the gyromagnetic ration of spin $j$, $r_{jk}$ is the distance beween spins $j$ and $k$, and ${\mbox{\boldmath $\sigma$}}$ is the vector of Pauli matrices. Introducing an anistropy tensor $g_{jk}^{\alpha \beta }$, this can be rewritten as: $$H_{I}=\sum_{j,k}\frac{\gamma _{j}\gamma _{k}}{r_{jk}^{3}}\sum_{\alpha ,\beta =-1}^{1}g_{jk}^{\alpha \beta }\left( \sigma _{j}^{\alpha }\otimes \sigma _{k}^{\beta }\right) Y_{2}^{-\alpha -\beta }, \label{eq:dipolar}$$ where $Y_{l}^{m}$ are the spherical harmonics, and $\sigma ^{0}\equiv \sigma ^{z}$. Eventhough only multiple-qubit terms appear here it is necessary to further impose anisotropy in order to obtain an example with a non-trivial DFS, as we discuss in more detail in Sec. \[Q2Z\]. This is the case, e.g., when only $\sigma _{j}^{z}\otimes \sigma _{k}^{z}$ terms remain (i.e., $ g_{jk}^{\alpha \beta }=\delta _{\alpha 0}\delta _{\beta 0}g_{jk}$), coupled to $Y_{2}^{0}$ rotations. With these observations, we are now ready to study the question of DFSs in open systems without spatial symmetry in the system-bath couplings. Decoherence-Free Subspaces from Subgroups of the Pauli Group {#DFS} ============================================================ We begin this section by recalling the condition for DFSs within the framework of the Kraus operator-sum representation, derived in Ref. [@Lidar:PRL99]. We then analyze the conditions for the appearance of DFSs when the errors are spanned by a subgroup of the Pauli group. The result is summarized by a theorem presented at the end of the section. Condition for Decoherence-Free Subspaces ---------------------------------------- A DFS is a subspace $\tilde{{\cal H}}={\rm Span}\{|\tilde{j}\rangle \}$ of the full system Hilbert space ${\cal H}_{K}$ over which the evolution of the density matrix is unitary. Necessary and sufficient conditions for a DFS were derived in the Markovian case in Ref. [@Lidar:PRL98] and in the exact (non-Markovian) case in Ref. [@Zanardi:97a]. A formulation of the exact DFS condition was given in terms of the operator sum representation in Ref. [@Lidar:PRL99], and will be briefly reviewed. Let $\{|\tilde{j}\rangle \}$ be a set of system states satisfying: $${\bf A}_{d}|\tilde{j}\rangle =c_{d}{\tilde{{\bf U}}}|\tilde{j}\rangle \qquad \forall d, \label{eq:DFS-cond}$$ where $\tilde{{\bf U}}$ is an arbitrary, $d$-independent but possibly time-dependent unitary transformation, and $c_{d}$ a complex constant. Under this condition, an initially pure state belonging to ${\rm Span}[\{|\tilde{j} \rangle \}]$, $$|\psi _{{\rm in}}\rangle =\sum_{j}\gamma _{j}|\tilde{j}\rangle ,$$ will be [*decoherence-free*]{}, since: $$|\phi _{d}\rangle ={\bf A}_{d}|\psi _{{\rm in}}\rangle =\sum_{j}\gamma _{j}c_{d}{\tilde{{\bf U}}}|\tilde{j}\rangle =c_{d}{\tilde{{\bf U}}}|\psi _{ {\rm in}}\rangle$$ so $$\rho _{{\rm out}}=\sum_{d}{\bf A}_{d}\tilde{\rho}_{{\rm in}}{\bf A} _{d}^{\dagger }=\sum_{d}c_{d}{\tilde{{\bf U}}}|\psi _{{\rm in}}\rangle \langle \psi _{{\rm in}}|{\tilde{{\bf U}}}^{\dagger }c_{d}^{\ast }={\tilde{ {\bf U}}}|\psi _{{\rm in}}\rangle \langle \psi _{{\rm in}}|{\tilde{{\bf U}}} ^{\dagger },$$ where we used the normalization of the Kraus operators \[Eq. (\[eq:OSRnorm\] )\] to set $\sum_{d}|c_{d}|^{2}=1$. This means that the time-evolved state $ \rho _{{\rm out}}$ is pure, and its evolution is governed by ${\tilde{{\bf U} }}$. This argument is easily generalized to an initial mixed state $\tilde{ \rho}_{{\rm in}}=\sum_{jj^{\prime }}\rho _{jj^{\prime }}|\tilde{j}\rangle \langle \tilde{j}^{\prime }|$, in which case $\rho _{{\rm out}}={\tilde{{\bf U}}}\tilde{\rho}_{{\rm in}}{\tilde{{\bf U}}}^{\dagger }$. The unitary transformation ${\ \tilde{{\bf U}}}$ is a “gauge freedom” which can be exploited in choosing a driving system Hamiltonian which implements a useful evolution on the DFS. In the interaction picture used in the previous section, ${\tilde{{\bf U}}}$ can be made to disappear by redefining all Kraus operators as ${\tilde{{\bf U}}}^{\dagger }{\bf A}_{d}$. The calculation above shows that Eq. (\[eq:DFS-cond\]) is a sufficient condition for a DFS. It follows from the results of Refs. [@Zanardi:97a; @Nielsen:97] that it is also a necessary condition for a DFS (under “generic” conditions – to be explained below). Eq. (\[eq:DFS-cond\]), however, does not seem to be a very useful characterization of a DFS if one does not know the explicit form of the Kraus operators (in general this cannot be found in closed analytical form, although they can be determined experimentally [@Chuang:97c]). When the Kraus operators derive from a Hamiltonian, as in Eq. (\[eq:Amunu\]), an equivalent DFS condition is [@Lidar:PRL99]: $$\begin{aligned} {\bf S}_{\alpha }|\tilde{j}\rangle =a_{\alpha }|\tilde{j}\rangle \qquad \forall \alpha ,\end{aligned}$$ where the system-bath interaction Hamiltonian is written as ${\bf H} _{I}=\sum_{\alpha }{\bf S}_{\alpha }\otimes {\bf B}_{\alpha }$ \[compare to Eq. (\[eq:H\_I\])\], with $\{{\bf S}_{\alpha }\}$ being the system operators. To make use of this last DFS condition, one needs to introduce assumptions about the structure of system-bath coupling, and this is how one is led to spatial symmetry considerations [@Lidar:PRL98]. Here, however, the DFS condition of Eq. (\[eq:DFS-cond\]) will be considered directly, based purely on the expansion of the Kraus operators in terms of the Pauli group elements, and without resorting to an explicit form for these operators. Representation Theory Construction of Decoherence-Free States {#reptheory} ------------------------------------------------------------- When the Kraus operators are viewed as operators in the algebra of the Pauli group, the DFS condition \[Eq. (\[eq:DFS-cond\])\] has a natural interpretation: the decoherence free states $\{|\tilde{j}\rangle \}$ belong to the one-dimensional irreps of the Pauli group. Motivated by this observation we now consider a group representation theory construction of decoherence-free states. The general criterion for the reducibility of a representation $\{\Gamma (G_{n})\}_{n=1}^{N}$of a finite group ${\cal G}=\{G_{n}\}$ of order $N$ is [@Cornwell:97]: $$\begin{aligned} \sum_{n=1}^{N}|\chi \lbrack \Gamma (G_{n}) \rbrack |^{2}>N,\end{aligned}$$ where $\chi $ is the character of the representation $\Gamma $ \[trace of the matrix $\Gamma (G_{n})$\]. If equality holds, then the representation is irreducible. The full Pauli group $P_{K}$ is irreducible over the Hilbert space ${\cal H} _{K}$ of $K$ qubits: since all Pauli matrices are traceless, only the four elements proportional to the identity matrix contribute (see also Appendix \[app1\]): $$\sum_{n=1}^{4^{K+1}}|\chi \lbrack {\bf p} _{n}]|^{2}=|2^{K}|^{2}+|-2^{K}|^{2}+|i2^{K}|^{2}+|-i2^{K}|^{2}=4^{K+1},$$ which is just the order of $P_{K}$ (generally the direct product representation of irreps of any direct product group is itself an irrep of that group [@Cornwell:97]). Now we come to the central assumption setting the stage for the DFSs considered in this paper: [*what if the Kraus operators belong to the group algebra of a [**subgroup**]{}* ]{}$Q$[* of* ]{}$P_{K}$? The motivation for this situation could be: (i) The case in which either only higher order errors occur, such that first-order terms of the form $I\otimes \cdots \otimes I\otimes \sigma _{i}^{\alpha }\otimes I\otimes \cdots \otimes I$ are absent in the Pauli-group expansion of the Kraus operators, or (ii) only errors of [*one*]{} kind, either $\sigma ^{x}$, $\sigma ^{y}$, or $\sigma ^{z}$ take place. Case (i) would imply that either: (a) There are certain cancellations involving bath matrix element terms such that first-order system operators are absent in the expansion of Eq. (\[eq:U\]). This would be a rather non-generic situation, involving a very special “friendly” bath; or (b) The system-bath Hamiltonian is in fact not of the form in Eq. ( \[eq:H\_I\]), but rather involves only second order terms such as $\sigma _{i}^{\alpha }\otimes \sigma _{j}^{\beta }$ (identity on all the rest). [^5] Case (ii) is applicable in, e.g., the case of pure phase damping (relevant to NMR [@Slichter:96]) and optical lattices using cold controlled collisions [@Jaksch:99]), where $\sigma ^{z}$ errors are dominant. In the subgroup case under consideration, we may find non-trivial irreducible representations (irreps) of $Q$ over ${\cal H}_{K}$ (a so-called “subduced” representation [@Schensted]). This situation can be interesting especially if there exist 1-dimensional irreps, as known from the general theory of DFSs [@Zanardi:97a; @Lidar:PRL98]. As will be shown next, the recipe for finding these DFSs uses the standard projection operators from elementary group representation theory. The projection is onto the subspace transforming according to a particular irrep. First, recall the multiplicity formula for unitary irreps (which we can always assume in this case since the Pauli group is finite): $$\begin{aligned} m_{k}=\frac{1}{N}\sum_{n=1}^{N}\chi \left[ \Gamma ^{k}(G_{n})\right] ^{\ast }\chi \left[ \Gamma (G_{n})\right] , \label{eq:mk}\end{aligned}$$ where $m_{k}$ is the number of times irrep $\Gamma ^{k}$ appears in the given reducible representation; $\chi \left[ \Gamma ^{k}(G_{n})\right] $ is the character of the $\Gamma ^{k}$ irrep on the group element $G_{n}$; and $ \chi \left[ \Gamma (G_{n})\right] $ is the character of $G_{n}$ in the given reducible representation $\Gamma $. We denote a set of (orthonormal) basis-states transforming according to an irrep $\Gamma^k$ by $\{|\psi_1^k \rangle, \ldots ,|\psi_{d_k}^k \rangle \}$. These states span the invariant subspace of the irrep $\Gamma^k$ and transform according to: $$\begin{aligned} G_{n}|\psi _{\mu }^{k}\rangle =\sum_{\nu =1}^{d_{k}}\Gamma ^{k}(G_{n})_{\nu \mu }|\psi _{\nu }^{k}\rangle . \label{eq:trafo}\end{aligned}$$ Furthermore they obey the orthogonality relation: $$\begin{aligned} \langle \psi _{\mu }^{l}|\psi _{\nu }^{k}\rangle =\delta _{lk}\delta _{\mu \nu } \label{eq:ortho}\end{aligned}$$ Next, a projection operator onto the subspace belonging to the $d_{k}$ -dimensional irrep $k$ is given by the appropriate sum over group elements [@Cornwell:97]: $$\begin{aligned} P_{\mu \nu }^{k}=\frac{d_{k}}{N}\sum_{n=1}^{N}\Gamma ^{k}(G_{n})_{\mu \nu }^{\ast }G_{n}\;;\qquad \mu ,\nu =1,...,d_{k} \label{eq:proj}\end{aligned}$$ and has the following properties: $$\begin{aligned} P_{\mu \nu }^{k}P_{\kappa \lambda }^{l} &=&\delta _{kl}\delta _{\nu \kappa }P_{\mu \lambda }^{k} \nonumber \\ P_{\mu \nu }^{l}|\psi _{\lambda }^{k}\rangle &=&\delta _{kl}\delta _{\nu \lambda }|\psi _{\mu }^{k}\rangle {.}\end{aligned}$$ To obtain a set of (orthonormal) basis states $\{|\psi _{1}^{k}\rangle,\ldots ,|\psi _{d_{k}}^{k}\rangle \}$ transforming as a set of partners in the basis for $\Gamma ^{k}$ from an arbitrary state $|\phi \rangle $, one can apply the set of operators $\{P_{\mu \nu }^{k}\}$ for a fixed $\nu $ (such that $P_{\nu \nu }^{k}|\phi \rangle \neq 0$) and renormalize the states thus obtained. Every state $|\phi \rangle $ can be expanded in terms of basis states for the constituting irreps $\Gamma ^{k}$ as $$\begin{aligned} |\phi \rangle =\sum_{k}\sum_{\nu =1}^{d_{k}}\theta _{\nu }^{k}|\psi _{\nu }^{k}\rangle , \label{eq:expand}\end{aligned}$$ where $P_{\nu \nu }^{k}|\phi \rangle =\theta _{\nu }^{k}|\psi _{\nu }^{k}\rangle $ and the summation over $k$ is over inequivalent irreps [@Cornwell:97]. Let us now consider the effect of applying the operators ${\bf A} _{d}=\sum_{n}a_{d,n}G_{n}$ from the group algebra to an arbitrary state $ |\phi \rangle $: $$\begin{aligned} {\bf A}_{d}|\phi \rangle =\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{k} {\bf A}_{d}|\psi _{\mu }^{k}\rangle =\sum_{n=1}^{N}a_{d,n}\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{k}\sum_{\nu =1}^{d_{k}} \Gamma ^{k}(G_{n}) _{\nu \mu}|\psi _{\nu}^{k}\rangle . \label{eq:Ai-proj}\end{aligned}$$ We would like to find the conditions such that this transforms into the DFS condition, Eq. (\[eq:DFS-cond\]). Consider the case when $\Gamma ^{k}$ are all [*1-dimensional irreps*]{}, possibly appearing with multiplicity $m_{k}$: $$\begin{aligned} \Gamma ^{k}(G_{n})_{\mu \nu }=\gamma _{n}^{k} \qquad \mu ,\nu =1. \label{eq:diag}\end{aligned}$$ In this case the indices $\mu ,\nu $ are irrelevant and we will omit them. Then: $$\begin{aligned} {\bf A}_{d}|\phi \rangle =\sum_{n=1}^{N}a_{d,n}\sum_{k}\gamma _{n}^{k}\theta ^{k}|\psi ^{k}\rangle .\end{aligned}$$ For $|\phi \rangle $ to be a decoherence free state, one would like to have this proportional to $|\phi \rangle =\sum_{k}\theta ^{k}|\psi ^{k}\rangle $ \[as in the original expansion of Eq. (\[eq:expand\])\]. However, this does not work because of the presence of $\gamma _{n}^{k}$ in the sum. We thus see that the initial function $|\phi \rangle $ must be restricted to be one of the basis-states $|\psi ^{k}\rangle $. Then, with $$\begin{aligned} c_{d}^{k}\equiv \sum_{n=1}^{N}a_{d,n}\gamma _{n}^{k}. \label{eq:c_d}\end{aligned}$$ we have finally: $$\begin{aligned} {\bf A}_{d}|\psi ^{k}\rangle =c_{d}^{k}|\psi ^{k}\rangle . \label{eq:phiDFS}\end{aligned}$$ At this point it is useful to introduce another index $z$ for the multiplicity of the irrep $k$, i.e. $z=1\ldots m_{k}$. The Hilbert space of $ K$-qubit states splits into invariant one-dimensional subspaces $V_{z}^{k}$ that are spanned by (fixed) basis states $|\psi _{z}^{k}\rangle $. Each of the $|\psi ^{k}\rangle $ in Eq. (\[eq:expand\]) is a linear combination of the $|\psi _{z}^{k}\rangle $: $$\begin{aligned} |\psi ^{k}\rangle =\sum_{z=1}^{m_{k}}\theta _{z}^{k}|\psi _{z}^{k}\rangle , \label{eq:sameirrep}\end{aligned}$$ \[Because of Eq. (\[eq:expand\]), the $\theta _{z}^{k}$ depend on the initial state $|\phi \rangle $.\] Thus for $|\phi \rangle $ to be a decoherence-free state, it is allowed to be an arbitrary superposition inside copies of a [*given*]{} irrep (different $z$’s), but not to be a superposition between different irreps (different $k$’s). In particular we have within each copy of the irrep $\Gamma ^{k}$: $$\begin{aligned} {\bf A}_{d}|\psi _{z}^{k}\rangle =c_{d}^{k}|\psi _{z}^{k}\rangle \;;\qquad z=1..m_{k}. \label{eq:cik-DFS}\end{aligned}$$ This is just the DFS condition, Eq. (\[eq:DFS-cond\]) with the $\{|\psi _{z}^{k}\rangle \}$ being the basis states for the DFS. Therefore Eq. (\[eq:diag\]) is a [*sufficient* ]{}condition for a DFS, provided that our initial state satisfies the condition that it is a superposition of states within a [*fixed*]{} irrep, Eq. (\[eq:sameirrep\]). It will now be shown that Eq. (\[eq:diag\]) is also a [*necessary* ]{} condition for a DFS under the “genericity” assumption that the error coefficients $\{a_{d,n}\}$ are arbitrary. In other words, it will be shown under these conditions that, if a set of basis-states $\{|\tilde{j}\rangle \} $ satisfies the DFS-condition Eq. (\[eq:DFS-cond\]) then the $\{|\tilde{j} \rangle \}$ belong to the invariant subspace of some one-dimensional irrep of our subgroup. Assume that the ${\bf A}_{d}$ have been redefined to incorporate the (constant) unitary transformation $\tilde{{\bf U}}$ such that Eq. (\[eq:DFS-cond\]) becomes ${\bf A}_{d}|\tilde{j}\rangle =c_{d}|\tilde{j}\rangle $. Expand the state $|\tilde{j}\rangle $ as in Eq. (\[eq:expand\]): [^6] $$\begin{aligned} |\tilde{j}\rangle =\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j} ,k}|\psi _{\mu }^{k}\rangle \label{eq:expandj}\end{aligned}$$ where $P_{\mu \mu }^{k}|\phi \rangle =\theta _{\mu }^{\tilde{j},k}|\psi _{\mu }^{k}\rangle $. Now using Eq. (\[eq:Ai-proj\]): $$\begin{aligned} {\bf A}_{d}|\tilde{j}\rangle &=&c_{d}|\tilde{j}\rangle =c_{d}\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j},k}|\psi _{\mu }^{k}\rangle \\ &=&\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j},k}{\bf A}_{d}|\psi _{\mu }^{k}\rangle =\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j} ,k}\sum_{n}a_{d,n}G_{n}|\psi _{\mu }^{k}\rangle \nonumber \\ &=&\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j},k}\sum_{n}a_{d,n} \sum_{\lambda =1}^{d_{k}}\Gamma ^{k}(G_{n})_{\lambda \mu }|\psi _{\lambda }^{k}\rangle \label{eq:long1}\end{aligned}$$ and taking inner products \[using (\[eq:ortho\])\] $$\begin{aligned} \langle \psi _{\sigma }^{l}|{\bf A}_{d}|\tilde{j}\rangle &=&c_{d}\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j},k}\langle \psi _{\sigma }^{l}|\psi _{\mu }^{k}\rangle =c_{d}\theta _{\sigma }^{\tilde{j },l} \nonumber \\ &=&\sum_{k}\sum_{\mu =1}^{d_{k}}\theta _{\mu }^{\tilde{j},k}\sum_{n}a_{d,n} \sum_{\lambda =1}^{d_{k}}\Gamma ^{k}(G_{n})_{\lambda \mu }\langle \psi _{\sigma }^{l}|\psi _{\lambda }^{k}\rangle =\sum_{\mu =1}^{d_{l}}\theta _{\mu }^{\tilde{j},l}\sum_{n}a_{d,n}\Gamma ^{l}(G_{n})_{\sigma \mu }. \label{eq:long}\end{aligned}$$ Using this result we would like to show that the $\Gamma ^{l}(G_{n})$ that appear here must be one-dimensional irreps. Let us establish “generic” conditions for this purpose. Eq. (\[eq:long\]) can be rewritten as an eigenvalue equation: $$\begin{aligned} {\cal A}_{d}^{l}\overrightarrow{\theta _{j}^{l}}=c_{d}\overrightarrow{\theta _{j}^{l}}, \label{eq:eigen}\end{aligned}$$ where $$\begin{aligned} {\cal A}_{d}^{l} &\equiv &\sum_{n}a_{d,n}\Gamma ^{l}(G_{n}) \\ \overrightarrow{\theta _{j}^{l}} &\equiv &(\theta _{1}^{\tilde{j},l},\ldots ,\theta _{d_{l^{\prime }}}^{\tilde{j},l}).\end{aligned}$$ The vector $\overrightarrow{\theta _{j}^{l}}$ may be zero for a given irrep $ \Gamma ^{l}$, in which case Eq. (\[eq:eigen\]) is trivially satisfied. Let us assume this is not the case for some $l$ \[it cannot be the case for [ *all*]{} $l$, by Eq. (\[eq:expandj\])\]. Then the most general way in which Eq. (\[eq:eigen\]) can be satisfied, is for $\overrightarrow{\theta _{j}^{l}}$ to be an eigenvector of ${\cal A}_{d}^{l}$ for all codewords $| \tilde{j}\rangle $, with eigenvalue $c_{d}$. However, while this is the most general condition, it is [*non-generic*]{}. By [*generic*]{} we mean that we take the errors to be [*arbitrary*]{}, i.e., we do not want to make any assumptions on the $a_{d,n}$. Now, if the eigenvalue eqnarray were to be satisfied, the vector of coefficients $\overrightarrow{\theta _{j}^{l}}$ would have to be “special”. In other words, [*it would have to be adjusted to be an eigenvector of*]{} ${\cal A}_{d}^{k}$. To make this adjustment would require two conditions: (i) Having a priori knowledge of the $a_{d,n}$, (ii) Being able to control $\overrightarrow{\theta _{j}^{l}}$ . We would like to avoid assuming (i) because fine-tuning the bath is physically unacceptable. In contrast, control of $\overrightarrow{\theta _{j}^{l}}$ is certainly desirable. However, we would like to avoid the situation where only certain special choices of $\overrightarrow{\theta _{j}^{l}}$, compatible with specific bath parameters, yield decoherence free states $|\tilde{j}\rangle $.[^7] We thus conclude that to avoid fine-tuning of the bath parameters and/or special initial conditions, ${\cal A}_{d}^{l}$ must be proportional to the identity. But since $\Gamma ^{l}$ is an irrep this is only possible if it is one-dimensional, i.e. $\Gamma ^{l}(G_{n})_{\mu \nu }=\gamma _{n}^{l}$, $\mu ,\nu =1$ and $ c_{d}=\sum_{n}a_{d,n}\gamma _{n}^{l}$. In addition we see that $c_{d}$ can only be $l$-independent if the DFS states $|\tilde{j}\rangle $ are spanned [*only*]{} by basis states of copies of the [*same*]{} irrep $\Gamma ^{l}$. Q.E.D. To summarize: [*Theorem 1*]{}: Suppose that the Kraus operators belong to the group algebra of some group ${\cal G}=\{G_{n}\}$, i.e., ${\bf A} _{d}=\sum_{n=1}^{N}a_{d,n}G_{n}$. If a set of states $\{|\tilde{j}\rangle \}$ belong to a given [*one-dimensional*]{} irrep $\Gamma ^{k}$ of ${\cal G}$, then the DFS condition ${\bf A}_{d}|\tilde{j}\rangle =c_{d}|\tilde{j}\rangle $ holds. If no assumptions are made on the bath coefficients $\{a_{d,n}\}$, then the DFS condition ${\bf A}_{d}|\tilde{j}\rangle =c_{d}|\tilde{j}\rangle $ implies that $|\tilde{j}\rangle $ belongs to a [*one-dimensional*]{} irrep $\Gamma ^{k}$ of ${\cal G}$. For completeness we give in Appendix \[app2\] an example of the “non-generic DFSs”, which result from “accidentally” satisfying Eq. (\[eq:eigen\]) with irreps of dimension greater than one. Examples of Subgroups with Decoherence Free States {#Examples} ================================================== The general considerations from the previous section will now be illustrated with some examples. To simplify the notation, let $X,Y,Z$ represent the $ \sigma ^{x},\sigma ^{y},\sigma ^{z}$ Pauli matrices, and let us drop the tensor product symbol (i.e., let $ZI\equiv Z\otimes I$, $X^{2}\equiv X\otimes X$, etc.). Also, we will ignore normalization factors in this section. Abelian Subgroups ----------------- The simplest non-trivial example of a subgroup is found already for $K=2$ qubits: $$\begin{aligned} Q_{Z}=\{I^{2},ZI,IZ,Z^{2}\}.\end{aligned}$$ This subgroup (generated by $ZI$ and $IZ$) describes phase damping. As another simple example, let $K=4$ qubits and consider the following subgroup: $$Q_{X}=\{I^{4},X^{2}I^{2},I^{2}X^{2},X^{4}\}.$$ Physically, this would correspond to the error process where bit flips happen on certain clusters of two or four qubits only (note that $XIXI$ and $ IXIX$ were left out – this case will be considered in the sequel paper [@Lidar:00b]). Another example is $$Q_{4}=\{I^{4},X^{4},Y^{4},Z^{4}\},$$ with all Pauli errors occurring just on clusters of 4 qubits. $Q_{Z}$, $Q_{X} $ and $Q_{4}$ are isomorphic and Abelian. All elements of these subgroups, except $I^{4}$, are traceless. $I^{4}$ has trace 16, so that $ \sum_{n=1}^{4}|\chi \lbrack \Gamma (G_{n})]|^{2}=256>4$ and thus the natural representation of these subgroups on $4$ qubits is reducible. Since they are Abelian, they have only 1-dimensional irreps. These irreps are given in the following table, expressed in terms of the elements of $Q_{X}$: $$\begin{aligned} \begin{tabular}{|r||r|r|r|r|} \hline & $I^{4}$ & $X^{2}I^{2}$ & $I^{2}X^{2}$ & $X^{4}$ \\ \hline\hline $\Gamma ^{1}$ & 1 & 1 & 1 & 1 \\ \hline $\Gamma ^{2}$ & 1 & 1 & -1 & -1 \\ \hline $\Gamma ^{3}$ & 1 & -1 & 1 & -1 \\ \hline $\Gamma ^{4}$ & 1 & -1 & -1 & 1 \\ \hline \end{tabular} \label{eq:mult-table}\end{aligned}$$ Motivated by Theorem 1, this reducibility implies the existence of DFSs, as long as the Kraus operators belong to the group algebra of these subgroups. ### The Subgroup $Q_{X}$ Consider the case of $Q_{X}$, i.e., assume that the Kraus operators can be written as $$\begin{aligned} {\bf A}_{d}=a_{d,0}I^{4}+a_{d,1}X^{2}I^{2}+a_{d,2}I^{2}X^{2}+a_{d,3}X^{4} \label{eq:generalA}\end{aligned}$$ \[the coefficients $a_{d,j}$ are of course constrained by the normalization condition Eq. (\[eq:OSRnorm\])\]. Using the general arguments of Sec. \[reptheory\] and in particular Eq. ( \[eq:proj\]), we can just read off the matrix elements of the 4 (1-dimensional) irreps from the table in Eq. (\[eq:mult-table\]). Thus the 4 projection operators are: $$\begin{aligned} P^{1} &=&I^{4}+X^{2}I^{2}+I^{2}X^{2}+X^{4}\qquad P^{2}=I^{4}+X^{2}I^{2}-I^{2}X^{2}-X^{4} \nonumber \\ P^{3} &=&I^{4}-X^{2}I^{2}+I^{2}X^{2}-X^{4}\qquad P^{4}=I^{4}-X^{2}I^{2}-I^{2}X^{2}+X^{4}\end{aligned}$$ The multiplicity of each of the four 1-dimensional irreps in the reducible representation generated here by the $K=4$ qubits, is 4. To see this, recall the multiplicity formula Eq. (\[eq:mk\]). In the present case, the given representation yields $\chi =\{16,0,0,0\}$ (for $ I^{4},X^{2}I^{2},I^{2}X^{2},X^{4}$ respectively) and so, with $\chi ^{k}(I^{4})=1$, $m_{k}=\frac{1}{4}\chi ^{k}(I^{4})16=4$ for all $k$. Now, let us explicitly find the decoherence free states. To do so we can pick an arbitrary, convenient 4-qubit state and project it onto a given irrep. For example, starting with $|0000\rangle $: $$\begin{aligned} P^{1}|0000\rangle &=&|0000\rangle +|1100\rangle +|0011\rangle +|1111\rangle \equiv |\psi _{1}^{1}\rangle \nonumber \\ P^{2}|0000\rangle &=&|0000\rangle +|1100\rangle -|0011\rangle -|1111\rangle \equiv |\psi _{1}^{2}\rangle \nonumber \\ P^{3}|0000\rangle &=&|0000\rangle -|1100\rangle +|0011\rangle -|1111\rangle \equiv |\psi _{1}^{3}\rangle \nonumber \\ P^{4}|0000\rangle &=&|0000\rangle -|1100\rangle -|0011\rangle +|1111\rangle \equiv |\psi _{1}^{4}\rangle .\end{aligned}$$ Each of these 4 states belongs to a different irrep, and thus to a different DFS, which can be verified by applying an arbitrary Kraus operator, as in Eq. (\[eq:generalA\]). E.g., $$\begin{aligned} {\bf A}_{d}|\psi _{1}^{1}\rangle &=&a_{d,0}\left( |0000\rangle +|1100\rangle +|0011\rangle +|1111\rangle \right) \nonumber \\ &&+a_{d,1}(|1100\rangle +|0000\rangle +|1111\rangle +|0011\rangle ) \nonumber \\ &&+a_{d,2}(|0011\rangle +|1111\rangle +|0000\rangle +|1100\rangle ) \nonumber \\ &&+a_{d,3}(|1111\rangle +|0011\rangle +|1100\rangle +|0000\rangle ) \nonumber \\ &=&\left( a_{d,0}+a_{d,1}+a_{d,2}+a_{d,3}\right) |\psi _{1}^{1}\rangle .\end{aligned}$$ Similarly: $$\begin{aligned} {\bf A}_{d}|\psi _{1}^{2}\rangle &=&\left( a_{d,0}+a_{d,1}-a_{d,2}-a_{d,3}\right) |\psi _{1}^{2}\rangle , \nonumber \\ {\bf A}_{d}|\psi _{1}^{3}\rangle &=&\left( a_{d,0}-a_{d,1}+a_{d,2}-a_{d,3}\right) |\psi _{1}^{3}\rangle , \nonumber \\ {\bf A}_{d}|\psi _{1}^{4}\rangle &=&\left( a_{d,0}-a_{d,1}-a_{d,2}+a_{d,3}\right) |\psi _{1}^{4}\rangle .\end{aligned}$$ This is in agreement with Eq. (\[eq:cik-DFS\]). Now, recall that each irrep appears 4 times. This means we should be able to find 3 more independent states belonging to each of the irreps. Indeed, by performing projections on the states $|0001\rangle $, $|0100\rangle $, $ |1001\rangle $ (using $|0010\rangle $ and $|1000\rangle $ does not produce new states) we obtain the complete basis for the DFSs. E.g., $$\begin{aligned} P^{1}|0001\rangle &=&|0001\rangle +|1101\rangle +|0010\rangle +|1110\rangle \equiv |\psi _{2}^{1}\rangle , \nonumber \\ P^{1}|0100\rangle &=&|0100\rangle +|1000\rangle +|0111\rangle +|1011\rangle \equiv |\psi _{3}^{1}\rangle , \nonumber \\ P^{1}|1001\rangle &=&|1001\rangle +|0101\rangle +|1010\rangle +|0110\rangle \equiv |\psi _{4}^{1}\rangle ,\end{aligned}$$ and again $$\begin{aligned} {\bf A}_{d}|\psi _{2}^{1}\rangle &=&\left[ a_{d,0}I^{4}+a_{d,1}X^{2}I^{2}+a_{d,2}I^{2}X^{2}+a_{d,3}X^{4}\right] \left( |0001\rangle +|1101\rangle +|0010\rangle +|1110\rangle \right) \nonumber \\ &=&\left( a_{d,0}+a_{d,1}+a_{d,2}+a_{d,3}\right) |\psi _{2}^{1}\rangle ,\end{aligned}$$ with similar results for the other states. All of this is in agreement with the general results of Sec. \[reptheory\]. Finally, we may consider an arbitrary superposition of decoherence free states taken from the multiple appearances of a given irrep, $|\phi ^{k}\rangle =\sum_{z=1}^{4}\theta _{z}^{k}|\psi _{z}^{k}\rangle $, and this will again be decoherence-free. ### The Subgroup $Q_{4}$ In this case the Kraus operators can be written as $$\begin{aligned} {\bf A}_{d}=a_{d,0}I^{4}+a_{d,1}X^{4}+a_{d,2}Y^{4}+a_{d,3}Z^{4}.\end{aligned}$$ Again, using the general arguments of Sec. \[reptheory\], in the case of $ Q_{4}$ we can just read off the matrix elements of the 4 (1-dimensional) irreps from the table in Eq. (\[eq:mult-table\]). Thus the 4 projection operators are: $$\begin{aligned} P^{1} &=&I^{4}+X^{4}+Y^{4}+Z^{4}\qquad P^{2}=I^{4}+X^{4}-Y^{4}-Z^{4} \nonumber \\ P^{3} &=&I^{4}-X^{4}+Y^{4}-Z^{4}\qquad P^{4}=I^{4}-X^{4}-Y^{4}+Z^{4}.\end{aligned}$$ Using the multiplicity formula, Eq. (\[eq:mk\]), the given representation again yields $\chi =\{16,0,0,0\}$ (for $I^{4},X^{4},Y^{4},Z^{4}$ respectively) and so once more $m_{k}=\frac{1}{4}\chi ^{k}(I^{4})16=4$ for all $k$. To find the decoherence free states let us start again with $|0000\rangle $. We find: $$\begin{aligned} P^{1}|0000\rangle &=& 2(|0000\rangle +|1111\rangle ) \equiv |\psi_{1}^{1}\rangle \nonumber \\ P^{2}|0000\rangle &=& |0000\rangle +|1111\rangle -|1111\rangle -|0000\rangle = 0 \nonumber \\ P^{3}|0000\rangle &=&|0000\rangle -|1111\rangle +|1111\rangle -|0000\rangle = 0 \nonumber \\ P^{4}|0000\rangle &=& 2(|0000\rangle -|1111\rangle ) \equiv |\psi _{1}^{4}\rangle .\end{aligned}$$ The vanishing of the projections of $P^{2}$ and $P^{3}$ implies that $ |0000\rangle $ has no components in the irreps $\Gamma ^{2}$ and $\Gamma ^{3} $. Thus a different starting state is needed, e.g., $|0001\rangle$. Then $$\begin{aligned} P^{2}|0001\rangle &=& 2(|0001\rangle +|1110\rangle ) \equiv |\psi _{1}^{2}\rangle \nonumber \\ P^{3}|0001\rangle &=& 2(|0001\rangle -|1110\rangle ) \equiv |\psi _{1}^{3}\rangle.\end{aligned}$$ That these states are decoherence free, is again easily verified by application of an arbitrary Kraus operator, e.g.: $$\begin{aligned} {\bf A}_{d}|\psi _{1}^{2}\rangle &=&\left[ a_{d,0}I^{4}+a_{d,1}X^{4}+a_{d,2}Y^{4}+a_{d,3}Z^{4}\right] 2\left( |0001\rangle +|1110\rangle \right) \nonumber \\ &=&\left( a_{d,0}+a_{d,1}-a_{d,2}-a_{d,3}\right) |\psi _{1}^{2}\rangle ,\end{aligned}$$ etc. The full DFS corresponding to the projection $P^{1}$ is found by applying $P^{1}$ to the initial states $|0011\rangle, |0101\rangle , |1001\rangle$: $$\begin{aligned} P^{1}|0011\rangle &=& 2(|0011\rangle +|1100\rangle ) \equiv |\psi _{2}^{1}\rangle \nonumber \\ P^{1}|0101\rangle &=& 2(|0101\rangle +|1010\rangle ) \equiv |\psi _{3}^{1}\rangle \nonumber \\ P^{1}|1001\rangle &=& 2(|1001\rangle +|0110\rangle ) \equiv |\psi _{4}^{1}\rangle,\end{aligned}$$ in addition to $|\psi _{1}^{1}\rangle$ above. Since the decoherence process described by $Q_{4}$ is different from that of $Q_{X}$, the decoherence free states are, not surprisingly, different in the two cases. ### The Subgroup $Q_{Z}$ As another example of an Abelian subgroup, assume now that the Kraus operators, for $K=2$ qubits, can be written as $$\begin{aligned} {\bf A}_{d}=a_{d,0}I^{2}+a_{d,1}ZI+a_{d,2}IZ+a_{d,3}Z^{2}.\end{aligned}$$ The 4 projection operators are thus: $$\begin{aligned} P^{1} &=&I^{2}+ZI+IZ+Z^{2}\qquad P^{2}=I^{2}+ZI-IZ-Z^{2} \nonumber \\ P^{3} &=&I^{2}-ZI+IZ-Z^{2}\qquad P^{4}=I^{2}-ZI-IZ+Z^{2}\end{aligned}$$ In this case, the given representation on 2 qubits yields $\chi =\{4,0,0,0\}$ (for $I^{2},ZI,IZ,Z^{2}$ respectively) and so $m_{k}=\frac{1}{4}\chi ^{k}(I^{2})4=1$ for all $k$. Thus as expected (since the representation is 4-dimensional), the multiplicity of each of the four 1-dimensional irreps is 1. Let us again explicitly find the decoherence free states: $$\begin{aligned} P^{1}|00\rangle &=&4|00\rangle \equiv |\psi ^{1}\rangle \nonumber \\ P^{2}|01\rangle &=&4|01\rangle \equiv |\psi ^{2}\rangle \nonumber \\ P^{3}|10\rangle &=&4|10\rangle \equiv |\psi ^{3}\rangle \nonumber \\ P^{4}|11\rangle &=&4|11\rangle \equiv |\psi ^{4}\rangle .\end{aligned}$$ And indeed: $${\bf A}_{d}|\psi ^{k}\rangle =\left( a_{d,0}+a_{d,1}+a_{d,2}+a_{d,3}\right) |\psi ^{k}\rangle ,\quad k=1,...,4.$$ This means that each of the 4 “computational basis states” $|\psi ^{k}\rangle $ is [*by itself*]{} a DFS. However, since these DFSs belong to different irreps, a superposition is not decoherence-free. This agrees with the well known fact that phase damping leads to decay of the off diagonal elements of the density matrix in the computational basis, but does not cause any population decay. ### The Subgroup $Q_{2Z}$ {#Q2Z} As a final example of an Abelian subgroup, let us return to the anistropic dipolar-coupling Hamiltonian \[Eq. (\[eq:dipolar\])\] discussed in Sec. \[Evolution\]. Note first that it is necessary to transform from the $ \sigma^\pm$ basis used there to $\sigma^{x,y}$ in order for our Pauli group-based discussion to apply. Having done that, it is clear that unless anisotropy is imposed this Hamiltonian generates the entire Pauli group, since all bilinear combinations $\sigma^\alpha \otimes \sigma^\beta$ appear in it. Assume therefore that we have a 4-spin molecule constrained to rotate only about the $z$-axis. This amounts to setting $g_{jk}^{\alpha \beta }=\delta _{\alpha 0}\delta _{\beta 0}g_{jk}$ in Eq. (\[eq:dipolar\]), so that only $\sigma _{j}^{z}\otimes \sigma _{k}^{z}$ terms remain. The corresponding subgroup is $$Q_{2Z}=\{IIII,ZZII,ZIIZ,IIZZ,ZIZI,IZZI,IZIZ,ZZZZ\}. \label{eq:Q2Z}$$ To find the DFS under $Q_{2Z}$, construct the projector $P^1=\frac{1}{8} \sum_{q \in Q_{2Z}}q$ corresponding to the identity irrep of $Q_{2Z}$. Applying this projector to the initial states $|0000\rangle$ and $ |1111\rangle$ we find a 2-dimensional DFS, spanned by these two states. This DFS thus encodes a single qubit. Non-Abelian Subgroups? ---------------------- It would have been interesting to find examples of non-Abelian subgroups which have 1-dimensional irreps and thus support a DFS. However, no such subgroups exist in the case of the Pauli group, as we now prove. Each two elements of the Pauli group $P_{K}$ either commute or anticommute (Appendix \[app1\]). Let $Q$ be a non-Abelian subgroup of $P_{K}$. Then there must be at least two elements of $Q$, say $q_{1}$ and $q_{2}$, that anticommute. Assume that the state $|i\rangle $ belongs to a 1-dimensional irrep $\tilde{\Gamma}$ of $Q$. Then $\tilde{\Gamma}(q_{1})|i\rangle =c_{1}|i\rangle $ and $\tilde{\Gamma}(q_{2})|i\rangle =c_{2}|i\rangle $, where $c_{1},c_{2}$ are numbers. Now, by assumption $\tilde{\Gamma} (q_{2}q_{1})=\tilde{\Gamma}(-q_{1}q_{2})$. Therefore $\tilde{\Gamma} (q_{1}q_{2})|i\rangle =\tilde{\Gamma}(q_{1})\tilde{\Gamma}(q_{2})|i\rangle =c_{1}c_{2}|i\rangle $, and also $\tilde{\Gamma}(q_{1}q_{2})|i\rangle = \tilde{\Gamma}(-q_{2}q_{1})|i\rangle =\tilde{\Gamma}(-q_{2})\tilde{\Gamma} (q_{1})|i\rangle =c_{1}\tilde{\Gamma}(-q_{2})|i\rangle $. If $\tilde{\Gamma} (-q_{2})=-\tilde{\Gamma}(q_{2})$ then we have $\tilde{\Gamma} (q_{1}q_{2})|i\rangle =-c_{1}c_{2}|i\rangle $ so that $c_{1}c_{2}=-c_{1}c_{2} $. This implies that at least one of $c_{1}$ and $c_{2}$ is zero. However, this cannot be true since the representation is [*unitary*]{}. Is there another possibility? Note that $\tilde{\Gamma}(-q_{2})=\tilde{\Gamma}(-{\bf I }q_{2})=\tilde{\Gamma}(-{\bf I})\tilde{\Gamma}(q_{2})$, so the question boils down to the value of $\alpha $ in $\tilde{\Gamma}(-{\bf I})=\alpha \tilde{\Gamma}({\bf I})$. But since $(-{\bf I})(-{\bf I})={\bf I}$ it follows that $\tilde{\Gamma}(-{\bf I})\tilde{\Gamma}(-{\bf I})=\tilde{\Gamma} ({\bf I})=1$, so that $\tilde{\Gamma}(-{\bf I})=\pm 1$. Assume then that the other case, $\tilde{\Gamma}(-{\bf I})=1$, holds. Let us use Eq. (\[eq:mk\] ) while recalling that only the four multiples of the identity have non-vanishing trace: $$\begin{aligned} m_k &=& \frac{1}{N}\sum_{n=1}^{N}\chi \left[ \Gamma ^{k}({\bf p}_{n})\right] ^* \chi \left[ \Gamma ({\bf p}_{n})\right] \nonumber \\ &=& \frac{1}{N} \left( \chi\left[\Gamma ^{k}({\bf I})\right]^* (2^K)+\chi \left[\Gamma ^{k}(-{\bf I})\right]^* (-2^K)+\chi\left[\Gamma ^{k}(i {\bf I}) \right]^* (i 2^K) +\chi\left[\Gamma ^{k}(-i {\bf I})\right]^* (-i 2^K) \right) . \label{eq:m_k-calc}\end{aligned}$$ Since the irrep $\Gamma^{k}$ is 1-dimensional, $\chi \left[\Gamma^k \right] = \Gamma^k$, i.e., the character is the element itself. Now let $\Gamma^{k}= \tilde{\Gamma}$. Then since $\tilde{\Gamma}(-{\bf I}) = 1$, and using $ \Gamma ^{k}(-i {\bf I}) = \tilde{\Gamma}(-{\bf I})\tilde{\Gamma}(i{\bf I})$, we find $\tilde{m}=0$. Therefore such irreps do not appear at all. Thus an anti-commuting pair of elements in $Q$ is incompatible with a 1-dimensional irrep, so that if $Q$ has a 1-dimensional irrep, it must be Abelian.[^8] Recall that the DFS condition of theorem 1 applies to arbitrary groups. Groups other than the Pauli group may support non-Abelian subgroups with 1-dimensional irreps (the above proof relied strongly on a property specific to the Pauli group, that its elements either commute or anticommute). However, at least within the Hamiltonian framework expounded in Secs. \[Hamiltonians\] and \[Evolution\], it is the Pauli group which appears naturally for the group algebra to which the Kraus operators belong. Dimension of the Decoherence-Free Subspaces {#dimension} =========================================== We showed in the previous section that for the Pauli group, DFSs can exist only for Abelian subgroups. This observation allows us to calculate the dimension of these DFSs. Recall from the general discussion in Sec. \[reptheory\] that in the generic case a superposition of states belonging to different irreps will decohere, whereas a superposition of states within copies of a given irrep will be decoherence-free (see also the examples in the previous section). Also, by the Abelian property, each such copy only supports a single decoherence free state. Hence [*the dimension of the DFS associated with a given irrep $\Gamma ^{k}$ is simply its multiplicity $ m_{k}$*]{}. Let $Q$ be an order-$N$ Abelian subgroup of the Pauli group on $K$ qubits. Using Eq. (\[eq:m\_k-calc\]) and ${\Gamma^k}(-{\bf I}) = \pm 1$ again, we have two (and only two) cases: (i) If $\Gamma ^{k}(-{\bf I}) = 1$ then $m_k=0 $, so such irreps do not support a DFS. (ii) If $\Gamma ^{k}(- {\bf I}) = -1$ then $$m_{k}=2^{K+2}/N.$$ This shows that all irreps that support a DFS have the same multiplicity, and thus all these DFSs have the same dimension. If the subgroup does not include elements with the $\pm 1,\pm i$ factors, as in the examples in Sec. \[Examples\], then only the term $\Gamma ^{k}({\bf I})$ appears in Eq. (\[eq:m\_k-calc\]), and consequently $$m_{k}=2^{K}/N\qquad {\rm no}\;\{\pm 1,\pm i\}\;{\rm {factors}.}$$ In any case, the dimension of the DFS is inversely proportional to the order of the subgroup. This implies a trade-off between the number of errors that can be dealt with by the code ($N$) and the number of decoherence free qubits ($\log _{2}m_{k}$). As an interesting corollary we see that largest Abelian subgroup of the Pauli group has order $2^{K+2}$ (since $m_{k}\geq 1$ implies $N\leq 2^{K+2}$ ). Examples of such subgroups are: - [The group generated by all the single qubit $X$’s (or $Y$’s or $Z$ ’s), with $\pm 1,\pm i$.]{} - [The group generated by $XXII...II$, $YYII...II$, $ZZII...II$, $ IIXXII...II$, $IIYYII...II$, $IIZZII...II$, ..., $II...IIZZ$, with $\pm 1,\pm i$.]{} These groups support only 1-dimensional DFSs. The last group is relevant for errors due to exchange on pairs of identical qubits [@Ruskai:99], and we see that the corresponding decoherence free state is automatically immune to exchange errors. (See Ref. [@Lidar:PRA99Exchange] for a discussion of protection of DFSs against exchange errors arising in the spatially correlated collective decoherence case.) Summary and Conclusions {#Summary} ======================= Decoherence-free subspaces (DFSs) are associated most commonly with the existence of a spatial symmetry in the system-bath coupling, as in the collective decoherence model. Here we have considered the case when no such symmetry is assumed, and have shown that one can nevertheless find DFSs under certain conditions. The essential assumptions are that either to lowest order only [*multiple*]{}-qubit errors are possible, meaning that the bath can only couple to multiple system excitations; or, that only one type of error process (such as phase-damping) occurs, which can be relevant for the NMR quantum computer schemes and optical lattices (or any other realization where scattering-induced phase-shifts are the dominant decoherence mechanism). In either case, instead of the full Pauli group of errors, only a subgroup needs to be considered. Barring certain non-generic cases, the DFSs then correspond to states that transform according to the $1$ -dimensional irreducible representations of such a subgroup. This characterization of DFSs, while formally similar to previous results, is different in that it trades the assumption of spatial symmetry for one of multiple-qubit coupling to the bath. We show in a sequel paper [@Lidar:00b] how to perform universal fault tolerant quantum computation on the DFSs found in this paper using only one- and two-body Hamiltonians. It would further be desirable to identify in detail the physical conditions under which the Pauli subgroup model is relevant for current proposals for quantum computers. An important example we have discussed is the dipolar-coupling induced decoherence in NMR. Acknowledgments =============== This material is based upon work supported by the U.S. Army Research Office under contract/grant number DAAG55-98-1-0371, and in part by NSF CHE-9616615. [10]{} , [*Introduction to Quantum Computation and Information*]{} (World Scientific, Singapore, 1999). C. Williams and S. Clearwater, [*Explorations in Quantum Computing*]{} (Springer-Verlag, New York, 1998). , Phys. Rev. A [**57**]{}, 737 (1998). L. Duan and G. Guo, Phys. Rev. A [**58**]{}, 3491 (1998), ANL Report No. quant-ph/9804014. , Phys. Rev. Lett. [**79**]{}, 3306 (1997), ANL Report No. quant-ph/9705044. , Mod. Phys. Lett. B [**11**]{}, 1085 (1997), ANL Report No. quant-ph/9710041. P. Zanardi, Phys. Rev. A [**57**]{}, 3276 (1998), ANL Report No. quant-ph/9705045. , Phys. Rev. Lett. [**81**]{}, 2594 (1998), ANL Report No. quant-ph/9807004. , Phys. Rev. Lett. [**82**]{}, 4556 (1999), ANL Report No. quant-ph/9809081. , Phys. Rev. A [**60**]{}, 1944 (1999), ANL Report No. quant-ph/9902041. , Phys. Rev. A [**52**]{}, 2493 (1995). , Phys. Rev. Lett. [**77**]{}, 793 (1996). , Phys. Rev. A [**54**]{}, 1862 (1996), ANL Report No. quant-ph/9604038. , Phys. Rev. A [**55**]{}, 900 (1997). L. Viola and S. Lloyd, Phys. Rev. A [**58**]{}, 2733 (1998). , Phys. Rev. Lett. [**82**]{}, 2417 (1999). L.-M. Duan and G. Guo, Pulse controlled noise suppressed quantum computation, ANL Report No. quant-ph/9807072. P. Zanardi, Phys. Lett. A [**258**]{}, 77 (1999), ANL Report No. quant-ph/9809064. , [Universal Fault-Tolerant Computation on Decoherence-Free Subspaces]{}, submitted to Phys. Rev. Lett. Available as ANL Report No. quant-ph/9909058. , [Theory of Decoherence-Free, Fault-Tolerant, Universal Quantum Computation]{}, submitted to Phys. Rev. A. Available as ANL Report No. quant-ph/0004064. , Phys. Rev. A [**61**]{}, 052307 (2000), ANL Report No. quant-ph/9907096. C. Slichter, [*Principles of Magnetic Resonance*]{}, No. 1 in [*Springer Series in Solid-State Sciences*]{} (Springer, Berlin, 1996). , in [*[Nonlinear Optics in Semiconductors]{}*]{}, edited by [R.K. Willardson and A.C. Beers]{} ([Academic Press]{}, [New York]{}, 1998). , [*[Theoretical Solid State Physics]{}*]{} ([Dover]{}, [New York]{}, 1985), Vol. 2. , ecoherence-Free Subspaces for Multiple-Qubit Errors: (II) Universal, Fault-Tolerant Quantum Computation, submitted to Phys. Rev. A. Available as ANL Report No. quant-ph/0007013. , Rev. Mod. Phys. [**59**]{}, 1 (1987). , Phys. Rev. A [**51**]{}, 1015 (1995). , Phys. Rev. Lett. [**75**]{}, 346 (1995). , Phys. Rev. A [**52**]{}, 3457 (1995). , [*[Matrix Analysis]{}*]{}, No. 169 in [*[Graduate Texts in Mathematics]{}*]{} (Springer-Verlag, [New York]{}, 1997). , Science [**275**]{}, 350 (1997). , in [*[Introduction to Quantum Computation and Information]{}*]{}, edited by [H.K. Lo, S. Popescu and T.P. Spiller]{} ([World Scientific]{}, Singapore, 1999), ANL Report No. quant-ph/9712048. , [On Universal and Fault-Tolerant Quantum Computing]{}, ANL Report No. quant-ph/9906054. , J. Mod. Optics [**44**]{}, 2455 (1997). C. Gardiner, [*Quantum Noise*]{}, Vol. 56 of [*Springer Series in Synergetics*]{} (Springer-Verlag, Berlin, 1991). , [*[States, Effects and Operations]{}*]{}, [*[Fundamental Notions of Quantum Theory]{}*]{} ([Academic]{}, Berlin, 1983). , Phys. Rev. A [**54**]{}, 2614 (1996). , [Classical interventions in quantum systems. I. The measuring process]{}, ANL Report No. quant-ph/9906023. , [*[The Many-Body Problem in Quantum Mechanics]{}*]{} ([Dover]{}, [New York]{}, 1995). L.-M. Duan and G.-C. Guo, Phys. Rev. A [**59**]{}, 4058 (1999). , Superlattices and Microstructures [**23**]{}, 445 (1998), further detail in LANL Report No. quant-ph/9704019. I. V. Schensted, [*A Course on the Application of Group Theory to Quantum Mechanics*]{} (NEO Press, Peaks Island, Maine, 1976). , Phys. Rev. A [**55**]{}, 2547 (1997). , [*[Group Theory in Physics: An Introduction]{}*]{} ([Academic Press]{}, San Diego, 1997). , Phys. Rev. Lett. [**82**]{}, 1975 (1999). , [Pauli Exchange Errors in Quantum Computation]{}, Phys. Rev. Lett. [**85**]{}, 194 (2000). ANL Report No. quant-ph/9906114. The Pauli Group {#app1} =============== The Pauli matrices are: $$\begin{aligned} \sigma _{0}\equiv I=\left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right) \qquad \sigma _{x}=\left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right) \qquad \sigma _{y}=\left( \begin{array}{cc} 0 & -i \\ i & 0 \end{array} \right) \qquad \sigma _{z}=\left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right) .\end{aligned}$$ They have the following properties: $$\begin{aligned} \sigma _{\alpha }^{2} &=&I\qquad \alpha =0,x,y,z \nonumber \\ \lbrack \sigma _{\alpha },\sigma _{\beta } \rbrack &=&2i\varepsilon _{\alpha \beta \gamma }\sigma _{\gamma } \nonumber \\ \{\sigma _{\alpha },\sigma _{\beta }\} &=&2\delta _{\alpha \beta }I \nonumber \\ \sigma _{\alpha }\sigma _{\beta } &=&i\varepsilon _{\alpha \beta \gamma }\sigma _{\gamma }+\delta _{\alpha \beta }I. \nonumber \\ {\rm Tr}(\sigma _{\alpha }) &=&0\qquad \alpha =x,y,z\end{aligned}$$ The [*Pauli group*]{} of order $K$ is the set of all $4^{K+1}$ possible tensor products of $K$ of the Pauli matrices and $\pm ,\pm i$: $$\begin{aligned} P_{K}=\pm ,\pm i\left\{ \bigotimes_{k=1}^{K}\sigma _{\alpha ,k}\right\} _{\alpha }.\end{aligned}$$ Some of its useful properties are - Let $p_{1},p_{2}\in P_{K}$. Since either $[\sigma _{\alpha ,k},\sigma _{\beta ,k}]=0$ or $\{\sigma _{\alpha ,k},\sigma _{\beta ,k}\}=0$ it follows that $$\begin{aligned} \text{either }[p_{1},p_{2}]=0\text{ or }\{p_{1},p_{2}\}=0.\end{aligned}$$ - Since $\sigma _{\alpha }$ are all unitary, so are all $p\in P_{K}$. - Since $\sigma _{\alpha }$ are all Hermitian but we allow for $\pm i$ factors, $p\in P_{K}$ is either Hermitian or anti-Hermitian. Thus if $p\in P_{K}$ then $p^{\dagger }\in P_{K}$. - Since ${\rm Tr}(A\otimes B)={\rm Tr}A\times {\rm Tr}B$, the only elements in $P_{K}$ which are not traceless are the four $\pm ,\pm i$ multiples of the identity, and each has trace $2^{K}$. Examples of Non-Generic Decoherence-Free Subspaces {#app2} ================================================== We will show here an example of a DFS that arises out of a two-dimensional irrep of a [*non-Abelian*]{} subgroup, in the “non-generic” case. Let us consider the non-Abelian 8-element subgroup $Q_{8}$ $=\{\pm III,\pm XXI,\pm IZZ,\pm iXYZ\}$. In this standard representation it is reducible and splits into 4 copies of a two-dimensional irreducible representation of $ Q_{8}$. Since there is just one irrep, we drop the irrep-index $k$ on $ \Gamma ^{k}$ etc. The two-dimensional representation of $Q_{8}$ is the following: $$\begin{aligned} \Gamma (\pm III) &=&\pm \left( \begin{array}{cc} 1 & 0 \\ 0 & 1 \end{array} \right) \qquad \Gamma (\pm XXI)=\pm \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right) \qquad \nonumber \\ \Gamma (\pm IZZ) &=&\pm \left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \end{array} \right) \qquad \Gamma (\pm iXYZ)=\pm \left( \begin{array}{cc} 0 & -1 \\ 1 & 0 \end{array} \right) .\end{aligned}$$ The 8-dimensional Hilbert space is split into 4 irreducible subspaces $V^{z}$ (corresponding to the 4 copies of $\Gamma $) spanned by $$\begin{aligned} V^{1} &=&(|\psi _{0}^{1}\rangle ,|\psi _{1}^{1}\rangle )=(|000\rangle ,|110\rangle ) \nonumber \\ V^{2} &=&(|\psi _{0}^{2}\rangle ,|\psi _{1}^{2}\rangle )=(|111\rangle ,|001\rangle ) \nonumber \\ V^{3} &=&(|\psi _{0}^{3}\rangle ,|\psi _{1}^{3}\rangle )=(|100\rangle ,|010\rangle ) \nonumber \\ V^{4} &=&(|\psi _{0}^{4}\rangle ,|\psi _{1}^{4}\rangle )=(|011\rangle ,|101\rangle ).\end{aligned}$$ On each of these two-dimensional subspaces the group acts like $\Gamma $. A codeword in the DFS can be expanded as $|\tilde{j}\rangle =\sum_{z=1}^{4}\sum_{\mu =0}^{1}\theta _{z,\mu }^{j}|\psi _{\mu }^{i}\rangle $.[^9] Let us take as our code just the first basis-vector of each irreducible subspace, i.e., $$\begin{aligned} {\cal C}=\{|\tilde{1}\rangle ,|\tilde{2}\rangle ,|\tilde{3}\rangle ,|\tilde{ 4 }\rangle \}\equiv \{|\psi _{0}^{z}\rangle :z=1\ldots 4\}=\{|000\rangle ,|111\rangle ,|100\rangle ,|011\rangle \}.\end{aligned}$$ Denoting the vector of coefficients as $\overrightarrow{\theta _{z}^{j}} =(\theta _{z,0}^{j},\theta _{z,1}^{j})$, this means that $\overrightarrow{ \theta _{z}^{z}}={(1,0)}$ and $\overrightarrow{\theta _{z}^{j\neq z}}$ $={\ (0,0)}$. In this case we can show that there are Kraus operators ${\bf A} _{d} $ that satisfy the DFS-condition on the code, by searching for matrices ${\cal A}_{d}$ which have $\overrightarrow{\theta _{z}^{j}}$ as eigenvectors. An example is: $$\begin{aligned} {\cal A}_{1}=\left( \begin{array}{cc} c_{1} & d_{1} \\ 0 & e_{1} \end{array} \right) \qquad {\cal A}_{2}=\left( \begin{array}{cc} c_{2} & d_{2} \\ 0 & e_{2} \end{array} \right) ,\end{aligned}$$ with the conditions $c_{1}^{\ast }d_{1}+c_{2}^{\ast }d_{2}=0$, $ |c_{1}|^{2}+|c_{2}|^{2}=1$ and $ |d_{1}|^{2}+|d_{2}|^{2}+|e_{1}|^{2}+|e_{2}|^{2}=1$ for normalization \[Eq. ( ref[eq:OSRnorm]{})\]. The corresponding Kraus-operators are: $$\begin{aligned} {\bf A}_{1} &=&\frac{c_{1}+e_{1}}{2}III+\frac{d_{1}}{2}XXI+\frac{c_{1}-e_{1} }{2}IZZ+\frac{d_{1}}{2}iXYZ \\ {\bf A}_{2} &=&\frac{c_{2}+e_{2}}{2}III+\frac{d_{2}}{2}XXI+\frac{c_{2}-e_{2} }{2}IZZ+\frac{d_{2}}{2}iXYZ. \nonumber\end{aligned}$$ The code ${\cal C}$ is a DFS. It is the particular equality (i.e., the “conspiring”, non-generic or accidental relationship) between the coefficients of the $XXI$ and $XYZ$ terms that is responsible for the existence of this DFS. [^1]: For a bath in thermal equilibrium, $|\nu \rangle $ would be an energy eigenstate with energy $E_{\nu }$, and $p_{\nu }=\exp (-\beta E_{\nu })/Z$, where $\beta $ is the inverse temperature and $Z={\rm Tr}[\exp (-\beta {\bf H }_{B})]$ is the canonical partition function. [^2]: See, however, Ref. [@Peres:99] for a discussion of Kraus operators represented by non-square matrices. [^3]: In fact [*spatially*]{} correlated errors can also be dealt with by QECCs [@Duan:99]. [^4]: We thank Prof. Dieter Suter for suggesting this example. [^5]: Note that in this case the expansion of the Kraus operators in terms of tensor products of Pauli matrices, Eq. (\[eq:Ai\]) remains valid. [^6]: For notational simplicity we avoid introducing another index for the multiplicity of the irrep here. That such superpositions are allowed for DF states is clear from Eq. (\[eq:sameirrep\]). [^7]: This statement of what are generic conditions that lead to a DFS is very similar to that in Ref. [@Lidar:PRL98]. [^8]: We thank Dr. P. Zanardi for discussions regarding this point. [^9]: Note that our indices here differ somewhat from the notations in Sec. \[reptheory\], because there we considered either one-dimensional irreps, or mostly avoided explicitly indicating superpositions between copies of a given irrep.

--- abstract: | The structure of the group $(\mathbb{Z}/n\mathbb{Z})^\star$ and Fermat’s little theorem are the basis for some of best-known primality testing algorithms. Many related concepts arise: Euler’s totient function and Carmichael’s lambda function, Fermat pseudoprimes, Carmichael and cyclic numbers, Lehmer’s totient problem, Giuga’s conjecture, etc. In this paper, we present and study analogues to some of the previous concepts arising when we consider the underlying group $\mathcal{G}_n:=\{a+bi\in\mathbb{Z}[i]/n\mathbb{Z}[i] : a^2+b^2\equiv 1\ \textrm{$\pmod n$}\}$. In particular we characterize Gaussian Carmichael numbers via a Korselt’s criterion and we present their relation with Gaussian cyclic numbers. Finally, we present the relation between Gaussian Carmichael number and 1-Williams numbers for numbers $n \equiv 3 \pmod{4}$. There are also no known composite numbers less than $10^{18}$ in this family that are both pseudoprime to base $1+2i$ and 2-pseudoprime. address: - | Departamento de Matemáticas, Universidad de Oviedo\ Avda. Calvo Sotelo, s/n, 33007 Oviedo, Spain - | Centro Universitario de la Defensa\ Ctra. de Heusca, s/n, 50090 Zaragoza, Spain - - | Departamento de Matemáticas, Estadística y Computación, Universidad de Cantabria\ F. Ciencias, Avda de los Castros s/n, 39005 Santander, Spain author: - 'J.M. Grau' - 'A. M. Oller-Marcén' - 'M. Rodríguez' - 'D. Sadornil' title: Fermat test with gaussian base and Gaussian pseudoprimes --- [AMS Mathematics Subject Classification: ]{} Introduction ============ Most of the classical primality tests are based on Fermat’s little theorem: let $p$ be a prime number and let $a$ be an integer such that $p\nmid a$, then $a^{p-1} \equiv 1 \pmod p$. This theorem gives a possible way to detect non-primes: if for a certain $a$ coprime to $n$, $a^{n-1}\not\equiv 1\pmod{n}$, then $n$ is not prime. The problem is that the converse is false and there exists composite numbers $n$ such that $a^{n-1}\equiv 1\pmod{n}$ for some $a$ coprime to $n$. In this situation $n$ is called pseudoprime with respect to base $a$ (or $a$-pseudoprime). A composite integer $n$ which is a pseudoprime to any base $a$ such that $\gcd(a,n)=1$ is called a Carmichael number (or absolut pseudoprime). Fermat theorem can be deduced from the fact that the units of $\mathbb{Z}/n\mathbb{Z}$ form a subgroup of order $n-1$ when $n$ is prime. Associated to the subgroup $(\mathbb{Z}/n\mathbb{Z})^\star$ we can define the well-know Euler’s totient function and Carmichael’s lambda function which are defined in the following way: $$\varphi(n):=|(\mathbb{Z}/n\mathbb{Z})^\star|, \quad \lambda(n):=\textrm{exp}(\mathbb{Z}/n\mathbb{Z})^\star.$$ It seems reasonable (and natural) to extend these ideas to other general groups $G_n$. This extension leads to composite/primality tests according to the following steps: - Compute $f(n)=|G_n|$ under the assumption that $n$ is prime. - Given $n$, if we can find $g\in G_n$ such that $|g|\nmid f(n)$, then $n$ is not prime. This idea is present in tests based in lucasian sequences [@WI] and elliptic curves [@SIL]. Recent works have developed these concepts in other contexts. Pinch [@PINCH] considers primality tests based on quadratic rings and discuss the absolute pseudoprimes for them. Shettler [@JORDAN] studies analogues to Lehmer’s Problem Totient and Carmichael numbers in a PID. Steele [@ST] generalizes Carmichael numbers to number rings introducing Carmichael ideals in number rings and proving an analogue to Korselt’s criterion for them. Following these approaches, in this paper we consider the groups $$\mathcal{G}_n:=\{a+bi\in\mathbb{Z}[i]/n\mathbb{Z}[i] : a^2+b^2 \equiv 1\ \textrm{(mod $n$)}\}.$$ For these groups, we define the corresponding Euler and Carmichael functions and we study some of their properties. We also present the concepts of Gaussian pseudoprime and Gaussian Carmichael number presenting an explicit Korselt’s criterion. Cyclic numbers, Lehmer’s Totient Problem [@BCF] and Giuga’s conjecture [@GIU] are also considered in this gaussian setting. It is known that Carmichael numbers have at least three prime factors. We show that Gaussian Carmichael numbers with only two prime factors exist and we determine their form. Moreover, although there are gaussian pseudoprimes with respect to any base, if we combine our ideas with a classical Fermat test, we show that no number of the form $4k+3$ smaller that than $10^{18}$ passes both tests (for some particular bases). This strength is possible due to a relationship with 1-Williams numbers [@WI] that we make explicit. Preliminaries ============= In this section we determine the order and structure of the group $\mathcal{G}_n$. We also show some elementary properties and relations between the Gaussian counterparts of Euler and Carmichael functions. For any positive integer $n$ we will denote by $\mathcal{I}_n$ the ring of gaussian integers modulo $n$; i.e., $$\mathcal{I}_n:=\{a+bi : a,b\in\mathbb{Z}/n\mathbb{Z}\}=\mathbb{Z}[i]/n\mathbb{Z}[i].$$ Further, we will consider the group $\mathcal{G}_n$ defined by $$\mathcal{G}_n:=\{a+bi\in\mathcal{I}_n : a^2+b^2\equiv 1\ \textrm{(mod $n$)}\}.$$ Once we have defined the group we can define the following arithmetic functions: $$\Phi(n):=|\mathcal{G}_n|,\ \ \rightthreetimes(n):=\textrm{exp}(\mathcal{G}_n).$$ Note that $\Phi$ and $\rightthreetimes$ are the analogues to Euler’s totient funtion and Carmichael’s lambda functions, respectively. It is quite clear that if $n=p_1^{r_1}\cdots p_s^{r_s}$, then $$\mathcal{G}_n\cong \mathcal{G}_{p_1^{r_1}}\times\cdots\times\mathcal{G}_{p_s^{r_s}}.$$ As a consequence, if $\gcd(m,n)=1$, $\Phi(mn)=\Phi(m)\Phi(n)$ and $\rightthreetimes(mn)=\textrm{lcm}(\rightthreetimes(m),\rightthreetimes(n))$. Hence, in order to study the group $\mathcal{G}_n$ we can restrict ourselves to the case when $n$ is a prime power. Let $p$ be a prime and let $k>0$ be an integer. Then: $$\mathcal{G}_{p^k}\cong\begin{cases} C_2, & \textrm{if $p=2$ and $k=1$};\\ C_{2^{k-2}}\times C_2\times C_4, & \textrm{if $p=2$ and $k\geq 2$};\\ C_{p^{k-1}}\times C_{p-1}, & \textrm{if $p\equiv 1$ (mod $4$)};\\ C_{p^{k-1}}\times C_{p+1}, & \textrm{if $p\equiv 3$ (mod $4$)}.\end{cases}$$ We will focus only on the case $p\equiv 3 \pmod 4$. In this case, it is well-known that $\mathcal{G}_p\cong GF(p^2)^{\star}$. Since $\mathcal{G}_p$ is a subgroup of $GF(p^2)^{\star}$, it must be cyclic. Moreover, counting quadratic residues it can be seen that $|\mathcal{G}_p|=p+1$ and, consequently, $\mathcal{G}_p\cong C_{p+1}$. We can now apply the Fundamental Lemma in [@GOL p. 587] to obtain that $|\mathcal{G}_{p^{k}}|=p^{k-1}(p+1)$. This means that, if $\Phi:\mathcal{G}_{p^k}\rightarrow \mathcal{G}_p$ is the $\pmod{p}$ group homomorphism, then $|\textrm{Ker}\ \Phi|=p^{k-1}$. Finally, observe that $\textrm{Ker}\ \Phi$ is an abelian $p$-group with exactly $p-1$ elements of order $p$, namely $\{1+Bp^{k-1}i\in\mathcal{G}_{p^{k}} : 1\leq B\leq p-1\}$. Consequently it must be cyclic and the proof is complete in this case. As a straightforward consequence we compute $\Phi(p^k)$ and $\rightthreetimes(p^k)$. \[corla2\] Let $p$ be a prime and let $k>0$ be an integer. Then $$\Phi(p^k)=\begin{cases} 2, & \textrm{if $p=2$ and $k=1$};\\ 2^{k+1}, & \textrm{if $p=2$ and $k>1$};\\ p^{k-1}(p-1), & \textrm{if $p\equiv 1$ (mod $4$)};\\ p^{k-1}(p+1), & \textrm{if $p\equiv 3$ (mod $4$)}.\end{cases}$$ $$\rightthreetimes(p^k)=\begin{cases} 2, & \textrm{if $p=2$ and $k=1$};\\ 4, & \textrm{if $p=2$ and $k=2,3,4$};\\ 2^{k-2}, & \textrm{if $p=2$ and $k\geq 5$}; \\ p^{k-1}(p-1), & \textrm{if $p\equiv 1$ (mod $4$)};\\ p^{k-1}(p+1), & \textrm{if $p\equiv 3$ (mod $4$)}.\end{cases}$$ For an odd prime number $p$, let us define $\beta(p)=\left(\frac{-1}{p}\right)$ and put $\beta(2)=0$. With this notation the following result is straightforward. \[Phin\] $$\Phi(n)=\begin{cases} \displaystyle{2n\prod_{p|n}\left(1-\frac{\beta(p)}{p}\right)}, & \textrm{if $4$ divides $n$}; \\ \displaystyle{n\prod_{p|n}\left(1-\frac{\beta(p)}{p}\right)}, & \textrm{otherwise}.\end{cases}$$ Recall that $\Phi(mn)=\Phi(m)\Phi(n)$ provided $\gcd(m,n)=1$. The following result describes the general situation. \[prod\] Let $m,n\in\mathbb{N}$. Then $$\Phi(nm)=\Phi(n)\Phi(m)\frac{\gcd(m,n)}{\Phi(\gcd(m,n))}.$$ It is enough to consider the prime power decomposition of $m$ and $n$. In particular, if we put $m=n$ we obtain the following. Let $n,s\in\mathbb{N}$. Then $$\Phi(n^m)=\begin{cases} n^{m-1}\Phi(n), & \textrm{if $n\not\equiv 2$ (mod $4$)};\\ 2n^{m-1}\Phi(n) & \textrm{if $n\equiv 2$ (mod $4$)}.\end{cases}$$ Let $m,n\in\mathbb{N}$. If $d=\textrm{gcd}(m,n)$ and $M= \textrm{lcm}(m,n)$, then: $$\Phi(d)\Phi(M)=\Phi(m)\Phi(n).$$ Recall that $M=n\frac{m}{d}$, where $\gcd(n,\frac{m}{d})=1$ and we can assume, without loss of generality, that $\gcd(m,d)=1$. Then, Proposition \[prod\] leads to: $$\Phi(M)=\Phi\left(n\frac{m}{d}\right)=\Phi(n)\Phi\left(\frac{m}{d}\right))\Phi(n)\frac{\Phi(m)}{\Phi(d)}$$ and the result follows. Recall that for the classical Euler and Carmichael functions, $\phi(n)=\lambda(n)$ if and only if $n=2$, $n=4$ or $n=p^r,2p^r$ for some odd prime $p$ and $r>0$. Note that in all these cases the group $(\mathbb{Z}/n\mathbb{Z})^\star$ is cyclic. For our recently defined functions $\Phi$ and $ \rightthreetimes$ we have the following: $\Phi(n)=\rightthreetimes(n)$ if and only if $n=2$ or $n=p^r$ for some odd prime $p$ and $r>0$. Just apply Corollary \[corla2\] and recall that if $\gcd(m,n)=1$, then $\Phi(mn)=\Phi(m)\Phi(n)$ while $\rightthreetimes(mn)=\textrm{lcm}(\rightthreetimes(m),\rightthreetimes(n))$. We end this section showing that the asymptotic behavior of $\Phi(n)$ is not exactly the same as that of his classical counterpart. $$\lim \inf \frac{\phi(n)}{n}=\lim \inf \frac{\Phi(n)}{n}=0$$ $$1=\lim \sup \frac{\phi(n)}{n} \neq \lim \sup \frac{\Phi(n)}{n}=\infty$$ For the asymptotic growth of Euler $\phi$ function and its limits see [@HW]. Now consider sequences $\{S_n\}$ and $\{L_n\}$ given by: $$S_n:= \prod_{\substack{p\leq n\\ p \equiv 3 \pmod 4}}p, \qquad L_n:= \prod_{\substack{p\leq n\\ p \equiv 1 \pmod 4}}p.$$ We have that $\Phi(p)=p+1$ for every odd prime $p \equiv 3 \pmod 4$, hence $$\lim_{n\rightarrow\infty}\frac{\Phi(S_n)}{S_n}=\lim_{n\rightarrow\infty} \prod_{\substack{p\leq n\\ p \equiv 3 \pmod 4}} \frac{p+1}{p}=\infty,$$ since $\prod \frac{p+1}{p}\geq 1+\sum 1/p$ and this series is divergent by the strong form of Dirichlet’s theorem. On the other hand, $$\lim_{n\rightarrow\infty}\frac{\Phi(L_n)}{L_n}= \lim_{n\rightarrow\infty}\prod_{\substack{p\leq n\\ p \equiv 1 \pmod 4}}\bigl(1-\frac{1}{p}\bigr).$$ Moreover, $$0\leq \prod_{\substack{p\leq n\\ p \equiv 1 \pmod 4}}\bigl(1-\frac{1}{p}\bigr) \leq \prod_{\substack{p\leq n\\ p \equiv 1 \pmod 4}} e^{-1/p}= e^{-\sum \frac{1}{p}},$$ where the sum in the exponent is taken over the primes $p\equiv 1 \pmod 4,\, p\leq n$. Again, by the strong form of Dirichlet’s theorem, this fuction tends to 0 and result holds. Gaussian Fermat pseudoprimes ============================ We start this section introducing the arithmetic function $\mathcal{F}$, which will play the same role as $n-1$ plays in the classical setting. $$\mathcal{F}(n)=\begin{cases} n-1, & \textrm{if $n\equiv 1$ (mod $4$)};\\ n+1, & \textrm{if $n\equiv 3$ (mod $4$)};\\ n, & \textrm{otherwise}.\end{cases}$$ Note that, if $n$ is prime, $\mathcal{F}(n)=|\mathcal{G}_n|$. We present the analogue to Fermat’s little theorem in this gaussian setting. \[FermatGaussiano\] Let $p$ be a prime number and let $z$ be a gaussian integer such that $p$ is coprime with $z \overline{z}$. Then: - $\left(\displaystyle{\frac{z}{\overline{z}}}\right)^{\mathcal{F}(p)}\equiv 1 \pmod{p}.$ - $Im(z^{\mathcal{F}(p)})\equiv 0 \pmod{p}.$ Note that if $z\in\mathbb{Z}[i]$ is such that $\gcd(n,z\overline{z})=1$, then $z/\overline{z}\in\mathcal{G}_n$. Hence, it is enough to apply Corollary \[corla2\]. Both conditions in Proposition \[FermatGaussiano\] are equivalent. We can consider the above result as a compositeness test for integers: if for some integer $n$ we find a gaussian integer $z$ such that either condition i) or ii) does not hold, then $n$ is a composite number. Nevertheless, like in the classical setting, the converse is not always true. This fact motivates the following definition: A composite integer $n$ is called a Gaussian Fermat pseudoprime (GFP) with respect to the base $z \in \mathbb{Z}[i]$ if $\gcd(n,z\overline{z})=1$ and condition i) (or equivalently ii)) from Proposition \[FermatGaussiano\] holds for $n$. In the classical setting the choice of different basis leads, in general, to different sets of associated Fermat pseudoprimes. In our case it is easy to describe a family of different basis leading to the same set of associated Gaussian Fermat pseudoprimes. Let $z,w$ be two gausian integers such that $|z|=|w|$. Then an integer $n$ is a Gaussian Fermat pseudoprime with respect to $z$ if and only if $n$ is a Gaussian Fermat pseudoprimes with respect to $w$. Assume that $n$ is a GFP with respect to $z$. Then $\gcd(n,z\overline{z})=1$ and $(z/\overline{z})^{\mathcal{F}(n)}\equiv 1\pmod{n}$. Now, since $|w|=|z|$ we have that $\gcd(n,w\overline{w})=\gcd(n,z\overline{z})=1$. Moreover, since $(z/\overline{z})^{\mathcal{F}(n)}\equiv 1\pmod{n}$ and $z/\overline{z}\in\mathcal{G}_n$ it follows that $\rightthreetimes(n)\mid \mathcal{F}(n)$. Hence, $(w/\overline{w})^{\mathcal{F}(n)}\equiv 1\pmod{n}$ because $w/\overline{w}\in\mathcal{G}_n$. The converse is clear since the roles of $z$ and $w$ are symmetric and the proof is complete. Gaussian Carmichael and cyclic numbers ====================================== An integer $n$ that is a Fermat pseudoprime for all bases coprime to $n$ is called a Carmichael number [@CARMI]. In the gaussian case there also exists composite numbers which are GFP with respect all bases. A composite number $n\in\mathbb{N}$ is a Gaussian Carmichael number ($G-$Carmichael) if it is a GFP to base $z$ for every gaussian integer $z$ such that $n$ is coprime to $z \overline{z}$. An alternative and equivalent definition of Carmichael numbers is given by Korselt’s criterion [@KORSELT] which states that a positive composite integer $n$ is a Carmichael number if and only if $n$ is square-free, and for every prime divisor $p$ of $n$, $p-1$ divides $n-1$. It follows from this characterization that all Carmichael numbers are odd. A similar characterization of $G-$Carmichael numbers can be given, showing that there are also even $G-$Carmichael numbers. \[car2\] For every composite integer $n$ the following are equivalent. - $n$ is $G-$Carmichael number. - $\rightthreetimes(n)$ divides $\mathcal{F}(n)$. - For every prime divisor $p$ of $n$, $\mathcal{F}(p)$ divides $\mathcal{F}(n)$ and one of the following conditions holds: - $n$ is odd and square-free, - $n$ is multiple of $4$ and $\frac{n}{4}=2,3,5$ or not a prime number. Since $\rightthreetimes(n)$ is the exponent of the group $\mathcal{G}_n$, i) and ii) are clearly equivalent. From Corollary \[corla2\] and the fact that $\rightthreetimes(mn)=\textrm{lcm}(\rightthreetimes(m),\rightthreetimes(n))$ if $\gcd(m,n)=1$, it is easy to see that iii) implies ii) when $n$ is a number that satisfies a) or b). Finally, assume now that $\rightthreetimes(n)$ divides $\mathcal{F}(n)$ and let be $n=2^ap_1^{r_1}\cdots p_s^{r_s}$. We have that $\rightthreetimes(n)=\textrm{lcm}(\rightthreetimes(2^a),\rightthreetimes(p_1^{r_1}),\dots,\rightthreetimes(p_s^{r_s}))$, so $\rightthreetimes(2^a)$ and $\rightthreetimes(p_i^{r_i})$ divides $\mathcal{F}(n)$. From Corollary \[corla2\], it is clear that for every prime $p$, $\rightthreetimes(p)=\mathcal{F}(p) $ divides $\rightthreetimes(p^k)$ with $k\geq 1$ and $\mathcal{F}(p)$ also divides $\mathcal{F}(n)$ as claimed. If $n$ is odd ($a=0$) and $r_i\geq 2$ for some $i\in\{1,\dots,s\}$ we get that $p_i$ divides $\rightthreetimes(n)$ and, consequently, also $\mathcal{F}(n)$. Thus, $p_i$ divides $n-1$ or $n+1$ which is a contradiction and $n$ must be square-free in this case. We now turn to the even case. If $a=1$ and $n$ is divisible by an other prime $p$ such that $p\equiv 1 \pmod{4}$, then $p-1$ divides $\mathcal{F}(n)=n$. Hence $n$ is a multiple of $4$, a contradiction. The same follows if there exist a prime $p\equiv 3 \pmod{4}$ dividing $n$ so we conclude that if $n\neq 2$ is even, it must be a multiple of $4$. Now, let be $n=4p$ with $p$ a prime. If $p=2$, then $n=8$ and we are done. If $p\equiv 1 \pmod{4}$, it follows that $p-1$ divides $n$; i.e., $p-1$ divides 4 so $p=5$ and $n=20$. Finally, if $p\equiv 3 \pmod{4}$, it follows that $p+1$ divides 4 so $p=3$ and $n=12$. Hence we have seen that if $4$ divides $n$ and $n\neq 8,12,20$, then $\frac{n}{4}$ is not prime and the proof in complete. In 1994 it was shown by Alford, Granville y Pomerance [@ALF] that there exist infinitely many Carmichael numbers. It is easy to see that every power of 2 is a $G-$Carmichael number, hence there are also infinitely many of them. However, if we restrict to odd $G-$Carmichael numbers, the problem seems to have at least the same difficulty as the classical case. Carmichael numbers have at least three prime factors. We know that 12 and 20 are only even $G-$Carmichael numbers with only two prime factors. The following result describes the family of odd $G-$Carmichael numbers with exactly two prime factors. \[Prop2p\] Let $p<q$ be odd primes. Then $n=pq$ is a gaussian Carmichael number if and only if $p$ and $q$ are twin primes such that $8$ divides $p+q$. Assume that $n=pq$ with $p<q$ odd primes is a $G-$Carmichael number. If $p,q\equiv 1 \pmod 4$ , $\mathcal{F}(n)=n-1=pq-1$. From proposition \[car2\], $p-1$ divides $pq-1=(p-1)(q+1)+q-p$. Hence $p-1$ divides $q-p$ and also $q-1=(q-p)+(p-1)$. In the same way $q-1$ divides $p-1$. So $p-1=q-1$ which is impossible. If $p,q\equiv 3 \pmod {4}$ we reach a similar contradiction using the same ideas. If $p\equiv 1 \pmod{4}$ and $q\equiv 3 \pmod{4}$ we obtain that $p=q+2$ which is impossible because $p<q$. Thus $p\equiv 3 \pmod{4}$ and $q\equiv 1 \pmod{4}$ and we have, whit similar reasoning that $q=p+2$. Moreover, $p+q=2p+2\equiv 0 \pmod{8}$. The converse is trivially true and the proof is complete. Recall that a positive integer $n$ which is coprime to $\phi(n)$ is called a cyclic number (sequence A003277 in [@OEIS]). This terminology comes from group theory since a number $n$ is cyclic if and only if any group of order $n$ is cyclic [@SZ]. From Korselt’s criterion it follows that any divisor of a Carmichael number is cyclic. In the gaussian setting we define Gaussian cyclic numbers in the following way. An integer $n$ is called $G-$cyclic if $\gcd(\Phi(n),n)=1$. The relationship between G-Carmichael and G-cyclic numbers is the same as in the setting, the proof being also quite similar. Any divisor of a odd $G-$Carmichael number is $G-$cyclic. Let $n$ be an odd $G-$Carmichael number. Since $n$ is square-free, $n=p_1p_2\cdots p_r$ and from proposition \[Phin\], $\Phi(n)=\prod (p_i-\beta(p_i))$. A divisor $d$ of $n$ is a product of some of these primes, that is, $d=\prod_{h \in J} p_h$, $J\subseteq \{1,2,\ldots, r\}$. If $GCD(\Phi(d),d)<>1$,, then there exist two indices $i\neq k$ in $J$ such that $p_i$ divides $p_k-\beta(p_k)$. As $n$ is a Carmichael number, we also have $p_k-\beta(p_k)$ divides $n-\beta(n)$. Hence, $p_i$ divides $n-\beta(n)$ which is absurd since $n$ is divisible by $p_i$ and $\beta(p_i)=\pm 1$. Around 1980, G. Michon conjectured that all odd cyclic numbers have Carmichael multiples. This can be reasonably extended to $G-$cyclic numbers and we can ask if all odd $G-$cyclic numbers have $G-$Carmichael multiples. Cyclic numbers can also be characterized in terms on congruences. A number $n$ is cyclic if and only if it satisfies $\phi(n)^{\phi(n)} \equiv 1 \pmod n$ or $\lambda(n)^{\lambda(n)} \equiv 1 \pmod n$. In our situation only one implication remains valid, namely. If $\Phi(n)^{\Phi(n)} \equiv 1 \pmod n$ or $\rightthreetimes(n)^{\rightthreetimes(n)} \equiv 1 \pmod n$, then $n$ is a $G-$cyclic number. Let $n$ be a positive integer such that $\Phi(n)^{\Phi(n)} \equiv 1 \pmod n$. Then, for any prime divisor $p$ of $n$ it holds that $\Phi(n)^{\Phi(n)} \equiv 1 \pmod p$. Now, if $n$ is not a $G-$cyclic number, there exists a prime $p$ with $p \mid \gcd(\Phi(n),n)$. Thus, $p$ divides $\Phi(n)$ and $\Phi(n)^{\Phi(n)} \equiv 0 \pmod p$, a contradiction. On the other hand, let $n=p_1^{e_1}p_2^{e_2}\cdots p_r^{e_r}$ be a positive integer. If $n$ is not a $G-$cyclic number then there exists a prime $p$ which $p \mid \gcd(\Phi(n),n)$. Since $p\mid \Phi(n)=\Phi(p_1^{e_1})\Phi(p_2^{e_2})\cdots \Phi(p_r^{e_r})$, there exists $1\leq j\leq r$ with $p\mid \Phi(p_j^{e_j})$. If $p_j=2$ then $p=2$ and $\rightthreetimes(n)$ is even. Otherwise, $\Phi(p_j^{e_j})=\rightthreetimes(p_j^{e_j})$ and $p$ divides $\rightthreetimes(n)$. So $\rightthreetimes(n)^{\rightthreetimes(n)} \equiv 0 \pmod p$ and $n$ do not satisfy the hypothesis. The converse of the previous proposition is no true. In fact there are $G-$cyclic numbers $n$ that do not satisfy any of the above conditions. The first of them being: $$77, 119, 133, 187, 217, 253, 287, 301, 319, 323, 341, 391,\ldots$$ $G-$Lehmer’s totient problem and $G-$Giuga’s conjecture ======================================================= Lehmer’s totient problem, named after D. H. Lehmer, asks whether there is any composite number $n$ such that $\phi(n)$ divides $n - 1$. This is true for every prime number, and Lehmer conjectured in 1932 [@LE] that the answer to his question was negative. He showed that if any such $n$ exists, it must be odd, square-free, and divisible by at least seven primes. This numbers, called Lehmer numbers, are clearly Carmichael numbers and, up to date, none has been found. It is known that these numbers have at least 15 prime factors and are greater than $10^{30}$. Moreover, if a Lehmer number is divisible by 3, the number of prime factors increases to $40000000$ with more than 360000000 digits (see [@BCF]). We now define our analogue concept. A composite number $n$ is a $G-$Lehmer number if $ \Phi (n) \mid \mathcal{F}(n)$. It is clear that every $G-$Lehmer number is a $G-$Carmichal number. Besides, it is easy to note that $G-$Lehmer numbers exist. Let $p<q$ be odd primes. Then $n=pq$ is a $G-$Lehmer number if and only if $p$ and $q$ are twin primes such that $8$ divides $p+q$. As $n$ must be a $G-$Carmichael number, $n=(4k-1)(4k+1)$ where $4k\pm 1$ are both primes. From this numbers $\Phi(n)=\mathcal{F}(n)=(4k)^2$, so $n$ is a $G-$Lehmer number. Note that, from Proposition \[Prop2p\] this result means that every odd $G-$Carmichael number with exactly 2 prime factors is a $G-$Lehmer number. Nevertheless, there are $G-$Lehmer numbers with more thatn 2 prime factors (A182221 in [@OEIS]): $$255, 385, 34561, 65535, 147455, 195841, \dots$$ This suggests an interesting question: Are there infinitely many $G-$Lehmer numbers? Furthermore, all known $G-$Lehmer numbers satisfy that $ \mathcal{F}(n) =\Phi(n) $. Hence, it is reasonable to propose the $G-$Lehmer’s Totient problem: Is there any number $n$ such that $\phi(n)$ is a proper divisor of $\mathcal{F}(n)$? In 1932, Giuga [@GIU] proposed another conjecture about prime numbers. He postulated that a number $p$ is prime if and only if $ \sum i^{p-1}\equiv -1 \pmod p $, where the sum is taken over all integers $1\leq i\leq p-1$. Giuga showed that there are no exceptions to his conjecture up to $10^{1000}$. This was later improved to $10^{13800}$ [@BOR]. A similar approach to Giugas’s conjecture, replacing $n-1$ by $\mathcal{F}(n)$, leads us to consider the following set, which contains all prime numbers. $$\mathfrak{G} := \{n \in \mathbb{N}:\sum_{z \in \mathcal{G}_n} z^{\mathcal{F}(n)} \equiv \mathcal{F}(n) \pmod{n}\}$$ However, this set also contains lots of composite numbers. For example, every power of 2 is in $\mathfrak{G}$. For odd integers we have the next result. Let be $n$ an odd integer. If $\Phi(n)=\mathcal{F}(n)$, then $n \in \mathfrak{G}$. Since $|\mathcal{G}_n|=\Phi(n)$, for all $z \in \mathcal{G}_n$, $z^{\Phi(n)} \equiv 1 \pmod n$. If $\Phi(n)=\mathcal{F}(n)$ then $$\sum_{z \in \mathcal{G}_n} z^{\Phi(n)} \equiv |\mathcal{G}_n| \equiv \Phi(n) \equiv \mathcal{F}(n) \pmod n,$$ and $n$ is in $\mathfrak{G}$. Thus, prime numbers and every known $G-$Lehmer numbers are in $\mathfrak{G}$. Furthermore, no other odd composite integer is known to be in $\mathfrak{G}$. So, we formulate the following conjecture regarding numbers in $\mathfrak{G}$. For every odd $n$, $n \in \mathfrak{G}$ if and only if $\Phi(n)=\mathcal{F}(n)$. Gaussian Fermat test for numbers of the form $4k+3$. ===================================================== The use of gaussian integers to perform the equivalent of Fermat’s little theorem to test primality is not just a mere theoretical speculation. Lucas pseudoprimes [@WI] for some particular sequences can be also seen as gaussian pseudoprimes. However, Gaussian integers, and the corresponding definition of peudoprimes using powers, is more similar to the classical one than the concept of Lucas sequences. As we have said before, we can take advantage of Proposition \[FermatGaussiano\] to test primality (more precisely compositeness) of a number. This is that we call the Gaussian Fermat Test with respect to the base $z$. Computational evidence reveals that this test, based on the structure of $\mathcal{G}_N$, is very powerful when it is combined with the classical one; i.e., there are very few common pseudoprimes. Furthermore, this combination is more stronger if we restrict to numbers of the form $4k +3 $. From the William Galway list [@GAL], we have checked that every Fermat pseudoprime number to base 2 less than $10^{18}$ and of the form $4k +3 $ is not a Gaussian pseudoprime to base $z=1+2i$. Baillie-PSW primality test [@PSW], used in a lot of computer algebra systems and software packages, is also a combination of two primality tests. More precisely it is a strong Fermat probable prime test to base 2 and a strong Lucas probable prime test. As the previous combination, no composite number below $10^{19}$ passes it, but it considers two strong type-test in contrast of our two basic Fermat type tests. On the other hand, there are integers of the form $4k+3$ which are both Fermat pseudoprimes to base 2 and Lucas pseudoprimes (see sequence A227905 in [@OEIS]). In general, combinations of two Fermat test with respect to two different prime basis (less than 30) present more than 10 (and a mean of 34) pseudoprimes lower than $ 4 \cdot 10^7 $ of the form $ 4k +3$. Even if we combine two basis to test if a number $n$ is a prime using the Gaussian Fermat Test, there are more pseudoprimes. However, there is no composite number of the form $4k+3$ less than $4 \cdot 10^7$ which is both a Gaussian pseudoprimes with respect to $1+2i$ and a Fermat pseudoprime with respect to a prime base less than 30. The lowest base to be used to find a Fermat pseudoprime with respect this base which is also a Gaussian Fermat pseudoprime to the base $ 1 +2 i $ is $10$. Also with other Gaussian basis the combination with a Fermat test is very strong as it is shown in the following table, which presents the number of composite integers less than $ 4 \cdot 10^7 $ which are simultaneously Gaussian Fermat pseudoprimes with respect to a base $z$ (horizontal) and Fermat pseudoprimes with respect to a base $a$ (vertical). $$ base 2 3 4 5 6 7 8 9 10 11 ----------- --- --- --- --- --- --- --- --- ---- ---- 1 + 2$i$ 0 0 0 0 0 0 0 0 1 0 1 + 4$i$ 0 0 1 0 0 0 0 0 0 1 1 + 6$i$ 0 1 2 0 2 0 0 1 0 1 1 + 10$i$ 0 1 1 0 0 0 2 1 2 1 2 + 5$i$ 0 0 1 0 1 0 0 0 0 1 2 + 7$i$ 0 0 1 0 1 0 2 1 0 1 3 + 8$i$ 0 1 2 1 0 0 1 1 0 1 3 + 10$i$ 0 1 2 0 1 0 2 1 1 1 4 + 5$i$ 0 0 1 0 0 0 1 0 0 1 4 + 9$i$ 0 0 1 0 0 0 1 0 0 1 $$ One of the reasons explaining this phenomenon is that Carmichael numbers, which always appear when combining two classical Fermat tests, are avoided when we combine a Fermat test and a Gaussian Fermat test, because Carmichael numbers are not necessarily $G-$Carmichael numbers and conversely. In fact, there are no Carmichael numbers of the form $4k +3 $ smaller than $ 10 ^{18} $ which are also $G-$Carmichael numbers. Recall that an integer $n$ is called an $r-$Williams number [@Ec; @WI] if $$p \mid n\Rightarrow (p+r) | (n+r)\textrm{ and } (p-r) | (n-r)$$ The following result relates our previous discussion with $1-$Williams numbers. An odd number $n \equiv 3 \pmod 4$ is simultaneously a Carmichael number and a $G-$Carmichael number if and only if $n$ is an $1-$Williams number and $p \equiv 3 \pmod 4$ for every $p$ dividing $n$. Let $n \equiv 3 \pmod 4$, then $\mathcal{F}(n)=n+1$. If $n$ is both a Carmichael and a $G-$Carmichael number we have that, for every $p$ dividing $n$: $$\begin{array}{ll} p-1 \mid n-1, & \\ p-1\mid n+1, & \hbox{if $p\equiv 1 \pmod{4}$,} \\ p+1\mid n+1, & \hbox{if $p\equiv 3 \pmod{4}$.} \end{array}$$ Now, if there exists a prime factor $p\equiv 1 \pmod{4}$ it follows that $n-1=(p-1)k\equiv 0\pmod{4}$, a contradiction. Hence, every prime factor is congruent with 3 modulo 4 and $n$ is a $1-$Williams number. On the other hand, if $n$ is a $1-$Williams number, then for each prime factor $p$ of $n$ we have $p-1\mid n-1$ and $p+1\mid n+1$, so $n$ is a Carmichael number. If $n$ is to be a $G-$Carmichael it is also necessary that every factor $p\equiv 1 \pmod{4}$ satisfies $p-1\mid n+1$. But, by hypothesis, $n$ does not have this kind of factors and the result follows. Thus, the search for a number of the form $n \equiv 3 \pmod 4$ which is both a $G-$Carmichael number and a Carmichael number is harder than to find a $1-$Williams number and, up to date, no $1-$Williams number is known [10]{} W.R. Alford, A. Granville and C. Pomerance, There are Infinitely Many Carmichael Numbers. Ann. Math. 139, 703-722, 1994. D. Borwein, C. Maitland and M. Skerritt Computation of an Improved Lower Bound to Giuga’s Primality Conjecture Integers, Electronic Journal of Combinatorial Number Theory. 13, 2013. P. Burcsi, S. Czirbusz and G. Farkas. Computational investigation of Lehmer’s totient problem. Ann. Univ. Sci. Budapest. Sect. Comput. 35, 43-49, 2011. R. D. Carmichael. Note on a new number theory function. Bull. Amer. Math. Soc., 16(5): 232-238, 1910 O. Echi. Williams numbers . Comptes rendus mathématiques de l’Académie des sciences Soc. Royal du Canada. 29, 2, 41–47. 2007. William Galway. http://www.cecm.sfu.ca/Pseudoprimes/ J. R. Goldman. Numbers of solutions of congruences: Poincare series for strongly nondegenerate forms. Proc. Amer. Math. Soc., 87(4), 586–590, 1983. G. Giuga. Su una presumibile proprieta caratteristica dei numeri primi. Ist. Lombardo Sci. Lett. Rend. A 83, 511-528, 1950. G. H. Hardy, and E. M. Wright, An introduction to the theory of numbers. Sixth edition. Oxford University Press, 2008. J. Schettler Lehmer’s totient problem and Carmichael numbers in a PID.\ (Parts I and II), Undergraduate Honors Seminar, University of Tennessee, November 2005. A. R. Korselt. Probleme chinois. L’intermediaire des mathematiciens, 6: 142-144, 1899. D. H. Lehmer. On Euler’s totient function. Bulletin of the American Mathematical Society 38: 745-751, 1932. OEIS. The On-Line Encyclopedia of Integer Sequences. http://www.oeis.org. R. Pinch. Absolute quadratic pseudoprimes. Proceedings Conference on Algorithmic Number Theory, Turku, May 2007. Turku Centre for Computer Science General Publications 46. C. Pomerance, J.L. Selfridge, and S.S. Wagstaff. The pseudoprimes to $25\cdot 10^9$. Mathematics of Computation 35, 1003–1026, 1980. J. H. Silverman. Elliptic Carmichael numbers and elliptic Korselt critria. Acta Arithmetica 155, 233–246, 2012. G.A. Steele. Carmichael numbers in number rings, J. Number Theory 128, No. 4, 910-917, 2008. T. Szele, Über die endlichen Ordnungszahlen zu denen nur eine Gruppe gehört, Commentarii. Math. Helv., 20, 265-67, 1947. H. C. Williams. On numbers analogous to the Carmichael numbers. Canad. Math. Bull. 20, 133 -143, 1977.

--- abstract: | Bilinear dynamical systems are commonly used in science and engineering because they form a bridge between linear and non-linear systems. However, simulating them is still a challenge because of their large size. Hence, a lot of research is currently being done for reducing such bilinear dynamical systems (termed as bilinear model order reduction or bilinear MOR). Bilinear iterative rational Krylov algorithm (BIRKA) is a very popular, standard and mathematically sound algorithm for bilinear MOR, which is based upon interpolatory projection technique. An efficient variant of BIRKA, Truncated BIRKA (or TBIRKA) has also been recently proposed. Like for any MOR algorithm, these two algorithms also require solving multiple linear systems as part of the model reduction process. For reducing very large dynamical systems, which is now-a-days becoming a norm, scaling of such linear systems with respect to input dynamical system size is a bottleneck. For efficiency, these linear systems are often solved by an iterative solver, which introduces approximation errors. Hence, stability analysis of MOR algorithms with respect to inexact linear solves is important. In our past work, we have shown that under mild conditions, BIRKA is stable (in the sense as discussed above). Here, we look at stability of TBIRKA in the same context. Besides deriving the conditions for a stable TBIRKA, our other novel contribution is the more intuitive methodology for achieving this. This approach exploits the fact that in TBIRKA a bilinear dynamical system can be represented by a finite set of functions, which was not possible in BIRKA (because infinite such functions were needed there). The stability analysis techniques that we propose here can be extended to many other methods for doing MOR of bilinear dynamical systems, e.g., using balanced truncation or the ADI methods. title: INEXACT LINEAR SOLVES IN MODEL REDUCTION OF BILINEAR DYNAMICAL SYSTEMS --- [**Rajendra Choudhary and Kapil Ahuja**]{} .5cm Introduction {#sec:Inroduction} ============ A dynamical system, usually represented by a set of differential equations, can be linear or non-linear [@Mathematics_and_Climate_dynamicalsystembook; @Differential_Dynamical_Systems_book; @dynamical_system_application_1986]. Linear dynamical systems have been studied more than the non-linear ones because of the obvious ease in working with them. Bilinear dynamical systems form a good bridge between the linear and the non-linear cases, and are usually approximated by a varying degree of bilinearity [@nonlinear_to_bilinear_paper_1974; @RughnonlinearsystemBook; @nonlinear_to_bilinear_paper_2015]. In this manuscript, we focus on bilinear dynamical systems. Dynamical systems coming from different real world applications are very large in size. Thus, simulation and computation with such systems is prohibitively expensive. Model Order Reduction (MOR) techniques provide a smaller system that besides being cheaper to work with, also replicates the input-output behaviour of the original system to a great extent [@AntoulasBook; @Peterbookdynamicalsystemreduction; @grimmephdthesis; @Zbaimorpaper; @k_willcox_model_reduction_paper; @serkanIRKAsiampaper; @mimosystem; @biliner_model_reduction_Zbai; @BIRKApaper; @AIRGApaper]. Since, bilinear dynamical systems have been recently studied, the techniques for reducing them are also recent. Out of the many methods available for performing bilinear MOR [@biliner_model_reduction_Zbai; @BIRKApaper; @bilinear_model_reduction_Mian_Ilyas; @bilinear_model_reduction_peter_benner_balanced_trunc; @bilinear_model_reduction_Antoulas_Loewner_framework; @bilinear_model_reduction_kang-li_xu_gramian_based; @garretphdthesis; @TBIRKApaper], we focus on a commonly used interpolatory projection method. BIRKA (Bilinear Iterative Rational Krylov Algorithm) [@BIRKApaper] is a very popular algorithm based upon this technique for reducing *first-order* bilinear dynamical systems[^1]. BIRKA’s biggest drawback is that it does not scale well in time (with respect to increase in the size of the input dynamical system). A cheaper variant of BIRKA, called TBIRKA (Truncated Bilinear Iterative Rational Krylov Algorithm) [@garretphdthesis; @TBIRKApaper] has also been proposed. Like in any other MOR algorithm, in BIRKA and TBIRKA also, people often use direct methods like LU-factorization, etc., to solve the arising linear systems, which are too expensive (i.e., computational complexity of $\mathcal{O}(n^3)$, where $n$ is the original system size). A common solution to this scaling problem is to use iterative methods like the Krylov subspace methods, etc., which have a reduced computational complexity (i.e., $\mathcal{O}(n \times nnz)$, where $nnz$ is the number of nonzeros in the system matrix) [@greenbaum_book; @SaaditerativeBook]. Although iterative methods are cheap, they are inexact too. Hence, studying stability of the underlying MOR algorithm (here BIRKA and TBIRKA) with respect to such approximate linear solves becomes important [@TrefethenBauBook; @Demmelbook]. One of the first works that performed such a stability analysis focused on popular MOR algorithms for *first-order* linear dynamical systems [@sarahinexactlaapaper]. Here, the authors briefly mention that their analysis would be easily carried from the *first-order* to the *second-order* case. A detailed stability analysis focusing on reducing *second-order* linear dynamical systems has been done in [@NavSISC]. A different kind of stability analysis for MOR of *second-order* linear dynamical systems has been done in [@TOAR_stability_analysis_ZBai]. In this, the authors first show that the SOAR algorithm (second order Arnoldi) is unstable with respect to the machine precision errors (and not inexact linear solves). Then, they propose a Two-level orthogonal Arnoldi (TOAR) algorithm that cures this instability of SOAR. An extended stability analysis for BIRKA (as above, a popular MOR algorithm for *first-order* bilinear dynamical systems) has been recently done in [@BIRKAstabilitylaapaper]. For rest of this manuscript, whenever stability analysis is referred, we mean it with respect to inexact linear solves. We follow the stability analysis framework of BIRKA from [@BIRKAstabilitylaapaper] and propose equivalent theorems for TBIRKA. The approach here is slightly different, which forms our most *novel* contribution. Norm of the dynamical system plays an important role in stability analysis (the kind of norm is discussed later). In BIRKA stability analysis, a single expression for bilinear dynamical system norm is used (involving a Volterra series). In TBIRKA stability analysis, a similar single expression (involving a truncated Volterra series) leads to complications. Alternatively, in TBIRKA, because of truncation, the bilinear dynamical system can be represented by a finite set of functions. This was not possible in BIRKA where infinite such functions were needed. Thus, in TBIRKA stability analysis, we use norm of all such functions leading to easier derivations. Our stability analysis, as done for BIRKA earlier and for TBIRKA here, can be easily extended to other MOR algorithms for bilinear dynamical systems, e.g., projection based [@biliner_model_reduction_Zbai], implicit Volterra series [@bilinear_model_reduction_Mian_Ilyas], balanced truncation [@bilinear_model_reduction_peter_benner_balanced_trunc], gramian based [@bilinear_model_reduction_kang-li_xu_gramian_based], etc. The rest of the paper is divided into three more parts. In Section \[sec:Background\], we first give a brief overview of MOR for bilinear dynamical systems using a projection method. Next, we review the stability analysis of BIRKA from [@BIRKAstabilitylaapaper]. A stability analysis typically involves satisfying two conditions. Hence, finally in this section, we study the first condition for a stable TBIRKA. We analyze the second condition of stability in TBIRKA’s context in Section \[sec:Stability\_Analysis\]. In Section \[sec:Conclusions Future Work\], we give conclusions and discuss the future work. For the rest of this paper, we use the terms and notations as listed below. 1. The $H_2-$norm is a functional norm defined as [@sarahinexactlaapaper; @garretphdthesis; @TBIRKApaper] $$\begin{aligned} {2}\label{eq:H_2normdefinitionsubsystem} \left \| H_k \right \|_{H_2}^2 = \left ( \frac{1}{2\pi } \right )^k \int_{-\infty}^{\infty} \ldots \int_{-\infty}^{\infty} \left \| H_k\left ( i\omega_1, \ \ldots, \ i \omega_k \right ) \right \|_F^2 d\omega_1\ldots d\omega_k, \end{aligned}$$ where $i$ denotes $\sqrt{-1}$. Here, we assume that all $H_2-$norms computed further exist. In other words, the improper integrals defined by the $H_2-$norm give finite value. This is a reasonable assumption because this happens often in practice (see [@sarahinexactlaapaper], where stability analysis of IRKA is done as well as [@BIRKAstabilitylaapaper], where stability analysis of BIRKA is done). 2. The $H_\infty-$norm is also a functional norm, defined as [@sarahinexactlaapaper; @garretphdthesis; @TBIRKApaper] $$\begin{aligned} {2} \left \| H_k \right \|_{H_\infty} = \underset{\omega_1,\ \ldots, \ \omega_k \in \mathbb{R}}{max} \left \| H_k \left ( i \omega_1,\ \ldots, \ i\omega_k \right ) \right \|_2. \notag \end{aligned}$$ 3. The Kronecker product between two matrices $P$ (of size $m\times n$), and $Q$ (of size $s \times t$) is defined as $$P\otimes Q =\begin{bmatrix} p_{11}Q & \cdots & p_{1n}Q\\ \vdots & \ddots & \vdots\\ p_{m1}Q & \cdots & p_{mn}Q \end{bmatrix}, \notag$$ where $p_{ij}$ is an element of matrix $P$ and order of $P\otimes Q$ is $ms\times nt$. 4. $vec$ operator on a matrix $P$ is defined as $$vec(P)= \left ( p_{11},\ \ldots,\ p_{m1},\ p_{12},\ \ldots,\ p_{m2},\ \ldots \ \ldots,\ p_{1n},\ \ldots,\ p_{mn} \right )^T. \notag$$ 5. Also, $I_n$ denotes an identity matrix of size $n \times n$ and $\mathbb{R}$ denotes the set of real numbers. Background {#sec:Background} ========== A *first-order* bilinear dynamical system is usually represented as [@biliner_model_reduction_Zbai; @BIRKApaper] $$\zeta: \left\{ \begin{array}{ll}\label{eq:originalbilinearequation} \dot{x}(t) = Ax(t) + N x(t) u(t)+bu(t), \\ y(t) = cx(t), \end{array} \right.$$ where $ A,N \in \mathbb{R}^{n \times n}$, $b\in\mathbb{R}^{n \times 1} \ \textnormal{and} \ c \in \mathbb{R}^{1\times n} $. Also, $u(t):\mathbb{R} \rightarrow \mathbb{R}, \ y(t):\mathbb{R} \rightarrow \mathbb{R} \ \textnormal{and} \ x(t): \mathbb{R} \rightarrow \mathbb{R}^n$, represent input, output and state of the bilinear dynamical system, respectively. This is a Single Input Single Output (SISO) system, which we have chosen for ease of our analysis. We plan to look at Multiple Input Multiple Output (MIMO) systems as part of our future work. A bilinear dynamical system can also be represented by a series of transfer functions, i.e., $$\begin{aligned} \label{eq:bilinear_series_representation_infinity} \zeta = \underset{k\rightarrow \infty }{Lim} \ \zeta^k, \end{aligned}$$ where $\zeta^k = \left \{H_{1}\left ( s_{1} \right ),\ H_{2}\left ( s_{1},\ s_{2} \right ),\ \ldots, \ H_k\left(s_1, \ s_2,\ \ldots, \ s_k\right) \right\}$. Here,\ $H_k\left(s_1, \ s_2,\ \ldots, \ s_k\right) $ is called the $k^{th}$ order transfer function of the bilinear dynamical system and is defined as $$\begin{aligned} {2}\label{eq:subsystem_transfer_function} H_{k}\left ( s_{1},\ \ldots,\ s_{k} \right ) =c\left ( s_{k}I-A \right )^{-1}N \left ( s_{k-1}I-A \right )^{-1}\ \ldots N\left ( s_{1}I-A \right )^{-1}b.\end{aligned}$$ After reduction, the bilinear dynamical system can be represented as $$\zeta_r : \left\{ \begin{array}{ll} \label{eq:reducedbilinearequation} \dot{x}_r(t)= A_r x_r (t) + N_r x_r(t) u(t)+ b_r u(t), \\ y_r(t) = c_r x_r(t), \end{array} \right.$$ where $A_r,\ N_r \in \mathbb{R}^{r \times r}$, $ b_r \in \mathbb{R}^{r\times 1} \ \textnormal{and} \ c_r\in \mathbb{R}^{1\times r} \ $ with $r \ll n$. The main goal of model reduction is to approximate $\zeta$ by $\zeta_r$ in an appropriate norm, such that for all admissible inputs, $y_r(t)$ is nearly same to $y(t)$. As mentioned earlier, we use interpolatory projection for performing model reduction. The two common and standard algorithms here, BIRKA and TBIRKA, use a Petrov-Galerkin projection. Let $\mathcal{V}_r$ and $\mathcal{W}_r$ be the two $r$-dimensional subspaces, such that $\mathcal{V}_r = Range (V_r)$ and $\mathcal{W}_r = Range (W_r)$, where $V_r \in \mathbb{R}^{n \times r}$ and $W_r \in \mathbb{R}^{n \times r}$ are matrices. Also, let $\left ( W_r^T V_r \right )$ be invertible[^2]. Applying the projection $x(t)= V_r x_r(t)$, and enforcing the Petrov-Galerkin conditions [@BIRKApaper; @TBIRKApaper] on the original bilinear dynamical system , we get the reduced system as $$\begin{aligned} {2} \begin{split} W_r ^T & \left( V_r \dot{x}_r(t)-A V_r x_r(t) - N V_r x_r(t) u(t)- bu(t) \right) = 0,\\ y(t) & = c V_rx_r(t). \end{split} \end{aligned}$$ Comparing the above two equations with their respective equations in , we get a relation between the original system matrices and the reduced system matrices, i.e., $$\begin{aligned} {2} \begin{split} A_{r}&= \left ( W_r^T V_r \right )^{-1} W_{r}^{T} A V_{r},\ N_r= \left ( W_r^T V_r \right )^{-1} W_{r}^{T} N V_{r}, \\ b_{r}&= \left ( W_r^T V_r \right )^{-1} W_{r}^{T} b, \ \textnormal{and} \ c_{r}= c V_{r}. \end{split}\end{aligned}$$ One way of obtaining subspaces $\mathcal{V}_r$ and $\mathcal{W}_r$ is to use Volterra series interpolation. Further, to decide where to interpolate so as to obtain an optimal reduced model, an $H_2-$optimization problem is commonly solved (Theorem 4.7 from [@garretphdthesis]). [@garretphdthesis] \[Theorem:Interpolation\_condition\_infinity\] Let $\zeta$ be a bilinear system of order n. Let $\zeta_r$ be an $H_2 -$ optimal approximation of order r. Then, $\zeta_r$ satisfies the following multi-point Volterra series interpolation conditions: $$\begin{aligned} {2} \sum_{k=1}^{\infty}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots, \ l_{k}} H_k \left( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k} \right) \ = \notag\\ \sum_{k=1}^{\infty}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} H_{r_k} \left( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k} \right),\quad \textnormal{and} \notag \\ \sum_{k=1}^{\infty}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} \left ( \sum_{j=1}^{k} \frac{\partial }{\partial s_j} H_k\left ( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k}\right ) \right ) = \notag \\ \sum_{k=1}^{\infty}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} \left ( \sum_{j=1}^{k} \frac{\partial }{\partial s_j} H_{r_k}\left ( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots, \ -\lambda_{l_k}\right ) \right ),\notag \end{aligned}$$ where $\phi_{l_1,\ l_2,\ \ldots,\ l_{k}}$ and $\lambda_{l_1},\ \lambda_{l_2},\ \ldots,\ \lambda_{l_k}$ are residues and poles of the transfer function $H_{r_k}$ associated with $\zeta_r$, respectively. BIRKA is designed in such a way that at convergence, the conditions of Theorem \[Theorem:Interpolation\_condition\_infinity\] are satisfied leading to a locally $H_2-$optimal reduced model. Algorithm \[BIRKAAlgo\] lists BIRKA. \ Given an input bilinear dynamical system $A,\ N , \ b,\ c$.\ Select an initial guess for the reduced system as $\check{A},\ \check{N},\ \check{b},\ \check{c}$. Also select stopping tolerance $btol$.\ While $\left ( \textnormal{relative change in eigenvalues of} \ \check{A} \geq btol \right )$ 1. $R \Lambda R^{-1} = \check{A},\ \check{\check{b}}= \check{b}^T R^{-T}, \ \check{\check{c}}=\check{c}R, \ \check{\check{N}}= R^T \check{N} R^{-T}$. 2. $vec\left ( \mathbf{V }\right ) = \left ( -\Lambda \otimes I_n - I_{r} \otimes A - \check{\check{N}}^T \otimes N \right )^{-1} \left ( \check{\check{b}}^T \otimes b\right )$. 3. $vec\left ( \mathbf{W}\right ) = \left ( -\Lambda \otimes I_n - I_{r} \otimes A^T - \check{\check{N}} \otimes N^T \right )^{-1} \left ( \check{\check{c}}^T \otimes c^T\right )$. 4. $\mathbf{V}_r = orth\left ( \mathbf{V } \right ) , \ \mathbf{W}_r = orth\left ( \mathbf{W} \right ) $. 5. $\check{A}= (\mathbf{W}^T_r \mathbf{V}_r)^{-1} \mathbf{W}^T_r A \mathbf{V}_r, \quad \check{N}= \left ( \mathbf{W}^T_r \mathbf{V}_r \right )^{-1} \mathbf{W}^T_r N \mathbf{V}_r,$ 6. $\check{b}=\left ( \mathbf{W}^T_r \mathbf{V}_r \right )^{-1} \mathbf{W}^T_r b, \quad \check{c}= c\mathbf{V}_r.$ \ $A_r = \check{A}, \quad N_r = \check{N}, \quad b_r =\check{b}, \quad c_r = \check{c}$. TBIRKA is similar to BIRKA in most aspects except that it performs a truncated Volterra series interpolation. Here, instead of $\zeta$ in -, they work with $\zeta^M$, which is defined as $$\begin{aligned} \label{eq:bilinear_series_representation_truncated} \zeta^M = \left\{ H_{1}\left ( s_{1} \right), \ H_{2}\left ( s_{1},\ s_{2} \right),\ H_{3}\left ( s_{1},\ s_{2},\ s_{3} \right),\ \ldots, \ H_{M}\left ( s_{1},\ \ldots, \ s_{M} \right ) \right \},\end{aligned}$$ with $H_{k}\left ( s_{1},\ \ldots, \ s_{k} \right )$ for $k=1, \ \ldots, \ M$ is given by . Equivalent of Theorem \[Theorem:Interpolation\_condition\_infinity\] above is as follows (Theorem 4.8 from [@garretphdthesis]): [@garretphdthesis] \[Theorem:Interpolation\_condition\_M\_term\] Let $\zeta=(A, N, b, c)$ be an order $n$ bilinear system and $\zeta^M$ be the polynomial system determined by $\zeta$. Let $\zeta_r=(A_r, N_r, b_r, c_r)$ be a bilinear system of order $r$, and define $\zeta_r^M$ as the polynomial system determined by $\zeta_r$. Suppose that $\zeta_r^M$ is an $H_2-$optimal approximation to $\zeta^M$. Then $\zeta_r^M$ satisfies $$\begin{aligned} {2} \sum_{k=1}^{M}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots, \ l_{k}} H_k \left( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k} \right) \ = \notag\\ \sum_{k=1}^{M}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} H_{r_k} \left( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k} \right),\quad \textnormal{and} \notag \end{aligned}$$ $$\begin{aligned} {2} \sum_{k=1}^{M}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} \left ( \sum_{j=1}^{k} \frac{\partial }{\partial s_j} H_k\left ( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots,\ -\lambda_{l_k}\right ) \right ) = \notag \\ \sum_{k=1}^{M}\sum_{l_1=1}^{r} \ldots \sum_{l_{k}=1}^{r} \phi_{l_1,\ l_2,\ \ldots,\ l_{k}} \left ( \sum_{j=1}^{k} \frac{\partial }{\partial s_j} H_{r_k}\left ( -\lambda_{l_1},\ -\lambda_{l_2},\ \ldots, \ -\lambda_{l_k}\right ) \right ),\notag \end{aligned}$$ where $\phi_{l_1,\ l_2,\ \ldots,\ l_{k}}$ and $\lambda_{l_1},\ \lambda_{l_2},\ \ldots,\ \lambda_{l_k}$ are residues and poles of the transfer function $H_{r_k}$ associated with $\zeta_r^M$, respectively. Algorithm \[TBIRKAAlgo\] lists TBIRKA. Both BIRKA and TBIRKA in turn require solving large sparse linear systems of equations. If we compare Algorithm \[BIRKAAlgo\] and \[TBIRKAAlgo\], we realize that the cost of linear solvers at each step of the `While` loop in the former is $\mathcal{O}\left(nr \times nr\right)$ and in the latter is $\mathcal{O}\left(M\times nr \times nr \right)$. This makes it seem that TBIRKA is more expensive than BIRKA. However, TBIRKA is implemented in such a way that the Kronecker products are avoided leading to cost of linear solves at each step of the `While` loop to be $\mathcal{O}\left(M \times r \times \left(n \times n\right)\right)$. Thus, if $M < r$, which is usually the case (e.g, $M=3 \ \textnormal{or}\ 4$ and $r \gg M$), then TBIRKA is more efficient than BIRKA. For further details on this see chapter 4 in [@garretphdthesis] and Section 5.3 in [@TBIRKApaper]. These implementation details do not affect our stability analysis, and hence, we use Algorithm \[TBIRKAAlgo\] in the current form as our base. \ Given an input bilinear dynamical system $A,\ N, \ b,\ c$.\ Select an initial guess for the reduced system as $ \check{A},\ \check{N},\ \check{b},\ \check{c}$. Also select the truncation index $M$ and stopping tolerance $tbtol$.\ While $\left ( \textnormal{relative change in eigenvalues of} \ \check{A} \geq tbtol \right )$ 1. $R \Lambda R^{-1} = \check{A},\ \check{\check{b}}= \check{b}^T R^{-T}, \ \check{\check{c}}=\check{c}R,\ \check{\check{N}}= R^T \check{N} R^{-T}$. 2. Compute $$\begin{aligned} & vec\left ( \mathbf{V}_{1} \right ) = \left ( -\Lambda\otimes I_{n} -I_{r}\otimes A \right )^{-1} \left(\check{\check{b}}^T \otimes b\right),\notag \\ & vec\left ( \mathbf{W}_{1} \right ) = \left ( -\Lambda\otimes I_{n} -I_{r}\otimes A^{T} \right )^{-1} \left(\check{\check{c}}^T \otimes c^T\right). \notag \end{aligned}$$ 3. For $j = 2, \ldots, M$, solve $$\begin{aligned} & vec\left ( \mathbf{V}_{j} \right ) = \left ( -\Lambda\otimes I_{n} -I_{r}\otimes A \right )^{-1} \left ( \check{\check{N}}^T \otimes N \right ) vec\left ( \mathbf{V}_{j-1} \right ), \notag \\ & vec\left ( \mathbf{W}_{j} \right ) = \left ( -\Lambda\otimes I_{n} -I_{r}\otimes A^{T} \right )^{-1} \left ( \check{\check{N}} \otimes N^{T} \right ) vec\left ( \mathbf{W}_{j-1} \right ). \notag \end{aligned}$$ 4. $\mathbf{V} = \sum\limits_{j=1}^{M} \mathbf{V}_j , \ \mathbf{W} = \sum\limits_{j=1}^{M} \mathbf{W}_j $. 5. $\mathbf{V}_r = orth\left ( \mathbf{V} \right ) , \ \mathbf{W}_r = orth\left ( \mathbf{W} \right ) $. 6. $\check{A}= (\mathbf{W}^T_r \mathbf{V}_r)^{-1} \mathbf{W}^T_r A \mathbf{V}_r,\quad \check{N}= \left ( \mathbf{W}^T_r \mathbf{V}_r \right )^{-1} \mathbf{W}^T_r N \mathbf{V}_r,$ 7. $\check{b}=\left ( \mathbf{W}^T_r \mathbf{V}_r \right )^{-1} \mathbf{W}^T_r b, \quad \check{c}= c\mathbf{V}_r$. \ $A_r = \check{A}, \quad N_{r} = \check{N}, \quad b_r =\check{b}, \quad c_r = \check{c}$. As mentioned earlier, using iterative methods for solving such linear systems introduces approximation errors. We have done a detailed stability analysis of BIRKA with respect to the inexact linear solves in [@BIRKAstabilitylaapaper], and we briefly revisit this next. Generally, accuracy is the metric that tells about the correctness in the output of an inexact algorithm. Due to unavailability of the exact output, it is not possible to determine accuracy [@TrefethenBauBook; @BIRKAstabilitylaapaper]. A more easier metric is stability. Backward stability is one such notation, which says “A backward stable algorithm gives exactly the right output to nearly the right input" [@TrefethenBauBook]. In our context, theoretically we obtain two reduced systems. One by applying an inexact MOR algorithm (with iterative linear solves) on the original full model, and other by applying the same MOR algorithm but exactly (with direct linear solves) on a perturbed full model (the perturbation is introduced in the original full model as part of stability analysis, and is an unknown quantity). If these two reduced systems are equal (*first condition*), with the difference between the original full model and the perturbed full model equal to the order of perturbation (*second condition*), then the MOR algorithm under consideration is called backward stable. The two theorems summarizing this stability analysis for BIRKA are listed below. [@BIRKAstabilitylaapaper] \[Theorem:BIRKA\_firstcondition\] If the inexact linear solves in BIRKA (lines 3b. and 3c. of Algorithm \[BIRKAAlgo\]) are solved using a Petrov-Galerkin framework, then BIRKA satisfies the first condition of backward stability with respect to these solves. [@BIRKAstabilitylaapaper]\[Theorem:BIRKA\_secondcondition\] Let $\widehat{Q} = \Biggl( - \begin{bmatrix} A & 0\\ 0 & A \end{bmatrix} \otimes \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} - \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} \otimes \begin{bmatrix} A & 0\\ 0 & A \end{bmatrix} - \begin{bmatrix} N & 0\\ 0 & N \end{bmatrix} \otimes \begin{bmatrix} N & 0\\ 0 & N \end{bmatrix} \Biggr)$, where $I_n$ is an identity matrix of size $n \times n$ and $\otimes$ denotes the standard Kronecker product. Also, let $\widehat{\widehat{F}} = \left(I_{2n} \otimes \widehat{F} + \widehat{F} \otimes I_{2n} \right)$ with $ \widehat{F} =\begin{bmatrix} 0 & 0\\ 0 & F \end{bmatrix} $, where $F$ is the perturbation introduced in $A$ matrix of the input dynamical system and $I_{2n}$ is an identity matrix of size $2n \times 2n$. If $\widehat{Q}$ is invertible, $\left \| \widehat{Q}^{-1} \right \|_2 < 1$, and $\left \| \widehat{\widehat{F}} \right \|_2 < 1$, then BIRKA satisfies the second condition of backward stability with respect to the inexact linear solves. Next, we analyze the backward stability of TBIRKA. Here, the *first condition* is satisfied in a way similar to that of BIRKA except that some extra orthogonality conditions are imposed on the linear solver (discussed below). \[Theorem:TBIRKA\_firstcondition\] Let the inexact linear solves in TBIRKA (lines 3b. and 3c. of Algorithm \[TBIRKAAlgo\]) are solved using a Petrov-Galerkin framework with $$\begin{aligned} {2} \label{eq:orthogonality_condition_theorem} \sum_{j=1}^{M} W_j \perp R_{b_{j}} \quad \textnormal{and} \quad \sum_{j=1}^{M} V_j \perp R_{c_{j}}, \end{aligned}$$ where $R_{b_{1}}$ and $R_{b_{j}}$ are the residuals in the first equations of 3b. and 3c. of Algorithm \[TBIRKAAlgo\], respectively; $R_{c_{1}}$ and $R_{c_{j}}$ are the residuals in the second equations of 3b. and 3c. of Algorithm \[TBIRKAAlgo\], respectively; and $j=2, \ \ldots, \ M.$ Then, TBIRKA satisfies the first condition of backward stability with respect to these solves. Follows the same pattern as the proof for Theorem 2.1 in [@BIRKAstabilitylaapaper]. From the above theorem, we infer that the underlying iterative solver should *firstly* be based upon a Petrov-Galerkin framework. Since BiConjugate Gradient (i.e., BiCG) is one such algorithm [@greenbaum_book; @SaaditerativeBook], we propose its use in TBIRKA. This is exactly same as for BIRKA. *Secondly,* this particular solver should also satisfy the extra orthogonalities . These orthogonalities have a form similar to the orthogonalities required while reducing second order linear dynamical systems ((23) and (24) in [@NavSISC]), and can be easily satisfied by using a recycling variant of the underlying iterative solver. In [@NavSISC], the ideal iterative solver to be used is Conjugate Gradient (i.e., CG) [@greenbaum_book; @SaaditerativeBook] (due to the use of Galerkin projection). Hence, to satisfy the similar orthogonalities there, without any extra cost, the authors use Recycling Conjugate Gradient (i.e., RCG) [@Michael_RCG]. Since here BiCG is the ideal iterative solver (as discussed above), we propose the use of Recycling BiConjugate Gradient (i.e., RBiCG) [@ahuja2012recyclingsiampaper; @kahujaphdthesis], which would ensure that orthogonalities given by are satisfied without any extra cost. To satisfy the *second condition* of backward stability of TBIRKA, we need to show that $$\begin{aligned} {2}\label{eq:second_condition_of_backward_stability_truncated_zeta} \left \| \zeta^M - \widetilde{\zeta}^M \right \|_{H_2} = \mathcal{O} \left ( \left \| F \right \|_2 \right ),\end{aligned}$$ where $\zeta^M$ is given by or $$\begin{aligned} &\zeta^M = \left\{ H_{1}\left ( s_{1} \right), \ H_{2}\left ( s_{1},\ s_{2} \right),\ \ldots, \ H_{M}\left ( s_{1},\ \ldots, \ s_{M} \right ) \right \},\label{eq:bilinear_series_representation_truncated_using_in_proof_original_system}\\ \intertext{with $H_{k}\left ( s_{1},\ \ldots,\ s_{k} \right )$ for $k=1,\ \ldots, \ M$ given by \eqref{eq:subsystem_transfer_function} or} \begin{split} & H_{k}\left ( s_{1},\ \ldots,\ s_{k} \right ) =c\left ( s_{k}I-A \right )^{-1}\\ &\qquad\qquad\qquad\qquad N \left ( s_{k-1}I-A \right )^{-1}\ \ldots N\left ( s_{1}I-A \right )^{-1}b,\label{eq:susbsytem_transfer_function_in_proof_original} \end{split} \end{aligned}$$ $$\begin{aligned} &\widetilde{\zeta}^M = \left\{ \widetilde{H}_{1}\left ( s_{1} \right), \ \widetilde{H}_{2}\left ( s_{1},\ s_{2} \right),\ \ldots, \ \widetilde{H}_{M}\left ( s_{1},\ \ldots, \ s_{M} \right ) \right \},\label{eq:bilinear_series_representation_truncated_using_in_proof_perturbed_system}\\ \intertext{with for $k=1,\ \ldots, \ M$} \begin{split} & \widetilde{H}_{k}\left ( s_{1},\ \ldots,\ s_{k} \right ) =c\left ( s_{k}I-\left(A+F\right) \right )^{-1}\\ & \qquad\qquad\qquad\qquad N \left ( s_{k-1}I-\left(A+F\right) \right )^{-1}\ \ldots N\left ( s_{1}I-\left(A+F\right) \right )^{-1}b,\label{eq:susbsytem_transfer_function_in_proof_perturbed} \end{split} \end{aligned}$$ and assuming perturbation in $A$ matrix of the input dynamical system (as for BIRKA stability; see Theorem \[Theorem:BIRKA\_secondcondition\] earlier). One way to satisfy is to use the definition of the $H_2-$norm of $\zeta^M-\widetilde{\zeta}^M$, i.e., from Lemma 5.1 of [@TBIRKApaper] $$\begin{aligned} {4} \begin{split} \label{eq:H_2_norm_error_system_truncated_Voterra} & \left \| \zeta^M -\widetilde{\zeta}^M \right \|_{H_2}^{2} = \left ( \begin{bmatrix} c & -c \end{bmatrix} \otimes \begin{bmatrix} c & -c \end{bmatrix} \right ) \sum_{j=0}^{M} \Biggl[ \biggl ( - \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} \otimes \\ & \quad \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} - \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} \otimes \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} \biggr )^{-1} \begin{bmatrix} N & 0 \\ 0 & N \end{bmatrix} \otimes \begin{bmatrix} N & 0 \\ 0 & N \end{bmatrix} \Biggr]^j \\ & \quad \biggl ( - \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} \otimes \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} - \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} \otimes \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} \biggr )^{-1} \\ & \quad \left ( \begin{bmatrix} b\\ b \end{bmatrix} \otimes \begin{bmatrix} b\\ b \end{bmatrix} \right ). \end{split}\end{aligned}$$ This approach is followed in satisfying the *second condition* of backward stability of BIRKA, but for TBIRKA it turns out to be more challenging. The reason for this is that the definition of the $H_2-$norm of $\zeta-\widetilde{\zeta}$ used in BIRKA is different from [^3], i.e., from Corollary 4.1 of [@BIRKApaper] or Theorem 4.5 of [@garretphdthesis] $$\begin{aligned} {4} %\begin{split} & \left \| \zeta -\widetilde{\zeta} \right \|_{H_2}^{2} = \left ( \begin{bmatrix} c & -c \end{bmatrix} \otimes \begin{bmatrix} c & -c \end{bmatrix} \right ) \biggl ( - \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} \otimes \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} - \\ & \qquad \begin{bmatrix} I_n & 0\\ 0 & I_n \end{bmatrix} \otimes \begin{bmatrix} A & 0\\ 0 & A+F \end{bmatrix} - \begin{bmatrix} N & 0\\ 0 & N \end{bmatrix} \otimes \begin{bmatrix} N & 0\\ 0 & N \end{bmatrix} \biggr )^{-1} \left ( \begin{bmatrix} b\\ b \end{bmatrix} \otimes \begin{bmatrix} b\\ b \end{bmatrix} \right ). \notag %\end{split}\end{aligned}$$ Another way to satisfy in case of TBIRKA, which turns out to be more easier, is to show that $$\begin{aligned} {2}\label{eq:second_condition_of_backward_stability_series} \begin{split} \left \| H_{1}\left ( s_{1} \right ) - \widetilde{H}_{1}\left ( s_{1} \right ) \right \|_{H_{2}} & \propto \mathcal{O}\left ( \left \| F \right \|_2 \right ), \\ \left \| H_{2}\left ( s_{1},\ s_{2} \right ) - \widetilde{H}_{2}\left ( s_{1},\ s_{2} \right ) \right \|_{H_{2}} &\propto \mathcal{O}\left ( \left \| F \right \|_2 \right ), \\ & \vdots\\ \left \| H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}} &\propto \mathcal{O}\left ( \left \| F \right \|_2 \right ), \end{split}\end{aligned}$$ where $ H_{k}\left ( s_{1},\ \ldots,\ s_{k} \right )$ for $k=1, \ \ldots,\ M$ is given by and , and $\widetilde{H}_{k}\left ( s_{1},\ \ldots,\ s_{k} \right )$ for $k=1, \ \ldots,\ M$ is given by . This way was not possible in BIRKA because there $M \rightarrow \infty$ (see -). Satisfying the Second Condition of Backward Stability {#sec:Stability_Analysis} ===================================================== We use induction to prove . To prove this condition, we first abstract out the term containing the perturbation $F$ from the normed difference between the transfer function of the original system $\left( H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right)$ and the transfer function of the perturbed system $\left( \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right)$ (in Lemma \[Lemma:Mth\_term\_H2\_norm\]). Next, we use induction to prove that this abstracted term is $\mathcal{O}\left ( \left \| F \right \|_2 \right )$ (in Lemma \[Lemma:induction\_proof\]). Finally, we use the result of these two lemmas to prove (in Theorem \[Theorem:second\_condition\_subsystem\]). Note, that in all our subsequent derivations, we assume that all inverses used exist. This is an acceptable assumption because the inverse of matrices arising here are of the form as in [@sarahinexactlaapaper] and [@BIRKAstabilitylaapaper] (the papers that discuss stability of IRKA and BIRKA, respectively). \[Lemma:Mth\_term\_H2\_norm\] If $$\begin{aligned} H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) =c\left ( s_{M}I_n-A \right )^{-1}N \left ( s_{M-1}I_n-A \right )^{-1} \ldots N\left ( s_{1}I_n-A \right )^{-1}b\end{aligned}$$ and $$\begin{aligned} \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) = & c\left ( s_{M}I_n-\left ( A+F \right ) \right )^{-1} \\ & \ N \left ( s_{M-1}I_n- \left ( A+F \right ) \right )^{-1} \ldots N\left ( s_{1}I_n- \left ( A+F \right ) \right )^{-1}b, \end{aligned}$$ then the $H_2-$norm of their difference $$\begin{aligned} {3} &\left \| H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}}^{2} \leq \left \| c \mathcal{K}^{-1}\left(s_{M}\right) \right \|_{H_{2}}^{2} \\ &\quad \left \| \mathcal{K}^{-1}\left(s_{M-1}\right) \right \|_{H_{2}}^{2} \ldots \left \| \mathcal{K}^{-1}\left(s_{1}\right) \right \|_{H_{2}}^{2} \left \| U(s_1,\ \ldots,\ s_M) \right \|_{H_{\infty}}^{2} \left \| \mathcal{K}^{-1}\left(s_1\right)b \right \|_{H_{\infty}}^{2} \end{aligned}$$ where, $ \mathcal{K}\left(s_i\right)=\left ( s_{i}I_{n}-A \right ) $ for $i= 1, \ \ldots, \ M$ and $$\begin{aligned} {2} \label{eq:U_expression} \begin{split} & U(s_1, \ldots,\ s_{M}) = \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left ( s_{M-1} \right ) \Bigg( N \mathcal{K}^{-1}\left(s_{M-1}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\qquad -\left ( I_n- F \mathcal{K}^{-1}\left(s_M\right) \right )^{-1} N \mathcal{K}^{-1}\left(s_{M-1}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M-1}\right) \right )^{-1} \ldots \\ &\qquad\quad N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ Using the definition of $H_2-$norm , we get $$\begin{aligned} {2} & \left \| H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}}^{2} \notag\\ & = \left(\frac{1}{2\pi }\right)^M \underset{m\rightarrow \infty}{Lim} \int_{-m}^{m} \ldots \int_{-m}^{m} \left \|c \mathcal{K}^{-1}\left(i\omega_M\right) N \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\ldots N\mathcal{K}^{-1}\left(i\omega_1\right)b \ \right.\notag\\ & \quad - c \left(\mathcal{K}\left ( i\omega_{M} \right)-F\right )^{-1} N \left(\mathcal{K}\left ( i\omega_{M-1} \right)-F\right )^{-1} \ldots N \left(\mathcal{K}\left ( i\omega_{2} \right)-F\right )^{-1} \notag\\ & \qquad \left. N \left(\mathcal{K}\left ( i\omega_{1}\right)-F\right )^{-1}b \right \|_{F}^{2} d\omega_{1} \ldots d\omega_{M} \notag \end{aligned}$$ $$\begin{aligned} {2} & = \left(\frac{1}{2\pi }\right)^M \underset{m\rightarrow \infty}{Lim} \int_{-m}^{m} \ldots \int_{-m}^{m} \left \| c \mathcal{K}^{-1}\left(i\omega_M\right)\Bigg( N \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\ldots N \mathcal{K}^{-1}\left(i\omega_2\right) N \ \right.\notag\\ & \quad -\left(I_n-F \mathcal{K}^{-1}\left(i\omega_M\right)\right )^{-1} N \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\left(I_n-F \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\right )^{-1} \ldots\notag\\ & \qquad \left. N \mathcal{K}^{-1}\left(i\omega_2\right)\left(I_n-F \mathcal{K}^{-1}\left(i\omega_2\right)\right )^{-1} N \left(I_n- \mathcal{K}^{-1}\left(i\omega_1\right)F\right )^{-1}\Bigg) \mathcal{K}^{-1}\left(i\omega_1\right) b \right \|_{F}^{2} \notag \\ & \qquad d\omega_{1} \ldots d\omega_{M} \notag\\ \begin{split}\notag & = \left(\frac{1}{2\pi }\right)^M \underset{m\rightarrow \infty}{Lim} \int_{-m}^{m} \ldots \int_{-m}^{m} \left \| c \mathcal{K}^{-1}\left(i\omega_M\right) \mathcal{K}^{-1}\left(i\omega_{M-1}\right) \ldots \mathcal{K}^{-1}\left(i\omega_1\right) \right.\\ &\qquad \mathcal{K}\left(i\omega_{1}\right)\ldots \mathcal{K}\left(i\omega_{M-1}\right) \Bigg( N \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\ldots N \mathcal{K}^{-1}\left(i\omega_2\right) N \ \\ & \quad - \left(I_n-F \mathcal{K}^{-1}\left(i\omega_M\right)\right )^{-1} N \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\left(I_n-F \mathcal{K}^{-1}\left(i\omega_{M-1}\right)\right )^{-1} \ldots \\ & \qquad \left. N \mathcal{K}^{-1}\left(i\omega_2\right)\left(I_n-F \mathcal{K}^{-1}\left(i\omega_2\right)\right )^{-1} N \left(I_n- \mathcal{K}^{-1}\left(i\omega_1\right)F\right )^{-1}\Bigg) \mathcal{K}^{-1}\left(i\omega_1\right) b \right \|_{F}^{2}\\ & \qquad d\omega_{1} \ldots d\omega_{M}. \end{split} \end{aligned}$$ Using $U\left(s_1, \ \ldots, \ s_M\right)$ given by , $\left \| X Y Z \right \|_{F} \leq \left \| X \right \|_{F} \left \| Y Z \right \|_{F}$, $\left \| Y Z \right \|_{F} \leq \left \| Y \right \|_{F} \left \| Z \right \|_{2}$, and comparison integral inequality[^4] [@Alen_Jeffrey__mean_value_theorem_book] for any matrices $X,\ Y$ and $Z$, in the above equation, we have $$\begin{aligned} {2}\label{eq:Mth_error_term_H2_norm_intermediate_limit} \begin{split} & \left \| H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}}^{2} \\ &\quad \leq \left(\frac{1}{2\pi }\right)^M \underset{m\rightarrow \infty}{Lim} \int_{-m}^{m} \ldots \int_{-m}^{m} \left \| c \mathcal{K}^{-1}\left(i\omega_M\right) \right\|_F^2 \left \| \mathcal{K}^{-1}\left(i\omega_{M-1}\right) \right\|_F^2 \ldots \\ & \qquad \left \| \mathcal{K}^{-1}\left(i\omega_1\right) \right\|_F^2 \left \| U \left(i\omega_{1}, \ \ldots, \ i \omega_{M}\right) \right\|_2^2 \left \| \mathcal{K}^{-1}\left(i\omega_1\right) b \right\|_2^2 d\omega_{1} \ldots d\omega_{M}. \end{split} \end{aligned}$$ From the mean value theorem of integration [@Alen_Jeffrey__mean_value_theorem_book] we know $$\begin{aligned} {4} \int_{-m}^{m} \int_{-m}^{m} f\left ( i\omega_{2} \right ) & g\left ( i\omega _{1},i\omega _{2} \right ) h\left ( i\omega _{1} \right ) d\omega _{1}d\omega _{2} \notag\\ = & \int_{-m}^{m} f\left ( i\omega_{2} \right ) \left ( \int_{-m}^{m} g\left ( i\omega _{1},i\omega _{2} \right ) h\left ( i\omega _{1} \right )d\omega _{1} \right) d\omega _{2}\notag \end{aligned}$$ $$\begin{aligned} {2} \leq & \int_{-m}^{m} f\left ( i\omega_{2} \right ) \left ( \underset{c\in \mathbb{R}}{max}\left ( g(ic,i\omega _{2}) \right ) \int_{-m}^{m} h\left ( i\omega _{1} \right )d\omega _{1} \right) d\omega _{2} \notag\\ \leq & \left ( \int_{m}^{m} f\left ( i\omega_{2} \right ) \underset{c\in \mathbb{R}}{max}\left ( g(ic,i\omega _{2}) \right ) d\omega _{2} \right ) \int_{-m}^{m} h\left ( i\omega _{1} \right )d\omega _{1} \notag\\ \leq & \ \underset{c,d\in \mathbb{R}}{max} (g(ic,id)) \int_{-m}^{m} f\left ( i\omega_{2} \right ) d\omega _{2} \int_{-m}^{m} h\left ( i\omega _{1} \right )d\omega _{1}. \notag \end{aligned}$$ Using this property in we get[^5] $$\begin{aligned} {2} & \left \| H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}}^{2} \\ &\qquad \leq \left(\frac{1}{2\pi }\right)^M \underset{m\rightarrow \infty}{Lim} \int_{-m}^{m} \ldots \int_{-m}^{m} \left \| c \mathcal{K}^{-1}\left(i\omega_M\right) \right\|_F^2 \left \| \mathcal{K}^{-1}\left(i\omega_{M-1}\right) \right\|_F^2 \ldots\\ & \qquad\qquad \left \| \mathcal{K}^{-1}\left(i\omega_1\right) \right\|_F^2 d\omega_{1} \ldots d\omega_{M} \underset{\omega_1,\ \ldots,\ \omega_M \in \mathbb{R}}{max} \left \| U \left(i\omega_{1}, \ \ldots, \ i \omega_{M}\right) \right\|_2^2 \ \\ &\qquad\qquad \underset{\omega_1 \in \mathbb{R}}{max} \left \| \mathcal{K}^{-1}\left(i\omega_1\right) b \right\|_2^2 \\ &\qquad \leq \left \| c \mathcal{K}^{-1}\left(s_M\right) \right\|_{H_2}^2 \left \| \mathcal{K}^{-1}\left(s_{M-1}\right) \right\|_{H_2}^2 \ldots \left \| \mathcal{K}^{-1}\left(s_1\right) \right\|_{H_2}^2 \\ & \qquad\qquad \left \| U \left(s_{1}, \ \ldots, \ s_{M}\right) \right\|_{H_{\infty}}^2 \left \| \mathcal{K}^{-1}\left(s_1\right) b \right\|_{H_{\infty}}^2. \end{aligned}$$ \[Lemma:induction\_proof\] If $U_{M} = U(s_1,\ \ldots,\ s_{M}) $ defined in , $\left \| F \right \|_2 < 1 $, and\ $\left \| \mathcal{K}^{-1} \left ( s \right ) \right \|_{H_{\infty}}<1$, where $\mathcal{K}\left(s\right) = \left(s I_n - A\right)$. Then we have $$\begin{aligned} \left\| U_{M} \right\|_{H_{\infty}} \propto \mathcal{O} \left( \left\| F \right\|_2 \right). \end{aligned}$$ We prove this by mathematical induction.\ 1. *First subsystem*\ This is the linear system case, which has been already proved in [@sarahinexactlaapaper] (see Theorem 4.3 of [@sarahinexactlaapaper]). 2. *Second subsystem*\ Substituting $M=2$ in , we get $$\begin{aligned} U_2 = \mathcal{K}\left(s_1\right)\left(N- \left ( I_n-F\mathcal{K}^{-1} \left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1} \right). \end{aligned}$$ If $ \left\| F\mathcal{K}^{-1} \left(s_2\right)\right\|_{H_{\infty}} < 1$ and $ \left\| \mathcal{K}^{-1} \left(s_1\right) F\right\|_{H_{\infty}}<1$, then by the Neumann series, we get[^6] $$\begin{aligned} {2} U_2 &= \mathcal{K}\left(s_1\right)\bigg(N- \left ( I_n+F\mathcal{K}^{-1}\left(s_2\right) + \left(F\mathcal{K}^{-1}\left(s_2\right)\right)^2 +\cdots \right ) N \\ &\quad \left ( I_n+ \mathcal{K}^{-1} \left(s_1\right) F + \left(\mathcal{K}^{-1} \left(s_1\right) F\right)^2 +\cdots \right ) \bigg)\\ &= \mathcal{K}\left(s_1\right)\bigg(N- N-N\mathcal{K}^{-1} \left(s_1\right)F\left ( I_n+ \mathcal{K}^{-1} \left(s_1\right) F + \cdots \right ) - \\ & \quad F \mathcal{K}^{-1} \left(s_2\right) \left ( I_n+F\mathcal{K}^{-1}\left(s_2\right) + \cdots \right ) N \\ & \quad \left ( I_n+ \mathcal{K}^{-1} \left(s_1\right) F + \left(\mathcal{K}^{-1} \left(s_1\right) F \right)^2 + \cdots \right ) \bigg)\\ &=\mathcal{K}\left(s_1\right)\bigg(-N\mathcal{K}^{-1} \left(s_1\right)F\left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1} -\\ & \quad F \mathcal{K}^{-1} \left(s_2\right) \left ( I_n-F\mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1} \bigg)\\ &=\mathcal{K}\left(s_1\right)\Big(-N\mathcal{K}^{-1} \left(s_1\right)F - F \mathcal{K}^{-1} \left(s_2\right) \left ( I_n-F\mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \Big)\\ & \quad \left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1}. \end{aligned}$$ Taking $H_{\infty}-$norm on both sides, we get $$\begin{aligned} {2} \left \| U_2 \right \|_{H_{\infty}} &= \left \| \mathcal{K}\left(s_1\right)\Big(-N\mathcal{K}^{-1} \left(s_1\right)F - F \mathcal{K}^{-1} \left(s_2\right) \left ( I_n-F\mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \Big) \right.\\ & \qquad \left. \left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1} \right \|_{H_{\infty}}\\ &= \underset{\omega_1, \omega_2 \in \mathbb{R}}{max} \left \| \mathcal{K}\left(i\omega_1\right)\Big(-N\mathcal{K}^{-1} \left(i\omega_1\right)F - F \mathcal{K}^{-1} \left(i\omega_2\right)\right.\\ &\qquad \left. \left ( I_n-F\mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} N \Big) \left ( I_n- \mathcal{K}^{-1} \left(i\omega_1\right) F \right )^{-1} \right \|_2. \end{aligned}$$ Using $\left\| X Y\right\|_2 \leq \left\| X\right\|_2 \left\| Y\right\|_2$ and $\left\| X+Y\right\|_2 \leq \left\| X\right\|_2 + \left\| Y\right\|_2$, for any two matrices $X$ and $Y$, in the above equation, we get $$\begin{aligned} \left \| U_2 \right \|_{H_{\infty}} \leq & \underset{\omega_1, \omega_2 \in \mathbb{R}}{max} \Bigg(\left \| \mathcal{K}\left(i\omega_1\right)\right \|_2 \bigg( \left \| N \right \|_2 \left \|\mathcal{K}^{-1} \left(i\omega_1\right)\right \|_2 \left \| F \right \|_2 + \notag\\ & \left \| F \right \|_2 \left \| \mathcal{K}^{-1} \left(i\omega_2\right) \right \|_2 \left \| \left ( I_n-F\mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} \right \|_2 \left \| N \right \|_2 \bigg) \notag\\ & \left \|\left ( I_n- \mathcal{K}^{-1} \left(i\omega_1\right) F \right )^{-1} \right \|_2 \Bigg) \notag\\ \begin{split} \label{eq:U_2_expression_inequality_max_inside} \leq &\left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \left \| N \right \|_{2} \left \| F \right \|_{2} \Bigg( \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} + \\ & \left \| \mathcal{K}^{-1} \left(s_2\right) \right \|_{H_{\infty}} \ \underset{\omega_2 \in \mathbb{R}}{max} \left \| \left ( I_n-F\mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} \right \|_2 \Bigg) \\ & \underset{\omega_1 \in \mathbb{R}}{max} \left \|\left ( I_n- \mathcal{K}^{-1} \left(i\omega_1\right) F \right )^{-1} \right \|_2. \end{split} % \leq &\left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \left \| N \right \|_{2} \left \| F \right \|_{2} \left( \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} + \left \| \mathcal{K}^{-1} \left(s_2\right) \right \|_{H_{\infty}} \dfrac{1}{1-\left \| F\mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}} } % \right)\\ % &\left( \dfrac{1}{1-\left \| \mathcal{K}^{-1} \left(s_1\right) F \right \|_{H_{\infty}} } \right) \ldots \ldots(\textnormal{Using lemma 2.3.3 in \cite{g_golub_book}}) . \end{aligned}$$ Technically by definition of the $H_{\infty}-$norm and how $\mathcal{K} \left(s\right)$ is defined in our hypotheses, $\left\| \mathcal{K} \left(s_1\right)\right\|_{H_{\infty}} = \left\| \mathcal{K} \left(s_2\right)\right\|_{H_{\infty}} = \left\| \mathcal{K} \left(s\right)\right\|_{H_{\infty}}$, however, for sake of exposition, we keep them separate. Similarly for the $H_{\infty}-$norm of inverses of $\mathcal{K} \left(s_1\right)$ and $\mathcal{K} \left(s_2\right)$. To abstract $\left\| F\right\|_2$ out from the above inequality, let us look at $\underset{\omega_2 \in \mathbb{R}}{max} \left \| \left ( I_n-F\mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} \right\|_2$ separately. Recall, while applying Neumann series we assumed that $\left \| F \mathcal{K}^{-1} \left ( s_2 \right ) \right \|_{H_{\infty}} <1$ or $ \underset{\omega_2 \in \mathbb{R}}{max} \left \| F \mathcal{K}^{-1} \left ( i\omega_2 \right ) \right \|_2 <1$. Since the maximum of such a norm is less than one, we have for all $\omega_2 \in \mathbb{R}$, $\left \| F \mathcal{K}^{-1} \left ( i\omega_2 \right ) \right \|_2 <1$. Using this along with Lemma 2.3.3 from [@g_golub_book][^7] in the above expression, we get $$\begin{aligned} {3} \underset{\omega_2 \in \mathbb{R}}{max} \left \| \left ( I_n-F\mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} \right\|_2 &\leq \underset{\omega_2 \in \mathbb{R}}{max} \ \frac{1}{1- \left \| F \mathcal{K}^{-1} \left ( i\omega_2 \right ) \right \|_2} \notag\\ &\leq \frac{1}{1- \underset{\omega_2 \in \mathbb{R}}{max} \left \| F \mathcal{K}^{-1} \left ( i\omega_2 \right ) \right \|_2} \notag\\ & \leq \frac{1}{1- \left \| F \mathcal{K}^{-1} \left ( s_2 \right ) \right \|_{H_{\infty}}}. \label{eq:max_to_H_infty_inequality_golub_s2} \end{aligned}$$ If we assume $\left \| F \right \|_2 < 1$ and $\left \| \mathcal{K}^{-1} \left(s\right) \right \|_{H_{\infty}} < 1$ (as in our hypotheses), then using earlier used matrix norm properties, we get $$\begin{aligned} {2} \left\|F\mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}} = \underset{\omega_2 \in \mathbb{R}}{max} \left\|F\mathcal{K}^{-1}\left(i\omega_2\right)\right\|_2 & \leq \left \|F\right \|_2 \underset{\omega_2 \in \mathbb{R}}{max} \left\|\mathcal{K}^{-1}\left(i\omega_2\right)\right\|_2 \\ &\leq \left \|F\right \|_2 \left\|\mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}} \\ &\leq 1, \end{aligned}$$ as assumed for applying Neumann series earlier as well as Lemma 2.3.3 from [@g_golub_book] above. Thus, no extra assumptions beyond those in hypotheses are needed. Further, we also get $$\begin{aligned} {2} 1- \left \|F\right \|_2 \left\|\mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}} & \leq 1- \left\|F \mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}} \qquad \textnormal{or} \notag\\ \frac{1}{ 1- \left\|F \mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}}} & \leq \frac{1}{1- \left \|F\right \|_2 \left\|\mathcal{K}^{-1}\left(s_2\right)\right\|_{H_{\infty}} }. \label{eq:golub_inequality_rational_form_individual_2_infty_norm_for_s2} \end{aligned}$$ Similarly, we can bound the last term of as follows: $$\begin{aligned} {3} \underset{\omega_1 \in \mathbb{R}}{max} \left \| \left ( I_n-\mathcal{K}^{-1}\left(i\omega_1\right)F \right )^{-1} \right\|_2 &\leq \frac{1}{1- \left \| \mathcal{K}^{-1} \left ( s_1 \right ) F \right \|_{H_{\infty}} } \qquad \textnormal{and} \label{eq:max_to_H_infty_inequality_golub_s1}\\ \frac{1}{ 1- \left\| \mathcal{K}^{-1}\left(s_1\right)F\right\|_{H_{\infty}}} & \leq \frac{1}{1- \left\|\mathcal{K}^{-1}\left(s_1\right)\right\|_{H_{\infty}} \left \|F\right \|_2 }. \label{eq:golub_inequality_rational_form_individual_2_infty_norm_for_s1} \end{aligned}$$ Substituting - and - in , we get $$\begin{aligned} \left \| U_2 \right \|_{H_{\infty}} \leq &\left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \left \| N \right \|_{2} \left \| F \right \|_{2} \Bigg[ \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} + \\ & \dfrac{\left \| \mathcal{K}^{-1} \left(s_2\right) \right \|_{H_{\infty}}}{1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}} \Bigg] \left( \dfrac{1}{1-\left \| \mathcal{K}^{-1} \left(s_1\right) \right \|_{H_{\infty}} \left \| F \right \|_{2}} \right). \end{aligned}$$ From the above inequality it is clear that if $\left\|F \right\|_{2} \left\|\mathcal{K}^{-1} \left(s_2\right) \right\|_{H_{\infty}} < 1$ and $\left\| \mathcal{K}^{-1} \left(s_1\right) \right\|_{H_{\infty}} \left\| F \right\|_{2}< 1$, which are true from our hypotheses, then $$\left \| U_2 \right \|_{H_{\infty}} = \mathcal{O} \left ( \left \| F \right \|_2 \right ).$$ 3. *Third subsystem*\ Using $M=3$ in , we get $$\begin{aligned} {2} U_3 = \ & \mathcal{K}\left(s_1\right) \mathcal{K}\left(s_2\right)\Big( N \mathcal{K}^{-1} \left(s_2\right) N - \left ( I_n-F\mathcal{K}^{-1} \left(s_3\right) \right )^{-1} N \mathcal{K}^{-1} \left(s_2\right) \\ & \left ( I_n-F\mathcal{K}^{-1} \left(s_2\right) \right )^{-1} N\left ( I_n- \mathcal{K}^{-1} \left(s_1\right) F \right )^{-1} \Big). \end{aligned}$$ Following the same steps as in the case of the second subsystem we get $$\begin{aligned} \left \| U_3 \right \|_{H_{\infty}} \leq &\left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \left \| \mathcal{K}\left(s_2\right)\right \|_{H_{\infty}} \left \| N \right \|_{2}^2 \left \| F \right \|_{2} \left \| \mathcal{K}^{-1} \left(s_2\right)\right \|_{H_{\infty}} \\ & \Bigg[ \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} + \dfrac{\left \| \mathcal{K}^{-1} \left(s_2\right) \right \|_{H_{\infty}}}{1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}} + \\ & \dfrac{\left \| \mathcal{K}^{-1} \left(s_3\right) \right \|_{H_{\infty}}}{\left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_3\right) \right \|_{H_{\infty}}\right) \left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}\right)} \Bigg] \\& \left( \dfrac{1}{1-\left \| \mathcal{K}^{-1} \left(s_1\right) \right \|_{H_{\infty}} \left \| F \right \|_{2}} \right). \end{aligned}$$ Again, from the above inequality it is clear that if $\left\|F \right\|_{2} \left\|\mathcal{K}^{-1} \left(s_2\right) \right\|_{H_{\infty}}$, $\left\|F \right\|_{2} \left\|\mathcal{K}^{-1} \left(s_3\right) \right\|_{H_{\infty}}$ and $\left\| \mathcal{K}^{-1} \left(s_1\right) \right\|_{H_{\infty}} \left\| F \right\|_{2}$ are all less than $1$, which are true from our hypotheses, then $$\begin{aligned} \left \| U_3 \right \|_{H_{\infty}} = \mathcal{O} \left ( \left \| F \right \|_2 \right ). \end{aligned}$$ \ From , we know $$\begin{aligned} {2}\label{eq:U_expression_induction} \begin{split} U_M &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left ( s_{M-1} \right ) \Bigg( N \mathcal{K}^{-1}\left(s_{M-1}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\quad -\left ( I_n- F \mathcal{K}^{-1}\left(s_M\right) \right )^{-1} N \mathcal{K}^{-1}\left(s_{M-1}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M-1}\right) \right )^{-1} \ldots \\ & \qquad \quad N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ Let $$\begin{aligned} \left \| U_M \right \|_{H_{\infty}} \leq &\left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \ldots \left \| \mathcal{K}\left(s_{M-1}\right)\right \|_{H_{\infty}} \left \| N \right \|_{2}^{M-1} \left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_{M-1}\right)\right \|_{H_{\infty}} \\ & \left \| \mathcal{K}^{-1} \left(s_2\right)\right \|_{H_{\infty}} \Bigg[ \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} + \dfrac{\left \| \mathcal{K}^{-1} \left(s_2\right) \right \|_{H_{\infty}}}{1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}} \\ %& + \dfrac{\left \| \mathcal{K}^{-1} \left(s_3\right) \right \|_{H_{\infty}}}{\left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_3\right) \right \|_{H_{\infty}}\right) \left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}\right)} & \quad \vdots \\ &+ \dfrac{\left \| \mathcal{K}^{-1} \left(s_M\right) \right \|_{H_{\infty}}}{\left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_M\right) \right \|_{H_{\infty}}\right) \ldots \left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}\right)} \Bigg] \\ & \left( \dfrac{1}{1-\left \| \mathcal{K}^{-1} \left(s_1\right) \right \|_{H_{\infty}} \left \| F \right \|_{2}} \right) \quad \textnormal{or}\\ \left \| U_{M} \right \|_{H_{\infty}} =& \ \mathcal{O} \left ( \left \| F \right \|_2 \right ). \end{aligned}$$\ Again, from , we know $$\begin{aligned} {2} \begin{split} U_{M+1} &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\quad -\left ( I_n- F \mathcal{K}^{-1}\left(s_{M+1}\right) \right )^{-1} N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots \\ &\qquad N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ We first write $U_{M+1}$ in terms of $U_M$. Thus, in the above equation using Neumann series, we get $$\begin{aligned} {2} U_{M+1} &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\quad -\left ( I_n + F \mathcal{K}^{-1}\left(s_{M+1}\right) + \left(F \mathcal{K}^{-1}\left(s_{M+1}\right)\right)^2 + \cdots \right ) \\ &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ & \qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg)\\ % &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) % \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N - \\ % &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ % &\qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}-\\ % &\qquad F \mathcal{K}^{-1}\left(s_{M+1}\right) \left ( I_n + F \mathcal{K}^{-1}\left(s_{M+1}\right) + \left(F \mathcal{K}^{-1}\left(s_{M+1}\right)\right)^2 + \cdots \right ) \\ % &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ % & \qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg) \\ &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\quad - N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ &\qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\\ &\quad - F \mathcal{K}^{-1}\left(s_{M+1}\right) \left ( I_n - F \mathcal{K}^{-1}\left(s_{M+1}\right) \right )^{-1} \\ &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ & \qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{aligned}$$ In the above equation, taking $N \mathcal{K}^{-1}\left(s_M\right)$ common from the first two terms of the bigger bracket, we have $$\begin{aligned} {2} \label{eq:U_M+1_expression_comparisson} \begin{split} &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \bigg( N \mathcal{K}^{-1}\left(s_{M-1}\right)\ldots N\mathcal{K}^{-1}\left(s_2\right)N\\ &\ - \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} N \mathcal{K}^{-1}\left(s_{M-1}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M-1}\right) \right )^{-1} \ldots \\ &\quad N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\bigg)\\ &\ - F \mathcal{K}^{-1}\left(s_{M+1}\right) \left ( I_n - F \mathcal{K}^{-1}\left(s_{M+1}\right) \right )^{-1} \\ &\quad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ & \quad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ Now we look at expression of $U_M$. Multiplying $\mathcal{K}^{-1} \left ( s_{M-1} \right )\ldots \mathcal{K}^{-1}\left ( s_{1} \right )$ on both the sides of from left, we get $$\begin{aligned} {2}\label{eq:U_M_expression_rewritten} \begin{split} &\mathcal{K}^{-1} \left ( s_{M-1} \right )\ldots \mathcal{K}^{-1}\left ( s_{1} \right ) U_M \\ &=\Bigg( N \mathcal{K}^{-1}\left(s_{M-1}\right) \ldots N\mathcal{K}^{-1}\left(s_2\right) N \\ &\ - \left ( I_n- F \mathcal{K}^{-1}\left(s_M\right) \right )^{-1} N \mathcal{K}^{-1}\left(s_{M-1}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M-1}\right) \right )^{-1} \ldots \\ &\qquad N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ Substituting in , we get $$\begin{aligned} {2} \begin{split} U_{M+1} &= \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \bigg(\mathcal{K}^{-1} \left ( s_{M-1} \right )\ldots \mathcal{K}^{-1}\left ( s_{1} \right ) U_M\bigg)\\ &\quad - F \mathcal{K}^{-1}\left(s_{M+1}\right) \left ( I_n - F \mathcal{K}^{-1}\left(s_{M+1}\right) \right )^{-1} \\ &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1} \\ & \qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg). \end{split} \end{aligned}$$ Taking $H_{\infty}-$norm on both sides, we get $$\begin{aligned} {2} &\left \| U_{M+1}\right \|_{H_{\infty}} = \\ & \quad \left \| \mathcal{K}\left ( s_1 \right )\ldots \mathcal{K}\left(s_M\right) \Bigg( N \mathcal{K}^{-1}\left(s_{M}\right) \mathcal{K}^{-1} \left ( s_{M-1} \right )\ldots \mathcal{K}^{-1}\left ( s_{1} \right ) U_M \right.\\ &\quad - F \mathcal{K}^{-1}\left(s_{M+1}\right) \left ( I_n - F \mathcal{K}^{-1}\left(s_{M+1}\right) \right )^{-1} \\ &\qquad N \mathcal{K}^{-1}\left(s_{M}\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_{M}\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(s_2\right) \left ( I_n- F \mathcal{K}^{-1}\left(s_2\right) \right )^{-1}\\ & \left. \qquad N \left ( I_n- \mathcal{K}^{-1}\left(s_1\right) F\right )^{-1}\Bigg)\right \|_{H_{\infty}}. % &= \underset{\omega_1, \ \ldots, \ \omega_{M+1} \in \mathbb{R}}{max} \left \| \mathcal{K}\left ( i\omega_1 \right )\ldots \mathcal{K}\left(i\omega_M\right) % \Bigg( N \mathcal{K}^{-1}\left(i\omega_M\right) \mathcal{K}^{-1} \left ( i\omega_{M-1} \right ) \right.\\ &\qquad \ldots \mathcal{K}^{-1}\left ( i\omega_1 \right ) U_M- F \mathcal{K}^{-1}\left(i\omega_{M+1}\right) \left ( I_n - F \mathcal{K}^{-1}\left(i\omega_{M+1}\right) \right )^{-1} \\ % &\qquad N \mathcal{K}^{-1}\left(i\omega_M\right) \left ( I_n- F \mathcal{K}^{-1}\left(i\omega_M\right) \right )^{-1} \ldots N \mathcal{K}^{-1}\left(i\omega_2\right) \\ % & \left. \qquad \left ( I_n- F \mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} N \left ( I_n- \mathcal{K}^{-1}\left(i\omega_1\right) F\right )^{-1}\Bigg)\right \|_2. \end{aligned}$$ As earlier, using the norm inequality properties in the above equation, we get $$\begin{aligned} {2} &\left \| U_{M+1} \right \|_{H_{\infty}} \leq \\ &\qquad \underset{\omega_1, \ \ldots, \ \omega_{M+1} \in \mathbb{R}}{max} \Bigg[ \left \| \mathcal{K}\left ( i\omega_1 \right ) \right \|_2 \ldots \left \| \mathcal{K}\left(i\omega_M\right) \right \|_2 \bigg( \left \| N \right \|_2 \left \| \mathcal{K}^{-1}\left(i\omega_M\right) \right \|_2 \ldots \\ &\qquad \left \| \mathcal{K}^{-1}\left ( i\omega_1 \right ) \right \|_2 \left \| U \left(i\omega_1, \ \ldots, \ i\omega_M\right)\right \|_2 + \left \| F \right \|_2 \left \| \mathcal{K}^{-1}\left(i\omega_{M+1}\right) \right \|_2 \\ &\qquad \left \| \left ( I_n - F \mathcal{K}^{-1}\left(i\omega_{M+1}\right) \right )^{-1} \right \|_2 \left \| N \right \|_2 \left \| \mathcal{K}^{-1}\left(i\omega_M\right) \right \|_2 \left \| \left ( I_n- F \mathcal{K}^{-1}\left(i\omega_M\right) \right )^{-1} \right \|_2 \\ &\qquad \ldots \left \| N \right \|_2 \left \| \mathcal{K}^{-1}\left(i\omega_2\right) \right \|_2 \left \| \left ( I_n- F \mathcal{K}^{-1}\left(i\omega_2\right) \right )^{-1} \right \|_2 \\ & \qquad \left \| N \right \|_2 \left \| \left ( I_n- \mathcal{K}^{-1}\left(i\omega_1\right) F\right )^{-1}\right \|_2 \bigg) \Bigg]. \end{aligned}$$ Similar to and , here also, using Lemma 2.3.3 from [@g_golub_book] we get $$\begin{aligned} \left \| U_{M+1} \right \|_{H_{\infty}} \leq & \left \| \mathcal{K}\left(s_1\right)\right \|_{H_{\infty}} \ldots \left \| \mathcal{K}\left(s_{M}\right)\right \|_{H_{\infty}} \left \| N \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_{M}\right)\right \|_{H_{\infty}} \ldots \\ & \left \| \mathcal{K}^{-1} \left(s_2\right)\right \|_{H_{\infty}} \Bigg[ \left \|\mathcal{K}^{-1} \left(s_1\right)\right \|_{H_{\infty}} \left \|U_M\right \|_{H_{\infty}} + \\ &\dfrac{ \left \|N\right \|_{2}^{M-1} \left \| \mathcal{K}^{-1}\left(s_{M+1}\right)\right \|_{H_{\infty}} }{\left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_{M+1}\right) \right \|_{H_{\infty}}\right) \ldots \left(1-\left \| F \right \|_{2} \left \| \mathcal{K}^{-1}\left(s_2\right) \right \|_{H_{\infty}}\right)} \cdot \\ &\dfrac{\left \| F \right \|_{2}}{1-\left \| \mathcal{K}^{-1} \left(s_1\right) \right \|_{H_{\infty}} \left \| F \right \|_{2}} \Bigg]. \end{aligned}$$ From induction hypothesis we know $\left \| U_{M} \right \|_{H_{\infty}} \propto \mathcal{O} \left ( \left \| F \right \|_2 \right )$. Using this we get $$\begin{aligned} \left \| U_{M+1} \right \|_{H_{\infty}} \propto \mathcal{O} \left ( \left \| F \right \|_2 \right ). \end{aligned}$$ \[Theorem:second\_condition\_subsystem\] If $$\begin{aligned} H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) = & \ c\left ( s_{M}I_n-A \right )^{-1}N \left ( s_{M-1}I_n-A \right )^{-1} \ldots N\left ( s_{1}I_n-A \right )^{-1}b,\\ \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) = & \ c\left ( s_{M}I_n-\left ( A+F \right ) \right )^{-1} \\ & \ N \left ( s_{M-1}I_n- \left ( A+F \right ) \right )^{-1} \ldots N\left ( s_{1}I_n- \left ( A+F \right ) \right )^{-1}b,\end{aligned}$$ $\left\|F\right\|_2 < 1$, and $\left \|\mathcal{K}^{-1}\left(s\right) \right \|_{H_{\infty}} <1$, where $\mathcal{K}\left(s\right)= \left(sI_n - A\right)$, then $$\begin{aligned} \left \|H_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) - \widetilde{H}_{M}\left ( s_{1},\ \ldots,\ s_{M} \right ) \right \|_{H_{2}}^2 = \mathcal{O}\left ( \left \| F \right \|_2^2\right ) \qquad \textnormal{or} \end{aligned}$$ TBIRKA satisfies the second condition of backward stability with respect to inexact linear solves. Conclusions & Future Work {#sec:Conclusions Future Work} ========================= In this paper, we apply iterative linear solvers during model order reduction (MOR) of bilinear dynamical systems. Since such solvers are inexact, the stability of the underlying MOR algorithm, with respect to these approximation errors, is important. Here, we extend the earlier stability analysis done for BIRKA in [@BIRKAstabilitylaapaper], which is a standard algorithm for obtaining a reduced bilinear dynamical system, to TBIRKA, a cheaper variant of BIRKA. Proving that an algorithm is stable, typically requires satisfying two conditions. In TBIRKA, fulfilling the first condition for stability leads to constraints on the iterative linear solver, which are similar to those obtained during BIRKA’s stability analysis. The second condition for a stable TBIRKA is satisfied using an approach different than the one used in BIRKA, and is more intuitive. Our first future direction is to extend our analysis from SISO (Single Input Single Output) to MIMO (Multiple Input Multiple Output) systems. The stability analysis as done for BIRKA earlier and TBIRKA here, all give us sufficiency conditions for a stable underlying MOR algorithm. Hence, second, we plan to derive the necessary conditions for the same. In recent years, there have been a lot of efforts in performing data-driven MOR algorithm (specially using Leowner framework [@bilinear_model_reduction_Antoulas_Loewner_framework]). Similar linear systems and challenges associated with the scaling arise here as well. Our third future direction is to apply this stability analysis to such classes of algorithms as well. [99]{} H. Kaper, H. Engler. . SIAM, Philadelphia, PA, USA, 2013. J. D. Meiss. . SIAM, Philadelphia, PA, USA, 2017. J. T. Stuart. . , 28(3):315–342, 1986. P. D’Alessandro, A. Isidori, and A. Ruberti. . , 12(3):517–535, 1974. R. J. Wilson. . Johns Hopkins Series in Information Sciences and Systems, Johns Hopkins University Press, Baltimore, 1981. F. Carravetta. . , 53(1):235–261, 2015. A. C. Antoulas. . Advances in Design and Control, SIAM, Philadelphia, PA, USA, 2005. P. Benner, D. C. Sorensen, and V. Mehrmann. . Lecture Notes in Computational Science and Engineering, Springer, Berlin, Germany, 2005. E. J. Grimme. Krylov Projection Methods for Model Reduction. , University of Illinois at Urbana-Champaign, Urbana, IL, USA, 1997. Z. Bai. Krylov subspace techniques for reduced-order modeling of large-scale dynamical systems. , 43(1):9–44, 2002. K. Willcox and J. Peraire. Balanced model reduction via the proper orthogonal decomposition. , 40(11):2323–2330, 2002. S. Gugercin, A. C. Antoulas, and C. Beattie. $\mathcal{H}_2$ model reduction for large-scale linear dynamical systems. , 30(2):609–638, 2008. A. Bunse-Gerstner, D. Kubali$\acute{\text{n}}$ska, G. Vossen, and D. Wilczek. $h_2$-norm optimal model reduction for large scale discrete dynamical MIMO systems. , 233(5):1202–1216, 2010. Z. Bai and D. Skoogh. . , 415(2-3):406–425, 2006. P. Benner and T. Breiten. Interpolation-based $\mathcal{H}_2$-model reduction of bilinear control systems. , 33(3):859–885, 2012. T. Bonin, H. Fa[ß]{}bender, A. Soppa, and M. Zaeh. A fully adaptive rational global Arnoldi method for the model-order reduction of second-order MIMO systems with proportional damping. , 122(C):1–19, 2016. M. I. Ahmad, U. Baur, and P. Benner. . , 316:15–28, 2017. P. Benner and T. Damm. . , 49(2):686–711, 2011. A. C. Antoulas, I. V. Gosea, and A. C. Ionita. . , 38(5):B889–B916, 2016. K. L. Xu, Y. L. Jiang, and Z. X. Yang. . , 90(3):616–626, 2017. G. M. Flagg. Interpolation Methods for the Model Reduction of Bilinear Systems. , Virginia Polytechnic Institute and State University, Blacksburg, VA, USA, 2012. G. M. Flagg and S. Gugercin. Multipoint Volterra series interpolation and $\mathcal{H}_2$ optimal model reduction of bilinear systems. , 36(2):549–579, 2015. A. Greenbaum. . SIAM, Philadelphia, PA, USA, 1997. Y. Saad. . SIAM, Philadelphia, PA, USA, 2003. L. N. Trefethen and D. Bau. . SIAM, Philadelphia, PA, USA, 1997. J. W. Demmel. . SIAM, Philadelphia, PA, USA, 1997. C. Beattie, S. Gugercin, and S. Wyatt. Inexact solves in interpolatory model reduction. , 436(8):2916–2943, 2012. N. P. Singh and K. Ahuja. Stability analysis of inexact solves in moment matching based model reduction. , 2018. D. Lu, Y. Su, and Z. Bai. . , 37(1):195–214, 2016. R. Choudhary and K. Ahuja. . , 538:56–88, 2018. K. Ahuja. Recycling Krylov subspaces and preconditioners. , Virginia Polytechnic Institute and State University, Blacksburg, VA, USA, 2011. K. Ahuja, E. D. Sturler, S. Gugercin, and E. R. Chang. Recycling BiCG with an application to model reduction. , 34(4):A1925–A1949, 2012. M. L. Parks, E. D. Sturler, G. Mackey, D. D. Johnson, and S. Maiti. Recycling [K]{}rylov subspaces for sequences of linear systems. , 28(5):1651-1674, 2006. A. Jeffrey. . Academic Press, 2003. C. D. Meyer. . SIAM, Philadelphia, PA, USA, 2000. G. H. Golub and C. F. Van Loan. . Johns Hopkins University Press, 2012. Rajendra Choudhary,\ Computational Science & Engineering Lab,\ Indian Institute of Technology Indore,\ Khandwa Road, Simrol, Indore-453552, India.\ Email: rajendracse46@gmail.com, phd1301201004@iiti.ac.in Kapil Ahuja,\ Computational Science & Engineering Lab,\ Indian Institute of Technology Indore,\ Khandwa Road, Simrol, Indore-453552, India.\ Email: kapsahuja22@gmail.com, kahuja@iiti.ac.in [^1]: *First-order* implies that the highest derivative of the state variable in the dynamical system is one. *Second-order* and *higher-orders* are similarly defined. [^2]: Obtaining such an invertible matrix is not difficult [@BIRKApaper; @TBIRKApaper]. [^3]: Recall, in BIRKA we work with $\zeta$ rather than $\zeta^M$. [^4]: \[footnote1\]This inequality says if $f\left(x\right)$ and $g\left(x\right)$ are integrable over $\left[a,b\right]$ and $f\left(x\right) \leq g\left(x\right)$, then $\int_{a}^{b} f \left ( x \right ) dx \leq \int_{a}^{b} g \left ( x \right ) dx$. Note that although we have improper integrals here, this inequality still holds because of the earlier assumption that such integrals give a finite value. [^5]: As mentioned in Footnote \[footnote1\], the improper integral does not affect application of this mean value theorem because all such integrals are assumed to give a finite value. [^6]: From [@Matrix_Analysis_carl_meyer_book page 527], we know $\left ( I - A \right )^{-1} = \sum\limits_{k=0}^{\infty} A^k$ when $ \left \| A \right \|<1$ for any matrix norm. Here, for the first inequality we have $\left \| F \mathcal{K}^{-1} \left ( s_2 \right ) \right \|_{H_{\infty}} <1$ or $\underset{\omega_2 \in \mathbb{R}}{max} \left \| F \mathcal{K}^{-1} \left ( i\omega_2 \right ) \right \|_2 <1$, and hence, the applicable matrix norm is $2-$norm. Similarly for the second inequality. [^7]: If $F \in \mathbb{R}^{n \times n}$ and $\left \| F \right \|_p <1$, then $I-F$ is nonsingular and $\left ( I-F \right )^{-1} =\sum\limits_{k=0}^{\infty} F^k $ with $ \left \| \left ( I-F \right )^{-1} \right \|_p \leq \dfrac{1}{1- \left \| F \right \|_p}$.

--- abstract: | In this paper, we will study the boundedness properties of multilinear Calderón–Zygmund operators and multilinear fractional integrals on products of weighted Morrey spaces with multiple weights.\ MSC(2010): 42B20; 42B35\ Keywords: Multilinear Calderón–Zygmund operators; multilinear fractional integrals; weighted Morrey spaces; multiple weights author: - | Hua Wang\ Department of Mathematics,\ Zhejiang University, Hangzhou 310027, P. R. China\ E-mail address: wanghua@pku.edu.cn.\ - | Wentan Yi\ Department of Applied Mathematics,\ Zhengzhou Information Science and Technology Institute,\ Zhengzhou 450002, P. R. China\ E-mail address: nlwt89@sina.com. title: Multilinear singular and fractional integral operators on weighted Morrey spaces --- Introduction and main results ============================= Multinear Calderón–Zygmund theory is a natural generalization of the linear case. The initial work on the class of multilinear Calderón–Zygmund operators was done by Coifman and Meyer in [@coifman], and was later systematically studied by Grafakos and Torres in [@grafakos2; @grafakos3; @grafakos4]. Let $\mathbb R^n$ be the $n$-dimensional Euclidean space and $(\mathbb R^n)^m=\mathbb R^n\times\cdots\times\mathbb R^n$ be the $m$-fold product space ($m\in\mathbb N$). We denote by $\mathscr S(\mathbb R^n)$ the space of all Schwartz functions on $\mathbb R^n$ and by $\mathscr S'(\mathbb R^n)$ its dual space, the set of all tempered distributions on $\mathbb R^n$. Let $m\ge2$ and $T$ be an $m$-linear operator initially defined on the $m$-fold product of Schwartz spaces and taking values into the space of tempered distributions, $$T:\mathscr S(\mathbb R^n)\times\cdots\times\mathscr S(\mathbb R^n)\to\mathscr S'(\mathbb R^n).$$ Following [@grafakos2], for given $\vec{f}=(f_1,\ldots,f_m)$, we say that $T$ is an $m$-linear Calderón–Zygmund operator if for some $q_1,\ldots,q_m\in[1,\infty)$ and $q\in(0,\infty)$ with $1/q=\sum_{k=1}^m 1/{q_k}$, it extends to a bounded multilinear operator from $L^{q_1}(\mathbb R^n)\times\cdots\times L^{q_m}(\mathbb R^n)$ into $L^q(\mathbb R^n)$, and if there exists a kernel function $K(x,y_1,\ldots,y_m)$ in the class $m$-$CZK(A,\varepsilon)$, defined away from the diagonal $x=y_1=\cdots=y_m$ in $(\mathbb R^n)^{m+1}$ such that $$T(\vec{f})(x)=T(f_1,\ldots,f_m)(x)=\int_{(\mathbb R^n)^m}K(x,y_1,\ldots,y_m) f_1(y_1)\cdots f_m(y_m)\,dy_1\cdots dy_m,$$ whenever $f_1,\ldots,f_m\in \mathscr S(\mathbb R^n)$ and $x\notin \cap_{k=1}^m$ supp$f_k$. We say that $K(x,y_1,\ldots,y_m)$ is a kernel in the class $m$-$CZK(A,\varepsilon)$, if it satisfies the size condition $$\big|K(x,y_1,\ldots,y_m)\big|\le\frac{A}{(|x-y_1|+\cdots+|x-y_m|)^{mn}},$$ for some $A>0$ and all $(x,y_1,\ldots,y_m)\in(\mathbb R^n)^{m+1}$ with $x\neq y_k$ for some $1\le k\le m$. Moreover, for some $\varepsilon>0$, it satisfies the regularity condition that $$\big|K(x,y_1,\ldots,y_m)-K(x',y_1,\ldots,y_m)\big|\le \frac{A\cdot|x-x'|^\varepsilon}{(|x-y_1|+\cdots+|x-y_m|)^{mn+\varepsilon}}$$ whenever $|x-x'|\le\frac12 \max_{1\le k\le m}|x-y_k|$, and also that for each fixed $k$ with $1\le k\le m$, $$\big|K(x,y_1,\ldots,y_k,\ldots,y_m)-K(x,y_1,\ldots,y'_k,\ldots,y_m)\big|\le \frac{A\cdot|y_k-y'_k|^\varepsilon}{(|x-y_1|+\cdots+|x-y_m|)^{mn+\varepsilon}}$$ whenever $|y_k-y'_k|\le\frac12 \max_{1\le i\le m}|x-y_i|$. In recent years, many authors have been interested in studying the boundedness of these operators on function spaces, see e.g.[@grafakos5; @hu; @li1; @li2]. In 2009, the weighted strong and weak type estimates of multilinear Calderón–Zygmund singular integral operators were established in [@lerner] by Lerner et al. New more refined multilinear maximal function was defined and used in [@lerner] to characterize the class of multiple $A_{\vec{P}}$ weights. Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in(1,\infty)$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)$ satisfy the $A_{\vec{P}}$ condition, then there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{L^p(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)$ satisfy the $A_{\vec{P}}$ condition, then there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{WL^p(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Let $m\ge2$ and $0<\alpha<mn$. For given $\vec{f}=(f_1,\ldots,f_m)$, the $m$-linear fractional integral operator is defined by $$I_{\alpha}(\vec{f})(x)=I_{\alpha}(f_1,\ldots,f_m)(x)= \int_{(\mathbb R^n)^m}\frac{f_1(y_1)\cdots f_m(y_m)}{|(x-y_1,\ldots,x-y_m)|^{mn-\alpha}}\,dy_1\cdots dy_m.$$ For the boundedness properties of multilinear fractional integrals on various function spaces, we refer the reader to [@grafakos1; @Iida1; @Iida2; @Iida3; @kenig; @pradolini; @tang]. In 2009, Moen [@moen] considered the weighted norm inequalities for multilinear fractional integral operators and constructed the class of multiple $A_{\vec{P},q}$ weights (see also [@chen]). Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in(1,\infty)$, $1/p=\sum_{k=1}^m 1/{p_k}$ and $1/q=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)$ satisfy the $A_{\vec{P},q}$ condition, then there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{L^q((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i}(w^{p_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$, $1/p=\sum_{k=1}^m 1/{p_k}$ and $1/q=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)$ satisfy the $A_{\vec{P},q}$ condition, then there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{WL^q((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i}(w^{p_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. On the other hand, the classical Morrey spaces $\mathcal L^{p,\lambda}$ were originally introduced by Morrey in [@morrey] to study the local behavior of solutions to second order elliptic partial differential equations. For the boundedness of the Hardy–Littlewood maximal operator, the fractional integral operator and the Calderón–Zygmund singular integral operator on these spaces, we refer the reader to [@adams; @chiarenza; @peetre]. For the properties and applications of classical Morrey spaces, one can see [@fan; @fazio1; @fazio2] and the references therein. In 2009, Komori and Shirai [@komori] first defined the weighted Morrey spaces $L^{p,\kappa}(w)$ which could be viewed as an extension of weighted Lebesgue spaces, and studied the boundedness of the above classical operators in Harmonic Analysis on these weighted spaces. Recently, in [@wang1; @wang2; @wang3; @wang4; @wang5; @wang6; @wang7; @wang8], we have established the continuity properties of some other operators and their commutators on the weighted Morrey spaces $L^{p,\kappa}(w)$. The main purpose of this paper is to establish the boundedness properties of multilinear Calderón–Zygmund operators and multilinear fractional integrals on products of weighted Morrey spaces with multiple weights. We now formulate our main results as follows. \[section\] Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in(1,\infty)$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P}}$ with $w_1,\ldots,w_m\in A_\infty$, then for any $0<\kappa<1$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{L^{p,\kappa}(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P}}$ with $w_1,\ldots,w_m\in A_\infty$, then for any $0<\kappa<1$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{WL^{p,\kappa}(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in(1,\infty)$, $1/p=\sum_{k=1}^m 1/{p_k}$, $1/{q_k}=1/{p_k}-\alpha/{mn}$ and $1/q=\sum_{k=1}^m 1/{q_k}=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P},q}$ with $w^{q_1}_1,\ldots,w^{q_m}_m\in A_\infty$, then for any $0<\kappa<p/q$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{L^{q,{\kappa q}/p}((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$, $1/p=\sum_{k=1}^m 1/{p_k}$, $1/{q_k}=1/{p_k}-\alpha/{mn}$ and $1/q=\sum_{k=1}^m 1/{q_k}=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P},q}$ with $w^{q_1}_1,\ldots,w^{q_m}_m\in A_\infty$, then for any $0<\kappa<p/q$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{WL^{q,{\kappa q}/p}((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. Notations and definitions ========================= The classical $A_p$ weight theory was first introduced by Muckenhoupt in the study of weighted $L^p$ boundedness of Hardy–Littlewood maximal functions in [@muckenhoupt1]. A weight $w$ is a nonnegative, locally integrable function on $\mathbb R^n$, $B=B(x_0,r_B)$ denotes the ball with the center $x_0$ and radius $r_B$. For $1<p<\infty$, a weight function $w$ is said to belong to $A_p$, if there is a constant $C>0$ such that for every ball $B\subseteq \mathbb R^n$, $$\left(\frac1{|B|}\int_B w(x)\,dx\right)\left(\frac1{|B|}\int_B w(x)^{-1/{(p-1)}}\,dx\right)^{p-1}\le C,$$ where $|B|$ denotes the Lebesgue measure of $B$. For the case $p=1$, $w\in A_1$, if there is a constant $C>0$ such that for every ball $B\subseteq \mathbb R^n$, $$\frac1{|B|}\int_B w(x)\,dx\le C\cdot\underset{x\in B}{\mbox{ess\,inf}}\;w(x).$$ A weight function $w\in A_\infty$ if it satisfies the $A_p$ condition for some $1<p<\infty$. We also need another weight class $A_{p,q}$ introduced by Muckenhoupt and Wheeden in [@muckenhoupt2]. A weight function $w$ belongs to $A_{p,q}$ for $1<p<q<\infty$ if there is a constant $C>0$ such that for every ball $B\subseteq \mathbb R^n$, $$\left(\frac{1}{|B|}\int_B w(x)^q\,dx\right)^{1/q}\left(\frac{1}{|B|}\int_B w(x)^{-p'}\,dx\right)^{1/{p'}}\le C.$$ When $p=1$, $w$ is in the class $A_{1,q}$ with $1<q<\infty$ if there is a constant $C>0$ such that for every ball $B\subseteq \mathbb R^n$, $$\left(\frac{1}{|B|}\int_B w(x)^q\,dx\right)^{1/q}\bigg(\underset{x\in B}{\mbox{ess\,sup}}\,\frac{1}{w(x)}\bigg)\le C.$$ Now let us recall the definitions of multiple weights. For $m$ exponents $p_1,\ldots,p_m$, we will write $\vec{P}$ for the vector $\vec{P}=(p_1,\ldots,p_m)$. Let $p_1,\ldots,p_m\in[1,\infty)$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$. Given $\vec{w}=(w_1,\ldots,w_m)$, set $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. We say that $\vec{w}$ satisfies the $A_{\vec{P}}$ condition if it satisfies $$\sup_B\left(\frac{1}{|B|}\int_B \nu_{\vec{w}}(x)\,dx\right)^{1/p}\prod_{i=1}^m\left(\frac{1}{|B|}\int_B w_i(x)^{1-p'_i}\,dx\right)^{1/{p'_i}}<\infty.$$ When $p_i=1$, $\big(\frac{1}{|B|}\int_B w_i(x)^{1-p'_i}\,dx\big)^{1/{p'_i}}$ is understood as $\big(\inf_{x\in B}w_i(x)\big)^{-1}$. Let $p_1,\ldots,p_m\in[1,\infty)$, $1/p=\sum_{k=1}^m 1/{p_k}$ and $q>0$. Given $\vec{w}=(w_1,\ldots,w_m)$, set $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. We say that $\vec{w}$ satisfies the $A_{\vec{P},q}$ condition if it satisfies $$\sup_B\left(\frac{1}{|B|}\int_B \nu_{\vec{w}}(x)^q\,dx\right)^{1/q}\prod_{i=1}^m\left(\frac{1}{|B|}\int_B w_i(x)^{-p'_i}\,dx\right)^{1/{p'_i}}<\infty.$$ When $p_i=1$, $\big(\frac{1}{|B|}\int_B w_i(x)^{-p'_i}\,dx\big)^{1/{p'_i}}$ is understood as $\big(\inf_{x\in B}w_i(x)\big)^{-1}$. Given a ball $B$ and $\lambda>0$, $\lambda B$ denotes the ball with the same center as $B$ whose radius is $\lambda$ times that of $B$. For a given weight function $w$ and a measurable set $E$, we also denote the Lebesgue measure of $E$ by $|E|$ and the weighted measure of $E$ by $w(E)$, where $w(E)=\int_E w(x)\,dx$. \[theorem\][Lemma]{} Let $w\in A_p$ with $1\le p<\infty$. Then, for any ball $B$, there exists an absolute constant $C>0$ such that $$w(2B)\le C\,w(B).$$ Let $w\in A_\infty$. Then for all balls $B\subseteq \mathbb R^n$, the following reverse Jensen inequality holds. $$\int_B w(x)\,dx\le C|B|\cdot\exp\left(\frac{1}{|B|}\int_B\log w(x)\,dx\right).$$ Let $w\in A_\infty$. Then for all balls $B$ and all measurable subsets $E$ of $B$, there exists $\delta>0$ such that $$\frac{w(E)}{w(B)}\le C\left(\frac{|E|}{|B|}\right)^\delta.$$ Let $p_1,\ldots,p_m\in[1,\infty)$ and $1/p=\sum_{k=1}^m 1/{p_k}$. Then $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P}}$ if and only if $$\left\{ \begin{aligned} &\nu_{\vec{w}}\in A_{mp},\\ &w_i^{1-p'_i}\in A_{mp'_i},\quad i=1,\ldots,m, \end{aligned}\right.$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$ and the condition $w_i^{1-p'_i}\in A_{mp'_i}$ in the case $p_i=1$ is understood as $w_i^{1/m}\in A_1$. Let $0<\alpha<mn$, $p_1,\ldots,p_m\in[1,\infty)$, $1/p=\sum_{k=1}^m 1/{p_k}$ and $1/q=1/p-\alpha/n$. Then $\vec{w}=(w_1,\ldots,w_m)\in A_{\vec{P},q}$ if and only if $$\left\{ \begin{aligned} &(\nu_{\vec{w}})^q\in A_{mq},\\ &w_i^{-p'_i}\in A_{mp'_i},\quad i=1,\ldots,m, \end{aligned}\right.$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. Given a weight function $w$ on $\mathbb R^n$, for $0<p<\infty$, the weighted Lebesgue space $L^p(w)$ defined as the set of all functions $f$ such that $$\big\|f\big\|_{L^p(w)}=\bigg(\int_{\mathbb R^n}|f(x)|^pw(x)\,dx\bigg)^{1/p}<\infty.$$ We also denote by $WL^p(w)$ the weighted weak space consisting of all measurable functions $f$ such that $$\big\|f\big\|_{WL^p(w)}=\sup_{\lambda>0}\lambda\cdot w\big(\big\{x\in\mathbb R^n:|f(x)|>\lambda \big\}\big)^{1/p}<\infty.$$ In 2009, Komori and Shirai [@komori] first defined the weighted Morrey spaces $L^{p,\kappa}(w)$ for $1\le p<\infty$. In order to deal with the multilinear case $m\ge2$, we shall define $L^{p,\kappa}(w)$ for all $0<p<\infty$. \[theorem\][Definition]{} Let $0<p<\infty$, $0<\kappa<1$ and $w$ be a weight function on $\mathbb R^n$. Then the weighted Morrey space is defined by $$L^{p,\kappa}(w)=\big\{f\in L^p_{loc}(w):\big\|f\big\|_{L^{p,\kappa}(w)}<\infty\big\},$$ where $$\big\|f\big\|_{L^{p,\kappa}(w)}=\sup_B\left(\frac{1}{w(B)^\kappa}\int_B|f(x)|^pw(x)\,dx\right)^{1/p}$$ and the supremum is taken over all balls $B$ in $\mathbb R^n$. Let $0<p<\infty$, $0<\kappa<1$ and $w$ be a weight function on $\mathbb R^n$. Then the weighted weak Morrey space is defined by $$WL^{p,\kappa}(w)=\big\{f \;\mbox{measurable}:\big\|f\big\|_{WL^{p,\kappa}(w)}<\infty\big\},$$ where $$\big\|f\big\|_{WL^{p,\kappa}(w)}=\sup_B\sup_{\lambda>0}\frac{1}{w(B)^{\kappa/p}}\lambda\cdot w\big(\big\{x\in B:|f(x)|>\lambda \big\}\big)^{1/p}.$$ Furthermore, in order to deal with the fractional order case, we need to consider the weighted Morrey spaces with two weights. Let $0<p<\infty$ and $0<\kappa<1$. Then for two weights $u$ and $v$, the weighted Morrey space is defined by $$L^{p,\kappa}(u,v)=\big\{f\in L^p_{loc}(u):\big\|f\big\|_{L^{p,\kappa}(u,v)}<\infty\big\},$$ where $$\big\|f\big\|_{L^{p,\kappa}(u,v)}=\sup_{B}\left(\frac{1}{v(B)^{\kappa}}\int_B|f(x)|^pu(x)\,dx\right)^{1/p}.$$ Throughout this article, we will use $C$ to denote a positive constant, which is independent of the main parameters and not necessarily the same at each occurrence. Moreover, we will denote the conjugate exponent of $p>1$ by $p'=p/{(p-1)}$. Proofs of Theorems 1.1 and 1.2 ============================== Before proving the main theorems of this section, we need to establish the following lemma. Let $m\ge2$, $p_1,\ldots,p_m\in[1,\infty)$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$. Assume that $w_1,\ldots,w_m\in A_\infty$ and $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$, then for any ball $B$, there exists a constant $C>0$ such that $$\prod_{i=1}^m\left(\int_B w_i(x)\,dx\right)^{p/{p_i}}\le C\int_B \nu_{\vec{w}}(x)\,dx.$$ Since $w_1,\ldots,w_m\in A_\infty$, then by using Lemma 2.2, we have $$\begin{split} \prod_{i=1}^m \left(\int_B w_i(x)\,dx\right)^{p/{p_i}}&\le C\prod_{i=1}^m \left(|B|\cdot\exp\bigg(\frac{1}{|B|}\int_B \log w_i(x)\,dx\bigg)\right)^{p/{p_i}}\\ &= C\prod_{i=1}^m \left(|B|^{p/{p_i}}\cdot\exp\bigg(\frac{1}{|B|}\int_B\log w_i(x)^{p/{p_i}}\,dx\bigg)\right)\\ &= C\cdot\big(|B|\big)^{\sum_{i=1}^m p/{p_i}}\cdot\exp\left(\sum_{i=1}^m\frac{1}{|B|}\int_B\log w_i(x)^{p/{p_i}}\,dx\right). \end{split}$$ Note that $\sum_{i=1}^m p/{p_i}=1$ and $\nu_{\vec{w}}(x)=\prod_{i=1}^m w_i(x)^{p/{p_i}}$. Then by Jensen inequality, we obtain $$\begin{split} \prod_{i=1}^m \left(\int_B w_i(x)\,dx\right)^{p/{p_i}}&\le C\cdot|B|\cdot\exp\left(\frac{1}{|B|}\int_B\log \nu_{\vec{w}}(x)\,dx\right)\\ &\le C\int_B \nu_{\vec{w}}(x)\,dx. \end{split}$$ We are done. For any ball $B=B(x_0,r_B)\subseteq\mathbb R^n$ and let $f_i=f^0_i+f^{\infty}_i$, where $f^0_i=f_i\chi_{2B}$, $i=1,\ldots,m$ and $\chi_{2B}$ denotes the characteristic function of $2B$. Then we write $$\begin{split} \prod_{i=1}^m f_i(y_i)&=\prod_{i=1}^m\Big(f^0_i(y_i)+f^{\infty}_i(y_i)\Big)\\ &=\sum_{\alpha_1,\ldots,\alpha_m\in\{0,\infty\}}f^{\alpha_1}_1(y_1)\cdots f^{\alpha_m}_m(y_m)\\ &=\prod_{i=1}^m f^0_i(y_i)+\sum\nolimits^\prime f^{\alpha_1}_1(y_1)\cdots f^{\alpha_m}_m(y_m), \end{split}$$ where each term of $\sum^\prime$ contains at least one $\alpha_i\neq0$. Since $T$ is an $m$-linear operator, then we have $$\begin{split} &\frac{1}{\nu_{\vec{w}}(B)^{\kappa/p}}\left(\int_B \big|T(f_1,\ldots,f_m)(x)\big|^p\nu_{\vec{w}}(x)\,dx\right)^{1/p}\\ \le& \frac{1}{\nu_{\vec{w}}(B)^{\kappa/p}}\left(\int_B \big|T(f^0_1,\ldots,f^0_m)(x)\big|^p\nu_{\vec{w}}(x)\,dx\right)^{1/p}\\ &+\sum\nolimits^\prime\frac{1}{\nu_{\vec{w}}(B)^{\kappa/p}}\left(\int_B \big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|^p\nu_{\vec{w}}(x)\,dx\right)^{1/p}\\ =& I^0+\sum\nolimits^\prime I^{\alpha_1,\ldots,\alpha_m}. \end{split}$$ In view of Lemma 2.4, we have that $\nu_{\vec{w}}\in A_{mp}$. Applying Theorem A, Lemma 3.1 and Lemma 2.1, we get $$\begin{split} I^0&\le C\cdot \frac{1}{\nu_{\vec{w}}(B)^{\kappa/p}}\prod_{i=1}^m\left(\int_{2B}|f_i(x)|^{p_i}w_i(x)\,dx\right)^{1/{p_i}}\\ &\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot \frac{\prod_{i=1}^m w_i(2B)^{\kappa/{p_i}}}{\nu_{\vec{w}}(B)^{\kappa/p}}\\ &\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot \frac{\nu_{\vec{w}}(2B)^{\kappa/p}}{\nu_{\vec{w}}(B)^{\kappa/p}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}. \end{split}$$ For the other terms, let us first consider the case when $\alpha_1=\cdots=\alpha_m=\infty$. By the size condition, for any $x\in B$, we obtain $$\begin{aligned} \big|T(f^\infty_1,\ldots,f^\infty_m)(x)\big|&\le C\int_{(\mathbb R^n)^m\backslash(2B)^m} \frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn}}dy_1\cdots dy_m\notag\\ &\le C\sum_{j=1}^\infty\int_{(2^{j+1}B)^m\backslash(2^{j}B)^m} \frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn}}dy_1\cdots dy_m\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\int_{2^{j+1}B\backslash 2^{j}B}\frac{|f_i(y_i)|}{|x-y_i|^n}dy_i\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i,\end{aligned}$$ where we have used the notation $E^m=E\times\cdots\times E$. Furthermore, by using Hölder’s inequality, the multiple $A_{\vec{P}}$ condition and Lemma 3.1, we deduce that $$\begin{split} \big|T(f^\infty_1,\ldots,f^\infty_m)(x)\big|&\le C\sum_{j=1}^\infty\prod_{i=1}^m \frac{1}{|2^{j+1}B|}\left(\int_{2^{j+1}B}\big|f_i(y_i)\big|^{p_i}w_i(y_i)\,dy_i\right)^{1/{p_i}} \left(\int_{2^{j+1}B}w_i(y_i)^{1-p'_i}\,dy_i\right)^{1/{p'_i}}\\ &\le C\sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^m}\cdot \frac{|2^{j+1}B|^{\frac 1p+\sum_{i=1}^m(1-\frac 1{p_i})}}{\nu_{\vec{w}}(2^{j+1}B)^{1/p}} \prod_{i=1}^m \left(\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}w_i\big(2^{j+1}B\big)^{\kappa/{p_i}}\right)\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\sum_{j=1}^\infty \left(\frac{\prod_{i=1}^m w_i(2^{j+1}B)^{\kappa/{p_i}}}{\nu_{\vec{w}}(2^{j+1}B)^{1/p}}\right)\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot \sum_{j=1}^\infty\nu_{\vec{w}}\big(2^{j+1}B\big)^{{(\kappa-1)}/p}. \end{split}$$ Since $\nu_{\vec{w}}\in A_{mp}\subset A_\infty$, then it follows directly from Lemma 2.3 that $$\frac{\nu_{\vec{w}}(B)}{\nu_{\vec{w}}(2^{j+1}B)}\le C\left(\frac{|B|}{|2^{j+1}B|}\right)^\delta.$$ Hence $$\begin{split} I^{\infty,\ldots,\infty}&\le\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}\big|T(f^\infty_1,\ldots,f^\infty_m)(x)\big|\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\sum_{j=1}^\infty \frac{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}}{\nu_{\vec{w}}(2^{j+1}B)^{{(1-\kappa)}/p}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\sum_{j=1}^\infty \left(\frac{|B|}{|2^{j+1}B|}\right)^{{\delta(1-\kappa)}/p}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}, \end{split}$$ where the last inequality holds since $0<\kappa<1$ and $\delta>0$. We now consider the case where exactly $\ell$ of the $\alpha_i$ are $\infty$ for some $1\le\ell<m$. We only give the arguments for one of these cases. The rest are similar and can easily be obtained from the arguments below by permuting the indices. Using the size condition again, we deduce that for any $x\in B$, $$\begin{aligned} \big|T(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|&\le C\int_{(\mathbb R^n)^{\ell}\backslash(2B)^{\ell}}\int_{(2B)^{m-\ell}}\frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn}}dy_1\cdots dy_m\notag\\ &\le C\prod_{i=\ell+1}^m\int_{2B}\big|f_i(y_i)\big|\,dy_i\notag\\ &\times \sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^m}\int_{(2^{j+1}B)^\ell\backslash(2^{j}B)^\ell} \big|f_1(y_1)\cdots f_{\ell}(y_\ell)\big|\,dy_1\cdots dy_\ell\notag\\ &\le C\prod_{i=\ell+1}^m\int_{2B}\big|f_i(y_i)\big|\,dy_i\times\sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^m}\prod_{i=1}^{\ell} \int_{2^{j+1}B\backslash 2^{j}B}\big|f_i(y_i)\big|\,dy_i\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i,\end{aligned}$$ and we arrived at the expression considered in the previous case. So for any $x\in B$, we also have $$\big|T(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot \sum_{j=1}^\infty\nu_{\vec{w}}\big(2^{j+1}B\big)^{{(\kappa-1)}/p}.$$ Therefore, by the inequality (3.2) and the above pointwise inequality, we have $$\begin{split} I^{\alpha_1,\ldots,\alpha_m}&\le\nu_{\vec{w}}(B)^{{(1-\kappa)}/p} \big|T(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\sum_{j=1}^\infty \frac{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}}{\nu_{\vec{w}}(2^{j+1}B)^{{(1-\kappa)}/p}}\\ \end{split}$$ $$\begin{split} &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\sum_{j=1}^\infty \left(\frac{|B|}{|2^{j+1}B|}\right)^{{\delta(1-\kappa)}/p}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}. \end{split}$$ Combining the above estimates and then taking the supremum over all balls $B\subseteq\mathbb R^n$, we complete the proof of Theorem 1.1. For any ball $B=B(x_0,r_B)\subseteq\mathbb R^n$ and decompose $f_i=f^0_i+f^{\infty}_i$, where $f^0_i=f_i\chi_{2B}$, $i=1,\ldots,m$. Then for any given $\lambda>0$, we can write $$\begin{split} &\nu_{\vec{w}}\big(\big\{x\in B:\big|T(f_1,\ldots,f_m)\big|>\lambda\big\}\big)^{1/p}\\ \le&\nu_{\vec{w}}\big(\big\{x\in B:\big|T(f^0_1,\ldots,f^0_m)\big|>\lambda/{2^m}\big\}\big)^{1/p} +\sum\nolimits^\prime \nu_{\vec{w}}\big(\big\{x\in B:\big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)\big|>\lambda/{2^m}\big\}\big)^{1/p}\\ =&I^0_*+\sum\nolimits^\prime I^{\alpha_1,\ldots,\alpha_m}_*, \end{split}$$ where each term of $\sum^\prime$ contains at least one $\alpha_i\neq0$. By Lemma 2.4 again, we know that $\nu_{\vec{w}}\in A_{mp}$ with $1\le mp<\infty$. Applying Theorem B, Lemma 3.1 and Lemma 2.1, we have $$\begin{split} I^0_*&\le\frac{C}{\lambda}\prod_{i=1}^m\left(\int_{2B}|f_i(x)|^{p_i}w_i(x)\,dx\right)^{1/{p_i}}\\ &\le \frac{C\cdot\prod_{i=1}^m w_i(2B)^{\kappa/{p_i}}}{\lambda}\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\\ &\le \frac{C\cdot\nu_{\vec{w}}(2B)^{\kappa/p}}{\lambda}\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\\ &\le \frac{C\cdot\nu_{\vec{w}}(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}. \end{split}$$ In the proof of Theorem 1.1, we have already showed the following pointwise estimate (see (3.1) and (3.3)). $$\big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i.$$ Without loss of generality, we may assume that $p_1=\cdots=p_{\ell}=\min\{p_1,\ldots,p_m\}=1$, and $p_{\ell+1},\ldots,p_m>1$. Using Hölder’s inequality, the multiple $A_{\vec{P}}$ condition and Lemma 3.1, we obtain $$\begin{split} \big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|&\le C\sum_{j=1}^\infty\prod_{i=1}^{\ell}\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i\times \prod_{i=\ell+1}^{m}\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^{\ell}\frac{1}{|2^{j+1}B|}\int_{2^{j+1}B}\big|f_i(y_i)\big|w_i(y_i)\,dy_i \left(\inf_{y_i\in 2^{j+1}B}w_i(y_i)\right)^{-1}\\ \end{split}$$ $$\begin{split} &\times\prod_{i=\ell+1}^{m}\frac{1}{|2^{j+1}B|} \left(\int_{2^{j+1}B}\big|f_i(y_i)\big|^{p_i}w_i(y_i)\,dy_i\right)^{1/{p_i}} \left(\int_{2^{j+1}B}w_i(y_i)^{1-p'_i}\,dy_i\right)^{1/{p'_i}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)} \sum_{j=1}^\infty\nu_{\vec{w}}\big(2^{j+1}B\big)^{{(\kappa-1)}/p}. \end{split}$$ Observe that $\nu_{\vec{w}}\in A_{mp}$ with $1\le mp<\infty$. Thus, it follows from the inequality (3.2) that for any $x\in B$, $$\begin{aligned} \big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big| &= C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\frac{1}{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}} \sum_{j=1}^\infty\frac{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}}{\nu_{\vec{w}}(2^{j+1}B)^{{(1-\kappa)}/p}}\notag\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\frac{1}{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}} \sum_{j=1}^\infty\left(\frac{|B|}{|2^{j+1}B|}\right)^{{\delta(1-\kappa)}/p}\notag\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\frac{1}{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}}.\end{aligned}$$ If $\big\{x\in B:\big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|>\lambda/{2^m}\big\}=\O$, then the inequality $$I^{\alpha_1,\ldots,\alpha_m}_*\le\frac{C\cdot\nu_{\vec{w}}(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}$$ holds trivially. Now if instead we suppose that $\big\{x\in B:\big|T(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|>\lambda/{2^m}\big\}\neq\O$, then by the pointwise inequality (3.5), we have $$\lambda< C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}\cdot\frac{1}{\nu_{\vec{w}}(B)^{{(1-\kappa)}/p}},$$ which is equivalent to $$\nu_{\vec{w}}(B)^{1/p}\le\frac{C\cdot\nu_{\vec{w}}(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}.$$ Therefore $$I^{\alpha_1,\ldots,\alpha_m}_*\le \nu_{\vec{w}}(B)^{1/p}\le\frac{C\cdot\nu_{\vec{w}}(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,\kappa}(w_i)}.$$ Summing up all the above estimates and then taking the supremum over all balls $B\subseteq\mathbb R^n$ and all $\lambda>0$, we complete the proof of Theorem 1.2. By using Hölder’s inequality, it is easy to check that if each $w_i$ is in $A_{p_i}$, then $$\prod_{i=1}^m A_{p_i}\subset A_{\vec{P}}.$$ and this inclusion is strict (see [@lerner]). Thus, as direct consequences of Theorems 1.1 and 1.2, we immediately obtain the following \[theorem\][Corollary]{} Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in(1,\infty)$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)\in \prod_{i=1}^m A_{p_i}$, then for any $0<\kappa<1$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{L^{p,\kappa}(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Let $m\ge2$ and $T$ be an $m$-linear Calderón–Zygmund operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$ and $p\in(0,\infty)$ with $1/p=\sum_{k=1}^m 1/{p_k}$, and $\vec{w}=(w_1,\ldots,w_m)\in \prod_{i=1}^m A_{p_i}$, then for any $0<\kappa<1$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|T(\vec{f})\big\|_{WL^{p,\kappa}(\nu_{\vec{w}})}\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,\kappa}(w_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i^{p/{p_i}}$. Proofs of Theorems 1.3 and 1.4 ============================== Following along the same lines as that of Lemma 3.1, we can also show the following result, which plays an important role in our proofs of Theorems 1.3 and 1.4. Let $m\ge2$, $q_1,\ldots,q_m\in[1,\infty)$ and $q\in(0,\infty)$ with $1/q=\sum_{k=1}^m 1/{q_k}$. Assume that $w^{q_1}_1,\ldots,w^{q_m}_m\in A_\infty$ and $\nu_{\vec{w}}=\prod_{i=1}^m w_i$, then for any ball $B$, there exists a constant $C>0$ such that $$\prod_{i=1}^m\left(\int_B w^{q_i}_i(x)\,dx\right)^{q/{q_i}}\le C\int_B\nu_{\vec{w}}(x)^q\,dx.$$ Arguing as in the proof of Theorem 1.1, fix a ball $B=B(x_0,r_B)\subseteq\mathbb R^n$ and decompose $f_i=f^0_i+f^{\infty}_i$, where $f^0_i=f_i\chi_{2B}$, $i=1,\ldots,m$. Since $I_\alpha$ is an $m$-linear operator, then we have $$\begin{split} &\frac{1}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\left(\int_B \big|I_\alpha(f_1,\ldots,f_m)(x)\big|^q\nu_{\vec{w}}(x)^q\,dx\right)^{1/q}\\ \le& \frac{1}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\left(\int_B \big|I_\alpha (f^0_1,\ldots,f^0_m)(x)\big|^q\nu_{\vec{w}}(x)^q\,dx\right)^{1/q}\\ &+\sum\nolimits^\prime\frac{1}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\left(\int_B \big|I_\alpha(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|^q\nu_{\vec{w}}(x)^q\,dx\right)^{1/q}\\ =& J^0+\sum\nolimits^\prime J^{\alpha_1,\ldots,\alpha_m}, \end{split}$$ where each term of $\sum^\prime$ contains at least one $\alpha_i\neq0$. In view of Lemma 2.5, we can see that $(\nu_{\vec{w}})^q\in A_{mq}$. Using Theorem C, Lemma 4.1 and Lemma 2.1, we get $$\begin{split} J^0&\le C\cdot \frac{1}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\prod_{i=1}^m\left(\int_{2B}|f_i(x)|^{p_i}w_i(x)^{p_i}\,dx\right)^{1/{p_i}}\\ &\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{\prod_{i=1}^m w^{q_i}_i(2B)^{{\kappa q}/{pq_i}}}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\\ &= C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{\Big(\prod_{i=1}^m w^{q_i}_i(2B)^{q/{q_i}}\Big)^{\kappa/p}}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\\ &\le C\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{\nu_{\vec{w}}^q(2B)^{\kappa/p}}{\nu_{\vec{w}}^q(B)^{\kappa/p}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}. \end{split}$$ For the other terms, let us first deal with the case when $\alpha_1=\cdots=\alpha_m=\infty$. By the definition of $I_\alpha$, for any $x\in B$, we obtain $$\begin{aligned} \big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_m)(x)\big|&= \int_{(\mathbb R^n)^m\backslash(2B)^m} \frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn-\alpha}}dy_1\cdots dy_m\notag\\ &= \sum_{j=1}^\infty\int_{(2^{j+1}B)^m\backslash(2^{j}B)^m} \frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn-\alpha}}dy_1\cdots dy_m\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\int_{2^{j+1}B\backslash 2^{j}B}\frac{|f_i(y_i)|}{|x-y_i|^{n-\alpha/m}}dy_i\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i.\end{aligned}$$ Moreover, by using Hölder’s inequality, the multiple $A_{\vec{P},q}$ condition and Lemma 4.1, we deduce that $$\begin{split} \big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_m)(x)\big|&\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}} \left(\int_{2^{j+1}B}\big|f_i(y_i)\big|^{p_i}w_i(y_i)^{p_i}\,dy_i\right)^{1/{p_i}}\\ &\times\left(\int_{2^{j+1}B}w_i(y_i)^{-p'_i}\,dy_i\right)^{1/{p'_i}}\\ &\le C\sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^{m-\alpha/{n}}}\cdot \frac{|2^{j+1}B|^{\frac 1q+\sum_{i=1}^m(1-\frac 1{p_i})}}{\nu_{\vec{w}}^q(2^{j+1}B)^{1/q}}\\ \end{split}$$ $$\begin{split} &\times\prod_{i=1}^m \left(\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}w^{q_i}_i\big(2^{j+1}B\big)^{{\kappa q}/{pq_i}}\right)\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \sum_{j=1}^\infty\Bigg[\frac{\Big(\prod_{i=1}^m w^{q_i}_i(2^{j+1}B)^{q/{q_i}}\Big)^{\kappa/p}}{\nu_{\vec{w}}^q(2^{j+1}B)^{1/q}}\Bigg]\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \sum_{j=1}^\infty\nu_{\vec{w}}^q\big(2^{j+1}B\big)^{\kappa/p-1/q}. \end{split}$$ Since $(\nu_{\vec{w}})^q\in A_{mq}\subset A_\infty$, then it follows immediately from Lemma 2.3 that $$\frac{\nu_{\vec{w}}^q(B)}{\nu_{\vec{w}}^q(2^{j+1}B)}\le C\left(\frac{|B|}{|2^{j+1}B|}\right)^{\delta^\prime}.$$ Hence $$\begin{split} J^{\infty,\ldots,\infty}&\le\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}\big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_m)(x)\big|\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot\sum_{j=1}^\infty \frac{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}}{\nu_{\vec{w}}^q(2^{j+1}B)^{1/q-\kappa/p}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot\sum_{j=1}^\infty \left(\frac{|B|}{|2^{j+1}B|}\right)^{\delta^\prime(1/q-\kappa/p)}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}, \end{split}$$ where in the last inequality we have used the fact that $0<\kappa<p/q$ and $\delta^\prime>0$. We now consider the case where exactly $\ell$ of the $\alpha_i$ are $\infty$ for some $1\le\ell<m$. We only give the arguments for one of these cases. The rest are similar and can easily be obtained from the arguments below by permuting the indices. Using the definition of $I_\alpha$ again, we can see that for any $x\in B$, $$\begin{aligned} \big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|&=\int_{(\mathbb R^n)^{\ell}\backslash(2B)^{\ell}}\int_{(2B)^{m-\ell}}\frac{|f_1(y_1)\cdots f_m(y_m)|}{(|x-y_1|+\cdots+|x-y_m|)^{mn-\alpha}}dy_1\cdots dy_m\notag\\ &\le C\prod_{i=\ell+1}^m\int_{2B}\big|f_i(y_i)\big|\,dy_i\notag\\ &\times \sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^{m-\alpha/n}}\int_{(2^{j+1}B)^\ell\backslash(2^{j}B)^\ell} \big|f_1(y_1)\cdots f_{\ell}(y_\ell)\big|\,dy_1\cdots dy_\ell\notag\\ &\le C\prod_{i=\ell+1}^m\int_{2B}\big|f_i(y_i)\big|\,dy_i\times \sum_{j=1}^\infty\frac{1}{|2^{j+1}B|^{m-\alpha/n}}\prod_{i=1}^{\ell} \int_{2^{j+1}B\backslash 2^{j}B}\big|f_i(y_i)\big|\,dy_i\notag\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i,\end{aligned}$$ and we arrived at the expression considered in the previous case. Thus, for any $x\in B$, we also have $$\big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \sum_{j=1}^\infty\nu_{\vec{w}}^q\big(2^{j+1}B\big)^{\kappa/p-1/q}.$$ Therefore, by the inequality (4.2) and the above pointwise inequality, we obtain $$\begin{split} J^{\alpha_1,\ldots,\alpha_m}&\le\nu_{\vec{w}}^q(B)^{1/q-\kappa/p} \big|I_{\alpha}(f^\infty_1,\ldots,f^\infty_\ell,f^0_{\ell+1},\ldots,f^0_m)(x)\big|\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot\sum_{j=1}^\infty \frac{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}}{\nu_{\vec{w}}^q(2^{j+1}B)^{1/q-\kappa/p}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot\sum_{j=1}^\infty \left(\frac{|B|}{|2^{j+1}B|}\right)^{\delta^\prime(1/q-\kappa/p)}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}. \end{split}$$ Summarizing the estimates derived above and then taking the supremum over all balls $B\subseteq\mathbb R^n$, we finish the proof of Theorem 1.3. As before, fix a ball $B=B(x_0,r_B)\subseteq\mathbb R^n$ and split $f_i$ into $f_i=f^0_i+f^{\infty}_i$, where $f^0_i=f_i\chi_{2B}$, $i=1,\ldots,m$. Then for each fixed $\lambda>0$, we can write $$\begin{split} &\nu_{\vec{w}}^q\big(\big\{x\in B:\big|I_{\alpha}(f_1,\ldots,f_m)\big|>\lambda\big\}\big)^{1/q}\\ \le&\nu_{\vec{w}}^q\big(\big\{x\in B:\big|I_{\alpha}(f^0_1,\ldots,f^0_m)\big|>\lambda/{2^m}\big\}\big)^{1/q} +\sum\nolimits^\prime \nu_{\vec{w}}^q\big(\big\{x\in B:\big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)\big|>\lambda/{2^m}\big\}\big)^{1/q}\\ =&J^0_*+\sum\nolimits^\prime J^{\alpha_1,\ldots,\alpha_m}_*, \end{split}$$ where each term of $\sum^\prime$ contains at least one $\alpha_i\neq0$. By Lemma 2.5 again, we know that $(\nu_{\vec{w}})^q\in A_{mq}$ with $1<mq<\infty$. Using Theorem D, Lemma 4.1 and Lemma 2.1, we have $$\begin{split} J^0_*&\le\frac{C}{\lambda}\prod_{i=1}^m\left(\int_{2B}|f_i(x)|^{p_i}w_i(x)^{p_i}\,dx\right)^{1/{p_i}}\\ &\le \frac{C\cdot\prod_{i=1}^m w^{q_i}_i(2B)^{{\kappa q}/{pq_i}}}{\lambda}\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\\ &\le \frac{C\cdot\nu_{\vec{w}}^q(2B)^{\kappa/p}}{\lambda}\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\\ &\le \frac{C\cdot\nu_{\vec{w}}^q(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m\big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}. \end{split}$$ In the proof of Theorem 1.3, we have already proved the following pointwise estimate (see (4.1) and (4.3)). $$\big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|\le C\sum_{j=1}^\infty\prod_{i=1}^m\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i.$$ Without loss of generality, we may assume that $p_1=\cdots=p_{\ell}=\min\{p_1,\ldots,p_m\}=1$, and $p_{\ell+1},\ldots,p_m>1$. By using Hölder’s inequality, the multiple $A_{\vec{P},q}$ condition and Lemma 4.1, we obtain $$\begin{split} \big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|&\le C\sum_{j=1}^\infty\prod_{i=1}^{\ell}\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i\\ &\times\prod_{i=\ell+1}^{m}\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|\,dy_i\\ &\le C\sum_{j=1}^\infty\prod_{i=1}^{\ell} \frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}}\int_{2^{j+1}B}\big|f_i(y_i)\big|w_i(y_i)\,dy_i \left(\inf_{y_i\in 2^{j+1}B}w_i(y_i)\right)^{-1}\\ &\times\prod_{i=\ell+1}^{m}\frac{1}{|2^{j+1}B|^{1-\alpha/{mn}}} \left(\int_{2^{j+1}B}\big|f_i(y_i)\big|^{p_i}w_i(y_i)^{p_i}\,dy_i\right)^{1/{p_i}} \left(\int_{2^{j+1}B}w_i(y_i)^{-p'_i}\,dy_i\right)^{1/{p'_i}}\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \sum_{j=1}^\infty\nu_{\vec{w}}^q\big(2^{j+1}B\big)^{\kappa/p-1/q}. \end{split}$$ Note that $(\nu_{\vec{w}})^q\in A_{mq}$ with $1<mq<\infty$. Hence, it follows from the inequality (4.2) that for any $x\in B$, $$\begin{aligned} \big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big| &= C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{1}{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}} \sum_{j=1}^\infty\frac{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}}{\nu_{\vec{w}}^q(2^{j+1}B)^{1/q-\kappa/p}}\notag\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{1}{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}}\sum_{j=1}^\infty \left(\frac{|B|}{|2^{j+1}B|}\right)^{\delta^\prime(1/q-\kappa/p)}\notag\\ &\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{1}{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}}.\end{aligned}$$ If $\big\{x\in B:\big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|>\lambda/{2^m}\big\}=\O$, then the inequality $$J^{\alpha_1,\ldots,\alpha_m}_*\le\frac{C\cdot\nu_{\vec{w}}^q(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}$$ holds trivially. Now if instead we assume that $\big\{x\in B:\big|I_{\alpha}(f^{\alpha_1}_1,\ldots,f^{\alpha_m}_m)(x)\big|>\lambda/{2^m}\big\}\neq\O$, then by the pointwise inequality (4.5), we get $$\lambda<C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}\cdot \frac{1}{\nu_{\vec{w}}^q(B)^{1/q-\kappa/p}},$$ which in turn gives that $$\nu_{\vec{w}}^q(B)^{1/q}\le\frac{C\cdot\nu_{\vec{w}}^q(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}.$$ Therefore $$J^{\alpha_1,\ldots,\alpha_m}_*\le\nu_{\vec{w}}^q(B)^{1/q} \le\frac{C\cdot\nu_{\vec{w}}^q(B)^{\kappa/p}}{\lambda}\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)}.$$ Collecting all the above estimates and then taking the supremum over all balls $B\subseteq\mathbb R^n$ and all $\lambda>0$, we conclude the proof of Theorem 1.4. By using Hölder’s inequality, it is easy to verify that if $1\le p_i<q_i$, $1/q=\sum_{k=1}^m 1/{q_k}$ and each $w_i$ is in $A_{p_i,q_i}$, then we have $$\prod_{i=1}^m A_{p_i,q_i}\subset A_{\vec{P},q}.$$ and this inclusion is strict (see [@moen]).Also recall that $w\in A_{p,q}$ if and only if $w^q\in A_{1+q/{p'}}\subset A_\infty$(see [@muckenhoupt2]). Thus, as straightforward consequences of Theorems 1.3 and 1.4, we finally obtain the following Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in(1,\infty)$, $1/p=\sum_{k=1}^m 1/{p_k}$, $1/{q_k}=1/{p_k}-\alpha/{mn}$ and $1/q=\sum_{k=1}^m 1/{q_k}=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)\in \prod_{i=1}^m A_{p_i,q_i}$, then for any $0<\kappa<p/q$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{L^{q,{\kappa q}/p}((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. Let $m\ge2$, $0<\alpha<mn$ and $I_\alpha$ be an $m$-linear fractional integral operator. If $p_1,\ldots,p_m\in[1,\infty)$, $\min\{p_1,\ldots,p_m\}=1$, $1/p=\sum_{k=1}^m 1/{p_k}$, $1/{q_k}=1/{p_k}-\alpha/{mn}$ and $1/q=\sum_{k=1}^m 1/{q_k}=1/p-\alpha/n$, and $\vec{w}=(w_1,\ldots,w_m)\in \prod_{i=1}^m A_{p_i,q_i}$, then for any $0<\kappa<p/q$, there exists a constant $C>0$ independent of $\vec{f}=(f_1,\ldots,f_m)$ such that $$\big\|I_\alpha(\vec{f})\big\|_{WL^{q,{\kappa q}/p}((\nu_{\vec{w}})^q)}\le C\prod_{i=1}^m \big\|f_i\big\|_{L^{p_i,{\kappa p_iq}/{pq_i}}(w^{p_i}_i,w^{q_i}_i)},$$ where $\nu_{\vec{w}}=\prod_{i=1}^m w_i$. [99]{} D. R. Adams, A note on Riesz potentials, Duke Math. J, **42**(1975), 765–778. X. Chen and Q. Y. Xue, Weighted estimates for a class of multilinear fractional type operators, J. Math. Anal. Appl, **362**(2010), 355–373. F. Chiarenza and M. Frasca, Morrey spaces and Hardy–Littlewood maximal function, Rend. Math. Appl, **7**(1987), 273–279. R. R. Coifman and Y. Meyer, On commutators of singular integrals and bilinear singular integrals, Trans. Amer. Math. Soc, **212**(1975), 315–331. J. Duoandikoetxea, Fourier Analysis, American Mathematical Society, Providence, Rhode Island, 2000. D. S. Fan, S. Z. Lu and D. C. Yang, Regularity in Morrey spaces of strong solutions to nondivergence elliptic equations with VMO coefficients, Georgian Math. J, **5**(1998), 425–440. G. Di Fazio and M. A. Ragusa, Interior estimates in Morrey spaces for strong solutions to nondivergence form equations with discontinuous coefficients, J. Funct. Anal, **112**(1993), 241–256. G. Di Fazio, D. K. Palagachev and M. A. Ragusa, Global Morrey regularity of strong solutions to the Dirichlet problem for elliptic equations with discontinuous coefficients, J. Funct. Anal, **166**(1999), 179–196. J. Garcia-Cuerva and J. L. Rubio de Francia, Weighted Norm Inequalities and Related Topics, North-Holland, Amsterdam, 1985. L. Grafakos, On multilinear fractional integrals, Studia Math., **102**(1992), 49–56. L. Grafakos and N. Kalton, Multilinear Calderón–Zygmund operators on Hardy spaces, Collect. Math., **52**(2001), 169–179. L. Grafakos and R. H. Torres, Multilinear Calderón–Zygmund theory, Adv. Math., **165**(2002), 124–164. L. Grafakos and R. H. Torres, Maximal operator and weighted norm inequalities for multilinear singular integrals, Indiana Univ. Math. J., **51**(2002), 1261–1276. L. Grafakos and R. H. Torres, On multilinear singular integrals of Calderón–Zygmund type, Publ. Mat., (2002), Vol. Extra, 57–91. G. E. Hu and Y. Meng, Multilinear Calderón–Zygmund operator on products of Hardy spaces, Acta Math. Sinica (Engl. Ser), **28**(2012), 281–294. T. Iida, E. Sato, Y. Sawano and H. Tanaka, Weighted norm inequalities for multilinear fractional operators on Morrey spaces, Studia Math., **205**(2011), 139–170. T. Iida, E. Sato, Y. Sawano and H. Tanaka, Multilinear fractional integrals on Morrey spaces, Acta Math. Sinica (Engl. Ser), **28**(2012), 1375–1384. T. Iida, E. Sato, Y. Sawano and H. Tanaka, Sharp bounds for multilinear fractional integral operators on Morrey type spaces, Positivity, **16**(2012), 339–358. C. E. Kenig and E. M. Stein, Multilinear estimates and fractional integration, Math. Res. Lett., **6**(1999), 1–15. Y. Komori and S. Shirai, Weighted Morrey spaces and a singular integral operator, Math. Nachr, **282**(2009), 219–231. A. K. Lerner, S. Ombrosi, C. Pérez, R. H. Torres and R. Trujillo-González, New maximal functions and multiple weights for the multilinear Calderón–Zygmund theory, Adv. Math., **220**(2009), 1222–1264. W. J. Li, Q. Y. Xue and K. Yabuta, Maximal operator for multilinear Calderón–Zygmund singular integral operators on weighted Hardy spaces, J. Math. Anal. Appl, **373**(2011), 384–392. W. J. Li, Q. Y. Xue and K. Yabuta, Multilinear Calderón–Zygmund operators on weighted Hardy spaces, Studia Math., **199**(2010), 1–16. K. Moen, Weighted inequalities for multilinear fractional integral operators, Collect. Math., **60**(2009), 213–238. C. B. Morrey, On the solutions of quasi-linear elliptic partial differential equations, Trans. Amer. Math. Soc, **43**(1938), 126–166. B. Muckenhoupt, Weighted norm inequalities for the Hardy maximal function, Trans. Amer. Math. Soc, **165**(1972), 207–226. B. Muckenhoupt and R. L. Wheeden, Weighted norm inequalities for fractional integrals, Trans. Amer. Math. Soc, **192**(1974), 261–274. J. Peetre, On the theory of $\mathcal L_{p,\lambda}$ spaces, J. Funct. Anal, **4**(1969), 71–87. G. Pradolini, Weighted inequalities and pointwise estimates for the multilinear fractional integral and maximal operators, J. Math. Anal. Appl, **367**(2010), 640–656. L. Tang, Endpoint estimates for multilinear fractional integrals, J. Aust. Math. Soc, **84**(2008), 419–429. H. Wang, Some estimates for commutators of Calderón–Zygmund operators on the weighted Morrey spaces, Sci. Sin. Math, **42**(2012), 31–45. H. Wang, The boundedness of some operators with rough kernel on the weighted Morrey spaces, Acta Math. Sinica (Chin. Ser), **55**(2012), 589–600. H. Wang, Intrinsic square functions on the weighted Morrey spaces, J. Math. Anal. Appl, **396**(2012), 302–314. H. Wang, Boundedness of fractional integral operators with rough kernels on weighted Morrey spaces, Acta Math. Sinica (Chin. Ser), **56**(2013), 175–186. H. Wang, Some estimates for commutators of fractional integral operators on weighted Morrey spaces, Acta Math. Sinica (Chin. Ser), to appear. H. Wang, Some estimates for commutators of fractional integrals associated to operators with Gaussian kernel bounds on weighted Morrey spaces, Anal. Theory Appl, to appear. H. Wang, Weak type estimates for intrinsic square functions on the weighted Morrey spaces, preprint, 2012. H. Wang and H. P. Liu, Some estimates for Bochner–Riesz operators on the weighted Morrey spaces, Acta Math. Sinica (Chin. Ser), **55**(2012), 551–560.

--- abstract: 'Transfinite patches provide a simple and elegant solution to the problem of representing non-four-sided continuous surfaces, which are useful in a variety of applications, such as curve network based design. Real-time responsiveness is essential in this context, and thus reducing the computation cost is an important concern. The Midpoint Coons (MC) patch presented in this paper is a fusion of two previous transfinite schemes, combining the speed of one with the superior control mechanism of the other. This is achieved using a new constrained parameterization based on generalized barycentric coordinates and transfinite blending functions.' author: - | Péter Salvi, István Kovács and Tamás Várady\ Budapest University of Technology and Economics bibliography: - 'cikkek.bib' - 'sajat.bib' title: Computationally efficient transfinite patches with fullness control --- \[sec:Introduction\]Introduction ================================ The representation of multi-sided surfaces is a difficult problem, with different solutions suited to different applications. Most CAD systems provide trimmed tensor product surfaces to reproduce patches of arbitrary sides in a mechanical model, while in computer graphics subdivision surfaces are the de facto standard. Splitting multi-sided regions into quadrilateral patchespreserving continuity along splitting curvesis also a well-researched approach. Transfinite interpolation has many advantages over the above methods: (i) it retains the continuity of its edge curves on the whole domain, while exactly interpolates $G^{1}$ boundary conditions; (ii) it depends only on the boundary data, without the need of additional control points or polyhedra; (iii) smooth connections to adjacent surfaces are easily ensured. There are also some limitations: (i) as the whole patch is being defined as a blend of its boundaries, there is little control over the center of the surface; (ii) computation costs are somewhat higher compared to conventional techniques; (iii) the computation of derivatives is complex (normally discrete approximations are used). One of these concerns, central control, has been alleviated recently by the addition of an extra degree of freedom (adjusting surface fullness) to the Gregory patch [@Gregory:1986], the most popular choice of multi-sided transfinite patch [@Salvi:2016:GrafGeo]. As for computational complexity, a practically equivalent, but much more efficient formulation has been given for the same surface [@Salvi:2014; @Salvi:2015:KEPAF]. These two modifications of the Gregory patch are, however, not compatible. The problem lies in the parameterization of the domain, which is enhanced in this paper in such a way as to accommodate for both adjustments, and thus create an efficient representation capable of fullness control. The rest of the paper is organized as follows. In Section \[sec:Previous-work\] we give a short review of related research. In Section \[sec:Preliminaries\] the necessary details of the above two transfinite surfaces are presented. The new Midpoint Coons (MC) patch is introduced in Section \[sec:Midpoint-Coons-patches\], and test results are shown in Section \[sec:Test-results\]. \[sec:Previous-work\]Previous work ================================== {width="\textwidth"} There has not been much work done on the interior control of transfinite interpolation surfaces, maybe because these were regarded as the means of filling multi-sided holes, not as a design tool. One exception is an earlier work of the authors [@Varady:2012], where the patch center can be adjusted via auxiliary surfaces. The formulation is a modification of Kato’s patch [@Kato:1991]. As briefly outlined above, this paper mainly draws on two previous representations: the Generalized Coons (GC) and Midpoint (MP) surfaces [@Salvi:2014; @Salvi:2016:GrafGeo]. The former is a multi-sided generalization of the Coons patch, which is shown to be virtually the same as the Gregory patch [@Salvi:2015:KEPAF], while the latter introduces a central control point for fullness control. Both of these are presented in detail in Section \[sec:Preliminaries\]. Our primary contribution, the constrained parameterization in Section \[subsec:Constrained-barycentric\], is based on generalized barycentric coordinates (see e.g. Floater [@Floater:2015]) and the blending function of Kato’s patch [@Kato:1991]. \[sec:Preliminaries\]Preliminaries ================================== In this section we will first look at the various constituents of transfinite patches, then review some concrete constructions: the Gregory patch, and its two enhancements, the Generalized Coons (GC) and Midpoint (MP) patches. Given a loop of 3D curves and cross-derivatives, we want to generate a surface that interpolates these boundary conditions (in a $G^{1}$ sense, i.e., only tangent planes are reproduced). Cross-derivatives can be defined automatically using a frame sweep (e.g. with a rotation-minimizing frame [@Wang:2008]), or semi-automatically by first fixing normal vectors at arbitrary points. We assume that the cross-derivative functions are twist-compatible; otherwise rational twists can be applied [@Gregory:1974]. Transfinite surfaces come in two flavors: side-based and corner-based schemes, see Figure \[fig:Side–and-corner-based\]. Side-based schemes blend together *side-interpolants* (four-sided surfaces that interpolate one side), while corner-based schemes use *corner-interpolants* (four-sided surfaces that interpolate two adjacent sides). In practice it is much easier to construct side-interpolants, so we will define corner-interpolants based on these, as well. A side-interpolant, or *ribbon*, for side $i$ is defined as $$R_{i}(s_{i},d_{i})=P_{i}(s_{i})+\gamma(d_{i})T_{i}(s_{i}),$$ where $P_{i}(s_{i})$ is the $i$-th boundary curve parameterized in $[0,1]$, $T_{i}$ is the corresponding cross-derivative, and $\gamma$ is a scaling function. (A recommended $\gamma$ function is $\gamma(d)=d/(2d+1)$, its derivation can be found in Salvi et al. [@Salvi:2014]) The arguments $s_{i}$ and $d_{i}$ are the side- and distance-parameters, with $$\begin{aligned} s_{i} & \in[0,1], & d_{i} & \geq0.\end{aligned}$$ The $n$-sided patch itself is defined over a convex polygonal domain, e.g. a regular $n$-sided polygon in the $(u,v)$ plane. A crucial component of a transfinite scheme is the parameter mapping from $(u,v)$ to $(s_{i},d_{i})$, i.e., from the $n$-sided polygon to each ribbon’s own parameterization. A basic constraint is that for a point on side $i$ of the domain polygon, $$\begin{aligned} s_{i-1} & =1, & s_{i+1} & =0, & d_{i} & =0.\end{aligned}$$ Also, the side parameter $s_{i}$ changes linearly from 0 to 1, and the distance parameter $d_{i}$ grows monotonically as we go inside the domain. Finally, we will need suitable blending functions that interpolate the ribbons at the boundaries, but blend them together inside the patch. \[subsec:Gregory-patch\]Gregory patch ------------------------------------- The classic Gregory patch is a corner-based[^1] scheme, so our first step is the creation of corner interpolants. These can be constructed as the Boolean sum of two adjacent ribbons, with the common part subtracted: $$\begin{aligned} I_{i,i-1}(u,v) & =R_{i-1}(s_{i-1},s_{i})+R_{i}(s_{i},1-s_{i-1})\nonumber \\ & -Q_{i,i-1}(s_{i},s_{i-1}),\end{aligned}$$ where the $Q_{i,i-1}$ correction patch is defined as $$\begin{aligned} Q_{i,i-1}(s_{i},s_{i-1}) & =P_{i}(0)+\gamma(1-s_{i-1})T_{i}(0)+\gamma(s_{i})T_{i-1}(1)\nonumber \\ & +\gamma(s_{i})\gamma(1-s_{i-1})W_{i,i-1}.\end{aligned}$$ Here $W_{i,i-1}$ is the (common) twist for the corner $(i,i-1)$. The patch equation is simply $$S_{CB}(u,v)=\sum_{i=1}^{n}I_{i,i-1}(u,v)B_{i,i-1}(u,v),$$ where $B_{i,i-1}$ is the blending function $$B_{i,i-1}(u,v)=\frac{D_{i,i-1}}{\sum_{j=1}^{n}D_{j,j-1}}=\frac{1/(d_{i}d_{i-1})^{2}}{\sum_{j=1}^{n}1/(d_{j}d_{j-1})^{2}}$$ with $D_{i,i-1}=\prod_{k\notin\{i,i-1\}}d_{k}^{2}$. For parameterization, radial side parameters and perpendicular distance parameters are used. As these are not relevant to the enhancements at hand, the reader is referred to the original paper [@Gregory:1986]. \[subsec:Generalized-Coons-patch\]Generalized Coons patch --------------------------------------------------------- This is a side-based formulation, similar in logic to the original four-sided Coons patch: $$\begin{aligned} S_{GC}(u,v) & =\sum_{i=1}^{n}R_{i}(s_{i},d_{i})B_{i}(u,v)\nonumber \\ & -\sum_{i=1}^{n}Q_{i,i-1}(s_{i},s_{i-1})B_{i,i-1}(u,v).\end{aligned}$$ Here $B_{i}(u,v)=B_{i,i-1}(u,v)+B_{i+1,i}(u,v)$ is a blending function assigned to the $i$-th side; everything else is as before. There are, however, more constraints on the parameters. For a point on the $i$-th side: $$\begin{aligned} d_{i-1} & =s_{i}, & d_{i+1} & =1-s_{i},\label{eq:gc-param-constr}\\ d_{i-1}' & =s_{i}', & d_{i+1}' & =-s_{i}',\label{eq:gc-param-constr-2}\end{aligned}$$ where the prime symbol means all directional derivatives. These are satisfied by a blended constructiondetails can be found in the original paper [@Salvi:2014]. Because of this constrained parameterization, there are less ribbon evaluations, and thus the computational cost of this surface is about 25% less than that of the Gregory patch, while there is no noticeable change in the surface [@Salvi:2015:KEPAF]. \[subsec:Midpoint-patch\]Midpoint patch --------------------------------------- Here a new degree of freedom was added to the Gregory patch, in form of a central control point $P_{0}$: $$S_{MP}(u,v)=\sum_{i=1}^{n}I_{i,i-1}(u,v)B_{i,i-1}^{*}(u,v)+P_{0}B_{0}^{*}(u,v).$$ The modified blending functions are defined as $$\begin{aligned} B_{i,i-1}^{*}(u,v) & =\frac{d_{i}H(1-s_{i-1})H(d_{i-1})+d_{i-1}H(s_{i})H(d_{i})}{d_{i}+d_{i-1}},\nonumber \\ B_{0}^{*}(u,v) & =1-\sum_{i=1}^{n}B_{i,i-1}^{*}(u,v),\end{aligned}$$ where $H(x)$ is a Hermite blend $$H(x)=(1-x)^{3}+3(1-x)^{2}x.$$ Note that by definition $\sum_{i=1}^{n}B_{i,i-1}^{*}(u,v)+B_{0}^{*}(u,v)=1$. The control point $P_{0}$ has the default location $$P_{0}=\frac{1}{n}\sum_{i=1}^{n}I_{i,i-1}(0.5,0.5),$$ but it can be used to move the center of the surface. In this scheme we also have a new constraint on the parameterization. For points on side $i$, it should satisfy $$d_{j}=1.\quad j\notin\{i-1,i,i+1\}\label{eq:mp-param-constr}$$ This can be achieved using generalized barycentric coordinates. Let us look at this in detail, as this will be the base of our new parameterization in Section \[subsec:Constrained-barycentric\]. \[subsec:Barycentric-parameterization\]Barycentric parameterization ------------------------------------------------------------------- Given a convex polygon with vertices $V_{i}$, the Wachspress coordinates $\{\lambda_{i}\}$ of a point $(u,v)$ have the following properties: $$\begin{aligned} \sum_{i=1}^{n}\lambda_{i}(u,v) & =1, & \lambda_{i}(V_{j}) & =\delta_{ij},\end{aligned}$$ and $\lambda_{i}$ decreases linearly on the adjacent domain edges. Now we can define side and distance parameters as $$\begin{aligned} s_{i} & =\lambda_{i}/\left(\lambda_{i-1}+\lambda_{i}\right), & d_{i} & =1-\lambda_{i-1}-\lambda_{i}.\end{aligned}$$ It is easy to see that this construction satisfies all requirements. An example is shown in Figure \[fig:Barycentric-parameterization\]. \[sec:Midpoint-Coons-patches\]Midpoint Coons patches ==================================================== We would like to combine the GC and MP patches, so that we have a computationally efficient transfinite patch with fullness control. The idea is simple: use the modified blending functions $B_{i,i-1}^{*}$ in the GC scheme: $$\begin{aligned} S_{MC}(u,v) & =\sum_{i=1}^{n}R_{i}(s_{i},d_{i})\left[B_{i,i-1}^{*}(u,v)+B_{i+1,i}^{*}(u,v)\right]\nonumber \\ & -\sum_{i=1}^{n}Q_{i,i-1}(s_{i},s_{i-1})B_{i,i-1}^{*}(u,v).\end{aligned}$$ Unfortunately this is not enough. The problem is in the parameterization: it has to satisfy both Eqs. (\[eq:gc-param-constr\])(\[eq:gc-param-constr-2\]) and (\[eq:mp-param-constr\]), which none of the original formulations were capable of. \[subsec:Constrained-barycentric\]Constrained barycentric parameterization -------------------------------------------------------------------------- The barycentric parameters described in Section \[subsec:Barycentric-parameterization\] already satisfy every constraint *except* Eq. (\[eq:gc-param-constr-2\]). This is evident in Figure \[fig:Barycentric-parameterization\], as the blue and red constant parameter lines do not have the same tangent as the green ones near the bottom. Looking at this figure, we can see that what we need is a new distance parameter $\hat{d}_{i}$ that “behaves” as $s_{i-1}$ when $s_{i}=0$, and as $s_{i+1}$ when $s_{i}=1$; but we want to retain the original $d_{i}$ for $d_{i}=0$ and $d_{i}=1$ (i.e., at the base side and at non-adjacent sides). Note that this is a very similar problem to transfinite surfaceswe have four one-dimensional boundary constraints to be interpolated, while for points inside the domain blended values are needed. Indeed, it can be solved using the blending functions of Kato’s patch [@Kato:1991]. Let the values at the boundaries be $$\begin{aligned} x_{1} & =d_{i}, & x_{2} & =s_{i+1}, & x_{3} & =d, & x_{4} & =1-s_{i-1},\end{aligned}$$ and the parameters in the corresponding square domain be $$\begin{aligned} t_{1} & =d_{i}, & t_{2} & =1-s_{i}, & t_{3} & =1-d_{i}, & t_{4} & =s_{i}.\end{aligned}$$ Then the new distance parameter is defined as $$\hat{d}_{i}=\sum_{j=1}^{4}x_{j}\hat{B}_{j}(u,v),$$ where $$\hat{B}_{j}(u,v)=\frac{\prod_{l\neq j}t_{l}^{2}}{\sum_{k=1}^{4}\prod_{l\neq k}t_{l}^{2}}=\frac{1/t_{j}^{2}}{\sum_{k=1}^{4}1/t_{k}^{2}}.$$ Note that this function is singular when $t_{i}=t_{j}=0$, $i\neq j$. This does not present a problem, as the parameterization is well-defined in these points. This $\hat{d}_{i}$, combined with the original side parameter $s_{i}$, gives a parameterization that satisfies all constraints. Figure \[fig:Constrained-barycentric\] shows the parameterization in a 6-sided domain. \[sec:Test-results\]Test results ================================ $n=3$ $n=4$ $n=5$ $n=6$ $n=7$ $n=8$ --------- ------- ------- ------- ------- ------- ------- CB 429ms 316ms 652ms 760ms 887ms 968ms GC 321ms 276ms 466ms 536ms 616ms 673ms MP 419ms 341ms 638ms 752ms 868ms 953ms MC 299ms 277ms 441ms 518ms 578ms 636ms Speedup 28.6% 18.8% 30.9% 31.1% 33.4% 33.3% ![\[fig:Deviation-between\]Deviations between 5-sided MP and MC patches. Full red color means $0.5\%$ of the bounding box axis. The green region has deviation less than $0.2\%$.](images/deviation){width="40.00000%"} As expected, the new MC patch shows an average of 30% speedup compared to the MP patch, see Table \[tab:Evaluation-speed\]. Figure \[fig:Deviation-between\] shows the deviation between an MP and an MC patch. The maximum deviation is $\approx0.4\%$ of the bounding box axis, with an average deviation of $\approx0.1\%$. (These values are even tighter, if the MP patch uses the same constrained parameterization.) The two patches are visually indistinguishable. In Figure \[fig:Changing-the-fullness\] we can see how the central control point affects the shape of the surface. This change has no effect on the $G^{1}$ interpolation properties, as can be seen from the contourings. Conclusion {#conclusion .unnumbered} ========== We have successfully combined two transfinite surface representations into a new one that takes the best of both worlds: fast evaluation and the ability to control the surface interior. Future work includes optimization with the GPU and the development of efficient derivative computation. [^1]: Because of this, it is also called “CB patch” [@Salvi:2014].

--- abstract: 'A lower bound of 135 MeV for the width of the $\xi$ meson is obtained from analyzing the $pp$ and ${\overline{p}p}$ interactions by use of Regge theory. The $pp$ data exclude a narrow $\xi$ as the latter would lead to $\sigma^{{\overline{p}p}}_{tot}$ far exceeding the measured ones. The broad width explains why the $\xi$ may not be seen in ${\overline{p}p}$ experiments.' address: | T-16, Theoretical Division, Los Alamos National Laboratory\ Los Alamos, NM 87545 U.S.A. author: - 'L.C. Liu' title: 'The width of the $\xi(2230)$ meson' --- Keywords: mesons, $J/\psi$ radiative decay, Regge theory. PACS: 12.40.Nn, 12.90.+b, 13.25.-k, 13.25.Gv One of the important predictions of the quantum chromodynamic (QCD) theory is the existence of bound states of the gluons (the glueballs), which arises from the self-coupling of the gluons. Calculations[@Morn] -[@Mich] have predicted that the lightest $2^{++}$ glueball has a mass about 2.2$\pm 0.3$ GeV. It is also generally believed that a pure glueball should decay flavor symmetric or flavor blind. The $\xi$ meson has been the center of attention in recent years because its reported mass and flavor-symmetric decay characteristics fit what one would expect from a glueball. This meson was first observed by the MARK III collaboration[@Balt] in the decay of $J/\psi$ to $K_{S}K_{S}$ and $K^{+}K^{-}$ channels and was found to have quantum numbers $J^{P}=(even)^{+}$, masses $M= 2230^{+6}_{-6} \pm 15 $ and $2232^{+7}_{-7} \pm7$ MeV, and widths $\Gamma = 18^{+23}_{-15} \pm 10$ and $26^{+26}_{-16} \pm 7$ MeV, respectively. The $\xi$ was also seen by the WA91 collaboration[@Brei], but with low statistics. Although the $\xi$ was not seen in a number of other experiments, in 1996 the BES collaboration[@Bai] reported the observation of $\xi$ in the radiative decay of the $J/\psi$ to ${\overline{p}p}, K^{+}K^{-}, K^{0}_{S}K^{0}_{S},$ and $ \pi^{+}\pi^{-}$ channels. The mass and the width of $\xi$ obtained in the BES-experiment agreed with those given in ref.[@Balt] and the data were found to be compatible with spin 2. Subsequently, the BES collaboration reported[@Bai2] [@Shen] the observation of $\xi{\rightarrow}\pi^{0}\pi^{0}, \eta\eta, \eta\eta', \eta'\eta'$ and concluded[@Shen] that the decays of $\xi$ to $K^{+}K^{-}, \pi^{+}\pi^{-}$, and $\pi^{0}\pi^{0}$ are flavor symmetric. The BES results raised, therefore, the expectation that the $\xi(2230)$ meson may be the lightest tensor glueball predicted by the QCD theory. As a result, there have been many experimental efforts aiming at verifying the BES results. One interesting experiment consisted in looking for a resonance structure in the ${\overline{p}p}{\rightarrow}\eta\eta, \pi^{0}\pi^{0}$ reactions. However, a systematic study in the mass region of 2222.7 to 2239.7 MeV led to negative results[@Seth], casting serious doubts on the very existence of the $\xi$. The challenge of understanding these conflicting experimental results has motivated this study. I will show how the high-energy ($>50$ GeV) $pp$ cross-section data can be used to set a limit for the width of the $\xi$. For notational convenience, let me denote the $s$-channel $pp{\rightarrow}pp$ and the $t$-channel ${p\overline{p}}{\rightarrow}{\overline{p}p}$ scatterings, respectively as $12{\rightarrow}34$ and $1\overline{3}{\rightarrow}\overline{2}4$. The total cross section of the $t$-channel ${\overline{p}p}{\rightarrow}\xi{\rightarrow}{\overline{p}p}$ scattering is given by $$\sigma_{tot}^{{\overline{p}p}}({\overline{s}}) = \frac{1}{8q^{2}}\sum_{{\lambda}'{\lambda}} {\cal I}m[A^{(t)}_{{\lambda}'{\lambda}}({\overline{s}}, {\overline{t}})]\mid_{{\overline{t}}=0} \label{eq:1}$$ where I have used the variables ${\overline{s}}$ and ${\overline{t}}$ to represent, respectively, the total c.m. energy and the momentum transfer in the $t$-channel. In eq.(\[eq:1\]) $q^{2}={\overline{s}}/4-m^{2}$ is the square of the c.m. three-momentum, with $m$ being the mass of the proton or antiproton. Furthermore, ${\lambda}'= {\lambda}_{\overline{2}}-{\lambda}_{4}$ and ${\lambda}= {\lambda}_{1}-{\lambda}_{\overline{3}}$, with $\lambda_{i}$ denoting the helicity of the particle $i$. In the helicity basis, the $t$-channel Feynman amplitude due to the exchange of $\xi$ can be written as $$A^{(t)}_{{\lambda}'{\lambda}}({\overline{s}},{\overline{t}})= \frac{-4\pi(2J+1)c_{{\lambda}';J} c_{J;{\lambda}} d^{J}_{{\lambda}'{\lambda}}(\theta_{t})} {{\overline{s}}-M^{2}+ iM\Gamma }\ , \label{eq:2}$$ where $J$, $M$, and $\Gamma$ are the spin, mass, and the width of the $\xi$. The $c_{{\lambda}';J} = 2m C_{I}G_{{\lambda}'}H_{{{\lambda}_{\overline{2}}{\lambda}_{4}};J}({\overline{s}})$ and $c_{J;{\lambda}} = 2m C_{I}G_{{\lambda}}H_{J;{{\lambda}_{1}{\lambda}_{\overline{3}}}}({\overline{s}})$ denote the coupling strength between the $\xi$ and the ${\overline{p}p}$ system, with $C_{I}=1/\sqrt{2}$ being the isospin coefficient for ${\overline{p}p}$ coupling to the isospin zero $\xi$ meson. The $H_{{{\lambda}_{\overline{2}}{\lambda}_{4}};J}$ and $H_{J;{{\lambda}_{1}{\lambda}_{\overline{3}}}}$ are the form factors for the $\xi{p\overline{p}}$ coupling vertex in the helicity basis with $G_{{\lambda}'}$ and $G_{{\lambda}}$ being the corresponding coupling constants. There are 16 helicity amplitudes but only 5 are independent. For forward ${\overline{p}p}$ scattering ($\theta_{t}=0$), $$\sigma^{{\overline{p}p}}_{tot}({\overline{s}}) = \frac{1}{8q^{2}} {\cal I}m[4A^{(t)}_{00}({\overline{s}}, {\overline{t}}=0)]\ . \label{eq:3}$$ The helicity-basis form factors are related to the canonical-basis form factors $F_{LS}$ by a unitary transformation [@Jaco]. For the ${\overline{p}p}$ system coupling to a $J^{P}=2^{+}$ state, the parity conservation leads to $L(= J\pm1)=1$ or 3, and $S=1$ only. Since only one value of $S$ is allowed, $F_{LS}$ will henceforth be denoted as $F_{L}$. The relations between the form factors in the above two bases are $G_{0}H_{{\uparrow}{\uparrow}} = G_{0}H_{{\downarrow}{\downarrow}} = \sqrt{2/5}g_{1}F_{1} -\sqrt{3/5}g_{3} F_{3}$ and $G_{1}H_{{\uparrow}{\downarrow}} = G_{-1}H_{{\downarrow}{\uparrow}} = \sqrt{3/5}g_{1}F_{1} + \sqrt{2/5}g_{3}F_{3}$. Here, ${\uparrow}$ and ${\downarrow}$ represent, respectively, ${\lambda}=+1/2$ and -1/2. The $F_{L}$ has the general form $F_{L}=q^{L}f(t)$. Here, the factor $q^{L}$ ensures the correct threshold behavior and $f(t)$ is an analytical function in both the $s$- and $t$-channels[@Liu]. The form factor can be normalized in such a way that $F_{L}(M^{2})$=1. Hence, at ${\overline{s}}=M^{2}$ and $J=2$, $A^{(t)}_{00}= i\ 40m^{2}\pi(\sqrt{2/5}g_{1} - \sqrt{3/5}g_{3})^{2}/M\Gamma$. The coupling constants $g_{1}$ and $g_{3}$ can be determined from the $pp$ total and elastic differential cross sections by use of Regge theory[@Regg][@Coll]. We recall that the Regge theory has been very successful in understanding high-energy hadronic scatterings. In recent years, Regge-theory based models have also been used to predict diffractive production of vector mesons[@Lee]. According to Regge theory, hadron-hadron scattering at very high energies can be described by the pole contribution of the Pomeron trajectory alone. This Pomeron dominance holds in the energy domain where $\sigma_{ph}(s) = \sigma_{p\overline{h}}(s)$. Here $h$ and $\overline{h}$ denote, respectively, a hadron and its antiparticle. Experiments have shown that this cross-section equality occurs[@Euro] at $\sqrt{s} > 40$ GeV. The Pomeron trajectory has the linear form $\alpha(t)=1.08 + \alpha'\ t$ at low $t$’s. Here, $t$ is the four-momentum transfer in the $s$-channel or the square of the mass Of the exchanged particle. Both the $\alpha$ and $\alpha'$ are complex functions of $t$, [*i.e.*]{}, $\alpha = \alpha_{_{R}} + i\alpha_{_{I}}$, and $\alpha' = \alpha'_{_{R}} + i\alpha'_{_{I}}$. For real $t$ lesser than the $t$-channel threshold, $\alpha'_{_{I}} = 0$ and, therefore, $\alpha(t) =$ $\alpha_{_{R}}(t) = 1.08 + \alpha'_{_{R}} t$. The Pomeron trajectory indicates that the mass of an exchanged particle having spin $J = \alpha_{R}(M^{2})$ is $M=[(J -1.08)/\alpha'_{R}]^{1/2}$. Model-independent analyses gave $\alpha'_{R}$ = $0.20\pm 0.02$ [@Coll] [@Bloc][@Note]. Hence, a particle of spin 2 will have a mass $M=2.05 - 2.26$ GeV, which overlaps with the mass range of the $\xi$. The Regge-pole amplitude $A^{(t)}_{{\lambda}'{\lambda}}$ due to the exchange of the Pomeron trajectory is given by[@Liu] $$A^{(t)}_{{\lambda_{\overline{2}}\lambda_{4};\lambda_{1}\lambda_{\overline{3}}}}(t,s)= C_{I} \frac{-4\pi^{2} (2\alpha +1)\beta_{{\lambda}'{\lambda}}(t) (-1)^{\alpha+{\lambda}}\frac{1}{2}[1+(-1)^{\alpha}]}{sin\pi(\alpha+{\lambda}')} d_{{\lambda}{\lambda}'}^{\alpha}(z_{t})\ , \label{eq:4}$$ where $\alpha\equiv \alpha(t)$ and $z_{t}\equiv cos\theta_{t}$. The variables $t$ and $s$ are the momentum transfer and energy in the $s-$channel. It is easy to verify that $t={\overline{s}}$ and $s={\overline{t}}$. The residue function $\beta_{{\lambda}'{\lambda}}$ is given by $$\beta_{{\lambda}'{\lambda}}(t)= 4\alpha_{R}'m^{2} G_{{\lambda}'}H_{{{\lambda}_{\overline{2}}{\lambda}_{4}};\alpha}(t) G_{{\lambda}}H_{\alpha;{{\lambda}_{1}{\lambda}_{\overline{3}}}}(t) \ . \label{eq:5}$$ Detailed expressions for the $pp$ total cross section and $pp$ elastic differential cross sections in terms of $A^{(t)}_{{\lambda}'{\lambda}}$ can be found in ref.[@Liu]. While the exponent $L$ in the threshold-behavior factor $q^{L}$ of the form factor was fixed with the integer values $J\pm1$ in [@Liu], in the present work I have used the relation $L = \alpha_{_{R}}(t) \pm 1$ to continue $L$ to noninteger values. The $g_{1}$ and $g_{3}$ obtained from fitting the $pp$ total cross sections and the diffractive peak of the $pp$ elastic differential cross sections[@Euro][@Bloc] at $\sqrt{s}= 53 $ and $62$ GeV are: $g_{1}= 1.55\pm0.06$ and $g_{3}= 0.24\pm0.02$, which are not too different from what was obtained with fixed integer $L$’s[@Note1]. Using these values in eq.(\[eq:3\]), I have calculated the $\sigma^{{\overline{p}p}} _{tot}({\overline{s}}=M^{2})$ as a function of $\Gamma$. These calculated cross sections are situated inside the zone between the two solid curves in fig.\[figw1\]. As one can see, small total widths lead to calculated cross sections that exceed the experimental $\sigma_{tot}^{{\overline{p}p}}$ (the dashed line). In order not to violate this experimental constraint, I deduce from figure \[figw1\] a lower bound of 135 MeV for the width of the $\xi$, which is much greater than the 22 MeV reported in the literature [@Balt][@Bai]. It is worth noting that the above broad width is in line with the finding of an earlier experiment by Alde [*et al.*]{}[@Alde] who observed a broad enhancement ($\Gamma \sim 140$ MeV) with spin 2 at $M=2220 $ MeV in the reaction $\pi^{-}p{\rightarrow}nX, X{\rightarrow}\eta\eta'$. The broad width given by the present analysis is also consistent with the non-observation of a resonance structure in the 2.23 GeV region in the ${\overline{p}p}{\rightarrow}\pi^{0}\pi^{0}, \eta\eta$ experiments of Seth [*et al.*]{}[@Seth]. Using the same notation as before, the ${\overline{p}p}{\rightarrow}h\overline{h}$ cross sections can be written as $$\left( \frac{d\sigma({\overline{s}},{\overline{t}})}{d{\overline{t}}}\right)_{a'a} = \frac{1}{256\pi q^{4}}\mid \sum_{{\lambda}'{\lambda}} \frac{-4\pi(2J+1)c'_{{\lambda}';J}c_{J;{\lambda}} d^{J}_{{\lambda}'{\lambda}}(\theta_{t}) }{{\overline{s}}- M^{2} + iM\Gamma} \mid^{2} \label{eq:6}$$ where the indices $a$ and $a'$ denote, respectively, the initial(${\overline{p}p}$) and the final ($\overline{h}h$) channels, with $c'_{{\lambda}';J}$ being the coupling strength of the final state $a'$ to the $\xi$. Clearly, a resonance structure will not show up in the energy dependence of the cross section if $c'_{{\lambda}';J}=0$, [*i.e.,*]{} if the resonance $\xi$ does not exist. However, I will show below that even if $\xi$ exists, a resonance structure may still not be seen when its width is broad. Let me introduce the angle-integrated cross section $$\Sigma({\overline{s}})= \int_{\overline{\Omega}}\left(\frac{d\sigma}{d{\overline{t}}}\right) _{a'a}d{\overline{t}}\label{eq:7}$$ and the variation of the cross sections $$V(\sqrt{{\overline{s}}_{\delta}})\equiv \frac{\Sigma(M^{2})}{\Sigma({\overline{s}}_{\delta})} \label{eq:8}$$ with $\overline{\Omega}$ being the solid angle over which the experimental differential cross sections were summed and $\sqrt{{\overline{s}}_{\delta}}\equiv M-\delta$. If $2\delta$ represents the mass interval covered by the measurement, then the greater the value of $V(\sqrt{{\overline{s}}_{\delta}})$, the more visible will be the resonance structure in this mass region. As the numerator in eq.(\[eq:6\]) does not vary rapidly with ${\overline{s}}$, we have $V(\sqrt{{\overline{s}}_{\delta}}) \simeq$ $ [({\overline{s}}_{\delta}-M^{2})^{2} + M^{2}\Gamma^{2} ] / M^{2}\Gamma^{2}$. For $M= 2.23$ GeV, $\sqrt{{\overline{s}}_{\delta}} = 2.223$ GeV, and $\Gamma\geq 135$ MeV, we have $V(2.223) \leq 1.01$. In other words, the energy dependence of the cross sections is practically flat in the mass region $2230\pm 7$ MeV. This was exactly the finding of ref.[@Seth]. In fact, even if $\Gamma$ were 50 MeV, one would still only have $V(2.223)=1.08$, i.e. only an 8% variation. Only when the width is 22 MeV, can a 40% variation be seen. Hence, the ${\overline{p}p}$ experiment of ref.[@Seth] can only confirm the existence of a narrow resonance. However, a non-observation of the resonance structure is not sufficient grounds for rejecting the existence of $\xi$. Had $\xi$ been an isolated resonance, it would have been possible to check the resonance by choosing a sufficiently large $\delta$. However, there are two isoscalar tensor resonances, the $f_{2}(2150)$ and $f_{2}(2300)$, situated only 80 MeV away and each of them has a width of about 150 to 160 MeV. These nearby resonances prevents us from probing the $\xi$ across a large mass interval. In conclusion, the high-energy $pp$ data and the experimental ${\overline{p}p}$ total cross section at c.m. energy $\sqrt{s_{{\overline{p}p}}} = 2.23$ GeV imply that the total width of the $\xi$ is at least 135 MeV. This broad width is compatible with the non-observation of the $\xi$ in the ${\overline{p}p}$ experiment of ref.[@Seth]. As the mass of the $\xi$ is situated in the region where the lightest $2^{++}$ tensor glueball is expected, it is of great importance to ascertain the existence of the $\xi$. We note that $q\overline{q}$ states can also have the quantum number $2^{++}$ and, therefore, can mix with the $2^{++}$ gluonium. In this latter case, an experimentally observed $2^{++}$ state can have a broad width. The experimental results of ref.[@Seth] and the present analysis all suggest that the $\xi$ meson, if it exists, should have a broad width. It was reported at the Hadron2001 Conference in Protvino, Russia, that the BES collaboration had repeated the measurement of $J/\psi$ radiative decays with much improved statistics. It will be of great interest to learn the new result. [7]{} Collin J. Morningstar and Mike Peardon [*Phys. Rev.D*]{}[**60**]{} (1999) 034509. Adam Szczepaniak, Eric S. Swanson, Chueng-Ryong Ji, and Stephen R. Cotanch, [*Phys. Rev. Lett.*]{}[**76**]{} (1996) 2011. C. Michael, [*AIP Conference Proc.*]{}[**432**]{} (1997) 657. R. M. Baltkusaitis et al., [*Phys. Rev. Lett.*]{}[**56**]{} (1986) 107. J. Breitweg et al., ZEUS collaboration, DESY 98-107; S. Abatzis et al., [*Phys. lett. B*]{}[**324**]{} (1994) 509. J. Z. Bai et al., [*Phys. Rev. Lett.*]{} [**76**]{} (1996) 3502. J. Z. Bai et al., [*Phys. Rev. Lett.*]{}[**81**]{} (1998) 1179. X. Shen, in [*Hadron Spectroscoy*]{}, ed. by S.U. Chung and H. Willutzki, [*AIP Conf. Proc.*]{} [**432**]{}, (1998) 47. K. K. Seth, [*Nucl. Phys. A*]{}[**675**]{} (2000) 25c. M. Jacob and G. C. Wick, [*Ann. Phys. (N.Y)*]{}[**7**]{}, (1959) 404. L. C. Liu and W. X. Ma, [*J. Phys. G*]{}[**26**]{} (2000) L59. T. Regge, [*Nuovo Cimento*]{} [**14**]{} (1959) 951; [**18**]{} (1959) 947. P. D. B. Collins, [*An Introduction to Regge Theory and High Energy Physics*]{}, Cambridge University Press, 1977. See section 6.8e for $\alpha'$. M. A. Pichowsky and T.-S. H. Lee, [*Phys. Lett. B*]{} [**379**]{} (1996) 1. Review of Particle Physics, [ *Euro. Phys. J.*]{} [**15**]{} (2000) 232-233. M. A. Block, K. Kang, and A. R. White, [*Int’l J. Mod. Phys. A*]{}[**7**]{} (1992) 4449. The uncertainties of $\alpha'$ were not given in the original articles (refs.[@Coll] and [@Bloc]); they are the results of fits to the same inputs by this author. The $g$ in table 1 of ref.[@Liu] should be multiplied by $\sqrt{2}$ to correct for the misprint. D. M. Alde et al., [*Phys. Lett. B*]{}[**177**]{} (1986) 120.

--- abstract: 'Narrowing was first studied in the context of equational E-unification and has been used in a wide range of applications. The classic completeness result due to Hullot states that any term rewriting derivation starting from an instance of an expression can be ‘lifted’ to a narrowing derivation, whenever the substitution employed is normalized. In this paper we adapt the generator-based extra-variables-elimination transformation used in functional-logic programming to overcome that limitation, so we are able to lift term rewriting derivations starting from arbitrary instances of expressions. The proposed technique is limited to left-linear constructor systems and to derivations reaching a ground expression. We also present a Maude-based implementation of the technique, using natural rewriting for the on-demand evaluation strategy.' author: - Adrián Riesco - 'Juan Rodríguez-Hortalá' bibliography: - 'references.bib' title: 'Lifting Term Rewriting Derivations in Constructor Systems by Using Generators[^1]' --- Introduction {#sect:intro} ============ Prelimininaries and formal setting {#sect:setting} ================================== The generators approach {#sect:generators} ======================= Maude prototype {#maude} =============== Concluding remarks and ongoing work {#sect:conclusions} =================================== [^1]: Research supported by MICINN Spanish project *StrongSoft* (TIN2012-39391-C04-04).

--- abstract: 'We study the space complexity of sketching cuts and Laplacian quadratic forms of graphs. We show that any data structure which approximately stores the sizes of all cuts in an undirected graph on $n$ vertices up to a $1+\epsilon$ error must use $\Omega(n\log n/\epsilon^2)$ bits of space in the worst case, improving the $\Omega(n/\epsilon^2)$ bound of [@andoni] and matching the best known upper bound achieved by spectral sparsifiers [@bss]. Our proof is based on a rigidity phenomenon for cut (and spectral) approximation which may be of independent interest: any two $d-$regular graphs which approximate each other’s cuts significantly better than a random graph approximates the complete graph must overlap in a constant fraction of their edges.' author: - 'Charles Carlson[^1]' - 'Alexandra Kolla [^2]' - 'Nikhil Srivastava[^3]' - 'Luca Trevisan[^4]' bibliography: - 'main.bib' title: Optimal Lower Bounds for Sketching Graph Cuts --- Introduction ============ An $\epsilon-$spectral sparsifier [@st04] of a weighted graph $G$ is a weighted graph $H$ such that for every $x\in{\mathbb{R}}^n$ we have the multiplicative guarantee: $$\label{eqn:specapprox} (1-\epsilon) x^TL_Gx \le x^TL_Hx \le (1+\epsilon)x^TL_Gx.$$ An $\epsilon-$cut sparsifier [@bk96] is a graph such that holds for all $x\in\{-1,1\}^n$; the former condition clearly implies the latter, and it is not hard to see that the converse is not true. There have been many papers on efficiently constructing sparsifiers with few edges (e.g. [@ss; @fhhp; @zhu; @lee2017sdp]). The best known sparsity-approximation tradeoff is achieved in [@bss], who showed that every graph on $n$ vertices has an $\epsilon$-spectral sparsifier with $O(n/\epsilon^2)$ edges (and this is also the best known upper bound for $\epsilon$-cut sparsification). Since every weighted edge can be stored using $O(\log n)$ bits[^5], storing such a sparsifier requires $O(n\log n/\epsilon^2)$ bits. The recent work [@andoni] studied the question of whether it is possible to use substantially fewer bits if one simply wants a data structure (not necessarily a graph) which answers cut and quadratic form queries. We will use the following definition, which is inspired by the one in [@sidford]. An [*$\epsilon-$spectral sketch*]{} of a graph $G$ is a function $f:{\mathbb{R}}^n\rightarrow{\mathbb{R}}$ such that for every $x\in{\mathbb{R}}^n$: $$\label{eqn:sketchapprox}(1-\epsilon)x^TL_Gx \le f(x)\le (1+\epsilon)x^TL_Gx.$$ An [*$\epsilon-$cut sketch*]{} is a function $f:\{-1,1\}^n\rightarrow{\mathbb{R}}$ such that holds for all $x\in\{-1,1\}^n$. A [*sketching scheme*]{} is a deterministic map ${\mathsf{sk}}$ from graphs on $n$ vertices to (${\epsilon}$-spectral or ${\epsilon}$-cut) sketches $f:{\mathbb{R}}^n\rightarrow{\mathbb{R}}$, along with specified procedures for storing the functions as bit strings and for evaluating any given $f$ on a query $x$ [^6]. The number of bits required to store $f$ is called its [*size*]{}. In [@andoni] it was shown that any ${\epsilon}$-cut sketch must use $\Omega(n/\epsilon^2)$ bits in the worst case, leaving a logarithmic gap between the best known upper and lower bounds. In this paper, we close this gap by showing that any $\epsilon$-cut sketching scheme must in fact use $\Omega(n\log n/\epsilon^2)$ bits whenever $\epsilon=\omega(n^{-1/4})$ (Theorem \[thm:cutlb\]), which means that it is not in general possible to obtain any asymptotic savings by considering sketches which are not graphs. We also give a lowerbound for $\epsilon$-spectral sketching with a simpler proof and slightly better constants (Theorem \[thm:speclb\]). Related Work ------------ The paper [@andoni] also studied the problem of producing a “for each” sketching algorithm, which has the property that for any particular query $x$, $f(x)$ approximates $x^TLx$ with constant probability. They showed that in this weaker model it is possible to obtain ${\epsilon}-$cut sketches of size $O(n{\mathrm{polylog}}(n) /\epsilon)$ and ${\epsilon}$-spectral sketches of size $O(n{\mathrm{polylog}}(n)/\epsilon^{1.6})$. The latter bound was recently improved to $O(n{\mathrm{polylog}}(n)/\epsilon)$ and generalized to include queries to the pseudoinverse in [@sidford]. In contrast, in this paper we study only the “for all” model, in which the sketch is required to work for all queries simultaneously. Techniques ---------- Our proof is based on the following rigidity phenomenon: there is a constant $c$ such that if two $d-$regular graphs $c/\sqrt{d}$-cut approximate each other then they must overlap in a constant fraction of edges (Lemmas \[lem:specrigid\] and \[lem:cutrigid\]). This allows us to argue that below this approximation threshold any sketch encodes a constant fraction of the information in the graph, and so any sketch must use $\Omega(dn\log n)=\Omega(n\log n/\epsilon^2)$ bits. Interestingly, this phenomenon is very sensitive to the value of the constant $c$ — in particular, a well-known theorem of Friedman [@friedman] says that the eigenvalues of the Laplacian of a random $d-$regular graph are contained in the range $[d-2\sqrt{d-1}-o(1),d+2\sqrt{d}+o(1)]$ with high probability, which implies that almost every $d-$regular graph has a $3/\sqrt{d}$-spectral sketch of constant size, namely the complete graph. Note that according to the Alon-Boppana bound [@alon] it is not possible to approximate $K_n$ by a $d-$regular graph with error less than $2/\sqrt{d}$. Thus, our result can be interpreted as saying that a $d-$regular graph can only be approximated by nearby graphs when the error is substantially below the Alon-Boppana bound. We remark that the proof of [@andoni] was based on showing a $\Omega(1/\epsilon^2)$ bit lower bound for a constant sized graph by reducing to the Gap-Hamming problem in communication complexity, and then concatenating $O(n)$ instances of this problem to obtain an instance of size $n$. This method is thereby inherently incapable of recovering the logarithmic factor that we obtain (in contrast, we work with random regular graphs). Notation and Organization ------------------------- We will denote $\epsilon$-spectral approximation as $$(1-\epsilon)L_G\preceq L_H\preceq (1+\epsilon)L_G$$ and $\epsilon-$cut approximation as $$(1-\epsilon)L_G\preceq_\square L_H\preceq_\square (1+\epsilon)L_G.$$ We will use $\lg$ and $\ln$ to denote binary and natural logarithms. We prove our lower bound for spectral sketching in Section 2, and the lower bound for cut sketching in Section 3. Although the latter is logically stronger than the former, we have chosen to present the spectral case first because it is extremely simple and introduces the conceptual ideas of the proof. Acknowledgments {#acknowledgments .unnumbered} --------------- We would like to thank MSRI and the Simons Institute for the Theory of Computing, where this work was carried out during the “Bridging Discrete and Continuous Optimization” program. Lower Bound for Spectral Sketches ================================= In this section, we prove the following theorem. \[thm:speclb\] For any $\epsilon=\omega(n^{-1/4})$, any $\epsilon$-spectral sketching scheme ${\mathsf{sk}}$ for graphs with $n$ vertices must use at least $\frac{n\lg n}{500{\epsilon}^2}\cdot (1-o_n(1))$ bits in the worst case. The main ingredient is the following lemma. \[lem:specrigid\] Suppose $G$ and $H$ are simple $d-$regular graphs such that $$(1-{\epsilon})L_G\preceq L_H\preceq (1+{\epsilon})L_G.$$ Then $G$ and $H$ must have at least $\frac{dn}{2}(1-\epsilon^2d/2)$ edges in common. Since $G$ and $H$ are $d-$regular, we have $L_H-L_G = (dI-A_H)-(dI-A_G)=A_G-A_H$ and $\|L_G\|\le 2d$. Thus the hypothesis implies that: $$-2d\epsilon I\preceq -\epsilon\cdot L_G\preceq L_H-L_G = A_G-A_H\preceq \epsilon\cdot L_G\preceq 2d\epsilon I,$$ which means $\|A_G-A_H\|\le 2\epsilon d$. Passing to the Frobenius norm, we find that $$\sum_{ij}(A_G(i,j)-A_H(i,j))^2=\|A_G-A_H\|_F^2\le n\|A_G-A_H\|^2\le 4\epsilon^2d^2 n.$$ The matrix $A_G-A_H$ has entries in $\{-1,0,1\}$, with exactly two nonzero entries for every edge in $E(G)\Delta E(H)$, so the left hand side is equal to $$2|E(G)\Delta E(H)| = 2(|E(G)|+|E(H)|-2|E(G)\cap E(H)|) = 2(dn-2|E(G)\cap E(H)|),$$ since $|E(G)|=|E(H)|=dn/2$. Rearranging yields the desired claim. Note that the above lemma is vacuous for $\epsilon \ge\sqrt{2/d}$ but indicates that $G$ and $H$ must share a fraction of their edges for $\epsilon$ below this threshold. \[lem:sketchcount\] For any function $f:{\mathbb{R}}^n\rightarrow {\mathbb{R}}$ and $\epsilon<1/2$: $$\lg |\{G\in G_{n,d}:\textrm{$G$ is $\epsilon-$spectrally approximated by $f$}\}|\le dn/2+5\epsilon^2d^2n\lg n,$$ where $G_{n,d}$ denotes the set of all $d-$regular graphs on $n$ vertices. If $f$ does not $\epsilon-$approximate any $d-$regular graph then we are done. Otherwise, let $H$ be the lexicographically (in some pre-determined ordering on $d-$regular graphs) first graph which $f$ $\epsilon$-approximates. Suppose $G$ is another graph that $f$ $\epsilon-$approximates. Notice that by applying twice, we have that for every $x\in{\mathbb{R}}^n$: $$\frac{1-\epsilon}{1+\epsilon}\le\frac{x^TL_Hx}{x^TL_Gx} = \frac{f(x)}{x^TL_Gx}\cdot \frac{x^TL_Hx}{f(x)}\le\frac{1+\epsilon}{1-\epsilon},$$ so $G$ $3\epsilon$-spectrally approximates $H$. By Lemma \[lem:specrigid\], $H$ and $G$ must share $$k:=\frac{dn}{2}(1-9\epsilon^2d/2)$$ edges. Thus, $G$ can be encoded by specifying: 1. Which edges of $H$ occur in $G$. This is a subset of the edges of $H$, which requires at most $dn/2$ bits. 2. The remaining $dn/2-k = 9\epsilon^2d^2n/4$ edges of $G$. Each edge requires at most $2\lg(n)$ bits to specify, so the number of bits needed is at most $18\epsilon^2d^2n\lg n/4$. Thus, the total number of bits required is at most $${dn/2+18\epsilon^2d^2n\lg n/4},$$ as desired. [([@wormald][Cor. 2.4]{})]{}\[thm:mckay\] For $d=o(\sqrt{n})$, the number of $d-$regular graphs on $n$ vertices is: $$\frac{(dn)!}{(dn/2)!2^{dn/2}(d!)^n}\exp\left(\frac{1-d^2}{4}-\frac{d^3}{12n}+O\left(\frac{d^2}{n}\right)\right) \\\ge \exp(dn\ln(n/d)/2\cdot(1-o_n(1))).$$ Let $N$ be the number of distinct sketches produced by ${\mathsf{sk}}$ and let $d=\lceil \frac{1}{25{\epsilon}^2}\rceil$. By Lemma \[lem:sketchcount\], the binary logarithm of the number of $d-$regular graphs $\epsilon$-spectrally approximated by any single sketch $f\in\mathrm{range}({\mathsf{sk}})$ is at most $${dn/2+dn\lg n/5}\le dn\lg n/5\cdot (1+o(1)).$$ On the other hand by Theorem \[thm:mckay\], since $d=o(\sqrt{n})$ we have $$\label{eqn:manygraphs}\lg|G_{n,d}|\ge dn\lg(n/d)/2\cdot(1-o(1))\ge dn\lg n/4\cdot (1-o(1)).$$ Since every $d-$regular graph receives a sketch, we must have $$(1+o(1))\lg N \ge dn\lg n/4-dn\lg n/5 = dn\lg n/20=n\lg n/500\epsilon^2,$$ as desired. The proof above actually shows that any $1/5\sqrt{d}$-spectral sketching scheme for $d-$regular graphs on $n$ vertices must use at least $\Omega(dn\lg n)$ bits [*on average*]{}, since the same proof goes through if we only insist that the sketches work for [*most*]{} graphs. The result of [@bss] produces $\epsilon$-spectral sparsifiers with $16n/\epsilon^2$ edges, which yield $\epsilon$-spectral sketches with $64n\lg n/\epsilon^2$ bits by discretizing the edge weights up to $1/n^2$ error, so the bound above is tight up to a factor of $64\cdot 500$. We have not made any attempt to optimize the constants. Lower Bound for Cut Sketches ============================ In this section we prove Theorem \[thm:cutlb\]. The new ingredient is a rigidity lemma for cuts, which may be seen as a discrete analogue of Lemma \[lem:specrigid\]. The lemma holds for bipartite graphs and is proven using a Goemans-Williamson [@goemans] style rounding argument. \[lem:cutrigid\] Suppose $G$ and $H$ are simple $d-$regular bipartite graphs with the same bipartition $L\cup R$, such that $$\label{eqn:cutappx} (1-{\epsilon})L_G\preceq_\square L_H\preceq_\square (1+{\epsilon})L_G.$$ Then $G$ and $H$ must have at least $\frac{dn}{2}(1-3\sqrt{d}\epsilon)$ edges in common. We will show the contrapositive. Assume $$|E(G)\setminus E(H)|=|E(H)\setminus E(G)|=\delta dn/2$$ for some $\delta\ge 3\sqrt{d}\epsilon$. To show that does not hold, it is sufficient to exhibit a vector $x\in\{-1,1\}^n$ such that $$x^TMx:=x^T(L_G - L_H)x > 2\epsilon dn \ge \epsilon \cdot x^T L_G x,$$ where $M=L_G-L_H=A_H-A_G$, since both graphs are $d-$regular and the latter inequality follows from $\|L_G\|\le 2d$. To find such an $x$ we will first construct $n$ vectors $y_1,\ldots,y_n\in{\mathbb{R}}^n$ such that $$\label{eqn:yiyj} \sum_{i,j = 1}^n M_{ij} \langle y_i, y_j \rangle > \pi \epsilon dn$$ and then use hyperplane rounding to find scalars $x_1,\ldots,x_n\in\{-1,1\}$ which satisfy: $$\label{eqn:xixj} \sum_{i,j =1}^n M_{ij} x_ix_j>2\epsilon dn.$$ Let $z_1, \ldots, z_n\in{\mathbb{R}}^n$ be the columns of $$I + \frac{M}{\sqrt{d}},$$ and note that $\|z_i\|^2\in [1,3]$ since $z_i(i)^2=1$ and every vertex is incident with at most $d$ edges in each of $E(H)\setminus E(G)$ and $E(G)\setminus E(H)$. The relevant inner products of the $z_i$ are easy to understand: $$M_{ij}\langle z_i,z_j\rangle = \begin{cases} 0 & \textrm{if $M_{ij}=0$} \\ \frac{2}{\sqrt{d}}=1\cdot(1\cdot \frac1{\sqrt{d}}+1\cdot \frac{1}{\sqrt{d}}) & \textrm{if }ij\in E(H)\setminus E(G)\\ \frac{2}{\sqrt{d}}=(-1)\cdot(1\cdot \frac{-1}{\sqrt{d}}+1\cdot \frac{-1}{\sqrt{d}}) & \textrm{if } ij\in E(G)\setminus E(H), \end{cases}$$ where in the latter two cases we have used the fact that $i$ and $j$ cannot have any common neighbors because they lie on different sides of the bipartition. Letting $y_i=z_i/\|z_i\|$, we therefore have $$\label{eqn:2eps} \sum_{i,j = 1}^n M_{ij} \langle y_i, y_j \rangle \geq 2\cdot \delta dn\cdot \frac{2}{3\sqrt{d}}>\pi\epsilon dn$$ by our choice of $\delta$. Let $w$ be a random unit vector and let $$x_i = \begin{cases} +1 &\textrm{if }\langle y_i,w\rangle \ge 0\\ -1 &\textrm{if }\langle y_i,w\rangle < 0.\end{cases}$$ We denote the angle between vectors $y_i, y_j \in Y$ as $\theta_{ij}$. We recall that $\theta_{ij} = \cos^{-1} \langle y_i, y_j \rangle$ since $y_i$ and $y_j$ have unit length. It follows that the probability that $x_i \not = x_j$ is equal to $\frac{\theta_{ij}}{\pi}$ and the probability of $x_i = x_j$ is equal to $(1 - \frac{\theta_{ij}}{\pi})$. Thus, $$\begin{aligned} {\mathbb{E}}\left[\sum_{i,j = 1}^{n} M_{ij} x_i x_j\right] &= \sum_{i,j = 1}^{n} M_{ij} {\mathbb{E}}[x_i x_j] \\ &= \sum_{i,j = 1}^{n} M_{ij} \left(\frac{\theta_{ij}}{\pi} (-1) + (1- \frac{\theta_{ij}}{\pi})(1)\right) \\ &= \sum_{i,j = 1}^{n} M_{ij} \left(1- \frac{2\theta_{ij}}{\pi}\right) \\ &= \frac{2}{\pi}\sum_{i,j = 1}^{n} M_{ij} \left(\frac{\pi}{2} - \theta_{ij}\right) \\ &= \frac{2}{\pi}\sum_{i,j = 1}^{n} M_{ij} \left(\frac{\pi}{2} - \cos^{-1} \langle y_i, y_j \rangle\right) \\ &= \frac{2}{\pi}\sum_{i,j = 1}^{n} M_{ij} \sin^{-1} \langle y_i, y_j \rangle\\ &\ge \frac{2}{\pi}\sum_{i,j = 1}^{n} M_{ij} \langle y_i, y_j \rangle\quad\textrm{since $\sin^{-1}(x)\ge x$}\\ &>2\epsilon dn,\end{aligned}$$ by , as desired. The analysis of the rounding scheme above can be improved from $2/\pi$ to the Goemans-Williamson constant $0.868\ldots$, but we have chosen not to do so for simplicity. \[thm:cutlb\] For any $\epsilon=\omega(n^{-1/4})$, any $\epsilon$-cut sketching scheme ${\mathsf{sk}}$ for graphs with $n$ vertices must use at least $\frac{n\lg n}{2304{\epsilon}^2}\cdot (1-o_n(1))$ bits in the worst case. Assume $n$ is divisible by $4$ (add a constant if this is not the case). Let $d=\frac{1}{16\cdot 9\epsilon^2}$ and let $B_{n,d}$ be the set of bipartite graphs on $n$ vertices with respect to a fixed bipartition. We proceed as in the proof of Theorem \[thm:speclb\]. Let $f:\{-1,1\}^n\rightarrow {\mathbb{R}}$ be any function which is an $\epsilon-$cut sketch for some graph $H$. Arguing as in Lemma \[lem:sketchcount\], any other graph $G\in B_{n,d}$ which has the same sketch must $3{\epsilon}$-cut approximate $H$, so by Lemma \[lem:cutrigid\], any such $G$ must have at most $9\sqrt{d}\epsilon\cdot dn/2$ edges which are not present in $H$. Thus, the encoding length of such $H$ is at most $$\ell:=dn/2 + 9d^{3/2}n\epsilon \lg n=dn\lg n/16\cdot(1+o(1))$$ bits, by our choice of $d$, so any particular $f$ can only be an $\epsilon$-cut sketch for $2^\ell$ graphs in $B_{n,d}$ On the other hand, $B_{n,d}$ is quite large. Recall that the [ *bipartite double cover*]{} of a graph $F$ on $n$ vertices is the graph on $2n$ vertices obtained by taking its tensor product with $K_2$, and two distinct graphs must have distinct double covers. Thus, by we have $$\lg |B_{n,d}|\ge \lg |G_{n/2,d}|\ge dn\lg n/8\cdot (1-o(1)).$$ Thus, if $N$ is the number of distinct sketches produced by ${\mathsf{sk}}$, we must have $$(1+o(1))\lg N \ge dn\lg n/8 - dn\lg n/16 = dn\lg n/16 = n\lg n/(16)^2\cdot 9\epsilon^2,$$ as desired. [^1]: [ccarlsn2@illinois.edu.]{} U.I.U.C. [^2]: [akolla@illinois.edu.]{} U.I.U.C. Supported by NSF Grant 1423452. [^3]: [nikhil@math.berkeley.edu.]{} [U.C. Berkeley]{}. Supported by NSF grant CCF-1553751 and a Sloan research fellowship. [^4]: [luca@berkeley.edu.]{} U.C. Berkeley. Supported by the National Science Foundation under Grants No. 1540685 and No. 1655215. [^5]: After discretizing the weights to, say, $1/n^4$ precision, which introduces negligible error. [^6]: We will not be concerned with the details of these procedures since we are only interested in the space used by the sketches and not computational parameters such as query time / success probability of randomized schemes.

--- author: - 'Shin-Ichi [Ikeda]{}, Naoki [Shirakawa]{}, Hiroshi [Bando]{} and Youiti [Ootuka]{}$^{1}$' title: | Orbital-Degenerate Paramagnetic Metal Sr$_{2}$MoO$_{4}$:\ An Electronic Analogue to Sr$_{2}$RuO$_{4}$ --- The manifestation of spin-triplet superconductivity (SC) in the quasi-two-dimensional Fermi liquid state of Sr$_{2}$RuO$_{4}$ has marked an epoch in the study of SC in strongly-correlated electron systems. [@ishida] One might expect that ferromagnetic spin fluctuations are essential for the origin of the spin-triplet SC, as observed in superfluid (p-wave) $^{3}$He. Rice and Sigrist suggested the p-wave pairing in Sr$_{2}$RuO$_{4}$ immediately after the discovery of SC. [@rice] Their idea was based on the existence of ferromagnetism in SrRuO$_{3}$, [@maeno1; @yoshimura] the three-dimensional (3D) analogue to Sr$_{2}$RuO$_{4}$, and similar values of the Landau parameters to those of Fermi liquid $^{3}$He. Concerning the electronic configuration in Sr$_{2}$RuO$_{4}$, the formal valence number of Ru ion is 4$+$ with four 4d electrons in three $t_{2g}$ orbitals (4d$_{xy}$, 4d$_{yz}$, 4d$_{zx}$). The strong crystalline field for 4d electrons causes the Ru ion to have a low spin state with spin degree of freedom S$=$1, which means triple degeneracy and ferromagnetic correlations due to the Hund coupling. The normal state properties of Sr$_{2}$RuO$_{4}$ are well described quantitatively by the Fermi surface parameters obtained from the quantum oscillation measurements. [@andy] They confirmed that two cylindrical sheets of 4d$_{yz}$ and 4d$_{zx}$ orbitals and a cylindrical sheet of 4d$_{xy}$ orbitals consist of four 4d electrons of Ru$^{4+}$, which had been derived from the band-structure calculation. [@oguchi] A recent inelastic neutron scattering measurement on Sr$_{2}$RuO$_{4}$ has shown that incommensurate peaks of dynamical magnetic susceptibility $\chi (\mib{Q})$ with $\mib{q}_0=(\pm 0.6\pi/a, \pm 0.6\pi/a, 0)$ exist instead of a ferromagnetic peak ($\mib{Q}\approx 0$). [@sidis] The incommensurate peaks are ascribed to nesting vectors between Fermi surfaces with one-dimensional (1D) 4d$_{yz}$ and 4d$_{zx}$ orbitals. [@mazin] This provides a new physical aspect, the relation between the spin-triplet SC and the incommensurate spin fluctuations. If the SC mainly corresponds to this incommensurate spin fluctuation, the 1D 4d$_{yz}$ and 4d$_{zx}$ orbitals may play a vital role in the spin-triplet SC of Sr$_{2}$RuO$_{4}$. On the other hand, if the incommensurate spin fluctuation contributes negligibly to the spin-triplet SC, the 2D 4d$_{xy}$ orbital is more important. Since the partial density of states for the 4d$_{xy}$ orbital is the highest among three t$_{2g}$ orbitals, [@andy] most theoretical studies on the spin-triplet SC are based on this 2D 4d$_{xy}$ orbital. [@ogata] Recently, Takimoto has indicated the possibility of spin-triplet pairing symmetry in Sr$_{2}$RuO$_{4}$ [@takimoto] originating from the orbital fluctuations together with the incommensurate spin fluctuation observed by the inelastic neutron scattering. [@sidis] The orbital fluctuations are supposed to be driven by the on-site Coulomb interaction between electrons in different 4d-t$_{2g}$ orbitals, which are almost degenerate. At present, the important problem is whether or not the incommensurate spin fluctuation can induce the spin-triplet pairing. In addition, the role of the orbital degeneracy must be understood to clarify the mechanism of the spin-triplet SC in Sr$_{2}$RuO$_{4}$. An alternative route to understand the spin-triplet SC is provided by studying other compounds related to Sr$_{2}$RuO$_{4}$. Several metallic ruthenates were expected to reveal SC and were intensively examined. [@ikeda; @nakatsuji; @nakatsuji2; @cao] However, such ruthenates have not shown SC, unlike cuprates, where many charge-transfer-type insulators can be made superconductive by ionic substitutions. Hence, it is intriguing to find a ruthenium-free electronic analogue to Sr$_{2}$RuO$_{4}$. In order to provide a new system to be compared with Sr$_{2}$RuO$_{4}$, we have focused on Sr$_{2}$MoO$_{4}$, because it is isostructural to Sr$_{2}$RuO$_{4}$ and has a similar electronic configuration. Mo ions in Sr$_{2}$MoO$_{4}$ formally possess the valence 4$+$ and have two 4d electrons in three $t_{2g}$ orbitals (4d$_{xy}$, 4d$_{yz}$, 4d$_{zx}$) with an equivalent orbital degeneracy (S$=$1) to Sr$_{2}$RuO$_{4}$. Moreover, Sr$_{2}$MoO$_{4}$ is expected to have three Fermi surfaces with the same topology as those of Sr$_{2}$RuO$_{4}$. [@hase] These Fermi surfaces for the 1D orbitals with smaller Fermi wave vectors than those of Sr$_{2}$RuO$_{4}$ may also produce incommensurate spin fluctuations at $\mib{q}$ closer to the ferromagntic wave vector ($\mib{Q}\approx 0$). Therefore, Sr$_{2}$MoO$_{4}$ is a good candidate not only for a new superconductor but also for a counterpart to clarify the mechanism of the spin-triplet SC in Sr$_{2}$RuO$_{4}$. In this Letter, we report the successful synthesis of polycrystalline Sr$_{2}$MoO$_{4}$ and the results of magnetic susceptibility $\chi$, electrical resistivity $\rho$ and specific heat $C$. We have no evidence for SC ascribable to Sr$_{2}$MoO$_{4}$ down to 25mK in the present results. As for the synthesis of Sr$_{2}$MoO$_{4}$, there was a controversy over whether or not the phase actually existed. To our knowledge, Balz and Plieth first reported on Sr$_{2}$MoO$_{4}$, [@balz] but McCarthy and Gooden could not obtain any Sr$_{2}$MoO$_{4}$ phase in their research. [@mccarthy] Lindblom and Rosen reported successful synthesis and deduced the lattice parameters by assuming I4/mmm symmetry. [@lindblom] Very recently, Steiner and Reichelt performed X-ray Rietveld analysis on a multi-phase sample including Sr$_{2}$MoO$_{4}$. [@steiner] They concluded that the crystal structure is the K$_{2}$NiF$_{4}$ type (Fig. 1), which is the same as that of Sr$_{2}$RuO$_{4}$. [@muller] Nevertheless, systematic investigations on physical properties of Sr$_{2}$MoO$_{4}$ as well as definitive structural refinements have not yet been carried out since the synthesis of single-phase samples is extremely difficult. The procedure for the sample synthesis is explained below. First, we prepared polycrystalline Sr$_{3}$MoO$_{6}$ from mixed raw materials of SrCO$_{3}$ (99.99%) and MoO$_{3}$ (99.99%) by firing in an Ar flow atmosphere at 1273K for 24 h with intermediate grindings. Then, Sr$_{3}$MoO$_{6}$ and Mo metal powder (99.9%) were ground together in a dry N$_{2}$ atmosphere so that the total composition was Sr$_{2}$MoO$_{4}$. The resultant mixture was pressed into pellets and then sintered in a quartz tube under controlled oxygen partial pressure at 1273K to 1473K. Details of the synthesis will be explained elsewhere. [@shirakawa; @shirakawa1] We characterized the obtained samples at room temperature by powder X-ray diffraction with Cu-K$\alpha$ radiation. DC magnetization above 2K was measured using a SQUID magnetometer (Quantum Design, MPMS) and that between 25mK and 2K was measured using a SQUID magnetometer (SHE) installed in a dilution refrigerator. [@ootuka] Data below 2K were collected by moving the sample through the pickup coil under a field of 0.1Oe. Electrical resistivity $\rho (T)$ was measured by the ac method from 80mK to 300K. The samples for $\rho (T)$ and magnetization below 1K were covered with epoxy resin in order to avoid exposure to air. Specific heat measurement between 0.7K and 60K were performed by the quasi-adiabatic heat-pulse method. In Fig. 2, we show the powder X-ray pattern of Sr$_{2}$MoO$_{4}$ along with that for a single phase of Sr$_{2}$RuO$_{4}$. The pattern of Sr$_{2}$MoO$_{4}$ without unidentified peaks indicates that the obtained polycrystal is single-phased, except for a very weak peak of Mo metal. Moreover, these patterns suggest that both compounds share a common crystal structure. Indeed, the structural symmetry of Sr$_{2}$MoO$_{4}$ has been concluded to be I4/mmm from the electron diffraction analysis by Shirakawa [*et al.*]{} [@shirakawa1] The calculated lattice parameters of tetragonal Sr$_{2}$MoO$_{4}$ for the powder X-ray pattern are $a=3.9168(4)$[Å]{} and $c= 12.859(2)$[Å]{} and those of Sr$_{2}$RuO$_{4}$ are $a=3.87073(2)$[Å]{} and $c= 12.7397(1)$[Å]{}. [@chmaissem] The values of Sr$_{2}$MoO$_{4}$ are in good agreement with those of previous work. [@steiner] Both $a$ and $c$ parameters are larger than those of Sr$_{2}$RuO$_{4}$ by about 1%, owing to the larger ionic radius of Mo$^{4+}$ than Ru$^{4+}$. We should note that the current samples still contain very small amounts of Mo metal and Mo$_{2}$C because the condition of the synthesis is highly subtle. [@shirakawa1] Nevertheless, we could synthesize the samples with the amount of those phases less than 1%. Temperature dependence of magnetic susceptibility $\chi(T)=M/H$ is shown in Fig. 3 for both Sr$_{2}$MoO$_{4}$ and Sr$_{2}$RuO$_{4}$ [@maeno2] between 2K and 300K. Sr$_{2}$MoO$_{4}$ reveals enhanced Pauli paramagnetic susceptibility, which is almost temperature independent except for the Curie-like increase with lowering temperature below 20K. This probably corresponds to the existence of a small amount of magnetic impurities. Such Curie-like behavior is also observed in CaVO$_{3}$. [@shirakawa2] The difference in the absolute value between Sr$_{2}$MoO$_{4}$ and Sr$_{2}$RuO$_{4}$ will be discussed later together with that of the electronic specific heat coefficient. No hysteresis is observed between zero-field-cooling (ZFC) and field-cooling (FC) sequences. Hence, a ferromagnetically ordering component or a spin-glass state does not exist. In addition, there is no indication of a long-range magnetic ordering. The temperature dependence of specific heat $C_{P}(T)$ for Sr$_{2}$MoO$_{4}$ is ordinary metallic behavior below 10K with a relatively large Sommerfeld coefficient $\gamma$, as shown in Fig. 4. The temperature dependence of $C_{P}(T)$ almost follows the relation $C_{P}(T)=\gamma T + \beta T^{3}$. The obtained electronic specific heat coefficient $\gamma$ is 12mJ/(K$^{2}$ Mo mol) for Sr$_{2}$MoO$_{4}$. Considering its Pauli paramagnetic susceptibility and normal metal behavior of $C_{P}(T)$, the ground state of Sr$_{2}$MoO$_{4}$ is probably dominated by renormalized quasi-particles, as observed in the Fermi liquid state. We did not observe the relevant anomaly in $C_{P}(T)$ below 10K to the Curie-type increase in $\chi(T)$. This is also consistent with the fact that the increase in $\chi(T)$ is due to a tiny amount of the magnetic impurity. For Sr$_{2}$MoO$_{4}$, we try to evaluate the Wilson ratio, a dimensionless parameter concerned with correlation among electrons. It is necessary to determine the precise contributions of the Landau diamagnetic susceptibility especially when we have small mass enhancement of electrons. Hase has calculated the band structure for Sr$_{2}$MoO$_{4}$ [@hase] based on the atomic coordinates obtained by the prior X-ray Rietveld analysis. [@steiner] According to their results, the density of states $D(E_{\rm F})$ and electronic specific heat coefficient $\gamma_{calc}$ are 2.1 states/eV and 5mJ/(K$^{2}$ Mo mol), respectively. [@hase] The ratio of $\gamma$/$\gamma_{calc}$ should be nearly equal to the mass enhancement $m^{\ast}/m$, thus $m^{\ast}/m \approx \gamma/ \gamma_{calc} \approx 2$. The reduction factor by the Landau diamagnetism $1-(1/3)(m/m^{\ast})^{2}$ is about 0.92 in the present case. Considering this diamagnetism and core diamagnetic susceptibility, we obtain the Pauli contribution $\chi_{\rm Pauli} = (\chi_{\rm obs}-\chi_{\rm core})/0.92 \approx 3.2 \times 10^{-4} $emu/mol, where $\chi_{\rm obs}$ is the observed susceptibility and $\chi_{\rm core} = - 0.95 \times 10^{-4} $emu/mol. [@selwood] This value of $\chi_{\rm Pauli}$ is three times smaller than that of Sr$_{2}$RuO$_{4}$. [@maeno2] We obtained the Wilson ratio $R_{\rm W}=7.3 \times 10^{4} \times \{\chi ({\rm emu/mol})\}\slash \{\gamma ({\rm mJ}/({\rm K}^{2}\,{\rm mol})\} = 1.9$, using the above value of $\chi_{\rm Pauli}$. This ratio is almost the same as that of Sr$_{2}$RuO$_{4}$. [@maeno2] The important point is that the value of $R_{\rm W}$ is considerably larger than unity, which is the value for the non-interacting free-electron system. This reflects the existence of substantial electron-electron correlation in Sr$_{2}$MoO$_{4}$, although it possesses smaller $D(E_{\rm F})$ than that of Sr$_{2}$RuO$_{4}$. The electrical resistivity $\rho(T)$ of Sr$_{2}$MoO$_{4}$ presented in Fig. 5 shows metallic (d$\rho$/d$T \ge$0) behavior for the entire temperature range (80mK $\le T \le$ 300K). The ground state is supposed to be Fermi liquid, as recognized in Sr$_{2}$RuO$_{4}$, [@maeno2] because of the temperature dependence of $C_{P}(T)$ and the Pauli paramagnetic $\chi(T)$. On the basis of the data of $ \rho(T)$ for polycrystals, we cannot draw any conclusions about its absolute value and the precise temperature dependence because of grain-boundary resistance and the mixing of both in-plane and out-of plane behaviors. It would be necessary to grow single-crystalline Sr$_{2}$MoO$_{4}$ to confirm whether the ground state is the Fermi liquid by observing the $T^{2}$-law in resistivity. We have paid a great deal of attention to the behavior of Sr$_{2}$MoO$_{4}$ below 4K to observe SC. In Fig. 6, we show $\rho(T)$ and $\chi(T)$ below 1K. Two drops of $\rho(T)$ at 0.9K and 0.4K are evident. The change at 0.9K is also present in the $\chi(T)$ data with a superconducting response. The volume fraction of this signal is, at most, about 1%. We have concluded that this superconducting signal is in accordance with the existence of Mo metal as a trace amount of impurity in the Sr$_{2}$MoO$_{4}$ sample. On the other hand, the other drop at 0.4K in $\rho(T)$ does not accompany any anomaly in $\chi(T)$. The negative value of $\chi(T)$ above 0.9K is probably due to superconducting molybdenum carbides with $T_{\rm c}=3\mbox{--}4$K, [@shirakawa] which was also found in $\chi(T)$ measured by the MPMS SQUID magnetometer. These results indicate that polycrystalline Sr$_{2}$MoO$_{4}$ is not superconducting above 25mK. Now we suggest the reasons why the polycrystalline Sr$_{2}$MoO$_{4}$ we obtained does not exhibit SC. The substantial difference between Sr$_{2}$MoO$_{4}$ and Sr$_{2}$RuO$_{4}$ is the density of states $D(E_{\rm F})$. The absolute values of $\chi(T)$ and $C_{P}(T)$, proportional to $D(E_{\rm F})$, of Sr$_{2}$MoO$_{4}$ are both about three times smaller than those of Sr$_{2}$RuO$_{4}$. This is nearly consistent with the preliminary results of band-structure calculations for Sr$_{2}$MoO$_{4}$. [@hase] The origin of the difference of $D(E_{\rm F})$ between Sr$_{2}$MoO$_{4}$ and Sr$_{2}$RuO$_{4}$ is proposed below. The ratio $d_{c}/d_{a}$, where $d_{c}$ ($d_{a}$) is the distance between the transition-metal ion and the neighboring oxygen ion along the $c$-axis ($a$-axis), plays an essential role in energy-level splitting of the three $t_{2g}$ orbitals. According to Steiner and Reichelt, [@steiner] the MoO$_{6}$ octahedron is more elongated along the $c$-axis than the RuO$_{6}$ octahedron in Sr$_{2}$RuO$_{4}$. The ratio $d_{c}/d_{a}$ for Sr$_{2}$MoO$_{4}$ is 1.13 [@steiner] while $d_{c}/d_{a}$ for Sr$_{2}$RuO$_{4}$ is 1.07. [@chmaissem] The larger $d_{c}/d_{a}$ will lead to the decreased population of 4d electrons in the 2D-4d$_{xy}$ orbital ($\gamma$ Fermi surface described in ref. 5). In other words, more dispersion along the $c$-axis is introduced and smaller $D(E_{\rm F})$ is expected. This smaller $D(E_{\rm F})$ is probably the main reason why Sr$_{2}$MoO$_{4}$ is not superconductive. Another important feature of strontium molybdates is paramagnetism in SrMoO$_{3}$ with a similar value of $\chi$ to that of Sr$_{2}$MoO$_{4}$. [@bouchard; @ikeda2] This shows a great contrast to the ferromagnetism in SrRuO$_{3}$, which satisfies the Stoner criterion with the largest $D(E_{\rm F})$ in its paramagnetic state among ruthenates. [@santi] The absence of significant ferromagnetic spin fluctuation in SrMoO$_{3}$ may be related to the absence of SC in Sr$_{2}$MoO$_{4}$. Despite our argument over the smaller $D(E_{\rm F})$ above, further investigations on Sr$_{2}$MoO$_{4}$, such as single-crystal studies are highly anticipated, since Sr$_{2}$MoO$_{4}$ is the best system to verify the importance of the orbital degeneracy or multiple Fermi surfaces to the spin-triplet SC in Sr$_{2}$RuO$_{4}$. In summary, we have succeeded in synthesizing single-phase polycrystalline Sr$_{2}$MoO$_{4}$. The first systematic studies of magnetic susceptibility, electrical resistivity and specific heat measurements have been performed down to low temperatures around $10^{-2}$K. We did not observe any signs of superconductivity intrinsic to Sr$_{2}$MoO$_{4}$. The value of the Wilson ratio shows remarkable electron-electron correlations. The authors are grateful to I. Hase, T. Yanagisawa, T. Takimoto, Y. Yamaguchi, C. H. Lee and M. Koyanagi for their helpful supports and fruitful advice. [99]{} K. Ishida, H. Mukuda, Y. Kitaoka, K. Asayama, Z. Q. Mao, Y. Mori and Y. Maeno: Nature (London) [**396**]{} (1998) 658. T. M. Rice and M. Sigrist: J. Phys. Condens. Matter [**7**]{} (1995) L643. Y. Maeno, S. Nakatsuji and S. Ikeda: Materials Science and Engineering B [**63**]{} (1999) 70. K. Yoshimura, T. Imai, T. Kiyama, K. R. Thurber, A. W. Hunt and K. Kosuge: Phys. Rev. Lett. [**83**]{} (1999) 4397. A. P. Mackenzie, S. R. Julian, A. J. Diver, G. J. McMullan, M. P. Ray, G. G. Lonzarich, Y. Maeno, S. Nishizaki and T. Fujita: Phys. Rev. Lett. [**76**]{} (1996) 3786. T. Oguchi: Phys. Rev. B [**51**]{} (1995) 1385. Y. Sidis, M. Braden, P. Bourges, B. Hennion, S. NishiZaki, Y. Maeno and Y. Mori: Phys. Rev. Lett. [**83**]{} (1999) 3320. I. I. Mazin and D. J. Singh: Phys. Rev. Lett. [**82**]{} (1999) 4324. T. Kuwabara and M. Ogata: to be published in Phys. Rev. Lett. T. Takimoto: unpublished. S. I. Ikeda, Y. Maeno, S. Nakatsuji, M. Kosaka and Y. Uwatoko: Phys. Rev. B [**62**]{} (2000) R6089. S. Nakatsuji, S. Ikeda and Y. Maeno: J. Phys. Soc. Jpn. (1997) 1868. S. Nakatsuji and Y. Maeno: Phys. Rev. Lett. (2000) 2666. G. Cao, S. McCall, J. E. Crow and R. P. Guertin: Phys. Rev. Lett. [**78**]{} (1997) 1751. I. Hase: unpublished. V. D. Balz and K. Plieth: Z. Elektrochem. [**59**]{} (1955) 545. G. J. McCarthy and C. E. Gooden: J. Inorg. Nucl. Chem. (1973) 2669. B. Lindblom and E. Rosen: Acta Chemica Scandinavica A [**40**]{} (1986) 452. U. Steiner and W. Reichelt: Z. Naturforsch. [**53 b**]{} (1998) 110. Hk. Müller-Buschbaum and J. Wilkens: Z. Anorg. Allg. Chem. [**591**]{} (1990) 161. N. Shirakawa and S. I. Ikeda: to be published in Physica C. N. Shirakawa, S. I. Ikeda and H. Matsuhata: unpublished. Y. Ootuka and N. Matsunaga: J. Phys. Soc. Jpn. [**59**]{} (1990) 1801. O. Chmaissem, J. D. Jorgensen, H. Shaked, S. Ikeda and Y. Maeno: Phys. Rev. B [**57**]{} (1998) 5067. Y. Maeno, K. Yoshida, H. Hashimoto, S. Nishizaki, S. Ikeda, M. Nohara, T. Fujita, A. P. Mackenzie, N. E. Hussey, J. G. Bednorz and F. Lichtenberg: J. Phys. Soc. Jpn. [**66**]{} (1997) 1405. N. Shirakawa, K. Murata, H. Makino, F. Iga and Y. Nishihara: J. Phys. Soc. Jpn. [**64**]{} (1995) 4824. P. W. Selwood: Magnetochemistry (Interscience Publishers, New York, 1956) 2nd ed., p. 78. G. H. Bouchard, Jr. and M. J. Sienko: Inorg. Chem. (1968) 441. S. I. Ikeda and N. Shirakawa: to be published in Physica C. G. Santi and T. Jarlborg: J. Phys.: Condens. Matter [**9**]{} (1997) 9563.

--- abstract: 'The three-state Potts model is numerically investigated on three-dimensional simple cubic lattices of up to $128^3$ volume, concentrating on the neighborhood of the first-order phase transition separating the ordered and disordered phases. In both phases clusters of like spins are observed with irregular boundaries. In the ordered phase the two different non-favored spins are found to attract each other with a long but finite range. As a result, the neighborhoods of the non-favored spins are interpreted as domains of the disordered phase. This explains why the first-order phase transitions associated with the global $Z_3$ symmetry, including the SU(3) pure-gauge one, are so weak.' --- [Phases of the three-state Potts model in three spatial dimensions]{}\ Shigemi Ohta\ The Institute of Physical and Chemical Research (RIKEN)\ Wako-shi, Saitama 351-01, Japan Finite-temperature systems with global $Z_3$ symmetry in three spatial dimensions have in common a weak first-order phase transition separating ordered and disordered phases. Lattice numerical calculations of the $SU(3)$ pure-gauge quantum chromodynamics (QCD) at finite temperature [@ref:puregauge] revealed that a first-order phase transition separates a low-temperature color-confining phase and a high-temperature non-confining phase. The confining phase is disordered with the Polyakov line order parameter taking zero expectation value. The non-confining phase is ordered and the Polyakov line takes a finite expectation value falling on one of the three $Z_3$ axes in the complex plane. The transition is considered weak because the latent heat is small and the Polyakov line correlation length grows. Numerical simulations of the three-state Potts model [@ref:Potts] showed there is a first-order phase transition separating a low-temperature ordered phase and a high-temperature disordered one. This transition is weak in the sense the latent heat is small. However, it is not yet understood why these phase transitions are so weak. This letter reports the author’s investigation of this problem. We consider the simplest form of the three-state Potts model on three-dimensional simple cubic lattices. At each site, $x$, of the lattice, a spin, $\sigma_x$, is defined. It takes one of the three possible states in the $Z_3$ group, $\sigma_x \in \{0, 1, 2\}$. The thermodynamics of the system is described by the partition function, $$Z = \exp(-H/T) = \exp(J \sum_{\langle xy \rangle} \delta(\sigma_x, \sigma_y)),$$ where the sum runs the nearest-neighbor pairs $\langle xy \rangle$ and Kronecker’s symbol is defined as $\delta(\sigma_x, \sigma_y) = 1$ if $\sigma_x = \sigma_y$ and 0 otherwise. The model is invariant under global $Z_3$ transformations. Here we consider only the “ferromagnetic” couplings, $J > 0$. The effective interactions among the Polyakov lines in the finite-temperature pure-gauge QCD are of this type [@ref:FOU]. Earlier numerical simulations showed that a weak first-order phase transition that separates a low-temperature ordered phase and a high-temperature disordered phase at $J \simeq 0.5505$. The phase transition is considered weak because the dimensionless internal energy density, $e \equiv \langle \delta (\sigma_x, \sigma_y) \rangle$, shows only a small gap at the phase transition, from about 0.58 to 0.53, compared with the maximum possible gap from 1 to 1/3. See Fig. \[fig:bothhist\] for distribution of this quantity. These results were, however, obtained on relatively small volumes such as $32^3$. The two broad peaks marked as “32” in Fig.\[fig:bothhist\] shows the distribution from such a small-volume simulation reproduced by the author. Although the two-peak structure is clear enough for establishing the first-order nature of this phase transition, the significant overlapping of the broad tails in the middle reflects the fact that flip-flop transitions between the two phases occur too frequently and it is hard to tell in which of the two phases the system resides at a given instant. Because of this difficulty, the natures of the individual phases could not be studied, nor the reason why the phase transition is so weak. To avoid this difficulty one has to simulate the system on larger volumes. The author wrote a heatbath simulation code for the model that runs on an experimental parallel computer, AP1000, built by Fujitsu Laboratory [@ref:AP1000]. On the largest existing configuration of the computer with $16 \times 32 = 512$ microcomputer “cells,” the code performs one million heatbath updates of a $128^3$ volume in about 32 hours. Typically a couple of million heatbath updates are made at four values of the coupling, $J = 0.55025$, 0.5505, 0.55075, and 0.551. At each of these coupling values, two independent simulations were made; one starting from a completely ordered configuration and the other starting from a completely disordered one. At the highest temperature, $J = 0.55025$, the ordered-start simulation quickly converged to the disordered phase within 100K heatbath updates, while the disordered-start simulation remained in the disordered phase indefinitely through the duration of the simulation for more than one million updates. In contrast, at the lowest temperature, $J = 0.551$, the disordered-start simulation quickly converged to the ordered phase while the ordered-start one remained in the ordered phase indefinitely. Coexistence of the two phases is observed at both of the intermediate temperatures, $J = 0.5505$ and 0.55075: The ordered-start simulations remained in the ordered phase and the disordered-start ones remained in the disordered phase through their durations of more than two million heatbath updates. The corresponding distribution of the internal energy density is shown in Fig. \[fig:bothhist\], marked as “128.” The two peaks are now completely separated, and there is no doubt about in which of the phases the system resides: We are now able to investigate the natures of the individual phases. The two-point correlation function, $C_{ij} (r)$, is defined as the probability to find a pair of spin $i$ and $j$ separated by the distance $r$. They are calculated once in every 100K heatbath updates, and are proven free of autocorrelation at least in the relevant range of $r \le 8$. In the disordered phase the correlations are expected to decay exponentially in Yukawa form $$C_{ij} (r) \rightarrow \alpha_{ij} \frac{\exp(-m_{ij}r)}{r} + p_i p_j,$$ as the distance $r$ increases toward infinity [@ref:Hintermann]. The constant term $p_i p_j$ is given by the product of probabilities, $p_i$ and $p_j$, to find a site with spin $i$ and $j$ respectively. In the disordered phase they should all approach 1/3 as the system size increases, because the $Z_3$ symmetry is preserved. For the same reason, there are only two different sets of correlations: the diagonal ones, $C_{00} = C_{11} = C_{22}$, and the off-diagonal ones, $C_{01} = C_{02} = C_{12}$. Moreover, because of the conservation of probability, $$C_{i0} (r) + C_{i1} (r) + C_{i2} (r) = 1, \label{eqn:sumrule}$$ for any $i$, all the correlation mass must agree with each other and the amplitudes obey a simple relation $\alpha_{00} = \alpha_{11} = \alpha_{22} = -2 \alpha_{01} = -2 \alpha_{02} = -2 \alpha_{12}$. We numerically confirmed that all the correlations are fitted well by the Yukawa form, as can be seen in Fig. \[fig:J5505d\_corr\] where the correlations are shown with the constant terms subtracted. Within statistical errors, the fitting parameters (see Table \[tab:J5505d\]) satisfy the above requirements. The positive $(ij)\protect$ ------------------ --------- ----------- (00), (11), (22) 0.15(1) 0.071(3) (01), (02), (12) 0.16(1) -0.037(1) : Two-point correlations in the disordered phase at $J = 0.5505\protect$. Two-parameter least-$\chi^2\protect$ fit to the Yukawa form for the range $r \ge 3\protect$.[]{data-label="tab:J5505d"} diagonal amplitudes, $\alpha_{ii} > 0$, shows the like spins attract each other, in accordance with the ferromagnetic interaction. The corresponding correlation mass of $m_{ii} \simeq 0.15$ suggests there exist clusters of like spins of the size of several lattice spacings. Indeed in the spin distributions one sees many such clusters. It is interesting to note that these clusters are neither smooth nor convex in shape, but are complex and concave. In the ordered phase there is one favored spin which we label as 0, and two non-favored ones which we label as 1 and 2. No analytic result is known about the behavior of the correlations. The current numerical results show the Yukawa form fits them well also in this case. See Fig. \[fig:J5505o\_dia\] and \[fig:J5505o\_off\], where the constant terms are subtracted again. The fitting parameters are given in Table \[tab:J5505o\]. $(ij)\protect$ ---------------- --------- ----------- (00) 0.14(2) 0.081(3) (11), (22) 0.18(3) 0.028(4) (01), (02) 0.15(1) -0.041(1) (12) 0.03(3) 0.010(1) : Two-point correlations in the ordered phase at $J = 0.5505\protect$. Two-parameter least-$\chi^2\protect$ fit to the Yukawa form for the range $r \ge 3\protect$. The favored spin is 0.[]{data-label="tab:J5505o"} The positive amplitudes of the diagonal correlations, $\alpha_{ii} > 0$, are in accordance with the ferromagnetic interaction. Smaller correlation mass for the favored spin, 0, compared with those of non-favored spins, 1 and 2, may not be statistically significant, but nonetheless is consistent with the fact that the favored spin dominates the volume in the ordered phase. At this coupling, $J = 0.5505$, about 60 % of the volume is covered by the favored spin, while the remaining is evenly split between the two non-favored ones. On the other hand the negative amplitudes of the correlations between the favored spin and either of the two non-favored spins, $\alpha_{01} < 0$ and $\alpha_{02} < 0$, means the favored spin repel the non-favored spins. These suggest that there form clusters of like spins, which are indeed observed in the spin distributions. Again the clusters have irregular boundaries [@ref:ohta92]. The correlation between the two different non-favored spins, $C_{12} (r)$, is the most interesting. It starts from zero at the origin as it should be, but then overshoots the asymptotic value $p_1 p_2$, and approaches it from above (see Fig. \[fig:J5505o\_off\].) This means an attractive force acts between the two different non-favored spins. The attractive part alone can be fitted by the Yukawa form if we limit the range of fitting to $r \ge 3$ (Fig.\[fig:J5505o\_off\].) Note that the correlation mass, $m_{12} = 0.03 \pm 0.03$, is much smaller than those for the other cases, the smallest of which is $m_{00} = 0.14 \pm 0.02$. It is not appropriate, however, to consider this as a truly long-range correlation with zero correlation mass. Again because of the conservation of probability (\[eqn:sumrule\]) the correlation mass $m_{12}$ in the long-distance limit $r \rightarrow \infty$ must be equal to the smaller of the masses $m_{01}$ and $m_{11}$. Much smaller value of $m_{12}$ we are seeing must come from subtle balance of the two terms $C_{01} (r)$ and $C_{11} (r)$ in the range we are looking at, and has to give way to the true correlation mass at some longer range. For this reason, we will call this attractive part as the “middle-range” attraction in the remainder of this letter. Yet it is useful to explore what would happen if we had a really long-range attraction between the two different non-favored spins: such an attraction is likely to cause instability of the ordered phase, because the repulsion against the non-favored spins from the favored ones can then be compensated by paring the two different non-favored spins attracting each other. Hence we expect that the smaller the correlation mass $m_{12}$, the weaker the phase transition. Indeed in the spin distribution of this phase one finds many clusters of the non-favored spins. And such a non-favored cluster almost always accompany clusters of the other non-favored spin in its neighborhood. It should be noted that such pairing of the clusters of the two different non-favored spins is a way to maintain the global $Z_3$ symmetry which requires the volumes occupied by the two non-favored spins must be equal. As was mentioned earlier, these clusters of the non-favored spins appear in irregular shapes. The neighborhood of such a cluster of complex-shaped clusters of the non-favored spins can be interpreted as an island of the disordered phase in the sea of the ordered phase. This is clearly seen by looking at the local-averaged internal energy density, $e_i (r)$, which is defined in the sphere of radius $r$ with spin $i$ at its center. See Fig. \[fig:J5505\_ie\] for its dependence on the radius. In the disordered phase, the quantity does not show any dependence and stays at its asymptotic value of $\simeq 0.53$. In the ordered phase, the one around the favored spin approaches the asymptotic value of $\simeq 0.58$ monotonously from above, while the other around the two non-favored spins starts from a value smaller than in the disordered phase and approaches the asymptotic value monotonously from below. These disordered domains increases the internal energy density of the ordered phase and explains why the latent heat is so small. Irregular boundaries of the non-favored clusters and disordered domains in the ordered phase near the transition may be consistent with very small surface tension at the confined-deconfined phase boundaries found by a MIT-bag calculation [@ref:Mardor] as well as by lattice QCD numerical calculations [@ref:BandH]. Note, however, that the irregular boundaries themselves suggests inadequacy of the sharp spherical boundaries assumed or imposed in these calculations. The “middle-range” attractive correlation between the two different non-favored spins remains attractive down to the lowest temperature simulated, $J = 0.551$. Its range, however, decreases to $m_{12} = 0.07(3)$ at $J = 0.55075$ and 0.11(3) at 0.551. This way the ordered phase away from the first-order transition consolidates itself. It would be interesting to check if the above mechanism of weakening the first-order phase transition works in the pure-gauge QCD thermodynamics. As a first step, the author generated several hundred pure-gauge configurations on a $16^3 \times 2$ lattice by the “APE6” computers in Rome, using a hybrid Monte Carlo algorithm written by the APE group [@ref:Donini]. From each gauge configuration a distribution of Polyakov lines is calculated. Near the first-order phase transition, most of the Poliyakov lines are found in the neighborhood of the three $Z_3$ axes with non-zero magnitude, even in the disordered phase. Each Polyakov line is projected onto the nearest of the $Z_3$ axes, resulting in distributions of the Potts-model spins. Correlations of the spins are then calculated, and found to behave in the same manner as their counterparts in the three-state Potts model. Most notably, the correlation between the two different non-favored spins in the ordered phase shows the same kind of “middle-range” attraction. This allows one to speculate that the deconfined phase of QCD near the phase transition is susceptible to mixture of complex-shaped “droplets” of the confined phase [@ref:Mardor]. If such is indeed the case, perturbation calculations of the QCD deconfined phase would be useless near the phase transition. Hence it is important to investigate this point in more detail. The author thanks Parallel Computing Research Facility, Fujitsu Laboratory, for computing time on their AP1000 parallel computer on which much of this work was done. He is also thankful for fruitful discussions with Norman Christ, Enzo Marinari, Giorgio Parisi, and Roberto Petronzio. Thanks are also due to the hospitality extended to him by the APE100 group, especially by Federico Rapuano and Raffaele Tripiccione. Finally, he warmly thanks Andrea Donini and Stefano Antonelli for kindly allowing him to use their hybrid Monte Carlo simulation code for the APE computer. [99]{} F.R. Brown et al., Phys. Rev. Lett. 61, 2058 (1988); F.R. Brown, Nucl. Phys. B (Proc. Suppl.) 17, 214 (1990); S. Cabasino et al., Nucl. Phys. B (Proc. Suppl.) 17, 218 (1990); M. Fukugita, M. Okawa, and A. Ukawa, Nucl. Phys. B337, 181 (1990); Y. Iwasaki et al., Phys. Rev. Lett. 67, 3343 (1991). H.W.J. Blöte and R.H. Swendsen, Phys. Rev.Lett. 43, 799 (1979); S.J. Knak Jensen and O.G. Mouritsen, Phys.Rev. Lett. 43, 1736 (1979); M. Fukugita and M. Okawa, Phys. Rev.Lett. 63, 13 (1989); A.J. Guttmann and I.G. Enting, Nucl. Phys. B (Proc. Suppl.) 17, 328 (1990); R.V. Gavai, ibid., 335. M. Fukugita, M. Okawa, and A. Ukawa, Phys. Rev.Lett. 63, 1768 (1989). S. Ohta, Nucl. Phys. B (Proc. Suppl.) 26, 647 (1992). A. Hintermann, H. Kunz, and F.Y. Wu, J.Stat. Phys. 19, 623 (1978). S. Ohta, in the proceedings of the International Symposium on Lattice Field Theories “Lattice ’92” in Amsterdam, September 1992, Nucl. Phys. B (Proc. Suppl.) to be published. I. Mardor and B. Svetitsy, Phys. Rev. D44, 878 (1991). J. Potvin and C. Rebbi, Nucl. Phys. B (Proc.Suppl.) 17, 317 (1990); K. Rumukainen, Nucl. Phys. B (Proc. Suppl.) 26, 320 (1992); Y. Aoki and K. Kanaya, talk presented at the 48th annual meeting of the Physical Society of Japan,in Sendai, March 1993. A. Donini and S. Antonelli, private communication.

--- abstract: 'The quad-curl term is an essential part in the resistive magnetohydrodynamic (MHD) equation and the fourth order inverse electromagnetic scattering problem which are both of great significance in science and engineering. It is desirable to develop efficient and practical numerical methods for the quad-curl problem. In this paper, we firstly present some new regularity results for the quad-curl problem on Lipschitz polyhedron domains, and then propose a mixed finite element method for solving the quad-curl problem. With a [*novel*]{} discrete Sobolev imbedding inequality for the piecewise polynomials, we obtain stability results and derive [*optimal*]{} error estimates based on a relatively low regularity assumption of the exact solution.' address: - 'School of Mathematical Sciences, University of Electronic Science and Technology of China, Sichuan 611731, China' - 'Department of Mathematics, City University of Hong Kong, 83 Tat Chee Avenue, Hong Kong, China' - 'School of Mathematical Sciences, University of Electronic Science and Technology of China, Sichuan 611731, China' author: - Gang Chen - Weifeng Qiu - Liwei Xu title: 'Analysis of a mixed finite element method for the quad-curl problem' --- Introduction {#intro} ============ Let $\Omega$ be a bounded simply connected Lipschitz polyhedron in $\mathbb{R}^3$ with connected boundary $\partial\Omega$. We consider the following quad-curl (fourth order) problem: find the vector $\bm u$ and the Lagrange mutiplier $p$ such that \[PDE:orignial\] $$\begin{aligned} \nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))+\nabla p&=\bm f&\text{ in }\Omega,\label{o1}\\ \nabla\cdot\bm u&=0&\text{ in }\Omega,\label{o2}\\ \bm n\times\bm u&=\bm 0 &\text{ on }\partial\Omega,\label{o3}\\ \bm n\times(\nabla\times\bm u)&=\bm 0&\text{ on }\partial\Omega,\label{o4}\\ p&=0&\text{ on }\partial\Omega.\label{o5} \end{aligned}$$ Here $\bm f\in [L^2(\Omega )]^3$, the vector $\bm n$ denotes the unit outer normal on $\partial\Omega$. This model problem arises in many different applications, such as in the resistive magnetohydrodynamics (MHD) and in the inverse electromagnetic scattering theory. The resistive MHD system reads ([@MR2813342; @MR3041683]): find the velocity $\bm u$, the pressure $p$ and the magnetic induction field $\bm B$ such that \[SMHD\] $$\begin{aligned} \rho(\bm u_t+(\bm u\cdot\nabla)\bm u)+\nabla p&=\frac{1}{\mu_0}(\nabla\times\bm B)+\mu\Delta\bm u&\text{in }\Omega,\\ \bm B_t-\nabla\times(\bm u\times\bm B)&=-\frac{\eta}{\mu_0}(\nabla\times(\nabla\times\bm B)) \\ &\quad-\frac{d_i}{\mu_0}\nabla\times((\nabla\times\bm B)\times \bm B )\nonumber\\ &\quad-\frac{\eta_2}{\mu_0}\nabla\times( \nabla\times(\nabla\times(\nabla\times\bm B)) )&\text{in }\Omega,\nonumber \\ \nabla\cdot\bm u&=0&\text{in }\Omega,\\ \nabla\cdot\bm B&=0&\text{in }\Omega, \end{aligned}$$ with some proper boundary conditions. Here, $\rho$ is the mass density, $\eta$ is the resistivity, $\eta_2$ is the hyper-resistivity, $\mu_0$ is the magnetic permeability of free space, and $\mu$ is the viscosity. In the inverse electromagnetic scattering theory, the transmission eigenvalue problem for the anisotropic Maxwell equations can be formulated in the following fourth order problem ([@MR2970278]): find the vector $\bm u$ and the number $k$ such that \[inverse\] $$\begin{aligned} (\nabla\times(\nabla\times)-k^2N)(N-I)^{-1} (\nabla\times(\nabla\times\bm u)-k^2\bm u)&=0&\text{in }\Omega,\\ \bm n\times\bm u&=\bm 0&\text{on }\partial\Omega,\\ \bm n\times(\nabla\times\bm u)&=\bm 0&\text{on }\partial\Omega, \end{aligned}$$ where $N$ is a given real matrix field and $I$ is the identity matrix. The leading term in both and is $\nabla\times( \nabla\times( \nabla\times(\nabla\times\bm u)) )$. There are vast literatures on numerical methods solving the MHD model [*without*]{} the quad-curl term $\nabla\times( \nabla\times( \nabla\times(\nabla\times\bm u)) )$, see [@MR1848902; @MR2740762; @MR3599563; @MR2290574; @2019-weifeng; @2017arXiv171111330H; @2018arXiv180101252G; @Xu-submit] and references therein for detailed information. However, when the quad-curl term $\nabla\times( \nabla\times( \nabla\times(\nabla\times\bm u)) )$ is present, the design and analysis of numerical methods for the MHD model becomes more difficult and challenging. Therefore, it is worth devising accurate and efficient numerical methods for the quad-curl problem, providing with substantial tools for the solution of the resistive MHD system and the fourth order inverse electromagnetic scattering problem. It is known that there is a strong correlation between the regularity of exact solutions and the extent of smoothness on the computational domain on which the quad-curl problem is imposed. At the continuous level of differential equations, the author proved in [@MR3760167] that: when the domain has no point and edge singularities, it holds that $\bm u\in [H^4(\Omega)]^3$; when the domain has point or edge singularities, $\bm u$ does [*not*]{} belong to $[H^3(\Omega)]^3$ in general. In [@MR3808156], the author proved that on convex polyhedral domains, if $\nabla\cdot\bm f=0$, there hold $$\begin{aligned} \bm u\in[H^2(\Omega)]^3,\qquad \nabla\times\bm u\in [H^2(\Omega)]^3,\qquad p=0.\end{aligned}$$ These results imply that a reasonable assumption on the regularity of the exact solution of the quad-curl problem, from which the stability and convergence results of numerics are derived, is highly desirable for the purpose of designing practical numerical methods. This is indeed one of our motivations for this work. There are already many works devoted to the numerical study on the quad-curl problem in the past decades. In [@MR2813342], a nonconforming finite element method was studied under the regularity assumption $$\begin{aligned} \bm u \in [H^4(\Omega)]^3.\end{aligned}$$ A discontinuous Galerkin (DG) method using $\bm H(\text{curl})$ conforming elements for the quad- model problem was investigated in [@MR3041683], where the following regularity requirements were assumed: $$\begin{aligned} \label{post-minimal} \bm u \in [H^2(\Omega)]^3,\qquad \nabla\times\bm u\in [H^2(\Omega)]^3.\end{aligned}$$ A mixed finite element method for the quad- eigenvalue problem was introduced and analyzed in [@MR3439219] given that the following set of regularities $$\begin{aligned} \bm u \in [H^3(\Omega)]^3,\qquad\nabla\times\bm u\in [H^3(\Omega)]^3\end{aligned}$$ holds. Instead of solving the quad-curl problem directly, through introducing extra unknowns, a mixed finite element method was proposed and analysed in [@MR3745016 method in (44)] based on a Helmholtz decomposition. The author proved the well-posedness and the stability of the method, and the optimal convergence rate in the energy-norm is also proved on the convex domain. A finite element method for the quad- problem in two dimensions was studied in [@MR3719596] based on the Hodge decomposition. Concerning conforming finite element methods, since the curl-curl conforming elements in three dimensions are still unknown (see [@2018arXiv180502962Z] for curl-curl conforming elements in two dimensions), it would be complicated and far from being obvious (since the conforming elements for the biharmonic problem are quite complicated even in two dimensions, see [@MR543934] for example) if the curl-curl conforming elements are considered. In this paper, we firstly present several regularity results of the quad-curl problem on general Lipschitz domains (might be non-convex) which have not been documented in literatures yet. Then, we introduce a new mixed finite element method solving the quad-curl model problem . Even though our numerical scheme shares some features with that proposed in [@MR3041683], the authors of [@MR3041683] dealt with a quad-curl problem with a reaction term which makes their theoretical analysis different from ours. Finally, we prove the corresponding stability and convergence results of the numerical solution of $\bm u$ under a relatively low regularity compared to that in existing works, i.e. $$\begin{aligned} \bm u\in[H^{r_{u_0}}(\Omega)]^3,\quad \nabla\times\bm u\in [H^{r_{u_1}}(\Omega)]^3,\quad \nabla\times(\nabla\times\bm u)\in[H^{r_{u_2}}(\Omega)]^3,\quad p\in H^{ r_p}(\Omega),\end{aligned}$$ where $r_{u_0}>\frac{1}{2}$, $r_{u_1}\ge 1$, $r_{u_2}>\frac{1}{2}$, and $r_p>\frac{3}{2}$. In addition, we establish a [*novel*]{} discrete Sobolev imbedding inequality in the following piecewise $H^1$ norm (see ): $$\begin{aligned} \sum_{K\in\mathcal{T}_h}\|\bm v_h\|_{1,K}^2\le C\left[ \sum_{K\in\mathcal{T}_h}\left( \|\nabla\times\bm v_h\|^2_{0,K} +\|\nabla\cdot\bm v_h\|^2_{0,K} \right) +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\bm v_h]\!]\|^2_{0,F} \right],\end{aligned}$$ where $\bm v_h$ is a piecewise polynomial of a fixed order. This inequality provides us with a fundamental tool to further achieve the discrete $H^1$ stability and $H^1$ error estimate for the approximation of $\nabla\times\bm u$. Moreover, turning to the inequality, we could apply our mixed method to solve eigenvalue problems and carry out the numerical analysis of nonlinear problems and high order problems in a low regularity region via a posteriori analysis techniques ([@MR2684360; @MR2846773]), and we will consider it in the future works. The rest of this paper is organized as follows. We present some regularity results for the partial differential equations (PDEs) with the quad-curl term in . A mixed method for the quad-curl problem is introduced in . We obtain a novel discrete Sobolev inequality and stability results for the underlying mixed method in . The convergence result is proved through an energy argument in . We give estimates in $\bm H({\rm curl})$ norm through dual arguments in . Regularity for PDEs\[sec:reg\] ============================== For any bounded domain $\Lambda \subset \mathbb{R}^s$ $(s=2,3)$, and any two functions $u,v\in L^2(\Lambda)$, we denote the $L^2(\Lambda)$ inner product and its norm by $$\begin{aligned} (u,v)_{\Lambda}:=\int_{\Lambda}uv\;{ \rm d}x,\qquad \|u\|_{0,\Lambda}:=(u,u)_{\Lambda}^{\frac{1}{2}},\end{aligned}$$ and when $\Lambda=\Omega$, we set $$\begin{aligned} (u,v):=(u,v)_{\Omega},\qquad \|u\|_{0}:=\|u\|_{0,\Omega},\end{aligned}$$ for simplicity. We denote the Sobolev spaces defined on $\Lambda$ by $W^{m,p}(\Lambda)$ and $W_0^{m,p}(\Lambda)$, and denote its semi-norm and norm by $|v|_{m,p,\Lambda}$, $\|v\|_{m,p,\Lambda}$, respectively. When $p=2$ we omit $p$ in $|v|_{m,p,\Lambda}$ and $\|v\|_{m,p,\Lambda}$; when $\Lambda=\Omega$, we omit $\Lambda$ in $|v|_{m,p,\Lambda}$ and $\|v\|_{m,p,\Lambda}$. For conventional notations, we denote $$\begin{aligned} &H^m(\Lambda):= W^{m,2}(\Lambda),\qquad H^m_0(\Lambda):= W_0^{m,2}(\Lambda).\end{aligned}$$ In particular, when $\Lambda\in \mathbb{R}^{2}$, we use $\langle\cdot,\cdot\rangle_{\Lambda}$ to replace $(\cdot,\cdot)_{\Lambda}$ for distinction. The bold face fonts will be used for vector (or tensor) analogues of the Sobolev spaces along with vector-valued (or tensor-valued) functions. We define the following function spaces $$\begin{aligned} \bm{H}(\text{div} ;\Omega)&:=\{\bm{v}\in [L^2(\Omega)]^3: \nabla\cdot\bm{v}\in L^2(\Omega) \},\\ \bm{H}(\text{curl};\Omega)&:=\{\bm{v}\in [L^2(\Omega)]^3: \nabla\times\bm{v}\in [L^2(\Omega)]^3 \},\\ \bm{H}^s(\text{curl};\Omega)&:=\{\bm{v}\in [H^s(\Omega)]^3: \nabla\times\bm{v}\in [H^s(\Omega)]^3 \text{ with }s\ge 0\},\\ \bm H({\rm curl}^2;\Omega) &:=\{ \bm v\in [L^2(\Omega)]^3:\nabla\times\bm v\in \bm{H}(\text{curl};\Omega) \}\end{aligned}$$ equipped with the graph norms $$\begin{aligned} &\|\bm v\|_{\bm H({\rm div};\Omega)}:=\left( \|\bm v\|^2_0+\|\nabla\cdot\bm v\|^2_0\right)^{\frac{1}{2}},\qquad\quad \|\bm v\|_{\bm H({\rm curl};\Omega)}:=\left( \|\bm v\|^2_0+\|\nabla\times\bm v\|^2_0\right)^{\frac{1}{2}},\\ &\|\bm v\|_{\bm H^s({\rm curl};\Omega)}:=\left( \|\bm v\|^2_s+\|\nabla\times\bm v\|^2_s\right)^{\frac{1}{2}},\qquad \|\bm v\|_{\bm H({\rm curl}^2;\Omega)}:=\left( \|\bm v\|^2_0+\|\nabla\times\bm v\|_{\bm H({\rm curl};\Omega)}^2 \right)^{\frac{1}{2}},\end{aligned}$$ respectively. Furthermore, we define $$\begin{aligned} \bm H_0({\rm div};\Omega)&:=\{ \bm v\in\bm H({\rm curl};\Omega):\bm n\cdot\bm v|_{\partial\Omega}= 0 \},\\ \bm H_0({\rm curl};\Omega)&:=\{ \bm v\in\bm H({\rm curl};\Omega):\bm n\times\bm v|_{\partial\Omega}=\bm 0 \},\\ \bm H_0^s({\rm curl};\Omega)&:=\{ \bm v\in\bm H^s({\rm curl};\Omega):\bm n\times\bm v|_{\partial\Omega}=\bm 0 \},\\ \bm H_0({\rm curl}^2;\Omega)&:=\{ \bm v\in\bm H({\rm curl}^2;\Omega):\bm n\times\bm v|_{\partial\Omega} =\bm n\times(\nabla\times\bm v)|_{\partial\Omega} =\bm 0 \}.\end{aligned}$$ Throughout this paper, we use $C$ to denote a positive constant independent of mesh size, not necessarily the same at its each occurrence. The following imbedding theory is standard but useful in the analysis of $\bm H(\text{curl})$ space. \[[ [@MR1626990 Proposition 3.7]]{}\]\[embed\] If the domain $\Omega$ is a Lipschitz polyhedron, then $\bm{X}_T(\Omega)$ and $\bm{X}_N(\Omega)$ are continuously imbedded in $[H^{\alpha}(\Omega)]^3$ for a real number $\alpha\in (\frac{1}{2},1]$, where the spaces $ \bm X_N(\Omega)$ and $ \bm X_T(\Omega)$ are defined as $$\begin{aligned} \bm{X} _N:=\bm{H}_0({\rm curl};\Omega)\cap \bm{H}({\rm div} ;\Omega),\qquad \bm{X} _T:=\bm{H}({\rm curl};\Omega)\cap \bm{H}_0({\rm div} ;\Omega). \end{aligned}$$ We define the weak formulation of as: Find $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ such that \[var\] $$\begin{aligned} (\nabla\times(\nabla\times\bm u),\nabla\times(\nabla\times\bm v))+(\nabla p,\bm v)&=(\bm f,\bm v),\label{var1}\\ (\bm u,\nabla q)&=0\label{var2} \end{aligned}$$ for all $(\bm v,q)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$. \[dense\] The space $[\mathcal C_0^{\infty}(\Omega)]^3$ is dense in the space $\bm H_0({\rm curl}^2;\Omega)$. For any $\bm v\in \bm H_0({\rm curl}^2;\Omega)$, we define $$\tilde{\bm v}=\left\{ \begin{aligned} \bm v(\bm x) & \text{ if }\bm x\in \Omega,\\ \bm 0\quad & \text{ if }\bm x\in \mathbb R^3/\Omega. \end{aligned} \right.$$ Since $[\mathcal C_0^{\infty}(\Omega)]^3$ is dense in $\bm H_0({\rm curl};\Omega)$, we have $\tilde{\bm v}\in \bm H({\rm curl};\mathbb R^3)$, thus $\nabla\times\tilde{\bm v}\in \bm L^2(\mathbb R^3)$ and $$\nabla\times\tilde{\bm v}=\left\{ \begin{aligned} \nabla\times\bm v(\bm x) & \text{ if }\bm x\in \Omega,\\ \bm 0\quad & \text{ if }\bm x\in \mathbb R^3/\Omega. \end{aligned} \right.$$ Again, due to the fact that $[\mathcal C_0^{\infty}(\Omega)]^3$ is dense in $\bm H_0({\rm curl};\Omega)$, we have $\nabla\times\widetilde{\bm v}\in \bm H({\rm curl};\mathbb R^3)$. Then we obtain that the space $[\mathcal C_0^{\infty}(\Omega)]^3$ is dense in space $\bm H_0({\rm curl}^2;\Omega)$ by following the proof of the part (ii) of [@MR1742312 Theorem 3.29]. Looking at the test function as $\mathcal C_{0}^{\infty}(\Omega)$ in , we know that the solution $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ of satisfies in the distribution sense. On the other hand, the solution $(\bm u,p)$ of with the low regularity $\bm u\in [L^2(\Omega)]^3$, $\nabla\times(\nabla\times\bm u)\in [L^2(\Omega)]^3$ and $p\in H_0^1(\Omega)$ satisfies . Now we are ready to prove the following regularity on Lipschitz polyhedron domains for the weak solution $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$. \[th:reg\] Let $\Omega$ be a simply connected Lipschitz domain in $\mathbb R^3$, for any given $\bm f\in [L^2(\Omega)]^3$, the problem has a unique weak solution $(\bm u,p)\in{\bm H}_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$. In addition, the following regularity estimate holds true: $$\begin{aligned} \|\bm u\|_{r_{u_0}} +\|\nabla\times\bm u\|_1 +\|\nabla\times(\nabla\times\bm u)\|_0 +\|\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))\|_{0} +\|\nabla p\|_0 \le C\|\bm f\|_0, \end{aligned}$$ where the regularity index $r_{u_0}\in (\frac{1}{2},1]$. Furthermore, if $\nabla\times(\nabla\times\bm u)\in [H^{r_{u_1-1}}(\Omega)]^3$, we have the following regularity $$\begin{aligned} \| \nabla\times\bm u\|_{r_{u_1}} \le C\|\nabla\times(\nabla\times\bm u)\|_{r_{u_1}-1}, \end{aligned}$$ where $r_{u_1}\in [1,\frac{3}{2})$, and it is close to $\frac{3}{2}$. Since is a linear square system, it is sufficient to prove the uniqueness, which is equivalent to the existence. First, we present our proof of the regularity estimate in following several steps: - Proof of $\|\nabla\times(\nabla\times\bm u)\|_0\le C\|\bm f\|_0$: Taking $\bm v=\bm u$ in and $q=-p$ in and adding them together, we have $$\begin{aligned} \|\nabla\times(\nabla\times\bm u)\|^2_0=(\bm f,\bm u). \end{aligned}$$ Considering the facts that $(\bm u,\nabla q)=0$ for all $q\in H_0^1(\Omega)$, $(\nabla\times\bm u,\nabla q)=0$ for all $q\in H_0^1(\Omega)$, and or [@MR2059447 Corollary 4.8], we obtain $$\begin{aligned} \|\nabla\times(\nabla\times\bm u)\|^2_0 \le \|\bm f\|_0\|\bm u\|_0\le C \|\bm f\|_0 \|\nabla\times \bm u\|_0\le C \|\bm f\|_0\|\nabla\times(\nabla\times\bm u)\|_0, \end{aligned}$$ which leads to $$\begin{aligned} \|\nabla\times(\nabla\times\bm u)\|_0\le C\|\bm f\|_0. \end{aligned}$$ - Proof of $ \|\bm u\|_{r_{u_0}}\le C\|f\|_0$: It can be obtained directly from , and the fact that $ \|\nabla\times(\nabla\times\bm u)\|_0\le C\|\bm f\|_0$. - Proof of $\|\nabla\times\bm u\|_{1} \le C\|\bm f\|_0$: We define $\bm \sigma:=\nabla\times\bm u$. From $\bm n\times\bm u|_{\partial\Omega}=\bm 0$ we have, $$\begin{aligned} \bm n\cdot\bm \sigma|_{\partial\Omega}=\bm n\cdot(\nabla\times\bm u)|_{\partial\Omega}=\nabla_{\partial\Omega}\cdot(\bm n\times\bm u)=0, \end{aligned}$$ which, together with $\bm n\times\bm{\sigma}|_{\partial\Omega}=\bm 0$ in terms of , leads to $$\begin{aligned} \label{ss1} \bm \sigma=\bm 0 \text{ on }\partial\Omega. \end{aligned}$$ Meanwhile, $\nabla\cdot\bm\sigma=0$, gives $$\begin{aligned} \label{ss2} \Delta \bm{\sigma}=\nabla(\nabla\cdot\bm{\sigma} )-\nabla\times(\nabla\times\bm {\sigma})=-\nabla\times(\nabla\times\bm {\sigma})\text{ in }\Omega. \end{aligned}$$ The combination of , , and the regularity for the elliptic problem in [@MR961439] yields $$\begin{aligned} \|\nabla\times\bm u\|_{1}=\|\bm\sigma\|_{1}&\le C\|\nabla\times(\nabla\times\bm \sigma)\|_{-1}\le C\|\nabla\times(\nabla\times\bm u)\|_0\le C\|\bm f\|_0. \end{aligned}$$ - Proof of $\|\nabla\times\bm u\|_{r_{u_1}} \le C\|\nabla\times(\nabla\times(\nabla\times\bm u))\|_{r_{u_1}-2}$: The combination of , and the regularity for the elliptic problem in [@MR1331981 Theorem 1.1], for some constant $r_{u_1}\in [1,\frac{3}{2})$ and it is colse to $\frac{3}{2}$, yields $$\begin{aligned} \|\nabla\times\bm u\|_{r_{u_1}}=\|\bm\sigma\|_{r_{u_1}}&\le C\|\nabla\times(\nabla\times\bm \sigma)\|_{r_{u_1}-2}\nonumber\\ &=C\|\nabla\times(\nabla\times(\nabla\times\bm u))\|_{r_{u_1}-2}\nonumber\\ &\le C\|\nabla\times(\nabla\times\bm u)\|_{r_{u_1}-1}. \end{aligned}$$ - Proof of $\|\nabla p\|_0\le \|\bm f\|_0$: Letting $\bm v=\nabla p$, it holds that $\nabla\times\bm v=\bm 0$ and $\bm v\in \bm H_0({\rm curl}^2;\Omega)$. We take $\bm v=\nabla p$ in to get $$\begin{aligned} \|\nabla p\|^2_0=(\bm f,\nabla p) \le \|\bm f\|_0\|\nabla p\|_0. \end{aligned}$$ Then $\|\nabla p\|_0\le \|\bm f\|_0$ follows immediately. - Proof of $ \|\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))\|_{0}\le C\|\bm f\|_0$: It follows from that $\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))=\bm f-\nabla p\in [L^2(\Omega)]^3$ in the distribution sense, therefore, $$\begin{aligned} \|\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))\|_{0} = \|\bm f-\nabla p\|_{0}\le \|\bm f\|_0+\|\nabla p\|_{0}\le 2\|\bm f\|_0. \end{aligned}$$ Finally, the uniqueness of the solution is followed from the regularity estimate, and thus the existence of the solution follows immediately. A Mixed finite element method\[sec:mixed\] ========================================== Let $\mathcal{T}_h$ be a quasi-uniform partition of the domain $\Omega$ consisting of tetrahedrons. For any element $K\in\mathcal{T}_h$, let $h_K$ be the infimum of the diameters of spheres containing $K$ and denote the mesh size $h:=\max_{K\in\mathcal{T}_h}h_K$. Let $\mathcal{E}_h$ be the set of all faces of the mesh $\mathcal{T}_h$ . For any element $K\in\mathcal{T}_h$ and face $F\in\mathcal{E}_h$, we denote by $ \bm{n}_K $ and $\bm{n}_F$ the unit outward normal vector to $\partial K$ and face $F$, respectively. Let $F=\partial K \cap \partial K{'}$ be an interior face shared by element $K$ and element $K{'}$, and $\bm n_F$ be pointing from $K$ to $K{'}$. Let $\bm{\phi}$ be a piecewise smooth function. We define the average and jump of $\bm{\phi}$ on $F$ as $$\begin{aligned} \color{black} \{\!\!\{ \bm{\phi}\}\!\!\}:=\frac{1}{2}(\bm{\phi}|_K+\bm{\phi}|_{K'}),\qquad [\![ \bm{\phi}]\!]:=\bm{\phi}|_K-\bm{\phi}|_{K'}.\end{aligned}$$ On a boundary face $F\subset\partial K\cap \partial\Omega$, we set $\{\!\!\{\bm{\phi}\}\!\!\}:=\bm\phi$, $[\![ \bm{\phi}]\!]:=\bm{\phi}$ and $\bm n_F=\bm n_K|_F$. We denote by $\mathcal P_{\ell}(\Lambda)$ the set of all polynomials with order at most $\ell$ on bounded domains $\Lambda$, and denote by $\mathcal P_{\ell}(\mathcal T_h)$ the set of all piecewise polynomials with order at most $\ell$ with respect to the decomposition $\mathcal{T}_h$. Mixed methods ------------- For any integer $ k\ge 1$, we define the finite element spaces $$\begin{aligned} \bm E_h:=\bm H_0(\text{curl};\Omega)\cap [\mathcal P_{k+1}(\mathcal{T}_h)]^3,\qquad Q_h:=H^1_0(\Omega)\cap \mathcal P_{k+2}(\mathcal{T}_h).\end{aligned}$$ Then our mixed method reads: Find $\bm u_h\in \bm E_h$ and $p_h\in Q_h$ such that \[FEM\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm u_h),\nabla\times(\nabla\times\bm v_h))_{K}+(\nabla p_h,\bm v_h)&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm u_h)\}\!\!\},\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F&\nonumber\\ \mp\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm v_h)\}\!\!\},\bm n\times[\![\nabla\times\bm u_h]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times\bm u_h]\!],\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F &=(\bm f,\bm v_h),\label{fem01}\\ (\bm u_h,\nabla q_h)&=0\label{fem02} \end{aligned}$$ hold for all $(\bm v_h,q_h)\in \bm E_h\times Q_h$. The stabilization parameter $\tau>0$ is independent of the mesh size. In the following text, we consider the analysis only for the symmetry case (i.e., we replace ‘$\mp$’ by ‘$-$’ in ), since the proof of the non-symmetry case is similar to the symmetry case. Interpolations -------------- For integer $\ell\ge1$, we denote by $\bm{\Pi}_{h,\ell}^{\text{curl}}$ the standard $\bm H({\rm curl})$-conforming interpolation of the second kind from $\bm H^s(\text{curl};\Omega)$ to $\bm H(\text{curl};\Omega)\cap [\mathcal P_{\ell}(\mathcal{T}_h)]^3$ with $s>\frac{1}{2}$, and thus also from $\bm H^s({\rm curl};$ $\Omega)\cap\bm H_0(\text{curl};\Omega)$ to $\bm H_0(\text{curl};\Omega)\cap [\mathcal P_{\ell}(\mathcal{T}_h)]^3$ with $s>\frac{1}{2}$. The following approximation properties hold ([@MR864305; @MR1609607; @MR2059447]) $$\begin{aligned} \label{app-u1} \|\bm u-\bm{\Pi}_{h,\ell}^{\text{curl}}\bm u\|_{0,K}&\le Ch_K^t(\|\bm u\|_{t,K}+\color{black}h_K\color{black}\|\nabla\times\bm u\|_{t,K}),\\ \|\nabla\times(\bm u-\bm{\Pi}_{h,\ell}^{\text{curl}}\bm u)\|_{0,K}&\le Ch_K^t\|\nabla\times\bm u\|_{t,K},\label{app-u2} \end{aligned}$$ where $\bm u\in \bm H^t({\rm curl};\Omega)$, and real number $t\in (\frac{1}{2},\ell]$, and $$\begin{aligned} \label{app-u4} \|\bm u-\bm{\Pi}_{h,\ell}^{\text{curl}}\bm u\|_{0,K}&\le Ch_K^m\|\bm u\|_{m,K}, \end{aligned}$$ where $\bm u\in \bm H^m(\Omega)$, and real number $m\in (1,\ell+1]$. We define the $L^2$-projection from $L^2(\Omega)$ onto $Q_h$ as: for any $p\in L^2(\Omega)$, find $\Pi_h^Qp\in Q_h$ such that $$\begin{aligned} (\Pi_h^Qp,q)=(p,q)\qquad\forall q\in Q_h.\end{aligned}$$ The following approximation result holds $$\begin{aligned} \label{app-p} |p-\Pi_h^Qp|_1\le Ch^{j-1}\|p\|_j\end{aligned}$$ for real number $j\in [1,k+3]$ and $p\in H^j(\Omega)$. Next, we introduce an interpolation ([@MR3771897; @2019-weifeng]): for any $\bm v\in \bm H^s({\rm curl};\Omega)$ with $s>\frac{1}{2}$, we define $\bm\Pi_h^{\bm E}\bm v\in \bm E_h$ such that $$\begin{aligned} \label{def:piE} \bm{\Pi}_h^{\bm E}\bm v=\bm{\Pi}_{h,k+1}^{\text{curl}}\bm v+\nabla \sigma_h,\end{aligned}$$ where $\sigma_h\in Q_h$ is the solution of the well-posed elliptic problem: $$\begin{aligned} (\nabla\sigma_h,\nabla q_h)=(\bm v-\bm{\Pi}_{h,k+1}^{\text{curl}}\bm v,\nabla q_h) \qquad \forall q_h\in Q_h.\end{aligned}$$ Utilizing and , we get the following result immediately. We have the following orthogonality $$\begin{aligned} \label{or-p} (\bm v-\bm{\Pi}_h^{\bm E}\bm v,\nabla q_h)=0 \end{aligned}$$ hold for all $\bm v\in\bm H^s({\rm curl};\Omega)$ with $s>\frac{1}{2}$ and $q_h\in Q_h$. In addition, there holds the approximation property $$\begin{aligned} \|\bm v-\bm{\Pi}_h^{\bm E}\bm v\|_0\le 2\|\bm v-\bm{\Pi}_{h,k+1}^{\rm curl}\bm v\|_0. \end{aligned}$$ Existence of a unique solution and stability for the mixed method\[sec:stability\] ================================================================================== In this section, we will derive stability results for the underlying mixed method. To this purpose, we firstly develop a novel discrete Sobolev imbedding inequality which is also an efficient tool for numerical analysis of nonlinear problems, and we will report it in future works. All results presented in are valid as $\Omega\in\mathbb{R}^3$ is a bounded multi-connected Lipschitz polyhedron even though we only consider in our presentation $\Omega\in\mathbb{R}^3$ to be a bounded simply connected Lipschitz polyhedron. A novel discrete Sobolev imbedding inequality\[s41\] ----------------------------------------------------- We present a novel discrete Sobolev inequality in the following subsection, which is the [*first*]{} to be reported in literatures. \[discrete-H1\] Assume that $\Omega$ is a bounded Lipschitz polyhedron [ (not necessarily simply-connected)]{} in $\mathbb{R}^3$, then there exists a positive constant $C>0$ such that $$\begin{aligned} \sum_{K\in\mathcal{T}_h}\|\bm v_h\|_{1,K}^2&\le C\left[ \sum_{K\in\mathcal{T}_h}\left( \|\nabla\times\bm v_h\|^2_{0,K} +\|\nabla\cdot\bm v_h\|^2_{0,K} \right) +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\bm v_h]\!]\|^2_{0,F} \right],\\ \|\bm v_h\|_{0,6}^2&\le C\left[ \sum_{K\in\mathcal{T}_h}\left( \|\nabla\times\bm v_h\|^2_{0,K} +\|\nabla\cdot\bm v_h\|^2_{0,K} \right) +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\bm v_h]\!]\|^2_{0,F} \right] \end{aligned}$$ hold for all $\bm v_h\in [\mathcal{P}_{\ell}(\mathcal{T}_h)]^3$ with $\ell\ge 1$ being an integer. We present in the next a Lemma concerning a continuous Sobolev imbedding inequality before the proof of \[continous-H1\] There exists a positive constant $C$ such that $$\begin{aligned} \|\bm v\|_1\le C\left( \|\nabla\times\bm v\|_0+\|\nabla\cdot\bm v\|_0 \right) \end{aligned}$$ holds for any $\bm v\in [H^1_0(\Omega)]^3$. Since $\bm v\in [H^1_0(\Omega)]^3$, it holds $\Delta\bm v\in [H^{-1}(\Omega)]^3$. The Poincar[é]{}’s inequality and integration by parts give that $$\begin{aligned} \|\bm v\|_1^2\le C(\nabla\bm v,\nabla \bm v)=C(-\Delta\bm v,\bm v)\le C\|\Delta\bm v\|_{-1}\|\bm v\|_{1}, \end{aligned}$$ and hence $$\begin{aligned} \label{l61} \|\bm v\|_1\le C\|\Delta \bm v\|_{-1}. \end{aligned}$$ Meanwhile, utilizing integration by parts and the Cauthy-Schwarz inequality, for any $\bm w\in [H^1_0(\Omega)]^3$, we have $$\begin{aligned} -(\color{black}\Delta\color{black}\bm v,\bm w)&=(\nabla\times(\nabla\times\bm v),\bm w)-(\color{black}\nabla\color{black}(\nabla\cdot\bm v),\bm w)\nonumber\\ &=(\nabla\times\bm v,\nabla\times\bm w)+(\nabla\cdot\bm v,\nabla\cdot\bm w)\nonumber\\ &\le C\left(\|\nabla\times\bm v\|_0 +\|\nabla\cdot\bm v\|_0 \right)\|\bm w\|_1. \end{aligned}$$ Then we arrive at $$\begin{aligned} \label{l62} \|\Delta\bm v\|_{-1}=\sup_{\bm 0\neq\bm w\in [H^1_0(\Omega)]^3}\frac{(\color{black}\Delta\color{black}\bm v,\bm w)}{\|\bm w\|_1}\le C\left(\|\nabla\times\bm v\|_0 +\|\nabla\cdot\bm v\|_0 \right). \end{aligned}$$ The proof can be obtained immediately from the combination of both and . Now, we are in the position to prove : By [@MR2034620 Theorem 2.2], for any $\bm v_h\in [\mathcal{P}_{\ell}(\mathcal{T}_h)]^3$ with $\ell\ge 1$, there exists an interpolation $\bm{\mathcal I}_{h,\ell}^{\rm c}$ such that $\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h\in [H^1_0(\Omega)]^3$, and $$\begin{aligned} \label{jc} \|\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h-\bm v_h\|_0 +h \left(\sum_{K\in\mathcal{T}_h}\|\nabla(\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h-\bm v_h)\|_{0,K}^2\right)^{\frac{1}{2}} \le C \left( \sum_{F\in\mathcal{E}_h}h_F\|[\![ \bm v_h ]\!]\|_{0,F}^2 \right)^{\frac{1}{2}}. \end{aligned}$$ Using , the triangle inequality and the estimate in , we obtain that $$\begin{aligned} \|\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h\|_1^2&\le C(\|\nabla\times\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h\|_0^2+\|\nabla\cdot\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h\|_0^2)\nonumber\\ &\le C \sum_{K\in\mathcal{T}_h} \left( \|\nabla\times\bm v_h\|_{0,K}^2+\|\nabla\cdot\bm v_h\|_{0,K}^2 +\|\nabla(\bm{\mathcal I}_{h,\ell}^{\rm c}\bm v_h-\bm v_h )\|_{0,K}^2\right) \nonumber\\ &\le C \left[ \sum_{K\in\mathcal{T}_h}\left( \|\nabla\times\bm v_h\|_{0,K}^2+\|\nabla\cdot\bm v_h\|_{0,K}^2 \right)+\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\bm v_h]\!]\|^2_{0,F}\right]. \end{aligned}$$ Due to the above estimate, the triangle inequality and the estimate in , the proof of the first inequality could be obtained immediately. The second inequality could be obtained through the combination of the discrete Sobolev imbedding inequality in [@MR2629994 Theorem 2.1] and the first inequality. \[th:embed\] For any $\bm v_h\in [\mathcal{P}_{k+1}(\mathcal{T}_h)]^3$ satisfying $$\begin{aligned} (\bm v_h,\nabla q_h)=0\, , \qquad\forall\, q_h\in Q_h= H_0^1(\Omega)\cap \mathcal{P}_{k+2}(\mathcal{T}_h), \end{aligned}$$ there exists a positive constant $C$ such that $$\begin{aligned} \|\bm v_h\|_{0,3}\le C \left( \sum_{K\in\mathcal{T}_h}\|\nabla\times\bm v_h\|_{0,K}^2 +\sum_{ F\in\mathcal{E}_h}\color{black} h_F^{-1}\color{black}\|\bm n\times[\![\bm v_h]\!]\|^2_{0,F} \right)^{\frac{1}{2}}. \end{aligned}$$ With the above lemma, we can derive discrete Sobolev inequalities for $\bm H({\rm curl})$-conforming functions. \[th:embed2\] For any $\bm v_h\in \bm E_h=\bm H_0({\rm curl};\Omega)\cap [\mathcal P_{k+1}(\mathcal{T}_h)]^3$, if there holds $$\begin{aligned} \label{conditon} (\bm v_h,\nabla q_h)=0\, ,\qquad\forall\, q_h\in Q_h= H_0^1(\Omega)\cap \mathcal{P}_{k+2}(\mathcal{T}_h), \end{aligned}$$ then we have $$\begin{aligned} \label{L31} \|\bm v_h\|_{0,3}\le C\|\nabla\times\bm v_h\|_{0}, \end{aligned}$$ and $$\begin{aligned} \label{L32} \|\nabla\times\bm v_h\|_{0,3}\le C \left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm v_h)\|_{0,K}^2 +\sum_{E\in\mathcal{F}_h}\color{black}h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm v_h]\!]\|^2_{0,F} \right)^{\frac{1}{2}}. \end{aligned}$$ The result follows from , , and the fact that $\bm v_h\in \bm H_0({\rm curl};\Omega)$. Since $\nabla\cdot(\nabla\times\bm v_h)=0$ holds for any $q_h\in Q_h$, we have that $$\begin{aligned} (\nabla\times\bm v_h,\nabla q_h)=-(\nabla\cdot(\nabla\times\bm v_h), q_h)=0 \end{aligned}$$ from which the result follows immediately because of . Existence of a unique solution\[s42\] ------------------------------------- In this subsection, we state the existence of a unique solution of . The variational equation admits a unique solution $(\bm u_h,p_h)\in \bm E_h\times Q_h$ as $\tau$ is sufficiently large. We define a mesh-dependent norm as $$\begin{aligned} \interleave\bm v_h \interleave^2&:=\|\bm v_h\|^2_0 +\|\nabla\times\bm v_h\|^2_0 +\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm v_h)\|^2_{0,K}+\sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\|\bm n\times[\![\nabla\times\bm v_h]\!]\|^2_{0,F}. \end{aligned}$$ For all $\bm u_h\in \bm E_h$ satisfying $$\begin{aligned} (\bm u_h,\nabla q_h)=0\, , \qquad \forall \,q_h\in Q_h, \end{aligned}$$ there holds, due to , $$\begin{aligned} \label{ellip1} \interleave \bm u_h \interleave^2&\le C\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F}\right), \end{aligned}$$ and $$\begin{aligned} \label{ellip2} & \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm u_h),\nabla\times(\nabla\times\bm u_h))_{K}\nonumber\\ &\quad-2\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm u_h)\}\!\!\},\bm n\times[\![\nabla\times\bm u_h]\!] \rangle_F\nonumber\\ &\quad +\sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\langle\bm n\times[\![\nabla\times\bm u_h]\!],\bm n\times[\![\nabla\times\bm u_h]\!] \rangle_F\nonumber\\ &\ge \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} -C\sum_{K\in\mathcal T_h}h_K^{-\frac{1}{2}}\|\nabla\times(\nabla\times\bm u_h)\|_{0,K} \|\bm n\times[\![\nabla\times\bm u_h]\!]\|_{0,\partial K}\nonumber\\ &\quad+\sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F}\nonumber\\ &\ge C\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F}\right). \end{aligned}$$ Here, $\tau$ is a sufficiently large constant. By taking $\bm v_h=\nabla q_h\in\bm E_h$ and noticing $\nabla\times(\nabla q_h)=0$, we get $$\begin{aligned} \label{LBB} \sup_{\bm 0\neq \bm v_h\in \bm E_h}\frac{(\bm v_h,\nabla q_h)}{ \interleave\bm v_h \interleave }\ge \|\nabla q_h\|_0. \end{aligned}$$ Then the existence of a unique solution is followed by , , , and the theory for the mixed problem [@MR1115205 Theorem 1.1]. $L^3$ stability of $u_h$ and $\nabla\times u_h$\[s43\] ------------------------------------------------------ \[stable0\] Let $(\bm u_h,p_h)\color{black}\in \bm E_h\times Q_h$ be the solution of , then when the constant $\tau>0$ is sufficient large, we have $$\begin{aligned} \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \le C(\bm f,\bm u_h). \end{aligned}$$ We take $\bm v_h=\bm u_h\in \bm E_h$ in , and $q_h=p_h\in Q_h$ in , and combine the corresponding equalities to obtain $$\begin{aligned} \label{PP1} \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} -2\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm u_h)\}\!\!\},\bm n\times[\![\nabla\times\bm u_h]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F}&=(\bm f,\bm u_h). \end{aligned}$$ Since $\tau>0$ is a sufficient large constant, we have $$\begin{aligned} \label{PP2} &\left| 2\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm u_h)\}\!\!\},\bm n\times[\![\nabla\times\bm u_h]\!] \rangle_F \right|\nonumber\\ &\qquad\le 2\sum_{F\in\mathcal{E}_h}\|\{\!\!\{ \nabla\times(\nabla\times\bm u_h)\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times\bm u_h]\!]\|_{0,F} \nonumber\\ &\qquad\le \frac{1}{2}\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\frac{1}{2}\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F}. \end{aligned}$$ Combining and yields $$\begin{aligned} \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \le C(\bm f,\bm u_h). \end{aligned}$$ With the above lemma, we are ready to prove the following stability result. \[th:stability\] Let $(\bm u_h,p_h)\color{black}\in \bm E_h\times Q_h$ be the solution of , then when the constant $\tau>0$ is sufficient large, we have the following stability result $$\begin{aligned} &\|\bm u_h\|_{0,3}+\|\nabla\times \bm u_h\|_{0,3} \nonumber\\ &+\left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \right)^{\frac{1}{2}} \le C\|\bm f\|_{0,\frac{3}{2}}. \end{aligned}$$ Using , one obtains $$\begin{aligned} \label{PP4} \|\bm u_h\|_{0,3}\le C\|\nabla\times\bm u_h\|_{0}\le C\|\nabla\times\bm u_h\|_{0,3}, \end{aligned}$$ and $$\begin{aligned} \label{PP3} \|\nabla\times\bm u_h\|_{0,3}\le C\left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \right)^{\frac{1}{2}}. \end{aligned}$$ Therefore, applying , , and the Cauthy-Schwarz inequality, we arrive at $$\begin{aligned} &\|\bm u_h\|_{0,3}+\|\nabla\times \bm u_h\|_{0,3}+\left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \right)^{\frac{1}{2}}\\ \nonumber &\le C\|\bm f\|_{0,\frac{3}{2}}. \end{aligned}$$ $L^6$ and discrete $H^1$ stability of $\nabla\times u_h$ \[s44\] ----------------------------------------------------------------- \[th:stability2\] Let $(\bm u_h,p_h)\color{black}\in \bm E_h\times Q_h$ be the solution of , when $\tau>0$ is a sufficient large constant, then we have the following stability result $$\begin{aligned} \|\nabla\times\bm u_h\|_{0,6}^2+ \sum_{K\in\mathcal{T}_h}\|\nabla(\nabla\times\bm u_h)\|_{0,K}^2 +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times\bm u_h]\!]\|_{0,F}^2&\nonumber\\ + \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times\bm u_h)\|_{0,K}^2 +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} &\le C\|\bm f\|_{0,\frac{3}{2}}^2. \end{aligned}$$ Since $\bm u_h\in \bm H_0({\rm curl};\Omega)$, there holds that $\nabla\times\bm u_h\in\bm H({\rm div};\Omega)$, and $\nabla\cdot(\nabla\times\bm u_h)=0$. Using the triangle inequality, one arrives at $$\begin{aligned} \|[\![\nabla\times\bm u_h]\!]\|_{0,F} &= \|(\bm n\times[\![\nabla\times\bm u_h]\!])\times\bm n+(\bm n\cdot[\![\nabla\times\bm u_h]\!])\color{black}\bm n\color{black}\|_{0,F} \nonumber\\ &\le \|(\bm n\times[\![\nabla\times\bm u_h]\!])\times\bm n\|_{0,F} + \|(\bm n\cdot[\![\nabla\times\bm u_h]\!])\color{black}\bm n\color{black}\|_{0,F} \nonumber\\ &\le C \|\bm n\times[\![\nabla\times\bm u_h]\!]\|_{0,F}. \end{aligned}$$ Making use of the above estimate, with $\bm v_h=\nabla\times\bm u_h$, and , we obtain $$\begin{aligned} & \|\nabla\times\bm u_h\|_{0,6}^2+ \sum_{K\in\mathcal{T}_h}\|\nabla(\nabla\times\bm u_h)\|_{0,K}^2 +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times\bm u_h]\!]\|_{0,F}^2 \nonumber\\ &\quad\le \sum_{K\in\mathcal{T}_h}\|\nabla(\nabla\times\bm u_h)\|_{0,K}^2 +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times\bm u_h]\!]\|_{0,F}^2 \nonumber\\ &\quad\le C\left[ \sum_{K\in\mathcal{T}_h}\left( \|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\|\nabla\cdot(\nabla\times\bm u_h)\|^2_{0,K} \right) +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \right]\nonumber\\ &\quad= C\left[ \sum_{K\in\mathcal{T}_h} \|\nabla\times(\nabla\times\bm u_h)\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|\bm n\times[\![\nabla\times\bm u_h]\!]\|^2_{0,F} \right]\nonumber\\ &\quad\le C(\bm f,\bm u_h)\nonumber\\ &\quad \le C\|\bm f\|_{0,\frac{3}{2}}\|\bm u_h\|_{0,3}~. \end{aligned}$$ Meanwhile, it holds from $$\begin{aligned} \|\bm u_h\|_{0,3}&\le C\|\nabla\times\bm u_h\|_{0}\le C\|\nabla\times\bm u_h\|_{0,3}\nonumber\\ &\le C\left( \sum_{K\in\mathcal{T}_h}\|\nabla(\nabla\times\bm u_h)\|_{0,K}^2 +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times\bm u_h]\!]\|_{0,F}^2 \right)^{\frac{1}{2}}. \end{aligned}$$ The combination of the above two inequities completes the proof of the theorem. Main error estimates\[sec:error\] ================================= We first give formulas for integration by parts on low regularity functions. \[aux1\] If $\phi\in H^{\frac{1}{2}+\delta}(\Omega)$ with $\delta\in (0,1/2]$, then $\frac{\partial\phi}{\partial x_i}\in H^{-\frac{1}{2}+\delta}(\Omega)$ for $i=1,2,3$, and there exists a constant $C>0$ such that $$\begin{aligned} \left\| \frac{\partial\phi}{\partial x_i}\right\|_{-\frac{1}{2}+\delta}\le C\| \phi\|_{\frac{1}{2}+\delta}~. \end{aligned}$$ \[th-int1\] If $\nabla\times(\nabla\times\bm u)\in [H^{\frac{1}{2}+\delta}(\Omega)]^3$ with $\delta\in (0,\frac{1}{2}]$, $\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u)))\in [L^2(\Omega)]^3$, and $\bm v_h\in \bm E_h$, then it holds $$\begin{aligned} \label{int} (\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u))),\bm v_h) =(\nabla\times(\nabla\times(\nabla\times\bm u)),\nabla\times\bm v_h)~, \end{aligned}$$ where $(\nabla\times(\nabla\times(\nabla\times\bm u)),\nabla\times\bm v_h)$ is regarded as a duality pairing between $H^{-\frac{1}{2} + \delta}(\Omega)$ and $H_{0}^{\frac{1}{2} - \delta}(\Omega)$. For all $\bm v_h\in \bm E_h$, it is easy to see that $\bm v_h\in [H^{\frac{1}{2}-\delta}(\Omega)]^3$, and $\nabla\times\bm v_h\in [H^{\frac{1}{2}-\delta}(\Omega)]^3$. Since $\nabla\times(\nabla\times\bm u)\in [H^{\frac{1}{2}+\delta}(\Omega)]^3$, according to , it holds $\nabla\times(\nabla\times(\nabla\times\bm u))\in [H^{-\frac{1}{2}+\delta}(\Omega)]^3$. Moreover, since $H^{\frac{1}{2}-\delta}(\Omega)=H_0^{\frac{1}{2}-\delta}(\Omega)$ for $\delta\in (0,\frac{1}{2}]$ ([@MR775683 Theorem 1.4.2.4] or [@MR1742312 Theorem 3.40]), $(\nabla\times(\nabla\times(\nabla\times\bm u)),\nabla\times\bm v_h)$ can be regarded as a duality pairing between $[H^{-\frac{1}{2}+\delta}(\Omega)]^3$ and $[H_0^{\frac{1}{2}-\delta}(\Omega)]^3$. Next, we are going to illustrate that $[\mathcal C_0^{\infty}(\Omega)]^3$ is dense in $\bm E_h$ with respect to the norm $\|\cdot\|_{\bm H^{\frac{1}{2}-\delta}({\rm curl};\Omega)}$. Let $\bm v_h\in \bm E_h$, we define $\tilde{\bm v}_{h}$ on $\mathbb R^{3}$ such that it is identical to $\bm v_{h}$ in $\Omega$, and is trivial outside $\Omega$. Obviously, $\tilde{\bm v}_{h}$ belongs to $\bm H({\rm curl}, \mathbb R^{3})$. It can also be shown that both $\tilde{\bm v}_{h}$ and $\nabla\times\tilde{\bm v}_{h}$ belong to $[H^{\frac{1}{2} - \delta}(\mathbb R^3)]^{3}$, where $\delta\in(0,\frac{1}{2}]$. As a result, we have, similar to the proof of part (ii) of [@MR1742312 Theorem 3.29], that there exists a sequence of vector fields in $[\mathcal C_{0}^{\infty}(\Omega)]^{3}$ which converges to $\bm v_{h}$ with respect to the norm $\|\cdot\|_{\bm H^{\frac{1}{2}-\delta}({\rm curl};\Omega)}$. Finally, the equation follows immediately from the standard duality argument and the fact that $$\begin{aligned} (\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u))),\bm v) &=((\nabla\times(\nabla\times(\nabla\times\bm u)),\nabla\times\bm v)&\forall \,\bm v\in \mathcal [C_0^{\infty}(\Omega)]^3. \end{aligned}$$ \[th-int2\] For all $\phi\in H^{\frac{1}{2}+\delta}(\Omega)$ with $\delta\in (0,\frac{1}{2}]$, and $\psi_h\in \mathcal P_{\ell}(\mathcal T_h)$ with integer $\ell\ge 0$, we have $$\begin{aligned} \label{33} \left(\frac{\partial\phi}{\partial x_i},\psi_h\right)=- \sum_{K\in\mathcal T_h}\left(\phi,\frac{\partial\psi_h}{\partial x_i}\right)_K +\sum_{K\in\mathcal T_h}\langle\phi,\psi_h n_i \rangle_{\partial K} \end{aligned}$$ for $i=1,2,3$, where $ \left(\frac{\partial\phi}{\partial x_i},\psi_h\right)$ denotes a duality pairing between $H^{-\frac{1}{2} + \delta}(\Omega)$ and $H_{0}^{\frac{1}{2} - \delta}(\Omega)$ . According to part (i) of [@MR1742312 Theorem 3.29], $\mathcal D(\bar{\Omega})$ is dense in $H^{\frac{1}{2}+\delta}(\Omega)$, where $$\begin{aligned} \mathcal D(\bar{\Omega})=\{v: v=\tilde{v}|_{\Omega} \text{ for some }\tilde v\in \mathcal C^{\infty}(\mathbb R^3)\}. \end{aligned}$$ We choose $\{\phi_n\}_{n=1}^{\infty}\subset \mathcal D(\bar{\Omega})$ such that $$\begin{aligned} \label{tozero} \|\phi_n-\phi\|_{\frac{1}{2}+\delta}\to 0 \text{ as }n\to\infty. \end{aligned}$$ According to , we have $$\begin{aligned} \left\|\frac{\partial \phi_n}{\partial x_i}-\frac{\partial\phi}{\partial x_i}\right\|_{-\frac{1}{2}+\delta}\to 0 \text{ as }n\to\infty. \end{aligned}$$ For any positive integer $n$, we have $$\begin{aligned} \left(\frac{\partial\phi_n}{\partial x_i},\psi_h\right)=- \sum_{K\in\mathcal T_h}\left(\phi_n,\frac{\partial\psi_h}{\partial x_i}\right)_K +\sum_{K\in\mathcal T_h}\langle\phi_n,\psi_h n_i \rangle_{\partial K}. \end{aligned}$$ Note that implies that $$\begin{aligned} \sum_{K\in\mathcal T_h}\|\phi_n-\phi\|^2_{L^2(\partial K)}\to 0 \text{ as }n\to\infty~, \end{aligned}$$ $\psi_h\in \mathcal P_{\ell}(\mathcal T_h)$ and $\delta\in (0,\frac{1}{2}]$, we have $\psi_h\in H^{\frac{1}{2}-\delta}(\Omega)=H_0^{\frac{1}{2}-\delta}(\Omega)$, so $\left(\frac{\partial\phi}{\partial x_i},\psi_h\right)$ can be regarded as a duality pairing between $H^{-\frac{1}{2}+\delta}(\Omega)$ and $H_0^{\frac{1}{2}-\delta}(\Omega)$. Then it holds $$\begin{aligned} \left|\left(\frac{\partial\phi}{\partial x_i},\psi_h\right) -\left(\frac{\partial\phi_n}{\partial x_i},\psi_h\right)\right|\le C \left\|\frac{\partial\phi}{\partial x_i}-\frac{\partial\phi_n}{\partial x_i}\right\|_{-\frac{1}{2}+\delta}\|\psi_h\|_{\frac{1}{2}-\delta}~, \end{aligned}$$ and follows immediately. \[ass\] In the rest part of this paper, we will assume that the following regularities hold true for the weak solution $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ of : $$\begin{aligned} &\bm u\in [H^{r_{u_0}}(\Omega)]^3,\quad \nabla\times \bm u \in [H^{r_{u_1}}(\Omega)]^3,\quad \nabla\times(\nabla\times \bm u)\in [H^{r_{u_2}}(\Omega)]^3,\quad p\in H^{r_p}(\Omega)~, \end{aligned}$$ where $r_{u_0}\in (\frac{1}{2},\infty)$, $r_{u_1}\in [1,\infty)$, $r_{u_2}\in (\frac{1}{2},\infty)$, and $r_{p}\in (\frac{3}{2},\infty)$. We also assume that $r_{u_0}\le r_{u_1}\le r_{u_0}+1$ in order to simplify the notations in the error analysis. \[pipi\] Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of and holds true, then \[pi\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm u),\nabla\times(\nabla\times\bm v_h))_{K}+(\nabla p,\bm v_h)&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm u)\}\!\!\},\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm v_h)\}\!\!\},\bm n\times[\![\nabla\times\bm u]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times\bm u]\!],\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F &=(\bm f,\bm v_h),\label{pi1}\\ (\bm u,\nabla q_h)&=0\label{pi2} \end{aligned}$$ are true for all $ (\bm v_h,q_h)\in \bm E_h\times Q_h$. For any $ \bm v_h\in \bm E_h$, since the weak solution $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ satisfies in the distribution sense, by , we have $\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u))) \in [L^2(\Omega)]^3$, which further leads to, together with the fact that $ \nabla p, \bm f\in [L^2(\Omega)]^3$, $$\begin{aligned} (\nabla\times(\nabla\times(\nabla\times(\nabla\times\bm u))),\bm v_h)+(\nabla p,\bm v_h)=(\bm f,\bm v_h). \end{aligned}$$ From , we have $$\begin{aligned} (\nabla\times(\nabla\times(\nabla\times\bm u)),\nabla\times\bm v_h)+(\nabla p,\bm v_h)=(\bm f,\bm v_h). \end{aligned}$$ According to , we have $$\begin{aligned} &\sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm u),\nabla\times(\nabla\times\bm v_h))_K\nonumber\\ &\qquad-\sum_{F\in\mathcal{E}_h}\langle \{\!\!\{\nabla\times(\nabla\times\bm u)\}\!\!\},\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F +(\nabla p,\bm v_h)=(\bm f,\bm v_h), \end{aligned}$$ where we have utilized the fact $\nabla\times(\nabla\times\bm u)=\{\!\!\{\nabla\times(\nabla\times\bm u)\}\!\!\}$ because of $\nabla\times(\nabla\times\bm u)\in [H^{s_{u_2}}(\Omega)]^3$ with $s_{u_2}>\frac{1}{2}$ . Moreover, since $\color{black}\bm n\times[\![\nabla\times\bm u]\!]=\bm 0$ because of $\nabla\times\bm u\in[H^{s_{u_1}}(\Omega)]^3$ with $s_{u_1}\ge 1$ and $\bm n\times(\nabla\times\bm u)|_{\partial\Omega}=\bm 0$, follows immediately. Finally, can be obtained due to the fact $Q_h\in H_0^1(\Omega)$. \[error estimates\] Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of and holds true; let $(\bm u_h, p_h)\in \bm E_h\times Q_h$ be the solution of . Then we have the following error estimates $$\begin{aligned} &\|\bm u-\bm u_h\|_0+\|\nabla\times(\bm u-\bm u_h)\|_0\nonumber\\ &\quad+\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times (\bm u-\bm u_h))\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times (\bm u-\bm u_h)]\!]\|^2_{0,F}\right)^{\frac{1}{2}} \nonumber\\ &\quad\quad \le C ( h^{s_{u_0}}\|\bm u\|_{s_{u_0}} + h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}}\color{black} + h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ),\\ & \|\nabla (p- p_h)\|_0\le Ch^{s_p-1}\|p\|_{s_p}, \end{aligned}$$ and the discrete $H^1$ norm error estimate for $\nabla\times(\bm u-\bm u_h)$ $$\begin{aligned} &\left(\sum_{K\in\mathcal{T}_h}\|\nabla(\nabla\times (\bm u-\bm u_h))\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}h_F^{-1}\|[\![\nabla\times (\bm u-\bm u_h)]\!]\|^2_{0,F}\right)^{\frac{1}{2}} \nonumber\\ &\quad\quad \le C ( h^{s_{u_0}}\|\bm u\|_{s_{u_0}} + h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} \color{black} + h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ), \end{aligned}$$ where $s_{u_0}\in(\frac{1}{2},\min(r_{u_0},\color{black}k+2)]$, $s_{u_1}\in [1,\min(r_{u_1},k+1)]$, $s_{u_2}\in [1,\min(r_{u_2},k+1)]$, and $s_{p} \in(\frac{3}{2},\min(r_{p},k+3)]$. We subtract from to get: for all $(\bm v_h,q_h)\in \bm E_h\times Q_h$, there hold \[error1\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm u-\bm u_h)),\nabla\times(\nabla\times\bm v_h))_{K}+(\nabla (p-p_h),\bm v_h)&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm v_h)\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!],\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F &=0,\label{error11}\\ (\bm u-\bm u_h,\nabla q_h)&=0.\label{error12} \end{aligned}$$ To simplify the notation, we define $$\begin{aligned} e_h^{\bm u}:=\bm{\Pi}_h^{\bm E}\bm u-\bm u_h,\qquad e_h^{p}:={\Pi}_h^{Q}p-p_h. \end{aligned}$$ By taking $\bm v_h=e_h^{\bm u}\in \bm E_h$ in and $q_h=e_h^p\in Q_h$ in , we can get \[error2\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm u-\bm u_h)),\nabla\times(\nabla\times e_h^{\bm u}))_{K}+(\nabla (p-p_h), e_h^{\bm u})&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times e_h^{\bm u})\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!],\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F &=0,\label{error21}\\ (\bm u-\bm u_h,\nabla e_h^p)&=0.\label{error22} \end{aligned}$$ Reformulating and noticing that $(\bm{\Pi}_h^{\bm E}\bm u-\bm u,\nabla e_h^p)=0$ from , we obtain that \[error3\] $$\begin{aligned} &\sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times e_h^{\bm u}),\nabla\times(\nabla\times e_h^{\bm u}))_{K}+(\nabla e_h^p, e_h^{\bm u})\nonumber\\ &\quad-2\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times e_h^{\bm u})\}\!\!\},\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\nonumber\\ &\quad+\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times e_h^{\bm u}]\!],\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F \nonumber\\ &\quad\quad= \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)),\nabla\times(\nabla\times e_h^{\bm u}))_{K}+(\nabla (\Pi_h^Qp-p), e_h^{\bm u})\nonumber\\ &\quad\quad\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u))\}\!\!\},\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\nonumber\\ &\quad\quad\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times e_h^{\bm u})\}\!\!\},\bm n\times[\![\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)]\!] \rangle_F\nonumber\\ &\quad\quad\quad+\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)]\!],\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F ,\label{error31}\\ &(e_h^{\bm u},\nabla e_h^p)=0.\label{error32} \end{aligned}$$ Utilizing and arguments similar to those in the proof of , we arrive at $$\begin{aligned} \label{R0} &\frac{1}{2}\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times e_h^{\bm u})\|^2_{0,K} +\frac{1}{2}\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F}\nonumber\\ &\qquad\le \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)),\nabla\times(\nabla\times e_h^{\bm u}))_{K}+(\nabla (\Pi_h^Qp-p), e_h^{\bm u})\nonumber\\ &\qquad\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u))\}\!\!\},\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\nonumber\\ &\qquad\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times e_h^{\bm u})\}\!\!\},\bm n\times[\![\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)]\!] \rangle_F\nonumber\\ &\qquad\quad+\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u)]\!],\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\nonumber\\ &\qquad=:R_1+R_2+R_3+R_4+R_5. \end{aligned}$$ Now, we make estimates $\{R_i\}_{i=1}^5$ individually. Let $\bm{\Pi}_{h,k}$ be the $L^2$-projection from $L^2(\Omega)$ to $[\mathcal{P}_k(\mathcal{T}_h)]^3$. Then, by the triangle inequality, the inverse inequality and the estimate , there holds $$\begin{aligned} \label{est-nabla} &\|\nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u)-\nabla\times(\nabla\times\bm u)\|_{0,K}\nonumber\\ &\quad\le \|\nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm{\Pi}_{h,k}(\nabla\times\bm u))\|_{0,K} + \|\nabla\times(\bm{\Pi}_{h,k}(\nabla\times\bm u)-\nabla\times\bm u)\|_{0,K}\nonumber\\ &\quad\le Ch_K^{-1}\|\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm{\Pi}_{h,k}(\nabla\times\bm u)\|_{0,K} + |\bm{\Pi}_{h,k}(\nabla\times\bm u)-\nabla\times\bm u|_{1,K}\nonumber\\ &\quad\le Ch_K^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1},K} .\nonumber\\ \end{aligned}$$ From and the above estimate, we get $$\begin{aligned} \label{R1} |R_1| &=\left|\sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u)-\nabla\times(\nabla\times\bm u),\nabla\times (\nabla\times e_h^{\bm u} ))_K\right|\nonumber\\ &\le Ch^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}}\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times (\nabla\times e_h^{\bm u})\|_{0,K}\right)^{\frac{1}{2}}. \end{aligned}$$ We use the approximation property for $\Pi_h^Q$ in and to obtain $$\begin{aligned} \label{R2} R_2&\le Ch^{s_p-1}\|p\|_{s_p}\|e_h^{\bm u}\|_0 \nonumber\\ &\le Ch^{s_p-1}\|p\|_{s_p}\|e_h^{\bm u}\|_{0,3} \nonumber\\ &\le Ch^{s_p-1}\|p\|_{s_p}\|\nabla\times e_h^{\bm u}\|_{0} \nonumber\\ &\le Ch^{s_p-1}\|p\|_{s_p}\|\nabla\times e_h^{\bm u}\|_{0,3} \nonumber\\ &\le Ch^{s_p-1}\|p\|_{s_p} \left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times e_h^{\bm u})\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F} \right)^{\frac{1}{2}} . \end{aligned}$$ Using the triangle inequality, the inverse inequality, and the approximation property of $\bm{\Pi}_{h,k+1}^{\text{curl}}$ in leads to $$\begin{aligned} &\sum_{F\in\mathcal{E}_h}h_F\|\{\!\!\{ \nabla\times(\nabla\times(\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm u))\}\!\!\}\|_{0,F}^2\nonumber\\ &\quad\le 2\sum_{F\in\mathcal{E}_h}h_F\|\{\!\!\{ \nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u)-\bm{\Pi}_{h,k}\nabla\times(\nabla\times\bm u)\}\!\!\}\|_{0,F}^2 \nonumber\\ &\qquad+2\sum_{F\in\mathcal{E}_h}h_F\|\{\!\!\{ \bm{\Pi}_{h,k}\nabla\times(\nabla\times\bm u)\}-\nabla\times(\nabla\times\bm u)\}\!\!\}\|_{0,F}^2 \nonumber\\ &\quad\le C\sum_{K\in\mathcal{T}_h}\|\ \nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u)-\bm{\Pi}_{h,k}\nabla\times(\nabla\times\bm u)\|_{0,K}^2 \nonumber\\ &\qquad+Ch^{2s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}}^2 \nonumber\\ &\quad\le C\sum_{K\in\mathcal{T}_h}\|\ \nabla\times(\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u)-\nabla\times(\nabla\times\bm u)\|_{0,K}^2\nonumber\\ &\qquad +C\sum_{K\in\mathcal{T}_h}\|\ \nabla\times(\nabla\times\bm u)-\bm{\Pi}_{h,k}\nabla\times(\nabla\times\bm u)\|_{0,K}^2 +Ch^{2s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}}^2 \nonumber\\ &\quad\le C\left(h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}}+h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}}\right)^2.\nonumber\\ \end{aligned}$$ We use the above estimate to get $$\begin{aligned} \label{R3} |R_3|&=\left| \sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times\nabla\times(\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm u)\}\!\!\},\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\right| \nonumber\\ &\le C\left(h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} +h^{s_{u_2}}\|\nabla\times\nabla\times\bm u\|_{s_{u_2}} \right)\nonumber\\ &\quad\times \left( \sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F} \right)^{\frac{1}{2}}. \end{aligned}$$ Again, by the triangle inequality, the approximation property of $\bm{\Pi}_{h,k+1}^{\text{curl}}$ in , we have $$\begin{aligned} &\sum_{F\in\mathcal{E}_h}h_F^{-1}\|\bm n\times[\![\nabla\times(\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm u)]\!]\|_{0,F}^2\nonumber\\ &\quad\le 2\sum_{F\in\mathcal{E}_h}h_F^{-1}\|\bm n\times[\![\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm{\Pi}_{h,k}\nabla\times\bm u]\!]\|_{0,F}^2\nonumber\\ &\qquad +2\sum_{F\in\mathcal{E}_h}h_F^{-1}\|\bm n\times[\![\bm{\Pi}_{h,k}\nabla\times\bm u-\nabla\times\bm u]\!]\|_{0,F}^2\nonumber\\ &\quad\le C \sum_{K\in\mathcal{T}_h}h_K^{-2}\|\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm{\Pi}_{h,k}\nabla\times\bm u\|_{0,K}^2+Ch^{2(s_{u_1}-1)} \|\nabla\times\bm u\|_{s_{u_1}}^2\nonumber\\ &\quad\le C \sum_{K\in\mathcal{T}_h}h_K^{-2}\|\nabla\times\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\nabla\times\bm u\|_{0,K}^2\nonumber\\ &\qquad +C \sum_{K\in\mathcal{T}_h}h_K^{-2}\|\nabla\times\bm u-\bm{\Pi}_{h,k}\nabla\times\bm u\|_{0,K}^2+Ch^{2(s_{u_1}-1)} \|\nabla\times\bm u\|_{s_{u_1}}^2\nonumber\\ &\quad\le Ch^{2(\color{black}s_{u_1}-1)}\|\nabla\times\bm u\|_{s_{u_1}}^2. \end{aligned}$$ Using the above estimate and the inverse inequality, we arrive at $$\begin{aligned} \label{R4} |R_4|&=\left| \sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times e_h^{\bm u})\}\!\!\},\bm n\times[\![\nabla\times(\bm{\Pi}_{h,k+1}^{\text{curl}}\bm u-\bm u)]\!] \rangle_F\right| \nonumber\\ &\le Ch^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}}\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times (\nabla\times e_h^{\bm u})\|_0\right)^{\frac{1}{2}},\\ |R_5|&=\left|\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm{\Pi}_{h,k+1}^{\rm curl}\bm u-\bm u)]\!],\bm n\times[\![\nabla\times e_h^{\bm u}]\!] \rangle_F\right| \nonumber\\ &\le Ch^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}}\left( \sum_{F\in\mathcal{E}_h}\tau h_F^{-1}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F} \right)^{\frac{1}{2}}. \end{aligned}$$ From , , , and , it gives us that $$\begin{aligned} &\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times e_h^{\bm u})\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F}\nonumber\\ &\qquad\le C ( h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} +h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ). \end{aligned}$$ The result for the estimates of $\bm u-\bm u_h$ can be concluded as follows $$\begin{aligned} &\|e_h^{\bm u}\|_0+\|\nabla\times e_h^{\bm u}\|_0\nonumber\\ &\qquad\le C\left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times e_h^{\bm u})\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times e_h^{\bm u}]\!]\|^2_{0,F} \right)^{\frac{1}{2}}, \end{aligned}$$ according to and the triangle inequality. Using with $\bm v_h=-\nabla e_h^p\in \bm E_h$ yields $$\begin{aligned} \|\nabla e_h^p\|_0^2 &= -(\nabla (p-\Pi_h^Qp),\nabla e_h^p)\le Ch^{s_p-1}\|p\|_{s_p}\|\nabla e_h^p\|_0, \end{aligned}$$ which further provides us with completeness of the proof of the estimate for $\nabla (p-p_h)$ by using the triangle inequality. $H({\rm curl})$ error estimate\[sec:dual\] ========================================== To derive the $\bm H({\rm curl})$ error estimate, we need the following dual problem: Find $(\bm\Phi,\Psi)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ such that \[PDE:dual\] $$\begin{aligned} \nabla\times(\nabla\times(\nabla\times(\nabla\times\bm \Phi)))+\nabla \Psi&=\bm \Theta&\text{ in }\Omega,\\ \nabla\cdot \bm\Phi&=0&\text{ in }\Omega,\\ \bm n\times \bm\Phi&=\bm 0 &\text{ on }\partial\Omega,\\ \bm n\times(\nabla\times \bm\Phi)&=\bm 0&\text{ on }\partial\Omega,\\ \Psi&=0&\text{ on }\partial\Omega. \end{aligned}$$ We notice that $\Psi=0$ when $\nabla\cdot\bm{\Theta}=0$. \[ass2\] We assume that the following regularity holds for the weak solution $(\bm{\Phi},\Psi)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ of $$\begin{aligned} \label{reg} \|\bm{\Phi}\|_{\beta} +\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} \color{black} +\|\nabla\times\bm\Phi\|_{1+\gamma} \le C_{\rm reg}\|\bm \Theta\|_0, \end{aligned}$$ where $\beta\in (\frac{1}{2},\color{black}1]$, $\gamma\in [0,1]$, $\gamma\le \beta$, and $C_{\rm reg}$ is a constant independent of mesh size. We notice that when $\Omega$ is convex, holds as $ \gamma=1, \beta=1$ from the regularity result in [@MR3808156 Theorem 11]. \[lemma:dual\] Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of , and let , hold true; let $(\bm u_h, p_h)\in \bm E_h\times Q_h$ be the solution of , and let $\nabla\cdot\bm \Theta=0$. Then we have the following error estimates $$\begin{aligned} (\bm \Theta,\bm u-\bm u_h) \le Ch^{{\sigma}} ( h^{s_{u_0}}\|\bm u\|_{s_{u_0}} + h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} + h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ) \|\bm{\Theta}\|_0, \end{aligned}$$ where $\sigma=\min(\beta,\gamma)$ with $\beta$, $\gamma$ being defined in . By arguments similar to those in the proof of , we have the following equations hold for the weak solution $(\bm{\Phi},\Psi)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ of : \[equa:dual\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times \bm\Phi),\nabla\times(\nabla\times(\bm u-\bm u_h)))_{K}+(\nabla \Psi,(\bm u-\bm u_h))&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm \Phi)\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times\bm \Phi]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times\bm \Phi]\!],\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F &=(\bm \Theta,(\bm u-\bm u_h)),\\ (\bm \Phi,\nabla q)&=0. \end{aligned}$$ We use the fact $\Psi=0$ to get $$\begin{aligned} \label{proof:dual1} (\bm \Theta,\bm u-\bm u_h) &=\sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm u-\bm u_h)),\nabla\times(\nabla\times \bm\Phi))_{K}\nonumber\\ &\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm \Phi)\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F\nonumber\\ &\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times\bm \Phi]\!] \rangle_F\nonumber\\ &\quad+\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times\bm \Phi]\!],\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F. \end{aligned}$$ Let $(\bm{\Phi}_h,\Psi_h)\in \bm E_h\times Q_h$ be the solution of the following system: for $\forall\, (\bm v_h,q_h)\in \bm E_h\times Q_h$ \[FEM2\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times\bm \Phi_h),\nabla\times(\nabla\times\bm v_h))_{K}+(\nabla \Psi_h,\bm v_h)&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm \Phi_h)\}\!\!\},\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm v_h)\}\!\!\},\bm n\times[\![\nabla\times\bm \Phi_h]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times\bm \Phi_h]\!],\bm n\times[\![\nabla\times\bm v_h]\!] \rangle_F &=(\bm f,\bm v_h),\\ (\bm \Phi_h,\nabla q_h)&=0 \end{aligned}$$ Using , and the fact $\Psi=0$, we can get the error estimate: $$\begin{aligned} \label{error:dual} &\|\bm \Phi-\bm \Phi_h\|_0+\|\nabla\times(\bm \Phi-\bm \Phi_h)\|_0 +\|\nabla (\Psi-\Psi_h)\|_0\nonumber\\ &\quad+\sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times (\bm \Phi-\bm \Phi_h))\|^2_{0,K} +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\|\bm n\times[\![\nabla\times (\bm \Phi-\bm \Phi_h)]\!]\|^2_{0,F} \nonumber\\ &\quad\quad \color{black} \le C\left(h^{\beta} \|\bm\Phi\|_{\beta} +h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma}\right). \end{aligned}$$ From , it follows that \[proof:dual2\] $$\begin{aligned} \sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm u-\bm u_h)),\nabla\times(\nabla\times\bm \Phi_h))_{K}+(\nabla (p-p_h),\bm \Phi_h)&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times\bm \Phi_h]\!] \rangle_F&\nonumber\\ -\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times\bm \Phi_h)\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F&\nonumber\\ +\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!],\bm n\times[\![\nabla\times\bm \Phi_h]\!] \rangle_F &=0,\\ (\bm u-\bm u_h,\nabla \Psi_h)&=0. \end{aligned}$$ In terms of and , and by the fact $(\nabla(p-p_h),\bm{\Phi}_h)=(\nabla(p-p_h),\bm{\Phi}_h-\bm{\Phi})$, we have $$\begin{aligned} \label{T0} (\bm \Theta,\bm u-\bm u_h)&=\sum_{K\in\mathcal{T}_h}(\nabla\times(\nabla\times(\bm u-\bm u_h)),\nabla\times(\nabla\times (\bm\Phi-\bm\Phi_h)))_{K}\nonumber\\ &\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm \Phi-\bm\Phi_h))\}\!\!\},\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F\nonumber\\ &\quad-\sum_{F\in\mathcal{E}_h}\langle\{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\},\bm n\times[\![\nabla\times(\bm \Phi-\bm\Phi_h)]\!] \rangle_F\nonumber\\ &\quad+\sum_{F\in\mathcal{E}_h}\color{black}\tau h_F^{-1}\color{black}\langle\bm n\times[\![\nabla\times(\bm \Phi-\bm\Phi_h)]\!],\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \rangle_F \nonumber\\ &\quad+(\nabla (p-p_h),\bm\Phi-\bm \Phi_h) \nonumber\\ &=:T_1+T_2+T_3+T_4+T_5. \end{aligned}$$ Now we make the estimate for $\{T_i\}_{i=1}^5$ individually. To simply the notation, we define $$\begin{aligned} \mathcal M:=( h^{s_{u_0}}\|\bm u\|_{s_{u_0}} + h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} \color{black} + h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ). \end{aligned}$$ We use the Cauthy-Schwarz inequality, the estimate in , the error estimate and the regularity to get $$\begin{aligned} \label{T1} |T_1|&\le \left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times(\bm u-\bm u_h) )\|_{0,K} \right)^{\frac{1}{2}} \left( \sum_{K\in\mathcal{T}_h}\|\nabla\times(\nabla\times(\bm \Phi-\bm \Phi_h) )\|_{0,K} \right)^{\frac{1}{2}} \nonumber\\ &\le C\mathcal M \color{black} \left( h^{\beta} \|\bm\Phi\|_{\beta} +h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma} \right)\nonumber\\ &\le Ch^{{\sigma}}\mathcal{M}\|\bm{\Theta}\|_0. \end{aligned}$$ Using the Cauthy-Schwarz inequality, the triangle inequality, the inverse inequality, the error estimates in , , and the regularity , we get $$\begin{aligned} \label{T2} |T_2|&\le \sum_{F\in \mathcal{E}_h} h_F^{\frac{1}{2}}\| \{\!\!\{ \nabla\times(\nabla\times(\bm \Phi-\bm\Phi_h))\}\!\!\}\|_{0,F} h_F^{-\frac{1}{2}}\|\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \|_{0,F}\nonumber\\ &\le \sum_{F\in \mathcal{E}_h} \| \{\!\!\{ \nabla\times(\nabla\times\bm\Phi-\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times\bm \Phi))\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \|_{0,F}\nonumber\\ &\quad+\sum_{F\in \mathcal{E}_h} \| \{\!\!\{ \nabla\times(\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times\bm \Phi)-\nabla\times\bm\Phi_h))\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times(\bm u-\bm u_h)]\!] \|_{0,F}\nonumber\\ &\le Ch^{\beta}\mathcal M\|\nabla\times(\nabla\times\bm\Phi)\|_{\beta} + Ch^{\gamma}\mathcal M\|\nabla\times\bm\Phi\|_{1+\gamma} \nonumber\\ &\quad +C\mathcal M\left( \sum_{K\in\mathcal{T}_h} \|\nabla\times(\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times\bm \Phi)-\nabla\times\bm\Phi_h))\|_{0,K} \right)^{\frac{1}{2}}\nonumber\\ &\le C\mathcal M \left( h^{\beta} \|\bm\Phi\|_{\beta}+h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma} \right)\nonumber\\ &\le Ch^{{\sigma}}\mathcal{M}\|\bm{\Theta}\|_0. \end{aligned}$$ Similar to , one can get $$\begin{aligned} \label{T3} |T_3|&\le \sum_{F\in \mathcal{E}_h} \| \{\!\!\{ \nabla\times(\nabla\times(\bm u-\bm u_h))\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times(\bm \Phi-\bm \Phi_h)]\!] \|_{0,F}\nonumber\\ &\le \sum_{F\in \mathcal{E}_h} \| \{\!\!\{ \nabla\times(\nabla\times\bm u-\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times\bm u))\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times(\bm \Phi-\bm \Phi_h)]\!] \|_{0,F}\nonumber\\ &\quad+\sum_{F\in \mathcal{E}_h} \| \{\!\!\{ \nabla\times(\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times\bm u)-\nabla\times\bm u_h))\}\!\!\}\|_{0,F} \|\bm n\times[\![\nabla\times(\bm \Phi-\bm \Phi_h)]\!] \|_{0,F}\nonumber\\ &\le C\mathcal M \color{black} \left( h^{\beta} \|\bm\Phi\|_{\beta} +h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma} \right)\nonumber\\ &\le Ch^{{\sigma}}\mathcal{M}\|\bm{\Theta}\|_0,\\ \label{T4} |T_4| &\le C\mathcal M \color{black} \left( h^{\beta} \|\bm\Phi\|_{\beta} +h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma} \right)\le Ch^{{\sigma}}\mathcal{M}\|\bm{\Theta}\|_0. \end{aligned}$$ Again, we use the Cauthy-Schwarz inequality, the estimate in , the error estimates in and the regularity to get $$\begin{aligned} \label{T5} |T_5|&\le \|\nabla p-\nabla p_h\|_0\|\bm{\Phi}-\bm\Phi_h\|_0\nonumber\\ &\le C\mathcal M \color{black} \left( h^{\beta}\|\bm{\Phi}\|_{\beta} +h^{\beta}\|\nabla\times(\nabla\times\bm{\Phi})\|_{\beta} +h^{\gamma}\|\nabla\times\bm{\Phi}\|_{1+\gamma} \right)\nonumber\\ &\le Ch^{{\sigma}}\mathcal{M}\|\bm{\Theta}\|_0. \end{aligned}$$ We thus complete the proof by applying , , , , and . In the following two subsections, we will give estimates for $\|\bm u-\bm u_h\|_0$ and $\|\nabla\times(\bm u-\bm u_h)\|_0$, respectively, and then obtain the error estimate in $\bm H({\rm curl})$ norm. $L^2$ error estimate -------------------- \[dual-u\] Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of , and let , hold true; let $(\bm u_h, p_h)\in \bm E_h\times Q_h$ be the solution of . Then we have the following error estimates $$\begin{aligned} \|\bm u-\bm u_h\|_0\le& Ch^{{\sigma}} ( h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} \color{black} +h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ) +Ch^{s_{u_0}}\|\bm u\|_{s_{u_0}}, \end{aligned}$$ where $\sigma=\min(\alpha,\beta,\gamma)$, $\alpha$ is defined in , $\beta$ and $\gamma$ are defined in . We take $\bm{\Theta}$ satisfying the following problem: Find $\bm \Theta\in \bm H({\rm curl};\Omega)\cap \bm H({\rm div};\Omega)$ such that $$\begin{aligned} \label{theta} \nabla\times\bm\Theta&=\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)&\text{in }\Omega,\\ \nabla\cdot\bm \Theta&=0&\text{in }\Omega,\\ \bm n\times\bm \Theta&=\bm 0&\text{on }\partial\Omega. \end{aligned}$$ It follows form and that $$\begin{aligned} (\bm{\Pi}_h^{\bm E}\bm u-\bm u_h,\nabla q_h)=(\bm u,\nabla q_h)=-(\nabla\cdot\bm u,q_h)=0 \end{aligned}$$ for all $q_h\in Q_h$. Due to the result in [@MR2009375 Lemma 4.5], one has $$\begin{aligned} \|(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)-\bm{\Theta}\|_0&\le Ch^{\alpha}\|\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)\|_{0},\label{theta02} \end{aligned}$$ where $\alpha$ is defined in . As a result, using the triangle inequality, the estimate , and $\sigma\le \alpha$, one can get $$\begin{aligned} \label{est:theta01} \|\bm\Theta\|_0&\le \|(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)-\bm{\Theta}\|_0+ \|\bm{\Pi}_h^{\bm E}\bm u-\bm u_h\|_0\nonumber\\ &\le Ch^{\sigma}\|\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)\|_{0}+\|\bm u-\bm u_h\|_0 \nonumber\\ &\quad +Ch^{s_{u_0}}(\|\bm u\|_{s_{u_0}} +\delta(s_{u_0})h\|\nabla\times\bm u\|_{s_{u_0}} )\nonumber\\ &\le Ch^{\sigma}\mathcal M+Ch^{s_{u_0}}(\|\bm u\|_{s_{u_0}} +\delta(s_{u_0})h\|\nabla\times\bm u\|_{s_{u_0}} )+\|\bm u-\bm u_h\|_0. \end{aligned}$$ Due to the triangle inequality, , and , one arrives at $$\begin{aligned} \|\bm u-\bm u_h\|_0^2&=(\bm u-\bm u_h,\bm u-\bm u_h)\nonumber\\ &=(\bm \Theta,\bm u-\bm u_h) +((\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)-\bm \Theta,\bm u-\bm u_h) +(\bm u-\bm{\Pi}_h^{\bm E}\bm u,\bm u-\bm u_h)\nonumber\\ &\le Ch^{\sigma}\mathcal{M}\|\bm{\Theta}\|_0 +Ch^{\sigma}\|\nabla\times(\bm{\Pi}_h^{\bm E}\bm u-\bm u_h)\|_{0}\|\bm u-\bm u_h\|_0 \nonumber\\ &\quad+Ch^{s_{u_0}}(\|\bm u\|_{s_{u_0}} +\delta(s_{u_0})h\|\nabla\times\bm u\|_{s_{u_0}} ) \|\bm u-\bm u_h\|_0\nonumber\\ &\le Ch^{2\sigma}\mathcal{M}^2 +Ch^{2s_{u_0}}(\|\bm u\|_{s_{u_0}} +\delta(s_{u_0})h\|\nabla\times\bm u\|_{s_{u_0}} )^2+\frac{1}{2}\|\bm u-\bm u_h\|_0^2, \end{aligned}$$ where we have defined $\delta(s_{u_0})=0$ when $s_{u_0}>1$ and $\delta(s_{u_0})=1$ when $s_{u_0}\in (\frac{1}{2},1]$. The above inequality further implies that $$\begin{aligned} \|\bm u-\bm u_h\|_0\le& C\left(h^{\sigma} ( h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} + h^{s_p-1}\|p\|_{s_p} )+h^{s_{u_0}}(\|\bm u\|_{s_{u_0}} +\delta(s_{u_0})h\|\nabla\times\bm u\|_{s_{u_0}} )\right)\nonumber\\ \le& C\left(h^{\sigma} ( h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} + h^{s_p-1}\|p\|_{s_p} )+h^{s_{u_0}}\|\bm u\|_{s_{u_0}}\right), \end{aligned}$$ and we complete the proof. Curl operator error estimate ---------------------------- We first introduce an interpolation denoted by $\bm{\Pi}_{h,\ell}^{\rm curl,c}$. Form [@MR2194528 Proposition 4.5], for any integer $\ell\ge 1$, let $\bm v_h\in [\mathcal P_{\ell}(\mathcal{T}_h)]^3$, there exists a function $\bm{\Pi}_{h,\ell}^{\rm curl,c}\bm v_h\in [\mathcal P_{\ell}(\mathcal{T}_h)]^3\cap \bm H_0({\rm curl};\Omega) $ such that $$\begin{aligned} \|\bm{\Pi}_{h,\ell}^{\rm curl,c}\bm v_h-\bm v_h\|_0 +h\left(\sum_{K\in\mathcal{T}_h}\|\nabla\times(\bm{\Pi}_{h,\ell}^{\rm curl,c}\bm v_h-\bm v_h)\|_{0,K}^2\right)^{\frac{1}{2}} \le C \left( \sum_{F\in\mathcal{E}_h}h_F\|\bm n\times[\![ \bm v_h ]\!]\|_{0,F}^2 \right)^{\frac{1}{2}}.\end{aligned}$$ Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of , and let , hold true; let $(\bm u_h, p_h)\in \bm E_h\times Q_h$ be the solution of . Then we have the following error estimates $$\begin{aligned} \small \|\nabla\times(\bm u-\bm u_h)\|_0\le Ch^{\sigma} ( h^{s_{u_1}-1}\|\nabla\times\bm u\|_{s_{u_1}} \color{black} + h^{s_{u_2}}\|\nabla\times(\nabla\times\bm u)\|_{s_{u_2}} + h^{s_p-1}\|p\|_{s_p} ) + Ch^{s_{u_0}}\|\bm u\|_{s_{u_0}}, \end{aligned}$$ where $\sigma=\min(\alpha,\beta,\gamma)$, $\alpha$ is defined in , $\beta$ and $\gamma$ are defined in . The proof of the theorem can be obtained through arguments similar to those in the proof of and through considering the problem with $\nabla\times(\bm{\Pi}_{h,k+1}^{\rm curl}(\nabla\times \bm u)-\bm{\Pi}_{h,k+1}^{\rm curl,c}(\nabla\times\bm u_h))$ being the right hand side term, and we skip the details of the proof. Finally, we present the optimal error estimates in $\bm H({\rm curl})$-norm as follows. Let $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ be the weak solution of , and $(\bm u_h, p_h)\in \bm E_h\times Q_h$ be the solution of . When $\Omega$ is convex and the weak solution $(\bm u,p)\in \bm H_0({\rm curl}^2;\Omega)\times H_0^1(\Omega)$ of is sufficiently smooth, then we have the following error estimates $$\begin{aligned} \|\bm u-\bm u_h\|_0+\|\nabla\times(\bm u-\bm u_h)\|_0\le Ch^{k+1} (\|\bm u\|_{k+1} +\|\nabla\times\bm u\|_{k+1}+\|p\|_{k+1} ). \end{aligned}$$ Acknowledgements {#acknowledgements .unnumbered} ================ [Gang Chen is supported by National Natural Science Foundation of China (NSFC) under grant no. 11801063, China Postdoctoral Science Foundation under grant no. 2018M633339, and Key Laboratory of Numerical Simulation of Sichuan Province (Neijiang, Sichuan Province) under grant no. 2017KF003. Weifeng Qiu is supported by Research Grants Council of the Hong Kong Special Administrative Region of China under grant no. CityU 11304017. Liwei Xu is supported by a Key Project of the Major Research Plan of NSFC under grant no. 91630205 and NSFC under grant no. 11771068. ]{} [10]{} , [*An optimal domain decomposition preconditioner for low-frequency time-harmonic [M]{}axwell equations*]{}, Math. Comp., 68 (1999), pp. 607–631. , [*Vector potentials in three-dimensional non-smooth domains*]{}, Math. Methods Appl. Sci., 21 (1998), pp. 823–864. , [*A finite element method for magnetohydrodynamics*]{}, Comput. Methods Appl. Mech. Engrg., 190 (2001), pp. 5867–5892. , [*Hodge decomposition methods for a quad-curl problem on planar domains*]{}, J. Sci. Comput., 73 (2017), pp. 495–513. , [*Mixed and hybrid finite element methods*]{}, vol. 15 of Springer Series in Computational Mathematics, Springer-Verlag, New York, 1991. , [*A priori and computable a posteriori error estimates for an [HDG]{} method for the coercive [M]{}axwell equations*]{}, Comput. Methods Appl. Mech. Engrg., 333 (2018), pp. 287–310. , [*Elliptic boundary value problems on corner domains*]{}, vol. 1341 of Lecture Notes in Mathematics, Springer-Verlag, Berlin, 1988. Smoothness and asymptotics of solutions. , [*Discrete functional analysis tools for discontinuous [G]{}alerkin methods with application to the incompressible [N]{}avier-[S]{}tokes equations*]{}, Math. Comp., 79 (2010), pp. 1303–1330. , [*A family of [$C\sp{1}$]{} finite elements with optimal approximation properties for various [G]{}alerkin methods for 2nd and 4th order problems*]{}, RAIRO Anal. Numér., 13 (1979), pp. 227–255. , [*A semi-implicit energy conserving finite element method for the dynamical incompressible magnetohydrodynamics equations*]{}, Computer Methods in Applied Mechanics and Engineering, (2018). , [*A mixed finite element method with exactly divergence-free velocities for incompressible magnetohydrodynamics*]{}, Comput. Methods Appl. Mech. Engrg., 199 (2010), pp. 2840–2855. , [*Elliptic problems in nonsmooth domains*]{}, vol. 24 of Monographs and Studies in Mathematics, Pitman (Advanced Publishing Program), Boston, MA, 1985. , [*A new error analysis for discontinuous finite element methods for linear elliptic problems*]{}, Math. Comp., 79 (2010), pp. 2169–2189. , [*An interior penalty method for a sixth-order elliptic equation*]{}, IMA J. Numer. Anal., 31 (2011), pp. 1734–1753. , [*An interior penalty [G]{}alerkin method for the [MHD]{} equations in heterogeneous domains*]{}, J. Comput. Phys., 221 (2007), pp. 349–369. , [*Finite elements in computational electromagnetism*]{}, Acta Numer., 11 (2002), pp. 237–339. , [*A discontinuous [G]{}alerkin method for the fourth-order curl problem*]{}, J. Comput. Math., 30 (2012), pp. 565–578. , [*Interior penalty method for the indefinite time-harmonic [M]{}axwell equations*]{}, Numer. Math., 100 (2005), pp. 485–518. , [*Stable finite element methods preserving [$\nabla\cdot B=0$]{} exactly for [MHD]{} models*]{}, Numer. Math., 135 (2017), pp. 371–396. , [*[Magnetic-Electric Formulations for Stationary Magnetohydrodynamics Models]{}*]{}, ArXiv e-prints, (2017). , [*The inhomogeneous [D]{}irichlet problem in [L]{}ipschitz domains*]{}, J. Funct. Anal., 130 (1995), pp. 161–219. , [*A posteriori error estimates for a discontinuous [G]{}alerkin approximation of second-order elliptic problems*]{}, SIAM J. Numer. Anal., 41 (2003), pp. 2374–2399. , [*A convergencet linerized [L]{}agrange finite element method for the magneto-hydrodynamic equations in [2D]{} nonsmooth and nonconvex domains*]{}, submmitted. , [*Strongly elliptic systems and boundary integral equations*]{}, Cambridge University Press, Cambridge, 2000. , [*Finite element methods for [M]{}axwell’s equations*]{}, Numerical Mathematics and Scientific Computation, Oxford University Press, New York, 2003. , [*Finite element methods for [M]{}axwell’s transmission eigenvalues*]{}, SIAM J. Sci. Comput., 34 (2012), pp. B247–B264. , [*A new family of mixed finite elements in [${\bf R}^3$]{}*]{}, Numer. Math., 50 (1986), pp. 57–81. , [*Singularities of the quad curl problem*]{}, J. Differential Equations, 264 (2018), pp. 5025–5069. , [*A mixed dg method and an hdg method for incompressible magnetohydrodynamics*]{}, IMA J. Numer. Anal., (2019), pp. 1–34. , [*A mixed [FEM]{} for the quad-curl eigenvalue problem*]{}, Numer. Math., 132 (2016), pp. 185–200. , [*[An $H^2$(curl)-conforming finite element in 2D and its applications to the quad-curl problem]{}*]{}, ArXiv e-prints, (2018). , [*Mixed schemes for quad-curl equations*]{}, ESAIM Math. Model. Numer. Anal., 52 (2018), pp. 147–161. height 2pt depth -1.6pt width 23pt, [*Regular decomposition and a framework of order reduced methods for fourth order problems*]{}, Numer. Math., 138 (2018), pp. 241–271. , [*A nonconforming finite element method for fourth order curl equations in [$\Bbb{R}^{3}$]{}*]{}, Math. Comp., 80 (2011), pp. 1871–1886.

--- abstract: 'This paper proposes a new deep neural network for object detection. The proposed network, termed ASSD, builds feature relations in the spatial space of the feature map. With the global relation information, ASSD learns to highlight useful regions on the feature maps while suppressing the irrelevant information, thereby providing reliable guidance for object detection. Compared to methods that rely on complicated CNN layers to refine the feature maps, ASSD is simple in design and is computationally efficient. Experimental results show that ASSD competes favorably with the state-of-the-arts, including SSD, DSSD, FSSD and RetinaNet. Code is available at: <https://github.com/yijingru/ASSD-Pytorch>.' author: - | Jingru Yi,Pengxiang Wu,Dimitris N. Metaxas\ Department of Computer Science, Rutgers University, Piscataway, NJ 08854, USA\ [jy486,pw241,dnm@cs.rutgers.edu]{} title: 'ASSD: Attentive Single Shot Multibox Detector' --- Introduction {#sec:intro} ============ In recent years, object detection has experienced a rapid development with the aid of convolutional neural networks (CNN). Generally, the CNN-based object detectors can be divided into two types: one-stage object detector and two-stage object detector. The two-stage object detectors, such as R-CNN [@girshick2014rich], Fast and Faster R-CNN [@Girshick_2015_ICCV; @ren2015faster] and SPPnet [@he2014spatial], are proposal driven, with a second stage for refining the detection. However, these two-stage object detectors are inefficient for real-time applications due to the decoupled multi-stage processing. In contrast, the one-stage object detectors, including YOLO [@redmon2016you], YOLO-v2 [@redmon2017yolo9000] and SSD [@liu2016ssd], propose to model the object detection as a simple regression problem and encapsulate all the computation in a single feed-forward CNN, thereby speeding up the detection to a large extent. However, the one-stage detectors are generally less accurate than the two-stage ones. The main reason would be the extreme foreground-background class imbalance of the dense anchor boxes [@lin2018focal]. To solve this issue, RetinaNet [@lin2018focal] proposes a focal loss to train its FPN-based [@lin2017feature] one-stage detector. However, the focal loss is parameter sensitive, and it would require exhaustive experiments to obtain the optimal parameters. ![The structures of different SSD-based detectors. (a) SSD [@liu2016ssd], (b) DSSD [@fu2017dssd], (c) FSSD [@li2017fssd], (d) ASSD (Ours).[]{data-label="fig:fig1"}](pic/fig1.pdf){width="43.00000%"} In this paper, we aim to improve the one-stage detectors from a different perspective. We propose to discover the intrinsic feature relations on the feature map to focus the detector on regions that are critical to the detection task. Our key motivation comes from the human vision system. When perceiving a scene, humans first glance at the scene and then instantly figure out the contents through global dependency analysis. Besides, when the eyeballs focus on a fixation point, the resolution of the neighboring regions decreases. To simulate such human vision mechanism, we design an *attention unit* that is capable of analyzing the importance of features at different positions, based on the global feature relations. The attention unit is fully differentiable and in-place. This design generates the *attention maps* which highlight the useful regions and suppress the irrelevant information. Compared to methods that only build relations among proposals [@Hu_2018_CVPR; @zeng2018crafting], our method considers the global feature correlations at pixel level and conforms to the visual mechanism of humans. We choose the SSD as our base one-stage detector, which provides the optimal trade-off among simplicity, speed and accuracy. Combined with the attention unit, we term the resulting object detector as Attentive SSD (ASSD). ASSD is simpler in design and more effective at refining the contextual semantics compared to the existing SSD-based detectors (see Fig. \[fig:fig1\]). In particular, DSSD [@fu2017dssd] relies on a complex feature pyramid to encourage the information flow among different layers. While achieving better accuracies than the original SSD, it is relatively more complex and thus computationally inefficient. Another recent approach, FSSD [@li2017fssd], builds additional fusion modules for multi-scale feature aggregation, but only achieves marginal improvements upon SSD. In contrast to these works, our ASSD retains the original structure of SSD and employs a single efficient attention unit to refine the object information from each layer (see Fig. \[fig:fig1\]d). This design preserves the advantages of the original SSD while being more effective at learning object features. We demonstrate the advantages of ASSD on a number of representative benchmark datasets, including PASCAL VOC [@Everingham15] and COCO [@lin2014microsoft]. Experimental results validate the superiority of ASSD compared to the state-of-the-arts in terms accuracy and efficiency. Our main contributions can be summarized as follows: 1. We propose to incorporate pixel-wise feature relations into the one-stage detector. Our design follows the human vision mechanism and facilitates the object feature learning. 2. The proposed network preserves the simplicity and efficiency of SSD while being more accurate. 3. We perform a series of experiments to validate the advantages of ASSD. The experimental results show that ASSD competes favorably with the state-of-the-arts in terms of accuracy and efficiency. Related Works {#sec:related_works} ============= {width="80.00000%"} Object Detection ---------------- Object detection involves localization and classification. From traditional hand-crafted feature-based methods (e.g., SIFT [@sermanet2013pedestrian] and HOG [@dalal2005histograms]) to recent CNN-based models, last decades have witnessed a significant development of object detection techniques. In recent years, CNN-based object detectors have gained remarkable success and generally can be divided into two categories: the proposal-driven two-stage detectors, and the regression-oriented one-stage detectors. The two-stage object detectors are composed of two decoupled operations: proposal generation and box refinement. The pioneering work, R-CNN [@girshick2014rich], utilizes selective search to generate region proposals and classifies them with class-specific linear SVM using the learned CNN features. The major weakness of R-CNN is that it needs to perform the forward pass for each proposal, leading to an extremely inefficient model. To solve this issue, SPPnet [@he2014spatial] suggests sharing the CNN computation for all proposals, whereas Fast R-CNN [@Girshick_2015_ICCV] replaces the SVM with fully-connected layers (FCs) to enable single-stage training without additional feature caching. Faster R-CNN [@ren2015faster] goes a step further and introduces a region proposal network (RPN) where the proposal computation is performed through shared CNN features, thereby largely speeding up the detection process. In a more aggressive manner, R-FCN [@dai2016r] replaces the FCs with position-sensitive score maps and encodes translation variance information into these maps, leading to a variance insensitive fully convolutional network (FCN) for accurate object detection. Another recent work, FPN [@lin2017feature], employs a top-down pyramid structure to reuse the higher-resolution features maps from the feature hierarchy and has achieved the state-of-the-art results. Two-stage object detectors are quite effective at object feature learning. However, they are generally inefficient in computation. Different from two-stage detectors, one-stage object detectors discard the region proposal stage, thereby making the detection more efficient. YOLO [@redmon2016you] proposes to use a single CNN to simultaneously predict multiple bounding boxes as well as their class probabilities. While being extremely fast, YOLO is far less accurate than the two-stage models. Instead of directly predicting the coordinates of bounding boxes, YOLOv2 [@redmon2017yolo9000] employs the anchor boxes to facilitate the detection and improves the accuracy a lot. From a different perspective, SSD [@liu2016ssd] builds a pyramid CNN network on top of the backbone, and detects objects of different scales from the multi-scale feature maps in a single forward pass. SSD has achieved better performance than YOLOv2. Based on SSD and similar to FPN, DSSD [@fu2017dssd] employs top-down pyramid CNN layers to improve the accuracy but at the cost of computational efficiency. FSSD [@li2017fssd] inserts a fusion module at the bottom of the feature pyramid to enhance the accuracy of SSD. While still being fast, FSSD only achieves marginal improvements upon SSD in accuracy. Other works, such as RefineDet [@zhang2018single], DSOD [@shen2017dsod] and STOD [@zhou2018scale], also improve the detection accuracy of SSD either through refinining the anchors or by aggregating the feature maps at different scales. CornerNet [@law2018cornernet] follows a different strategy and improves the detection accuracy with keypoint-based object detectors. The recent work, RetinaNet [@lin2018focal], builds the one-stage detector based on FPN and proposes a focal loss for better training. RetinaNet is efficient in inference; however, it requires a large effort for loss function parameter tuning. In this work, we show that by explicitly modeling the feature relations, our ASSD model competes favorably with RetinaNet without heavy tuning of parameters. Visual Attention ---------------- Visual attention mechanism is generally used to exploit the salient visual information and facilitate visual tasks such as object recognition. There are many visual attention methods in the literature. For example, the saliency-based visual attention model [@itti1998model] selects attended locations from saliency maps. In contrast, RAM [@mnih2014recurrent], AttentionNet [@yoo2015attentionnet] and RA-CNN [@fu2017look] search and crop the useful regions recurrently. In particular, RAM employs Recurrent Neural Network (RNN) and reinforcement learning to discover the target. AttentionNet explores the direction that leads to the real object through CNN classification. RA-CNN also uses reinforcement learning to learn the discriminative region attention and region-based feature representation. The common characteristic of these methods is that they only focus on single instance problems. For multi-object recognition, AC-CNN [@li2017attentive], LPA [@jetley2018learn] and RelationNet [@Hu_2018_CVPR] have been proposed to discover a global contextual guidance. AC-CNN examines the global context through the stacked Long Short-Term Memory (LSTM) units. LPA learns the attention maps from the compatibility scores between the shallow and deep layers. RelationNet correlates the geometry features and appearance information between proposals to generate and forward the attentive features, and it is designed specifically for the two-stage object detectors. In practice, RelationNet only achieves a slight improvement. Self-Attention -------------- The self-attention mechanism has been widely used in natural language processing (NLP) field to model long-range dependencies of a sentence. LSTMN [@cheng2016long] develops an attention memory network that discovers the relations between tokens to enhance the memorization capability of LSTM. Structured self-attentive sentence embedding [@lin2017structured] introduces self-attention in the bidirectional LSTM to generate a 2-D matrix representation of the embeddings, where each row attends to a different part of the sentence. Transformer [@vaswani2017attention] draws global dependencies between input and output based solely on attention mechanisms. Inspired by Transformer, in this work we build the long-range dependencies among all feature pixels within the feature map itself. In a similar spirit to Transformer, our ASSD is capable of attending to different regions for more effective object detection. Attentive SSD {#sec:method} ============= SSD [@liu2016ssd] performs the detection on multi-scale feature maps to handle various object sizes effectively. However, the shallow layer lacks semantic information and is therefore insufficient for detecting small objects. One way to solving this problem is to build more CNN layers to make further refinements of the feature maps or inject semantics from deep layers to the shallow ones exhaustively. Considering that speed is the key advantage of one-stage object detectors, we aim to improve the SSD accuracy with small extra computational cost. To this end, we construct a small network, namely attention unit, and embed it into SSD to improve the detection accuracy. Our ASSD network architecture is illustrated in Fig. \[fig:fig2\]. Specifically, we use ResNet101 (conv1-5) [@he2016deep] as the backbone. The pyramid convolutional blocks (conv6-9) follow the same design as the original SSD [@liu2016ssd]. The feature maps from conv3-9 are used to detect objects with different scales. ASSD places the attention unit between the feature map and the prediction module, where the box regression and object classification are performed. [lcc]{} Layer Name& Output Size & Specifications\ ------------------------------------------------------------------------ conv1 ------------------------------------------------------------------------ & 256$\times$256 & 7$\times$7, 64, stride 2\ conv2\_x & 128$\times$128 & $\begin{bmatrix} 1\times1, 64 \\ 3\times3, 64 \\ 1\times1, 256 \end{bmatrix}\times 3$\ conv3\_x & 64$\times$64 & $\begin{bmatrix} 1\times1, 128 \\ 3\times3, 128 \\ 1\times1, 512 \end{bmatrix}\times 4$\ conv4\_x & 32$\times$32 & $\begin{bmatrix} 1\times1, 256 \\ 3\times3, 256 \\ 1\times1, 1024 \end{bmatrix}\times 23$\ conv5\_x & 8$\times$8 & $\begin{bmatrix} 1\times1, 512 \\ 3\times3, 512 \\ 1\times1, 2048 \end{bmatrix}\times 3$\ Attention Unit -------------- We adapt the self-attention mechanism from the sequence transduction problem [@vaswani2017attention] to our task. In sequence transduction, self-attention mechanism draws global dependencies between the input and output sequences by an attention function, which maps a query and a set of key-value pairs to an output. In self-attention, the attention is motivated by the input features and used for refining these features. Here we repurpose our problem as a similar query problem that estimates the relevant information from the input features in order to build global pixel-level feature correlations. Suppose $ \mathbf{x^s} \in \mathbb{R}^{C^s\times N^s} $ is the feature map at a given scale $ s\in \{1,\cdots,S\} $, with $ C $ and $ N $ representing the number of channels and total spatial locations in the feature map, respectively. We first linearly transform the feature map $ \mathbf{x^s} $ into three different feature spaces $ \mathbf{q}, \mathbf{k} $ and $ \mathbf{v} $, i.e., $ \mathbf{q}(\mathbf{x^s}) = \mathbf{W_q^s}^{\top}\mathbf{x^s} $, $ \mathbf{k}(\mathbf{x^s}) = \mathbf{W_k^s}^{\top}\mathbf{x^s} $, and $ \mathbf{v}(\mathbf{x^s}) = \mathbf{W_v^s}^\top\mathbf{x^s} $, where $\mathbf{W_q^s},\mathbf{W_k^s}\in \mathbb{R}^{C^s\times C^\prime} $ and $ \mathbf{W_v^s}\in \mathbb{R}^{C^s\times C^s} $ with $C^\prime = C^s/8$. The attention score matrix $ \mathbf{a^s}\in \mathbb{R}^{N^s\times N^s} $ is then calculated by the matrix multiplication of $ \mathbf{q}(\mathbf{x^s}) $ and $ \mathbf{k}(\mathbf{x^s}) $, as shown in Fig. \[fig:fig2\]. Each row of the attention score matrix is normalized by a softmax operation: $$\label{eq1} \begin{split} \bar{a}^s_{ij}&= \frac{\exp(a^s_{ij})}{\sum_{j}^{N^s}\exp(a^s_{ij})}, i,j=1,2,\cdots,N^s, \\ \mathbf{a^s} &= \mathbf{q(x^s)}^\top\mathbf{k(x^s)}, \end{split}$$ where $ \mathbf{\bar{a}^s_i} $ describes the pixel relations when querying the $ i $-th location of the feature map. We call $ \mathbf{\bar{a}^s_i} $ as an attention map. Note that, the reason we transform the input feature $\mathbf{x^s}$ into $\mathbf{q}$ and $\mathbf{k}$ is to reduce computational cost. The matrix computation of $ \mathbf{q}(\mathbf{x^s}) $ and $ \mathbf{k}(\mathbf{x^s}) $ calculates the feature similarities and creates an $N\times N$ attention map that reveals the feature relations. Note that such pixel-wise relations are learned through the network. Next, we apply a matrix multiplication between $ \mathbf{v}(\mathbf{x^s}) $ and the attention maps $ \mathbf{\bar{a}^s} $. In this way we compute an updated feature map as the weighted sums of individual features at each location. Finally, we add the matrix multiplication result back to the input feature map $ \mathbf{x^s} $: $$\label{eq2} \mathbf{x^{s^\prime}} = \mathbf{x^{s}}+(\mathbf{\bar{a}^s}\mathbf{v}(\mathbf{x^s})^\top)^{\top}.$$ Attention map $ \mathbf{\bar{a}^s} $ relates the long-range dependencies of features at all positions and therefore learns global contexts of the feature map. It highlights the relevant parts of the feature map and guides the detection with refined information. Semantic Fusion --------------- Motivated by FSSD [@li2017fssd], we fuse the contextual information from layer4 and layer5 into layer3 to enrich its semantics. In our experiment, we find the fusion operation alone does not notably improve the detection accuracy (see Table \[tab:Ablation\]). Instead, it even decreases the accuracy a bit with more computational cost. The reason would be that the three layers possess different receptive fields and have different capabilities; further, the concatenation and $ 1\times1 $ conv transformation would possibly neutralize the relative importance of the three layers and suppress the critical features in original layer3. However, when we place the attention unit after the fusion operation, there is a noticeable improvement (see Table \[tab:Ablation\]). It is possible that semantics from the deep layers help the attention unit to discover useful information that resides in the original layer3. Finally, when only applying the attention unit, we observe inferior performance in contrast to the model with both fusion and attention mechanisms. This indicates that the feature fusion and attention are complementary to each other. The semantic fusion process can be formulated as: $$\mathbf{x^3} = \mathbf{W^3}Concat\{\mathbf{x^3},\mathbf{x^4},\mathbf{x^5}\}+\mathbf{b^3},$$ where $ \mathbf{x^s}\in \mathbb{R}^{C^s\times N^s} $ is the feature map at layer $ s $, $ \mathbf{W^3}\in \mathbb{R}^{C^3\times C^{\prime}} $ and $ \mathbf{b^3}\in \mathbb{R}^{C^3} $. In the concatenation operation, layer4 and layer5 are upsampled through bilinear interpolation in order to align their sizes with that of layer3. {width="100.00000%"} Implementation Details ====================== We follow the same anchor box generating method as SSD [@liu2016ssd]. Specifically, we use aspect ratio $a_r = \{1,2,1/2\} $ for anchor boxes on feature maps conv3,8,9 and $a_r=\{1,2,1/2,3,1/3\}$ for anchor boxes on feature maps of conv4-7. Each box has a minimum scale $s_{\min}$ and a maximum scale $s_{\max}$, where the scale $s_{\min}$ is regularly spaced over the feature map layers and $s_{\max}$ is the $s_{\min}$ of next layer. The normalized width and height of an anchor box are calculated by $w=s\sqrt{a_r}$ and $h=s/\sqrt{a_r}$, where $s=\sqrt{s_{\min}s_{\max}}$ for $a_r=1$, otherwise $s=s_{\min}$. We use hard negative mining to solve the positive-negative box class imbalance problem as in the original SSD [@liu2016ssd]. Also, we employ the same data augmentations and the same loss functions as SSD. Our model is implemented with Pytorch [@paszke2017automatic] and trained on 8 NVIDIA Tesla K80 GPUs. The weights of ResNet101 backbone are pretrained on ImageNet. We use Stochastic Gradient Descent (SGD) algorithm to optimize ASSD weights, with a momentum of 0.9, a decay of 0.0005 and an initial learning rate of 0.001. Following the settings of SSD, DSSD and FSSD, we train and evaluate ASSD on two input resolution images: $321\times 321$ and $513\times 513$. In particular, we set the mini-batch size to 10 images per GPU for ASSD321 and 8 images per GPU for ASSD512. Experiments =========== We conduct experiments on two common datasets: PASCAL VOC [@Everingham15] and COCO [@lin2014microsoft]. The PASCAL VOC dataset contains 20 object classes for object detection challenge. We evaluate ASSD on the PASCAL VOC 2007/2012 test set. The COCO dataset includes 80 object categories. In this work, we use COCO 2017 dataset, which has the same train, validation and test images as COCO 2014. Hence we have a fair comparison with the state-of-the-art methods. Note that RetinaNet [@lin2018focal] does not have PASCAL VOC detection results. Therefore we only compare the accuracy and speed of RetinaNet on COCO dataset. PASCAL VOC 2007 --------------- We first evaluate our ASSD on PASCAL VOC 2007 test set with a primary goal of comparing the speed and accuracy of ASSD with state-of-the-art methods. The training dataset we use here is a union of 2007 trainval and 2012 trainval. We train ASSD321 for 280 epochs, where the initial learning rate of 0.001 decreases by 0.1 at the 200th epoch and the 250th epoch. For ASSD513, we train for 180 epochs, with a learning rate decay of 0.1 at the 120th and 170th epochs. As shown in Table \[tab:VOC2007\], with a comparable fast speed, ASSD achieves a large improvement in accuracy compared to SSD, DSSD, and FSSD. Ablation study on PASCAL VOC 2007 --------------------------------- We perform ablation study to explore the effects of attention unit and semantic fusion on detection accuracy and speed. Here we investigate four models, SSD513, SSD513+fusion, SSD513+att, SSD513+fusion+att, on the PASCAL VOC 2007 test set. It can be observed from Table \[tab:Ablation\] that the fusion module alone does not show noticeable accuracy improvement. On the contrary, it brings a little more computational overhead. In contrast, attention unit alone leads to a significant performance improvement. When combining the attention unit with the fusion module, we observe further boost of performance. We conjecture that the attention unit may have the ability to analyze the contextual semantics at different levels and select the useful information for guiding a better detection. PASCAL VOC 2012 --------------- We compare the detection accuracy of ASSD with the state-of-the-art methods on the PASCAL VOC 2012 test set. The mAP is evaluated by online PASCAL VOC evaluation server. We present a detailed comparison of average precision (AP) for each class in Table \[tab:VOC2012\]. The training dataset contains 2007 trainval+test and 2012 trainval. We follow similar training settings as PASCAL VOC 2007. From Table \[tab:VOC2012\], it can be seen that ASSD513 improves the detection accuracy for most of the classes. The reason would be that the attention unit figures out the pixel-level feature relationships and therefore enhances the model ability to distinguish objects of different classes. COCO ---- We train and validate ASSD on COCO training dataset (118k) and validation dataset (5k). We compare with the state-of-the-art methods on COCO test-dev. The detection performance is evaluated by the online evaluation server. We train ASSD321 for 160 epochs with a learning rate decay of 0.1 at the 100th epoch and the 150th epoch. ASSD513 is trained for 140 epochs, and the learning rate decreases after 80 and 130 epochs. As illustrated in Table \[tab:COCO\], ASSD achieves a large improvement over SSD, DSSD and FSSD. Besides, at a similar input resolution, ASSD513 obtains better accuracies than RetinaNet500, especially for AP at different object area thresholds. In particular, when the intersection over union (IoU) is higher than 0.5, ASSD513 has a 2.4% improvement compared to RetinaNet500. Furthermore, from Table \[tab:COCO\] it can also be observed that ASSD is more effective at detecting the small, medium and large objects. Note that, with the above superiority in detection accuracy, ASSD513 (6.1FPS K40) still achieves comparable speed as RetinaNet500 (6.8FPS K40). Attention Visualization ----------------------- To better investigate the attention mechanism, we visualize the attention maps of different scales. In particular, we project the attention maps onto the original images. Here we utilize the PASCAL VOC 2007 test set, which contains 20 classes. From Fig. \[fig:fig3\], we observe that the attention maps highlight the crucial locations of objects, indicating the feature relations help the model concentrate on useful regions. At shallow layers, the attention map guides the model to focus on small objects; while at deep layers, the attention map highlights objects with large sizes. Moreover, it can also be observed that the attention map suppresses the negative regions, which would be of great help for fast determination of negative anchor boxes. Conclusion ========== In this paper, we propose an attentive single shot multibox detector, termed ASSD, for more effective object detection. Specifically, ASSD utilizes a fast and light-weight attention unit to help discover feature dependencies and focus the model on useful and relevant regions. ASSD improves the accuracy of SSD by a large margin at a small extra cost of computation. Moreover, ASSD competes favorably with the other state-of-the-art methods. In particular, it achieves better performance than the one-stage detector RetinaNet, while being easier to train without the need to heavily tune the loss parameters.

--- abstract: 'The quantification of controllability and observability has recently received new interest in the context of large, complex networks of dynamical systems. A fundamental but computationally difficult problem is the placement or selection of actuators and sensors that optimize real-valued controllability and observability metrics of the network. We show that several classes of energy related metrics associated with the controllability Gramian in linear dynamical systems have a strong structural property, called *submodularity*. This property allows for an approximation guarantee by using a simple greedy heuristic for their maximization. The results are illustrated for randomly generated systems and for placement of power electronic actuators in a model of the European power grid.' author: - 'Fabrizio L. Cortesi, Tyler H. Summers, and John Lygeros[^1]' bibliography: - 'bibtex/ref.bib' title: | Submodularity of Energy Related\ Controllability Metrics --- Introduction ============ The concepts of controllability and observability have long been considered fundamental structural system properties since Kalman’s seminal work over half a century ago [@kalman1959general]. Recently though they have seen renewed interest in the context of large, complex networks. There are many engineering examples such as the electric power grid or transportation networks, but also biological examples like neural or metabolic networks. Many efforts have also been made to understand the complex dynamics in social and economic networks, where there is much interest in controlling the spreading of contagion, such as shocks (e.g., defaults) in financial networks [@gai2010contagion_finance]. In the context of such networks, an interesting problem to consider before the design of a control law is the selection or placement of sensors and actuators in the network. Many approaches for sensor and actuator selection problems have been proposed [@van2001review], but the corresponding combinatorial optimization problems are generally considered to be computationally difficult. Here we will consider the combinatorial structure of these problems that allows approximation guarantees using metrics based on the controllability and observability Gramians. In a prominent article [@liu2011controllability], Kalman’s rank criterion together with the concept of structural controllability [@lin1974structural] is used to find the minimal set of driver nodes, the smallest set of nodes that have to be controlled individually in order to achieve full control over the entire network. But this results in a rather simplified and sometimes misleading view, as the rank criterion can only considers whether a system is completely controllable/observable, regardless of other considerations such as the energy required to actually drive the system around the state space. The controllability Gramian quantifies the energy required to move the system around the state space. One can consider several scalarizations, such as the maximum eigenvalue, trace and determinant of the inverse Gramian [@muller1972analysis]. Recently [@tylerquote] used the minimum eigenvalue to select which nodes to control, and also a trade-off between control energy and the minimum number of driver nodes is presented. The Gramian can be computed by solving a Lyapunov equation. There are specialized methods[@hammarling1982numerical], including recent approximations based on Quantized Tensor Trains [@nipdirect], that enable computation for very large networks. Although selecting subsets of possible actuators to maximize these metrics is typically computationally hard, some combinatorial structure can improve the situation. For example, the trace of the controllability Gramian was recently found to be a modular set function [@summers2013optimal], which allows globally optimal subsets of actuators to be obtained. In the present paper, we identify what several other metrics of the controllability Gramian that have a strong structural property that give approximation guarantees. Our main result is that 1. $-\operatorname{tr}(W_c^{-1})$ 2. $\log \det W_c$ 3. $\operatorname{rank}(W_c)$ are submodular set functions. Submodularity has been recognized in many areas in machine learning and computer science, such as [@krause2007near]. Together with [@summers2013optimal], the present paper is among the first to study submodularity in controllability problems. This paper is structured as follows: Section II recalls some basics of linear systems theory and discusses several energy related measures for controllability. Section III gives a brief introduction into submodular functions and their maximization under cardinality constraint. Then a framework for actuator selection in the context of set functions is presented, and we present our main results on submodularity of controllability metrics. Finally in section V we illustrate the results on numerical examples. Energy-related Controllability Measures ======================================= This section reviews several classes of energy-related controllability metrics for dynamical networks based on the controllability Gramian of a linear model for the network dynamics. Of course, real networks have nonlinear dynamics; however, the combinatorial properties of optimal sensor and actuator selection problems, which is the main topic of this paper, are not understood even for linear systems. Furthermore, linear models are often used to approximate nonlinear models near an equilibrium point and have been used as a starting point in many recent papers on controllability in networks. Because controllability and observability are dual properties [@kalman1959general], we will only discuss controllability; analogous results and interpretations for observability follow directly. Linear System Dynamics ---------------------- We consider linear, continuous time, time-invariant dynamical systems of the form $$\label{eq:sys} \begin{aligned} \dot{x}(t) &= Ax(t) + Bu(t) \end{aligned}$$ where $A\in \mathbb{R}^{n\times n}$ is the dynamic matrix and $B\in \mathbb{R}^{n\times m}$ is the input matrix. The basic problem we consider can be formulated as follows. Let $V = \{ b_1,...,b_M\} $ be a set of possible columns corresponding to actuators that could be selected to form the input matrix. We consider the set function optimization problem $$\label{eq:maxsubmod} \max_{S\subseteq V} f(S)\qquad \text{subject to } |S|\leq k,$$ where $f : 2^V \rightarrow \mathbb{R}$ is a metric that quantifies to the controllability of the system when choosing subset $S \subseteq V$. Actuators in real systems are usually energy limited, so an important class of metrics of controllability deals with the amount of input energy required to reach a given state from the origin. In particular, we can pose the following optimal control problem seeking the minimum energy input that drives the system from the origin to a final state $x_f$ at time $t$: $$\label{optprob} \begin{aligned} \underset{{u(\cdot) \in \mathcal{L}_2 }}{\text{minimize}} & &&\int_0^t \Vert u(\tau) \Vert^2 d\tau \\ \text{subject to} & && \dot{x}(t) = Ax(t) + Bu(t), \\ & && x(0) = 0, \quad x(t) = x_f \end{aligned}$$ Standard methods from optimal control theory can be used to derive the solution. If the system is controllable, the optimal input has the form $$\begin{aligned} u^*(\tau) = B^T e^{A^T(t-\tau)} &\left( \int_0^t e^{A \sigma} B B^T e^{A^T \sigma} d\sigma \right)^{-1} x_f, \\ &0 \leq \tau \leq t \end{aligned}$$ and the resulting minimum energy is $$\int_0^t \Vert u^*(\tau) \Vert^2 d\tau = x_f^T \left( \int_0^t e^{A \sigma} B B^T e^{A^T \sigma} d\sigma \right)^{-1} x_f.$$ The symmetric positive semidefinite matrix $$W_c(t) = \int_0^t e^{A \tau} B B^T e^{A^T \tau} d\tau \ \in \mathbf{R}^{n\times n}$$ is called the *controllability Gramian* at time $t$. The controllability Gramian has the same rank as $[B, AB, ..., A^{n-1} B]$ and defines an ellipsoid in the state space $$\mathcal{E}_{min}(t) = \left\{ x \in \mathbf{R}^n \mid x^T W_c(t)^{-1} x \leq 1 \right\}$$ that contains the set of states reachable in $t$ seconds with one unit or less of input energy. The eigenvectors and corresponding eigenvalues of $W_c$ quantify energy required to move the system in various state space directions. For stable systems, the state transition matrix $e^{At}$ consists of decaying exponentials, so a finite positive definite limit of the controllability Gramian always exists and is given by $$W_c = \int_0^\infty e^{A \tau} B B^T e^{A^T \tau} d\tau \ \in \mathbf{R}^{n\times n}$$ This matrix defines an ellipsoid in the state space that gives the states reachable with one unit or less of energy, regardless of time. This infinite-horizon controllability Gramian can be computed by solving a Lyapunov equation $$A W_c + W_c A^T + B B^T = 0,$$ which is a system of linear equations and is therefore easily solvable, even for large systems. Specialized algorithms have been developed to compute the solution [@hammarling1982numerical]. For unstable system, an alternative definition of the controllability Gramian can be used [@balancedgramian]. We still need scalarize $W_c$, which is a positive semidefinite matrix. We want $W_c$ “large” so that $W_c^{-1}$ is “small”, requiring small amount of input energy to move around the state space. There are a number of possible metrics for the size of $W_c$, several of which we now discuss. ### $\mathbf{tr}(W_c)$ The trace of the Gramian is inversely related to the average energy and can be interpreted as the average controllability in all directions in the state space. It is also closely related to the system $H_2$ norm: $$\begin{aligned} \Vert H \Vert_2^2 &= \textbf{tr} \left( C \int_0^\infty e^{A t} B B^T e^{A^T t} dt C^T \right) \\ &= \textbf{tr} (C W_c C^T) \end{aligned}$$ i.e., the system $H_2$ norm is a weighted trace of the controllability Gramian. ### $\mathbf{tr}(W_c^{-1})$ The trace of the inverse controllability Gramian is proprtional to the energy needed on average to move the system around on the state space. Note that when the system is uncontrollable, the inverse Gramian does not exist and the average energy is infinite because there is at least one direction in which it is impossible to move the system using the inputs. In this case, one could consider the trace of the pseudoinverse, which is the average energy required to move the system around the controllable subspace. ### $\log \det W_c$ The determinant of the Gramian is related to the volume enclosed by the ellipse it defines $$V(\mathcal{E}_{min}) = \frac{\pi^{n/2}}{\Gamma(n/2+1)}\sqrt[n]{\det W_c},$$ where $\Gamma$ is the Gamma function. This means that the determinant is a volumetric measure of the set of states that can be reached with one unit or less of input energy. Since determinant is numerically problematic in high dimensions, and because the matrix logarithm is a monotone matrix function that preserves the semidefinite order, we will consider optimizing $\log \det W_c$. Note that for uncontrollable systems, the ellipsoid volume is zero, so $\log \det W_c = -\infty$. In this case, one could consider the associated volume in the controllable subspace. ### $\lambda_{min} (W_c)$ The smallest eigenvalue of the Gramian is a worst-case metric inversely related to the amount of energy required to move the system in the direction in the state space that is most difficult to control. ### $\operatorname{rank}(W_c$) The rank of the Gramian is the dimension of the controllable subspace. Optimal Actuator Placement ========================== In this section, we briefly review submodularity and consider which of the above controllability metrics are submodular, which provides approximation guarantees for associated actuator selection problems. Detailed treatments of the topic are provided by many texts [@lovasz83; @fujishige2005submodular; @schrijver2003combinatorial; @iwata2008submodular], as well as by recent survey [@google] which also focuses on maximization. Submodularity Basics -------------------- A set function $f\colon 2^V \to \mathbb{R}$ assigns a real value to every subset $S\subseteq V$. Submodularity is defined as follows. \[def:submod\] A set function $f\colon 2^V \to \mathbb{R}$ is submodular if and only if $$f(A\cup \{a\})-f(A)\geq f(B\cup \{a\})-f(B),$$ for any subsets $A\subseteq B\subseteq V$ and $\{a\}\in V\backslash B$. This is a diminishing gains property: adding an element $a$ to a larger set $B$ will result in a smaller gain than adding the same element to a subset $A$. We will denote this the marginal gain as ${\Delta(\,a\,|\,S\,)}=f(S\cup\{a\})-f(S).$ A useful consequence of this definition is expressed in the following theorem A set function $f: 2^V \rightarrow \mathbf{R}$ is submodular if and only if the derived set functions $f_a : 2^{V - \{ a \} } \rightarrow \mathbf{R}$ $$f_a (X) = f(X \cup \{a\}) - f(X)$$ are monotone decreasing for all $a \in V$. A set function $f_a(S)$ is called monotone decreasing if for all $S_1$, $S_2\subseteq V$ holds, that $$S_1\subseteq S_2 \quad \Leftrightarrow\quad f_a(S_1)\geq f_a(S_2).$$ Submodular Function Maximization -------------------------------- We consider set function optimization problems of the form $$\label{eq:maxsubmod} \max_{S\subseteq V} f(S)\qquad \text{subject to } |S|\leq k.$$ Maximization of monotone submodular functions is NP-hard, but the so-called greedy heuristic can be used to obtain a solution that is provably close to the optimal solution. The greedy algorithm for starts with an empty set, $S_0 \leftarrow \emptyset$, computes the gain ${\Delta(\,a\,|\,S_i\,)} = f(S_{i}\cup \{a\})-f(S_{i})$ for all elements $a\in V\backslash S_{i}$ and adds the element with the highest gain: $$S_{i+1} \leftarrow S_{i}\cup \{\arg \max_{a} {\Delta(\,a\,|\,S_i\,)}\; |\; a\in V\backslash S_{i}\}.$$ The algorithm terminates after $k$ iterations. Performance for a submodular, non-negative, monotone set function is guaranteed by the well known bound [@greedybound]: $$\dfrac{f(S_{greedy})}{f(S_{optimal})} \geq 1-\left(\frac{k-1}{k}\right)^k \geq \frac{e-1}{e} \approx 0.63, \label{eq:greedy_bound}$$ where $S_{optimal}$ is the optimal subset and $S_{greedy}$ is the subset obtained by the greedy algorithm. This is the best any polynomial time algorithm can achieve [@feige1998threshold], assuming $P\neq NP$. Note that this is a worst-case bound; the greedy algorithm often performs much better than the bound in practice. The space of symmetric $n\times n$ matrices $\mathcal{S}^n$ has a partial semidefinite order: $W_1\succeq W_2$ if $W_1-W_2\succeq 0$. The space of symmetric positive definite matrices is denoted $\mathcal{S}_{++}^n$ and the space of symmetric positive semidefinite matrices is denoted $\mathcal{S}^n_{+}$. We now demonstrate the submodularity of several classes of controllability metrics involving functions of the controllability Gramian. Trace of the inverse Gramian ---------------------------- Suppose $A \in \mathbf{R}^{n\times n}$ is a stable system dynamics matrix and $V = \{b_1,..., b_M \} $ is a set of possible columns that can be used to form or modify the system input matrix $B$. The problem is to choose a subset of $V$ to maximize a metric of controllability. For a given $S \subseteq V$, we form $B_S = [B_0 \quad b_s]$ given a (possibly empty) existing matrix $B_0$ and using the associated columns defined by $s \in S$ and the associated controllability Gramian $W_S = \int_0^\infty e^{A \tau} B_S B_S^T e^{A^T \tau} d\tau =\sum_{s\in S}W_{\{s\}}$, which is the unique positive semidefinite solution the Lyapunov equation $$AW_S + W_S A^T + B_S B_S^T = 0.$$ To simplify notation, we write $W_s$ for $W_{\{s\} }$. We first consider the trace of the inverse of the controllability Gramian. We assume in this subsection that for any $S \subseteq V$ the associated Gramian $W_S$ is invertible. This is the case, for example, if the network already has a set of actuators that provide controllability and we would like to add additional actuators to improve controllability. We discuss how to deal with non-invertibility of the Gramian subsequently. \[theorem:trace\] Let $V = \{ b_1,...,b_M\}$ be a set of possible input matrix columns and $W_S$ the controllability Gramian associated with $S \subseteq V$. The set function $f : 2^{V}\to \mathbf{R}$ defined as $$f(S) = -\operatorname{tr}(W_{S}^{-1})$$ is submodular. Before proving Theorem \[theorem:trace\], we need a result on differentiable matrix functions that are monotone with respect to the semidefinite order. \[def:mon\] A matrix function $f : \mathbf{S}^n_{++}\to\mathbf{R}$ is called monotone decreasing if $$W_1\preceq W_2 \quad \Rightarrow \quad f(W_1)\geq f(W_2) \qquad \forall\, W_1, W_2 \in \mathbf{S}^n_{++}.$$ \[theorem:mono\] Let $f\colon\mathbf{S}^n_{++}\to\mathbf{R}$ be continuously differentiable matrix function and $X(t)=A+tB$ a ray in the space of symmetric positive definite matrices with arbitrary $A \in \mathbf{S}^n_{++}$, $B\in \mathbf{S}^n_+$, and $t > 0$. Then $f$ is monotone decreasing if $$B\succeq 0 \quad \Rightarrow \quad \frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}f(X(t))\leq 0\qquad \forall t.\label{eq:mono}$$ *Proof:* Take any $W_1$, $W_2\in\mathbf{S}^n_{++}$ and assume $W_1\preceq W_2$. Then there exist $t_1, t_2\in \mathbf{R}$, $A \succ 0$, $B \succeq 0$, and $t_1\leq t_2$ such that $X(t_1)=W_1$ and $X(t_2)=W_2$. If $B\succeq 0$ implies $\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}f(X(t))\leq 0$, then from $$f(X(t_2)) = f(X(t_1)) +\int_{t_1}^{t_2}\!\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}f(X(t)){\mathop{}\!\mathrm{d}}t$$ it follows that $f(W_1)\geq f(W_2)$. *Proof of Theorem \[theorem:trace\]:* Consider the derived set functions $f_a: 2^{V-\{a\}} \to \mathbf{R}$ $$\begin{aligned} f_a(S) &= -\operatorname{tr}((W_{S\cup \{a\} })^{-1}) + \operatorname{tr}((W_{S})^{-1}) \\ &= -\operatorname{tr}((W_{S}+W_{a})^{-1}) + \operatorname{tr}((W_{S})^{-1}). \end{aligned}$$ and the derived matrix functions $\hat{f}_a : \mathbf{S}_{++}^n \to \mathbf{R}$ defined by $\hat{f}_a(W_{S}) = f_a(S).$ Since $S_1\subseteq S_2\Leftrightarrow W_{S_1} \preceq W_{S_2}$, it is clear that $f_a(S_1)\geq f_a(S_2)\Leftrightarrow \hat{f}_a(W_{S_1}) \geq \hat{f}_a(W_{S_2})$ and therefore if $\hat{f}_a$ is monotone decreasing, then so is $f_a$. Let $X(t)=A+tB$ with $A\succ 0$, $B\succeq 0$ and $t>0$. We have $$\begin{aligned} &\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\hat{f}_a\left(X(t)\right) \\ &=-\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\left[ \operatorname{tr}((A+tB+W_{a})^{-1}) + \operatorname{tr}((A+tB)^{-1}) \right]\\ &=\operatorname{tr}\left((A+tB+W_{a})^{-1}B(A+tB+W_{a})^{-1} \right) \\ & \quad - \operatorname{tr}\left((A+tB)^{-1}B(A+tB)^{-1}\right)\\ &=\operatorname{tr}\bigg( \left((A+tB+W_{a})^{-2}-(A+tB)^{-2}\right)B \bigg)\,. \end{aligned}$$ where we used a standard matrix derivative formula and the cyclic property of trace to obtain the second and third equalities. Since $A+tB+W_{a}\succeq A+tB$, it follows that $(A+tB+W_{a})^{-2}\preceq(A+tB)^{-2}$. Thus, $$\Gamma=(A+tB+W_{a})^{-2}-(A+tB)^{-2}\preceq 0.$$ Since the trace of the product of a positive and negative semidefinite matrix is non-positive, we have $$\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\hat{f}_a\left(X(t)\right)=\operatorname{tr}(\Gamma B)\leq 0.$$ Thus, $\hat{f}_a$ and hence $f_a$ are monotone decreasing, and $f$ is submodular by Theorem \[theorem:mono\]. Log determinant of the Gramian ------------------------------ We now consider the log determinant of the controllability Gramian. We assume again that for any $S \subseteq V$ the associated Gramian is invertible. We have the following result. \[theorem:determinant\] Let $V = \{ b_1,...,b_M\}$ be a set of possible input matrix columns and $W_S$ the controllability Gramian associated with $S \subseteq V$. The set function $f\colon 2^{V}\to \mathbf{R}$, defined as $$f(S) = \log \det W_{S}$$ is submodular. The proof uses the same idea as before, namely, showing monotonicity of a family of derived matrix functions. Consider the derived set functions $f_a: 2^{V-\{a\}} \to \mathbf{R}$ $$\begin{aligned} f_a(S) &= \log \det W_{S \cup \{a\} } - \log \det W_{S} \\ &= \log\det (W_{S}+W_{a}) - \log\det W_{S} \end{aligned}$$ and the associated matrix functions $\hat{f}_a : \mathbf{S}_{++}^n \to \mathbf{R}$ defined by $\hat{f}_a(W_{S}) = f_a(S).$ Let $X(t)=A+tB$ with $A\succ 0$, $B\succeq 0$ and $t>0$. We have $$\begin{aligned} \frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t} &\hat{f}_a(X(t)) \\ &=\frac{{\mathop{}\!\mathrm{d}}}{{\mathop{}\!\mathrm{d}}t}\left[ \log \det (A+tB + W_a) - \log \det (A+tB) \right] \\ &=\operatorname{tr}((A+tB+W_{a})^{-1}B)-\operatorname{tr}((A+tB)^{-1}B)\\ &=\operatorname{tr}(((A+tB+W_{a})^{-1}-(A+tB)^{-1})B)\\ &\leq 0.\end{aligned}$$ where we used again a standard matrix derivative formula and the cyclic property of trace to obtain the second and third equalities. The remainder of the proof follows the previous proof. The related set function $g: 2^V \rightarrow \mathbf{R}$ defined by $g(S)=\log\sqrt[n]{\det{W_S}}$ is submodular. *Proof:* We have $$g(S)= \frac{1}{n} \log \det W_S$$ Thus, from Theorem \[theorem:determinant\] $g$ is a non-negatively scaled version of a submodular function and therefore also submodular. Not all directions in the state space may be of equal importance, one might want to use a weight matrix as an additional design parameter for an actuator selection problem. In a simple case, the weight matrix could be a diagonal matrix, assigning a relative weight to every state. We have the following corollary; the proof follows exactly the same argument as in the previous theorem. Let $V = \{ b_1,...,b_M\}$ be a set of possible input matrix columns and $W_S$ the controllability Gramian associated with $S \subseteq V$. The set function $g\colon 2^V \to \mathbf{R}$ defined as $$g(S)=\log\det(QW_SQ^T),$$ where $Q\in\mathbf{R}^{m\times n}$ with $m\leq n$ and $\operatorname{rank}(Q)=m$, is submodular. Rank of the Gramian ------------------- The controllability metrics $-\operatorname{tr}(W_S^{-1})$ and $\log \det W_S$ fail to distinguish amongst subsets of actuators that do not yield a fully controllable system. In particular, these functions are undefined, or are interpreted to return $-\infty$, when the Gramian is not full rank. One way to handle these cases is to first consider the rank of the controllability Gramian. The following result shows that this is also a submodular set function. Let $V = \{ b_1,...,b_M\}$ be a set of possible input matrix columns and $W_S$ the controllability Gramian associated with $S \subseteq V$. The set function $f\colon 2^V \to \mathbf{R}$, defined as $$f(S) = \operatorname{rank}(W_S)$$ is submodular. For two linear transformations $V_1$, $V_2 \in \mathbf{R}^{n\times n}$, we have $$\begin{aligned} \operatorname{rank}& (V_1+V_2) \\ &= \operatorname{rank}(V_1) + \operatorname{rank}(V_2) -\dim(\operatorname{im}(V_1) \cap \operatorname{im}(V_2)). \end{aligned}$$ We can form gain functions $f_a: 2^{V-\{a\} } \rightarrow \mathbf{R}$ $$\begin{aligned} f_a(S)&=\operatorname{rank}(W_{S \cup \{a\} }) - \operatorname{rank}(W_S) \\ &=\operatorname{rank}(W_{a}) - \dim(\operatorname{im}(W_S)\cap\operatorname{im}(W_{a})) \end{aligned}$$ It is now easy to see that $f_a$ is monotone decreasing: the first term in the second line is constant and the second term decreases because $\operatorname{im}(W_S)$ only increases with $S$. Another way to handle uncontrollable systems is to work with related continuous metrics defined for uncontrollable systems, such as the trace of the pseudoinverse $\operatorname{tr}(W_S^\dag)$, which corresponds to the average energy required to move the system around the controllable subspace, or the log product of non-zero eigenvalues $\log\Pi_{i=1}^{\operatorname{rank}W_S} \lambda_i(W_S)$, which relates to the volume of the subspace reachable with one unit of input energy. Smallest eigenvalue of the Gramian ---------------------------------- We have seen so far that the trace of the Gramian is a modular (and thus both sub- and supermodular) set function of actuator subsets and that the trace of the inverse Gramian and the log determinant of the Gramian are submodular set functions. These functions are also all concave matrix functions of the Gramian. Given the connections between submodular functions and concave functions, one might be tempted to conjecture that any concave function of the Gramian, e.g. the smallest eigenvalue $\lambda_{min}(W_S)$, corresponds to a submodular function of actuator subsets. However, we now show by counterexample that this is false. We show an example where this function violates the diminishing gains property of a set function $f(S)$ $$\label{eq:supermodular} {\Delta(\,s\,|\,A\,)}\geq{\Delta(\,s\,|\,B\,)}, \quad A\subseteq B \subseteq V,\; s\notin B,$$ where ${\Delta(\,s\,|\,A\,)} = f(A \cup \{s\}) - f(A).$ Consider the system defined by $$A = \begin{bmatrix} -8 &0 & -2\\ 0 & -2 & -8\\ 7& 0 & -3 \end{bmatrix},\qquad B_V=[b_V]=I_3.$$ We see that the diminishing returns property holds in some cases, e.g., $${\Delta(\,b_3\,|\,\{b_1\}\,)}=0.037 \geq 0.033 = {\Delta(\,b_3\,|\,\{b_1,b_2\}\,)},$$ but is violated in others $${\Delta(\,b_3\,|\,\{b_2\}\,)}=0.001 \leq 0.033 = {\Delta(\,b_3\,|\,\{b_1,b_2\}\,)}.$$ Illustrative Numerical Examples {#sec:examples} =============================== In this section, we illustrate the results on randomly generated systems and for placement of power electronic actuators in a model of the European power grid. The problem data is a system dynamics matrix $A \in \mathbf{R}^{n\times n}$, a set of possible input matrix columns $V = \{ b_1,...,b_M \}$, and an integer number $k$ of actuators to choose from this set to form an input matrix that maximizes a controllability metric. Since the trace of the inverse Gramian and the log determinant do not distinguish amongst actuator subsets that result in uncontrollable systems, we use for these metrics a two-stage greedy algorithm that first greedily optimizes the rank of the Gramian until controllability is achieved. During this stage, if two or more actuators yield a Gramian with the same rank, the trace of the pseudoinverse $\operatorname{tr}(W_S^\dag)$ or the log product of non-zero eigenvalues $\log\Pi_{i=1}^{\operatorname{rank}W_S} \lambda_i(W_S)$ is used to determine preference. After controllability it achieved, the trace of the inverse and log determinant are used in a second stage until the desired number of actuators has been selected.. We assume that there exists a subset of $V$ of size $k$ that renders the system fully controllable. However, although a real-valued controllability metric is optimized, the greedy algorithm does not necessarily guarantee that full controllability is achieved after $k$ iterations; if and how rank constraints could be incorporated is an interesting topic for future work. Greedy performance on a random system ------------------------------------- To evaluate performance of the greedy algorithm and to compare the various controllability metrics, we first consider randomly generated data. We use Matlab’s `rss` routine to generate a stable dynamics matrix with random stable eigenvalues. We use $V=\{e_1,...,e_n\}$, where $e_i$ is the $i$th unit vector in $\mathbf{R}^n$, i.e., we assume one can choose states in which a control input can be injected. Figure \[fig:greedy\_spread\] shows the result of applying the greedy algorithm to maximize the log determinant metric with $n=25$ and $k=7$. This problem is small enough to evaluate every possible 7-element actuator subset, and this result is also shown in a histogram. Our algorithm finds a good set $S_{greedy}$ scoring $$\frac{f(S_{greedy})}{f(S_{opt})}\approx98\%$$ of the absolute optimum value, which is better than 99.93% of all other possible choices. Those few actuator choices that score better though, are significantly better with respect to the reachable subspace volume, where we only score $$\frac{V_{greedy}}{V_{opt}} =\sqrt{\frac{e^{f(S_{greedy})}}{e^{f(S_{opt})}}} \approx 68.1\%.$$ Similar results are obtained with other randomly generated data. ![\[fig:greedy\_spread\]Histogram displaying the positive log determinant metric $f'(S)$ for all possible selection of 7 actuators from a set of possible 25. The result achieved by greedy optimization is displayed by the red line.](greedy-spread){width="\linewidth" height="4cm"} Power Electronic Actuator Selection ----------------------------------- In this second example we consider the placement of HVDC (high voltage direct current) links in a simplified model of the Continental European power grid. HVDC lines connect two points in the alternating current grid via a direct current line which allows the injection of both active and reactive power into the grid on both ends of the line. They can be used to increase transport capacity and to improve transient stability, power oscillation damping, and voltage stability control [@hvdc]. The system dynamics are based on the well known swing equations, which describe the generator rotor dynamics [@bookPowerSystems]. The model consists of 74 buses which are each connected to a generator and a constant impedance load. The system is based on a midday operation snapshot of the actual grid and a simplification of the model presented in [@haase]. The HVDC links are modeled as two ideal current sources, one on each end, allowing for three degrees of freedom to influence the frequency dynamics at the terminals; for modeling details see [@g3; @g1; @g2]. The swing equations are linearized around a nominal operating point. Each actuator results in a slightly different linearized dynamics matrix, but all are within 1% in Frobenius norm. Therefore, we use a constant $A$ taken from one of the linearizations. ![Map of the European power grid model, red nodes represent the buses that are interconnected via the blue HVDC lines placed to maximise $\log \det W_c$.[]{data-label="fig:map_res"}](map_lines_paper){width="\linewidth"} An HVDC link can connect any of the 74 nodes leaving a choice of 2701 possible HVDC locations. Choosing a subset of size 10 results in approximately $5.6 \times 10^{27}$ possible combinations, which is far too many for brute force evaluation. Figure \[fig:map\_res\] show the placement obtained by using the two-stage greedy algorithm with the log determinant metric. Compared to the trace metric used in [@summers2013optimal], we see that the lines are in general longer, connecting buses that are further apart, and more evenly distributed in the network. Also, no node is part of more than one HVDC line. This is due to focus of the trace metric mainly on the few directions in which controllability can be increased the most, whereas the determinant is focused more on the overall performance in all directions. We emphasize that this model does not consider any other factors such as political, geographical or financial aspects. The greedy algorithm was implemented in Matlab, building upon the Submodular Function Optimization toolbox [@toolbox]. Conclusion ========== We have considered optimal actuator placement problems in complex dynamical networks showed that several important class of metrics related to the controllability Gramians, viz. the log determinant and rank for the Gramian, and the trace of the inverse Gramian, result in submodular set functions. This allows approximation guarantees to be obtained with a simple greedy algorithm. By duality, all of the results hold for corresponding sensor placement problems using metrics of the observability Gramian. The results were illustrated in problems with randomly generated data and via placement of power electronic actuators in a model of the European power grid. Acknowledgment {#acknowledgment .unnumbered} ============== The authors would like to thank Dr. Alexander Fuchs for providing details and helpful discussion about the power grid model discussed in section \[sec:examples\]. [^1]: The authors are with the Automatic Control Laboratory, ETH Zurich, Switzerland.

--- author: - 'Alejandro Borlaff, M. Carmen Eliche-Moral, John Beckman and Joan Font' bibliography: - 'borlaff\_14.bib' title: 'Type-II surface brightness profiles in edge-on galaxies produced by flares' --- Introduction ============ {width="\textwidth"} \[fig:Str\_photo\_params\] Disc galaxies often do not show a single radial exponential surface brightness profile [@1970ApJ...160..811F]. and @2008AJ....135...20E classified the discs of galaxies into three classes according to the shape of their profiles. Type-I discs are well modelled with a single exponential profile. Type-II discs present a down-bending profile, i.e. a brightness deficit in the outer parts of the disc with respect to the extrapolation of the inner profile outside a given “break radius”. Finally, the profiles of Type-III discs are shallower outside the break radius than the extrapolation of the inner exponential profile (antitruncation). In the present study we focus on Type-II profiles. Sharp truncations in galaxy disc profiles were first reported in edge-on galaxies . Owing to projection effects, edge-on galaxies present a larger number of truncations and Type-II breaks than their face-on counterparts [@2007MNRAS.378..594P]. Recent studies associate sharp truncations with stellar density cut-offs and inner down-bending breaks to different stellar formation thresholds [@2012MNRAS.427.1102M]. @2014MNRAS.441.1992L found an interesting correlation between the break in Type-II profiles and the characteristic radii of structures in the galaxy: lenses, rings or pseudorings. This means that these morphological features may be associated with the break in some Type-II profiles. Previous studies also explored the possibility that flares (i.e. an increase in the scale height of the stellar disc with galactocentric radius) may be responsible for the down-bending profiles detected in highly inclined galaxies. More than a decade ago, @2002MNRAS.334..646K performed 2D decompositions of 34 edge-on late-type galaxies; 60% of them showed Type-II profiles. The authors studied whether a linear radial increase in the scale height of the disc could produce some of the detected truncations if the discs were observed at high inclination. They created a set of artificial images ($\sim$50 simulations) using a random uniform distribution covering the observed range of values of the disc central surface brightness ([$\mu_\mathrm{0}$]{}), disc scale length ([$h_{R}$]{}), disc scale height ([$h_{z}$]{}), bulge effective surface brightness ([$\mu_\mathrm{e}$]{}), bulge effective radius ([$r_\mathrm{e}$]{}), and bulge axial ratio ($q$) of their data. They also restricted the bulge-to-disc ratio to ${\ensuremath{B/D}}< 2$. Finally, they concluded that values of the radial gradient of the scale height larger than 1 [$h_{z}$]{} per [$h_{R}$]{} are needed – according to their models – to reproduce the characteristics of the observed truncations. However, this gradient in late-type galaxies is typically very low , so @2002MNRAS.334..646KVa concluded that flaring is a very unlikely explanation for the breaks observed in late-type galaxies. Radial variations of the scale height of galaxy discs are difficult to measure [@2015MNRAS.451.2376M], but as better and deeper data are obtained, higher values of the radial gradient of the scale height ([$\nabla_{R} h_{z}$]{}) are being derived even for late-type galaxies [@2014ApJ...787...24B]. The Milky Way (MW, which is thought to be an SBbc galaxy) exhibits a thin+thick disc structure, both affected by a particularly strong flare . Cosmological simulations also predict much stronger flares for old stellar discs than current data reveal [@2015ApJ...804L...9M]. These authors have pointed out that the phenomenon may be unavoidable on the outskirts of galaxies. Consequently, the effects of flares on the disc profiles of edge-on galaxies may have systematically been underestimated in previous studies. No study has analysed systematically whether flares can produce Type-II profiles in highly inclined discs, using realistic models for the bulge, the disc, and (especially) the flare. So we have revisited this question to quantify the possible effects of realistic flares on the generation of breaks in edge-on disc profiles. To this end, we used observations for different morphological types and masses to create 3D models of disc galaxies assuming distributions of [$B/D$]{}, bulge and disc photometric parameters ([$\mu_\mathrm{0}$]{}, [$h_{R}$]{}, [$\mu_\mathrm{e}$]{}, [$r_\mathrm{e}$]{}, Sérsic index $n$), scale heights of the discs, and their trend with the galactocentric radius. To properly sample the complete range of parameters, we performed 10,000 simulations of realistic flared disc galaxies. We derived edge-on and face-on surface brightness profiles for the resulting flared galaxy models. The surface brightness profile of each galaxy model was analysed to find out whether the flare produces a significant break in the disc in the edge-on view compared with the face-on view. Methods ======= We created 10,000 3D models of galaxies, each with an exponential disc plus a Sérsic bulge. We adopted the following functional form for the exponential disc : $$\label{eq:Corredoira} \rho(R,z)=\rho_{\mathrm{0}}\cdot\exp{\Bigg(\frac{-R}{h_{R}}\Bigg)}\cdot \exp{\Bigg(\frac{-|z|}{h_{z}(R)}\Bigg)}\cdot \frac{h_{z}(R)}{h_{z}(0)} \ .$$ Following the observations, we explored the effect of a linear increase in the vertical scale height, as follows: $$\label{eq:Flares_linear} h_{z}(R) = \begin{cases} h_{z}(0) & \text{if } R \leq R_{\mathrm{flare}} \\ h_{z}(0) + {\ensuremath{\nabla_{R} h_{z}}}\cdot R & \text{if } R > R_{\mathrm{flare}}. \end{cases}$$ @2001AJ....121..820G analysed a sample of 86 face-on disc-dominated galaxies previously selected by . This author performed a bulge + disc decomposition for 69 galaxies in the $I$ band, correcting for the effects of the internal extinction, Galactic extinction, inclination, and cosmological dimming [@2003AJ....125.3398G], that we used as a reference for our models. We estimated stellar masses using the relationship between the $V$-band mass-to-light ratio of galaxies and their dust-corrected rest-frame colours derived by @2013MNRAS.431..430W. According to these authors, for $z<0.1$ the optimal observed colour is $(B-V)$, so we have estimated this colour for the 69 galaxies of @2003AJ....125.3398G from HyperLeda data[^1], and estimated the stellar mass of each galaxy using the relations in @2013MNRAS.431..430W For those objects without $(B-V)$ available in HyperLeda, we estimated them from their SDSS $(g-r)$ colour following the transformations published in @2005AJ....130..873J. To simulate realistic images of the disc galaxies, we adopted the observational $I$-band distributions of @2001AJ....121..820G [@2003AJ....125.3398G] for the photometric parameters ([$r_\mathrm{e}$]{}, Sérsic index $n$, [$h_{R}$]{}, [$\mu_\mathrm{0}$]{}, [$\mu_\mathrm{e}$]{}, $B/T$, the absolute magnitudes of the disc $M_{abs,disc}$ and the bulge $M_{abs,bulge}$), and four morphological type bins (S0–Sa, Sb–Sbc, Sc–Scd and Sd–Sdm), in three mass bins ($10 < \log_{10}M/M_{\odot} < 10.7, 10.7 < \log_{10}M/M_{\odot} < 11$ and $\log_{10}M/M_{\odot} > 11$) in order to explore realistic mass distributions for the morphological type bins. In Fig.\[fig:Str\_photo\_params\] we represent the distributions of the structural and photometric parameters from which we created the models and compared them to the observations they are based on. For each morphological type bin we randomly chose the ratio of scale height to scale length ([$\nicefrac{h_{z}}{h_{R}}$]{}) from the observational range of values corresponding to each type reported by @2002MNRAS.334..646K and @2015MNRAS.451.2376M in the $I$ band, as shown in Fig.\[fig:Str\_photo\_params\]. ![**Top panels:** Edge-on $I$-band images for the flared and non-flared models. **Intermediate panels:** The same for the face-on discs. **Bottom panels:** Surface brightness profiles in the major axis of the galaxy for edge-on (*left*) and the face-on (*right*) views of the two models (*red*: non-flared model, *blue*: flared model). *Blue vertical solid line*: flare radius $R_{\mathrm{flare}}$. *Black dot-dashed line*: bulge component. *Solid blue and red lines*: linear fits performed to the inner regions of the disc for the flared and non-flared models. *Dashed blue and red lines*: The same for the outer regions of the disc. *Black dotted lines*: Bulge [$r_\mathrm{e}$]{} and [$\mu_\mathrm{e}$]{}values. *Horizontal dashed black line*: Limiting magnitude $\mu_{lim}=25$ mag arcsec$^{\mathrm{-2}}$. *Vertical dashed magenta line*: Inner limit for the inner profile. We note that the disc profile in the flared case exhibits a clear Type-II break in the edge-on view, whereas the non-flared model keeps the exponential profile that both models exhibit in the face-on view. See the legend for details.[]{data-label="fig:Ima_profiles"}](IMAGES/TRON100531_images.png "fig:"){width="0.95\linewidth"}\ ![**Top panels:** Edge-on $I$-band images for the flared and non-flared models. **Intermediate panels:** The same for the face-on discs. **Bottom panels:** Surface brightness profiles in the major axis of the galaxy for edge-on (*left*) and the face-on (*right*) views of the two models (*red*: non-flared model, *blue*: flared model). *Blue vertical solid line*: flare radius $R_{\mathrm{flare}}$. *Black dot-dashed line*: bulge component. *Solid blue and red lines*: linear fits performed to the inner regions of the disc for the flared and non-flared models. *Dashed blue and red lines*: The same for the outer regions of the disc. *Black dotted lines*: Bulge [$r_\mathrm{e}$]{} and [$\mu_\mathrm{e}$]{}values. *Horizontal dashed black line*: Limiting magnitude $\mu_{lim}=25$ mag arcsec$^{\mathrm{-2}}$. *Vertical dashed magenta line*: Inner limit for the inner profile. We note that the disc profile in the flared case exhibits a clear Type-II break in the edge-on view, whereas the non-flared model keeps the exponential profile that both models exhibit in the face-on view. See the legend for details.[]{data-label="fig:Ima_profiles"}](IMAGES/TRON100531_profiles.png "fig:"){width="0.95\linewidth"} We explored the range of [$\nabla_{R} h_{z}$]{} values derived by recent observational studies for galaxies of different types (see also Fig.\[fig:Str\_photo\_params\]) in the $I$ band: $0<{\ensuremath{\nabla_{R} h_{z}}}<0.15$ [@2014ApJ...787...24B; @2015MNRAS.451.2376M]. These values agree with those measured in the MW by several authors . @2015MNRAS.451.2376M argued that the scale heights derived from 1D decompositions may be biased to higher values than reality owing to bulge contamination (up to 10% in galaxies with $B/T\sim0.4$). Therefore, we restricted the sample from which we obtain the scale-height gradients to objects with $B/T < 0.4$ to avoid bulge contamination. The radius at which the flare starts ($R_{\mathrm{flare}}$) is randomly chosen following a uniform distribution between 2 and 5 times the scale length of the disc. The lowest value is set to avoid bulge contamination from the inner regions of the profile , while the highest value is the typical maximum radius where the limiting magnitude of both edge-on and face-on images is reached: $\mu_{lim}=25$ mag arcsec$^{\mathrm{-2}}$. At higher radial distances the flare would not have any significant effect on our profiles. Finally, once a set of parameters describing a realistic galaxy model is created, we check that the chosen values of $M_{abs,disc}$ and $M_{abs,bulge}$ provide a $B/T$ ratio within the observed distribution of $B/T$ values for the stellar mass and morphological type of the galaxy. We created a non-flared analogue of each flared model in order to compare their edge-on profiles. The 3D models of luminosity distribution in the $I$ band have a spatial resolution of 1 arcsec, with a total size of 301 arcsec in each direction. The adopted resolution is similar to the typical FWHM achieved with ground-based seeing-limited observations. We assumed a distance of $100$ Mpc to the simulated object, which is the upper limit of the distance distribution of the galaxies in @2001AJ....121..820G. Owing to the increase in the thickness of the disc in the flared models, the simulations might suffer from flux loss because of the limited box size. To prevent this, we automatically removed any models with differences greater than 1% in total flux between the flared and non-flared models. We created face-on and edge-on density images by integrating along the line of sight. The images are then convolved with a Gaussian kernel of $\sigma = 1$ arcsec. The surface brightness profile for each galaxy model has been derived for edge-on and face-on views, by locating a 1 arcsec wide single slit along the major axis in the first case and with the task in IRAF in the second. We included Poissonian noise across the whole profile. We assumed Planck cosmology when required [$\Omega_{M} = 0.31, \Omega_{\Lambda}=0.69$, H$_{\ensuremath{0}}$ = 67.8 km s$^{\ensuremath{-1}}$ Mpc$^{\ensuremath{-1}}$; @2015arXiv150201589P]. We took the model of the MW presented in as reference. These authors fitted SDSS-SEGUE MW data to a flared thin+thick disc model, obtaining compatible and deeper results than previous studies . They provide the 3D distribution, [$h_{z}$]{} distribution, as well as the fitted parameters for the thin and thick discs and both flares. We built up this model following the described procedure. For this model, we also created a non-flared analogue MW model by assuming the scale height at the sun’s radius to be constant for the whole disc. Results and discussion ====================== In Fig.\[fig:Ima\_profiles\] we present one example of the realistic flared and non-flared models that we have created. The top and middle panels represent the simulated images for the flared and non-flared model of a galaxy with a stellar mass of $10^{\ensuremath{10}} M_{\odot}$ and type Sd-Sdm, for edge-on and face-on views. The lower panels show the surface brightness profiles for the two models and for both inclinations (see the caption for details). The main result is that there is no significant difference between the flared and non-flared models in face-on orientation, either in the images or in the surface brightness profiles. On the contrary, when we compare the models in edge-on orientation, we clearly detect a break in the surface brightness profile of the flared model that does not appear in the non-flared analogue. We have found detectable breaks in the edge-on profiles of $\sim42\%$ of all our models including a flare, with a mean strength[^2] of $S=-0.15$, and 98% of them with $-0.21 < S < -0.11$. We obtain higher detection rates in brighter discs (i.e. higher masses), and in those models with values for ${\ensuremath{\nabla_{R} h_{z}}}$ greater than $\sim 0.05$. This is expected if a fixed limiting magnitude and spatial resolution is assumed. This means that realistic flares can frequently produce noticeable Type-II breaks in highly inclined galaxies, contrary to previous claims. Concerning the MW reference model, we find a significant drop in its surface brightness profile at $8.58^{+0.43}_{-0.31}$ kpc in the edge-on profile in the flared model, which is not observed either in the face-on profile or in the edge-on profile of its non-flared analogue. We obtained a break strength $S=-0.22$ for the break of the MW flared model. Therefore, we conclude that a flare such as the one reported for the MW in many studies, would produce a Type-II profile if the MW were observed edge-on. ![Break radius normalized by the inner disc scale length vs. the break strength $S$. Our simulations are plotted with colour-coded circles according to [$\nabla_{R} h_{z}$]{}. The edge-on MW model is overplotted with a green diamond. The asterisks represent the Type-II breaks observed by @2014MNRAS.441.1992L from the S$^{4}$G survey. See the legend in the figure.](IMAGES/flare_laine_comp_rbrkhi_T.png){width="0.94\linewidth"} \[fig:Laine\_comp\] ![Histograms comparing the distribution of Type-II profiles from the present study and @2012MNRAS.427.1102M, as a function of the ratio between the inner and the outer scale length [$h_{\mathrm{i}}$]{}/[$h_{\mathrm{o}}$]{} (upper panel) and of the break radius normalized by the outer scale length [$R_\mathrm{brk II}$]{}/[$h_{\mathrm{o}}$]{} (lower panel). The vertical green lines mark the values for the MW reference model. See the legend in the figure.](IMAGES/flare_MN2012_hist.png){width="0.94\linewidth"} \[fig:MN2012\_comp\] In Fig.\[fig:Laine\_comp\] we represent the ratio between the break radius and the scale length of the inner disc (${\ensuremath{R_\mathrm{brk II}}}/ {\ensuremath{h_{\mathrm{i}}}}$) vs. the break strength ($S$). The flared models reproduce the observations by @2014MNRAS.441.1992L in the lower break strength part of the real distribution ($-0.25<S<-0.1$). We note that our 3D models are dust-free by construction. We can compare them because the authors use data in the $K$ and 3.6 microns bands; these data are negligibly affected by dust extinction. We find a significant correlation between $S$ and [$\nabla_{R} h_{z}$]{} ($\rho_{Spearman} = -0.32,$ p-value $< 10^{\mathrm{-5}}$), which is not found between $S$ and the morphological type (see Fig. A.1, available in the online version). In Fig.\[fig:MN2012\_comp\], we compare characteristics of the Type-II profiles from the observational sample by @2012MNRAS.427.1102M with those of our flared models. These authors identify two types of Type-II profiles: the *breaks*, which appear as down-bending transitions, and the *truncations*, much steeper and located much further out. In order to reproduce the latter type, often classified as *sharp* truncations, we would require significantly greater values of [$\nabla_{R} h_{z}$]{} than those reported in galaxies by current observational studies, including the MW. This contrasts with the *breaks*, whose distribution can be partially reproduced by flared discs, as demonstrated by our models. We have also studied the surface brightness profiles as a function of the distance from the galactic plane in the flared models that exhibit breaks in their edge-on profiles. We find an increase in the values of [$h_{\mathrm{i}}$]{} and [$h_{\mathrm{o}}$]{} with $z$ equal to factor of $1.15^{+0.23}_{-0.12}$ for the inner disc and $2.19^{+1.21}_{-0.72}$ for the outer disc up to a maximum height of $z/{\ensuremath{h_{z}}}\leq 5$, such that [$h_{\mathrm{o}}$]{} becomes closer to [$h_{\mathrm{i}}$]{}. This agrees with the results of @2007MNRAS.378..594P, where the authors found a significant weakening of the breaks with increasing distance from the plane for both inner and outer profiles, with factors of 1.1-1.4 and 1.4-2.4, respectively. *So we can conclude that the Type-II discs produced by flares in our simulations can reproduce the global properties of real edge-on Type-II discs with low to intermediate break strengths.* Conclusions =========== The present realistic 3D models of flared disc galaxies demonstrate that, contrary to previous claims, flares can produce Type-II profiles in galaxies when viewed edge-on, especially for breaks with intermediate to low strength ($-0.25<S<-0.05$). We also find a significant weakening of the breaks with the distance from the plane, which agrees with observations. We do not find any correlation between the existence of a significant bulge component and the presence of a Type-II break. The authors thank to the anonymous referee for the useful criticisms that helped to improve this publication significantly and Martín López-Corredoira for his kind support with the MW models. This research has been supported by the Ministerio de Economía y Competitividad del Gobierno de España (MINECO) under project AYA2012-31277, and by the Instituto de Astrofísica de Canarias under project P3/86. We acknowledge the usage of the HyperLeda database (http://leda.univ-lyon1.fr). This research has made use of the NASA Astrophysics Data System and NASA/IPAC Extragalactic Database (NED) and the $R$ environment for statistical computing. Online version material ======================= ![Break radius normalized by the inner disc scale length vs. the break strength $S$. Our simulations are plotted with coloured circles, colour-coded according to the morphological type of the object simulated object. The edge-on MW model is overplotted with a green diamond. The asterisks represent the Type-II breaks observed by @2014MNRAS.441.1992L from the S$^{4}$G survey. See the legend in the figure. \[Figure available in the online version.\]](IMAGES/flare_laine_comp_rbrkhi_T_Type.png){width="0.99\linewidth"} \[fig:Laine\_comp\_2\] [^1]: HyperLeda database available at: http://leda.univ-lyon1.fr/ [^2]: The strength of a disc break is defined as $S=\log_{10}{({\ensuremath{h_{\mathrm{o}}}}/{\ensuremath{h_{\mathrm{i}}}})}$, with [$h_{\mathrm{i}}$]{} and [$h_{\mathrm{o}}$]{} the scale lengths of the inner and outer discs.

--- abstract: 'We consider the possibility of constraining decaying dark matter by looking out through the Milky Way halo. Specifically we use [*Chandra*]{} blank sky observations to constrain the parameter space of sterile neutrinos. We find that a broad band in parameter space is still open, leaving the sterile neutrino as an excellent dark matter candidate.' author: - 'Signe Riemer-S[ø]{}rensen$^1$, Steen H. Hansen$^2$, and Kristian Pedersen$^1$' title: | Sterile neutrinos in the Milky Way:\ Observational constraints --- INTRODUCTION ============ The particle nature of the dark matter remains a mystery. This is contrasted with the firmly established dark matter abundance, which has been measured using the cosmic microwave background and large scale structure observations [@spergel; @uros]. A wide range of dark matter particle candidates have been proposed, including axions, the lightest supersymmetric particle, and sterile neutrinos. Recently, sterile neutrinos have gained renewed interest, since theoretical considerations have shown them to be a natural part of a minimally extended standard model, the $\nu$MSM. This $\nu$MSM, which is the standard model extended with only 3 sterile neutrinos, manages to explain the masses of the active neutrinos [@asaka0503065], the baryon asymmetry of the universe [@asaka0505013], and the abundance of dark matter [@dodelson]. Arguably, one of the most appealing aspects of sterile neutrinos as dark matter candidates is, that they will very likely either be detected or rejected during the next couple of years. This is because the sterile neutrinos decay virtually at rest into a photon and an active neutrino producing a sharp decay line at $E=m_s/2$ where $m_s$ is the mass of the sterile neutrinos [@warm]. This photon line can now be searched for in cosmic high energy data of the background radiation (for recent studies see @mapelli [@boyarsky]), towards massive structures like galaxy clusters [@tucker; @boyarsky06], or towards dark matter structures with a very low fraction of baryons [@hansen02]. The sterile neutrino was originally introduced in order to alleviate discrepancies between the theory of cold dark matter structure formation and observations [@dodelson; @warm], which indicated that the dark matter particle should be warm. Traditionally “warm” means that the mass of the dark matter particle should be in the keV range, in order to suppress the formation of small scale structures. While some of these problems have found other solutions, sterile neutrinos have found other uses, including an explanation for the pulsar peculiar velocities [@kusenko; @fuller], synthesising early star formation [@biermann], and, as mentioned above, as an explanation for the masses of the active neutrinos and the baryon asymmetry [@asaka0503065; @asaka0505013]. Also constraints on doube $\beta$ decays have been derived [@beruzkov]. We will here follow the idea of @hansen02 and look for a decay line towards a region with very little baryons. Specifically we will study X-ray “blank” regions in the sky with no known X-ray sources. By doing so we will search for a signal from the dark matter halo of our own Galaxy. We will see that the signal is very low for photons with energies of the order 0.1–10 keV, which allows us to exclude part of the sterile neutrino parameter space. ![The blank sky ACIS-S3 spectrum – i.e. a view through the Milky Way halo. Also shown is the maximal single gaussian emission line from decaying dark matter at a specific energy as discussed in [ Section  \[sec:analysis\]]{}.[]{data-label="fig:gaussian"}](./spectra_gauss.eps){width="49.00000%"} X-RAY DATA ANALYSIS {#sec:analysis} =================== We have analysed a 0.3–9 keV spectrum extracted from many combined [*Chandra*]{} ACIS-S3 blank sky observations[^1] (shown in [  \[fig:gaussian\]]{}) with CIAO 3.3[^2], and fitted spectral models with the spectral fitting package Xspec [@xspec]. The line features at 1.74 keV, 2.1–2.2 keV, and 7.48 keV are Si $K\alpha$, the Au $M\alpha\beta$ complex, and Ni $K\alpha$ respectively, originating from fluorescence of material in the telescope and focal plane (i.e. they are seen in spectra obtained when ACIS-S3 is stowed and not pointing at the sky[^3]). However, a decay line from dark matter could “hide” under these prominent lines. In this case the decay line should be redshifted by a factor of $1+z$ where $z$ is the redshift of the dark matter emitter. Dark matter concentrations located at different distances should thus give rise to lines at different observed energies. In order to test for this, we extracted spectra from ACIS-S3 data of two galaxy clusters, A383 and A478, at redshifts, $z=0.1883$ and $0.0881$, respectively. We optimized the ratio between expected dark matter signal and emission from hot intracluster gas by extracting spectra from the cluster outskirts. The X-ray emission from hot intracluster gas is proportional to the gas density squared, whereas the emission of photons from dark matter is directly proportional to the dark matter density. Hence, in the cluster outskirts the gas surface brightness falls off with projected cluster-centric distance, $b$ as, $S_{gas} \propto b^{-3}$ (assuming a typical $\beta$-model) whereas the dark matter surface brightness is expected to fall off as $S_{DM} \propto b^{-1}$ (assuming an isothermal sphere). Specifically, we have extracted spectra from an outer radial bin, defined by optimizing the ratio of matter in the field of view (proportional to the signal from decaying dark matter) to the X-ray emission from the intracluster gas (“noise”) described by a $\beta$-profile. The outer radius was chosen as the radius with the optimal ratio (which is very close to the edge of the ACIS-S3 chip) and the inner radius was chosen so the signal to noise ratio had a value of half the optimal value. No obvious line features was detected at the corresponding redshifted energies[^4]. We then proceeded to firmly constrain the flux from any dark matter decay line in the full energy interval spanned by the blank sky spectrum. The blank sky spectrum was well fitted (reduced $\chi ^2 = 1.1$ for 540 d.o.f.) by a composite model consisting of an exponential plus a power law plus four gaussians for the most prominent lines. A hypothetical mono-energetic emission line in the spectrum was represented by a gaussian, centered at the line energy, and with a width, $\sigma$, given by the instrument spectral resolution: $\sigma \approx 0.1$ keV for $0.3\, {\mathrm{keV}}\leq E < 6.0$ keV and $\sigma \approx 0.15$ keV for $6.0\, {\mathrm{keV}}\leq E < 8.0$ keV[^5]. (The velocity of the halo dark matter is producing negligible line broadening, $v/c \sim 10^{-3}-10^{-4}$). For each energy in steps of 0.05 keV, we defined a gaussian with the instrumental width and maximum at the model value of the fit to the broad-band spectrum. The flux of this gaussian was then calculated providing an upper limit of emission from decaying dark matter at the given energy. This is a model independent and very conservative analysis, since any X-ray emitters within the field of view, and background, contributes to the extracted spectrum. If we instead only allowed the decay signal from potential decaying dark matter to produce a bump above the “base line” spectrum, then we could improve the bounds by a large factor [@boyarsky]. However, this approach would depend sensitively on the robustness of modelling background and emissions from sources contributing to the “base line” spectrum. DECAY RATE OF STERILE NEUTRINOS IN THE MILKY WAY HALO ===================================================== The amount of dark matter from our Galaxy within the observed field of view is only a minute fraction of the total mass of the halo. We use the model for the Milky Way presented in @klypin, namely an NFW dark matter profile with $M_{vir} = 10^{12} M_\sun$, virial radius $r_{vir} = 258$ kpc, concentration $c=12$, and solar distance $R_\sun=8$ kpc. We then integrate out through the Milky Way halo to find the expected signal. Changing the details of this model does not significantly affrect our results (e.g. changing the inner density slope between $-1$ and zero, or changing the outer slope between $-3$ and $-4$, changes the predicted signal by less than a factor of 2). There is also an uncertainty, possibly as large as a factor of 2, arising from using the value of $M_{vir} = 10^{12} M_\sun$. The blank sky spectrum was extracted from the full angular opening of the [*Chandra*]{} ACIS-S3 chip which is $8.4 \arcmin$ corresponding to a mass within the field of view of $M_{fov}=1.5\times10^{-7}M_{vir}$ at a mean luminosity distance of $D_L = 35$ kpc, derived from the model mentioned above. The luminosity from decaying sterile neutrinos is $\mathcal{L}=E_{\gamma} N \Gamma _{\gamma}$, where $E_{\gamma}$ is the energy of the photons, $N$ is the number of sterile neutrinos particles in the field of view ($N = M_{fov}/m_s$) and $\Gamma _{\gamma}$ is the decay rate. The flux at a luminosity distance, $D_L$, is: $$\label{eqn:flux} F = \frac{\mathcal{L}}{4 \pi D_L^2}= \frac{E_{\gamma} N \Gamma _{\gamma}}{4\pi D_L^2} \, .$$ The observed flux, $F$, at a given energy ($E_{\gamma}=m_s/2$) yields a conservative upper limit on the flux from decaying sterile neutrinos, so [ eq. \[eqn:flux\]]{} can be rewritten as: $$\label{eqn:maxdecay} \Gamma _{\gamma} \leq \frac{8 \pi F D_L^2}{M_{fov}} \, . $$ From the blank sky flux, an upper limit on the decay rate has been derived as function of line energy and plotted in [  \[fig:decayrate\]]{}. ![The decay rate as function of photon energy. The solid line is the upper limit from the [*Chandra*]{} blank sky observations viewing out through the Milky Way halo.[]{data-label="fig:decayrate"}](./decayrate.eps){width="49.00000%"} CONFRONTING THE MODEL WITH DATA =============================== The neutrinos can decay via various channels, and the lifetime of the sterile neutrinos is given by [@barger:1995; @heavy]: $$\label{eqn:tau} \tau = \frac{1}{\Gamma _{tot}}=\frac{f(m_s)\times 10^{20}}{\left( m_s/{\mathrm{keV}}\right)^5 \sin^2(2\theta)} \, ,$$ $\sin^2(2\theta)$ is the mixing angle with the active neutrino, and $f(m_s)$ takes into account the open decay channels so that for $m_s < 1 {\mathrm{MeV}}$, where only the neutrino channel is open, $f(m_s)=0.86$. The mixing angle is given by [@dolgov:2002 eqs. 209, 210]: $$\begin{aligned} \sin ^2 (2\theta) &\approx& A \left( \frac{{\mathrm{keV}}}{m_s} \right)^2 \left( \frac{\Omega _{DM}}{0.3} \right)\left( \frac{h}{0.65} \right)^2 \nonumber \\ && S \left( \frac{g_*(T_{produced})}{10.75} \right)^{3/2} \, , \label{eqn:sin}\end{aligned}$$ where $g_*(T_{produced})$ is the number of relativistic degrees of freedom at the temperature at which the sterile neutrinos are produced. For neutrino masses of $m_s \approx$ keV, the neutrinos are produced near the QCD phase transition [@dodelson], making the distribution somewhat non-thermal [@warm]. The value of $g_*$ for such neutrinos is very conservatively considered to be between 10.75 and 20, depending on the details of the QCD phase transition. We will here use $g_*(T_{produced})=15$ as a reference value. A numerical calculation including the details of the QCD phase-transition [@kev06] has confirmed that the choice $g_*=15$ gives good agreement with the analytical results presented by @dolgov:2002. S is a free parameter taking into account that additional entropy may be produced after the sterile neutrinos have been created [@asaka06]. This parameter was suggested to be in the range between 1 and 100 [@asaka06]. $A$ is a constant depending on which active neutrino the sterile neutrinos are assumed to mix with. It takes the values $A_{se}=6.7\times 10^{-8}$ for $\nu _s$ mixing with $\nu _e$ and $A_{s\mu}=4.8\times 10^{-8}$ for $\nu _{\mu, \tau}$. The sterile neutrinos are assumed to account for all dark matter, i.e. $\Omega _{DM}=0.30$, and the Hubble parameter at the present time, $H_0$, given as $h=H_0/(100{\mathrm{km/s/Mpc}})$. @barger:1995 derived the branching ratio for the radiative decay, $\nu_s \rightarrow \nu_\alpha + \gamma$, to be $\Gamma _{\gamma}/\Gamma _{tot} = (27\alpha)/(8\pi) \approx 1/128$. Now the lifetime can be rewritten as: $$\begin{aligned} \tau &=& \left(\frac{E}{{\mathrm{keV}}} \right)^{-3} \left(\frac{g_*}{15} \right)^{-3/2} \left(\frac{\Omega _{DM}}{0.3} \right)^{-1} \left(\frac{h}{0.7} \right)^{-2} \nonumber \\ && \times \, \frac{f(m_s) 6.5\times10^{18} sec}{A S} \, .\end{aligned}$$ An upper limit on the decay rate is constrained from [ eq. \[eqn:maxdecay\]]{} and $$\tau = \frac{1}{\Gamma _{tot}} \approx \frac{1}{128 \Gamma _\gamma} \, .$$ The lower limit on the lifetime from the data and the model predictions are plotted in [  \[fig:lifetime\]]{} for various values of $S$ and $g_*$. We see from [  \[fig:lifetime\]]{} that the blank sky data (solid red line) is at least one order of magnitude less restrictive than the simplest sterile model predictions (hatched region). Models with significant entropy production, $S \approx 30$, are excluded in the mass range $4.2\, {\mathrm{keV}}< m < 7 {\mathrm{keV}}$. For $S>100$ we exclude $2\, {\mathrm{keV}}< m < 16 {\mathrm{keV}}$. These constraints are more restrictive than the ones obtained for the Coma cluster periphery [@boyarsky06] for masses less than approximately $8$ keV. ![The lifetime constrained from the flux of the [*Chandra*]{} blank sky data (red) and A383 (green). The $\nu$MSM prediction for $S=1$ and $g_*=10-20$ (hatched) and several variations of $S$ and $g_*$ (black) have been overplotted.[]{data-label="fig:lifetime"}](./lifetime_gaussian.eps){width="49.00000%"} OTHER CONSTRAINTS ================= The simplest models with $S=1$ are bounded from below, $m_s>2\, {\mathrm{keV}}$, using Ly$\alpha$ observations [@viel], and possibly even $m_s>14$ keV according to a recent analysis [@seljak]. These constraints are weakned when allowing for $S>1$ [@asaka06], since additional entropy dilutes the momentum of the sterile neutrinos. Upper limits on the sterile neutrino mass are derived from the flux of the diffuse photon background [@warm], and are of the order $m_s<15$ keV for $S=1$ [@mapelli; @boyarsky]. These bounds are strengthened for $S>1$, since the sterile neutrinos then will decay faster. A strict lower bound on the sterile neutrino mass arises since Big Bang nucleosynthesis only allows approximately $0.3$ additional relativistic neutrino species to be populated at $\sim$ MeV temperatures. This lower bound is approximately $m_s>50$ eV. The Tremaine-Gunn bound applied to dwarf speroidals give $m_s>0.5$ keV [@tremaine; @lin]. The strong claim of an upper mass limit presented in [@tucker], under-estimated the flux from Virgo cluster by 2 orders of magnitude, and is unreliable [@boyarsky06]. Comparing these existing bounds with our findings, we conclude that a broad band in neutrino parameter space is still open as shown in [  \[fig:m\_S\]]{}. Taking the results of [@seljak] into account, then only masses above approximately $5$ keV are allowed. Ignoring the results of [@seljak], then another band opens in parameter space, for small masses and large entropy production. ![The $m_s-S$ parameter space with constraints from the [*Chandra*]{} blank sky data (red, fat, diagonal, this paper), Ly-$\alpha$ observations (blue, diagonal, conservative [@viel], and green, vertical, ambitious [@seljak]), Tremaine-Gunn bound $m>0.5$ keV (black, horizontal). []{data-label="fig:m_S"}](./m_S.eps){width="49.00000%"} CONCLUSIONS =========== We have used [*[Chandra]{}*]{} blank sky observations, i.e. viewing out through the Milky Way halo, to search for a possible decay line from dark matter particles. No obvious decay line was identified. The blank sky 0.3–9 keV X-ray flux is used to constrain the parameter space of sterile neutrinos, predicted to decay at rest to a detectable photon and an active neutrino. We find that the entropy production must be limited, $S<100$ for the mass range, $2\, {\mathrm{keV}}< m_s < 16$ keV, and even $S<10$ for masses near $m_s=10$ keV. This leaves a wide allowed space in parameter space, and the sterile neutrino is therefore still a viable dark matter candidate. If all Ly$\alpha$ constraints are considered, then it appears that the sterile neutrino is either rather massive, or its momenta are diluted by the entropy production, rendering it a cold dark matter candidate. If some of the Ly$\alpha$ constraints are relaxed, then there is a broad allowed band for smaller masses where the sterile neutrino is still a warm dark matter candidate. Our analysis is conservative and assumes that the entire blank sky signal at a given energy is arising from a decaying neutrino. By looking for decay lines above the expected blank sky emission, the results obtained here may be improved significantly. SHH is pleased to thank Ben Moore for inspiring discussions. We thank Alex Kusenko for useful comments. The Dark Cosmology Centre is funded by the Danish National Research Foundation. SHH is supported by the Swiss National Foundation. This research has made use of data obtained through the High Energy Astrophysics Science Archive Research Center Online Service, provided by the NASA/Goddard Space Flight Center.\ [*Note added:*]{}\ After finishing this paper we became aware of an independent analysis, where the possibility of measuring decaying dark matter particles from our Galaxy, using the blank sky observations with [*XMM-Newton*]{}, was considered [@bnrst2006]. The conclusions reached in that paper are qualitatively similar to ours.\ Facilities: Abazajian, K., 2005, astro-ph/0511630 Abazajian, K., Fuller G. M. & Tucker, W. H. 2001, Astrophys. J., 562, 593 Arnaud, K.A. 1996, in ASP Conf. Ser. 101: Astronomical Data Analysis Software and Systems V, 17 Asaka, T. Blanchet, S. & Shaposhnikov, M. 2005, Phys. Lett. B631, 151 Asaka, T. & Shaposhnikov, M. 2005, Phys. Lett., B620, 17 Asaka, T., Kusenko, A. & Shaposhnikov, M. 2006, arXiv:hep-ph/0602150 Barger, V. D., Phillips, R. J. N., Sarkar, S 1995, Phys. Lett. B, 352, 365 Bezrukov, F. 2005, Phys. Rev. D 72, 071303 Biermann, P. L. & Kusenko, A. 2006, Phys. Rev. Lett., 96, 091301 Boyarsky, A., Neronov, A., Ruchayskiy, O. & Shaposhnikov, M. 2005 arXiv:astro-ph/0512509. Boyarsky, A., Neronov, A., Ruchayskiy, O. & Shaposhnikov, M. 2006, arXiv:astro-ph/0603368. Boyarsky A., Neronov A., Ruchayskiy O., Shaposhnikov M. & Tkachev I. 2006, astro-ph/0603660 Dodelson S. & Widrow L. M. 1994, Phys. Rev. Lett. 72, 17 \[arXiv:hep-ph/9303287\]. Dolgov, A. D., Hansen, S. H., Raffelt, G. & Semikoz, D. V. 2000, Nucl. Phys. B590, 562 Dolgov, A. D. & Hansen, S. H. 2002 Astropart. Phys., 16, 339 Dolgov, A.D. 2002, Phys. Rept. 370, 333 Fuller G. M., Kusenko A., Mocioiu I. & Pascoli S. 2003, Phys. Rev. D 68, 103002 Hansen S. H., Lesgourgues, J., Pastor, S. & Silk, J. 2002 MNRAS, 333, 544 Klypin, A., Zhao, H. & Somerville, R. S. 2002, ApJ, 573, 597 Kusenko, A. & Segre, Y. 1997, Phys. Lett. B396, 197 Lin D. N. C. & Faber S. M. 1983, Astrophys. J. 266, L21 Mapelli, M. & Ferrara, A. 2005, MNRAS, 364, 2 Seljak, U. et al. 2005, Phys. Rev. D 71, 103515 Seljak, U. 2006, arXiv:astro-ph/0602430 Spergel, D. N. et al. 2003, astro-ph/0603449 Tremaine S. & Gunn J. E. 1979, Phys. Rev. Lett. 42, 407 Viel, M. 2005, Phys. Rev. D 71, 063534 Vikhlinin A. et al. 2005, Astrophys. J. 640, 691 [^1]: http://cxc.harvard.edu/contrib/maxim [^2]: http://cxc.harvard.edu/ciao [^3]: http://cxc.harvard.edu/proposer/POG/html/node1.html [^4]: After finisihing this analysis a related paper appeared [@boyarsky06] where an analysis using exactly this method of looking at the outer cluster region was made. [^5]: http://acis.mit.edu/acis/spect/spect.html